diff options
Diffstat (limited to 'erts')
62 files changed, 2285 insertions, 1310 deletions
diff --git a/erts/configure.in b/erts/configure.in index f15bb56435..b2435f1ac9 100644 --- a/erts/configure.in +++ b/erts/configure.in @@ -1760,6 +1760,25 @@ if test $ac_cv_sizeof_void_p = 8; then fi AC_SUBST(BITS64) +AC_MSG_CHECKING([for C compiler 'restrict' support]) +restrict_keyword="" +for x in restrict __restrict; do + AC_TRY_COMPILE([int * $x foo(int * $x arg); + int * $x foo(int * $x arg) + { int * $x var=arg; return var;} + ],[], + [restrict_keyword=$x],[]) + if test "x$restrict_keyword" != "x"; then + break + fi +done +AC_DEFINE_UNQUOTED(ERTS_RESTRICT,[$restrict_keyword],[Type qualifier restrict]) +if test "x$restrict_keyword" != "x"; then + AC_MSG_RESULT(yes) +else + AC_MSG_RESULT(no) +fi + if test "x$ac_compiler_gnu" = "xyes"; then AC_MSG_CHECKING([if we should add -fno-tree-copyrename to CFLAGS for computed gotos to work properly]) AC_TRY_COMPILE([],[ @@ -1779,8 +1798,6 @@ else AC_MSG_RESULT(no) fi - - AC_MSG_CHECKING([for broken gcc-4.3.0 compiler]) AC_TRY_RUN([ /* pr36339.c */ @@ -2749,23 +2766,6 @@ if test "$cross_compiling" != "yes" && test X${enable_hipe} != Xno; then fi fi -dnl Check to disable -fPIE and friends for HiPE on amd64 -if test X${enable_hipe} = Xyes && test X$ARCH = Xamd64; then - AC_TRY_COMPILE(, [#if defined(__pie__) || defined(__PIE__) - #error -fPIE is enabled by default - #endif], - [AC_MSG_NOTICE([No -fPIE enabled by default])], - [AC_MSG_WARN([Security feature -fPIE will be disabled for HiPE]) - STATIC_CFLAGS="-fno-PIE $STATIC_CFLAGS" - saved_LDFLAGS=$LDFLAGS - LDFLAGS="-no-pie $LDFLAGS" - AC_TRY_LINK(,, [], - [LDFLAGS="-fno-PIE $saved_LDFLAGS" - AC_TRY_LINK(,, [], - [AC_MSG_WARN([Linked does not accept option -no-pie nor -fno-PIE]) - LDFLAGS=$saved_LDFLAGS])])]) -fi - if test X${enable_hipe} = Xyes; then case $OPSYS in linux) diff --git a/erts/doc/src/absform.xml b/erts/doc/src/absform.xml index e49c8c32e9..2ada903edb 100644 --- a/erts/doc/src/absform.xml +++ b/erts/doc/src/absform.xml @@ -4,7 +4,7 @@ <chapter> <header> <copyright> - <year>2001</year><year>2017</year> + <year>2001</year><year>2018</year> <holder>Ericsson AB. All Rights Reserved.</holder> </copyright> <legalnotice> @@ -614,26 +614,58 @@ <item> <p>If C is a catch clause <c>P -> B</c>, where <c>P</c> is a pattern and <c>B</c> is a body, then - Rep(C) = <c>{clause,LINE,[Rep({throw,P,_})],[],Rep(B)}</c>.</p> + Rep(C) = <c>{clause,LINE,[Rep({throw,P,_})],[],Rep(B)}</c>, + that is, a catch clause with an explicit exception class + <c>throw</c> and with or without an explicit stacktrace + variable <c>_</c> cannot be distinguished from a catch clause + without an explicit exception class and without an explicit + stacktrace variable.</p> </item> <item> <p>If C is a catch clause <c>X : P -> B</c>, where <c>X</c> is an atomic literal or a variable pattern, <c>P</c> is a pattern, and <c>B</c> is a body, then - Rep(C) = <c>{clause,LINE,[Rep({X,P,_})],[],Rep(B)}</c>.</p> + Rep(C) = <c>{clause,LINE,[Rep({X,P,_})],[],Rep(B)}</c>, + that is, a catch clause with an explicit exception class and + with an explicit stacktrace variable <c>_</c> cannot be + distinguished from a catch clause with an explicit exception + class and without an explicit stacktrace variable.</p> + </item> + <item> + <p>If C is a catch clause <c>X : P : S -> B</c>, + where <c>X</c> is an atomic literal or a variable pattern, + <c>P</c> is a pattern, <c>S</c> is a variable, and <c>B</c> + is a body, then + Rep(C) = <c>{clause,LINE,[Rep({X,P,S})],[],Rep(B)}</c>.</p> </item> <item> <p>If C is a catch clause <c>P when Gs -> B</c>, where <c>P</c> is a pattern, <c>Gs</c> is a guard sequence, and <c>B</c> is a body, then - Rep(C) = <c>{clause,LINE,[Rep({throw,P,_})],Rep(Gs),Rep(B)}</c>.</p> + Rep(C) = <c>{clause,LINE,[Rep({throw,P,_})],Rep(Gs),Rep(B)}</c>, + that is, a catch clause with an explicit exception class + <c>throw</c> and with or without an explicit stacktrace + variable <c>_</c> cannot be distinguished from a catch clause + without an explicit exception class and without an explicit + stacktrace variable.</p> </item> <item> <p>If C is a catch clause <c>X : P when Gs -> B</c>, where <c>X</c> is an atomic literal or a variable pattern, <c>P</c> is a pattern, <c>Gs</c> is a guard sequence, and <c>B</c> is a body, then - Rep(C) = <c>{clause,LINE,[Rep({X,P,_})],Rep(Gs),Rep(B)}</c>.</p> + Rep(C) = <c>{clause,LINE,[Rep({X,P,_})],Rep(Gs),Rep(B)}</c>, + that is, a catch clause with an explicit exception class and + with an explicit stacktrace variable <c>_</c> cannot be + distinguished from a catch clause with an explicit exception + class and without an explicit stacktrace variable.</p> + </item> + <item> + <p>If C is a catch clause <c>X : P : S when Gs -> B</c>, + where <c>X</c> is an atomic literal or a variable pattern, + <c>P</c> is a pattern, <c>Gs</c> is a guard sequence, + <c>S</c> is a variable, and <c>B</c> is a body, then + Rep(C) = <c>{clause,LINE,[Rep({X,P,S})],Rep(Gs),Rep(B)}</c>.</p> </item> <item> <p>If C is a function clause <c>( Ps ) -> B</c>, diff --git a/erts/doc/src/erl.xml b/erts/doc/src/erl.xml index bd824b3405..99f0421080 100644 --- a/erts/doc/src/erl.xml +++ b/erts/doc/src/erl.xml @@ -360,11 +360,12 @@ </item> <tag><c><![CDATA[-mode interactive | embedded]]></c></tag> <item> - <p>Indicates if the system is to load code dynamically - (<c><![CDATA[interactive]]></c>), or if all code is to be loaded - during system initialization (<c><![CDATA[embedded]]></c>); see - <seealso marker="kernel:code"><c>code(3)</c></seealso>. - Defaults to <c><![CDATA[interactive]]></c>.</p> + <p>Modules are auto loaded when they are first referenced if the + runtime system runs in <c><![CDATA[interactive]]></c> mode, which is + the default. In <c><![CDATA[embedded]]></c> mode modules are not auto + loaded. The latter is recommended when the boot script preloads all + modules, as conventionally happens in OTP releases. See + <seealso marker="kernel:code"><c>code(3)</c></seealso></p>. </item> <tag><c><![CDATA[-name Name]]></c></tag> <item> @@ -1693,4 +1694,3 @@ code:load_abs("..../user_default"). ]]></code> <seealso marker="tools:make"><c>make(3)</c></seealso></p> </section> </comref> - diff --git a/erts/doc/src/erl_nif.xml b/erts/doc/src/erl_nif.xml index 23f0c66429..f2f380ca4e 100644 --- a/erts/doc/src/erl_nif.xml +++ b/erts/doc/src/erl_nif.xml @@ -974,6 +974,10 @@ typedef struct { <seealso marker="#enif_make_binary"><c>enif_make_binary</c></seealso>. An allocated (and owned) <c>ErlNifBinary</c> can be kept between NIF calls.</p> + <p>If you do not need to reallocate or keep the data alive across NIF + calls, consider using <seealso marker="#enif_make_new_binary"> + <c>enif_make_new_binary</c></seealso> instead as it will allocate + small binaries on the process heap when possible.</p> <p>Returns <c>true</c> on success, or <c>false</c> if allocation fails.</p> </desc> @@ -3060,9 +3064,10 @@ if (retval & ERL_NIF_SELECT_STOP_CALLED) { </name> <fsummary>Get the pid of the calling process.</fsummary> <desc> - <p>Initializes the pid variable <c>*pid</c> to represent the - calling process.</p> - <p>Returns <c>pid</c>.</p> + <p>Initializes the <seealso marker="#ErlNifPid"><c>ErlNifPid</c></seealso> + variable at <c>*pid</c> to represent the calling process.</p> + <p>Returns <c>pid</c> if successful, or NULL if <c>caller_env</c> is not + a <seealso marker="#ErlNifEnv">process-bound environment</seealso>.</p> </desc> </func> diff --git a/erts/doc/src/erlang.xml b/erts/doc/src/erlang.xml index 24e0773e41..771897ba94 100644 --- a/erts/doc/src/erlang.xml +++ b/erts/doc/src/erlang.xml @@ -7116,6 +7116,23 @@ ok <func> <name name="system_flag" arity="2" clause_i="4"/> + <fsummary>Set system flag for erts_alloc.</fsummary> + <desc> + <p>Sets system flags for + <seealso marker="erts:erts_alloc"><c>erts_alloc(3)</c></seealso>. + <c><anno>Alloc</anno></c> is the allocator to affect, for example + <c>binary_alloc</c>. <c><anno>F</anno></c> is the flag to change and + <c><anno>V</anno></c> is the new value.</p> + <p>Only a subset of all <c>erts_alloc</c> flags can be changed + at run time. This subset is currently only the flag + <seealso marker="erts:erts_alloc#M_sbct"><c>sbct</c></seealso>.</p> + <p>Returns <c>ok</c> if the flag was set or <c>notsup</c> if not + supported by <c>erts_alloc</c>.</p> + </desc> + </func> + + <func> + <name name="system_flag" arity="2" clause_i="5"/> <fsummary>Set system flag fullsweep_after.</fsummary> <desc> <p>Sets system flag <c>fullsweep_after</c>. @@ -7134,7 +7151,7 @@ ok </func> <func> - <name name="system_flag" arity="2" clause_i="5" + <name name="system_flag" arity="2" clause_i="6" anchor="system_flag_microstate_accounting"/> <fsummary>Set system flag microstate_accounting.</fsummary> <desc> @@ -7148,7 +7165,7 @@ ok </func> <func> - <name name="system_flag" arity="2" clause_i="6"/> + <name name="system_flag" arity="2" clause_i="7"/> <fsummary>Set system flag min_heap_size.</fsummary> <desc> <p>Sets the default minimum heap size for processes. The size @@ -7163,7 +7180,7 @@ ok </func> <func> - <name name="system_flag" arity="2" clause_i="7"/> + <name name="system_flag" arity="2" clause_i="8"/> <fsummary>Set system flag min_bin_vheap_size.</fsummary> <desc> <p>Sets the default minimum binary virtual heap size for @@ -7180,7 +7197,7 @@ ok </func> <func> - <name name="system_flag" arity="2" clause_i="8" + <name name="system_flag" arity="2" clause_i="9" anchor="system_flag_max_heap_size"/> <fsummary>Set system flag max_heap_size.</fsummary> <type name="max_heap_size"/> @@ -7198,7 +7215,7 @@ ok </func> <func> - <name name="system_flag" arity="2" clause_i="9" + <name name="system_flag" arity="2" clause_i="10" anchor="system_flag_multi_scheduling"/> <fsummary>Set system flag multi_scheduling.</fsummary> <desc> @@ -7254,7 +7271,7 @@ ok </func> <func> - <name name="system_flag" arity="2" clause_i="10" + <name name="system_flag" arity="2" clause_i="11" anchor="system_flag_scheduler_bind_type"/> <fsummary>Set system flag scheduler_bind_type.</fsummary> <type name="scheduler_bind_type"/> @@ -7381,7 +7398,7 @@ ok </func> <func> - <name name="system_flag" arity="2" clause_i="11" + <name name="system_flag" arity="2" clause_i="12" anchor="system_flag_scheduler_wall_time"/> <fsummary>Set system flag scheduler_wall_time.</fsummary> <desc> @@ -7394,7 +7411,7 @@ ok </func> <func> - <name name="system_flag" arity="2" clause_i="12" + <name name="system_flag" arity="2" clause_i="13" anchor="system_flag_schedulers_online"/> <fsummary>Set system flag schedulers_online.</fsummary> <desc> @@ -7423,7 +7440,7 @@ ok </func> <func> - <name name="system_flag" arity="2" clause_i="13"/> + <name name="system_flag" arity="2" clause_i="14"/> <fsummary>Set system flag trace_control_word.</fsummary> <desc> <p>Sets the value of the node trace control word to @@ -7437,7 +7454,7 @@ ok </func> <func> - <name name="system_flag" arity="2" clause_i="14" + <name name="system_flag" arity="2" clause_i="15" anchor="system_flag_time_offset"/> <fsummary>Finalize the time offset.</fsummary> <desc> diff --git a/erts/doc/src/erts_alloc.xml b/erts/doc/src/erts_alloc.xml index 580780e73b..049ffe560a 100644 --- a/erts/doc/src/erts_alloc.xml +++ b/erts/doc/src/erts_alloc.xml @@ -225,6 +225,33 @@ used. The time complexity is proportional to log N, where N is the number of free blocks.</p> </item> + <tag>Age order first fit carrier address order first fit</tag> + <item> + <p>Strategy: Find the <em>oldest carrier</em> that + can satisfy the requested block size, then find a block within + that carrier using the "address order first fit" strategy.</p> + <p>Implementation: A balanced binary search tree is + used. The time complexity is proportional to log N, where + N is the number of free blocks.</p> + </item> + <tag>Age order first fit carrier best fit</tag> + <item> + <p>Strategy: Find the <em>oldest carrier</em> that + can satisfy the requested block size, then find a block within + that carrier using the "best fit" strategy.</p> + <p>Implementation: Balanced binary search trees are + used. The time complexity is proportional to log N, where + N is the number of free blocks.</p> + </item> + <tag>Age order first fit carrier address order best fit</tag> + <item> + <p>Strategy: Find the <em>oldest carrier</em> that + can satisfy the requested block size, then find a block within + that carrier using the "address order best fit" strategy.</p> + <p>Implementation: Balanced binary search trees are + used. The time complexity is proportional to log N, where + N is the number of free blocks.</p> + </item> <tag>Good fit</tag> <item> <p>Strategy: Try to find the best fit, but settle for the best fit @@ -467,7 +494,8 @@ fetched, it will function as an ordinary carrier. This feature has special requirements on the <seealso marker="#M_as">allocation strategy</seealso> used. Only - the strategies <c>aoff</c>, <c>aoffcbf</c>, and <c>aoffcaobf</c> + the strategies <c>aoff</c>, <c>aoffcbf</c>, <c>aoffcaobf</c>, + <c>ageffcaoff</c>m, <c>ageffcbf</c> and <c>ageffcaobf</c> support abandoned carriers.</p> <p>This feature also requires <seealso marker="#M_t">multiple thread specific instances</seealso> @@ -478,8 +506,30 @@ allocators based on the <c>alloc_util</c> framework, except <c>temp_alloc</c> (which would be pointless).</p> </item> + + <tag><marker id="M_acfml"/><c><![CDATA[+M<S>acfml <bytes>]]></c> + </tag> + <item> + <p>Abandon carrier free block min limit. A valid <c><![CDATA[<bytes>]]></c> + is a positive integer representing a block size limit. The largest + free block in a carrier must be at least <c>bytes</c> large, for the + carrier to be abandoned. The default is zero but can be changed + in the future.</p> + <p>See also <seealso marker="#M_acul"><c>acul</c></seealso>.</p> + </item> + + <tag><marker id="M_acnl"/><c><![CDATA[+M<S>acnl <amount>]]></c> + </tag> + <item> + <p>Abandon carrier number limit. A valid <c><![CDATA[<amount>]]></c> + is a positive integer representing max number of abandoned carriers per + allocator instance. Defaults to 1000 which will practically disable + the limit, but this can be changed in the future.</p> + <p>See also <seealso marker="#M_acul"><c>acul</c></seealso>.</p> + </item> + <tag><marker id="M_as"/> - <c><![CDATA[+M<S>as bf|aobf|aoff|aoffcbf|aoffcaobf|gf|af]]></c></tag> + <c><![CDATA[+M<S>as bf|aobf|aoff|aoffcbf|aoffcaobf|ageffcaoff|ageffcbf|ageffcaobf|gf|af]]></c></tag> <item> <p>Allocation strategy. The following strategies are valid:</p> <list type="bulleted"> @@ -490,6 +540,11 @@ </item> <item><c>aoffcaobf</c> (address order first fit carrier address order best fit)</item> + <item><c>ageffcaoff</c> (age order first fit carrier address order first fit)</item> + <item><c>ageffcbf</c> (age order first fit carrier best fit) + </item> + <item><c>ageffcaobf</c> (age order first fit carrier address + order best fit)</item> <item><c>gf</c> (good fit)</item> <item><c>af</c> (a fit)</item> </list> diff --git a/erts/doc/src/notes.xml b/erts/doc/src/notes.xml index e91e56e581..3109da6738 100644 --- a/erts/doc/src/notes.xml +++ b/erts/doc/src/notes.xml @@ -31,6 +31,48 @@ </header> <p>This document describes the changes made to the ERTS application.</p> +<section><title>Erts 9.2.1</title> + + <section><title>Improvements and New Features</title> + <list> + <item> + <p> + Improve search algorithm of abandoned memory carriers. + Instead of limited linear search, each allocator instance + maintain a balanced search tree of all its abandoned + carriers for faster and more exhaustive search.</p> + <p> + Own Id: OTP-14915 Aux Id: ERIERL-88 </p> + </item> + <item> + <p> + New erts_alloc command line options <c>+M_acnl</c> and + <c>+M_acfml</c> to limit carrier abandonment.</p> + <p> + Own Id: OTP-14916 Aux Id: ERIERL-88 </p> + </item> + <item> + <p> + New family of <c>erts_alloc</c> strategies: Age Order + First Fit. Similar to "address order", but instead the + oldest possible carrier is always chosen for allocation.</p> + <p> + Own Id: OTP-14917 Aux Id: ERIERL-88 </p> + </item> + <item> + <p> + Add possibility to change allocator options at runtime + with <c>system_info(erts_alloc, ...)</c>. Only option + <c>sbct</c> (single block carrier threshold) is currently + supported via this interface.</p> + <p> + Own Id: OTP-14918 Aux Id: ERIERL-88 </p> + </item> + </list> + </section> + +</section> + <section><title>Erts 9.2</title> <section><title>Fixed Bugs and Malfunctions</title> diff --git a/erts/emulator/beam/atom.names b/erts/emulator/beam/atom.names index cb4fab51f1..42a368cdd8 100644 --- a/erts/emulator/beam/atom.names +++ b/erts/emulator/beam/atom.names @@ -124,7 +124,6 @@ atom big atom bif_return_trap atom bif_timer_server atom binary -atom binary_bin_to_list_trap atom binary_copy_trap atom binary_find_trap atom binary_longest_prefix_trap @@ -577,7 +576,8 @@ atom running_procs atom runtime atom safe atom save_calls -atom scheduler +atom sbct +atom scheduler atom scheduler_id atom scheduler_wall_time atom scheduler_wall_time_all diff --git a/erts/emulator/beam/bif.c b/erts/emulator/beam/bif.c index f086c434ea..d68ccc3028 100644 --- a/erts/emulator/beam/bif.c +++ b/erts/emulator/beam/bif.c @@ -4805,6 +4805,8 @@ BIF_RETTYPE system_flag_2(BIF_ALIST_2) "scheduled for removal in Erlang/OTP 18. For more\n" "information see the erlang:system_flag/2 documentation.\n"); return erts_bind_schedulers(BIF_P, BIF_ARG_2); + } else if (ERTS_IS_ATOM_STR("erts_alloc", BIF_ARG_1)) { + return erts_alloc_set_dyn_param(BIF_P, BIF_ARG_2); } error: BIF_ERROR(BIF_P, BADARG); diff --git a/erts/emulator/beam/bif.tab b/erts/emulator/beam/bif.tab index c0d5e8ce74..0d1166f6ed 100644 --- a/erts/emulator/beam/bif.tab +++ b/erts/emulator/beam/bif.tab @@ -551,9 +551,6 @@ bif binary:last/1 bif binary:at/2 bif binary:part/2 binary_binary_part_2 bif binary:part/3 binary_binary_part_3 -bif binary:bin_to_list/1 -bif binary:bin_to_list/2 -bif binary:bin_to_list/3 bif binary:list_to_bin/1 bif binary:copy/1 bif binary:copy/2 @@ -696,3 +693,4 @@ bif erlang:iolist_to_iovec/1 bif erts_internal:new_connection/1 bif erts_internal:abort_connection/2 bif erts_internal:map_next/3 +bif ets:whereis/1 diff --git a/erts/emulator/beam/binary.c b/erts/emulator/beam/binary.c index d4db1a4188..95d324d2c1 100644 --- a/erts/emulator/beam/binary.c +++ b/erts/emulator/beam/binary.c @@ -953,7 +953,10 @@ HIPE_WRAPPER_BIF_DISABLE_GC(iolist_to_binary, 1) BIF_RETTYPE iolist_to_binary_1(BIF_ALIST_1) { if (is_binary(BIF_ARG_1)) { - BIF_RET(BIF_ARG_1); + if (binary_bitsize(BIF_ARG_1) == 0) { + BIF_RET(BIF_ARG_1); + } + BIF_ERROR(BIF_P, BADARG); } return erts_list_to_binary_bif(BIF_P, BIF_ARG_1, bif_export[BIF_iolist_to_binary_1]); } diff --git a/erts/emulator/beam/break.c b/erts/emulator/beam/break.c index e7acea0c5f..b9b70cd8ef 100644 --- a/erts/emulator/beam/break.c +++ b/erts/emulator/beam/break.c @@ -336,6 +336,12 @@ print_process_info(fmtfn_t to, void *to_arg, Process *p) erts_print(to, to_arg, "Heap unused: %bpu\n", (p->hend - p->htop)); erts_print(to, to_arg, "OldHeap unused: %bpu\n", (OLD_HEAP(p) == NULL) ? 0 : (OLD_HEND(p) - OLD_HTOP(p)) ); + erts_print(to, to_arg, "BinVHeap: %b64u\n", p->off_heap.overhead); + erts_print(to, to_arg, "OldBinVHeap: %b64u\n", BIN_OLD_VHEAP(p)); + erts_print(to, to_arg, "BinVHeap unused: %b64u\n", + BIN_VHEAP_SZ(p) - p->off_heap.overhead); + erts_print(to, to_arg, "OldBinVHeap unused: %b64u\n", + BIN_OLD_VHEAP_SZ(p) - BIN_OLD_VHEAP(p)); erts_print(to, to_arg, "Memory: %beu\n", erts_process_memory(p, !0)); if (garbing) { @@ -891,6 +897,21 @@ erl_crash_dump_v(char *file, int line, char* fmt, va_list args) erts_print_scheduler_info(to, to_arg, ERTS_SCHEDULER_IX(i)), erts_cbprintf(to, to_arg, "** crashed **\n")); } + for (i = 0; i < erts_no_dirty_cpu_schedulers; i++) { + ERTS_SYS_TRY_CATCH( + erts_print_scheduler_info(to, to_arg, ERTS_DIRTY_CPU_SCHEDULER_IX(i)), + erts_cbprintf(to, to_arg, "** crashed **\n")); + } + erts_cbprintf(to, to_arg, "=dirty_cpu_run_queue\n"); + erts_print_run_queue_info(to, to_arg, ERTS_DIRTY_CPU_RUNQ); + + for (i = 0; i < erts_no_dirty_io_schedulers; i++) { + ERTS_SYS_TRY_CATCH( + erts_print_scheduler_info(to, to_arg, ERTS_DIRTY_IO_SCHEDULER_IX(i)), + erts_cbprintf(to, to_arg, "** crashed **\n")); + } + erts_cbprintf(to, to_arg, "=dirty_io_run_queue\n"); + erts_print_run_queue_info(to, to_arg, ERTS_DIRTY_IO_RUNQ); #endif diff --git a/erts/emulator/beam/bs_instrs.tab b/erts/emulator/beam/bs_instrs.tab index b11903a47b..94e0000c8b 100644 --- a/erts/emulator/beam/bs_instrs.tab +++ b/erts/emulator/beam/bs_instrs.tab @@ -990,6 +990,9 @@ ctx_to_bin.execute() { Uint hole_size; Uint orig = mb->orig; ErlSubBin* sb = (ErlSubBin *) boxed_val(context); + /* Since we're going to overwrite the match state with the result, an + * ErlBinMatchState must be at least as large as an ErlSubBin. */ + ERTS_CT_ASSERT(sizeof(ErlSubBin) <= sizeof(ErlBinMatchState)); hole_size = 1 + header_arity(sb->thing_word) - ERL_SUB_BIN_SIZE; sb->thing_word = HEADER_SUB_BIN; sb->size = BYTE_OFFSET(size); diff --git a/erts/emulator/beam/copy.c b/erts/emulator/beam/copy.c index 10bf197405..7769a914db 100644 --- a/erts/emulator/beam/copy.c +++ b/erts/emulator/beam/copy.c @@ -611,7 +611,7 @@ Eterm copy_struct_x(Eterm obj, Uint sz, Eterm** hpp, ErlOffHeap* off_heap, Uint Eterm* htop; Eterm* hbot; Eterm* hp; - Eterm* objp; + Eterm* ERTS_RESTRICT objp; Eterm* tp; Eterm res; Eterm elem; @@ -1821,7 +1821,8 @@ all_clean: * * NOTE: Assumes that term is a tuple (ptr is an untagged tuple ptr). */ -Eterm copy_shallow(Eterm* ptr, Uint sz, Eterm** hpp, ErlOffHeap* off_heap) +Eterm copy_shallow(Eterm* ERTS_RESTRICT ptr, Uint sz, Eterm** hpp, + ErlOffHeap* off_heap) { Eterm* tp = ptr; Eterm* hp = *hpp; @@ -1985,7 +1986,7 @@ move_one_frag(Eterm** hpp, ErlHeapFragment* frag, ErlOffHeap* off_heap, int lite if (is_header(val)) { struct erl_off_heap_header* hdr = (struct erl_off_heap_header*)hp; ASSERT(ptr + header_arity(val) < end); - move_boxed(&ptr, val, &hp, &dummy_ref); + ptr = move_boxed(ptr, val, &hp, &dummy_ref); switch (val & _HEADER_SUBTAG_MASK) { case REF_SUBTAG: if (is_ordinary_ref_thing(hdr)) @@ -2002,7 +2003,7 @@ move_one_frag(Eterm** hpp, ErlHeapFragment* frag, ErlOffHeap* off_heap, int lite } else { /* must be a cons cell */ ASSERT(ptr+1 < end); - move_cons(&ptr, val, &hp, &dummy_ref); + move_cons(ptr, val, &hp, &dummy_ref); ptr += 2; } } diff --git a/erts/emulator/beam/erl_alloc.c b/erts/emulator/beam/erl_alloc.c index 88285d8be6..239dda6bcf 100644 --- a/erts/emulator/beam/erl_alloc.c +++ b/erts/emulator/beam/erl_alloc.c @@ -160,7 +160,7 @@ enum allctr_type { GOODFIT, BESTFIT, AFIT, - AOFIRSTFIT + FIRSTFIT }; struct au_init { @@ -500,8 +500,9 @@ set_default_test_alloc_opts(struct au_init *ip) SET_DEFAULT_ALLOC_OPTS(ip); ip->enable = 0; /* Disabled by default */ ip->thr_spec = -1 * erts_no_schedulers; - ip->atype = AOFIRSTFIT; - ip->init.aoff.flavor = AOFF_BF; + ip->atype = FIRSTFIT; + ip->init.aoff.crr_order = FF_AOFF; + ip->init.aoff.blk_order = FF_BF; ip->init.util.name_prefix = "test_"; ip->init.util.alloc_no = ERTS_ALC_A_TEST; ip->init.util.mmbcs = 0; /* Main carrier size */ @@ -599,10 +600,10 @@ static ERTS_INLINE int strategy_support_carrier_migration(struct au_init *auip) { /* - * Currently only aoff, aoffcbf and aoffcaobf support carrier + * Currently only aoff* and ageff* support carrier * migration, i.e, type AOFIRSTFIT. */ - return auip->atype == AOFIRSTFIT; + return auip->atype == FIRSTFIT; } static ERTS_INLINE void @@ -617,8 +618,9 @@ adjust_carrier_migration_support(struct au_init *auip) */ if (!strategy_support_carrier_migration(auip)) { /* Default to aoffcbf */ - auip->atype = AOFIRSTFIT; - auip->init.aoff.flavor = AOFF_BF; + auip->atype = FIRSTFIT; + auip->init.aoff.crr_order = FF_AOFF; + auip->init.aoff.blk_order = FF_BF; } } } @@ -1132,7 +1134,7 @@ start_au_allocator(ErtsAlcType_t alctr_n, &init->init.af, &init->init.util); break; - case AOFIRSTFIT: + case FIRSTFIT: as = erts_aoffalc_start((AOFFAllctr_t *) as0, &init->init.aoff, &init->init.util); @@ -1226,22 +1228,32 @@ get_bool_value(char *param_end, char** argv, int* ip) return -1; } +static Uint kb_to_bytes(Sint kb, Uint *bytes) +{ + const Uint max = ((~((Uint) 0))/1024) + 1; + + if (kb < 0 || (Uint)kb > max) + return 0; + if ((Uint)kb == max) + *bytes = ~((Uint) 0); + else + *bytes = ((Uint) kb)*1024; + return 1; +} + static Uint get_kb_value(char *param_end, char** argv, int* ip) { Sint tmp; - Uint max = ((~((Uint) 0))/1024) + 1; + Uint bytes = 0; char *rest; char *param = argv[*ip]+1; char *value = get_value(param_end, argv, ip); errno = 0; tmp = (Sint) ErtsStrToSint(value, &rest, 10); - if (errno != 0 || rest == value || tmp < 0 || max < ((Uint) tmp)) + if (errno != 0 || rest == value || !kb_to_bytes(tmp, &bytes)) bad_value(param, param_end, value); - if (max == (Uint) tmp) - return ~((Uint) 0); - else - return ((Uint) tmp)*1024; + return bytes; } static UWord @@ -1328,18 +1340,30 @@ handle_au_arg(struct au_init *auip, switch (sub_param[0]) { case 'a': - if (has_prefix("acul", sub_param)) { - if (!auip->carrier_migration_allowed) { - if (!u_switch) - goto bad_switch; - else { - /* ignore */ - (void) get_acul_value(auip, sub_param + 4, argv, ip); - break; - } - } - auip->init.util.acul = get_acul_value(auip, sub_param + 4, argv, ip); - } + if (sub_param[1] == 'c') { /* Migration parameters "ac*" */ + UWord value; + UWord* wp; + if (!auip->carrier_migration_allowed && !u_switch) + goto bad_switch; + + if (has_prefix("acul", sub_param)) { + value = get_acul_value(auip, sub_param + 4, argv, ip); + wp = &auip->init.util.acul; + } + else if (has_prefix("acnl", sub_param)) { + value = get_amount_value(sub_param + 4, argv, ip); + wp = &auip->init.util.acnl; + } + else if (has_prefix("acfml", sub_param)) { + value = get_amount_value(sub_param + 5, argv, ip); + wp = &auip->init.util.acfml; + } + else + goto bad_switch; + + if (auip->carrier_migration_allowed) + *wp = value; + } else if(has_prefix("asbcst", sub_param)) { auip->init.util.asbcst = get_kb_value(sub_param + 6, argv, ip); } @@ -1360,17 +1384,35 @@ handle_au_arg(struct au_init *auip, auip->atype = AFIT; } else if (strcmp("aoff", alg) == 0) { - auip->atype = AOFIRSTFIT; - auip->init.aoff.flavor = AOFF_AOFF; + auip->atype = FIRSTFIT; + auip->init.aoff.crr_order = FF_AOFF; + auip->init.aoff.blk_order = FF_AOFF; } else if (strcmp("aoffcbf", alg) == 0) { - auip->atype = AOFIRSTFIT; - auip->init.aoff.flavor = AOFF_BF; + auip->atype = FIRSTFIT; + auip->init.aoff.crr_order = FF_AOFF; + auip->init.aoff.blk_order = FF_BF; } else if (strcmp("aoffcaobf", alg) == 0) { - auip->atype = AOFIRSTFIT; - auip->init.aoff.flavor = AOFF_AOBF; + auip->atype = FIRSTFIT; + auip->init.aoff.crr_order = FF_AOFF; + auip->init.aoff.blk_order = FF_AOBF; } + else if (strcmp("ageffcaoff", alg) == 0) { + auip->atype = FIRSTFIT; + auip->init.aoff.crr_order = FF_AGEFF; + auip->init.aoff.blk_order = FF_AOFF; + } + else if (strcmp("ageffcbf", alg) == 0) { + auip->atype = FIRSTFIT; + auip->init.aoff.crr_order = FF_AGEFF; + auip->init.aoff.blk_order = FF_BF; + } + else if (strcmp("ageffcaobf", alg) == 0) { + auip->atype = FIRSTFIT; + auip->init.aoff.crr_order = FF_AGEFF; + auip->init.aoff.blk_order = FF_AOBF; + } else { bad_value(param, sub_param + 1, alg); } @@ -3392,6 +3434,65 @@ erts_request_alloc_info(struct process *c_p, return 1; } +Eterm erts_alloc_set_dyn_param(Process* c_p, Eterm tuple) +{ + ErtsAllocatorThrSpec_t *tspec; + ErtsAlcType_t ai; + Allctr_t* allctr; + Eterm* tp; + Eterm res; + + if (!is_tuple_arity(tuple, 3)) + goto badarg; + + tp = tuple_val(tuple); + + /* + * Ex: {ets_alloc, sbct, 256000} + */ + if (!is_atom(tp[1]) || !is_atom(tp[2]) || !is_integer(tp[3])) + goto badarg; + + for (ai = ERTS_ALC_A_MIN; ai <= ERTS_ALC_A_MAX; ai++) + if (erts_is_atom_str(erts_alc_a2ad[ai], tp[1], 0)) + break; + + if (ai > ERTS_ALC_A_MAX) + goto badarg; + + if (!erts_allctrs_info[ai].enabled || + !erts_allctrs_info[ai].alloc_util) { + return am_notsup; + } + + if (tp[2] == am_sbct) { + Uint sbct; + int i, ok; + + if (!term_to_Uint(tp[3], &sbct)) + goto badarg; + + tspec = &erts_allctr_thr_spec[ai]; + if (tspec->enabled) { + ok = 0; + for (i = 0; i < tspec->size; i++) { + allctr = tspec->allctr[i]; + ok |= allctr->try_set_dyn_param(allctr, am_sbct, sbct); + } + } + else { + allctr = erts_allctrs_info[ai].extra; + ok = allctr->try_set_dyn_param(allctr, am_sbct, sbct); + } + return ok ? am_ok : am_notsup; + } + return am_notsup; + +badarg: + ERTS_BIF_PREP_ERROR(res, c_p, EXC_BADARG); + return res; +} + /* * The allocator wrapper prelocking stuff below is about the locking order. * It only affects wrappers (erl_mtrace.c and erl_instrument.c) that keep locks @@ -3528,7 +3629,7 @@ UWord erts_alc_test(UWord op, UWord a1, UWord a2, UWord a3) &init.init.af, &init.init.util); break; - case AOFIRSTFIT: + case FIRSTFIT: allctr = erts_aoffalc_start((AOFFAllctr_t *) erts_alloc(ERTS_ALC_T_UNDEF, sizeof(AOFFAllctr_t)), @@ -3622,7 +3723,9 @@ UWord erts_alc_test(UWord op, UWord a1, UWord a2, UWord a3) case 0xf15: erts_free(ERTS_ALC_T_TEST, (void*)a1); return 0; - case 0xf16: { + case 0xf16: return (UWord) erts_realloc(ERTS_ALC_T_TEST, (void*)a1, (Uint)a2); + + case 0xf17: { Uint extra_hdr_sz = UNIT_CEILING((Uint)a1); ErtsAllocatorThrSpec_t* ts = &erts_allctr_thr_spec[ERTS_ALC_A_TEST]; Uint offset = ts->allctr[0]->mbc_header_size; @@ -3649,7 +3752,7 @@ UWord erts_alc_test(UWord op, UWord a1, UWord a2, UWord a3) *(void**)a3 = orig_destroying_mbc; return offset; } - case 0xf17: { + case 0xf18: { ErtsAllocatorThrSpec_t* ts = &erts_allctr_thr_spec[ERTS_ALC_A_TEST]; return ts->allctr[0]->largest_mbc_size; } diff --git a/erts/emulator/beam/erl_alloc.h b/erts/emulator/beam/erl_alloc.h index 117f96a4ad..174bf7a80a 100644 --- a/erts/emulator/beam/erl_alloc.h +++ b/erts/emulator/beam/erl_alloc.h @@ -126,8 +126,10 @@ typedef struct { void *extra; } ErtsAllocatorFunctions_t; -extern ErtsAllocatorFunctions_t erts_allctrs[ERTS_ALC_A_MAX+1]; -extern ErtsAllocatorInfo_t erts_allctrs_info[ERTS_ALC_A_MAX+1]; +extern ErtsAllocatorFunctions_t + ERTS_WRITE_UNLIKELY(erts_allctrs[ERTS_ALC_A_MAX+1]); +extern ErtsAllocatorInfo_t + ERTS_WRITE_UNLIKELY(erts_allctrs_info[ERTS_ALC_A_MAX+1]); typedef struct { int enabled; @@ -144,7 +146,7 @@ typedef struct ErtsAllocatorWrapper_t_ { void (*unlock)(void); struct ErtsAllocatorWrapper_t_* next; }ErtsAllocatorWrapper_t; -ErtsAllocatorWrapper_t *erts_allctr_wrappers; +extern ErtsAllocatorWrapper_t *erts_allctr_wrappers; extern int erts_allctr_wrapper_prelocked; extern erts_tsd_key_t erts_allctr_prelock_tsd_key; void erts_allctr_wrapper_prelock_init(ErtsAllocatorWrapper_t* wrapper); @@ -169,6 +171,8 @@ __decl_noreturn void erts_realloc_n_enomem(ErtsAlcType_t,void*,Uint) __decl_noreturn void erts_alc_fatal_error(int,int,ErtsAlcType_t,...) __noreturn; +Eterm erts_alloc_set_dyn_param(struct process*, Eterm); + #undef ERTS_HAVE_IS_IN_LITERAL_RANGE #if defined(ARCH_32) || defined(ERTS_HAVE_OS_PHYSICAL_MEMORY_RESERVATION) # define ERTS_HAVE_IS_IN_LITERAL_RANGE diff --git a/erts/emulator/beam/erl_alloc_util.c b/erts/emulator/beam/erl_alloc_util.c index 4d4bddb93f..fa97ead908 100644 --- a/erts/emulator/beam/erl_alloc_util.c +++ b/erts/emulator/beam/erl_alloc_util.c @@ -362,8 +362,10 @@ do { \ #define ERTS_CRR_ALCTR_FLG_IN_POOL (((erts_aint_t) 1) << 0) #define ERTS_CRR_ALCTR_FLG_BUSY (((erts_aint_t) 1) << 1) +#define ERTS_CRR_ALCTR_FLG_HOMECOMING (((erts_aint_t) 1) << 2) #define ERTS_CRR_ALCTR_FLG_MASK (ERTS_CRR_ALCTR_FLG_IN_POOL | \ - ERTS_CRR_ALCTR_FLG_BUSY) + ERTS_CRR_ALCTR_FLG_BUSY | \ + ERTS_CRR_ALCTR_FLG_HOMECOMING) #define SBC_HEADER_SIZE \ (UNIT_CEILING(offsetof(Carrier_t, cpool) \ @@ -563,7 +565,7 @@ do { \ DEBUG_CHECK_CARRIER_NO_SZ((AP)); \ } while (0) -#define STAT_MBC_CPOOL_INSERT(AP, CRR) \ +#define STAT_MBC_ABANDON(AP, CRR) \ do { \ UWord csz__ = CARRIER_SZ((CRR)); \ if (IS_MSEG_CARRIER((CRR))) \ @@ -1153,89 +1155,23 @@ unlink_carrier(CarrierList_t *cl, Carrier_t *crr) ASSERT(crr->next); crr->next->prev = crr->prev; } -} - - #ifdef DEBUG -static int is_in_list(ErtsDoubleLink_t* sentinel, ErtsDoubleLink_t* node) -{ - ErtsDoubleLink_t* p; - - ASSERT(node != sentinel); - for (p = sentinel->next; p != sentinel; p = p->next) { - if (p == node) - return 1; - } - return 0; -} -#endif /* DEBUG */ - -static ERTS_INLINE void -link_edl_after(ErtsDoubleLink_t* after_me, ErtsDoubleLink_t* node) -{ - ErtsDoubleLink_t* before_me = after_me->next; - ASSERT(node != after_me && node != before_me); - node->next = before_me; - node->prev = after_me; - before_me->prev = node; - after_me->next = node; -} - -static ERTS_INLINE void -link_edl_before(ErtsDoubleLink_t* before_me, ErtsDoubleLink_t* node) -{ - ErtsDoubleLink_t* after_me = before_me->prev; - ASSERT(node != before_me && node != after_me); - node->next = before_me; - node->prev = after_me; - before_me->prev = node; - after_me->next = node; -} - -static ERTS_INLINE void -unlink_edl(ErtsDoubleLink_t* node) -{ - node->next->prev = node->prev; - node->prev->next = node->next; -} - -static ERTS_INLINE void -relink_edl_before(ErtsDoubleLink_t* before_me, ErtsDoubleLink_t* node) -{ - if (node != before_me && node != before_me->prev) { - unlink_edl(node); - link_edl_before(before_me, node); - } + crr->next = crr; + crr->prev = crr; +#endif } static ERTS_INLINE int is_abandoned(Carrier_t *crr) { - return crr->cpool.abandoned.next != NULL; -} - -static ERTS_INLINE void -link_abandoned_carrier(ErtsDoubleLink_t* list, Carrier_t *crr) -{ - ASSERT(!is_abandoned(crr)); - - link_edl_after(list, &crr->cpool.abandoned); - - ASSERT(crr->cpool.abandoned.next != &crr->cpool.abandoned); - ASSERT(crr->cpool.abandoned.prev != &crr->cpool.abandoned); + return crr->cpool.state != ERTS_MBC_IS_HOME; } static ERTS_INLINE void unlink_abandoned_carrier(Carrier_t *crr) { - ASSERT(is_in_list(&crr->cpool.orig_allctr->cpool.pooled_list, - &crr->cpool.abandoned) || - is_in_list(&crr->cpool.orig_allctr->cpool.traitor_list, - &crr->cpool.abandoned)); - - unlink_edl(&crr->cpool.abandoned); - - crr->cpool.abandoned.next = NULL; - crr->cpool.abandoned.prev = NULL; + if (crr->cpool.state == ERTS_MBC_WAS_POOLED) { + aoff_remove_pooled_mbc(crr->cpool.orig_allctr, crr); + } } static ERTS_INLINE void @@ -1243,24 +1179,19 @@ clear_busy_pool_carrier(Allctr_t *allctr, Carrier_t *crr) { if (crr) { erts_aint_t max_size; - erts_aint_t new_val; + erts_aint_t iallctr; max_size = (erts_aint_t) allctr->largest_fblk_in_mbc(allctr, crr); erts_atomic_set_nob(&crr->cpool.max_size, max_size); - new_val = (((erts_aint_t) allctr)|ERTS_CRR_ALCTR_FLG_IN_POOL); + iallctr = erts_atomic_read_nob(&crr->allctr); + ERTS_ALC_CPOOL_ASSERT((iallctr & ~ERTS_CRR_ALCTR_FLG_HOMECOMING) + == ((erts_aint_t)allctr | + ERTS_CRR_ALCTR_FLG_IN_POOL | + ERTS_CRR_ALCTR_FLG_BUSY)); -#ifdef ERTS_ALC_CPOOL_DEBUG - { - erts_aint_t old_val = new_val|ERTS_CRR_ALCTR_FLG_BUSY; - - ERTS_ALC_CPOOL_ASSERT(old_val - == erts_atomic_xchg_relb(&crr->allctr, - new_val)); - } -#else - erts_atomic_set_relb(&crr->allctr, new_val); -#endif + iallctr &= ~ERTS_CRR_ALCTR_FLG_BUSY; + erts_atomic_set_relb(&crr->allctr, iallctr); } } @@ -1658,6 +1589,11 @@ dealloc_mbc(Allctr_t *allctr, Carrier_t *crr) } +static void set_new_allctr_abandon_limit(Allctr_t*); +static void abandon_carrier(Allctr_t*, Carrier_t*); +static void poolify_my_carrier(Allctr_t*, Carrier_t*); +static void enqueue_homecoming(Allctr_t*, Carrier_t*); + static ERTS_INLINE Allctr_t* get_pref_allctr(void *extra) { @@ -1724,9 +1660,23 @@ get_used_allctr(Allctr_t *pref_allctr, int pref_lock, void *p, UWord *sizep, erts_aint_t act; ERTS_ALC_CPOOL_ASSERT(!(iallctr & ERTS_CRR_ALCTR_FLG_BUSY)); - act = erts_atomic_cmpxchg_ddrb(&crr->allctr, - iallctr|ERTS_CRR_ALCTR_FLG_BUSY, - iallctr); + if (iallctr & ERTS_CRR_ALCTR_FLG_HOMECOMING) { + /* + * This carrier has just been given back to us by writing + * to crr->allctr with a write barrier (see abandon_carrier). + * + * We need a mathing read barrier to guarantee a correct view + * of the carrier for deallocation work. + */ + act = erts_atomic_cmpxchg_rb(&crr->allctr, + iallctr|ERTS_CRR_ALCTR_FLG_BUSY, + iallctr); + } + else { + act = erts_atomic_cmpxchg_ddrb(&crr->allctr, + iallctr|ERTS_CRR_ALCTR_FLG_BUSY, + iallctr); + } if (act == iallctr) { *busy_pcrr_pp = crr; break; @@ -1742,13 +1692,6 @@ get_used_allctr(Allctr_t *pref_allctr, int pref_lock, void *p, UWord *sizep, erts_mtx_unlock(&pref_allctr->mutex); } } - - ERTS_ALC_CPOOL_ASSERT( - (((iallctr & ~ERTS_CRR_ALCTR_FLG_MASK) == (erts_aint_t) pref_allctr) - ? (((iallctr & ERTS_CRR_ALCTR_FLG_MASK) == ERTS_CRR_ALCTR_FLG_IN_POOL) - || ((iallctr & ERTS_CRR_ALCTR_FLG_MASK) == 0)) - : 1)); - return used_allctr; } } @@ -2000,9 +1943,9 @@ handle_delayed_fix_dealloc(Allctr_t *allctr, void *ptr) /* Carrier migrated; need to redirect block to new owner... */ int cinit = used_allctr->dd.ix - allctr->dd.ix; - ERTS_ALC_CPOOL_ASSERT(!busy_pcrr_p); + ERTS_ALC_CPOOL_ASSERT(!busy_pcrr_p); - DEC_CC(allctr->calls.this_free); + DEC_CC(allctr->calls.this_free); ((ErtsAllctrFixDDBlock_t *) ptr)->fix_type = type; if (ddq_enqueue(&used_allctr->dd.q, ptr, cinit)) @@ -2011,8 +1954,9 @@ handle_delayed_fix_dealloc(Allctr_t *allctr, void *ptr) } } -static void -schedule_dealloc_carrier(Allctr_t *allctr, Carrier_t *crr); +static void schedule_dealloc_carrier(Allctr_t*, Carrier_t*); +static void dealloc_my_carrier(Allctr_t*, Carrier_t*); + static ERTS_INLINE int handle_delayed_dealloc(Allctr_t *allctr, @@ -2074,39 +2018,61 @@ handle_delayed_dealloc(Allctr_t *allctr, res = 1; blk = UMEM2BLK(ptr); - if (IS_FREE_LAST_MBC_BLK(blk)) { + if (blk->bhdr == HOMECOMING_MBC_BLK_HDR) { /* * A multiblock carrier that previously has been migrated away - * from us and now is back to be deallocated. For more info - * see schedule_dealloc_carrier(). - * - * Note that we cannot use FBLK_TO_MBC(blk) since it - * data has been overwritten by the queue. + * from us, was sent back to us either because + * - it became empty and we need to deallocated it, or + * - it was inserted into the pool and we need to update our pooled_tree */ - Carrier_t *crr = FIRST_BLK_TO_MBC(allctr, blk); - - /* Restore word overwritten by the dd-queue as it will be read - * if this carrier is pulled from dc_list by cpool_fetch() - */ - ERTS_ALC_CPOOL_ASSERT(FBLK_TO_MBC(blk) != crr); - ERTS_CT_ASSERT(sizeof(ErtsAllctrDDBlock_t) == sizeof(void*)); -#ifdef MBC_ABLK_OFFSET_BITS - blk->u.carrier = crr; -#else - blk->carrier = crr; -#endif + Carrier_t *crr = ErtsContainerStruct(blk, Carrier_t, + cpool.homecoming_dd.blk); + Block_t* first_blk = MBC_TO_FIRST_BLK(allctr, crr); + erts_aint_t iallctr; ERTS_ALC_CPOOL_ASSERT(ERTS_ALC_IS_CPOOL_ENABLED(allctr)); ERTS_ALC_CPOOL_ASSERT(allctr == crr->cpool.orig_allctr); - ERTS_ALC_CPOOL_ASSERT(((erts_aint_t) allctr) - != (erts_atomic_read_nob(&crr->allctr) - & ~ERTS_CRR_ALCTR_FLG_MASK)); - - erts_atomic_set_nob(&crr->allctr, ((erts_aint_t) allctr)); - schedule_dealloc_carrier(allctr, crr); + iallctr = erts_atomic_read_nob(&crr->allctr); + ASSERT(iallctr & ERTS_CRR_ALCTR_FLG_HOMECOMING); + while (1) { + if ((iallctr & (~ERTS_CRR_ALCTR_FLG_MASK | + ERTS_CRR_ALCTR_FLG_IN_POOL)) + == (erts_aint_t)allctr) { + /* + * Carrier is home (mine and not in pool) + */ + ASSERT(!(iallctr & ERTS_CRR_ALCTR_FLG_BUSY)); + erts_atomic_set_nob(&crr->allctr, (erts_aint_t)allctr); + if (IS_FREE_LAST_MBC_BLK(first_blk)) + dealloc_my_carrier(allctr, crr); + else + ASSERT(crr->cpool.state == ERTS_MBC_IS_HOME); + } + else { + erts_aint_t exp = iallctr; + erts_aint_t want = iallctr & ~ERTS_CRR_ALCTR_FLG_HOMECOMING; + + iallctr = erts_atomic_cmpxchg_nob(&crr->allctr, + want, + exp); + if (iallctr != exp) + continue; /* retry */ + + ASSERT(crr->cpool.state != ERTS_MBC_IS_HOME); + unlink_abandoned_carrier(crr); + if (iallctr & ERTS_CRR_ALCTR_FLG_IN_POOL) + poolify_my_carrier(allctr, crr); + else + crr->cpool.state = ERTS_MBC_WAS_TRAITOR; + } + break; + } } else { + ASSERT(IS_SBC_BLK(blk) || (ABLK_TO_MBC(blk) != + ErtsContainerStruct(blk, Carrier_t, + cpool.homecoming_dd.blk))); INC_CC(allctr->calls.this_free); @@ -2148,20 +2114,26 @@ enqueue_dealloc_other_instance(ErtsAlcType_t type, erts_alloc_notify_delayed_dealloc(allctr->ix); } - -static void -set_new_allctr_abandon_limit(Allctr_t *allctr); -static void -abandon_carrier(Allctr_t *allctr, Carrier_t *crr); - +static ERTS_INLINE void +update_pooled_tree(Allctr_t *allctr, Carrier_t *crr, Uint blk_sz) +{ + if (allctr == crr->cpool.orig_allctr && crr->cpool.state == ERTS_MBC_WAS_POOLED) { + /* + * Update pooled_tree with a potentially new (larger) max_sz + */ + AOFF_RBTree_t* crr_node = &crr->cpool.pooled; + if (blk_sz > crr_node->hdr.bhdr) { + crr_node->hdr.bhdr = blk_sz; + erts_aoff_larger_max_size(crr_node); + } + } +} static ERTS_INLINE void check_abandon_carrier(Allctr_t *allctr, Block_t *fblk, Carrier_t **busy_pcrr_pp) { Carrier_t *crr; - - if (busy_pcrr_pp && *busy_pcrr_pp) - return; + UWord ncrr_in_pool, largest_fblk; if (!ERTS_ALC_IS_CPOOL_ENABLED(allctr)) return; @@ -2170,8 +2142,7 @@ check_abandon_carrier(Allctr_t *allctr, Block_t *fblk, Carrier_t **busy_pcrr_pp) if (--allctr->cpool.check_limit_count <= 0) set_new_allctr_abandon_limit(allctr); - if (!erts_thr_progress_is_managed_thread()) - return; + ASSERT(erts_thr_progress_is_managed_thread()); if (allctr->cpool.disable_abandon) return; @@ -2179,6 +2150,9 @@ check_abandon_carrier(Allctr_t *allctr, Block_t *fblk, Carrier_t **busy_pcrr_pp) if (allctr->mbcs.blocks.curr.size > allctr->cpool.abandon_limit) return; + ncrr_in_pool = erts_atomic_read_nob(&allctr->cpool.stat.no_carriers); + if (ncrr_in_pool >= allctr->cpool.in_pool_limit) + return; crr = FBLK_TO_MBC(fblk); @@ -2189,9 +2163,14 @@ check_abandon_carrier(Allctr_t *allctr, Block_t *fblk, Carrier_t **busy_pcrr_pp) return; if (crr->cpool.thr_prgr != ERTS_THR_PRGR_INVALID - && !erts_thr_progress_has_reached(crr->cpool.thr_prgr)) - return; + && !erts_thr_progress_has_reached(crr->cpool.thr_prgr)) + return; + + largest_fblk = allctr->largest_fblk_in_mbc(allctr, crr); + if (largest_fblk < allctr->cpool.fblk_min_limit) + return; + erts_atomic_set_nob(&crr->cpool.max_size, largest_fblk); abandon_carrier(allctr, crr); } @@ -2237,6 +2216,7 @@ dealloc_block(Allctr_t *allctr, void *ptr, ErtsAlcFixList_t *fix, int dec_cc_on_ else { Carrier_t *busy_pcrr_p; Allctr_t *used_allctr; + used_allctr = get_used_allctr(allctr, ERTS_ALC_TS_PREF_LOCK_NO, ptr, NULL, &busy_pcrr_p); if (used_allctr == allctr) { @@ -2253,10 +2233,10 @@ dealloc_block(Allctr_t *allctr, void *ptr, ErtsAlcFixList_t *fix, int dec_cc_on_ /* Carrier migrated; need to redirect block to new owner... */ int cinit = used_allctr->dd.ix - allctr->dd.ix; - ERTS_ALC_CPOOL_ASSERT(!busy_pcrr_p); + ERTS_ALC_CPOOL_ASSERT(!busy_pcrr_p); - if (dec_cc_on_redirect) - DEC_CC(allctr->calls.this_free); + if (dec_cc_on_redirect) + DEC_CC(allctr->calls.this_free); if (ddq_enqueue(&used_allctr->dd.q, ptr, cinit)) erts_alloc_notify_delayed_dealloc(used_allctr->ix); } @@ -2500,15 +2480,16 @@ mbc_free(Allctr_t *allctr, void *p, Carrier_t **busy_pcrr_pp) ASSERT(blk_sz % sizeof(Unit_t) == 0); ASSERT(IS_MBC_BLK(blk)); - if (is_first_blk - && is_last_blk - && allctr->main_carrier != FIRST_BLK_TO_MBC(allctr, blk)) { - destroy_carrier(allctr, blk, busy_pcrr_pp); + if (is_first_blk && is_last_blk && crr != allctr->main_carrier) { + destroy_carrier(allctr, blk, busy_pcrr_pp); } else { (*allctr->link_free_block)(allctr, blk); HARD_CHECK_BLK_CARRIER(allctr, blk); - check_abandon_carrier(allctr, blk, busy_pcrr_pp); + if (busy_pcrr_pp && *busy_pcrr_pp) + update_pooled_tree(allctr, crr, blk_sz); + else + check_abandon_carrier(allctr, blk, busy_pcrr_pp); } } @@ -2542,8 +2523,19 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, Uint32 alcu_flgs, return NULL; #else /* !MBC_REALLOC_ALWAYS_MOVES */ - if (busy_pcrr_pp && *busy_pcrr_pp) - goto realloc_move; /* Don't want to use carrier in pool */ + if (busy_pcrr_pp && *busy_pcrr_pp) { + /* + * Don't want to use carrier in pool + */ + new_p = mbc_alloc(allctr, size); + if (!new_p) + return NULL; + new_blk = UMEM2BLK(new_p); + ASSERT(!(IS_MBC_BLK(new_blk) && ABLK_TO_MBC(new_blk) == *busy_pcrr_pp)); + sys_memcpy(new_p, p, MIN(size, old_blk_sz - ABLK_HDR_SZ)); + mbc_free(allctr, p, busy_pcrr_pp); + return new_p; + } get_blk_sz = blk_sz = UMEMSZ2BLKSZ(allctr, size); @@ -2776,7 +2768,7 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, Uint32 alcu_flgs, if (cand_blk_sz < get_blk_sz) { /* We wont fit in cand_blk get a new one */ - realloc_move: + #endif /* !MBC_REALLOC_ALWAYS_MOVES */ new_p = mbc_alloc(allctr, size); @@ -2880,8 +2872,7 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, Uint32 alcu_flgs, #define ERTS_ALC_MAX_DEALLOC_CARRIER 10 -#define ERTS_ALC_CPOOL_MAX_FETCH_INSPECT 20 -#define ERTS_ALC_CPOOL_MAX_TRAITOR_INSPECT 10 +#define ERTS_ALC_CPOOL_MAX_FETCH_INSPECT 100 #define ERTS_ALC_CPOOL_CHECK_LIMIT_COUNT 100 #define ERTS_ALC_CPOOL_MAX_FAILED_STAT_READS 3 @@ -3045,19 +3036,18 @@ cpool_insert(Allctr_t *allctr, Carrier_t *crr) ErtsAlcCPoolData_t *cpd1p, *cpd2p; erts_aint_t val; ErtsAlcCPoolData_t *sentinel = &carrier_pool[allctr->alloc_no].sentinel; + Allctr_t *orig_allctr = crr->cpool.orig_allctr; ERTS_ALC_CPOOL_ASSERT(allctr->alloc_no == ERTS_ALC_A_INVALID /* testcase */ || erts_thr_progress_is_managed_thread()); - ERTS_ALC_CPOOL_ASSERT(erts_atomic_read_nob(&crr->allctr) - == (erts_aint_t) allctr); - erts_atomic_add_nob(&allctr->cpool.stat.blocks_size, + erts_atomic_add_nob(&orig_allctr->cpool.stat.blocks_size, (erts_aint_t) crr->cpool.blocks_size); - erts_atomic_add_nob(&allctr->cpool.stat.no_blocks, + erts_atomic_add_nob(&orig_allctr->cpool.stat.no_blocks, (erts_aint_t) crr->cpool.blocks); - erts_atomic_add_nob(&allctr->cpool.stat.carriers_size, + erts_atomic_add_nob(&orig_allctr->cpool.stat.carriers_size, (erts_aint_t) CARRIER_SZ(crr)); - erts_atomic_inc_nob(&allctr->cpool.stat.no_carriers); + erts_atomic_inc_nob(&orig_allctr->cpool.stat.no_carriers); /* * We search in 'next' direction and begin by passing @@ -3118,8 +3108,6 @@ cpool_insert(Allctr_t *allctr, Carrier_t *crr) (erts_aint_t) &crr->cpool, (erts_aint_t) cpd1p); - erts_atomic_set_wb(&crr->allctr, - ((erts_aint_t) allctr)|ERTS_CRR_ALCTR_FLG_IN_POOL); LTTNG3(carrier_pool_put, ERTS_ALC_A2AD(allctr->alloc_no), allctr->ix, CARRIER_SZ(crr)); } @@ -3221,130 +3209,126 @@ cpool_delete(Allctr_t *allctr, Allctr_t *prev_allctr, Carrier_t *crr) static Carrier_t * cpool_fetch(Allctr_t *allctr, UWord size) { - int i, i_stop, has_passed_sentinel; + enum { IGNORANT, HAS_SEEN_SENTINEL, THE_LAST_ONE } loop_state; + int i; Carrier_t *crr; + Carrier_t *reinsert_crr = NULL; ErtsAlcCPoolData_t *cpdp; - ErtsAlcCPoolData_t *cpool_entrance; + ErtsAlcCPoolData_t *cpool_entrance = NULL; ErtsAlcCPoolData_t *sentinel; - ErtsDoubleLink_t* dl; - ErtsDoubleLink_t* first_old_traitor; ERTS_ALC_CPOOL_ASSERT(allctr->alloc_no == ERTS_ALC_A_INVALID /* testcase */ || erts_thr_progress_is_managed_thread()); i = ERTS_ALC_CPOOL_MAX_FETCH_INSPECT; - first_old_traitor = allctr->cpool.traitor_list.next; - cpool_entrance = NULL; LTTNG3(carrier_pool_get, ERTS_ALC_A2AD(allctr->alloc_no), allctr->ix, (unsigned long)size); /* - * Search my own pooled_list, + * Search my own pooled_tree, * i.e my abandoned carriers that were in the pool last time I checked. */ + do { + erts_aint_t exp, act; + + crr = aoff_lookup_pooled_mbc(allctr, size); + if (!crr) + break; + + ASSERT(crr->cpool.state == ERTS_MBC_WAS_POOLED); + ASSERT(crr->cpool.orig_allctr == allctr); + + aoff_remove_pooled_mbc(allctr, crr); + + exp = erts_atomic_read_nob(&crr->allctr); + if (exp & ERTS_CRR_ALCTR_FLG_IN_POOL) { + ASSERT((exp & ~ERTS_CRR_ALCTR_FLG_MASK) == (erts_aint_t)allctr); + if (erts_atomic_read_nob(&crr->cpool.max_size) < size) { + /* + * This carrier has been fetched and inserted back again + * by a foreign allocator. That's why it has a stale search size. + */ + ASSERT(exp & ERTS_CRR_ALCTR_FLG_HOMECOMING); + crr->cpool.pooled.hdr.bhdr = erts_atomic_read_nob(&crr->cpool.max_size); + aoff_add_pooled_mbc(allctr, crr); + INC_CC(allctr->cpool.stat.skip_size); + continue; + } + else if (exp & ERTS_CRR_ALCTR_FLG_BUSY) { + /* + * This must be our own carrier as part of a realloc call. + * Skip it to make things simpler. + * Must wait to re-insert to not be found again by lookup. + */ + ASSERT(!reinsert_crr); + reinsert_crr = crr; + INC_CC(allctr->cpool.stat.skip_busy); + continue; + } - dl = allctr->cpool.pooled_list.next; - while(dl != &allctr->cpool.pooled_list) { - erts_aint_t exp, act; - crr = (Carrier_t *) (((char *) dl) - offsetof(Carrier_t, cpool.abandoned)); - - ASSERT(!is_in_list(&allctr->cpool.traitor_list, dl)); - ASSERT(crr->cpool.orig_allctr == allctr); - dl = dl->next; - exp = erts_atomic_read_rb(&crr->allctr); - if ((exp & ERTS_CRR_ALCTR_FLG_MASK) == ERTS_CRR_ALCTR_FLG_IN_POOL - && erts_atomic_read_nob(&crr->cpool.max_size) >= size) { - /* Try to fetch it... */ - act = erts_atomic_cmpxchg_mb(&crr->allctr, - (erts_aint_t) allctr, - exp); - if (act == exp) { - cpool_delete(allctr, ((Allctr_t *) (act & ~ERTS_CRR_ALCTR_FLG_MASK)), crr); - unlink_abandoned_carrier(crr); - - /* Move sentinel to continue next search from here */ - relink_edl_before(dl, &allctr->cpool.pooled_list); - return crr; - } - exp = act; - } - if (exp & ERTS_CRR_ALCTR_FLG_IN_POOL) { - if (!cpool_entrance) - cpool_entrance = &crr->cpool; - } - else { /* Not in pool, move to traitor_list */ - unlink_abandoned_carrier(crr); - link_abandoned_carrier(&allctr->cpool.traitor_list, crr); - } - if (--i <= 0) { - /* Move sentinel to continue next search from here */ - relink_edl_before(dl, &allctr->cpool.pooled_list); - return NULL; - } - } + /* Try to fetch it... */ + act = erts_atomic_cmpxchg_mb(&crr->allctr, + exp & ~ERTS_CRR_ALCTR_FLG_IN_POOL, + exp); + if (act == exp) { + cpool_delete(allctr, allctr, crr); + crr->cpool.state = ERTS_MBC_IS_HOME; + + if (reinsert_crr) + aoff_add_pooled_mbc(allctr, reinsert_crr); + return crr; + } + exp = act; + INC_CC(allctr->cpool.stat.skip_race); + } + else + INC_CC(allctr->cpool.stat.skip_not_pooled); - /* Now search traitor_list. - * i.e carriers employed by other allocators last time I checked. - * They might have been abandoned since then. - */ + /* Not in pool anymore */ + ASSERT(!(exp & ERTS_CRR_ALCTR_FLG_BUSY)); + crr->cpool.state = ERTS_MBC_WAS_TRAITOR; - i_stop = (i < ERTS_ALC_CPOOL_MAX_TRAITOR_INSPECT ? - 0 : i - ERTS_ALC_CPOOL_MAX_TRAITOR_INSPECT); - dl = first_old_traitor; - while(dl != &allctr->cpool.traitor_list) { - erts_aint_t exp, act; - crr = (Carrier_t *) (((char *) dl) - offsetof(Carrier_t, cpool.abandoned)); - ASSERT(dl != &allctr->cpool.pooled_list); - ASSERT(crr->cpool.orig_allctr == allctr); - dl = dl->next; - exp = erts_atomic_read_rb(&crr->allctr); - if (exp & ERTS_CRR_ALCTR_FLG_IN_POOL) { - if (!