diff --git a/contrib/jemalloc/.clang-format b/contrib/jemalloc/.clang-format new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/.clang-format @@ -0,0 +1,122 @@ +# jemalloc targets clang-format version 8. We include every option it supports +# here, but comment out the ones that aren't relevant for us. +--- +# AccessModifierOffset: -2 +AlignAfterOpenBracket: DontAlign +AlignConsecutiveAssignments: false +AlignConsecutiveDeclarations: false +AlignEscapedNewlines: Right +AlignOperands: false +AlignTrailingComments: false +AllowAllParametersOfDeclarationOnNextLine: true +AllowShortBlocksOnASingleLine: false +AllowShortCaseLabelsOnASingleLine: false +AllowShortFunctionsOnASingleLine: Empty +AllowShortIfStatementsOnASingleLine: false +AllowShortLoopsOnASingleLine: false +AlwaysBreakAfterReturnType: AllDefinitions +AlwaysBreakBeforeMultilineStrings: true +# AlwaysBreakTemplateDeclarations: Yes +BinPackArguments: true +BinPackParameters: true +BraceWrapping: + AfterClass: false + AfterControlStatement: false + AfterEnum: false + AfterFunction: false + AfterNamespace: false + AfterObjCDeclaration: false + AfterStruct: false + AfterUnion: false + BeforeCatch: false + BeforeElse: false + IndentBraces: false +# BreakAfterJavaFieldAnnotations: true +BreakBeforeBinaryOperators: NonAssignment +BreakBeforeBraces: Attach +BreakBeforeTernaryOperators: true +# BreakConstructorInitializers: BeforeColon +# BreakInheritanceList: BeforeColon +BreakStringLiterals: false +ColumnLimit: 80 +# CommentPragmas: '' +# CompactNamespaces: true +# ConstructorInitializerAllOnOneLineOrOnePerLine: true +# ConstructorInitializerIndentWidth: 4 +ContinuationIndentWidth: 2 +Cpp11BracedListStyle: true +DerivePointerAlignment: false +DisableFormat: false +ExperimentalAutoDetectBinPacking: false +FixNamespaceComments: true +ForEachMacros: [ ql_foreach, qr_foreach, ] +# IncludeBlocks: Preserve +# IncludeCategories: +# - Regex: '^<.*\.h(pp)?>' +# Priority: 1 +# IncludeIsMainRegex: '' +IndentCaseLabels: false +IndentPPDirectives: AfterHash +IndentWidth: 4 +IndentWrappedFunctionNames: false +# JavaImportGroups: [] +# JavaScriptQuotes: Leave +# JavaScriptWrapImports: True +KeepEmptyLinesAtTheStartOfBlocks: false +Language: Cpp +MacroBlockBegin: '' +MacroBlockEnd: '' +MaxEmptyLinesToKeep: 1 +# NamespaceIndentation: None +# ObjCBinPackProtocolList: Auto +# ObjCBlockIndentWidth: 2 +# ObjCSpaceAfterProperty: false +# ObjCSpaceBeforeProtocolList: false + +PenaltyBreakAssignment: 2 +PenaltyBreakBeforeFirstCallParameter: 1 +PenaltyBreakComment: 300 +PenaltyBreakFirstLessLess: 120 +PenaltyBreakString: 1000 +# PenaltyBreakTemplateDeclaration: 10 +PenaltyExcessCharacter: 1000000 +PenaltyReturnTypeOnItsOwnLine: 60 +PointerAlignment: Right +# RawStringFormats: +# - Language: TextProto +# Delimiters: +# - 'pb' +# - 'proto' +# EnclosingFunctions: +# - 'PARSE_TEXT_PROTO' +# BasedOnStyle: google +# - Language: Cpp +# Delimiters: +# - 'cc' +# - 'cpp' +# BasedOnStyle: llvm +# CanonicalDelimiter: 'cc' +ReflowComments: true +SortIncludes: false +SpaceAfterCStyleCast: false +# SpaceAfterTemplateKeyword: true +SpaceBeforeAssignmentOperators: true +# SpaceBeforeCpp11BracedList: false +# SpaceBeforeCtorInitializerColon: true +# SpaceBeforeInheritanceColon: true +SpaceBeforeParens: ControlStatements +# SpaceBeforeRangeBasedForLoopColon: true +SpaceInEmptyParentheses: false +SpacesBeforeTrailingComments: 2 +SpacesInAngles: false +SpacesInCStyleCastParentheses: false +# SpacesInContainerLiterals: false +SpacesInParentheses: false +SpacesInSquareBrackets: false +# Standard: Cpp11 +# This is nominally supported in clang-format version 8, but not in the build +# used by some of the core jemalloc developers. +# StatementMacros: [] +TabWidth: 8 +UseTab: Never +... diff --git a/contrib/jemalloc/ChangeLog b/contrib/jemalloc/ChangeLog --- a/contrib/jemalloc/ChangeLog +++ b/contrib/jemalloc/ChangeLog @@ -4,6 +4,106 @@ https://github.com/jemalloc/jemalloc +* 5.3.0 (May 6, 2022) + + This release contains many speed and space optimizations, from micro + optimizations on common paths to rework of internal data structures and + locking schemes, and many more too detailed to list below. Multiple percent + of system level metric improvements were measured in tested production + workloads. The release has gone through large-scale production testing. + + New features: + - Add the thread.idle mallctl which hints that the calling thread will be + idle for a nontrivial period of time. (@davidtgoldblatt) + - Allow small size classes to be the maximum size class to cache in the + thread-specific cache, through the opt.[lg_]tcache_max option. (@interwq, + @jordalgo) + - Make the behavior of realloc(ptr, 0) configurable with opt.zero_realloc. + (@davidtgoldblatt) + - Add 'make uninstall' support. (@sangshuduo, @Lapenkov) + - Support C++17 over-aligned allocation. (@marksantaniello) + - Add the thread.peak mallctl for approximate per-thread peak memory tracking. + (@davidtgoldblatt) + - Add interval-based stats output opt.stats_interval. (@interwq) + - Add prof.prefix to override filename prefixes for dumps. (@zhxchen17) + - Add high resolution timestamp support for profiling. (@tyroguru) + - Add the --collapsed flag to jeprof for flamegraph generation. + (@igorwwwwwwwwwwwwwwwwwwww) + - Add the --debug-syms-by-id option to jeprof for debug symbols discovery. + (@DeannaGelbart) + - Add the opt.prof_leak_error option to exit with error code when leak is + detected using opt.prof_final. (@yunxuo) + - Add opt.cache_oblivious as an runtime alternative to config.cache_oblivious. + (@interwq) + - Add mallctl interfaces: + + opt.zero_realloc (@davidtgoldblatt) + + opt.cache_oblivious (@interwq) + + opt.prof_leak_error (@yunxuo) + + opt.stats_interval (@interwq) + + opt.stats_interval_opts (@interwq) + + opt.tcache_max (@interwq) + + opt.trust_madvise (@azat) + + prof.prefix (@zhxchen17) + + stats.zero_reallocs (@davidtgoldblatt) + + thread.idle (@davidtgoldblatt) + + thread.peak.{read,reset} (@davidtgoldblatt) + + Bug fixes: + - Fix the synchronization around explicit tcache creation which could cause + invalid tcache identifiers. This regression was first released in 5.0.0. + (@yoshinorim, @davidtgoldblatt) + - Fix a profiling biasing issue which could cause incorrect heap usage and + object counts. This issue existed in all previous releases with the heap + profiling feature. (@davidtgoldblatt) + - Fix the order of stats counter updating on large realloc which could cause + failed assertions. This regression was first released in 5.0.0. (@azat) + - Fix the locking on the arena destroy mallctl, which could cause concurrent + arena creations to fail. This functionality was first introduced in 5.0.0. + (@interwq) + + Portability improvements: + - Remove nothrow from system function declarations on macOS and FreeBSD. + (@davidtgoldblatt, @fredemmott, @leres) + - Improve overcommit and page alignment settings on NetBSD. (@zoulasc) + - Improve CPU affinity support on BSD platforms. (@devnexen) + - Improve utrace detection and support. (@devnexen) + - Improve QEMU support with MADV_DONTNEED zeroed pages detection. (@azat) + - Add memcntl support on Solaris / illumos. (@devnexen) + - Improve CPU_SPINWAIT on ARM. (@AWSjswinney) + - Improve TSD cleanup on FreeBSD. (@Lapenkov) + - Disable percpu_arena if the CPU count cannot be reliably detected. (@azat) + - Add malloc_size(3) override support. (@devnexen) + - Add mmap VM_MAKE_TAG support. (@devnexen) + - Add support for MADV_[NO]CORE. (@devnexen) + - Add support for DragonFlyBSD. (@devnexen) + - Fix the QUANTUM setting on MIPS64. (@brooksdavis) + - Add the QUANTUM setting for ARC. (@vineetgarc) + - Add the QUANTUM setting for LoongArch. (@wangjl-uos) + - Add QNX support. (@jqian-aurora) + - Avoid atexit(3) calls unless the relevant profiling features are enabled. + (@BusyJay, @laiwei-rice, @interwq) + - Fix unknown option detection when using Clang. (@Lapenkov) + - Fix symbol conflict with musl libc. (@georgthegreat) + - Add -Wimplicit-fallthrough checks. (@nickdesaulniers) + - Add __forceinline support on MSVC. (@santagada) + - Improve FreeBSD and Windows CI support. (@Lapenkov) + - Add CI support for PPC64LE architecture. (@ezeeyahoo) + + Incompatible changes: + - Maximum size class allowed in tcache (opt.[lg_]tcache_max) now has an upper + bound of 8MiB. (@interwq) + + Optimizations and refactors (@davidtgoldblatt, @Lapenkov, @interwq): + - Optimize the common cases of the thread cache operations. + - Optimize internal data structures, including RB tree and pairing heap. + - Optimize the internal locking on extent management. + - Extract and refactor the internal page allocator and interface modules. + + Documentation: + - Fix doc build with --with-install-suffix. (@lawmurray, @interwq) + - Add PROFILING_INTERNALS.md. (@davidtgoldblatt) + - Ensure the proper order of doc building and installation. (@Mingli-Yu) + * 5.2.1 (August 5, 2019) This release is primarily about Windows. A critical virtual memory leak is diff --git a/contrib/jemalloc/FREEBSD-Xlist b/contrib/jemalloc/FREEBSD-Xlist --- a/contrib/jemalloc/FREEBSD-Xlist +++ b/contrib/jemalloc/FREEBSD-Xlist @@ -11,6 +11,7 @@ doc/*.in doc/*.xml doc/*.xsl +doc_internal/ include/jemalloc/internal/atomic_msvc.h include/jemalloc/internal/jemalloc_internal_defs.h.in include/jemalloc/internal/jemalloc_preamble.h.in @@ -45,5 +46,6 @@ run_tests.sh scripts/ src/jemalloc_cpp.cpp +src/ticker.py src/zone.c test/ diff --git a/contrib/jemalloc/FREEBSD-diffs b/contrib/jemalloc/FREEBSD-diffs --- a/contrib/jemalloc/FREEBSD-diffs +++ b/contrib/jemalloc/FREEBSD-diffs @@ -1,5 +1,5 @@ diff --git a/doc/jemalloc.xml.in b/doc/jemalloc.xml.in -index 7fecda7c..d5ca5e86 100644 +index e28e8f38..4f5d2799 100644 --- a/doc/jemalloc.xml.in +++ b/doc/jemalloc.xml.in @@ -53,11 +53,22 @@ @@ -26,7 +26,7 @@ Standard API -@@ -3510,4 +3521,18 @@ malloc_conf = "narenas:1";]]> +@@ -3760,4 +3771,18 @@ malloc_conf = "narenas:1";]]> The posix_memalign() function conforms to IEEE Std 1003.1-2001 (POSIX.1). @@ -46,7 +46,7 @@ + diff --git a/include/jemalloc/internal/jemalloc_internal_decls.h b/include/jemalloc/internal/jemalloc_internal_decls.h -index 7d6053e2..a0e4f5af 100644 +index 983027c8..3cbed72b 100644 --- a/include/jemalloc/internal/jemalloc_internal_decls.h +++ b/include/jemalloc/internal/jemalloc_internal_decls.h @@ -1,6 +1,9 @@ @@ -75,11 +75,11 @@ +# undef JEMALLOC_GCC_U8_SYNC_ATOMICS +#endif diff --git a/include/jemalloc/internal/jemalloc_preamble.h.in b/include/jemalloc/internal/jemalloc_preamble.h.in -index 3418cbfa..53e30dc4 100644 +index 5ce77d96..d970370c 100644 --- a/include/jemalloc/internal/jemalloc_preamble.h.in +++ b/include/jemalloc/internal/jemalloc_preamble.h.in -@@ -8,6 +8,9 @@ - #include +@@ -14,6 +14,9 @@ + # endif #endif +#include "un-namespace.h" @@ -88,7 +88,7 @@ #define JEMALLOC_NO_DEMANGLE #ifdef JEMALLOC_JET # undef JEMALLOC_IS_MALLOC -@@ -79,13 +82,7 @@ static const bool config_fill = +@@ -85,13 +88,7 @@ static const bool config_fill = false #endif ; @@ -104,10 +104,10 @@ static const bool config_prof = #ifdef JEMALLOC_PROF diff --git a/include/jemalloc/internal/mutex.h b/include/jemalloc/internal/mutex.h -index 7c24f072..94af1618 100644 +index 63a0b1b3..8468e7bc 100644 --- a/include/jemalloc/internal/mutex.h +++ b/include/jemalloc/internal/mutex.h -@@ -135,9 +135,6 @@ struct malloc_mutex_s { +@@ -131,9 +131,6 @@ struct malloc_mutex_s { #ifdef JEMALLOC_LAZY_LOCK extern bool isthreaded; @@ -117,7 +117,7 @@ #endif bool malloc_mutex_init(malloc_mutex_t *mutex, const char *name, -@@ -145,6 +142,7 @@ bool malloc_mutex_init(malloc_mutex_t *mutex, const char *name, +@@ -141,6 +138,7 @@ bool malloc_mutex_init(malloc_mutex_t *mutex, const char *name, void malloc_mutex_prefork(tsdn_t *tsdn, malloc_mutex_t *mutex); void malloc_mutex_postfork_parent(tsdn_t *tsdn, malloc_mutex_t *mutex); void malloc_mutex_postfork_child(tsdn_t *tsdn, malloc_mutex_t *mutex); @@ -125,31 +125,37 @@ bool malloc_mutex_boot(void); void malloc_mutex_prof_data_reset(tsdn_t *tsdn, malloc_mutex_t *mutex); -diff --git a/include/jemalloc/internal/test_hooks.h b/include/jemalloc/internal/test_hooks.h -index a6351e59..0780c52f 100644 ---- a/include/jemalloc/internal/test_hooks.h -+++ b/include/jemalloc/internal/test_hooks.h -@@ -6,13 +6,6 @@ extern JEMALLOC_EXPORT void (*test_hooks_libc_hook)(); +diff --git a/include/jemalloc/internal/pac.h b/include/jemalloc/internal/pac.h +index 01c4e6af..9122aa0f 100644 +--- a/include/jemalloc/internal/pac.h ++++ b/include/jemalloc/internal/pac.h +@@ -74,7 +74,6 @@ struct pac_stats_s { + atomic_zu_t abandoned_vm; + }; - #define JEMALLOC_HOOK(fn, hook) ((void)(hook != NULL && (hook(), 0)), fn) +-typedef struct pac_s pac_t; + struct pac_s { + /* + * Must be the first member (we convert it to a PAC given only a +diff --git a/include/jemalloc/internal/san_bump.h b/include/jemalloc/internal/san_bump.h +index 8ec4a710..d5902ba7 100644 +--- a/include/jemalloc/internal/san_bump.h ++++ b/include/jemalloc/internal/san_bump.h +@@ -9,7 +9,6 @@ --#define open JEMALLOC_HOOK(open, test_hooks_libc_hook) --#define read JEMALLOC_HOOK(read, test_hooks_libc_hook) --#define write JEMALLOC_HOOK(write, test_hooks_libc_hook) --#define readlink JEMALLOC_HOOK(readlink, test_hooks_libc_hook) --#define close JEMALLOC_HOOK(close, test_hooks_libc_hook) --#define creat JEMALLOC_HOOK(creat, test_hooks_libc_hook) --#define secure_getenv JEMALLOC_HOOK(secure_getenv, test_hooks_libc_hook) - /* Note that this is undef'd and re-define'd in src/prof.c. */ - #define _Unwind_Backtrace JEMALLOC_HOOK(_Unwind_Backtrace, test_hooks_libc_hook) + extern bool opt_retain; +-typedef struct ehooks_s ehooks_t; + typedef struct pac_s pac_t; + + typedef struct san_bump_alloc_s san_bump_alloc_t; diff --git a/include/jemalloc/internal/tsd.h b/include/jemalloc/internal/tsd.h -index 9ba26004..ecfda5d6 100644 +index 66d68822..6cd52aee 100644 --- a/include/jemalloc/internal/tsd.h +++ b/include/jemalloc/internal/tsd.h -@@ -198,7 +198,8 @@ struct tsd_s { - t TSD_MANGLE(n); - MALLOC_TSD +@@ -256,7 +256,8 @@ struct tsd_s { + TSD_DATA_FAST + TSD_DATA_SLOWER #undef O -}; +/* AddressSanitizer requires TLS data to be aligned to at least 8 bytes. */ @@ -159,10 +165,10 @@ tsd_state_get(tsd_t *tsd) { diff --git a/include/jemalloc/jemalloc_FreeBSD.h b/include/jemalloc/jemalloc_FreeBSD.h new file mode 100644 -index 00000000..b752b0e7 +index 00000000..b23bdba0 --- /dev/null +++ b/include/jemalloc/jemalloc_FreeBSD.h -@@ -0,0 +1,185 @@ +@@ -0,0 +1,192 @@ +/* + * Override settings that were generated in jemalloc_defs.h as necessary. + */ @@ -228,15 +234,21 @@ +#ifdef __powerpc64__ +# define LG_VADDR 64 +# define LG_SIZEOF_PTR 3 ++# define JEMALLOC_TLS_MODEL __attribute__((tls_model("initial-exec"))) +#elif defined(__powerpc__) +# define LG_VADDR 32 +# define LG_SIZEOF_PTR 2 ++# define JEMALLOC_TLS_MODEL __attribute__((tls_model("initial-exec"))) +#endif +#ifdef __riscv +# define LG_VADDR 48 +# define LG_SIZEOF_PTR 3 +#endif + ++#if LG_VADDR > 32 ++# define JEMALLOC_RETAIN ++#endif ++ +#ifndef JEMALLOC_TLS_MODEL +# define JEMALLOC_TLS_MODEL /* Default. */ +#endif @@ -318,6 +330,7 @@ +#define pthread_cond_wait _pthread_cond_wait +#define pthread_cond_timedwait _pthread_cond_timedwait +#define pthread_cond_signal _pthread_cond_signal ++#define pthread_getname_np _pthread_getname_np + +#ifdef JEMALLOC_C_ +/* @@ -360,10 +373,10 @@ +#include "jemalloc_FreeBSD.h" EOF diff --git a/src/jemalloc.c b/src/jemalloc.c -index ed13718d..fefb719a 100644 +index 7655de4e..e4b183d1 100644 --- a/src/jemalloc.c +++ b/src/jemalloc.c -@@ -23,6 +23,10 @@ +@@ -29,6 +29,10 @@ /******************************************************************************/ /* Data. */ @@ -374,33 +387,7 @@ /* Runtime configuration options. */ const char *je_malloc_conf #ifndef _WIN32 -@@ -2660,25 +2664,6 @@ je_realloc(void *ptr, size_t arg_size) { - LOG("core.realloc.entry", "ptr: %p, size: %zu\n", ptr, size); - - if (unlikely(size == 0)) { -- if (ptr != NULL) { -- /* realloc(ptr, 0) is equivalent to free(ptr). */ -- UTRACE(ptr, 0, 0); -- tcache_t *tcache; -- tsd_t *tsd = tsd_fetch(); -- if (tsd_reentrancy_level_get(tsd) == 0) { -- tcache = tcache_get(tsd); -- } else { -- tcache = NULL; -- } -- -- uintptr_t args[3] = {(uintptr_t)ptr, size}; -- hook_invoke_dalloc(hook_dalloc_realloc, ptr, args); -- -- ifree(tsd, ptr, tcache, true); -- -- LOG("core.realloc.exit", "result: %p", NULL); -- return NULL; -- } - size = 1; - } - -@@ -3750,6 +3735,103 @@ je_malloc_usable_size(JEMALLOC_USABLE_SIZE_CONST void *ptr) { +@@ -4297,6 +4301,103 @@ label_done: * End non-standard functions. */ /******************************************************************************/ @@ -504,7 +491,7 @@ /* * The following functions are used by threading libraries for protection of * malloc during fork(). -@@ -3919,4 +4001,11 @@ jemalloc_postfork_child(void) { +@@ -4473,4 +4574,11 @@ jemalloc_postfork_child(void) { ctl_postfork_child(tsd_tsdn(tsd)); } @@ -517,10 +504,10 @@ + /******************************************************************************/ diff --git a/src/malloc_io.c b/src/malloc_io.c -index d7cb0f52..cda589c4 100644 +index b76885cb..93a30497 100644 --- a/src/malloc_io.c +++ b/src/malloc_io.c -@@ -75,6 +75,20 @@ wrtmessage(void *cbopaque, const char *s) { +@@ -73,6 +73,20 @@ wrtmessage(void *cbopaque, const char *s) { JEMALLOC_EXPORT void (*je_malloc_message)(void *, const char *s); @@ -542,10 +529,10 @@ * Wrapper around malloc_message() that avoids the need for * je_malloc_message(...) throughout the code. diff --git a/src/mutex.c b/src/mutex.c -index 3f920f5b..88a7730c 100644 +index 0b3547a8..1eedd626 100644 --- a/src/mutex.c +++ b/src/mutex.c -@@ -41,6 +41,17 @@ pthread_create(pthread_t *__restrict thread, +@@ -46,6 +46,17 @@ pthread_create(pthread_t *__restrict thread, #ifdef JEMALLOC_MUTEX_INIT_CB JEMALLOC_EXPORT int _pthread_mutex_init_calloc_cb(pthread_mutex_t *mutex, void *(calloc_cb)(size_t, size_t)); @@ -563,7 +550,7 @@ #endif void -@@ -131,6 +142,16 @@ mutex_addr_comp(const witness_t *witness1, void *mutex1, +@@ -136,6 +147,16 @@ mutex_addr_comp(const witness_t *witness1, void *mutex1, } } @@ -580,3 +567,45 @@ bool malloc_mutex_init(malloc_mutex_t *mutex, const char *name, witness_rank_t rank, malloc_mutex_lock_order_t lock_order) { +diff --git a/src/pages.c b/src/pages.c +index 8c83a7de..9ccff113 100644 +--- a/src/pages.c ++++ b/src/pages.c +@@ -10,7 +10,9 @@ + #ifdef JEMALLOC_SYSCTL_VM_OVERCOMMIT + #include + #ifdef __FreeBSD__ ++#include + #include ++#include + #endif + #endif + #ifdef __NetBSD__ +@@ -592,6 +594,13 @@ os_overcommits_sysctl(void) { + int vm_overcommit; + size_t sz; + ++#ifdef ELF_BSDF_VMNOOVERCOMMIT ++ int bsdflags; ++ ++ if (_elf_aux_info(AT_BSDFLAGS, &bsdflags, sizeof(bsdflags)) == 0) ++ return ((bsdflags & ELF_BSDF_VMNOOVERCOMMIT) == 0); ++#endif ++ + sz = sizeof(vm_overcommit); + #if defined(__FreeBSD__) && defined(VM_OVERCOMMIT) + int mib[2]; +@@ -607,7 +616,12 @@ os_overcommits_sysctl(void) { + } + #endif + +- return ((vm_overcommit & 0x3) == 0); ++#ifndef SWAP_RESERVE_FORCE_ON ++#define SWAP_RESERVE_FORCE_ON (1 << 0) ++#define SWAP_RESERVE_RLIMIT_ON (1 << 1) ++#endif ++ return ((vm_overcommit & (SWAP_RESERVE_FORCE_ON | ++ SWAP_RESERVE_RLIMIT_ON)) == 0); + } + #endif + diff --git a/contrib/jemalloc/VERSION b/contrib/jemalloc/VERSION --- a/contrib/jemalloc/VERSION +++ b/contrib/jemalloc/VERSION @@ -1 +1 @@ -5.2.1-0-gea6b3e973b477b8061e0076bb257dbd7f3faa756 +5.3.0-0-g54eaed1d8b56b1aa528be3bdd1877e59c56fa90c diff --git a/contrib/jemalloc/include/jemalloc/internal/activity_callback.h b/contrib/jemalloc/include/jemalloc/internal/activity_callback.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/activity_callback.h @@ -0,0 +1,23 @@ +#ifndef JEMALLOC_INTERNAL_ACTIVITY_CALLBACK_H +#define JEMALLOC_INTERNAL_ACTIVITY_CALLBACK_H + +/* + * The callback to be executed "periodically", in response to some amount of + * allocator activity. + * + * This callback need not be computing any sort of peak (although that's the + * intended first use case), but we drive it from the peak counter, so it's + * keeps things tidy to keep it here. + * + * The calls to this thunk get driven by the peak_event module. + */ +#define ACTIVITY_CALLBACK_THUNK_INITIALIZER {NULL, NULL} +typedef void (*activity_callback_t)(void *uctx, uint64_t allocated, + uint64_t deallocated); +typedef struct activity_callback_thunk_s activity_callback_thunk_t; +struct activity_callback_thunk_s { + activity_callback_t callback; + void *uctx; +}; + +#endif /* JEMALLOC_INTERNAL_ACTIVITY_CALLBACK_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/arena_externs.h b/contrib/jemalloc/include/jemalloc/internal/arena_externs.h --- a/contrib/jemalloc/include/jemalloc/internal/arena_externs.h +++ b/contrib/jemalloc/include/jemalloc/internal/arena_externs.h @@ -2,59 +2,67 @@ #define JEMALLOC_INTERNAL_ARENA_EXTERNS_H #include "jemalloc/internal/bin.h" +#include "jemalloc/internal/div.h" #include "jemalloc/internal/extent_dss.h" #include "jemalloc/internal/hook.h" #include "jemalloc/internal/pages.h" #include "jemalloc/internal/stats.h" +/* + * When the amount of pages to be purged exceeds this amount, deferred purge + * should happen. + */ +#define ARENA_DEFERRED_PURGE_NPAGES_THRESHOLD UINT64_C(1024) + extern ssize_t opt_dirty_decay_ms; extern ssize_t opt_muzzy_decay_ms; extern percpu_arena_mode_t opt_percpu_arena; extern const char *percpu_arena_mode_names[]; -extern const uint64_t h_steps[SMOOTHSTEP_NSTEPS]; +extern div_info_t arena_binind_div_info[SC_NBINS]; + extern malloc_mutex_t arenas_lock; +extern emap_t arena_emap_global; extern size_t opt_oversize_threshold; extern size_t oversize_threshold; +/* + * arena_bin_offsets[binind] is the offset of the first bin shard for size class + * binind. + */ +extern uint32_t arena_bin_offsets[SC_NBINS]; + void arena_basic_stats_merge(tsdn_t *tsdn, arena_t *arena, unsigned *nthreads, const char **dss, ssize_t *dirty_decay_ms, ssize_t *muzzy_decay_ms, size_t *nactive, size_t *ndirty, size_t *nmuzzy); void arena_stats_merge(tsdn_t *tsdn, arena_t *arena, unsigned *nthreads, const char **dss, ssize_t *dirty_decay_ms, ssize_t *muzzy_decay_ms, size_t *nactive, size_t *ndirty, size_t *nmuzzy, arena_stats_t *astats, - bin_stats_t *bstats, arena_stats_large_t *lstats, - arena_stats_extents_t *estats); -void arena_extents_dirty_dalloc(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, extent_t *extent); -#ifdef JEMALLOC_JET -size_t arena_slab_regind(extent_t *slab, szind_t binind, const void *ptr); -#endif -extent_t *arena_extent_alloc_large(tsdn_t *tsdn, arena_t *arena, - size_t usize, size_t alignment, bool *zero); + bin_stats_data_t *bstats, arena_stats_large_t *lstats, + pac_estats_t *estats, hpa_shard_stats_t *hpastats, sec_stats_t *secstats); +void arena_handle_deferred_work(tsdn_t *tsdn, arena_t *arena); +edata_t *arena_extent_alloc_large(tsdn_t *tsdn, arena_t *arena, + size_t usize, size_t alignment, bool zero); void arena_extent_dalloc_large_prep(tsdn_t *tsdn, arena_t *arena, - extent_t *extent); + edata_t *edata); void arena_extent_ralloc_large_shrink(tsdn_t *tsdn, arena_t *arena, - extent_t *extent, size_t oldsize); + edata_t *edata, size_t oldsize); void arena_extent_ralloc_large_expand(tsdn_t *tsdn, arena_t *arena, - extent_t *extent, size_t oldsize); -ssize_t arena_dirty_decay_ms_get(arena_t *arena); -bool arena_dirty_decay_ms_set(tsdn_t *tsdn, arena_t *arena, ssize_t decay_ms); -ssize_t arena_muzzy_decay_ms_get(arena_t *arena); -bool arena_muzzy_decay_ms_set(tsdn_t *tsdn, arena_t *arena, ssize_t decay_ms); + edata_t *edata, size_t