diff options
Diffstat (limited to 'erts/emulator/beam')
63 files changed, 4693 insertions, 2210 deletions
diff --git a/erts/emulator/beam/atom.c b/erts/emulator/beam/atom.c index b97705ed96..d7c7f117cf 100644 --- a/erts/emulator/beam/atom.c +++ b/erts/emulator/beam/atom.c @@ -1,7 +1,7 @@ /* * %CopyrightBegin% * - * Copyright Ericsson AB 1996-2010. All Rights Reserved. + * Copyright Ericsson AB 1996-2011. All Rights Reserved. * * The contents of this file are subject to the Erlang Public License, * Version 1.1, (the "License"); you may not use this file except in @@ -75,7 +75,7 @@ void atom_info(int to, void *to_arg) index_info(to, to_arg, &erts_atom_table); #ifdef ERTS_ATOM_PUT_OPS_STAT erts_print(to, to_arg, "atom_put_ops: %ld\n", - erts_smp_atomic_read(&atom_put_ops)); + erts_smp_atomic_read_nob(&atom_put_ops)); #endif if (lock) @@ -213,7 +213,7 @@ am_atom_put(const char* name, int len) len = MAX_ATOM_LENGTH; } #ifdef ERTS_ATOM_PUT_OPS_STAT - erts_smp_atomic_inc(&atom_put_ops); + erts_smp_atomic_inc_nob(&atom_put_ops); #endif a.len = len; a.name = (byte*)name; @@ -309,7 +309,7 @@ init_atom_table(void) rwmtx_opt.lived = ERTS_SMP_RWMTX_LONG_LIVED; #ifdef ERTS_ATOM_PUT_OPS_STAT - erts_smp_atomic_init(&atom_put_ops, 0); + erts_smp_atomic_init_nob(&atom_put_ops, 0); #endif erts_smp_rwmtx_init_opt(&atom_table_lock, &rwmtx_opt, "atom_tab"); diff --git a/erts/emulator/beam/atom.names b/erts/emulator/beam/atom.names index 68d64fb7b0..e7308dbf43 100644 --- a/erts/emulator/beam/atom.names +++ b/erts/emulator/beam/atom.names @@ -156,6 +156,8 @@ atom cr atom crlf atom creation atom current_function +atom current_location +atom current_stacktrace atom data atom debug_flags atom delay_trap diff --git a/erts/emulator/beam/beam_bif_load.c b/erts/emulator/beam/beam_bif_load.c index d76a7d8e9f..2561d7a630 100644 --- a/erts/emulator/beam/beam_bif_load.c +++ b/erts/emulator/beam/beam_bif_load.c @@ -162,6 +162,23 @@ BIF_RETTYPE code_make_stub_module_3(BIF_ALIST_3) return res; } +BIF_RETTYPE +check_old_code_1(BIF_ALIST_1) +{ + Module* modp; + + if (is_not_atom(BIF_ARG_1)) { + BIF_ERROR(BIF_P, BADARG); + } + modp = erts_get_module(BIF_ARG_1); + if (modp == NULL) { /* Doesn't exist. */ + BIF_RET(am_false); + } else if (modp->old_code == NULL) { /* No old code. */ + BIF_RET(am_false); + } + BIF_RET(am_true); +} + Eterm check_process_code_2(BIF_ALIST_2) { @@ -175,6 +192,13 @@ check_process_code_2(BIF_ALIST_2) Eterm res; if (internal_pid_index(BIF_ARG_1) >= erts_max_processes) goto error; + modp = erts_get_module(BIF_ARG_2); + if (modp == NULL) { /* Doesn't exist. */ + return am_false; + } else if (modp->old_code == NULL) { /* No old code. */ + return am_false; + } + #ifdef ERTS_SMP rp = erts_pid2proc_suspend(BIF_P, ERTS_PROC_LOCK_MAIN, BIF_ARG_1, ERTS_PROC_LOCK_MAIN); @@ -188,7 +212,6 @@ check_process_code_2(BIF_ALIST_2) ERTS_BIF_YIELD2(bif_export[BIF_check_process_code_2], BIF_P, BIF_ARG_1, BIF_ARG_2); } - modp = erts_get_module(BIF_ARG_2); res = check_process_code(rp, modp); #ifdef ERTS_SMP if (BIF_P != rp) { @@ -412,11 +435,6 @@ check_process_code(Process* rp, Module* modp) #endif #define INSIDE(a) (start <= (a) && (a) < end) - if (modp == NULL) { /* Doesn't exist. */ - return am_false; - } else if (modp->old_code == NULL) { /* No old code. */ - return am_false; - } /* * Pick up limits for the module. diff --git a/erts/emulator/beam/beam_bp.c b/erts/emulator/beam/beam_bp.c index 31910888d1..773baad01f 100644 --- a/erts/emulator/beam/beam_bp.c +++ b/erts/emulator/beam/beam_bp.c @@ -1,7 +1,7 @@ /* * %CopyrightBegin% * - * Copyright Ericsson AB 2000-2010. All Rights Reserved. + * Copyright Ericsson AB 2000-2011. All Rights Reserved. * * The contents of this file are subject to the Erlang Public License, * Version 1.1, (the "License"); you may not use this file except in @@ -408,7 +408,7 @@ erts_is_count_break(BeamInstr *pc, Sint *count_ret) { if (bdc) { if (count_ret) { - *count_ret = (Sint) erts_smp_atomic_read(&bdc->acount); + *count_ret = (Sint) erts_smp_atomic_read_nob(&bdc->acount); } return !0; } @@ -958,17 +958,17 @@ static int set_function_break(Module *modp, BeamInstr *pc, int bif, if (break_op == (BeamInstr) BeamOp(op_i_count_breakpoint)) { if (count_op == erts_break_stop) { - count = erts_smp_atomic_read(&bdc->acount); + count = erts_smp_atomic_read_nob(&bdc->acount); if (count >= 0) { while(1) { - res = erts_smp_atomic_cmpxchg(&bdc->acount, -count - 1, count); + res = erts_smp_atomic_cmpxchg_nob(&bdc->acount, -count - 1, count); if ((res == count) || count < 0) break; count = res; } } } else { /* Reset call counter */ - erts_smp_atomic_set(&bdc->acount, 0); + erts_smp_atomic_set_nob(&bdc->acount, 0); } } else if (break_op == (BeamInstr) BeamOp(op_i_time_breakpoint)) { @@ -1097,7 +1097,7 @@ static int set_function_break(Module *modp, BeamInstr *pc, int bif, } } else if (break_op == (BeamInstr) BeamOp(op_i_count_breakpoint)) { BpDataCount *bdc = (BpDataCount *) bd; - erts_smp_atomic_init(&bdc->acount, 0); + erts_smp_atomic_init_nob(&bdc->acount, 0); } if (bif == BREAK_IS_ERL) { diff --git a/erts/emulator/beam/beam_bp.h b/erts/emulator/beam/beam_bp.h index bd8a7249a7..2ec5818688 100644 --- a/erts/emulator/beam/beam_bp.h +++ b/erts/emulator/beam/beam_bp.h @@ -1,7 +1,7 @@ /* * %CopyrightBegin% * - * Copyright Ericsson AB 2000-2010. All Rights Reserved. + * Copyright Ericsson AB 2000-2011. All Rights Reserved. * * The contents of this file are subject to the Erlang Public License, * Version 1.1, (the "License"); you may not use this file except in @@ -165,8 +165,8 @@ do { \ bdc = (BpDataCount *) bdc->next; \ ASSERT(bdc); \ bds[ix] = (BpData *) bdc; \ - count = erts_smp_atomic_read(&bdc->acount); \ - if (count >= 0) erts_smp_atomic_inc(&bdc->acount); \ + count = erts_smp_atomic_read_nob(&bdc->acount); \ + if (count >= 0) erts_smp_atomic_inc_nob(&bdc->acount); \ *(instr_result) = bdc->orig_instr; \ } while (0) diff --git a/erts/emulator/beam/beam_emu.c b/erts/emulator/beam/beam_emu.c index fb90a7d4f7..5691f7aec1 100644 --- a/erts/emulator/beam/beam_emu.c +++ b/erts/emulator/beam/beam_emu.c @@ -1145,26 +1145,11 @@ void process_main(void) Eterm *tmp_big; /* Temporary buffer for small bignums if !HEAP_ON_C_STACK. */ #endif -#ifndef ERTS_SMP -#if !HALFWORD_HEAP - static Eterm save_reg[ERTS_X_REGS_ALLOCATED]; - /* X registers -- not used directly, but - * through 'reg', because using it directly - * needs two instructions on a SPARC, - * while using it through reg needs only - * one. - */ -#endif /* - * Floating point registers. - */ - static FloatDef freg[MAX_REG]; -#else - /* X regisers and floating point registers are located in + * X registers and floating point registers are located in * scheduler specific data. */ register FloatDef *freg; -#endif /* * For keeping the negative old value of 'reds' when call saving is active. @@ -1201,14 +1186,6 @@ void process_main(void) init_done = 1; goto init_emulator; } -#ifndef ERTS_SMP -#if !HALFWORD_HEAP - reg = save_reg; /* XXX: probably wastes a register on x86 */ -#else - /* Registers need to be heap allocated (correct memory range) for tracing to work */ - reg = erts_alloc(ERTS_ALC_T_BEAM_REGISTER, ERTS_X_REGS_ALLOCATED * sizeof(Eterm)); -#endif -#endif c_p = NULL; reds_used = 0; goto do_schedule1; @@ -1229,10 +1206,8 @@ void process_main(void) #endif ERTS_SMP_REQ_PROC_MAIN_LOCK(c_p); PROCESS_MAIN_CHK_LOCKS(c_p); -#ifdef ERTS_SMP - reg = c_p->scheduler_data->save_reg; - freg = c_p->scheduler_data->freg; -#endif + reg = ERTS_PROC_GET_SCHDATA(c_p)->x_reg_array; + freg = ERTS_PROC_GET_SCHDATA(c_p)->f_reg_array; #if !HEAP_ON_C_STACK tmp_big = ERTS_PROC_GET_SCHDATA(c_p)->beam_emu_tmp_heap; #endif @@ -3561,7 +3536,7 @@ void process_main(void) * Operands: NotUsed Live Dst */ do_bs_init_bits_known: - num_bytes = (num_bits+7) >> 3; + num_bytes = ((Uint64)num_bits+(Uint64)7) >> 3; if (num_bits & 7) { alloc += ERL_SUB_BIN_SIZE; } @@ -4888,92 +4863,6 @@ void process_main(void) } /* - * Instructions for allocating on the message area. - */ - - OpCase(i_global_cons): - { - BeamInstr *next; -#ifdef HYBRID - Eterm *hp; - - PreFetch(0,next); - TestGlobalHeap(2,2,hp); - hp[0] = r(0); - hp[1] = x(1); - r(0) = make_list(hp); -#ifndef INCREMENTAL - global_htop += 2; -#endif - NextPF(0,next); -#else - PreFetch(0,next); - c_p->freason = EXC_INTERNAL_ERROR; - goto find_func_info; -#endif - } - - OpCase(i_global_tuple): - { - BeamInstr *next; - int len; -#ifdef HYBRID - Eterm list; - Eterm *hp; -#endif - - if ((len = list_length(r(0))) < 0) { - goto badarg; - } - - PreFetch(0,next); -#ifdef HYBRID - TestGlobalHeap(len + 1,1,hp); - list = r(0); - r(0) = make_tuple(hp); - *hp++ = make_arityval(len); - while(is_list(list)) - { - Eterm* cons = list_val(list); - *hp++ = CAR(cons); - list = CDR(cons); - } -#ifndef INCREMENTAL - global_htop += len + 1; -#endif - NextPF(0,next); -#else - c_p->freason = EXC_INTERNAL_ERROR; - goto find_func_info; -#endif - } - - OpCase(i_global_copy): - { - BeamInstr *next; - PreFetch(0,next); -#ifdef HYBRID - if (!IS_CONST(r(0))) - { - BM_SWAP_TIMER(system,copy); - SWAPOUT; - reg[0] = r(0); - reg[1] = NIL; - r(0) = copy_struct_lazy(c_p,r(0),0); - ASSERT(ma_src_top == 0); - ASSERT(ma_dst_top == 0); - ASSERT(ma_offset_top == 0); - SWAPIN; - BM_SWAP_TIMER(copy,system); - } - NextPF(0,next); -#else - c_p->freason = EXC_INTERNAL_ERROR; - goto find_func_info; -#endif - } - - /* * New floating point instructions. */ @@ -5151,10 +5040,8 @@ void process_main(void) c_p->def_arg_reg[4] = -neg_o_reds; reg[0] = r(0); c_p = hipe_mode_switch(c_p, cmd, reg); -#ifdef ERTS_SMP - reg = c_p->scheduler_data->save_reg; - freg = c_p->scheduler_data->freg; -#endif + reg = ERTS_PROC_GET_SCHDATA(c_p)->x_reg_array; + freg = ERTS_PROC_GET_SCHDATA(c_p)->f_reg_array; ERL_BITS_RELOAD_STATEP(c_p); neg_o_reds = -c_p->def_arg_reg[4]; FCALLS = c_p->fcalls; @@ -5268,8 +5155,8 @@ void process_main(void) OpCase(int_code_end): OpCase(label_L): - OpCase(too_old_compiler): OpCase(on_load): + OpCase(line_I): erl_exit(1, "meta op\n"); /* @@ -5686,6 +5573,25 @@ expand_error_value(Process* c_p, Uint freason, Eterm Value) { * that c_p->ftrace will point to a cons cell which holds the given args * and the saved data (encoded as a bignum). * + * There is an issue with line number information. Line number + * information is associated with the address *before* an operation + * that may fail or be stored stored on the stack. But continuation + * pointers point after its call instruction, not before. To avoid + * finding the wrong line number, we'll need to adjust them so that + * they point at the beginning of the call instruction or inside the + * call instruction. Since its impractical to point at the beginning, + * we'll do the simplest thing and decrement the continuation pointers + * by one. + * + * Here is an example of what can go wrong. Without the adjustment + * of continuation pointers, the call at line 42 below would seem to + * be at line 43: + * + * line 42 + * call ... + * line 43 + * gc_bif ... + * * (It would be much better to put the arglist - when it exists - in the * error value instead of in the actual trace; e.g. '{badarg, Args}' * instead of using 'badarg' with Args in the trace. The arglist may @@ -5752,7 +5658,7 @@ save_stacktrace(Process* c_p, BeamInstr* pc, Eterm* reg, BifFunction bf, } /* Save second stack entry if CP is valid and different from pc */ if (depth > 0 && c_p->cp != 0 && c_p->cp != pc) { - s->trace[s->depth++] = c_p->cp; + s->trace[s->depth++] = c_p->cp - 1; depth--; } s->pc = NULL; @@ -5772,13 +5678,13 @@ save_stacktrace(Process* c_p, BeamInstr* pc, Eterm* reg, BifFunction bf, /* Save first stack entry */ ASSERT(c_p->cp); if (depth > 0) { - s->trace[s->depth++] = c_p->cp; + s->trace[s->depth++] = c_p->cp - 1; depth--; } s->pc = NULL; /* Ignore pc */ } else { if (depth > 0 && c_p->cp != 0 && c_p->cp != pc) { - s->trace[s->depth++] = c_p->cp; + s->trace[s->depth++] = c_p->cp - 1; depth--; } s->pc = pc; @@ -5793,24 +5699,31 @@ save_stacktrace(Process* c_p, BeamInstr* pc, Eterm* reg, BifFunction bf, } /* Save the actual stack trace */ + erts_save_stacktrace(c_p, s, depth); +} + +void +erts_save_stacktrace(Process* p, struct StackTrace* s, int depth) +{ if (depth > 0) { Eterm *ptr; BeamInstr *prev = s->depth ? s->trace[s->depth-1] : NULL; BeamInstr i_return_trace = beam_return_trace[0]; BeamInstr i_return_to_trace = beam_return_to_trace[0]; + /* * Traverse the stack backwards and add all unique continuation * pointers to the buffer, up to the maximum stack trace size. * * Skip trace stack frames. */ - ptr = c_p->stop; - if (ptr < STACK_START(c_p) - && (is_not_CP(*ptr)|| (*cp_val(*ptr) != i_return_trace && - *cp_val(*ptr) != i_return_to_trace)) - && c_p->cp) { - /* Can not follow cp here - code may be unloaded */ - BeamInstr *cpp = c_p->cp; + ptr = p->stop; + if (ptr < STACK_START(p) && + (is_not_CP(*ptr)|| (*cp_val(*ptr) != i_return_trace && + *cp_val(*ptr) != i_return_to_trace)) && + p->cp) { + /* Cannot follow cp here - code may be unloaded */ + BeamInstr *cpp = p->cp; if (cpp == beam_exception_trace || cpp == beam_return_trace) { /* Skip return_trace parameters */ ptr += 2; @@ -5819,7 +5732,7 @@ save_stacktrace(Process* c_p, BeamInstr* pc, Eterm* reg, BifFunction bf, ptr += 1; } } - while (ptr < STACK_START(c_p) && depth > 0) { + while (ptr < STACK_START(p) && depth > 0) { if (is_CP(*ptr)) { if (*cp_val(*ptr) == i_return_trace) { /* Skip stack frame variables */ @@ -5834,7 +5747,7 @@ save_stacktrace(Process* c_p, BeamInstr* pc, Eterm* reg, BifFunction bf, if (cp != prev) { /* Record non-duplicates only */ prev = cp; - s->trace[s->depth++] = cp; + s->trace[s->depth++] = cp - 1; depth--; } ptr++; @@ -5902,9 +5815,14 @@ build_stacktrace(Process* c_p, Eterm exc) { struct StackTrace* s; Eterm args; int depth; - BeamInstr* current; - Eterm Where = NIL; - Eterm *next_p = &Where; + FunctionInfo fi; + FunctionInfo* stk; + FunctionInfo* stkp; + Eterm res = NIL; + Uint heap_size; + Eterm* hp; + Eterm mfa; + int i; if (! (s = get_trace_from_exc(exc))) { return NIL; @@ -5923,64 +5841,56 @@ build_stacktrace(Process* c_p, Eterm exc) { * saved s->current should already contain the proper value. */ if (s->pc != NULL) { - current = find_function_from_pc(s->pc); + erts_lookup_function_info(&fi, s->pc, 1); + } else if (GET_EXC_INDEX(s->freason) == + GET_EXC_INDEX(EXC_FUNCTION_CLAUSE)) { + erts_lookup_function_info(&fi, s->current, 1); } else { - current = s->current; + erts_set_current_function(&fi, s->current); } + /* - * If current is still NULL, default to the initial function + * If fi.current is still NULL, default to the initial function * (e.g. spawn_link(erlang, abs, [1])). */ - if (current == NULL) { - current = c_p->initial; + if (fi.current == NULL) { + erts_set_current_function(&fi, c_p->initial); args = am_true; /* Just in case */ } else { args = get_args_from_exc(exc); } - depth = s->depth; - /* - * Add the {M,F,A} for the current function - * (where A is arity or [Argument]). + * Look up all saved continuation pointers and calculate + * needed heap space. */ - { - int i; - Eterm mfa; - Uint heap_size = 6*(depth+1); - Eterm* hp = HAlloc(c_p, heap_size); - Eterm* hp_end = hp + heap_size; - - if (args != am_true) { - /* We have an arglist - use it */ - mfa = TUPLE3(hp, current[0], current[1], args); - } else { - Eterm arity = make_small(current[2]); - mfa = TUPLE3(hp, current[0], current[1], arity); + depth = s->depth; + stk = stkp = (FunctionInfo *) erts_alloc(ERTS_ALC_T_TMP, + depth*sizeof(FunctionInfo)); + heap_size = fi.needed + 2; + for (i = 0; i < depth; i++) { + erts_lookup_function_info(stkp, s->trace[i], 1); + if (stkp->current) { + heap_size += stkp->needed + 2; + stkp++; } - hp += 4; - ASSERT(*next_p == NIL); - *next_p = CONS(hp, mfa, NIL); - next_p = &CDR(list_val(*next_p)); - hp += 2; + } - /* - * Finally, we go through the saved continuation pointers. - */ - for (i = 0; i < depth; i++) { - BeamInstr *fi = find_function_from_pc((BeamInstr *) s->trace[i]); - if (fi == NULL) continue; - mfa = TUPLE3(hp, fi[0], fi[1], make_small(fi[2])); - hp += 4; - ASSERT(*next_p == NIL); - *next_p = CONS(hp, mfa, NIL); - next_p = &CDR(list_val(*next_p)); - hp += 2; - } - ASSERT(hp <= hp_end); - HRelease(c_p, hp_end, hp); + /* + * Allocate heap space and build the stacktrace. + */ + hp = HAlloc(c_p, heap_size); + while (stkp > stk) { + stkp--; + hp = erts_build_mfa_item(stkp, hp, am_true, &mfa); + res = CONS(hp, mfa, res); + hp += 2; } - return Where; + hp = erts_build_mfa_item(&fi, hp, args, &mfa); + res = CONS(hp, mfa, res); + + erts_free(ERTS_ALC_T_TMP, (void *) stk); + return res; } diff --git a/erts/emulator/beam/beam_load.c b/erts/emulator/beam/beam_load.c index 57fe25453d..de4b32b238 100644 --- a/erts/emulator/beam/beam_load.c +++ b/erts/emulator/beam/beam_load.c @@ -158,6 +158,7 @@ typedef struct { #define LITERAL_CHUNK 6 #define ATTR_CHUNK 7 #define COMPILE_CHUNK 8 +#define LINE_CHUNK 9 #define NUM_CHUNK_TYPES (sizeof(chunk_types)/sizeof(chunk_types[0])) @@ -182,6 +183,7 @@ static Uint chunk_types[] = { MakeIffId('L', 'i', 't', 'T'), /* 6 */ MakeIffId('A', 't', 't', 'r'), /* 7 */ MakeIffId('C', 'I', 'n', 'f'), /* 8 */ + MakeIffId('L', 'i', 'n', 'e'), /* 9 */ }; /* @@ -231,6 +233,15 @@ struct string_patch { }; /* + * This structure associates a code offset with a source code location. + */ + +typedef struct { + int pos; /* Position in code */ + Uint32 loc; /* Location in source code */ +} LineInstr; + +/* * This structure contains all information about the module being loaded. */ @@ -325,6 +336,19 @@ typedef struct { Literal* literals; /* Array of literals. */ LiteralPatch* literal_patches; /* Operands that need to be patched. */ Uint total_literal_size; /* Total heap size for all literals. */ + + /* + * Line table. + */ + BeamInstr* line_item; /* Line items from the BEAM file. */ + int num_line_items; /* Number of line items. */ + LineInstr* line_instr; /* Line instructions */ + int num_line_instrs; /* Maximum number of line instructions */ + int current_li; /* Current line instruction */ + int* func_line; /* Mapping from function to first line instr */ + Eterm* fname; /* List of file names */ + int num_fnames; /* Number of filenames in fname table */ + int loc_size; /* Size of location info in bytes (2/4) */ } LoaderState; typedef struct { @@ -332,20 +356,43 @@ typedef struct { Eterm* func_tab[1]; /* Pointers to each function. */ } LoadedCode; -#define GetTagAndValue(Stp, Tag, Val) \ - do { \ - BeamInstr __w; \ - GetByte(Stp, __w); \ - Tag = __w & 0x07; \ - if ((__w & 0x08) == 0) { \ - Val = __w >> 4; \ - } else if ((__w & 0x10) == 0) { \ - Val = ((__w >> 5) << 8); \ - GetByte(Stp, __w); \ - Val |= __w; \ - } else { \ - if (!get_int_val(Stp, __w, &(Val))) goto load_error; \ - } \ +/* + * Layout of the line table. + */ + +#define MI_LINE_FNAME_PTR 0 +#define MI_LINE_LOC_TAB 1 +#define MI_LINE_LOC_SIZE 2 +#define MI_LINE_FUNC_TAB 3 + +#define LINE_INVALID_LOCATION (0) + +/* + * Macros for manipulating locations. + */ + +#define IS_VALID_LOCATION(File, Line) \ + ((unsigned) (File) < 255 && (unsigned) (Line) < ((1 << 24) - 1)) +#define MAKE_LOCATION(File, Line) (((File) << 24) | (Line)) +#define LOC_FILE(Loc) ((Loc) >> 24) +#define LOC_LINE(Loc) ((Loc) & ((1 << 24)-1)) + +#define GetTagAndValue(Stp, Tag, Val) \ + do { \ + BeamInstr __w; \ + GetByte(Stp, __w); \ + Tag = __w & 0x07; \ + if ((__w & 0x08) == 0) { \ + Val = __w >> 4; \ + } else if ((__w & 0x10) == 0) { \ + Val = ((__w >> 5) << 8); \ + GetByte(Stp, __w); \ + Val |= __w; \ + } else { \ + int __res = get_tag_and_value(Stp, __w, (Tag), &(Val)); \ + if (__res < 0) goto load_error; \ + Tag = (unsigned) __res; \ + } \ } while (0) @@ -466,6 +513,7 @@ static int load_import_table(LoaderState* stp); static int read_export_table(LoaderState* stp); static int read_lambda_table(LoaderState* stp); static int read_literal_table(LoaderState* stp); +static int read_line_table(LoaderState* stp); static int read_code_header(LoaderState* stp); static int load_code(LoaderState* stp); static GenOp* gen_element(LoaderState* stp, GenOpArg Fail, GenOpArg Index, @@ -489,8 +537,8 @@ static void load_printf(int line, LoaderState* context, char *fmt, ...); static int transform_engine(LoaderState* st); static void id_to_string(Uint id, char* s); static void new_genop(LoaderState* stp); -static int get_int_val(LoaderState* stp, Uint len_code, BeamInstr* result); -static int get_erlang_integer(LoaderState* stp, Uint len_code, BeamInstr* result); +static int get_tag_and_value(LoaderState* stp, Uint len_code, + unsigned tag, BeamInstr* result); static int new_label(LoaderState* stp); static void new_literal_patch(LoaderState* stp, int pos); static void new_string_patch(LoaderState* stp, int pos); @@ -504,6 +552,8 @@ static Eterm native_addresses(Process* p, Eterm mod); int patch_funentries(Eterm Patchlist); int patch(Eterm Addresses, Uint fe); static int safe_mul(UWord a, UWord b, UWord* resp); +static void lookup_loc(FunctionInfo* fi, BeamInstr* pc, + BeamInstr* modp, int idx); static int must_swap_floats; @@ -633,6 +683,25 @@ bin_load(Process *c_p, ErtsProcLocks c_p_locks, } /* + * Initialize code area. + */ + state.code_buffer_size = erts_next_heap_size(2048 + state.num_functions, 0); + state.code = (BeamInstr *) erts_alloc(ERTS_ALC_T_CODE, + sizeof(BeamInstr) * state.code_buffer_size); + + state.code[MI_NUM_FUNCTIONS] = state.num_functions; + state.ci = MI_FUNCTIONS + state.num_functions + 1; + + state.code[MI_ATTR_PTR] = 0; + state.code[MI_ATTR_SIZE] = 0; + state.code[MI_ATTR_SIZE_ON_HEAP] = 0; + state.code[MI_COMPILE_PTR] = 0; + state.code[MI_COMPILE_SIZE] = 0; + state.code[MI_COMPILE_SIZE_ON_HEAP] = 0; + state.code[MI_NUM_BREAKPOINTS] = 0; + + + /* * Read the atom table. */ @@ -677,6 +746,18 @@ bin_load(Process *c_p, ErtsProcLocks c_p_locks, } /* + * Read the line table (if present). + */ + + CHKBLK(ERTS_ALC_T_CODE,state.code); + if (state.chunks[LINE_CHUNK].size > 0) { + define_file(&state, "line table", LINE_CHUNK); + if (!read_line_table(&state)) { + goto load_error; + } + } + + /* * Load the code chunk. */ @@ -784,6 +865,22 @@ bin_load(Process *c_p, ErtsProcLocks c_p_locks, state.genop_blocks = next; } + if (state.line_item != 0) { + erts_free(ERTS_ALC_T_LOADER_TMP, state.line_item); + } + + if (state.line_instr != 0) { + erts_free(ERTS_ALC_T_LOADER_TMP, state.line_instr); + } + + if (state.func_line != 0) { + erts_free(ERTS_ALC_T_LOADER_TMP, state.func_line); + } + + if (state.fname != 0) { + erts_free(ERTS_ALC_T_LOADER_TMP, state.fname); + } + return rval; } @@ -814,6 +911,10 @@ init_state(LoaderState* stp) stp->string_patches = 0; stp->may_load_nif = 0; stp->on_load = 0; + stp->line_item = 0; + stp->line_instr = 0; + stp->func_line = 0; + stp->fname = 0; } static int @@ -1303,6 +1404,138 @@ read_literal_table(LoaderState* stp) return 0; } +static int +read_line_table(LoaderState* stp) +{ + unsigned version; + unsigned flags; + int num_line_items; + BeamInstr* lp; + int i; + BeamInstr fname_index; + BeamInstr tag; + + /* + * If the emulator flag ignoring the line information was given, + * return immediately. + */ + + if (erts_no_line_info) { + return 1; + } + + /* + * Check version of line table. + */ + + GetInt(stp, 4, version); + if (version != 0) { + /* + * Wrong version. Silently ignore the line number chunk. + */ + return 1; + } + + /* + * Read the remaining header words. The flag word is reserved + * for possible future use; for the moment we ignore it. + */ + GetInt(stp, 4, flags); + GetInt(stp, 4, stp->num_line_instrs); + GetInt(stp, 4, num_line_items); + GetInt(stp, 4, stp->num_fnames); + + /* + * Calculate space and allocate memory for the line item table. + */ + + num_line_items++; + lp = (BeamInstr *) erts_alloc(ERTS_ALC_T_LOADER_TMP, + num_line_items * sizeof(BeamInstr)); + stp->line_item = lp; + stp->num_line_items = num_line_items; + + /* + * The zeroth entry in the line item table is special. + * It contains the undefined location. + */ + + *lp++ = LINE_INVALID_LOCATION; + num_line_items--; + + /* + * Read all the line items. + */ + + stp->loc_size = stp->num_fnames ? 4 : 2; + fname_index = 0; + while (num_line_items-- > 0) { + BeamInstr val; + BeamInstr loc; + + GetTagAndValue(stp, tag, val); + if (tag == TAG_i) { + if (IS_VALID_LOCATION(fname_index, val)) { + loc = MAKE_LOCATION(fname_index, val); + } else { + /* + * Too many files or huge line number. Silently invalidate + * the location. + */ + loc = LINE_INVALID_LOCATION; + } + *lp++ = loc; + if (val > 0xFFFF) { + stp->loc_size = 4; + } + } else if (tag == TAG_a) { + if (val > stp->num_fnames) { + LoadError2(stp, "file index overflow (%d/%d)", + val, stp->num_fnames); + } + fname_index = val; + num_line_items++; + } else { + LoadError1(stp, "bad tag '%c' (expected 'a' or 'i')", + tag_to_letter[tag]); + } + } + + /* + * Read all filenames. + */ + + if (stp->num_fnames != 0) { + stp->fname = (Eterm *) erts_alloc(ERTS_ALC_T_LOADER_TMP, + stp->num_fnames * + sizeof(Eterm)); + for (i = 0; i < stp->num_fnames; i++) { + byte* fname; + Uint n; + + GetInt(stp, 2, n); + GetString(stp, fname, n); + stp->fname[i] = am_atom_put((char*)fname, n); + } + } + + /* + * Allocate the arrays to be filled while code is being loaded. + */ + stp->line_instr = (LineInstr *) erts_alloc(ERTS_ALC_T_LOADER_TMP, + stp->num_line_instrs * + sizeof(LineInstr)); + stp->current_li = 0; + stp->func_line = (int *) erts_alloc(ERTS_ALC_T_LOADER_TMP, + stp->num_functions * + sizeof(int)); + + return 1; + + load_error: + return 0; +} + static int read_code_header(LoaderState* stp) @@ -1361,24 +1594,6 @@ read_code_header(LoaderState* stp) #endif } - /* - * Initialize code area. - */ - stp->code_buffer_size = erts_next_heap_size(2048 + stp->num_functions, 0); - stp->code = (BeamInstr *) erts_alloc(ERTS_ALC_T_CODE, - sizeof(BeamInstr) * stp->code_buffer_size); - - stp->code[MI_NUM_FUNCTIONS] = stp->num_functions; - stp->ci = MI_FUNCTIONS + stp->num_functions + 1; - - stp->code[MI_ATTR_PTR] = 0; - stp->code[MI_ATTR_SIZE] = 0; - stp->code[MI_ATTR_SIZE_ON_HEAP] = 0; - stp->code[MI_COMPILE_PTR] = 0; - stp->code[MI_COMPILE_SIZE] = 0; - stp->code[MI_COMPILE_SIZE_ON_HEAP] = 0; - stp->code[MI_NUM_BREAKPOINTS] = 0; - stp->new_bs_put_strings = 0; stp->catches = 0; return 1; @@ -1412,7 +1627,7 @@ load_code(LoaderState* stp) { int i; int ci; - int last_func_start = 0; + int last_func_start = 0; /* Needed by nif loading and line instructions */ char* sign; int arg; /* Number of current argument. */ int num_specific; /* Number of specific ops for current. */ @@ -1425,6 +1640,14 @@ load_code(LoaderState* stp) GenOp** last_op_next = NULL; int arity; + /* + * The size of the loaded func_info instruction is needed + * by both the nif functionality and line instructions. + */ + enum { + FUNC_INFO_SZ = 5 + }; + code = stp->code; code_buffer_size = stp->code_buffer_size; ci = stp->ci; @@ -1470,46 +1693,15 @@ load_code(LoaderState* stp) last_op->arity = 0; ASSERT(arity <= MAX_OPARGS); -#define GetValue(Stp, First, Val) \ - do { \ - if (((First) & 0x08) == 0) { \ - Val = (First) >> 4; \ - } else if (((First) & 0x10) == 0) { \ - BeamInstr __w; \ - GetByte(Stp, __w); \ - Val = (((First) >> 5) << 8) | __w; \ - } else { \ - if (!get_int_val(Stp, (First), &(Val))) goto load_error; \ - } \ - } while (0) - for (arg = 0; arg < arity; arg++) { - BeamInstr first; - - GetByte(stp, first); - last_op->a[arg].type = first & 0x07; + GetTagAndValue(stp, last_op->a[arg].type, last_op->a[arg].val); switch (last_op->a[arg].type) { case TAG_i: - if ((first & 0x08) == 0) { - last_op->a[arg].val = first >> 4; - } else if ((first & 0x10) == 0) { - BeamInstr w; - GetByte(stp, w); - ASSERT(first < 0x800); - last_op->a[arg].val = ((first >> 5) << 8) | w; - } else { - int i = get_erlang_integer(stp, first, &(last_op->a[arg].val)); - if (i < 0) { - goto load_error; - } - last_op->a[arg].type = i; - } - break; case TAG_u: - GetValue(stp, first, last_op->a[arg].val); + case TAG_q: + case TAG_o: break; case TAG_x: - GetValue(stp, first, last_op->a[arg].val); if (last_op->a[arg].val == 0) { last_op->a[arg].type = TAG_r; } else if (last_op->a[arg].val >= MAX_REG) { @@ -1518,7 +1710,6 @@ load_code(LoaderState* stp) } break; case TAG_y: - GetValue(stp, first, last_op->a[arg].val); if (last_op->a[arg].val >= MAX_REG) { LoadError1(stp, "invalid y register number: %u", last_op->a[arg].val); @@ -1526,7 +1717,6 @@ load_code(LoaderState* stp) last_op->a[arg].val += CP_SIZE; break; case TAG_a: - GetValue(stp, first, last_op->a[arg].val); if (last_op->a[arg].val == 0) { last_op->a[arg].type = TAG_n; } else if (last_op->a[arg].val >= stp->num_atoms) { @@ -1536,7 +1726,6 @@ load_code(LoaderState* stp) } break; case TAG_f: - GetValue(stp, first, last_op->a[arg].val); if (last_op->a[arg].val == 0) { last_op->a[arg].type = TAG_p; } else if (last_op->a[arg].val >= stp->num_labels) { @@ -1544,7 +1733,6 @@ load_code(LoaderState* stp) } break; case TAG_h: - GetValue(stp, first, last_op->a[arg].val); if (last_op->a[arg].val > 65535) { LoadError1(stp, "invalid range for character data type: %u", last_op->a[arg].val); @@ -1552,11 +1740,9 @@ load_code(LoaderState* stp) break; case TAG_z: { - BeamInstr ext_tag; unsigned tag; - GetValue(stp, first, ext_tag); - switch (ext_tag) { + switch (last_op->a[arg].val) { case 0: /* Floating point number */ { Eterm* hp; @@ -1648,7 +1834,8 @@ load_code(LoaderState* stp) break; } default: - LoadError1(stp, "invalid extended tag %d", ext_tag); + LoadError1(stp, "invalid extended tag %d", + last_op->a[arg].val); break; } } @@ -1659,7 +1846,6 @@ load_code(LoaderState* stp) } last_op->arity++; } -#undef GetValue ASSERT(arity == last_op->arity); @@ -1701,14 +1887,6 @@ load_code(LoaderState* stp) } /* - * Special error message instruction. - */ - if (stp->genop->op == genop_too_old_compiler_0) { - LoadError0(stp, "please re-compile this module with an " - ERLANG_OTP_RELEASE " compiler"); - } - - /* * From the collected generic instruction, find the specific * instruction. */ @@ -1759,7 +1937,27 @@ load_code(LoaderState* stp) ERLANG_OTP_RELEASE " compiler "); } - LoadError0(stp, "no specific operation found"); + /* + * Some generic instructions should have a special + * error message. + */ + switch (stp->genop->op) { + case genop_too_old_compiler_0: + LoadError0(stp, "please re-compile this module with an " + ERLANG_OTP_RELEASE " compiler"); + case genop_unsupported_guard_bif_3: + { + Eterm Mod = (Eterm) stp->genop->a[0].val; + Eterm Name = (Eterm) stp->genop->a[1].val; + Uint arity = (Uint) stp->genop->a[2].val; + FREE_GENOP(stp, stp->genop); + stp->genop = 0; + LoadError3(stp, "unsupported guard BIF: %T:%T/%d\n", + Mod, Name, arity); + } + default: + LoadError0(stp, "no specific operation found"); + } } stp->specific_op = specific; @@ -2048,7 +2246,6 @@ load_code(LoaderState* stp) case op_i_func_info_IaaI: { Uint offset; - enum { FINFO_SZ = 5 }; if (function_number >= stp->num_functions) { LoadError1(stp, "too many functions in module (header said %d)", @@ -2056,27 +2253,37 @@ load_code(LoaderState* stp) } if (stp->may_load_nif) { - const int finfo_ix = ci - FINFO_SZ; + const int finfo_ix = ci - FUNC_INFO_SZ; enum { MIN_FUNC_SZ = 3 }; if (finfo_ix - last_func_start < MIN_FUNC_SZ && last_func_start) { /* Must make room for call_nif op */ int pad = MIN_FUNC_SZ - (finfo_ix - last_func_start); ASSERT(pad > 0 && pad < MIN_FUNC_SZ); CodeNeed(pad); - sys_memmove(&code[finfo_ix+pad], &code[finfo_ix], FINFO_SZ*sizeof(BeamInstr)); + sys_memmove(&code[finfo_ix+pad], &code[finfo_ix], + FUNC_INFO_SZ*sizeof(BeamInstr)); sys_memset(&code[finfo_ix], 0, pad*sizeof(BeamInstr)); ci += pad; stp->labels[last_label].value += pad; } } last_func_start = ci; + + /* + * Save current offset of into the line instruction array. + */ + + if (stp->func_line) { + stp->func_line[function_number] = stp->current_li; + } + /* * Save context for error messages. */ stp->function = code[ci-2]; stp->arity = code[ci-1]; - ASSERT(stp->labels[last_label].value == ci - FINFO_SZ); + ASSERT(stp->labels[last_label].value == ci - FUNC_INFO_SZ); offset = MI_FUNCTIONS + function_number; code[offset] = stp->labels[last_label].patches; stp->labels[last_label].patches = offset; @@ -2139,6 +2346,45 @@ load_code(LoaderState* stp) stp->catches = ci-3; break; + case op_line_I: + if (stp->line_item) { + BeamInstr item = code[ci-1]; + BeamInstr loc; + int li; + if (item >= stp->num_line_items) { + LoadError2(stp, "line instruction index overflow (%d/%d)", + item, stp->num_line_items); + } + li = stp->current_li; + if (li >= stp->num_line_instrs) { + LoadError2(stp, "line instruction table overflow (%d/%d)", + li, stp->num_line_instrs); + } + loc = stp->line_item[item]; + + if (ci - 2 == last_func_start) { + /* + * This line instruction directly follows the func_info + * instruction. Its address must be adjusted to point to + * func_info instruction. + */ + stp->line_instr[li].pos = last_func_start - FUNC_INFO_SZ; + stp->line_instr[li].loc = stp->line_item[item]; + stp->current_li++; + } else if (li <= stp->func_line[function_number-1] || + stp->line_instr[li-1].loc != loc) { + /* + * Only store the location if it is different + * from the previous location in the same function. + */ + stp->line_instr[li].pos = ci - 2; + stp->line_instr[li].loc = stp->line_item[item]; + stp->current_li++; + } + } + ci -= 2; /* Get rid of the instruction */ + break; + /* * End of code found. */ @@ -2175,6 +2421,8 @@ load_code(LoaderState* stp) #define no_fpe_signals(St) 0 #endif +#define never(St) 0 + /* * Predicate that tests whether a jump table can be used. */ @@ -2562,13 +2810,8 @@ should_gen_heap_bin(LoaderState* stp, GenOpArg Src) static int binary_too_big(LoaderState* stp, GenOpArg Size) { - return Size.type == TAG_u && ((Size.val >> (8*sizeof(Uint)-3)) != 0); -} - -static int -binary_too_big_bits(LoaderState* stp, GenOpArg Size) -{ - return Size.type == TAG_u && (((Size.val+7)/8) >> (8*sizeof(Uint)-3) != 0); + return Size.type == TAG_o || + (Size.type == TAG_u && ((Size.val >> (8*sizeof(Uint)-3)) != 0)); } static GenOp* @@ -3435,10 +3678,7 @@ gen_guard_bif1(LoaderState* stp, GenOpArg Fail, GenOpArg Live, GenOpArg Bif, BifFunction bf; NEW_GENOP(stp, op); - op->op = genop_i_gc_bif1_5; - op->arity = 5; - op->a[0] = Fail; - op->a[1].type = TAG_u; + op->next = NULL; bf = stp->import[Bif.val].bf; /* The translations here need to have a reverse counterpart in beam_emu.c:translate_gc_bif for error handling to work properly. */ @@ -3459,19 +3699,30 @@ gen_guard_bif1(LoaderState* stp, GenOpArg Fail, GenOpArg Live, GenOpArg Bif, } else if (bf == trunc_1) { op->a[1].val = (BeamInstr) (void *) erts_gc_trunc_1; } else { - abort(); + op->op = genop_unsupported_guard_bif_3; + op->arity = 3; + op->a[0].type = TAG_a; + op->a[0].val = stp->import[Bif.val].module; + op->a[1].type = TAG_a; + op->a[1].val = stp->import[Bif.val].function; + op->a[2].type = TAG_u; + op->a[2].val = stp->import[Bif.val].arity; + return op; } + op->op = genop_i_gc_bif1_5; + op->arity = 5; + op->a[0] = Fail; + op->a[1].type = TAG_u; op->a[2] = Src; op->a[3] = Live; op->a[4] = Dst; - op->next = NULL; return op; } /* - * This is used by the ops.tab rule that rewrites gc_bifs with two parameters + * This is used by the ops.tab rule that rewrites gc_bifs with two parameters. * The instruction returned is then again rewritten to an i_load instruction - * folowed by i_gc_bif2_jIId, to handle literals properly. + * followed by i_gc_bif2_jIId, to handle literals properly. * As opposed to the i_gc_bif1_jIsId, the instruction i_gc_bif2_jIId is * always rewritten, regardless of if there actually are any literals. */ @@ -3483,31 +3734,39 @@ gen_guard_bif2(LoaderState* stp, GenOpArg Fail, GenOpArg Live, GenOpArg Bif, BifFunction bf; NEW_GENOP(stp, op); - op->op = genop_ii_gc_bif2_6; - op->arity = 6; - op->a[0] = Fail; - op->a[1].type = TAG_u; + op->next = NULL; bf = stp->import[Bif.val].bf; /* The translations here need to have a reverse counterpart in beam_emu.c:translate_gc_bif for error handling to work properly. */ if (bf == binary_part_2) { op->a[1].val = (BeamInstr) (void *) erts_gc_binary_part_2; } else { - abort(); + op->op = genop_unsupported_guard_bif_3; + op->arity = 3; + op->a[0].type = TAG_a; + op->a[0].val = stp->import[Bif.val].module; + op->a[1].type = TAG_a; + op->a[1].val = stp->import[Bif.val].function; + op->a[2].type = TAG_u; + op->a[2].val = stp->import[Bif.val].arity; + return op; } + op->op = genop_ii_gc_bif2_6; + op->arity = 6; + op->a[0] = Fail; + op->a[1].type = TAG_u; op->a[2] = S1; op->a[3] = S2; op->a[4] = Live; op->a[5] = Dst; - op->next = NULL; return op; } /* - * This is used by the ops.tab rule that rewrites gc_bifs with three parameters + * This is used by the ops.tab rule that rewrites gc_bifs with three parameters. * The instruction returned is then again rewritten to a move instruction that * uses r[0] for temp storage, followed by an i_load instruction, - * folowed by i_gc_bif3_jIsId, to handle literals properly. Rewriting + * followed by i_gc_bif3_jIsId, to handle literals properly. Rewriting * always occur, as with the gc_bif2 counterpart. */ static GenOp* @@ -3518,18 +3777,27 @@ gen_guard_bif3(LoaderState* stp, GenOpArg Fail, GenOpArg Live, GenOpArg Bif, BifFunction bf; NEW_GENOP(stp, op); - op->op = genop_ii_gc_bif3_7; - op->arity = 7; - op->a[0] = Fail; - op->a[1].type = TAG_u; + op->next = NULL; bf = stp->import[Bif.val].bf; /* The translations here need to have a reverse counterpart in beam_emu.c:translate_gc_bif for error handling to work properly. */ if (bf == binary_part_3) { op->a[1].val = (BeamInstr) (void *) erts_gc_binary_part_3; } else { - abort(); + op->op = genop_unsupported_guard_bif_3; + op->arity = 3; + op->a[0].type = TAG_a; + op->a[0].val = stp->import[Bif.val].module; + op->a[1].type = TAG_a; + op->a[1].val = stp->import[Bif.val].function; + op->a[2].type = TAG_u; + op->a[2].val = stp->import[Bif.val].arity; + return op; } + op->op = genop_ii_gc_bif3_7; + op->arity = 7; + op->a[0] = Fail; + op->a[1].type = TAG_u; op->a[2] = S1; op->a[3] = S2; op->a[4] = S3; @@ -3609,6 +3877,7 @@ freeze_code(LoaderState* stp) Uint size; unsigned catches; Sint decoded_size; + Uint line_size; /* * Verify that there was a correct 'FunT' chunk if there were @@ -3619,13 +3888,19 @@ freeze_code(LoaderState* stp) LoadError0(stp, stp->lambda_error); } - /* * Calculate the final size of the code. */ - - size = (stp->ci * sizeof(BeamInstr)) + (stp->total_literal_size * sizeof(Eterm)) + - strtab_size + attr_size + compile_size; + if (stp->line_instr == 0) { + line_size = 0; + } else { + line_size = (MI_LINE_FUNC_TAB + (stp->num_functions + 1) + + (stp->current_li+1) + stp->num_fnames) * + sizeof(Eterm) + (stp->current_li+1) * stp->loc_size; + } + size = (stp->ci * sizeof(BeamInstr)) + + (stp->total_literal_size * sizeof(Eterm)) + + strtab_size + attr_size + compile_size + line_size; /* * Move the code to its final location. @@ -3713,15 +3988,66 @@ freeze_code(LoaderState* stp) } literal_end += stp->total_literal_size; } - + CHKBLK(ERTS_ALC_T_CODE,code); + /* - * Place the string table and, optionally, attributes, after the literal heap. + * If there is line information, place it here. */ - CHKBLK(ERTS_ALC_T_CODE,code); + if (stp->line_instr == 0) { + code[MI_LINE_TABLE] = (BeamInstr) 0; + str_table = (byte *) literal_end; + } else { + Eterm* line_tab = (Eterm *) literal_end; + Eterm* p; + int ftab_size = stp->num_functions; + int num_instrs = stp->current_li; + Eterm* first_line_item; + + code[MI_LINE_TABLE] = (BeamInstr) line_tab; + p = line_tab + MI_LINE_FUNC_TAB; + + first_line_item = (p + ftab_size + 1); + for (i = 0; i < ftab_size; i++) { + *p++ = (Eterm) (BeamInstr) (first_line_item + stp->func_line[i]); + } + *p++ = (Eterm) (BeamInstr) (first_line_item + num_instrs); + ASSERT(p == first_line_item); + for (i = 0; i < num_instrs; i++) { + *p++ = (Eterm) (BeamInstr) (code + stp->line_instr[i].pos); + } + *p++ = (Eterm) (BeamInstr) (code + stp->ci - 1); + + line_tab[MI_LINE_FNAME_PTR] = (Eterm) (BeamInstr) p; + memcpy(p, stp->fname, stp->num_fnames*sizeof(Eterm)); + p += stp->num_fnames; + + line_tab[MI_LINE_LOC_TAB] = (Eterm) (BeamInstr) p; + line_tab[MI_LINE_LOC_SIZE] = stp->loc_size; + if (stp->loc_size == 2) { + Uint16* locp = (Uint16 *) p; + for (i = 0; i < num_instrs; i++) { + *locp++ = (Uint16) stp->line_instr[i].loc; + } + *locp++ = LINE_INVALID_LOCATION; + str_table = (byte *) locp; + } else { + Uint32* locp = (Uint32 *) p; + ASSERT(stp->loc_size == 4); + for (i = 0; i < num_instrs; i++) { + *locp++ = stp->line_instr[i].loc; + } + *locp++ = LINE_INVALID_LOCATION; + str_table = (byte *) locp; + } + + CHKBLK(ERTS_ALC_T_CODE,code); + } - sys_memcpy(literal_end, stp->chunks[STR_CHUNK].start, strtab_size); + /* + * Place the string table and, optionally, attributes here. + */ + sys_memcpy(str_table, stp->chunks[STR_CHUNK].start, strtab_size); CHKBLK(ERTS_ALC_T_CODE,code); - str_table = (byte *) literal_end; if (attr_size) { byte* attr = str_table + strtab_size; sys_memcpy(attr, stp->chunks[ATTR_CHUNK].start, stp->chunks[ATTR_CHUNK].size); @@ -3938,6 +4264,7 @@ transform_engine(LoaderState* st) GenOp* instr; Uint* pc; int rval; + static Uint restart_fail[1] = {TOP_fail}; ASSERT(gen_opc[st->genop->op].transform != -1); pc = op_transform + gen_opc[st->genop->op].transform; @@ -3951,7 +4278,6 @@ transform_engine(LoaderState* st) ASSERT(restart != NULL); pc = restart; ASSERT(*pc < NUM_TOPS); /* Valid instruction? */ - ASSERT(*pc == TOP_try_me_else || *pc == TOP_fail); instr = st->genop; #define RETURN(r) rval = (r); goto do_return; @@ -3964,7 +4290,9 @@ transform_engine(LoaderState* st) op = *pc++; switch (op) { - case TOP_is_op: + case TOP_next_instr: + instr = instr->next; + ap = 0; if (instr == NULL) { /* * We'll need at least one more instruction to decide whether @@ -4151,10 +4479,6 @@ transform_engine(LoaderState* st) case TOP_next_arg: ap++; break; - case TOP_next_instr: - instr = instr->next; - ap = 0; - break; case TOP_commit: instr = instr->next; /* The next_instr was optimized away. */ @@ -4172,8 +4496,8 @@ transform_engine(LoaderState* st) #endif break; -#if defined(TOP_call) - case TOP_call: +#if defined(TOP_call_end) + case TOP_call_end: { GenOp** lastp; GenOp* new_instr; @@ -4210,7 +4534,7 @@ transform_engine(LoaderState* st) *lastp = st->genop; st->genop = new_instr; } - break; + RETURN(TE_OK); #endif case TOP_new_instr: /* @@ -4219,12 +4543,10 @@ transform_engine(LoaderState* st) NEW_GENOP(st, instr); instr->next = st->genop; st->genop = instr; + instr->op = op = *pc++; + instr->arity = gen_opc[op].arity; ap = 0; break; - case TOP_store_op: - instr->op = *pc++; - instr->arity = *pc++; - break; case TOP_store_type: i = *pc++; instr->a[ap].type = i; @@ -4234,21 +4556,25 @@ transform_engine(LoaderState* st) i = *pc++; instr->a[ap].val = i; break; - case TOP_store_var: + case TOP_store_var_next_arg: i = *pc++; ASSERT(i < TE_MAX_VARS); instr->a[ap].type = var[i].type; instr->a[ap].val = var[i].val; + ap++; break; case TOP_try_me_else: restart = pc + 1; restart += *pc++; ASSERT(*pc < NUM_TOPS); /* Valid instruction? */ break; + case TOP_try_me_else_fail: + restart = restart_fail; + break; case TOP_end: RETURN(TE_OK); case TOP_fail: - RETURN(TE_FAIL) + RETURN(TE_FAIL); default: ASSERT(0); } @@ -4317,41 +4643,9 @@ load_printf(int line, LoaderState* context, char *fmt,...) erts_send_error_to_logger(context->group_leader, dsbufp); } - -static int -get_int_val(LoaderState* stp, Uint len_code, BeamInstr* result) -{ - Uint count; - Uint val; - - len_code >>= 5; - ASSERT(len_code < 8); - if (len_code == 7) { - LoadError0(stp, "can't load integers bigger than 8 bytes yet\n"); - } - count = len_code + 2; - if (count == 5) { - Uint msb; - GetByte(stp, msb); - if (msb == 0) { - count--; - } - GetInt(stp, 4, *result); - } else if (count <= 4) { - GetInt(stp, count, val); - *result = ((val << 8*(sizeof(val)-count)) >> 8*(sizeof(val)-count)); - } else { - LoadError1(stp, "too big integer; %d bytes\n", count); - } - return 1; - - load_error: - return 0; -} - - static int -get_erlang_integer(LoaderState* stp, Uint len_code, BeamInstr* result) +get_tag_and_value(LoaderState* stp, Uint len_code, + unsigned tag, BeamInstr* result) { Uint count; Sint val; @@ -4371,17 +4665,62 @@ get_erlang_integer(LoaderState* stp, Uint len_code, BeamInstr* result) if (len_code < 7) { count = len_code + 2; } else { - Uint tag; + unsigned sztag; UWord len_word; ASSERT(len_code == 7); - GetTagAndValue(stp, tag, len_word); - VerifyTag(stp, TAG_u, tag); + GetTagAndValue(stp, sztag, len_word); + VerifyTag(stp, sztag, TAG_u); count = len_word + 9; } /* - * Handle values up to the size of an int, meaning either a small or bignum. + * The value for tags except TAG_i must be an unsigned integer + * fitting in an Uint. If it does not fit, we'll indicate overflow + * by changing the tag to TAG_o. + */ + + if (tag != TAG_i) { + if (count == sizeof(Uint)+1) { + Uint msb; + + /* + * The encoded value has one more byte than an Uint. + * It will still fit in an Uint if the most significant + * byte is 0. + */ + GetByte(stp, msb); + GetInt(stp, sizeof(Uint), *result); + if (msb != 0) { + /* Overflow: Negative or too big. */ + return TAG_o; + } + } else if (count == sizeof(Uint)) { + /* + * The value must be positive (or the encoded value would + * have been one byte longer). + */ + GetInt(stp, count, *result); + } else if (count < sizeof(Uint)) { + GetInt(stp, count, *result); + + /* + * If the sign bit is set, the value is negative + * (not allowed). + */ + if (*result & ((Uint)1 << (count*8-1))) { + return TAG_o; + } + } else { + GetInt(stp, count, *result); + return TAG_o; + } + return tag; + } + + /* + * TAG_i: First handle values up to the size of an Uint (i.e. either + * a small or a bignum). */ if (count <= sizeof(val)) { @@ -4836,17 +5175,24 @@ compilation_info_for_module(Process* p, /* Process whose heap to use. */ return result; } - /* - * Returns a pointer to {module, function, arity}, or NULL if not found. + * Find a function from the given pc and fill information in + * the FunctionInfo struct. If the full_info is non-zero, fill + * in all available information (including location in the + * source code). If no function is found, the 'current' field + * will be set to NULL. */ -BeamInstr * -find_function_from_pc(BeamInstr* pc) + +void +erts_lookup_function_info(FunctionInfo* fi, BeamInstr* pc, int full_info) { Range* low = modules; Range* high = low + num_loaded_modules; Range* mid = mid_module; + fi->current = NULL; + fi->needed = 5; + fi->loc = LINE_INVALID_LOCATION; while (low < high) { if (pc < mid->start) { high = mid; @@ -4863,16 +5209,147 @@ find_function_from_pc(BeamInstr* pc) high1 = mid1; } else if (pc < mid1[1]) { mid_module = mid; - return mid1[0]+2; + fi->current = mid1[0]+2; + if (full_info) { + BeamInstr** fp = (BeamInstr **) (mid->start + + MI_FUNCTIONS); + int idx = mid1 - fp; + lookup_loc(fi, pc, mid->start, idx); + } + return; } else { low1 = mid1 + 1; } } - return NULL; + return; } mid = low + (high-low) / 2; } - return NULL; +} + +static void +lookup_loc(FunctionInfo* fi, BeamInstr* orig_pc, BeamInstr* modp, int idx) +{ + Eterm* line = (Eterm *) modp[MI_LINE_TABLE]; + Eterm* low; + Eterm* high; + Eterm* mid; + Eterm pc; + + if (line == 0) { + return; + } + + pc = (Eterm) (BeamInstr) orig_pc; + fi->fname_ptr = (Eterm *) (BeamInstr) line[MI_LINE_FNAME_PTR]; + low = (Eterm *) (BeamInstr) line[MI_LINE_FUNC_TAB+idx]; + high = (Eterm *) (BeamInstr) line[MI_LINE_FUNC_TAB+idx+1]; + while (high > low) { + mid = low + (high-low) / 2; + if (pc < mid[0]) { + high = mid; + } else if (pc < mid[1]) { + int file; + int index = mid - (Eterm *) (BeamInstr) line[MI_LINE_FUNC_TAB]; + + if (line[MI_LINE_LOC_SIZE] == 2) { + Uint16* loc_table = + (Uint16 *) (BeamInstr) line[MI_LINE_LOC_TAB]; + fi->loc = loc_table[index]; + } else { + Uint32* loc_table = + (Uint32 *) (BeamInstr) line[MI_LINE_LOC_TAB]; + ASSERT(line[MI_LINE_LOC_SIZE] == 4); + fi->loc = loc_table[index]; + } + if (fi->loc == LINE_INVALID_LOCATION) { + return; + } + fi->needed += 3+2+3+2; + file = LOC_FILE(fi->loc); + if (file == 0) { + /* Special case: Module name with ".erl" appended */ + Atom* mod_atom = atom_tab(atom_val(fi->current[0])); + fi->needed += 2*(mod_atom->len+4); + } else { + Atom* ap = atom_tab(atom_val((fi->fname_ptr)[file-1])); + fi->needed += 2*ap->len; + } + return; + } else { + low = mid + 1; + } + } +} + +/* + * Build a single {M,F,A,Loction} item to be part of + * a stack trace. + */ +Eterm* +erts_build_mfa_item(FunctionInfo* fi, Eterm* hp, Eterm args, Eterm* mfa_p) +{ + BeamInstr* current = fi->current; + Eterm loc = NIL; + + if (fi->loc != LINE_INVALID_LOCATION) { + Eterm tuple; + int line = LOC_LINE(fi->loc); + int file = LOC_FILE(fi->loc); + Eterm file_term = NIL; + + if (file == 0) { + Atom* ap = atom_tab(atom_val(fi->current[0])); + file_term = buf_to_intlist(&hp, ".erl", 4, NIL); + file_term = buf_to_intlist(&hp, (char*)ap->name, ap->len, file_term); + } else { + Atom* ap = atom_tab(atom_val((fi->fname_ptr)[file-1])); + file_term = buf_to_intlist(&hp, (char*)ap->name, ap->len, NIL); + } + + tuple = TUPLE2(hp, am_line, make_small(line)); + hp += 3; + loc = CONS(hp, tuple, loc); + hp += 2; + tuple = TUPLE2(hp, am_file, file_term); + hp += 3; + loc = CONS(hp, tuple, loc); + hp += 2; + } + + if (is_list(args) || is_nil(args)) { + *mfa_p = TUPLE4(hp, current[0], current[1], args, loc); + } else { + Eterm arity = make_small(current[2]); + *mfa_p = TUPLE4(hp, current[0], current[1], arity, loc); + } + return hp + 5; +} + +/* + * Force setting of the current function in a FunctionInfo + * structure. No source code location will be associated with + * the function. + */ +void +erts_set_current_function(FunctionInfo* fi, BeamInstr* current) +{ + fi->current = current; + fi->needed = 5; + fi->loc = LINE_INVALID_LOCATION; +} + + +/* + * Returns a pointer to {module, function, arity}, or NULL if not found. + */ +BeamInstr* +find_function_from_pc(BeamInstr* pc) +{ + FunctionInfo fi; + + erts_lookup_function_info(&fi, pc, 0); + return fi.current; } /* diff --git a/erts/emulator/beam/beam_load.h b/erts/emulator/beam/beam_load.h index 26e3054c4b..9d4a60fed1 100644 --- a/erts/emulator/beam/beam_load.h +++ b/erts/emulator/beam/beam_load.h @@ -108,6 +108,11 @@ extern Uint erts_total_code_size; #define MI_ON_LOAD_FUNCTION_PTR 10 /* + * Pointer to the line table (or NULL if none). + */ +#define MI_LINE_TABLE 11 + +/* * Start of function pointer table. This table contains pointers to * all functions in the module plus an additional pointer just beyond * the end of the last function. @@ -116,5 +121,5 @@ extern Uint erts_total_code_size; * this table. */ -#define MI_FUNCTIONS 11 +#define MI_FUNCTIONS 12 #endif /* _BEAM_LOAD_H */ diff --git a/erts/emulator/beam/bif.c b/erts/emulator/beam/bif.c index 68b3350d7f..5b3261077b 100644 --- a/erts/emulator/beam/bif.c +++ b/erts/emulator/beam/bif.c @@ -811,7 +811,7 @@ BIF_RETTYPE spawn_opt_1(BIF_ALIST_1) so.min_heap_size = H_MIN_SIZE; so.min_vheap_size = BIN_VH_MIN_SIZE; so.priority = PRIORITY_NORMAL; - so.max_gen_gcs = (Uint16) erts_smp_atomic32_read(&erts_max_gen_gcs); + so.max_gen_gcs = (Uint16) erts_smp_atomic32_read_nob(&erts_max_gen_gcs); so.scheduler = 0; /* @@ -1189,8 +1189,9 @@ raise_3(Process *c_p, Eterm class, Eterm value, Eterm stacktrace) { Eterm l, *hp, *hp_end, *tp; int depth, cnt; size_t sz; + int must_copy = 0; struct StackTrace *s; - + if (class == am_error) { c_p->fvalue = value; reason = EXC_ERROR; @@ -1206,35 +1207,74 @@ raise_3(Process *c_p, Eterm class, Eterm value, Eterm stacktrace) { /* Check syntax of stacktrace, and count depth. * Accept anything that can be returned from erlang:get_stacktrace/0, * as well as a 2-tuple with a fun as first element that the - * error_handler may need to give us. + * error_handler may need to give us. Also allow old-style + * MFA three-tuples. */ for (l = stacktrace, depth = 0; is_list(l); l = CDR(list_val(l)), depth++) { Eterm t = CAR(list_val(l)); - int arity; + Eterm location = NIL; + if (is_not_tuple(t)) goto error; tp = tuple_val(t); - arity = arityval(tp[0]); - if ((arity == 3) && is_atom(tp[1]) && is_atom(tp[2])) continue; - if ((arity == 2) && is_fun(tp[1])) continue; - goto error; + switch (arityval(tp[0])) { + case 2: + /* {Fun,Args} */ + if (is_fun(tp[1])) { + must_copy = 1; + } else { + goto error; + } + break; + case 3: + /* + * One of: + * {Fun,Args,Location} + * {M,F,A} + */ + if (is_fun(tp[1])) { + location = tp[3]; + } else if (is_atom(tp[1]) && is_atom(tp[2])) { + must_copy = 1; + } else { + goto error; + } + break; + case 4: + if (!