(exp & ERTS_CRR_ALCTR_FLG_BUSY) - && erts_atomic_read_nob(&crr->cpool.max_size) >= size) { - /* Try to fetch it... */ - act = erts_atomic_cmpxchg_mb(&crr->allctr, - (erts_aint_t) allctr, - exp); - if (act == exp) { - cpool_delete(allctr, ((Allctr_t *) (act & ~ERTS_CRR_ALCTR_FLG_MASK)), crr); - unlink_abandoned_carrier(crr); + }while (--i > 0); - /* Move sentinel to continue next search from here */ - relink_edl_before(dl, &allctr->cpool.traitor_list); - return crr; - } - exp = act; - } - if (exp & ERTS_CRR_ALCTR_FLG_IN_POOL) { - if (!cpool_entrance) - cpool_entrance = &crr->cpool; + if (reinsert_crr) + aoff_add_pooled_mbc(allctr, reinsert_crr); - /* Move to pooled_list */ - unlink_abandoned_carrier(crr); - link_abandoned_carrier(&allctr->cpool.pooled_list, crr); - } - } - if (--i <= i_stop) { - /* Move sentinel to continue next search from here */ - relink_edl_before(dl, &allctr->cpool.traitor_list); - if (i > 0) - break; - else - return NULL; - } + /* + * Try find a nice cpool_entrance + */ + while (allctr->cpool.pooled_tree) { + erts_aint_t iallctr; + + crr = ErtsContainerStruct(allctr->cpool.pooled_tree, Carrier_t, cpool.pooled); + iallctr = erts_atomic_read_nob(&crr->allctr); + if (iallctr & ERTS_CRR_ALCTR_FLG_IN_POOL) { + cpool_entrance = &crr->cpool; + break; + } + /* Not in pool anymore */ + ASSERT(!(iallctr & ERTS_CRR_ALCTR_FLG_BUSY)); + aoff_remove_pooled_mbc(allctr, crr); + crr->cpool.state = ERTS_MBC_WAS_TRAITOR; + + if (--i <= 0) { + INC_CC(allctr->cpool.stat.fail_pooled); + return NULL; + } } + /* * Finally search the shared pool and try employ foreign carriers */ - sentinel = &carrier_pool[allctr->alloc_no].sentinel; if (cpool_entrance) { - /* We saw a pooled carried above, use it as entrance into the pool + /* + * We saw a pooled carried above, use it as entrance into the pool */ cpdp = cpool_entrance; } else { - /* No pooled carried seen above. Start search at cpool sentinel, + /* + * No pooled carried seen above. Start search at cpool sentinel, * but begin by passing one element before trying to fetch. * This in order to avoid contention with threads inserting elements. */ @@ -3354,8 +3338,8 @@ cpool_fetch(Allctr_t *allctr, UWord size) goto check_dc_list; } - has_passed_sentinel = 0; - while (1) { + loop_state = IGNORANT; + do { erts_aint_t exp; cpdp = cpool_aint2cpd(cpool_read(&cpdp->prev)); if (cpdp == cpool_entrance) { @@ -3364,38 +3348,52 @@ cpool_fetch(Allctr_t *allctr, UWord size) if (cpdp == sentinel) break; } - i = 0; /* Last one to inspect */ + loop_state = THE_LAST_ONE; } else if (cpdp == sentinel) { - if (has_passed_sentinel) { + if (loop_state == HAS_SEEN_SENTINEL) { /* We been here before. cpool_entrance must have been removed */ + INC_CC(allctr->cpool.stat.entrance_removed); break; } cpdp = cpool_aint2cpd(cpool_read(&cpdp->prev)); if (cpdp == sentinel) break; - has_passed_sentinel = 1; + loop_state = HAS_SEEN_SENTINEL; } - crr = (Carrier_t *)(((char *)cpdp) - offsetof(Carrier_t, cpool)); + crr = ErtsContainerStruct(cpdp, Carrier_t, cpool); exp = erts_atomic_read_rb(&crr->allctr); - if (((exp & (ERTS_CRR_ALCTR_FLG_MASK)) == ERTS_CRR_ALCTR_FLG_IN_POOL) - && (erts_atomic_read_nob(&cpdp->max_size) >= size)) { + + if (erts_atomic_read_nob(&cpdp->max_size) < size) { + INC_CC(allctr->cpool.stat.skip_size); + } + else if ((exp & (ERTS_CRR_ALCTR_FLG_IN_POOL | ERTS_CRR_ALCTR_FLG_BUSY)) + == ERTS_CRR_ALCTR_FLG_IN_POOL) { erts_aint_t act; - /* Try to fetch it... */ - act = erts_atomic_cmpxchg_mb(&crr->allctr, - (erts_aint_t) allctr, - exp); + erts_aint_t want = (((erts_aint_t) allctr) + | (exp & ERTS_CRR_ALCTR_FLG_HOMECOMING)); + /* Try to fetch it... */ + act = erts_atomic_cmpxchg_mb(&crr->allctr, want, exp); if (act == exp) { cpool_delete(allctr, ((Allctr_t *) (act & ~ERTS_CRR_ALCTR_FLG_MASK)), crr); if (crr->cpool.orig_allctr == allctr) { unlink_abandoned_carrier(crr); - } + crr->cpool.state = ERTS_MBC_IS_HOME; + } return crr; } } - if (--i <= 0) + + if (exp & ERTS_CRR_ALCTR_FLG_BUSY) + INC_CC(allctr->cpool.stat.skip_busy); + else + INC_CC(allctr->cpool.stat.skip_race); + + if (--i <= 0) { + INC_CC(allctr->cpool.stat.fail_shared); return NULL; - } + } + }while (loop_state != THE_LAST_ONE); check_dc_list: /* Last; check our own pending dealloc carrier list... */ @@ -3404,23 +3402,23 @@ check_dc_list: if (erts_atomic_read_nob(&crr->cpool.max_size) >= size) { Block_t* blk; unlink_carrier(&allctr->cpool.dc_list, crr); -#ifdef ERTS_ALC_CPOOL_DEBUG - ERTS_ALC_CPOOL_ASSERT(erts_atomic_xchg_nob(&crr->allctr, - ((erts_aint_t) allctr)) - == (((erts_aint_t) allctr) & ~ERTS_CRR_ALCTR_FLG_MASK)); -#else - erts_atomic_set_nob(&crr->allctr, ((erts_aint_t) allctr)); -#endif + ERTS_ALC_CPOOL_ASSERT(erts_atomic_read_nob(&crr->allctr) + == ((erts_aint_t) allctr)); blk = MBC_TO_FIRST_BLK(allctr, crr); ASSERT(FBLK_TO_MBC(blk) == crr); allctr->link_free_block(allctr, blk); return crr; } crr = crr->prev; - if (--i <= 0) + if (--i <= 0) { + INC_CC(allctr->cpool.stat.fail_pend_dealloc); return NULL; + } } + if (i != ERTS_ALC_CPOOL_MAX_FETCH_INSPECT) + INC_CC(allctr->cpool.stat.fail); + return NULL; } @@ -3475,9 +3473,6 @@ static void schedule_dealloc_carrier(Allctr_t *allctr, Carrier_t *crr) { Allctr_t *orig_allctr; - Block_t *blk; - int check_pending_dealloc; - erts_aint_t max_size; ASSERT(IS_MB_CARRIER(crr)); @@ -3488,9 +3483,17 @@ schedule_dealloc_carrier(Allctr_t *allctr, Carrier_t *crr) orig_allctr = crr->cpool.orig_allctr; - if (allctr != orig_allctr) { - int cinit = orig_allctr->dd.ix - allctr->dd.ix; - + if (allctr == orig_allctr) { + if (!(erts_atomic_read_nob(&crr->allctr) & ERTS_CRR_ALCTR_FLG_HOMECOMING)) { + dealloc_my_carrier(allctr, crr); + } + /*else + * Carrier was abandoned earlier by other thread and + * is still waiting for us in dd-queue. + * handle_delayed_dealloc() will handle it when crr is dequeued. + */ + } + else { /* * We send the carrier to its origin for deallocation. * This in order: @@ -3499,29 +3502,39 @@ schedule_dealloc_carrier(Allctr_t *allctr, Carrier_t *crr) * - to ensure that we always only reuse empty carriers * originating from our own thread specific mseg_alloc * instance which is beneficial on NUMA systems. - * - * The receiver will recognize that this is a carrier to - * deallocate (and not a block which is the common case) - * since the block is an mbc block that is free and last - * in the carrier. */ - blk = MBC_TO_FIRST_BLK(allctr, crr); - ERTS_ALC_CPOOL_ASSERT(IS_FREE_LAST_MBC_BLK(blk)); - - ERTS_ALC_CPOOL_ASSERT(IS_MBC_FIRST_ABLK(allctr, blk)); - ERTS_ALC_CPOOL_ASSERT(crr == FBLK_TO_MBC(blk)); - ERTS_ALC_CPOOL_ASSERT(crr == FIRST_BLK_TO_MBC(allctr, blk)); - ERTS_ALC_CPOOL_ASSERT(((erts_aint_t) allctr) - == (erts_atomic_read_nob(&crr->allctr) - & ~ERTS_CRR_ALCTR_FLG_MASK)); + erts_aint_t iallctr; +#ifdef ERTS_ALC_CPOOL_DEBUG + Block_t* first_blk = MBC_TO_FIRST_BLK(allctr, crr); + ERTS_ALC_CPOOL_ASSERT(IS_FREE_LAST_MBC_BLK(first_blk)); + + ERTS_ALC_CPOOL_ASSERT(IS_MBC_FIRST_ABLK(allctr, first_blk)); + ERTS_ALC_CPOOL_ASSERT(crr == FBLK_TO_MBC(first_blk)); + ERTS_ALC_CPOOL_ASSERT(crr == FIRST_BLK_TO_MBC(allctr, first_blk)); + ERTS_ALC_CPOOL_ASSERT((erts_atomic_read_nob(&crr->allctr) + & ~ERTS_CRR_ALCTR_FLG_HOMECOMING) + == (erts_aint_t) allctr); +#endif - if (ddq_enqueue(&orig_allctr->dd.q, BLK2UMEM(blk), cinit)) - erts_alloc_notify_delayed_dealloc(orig_allctr->ix); - return; + iallctr = (erts_aint_t)orig_allctr | ERTS_CRR_ALCTR_FLG_HOMECOMING; + if (!(erts_atomic_xchg_nob(&crr->allctr, iallctr) + & ERTS_CRR_ALCTR_FLG_HOMECOMING)) { + enqueue_homecoming(allctr, crr); + } } +} + +static void dealloc_my_carrier(Allctr_t *allctr, Carrier_t *crr) +{ + Block_t *blk; + int check_pending_dealloc; + erts_aint_t max_size; - if (is_abandoned(crr)) - unlink_abandoned_carrier(crr); + ERTS_ALC_CPOOL_ASSERT(allctr == crr->cpool.orig_allctr); + if (is_abandoned(crr)) { + unlink_abandoned_carrier(crr); + crr->cpool.state = ERTS_MBC_IS_HOME; + } if (crr->cpool.thr_prgr == ERTS_THR_PRGR_INVALID || erts_thr_progress_has_reached(crr->cpool.thr_prgr)) { @@ -3553,6 +3566,7 @@ schedule_dealloc_carrier(Allctr_t *allctr, Carrier_t *crr) static ERTS_INLINE void cpool_init_carrier_data(Allctr_t *allctr, Carrier_t *crr) { + crr->cpool.homecoming_dd.blk.bhdr = HOMECOMING_MBC_BLK_HDR; erts_atomic_init_nob(&crr->cpool.next, ERTS_AINT_NULL); erts_atomic_init_nob(&crr->cpool.prev, ERTS_AINT_NULL); crr->cpool.orig_allctr = allctr; @@ -3571,8 +3585,7 @@ cpool_init_carrier_data(Allctr_t *allctr, Carrier_t *crr) limit = (csz/100)*allctr->cpool.util_limit; crr->cpool.abandon_limit = limit; } - crr->cpool.abandoned.next = NULL; - crr->cpool.abandoned.prev = NULL; + crr->cpool.state = ERTS_MBC_IS_HOME; } static void @@ -3598,23 +3611,62 @@ set_new_allctr_abandon_limit(Allctr_t *allctr) static void abandon_carrier(Allctr_t *allctr, Carrier_t *crr) { - erts_aint_t max_size; + erts_aint_t iallctr; - STAT_MBC_CPOOL_INSERT(allctr, crr); + STAT_MBC_ABANDON(allctr, crr); unlink_carrier(&allctr->mbc_list, crr); - if (crr->cpool.orig_allctr == allctr) { - link_abandoned_carrier(&allctr->cpool.pooled_list, crr); + allctr->remove_mbc(allctr, crr); + set_new_allctr_abandon_limit(allctr); + + cpool_insert(allctr, crr); + + + iallctr = erts_atomic_read_nob(&crr->allctr); + if (allctr == crr->cpool.orig_allctr) { + /* preserve HOMECOMING flag */ + ASSERT((iallctr & ~ERTS_CRR_ALCTR_FLG_HOMECOMING) == (erts_aint_t)allctr); + erts_atomic_set_wb(&crr->allctr, iallctr | ERTS_CRR_ALCTR_FLG_IN_POOL); + poolify_my_carrier(allctr, crr); } + else { + ASSERT((iallctr & ~ERTS_CRR_ALCTR_FLG_HOMECOMING) == (erts_aint_t)allctr); + iallctr = ((erts_aint_t)crr->cpool.orig_allctr | + ERTS_CRR_ALCTR_FLG_HOMECOMING | + ERTS_CRR_ALCTR_FLG_IN_POOL); + if (!(erts_atomic_xchg_wb(&crr->allctr, iallctr) + & ERTS_CRR_ALCTR_FLG_HOMECOMING)) { + + enqueue_homecoming(allctr, crr); + } + } +} - allctr->remove_mbc(allctr, crr); +static void +enqueue_homecoming(Allctr_t* allctr, Carrier_t* crr) +{ + Allctr_t* orig_allctr = crr->cpool.orig_allctr; + const int cinit = orig_allctr->dd.ix - allctr->dd.ix; + Block_t* dd_blk = &crr->cpool.homecoming_dd.blk; - max_size = (erts_aint_t) allctr->largest_fblk_in_mbc(allctr, crr); - erts_atomic_set_nob(&crr->cpool.max_size, max_size); + /* + * The receiver will recognize this as a carrier + * (and not a block which is the common case) + * since the block header is HOMECOMING_MBC_BLK_HDR. + */ + ASSERT(dd_blk->bhdr == HOMECOMING_MBC_BLK_HDR); + if (ddq_enqueue(&orig_allctr->dd.q, BLK2UMEM(dd_blk), cinit)) + erts_alloc_notify_delayed_dealloc(orig_allctr->ix); +} - cpool_insert(allctr, crr); +static void +poolify_my_carrier(Allctr_t *allctr, Carrier_t *crr) +{ + ERTS_ALC_CPOOL_ASSERT(allctr == crr->cpool.orig_allctr); - set_new_allctr_abandon_limit(allctr); + crr->cpool.pooled.hdr.bhdr = erts_atomic_read_nob(&crr->cpool.max_size); + aoff_add_pooled_mbc(allctr, crr); + crr->cpool.state = ERTS_MBC_WAS_POOLED; } static void @@ -3771,6 +3823,7 @@ create_carrier(Allctr_t *allctr, Uint umem_sz, UWord flags) crr = cpool_fetch(allctr, blk_sz); if (crr) { STAT_MBC_CPOOL_FETCH(allctr, crr); + INC_CC(allctr->cpool.stat.fetch); link_carrier(&allctr->mbc_list, crr); (*allctr->add_mbc)(allctr, crr); blk = (*allctr->get_free_block)(allctr, blk_sz, NULL, 0); @@ -4128,13 +4181,18 @@ destroy_carrier(Allctr_t *allctr, Block_t *blk, Carrier_t **busy_pcrr_pp) #endif if (busy_pcrr_pp && *busy_pcrr_pp) { + erts_aint_t iallctr = erts_atomic_read_nob(&crr->allctr); ERTS_ALC_CPOOL_ASSERT(*busy_pcrr_pp == crr); - *busy_pcrr_pp = NULL; - ERTS_ALC_CPOOL_ASSERT(erts_atomic_read_nob(&crr->allctr) - == (((erts_aint_t) allctr) - | ERTS_CRR_ALCTR_FLG_IN_POOL - | ERTS_CRR_ALCTR_FLG_BUSY)); - erts_atomic_set_nob(&crr->allctr, ((erts_aint_t) allctr)); + ERTS_ALC_CPOOL_ASSERT((iallctr & ~ERTS_CRR_ALCTR_FLG_HOMECOMING) + == (((erts_aint_t) allctr) + | ERTS_CRR_ALCTR_FLG_IN_POOL + | ERTS_CRR_ALCTR_FLG_BUSY)); + ERTS_ALC_CPOOL_ASSERT(allctr == crr->cpool.orig_allctr); + + *busy_pcrr_pp = NULL; + erts_atomic_set_nob(&crr->allctr, + (iallctr & ~(ERTS_CRR_ALCTR_FLG_IN_POOL | + ERTS_CRR_ALCTR_FLG_BUSY))); cpool_delete(allctr, allctr, crr); } else @@ -4184,7 +4242,6 @@ static struct { Eterm e; Eterm t; Eterm ramv; - Eterm sbct; #if HAVE_ERTS_MSEG Eterm asbcst; Eterm rsbcst; @@ -4201,6 +4258,8 @@ static struct { Eterm smbcs; Eterm mbcgs; Eterm acul; + Eterm acnl; + Eterm acfml; #if HAVE_ERTS_MSEG Eterm mmc; @@ -4212,6 +4271,17 @@ static struct { Eterm mbcs; Eterm mbcs_pool; + Eterm fetch; + Eterm fail_pooled; + Eterm fail_shared; + Eterm fail_pend_dealloc; + Eterm fail; + Eterm skip_size; + Eterm skip_busy; + Eterm skip_not_pooled; + Eterm skip_homecoming; + Eterm skip_race; + Eterm entrance_removed; Eterm sbcs; Eterm sys_alloc_carriers_size; @@ -4272,7 +4342,6 @@ init_atoms(Allctr_t *allctr) AM_INIT(e); AM_INIT(t); AM_INIT(ramv); - AM_INIT(sbct); #if HAVE_ERTS_MSEG AM_INIT(asbcst); AM_INIT(rsbcst); @@ -4289,6 +4358,8 @@ init_atoms(Allctr_t *allctr) AM_INIT(smbcs); AM_INIT(mbcgs); AM_INIT(acul); + AM_INIT(acnl); + AM_INIT(acfml); #if HAVE_ERTS_MSEG AM_INIT(mmc); @@ -4300,6 +4371,17 @@ init_atoms(Allctr_t *allctr) AM_INIT(mbcs); AM_INIT(mbcs_pool); + AM_INIT(fetch); + AM_INIT(fail_pooled); + AM_INIT(fail_shared); + AM_INIT(fail_pend_dealloc); + AM_INIT(fail); + AM_INIT(skip_size); + AM_INIT(skip_busy); + AM_INIT(skip_not_pooled); + AM_INIT(skip_homecoming); + AM_INIT(skip_race); + AM_INIT(entrance_removed); AM_INIT(sbcs); AM_INIT(sys_alloc_carriers_size); @@ -4583,9 +4665,56 @@ info_cpool(Allctr_t *allctr, if (hpp || szp) { res = NIL; + + if (!sz_only) { + add_3tup(hpp, szp, &res, am.fail_pooled, + bld_unstable_uint(hpp, szp, ERTS_ALC_CC_GIGA_VAL(allctr->cpool.stat.fail_pooled)), + bld_unstable_uint(hpp, szp, ERTS_ALC_CC_VAL(allctr->cpool.stat.fail_pooled))); + + add_3tup(hpp, szp, &res, am.fail_shared, + bld_unstable_uint(hpp, szp, ERTS_ALC_CC_GIGA_VAL(allctr->cpool.stat.fail_shared)), + bld_unstable_uint(hpp, szp, ERTS_ALC_CC_VAL(allctr->cpool.stat.fail_shared))); + + add_3tup(hpp, szp, &res, am.fail_pend_dealloc, + bld_unstable_uint(hpp, szp, ERTS_ALC_CC_GIGA_VAL(allctr->cpool.stat.fail_pend_dealloc)), + bld_unstable_uint(hpp, szp, ERTS_ALC_CC_VAL(allctr->cpool.stat.fail_pend_dealloc))); + + add_3tup(hpp, szp, &res, am.fail, + bld_unstable_uint(hpp, szp, ERTS_ALC_CC_GIGA_VAL(allctr->cpool.stat.fail)), + bld_unstable_uint(hpp, szp, ERTS_ALC_CC_VAL(allctr->cpool.stat.fail))); + + add_3tup(hpp, szp, &res, am.fetch, + bld_unstable_uint(hpp, szp, ERTS_ALC_CC_GIGA_VAL(allctr->cpool.stat.fetch)), + bld_unstable_uint(hpp, szp, ERTS_ALC_CC_VAL(allctr->cpool.stat.fetch))); + + add_3tup(hpp, szp, &res, am.skip_size, + bld_unstable_uint(hpp, szp, ERTS_ALC_CC_GIGA_VAL(allctr->cpool.stat.skip_size)), + bld_unstable_uint(hpp, szp, ERTS_ALC_CC_VAL(allctr->cpool.stat.skip_size))); + + add_3tup(hpp, szp, &res, am.skip_busy, + bld_unstable_uint(hpp, szp, ERTS_ALC_CC_GIGA_VAL(allctr->cpool.stat.skip_busy)), + bld_unstable_uint(hpp, szp, ERTS_ALC_CC_VAL(allctr->cpool.stat.skip_busy))); + + add_3tup(hpp, szp, &res, am.skip_not_pooled, + bld_unstable_uint(hpp, szp, ERTS_ALC_CC_GIGA_VAL(allctr->cpool.stat.skip_not_pooled)), + bld_unstable_uint(hpp, szp, ERTS_ALC_CC_VAL(allctr->cpool.stat.skip_not_pooled))); + + add_3tup(hpp, szp, &res, am.skip_homecoming, + bld_unstable_uint(hpp, szp, ERTS_ALC_CC_GIGA_VAL(allctr->cpool.stat.skip_homecoming)), + bld_unstable_uint(hpp, szp, ERTS_ALC_CC_VAL(allctr->cpool.stat.skip_homecoming))); + + add_3tup(hpp, szp, &res, am.skip_race, + bld_unstable_uint(hpp, szp, ERTS_ALC_CC_GIGA_VAL(allctr->cpool.stat.skip_race)), + bld_unstable_uint(hpp, szp, ERTS_ALC_CC_VAL(allctr->cpool.stat.skip_race))); + + add_3tup(hpp, szp, &res, am.entrance_removed, + bld_unstable_uint(hpp, szp, ERTS_ALC_CC_GIGA_VAL(allctr->cpool.stat.entrance_removed)), + bld_unstable_uint(hpp, szp, ERTS_ALC_CC_VAL(allctr->cpool.stat.entrance_removed))); + add_2tup(hpp, szp, &res, am.carriers_size, bld_unstable_uint(hpp, szp, csz)); + } if (!sz_only) add_2tup(hpp, szp, &res, am.carriers, @@ -4844,7 +4973,7 @@ info_options(Allctr_t *allctr, Uint *szp) { Eterm res = THE_NON_VALUE; - int acul; + UWord acul, acnl, acfml; if (!allctr) { if (print_to_p) @@ -4857,6 +4986,8 @@ info_options(Allctr_t *allctr, } acul = allctr->cpool.util_limit; + acnl = allctr->cpool.in_pool_limit; + acfml = allctr->cpool.fblk_min_limit; if (print_to_p) { char topt[21]; /* Enough for any 64-bit integer */ @@ -4884,7 +5015,7 @@ info_options(Allctr_t *allctr, "option lmbcs: %beu\n" "option smbcs: %beu\n" "option mbcgs: %beu\n" - "option acul: %d\n", + "option acul: %bpu\n", topt, allctr->ramv ? "true" : "false", allctr->sbc_threshold, @@ -4909,9 +5040,15 @@ info_options(Allctr_t *allctr, hpp, szp); if (hpp || szp) { + add_2tup(hpp, szp, &res, + am.acfml, + bld_uint(hpp, szp, acfml)); + add_2tup(hpp, szp, &res, + am.acnl, + bld_uint(hpp, szp, acnl)); add_2tup(hpp, szp, &res, am.acul, - bld_uint(hpp, szp, (UWord) acul)); + bld_uint(hpp, szp, acul)); add_2tup(hpp, szp, &res, am.mbcgs, bld_uint(hpp, szp, allctr->mbc_growth_stages)); @@ -4947,7 +5084,7 @@ info_options(Allctr_t *allctr, bld_uint(hpp, szp, allctr->mseg_opt.abs_shrink_th)); #endif add_2tup(hpp, szp, &res, - am.sbct, + 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.t, (allctr->t ? am_true : am_false)); @@ -5481,12 +5618,13 @@ erts_alcu_free_thr_pref(ErtsAlcType_t type, void *extra, void *p) pref_allctr = get_pref_allctr(extra); used_allctr = get_used_allctr(pref_allctr, ERTS_ALC_TS_PREF_LOCK_IF_USED, p, NULL, &busy_pcrr_p); - if (pref_allctr != used_allctr) + if (pref_allctr != used_allctr) { enqueue_dealloc_other_instance(type, - used_allctr, - p, - (used_allctr->dd.ix - - pref_allctr->dd.ix)); + used_allctr, + p, + (used_allctr->dd.ix + - pref_allctr->dd.ix)); + } else { ERTS_ALCU_DBG_CHK_THR_ACCESS(used_allctr); do_erts_alcu_free(type, used_allctr, p, &busy_pcrr_p); @@ -5854,6 +5992,37 @@ erts_alcu_realloc_mv_thr_pref(ErtsAlcType_t type, void *extra, +static Uint adjust_sbct(Allctr_t* allctr, Uint sbct) +{ +#ifndef ARCH_64 + if (sbct > 0) { + Uint max_mbc_block_sz = UNIT_CEILING(sbct - 1 + ABLK_HDR_SZ); + if (max_mbc_block_sz + UNIT_FLOOR(allctr->min_block_size - 1) > MBC_ABLK_SZ_MASK + || max_mbc_block_sz < sbct) { /* wrap around */ + /* + * By limiting sbc_threshold to (hard limit - min_block_size) + * we avoid having to split off free "residue blocks" + * smaller than min_block_size. + */ + max_mbc_block_sz = MBC_ABLK_SZ_MASK - UNIT_FLOOR(allctr->min_block_size - 1); + sbct = max_mbc_block_sz - ABLK_HDR_SZ + 1; + } + } +#endif + return sbct; +} + +int erts_alcu_try_set_dyn_param(Allctr_t* allctr, Eterm param, Uint value) +{ + const Uint MIN_DYN_SBCT = 4000; /* a lame catastrophe prevention */ + + if (param == am_sbct && value >= MIN_DYN_SBCT) { + allctr->sbc_threshold = adjust_sbct(allctr, value); + return 1; + } + return 0; +} + /* ------------------------------------------------------------------------- */ int @@ -5941,10 +6110,7 @@ erts_alcu_start(Allctr_t *allctr, AllctrInit_t *init) allctr->min_block_size = sz; } - allctr->cpool.pooled_list.next = &allctr->cpool.pooled_list; - allctr->cpool.pooled_list.prev = &allctr->cpool.pooled_list; - allctr->cpool.traitor_list.next = &allctr->cpool.traitor_list; - allctr->cpool.traitor_list.prev = &allctr->cpool.traitor_list; + allctr->cpool.pooled_tree = NULL; allctr->cpool.dc_list.first = NULL; allctr->cpool.dc_list.last = NULL; allctr->cpool.abandon_limit = 0; @@ -5954,24 +6120,18 @@ erts_alcu_start(Allctr_t *allctr, AllctrInit_t *init) erts_atomic_init_nob(&allctr->cpool.stat.carriers_size, 0); erts_atomic_init_nob(&allctr->cpool.stat.no_carriers, 0); allctr->cpool.check_limit_count = ERTS_ALC_CPOOL_CHECK_LIMIT_COUNT; - allctr->cpool.util_limit = init->ts ? 0 : init->acul; - - allctr->sbc_threshold = init->sbct; -#ifndef ARCH_64 - if (allctr->sbc_threshold > 0) { - Uint max_mbc_block_sz = UNIT_CEILING(allctr->sbc_threshold - 1 + ABLK_HDR_SZ); - if (max_mbc_block_sz + UNIT_FLOOR(allctr->min_block_size - 1) > MBC_ABLK_SZ_MASK - || max_mbc_block_sz < allctr->sbc_threshold) { /* wrap around */ - /* - * By limiting sbc_threshold to (hard limit - min_block_size) - * we avoid having to split off free "residue blocks" - * smaller than min_block_size. - */ - max_mbc_block_sz = MBC_ABLK_SZ_MASK - UNIT_FLOOR(allctr->min_block_size - 1); - allctr->sbc_threshold = max_mbc_block_sz - ABLK_HDR_SZ + 1; - } + if (!init->ts && init->acul && init->acnl) { + allctr->cpool.util_limit = init->acul; + allctr->cpool.in_pool_limit = init->acnl; + allctr->cpool.fblk_min_limit = init->acfml; } -#endif + else { + allctr->cpool.util_limit = 0; + allctr->cpool.in_pool_limit = 0; + allctr->cpool.fblk_min_limit = 0; + } + + allctr->sbc_threshold = adjust_sbct(allctr, init->sbct); #if HAVE_ERTS_MSEG if (allctr->mseg_opt.abs_shrink_th > ~((UWord) 0) / 100) @@ -6022,6 +6182,9 @@ erts_alcu_start(Allctr_t *allctr, AllctrInit_t *init) allctr->sys_realloc = &erts_alcu_sys_realloc; allctr->sys_dealloc = &erts_alcu_sys_dealloc; } + + allctr->try_set_dyn_param = &erts_alcu_try_set_dyn_param; + #if HAVE_ERTS_MSEG if (init->mseg_alloc) { ASSERT(init->mseg_realloc && init->mseg_dealloc); @@ -6036,6 +6199,7 @@ erts_alcu_start(Allctr_t *allctr, AllctrInit_t *init) allctr->mseg_realloc = &erts_alcu_mseg_realloc; allctr->mseg_dealloc = &erts_alcu_mseg_dealloc; } + /* If a custom carrier alloc function is specified, make sure it's used */ if (init->mseg_alloc && !init->sys_alloc) { allctr->crr_set_flgs = CFLG_FORCE_MSEG; diff --git a/erts/emulator/beam/erl_alloc_util.h b/erts/emulator/beam/erl_alloc_util.h index faeb5ef368..9a6de2bb75 100644 --- a/erts/emulator/beam/erl_alloc_util.h +++ b/erts/emulator/beam/erl_alloc_util.h @@ -61,7 +61,9 @@ typedef struct { UWord lmbcs; UWord smbcs; UWord mbcgs; - int acul; + UWord acul; + UWord acnl; + UWord acfml; void *fix; size_t *fix_type_size; @@ -116,6 +118,8 @@ typedef struct { 1024*1024, /* (bytes) smbcs: smallest mbc size */\ 10, /* (amount) mbcgs: mbc growth stages */\ 0, /* (%) acul: abandon carrier utilization limit */\ + 1000, /* (amount) acnl: abandoned carriers number limit */\ + 0, /* (bytes) acfml: abandoned carrier fblk min limit */\ /* --- Data not options -------------------------------------------- */\ NULL, /* (ptr) fix */\ NULL /* (ptr) fix_type_size */\ @@ -149,6 +153,8 @@ typedef struct { 128*1024, /* (bytes) smbcs: smallest mbc size */\ 10, /* (amount) mbcgs: mbc growth stages */\ 0, /* (%) acul: abandon carrier utilization limit */\ + 1000, /* (amount) acnl: abandoned carriers number limit */\ + 0, /* (bytes) acfml: abandoned carrier fblk min limit */\ /* --- Data not options -------------------------------------------- */\ NULL, /* (ptr) fix */\ NULL /* (ptr) fix_type_size */\ @@ -216,6 +222,8 @@ void erts_alcu_literal_32_sys_dealloc(Allctr_t*, void *ptr, Uint size, int supe void erts_lcnt_update_allocator_locks(int enable); #endif +int erts_alcu_try_set_dyn_param(Allctr_t*, Eterm param, Uint value); + #endif /* !ERL_ALLOC_UTIL__ */ #if defined(GET_ERL_ALLOC_UTIL_IMPL) && !defined(ERL_ALLOC_UTIL_IMPL__) @@ -296,41 +304,7 @@ void erts_lcnt_update_allocator_locks(int enable); typedef union {char c[ERTS_ALLOC_ALIGN_BYTES]; long l; double d;} Unit_t; - -typedef struct ErtsDoubleLink_t_ { - struct ErtsDoubleLink_t_ *next; - struct ErtsDoubleLink_t_ *prev; -}ErtsDoubleLink_t; - -typedef struct { - erts_atomic_t next; - erts_atomic_t prev; - Allctr_t *orig_allctr; /* read-only while carrier is alive */ - ErtsThrPrgrVal thr_prgr; - erts_atomic_t max_size; - UWord abandon_limit; - UWord blocks; - UWord blocks_size; - ErtsDoubleLink_t abandoned; /* node in pooled_list or traitor_list */ -} ErtsAlcCPoolData_t; - - typedef struct Carrier_t_ Carrier_t; -struct Carrier_t_ { - UWord chdr; - Carrier_t *next; - Carrier_t *prev; - erts_atomic_t allctr; - ErtsAlcCPoolData_t cpool; /* Overwritten by block if sbc */ -}; - -#define ERTS_ALC_CARRIER_TO_ALLCTR(C) \ - ((Allctr_t *) (erts_atomic_read_nob(&(C)->allctr) & ~FLG_MASK)) - -typedef struct { - Carrier_t *first; - Carrier_t *last; -} CarrierList_t; typedef struct { UWord bhdr; @@ -344,6 +318,22 @@ typedef struct { #endif } Block_t; +typedef union ErtsAllctrDDBlock_t_ ErtsAllctrDDBlock_t; + +union ErtsAllctrDDBlock_t_ { + erts_atomic_t atmc_next; + ErtsAllctrDDBlock_t *ptr_next; +}; + +typedef struct { + Block_t blk; +#if !MBC_ABLK_OFFSET_BITS + ErtsAllctrDDBlock_t umem_; +#endif +} ErtsFakeDDBlock_t; + + + #define THIS_FREE_BLK_HDR_FLG (((UWord) 1) << 0) #define PREV_FREE_BLK_HDR_FLG (((UWord) 1) << 1) #define LAST_BLK_HDR_FLG (((UWord) 1) << 2) @@ -352,14 +342,13 @@ typedef struct { (THIS_FREE_BLK_HDR_FLG | PREV_FREE_BLK_HDR_FLG | LAST_BLK_HDR_FLG) /* - * FREE_LAST_MBC_BLK_HDR_FLGS is a special flag combo used for - * distinguishing empty mbc's from allocated blocks in - * handle_delayed_dealloc(). + * HOMECOMING_MBC_BLK_HDR is a special block header combo used for + * distinguishing MBC's from allocated blocks in handle_delayed_dealloc(). */ -#define FREE_LAST_MBC_BLK_HDR_FLGS (THIS_FREE_BLK_HDR_FLG | LAST_BLK_HDR_FLG) +#define HOMECOMING_MBC_BLK_HDR (THIS_FREE_BLK_HDR_FLG | LAST_BLK_HDR_FLG) #define IS_FREE_LAST_MBC_BLK(B) \ - (((B)->bhdr & FLG_MASK) == FREE_LAST_MBC_BLK_HDR_FLGS) + (((B)->bhdr & FLG_MASK) == (THIS_FREE_BLK_HDR_FLG | LAST_BLK_HDR_FLG)) #define IS_SBC_BLK(B) (((B)->bhdr & FLG_MASK) == SBC_BLK_HDR_FLG) #define IS_MBC_BLK(B) (!IS_SBC_BLK((B))) @@ -383,6 +372,57 @@ typedef struct { typedef UWord FreeBlkFtr_t; /* Footer of a free block */ +/* This AOFF stuff really belong in erl_ao_firstfit_alloc.h */ +typedef struct AOFF_RBTree_t_ AOFF_RBTree_t; +struct AOFF_RBTree_t_ { + Block_t hdr; + AOFF_RBTree_t *parent; + AOFF_RBTree_t *left; + AOFF_RBTree_t *right; + Uint32 flags; + Uint32 max_sz; /* of all blocks in this sub-tree */ +}; + +void aoff_add_pooled_mbc(Allctr_t*, Carrier_t*); +void