oldsize); +bool arena_decay_ms_set(tsdn_t *tsdn, arena_t *arena, extent_state_t state, + ssize_t decay_ms); +ssize_t arena_decay_ms_get(arena_t *arena, extent_state_t state); void arena_decay(tsdn_t *tsdn, arena_t *arena, bool is_background_thread, bool all); +uint64_t arena_time_until_deferred(tsdn_t *tsdn, arena_t *arena); +void arena_do_deferred_work(tsdn_t *tsdn, arena_t *arena); void arena_reset(tsd_t *tsd, arena_t *arena); void arena_destroy(tsd_t *tsd, arena_t *arena); -void arena_tcache_fill_small(tsdn_t *tsdn, arena_t *arena, tcache_t *tcache, - cache_bin_t *tbin, szind_t binind, uint64_t prof_accumbytes); -void arena_alloc_junk_small(void *ptr, const bin_info_t *bin_info, - bool zero); - -typedef void (arena_dalloc_junk_small_t)(void *, const bin_info_t *); -extern arena_dalloc_junk_small_t *JET_MUTABLE arena_dalloc_junk_small; +void arena_cache_bin_fill_small(tsdn_t *tsdn, arena_t *arena, + cache_bin_t *cache_bin, cache_bin_info_t *cache_bin_info, szind_t binind, + const unsigned nfill); void *arena_malloc_hard(tsdn_t *tsdn, arena_t *arena, size_t size, szind_t ind, bool zero); @@ -63,8 +71,12 @@ void arena_prof_promote(tsdn_t *tsdn, void *ptr, size_t usize); void arena_dalloc_promoted(tsdn_t *tsdn, void *ptr, tcache_t *tcache, bool slow_path); -void arena_dalloc_bin_junked_locked(tsdn_t *tsdn, arena_t *arena, bin_t *bin, - szind_t binind, extent_t *extent, void *ptr); +void arena_slab_dalloc(tsdn_t *tsdn, arena_t *arena, edata_t *slab); + +void arena_dalloc_bin_locked_handle_newly_empty(tsdn_t *tsdn, arena_t *arena, + edata_t *slab, bin_t *bin); +void arena_dalloc_bin_locked_handle_newly_nonempty(tsdn_t *tsdn, arena_t *arena, + edata_t *slab, bin_t *bin); void arena_dalloc_small(tsdn_t *tsdn, void *ptr); bool arena_ralloc_no_move(tsdn_t *tsdn, void *ptr, size_t oldsize, size_t size, size_t extra, bool zero, size_t *newsize); @@ -72,6 +84,9 @@ size_t size, size_t alignment, bool zero, tcache_t *tcache, hook_ralloc_args_t *hook_args); dss_prec_t arena_dss_prec_get(arena_t *arena); +ehooks_t *arena_get_ehooks(arena_t *arena); +extent_hooks_t *arena_set_extent_hooks(tsd_t *tsd, arena_t *arena, + extent_hooks_t *extent_hooks); bool arena_dss_prec_set(arena_t *arena, dss_prec_t dss_prec); ssize_t arena_dirty_decay_ms_default_get(void); bool arena_dirty_decay_ms_default_set(ssize_t decay_ms); @@ -82,14 +97,15 @@ unsigned arena_nthreads_get(arena_t *arena, bool internal); void arena_nthreads_inc(arena_t *arena, bool internal); void arena_nthreads_dec(arena_t *arena, bool internal); -size_t arena_extent_sn_next(arena_t *arena); -arena_t *arena_new(tsdn_t *tsdn, unsigned ind, extent_hooks_t *extent_hooks); +arena_t *arena_new(tsdn_t *tsdn, unsigned ind, const arena_config_t *config); bool arena_init_huge(void); bool arena_is_huge(unsigned arena_ind); arena_t *arena_choose_huge(tsd_t *tsd); -bin_t *arena_bin_choose_lock(tsdn_t *tsdn, arena_t *arena, szind_t binind, +bin_t *arena_bin_choose(tsdn_t *tsdn, arena_t *arena, szind_t binind, unsigned *binshard); -void arena_boot(sc_data_t *sc_data); +size_t arena_fill_small_fresh(tsdn_t *tsdn, arena_t *arena, szind_t binind, + void **ptrs, size_t nfill, bool zero); +bool arena_boot(sc_data_t *sc_data, base_t *base, bool hpa); void arena_prefork0(tsdn_t *tsdn, arena_t *arena); void arena_prefork1(tsdn_t *tsdn, arena_t *arena); void arena_prefork2(tsdn_t *tsdn, arena_t *arena); @@ -98,6 +114,7 @@ void arena_prefork5(tsdn_t *tsdn, arena_t *arena); void arena_prefork6(tsdn_t *tsdn, arena_t *arena); void arena_prefork7(tsdn_t *tsdn, arena_t *arena); +void arena_prefork8(tsdn_t *tsdn, arena_t *arena); void arena_postfork_parent(tsdn_t *tsdn, arena_t *arena); void arena_postfork_child(tsdn_t *tsdn, arena_t *arena); diff --git a/contrib/jemalloc/include/jemalloc/internal/arena_inlines_a.h b/contrib/jemalloc/include/jemalloc/internal/arena_inlines_a.h --- a/contrib/jemalloc/include/jemalloc/internal/arena_inlines_a.h +++ b/contrib/jemalloc/include/jemalloc/internal/arena_inlines_a.h @@ -3,7 +3,7 @@ static inline unsigned arena_ind_get(const arena_t *arena) { - return base_ind_get(arena->base); + return arena->ind; } static inline void @@ -21,37 +21,4 @@ return atomic_load_zu(&arena->stats.internal, ATOMIC_RELAXED); } -static inline bool -arena_prof_accum(tsdn_t *tsdn, arena_t *arena, uint64_t accumbytes) { - cassert(config_prof); - - if (likely(prof_interval == 0 || !prof_active_get_unlocked())) { - return false; - } - - return prof_accum_add(tsdn, &arena->prof_accum, accumbytes); -} - -static inline void -percpu_arena_update(tsd_t *tsd, unsigned cpu) { - assert(have_percpu_arena); - arena_t *oldarena = tsd_arena_get(tsd); - assert(oldarena != NULL); - unsigned oldind = arena_ind_get(oldarena); - - if (oldind != cpu) { - unsigned newind = cpu; - arena_t *newarena = arena_get(tsd_tsdn(tsd), newind, true); - assert(newarena != NULL); - - /* Set new arena/tcache associations. */ - arena_migrate(tsd, oldind, newind); - tcache_t *tcache = tcache_get(tsd); - if (tcache != NULL) { - tcache_arena_reassociate(tsd_tsdn(tsd), tcache, - newarena); - } - } -} - #endif /* JEMALLOC_INTERNAL_ARENA_INLINES_A_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/arena_inlines_b.h b/contrib/jemalloc/include/jemalloc/internal/arena_inlines_b.h --- a/contrib/jemalloc/include/jemalloc/internal/arena_inlines_b.h +++ b/contrib/jemalloc/include/jemalloc/internal/arena_inlines_b.h @@ -1,16 +1,20 @@ #ifndef JEMALLOC_INTERNAL_ARENA_INLINES_B_H #define JEMALLOC_INTERNAL_ARENA_INLINES_B_H +#include "jemalloc/internal/div.h" +#include "jemalloc/internal/emap.h" #include "jemalloc/internal/jemalloc_internal_types.h" #include "jemalloc/internal/mutex.h" #include "jemalloc/internal/rtree.h" +#include "jemalloc/internal/safety_check.h" #include "jemalloc/internal/sc.h" #include "jemalloc/internal/sz.h" #include "jemalloc/internal/ticker.h" -JEMALLOC_ALWAYS_INLINE bool -arena_has_default_hooks(arena_t *arena) { - return (extent_hooks_get(arena) == &extent_hooks_default); +static inline arena_t * +arena_get_from_edata(edata_t *edata) { + return (arena_t *)atomic_load_p(&arenas[edata_arena_ind_get(edata)], + ATOMIC_RELAXED); } JEMALLOC_ALWAYS_INLINE arena_t * @@ -34,127 +38,109 @@ return arena_choose(tsd, NULL); } -JEMALLOC_ALWAYS_INLINE prof_tctx_t * -arena_prof_tctx_get(tsdn_t *tsdn, const void *ptr, alloc_ctx_t *alloc_ctx) { +JEMALLOC_ALWAYS_INLINE void +arena_prof_info_get(tsd_t *tsd, const void *ptr, emap_alloc_ctx_t *alloc_ctx, + prof_info_t *prof_info, bool reset_recent) { cassert(config_prof); assert(ptr != NULL); + assert(prof_info != NULL); + + edata_t *edata = NULL; + bool is_slab; /* Static check. */ if (alloc_ctx == NULL) { - const extent_t *extent = iealloc(tsdn, ptr); - if (unlikely(!extent_slab_get(extent))) { - return large_prof_tctx_get(tsdn, extent); - } + edata = emap_edata_lookup(tsd_tsdn(tsd), &arena_emap_global, + ptr); + is_slab = edata_slab_get(edata); + } else if (unlikely(!(is_slab = alloc_ctx->slab))) { + edata = emap_edata_lookup(tsd_tsdn(tsd), &arena_emap_global, + ptr); + } + + if (unlikely(!is_slab)) { + /* edata must have been initialized at this point. */ + assert(edata != NULL); + large_prof_info_get(tsd, edata, prof_info, reset_recent); } else { - if (unlikely(!alloc_ctx->slab)) { - return large_prof_tctx_get(tsdn, iealloc(tsdn, ptr)); - } + prof_info->alloc_tctx = (prof_tctx_t *)(uintptr_t)1U; + /* + * No need to set other fields in prof_info; they will never be + * accessed if (uintptr_t)alloc_tctx == (uintptr_t)1U. + */ } - return (prof_tctx_t *)(uintptr_t)1U; } JEMALLOC_ALWAYS_INLINE void -arena_prof_tctx_set(tsdn_t *tsdn, const void *ptr, size_t usize, - alloc_ctx_t *alloc_ctx, prof_tctx_t *tctx) { +arena_prof_tctx_reset(tsd_t *tsd, const void *ptr, + emap_alloc_ctx_t *alloc_ctx) { cassert(config_prof); assert(ptr != NULL); /* Static check. */ if (alloc_ctx == NULL) { - extent_t *extent = iealloc(tsdn, ptr); - if (unlikely(!extent_slab_get(extent))) { - large_prof_tctx_set(tsdn, extent, tctx); + edata_t *edata = emap_edata_lookup(tsd_tsdn(tsd), + &arena_emap_global, ptr); + if (unlikely(!edata_slab_get(edata))) { + large_prof_tctx_reset(edata); } } else { if (unlikely(!alloc_ctx->slab)) { - large_prof_tctx_set(tsdn, iealloc(tsdn, ptr), tctx); + edata_t *edata = emap_edata_lookup(tsd_tsdn(tsd), + &arena_emap_global, ptr); + large_prof_tctx_reset(edata); } } } -static inline void -arena_prof_tctx_reset(tsdn_t *tsdn, const void *ptr, prof_tctx_t *tctx) { +JEMALLOC_ALWAYS_INLINE void +arena_prof_tctx_reset_sampled(tsd_t *tsd, const void *ptr) { cassert(config_prof); assert(ptr != NULL); - extent_t *extent = iealloc(tsdn, ptr); - assert(!extent_slab_get(extent)); + edata_t *edata = emap_edata_lookup(tsd_tsdn(tsd), &arena_emap_global, + ptr); + assert(!edata_slab_get(edata)); - large_prof_tctx_reset(tsdn, extent); -} - -JEMALLOC_ALWAYS_INLINE nstime_t -arena_prof_alloc_time_get(tsdn_t *tsdn, const void *ptr, - alloc_ctx_t *alloc_ctx) { - cassert(config_prof); - assert(ptr != NULL); - - extent_t *extent = iealloc(tsdn, ptr); - /* - * Unlike arena_prof_prof_tctx_{get, set}, we only call this once we're - * sure we have a sampled allocation. - */ - assert(!extent_slab_get(extent)); - return large_prof_alloc_time_get(extent); + large_prof_tctx_reset(edata); } JEMALLOC_ALWAYS_INLINE void -arena_prof_alloc_time_set(tsdn_t *tsdn, const void *ptr, alloc_ctx_t *alloc_ctx, - nstime_t t) { +arena_prof_info_set(tsd_t *tsd, edata_t *edata, prof_tctx_t *tctx, + size_t size) { cassert(config_prof); - assert(ptr != NULL); - extent_t *extent = iealloc(tsdn, ptr); - assert(!extent_slab_get(extent)); - large_prof_alloc_time_set(extent, t); + assert(!edata_slab_get(edata)); + large_prof_info_set(edata, tctx, size); } JEMALLOC_ALWAYS_INLINE void arena_decay_ticks(tsdn_t *tsdn, arena_t *arena, unsigned nticks) { - tsd_t *tsd; - ticker_t *decay_ticker; - if (unlikely(tsdn_null(tsdn))) { return; } - tsd = tsdn_tsd(tsdn); - decay_ticker = decay_ticker_get(tsd, arena_ind_get(arena)); - if (unlikely(decay_ticker == NULL)) { - return; - } - if (unlikely(ticker_ticks(decay_ticker, nticks))) { + tsd_t *tsd = tsdn_tsd(tsdn); + /* + * We use the ticker_geom_t to avoid having per-arena state in the tsd. + * Instead of having a countdown-until-decay timer running for every + * arena in every thread, we flip a coin once per tick, whose + * probability of coming up heads is 1/nticks; this is effectively the + * operation of the ticker_geom_t. Each arena has the same chance of a + * coinflip coming up heads (1/ARENA_DECAY_NTICKS_PER_UPDATE), so we can + * use a single ticker for all of them. + */ + ticker_geom_t *decay_ticker = tsd_arena_decay_tickerp_get(tsd); + uint64_t *prng_state = tsd_prng_statep_get(tsd); + if (unlikely(ticker_geom_ticks(decay_ticker, prng_state, nticks))) { arena_decay(tsdn, arena, false, false); } } JEMALLOC_ALWAYS_INLINE void arena_decay_tick(tsdn_t *tsdn, arena_t *arena) { - malloc_mutex_assert_not_owner(tsdn, &arena->decay_dirty.mtx); - malloc_mutex_assert_not_owner(tsdn, &arena->decay_muzzy.mtx); - arena_decay_ticks(tsdn, arena, 1); } -/* Purge a single extent to retained / unmapped directly. */ -JEMALLOC_ALWAYS_INLINE void -arena_decay_extent(tsdn_t *tsdn,arena_t *arena, extent_hooks_t **r_extent_hooks, - extent_t *extent) { - size_t extent_size = extent_size_get(extent); - extent_dalloc_wrapper(tsdn, arena, - r_extent_hooks, extent); - if (config_stats) { - /* Update stats accordingly. */ - arena_stats_lock(tsdn, &arena->stats); - arena_stats_add_u64(tsdn, &arena->stats, - &arena->decay_dirty.stats->nmadvise, 1); - arena_stats_add_u64(tsdn, &arena->stats, - &arena->decay_dirty.stats->purged, extent_size >> LG_PAGE); - arena_stats_sub_zu(tsdn, &arena->stats, &arena->stats.mapped, - extent_size); - arena_stats_unlock(tsdn, &arena->stats); - } -} - JEMALLOC_ALWAYS_INLINE void * arena_malloc(tsdn_t *tsdn, arena_t *arena, size_t size, szind_t ind, bool zero, tcache_t *tcache, bool slow_path) { @@ -178,21 +164,19 @@ JEMALLOC_ALWAYS_INLINE arena_t * arena_aalloc(tsdn_t *tsdn, const void *ptr) { - return extent_arena_get(iealloc(tsdn, ptr)); + edata_t *edata = emap_edata_lookup(tsdn, &arena_emap_global, ptr); + unsigned arena_ind = edata_arena_ind_get(edata); + return (arena_t *)atomic_load_p(&arenas[arena_ind], ATOMIC_RELAXED); } JEMALLOC_ALWAYS_INLINE size_t arena_salloc(tsdn_t *tsdn, const void *ptr) { assert(ptr != NULL); + emap_alloc_ctx_t alloc_ctx; + emap_alloc_ctx_lookup(tsdn, &arena_emap_global, ptr, &alloc_ctx); + assert(alloc_ctx.szind != SC_NSIZES); - rtree_ctx_t rtree_ctx_fallback; - rtree_ctx_t *rtree_ctx = tsdn_rtree_ctx(tsdn, &rtree_ctx_fallback); - - szind_t szind = rtree_szind_read(tsdn, &extents_rtree, rtree_ctx, - (uintptr_t)ptr, true); - assert(szind != SC_NSIZES); - - return sz_index2size(szind); + return sz_index2size(alloc_ctx.szind); } JEMALLOC_ALWAYS_INLINE size_t @@ -206,26 +190,53 @@ * failure. */ - rtree_ctx_t rtree_ctx_fallback; - rtree_ctx_t *rtree_ctx = tsdn_rtree_ctx(tsdn, &rtree_ctx_fallback); - - extent_t *extent; - szind_t szind; - if (rtree_extent_szind_read(tsdn, &extents_rtree, rtree_ctx, - (uintptr_t)ptr, false, &extent, &szind)) { + emap_full_alloc_ctx_t full_alloc_ctx; + bool missing = emap_full_alloc_ctx_try_lookup(tsdn, &arena_emap_global, + ptr, &full_alloc_ctx); + if (missing) { return 0; } - if (extent == NULL) { + if (full_alloc_ctx.edata == NULL) { return 0; } - assert(extent_state_get(extent) == extent_state_active); + assert(edata_state_get(full_alloc_ctx.edata) == extent_state_active); /* Only slab members should be looked up via interior pointers. */ - assert(extent_addr_get(extent) == ptr || extent_slab_get(extent)); + assert(edata_addr_get(full_alloc_ctx.edata) == ptr + || edata_slab_get(full_alloc_ctx.edata)); + + assert(full_alloc_ctx.szind != SC_NSIZES); + + return sz_index2size(full_alloc_ctx.szind); +} - assert(szind != SC_NSIZES); +JEMALLOC_ALWAYS_INLINE bool +large_dalloc_safety_checks(edata_t *edata, void *ptr, szind_t szind) { + if (!config_opt_safety_checks) { + return false; + } + + /* + * Eagerly detect double free and sized dealloc bugs for large sizes. + * The cost is low enough (as edata will be accessed anyway) to be + * enabled all the time. + */ + if (unlikely(edata == NULL || + edata_state_get(edata) != extent_state_active)) { + safety_check_fail("Invalid deallocation detected: " + "pages being freed (%p) not currently active, " + "possibly caused by double free bugs.", + (uintptr_t)edata_addr_get(edata)); + return true; + } + size_t input_size = sz_index2size(szind); + if (unlikely(input_size != edata_usize_get(edata))) { + safety_check_fail_sized_dealloc(/* current_dealloc */ true, ptr, + /* true_size */ edata_usize_get(edata), input_size); + return true; + } - return sz_index2size(szind); + return false; } static inline void @@ -233,8 +244,13 @@ if (config_prof && unlikely(szind < SC_NBINS)) { arena_dalloc_promoted(tsdn, ptr, NULL, true); } else { - extent_t *extent = iealloc(tsdn, ptr); - large_dalloc(tsdn, extent); + edata_t *edata = emap_edata_lookup(tsdn, &arena_emap_global, + ptr); + if (large_dalloc_safety_checks(edata, ptr, szind)) { + /* See the comment in isfree. */ + return; + } + large_dalloc(tsdn, edata); } } @@ -242,27 +258,22 @@ arena_dalloc_no_tcache(tsdn_t *tsdn, void *ptr) { assert(ptr != NULL); - rtree_ctx_t rtree_ctx_fallback; - rtree_ctx_t *rtree_ctx = tsdn_rtree_ctx(tsdn, &rtree_ctx_fallback); - - szind_t szind; - bool slab; - rtree_szind_slab_read(tsdn, &extents_rtree, rtree_ctx, (uintptr_t)ptr, - true, &szind, &slab); + emap_alloc_ctx_t alloc_ctx; + emap_alloc_ctx_lookup(tsdn, &arena_emap_global, ptr, &alloc_ctx); if (config_debug) { - extent_t *extent = rtree_extent_read(tsdn, &extents_rtree, - rtree_ctx, (uintptr_t)ptr, true); - assert(szind == extent_szind_get(extent)); - assert(szind < SC_NSIZES); - assert(slab == extent_slab_get(extent)); + edata_t *edata = emap_edata_lookup(tsdn, &arena_emap_global, + ptr); + assert(alloc_ctx.szind == edata_szind_get(edata)); + assert(alloc_ctx.szind < SC_NSIZES); + assert(alloc_ctx.slab == edata_slab_get(edata)); } - if (likely(slab)) { + if (likely(alloc_ctx.slab)) { /* Small allocation. */ arena_dalloc_small(tsdn, ptr); } else { - arena_dalloc_large_no_tcache(tsdn, ptr, szind); + arena_dalloc_large_no_tcache(tsdn, ptr, alloc_ctx.szind); } } @@ -277,14 +288,19 @@ slow_path); } } else { - extent_t *extent = iealloc(tsdn, ptr); - large_dalloc(tsdn, extent); + edata_t *edata = emap_edata_lookup(tsdn, &arena_emap_global, + ptr); + if (large_dalloc_safety_checks(edata, ptr, szind)) { + /* See the comment in isfree. */ + return; + } + large_dalloc(tsdn, edata); } } JEMALLOC_ALWAYS_INLINE void arena_dalloc(tsdn_t *tsdn, void *ptr, tcache_t *tcache, - alloc_ctx_t *alloc_ctx, bool slow_path) { + emap_alloc_ctx_t *caller_alloc_ctx, bool slow_path) { assert(!tsdn_null(tsdn) || tcache == NULL); assert(ptr != NULL); @@ -293,34 +309,30 @@ return; } - szind_t szind; - bool slab; - rtree_ctx_t *rtree_ctx; - if (alloc_ctx != NULL) { - szind = alloc_ctx->szind; - slab = alloc_ctx->slab; - assert(szind != SC_NSIZES); + emap_alloc_ctx_t alloc_ctx; + if (caller_alloc_ctx != NULL) { + alloc_ctx = *caller_alloc_ctx; } else { - rtree_ctx = tsd_rtree_ctx(tsdn_tsd(tsdn)); - rtree_szind_slab_read(tsdn, &extents_rtree, rtree_ctx, - (uintptr_t)ptr, true, &szind, &slab); + util_assume(!tsdn_null(tsdn)); + emap_alloc_ctx_lookup(tsdn, &arena_emap_global, ptr, + &alloc_ctx); } if (config_debug) { - rtree_ctx = tsd_rtree_ctx(tsdn_tsd(tsdn)); - extent_t *extent = rtree_extent_read(tsdn, &extents_rtree, - rtree_ctx, (uintptr_t)ptr, true); - assert(szind == extent_szind_get(extent)); - assert(szind < SC_NSIZES); - assert(slab == extent_slab_get(extent)); + edata_t *edata = emap_edata_lookup(tsdn, &arena_emap_global, + ptr); + assert(alloc_ctx.szind == edata_szind_get(edata)); + assert(alloc_ctx.szind < SC_NSIZES); + assert(alloc_ctx.slab == edata_slab_get(edata)); } - if (likely(slab)) { + if (likely(alloc_ctx.slab)) { /* Small allocation. */ - tcache_dalloc_small(tsdn_tsd(tsdn), tcache, ptr, szind, - slow_path); + tcache_dalloc_small(tsdn_tsd(tsdn), tcache, ptr, + alloc_ctx.szind, slow_path); } else { - arena_dalloc_large(tsdn, ptr, tcache, szind, slow_path); + arena_dalloc_large(tsdn, ptr, tcache, alloc_ctx.szind, + slow_path); } } @@ -329,47 +341,43 @@ assert(ptr != NULL); assert(size <= SC_LARGE_MAXCLASS); - szind_t szind; - bool slab; + emap_alloc_ctx_t alloc_ctx; if (!config_prof || !opt_prof) { /* * There is no risk of being confused by a promoted sampled * object, so base szind and slab on the given size. */ - szind = sz_size2index(size); - slab = (szind < SC_NBINS); + alloc_ctx.szind = sz_size2index(size); + alloc_ctx.slab = (alloc_ctx.szind < SC_NBINS); } if ((config_prof && opt_prof) || config_debug) { - rtree_ctx_t rtree_ctx_fallback; - rtree_ctx_t *rtree_ctx = tsdn_rtree_ctx(tsdn, - &rtree_ctx_fallback); - - rtree_szind_slab_read(tsdn, &extents_rtree, rtree_ctx, - (uintptr_t)ptr, true, &szind, &slab); + emap_alloc_ctx_lookup(tsdn, &arena_emap_global, ptr, + &alloc_ctx); - assert(szind == sz_size2index(size)); - assert((config_prof && opt_prof) || slab == (szind < SC_NBINS)); + assert(alloc_ctx.szind == sz_size2index(size)); + assert((config_prof && opt_prof) + || alloc_ctx.slab == (alloc_ctx.szind < SC_NBINS)); if (config_debug) { - extent_t *extent = rtree_extent_read(tsdn, - &extents_rtree, rtree_ctx, (uintptr_t)ptr, true); - assert(szind == extent_szind_get(extent)); - assert(slab == extent_slab_get(extent)); + edata_t *edata = emap_edata_lookup(tsdn, + &arena_emap_global, ptr); + assert(alloc_ctx.szind == edata_szind_get(edata)); + assert(alloc_ctx.slab == edata_slab_get(edata)); } } - if (likely(slab)) { + if (likely(alloc_ctx.slab)) { /* Small allocation. */ arena_dalloc_small(tsdn, ptr); } else { - arena_dalloc_large_no_tcache(tsdn, ptr, szind); + arena_dalloc_large_no_tcache(tsdn, ptr, alloc_ctx.szind); } } JEMALLOC_ALWAYS_INLINE void arena_sdalloc(tsdn_t *tsdn, void *ptr, size_t size, tcache_t *tcache, - alloc_ctx_t *alloc_ctx, bool slow_path) { + emap_alloc_ctx_t *caller_alloc_ctx, bool slow_path) { assert(!tsdn_null(tsdn) || tcache == NULL); assert(ptr != NULL); assert(size <= SC_LARGE_MAXCLASS); @@ -379,49 +387,164 @@ return; } - szind_t szind; - bool slab; - alloc_ctx_t local_ctx; + emap_alloc_ctx_t alloc_ctx; if (config_prof && opt_prof) { - if (alloc_ctx == NULL) { + if (caller_alloc_ctx == NULL) { /* Uncommon case and should be a static check. */ - rtree_ctx_t rtree_ctx_fallback; - rtree_ctx_t *rtree_ctx = tsdn_rtree_ctx(tsdn, - &rtree_ctx_fallback); - rtree_szind_slab_read(tsdn, &extents_rtree, rtree_ctx, - (uintptr_t)ptr, true, &local_ctx.szind, - &local_ctx.slab); - assert(local_ctx.szind == sz_size2index(size)); - alloc_ctx = &local_ctx; + emap_alloc_ctx_lookup(tsdn, &arena_emap_global, ptr, + &alloc_ctx); + assert(alloc_ctx.szind == sz_size2index(size)); + } else { + alloc_ctx = *caller_alloc_ctx; } - slab = alloc_ctx->slab; - szind = alloc_ctx->szind; } else { /* * There is no risk of being confused by a promoted sampled * object, so base szind and slab on the given size. */ - szind = sz_size2index(size); - slab = (szind < SC_NBINS); + alloc_ctx.szind = sz_size2index(size); + alloc_ctx.slab = (alloc_ctx.szind < SC_NBINS); } if (config_debug) { - rtree_ctx_t *rtree_ctx = tsd_rtree_ctx(tsdn_tsd(tsdn)); - rtree_szind_slab_read(tsdn, &extents_rtree, rtree_ctx, - (uintptr_t)ptr, true, &szind, &slab); - extent_t *extent = rtree_extent_read(tsdn, - &extents_rtree, rtree_ctx, (uintptr_t)ptr, true); - assert(szind == extent_szind_get(extent)); - assert(slab == extent_slab_get(extent)); + edata_t *edata = emap_edata_lookup(tsdn, &arena_emap_global, + ptr); + assert(alloc_ctx.szind == edata_szind_get(edata)); + assert(alloc_ctx.slab == edata_slab_get(edata)); } - if (likely(slab)) { + if (likely(alloc_ctx.slab)) { /* Small allocation. */ - tcache_dalloc_small(tsdn_tsd(tsdn), tcache, ptr, szind, - slow_path); + tcache_dalloc_small(tsdn_tsd(tsdn), tcache, ptr, + alloc_ctx.szind, slow_path); } else { - arena_dalloc_large(tsdn, ptr, tcache, szind, slow_path); + arena_dalloc_large(tsdn, ptr, tcache, alloc_ctx.szind, + slow_path); + } +} + +static inline void +arena_cache_oblivious_randomize(tsdn_t *tsdn, arena_t *arena, edata_t *edata, + size_t alignment) { + assert(edata_base_get(edata) == edata_addr_get(edata)); + + if (alignment < PAGE) { + unsigned lg_range = LG_PAGE - + lg_floor(CACHELINE_CEILING(alignment)); + size_t r; + if (!tsdn_null(tsdn)) { + tsd_t *tsd = tsdn_tsd(tsdn); + r = (size_t)prng_lg_range_u64( + tsd_prng_statep_get(tsd), lg_range); + } else { + uint64_t stack_value = (uint64_t)(uintptr_t)&r; + r = (size_t)prng_lg_range_u64(&stack_value, lg_range); + } + uintptr_t random_offset = ((uintptr_t)r) << (LG_PAGE - + lg_range); + edata->e_addr = (void *)((uintptr_t)edata->e_addr + + random_offset); + assert(ALIGNMENT_ADDR2BASE(edata->e_addr, alignment) == + edata->e_addr); + } +} + +/* + * The dalloc bin info contains just the information that the common paths need + * during tcache flushes. By force-inlining these paths, and using local copies + * of data (so that the compiler knows it's constant), we avoid a whole bunch of + * redundant loads and stores by leaving this information in registers. + */ +typedef struct arena_dalloc_bin_locked_info_s arena_dalloc_bin_locked_info_t; +struct arena_dalloc_bin_locked_info_s { + div_info_t div_info; + uint32_t nregs; + uint64_t ndalloc; +}; + +JEMALLOC_ALWAYS_INLINE size_t +arena_slab_regind(arena_dalloc_bin_locked_info_t *info, szind_t binind, + edata_t *slab, const void *ptr) { + size_t diff, regind; + + /* Freeing