(is_atom(tp[1]) && is_atom(tp[2]))) { + goto error; + } + location = tp[4]; + break; + default: + goto error; + } + if (is_not_list(location) && is_not_nil(location)) { + goto error; + } } if (is_not_nil(l)) goto error; /* Create stacktrace and store */ - if (depth <= erts_backtrace_depth) { + if (erts_backtrace_depth < depth) { + depth = erts_backtrace_depth; + must_copy = 1; + } + if (must_copy) { + cnt = depth; + c_p->ftrace = NIL; + } else { + /* No need to copy the stacktrace */ cnt = 0; c_p->ftrace = stacktrace; - } else { - cnt = depth = erts_backtrace_depth; - c_p->ftrace = NIL; } + tp = &c_p->ftrace; sz = (offsetof(struct StackTrace, trace) + sizeof(Eterm) - 1) / sizeof(Eterm); - hp = HAlloc(c_p, sz + 2*(cnt + 1)); - hp_end = hp + sz + 2*(cnt + 1); + hp = HAlloc(c_p, sz + (2+6)*(cnt + 1)); + hp_end = hp + sz + (2+6)*(cnt + 1); s = (struct StackTrace *) hp; s->header = make_neg_bignum_header(sz - 1); s->freason = reason; @@ -1242,13 +1282,29 @@ raise_3(Process *c_p, Eterm class, Eterm value, Eterm stacktrace) { s->current = NULL; s->depth = 0; hp += sz; - if (cnt > 0) { + if (must_copy) { + int cnt; + /* Copy list up to depth */ for (cnt = 0, l = stacktrace; cnt < depth; cnt++, l = CDR(list_val(l))) { + Eterm t; + Eterm *tpp; + int arity; + ASSERT(*tp == NIL); - *tp = CONS(hp, CAR(list_val(l)), *tp); + t = CAR(list_val(l)); + tpp = tuple_val(t); + arity = arityval(tpp[0]); + if (arity == 2) { + t = TUPLE3(hp, tpp[1], tpp[2], NIL); + hp += 4; + } else if (arity == 3 && is_atom(tpp[1])) { + t = TUPLE4(hp, tpp[1], tpp[2], tpp[3], NIL); + hp += 5; + } + *tp = CONS(hp, t, *tp); tp = &CDR(list_val(*tp)); hp += 2; } @@ -1256,7 +1312,7 @@ raise_3(Process *c_p, Eterm class, Eterm value, Eterm stacktrace) { c_p->ftrace = CONS(hp, c_p->ftrace, make_big((Eterm *) s)); hp += 2; ASSERT(hp <= hp_end); - + HRelease(c_p, hp_end, hp); BIF_ERROR(c_p, reason); error: @@ -3417,10 +3473,10 @@ BIF_RETTYPE ports_0(BIF_ALIST_0) erts_smp_mtx_lock(&ports_snapshot_mtx); /* One snapshot at a time */ - erts_smp_atomic_set(&erts_dead_ports_ptr, - (erts_aint_t) (port_buf + erts_max_ports)); + erts_smp_atomic_set_nob(&erts_dead_ports_ptr, + (erts_aint_t) (port_buf + erts_max_ports)); - next_ss = erts_smp_atomic32_inctest(&erts_ports_snapshot); + next_ss = erts_smp_atomic32_inc_read_relb(&erts_ports_snapshot); for (i = erts_max_ports-1; i >= 0; i--) { Port* prt = &erts_port[i]; @@ -3434,8 +3490,8 @@ BIF_RETTYPE ports_0(BIF_ALIST_0) erts_smp_port_state_unlock(prt); } - dead_ports = (Eterm*)erts_smp_atomic_xchg(&erts_dead_ports_ptr, - (erts_aint_t) NULL); + dead_ports = (Eterm*)erts_smp_atomic_xchg_nob(&erts_dead_ports_ptr, + (erts_aint_t) NULL); erts_smp_mtx_unlock(&ports_snapshot_mtx); ASSERT(pp <= dead_ports); @@ -3942,8 +3998,8 @@ BIF_RETTYPE system_flag_2(BIF_ALIST_2) goto error; } nval = (n > (Sint) ((Uint16) -1)) ? ((Uint16) -1) : ((Uint16) n); - oval = (Uint) erts_smp_atomic32_xchg(&erts_max_gen_gcs, - (erts_aint32_t) nval); + oval = (Uint) erts_smp_atomic32_xchg_nob(&erts_max_gen_gcs, + (erts_aint32_t) nval); BIF_RET(make_small(oval)); } else if (BIF_ARG_1 == am_min_heap_size) { int oval = H_MIN_SIZE; @@ -4286,7 +4342,7 @@ void erts_init_bif(void) erts_smp_spinlock_init(&make_ref_lock, "make_ref"); erts_smp_mtx_init(&ports_snapshot_mtx, "ports_snapshot"); - erts_smp_atomic_init(&erts_dead_ports_ptr, (erts_aint_t) NULL); + erts_smp_atomic_init_nob(&erts_dead_ports_ptr, (erts_aint_t) NULL); /* * bif_return_trap/1 is a hidden BIF that bifs that need to diff --git a/erts/emulator/beam/bif.tab b/erts/emulator/beam/bif.tab index d9dd80fa8b..ba30fa85b8 100644 --- a/erts/emulator/beam/bif.tab +++ b/erts/emulator/beam/bif.tab @@ -1,7 +1,7 @@ # # %CopyrightBegin% # -# Copyright Ericsson AB 1996-2010. All Rights Reserved. +# Copyright Ericsson AB 1996-2011. All Rights Reserved. # # The contents of this file are subject to the Erlang Public License, # Version 1.1, (the "License"); you may not use this file except in @@ -87,6 +87,8 @@ bif erlang:exit/2 bif 'erl.lang.proc':signal/2 ebif_signal_2 exit_2 bif erlang:external_size/1 bif 'erl.lang.term':external_size/1 ebif_external_size_1 +bif erlang:external_size/2 +bif 'erl.lang.term':external_size/2 ebif_external_size_2 ubif erlang:float/1 ubif 'erl.lang.number':to_float/1 ebif_to_float_1 float_1 bif erlang:float_to_list/1 @@ -802,6 +804,12 @@ bif prim_file:internal_name2native/1 bif prim_file:internal_native2name/1 bif prim_file:internal_normalize_utf8/1 bif file:native_name_encoding/0 + +# +# New in R14B04. +# +bif erlang:check_old_code/1 + # # Obsolete # diff --git a/erts/emulator/beam/break.c b/erts/emulator/beam/break.c index b8889e6206..432b3d0780 100644 --- a/erts/emulator/beam/break.c +++ b/erts/emulator/beam/break.c @@ -626,7 +626,7 @@ bin_check(void) erts_printf("%p orig_size: %bpd, norefs = %bpd\n", bp->val, bp->val->orig_size, - erts_smp_atomic_read(&bp->val->refc)); + erts_smp_atomic_read_nob(&bp->val->refc)); } } if (printed) { @@ -650,7 +650,7 @@ erl_crash_dump_v(char *file, int line, char* fmt, va_list args) char dumpnamebuf[MAXPATHLEN]; char* dumpname; - if (ERTS_IS_CRASH_DUMPING) + if (ERTS_SOMEONE_IS_CRASH_DUMPING) return; /* Wait for all threads to block. If all threads haven't blocked @@ -667,7 +667,8 @@ erl_crash_dump_v(char *file, int line, char* fmt, va_list args) /* Allow us to pass certain places without locking... */ #ifdef ERTS_SMP - erts_smp_atomic_inc(&erts_writing_erl_crash_dump); + erts_smp_atomic32_set_mb(&erts_writing_erl_crash_dump, 1); + erts_smp_tsd_set(erts_is_crash_dumping_key, (void *) 1); #else erts_writing_erl_crash_dump = 1; #endif diff --git a/erts/emulator/beam/dist.c b/erts/emulator/beam/dist.c index b1cdd0660a..ad042ec088 100644 --- a/erts/emulator/beam/dist.c +++ b/erts/emulator/beam/dist.c @@ -128,8 +128,8 @@ delete_cache(ErtsAtomCache *cache) { if (cache) { erts_free(ERTS_ALC_T_DCACHE, (void *) cache); - ASSERT(erts_smp_atomic_read(&no_caches) > 0); - erts_smp_atomic_dec(&no_caches); + ASSERT(erts_smp_atomic_read_nob(&no_caches) > 0); + erts_smp_atomic_dec_nob(&no_caches); } } @@ -147,7 +147,7 @@ create_cache(DistEntry *dep) dep->cache = cp = (ErtsAtomCache*) erts_alloc(ERTS_ALC_T_DCACHE, sizeof(ErtsAtomCache)); - erts_smp_atomic_inc(&no_caches); + erts_smp_atomic_inc_nob(&no_caches); for (i = 0; i < sizeof(cp->in_arr)/sizeof(cp->in_arr[0]); i++) { cp->in_arr[i] = THE_NON_VALUE; cp->out_arr[i] = THE_NON_VALUE; @@ -156,7 +156,7 @@ create_cache(DistEntry *dep) Uint erts_dist_cache_size(void) { - return (Uint) erts_smp_atomic_read(&no_caches)*sizeof(ErtsAtomCache); + return (Uint) erts_smp_atomic_read_mb(&no_caches)*sizeof(ErtsAtomCache); } static ErtsProcList * @@ -444,7 +444,7 @@ int erts_do_net_exits(DistEntry *dep, Eterm reason) ErtsMonitor *monitors; Uint32 flags; - erts_smp_atomic_set(&dep->dist_cmd_scheduled, 1); + erts_smp_atomic_set_mb(&dep->dist_cmd_scheduled, 1); erts_smp_de_rwlock(dep); ERTS_SMP_LC_ASSERT(is_internal_port(dep->cid) @@ -510,7 +510,7 @@ void init_dist(void) { init_nodes_monitors(); - erts_smp_atomic_init(&no_caches, 0); + erts_smp_atomic_init_nob(&no_caches, 0); /* Lookup/Install all references to trap functions */ dsend2_trap = trap_function(am_dsend,2); @@ -596,7 +596,7 @@ static void clear_dist_entry(DistEntry *dep) suspendees = get_suspended_on_de(dep, ERTS_DE_QFLGS_ALL); erts_smp_mtx_unlock(&dep->qlock); - erts_smp_atomic_set(&dep->dist_cmd_scheduled, 0); + erts_smp_atomic_set_nob(&dep->dist_cmd_scheduled, 0); dep->send = NULL; erts_smp_de_rwunlock(dep); @@ -1775,7 +1775,7 @@ erts_dist_command(Port *prt, int reds_limit) erts_refc_inc(&dep->refc, 1); /* Otherwise dist_entry might be removed if port command fails */ - erts_smp_atomic_xchg(&dep->dist_cmd_scheduled, 0); + erts_smp_atomic_set_mb(&dep->dist_cmd_scheduled, 0); erts_smp_de_rlock(dep); flags = dep->flags; diff --git a/erts/emulator/beam/dist.h b/erts/emulator/beam/dist.h index 695a4fc3fe..845151c895 100644 --- a/erts/emulator/beam/dist.h +++ b/erts/emulator/beam/dist.h @@ -1,7 +1,7 @@ /* * %CopyrightBegin% * - * Copyright Ericsson AB 1996-2010. All Rights Reserved. + * Copyright Ericsson AB 1996-2011. All Rights Reserved. * * The contents of this file are subject to the Erlang Public License, * Version 1.1, (the "License"); you may not use this file except in @@ -203,7 +203,7 @@ void erts_schedule_dist_command(Port *prt, DistEntry *dist_entry) id = dep->cid; } - if (!erts_smp_atomic_xchg(&dep->dist_cmd_scheduled, 1)) { + if (!erts_smp_atomic_xchg_mb(&dep->dist_cmd_scheduled, 1)) { (void) erts_port_task_schedule(id, &dep->dist_cmd, ERTS_PORT_TASK_DIST_CMD, diff --git a/erts/emulator/beam/erl_afit_alloc.c b/erts/emulator/beam/erl_afit_alloc.c index e8b594bb47..bcc7ea04ae 100644 --- a/erts/emulator/beam/erl_afit_alloc.c +++ b/erts/emulator/beam/erl_afit_alloc.c @@ -1,7 +1,7 @@ /* * %CopyrightBegin% * - * Copyright Ericsson AB 2003-2009. All Rights Reserved. + * Copyright Ericsson AB 2003-2011. All Rights Reserved. * * The contents of this file are subject to the Erlang Public License, * Version 1.1, (the "License"); you may not use this file except in @@ -43,9 +43,9 @@ /* Prototypes of callback functions */ static Block_t * get_free_block (Allctr_t *, Uint, - Block_t *, Uint); -static void link_free_block (Allctr_t *, Block_t *); -static void unlink_free_block (Allctr_t *, Block_t *); + Block_t *, Uint, Uint32); +static void link_free_block (Allctr_t *, Block_t *, Uint32); +static void unlink_free_block (Allctr_t *, Block_t *, Uint32); static Eterm info_options (Allctr_t *, char *, int *, @@ -72,6 +72,8 @@ erts_afalc_start(AFAllctr_t *afallctr, is a struct). */ Allctr_t *allctr = (Allctr_t *) afallctr; + init->sbmbct = 0; /* Small mbc not supported by afit */ + sys_memcpy((void *) afallctr, (void *) &nulled_state, sizeof(AFAllctr_t)); allctr->mbc_header_size = sizeof(Carrier_t); @@ -105,7 +107,8 @@ erts_afalc_start(AFAllctr_t *afallctr, } static Block_t * -get_free_block(Allctr_t *allctr, Uint size, Block_t *cand_blk, Uint cand_size) +get_free_block(Allctr_t *allctr, Uint size, Block_t *cand_blk, Uint cand_size, + Uint32 flags) { AFAllctr_t *afallctr = (AFAllctr_t *) allctr; @@ -123,7 +126,7 @@ get_free_block(Allctr_t *allctr, Uint size, Block_t *cand_blk, Uint cand_size) } static void -link_free_block(Allctr_t *allctr, Block_t *block) +link_free_block(Allctr_t *allctr, Block_t *block, Uint32 flags) { AFFreeBlock_t *blk = (AFFreeBlock_t *) block; AFAllctr_t *afallctr = (AFAllctr_t *) allctr; @@ -144,7 +147,7 @@ link_free_block(Allctr_t *allctr, Block_t *block) } static void -unlink_free_block(Allctr_t *allctr, Block_t *block) +unlink_free_block(Allctr_t *allctr, Block_t *block, Uint32 flags) { AFFreeBlock_t *blk = (AFFreeBlock_t *) block; AFAllctr_t *afallctr = (AFAllctr_t *) allctr; diff --git a/erts/emulator/beam/erl_alloc.c b/erts/emulator/beam/erl_alloc.c index cda404af5e..9af80dd7a9 100644 --- a/erts/emulator/beam/erl_alloc.c +++ b/erts/emulator/beam/erl_alloc.c @@ -50,6 +50,9 @@ #include "erl_bestfit_alloc.h" #define GET_ERL_AF_ALLOC_IMPL #include "erl_afit_alloc.h" +#define GET_ERL_AOFF_ALLOC_IMPL +#include "erl_ao_firstfit_alloc.h" + #define ERTS_ALC_DEFAULT_MAX_THR_PREF 16 @@ -85,15 +88,19 @@ typedef union { char align_bfa[ERTS_ALC_CACHE_LINE_ALIGN_SIZE(sizeof(BFAllctr_t))]; AFAllctr_t afa; char align_afa[ERTS_ALC_CACHE_LINE_ALIGN_SIZE(sizeof(AFAllctr_t))]; + AOFFAllctr_t aoffa; + char align_aoffa[ERTS_ALC_CACHE_LINE_ALIGN_SIZE(sizeof(AOFFAllctr_t))]; } ErtsAllocatorState_t; -static ErtsAllocatorState_t sl_alloc_state; +static ErtsAllocatorState_t sbmbc_alloc_state; static ErtsAllocatorState_t std_alloc_state; static ErtsAllocatorState_t ll_alloc_state; #if HALFWORD_HEAP -static ErtsAllocatorState_t std_alloc_low_state; -static ErtsAllocatorState_t ll_alloc_low_state; +static ErtsAllocatorState_t sbmbc_low_alloc_state; +static ErtsAllocatorState_t std_low_alloc_state; +static ErtsAllocatorState_t ll_low_alloc_state; #endif +static ErtsAllocatorState_t sl_alloc_state; static ErtsAllocatorState_t temp_alloc_state; static ErtsAllocatorState_t eheap_alloc_state; static ErtsAllocatorState_t binary_alloc_state; @@ -120,7 +127,8 @@ static void *fix_core_alloc(Uint size) enum allctr_type { GOODFIT, BESTFIT, - AFIT + AFIT, + AOFIRSTFIT }; struct au_init { @@ -132,6 +140,7 @@ struct au_init { GFAllctrInit_t gf; BFAllctrInit_t bf; AFAllctrInit_t af; + AOFFAllctrInit_t aoff; } init; struct { int mmbcs; @@ -145,7 +154,8 @@ struct au_init { ERTS_DEFAULT_ALLCTR_INIT, \ ERTS_DEFAULT_GF_ALLCTR_INIT, \ ERTS_DEFAULT_BF_ALLCTR_INIT, \ - ERTS_DEFAULT_AF_ALLCTR_INIT \ + ERTS_DEFAULT_AF_ALLCTR_INIT, \ + ERTS_DEFAULT_AOFF_ALLCTR_INIT \ } typedef struct { @@ -162,6 +172,7 @@ typedef struct { char *mtrace; char *nodename; } instr; + struct au_init sbmbc_alloc; struct au_init sl_alloc; struct au_init std_alloc; struct au_init ll_alloc; @@ -171,8 +182,9 @@ typedef struct { struct au_init ets_alloc; struct au_init driver_alloc; #if HALFWORD_HEAP - struct au_init std_alloc_low; - struct au_init ll_alloc_low; + struct au_init sbmbc_low_alloc; + struct au_init std_low_alloc; + struct au_init ll_low_alloc; #endif } erts_alc_hndl_args_init_t; @@ -185,6 +197,34 @@ do { \ } while (0) static void +set_default_sbmbc_alloc_opts(struct au_init *ip) +{ + SET_DEFAULT_ALLOC_OPTS(ip); + ip->enable = 0; + ip->thr_spec = 0; + ip->atype = BESTFIT; + ip->init.bf.ao = 1; + ip->init.util.ramv = 0; + ip->init.util.mmsbc = 0; + ip->init.util.mmmbc = 500; + ip->init.util.sbct = ~((UWord) 0); + ip->init.util.name_prefix = "sbmbc_"; + ip->init.util.alloc_no = ERTS_ALC_A_SBMBC; +#ifndef SMALL_MEMORY + ip->init.util.mmbcs = 2*1024*1024; /* Main carrier size */ +#else + ip->init.util.mmbcs = 1*1024*1024; /* Main carrier size */ +#endif + ip->init.util.ts = ERTS_ALC_MTA_SBMBC; + ip->init.util.asbcst = 0; + ip->init.util.rsbcst = 0; + ip->init.util.rsbcmt = 0; + ip->init.util.rmbcmt = 0; + ip->init.util.sbmbct = 0; + ip->init.util.sbmbcs = 0; +} + +static void set_default_sl_alloc_opts(struct au_init *ip) { SET_DEFAULT_ALLOC_OPTS(ip); @@ -202,6 +242,7 @@ set_default_sl_alloc_opts(struct au_init *ip) ip->init.util.ts = ERTS_ALC_MTA_SHORT_LIVED; ip->init.util.rsbcst = 80; #if HALFWORD_HEAP + ip->init.util.force = 1; ip->init.util.low_mem = 1; #endif @@ -249,6 +290,8 @@ set_default_ll_alloc_opts(struct au_init *ip) ip->init.util.rsbcst = 0; ip->init.util.rsbcmt = 0; ip->init.util.rmbcmt = 0; + ip->init.util.sbmbct = 0; + ip->init.util.sbmbcs = 0; } static void @@ -269,6 +312,7 @@ set_default_temp_alloc_opts(struct au_init *ip) ip->init.util.rsbcst = 90; ip->init.util.rmbcmt = 100; #if HALFWORD_HEAP + ip->init.util.force = 1; ip->init.util.low_mem = 1; #endif } @@ -291,6 +335,7 @@ set_default_eheap_alloc_opts(struct au_init *ip) ip->init.util.ts = ERTS_ALC_MTA_EHEAP; ip->init.util.rsbcst = 50; #if HALFWORD_HEAP + ip->init.util.force = 1; ip->init.util.low_mem = 1; #endif } @@ -436,10 +481,13 @@ erts_alloc_init(int *argc, char **argv, ErtsAllocInitOpts *eaiop) hdbg_init(); #endif + erts_have_sbmbc_alloc = 0; + erts_sys_alloc_init(); init_thr_ix(erts_no_schedulers); erts_init_utils_mem(); + set_default_sbmbc_alloc_opts(&init.sbmbc_alloc); set_default_sl_alloc_opts(&init.sl_alloc); set_default_std_alloc_opts(&init.std_alloc); set_default_ll_alloc_opts(&init.ll_alloc); @@ -453,6 +501,7 @@ erts_alloc_init(int *argc, char **argv, ErtsAllocInitOpts *eaiop) handle_args(argc, argv, &init); if (erts_no_schedulers <= 1) { + init.sbmbc_alloc.thr_spec = 0; init.sl_alloc.thr_spec = 0; init.std_alloc.thr_spec = 0; init.ll_alloc.thr_spec = 0; @@ -464,6 +513,7 @@ erts_alloc_init(int *argc, char **argv, ErtsAllocInitOpts *eaiop) if (init.erts_alloc_config) { /* Adjust flags that erts_alloc_config won't like */ + init.sbmbc_alloc.thr_spec = 0; init.temp_alloc.thr_spec = 0; init.sl_alloc.thr_spec = 0; init.std_alloc.thr_spec = 0; @@ -480,6 +530,7 @@ erts_alloc_init(int *argc, char **argv, ErtsAllocInitOpts *eaiop) init.temp_alloc.thr_spec = erts_no_schedulers; /* Others must use thread preferred interface */ + adjust_tpref(&init.sbmbc_alloc, erts_no_schedulers); adjust_tpref(&init.sl_alloc, erts_no_schedulers); adjust_tpref(&init.std_alloc, erts_no_schedulers); adjust_tpref(&init.ll_alloc, erts_no_schedulers); @@ -497,6 +548,7 @@ erts_alloc_init(int *argc, char **argv, ErtsAllocInitOpts *eaiop) * The following allocators cannot be run with afit strategy. * Make sure they don't... */ + refuse_af_strategy(&init.sbmbc_alloc); refuse_af_strategy(&init.sl_alloc); refuse_af_strategy(&init.std_alloc); refuse_af_strategy(&init.ll_alloc); @@ -518,6 +570,7 @@ erts_alloc_init(int *argc, char **argv, ErtsAllocInitOpts *eaiop) erts_afalc_init(); erts_bfalc_init(); erts_gfalc_init(); + erts_aoffalc_init(); for (i = ERTS_ALC_A_MIN; i <= ERTS_ALC_A_MAX; i++) { erts_allctrs[i].alloc = NULL; @@ -551,19 +604,30 @@ erts_alloc_init(int *argc, char **argv, ErtsAllocInitOpts *eaiop) #if HALFWORD_HEAP /* Init low memory variants by cloning */ - init.std_alloc_low = init.std_alloc; - init.std_alloc_low.init.util.alloc_no = ERTS_ALC_A_STANDARD_LOW; - init.std_alloc_low.init.util.low_mem = 1; - - init.ll_alloc_low = init.ll_alloc; - init.ll_alloc_low.init.util.alloc_no = ERTS_ALC_A_LONG_LIVED_LOW; - init.ll_alloc_low.init.util.low_mem = 1; - - set_au_allocator(ERTS_ALC_A_STANDARD_LOW, &init.std_alloc_low); - set_au_allocator(ERTS_ALC_A_LONG_LIVED_LOW, &init.ll_alloc_low); + init.sbmbc_low_alloc = init.sbmbc_alloc; + init.sbmbc_low_alloc.init.util.name_prefix = "sbmbc_low_"; + init.sbmbc_low_alloc.init.util.alloc_no = ERTS_ALC_A_SBMBC_LOW; + init.sbmbc_low_alloc.init.util.low_mem = 1; + + init.std_low_alloc = init.std_alloc; + init.std_low_alloc.init.util.name_prefix = "std_low_"; + init.std_low_alloc.init.util.alloc_no = ERTS_ALC_A_STANDARD_LOW; + init.std_low_alloc.init.util.force = 1; + init.std_low_alloc.init.util.low_mem = 1; + + init.ll_low_alloc = init.ll_alloc; + init.ll_low_alloc.init.util.name_prefix = "ll_low_"; + init.ll_low_alloc.init.util.alloc_no = ERTS_ALC_A_LONG_LIVED_LOW; + init.ll_low_alloc.init.util.force = 1; + init.ll_low_alloc.init.util.low_mem = 1; + + set_au_allocator(ERTS_ALC_A_SBMBC_LOW, &init.sbmbc_low_alloc); + set_au_allocator(ERTS_ALC_A_STANDARD_LOW, &init.std_low_alloc); + set_au_allocator(ERTS_ALC_A_LONG_LIVED_LOW, &init.ll_low_alloc); #endif /* HALFWORD */ set_au_allocator(ERTS_ALC_A_TEMPORARY, &init.temp_alloc); + set_au_allocator(ERTS_ALC_A_SBMBC, &init.sbmbc_alloc); set_au_allocator(ERTS_ALC_A_SHORT_LIVED, &init.sl_alloc); set_au_allocator(ERTS_ALC_A_STANDARD, &init.std_alloc); set_au_allocator(ERTS_ALC_A_LONG_LIVED, &init.ll_alloc); @@ -593,6 +657,20 @@ erts_alloc_init(int *argc, char **argv, ErtsAllocInitOpts *eaiop) erts_mtrace_init(init.instr.mtrace, init.instr.nodename); + /* sbmbc_alloc() needs to be started first */ + start_au_allocator(ERTS_ALC_A_SBMBC, + &init.sbmbc_alloc, + &sbmbc_alloc_state); +#if HALFWORD_HEAP + start_au_allocator(ERTS_ALC_A_SBMBC_LOW, + &init.sbmbc_low_alloc, + &sbmbc_low_alloc_state); + erts_have_sbmbc_alloc = (init.sbmbc_alloc.enable + && init.sbmbc_low_alloc.enable); +#else + erts_have_sbmbc_alloc = init.sbmbc_alloc.enable; +#endif + start_au_allocator(ERTS_ALC_A_TEMPORARY, &init.temp_alloc, &temp_alloc_state); @@ -610,11 +688,11 @@ erts_alloc_init(int *argc, char **argv, ErtsAllocInitOpts *eaiop) &ll_alloc_state); #if HALFWORD_HEAP start_au_allocator(ERTS_ALC_A_LONG_LIVED_LOW, - &init.ll_alloc_low, - &ll_alloc_low_state); + &init.ll_low_alloc, + &ll_low_alloc_state); start_au_allocator(ERTS_ALC_A_STANDARD_LOW, - &init.std_alloc_low, - &std_alloc_low_state); + &init.std_low_alloc, + &std_low_alloc_state); #endif start_au_allocator(ERTS_ALC_A_EHEAP, &init.eheap_alloc, @@ -680,14 +758,11 @@ set_au_allocator(ErtsAlcType_t alctr_n, struct au_init *init) ErtsAllocatorInfo_t *ai = &erts_allctrs_info[alctr_n]; ErtsAllocatorThrSpec_t *tspec = &erts_allctr_thr_spec[alctr_n]; -#if HALFWORD_HEAP - /* If halfword heap, silently ignore any disabling of internal - * allocators for low memory + /* + * Some allocators are forced on if halfword heap is used. */ - if (init->init.util.low_mem) { + if (init->init.util.force) init->enable = 1; - } -#endif if (!init->enable) { af->alloc = erts_sys_alloc; @@ -837,6 +912,12 @@ start_au_allocator(ErtsAlcType_t alctr_n, &init->init.af, &init->init.util); break; + case AOFIRSTFIT: + as = (void *) erts_aoffalc_start((AOFFAllctr_t *) as0, + &init->init.aoff, + &init->init.util); + break; + default: as = NULL; ASSERT(0); @@ -947,6 +1028,20 @@ get_kb_value(char *param_end, char** argv, int* ip) } static Uint +get_byte_value(char *param_end, char** argv, int* ip) +{ + Sint tmp; + char *rest; + char *param = argv[*ip]+1; + char *value = get_value(param_end, argv, ip); + errno = 0; + tmp = (Sint) strtol(value, &rest, 10); + if (errno != 0 || rest == value || tmp < 0) + bad_value(param, param_end, value); + return (Uint) tmp; +} + +static Uint get_amount_value(char *param_end, char** argv, int* ip) { Sint tmp; @@ -1017,6 +1112,9 @@ handle_au_arg(struct au_init *auip, else if (strcmp("af", alg) == 0) { auip->atype = AFIT; } + else if (strcmp("aoff", alg) == 0) { + auip->atype = AOFIRSTFIT; + } else { bad_value(param, sub_param + 1, alg); } @@ -1085,6 +1183,12 @@ handle_au_arg(struct au_init *auip, if(has_prefix("sbct", sub_param)) { auip->init.util.sbct = get_kb_value(sub_param + 4, argv, ip); } + else if (has_prefix("sbmbcs", sub_param)) { + auip->init.util.sbmbcs = get_byte_value(sub_param + 6, argv, ip); + } + else if (has_prefix("sbmbct", sub_param)) { + auip->init.util.sbmbct = get_byte_value(sub_param + 6, argv, ip); + } else if (has_prefix("smbcs", sub_param)) { auip->default_.smbcs = 0; auip->init.util.smbcs = get_kb_value(sub_param + 5, argv, ip); @@ -1123,6 +1227,7 @@ static void handle_args(int *argc, char **argv, erts_alc_hndl_args_init_t *init) { struct au_init *aui[] = { + &init->sbmbc_alloc, &init->binary_alloc, &init->std_alloc, &init->ets_alloc, @@ -1150,6 +1255,9 @@ handle_args(int *argc, char **argv, erts_alc_hndl_args_init_t *init) case 'B': handle_au_arg(&init->binary_alloc, &argv[i][3], argv, &i); break; + case 'C': + handle_au_arg(&init->sbmbc_alloc, &argv[i][3], argv, &i); + break; case 'D': handle_au_arg(&init->std_alloc, &argv[i][3], argv, &i); break; @@ -1856,12 +1964,16 @@ erts_memory(int *print_to_p, void *print_to_arg, void *proc, Eterm earg) return am_badarg; } - /* All alloc_util allocators *have* to be enabled */ + /* All alloc_util allocators except sbmbc_alloc *have* to be enabled */ for (ai = ERTS_ALC_A_MIN; ai <= ERTS_ALC_A_MAX; ai++) { switch (ai) { case ERTS_ALC_A_SYSTEM: case ERTS_ALC_A_FIXED_SIZE: + case ERTS_ALC_A_SBMBC: +#if HALFWORD_HEAP + case ERTS_ALC_A_SBMBC_LOW: +#endif break; default: if (!erts_allctrs_info[ai].enabled @@ -1901,6 +2013,12 @@ erts_memory(int *print_to_p, void *print_to_arg, void *proc, Eterm earg) * Often not thread safe and usually never * contain any allocated memory. */ + case ERTS_ALC_A_SBMBC: + /* Included in other allocators */ +#if HALFWORD_HEAP + case ERTS_ALC_A_SBMBC_LOW: + /* Included in other allocators */ +#endif continue; case ERTS_ALC_A_EHEAP: save = &size.processes; @@ -2842,6 +2960,29 @@ erts_allocator_options(void *proc) return res; } +void *erts_alloc_permanent_cache_aligned(ErtsAlcType_t type, Uint size) +{ + UWord v = (UWord) erts_alloc(type, size + (ERTS_CACHE_LINE_SIZE-1)); + +#ifdef VALGRIND + { /* Avoid Leak_PossiblyLost */ + static UWord vg_root_set[10]; + static unsigned ix = 0; + if (ix >= sizeof(vg_root_set) / sizeof(*vg_root_set)) { + erl_exit(ERTS_ABORT_EXIT, "Too many erts_alloc_permanent_cache_aligned's\n"); + } + vg_root_set[ix++] = v; /* not thread safe */ + } +#endif + + if (v & ERTS_CACHE_LINE_MASK) { + v = (v & ~ERTS_CACHE_LINE_MASK) + ERTS_CACHE_LINE_SIZE; + } + ASSERT((v & ERTS_CACHE_LINE_MASK) == 0); + return (void*)v; +} + + /* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *\ * Deprecated functions * * * @@ -2882,6 +3023,7 @@ unsigned long erts_alc_test(unsigned long op, case 0x2: return erts_bfalc_test(op, a1, a2); case 0x3: return erts_afalc_test(op, a1, a2); case 0x4: return erts_mseg_test(op, a1, a2, a3); + case 0x5: return erts_aoffalc_test(op, a1, a2); case 0xf: switch (op) { case 0xf00: @@ -2925,6 +3067,7 @@ unsigned long erts_alc_test(unsigned long op, init.atype = GOODFIT; init.init.util.name_prefix = (char *) a1; init.init.util.ts = a2 ? 1 : 0; + init.init.util.sbmbct = 0; if ((char **) a3) { char **argv = (char **) a3; @@ -2960,6 +3103,14 @@ unsigned long erts_alc_test(unsigned long op, &init.init.af, &init.init.util); break; + case AOFIRSTFIT: + allctr = erts_aoffalc_start((AOFFAllctr_t *) + erts_alloc(ERTS_ALC_T_UNDEF, + sizeof(AOFFAllctr_t)), + &init.init.aoff, + &init.init.util); + break; + default: ASSERT(0); allctr = NULL; @@ -3455,6 +3606,4 @@ install_debug_functions(void) return FENCE_SZ; } - - #endif /* #ifdef DEBUG */ diff --git a/erts/emulator/beam/erl_alloc.h b/erts/emulator/beam/erl_alloc.h index ce792d4d17..80cb82c393 100644 --- a/erts/emulator/beam/erl_alloc.h +++ b/erts/emulator/beam/erl_alloc.h @@ -99,6 +99,14 @@ unsigned long erts_alc_test(unsigned long, #define ERTS_ALC_MIN_LONG_LIVED_TIME (10*60*1000) +#if HALFWORD_HEAP +#define ERTS_IS_SBMBC_ALLOCATOR_NO__(NO) \ + ((NO) == ERTS_ALC_A_SBMBC || (NO) == ERTS_ALC_A_SBMBC_LOW) +#else +#define ERTS_IS_SBMBC_ALLOCATOR_NO__(NO) \ + ((NO) == ERTS_ALC_A_SBMBC) +#endif + typedef struct { int alloc_util; int enabled; @@ -172,11 +180,11 @@ void *erts_realloc(ErtsAlcType_t type, void *ptr, Uint size); void erts_free(ErtsAlcType_t type, void *ptr); void *erts_alloc_fnf(ErtsAlcType_t type, Uint size); void *erts_realloc_fnf(ErtsAlcType_t type, void *ptr, Uint size); -void *erts_alloc_permanent_cache_aligned(ErtsAlcType_t type, Uint size); - #endif /* #if !ERTS_ALC_DO_INLINE */ +void *erts_alloc_permanent_cache_aligned(ErtsAlcType_t type, Uint size); + #ifndef ERTS_CACHE_LINE_SIZE /* Assume a cache line size of 64 bytes */ # define ERTS_CACHE_LINE_SIZE ((UWord) 64) @@ -242,18 +250,6 @@ void *erts_realloc_fnf(ErtsAlcType_t type, void *ptr, Uint size) size); } -ERTS_ALC_INLINE -void *erts_alloc_permanent_cache_aligned(ErtsAlcType_t type, Uint size) -{ - UWord v = (UWord) erts_alloc(type, size + (ERTS_CACHE_LINE_SIZE-1)); - - if (v & ERTS_CACHE_LINE_MASK) { - v = (v & ~ERTS_CACHE_LINE_MASK) + ERTS_CACHE_LINE_SIZE; - } - ASSERT((v & ERTS_CACHE_LINE_MASK) == 0); - return (void*)v; -} - #endif /* #if ERTS_ALC_DO_INLINE || defined(ERTS_ALC_INTERNAL__) */ typedef void (*erts_alloc_verify_func_t)(Allctr_t *); diff --git a/erts/emulator/beam/erl_alloc.types b/erts/emulator/beam/erl_alloc.types index c6cc0e1fac..eda0831441 100644 --- a/erts/emulator/beam/erl_alloc.types +++ b/erts/emulator/beam/erl_alloc.types @@ -65,6 +65,11 @@ allocator SYSTEM true sys_alloc +allocator SBMBC true sbmbc_alloc ++if halfword +allocator SBMBC_LOW true sbmbc_low_alloc ++endif + +if smp allocator TEMPORARY true temp_alloc @@ -76,8 +81,8 @@ allocator ETS true ets_alloc allocator FIXED_SIZE true fix_alloc +if halfword -allocator LONG_LIVED_LOW true ll_alloc_low -allocator STANDARD_LOW true std_alloc_low +allocator LONG_LIVED_LOW true ll_low_alloc +allocator STANDARD_LOW true std_low_alloc +endif +else # Non smp build @@ -91,8 +96,8 @@ allocator ETS false ets_alloc allocator FIXED_SIZE false fix_alloc +if halfword -allocator LONG_LIVED_LOW false ll_alloc_low -allocator STANDARD_LOW false std_alloc_low +allocator LONG_LIVED_LOW false ll_low_alloc +allocator STANDARD_LOW false std_low_alloc +endif +endif @@ -134,6 +139,7 @@ class SYSTEM system_data # # <TYPE> <ALLOCATOR> <CLASS> <DESCRIPTION> +type SBMBC SBMBC SYSTEM small_block_mbc type PROC FIXED_SIZE PROCESSES proc type ATOM FIXED_SIZE ATOM atom_entry type MODULE FIXED_SIZE CODE module_entry @@ -330,6 +336,7 @@ type SSB SHORT_LIVED PROCESSES ssb +if halfword +type SBMBC_LOW SBMBC_LOW SYSTEM small_block_mbc_low type DDLL_PROCESS STANDARD_LOW SYSTEM ddll_processes type MONITOR_LH STANDARD_LOW PROCESSES monitor_lh type NLINK_LH STANDARD_LOW PROCESSES nlink_lh diff --git a/erts/emulator/beam/erl_alloc_util.c b/erts/emulator/beam/erl_alloc_util.c index cc04ef65bf..d51ed0c36d 100644 --- a/erts/emulator/beam/erl_alloc_util.c +++ b/erts/emulator/beam/erl_alloc_util.c @@ -66,6 +66,7 @@ static int atoms_initialized = 0; static int initialized = 0; +int erts_have_sbmbc_alloc; #if HAVE_ERTS_MSEG @@ -85,8 +86,6 @@ static int initialized = 0; #undef ASSERT #define ASSERT ASSERT_EXPR -#define ERTS_ALCU_FLG_FAIL_REALLOC_MOVE ((UWord) 1) - #if 0 /* Can be useful for debugging */ #define MBC_REALLOC_ALWAYS_MOVES @@ -275,14 +274,26 @@ static void check_blk_carrier(Allctr_t *, Block_t *); #ifdef DEBUG #define DEBUG_CHECK_CARRIER_NO_SZ(AP) \ - ASSERT(((AP)->sbcs.curr_mseg.no && (AP)->sbcs.curr_mseg.size) \ - || (!(AP)->sbcs.curr_mseg.no && !(AP)->sbcs.curr_mseg.size));\ - ASSERT(((AP)->sbcs.curr_sys_alloc.no && (AP)->sbcs.curr_sys_alloc.size)\ - || (!(AP)->sbcs.curr_sys_alloc.no && !(AP)->sbcs.curr_sys_alloc.size));\ - ASSERT(((AP)->mbcs.curr_mseg.no && (AP)->mbcs.curr_mseg.size) \ - || (!(AP)->mbcs.curr_mseg.no && !(AP)->mbcs.curr_mseg.size));\ - ASSERT(((AP)->mbcs.curr_sys_alloc.no && (AP)->mbcs.curr_sys_alloc.size)\ - || (!(AP)->mbcs.curr_sys_alloc.no && !(AP)->mbcs.curr_sys_alloc.size)) + ASSERT(((AP)->sbcs.curr.norm.mseg.no \ + && (AP)->sbcs.curr.norm.mseg.size) \ + || (!(AP)->sbcs.curr.norm.mseg.no \ + && !(AP)->sbcs.curr.norm.mseg.size)); \ + ASSERT(((AP)->sbcs.curr.norm.sys_alloc.no \ + && (AP)->sbcs.curr.norm.sys_alloc.size) \ + || (!(AP)->sbcs.curr.norm.sys_alloc.no \ + && !(AP)->sbcs.curr.norm.sys_alloc.size)); \ + ASSERT(((AP)->mbcs.curr.norm.mseg.no \ + && (AP)->mbcs.curr.norm.mseg.size) \ + || (!(AP)->mbcs.curr.norm.mseg.no \ + && !(AP)->mbcs.curr.norm.mseg.size)); \ + ASSERT(((AP)->mbcs.curr.norm.sys_alloc.no \ + && (AP)->mbcs.curr.norm.sys_alloc.size) \ + || (!(AP)->mbcs.curr.norm.sys_alloc.no \ + && !(AP)->mbcs.curr.norm.sys_alloc.size)); \ + ASSERT(((AP)->sbmbcs.curr.small_block.no \ + && (AP)->sbmbcs.curr.small_block.size) \ + || (!(AP)->sbmbcs.curr.small_block.no \ + && !(AP)->sbmbcs.curr.small_block.size)) #else #define DEBUG_CHECK_CARRIER_NO_SZ(AP) @@ -292,27 +303,27 @@ static void check_blk_carrier(Allctr_t *, Block_t *); (AP)->sbcs.blocks.curr.size += (BSZ); \ if ((AP)->sbcs.blocks.max.size < (AP)->sbcs.blocks.curr.size) \ (AP)->sbcs.blocks.max.size = (AP)->sbcs.blocks.curr.size; \ - if ((AP)->sbcs.max.no < ((AP)->sbcs.curr_mseg.no \ - + (AP)->sbcs.curr_sys_alloc.no)) \ - (AP)->sbcs.max.no = ((AP)->sbcs.curr_mseg.no \ - + (AP)->sbcs.curr_sys_alloc.no); \ - if ((AP)->sbcs.max.size < ((AP)->sbcs.curr_mseg.size \ - + (AP)->sbcs.curr_sys_alloc.size)) \ - (AP)->sbcs.max.size = ((AP)->sbcs.curr_mseg.size \ - + (AP)->sbcs.curr_sys_alloc.size) + if ((AP)->sbcs.max.no < ((AP)->sbcs.curr.norm.mseg.no \ + + (AP)->sbcs.curr.norm.sys_alloc.no)) \ + (AP)->sbcs.max.no = ((AP)->sbcs.curr.norm.mseg.no \ + + (AP)->sbcs.curr.norm.sys_alloc.no); \ + if ((AP)->sbcs.max.size < ((AP)->sbcs.curr.norm.mseg.size \ + + (AP)->sbcs.curr.norm.sys_alloc.size)) \ + (AP)->sbcs.max.size = ((AP)->sbcs.curr.norm.mseg.size \ + + (AP)->sbcs.curr.norm.sys_alloc.size) #define STAT_MSEG_SBC_ALLOC(AP, CSZ, BSZ) \ do { \ - (AP)->sbcs.curr_mseg.no++; \ - (AP)->sbcs.curr_mseg.size += (CSZ); \ + (AP)->sbcs.curr.norm.mseg.no++; \ + (AP)->sbcs.curr.norm.mseg.size += (CSZ); \ STAT_SBC_ALLOC((AP), (BSZ)); \ DEBUG_CHECK_CARRIER_NO_SZ((AP)); \ } while (0) #define STAT_SYS_ALLOC_SBC_ALLOC(AP, CSZ, BSZ) \ do { \ - (AP)->sbcs.curr_sys_alloc.no++; \ - (AP)->sbcs.curr_sys_alloc.size += (CSZ); \ + (AP)->sbcs.curr.norm.sys_alloc.no++; \ + (AP)->sbcs.curr.norm.sys_alloc.size += (CSZ); \ STAT_SBC_ALLOC((AP), (BSZ)); \ DEBUG_CHECK_CARRIER_NO_SZ((AP)); \ } while (0) @@ -324,85 +335,111 @@ do { \ #define STAT_MSEG_SBC_FREE(AP, CSZ, BSZ) \ do { \ - ASSERT((AP)->sbcs.curr_mseg.no > 0); \ - (AP)->sbcs.curr_mseg.no--; \ - ASSERT((AP)->sbcs.curr_mseg.size >= (CSZ)); \ - (AP)->sbcs.curr_mseg.size -= (CSZ); \ + ASSERT((AP)->sbcs.curr.norm.mseg.no > 0); \ + (AP)->sbcs.curr.norm.mseg.no--; \ + ASSERT((AP)->sbcs.curr.norm.mseg.size >= (CSZ)); \ + (AP)->sbcs.curr.norm.mseg.size -= (CSZ); \ STAT_SBC_FREE((AP), (BSZ)); \ DEBUG_CHECK_CARRIER_NO_SZ((AP)); \ } while (0) #define STAT_SYS_ALLOC_SBC_FREE(AP, CSZ, BSZ) \ do { \ - ASSERT((AP)->sbcs.curr_sys_alloc.no > 0); \ - (AP)->sbcs.curr_sys_alloc.no--; \ - ASSERT((AP)->sbcs.curr_sys_alloc.size >= (CSZ)); \ - (AP)->sbcs.curr_sys_alloc.size -= (CSZ); \ + ASSERT((AP)->sbcs.curr.norm.sys_alloc.no > 0); \ + (AP)->sbcs.curr.norm.sys_alloc.no--; \ + ASSERT((AP)->sbcs.curr.norm.sys_alloc.size >= (CSZ)); \ + (AP)->sbcs.curr.norm.sys_alloc.size -= (CSZ); \ STAT_SBC_FREE((AP), (BSZ)); \ DEBUG_CHECK_CARRIER_NO_SZ((AP)); \ } while (0) #define STAT_MBC_ALLOC(AP) \ - if ((AP)->mbcs.max.no < ((AP)->mbcs.curr_mseg.no \ - + (AP)->mbcs.curr_sys_alloc.no)) \ - (AP)->mbcs.max.no = ((AP)->mbcs.curr_mseg.no \ - + (AP)->mbcs.curr_sys_alloc.no); \ - if ((AP)->mbcs.max.size < ((AP)->mbcs.curr_mseg.size \ - + (AP)->mbcs.curr_sys_alloc.size)) \ - (AP)->mbcs.max.size = ((AP)->mbcs.curr_mseg.size \ - + (AP)->mbcs.curr_sys_alloc.size) + if ((AP)->mbcs.max.no < ((AP)->mbcs.curr.norm.mseg.no \ + + (AP)->mbcs.curr.norm.sys_alloc.no)) \ + (AP)->mbcs.max.no = ((AP)->mbcs.curr.norm.mseg.no \ + + (AP)->mbcs.curr.norm.sys_alloc.no); \ + if ((AP)->mbcs.max.size < ((AP)->mbcs.curr.norm.mseg.size \ + + (AP)->mbcs.curr.norm.sys_alloc.size)) \ + (AP)->mbcs.max.size = ((AP)->mbcs.curr.norm.mseg.size \ + + (AP)->mbcs.curr.norm.sys_alloc.size) + +#define STAT_SBMBC_ALLOC(AP, CSZ) \ +do { \ + (AP)->sbmbcs.curr.small_block.no++; \ + (AP)->sbmbcs.curr.small_block.size += (CSZ); \ + if ((AP)->sbmbcs.max.no < (AP)->sbmbcs.curr.small_block.no) \ + (AP)->sbmbcs.max.no = (AP)->sbmbcs.curr.small_block.no; \ + if ((AP)->sbmbcs.max.size < (AP)->sbmbcs.curr.small_block.size) \ + (AP)->sbmbcs.max.size = (AP)->sbmbcs.curr.small_block.size; \ + DEBUG_CHECK_CARRIER_NO_SZ((AP)); \ +} while (0) #define STAT_MSEG_MBC_ALLOC(AP, CSZ) \ do { \ - (AP)->mbcs.curr_mseg.no++; \ - (AP)->mbcs.curr_mseg.size += (CSZ); \ + (AP)->mbcs.curr.norm.mseg.no++; \ + (AP)->mbcs.curr.norm.mseg.size += (CSZ); \ STAT_MBC_ALLOC((AP)); \ DEBUG_CHECK_CARRIER_NO_SZ((AP)); \ } while (0) #define STAT_SYS_ALLOC_MBC_ALLOC(AP, CSZ) \ do { \ - (AP)->mbcs.curr_sys_alloc.no++; \ - (AP)->mbcs.curr_sys_alloc.size += (CSZ); \ + (AP)->mbcs.curr.norm.sys_alloc.no++; \ + (AP)->mbcs.curr.norm.sys_alloc.size += (CSZ); \ STAT_MBC_ALLOC((AP)); \ DEBUG_CHECK_CARRIER_NO_SZ((AP)); \ } while (0) +#define STAT_SBMBC_FREE(AP, CSZ) \ +do { \ + ASSERT((AP)->sbmbcs.curr.small_block.no > 0); \ + (AP)->sbmbcs.curr.small_block.no--; \ + ASSERT((AP)->sbmbcs.curr.small_block.size >= (CSZ)); \ + (AP)->sbmbcs.curr.small_block.size -= (CSZ); \ + DEBUG_CHECK_CARRIER_NO_SZ((AP)); \ +} while (0) + #define STAT_MSEG_MBC_FREE(AP, CSZ) \ do { \ - ASSERT((AP)->mbcs.curr_mseg.no > 0); \ - (AP)->mbcs.curr_mseg.no--; \ - ASSERT((AP)->mbcs.curr_mseg.size >= (CSZ)); \ - (AP)->mbcs.curr_mseg.size -= (CSZ); \ + ASSERT((AP)->mbcs.curr.norm.mseg.no > 0); \ + (AP)->mbcs.curr.norm.mseg.no--; \ + ASSERT((AP)->mbcs.curr.norm.mseg.size >= (CSZ)); \ + (AP)->mbcs.curr.norm.mseg.size -= (CSZ); \ DEBUG_CHECK_CARRIER_NO_SZ((AP)); \ } while (0) #define STAT_SYS_ALLOC_MBC_FREE(AP, CSZ) \ do { \ - ASSERT((AP)->mbcs.curr_sys_alloc.no > 0); \ - (AP)->mbcs.curr_sys_alloc.no--; \ - ASSERT((AP)->mbcs.curr_sys_alloc.size >= (CSZ)); \ - (AP)->mbcs.curr_sys_alloc.size -= (CSZ); \ + ASSERT((AP)->mbcs.curr.norm.sys_alloc.no > 0); \ + (AP)->mbcs.curr.norm.sys_alloc.no--; \ + ASSERT((AP)->mbcs.curr.norm.sys_alloc.size >= (CSZ)); \ + (AP)->mbcs.curr.norm.sys_alloc.size -= (CSZ); \ DEBUG_CHECK_CARRIER_NO_SZ((AP)); \ } while (0) -#define STAT_MBC_BLK_ALLOC(AP, BSZ) \ +#define STAT_MBC_BLK_ALLOC(AP, BSZ, FLGS) \ do { \ - (AP)->mbcs.blocks.curr.no++; \ - if ((AP)->mbcs.blocks.max.no < (AP)->mbcs.blocks.curr.no) \ - (AP)->mbcs.blocks.max.no = (AP)->mbcs.blocks.curr.no; \ - (AP)->mbcs.blocks.curr.size += (BSZ); \ - if ((AP)->mbcs.blocks.max.size < (AP)->mbcs.blocks.curr.size) \ - (AP)->mbcs.blocks.max.size = (AP)->mbcs.blocks.curr.size; \ + CarriersStats_t *cstats__ = (((FLGS) & ERTS_ALCU_FLG_SBMBC) \ + ? &(AP)->sbmbcs \ + : &(AP)->mbcs); \ + cstats__->blocks.curr.no++; \ + if (cstats__->blocks.max.no < cstats__->blocks.curr.no) \ + cstats__->blocks.max.no = cstats__->blocks.curr.no; \ + cstats__->blocks.curr.size += (BSZ); \ + if (cstats__->blocks.max.size < cstats__->blocks.curr.size) \ + cstats__->blocks.max.size = cstats__->blocks.curr.size; \ } while (0) -#define STAT_MBC_BLK_FREE(AP, BSZ) \ +#define STAT_MBC_BLK_FREE(AP, BSZ, FLGS) \ do { \ - ASSERT((AP)->mbcs.blocks.curr.no > 0); \ - (AP)->mbcs.blocks.curr.no--; \ - ASSERT((AP)->mbcs.blocks.curr.size >= (BSZ)); \ - (AP)->mbcs.blocks.curr.size -= (BSZ); \ + CarriersStats_t *cstats__ = (((FLGS) & ERTS_ALCU_FLG_SBMBC) \ + ? &(AP)->sbmbcs \ + : &(AP)->mbcs); \ + ASSERT(cstats__->blocks.curr.no > 0); \ + cstats__->blocks.curr.no--; \ + ASSERT(cstats__->blocks.curr.size >= (BSZ)); \ + cstats__->blocks.curr.size -= (BSZ); \ } while (0) /* Debug stuff... */ @@ -410,7 +447,7 @@ do { \ static UWord carrier_alignment; #define DEBUG_SAVE_ALIGNMENT(C) \ do { \ - UWord algnmnt__ = sizeof(Unit_t) - (((UWord) (C)) % sizeof(Unit_t)); \ + UWord algnmnt__ = sizeof(Unit_t) - (((UWord) (C)) % sizeof(Unit_t));\ carrier_alignment = MIN(carrier_alignment, algnmnt__); \ ASSERT(((UWord) (C)) % sizeof(UWord) == 0); \ } while (0) @@ -524,8 +561,8 @@ static Uint get_next_mbc_size(Allctr_t *allctr) { Uint size; - int cs = (allctr->mbcs.curr_mseg.no - + allctr->mbcs.curr_sys_alloc.no + int cs = (allctr->mbcs.curr.norm.mseg.no + + allctr->mbcs.curr.norm.sys_alloc.no - (allctr->main_carrier ? 