aoff_remove_pooled_mbc(Allctr_t*, Carrier_t*); +Carrier_t* aoff_lookup_pooled_mbc(Allctr_t*, Uint size); +void erts_aoff_larger_max_size(AOFF_RBTree_t *node); + +typedef struct { + ErtsFakeDDBlock_t homecoming_dd; + erts_atomic_t next; + erts_atomic_t prev; + Allctr_t *orig_allctr; /* read-only while carrier is alive */ + ErtsThrPrgrVal thr_prgr; + erts_atomic_t max_size; + UWord abandon_limit; + UWord blocks; + UWord blocks_size; + enum { + ERTS_MBC_IS_HOME, + ERTS_MBC_WAS_POOLED, + ERTS_MBC_WAS_TRAITOR + } state; + AOFF_RBTree_t pooled; /* node in pooled_tree */ +} ErtsAlcCPoolData_t; + +struct Carrier_t_ { + UWord chdr; + Carrier_t *next; + Carrier_t *prev; + erts_atomic_t allctr; + ErtsAlcCPoolData_t cpool; /* Overwritten by block if sbc */ +}; + +#define ERTS_ALC_CARRIER_TO_ALLCTR(C) \ + ((Allctr_t *) (erts_atomic_read_nob(&(C)->allctr) & ~FLG_MASK)) + +typedef struct { + Carrier_t *first; + Carrier_t *last; +} CarrierList_t; + + typedef Uint64 CallCounter_t; typedef struct { @@ -419,13 +459,6 @@ typedef struct { #endif -typedef union ErtsAllctrDDBlock_t_ ErtsAllctrDDBlock_t; - -union ErtsAllctrDDBlock_t_ { - erts_atomic_t atmc_next; - ErtsAllctrDDBlock_t *ptr_next; -}; - typedef struct { ErtsAllctrDDBlock_t marker; erts_atomic_t last; @@ -537,25 +570,37 @@ struct Allctr_t_ { UWord crr_set_flgs; UWord crr_clr_flgs; - /* Carriers */ + /* Carriers *employed* by this allocator */ CarrierList_t mbc_list; CarrierList_t sbc_list; struct { - /* pooled_list, traitor list and dc_list contain only - carriers _created_ by this allocator */ - ErtsDoubleLink_t pooled_list; - ErtsDoubleLink_t traitor_list; + /* pooled_tree and dc_list contain only + carriers *created* by this allocator */ + AOFF_RBTree_t* pooled_tree; CarrierList_t dc_list; UWord abandon_limit; int disable_abandon; int check_limit_count; - int util_limit; + UWord util_limit; /* acul */ + UWord in_pool_limit; /* acnl */ + UWord fblk_min_limit; /* acmfl */ struct { erts_atomic_t blocks_size; erts_atomic_t no_blocks; erts_atomic_t carriers_size; erts_atomic_t no_carriers; + CallCounter_t fail_pooled; + CallCounter_t fail_shared; + CallCounter_t fail_pend_dealloc; + CallCounter_t fail; + CallCounter_t fetch; + CallCounter_t skip_size; + CallCounter_t skip_busy; + CallCounter_t skip_not_pooled; + CallCounter_t skip_homecoming; + CallCounter_t skip_race; + CallCounter_t entrance_removed; } stat; } cpool; @@ -589,6 +634,8 @@ struct Allctr_t_ { void* (*sys_realloc)(Allctr_t *allctr, void *ptr, Uint *size_p, Uint old_size, int superalign); void (*sys_dealloc)(Allctr_t *allctr, void *ptr, Uint size, int superalign); + int (*try_set_dyn_param)(Allctr_t*, Eterm param, Uint value); + void (*init_atoms) (void); #ifdef ERTS_ALLOC_UTIL_HARD_DEBUG diff --git a/erts/emulator/beam/erl_ao_firstfit_alloc.c b/erts/emulator/beam/erl_ao_firstfit_alloc.c index 05ba1f9891..f8a6101b95 100644 --- a/erts/emulator/beam/erl_ao_firstfit_alloc.c +++ b/erts/emulator/beam/erl_ao_firstfit_alloc.c @@ -20,7 +20,7 @@ /* - * Description: An "address order first fit" allocator + * Description: A family of "first fit" allocator strategies * 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. @@ -40,6 +40,10 @@ * sorting order. Blocks within the same carrier are sorted * wrt size instead of address. The 'max_sz' field is maintained * in order to dismiss entire carriers with too small blocks. + * Age Order: + * Carriers are ordered by creation time instead of address. + * Oldest carrier with a large enough free block is chosen. + * No age order supported for blocks. * * Authors: Rickard Green/Sverker Eriksson */ @@ -53,10 +57,12 @@ #include "erl_ao_firstfit_alloc.h" #ifdef DEBUG +# define IS_DEBUG 1 #if 0 #define HARD_DEBUG #endif #else +# define IS_DEBUG 0 #undef HARD_DEBUG #endif @@ -92,18 +98,6 @@ #define RBT_ASSERT(x) #endif - -/* Types... */ -typedef struct AOFF_RBTree_t_ AOFF_RBTree_t; - -struct AOFF_RBTree_t_ { - Block_t hdr; - AOFF_RBTree_t *parent; - AOFF_RBTree_t *left; - AOFF_RBTree_t *right; - Uint32 flags; - Uint32 max_sz; /* of all blocks in this sub-tree */ -}; #define AOFF_BLK_SZ(B) MBC_FBLK_SZ(&(B)->hdr) /* BF block nodes keeps list of all with equal size @@ -121,6 +115,7 @@ typedef struct AOFF_Carrier_t_ AOFF_Carrier_t; struct AOFF_Carrier_t_ { Carrier_t crr; AOFF_RBTree_t rbt_node; /* My node in the carrier tree */ + Sint64 birth_time; AOFF_RBTree_t* root; /* Root of my block tree */ }; #define RBT_NODE_TO_MBC(PTR) ErtsContainerStruct((PTR), AOFF_Carrier_t, rbt_node) @@ -136,12 +131,12 @@ struct AOFF_Carrier_t_ { */ #ifdef HARD_DEBUG -# define HARD_CHECK_IS_MEMBER(ROOT,NODE) rbt_assert_is_member(ROOT,NODE) -# define HARD_CHECK_TREE(CRR,FLV,ROOT,SZ) check_tree(CRR, FLV, ROOT, SZ) -static AOFF_RBTree_t * check_tree(Carrier_t* within_crr, enum AOFF_Flavor flavor, AOFF_RBTree_t* root, Uint); +# define HARD_CHECK_IS_MEMBER(ROOT,NODE) ASSERT(rbt_is_member(ROOT,NODE)) +# define HARD_CHECK_TREE(CRR,ORDER,ROOT,SZ) check_tree(CRR, ORDER, ROOT, SZ) +static AOFF_RBTree_t * check_tree(Carrier_t*, enum AOFFSortOrder, AOFF_RBTree_t*, Uint); #else # define HARD_CHECK_IS_MEMBER(ROOT,NODE) -# define HARD_CHECK_TREE(CRR,FLV,ROOT,SZ) +# define HARD_CHECK_TREE(CRR,ORDER,ROOT,SZ) #endif @@ -179,25 +174,63 @@ static ERTS_INLINE void lower_max_size(AOFF_RBTree_t *node, else ASSERT(new_max == old_max); } -static ERTS_INLINE SWord cmp_blocks(enum AOFF_Flavor flavor, +/* + * Set possibly new larger 'max_sz' of node and propagate change toward root + */ +void erts_aoff_larger_max_size(AOFF_RBTree_t *node) +{ + AOFF_RBTree_t* x = node; + const Uint new_sz = node->hdr.bhdr; + + ASSERT(!x->left || x->left->max_sz <= x->max_sz); + ASSERT(!x->right || x->right->max_sz <= x->max_sz); + + while (new_sz > x->max_sz) { + x->max_sz = new_sz; + x = x->parent; + if (!x) + break; + } +} + +/* Compare nodes for both carrier and block trees */ +static ERTS_INLINE SWord cmp_blocks(enum AOFFSortOrder order, AOFF_RBTree_t* lhs, AOFF_RBTree_t* rhs) { ASSERT(lhs != rhs); - ASSERT(flavor == AOFF_AOFF || FBLK_TO_MBC(&lhs->hdr) == FBLK_TO_MBC(&rhs->hdr)); - if (flavor != AOFF_AOFF) { - SWord diff = (SWord)AOFF_BLK_SZ(lhs) - (SWord)AOFF_BLK_SZ(rhs); - if (diff || flavor == AOFF_BF) return diff; + if (order == FF_AGEFF) { + AOFF_Carrier_t* lc = RBT_NODE_TO_MBC(lhs); + AOFF_Carrier_t* rc = RBT_NODE_TO_MBC(rhs); + Sint64 diff = lc->birth_time - rc->birth_time; + #ifdef ARCH_64 + if (diff) + return diff; + #else + if (diff < 0) + return -1; + else if (diff > 0) + return 1; + #endif + } + else { + ASSERT(order == FF_AOFF || FBLK_TO_MBC(&lhs->hdr) == FBLK_TO_MBC(&rhs->hdr)); + if (order != FF_AOFF) { + SWord diff = (SWord)AOFF_BLK_SZ(lhs) - (SWord)AOFF_BLK_SZ(rhs); + if (diff || order == FF_BF) return diff; + } } return (char*)lhs - (char*)rhs; } -static ERTS_INLINE SWord cmp_cand_blk(enum AOFF_Flavor flavor, +/* Compare candidate block. Only for block tree */ +static ERTS_INLINE SWord cmp_cand_blk(enum AOFFSortOrder order, Block_t* cand_blk, AOFF_RBTree_t* rhs) { - if (flavor != AOFF_AOFF) { + ASSERT(order != FF_AGEFF); + if (order != FF_AOFF) { if (BLK_TO_MBC(cand_blk) == FBLK_TO_MBC(&rhs->hdr)) { SWord diff = (SWord)MBC_BLK_SZ(cand_blk) - (SWord)MBC_FBLK_SZ(&rhs->hdr); - if (diff || flavor == AOFF_BF) return diff; + if (diff || order == FF_BF) return diff; } } return (char*)cand_blk - (char*)rhs; @@ -218,11 +251,8 @@ static UWord aoff_largest_fblk_in_mbc(Allctr_t*, Carrier_t*); /* Generic tree functions used by both carrier and block trees. */ static void rbt_delete(AOFF_RBTree_t** root, AOFF_RBTree_t* del); -static void rbt_insert(enum AOFF_Flavor flavor, AOFF_RBTree_t** root, AOFF_RBTree_t* blk); +static void rbt_insert(enum AOFFSortOrder, AOFF_RBTree_t** root, AOFF_RBTree_t* blk); static AOFF_RBTree_t* rbt_search(AOFF_RBTree_t* root, Uint size); -#ifdef HARD_DEBUG -static int rbt_assert_is_member(AOFF_RBTree_t* root, AOFF_RBTree_t* node); -#endif static Eterm info_options(Allctr_t *, char *, fmtfn_t *, void *, Uint **, Uint *); static void init_atoms(void); @@ -230,10 +260,17 @@ static void init_atoms(void); static int atoms_initialized = 0; +#ifndef ERTS_HAVE_OS_MONOTONIC_TIME_SUPPORT +static erts_atomic64_t birth_time_counter; +#endif + void erts_aoffalc_init(void) { atoms_initialized = 0; +#ifndef ERTS_HAVE_OS_MONOTONIC_TIME_SUPPORT + erts_atomic64_init_nob(&birth_time_counter, 0); +#endif } Allctr_t * @@ -254,11 +291,12 @@ erts_aoffalc_start(AOFFAllctr_t *alc, sys_memcpy((void *) alc, (void *) &zero.allctr, sizeof(AOFFAllctr_t)); - alc->flavor = aoffinit->flavor; + alc->blk_order = aoffinit->blk_order; + alc->crr_order = aoffinit->crr_order; allctr->mbc_header_size = sizeof(AOFF_Carrier_t); allctr->min_mbc_size = MIN_MBC_SZ; allctr->min_mbc_first_free_size = MIN_MBC_FIRST_FREE_SZ; - allctr->min_block_size = (aoffinit->flavor == AOFF_BF ? + allctr->min_block_size = (aoffinit->blk_order == FF_BF ? sizeof(AOFF_RBTreeList_t):sizeof(AOFF_RBTree_t)); allctr->vsn_str = ERTS_ALC_AOFF_ALLOC_VSN_STR; @@ -487,9 +525,9 @@ aoff_unlink_free_block(Allctr_t *allctr, Block_t *blk) AOFF_Carrier_t *crr = (AOFF_Carrier_t*) FBLK_TO_MBC(&del->hdr); ASSERT(crr->rbt_node.hdr.bhdr == crr->root->max_sz); - HARD_CHECK_TREE(&crr->crr, alc->flavor, crr->root, 0); + HARD_CHECK_TREE(&crr->crr, alc->blk_order, crr->root, 0); - if (alc->flavor == AOFF_BF) { + if (alc->blk_order == FF_BF) { ASSERT(del->flags & IS_BF_FLG); if (IS_LIST_ELEM(del)) { /* Remove from list */ @@ -510,14 +548,14 @@ aoff_unlink_free_block(Allctr_t *allctr, Block_t *blk) replace(&crr->root, (AOFF_RBTree_t*)del, LIST_NEXT(del)); - HARD_CHECK_TREE(&crr->crr, alc->flavor, crr->root, 0); + HARD_CHECK_TREE(&crr->crr, alc->blk_order, crr->root, 0); return; } } rbt_delete(&crr->root, (AOFF_RBTree_t*)del); - HARD_CHECK_TREE(&crr->crr, alc->flavor, crr->root, 0); + HARD_CHECK_TREE(&crr->crr, alc->blk_order, crr->root, 0); /* Update the carrier tree with a potentially new (lower) max_sz */ @@ -715,32 +753,33 @@ aoff_link_free_block(Allctr_t *allctr, Block_t *block) ASSERT(allctr == ERTS_ALC_CARRIER_TO_ALLCTR(&blk_crr->crr)); ASSERT(blk_crr->rbt_node.hdr.bhdr == (blk_crr->root ? blk_crr->root->max_sz : 0)); - HARD_CHECK_TREE(&blk_crr->crr, alc->flavor, blk_crr->root, 0); + HARD_CHECK_TREE(&blk_crr->crr, alc->blk_order, blk_crr->root, 0); - rbt_insert(alc->flavor, &blk_crr->root, blk); + rbt_insert(alc->blk_order, &blk_crr->root, blk); - /* Update the carrier tree with a potentially new (larger) max_sz - */ + /* + * Update carrier tree with a potentially new (larger) max_sz + */ crr_node = &blk_crr->rbt_node; if (blk_sz > crr_node->hdr.bhdr) { - ASSERT(blk_sz == blk_crr->root->max_sz); - crr_node->hdr.bhdr = blk_sz; - while (blk_sz > crr_node->max_sz) { - crr_node->max_sz = blk_sz; - crr_node = crr_node->parent; - if (!crr_node) break; - } + ASSERT(blk_sz == blk_crr->root->max_sz); + crr_node->hdr.bhdr = blk_sz; + while (blk_sz > crr_node->max_sz) { + crr_node->max_sz = blk_sz; + crr_node = crr_node->parent; + if (!crr_node) break; + } } - HARD_CHECK_TREE(&blk_crr->crr, alc->flavor, blk_crr->root, 0); + HARD_CHECK_TREE(NULL, alc->crr_order, alc->mbc_root, 0); } static void -rbt_insert(enum AOFF_Flavor flavor, AOFF_RBTree_t** root, AOFF_RBTree_t* blk) +rbt_insert(enum AOFFSortOrder order, AOFF_RBTree_t** root, AOFF_RBTree_t* blk) { Uint blk_sz = AOFF_BLK_SZ(blk); #ifdef DEBUG - blk->flags = (flavor == AOFF_BF) ? IS_BF_FLG : 0; + blk->flags = (order == FF_BF) ? IS_BF_FLG : 0; #else blk->flags = 0; #endif @@ -760,7 +799,7 @@ rbt_insert(enum AOFF_Flavor flavor, AOFF_RBTree_t** root, AOFF_RBTree_t* blk) if (x->max_sz < blk_sz) { x->max_sz = blk_sz; } - diff = cmp_blocks(flavor, blk, x); + diff = cmp_blocks(order, blk, x); if (diff < 0) { if (!x->left) { blk->parent = x; @@ -778,7 +817,7 @@ rbt_insert(enum AOFF_Flavor flavor, AOFF_RBTree_t** root, AOFF_RBTree_t* blk) x = x->right; } else { - ASSERT(flavor == AOFF_BF); + ASSERT(order == FF_BF); ASSERT(blk->flags & IS_BF_FLG); ASSERT(x->flags & IS_BF_FLG); SET_LIST_ELEM(blk); @@ -798,7 +837,7 @@ rbt_insert(enum AOFF_Flavor flavor, AOFF_RBTree_t** root, AOFF_RBTree_t* blk) if (IS_RED(blk->parent)) tree_insert_fixup(root, blk); } - if (flavor == AOFF_BF) { + if (order == FF_BF) { SET_TREE_NODE(blk); LIST_NEXT(blk) = NULL; } @@ -826,6 +865,16 @@ rbt_search(AOFF_RBTree_t* root, Uint size) } } +Carrier_t* aoff_lookup_pooled_mbc(Allctr_t* allctr, Uint size) +{ + AOFF_RBTree_t* node; + + if (!allctr->cpool.pooled_tree) + return NULL; + node = rbt_search(allctr->cpool.pooled_tree, size); + return node ? ErtsContainerStruct(node, Carrier_t, cpool.pooled) : NULL; +} + static Block_t * aoff_get_free_block(Allctr_t *allctr, Uint size, Block_t *cand_blk, Uint cand_size) @@ -850,7 +899,7 @@ aoff_get_free_block(Allctr_t *allctr, Uint size, /* Get block within carrier tree */ #ifdef HARD_DEBUG - dbg_blk = HARD_CHECK_TREE(&crr->crr, alc->flavor, crr->root, size); + dbg_blk = HARD_CHECK_TREE(&crr->crr, alc->blk_order, crr->root, size); #endif blk = rbt_search(crr->root, size); @@ -863,7 +912,7 @@ aoff_get_free_block(Allctr_t *allctr, Uint size, if (!blk) return NULL; - if (cand_blk && cmp_cand_blk(alc->flavor, cand_blk, blk) < 0) { + if (cand_blk && cmp_cand_blk(alc->blk_order, cand_blk, blk) < 0) { return NULL; /* cand_blk was better */ } @@ -872,23 +921,32 @@ aoff_get_free_block(Allctr_t *allctr, Uint size, return (Block_t *) blk; } +static ERTS_INLINE Sint64 get_birth_time(void) +{ +#ifdef ERTS_HAVE_OS_MONOTONIC_TIME_SUPPORT + return (Sint64) erts_os_monotonic_time(); +#else + return (Sint64) erts_atomic64_inc_read_nob(&birth_time_counter); +#endif +} + static void aoff_creating_mbc(Allctr_t *allctr, Carrier_t *carrier) { AOFFAllctr_t *alc = (AOFFAllctr_t *) allctr; AOFF_Carrier_t *crr = (AOFF_Carrier_t*) carrier; AOFF_RBTree_t **root = &alc->mbc_root; - HARD_CHECK_TREE(NULL, 0, *root, 0); + HARD_CHECK_TREE(NULL, alc->crr_order, *root, 0); - /* Link carrier in address order tree - */ crr->rbt_node.hdr.bhdr = 0; - rbt_insert(AOFF_AOFF, root, &crr->rbt_node); + if (alc->crr_order == FF_AGEFF || IS_DEBUG) + crr->birth_time = get_birth_time(); + rbt_insert(alc->crr_order, root, &crr->rbt_node); /* aoff_link_free_block will add free block later */ crr->root = NULL; - HARD_CHECK_TREE(NULL, 0, *root, 0); + HARD_CHECK_TREE(NULL, alc->crr_order, *root, 0); } #define IS_CRR_IN_TREE(CRR,ROOT) \ @@ -911,27 +969,38 @@ static void aoff_add_mbc(Allctr_t *allctr, Carrier_t *carrier) AOFF_RBTree_t **root = &alc->mbc_root; ASSERT(!IS_CRR_IN_TREE(crr, *root)); - HARD_CHECK_TREE(NULL, 0, *root, 0); + HARD_CHECK_TREE(NULL, alc->crr_order, *root, 0); + + rbt_insert(alc->crr_order, root, &crr->rbt_node); + + HARD_CHECK_TREE(NULL, alc->crr_order, *root, 0); +} + +void aoff_add_pooled_mbc(Allctr_t *allctr, Carrier_t *crr) +{ + AOFF_RBTree_t **root = &allctr->cpool.pooled_tree; + + ASSERT(allctr == crr->cpool.orig_allctr); + HARD_CHECK_TREE(NULL, 0, *root, 0); /* Link carrier in address order tree */ - rbt_insert(AOFF_AOFF, root, &crr->rbt_node); + rbt_insert(FF_AOFF, root, &crr->cpool.pooled); HARD_CHECK_TREE(NULL, 0, *root, 0); } static void aoff_remove_mbc(Allctr_t *allctr, Carrier_t *carrier) { - AOFFAllctr_t *alc = (AOFFAllctr_t *) allctr; - AOFF_Carrier_t *crr = (AOFF_Carrier_t*) carrier; - AOFF_RBTree_t **root = &alc->mbc_root; + AOFF_RBTree_t **root = &((AOFFAllctr_t*)allctr)->mbc_root; + AOFF_Carrier_t *crr = (AOFF_Carrier_t*)carrier; ASSERT(allctr == ERTS_ALC_CARRIER_TO_ALLCTR(carrier)); if (!IS_CRR_IN_TREE(crr,*root)) - return; + return; - HARD_CHECK_TREE(NULL, 0, *root, 0); + HARD_CHECK_TREE(NULL, alc->crr_order, *root, 0); rbt_delete(root, &crr->rbt_node); crr->rbt_node.parent = NULL; @@ -939,9 +1008,27 @@ static void aoff_remove_mbc(Allctr_t *allctr, Carrier_t *carrier) crr->rbt_node.right = NULL; crr->rbt_node.max_sz = crr->rbt_node.hdr.bhdr; - HARD_CHECK_TREE(NULL, 0, *root, 0); + HARD_CHECK_TREE(NULL, alc->crr_order, *root, 0); +} + +void aoff_remove_pooled_mbc(Allctr_t *allctr, Carrier_t *crr) +{ + ASSERT(allctr == crr->cpool.orig_allctr); + + HARD_CHECK_TREE(NULL, 0, allctr->cpool.pooled_tree, 0); + + rbt_delete(&allctr->cpool.pooled_tree, &crr->cpool.pooled); +#ifdef DEBUG + crr->cpool.pooled.parent = NULL; + crr->cpool.pooled.left = NULL; + crr->cpool.pooled.right = NULL; + crr->cpool.pooled.max_sz = 0; +#endif + HARD_CHECK_TREE(NULL, 0, allctr->cpool.pooled_tree, 0); + } + static UWord aoff_largest_fblk_in_mbc(Allctr_t* allctr, Carrier_t* carrier) { AOFF_Carrier_t *crr = (AOFF_Carrier_t*) carrier; @@ -955,17 +1042,17 @@ static UWord aoff_largest_fblk_in_mbc(Allctr_t* allctr, Carrier_t* carrier) * info_options() */ +static const char* flavor_str[2][3] = { + {"ageffcaoff", "ageffcaobf", "ageffcbf"}, + { "aoff", "aoffcaobf", "aoffcbf"} +}; +static Eterm flavor_atoms[2][3]; + static struct { Eterm as; - Eterm aoff; - Eterm aoffcaobf; - Eterm aoffcbf; -#ifdef DEBUG - Eterm end_of_atoms; -#endif } am; -static void ERTS_INLINE atom_init(Eterm *atom, char *name) +static void ERTS_INLINE atom_init(Eterm *atom, const char *name) { *atom = am_atom_put(name, strlen(name)); } @@ -974,28 +1061,16 @@ static void ERTS_INLINE atom_init(Eterm *atom, char *name) static void init_atoms(void) { -#ifdef DEBUG - Eterm *atom; -#endif + int i, j; 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); - AM_INIT(aoffcaobf); - AM_INIT(aoffcbf); -#ifdef DEBUG - for (atom = (Eterm *) &am; atom < &am.end_of_atoms; atom++) { - ASSERT(*atom != THE_NON_VALUE); - } -#endif + for (i = 0; i < 2; i++) + for (j = 0; j < 3; j++) + atom_init(&flavor_atoms[i][j], flavor_str[i][j]); atoms_initialized = 1; } @@ -1021,15 +1096,16 @@ info_options(Allctr_t *allctr, { AOFFAllctr_t* alc = (AOFFAllctr_t*) allctr; Eterm res = THE_NON_VALUE; - const char* flavor_str[3] = {"aoff", "aoffcaobf", "aoffcbf"}; - Eterm flavor_atom[3] = {am.aoff, am.aoffcaobf, am.aoffcbf}; + + ASSERT(alc->crr_order >= 0 && alc->crr_order <= 1); + ASSERT(alc->blk_order >= 1 && alc->blk_order <= 3); if (print_to_p) { erts_print(*print_to_p, print_to_arg, "%sas: %s\n", prefix, - flavor_str[alc->flavor]); + flavor_str[alc->crr_order][alc->blk_order-1]); } if (hpp || szp) { @@ -1039,7 +1115,8 @@ info_options(Allctr_t *allctr, __FILE__, __LINE__);; res = NIL; - add_2tup(hpp, szp, &res, am.as, flavor_atom[alc->flavor]); + add_2tup(hpp, szp, &res, am.as, + flavor_atoms[alc->crr_order][alc->blk_order-1]); } return res; @@ -1057,7 +1134,7 @@ UWord erts_aoffalc_test(UWord op, UWord a1, UWord a2) { switch (op) { - case 0x500: return (UWord) ((AOFFAllctr_t *) a1)->flavor == AOFF_AOBF; + case 0x500: return (UWord) ((AOFFAllctr_t *) a1)->blk_order == FF_AOBF; case 0x501: { AOFF_RBTree_t *node = ((AOFFAllctr_t *) a1)->mbc_root; Uint size = (Uint) a2; @@ -1072,7 +1149,7 @@ erts_aoffalc_test(UWord op, UWord a1, UWord a2) case 0x507: return (UWord) IS_TREE_NODE((AOFF_RBTree_t *) a1); case 0x508: return (UWord) 0; /* IS_BF_ALGO */ case 0x509: return (UWord) ((AOFF_RBTree_t *) a1)->max_sz; - case 0x50a: return (UWord) ((AOFFAllctr_t *) a1)->flavor == AOFF_BF; + case 0x50a: return (UWord) ((AOFFAllctr_t *) a1)->blk_order == FF_BF; case 0x50b: return (UWord) LIST_PREV(a1); default: ASSERT(0); return ~((UWord) 0); } @@ -1085,12 +1162,13 @@ erts_aoffalc_test(UWord op, UWord a1, UWord a2) #ifdef HARD_DEBUG - -static int rbt_assert_is_member(AOFF_RBTree_t* root, AOFF_RBTree_t* node) +static int rbt_is_member(AOFF_RBTree_t* root, AOFF_RBTree_t* node) { while (node != root) { - ASSERT(node->parent); - ASSERT(node->parent->left == node || node->parent->right == node); + if (!node->parent || (node->parent->left != node && + node->parent->right != node)) { + return 0; + } node = node->parent; } return 1; @@ -1132,7 +1210,7 @@ static void print_tree(AOFF_RBTree_t*); */ static AOFF_RBTree_t * -check_tree(Carrier_t* within_crr, enum AOFF_Flavor flavor, AOFF_RBTree_t* root, Uint size) +check_tree(Carrier_t* within_crr, enum AOFFSortOrder order, AOFF_RBTree_t* root, Uint size) { AOFF_RBTree_t *res = NULL; Sint blacks; @@ -1144,7 +1222,8 @@ check_tree(Carrier_t* within_crr, enum AOFF_Flavor flavor, AOFF_RBTree_t* root, #ifdef PRINT_TREE print_tree(root); #endif - ASSERT(within_crr || flavor == AOFF_AOFF); + ASSERT((within_crr && order >= FF_AOFF) || + (!within_crr && order <= FF_AOFF)); if (!root) return res; @@ -1202,7 +1281,7 @@ check_tree(Carrier_t* within_crr, enum AOFF_Flavor flavor, AOFF_RBTree_t* root, ASSERT(((char*)x + AOFF_BLK_SZ(x)) <= ((char*)crr + CARRIER_SZ(crr))); } - if (flavor == AOFF_BF) { + if (order == FF_BF) { AOFF_RBTree_t* y = x; AOFF_RBTree_t* nxt = LIST_NEXT(y); ASSERT(IS_TREE_NODE(x)); @@ -1225,13 +1304,13 @@ check_tree(Carrier_t* within_crr, enum AOFF_Flavor flavor, AOFF_RBTree_t* root, if (x->left) { ASSERT(x->left->parent == x); - ASSERT(cmp_blocks(flavor, x->left, x) < 0); + ASSERT(cmp_blocks(order, x->left, x) < 0); ASSERT(x->left->max_sz <= x->max_sz); } if (x->right) { ASSERT(x->right->parent == x); - ASSERT(cmp_blocks(flavor, x->right, x) > 0); + ASSERT(cmp_blocks(order, x->right, x) > 0); ASSERT(x->right->max_sz <= x->max_sz); } ASSERT(x->max_sz >= AOFF_BLK_SZ(x)); @@ -1240,7 +1319,7 @@ check_tree(Carrier_t* within_crr, enum AOFF_Flavor flavor, AOFF_RBTree_t* root, || x->max_sz == (x->right ? x->right->max_sz : 0)); if (size && AOFF_BLK_SZ(x) >= size) { - if (!res || cmp_blocks(flavor, x, res) < 0) { + if (!res || cmp_blocks(order, x, res) < 0) { res = x; } } diff --git a/erts/emulator/beam/erl_ao_firstfit_alloc.h b/erts/emulator/beam/erl_ao_firstfit_alloc.h index 7349c6ab19..9cf4fc81a8 100644 --- a/erts/emulator/beam/erl_ao_firstfit_alloc.h +++ b/erts/emulator/beam/erl_ao_firstfit_alloc.h @@ -28,14 +28,16 @@ typedef struct AOFFAllctr_t_ AOFFAllctr_t; -enum AOFF_Flavor { - AOFF_AOFF = 0, - AOFF_AOBF = 1, - AOFF_BF = 2 +enum AOFFSortOrder { + FF_AGEFF = 0, + FF_AOFF = 1, + FF_AOBF = 2, + FF_BF = 3 }; typedef struct { - enum AOFF_Flavor flavor; + enum AOFFSortOrder blk_order; + enum AOFFSortOrder crr_order; } AOFFAllctrInit_t; #define ERTS_DEFAULT_AOFF_ALLCTR_INIT {0/*dummy*/} @@ -53,12 +55,12 @@ Allctr_t *erts_aoffalc_start(AOFFAllctr_t *, AOFFAllctrInit_t*, AllctrInit_t *); #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; - enum AOFF_Flavor flavor; + enum AOFFSortOrder blk_order; + enum AOFFSortOrder crr_order; }; UWord erts_aoffalc_test(UWord, UWord, UWord); diff --git a/erts/emulator/beam/erl_bif_binary.c b/erts/emulator/beam/erl_bif_binary.c index 41c2ae08d3..33bc189182 100644 --- a/erts/emulator/beam/erl_bif_binary.c +++ b/erts/emulator/beam/erl_bif_binary.c @@ -61,8 +61,6 @@ static Export binary_longest_prefix_trap_export; static BIF_RETTYPE binary_longest_prefix_trap(BIF_ALIST_3); static Export binary_longest_suffix_trap_export; static BIF_RETTYPE binary_longest_suffix_trap(BIF_ALIST_3); -static Export binary_bin_to_list_trap_export; -static BIF_RETTYPE binary_bin_to_list_trap(BIF_ALIST_3); static Export binary_copy_trap_export; static BIF_RETTYPE binary_copy_trap(BIF_ALIST_2); static Uint max_loop_limit; @@ -86,10 +84,6 @@ void erts_init_bif_binary(void) am_erlang, am_binary_longest_suffix_trap, 3, &binary_longest_suffix_trap); - erts_init_trap_export(&binary_bin_to_list_trap_export, - am_erlang, am_binary_bin_to_list_trap, 3, - &binary_bin_to_list_trap); - erts_init_trap_export(&binary_copy_trap_export, am_erlang, am_binary_copy_trap, 2, &binary_copy_trap); @@ -2440,191 +2434,6 @@ BIF_RETTYPE binary_at_2(BIF_ALIST_2) BIF_ERROR(BIF_P,BADARG); } -#define BIN_TO_LIST_OK 0 -#define BIN_TO_LIST_TRAP 1 -/* No badarg, checked before call */ - -#define BIN_TO_LIST_LOOP_FACTOR 10 - -static int do_bin_to_list(Process *p, byte *bytes, Uint bit_offs, - Uint start, Sint *lenp, Eterm *termp) -{ - Uint reds = get_reds(p, BIN_TO_LIST_LOOP_FACTOR); /* reds can never be 0 */ - Uint len = *lenp; - Uint loops; - Eterm *hp; - Eterm term = *termp; - Uint n; - - ASSERT(reds > 0); - - loops = MIN(reds,len); - - BUMP_REDS(p, loops / BIN_TO_LIST_LOOP_FACTOR); - - hp = HAlloc(p,2*loops); - while (loops--) { - --len; - if (bit_offs) { - n = ((((Uint) bytes[start+len]) << bit_offs) | - (((Uint) bytes[start+len+1]) >> (8-bit_offs))) & 0xFF; - } else { - n = bytes[start+len]; - } - - term = CONS(hp,make_small(n),term); - hp +=2; - } - *termp = term; - *lenp = len; - if (len) { - BUMP_ALL_REDS(p); - return BIN_TO_LIST_TRAP; - } - return BIN_TO_LIST_OK; -} - - -static BIF_RETTYPE do_trap_bin_to_list(Process *p, Eterm binary, - Uint start, Sint len, Eterm sofar) -{ - Eterm *hp; - Eterm blob; - - hp = HAlloc(p,3); - hp[0] = make_pos_bignum_header(2); - hp[1] = start; - hp[2] = (Uint) len; - blob = make_big(hp); - BIF_TRAP3(&binary_bin_to_list_trap_export, p, binary, blob, sofar); -} - -static BIF_RETTYPE binary_bin_to_list_trap(BIF_ALIST_3) -{ - Eterm *ptr; - Uint start; - Sint len; - byte *bytes; - Uint bit_offs; - Uint bit_size; - Eterm res = BIF_ARG_3; - - ptr = big_val(BIF_ARG_2); - start = ptr[1]; - len = (Sint) ptr[2]; - - ERTS_GET_BINARY_BYTES(BIF_ARG_1,bytes,bit_offs,bit_size); - if (do_bin_to_list(BIF_P, bytes, bit_offs, start, &len, &res) == - BIN_TO_LIST_OK) { - BIF_RET(res); - } - return do_trap_bin_to_list(BIF_P,BIF_ARG_1,start,len,res); -} - -static BIF_RETTYPE binary_bin_to_list_common(Process *p, - Eterm bin, - Eterm epos, - Eterm elen) -{ - Uint pos; - Sint len; - size_t sz; - byte *bytes; - Uint bit_offs; - Uint bit_size; - Eterm res = NIL; - - if (is_not_binary(bin)) { - goto badarg; - } - if (!term_to_Uint(epos, &pos)) { - goto badarg; - } - if (!term_to_Sint(elen, &len)) { - goto badarg; - } - if (len < 0) { - Uint lentmp = -(Uint)len; - /* overflow */ - if ((Sint)lentmp < 0) { - goto badarg; - } - len = lentmp; - if (len > pos) { - goto badarg; - } - pos -= len; - } - /* overflow */ - if ((pos + len) < pos || (len > 0 && (pos + len) == pos)) { - goto badarg; - } - sz = binary_size(bin); - - if (pos+len > sz) { - goto badarg; - } - ERTS_GET_BINARY_BYTES(bin,bytes,bit_offs,bit_size); - if (bit_size != 0) { - goto badarg; - } - if(do_bin_to_list(p, bytes, bit_offs, pos, &len, &res) == - BIN_TO_LIST_OK) { - BIF_RET(res); - } - return do_trap_bin_to_list(p,bin,pos,len,res); - - badarg: - BIF_ERROR(p,BADARG); -} - -BIF_RETTYPE binary_bin_to_list_3(BIF_ALIST_3) -{ - return binary_bin_to_list_common(BIF_P,BIF_ARG_1,BIF_ARG_2,BIF_ARG_3); -} - -BIF_RETTYPE binary_bin_to_list_2(BIF_ALIST_2) -{ - Eterm *tp; - - if (is_not_tuple(BIF_ARG_2)) { - goto badarg; - } - tp = tuple_val(BIF_ARG_2); - if (arityval(*tp) != 2) { - goto badarg; - } - return binary_bin_to_list_common(BIF_P,BIF_ARG_1,tp[1],tp[2]); - badarg: - BIF_ERROR(BIF_P,BADARG); -} - -BIF_RETTYPE binary_bin_to_list_1(BIF_ALIST_1) -{ - Uint pos = 0; - Sint len; - byte *bytes; - Uint bit_offs; - Uint bit_size; - Eterm res = NIL; - - if (is_not_binary(BIF_ARG_1)) { - goto badarg; - } - len = binary_size(BIF_ARG_1); - ERTS_GET_BINARY_BYTES(BIF_ARG_1,bytes,bit_offs,bit_size); - if (bit_size != 0) { - goto badarg; - } - if(do_bin_to_list(BIF_P, bytes, bit_offs, pos, &len, &res) == - BIN_TO_LIST_OK) { - BIF_RET(res); - } - return do_trap_bin_to_list(BIF_P,BIF_ARG_1,pos,len,res); - badarg: - BIF_ERROR(BIF_P,BADARG); -} - HIPE_WRAPPER_BIF_DISABLE_GC(binary_list_to_bin, 1) BIF_RETTYPE binary_list_to_bin_1(BIF_ALIST_1) diff --git a/erts/emulator/beam/erl_bif_port.c b/erts/emulator/beam/erl_bif_port.c index 9f0c90ff7b..b184adedee 100644 --- a/erts/emulator/beam/erl_bif_port.c +++ b/erts/emulator/beam/erl_bif_port.c @@ -639,6 +639,27 @@ BIF_RETTYPE port_get_data_1(BIF_ALIST_1) BIF_RET(res); } +Eterm erts_port_data_read(Port* prt) +{ + Eterm res; + erts_aint_t data; + + data = erts_atomic_read_ddrb(&prt->data); + if (data == (erts_aint_t)NULL) + return am_undefined; /* Port terminated by racing thread */ + + if ((data & 0x3) != 0) { + res = (Eterm) (UWord) data; + ASSERT(is_immed(res)); + } + else { + ErtsPortDataHeap *pdhp = (ErtsPortDataHeap *) data; + res = pdhp->data; + } + return res; +} + + /* * Open a port. Most of the work is not done here but rather in * the file io.c. diff --git a/erts/emulator/beam/erl_db.c b/erts/emulator/beam/erl_db.c index 3ba0886464..1ab1c4a363 100644 --- a/erts/emulator/beam/erl_db.c +++ b/erts/emulator/beam/erl_db.c @@ -1753,6 +1753,28 @@ BIF_RETTYPE ets_new_2(BIF_ALIST_2) BIF_RET(ret); } +/* +** Retrieves the tid() of a named ets table. +*/ +BIF_RETTYPE ets_whereis_1(BIF_ALIST_1) +{ + DbTable* tb; + Eterm res; + + if (is_not_atom(BIF_ARG_1)) { + BIF_ERROR(BIF_P, BADARG); + } + + if ((tb = db_get_table(BIF_P, BIF_ARG_1, DB_INFO, LCK_READ)) == NULL) { + BIF_RET(am_undefined); + } + + res = make_tid(BIF_P, tb); + db_unlock(tb, LCK_READ); + + BIF_RET(res); +} + /* ** The lookup BIF */ @@ -3126,7 +3148,8 @@ BIF_RETTYPE ets_info_1(BIF_ALIST_1) static Eterm fields[] = {am_protection, am_keypos, am_type, am_named_table, am_node, am_size, am_name, am_heir, am_owner, am_memory, am_compressed, am_write_concurrency, - am_read_concurrency}; + am_read_concurrency, + am_id}; Eterm results[sizeof(fields)/sizeof(Eterm)]; DbTable* tb; Eterm res; @@ -4016,7 +4039,10 @@ static Eterm table_info(Process* p, DbTable* tb, Eterm What) ret = is_table_named(tb) ? am_true : am_false; } else if (What == am_compressed) { ret = tb->common.compress ? am_true : am_false; + } else if (What == am_id) { + ret = make_tid(p, tb); } + /* * For debugging purposes */ diff --git a/erts/emulator/beam/erl_gc.c b/erts/emulator/beam/erl_gc.c index 0da4468f9c..1c64644efc 100644 --- a/erts/emulator/beam/erl_gc.c +++ b/erts/emulator/beam/erl_gc.c @@ -116,6 +116,7 @@ typedef struct { static Uint setup_rootset(Process*, Eterm*, int, Rootset*); static void cleanup_rootset(Rootset *rootset); static Eterm *full_sweep_heaps(Process *p, + ErlHeapFragment *live_hf_end, int hibernate, Eterm *n_heap, Eterm* n_htop, char *oh, Uint oh_size, @@ -142,7 +143,7 @@ static Eterm* sweep_literal_area(Eterm* n_hp, Eterm* n_htop, static Eterm* sweep_literals_to_old_heap(Eterm* heap_ptr, Eterm* heap_end, Eterm* htop, char* src, Uint src_size); static Eterm* collect_live_heap_frags(Process* p, ErlHeapFragment *live_hf_end, - Eterm* heap, Eterm* htop, Eterm* objv, int nobj); + Eterm* htop); static int adjust_after_fullsweep(Process *p, 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); @@ -917,6 +918,7 @@ garbage_collect_hibernate(Process* p, int check_long_gc) htop = heap; htop = full_sweep_heaps(p, + ERTS_INVALID_HFRAG_PTR, 1, heap, htop, @@ -1161,7 +1163,7 @@ erts_garbage_collect_literals(Process* p, Eterm* literals, roots++; - while (g_sz--) { + for ( ; g_sz--; g_ptr++) { Eterm gval = *g_ptr; switch (primary_tag(gval)) { @@ -1170,26 +1172,21 @@ erts_garbage_collect_literals(Process* p, Eterm* literals, val = *ptr; if (IS_MOVED_BOXED(val)) { ASSERT(is_boxed(val)); - *g_ptr++ = val; + *g_ptr = val; } else if (ErtsInArea(ptr, area, area_size)) { - move_boxed(&ptr,val,&old_htop,g_ptr++); - } else { - g_ptr++; + move_boxed(ptr,val,&old_htop,g_ptr); } break; case TAG_PRIMARY_LIST: ptr = list_val(gval); val = *ptr; if (IS_MOVED_CONS(val)) { /* Moved */ - *g_ptr++ = ptr[1]; + *g_ptr = ptr[1]; } else if (ErtsInArea(ptr, area, area_size)) { - move_cons(&ptr,val,&old_htop,g_ptr++); - } else { - g_ptr++; - } + move_cons(ptr,val,&old_htop,g_ptr); + } break; default: - g_ptr++; break; } } @@ -1479,25 +1476,29 @@ do_minor(Process *p, ErlHeapFragment *live_hf_end, n_htop = n_heap = (Eterm*) ERTS_HEAP_ALLOC(ERTS_ALC_T_HEAP, sizeof(Eterm)*new_sz); + n = setup_rootset(p, objv, nobj, &rootset); + roots = rootset.roots; + + /* + * All allocations done. Start defile heap with move markers. + * A crash dump due to allocation failure above will see a healthy heap. + */ + if (live_hf_end != ERTS_INVALID_HFRAG_PTR) { /* * Move heap frags that we know are completely live * directly into the new young heap generation. */ - n_htop = collect_live_heap_frags(p, live_hf_end, n_heap, n_htop, - objv, nobj); + n_htop = collect_live_heap_frags(p, live_hf_end, n_htop); } - n = setup_rootset(p, objv, nobj, &rootset); - roots = rootset.roots; - GENSWEEP_NSTACK(p, old_htop, n_htop); while (n--) { Eterm* g_ptr = roots->v; Uint g_sz = roots->sz; roots++; - while (g_sz--) { + for ( ; g_sz--; g_ptr++) { gval = *g_ptr; switch (primary_tag(gval)) { @@ -1507,14 +1508,12 @@ do_minor(Process *p, ErlHeapFragment *live_hf_end, val = *ptr; if (IS_MOVED_BOXED(val)) { ASSERT(is_boxed(val)); - *g_ptr++ = val; + *g_ptr = val; } else if (ErtsInArea(ptr, mature, mature_size)) { - move_boxed(&ptr,val,&old_htop,g_ptr++); + move_boxed(ptr,val,&old_htop,g_ptr); } else if (ErtsInYoungGen(gval, ptr, oh, oh_size)) { - move_boxed(&ptr,val,&n_htop,g_ptr++); - } else { - g_ptr++; - } + move_boxed(ptr,val,&n_htop,g_ptr); + } break; } @@ -1522,19 +1521,15 @@ do_minor(Process *p, ErlHeapFragment *live_hf_end, ptr = list_val(gval); val = *ptr; if (IS_MOVED_CONS(val)) { /* Moved */ - *g_ptr++ = ptr[1]; + *g_ptr = ptr[1]; } else if (ErtsInArea(ptr, mature, mature_size)) { - move_cons(&ptr,val,&old_htop,g_ptr++); + move_cons(ptr,val,&old_htop,g_ptr); } else if (ErtsInYoungGen(gval, ptr, oh, oh_size)) { - move_cons(&ptr,val,&n_htop,g_ptr++); - } else { - g_ptr++; - } + move_cons(ptr,val,&n_htop,g_ptr); + } break; } - default: - g_ptr++; break; } } @@ -1568,9 +1563,9 @@ do_minor(Process *p, ErlHeapFragment *live_hf_end, ASSERT(is_boxed(val)); *n_hp++ = val; } else if (ErtsInArea(ptr, mature, mature_size)) { - move_boxed(&ptr,val,&old_htop,n_hp++); + move_boxed(ptr,val,&old_htop,n_hp++); } else if (ErtsInYoungGen(gval, ptr, oh, oh_size)) { - move_boxed(&ptr,val,&n_htop,n_hp++); + move_boxed(ptr,val,&n_htop,n_hp++); } else { n_hp++; } @@ -1582,9 +1577,9 @@ do_minor(Process *p, ErlHeapFragment *live_hf_end, if (IS_MOVED_CONS(val)) { *n_hp++ = ptr[1]; } else if (ErtsInArea(ptr, mature, mature_size)) { - move_cons(&ptr,val,&old_htop,n_hp++); + move_cons(ptr,val,&old_htop,n_hp++); } else if (ErtsInYoungGen(gval, ptr, oh, oh_size)) { - move_cons(&ptr,val,&n_htop,n_hp++); + move_cons(ptr,val,&n_htop,n_hp++); } else { n_hp++; } @@ -1604,10 +1599,10 @@ do_minor(Process *p, ErlHeapFragment *live_hf_end, *origptr = val; mb->base = binary_bytes(val); } else if (ErtsInArea(ptr, mature, mature_size)) { - move_boxed(&ptr,val,&old_htop,origptr); + move_boxed(ptr,val,&old_htop,origptr); mb->base = binary_bytes(mb->orig); } else if (ErtsInYoungGen(*origptr, ptr, oh, oh_size)) { - move_boxed(&ptr,val,&n_htop,origptr); + move_boxed(ptr,val,&n_htop,origptr); mb->base = binary_bytes(mb->orig); } } @@ -1733,16 +1728,8 @@ major_collection(Process* p, ErlHeapFragment *live_hf_end, n_htop = n_heap = (Eterm *) ERTS_HEAP_ALLOC(ERTS_ALC_T_HEAP, sizeof(Eterm)*new_sz); - if (live_hf_end != ERTS_INVALID_HFRAG_PTR) { - /* - * Move heap frags that we know are completely live - * directly into the heap. - */ - n_htop = collect_live_heap_frags(p, live_hf_end, n_heap, n_htop, - objv, nobj); - } - - n_htop = full_sweep_heaps(p, 0, n_heap, n_htop, oh, oh_size, objv, nobj); + n_htop = full_sweep_heaps(p, live_hf_end, 0, n_heap, n_htop, oh, oh_size, + objv, nobj); /* Move the stack to the end of the heap */ stk_sz = HEAP_END(p) - p->stop; @@ -1789,6 +1776,7 @@ major_collection(Process* p, ErlHeapFragment *live_hf_end, static Eterm * full_sweep_heaps(Process *p, + ErlHeapFragment *live_hf_end, int hibernate, Eterm *n_heap, Eterm* n_htop, char *oh, Uint oh_size, @@ -1805,6 +1793,19 @@ full_sweep_heaps(Process *p, n = setup_rootset(p, objv, nobj, &rootset); + /* + * All allocations done. Start defile heap with move markers. + * A crash dump due to allocation failure above will see a healthy heap. + */ + + if (live_hf_end != ERTS_INVALID_HFRAG_PTR) { + /* + * Move heap frags that we know are completely live + * directly into the heap. + */ + n_htop = collect_live_heap_frags(p, live_hf_end, n_htop); + } + #ifdef HIPE if (hibernate) hipe_empty_nstack(p); @@ -1818,7 +1819,7 @@ full_sweep_heaps(Process *p, Eterm g_sz = roots->sz; roots++; - while (g_sz--) { + for ( ; g_sz--; g_ptr++) { Eterm* ptr; Eterm val; Eterm gval = *g_ptr; @@ -1830,32 +1831,26 @@ full_sweep_heaps(Process *p, val = *ptr; if (IS_MOVED_BOXED(val)) { ASSERT(is_boxed(val)); - *g_ptr++ = val; + *g_ptr = val; } else if (!erts_is_literal(gval, ptr)) { - move_boxed(&ptr,val,&n_htop,g_ptr++); - } else { - g_ptr++; + move_boxed(ptr,val,&n_htop,g_ptr); } - continue; + break; } case TAG_PRIMARY_LIST: { ptr = list_val(gval); val = *ptr; if (IS_MOVED_CONS(val)) { - *g_ptr++ = ptr[1]; + *g_ptr = ptr[1]; } else if (!erts_is_literal(gval, ptr)) { - move_cons(&ptr,val,&n_htop,g_ptr++); - } else { - g_ptr++; + move_cons(ptr,val,&n_htop,g_ptr); } - continue; + break; } - default: { - g_ptr++; - continue; - } + default: + break; } } } @@ -2134,7 +2129,7 @@ sweep(Eterm *n_hp, Eterm *n_htop, ASSERT(is_boxed(val)); *n_hp++ = val; } else if (ERTS_IS_IN_SWEEP_AREA(gval, ptr)) { - move_boxed(&ptr,val,&n_htop,n_hp++); + move_boxed(ptr,val,&n_htop,n_hp++); } else { n_hp++; } @@ -2146,7 +2141,7 @@ sweep(Eterm *n_hp, Eterm *n_htop, if (IS_MOVED_CONS(val)) { *n_hp++ = ptr[1]; } else if (ERTS_IS_IN_SWEEP_AREA(gval, ptr)) { - move_cons(&ptr,val,&n_htop,n_hp++); + move_cons(ptr,val,&n_htop,n_hp++); } else { n_hp++; } @@ -2167,7 +2162,7 @@ sweep(Eterm *n_hp, Eterm *n_htop, *origptr = val; mb->base = binary_bytes(*origptr); } else if (ERTS_IS_IN_SWEEP_AREA(*origptr, ptr)) { - move_boxed(&ptr,val,&n_htop,origptr); + move_boxed(ptr,val,&n_htop,origptr); mb->base = binary_bytes(*origptr); } } @@ -2230,7 +2225,7 @@ sweep_literals_to_old_heap(Eterm* heap_ptr, Eterm* heap_end, Eterm* htop, ASSERT(is_boxed(val)); *heap_ptr++ = val; } else if (ErtsInArea(ptr, src, src_size)) { - move_boxed(&ptr,val,&htop,heap_ptr++); + move_boxed(ptr,val,&htop,heap_ptr++); } else { heap_ptr++; } @@ -2242,7 +2237,7 @@ sweep_literals_to_old_heap(Eterm* heap_ptr, Eterm* heap_end, Eterm* htop, if (IS_MOVED_CONS(val)) { *heap_ptr++ = ptr[1]; } else if (ErtsInArea(ptr, src, src_size)) { - move_cons(&ptr,val,&htop,heap_ptr++); + move_cons(ptr,val,&htop,heap_ptr++); } else { heap_ptr++; } @@ -2263,7 +2258,7 @@ sweep_literals_to_old_heap(Eterm* heap_ptr, Eterm* heap_end, Eterm* htop, *origptr = val; mb->base = binary_bytes(*origptr); } else if (ErtsInArea(ptr, src, src_size)) { - move_boxed(&ptr,val,&htop,origptr); + move_boxed(ptr,val,&htop,origptr); mb->base = binary_bytes(*origptr); } } @@ -2296,11 +2291,11 @@ move_one_area(Eterm* n_htop, char* src, Uint src_size) ASSERT(val != ERTS_HOLE_MARKER); if (is_header(val)) { ASSERT(ptr + header_arity(val) < end); - move_boxed(&ptr, val, &n_htop, &dummy_ref); + ptr = move_boxed(ptr, val, &n_htop, &dummy_ref); } else { /* must be a cons cell */ ASSERT(ptr+1 < end); - move_cons(&ptr, val, &n_htop, &dummy_ref); + move_cons(ptr, val, &n_htop, &dummy_ref); ptr += 2; } } @@ -2313,9 +2308,7 @@ move_one_area(Eterm* n_htop, char* src, Uint src_size) */ static Eterm* -collect_live_heap_frags(Process* p, ErlHeapFragment *live_hf_end, - Eterm* n_hstart, Eterm* n_htop, - Eterm* objv, int nobj) +collect_live_heap_frags(Process* p, ErlHeapFragment *live_hf_end, Eterm* n_htop) { ErlHeapFragment* qb; char* frag_begin; @@ -3281,8 +3274,7 @@ reply_gc_info(void *vgcirp) gcireq_free(vgcirp); } -void erts_sub_binary_to_heap_binary(Eterm **pp, Eterm **hpp, Eterm *orig) { - Eterm *ptr = *pp; +Eterm* erts_sub_binary_to_heap_binary(Eterm *ptr, Eterm **hpp, Eterm *orig) { Eterm *htop = *hpp; Eterm gval; ErlSubBin *sb = (ErlSubBin *)ptr; @@ -3310,7 +3302,7 @@ void erts_sub_binary_to_heap_binary(Eterm **pp, Eterm **hpp, Eterm *orig) { htop += heap_bin_size(sb->size); *hpp = htop; - *pp = ptr; + return ptr; } diff --git a/erts/emulator/beam/erl_gc.h b/erts/emulator/beam/erl_gc.h index 6a529b8443..dec0ab1143 100644 --- a/erts/emulator/beam/erl_gc.h +++ b/erts/emulator/beam/erl_gc.h @@ -33,14 +33,15 @@ #define IS_MOVED_BOXED(x) (!is_header((x))) #define IS_MOVED_CONS(x) (is_non_value((x))) -void erts_sub_binary_to_heap_binary(Eterm **pp, Eterm **hpp, Eterm *orig); +Eterm* erts_sub_binary_to_heap_binary(Eterm *ptr, Eterm **hpp, Eterm *orig); -ERTS_GLB_INLINE void move_cons(Eterm **pp, Eterm car, Eterm **hpp, Eterm *orig); +ERTS_GLB_INLINE void move_cons(Eterm *ERTS_RESTRICT ptr, Eterm car, Eterm **hpp, + Eterm *orig); #if ERTS_GLB_INLINE_INCL_FUNC_DEF -ERTS_GLB_INLINE void move_cons(Eterm **pp, Eterm car, Eterm **hpp, Eterm *orig) +ERTS_GLB_INLINE void move_cons(Eterm *ERTS_RESTRICT ptr, Eterm car, Eterm **hpp, + Eterm *orig) { - Eterm *ptr = *pp; - Eterm *htop = *hpp; + Eterm *ERTS_RESTRICT htop = *hpp; Eterm gval; htop[0] = car; /* copy car */ @@ -53,14 +54,15 @@ ERTS_GLB_INLINE void move_cons(Eterm **pp, Eterm car, Eterm **hpp, Eterm *orig) } #endif -ERTS_GLB_INLINE void move_boxed(Eterm **pp, Eterm hdr, Eterm **hpp, Eterm *orig); +ERTS_GLB_INLINE Eterm* move_boxed(Eterm *ERTS_RESTRICT ptr, Eterm hdr, Eterm **hpp, + Eterm *orig); #if ERTS_GLB_INLINE_INCL_FUNC_DEF -ERTS_GLB_INLINE void move_boxed(Eterm **pp, Eterm hdr, Eterm **hpp, Eterm *orig) +ERTS_GLB_INLINE Eterm* move_boxed(Eterm *ERTS_RESTRICT ptr, Eterm hdr, Eterm **hpp, + Eterm *orig) { Eterm gval; Sint nelts; - Eterm *ptr = *pp; - Eterm *htop = *hpp; + Eterm *ERTS_RESTRICT htop = *hpp; ASSERT(is_header(hdr)); nelts = header_arity(hdr); @@ -71,8 +73,7 @@ ERTS_GLB_INLINE void move_boxed(Eterm **pp, Eterm hdr, Eterm **hpp, Eterm *orig) /* convert sub-binary to heap-binary if applicable */ if (sb->bitsize == 0 && sb->bitoffs == 0 && sb->is_writable == 0 && sb->size <= sizeof(Eterm) * 3) { - erts_sub_binary_to_heap_binary(pp, hpp, orig); - return; + return erts_sub_binary_to_heap_binary(ptr, hpp, orig); } } nelts++; @@ -90,7 +91,7 @@ ERTS_GLB_INLINE void move_boxed(Eterm **pp, Eterm hdr, Eterm **hpp, Eterm *orig) while (nelts--) *htop++ = *ptr++; *hpp = htop; - *pp = ptr; + return ptr; } #endif diff --git a/erts/emulator/beam/erl_goodfit_alloc.c b/erts/emulator/beam/erl_goodfit_alloc.c index 50aa41b4d2..a38f6c7daf 100644 --- a/erts/emulator/beam/erl_goodfit_alloc.c +++ b/erts/emulator/beam/erl_goodfit_alloc.c @@ -170,6 +170,7 @@ static void unlink_free_block (Allctr_t *, Block_t *); static void update_last_aux_mbc (Allctr_t *, Carrier_t *); static Eterm info_options (Allctr_t *, char *, fmtfn_t *, void *, Uint **, Uint *); +static int gfalc_try_set_dyn_param(Allctr_t*, Eterm param, Uint value); static void init_atoms (void); #ifdef ERTS_ALLOC_UTIL_HARD_DEBUG @@ -250,6 +251,8 @@ erts_gfalc_start(GFAllctr_t *gfallctr, if (!erts_alcu_start(allctr, init)) return NULL; + allctr->try_set_dyn_param = gfalc_try_set_dyn_param; + if (allctr->min_block_size != MIN_BLK_SZ) return NULL; @@ -584,6 +587,15 @@ info_options(Allctr_t *allctr, return res; } +static int gfalc_try_set_dyn_param(Allctr_t* allctr, Eterm param, Uint value) +{ + if (param == am_sbct) { + /* Cannot change 'sbct' without rearranging buckets */ + return 0; + } + return erts_alcu_try_set_dyn_param(allctr, param, value); +} + /* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *\ * NOTE: erts_gfalc_test() is only supposed to be used for testing. * * * diff --git a/erts/emulator/beam/erl_init.c b/erts/emulator/beam/erl_init.c index 4846ccd2d3..e8048cfdfc 100644 --- a/erts/emulator/beam/erl_init.c +++ b/erts/emulator/beam/erl_init.c @@ -1746,6 +1746,7 @@ erl_start(int argc, char **argv) } else if (has_prefix("ecio", sub_param)) { /* ignore argument, eager check io no longer used */ + arg = get_arg(sub_param+4, argv[i+1], &i); } else if (has_prefix("pp", sub_param)) { arg = get_arg(sub_param+2, argv[i+1], &i); diff --git a/erts/emulator/beam/erl_io_queue.c b/erts/emulator/beam/erl_io_queue.c index c93b5248d9..d779d1031a 100644 --- a/erts/emulator/beam/erl_io_queue.c +++ b/erts/emulator/beam/erl_io_queue.c @@ -801,12 +801,11 @@ static Eterm iol2v_make_sub_bin(iol2v_state_t *state, Eterm bin_term, ERTS_GET_REAL_BIN(bin_term, orig_pb_term, byte_offset, bit_offset, bit_size); - (void)bit_offset; - (void)bit_size; + ASSERT(bit_size == 0); sb->thing_word = HEADER_SUB_BIN; + sb->bitoffs = bit_offset; sb->bitsize = 0; - sb->bitoffs = 0; sb->orig = orig_pb_term; sb->is_writable = 0; @@ -975,7 +974,7 @@ static int iol2v_append_binary(iol2v_state_t *state, Eterm bin_term) { parent_header = binary_val(parent_binary); binary_size = binary_size(bin_term); - if (bit_offset != 0 || bit_size != 0) { + if (bit_size != 0) { return 0; } else if (binary_size == 0) { state->bytereds_spent += 1; @@ -1017,8 +1016,16 @@ static int iol2v_append_binary(iol2v_state_t *state, Eterm bin_term) { * then just copy it into the accumulator. */ iol2v_expand_acc(state, binary_size); - sys_memcpy(&(state->acc)->orig_bytes[state->acc_size], - binary_data, binary_size); + if (ERTS_LIKELY(bit_offset == 0)) { + sys_memcpy(&(state->acc)->orig_bytes[state->acc_size], + binary_data, binary_size); + } else { + ASSERT(binary_size <= ERTS_UWORD_MAX / 8); + + erts_copy_bits(binary_data, bit_offset, 1, + (byte*)&(state->acc)->orig_bytes[state->acc_size], 0, 1, + binary_size * 8); + } state->acc_size += binary_size; } else { @@ -1029,8 +1036,16 @@ static int iol2v_append_binary(iol2v_state_t *state, Eterm bin_term) { iol2v_expand_acc(state, spill); - sys_memcpy(&(state->acc)->orig_bytes[state->acc_size], - binary_data, spill); + if (ERTS_LIKELY(bit_offset == 0)) { + sys_memcpy(&(state->acc)->orig_bytes[state->acc_size], + binary_data, spill); + } else { + ASSERT(binary_size <= ERTS_UWORD_MAX / 8); + + erts_copy_bits(binary_data, bit_offset, 1, + (byte*)&(state->acc)->orig_bytes[state->acc_size], 0, 1, + spill * 8); + } state->acc_size += spill; @@ -1179,7 +1194,10 @@ BIF_RETTYPE iolist_to_iovec_1(BIF_ALIST_1) { if (is_nil(BIF_ARG_1)) { BIF_RET(NIL); } else if (is_binary(BIF_ARG_1)) { - if (binary_size(BIF_ARG_1) != 0) { + if (binary_bitsize(BIF_ARG_1) != 0) { + ASSERT(!(BIF_P->flags & F_DISABLE_GC)); + BIF_ERROR(BIF_P, BADARG); + } else if (binary_size(BIF_ARG_1) != 0) { Eterm *hp = HAlloc(BIF_P, 2); BIF_RET(CONS(hp, BIF_ARG_1, NIL)); diff --git a/erts/emulator/beam/erl_msacc.h b/erts/emulator/beam/erl_msacc.h index 2588dec903..895b1ae319 100644 --- a/erts/emulator/beam/erl_msacc.h +++ b/erts/emulator/beam/erl_msacc.h @@ -159,12 +159,12 @@ struct erl_msacc_t_ { #ifdef ERTS_ENABLE_MSACC -extern erts_tsd_key_t erts_msacc_key; +extern erts_tsd_key_t ERTS_WRITE_UNLIKELY(erts_msacc_key); #ifdef ERTS_MSACC_ALWAYS_ON #define erts_msacc_enabled 1 #else -extern int erts_msacc_enabled; +extern int ERTS_WRITE_UNLIKELY(erts_msacc_enabled); #endif #define ERTS_MSACC_TSD_GET() erts_tsd_get(erts_msacc_key) diff --git a/erts/emulator/beam/erl_nif.c b/erts/emulator/beam/erl_nif.c index 94ebf88b56..d2000aa71e 100644 --- a/erts/emulator/beam/erl_nif.c +++ b/erts/emulator/beam/erl_nif.c @@ -237,9 +237,11 @@ static void cache_env(ErlNifEnv* env); static void full_flush_env(ErlNifEnv *env); static void flush_env(ErlNifEnv* env); -/* Temporary object header, auto-deallocated when NIF returns - * or when independent environment is cleared. - */ +/* Temporary object header, auto-deallocated when NIF returns or when + * independent environment is cleared. + * + * The payload can be accessed with &tmp_obj_ptr[1] but keep in mind that its + * first element must not require greater alignment than `next`. */ struct enif_tmp_obj_t { struct enif_tmp_obj_t* next; void (*dtor)(struct enif_tmp_obj_t*); @@ -256,6 +258,46 @@ static ERTS_INLINE void free_tmp_objs(ErlNifEnv* env) } } +/* Whether the given environment is bound to a process and will be cleaned up + * when the NIF returns. It's safe to use temp_alloc for objects in + * env->tmp_obj_list when this is true. */ +static ERTS_INLINE int is_proc_bound(ErlNifEnv *env) +{ + return env->mod_nif != NULL; +} + +/* Allocates and attaches an object to the given environment, running its + * destructor when the environment is cleared. To avoid temporary variables the + * address of the allocated object is returned instead of the enif_tmp_obj_t. + * + * The destructor *must* call `erts_free(tmp_obj->allocator, tmp_obj)` to free + * the object. If the destructor needs to refer to the allocated object its + * address will be &tmp_obj[1]. */ +static ERTS_INLINE void *alloc_tmp_obj(ErlNifEnv *env, size_t size, + void (*dtor)(struct enif_tmp_obj_t*)) { + struct enif_tmp_obj_t *tmp_obj; + ErtsAlcType_t allocator; + + allocator = is_proc_bound(env) ? ERTS_ALC_T_TMP : ERTS_ALC_T_NIF; + + tmp_obj = erts_alloc(allocator, sizeof(struct enif_tmp_obj_t) + MAX(1, size)); + + tmp_obj->next = env->tmp_obj_list; + tmp_obj->allocator = allocator; + tmp_obj->dtor = dtor; + + env->tmp_obj_list = tmp_obj; + + return (void*)&tmp_obj[1]; +} + +/* Generic destructor for objects allocated through alloc_tmp_obj that don't + * care about their payload. */ +static void tmp_alloc_dtor(struct enif_tmp_obj_t *tmp_obj) +{ + erts_free(tmp_obj->allocator, tmp_obj); +} + void erts_post_nif(ErlNifEnv* env) { erts_unblock_fpe(env->fpe_was_unmasked); @@ -446,6 +488,7 @@ static void cache_env(ErlNifEnv* env) env->hp_end = env->heap_frag->mem + env->heap_frag->alloc_size; } } + void* enif_priv_data(ErlNifEnv* env) { return env->mod_nif->priv_data; @@ -1019,11 +1062,6 @@ int enif_is_number(ErlNifEnv* env, ERL_NIF_TERM term) return is_number(term); } -static ERTS_INLINE int is_proc_bound(ErlNifEnv* env) -{ - return env->mod_nif != NULL; -} - static void aligned_binary_dtor(struct enif_tmp_obj_t* obj) { erts_free_aligned_binary_bytes_extra((byte*)obj, obj->allocator); @@ -1058,22 +1096,14 @@ int enif_inspect_binary(ErlNifEnv* env, Eterm bin_term, ErlNifBinary* bin) u.tmp->dtor = &aligned_binary_dtor; env->tmp_obj_list = u.tmp; } - bin->bin_term = bin_term; bin->size = binary_size(bin_term); bin->ref_bin = NULL; ADD_READONLY_CHECK(env, bin->data, bin->size); return 1; } -static void tmp_alloc_dtor(struct enif_tmp_obj_t* obj) -{ - erts_free(obj->allocator, obj); -} - int enif_inspect_iolist_as_binary(ErlNifEnv* env, Eterm term, ErlNifBinary* bin) { - struct enif_tmp_obj_t* tobj; - ErtsAlcType_t allocator; ErlDrvSizeT sz; if (is_binary(term)) { return enif_inspect_binary(env,term,bin); @@ -1081,7 +1111,6 @@ int enif_inspect_iolist_as_binary(ErlNifEnv* env, Eterm term, ErlNifBinary* bin) if (is_nil(term)) { bin->data = (unsigned char*) &bin->data; /* dummy non-NULL */ bin->size = 0; - bin->bin_term = THE_NON_VALUE; bin->ref_bin = NULL; return 1; } @@ -1089,16 +1118,8 @@ int enif_inspect_iolist_as_binary(ErlNifEnv* env, Eterm term, ErlNifBinary* bin) return 0; } - allocator = is_proc_bound(env) ? ERTS_ALC_T_TMP : ERTS_ALC_T_NIF; - tobj = erts_alloc(allocator, sz + sizeof(struct enif_tmp_obj_t)); - tobj->allocator = allocator; - tobj->next = env->tmp_obj_list; - tobj->dtor = &tmp_alloc_dtor; - env->tmp_obj_list = tobj; - - bin->data = (unsigned char*) &tobj[1]; + bin->data = alloc_tmp_obj(env, sz, &tmp_alloc_dtor); bin->size = sz; - bin->bin_term = THE_NON_VALUE; bin->ref_bin = NULL; erts_iolist_to_buf(term, (char*) bin->data, sz); ADD_READONLY_CHECK(env, bin->data, bin->size); @@ -1116,7 +1137,6 @@ int enif_alloc_binary(size_t size, ErlNifBinary* bin) bin->size = size; bin->data = (unsigned char*) refbin->orig_bytes; - bin->bin_term = THE_NON_VALUE; bin->ref_bin = refbin; return 1; } @@ -1150,12 +1170,10 @@ void enif_release_binary(ErlNifBinary* bin) { if (bin->ref_bin != NULL) { Binary* refbin = bin->ref_bin; - ASSERT(bin->bin_term == THE_NON_VALUE); erts_bin_release(refbin); } #ifdef DEBUG bin->data = NULL; - bin->bin_term = THE_NON_VALUE; bin->ref_bin = NULL; #endif } @@ -1312,29 +1330,51 @@ int enif_get_string(ErlNifEnv *env, ERL_NIF_TERM list, char* buf, unsigned len, Eterm enif_make_binary(ErlNifEnv* env, ErlNifBinary* bin) { - if (bin->bin_term != THE_NON_VALUE) { - return bin->bin_term; - } - else if (bin->ref_bin != NULL) { - Binary* bptr = bin->ref_bin; - Eterm bin_term; - - bin_term = erts_build_proc_bin(&MSO(env->proc), - alloc_heap(env, PROC_BIN_SIZE), - bptr); - if (erts_refc_read(&bptr->intern.refc, 1) == 1) { - /* Total ownership transfer */ - bin->ref_bin = NULL; - bin->bin_term = bin_term; - } - return bin_term; - } - else { - flush_env(env); - bin->bin_term = new_binary(env->proc, bin->data, bin->size); - cache_env(env); - return bin->bin_term; + Eterm bin_term; + + if (bin->ref_bin != NULL) { + Binary* binary = bin->ref_bin; + + /* If the binary is smaller than the heap binary limit we'll return a + * heap binary to reduce the number of small refc binaries in the + * system. We can't simply release the refc binary right away however; + * the documentation states that the binary should be considered + * read-only from this point on, which implies that it should still be + * readable. + * + * We could keep it alive until we return by adding it to the temporary + * object list, but that requires an off-heap allocation which is + * potentially quite slow, so we create a dummy ProcBin instead and + * rely on the next minor GC to get rid of it. */ + if (bin->size <= ERL_ONHEAP_BIN_LIMIT) { + ErlHeapBin* hb; + + hb = (ErlHeapBin*)alloc_heap(env, heap_bin_size(bin->size)); + hb->thing_word = header_heap_bin(bin->size); + hb->size = bin->size; + + sys_memcpy(hb->data, bin->data, bin->size); + + erts_build_proc_bin(&MSO(env->proc), + alloc_heap(env, PROC_BIN_SIZE), + binary); + + bin_term = make_binary(hb); + } else { + bin_term = erts_build_proc_bin(&MSO(env->proc), + alloc_heap(env, PROC_BIN_SIZE), + binary); + } + + /* Our (possibly shared) ownership has been transferred to the term. */ + bin->ref_bin = NULL; + } else { + flush_env(env); + bin_term = new_binary(env->proc, bin->data, bin->size); + cache_env(env); } + + return bin_term; } Eterm enif_make_sub_binary(ErlNifEnv* env, ERL_NIF_TERM bin_term, @@ -3284,8 +3324,8 @@ typedef struct { Eterm sublist_start; Eterm sublist_end; - UWord offheap_size; - UWord onheap_size; + UWord referenced_size; + UWord copied_size; UWord iovec_len; } iovec_slice_t; @@ -3295,16 +3335,16 @@ static int examine_iovec_term(Eterm list, UWord max_length, iovec_slice_t *resul result->sublist_start = list; result->sublist_length = 0; - result->offheap_size = 0; - result->onheap_size = 0; + result->referenced_size = 0; + result->copied_size = 0; result->iovec_len = 0; lookahead = result->sublist_start; while (is_list(lookahead)) { - Eterm *binary_header, binary; + UWord byte_size; + Eterm binary; Eterm *cell; - UWord size; cell = list_val(lookahead); binary = CAR(cell); @@ -3313,35 +3353,36 @@ static int examine_iovec_term(Eterm list, UWord max_length, iovec_slice_t *resul return 0; } - size = binary_size(binary); - binary_header = binary_val(binary); + byte_size = binary_size(binary); + + if (byte_size > 0) { + int bit_offset, bit_size; + Eterm parent_binary; + UWord byte_offset; - if (size > 0) { - /* If we're a sub-binary we'll need to check our underlying binary - * to determine whether we're on-heap or not. */ - if (thing_subtag(*binary_header) == SUB_BINARY_SUBTAG) { - ErlSubBin *sb = (ErlSubBin*)binary_header; + int requires_copying; - /* Reject bitstrings */ - if((sb->bitoffs + sb->bitsize) > 0) { - return 0; - } + ERTS_GET_REAL_BIN(binary, parent_binary, byte_offset, + bit_offset, bit_size); - ASSERT(size <= binary_size(sb->orig)); - binary_header = binary_val(sb->orig); + (void)byte_offset; + + if (bit_size != 0) { + return 0; } - if (thing_subtag(*binary_header) == HEAP_BINARY_SUBTAG) { - ASSERT(size <= ERL_ONHEAP_BIN_LIMIT); + /* If we're unaligned or an on-heap binary we'll need to copy + * ourselves over to a temporary buffer. */ + requires_copying = (bit_offset != 0) || + thing_subtag(*binary_val(parent_binary)) == HEAP_BINARY_SUBTAG; - result->iovec_len += 1; - result->onheap_size += size; + if (requires_copying) { + result->copied_size += byte_size; } else { - ASSERT(thing_subtag(*binary_header) == REFC_BINARY_SUBTAG); - - result->iovec_len += 1 + size / MAX_SYSIOVEC_IOVLEN; - result->offheap_size += size; + result->referenced_size += byte_size; } + + result->iovec_len += 1 + byte_size / MAX_SYSIOVEC_IOVLEN; } result->sublist_length += 1; @@ -3361,7 +3402,9 @@ static int examine_iovec_term(Eterm list, UWord max_length, iovec_slice_t *resul return 1; } -static void inspect_raw_binary_data(Eterm binary, ErlNifBinary *result) { +static void marshal_iovec_binary(Eterm binary, ErlNifBinary *copy_buffer, + UWord *copy_offset, ErlNifBinary *result) { + Eterm *parent_header; Eterm parent_binary; @@ -3372,10 +3415,11 @@ static void inspect_raw_binary_data(Eterm binary, ErlNifBinary *result) { ERTS_GET_REAL_BIN(binary, parent_binary, byte_offset, bit_offset, bit_size); + ASSERT(bit_size == 0); + parent_header = binary_val(parent_binary); result->size = binary_size(binary); - result->bin_term = binary; if (thing_subtag(*parent_header) == REFC_BINARY_SUBTAG) { ProcBin *pb = (ProcBin*)parent_header; @@ -3398,24 +3442,48 @@ static void inspect_raw_binary_data(Eterm binary, ErlNifBinary *result) { result->data = &((unsigned char*)&hb->data)[byte_offset]; result->ref_bin = NULL; } + + /* If this isn't an *aligned* refc binary, copy its contents to the buffer + * and reference that instead. */ + + if (result->ref_bin == NULL || bit_offset != 0) { + ASSERT(copy_buffer->ref_bin != NULL && copy_buffer->data != NULL); + ASSERT(result->size <= (copy_buffer->size - *copy_offset)); + + if (bit_offset == 0) { + sys_memcpy(©_buffer->data[*copy_offset], + result->data, result->size); + } else { + erts_copy_bits(result->data, bit_offset, 1, + (byte*)©_buffer->data[*copy_offset], 0, 1, + result->size * 8); + } + + result->data = ©_buffer->data[*copy_offset]; + result->ref_bin = copy_buffer->ref_bin; + + *copy_offset += result->size; + } } static int fill_iovec_with_slice(ErlNifEnv *env, iovec_slice_t *slice, ErlNifIOVec *iovec) { - UWord onheap_offset, iovec_idx; - ErlNifBinary onheap_data; + ErlNifBinary copy_buffer = {0}; + UWord copy_offset, iovec_idx; Eterm sublist_iterator; - /* Set up a common refc binary for all on-heap binaries. */ - if (slice->onheap_size > 0) { - if (!enif_alloc_binary(slice->onheap_size, &onheap_data)) { + /* Set up a common refc binary for all on-heap and unaligned binaries. */ + if (slice->copied_size > 0) { + if (!enif_alloc_binary(slice->copied_size, ©_buffer)) { return 0; } + + ASSERT(copy_buffer.ref_bin != NULL); } sublist_iterator = slice->sublist_start; - onheap_offset = 0; + copy_offset = 0; iovec_idx = 0; while (sublist_iterator != slice->sublist_end) { @@ -3423,20 +3491,7 @@ static int fill_iovec_with_slice(ErlNifEnv *env, Eterm *cell; cell = list_val(sublist_iterator); - inspect_raw_binary_data(CAR(cell), &raw_data); - - /* If this isn't a refc binary, copy its contents to the onheap buffer - * and reference that instead. */ - if (raw_data.size > 0 && raw_data.ref_bin == NULL) { - ASSERT(onheap_offset < onheap_data.size); - ASSERT(slice->onheap_size > 0); - - sys_memcpy(&onheap_data.data[onheap_offset], - raw_data.data, raw_data.size); - - raw_data.data = &onheap_data.data[onheap_offset]; - raw_data.ref_bin = onheap_data.ref_bin; - } + marshal_iovec_binary(CAR(cell), ©_buffer, ©_offset, &raw_data); while (raw_data.size > 0) { UWord chunk_len = MIN(raw_data.size, MAX_SYSIOVEC_IOVLEN); @@ -3467,16 +3522,18 @@ static int fill_iovec_with_slice(ErlNifEnv *env, erts_refc_inc(&refc_binary->intern.refc, 1); } - if (slice->onheap_size > 0) { + if (slice->copied_size > 0) { /* Transfer ownership to the iovec; we've taken references to it in * the above loop. */ - enif_release_binary(&onheap_data); + enif_release_binary(©_buffer); } } else { - if (slice->onheap_size > 0) { - /* Attach the binary to our environment and let the GC take care of - * it after returning. */ - enif_make_binary(env, &onheap_data); + if (slice->copied_size > 0) { + /* Attach the binary to our environment and let the next minor GC + * get rid of it. This is slightly faster than using the tmp object + * list since it avoids off-heap allocations. */ + erts_build_proc_bin(&MSO(env->proc), + alloc_heap(env, PROC_BIN_SIZE), copy_buffer.ref_bin); } } @@ -3502,19 +3559,14 @@ static int create_iovec_from_slice(ErlNifEnv *env, alloc_size = binv_offset; alloc_size += slice->iovec_len * sizeof(Binary*); - /* If we have an environment we'll attach the allocated data to it. The - * GC will take care of releasing it later on. */ + /* When the user passes an environment, we attach the iovec to it so + * the user won't have to bother managing it (similar to + * enif_inspect_binary). It'll disappear once the environment is + * cleaned up. */ if (env != NULL) { - ErlNifBinary gc_bin; - - if (!enif_alloc_binary(alloc_size, &gc_bin)) { - return 0; - } - - alloc_base = (char*)gc_bin.data; - enif_make_binary(env, &gc_bin); + alloc_base = alloc_tmp_obj(env, alloc_size, &tmp_alloc_dtor); } else { - alloc_base = enif_alloc(alloc_size); + alloc_base = erts_alloc(ERTS_ALC_T_NIF, alloc_size); } iovec = (ErlNifIOVec*)alloc_base; @@ -3523,12 +3575,12 @@ static int create_iovec_from_slice(ErlNifEnv *env, iovec->flags = 0; } - iovec->size = slice->offheap_size + slice->onheap_size; + iovec->size = slice->referenced_size + slice->copied_size; iovec->iovcnt = slice->iovec_len; if(!fill_iovec_with_slice(env, slice, iovec)) { if (env == NULL && !(iovec->flags & ERL_NIF_IOVEC_FLAGS_PREALLOC)) { - enif_free(iovec); + erts_free(ERTS_ALC_T_NIF, iovec); } return 0; @@ -4218,34 +4270,31 @@ static unsigned calc_checksum(unsigned char* ptr, unsigned size); struct readonly_check_t { - struct enif_tmp_obj_t hdr; unsigned char* ptr; unsigned size; unsigned checksum; }; static void add_readonly_check(ErlNifEnv* env, unsigned char* ptr, unsigned sz) { - ErtsAlcType_t allocator = is_proc_bound(env) ? ERTS_ALC_T_TMP : ERTS_ALC_T_NIF; - struct readonly_check_t* obj = erts_alloc(allocator, - sizeof(struct readonly_check_t)); - obj->hdr.allocator = allocator; - obj->hdr.next = env->tmp_obj_list; - env->tmp_obj_list = &obj->hdr; - obj->hdr.dtor = &readonly_check_dtor; + struct readonly_check_t* obj; + + obj = alloc_tmp_obj(env, sizeof(struct readonly_check_t), + &readonly_check_dtor); + obj->ptr = ptr; obj->size = sz; - obj->checksum = calc_checksum(ptr, sz); + obj->checksum = calc_checksum(ptr, sz); } -static void readonly_check_dtor(struct enif_tmp_obj_t* o) +static void readonly_check_dtor(struct enif_tmp_obj_t* tmp_obj) { - struct readonly_check_t* obj = (struct readonly_check_t*) o; - unsigned chksum = calc_checksum(obj->ptr, obj->size); - if (chksum != obj->checksum) { + struct readonly_check_t* ro_check = (struct readonly_check_t*)&tmp_obj[1]; + unsigned chksum = calc_checksum(ro_check->ptr, ro_check->size); + if (chksum != ro_check->checksum) { fprintf(stderr, "\r\nReadonly data written by NIF, checksums differ" - " %x != %x\r\nABORTING\r\n", chksum, obj->checksum); + " %x != %x\r\nABORTING\r\n", chksum, ro_check->checksum); abort(); } - erts_free(obj->hdr.allocator, obj); + erts_free(tmp_obj->allocator, tmp_obj); } static unsigned calc_checksum(unsigned char* ptr, unsigned size) { diff --git a/erts/emulator/beam/erl_nif.h b/erts/emulator/beam/erl_nif.h index 053f7673c4..a99b4db705 100644 --- a/erts/emulator/beam/erl_nif.h +++ b/erts/emulator/beam/erl_nif.h @@ -137,8 +137,9 @@ typedef struct unsigned char* data; /* Internals (avert your eyes) */ - ERL_NIF_TERM bin_term; void* ref_bin; + /* for future additions to be ABI compatible (same struct size) */ + void* __spare__[2]; }ErlNifBinary; #if (defined(__WIN32__) || defined(_WIN32) || defined(_WIN32_)) diff --git a/erts/emulator/beam/erl_port.h b/erts/emulator/beam/erl_port.h index 9117eb1f72..6b90187d60 100644 --- a/erts/emulator/beam/erl_port.h +++ b/erts/emulator/beam/erl_port.h @@ -180,6 +180,7 @@ void erts_init_port_data(Port *); void erts_cleanup_port_data(Port *); Uint erts_port_data_size(Port *); ErlOffHeap *erts_port_data_offheap(Port *); +Eterm erts_port_data_read(Port* prt); #define ERTS_PORT_GET_CONNECTED(PRT) \ ((Eterm) erts_atomic_read_nob(&(PRT)->connected)) diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c index a807d60ec7..3baca6ce79 100644 --- a/erts/emulator/beam/erl_process.c +++ b/erts/emulator/beam/erl_process.c @@ -1538,6 +1538,20 @@ erts_proclist_destroy(ErtsProcList *plp) proclist_destroy(plp); } +void +erts_proclist_dump(fmtfn_t to, void *to_arg, ErtsProcList *plp) +{ + ErtsProcList *first = plp; + + while (plp) { + erts_print(to, to_arg, "%T", plp->pid); + plp = plp->next; + if (plp == first) + break; + } + erts_print(to, to_arg, "\n"); +} + void * erts_psd_set_init(Process *p, int ix, void *data) { @@ -13398,16 +13412,33 @@ stack_element_dump(fmtfn_t to, void *to_arg, Eterm* sp, int yreg) return yreg; } +static void print_current_process_info(fmtfn_t, void *to_arg, ErtsSchedulerData*); + /* * Print scheduler information */ void -erts_print_scheduler_info(fmtfn_t to, void *to_arg, ErtsSchedulerData *esdp) { +erts_print_scheduler_info(fmtfn_t to, void *to_arg, ErtsSchedulerData *esdp) +{ int i; erts_aint32_t flg; - Process *p; - erts_print(to, to_arg, "=scheduler:%u\n", esdp->no); + switch (esdp->type) { + case ERTS_SCHED_NORMAL: + erts_print(to, to_arg, "=scheduler:%u\n", esdp->no); + break; + case ERTS_SCHED_DIRTY_CPU: + erts_print(to, to_arg, "=dirty_cpu_scheduler:%u\n", + (esdp->dirty_no + erts_no_schedulers)); + break; + case ERTS_SCHED_DIRTY_IO: + erts_print(to, to_arg, "=dirty_io_scheduler:%u\n", + (esdp->dirty_no + erts_no_schedulers + erts_no_dirty_cpu_schedulers)); + break; + default: + erts_print(to, to_arg, "=unknown_scheduler_type:%u\n", esdp->type); + break; + } flg = erts_atomic32_read_dirty(&esdp->ssi->flags); erts_print(to, to_arg, "Scheduler Sleep Info Flags: "); @@ -13453,10 +13484,24 @@ erts_print_scheduler_info(fmtfn_t to, void *to_arg, ErtsSchedulerData *esdp) { } erts_print(to, to_arg, "\n"); - erts_print(to, to_arg, "Current Port: "); - if (esdp->current_port) - erts_print(to, to_arg, "%T", esdp->current_port->common.id); - erts_print(to, to_arg, "\n"); + if (esdp->type == ERTS_SCHED_NORMAL) { + erts_print(to, to_arg, "Current Port: "); + if (esdp->current_port) + erts_print(to, to_arg, "%T", esdp->current_port->common.id); + erts_print(to, to_arg, "\n"); + + erts_print_run_queue_info(to, to_arg, esdp->run_queue); + } + + /* This *MUST* to be the last information in scheduler block */ + print_current_process_info(to, to_arg, esdp); +} + +void erts_print_run_queue_info(fmtfn_t to, void *to_arg, + ErtsRunQueue *run_queue) +{ + erts_aint32_t flg; + int i; for (i = 0; i < ERTS_NO_PROC_PRIO_LEVELS; i++) { erts_print(to, to_arg, "Run Queue "); @@ -13478,12 +13523,12 @@ erts_print_scheduler_info(fmtfn_t to, void *to_arg, ErtsSchedulerData *esdp) { break; } erts_print(to, to_arg, "Length: %d\n", - erts_atomic32_read_dirty(&esdp->run_queue->procs.prio_info[i].len)); + erts_atomic32_read_dirty(&run_queue->procs.prio_info[i].len)); } erts_print(to, to_arg, "Run Queue Port Length: %d\n", - erts_atomic32_read_dirty(&esdp->run_queue->ports.info.len)); + erts_atomic32_read_dirty(&run_queue->ports.info.len)); - flg = erts_atomic32_read_dirty(&esdp->run_queue->flags); + flg = erts_atomic32_read_dirty(&run_queue->flags); erts_print(to, to_arg, "Run Queue Flags: "); for (i = 0; i < ERTS_RUNQ_FLG_MAX && flg; i++) { erts_aint32_t chk = (1 << i); @@ -13550,9 +13595,15 @@ erts_print_scheduler_info(fmtfn_t to, void *to_arg, ErtsSchedulerData *esdp) { } } erts_print(to, to_arg, "\n"); +} + + +static void print_current_process_info(fmtfn_t to, void *to_arg, + ErtsSchedulerData* esdp) +{ + Process *p = esdp->current_process; + erts_aint32_t flg; - /* This *MUST* to be the last information in scheduler block */ - p = esdp->current_process; erts_print(to, to_arg, "Current Process: "); if (esdp->current_process && !(ERTS_TRACE_FLAGS(p) & F_SENSITIVE)) { flg = erts_atomic32_read_dirty(&p->state); diff --git a/erts/emulator/beam/erl_process.h b/erts/emulator/beam/erl_process.h index 66d7848f89..4f200b13ec 100644 --- a/erts/emulator/beam/erl_process.h +++ b/erts/emulator/beam/erl_process.h @@ -105,13 +105,13 @@ struct saved_calls { }; extern Export exp_send, exp_receive, exp_timeout; -extern int erts_sched_compact_load; -extern int erts_sched_balance_util; -extern Uint erts_no_schedulers; -extern Uint erts_no_total_schedulers; -extern Uint erts_no_dirty_cpu_schedulers; -extern Uint erts_no_dirty_io_schedulers; -extern Uint erts_no_run_queues; +extern int ERTS_WRITE_UNLIKELY(erts_sched_compact_load); +extern int ERTS_WRITE_UNLIKELY(erts_sched_balance_util); +extern Uint ERTS_WRITE_UNLIKELY(erts_no_schedulers); +extern Uint ERTS_WRITE_UNLIKELY(erts_no_total_schedulers); +extern Uint ERTS_WRITE_UNLIKELY(erts_no_dirty_cpu_schedulers); +extern Uint ERTS_WRITE_UNLIKELY(erts_no_dirty_io_schedulers); +extern Uint ERTS_WRITE_UNLIKELY(erts_no_run_queues); extern int erts_sched_thread_suggested_stack_size; extern int erts_dcpu_sched_thread_suggested_stack_size; extern int erts_dio_sched_thread_suggested_stack_size; @@ -522,7 +522,7 @@ typedef union { char align[ERTS_ALC_CACHE_LINE_ALIGN_SIZE(sizeof(ErtsRunQueue))]; } ErtsAlignedRunQueue; -extern ErtsAlignedRunQueue *erts_aligned_run_queues; +extern ErtsAlignedRunQueue * ERTS_WRITE_UNLIKELY(erts_aligned_run_queues); #define ERTS_PROC_REDUCTIONS_EXECUTED(SD, RQ, PRIO, REDS, AREDS)\ do { \ @@ -675,9 +675,9 @@ typedef union { char align[ERTS_ALC_CACHE_LINE_ALIGN_SIZE(sizeof(ErtsSchedulerData))]; } ErtsAlignedSchedulerData; -extern ErtsAlignedSchedulerData *erts_aligned_scheduler_data; -extern ErtsAlignedSchedulerData *erts_aligned_dirty_cpu_scheduler_data; -extern ErtsAlignedSchedulerData *erts_aligned_dirty_io_scheduler_data; +extern ErtsAlignedSchedulerData * ERTS_WRITE_UNLIKELY(erts_aligned_scheduler_data); +extern ErtsAlignedSchedulerData * ERTS_WRITE_UNLIKELY(erts_aligned_dirty_cpu_scheduler_data); +extern ErtsAlignedSchedulerData * ERTS_WRITE_UNLIKELY(erts_aligned_dirty_io_scheduler_data); #if defined(ERTS_ENABLE_LOCK_CHECK) @@ -1270,7 +1270,7 @@ void erts_check_for_holes(Process* p); #define SPO_OFF_HEAP_MSGQ 16 #define SPO_ON_HEAP_MSGQ 32 -extern int erts_default_spo_flags; +extern int ERTS_WRITE_UNLIKELY(erts_default_spo_flags); /* * The following struct contains options for a process to be spawned. @@ -1326,10 +1326,10 @@ extern erts_rwmtx_t erts_cpu_bind_rwmtx; ** erts_system_monitor must be != NIL, to allow testing on just ** the erts_system_monitor_* variables. */ -extern Eterm erts_system_monitor; -extern Uint erts_system_monitor_long_gc; -extern Uint erts_system_monitor_long_schedule; -extern Uint erts_system_monitor_large_heap; +extern Eterm ERTS_WRITE_UNLIKELY(erts_system_monitor); +extern Uint ERTS_WRITE_UNLIKELY(erts_system_monitor_long_gc); +extern Uint ERTS_WRITE_UNLIKELY(erts_system_monitor_long_schedule); +extern Uint ERTS_WRITE_UNLIKELY(erts_system_monitor_large_heap); struct erts_system_monitor_flags_t { unsigned int busy_port : 1; unsigned int busy_dist_port : 1; @@ -1546,6 +1546,7 @@ Uint64 erts_step_proc_interval(void); ErtsProcList *erts_proclist_create(Process *); ErtsProcList *erts_proclist_copy(ErtsProcList *); void erts_proclist_destroy(ErtsProcList *); +void erts_proclist_dump(fmtfn_t to, void *to_arg, ErtsProcList*); ERTS_GLB_INLINE int erts_proclist_same(ErtsProcList *, Process *); ERTS_GLB_INLINE void erts_proclist_store_first(ErtsProcList **, ErtsProcList *); @@ -1792,6 +1793,7 @@ void erts_stack_dump(fmtfn_t to, void *to_arg, Process *); void erts_limited_stack_trace(fmtfn_t to, void *to_arg, Process *); void erts_program_counter_info(fmtfn_t to, void *to_arg, Process *); void erts_print_scheduler_info(fmtfn_t to, void *to_arg, ErtsSchedulerData *esdp); +void erts_print_run_queue_info(fmtfn_t, void *to_arg, ErtsRunQueue*); void erts_dump_extended_process_state(fmtfn_t to, void *to_arg, erts_aint32_t psflg); void erts_dump_process_state(fmtfn_t to, void *to_arg, erts_aint32_t psflg); diff --git a/erts/emulator/beam/erl_process_dump.c b/erts/emulator/beam/erl_process_dump.c index f562fc961b..0d76a4ef53 100644 --- a/erts/emulator/beam/erl_process_dump.c +++ b/erts/emulator/beam/erl_process_dump.c @@ -78,8 +78,17 @@ erts_deep_process_dump(fmtfn_t to, void *to_arg) Process *p = erts_pix2proc(i); if (p && p->i != ENULL) { erts_aint32_t state = erts_atomic32_read_acqb(&p->state); - if (!(state & (ERTS_PSFLG_EXITING|ERTS_PSFLG_GC))) - dump_process_info(to, to_arg, p); + if (state & ERTS_PSFLG_EXITING) + continue; + if (state & ERTS_PSFLG_GC) { + ErtsSchedulerData *sdp = erts_get_scheduler_data(); + if (!sdp || p != sdp->current_process) + continue; + + /* We want to dump the garbing process that caused the dump */ + } + + dump_process_info(to, to_arg, p); } } @@ -135,9 +144,12 @@ dump_process_info(fmtfn_t to, void *to_arg, Process *p) ErtsMessage* mp; int yreg = -1; + if (ERTS_TRACE_FLAGS(p) & F_SENSITIVE) + return; + ERTS_MSGQ_MV_INQ2PRIVQ(p); - if ((ERTS_TRACE_FLAGS(p) & F_SENSITIVE) == 0 && p->msg.first) { + if (p->msg.first) { erts_print(to, to_arg, "=proc_messages:%T\n", p->common.id); for (mp = p->msg.first; mp != NULL; mp = mp->next) { Eterm mesg = ERL_MESSAGE_TERM(mp); @@ -152,38 +164,34 @@ dump_process_info(fmtfn_t to, void *to_arg, Process *p) } } - if ((ERTS_TRACE_FLAGS(p) & F_SENSITIVE) == 0) { - if (p->dictionary) { - erts_print(to, to_arg, "=proc_dictionary:%T\n", p->common.id); - erts_deep_dictionary_dump(to, to_arg, - p->dictionary, dump_element_nl); - } + if (p->dictionary) { + erts_print(to, to_arg, "=proc_dictionary:%T\n", p->common.id); + erts_deep_dictionary_dump(to, to_arg, + p->dictionary, dump_element_nl); } - if ((ERTS_TRACE_FLAGS(p) & F_SENSITIVE) == 0) { - erts_print(to, to_arg, "=proc_stack:%T\n", p->common.id); - for (sp = p->stop; sp < STACK_START(p); sp++) { - yreg = stack_element_dump(to, to_arg, sp, yreg); - } + erts_print(to, to_arg, "=proc_stack:%T\n", p->common.id); + for (sp = p->stop; sp < STACK_START(p); sp++) { + yreg = stack_element_dump(to, to_arg, sp, yreg); + } - erts_print(to, to_arg, "=proc_heap:%T\n", p->common.id); - for (sp = p->stop; sp < STACK_START(p); sp++) { - Eterm term = *sp; - - if (!is_catch(term) && !is_CP(term)) { - heap_dump(to, to_arg, term); - } - } - for (mp = p->msg.first; mp != NULL; mp = mp->next) { - Eterm mesg = ERL_MESSAGE_TERM(mp); - if (is_value(mesg)) - heap_dump(to, to_arg, mesg); - mesg = ERL_MESSAGE_TOKEN(mp); - heap_dump(to, to_arg, mesg); - } - if (p->dictionary) { - erts_deep_dictionary_dump(to, to_arg, p->dictionary, heap_dump); - } + erts_print(to, to_arg, "=proc_heap:%T\n", p->common.id); + for (sp = p->stop; sp < STACK_START(p); sp++) { + Eterm term = *sp; + + if (!is_catch(term) && !is_CP(term)) { + heap_dump(to, to_arg, term); + } + } + for (mp = p->msg.first; mp != NULL; mp = mp->next) { + Eterm mesg = ERL_MESSAGE_TERM(mp); + if (is_value(mesg)) + heap_dump(to, to_arg, mesg); + mesg = ERL_MESSAGE_TOKEN(mp); + heap_dump(to, to_arg, mesg); + } + if (p->dictionary) { + erts_deep_dictionary_dump(to, to_arg, p->dictionary, heap_dump); } } diff --git a/erts/emulator/beam/erl_trace.c b/erts/emulator/beam/erl_trace.c index 4b996d8fc2..d2495479ab 100644 --- a/erts/emulator/beam/erl_trace.c +++ b/erts/emulator/beam/erl_trace.c @@ -2783,24 +2783,28 @@ is_tracer_enabled(Process* c_p, ErtsProcLocks c_p_locks, ASSERT(0); } - /* Only remove tracer on self() and ports */ + /* Only remove tracer on (self() or ports) AND we are on a normal scheduler */ if (is_internal_port(t_p->id) || (c_p && c_p->common.id == t_p->id)) { + ErtsSchedulerData *esdp = erts_get_scheduler_data(); ErtsProcLocks c_p_xlocks = 0; - if (is_internal_pid(t_p->id)) { - ERTS_LC_ASSERT(erts_proc_lc_my_proc_locks(c_p) & ERTS_PROC_LOCK_MAIN); - if (c_p_locks != ERTS_PROC_LOCKS_ALL) { - c_p_xlocks = ~c_p_locks & ERTS_PROC_LOCKS_ALL; - if (erts_proc_trylock(c_p, c_p_xlocks) == EBUSY) { - erts_proc_unlock(c_p, c_p_locks & ~ERTS_PROC_LOCK_MAIN); - erts_proc_lock(c_p, ERTS_PROC_LOCKS_ALL_MINOR); + if (esdp && !ERTS_SCHEDULER_IS_DIRTY(esdp)) { + if (is_internal_pid(t_p->id)) { + ERTS_LC_ASSERT(erts_proc_lc_my_proc_locks(c_p) & ERTS_PROC_LOCK_MAIN); + if (c_p_locks != ERTS_PROC_LOCKS_ALL) { + c_p_xlocks = ~c_p_locks & ERTS_PROC_LOCKS_ALL; + if (erts_proc_trylock(c_p, c_p_xlocks) == EBUSY) { + erts_proc_unlock(c_p, c_p_locks & ~ERTS_PROC_LOCK_MAIN); + erts_proc_lock(c_p, ERTS_PROC_LOCKS_ALL_MINOR); + } } } - } - erts_tracer_replace(t_p, erts_tracer_nil); - t_p->trace_flags &= ~TRACEE_FLAGS; - if (c_p_xlocks) - erts_proc_unlock(c_p, c_p_xlocks); + erts_tracer_replace(t_p, erts_tracer_nil); + t_p->trace_flags &= ~TRACEE_FLAGS; + + if (c_p_xlocks) + erts_proc_unlock(c_p, c_p_xlocks); + } } return 0; diff --git a/erts/emulator/beam/global.h b/erts/emulator/beam/global.h index 86e2c351af..d757651933 100644 --- a/erts/emulator/beam/global.h +++ b/erts/emulator/beam/global.h @@ -948,8 +948,8 @@ void erts_update_ranges(BeamInstr* code, Uint size); void erts_remove_from_ranges(BeamInstr* code); UWord erts_ranges_sz(void); void erts_lookup_function_info(FunctionInfo* fi, BeamInstr* pc, int full_info); -ErtsLiteralArea** erts_dump_lit_areas; -Uint erts_dump_num_lit_areas; +extern ErtsLiteralArea** erts_dump_lit_areas; +extern Uint erts_dump_num_lit_areas; /* break.c */ void init_break_handler(void); @@ -1066,7 +1066,7 @@ Eterm