a pointer outside the slab can cause assertion failure. */ + assert((uintptr_t)ptr >= (uintptr_t)edata_addr_get(slab)); + assert((uintptr_t)ptr < (uintptr_t)edata_past_get(slab)); + /* Freeing an interior pointer can cause assertion failure. */ + assert(((uintptr_t)ptr - (uintptr_t)edata_addr_get(slab)) % + (uintptr_t)bin_infos[binind].reg_size == 0); + + diff = (size_t)((uintptr_t)ptr - (uintptr_t)edata_addr_get(slab)); + + /* Avoid doing division with a variable divisor. */ + regind = div_compute(&info->div_info, diff); + + assert(regind < bin_infos[binind].nregs); + + return regind; +} + +JEMALLOC_ALWAYS_INLINE void +arena_dalloc_bin_locked_begin(arena_dalloc_bin_locked_info_t *info, + szind_t binind) { + info->div_info = arena_binind_div_info[binind]; + info->nregs = bin_infos[binind].nregs; + info->ndalloc = 0; +} + +/* + * Does the deallocation work associated with freeing a single pointer (a + * "step") in between a arena_dalloc_bin_locked begin and end call. + * + * Returns true if arena_slab_dalloc must be called on slab. Doesn't do + * stats updates, which happen during finish (this lets running counts get left + * in a register). + */ +JEMALLOC_ALWAYS_INLINE bool +arena_dalloc_bin_locked_step(tsdn_t *tsdn, arena_t *arena, bin_t *bin, + arena_dalloc_bin_locked_info_t *info, szind_t binind, edata_t *slab, + void *ptr) { + const bin_info_t *bin_info = &bin_infos[binind]; + size_t regind = arena_slab_regind(info, binind, slab, ptr); + slab_data_t *slab_data = edata_slab_data_get(slab); + + assert(edata_nfree_get(slab) < bin_info->nregs); + /* Freeing an unallocated pointer can cause assertion failure. */ + assert(bitmap_get(slab_data->bitmap, &bin_info->bitmap_info, regind)); + + bitmap_unset(slab_data->bitmap, &bin_info->bitmap_info, regind); + edata_nfree_inc(slab); + + if (config_stats) { + info->ndalloc++; + } + + unsigned nfree = edata_nfree_get(slab); + if (nfree == bin_info->nregs) { + arena_dalloc_bin_locked_handle_newly_empty(tsdn, arena, slab, + bin); + return true; + } else if (nfree == 1 && slab != bin->slabcur) { + arena_dalloc_bin_locked_handle_newly_nonempty(tsdn, arena, slab, + bin); } + return false; +} + +JEMALLOC_ALWAYS_INLINE void +arena_dalloc_bin_locked_finish(tsdn_t *tsdn, arena_t *arena, bin_t *bin, + arena_dalloc_bin_locked_info_t *info) { + if (config_stats) { + bin->stats.ndalloc += info->ndalloc; + assert(bin->stats.curregs >= (size_t)info->ndalloc); + bin->stats.curregs -= (size_t)info->ndalloc; + } +} + +static inline bin_t * +arena_get_bin(arena_t *arena, szind_t binind, unsigned binshard) { + bin_t *shard0 = (bin_t *)((uintptr_t)arena + arena_bin_offsets[binind]); + return shard0 + binshard; } #endif /* JEMALLOC_INTERNAL_ARENA_INLINES_B_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/arena_stats.h b/contrib/jemalloc/include/jemalloc/internal/arena_stats.h --- a/contrib/jemalloc/include/jemalloc/internal/arena_stats.h +++ b/contrib/jemalloc/include/jemalloc/internal/arena_stats.h @@ -2,77 +2,41 @@ #define JEMALLOC_INTERNAL_ARENA_STATS_H #include "jemalloc/internal/atomic.h" +#include "jemalloc/internal/lockedint.h" #include "jemalloc/internal/mutex.h" #include "jemalloc/internal/mutex_prof.h" +#include "jemalloc/internal/pa.h" #include "jemalloc/internal/sc.h" JEMALLOC_DIAGNOSTIC_DISABLE_SPURIOUS -/* - * In those architectures that support 64-bit atomics, we use atomic updates for - * our 64-bit values. Otherwise, we use a plain uint64_t and synchronize - * externally. - */ -#ifdef JEMALLOC_ATOMIC_U64 -typedef atomic_u64_t arena_stats_u64_t; -#else -/* Must hold the arena stats mutex while reading atomically. */ -typedef uint64_t arena_stats_u64_t; -#endif - typedef struct arena_stats_large_s arena_stats_large_t; struct arena_stats_large_s { /* * Total number of allocation/deallocation requests served directly by * the arena. */ - arena_stats_u64_t nmalloc; - arena_stats_u64_t ndalloc; + locked_u64_t nmalloc; + locked_u64_t ndalloc; /* * Number of allocation requests that correspond to this size class. * This includes requests served by tcache, though tcache only * periodically merges into this counter. */ - arena_stats_u64_t nrequests; /* Partially derived. */ + locked_u64_t nrequests; /* Partially derived. */ /* * Number of tcache fills / flushes for large (similarly, periodically * merged). Note that there is no large tcache batch-fill currently * (i.e. only fill 1 at a time); however flush may be batched. */ - arena_stats_u64_t nfills; /* Partially derived. */ - arena_stats_u64_t nflushes; /* Partially derived. */ + locked_u64_t nfills; /* Partially derived. */ + locked_u64_t nflushes; /* Partially derived. */ /* Current number of allocations of this size class. */ size_t curlextents; /* Derived. */ }; -typedef struct arena_stats_decay_s arena_stats_decay_t; -struct arena_stats_decay_s { - /* Total number of purge sweeps. */ - arena_stats_u64_t npurge; - /* Total number of madvise calls made. */ - arena_stats_u64_t nmadvise; - /* Total number of pages purged. */ - arena_stats_u64_t purged; -}; - -typedef struct arena_stats_extents_s arena_stats_extents_t; -struct arena_stats_extents_s { - /* - * Stats for a given index in the range [0, SC_NPSIZES] in an extents_t. - * We track both bytes and # of extents: two extents in the same bucket - * may have different sizes if adjacent size classes differ by more than - * a page, so bytes cannot always be derived from # of extents. - */ - atomic_zu_t ndirty; - atomic_zu_t dirty_bytes; - atomic_zu_t nmuzzy; - atomic_zu_t muzzy_bytes; - atomic_zu_t nretained; - atomic_zu_t retained_bytes; -}; - /* * Arena stats. Note that fields marked "derived" are not directly maintained * within the arena code; rather their values are derived during stats merge @@ -80,43 +44,36 @@ */ typedef struct arena_stats_s arena_stats_t; struct arena_stats_s { -#ifndef JEMALLOC_ATOMIC_U64 - malloc_mutex_t mtx; -#endif - - /* Number of bytes currently mapped, excluding retained memory. */ - atomic_zu_t mapped; /* Partially derived. */ + LOCKEDINT_MTX_DECLARE(mtx) /* - * Number of unused virtual memory bytes currently retained. Retained - * bytes are technically mapped (though always decommitted or purged), - * but they are excluded from the mapped statistic (above). + * resident includes the base stats -- that's why it lives here and not + * in pa_shard_stats_t. */ - atomic_zu_t retained; /* Derived. */ - - /* Number of extent_t structs allocated by base, but not being used. */ - atomic_zu_t extent_avail; - - arena_stats_decay_t decay_dirty; - arena_stats_decay_t decay_muzzy; + size_t base; /* Derived. */ + size_t resident; /* Derived. */ + size_t metadata_thp; /* Derived. */ + size_t mapped; /* Derived. */ - atomic_zu_t base; /* Derived. */ atomic_zu_t internal; - atomic_zu_t resident; /* Derived. */ - atomic_zu_t metadata_thp; - atomic_zu_t allocated_large; /* Derived. */ - arena_stats_u64_t nmalloc_large; /* Derived. */ - arena_stats_u64_t ndalloc_large; /* Derived. */ - arena_stats_u64_t nfills_large; /* Derived. */ - arena_stats_u64_t nflushes_large; /* Derived. */ - arena_stats_u64_t nrequests_large; /* Derived. */ + size_t allocated_large; /* Derived. */ + uint64_t nmalloc_large; /* Derived. */ + uint64_t ndalloc_large; /* Derived. */ + uint64_t nfills_large; /* Derived. */ + uint64_t nflushes_large; /* Derived. */ + uint64_t nrequests_large; /* Derived. */ - /* VM space had to be leaked (undocumented). Normally 0. */ - atomic_zu_t abandoned_vm; + /* + * The stats logically owned by the pa_shard in the same arena. This + * lives here only because it's convenient for the purposes of the ctl + * module -- it only knows about the single arena_stats. + */ + pa_shard_stats_t pa_shard_stats; /* Number of bytes cached in tcache associated with this arena. */ - atomic_zu_t tcache_bytes; /* Derived. */ + size_t tcache_bytes; /* Derived. */ + size_t tcache_stashed_bytes; /* Derived. */ mutex_prof_data_t mutex_prof_data[mutex_prof_num_arena_mutexes]; @@ -134,138 +91,24 @@ assert(((char *)arena_stats)[i] == 0); } } -#ifndef JEMALLOC_ATOMIC_U64 - if (malloc_mutex_init(&arena_stats->mtx, "arena_stats", + if (LOCKEDINT_MTX_INIT(arena_stats->mtx, "arena_stats", WITNESS_RANK_ARENA_STATS, malloc_mutex_rank_exclusive)) { return true; } -#endif /* Memory is zeroed, so there is no need to clear stats. */ return false; } -static inline void -arena_stats_lock(tsdn_t *tsdn, arena_stats_t *arena_stats) { -#ifndef JEMALLOC_ATOMIC_U64 - malloc_mutex_lock(tsdn, &arena_stats->mtx); -#endif -} - -static inline void -arena_stats_unlock(tsdn_t *tsdn, arena_stats_t *arena_stats) { -#ifndef JEMALLOC_ATOMIC_U64 - malloc_mutex_unlock(tsdn, &arena_stats->mtx); -#endif -} - -static inline uint64_t -arena_stats_read_u64(tsdn_t *tsdn, arena_stats_t *arena_stats, - arena_stats_u64_t *p) { -#ifdef JEMALLOC_ATOMIC_U64 - return atomic_load_u64(p, ATOMIC_RELAXED); -#else - malloc_mutex_assert_owner(tsdn, &arena_stats->mtx); - return *p; -#endif -} - -static inline void -arena_stats_add_u64(tsdn_t *tsdn, arena_stats_t *arena_stats, - arena_stats_u64_t *p, uint64_t x) { -#ifdef JEMALLOC_ATOMIC_U64 - atomic_fetch_add_u64(p, x, ATOMIC_RELAXED); -#else - malloc_mutex_assert_owner(tsdn, &arena_stats->mtx); - *p += x; -#endif -} - -static inline void -arena_stats_sub_u64(tsdn_t *tsdn, arena_stats_t *arena_stats, - arena_stats_u64_t *p, uint64_t x) { -#ifdef JEMALLOC_ATOMIC_U64 - uint64_t r = atomic_fetch_sub_u64(p, x, ATOMIC_RELAXED); - assert(r - x <= r); -#else - malloc_mutex_assert_owner(tsdn, &arena_stats->mtx); - *p -= x; - assert(*p + x >= *p); -#endif -} - -/* - * Non-atomically sets *dst += src. *dst needs external synchronization. - * This lets us avoid the cost of a fetch_add when its unnecessary (note that - * the types here are atomic). - */ -static inline void -arena_stats_accum_u64(arena_stats_u64_t *dst, uint64_t src) { -#ifdef JEMALLOC_ATOMIC_U64 - uint64_t cur_dst = atomic_load_u64(dst, ATOMIC_RELAXED); - atomic_store_u64(dst, src + cur_dst, ATOMIC_RELAXED); -#else - *dst += src; -#endif -} - -static inline size_t -arena_stats_read_zu(tsdn_t *tsdn, arena_stats_t *arena_stats, - atomic_zu_t *p) { -#ifdef JEMALLOC_ATOMIC_U64 - return atomic_load_zu(p, ATOMIC_RELAXED); -#else - malloc_mutex_assert_owner(tsdn, &arena_stats->mtx); - return atomic_load_zu(p, ATOMIC_RELAXED); -#endif -} - -static inline void -arena_stats_add_zu(tsdn_t *tsdn, arena_stats_t *arena_stats, - atomic_zu_t *p, size_t x) { -#ifdef JEMALLOC_ATOMIC_U64 - atomic_fetch_add_zu(p, x, ATOMIC_RELAXED); -#else - malloc_mutex_assert_owner(tsdn, &arena_stats->mtx); - size_t cur = atomic_load_zu(p, ATOMIC_RELAXED); - atomic_store_zu(p, cur + x, ATOMIC_RELAXED); -#endif -} - -static inline void -arena_stats_sub_zu(tsdn_t *tsdn, arena_stats_t *arena_stats, - atomic_zu_t *p, size_t x) { -#ifdef JEMALLOC_ATOMIC_U64 - size_t r = atomic_fetch_sub_zu(p, x, ATOMIC_RELAXED); - assert(r - x <= r); -#else - malloc_mutex_assert_owner(tsdn, &arena_stats->mtx); - size_t cur = atomic_load_zu(p, ATOMIC_RELAXED); - atomic_store_zu(p, cur - x, ATOMIC_RELAXED); -#endif -} - -/* Like the _u64 variant, needs an externally synchronized *dst. */ -static inline void -arena_stats_accum_zu(atomic_zu_t *dst, size_t src) { - size_t cur_dst = atomic_load_zu(dst, ATOMIC_RELAXED); - atomic_store_zu(dst, src + cur_dst, ATOMIC_RELAXED); -} - static inline void arena_stats_large_flush_nrequests_add(tsdn_t *tsdn, arena_stats_t *arena_stats, szind_t szind, uint64_t nrequests) { - arena_stats_lock(tsdn, arena_stats); + LOCKEDINT_MTX_LOCK(tsdn, arena_stats->mtx); arena_stats_large_t *lstats = &arena_stats->lstats[szind - SC_NBINS]; - arena_stats_add_u64(tsdn, arena_stats, &lstats->nrequests, nrequests); - arena_stats_add_u64(tsdn, arena_stats, &lstats->nflushes, 1); - arena_stats_unlock(tsdn, arena_stats); -} - -static inline void -arena_stats_mapped_add(tsdn_t *tsdn, arena_stats_t *arena_stats, size_t size) { - arena_stats_lock(tsdn, arena_stats); - arena_stats_add_zu(tsdn, arena_stats, &arena_stats->mapped, size); - arena_stats_unlock(tsdn, arena_stats); + locked_inc_u64(tsdn, LOCKEDINT_MTX(arena_stats->mtx), + &lstats->nrequests, nrequests); + locked_inc_u64(tsdn, LOCKEDINT_MTX(arena_stats->mtx), + &lstats->nflushes, 1); + LOCKEDINT_MTX_UNLOCK(tsdn, arena_stats->mtx); } #endif /* JEMALLOC_INTERNAL_ARENA_STATS_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/arena_structs.h b/contrib/jemalloc/include/jemalloc/internal/arena_structs.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/arena_structs.h @@ -0,0 +1,101 @@ +#ifndef JEMALLOC_INTERNAL_ARENA_STRUCTS_H +#define JEMALLOC_INTERNAL_ARENA_STRUCTS_H + +#include "jemalloc/internal/arena_stats.h" +#include "jemalloc/internal/atomic.h" +#include "jemalloc/internal/bin.h" +#include "jemalloc/internal/bitmap.h" +#include "jemalloc/internal/counter.h" +#include "jemalloc/internal/ecache.h" +#include "jemalloc/internal/edata_cache.h" +#include "jemalloc/internal/extent_dss.h" +#include "jemalloc/internal/jemalloc_internal_types.h" +#include "jemalloc/internal/mutex.h" +#include "jemalloc/internal/nstime.h" +#include "jemalloc/internal/pa.h" +#include "jemalloc/internal/ql.h" +#include "jemalloc/internal/sc.h" +#include "jemalloc/internal/ticker.h" + +struct arena_s { + /* + * Number of threads currently assigned to this arena. Each thread has + * two distinct assignments, one for application-serving allocation, and + * the other for internal metadata allocation. Internal metadata must + * not be allocated from arenas explicitly created via the arenas.create + * mallctl, because the arena..reset mallctl indiscriminately + * discards all allocations for the affected arena. + * + * 0: Application allocation. + * 1: Internal metadata allocation. + * + * Synchronization: atomic. + */ + atomic_u_t nthreads[2]; + + /* Next bin shard for binding new threads. Synchronization: atomic. */ + atomic_u_t binshard_next; + + /* + * When percpu_arena is enabled, to amortize the cost of reading / + * updating the current CPU id, track the most recent thread accessing + * this arena, and only read CPU if there is a mismatch. + */ + tsdn_t *last_thd; + + /* Synchronization: internal. */ + arena_stats_t stats; + + /* + * Lists of tcaches and cache_bin_array_descriptors for extant threads + * associated with this arena. Stats from these are merged + * incrementally, and at exit if opt_stats_print is enabled. + * + * Synchronization: tcache_ql_mtx. + */ + ql_head(tcache_slow_t) tcache_ql; + ql_head(cache_bin_array_descriptor_t) cache_bin_array_descriptor_ql; + malloc_mutex_t tcache_ql_mtx; + + /* + * Represents a dss_prec_t, but atomically. + * + * Synchronization: atomic. + */ + atomic_u_t dss_prec; + + /* + * Extant large allocations. + * + * Synchronization: large_mtx. + */ + edata_list_active_t large; + /* Synchronizes all large allocation/update/deallocation. */ + malloc_mutex_t large_mtx; + + /* The page-level allocator shard this arena uses. */ + pa_shard_t pa_shard; + + /* + * A cached copy of base->ind. This can get accessed on hot paths; + * looking it up in base requires an extra pointer hop / cache miss. + */ + unsigned ind; + + /* + * Base allocator, from which arena metadata are allocated. + * + * Synchronization: internal. + */ + base_t *base; + /* Used to determine uptime. Read-only after initialization. */ + nstime_t create_time; + + /* + * The arena is allocated alongside its bins; really this is a + * dynamically sized array determined by the binshard settings. + */ + bin_t bins[0]; +}; + +#endif /* JEMALLOC_INTERNAL_ARENA_STRUCTS_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/arena_structs_a.h b/contrib/jemalloc/include/jemalloc/internal/arena_structs_a.h deleted file mode 100644 --- a/contrib/jemalloc/include/jemalloc/internal/arena_structs_a.h +++ /dev/null @@ -1,11 +0,0 @@ -#ifndef JEMALLOC_INTERNAL_ARENA_STRUCTS_A_H -#define JEMALLOC_INTERNAL_ARENA_STRUCTS_A_H - -#include "jemalloc/internal/bitmap.h" - -struct arena_slab_data_s { - /* Per region allocated/deallocated bitmap. */ - bitmap_t bitmap[BITMAP_GROUPS_MAX]; -}; - -#endif /* JEMALLOC_INTERNAL_ARENA_STRUCTS_A_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/arena_structs_b.h b/contrib/jemalloc/include/jemalloc/internal/arena_structs_b.h deleted file mode 100644 --- a/contrib/jemalloc/include/jemalloc/internal/arena_structs_b.h +++ /dev/null @@ -1,232 +0,0 @@ -#ifndef JEMALLOC_INTERNAL_ARENA_STRUCTS_B_H -#define JEMALLOC_INTERNAL_ARENA_STRUCTS_B_H - -#include "jemalloc/internal/arena_stats.h" -#include "jemalloc/internal/atomic.h" -#include "jemalloc/internal/bin.h" -#include "jemalloc/internal/bitmap.h" -#include "jemalloc/internal/extent_dss.h" -#include "jemalloc/internal/jemalloc_internal_types.h" -#include "jemalloc/internal/mutex.h" -#include "jemalloc/internal/nstime.h" -#include "jemalloc/internal/ql.h" -#include "jemalloc/internal/sc.h" -#include "jemalloc/internal/smoothstep.h" -#include "jemalloc/internal/ticker.h" - -struct arena_decay_s { - /* Synchronizes all non-atomic fields. */ - malloc_mutex_t mtx; - /* - * True if a thread is currently purging the extents associated with - * this decay structure. - */ - bool purging; - /* - * Approximate time in milliseconds from the creation of a set of unused - * dirty pages until an equivalent set of unused dirty pages is purged - * and/or reused. - */ - atomic_zd_t time_ms; - /* time / SMOOTHSTEP_NSTEPS. */ - nstime_t interval; - /* - * Time at which the current decay interval logically started. We do - * not actually advance to a new epoch until sometime after it starts - * because of scheduling and computation delays, and it is even possible - * to completely skip epochs. In all cases, during epoch advancement we - * merge all relevant activity into the most recently recorded epoch. - */ - nstime_t epoch; - /* Deadline randomness generator. */ - uint64_t jitter_state; - /* - * Deadline for current epoch. This is the sum of interval and per - * epoch jitter which is a uniform random variable in [0..interval). - * Epochs always advance by precise multiples of interval, but we - * randomize the deadline to reduce the likelihood of arenas purging in - * lockstep. - */ - nstime_t deadline; - /* - * Number of unpurged pages at beginning of current epoch. During epoch - * advancement we use the delta between arena->decay_*.nunpurged and - * extents_npages_get(&arena->extents_*) to determine how many dirty - * pages, if any, were generated. - */ - size_t nunpurged; - /* - * Trailing log of how many unused dirty pages were generated during - * each of the past SMOOTHSTEP_NSTEPS decay epochs, where the last - * element is the most recent epoch. Corresponding epoch times are - * relative to epoch. - */ - size_t backlog[SMOOTHSTEP_NSTEPS]; - - /* - * Pointer to associated stats. These stats are embedded directly in - * the arena's stats due to how stats structures are shared between the - * arena and ctl code. - * - * Synchronization: Same as associated arena's stats field. */ - arena_stats_decay_t *stats; - /* Peak number of pages in associated extents. Used for debug only. */ - uint64_t ceil_npages; -}; - -struct arena_s { - /* - * Number of threads currently assigned to this arena. Each thread has - * two distinct assignments, one for application-serving allocation, and - * the other for internal metadata allocation. Internal metadata must - * not be allocated from arenas explicitly created via the arenas.create - * mallctl, because the arena..reset mallctl indiscriminately - * discards all allocations for the affected arena. - * - * 0: Application allocation. - * 1: Internal metadata allocation. - * - * Synchronization: atomic. - */ - atomic_u_t nthreads[2]; - - /* Next bin shard for binding new threads. Synchronization: atomic. */ - atomic_u_t binshard_next; - - /* - * When percpu_arena is enabled, to amortize the cost of reading / - * updating the current CPU id, track the most recent thread accessing - * this arena, and only read CPU if there is a mismatch. - */ - tsdn_t *last_thd; - - /* Synchronization: internal. */ - arena_stats_t stats; - - /* - * Lists of tcaches and cache_bin_array_descriptors for extant threads - * associated with this arena. Stats from these are merged - * incrementally, and at exit if opt_stats_print is enabled. - * - * Synchronization: tcache_ql_mtx. - */ - ql_head(tcache_t) tcache_ql; - ql_head(cache_bin_array_descriptor_t) cache_bin_array_descriptor_ql; - malloc_mutex_t tcache_ql_mtx; - - /* Synchronization: internal. */ - prof_accum_t prof_accum; - - /* - * PRNG state for cache index randomization of large allocation base - * pointers. - * - * Synchronization: atomic. - */ - atomic_zu_t offset_state; - - /* - * Extent serial number generator state. - * - * Synchronization: atomic. - */ - atomic_zu_t extent_sn_next; - - /* - * Represents a dss_prec_t, but atomically. - * - * Synchronization: atomic. - */ - atomic_u_t dss_prec; - - /* - * Number of pages in active extents. - * - * Synchronization: atomic. - */ - atomic_zu_t nactive; - - /* - * Extant large allocations. - * - * Synchronization: large_mtx. - */ - extent_list_t large; - /* Synchronizes all large allocation/update/deallocation. */ - malloc_mutex_t large_mtx; - - /* - * Collections of extents that were previously allocated. These are - * used when allocating extents, in an attempt to re-use address space. - * - * Synchronization: internal. - */ - extents_t extents_dirty; - extents_t extents_muzzy; - extents_t extents_retained; - - /* - * Decay-based purging state, responsible for scheduling extent state - * transitions. - * - * Synchronization: internal. - */ - arena_decay_t decay_dirty; /* dirty --> muzzy */ - arena_decay_t decay_muzzy; /* muzzy --> retained */ - - /* - * Next extent size class in a growing series to use when satisfying a - * request via the extent hooks (only if opt_retain). This limits the - * number of disjoint virtual memory ranges so that extent merging can - * be effective even if multiple arenas' extent allocation requests are - * highly interleaved. - * - * retain_grow_limit is the max allowed size ind to expand (unless the - * required size is greater). Default is no limit, and controlled - * through mallctl only. - * - * Synchronization: extent_grow_mtx - */ - pszind_t extent_grow_next; - pszind_t retain_grow_limit; - malloc_mutex_t extent_grow_mtx; - - /* - * Available extent structures that were allocated via - * base_alloc_extent(). - * - * Synchronization: extent_avail_mtx. - */ - extent_tree_t extent_avail; - atomic_zu_t extent_avail_cnt; - malloc_mutex_t extent_avail_mtx; - - /* - * bins is used to store heaps of free regions. - * - * Synchronization: internal. - */ - bins_t bins[SC_NBINS]; - - /* - * Base allocator, from which arena metadata are allocated. - * - * Synchronization: internal. - */ - base_t *base; - /* Used to determine uptime. Read-only after initialization. */ - nstime_t create_time; -}; - -/* Used in conjunction with tsd for fast arena-related context lookup. */ -struct arena_tdata_s { - ticker_t decay_ticker; -}; - -/* Used to pass rtree lookup context down the path. */ -struct alloc_ctx_s { - szind_t szind; - bool slab; -}; - -#endif /* JEMALLOC_INTERNAL_ARENA_STRUCTS_B_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/arena_types.h b/contrib/jemalloc/include/jemalloc/internal/arena_types.h --- a/contrib/jemalloc/include/jemalloc/internal/arena_types.h +++ b/contrib/jemalloc/include/jemalloc/internal/arena_types.h @@ -3,21 +3,14 @@ #include "jemalloc/internal/sc.h" -/* Maximum number of regions in one slab. */ -#define LG_SLAB_MAXREGS (LG_PAGE - SC_LG_TINY_MIN) -#define SLAB_MAXREGS (1U << LG_SLAB_MAXREGS) - /* Default decay