1 : 0)); ASSERT(cs >= 0); @@ -609,7 +646,8 @@ unlink_carrier(CarrierList_t *cl, Carrier_t *crr) } } - +static Block_t *create_sbmbc(Allctr_t *allctr, Uint umem_sz); +static void destroy_sbmbc(Allctr_t *allctr, Block_t *blk); static Block_t *create_carrier(Allctr_t *, Uint, UWord); static void destroy_carrier(Allctr_t *, Block_t *); @@ -619,39 +657,57 @@ static void destroy_carrier(Allctr_t *, Block_t *); * block in a sbc. */ static ERTS_INLINE void * -mbc_alloc_block(Allctr_t *allctr, Uint size, Uint *blk_szp) +mbc_alloc_block(Allctr_t *allctr, Uint size, Uint *blk_szp, Uint32 *alcu_flgsp) { Block_t *blk; + Uint get_blk_sz; + Uint sbmbct; ASSERT(size); ASSERT(size < allctr->sbc_threshold); - *blk_szp = UMEMSZ2BLKSZ(allctr, size); + *blk_szp = get_blk_sz = UMEMSZ2BLKSZ(allctr, size); + + sbmbct = allctr->sbmbc_threshold; + if (sbmbct) { + if (get_blk_sz < sbmbct) { + *alcu_flgsp |= ERTS_ALCU_FLG_SBMBC; + if (get_blk_sz + allctr->min_block_size > sbmbct) { + /* Since we use block size to determine if blocks are + located in sbmbc or not... */ + get_blk_sz += allctr->min_block_size; + } + } + } - blk = (*allctr->get_free_block)(allctr, *blk_szp, NULL, 0); + blk = (*allctr->get_free_block)(allctr, get_blk_sz, NULL, 0, *alcu_flgsp); -#if HALFWORD_HEAP if (!blk) { - blk = create_carrier(allctr, *blk_szp, CFLG_MBC|CFLG_FORCE_MSEG); - } + if ((*alcu_flgsp) & ERTS_ALCU_FLG_SBMBC) + blk = create_sbmbc(allctr, get_blk_sz); + else { +#if HALFWORD_HEAP + blk = create_carrier(allctr, get_blk_sz, CFLG_MBC|CFLG_FORCE_MSEG); #else - if (!blk) { - blk = create_carrier(allctr, *blk_szp, CFLG_MBC); - if (!blk) { - /* Emergency! We couldn't create the carrier as we wanted. - Try to place it in a sys_alloced sbc. */ - blk = create_carrier(allctr, - size, - CFLG_SBC|CFLG_FORCE_SIZE|CFLG_FORCE_SYS_ALLOC); + blk = create_carrier(allctr, get_blk_sz, CFLG_MBC); + if (!blk) { + /* Emergency! We couldn't create the carrier as we wanted. + Try to place it in a sys_alloced sbc. */ + blk = create_carrier(allctr, + size, + (CFLG_SBC + | CFLG_FORCE_SIZE + | CFLG_FORCE_SYS_ALLOC)); + } +#endif } } -#endif #ifdef ERTS_ALLOC_UTIL_HARD_DEBUG if (IS_MBC_BLK(blk)) { - (*allctr->link_free_block)(allctr, blk); + (*allctr->link_free_block)(allctr, blk, *alcu_flgsp); HARD_CHECK_BLK_CARRIER(allctr, blk); - (*allctr->unlink_free_block)(allctr, blk); + (*allctr->unlink_free_block)(allctr, blk, *alcu_flgsp); } #endif @@ -664,7 +720,8 @@ mbc_alloc_finalize(Allctr_t *allctr, Uint org_blk_sz, UWord flags, Uint want_blk_sz, - int valid_blk_info) + int valid_blk_info, + Uint32 alcu_flgs) { Uint blk_sz; Uint nxt_blk_sz; @@ -700,7 +757,7 @@ mbc_alloc_finalize(Allctr_t *allctr, SET_PREV_BLK_FREE(nxt_nxt_blk); } } - (*allctr->link_free_block)(allctr, nxt_blk); + (*allctr->link_free_block)(allctr, nxt_blk, alcu_flgs); ASSERT(IS_NOT_LAST_BLK(blk)); ASSERT(IS_FREE_BLK(nxt_blk)); @@ -741,7 +798,7 @@ mbc_alloc_finalize(Allctr_t *allctr, : IS_NOT_LAST_BLK(blk)); } - STAT_MBC_BLK_ALLOC(allctr, blk_sz); + STAT_MBC_BLK_ALLOC(allctr, blk_sz, alcu_flgs); ASSERT(IS_ALLOCED_BLK(blk)); ASSERT(blk_sz == BLK_SZ(blk)); @@ -761,7 +818,8 @@ mbc_alloc(Allctr_t *allctr, Uint size) { Block_t *blk; Uint blk_sz; - blk = mbc_alloc_block(allctr, size, &blk_sz); + Uint32 alcu_flgs = 0; + blk = mbc_alloc_block(allctr, size, &blk_sz, &alcu_flgs); if (!blk) return NULL; if (IS_MBC_BLK(blk)) @@ -770,7 +828,8 @@ mbc_alloc(Allctr_t *allctr, Uint size) BLK_SZ(blk), GET_BLK_HDR_FLGS(blk), blk_sz, - 1); + 1, + alcu_flgs); return BLK2UMEM(blk); } @@ -779,6 +838,7 @@ mbc_free(Allctr_t *allctr, void *p) { Uint is_first_blk; Uint is_last_blk; + Uint32 alcu_flgs = 0; Uint blk_sz; Block_t *blk; Block_t *nxt_blk; @@ -788,13 +848,15 @@ mbc_free(Allctr_t *allctr, void *p) blk = UMEM2BLK(p); blk_sz = BLK_SZ(blk); + if (blk_sz < allctr->sbmbc_threshold) + alcu_flgs |= ERTS_ALCU_FLG_SBMBC; ASSERT(IS_MBC_BLK(blk)); ASSERT(blk_sz >= allctr->min_block_size); HARD_CHECK_BLK_CARRIER(allctr, blk); - STAT_MBC_BLK_FREE(allctr, blk_sz); + STAT_MBC_BLK_FREE(allctr, blk_sz, alcu_flgs); is_first_blk = IS_FIRST_BLK(blk); is_last_blk = IS_LAST_BLK(blk); @@ -802,7 +864,7 @@ mbc_free(Allctr_t *allctr, void *p) if (!is_first_blk && IS_PREV_BLK_FREE(blk)) { /* Coalesce with previous block... */ blk = PREV_BLK(blk); - (*allctr->unlink_free_block)(allctr, blk); + (*allctr->unlink_free_block)(allctr, blk, alcu_flgs); blk_sz += BLK_SZ(blk); is_first_blk = IS_FIRST_BLK(blk); @@ -818,7 +880,7 @@ mbc_free(Allctr_t *allctr, void *p) nxt_blk = NXT_BLK(blk); if (IS_FREE_BLK(nxt_blk)) { /* Coalesce with next block... */ - (*allctr->unlink_free_block)(allctr, nxt_blk); + (*allctr->unlink_free_block)(allctr, nxt_blk, alcu_flgs); blk_sz += BLK_SZ(nxt_blk); SET_BLK_SZ(blk, blk_sz); @@ -850,16 +912,20 @@ mbc_free(Allctr_t *allctr, void *p) if (is_first_blk && is_last_blk - && allctr->main_carrier != FBLK2MBC(allctr, blk)) - destroy_carrier(allctr, blk); + && allctr->main_carrier != FBLK2MBC(allctr, blk)) { + if (alcu_flgs & ERTS_ALCU_FLG_SBMBC) + destroy_sbmbc(allctr, blk); + else + destroy_carrier(allctr, blk); + } else { - (*allctr->link_free_block)(allctr, blk); + (*allctr->link_free_block)(allctr, blk, alcu_flgs); HARD_CHECK_BLK_CARRIER(allctr, blk); } } static void * -mbc_realloc(Allctr_t *allctr, void *p, Uint size, UWord flgs) +mbc_realloc(Allctr_t *allctr, void *p, Uint size, Uint32 alcu_flgs) { void *new_p; Uint old_blk_sz; @@ -867,7 +933,7 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, UWord flgs) #ifndef MBC_REALLOC_ALWAYS_MOVES Block_t *new_blk, *cand_blk; Uint cand_blk_sz; - Uint blk_sz; + Uint blk_sz, get_blk_sz; Block_t *nxt_blk; Uint nxt_blk_sz; Uint is_last_blk; @@ -883,10 +949,16 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, UWord flgs) ASSERT(old_blk_sz >= allctr->min_block_size); #ifdef MBC_REALLOC_ALWAYS_MOVES - if (flgs & ERTS_ALCU_FLG_FAIL_REALLOC_MOVE) + if (alcu_flgs & ERTS_ALCU_FLG_FAIL_REALLOC_MOVE) return NULL; #else /* !MBC_REALLOC_ALWAYS_MOVES */ - blk_sz = UMEMSZ2BLKSZ(allctr, size); + get_blk_sz = blk_sz = UMEMSZ2BLKSZ(allctr, size); + if ((alcu_flgs & ERTS_ALCU_FLG_SBMBC) + && (blk_sz + allctr->min_block_size > allctr->sbmbc_threshold)) { + /* Since we use block size to determine if blocks are + located in sbmbc or not... */ + get_blk_sz = blk_sz + allctr->min_block_size; + } ASSERT(IS_ALLOCED_BLK(blk)); ASSERT(IS_MBC_BLK(blk)); @@ -901,6 +973,9 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, UWord flgs) Uint diff_sz_val = old_blk_sz - blk_sz; Uint old_blk_sz_val = old_blk_sz; + if (get_blk_sz >= old_blk_sz) + return p; + if (diff_sz_val >= (~((Uint) 0) / 100)) { /* div both by 128 */ old_blk_sz_val >>= 7; @@ -909,7 +984,7 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, UWord flgs) /* Avoid fragmentation by moving the block if it is shrunk much */ if (100*diff_sz_val > allctr->mbc_move_threshold*old_blk_sz_val) { - if (flgs & ERTS_ALCU_FLG_FAIL_REALLOC_MOVE) + if (alcu_flgs & ERTS_ALCU_FLG_FAIL_REALLOC_MOVE) return NULL; cand_blk_sz = old_blk_sz; @@ -926,9 +1001,10 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, UWord flgs) } new_blk = (*allctr->get_free_block)(allctr, - blk_sz, + get_blk_sz, cand_blk, - cand_blk_sz); + cand_blk_sz, + alcu_flgs); if (new_blk || cand_blk != blk) goto move_into_new_blk; @@ -952,8 +1028,8 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, UWord flgs) nxt_blk_sz, SBH_THIS_FREE|SBH_PREV_ALLOCED|SBH_NOT_LAST_BLK); - STAT_MBC_BLK_FREE(allctr, old_blk_sz); - STAT_MBC_BLK_ALLOC(allctr, blk_sz); + STAT_MBC_BLK_FREE(allctr, old_blk_sz, alcu_flgs); + STAT_MBC_BLK_ALLOC(allctr, blk_sz, alcu_flgs); ASSERT(BLK_SZ(blk) >= allctr->min_block_size); @@ -964,7 +1040,7 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, UWord flgs) if (IS_FREE_BLK(nxt_nxt_blk)) { /* Coalesce with next free block... */ nxt_blk_sz += BLK_SZ(nxt_nxt_blk); - (*allctr->unlink_free_block)(allctr, nxt_nxt_blk); + (*allctr->unlink_free_block)(allctr, nxt_nxt_blk, alcu_flgs); SET_BLK_SZ(nxt_blk, nxt_blk_sz); is_last_blk = IS_LAST_BLK(nxt_nxt_blk); @@ -979,7 +1055,7 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, UWord flgs) } } - (*allctr->link_free_block)(allctr, nxt_blk); + (*allctr->link_free_block)(allctr, nxt_blk, alcu_flgs); ASSERT(IS_ALLOCED_BLK(blk)); @@ -1009,12 +1085,12 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, UWord flgs) if (!is_last_blk) { nxt_blk = NXT_BLK(blk); nxt_blk_sz = BLK_SZ(nxt_blk); - if (IS_FREE_BLK(nxt_blk) && blk_sz <= old_blk_sz + nxt_blk_sz) { + if (IS_FREE_BLK(nxt_blk) && get_blk_sz <= old_blk_sz + nxt_blk_sz) { /* Grow into next block... */ HARD_CHECK_BLK_CARRIER(allctr, blk); - (*allctr->unlink_free_block)(allctr, nxt_blk); + (*allctr->unlink_free_block)(allctr, nxt_blk, alcu_flgs); nxt_blk_sz -= blk_sz - old_blk_sz; is_last_blk = IS_LAST_BLK(nxt_blk); @@ -1051,13 +1127,13 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, UWord flgs) else SET_BLK_SZ_FTR(nxt_blk, nxt_blk_sz); - (*allctr->link_free_block)(allctr, nxt_blk); + (*allctr->link_free_block)(allctr, nxt_blk, alcu_flgs); ASSERT(IS_FREE_BLK(nxt_blk)); } - STAT_MBC_BLK_FREE(allctr, old_blk_sz); - STAT_MBC_BLK_ALLOC(allctr, blk_sz); + STAT_MBC_BLK_FREE(allctr, old_blk_sz, alcu_flgs); + STAT_MBC_BLK_ALLOC(allctr, blk_sz, alcu_flgs); ASSERT(IS_ALLOCED_BLK(blk)); @@ -1088,7 +1164,7 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, UWord flgs) } } - if (flgs & ERTS_ALCU_FLG_FAIL_REALLOC_MOVE) + if (alcu_flgs & ERTS_ALCU_FLG_FAIL_REALLOC_MOVE) return NULL; /* Need to grow in another block */ @@ -1108,7 +1184,7 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, UWord flgs) } } - if (cand_blk_sz < blk_sz) { + if (cand_blk_sz < get_blk_sz) { /* We wont fit in cand_blk get a new one */ #endif /* !MBC_REALLOC_ALWAYS_MOVES */ @@ -1127,9 +1203,10 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, UWord flgs) /* We will at least fit in cand_blk */ new_blk = (*allctr->get_free_block)(allctr, - blk_sz, + get_blk_sz, cand_blk, - cand_blk_sz); + cand_blk_sz, + alcu_flgs); move_into_new_blk: /* * new_blk, and cand_blk have to be correctly set @@ -1142,7 +1219,8 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, UWord flgs) BLK_SZ(new_blk), GET_BLK_HDR_FLGS(new_blk), blk_sz, - 1); + 1, + alcu_flgs); new_p = BLK2UMEM(new_blk); sys_memcpy(new_p, p, MIN(size, old_blk_sz - ABLK_HDR_SZ)); mbc_free(allctr, p); @@ -1164,7 +1242,7 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, UWord flgs) HARD_CHECK_BLK_CARRIER(allctr, blk); - (*allctr->unlink_free_block)(allctr, new_blk); /* prev */ + (*allctr->unlink_free_block)(allctr, new_blk, alcu_flgs); /* prev */ if (is_last_blk) new_blk_flgs |= LAST_BLK_HDR_FLG; @@ -1173,7 +1251,7 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, UWord flgs) if (IS_FREE_BLK(nxt_blk)) { new_blk_flgs |= GET_LAST_BLK_HDR_FLG(nxt_blk); new_blk_sz += BLK_SZ(nxt_blk); - (*allctr->unlink_free_block)(allctr, nxt_blk); + (*allctr->unlink_free_block)(allctr, nxt_blk, alcu_flgs); } } @@ -1196,9 +1274,10 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, UWord flgs) new_blk_sz, new_blk_flgs, blk_sz, - 0); + 0, + alcu_flgs); - STAT_MBC_BLK_FREE(allctr, old_blk_sz); + STAT_MBC_BLK_FREE(allctr, old_blk_sz, alcu_flgs); return new_p; } @@ -1243,6 +1322,100 @@ do { \ #define CHECK_1BLK_CARRIER(A, SBC, MSEGED, C, CSZ, B, BSZ) #endif +static Block_t * +create_sbmbc(Allctr_t *allctr, Uint umem_sz) +{ + Block_t *blk; + Uint blk_sz; + Uint crr_sz = allctr->sbmbc_size; + Carrier_t *crr; + +#if HALFWORD_HEAP + if (allctr->mseg_opt.low_mem) + crr = erts_alloc(ERTS_ALC_T_SBMBC_LOW, crr_sz); + else +#endif + crr = erts_alloc(ERTS_ALC_T_SBMBC, crr_sz); + + INC_CC(allctr->calls.sbmbc_alloc); + SET_CARRIER_HDR(crr, crr_sz, SCH_SYS_ALLOC|SCH_MBC); + + blk = MBC2FBLK(allctr, crr); + +#ifdef ERTS_ALLOC_UTIL_HARD_DEBUG + if (allctr->mbc_header_size % sizeof(Unit_t) == 0) + crr_sz -= sizeof(UWord); +#endif + + blk_sz = UNIT_FLOOR(crr_sz - allctr->mbc_header_size); + + SET_MBC_BLK_FTR(((UWord *) blk)[-1]); + SET_BLK_HDR(blk, blk_sz, SBH_THIS_FREE|SBH_PREV_FREE|SBH_LAST_BLK); + +#ifdef ERTS_ALLOC_UTIL_HARD_DEBUG + *((Carrier_t **) NXT_BLK(blk)) = crr; +#endif + + link_carrier(&allctr->sbmbc_list, crr); + +#ifdef ERTS_ALLOC_UTIL_HARD_DEBUG + if (allctr->mbc_header_size % sizeof(Unit_t) == 0) + crr_sz += sizeof(UWord); +#endif + + STAT_SBMBC_ALLOC(allctr, crr_sz); + CHECK_1BLK_CARRIER(allctr, 0, 0, crr, crr_sz, blk, blk_sz); +#ifdef ERTS_ALLOC_UTIL_HARD_DEBUG + if (allctr->mbc_header_size % sizeof(Unit_t) == 0) + crr_sz -= sizeof(UWord); +#endif + if (allctr->creating_mbc) + (*allctr->creating_mbc)(allctr, crr, ERTS_ALCU_FLG_SBMBC); + + DEBUG_SAVE_ALIGNMENT(crr); + return blk; +} + +static void +destroy_sbmbc(Allctr_t *allctr, Block_t *blk) +{ + Uint crr_sz; + Carrier_t *crr; + + ASSERT(IS_FIRST_BLK(blk)); + + ASSERT(IS_MBC_BLK(blk)); + + crr = FBLK2MBC(allctr, blk); + crr_sz = CARRIER_SZ(crr); + +#ifdef DEBUG + if (!allctr->stopped) { + ASSERT(IS_LAST_BLK(blk)); + +#ifdef ERTS_ALLOC_UTIL_HARD_DEBUG + (*allctr->link_free_block)(allctr, blk, ERTS_ALCU_FLG_SBMBC); + HARD_CHECK_BLK_CARRIER(allctr, blk); + (*allctr->unlink_free_block)(allctr, blk, ERTS_ALCU_FLG_SBMBC); +#endif + } +#endif + + STAT_SBMBC_FREE(allctr, crr_sz); + + unlink_carrier(&allctr->sbmbc_list, crr); + if (allctr->destroying_mbc) + (*allctr->destroying_mbc)(allctr, crr, ERTS_ALCU_FLG_SBMBC); + + INC_CC(allctr->calls.sbmbc_free); + +#if HALFWORD_HEAP + if (allctr->mseg_opt.low_mem) + erts_free(ERTS_ALC_T_SBMBC_LOW, crr); + else +#endif + erts_free(ERTS_ALC_T_SBMBC, crr); +} static Block_t * create_carrier(Allctr_t *allctr, Uint umem_sz, UWord flags) @@ -1271,11 +1444,11 @@ create_carrier(Allctr_t *allctr, Uint umem_sz, UWord flags) if (erts_mseg_no() >= max_mseg_carriers) goto try_sys_alloc; if (flags & CFLG_SBC) { - if (allctr->sbcs.curr_mseg.no >= allctr->max_mseg_sbcs) + if (allctr->sbcs.curr.norm.mseg.no >= allctr->max_mseg_sbcs) goto try_sys_alloc; } else { - if (allctr->mbcs.curr_mseg.no >= allctr->max_mseg_mbcs) + if (allctr->mbcs.curr.norm.mseg.no >= allctr->max_mseg_mbcs) goto try_sys_alloc; } @@ -1289,7 +1462,7 @@ create_carrier(Allctr_t *allctr, Uint umem_sz, UWord flags) if (crr_sz < allctr->mbc_header_size + blk_sz) crr_sz = allctr->mbc_header_size + blk_sz; #ifdef ERTS_ALLOC_UTIL_HARD_DEBUG - if (sizeof(Unit_t) == sizeof(UWord)) + if (allctr->mbc_header_size % sizeof(Unit_t) == 0) crr_sz += sizeof(UWord); #endif } @@ -1330,7 +1503,7 @@ create_carrier(Allctr_t *allctr, Uint umem_sz, UWord flags) && bcrr_sz < allctr->smallest_mbc_size) bcrr_sz = allctr->smallest_mbc_size; #ifdef ERTS_ALLOC_UTIL_HARD_DEBUG - if (sizeof(Unit_t) == sizeof(UWord)) + if (allctr->mbc_header_size % sizeof(Unit_t) == 0) bcrr_sz += sizeof(UWord); #endif @@ -1385,7 +1558,7 @@ create_carrier(Allctr_t *allctr, Uint umem_sz, UWord flags) blk = MBC2FBLK(allctr, crr); #ifdef ERTS_ALLOC_UTIL_HARD_DEBUG - if (sizeof(Unit_t) == sizeof(UWord)) + if (allctr->mbc_header_size % sizeof(Unit_t) == 0) crr_sz -= sizeof(UWord); #endif @@ -1406,16 +1579,16 @@ create_carrier(Allctr_t *allctr, Uint umem_sz, UWord flags) link_carrier(&allctr->mbc_list, crr); #ifdef ERTS_ALLOC_UTIL_HARD_DEBUG - if (sizeof(Unit_t) == sizeof(UWord)) + if (allctr->mbc_header_size % sizeof(Unit_t) == 0) crr_sz += sizeof(UWord); #endif CHECK_1BLK_CARRIER(allctr, 0, is_mseg, crr, crr_sz, blk, blk_sz); #ifdef ERTS_ALLOC_UTIL_HARD_DEBUG - if (sizeof(Unit_t) == sizeof(UWord)) + if (allctr->mbc_header_size % sizeof(Unit_t) == 0) crr_sz -= sizeof(UWord); #endif if (allctr->creating_mbc) - (*allctr->creating_mbc)(allctr, crr); + (*allctr->creating_mbc)(allctr, crr, 0); } @@ -1595,9 +1768,9 @@ destroy_carrier(Allctr_t *allctr, Block_t *blk) ASSERT(IS_LAST_BLK(blk)); #ifdef ERTS_ALLOC_UTIL_HARD_DEBUG - (*allctr->link_free_block)(allctr, blk); + (*allctr->link_free_block)(allctr, blk, 0); HARD_CHECK_BLK_CARRIER(allctr, blk); - (*allctr->unlink_free_block)(allctr, blk); + (*allctr->unlink_free_block)(allctr, blk, 0); #endif } #endif @@ -1614,7 +1787,7 @@ destroy_carrier(Allctr_t *allctr, Block_t *blk) unlink_carrier(&allctr->mbc_list, crr); if (allctr->destroying_mbc) - (*allctr->destroying_mbc)(allctr, crr); + (*allctr->destroying_mbc)(allctr, crr, 0); } @@ -1658,12 +1831,15 @@ static struct { Eterm lmbcs; Eterm smbcs; Eterm mbcgs; + Eterm sbmbcs; + Eterm sbmbct; #if HAVE_ERTS_MSEG Eterm mmc; #endif Eterm ycs; + /* Eterm sbmbcs; */ Eterm mbcs; Eterm sbcs; Eterm sys_alloc_carriers_size; @@ -1688,6 +1864,8 @@ static struct { Eterm mseg_dealloc; Eterm mseg_realloc; #endif + Eterm sbmbc_alloc; + Eterm sbmbc_free; #ifdef DEBUG Eterm end_of_atoms; #endif @@ -1746,12 +1924,15 @@ init_atoms(Allctr_t *allctr) AM_INIT(lmbcs); AM_INIT(smbcs); AM_INIT(mbcgs); + AM_INIT(sbmbcs); + AM_INIT(sbmbct); #if HAVE_ERTS_MSEG AM_INIT(mmc); #endif AM_INIT(ycs); + /*AM_INIT(sbmbcs);*/ AM_INIT(mbcs); AM_INIT(sbcs); AM_INIT(sys_alloc_carriers_size); @@ -1776,6 +1957,8 @@ init_atoms(Allctr_t *allctr) AM_INIT(mseg_dealloc); AM_INIT(mseg_realloc); #endif + AM_INIT(sbmbc_free); + AM_INIT(sbmbc_alloc); #ifdef DEBUG for (atom = (Eterm *) &am; atom < &am.end_of_atoms; atom++) { @@ -1869,7 +2052,9 @@ sz_info_carriers(Allctr_t *allctr, Uint *szp) { Eterm res = THE_NON_VALUE; - Uint curr_size = cs->curr_mseg.size + cs->curr_sys_alloc.size; + Uint curr_size = (cs == &allctr->sbmbcs + ? cs->curr.small_block.size + : cs->curr.norm.mseg.size + cs->curr.norm.sys_alloc.size); if (print_to_p) { int to = *print_to_p; @@ -1917,8 +2102,17 @@ info_carriers(Allctr_t *allctr, Uint *szp) { Eterm res = THE_NON_VALUE; - Uint curr_no = cs->curr_mseg.no + cs->curr_sys_alloc.no; - Uint curr_size = cs->curr_mseg.size + cs->curr_sys_alloc.size; + Uint curr_no, curr_size; + int small_block = cs == &allctr->sbmbcs; + + if (small_block) { + curr_no = cs->curr.small_block.no; + curr_size = cs->curr.small_block.size; + } + else { + curr_no = cs->curr.norm.mseg.no + cs->curr.norm.sys_alloc.no; + curr_size = cs->curr.norm.mseg.size + cs->curr.norm.sys_alloc.size; + } if (print_to_p) { int to = *print_to_p; @@ -1944,18 +2138,20 @@ info_carriers(Allctr_t *allctr, curr_no, cs->max.no, cs->max_ever.no); + if (!small_block) { #if HAVE_ERTS_MSEG - erts_print(to, - arg, - "%smseg carriers: %bpu\n", - prefix, - cs->curr_mseg.no); -#endif - erts_print(to, - arg, - "%ssys_alloc carriers: %bpu\n", - prefix, - cs->curr_sys_alloc.no); + erts_print(to, + arg, + "%smseg carriers: %bpu\n", + prefix, + cs->curr.norm.mseg.no); +#endif + erts_print(to, + arg, + "%ssys_alloc carriers: %bpu\n", + prefix, + cs->curr.norm.sys_alloc.no); + } erts_print(to, arg, "%scarriers size: %beu %bpu %bpu\n", @@ -1963,43 +2159,49 @@ info_carriers(Allctr_t *allctr, curr_size, cs->max.size, cs->max_ever.size); + if (!small_block) { #if HAVE_ERTS_MSEG - erts_print(to, - arg, - "%smseg carriers size: %bpu\n", - prefix, - cs->curr_mseg.size); -#endif - erts_print(to, - arg, - "%ssys_alloc carriers size: %bpu\n", - prefix, - cs->curr_sys_alloc.size); + erts_print(to, + arg, + "%smseg carriers size: %bpu\n", + prefix, + cs->curr.norm.mseg.size); +#endif + erts_print(to, + arg, + "%ssys_alloc carriers size: %bpu\n", + prefix, + cs->curr.norm.sys_alloc.size); + } } if (hpp || szp) { res = NIL; - add_2tup(hpp, szp, &res, - am.sys_alloc_carriers_size, - bld_unstable_uint(hpp, szp, cs->curr_sys_alloc.size)); + if (!small_block) { + add_2tup(hpp, szp, &res, + am.sys_alloc_carriers_size, + bld_unstable_uint(hpp, szp, cs->curr.norm.sys_alloc.size)); #if HAVE_ERTS_MSEG - add_2tup(hpp, szp, &res, - am.mseg_alloc_carriers_size, - bld_unstable_uint(hpp, szp, cs->curr_mseg.size)); + add_2tup(hpp, szp, &res, + am.mseg_alloc_carriers_size, + bld_unstable_uint(hpp, szp, cs->curr.norm.mseg.size)); #endif + } add_4tup(hpp, szp, &res, am.carriers_size, bld_unstable_uint(hpp, szp, curr_size), bld_unstable_uint(hpp, szp, cs->max.size), bld_unstable_uint(hpp, szp, cs->max_ever.size)); - add_2tup(hpp, szp, &res, - am.sys_alloc_carriers, - bld_unstable_uint(hpp, szp, cs->curr_sys_alloc.no)); + if (!small_block) { + add_2tup(hpp, szp, &res, + am.sys_alloc_carriers, + bld_unstable_uint(hpp, szp, cs->curr.norm.sys_alloc.no)); #if HAVE_ERTS_MSEG - add_2tup(hpp, szp, &res, - am.mseg_alloc_carriers, - bld_unstable_uint(hpp, szp, cs->curr_mseg.no)); + add_2tup(hpp, szp, &res, + am.mseg_alloc_carriers, + bld_unstable_uint(hpp, szp, cs->curr.norm.mseg.no)); #endif + } add_4tup(hpp, szp, &res, am.carriers, bld_unstable_uint(hpp, szp, curr_no), @@ -2077,6 +2279,9 @@ info_calls(Allctr_t *allctr, PRINT_CC_5(to, arg, prefix, "free", allctr->calls.this_free); PRINT_CC_5(to, arg, prefix, "realloc", allctr->calls.this_realloc); + PRINT_CC_4(to, arg, "sbmbc_alloc", allctr->calls.sbmbc_alloc); + PRINT_CC_4(to, arg, "sbmbc_free", allctr->calls.sbmbc_free); + #if HAVE_ERTS_MSEG PRINT_CC_4(to, arg, "mseg_alloc", allctr->calls.mseg_alloc); PRINT_CC_4(to, arg, "mseg_dealloc", allctr->calls.mseg_dealloc); @@ -2128,6 +2333,14 @@ info_calls(Allctr_t *allctr, bld_unstable_uint(hpp, szp, allctr->calls.mseg_alloc.no)); #endif add_3tup(hpp, szp, &res, + am.sbmbc_free, + bld_unstable_uint(hpp, szp, allctr->calls.sbmbc_free.giga_no), + bld_unstable_uint(hpp, szp, allctr->calls.sbmbc_free.no)); + add_3tup(hpp, szp, &res, + am.sbmbc_alloc, + bld_unstable_uint(hpp, szp, allctr->calls.sbmbc_alloc.giga_no), + bld_unstable_uint(hpp, szp, allctr->calls.sbmbc_alloc.no)); + add_3tup(hpp, szp, &res, allctr->name.realloc, bld_unstable_uint(hpp, szp, allctr->calls.this_realloc.giga_no), bld_unstable_uint(hpp, szp, allctr->calls.this_realloc.no)); @@ -2191,7 +2404,9 @@ info_options(Allctr_t *allctr, #endif "option lmbcs: %beu\n" "option smbcs: %beu\n" - "option mbcgs: %beu\n", + "option mbcgs: %beu\n" + "option sbmbcs: %beu\n" + "option sbmbct: %beu\n", topt, allctr->ramv ? "true" : "false", #if HALFWORD_HEAP @@ -2211,7 +2426,9 @@ info_options(Allctr_t *allctr, #endif allctr->largest_mbc_size, allctr->smallest_mbc_size, - allctr->mbc_growth_stages); + allctr->mbc_growth_stages, + allctr->sbmbc_size, + allctr->sbmbc_threshold); } res = (*allctr->info_options)(allctr, "option ", print_to_p, print_to_arg, @@ -2219,6 +2436,12 @@ info_options(Allctr_t *allctr, if (hpp || szp) { add_2tup(hpp, szp, &res, + am.sbmbct, + bld_uint(hpp, szp, allctr->sbmbc_threshold)); + add_2tup(hpp, szp, &res, + am.sbmbcs, + bld_uint(hpp, szp, allctr->sbmbc_size)); + add_2tup(hpp, szp, &res, am.mbcgs, bld_uint(hpp, szp, allctr->mbc_growth_stages)); add_2tup(hpp, szp, &res, @@ -2285,10 +2508,10 @@ update_max_ever_values(CarriersStats_t *cs) static ERTS_INLINE void reset_max_values(CarriersStats_t *cs) { - cs->max.no = cs->curr_mseg.no + cs->curr_sys_alloc.no; - cs->max.size = cs->curr_mseg.size + cs->curr_sys_alloc.size; - cs->blocks.max.no = cs->blocks.curr.no; - cs->blocks.max.size = cs->blocks.curr.size; + cs->max.no = cs->curr.norm.mseg.no + cs->curr.norm.sys_alloc.no; + cs->max.size = cs->curr.norm.mseg.size + cs->curr.norm.sys_alloc.size; + cs->blocks.max.no = cs->blocks.curr.no; + cs->blocks.max.size = cs->blocks.curr.size; } @@ -2367,7 +2590,7 @@ erts_alcu_sz_info(Allctr_t *allctr, Uint **hpp, Uint *szp) { - Eterm res, mbcs, sbcs; + Eterm res, sbmbcs, mbcs, sbcs; res = THE_NON_VALUE; @@ -2389,24 +2612,29 @@ erts_alcu_sz_info(Allctr_t *allctr, /* Update sbc values not continously updated */ allctr->sbcs.blocks.curr.no - = allctr->sbcs.curr_mseg.no + allctr->sbcs.curr_sys_alloc.no; + = allctr->sbcs.curr.norm.mseg.no + allctr->sbcs.curr.norm.sys_alloc.no; allctr->sbcs.blocks.max.no = allctr->sbcs.max.no; + update_max_ever_values(&allctr->sbmbcs); update_max_ever_values(&allctr->mbcs); update_max_ever_values(&allctr->sbcs); - mbcs = sz_info_carriers(allctr, &allctr->mbcs, "mbcs ", print_to_p, - print_to_arg, hpp, szp); - sbcs = sz_info_carriers(allctr, &allctr->sbcs, "sbcs ", print_to_p, - print_to_arg, hpp, szp); + sbmbcs = sz_info_carriers(allctr, &allctr->sbmbcs, "sbmbcs ", print_to_p, + print_to_arg, hpp, szp); + mbcs = sz_info_carriers(allctr, &allctr->mbcs, "mbcs ", print_to_p, + print_to_arg, hpp, szp); + sbcs = sz_info_carriers(allctr, &allctr->sbcs, "sbcs ", print_to_p, + print_to_arg, hpp, szp); if (hpp || szp) { res = NIL; add_2tup(hpp, szp, &res, am.sbcs, sbcs); add_2tup(hpp, szp, &res, am.mbcs, mbcs); + add_2tup(hpp, szp, &res, am.sbmbcs, sbmbcs); } if (begin_max_period) { + reset_max_values(&allctr->sbmbcs); reset_max_values(&allctr->mbcs); reset_max_values(&allctr->sbcs); } @@ -2428,7 +2656,7 @@ erts_alcu_info(Allctr_t *allctr, Uint **hpp, Uint *szp) { - Eterm res, sett, mbcs, sbcs, calls; + Eterm res, sett, sbmbcs, mbcs, sbcs, calls; res = THE_NON_VALUE; @@ -2450,9 +2678,10 @@ erts_alcu_info(Allctr_t *allctr, /* Update sbc values not continously updated */ allctr->sbcs.blocks.curr.no - = allctr->sbcs.curr_mseg.no + allctr->sbcs.curr_sys_alloc.no; + = allctr->sbcs.curr.norm.mseg.no + allctr->sbcs.curr.norm.sys_alloc.no; allctr->sbcs.blocks.max.no = allctr->sbcs.max.no; + update_max_ever_values(&allctr->sbmbcs); update_max_ever_values(&allctr->mbcs); update_max_ever_values(&allctr->sbcs); @@ -2464,11 +2693,13 @@ erts_alcu_info(Allctr_t *allctr, ERTS_ALCU_VSN_STR); } - sett = info_options(allctr, print_to_p, print_to_arg, hpp, szp); - mbcs = info_carriers(allctr, &allctr->mbcs, "mbcs ", print_to_p, - print_to_arg, hpp, szp); - sbcs = info_carriers(allctr, &allctr->sbcs, "sbcs ", print_to_p, - print_to_arg, hpp, szp); + sett = info_options(allctr, print_to_p, print_to_arg, hpp, szp); + sbmbcs = info_carriers(allctr, &allctr->sbmbcs, "sbmbcs ", print_to_p, + print_to_arg, hpp, szp); + mbcs = info_carriers(allctr, &allctr->mbcs, "mbcs ", print_to_p, + print_to_arg, hpp, szp); + sbcs = info_carriers(allctr, &allctr->sbcs, "sbcs ", print_to_p, + print_to_arg, hpp, szp); calls = info_calls(allctr, print_to_p, print_to_arg, hpp, szp); if (hpp || szp) { @@ -2477,6 +2708,7 @@ erts_alcu_info(Allctr_t *allctr, add_2tup(hpp, szp, &res, am.calls, calls); add_2tup(hpp, szp, &res, am.sbcs, sbcs); add_2tup(hpp, szp, &res, am.mbcs, mbcs); + add_2tup(hpp, szp, &res, am.sbmbcs, sbmbcs); add_2tup(hpp, szp, &res, am.options, sett); add_3tup(hpp, szp, &res, am.versions, @@ -2485,6 +2717,7 @@ erts_alcu_info(Allctr_t *allctr, } if (begin_max_period) { + reset_max_values(&allctr->sbmbcs); reset_max_values(&allctr->mbcs); reset_max_values(&allctr->sbcs); } @@ -2508,12 +2741,14 @@ erts_alcu_current_size(Allctr_t *allctr, AllctrSize_t *size) erts_mtx_lock(&allctr->mutex); #endif - size->carriers = allctr->mbcs.curr_mseg.size; - size->carriers += allctr->mbcs.curr_sys_alloc.size; - size->carriers += allctr->sbcs.curr_mseg.size; - size->carriers += allctr->sbcs.curr_sys_alloc.size; + size->carriers = allctr->mbcs.curr.norm.mseg.size; + size->carriers += allctr->mbcs.curr.norm.sys_alloc.size; + size->carriers += allctr->sbmbcs.curr.small_block.size; + size->carriers += allctr->sbcs.curr.norm.mseg.size; + size->carriers += allctr->sbcs.curr.norm.sys_alloc.size; size->blocks = allctr->mbcs.blocks.curr.size; + size->blocks += allctr->sbmbcs.blocks.curr.size; size->blocks += allctr->sbcs.blocks.curr.size; #ifdef USE_THREADS @@ -2725,7 +2960,7 @@ do_erts_alcu_realloc(ErtsAlcType_t type, void *extra, void *p, Uint size, - UWord flgs) + Uint32 alcu_flgs) { Allctr_t *allctr = (Allctr_t *) extra; Block_t *blk; @@ -2758,9 +2993,32 @@ do_erts_alcu_realloc(ErtsAlcType_t type, blk = UMEM2BLK(p); + if (allctr->sbmbc_threshold > 0) { + Uint old_sz, new_sz, lim; + lim = allctr->sbmbc_threshold; + old_sz = BLK_SZ(blk); + new_sz = UMEMSZ2BLKSZ(allctr, size); + if ((old_sz < lim && lim <= new_sz) + || (new_sz < lim && lim <= old_sz)) { + /* *Need* to move it... */ + + INC_CC(allctr->calls.this_realloc); + res = do_erts_alcu_alloc(type, extra, size); + DEC_CC(allctr->calls.this_alloc); + + sys_memcpy(res, p, MIN(size, old_sz - ABLK_HDR_SZ)); + + do_erts_alcu_free(type, extra, p); + DEC_CC(allctr->calls.this_free); + return res; + } + if (old_sz < lim) + alcu_flgs |= ERTS_ALCU_FLG_SBMBC; + } + if (size < allctr->sbc_threshold) { if (IS_MBC_BLK(blk)) - res = mbc_realloc(allctr, p, size, flgs); + res = mbc_realloc(allctr, p, size, alcu_flgs); else { Uint used_sz = allctr->sbc_header_size + ABLK_HDR_SZ + size; Uint crr_sz; @@ -2791,7 +3049,7 @@ do_erts_alcu_realloc(ErtsAlcType_t type, if (100*diff_sz_val < allctr->sbc_move_threshold*crr_sz_val) /* Data won't be copied into a new carrier... */ goto do_carrier_resize; - else if (flgs & ERTS_ALCU_FLG_FAIL_REALLOC_MOVE) + else if (alcu_flgs & ERTS_ALCU_FLG_FAIL_REALLOC_MOVE) return NULL; res = mbc_alloc(allctr, size); @@ -2814,7 +3072,7 @@ do_erts_alcu_realloc(ErtsAlcType_t type, #endif res = new_blk ? BLK2UMEM(new_blk) : NULL; } - else if (flgs & ERTS_ALCU_FLG_FAIL_REALLOC_MOVE) + else if (alcu_flgs & ERTS_ALCU_FLG_FAIL_REALLOC_MOVE) return NULL; else { #if HALFWORD_HEAP @@ -3174,6 +3432,26 @@ erts_alcu_start(Allctr_t *allctr, AllctrInit_t *init) allctr->min_block_size = UNIT_CEILING(allctr->min_block_size + sizeof(UWord)); + + allctr->sbmbc_threshold = init->sbmbct; + + if (!erts_have_sbmbc_alloc + || ERTS_IS_SBMBC_ALLOCATOR_NO__(allctr->alloc_no)) + allctr->sbmbc_threshold = 0; + + if (!allctr->sbmbc_threshold) + allctr->sbmbc_size = 0; + else { + Uint min_size; + allctr->sbmbc_size = init->sbmbcs; + min_size = allctr->sbmbc_threshold; + min_size += allctr->min_block_size; + min_size += allctr->mbc_header_size; + if (allctr->sbmbc_size < min_size) + allctr->sbmbc_size = min_size; + } + + #if HAVE_ERTS_MSEG if (allctr->mseg_opt.abs_shrink_th > ~((UWord) 0) / 100) allctr->mseg_opt.abs_shrink_th = ~((UWord) 0) / 100; @@ -3185,12 +3463,16 @@ erts_alcu_start(Allctr_t *allctr, AllctrInit_t *init) #ifdef ERTS_ENABLE_LOCK_COUNT erts_mtx_init_x_opt(&allctr->mutex, - "alcu_allocator", - make_small(allctr->alloc_no), - ERTS_LCNT_LT_ALLOC); + ERTS_IS_SBMBC_ALLOCATOR_NO__(allctr->alloc_no) + ? "sbmbc_alloc" + : "alcu_allocator", + make_small(allctr->alloc_no), + ERTS_LCNT_LT_ALLOC); #else erts_mtx_init_x(&allctr->mutex, - "alcu_allocator", + ERTS_IS_SBMBC_ALLOCATOR_NO__(allctr->alloc_no) + ? "sbmbc_alloc" + : "alcu_allocator", make_small(allctr->alloc_no)); #endif /*ERTS_ENABLE_LOCK_COUNT*/ @@ -3260,7 +3542,7 @@ erts_alcu_start(Allctr_t *allctr, AllctrInit_t *init) if (!blk) goto error; - (*allctr->link_free_block)(allctr, blk); + (*allctr->link_free_block)(allctr, blk, 0); HARD_CHECK_BLK_CARRIER(allctr, blk); @@ -3290,6 +3572,8 @@ erts_alcu_stop(Allctr_t *allctr) destroy_carrier(allctr, SBC2BLK(allctr, allctr->sbc_list.first)); while (allctr->mbc_list.first) destroy_carrier(allctr, MBC2FBLK(allctr, allctr->mbc_list.first)); + while (allctr->sbmbc_list.first) + destroy_sbmbc(allctr, MBC2FBLK(allctr, allctr->sbmbc_list.first)); #ifdef USE_THREADS if (allctr->thread_safe) @@ -3387,13 +3671,15 @@ erts_alcu_verify_unused(Allctr_t *allctr) { UWord no; - no = allctr->sbcs.curr_mseg.no; - no += allctr->sbcs.curr_sys_alloc.no; + no = allctr->sbcs.curr.norm.mseg.no; + no += allctr->sbcs.curr.norm.sys_alloc.no; no += allctr->mbcs.blocks.curr.no; + no += allctr->sbmbcs.blocks.curr.no; if (no) { UWord sz = allctr->sbcs.blocks.curr.size; sz += allctr->mbcs.blocks.curr.size; + sz += allctr->sbmbcs.blocks.curr.size; erl_exit(ERTS_ABORT_EXIT, "%salloc() used when expected to be unused!\n" "Total amount of blocks allocated: %bpu\n" @@ -3492,7 +3778,7 @@ check_blk_carrier(Allctr_t *allctr, Block_t *iblk) (*allctr->check_block)(allctr, blk, (int) is_free_blk); if (IS_LAST_BLK(blk)) { - carrier_end = ((char *) NXT_BLK(blk)) + sizeof(UWord); + carrier_end = ((char *) NXT_BLK(blk)); mbc = *((Carrier_t **) NXT_BLK(blk)); prev_blk = NULL; blk = MBC2FBLK(allctr, mbc); @@ -3507,9 +3793,9 @@ check_blk_carrier(Allctr_t *allctr, Block_t *iblk) ASSERT(IS_MB_CARRIER(mbc)); ASSERT((((char *) mbc) + allctr->mbc_header_size - + tot_blk_sz - + sizeof(UWord)) == carrier_end); - ASSERT(((char *) mbc) + CARRIER_SZ(mbc) == carrier_end); + + tot_blk_sz) == carrier_end); + ASSERT(((char *) mbc) + CARRIER_SZ(mbc) - sizeof(Unit_t) <= carrier_end + && carrier_end <= ((char *) mbc) + CARRIER_SZ(mbc)); if (allctr->check_mbc) (*allctr->check_mbc)(allctr, mbc); @@ -3523,6 +3809,7 @@ check_blk_carrier(Allctr_t *allctr, Block_t *iblk) cl = &allctr->mbc_list; } +#if 0 /* FIXIT sbmbc */ if (cl->first == crr) { ASSERT(!crr->prev); } @@ -3537,6 +3824,7 @@ check_blk_carrier(Allctr_t *allctr, Block_t *iblk) ASSERT(crr->next); ASSERT(crr->next->prev == crr); } +#endif } #endif diff --git a/erts/emulator/beam/erl_alloc_util.h b/erts/emulator/beam/erl_alloc_util.h index ddf84c086c..fed4d3dbe6 100644 --- a/erts/emulator/beam/erl_alloc_util.h +++ b/erts/emulator/beam/erl_alloc_util.h @@ -34,6 +34,7 @@ typedef struct { typedef struct { char *name_prefix; ErtsAlcType_t alloc_no; + int force; int ts; int tspec; int tpref; @@ -50,6 +51,8 @@ typedef struct { UWord lmbcs; UWord smbcs; UWord mbcgs; + UWord sbmbct; + UWord sbmbcs; } AllctrInit_t; typedef struct { @@ -67,6 +70,7 @@ typedef struct { #define ERTS_DEFAULT_ALLCTR_INIT { \ NULL, \ ERTS_ALC_A_INVALID, /* (number) alloc_no: allocator number */\ + 0, /* (bool) force: force enabled */\ 1, /* (bool) ts: thread safe */\ 0, /* (bool) tspec: thread specific */\ 0, /* (bool) tpref: thread preferred */\ @@ -82,7 +86,9 @@ typedef struct { 10, /* (amount) mmmbc: max mseg mbcs */\ 10*1024*1024, /* (bytes) lmbcs: largest mbc size */\ 1024*1024, /* (bytes) smbcs: smallest mbc size */\ - 10 /* (amount) mbcgs: mbc growth stages */\ + 10, /* (amount) mbcgs: mbc growth stages */\ + 256, /* (bytes) sbmbct: small block mbc threshold */\ + 8*1024 /* (bytes) sbmbcs: small block mbc size */\ } #else /* if SMALL_MEMORY */ @@ -95,6 +101,7 @@ typedef struct { #define ERTS_DEFAULT_ALLCTR_INIT { \ NULL, \ ERTS_ALC_A_INVALID, /* (number) alloc_no: allocator number */\ + 0, /* (bool) force: force enabled */\ 1, /* (bool) ts: thread safe */\ 0, /* (bool) tspec: thread specific */\ 0, /* (bool) tpref: thread preferred */\ @@ -109,7 +116,9 @@ typedef struct { 10, /* (amount) mmmbc: max mseg mbcs */\ 1024*1024, /* (bytes) lmbcs: largest mbc size */\ 128*1024, /* (bytes) smbcs: smallest mbc size */\ - 10 /* (amount) mbcgs: mbc growth stages */\ + 10, /* (amount) mbcgs: mbc growth stages */\ + 256, /* (bytes) sbmbct: small block mbc threshold */\ + 8*1024 /* (bytes) sbmbcs: small block mbc size */\ } #endif @@ -144,6 +153,9 @@ void erts_alcu_current_size(Allctr_t *, AllctrSize_t *); #if defined(GET_ERL_ALLOC_UTIL_IMPL) && !defined(ERL_ALLOC_UTIL_IMPL__) #define ERL_ALLOC_UTIL_IMPL__ +#define ERTS_ALCU_FLG_FAIL_REALLOC_MOVE (((Uint32) 1) << 0) +#define ERTS_ALCU_FLG_SBMBC (((Uint32) 1) << 1) + #ifdef USE_THREADS #define ERL_THREADS_EMU_INTERNAL__ #include "erl_threads.h" @@ -188,6 +200,8 @@ void erts_alcu_current_size(Allctr_t *, AllctrSize_t *); #define CARRIER_SZ(C) \ ((C)->chdr & SZ_MASK) +extern int erts_have_sbmbc_alloc; + typedef union {char c[8]; long l; double d;} Unit_t; typedef struct Carrier_t_ Carrier_t; @@ -216,8 +230,13 @@ typedef struct { } StatValues_t; typedef struct { - StatValues_t curr_mseg; - StatValues_t curr_sys_alloc; + union { + struct { + StatValues_t mseg; + StatValues_t sys_alloc; + } norm; + StatValues_t small_block; + } curr; StatValues_t max; StatValues_t max_ever; struct { @@ -257,6 +276,8 @@ struct Allctr_t_ { Uint largest_mbc_size; Uint smallest_mbc_size; Uint mbc_growth_stages; + Uint sbmbc_threshold; + Uint sbmbc_size; #if HAVE_ERTS_MSEG ErtsMsegOpt_t mseg_opt; #endif @@ -269,6 +290,7 @@ struct Allctr_t_ { Uint min_block_size; /* Carriers */ + CarrierList_t sbmbc_list; CarrierList_t mbc_list; CarrierList_t sbc_list; @@ -277,15 +299,15 @@ struct Allctr_t_ { /* Callback functions (first 4 are mandatory) */ Block_t * (*get_free_block) (Allctr_t *, Uint, - Block_t *, Uint); - void (*link_free_block) (Allctr_t *, Block_t *); - void (*unlink_free_block) (Allctr_t *, Block_t *); + Block_t *, Uint, Uint32); + void (*link_free_block) (Allctr_t *, Block_t *, Uint32); + void (*unlink_free_block) (Allctr_t *, Block_t *, Uint32); Eterm (*info_options) (Allctr_t *, char *, int *, void *, Uint **, Uint *); Uint (*get_next_mbc_size) (Allctr_t *); - void (*creating_mbc) (Allctr_t *, Carrier_t *); - void (*destroying_mbc) (Allctr_t *, Carrier_t *); + void (*creating_mbc) (Allctr_t *, Carrier_t *, Uint32); + void (*destroying_mbc) (Allctr_t *, Carrier_t *, Uint32); void (*init_atoms) (void); #ifdef ERTS_ALLOC_UTIL_HARD_DEBUG @@ -312,6 +334,8 @@ struct Allctr_t_ { CallCounter_t this_alloc; CallCounter_t this_free; CallCounter_t this_realloc; + CallCounter_t sbmbc_alloc; + CallCounter_t sbmbc_free; CallCounter_t mseg_alloc; CallCounter_t mseg_dealloc; CallCounter_t mseg_realloc; @@ -322,6 +346,7 @@ struct Allctr_t_ { CarriersStats_t sbcs; CarriersStats_t mbcs; + CarriersStats_t sbmbcs; #ifdef DEBUG #ifdef USE_THREADS diff --git a/erts/emulator/beam/erl_ao_firstfit_alloc.c b/erts/emulator/beam/erl_ao_firstfit_alloc.c new file mode 100644 index 0000000000..90d8ea7300 --- /dev/null +++ b/erts/emulator/beam/erl_ao_firstfit_alloc.c @@ -0,0 +1,972 @@ +/* + * %CopyrightBegin% + * + * Copyright Ericsson AB 2003-2011. All Rights Reserved. + * + * The contents of this file are subject to the Erlang Public License, + * Version 1.1, (the "License"); you may not use this file except in + * compliance with the License. You should have received a copy of the + * Erlang Public License along with this software. If not, it can be + * retrieved online at http://www.erlang.org/. + * + * Software distributed under the License is distributed on an "AS IS" + * basis, WITHOUT WARRANTY OF ANY KIND, either express or implied. See + * the License for the specific language governing rights and limitations + * under the License. + * + * %CopyrightEnd% + */ + + +/* + * Description: An "address order first fit" allocator + * based on a Red-Black (binary search) Tree. The search, + * insert, and delete operations are all O(log n) operations + * on a Red-Black Tree. + * Red-Black Trees are described in "Introduction to Algorithms", + * by Thomas H. Cormen, Charles E. Leiserson, and Ronald L. Riverest. + * + * This module is a callback-module for erl_alloc_util.c + * + * Algorithm: The tree nodes are free-blocks ordered in address order. + * Every node also keeps the size of the largest block in its + * sub-tree ('max_size'). By that we can start from root and keep + * left (for low addresses) while dismissing entire sub-trees with + * too small blocks. + * + * Authors: Rickard Green/Sverker Eriksson + */ + + +#ifdef HAVE_CONFIG_H +# include "config.h" +#endif +#include "global.h" +#define GET_ERL_AOFF_ALLOC_IMPL +#include "erl_ao_firstfit_alloc.h" + +#ifdef DEBUG +#if 0 +#define HARD_DEBUG +#endif +#else +#undef HARD_DEBUG +#endif + +#define MIN_MBC_SZ (16*1024) +#define MIN_MBC_FIRST_FREE_SZ (4*1024) + +#define TREE_NODE_FLG (((Uint) 1) << 0) +#define RED_FLG (((Uint) 1) << 1) +#ifdef HARD_DEBUG +# define LEFT_VISITED_FLG (((Uint) 1) << 2) +# define RIGHT_VISITED_FLG (((Uint) 1) << 3) +#endif + +#define IS_RED(N) (((AOFF_RBTree_t *) (N)) \ + && ((AOFF_RBTree_t *) (N))->flags & RED_FLG) +#define IS_BLACK(N) (!IS_RED(((AOFF_RBTree_t *) (N)))) + +#define SET_RED(N) (((AOFF_RBTree_t *) (N))->flags |= RED_FLG) +#define SET_BLACK(N) (((AOFF_RBTree_t *) (N))->flags &= ~RED_FLG) + +#undef ASSERT +#define ASSERT ASSERT_EXPR + +#if 1 +#define RBT_ASSERT ASSERT +#else +#define RBT_ASSERT(x) +#endif + + +/* Types... */ +typedef struct AOFF_RBTree_t_ AOFF_RBTree_t; + +struct AOFF_RBTree_t_ { + Block_t hdr; + Uint flags; + AOFF_RBTree_t *parent; + AOFF_RBTree_t *left; + AOFF_RBTree_t *right; + Uint max_sz; /* of all blocks in this sub-tree */ +}; + +#ifdef HARD_DEBUG +static AOFF_RBTree_t * check_tree(AOFF_RBTree_t* root, Uint); +#endif + + +/* Calculate 'max_size' of tree node x by only looking at the direct children + * of x and x itself. + */ +static ERTS_INLINE Uint node_max_size(AOFF_RBTree_t *x) +{ + Uint sz = BLK_SZ(x); + if (x->left && x->left->max_sz > sz) { + sz = x->left->max_sz; + } + if (x->right && x->right->max_sz > sz) { + sz = x->right->max_sz; + } + return sz; +} + +/* Set new possibly lower 'max_size' of node and propagate change toward root +*/ +static ERTS_INLINE void lower_max_size(AOFF_RBTree_t *node, + AOFF_RBTree_t* stop_at) +{ + AOFF_RBTree_t* x = node; + Uint old_max = x->max_sz; + Uint new_max = node_max_size(x); + + if (new_max < old_max) { + x->max_sz = new_max; + while ((x=x->parent) != stop_at && x->max_sz == old_max) { + x->max_sz = node_max_size(x); + } + ASSERT(x == stop_at || x->max_sz > old_max); + } + else ASSERT(new_max == old_max); +} + + +/* Prototypes of callback functions */ +static Block_t* aoff_get_free_block(Allctr_t *, Uint, Block_t *, Uint, Uint32 flags); +static void aoff_link_free_block(Allctr_t *, Block_t*, Uint32 flags); +static void aoff_unlink_free_block(Allctr_t *allctr, Block_t *del, Uint32 flags); + +static Eterm info_options(Allctr_t *, char *, int *, void *, Uint **, Uint *); +static void init_atoms(void); + + + +#ifdef DEBUG + +/* Destroy all tree fields */ +#define DESTROY_TREE_NODE(N) \ + sys_memset((void *) (((Block_t *) (N)) + 1), \ + 0xff, \ + (sizeof(AOFF_RBTree_t) - sizeof(Block_t))) + +#else + +#define DESTROY_TREE_NODE(N) + +#endif + + +static int atoms_initialized = 0; + +void +erts_aoffalc_init(void) +{ + atoms_initialized = 0; +} + +Allctr_t * +erts_aoffalc_start(AOFFAllctr_t *alc, + AOFFAllctrInit_t* aoffinit, + AllctrInit_t *init) +{ + AOFFAllctr_t nulled_state = {{0}}; + /* {{0}} is used instead of {0}, in order to avoid (an incorrect) gcc + warning. gcc warns if {0} is used as initializer of a struct when + the first member is a struct (not if, for example, the third member + is a struct). */ + Allctr_t *allctr = (Allctr_t *) alc; + + sys_memcpy((void *) alc, (void *) &nulled_state, sizeof(AOFFAllctr_t)); + + allctr->mbc_header_size = sizeof(Carrier_t); + allctr->min_mbc_size = MIN_MBC_SZ; + allctr->min_mbc_first_free_size = MIN_MBC_FIRST_FREE_SZ; + allctr->min_block_size = sizeof(AOFF_RBTree_t); + + allctr->vsn_str = ERTS_ALC_AOFF_ALLOC_VSN_STR; + + + /* Callback functions */ + + allctr->get_free_block = aoff_get_free_block; + allctr->link_free_block = aoff_link_free_block; + allctr->unlink_free_block = aoff_unlink_free_block; + allctr->info_options = info_options; + + allctr->get_next_mbc_size = NULL; + allctr->creating_mbc = NULL; + allctr->destroying_mbc = NULL; + allctr->init_atoms = init_atoms; + +#ifdef ERTS_ALLOC_UTIL_HARD_DEBUG + allctr->check_block = NULL; + allctr->check_mbc = NULL; +#endif + + allctr->atoms_initialized = 0; + + if (!erts_alcu_start(allctr, init)) + return NULL; + + return allctr; +} + +/* + * Red-Black Tree operations needed + */ + +static ERTS_INLINE void +left_rotate(AOFF_RBTree_t **root, AOFF_RBTree_t *x) +{ + AOFF_RBTree_t *y = x->right; + x->right = y->left; + if (y->left) + y->left->parent = x; + y->parent = x->parent; + if (!y->parent) { + RBT_ASSERT(*root == x); + *root = y; + } + else if (x == x->parent->left) + x->parent->left = y; + else { + RBT_ASSERT(x == x->parent->right); + x->parent->right = y; + } + y->left = x; + x->parent = y; + + y->max_sz = x->max_sz; + x->max_sz = node_max_size(x); + ASSERT(y->max_sz >= x->max_sz); +} + +static ERTS_INLINE void +right_rotate(AOFF_RBTree_t **root, AOFF_RBTree_t *x) +{ + AOFF_RBTree_t *y = x->left; + x->left = y->right; + if (y->right) + y->right->parent = x; + y->parent = x->parent; + if (!y->parent) { + RBT_ASSERT(*root == x); + *root = y; + } + else if (x == x->parent->right) + x->parent->right = y; + else { + RBT_ASSERT(x == x->parent->left); + x->parent->left = y; + } + y->right = x; + x->parent = y; + y->max_sz = x->max_sz; + x->max_sz = node_max_size(x); + ASSERT(y->max_sz >= x->max_sz); +} + + +/* + * Replace node x with node y + * NOTE: block header of y is not changed + */ +static ERTS_INLINE void +replace(AOFF_RBTree_t **root, AOFF_RBTree_t *x, AOFF_RBTree_t *y) +{ + + if (!x->parent) { + RBT_ASSERT(*root == x); + *root = y; + } + else if (x == x->parent->left) + x->parent->left = y; + else { + RBT_ASSERT(x == x->parent->right); + x->parent->right = y; + } + if (x->left) { + RBT_ASSERT(x->left->parent == x); + x->left->parent = y; + } + if (x->right) { + RBT_ASSERT(x->right->parent == x); + x->right->parent = y; + } + + y->flags = x->flags; + y->parent = x->parent; + y->right = x->right; + y->left = x->left; + + y->max_sz = x->max_sz; + lower_max_size(y, NULL); + DESTROY_TREE_NODE(x); +} + +static void +tree_insert_fixup(AOFF_RBTree_t** root, AOFF_RBTree_t *blk) +{ + AOFF_RBTree_t *x = blk, *y; + + /* + * Rearrange the tree so that it satisfies the Red-Black Tree properties + */ + + RBT_ASSERT(x != *root && IS_RED(x->parent)); + do { + + /* + * x and its parent are both red. Move the red pair up the tree + * until we get to the root or until we can separate them. + */ + + RBT_ASSERT(IS_RED(x)); + RBT_ASSERT(IS_BLACK(x->parent->parent)); + RBT_ASSERT(x->parent->parent); + + if (x->parent == x->parent->parent->left) { + y = x->parent->parent->right; + if (IS_RED(y)) { + SET_BLACK(y); + x = x->parent; + SET_BLACK(x); + x = x->parent; + SET_RED(x); + } + else { + + if (x == x->parent->right) { + x = x->parent; + left_rotate(root, x); + } + + RBT_ASSERT(x == x->parent->parent->left->left); + RBT_ASSERT(IS_RED(x)); + RBT_ASSERT(IS_RED(x->parent)); + RBT_ASSERT(IS_BLACK(x->parent->parent)); + RBT_ASSERT(IS_BLACK(y)); + + SET_BLACK(x->parent); + SET_RED(x->parent->parent); + right_rotate(root, x->parent->parent); + + RBT_ASSERT(x == x->parent->left); + RBT_ASSERT(IS_RED(x)); + RBT_ASSERT(IS_RED(x->parent->right)); + RBT_ASSERT(IS_BLACK(x->parent)); + break; + } + } + else { + RBT_ASSERT(x->parent == x->parent->parent->right); + y = x->parent->parent->left; + if (IS_RED(y)) { + SET_BLACK(y); + x = x->parent; + SET_BLACK(x); + x = x->parent; + SET_RED(x); + } + else { + + if (x == x->parent->left) { + x = x->parent; + right_rotate(root, x); + } + + RBT_ASSERT(x == x->parent->parent->right->right); + RBT_ASSERT(IS_RED(x)); + RBT_ASSERT(IS_RED(x->parent)); + RBT_ASSERT(IS_BLACK(x->parent->parent)); + RBT_ASSERT(IS_BLACK(y)); + + SET_BLACK(x->parent); + SET_RED(x->parent->parent); + left_rotate(root, x->parent->parent); + + RBT_ASSERT(x == x->parent->right); + RBT_ASSERT(IS_RED(x)); + RBT_ASSERT(IS_RED(x->parent->left)); + RBT_ASSERT(IS_BLACK(x->parent)); + break; + } + } + } while (x != *root && IS_RED(x->parent)); + + SET_BLACK(*root); +} + +static void +aoff_unlink_free_block(Allctr_t *allctr, Block_t *del, Uint32 flags) +{ + AOFFAllctr_t *alc = (AOFFAllctr_t *) allctr; + AOFF_RBTree_t **root = ((flags & ERTS_ALCU_FLG_SBMBC) + ? &alc->sbmbc_root : &alc->mbc_root); + Uint spliced_is_black; + AOFF_RBTree_t *x, *y, *z = (AOFF_RBTree_t *) del; + AOFF_RBTree_t null_x; /* null_x is used to get the fixup started when we + splice out a node without children. */ + + null_x.parent = NULL; + +#ifdef HARD_DEBUG + check_tree(*root, 0); +#endif + + /* Remove node from tree... */ + + /* Find node to splice out */ + if (!z->left || !z->right) + y = z; + else + /* Set y to z:s successor */ + for(y = z->right; y->left; y = y->left); + /* splice out y */ + x = y->left ? y->left : y->right; + spliced_is_black = IS_BLACK(y); + if (x) { + x->parent = y->parent; + } + else if (spliced_is_black) { + x = &null_x; + x->flags = 0; + SET_BLACK(x); + x->right = x->left = NULL; + x->max_sz = 0; + x->parent = y->parent; + y->left = x; + } + + if (!y->parent) { + RBT_ASSERT(*root == y); + *root = x; + } + else { + if (y == y->parent->left) { + y->parent->left = x; + } + else { + RBT_ASSERT(y == y->parent->right); + y->parent->right = x; + } + if (y->parent != z) { + lower_max_size(y->parent, (y==z ? NULL : z)); + } + } + if (y != z) { + /* We spliced out the successor of z; replace z by the successor */ + replace(root, z, y); + } + + if (spliced_is_black) { + /* We removed a black node which makes the resulting tree + violate the Red-Black Tree properties. Fixup tree... */ + + while (IS_BLACK(x) && x->parent) { + + /* + * x has an "extra black" which we move up the tree + * until we reach the root or until we can get rid of it. + * + * y is the sibbling of x + */ + + if (x == x->parent->left) { + y = x->parent->right; + RBT_ASSERT(y); + if (IS_RED(y)) { + RBT_ASSERT(y->right); + RBT_ASSERT(y->left); + SET_BLACK(y); + RBT_ASSERT(IS_BLACK(x->parent)); + SET_RED(x->parent); + left_rotate(root, x->parent); + y = x->parent->right; + } + RBT_ASSERT(y); + RBT_ASSERT(IS_BLACK(y)); + if (IS_BLACK(y->left) && IS_BLACK(y->right)) { + SET_RED(y); + x = x->parent; + } + else { + if (IS_BLACK(y->right)) { + SET_BLACK(y->left); + SET_RED(y); + right_rotate(root, y); + y = x->parent->right; + } + RBT_ASSERT(y); + if (IS_RED(x->parent)) { + + SET_BLACK(x->parent); + SET_RED(y); + } + RBT_ASSERT(y->right); + SET_BLACK(y->right); + left_rotate(root, x->parent); + x = *root; + break; + } + } + else { + RBT_ASSERT(x == x->parent->right); + y = x->parent->left; + RBT_ASSERT(y); + if (IS_RED(y)) { + RBT_ASSERT(y->right); + RBT_ASSERT(y->left); + SET_BLACK(y); + RBT_ASSERT(IS_BLACK(x->parent)); + SET_RED(x->parent); + right_rotate(root, x->parent); + y = x->parent->left; + } + RBT_ASSERT(y); + RBT_ASSERT(IS_BLACK(y)); + if (IS_BLACK(y->right) && IS_BLACK(y->left)) { + SET_RED(y); + x = x->parent; + } + else { + if (IS_BLACK(y->left)) { + SET_BLACK(y->right); + SET_RED(y); + left_rotate(root, y); + y = x->parent->left; + } + RBT_ASSERT(y); + if (IS_RED(x->parent)) { + SET_BLACK(x->parent); + SET_RED(y); + } + RBT_ASSERT(y->left); + SET_BLACK(y->left); + right_rotate(root, x->parent); + x = *root; + break; + } + } + } + SET_BLACK(x); + + if (null_x.parent) { + if (null_x.parent->left == &null_x) + null_x.parent->left = NULL; + else { + RBT_ASSERT(null_x.parent->right == &null_x); + null_x.parent->right = NULL; + } + RBT_ASSERT(!null_x.left); + RBT_ASSERT(!null_x.right); + } + else if (*root == &null_x) { + *root = NULL; + RBT_ASSERT(!null_x.left); + RBT_ASSERT(!null_x.right); + } + } + + DESTROY_TREE_NODE(del); + +#ifdef HARD_DEBUG + check_tree(*root, 0); +#endif +} + +static void +aoff_link_free_block(Allctr_t *allctr, Block_t *block, Uint32 flags) +{ + AOFFAllctr_t *alc = (AOFFAllctr_t *) allctr; + AOFF_RBTree_t *blk = (AOFF_RBTree_t *) block; + AOFF_RBTree_t **root = ((flags & ERTS_ALCU_FLG_SBMBC) + ? &alc->sbmbc_root : &alc->mbc_root); + Uint blk_sz = BLK_SZ(blk); + +#ifdef HARD_DEBUG + check_tree(*root, 0); +#endif + + blk->flags = 0; + blk->left = NULL; + blk->right = NULL; + blk->max_sz = blk_sz; + + if (!