copy_struct_x(Eterm, Uint, Eterm**, ErlOffHeap*, Uint*, erts_literal_area_ #define copy_struct_litopt(Obj,Sz,HPP,OH,LitArea) \ copy_struct_x(Obj,Sz,HPP,OH,NULL,LitArea) -Eterm copy_shallow(Eterm*, Uint, Eterm**, ErlOffHeap*); +Eterm copy_shallow(Eterm* ERTS_RESTRICT, Uint, Eterm**, ErlOffHeap*); void erts_move_multi_frags(Eterm** hpp, ErlOffHeap*, ErlHeapFragment* first, Eterm* refs, unsigned nrefs, int literals); diff --git a/erts/emulator/beam/instrs.tab b/erts/emulator/beam/instrs.tab index f19027b1ec..6a531fcc09 100644 --- a/erts/emulator/beam/instrs.tab +++ b/erts/emulator/beam/instrs.tab @@ -322,6 +322,16 @@ get_list(Src, Hd, Tl) { $Tl = tl; } +get_hd(Src, Hd) { + Eterm* tmp_ptr = list_val($Src); + $Hd = CAR(tmp_ptr); +} + +get_tl(Src, Tl) { + Eterm* tmp_ptr = list_val($Src); + $Tl = CDR(tmp_ptr); +} + i_get(Src, Dst) { $Dst = erts_pd_hash_get(c_p, $Src); } diff --git a/erts/emulator/beam/io.c b/erts/emulator/beam/io.c index 2c1b7871c4..4bb2c0cb01 100644 --- a/erts/emulator/beam/io.c +++ b/erts/emulator/beam/io.c @@ -5075,6 +5075,93 @@ static void prt_one_lnk(ErtsLink *lnk, void *vprtd) erts_print(prtd->to, prtd->arg, "%T", lnk->pid); } +static void dump_port_state(fmtfn_t to, void *arg, erts_aint32_t state) +{ + erts_aint32_t rest; + int unknown = 0; + char delim = ' '; + + erts_print(to, arg, "State:"); + + rest = state; + while (rest) { + erts_aint32_t chk = (rest ^ (rest-1)) & rest; /* lowest set bit */ + char* s; + + rest &= ~chk; + switch (chk) { + case ERTS_PORT_SFLG_CONNECTED: s = "CONNECTED"; break; + case ERTS_PORT_SFLG_EXITING: s = "EXITING"; break; + case ERTS_PORT_SFLG_DISTRIBUTION: s = "DISTR"; break; + case ERTS_PORT_SFLG_BINARY_IO: s = "BINARY_IO"; break; + case ERTS_PORT_SFLG_SOFT_EOF: s = "SOFT_EOF"; break; + case ERTS_PORT_SFLG_CLOSING: s = "CLOSING"; break; + case ERTS_PORT_SFLG_SEND_CLOSED: s = "SEND_CLOSED"; break; + case ERTS_PORT_SFLG_LINEBUF_IO: s = "LINEBUF_IO"; break; + case ERTS_PORT_SFLG_FREE: s = "FREE"; break; + case ERTS_PORT_SFLG_INITIALIZING: s = "INITIALIZING"; break; + case ERTS_PORT_SFLG_PORT_SPECIFIC_LOCK: s = "PORT_LOCK"; break; + case ERTS_PORT_SFLG_INVALID: s = "INVALID"; break; + case ERTS_PORT_SFLG_HALT: s = "HALT"; break; +#ifdef DEBUG + case ERTS_PORT_SFLG_PORT_DEBUG: s = "DEBUG"; break; +#endif + default: + unknown = 1; + continue; + } + erts_print(to, arg, "%c%s", delim, s); + delim = '|'; + } + if (unknown || !state) + erts_print(to, arg, "%c0x%x\n", delim, state); + else + erts_print(to, arg, "\n"); +} + +static void dump_port_task_flags(fmtfn_t to, void *arg, Port* p) +{ + erts_aint32_t flags = erts_atomic32_read_nob(&p->sched.flags); + erts_aint32_t unknown = 0; + char delim = ' '; + + if (!flags) + return; + + erts_print(to, arg, "Task Flags:"); + + while (flags) { + erts_aint32_t chk = (flags ^ (flags-1)) & flags; /* lowest set bit */ + char* s; + + flags &= ~chk; + switch (chk) { + case ERTS_PTS_FLG_IN_RUNQ: s = "IN_RUNQ"; break; + case ERTS_PTS_FLG_EXEC: s = "EXEC"; break; + case ERTS_PTS_FLG_HAVE_TASKS: s = "HAVE_TASKS"; break; + case ERTS_PTS_FLG_EXIT: s = "EXIT"; break; + case ERTS_PTS_FLG_BUSY_PORT: s = "BUSY_PORT"; break; + case ERTS_PTS_FLG_BUSY_PORT_Q: s = "BUSY_Q"; break; + case ERTS_PTS_FLG_CHK_UNSET_BUSY_PORT_Q: s = "CHK_UNSET_BUSY_Q"; break; + case ERTS_PTS_FLG_HAVE_BUSY_TASKS: s = "BUSY_TASKS"; break; + case ERTS_PTS_FLG_HAVE_NS_TASKS: s = "NS_TASKS"; break; + case ERTS_PTS_FLG_PARALLELISM: s = "PARALLELISM"; break; + case ERTS_PTS_FLG_FORCE_SCHED: s = "FORCE_SCHED"; break; + case ERTS_PTS_FLG_EXITING: s = "EXITING"; break; + case ERTS_PTS_FLG_EXEC_IMM: s = "EXEC_IMM"; break; + default: + unknown |= chk; + continue; + } + erts_print(to, arg, "%c%s", delim, s); + delim = '|'; + } + if (unknown) + erts_print(to, arg, "%cUNKNOWN(0x%x)\n", delim, unknown); + else + erts_print(to, arg, "\n"); +} + void print_port_info(Port *p, fmtfn_t to, void *arg) { @@ -5084,6 +5171,8 @@ print_port_info(Port *p, fmtfn_t to, void *arg) return; erts_print(to, arg, "=port:%T\n", p->common.id); + dump_port_state(to, arg, state); + dump_port_task_flags(to, arg, p); erts_print(to, arg, "Slot: %d\n", internal_port_index(p->common.id)); if (state & ERTS_PORT_SFLG_CONNECTED) { erts_print(to, arg, "Connected: %T", ERTS_PORT_GET_CONNECTED(p)); @@ -5106,6 +5195,10 @@ print_port_info(Port *p, fmtfn_t to, void *arg) erts_doforall_monitors(ERTS_P_MONITORS(p), &prt_one_monitor, &prtd); erts_print(to, arg, "\n"); } + if (p->suspended) { + erts_print(to, arg, "Suspended: "); + erts_proclist_dump(to, arg, p->suspended); + } if (p->common.u.alive.reg != NULL) erts_print(to, arg, "Registered as: %T\n", p->common.u.alive.reg->name); @@ -5123,6 +5216,14 @@ print_port_info(Port *p, fmtfn_t to, void *arg) } else { erts_print(to, arg, "Port controls linked-in driver: %s\n",p->name); } + erts_print(to, arg, "Input: %beu\n", p->bytes_in); + erts_print(to, arg, "Output: %beu\n", p->bytes_out); + erts_print(to, arg, "Queue: %beu\n", erts_ioq_size(&p->ioq)); + { + Eterm port_data = erts_port_data_read(p); + if (port_data != am_undefined) + erts_print(to, arg, "Port Data: %T\n", port_data); + } } void diff --git a/erts/emulator/beam/ops.tab b/erts/emulator/beam/ops.tab index 1f4a8eadb0..77e375f2c0 100644 --- a/erts/emulator/beam/ops.tab +++ b/erts/emulator/beam/ops.tab @@ -182,6 +182,9 @@ get_list r x y get_list r y r get_list r x r +get_hd xy xy +get_tl xy xy + # Old-style catch. catch y f catch_end y diff --git a/erts/emulator/beam/utils.c b/erts/emulator/beam/utils.c index fe9f1c7606..4bf60619ba 100644 --- a/erts/emulator/beam/utils.c +++ b/erts/emulator/beam/utils.c @@ -3152,6 +3152,9 @@ tailrecur_ne: int cmp; byte* a_ptr; byte* b_ptr; + if (eq_only && a_size != b_size) { + RETURN_NEQ(a_size - b_size); + } ERTS_GET_BINARY_BYTES(a, a_ptr, a_bitoffs, a_bitsize); ERTS_GET_BINARY_BYTES(b, b_ptr, b_bitoffs, b_bitsize); if ((a_bitsize | b_bitsize | a_bitoffs | b_bitoffs) == 0) { diff --git a/erts/emulator/hipe/hipe_amd64.c b/erts/emulator/hipe/hipe_amd64.c index e3cff4a4ba..f23f341e6d 100644 --- a/erts/emulator/hipe/hipe_amd64.c +++ b/erts/emulator/hipe/hipe_amd64.c @@ -28,6 +28,7 @@ #include "error.h" #include "bif.h" #include "big.h" /* term_to_Sint() */ +#include "erl_binary.h" #include "hipe_arch.h" #include "hipe_bif0.h" @@ -38,6 +39,8 @@ #undef ERL_FUN_SIZE #include "hipe_literals.h" +static void patch_trampoline(void *trampoline, void *destAddress); + const Uint sse2_fnegate_mask[2] = {0x8000000000000000,0}; void hipe_patch_load_fe(Uint64 *address, Uint64 value) @@ -52,9 +55,9 @@ int hipe_patch_insn(void *address, Uint64 value, Eterm type) switch (type) { case am_closure: case am_constant: + case am_c_const: *(Uint64*)address = value; break; - case am_c_const: case am_atom: /* check that value fits in an unsigned imm32 */ /* XXX: are we sure it's not really a signed imm32? */ @@ -71,14 +74,18 @@ int hipe_patch_insn(void *address, Uint64 value, Eterm type) int hipe_patch_call(void *callAddress, void *destAddress, void *trampoline) { - Sint rel32; + Sint64 destOffset = (Sint64)destAddress - (Sint64)callAddress - 4; - ASSERT(trampoline == NULL); + if ((destOffset < -0x80000000L) || (destOffset >= 0x80000000L)) { + destOffset = (Sint64)trampoline - (Sint64)callAddress - 4; - rel32 = (Sint)destAddress - (Sint)callAddress - 4; - if ((Sint)(Sint32)rel32 != rel32) - return -1; - *(Uint32*)callAddress = (Uint32)rel32; + if ((destOffset < -0x80000000L) || (destOffset >= 0x80000000L)) + return -1; + + patch_trampoline(trampoline, destAddress); + } + + *(Uint32*)callAddress = (Uint32)destOffset; hipe_flush_icache_word(callAddress); return 0; } @@ -96,12 +103,80 @@ static void *alloc_code(unsigned int alloc_bytes) return erts_alloc(ERTS_ALC_T_HIPE_EXEC, alloc_bytes); } +static int check_callees(Eterm callees) +{ + Eterm *tuple; + Uint arity; + Uint i; + + if (is_not_tuple(callees)) + return -1; + tuple = tuple_val(callees); + arity = arityval(tuple[0]); + for (i = 1; i <= arity; ++i) { + Eterm mfa = tuple[i]; + if (is_atom(mfa)) + continue; + if (is_not_tuple(mfa) || + tuple_val(mfa)[0] != make_arityval(3) || + is_not_atom(tuple_val(mfa)[1]) || + is_not_atom(tuple_val(mfa)[2]) || + is_not_small(tuple_val(mfa)[3]) || + unsigned_val(tuple_val(mfa)[3]) > 255) + return -1; + } + return arity; +} + +#define TRAMPOLINE_BYTES 12 + +static void generate_trampolines(unsigned char *address, + int nrcallees, Eterm callees, + unsigned char **trampvec) +{ + unsigned char *trampoline = address; + int i; + + for(i = 0; i < nrcallees; ++i) { + trampoline[0] = 0x48; /* movabsq $..., %rax; */ + trampoline[1] = 0xb8; + *(void**)(trampoline+2) = NULL; /* callee's address */ + trampoline[10] = 0xff; /* jmpq *%rax */ + trampoline[11] = 0xe0; + trampvec[i] = trampoline; + trampoline += TRAMPOLINE_BYTES; + } + hipe_flush_icache_range(address, nrcallees*TRAMPOLINE_BYTES); +} + +static void patch_trampoline(void *trampoline, void *destAddress) +{ + unsigned char *tp = (unsigned char*) trampoline; + + ASSERT(tp[0] == 0x48 && tp[1] == 0xb8); + + *(void**)(tp+2) = destAddress; /* callee's address */ + hipe_flush_icache_word(tp+2); +} + void *hipe_alloc_code(Uint nrbytes, Eterm callees, Eterm *trampolines, Process *p) { - if (is_not_nil(callees)) + int nrcallees; + Eterm trampvecbin; + unsigned char **trampvec; + unsigned char *address; + + nrcallees = check_callees(callees); + if (nrcallees < 0) return NULL; - *trampolines = NIL; - return alloc_code(nrbytes); + + trampvecbin = new_binary(p, NULL, nrcallees*sizeof(unsigned char*)); + trampvec = (unsigned char **)binary_bytes(trampvecbin); + + address = alloc_code(nrbytes + nrcallees*TRAMPOLINE_BYTES); + generate_trampolines(address + nrbytes, nrcallees, callees, trampvec); + *trampolines = trampvecbin; + return address; } void hipe_free_code(void* code, unsigned int bytes) @@ -129,10 +204,9 @@ void *hipe_make_native_stub(void *callee_exp, unsigned int beamArity) */ unsigned int codeSize; unsigned char *code, *codep; - unsigned int callEmuOffset; - codeSize = /* 23, 26, 29, or 32 bytes */ - 23 + /* 23 when all offsets are 8-bit */ + codeSize = /* 30, 33, 36, or 39 bytes */ + 30 + /* 30 when all offsets are 8-bit */ (P_CALLEE_EXP >= 128 ? 3 : 0) + ((P_CALLEE_EXP + 4) >= 128 ? 3 : 0) + (P_ARITY >= 128 ? 3 : 0); @@ -197,14 +271,15 @@ void *hipe_make_native_stub(void *callee_exp, unsigned int beamArity) codep[0] = beamArity; codep += 1; - /* jmp callemu; 5 bytes */ - callEmuOffset = (unsigned char*)nbif_callemu - (code + codeSize); - codep[0] = 0xe9; - codep[1] = callEmuOffset & 0xFF; - codep[2] = (callEmuOffset >> 8) & 0xFF; - codep[3] = (callEmuOffset >> 16) & 0xFF; - codep[4] = (callEmuOffset >> 24) & 0xFF; - codep += 5; + /* jmp callemu; 12 bytes */ + codep[0] = 0x48; + codep[1] = 0xb8; + codep += 2; + *(Uint64*)codep = (Uint64)nbif_callemu; + codep += 8; + codep[0] = 0xff; + codep[1] = 0xe0; + codep += 2; ASSERT(codep == code + codeSize); diff --git a/erts/emulator/hipe/hipe_bif0.c b/erts/emulator/hipe/hipe_bif0.c index e477c4cdea..a8be64e08d 100644 --- a/erts/emulator/hipe/hipe_bif0.c +++ b/erts/emulator/hipe/hipe_bif0.c @@ -1112,7 +1112,7 @@ static struct hipe_mfa_info* mod2mfa_put(struct hipe_mfa_info* mfa) struct hipe_ref { struct hipe_ref_head head; /* list of refs to same calleee */ void *address; -#if defined(__arm__) || defined(__powerpc__) || defined(__ppc__) || defined(__powerpc64__) +#if defined(__x86_64__) || defined(__arm__) || defined(__powerpc__) || defined(__ppc__) || defined(__powerpc64__) void *trampoline; #endif unsigned int flags; @@ -1543,7 +1543,7 @@ BIF_RETTYPE hipe_bifs_add_ref_2(BIF_ALIST_2) ref = erts_alloc(ERTS_ALC_T_HIPE_LL, sizeof(struct hipe_ref)); ref->address = address; -#if defined(__arm__) || defined(__powerpc__) || defined(__ppc__) || defined(__powerpc64__) +#if defined(__x86_64__) || defined(__arm__) || defined(__powerpc__) || defined(__ppc__) || defined(__powerpc64__) ref->trampoline = trampoline; #endif ref->flags = flags; @@ -1819,7 +1819,7 @@ void hipe_redirect_to_module(Module* modp) if (ref->flags & REF_FLAG_IS_LOAD_MFA) res = hipe_patch_insn(ref->address, (Uint)p->remote_address, am_load_mfa); else { -#if defined(__arm__) || defined(__powerpc__) || defined(__ppc__) || defined(__powerpc64__) +#if defined(__x86_64__) || defined(__arm__) || defined(__powerpc__) || defined(__ppc__) || defined(__powerpc64__) void* trampoline = ref->trampoline; #else void* trampoline = NULL; diff --git a/erts/emulator/hipe/hipe_gc.c b/erts/emulator/hipe/hipe_gc.c index 1a4a4c7952..aaedba1afd 100644 --- a/erts/emulator/hipe/hipe_gc.c +++ b/erts/emulator/hipe/hipe_gc.c @@ -91,7 +91,7 @@ Eterm *fullsweep_nstack(Process *p, Eterm *n_htop) ASSERT(is_boxed(val)); *nsp_i = val; } else if (!erts_is_literal(gval, ptr)) { - move_boxed(&ptr, val, &n_htop, nsp_i); + move_boxed(ptr, val, &n_htop, nsp_i); } } else if (is_list(gval)) { Eterm *ptr = list_val(gval); @@ -100,7 +100,7 @@ Eterm *fullsweep_nstack(Process *p, Eterm *n_htop) *nsp_i = ptr[1]; } else if (!erts_is_literal(gval, ptr)) { ASSERT(erts_dbg_within_proc(ptr, p, NULL)); - move_cons(&ptr, val, &n_htop, nsp_i); + move_cons(ptr, val, &n_htop, nsp_i); } } } @@ -206,10 +206,10 @@ void gensweep_nstack(Process *p, Eterm **ptr_old_htop, Eterm **ptr_n_htop) ASSERT(is_boxed(val)); *nsp_i = val; } else if (ErtsInArea(ptr, mature, mature_size)) { - move_boxed(&ptr, val, &old_htop, nsp_i); + move_boxed(ptr, val, &old_htop, nsp_i); } else if (ErtsInYoungGen(gval, ptr, oh, oh_size)) { ASSERT(erts_dbg_within_proc(ptr, p, NULL)); - move_boxed(&ptr, val, &n_htop, nsp_i); + move_boxed(ptr, val, &n_htop, nsp_i); } } else if (is_list(gval)) { Eterm *ptr = list_val(gval); @@ -217,10 +217,10 @@ void gensweep_nstack(Process *p, Eterm **ptr_old_htop, Eterm **ptr_n_htop) if (IS_MOVED_CONS(val)) { *nsp_i = ptr[1]; } else if (ErtsInArea(ptr, mature, mature_size)) { - move_cons(&ptr, val, &old_htop, nsp_i); + move_cons(ptr, val, &old_htop, nsp_i); } else if (ErtsInYoungGen(gval, ptr, oh, oh_size)) { ASSERT(erts_dbg_within_proc(ptr, p, NULL)); - move_cons(&ptr, val, &n_htop, nsp_i); + move_cons(ptr, val, &n_htop, nsp_i); } } } @@ -278,7 +278,7 @@ Eterm *sweep_literals_nstack(Process *p, Eterm *old_htop, char *area, ASSERT(is_boxed(val)); *nsp_i = val; } else if (ErtsInArea(ptr, area, area_size)) { - move_boxed(&ptr, val, &old_htop, nsp_i); + move_boxed(ptr, val, &old_htop, nsp_i); } } else if (is_list(gval)) { Eterm *ptr = list_val(gval); @@ -286,7 +286,7 @@ Eterm *sweep_literals_nstack(Process *p, Eterm *old_htop, char *area, if (IS_MOVED_CONS(val)) { *nsp_i = ptr[1]; } else if (ErtsInArea(ptr, area, area_size)) { - move_cons(&ptr, val, &old_htop, nsp_i); + move_cons(ptr, val, &old_htop, nsp_i); } } } diff --git a/erts/emulator/internal_doc/CarrierMigration.md b/erts/emulator/internal_doc/CarrierMigration.md index 2a9594db25..3a796d11b7 100644 --- a/erts/emulator/internal_doc/CarrierMigration.md +++ b/erts/emulator/internal_doc/CarrierMigration.md @@ -3,17 +3,17 @@ Carrier Migration The ERTS memory allocators manage memory blocks in two types of raw memory chunks. We call these chunks of raw memory -*carriers*. Singleblock carriers which only contain one large block, -and multiblock carriers which contain multiple blocks. A carrier is +*carriers*. Single-block carriers which only contain one large block, +and multi-block carriers which contain multiple blocks. A carrier is typically created using `mmap()` on unix systems. However, how a carrier is created is of minor importance. An allocator instance -typically manages a mixture of single- and multiblock carriers. +typically manages a mixture of single- and multi-block carriers. Problem ------- When a carrier is empty, i.e. contains only one large free block, it -is deallocated. Since multiblock carriers can contain both allocated +is deallocated. Since multi-block carriers can contain both allocated blocks and free blocks at the same time, an allocator instance might be stuck with a large amount of poorly utilized carriers if the memory load decreases. After a peak in memory usage it is expected that not @@ -23,9 +23,9 @@ can usually be reused if the memory load increases again. However, since each scheduler thread manages its own set of allocator instances, and memory load is not necessarily correlated to CPU load, we might get into a situation where there are lots of poorly utilized -multiblock carriers on some allocator instances while we need to -allocate new multiblock carriers on other allocator instances. In -scenarios like this, the demand for multiblock carriers in the system +multi-block carriers on some allocator instances while we need to +allocate new multi-block carriers on other allocator instances. In +scenarios like this, the demand for multi-block carriers in the system might increase at the same time as the actual memory demand in the system has decreased which is both unwanted and quite unexpected for the end user. @@ -34,7 +34,7 @@ Solution -------- In order to prevent scenarios like this we've implemented support for -migration of multiblock carriers between allocator instances of the +migration of multi-block carriers between allocator instances of the same type. ### Management of Free Blocks ### @@ -44,7 +44,7 @@ and add it to another we need to be able to move references to the free blocks of the carrier between the allocator instances. The allocator instance specific data structure referring to the free blocks it manages often refers to the same carrier from multiple -places. For example, when the address order bestfit strategy is used +places. For example, when the address order best-fit strategy is used this data structure is a binary search tree spanning all carriers that the allocator instance manages. Free blocks in one specific carrier can be referred to from potentially every other carrier that is @@ -135,7 +135,7 @@ carriers between scheduler specific allocator instances of the same allocator type. Each allocator instance keeps track of the current utilization of its -multiblock carriers. When the total utilization falls below the "abandon +multi-block carriers. When the total utilization falls below the "abandon carrier utilization limit" it starts to inspect the utilization of the current carrier when deallocations are made. If also the utilization of the carrier falls below the "abandon carrier utilization limit" it @@ -144,31 +144,45 @@ and inserts the carrier into the pool. Since the carrier has been unlinked from the data structure of available free blocks, no more allocations will be made in the -carrier. The allocator instance putting the carrier into the pool, -however, still has the responsibility of performing deallocations in -it while it remains in the pool. The allocator instance with this -deallocation responsibility is here called the **employer**. - -Each carrier has a flag field containing information about the -employing allocator instance, a flag indicating if the carrier is in -the pool or not, and a flag indicating if it is busy or not. When the -carrier is in the pool, the employing allocator instance needs to mark it -as busy while operating on it. If another thread inspects it in order -to try to fetch it from the pool, it will skip it if it is busy. When -fetching the carrier from the pool, employment will change and further +carrier. + +The allocator instance that created a carrier is called its **owner**. +Ownership never changes. + +The allocator instance that has the responsibility to perform deallocations in a +carrier is called its **employer**. The employer may also perform allocations if +the carrier is not in the pool. Employment may change when a carrier is fetched from +or inserted into the pool. + +Deallocations in a carrier, while it remains in the pool, is always performed +the owner. That is, all pooled carriers are employed by their owners. + +Each carrier has an atomic word containing a pointer to the employing allocator +instance and three bit flags; IN_POOL, BUSY and HOMECOMING. + +When fetching a carrier from the pool, employment may change and further deallocations in the carrier will be redirected to the new employer using the delayed dealloc functionality. -If a carrier in the pool becomes empty, it will be withdrawn from the -pool. All carriers that become empty are also always passed to its -**owning** allocator instance for deallocation using the delayed -dealloc functionality. Since carriers this way always will be -deallocated by the owner that allocated the carrier, the +When a foreign allocator instance abandons a carrier back into the pool, it will +also pass it back to its **owner** using the delayed dealloc queue. When doing +this it will set the HOMECOMING bit flag to mark it as "enqueued". The owner +will later clear the HOMECOMING bit when the carrier is dequeued. This mechanism +prevents a carrier from being enqueued again before it has been dequeued. + +When a carrier becomes empty, it will be deallocated. Carrier deallocation is +always done by the owner that allocated the carrier. By doing this, the underlying functionality of allocating and deallocating carriers can remain simple and doesn't have to bother about multiple threads. In a NUMA system we will also not mix carriers originating from multiple NUMA nodes. +If a carrier in the pool becomes empty, it will be withdrawn from the +pool and be deallocated by the owner which already employs it. + +If a carrier employed by a foreign allocator becomes empty, it will be passed +back to the owner for deallocation using the delayed dealloc functionality. + In short: * The allocator instance that created a carrier **owns** it. @@ -177,34 +191,31 @@ In short: * The allocator instance that uses a carrier **employs** it. * An **employer** can abandon a carrier into the pool. * Pooled carriers are not allocated from. -* Deallocation in a pooled carrier is still performed by its **employer**. -* **Employment** can only change when a carrier is fetched from the pool. +* Pooled carriers are always **employed** by their **owner**. +* **Employment** can only change from **owner** to a foreign allocator + when a carrier is fetched from the pool. + ### Searching the pool ### +When an allocator instance needs more carrier space, it inspects the pool. If no +carrier could be fetched from the pool, it will allocate a new +carrier. Regardless of where the allocator instance gets the carrier from, it +just links in the carrier into its data structure of free blocks. + To harbor real time characteristics, searching the pool is limited. We only inspect a limited number of carriers. If none of those carriers had a free block large enough to satisfy the allocation -request, the search will fail. A carrier in the pool can also be busy +request, the search will fail. A carrier in the pool can also be BUSY if another thread is currently doing block deallocation work on the -carrier. A busy carrier will also be skipped by the search as it can +carrier. A BUSY carrier will also be skipped by the search as it can not satisfy the request. The pool is lock-free and we do not want to block, waiting for the other thread to finish. -#### Before OTP 17.4 #### +### The bad cluster problem ### -When an allocator instance needs more carrier space, it always begins -by inspecting its own carriers that are waiting for thread progress -before they can be deallocated. If no such carrier could be found, it -then inspects the pool. If no carrier could be fetched from the pool, -it will allocate a new carrier. Regardless of where the allocator -instance gets the carrier from it the just links in the carrier into -its data structure of free blocks. - -#### After OTP 17.4 #### - -The old search algorithm had a problem as the search always started at -the same position in the pool, the sentinel. This could lead to +Before OTP-17.4 the search algorithm had a problem as the search always started +at the same position in the pool, the sentinel. This could lead to contention from concurrent searching processes. But even worse, it could lead to a "bad" state when searches fail with a high rate leading to new carriers instead being allocated. These new carriers @@ -236,26 +247,27 @@ The result is that we prefer carriers created by the thread itself, which is good for NUMA performance. And we get more entry points when searching the pool, which will ease contention and clustering. +### Our own pooled tree ### + To do the first search among own carriers, every allocator instance -has two new lists: `pooled_list` and `traitor_list`. These lists are only -accessed by the allocator itself and they only contain the allocator's -own carriers. When an owned carrier is abandoned and put in the -pool, it is also linked into `pooled_list`. When we search our -`pooled_list` and find a carrier that is no longer in the pool, we -move that carrier from `pooled_list` to `traitor_list` as it is now -employed by another allocator. If searching `pooled_list` fails, we -also do a limited search of `traitor_list`. When finding an abandoned -carrier in `traitor_list` it is either employed or moved back to -`pooled_list` if it could not satisfy the allocation request. - -When searching `pooled_list` and `traitor_list` we always start at the -point where the last search ended. This to avoid clustering -problems and increase the probability to find a "good" carrier. As -`pooled_list` and `traitor_list` are only accessed by the owning -allocator instance, they need no thread synchronization at all. +has a `pooled_tree` of carriers. This tree is only accessed by the allocator +itself and can only contain its own carriers. When