times in milliseconds. */ #define DIRTY_DECAY_MS_DEFAULT ZD(10 * 1000) #define MUZZY_DECAY_MS_DEFAULT (0) /* Number of event ticks between time checks. */ -#define DECAY_NTICKS_PER_UPDATE 1000 +#define ARENA_DECAY_NTICKS_PER_UPDATE 1000 -typedef struct arena_slab_data_s arena_slab_data_t; typedef struct arena_decay_s arena_decay_t; typedef struct arena_s arena_t; -typedef struct arena_tdata_s arena_tdata_t; -typedef struct alloc_ctx_s alloc_ctx_t; typedef enum { percpu_arena_mode_names_base = 0, /* Used for options processing. */ @@ -48,4 +41,18 @@ */ #define OVERSIZE_THRESHOLD_DEFAULT (8 << 20) +struct arena_config_s { + /* extent hooks to be used for the arena */ + extent_hooks_t *extent_hooks; + + /* + * Use extent hooks for metadata (base) allocations when true. + */ + bool metadata_use_hooks; +}; + +typedef struct arena_config_s arena_config_t; + +extern const arena_config_t arena_config_default; + #endif /* JEMALLOC_INTERNAL_ARENA_TYPES_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/atomic.h b/contrib/jemalloc/include/jemalloc/internal/atomic.h --- a/contrib/jemalloc/include/jemalloc/internal/atomic.h +++ b/contrib/jemalloc/include/jemalloc/internal/atomic.h @@ -51,6 +51,27 @@ #define ATOMIC_ACQ_REL atomic_memory_order_acq_rel #define ATOMIC_SEQ_CST atomic_memory_order_seq_cst +/* + * Another convenience -- simple atomic helper functions. + */ +#define JEMALLOC_GENERATE_EXPANDED_INT_ATOMICS(type, short_type, \ + lg_size) \ + JEMALLOC_GENERATE_INT_ATOMICS(type, short_type, lg_size) \ + ATOMIC_INLINE void \ + atomic_load_add_store_##short_type(atomic_##short_type##_t *a, \ + type inc) { \ + type oldval = atomic_load_##short_type(a, ATOMIC_RELAXED); \ + type newval = oldval + inc; \ + atomic_store_##short_type(a, newval, ATOMIC_RELAXED); \ + } \ + ATOMIC_INLINE void \ + atomic_load_sub_store_##short_type(atomic_##short_type##_t *a, \ + type inc) { \ + type oldval = atomic_load_##short_type(a, ATOMIC_RELAXED); \ + type newval = oldval - inc; \ + atomic_store_##short_type(a, newval, ATOMIC_RELAXED); \ + } + /* * Not all platforms have 64-bit atomics. If we do, this #define exposes that * fact. @@ -67,18 +88,18 @@ */ JEMALLOC_GENERATE_ATOMICS(bool, b, 0) -JEMALLOC_GENERATE_INT_ATOMICS(unsigned, u, LG_SIZEOF_INT) +JEMALLOC_GENERATE_EXPANDED_INT_ATOMICS(unsigned, u, LG_SIZEOF_INT) -JEMALLOC_GENERATE_INT_ATOMICS(size_t, zu, LG_SIZEOF_PTR) +JEMALLOC_GENERATE_EXPANDED_INT_ATOMICS(size_t, zu, LG_SIZEOF_PTR) -JEMALLOC_GENERATE_INT_ATOMICS(ssize_t, zd, LG_SIZEOF_PTR) +JEMALLOC_GENERATE_EXPANDED_INT_ATOMICS(ssize_t, zd, LG_SIZEOF_PTR) -JEMALLOC_GENERATE_INT_ATOMICS(uint8_t, u8, 0) +JEMALLOC_GENERATE_EXPANDED_INT_ATOMICS(uint8_t, u8, 0) -JEMALLOC_GENERATE_INT_ATOMICS(uint32_t, u32, 2) +JEMALLOC_GENERATE_EXPANDED_INT_ATOMICS(uint32_t, u32, 2) #ifdef JEMALLOC_ATOMIC_U64 -JEMALLOC_GENERATE_INT_ATOMICS(uint64_t, u64, 3) +JEMALLOC_GENERATE_EXPANDED_INT_ATOMICS(uint64_t, u64, 3) #endif #undef ATOMIC_INLINE diff --git a/contrib/jemalloc/include/jemalloc/internal/background_thread_externs.h b/contrib/jemalloc/include/jemalloc/internal/background_thread_externs.h --- a/contrib/jemalloc/include/jemalloc/internal/background_thread_externs.h +++ b/contrib/jemalloc/include/jemalloc/internal/background_thread_externs.h @@ -12,8 +12,9 @@ bool background_thread_create(tsd_t *tsd, unsigned arena_ind); bool background_threads_enable(tsd_t *tsd); bool background_threads_disable(tsd_t *tsd); -void background_thread_interval_check(tsdn_t *tsdn, arena_t *arena, - arena_decay_t *decay, size_t npages_new); +bool background_thread_is_started(background_thread_info_t* info); +void background_thread_wakeup_early(background_thread_info_t *info, + nstime_t *remaining_sleep); void background_thread_prefork0(tsdn_t *tsdn); void background_thread_prefork1(tsdn_t *tsdn); void background_thread_postfork_parent(tsdn_t *tsdn); @@ -27,6 +28,6 @@ void *(*)(void *), void *__restrict); #endif bool background_thread_boot0(void); -bool background_thread_boot1(tsdn_t *tsdn); +bool background_thread_boot1(tsdn_t *tsdn, base_t *base); #endif /* JEMALLOC_INTERNAL_BACKGROUND_THREAD_EXTERNS_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/background_thread_inlines.h b/contrib/jemalloc/include/jemalloc/internal/background_thread_inlines.h --- a/contrib/jemalloc/include/jemalloc/internal/background_thread_inlines.h +++ b/contrib/jemalloc/include/jemalloc/internal/background_thread_inlines.h @@ -45,18 +45,4 @@ return atomic_load_b(&info->indefinite_sleep, ATOMIC_ACQUIRE); } -JEMALLOC_ALWAYS_INLINE void -arena_background_thread_inactivity_check(tsdn_t *tsdn, arena_t *arena, - bool is_background_thread) { - if (!background_thread_enabled() || is_background_thread) { - return; - } - background_thread_info_t *info = - arena_background_thread_info_get(arena); - if (background_thread_indefinite_sleep(info)) { - background_thread_interval_check(tsdn, arena, - &arena->decay_dirty, 0); - } -} - #endif /* JEMALLOC_INTERNAL_BACKGROUND_THREAD_INLINES_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/background_thread_structs.h b/contrib/jemalloc/include/jemalloc/internal/background_thread_structs.h --- a/contrib/jemalloc/include/jemalloc/internal/background_thread_structs.h +++ b/contrib/jemalloc/include/jemalloc/internal/background_thread_structs.h @@ -11,6 +11,17 @@ #define MAX_BACKGROUND_THREAD_LIMIT MALLOCX_ARENA_LIMIT #define DEFAULT_NUM_BACKGROUND_THREAD 4 +/* + * These exist only as a transitional state. Eventually, deferral should be + * part of the PAI, and each implementation can indicate wait times with more + * specificity. + */ +#define BACKGROUND_THREAD_HPA_INTERVAL_MAX_UNINITIALIZED (-2) +#define BACKGROUND_THREAD_HPA_INTERVAL_MAX_DEFAULT_WHEN_ENABLED 5000 + +#define BACKGROUND_THREAD_DEFERRED_MIN UINT64_C(0) +#define BACKGROUND_THREAD_DEFERRED_MAX UINT64_MAX + typedef enum { background_thread_stopped, background_thread_started, @@ -48,6 +59,7 @@ size_t num_threads; uint64_t num_runs; nstime_t run_interval; + mutex_prof_data_t max_counter_per_bg_thd; }; typedef struct background_thread_stats_s background_thread_stats_t; diff --git a/contrib/jemalloc/include/jemalloc/internal/base.h b/contrib/jemalloc/include/jemalloc/internal/base.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/base.h @@ -0,0 +1,110 @@ +#ifndef JEMALLOC_INTERNAL_BASE_H +#define JEMALLOC_INTERNAL_BASE_H + +#include "jemalloc/internal/edata.h" +#include "jemalloc/internal/ehooks.h" +#include "jemalloc/internal/mutex.h" + +enum metadata_thp_mode_e { + metadata_thp_disabled = 0, + /* + * Lazily enable hugepage for metadata. To avoid high RSS caused by THP + * + low usage arena (i.e. THP becomes a significant percentage), the + * "auto" option only starts using THP after a base allocator used up + * the first THP region. Starting from the second hugepage (in a single + * arena), "auto" behaves the same as "always", i.e. madvise hugepage + * right away. + */ + metadata_thp_auto = 1, + metadata_thp_always = 2, + metadata_thp_mode_limit = 3 +}; +typedef enum metadata_thp_mode_e metadata_thp_mode_t; + +#define METADATA_THP_DEFAULT metadata_thp_disabled +extern metadata_thp_mode_t opt_metadata_thp; +extern const char *metadata_thp_mode_names[]; + + +/* Embedded at the beginning of every block of base-managed virtual memory. */ +typedef struct base_block_s base_block_t; +struct base_block_s { + /* Total size of block's virtual memory mapping. */ + size_t size; + + /* Next block in list of base's blocks. */ + base_block_t *next; + + /* Tracks unused trailing space. */ + edata_t edata; +}; + +typedef struct base_s base_t; +struct base_s { + /* + * User-configurable extent hook functions. + */ + ehooks_t ehooks; + + /* + * User-configurable extent hook functions for metadata allocations. + */ + ehooks_t ehooks_base; + + /* Protects base_alloc() and base_stats_get() operations. */ + malloc_mutex_t mtx; + + /* Using THP when true (metadata_thp auto mode). */ + bool auto_thp_switched; + /* + * Most recent size class in the series of increasingly large base + * extents. Logarithmic spacing between subsequent allocations ensures + * that the total number of distinct mappings remains small. + */ + pszind_t pind_last; + + /* Serial number generation state. */ + size_t extent_sn_next; + + /* Chain of all blocks associated with base. */ + base_block_t *blocks; + + /* Heap of extents that track unused trailing space within blocks. */ + edata_heap_t avail[SC_NSIZES]; + + /* Stats, only maintained if config_stats. */ + size_t allocated; + size_t resident; + size_t mapped; + /* Number of THP regions touched. */ + size_t n_thp; +}; + +static inline unsigned +base_ind_get(const base_t *base) { + return ehooks_ind_get(&base->ehooks); +} + +static inline bool +metadata_thp_enabled(void) { + return (opt_metadata_thp != metadata_thp_disabled); +} + +base_t *b0get(void); +base_t *base_new(tsdn_t *tsdn, unsigned ind, + const extent_hooks_t *extent_hooks, bool metadata_use_hooks); +void base_delete(tsdn_t *tsdn, base_t *base); +ehooks_t *base_ehooks_get(base_t *base); +ehooks_t *base_ehooks_get_for_metadata(base_t *base); +extent_hooks_t *base_extent_hooks_set(base_t *base, + extent_hooks_t *extent_hooks); +void *base_alloc(tsdn_t *tsdn, base_t *base, size_t size, size_t alignment); +edata_t *base_alloc_edata(tsdn_t *tsdn, base_t *base); +void base_stats_get(tsdn_t *tsdn, base_t *base, size_t *allocated, + size_t *resident, size_t *mapped, size_t *n_thp); +void base_prefork(tsdn_t *tsdn, base_t *base); +void base_postfork_parent(tsdn_t *tsdn, base_t *base); +void base_postfork_child(tsdn_t *tsdn, base_t *base); +bool base_boot(tsdn_t *tsdn); + +#endif /* JEMALLOC_INTERNAL_BASE_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/base_externs.h b/contrib/jemalloc/include/jemalloc/internal/base_externs.h deleted file mode 100644 --- a/contrib/jemalloc/include/jemalloc/internal/base_externs.h +++ /dev/null @@ -1,22 +0,0 @@ -#ifndef JEMALLOC_INTERNAL_BASE_EXTERNS_H -#define JEMALLOC_INTERNAL_BASE_EXTERNS_H - -extern metadata_thp_mode_t opt_metadata_thp; -extern const char *metadata_thp_mode_names[]; - -base_t *b0get(void); -base_t *base_new(tsdn_t *tsdn, unsigned ind, extent_hooks_t *extent_hooks); -void base_delete(tsdn_t *tsdn, base_t *base); -extent_hooks_t *base_extent_hooks_get(base_t *base); -extent_hooks_t *base_extent_hooks_set(base_t *base, - extent_hooks_t *extent_hooks); -void *base_alloc(tsdn_t *tsdn, base_t *base, size_t size, size_t alignment); -extent_t *base_alloc_extent(tsdn_t *tsdn, base_t *base); -void base_stats_get(tsdn_t *tsdn, base_t *base, size_t *allocated, - size_t *resident, size_t *mapped, size_t *n_thp); -void base_prefork(tsdn_t *tsdn, base_t *base); -void base_postfork_parent(tsdn_t *tsdn, base_t *base); -void base_postfork_child(tsdn_t *tsdn, base_t *base); -bool base_boot(tsdn_t *tsdn); - -#endif /* JEMALLOC_INTERNAL_BASE_EXTERNS_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/base_inlines.h b/contrib/jemalloc/include/jemalloc/internal/base_inlines.h deleted file mode 100644 --- a/contrib/jemalloc/include/jemalloc/internal/base_inlines.h +++ /dev/null @@ -1,13 +0,0 @@ -#ifndef JEMALLOC_INTERNAL_BASE_INLINES_H -#define JEMALLOC_INTERNAL_BASE_INLINES_H - -static inline unsigned -base_ind_get(const base_t *base) { - return base->ind; -} - -static inline bool -metadata_thp_enabled(void) { - return (opt_metadata_thp != metadata_thp_disabled); -} -#endif /* JEMALLOC_INTERNAL_BASE_INLINES_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/base_structs.h b/contrib/jemalloc/include/jemalloc/internal/base_structs.h deleted file mode 100644 --- a/contrib/jemalloc/include/jemalloc/internal/base_structs.h +++ /dev/null @@ -1,59 +0,0 @@ -#ifndef JEMALLOC_INTERNAL_BASE_STRUCTS_H -#define JEMALLOC_INTERNAL_BASE_STRUCTS_H - -#include "jemalloc/internal/jemalloc_internal_types.h" -#include "jemalloc/internal/mutex.h" -#include "jemalloc/internal/sc.h" - -/* Embedded at the beginning of every block of base-managed virtual memory. */ -struct base_block_s { - /* Total size of block's virtual memory mapping. */ - size_t size; - - /* Next block in list of base's blocks. */ - base_block_t *next; - - /* Tracks unused trailing space. */ - extent_t extent; -}; - -struct base_s { - /* Associated arena's index within the arenas array. */ - unsigned ind; - - /* - * User-configurable extent hook functions. Points to an - * extent_hooks_t. - */ - atomic_p_t extent_hooks; - - /* Protects base_alloc() and base_stats_get() operations. */ - malloc_mutex_t mtx; - - /* Using THP when true (metadata_thp auto mode). */ - bool auto_thp_switched; - /* - * Most recent size class in the series of increasingly large base - * extents. Logarithmic spacing between subsequent allocations ensures - * that the total number of distinct mappings remains small. - */ - pszind_t pind_last; - - /* Serial number generation state. */ - size_t extent_sn_next; - - /* Chain of all blocks associated with base. */ - base_block_t *blocks; - - /* Heap of extents that track unused trailing space within blocks. */ - extent_heap_t avail[SC_NSIZES]; - - /* Stats, only maintained if config_stats. */ - size_t allocated; - size_t resident; - size_t mapped; - /* Number of THP regions touched. */ - size_t n_thp; -}; - -#endif /* JEMALLOC_INTERNAL_BASE_STRUCTS_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/base_types.h b/contrib/jemalloc/include/jemalloc/internal/base_types.h deleted file mode 100644 --- a/contrib/jemalloc/include/jemalloc/internal/base_types.h +++ /dev/null @@ -1,33 +0,0 @@ -#ifndef JEMALLOC_INTERNAL_BASE_TYPES_H -#define JEMALLOC_INTERNAL_BASE_TYPES_H - -typedef struct base_block_s base_block_t; -typedef struct base_s base_t; - -#define METADATA_THP_DEFAULT metadata_thp_disabled - -/* - * In auto mode, arenas switch to huge pages for the base allocator on the - * second base block. a0 switches to thp on the 5th block (after 20 megabytes - * of metadata), since more metadata (e.g. rtree nodes) come from a0's base. - */ - -#define BASE_AUTO_THP_THRESHOLD 2 -#define BASE_AUTO_THP_THRESHOLD_A0 5 - -typedef enum { - metadata_thp_disabled = 0, - /* - * Lazily enable hugepage for metadata. To avoid high RSS caused by THP - * + low usage arena (i.e. THP becomes a significant percentage), the - * "auto" option only starts using THP after a base allocator used up - * the first THP region. Starting from the second hugepage (in a single - * arena), "auto" behaves the same as "always", i.e. madvise hugepage - * right away. - */ - metadata_thp_auto = 1, - metadata_thp_always = 2, - metadata_thp_mode_limit = 3 -} metadata_thp_mode_t; - -#endif /* JEMALLOC_INTERNAL_BASE_TYPES_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/bin.h b/contrib/jemalloc/include/jemalloc/internal/bin.h --- a/contrib/jemalloc/include/jemalloc/internal/bin.h +++ b/contrib/jemalloc/include/jemalloc/internal/bin.h @@ -3,8 +3,7 @@ #include "jemalloc/internal/bin_stats.h" #include "jemalloc/internal/bin_types.h" -#include "jemalloc/internal/extent_types.h" -#include "jemalloc/internal/extent_structs.h" +#include "jemalloc/internal/edata.h" #include "jemalloc/internal/mutex.h" #include "jemalloc/internal/sc.h" @@ -12,74 +11,34 @@ * A bin contains a set of extents that are currently being used for slab * allocations. */ - -/* - * Read-only information associated with each element of arena_t's bins array - * is stored separately, partly to reduce memory usage (only one copy, rather - * than one per arena), but mainly to avoid false cacheline sharing. - * - * Each slab has the following layout: - * - * /--------------------\ - * | region 0 | - * |--------------------| - * | region 1 | - * |--------------------| - * | ... | - * | ... | - * | ... | - * |--------------------| - * | region nregs-1 | - * \--------------------/ - */ -typedef struct bin_info_s bin_info_t; -struct bin_info_s { - /* Size of regions in a slab for this bin's size class. */ - size_t reg_size; - - /* Total size of a slab for this bin's size class. */ - size_t slab_size; - - /* Total number of regions in a slab for this bin's size class. */ - uint32_t nregs; - - /* Number of sharded bins in each arena for this size class. */ - uint32_t n_shards; - - /* - * Metadata used to manipulate bitmaps for slabs associated with this - * bin. - */ - bitmap_info_t bitmap_info; -}; - -extern bin_info_t bin_infos[SC_NBINS]; - typedef struct bin_s bin_t; struct bin_s { /* All operations on bin_t fields require lock ownership. */ malloc_mutex_t lock; + /* + * Bin statistics. These get touched every time the lock is acquired, + * so put them close by in the hopes of getting some cache locality. + */ + bin_stats_t stats; + /* * Current slab being used to service allocations of this bin's size * class. slabcur is independent of slabs_{nonfull,full}; whenever * slabcur is reassigned, the previous slab must be deallocated or * inserted into slabs_{nonfull,full}. */ - extent_t *slabcur; + edata_t *slabcur; /* * Heap of non-full slabs. This heap is used to assure that new * allocations come from the non-full slab that is oldest/lowest in * memory. */ - extent_heap_t slabs_nonfull; + edata_heap_t slabs_nonfull; /* List used to track full slabs. */ - extent_list_t slabs_full; - - /* Bin statistics. */ - bin_stats_t stats; + edata_list_active_t slabs_full; }; /* A set of sharded bins of the same size class. */ @@ -92,7 +51,6 @@ void bin_shard_sizes_boot(unsigned bin_shards[SC_NBINS]); bool bin_update_shard_size(unsigned bin_shards[SC_NBINS], size_t start_size, size_t end_size, size_t nshards); -void bin_boot(sc_data_t *sc_data, unsigned bin_shard_sizes[SC_NBINS]); /* Initializes a bin to empty. Returns true on error. */ bool bin_init(bin_t *bin); @@ -104,19 +62,20 @@ /* Stats. */ static inline void -bin_stats_merge(tsdn_t *tsdn, bin_stats_t *dst_bin_stats, bin_t *bin) { +bin_stats_merge(tsdn_t *tsdn, bin_stats_data_t *dst_bin_stats, bin_t *bin) { malloc_mutex_lock(tsdn, &bin->lock); malloc_mutex_prof_accum(tsdn, &dst_bin_stats->mutex_data, &bin->lock); - dst_bin_stats->nmalloc += bin->stats.nmalloc; - dst_bin_stats->ndalloc += bin->stats.ndalloc; - dst_bin_stats->nrequests += bin->stats.nrequests; - dst_bin_stats->curregs += bin->stats.curregs; - dst_bin_stats->nfills += bin->stats.nfills; - dst_bin_stats->nflushes += bin->stats.nflushes; - dst_bin_stats->nslabs += bin->stats.nslabs; - dst_bin_stats->reslabs += bin->stats.reslabs; - dst_bin_stats->curslabs += bin->stats.curslabs; - dst_bin_stats->nonfull_slabs += bin->stats.nonfull_slabs; + bin_stats_t *stats = &dst_bin_stats->stats_data; + stats->nmalloc += bin->stats.nmalloc; + stats->ndalloc += bin->stats.ndalloc; + stats->nrequests += bin->stats.nrequests; + stats->curregs += bin->stats.curregs; + stats->nfills += bin->stats.nfills; + stats->nflushes += bin->stats.nflushes; + stats->nslabs += bin->stats.nslabs; + stats->reslabs += bin->stats.reslabs; + stats->curslabs += bin->stats.curslabs; + stats->nonfull_slabs += bin->stats.nonfull_slabs; malloc_mutex_unlock(tsdn, &bin->lock); } diff --git a/contrib/jemalloc/include/jemalloc/internal/bin_info.h b/contrib/jemalloc/include/jemalloc/internal/bin_info.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/bin_info.h @@ -0,0 +1,50 @@ +#ifndef JEMALLOC_INTERNAL_BIN_INFO_H +#define JEMALLOC_INTERNAL_BIN_INFO_H + +#include "jemalloc/internal/bitmap.h" + +/* + * Read-only information associated with each element of arena_t's bins array + * is stored separately, partly to reduce memory usage (only one copy, rather + * than one per arena), but mainly to avoid false cacheline sharing. + * + * Each slab has the following layout: + * + * /--------------------\ + * | region 0 | + * |--------------------| + * | region 1 | + * |--------------------| + * | ... | + * | ... | + * | ... | + * |--------------------| + * | region nregs-1 | + * \--------------------/ + */ +typedef struct bin_info_s bin_info_t; +struct bin_info_s { + /* Size of regions in a slab for this bin's size class. */ + size_t reg_size; + + /* Total size of a slab for this bin's size class. */ + size_t slab_size; + + /* Total number of regions in a slab for this bin's size class. */ + uint32_t nregs; + + /* Number of sharded bins in each arena for this size class. */ + uint32_t n_shards; + + /* + * Metadata used to manipulate bitmaps for slabs associated with this + * bin. + */ + bitmap_info_t bitmap_info; +}; + +extern bin_info_t bin_infos[SC_NBINS]; + +void bin_info_boot(sc_data_t *sc_data, unsigned bin_shard_sizes[SC_NBINS]); + +#endif /* JEMALLOC_INTERNAL_BIN_INFO_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/bin_stats.h b/contrib/jemalloc/include/jemalloc/internal/bin_stats.h --- a/contrib/jemalloc/include/jemalloc/internal/bin_stats.h +++ b/contrib/jemalloc/include/jemalloc/internal/bin_stats.h @@ -47,8 +47,11 @@ /* Current size of nonfull slabs heap in this bin. */ size_t nonfull_slabs; +}; +typedef struct bin_stats_data_s bin_stats_data_t; +struct bin_stats_data_s { + bin_stats_t stats_data; mutex_prof_data_t mutex_data; }; - #endif /* JEMALLOC_INTERNAL_BIN_STATS_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/bin_types.h b/contrib/jemalloc/include/jemalloc/internal/bin_types.h --- a/contrib/jemalloc/include/jemalloc/internal/bin_types.h +++ b/contrib/jemalloc/include/jemalloc/internal/bin_types.h @@ -3,7 +3,7 @@ #include "jemalloc/internal/sc.h" -#define BIN_SHARDS_MAX (1 << EXTENT_BITS_BINSHARD_WIDTH) +#define BIN_SHARDS_MAX (1 << EDATA_BITS_BINSHARD_WIDTH) #define N_BIN_SHARDS_DEFAULT 1 /* Used in TSD static initializer only. Real init in arena_bind(). */ diff --git a/contrib/jemalloc/include/jemalloc/internal/bit_util.h b/contrib/jemalloc/include/jemalloc/internal/bit_util.h --- a/contrib/jemalloc/include/jemalloc/internal/bit_util.h +++ b/contrib/jemalloc/include/jemalloc/internal/bit_util.h @@ -3,144 +3,383 @@ #include "jemalloc/internal/assert.h" -#define BIT_UTIL_INLINE static inline - /* Sanity check. */ #if !defined(JEMALLOC_INTERNAL_FFSLL) || !defined(JEMALLOC_INTERNAL_FFSL) \ || !defined(JEMALLOC_INTERNAL_FFS) # error JEMALLOC_INTERNAL_FFS{,L,LL} should have been defined by configure #endif +/* + * Unlike the builtins and posix ffs functions, our ffs requires a non-zero + * input, and returns the position of the lowest bit set (as opposed to the + * posix versions, which return 1 larger than that position and use a return + * value of zero as a sentinel. This tends to simplify logic in callers, and + * allows for consistency with the builtins we build fls on top of. + */ +static inline unsigned +ffs_llu(unsigned long long x) { + util_assume(x != 0); + return JEMALLOC_INTERNAL_FFSLL(x) - 1; +} -BIT_UTIL_INLINE unsigned -ffs_llu(unsigned long long bitmap) { - return JEMALLOC_INTERNAL_FFSLL(bitmap); +static inline unsigned +ffs_lu(unsigned long x) { + util_assume(x != 0); + return JEMALLOC_INTERNAL_FFSL(x) - 1; } -BIT_UTIL_INLINE unsigned -ffs_lu(unsigned long bitmap) { - return JEMALLOC_INTERNAL_FFSL(bitmap); +static inline unsigned +ffs_u(unsigned x) { + util_assume(x != 0); + return JEMALLOC_INTERNAL_FFS(x) - 1; } -BIT_UTIL_INLINE unsigned -ffs_u(unsigned bitmap) { - return JEMALLOC_INTERNAL_FFS(bitmap); +#define DO_FLS_SLOW(x, suffix) do { \ + util_assume(x != 0); \ + x |= (x >> 1); \ + x |= (x >> 2); \ + x |= (x >> 4); \ + x |= (x >> 8); \ + x |= (x >> 16); \ + if (sizeof(x) > 4) { \ + /* \ + * If sizeof(x) is 4, then the expression "x >> 32" \ + * will generate compiler warnings even if the code \ + * never executes. This circumvents the warning, and \ + * gets compiled out in optimized builds. \ + */ \ + int constant_32 = sizeof(x) * 4; \ + x |= (x >> constant_32); \ + } \ + x++; \ + if (x == 0) { \ + return 8 * sizeof(x) - 1; \ + } \ + return ffs_##suffix(x) - 1; \ +} while(0) + +static inline unsigned +fls_llu_slow(unsigned long long x) { + DO_FLS_SLOW(x, llu); } -#ifdef JEMALLOC_INTERNAL_POPCOUNTL -BIT_UTIL_INLINE unsigned +static inline unsigned +fls_lu_slow(unsigned long x) { + DO_FLS_SLOW(x, lu); +} + +static inline unsigned +fls_u_slow(unsigned x) { + DO_FLS_SLOW(x, u); +} + +#undef DO_FLS_SLOW + +#ifdef JEMALLOC_HAVE_BUILTIN_CLZ +static inline unsigned +fls_llu(unsigned long long x) { + util_assume(x != 0); + /* + * Note that the xor here is more naturally written as subtraction; the + * last bit set is the number of bits in the type minus the number of + * leading zero bits. But GCC implements that as: + * bsr edi, edi + * mov eax, 31 + * xor edi, 31 + * sub eax, edi + * If we write it as xor instead, then we get + * bsr eax, edi + * as desired. + */ + return (8 * sizeof(x) - 1) ^ __builtin_clzll(x); +} + +static inline unsigned +fls_lu(unsigned long x) { + util_assume(x != 0); + return (8 * sizeof(x) - 1) ^ __builtin_clzl(x); +} + +static inline unsigned +fls_u(unsigned x) { + util_assume(x != 0); + return (8 * sizeof(x) - 1) ^ __builtin_clz(x); +} +#elif defined(_MSC_VER) + +#if LG_SIZEOF_PTR == 3 +#define DO_BSR64(bit, x) _BitScanReverse64(&bit, x) +#else +/* + * This never actually runs; we're just dodging a compiler error for the + * never-taken branch where sizeof(void *) == 8. + */ +#define DO_BSR64(bit, x) bit = 0; unreachable() +#endif + +#define DO_FLS(x) do { \ + if (x == 0) { \ + return 8 * sizeof(x); \ + } \ + unsigned long bit; \ + if (sizeof(x) == 4) { \ + _BitScanReverse(&bit, (unsigned)x); \ + return (unsigned)bit; \ + } \ + if (sizeof(x) == 8 && sizeof(void *) == 8) { \ + DO_BSR64(bit, x); \ + return (unsigned)bit; \ + } \ + if (sizeof(x) == 8 && sizeof(void *) == 4) { \ + /* Dodge a compiler warning, as above. */ \ + int constant_32 = sizeof(x) * 4; \ + if (_BitScanReverse(&bit, \ + (unsigned)(x >> constant_32))) { \ + return 32 + (unsigned)bit; \ + } else { \ + _BitScanReverse(&bit, (unsigned)x); \ + return (unsigned)bit; \ + } \ + } \ + unreachable(); \ +} while (0) + +static inline unsigned +fls_llu(unsigned long long x) { + DO_FLS(x); +} + +static inline unsigned +fls_lu(unsigned long x) { + DO_FLS(x); +} + +static inline unsigned +fls_u(unsigned x) { + DO_FLS(x); +} + +#undef DO_FLS +#undef DO_BSR64 +#else + +static inline unsigned +fls_llu(unsigned long long x) { + return fls_llu_slow(x); +} + +static inline unsigned +fls_lu(unsigned long x) { + return fls_lu_slow(x); +} + +static inline unsigned +fls_u(unsigned x) { + return fls_u_slow(x); +} +#endif + +#if LG_SIZEOF_LONG_LONG > 3 +# error "Haven't implemented popcount for 16-byte ints." +#endif + +#define DO_POPCOUNT(x, type) do { \ + /* \ + * Algorithm from an old AMD optimization reference manual. \ + * We're putting a little bit more work than you might expect \ + * into the no-instrinsic case, since we only support the \ + * GCC intrinsics spelling of popcount (for now). Detecting \ + * whether or not the popcount builtin is actually useable in \ + * MSVC is nontrivial. \ + */ \ + \ + type bmul = (type)0x0101010101010101ULL; \ + \ + /* \ + * Replace each 2 bits with the sideways sum of the original \ + * values. 