*root) { + blk->parent = NULL; + SET_BLACK(blk); + *root = blk; + } + else { + AOFF_RBTree_t *x = *root; + while (1) { + if (x->max_sz < blk_sz) { + x->max_sz = blk_sz; + } + if (blk < x) { + if (!x->left) { + blk->parent = x; + x->left = blk; + break; + } + x = x->left; + } + else { + if (!x->right) { + blk->parent = x; + x->right = blk; + break; + } + x = x->right; + } + + } + + /* Insert block into size tree */ + RBT_ASSERT(blk->parent); + + SET_RED(blk); + if (IS_RED(blk->parent)) + tree_insert_fixup(root, blk); + } + +#ifdef HARD_DEBUG + check_tree(*root, 0); +#endif +} + +static Block_t * +aoff_get_free_block(Allctr_t *allctr, Uint size, + Block_t *cand_blk, Uint cand_size, Uint32 flags) +{ + AOFFAllctr_t *alc = (AOFFAllctr_t *) allctr; + AOFF_RBTree_t *x = ((flags & ERTS_ALCU_FLG_SBMBC) + ? alc->sbmbc_root : alc->mbc_root); + AOFF_RBTree_t *blk = NULL; +#ifdef HARD_DEBUG + AOFF_RBTree_t* dbg_blk = check_tree(x, size); +#endif + + ASSERT(!cand_blk || cand_size >= size); + + while (x) { + if (x->left && x->left->max_sz >= size) { + x = x->left; + } + else if (BLK_SZ(x) >= size) { + blk = x; + break; + } + else { + x = x->right; + } + } + +#ifdef HARD_DEBUG + ASSERT(blk == dbg_blk); +#endif + + if (!blk) + return NULL; + + if (cand_blk && cand_blk < &blk->hdr) { + return NULL; /* cand_blk was better */ + } + + aoff_unlink_free_block(allctr, (Block_t *) blk, flags); + + return (Block_t *) blk; +} + + +/* + * info_options() + */ + +static struct { + Eterm as; + Eterm aoff; +#ifdef DEBUG + Eterm end_of_atoms; +#endif +} am; + +static void ERTS_INLINE atom_init(Eterm *atom, char *name) +{ + *atom = am_atom_put(name, strlen(name)); +} +#define AM_INIT(AM) atom_init(&am.AM, #AM) + +static void +init_atoms(void) +{ +#ifdef DEBUG + Eterm *atom; +#endif + + if (atoms_initialized) + return; + +#ifdef DEBUG + for (atom = (Eterm *) &am; atom <= &am.end_of_atoms; atom++) { + *atom = THE_NON_VALUE; + } +#endif + AM_INIT(as); + AM_INIT(aoff); + +#ifdef DEBUG + for (atom = (Eterm *) &am; atom < &am.end_of_atoms; atom++) { + ASSERT(*atom != THE_NON_VALUE); + } +#endif + + atoms_initialized = 1; +} + + +#define bld_uint erts_bld_uint +#define bld_cons erts_bld_cons +#define bld_tuple erts_bld_tuple + +static ERTS_INLINE void +add_2tup(Uint **hpp, Uint *szp, Eterm *lp, Eterm el1, Eterm el2) +{ + *lp = bld_cons(hpp, szp, bld_tuple(hpp, szp, 2, el1, el2), *lp); +} + +static Eterm +info_options(Allctr_t *allctr, + char *prefix, + int *print_to_p, + void *print_to_arg, + Uint **hpp, + Uint *szp) +{ + Eterm res = THE_NON_VALUE; + + if (print_to_p) { + erts_print(*print_to_p, + print_to_arg, + "%sas: %s\n", + prefix, + "aoff"); + } + + if (hpp || szp) { + + if (!atoms_initialized) + erl_exit(1, "%s:%d: Internal error: Atoms not initialized", + __FILE__, __LINE__);; + + res = NIL; + add_2tup(hpp, szp, &res, am.as, am.aoff); + } + + return res; +} + + +/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *\ + * NOTE: erts_aoffalc_test() is only supposed to be used for testing. * + * * + * Keep alloc_SUITE_data/allocator_test.h updated if changes are made * + * to erts_aoffalc_test() * +\* */ + +unsigned long +erts_aoffalc_test(unsigned long op, unsigned long a1, unsigned long a2) +{ + switch (op) { + case 0x500: return (unsigned long) 0; /* IS_AOBF */ + case 0x501: return (unsigned long) ((AOFFAllctr_t *) a1)->mbc_root; + case 0x502: return (unsigned long) ((AOFF_RBTree_t *) a1)->parent; + case 0x503: return (unsigned long) ((AOFF_RBTree_t *) a1)->left; + case 0x504: return (unsigned long) ((AOFF_RBTree_t *) a1)->right; + case 0x506: return (unsigned long) IS_BLACK((AOFF_RBTree_t *) a1); + case 0x508: return (unsigned long) 1; /* IS_AOFF */ + case 0x509: return (unsigned long) ((AOFF_RBTree_t *) a1)->max_sz; + default: ASSERT(0); return ~((unsigned long) 0); + } +} + + +/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *\ + * Debug functions * +\* */ + + +#ifdef HARD_DEBUG + +#define IS_LEFT_VISITED(FB) ((FB)->flags & LEFT_VISITED_FLG) +#define IS_RIGHT_VISITED(FB) ((FB)->flags & RIGHT_VISITED_FLG) + +#define SET_LEFT_VISITED(FB) ((FB)->flags |= LEFT_VISITED_FLG) +#define SET_RIGHT_VISITED(FB) ((FB)->flags |= RIGHT_VISITED_FLG) + +#define UNSET_LEFT_VISITED(FB) ((FB)->flags &= ~LEFT_VISITED_FLG) +#define UNSET_RIGHT_VISITED(FB) ((FB)->flags &= ~RIGHT_VISITED_FLG) + + +#if 0 +# define PRINT_TREE +#else +# undef PRINT_TREE +#endif + +#ifdef PRINT_TREE +static void print_tree(AOFF_RBTree_t*); +#endif + +/* + * Checks that the order between parent and children are correct, + * and that the Red-Black Tree properies are satisfied. if size > 0, + * check_tree() returns the node that satisfies "address order first fit" + * + * The Red-Black Tree properies are: + * 1. Every node is either red or black. + * 2. Every leaf (NIL) is black. + * 3. If a node is red, then both its children are black. + * 4. Every simple path from a node to a descendant leaf + * contains the same number of black nodes. + * + * + own.max_size == MAX(own.size, left.max_size, right.max_size) + */ + +static AOFF_RBTree_t * +check_tree(AOFF_RBTree_t* root, Uint size) +{ + AOFF_RBTree_t *res = NULL; + Sint blacks; + Sint curr_blacks; + AOFF_RBTree_t *x; + +#ifdef PRINT_TREE + print_tree(root); +#endif + + if (!root) + return res; + + x = root; + ASSERT(IS_BLACK(x)); + ASSERT(!x->parent); + curr_blacks = 1; + blacks = -1; + + while (x) { + if (!IS_LEFT_VISITED(x)) { + SET_LEFT_VISITED(x); + if (x->left) { + x = x->left; + if (IS_BLACK(x)) + curr_blacks++; + continue; + } + else { + if (blacks < 0) + blacks = curr_blacks; + ASSERT(blacks == curr_blacks); + } + } + + if (!IS_RIGHT_VISITED(x)) { + SET_RIGHT_VISITED(x); + if (x->right) { + x = x->right; + if (IS_BLACK(x)) + curr_blacks++; + continue; + } + else { + if (blacks < 0) + blacks = curr_blacks; + ASSERT(blacks == curr_blacks); + } + } + + + if (IS_RED(x)) { + ASSERT(IS_BLACK(x->right)); + ASSERT(IS_BLACK(x->left)); + } + + ASSERT(x->parent || x == root); + + if (x->left) { + ASSERT(x->left->parent == x); + ASSERT(x->left < x); + ASSERT(x->left->max_sz <= x->max_sz); + } + + if (x->right) { + ASSERT(x->right->parent == x); + ASSERT(x->right > x); + ASSERT(x->right->max_sz <= x->max_sz); + } + ASSERT(x->max_sz >= BLK_SZ(x)); + ASSERT(x->max_sz == BLK_SZ(x) + || x->max_sz == (x->left ? x->left->max_sz : 0) + || x->max_sz == (x->right ? x->right->max_sz : 0)); + + if (size && BLK_SZ(x) >= size) { + if (!res || x < res) { + res = x; + } + } + + UNSET_LEFT_VISITED(x); + UNSET_RIGHT_VISITED(x); + if (IS_BLACK(x)) + curr_blacks--; + x = x->parent; + + } + + ASSERT(curr_blacks == 0); + + UNSET_LEFT_VISITED(root); + UNSET_RIGHT_VISITED(root); + + return res; + +} + + +#ifdef PRINT_TREE +#define INDENT_STEP 2 + +#include <stdio.h> + +static void +print_tree_aux(AOFF_RBTree_t *x, int indent) +{ + int i; + + if (x) { + print_tree_aux(x->right, indent + INDENT_STEP); + for (i = 0; i < indent; i++) { + putc(' ', stderr); + } + fprintf(stderr, "%s: sz=%lu addr=0x%lx max_size=%lu\r\n", + IS_BLACK(x) ? "BLACK" : "RED", + BLK_SZ(x), (Uint)x, x->max_sz); + print_tree_aux(x->left, indent + INDENT_STEP); + } +} + + +static void +print_tree(AOFF_RBTree_t* root) +{ + fprintf(stderr, " --- AOFF tree begin ---\r\n"); + print_tree_aux(root, 0); + fprintf(stderr, " --- AOFF tree end ---\r\n"); +} + +#endif /* PRINT_TREE */ + +#endif /* HARD_DEBUG */ + diff --git a/erts/emulator/beam/erl_ao_firstfit_alloc.h b/erts/emulator/beam/erl_ao_firstfit_alloc.h new file mode 100644 index 0000000000..6fa626f723 --- /dev/null +++ b/erts/emulator/beam/erl_ao_firstfit_alloc.h @@ -0,0 +1,60 @@ +/* + * %CopyrightBegin% + * + * Copyright Ericsson AB 2003-2011. All Rights Reserved. + * + * The contents of this file are subject to the Erlang Public License, + * Version 1.1, (the "License"); you may not use this file except in + * compliance with the License. You should have received a copy of the + * Erlang Public License along with this software. If not, it can be + * retrieved online at http://www.erlang.org/. + * + * Software distributed under the License is distributed on an "AS IS" + * basis, WITHOUT WARRANTY OF ANY KIND, either express or implied. See + * the License for the specific language governing rights and limitations + * under the License. + * + * %CopyrightEnd% + */ + + +#ifndef ERL_AO_FIRSTFIT_ALLOC__ +#define ERL_AO_FIRSTFIT_ALLOC__ + +#include "erl_alloc_util.h" + +#define ERTS_ALC_AOFF_ALLOC_VSN_STR "0.9" + +typedef struct AOFFAllctr_t_ AOFFAllctr_t; + +typedef struct { + int dummy; +} AOFFAllctrInit_t; + +#define ERTS_DEFAULT_AOFF_ALLCTR_INIT {0/*dummy*/} + +void erts_aoffalc_init(void); +Allctr_t *erts_aoffalc_start(AOFFAllctr_t *, AOFFAllctrInit_t*, AllctrInit_t *); + +#endif /* #ifndef ERL_AO_FIRSTFIT_ALLOC__ */ + + + +#if defined(GET_ERL_AOFF_ALLOC_IMPL) && !defined(ERL_AOFF_ALLOC_IMPL__) +#define ERL_AOFF_ALLOC_IMPL__ + +#define GET_ERL_ALLOC_UTIL_IMPL +#include "erl_alloc_util.h" + + +struct AOFFAllctr_t_ { + Allctr_t allctr; /* Has to be first! */ + + struct AOFF_RBTree_t_* mbc_root; + struct AOFF_RBTree_t_* sbmbc_root; +}; + +unsigned long erts_aoffalc_test(unsigned long, unsigned long, unsigned long); + +#endif /* #if defined(GET_ERL_AOFF_ALLOC_IMPL) + && !defined(ERL_AOFF_ALLOC_IMPL__) */ diff --git a/erts/emulator/beam/erl_bestfit_alloc.c b/erts/emulator/beam/erl_bestfit_alloc.c index 3035e5df16..f2199d41a1 100644 --- a/erts/emulator/beam/erl_bestfit_alloc.c +++ b/erts/emulator/beam/erl_bestfit_alloc.c @@ -1,7 +1,7 @@ /* * %CopyrightBegin% * - * Copyright Ericsson AB 2003-2009. All Rights Reserved. + * Copyright Ericsson AB 2003-2011. All Rights Reserved. * * The contents of this file are subject to the Erlang Public License, * Version 1.1, (the "License"); you may not use this file except in @@ -84,24 +84,24 @@ #ifdef HARD_DEBUG -static RBTree_t * check_tree(BFAllctr_t *, Uint); +static RBTree_t * check_tree(RBTree_t, int, Uint); #endif -static void tree_delete(Allctr_t *allctr, Block_t *del); +static void tree_delete(Allctr_t *allctr, Block_t *del, Uint32 flags); /* Prototypes of callback functions */ /* "address order best fit" specific callback functions */ static Block_t * aobf_get_free_block (Allctr_t *, Uint, - Block_t *, Uint); -static void aobf_link_free_block (Allctr_t *, Block_t *); + Block_t *, Uint, Uint32); +static void aobf_link_free_block (Allctr_t *, Block_t *, Uint32); #define aobf_unlink_free_block tree_delete /* "best fit" specific callback functions */ static Block_t * bf_get_free_block (Allctr_t *, Uint, - Block_t *, Uint); -static void bf_link_free_block (Allctr_t *, Block_t *); -static ERTS_INLINE void bf_unlink_free_block (Allctr_t *, Block_t *); + Block_t *, Uint, Uint32); +static void bf_link_free_block (Allctr_t *, Block_t *, Uint32); +static ERTS_INLINE void bf_unlink_free_block (Allctr_t *, Block_t *, Uint32); static Eterm info_options (Allctr_t *, char *, int *, @@ -303,7 +303,7 @@ replace(RBTree_t **root, RBTree_t *x, RBTree_t *y) } static void -tree_insert_fixup(BFAllctr_t *bfallctr, RBTree_t *blk) +tree_insert_fixup(RBTree_t **root, RBTree_t *blk) { RBTree_t *x = blk, *y; @@ -311,7 +311,7 @@ tree_insert_fixup(BFAllctr_t *bfallctr, RBTree_t *blk) * Rearrange the tree so that it satisfies the Red-Black Tree properties */ - RBT_ASSERT(x != bfallctr->root && IS_RED(x->parent)); + RBT_ASSERT(x != *root && IS_RED(x->parent)); do { /* @@ -336,7 +336,7 @@ tree_insert_fixup(BFAllctr_t *bfallctr, RBTree_t *blk) if (x == x->parent->right) { x = x->parent; - left_rotate(&bfallctr->root, x); + left_rotate(root, x); } RBT_ASSERT(x == x->parent->parent->left->left); @@ -347,7 +347,7 @@ tree_insert_fixup(BFAllctr_t *bfallctr, RBTree_t *blk) SET_BLACK(x->parent); SET_RED(x->parent->parent); - right_rotate(&bfallctr->root, x->parent->parent); + right_rotate(root, x->parent->parent); RBT_ASSERT(x == x->parent->left); RBT_ASSERT(IS_RED(x)); @@ -370,7 +370,7 @@ tree_insert_fixup(BFAllctr_t *bfallctr, RBTree_t *blk) if (x == x->parent->left) { x = x->parent; - right_rotate(&bfallctr->root, x); + right_rotate(root, x); } RBT_ASSERT(x == x->parent->parent->right->right); @@ -381,7 +381,7 @@ tree_insert_fixup(BFAllctr_t *bfallctr, RBTree_t *blk) SET_BLACK(x->parent); SET_RED(x->parent->parent); - left_rotate(&bfallctr->root, x->parent->parent); + left_rotate(root, x->parent->parent); RBT_ASSERT(x == x->parent->right); RBT_ASSERT(IS_RED(x)); @@ -390,9 +390,9 @@ tree_insert_fixup(BFAllctr_t *bfallctr, RBTree_t *blk) break; } } - } while (x != bfallctr->root && IS_RED(x->parent)); + } while (x != *root && IS_RED(x->parent)); - SET_BLACK(bfallctr->root); + SET_BLACK(*root); } @@ -402,18 +402,22 @@ tree_insert_fixup(BFAllctr_t *bfallctr, RBTree_t *blk) * callback function in the address order case. */ static void -tree_delete(Allctr_t *allctr, Block_t *del) +tree_delete(Allctr_t *allctr, Block_t *del, Uint32 flags) { BFAllctr_t *bfallctr = (BFAllctr_t *) allctr; Uint spliced_is_black; + RBTree_t **root = ((flags & ERTS_ALCU_FLG_SBMBC) + ? &bfallctr->sbmbc_root + : &bfallctr->mbc_root); RBTree_t *x, *y, *z = (RBTree_t *) del; RBTree_t null_x; /* null_x is used to get the fixup started when we splice out a node without children. */ null_x.parent = NULL; + #ifdef HARD_DEBUG - check_tree(bfallctr, 0); + check_tree(*root, bfallctr->address_order, 0); #endif /* Remove node from tree... */ @@ -440,8 +444,8 @@ tree_delete(Allctr_t *allctr, Block_t *del) } if (!y->parent) { - RBT_ASSERT(bfallctr->root == y); - bfallctr->root = x; + RBT_ASSERT(*root == y); + *root = x; } else if (y == y->parent->left) y->parent->left = x; @@ -451,7 +455,7 @@ tree_delete(Allctr_t *allctr, Block_t *del) } if (y != z) { /* We spliced out the successor of z; replace z by the successor */ - replace(&bfallctr->root, z, y); + replace(root, z, y); } if (spliced_is_black) { @@ -476,7 +480,7 @@ tree_delete(Allctr_t *allctr, Block_t *del) SET_BLACK(y); RBT_ASSERT(IS_BLACK(x->parent)); SET_RED(x->parent); - left_rotate(&bfallctr->root, x->parent); + left_rotate(root, x->parent); y = x->parent->right; } RBT_ASSERT(y); @@ -489,7 +493,7 @@ tree_delete(Allctr_t *allctr, Block_t *del) if (IS_BLACK(y->right)) { SET_BLACK(y->left); SET_RED(y); - right_rotate(&bfallctr->root, y); + right_rotate(root, y); y = x->parent->right; } RBT_ASSERT(y); @@ -500,8 +504,8 @@ tree_delete(Allctr_t *allctr, Block_t *del) } RBT_ASSERT(y->right); SET_BLACK(y->right); - left_rotate(&bfallctr->root, x->parent); - x = bfallctr->root; + left_rotate(root, x->parent); + x = *root; break; } } @@ -515,7 +519,7 @@ tree_delete(Allctr_t *allctr, Block_t *del) SET_BLACK(y); RBT_ASSERT(IS_BLACK(x->parent)); SET_RED(x->parent); - right_rotate(&bfallctr->root, x->parent); + right_rotate(root, x->parent); y = x->parent->left; } RBT_ASSERT(y); @@ -528,7 +532,7 @@ tree_delete(Allctr_t *allctr, Block_t *del) if (IS_BLACK(y->left)) { SET_BLACK(y->right); SET_RED(y); - left_rotate(&bfallctr->root, y); + left_rotate(root, y); y = x->parent->left; } RBT_ASSERT(y); @@ -538,8 +542,8 @@ tree_delete(Allctr_t *allctr, Block_t *del) } RBT_ASSERT(y->left); SET_BLACK(y->left); - right_rotate(&bfallctr->root, x->parent); - x = bfallctr->root; + right_rotate(root, x->parent); + x = *root; break; } } @@ -556,8 +560,8 @@ tree_delete(Allctr_t *allctr, Block_t *del) RBT_ASSERT(!null_x.left); RBT_ASSERT(!null_x.right); } - else if (bfallctr->root == &null_x) { - bfallctr->root = NULL; + else if (*root == &null_x) { + *root = NULL; RBT_ASSERT(!null_x.left); RBT_ASSERT(!null_x.right); } @@ -567,7 +571,7 @@ tree_delete(Allctr_t *allctr, Block_t *del) DESTROY_TREE_NODE(del); #ifdef HARD_DEBUG - check_tree(bfallctr, 0); + check_tree(root, bfallctr->address_order, 0); #endif } @@ -577,23 +581,28 @@ tree_delete(Allctr_t *allctr, Block_t *del) \* */ static void -aobf_link_free_block(Allctr_t *allctr, Block_t *block) +aobf_link_free_block(Allctr_t *allctr, Block_t *block, Uint32 flags) { BFAllctr_t *bfallctr = (BFAllctr_t *) allctr; + RBTree_t **root = ((flags & ERTS_ALCU_FLG_SBMBC) + ? &bfallctr->sbmbc_root + : &bfallctr->mbc_root); RBTree_t *blk = (RBTree_t *) block; Uint blk_sz = BLK_SZ(blk); + + blk->flags = 0; blk->left = NULL; blk->right = NULL; - if (!bfallctr->root) { + if (!*root) { blk->parent = NULL; SET_BLACK(blk); - bfallctr->root = blk; + *root = blk; } else { - RBTree_t *x = bfallctr->root; + RBTree_t *x = *root; while (1) { Uint size; @@ -623,28 +632,32 @@ aobf_link_free_block(Allctr_t *allctr, Block_t *block) SET_RED(blk); if (IS_RED(blk->parent)) - tree_insert_fixup(bfallctr, blk); + tree_insert_fixup(root, blk); } #ifdef HARD_DEBUG - check_tree(bfallctr, 0); + check_tree(root, 1, 0); #endif } #if 0 /* tree_delete() is directly used instead */ static void -aobf_unlink_free_block(Allctr_t *allctr, Block_t *block) +aobf_unlink_free_block(Allctr_t *allctr, Block_t *block, Uint32 flags) { - tree_delete(allctr, block); + tree_delete(allctr, block, flags); } #endif static Block_t * aobf_get_free_block(Allctr_t *allctr, Uint size, - Block_t *cand_blk, Uint cand_size) + Block_t *cand_blk, Uint cand_size, + Uint32 flags) { BFAllctr_t *bfallctr = (BFAllctr_t *) allctr; - RBTree_t *x = bfallctr->root; + RBTree_t **root = ((flags & ERTS_ALCU_FLG_SBMBC) + ? &bfallctr->sbmbc_root + : &bfallctr->mbc_root); + RBTree_t *x = *root; RBTree_t *blk = NULL; Uint blk_sz; @@ -665,7 +678,7 @@ aobf_get_free_block(Allctr_t *allctr, Uint size, return NULL; #ifdef HARD_DEBUG - ASSERT(blk == check_tree(bfallctr, size)); + ASSERT(blk == check_tree(root, 1, size)); #endif if (cand_blk) { @@ -676,7 +689,7 @@ aobf_get_free_block(Allctr_t *allctr, Uint size, return NULL; /* cand_blk was better */ } - aobf_unlink_free_block(allctr, (Block_t *) blk); + aobf_unlink_free_block(allctr, (Block_t *) blk, flags); return (Block_t *) blk; } @@ -687,9 +700,12 @@ aobf_get_free_block(Allctr_t *allctr, Uint size, \* */ static void -bf_link_free_block(Allctr_t *allctr, Block_t *block) +bf_link_free_block(Allctr_t *allctr, Block_t *block, Uint32 flags) { BFAllctr_t *bfallctr = (BFAllctr_t *) allctr; + RBTree_t **root = ((flags & ERTS_ALCU_FLG_SBMBC) + ? &bfallctr->sbmbc_root + : &bfallctr->mbc_root); RBTree_t *blk = (RBTree_t *) block; Uint blk_sz = BLK_SZ(blk); @@ -700,13 +716,13 @@ bf_link_free_block(Allctr_t *allctr, Block_t *block) blk->left = NULL; blk->right = NULL; - if (!bfallctr->root) { + if (!*root) { blk->parent = NULL; SET_BLACK(blk); - bfallctr->root = blk; + *root = blk; } else { - RBTree_t *x = bfallctr->root; + RBTree_t *x = *root; while (1) { Uint size; @@ -745,7 +761,7 @@ bf_link_free_block(Allctr_t *allctr, Block_t *block) SET_RED(blk); if (IS_RED(blk->parent)) - tree_insert_fixup(bfallctr, blk); + tree_insert_fixup(root, blk); } @@ -753,14 +769,17 @@ bf_link_free_block(Allctr_t *allctr, Block_t *block) LIST_NEXT(blk) = NULL; #ifdef HARD_DEBUG - check_tree(bfallctr, 0); + check_tree(root, 0, 0); #endif } static ERTS_INLINE void -bf_unlink_free_block(Allctr_t *allctr, Block_t *block) +bf_unlink_free_block(Allctr_t *allctr, Block_t *block, Uint32 flags) { BFAllctr_t *bfallctr = (BFAllctr_t *) allctr; + RBTree_t **root = ((flags & ERTS_ALCU_FLG_SBMBC) + ? &bfallctr->sbmbc_root + : &bfallctr->mbc_root); RBTree_t *x = (RBTree_t *) block; if (IS_LIST_ELEM(x)) { @@ -778,9 +797,9 @@ bf_unlink_free_block(Allctr_t *allctr, Block_t *block) ASSERT(IS_LIST_ELEM(LIST_NEXT(x))); #ifdef HARD_DEBUG - check_tree(bfallctr, 0); + check_tree(root, 0, 0); #endif - replace(&bfallctr->root, x, LIST_NEXT(x)); + replace(root, x, LIST_NEXT(x)); #ifdef HARD_DEBUG check_tree(bfallctr, 0); @@ -788,7 +807,7 @@ bf_unlink_free_block(Allctr_t *allctr, Block_t *block) } else { /* Remove from tree */ - tree_delete(allctr, block); + tree_delete(allctr, block, flags); } DESTROY_LIST_ELEM(x); @@ -797,10 +816,14 @@ bf_unlink_free_block(Allctr_t *allctr, Block_t *block) static Block_t * bf_get_free_block(Allctr_t *allctr, Uint size, - Block_t *cand_blk, Uint cand_size) + Block_t *cand_blk, Uint cand_size, + Uint32 flags) { BFAllctr_t *bfallctr = (BFAllctr_t *) allctr; - RBTree_t *x = bfallctr->root; + RBTree_t **root = ((flags & ERTS_ALCU_FLG_SBMBC) + ? &bfallctr->sbmbc_root + : &bfallctr->mbc_root); + RBTree_t *x = *root; RBTree_t *blk = NULL; Uint blk_sz; @@ -827,7 +850,7 @@ bf_get_free_block(Allctr_t *allctr, Uint size, #ifdef HARD_DEBUG { - RBTree_t *ct_blk = check_tree(bfallctr, size); + RBTree_t *ct_blk = check_tree(root, 0, size); ASSERT(BLK_SZ(ct_blk) == BLK_SZ(blk)); } #endif @@ -839,7 +862,7 @@ bf_get_free_block(Allctr_t *allctr, Uint size, the tree node */ blk = LIST_NEXT(blk) ? LIST_NEXT(blk) : blk; - bf_unlink_free_block(allctr, (Block_t *) blk); + bf_unlink_free_block(allctr, (Block_t *) blk, flags); return (Block_t *) blk; } @@ -949,13 +972,14 @@ erts_bfalc_test(unsigned long op, unsigned long a1, unsigned long a2) { switch (op) { case 0x200: return (unsigned long) ((BFAllctr_t *) a1)->address_order; - case 0x201: return (unsigned long) ((BFAllctr_t *) a1)->root; + case 0x201: return (unsigned long) ((BFAllctr_t *) a1)->mbc_root; case 0x202: return (unsigned long) ((RBTree_t *) a1)->parent; case 0x203: return (unsigned long) ((RBTree_t *) a1)->left; case 0x204: return (unsigned long) ((RBTree_t *) a1)->right; case 0x205: return (unsigned long) ((RBTreeList_t *) a1)->next; case 0x206: return (unsigned long) IS_BLACK((RBTree_t *) a1); case 0x207: return (unsigned long) IS_TREE_NODE((RBTree_t *) a1); + case 0x208: return (unsigned long) 0; /* IS_AOFF */ default: ASSERT(0); return ~((unsigned long) 0); } } @@ -985,7 +1009,7 @@ erts_bfalc_test(unsigned long op, unsigned long a1, unsigned long a2) #endif #ifdef PRINT_TREE -static void print_tree(BFAllctr_t *); +static void print_tree(RBTree_t *, int); #endif /* @@ -1003,7 +1027,7 @@ static void print_tree(BFAllctr_t *); */ static RBTree_t * -check_tree(BFAllctr_t *bfallctr, Uint size) +check_tree(RBTree_t *root, int ao, Uint size) { RBTree_t *res = NULL; Sint blacks; @@ -1011,13 +1035,13 @@ check_tree(BFAllctr_t *bfallctr, Uint size) RBTree_t *x; #ifdef PRINT_TREE - print_tree(bfallctr); + print_tree(root, ao); #endif - if (!bfallctr->root) + if (!root) return res; - x = bfallctr->root; + x = root; ASSERT(IS_BLACK(x)); ASSERT(!x->parent); curr_blacks = 1; @@ -1060,11 +1084,11 @@ check_tree(BFAllctr_t *bfallctr, Uint size) ASSERT(IS_BLACK(x->left)); } - ASSERT(x->parent || x == bfallctr->root); + ASSERT(x->parent || x == root); if (x->left) { ASSERT(x->left->parent == x); - if (bfallctr->address_order) { + if (ao) { ASSERT(BLK_SZ(x->left) < BLK_SZ(x) || (BLK_SZ(x->left) == BLK_SZ(x) && x->left < x)); } @@ -1076,7 +1100,7 @@ check_tree(BFAllctr_t *bfallctr, Uint size) if (x->right) { ASSERT(x->right->parent == x); - if (bfallctr->address_order) { + if (ao) { ASSERT(BLK_SZ(x->right) > BLK_SZ(x) || (BLK_SZ(x->right) == BLK_SZ(x) && x->right > x)); } @@ -1087,7 +1111,7 @@ check_tree(BFAllctr_t *bfallctr, Uint size) } if (size && BLK_SZ(x) >= size) { - if (bfallctr->address_order) { + if (ao) { if (!res || BLK_SZ(x) < BLK_SZ(res) || (BLK_SZ(x) == BLK_SZ(res) && x < res)) @@ -1109,8 +1133,8 @@ check_tree(BFAllctr_t *bfallctr, Uint size) ASSERT(curr_blacks == 0); - UNSET_LEFT_VISITED(bfallctr->root); - UNSET_RIGHT_VISITED(bfallctr->root); + UNSET_LEFT_VISITED(root); + UNSET_RIGHT_VISITED(root); return res; @@ -1148,11 +1172,11 @@ print_tree_aux(RBTree_t *x, int indent) static void -print_tree(BFAllctr_t *bfallctr) +print_tree(RBTree_t *root, int ao) { - char *type = bfallctr->address_order ? "Size-Adress" : "Size"; + char *type = ao ? "Size-Adress" : "Size"; fprintf(stderr, " --- %s tree begin ---\r\n", type); - print_tree_aux(bfallctr->root, 0); + print_tree_aux(root, 0); fprintf(stderr, " --- %s tree end ---\r\n", type); } diff --git a/erts/emulator/beam/erl_bestfit_alloc.h b/erts/emulator/beam/erl_bestfit_alloc.h index cb35e21e57..0c29662852 100644 --- a/erts/emulator/beam/erl_bestfit_alloc.h +++ b/erts/emulator/beam/erl_bestfit_alloc.h @@ -1,7 +1,7 @@ /* * %CopyrightBegin% * - * Copyright Ericsson AB 2003-2009. All Rights Reserved. + * Copyright Ericsson AB 2003-2011. All Rights Reserved. * * The contents of this file are subject to the Erlang Public License, * Version 1.1, (the "License"); you may not use this file except in @@ -54,7 +54,8 @@ typedef struct RBTree_t_ RBTree_t; struct BFAllctr_t_ { Allctr_t allctr; /* Has to be first! */ - RBTree_t * root; + RBTree_t * mbc_root; + RBTree_t * sbmbc_root; int address_order; }; diff --git a/erts/emulator/beam/erl_bif_ddll.c b/erts/emulator/beam/erl_bif_ddll.c index 9631fb50db..d714eacd06 100644 --- a/erts/emulator/beam/erl_bif_ddll.c +++ b/erts/emulator/beam/erl_bif_ddll.c @@ -369,7 +369,7 @@ BIF_RETTYPE erl_ddll_try_load_3(Process *p, Eterm path_term, if (!(prt->status & FREE_PORT_FLAGS) && prt->drv_ptr->handle == dh) { #if DDLL_SMP - erts_smp_atomic_inc(&prt->refc); + erts_smp_atomic_inc_nob(&prt->refc); /* Extremely rare spinlock */ while(prt->status & ERTS_PORT_SFLG_INITIALIZING) { erts_smp_port_state_unlock(prt); @@ -597,7 +597,7 @@ done: if (!(prt->status & FREE_PORT_FLAGS) && prt->drv_ptr->handle == dh) { #if DDLL_SMP - erts_smp_atomic_inc(&prt->refc); + erts_smp_atomic_inc_nob(&prt->refc); /* Extremely rare spinlock */ while(prt->status & ERTS_PORT_SFLG_INITIALIZING) { erts_smp_port_state_unlock(prt); @@ -1054,7 +1054,7 @@ void erts_ddll_proc_dead(Process *p, ErtsProcLocks plocks) if (!(prt->status & FREE_PORT_FLAGS) && prt->drv_ptr->handle == dh) { #if DDLL_SMP - erts_smp_atomic_inc(&prt->refc); + erts_smp_atomic_inc_nob(&prt->refc); while(prt->status & ERTS_PORT_SFLG_INITIALIZING) { erts_smp_port_state_unlock(prt); erts_smp_port_state_lock(prt); @@ -1602,7 +1602,7 @@ static int do_load_driver_entry(DE_Handle *dh, char *path, char *name) erts_sys_ddll_close(dh->handle); return ERL_DE_LOAD_ERROR_BAD_NAME; } - erts_smp_atomic_init(&(dh->refc), (erts_aint_t) 0); + erts_smp_atomic_init_nob(&(dh->refc), (erts_aint_t) 0); dh->port_count = 0; dh->full_path = erts_alloc(ERTS_ALC_T_DDLL_HANDLE, sys_strlen(path) + 1); sys_strcpy(dh->full_path, path); diff --git a/erts/emulator/beam/erl_bif_info.c b/erts/emulator/beam/erl_bif_info.c index f264bf44df..e17325b64f 100644 --- a/erts/emulator/beam/erl_bif_info.c +++ b/erts/emulator/beam/erl_bif_info.c @@ -120,6 +120,10 @@ static char erts_system_version[] = ("Erlang " ERLANG_OTP_RELEASE #endif static Eterm +current_function(Process* p, Process* rp, Eterm** hpp, int full_info); +static Eterm current_stacktrace(Process* p, Process* rp, Eterm** hpp); + +static Eterm bld_bin_list(Uint **hpp, Uint *szp, ErlOffHeap* oh) { struct erl_off_heap_header* ohh; @@ -135,7 +139,7 @@ bld_bin_list(Uint **hpp, Uint *szp, ErlOffHeap* oh) if (szp) *szp += 4+2; if (hpp) { - Uint refc = (Uint) erts_smp_atomic_read(&pb->val->refc); + Uint refc = (Uint) erts_smp_atomic_read_nob(&pb->val->refc); tuple = TUPLE3(*hpp, val, orig_size, make_small(refc)); res = CONS(*hpp + 4, tuple, res); *hpp += 4+2; @@ -554,6 +558,8 @@ static Eterm pi_args[] = { am_suspending, am_min_heap_size, am_min_bin_vheap_size, + am_current_location, + am_current_stacktrace, #ifdef HYBRID am_message_binary #endif @@ -602,8 +608,10 @@ pi_arg2ix(Eterm arg) case am_suspending: return 26; case am_min_heap_size: return 27; case am_min_bin_vheap_size: return 28; + case am_current_location: return 29; + case am_current_stacktrace: return 30; #ifdef HYBRID - case am_message_binary: return 29; + case am_message_binary: return 31; #endif default: return -1; } @@ -1006,35 +1014,15 @@ process_info_aux(Process *BIF_P, break; case am_current_function: - if (rp->current == NULL) { - rp->current = find_function_from_pc(rp->i); - } - if (rp->current == NULL) { - hp = HAlloc(BIF_P, 3); - res = am_undefined; - } else { - BeamInstr* current; - - if (rp->current[0] == am_erlang && - rp->current[1] == am_process_info && - (rp->current[2] == 1 || rp->current[2] == 2) && - (current = find_function_from_pc(rp->cp)) != NULL) { - - /* - * The current function is erlang:process_info/2, - * which is not the answer that the application want. - * We will use the function pointed into by rp->cp - * instead. - */ + res = current_function(BIF_P, rp, &hp, 0); + break; - rp->current = current; - } + case am_current_location: + res = current_function(BIF_P, rp, &hp, 1); + break; - hp = HAlloc(BIF_P, 3+4); - res = TUPLE3(hp, rp->current[0], - rp->current[1], make_small(rp->current[2])); - hp += 4; - } + case am_current_stacktrace: + res = current_stacktrace(BIF_P, rp, &hp); break; case am_initial_call: @@ -1608,6 +1596,113 @@ process_info_aux(Process *BIF_P, } #undef MI_INC +static Eterm +current_function(Process* BIF_P, Process* rp, Eterm** hpp, int full_info) +{ + Eterm* hp; + Eterm res; + FunctionInfo fi; + + if (rp->current == NULL) { + erts_lookup_function_info(&fi, rp->i, full_info); + rp->current = fi.current; + } else if (full_info) { + erts_lookup_function_info(&fi, rp->i, full_info); + if (fi.current == NULL) { + /* Use the current function without location info */ + erts_set_current_function(&fi, rp->current); + } + } + + if (BIF_P->id == rp->id) { + FunctionInfo fi2; + + /* + * The current function is erlang:process_info/{1,2}, + * which is not the answer that the application want. + * We will use the function pointed into by rp->cp + * instead if it can be looked up. + */ + erts_lookup_function_info(&fi2, rp->cp, full_info); + if (fi2.current) { + fi = fi2; + rp->current = fi2.current; + } + } + + /* + * Return the result. + */ + if (rp->current == NULL) { + hp = HAlloc(BIF_P, 3); + res = am_undefined; + } else if (full_info) { + hp = HAlloc(BIF_P, 3+fi.needed); + hp = erts_build_mfa_item(&fi, hp, am_true, &res); + } else { + hp = HAlloc(BIF_P, 3+4); + res = TUPLE3(hp, rp->current[0], + rp->current[1], make_small(rp->current[2])); + hp += 4; + } + *hpp = hp; + return res; +} + +static Eterm +current_stacktrace(Process* p, Process* rp, Eterm** hpp) +{ + Uint sz; + struct StackTrace* s; + int depth; + FunctionInfo* stk; + FunctionInfo* stkp; + Uint heap_size; + int i; + Eterm* hp = *hpp; + Eterm mfa; + Eterm res = NIL; + + depth = 8; + sz = offsetof(struct StackTrace, trace) + sizeof(BeamInstr *)*depth; + s = (struct StackTrace *) erts_alloc(ERTS_ALC_T_TMP, sz); + s->depth = 0; + if (rp->i) { + s->trace[s->depth++] = rp->i; + depth--; + } + if (depth > 0 && rp->cp != 0) { + s->trace[s->depth++] = rp->cp - 1; + depth--; + } + erts_save_stacktrace(rp, s, depth); + + depth = s->depth; + stk = stkp = (FunctionInfo *) erts_alloc(ERTS_ALC_T_TMP, + depth*sizeof(FunctionInfo)); + heap_size = 3; + for (i = 0; i < depth; i++) { + erts_lookup_function_info(stkp, s->trace[i], 1); + if (stkp->current) { + heap_size += stkp->needed + 2; + stkp++; + } + } + + hp = HAlloc(p, heap_size); + while (stkp > stk) { + stkp--; + hp = erts_build_mfa_item(stkp, hp, am_true, &mfa); + res = CONS(hp, mfa, res); + hp += 2; + } + + erts_free(ERTS_ALC_T_TMP, stk); + erts_free(ERTS_ALC_T_TMP, s); + *hpp = hp; + return res; +} + #if defined(VALGRIND) static int check_if_xml(void) { @@ -2026,7 +2121,7 @@ BIF_RETTYPE system_info_1(BIF_ALIST_1) res = TUPLE2(hp, am_sequential_tracer, val); BIF_RET(res); } else if (BIF_ARG_1 == am_garbage_collection){ - Uint val = (Uint) erts_smp_atomic32_read(&erts_max_gen_gcs); + Uint val = (Uint) erts_smp_atomic32_read_nob(&erts_max_gen_gcs); Eterm tup; hp = HAlloc(BIF_P, 3+2 + 3+2 + 3+2); @@ -2041,7 +2136,7 @@ BIF_RETTYPE system_info_1(BIF_ALIST_1) BIF_RET(res); } else if (BIF_ARG_1 == am_fullsweep_after){ - Uint val = (Uint) erts_smp_atomic32_read(&erts_max_gen_gcs); + Uint val = (Uint) erts_smp_atomic32_read_nob(&erts_max_gen_gcs); hp = HAlloc(BIF_P, 3); res = TUPLE2(hp, am_fullsweep_after, make_small(val)); BIF_RET(res); @@ -2545,6 +2640,70 @@ BIF_RETTYPE system_info_1(BIF_ALIST_1) hp = hsz ? HAlloc(BIF_P, hsz) : NULL; res = erts_bld_uint(&hp, NULL, erts_dist_buf_busy_limit); BIF_RET(res); + } else if (ERTS_IS_ATOM_STR("print_ethread_info", BIF_ARG_1)) { + int i; + char **str; +#ifdef ETHR_NATIVE_ATOMIC32_IMPL + erts_printf("32-bit native atomics: %s\n", + ETHR_NATIVE_ATOMIC32_IMPL); + str = ethr_native_atomic32_ops(); + for (i = 0; str[i]; i++) + erts_printf("ethr_native_atomic32_%s()\n", str[i]); +#endif +#ifdef ETHR_NATIVE_ATOMIC64_IMPL + erts_printf("64-bit native atomics: %s\n", + ETHR_NATIVE_ATOMIC64_IMPL); + str = ethr_native_atomic64_ops(); + for (i = 0; str[i]; i++) + erts_printf("ethr_native_atomic64_%s()\n", str[i]); +#endif +#ifdef ETHR_NATIVE_DW_ATOMIC_IMPL + if (ethr_have_native_dw_atomic()) { + erts_printf("Double word native atomics: %s\n", + ETHR_NATIVE_DW_ATOMIC_IMPL); + str = ethr_native_dw_atomic_ops(); + for (i = 0; str[i]; i++) + erts_printf("ethr_native_dw_atomic_%s()\n", str[i]); + str = ethr_native_su_dw_atomic_ops(); + for (i = 0; str[i]; i++) + erts_printf("ethr_native_su_dw_atomic_%s()\n", str[i]); + } +#endif +#ifdef ETHR_NATIVE_SPINLOCK_IMPL + erts_printf("Native spin-locks: %s\n", ETHR_NATIVE_SPINLOCK_IMPL); +#endif +#ifdef ETHR_NATIVE_RWSPINLOCK_IMPL + erts_printf("Native rwspin-locks: %s\n", ETHR_NATIVE_RWSPINLOCK_IMPL); +#endif +#ifdef ETHR_X86_RUNTIME_CONF_HAVE_SSE2__ + erts_printf("SSE2 support: %s\n", (ETHR_X86_RUNTIME_CONF_HAVE_SSE2__ + ? "yes" : "no")); +#endif +#ifdef ETHR_X86_OUT_OF_ORDER + erts_printf("x86" +#ifdef ARCH_64 + "_64" +#endif + " out of order\n"); +#endif +#ifdef ETHR_SPARC_TSO + erts_printf("Sparc TSO\n"); +#endif +#ifdef ETHR_SPARC_PSO + erts_printf("Sparc PSO\n"); +#endif +#ifdef ETHR_SPARC_RMO + erts_printf("Sparc RMO\n"); +#endif +#if defined(ETHR_PPC_HAVE_LWSYNC) + erts_printf("Have lwsync instruction: yes\n"); +#elif defined(ETHR_PPC_HAVE_NO_LWSYNC) + erts_printf("Have lwsync instruction: no\n"); +#elif defined(ETHR_PPC_RUNTIME_CONF_HAVE_LWSYNC__) + erts_printf("Have lwsync instruction: %s (runtime test)\n", + ETHR_PPC_RUNTIME_CONF_HAVE_LWSYNC__ ? "yes" : "no"); +#endif + BIF_RET(am_true); } BIF_ERROR(BIF_P, BADARG); @@ -2845,7 +3004,7 @@ fun_info_2(Process* p, Eterm fun, Eterm what) } break; case am_refc: - val = erts_make_integer(erts_smp_atomic_read(&funp->fe->refc), p); + val = erts_make_integer(erts_smp_atomic_read_nob(&funp->fe->refc), p); hp = HAlloc(p, 3); break; case am_arity: @@ -3065,8 +3224,8 @@ BIF_RETTYPE statistics_1(BIF_ALIST_1) Eterm r1, r2; Eterm in, out; Uint hsz = 9; - Uint bytes_in = (Uint) erts_smp_atomic_read(&erts_bytes_in); - Uint bytes_out = (Uint) erts_smp_atomic_read(&erts_bytes_out); + Uint bytes_in = (Uint) erts_smp_atomic_read_nob(&erts_bytes_in); + Uint bytes_out = (Uint) erts_smp_atomic_read_nob(&erts_bytes_out); (void) erts_bld_uint(NULL, &hsz, bytes_in); (void) erts_bld_uint(NULL, &hsz, bytes_out); @@ -3139,7 +3298,7 @@ BIF_RETTYPE erts_debug_get_internal_state_1(BIF_ALIST_1) * NOTE: Only supposed to be used for testing, and debugging. */ - if (!erts_smp_atomic_read(&available_internal_state)) { + if (!erts_smp_atomic_read_nob(&available_internal_state)) { BIF_ERROR(BIF_P, EXC_UNDEF); } @@ -3437,7 +3596,7 @@ BIF_RETTYPE erts_debug_set_internal_state_2(BIF_ALIST_2) if (ERTS_IS_ATOM_STR("available_internal_state", BIF_ARG_1) && (BIF_ARG_2 == am_true || BIF_ARG_2 == am_false)) { erts_aint_t on = (erts_aint_t) (BIF_ARG_2 == am_true); - erts_aint_t prev_on = erts_smp_atomic_xchg(&available_internal_state, on); + erts_aint_t prev_on = erts_smp_atomic_xchg_nob(&available_internal_state, on); if (on) { erts_dsprintf_buf_t *dsbufp = erts_create_logger_dsbuf(); erts_dsprintf(dsbufp, "Process %T ", BIF_P->id); @@ -3453,7 +3612,7 @@ BIF_RETTYPE erts_debug_set_internal_state_2(BIF_ALIST_2) BIF_RET(prev_on ? am_true : am_false); } - if (!erts_smp_atomic_read(&available_internal_state)) { + if (!erts_smp_atomic_read_nob(&available_internal_state)) { BIF_ERROR(BIF_P, EXC_UNDEF); } @@ -3634,14 +3793,14 @@ BIF_RETTYPE erts_debug_set_internal_state_2(BIF_ALIST_2) } else if (ERTS_IS_ATOM_STR("hipe_test_reschedule_suspend", BIF_ARG_1)) { /* Used by hipe test suites */ - erts_aint_t flag = erts_smp_atomic_read(&hipe_test_reschedule_flag); + erts_aint_t flag = erts_smp_atomic_read_nob(&hipe_test_reschedule_flag); if (!flag && BIF_ARG_2 != am_false) { - erts_smp_atomic_set(&hipe_test_reschedule_flag, 1); + erts_smp_atomic_set_nob(&hipe_test_reschedule_flag, 1); erts_suspend(BIF_P, ERTS_PROC_LOCK_MAIN, NULL); ERTS_BIF_YIELD2(bif_export[BIF_erts_debug_set_internal_state_2], BIF_P, BIF_ARG_1, BIF_ARG_2); } - erts_smp_atomic_set(&hipe_test_reschedule_flag, !flag); + erts_smp_atomic_set_nob(&hipe_test_reschedule_flag, !flag); BIF_RET(NIL); } else if (ERTS_IS_ATOM_STR("hipe_test_reschedule_resume", BIF_ARG_1)) { @@ -3951,8 +4110,8 @@ BIF_RETTYPE erts_debug_lock_counters_1(BIF_ALIST_1) void erts_bif_info_init(void) { - erts_smp_atomic_init(&available_internal_state, 0); - erts_smp_atomic_init(&hipe_test_reschedule_flag, 0); + erts_smp_atomic_init_nob(&available_internal_state, 0); + erts_smp_atomic_init_nob(&hipe_test_reschedule_flag, 0); process_info_init(); } diff --git a/erts/emulator/beam/erl_bif_trace.c b/erts/emulator/beam/erl_bif_trace.c index 0509e51a6f..7a08182e18 100644 --- a/erts/emulator/beam/erl_bif_trace.c +++ b/erts/emulator/beam/erl_bif_trace.c @@ -1,7 +1,7 @@ /* * %CopyrightBegin% * - * Copyright Ericsson AB 1999-2010. All Rights Reserved. + * Copyright Ericsson AB 1999-2011. All Rights Reserved. * * The contents of this file are subject to the Erlang Public License, * Version 1.1, (the "License"); you may not use this file except in @@ -679,7 +679,7 @@ trace_3(Process* p, Eterm pid_spec, Eterm how, Eterm list) } else if (tracer != NIL) { tracee_port->tracer_proc = tracer; } - /* matches are not counted for ports since it would violate compability */ + /* matches are not counted for ports since it would violate compatibility */ /* This could be a reason to modify this function or make a new one. */ } } diff --git a/erts/emulator/beam/erl_bits.c b/erts/emulator/beam/erl_bits.c index e56084b9cb..326a5c136b 100644 --- a/erts/emulator/beam/erl_bits.c +++ b/erts/emulator/beam/erl_bits.c @@ -76,14 +76,12 @@ struct erl_bits_state ErlBitsState; #define byte_buf (ErlBitsState.byte_buf_) #define byte_buf_len (ErlBitsState.byte_buf_len_) -#ifdef ERTS_SMP static erts_smp_atomic_t bits_bufs_size; -#endif Uint erts_bits_bufs_size(void) { - return 0; + return (Uint) erts_smp_atomic_read_nob(&bits_bufs_size); } #if !defined(ERTS_SMP) @@ -109,8 +107,8 @@ erts_bits_destroy_state(ERL_BITS_PROTO_0) void erts_init_bits(void) { + erts_smp_atomic_init_nob(&bits_bufs_size, 0); #if defined(ERTS_SMP) - erts_smp_atomic_init(&bits_bufs_size, 0); /* erl_process.c calls erts_bits_init_state() on all state instances */ #else ERL_BITS_DECLARE_STATEP; @@ -713,9 +711,7 @@ static void ERTS_INLINE need_byte_buf(ERL_BITS_PROTO_1(int need)) { if (byte_buf_len < need) { -#ifdef ERTS_SMP - erts_smp_atomic_add(&bits_bufs_size, need - byte_buf_len); -#endif + erts_smp_atomic_add_nob(&bits_bufs_size, need - byte_buf_len); byte_buf_len = need; byte_buf = erts_realloc(ERTS_ALC_T_BITS_BUF, byte_buf, byte_buf_len); } diff --git a/erts/emulator/beam/erl_bits.h b/erts/emulator/beam/erl_bits.h index 0f67733fa4..388d943755 100644 --- a/erts/emulator/beam/erl_bits.h +++ b/erts/emulator/beam/erl_bits.h @@ -1,7 +1,7 @@ /* * %CopyrightBegin% * - * Copyright Ericsson AB 1999-2010. All Rights Reserved. + * Copyright Ericsson AB 1999-2011. All Rights Reserved. * * The contents of this file are subject to the Erlang Public License, * Version 1.1, (the "License"); you may not use this file except in @@ -150,7 +150,7 @@ void erts_bits_destroy_state(ERL_BITS_PROTO_0); * NBYTES(x) returns the number of bytes needed to store x bits. */ -#define NBYTES(x) (((x) + 7) >> 3) +#define NBYTES(x) (((Uint64)(x) + (Uint64) 7) >> 3) #define BYTE_OFFSET(ofs) ((Uint) (ofs) >> 3) #define BIT_OFFSET(ofs) ((ofs) & 7) diff --git a/erts/emulator/beam/erl_cpu_topology.c b/erts/emulator/beam/erl_cpu_topology.c index bcf8bcf270..03c0ef904a 100644 --- a/erts/emulator/beam/erl_cpu_topology.c +++ b/erts/emulator/beam/erl_cpu_topology.c @@ -487,7 +487,7 @@ erts_sched_check_cpu_bind_post_suspend(ErtsSchedulerData *esdp) /* Make sure we check if we should bind to a cpu or not... */ if (esdp->run_queue->flags & ERTS_RUNQ_FLG_SHARED_RUNQ) - erts_smp_atomic32_set(&esdp->chk_cpu_bind, 1); + erts_smp_atomic32_set_nob(&esdp->chk_cpu_bind, 1); else esdp->run_queue->flags |= ERTS_RUNQ_FLG_CHK_CPU_BIND; } @@ -503,7 +503,7 @@ erts_sched_check_cpu_bind(ErtsSchedulerData *esdp) erts_cpu_groups_callback_call_t *cgcc; #ifdef ERTS_SMP if (erts_common_run_queue) - erts_smp_atomic32_set(&esdp->chk_cpu_bind, 0); + erts_smp_atomic32_set_nob(&esdp->chk_cpu_bind, 0); else { esdp->run_queue->flags &= ~ERTS_RUNQ_FLG_CHK_CPU_BIND; } diff --git a/erts/emulator/beam/erl_db.c b/erts/emulator/beam/erl_db.c index e0a6aa05c6..0327850cb9 100644 --- a/erts/emulator/beam/erl_db.c +++ b/erts/emulator/beam/erl_db.c @@ -224,21 +224,21 @@ static void free_dbtable(DbTable* tb) { #ifdef HARDDEBUG - if (erts_smp_atomic_read(&tb->common.memory_size) != sizeof(DbTable)) { + if (erts_smp_atomic_read_nob(&tb->common.memory_size) != sizeof(DbTable)) { erts_fprintf(stderr, "ets: free_dbtable memory remain=%ld fix=%x\n", - erts_smp_atomic_read(&tb->common.memory_size)-sizeof(DbTable), + erts_smp_atomic_read_nob(&tb->common.memory_size)-sizeof(DbTable), tb->common.fixations); } erts_fprintf(stderr, "ets: free_dbtable(%T) deleted!!!\r\n", tb->common.id); erts_fprintf(stderr, "ets: free_dbtable: meta_pid_to_tab common.memory_size = %ld\n", - erts_smp_atomic_read(&meta_pid_to_tab->common.memory_size)); + erts_smp_atomic_read_nob(&meta_pid_to_tab->common.memory_size)); print_table(ERTS_PRINT_STDOUT, NULL, 1, meta_pid_to_tab); erts_fprintf(stderr, "ets: free_dbtable: meta_pid_to_fixed_tab common.memory_size = %ld\n", - erts_smp_atomic_read(&meta_pid_to_fixed_tab->common.memory_size)); + erts_smp_atomic_read_nob(&meta_pid_to_fixed_tab->common.memory_size)); print_table(ERTS_PRINT_STDOUT, NULL, 1, meta_pid_to_fixed_tab); #endif #ifdef ERTS_SMP @@ -248,6 +248,7 @@ free_dbtable(DbTable* tb) ASSERT(is_immed(tb->common.heir_data)); erts_db_free(ERTS_ALC_T_DB_TABLE, tb, (void *) tb, sizeof(DbTable)); ERTS_ETS_MISC_MEM_ADD(-sizeof(DbTable)); + ERTS_THR_MEMORY_BARRIER; } #ifdef ERTS_SMP @@ -338,13 +339,13 @@ static ERTS_INLINE void db_unlock(DbTable* tb, db_lock_kind_t kind) ASSERT(tb != meta_pid_to_tab && tb != meta_pid_to_fixed_tab); if (tb->common.type & DB_FINE_LOCKED) { - if (tb->common.is_thread_safe) { - ASSERT(kind == LCK_WRITE); + if (kind == LCK_WRITE) { + ASSERT(tb->common.is_thread_safe); tb->common.is_thread_safe = 0; erts_smp_rwmtx_rwunlock(&tb->common.rwlock); } else { - ASSERT(kind != LCK_WRITE); + ASSERT(!tb->common.is_thread_safe); erts_smp_rwmtx_runlock(&tb->common.rwlock); } } @@ -543,9 +544,9 @@ static int remove_named_tab(DbTable *tb, int have_lock) * We keep our increased refc over this op in order to * prevent the table from disapearing. */ - erts_smp_rwmtx_rwunlock(&tb->common.rwlock); + db_unlock(tb, LCK_WRITE); erts_smp_rwmtx_rwlock(rwlock); - erts_smp_rwmtx_rwlock(&tb->common.rwlock); + db_lock(tb, LCK_WRITE); } #endif @@ -1417,12 +1418,12 @@ BIF_RETTYPE ets_new_2(BIF_ALIST_2) { DbTable init_tb; - erts_smp_atomic_init(&init_tb.common.memory_size, 0); + erts_smp_atomic_init_nob(&init_tb.common.memory_size, 0); tb = (DbTable*) erts_db_alloc(ERTS_ALC_T_DB_TABLE, &init_tb, sizeof(DbTable)); ERTS_ETS_MISC_MEM_ADD(sizeof(DbTable)); - erts_smp_atomic_init(&tb->common.memory_size, - erts_smp_atomic_read(&init_tb.common.memory_size)); + erts_smp_atomic_init_nob(&tb->common.memory_size, + erts_smp_atomic_read_nob(&init_tb.common.memory_size)); } tb->common.meth = meth; @@ -1439,7 +1440,7 @@ BIF_RETTYPE ets_new_2(BIF_ALIST_2) tb->common.owner = BIF_P->id; set_heir(BIF_P, tb, heir, heir_data); - erts_smp_atomic_init(&tb->common.nitems, 0); + erts_smp_atomic_init_nob(&tb->common.nitems, 0); tb->common.fixations = NULL; tb->common.compress = is_compressed; @@ -1505,9 +1506,9 @@ BIF_RETTYPE ets_new_2(BIF_ALIST_2) BIF_ARG_1, BIF_ARG_2, ret, BIF_P->id, BIF_P->initial[0], BIF_P->initial[1], BIF_P->initial[2]); erts_fprintf(stderr, "ets: new: meta_pid_to_tab common.memory_size = %ld\n", - erts_smp_atomic_read(&meta_pid_to_tab->common.memory_size)); + erts_smp_atomic_read_nob(&meta_pid_to_tab->common.memory_size)); erts_fprintf(stderr, "ets: new: meta_pid_to_fixed_tab common.memory_size = %ld\n", - erts_smp_atomic_read(&meta_pid_to_fixed_tab->common.memory_size)); + erts_smp_atomic_read_nob(&meta_pid_to_fixed_tab->common.memory_size)); #endif UseTmpHeap(3,BIF_P); @@ -1650,24 +1651,6 @@ BIF_RETTYPE ets_delete_1(BIF_ALIST_1) tb->common.status &= ~(DB_PROTECTED|DB_PUBLIC|DB_PRIVATE); tb->common.status |= DB_DELETE; - mmtl = get_meta_main_tab_lock(tb->common.slot); -#ifdef ERTS_SMP - if (erts_smp_rwmtx_tryrwlock(mmtl) == EBUSY) { - /* - * We keep our increased refc over this op in order to - * prevent the table from disapearing. - */ - erts_smp_rwmtx_rwunlock(&tb->common.rwlock); - erts_smp_rwmtx_rwlock(mmtl); - erts_smp_rwmtx_rwlock(&tb->common.rwlock); - } -#endif - /* We must keep the slot, to be found by db_proc_dead() if process dies */ - MARK_SLOT_DEAD(tb->common.slot); - erts_smp_rwmtx_rwunlock(mmtl); - if (is_atom(tb->common.id)) - remove_named_tab(tb, 0); - if (tb->common.owner != BIF_P->id) { DeclareTmpHeap(meta_tuple,3,BIF_P); @@ -1691,6 +1674,25 @@ BIF_RETTYPE ets_delete_1(BIF_ALIST_1) db_meta_unlock(meta_pid_to_tab, LCK_WRITE_REC); UnUseTmpHeap(3,BIF_P); } + + mmtl = get_meta_main_tab_lock(tb->common.slot); +#ifdef ERTS_SMP + if (erts_smp_rwmtx_tryrwlock(mmtl) == EBUSY) { + /* + * We keep our increased refc over this op in order to + * prevent the table from disapearing. + */ + db_unlock(tb, LCK_WRITE); + erts_smp_rwmtx_rwlock(mmtl); + db_lock(tb, LCK_WRITE); + } +#endif + /* We must keep the slot, to be found by db_proc_dead() if process dies */ + MARK_SLOT_DEAD(tb->common.slot); + erts_smp_rwmtx_rwunlock(mmtl); + if (is_atom(tb->common.id)) + remove_named_tab(tb, 0); + /* disable inheritance */ free_heir_data(tb); tb->common.heir = am_none; @@ -1995,7 +1997,7 @@ BIF_RETTYPE ets_select_delete_2(BIF_ALIST_2) if ((tb = db_get_table(BIF_P, BIF_ARG_1, DB_WRITE, LCK_WRITE)) == NULL) { BIF_ERROR(BIF_P, BADARG); } - nitems = erts_smp_atomic_read(&tb->common.nitems); + nitems = erts_smp_atomic_read_nob(&tb->common.nitems); tb->common.meth->db_delete_all_objects(BIF_P, tb); db_unlock(tb, LCK_WRITE); BIF_RET(erts_make_integer(nitems,BIF_P)); @@ -2789,7 +2791,7 @@ void init_db(void) } #endif - erts_smp_atomic_init(&erts_ets_misc_mem_size, 0); + erts_smp_atomic_init_nob(&erts_ets_misc_mem_size, 0); db_initialize_util(); if (user_requested_db_max_tabs < DB_DEF_MAX_TABS) @@ -2831,13 +2833,13 @@ void init_db(void) /*TT*/ /* Create meta table invertion. */ - erts_smp_atomic_init(&init_tb.common.memory_size, 0); + erts_smp_atomic_init_nob(&init_tb.common.memory_size, 0); meta_pid_to_tab = (DbTable*) erts_db_alloc(ERTS_ALC_T_DB_TABLE, &init_tb, sizeof(DbTable)); ERTS_ETS_MISC_MEM_ADD(sizeof(DbTable)); - erts_smp_atomic_init(&meta_pid_to_tab->common.memory_size, - erts_smp_atomic_read(&init_tb.common.memory_size)); + erts_smp_atomic_init_nob(&meta_pid_to_tab->common.memory_size, + erts_smp_atomic_read_nob(&init_tb.common.memory_size)); meta_pid_to_tab->common.id = NIL; meta_pid_to_tab->common.the_name = am_true; @@ -2850,7 +2852,7 @@ void init_db(void) #endif meta_pid_to_tab->common.keypos = 1; meta_pid_to_tab->common.owner = NIL; - erts_smp_atomic_init(&meta_pid_to_tab->common.nitems, 0); + erts_smp_atomic_init_nob(&meta_pid_to_tab->common.nitems, 0); meta_pid_to_tab->common.slot = -1; meta_pid_to_tab->common.meth = &db_hash; meta_pid_to_tab->common.compress = 0; @@ -2863,13 +2865,13 @@ void init_db(void) erl_exit(1,"Unable to create ets metadata tables."); } - erts_smp_atomic_set(&init_tb.common.memory_size, 0); + erts_smp_atomic_set_nob(&init_tb.common.memory_size, 0); meta_pid_to_fixed_tab = (DbTable*) erts_db_alloc(ERTS_ALC_T_DB_TABLE, &init_tb, sizeof(DbTable)); ERTS_ETS_MISC_MEM_ADD(sizeof(DbTable)); - erts_smp_atomic_init(&meta_pid_to_fixed_tab->common.memory_size, - erts_smp_atomic_read(&init_tb.common.memory_size)); + erts_smp_atomic_init_nob(&meta_pid_to_fixed_tab->common.memory_size, + erts_smp_atomic_read_nob(&init_tb.common.memory_size)); meta_pid_to_fixed_tab->common.id = NIL; meta_pid_to_fixed_tab->common.the_name = am_true; @@ -2882,7 +2884,7 @@ void init_db(void) #endif meta_pid_to_fixed_tab->common.keypos = 1; meta_pid_to_fixed_tab->common.owner = NIL; - erts_smp_atomic_init(&meta_pid_to_fixed_tab->common.nitems, 0); + erts_smp_atomic_init_nob(&meta_pid_to_fixed_tab->common.nitems, 0); meta_pid_to_fixed_tab->common.slot = -1; meta_pid_to_fixed_tab->common.meth = &db_hash; meta_pid_to_fixed_tab->common.compress = 0; @@ -3421,7 +3423,7 @@ static void unfix_table_locked(Process* p, DbTable* tb, unlocked: if (!IS_FIXED(tb) && IS_HASH_TABLE(tb->common.status) - && erts_smp_atomic_read(&tb->hash.fixdel) != (erts_aint_t)NULL) { + && erts_smp_atomic_read_nob(&tb->hash.fixdel) != (erts_aint_t)NULL) { #ifdef ERTS_SMP if (*kind_p == LCK_READ && tb->common.is_thread_safe) { /* Must have write lock while purging pseudo-deleted (OTP-8166) */ @@ -3606,7 +3608,7 @@ static Eterm table_info(Process* p, DbTable* tb, Eterm What) Eterm ret = THE_NON_VALUE; if (What == am_size) { - ret = make_small(erts_smp_atomic_read(&tb->common.nitems)); + ret = make_small(erts_smp_atomic_read_nob(&tb->common.nitems)); } else if (What == am_type) { if (tb->common.status & DB_SET) { ret = am_set; @@ -3619,7 +3621,7 @@ static Eterm table_info(Process* p, DbTable* tb, Eterm What) ret = am_bag; } } else if (What == am_memory) { - Uint words = (Uint) ((erts_smp_atomic_read(&tb->common.memory_size) + Uint words = (Uint) ((erts_smp_atomic_read_nob(&tb->common.memory_size) + sizeof(Uint) - 1) / sizeof(Uint)); @@ -3657,9 +3659,6 @@ static Eterm table_info(Process* p, DbTable* tb, Eterm What) ret = am_true; else ret = am_false; - } else if (What == am_atom_put("kept_objects",12)) { - ret = make_small(IS_HASH_TABLE(tb->common.status) - ? db_kept_items_hash(&tb->hash) : 0); } else if (What == am_atom_put("safe_fixed",10)) { #ifdef ERTS_SMP erts_smp_mtx_lock(&tb->common.fixlock); @@ -3701,7 +3700,7 @@ static Eterm table_info(Process* p, DbTable* tb, Eterm What) Eterm* hp; db_calc_stats_hash(&tb->hash, &stats); - hp = HAlloc(p, 1 + 6 + FLOAT_SIZE_OBJECT*3); + hp = HAlloc(p, 1 + 7 + FLOAT_SIZE_OBJECT*3); f.fd = stats.avg_chain_len; avg = make_float(hp); PUT_DOUBLE(f, hp); @@ -3716,10 +3715,11 @@ static Eterm table_info(Process* p, DbTable* tb, Eterm What) std_dev_exp = make_float(hp); PUT_DOUBLE(f, hp); hp += FLOAT_SIZE_OBJECT; - ret = TUPLE6(hp, make_small(erts_smp_atomic_read(&tb->hash.nactive)), + ret = TUPLE7(hp, make_small(erts_smp_atomic_read_nob(&tb->hash.nactive)), avg, std_dev_real, std_dev_exp, make_small(stats.min_chain_len), - make_small(stats.max_chain_len)); + make_small(stats.max_chain_len), + make_small(db_kept_items_hash(&tb->hash))); } else { ret = am_false; @@ -3735,9 +3735,9 @@ static void print_table(int to, void *to_arg, int show, DbTable* tb) tb->common.meth->db_print(to, to_arg, show, tb); - erts_print(to, to_arg, "Objects: %d\n", (int)erts_smp_atomic_read(&tb->common.nitems)); + erts_print(to, to_arg, "Objects: %d\n", (int)erts_smp_atomic_read_nob(&tb->common.nitems)); erts_print(to, to_arg, "Words: %bpu\n", - (UWord) ((erts_smp_atomic_read(&tb->common.memory_size) + (Uint) ((erts_smp_atomic_read_nob(&tb->common.memory_size) + sizeof(Uint) - 1) / sizeof(Uint))); @@ -3763,8 +3763,9 @@ void db_info(int to, void *to_arg, int show) /* Called by break handler */ Uint erts_get_ets_misc_mem_size(void) { + ERTS_THR_MEMORY_BARRIER; /* Memory not allocated in ets_alloc */ - return (Uint) erts_smp_atomic_read(&erts_ets_misc_mem_size); + return (Uint) erts_smp_atomic_read_nob(&erts_ets_misc_mem_size); } /* SMP Note: May only be used when system is locked */ diff --git a/erts/emulator/beam/erl_db.h b/erts/emulator/beam/erl_db.h index e0bdebcb01..2e5deaf338 100644 --- a/erts/emulator/beam/erl_db.h +++ b/erts/emulator/beam/erl_db.h @@ -1,7 +1,7 @@ /* * %CopyrightBegin% * - * Copyright Ericsson AB 1996-2010. All Rights Reserved. + * Copyright Ericsson AB 1996-2011. All Rights Reserved. * * The contents of this file are subject to the Erlang Public License, * Version 1.1, (the "License"); you may not use this file except in @@ -86,11 +86,11 @@ do { \ erts_aint_t sz__ = (((erts_aint_t) (ALLOC_SZ)) \ - ((erts_aint_t) (FREE_SZ))); \ ASSERT((TAB)); \ - erts_smp_atomic_add(&(TAB)->common.memory_size, sz__); \ + erts_smp_atomic_add_nob(&(TAB)->common.memory_size, sz__); \ } while (0) #define ERTS_ETS_MISC_MEM_ADD(SZ) \ - erts_smp_atomic_add(&erts_ets_misc_mem_size, (SZ)); + erts_smp_atomic_add_nob(&erts_ets_misc_mem_size, (SZ)); ERTS_GLB_INLINE void *erts_db_alloc(ErtsAlcType_t type, DbTable *tab, @@ -227,7 +227,7 @@ erts_db_free(ErtsAlcType_t type, DbTable *tab, void *ptr, Uint size) ERTS_DB_ALC_MEM_UPDATE_(tab, size, 0); ASSERT(((void *) tab) != ptr - || erts_smp_atomic_read(&tab->common.memory_size) == 0); + || erts_smp_atomic_read_nob(&tab->common.memory_size) == 0); erts_free(type, ptr); } diff --git a/erts/emulator/beam/erl_db_hash.c b/erts/emulator/beam/erl_db_hash.c index fdc82c8b88..e3380a57b2 100644 --- a/erts/emulator/beam/erl_db_hash.c +++ b/erts/emulator/beam/erl_db_hash.c @@ -105,9 +105,13 @@ #define NSEG_2 256 /* Size of second segment table */ #define NSEG_INC 128 /* Number of segments to grow after that */ -#define SEGTAB(tb) ((struct segment**)erts_smp_atomic_read_acqb(&(tb)->segtab)) -#define NACTIVE(tb) ((int)erts_smp_atomic_read(&(tb)->nactive)) -#define NITEMS(tb) ((int)erts_smp_atomic_read(&(tb)->common.nitems)) +#ifdef ETHR_ORDERED_READ_DEPEND +#define SEGTAB(tb) ((struct segment**)erts_smp_atomic_read_nob(&(tb)->segtab)) +#else +#define SEGTAB(tb) ((struct segment**)erts_smp_atomic_read_rb(&(tb)->segtab)) +#endif +#define NACTIVE(tb) ((int)erts_smp_atomic_read_nob(&(tb)->nactive)) +#define NITEMS(tb) ((int)erts_smp_atomic_read_nob(&(tb)->common.nitems)) #define BUCKET(tb, i) SEGTAB(tb)[(i) >> SEGSZ_EXP]->buckets[(i) & SEGSZ_MASK] @@ -123,10 +127,10 @@ static ERTS_INLINE Uint hash_to_ix(DbTableHash* tb, HashValue hval) { Uint mask = erts_smp_atomic_read_acqb(&tb->szm); - Uint ix = hval & mask; - if (ix >= erts_smp_atomic_read(&tb->nactive)) { + Uint ix = hval & mask; + if (ix >= erts_smp_atomic_read_nob(&tb->nactive)) { ix &= mask>>1; - ASSERT(ix < erts_smp_atomic_read(&tb->nactive)); + ASSERT(ix < erts_smp_atomic_read_nob(&tb->nactive)); } return ix; } @@ -141,14 +145,14 @@ static ERTS_INLINE void add_fixed_deletion(DbTableHash* tb, int ix) (DbTable *) tb, sizeof(FixedDeletion)); ERTS_ETS_MISC_MEM_ADD(sizeof(FixedDeletion)); - fixd->slot = ix; - was_next = erts_smp_atomic_read(&tb->fixdel); + fixd->slot = ix; + was_next = erts_smp_atomic_read_acqb(&tb->fixdel); do { /* Lockless atomic insertion in linked list: */ exp_next = was_next; fixd->next = (FixedDeletion*) exp_next; - was_next = erts_smp_atomic_cmpxchg(&tb->fixdel, - (erts_aint_t) fixd, - exp_next); + was_next = erts_smp_atomic_cmpxchg_relb(&tb->fixdel, + (erts_aint_t) fixd, + exp_next); }while (was_next != exp_next); } @@ -540,22 +544,24 @@ static void restore_fixdel(DbTableHash* tb, FixedDeletion* fixdel) { /*int tries = 0;*/ DEBUG_WAIT(); - if (erts_smp_atomic_cmpxchg(&tb->fixdel, (erts_aint_t)fixdel, - (erts_aint_t)NULL) != (erts_aint_t)NULL) { + if (erts_smp_atomic_cmpxchg_relb(&tb->fixdel, + (erts_aint_t) fixdel, + (erts_aint_t) NULL) != (erts_aint_t) NULL) { /* Oboy, must join lists */ FixedDeletion* last = fixdel; erts_aint_t was_tail; erts_aint_t exp_tail; - while (last->next != NULL) last = last->next; - was_tail = erts_smp_atomic_read(&tb->fixdel); + while (last->next != NULL) last = last->next; + was_tail = erts_smp_atomic_read_acqb(&tb->fixdel); do { /* Lockless atomic list insertion */ exp_tail = was_tail; last->next = (FixedDeletion*) exp_tail; /*++tries;*/ DEBUG_WAIT(); - was_tail = erts_smp_atomic_cmpxchg(&tb->fixdel, (erts_aint_t)fixdel, - exp_tail); + was_tail = erts_smp_atomic_cmpxchg_relb(&tb->fixdel, + (erts_aint_t) fixdel, + exp_tail); }while (was_tail != exp_tail); } /*erts_fprintf(stderr,"erl_db_hash: restore_fixdel tries=%d\r\n", tries);*/ @@ -572,7 +578,8 @@ void db_unfix_table_hash(DbTableHash *tb) || (erts_smp_lc_rwmtx_is_rlocked(&tb->common.rwlock) && !tb->common.is_thread_safe)); restart: - fixdel = (FixedDeletion*) erts_smp_atomic_xchg(&tb->fixdel, (erts_aint_t)NULL); + fixdel = (FixedDeletion*) erts_smp_atomic_xchg_acqb(&tb->fixdel, + (erts_aint_t) NULL); while (fixdel != NULL) { FixedDeletion *fx = fixdel; int ix = fx->slot; @@ -639,14 +646,14 @@ int db_create_hash(Process *p, DbTable *tbl) { DbTableHash *tb = &tbl->hash; - erts_smp_atomic_init(&tb->szm, SEGSZ_MASK); - erts_smp_atomic_init(&tb->nactive, SEGSZ); - erts_smp_atomic_init(&tb->fixdel, (erts_aint_t)NULL); - erts_smp_atomic_init(&tb->segtab, (erts_aint_t) alloc_ext_seg(tb,0,NULL)->segtab); + erts_smp_atomic_init_nob(&tb->szm, SEGSZ_MASK); + erts_smp_atomic_init_nob(&tb->nactive, SEGSZ); + erts_smp_atomic_init_nob(&tb->fixdel, (erts_aint_t)NULL); + erts_smp_atomic_init_nob(&tb->segtab, (erts_aint_t) alloc_ext_seg(tb,0,NULL)->segtab); tb->nsegs = NSEG_1; tb->nslots = SEGSZ; - erts_smp_atomic_init(&tb->is_resizing, 0); + erts_smp_atomic_init_nob(&tb->is_resizing, 0); #ifdef ERTS_SMP if (tb->common.type & DB_FINE_LOCKED) { erts_smp_rwmtx_opt_t rwmtx_opt = ERTS_SMP_RWMTX_OPT_DEFAULT_INITER; @@ -663,7 +670,7 @@ int db_create_hash(Process *p, DbTable *tbl) /* This important property is needed to guarantee that the buckets * involved in a grow/shrink operation it protected by the same lock: */ - ASSERT(erts_smp_atomic_read(&tb->nactive) % DB_HASH_LOCK_CNT == 0); + ASSERT(erts_smp_atomic_read_nob(&tb->nactive) % DB_HASH_LOCK_CNT == 0); } else { /* coarse locking */ tb->locks = NULL; @@ -783,7 +790,7 @@ int db_put_hash(DbTable *tbl, Eterm obj, int key_clash_fail) if (tb->common.status & DB_SET) { HashDbTerm* bnext = b->next; if (b->hvalue == INVALID_HASH) { - erts_smp_atomic_inc(&tb->common.nitems); + erts_smp_atomic_inc_nob(&tb->common.nitems); } else if (key_clash_fail) { ret = DB_ERROR_BADKEY; @@ -811,7 +818,7 @@ int db_put_hash(DbTable *tbl, Eterm obj, int key_clash_fail) do { if (db_eq(&tb->common,obj,&q->dbterm)) { if (q->hvalue == INVALID_HASH) { - erts_smp_atomic_inc(&tb->common.nitems); + erts_smp_atomic_inc_nob(&tb->common.nitems); q->hvalue = hval; if (q != b) { /* must move to preserve key insertion order */ *qp = q->next; @@ -832,7 +839,7 @@ Lnew: q->hvalue = hval; q->next = b; *bp = q; - nitems = erts_smp_atomic_inctest(&tb->common.nitems); + nitems = erts_smp_atomic_inc_read_nob(&tb->common.nitems); WUNLOCK_HASH(lck); { int nactive = NACTIVE(tb); @@ -1069,7 +1076,7 @@ int db_erase_bag_exact2(DbTable *tbl, Eterm key, Eterm value) EQ(value, b->dbterm.tpl[2])) { *bp = b->next; free_term(tb, b); - erts_smp_atomic_dec(&tb->common.nitems); + erts_smp_atomic_dec_nob(&tb->common.nitems); b = *bp; break; } @@ -1128,7 +1135,7 @@ int db_erase_hash(DbTable *tbl, Eterm key, Eterm *ret) } WUNLOCK_HASH(lck); if (nitems_diff) { - erts_smp_atomic_add(&tb->common.nitems, nitems_diff); + erts_smp_atomic_add_nob(&tb->common.nitems, nitems_diff); try_shrink(tb); } *ret = am_true; @@ -1187,7 +1194,7 @@ static int db_erase_object_hash(DbTable *tbl, Eterm object, Eterm *ret) } WUNLOCK_HASH(lck); if (nitems_diff) { - erts_smp_atomic_add(&tb->common.nitems, nitems_diff); + erts_smp_atomic_add_nob(&tb->common.nitems, nitems_diff); try_shrink(tb); } *ret = am_true; @@ -1798,7 +1805,7 @@ static int db_select_delete_hash(Process *p, free_term(tb, del); did_erase = 1; } - erts_smp_atomic_dec(&tb->common.nitems); + erts_smp_atomic_dec_nob(&tb->common.nitems); ++got; } --num_left; @@ -1909,7 +1916,7 @@ static int db_select_delete_continue_hash(Process *p, free_term(tb, del); did_erase = 1; } - erts_smp_atomic_dec(&tb->common.nitems); + erts_smp_atomic_dec_nob(&tb->common.nitems); ++got; } @@ -2064,7 +2071,7 @@ int db_mark_all_deleted_hash(DbTable *tbl) }while(list != NULL); } } - erts_smp_atomic_set(&tb->common.nitems, 0); + erts_smp_atomic_set_nob(&tb->common.nitems, 0); return DB_ERROR_NONE; } @@ -2115,7 +2122,7 @@ static int db_free_table_continue_hash(DbTable *tbl) { DbTableHash *tb = &tbl->hash; int done; - FixedDeletion* fixdel = (FixedDeletion*) erts_smp_atomic_read(&tb->fixdel); + FixedDeletion* fixdel = (FixedDeletion*) erts_smp_atomic_read_acqb(&tb->fixdel); ERTS_SMP_LC_ASSERT(IS_TAB_WLOCKED(tb)); done = 0; @@ -2129,11 +2136,11 @@ static int db_free_table_continue_hash(DbTable *tbl) sizeof(FixedDeletion)); ERTS_ETS_MISC_MEM_ADD(-sizeof(FixedDeletion)); if (++done >= 2*DELETE_RECORD_LIMIT) { - erts_smp_atomic_set(&tb->fixdel, (erts_aint_t)fixdel); + erts_smp_atomic_set_relb(&tb->fixdel, (erts_aint_t)fixdel); return 0; /* Not done */ } } - erts_smp_atomic_set(&tb->fixdel, (erts_aint_t)NULL); + erts_smp_atomic_set_relb(&tb->fixdel, (erts_aint_t)NULL); done /= 2; while(tb->nslots != 0) { @@ -2157,7 +2164,7 @@ static int db_free_table_continue_hash(DbTable *tbl) tb->locks = NULL; } #endif - ASSERT(erts_smp_atomic_read(&tb->common.memory_size) == sizeof(DbTable)); + ASSERT(erts_smp_atomic_read_nob(&tb->common.memory_size) == sizeof(DbTable)); return 1; /* Done */ } @@ -2350,7 +2357,7 @@ static int alloc_seg(DbTableHash *tb) struct ext_segment* eseg; eseg = (struct ext_segment*) SEGTAB(tb)[seg_ix-1]; MY_ASSERT(eseg!=NULL && eseg->s.is_ext_segment); - erts_smp_atomic_set_relb(&tb->segtab, (erts_aint_t) eseg->segtab); + erts_smp_atomic_set_wb(&tb->segtab, (erts_aint_t) eseg->segtab); tb->nsegs = eseg->nsegs; } ASSERT(seg_ix < tb->nsegs); @@ -2422,7 +2429,7 @@ static int free_seg(DbTableHash *tb, int free_records) MY_ASSERT(newtop->s.is_ext_segment); if (newtop->prev_segtab != NULL) { /* Time to use a smaller segtab */ - erts_smp_atomic_set_relb(&tb->segtab, (erts_aint_t)newtop->prev_segtab); + erts_smp_atomic_set_wb(&tb->segtab, (erts_aint_t)newtop->prev_segtab); tb->nsegs = seg_ix; ASSERT(tb->nsegs == EXTSEG(SEGTAB(tb))->nsegs); } @@ -2439,7 +2446,7 @@ static int free_seg(DbTableHash *tb, int free_records) if (seg_ix > 0) { if (seg_ix < tb->nsegs) SEGTAB(tb)[seg_ix] = NULL; } else { - erts_smp_atomic_set_relb(&tb->segtab, (erts_aint_t)NULL); + erts_smp_atomic_set_wb(&tb->segtab, (erts_aint_t)NULL); } #endif tb->nslots -= SEGSZ; @@ -2500,7 +2507,7 @@ static void grow(DbTableHash* tb, int nactive) int from_ix; int szm; - if (erts_smp_atomic_xchg(&tb->is_resizing, 1)) { + if (erts_smp_atomic_xchg_acqb(&tb->is_resizing, 1)) { return; /* already in progress */ } if (NACTIVE(tb) != nactive) { @@ -2515,7 +2522,7 @@ static void grow(DbTableHash* tb, int nactive) } ASSERT(nactive < tb->nslots); - szm = erts_smp_atomic_read(&tb->szm); + szm = erts_smp_atomic_read_nob(&tb->szm); if (nactive <= szm) { from_ix = nactive & (szm >> 1); } else { @@ -2532,7 +2539,7 @@ static void grow(DbTableHash* tb, int nactive) WUNLOCK_HASH(lck); goto abort; } - erts_smp_atomic_inc(&tb->nactive); + erts_smp_atomic_inc_nob(&tb->nactive); if (from_ix == 0) { erts_smp_atomic_set_relb(&tb->szm, szm); } @@ -2577,13 +2584,13 @@ abort: */ static void shrink(DbTableHash* tb, int nactive) { - if (erts_smp_atomic_xchg(&tb->is_resizing, 1)) { + if (erts_smp_atomic_xchg_acqb(&tb->is_resizing, 1)) { return; /* already in progress */ } if (NACTIVE(tb) == nactive) { erts_smp_rwmtx_t* lck; int src_ix = nactive - 1; - int low_szm = erts_smp_atomic_read(&tb->szm) >> 1; + int low_szm = erts_smp_atomic_read_nob(&tb->szm) >> 1; int dst_ix = src_ix & low_szm; ASSERT(dst_ix < src_ix); @@ -2610,7 +2617,7 @@ static void shrink(DbTableHash* tb, int nactive) *dst_bp = *src_bp; *src_bp = NULL; - erts_smp_atomic_set(&tb->nactive, src_ix); + erts_smp_atomic_set_nob(&tb->nactive, src_ix); if (dst_ix == 0) { erts_smp_atomic_set_relb(&tb->szm, low_szm); } @@ -2746,7 +2753,7 @@ static int db_delete_all_objects_hash(Process* p, DbTable* tbl) } else { db_free_table_hash(tbl); db_create_hash(p, tbl); - erts_smp_atomic_set(&tbl->hash.common.nitems, 0); + erts_smp_atomic_set_nob(&tbl->hash.common.nitems, 0); } return 0; } diff --git a/erts/emulator/beam/erl_db_tree.c b/erts/emulator/beam/erl_db_tree.c index 9a0ba3a418..c6f0d80e32 100644 --- a/erts/emulator/beam/erl_db_tree.c +++ b/erts/emulator/beam/erl_db_tree.c @@ -49,7 +49,7 @@ #include "erl_db_tree.h" #define GETKEY_WITH_POS(Keypos, Tplp) (*((Tplp) + Keypos)) -#define NITEMS(tb) ((int)erts_smp_atomic_read(&(tb)->common.nitems)) +#define NITEMS(tb) ((int)erts_smp_atomic_read_nob(&(tb)->common.nitems)) /* ** A stack of this size is enough for an AVL tree with more than @@ -84,7 +84,7 @@ */ static DbTreeStack* get_static_stack(DbTableTree* tb) { - if (!erts_smp_atomic_xchg(&tb->is_stack_busy, 1)) { + if (!erts_smp_atomic_xchg_acqb(&tb->is_stack_busy, 1)) { return &tb->static_stack; } return NULL; @@ -96,7 +96,7 @@ static DbTreeStack* get_static_stack(DbTableTree* tb) static DbTreeStack* get_any_stack(DbTableTree* tb) { DbTreeStack* stack; - if (!erts_smp_atomic_xchg(&tb->is_stack_busy, 1)) { + if (!erts_smp_atomic_xchg_acqb(&tb->is_stack_busy, 1)) { return &tb->static_stack; } stack = erts_db_alloc(ERTS_ALC_T_DB_STK, (DbTable *) tb, @@ -110,7 +110,7 @@ static DbTreeStack* get_any_stack(DbTableTree* tb) static void release_stack(DbTableTree* tb, DbTreeStack* stack) { if (stack == &tb->static_stack) { - ASSERT(erts_smp_atomic_read(&tb->is_stack_busy) == 1); + ASSERT(erts_smp_atomic_read_nob(&tb->is_stack_busy) == 1); erts_smp_atomic_set_relb(&tb->is_stack_busy, 0); } else { @@ -478,7 +478,7 @@ int db_create_tree(Process *p, DbTable *tbl) sizeof(TreeDbTerm *) * STACK_NEED); tb->static_stack.pos = 0; tb->static_stack.slot = 0; - erts_smp_atomic_init(&tb->is_stack_busy, 0); + erts_smp_atomic_init_nob(&tb->is_stack_busy, 0); tb->deletion = 0; return DB_ERROR_NONE; } @@ -613,8 +613,8 @@ static int db_put_tree(DbTable *tbl, Eterm obj, int key_clash_fail) for (;;) if (!*this) { /* Found our place */ state = 1; - if (erts_smp_atomic_inctest(&tb->common.nitems) >= TREE_MAX_ELEMENTS) { - erts_smp_atomic_dec(&tb->common.nitems); + if (erts_smp_atomic_inc_read_nob(&tb->common.nitems) >= TREE_MAX_ELEMENTS) { + erts_smp_atomic_dec_nob(&tb->common.nitems); return DB_ERROR_SYSRES; } *this = new_dbterm(tb, obj); @@ -1583,7 +1583,7 @@ static int db_select_delete_continue_tree(Process *p, sc.max = 1000; sc.keypos = tb->common.keypos; - ASSERT(!erts_smp_atomic_read(&tb->is_stack_busy)); + ASSERT(!erts_smp_atomic_read_nob(&tb->is_stack_busy)); traverse_backwards(tb, &tb->static_stack, lastkey, NULL, &doit_select_delete, &sc); BUMP_REDS(p, 1000 - sc.max); @@ -1774,7 +1774,7 @@ static int db_free_table_continue_tree(DbTable *tbl) (DbTable *) tb, (void *) tb->static_stack.array, sizeof(TreeDbTerm *) * STACK_NEED); - ASSERT(erts_smp_atomic_read(&tb->common.memory_size) + ASSERT(erts_smp_atomic_read_nob(&tb->common.memory_size) == sizeof(DbTable)); } return result; @@ -1784,7 +1784,7 @@ static int db_delete_all_objects_tree(Process* p, DbTable* tbl) { db_free_table_tree(tbl); db_create_tree(p, tbl); - erts_smp_atomic_set(&tbl->tree.common.nitems, 0); + erts_smp_atomic_set_nob(&tbl->tree.common.nitems, 0); return 0; } @@ -1866,7 +1866,7 @@ static TreeDbTerm *linkout_tree(DbTableTree *tb, tstack[tpos++] = this; state = delsub(this); } - erts_smp_atomic_dec(&tb->common.nitems); + erts_smp_atomic_dec_nob(&tb->common.nitems); break; } } @@ -1933,7 +1933,7 @@ static TreeDbTerm *linkout_object_tree(DbTableTree *tb, tstack[tpos++] = this; state = delsub(this); } - erts_smp_atomic_dec(&tb->common.nitems); + erts_smp_atomic_dec_nob(&tb->common.nitems); break; } } diff --git a/erts/emulator/beam/erl_db_util.c b/erts/emulator/beam/erl_db_util.c index c3b074f782..7dfbb2ed02 100644 --- a/erts/emulator/beam/erl_db_util.c +++ b/erts/emulator/beam/erl_db_util.c @@ -491,7 +491,7 @@ erts_match_set_release_result(Process* c_p) /* The trace control word. */ -static erts_smp_atomic_t trace_control_word; +static erts_smp_atomic32_t trace_control_word; /* This needs to be here, before the bif table... */ @@ -911,7 +911,7 @@ static void db_free_tmp_uncompressed(DbTerm* obj); BIF_RETTYPE db_get_trace_control_word_0(Process *p) { - Uint32 tcw = (Uint32) erts_smp_atomic_read(&trace_control_word); + Uint32 tcw = (Uint32) erts_smp_atomic32_read_acqb(&trace_control_word); BIF_RET(erts_make_integer((Uint) tcw, p)); } @@ -924,7 +924,8 @@ BIF_RETTYPE db_set_trace_control_word_1(Process *p, Eterm new) if (val != ((Uint32)val)) BIF_ERROR(p, BADARG); - old_tcw = (Uint32) erts_smp_atomic_xchg(&trace_control_word, (erts_aint_t) val); + old_tcw = (Uint32) erts_smp_atomic32_xchg_relb(&trace_control_word, + (erts_aint32_t) val); BIF_RET(erts_make_integer((Uint) old_tcw, p)); } @@ -1249,7 +1250,7 @@ void db_initialize_util(void){ sizeof(DMCGuardBif), (int (*)(const void *, const void *)) &cmp_guard_bif); match_pseudo_process_init(); - erts_smp_atomic_init(&trace_control_word, 0); + erts_smp_atomic32_init_nob(&trace_control_word, 0); } @@ -1731,6 +1732,7 @@ Eterm db_prog_match(Process *c_p, Binary *bprog, #define BEGIN_ATOMIC_TRACE(p) \ do { \ if (! atomic_trace) { \ + erts_refc_inc(&bprog->refc, 2); \ erts_smp_proc_unlock((p), ERTS_PROC_LOCK_MAIN); \ erts_smp_block_system(0); \ atomic_trace = !0; \ @@ -1741,6 +1743,9 @@ Eterm db_prog_match(Process *c_p, Binary *bprog, if (atomic_trace) { \ erts_smp_release_system(); \ erts_smp_proc_lock((p), ERTS_PROC_LOCK_MAIN); \ + if (erts_refc_dectest(&bprog->refc, 0) == 0) {\ + erts_bin_free(bprog); \ + } \ atomic_trace = 0; \ } \ } while (0) diff --git a/erts/emulator/beam/erl_gc.c b/erts/emulator/beam/erl_gc.c index 5edcd667e7..e3445bcdc5 100644 --- a/erts/emulator/beam/erl_gc.c +++ b/erts/emulator/beam/erl_gc.c @@ -100,14 +100,14 @@ static Uint combined_message_size(Process* p); static void remove_message_buffers(Process* p); static int major_collection(Process* p, int need, Eterm* objv, int nobj, Uint *recl); static int minor_collection(Process* p, int need, Eterm* objv, int nobj, Uint *recl); -static void do_minor(Process *p, int new_sz, Eterm* objv, int nobj); +static void do_minor(Process *p, Uint new_sz, Eterm* objv, int nobj); static Eterm* sweep_rootset(Rootset *rootset, Eterm* htop, char* src, Uint src_size); static Eterm* sweep_one_area(Eterm* n_hp, Eterm* n_htop, char* src, Uint src_size); static Eterm* sweep_one_heap(Eterm* heap_ptr, Eterm* heap_end, Eterm* htop, char* src, Uint src_size); static Eterm* collect_heap_frags(Process* p, Eterm* heap, Eterm* htop, Eterm* objv, int nobj); -static Uint adjust_after_fullsweep(Process *p, int size_before, +static Uint adjust_after_fullsweep(Process *p, Uint size_before, int need, Eterm *objv, int nobj); static void shrink_new_heap(Process *p, Uint new_sz, Eterm *objv, int nobj); static void grow_new_heap(Process *p, Uint new_sz, Eterm* objv, int nobj); @@ -441,7 +441,15 @@ erts_garbage_collect(Process* p, int need, Eterm* objv, int nobj) p->last_old_htop = p->old_htop; #endif - return ((int) (HEAP_TOP(p) - HEAP_START(p)) / 10); + /* FIXME: This function should really return an Sint, i.e., a possibly + 64 bit wide signed integer, but that requires updating all the code + that calls it. For now, we just return INT_MAX if the result is too + large for an int. */ + { + Sint result = (HEAP_TOP(p) - HEAP_START(p)) / 10; + if (result >= INT_MAX) return INT_MAX; + else return (int) result; + } } /* @@ -599,7 +607,7 @@ erts_garbage_collect_literals(Process* p, Eterm* literals, Uint lit_size) char* area; Uint area_size; Eterm* old_htop; - int n; + Uint n; /* * Set GC state. @@ -731,7 +739,7 @@ minor_collection(Process* p, int need, Eterm* objv, int nobj, Uint *recl) * This improved Estone by more than 1200 estones on my computer * (Ultra Sparc 10). */ - size_t new_sz = erts_next_heap_size(HEAP_TOP(p) - HEAP_START(p), 1); + Uint new_sz = erts_next_heap_size(HEAP_TOP(p) - HEAP_START(p), 1); /* Create new, empty old_heap */ n_old = (Eterm *) ERTS_HEAP_ALLOC(ERTS_ALC_T_OLD_HEAP, @@ -871,12 +879,12 @@ minor_collection(Process* p, int need, Eterm* objv, int nobj, Uint *recl) #endif /* HIPE */ static void -do_minor(Process *p, int new_sz, Eterm* objv, int nobj) +do_minor(Process *p, Uint new_sz, Eterm* objv, int nobj) { Rootset rootset; /* Rootset for GC (stack, dictionary, etc). */ Roots* roots; Eterm* n_htop; - int n; + Uint n; Eterm* ptr; Eterm val; Eterm gval; @@ -1079,14 +1087,14 @@ major_collection(Process* p, int need, Eterm* objv, int nobj, Uint *recl) { Rootset rootset; Roots* roots; - int size_before; + Uint size_before; Eterm* n_heap; Eterm* n_htop; char* src = (char *) HEAP_START(p); Uint src_size = (char *) HEAP_TOP(p) - src; char* oh = (char *) OLD_HEAP(p); Uint oh_size = (char *) OLD_HTOP(p) - oh; - int n; + Uint n; Uint new_sz; Uint fragments = MBUF_SIZE(p) + combined_message_size(p); ErlMessage *msgp; @@ -1312,10 +1320,10 @@ major_collection(Process* p, int need, Eterm* objv, int nobj, Uint *recl) } static Uint -adjust_after_fullsweep(Process *p, int size_before, int need, Eterm *objv, int nobj) +adjust_after_fullsweep(Process *p, Uint size_before, int need, Eterm *objv, int nobj) { - int wanted, sz, size_after, need_after; - int stack_size = STACK_SZ_ON_HEAP(p); + Uint wanted, sz, size_after, need_after; + Uint stack_size = STACK_SZ_ON_HEAP(p); Uint reclaimed_now; size_after = (HEAP_TOP(p) - HEAP_START(p)); @@ -1915,8 +1923,8 @@ static void grow_new_heap(Process *p, Uint new_sz, Eterm* objv, int nobj) { Eterm* new_heap; - int heap_size = HEAP_TOP(p) - HEAP_START(p); - int stack_size = p->hend - p->stop; + Uint heap_size = HEAP_TOP(p) - HEAP_START(p); + Uint stack_size = p->hend - p->stop; Sint offs; ASSERT(HEAP_SIZE(p) < new_sz); @@ -1954,10 +1962,10 @@ static void shrink_new_heap(Process *p, Uint new_sz, Eterm *objv, int nobj) { Eterm* new_heap; - int heap_size = HEAP_TOP(p) - HEAP_START(p); + Uint heap_size = HEAP_TOP(p) - HEAP_START(p); Sint offs; - int stack_size = p->hend - p->stop; + Uint stack_size = p->hend - p->stop; ASSERT(new_sz < p->heap_sz); sys_memmove(p->heap + new_sz - stack_size, p->stop, stack_size * diff --git a/erts/emulator/beam/erl_goodfit_alloc.c b/erts/emulator/beam/erl_goodfit_alloc.c index 76b206d76f..8322b233ac 100644 --- a/erts/emulator/beam/erl_goodfit_alloc.c +++ b/erts/emulator/beam/erl_goodfit_alloc.c @@ -1,7 +1,7 @@ /* * %CopyrightBegin% * - * Copyright Ericsson AB 2003-2010. All Rights Reserved. + * Copyright Ericsson AB 2003-2011. All Rights Reserved. * * The contents of this file are subject to the Erlang Public License, * Version 1.1, (the "License"); you may not use this file except in @@ -163,10 +163,10 @@ BKT_MIN_SZ(GFAllctr_t *gfallctr, int ix) /* Prototypes of callback functions */ static Block_t * get_free_block (Allctr_t *, Uint, - Block_t *, Uint); -static void link_free_block (Allctr_t *, Block_t *); -static void unlink_free_block (Allctr_t *, Block_t *); -static void update_last_aux_mbc (Allctr_t *, Carrier_t *); + Block_t *, Uint, Uint32); +static void link_free_block (Allctr_t *, Block_t *, Uint32); +static void unlink_free_block (Allctr_t *, Block_t *, Uint32); +static void update_last_aux_mbc (Allctr_t *, Carrier_t *, Uint32); static Eterm info_options (Allctr_t *, char *, int *, void *, Uint **, Uint *); static void init_atoms (void); @@ -197,6 +197,8 @@ erts_gfalc_start(GFAllctr_t *gfallctr, is a struct). */ Allctr_t *allctr = (Allctr_t *) gfallctr; + init->sbmbct = 0; /* Small mbc not yet supported by goodfit */ + sys_memcpy((void *) gfallctr, (void *) &nulled_state, sizeof(GFAllctr_t)); allctr->mbc_header_size = sizeof(Carrier_t); @@ -379,7 +381,7 @@ search_bucket(Allctr_t *allctr, int ix, Uint size) static Block_t * get_free_block(Allctr_t *allctr, Uint size, - Block_t *cand_blk, Uint cand_size) + Block_t *cand_blk, Uint cand_size, Uint32 flags) { GFAllctr_t *gfallctr = (GFAllctr_t *) allctr; int unsafe_bi, min_bi; @@ -398,7 +400,7 @@ get_free_block(Allctr_t *allctr, Uint size, if (blk) { if (cand_blk && cand_size <= BLK_SZ(blk)) return NULL; /* cand_blk was better */ - unlink_free_block(allctr, blk); + unlink_free_block(allctr, blk, flags); return blk; } if (min_bi < NO_OF_BKTS - 1) { @@ -418,14 +420,14 @@ get_free_block(Allctr_t *allctr, Uint size, ASSERT(blk); if (cand_blk && cand_size <= BLK_SZ(blk)) return NULL; /* cand_blk was better */ - unlink_free_block(allctr, blk); + unlink_free_block(allctr, blk, flags); return blk; } static void -link_free_block(Allctr_t *allctr, Block_t *block) +link_free_block(Allctr_t *allctr, Block_t *block, Uint32 flags) { GFAllctr_t *gfallctr = (GFAllctr_t *) allctr; GFFreeBlock_t *blk = (GFFreeBlock_t *) block; @@ -446,7 +448,7 @@ link_free_block(Allctr_t *allctr, Block_t *block) } static void -unlink_free_block(Allctr_t *allctr, Block_t *block) +unlink_free_block(Allctr_t *allctr, Block_t *block, Uint32 flags) { GFAllctr_t *gfallctr = (GFAllctr_t *) allctr; GFFreeBlock_t *blk = (GFFreeBlock_t *) block; @@ -467,7 +469,7 @@ unlink_free_block(Allctr_t *allctr, Block_t *block) } static void -update_last_aux_mbc(Allctr_t *allctr, Carrier_t *mbc) +update_last_aux_mbc(Allctr_t *allctr, Carrier_t *mbc, Uint32 flags) { GFAllctr_t *gfallctr = (GFAllctr_t *) allctr; diff --git a/erts/emulator/beam/erl_init.c b/erts/emulator/beam/erl_init.c index 0a57eb6d88..8a297cded0 100644 --- a/erts/emulator/beam/erl_init.c +++ b/erts/emulator/beam/erl_init.c @@ -1,7 +1,7 @@ /* * %CopyrightBegin% * - * Copyright Ericsson AB 1997-2010. All Rights Reserved. + * Copyright Ericsson AB 1997-2011. All Rights Reserved. * * The contents of this file are subject to the Erlang Public License, * Version 1.1, (the "License"); you may not use this file except in @@ -69,7 +69,8 @@ static void erl_init(int ncpu); #define ERTS_MIN_COMPAT_REL 7 #ifdef ERTS_SMP -erts_smp_atomic_t erts_writing_erl_crash_dump; +erts_smp_atomic32_t erts_writing_erl_crash_dump; +erts_tsd_key_t erts_is_crash_dumping_key; #else volatile int erts_writing_erl_crash_dump = 0; #endif @@ -126,6 +127,8 @@ int erts_modified_timing_level; int erts_no_crash_dump = 0; /* Use -d to suppress crash dump. */ +int erts_no_line_info = 0; /* -L: Don't load line information */ + /* * Other global variables. */ @@ -323,7 +326,7 @@ init_shared_memory(int argc, char **argv) #endif global_gen_gcs = 0; - global_max_gen_gcs = (Uint16) erts_smp_atomic32_read(&erts_max_gen_gcs); + global_max_gen_gcs = (Uint16) erts_smp_atomic32_read_nob(&erts_max_gen_gcs); global_gc_flags = erts_default_process_flags; erts_global_offheap.mso = NULL; @@ -646,12 +649,14 @@ early_init(int *argc, char **argv) /* erts_lc_init(); #endif #ifdef ERTS_SMP - erts_smp_atomic_init(&erts_writing_erl_crash_dump, 0L); + erts_smp_atomic32_init_nob(&erts_writing_erl_crash_dump, 0L); + erts_tsd_key_create(&erts_is_crash_dumping_key); #else erts_writing_erl_crash_dump = 0; #endif - erts_smp_atomic32_init(&erts_max_gen_gcs, (erts_aint32_t) ((Uint16) -1)); + erts_smp_atomic32_init_nob(&erts_max_gen_gcs, + (erts_aint32_t) ((Uint16) -1)); erts_pre_init_process(); #if defined(USE_THREADS) && !defined(ERTS_SMP) @@ -803,10 +808,12 @@ early_init(int *argc, char **argv) /* #if defined(HIPE) hipe_signal_init(); /* must be done very early */ #endif - erl_sys_init(); erl_sys_args(argc, argv); + /* Creates threads on Windows that depend on the arguments, so has to be after erl_sys_args */ + erl_sys_init(); + erts_ets_realloc_always_moves = 0; erts_ets_always_compress = 0; erts_dist_buf_busy_limit = ERTS_DE_BUSY_LIMIT; @@ -856,7 +863,8 @@ erl_start(int argc, char **argv) envbufsz = sizeof(envbuf); if (erts_sys_getenv("ERL_FULLSWEEP_AFTER", envbuf, &envbufsz) == 0) { Uint16 max_gen_gcs = atoi(envbuf); - erts_smp_atomic32_set(&erts_max_gen_gcs, (erts_aint32_t) max_gen_gcs); + erts_smp_atomic32_set_nob(&erts_max_gen_gcs, + (erts_aint32_t) max_gen_gcs); } envbufsz = sizeof(envbuf); @@ -932,7 +940,9 @@ erl_start(int argc, char **argv) case 'l': display_loads++; break; - + case 'L': + erts_no_line_info = 1; + break; case 'v': #ifdef DEBUG if (argv[i][2] == '\0') { diff --git a/erts/emulator/beam/erl_instrument.c b/erts/emulator/beam/erl_instrument.c index f3f3c22933..963c8b3c58 100644 --- a/erts/emulator/beam/erl_instrument.c +++ b/erts/emulator/beam/erl_instrument.c @@ -1,7 +1,7 @@ /* * %CopyrightBegin% * - * Copyright Ericsson AB 2003-2010. All Rights Reserved. + * Copyright Ericsson AB 2003-2011. All Rights Reserved. * * The contents of this file are subject to the Erlang Public License, * Version 1.1, (the "License"); you may not use this file except in @@ -1186,6 +1186,8 @@ erts_instr_init(int stat, int map_stat) sys_memzero((void *) stats->n, sizeof(Stat_t)*(ERTS_ALC_N_MAX+1)); for (i = ERTS_ALC_A_MIN; i <= ERTS_ALC_A_MAX; i++) { + if (ERTS_IS_SBMBC_ALLOCATOR_NO__(i)) + continue; if (erts_allctrs_info[i].enabled) stats->ap[i] = &stats->a[i]; else @@ -1199,6 +1201,8 @@ erts_instr_init(int stat, int map_stat) erts_instr_memory_map = 1; erts_instr_stat = 1; for (i = ERTS_ALC_A_MIN; i <= ERTS_ALC_A_MAX; i++) { + if (ERTS_IS_SBMBC_ALLOCATOR_NO__(i)) + continue; erts_allctrs[i].alloc = map_stat_alloc; erts_allctrs[i].realloc = map_stat_realloc; erts_allctrs[i].free = map_stat_free; @@ -1209,6 +1213,8 @@ erts_instr_init(int stat, int map_stat) else { erts_instr_stat = 1; for (i = ERTS_ALC_A_MIN; i <= ERTS_ALC_A_MAX; i++) { + if (ERTS_IS_SBMBC_ALLOCATOR_NO__(i)) + continue; erts_allctrs[i].alloc = stat_alloc; erts_allctrs[i].realloc = stat_realloc; erts_allctrs[i].free = stat_free; diff --git a/erts/emulator/beam/erl_lock_check.c b/erts/emulator/beam/erl_lock_check.c index 9180508a49..587d82f2bb 100644 --- a/erts/emulator/beam/erl_lock_check.c +++ b/erts/emulator/beam/erl_lock_check.c @@ -153,6 +153,7 @@ static erts_lc_lock_order_t erts_lock_order[] = { { "instr", NULL }, { "fix_alloc", "index" }, { "alcu_allocator", "index" }, + { "sbmbc_alloc", "index" }, { "alcu_delayed_free", "index" }, { "mseg", NULL }, #if HALFWORD_HEAP diff --git a/erts/emulator/beam/erl_monitors.c b/erts/emulator/beam/erl_monitors.c index 9751b5d77c..47597f302b 100644 --- a/erts/emulator/beam/erl_monitors.c +++ b/erts/emulator/beam/erl_monitors.c @@ -125,7 +125,7 @@ static ErtsMonitor *create_monitor(Uint type, Eterm ref, Eterm pid, Eterm name) } else { n = (ErtsMonitor *) erts_alloc(ERTS_ALC_T_MONITOR_LH, mon_size*sizeof(Uint)); - erts_smp_atomic_add(&tot_link_lh_size, mon_size*sizeof(Uint)); + erts_smp_atomic_add_nob(&tot_link_lh_size, mon_size*sizeof(Uint)); } hp = n->heap; @@ -156,7 +156,7 @@ static ErtsLink *create_link(Uint type, Eterm pid) } else { n = (ErtsLink *) erts_alloc(ERTS_ALC_T_NLINK_LH, lnk_size*sizeof(Uint)); - erts_smp_atomic_add(&tot_link_lh_size, lnk_size*sizeof(Uint)); + erts_smp_atomic_add_nob(&tot_link_lh_size, lnk_size*sizeof(Uint)); } hp = n->heap; @@ -191,13 +191,13 @@ static ErtsSuspendMonitor *create_suspend_monitor(Eterm pid) void erts_init_monitors(void) { - erts_smp_atomic_init(&tot_link_lh_size, 0); + erts_smp_atomic_init_nob(&tot_link_lh_size, 0); } Uint erts_tot_link_lh_size(void) { - return (Uint) erts_smp_atomic_read(&tot_link_lh_size); + return (Uint) erts_smp_atomic_read_nob(&tot_link_lh_size); } void erts_destroy_monitor(ErtsMonitor *mon) @@ -222,7 +222,7 @@ void erts_destroy_monitor(ErtsMonitor *mon) erts_free(ERTS_ALC_T_MONITOR_SH, (void *) mon); } else { erts_free(ERTS_ALC_T_MONITOR_LH, (void *) mon); - erts_smp_atomic_add(&tot_link_lh_size, -1*mon_size*sizeof(Uint)); + erts_smp_atomic_add_nob(&tot_link_lh_size, -1*mon_size*sizeof(Uint)); } } @@ -244,7 +244,7 @@ void erts_destroy_link(ErtsLink *lnk) erts_free(ERTS_ALC_T_NLINK_SH, (void *) lnk); } else { erts_free(ERTS_ALC_T_NLINK_LH, (void *) lnk); - erts_smp_atomic_add(&tot_link_lh_size, -1*lnk_size*sizeof(Uint)); + erts_smp_atomic_add_nob(&tot_link_lh_size, -1*lnk_size*sizeof(Uint)); } } diff --git a/erts/emulator/beam/erl_nif.c b/erts/emulator/beam/erl_nif.c index 7ea0e1f6b2..51f1fad811 100644 --- a/erts/emulator/beam/erl_nif.c +++ b/erts/emulator/beam/erl_nif.c @@ -583,7 +583,15 @@ int enif_is_identical(Eterm lhs, Eterm rhs) int enif_compare(Eterm lhs, Eterm rhs) { - return CMP(lhs,rhs); + Sint result = CMP(lhs,rhs); + + if (result < 0) { + return -1; + } else if (result > 0) { + return 1; + } + + return result; } int enif_get_tuple(ErlNifEnv* env, Eterm tpl, int* arity, const Eterm** array) @@ -838,8 +846,11 @@ ERL_NIF_TERM enif_make_uint(ErlNifEnv* env, unsigned i) ERL_NIF_TERM enif_make_long(ErlNifEnv* env, long i) { + if (IS_SSMALL(i)) { + return make_small(i); + } #if SIZEOF_LONG == ERTS_SIZEOF_ETERM - return IS_SSMALL(i) ? make_small(i) : small_to_big(i, alloc_heap(env,2)); + return small_to_big(i, alloc_heap(env,2)); #elif SIZEOF_LONG == 8 ensure_heap(env,3); return erts_sint64_to_big(i, &env->hp); @@ -848,8 +859,11 @@ ERL_NIF_TERM enif_make_long(ErlNifEnv* env, long i) ERL_NIF_TERM enif_make_ulong(ErlNifEnv* env, unsigned long i) { + if (IS_USMALL(0,i)) { + return make_small(i); + } #if SIZEOF_LONG == ERTS_SIZEOF_ETERM - return IS_USMALL(0,i) ? make_small(i) : uint_to_big(i,alloc_heap(env,2)); + return uint_to_big(i,alloc_heap(env,2)); #elif SIZEOF_LONG == 8 ensure_heap(env,3); return erts_uint64_to_big(i, &env->hp); @@ -1012,6 +1026,29 @@ void enif_system_info(ErlNifSysInfo *sip, size_t si_size) driver_system_info(sip, si_size); } +int enif_make_reverse_list(ErlNifEnv* env, ERL_NIF_TERM term, ERL_NIF_TERM *list) { + Eterm *listptr, ret = NIL, *hp; + + if (is_nil(term)) { + *list = term; + return 1; + } + + ret = NIL; + + while (is_not_nil(term)) { + if (is_not_list(term)) { + return 0; + } + hp = alloc_heap(env, 2); + listptr = list_val(term); + ret = CONS(hp, CAR(listptr), ret); + term = CDR(listptr); + } + *list = ret; + return 1; +} + ErlNifMutex* enif_mutex_create(char *name) { return erl_drv_mutex_create(name); } void enif_mutex_destroy(ErlNifMutex *mtx) { erl_drv_mutex_destroy(mtx); } diff --git a/erts/emulator/beam/erl_nif.h b/erts/emulator/beam/erl_nif.h index d028567faf..fea527f954 100644 --- a/erts/emulator/beam/erl_nif.h +++ b/erts/emulator/beam/erl_nif.h @@ -32,9 +32,10 @@ ** 2.0: R14A ** 2.1: R14B02 "vm_variant" ** 2.2: R14B03 enif_is_exception +** 2.3: R15 enif_make_reverse_list */ #define ERL_NIF_MAJOR_VERSION 2 -#define ERL_NIF_MINOR_VERSION 2 +#define ERL_NIF_MINOR_VERSION 3 #include <stdlib.h> diff --git a/erts/emulator/beam/erl_nif_api_funcs.h b/erts/emulator/beam/erl_nif_api_funcs.h index 43d706dc72..18f2c022eb 100644 --- a/erts/emulator/beam/erl_nif_api_funcs.h +++ b/erts/emulator/beam/erl_nif_api_funcs.h @@ -136,6 +136,7 @@ ERL_NIF_API_FUNC_DECL(ERL_NIF_TERM,enif_make_int64,(ErlNifEnv*, ErlNifSInt64)); ERL_NIF_API_FUNC_DECL(ERL_NIF_TERM,enif_make_uint64,(ErlNifEnv*, ErlNifUInt64)); #endif ERL_NIF_API_FUNC_DECL(int,enif_is_exception,(ErlNifEnv*, ERL_NIF_TERM term)); +ERL_NIF_API_FUNC_DECL(int,enif_make_reverse_list,(ErlNifEnv*, ERL_NIF_TERM term, ERL_NIF_TERM *list)); ERL_NIF_API_FUNC_DECL(int,enif_is_number,(ErlNifEnv*, ERL_NIF_TERM term)); /* @@ -257,6 +258,7 @@ ERL_NIF_API_FUNC_DECL(int,enif_is_number,(ErlNifEnv*, ERL_NIF_TERM term)); #endif # define enif_is_exception ERL_NIF_API_FUNC_MACRO(enif_is_exception) +# define enif_make_reverse_list ERL_NIF_API_FUNC_MACRO(enif_make_reverse_list) # define enif_is_number ERL_NIF_API_FUNC_MACRO(enif_is_number) /* diff --git a/erts/emulator/beam/erl_node_tables.c b/erts/emulator/beam/erl_node_tables.c index 6daa127d23..af3873995e 100644 --- a/erts/emulator/beam/erl_node_tables.c +++ b/erts/emulator/beam/erl_node_tables.c @@ -1,7 +1,7 @@ /* * %CopyrightBegin% * - * Copyright Ericsson AB 2001-2010. All Rights Reserved. + * Copyright Ericsson AB 2001-2011. All Rights Reserved. * * The contents of this file are subject to the Erlang Public License, * Version 1.1, (the "License"); you may not use this file except in @@ -118,7 +118,7 @@ dist_table_alloc(void *dep_tmpl) dep->finalized_out_queue.first = NULL; dep->finalized_out_queue.last = NULL; - erts_smp_atomic_init(&dep->dist_cmd_scheduled, 0); + erts_smp_atomic_init_nob(&dep->dist_cmd_scheduled, 0); erts_port_task_handle_init(&dep->dist_cmd); dep->send = NULL; dep->cache = NULL; @@ -767,7 +767,7 @@ void erts_init_node_tables(void) erts_this_dist_entry->finalized_out_queue.first = NULL; erts_this_dist_entry->finalized_out_queue.last = NULL; - erts_smp_atomic_init(&erts_this_dist_entry->dist_cmd_scheduled, 0); + erts_smp_atomic_init_nob(&erts_this_dist_entry->dist_cmd_scheduled, 0); erts_port_task_handle_init(&erts_this_dist_entry->dist_cmd); erts_this_dist_entry->send = NULL; erts_this_dist_entry->cache = NULL; diff --git a/erts/emulator/beam/erl_port_task.c b/erts/emulator/beam/erl_port_task.c index e6b55c45e4..6aa5161b08 100644 --- a/erts/emulator/beam/erl_port_task.c +++ b/erts/emulator/beam/erl_port_task.c @@ -121,7 +121,7 @@ ERTS_SCHED_PREF_QUICK_ALLOC_IMPL(port_taskq, static ERTS_INLINE ErtsPortTask * handle2task(ErtsPortTaskHandle *pthp) { - return (ErtsPortTask *) erts_smp_atomic_read(pthp); + return (ErtsPortTask *) erts_smp_atomic_read_nob(pthp); } static ERTS_INLINE void @@ -129,7 +129,7 @@ reset_handle(ErtsPortTask *ptp) { if (ptp->handle) { ASSERT(ptp == handle2task(ptp->handle)); - erts_smp_atomic_set(ptp->handle, (erts_aint_t) NULL); + erts_smp_atomic_set_nob(ptp->handle, (erts_aint_t) NULL); } } @@ -138,7 +138,7 @@ set_handle(ErtsPortTask *ptp, ErtsPortTaskHandle *pthp) { ptp->handle = pthp; if (pthp) { - erts_smp_atomic_set(pthp, (erts_aint_t) ptp); + erts_smp_atomic_set_nob(pthp, (erts_aint_t) ptp); ASSERT(ptp == handle2task(ptp->handle)); } } @@ -479,8 +479,8 @@ erts_port_task_abort(Eterm id, ErtsPortTaskHandle *pthp) case ERTS_PORT_TASK_INPUT: case ERTS_PORT_TASK_OUTPUT: case ERTS_PORT_TASK_EVENT: - ASSERT(erts_smp_atomic_read(&erts_port_task_outstanding_io_tasks) > 0); - erts_smp_atomic_dec(&erts_port_task_outstanding_io_tasks); + ASSERT(erts_smp_atomic_read_nob(&erts_port_task_outstanding_io_tasks) > 0); + erts_smp_atomic_dec_relb(&erts_port_task_outstanding_io_tasks); break; default: break; @@ -568,7 +568,7 @@ erts_port_task_schedule(Eterm id, ErtsRunQueue *xrunq = erts_check_emigration_need(runq, ERTS_PORT_PRIO_LEVEL); if (xrunq) { /* Port emigrated ... */ - erts_smp_atomic_set(&pp->run_queue, (erts_aint_t) xrunq); + erts_smp_atomic_set_nob(&pp->run_queue, (erts_aint_t) xrunq); erts_smp_runq_unlock(runq); runq = xrunq; } @@ -594,7 +594,7 @@ erts_port_task_schedule(Eterm id, case ERTS_PORT_TASK_INPUT: case ERTS_PORT_TASK_OUTPUT: case ERTS_PORT_TASK_EVENT: - erts_smp_atomic_inc(&erts_port_task_outstanding_io_tasks); + erts_smp_atomic_inc_relb(&erts_port_task_outstanding_io_tasks); /* Fall through... */ default: enqueue_task(pp->sched.taskq, ptp); @@ -662,7 +662,7 @@ erts_port_task_free_port(Port *pp) pp->status |= ERTS_PORT_SFLG_FREE_SCHEDULED; erts_may_save_closed_port(pp); erts_smp_port_state_unlock(pp); - ERTS_SMP_LC_ASSERT(erts_smp_atomic_read(&pp->refc) > 1); + ERTS_SMP_LC_ASSERT(erts_smp_atomic_read_nob(&pp->refc) > 1); ptp->type = ERTS_PORT_TASK_FREE; ptp->event = (ErlDrvEvent) -1; ptp->event_data = NULL; @@ -684,9 +684,9 @@ erts_port_task_free_port(Port *pp) erts_may_save_closed_port(pp); erts_smp_port_state_unlock(pp); #ifdef ERTS_SMP - erts_smp_atomic_dec(&pp->refc); /* Not alive */ + erts_smp_atomic_dec_nob(&pp->refc); /* Not alive */ #endif - ERTS_SMP_LC_ASSERT(erts_smp_atomic_read(&pp->refc) > 0); /* Lock */ + ERTS_SMP_LC_ASSERT(erts_smp_atomic_read_nob(&pp->refc) > 0); /* Lock */ handle_remaining_tasks(runq, pp); /* May release runq lock */ ASSERT(!pp->sched.exe_taskq && (!ptqp || !ptqp->first)); pp->sched.taskq = NULL; @@ -724,7 +724,7 @@ resume_after_block(void *vd) ErtsPortTaskExeBlockData *d = (ErtsPortTaskExeBlockData *) vd; erts_smp_runq_lock(d->runq); if (d->resp) - *d->resp = (erts_smp_atomic_read(&erts_port_task_outstanding_io_tasks) + *d->resp = (erts_smp_atomic_read_nob(&erts_port_task_outstanding_io_tasks) != (erts_aint_t) 0); } @@ -832,8 +832,8 @@ erts_port_task_execute(ErtsRunQueue *runq, Port **curr_port_pp) ASSERT(!ptqp->first && (!pp->sched.taskq || !pp->sched.taskq->first)); #ifdef ERTS_SMP - erts_smp_atomic_dec(&pp->refc); /* Not alive */ - ERTS_SMP_LC_ASSERT(erts_smp_atomic_read(&pp->refc) > 0); /* Lock */ + erts_smp_atomic_dec_nob(&pp->refc); /* Not alive */ + ERTS_SMP_LC_ASSERT(erts_smp_atomic_read_nob(&pp->refc) > 0); /* Lock */ #else erts_port_status_bor_set(pp, ERTS_PORT_SFLG_FREE); #endif @@ -906,14 +906,16 @@ erts_port_task_execute(ErtsRunQueue *runq, Port **curr_port_pp) erts_unblock_fpe(fpe_was_unmasked); if (io_tasks_executed) { - ASSERT(erts_smp_atomic_read(&erts_port_task_outstanding_io_tasks) >= io_tasks_executed); - erts_smp_atomic_add(&erts_port_task_outstanding_io_tasks, -1*io_tasks_executed); + ASSERT(erts_smp_atomic_read_nob(&erts_port_task_outstanding_io_tasks) + >= io_tasks_executed); + erts_smp_atomic_add_relb(&erts_port_task_outstanding_io_tasks, + -1*io_tasks_executed); } *curr_port_pp = NULL; #ifdef ERTS_SMP - ASSERT(runq == (ErtsRunQueue *) erts_smp_atomic_read(&pp->run_queue)); + ASSERT(runq == (ErtsRunQueue *) erts_smp_atomic_read_nob(&pp->run_queue)); #endif if (!pp->sched.taskq) { @@ -940,7 +942,7 @@ erts_port_task_execute(ErtsRunQueue *runq, Port **curr_port_pp) } else { /* Port emigrated ... */ - erts_smp_atomic_set(&pp->run_queue, (erts_aint_t) xrunq); + erts_smp_atomic_set_nob(&pp->run_queue, (erts_aint_t) xrunq); enqueue_port(xrunq, pp); ASSERT(pp->sched.exe_taskq); pp->sched.exe_taskq = NULL; @@ -951,7 +953,7 @@ erts_port_task_execute(ErtsRunQueue *runq, Port **curr_port_pp) port_was_enqueued = 1; } - res = (erts_smp_atomic_read(&erts_port_task_outstanding_io_tasks) + res = (erts_smp_atomic_read_nob(&erts_port_task_outstanding_io_tasks) != (erts_aint_t) 0); ERTS_PT_CHK_PRES_PORTQ(runq, pp); @@ -972,13 +974,13 @@ erts_port_task_execute(ErtsRunQueue *runq, Port **curr_port_pp) { erts_aint_t refc; erts_smp_mtx_unlock(pp->lock); - refc = erts_smp_atomic_dectest(&pp->refc); + refc = erts_smp_atomic_dec_read_nob(&pp->refc); ASSERT(refc >= 0); if (refc == 0) { erts_smp_runq_unlock(runq); erts_port_cleanup(pp); /* Might aquire runq lock */ erts_smp_runq_lock(runq); - res = (erts_smp_atomic_read(&erts_port_task_outstanding_io_tasks) + res = (erts_smp_atomic_read_nob(&erts_port_task_outstanding_io_tasks) != (erts_aint_t) 0); } } @@ -1107,12 +1109,12 @@ erts_port_migrate(Port *prt, int *prt_locked, /* Refuse to migrate to a suspended run queue */ if (to_rq->flags & ERTS_RUNQ_FLG_SUSPENDED) return ERTS_MIGRATE_FAILED_RUNQ_SUSPENDED; - if (from_rq != (ErtsRunQueue *) erts_smp_atomic_read(&prt->run_queue)) + if (from_rq != (ErtsRunQueue *) erts_smp_atomic_read_nob(&prt->run_queue)) return ERTS_MIGRATE_FAILED_RUNQ_CHANGED; if (!ERTS_PORT_IS_IN_RUNQ(from_rq, prt)) return ERTS_MIGRATE_FAILED_NOT_IN_RUNQ; dequeue_port(from_rq, prt); - erts_smp_atomic_set(&prt->run_queue, (erts_aint_t) to_rq); + erts_smp_atomic_set_nob(&prt->run_queue, (erts_aint_t) to_rq); enqueue_port(to_rq, prt); return ERTS_MIGRATE_SUCCESS; } @@ -1125,7 +1127,8 @@ erts_port_migrate(Port *prt, int *prt_locked, void erts_port_task_init(void) { - erts_smp_atomic_init(&erts_port_task_outstanding_io_tasks, (erts_aint_t) 0); + erts_smp_atomic_init_nob(&erts_port_task_outstanding_io_tasks, + (erts_aint_t) 0); init_port_task_alloc(); init_port_taskq_alloc(); } diff --git a/erts/emulator/beam/erl_port_task.h b/erts/emulator/beam/erl_port_task.h index 3e2c5f07ab..d7104e1143 100644 --- a/erts/emulator/beam/erl_port_task.h +++ b/erts/emulator/beam/erl_port_task.h @@ -79,13 +79,13 @@ ERTS_GLB_INLINE int erts_port_task_have_outstanding_io_tasks(void); ERTS_GLB_INLINE void erts_port_task_handle_init(ErtsPortTaskHandle *pthp) { - erts_smp_atomic_init(pthp, (erts_aint_t) NULL); + erts_smp_atomic_init_nob(pthp, (erts_aint_t) NULL); } ERTS_GLB_INLINE int erts_port_task_is_scheduled(ErtsPortTaskHandle *pthp) { - return ((void *) erts_smp_atomic_read(pthp)) != NULL; + return ((void *) erts_smp_atomic_read_nob(pthp)) != NULL; } ERTS_GLB_INLINE void @@ -102,8 +102,8 @@ erts_port_task_init_sched(ErtsPortTaskSched *ptsp) ERTS_GLB_INLINE int erts_port_task_have_outstanding_io_tasks(void) { - ERTS_THR_MEMORY_BARRIER; - return erts_smp_atomic_read(&erts_port_task_outstanding_io_tasks) != 0; + return (erts_smp_atomic_read_acqb(&erts_port_task_outstanding_io_tasks) + != 0); } #endif /* ERTS_INCLUDE_SCHEDULER_INTERNALS */ diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c index 2704359a8f..5ceb4ce9a8 100644 --- a/erts/emulator/beam/erl_process.c +++ b/erts/emulator/beam/erl_process.c @@ -134,15 +134,15 @@ int erts_disable_proc_not_running_opt; #ifndef DEBUG #define ERTS_SCHDLR_SSPND_CHNG_SET(VAL, OLD_VAL) \ - erts_smp_atomic32_set(&schdlr_sspnd.changing, (VAL)) + erts_smp_atomic32_set_nob(&schdlr_sspnd.changing, (VAL)) #else #define ERTS_SCHDLR_SSPND_CHNG_SET(VAL, OLD_VAL) \ do { \ erts_aint32_t old_val__; \ - old_val__ = erts_smp_atomic32_xchg(&schdlr_sspnd.changing, \ - (VAL)); \ + old_val__ = erts_smp_atomic32_xchg_nob(&schdlr_sspnd.changing, \ + (VAL)); \ ASSERT(old_val__ == (OLD_VAL)); \ } while (0) @@ -158,7 +158,7 @@ static struct { erts_smp_atomic32_t changing; erts_smp_atomic32_t active; struct { - erts_smp_atomic32_t ongoing; + int ongoing; long wait_active; ErtsProcList *procs; } msb; /* Multi Scheduling Block */ @@ -410,7 +410,7 @@ erts_init_process(int ncpu) init_proclist_alloc(); - erts_smp_atomic32_init(&process_count, 0); + erts_smp_atomic32_init_nob(&process_count, 0); if (erts_use_r9_pids_ports) { proc_bits = ERTS_R9_PROC_BITS; @@ -694,8 +694,8 @@ erts_smp_schedule_misc_aux_work(int ignore_self, erts_smp_mtx_unlock(&misc_aux_work_queues[ix].data.mtx); ssi = ERTS_SCHED_SLEEP_INFO_IX(ix); - aux_work = erts_smp_atomic32_bor(&ssi->aux_work, - ERTS_SSI_AUX_WORK_MISC); + aux_work = erts_smp_atomic32_read_bor_nob(&ssi->aux_work, + ERTS_SSI_AUX_WORK_MISC); if ((aux_work & ERTS_SSI_AUX_WORK_MISC) == 0) erts_sched_poke(ssi); } @@ -711,8 +711,8 @@ erts_smp_notify_check_children_needed(void) erts_aint32_t aux_work; ErtsSchedulerSleepInfo *ssi; ssi = ERTS_SCHED_SLEEP_INFO_IX(i); - aux_work = erts_smp_atomic32_bor(&ssi->aux_work, - ERTS_SSI_AUX_WORK_CHECK_CHILDREN); + aux_work = erts_smp_atomic32_read_bor_nob(&ssi->aux_work, + ERTS_SSI_AUX_WORK_CHECK_CHILDREN); if (!(aux_work & ERTS_SSI_AUX_WORK_CHECK_CHILDREN)) erts_sched_poke(ssi); } @@ -727,15 +727,15 @@ blockable_aux_work(ErtsSchedulerData *esdp, { if (aux_work & ERTS_SSI_BLOCKABLE_AUX_WORK_MASK) { if (aux_work & ERTS_SSI_AUX_WORK_MISC) { - aux_work = erts_smp_atomic32_band(&ssi->aux_work, - ~ERTS_SSI_AUX_WORK_MISC); + aux_work = erts_smp_atomic32_read_band_nob(&ssi->aux_work, + ~ERTS_SSI_AUX_WORK_MISC); aux_work &= ~ERTS_SSI_AUX_WORK_MISC; handle_misc_aux_work(esdp); } #ifdef ERTS_SMP_SCHEDULERS_NEED_TO_CHECK_CHILDREN if (aux_work & ERTS_SSI_AUX_WORK_CHECK_CHILDREN) { - aux_work = erts_smp_atomic32_band(&ssi->aux_work, - ~ERTS_SSI_AUX_WORK_CHECK_CHILDREN); + aux_work = erts_smp_atomic32_band_nob(&ssi->aux_work, + ~ERTS_SSI_AUX_WORK_CHECK_CHILDREN); aux_work &= ~ERTS_SSI_AUX_WORK_CHECK_CHILDREN; erts_check_children(); } @@ -815,7 +815,7 @@ erts_active_schedulers(void) static ERTS_INLINE void clear_sys_scheduling(void) { - erts_smp_atomic32_set_relb(&doing_sys_schedule, 0); + erts_smp_atomic32_set_mb(&doing_sys_schedule, 0); } static ERTS_INLINE int @@ -882,42 +882,43 @@ sched_active(Uint no, ErtsRunQueue *rq) static int ERTS_INLINE ongoing_multi_scheduling_block(void) { - return erts_smp_atomic32_read(&schdlr_sspnd.msb.ongoing) != 0; + ERTS_SMP_LC_ASSERT(erts_lc_mtx_is_locked(&schdlr_sspnd.mtx)); + return schdlr_sspnd.msb.ongoing; } static ERTS_INLINE void empty_runq(ErtsRunQueue *rq) { - erts_aint32_t oifls = erts_smp_atomic32_band(&rq->info_flags, - ~ERTS_RUNQ_IFLG_NONEMPTY); + erts_aint32_t oifls = erts_smp_atomic32_read_band_nob(&rq->info_flags, + ~ERTS_RUNQ_IFLG_NONEMPTY); if (oifls & ERTS_RUNQ_IFLG_NONEMPTY) { #ifdef DEBUG - erts_aint32_t empty = erts_smp_atomic32_read(&no_empty_run_queues); + erts_aint32_t empty = erts_smp_atomic32_read_nob(&no_empty_run_queues); /* * For a short period of time no_empty_run_queues may have * been increased twice for a specific run queue. */ ASSERT(0 <= empty && empty < 2*erts_no_run_queues); #endif - erts_smp_atomic32_inc(&no_empty_run_queues); + erts_smp_atomic32_inc_relb(&no_empty_run_queues); } } static ERTS_INLINE void non_empty_runq(ErtsRunQueue *rq) { - erts_aint32_t oifls = erts_smp_atomic32_bor(&rq->info_flags, - ERTS_RUNQ_IFLG_NONEMPTY); + erts_aint32_t oifls = erts_smp_atomic32_read_bor_nob(&rq->info_flags, + ERTS_RUNQ_IFLG_NONEMPTY); if (!