diff options
Diffstat (limited to 'erts/emulator/beam/erl_alloc_util.c')
| -rw-r--r-- | erts/emulator/beam/erl_alloc_util.c | 1384 |
1 files changed, 1333 insertions, 51 deletions
diff --git a/erts/emulator/beam/erl_alloc_util.c b/erts/emulator/beam/erl_alloc_util.c index fed51eb200..fdf355d503 100644 --- a/erts/emulator/beam/erl_alloc_util.c +++ b/erts/emulator/beam/erl_alloc_util.c @@ -48,6 +48,8 @@ #include "erl_mseg.h" #include "erl_threads.h" #include "erl_thr_progress.h" +#include "erl_bif_unique.h" +#include "erl_nif.h" #ifdef ERTS_ENABLE_LOCK_COUNT #include "erl_lock_count.h" @@ -139,6 +141,33 @@ MBC after deallocating first block: [Carrier_t|pad|Block_t 111| udata... ] */ +/* Allocation tags ... + * + * These are added to the footer of every block when enabled. Currently they + * consist of the allocation type and an atom identifying the allocating + * driver/nif (or 'system' if that can't be determined), but the format is not + * supposed to be set in stone. + * + * The packing scheme requires that the atom values are small enough to fit + * into a word with ERTS_ALC_N_BITS to spare. Users must check for overflow + * before MAKE_ATAG(). */ + +typedef UWord alcu_atag_t; + +#define MAKE_ATAG(IdAtom, Type) \ + (ASSERT((Type) >= ERTS_ALC_N_MIN && (Type) <= ERTS_ALC_N_MAX), \ + ASSERT(atom_val(IdAtom) <= MAX_ATAG_ATOM_ID), \ + (atom_val(IdAtom) << ERTS_ALC_N_BITS) | (Type)) + +#define ATAG_ID(AT) (make_atom((AT) >> ERTS_ALC_N_BITS)) +#define ATAG_TYPE(AT) ((AT) & ERTS_ALC_N_MASK) + +#define MAX_ATAG_ATOM_ID (ERTS_UWORD_MAX >> ERTS_ALC_N_BITS) + +#define DBG_IS_VALID_ATAG(Allocator, AT) \ + (ATAG_TYPE(AT) >= ERTS_ALC_N_MIN && \ + ATAG_TYPE(AT) <= ERTS_ALC_N_MAX && \ + (Allocator)->alloc_no == ERTS_ALC_T2A(ERTS_ALC_N2T(ATAG_TYPE(AT)))) /* Blocks ... */ @@ -153,10 +182,17 @@ MBC after deallocating first block: #endif #define FBLK_FTR_SZ (sizeof(FreeBlkFtr_t)) +#define GET_BLK_ATAG(B) \ + (((alcu_atag_t *) (((char *) (B)) + (BLK_SZ(B))))[-1]) +#define SET_BLK_ATAG(B, T) \ + (((alcu_atag_t *) (((char *) (B)) + (BLK_SZ(B))))[-1] = (T)) + +#define BLK_ATAG_SZ(AP) ((AP)->atags ? sizeof(alcu_atag_t) : 0) + #define UMEMSZ2BLKSZ(AP, SZ) \ - (ABLK_HDR_SZ + (SZ) <= (AP)->min_block_size \ + (ABLK_HDR_SZ + BLK_ATAG_SZ(AP) + (SZ) <= (AP)->min_block_size \ ? (AP)->min_block_size \ - : UNIT_CEILING(ABLK_HDR_SZ + (SZ))) + : UNIT_CEILING(ABLK_HDR_SZ + BLK_ATAG_SZ(AP) + (SZ))) #define UMEM2BLK(P) ((Block_t *) (((char *) (P)) - ABLK_HDR_SZ)) #define BLK2UMEM(P) ((void *) (((char *) (P)) + ABLK_HDR_SZ)) @@ -688,6 +724,62 @@ static void destroy_carrier(Allctr_t *, Block_t *, Carrier_t **); static void mbc_free(Allctr_t *allctr, void *p, Carrier_t **busy_pcrr_pp); static void dealloc_block(Allctr_t *, void *, ErtsAlcFixList_t *, int); +static alcu_atag_t determine_alloc_tag(Allctr_t *allocator, ErtsAlcType_t type) +{ + ErtsSchedulerData *esdp; + Eterm id; + + ERTS_CT_ASSERT(_unchecked_atom_val(am_system) <= MAX_ATAG_ATOM_ID); + ASSERT(allocator->atags); + + esdp = erts_get_scheduler_data(); + id = am_system; + + if (esdp) { + if (esdp->current_nif) { + Module *mod = erts_nif_get_module((esdp->current_nif)->mod_nif); + + /* Mod can be NULL if a resource destructor allocates memory after + * the module has been unloaded. */ + if (mod) { + id = make_atom(mod->module); + } + } else if (esdp->current_port) { + Port *p = esdp->current_port; + id = (p->drv_ptr)->name_atom; + } + + /* We fall back to 'system' if we can't pack the driver/NIF name into + * the tag. This may be a bit misleading but we've made no promises + * that the information is complete. + * + * This can only happen on 32-bit emulators when a new driver/NIF has + * been loaded *after* 16 million atoms have been used, and supporting + * that fringe case is not worth an extra word. 64-bit emulators are + * unaffected since the atom cache limits atom indexes to 32 bits. */ + if(MAX_ATOM_TABLE_SIZE > MAX_ATAG_ATOM_ID) { + if (atom_val(id) > MAX_ATAG_ATOM_ID) { + id = am_system; + } + } + } + + return MAKE_ATAG(id, type); +} + +static void set_alloc_tag(Allctr_t *allocator, void *p, alcu_atag_t tag) +{ + Block_t *block; + + ASSERT(DBG_IS_VALID_ATAG(allocator, tag)); + ASSERT(allocator->atags && p); + (void)allocator; + + block = UMEM2BLK(p); + + SET_BLK_ATAG(block, tag); +} + /* internal data... */ #if 0 @@ -4242,6 +4334,7 @@ static struct { Eterm e; Eterm t; Eterm ramv; + Eterm atags; #if HAVE_ERTS_MSEG Eterm asbcst; Eterm rsbcst; @@ -4311,7 +4404,7 @@ static struct { #endif } am; -static Eterm fix_type_atoms[ERTS_ALC_NO_FIXED_SIZES]; +static Eterm alloc_type_atoms[ERTS_ALC_N_MAX + 1]; static ERTS_INLINE void atom_init(Eterm *atom, char *name) { @@ -4342,6 +4435,7 @@ init_atoms(Allctr_t *allctr) AM_INIT(e); AM_INIT(t); AM_INIT(ramv); + AM_INIT(atags); #if HAVE_ERTS_MSEG AM_INIT(asbcst); AM_INIT(rsbcst); @@ -4413,12 +4507,12 @@ init_atoms(Allctr_t *allctr) } #endif - for (ix = 0; ix < ERTS_ALC_NO_FIXED_SIZES; ix++) { - ErtsAlcType_t n = ERTS_ALC_N_MIN_A_FIXED_SIZE + ix; - char *name = (char *) ERTS_ALC_N2TD(n); - size_t len = sys_strlen(name); - fix_type_atoms[ix] = am_atom_put(name, len); - } + for (ix = ERTS_ALC_N_MIN; ix <= ERTS_ALC_N_MAX; ix++) { + const char *name = ERTS_ALC_N2TD(ix); + size_t len = sys_strlen(name); + + alloc_type_atoms[ix] = am_atom_put(name, len); + } } if (allctr && !allctr->atoms_initialized) { @@ -4531,6 +4625,7 @@ sz_info_fix(Allctr_t *allctr, ErtsAlcFixList_t *fix = &allctr->fix[ix]; UWord alloced = fix->type_size * fix->u.cpool.allocated; UWord used = fix->type_size * fix->u.cpool.used; + ErtsAlcType_t n = ERTS_ALC_N_MIN_A_FIXED_SIZE + ix; if (print_to_p) { fmtfn_t to = *print_to_p; @@ -4538,15 +4633,14 @@ sz_info_fix(Allctr_t *allctr, erts_print(to, arg, "fix type internal: %s %bpu %bpu\n", - (char *) ERTS_ALC_N2TD(ERTS_ALC_N_MIN_A_FIXED_SIZE - + ix), + (char *) ERTS_ALC_N2TD(n), alloced, used); } if (hpp || szp) { add_3tup(hpp, szp, &res, - fix_type_atoms[ix], + alloc_type_atoms[n], bld_unstable_uint(hpp, szp, alloced), bld_unstable_uint(hpp, szp, used)); } @@ -4559,6 +4653,7 @@ sz_info_fix(Allctr_t *allctr, ErtsAlcFixList_t *fix = &allctr->fix[ix]; UWord alloced = fix->type_size * fix->u.nocpool.allocated; UWord used = fix->type_size*fix->u.nocpool.used; + ErtsAlcType_t n = ERTS_ALC_N_MIN_A_FIXED_SIZE + ix; if (print_to_p) { fmtfn_t to = *print_to_p; @@ -4566,15 +4661,14 @@ sz_info_fix(Allctr_t *allctr, erts_print(to, arg, "fix type: %s %bpu %bpu\n", - (char *) ERTS_ALC_N2TD(ERTS_ALC_N_MIN_A_FIXED_SIZE - + ix), + (char *) ERTS_ALC_N2TD(n), alloced, used); } if (hpp || szp) { add_3tup(hpp, szp, &res, - fix_type_atoms[ix], + alloc_type_atoms[n], bld_unstable_uint(hpp, szp, alloced), bld_unstable_uint(hpp, szp, used)); } @@ -5000,6 +5094,7 @@ info_options(Allctr_t *allctr, "option e: true\n" "option t: %s\n" "option ramv: %s\n" + "option atags: %s\n" "option sbct: %beu\n" #if HAVE_ERTS_MSEG "option asbcst: %bpu\n" @@ -5018,6 +5113,7 @@ info_options(Allctr_t *allctr, "option acul: %bpu\n", topt, allctr->ramv ? "true" : "false", + allctr->atags ? "true" : "false", allctr->sbc_threshold, #if HAVE_ERTS_MSEG allctr->mseg_opt.abs_shrink_th, @@ -5087,6 +5183,7 @@ info_options(Allctr_t *allctr, am_sbct, bld_uint(hpp, szp, allctr->sbc_threshold)); add_2tup(hpp, szp, &res, am.ramv, allctr->ramv ? am_true : am_false); + add_2tup(hpp, szp, &res, am.atags, allctr->atags ? am_true : am_false); add_2tup(hpp, szp, &res, am.t, (allctr->t ? am_true : am_false)); add_2tup(hpp, szp, &res, am.e, am_true); } @@ -5408,9 +5505,8 @@ erts_alcu_current_size(Allctr_t *allctr, AllctrSize_t *size, ErtsAlcUFixInfo_t * /* ----------------------------------------------------------------------- */ static ERTS_INLINE void * -do_erts_alcu_alloc(ErtsAlcType_t type, void *extra, Uint size) +do_erts_alcu_alloc(ErtsAlcType_t type, Allctr_t *allctr, Uint size) { - Allctr_t *allctr = (Allctr_t *) extra; void *res; ASSERT(initialized); @@ -5449,10 +5545,19 @@ do_erts_alcu_alloc(ErtsAlcType_t type, void *extra, Uint size) void *erts_alcu_alloc(ErtsAlcType_t type, void *extra, Uint size) { + Allctr_t *allctr = (Allctr_t *) extra; void *res; + ASSERT(!"This is not thread safe"); - res = do_erts_alcu_alloc(type, extra, size); + + res = do_erts_alcu_alloc(type, allctr, size); + + if (allctr->atags && res) { + set_alloc_tag(allctr, res, determine_alloc_tag(allctr, type)); + } + DEBUG_CHECK_ALIGNMENT(res); + return res; } @@ -5462,13 +5567,25 @@ void * erts_alcu_alloc_ts(ErtsAlcType_t type, void *extra, Uint size) { Allctr_t *allctr = (Allctr_t *) extra; + alcu_atag_t tag = 0; void *res; + + if (allctr->atags) { + tag = determine_alloc_tag(allctr, type); + } + erts_mtx_lock(&allctr->mutex); - res = do_erts_alcu_alloc(type, extra, size); - DEBUG_CHECK_ALIGNMENT(res); + res = do_erts_alcu_alloc(type, allctr, size); + + if (allctr->atags && res) { + set_alloc_tag(allctr, res, tag); + } erts_mtx_unlock(&allctr->mutex); + + DEBUG_CHECK_ALIGNMENT(res); + return res; } @@ -5478,6 +5595,7 @@ erts_alcu_alloc_thr_spec(ErtsAlcType_t type, void *extra, Uint size) { ErtsAllocatorThrSpec_t *tspec = (ErtsAllocatorThrSpec_t *) extra; int ix; + alcu_atag_t tag = 0; Allctr_t *allctr; void *res; @@ -5487,11 +5605,19 @@ erts_alcu_alloc_thr_spec(ErtsAlcType_t type, void *extra, Uint size) allctr = tspec->allctr[ix]; + if (allctr->atags) { + tag = determine_alloc_tag(allctr, type); + } + if (allctr->thread_safe) erts_mtx_lock(&allctr->mutex); res = do_erts_alcu_alloc(type, allctr, size); + if (allctr->atags && res) { + set_alloc_tag(allctr, res, tag); + } + if (allctr->thread_safe) erts_mtx_unlock(&allctr->mutex); @@ -5504,10 +5630,15 @@ void * erts_alcu_alloc_thr_pref(ErtsAlcType_t type, void *extra, Uint size) { Allctr_t *pref_allctr; + alcu_atag_t tag = 0; void *res; pref_allctr = get_pref_allctr(extra); + if (pref_allctr->atags) { + tag = determine_alloc_tag(pref_allctr, type); + } + if (pref_allctr->thread_safe) erts_mtx_lock(&pref_allctr->mutex); @@ -5523,12 +5654,15 @@ erts_alcu_alloc_thr_pref(ErtsAlcType_t type, void *extra, Uint size) res = do_erts_alcu_alloc(type, pref_allctr, size); } + if (pref_allctr->atags && res) { + set_alloc_tag(pref_allctr, res, tag); + } + if (pref_allctr->thread_safe) erts_mtx_unlock(&pref_allctr->mutex); DEBUG_CHECK_ALIGNMENT(res); - return res; } @@ -5537,10 +5671,9 @@ erts_alcu_alloc_thr_pref(ErtsAlcType_t type, void *extra, Uint size) /* ------------------------------------------------------------------------- */ static ERTS_INLINE void -do_erts_alcu_free(ErtsAlcType_t type, void *extra, void *p, +do_erts_alcu_free(ErtsAlcType_t type, Allctr_t *allctr, void *p, Carrier_t **busy_pcrr_pp) { - Allctr_t *allctr = (Allctr_t *) extra; ASSERT(initialized); ASSERT(allctr); @@ -5572,7 +5705,8 @@ do_erts_alcu_free(ErtsAlcType_t type, void *extra, void *p, void erts_alcu_free(ErtsAlcType_t type, void *extra, void *p) { - do_erts_alcu_free(type, extra, p, NULL); + Allctr_t *allctr = (Allctr_t *) extra; + do_erts_alcu_free(type, allctr, p, NULL); } @@ -5581,7 +5715,7 @@ erts_alcu_free_ts(ErtsAlcType_t type, void *extra, void *p) { Allctr_t *allctr = (Allctr_t *) extra; erts_mtx_lock(&allctr->mutex); - do_erts_alcu_free(type, extra, p, NULL); + do_erts_alcu_free(type, allctr, p, NULL); erts_mtx_unlock(&allctr->mutex); } @@ -5641,13 +5775,12 @@ erts_alcu_free_thr_pref(ErtsAlcType_t type, void *extra, void *p) static ERTS_INLINE void * do_erts_alcu_realloc(ErtsAlcType_t type, - void *extra, + Allctr_t *allctr, void *p, Uint size, Uint32 alcu_flgs, Carrier_t **busy_pcrr_pp) { - Allctr_t *allctr = (Allctr_t *) extra; Block_t *blk; void *res; @@ -5661,7 +5794,7 @@ do_erts_alcu_realloc(ErtsAlcType_t type, ERTS_ALCU_DBG_CHK_THR_ACCESS(allctr); if (!p) { - res = do_erts_alcu_alloc(type, extra, size); + res = do_erts_alcu_alloc(type, allctr, size); INC_CC(allctr->calls.this_realloc); DEC_CC(allctr->calls.this_alloc); return res; @@ -5670,7 +5803,7 @@ do_erts_alcu_realloc(ErtsAlcType_t type, #if ALLOC_ZERO_EQ_NULL if (!size) { ASSERT(p); - do_erts_alcu_free(type, extra, p, busy_pcrr_pp); + do_erts_alcu_free(type, allctr, p, busy_pcrr_pp); INC_CC(allctr->calls.this_realloc); DEC_CC(allctr->calls.this_free); return NULL; @@ -5755,19 +5888,29 @@ do_erts_alcu_realloc(ErtsAlcType_t type, void * erts_alcu_realloc(ErtsAlcType_t type, void *extra, void *p, Uint size) { + Allctr_t *allctr = (Allctr_t *)extra; void *res; - res = do_erts_alcu_realloc(type, extra, p, size, 0, NULL); + + res = do_erts_alcu_realloc(type, allctr, p, size, 0, NULL); + DEBUG_CHECK_ALIGNMENT(res); + + if (allctr->atags && res) { + set_alloc_tag(allctr, res, determine_alloc_tag(allctr, type)); + } + return res; } void * erts_alcu_realloc_mv(ErtsAlcType_t type, void *extra, void *p, Uint size) { + Allctr_t *allctr = (Allctr_t *)extra; void *res; - res = do_erts_alcu_alloc(type, extra, size); + + res = do_erts_alcu_alloc(type, allctr, size); if (!res) - res = erts_alcu_realloc(type, extra, p, size); + res = do_erts_alcu_realloc(type, allctr, p, size, 0, NULL); else { Block_t *blk; size_t cpy_size; @@ -5777,23 +5920,42 @@ erts_alcu_realloc_mv(ErtsAlcType_t type, void *extra, void *p, Uint size) if (cpy_size > size) cpy_size = size; sys_memcpy(res, p, cpy_size); - do_erts_alcu_free(type, extra, p, NULL); + do_erts_alcu_free(type, allctr, p, NULL); } + DEBUG_CHECK_ALIGNMENT(res); + + if (allctr->atags && res) { + set_alloc_tag(allctr, res, determine_alloc_tag(allctr, type)); + } + return res; } - void * erts_alcu_realloc_ts(ErtsAlcType_t type, void *extra, void *ptr, Uint size) { Allctr_t *allctr = (Allctr_t *) extra; + alcu_atag_t tag = 0; void *res; + + if (allctr->atags) { + tag = determine_alloc_tag(allctr, type); + } + erts_mtx_lock(&allctr->mutex); - res = do_erts_alcu_realloc(type, extra, ptr, size, 0, NULL); + + res = do_erts_alcu_realloc(type, allctr, ptr, size, 0, NULL); + + if (allctr->atags && res) { + set_alloc_tag(allctr, res, tag); + } + erts_mtx_unlock(&allctr->mutex); + DEBUG_CHECK_ALIGNMENT(res); + return res; } @@ -5801,11 +5963,17 @@ void * erts_alcu_realloc_mv_ts(ErtsAlcType_t type, void *extra, void *p, Uint size) { Allctr_t *allctr = (Allctr_t *) extra; + alcu_atag_t tag = 0; void *res; + + if (allctr->atags) { + tag = determine_alloc_tag(allctr, type); + } + erts_mtx_lock(&allctr->mutex); - res = do_erts_alcu_alloc(type, extra, size); + res = do_erts_alcu_alloc(type, allctr, size); if (!res) - res = do_erts_alcu_realloc(type, extra, p, size, 0, NULL); + res = do_erts_alcu_realloc(type, allctr, p, size, 0, NULL); else { Block_t *blk; size_t cpy_size; @@ -5815,10 +5983,17 @@ erts_alcu_realloc_mv_ts(ErtsAlcType_t type, void *extra, void *p, Uint size) if (cpy_size > size) cpy_size = size; sys_memcpy(res, p, cpy_size); - do_erts_alcu_free(type, extra, p, NULL); + do_erts_alcu_free(type, allctr, p, NULL); } + + if (allctr->atags && res) { + set_alloc_tag(allctr, res, tag); + } + erts_mtx_unlock(&allctr->mutex); + DEBUG_CHECK_ALIGNMENT(res); + return res; } @@ -5829,6 +6004,7 @@ erts_alcu_realloc_thr_spec(ErtsAlcType_t type, void *extra, { ErtsAllocatorThrSpec_t *tspec = (ErtsAllocatorThrSpec_t *) extra; int ix; + alcu_atag_t tag = 0; Allctr_t *allctr; void *res; @@ -5838,11 +6014,19 @@ erts_alcu_realloc_thr_spec(ErtsAlcType_t type, void *extra, allctr = tspec->allctr[ix]; + if (allctr->atags) { + tag = determine_alloc_tag(allctr, type); + } + if (allctr->thread_safe) erts_mtx_lock(&allctr->mutex); res = do_erts_alcu_realloc(type, allctr, ptr, size, 0, NULL); + if (allctr->atags && res) { + set_alloc_tag(allctr, res, tag); + } + if (allctr->thread_safe) erts_mtx_unlock(&allctr->mutex); @@ -5857,6 +6041,7 @@ erts_alcu_realloc_mv_thr_spec(ErtsAlcType_t type, void *extra, { ErtsAllocatorThrSpec_t *tspec = (ErtsAllocatorThrSpec_t *) extra; int ix; + alcu_atag_t tag = 0; Allctr_t *allctr; void *res; @@ -5866,14 +6051,16 @@ erts_alcu_realloc_mv_thr_spec(ErtsAlcType_t type, void *extra, allctr = tspec->allctr[ix]; + if (allctr->atags) { + tag = determine_alloc_tag(allctr, type); + } + if (allctr->thread_safe) erts_mtx_lock(&allctr->mutex); res = do_erts_alcu_alloc(type, allctr, size); if (!res) { - if (allctr->thread_safe) - erts_mtx_unlock(&allctr->mutex); - res = erts_alcu_realloc_thr_spec(type, allctr, ptr, size); + res = do_erts_alcu_realloc(type, allctr, ptr, size, 0, NULL); } else { Block_t *blk; @@ -5885,29 +6072,34 @@ erts_alcu_realloc_mv_thr_spec(ErtsAlcType_t type, void *extra, cpy_size = size; sys_memcpy(res, ptr, cpy_size); do_erts_alcu_free(type, allctr, ptr, NULL); - if (allctr->thread_safe) - erts_mtx_unlock(&allctr->mutex); } + if (allctr->atags && res) { + set_alloc_tag(allctr, res, tag); + } + + if (allctr->thread_safe) + erts_mtx_unlock(&allctr->mutex); + DEBUG_CHECK_ALIGNMENT(res); return res; } static ERTS_INLINE void * -realloc_thr_pref(ErtsAlcType_t type, void *extra, void *p, Uint size, +realloc_thr_pref(ErtsAlcType_t type, Allctr_t *pref_allctr, void *p, Uint size, int force_move) { void *res; - Allctr_t *pref_allctr, *used_allctr; + Allctr_t *used_allctr; UWord old_user_size; Carrier_t *busy_pcrr_p; + alcu_atag_t tag = 0; int retried; - if (!p) - return erts_alcu_alloc_thr_pref(type, extra, size); - - pref_allctr = get_pref_allctr(extra); + if (pref_allctr->atags) { + tag = determine_alloc_tag(pref_allctr, type); + } if (pref_allctr->thread_safe) erts_mtx_lock(&pref_allctr->mutex); @@ -5936,6 +6128,11 @@ restart: retried = 1; goto restart; } + + if (pref_allctr->atags && res) { + set_alloc_tag(pref_allctr, res, tag); + } + if (pref_allctr->thread_safe) erts_mtx_unlock(&pref_allctr->mutex); } @@ -5944,6 +6141,9 @@ restart: if (!res) goto unlock_ts_return; else { + if (pref_allctr->atags) { + set_alloc_tag(pref_allctr, res, tag); + } DEBUG_CHECK_ALIGNMENT(res); @@ -5974,20 +6174,34 @@ restart: } } + DEBUG_CHECK_ALIGNMENT(res); + return res; } void * erts_alcu_realloc_thr_pref(ErtsAlcType_t type, void *extra, void *p, Uint size) { - return realloc_thr_pref(type, extra, p, size, 0); + if (p) { + Allctr_t *pref_allctr = get_pref_allctr(extra); + + return realloc_thr_pref(type, pref_allctr, p, size, 0); + } + + return erts_alcu_alloc_thr_pref(type, extra, size); } void * erts_alcu_realloc_mv_thr_pref(ErtsAlcType_t type, void *extra, void *p, Uint size) { - return realloc_thr_pref(type, extra, p, size, 1); + if (p) { + Allctr_t *pref_allctr = get_pref_allctr(extra); + + return realloc_thr_pref(type, pref_allctr, p, size, 1); + } + + return erts_alcu_alloc_thr_pref(type, extra, size); } @@ -6071,6 +6285,7 @@ erts_alcu_start(Allctr_t *allctr, AllctrInit_t *init) allctr->t = 0; allctr->ramv = init->ramv; + allctr->atags = init->atags; allctr->main_carrier_size = init->mmbcs; #if HAVE_ERTS_MSEG @@ -6320,6 +6535,1072 @@ erts_alcu_init(AlcUInit_t *init) initialized = 1; } +/* ------------------------------------------------------------------------- */ + +/* Allocation histograms and carrier information is gathered by walking through + * all carriers associated with each allocator instance. This is done as + * aux_yield_work on the scheduler that owns each instance. + * + * Yielding is implemented by temporarily inserting a "dummy carrier" at the + * last position. It's permanently "busy" so it won't get picked up by someone + * else when in the carrier pool, and we never make the employer aware of it + * through callbacks so we can't accidentally allocate on it. + * + * Plain malloc/free is used to guarantee we won't allocate with the allocator + * we're scanning. */ + +/* Yield between carriers once this many blocks have been processed. Note that + * a single carrier scan may exceed this figure. */ +#ifndef DEBUG + #define BLOCKSCAN_REDUCTIONS (8000) +#else + #define BLOCKSCAN_REDUCTIONS (400) +#endif + +/* Abort a single carrier scan after this many blocks to prevent really large + * MBCs from blocking forever. */ +#define BLOCKSCAN_BAILOUT_THRESHOLD (16000) + +typedef struct alcu_blockscan { + /* A per-scheduler list used when multiple scans have been queued. The + * current scanner will always run until completion/abort before moving on + * to the next. */ + struct alcu_blockscan *scanner_queue; + + Allctr_t *allocator; + Process *process; + + int (*current_op)(struct alcu_blockscan *scanner); + int (*next_op)(struct alcu_blockscan *scanner); + int reductions; + + ErtsAlcCPoolData_t *cpool_cursor; + CarrierList_t *current_clist; + Carrier_t *clist_cursor; + Carrier_t dummy_carrier; + + /* Called if the process that started this job dies before we're done. */ + void (*abort)(void *user_data); + + /* Called on each carrier. The callback must return the number of blocks + * scanned to yield properly between carriers. + * + * Note that it's not possible to "yield back" into a carrier. */ + int (*scan)(Allctr_t *, void *user_data, Carrier_t *); + + /* Called when all carriers have been scanned. The callback may return + * non-zero to yield. */ + int (*finish)(void *user_data); + + void *user_data; +} blockscan_t; + +static Carrier_t *blockscan_restore_clist_cursor(blockscan_t *state) +{ + Carrier_t *cursor = state->clist_cursor; + + ASSERT(state->clist_cursor == (state->current_clist)->first || + state->clist_cursor == &state->dummy_carrier); + + if (cursor == &state->dummy_carrier) { + cursor = cursor->next; + + unlink_carrier(state->current_clist, state->clist_cursor); + } + + return cursor; +} + +static void blockscan_save_clist_cursor(blockscan_t *state, Carrier_t *after) +{ + ASSERT(state->clist_cursor == (state->current_clist)->first || + state->clist_cursor == &state->dummy_carrier); + + state->clist_cursor = &state->dummy_carrier; + + (state->clist_cursor)->next = after->next; + (state->clist_cursor)->prev = after; + + relink_carrier(state->current_clist, state->clist_cursor); +} + +static int blockscan_clist_yielding(blockscan_t *state) +{ + Carrier_t *cursor = blockscan_restore_clist_cursor(state); + + if (ERTS_PROC_IS_EXITING(state->process)) { + return 0; + } + + while (cursor) { + /* Skip dummy carriers inserted by another (concurrent) block scan. + * This can happen when scanning thread-safe allocators from multiple + * schedulers. */ + if (CARRIER_SZ(cursor) > 0) { + int blocks_scanned = state->scan(state->allocator, + state->user_data, + cursor); + + state->reductions -= blocks_scanned; + + if (state->reductions <= 0) { + blockscan_save_clist_cursor(state, cursor); + return 1; + } + } + + cursor = cursor->next; + } + + return 0; +} + +static ErtsAlcCPoolData_t *blockscan_restore_cpool_cursor(blockscan_t *state) +{ + ErtsAlcCPoolData_t *cursor; + + cursor = cpool_aint2cpd(cpool_read(&(state->cpool_cursor)->next)); + + if (state->cpool_cursor == &state->dummy_carrier.cpool) { + cpool_delete(state->allocator, state->allocator, &state->dummy_carrier); + } + + return cursor; +} + +static void blockscan_save_cpool_cursor(blockscan_t *state, + ErtsAlcCPoolData_t *after) +{ + ErtsAlcCPoolData_t *dummy_carrier, *prev_carrier, *next_carrier; + + dummy_carrier = &state->dummy_carrier.cpool; + + next_carrier = cpool_aint2cpd(cpool_mod_mark(&after->next)); + prev_carrier = cpool_aint2cpd(cpool_mod_mark(&next_carrier->prev)); + + cpool_init(&dummy_carrier->next, (erts_aint_t)next_carrier); + cpool_init(&dummy_carrier->prev, (erts_aint_t)prev_carrier); + + cpool_set_mod_marked(&prev_carrier->next, + (erts_aint_t)dummy_carrier, + (erts_aint_t)next_carrier); + cpool_set_mod_marked(&next_carrier->prev, + (erts_aint_t)dummy_carrier, + (erts_aint_t)prev_carrier); + + state->cpool_cursor = dummy_carrier; +} + +static int blockscan_cpool_yielding(blockscan_t *state) +{ + ErtsAlcCPoolData_t *sentinel, *cursor; + + sentinel = &carrier_pool[(state->allocator)->alloc_no].sentinel; + cursor = blockscan_restore_cpool_cursor(state); + + if (ERTS_PROC_IS_EXITING(state->process)) { + return 0; + } + + while (cursor != sentinel) { + Carrier_t *carrier; + erts_aint_t exp; + + /* When a deallocation happens on a pooled carrier it will be routed to + * its owner, so the only way to be sure that it isn't modified while + * scanning is to skip all carriers that aren't ours. The deallocations + * deferred to us will get handled when we're done. */ + while (cursor->orig_allctr != state->allocator) { + cursor = cpool_aint2cpd(cpool_read(&cursor->next)); + + if (cursor == sentinel) { + return 0; + } + } + + carrier = ErtsContainerStruct(cursor, Carrier_t, cpool); + exp = erts_atomic_read_rb(&carrier->allctr); + + if (exp & ERTS_CRR_ALCTR_FLG_IN_POOL) { + ASSERT(state->allocator == (Allctr_t*)(exp & ~ERTS_CRR_ALCTR_FLG_MASK)); + ASSERT(!(exp & ERTS_CRR_ALCTR_FLG_BUSY)); + + if (erts_atomic_cmpxchg_acqb(&carrier->allctr, + exp | ERTS_CRR_ALCTR_FLG_BUSY, + exp) == exp) { + /* Skip dummy carriers inserted by another (concurrent) block + * scan. This can happen when scanning thread-safe allocators + * from multiple schedulers. */ + if (CARRIER_SZ(carrier) > 0) { + int blocks_scanned = state->scan(state->allocator, + state->user_data, + carrier); + + state->reductions -= blocks_scanned; + + if (state->reductions <= 0) { + blockscan_save_cpool_cursor(state, cursor); + erts_atomic_set_relb(&carrier->allctr, exp); + + return 1; + } + } + + erts_atomic_set_relb(&carrier->allctr, exp); + } + } + + cursor = cpool_aint2cpd(cpool_read(&cursor->next)); + } + + return 0; +} + +static int blockscan_yield_helper(blockscan_t *state, + int (*yielding_op)(blockscan_t*)) +{ + /* Note that we don't check whether to abort here; only yielding_op knows + * whether the carrier is still in the list/pool. */ + + if ((state->allocator)->thread_safe) { + /* Locked scans have to be as short as possible. */ + state->reductions = 1; + + erts_mtx_lock(&(state->allocator)->mutex); + } else { + state->reductions = BLOCKSCAN_REDUCTIONS; + } + + if (yielding_op(state)) { + state->next_op = state->current_op; + } + + if ((state->allocator)->thread_safe) { + erts_mtx_unlock(&(state->allocator)->mutex); + } + + return 1; +} + +/* */ + +static int blockscan_finish(blockscan_t *state) +{ + if (ERTS_PROC_IS_EXITING(state->process)) { + state->abort(state->user_data); + return 0; + } + + state->current_op = blockscan_finish; + + return state->finish(state->user_data); +} + +static int blockscan_sweep_sbcs(blockscan_t *state) +{ + if (state->current_op != blockscan_sweep_sbcs) { + SET_CARRIER_HDR(&state->dummy_carrier, 0, SCH_SBC, state->allocator); + state->current_clist = &(state->allocator)->sbc_list; + state->clist_cursor = (state->current_clist)->first; + } + + state->current_op = blockscan_sweep_sbcs; + state->next_op = blockscan_finish; + + return blockscan_yield_helper(state, blockscan_clist_yielding); +} + +static int blockscan_sweep_mbcs(blockscan_t *state) +{ + if (state->current_op != blockscan_sweep_mbcs) { + SET_CARRIER_HDR(&state->dummy_carrier, 0, SCH_MBC, state->allocator); + state->current_clist = &(state->allocator)->mbc_list; + state->clist_cursor = (state->current_clist)->first; + } + + state->current_op = blockscan_sweep_mbcs; + state->next_op = blockscan_sweep_sbcs; + + return blockscan_yield_helper(state, blockscan_clist_yielding); +} + +static int blockscan_sweep_cpool(blockscan_t *state) +{ + if (state->current_op != blockscan_sweep_cpool) { + ErtsAlcCPoolData_t *sentinel; + + SET_CARRIER_HDR(&state->dummy_carrier, 0, SCH_MBC, state->allocator); + sentinel = &carrier_pool[(state->allocator)->alloc_no].sentinel; + state->cpool_cursor = sentinel; + } + + state->current_op = blockscan_sweep_cpool; + state->next_op = blockscan_sweep_mbcs; + + return blockscan_yield_helper(state, blockscan_cpool_yielding); +} + +static int blockscan_get_specific_allocator(int allocator_num, + int sched_id, + Allctr_t **out) +{ + ErtsAllocatorInfo_t *ai; + Allctr_t *allocator; + + ASSERT(allocator_num >= ERTS_ALC_A_MIN && + allocator_num <= ERTS_ALC_A_MAX); + ASSERT(sched_id >= 0 && sched_id <= erts_no_schedulers); + + ai = &erts_allctrs_info[allocator_num]; + + if (!ai->enabled || !ai->alloc_util) { + return 0; + } + + if (!ai->thr_spec) { + if (sched_id != 0) { + /* Only thread-specific allocators can be scanned on a specific + * scheduler. */ + return 0; + } + + allocator = (Allctr_t*)ai->extra; + ASSERT(allocator->thread_safe); + } else { + ErtsAllocatorThrSpec_t *tspec = (ErtsAllocatorThrSpec_t*)ai->extra; + + ASSERT(sched_id < tspec->size); + + allocator = tspec->allctr[sched_id]; + } + + *out = allocator; + + return 1; +} + +static void blockscan_sched_trampoline(void *arg) +{ + ErtsAlcuBlockscanYieldData *yield; + ErtsSchedulerData *esdp; + blockscan_t *scanner; + + esdp = erts_get_scheduler_data(); + scanner = (blockscan_t*)arg; + + yield = ERTS_SCHED_AUX_YIELD_DATA(esdp, alcu_blockscan); + + ASSERT((yield->last == NULL) == (yield->current == NULL)); + + if (yield->last != NULL) { + blockscan_t *prev_scanner = yield->last; + + ASSERT(prev_scanner->scanner_queue == NULL); + + prev_scanner->scanner_queue = scanner; + } else { + yield->current = scanner; + } + + scanner->scanner_queue = NULL; + yield->last = scanner; + + erts_notify_new_aux_yield_work(esdp); +} + +static void blockscan_dispatch(blockscan_t *scanner, Process *owner, + Allctr_t *allocator, int sched_id) +{ + ASSERT(erts_get_scheduler_id() != 0); + + if (sched_id == 0) { + /* Global instances are always handled on the current scheduler. */ + sched_id = ERTS_ALC_GET_THR_IX(); + ASSERT(allocator->thread_safe); + } + + scanner->allocator = allocator; + scanner->process = owner; + + erts_proc_inc_refc(scanner->process); + + cpool_init_carrier_data(scanner->allocator, &scanner->dummy_carrier); + erts_atomic_init_nob(&(scanner->dummy_carrier).allctr, + (erts_aint_t)allocator | ERTS_CRR_ALCTR_FLG_BUSY); + + if (ERTS_ALC_IS_CPOOL_ENABLED(scanner->allocator)) { + scanner->next_op = blockscan_sweep_cpool; + } else { + scanner->next_op = blockscan_sweep_mbcs; + } + + /* Aux yield jobs can only be set up while running on the scheduler that + * services them, so we move there before continuing. + * + * We can't drive the scan itself through this since the scheduler will + * always finish *all* misc aux work in one go which makes it impossible to + * yield. */ + erts_schedule_misc_aux_work(sched_id, blockscan_sched_trampoline, scanner); +} + +int erts_handle_yielded_alcu_blockscan(ErtsSchedulerData *esdp, + ErtsAlcuBlockscanYieldData *yield) +{ + blockscan_t *scanner = yield->current; + + (void)esdp; + + ASSERT((yield->last == NULL) == (yield->current == NULL)); + + if (scanner) { + if (scanner->next_op(scanner)) { + return 1; + } + + ASSERT(ERTS_PROC_IS_EXITING(scanner->process) || + scanner->current_op == blockscan_finish); + + yield->current = scanner->scanner_queue; + + if (yield->current == NULL) { + ASSERT(scanner == yield->last); + yield->last = NULL; + } + + erts_proc_dec_refc(scanner->process); + + /* Plain free is intentional. */ + free(scanner); + + return yield->current != NULL; + } + + return 0; +} + +void erts_alcu_sched_spec_data_init(ErtsSchedulerData *esdp) +{ + ErtsAlcuBlockscanYieldData *yield; + + yield = ERTS_SCHED_AUX_YIELD_DATA(esdp, alcu_blockscan); + + yield->current = NULL; + yield->last = NULL; +} + +/* ------------------------------------------------------------------------- */ + +static ERTS_INLINE int u64_log2(Uint64 v) +{ + static const int log2_tab64[64] = { + 63, 0, 58, 1, 59, 47, 53, 2, + 60, 39, 48, 27, 54, 33, 42, 3, + 61, 51, 37, 40, 49, 18, 28, 20, + 55, 30, 34, 11, 43, 14, 22, 4, + 62, 57, 46, 52, 38, 26, 32, 41, + 50, 36, 17, 19, 29, 10, 13, 21, + 56, 45, 25, 31, 35, 16, 9, 12, + 44, 24, 15, 8, 23, 7, 6, 5}; + + v |= v >> 1; + v |= v >> 2; + v |= v >> 4; + v |= v >> 8; + v |= v >> 16; + v |= v >> 32; + + return log2_tab64[((Uint64)((v - (v >> 1))*0x07EDD5E59A4E28C2)) >> 58]; +} + +/* ------------------------------------------------------------------------- */ + +typedef struct hist_tree__ { + struct hist_tree__ *parent; + struct hist_tree__ *left; + struct hist_tree__ *right; + + int is_red; + + alcu_atag_t tag; + UWord histogram[1]; +} hist_tree_t; + +#define ERTS_RBT_PREFIX hist_tree +#define ERTS_RBT_T hist_tree_t +#define ERTS_RBT_KEY_T UWord +#define ERTS_RBT_FLAGS_T int +#define ERTS_RBT_INIT_EMPTY_TNODE(T) ((void)0) +#define ERTS_RBT_IS_RED(T) ((T)->is_red) +#define ERTS_RBT_SET_RED(T) ((T)->is_red = 1) +#define ERTS_RBT_IS_BLACK(T) (!ERTS_RBT_IS_RED(T)) +#define ERTS_RBT_SET_BLACK(T) ((T)->is_red = 0) +#define ERTS_RBT_GET_FLAGS(T) ((T)->is_red) +#define ERTS_RBT_SET_FLAGS(T, F) ((T)->is_red = F) +#define ERTS_RBT_GET_PARENT(T) ((T)->parent) +#define ERTS_RBT_SET_PARENT(T, P) ((T)->parent = P) +#define ERTS_RBT_GET_RIGHT(T) ((T)->right) +#define ERTS_RBT_SET_RIGHT(T, R) ((T)->right = (R)) +#define ERTS_RBT_GET_LEFT(T) ((T)->left) +#define ERTS_RBT_SET_LEFT(T, L) ((T)->left = (L)) +#define ERTS_RBT_GET_KEY(T) ((T)->tag) +#define ERTS_RBT_IS_LT(KX, KY) (KX < KY) +#define ERTS_RBT_IS_EQ(KX, KY) (KX == KY) +#define ERTS_RBT_WANT_FOREACH_DESTROY_YIELDING +#define ERTS_RBT_WANT_FOREACH_DESTROY +#define ERTS_RBT_WANT_INSERT +#define ERTS_RBT_WANT_LOOKUP +#define ERTS_RBT_UNDEF + +#include "erl_rbtree.h" + +typedef struct { + blockscan_t common; + + ErtsIRefStorage iref; + Process *process; + + hist_tree_rbt_yield_state_t hist_tree_yield; + hist_tree_t *hist_tree; + UWord hist_count; + + UWord hist_slot_start; + int hist_slot_count; + + UWord unscanned_size; + + ErtsHeapFactory msg_factory; + int building_result; + Eterm result_list; +} gather_ahist_t; + +static void gather_ahist_update(gather_ahist_t *state, UWord tag, UWord size) +{ + hist_tree_t *hist_node; + UWord size_interval; + int hist_slot; + + hist_node = hist_tree_rbt_lookup(state->hist_tree, tag); + + if (hist_node == NULL) { + /* Plain calloc is intentional. */ + hist_node = (hist_tree_t*)calloc(1, sizeof(hist_tree_t) + + (state->hist_slot_count - 1) * + sizeof(hist_node->histogram[0])); + hist_node->tag = tag; + + hist_tree_rbt_insert(&state->hist_tree, hist_node); + state->hist_count++; + } + + size_interval = (size / state->hist_slot_start); + size_interval = u64_log2(size_interval + 1); + + hist_slot = MIN(size_interval, state->hist_slot_count - 1); + + hist_node->histogram[hist_slot]++; +} + +static int gather_ahist_scan(Allctr_t *allocator, + void *user_data, + Carrier_t *carrier) +{ + gather_ahist_t *state; + int blocks_scanned; + Block_t *block; + + state = (gather_ahist_t*)user_data; + blocks_scanned = 1; + + if (IS_SB_CARRIER(carrier)) { + alcu_atag_t tag; + + block = SBC2BLK(allocator, carrier); + tag = GET_BLK_ATAG(block); + + ASSERT(DBG_IS_VALID_ATAG(allocator, tag)); + + gather_ahist_update(state, tag, SBC_BLK_SZ(block)); + } else { + UWord scanned_bytes = MBC_HEADER_SIZE(allocator); + + ASSERT(IS_MB_CARRIER(carrier)); + + block = MBC_TO_FIRST_BLK(allocator, carrier); + + while (1) { + UWord block_size = MBC_BLK_SZ(block); + + if (IS_ALLOCED_BLK(block)) { + alcu_atag_t tag = GET_BLK_ATAG(block); + + ASSERT(DBG_IS_VALID_ATAG(allocator, tag)); + + gather_ahist_update(state, tag, block_size); + } + + scanned_bytes += block_size; + + if (blocks_scanned >= BLOCKSCAN_BAILOUT_THRESHOLD) { + state->unscanned_size += CARRIER_SZ(carrier) - scanned_bytes; + break; + } else if (IS_LAST_BLK(block)) { + break; + } + + block = NXT_BLK(block); + blocks_scanned++; + } + } + + return blocks_scanned; +} + +static void gather_ahist_append_result(hist_tree_t *node, void *arg) +{ + gather_ahist_t *state = (gather_ahist_t*)arg; + + Eterm histogram_tuple, tag_tuple; + + Eterm *hp; + int ix; + + ASSERT(state->building_result); + + hp = erts_produce_heap(&state->msg_factory, 7 + state->hist_slot_count, 0); + + hp[0] = make_arityval(state->hist_slot_count); + + for (ix = 0; ix < state->hist_slot_count; ix++) { + hp[1 + ix] = make_small(node->histogram[ix]); + } + + histogram_tuple = make_tuple(hp); + hp += 1 + state->hist_slot_count; + + hp[0] = make_arityval(3); + hp[1] = ATAG_ID(node->tag); + hp[2] = alloc_type_atoms[ATAG_TYPE(node->tag)]; + hp[3] = histogram_tuple; + + tag_tuple = make_tuple(hp); + hp += 4; + + state->result_list = CONS(hp, tag_tuple, state->result_list); + + /* Plain free is intentional. */ + free(node); +} + +static void gather_ahist_send(gather_ahist_t *state) +{ + Eterm result_tuple, unscanned_size, task_ref; + + Uint term_size; + Eterm *hp; + + ASSERT((state->result_list == NIL) ^ (state->hist_count > 0)); + ASSERT(state->building_result); + + term_size = 4 + erts_iref_storage_heap_size(&state->iref); + term_size += IS_USMALL(0, state->unscanned_size) ? 0 : BIG_UINT_HEAP_SIZE; + + hp = erts_produce_heap(&state->msg_factory, term_size, 0); + + task_ref = erts_iref_storage_make_ref(&state->iref, &hp, + &(state->msg_factory.message)->hfrag.off_heap, 0); + + unscanned_size = bld_unstable_uint(&hp, NULL, state->unscanned_size); + + hp[0] = make_arityval(3); + hp[1] = task_ref; + hp[2] = unscanned_size; + hp[3] = state->result_list; + + result_tuple = make_tuple(hp); + + erts_factory_trim_and_close(&state->msg_factory, &result_tuple, 1); + + erts_queue_message(state->process, 0, state->msg_factory.message, + result_tuple, am_system); +} + +static int gather_ahist_finish(void *arg) +{ + gather_ahist_t *state = (gather_ahist_t*)arg; + + if (!state->building_result) { + ErtsMessage *message; + Uint minimum_size; + Eterm *hp; + + /* {Ref, unscanned size, [{Tag, {Histogram}} | Rest]} */ + minimum_size = 4 + erts_iref_storage_heap_size(&state->iref) + + state->hist_count * (7 + state->hist_slot_count); + + message = erts_alloc_message(minimum_size, &hp); + erts_factory_selfcontained_message_init(&state->msg_factory, + message, hp); + + ERTS_RBT_YIELD_STAT_INIT(&state->hist_tree_yield); + + state->result_list = NIL; + state->building_result = 1; + } + + if (hist_tree_rbt_foreach_destroy_yielding(&state->hist_tree, + &gather_ahist_append_result, + state, + &state->hist_tree_yield, + BLOCKSCAN_REDUCTIONS)) { + return 1; + } + + gather_ahist_send(state); + + return 0; +} + +static void gather_ahist_destroy_result(hist_tree_t *node, void *arg) +{ + (void)arg; + free(node); +} + +static void gather_ahist_abort(void *arg) +{ + gather_ahist_t *state = (gather_ahist_t*)arg; + + if (state->building_result) { + erts_factory_undo(&state->msg_factory); + } + + hist_tree_rbt_foreach_destroy(&state->hist_tree, + &gather_ahist_destroy_result, + NULL); +} + +int erts_alcu_gather_alloc_histograms(Process *p, int allocator_num, + int sched_id, int hist_width, + UWord hist_start, Eterm ref) +{ + gather_ahist_t *gather_state; + blockscan_t *scanner; + Allctr_t *allocator; + + ASSERT(is_internal_ref(ref)); + + if (!blockscan_get_specific_allocator(allocator_num, + sched_id, + &allocator)) { + return 0; + } else if (!allocator->atags) { + return 0; + } + + ensure_atoms_initialized(allocator); + + /* Plain calloc is intentional. */ + gather_state = (gather_ahist_t*)calloc(1, sizeof(gather_ahist_t)); + scanner = &gather_state->common; + + scanner->abort = gather_ahist_abort; + scanner->scan = gather_ahist_scan; + scanner->finish = gather_ahist_finish; + scanner->user_data = gather_state; + + erts_iref_storage_save(&gather_state->iref, ref); + gather_state->hist_slot_start = hist_start; + gather_state->hist_slot_count = hist_width; + gather_state->process = p; + + blockscan_dispatch(scanner, p, allocator, sched_id); + + return 1; +} + +/* ------------------------------------------------------------------------- */ + +typedef struct chist_node__ { + struct chist_node__ *next; + + UWord carrier_size; + UWord unscanned_size; + UWord allocated_size; + + /* BLOCKSCAN_BAILOUT_THRESHOLD guarantees we won't overflow this or the + * counters in the free block histogram. */ + int allocated_count; + int flags; + + int histogram[1]; +} chist_node_t; + +typedef struct { + blockscan_t common; + + ErtsIRefStorage iref; + Process *process; + + Eterm allocator_desc; + + chist_node_t *info_list; + UWord info_count; + + UWord hist_slot_start; + int hist_slot_count; + + ErtsHeapFactory msg_factory; + int building_result; + Eterm result_list; +} gather_cinfo_t; + +static int gather_cinfo_scan(Allctr_t *allocator, + void *user_data, + Carrier_t *carrier) +{ + gather_cinfo_t *state; + chist_node_t *node; + int blocks_scanned; + Block_t *block; + + state = (gather_cinfo_t*)user_data; + node = calloc(1, sizeof(chist_node_t) + + (state->hist_slot_count - 1) * + sizeof(node->histogram[0])); + blocks_scanned = 1; + + /* ERTS_CRR_ALCTR_FLG_BUSY is ignored since we've set it ourselves and it + * would be misleading to include it. */ + node->flags = erts_atomic_read_rb(&carrier->allctr) & + (ERTS_CRR_ALCTR_FLG_MASK & ~ERTS_CRR_ALCTR_FLG_BUSY); + node->carrier_size = CARRIER_SZ(carrier); + + if (IS_SB_CARRIER(carrier)) { + UWord block_size; + + block = SBC2BLK(allocator, carrier); + block_size = SBC_BLK_SZ(block); + + node->allocated_size = block_size; + node->allocated_count = 1; + } else { + UWord scanned_bytes = MBC_HEADER_SIZE(allocator); + + block = MBC_TO_FIRST_BLK(allocator, carrier); + + while (1) { + UWord block_size = MBC_BLK_SZ(block); + + scanned_bytes += block_size; + + if (IS_ALLOCED_BLK(block)) { + node->allocated_size += block_size; + node->allocated_count++; + } else { + UWord size_interval; + int hist_slot; + + size_interval = (block_size / state->hist_slot_start); + size_interval = u64_log2(size_interval + 1); + + hist_slot = MIN(size_interval, state->hist_slot_count - 1); + + node->histogram[hist_slot]++; + } + + if (blocks_scanned >= BLOCKSCAN_BAILOUT_THRESHOLD) { + node->unscanned_size += CARRIER_SZ(carrier) - scanned_bytes; + break; + } else if (IS_LAST_BLK(block)) { + break; + } + + block = NXT_BLK(block); + blocks_scanned++; + } + } + + node->next = state->info_list; + state->info_list = node; + state->info_count++; + + return blocks_scanned; +} + +static void gather_cinfo_append_result(gather_cinfo_t *state, + chist_node_t *info) +{ + Eterm carrier_size, unscanned_size, allocated_size; + Eterm histogram_tuple, carrier_tuple; + + Uint term_size; + Eterm *hp; + int ix; + + ASSERT(state->building_result); + + term_size = 11 + state->hist_slot_count; + term_size += IS_USMALL(0, info->carrier_size) ? 0 : BIG_UINT_HEAP_SIZE; + term_size += IS_USMALL(0, info->unscanned_size) ? 0 : BIG_UINT_HEAP_SIZE; + term_size += IS_USMALL(0, info->allocated_size) ? 0 : BIG_UINT_HEAP_SIZE; + + hp = erts_produce_heap(&state->msg_factory, term_size, 0); + + hp[0] = make_arityval(state->hist_slot_count); + + for (ix = 0; ix < state->hist_slot_count; ix++) { + hp[1 + ix] = make_small(info->histogram[ix]); + } + + histogram_tuple = make_tuple(hp); + hp += 1 + state->hist_slot_count; + + carrier_size = bld_unstable_uint(&hp, NULL, info->carrier_size); + unscanned_size = bld_unstable_uint(&hp, NULL, info->unscanned_size); + allocated_size = bld_unstable_uint(&hp, NULL, info->allocated_size); + + hp[0] = make_arityval(7); + hp[1] = state->allocator_desc; + hp[2] = carrier_size; + hp[3] = unscanned_size; + hp[4] = allocated_size; + hp[5] = make_small(info->allocated_count); + hp[6] = (info->flags & ERTS_CRR_ALCTR_FLG_IN_POOL) ? am_true : am_false; + hp[7] = histogram_tuple; + + carrier_tuple = make_tuple(hp); + hp += 8; + + state->result_list = CONS(hp, carrier_tuple, state->result_list); + + free(info); +} + +static void gather_cinfo_send(gather_cinfo_t *state) +{ + Eterm result_tuple, task_ref; + + int term_size; + Eterm *hp; + + ASSERT((state->result_list == NIL) ^ (state->info_count > 0)); + ASSERT(state->building_result); + + term_size = 3 + erts_iref_storage_heap_size(&state->iref); + hp = erts_produce_heap(&state->msg_factory, term_size, 0); + + task_ref = erts_iref_storage_make_ref(&state->iref, &hp, + &(state->msg_factory.message)->hfrag.off_heap, 0); + + hp[0] = make_arityval(2); + hp[1] = task_ref; + hp[2] = state->result_list; + + result_tuple = make_tuple(hp); + + erts_factory_trim_and_close(&state->msg_factory, &result_tuple, 1); + + erts_queue_message(state->process, 0, state->msg_factory.message, + result_tuple, am_system); +} + +static int gather_cinfo_finish(void *arg) +{ + gather_cinfo_t *state = (gather_cinfo_t*)arg; + int reductions = BLOCKSCAN_REDUCTIONS; + + if (!state->building_result) { + ErtsMessage *message; + Uint minimum_size; + Eterm *hp; + + /* {Ref, [{Carrier size, unscanned size, allocated size, + * allocated block count, {Free block histogram}} | Rest]} */ + minimum_size = 3 + erts_iref_storage_heap_size(&state->iref) + + state->info_count * (11 + state->hist_slot_count); + + message = erts_alloc_message(minimum_size, &hp); + erts_factory_selfcontained_message_init(&state->msg_factory, + message, hp); + + state->result_list = NIL; + state->building_result = 1; + } + + while (state->info_list) { + chist_node_t *current = state->info_list; + state->info_list = current->next; + + gather_cinfo_append_result(state, current); + + if (reductions-- <= 0) { + return 1; + } + } + + gather_cinfo_send(state); + + return 0; +} + +static void gather_cinfo_abort(void *arg) +{ + gather_cinfo_t *state = (gather_cinfo_t*)arg; + + if (state->building_result) { + erts_factory_undo(&state->msg_factory); + } + + while (state->info_list) { + chist_node_t *current = state->info_list; + state->info_list = current->next; + + free(current); + } +} + +int erts_alcu_gather_carrier_info(struct process *p, int allocator_num, + int sched_id, int hist_width, + UWord hist_start, Eterm ref) +{ + gather_cinfo_t *gather_state; + blockscan_t *scanner; + + const char *allocator_desc; + Allctr_t *allocator; + + ASSERT(is_internal_ref(ref)); + + if (!blockscan_get_specific_allocator(allocator_num, + sched_id, + &allocator)) { + return 0; + } + + allocator_desc = ERTS_ALC_A2AD(allocator_num); + + /* Plain calloc is intentional. */ + gather_state = (gather_cinfo_t*)calloc(1, sizeof(gather_cinfo_t)); + scanner = &gather_state->common; + + scanner->abort = gather_cinfo_abort; + scanner->scan = gather_cinfo_scan; + scanner->finish = gather_cinfo_finish; + scanner->user_data = gather_state; + + gather_state->allocator_desc = erts_atom_put((byte *)allocator_desc, + sys_strlen(allocator_desc), + ERTS_ATOM_ENC_LATIN1, 1); + erts_iref_storage_save(&gather_state->iref, ref); + gather_state->hist_slot_start = hist_start * 2; + gather_state->hist_slot_count = hist_width; + gather_state->process = p; + + blockscan_dispatch(scanner, p, allocator, sched_id); + + return 1; +} + /* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *\ * NOTE: erts_alcu_test() is only supposed to be used for testing. * @@ -6441,6 +7722,7 @@ erts_alcu_verify_unused_ts(Allctr_t *allctr) erts_mtx_unlock(&allctr->mutex); } + #ifdef DEBUG int is_sbc_blk(Block_t* blk) { |