0x5 = 0b0101. \ + * \ + * You might expect this to be: \ + * x = (x & 0x55...) + ((x >> 1) & 0x55...). \ + * That costs an extra mask relative to this, though. \ + */ \ + x = x - ((x >> 1) & (0x55U * bmul)); \ + /* Replace each 4 bits with their sideays sum. 0x3 = 0b0011. */\ + x = (x & (bmul * 0x33U)) + ((x >> 2) & (bmul * 0x33U)); \ + /* \ + * Replace each 8 bits with their sideways sum. Note that we \ + * can't overflow within each 4-bit sum here, so we can skip \ + * the initial mask. \ + */ \ + x = (x + (x >> 4)) & (bmul * 0x0FU); \ + /* \ + * None of the partial sums in this multiplication (viewed in \ + * base-256) can overflow into the next digit. So the least \ + * significant byte of the product will be the least \ + * significant byte of the original value, the second least \ + * significant byte will be the sum of the two least \ + * significant bytes of the original value, and so on. \ + * Importantly, the high byte will be the byte-wise sum of all \ + * the bytes of the original value. \ + */ \ + x = x * bmul; \ + x >>= ((sizeof(x) - 1) * 8); \ + return (unsigned)x; \ +} while(0) + +static inline unsigned +popcount_u_slow(unsigned bitmap) { + DO_POPCOUNT(bitmap, unsigned); +} + +static inline unsigned +popcount_lu_slow(unsigned long bitmap) { + DO_POPCOUNT(bitmap, unsigned long); +} + +static inline unsigned +popcount_llu_slow(unsigned long long bitmap) { + DO_POPCOUNT(bitmap, unsigned long long); +} + +#undef DO_POPCOUNT + +static inline unsigned +popcount_u(unsigned bitmap) { +#ifdef JEMALLOC_INTERNAL_POPCOUNT + return JEMALLOC_INTERNAL_POPCOUNT(bitmap); +#else + return popcount_u_slow(bitmap); +#endif +} + +static inline unsigned popcount_lu(unsigned long bitmap) { - return JEMALLOC_INTERNAL_POPCOUNTL(bitmap); +#ifdef JEMALLOC_INTERNAL_POPCOUNTL + return JEMALLOC_INTERNAL_POPCOUNTL(bitmap); +#else + return popcount_lu_slow(bitmap); +#endif } + +static inline unsigned +popcount_llu(unsigned long long bitmap) { +#ifdef JEMALLOC_INTERNAL_POPCOUNTLL + return JEMALLOC_INTERNAL_POPCOUNTLL(bitmap); +#else + return popcount_llu_slow(bitmap); #endif +} /* * Clears first unset bit in bitmap, and returns * place of bit. bitmap *must not* be 0. */ -BIT_UTIL_INLINE size_t +static inline size_t cfs_lu(unsigned long* bitmap) { - size_t bit = ffs_lu(*bitmap) - 1; + util_assume(*bitmap != 0); + size_t bit = ffs_lu(*bitmap); *bitmap ^= ZU(1) << bit; return bit; } -BIT_UTIL_INLINE unsigned -ffs_zu(size_t bitmap) { +static inline unsigned +ffs_zu(size_t x) { #if LG_SIZEOF_PTR == LG_SIZEOF_INT - return ffs_u(bitmap); + return ffs_u(x); #elif LG_SIZEOF_PTR == LG_SIZEOF_LONG - return ffs_lu(bitmap); + return ffs_lu(x); #elif LG_SIZEOF_PTR == LG_SIZEOF_LONG_LONG - return ffs_llu(bitmap); + return ffs_llu(x); #else #error No implementation for size_t ffs() #endif } -BIT_UTIL_INLINE unsigned -ffs_u64(uint64_t bitmap) { +static inline unsigned +fls_zu(size_t x) { +#if LG_SIZEOF_PTR == LG_SIZEOF_INT + return fls_u(x); +#elif LG_SIZEOF_PTR == LG_SIZEOF_LONG + return fls_lu(x); +#elif LG_SIZEOF_PTR == LG_SIZEOF_LONG_LONG + return fls_llu(x); +#else +#error No implementation for size_t fls() +#endif +} + + +static inline unsigned +ffs_u64(uint64_t x) { #if LG_SIZEOF_LONG == 3 - return ffs_lu(bitmap); + return ffs_lu(x); #elif LG_SIZEOF_LONG_LONG == 3 - return ffs_llu(bitmap); + return ffs_llu(x); #else #error No implementation for 64-bit ffs() #endif } -BIT_UTIL_INLINE unsigned -ffs_u32(uint32_t bitmap) { +static inline unsigned +fls_u64(uint64_t x) { +#if LG_SIZEOF_LONG == 3 + return fls_lu(x); +#elif LG_SIZEOF_LONG_LONG == 3 + return fls_llu(x); +#else +#error No implementation for 64-bit fls() +#endif +} + +static inline unsigned +ffs_u32(uint32_t x) { #if LG_SIZEOF_INT == 2 - return ffs_u(bitmap); + return ffs_u(x); #else #error No implementation for 32-bit ffs() #endif - return ffs_u(bitmap); + return ffs_u(x); +} + +static inline unsigned +fls_u32(uint32_t x) { +#if LG_SIZEOF_INT == 2 + return fls_u(x); +#else +#error No implementation for 32-bit fls() +#endif + return fls_u(x); } -BIT_UTIL_INLINE uint64_t +static inline uint64_t pow2_ceil_u64(uint64_t x) { -#if (defined(__amd64__) || defined(__x86_64__) || defined(JEMALLOC_HAVE_BUILTIN_CLZ)) - if(unlikely(x <= 1)) { + if (unlikely(x <= 1)) { return x; } - size_t msb_on_index; -#if (defined(__amd64__) || defined(__x86_64__)) - asm ("bsrq %1, %0" - : "=r"(msb_on_index) // Outputs. - : "r"(x-1) // Inputs. - ); -#elif (defined(JEMALLOC_HAVE_BUILTIN_CLZ)) - msb_on_index = (63 ^ __builtin_clzll(x - 1)); -#endif + size_t msb_on_index = fls_u64(x - 1); + /* + * Range-check; it's on the callers to ensure that the result of this + * call won't overflow. + */ assert(msb_on_index < 63); return 1ULL << (msb_on_index + 1); -#else - x--; - x |= x >> 1; - x |= x >> 2; - x |= x >> 4; - x |= x >> 8; - x |= x >> 16; - x |= x >> 32; - x++; - return x; -#endif } -BIT_UTIL_INLINE uint32_t +static inline uint32_t pow2_ceil_u32(uint32_t x) { -#if ((defined(__i386__) || defined(JEMALLOC_HAVE_BUILTIN_CLZ)) && (!defined(__s390__))) - if(unlikely(x <= 1)) { - return x; + if (unlikely(x <= 1)) { + return x; } - size_t msb_on_index; -#if (defined(__i386__)) - asm ("bsr %1, %0" - : "=r"(msb_on_index) // Outputs. - : "r"(x-1) // Inputs. - ); -#elif (defined(JEMALLOC_HAVE_BUILTIN_CLZ)) - msb_on_index = (31 ^ __builtin_clz(x - 1)); -#endif + size_t msb_on_index = fls_u32(x - 1); + /* As above. */ assert(msb_on_index < 31); return 1U << (msb_on_index + 1); -#else - x--; - x |= x >> 1; - x |= x >> 2; - x |= x >> 4; - x |= x >> 8; - x |= x >> 16; - x++; - return x; -#endif } /* Compute the smallest power of 2 that is >= x. */ -BIT_UTIL_INLINE size_t +static inline size_t pow2_ceil_zu(size_t x) { #if (LG_SIZEOF_PTR == 3) return pow2_ceil_u64(x); @@ -149,77 +388,21 @@ #endif } -#if (defined(__i386__) || defined(__amd64__) || defined(__x86_64__)) -BIT_UTIL_INLINE unsigned -lg_floor(size_t x) { - size_t ret; - assert(x != 0); - - asm ("bsr %1, %0" - : "=r"(ret) // Outputs. - : "r"(x) // Inputs. - ); - assert(ret < UINT_MAX); - return (unsigned)ret; -} -#elif (defined(_MSC_VER)) -BIT_UTIL_INLINE unsigned +static inline unsigned lg_floor(size_t x) { - unsigned long ret; - - assert(x != 0); - + util_assume(x != 0); #if (LG_SIZEOF_PTR == 3) - _BitScanReverse64(&ret, x); -#elif (LG_SIZEOF_PTR == 2) - _BitScanReverse(&ret, x); + return fls_u64(x); #else -# error "Unsupported type size for lg_floor()" + return fls_u32(x); #endif - assert(ret < UINT_MAX); - return (unsigned)ret; } -#elif (defined(JEMALLOC_HAVE_BUILTIN_CLZ)) -BIT_UTIL_INLINE unsigned -lg_floor(size_t x) { - assert(x != 0); -#if (LG_SIZEOF_PTR == LG_SIZEOF_INT) - return ((8 << LG_SIZEOF_PTR) - 1) - __builtin_clz(x); -#elif (LG_SIZEOF_PTR == LG_SIZEOF_LONG) - return ((8 << LG_SIZEOF_PTR) - 1) - __builtin_clzl(x); -#else -# error "Unsupported type size for lg_floor()" -#endif -} -#else -BIT_UTIL_INLINE unsigned -lg_floor(size_t x) { - assert(x != 0); - - x |= (x >> 1); - x |= (x >> 2); - x |= (x >> 4); - x |= (x >> 8); - x |= (x >> 16); -#if (LG_SIZEOF_PTR == 3) - x |= (x >> 32); -#endif - if (x == SIZE_T_MAX) { - return (8 << LG_SIZEOF_PTR) - 1; - } - x++; - return ffs_zu(x) - 2; -} -#endif - -BIT_UTIL_INLINE unsigned +static inline unsigned lg_ceil(size_t x) { return lg_floor(x) + ((x & (x - 1)) == 0 ? 0 : 1); } -#undef BIT_UTIL_INLINE - /* A compile-time version of lg_floor and lg_ceil. */ #define LG_FLOOR_1(x) 0 #define LG_FLOOR_2(x) (x < (1ULL << 1) ? LG_FLOOR_1(x) : 1 + LG_FLOOR_1(x >> 1)) diff --git a/contrib/jemalloc/include/jemalloc/internal/bitmap.h b/contrib/jemalloc/include/jemalloc/internal/bitmap.h --- a/contrib/jemalloc/include/jemalloc/internal/bitmap.h +++ b/contrib/jemalloc/include/jemalloc/internal/bitmap.h @@ -1,7 +1,6 @@ #ifndef JEMALLOC_INTERNAL_BITMAP_H #define JEMALLOC_INTERNAL_BITMAP_H -#include "jemalloc/internal/arena_types.h" #include "jemalloc/internal/bit_util.h" #include "jemalloc/internal/sc.h" @@ -9,9 +8,9 @@ #define LG_SIZEOF_BITMAP LG_SIZEOF_LONG /* Maximum bitmap bit count is 2^LG_BITMAP_MAXBITS. */ -#if LG_SLAB_MAXREGS > LG_CEIL(SC_NSIZES) +#if SC_LG_SLAB_MAXREGS > LG_CEIL(SC_NSIZES) /* Maximum bitmap bit count is determined by maximum regions per slab. */ -# define LG_BITMAP_MAXBITS LG_SLAB_MAXREGS +# define LG_BITMAP_MAXBITS SC_LG_SLAB_MAXREGS #else /* Maximum bitmap bit count is determined by number of extent size classes. */ # define LG_BITMAP_MAXBITS LG_CEIL(SC_NSIZES) @@ -273,7 +272,7 @@ } return bitmap_ffu(bitmap, binfo, sib_base); } - bit += ((size_t)(ffs_lu(group_masked) - 1)) << + bit += ((size_t)ffs_lu(group_masked)) << (lg_bits_per_group - LG_BITMAP_GROUP_NBITS); } assert(bit >= min_bit); @@ -285,9 +284,9 @@ - 1); size_t bit; do { - bit = ffs_lu(g); - if (bit != 0) { - return (i << LG_BITMAP_GROUP_NBITS) + (bit - 1); + if (g != 0) { + bit = ffs_lu(g); + return (i << LG_BITMAP_GROUP_NBITS) + bit; } i++; g = bitmap[i]; @@ -308,20 +307,20 @@ #ifdef BITMAP_USE_TREE i = binfo->nlevels - 1; g = bitmap[binfo->levels[i].group_offset]; - bit = ffs_lu(g) - 1; + bit = ffs_lu(g); while (i > 0) { i--; g = bitmap[binfo->levels[i].group_offset + bit]; - bit = (bit << LG_BITMAP_GROUP_NBITS) + (ffs_lu(g) - 1); + bit = (bit << LG_BITMAP_GROUP_NBITS) + ffs_lu(g); } #else i = 0; g = bitmap[0]; - while ((bit = ffs_lu(g)) == 0) { + while (g == 0) { i++; g = bitmap[i]; } - bit = (i << LG_BITMAP_GROUP_NBITS) + (bit - 1); + bit = (i << LG_BITMAP_GROUP_NBITS) + ffs_lu(g); #endif bitmap_set(bitmap, binfo, bit); return bit; diff --git a/contrib/jemalloc/include/jemalloc/internal/buf_writer.h b/contrib/jemalloc/include/jemalloc/internal/buf_writer.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/buf_writer.h @@ -0,0 +1,32 @@ +#ifndef JEMALLOC_INTERNAL_BUF_WRITER_H +#define JEMALLOC_INTERNAL_BUF_WRITER_H + +/* + * Note: when using the buffered writer, cbopaque is passed to write_cb only + * when the buffer is flushed. It would make a difference if cbopaque points + * to something that's changing for each write_cb call, or something that + * affects write_cb in a way dependent on the content of the output string. + * However, the most typical usage case in practice is that cbopaque points to + * some "option like" content for the write_cb, so it doesn't matter. + */ + +typedef struct { + write_cb_t *write_cb; + void *cbopaque; + char *buf; + size_t buf_size; + size_t buf_end; + bool internal_buf; +} buf_writer_t; + +bool buf_writer_init(tsdn_t *tsdn, buf_writer_t *buf_writer, + write_cb_t *write_cb, void *cbopaque, char *buf, size_t buf_len); +void buf_writer_flush(buf_writer_t *buf_writer); +write_cb_t buf_writer_cb; +void buf_writer_terminate(tsdn_t *tsdn, buf_writer_t *buf_writer); + +typedef ssize_t (read_cb_t)(void *read_cbopaque, void *buf, size_t limit); +void buf_writer_pipe(buf_writer_t *buf_writer, read_cb_t *read_cb, + void *read_cbopaque); + +#endif /* JEMALLOC_INTERNAL_BUF_WRITER_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/cache_bin.h b/contrib/jemalloc/include/jemalloc/internal/cache_bin.h --- a/contrib/jemalloc/include/jemalloc/internal/cache_bin.h +++ b/contrib/jemalloc/include/jemalloc/internal/cache_bin.h @@ -2,6 +2,7 @@ #define JEMALLOC_INTERNAL_CACHE_BIN_H #include "jemalloc/internal/ql.h" +#include "jemalloc/internal/sz.h" /* * The cache_bins are the mechanism that the tcache and the arena use to @@ -13,14 +14,38 @@ * of the tcache at all. */ +/* + * The size in bytes of each cache bin stack. We also use this to indicate + * *counts* of individual objects. + */ +typedef uint16_t cache_bin_sz_t; /* - * The count of the number of cached allocations in a bin. We make this signed - * so that negative numbers can encode "invalid" states (e.g. a low water mark - * of -1 for a cache that has been depleted). + * Leave a noticeable mark pattern on the cache bin stack boundaries, in case a + * bug starts leaking those. Make it look like the junk pattern but be distinct + * from it. */ -typedef int32_t cache_bin_sz_t; +static const uintptr_t cache_bin_preceding_junk = + (uintptr_t)0x7a7a7a7a7a7a7a7aULL; +/* Note: a7 vs. 7a above -- this tells you which pointer leaked. */ +static const uintptr_t cache_bin_trailing_junk = + (uintptr_t)0xa7a7a7a7a7a7a7a7ULL; +/* + * That implies the following value, for the maximum number of items in any + * individual bin. The cache bins track their bounds looking just at the low + * bits of a pointer, compared against a cache_bin_sz_t. So that's + * 1 << (sizeof(cache_bin_sz_t) * 8) + * bytes spread across pointer sized objects to get the maximum. + */ +#define CACHE_BIN_NCACHED_MAX (((size_t)1 << sizeof(cache_bin_sz_t) * 8) \ + / sizeof(void *) - 1) + +/* + * This lives inside the cache_bin (for locality reasons), and is initialized + * alongside it, but is otherwise not modified by any cache bin operations. + * It's logically public and maintained by its callers. + */ typedef struct cache_bin_stats_s cache_bin_stats_t; struct cache_bin_stats_s { /* @@ -36,34 +61,75 @@ */ typedef struct cache_bin_info_s cache_bin_info_t; struct cache_bin_info_s { - /* Upper limit on ncached. */ cache_bin_sz_t ncached_max; }; +/* + * Responsible for caching allocations associated with a single size. + * + * Several pointers are used to track the stack. To save on metadata bytes, + * only the stack_head is a full sized pointer (which is dereferenced on the + * fastpath), while the others store only the low 16 bits -- this is correct + * because a single stack never takes more space than 2^16 bytes, and at the + * same time only equality checks are performed on the low bits. + * + * (low addr) (high addr) + * |------stashed------|------available------|------cached-----| + * ^ ^ ^ ^ + * low_bound(derived) low_bits_full stack_head low_bits_empty + */ typedef struct cache_bin_s cache_bin_t; struct cache_bin_s { - /* Min # cached since last GC. */ - cache_bin_sz_t low_water; - /* # of cached objects. */ - cache_bin_sz_t ncached; /* - * ncached and stats are both modified frequently. Let's keep them + * The stack grows down. Whenever the bin is nonempty, the head points + * to an array entry containing a valid allocation. When it is empty, + * the head points to one element past the owned array. + */ + void **stack_head; + /* + * cur_ptr and stats are both modified frequently. Let's keep them * close so that they have a higher chance of being on the same * cacheline, thus less write-backs. */ cache_bin_stats_t tstats; + /* - * Stack of available objects. + * The low bits of the address of the first item in the stack that + * hasn't been used since the last GC, to track the low water mark (min + * # of cached items). * - * To make use of adjacent cacheline prefetch, the items in the avail - * stack goes to higher address for newer allocations. avail points - * just above the available space, which means that - * avail[-ncached, ... -1] are available items and the lowest item will - * be allocated first. + * Since the stack grows down, this is a higher address than + * low_bits_full. */ - void **avail; + uint16_t low_bits_low_water; + + /* + * The low bits of the value that stack_head will take on when the array + * is full (of cached & stashed items). But remember that stack_head + * always points to a valid item when the array is nonempty -- this is + * in the array. + * + * Recall that since the stack grows down, this is the lowest available + * address in the array for caching. Only adjusted when stashing items. + */ + uint16_t low_bits_full; + + /* + * The low bits of the value that stack_head will take on when the array + * is empty. + * + * The stack grows down -- this is one past the highest address in the + * array. Immutable after initialization. + */ + uint16_t low_bits_empty; }; +/* + * The cache_bins live inside the tcache, but the arena (by design) isn't + * supposed to know much about tcache internals. To let the arena iterate over + * associated bins, we keep (with the tcache) a linked list of + * cache_bin_array_descriptor_ts that tell the arena how to find the bins. + */ typedef struct cache_bin_array_descriptor_s cache_bin_array_descriptor_t; struct cache_bin_array_descriptor_s { /* @@ -72,37 +138,214 @@ */ ql_elm(cache_bin_array_descriptor_t) link; /* Pointers to the tcache bins. */ - cache_bin_t *bins_small; - cache_bin_t *bins_large; + cache_bin_t *bins; }; static inline void cache_bin_array_descriptor_init(cache_bin_array_descriptor_t *descriptor, - cache_bin_t *bins_small, cache_bin_t *bins_large) { + cache_bin_t *bins) { ql_elm_new(descriptor, link); - descriptor->bins_small = bins_small; - descriptor->bins_large = bins_large; + descriptor->bins = bins; } -JEMALLOC_ALWAYS_INLINE void * -cache_bin_alloc_easy(cache_bin_t *bin, bool *success) { - void *ret; +JEMALLOC_ALWAYS_INLINE bool +cache_bin_nonfast_aligned(const void *ptr) { + if (!config_uaf_detection) { + return false; + } + /* + * Currently we use alignment to decide which pointer to junk & stash on + * dealloc (for catching use-after-free). In some common cases a + * page-aligned check is needed already (sdalloc w/ config_prof), so we + * are getting it more or less for free -- no added instructions on + * free_fastpath. + * + * Another way of deciding which pointer to sample, is adding another + * thread_event to pick one every N bytes. That also adds no cost on + * the fastpath, however it will tend to pick large allocations which is + * not the desired behavior. + */ + return ((uintptr_t)ptr & san_cache_bin_nonfast_mask) == 0; +} + +/* Returns ncached_max: Upper limit on ncached. */ +static inline cache_bin_sz_t +cache_bin_info_ncached_max(cache_bin_info_t *info) { + return info->ncached_max; +} + +/* + * Internal. + * + * Asserts that the pointer associated with earlier is <= the one associated + * with later. + */ +static inline void +cache_bin_assert_earlier(cache_bin_t *bin, uint16_t earlier, uint16_t later) { + if (earlier > later) { + assert(bin->low_bits_full > bin->low_bits_empty); + } +} - bin->ncached--; +/* + * Internal. + * + * Does difference calculations that handle wraparound correctly. Earlier must + * be associated with the position earlier in memory. + */ +static inline uint16_t +cache_bin_diff(cache_bin_t *bin, uint16_t earlier, uint16_t later, bool racy) { + /* + * When it's racy, bin->low_bits_full can be modified concurrently. It + * can cross the uint16_t max value and become less than + * bin->low_bits_empty at the time of the check. + */ + if (!racy) { + cache_bin_assert_earlier(bin, earlier, later); + } + return later - earlier; +} +/* + * Number of items currently cached in the bin, without checking ncached_max. + * We require specifying whether or not the request is racy or not (i.e. whether + * or not concurrent modifications are possible). + */ +static inline cache_bin_sz_t +cache_bin_ncached_get_internal(cache_bin_t *bin, bool racy) { + cache_bin_sz_t diff = cache_bin_diff(bin, + (uint16_t)(uintptr_t)bin->stack_head, bin->low_bits_empty, racy); + cache_bin_sz_t n = diff / sizeof(void *); /* - * Check for both bin->ncached == 0 and ncached < low_water - * in a single branch. + * We have undefined behavior here; if this function is called from the + * arena stats updating code, then stack_head could change from the + * first line to the next one. Morally, these loads should be atomic, + * but compilers won't currently generate comparisons with in-memory + * operands against atomics, and these variables get accessed on the + * fast paths. This should still be "safe" in the sense of generating + * the correct assembly for the foreseeable future, though. */ - if (unlikely(bin->ncached <= bin->low_water)) { - bin->low_water = bin->ncached; - if (bin->ncached == -1) { - bin->ncached = 0; - *success = false; - return NULL; - } + assert(n == 0 || *(bin->stack_head) != NULL || racy); + return n; +} + +/* + * Number of items currently cached in the bin, with checking ncached_max. The + * caller must know that no concurrent modification of the cache_bin is + * possible. + */ +static inline cache_bin_sz_t +cache_bin_ncached_get_local(cache_bin_t *bin, cache_bin_info_t *info) { + cache_bin_sz_t n = cache_bin_ncached_get_internal(bin, + /* racy */ false); + assert(n <= cache_bin_info_ncached_max(info)); + return n; +} + +/* + * Internal. + * + * A pointer to the position one past the end of the backing array. + * + * Do not call if racy, because both 'bin->stack_head' and 'bin->low_bits_full' + * are subject to