(oifls & ERTS_RUNQ_IFLG_NONEMPTY)) { #ifdef DEBUG - erts_aint32_t empty = erts_smp_atomic32_read(&no_empty_run_queues); + erts_aint32_t empty = erts_smp_atomic32_read_nob(&no_empty_run_queues); /* * For a short period of time no_empty_run_queues may have * been increased twice for a specific run queue. */ ASSERT(0 < empty && empty <= 2*erts_no_run_queues); #endif - erts_smp_atomic32_dec(&no_empty_run_queues); + erts_smp_atomic32_dec_relb(&no_empty_run_queues); } } @@ -930,7 +931,7 @@ sched_prep_spin_wait(ErtsSchedulerSleepInfo *ssi) erts_aint32_t xflgs = 0; do { - oflgs = erts_smp_atomic32_cmpxchg(&ssi->flags, nflgs, xflgs); + oflgs = erts_smp_atomic32_cmpxchg_acqb(&ssi->flags, nflgs, xflgs); if (oflgs == xflgs) return nflgs; xflgs = oflgs; @@ -947,7 +948,7 @@ sched_prep_cont_spin_wait(ErtsSchedulerSleepInfo *ssi) erts_aint32_t xflgs = ERTS_SSI_FLG_WAITING; do { - oflgs = erts_smp_atomic32_cmpxchg(&ssi->flags, nflgs, xflgs); + oflgs = erts_smp_atomic32_cmpxchg_acqb(&ssi->flags, nflgs, xflgs); if (oflgs == xflgs) return nflgs; xflgs = oflgs; @@ -989,7 +990,7 @@ sched_set_sleeptype(ErtsSchedulerSleepInfo *ssi, erts_aint32_t sleep_type) erts_tse_reset(ssi->event); while (1) { - oflgs = erts_smp_atomic32_cmpxchg(&ssi->flags, nflgs, xflgs); + oflgs = erts_smp_atomic32_cmpxchg_acqb(&ssi->flags, nflgs, xflgs); if (oflgs == xflgs) return nflgs; if ((oflgs & (ERTS_SSI_FLG_SLEEPING|ERTS_SSI_FLG_WAITING)) @@ -1049,7 +1050,7 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) tse_wait: #ifdef ERTS_SCHED_NEED_BLOCKABLE_AUX_WORK - aux_work = erts_smp_atomic32_read(&ssi->aux_work); + aux_work = erts_smp_atomic32_read_nob(&ssi->aux_work); tse_blockable_aux_work: aux_work = blockable_aux_work(esdp, ssi, aux_work); #endif @@ -1059,7 +1060,7 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) #ifdef ERTS_SCHED_NEED_NONBLOCKABLE_AUX_WORK #ifndef ERTS_SCHED_NEED_BLOCKABLE_AUX_WORK - aux_work = erts_smp_atomic32_read(&ssi->aux_work); + aux_work = erts_smp_atomic32_read_nob(&ssi->aux_work); #endif nonblockable_aux_work(esdp, ssi, aux_work); #endif @@ -1092,7 +1093,7 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) } #ifdef ERTS_SCHED_NEED_BLOCKABLE_AUX_WORK - aux_work = erts_smp_atomic32_read(&ssi->aux_work); + aux_work = erts_smp_atomic32_read_nob(&ssi->aux_work); if (aux_work & ERTS_SSI_BLOCKABLE_AUX_WORK_MASK) { erts_smp_activity_end(ERTS_ACTIVITY_WAIT, NULL, NULL, NULL); goto tse_blockable_aux_work; @@ -1104,7 +1105,7 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) erts_smp_activity_end(ERTS_ACTIVITY_WAIT, NULL, NULL, NULL); if (flgs & ~ERTS_SSI_FLG_SUSPENDED) - erts_smp_atomic32_band(&ssi->flags, ERTS_SSI_FLG_SUSPENDED); + erts_smp_atomic32_read_band_nob(&ssi->flags, ERTS_SSI_FLG_SUSPENDED); erts_smp_runq_lock(rq); sched_active(esdp->no, rq); @@ -1136,12 +1137,12 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) sys_aux_work: #ifdef ERTS_SCHED_NEED_BLOCKABLE_AUX_WORK - aux_work = erts_smp_atomic32_read(&ssi->aux_work); + aux_work = erts_smp_atomic32_read_nob(&ssi->aux_work); aux_work = blockable_aux_work(esdp, ssi, aux_work); #endif #ifdef ERTS_SCHED_NEED_NONBLOCKABLE_AUX_WORK #ifndef ERTS_SCHED_NEED_BLOCKABLE_AUX_WORK - aux_work = erts_smp_atomic32_read(&ssi->aux_work); + aux_work = erts_smp_atomic32_read_nob(&ssi->aux_work); #endif nonblockable_aux_work(esdp, ssi, aux_work); #endif @@ -1151,13 +1152,6 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) ASSERT(!(flgs & ERTS_SSI_FLG_SLEEPING)); goto sys_woken; } - if (!(flgs & ERTS_SSI_FLG_SLEEPING)) { - flgs = sched_prep_cont_spin_wait(ssi); - if (!(flgs & ERTS_SSI_FLG_WAITING)) { - ASSERT(!(flgs & ERTS_SSI_FLG_SLEEPING)); - goto sys_woken; - } - } /* * If we got new I/O tasks we aren't allowed to @@ -1239,7 +1233,7 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) sys_locked_woken: clear_sys_scheduling(); if (flgs & ~ERTS_SSI_FLG_SUSPENDED) - erts_smp_atomic32_band(&ssi->flags, ERTS_SSI_FLG_SUSPENDED); + erts_smp_atomic32_read_band_nob(&ssi->flags, ERTS_SSI_FLG_SUSPENDED); sched_active_sys(esdp->no, rq); } } @@ -1255,7 +1249,7 @@ ssi_flags_set_wake(ErtsSchedulerSleepInfo *ssi) erts_aint32_t nflgs = 0; erts_aint32_t xflgs = ERTS_SSI_FLG_SLEEPING|ERTS_SSI_FLG_WAITING; while (1) { - oflgs = erts_smp_atomic32_cmpxchg(&ssi->flags, nflgs, xflgs); + oflgs = erts_smp_atomic32_cmpxchg_relb(&ssi->flags, nflgs, xflgs); if (oflgs == xflgs) return oflgs; nflgs = oflgs & ERTS_SSI_FLG_SUSPENDED; @@ -1298,7 +1292,6 @@ wake_scheduler(ErtsRunQueue *rq, int incq, int one) erts_smp_spin_unlock(&sl->lock); - ERTS_THR_MEMORY_BARRIER; flgs = ssi_flags_set_wake(ssi); erts_sched_finish_poke(ssi, flgs); @@ -1344,13 +1337,13 @@ init_no_runqs(int active, int used) { erts_aint32_t no_runqs = (erts_aint32_t) (active & ERTS_NO_RUNQS_MASK); no_runqs |= (erts_aint32_t) ((used & ERTS_NO_RUNQS_MASK) << ERTS_NO_USED_RUNQS_SHIFT); - erts_smp_atomic32_init(&balance_info.no_runqs, no_runqs); + erts_smp_atomic32_init_nob(&balance_info.no_runqs, no_runqs); } static ERTS_INLINE void get_no_runqs(int *active, int *used) { - erts_aint32_t no_runqs = erts_smp_atomic32_read(&balance_info.no_runqs); + erts_aint32_t no_runqs = erts_smp_atomic32_read_nob(&balance_info.no_runqs); if (active) *active = (int) (no_runqs & ERTS_NO_RUNQS_MASK); if (used) @@ -1360,11 +1353,12 @@ get_no_runqs(int *active, int *used) static ERTS_INLINE void set_no_used_runqs(int used) { - erts_aint32_t exp = erts_smp_atomic32_read(&balance_info.no_runqs); + erts_aint32_t exp = erts_smp_atomic32_read_nob(&balance_info.no_runqs); while (1) { erts_aint32_t act, new; - new = (used << ERTS_NO_USED_RUNQS_SHIFT) | (exp & ERTS_NO_RUNQS_MASK); - act = erts_smp_atomic32_cmpxchg(&balance_info.no_runqs, new, exp); + new = (used & ERTS_NO_RUNQS_MASK) << ERTS_NO_USED_RUNQS_SHIFT; + new |= exp & ERTS_NO_RUNQS_MASK; + act = erts_smp_atomic32_cmpxchg_nob(&balance_info.no_runqs, new, exp); if (act == exp) break; exp = act; @@ -1374,11 +1368,12 @@ set_no_used_runqs(int used) static ERTS_INLINE void set_no_active_runqs(int active) { - erts_aint32_t exp = erts_smp_atomic32_read(&balance_info.no_runqs); + erts_aint32_t exp = erts_smp_atomic32_read_nob(&balance_info.no_runqs); while (1) { erts_aint32_t act, new; - new = (exp & (ERTS_NO_RUNQS_MASK << ERTS_NO_USED_RUNQS_SHIFT)) | active; - act = erts_smp_atomic32_cmpxchg(&balance_info.no_runqs, new, exp); + new = exp & (ERTS_NO_RUNQS_MASK << ERTS_NO_USED_RUNQS_SHIFT); + new |= active & ERTS_NO_RUNQS_MASK; + act = erts_smp_atomic32_cmpxchg_nob(&balance_info.no_runqs, new, exp); if (act == exp) break; exp = act; @@ -1388,13 +1383,14 @@ set_no_active_runqs(int active) static ERTS_INLINE int try_inc_no_active_runqs(int active) { - erts_aint32_t exp = erts_smp_atomic32_read(&balance_info.no_runqs); + erts_aint32_t exp = erts_smp_atomic32_read_nob(&balance_info.no_runqs); if (((exp >> ERTS_NO_USED_RUNQS_SHIFT) & ERTS_NO_RUNQS_MASK) < active) return 0; if ((exp & ERTS_NO_RUNQS_MASK) + 1 == active) { erts_aint32_t new, act; - new = (exp & ~ERTS_NO_RUNQS_MASK) | active; - act = erts_smp_atomic32_cmpxchg(&balance_info.no_runqs, new, exp); + new = exp & (ERTS_NO_RUNQS_MASK << ERTS_NO_USED_RUNQS_SHIFT); + new |= active & ERTS_NO_RUNQS_MASK; + act = erts_smp_atomic32_cmpxchg_nob(&balance_info.no_runqs, new, exp); if (act == exp) return 1; } @@ -1410,7 +1406,7 @@ chk_wake_sched(ErtsRunQueue *crq, int ix, int activate) if (crq->ix == ix) return 0; wrq = ERTS_RUNQ_IX(ix); - iflgs = erts_smp_atomic32_read(&wrq->info_flags); + iflgs = erts_smp_atomic32_read_nob(&wrq->info_flags); if (!(iflgs & (ERTS_RUNQ_IFLG_SUSPENDED|ERTS_RUNQ_IFLG_NONEMPTY))) { if (activate) { if (try_inc_no_active_runqs(ix+1)) { @@ -1652,15 +1648,15 @@ evacuate_run_queue(ErtsRunQueue *evac_rq, ErtsRunQueue *rq) erts_smp_runq_lock(evac_rq); - erts_smp_atomic32_bor(&evac_rq->scheduler->ssi->flags, - ERTS_SSI_FLG_SUSPENDED); + erts_smp_atomic32_read_bor_nob(&evac_rq->scheduler->ssi->flags, + ERTS_SSI_FLG_SUSPENDED); evac_rq->flags &= ~ERTS_RUNQ_FLGS_IMMIGRATE_QMASK; evac_rq->flags |= (ERTS_RUNQ_FLGS_EMIGRATE_QMASK | ERTS_RUNQ_FLGS_EVACUATE_QMASK | ERTS_RUNQ_FLG_SUSPENDED); - erts_smp_atomic32_bor(&evac_rq->info_flags, ERTS_RUNQ_IFLG_SUSPENDED); + erts_smp_atomic32_read_bor_nob(&evac_rq->info_flags, ERTS_RUNQ_IFLG_SUSPENDED); /* * Need to set up evacuation paths first since we * may release the run queue lock on evac_rq @@ -1909,7 +1905,7 @@ static ERTS_INLINE int check_possible_steal_victim(ErtsRunQueue *rq, int *rq_lockedp, int vix) { ErtsRunQueue *vrq = ERTS_RUNQ_IX(vix); - erts_aint32_t iflgs = erts_smp_atomic32_read(&vrq->info_flags); + erts_aint32_t iflgs = erts_smp_atomic32_read_nob(&vrq->info_flags); if (iflgs & ERTS_RUNQ_IFLG_NONEMPTY) return try_steal_task_from_victim(rq, rq_lockedp, vrq); else @@ -2061,7 +2057,7 @@ check_balance(ErtsRunQueue *c_rq) int forced, active, current_active, oowc, half_full_scheds, full_scheds, mmax_len, blnc_no_rqs, qix, pix, freds_hist_ix; - if (erts_smp_atomic32_xchg(&balance_info.checking_balance, 1)) { + if (erts_smp_atomic32_xchg_nob(&balance_info.checking_balance, 1)) { c_rq->check_balance_reds = INT_MAX; return; } @@ -2069,7 +2065,7 @@ check_balance(ErtsRunQueue *c_rq) get_no_runqs(NULL, &blnc_no_rqs); if (blnc_no_rqs == 1) { c_rq->check_balance_reds = INT_MAX; - erts_smp_atomic32_set(&balance_info.checking_balance, 0); + erts_smp_atomic32_set_nob(&balance_info.checking_balance, 0); return; } @@ -2077,7 +2073,7 @@ check_balance(ErtsRunQueue *c_rq) if (balance_info.halftime) { balance_info.halftime = 0; - erts_smp_atomic32_set(&balance_info.checking_balance, 0); + erts_smp_atomic32_set_nob(&balance_info.checking_balance, 0); ERTS_FOREACH_RUNQ(rq, { if (rq->waiting) @@ -2111,7 +2107,7 @@ check_balance(ErtsRunQueue *c_rq) erts_smp_mtx_unlock(&balance_info.update_mtx); erts_smp_runq_lock(c_rq); c_rq->check_balance_reds = INT_MAX; - erts_smp_atomic32_set(&balance_info.checking_balance, 0); + erts_smp_atomic32_set_nob(&balance_info.checking_balance, 0); return; } @@ -2456,7 +2452,7 @@ erts_fprintf(stderr, "--------------------------------\n"); set_no_active_runqs(active); balance_info.halftime = 1; - erts_smp_atomic32_set(&balance_info.checking_balance, 0); + erts_smp_atomic32_set_nob(&balance_info.checking_balance, 0); /* Write migration paths and reset balance statistics in all queues */ for (qix = 0; qix < blnc_no_rqs; qix++) { @@ -2598,7 +2594,7 @@ erts_init_scheduling(int mrq, int no_schedulers, int no_schedulers_online) erts_alloc_permanent_cache_aligned(ERTS_ALC_T_RUNQS, sizeof(ErtsAlignedRunQueue) * n); #ifdef ERTS_SMP - erts_smp_atomic32_init(&no_empty_run_queues, 0); + erts_smp_atomic32_init_nob(&no_empty_run_queues, 0); #endif erts_no_run_queues = n; @@ -2608,7 +2604,7 @@ erts_init_scheduling(int mrq, int no_schedulers, int no_schedulers_online) ErtsRunQueue *rq = ERTS_RUNQ_IX(ix); rq->ix = ix; - erts_smp_atomic32_init(&rq->info_flags, ERTS_RUNQ_IFLG_NONEMPTY); + erts_smp_atomic32_init_nob(&rq->info_flags, ERTS_RUNQ_IFLG_NONEMPTY); /* make sure that the "extra" id correponds to the schedulers * id if the esdp->no <-> ix+1 mapping change. @@ -2701,9 +2697,9 @@ erts_init_scheduling(int mrq, int no_schedulers, int no_schedulers_online) ssi->next = NULL; ssi->prev = NULL; #endif - erts_smp_atomic32_init(&ssi->flags, 0); + erts_smp_atomic32_init_nob(&ssi->flags, 0); ssi->event = NULL; /* initialized in sched_thread_func */ - erts_smp_atomic32_init(&ssi->aux_work, 0); + erts_smp_atomic32_init_nob(&ssi->aux_work, 0); } #endif @@ -2720,11 +2716,14 @@ erts_init_scheduling(int mrq, int no_schedulers, int no_schedulers_online) esdp->match_pseudo_process = NULL; esdp->ssi = ERTS_SCHED_SLEEP_INFO_IX(ix); esdp->free_process = NULL; -#if HALFWORD_HEAP - /* Registers need to be heap allocated (correct memory range) for tracing to work */ - esdp->save_reg = erts_alloc(ERTS_ALC_T_BEAM_REGISTER, ERTS_X_REGS_ALLOCATED * sizeof(Eterm)); -#endif #endif + esdp->x_reg_array = + erts_alloc_permanent_cache_aligned(ERTS_ALC_T_BEAM_REGISTER, + ERTS_X_REGS_ALLOCATED * + sizeof(Eterm)); + esdp->f_reg_array = + erts_alloc_permanent_cache_aligned(ERTS_ALC_T_BEAM_REGISTER, + MAX_REG * sizeof(FloatDef)); #if !HEAP_ON_C_STACK esdp->num_tmp_heap_used = 0; #endif @@ -2747,7 +2746,7 @@ erts_init_scheduling(int mrq, int no_schedulers, int no_schedulers_online) } #ifdef ERTS_SMP - erts_smp_atomic32_init(&esdp->chk_cpu_bind, 0); + erts_smp_atomic32_init_nob(&esdp->chk_cpu_bind, 0); #endif } @@ -2755,11 +2754,11 @@ erts_init_scheduling(int mrq, int no_schedulers, int no_schedulers_online) erts_smp_mtx_init(&schdlr_sspnd.mtx, "schdlr_sspnd"); erts_smp_cnd_init(&schdlr_sspnd.cnd); - erts_smp_atomic32_init(&schdlr_sspnd.changing, 0); + erts_smp_atomic32_init_nob(&schdlr_sspnd.changing, 0); schdlr_sspnd.online = no_schedulers_online; schdlr_sspnd.curr_online = no_schedulers; - erts_smp_atomic32_init(&schdlr_sspnd.msb.ongoing, 0); - erts_smp_atomic32_init(&schdlr_sspnd.active, no_schedulers); + schdlr_sspnd.msb.ongoing = 0; + erts_smp_atomic32_init_nob(&schdlr_sspnd.active, no_schedulers); schdlr_sspnd.msb.procs = NULL; init_no_runqs(no_schedulers, erts_common_run_queue ? 1 : no_schedulers_online); @@ -2768,7 +2767,7 @@ erts_init_scheduling(int mrq, int no_schedulers, int no_schedulers_online) balance_info.forced_check_balance = 0; balance_info.halftime = 1; balance_info.full_reds_history_index = 0; - erts_smp_atomic32_init(&balance_info.checking_balance, 0); + erts_smp_atomic32_init_nob(&balance_info.checking_balance, 0); balance_info.prev_rise.active_runqs = 0; balance_info.prev_rise.max_len = 0; balance_info.prev_rise.reds = 0; @@ -2777,8 +2776,8 @@ erts_init_scheduling(int mrq, int no_schedulers, int no_schedulers_online) if (no_schedulers_online < no_schedulers) { if (erts_common_run_queue) { for (ix = no_schedulers_online; ix < no_schedulers; ix++) - erts_smp_atomic32_bor(&ERTS_SCHED_SLEEP_INFO_IX(ix)->flags, - ERTS_SSI_FLG_SUSPENDED); + erts_smp_atomic32_read_bor_nob(&ERTS_SCHED_SLEEP_INFO_IX(ix)->flags, + ERTS_SSI_FLG_SUSPENDED); } else { for (ix = no_schedulers_online; ix < erts_no_run_queues; ix++) @@ -2792,7 +2791,7 @@ erts_init_scheduling(int mrq, int no_schedulers, int no_schedulers_online) ERTS_SCHDLR_SSPND_CHNG_SET((ERTS_SCHDLR_SSPND_CHNG_ONLN | ERTS_SCHDLR_SSPND_CHNG_WAITER), 0); - erts_smp_atomic32_init(&doing_sys_schedule, 0); + erts_smp_atomic32_init_nob(&doing_sys_schedule, 0); init_misc_aux_work(); @@ -2808,7 +2807,7 @@ erts_init_scheduling(int mrq, int no_schedulers, int no_schedulers_online) erts_no_schedulers = 1; #endif - erts_smp_atomic32_init(&function_calls, 0); + erts_smp_atomic32_init_nob(&function_calls, 0); /* init port tasks */ erts_port_task_init(); @@ -2935,10 +2934,10 @@ int erts_get_max_no_executing_schedulers(void) { #ifdef ERTS_SMP - if (erts_smp_atomic32_read(&schdlr_sspnd.changing)) + if (erts_smp_atomic32_read_nob(&schdlr_sspnd.changing)) return (int) erts_no_schedulers; ERTS_THR_MEMORY_BARRIER; - return (int) erts_smp_atomic32_read(&schdlr_sspnd.active); + return (int) erts_smp_atomic32_read_nob(&schdlr_sspnd.active); #else return 1; #endif @@ -2968,7 +2967,7 @@ scheduler_ix_resume_wake(Uint ix) | ERTS_SSI_FLG_SUSPENDED); erts_aint32_t oflgs; do { - oflgs = erts_smp_atomic32_cmpxchg(&ssi->flags, 0, xflgs); + oflgs = erts_smp_atomic32_cmpxchg_relb(&ssi->flags, 0, xflgs); if (oflgs == xflgs) { erts_sched_finish_poke(ssi, oflgs); break; @@ -2987,7 +2986,7 @@ sched_prep_spin_suspended(ErtsSchedulerSleepInfo *ssi, erts_aint32_t xpct) erts_aint32_t xflgs = xpct; do { - oflgs = erts_smp_atomic32_cmpxchg(&ssi->flags, nflgs, xflgs); + oflgs = erts_smp_atomic32_cmpxchg_acqb(&ssi->flags, nflgs, xflgs); if (oflgs == xflgs) return nflgs; xflgs = oflgs; @@ -3037,7 +3036,7 @@ sched_set_suspended_sleeptype(ErtsSchedulerSleepInfo *ssi) erts_tse_reset(ssi->event); while (1) { - oflgs = erts_smp_atomic32_cmpxchg(&ssi->flags, nflgs, xflgs); + oflgs = erts_smp_atomic32_cmpxchg_acqb(&ssi->flags, nflgs, xflgs); if (oflgs == xflgs) return nflgs; if ((oflgs & (ERTS_SSI_FLG_SLEEPING @@ -3092,15 +3091,15 @@ suspend_scheduler(ErtsSchedulerData *esdp) flgs = sched_prep_spin_suspended(ssi, ERTS_SSI_FLG_SUSPENDED); if (flgs & ERTS_SSI_FLG_SUSPENDED) { - active_schedulers = erts_smp_atomic32_dectest(&schdlr_sspnd.active); + active_schedulers = erts_smp_atomic32_dec_read_nob(&schdlr_sspnd.active); ASSERT(active_schedulers >= 1); - changing = erts_smp_atomic32_read(&schdlr_sspnd.changing); + changing = erts_smp_atomic32_read_nob(&schdlr_sspnd.changing); if (changing & ERTS_SCHDLR_SSPND_CHNG_MSB) { if (active_schedulers == schdlr_sspnd.msb.wait_active) wake = 1; if (active_schedulers == 1) { - changing = erts_smp_atomic32_band(&schdlr_sspnd.changing, - ~ERTS_SCHDLR_SSPND_CHNG_MSB); + changing = erts_smp_atomic32_read_band_nob(&schdlr_sspnd.changing, + ~ERTS_SCHDLR_SSPND_CHNG_MSB); changing &= ~ERTS_SCHDLR_SSPND_CHNG_MSB; } } @@ -3122,8 +3121,8 @@ suspend_scheduler(ErtsSchedulerData *esdp) && schdlr_sspnd.curr_online == schdlr_sspnd.wait_curr_online) wake = 1; if (schdlr_sspnd.online == schdlr_sspnd.curr_online) { - changing = erts_smp_atomic32_band(&schdlr_sspnd.changing, - ~ERTS_SCHDLR_SSPND_CHNG_ONLN); + changing = erts_smp_atomic32_read_band_nob(&schdlr_sspnd.changing, + ~ERTS_SCHDLR_SSPND_CHNG_ONLN); changing &= ~ERTS_SCHDLR_SSPND_CHNG_ONLN; } } @@ -3140,7 +3139,7 @@ suspend_scheduler(ErtsSchedulerData *esdp) #ifdef ERTS_SCHED_NEED_BLOCKABLE_AUX_WORK - aux_work = erts_smp_atomic32_read(&ssi->aux_work); + aux_work = erts_smp_atomic32_read_nob(&ssi->aux_work); blockable_aux_work: blockable_aux_work(esdp, ssi, aux_work); #endif @@ -3176,13 +3175,13 @@ suspend_scheduler(ErtsSchedulerData *esdp) | ERTS_SSI_FLG_SUSPENDED)); if (!(flgs & ERTS_SSI_FLG_SUSPENDED)) break; - changing = erts_smp_atomic32_read(&schdlr_sspnd.changing); + changing = erts_smp_atomic32_read_nob(&schdlr_sspnd.changing); if (changing & ~ERTS_SCHDLR_SSPND_CHNG_WAITER) break; #ifdef ERTS_SCHED_NEED_BLOCKABLE_AUX_WORK - aux_work = erts_smp_atomic32_read(&ssi->aux_work); + aux_work = erts_smp_atomic32_read_nob(&ssi->aux_work); if (aux_work & ERTS_SSI_BLOCKABLE_AUX_WORK_MASK) { erts_smp_activity_end(ERTS_ACTIVITY_WAIT, NULL, NULL, NULL); goto blockable_aux_work; @@ -3194,19 +3193,19 @@ suspend_scheduler(ErtsSchedulerData *esdp) erts_smp_activity_end(ERTS_ACTIVITY_WAIT, NULL, NULL, NULL); erts_smp_mtx_lock(&schdlr_sspnd.mtx); - changing = erts_smp_atomic32_read(&schdlr_sspnd.changing); + changing = erts_smp_atomic32_read_nob(&schdlr_sspnd.changing); } - active_schedulers = erts_smp_atomic32_inctest(&schdlr_sspnd.active); - changing = erts_smp_atomic32_read(&schdlr_sspnd.changing); + active_schedulers = erts_smp_atomic32_inc_read_nob(&schdlr_sspnd.active); + changing = erts_smp_atomic32_read_nob(&schdlr_sspnd.changing); if ((changing & ERTS_SCHDLR_SSPND_CHNG_MSB) && schdlr_sspnd.online == active_schedulers) { - erts_smp_atomic32_band(&schdlr_sspnd.changing, - ~ERTS_SCHDLR_SSPND_CHNG_MSB); + erts_smp_atomic32_read_band_nob(&schdlr_sspnd.changing, + ~ERTS_SCHDLR_SSPND_CHNG_MSB); } ASSERT(no <= schdlr_sspnd.online); - ASSERT(!erts_smp_atomic32_read(&schdlr_sspnd.msb.ongoing)); + ASSERT(!ongoing_multi_scheduling_block()); } @@ -3235,7 +3234,7 @@ do { \ (RQ)->flags |= (ERTS_RUNQ_FLG_OUT_OF_WORK \ | ERTS_RUNQ_FLG_HALFTIME_OUT_OF_WORK); \ (RQ)->check_balance_reds = ERTS_RUNQ_CALL_CHECK_BALANCE_REDS; \ - erts_smp_atomic32_band(&(RQ)->info_flags, ~ERTS_RUNQ_IFLG_SUSPENDED);\ + erts_smp_atomic32_read_band_nob(&(RQ)->info_flags, ~ERTS_RUNQ_IFLG_SUSPENDED);\ for (pix__ = 0; pix__ < ERTS_NO_PROC_PRIO_LEVELS; pix__++) { \ (RQ)->procs.prio_info[pix__].max_len = 0; \ (RQ)->procs.prio_info[pix__].reds = 0; \ @@ -3279,7 +3278,7 @@ erts_schedulers_state(Uint *total, int res; erts_aint32_t changing; erts_smp_mtx_lock(&schdlr_sspnd.mtx); - changing = erts_smp_atomic32_read(&schdlr_sspnd.changing); + changing = erts_smp_atomic32_read_nob(&schdlr_sspnd.changing); if (yield_allowed && (changing & ~ERTS_SCHDLR_SSPND_CHNG_WAITER)) res = ERTS_SCHDLR_SSPND_YIELD_RESTART; else { @@ -3310,7 +3309,7 @@ erts_set_schedulers_online(Process *p, have_unlocked_plocks = 0; no = (int) new_no; - changing = erts_smp_atomic32_read(&schdlr_sspnd.changing); + changing = erts_smp_atomic32_read_nob(&schdlr_sspnd.changing); if (changing) { res = ERTS_SCHDLR_SSPND_YIELD_RESTART; } @@ -3385,8 +3384,8 @@ erts_set_schedulers_online(Process *p, for (ix = no; ix < online; ix++) { ErtsSchedulerSleepInfo *ssi; ssi = ERTS_SCHED_SLEEP_INFO_IX(ix); - erts_smp_atomic32_bor(&ssi->flags, - ERTS_SSI_FLG_SUSPENDED); + erts_smp_atomic32_read_bor_nob(&ssi->flags, + ERTS_SSI_FLG_SUSPENDED); } wake_all_schedulers(); } @@ -3433,11 +3432,11 @@ erts_set_schedulers_online(Process *p, NULL); ASSERT(res != ERTS_SCHDLR_SSPND_DONE ? (ERTS_SCHDLR_SSPND_CHNG_WAITER - & erts_smp_atomic32_read(&schdlr_sspnd.changing)) + & erts_smp_atomic32_read_nob(&schdlr_sspnd.changing)) : (ERTS_SCHDLR_SSPND_CHNG_WAITER - == erts_smp_atomic32_read(&schdlr_sspnd.changing))); - erts_smp_atomic32_band(&schdlr_sspnd.changing, - ~ERTS_SCHDLR_SSPND_CHNG_WAITER); + == erts_smp_atomic32_read_nob(&schdlr_sspnd.changing))); + erts_smp_atomic32_read_band_nob(&schdlr_sspnd.changing, + ~ERTS_SCHDLR_SSPND_CHNG_WAITER); } } @@ -3456,7 +3455,7 @@ erts_block_multi_scheduling(Process *p, ErtsProcLocks plocks, int on, int all) ErtsProcList *plp; erts_smp_mtx_lock(&schdlr_sspnd.mtx); - changing = erts_smp_atomic32_read(&schdlr_sspnd.changing); + changing = erts_smp_atomic32_read_nob(&schdlr_sspnd.changing); if (changing) { res = ERTS_SCHDLR_SSPND_YIELD_RESTART; /* Yield */ } @@ -3466,7 +3465,7 @@ erts_block_multi_scheduling(Process *p, ErtsProcLocks plocks, int on, int all) plp->next = schdlr_sspnd.msb.procs; schdlr_sspnd.msb.procs = plp; p->flags |= F_HAVE_BLCKD_MSCHED; - ASSERT(erts_smp_atomic32_read(&schdlr_sspnd.active) == 1); + ASSERT(erts_smp_atomic32_read_nob(&schdlr_sspnd.active) == 1); ASSERT(p->scheduler_data->no == 1); res = ERTS_SCHDLR_SSPND_DONE_MSCHED_BLOCKED; } @@ -3477,11 +3476,11 @@ erts_block_multi_scheduling(Process *p, ErtsProcLocks plocks, int on, int all) have_unlocked_plocks = 1; erts_smp_proc_unlock(p, plocks); } - ASSERT(0 == erts_smp_atomic32_read(&schdlr_sspnd.msb.ongoing)); - erts_smp_atomic32_set(&schdlr_sspnd.msb.ongoing, 1); + ASSERT(!ongoing_multi_scheduling_block()); + schdlr_sspnd.msb.ongoing = 1; if (online == 1) { res = ERTS_SCHDLR_SSPND_DONE_MSCHED_BLOCKED; - ASSERT(erts_smp_atomic32_read(&schdlr_sspnd.active) == 1); + ASSERT(erts_smp_atomic32_read_nob(&schdlr_sspnd.active) == 1); ASSERT(p->scheduler_data->no == 1); } else { @@ -3501,8 +3500,8 @@ erts_block_multi_scheduling(Process *p, ErtsProcLocks plocks, int on, int all) } if (erts_common_run_queue) { for (ix = 1; ix < online; ix++) - erts_smp_atomic32_bor(&ERTS_SCHED_SLEEP_INFO_IX(ix)->flags, - ERTS_SSI_FLG_SUSPENDED); + erts_smp_atomic32_read_bor_nob(&ERTS_SCHED_SLEEP_INFO_IX(ix)->flags, + ERTS_SSI_FLG_SUSPENDED); wake_all_schedulers(); } else { @@ -3530,7 +3529,7 @@ erts_block_multi_scheduling(Process *p, ErtsProcLocks plocks, int on, int all) susp_sched_prep_block, susp_sched_resume_block, NULL); - while (erts_smp_atomic32_read(&schdlr_sspnd.active) + while (erts_smp_atomic32_read_nob(&schdlr_sspnd.active) != schdlr_sspnd.msb.wait_active) erts_smp_cnd_wait(&schdlr_sspnd.cnd, &schdlr_sspnd.mtx); erts_smp_activity_end(ERTS_ACTIVITY_WAIT, @@ -3539,11 +3538,11 @@ erts_block_multi_scheduling(Process *p, ErtsProcLocks plocks, int on, int all) NULL); ASSERT(res != ERTS_SCHDLR_SSPND_DONE_MSCHED_BLOCKED ? (ERTS_SCHDLR_SSPND_CHNG_WAITER - & erts_smp_atomic32_read(&schdlr_sspnd.changing)) + & erts_smp_atomic32_read_nob(&schdlr_sspnd.changing)) : (ERTS_SCHDLR_SSPND_CHNG_WAITER - == erts_smp_atomic32_read(&schdlr_sspnd.changing))); - erts_smp_atomic32_band(&schdlr_sspnd.changing, - ~ERTS_SCHDLR_SSPND_CHNG_WAITER); + == erts_smp_atomic32_read_nob(&schdlr_sspnd.changing))); + erts_smp_atomic32_read_band_nob(&schdlr_sspnd.changing, + ~ERTS_SCHDLR_SSPND_CHNG_WAITER); } plp = proclist_create(p); plp->next = schdlr_sspnd.msb.procs; @@ -3610,16 +3609,16 @@ erts_block_multi_scheduling(Process *p, ErtsProcLocks plocks, int on, int all) }); #endif p->flags &= ~F_HAVE_BLCKD_MSCHED; - erts_smp_atomic32_set(&schdlr_sspnd.msb.ongoing, 0); + schdlr_sspnd.msb.ongoing = 0; if (schdlr_sspnd.online == 1) { /* No schedulers to resume */ - ASSERT(erts_smp_atomic32_read(&schdlr_sspnd.active) == 1); + ASSERT(erts_smp_atomic32_read_nob(&schdlr_sspnd.active) == 1); ERTS_SCHDLR_SSPND_CHNG_SET(0, ERTS_SCHDLR_SSPND_CHNG_MSB); } else if (erts_common_run_queue) { for (ix = 1; ix < schdlr_sspnd.online; ix++) - erts_smp_atomic32_band(&ERTS_SCHED_SLEEP_INFO_IX(ix)->flags, - ~ERTS_SSI_FLG_SUSPENDED); + erts_smp_atomic32_read_band_nob(&ERTS_SCHED_SLEEP_INFO_IX(ix)->flags, + ~ERTS_SSI_FLG_SUSPENDED); wake_all_schedulers(); } else { @@ -3669,7 +3668,7 @@ void erts_dbg_multi_scheduling_return_trap(Process *p, Eterm return_value) { if (return_value == am_blocked) { - erts_aint32_t active = erts_smp_atomic32_read(&schdlr_sspnd.active); + erts_aint32_t active = erts_smp_atomic32_read_nob(&schdlr_sspnd.active); ASSERT(1 <= active && active <= 2); ASSERT(ERTS_PROC_GET_SCHDATA(p)->no == 1); } @@ -3752,12 +3751,12 @@ sched_thread_func(void *vesdp) erts_thread_init_float(); erts_smp_mtx_lock(&schdlr_sspnd.mtx); - ASSERT(erts_smp_atomic32_read(&schdlr_sspnd.changing) + ASSERT(erts_smp_atomic32_read_nob(&schdlr_sspnd.changing) & ERTS_SCHDLR_SSPND_CHNG_ONLN); if (--schdlr_sspnd.curr_online == schdlr_sspnd.wait_curr_online) { - erts_smp_atomic32_band(&schdlr_sspnd.changing, - ~ERTS_SCHDLR_SSPND_CHNG_ONLN); + erts_smp_atomic32_read_band_nob(&schdlr_sspnd.changing, + ~ERTS_SCHDLR_SSPND_CHNG_ONLN); if (((ErtsSchedulerData *) vesdp)->no != 1) erts_smp_cnd_signal(&schdlr_sspnd.cnd); } @@ -5214,7 +5213,7 @@ Process *schedule(Process *p, int calls) reds = ERTS_PROC_MIN_CONTEXT_SWITCH_REDS_COST; esdp->virtual_reds = 0; - fcalls = (int) erts_smp_atomic32_addtest(&function_calls, reds); + fcalls = (int) erts_smp_atomic32_add_read_acqb(&function_calls, reds); ASSERT(esdp && esdp == erts_get_scheduler_data()); rq = erts_get_runq_current(esdp); @@ -5349,7 +5348,7 @@ Process *schedule(Process *p, int calls) if ((rq->flags & ERTS_RUNQ_FLG_SUSPENDED) || (erts_smp_atomic32_read_acqb(&esdp->ssi->flags) & ERTS_SSI_FLG_SUSPENDED)) { - ASSERT(erts_smp_atomic32_read(&esdp->ssi->flags) + ASSERT(erts_smp_atomic32_read_nob(&esdp->ssi->flags) & ERTS_SSI_FLG_SUSPENDED); suspend_scheduler(esdp); } @@ -5363,7 +5362,7 @@ Process *schedule(Process *p, int calls) || defined(ERTS_SCHED_NEED_NONBLOCKABLE_AUX_WORK) { ErtsSchedulerSleepInfo *ssi = esdp->ssi; - erts_aint32_t aux_work = erts_smp_atomic32_read(&ssi->aux_work); + erts_aint32_t aux_work = erts_smp_atomic32_read_nob(&ssi->aux_work); if (aux_work) { erts_smp_runq_unlock(rq); #ifdef ERTS_SCHED_NEED_BLOCKABLE_AUX_WORK @@ -5407,7 +5406,7 @@ Process *schedule(Process *p, int calls) if ((rq->flags & ERTS_RUNQ_FLG_SUSPENDED) || (erts_smp_atomic32_read_acqb(&esdp->ssi->flags) & ERTS_SSI_FLG_SUSPENDED)) { - ASSERT(erts_smp_atomic32_read(&esdp->ssi->flags) + ASSERT(erts_smp_atomic32_read_nob(&esdp->ssi->flags) & ERTS_SSI_FLG_SUSPENDED); non_empty_runq(rq); goto continue_check_activities_to_run; @@ -5782,10 +5781,13 @@ erts_sched_stat_term(Process *p, int total) void erts_schedule_misc_op(void (*func)(void *), void *arg) { - ErtsRunQueue *rq = erts_get_runq_current(NULL); + ErtsRunQueue *rq; ErtsMiscOpList *molp = misc_op_list_alloc(); + ErtsSchedulerData *esdp = erts_get_scheduler_data(); - if (!rq) { + if (esdp) { + rq = esdp->run_queue; + } else { /* * This can only happen when the sys msg dispatcher * thread schedules misc ops (this happens *very* @@ -5948,7 +5950,7 @@ erts_test_next_pid(int set, Uint next) Uint erts_process_count(void) { - erts_aint32_t res = erts_smp_atomic32_read(&process_count); + erts_aint32_t res = erts_smp_atomic32_read_nob(&process_count); ASSERT(res >= 0); return (Uint) res; } @@ -5997,7 +5999,7 @@ alloc_process(void) ASSERT(!process_tab[p_next]); process_tab[p_next] = p; - erts_smp_atomic32_inc(&process_count); + erts_smp_atomic32_inc_nob(&process_count); p->id = make_internal_pid(p_serial << p_serial_shift | p_next); if (p->id == ERTS_INVALID_PID) { /* Do not use the invalid pid; change serial */ @@ -6123,7 +6125,7 @@ erl_create_process(Process* parent, /* Parent of process (default group leader). p->min_heap_size = H_MIN_SIZE; p->min_vheap_size = BIN_VH_MIN_SIZE; p->prio = PRIORITY_NORMAL; - p->max_gen_gcs = (Uint16) erts_smp_atomic32_read(&erts_max_gen_gcs); + p->max_gen_gcs = (Uint16) erts_smp_atomic32_read_nob(&erts_max_gen_gcs); } p->skipped = 0; ASSERT(p->min_heap_size == erts_next_heap_size(p->min_heap_size, 0)); @@ -7580,8 +7582,8 @@ continue_exit_process(Process *p p->status_flags = 0; #endif process_tab[pix] = NULL; /* Time of death! */ - ASSERT(erts_smp_atomic32_read(&process_count) > 0); - erts_smp_atomic32_dec(&process_count); + ASSERT(erts_smp_atomic32_read_nob(&process_count) > 0); + erts_smp_atomic32_dec_nob(&process_count); #ifdef ERTS_SMP erts_pix_unlock(pix_lock); diff --git a/erts/emulator/beam/erl_process.h b/erts/emulator/beam/erl_process.h index 296acc7367..627f10b142 100644 --- a/erts/emulator/beam/erl_process.h +++ b/erts/emulator/beam/erl_process.h @@ -387,19 +387,15 @@ do { \ } while (0) struct ErtsSchedulerData_ { - -#ifdef ERTS_SMP /* * Keep X registers first (so we get as many low * numbered registers as possible in the same cache * line). */ -#if !HALFWORD_HEAP - Eterm save_reg[ERTS_X_REGS_ALLOCATED]; /* X registers */ -#else - Eterm *save_reg; -#endif - FloatDef freg[MAX_REG]; /* Floating point registers. */ + Eterm* x_reg_array; /* X registers */ + FloatDef* f_reg_array; /* Floating point registers. */ + +#ifdef ERTS_SMP ethr_tid tid; /* Thread id */ struct erl_bits_state erl_bits_state; /* erl_bits.c state */ void *match_pseudo_process; /* erl_db_util.c:db_prog_match() */ @@ -1599,11 +1595,11 @@ erts_sched_poke(ErtsSchedulerSleepInfo *ssi) { erts_aint32_t flags; ERTS_THR_MEMORY_BARRIER; - flags = erts_smp_atomic32_read(&ssi->flags); + flags = erts_smp_atomic32_read_nob(&ssi->flags); ASSERT(!(flags & ERTS_SSI_FLG_SLEEPING) || (flags & ERTS_SSI_FLG_WAITING)); if (flags & ERTS_SSI_FLG_SLEEPING) { - flags = erts_smp_atomic32_band(&ssi->flags, ~ERTS_SSI_FLGS_SLEEP); + flags = erts_smp_atomic32_read_band_nob(&ssi->flags, ~ERTS_SSI_FLGS_SLEEP); erts_sched_finish_poke(ssi, flags); } } diff --git a/erts/emulator/beam/erl_process_dump.c b/erts/emulator/beam/erl_process_dump.c index 5410bcd495..3550f1396c 100644 --- a/erts/emulator/beam/erl_process_dump.c +++ b/erts/emulator/beam/erl_process_dump.c @@ -350,7 +350,7 @@ heap_dump(int to, void *to_arg, Eterm x) ProcBin* pb = (ProcBin *) binary_val(x); Binary* val = pb->val; - if (erts_smp_atomic_xchg(&val->refc, 0) != 0) { + if (erts_smp_atomic_xchg_nob(&val->refc, 0) != 0) { val->flags = (UWord) all_binaries; all_binaries = val; } diff --git a/erts/emulator/beam/erl_process_lock.c b/erts/emulator/beam/erl_process_lock.c index 72560aa124..83379d7352 100644 --- a/erts/emulator/beam/erl_process_lock.c +++ b/erts/emulator/beam/erl_process_lock.c @@ -1,7 +1,7 @@ /* * %CopyrightBegin% * - * Copyright Ericsson AB 2007-2010. All Rights Reserved. + * Copyright Ericsson AB 2007-2011. All Rights Reserved. * * The contents of this file are subject to the Erlang Public License, * Version 1.1, (the "License"); you may not use this file except in @@ -316,7 +316,7 @@ try_aquire(erts_proc_lock_t *lck, erts_tse_t *wtr) break; } wflg = lock << ERTS_PROC_LOCK_WAITER_SHIFT; - old_lflgs = ERTS_PROC_LOCK_FLGS_BOR_(lck, wflg | lock); + old_lflgs = ERTS_PROC_LOCK_FLGS_BOR_ACQB_(lck, wflg | lock); if (old_lflgs & lock) { /* Didn't get the lock */ goto enqueue; @@ -413,7 +413,7 @@ transfer_locks(Process *p, do { erts_tse_t *tmp = wake; wake = wake->next; - erts_atomic32_set(&tmp->uaflgs, 0); + erts_atomic32_set_nob(&tmp->uaflgs, 0); erts_tse_set(tmp); } while (wake); @@ -509,14 +509,14 @@ wait_for_locks(Process *p, ASSERT((wtr->uflgs & ~ERTS_PROC_LOCKS_ALL) == 0); - erts_atomic32_set(&wtr->uaflgs, 1); + erts_atomic32_set_nob(&wtr->uaflgs, 1); erts_pix_unlock(pix_lock); while (1) { int res; erts_tse_reset(wtr); - if (erts_atomic32_read(&wtr->uaflgs) == 0) + if (erts_atomic32_read_nob(&wtr->uaflgs) == 0) break; /* @@ -955,7 +955,8 @@ erts_proc_lock_init(Process *p) { /* We always start with all locks locked */ #if ERTS_PROC_LOCK_ATOMIC_IMPL - erts_smp_atomic32_init(&p->lock.flags, (erts_aint32_t) ERTS_PROC_LOCKS_ALL); + erts_smp_atomic32_init_nob(&p->lock.flags, + (erts_aint32_t) ERTS_PROC_LOCKS_ALL); #else p->lock.flags = ERTS_PROC_LOCKS_ALL; #endif @@ -974,7 +975,7 @@ erts_proc_lock_init(Process *p) { int i; for (i = 0; i <= ERTS_PROC_LOCK_MAX_BIT; i++) - erts_smp_atomic32_init(&p->lock.locked[i], (erts_aint32_t) 1); + erts_smp_atomic32_init_nob(&p->lock.locked[i], (erts_aint32_t) 1); } #endif } diff --git a/erts/emulator/beam/erl_process_lock.h b/erts/emulator/beam/erl_process_lock.h index 355179f084..cd3b2182fd 100644 --- a/erts/emulator/beam/erl_process_lock.h +++ b/erts/emulator/beam/erl_process_lock.h @@ -1,7 +1,7 @@ /* * %CopyrightBegin% * - * Copyright Ericsson AB 2007-2010. All Rights Reserved. + * Copyright Ericsson AB 2007-2011. All Rights Reserved. * * The contents of this file are subject to the Erlang Public License, * Version 1.1, (the "License"); you may not use this file except in @@ -41,10 +41,10 @@ #define ERTS_PROC_LOCK_SPINLOCK_IMPL 0 #define ERTS_PROC_LOCK_MUTEX_IMPL 0 -#if defined(ETHR_HAVE_OPTIMIZED_ATOMIC_OPS) +#if defined(ETHR_HAVE_32BIT_NATIVE_ATOMIC_OPS) # undef ERTS_PROC_LOCK_ATOMIC_IMPL # define ERTS_PROC_LOCK_ATOMIC_IMPL 1 -#elif defined(ETHR_HAVE_OPTIMIZED_SPINLOCK) +#elif defined(ETHR_HAVE_NATIVE_SPINLOCKS) # undef ERTS_PROC_LOCK_SPINLOCK_IMPL # define ERTS_PROC_LOCK_SPINLOCK_IMPL 1 #else @@ -270,9 +270,11 @@ typedef struct { #if ERTS_PROC_LOCK_ATOMIC_IMPL #define ERTS_PROC_LOCK_FLGS_BAND_(L, MSK) \ - ((ErtsProcLocks) erts_smp_atomic32_band(&(L)->flags, (erts_aint32_t) (MSK))) -#define ERTS_PROC_LOCK_FLGS_BOR_(L, MSK) \ - ((ErtsProcLocks) erts_smp_atomic32_bor(&(L)->flags, (erts_aint32_t) (MSK))) + ((ErtsProcLocks) erts_smp_atomic32_read_band_nob(&(L)->flags, \ + (erts_aint32_t) (MSK))) +#define ERTS_PROC_LOCK_FLGS_BOR_ACQB_(L, MSK) \ + ((ErtsProcLocks) erts_smp_atomic32_read_bor_acqb(&(L)->flags, \ + (erts_aint32_t) (MSK))) #define ERTS_PROC_LOCK_FLGS_CMPXCHG_ACQB_(L, NEW, EXPECTED) \ ((ErtsProcLocks) erts_smp_atomic32_cmpxchg_acqb(&(L)->flags, \ (erts_aint32_t) (NEW), \ @@ -282,7 +284,7 @@ typedef struct { (erts_aint32_t) (NEW), \ (erts_aint32_t) (EXPECTED))) #define ERTS_PROC_LOCK_FLGS_READ_(L) \ - ((ErtsProcLocks) erts_smp_atomic32_read(&(L)->flags)) + ((ErtsProcLocks) erts_smp_atomic32_read_nob(&(L)->flags)) #else /* no opt atomic ops */ @@ -325,7 +327,7 @@ erts_proc_lock_flags_cmpxchg(erts_proc_lock_t *lck, ErtsProcLocks new, #endif #define ERTS_PROC_LOCK_FLGS_BAND_(L, MSK) erts_proc_lock_flags_band((L), (MSK)) -#define ERTS_PROC_LOCK_FLGS_BOR_(L, MSK) erts_proc_lock_flags_bor((L), (MSK)) +#define ERTS_PROC_LOCK_FLGS_BOR_ACQB_(L, MSK) erts_proc_lock_flags_bor((L), (MSK)) #define ERTS_PROC_LOCK_FLGS_CMPXCHG_ACQB_(L, NEW, EXPECTED) \ erts_proc_lock_flags_cmpxchg((L), (NEW), (EXPECTED)) #define ERTS_PROC_LOCK_FLGS_CMPXCHG_RELB_(L, NEW, EXPECTED) \ @@ -623,11 +625,11 @@ erts_proc_lock_op_debug(Process *p, ErtsProcLocks locks, int locked) if (locks & lock) { erts_aint32_t lock_count; if (locked) { - lock_count = erts_smp_atomic32_inctest(&p->lock.locked[i]); + lock_count = erts_smp_atomic32_inc_read_nob(&p->lock.locked[i]); ERTS_LC_ASSERT(lock_count == 1); } else { - lock_count = erts_smp_atomic32_dectest(&p->lock.locked[i]); + lock_count = erts_smp_atomic32_dec_read_nob(&p->lock.locked[i]); ERTS_LC_ASSERT(lock_count == 0); } } diff --git a/erts/emulator/beam/erl_smp.h b/erts/emulator/beam/erl_smp.h index 287327bfe1..a89ddfbcc1 100644 --- a/erts/emulator/beam/erl_smp.h +++ b/erts/emulator/beam/erl_smp.h @@ -1,7 +1,7 @@ /* * %CopyrightBegin% * - * Copyright Ericsson AB 2005-2010. All Rights Reserved. + * Copyright Ericsson AB 2005-2011. All Rights Reserved. * * The contents of this file are subject to the Erlang Public License, * Version 1.1, (the "License"); you may not use this file except in @@ -54,8 +54,9 @@ typedef erts_cnd_t erts_smp_cnd_t; typedef erts_rwmtx_opt_t erts_smp_rwmtx_opt_t; typedef erts_rwmtx_t erts_smp_rwmtx_t; typedef erts_tsd_key_t erts_smp_tsd_key_t; -typedef erts_atomic_t erts_smp_atomic_t; -typedef erts_atomic32_t erts_smp_atomic32_t; +#define erts_smp_dw_atomic_t erts_dw_atomic_t +#define erts_smp_atomic_t erts_atomic_t +#define erts_smp_atomic32_t erts_atomic32_t typedef erts_spinlock_t erts_smp_spinlock_t; typedef erts_rwlock_t erts_smp_rwlock_t; void erts_thr_fatal_error(int, char *); /* implemented in erl_init.c */ @@ -83,8 +84,9 @@ typedef struct { } erts_smp_rwmtx_opt_t; typedef int erts_smp_rwmtx_t; typedef int erts_smp_tsd_key_t; -typedef SWord erts_smp_atomic_t; -typedef Uint32 erts_smp_atomic32_t; +#define erts_smp_dw_atomic_t erts_no_dw_atomic_t +#define erts_smp_atomic_t erts_no_atomic_t +#define erts_smp_atomic32_t erts_no_atomic32_t #if __GNUC__ > 2 typedef struct { } erts_smp_spinlock_t; typedef struct { } erts_smp_rwlock_t; @@ -160,82 +162,6 @@ ERTS_GLB_INLINE int erts_smp_rwmtx_tryrwlock(erts_smp_rwmtx_t *rwmtx); ERTS_GLB_INLINE void erts_smp_rwmtx_rwunlock(erts_smp_rwmtx_t *rwmtx); ERTS_GLB_INLINE int erts_smp_lc_rwmtx_is_rlocked(erts_smp_rwmtx_t *mtx); ERTS_GLB_INLINE int erts_smp_lc_rwmtx_is_rwlocked(erts_smp_rwmtx_t *mtx); -ERTS_GLB_INLINE void erts_smp_atomic_init(erts_smp_atomic_t *var, - erts_aint_t i); -ERTS_GLB_INLINE void erts_smp_atomic_set(erts_smp_atomic_t *var, erts_aint_t i); -ERTS_GLB_INLINE erts_aint_t erts_smp_atomic_read(erts_smp_atomic_t *var); -ERTS_GLB_INLINE erts_aint_t erts_smp_atomic_inctest(erts_smp_atomic_t *incp); -ERTS_GLB_INLINE erts_aint_t erts_smp_atomic_dectest(erts_smp_atomic_t *decp); -ERTS_GLB_INLINE void erts_smp_atomic_inc(erts_smp_atomic_t *incp); -ERTS_GLB_INLINE void erts_smp_atomic_dec(erts_smp_atomic_t *decp); -ERTS_GLB_INLINE erts_aint_t erts_smp_atomic_addtest(erts_smp_atomic_t *addp, - erts_aint_t i); -ERTS_GLB_INLINE void erts_smp_atomic_add(erts_smp_atomic_t *addp, - erts_aint_t i); -ERTS_GLB_INLINE erts_aint_t erts_smp_atomic_xchg(erts_smp_atomic_t *xchgp, - erts_aint_t new); -ERTS_GLB_INLINE erts_aint_t erts_smp_atomic_cmpxchg(erts_smp_atomic_t *xchgp, - erts_aint_t new, - erts_aint_t expected); -ERTS_GLB_INLINE erts_aint_t erts_smp_atomic_bor(erts_smp_atomic_t *var, - erts_aint_t mask); -ERTS_GLB_INLINE erts_aint_t erts_smp_atomic_band(erts_smp_atomic_t *var, - erts_aint_t mask); -ERTS_GLB_INLINE erts_aint_t erts_smp_atomic_read_acqb(erts_smp_atomic_t *var); -ERTS_GLB_INLINE void erts_smp_atomic_set_relb(erts_smp_atomic_t *var, - erts_aint_t i); -ERTS_GLB_INLINE void erts_smp_atomic_dec_relb(erts_smp_atomic_t *decp); -ERTS_GLB_INLINE erts_aint_t erts_smp_atomic_dectest_relb(erts_smp_atomic_t *decp); -ERTS_GLB_INLINE erts_aint_t erts_smp_atomic_cmpxchg_acqb(erts_smp_atomic_t *xchgp, - erts_aint_t new, - erts_aint_t exp); -ERTS_GLB_INLINE erts_aint_t erts_smp_atomic_cmpxchg_relb(erts_smp_atomic_t *xchgp, - erts_aint_t new, - erts_aint_t exp); -ERTS_GLB_INLINE void -erts_smp_atomic32_init(erts_smp_atomic32_t *var, erts_aint32_t i); -ERTS_GLB_INLINE void -erts_smp_atomic32_set(erts_smp_atomic32_t *var, erts_aint32_t i); -ERTS_GLB_INLINE erts_aint32_t -erts_smp_atomic32_read(erts_smp_atomic32_t *var); -ERTS_GLB_INLINE erts_aint32_t -erts_smp_atomic32_inctest(erts_smp_atomic32_t *incp); -ERTS_GLB_INLINE erts_aint32_t -erts_smp_atomic32_dectest(erts_smp_atomic32_t *decp); -ERTS_GLB_INLINE void -erts_smp_atomic32_inc(erts_smp_atomic32_t *incp); -ERTS_GLB_INLINE void -erts_smp_atomic32_dec(erts_smp_atomic32_t *decp); -ERTS_GLB_INLINE erts_aint32_t -erts_smp_atomic32_addtest(erts_smp_atomic32_t *addp, erts_aint32_t i); -ERTS_GLB_INLINE void -erts_smp_atomic32_add(erts_smp_atomic32_t *addp, erts_aint32_t i); -ERTS_GLB_INLINE erts_aint32_t -erts_smp_atomic32_xchg(erts_smp_atomic32_t *xchgp, erts_aint32_t new); -ERTS_GLB_INLINE erts_aint32_t -erts_smp_atomic32_cmpxchg(erts_smp_atomic32_t *xchgp, - erts_aint32_t new, - erts_aint32_t expected); -ERTS_GLB_INLINE erts_aint32_t -erts_smp_atomic32_bor(erts_smp_atomic32_t *var, erts_aint32_t mask); -ERTS_GLB_INLINE erts_aint32_t -erts_smp_atomic32_band(erts_smp_atomic32_t *var, erts_aint32_t mask); -ERTS_GLB_INLINE erts_aint32_t -erts_smp_atomic32_read_acqb(erts_smp_atomic32_t *var); -ERTS_GLB_INLINE void -erts_smp_atomic32_set_relb(erts_smp_atomic32_t *var, erts_aint32_t i); -ERTS_GLB_INLINE void -erts_smp_atomic32_dec_relb(erts_smp_atomic32_t *decp); -ERTS_GLB_INLINE erts_aint32_t -erts_smp_atomic32_dectest_relb(erts_smp_atomic32_t *decp); -ERTS_GLB_INLINE erts_aint32_t -erts_smp_atomic32_cmpxchg_acqb(erts_smp_atomic32_t *xchgp, - erts_aint32_t new, - erts_aint32_t exp); -ERTS_GLB_INLINE erts_aint32_t -erts_smp_atomic32_cmpxchg_relb(erts_smp_atomic32_t *xchgp, - erts_aint32_t new, - erts_aint32_t exp); ERTS_GLB_INLINE void erts_smp_spinlock_init_x(erts_smp_spinlock_t *lock, char *name, Eterm extra); @@ -279,6 +205,429 @@ ERTS_GLB_INLINE void erts_smp_thr_sigmask(int how, ERTS_GLB_INLINE void erts_smp_thr_sigwait(const sigset_t *set, int *sig); #endif /* #ifdef ERTS_THR_HAVE_SIG_FUNCS */ +/* + * Functions implementing atomic operations with with no (nob), + * full (mb), acquire (acqb), release (relb), read (rb), and + * write (wb) memory barriers. + * + * If SMP support has been disabled, they are mapped to functions + * that performs the same operation, but aren't atomic and don't + * imply memory barriers. + */ + +#ifdef ERTS_SMP + +/* Double word size atomics */ + +#define erts_smp_dw_atomic_init_nob erts_dw_atomic_init_nob +#define erts_smp_dw_atomic_set_nob erts_dw_atomic_set_nob +#define erts_smp_dw_atomic_read_nob erts_dw_atomic_read_nob +#define erts_smp_dw_atomic_cmpxchg_nob erts_dw_atomic_cmpxchg_nob + +#define erts_smp_dw_atomic_init_mb erts_dw_atomic_init_mb +#define erts_smp_dw_atomic_set_mb erts_dw_atomic_set_mb +#define erts_smp_dw_atomic_read_mb erts_dw_atomic_read_mb +#define erts_smp_dw_atomic_cmpxchg_mb erts_dw_atomic_cmpxchg_mb + +#define erts_smp_dw_atomic_init_acqb erts_dw_atomic_init_acqb +#define erts_smp_dw_atomic_set_acqb erts_dw_atomic_set_acqb +#define erts_smp_dw_atomic_read_acqb erts_dw_atomic_read_acqb +#define erts_smp_dw_atomic_cmpxchg_acqb erts_dw_atomic_cmpxchg_acqb + +#define erts_smp_dw_atomic_init_relb erts_dw_atomic_init_relb +#define erts_smp_dw_atomic_set_relb erts_dw_atomic_set_relb +#define erts_smp_dw_atomic_read_relb erts_dw_atomic_read_relb +#define erts_smp_dw_atomic_cmpxchg_relb erts_dw_atomic_cmpxchg_relb + +#define erts_smp_dw_atomic_init_rb erts_dw_atomic_init_rb +#define erts_smp_dw_atomic_set_rb erts_dw_atomic_set_rb +#define erts_smp_dw_atomic_read_rb erts_dw_atomic_read_rb +#define erts_smp_dw_atomic_cmpxchg_rb erts_dw_atomic_cmpxchg_rb + +#define erts_smp_dw_atomic_init_wb erts_dw_atomic_init_wb +#define erts_smp_dw_atomic_set_wb erts_dw_atomic_set_wb +#define erts_smp_dw_atomic_read_wb erts_dw_atomic_read_wb +#define erts_smp_dw_atomic_cmpxchg_wb erts_dw_atomic_cmpxchg_wb + +/* Word size atomics */ + +#define erts_smp_atomic_init_nob erts_atomic_init_nob +#define erts_smp_atomic_set_nob erts_atomic_set_nob +#define erts_smp_atomic_read_nob erts_atomic_read_nob +#define erts_smp_atomic_inc_read_nob erts_atomic_inc_read_nob +#define erts_smp_atomic_dec_read_nob erts_atomic_dec_read_nob +#define erts_smp_atomic_inc_nob erts_atomic_inc_nob +#define erts_smp_atomic_dec_nob erts_atomic_dec_nob +#define erts_smp_atomic_add_read_nob erts_atomic_add_read_nob +#define erts_smp_atomic_add_nob erts_atomic_add_nob +#define erts_smp_atomic_read_bor_nob erts_atomic_read_bor_nob +#define erts_smp_atomic_read_band_nob erts_atomic_read_band_nob +#define erts_smp_atomic_xchg_nob erts_atomic_xchg_nob +#define erts_smp_atomic_cmpxchg_nob erts_atomic_cmpxchg_nob + +#define erts_smp_atomic_init_mb erts_atomic_init_mb +#define erts_smp_atomic_set_mb erts_atomic_set_mb +#define erts_smp_atomic_read_mb erts_atomic_read_mb +#define erts_smp_atomic_inc_read_mb erts_atomic_inc_read_mb +#define erts_smp_atomic_dec_read_mb erts_atomic_dec_read_mb +#define erts_smp_atomic_inc_mb erts_atomic_inc_mb +#define erts_smp_atomic_dec_mb erts_atomic_dec_mb +#define erts_smp_atomic_add_read_mb erts_atomic_add_read_mb +#define erts_smp_atomic_add_mb erts_atomic_add_mb +#define erts_smp_atomic_read_bor_mb erts_atomic_read_bor_mb +#define erts_smp_atomic_read_band_mb erts_atomic_read_band_mb +#define erts_smp_atomic_xchg_mb erts_atomic_xchg_mb +#define erts_smp_atomic_cmpxchg_mb erts_atomic_cmpxchg_mb + +#define erts_smp_atomic_init_acqb erts_atomic_init_acqb +#define erts_smp_atomic_set_acqb erts_atomic_set_acqb +#define erts_smp_atomic_read_acqb erts_atomic_read_acqb +#define erts_smp_atomic_inc_read_acqb erts_atomic_inc_read_acqb +#define erts_smp_atomic_dec_read_acqb erts_atomic_dec_read_acqb +#define erts_smp_atomic_inc_acqb erts_atomic_inc_acqb +#define erts_smp_atomic_dec_acqb erts_atomic_dec_acqb +#define erts_smp_atomic_add_read_acqb erts_atomic_add_read_acqb +#define erts_smp_atomic_add_acqb erts_atomic_add_acqb +#define erts_smp_atomic_read_bor_acqb erts_atomic_read_bor_acqb +#define erts_smp_atomic_read_band_acqb erts_atomic_read_band_acqb +#define erts_smp_atomic_xchg_acqb erts_atomic_xchg_acqb +#define erts_smp_atomic_cmpxchg_acqb erts_atomic_cmpxchg_acqb + +#define erts_smp_atomic_init_relb erts_atomic_init_relb +#define erts_smp_atomic_set_relb erts_atomic_set_relb +#define erts_smp_atomic_read_relb erts_atomic_read_relb +#define erts_smp_atomic_inc_read_relb erts_atomic_inc_read_relb +#define erts_smp_atomic_dec_read_relb erts_atomic_dec_read_relb +#define erts_smp_atomic_inc_relb erts_atomic_inc_relb +#define erts_smp_atomic_dec_relb erts_atomic_dec_relb +#define erts_smp_atomic_add_read_relb erts_atomic_add_read_relb +#define erts_smp_atomic_add_relb erts_atomic_add_relb +#define erts_smp_atomic_read_bor_relb erts_atomic_read_bor_relb +#define erts_smp_atomic_read_band_relb erts_atomic_read_band_relb +#define erts_smp_atomic_xchg_relb erts_atomic_xchg_relb +#define erts_smp_atomic_cmpxchg_relb erts_atomic_cmpxchg_relb + +#define erts_smp_atomic_init_rb erts_atomic_init_rb +#define erts_smp_atomic_set_rb erts_atomic_set_rb +#define erts_smp_atomic_read_rb erts_atomic_read_rb +#define erts_smp_atomic_inc_read_rb erts_atomic_inc_read_rb +#define erts_smp_atomic_dec_read_rb erts_atomic_dec_read_rb +#define erts_smp_atomic_inc_rb erts_atomic_inc_rb +#define erts_smp_atomic_dec_rb erts_atomic_dec_rb +#define erts_smp_atomic_add_read_rb erts_atomic_add_read_rb +#define erts_smp_atomic_add_rb erts_atomic_add_rb +#define erts_smp_atomic_read_bor_rb erts_atomic_read_bor_rb +#define erts_smp_atomic_read_band_rb erts_atomic_read_band_rb +#define erts_smp_atomic_xchg_rb erts_atomic_xchg_rb +#define erts_smp_atomic_cmpxchg_rb erts_atomic_cmpxchg_rb + +#define erts_smp_atomic_init_wb erts_atomic_init_wb +#define erts_smp_atomic_set_wb erts_atomic_set_wb +#define erts_smp_atomic_read_wb erts_atomic_read_wb +#define erts_smp_atomic_inc_read_wb erts_atomic_inc_read_wb +#define erts_smp_atomic_dec_read_wb erts_atomic_dec_read_wb +#define erts_smp_atomic_inc_wb erts_atomic_inc_wb +#define erts_smp_atomic_dec_wb erts_atomic_dec_wb +#define erts_smp_atomic_add_read_wb erts_atomic_add_read_wb +#define erts_smp_atomic_add_wb erts_atomic_add_wb +#define erts_smp_atomic_read_bor_wb erts_atomic_read_bor_wb +#define erts_smp_atomic_read_band_wb erts_atomic_read_band_wb +#define erts_smp_atomic_xchg_wb erts_atomic_xchg_wb +#define erts_smp_atomic_cmpxchg_wb erts_atomic_cmpxchg_wb + +/* 32-bit atomics */ + +#define erts_smp_atomic32_init_nob erts_atomic32_init_nob +#define erts_smp_atomic32_set_nob erts_atomic32_set_nob +#define erts_smp_atomic32_read_nob erts_atomic32_read_nob +#define erts_smp_atomic32_inc_read_nob erts_atomic32_inc_read_nob +#define erts_smp_atomic32_dec_read_nob erts_atomic32_dec_read_nob +#define erts_smp_atomic32_inc_nob erts_atomic32_inc_nob +#define erts_smp_atomic32_dec_nob erts_atomic32_dec_nob +#define erts_smp_atomic32_add_read_nob erts_atomic32_add_read_nob +#define erts_smp_atomic32_add_nob erts_atomic32_add_nob +#define erts_smp_atomic32_read_bor_nob erts_atomic32_read_bor_nob +#define erts_smp_atomic32_read_band_nob erts_atomic32_read_band_nob +#define erts_smp_atomic32_xchg_nob erts_atomic32_xchg_nob +#define erts_smp_atomic32_cmpxchg_nob erts_atomic32_cmpxchg_nob + +#define erts_smp_atomic32_init_mb erts_atomic32_init_mb +#define erts_smp_atomic32_set_mb erts_atomic32_set_mb +#define erts_smp_atomic32_read_mb erts_atomic32_read_mb +#define erts_smp_atomic32_inc_read_mb erts_atomic32_inc_read_mb +#define erts_smp_atomic32_dec_read_mb erts_atomic32_dec_read_mb +#define erts_smp_atomic32_inc_mb erts_atomic32_inc_mb +#define erts_smp_atomic32_dec_mb erts_atomic32_dec_mb +#define erts_smp_atomic32_add_read_mb erts_atomic32_add_read_mb +#define erts_smp_atomic32_add_mb erts_atomic32_add_mb +#define erts_smp_atomic32_read_bor_mb erts_atomic32_read_bor_mb +#define erts_smp_atomic32_read_band_mb erts_atomic32_read_band_mb +#define erts_smp_atomic32_xchg_mb erts_atomic32_xchg_mb +#define erts_smp_atomic32_cmpxchg_mb erts_atomic32_cmpxchg_mb + +#define erts_smp_atomic32_init_acqb erts_atomic32_init_acqb +#define erts_smp_atomic32_set_acqb erts_atomic32_set_acqb +#define erts_smp_atomic32_read_acqb erts_atomic32_read_acqb +#define erts_smp_atomic32_inc_read_acqb erts_atomic32_inc_read_acqb +#define erts_smp_atomic32_dec_read_acqb erts_atomic32_dec_read_acqb +#define erts_smp_atomic32_inc_acqb erts_atomic32_inc_acqb +#define erts_smp_atomic32_dec_acqb erts_atomic32_dec_acqb +#define erts_smp_atomic32_add_read_acqb erts_atomic32_add_read_acqb +#define erts_smp_atomic32_add_acqb erts_atomic32_add_acqb +#define erts_smp_atomic32_read_bor_acqb erts_atomic32_read_bor_acqb +#define erts_smp_atomic32_read_band_acqb erts_atomic32_read_band_acqb +#define erts_smp_atomic32_xchg_acqb erts_atomic32_xchg_acqb +#define erts_smp_atomic32_cmpxchg_acqb erts_atomic32_cmpxchg_acqb + +#define erts_smp_atomic32_init_relb erts_atomic32_init_relb +#define erts_smp_atomic32_set_relb erts_atomic32_set_relb +#define erts_smp_atomic32_read_relb erts_atomic32_read_relb +#define erts_smp_atomic32_inc_read_relb erts_atomic32_inc_read_relb +#define erts_smp_atomic32_dec_read_relb erts_atomic32_dec_read_relb +#define erts_smp_atomic32_inc_relb erts_atomic32_inc_relb +#define erts_smp_atomic32_dec_relb erts_atomic32_dec_relb +#define erts_smp_atomic32_add_read_relb erts_atomic32_add_read_relb +#define erts_smp_atomic32_add_relb erts_atomic32_add_relb +#define erts_smp_atomic32_read_bor_relb erts_atomic32_read_bor_relb +#define erts_smp_atomic32_read_band_relb erts_atomic32_read_band_relb +#define erts_smp_atomic32_xchg_relb erts_atomic32_xchg_relb +#define erts_smp_atomic32_cmpxchg_relb erts_atomic32_cmpxchg_relb + +#define erts_smp_atomic32_init_rb erts_atomic32_init_rb +#define erts_smp_atomic32_set_rb erts_atomic32_set_rb +#define erts_smp_atomic32_read_rb erts_atomic32_read_rb +#define erts_smp_atomic32_inc_read_rb erts_atomic32_inc_read_rb +#define erts_smp_atomic32_dec_read_rb erts_atomic32_dec_read_rb +#define erts_smp_atomic32_inc_rb erts_atomic32_inc_rb +#define erts_smp_atomic32_dec_rb erts_atomic32_dec_rb +#define erts_smp_atomic32_add_read_rb erts_atomic32_add_read_rb +#define erts_smp_atomic32_add_rb erts_atomic32_add_rb +#define erts_smp_atomic32_read_bor_rb erts_atomic32_read_bor_rb +#define erts_smp_atomic32_read_band_rb erts_atomic32_read_band_rb +#define erts_smp_atomic32_xchg_rb erts_atomic32_xchg_rb +#define erts_smp_atomic32_cmpxchg_rb erts_atomic32_cmpxchg_rb + +#define erts_smp_atomic32_init_wb erts_atomic32_init_wb +#define erts_smp_atomic32_set_wb erts_atomic32_set_wb +#define erts_smp_atomic32_read_wb erts_atomic32_read_wb +#define erts_smp_atomic32_inc_read_wb erts_atomic32_inc_read_wb +#define erts_smp_atomic32_dec_read_wb erts_atomic32_dec_read_wb +#define erts_smp_atomic32_inc_wb erts_atomic32_inc_wb +#define erts_smp_atomic32_dec_wb erts_atomic32_dec_wb +#define erts_smp_atomic32_add_read_wb erts_atomic32_add_read_wb +#define erts_smp_atomic32_add_wb erts_atomic32_add_wb +#define erts_smp_atomic32_read_bor_wb erts_atomic32_read_bor_wb +#define erts_smp_atomic32_read_band_wb erts_atomic32_read_band_wb +#define erts_smp_atomic32_xchg_wb erts_atomic32_xchg_wb +#define erts_smp_atomic32_cmpxchg_wb erts_atomic32_cmpxchg_wb + +#else /* !ERTS_SMP */ + +/* Double word size atomics */ + +#define erts_smp_dw_atomic_init_nob erts_no_dw_atomic_set +#define erts_smp_dw_atomic_set_nob erts_no_dw_atomic_set +#define erts_smp_dw_atomic_read_nob erts_no_dw_atomic_read +#define erts_smp_dw_atomic_cmpxchg_nob erts_no_dw_atomic_cmpxchg + +#define erts_smp_dw_atomic_init_mb erts_no_dw_atomic_init +#define erts_smp_dw_atomic_set_mb erts_no_dw_atomic_set +#define erts_smp_dw_atomic_read_mb erts_no_dw_atomic_read +#define erts_smp_dw_atomic_cmpxchg_mb erts_no_dw_atomic_cmpxchg + +#define erts_smp_dw_atomic_init_acqb erts_no_dw_atomic_init +#define erts_smp_dw_atomic_set_acqb erts_no_dw_atomic_set +#define erts_smp_dw_atomic_read_acqb erts_no_dw_atomic_read +#define erts_smp_dw_atomic_cmpxchg_acqb erts_no_dw_atomic_cmpxchg + +#define erts_smp_dw_atomic_init_relb erts_no_dw_atomic_init +#define erts_smp_dw_atomic_set_relb erts_no_dw_atomic_set +#define erts_smp_dw_atomic_read_relb erts_no_dw_atomic_read +#define erts_smp_dw_atomic_cmpxchg_relb erts_no_dw_atomic_cmpxchg + +#define erts_smp_dw_atomic_init_rb erts_no_dw_atomic_init +#define erts_smp_dw_atomic_set_rb erts_no_dw_atomic_set +#define erts_smp_dw_atomic_read_rb erts_no_dw_atomic_read +#define erts_smp_dw_atomic_cmpxchg_rb erts_no_dw_atomic_cmpxchg + +#define erts_smp_dw_atomic_init_wb erts_no_dw_atomic_init +#define erts_smp_dw_atomic_set_wb erts_no_dw_atomic_set +#define erts_smp_dw_atomic_read_wb erts_no_dw_atomic_read +#define erts_smp_dw_atomic_cmpxchg_wb erts_no_dw_atomic_cmpxchg + +/* Word size atomics */ + +#define erts_smp_atomic_init_nob erts_no_atomic_set +#define erts_smp_atomic_set_nob erts_no_atomic_set +#define erts_smp_atomic_read_nob erts_no_atomic_read +#define erts_smp_atomic_inc_read_nob erts_no_atomic_inc_read +#define erts_smp_atomic_dec_read_nob erts_no_atomic_dec_read +#define erts_smp_atomic_inc_nob erts_no_atomic_inc +#define erts_smp_atomic_dec_nob erts_no_atomic_dec +#define erts_smp_atomic_add_read_nob erts_no_atomic_add_read +#define erts_smp_atomic_add_nob erts_no_atomic_add +#define erts_smp_atomic_read_bor_nob erts_no_atomic_read_bor +#define erts_smp_atomic_read_band_nob erts_no_atomic_read_band +#define erts_smp_atomic_xchg_nob erts_no_atomic_xchg +#define erts_smp_atomic_cmpxchg_nob erts_no_atomic_cmpxchg + +#define erts_smp_atomic_init_mb erts_no_atomic_set +#define erts_smp_atomic_set_mb erts_no_atomic_set +#define erts_smp_atomic_read_mb erts_no_atomic_read +#define erts_smp_atomic_inc_read_mb erts_no_atomic_inc_read +#define erts_smp_atomic_dec_read_mb erts_no_atomic_dec_read +#define erts_smp_atomic_inc_mb erts_no_atomic_inc +#define erts_smp_atomic_dec_mb erts_no_atomic_dec +#define erts_smp_atomic_add_read_mb erts_no_atomic_add_read +#define erts_smp_atomic_add_mb erts_no_atomic_add +#define erts_smp_atomic_read_bor_mb erts_no_atomic_read_bor +#define erts_smp_atomic_read_band_mb erts_no_atomic_read_band +#define erts_smp_atomic_xchg_mb erts_no_atomic_xchg +#define erts_smp_atomic_cmpxchg_mb erts_no_atomic_cmpxchg + +#define erts_smp_atomic_init_acqb erts_no_atomic_set +#define erts_smp_atomic_set_acqb erts_no_atomic_set +#define erts_smp_atomic_read_acqb erts_no_atomic_read +#define erts_smp_atomic_inc_read_acqb erts_no_atomic_inc_read +#define erts_smp_atomic_dec_read_acqb erts_no_atomic_dec_read +#define erts_smp_atomic_inc_acqb erts_no_atomic_inc +#define erts_smp_atomic_dec_acqb erts_no_atomic_dec +#define erts_smp_atomic_add_read_acqb erts_no_atomic_add_read +#define erts_smp_atomic_add_acqb erts_no_atomic_add +#define erts_smp_atomic_read_bor_acqb erts_no_atomic_read_bor +#define erts_smp_atomic_read_band_acqb erts_no_atomic_read_band +#define erts_smp_atomic_xchg_acqb erts_no_atomic_xchg +#define erts_smp_atomic_cmpxchg_acqb erts_no_atomic_cmpxchg + +#define erts_smp_atomic_init_relb erts_no_atomic_set +#define erts_smp_atomic_set_relb erts_no_atomic_set +#define erts_smp_atomic_read_relb erts_no_atomic_read +#define erts_smp_atomic_inc_read_relb erts_no_atomic_inc_read +#define erts_smp_atomic_dec_read_relb erts_no_atomic_dec_read +#define erts_smp_atomic_inc_relb erts_no_atomic_inc +#define erts_smp_atomic_dec_relb erts_no_atomic_dec +#define erts_smp_atomic_add_read_relb erts_no_atomic_add_read +#define erts_smp_atomic_add_relb erts_no_atomic_add +#define erts_smp_atomic_read_bor_relb erts_no_atomic_read_bor +#define erts_smp_atomic_read_band_relb erts_no_atomic_read_band +#define erts_smp_atomic_xchg_relb erts_no_atomic_xchg +#define erts_smp_atomic_cmpxchg_relb erts_no_atomic_cmpxchg + +#define erts_smp_atomic_init_rb erts_no_atomic_set +#define erts_smp_atomic_set_rb erts_no_atomic_set +#define erts_smp_atomic_read_rb erts_no_atomic_read +#define erts_smp_atomic_inc_read_rb erts_no_atomic_inc_read +#define erts_smp_atomic_dec_read_rb erts_no_atomic_dec_read +#define erts_smp_atomic_inc_rb erts_no_atomic_inc +#define erts_smp_atomic_dec_rb erts_no_atomic_dec +#define erts_smp_atomic_add_read_rb erts_no_atomic_add_read +#define erts_smp_atomic_add_rb erts_no_atomic_add +#define erts_smp_atomic_read_bor_rb erts_no_atomic_read_bor +#define erts_smp_atomic_read_band_rb erts_no_atomic_read_band +#define erts_smp_atomic_xchg_rb erts_no_atomic_xchg +#define erts_smp_atomic_cmpxchg_rb erts_no_atomic_cmpxchg + +#define erts_smp_atomic_init_wb erts_no_atomic_set +#define erts_smp_atomic_set_wb erts_no_atomic_set +#define erts_smp_atomic_read_wb erts_no_atomic_read +#define erts_smp_atomic_inc_read_wb erts_no_atomic_inc_read +#define erts_smp_atomic_dec_read_wb erts_no_atomic_dec_read +#define erts_smp_atomic_inc_wb erts_no_atomic_inc +#define erts_smp_atomic_dec_wb erts_no_atomic_dec +#define erts_smp_atomic_add_read_wb erts_no_atomic_add_read +#define erts_smp_atomic_add_wb erts_no_atomic_add +#define erts_smp_atomic_read_bor_wb erts_no_atomic_read_bor +#define erts_smp_atomic_read_band_wb erts_no_atomic_read_band +#define erts_smp_atomic_xchg_wb erts_no_atomic_xchg +#define erts_smp_atomic_cmpxchg_wb erts_no_atomic_cmpxchg + +/* 32-bit atomics */ + +#define erts_smp_atomic32_init_nob erts_no_atomic32_set +#define erts_smp_atomic32_set_nob erts_no_atomic32_set +#define erts_smp_atomic32_read_nob erts_no_atomic32_read +#define erts_smp_atomic32_inc_read_nob erts_no_atomic32_inc_read +#define erts_smp_atomic32_dec_read_nob erts_no_atomic32_dec_read +#define erts_smp_atomic32_inc_nob erts_no_atomic32_inc +#define erts_smp_atomic32_dec_nob erts_no_atomic32_dec +#define erts_smp_atomic32_add_read_nob erts_no_atomic32_add_read +#define erts_smp_atomic32_add_nob erts_no_atomic32_add +#define erts_smp_atomic32_read_bor_nob erts_no_atomic32_read_bor +#define erts_smp_atomic32_read_band_nob erts_no_atomic32_read_band +#define erts_smp_atomic32_xchg_nob erts_no_atomic32_xchg +#define erts_smp_atomic32_cmpxchg_nob erts_no_atomic32_cmpxchg + +#define erts_smp_atomic32_init_mb erts_no_atomic32_set +#define erts_smp_atomic32_set_mb erts_no_atomic32_set +#define erts_smp_atomic32_read_mb erts_no_atomic32_read +#define erts_smp_atomic32_inc_read_mb erts_no_atomic32_inc_read +#define erts_smp_atomic32_dec_read_mb erts_no_atomic32_dec_read +#define erts_smp_atomic32_inc_mb erts_no_atomic32_inc +#define erts_smp_atomic32_dec_mb erts_no_atomic32_dec +#define erts_smp_atomic32_add_read_mb erts_no_atomic32_add_read +#define erts_smp_atomic32_add_mb erts_no_atomic32_add +#define erts_smp_atomic32_read_bor_mb erts_no_atomic32_read_bor +#define erts_smp_atomic32_read_band_mb erts_no_atomic32_read_band +#define erts_smp_atomic32_xchg_mb erts_no_atomic32_xchg +#define erts_smp_atomic32_cmpxchg_mb erts_no_atomic32_cmpxchg + +#define erts_smp_atomic32_init_acqb erts_no_atomic32_set +#define erts_smp_atomic32_set_acqb erts_no_atomic32_set +#define erts_smp_atomic32_read_acqb erts_no_atomic32_read +#define erts_smp_atomic32_inc_read_acqb erts_no_atomic32_inc_read +#define erts_smp_atomic32_dec_read_acqb erts_no_atomic32_dec_read +#define erts_smp_atomic32_inc_acqb erts_no_atomic32_inc +#define erts_smp_atomic32_dec_acqb erts_no_atomic32_dec +#define erts_smp_atomic32_add_read_acqb erts_no_atomic32_add_read +#define erts_smp_atomic32_add_acqb erts_no_atomic32_add +#define erts_smp_atomic32_read_bor_acqb erts_no_atomic32_read_bor +#define erts_smp_atomic32_read_band_acqb erts_no_atomic32_read_band +#define erts_smp_atomic32_xchg_acqb erts_no_atomic32_xchg +#define erts_smp_atomic32_cmpxchg_acqb erts_no_atomic32_cmpxchg + +#define erts_smp_atomic32_init_relb erts_no_atomic32_set +#define erts_smp_atomic32_set_relb erts_no_atomic32_set +#define erts_smp_atomic32_read_relb erts_no_atomic32_read +#define erts_smp_atomic32_inc_read_relb erts_no_atomic32_inc_read +#define erts_smp_atomic32_dec_read_relb erts_no_atomic32_dec_read +#define erts_smp_atomic32_inc_relb erts_no_atomic32_inc +#define erts_smp_atomic32_dec_relb erts_no_atomic32_dec +#define erts_smp_atomic32_add_read_relb erts_no_atomic32_add_read +#define erts_smp_atomic32_add_relb erts_no_atomic32_add +#define erts_smp_atomic32_read_bor_relb erts_no_atomic32_read_bor +#define erts_smp_atomic32_read_band_relb erts_no_atomic32_read_band +#define erts_smp_atomic32_xchg_relb erts_no_atomic32_xchg +#define erts_smp_atomic32_cmpxchg_relb erts_no_atomic32_cmpxchg + +#define erts_smp_atomic32_init_rb erts_no_atomic32_set +#define erts_smp_atomic32_set_rb erts_no_atomic32_set +#define erts_smp_atomic32_read_rb erts_no_atomic32_read +#define erts_smp_atomic32_inc_read_rb erts_no_atomic32_inc_read +#define erts_smp_atomic32_dec_read_rb erts_no_atomic32_dec_read +#define erts_smp_atomic32_inc_rb erts_no_atomic32_inc +#define erts_smp_atomic32_dec_rb erts_no_atomic32_dec +#define erts_smp_atomic32_add_read_rb erts_no_atomic32_add_read +#define erts_smp_atomic32_add_rb erts_no_atomic32_add +#define erts_smp_atomic32_read_bor_rb erts_no_atomic32_read_bor +#define erts_smp_atomic32_read_band_rb erts_no_atomic32_read_band +#define erts_smp_atomic32_xchg_rb erts_no_atomic32_xchg +#define erts_smp_atomic32_cmpxchg_rb erts_no_atomic32_cmpxchg + +#define erts_smp_atomic32_init_wb erts_no_atomic32_set +#define erts_smp_atomic32_set_wb erts_no_atomic32_set +#define erts_smp_atomic32_read_wb erts_no_atomic32_read +#define erts_smp_atomic32_inc_read_wb erts_no_atomic32_inc_read +#define erts_smp_atomic32_dec_read_wb erts_no_atomic32_dec_read +#define erts_smp_atomic32_inc_wb erts_no_atomic32_inc +#define erts_smp_atomic32_dec_wb erts_no_atomic32_dec +#define erts_smp_atomic32_add_read_wb erts_no_atomic32_add_read +#define erts_smp_atomic32_add_wb erts_no_atomic32_add +#define erts_smp_atomic32_read_bor_wb erts_no_atomic32_read_bor +#define erts_smp_atomic32_read_band_wb erts_no_atomic32_read_band +#define erts_smp_atomic32_xchg_wb erts_no_atomic32_xchg +#define erts_smp_atomic32_cmpxchg_wb erts_no_atomic32_cmpxchg + +#endif /* !ERTS_SMP */ #if ERTS_GLB_INLINE_INCL_FUNC_DEF @@ -655,434 +1004,6 @@ erts_smp_lc_rwmtx_is_rwlocked(erts_smp_rwmtx_t *mtx) } ERTS_GLB_INLINE void -erts_smp_atomic_init(erts_smp_atomic_t *var, erts_aint_t i) -{ -#ifdef ERTS_SMP - erts_atomic_init(var, i); -#else - *var = i; -#endif -} - -ERTS_GLB_INLINE void -erts_smp_atomic_set(erts_smp_atomic_t *var, erts_aint_t i) -{ -#ifdef ERTS_SMP - erts_atomic_set(var, i); -#else - *var = i; -#endif -} - -ERTS_GLB_INLINE erts_aint_t -erts_smp_atomic_read(erts_smp_atomic_t *var) -{ -#ifdef ERTS_SMP - return erts_atomic_read(var); -#else - return *var; -#endif -} - -ERTS_GLB_INLINE erts_aint_t -erts_smp_atomic_inctest(erts_smp_atomic_t *incp) -{ -#ifdef ERTS_SMP - return erts_atomic_inctest(incp); -#else - return ++(*incp); -#endif -} - -ERTS_GLB_INLINE erts_aint_t -erts_smp_atomic_dectest(erts_smp_atomic_t *decp) -{ -#ifdef ERTS_SMP - return erts_atomic_dectest(decp); -#else - return --(*decp); -#endif -} - -ERTS_GLB_INLINE void -erts_smp_atomic_inc(erts_smp_atomic_t *incp) -{ -#ifdef ERTS_SMP - erts_atomic_inc(incp); -#else - ++(*incp); -#endif -} - -ERTS_GLB_INLINE void -erts_smp_atomic_dec(erts_smp_atomic_t *decp) -{ -#ifdef ERTS_SMP - erts_atomic_dec(decp); -#else - --(*decp); -#endif -} - -ERTS_GLB_INLINE erts_aint_t -erts_smp_atomic_addtest(erts_smp_atomic_t *addp, erts_aint_t i) -{ -#ifdef ERTS_SMP - return erts_atomic_addtest(addp, i); -#else - return *addp += i; -#endif -} - -ERTS_GLB_INLINE void -erts_smp_atomic_add(erts_smp_atomic_t *addp, erts_aint_t i) -{ -#ifdef ERTS_SMP - erts_atomic_add(addp, i); -#else - *addp += i; -#endif -} - -ERTS_GLB_INLINE erts_aint_t -erts_smp_atomic_xchg(erts_smp_atomic_t *xchgp, erts_aint_t new) -{ -#ifdef ERTS_SMP - return erts_atomic_xchg(xchgp, new); -#else - erts_aint_t old; - old = *xchgp; - *xchgp = new; - return old; -#endif -} - -ERTS_GLB_INLINE erts_aint_t -erts_smp_atomic_cmpxchg(erts_smp_atomic_t *xchgp, - erts_aint_t new, - erts_aint_t expected) -{ -#ifdef ERTS_SMP - return erts_atomic_cmpxchg(xchgp, new, expected); -#else - erts_aint_t old = *xchgp; - if (old == expected) - *xchgp = new; - return old; -#endif -} - -ERTS_GLB_INLINE erts_aint_t -erts_smp_atomic_bor(erts_smp_atomic_t *var, erts_aint_t mask) -{ -#ifdef ERTS_SMP - return erts_atomic_bor(var, mask); -#else - erts_aint_t old; - old = *var; - *var |= mask; - return old; -#endif -} - -ERTS_GLB_INLINE erts_aint_t -erts_smp_atomic_band(erts_smp_atomic_t *var, erts_aint_t mask) -{ -#ifdef ERTS_SMP - return erts_atomic_band(var, mask); -#else - erts_aint_t old; - old = *var; - *var &= mask; - return old; -#endif -} - -ERTS_GLB_INLINE erts_aint_t -erts_smp_atomic_read_acqb(erts_smp_atomic_t *var) -{ -#ifdef ERTS_SMP - return erts_atomic_read_acqb(var); -#else - return *var; -#endif -} - -ERTS_GLB_INLINE void -erts_smp_atomic_set_relb(erts_smp_atomic_t *var, erts_aint_t i) -{ -#ifdef ERTS_SMP - erts_atomic_set_relb(var, i); -#else - *var = i; -#endif -} - -ERTS_GLB_INLINE void -erts_smp_atomic_dec_relb(erts_smp_atomic_t *decp) -{ -#ifdef ERTS_SMP - erts_atomic_dec_relb(decp); -#else - --(*decp); -#endif -} - -ERTS_GLB_INLINE erts_aint_t -erts_smp_atomic_dectest_relb(erts_smp_atomic_t *decp) -{ -#ifdef ERTS_SMP - return erts_atomic_dectest_relb(decp); -#else - return --(*decp); -#endif -} - -ERTS_GLB_INLINE erts_aint_t -erts_smp_atomic_cmpxchg_acqb(erts_smp_atomic_t *xchgp, - erts_aint_t new, - erts_aint_t exp) -{ -#ifdef ERTS_SMP - return erts_atomic_cmpxchg_acqb(xchgp, new, exp); -#else - erts_aint_t old = *xchgp; - if (old == exp) - *xchgp = new; - return old; -#endif -} - -ERTS_GLB_INLINE erts_aint_t -erts_smp_atomic_cmpxchg_relb(erts_smp_atomic_t *xchgp, - erts_aint_t new, - erts_aint_t exp) -{ -#ifdef ERTS_SMP - return erts_atomic_cmpxchg_relb(xchgp, new, exp); -#else - erts_aint_t old = *xchgp; - if (old == exp) - *xchgp = new; - return old; -#endif -} - -ERTS_GLB_INLINE void -erts_smp_atomic32_init(erts_smp_atomic32_t *var, erts_aint32_t i) -{ -#ifdef ERTS_SMP - erts_atomic32_init(var, i); -#else - *var = i; -#endif -} - -ERTS_GLB_INLINE void -erts_smp_atomic32_set(erts_smp_atomic32_t *var, erts_aint32_t i) -{ -#ifdef ERTS_SMP - erts_atomic32_set(var, i); -#else - *var = i; -#endif -} - -ERTS_GLB_INLINE erts_aint32_t -erts_smp_atomic32_read(erts_smp_atomic32_t *var) -{ -#ifdef ERTS_SMP - return erts_atomic32_read(var); -#else - return *var; -#endif -} - -ERTS_GLB_INLINE erts_aint32_t -erts_smp_atomic32_inctest(erts_smp_atomic32_t *incp) -{ -#ifdef ERTS_SMP - return erts_atomic32_inctest(incp); -#else - return ++(*incp); -#endif -} - -ERTS_GLB_INLINE erts_aint32_t -erts_smp_atomic32_dectest(erts_smp_atomic32_t *decp) -{ -#ifdef ERTS_SMP - return erts_atomic32_dectest(decp); -#else - return --(*decp); -#endif -} - -ERTS_GLB_INLINE void -erts_smp_atomic32_inc(erts_smp_atomic32_t *incp) -{ -#ifdef ERTS_SMP - erts_atomic32_inc(incp); -#else - ++(*incp); -#endif -} - -ERTS_GLB_INLINE void -erts_smp_atomic32_dec(erts_smp_atomic32_t *decp) -{ -#ifdef ERTS_SMP - erts_atomic32_dec(decp); -#else - --(*decp); -#endif -} - -ERTS_GLB_INLINE erts_aint32_t -erts_smp_atomic32_addtest(erts_smp_atomic32_t *addp, erts_aint32_t i) -{ -#ifdef ERTS_SMP - return erts_atomic32_addtest(addp, i); -#else - return *addp += i; -#endif -} - -ERTS_GLB_INLINE void -erts_smp_atomic32_add(erts_smp_atomic32_t *addp, erts_aint32_t i) -{ -#ifdef ERTS_SMP - erts_atomic32_add(addp, i); -#else - *addp += i; -#endif -} - -ERTS_GLB_INLINE erts_aint32_t -erts_smp_atomic32_xchg(erts_smp_atomic32_t *xchgp, erts_aint32_t new) -{ -#ifdef ERTS_SMP - return erts_atomic32_xchg(xchgp, new); -#else - erts_aint32_t old; - old = *xchgp; - *xchgp = new; - return old; -#endif -} - -ERTS_GLB_INLINE erts_aint32_t -erts_smp_atomic32_cmpxchg(erts_smp_atomic32_t *xchgp, - erts_aint32_t new, - erts_aint32_t expected) -{ -#ifdef ERTS_SMP - return erts_atomic32_cmpxchg(xchgp, new, expected); -#else - erts_aint32_t old = *xchgp; - if (old == expected) - *xchgp = new; - return old; -#endif -} - -ERTS_GLB_INLINE erts_aint32_t -erts_smp_atomic32_bor(erts_smp_atomic32_t *var, erts_aint32_t mask) -{ -#ifdef ERTS_SMP - return erts_atomic32_bor(var, mask); -#else - erts_aint32_t old; - old = *var; - *var |= mask; - return old; -#endif -} - -ERTS_GLB_INLINE erts_aint32_t -erts_smp_atomic32_band(erts_smp_atomic32_t *var, erts_aint32_t mask) -{ -#ifdef ERTS_SMP - return erts_atomic32_band(var, mask); -#else - erts_aint32_t old; - old = *var; - *var &= mask; - return old; -#endif -} - -ERTS_GLB_INLINE erts_aint32_t -erts_smp_atomic32_read_acqb(erts_smp_atomic32_t *var) -{ -#ifdef ERTS_SMP - return erts_atomic32_read_acqb(var); -#else - return *var; -#endif -} - -ERTS_GLB_INLINE void -erts_smp_atomic32_set_relb(erts_smp_atomic32_t *var, erts_aint32_t i) -{ -#ifdef ERTS_SMP - erts_atomic32_set_relb(var, i); -#else - *var = i; -#endif -} - -ERTS_GLB_INLINE void -erts_smp_atomic32_dec_relb(erts_smp_atomic32_t *decp) -{ -#ifdef ERTS_SMP - erts_atomic32_dec_relb(decp); -#else - --(*decp); -#endif -} - -ERTS_GLB_INLINE erts_aint32_t -erts_smp_atomic32_dectest_relb(erts_smp_atomic32_t *decp) -{ -#ifdef ERTS_SMP - return erts_atomic32_dectest_relb(decp); -#else - return --(*decp); -#endif -} - -ERTS_GLB_INLINE erts_aint32_t -erts_smp_atomic32_cmpxchg_acqb(erts_smp_atomic32_t *xchgp, - erts_aint32_t new, - erts_aint32_t exp) -{ -#ifdef ERTS_SMP - return erts_atomic32_cmpxchg_acqb(xchgp, new, exp); -#else - erts_aint32_t old = *xchgp; - if (old == exp) - *xchgp = new; - return old; -#endif -} - -ERTS_GLB_INLINE erts_aint32_t -erts_smp_atomic32_cmpxchg_relb(erts_smp_atomic32_t *xchgp, - erts_aint32_t new, - erts_aint32_t exp) -{ -#ifdef ERTS_SMP - return erts_atomic32_cmpxchg_relb(xchgp, new, exp); -#else - erts_aint32_t old = *xchgp; - if (old == exp) - *xchgp = new; - return old; -#endif -} - -ERTS_GLB_INLINE void erts_smp_spinlock_init_x(erts_smp_spinlock_t *lock, char *name, Eterm extra) { #ifdef ERTS_SMP @@ -1308,3 +1229,37 @@ erts_smp_thr_sigwait(const sigset_t *set, int *sig) #endif /* #if ERTS_GLB_INLINE_INCL_FUNC_DEF */ #endif /* ERL_SMP_H */ + +#ifdef ERTS_UNDEF_DEPRECATED_ATOMICS + +/* Deprecated functions to replace */ + +#undef erts_smp_atomic_init +#undef erts_smp_atomic_set +#undef erts_smp_atomic_read +#undef erts_smp_atomic_inctest +#undef erts_smp_atomic_dectest +#undef erts_smp_atomic_inc +#undef erts_smp_atomic_dec +#undef erts_smp_atomic_addtest +#undef erts_smp_atomic_add +#undef erts_smp_atomic_xchg +#undef erts_smp_atomic_cmpxchg +#undef erts_smp_atomic_bor +#undef erts_smp_atomic_band + +#undef erts_smp_atomic32_init +#undef erts_smp_atomic32_set +#undef erts_smp_atomic32_read +#undef erts_smp_atomic32_inctest +#undef erts_smp_atomic32_dectest +#undef erts_smp_atomic32_inc +#undef erts_smp_atomic32_dec +#undef erts_smp_atomic32_addtest +#undef erts_smp_atomic32_add +#undef erts_smp_atomic32_xchg +#undef erts_smp_atomic32_cmpxchg +#undef erts_smp_atomic32_bor +#undef erts_smp_atomic32_band + +#endif diff --git a/erts/emulator/beam/erl_threads.h b/erts/emulator/beam/erl_threads.h index 8c9cace0c5..12eaf39ec7 100644 --- a/erts/emulator/beam/erl_threads.h +++ b/erts/emulator/beam/erl_threads.h @@ -28,6 +28,11 @@ #define ERTS_SPIN_BODY ETHR_SPIN_BODY #include "sys.h" + +typedef struct { SWord sint[2]; } erts_no_dw_atomic_t; +typedef SWord erts_no_atomic_t; +typedef Sint32 erts_no_atomic32_t; + #ifdef USE_THREADS #define ETHR_TRY_INLINE_FUNCS @@ -99,10 +104,12 @@ typedef ethr_rwmutex_opt erts_rwmtx_opt_t; typedef ethr_tsd_key erts_tsd_key_t; typedef ethr_ts_event erts_tse_t; -typedef ethr_sint_t erts_aint_t; -typedef ethr_atomic_t erts_atomic_t; -typedef ethr_sint32_t erts_aint32_t; -typedef ethr_atomic32_t erts_atomic32_t; +#define erts_dw_aint_t ethr_dw_sint_t +#define erts_dw_atomic_t ethr_dw_atomic_t +#define erts_aint_t ethr_sint_t +#define erts_atomic_t ethr_atomic_t +#define erts_aint32_t ethr_sint32_t +#define erts_atomic32_t ethr_atomic32_t /* spinlock */ typedef struct { @@ -164,10 +171,12 @@ typedef struct { typedef int erts_rwmtx_t; typedef int erts_tsd_key_t; typedef int erts_tse_t; -typedef SWord erts_aint_t; -typedef SWord erts_atomic_t; -typedef SWord erts_aint32_t; -typedef SWord erts_atomic32_t; +#define erts_dw_aint_t erts_no_dw_atomic_t +#define erts_dw_atomic_t erts_no_dw_atomic_t +#define erts_aint_t SWord +#define erts_atomic_t erts_no_atomic_t +#define erts_aint32_t Sint32 +#define erts_atomic32_t erts_no_atomic32_t #if __GNUC__ > 2 typedef struct { } erts_spinlock_t; typedef struct { } erts_rwlock_t; @@ -247,65 +256,51 @@ ERTS_GLB_INLINE int erts_rwmtx_tryrwlock(erts_rwmtx_t *rwmtx); ERTS_GLB_INLINE void erts_rwmtx_rwunlock(erts_rwmtx_t *rwmtx); ERTS_GLB_INLINE int erts_lc_rwmtx_is_rlocked(erts_rwmtx_t *mtx); ERTS_GLB_INLINE int erts_lc_rwmtx_is_rwlocked(erts_rwmtx_t *mtx); -ERTS_GLB_INLINE void erts_atomic_init(erts_atomic_t *var, erts_aint_t i); -ERTS_GLB_INLINE void erts_atomic_set(erts_atomic_t *var, erts_aint_t i); -ERTS_GLB_INLINE erts_aint_t erts_atomic_read(erts_atomic_t *var); -ERTS_GLB_INLINE erts_aint_t erts_atomic_inctest(erts_atomic_t *incp); -ERTS_GLB_INLINE erts_aint_t erts_atomic_dectest(erts_atomic_t *decp); -ERTS_GLB_INLINE void erts_atomic_inc(erts_atomic_t *incp); -ERTS_GLB_INLINE void erts_atomic_dec(erts_atomic_t *decp); -ERTS_GLB_INLINE erts_aint_t erts_atomic_addtest(erts_atomic_t *addp, - erts_aint_t i); -ERTS_GLB_INLINE void erts_atomic_add(erts_atomic_t *addp, erts_aint_t i); -ERTS_GLB_INLINE erts_aint_t erts_atomic_xchg(erts_atomic_t *xchgp, - erts_aint_t new); -ERTS_GLB_INLINE erts_aint_t erts_atomic_cmpxchg(erts_atomic_t *xchgp, - erts_aint_t new, - erts_aint_t expected); -ERTS_GLB_INLINE erts_aint_t erts_atomic_bor(erts_atomic_t *var, - erts_aint_t mask); -ERTS_GLB_INLINE erts_aint_t erts_atomic_band(erts_atomic_t *var, - erts_aint_t mask); -ERTS_GLB_INLINE erts_aint_t erts_atomic_read_acqb(erts_atomic_t *var); -ERTS_GLB_INLINE void erts_atomic_set_relb(erts_atomic_t *var, erts_aint_t i); -ERTS_GLB_INLINE void erts_atomic_dec_relb(erts_atomic_t *decp); -ERTS_GLB_INLINE erts_aint_t erts_atomic_dectest_relb(erts_atomic_t *decp); -ERTS_GLB_INLINE erts_aint_t erts_atomic_cmpxchg_acqb(erts_atomic_t *xchgp, - erts_aint_t new, - erts_aint_t exp); -ERTS_GLB_INLINE erts_aint_t erts_atomic_cmpxchg_relb(erts_atomic_t *xchgp, - erts_aint_t new, - erts_aint_t exp); -ERTS_GLB_INLINE void erts_atomic32_init(erts_atomic32_t *var, erts_aint32_t i); -ERTS_GLB_INLINE void erts_atomic32_set(erts_atomic32_t *var, erts_aint32_t i); -ERTS_GLB_INLINE erts_aint32_t erts_atomic32_read(erts_atomic32_t *var); -ERTS_GLB_INLINE erts_aint32_t erts_atomic32_inctest(erts_atomic32_t *incp); -ERTS_GLB_INLINE erts_aint32_t erts_atomic32_dectest(erts_atomic32_t *decp); -ERTS_GLB_INLINE void erts_atomic32_inc(erts_atomic32_t *incp); -ERTS_GLB_INLINE void erts_atomic32_dec(erts_atomic32_t *decp); -ERTS_GLB_INLINE erts_aint32_t erts_atomic32_addtest(erts_atomic32_t *addp, - erts_aint32_t i); -ERTS_GLB_INLINE void erts_atomic32_add(erts_atomic32_t *addp, erts_aint32_t i); -ERTS_GLB_INLINE erts_aint32_t erts_atomic32_xchg(erts_atomic32_t *xchgp, - erts_aint32_t new); -ERTS_GLB_INLINE erts_aint32_t erts_atomic32_cmpxchg(erts_atomic32_t *xchgp, - erts_aint32_t new, - erts_aint32_t expected); -ERTS_GLB_INLINE erts_aint32_t erts_atomic32_bor(erts_atomic32_t *var, - erts_aint32_t mask); -ERTS_GLB_INLINE erts_aint32_t erts_atomic32_band(erts_atomic32_t *var, - erts_aint32_t mask); -ERTS_GLB_INLINE erts_aint32_t erts_atomic32_read_acqb(erts_atomic32_t *var); -ERTS_GLB_INLINE void erts_atomic32_set_relb(erts_atomic32_t *var, - erts_aint32_t i); -ERTS_GLB_INLINE void erts_atomic32_dec_relb(erts_atomic32_t *decp); -ERTS_GLB_INLINE erts_aint32_t erts_atomic32_dectest_relb(erts_atomic32_t *decp); -ERTS_GLB_INLINE erts_aint32_t erts_atomic32_cmpxchg_acqb(erts_atomic32_t *xchgp, - erts_aint32_t new, - erts_aint32_t exp); -ERTS_GLB_INLINE erts_aint32_t erts_atomic32_cmpxchg_relb(erts_atomic32_t *xchgp, - erts_aint32_t new, - erts_aint32_t exp); + +ERTS_GLB_INLINE void erts_no_dw_atomic_set(erts_no_dw_atomic_t *var, erts_no_dw_atomic_t *val); +ERTS_GLB_INLINE void erts_no_dw_atomic_read(erts_no_dw_atomic_t *var, erts_no_dw_atomic_t *val); +ERTS_GLB_INLINE int erts_no_dw_atomic_cmpxchg(erts_no_dw_atomic_t *var, + erts_no_dw_atomic_t *val, + erts_no_dw_atomic_t *old_val); +ERTS_GLB_INLINE void erts_no_atomic_set(erts_no_atomic_t *var, erts_aint_t i); +ERTS_GLB_INLINE erts_aint_t erts_no_atomic_read(erts_no_atomic_t *var); +ERTS_GLB_INLINE erts_aint_t erts_no_atomic_inc_read(erts_no_atomic_t *incp); +ERTS_GLB_INLINE erts_aint_t erts_no_atomic_dec_read(erts_no_atomic_t *decp); +ERTS_GLB_INLINE void erts_no_atomic_inc(erts_no_atomic_t *incp); +ERTS_GLB_INLINE void erts_no_atomic_dec(erts_no_atomic_t *decp); +ERTS_GLB_INLINE erts_aint_t erts_no_atomic_add_read(erts_no_atomic_t *addp, + erts_aint_t i); +ERTS_GLB_INLINE void erts_no_atomic_add(erts_no_atomic_t *addp, erts_aint_t i); +ERTS_GLB_INLINE erts_aint_t erts_no_atomic_read_bor(erts_no_atomic_t *var, + erts_aint_t mask); +ERTS_GLB_INLINE erts_aint_t erts_no_atomic_read_band(erts_no_atomic_t *var, + erts_aint_t mask); +ERTS_GLB_INLINE erts_aint_t erts_no_atomic_xchg(erts_no_atomic_t *xchgp, + erts_aint_t new); +ERTS_GLB_INLINE erts_aint_t erts_no_atomic_cmpxchg(erts_no_atomic_t *xchgp, + erts_aint_t new, + erts_aint_t expected); +ERTS_GLB_INLINE void erts_no_atomic32_set(erts_no_atomic32_t *var, + erts_aint32_t i); +ERTS_GLB_INLINE erts_aint32_t erts_no_atomic32_read(erts_no_atomic32_t *var); +ERTS_GLB_INLINE erts_aint32_t erts_no_atomic32_inc_read(erts_no_atomic32_t *incp); +ERTS_GLB_INLINE erts_aint32_t erts_no_atomic32_dec_read(erts_no_atomic32_t *decp); +ERTS_GLB_INLINE void erts_no_atomic32_inc(erts_no_atomic32_t *incp); +ERTS_GLB_INLINE void erts_no_atomic32_dec(erts_no_atomic32_t *decp); +ERTS_GLB_INLINE erts_aint32_t erts_no_atomic32_add_read(erts_no_atomic32_t *addp, + erts_aint32_t i); +ERTS_GLB_INLINE void erts_no_atomic32_add(erts_no_atomic32_t *addp, + erts_aint32_t i); +ERTS_GLB_INLINE erts_aint32_t erts_no_atomic32_read_bor(erts_no_atomic32_t *var, + erts_aint32_t mask); +ERTS_GLB_INLINE erts_aint32_t erts_no_atomic32_read_band(erts_no_atomic32_t *var, + erts_aint32_t mask); +ERTS_GLB_INLINE erts_aint32_t erts_no_atomic32_xchg(erts_no_atomic32_t *xchgp, + erts_aint32_t new); +ERTS_GLB_INLINE erts_aint32_t erts_no_atomic32_cmpxchg(erts_no_atomic32_t *xchgp, + erts_aint32_t new, + erts_aint32_t expected); + ERTS_GLB_INLINE void erts_spinlock_init_x_opt(erts_spinlock_t *lock, char *name, Eterm extra, @@ -362,6 +357,430 @@ ERTS_GLB_INLINE void erts_thr_sigmask(int how, const sigset_t *set, ERTS_GLB_INLINE void erts_thr_sigwait(const sigset_t *set, int *sig); #endif /* #ifdef HAVE_ETHR_SIG_FUNCS */ +/* + * Functions implementing atomic operations with with no (nob), + * full (mb), acquire (acqb), release (relb), read (rb), and + * write (wb) memory barriers. + * + * If thread support has been disabled, they are mapped to + * functions that performs the same operation, but aren't atomic + * and don't imply memory barriers. + */ + +#ifdef USE_THREADS + +/* Double word size atomics */ + +#define erts_dw_atomic_init_nob ethr_dw_atomic_init +#define erts_dw_atomic_set_nob ethr_dw_atomic_set +#define erts_dw_atomic_read_nob ethr_dw_atomic_read +#define erts_dw_atomic_cmpxchg_nob ethr_dw_atomic_cmpxchg + +#define erts_dw_atomic_init_mb ethr_dw_atomic_init_mb +#define erts_dw_atomic_set_mb ethr_dw_atomic_set_mb +#define erts_dw_atomic_read_mb ethr_dw_atomic_read_mb +#define erts_dw_atomic_cmpxchg_mb ethr_dw_atomic_cmpxchg_mb + +#define erts_dw_atomic_init_acqb ethr_dw_atomic_init_acqb +#define erts_dw_atomic_set_acqb ethr_dw_atomic_set_acqb +#define erts_dw_atomic_read_acqb ethr_dw_atomic_read_acqb +#define erts_dw_atomic_cmpxchg_acqb ethr_dw_atomic_cmpxchg_acqb + +#define erts_dw_atomic_init_relb ethr_dw_atomic_init_relb +#define erts_dw_atomic_set_relb ethr_dw_atomic_set_relb +#define erts_dw_atomic_read_relb ethr_dw_atomic_read_relb +#define erts_dw_atomic_cmpxchg_relb ethr_dw_atomic_cmpxchg_relb + +#define erts_dw_atomic_init_rb ethr_dw_atomic_init_rb +#define erts_dw_atomic_set_rb ethr_dw_atomic_set_rb +#define erts_dw_atomic_read_rb ethr_dw_atomic_read_rb +#define erts_dw_atomic_cmpxchg_rb ethr_dw_atomic_cmpxchg_rb + +#define erts_dw_atomic_init_wb ethr_dw_atomic_init_wb +#define erts_dw_atomic_set_wb ethr_dw_atomic_set_wb +#define erts_dw_atomic_read_wb ethr_dw_atomic_read_wb +#define erts_dw_atomic_cmpxchg_wb ethr_dw_atomic_cmpxchg_wb + +/* Word size atomics */ + +#define erts_atomic_init_nob ethr_atomic_init +#define erts_atomic_set_nob ethr_atomic_set +#define erts_atomic_read_nob ethr_atomic_read +#define erts_atomic_inc_read_nob ethr_atomic_inc_read +#define erts_atomic_dec_read_nob ethr_atomic_dec_read +#define erts_atomic_inc_nob ethr_atomic_inc +#define erts_atomic_dec_nob ethr_atomic_dec +#define erts_atomic_add_read_nob ethr_atomic_add_read +#define erts_atomic_add_nob ethr_atomic_add +#define erts_atomic_read_bor_nob ethr_atomic_read_bor +#define erts_atomic_read_band_nob ethr_atomic_read_band +#define erts_atomic_xchg_nob ethr_atomic_xchg +#define erts_atomic_cmpxchg_nob ethr_atomic_cmpxchg + +#define erts_atomic_init_mb ethr_atomic_init_mb +#define erts_atomic_set_mb ethr_atomic_set_mb +#define erts_atomic_read_mb ethr_atomic_read_mb +#define erts_atomic_inc_read_mb ethr_atomic_inc_read_mb +#define erts_atomic_dec_read_mb ethr_atomic_dec_read_mb +#define erts_atomic_inc_mb ethr_atomic_inc_mb +#define erts_atomic_dec_mb ethr_atomic_dec_mb +#define erts_atomic_add_read_mb ethr_atomic_add_read_mb +#define erts_atomic_add_mb ethr_atomic_add_mb +#define erts_atomic_read_bor_mb ethr_atomic_read_bor_mb +#define erts_atomic_read_band_mb ethr_atomic_read_band_mb +#define erts_atomic_xchg_mb ethr_atomic_xchg_mb +#define erts_atomic_cmpxchg_mb ethr_atomic_cmpxchg_mb + +#define erts_atomic_init_acqb ethr_atomic_init_acqb +#define erts_atomic_set_acqb ethr_atomic_set_acqb +#define erts_atomic_read_acqb ethr_atomic_read_acqb +#define erts_atomic_inc_read_acqb ethr_atomic_inc_read_acqb +#define erts_atomic_dec_read_acqb ethr_atomic_dec_read_acqb +#define erts_atomic_inc_acqb ethr_atomic_inc_acqb +#define erts_atomic_dec_acqb ethr_atomic_dec_acqb +#define erts_atomic_add_read_acqb ethr_atomic_add_read_acqb +#define erts_atomic_add_acqb ethr_atomic_add_acqb +#define erts_atomic_read_bor_acqb ethr_atomic_read_bor_acqb +#define erts_atomic_read_band_acqb ethr_atomic_read_band_acqb +#define erts_atomic_xchg_acqb ethr_atomic_xchg_acqb +#define erts_atomic_cmpxchg_acqb ethr_atomic_cmpxchg_acqb + +#define erts_atomic_init_relb ethr_atomic_init_relb +#define erts_atomic_set_relb ethr_atomic_set_relb +#define erts_atomic_read_relb ethr_atomic_read_relb +#define erts_atomic_inc_read_relb ethr_atomic_inc_read_relb +#define erts_atomic_dec_read_relb ethr_atomic_dec_read_relb +#define erts_atomic_inc_relb ethr_atomic_inc_relb +#define erts_atomic_dec_relb ethr_atomic_dec_relb +#define erts_atomic_add_read_relb ethr_atomic_add_read_relb +#define erts_atomic_add_relb ethr_atomic_add_relb +#define erts_atomic_read_bor_relb ethr_atomic_read_bor_relb +#define erts_atomic_read_band_relb ethr_atomic_read_band_relb +#define erts_atomic_xchg_relb ethr_atomic_xchg_relb +#define erts_atomic_cmpxchg_relb ethr_atomic_cmpxchg_relb + +#define erts_atomic_init_rb ethr_atomic_init_rb +#define erts_atomic_set_rb ethr_atomic_set_rb +#define erts_atomic_read_rb ethr_atomic_read_rb +#define erts_atomic_inc_read_rb ethr_atomic_inc_read_rb +#define erts_atomic_dec_read_rb ethr_atomic_dec_read_rb +#define erts_atomic_inc_rb ethr_atomic_inc_rb +#define erts_atomic_dec_rb ethr_atomic_dec_rb +#define erts_atomic_add_read_rb ethr_atomic_add_read_rb +#define erts_atomic_add_rb ethr_atomic_add_rb +#define erts_atomic_read_bor_rb ethr_atomic_read_bor_rb +#define erts_atomic_read_band_rb ethr_atomic_read_band_rb +#define erts_atomic_xchg_rb ethr_atomic_xchg_rb +#define erts_atomic_cmpxchg_rb ethr_atomic_cmpxchg_rb + +#define erts_atomic_init_wb ethr_atomic_init_wb +#define erts_atomic_set_wb ethr_atomic_set_wb +#define erts_atomic_read_wb ethr_atomic_read_wb +#define erts_atomic_inc_read_wb ethr_atomic_inc_read_wb +#define erts_atomic_dec_read_wb ethr_atomic_dec_read_wb +#define erts_atomic_inc_wb ethr_atomic_inc_wb +#define erts_atomic_dec_wb ethr_atomic_dec_wb +#define erts_atomic_add_read_wb ethr_atomic_add_read_wb +#define erts_atomic_add_wb ethr_atomic_add_wb +#define erts_atomic_read_bor_wb ethr_atomic_read_bor_wb +#define erts_atomic_read_band_wb ethr_atomic_read_band_wb +#define erts_atomic_xchg_wb ethr_atomic_xchg_wb +#define erts_atomic_cmpxchg_wb ethr_atomic_cmpxchg_wb + +/* 32-bit atomics */ + +#define erts_atomic32_init_nob ethr_atomic32_init +#define erts_atomic32_set_nob ethr_atomic32_set +#define erts_atomic32_read_nob ethr_atomic32_read +#define erts_atomic32_inc_read_nob ethr_atomic32_inc_read +#define erts_atomic32_dec_read_nob ethr_atomic32_dec_read +#define erts_atomic32_inc_nob ethr_atomic32_inc +#define erts_atomic32_dec_nob ethr_atomic32_dec +#define erts_atomic32_add_read_nob ethr_atomic32_add_read +#define erts_atomic32_add_nob ethr_atomic32_add +#define erts_atomic32_read_bor_nob ethr_atomic32_read_bor +#define erts_atomic32_read_band_nob ethr_atomic32_read_band +#define erts_atomic32_xchg_nob ethr_atomic32_xchg +#define erts_atomic32_cmpxchg_nob ethr_atomic32_cmpxchg + +#define erts_atomic32_init_mb ethr_atomic32_init_mb +#define erts_atomic32_set_mb ethr_atomic32_set_mb +#define erts_atomic32_read_mb ethr_atomic32_read_mb +#define erts_atomic32_inc_read_mb ethr_atomic32_inc_read_mb +#define erts_atomic32_dec_read_mb ethr_atomic32_dec_read_mb +#define erts_atomic32_inc_mb ethr_atomic32_inc_mb +#define erts_atomic32_dec_mb ethr_atomic32_dec_mb +#define erts_atomic32_add_read_mb ethr_atomic32_add_read_mb +#define erts_atomic32_add_mb ethr_atomic32_add_mb +#define erts_atomic32_read_bor_mb ethr_atomic32_read_bor_mb +#define erts_atomic32_read_band_mb ethr_atomic32_read_band_mb +#define erts_atomic32_xchg_mb ethr_atomic32_xchg_mb +#define erts_atomic32_cmpxchg_mb ethr_atomic32_cmpxchg_mb + +#define erts_atomic32_init_acqb ethr_atomic32_init_acqb +#define erts_atomic32_set_acqb ethr_atomic32_set_acqb +#define erts_atomic32_read_acqb ethr_atomic32_read_acqb +#define erts_atomic32_inc_read_acqb ethr_atomic32_inc_read_acqb +#define erts_atomic32_dec_read_acqb ethr_atomic32_dec_read_acqb +#define erts_atomic32_inc_acqb ethr_atomic32_inc_acqb +#define erts_atomic32_dec_acqb ethr_atomic32_dec_acqb +#define erts_atomic32_add_read_acqb ethr_atomic32_add_read_acqb +#define erts_atomic32_add_acqb ethr_atomic32_add_acqb +#define erts_atomic32_read_bor_acqb ethr_atomic32_read_bor_acqb +#define erts_atomic32_read_band_acqb ethr_atomic32_read_band_acqb +#define erts_atomic32_xchg_acqb ethr_atomic32_xchg_acqb +#define erts_atomic32_cmpxchg_acqb ethr_atomic32_cmpxchg_acqb + +#define erts_atomic32_init_relb ethr_atomic32_init_relb +#define erts_atomic32_set_relb ethr_atomic32_set_relb +#define erts_atomic32_read_relb ethr_atomic32_read_relb +#define erts_atomic32_inc_read_relb ethr_atomic32_inc_read_relb +#define erts_atomic32_dec_read_relb ethr_atomic32_dec_read_relb +#define erts_atomic32_inc_relb ethr_atomic32_inc_relb +#define erts_atomic32_dec_relb ethr_atomic32_dec_relb +#define erts_atomic32_add_read_relb ethr_atomic32_add_read_relb +#define erts_atomic32_add_relb ethr_atomic32_add_relb +#define erts_atomic32_read_bor_relb ethr_atomic32_read_bor_relb +#define erts_atomic32_read_band_relb ethr_atomic32_read_band_relb +#define erts_atomic32_xchg_relb ethr_atomic32_xchg_relb +#define erts_atomic32_cmpxchg_relb ethr_atomic32_cmpxchg_relb + +#define erts_atomic32_init_rb ethr_atomic32_init_rb +#define erts_atomic32_set_rb ethr_atomic32_set_rb +#define erts_atomic32_read_rb ethr_atomic32_read_rb +#define erts_atomic32_inc_read_rb ethr_atomic32_inc_read_rb +#define erts_atomic32_dec_read_rb ethr_atomic32_dec_read_rb +#define erts_atomic32_inc_rb ethr_atomic32_inc_rb +#define erts_atomic32_dec_rb ethr_atomic32_dec_rb +#define erts_atomic32_add_read_rb ethr_atomic32_add_read_rb +#define erts_atomic32_add_rb ethr_atomic32_add_rb +#define erts_atomic32_read_bor_rb ethr_atomic32_read_bor_rb +#define erts_atomic32_read_band_rb ethr_atomic32_read_band_rb +#define erts_atomic32_xchg_rb ethr_atomic32_xchg_rb +#define erts_atomic32_cmpxchg_rb ethr_atomic32_cmpxchg_rb + +#define erts_atomic32_init_wb ethr_atomic32_init_wb +#define erts_atomic32_set_wb ethr_atomic32_set_wb +#define erts_atomic32_read_wb ethr_atomic32_read_wb +#define erts_atomic32_inc_read_wb ethr_atomic32_inc_read_wb +#define erts_atomic32_dec_read_wb ethr_atomic32_dec_read_wb +#define erts_atomic32_inc_wb ethr_atomic32_inc_wb +#define erts_atomic32_dec_wb ethr_atomic32_dec_wb +#define erts_atomic32_add_read_wb ethr_atomic32_add_read_wb +#define erts_atomic32_add_wb ethr_atomic32_add_wb +#define erts_atomic32_read_bor_wb ethr_atomic32_read_bor_wb +#define erts_atomic32_read_band_wb ethr_atomic32_read_band_wb +#define erts_atomic32_xchg_wb ethr_atomic32_xchg_wb +#define erts_atomic32_cmpxchg_wb ethr_atomic32_cmpxchg_wb + +#else /* !USE_THREADS */ + +/* Double word size atomics */ + +#define erts_dw_atomic_init_nob erts_no_dw_atomic_set +#define erts_dw_atomic_set_nob erts_no_dw_atomic_set +#define erts_dw_atomic_read_nob erts_no_dw_atomic_read +#define erts_dw_atomic_cmpxchg_nob erts_no_dw_atomic_cmpxchg + +#define erts_dw_atomic_init_mb erts_no_dw_atomic_init +#define erts_dw_atomic_set_mb erts_no_dw_atomic_set +#define erts_dw_atomic_read_mb erts_no_dw_atomic_read +#define erts_dw_atomic_cmpxchg_mb erts_no_dw_atomic_cmpxchg + +#define erts_dw_atomic_init_acqb erts_no_dw_atomic_init +#define erts_dw_atomic_set_acqb erts_no_dw_atomic_set +#define erts_dw_atomic_read_acqb erts_no_dw_atomic_read +#define erts_dw_atomic_cmpxchg_acqb erts_no_dw_atomic_cmpxchg + +#define erts_dw_atomic_init_relb erts_no_dw_atomic_init +#define erts_dw_atomic_set_relb erts_no_dw_atomic_set +#define erts_dw_atomic_read_relb erts_no_dw_atomic_read +#define erts_dw_atomic_cmpxchg_relb erts_no_dw_atomic_cmpxchg + +#define erts_dw_atomic_init_rb erts_no_dw_atomic_init +#define erts_dw_atomic_set_rb erts_no_dw_atomic_set +#define erts_dw_atomic_read_rb erts_no_dw_atomic_read +#define erts_dw_atomic_cmpxchg_rb erts_no_dw_atomic_cmpxchg + +#define erts_dw_atomic_init_wb erts_no_dw_atomic_init +#define erts_dw_atomic_set_wb erts_no_dw_atomic_set +#define erts_dw_atomic_read_wb erts_no_dw_atomic_read +#define erts_dw_atomic_cmpxchg_wb erts_no_dw_atomic_cmpxchg + +/* Word size atomics */ + +#define erts_atomic_init_nob erts_no_atomic_set +#define erts_atomic_set_nob erts_no_atomic_set +#define erts_atomic_read_nob erts_no_atomic_read +#define erts_atomic_inc_read_nob erts_no_atomic_inc_read +#define erts_atomic_dec_read_nob erts_no_atomic_dec_read +#define erts_atomic_inc_nob erts_no_atomic_inc +#define erts_atomic_dec_nob erts_no_atomic_dec +#define erts_atomic_add_read_nob erts_no_atomic_add_read +#define erts_atomic_add_nob erts_no_atomic_add +#define erts_atomic_read_bor_nob erts_no_atomic_read_bor +#define erts_atomic_read_band_nob erts_no_atomic_read_band +#define erts_atomic_xchg_nob erts_no_atomic_xchg +#define erts_atomic_cmpxchg_nob erts_no_atomic_cmpxchg + +#define erts_atomic_init_mb erts_no_atomic_set +#define erts_atomic_set_mb erts_no_atomic_set +#define erts_atomic_read_mb erts_no_atomic_read +#define erts_atomic_inc_read_mb erts_no_atomic_inc_read +#define erts_atomic_dec_read_mb erts_no_atomic_dec_read +#define erts_atomic_inc_mb erts_no_atomic_inc +#define erts_atomic_dec_mb erts_no_atomic_dec +#define erts_atomic_add_read_mb erts_no_atomic_add_read +#define erts_atomic_add_mb erts_no_atomic_add +#define erts_atomic_read_bor_mb erts_no_atomic_read_bor +#define erts_atomic_read_band_mb erts_no_atomic_read_band +#define erts_atomic_xchg_mb erts_no_atomic_xchg +#define erts_atomic_cmpxchg_mb erts_no_atomic_cmpxchg + +#define erts_atomic_init_acqb erts_no_atomic_set +#define erts_atomic_set_acqb erts_no_atomic_set +#define erts_atomic_read_acqb erts_no_atomic_read +#define erts_atomic_inc_read_acqb erts_no_atomic_inc_read +#define erts_atomic_dec_read_acqb erts_no_atomic_dec_read +#define erts_atomic_inc_acqb erts_no_atomic_inc +#define erts_atomic_dec_acqb erts_no_atomic_dec +#define erts_atomic_add_read_acqb erts_no_atomic_add_read +#define erts_atomic_add_acqb erts_no_atomic_add +#define erts_atomic_read_bor_acqb erts_no_atomic_read_bor +#define erts_atomic_read_band_acqb erts_no_atomic_read_band +#define erts_atomic_xchg_acqb erts_no_atomic_xchg +#define erts_atomic_cmpxchg_acqb erts_no_atomic_cmpxchg + +#define erts_atomic_init_relb erts_no_atomic_set +#define erts_atomic_set_relb erts_no_atomic_set +#define erts_atomic_read_relb erts_no_atomic_read +#define erts_atomic_inc_read_relb erts_no_atomic_inc_read +#define erts_atomic_dec_read_relb erts_no_atomic_dec_read +#define erts_atomic_inc_relb erts_no_atomic_inc +#define erts_atomic_dec_relb erts_no_atomic_dec +#define erts_atomic_add_read_relb erts_no_atomic_add_read +#define erts_atomic_add_relb erts_no_atomic_add +#define erts_atomic_read_bor_relb erts_no_atomic_read_bor +#define erts_atomic_read_band_relb erts_no_atomic_read_band +#define erts_atomic_xchg_relb erts_no_atomic_xchg +#define erts_atomic_cmpxchg_relb erts_no_atomic_cmpxchg + +#define erts_atomic_init_rb erts_no_atomic_set +#define erts_atomic_set_rb erts_no_atomic_set +#define erts_atomic_read_rb erts_no_atomic_read +#define erts_atomic_inc_read_rb erts_no_atomic_inc_read +#define erts_atomic_dec_read_rb erts_no_atomic_dec_read +#define erts_atomic_inc_rb erts_no_atomic_inc +#define erts_atomic_dec_rb erts_no_atomic_dec +#define erts_atomic_add_read_rb erts_no_atomic_add_read +#define erts_atomic_add_rb erts_no_atomic_add +#define erts_atomic_read_bor_rb erts_no_atomic_read_bor +#define erts_atomic_read_band_rb erts_no_atomic_read_band +#define erts_atomic_xchg_rb erts_no_atomic_xchg +#define erts_atomic_cmpxchg_rb erts_no_atomic_cmpxchg + +#define erts_atomic_init_wb erts_no_atomic_set +#define erts_atomic_set_wb erts_no_atomic_set +#define erts_atomic_read_wb erts_no_atomic_read +#define erts_atomic_inc_read_wb erts_no_atomic_inc_read +#define erts_atomic_dec_read_wb erts_no_atomic_dec_read +#define erts_atomic_inc_wb erts_no_atomic_inc +#define erts_atomic_dec_wb erts_no_atomic_dec +#define erts_atomic_add_read_wb erts_no_atomic_add_read +#define erts_atomic_add_wb erts_no_atomic_add +#define erts_atomic_read_bor_wb erts_no_atomic_read_bor +#define erts_atomic_read_band_wb erts_no_atomic_read_band +#define erts_atomic_xchg_wb erts_no_atomic_xchg +#define erts_atomic_cmpxchg_wb erts_no_atomic_cmpxchg + +/* 32-bit atomics */ + +#define erts_atomic32_init_nob erts_no_atomic32_set +#define erts_atomic32_set_nob erts_no_atomic32_set +#define erts_atomic32_read_nob erts_no_atomic32_read +#define erts_atomic32_inc_read_nob erts_no_atomic32_inc_read +#define erts_atomic32_dec_read_nob erts_no_atomic32_dec_read +#define erts_atomic32_inc_nob erts_no_atomic32_inc +#define erts_atomic32_dec_nob erts_no_atomic32_dec +#define erts_atomic32_add_read_nob erts_no_atomic32_add_read +#define erts_atomic32_add_nob erts_no_atomic32_add +#define erts_atomic32_read_bor_nob erts_no_atomic32_read_bor +#define erts_atomic32_read_band_nob erts_no_atomic32_read_band +#define erts_atomic32_xchg_nob erts_no_atomic32_xchg +#define