a carrier is +abandoned and put in the pool, it is also inserted into `pooled_tree`. This is +either done direct, if the carrier was already employed by its owner, or by +first passing it back to the owner via the delayed dealloc queue. + +When we search our `pooled_tree` and find a carrier that is no longer in the +pool, we remove that carrier from `pooled_tree` and mark it as TRAITOR, as it is +now employed by a foreign allocator. We will not find any carriers in +`pooled_tree` that are marked as BUSY by other threads. + +If no carrier in `pooled_tree` had a large enough free block, we search it again +to find any carrier that may act as an entry point into the shared list of all +pooled carriers. This in order to, if possible, avoid starting at the sentinel +and thereby ease the "bad clustering" problem. Furthermore, the search for own carriers that are scheduled -for deallocation is now done as the last search option. The idea is +for deallocation is done as the last search option. The idea is that it is better to reuse a poorly utilized carrier than to resurrect an empty carrier that was just about to be released back to the OS. @@ -271,14 +283,14 @@ load did not. When using the `aoffcaobf` or `aoff` strategies compared to `gf` or `bf`, we loose some performance since we get more modifications in the data structure of free blocks. This performance penalty is however -reduced using the `aoffcbf` strategy. A tradeoff between memory +reduced using the `aoffcbf` strategy. A trade off between memory consumption and performance is however inevitable, and it is up to the user to decide what is most important. Further work ------------ -It would be quite easy to extend this to allow migration of multiblock +It would be quite easy to extend this to allow migration of multi-block carriers between all allocator types. More or less the only obstacle is maintenance of the statistics information. diff --git a/erts/emulator/nifs/unix/unix_prim_file.c b/erts/emulator/nifs/unix/unix_prim_file.c index 57c8ef62e1..4a6c476882 100644 --- a/erts/emulator/nifs/unix/unix_prim_file.c +++ b/erts/emulator/nifs/unix/unix_prim_file.c @@ -125,24 +125,11 @@ static int open_file_type_check(const efile_path_t *path, int fd) { * immediately in a read within the call, but the new implementation * never does that. */ return 1; - } else { - /* The old driver tolerated opening /dev/null despite the "no devices" - * limitation. It provided no explanation for this but we still need - * to match the behavior. We're checking through stat(2) instead of - * comparing the name to account for links. */ - struct stat null_device_info; - int is_dev_null; - - is_dev_null = (stat("/dev/null", &null_device_info) == 0); - is_dev_null &= (file_info.st_ino == null_device_info.st_ino); - is_dev_null &= (file_info.st_dev == null_device_info.st_dev); - - if(is_dev_null) { - return 1; - } } - if(!S_ISREG(file_info.st_mode)) { + /* Allow everything that isn't a directory, and error out on the next call + * if it's unsupported. */ + if(S_ISDIR(file_info.st_mode)) { return 0; } diff --git a/erts/emulator/test/alloc_SUITE.erl b/erts/emulator/test/alloc_SUITE.erl index 5e5cd6f578..88ff2a7a92 100644 --- a/erts/emulator/test/alloc_SUITE.erl +++ b/erts/emulator/test/alloc_SUITE.erl @@ -31,6 +31,7 @@ mseg_clear_cache/1, erts_mmap/1, cpool/1, + set_dyn_param/1, migration/1]). -include_lib("common_test/include/ct.hrl"). @@ -41,6 +42,7 @@ suite() -> all() -> [basic, coalesce, threads, realloc_copy, bucket_index, + set_dyn_param, bucket_mask, rbtree, mseg_clear_cache, erts_mmap, cpool, migration]. init_per_testcase(Case, Config) when is_list(Config) -> @@ -65,7 +67,11 @@ mseg_clear_cache(Cfg) -> drv_case(Cfg). cpool(Cfg) -> drv_case(Cfg). migration(Cfg) -> - drv_case(Cfg, concurrent, "+MZe true"). + %% Enable test_alloc. + %% Disable driver_alloc to avoid recursive alloc_util calls + %% through enif_mutex_create() in my_creating_mbc(). + drv_case(Cfg, concurrent, "+MZe true +MRe false"), + drv_case(Cfg, concurrent, "+MZe true +MRe false +MZas ageffcbf"). erts_mmap(Config) when is_list(Config) -> case {os:type(), mmsc_flags()} of @@ -110,7 +116,7 @@ erts_mmap_do(Config, SCO, SCRPM, SCRFSD) -> 0 -> O1; _ -> O1 ++ " +MMscrfsd"++integer_to_list(SCRFSD) end, - {ok, Node} = start_node(Config, Opts), + {ok, Node} = start_node(Config, Opts, []), Self = self(), Ref = make_ref(), F = fun() -> @@ -140,6 +146,82 @@ erts_mmap_do(Config, SCO, SCRPM, SCRFSD) -> Result. +%% Test erlang:system_flag(erts_alloc, ...) +set_dyn_param(_Config) -> + {_, _, _, AlcList} = erlang:system_info(allocator), + + {Enabled, Disabled, Others} = + lists:foldl(fun({sys_alloc,_}, {Es, Ds, Os}) -> + {Es, [sys_alloc | Ds], Os}; + + ({AT, Opts}, {Es, Ds, Os}) when is_list(Opts) -> + case lists:keyfind(e, 1, Opts) of + {e, true} -> + {[AT | Es], Ds, Os}; + {e, false} -> + {Es, [AT | Ds], Os}; + false -> + {Es, Ds, [AT | Os]} + end; + + (_, Acc) -> Acc + end, + {[], [], []}, + AlcList), + + Param = sbct, + lists:foreach(fun(AT) -> set_dyn_param_enabled(AT, Param) end, + Enabled), + + lists:foreach(fun(AT) -> + Tpl = {AT, Param, 12345}, + io:format("~p\n", [Tpl]), + notsup = erlang:system_flag(erts_alloc, Tpl) + end, + Disabled), + + lists:foreach(fun(AT) -> + Tpl = {AT, Param, 12345}, + io:format("~p\n", [Tpl]), + {'EXIT',{badarg,_}} = + (catch erlang:system_flag(erts_alloc, Tpl)) + end, + Others), + ok. + +set_dyn_param_enabled(AT, Param) -> + OldVal = get_alc_param(AT, Param), + + Val1 = OldVal div 2, + Tuple = {AT, Param, Val1}, + io:format("~p\n", [Tuple]), + ok = erlang:system_flag(erts_alloc, Tuple), + Val1 = get_alc_param(AT, Param), + + ok = erlang:system_flag(erts_alloc, {AT, Param, OldVal}), + OldVal = get_alc_param(AT, Param), + ok. + +get_alc_param(AT, Param) -> + lists:foldl(fun({instance,_,Istats}, Acc) -> + {options,Opts} = lists:keyfind(options, 1, Istats), + {Param,Val} = lists:keyfind(Param, 1, Opts), + {as,Strategy} = lists:keyfind(as, 1, Opts), + + case {param_for_strat(Param, Strategy), Acc} of + {false, _} -> Acc; + {true, undefined} -> Val; + {true, _} -> + Val = Acc + end + end, + undefined, + erlang:system_info({allocator, AT})). + +param_for_strat(sbct, gf) -> false; +param_for_strat(_, _) -> true. + + %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% %% %% %% Internal functions %% @@ -151,7 +233,9 @@ drv_case(Config) -> drv_case(Config, Mode, NodeOpts) when is_list(Config) -> case os:type() of {Family, _} when Family == unix; Family == win32 -> - {ok, Node} = start_node(Config, NodeOpts), + %%Prog = {prog,"/my/own/otp/bin/cerl -debug"}, + Prog = [], + {ok, Node} = start_node(Config, NodeOpts, Prog), Self = self(), Ref = make_ref(), spawn_link(Node, @@ -217,19 +301,35 @@ wait_for_memory_deallocations() -> end. print_stats(migration) -> - {Btot,Ctot} = lists:foldl(fun({instance,Inr,Istats}, {Bacc,Cacc}) -> - {mbcs,MBCS} = lists:keyfind(mbcs, 1, Istats), - Btup = lists:keyfind(blocks, 1, MBCS), - Ctup = lists:keyfind(carriers, 1, MBCS), - io:format("{instance,~p,~p,~p}\n", [Inr, Btup, Ctup]), - {tuple_add(Bacc,Btup),tuple_add(Cacc,Ctup)}; - (_, Acc) -> Acc - end, - {{blocks,0,0,0},{carriers,0,0,0}}, - erlang:system_info({allocator,test_alloc})), - + IFun = fun({instance,Inr,Istats}, {Bacc,Cacc,Pacc}) -> + {mbcs,MBCS} = lists:keyfind(mbcs, 1, Istats), + Btup = lists:keyfind(blocks, 1, MBCS), + Ctup = lists:keyfind(carriers, 1, MBCS), + + Ptup = case lists:keyfind(mbcs_pool, 1, Istats) of + {mbcs_pool,POOL} -> + {blocks, Bpool} = lists:keyfind(blocks, 1, POOL), + {carriers, Cpool} = lists:keyfind(carriers, 1, POOL), + {pool, Bpool, Cpool}; + false -> + {pool, 0, 0} + end, + io:format("{instance,~p,~p,~p,~p}}\n", + [Inr, Btup, Ctup, Ptup]), + {tuple_add(Bacc,Btup),tuple_add(Cacc,Ctup), + tuple_add(Pacc,Ptup)}; + (_, Acc) -> Acc + end, + + {Btot,Ctot,Ptot} = lists:foldl(IFun, + {{blocks,0,0,0},{carriers,0,0,0},{pool,0,0}}, + erlang:system_info({allocator,test_alloc})), + + {pool, PBtot, PCtot} = Ptot, io:format("Number of blocks : ~p\n", [Btot]), - io:format("Number of carriers: ~p\n", [Ctot]); + io:format("Number of carriers: ~p\n", [Ctot]), + io:format("Number of pooled blocks : ~p\n", [PBtot]), + io:format("Number of pooled carriers: ~p\n", [PCtot]); print_stats(_) -> ok. tuple_add(T1, T2) -> @@ -326,13 +426,13 @@ handle_result(_State, Result0) -> continue end. -start_node(Config, Opts) when is_list(Config), is_list(Opts) -> +start_node(Config, Opts, Prog) when is_list(Config), is_list(Opts) -> case proplists:get_value(debug,Config) of true -> {ok, node()}; - _ -> start_node_1(Config, Opts) + _ -> start_node_1(Config, Opts, Prog) end. -start_node_1(Config, Opts) -> +start_node_1(Config, Opts, Prog) -> Pa = filename:dirname(code:which(?MODULE)), Name = list_to_atom(atom_to_list(?MODULE) ++ "-" @@ -341,7 +441,11 @@ start_node_1(Config, Opts) -> ++ integer_to_list(erlang:system_time(second)) ++ "-" ++ integer_to_list(erlang:unique_integer([positive]))), - test_server:start_node(Name, slave, [{args, Opts++" -pa "++Pa}]). + ErlArg = case Prog of + [] -> []; + _ -> [{erl,[Prog]}] + end, + test_server:start_node(Name, slave, [{args, Opts++" -pa "++Pa} | ErlArg]). stop_node(Node) when Node =:= node() -> ok; stop_node(Node) -> diff --git a/erts/emulator/test/alloc_SUITE_data/allocator_test.h b/erts/emulator/test/alloc_SUITE_data/allocator_test.h index 97ee58cdad..5272f86c98 100644 --- a/erts/emulator/test/alloc_SUITE_data/allocator_test.h +++ b/erts/emulator/test/alloc_SUITE_data/allocator_test.h @@ -156,7 +156,8 @@ typedef void* erts_cond; #define IS_SMP_ENABLED ((int) ALC_TEST0(0xf13)) #define ALLOC_TEST(S) ((void*) ALC_TEST1(0xf14, (S))) #define FREE_TEST(P) ((void) ALC_TEST1(0xf15, (P))) -#define SET_TEST_MBC_USER_HEADER(SZ,CMBC,DMBC) ((int)ALC_TEST3(0xf16, (SZ), (CMBC), (DMBC))) -#define GET_TEST_MBC_SIZE() ((int) ALC_TEST0(0xf17)) +#define REALLOC_TEST(P,S) ((void*) ALC_TEST2(0xf16, (P), (S))) +#define SET_TEST_MBC_USER_HEADER(SZ,CMBC,DMBC) ((int)ALC_TEST3(0xf17, (SZ), (CMBC), (DMBC))) +#define GET_TEST_MBC_SIZE() ((int) ALC_TEST0(0xf18)) #endif diff --git a/erts/emulator/test/alloc_SUITE_data/migration.c b/erts/emulator/test/alloc_SUITE_data/migration.c index b9a4de03b3..1d974225fc 100644 --- a/erts/emulator/test/alloc_SUITE_data/migration.c +++ b/erts/emulator/test/alloc_SUITE_data/migration.c @@ -223,6 +223,42 @@ static int rand_int(MigrationState* state, int low, int high) return low + (x % (high+1-low)); } +enum Operation +{ + ALLOCATE_OP, + FREE_OP, + REALLOC_OP, + CLEANUP_OP +}; + +static enum Operation rand_op(MigrationState* state) +{ + int r = rand_int(state, 1, 100); + switch (state->phase) { + case GROWING: + FATAL_ASSERT(state->nblocks < state->max_nblocks); + if (r > 10 || state->nblocks == 0) + return ALLOCATE_OP; + else if (r > 5) + return FREE_OP; + else + return REALLOC_OP; + + case SHRINKING: + FATAL_ASSERT(state->nblocks > 0); + if (r > 10 || state->nblocks == state->max_nblocks) + return FREE_OP; + else if (r > 5) + return ALLOCATE_OP; + else + return REALLOC_OP; + + case CLEANUP: + return CLEANUP_OP; + default: + FATAL_ASSERT(!"Invalid op phase"); + } +} static void do_cleanup(TestCaseState_t *tcs, MigrationState* state) { @@ -275,53 +311,75 @@ testcase_run(TestCaseState_t *tcs) state->goal_nblocks = rand_int(state, 1, state->max_nblocks); } - switch (state->phase) { - case GROWING: { + switch (rand_op(state)) { + case ALLOCATE_OP: { MyBlock* p; FATAL_ASSERT(!state->blockv[state->nblocks]); - p = ALLOC_TEST(rand_int(state, state->block_size/2, state->block_size)); + p = ALLOC_TEST(rand_int(state, state->block_size/2, state->block_size)); FATAL_ASSERT(p); add_block(p, state); - state->blockv[state->nblocks] = p; - if (++state->nblocks >= state->goal_nblocks) { - /*testcase_printf(tcs, "%d: Grown to %d blocks", tcs->thr_nr, state->nblocks);*/ - state->phase = SHRINKING; - state->goal_nblocks = rand_int(state, 0, state->goal_nblocks-1); - } - else - FATAL_ASSERT(!state->blockv[state->nblocks]); + state->blockv[state->nblocks++] = p; break; } - case SHRINKING: { + case FREE_OP: { int ix = rand_int(state, 0, state->nblocks-1); FATAL_ASSERT(state->blockv[ix]); remove_block(state->blockv[ix]); FREE_TEST(state->blockv[ix]); state->blockv[ix] = state->blockv[--state->nblocks]; state->blockv[state->nblocks] = NULL; - - if (state->nblocks <= state->goal_nblocks) { - /*testcase_printf(tcs, "%d: Shrunk to %d blocks", tcs->thr_nr, state->nblocks);*/ - if (++state->round >= MAX_ROUNDS) { - state->phase = CLEANUP; - } else { - state->phase = GROWING; - state->goal_nblocks = rand_int(state, state->goal_nblocks+1, state->max_nblocks); - } - } break; } + case REALLOC_OP: { + int ix = rand_int(state, 0, state->nblocks-1); + MyBlock* p; + FATAL_ASSERT(state->blockv[ix]); + remove_block(state->blockv[ix]); + p = REALLOC_TEST(state->blockv[ix], rand_int(state, state->block_size/2, state->block_size)); + FATAL_ASSERT(p); + add_block(p, state); + state->blockv[ix] = p; + break; + } + case CLEANUP_OP: + do_cleanup(tcs, state); + break; + default: + FATAL_ASSERT(!"Invalid operation"); + } + + switch (state->phase) { + case GROWING: { + if (state->nblocks >= state->goal_nblocks) { + /*testcase_printf(tcs, "%d: Grown to %d blocks", tcs->thr_nr, state->nblocks);*/ + state->phase = SHRINKING; + state->goal_nblocks = rand_int(state, 0, state->goal_nblocks-1); + } + else + FATAL_ASSERT(!state->blockv[state->nblocks]); + break; + } + case SHRINKING: { + if (state->nblocks <= state->goal_nblocks) { + /*testcase_printf(tcs, "%d: Shrunk to %d blocks", tcs->thr_nr, state->nblocks);*/ + if (++state->round >= MAX_ROUNDS) { + state->phase = CLEANUP; + } else { + state->phase = GROWING; + state->goal_nblocks = rand_int(state, state->goal_nblocks+1, state->max_nblocks); + } + } + break; + } case CLEANUP: - do_cleanup(tcs, state); - break; + case DONE: + break; default: FATAL_ASSERT(!"Invalid phase"); } - if (state->phase == DONE) { - } - else { + if (state->phase != DONE) { testcase_continue(tcs); } } diff --git a/erts/emulator/test/binary_SUITE.erl b/erts/emulator/test/binary_SUITE.erl index 4bc1838139..a3c3daac15 100644 --- a/erts/emulator/test/binary_SUITE.erl +++ b/erts/emulator/test/binary_SUITE.erl @@ -262,6 +262,7 @@ test_deep_bitstr(List) -> {Bin,bitstring_to_list(Bin)}. bad_list_to_binary(Config) when is_list(Config) -> + test_bad_bin(<<1:1>>), test_bad_bin(atom), test_bad_bin(42), test_bad_bin([1|2]), diff --git a/erts/emulator/test/efile_SUITE.erl b/erts/emulator/test/efile_SUITE.erl index 16d581a567..821381bf0d 100644 --- a/erts/emulator/test/efile_SUITE.erl +++ b/erts/emulator/test/efile_SUITE.erl @@ -46,14 +46,14 @@ iter_max_files_1(Config) -> DataDir = proplists:get_value(data_dir,Config), TestFile = filename:join(DataDir, "existing_file"), N = 10, - %% Run on a different node in order to set the max ports + %% Run on a different node in order to make the test more stable. Dir = filename:dirname(code:which(?MODULE)), {ok,Node} = test_server:start_node(test_iter_max_files,slave, - [{args,"+Q 1524 -pa " ++ Dir}]), + [{args,"-pa " ++ Dir}]), L = rpc:call(Node,?MODULE,do_iter_max_files,[N, TestFile]), test_server:stop_node(Node), io:format("Number of files opened in each test:~n~w\n", [L]), - all_equal(L), + verify_max_files(L), Head = hd(L), if Head >= 2 -> ok; true -> ct:fail(too_few_files) @@ -65,12 +65,15 @@ do_iter_max_files(N, Name) when N > 0 -> do_iter_max_files(_, _) -> []. -all_equal([E, E| T]) -> - all_equal([E| T]); -all_equal([_]) -> - ok; -all_equal([]) -> - ok. +%% The attempts shouldn't vary too much; we used to require that they were all +%% exactly equal, but after we reimplemented the file driver as a NIF we +%% noticed that the only reason it was stable on Darwin was because the port +%% limit was hit before ulimit. +verify_max_files(Attempts) -> + N = length(Attempts), + Mean = lists:sum(Attempts) / N, + Variance = lists:sum([(X - Mean) * (X - Mean) || X <- Attempts]) / N, + true = math:sqrt(Variance) =< 1 + (Mean / 1000). max_files(Name) -> Fds = open_files(Name), diff --git a/erts/emulator/test/iovec_SUITE.erl b/erts/emulator/test/iovec_SUITE.erl index 49dc64b0d2..963b7e2501 100644 --- a/erts/emulator/test/iovec_SUITE.erl +++ b/erts/emulator/test/iovec_SUITE.erl @@ -25,7 +25,8 @@ -export([integer_lists/1, binary_lists/1, empty_lists/1, empty_binary_lists/1, mixed_lists/1, improper_lists/1, illegal_lists/1, cons_bomb/1, sub_binary_lists/1, iolist_to_iovec_idempotence/1, - iolist_to_iovec_correctness/1]). + iolist_to_iovec_correctness/1, unaligned_sub_binaries/1, + direct_binary_arg/1]). -include_lib("common_test/include/ct.hrl"). @@ -36,7 +37,8 @@ suite() -> all() -> [integer_lists, binary_lists, empty_lists, empty_binary_lists, mixed_lists, sub_binary_lists, illegal_lists, improper_lists, cons_bomb, - iolist_to_iovec_idempotence, iolist_to_iovec_correctness]. + iolist_to_iovec_idempotence, iolist_to_iovec_correctness, + unaligned_sub_binaries, direct_binary_arg]. init_per_suite(Config) -> Config. @@ -78,7 +80,7 @@ illegal_lists(Config) when is_list(Config) -> BitStrs = gen_variations(["gurka", <<1:1>>, "gaffel"]), BadInts = gen_variations(["gurka", 890, "gaffel"]), Atoms = gen_variations([gurka, "gaffel"]), - BadTails = [["test" | 0], ["gurka", gaffel]], + BadTails = [["test" | 0], ["gurka" | gaffel], ["gaffel" | <<1:1>>]], Variations = BitStrs ++ BadInts ++ Atoms ++ BadTails, @@ -98,14 +100,7 @@ cons_bomb(Config) when is_list(Config) -> BinBase = gen_variations([<<I:8>> || I <- lists:seq(1, 255)]), MixBase = gen_variations([<<12, 45, 78>>, lists:seq(1, 255)]), - Rounds = - case system_mem_size() of - Mem when Mem >= (16 bsl 30) -> 32; - Mem when Mem >= (3 bsl 30) -> 28; - _ -> 20 - end, - - Variations = gen_variations([IntBase, BinBase, MixBase], Rounds), + Variations = gen_variations([IntBase, BinBase, MixBase], 16), equivalence_test(fun erlang:iolist_to_iovec/1, Variations). iolist_to_iovec_idempotence(Config) when is_list(Config) -> @@ -130,6 +125,21 @@ iolist_to_iovec_correctness(Config) when is_list(Config) -> true = is_iolist_equal(Optimized, Variations), ok. +unaligned_sub_binaries(Config) when is_list(Config) -> + UnalignedBins = [gen_unaligned_binary(I) || I <- lists:seq(32, 4 bsl 10, 512)], + UnalignedVariations = gen_variations(UnalignedBins), + + Optimized = erlang:iolist_to_iovec(UnalignedVariations), + + true = is_iolist_equal(Optimized, UnalignedVariations), + ok. + +direct_binary_arg(Config) when is_list(Config) -> + {'EXIT',{badarg, _}} = (catch erlang:iolist_to_iovec(<<1:1>>)), + [<<1>>] = erlang:iolist_to_iovec(<<1>>), + [] = erlang:iolist_to_iovec(<<>>), + ok. + illegality_test(Fun, Variations) -> [{'EXIT',{badarg, _}} = (catch Fun(Variation)) || Variation <- Variations], ok. @@ -145,11 +155,18 @@ equivalence_test(Fun, [Head | _] = Variations) -> is_iolist_equal(A, B) -> iolist_to_binary(A) =:= iolist_to_binary(B). +gen_unaligned_binary(Size) -> + Bin0 = << <<I>> || I <- lists:seq(1, Size) >>, + <<0:3,Bin:Size/binary,31:5>> = id(<<0:3,Bin0/binary,31:5>>), + Bin. + +id(I) -> I. + %% Generates a bunch of lists whose contents will be equal to Base repeated a %% few times. The lists only differ by their structure, so their reduction to %% a simpler format should yield the same result. gen_variations(Base) -> - gen_variations(Base, 16). + gen_variations(Base, 12). gen_variations(Base, N) -> [gen_flat_list(Base, N), gen_nested_list(Base, N), @@ -169,8 +186,3 @@ gen_nasty_list_1([Head | Base], Result) when is_list(Head) -> gen_nasty_list_1(Base, [[Result], [gen_nasty_list_1(Head, [])]]); gen_nasty_list_1([Head | Base], Result) -> gen_nasty_list_1(Base, [[Result], [Head]]). - -system_mem_size() -> - application:ensure_all_started(os_mon), - {Tot,_Used,_} = memsup:get_memory_data(), - Tot. diff --git a/erts/emulator/test/nif_SUITE.erl b/erts/emulator/test/nif_SUITE.erl index 85c302e310..b6e15ababb 100644 --- a/erts/emulator/test/nif_SUITE.erl +++ b/erts/emulator/test/nif_SUITE.erl @@ -3049,8 +3049,24 @@ nif_ioq(Config) -> {peek_head, a}, {peek, a}, - %% Test to enqueue a bunch of refc binaries + %% Ensure that enqueued refc binaries are intact after a roundtrip. + %% + %% This test and the ones immediately following it does not go through + %% erlang:iolist_to_iovec/1 {enqv, a, [nif_ioq_payload(refcbin) || _ <- lists:seq(1,20)], 0}, + {peek, a}, + + %% ... heap binaries + {enqv, a, [nif_ioq_payload(heapbin) || _ <- lists:seq(1,20)], 0}, + {peek, a}, + + %% ... plain sub-binaries + {enqv, a, [nif_ioq_payload(subbin) || _ <- lists:seq(1,20)], 0}, + {peek, a}, + + %% ... unaligned binaries + {enqv, a, [nif_ioq_payload(unaligned_bin) || _ <- lists:seq(1,20)], 0}, + {peek, a}, %% Enq stuff to destroy with data in queue {enqv, a, 2, 100}, @@ -3206,13 +3222,16 @@ nif_ioq_payload(N) when is_integer(N) -> Tail = if N > 3 -> nif_ioq_payload(N-3); true -> [] end, Head = element(1, lists:split(N,[nif_ioq_payload(subbin), nif_ioq_payload(heapbin), - nif_ioq_payload(refcbin) | Tail])), + nif_ioq_payload(refcbin), + nif_ioq_payload(unaligned_bin) | Tail])), erlang:iolist_to_iovec(Head); nif_ioq_payload(subbin) -> Bin = nif_ioq_payload(refcbin), Sz = size(Bin) - 1, <<_:8,SubBin:Sz/binary,_/bits>> = Bin, SubBin; +nif_ioq_payload(unaligned_bin) -> + make_unaligned_binary(<< <<I>> || I <- lists:seq(1, 255) >>); nif_ioq_payload(heapbin) -> <<"a literal heap binary">>; nif_ioq_payload(refcbin) -> @@ -3220,6 +3239,13 @@ nif_ioq_payload(refcbin) -> nif_ioq_payload(Else) -> Else. +make_unaligned_binary(Bin0) -> + Size = byte_size(Bin0), + <<0:3,Bin:Size/binary,31:5>> = id(<<0:3,Bin0/binary,31:5>>), + Bin. + +id(I) -> I. + %% The NIFs: lib_version() -> undefined. call_history() -> ?nif_stub. diff --git a/erts/etc/common/erlexec.c b/erts/etc/common/erlexec.c index f7c8c27d53..f7774c6e2e 100644 --- a/erts/etc/common/erlexec.c +++ b/erts/etc/common/erlexec.c @@ -80,6 +80,8 @@ static char *plusM_au_alloc_switches[] = { "as", "asbcst", "acul", + "acnl", + "acfml", "e", "t", "lmbcs", diff --git a/erts/lib_src/common/erl_printf_format.c b/erts/lib_src/common/erl_printf_format.c index 3302083288..5a680d6f9d 100644 --- a/erts/lib_src/common/erl_printf_format.c +++ b/erts/lib_src/common/erl_printf_format.c @@ -326,7 +326,7 @@ static int fmt_double(fmtfn_t fn,void*arg,double val, { int res; int fi = 0; - char format_str[7]; + char format_str[8]; char sbuf[32]; char *bufp = sbuf; double dexp; diff --git a/erts/preloaded/ebin/erl_prim_loader.beam b/erts/preloaded/ebin/erl_prim_loader.beam Binary files differindex 2d86dd3843..02e77bfbb2 100644 --- a/erts/preloaded/ebin/erl_prim_loader.beam +++ b/erts/preloaded/ebin/erl_prim_loader.beam diff --git a/erts/preloaded/ebin/erlang.beam b/erts/preloaded/ebin/erlang.beam Binary files differindex f0da8b5869..cf4b5c56ea 100644 --- a/erts/preloaded/ebin/erlang.beam +++ b/erts/preloaded/ebin/erlang.beam diff --git a/erts/preloaded/src/erl_prim_loader.erl b/erts/preloaded/src/erl_prim_loader.erl index d16fe5a0bc..b82b9f5b46 100644 --- a/erts/preloaded/src/erl_prim_loader.erl +++ b/erts/preloaded/src/erl_prim_loader.erl @@ -56,7 +56,7 @@ -export([purge_archive_cache/0]). %% Used by init and the code server. --export([get_modules/2,get_modules/3]). +-export([get_modules/2,get_modules/3, is_basename/1]). -include_lib("kernel/include/file.hrl"). diff --git a/erts/preloaded/src/erlang.erl b/erts/preloaded/src/erlang.erl index 278a485a01..d352e11751 100644 --- a/erts/preloaded/src/erlang.erl +++ b/erts/preloaded/src/erlang.erl @@ -2403,6 +2403,10 @@ subtract(_,_) -> OldDirtyCPUSchedulersOnline when DirtyCPUSchedulersOnline :: pos_integer(), OldDirtyCPUSchedulersOnline :: pos_integer(); + (erts_alloc, {Alloc, F, V}) -> ok | notsup when + Alloc :: atom(), + F :: atom(), + V :: integer(); (fullsweep_after, Number) -> OldNumber when Number :: non_neg_integer(), OldNumber :: non_neg_integer(); @@ -3730,15 +3734,14 @@ memory_is_supported() -> get_blocks_size([{blocks_size, Sz, _, _} | Rest], Acc) -> get_blocks_size(Rest, Acc+Sz); -get_blocks_size([{_, _, _, _} | Rest], Acc) -> - get_blocks_size(Rest, Acc); get_blocks_size([{blocks_size, Sz} | Rest], Acc) -> get_blocks_size(Rest, Acc+Sz); -get_blocks_size([{_, _} | Rest], Acc) -> +get_blocks_size([_ | Rest], Acc) -> get_blocks_size(Rest, Acc); get_blocks_size([], Acc) -> Acc. + blocks_size([{Carriers, SizeList} | Rest], Acc) when Carriers == mbcs; Carriers == mbcs_pool; Carriers == sbcs -> diff --git a/erts/preloaded/src/init.erl b/erts/preloaded/src/init.erl index 679a2241d2..e0ae6b1656 100644 --- a/erts/preloaded/src/init.erl +++ b/erts/preloaded/src/init.erl @@ -32,8 +32,8 @@ %% (Optional - default efile) %% -hosts [Node] : List of hosts from which we can boot. %% (Mandatory if -loader inet) -%% -mode embedded : Load all modules at startup, no automatic loading -%% -mode interactive : Auto load modules (default system behaviour). +%% -mode interactive : Auto load modules not needed at startup (default system behaviour). +%% -mode embedded : Load all modules in the boot script, disable auto loading. %% -path : Override path in bootfile. %% -pa Path+ : Add my own paths first. %% -pz Path+ : Add my own paths last. diff --git a/erts/test/z_SUITE.erl b/erts/test/z_SUITE.erl index b3d0ba20bc..bb1afecafc 100644 --- a/erts/test/z_SUITE.erl +++ b/erts/test/z_SUITE.erl @@ -40,7 +40,7 @@ suite() -> [{ct_hooks,[ts_install_cth]}, - {timetrap, {minutes, 5}}]. + {timetrap, {minutes, 10}}]. all() -> [core_files]. diff --git a/erts/vsn.mk b/erts/vsn.mk index 8b22afea3b..50b77fdbc0 100644 --- a/erts/vsn.mk +++ b/erts/vsn.mk @@ -18,7 +18,7 @@ # %CopyrightEnd% # -VSN = 9.2 +VSN = 9.2.1 # Port number 4365 in 4.2 # Port number 4366 in 4.3 |