concurrent modifications. + */ +static inline void ** +cache_bin_empty_position_get(cache_bin_t *bin) { + cache_bin_sz_t diff = cache_bin_diff(bin, + (uint16_t)(uintptr_t)bin->stack_head, bin->low_bits_empty, + /* racy */ false); + uintptr_t empty_bits = (uintptr_t)bin->stack_head + diff; + void **ret = (void **)empty_bits; + + assert(ret >= bin->stack_head); + + return ret; +} + +/* + * Internal. + * + * Calculates low bits of the lower bound of the usable cache bin's range (see + * cache_bin_t visual representation above). + * + * No values are concurrently modified, so should be safe to read in a + * multithreaded environment. Currently concurrent access happens only during + * arena statistics collection. + */ +static inline uint16_t +cache_bin_low_bits_low_bound_get(cache_bin_t *bin, cache_bin_info_t *info) { + return (uint16_t)bin->low_bits_empty - + info->ncached_max * sizeof(void *); +} + +/* + * Internal. + * + * A pointer to the position with the lowest address of the backing array. + */ +static inline void ** +cache_bin_low_bound_get(cache_bin_t *bin, cache_bin_info_t *info) { + cache_bin_sz_t ncached_max = cache_bin_info_ncached_max(info); + void **ret = cache_bin_empty_position_get(bin) - ncached_max; + assert(ret <= bin->stack_head); + + return ret; +} + +/* + * As the name implies. This is important since it's not correct to try to + * batch fill a nonempty cache bin. + */ +static inline void +cache_bin_assert_empty(cache_bin_t *bin, cache_bin_info_t *info) { + assert(cache_bin_ncached_get_local(bin, info) == 0); + assert(cache_bin_empty_position_get(bin) == bin->stack_head); +} + +/* + * Get low water, but without any of the correctness checking we do for the + * caller-usable version, if we are temporarily breaking invariants (like + * ncached >= low_water during flush). + */ +static inline cache_bin_sz_t +cache_bin_low_water_get_internal(cache_bin_t *bin) { + return cache_bin_diff(bin, bin->low_bits_low_water, + bin->low_bits_empty, /* racy */ false) / sizeof(void *); +} + +/* Returns the numeric value of low water in [0, ncached]. */ +static inline cache_bin_sz_t +cache_bin_low_water_get(cache_bin_t *bin, cache_bin_info_t *info) { + cache_bin_sz_t low_water = cache_bin_low_water_get_internal(bin); + assert(low_water <= cache_bin_info_ncached_max(info)); + assert(low_water <= cache_bin_ncached_get_local(bin, info)); + + cache_bin_assert_earlier(bin, (uint16_t)(uintptr_t)bin->stack_head, + bin->low_bits_low_water); + + return low_water; +} + +/* + * Indicates that the current cache bin position should be the low water mark + * going forward. + */ +static inline void +cache_bin_low_water_set(cache_bin_t *bin) { + bin->low_bits_low_water = (uint16_t)(uintptr_t)bin->stack_head; +} + +static inline void +cache_bin_low_water_adjust(cache_bin_t *bin) { + if (cache_bin_ncached_get_internal(bin, /* racy */ false) + < cache_bin_low_water_get_internal(bin)) { + cache_bin_low_water_set(bin); } +} +JEMALLOC_ALWAYS_INLINE void * +cache_bin_alloc_impl(cache_bin_t *bin, bool *success, bool adjust_low_water) { /* * success (instead of ret) should be checked upon the return of this * function. We avoid checking (ret == NULL) because there is never a @@ -110,22 +353,318 @@ * and eagerly checking ret would cause pipeline stall (waiting for the * cacheline). */ - *success = true; - ret = *(bin->avail - (bin->ncached + 1)); - return ret; + /* + * This may read from the empty position; however the loaded value won't + * be used. It's safe because the stack has one more slot reserved. + */ + void *ret = *bin->stack_head; + uint16_t low_bits = (uint16_t)(uintptr_t)bin->stack_head; + void **new_head = bin->stack_head + 1; + + /* + * Note that the low water mark is at most empty; if we pass this check, + * we know we're non-empty. + */ + if (likely(low_bits != bin->low_bits_low_water)) { + bin->stack_head = new_head; + *success = true; + return ret; + } + if (!adjust_low_water) { + *success = false; + return NULL; + } + /* + * In the fast-path case where we call alloc_easy and then alloc, the + * previous checking and computation is optimized away -- we didn't + * actually commit any of our operations. + */ + if (likely(low_bits != bin->low_bits_empty)) { + bin->stack_head = new_head; + bin->low_bits_low_water = (uint16_t)(uintptr_t)new_head; + *success = true; + return ret; + } + *success = false; + return NULL; +} + +/* + * Allocate an item out of the bin, failing if we're at the low-water mark. + */ +JEMALLOC_ALWAYS_INLINE void * +cache_bin_alloc_easy(cache_bin_t *bin, bool *success) { + /* We don't look at info if we're not adjusting low-water. */ + return cache_bin_alloc_impl(bin, success, false); +} + +/* + * Allocate an item out of the bin, even if we're currently at the low-water + * mark (and failing only if the bin is empty). + */ +JEMALLOC_ALWAYS_INLINE void * +cache_bin_alloc(cache_bin_t *bin, bool *success) { + return cache_bin_alloc_impl(bin, success, true); +} + +JEMALLOC_ALWAYS_INLINE cache_bin_sz_t +cache_bin_alloc_batch(cache_bin_t *bin, size_t num, void **out) { + cache_bin_sz_t n = cache_bin_ncached_get_internal(bin, + /* racy */ false); + if (n > num) { + n = (cache_bin_sz_t)num; + } + memcpy(out, bin->stack_head, n * sizeof(void *)); + bin->stack_head += n; + cache_bin_low_water_adjust(bin); + + return n; } JEMALLOC_ALWAYS_INLINE bool -cache_bin_dalloc_easy(cache_bin_t *bin, cache_bin_info_t *bin_info, void *ptr) { - if (unlikely(bin->ncached == bin_info->ncached_max)) { +cache_bin_full(cache_bin_t *bin) { + return ((uint16_t)(uintptr_t)bin->stack_head == bin->low_bits_full); +} + +/* + * Free an object into the given bin. Fails only if the bin is full. + */ +JEMALLOC_ALWAYS_INLINE bool +cache_bin_dalloc_easy(cache_bin_t *bin, void *ptr) { + if (unlikely(cache_bin_full(bin))) { return false; } - assert(bin->ncached < bin_info->ncached_max); - bin->ncached++; - *(bin->avail - bin->ncached) = ptr; + + bin->stack_head--; + *bin->stack_head = ptr; + cache_bin_assert_earlier(bin, bin->low_bits_full, + (uint16_t)(uintptr_t)bin->stack_head); return true; } +/* Returns false if failed to stash (i.e. bin is full). */ +JEMALLOC_ALWAYS_INLINE bool +cache_bin_stash(cache_bin_t *bin, void *ptr) { + if (cache_bin_full(bin)) { + return false; + } + + /* Stash at the full position, in the [full, head) range. */ + uint16_t low_bits_head = (uint16_t)(uintptr_t)bin->stack_head; + /* Wraparound handled as well. */ + uint16_t diff = cache_bin_diff(bin, bin->low_bits_full, low_bits_head, + /* racy */ false); + *(void **)((uintptr_t)bin->stack_head - diff) = ptr; + + assert(!cache_bin_full(bin)); + bin->low_bits_full += sizeof(void *); + cache_bin_assert_earlier(bin, bin->low_bits_full, low_bits_head); + + return true; +} + +/* + * Get the number of stashed pointers. + * + * When called from a thread not owning the TLS (i.e. racy = true), it's + * important to keep in mind that 'bin->stack_head' and 'bin->low_bits_full' can + * be modified concurrently and almost none assertions about their values can be + * made. + */ +JEMALLOC_ALWAYS_INLINE cache_bin_sz_t +cache_bin_nstashed_get_internal(cache_bin_t *bin, cache_bin_info_t *info, + bool racy) { + cache_bin_sz_t ncached_max = cache_bin_info_ncached_max(info); + uint16_t low_bits_low_bound = cache_bin_low_bits_low_bound_get(bin, + info); + + cache_bin_sz_t n = cache_bin_diff(bin, low_bits_low_bound, + bin->low_bits_full, racy) / sizeof(void *); + assert(n <= ncached_max); + + if (!racy) { + /* Below are for assertions only. */ + void **low_bound = cache_bin_low_bound_get(bin, info); + + assert((uint16_t)(uintptr_t)low_bound == low_bits_low_bound); + void *stashed = *(low_bound + n - 1); + bool aligned = cache_bin_nonfast_aligned(stashed); +#ifdef JEMALLOC_JET + /* Allow arbitrary pointers to be stashed in tests. */ + aligned = true; +#endif + assert(n == 0 || (stashed != NULL && aligned)); + } + + return n; +} + +JEMALLOC_ALWAYS_INLINE cache_bin_sz_t +cache_bin_nstashed_get_local(cache_bin_t *bin, cache_bin_info_t *info) { + cache_bin_sz_t n = cache_bin_nstashed_get_internal(bin, info, + /* racy */ false); + assert(n <= cache_bin_info_ncached_max(info)); + return n; +} + +/* + * Obtain a racy view of the number of items currently in the cache bin, in the + * presence of possible concurrent modifications. + */ +static inline void +cache_bin_nitems_get_remote(cache_bin_t *bin, cache_bin_info_t *info, + cache_bin_sz_t *ncached, cache_bin_sz_t *nstashed) { + cache_bin_sz_t n = cache_bin_ncached_get_internal(bin, /* racy */ true); + assert(n <= cache_bin_info_ncached_max(info)); + *ncached = n; + + n = cache_bin_nstashed_get_internal(bin, info, /* racy */ true); + assert(n <= cache_bin_info_ncached_max(info)); + *nstashed = n; + /* Note that cannot assert ncached + nstashed <= ncached_max (racy). */ +} + +/* + * Filling and flushing are done in batch, on arrays of void *s. For filling, + * the arrays go forward, and can be accessed with ordinary array arithmetic. + * For flushing, we work from the end backwards, and so need to use special + * accessors that invert the usual ordering. + * + * This is important for maintaining first-fit; the arena code fills with + * earliest objects first, and so those are the ones we should return first for + * cache_bin_alloc calls. When flushing, we should flush the objects that we + * wish to return later; those at the end of the array. This is better for the + * first-fit heuristic as well as for cache locality; the most recently freed + * objects are the ones most likely to still be in cache. + * + * This all sounds very hand-wavey and theoretical, but reverting the ordering + * on one or the other pathway leads to measurable slowdowns. + */ + +typedef struct cache_bin_ptr_array_s cache_bin_ptr_array_t; +struct cache_bin_ptr_array_s { + cache_bin_sz_t n; + void **ptr; +}; + +/* + * Declare a cache_bin_ptr_array_t sufficient for nval items. + * + * In the current implementation, this could be just part of a + * cache_bin_ptr_array_init_... call, since we reuse the cache bin stack memory. + * Indirecting behind a macro, though, means experimenting with linked-list + * representations is easy (since they'll require an alloca in the calling + * frame). + */ +#define CACHE_BIN_PTR_ARRAY_DECLARE(name, nval) \ + cache_bin_ptr_array_t name; \ + name.n = (nval) + +/* + * Start a fill. The bin must be empty, and This must be followed by a + * finish_fill call before doing any alloc/dalloc operations on the bin. + */ +static inline void +cache_bin_init_ptr_array_for_fill(cache_bin_t *bin, cache_bin_info_t *info, + cache_bin_ptr_array_t *arr, cache_bin_sz_t nfill) { + cache_bin_assert_empty(bin, info); + arr->ptr = cache_bin_empty_position_get(bin) - nfill; +} + +/* + * While nfill in cache_bin_init_ptr_array_for_fill is the number we *intend* to + * fill, nfilled here is the number we actually filled (which may be less, in + * case of OOM. + */ +static inline void +cache_bin_finish_fill(cache_bin_t *bin, cache_bin_info_t *info, + cache_bin_ptr_array_t *arr, cache_bin_sz_t nfilled) { + cache_bin_assert_empty(bin, info); + void **empty_position = cache_bin_empty_position_get(bin); + if (nfilled < arr->n) { + memmove(empty_position - nfilled, empty_position - arr->n, + nfilled * sizeof(void *)); + } + bin->stack_head = empty_position - nfilled; +} + +/* + * Same deal, but with flush. Unlike fill (which can fail), the user must flush + * everything we give them. + */ +static inline void +cache_bin_init_ptr_array_for_flush(cache_bin_t *bin, cache_bin_info_t *info, + cache_bin_ptr_array_t *arr, cache_bin_sz_t nflush) { + arr->ptr = cache_bin_empty_position_get(bin) - nflush; + assert(cache_bin_ncached_get_local(bin, info) == 0 + || *arr->ptr != NULL); +} + +static inline void +cache_bin_finish_flush(cache_bin_t *bin, cache_bin_info_t *info, + cache_bin_ptr_array_t *arr, cache_bin_sz_t nflushed) { + unsigned rem = cache_bin_ncached_get_local(bin, info) - nflushed; + memmove(bin->stack_head + nflushed, bin->stack_head, + rem * sizeof(void *)); + bin->stack_head = bin->stack_head + nflushed; + cache_bin_low_water_adjust(bin); +} + +static inline void +cache_bin_init_ptr_array_for_stashed(cache_bin_t *bin, szind_t binind, + cache_bin_info_t *info, cache_bin_ptr_array_t *arr, + cache_bin_sz_t nstashed) { + assert(nstashed > 0); + assert(cache_bin_nstashed_get_local(bin, info) == nstashed); + + void **low_bound = cache_bin_low_bound_get(bin, info); + arr->ptr = low_bound; + assert(*arr->ptr != NULL); +} + +static inline void +cache_bin_finish_flush_stashed(cache_bin_t *bin, cache_bin_info_t *info) { + void **low_bound = cache_bin_low_bound_get(bin, info); + + /* Reset the bin local full position. */ + bin->low_bits_full = (uint16_t)(uintptr_t)low_bound; + assert(cache_bin_nstashed_get_local(bin, info) == 0); +} + +/* + * Initialize a cache_bin_info to represent up to the given number of items in + * the cache_bins it is associated with. + */ +void cache_bin_info_init(cache_bin_info_t *bin_info, + cache_bin_sz_t ncached_max); +/* + * Given an array of initialized cache_bin_info_ts, determine how big an + * allocation is required to initialize a full set of cache_bin_ts. + */ +void cache_bin_info_compute_alloc(cache_bin_info_t *infos, szind_t ninfos, + size_t *size, size_t *alignment); + +/* + * Actually initialize some cache bins. Callers should allocate the backing + * memory indicated by a call to cache_bin_compute_alloc. They should then + * preincrement, call init once for each bin and info, and then call + * cache_bin_postincrement. *alloc_cur will then point immediately past the end + * of the allocation. + */ +void cache_bin_preincrement(cache_bin_info_t *infos, szind_t ninfos, + void *alloc, size_t *cur_offset); +void cache_bin_postincrement(cache_bin_info_t *infos, szind_t ninfos, + void *alloc, size_t *cur_offset); +void cache_bin_init(cache_bin_t *bin, cache_bin_info_t *info, void *alloc, + size_t *cur_offset); + +/* + * If a cache bin was zero initialized (either because it lives in static or + * thread-local storage, or was memset to 0), this function indicates whether or + * not cache_bin_init was called on it. + */ +bool cache_bin_still_zero_initialized(cache_bin_t *bin); + #endif /* JEMALLOC_INTERNAL_CACHE_BIN_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/counter.h b/contrib/jemalloc/include/jemalloc/internal/counter.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/counter.h @@ -0,0 +1,34 @@ +#ifndef JEMALLOC_INTERNAL_COUNTER_H +#define JEMALLOC_INTERNAL_COUNTER_H + +#include "jemalloc/internal/mutex.h" + +typedef struct counter_accum_s { + LOCKEDINT_MTX_DECLARE(mtx) + locked_u64_t accumbytes; + uint64_t interval; +} counter_accum_t; + +JEMALLOC_ALWAYS_INLINE bool +counter_accum(tsdn_t *tsdn, counter_accum_t *counter, uint64_t bytes) { + uint64_t interval = counter->interval; + assert(interval > 0); + LOCKEDINT_MTX_LOCK(tsdn, counter->mtx); + /* + * If the event moves fast enough (and/or if the event handling is slow + * enough), extreme overflow can cause counter trigger coalescing. + * This is an intentional mechanism that avoids rate-limiting + * allocation. + */ + bool overflow = locked_inc_mod_u64(tsdn, LOCKEDINT_MTX(counter->mtx), + &counter->accumbytes, bytes, interval); + LOCKEDINT_MTX_UNLOCK(tsdn, counter->mtx); + return overflow; +} + +bool counter_accum_init(counter_accum_t *counter, uint64_t interval); +void counter_prefork(tsdn_t *tsdn, counter_accum_t *counter); +void counter_postfork_parent(tsdn_t *tsdn, counter_accum_t *counter); +void counter_postfork_child(tsdn_t *tsdn, counter_accum_t *counter); + +#endif /* JEMALLOC_INTERNAL_COUNTER_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/ctl.h b/contrib/jemalloc/include/jemalloc/internal/ctl.h --- a/contrib/jemalloc/include/jemalloc/internal/ctl.h +++ b/contrib/jemalloc/include/jemalloc/internal/ctl.h @@ -42,9 +42,11 @@ uint64_t nfills_small; uint64_t nflushes_small; - bin_stats_t bstats[SC_NBINS]; + bin_stats_data_t bstats[SC_NBINS]; arena_stats_large_t lstats[SC_NSIZES - SC_NBINS]; - arena_stats_extents_t estats[SC_NPSIZES]; + pac_estats_t estats[SC_NPSIZES]; + hpa_shard_stats_t hpastats; + sec_stats_t secstats; } ctl_arena_stats_t; typedef struct ctl_stats_s { @@ -96,13 +98,17 @@ int ctl_byname(tsd_t *tsd, const char *name, void *oldp, size_t *oldlenp, void *newp, size_t newlen); int ctl_nametomib(tsd_t *tsd, const char *name, size_t *mibp, size_t *miblenp); - int ctl_bymib(tsd_t *tsd, const size_t *mib, size_t miblen, void *oldp, size_t *oldlenp, void *newp, size_t newlen); +int ctl_mibnametomib(tsd_t *tsd, size_t *mib, size_t miblen, const char *name, + size_t *miblenp); +int ctl_bymibname(tsd_t *tsd, size_t *mib, size_t miblen, const char *name, + size_t *miblenp, void *oldp, size_t *oldlenp, void *newp, size_t newlen); bool ctl_boot(void); void ctl_prefork(tsdn_t *tsdn); void ctl_postfork_parent(tsdn_t *tsdn); void ctl_postfork_child(tsdn_t *tsdn); +void ctl_mtx_assert_held(tsdn_t *tsdn); #define xmallctl(name, oldp, oldlenp, newp, newlen) do { \ if (je_mallctl(name, oldp, oldlenp, newp, newlen) \ @@ -131,4 +137,23 @@ } \ } while (0) +#define xmallctlmibnametomib(mib, miblen, name, miblenp) do { \ + if (ctl_mibnametomib(tsd_fetch(), mib, miblen, name, miblenp) \ + != 0) { \ + malloc_write( \ + ": Failure in ctl_mibnametomib()\n"); \ + abort(); \ + } \ +} while (0) + +#define xmallctlbymibname(mib, miblen, name, miblenp, oldp, oldlenp, \ + newp, newlen) do { \ + if (ctl_bymibname(tsd_fetch(), mib, miblen, name, miblenp, \ + oldp, oldlenp, newp, newlen) != 0) { \ + malloc_write( \ + ": Failure in ctl_bymibname()\n"); \ + abort(); \ + } \ +} while (0) + #endif /* JEMALLOC_INTERNAL_CTL_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/decay.h b/contrib/jemalloc/include/jemalloc/internal/decay.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/decay.h @@ -0,0 +1,186 @@ +#ifndef JEMALLOC_INTERNAL_DECAY_H +#define JEMALLOC_INTERNAL_DECAY_H + +#include "jemalloc/internal/smoothstep.h" + +#define DECAY_UNBOUNDED_TIME_TO_PURGE ((uint64_t)-1) + +/* + * The decay_t computes the number of pages we should purge at any given time. + * Page allocators inform a decay object when pages enter a decay-able state + * (i.e. dirty or muzzy), and query it to determine how many pages should be + * purged at any given time. + * + * This is mostly a single-threaded data structure and doesn't care about + * synchronization at all; it's the caller's responsibility to manage their + * synchronization on their own. There are two exceptions: + * 1) It's OK to racily call decay_ms_read (i.e. just the simplest state query). + * 2) The mtx and purging fields live (and are initialized) here, but are + * logically owned by the page allocator. This is just a convenience (since + * those fields would be duplicated for both the dirty and muzzy states + * otherwise). + */ +typedef struct decay_s decay_t; +struct decay_s { + /* Synchronizes all non-atomic fields. */ + malloc_mutex_t mtx; + /* + * True if a thread is currently purging the extents associated with + * this decay structure. + */ + bool purging; + /* + * Approximate time in milliseconds from the creation of a set of unused + * dirty pages until an equivalent set of unused dirty pages is purged + * and/or reused. + */ + atomic_zd_t time_ms; + /* time / SMOOTHSTEP_NSTEPS. */ + nstime_t interval; + /* + * Time at which the current decay interval logically started. We do + * not actually advance to a new epoch until sometime after it starts + * because of scheduling and computation delays, and it is even possible + * to completely skip epochs. In all cases, during epoch advancement we + * merge all relevant activity into the most recently recorded epoch. + */ + nstime_t epoch; + /* Deadline randomness generator. */ + uint64_t jitter_state; + /* + * Deadline for current epoch. This is the sum of interval and per + * epoch jitter which is a uniform random variable in [0..interval). + * Epochs always advance by precise multiples of interval, but we + * randomize the deadline to reduce the likelihood of arenas purging in + * lockstep. + */ + nstime_t deadline; + /* + * The number of pages we cap ourselves at in the current epoch, per + * decay policies. Updated on an epoch change. After an epoch change, + * the caller should take steps to try to purge down to this amount. + */ + size_t npages_limit; + /* + * Number of unpurged pages at beginning of current epoch. During epoch + * advancement we use the delta between arena->decay_*.nunpurged and + * ecache_npages_get(&arena->ecache_*) to determine how many dirty pages, + * if any, were generated. + */ + size_t nunpurged; + /* + * Trailing log of how many unused dirty pages were generated during + * each of the past SMOOTHSTEP_NSTEPS decay epochs, where the last + * element is the most recent epoch. Corresponding epoch times are + * relative to epoch. + * + * Updated only on epoch advance, triggered by + * decay_maybe_advance_epoch, below. + */ + size_t backlog[SMOOTHSTEP_NSTEPS]; + + /* Peak number of pages in associated extents. Used for debug only. */ + uint64_t ceil_npages; +}; + +/* + * The current decay time setting. This is the only public access to a decay_t + * that's allowed without holding mtx. + */ +static inline ssize_t +decay_ms_read(const decay_t *decay) { + return atomic_load_zd(&decay->time_ms, ATOMIC_RELAXED); +} + +/* + * See the comment on the struct field -- the limit on pages we should allow in + * this decay state this epoch. + */ +static inline size_t +decay_npages_limit_get(const decay_t *decay) { + return decay->npages_limit; +} + +/* How many unused dirty pages were generated during the last epoch. */ +static inline size_t +decay_epoch_npages_delta(const decay_t *decay) { + return decay->backlog[SMOOTHSTEP_NSTEPS - 1]; +} + +/* + * Current epoch duration, in nanoseconds. Given that new epochs are started + * somewhat haphazardly, this is not necessarily exactly the time between any + * two calls to decay_maybe_advance_epoch; see the comments on fields in the + * decay_t. + */ +static inline uint64_t +decay_epoch_duration_ns(const decay_t *decay) { + return nstime_ns(&decay->interval); +} + +static inline bool +decay_immediately(const decay_t *decay) { + ssize_t decay_ms = decay_ms_read(decay); + return decay_ms == 0; +} + +static inline bool +decay_disabled(const decay_t *decay) { + ssize_t decay_ms = decay_ms_read(decay); + return decay_ms < 0; +} + +/* Returns true if decay is enabled and done gradually. */ +static inline bool +decay_gradually(const decay_t *decay) { + ssize_t decay_ms = decay_ms_read(decay); + return decay_ms > 0; +} + +/* + * Returns true if the passed in decay time setting is valid. + * < -1 : invalid + * -1 : never decay + * 0 : decay immediately + * > 0 : some positive decay time, up to a maximum allowed value of + * NSTIME_SEC_MAX * 1000, which corresponds to decaying somewhere in the early + * 27th century. By that time, we expect to have implemented alternate purging + * strategies. + */ +bool decay_ms_valid(ssize_t decay_ms); + +/* + * As a precondition, the decay_t must be zeroed out (as if with memset). + * + * Returns true on error. + */ +bool decay_init(decay_t *decay, nstime_t *cur_time, ssize_t decay_ms); + +/* + * Given an already-initialized decay_t, reinitialize it with the given decay + * time. The decay_t must have previously been initialized (and should not then + * be zeroed). + */ +void decay_reinit(decay_t *decay, nstime_t *cur_time, ssize_t decay_ms); + +/* + * Compute how many of 'npages_new' pages we would need to purge in 'time'. + */ +uint64_t decay_npages_purge_in(decay_t *decay, nstime_t *time, + size_t npages_new); + +/* Returns true if the epoch advanced and there are pages to purge. */ +bool decay_maybe_advance_epoch(decay_t *decay, nstime_t *new_time, + size_t current_npages); + +/* + * Calculates wait time until a number of pages in the interval + * [0.5 * npages_threshold .. 