erts_atomic32_cmpxchg_nob erts_no_atomic32_cmpxchg + +#define erts_atomic32_init_mb erts_no_atomic32_set +#define erts_atomic32_set_mb erts_no_atomic32_set +#define erts_atomic32_read_mb erts_no_atomic32_read +#define erts_atomic32_inc_read_mb erts_no_atomic32_inc_read +#define erts_atomic32_dec_read_mb erts_no_atomic32_dec_read +#define erts_atomic32_inc_mb erts_no_atomic32_inc +#define erts_atomic32_dec_mb erts_no_atomic32_dec +#define erts_atomic32_add_read_mb erts_no_atomic32_add_read +#define erts_atomic32_add_mb erts_no_atomic32_add +#define erts_atomic32_read_bor_mb erts_no_atomic32_read_bor +#define erts_atomic32_read_band_mb erts_no_atomic32_read_band +#define erts_atomic32_xchg_mb erts_no_atomic32_xchg +#define erts_atomic32_cmpxchg_mb erts_no_atomic32_cmpxchg + +#define erts_atomic32_init_acqb erts_no_atomic32_set +#define erts_atomic32_set_acqb erts_no_atomic32_set +#define erts_atomic32_read_acqb erts_no_atomic32_read +#define erts_atomic32_inc_read_acqb erts_no_atomic32_inc_read +#define erts_atomic32_dec_read_acqb erts_no_atomic32_dec_read +#define erts_atomic32_inc_acqb erts_no_atomic32_inc +#define erts_atomic32_dec_acqb erts_no_atomic32_dec +#define erts_atomic32_add_read_acqb erts_no_atomic32_add_read +#define erts_atomic32_add_acqb erts_no_atomic32_add +#define erts_atomic32_read_bor_acqb erts_no_atomic32_read_bor +#define erts_atomic32_read_band_acqb erts_no_atomic32_read_band +#define erts_atomic32_xchg_acqb erts_no_atomic32_xchg +#define erts_atomic32_cmpxchg_acqb erts_no_atomic32_cmpxchg + +#define erts_atomic32_init_relb erts_no_atomic32_set +#define erts_atomic32_set_relb erts_no_atomic32_set +#define erts_atomic32_read_relb erts_no_atomic32_read +#define erts_atomic32_inc_read_relb erts_no_atomic32_inc_read +#define erts_atomic32_dec_read_relb erts_no_atomic32_dec_read +#define erts_atomic32_inc_relb erts_no_atomic32_inc +#define erts_atomic32_dec_relb erts_no_atomic32_dec +#define erts_atomic32_add_read_relb erts_no_atomic32_add_read +#define erts_atomic32_add_relb erts_no_atomic32_add +#define erts_atomic32_read_bor_relb erts_no_atomic32_read_bor +#define erts_atomic32_read_band_relb erts_no_atomic32_read_band +#define erts_atomic32_xchg_relb erts_no_atomic32_xchg +#define erts_atomic32_cmpxchg_relb erts_no_atomic32_cmpxchg + +#define erts_atomic32_init_rb erts_no_atomic32_set +#define erts_atomic32_set_rb erts_no_atomic32_set +#define erts_atomic32_read_rb erts_no_atomic32_read +#define erts_atomic32_inc_read_rb erts_no_atomic32_inc_read +#define erts_atomic32_dec_read_rb erts_no_atomic32_dec_read +#define erts_atomic32_inc_rb erts_no_atomic32_inc +#define erts_atomic32_dec_rb erts_no_atomic32_dec +#define erts_atomic32_add_read_rb erts_no_atomic32_add_read +#define erts_atomic32_add_rb erts_no_atomic32_add +#define erts_atomic32_read_bor_rb erts_no_atomic32_read_bor +#define erts_atomic32_read_band_rb erts_no_atomic32_read_band +#define erts_atomic32_xchg_rb erts_no_atomic32_xchg +#define erts_atomic32_cmpxchg_rb erts_no_atomic32_cmpxchg + +#define erts_atomic32_init_wb erts_no_atomic32_set +#define erts_atomic32_set_wb erts_no_atomic32_set +#define erts_atomic32_read_wb erts_no_atomic32_read +#define erts_atomic32_inc_read_wb erts_no_atomic32_inc_read +#define erts_atomic32_dec_read_wb erts_no_atomic32_dec_read +#define erts_atomic32_inc_wb erts_no_atomic32_inc +#define erts_atomic32_dec_wb erts_no_atomic32_dec +#define erts_atomic32_add_read_wb erts_no_atomic32_add_read +#define erts_atomic32_add_wb erts_no_atomic32_add +#define erts_atomic32_read_bor_wb erts_no_atomic32_read_bor +#define erts_atomic32_read_band_wb erts_no_atomic32_read_band +#define erts_atomic32_xchg_wb erts_no_atomic32_xchg +#define erts_atomic32_cmpxchg_wb erts_no_atomic32_cmpxchg + +#endif /* !USE_THREADS */ + #if ERTS_GLB_INLINE_INCL_FUNC_DEF ERTS_GLB_INLINE void @@ -571,8 +990,9 @@ erts_mtx_destroy(erts_mtx_t *mtx) "Most likely a bug in pthread implementation."; erts_send_warning_to_logger_str_nogl(warn); } + else #endif - erts_thr_fatal_error(res, "destroy mutex"); + erts_thr_fatal_error(res, "destroy mutex"); } #endif } @@ -675,8 +1095,9 @@ erts_cnd_destroy(erts_cnd_t *cnd) "Most likely a bug in pthread implementation."; erts_send_warning_to_logger_str_nogl(warn); } + else #endif - erts_thr_fatal_error(res, "destroy condition variable"); + erts_thr_fatal_error(res, "destroy condition variable"); } #endif } @@ -810,8 +1231,9 @@ erts_rwmtx_destroy(erts_rwmtx_t *rwmtx) "Most likely a bug in pthread implementation."; erts_send_warning_to_logger_str_nogl(warn); } + else #endif - erts_thr_fatal_error(res, "destroy rwmutex"); + erts_thr_fatal_error(res, "destroy rwmutex"); } #endif } @@ -995,428 +1417,206 @@ erts_lc_rwmtx_is_rwlocked(erts_rwmtx_t *mtx) #endif } +/* No atomic ops */ + ERTS_GLB_INLINE void -erts_atomic_init(erts_atomic_t *var, erts_aint_t i) +erts_no_dw_atomic_set(erts_no_dw_atomic_t *var, erts_no_dw_atomic_t *val) { -#ifdef USE_THREADS - ethr_atomic_init(var, i); -#else - *var = i; -#endif + var->sint[0] = val->sint[0]; + var->sint[1] = val->sint[1]; } ERTS_GLB_INLINE void -erts_atomic_set(erts_atomic_t *var, erts_aint_t i) +erts_no_dw_atomic_read(erts_no_dw_atomic_t *var, erts_no_dw_atomic_t *val) +{ + val->sint[0] = var->sint[0]; + val->sint[1] = var->sint[1]; +} + +ERTS_GLB_INLINE int erts_no_dw_atomic_cmpxchg(erts_no_dw_atomic_t *var, + erts_no_dw_atomic_t *new_val, + erts_no_dw_atomic_t *old_val) +{ + if (var->sint[0] != old_val->sint[0] || var->sint[1] != old_val->sint[1]) { + erts_no_dw_atomic_read(var, old_val); + return 0; + } + else { + erts_no_dw_atomic_set(var, new_val); + return !0; + } +} + +ERTS_GLB_INLINE void +erts_no_atomic_set(erts_no_atomic_t *var, erts_aint_t i) { -#ifdef USE_THREADS - ethr_atomic_set(var, i); -#else *var = i; -#endif } ERTS_GLB_INLINE erts_aint_t -erts_atomic_read(erts_atomic_t *var) +erts_no_atomic_read(erts_no_atomic_t *var) { -#ifdef USE_THREADS - return ethr_atomic_read(var); -#else return *var; -#endif } ERTS_GLB_INLINE erts_aint_t -erts_atomic_inctest(erts_atomic_t *incp) +erts_no_atomic_inc_read(erts_no_atomic_t *incp) { -#ifdef USE_THREADS - return ethr_atomic_inc_read(incp); -#else return ++(*incp); -#endif } ERTS_GLB_INLINE erts_aint_t -erts_atomic_dectest(erts_atomic_t *decp) +erts_no_atomic_dec_read(erts_no_atomic_t *decp) { -#ifdef USE_THREADS - return ethr_atomic_dec_read(decp); -#else return --(*decp); -#endif } ERTS_GLB_INLINE void -erts_atomic_inc(erts_atomic_t *incp) +erts_no_atomic_inc(erts_no_atomic_t *incp) { -#ifdef USE_THREADS - ethr_atomic_inc(incp); -#else ++(*incp); -#endif } ERTS_GLB_INLINE void -erts_atomic_dec(erts_atomic_t *decp) +erts_no_atomic_dec(erts_no_atomic_t *decp) { -#ifdef USE_THREADS - ethr_atomic_dec(decp); -#else --(*decp); -#endif } ERTS_GLB_INLINE erts_aint_t -erts_atomic_addtest(erts_atomic_t *addp, erts_aint_t i) +erts_no_atomic_add_read(erts_no_atomic_t *addp, erts_aint_t i) { -#ifdef USE_THREADS - return ethr_atomic_add_read(addp, i); -#else return *addp += i; -#endif } ERTS_GLB_INLINE void -erts_atomic_add(erts_atomic_t *addp, erts_aint_t i) +erts_no_atomic_add(erts_no_atomic_t *addp, erts_aint_t i) { -#ifdef USE_THREADS - ethr_atomic_add(addp, i); -#else *addp += i; -#endif -} - -ERTS_GLB_INLINE erts_aint_t -erts_atomic_xchg(erts_atomic_t *xchgp, erts_aint_t new) -{ -#ifdef USE_THREADS - return ethr_atomic_xchg(xchgp, new); -#else - erts_aint_t old = *xchgp; - *xchgp = new; - return old; -#endif -} - -ERTS_GLB_INLINE erts_aint_t -erts_atomic_cmpxchg(erts_atomic_t *xchgp, erts_aint_t new, erts_aint_t expected) -{ -#ifdef USE_THREADS - return ethr_atomic_cmpxchg(xchgp, new, expected); -#else - erts_aint_t old = *xchgp; - if (old == expected) - *xchgp = new; - return old; -#endif } ERTS_GLB_INLINE erts_aint_t -erts_atomic_bor(erts_atomic_t *var, erts_aint_t mask) +erts_no_atomic_read_bor(erts_no_atomic_t *var, erts_aint_t mask) { -#ifdef USE_THREADS - return ethr_atomic_read_bor(var, mask); -#else erts_aint_t old; old = *var; *var |= mask; return old; -#endif } ERTS_GLB_INLINE erts_aint_t -erts_atomic_band(erts_atomic_t *var, erts_aint_t mask) +erts_no_atomic_read_band(erts_no_atomic_t *var, erts_aint_t mask) { -#ifdef USE_THREADS - return ethr_atomic_read_band(var, mask); -#else erts_aint_t old; old = *var; *var &= mask; return old; -#endif } ERTS_GLB_INLINE erts_aint_t -erts_atomic_read_acqb(erts_atomic_t *var) +erts_no_atomic_xchg(erts_no_atomic_t *xchgp, erts_aint_t new) { -#ifdef USE_THREADS - return ethr_atomic_read_acqb(var); -#else - return *var; -#endif -} - -ERTS_GLB_INLINE void -erts_atomic_set_relb(erts_atomic_t *var, erts_aint_t i) -{ -#ifdef USE_THREADS - ethr_atomic_set_relb(var, i); -#else - *var = i; -#endif -} - -ERTS_GLB_INLINE void -erts_atomic_dec_relb(erts_atomic_t *decp) -{ -#ifdef USE_THREADS - ethr_atomic_dec_relb(decp); -#else - --(*decp); -#endif -} - -ERTS_GLB_INLINE erts_aint_t -erts_atomic_dectest_relb(erts_atomic_t *decp) -{ -#ifdef USE_THREADS - return ethr_atomic_dec_read_relb(decp); -#else - return --(*decp); -#endif -} - -ERTS_GLB_INLINE erts_aint_t erts_atomic_cmpxchg_acqb(erts_atomic_t *xchgp, - erts_aint_t new, - erts_aint_t exp) -{ -#ifdef USE_THREADS - return ethr_atomic_cmpxchg_acqb(xchgp, new, exp); -#else erts_aint_t old = *xchgp; - if (old == exp) - *xchgp = new; + *xchgp = new; return old; -#endif } -ERTS_GLB_INLINE erts_aint_t erts_atomic_cmpxchg_relb(erts_atomic_t *xchgp, - erts_aint_t new, - erts_aint_t exp) +ERTS_GLB_INLINE erts_aint_t +erts_no_atomic_cmpxchg(erts_no_atomic_t *xchgp, + erts_aint_t new, + erts_aint_t expected) { -#ifdef USE_THREADS - return ethr_atomic_cmpxchg_relb(xchgp, new, exp); -#else erts_aint_t old = *xchgp; - if (old == exp) + if (old == expected) *xchgp = new; return old; -#endif } /* atomic32 */ ERTS_GLB_INLINE void -erts_atomic32_init(erts_atomic32_t *var, erts_aint32_t i) -{ -#ifdef USE_THREADS - ethr_atomic32_init(var, i); -#else - *var = i; -#endif -} - -ERTS_GLB_INLINE void -erts_atomic32_set(erts_atomic32_t *var, erts_aint32_t i) +erts_no_atomic32_set(erts_no_atomic32_t *var, erts_aint32_t i) { -#ifdef USE_THREADS - ethr_atomic32_set(var, i); -#else *var = i; -#endif } ERTS_GLB_INLINE erts_aint32_t -erts_atomic32_read(erts_atomic32_t *var) +erts_no_atomic32_read(erts_no_atomic32_t *var) { -#ifdef USE_THREADS - return ethr_atomic32_read(var); -#else return *var; -#endif } ERTS_GLB_INLINE erts_aint32_t -erts_atomic32_inctest(erts_atomic32_t *incp) +erts_no_atomic32_inc_read(erts_no_atomic32_t *incp) { -#ifdef USE_THREADS - return ethr_atomic32_inc_read(incp); -#else return ++(*incp); -#endif } ERTS_GLB_INLINE erts_aint32_t -erts_atomic32_dectest(erts_atomic32_t *decp) +erts_no_atomic32_dec_read(erts_no_atomic32_t *decp) { -#ifdef USE_THREADS - return ethr_atomic32_dec_read(decp); -#else return --(*decp); -#endif } ERTS_GLB_INLINE void -erts_atomic32_inc(erts_atomic32_t *incp) +erts_no_atomic32_inc(erts_no_atomic32_t *incp) { -#ifdef USE_THREADS - ethr_atomic32_inc(incp); -#else ++(*incp); -#endif } ERTS_GLB_INLINE void -erts_atomic32_dec(erts_atomic32_t *decp) +erts_no_atomic32_dec(erts_no_atomic32_t *decp) { -#ifdef USE_THREADS - ethr_atomic32_dec(decp); -#else --(*decp); -#endif } ERTS_GLB_INLINE erts_aint32_t -erts_atomic32_addtest(erts_atomic32_t *addp, erts_aint32_t i) +erts_no_atomic32_add_read(erts_no_atomic32_t *addp, erts_aint32_t i) { -#ifdef USE_THREADS - return ethr_atomic32_add_read(addp, i); -#else return *addp += i; -#endif } ERTS_GLB_INLINE void -erts_atomic32_add(erts_atomic32_t *addp, erts_aint32_t i) +erts_no_atomic32_add(erts_no_atomic32_t *addp, erts_aint32_t i) { -#ifdef USE_THREADS - ethr_atomic32_add(addp, i); -#else *addp += i; -#endif -} - -ERTS_GLB_INLINE erts_aint32_t -erts_atomic32_xchg(erts_atomic32_t *xchgp, erts_aint32_t new) -{ -#ifdef USE_THREADS - return ethr_atomic32_xchg(xchgp, new); -#else - erts_aint32_t old = *xchgp; - *xchgp = new; - return old; -#endif -} - -ERTS_GLB_INLINE erts_aint32_t -erts_atomic32_cmpxchg(erts_atomic32_t *xchgp, - erts_aint32_t new, - erts_aint32_t expected) -{ -#ifdef USE_THREADS - return ethr_atomic32_cmpxchg(xchgp, new, expected); -#else - erts_aint32_t old = *xchgp; - if (old == expected) - *xchgp = new; - return old; -#endif } ERTS_GLB_INLINE erts_aint32_t -erts_atomic32_bor(erts_atomic32_t *var, erts_aint32_t mask) +erts_no_atomic32_read_bor(erts_no_atomic32_t *var, erts_aint32_t mask) { -#ifdef USE_THREADS - return ethr_atomic32_read_bor(var, mask); -#else erts_aint32_t old; old = *var; *var |= mask; return old; -#endif } ERTS_GLB_INLINE erts_aint32_t -erts_atomic32_band(erts_atomic32_t *var, erts_aint32_t mask) +erts_no_atomic32_read_band(erts_no_atomic32_t *var, erts_aint32_t mask) { -#ifdef USE_THREADS - return ethr_atomic32_read_band(var, mask); -#else erts_aint32_t old; old = *var; *var &= mask; return old; -#endif } ERTS_GLB_INLINE erts_aint32_t -erts_atomic32_read_acqb(erts_atomic32_t *var) -{ -#ifdef USE_THREADS - return ethr_atomic32_read_acqb(var); -#else - return *var; -#endif -} - -ERTS_GLB_INLINE void -erts_atomic32_set_relb(erts_atomic32_t *var, erts_aint32_t i) -{ -#ifdef USE_THREADS - ethr_atomic32_set_relb(var, i); -#else - *var = i; -#endif -} - -ERTS_GLB_INLINE void -erts_atomic32_dec_relb(erts_atomic32_t *decp) +erts_no_atomic32_xchg(erts_no_atomic32_t *xchgp, erts_aint32_t new) { -#ifdef USE_THREADS - ethr_atomic32_dec_relb(decp); -#else - --(*decp); -#endif -} - -ERTS_GLB_INLINE erts_aint32_t -erts_atomic32_dectest_relb(erts_atomic32_t *decp) -{ -#ifdef USE_THREADS - return ethr_atomic32_dec_read_relb(decp); -#else - return --(*decp); -#endif -} - -ERTS_GLB_INLINE erts_aint32_t -erts_atomic32_cmpxchg_acqb(erts_atomic32_t *xchgp, - erts_aint32_t new, - erts_aint32_t exp) -{ -#ifdef USE_THREADS - return ethr_atomic32_cmpxchg_acqb(xchgp, new, exp); -#else erts_aint32_t old = *xchgp; - if (old == exp) - *xchgp = new; + *xchgp = new; return old; -#endif } ERTS_GLB_INLINE erts_aint32_t -erts_atomic32_cmpxchg_relb(erts_atomic32_t *xchgp, - erts_aint32_t new, - erts_aint32_t exp) +erts_no_atomic32_cmpxchg(erts_no_atomic32_t *xchgp, + erts_aint32_t new, + erts_aint32_t expected) { -#ifdef USE_THREADS - return ethr_atomic32_cmpxchg_relb(xchgp, new, exp); -#else erts_aint32_t old = *xchgp; - if (old == exp) + if (old == expected) *xchgp = new; return old; -#endif } /* spinlock */ @@ -1496,8 +1696,9 @@ erts_spinlock_destroy(erts_spinlock_t *lock) "Most likely a bug in pthread implementation."; erts_send_warning_to_logger_str_nogl(warn); } + else #endif - erts_thr_fatal_error(res, "destroy rwlock"); + erts_thr_fatal_error(res, "destroy rwlock"); } #else (void)lock; @@ -1614,8 +1815,9 @@ erts_rwlock_destroy(erts_rwlock_t *lock) "Most likely a bug in pthread implementation."; erts_send_warning_to_logger_str_nogl(warn); } + else #endif - erts_thr_fatal_error(res, "destroy rwlock"); + erts_thr_fatal_error(res, "destroy rwlock"); } #else (void)lock; @@ -1887,3 +2089,37 @@ erts_thr_sigwait(const sigset_t *set, int *sig) #endif /* #if ERTS_GLB_INLINE_INCL_FUNC_DEF */ #endif /* #ifndef ERL_THREAD_H__ */ + +#ifdef ERTS_UNDEF_DEPRECATED_ATOMICS + +/* Deprecated functions to replace */ + +#undef erts_atomic_init +#undef erts_atomic_set +#undef erts_atomic_read +#undef erts_atomic_inctest +#undef erts_atomic_dectest +#undef erts_atomic_inc +#undef erts_atomic_dec +#undef erts_atomic_addtest +#undef erts_atomic_add +#undef erts_atomic_xchg +#undef erts_atomic_cmpxchg +#undef erts_atomic_bor +#undef erts_atomic_band + +#undef erts_atomic32_init +#undef erts_atomic32_set +#undef erts_atomic32_read +#undef erts_atomic32_inctest +#undef erts_atomic32_dectest +#undef erts_atomic32_inc +#undef erts_atomic32_dec +#undef erts_atomic32_addtest +#undef erts_atomic32_add +#undef erts_atomic32_xchg +#undef erts_atomic32_cmpxchg +#undef erts_atomic32_bor +#undef erts_atomic32_band + +#endif diff --git a/erts/emulator/beam/erl_time.h b/erts/emulator/beam/erl_time.h index d0ad73cd81..7a09d30ff6 100644 --- a/erts/emulator/beam/erl_time.h +++ b/erts/emulator/beam/erl_time.h @@ -85,8 +85,8 @@ ERTS_GLB_INLINE void erts_do_time_add(long); #if ERTS_GLB_INLINE_INCL_FUNC_DEF -ERTS_GLB_INLINE erts_aint_t erts_do_time_read_and_reset(void) { return erts_smp_atomic_xchg(&do_time, 0L); } -ERTS_GLB_INLINE void erts_do_time_add(long elapsed) { erts_smp_atomic_add(&do_time, elapsed); } +ERTS_GLB_INLINE erts_aint_t erts_do_time_read_and_reset(void) { return erts_smp_atomic_xchg_acqb(&do_time, 0L); } +ERTS_GLB_INLINE void erts_do_time_add(long elapsed) { erts_smp_atomic_add_relb(&do_time, elapsed); } #endif /* #if ERTS_GLB_INLINE_INCL_FUNC_DEF */ diff --git a/erts/emulator/beam/external.c b/erts/emulator/beam/external.c index 1a102f7187..6953e7fe7d 100644 --- a/erts/emulator/beam/external.c +++ b/erts/emulator/beam/external.c @@ -459,6 +459,12 @@ Uint erts_encode_ext_size(Eterm term) + 1 /* VERSION_MAGIC */; } +Uint erts_encode_ext_size_2(Eterm term, unsigned dflags) +{ + return encode_size_struct2(NULL, term, TERM_TO_BINARY_DFLAGS|dflags) + + 1 /* VERSION_MAGIC */; +} + Uint erts_encode_ext_size_ets(Eterm term) { return encode_size_struct2(NULL, term, TERM_TO_BINARY_DFLAGS|DFLAGS_INTERNAL_TAGS); @@ -1262,6 +1268,49 @@ external_size_1(Process* p, Eterm Term) } Eterm +external_size_2(Process* p, Eterm Term, Eterm Flags) +{ + Uint size; + Uint flags = TERM_TO_BINARY_DFLAGS; + + while (is_list(Flags)) { + Eterm arg = CAR(list_val(Flags)); + Eterm* tp; + + if (is_tuple(arg) && *(tp = tuple_val(arg)) == make_arityval(2)) { + if (tp[1] == am_minor_version && is_small(tp[2])) { + switch (signed_val(tp[2])) { + case 0: + break; + case 1: + flags |= DFLAG_NEW_FLOATS; + break; + default: + goto error; + } + } else { + goto error; + } + } else { + error: + BIF_ERROR(p, BADARG); + } + Flags = CDR(list_val(Flags)); + } + if (is_not_nil(Flags)) { + goto error; + } + + size = erts_encode_ext_size_2(Term, flags); + if (IS_USMALL(0, size)) { + BIF_RET(make_small(size)); + } else { + Eterm* hp = HAlloc(p, BIG_UINT_HEAP_SIZE); + BIF_RET(uint_to_big(size, hp)); + } +} + +Eterm erts_term_to_binary(Process* p, Eterm Term, int level, Uint flags) { Uint size; diff --git a/erts/emulator/beam/external.h b/erts/emulator/beam/external.h index d8287b96a4..671b8b8781 100644 --- a/erts/emulator/beam/external.h +++ b/erts/emulator/beam/external.h @@ -1,7 +1,7 @@ /* * %CopyrightBegin% * - * Copyright Ericsson AB 1996-2010. All Rights Reserved. + * Copyright Ericsson AB 1996-2011. All Rights Reserved. * * The contents of this file are subject to the Erlang Public License, * Version 1.1, (the "License"); you may not use this file except in @@ -160,6 +160,7 @@ Uint erts_encode_dist_ext_size(Eterm, Uint32, ErtsAtomCacheMap *); void erts_encode_dist_ext(Eterm, byte **, Uint32, ErtsAtomCacheMap *); Uint erts_encode_ext_size(Eterm); +Uint erts_encode_ext_size_2(Eterm, unsigned); Uint erts_encode_ext_size_ets(Eterm); void erts_encode_ext(Eterm, byte **); byte* erts_encode_ext_ets(Eterm, byte *, struct erl_off_heap_header** ext_off_heap); diff --git a/erts/emulator/beam/global.h b/erts/emulator/beam/global.h index 499bdd77ba..a967aa0e3e 100644 --- a/erts/emulator/beam/global.h +++ b/erts/emulator/beam/global.h @@ -37,6 +37,7 @@ #include "erl_process.h" #include "erl_sys_driver.h" #include "erl_debug.h" +#include "error.h" typedef struct port Port; #include "erl_port_task.h" @@ -200,10 +201,10 @@ erts_port_runq(Port *prt) { #ifdef ERTS_SMP ErtsRunQueue *rq1, *rq2; - rq1 = (ErtsRunQueue *) erts_smp_atomic_read(&prt->run_queue); + rq1 = (ErtsRunQueue *) erts_smp_atomic_read_nob(&prt->run_queue); while (1) { erts_smp_runq_lock(rq1); - rq2 = (ErtsRunQueue *) erts_smp_atomic_read(&prt->run_queue); + rq2 = (ErtsRunQueue *) erts_smp_atomic_read_nob(&prt->run_queue); if (rq1 == rq2) return rq1; erts_smp_runq_unlock(rq1); @@ -542,10 +543,11 @@ ERTS_GLB_INLINE void erts_may_save_closed_port(Port *prt) ERTS_SMP_LC_ASSERT(erts_smp_lc_spinlock_is_locked(&prt->state_lck)); if (prt->snapshot != erts_smp_atomic32_read_acqb(&erts_ports_snapshot)) { /* Dead ports are added from the end of the snapshot buffer */ - Eterm* tombstone = (Eterm*) erts_smp_atomic_addtest(&erts_dead_ports_ptr, - -(erts_aint_t)sizeof(Eterm)); + Eterm* tombstone; + tombstone = (Eterm*) erts_smp_atomic_add_read_nob(&erts_dead_ports_ptr, + -(erts_aint_t)sizeof(Eterm)); ASSERT(tombstone+1 != NULL); - ASSERT(prt->snapshot == erts_smp_atomic32_read(&erts_ports_snapshot) - 1); + ASSERT(prt->snapshot == erts_smp_atomic_read_nob(&erts_ports_snapshot) - 1); *tombstone = prt->id; } /*else no ongoing snapshot or port was already included or created after snapshot */ @@ -858,10 +860,21 @@ void erts_system_monitor_clear(Process *c_p); void erts_system_profile_clear(Process *c_p); /* beam_load.c */ +typedef struct { + BeamInstr* current; /* Pointer to: Mod, Name, Arity */ + Uint needed; /* Heap space needed for entire tuple */ + Uint32 loc; /* Location in source code */ + Eterm* fname_ptr; /* Pointer to fname table */ +} FunctionInfo; + int erts_load_module(Process *c_p, ErtsProcLocks c_p_locks, Eterm group_leader, Eterm* mod, byte* code, int size); void init_load(void); BeamInstr* find_function_from_pc(BeamInstr* pc); +Eterm* erts_build_mfa_item(FunctionInfo* fi, Eterm* hp, + Eterm args, Eterm* mfa_p); +void erts_lookup_function_info(FunctionInfo* fi, BeamInstr* pc, int full_info); +void erts_set_current_function(FunctionInfo* fi, BeamInstr* current); Eterm erts_module_info_0(Process* p, Eterm module); Eterm erts_module_info_1(Process* p, Eterm module, Eterm what); Eterm erts_make_stub_module(Process* p, Eterm Mod, Eterm Beam, Eterm Info); @@ -1052,6 +1065,7 @@ void init_emulator(void); void process_main(void); Eterm build_stacktrace(Process* c_p, Eterm exc); Eterm expand_error_value(Process* c_p, Uint freason, Eterm Value); +void erts_save_stacktrace(Process* p, struct StackTrace* s, int depth); /* erl_init.c */ @@ -1073,6 +1087,7 @@ extern ErtsModifiedTimings erts_modified_timings[]; #define ERTS_MODIFIED_TIMING_INPUT_REDS \ (erts_modified_timings[erts_modified_timing_level].input_reds) +extern int erts_no_line_info; extern Eterm erts_error_logger_warnings; extern int erts_initialized; extern int erts_compat_rel; @@ -1200,11 +1215,11 @@ erts_smp_port_trylock(Port *prt) #ifdef ERTS_SMP int res; - ASSERT(erts_smp_atomic_read(&prt->refc) > 0); - erts_smp_atomic_inc(&prt->refc); + ASSERT(erts_smp_atomic_read_nob(&prt->refc) > 0); + erts_smp_atomic_inc_nob(&prt->refc); res = erts_smp_mtx_trylock(prt->lock); if (res == EBUSY) { - erts_smp_atomic_dec(&prt->refc); + erts_smp_atomic_dec_nob(&prt->refc); } return res; @@ -1217,8 +1232,8 @@ ERTS_GLB_INLINE void erts_smp_port_lock(Port *prt) { #ifdef ERTS_SMP - ASSERT(erts_smp_atomic_read(&prt->refc) > 0); - erts_smp_atomic_inc(&prt->refc); + ASSERT(erts_smp_atomic_read_nob(&prt->refc) > 0); + erts_smp_atomic_inc_nob(&prt->refc); erts_smp_mtx_lock(prt->lock); #endif } @@ -1229,7 +1244,7 @@ erts_smp_port_unlock(Port *prt) #ifdef ERTS_SMP erts_aint_t refc; erts_smp_mtx_unlock(prt->lock); - refc = erts_smp_atomic_dectest(&prt->refc); + refc = erts_smp_atomic_dec_read_nob(&prt->refc); ASSERT(refc >= 0); if (refc == 0) erts_port_cleanup(prt); @@ -1298,7 +1313,7 @@ erts_id2port_sflgs(Eterm id, Process *c_p, ErtsProcLocks c_p_locks, Uint32 sflgs } #ifdef ERTS_SMP else { - erts_smp_atomic_inc(&prt->refc); + erts_smp_atomic_inc_nob(&prt->refc); erts_smp_port_state_unlock(prt); if (no_proc_locks) diff --git a/erts/emulator/beam/io.c b/erts/emulator/beam/io.c index df5f8b22a3..151c776a3d 100644 --- a/erts/emulator/beam/io.c +++ b/erts/emulator/beam/io.c @@ -244,8 +244,8 @@ get_free_port(void) } port->status = ERTS_PORT_SFLG_INITIALIZING; #ifdef ERTS_SMP - ERTS_SMP_LC_ASSERT(erts_smp_atomic_read(&port->refc) == 0); - erts_smp_atomic_set(&port->refc, 2); /* Port alive + lock */ + ERTS_SMP_LC_ASSERT(erts_smp_atomic_read_nob(&port->refc) == 0); + erts_smp_atomic_set_nob(&port->refc, 2); /* Port alive + lock */ #endif erts_smp_port_state_unlock(port); return num & port_num_mask; @@ -327,7 +327,7 @@ port_cleanup(Port *prt) #ifdef ERTS_SMP ASSERT(prt->status & ERTS_PORT_SFLG_FREE_SCHEDULED); - ERTS_SMP_LC_ASSERT(erts_smp_atomic_read(&prt->refc) == 0); + ERTS_SMP_LC_ASSERT(erts_smp_atomic_read_nob(&prt->refc) == 0); port_specific = (prt->status & ERTS_PORT_SFLG_PORT_SPECIFIC_LOCK); @@ -425,11 +425,11 @@ setup_port(Port* prt, Eterm pid, erts_driver_t *driver, erts_smp_runq_lock(runq); erts_smp_port_state_lock(prt); prt->status = ERTS_PORT_SFLG_CONNECTED | xstatus; - prt->snapshot = erts_smp_atomic32_read(&erts_ports_snapshot); + prt->snapshot = erts_smp_atomic32_read_nob(&erts_ports_snapshot); old_name = prt->name; prt->name = new_name; #ifdef ERTS_SMP - erts_smp_atomic_set(&prt->run_queue, (erts_aint_t) runq); + erts_smp_atomic_set_nob(&prt->run_queue, (erts_aint_t) runq); #endif ASSERT(!prt->drv_ptr); prt->drv_ptr = driver; @@ -590,8 +590,8 @@ erts_open_driver(erts_driver_t* driver, /* Pointer to driver. */ erts_smp_port_state_lock(port); port->status = ERTS_PORT_SFLG_FREE; #ifdef ERTS_SMP - ERTS_SMP_LC_ASSERT(erts_smp_atomic_read(&port->refc) == 2); - erts_smp_atomic_set(&port->refc, 0); + ERTS_SMP_LC_ASSERT(erts_smp_atomic_read_nob(&port->refc) == 2); + erts_smp_atomic_set_nob(&port->refc, 0); #endif erts_smp_port_state_unlock(port); return -3; @@ -1206,7 +1206,7 @@ int erts_write_to_port(Eterm caller_id, Port *p, Eterm list) } } p->bytes_out += size; - erts_smp_atomic_add(&erts_bytes_out, size); + erts_smp_atomic_add_nob(&erts_bytes_out, size); #ifdef ERTS_SMP if (p->xports) @@ -1277,13 +1277,13 @@ void init_io(void) erts_port = (Port *) erts_alloc(ERTS_ALC_T_PORT_TABLE, erts_max_ports * sizeof(Port)); - erts_smp_atomic_init(&erts_bytes_out, 0); - erts_smp_atomic_init(&erts_bytes_in, 0); + erts_smp_atomic_init_nob(&erts_bytes_out, 0); + erts_smp_atomic_init_nob(&erts_bytes_in, 0); for (i = 0; i < erts_max_ports; i++) { erts_port_task_init_sched(&erts_port[i].sched); #ifdef ERTS_SMP - erts_smp_atomic_init(&erts_port[i].refc, 0); + erts_smp_atomic_init_nob(&erts_port[i].refc, 0); erts_port[i].lock = NULL; erts_port[i].xports = NULL; erts_smp_spinlock_init_x(&erts_port[i].state_lck, "port_state", make_small(i)); @@ -1300,7 +1300,7 @@ void init_io(void) erts_port[i].port_data_lock = NULL; } - erts_smp_atomic32_init(&erts_ports_snapshot, (erts_aint32_t) 0); + erts_smp_atomic32_init_nob(&erts_ports_snapshot, (erts_aint32_t) 0); last_port_num = 0; erts_smp_spinlock_init(&get_free_port_lck, "get_free_port"); @@ -3253,7 +3253,7 @@ int driver_output_binary(ErlDrvPort ix, char* hbuf, int hlen, return 0; prt->bytes_in += (hlen + len); - erts_smp_atomic_add(&erts_bytes_in, (erts_aint_t) (hlen + len)); + erts_smp_atomic_add_nob(&erts_bytes_in, (erts_aint_t) (hlen + len)); if (prt->status & ERTS_PORT_SFLG_DISTRIBUTION) { return erts_net_message(prt, prt->dist_entry, @@ -3288,7 +3288,7 @@ int driver_output2(ErlDrvPort ix, char* hbuf, int hlen, char* buf, int len) return 0; prt->bytes_in += (hlen + len); - erts_smp_atomic_add(&erts_bytes_in, (erts_aint_t) (hlen + len)); + erts_smp_atomic_add_nob(&erts_bytes_in, (erts_aint_t) (hlen + len)); if (prt->status & ERTS_PORT_SFLG_DISTRIBUTION) { if (len == 0) return erts_net_message(prt, @@ -3365,7 +3365,7 @@ int driver_outputv(ErlDrvPort ix, char* hbuf, int hlen, ErlIOVec* vec, int skip) /* XXX handle distribution !!! */ prt->bytes_in += (hlen + size); - erts_smp_atomic_add(&erts_bytes_in, (erts_aint_t) (hlen + size)); + erts_smp_atomic_add_nob(&erts_bytes_in, (erts_aint_t) (hlen + size)); deliver_vec_message(prt, prt->connected, hbuf, hlen, binv, iov, n, size); return 0; } @@ -3539,13 +3539,13 @@ pdl_init(void) static ERTS_INLINE void pdl_init_refc(ErlDrvPDL pdl) { - erts_atomic_init(&pdl->refc, 1); + erts_atomic_init_nob(&pdl->refc, 1); } static ERTS_INLINE ErlDrvSInt pdl_read_refc(ErlDrvPDL pdl) { - erts_aint_t refc = erts_atomic_read(&pdl->refc); + erts_aint_t refc = erts_atomic_read_nob(&pdl->refc); ERTS_LC_ASSERT(refc >= 0); return (ErlDrvSInt) refc; } @@ -3553,14 +3553,14 @@ pdl_read_refc(ErlDrvPDL pdl) static ERTS_INLINE void pdl_inc_refc(ErlDrvPDL pdl) { - erts_atomic_inc(&pdl->refc); + erts_atomic_inc_nob(&pdl->refc); ERTS_LC_ASSERT(driver_pdl_get_refc(pdl) > 1); } static ERTS_INLINE ErlDrvSInt pdl_inctest_refc(ErlDrvPDL pdl) { - erts_aint_t refc = erts_atomic_inctest(&pdl->refc); + erts_aint_t refc = erts_atomic_inc_read_nob(&pdl->refc); ERTS_LC_ASSERT(refc > 1); return (ErlDrvSInt) refc; } @@ -3569,7 +3569,7 @@ pdl_inctest_refc(ErlDrvPDL pdl) static ERTS_INLINE void pdl_dec_refc(ErlDrvPDL pdl) { - erts_atomic_dec(&pdl->refc); + erts_atomic_dec_nob(&pdl->refc); ERTS_LC_ASSERT(driver_pdl_get_refc(pdl) > 0); } #endif @@ -3577,7 +3577,7 @@ pdl_dec_refc(ErlDrvPDL pdl) static ERTS_INLINE ErlDrvSInt pdl_dectest_refc(ErlDrvPDL pdl) { - erts_aint_t refc = erts_atomic_dectest(&pdl->refc); + erts_aint_t refc = erts_atomic_dec_read_nob(&pdl->refc); ERTS_LC_ASSERT(refc >= 0); return (ErlDrvSInt) refc; } diff --git a/erts/emulator/beam/ops.tab b/erts/emulator/beam/ops.tab index 8a5763b4bb..34bd5d0653 100644 --- a/erts/emulator/beam/ops.tab +++ b/erts/emulator/beam/ops.tab @@ -25,30 +25,12 @@ # instruction transformations; thus, they never occur in BEAM files. # -# Special instruction used to generate an error message when -# trying to load a module compiled by the V1 compiler (R5 & R6). -# (Specially treated in beam_load.c.) +# The too_old_compiler/0 instruction is specially handled in beam_load.c +# to produce a user-friendly message informing the user that the module +# needs to be re-compiled with a modern compiler. too_old_compiler/0 -too_old_compiler - -# -# Obsolete instruction usage follow. (Nowdays we use f with -# a zero label instead of p.) -# - -is_list p S => too_old_compiler -is_nonempty_list p R => too_old_compiler -is_nil p R => too_old_compiler - -is_tuple p S => too_old_compiler -test_arity p S Arity => too_old_compiler - -is_integer p R => too_old_compiler -is_float p R => too_old_compiler -is_atom p R => too_old_compiler - -is_eq_exact p S1 S2 => too_old_compiler +too_old_compiler | never() => # In R9C and earlier, the loader used to insert special instructions inside # the module_info/0,1 functions. (In R10B and later, the compiler inserts @@ -88,12 +70,42 @@ i_time_breakpoint i_return_time_trace i_return_to_trace i_yield -i_global_cons -i_global_tuple -i_global_copy return +# +# To ensure that a "move Src x(0)" instruction can be combined +# with the following call instruction, we need to make sure that +# there is no line/1 instruction between the move and the call. +# + +move S r | line Loc | call_ext Ar Func => \ + line Loc | move S r | call_ext Ar Func +move S r | line Loc | call_ext_last Ar Func=u$is_bif D => \ + line Loc | move S r | call_ext_last Ar Func D +move S r | line Loc | call_ext_only Ar Func=u$is_bif => \ + line Loc | move S r | call_ext_only Ar Func +move S r | line Loc | call Ar Func => \ + line Loc | move S r | call Ar Func + +# +# A tail-recursive call to an external function (non-BIF) will +# never be saved on the stack, so there is no reason to keep +# the line instruction. (The compiler did not remove the line +# instruction because it cannot tell the difference between +# BIFs and ordinary Erlang functions.) +# + +line Loc | call_ext_last Ar Func=u$is_not_bif D => \ + call_ext_last Ar Func D +line Loc | call_ext_only Ar Func=u$is_not_bif => \ + call_ext_only Ar Func + +line Loc | func_info M F A => func_info M F A | line Loc + +line I + + %macro: allocate Allocate -pack %macro: allocate_zero AllocateZero -pack %macro: allocate_heap AllocateHeap -pack @@ -277,8 +289,6 @@ raise s s badarg j system_limit j -move R R => - move C=cxy r | jump Lbl => move_jump Lbl C %macro: move_jump MoveJump -nonext @@ -585,8 +595,6 @@ get_tuple_element Reg P Dst => i_get_tuple_element Reg P Dst | original_reg Reg original_reg Reg Pos => -get_tuple_element Reg P Dst => i_get_tuple_element Reg P Dst - original_reg/2 extract_next_element D1=xy | original_reg Reg P1 | get_tuple_element Reg P2 D2=xy | \ @@ -875,23 +883,6 @@ call_ext_last u==3 u$func:erlang:hibernate/3 D => i_hibernate call_ext_only u==3 u$func:erlang:hibernate/3 => i_hibernate # -# Hybrid memory architecture need special cons and tuple instructions -# that allocate on the message area. These looks like BIFs in the BEAM code. -# - -call_ext u==2 u$func:hybrid:cons/2 => i_global_cons -call_ext_last u==2 u$func:hybrid:cons/2 D => i_global_cons | deallocate_return D -call_ext_only Ar=u==2 u$func:hybrid:cons/2 => i_global_cons | return - -call_ext u==1 u$func:hybrid:tuple/1 => i_global_tuple -call_ext_last u==1 u$func:hybrid:tuple/1 D => i_global_tuple | deallocate_return D -call_ext_only Ar=u==1 u$func:hybrid:tuple/1 => i_global_tuple | return - -call_ext u==1 u$func:hybrid:copy/1 => i_global_copy -call_ext_last u==1 u$func:hybrid:copy/1 D => i_global_copy | deallocate_return D -call_ext_only u==1 Ar=u$func:hybrid:copy/1 => i_global_copy | return - -# # The general case for BIFs that have no special instructions. # A BIF used in the tail must be followed by a return instruction. # @@ -928,9 +919,9 @@ move S=c r | call_ext Ar=u Func=u$is_not_bif => i_move_call_ext S r Func move S=c r | call_ext_last Ar=u Func=u$is_not_bif D => i_move_call_ext_last Func D S r move S=c r | call_ext_only Ar=u Func=u$is_not_bif => i_move_call_ext_only Func S r -call_ext Ar=u Func => i_call_ext Func -call_ext_last Ar=u Func D => i_call_ext_last Func D -call_ext_only Ar=u Func => i_call_ext_only Func +call_ext Ar Func => i_call_ext Func +call_ext_last Ar Func D => i_call_ext_last Func D +call_ext_only Ar Func => i_call_ext_only Func i_apply i_apply_last P @@ -964,7 +955,7 @@ bif1 p Bif S1 Dst => bif1_body Bif S1 Dst bif1_body Bif Literal=q Dst => move Literal x | bif1_body Bif x Dst bif2 p Bif S1 S2 Dst => i_fetch S1 S2 | i_bif2_body Bif Dst -bif2 Fail=f Bif S1 S2 Dst => i_fetch S1 S2 | i_bif2 Fail Bif Dst +bif2 Fail Bif S1 S2 Dst => i_fetch S1 S2 | i_bif2 Fail Bif Dst i_get s d @@ -1047,8 +1038,8 @@ i_move_call_ext_only e c r # Fun calls. -call_fun Arity=u | deallocate D | return => i_call_fun_last Arity D -call_fun Arity=u => i_call_fun Arity +call_fun Arity | deallocate D | return => i_call_fun_last Arity D +call_fun Arity => i_call_fun Arity i_call_fun I i_call_fun_last I P @@ -1236,7 +1227,7 @@ i_bs_init_heap I I I d i_bs_init_heap_bin_heap I I I d -bs_init_bits Fail Sz Words Regs Flags Dst | binary_too_big_bits(Sz) => system_limit Fail +bs_init_bits Fail Sz=o Words Regs Flags Dst => system_limit Fail bs_init_bits Fail Sz=u Words=u==0 Regs Flags Dst => i_bs_init_bits Sz Regs Dst bs_init_bits Fail Sz=u Words Regs Flags Dst => i_bs_init_bits_heap Sz Words Regs Dst @@ -1304,13 +1295,13 @@ i_bs_utf16_size s d bs_put_utf8 Fail=j Flags=u Literal=q => \ move Literal x | bs_put_utf8 Fail Flags x -bs_put_utf8 Fail=j u Src=s => i_bs_put_utf8 Fail Src +bs_put_utf8 Fail u Src=s => i_bs_put_utf8 Fail Src i_bs_put_utf8 j s bs_put_utf16 Fail=j Flags=u Literal=q => \ move Literal x | bs_put_utf16 Fail Flags x -bs_put_utf16 Fail=j Flags=u Src=s => i_bs_put_utf16 Fail Flags Src +bs_put_utf16 Fail Flags=u Src=s => i_bs_put_utf16 Fail Flags Src i_bs_put_utf16 j I s @@ -1475,34 +1466,13 @@ bif1 Fail u$bif:erlang:trunc/1 s d => too_old_compiler # # Guard BIFs. # -gc_bif1 Fail I Bif=u$bif:erlang:length/1 Src Dst=d => \ - gen_guard_bif1(Fail, I, Bif, Src, Dst) - -gc_bif1 Fail I Bif=u$bif:erlang:size/1 Src Dst=d => \ - gen_guard_bif1(Fail, I, Bif, Src, Dst) - -gc_bif1 Fail I Bif=u$bif:erlang:bit_size/1 Src Dst=d => \ - gen_guard_bif1(Fail, I, Bif, Src, Dst) - -gc_bif1 Fail I Bif=u$bif:erlang:byte_size/1 Src Dst=d => \ - gen_guard_bif1(Fail, I, Bif, Src, Dst) - -gc_bif1 Fail I Bif=u$bif:erlang:abs/1 Src Dst=d => \ - gen_guard_bif1(Fail, I, Bif, Src, Dst) - -gc_bif1 Fail I Bif=u$bif:erlang:float/1 Src Dst=d => \ +gc_bif1 Fail I Bif Src Dst => \ gen_guard_bif1(Fail, I, Bif, Src, Dst) -gc_bif1 Fail I Bif=u$bif:erlang:round/1 Src Dst=d => \ - gen_guard_bif1(Fail, I, Bif, Src, Dst) - -gc_bif1 Fail I Bif=u$bif:erlang:trunc/1 Src Dst=d => \ - gen_guard_bif1(Fail, I, Bif, Src, Dst) - -gc_bif2 Fail I Bif=u$bif:erlang:binary_part/2 S1 S2 Dst=d => \ +gc_bif2 Fail I Bif S1 S2 Dst => \ gen_guard_bif2(Fail, I, Bif, S1, S2, Dst) -gc_bif3 Fail I Bif=u$bif:erlang:binary_part/3 S1 S2 S3 Dst=d => \ +gc_bif3 Fail I Bif S1 S2 S3 Dst => \ gen_guard_bif3(Fail, I, Bif, S1, S2, S3, Dst) i_gc_bif1 Fail Bif V=q Live D => move V x | i_gc_bif1 Fail Bif x Live D @@ -1520,6 +1490,15 @@ ii_gc_bif3/7 ii_gc_bif3 Fail Bif S1 S2 S3 Live D => move S1 x | i_fetch S2 S3 | i_gc_bif3 Fail Bif x Live D i_gc_bif3 j I s I d + +# +# The following instruction is specially handled in beam_load.c +# to produce a user-friendly message if an unsupported guard BIF is +# encountered. +# +unsupported_guard_bif/3 +unsupported_guard_bif A B C | never() => + # # R13B03 # diff --git a/erts/emulator/beam/safe_hash.c b/erts/emulator/beam/safe_hash.c index 4c54e19cdb..3326e5cc2a 100644 --- a/erts/emulator/beam/safe_hash.c +++ b/erts/emulator/beam/safe_hash.c @@ -61,7 +61,7 @@ static ERTS_INLINE int align_up_pow2(int val) */ static void rehash(SafeHash* h, int grow_limit) { - if (erts_smp_atomic_xchg(&h->is_rehashing, 1) != 0) { + if (erts_smp_atomic_xchg_acqb(&h->is_rehashing, 1) != 0) { return; /* already in progress */ } if (h->grow_limit == grow_limit) { @@ -166,8 +166,8 @@ SafeHash* safe_hash_init(ErtsAlcType_t type, SafeHash* h, char* name, int size, h->name = name; h->fun = fun; set_size(h,size); - erts_smp_atomic_init(&h->is_rehashing, 0); - erts_smp_atomic_init(&h->nitems, 0); + erts_smp_atomic_init_nob(&h->is_rehashing, 0); + erts_smp_atomic_init_nob(&h->nitems, 0); for (i=0; i<SAFE_HASH_LOCK_CNT; i++) { erts_smp_mtx_init(&h->lock_vec[i].mtx,"safe_hash"); } @@ -222,7 +222,7 @@ void* safe_hash_put(SafeHash* h, void* tmpl) *head = b; grow_limit = h->grow_limit; erts_smp_mtx_unlock(lock); - if (erts_smp_atomic_inctest(&h->nitems) > grow_limit) { + if (erts_smp_atomic_inc_read_nob(&h->nitems) > grow_limit) { rehash(h, grow_limit); } return (void*) b; @@ -245,7 +245,7 @@ void* safe_hash_erase(SafeHash* h, void* tmpl) if ((b->hvalue == hval) && (h->fun.cmp(tmpl, (void*)b) == 0)) { *prevp = b->next; erts_smp_mtx_unlock(lock); - erts_smp_atomic_dec(&h->nitems); + erts_smp_atomic_dec_nob(&h->nitems); h->fun.free((void*)b); return tmpl; } diff --git a/erts/emulator/beam/sys.h b/erts/emulator/beam/sys.h index e64c43de6e..669a601b35 100644 --- a/erts/emulator/beam/sys.h +++ b/erts/emulator/beam/sys.h @@ -340,7 +340,8 @@ int erts_send_warning_to_logger_str_nogl(char *); #ifdef ERTS_WANT_BREAK_HANDLING # ifdef ERTS_SMP extern erts_smp_atomic32_t erts_break_requested; -# define ERTS_BREAK_REQUESTED ((int) erts_smp_atomic32_read(&erts_break_requested)) +# define ERTS_BREAK_REQUESTED \ + ((int) erts_smp_atomic32_read_nob(&erts_break_requested)) # else extern volatile int erts_break_requested; # define ERTS_BREAK_REQUESTED erts_break_requested @@ -354,7 +355,7 @@ void erts_do_break_handling(void); # else # ifdef ERTS_SMP extern erts_smp_atomic32_t erts_got_sigusr1; -# define ERTS_GOT_SIGUSR1 ((int) erts_smp_atomic32_read(&erts_got_sigusr1)) +# define ERTS_GOT_SIGUSR1 ((int) erts_smp_atomic32_read_mb(&erts_got_sigusr1)) # else extern volatile int erts_got_sigusr1; # define ERTS_GOT_SIGUSR1 erts_got_sigusr1 @@ -363,11 +364,15 @@ extern volatile int erts_got_sigusr1; #endif #ifdef ERTS_SMP -extern erts_smp_atomic_t erts_writing_erl_crash_dump; +extern erts_smp_atomic32_t erts_writing_erl_crash_dump; +extern erts_tsd_key_t erts_is_crash_dumping_key; +#define ERTS_SOMEONE_IS_CRASH_DUMPING \ + ((int) erts_smp_atomic32_read_mb(&erts_writing_erl_crash_dump)) #define ERTS_IS_CRASH_DUMPING \ - ((int) erts_smp_atomic_read(&erts_writing_erl_crash_dump)) + ((int) (SWord) erts_tsd_get(erts_is_crash_dumping_key)) #else extern volatile int erts_writing_erl_crash_dump; +#define ERTS_SOMEONE_IS_CRASH_DUMPING erts_writing_erl_crash_dump #define ERTS_IS_CRASH_DUMPING erts_writing_erl_crash_dump #endif @@ -877,7 +882,7 @@ ERTS_GLB_INLINE int erts_smp_pending_system_block(void) { #ifdef ERTS_SMP - return (int) erts_smp_atomic32_read(&erts_system_block_state.do_block); + return (int) erts_smp_atomic32_read_nob(&erts_system_block_state.do_block); #else return 0; #endif @@ -913,7 +918,7 @@ erts_smp_set_activity(erts_activity_t old_activity, case ERTS_ACTIVITY_UNDEFINED: break; case ERTS_ACTIVITY_WAIT: - erts_smp_atomic32_dec(&erts_system_block_state.in_activity.wait); + erts_smp_atomic32_dec_acqb(&erts_system_block_state.in_activity.wait); if (locked) { /* You are not allowed to leave activity waiting * without supplying the possibility to block @@ -924,10 +929,10 @@ erts_smp_set_activity(erts_activity_t old_activity, } break; case ERTS_ACTIVITY_GC: - erts_smp_atomic32_dec(&erts_system_block_state.in_activity.gc); + erts_smp_atomic32_dec_acqb(&erts_system_block_state.in_activity.gc); break; case ERTS_ACTIVITY_IO: - erts_smp_atomic32_dec(&erts_system_block_state.in_activity.io); + erts_smp_atomic32_dec_acqb(&erts_system_block_state.in_activity.io); break; default: erts_set_activity_error(ERTS_ACT_ERR_LEAVE_UNKNOWN_ACTIVITY, @@ -943,13 +948,13 @@ erts_smp_set_activity(erts_activity_t old_activity, case ERTS_ACTIVITY_UNDEFINED: break; case ERTS_ACTIVITY_WAIT: - erts_smp_atomic32_inc(&erts_system_block_state.in_activity.wait); + erts_smp_atomic32_inc_mb(&erts_system_block_state.in_activity.wait); break; case ERTS_ACTIVITY_GC: - erts_smp_atomic32_inc(&erts_system_block_state.in_activity.gc); + erts_smp_atomic32_inc_mb(&erts_system_block_state.in_activity.gc); break; case ERTS_ACTIVITY_IO: - erts_smp_atomic32_inc(&erts_system_block_state.in_activity.io); + erts_smp_atomic32_inc_mb(&erts_system_block_state.in_activity.io); break; default: erts_set_activity_error(ERTS_ACT_ERR_ENTER_UNKNOWN_ACTIVITY, @@ -1001,27 +1006,27 @@ ERTS_GLB_INLINE erts_aint_t erts_refc_read(erts_refc_t *refcp, ERTS_GLB_INLINE void erts_refc_init(erts_refc_t *refcp, erts_aint_t val) { - erts_smp_atomic_init((erts_smp_atomic_t *) refcp, val); + erts_smp_atomic_init_nob((erts_smp_atomic_t *) refcp, val); } ERTS_GLB_INLINE void erts_refc_inc(erts_refc_t *refcp, erts_aint_t min_val) { #ifdef ERTS_REFC_DEBUG - erts_aint_t val = erts_smp_atomic_inctest((erts_smp_atomic_t *) refcp); + erts_aint_t val = erts_smp_atomic_inc_read_nob((erts_smp_atomic_t *) refcp); if (val < min_val) erl_exit(ERTS_ABORT_EXIT, "erts_refc_inc(): Bad refc found (refc=%ld < %ld)!\n", val, min_val); #else - erts_smp_atomic_inc((erts_smp_atomic_t *) refcp); + erts_smp_atomic_inc_nob((erts_smp_atomic_t *) refcp); #endif } ERTS_GLB_INLINE erts_aint_t erts_refc_inctest(erts_refc_t *refcp, erts_aint_t min_val) { - erts_aint_t val = erts_smp_atomic_inctest((erts_smp_atomic_t *) refcp); + erts_aint_t val = erts_smp_atomic_inc_read_nob((erts_smp_atomic_t *) refcp); #ifdef ERTS_REFC_DEBUG if (val < min_val) erl_exit(ERTS_ABORT_EXIT, @@ -1035,20 +1040,20 @@ ERTS_GLB_INLINE void erts_refc_dec(erts_refc_t *refcp, erts_aint_t min_val) { #ifdef ERTS_REFC_DEBUG - erts_aint_t val = erts_smp_atomic_dectest((erts_smp_atomic_t *) refcp); + erts_aint_t val = erts_smp_atomic_dec_read_nob((erts_smp_atomic_t *) refcp); if (val < min_val) erl_exit(ERTS_ABORT_EXIT, "erts_refc_dec(): Bad refc found (refc=%ld < %ld)!\n", val, min_val); #else - erts_smp_atomic_dec((erts_smp_atomic_t *) refcp); + erts_smp_atomic_dec_nob((erts_smp_atomic_t *) refcp); #endif } ERTS_GLB_INLINE erts_aint_t erts_refc_dectest(erts_refc_t *refcp, erts_aint_t min_val) { - erts_aint_t val = erts_smp_atomic_dectest((erts_smp_atomic_t *) refcp); + erts_aint_t val = erts_smp_atomic_dec_read_nob((erts_smp_atomic_t *) refcp); #ifdef ERTS_REFC_DEBUG if (val < min_val) erl_exit(ERTS_ABORT_EXIT, @@ -1062,20 +1067,20 @@ ERTS_GLB_INLINE void erts_refc_add(erts_refc_t *refcp, erts_aint_t diff, erts_aint_t min_val) { #ifdef ERTS_REFC_DEBUG - erts_aint_t val = erts_smp_atomic_addtest((erts_smp_atomic_t *) refcp, diff); + erts_aint_t val = erts_smp_atomic_add_read_nob((erts_smp_atomic_t *) refcp, diff); if (val < min_val) erl_exit(ERTS_ABORT_EXIT, "erts_refc_add(%ld): Bad refc found (refc=%ld < %ld)!\n", diff, val, min_val); #else - erts_smp_atomic_add((erts_smp_atomic_t *) refcp, diff); + erts_smp_atomic_add_nob((erts_smp_atomic_t *) refcp, diff); #endif } ERTS_GLB_INLINE erts_aint_t erts_refc_read(erts_refc_t *refcp, erts_aint_t min_val) { - erts_aint_t val = erts_smp_atomic_read((erts_smp_atomic_t *) refcp); + erts_aint_t val = erts_smp_atomic_read_nob((erts_smp_atomic_t *) refcp); #ifdef ERTS_REFC_DEBUG if (val < min_val) erl_exit(ERTS_ABORT_EXIT, diff --git a/erts/emulator/beam/time.c b/erts/emulator/beam/time.c index a00faff912..8fa8c1cfe0 100644 --- a/erts/emulator/beam/time.c +++ b/erts/emulator/beam/time.c @@ -108,9 +108,9 @@ static ErlTimer *tiw_min_ptr; static int itime; /* Constant after init */ erts_smp_atomic_t do_time; /* set at clock interrupt */ -static ERTS_INLINE erts_aint_t do_time_read(void) { return erts_smp_atomic_read(&do_time); } +static ERTS_INLINE erts_aint_t do_time_read(void) { return erts_smp_atomic_read_acqb(&do_time); } static ERTS_INLINE erts_aint_t do_time_update(void) { return do_time_read(); } -static ERTS_INLINE void do_time_init(void) { erts_smp_atomic_init(&do_time, 0L); } +static ERTS_INLINE void do_time_init(void) { erts_smp_atomic_init_nob(&do_time, 0L); } /* get the time (in units of itime) to the next timeout, or -1 if there are no timeouts */ diff --git a/erts/emulator/beam/utils.c b/erts/emulator/beam/utils.c index a17de717bc..3f6accba2d 100644 --- a/erts/emulator/beam/utils.c +++ b/erts/emulator/beam/utils.c @@ -3689,14 +3689,16 @@ threads_not_under_control(void) { erts_aint32_t res = system_block_state.threads_to_block; + ERTS_THR_MEMORY_BARRIER; + /* Waiting is always an allowed activity... */ - res -= erts_smp_atomic32_read(&erts_system_block_state.in_activity.wait); + res -= erts_smp_atomic32_read_nob(&erts_system_block_state.in_activity.wait); if (system_block_state.allowed_activities & ERTS_BS_FLG_ALLOW_GC) - res -= erts_smp_atomic32_read(&erts_system_block_state.in_activity.gc); + res -= erts_smp_atomic32_read_nob(&erts_system_block_state.in_activity.gc); if (system_block_state.allowed_activities & ERTS_BS_FLG_ALLOW_IO) - res -= erts_smp_atomic32_read(&erts_system_block_state.in_activity.io); + res -= erts_smp_atomic32_read_nob(&erts_system_block_state.in_activity.io); if (res < 0) { ASSERT(0); @@ -3756,7 +3758,7 @@ erts_block_system(Uint32 allowed_activities) } else { - erts_smp_atomic32_inc(&erts_system_block_state.do_block); + erts_smp_atomic32_inc_nob(&erts_system_block_state.do_block); /* Someone else might be waiting for us to block... */ if (do_block) { @@ -3808,11 +3810,11 @@ erts_emergency_block_system(long timeout, Uint32 allowed_activities) another_blocker = erts_smp_pending_system_block(); system_block_state.emergency = 1; - erts_smp_atomic32_inc(&erts_system_block_state.do_block); + erts_smp_atomic32_inc_nob(&erts_system_block_state.do_block); if (another_blocker) { if (is_blocker()) { - erts_smp_atomic32_dec(&erts_system_block_state.do_block); + erts_smp_atomic32_dec_nob(&erts_system_block_state.do_block); res = 0; goto done; } @@ -3869,7 +3871,7 @@ erts_release_system(void) if (system_block_state.recursive_block) system_block_state.recursive_block--; else { - do_block = erts_smp_atomic32_dectest(&erts_system_block_state.do_block); + do_block = erts_smp_atomic32_dec_read_nob(&erts_system_block_state.do_block); system_block_state.have_blocker = 0; if (is_blockable_thread()) system_block_state.threads_to_block++; @@ -4004,10 +4006,10 @@ erts_system_block_init(void) /* Global state... */ - erts_smp_atomic32_init(&erts_system_block_state.do_block, 0); - erts_smp_atomic32_init(&erts_system_block_state.in_activity.wait, 0); - erts_smp_atomic32_init(&erts_system_block_state.in_activity.gc, 0); - erts_smp_atomic32_init(&erts_system_block_state.in_activity.io, 0); + erts_smp_atomic32_init_nob(&erts_system_block_state.do_block, 0); + erts_smp_atomic32_init_nob(&erts_system_block_state.in_activity.wait, 0); + erts_smp_atomic32_init_nob(&erts_system_block_state.in_activity.gc, 0); + erts_smp_atomic32_init_nob(&erts_system_block_state.in_activity.io, 0); /* Make sure blockable threads unregister when exiting... */ erts_smp_install_exit_handler(erts_unregister_blockable_thread); |