1.5 * npages_threshold] should be purged. + * + * Returns number of nanoseconds or DECAY_UNBOUNDED_TIME_TO_PURGE in case of + * indefinite wait. + */ +uint64_t decay_ns_until_purge(decay_t *decay, size_t npages_current, + uint64_t npages_threshold); + +#endif /* JEMALLOC_INTERNAL_DECAY_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/ecache.h b/contrib/jemalloc/include/jemalloc/internal/ecache.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/ecache.h @@ -0,0 +1,55 @@ +#ifndef JEMALLOC_INTERNAL_ECACHE_H +#define JEMALLOC_INTERNAL_ECACHE_H + +#include "jemalloc/internal/eset.h" +#include "jemalloc/internal/san.h" +#include "jemalloc/internal/mutex.h" + +typedef struct ecache_s ecache_t; +struct ecache_s { + malloc_mutex_t mtx; + eset_t eset; + eset_t guarded_eset; + /* All stored extents must be in the same state. */ + extent_state_t state; + /* The index of the ehooks the ecache is associated with. */ + unsigned ind; + /* + * If true, delay coalescing until eviction; otherwise coalesce during + * deallocation. + */ + bool delay_coalesce; +}; + +static inline size_t +ecache_npages_get(ecache_t *ecache) { + return eset_npages_get(&ecache->eset) + + eset_npages_get(&ecache->guarded_eset); +} + +/* Get the number of extents in the given page size index. */ +static inline size_t +ecache_nextents_get(ecache_t *ecache, pszind_t ind) { + return eset_nextents_get(&ecache->eset, ind) + + eset_nextents_get(&ecache->guarded_eset, ind); +} + +/* Get the sum total bytes of the extents in the given page size index. */ +static inline size_t +ecache_nbytes_get(ecache_t *ecache, pszind_t ind) { + return eset_nbytes_get(&ecache->eset, ind) + + eset_nbytes_get(&ecache->guarded_eset, ind); +} + +static inline unsigned +ecache_ind_get(ecache_t *ecache) { + return ecache->ind; +} + +bool ecache_init(tsdn_t *tsdn, ecache_t *ecache, extent_state_t state, + unsigned ind, bool delay_coalesce); +void ecache_prefork(tsdn_t *tsdn, ecache_t *ecache); +void ecache_postfork_parent(tsdn_t *tsdn, ecache_t *ecache); +void ecache_postfork_child(tsdn_t *tsdn, ecache_t *ecache); + +#endif /* JEMALLOC_INTERNAL_ECACHE_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/edata.h b/contrib/jemalloc/include/jemalloc/internal/edata.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/edata.h @@ -0,0 +1,698 @@ +#ifndef JEMALLOC_INTERNAL_EDATA_H +#define JEMALLOC_INTERNAL_EDATA_H + +#include "jemalloc/internal/atomic.h" +#include "jemalloc/internal/bin_info.h" +#include "jemalloc/internal/bit_util.h" +#include "jemalloc/internal/hpdata.h" +#include "jemalloc/internal/nstime.h" +#include "jemalloc/internal/ph.h" +#include "jemalloc/internal/ql.h" +#include "jemalloc/internal/sc.h" +#include "jemalloc/internal/slab_data.h" +#include "jemalloc/internal/sz.h" +#include "jemalloc/internal/typed_list.h" + +/* + * sizeof(edata_t) is 128 bytes on 64-bit architectures. Ensure the alignment + * to free up the low bits in the rtree leaf. + */ +#define EDATA_ALIGNMENT 128 + +enum extent_state_e { + extent_state_active = 0, + extent_state_dirty = 1, + extent_state_muzzy = 2, + extent_state_retained = 3, + extent_state_transition = 4, /* States below are intermediate. */ + extent_state_merging = 5, + extent_state_max = 5 /* Sanity checking only. */ +}; +typedef enum extent_state_e extent_state_t; + +enum extent_head_state_e { + EXTENT_NOT_HEAD, + EXTENT_IS_HEAD /* See comments in ehooks_default_merge_impl(). */ +}; +typedef enum extent_head_state_e extent_head_state_t; + +/* + * Which implementation of the page allocator interface, (PAI, defined in + * pai.h) owns the given extent? + */ +enum extent_pai_e { + EXTENT_PAI_PAC = 0, + EXTENT_PAI_HPA = 1 +}; +typedef enum extent_pai_e extent_pai_t; + +struct e_prof_info_s { + /* Time when this was allocated. */ + nstime_t e_prof_alloc_time; + /* Allocation request size. */ + size_t e_prof_alloc_size; + /* Points to a prof_tctx_t. */ + atomic_p_t e_prof_tctx; + /* + * Points to a prof_recent_t for the allocation; NULL + * means the recent allocation record no longer exists. + * Protected by prof_recent_alloc_mtx. + */ + atomic_p_t e_prof_recent_alloc; +}; +typedef struct e_prof_info_s e_prof_info_t; + +/* + * The information about a particular edata that lives in an emap. Space is + * more precious there (the information, plus the edata pointer, has to live in + * a 64-bit word if we want to enable a packed representation. + * + * There are two things that are special about the information here: + * - It's quicker to access. You have one fewer pointer hop, since finding the + * edata_t associated with an item always requires accessing the rtree leaf in + * which this data is stored. + * - It can be read unsynchronized, and without worrying about lifetime issues. + */ +typedef struct edata_map_info_s edata_map_info_t; +struct edata_map_info_s { + bool slab; + szind_t szind; +}; + +typedef struct edata_cmp_summary_s edata_cmp_summary_t; +struct edata_cmp_summary_s { + uint64_t sn; + uintptr_t addr; +}; + +/* Extent (span of pages). Use accessor functions for e_* fields. */ +typedef struct edata_s edata_t; +ph_structs(edata_avail, edata_t); +ph_structs(edata_heap, edata_t); +struct edata_s { + /* + * Bitfield containing several fields: + * + * a: arena_ind + * b: slab + * c: committed + * p: pai + * z: zeroed + * g: guarded + * t: state + * i: szind + * f: nfree + * s: bin_shard + * + * 00000000 ... 0000ssss ssffffff ffffiiii iiiitttg zpcbaaaa aaaaaaaa + * + * arena_ind: Arena from which this extent came, or all 1 bits if + * unassociated. + * + * slab: The slab flag indicates whether the extent is used for a slab + * of small regions. This helps differentiate small size classes, + * and it indicates whether interior pointers can be looked up via + * iealloc(). + * + * committed: The committed flag indicates whether physical memory is + * committed to the extent, whether explicitly or implicitly + * as on a system that overcommits and satisfies physical + * memory needs on demand via soft page faults. + * + * pai: The pai flag is an extent_pai_t. + * + * zeroed: The zeroed flag is used by extent recycling code to track + * whether memory is zero-filled. + * + * guarded: The guarded flag is use by the sanitizer to track whether + * the extent has page guards around it. + * + * state: The state flag is an extent_state_t. + * + * szind: The szind flag indicates usable size class index for + * allocations residing in this extent, regardless of whether the + * extent is a slab. Extent size and usable size often differ + * even for non-slabs, either due to sz_large_pad or promotion of + * sampled small regions. + * + * nfree: Number of free regions in slab. + * + * bin_shard: the shard of the bin from which this extent came. + */ + uint64_t e_bits; +#define MASK(CURRENT_FIELD_WIDTH, CURRENT_FIELD_SHIFT) ((((((uint64_t)0x1U) << (CURRENT_FIELD_WIDTH)) - 1)) << (CURRENT_FIELD_SHIFT)) + +#define EDATA_BITS_ARENA_WIDTH MALLOCX_ARENA_BITS +#define EDATA_BITS_ARENA_SHIFT 0 +#define EDATA_BITS_ARENA_MASK MASK(EDATA_BITS_ARENA_WIDTH, EDATA_BITS_ARENA_SHIFT) + +#define EDATA_BITS_SLAB_WIDTH 1 +#define EDATA_BITS_SLAB_SHIFT (EDATA_BITS_ARENA_WIDTH + EDATA_BITS_ARENA_SHIFT) +#define EDATA_BITS_SLAB_MASK MASK(EDATA_BITS_SLAB_WIDTH, EDATA_BITS_SLAB_SHIFT) + +#define EDATA_BITS_COMMITTED_WIDTH 1 +#define EDATA_BITS_COMMITTED_SHIFT (EDATA_BITS_SLAB_WIDTH + EDATA_BITS_SLAB_SHIFT) +#define EDATA_BITS_COMMITTED_MASK MASK(EDATA_BITS_COMMITTED_WIDTH, EDATA_BITS_COMMITTED_SHIFT) + +#define EDATA_BITS_PAI_WIDTH 1 +#define EDATA_BITS_PAI_SHIFT (EDATA_BITS_COMMITTED_WIDTH + EDATA_BITS_COMMITTED_SHIFT) +#define EDATA_BITS_PAI_MASK MASK(EDATA_BITS_PAI_WIDTH, EDATA_BITS_PAI_SHIFT) + +#define EDATA_BITS_ZEROED_WIDTH 1 +#define EDATA_BITS_ZEROED_SHIFT (EDATA_BITS_PAI_WIDTH + EDATA_BITS_PAI_SHIFT) +#define EDATA_BITS_ZEROED_MASK MASK(EDATA_BITS_ZEROED_WIDTH, EDATA_BITS_ZEROED_SHIFT) + +#define EDATA_BITS_GUARDED_WIDTH 1 +#define EDATA_BITS_GUARDED_SHIFT (EDATA_BITS_ZEROED_WIDTH + EDATA_BITS_ZEROED_SHIFT) +#define EDATA_BITS_GUARDED_MASK MASK(EDATA_BITS_GUARDED_WIDTH, EDATA_BITS_GUARDED_SHIFT) + +#define EDATA_BITS_STATE_WIDTH 3 +#define EDATA_BITS_STATE_SHIFT (EDATA_BITS_GUARDED_WIDTH + EDATA_BITS_GUARDED_SHIFT) +#define EDATA_BITS_STATE_MASK MASK(EDATA_BITS_STATE_WIDTH, EDATA_BITS_STATE_SHIFT) + +#define EDATA_BITS_SZIND_WIDTH LG_CEIL(SC_NSIZES) +#define EDATA_BITS_SZIND_SHIFT (EDATA_BITS_STATE_WIDTH + EDATA_BITS_STATE_SHIFT) +#define EDATA_BITS_SZIND_MASK MASK(EDATA_BITS_SZIND_WIDTH, EDATA_BITS_SZIND_SHIFT) + +#define EDATA_BITS_NFREE_WIDTH (SC_LG_SLAB_MAXREGS + 1) +#define EDATA_BITS_NFREE_SHIFT (EDATA_BITS_SZIND_WIDTH + EDATA_BITS_SZIND_SHIFT) +#define EDATA_BITS_NFREE_MASK MASK(EDATA_BITS_NFREE_WIDTH, EDATA_BITS_NFREE_SHIFT) + +#define EDATA_BITS_BINSHARD_WIDTH 6 +#define EDATA_BITS_BINSHARD_SHIFT (EDATA_BITS_NFREE_WIDTH + EDATA_BITS_NFREE_SHIFT) +#define EDATA_BITS_BINSHARD_MASK MASK(EDATA_BITS_BINSHARD_WIDTH, EDATA_BITS_BINSHARD_SHIFT) + +#define EDATA_BITS_IS_HEAD_WIDTH 1 +#define EDATA_BITS_IS_HEAD_SHIFT (EDATA_BITS_BINSHARD_WIDTH + EDATA_BITS_BINSHARD_SHIFT) +#define EDATA_BITS_IS_HEAD_MASK MASK(EDATA_BITS_IS_HEAD_WIDTH, EDATA_BITS_IS_HEAD_SHIFT) + + /* Pointer to the extent that this structure is responsible for. */ + void *e_addr; + + union { + /* + * Extent size and serial number associated with the extent + * structure (different than the serial number for the extent at + * e_addr). + * + * ssssssss [...] ssssssss ssssnnnn nnnnnnnn + */ + size_t e_size_esn; + #define EDATA_SIZE_MASK ((size_t)~(PAGE-1)) + #define EDATA_ESN_MASK ((size_t)PAGE-1) + /* Base extent size, which may not be a multiple of PAGE. */ + size_t e_bsize; + }; + + /* + * If this edata is a user allocation from an HPA, it comes out of some + * pageslab (we don't yet support huegpage allocations that don't fit + * into pageslabs). This tracks it. + */ + hpdata_t *e_ps; + + /* + * Serial number. These are not necessarily unique; splitting an extent + * results in two extents with the same serial number. + */ + uint64_t e_sn; + + union { + /* + * List linkage used when the edata_t is active; either in + * arena's large allocations or bin_t's slabs_full. + */ + ql_elm(edata_t) ql_link_active; + /* + * Pairing heap linkage. Used whenever the extent is inactive + * (in the page allocators), or when it is active and in + * slabs_nonfull, or when the edata_t is unassociated with an + * extent and sitting in an edata_cache. + */ + union { + edata_heap_link_t heap_link; + edata_avail_link_t avail_link; + }; + }; + + union { + /* + * List linkage used when the extent is inactive: + * - Stashed dirty extents + * - Ecache LRU functionality. + */ + ql_elm(edata_t) ql_link_inactive; + /* Small region slab metadata. */ + slab_data_t e_slab_data; + + /* Profiling data, used for large objects. */ + e_prof_info_t e_prof_info; + }; +}; + +TYPED_LIST(edata_list_active, edata_t, ql_link_active) +TYPED_LIST(edata_list_inactive, edata_t, ql_link_inactive) + +static inline unsigned +edata_arena_ind_get(const edata_t *edata) { + unsigned arena_ind = (unsigned)((edata->e_bits & + EDATA_BITS_ARENA_MASK) >> EDATA_BITS_ARENA_SHIFT); + assert(arena_ind < MALLOCX_ARENA_LIMIT); + + return arena_ind; +} + +static inline szind_t +edata_szind_get_maybe_invalid(const edata_t *edata) { + szind_t szind = (szind_t)((edata->e_bits & EDATA_BITS_SZIND_MASK) >> + EDATA_BITS_SZIND_SHIFT); + assert(szind <= SC_NSIZES); + return szind; +} + +static inline szind_t +edata_szind_get(const edata_t *edata) { + szind_t szind = edata_szind_get_maybe_invalid(edata); + assert(szind < SC_NSIZES); /* Never call when "invalid". */ + return szind; +} + +static inline size_t +edata_usize_get(const edata_t *edata) { + return sz_index2size(edata_szind_get(edata)); +} + +static inline unsigned +edata_binshard_get(const edata_t *edata) { + unsigned binshard = (unsigned)((edata->e_bits & + EDATA_BITS_BINSHARD_MASK) >> EDATA_BITS_BINSHARD_SHIFT); + assert(binshard < bin_infos[edata_szind_get(edata)].n_shards); + return binshard; +} + +static inline uint64_t +edata_sn_get(const edata_t *edata) { + return edata->e_sn; +} + +static inline extent_state_t +edata_state_get(const edata_t *edata) { + return (extent_state_t)((edata->e_bits & EDATA_BITS_STATE_MASK) >> + EDATA_BITS_STATE_SHIFT); +} + +static inline bool +edata_guarded_get(const edata_t *edata) { + return (bool)((edata->e_bits & EDATA_BITS_GUARDED_MASK) >> + EDATA_BITS_GUARDED_SHIFT); +} + +static inline bool +edata_zeroed_get(const edata_t *edata) { + return (bool)((edata->e_bits & EDATA_BITS_ZEROED_MASK) >> + EDATA_BITS_ZEROED_SHIFT); +} + +static inline bool +edata_committed_get(const edata_t *edata) { + return (bool)((edata->e_bits & EDATA_BITS_COMMITTED_MASK) >> + EDATA_BITS_COMMITTED_SHIFT); +} + +static inline extent_pai_t +edata_pai_get(const edata_t *edata) { + return (extent_pai_t)((edata->e_bits & EDATA_BITS_PAI_MASK) >> + EDATA_BITS_PAI_SHIFT); +} + +static inline bool +edata_slab_get(const edata_t *edata) { + return (bool)((edata->e_bits & EDATA_BITS_SLAB_MASK) >> + EDATA_BITS_SLAB_SHIFT); +} + +static inline unsigned +edata_nfree_get(const edata_t *edata) { + assert(edata_slab_get(edata)); + return (unsigned)((edata->e_bits & EDATA_BITS_NFREE_MASK) >> + EDATA_BITS_NFREE_SHIFT); +} + +static inline void * +edata_base_get(const edata_t *edata) { + assert(edata->e_addr == PAGE_ADDR2BASE(edata->e_addr) || + !edata_slab_get(edata)); + return PAGE_ADDR2BASE(edata->e_addr); +} + +static inline void * +edata_addr_get(const edata_t *edata) { + assert(edata->e_addr == PAGE_ADDR2BASE(edata->e_addr) || + !edata_slab_get(edata)); + return edata->e_addr; +} + +static inline size_t +edata_size_get(const edata_t *edata) { + return (edata->e_size_esn & EDATA_SIZE_MASK); +} + +static inline size_t +edata_esn_get(const edata_t *edata) { + return (edata->e_size_esn & EDATA_ESN_MASK); +} + +static inline size_t +edata_bsize_get(const edata_t *edata) { + return edata->e_bsize; +} + +static inline hpdata_t * +edata_ps_get(const edata_t *edata) { + assert(edata_pai_get(edata) == EXTENT_PAI_HPA); + return edata->e_ps; +} + +static inline void * +edata_before_get(const edata_t *edata) { + return (void *)((uintptr_t)edata_base_get(edata) - PAGE); +} + +static inline void * +edata_last_get(const edata_t *edata) { + return (void *)((uintptr_t)edata_base_get(edata) + + edata_size_get(edata) - PAGE); +} + +static inline void * +edata_past_get(const edata_t *edata) { + return (void *)((uintptr_t)edata_base_get(edata) + + edata_size_get(edata)); +} + +static inline slab_data_t * +edata_slab_data_get(edata_t *edata) { + assert(edata_slab_get(edata)); + return &edata->e_slab_data; +} + +static inline const slab_data_t * +edata_slab_data_get_const(const edata_t *edata) { + assert(edata_slab_get(edata)); + return &edata->e_slab_data; +} + +static inline prof_tctx_t * +edata_prof_tctx_get(const edata_t *edata) { + return (prof_tctx_t *)atomic_load_p(&edata->e_prof_info.e_prof_tctx, + ATOMIC_ACQUIRE); +} + +static inline const nstime_t * +edata_prof_alloc_time_get(const edata_t *edata) { + return &edata->e_prof_info.e_prof_alloc_time; +} + +static inline size_t +edata_prof_alloc_size_get(const edata_t *edata) { + return edata->e_prof_info.e_prof_alloc_size; +} + +static inline prof_recent_t * +edata_prof_recent_alloc_get_dont_call_directly(const edata_t *edata) { + return (prof_recent_t *)atomic_load_p( + &edata->e_prof_info.e_prof_recent_alloc, ATOMIC_RELAXED); +} + +static inline void +edata_arena_ind_set(edata_t *edata, unsigned arena_ind) { + edata->e_bits = (edata->e_bits & ~EDATA_BITS_ARENA_MASK) | + ((uint64_t)arena_ind << EDATA_BITS_ARENA_SHIFT); +} + +static inline void +edata_binshard_set(edata_t *edata, unsigned binshard) { + /* The assertion assumes szind is set already. */ + assert(binshard < bin_infos[edata_szind_get(edata)].n_shards); + edata->e_bits = (edata->e_bits & ~EDATA_BITS_BINSHARD_MASK) | + ((uint64_t)binshard << EDATA_BITS_BINSHARD_SHIFT); +} + +static inline void +edata_addr_set(edata_t *edata, void *addr) { + edata->e_addr = addr; +} + +static inline void +edata_size_set(edata_t *edata, size_t size) { + assert((size & ~EDATA_SIZE_MASK) == 0); + edata->e_size_esn = size | (edata->e_size_esn & ~EDATA_SIZE_MASK); +} + +static inline void +edata_esn_set(edata_t *edata, size_t esn) { + edata->e_size_esn = (edata->e_size_esn & ~EDATA_ESN_MASK) | (esn & + EDATA_ESN_MASK); +} + +static inline void +edata_bsize_set(edata_t *edata, size_t bsize) { + edata->e_bsize = bsize; +} + +static inline void +edata_ps_set(edata_t *edata, hpdata_t *ps) { + assert(edata_pai_get(edata) == EXTENT_PAI_HPA); + edata->e_ps = ps; +} + +static inline void +edata_szind_set(edata_t *edata, szind_t szind) { + assert(szind <= SC_NSIZES); /* SC_NSIZES means "invalid". */ + edata->e_bits = (edata->e_bits & ~EDATA_BITS_SZIND_MASK) | + ((uint64_t)szind << EDATA_BITS_SZIND_SHIFT); +} + +static inline void +edata_nfree_set(edata_t *edata, unsigned nfree) { + assert(edata_slab_get(edata)); + edata->e_bits = (edata->e_bits & ~EDATA_BITS_NFREE_MASK) | + ((uint64_t)nfree << EDATA_BITS_NFREE_SHIFT); +} + +static inline void +edata_nfree_binshard_set(edata_t *edata, unsigned nfree, unsigned binshard) { + /* The assertion assumes szind is set already. */ + assert(binshard < bin_infos[edata_szind_get(edata)].n_shards); + edata->e_bits = (edata->e_bits & + (~EDATA_BITS_NFREE_MASK & ~EDATA_BITS_BINSHARD_MASK)) | + ((uint64_t)binshard << EDATA_BITS_BINSHARD_SHIFT) | + ((uint64_t)nfree << EDATA_BITS_NFREE_SHIFT); +} + +static inline void +edata_nfree_inc(edata_t *edata) { + assert(edata_slab_get(edata)); + edata->e_bits += ((uint64_t)1U << EDATA_BITS_NFREE_SHIFT); +} + +static inline void +edata_nfree_dec(edata_t *edata) { + assert(edata_slab_get(edata)); + edata->e_bits -= ((uint64_t)1U << EDATA_BITS_NFREE_SHIFT); +} + +static inline void +edata_nfree_sub(edata_t *edata, uint64_t n) { + assert(edata_slab_get(edata)); + edata->e_bits -= (n << EDATA_BITS_NFREE_SHIFT); +} + +static inline void +edata_sn_set(edata_t *edata, uint64_t sn) { + edata->e_sn = sn; +} + +static inline void +edata_state_set(edata_t *edata, extent_state_t state) { + edata->e_bits = (edata->e_bits & ~EDATA_BITS_STATE_MASK) | + ((uint64_t)state << EDATA_BITS_STATE_SHIFT); +} + +static inline void +edata_guarded_set(edata_t *edata, bool guarded) { + edata->e_bits = (edata->e_bits & ~EDATA_BITS_GUARDED_MASK) | + ((uint64_t)guarded << EDATA_BITS_GUARDED_SHIFT); +} + +static inline void +edata_zeroed_set(edata_t *edata, bool zeroed) { + edata->e_bits = (edata->e_bits & ~EDATA_BITS_ZEROED_MASK) | + ((uint64_t)zeroed << EDATA_BITS_ZEROED_SHIFT); +} + +static inline void +edata_committed_set(edata_t *edata, bool committed) { + edata->e_bits = (edata->e_bits & ~EDATA_BITS_COMMITTED_MASK) | + ((uint64_t)committed << EDATA_BITS_COMMITTED_SHIFT); +} + +static inline void +edata_pai_set(edata_t *edata, extent_pai_t pai) { + edata->e_bits = (edata->e_bits & ~EDATA_BITS_PAI_MASK) | + ((uint64_t)pai << EDATA_BITS_PAI_SHIFT); +} + +static inline void +edata_slab_set(edata_t *edata, bool slab) { + edata->e_bits = (edata->e_bits & ~EDATA_BITS_SLAB_MASK) | + ((uint64_t)slab << EDATA_BITS_SLAB_SHIFT); +} + +static inline void +edata_prof_tctx_set(edata_t *edata, prof_tctx_t *tctx) { + atomic_store_p(&edata->e_prof_info.e_prof_tctx, tctx, ATOMIC_RELEASE); +} + +static inline void +edata_prof_alloc_time_set(edata_t *edata, nstime_t *t) { + nstime_copy(&edata->e_prof_info.e_prof_alloc_time, t); +} + +static inline void +edata_prof_alloc_size_set(edata_t *edata, size_t size) { + edata->e_prof_info.e_prof_alloc_size = size; +} + +static inline void +edata_prof_recent_alloc_set_dont_call_directly(edata_t *edata, + prof_recent_t *recent_alloc) { + atomic_store_p(&edata->e_prof_info.e_prof_recent_alloc, recent_alloc, + ATOMIC_RELAXED); +} + +static inline bool +edata_is_head_get(edata_t *edata) { + return (bool)((edata->e_bits & EDATA_BITS_IS_HEAD_MASK) >> + EDATA_BITS_IS_HEAD_SHIFT); +} + +static inline void +edata_is_head_set(edata_t *edata, bool is_head) { + edata->e_bits = (edata->e_bits & ~EDATA_BITS_IS_HEAD_MASK) | + ((uint64_t)is_head << EDATA_BITS_IS_HEAD_SHIFT); +} + +static inline bool +edata_state_in_transition(extent_state_t state) { + return state >= extent_state_transition; +} + +/* + * Because this function is implemented as a sequence of bitfield modifications, + * even though each individual bit is properly initialized, we technically read + * uninitialized data within it. This is mostly fine, since most callers get + * their edatas from zeroing sources, but callers who make stack edata_ts need + * to manually zero them. + */ +static inline void +edata_init(edata_t *edata, unsigned arena_ind, void *addr, size_t size, + bool slab, szind_t szind, uint64_t sn, extent_state_t state, bool zeroed, + bool committed, extent_pai_t pai, extent_head_state_t is_head) { + assert(addr == PAGE_ADDR2BASE(addr) || !slab); + + edata_arena_ind_set(edata, arena_ind); + edata_addr_set(edata, addr); + edata_size_set(edata, size); + edata_slab_set(edata, slab); + edata_szind_set(edata, szind); + edata_sn_set(edata, sn); + edata_state_set(edata, state); + edata_guarded_set(edata, false); + edata_zeroed_set(edata, zeroed); + edata_committed_set(edata, committed); + edata_pai_set(edata, pai); + edata_is_head_set(edata, is_head == EXTENT_IS_HEAD); + if (config_prof) { + edata_prof_tctx_set(edata, NULL); + } +} + +static inline void +edata_binit(edata_t *edata, void *addr, size_t bsize, uint64_t sn) { + edata_arena_ind_set(edata, (1U << MALLOCX_ARENA_BITS) - 1); + edata_addr_set(edata, addr); + edata_bsize_set(edata, bsize); + edata_slab_set(edata, false); + edata_szind_set(edata, SC_NSIZES); + edata_sn_set(edata, sn); + edata_state_set(edata, extent_state_active); + edata_guarded_set(edata, false); + edata_zeroed_set(edata, true); + edata_committed_set(edata, true); + /* + * This isn't strictly true, but base allocated extents never get + * deallocated and can't be looked up in the emap, but no sense in + * wasting a state bit to encode this fact. + */ + edata_pai_set(edata, EXTENT_PAI_PAC); +} + +static inline int +edata_esn_comp(const edata_t *a, const edata_t *b) { + size_t a_esn = edata_esn_get(a); + size_t b_esn = edata_esn_get(b); + + return (a_esn > b_esn) - (a_esn < b_esn); +} + +static inline int +edata_ead_comp(const edata_t *a, const edata_t *b) { + uintptr_t a_eaddr = (uintptr_t)a; + uintptr_t b_eaddr = (uintptr_t)b; + + return (a_eaddr > b_eaddr) - (a_eaddr < b_eaddr); +} + +static inline edata_cmp_summary_t +edata_cmp_summary_get(const edata_t *edata) { + return (edata_cmp_summary_t){edata_sn_get(edata), + (uintptr_t)edata_addr_get(edata)}; +} + +static inline int +edata_cmp_summary_comp(edata_cmp_summary_t a, edata_cmp_summary_t b) { + int ret; + ret = (a.sn > b.sn) - (a.sn < b.sn); + if (ret != 0) { + return ret; + } + ret = (a.addr > b.addr) - (a.addr < b.addr); + return ret; +} + +static inline int +edata_snad_comp(const edata_t *a, const edata_t *b) { + edata_cmp_summary_t a_cmp = edata_cmp_summary_get(a); + edata_cmp_summary_t b_cmp = edata_cmp_summary_get(b); + + return edata_cmp_summary_comp(a_cmp, b_cmp); +} + +static inline int +edata_esnead_comp(const edata_t *a, const edata_t *b) { + int ret; + + ret = edata_esn_comp(a, b); + if (ret != 0) { + return ret; + } + + ret = edata_ead_comp(a, b); + return ret; +} + +ph_proto(, edata_avail, edata_t) +ph_proto(, edata_heap, edata_t) + +#endif /* JEMALLOC_INTERNAL_EDATA_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/edata_cache.h b/contrib/jemalloc/include/jemalloc/internal/edata_cache.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/edata_cache.h @@ -0,0 +1,49 @@ +#ifndef JEMALLOC_INTERNAL_EDATA_CACHE_H +#define JEMALLOC_INTERNAL_EDATA_CACHE_H + +#include "jemalloc/internal/base.h" + +/* For tests only. */ +#define EDATA_CACHE_FAST_FILL 4 + +/* + * A cache of edata_t structures allocated via base_alloc_edata (as opposed to + * the underlying extents they describe). The contents of returned edata_t + * objects are garbage and cannot be relied upon. + */ + +typedef struct edata_cache_s edata_cache_t; +struct edata_cache_s { + edata_avail_t avail; + atomic_zu_t count; + malloc_mutex_t mtx; + base_t *base; +}; + +bool edata_cache_init(edata_cache_t *edata_cache, base_t *base); +edata_t *edata_cache_get(tsdn_t *tsdn, edata_cache_t *edata_cache); +void edata_cache_put(tsdn_t *tsdn, edata_cache_t *edata_cache, edata_t *edata); + +void edata_cache_prefork(tsdn_t *tsdn, edata_cache_t *edata_cache); +void edata_cache_postfork_parent(tsdn_t *tsdn, edata_cache_t *edata_cache); +void edata_cache_postfork_child(tsdn_t *tsdn, edata_cache_t *edata_cache); + +/* + * An edata_cache_small is like an edata_cache, but it relies on external + * synchronization and avoids first-fit strategies. + */ + +typedef struct edata_cache_fast_s edata_cache_fast_t; +struct edata_cache_fast_s { + edata_list_inactive_t list; + edata_cache_t *fallback; + bool disabled; +}; + +void edata_cache_fast_init(edata_cache_fast_t *ecs, edata_cache_t *fallback); +edata_t *edata_cache_fast_get(tsdn_t *tsdn, edata_cache_fast_t *ecs); +void edata_cache_fast_put(tsdn_t *tsdn, edata_cache_fast_t *ecs, + edata_t *edata); +void edata_cache_fast_disable(tsdn_t *tsdn, edata_cache_fast_t *ecs); + +#endif /* JEMALLOC_INTERNAL_EDATA_CACHE_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/ehooks.h b/contrib/jemalloc/include/jemalloc/internal/ehooks.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/ehooks.h @@ -0,0 +1,412 @@ +#ifndef JEMALLOC_INTERNAL_EHOOKS_H +#define JEMALLOC_INTERNAL_EHOOKS_H + +#include "jemalloc/internal/atomic.h" +#include "jemalloc/internal/extent_mmap.h" + +/* + * This module is the internal interface to the extent hooks (both + * user-specified and external). Eventually, this will give us the flexibility + * to use multiple different versions of user-visible extent-hook APIs under a + * single user interface. + * + * Current API expansions (not available to anyone but the default hooks yet): + * - Head state tracking. Hooks can decide whether or not to merge two + * extents based on whether or not one of them is the head (i.e. was + * allocated on its own). The later extent loses its "head" status. + */ + +extern const extent_hooks_t ehooks_default_extent_hooks; + +typedef struct ehooks_s ehooks_t; +struct ehooks_s { + /* + * The user-visible id that goes with the ehooks (i.e. that of the base + * they're a part of, the associated arena's index within the arenas + * array). + */ + unsigned ind; + /* Logically an extent_hooks_t *. */ + atomic_p_t ptr; +}; + +extern const extent_hooks_t ehooks_default_extent_hooks; + +/* + * These are not really part of the public API. Each hook has a fast-path for + * the default-hooks case that can avoid various small inefficiencies: + * - Forgetting tsd and then calling tsd_get within the hook. + * - Getting more state than necessary out of the extent_t. + * - Doing arena_ind -> arena -> arena_ind lookups. + * By making the calls to these functions visible to the compiler, it can move + * those extra bits of computation down below the fast-paths where they get ignored. + */ +void *ehooks_default_alloc_impl(tsdn_t *tsdn, void *new_addr, size_t size, + size_t alignment, bool *zero, bool *commit, unsigned arena_ind); +bool ehooks_default_dalloc_impl(void *addr, size_t size); +void ehooks_default_destroy_impl(void *addr, size_t size); +bool ehooks_default_commit_impl(void *addr, size_t offset, size_t length); +bool ehooks_default_decommit_impl(void *addr, size_t offset, size_t length); +#ifdef PAGES_CAN_PURGE_LAZY +bool ehooks_default_purge_lazy_impl(void *addr, size_t offset, size_t length); +#endif +#ifdef PAGES_CAN_PURGE_FORCED +bool ehooks_default_purge_forced_impl(void *addr, size_t offset, size_t length); +#endif +bool ehooks_default_split_impl(); +/* + * Merge is the only default extent hook we declare -- see the comment in + * ehooks_merge. + */ +bool ehooks_default_merge(extent_hooks_t *extent_hooks, void *addr_a, + size_t size_a, void *addr_b, size_t size_b, bool committed, + unsigned arena_ind); +bool ehooks_default_merge_impl(tsdn_t *tsdn, void *addr_a, void *addr_b); +void ehooks_default_zero_impl(void *addr, size_t size); +void ehooks_default_guard_impl(void *guard1, void *guard2); +void ehooks_default_unguard_impl(void *guard1, void *guard2); + +/* + * We don't officially support reentrancy from wtihin the extent hooks. But + * various people who sit within throwing distance of the jemalloc team want + * that functionality in certain limited cases. The default reentrancy guards + * assert that we're not reentrant from a0 (since it's the bootstrap arena, + * where reentrant allocations would be redirected), which we would incorrectly + * trigger in cases where a0 has extent hooks (those hooks themselves can't be + * reentrant, then, but there are reasonable uses for such functionality, like + * putting internal metadata on hugepages). Therefore, we use the raw + * reentrancy guards. + * + * Eventually, we need to think more carefully about whether and where we + * support allocating from within extent hooks (and what that means for things + * like profiling, stats collection, etc.), and document what the guarantee is. + */ +static inline void +ehooks_pre_reentrancy(tsdn_t *tsdn) { + tsd_t *tsd = tsdn_null(tsdn) ? tsd_fetch() : tsdn_tsd(tsdn); + tsd_pre_reentrancy_raw(tsd); +} + +static inline void +ehooks_post_reentrancy(tsdn_t *tsdn) { + tsd_t *tsd = tsdn_null(tsdn) ? tsd_fetch() : tsdn_tsd(tsdn); + tsd_post_reentrancy_raw(tsd); +} + +/* Beginning of the public API. */ +void ehooks_init(ehooks_t *ehooks, extent_hooks_t *extent_hooks, unsigned ind); + +static inline unsigned +ehooks_ind_get(const ehooks_t *ehooks) { + return ehooks->ind; +} + +static inline void +ehooks_set_extent_hooks_ptr(ehooks_t *ehooks, extent_hooks_t *extent_hooks) { + atomic_store_p(&ehooks->ptr, extent_hooks, ATOMIC_RELEASE); +} + +static inline extent_hooks_t * +ehooks_get_extent_hooks_ptr(ehooks_t *ehooks) { + return (extent_hooks_t *)atomic_load_p(&ehooks->ptr, ATOMIC_ACQUIRE); +} + +static inline bool +ehooks_are_default(ehooks_t *ehooks) { + return ehooks_get_extent_hooks_ptr(ehooks) == + &ehooks_default_extent_hooks; +} + +/* + * In some cases, a caller needs to allocate resources before attempting to call + * a hook. If that hook is doomed to fail, this is wasteful. We therefore + * include some checks for such cases. + */ +static inline bool +ehooks_dalloc_will_fail(ehooks_t *ehooks) { + if (ehooks_are_default(ehooks)) { + return opt_retain; + } else { + return ehooks_get_extent_hooks_ptr(ehooks)->dalloc == NULL; + } +} + +static inline bool +ehooks_split_will_fail(ehooks_t *ehooks) { + return ehooks_get_extent_hooks_ptr(ehooks)->split == NULL; +} + +static inline bool +ehooks_merge_will_fail(ehooks_t *ehooks) { + return ehooks_get_extent_hooks_ptr(ehooks)->merge == NULL; +} + +static inline bool +ehooks_guard_will_fail(ehooks_t *ehooks) { + /* + * Before the guard hooks are officially introduced, limit the use to + * the default hooks only. + */ + return !ehooks_are_default(ehooks); +} + +/* + * Some hooks are required to return zeroed memory in certain situations. In + * debug mode, we do some heuristic checks that they did what they were supposed + * to. + * + * This isn't really ehooks-specific (i.e. anyone can check for zeroed memory). + * But incorrect zero information indicates an ehook bug. + */ +static inline void +ehooks_debug_zero_check(void *addr, size_t size) { + assert(((uintptr_t)addr & PAGE_MASK) == 0); + assert((size & PAGE_MASK) == 0); + assert(size > 0); + if (config_debug) { + /* Check the whole first page. */ + size_t *p = (size_t *)addr; + for (size_t i = 0; i < PAGE / sizeof(size_t); i++) { + assert(p[i] == 0); + } + /* + * And 4 spots within. There's a tradeoff here; the larger + * this number, the more likely it is that we'll catch a bug + * where ehooks return a sparsely non-zero range. But + * increasing the number of checks also increases the number of + * page faults in debug mode. FreeBSD does much of their + * day-to-day development work in debug mode, so we don't want + * even the debug builds to be too slow. + */ + const size_t nchecks = 4; + assert(PAGE >= sizeof(size_t) * nchecks); + for (size_t i = 0; i < nchecks; ++i) { + assert(p[i * (size / sizeof(size_t) / nchecks)] == 0); + } + } +} + + +static inline void * +ehooks_alloc(tsdn_t *tsdn, ehooks_t *ehooks, void *new_addr, size_t size, + size_t alignment, bool *zero, bool *commit) { + bool orig_zero = *zero; + void *ret; + extent_hooks_t *extent_hooks = ehooks_get_extent_hooks_ptr(ehooks); + if (extent_hooks == &ehooks_default_extent_hooks) { + ret = ehooks_default_alloc_impl(tsdn, new_addr, size, + alignment, zero, commit, ehooks_ind_get(ehooks)); + } else { + ehooks_pre_reentrancy(tsdn); + ret = extent_hooks->alloc(extent_hooks, new_addr, size, + alignment, zero, commit, ehooks_ind_get(ehooks)); + ehooks_post_reentrancy(tsdn); + } + assert(new_addr == NULL || ret == NULL || new_addr == ret); + assert(!orig_zero || *zero); + if (*zero && ret != NULL) { + ehooks_debug_zero_check(ret, size); + } + return ret; +} + +static inline bool +ehooks_dalloc(tsdn_t *tsdn, ehooks_t *ehooks, void *addr, size_t size, + bool committed) { + extent_hooks_t *extent_hooks = ehooks_get_extent_hooks_ptr(ehooks); + if (extent_hooks == &ehooks_default_extent_hooks) { + return ehooks_default_dalloc_impl(addr, size); + } else if (extent_hooks->dalloc == NULL) { + return true; + } else { + ehooks_pre_reentrancy(tsdn); + bool err = extent_hooks->dalloc(extent_hooks, addr, size, + committed, ehooks_ind_get(ehooks)); + ehooks_post_reentrancy(tsdn); + return err; + } +} + +static inline void +ehooks_destroy(tsdn_t *tsdn, ehooks_t *ehooks, void *addr, size_t size, + bool committed) { + extent_hooks_t *extent_hooks = ehooks_get_extent_hooks_ptr(ehooks); + if (extent_hooks == &ehooks_default_extent_hooks) { + ehooks_default_destroy_impl(addr, size); + } else if (extent_hooks->destroy == NULL) { + /* Do nothing. */ + } else { + ehooks_pre_reentrancy(tsdn); + extent_hooks->destroy(extent_hooks, addr, size, committed, + ehooks_ind_get(ehooks)); + ehooks_post_reentrancy(tsdn); + } +} + +static inline bool +ehooks_commit(tsdn_t *tsdn, ehooks_t *ehooks, void *addr, size_t size, + size_t offset, size_t length) { + extent_hooks_t *extent_hooks = ehooks_get_extent_hooks_ptr(ehooks); + bool err; + if (extent_hooks == &ehooks_default_extent_hooks) { + err = ehooks_default_commit_impl(addr, offset, length); + } else if (extent_hooks->commit == NULL) { + err = true; + } else { + ehooks_pre_reentrancy(tsdn); + err = extent_hooks->commit(extent_hooks, addr, size, + offset, length, ehooks_ind_get(ehooks)); + ehooks_post_reentrancy(tsdn); + } + if (!err) { + ehooks_debug_zero_check(addr, size); + } + return err; +} + +static inline bool +ehooks_decommit(tsdn_t *tsdn, ehooks_t *ehooks, void *addr, size_t size, + size_t offset, size_t length) { + extent_hooks_t *extent_hooks = ehooks_get_extent_hooks_ptr(ehooks); + if (extent_hooks == &ehooks_default_extent_hooks) { + return ehooks_default_decommit_impl(addr, offset, length); + } else if (extent_hooks->decommit == NULL) { + return true; + } else { + ehooks_pre_reentrancy(tsdn); + bool err = extent_hooks->decommit(extent_hooks, addr, size, + offset, length, ehooks_ind_get(ehooks)); + ehooks_post_reentrancy(tsdn); + return err; + } +} + +static inline bool +ehooks_purge_lazy(tsdn_t *tsdn, ehooks_t *ehooks, void *addr, size_t size, + size_t offset, size_t length) { + extent_hooks_t *extent_hooks = ehooks_get_extent_hooks_ptr(ehooks); +#ifdef PAGES_CAN_PURGE_LAZY + if (extent_hooks == &ehooks_default_extent_hooks) { + return ehooks_default_purge_lazy_impl(addr, offset, length); + } +#endif + if (extent_hooks->purge_lazy == NULL) { + return true; + } else { + ehooks_pre_reentrancy(tsdn); + bool err = extent_hooks->purge_lazy(extent_hooks, addr, size, + offset, length, ehooks_ind_get(ehooks)); + ehooks_post_reentrancy(tsdn); + return err; + } +} + +static inline bool +ehooks_purge_forced(tsdn_t *tsdn, ehooks_t *ehooks, void *addr, size_t size, + size_t offset, size_t length) { + extent_hooks_t *extent_hooks = ehooks_get_extent_hooks_ptr(ehooks); + /* + * It would be correct to have a ehooks_debug_zero_check call at the end + * of this function; purge_forced is required to zero. But checking + * would touch the page in question, which may have performance + * consequences (imagine the hooks are using hugepages, with a global + * zero page off). Even in debug mode, it's usually a good idea to + * avoid cases that can dramatically increase memory consumption. + */ +#ifdef PAGES_CAN_PURGE_FORCED + if (extent_hooks == &ehooks_default_extent_hooks) { + return ehooks_default_purge_forced_impl(addr, offset, length); + } +#endif + if (extent_hooks->purge_forced == NULL) { + return true; + } else { + ehooks_pre_reentrancy(tsdn); + bool err = extent_hooks->purge_forced(extent_hooks, addr, size, + offset, length, ehooks_ind_get(ehooks)); + ehooks_post_reentrancy(tsdn); + return err; + } +} + +static inline bool +ehooks_split(tsdn_t *tsdn, ehooks_t *ehooks, void *addr, size_t size, + size_t size_a, size_t size_b, bool committed) { + extent_hooks_t *extent_hooks = ehooks_get_extent_hooks_ptr(ehooks); + if (ehooks_are_default(ehooks)) { + return ehooks_default_split_impl(); + } else if (extent_hooks->split == NULL) { + return true; + } else { + ehooks_pre_reentrancy(tsdn); + bool err = extent_hooks->split(extent_hooks, addr, size, size_a, + size_b, committed, ehooks_ind_get(ehooks)); + ehooks_post_reentrancy(tsdn); + return err; + } +} + +static inline bool +ehooks_merge(tsdn_t *tsdn, ehooks_t *ehooks, void *addr_a, size_t size_a, + void *addr_b, size_t size_b, bool committed) { + extent_hooks_t *extent_hooks = ehooks_get_extent_hooks_ptr(ehooks); + if (extent_hooks == &ehooks_default_extent_hooks) { + return ehooks_default_merge_impl(tsdn, addr_a, addr_b); + } else if (extent_hooks->merge == NULL) { + return true; + } else { + ehooks_pre_reentrancy(tsdn); + bool err = extent_hooks->merge(extent_hooks, addr_a, size_a, + addr_b, size_b, committed, ehooks_ind_get(ehooks)); + ehooks_post_reentrancy(tsdn); + return err; + } +} + +static inline void +ehooks_zero(tsdn_t *tsdn, ehooks_t *ehooks, void *addr, size_t size) { + extent_hooks_t *extent_hooks = ehooks_get_extent_hooks_ptr(ehooks); + if (extent_hooks == &ehooks_default_extent_hooks) { + ehooks_default_zero_impl(addr, size); + } else { + /* + * It would be correct to try using the user-provided purge + * hooks (since they are required to have zeroed the extent if + * they indicate success), but we don't necessarily know their + * cost. We'll be conservative and use memset. + */ + memset(addr, 0, size); + } +} + +static inline bool +ehooks_guard(tsdn_t *tsdn, ehooks_t *ehooks, void *guard1, void *guard2) { + bool err; + extent_hooks_t *extent_hooks = ehooks_get_extent_hooks_ptr(ehooks); + + if (extent_hooks == &ehooks_default_extent_hooks) { + ehooks_default_guard_impl(guard1, guard2); + err = false; + } else { + err = true; + } + + return err; +} + +static inline bool +ehooks_unguard(tsdn_t *tsdn, ehooks_t *ehooks, void *guard1, void *guard2) { + bool err; + extent_hooks_t *extent_hooks = ehooks_get_extent_hooks_ptr(ehooks); + + if (extent_hooks == &ehooks_default_extent_hooks) { + ehooks_default_unguard_impl(guard1, guard2); + err = false; + } else { + err = true; + } + + return err; +} + +#endif /* JEMALLOC_INTERNAL_EHOOKS_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/emap.h b/contrib/jemalloc/include/jemalloc/internal/emap.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/emap.h @@ -0,0 +1,357 @@ +#ifndef JEMALLOC_INTERNAL_EMAP_H +#define JEMALLOC_INTERNAL_EMAP_H + +#include "jemalloc/internal/base.h" +#include "jemalloc/internal/rtree.h" + +/* + * Note: Ends without at semicolon, so that + * EMAP_DECLARE_RTREE_CTX; + * in uses will avoid empty-statement warnings. + */ +#define EMAP_DECLARE_RTREE_CTX \ + rtree_ctx_t rtree_ctx_fallback; \ + rtree_ctx_t *rtree_ctx = tsdn_rtree_ctx(tsdn, &rtree_ctx_fallback) + +typedef struct emap_s emap_t; +struct emap_s { + rtree_t rtree; +}; + +/* Used to pass rtree lookup context down the path. */ +typedef struct emap_alloc_ctx_t emap_alloc_ctx_t; +struct emap_alloc_ctx_t { + szind_t szind; + bool slab; +}; + +typedef struct emap_full_alloc_ctx_s emap_full_alloc_ctx_t; +struct emap_full_alloc_ctx_s { + szind_t szind; + bool slab; + edata_t *edata; +}; + +bool emap_init(emap_t *emap, base_t *base, bool zeroed); + +void emap_remap(tsdn_t *tsdn, emap_t *emap, edata_t *edata, szind_t szind, + bool slab); + +void emap_update_edata_state(tsdn_t *tsdn, emap_t *emap, edata_t *edata, + extent_state_t state); + +/* + * The two acquire functions below allow accessing neighbor edatas, if it's safe + * and valid to do so (i.e. from the same arena, of the same state, etc.). This + * is necessary because the ecache locks are state based, and only protect + * edatas with the same state. Therefore the neighbor edata's state needs to be + * verified first, before chasing the edata pointer. The returned edata will be + * in an acquired state, meaning other threads will be prevented from accessing + * it, even if technically the edata can still be discovered from the rtree. + * + * This means, at any moment when holding pointers to edata, either one of the + * state based locks is held (and the edatas are all of the protected state), or + * the edatas are in an acquired state (e.g. in active or merging state). The + * acquire operation itself (changing the edata to an acquired state) is done + * under the state locks. + */ +edata_t *emap_try_acquire_edata_neighbor(tsdn_t *tsdn, emap_t *emap, + edata_t *edata, extent_pai_t pai, extent_state_t expected_state, + bool forward); +edata_t *emap_try_acquire_edata_neighbor_expand(tsdn_t *tsdn, emap_t *emap, + edata_t *edata, extent_pai_t pai, extent_state_t expected_state); +void emap_release_edata(tsdn_t *tsdn, emap_t *emap, edata_t *edata, + extent_state_t new_state); + +/* + * Associate the given edata with its beginning and end address, setting the + * szind and slab info appropriately. + * Returns true on error (i.e. resource exhaustion). + */ +bool emap_register_boundary(tsdn_t *tsdn, emap_t *emap, edata_t *edata, + szind_t szind, bool slab); + +/* + * Does the same thing, but with the interior of the range, for slab + * allocations. + * + * You might wonder why we don't just have a single emap_register function that + * does both depending on the value of 'slab'. The answer is twofold: + * - As a practical matter, in places like the extract->split->commit pathway, + * we defer the interior operation until we're sure that the commit won't fail + * (but we have to register the split boundaries there). + * - In general, we're trying to move to a world where the page-specific + * allocator doesn't know as much about how the pages it allocates will be + * used, and passing a 'slab' parameter everywhere makes that more + * complicated. + * + * Unlike the boundary version, this function can't fail; this is because slabs + * can't get big enough to touch a new page that neither of the boundaries + * touched, so no allocation is necessary to fill the interior once the boundary + * has been touched. + */ +void emap_register_interior(tsdn_t *tsdn, emap_t *emap, edata_t *edata, + szind_t szind); + +void emap_deregister_boundary(tsdn_t *tsdn, emap_t *emap, edata_t *edata); +void emap_deregister_interior(tsdn_t *tsdn, emap_t *emap, edata_t *edata); + +typedef struct emap_prepare_s emap_prepare_t; +struct emap_prepare_s { + rtree_leaf_elm_t *lead_elm_a; + rtree_leaf_elm_t *lead_elm_b; + rtree_leaf_elm_t *trail_elm_a; + rtree_leaf_elm_t *trail_elm_b; +}; + +/** + * These functions the emap metadata management for merging, splitting, and + * reusing extents. In particular, they set the boundary mappings from + * addresses to edatas. If the result is going to be used as a slab, you + * still need to call emap_register_interior on it, though. + * + * Remap simply changes the szind and slab status of an extent's boundary + * mappings. If the extent is not a slab, it doesn't bother with updating the + * end mapping (since lookups only occur in the interior of an extent for + * slabs). Since the szind and slab status only make sense for active extents, + * this should only be called while activating or deactivating an extent. + * + * Split and merge have a "prepare" and a "commit" portion. The prepare portion + * does the operations that can be done without exclusive access to the extent + * in question, while the commit variant requires exclusive access to maintain + * the emap invariants. The only function that can fail is emap_split_prepare, + * and it returns true on failure (at which point the caller shouldn't commit). + * + * In all cases, "lead" refers to the lower-addressed extent, and trail to the + * higher-addressed one. It's the caller's responsibility to set the edata + * state appropriately. + */ +bool emap_split_prepare(tsdn_t *tsdn, emap_t *emap, emap_prepare_t *prepare, + edata_t *edata, size_t size_a, edata_t *trail, size_t size_b); +void emap_split_commit(tsdn_t *tsdn, emap_t *emap, emap_prepare_t *prepare, + edata_t *lead, size_t size_a, edata_t *trail, size_t size_b); +void emap_merge_prepare(tsdn_t *tsdn, emap_t *emap, emap_prepare_t *prepare, + edata_t *lead, edata_t *trail); +void emap_merge_commit(tsdn_t *tsdn, emap_t *emap, emap_prepare_t *prepare, + edata_t *lead, edata_t *trail); + +/* Assert that the emap's view of the given edata matches the edata's view. */ +void emap_do_assert_mapped(tsdn_t *tsdn, emap_t *emap, edata_t *edata); +static inline void +emap_assert_mapped(tsdn_t *tsdn, emap_t *emap, edata_t *edata) { + if (config_debug) { + emap_do_assert_mapped(tsdn, emap, edata); + } +} + +/* Assert that the given edata isn't in the map. */ +void emap_do_assert_not_mapped(tsdn_t *tsdn, emap_t *emap, edata_t *edata); +static inline void +emap_assert_not_mapped(tsdn_t *tsdn, emap_t *emap, edata_t *edata) { + if (config_debug) { + emap_do_assert_not_mapped(tsdn, emap, edata); + } +} + +JEMALLOC_ALWAYS_INLINE bool +emap_edata_in_transition(tsdn_t *tsdn, emap_t *emap, edata_t *edata) { + assert(config_debug); + emap_assert_mapped(tsdn, emap, edata); + + EMAP_DECLARE_RTREE_CTX; + rtree_contents_t contents = rtree_read(tsdn, &emap->rtree, rtree_ctx, + (uintptr_t)edata_base_get(edata)); + + return edata_state_in_transition(contents.metadata.state); +} + +JEMALLOC_ALWAYS_INLINE bool +emap_edata_is_acquired(tsdn_t *tsdn, emap_t *emap, edata_t *edata) { + if (!config_debug) { + /* For assertions only. */ + return false; + } + + /* + * The edata is considered acquired if no other threads will attempt to + * read / write any fields from it. This includes a few cases: + * + * 1) edata not hooked into emap yet -- This implies the edata just got + * allocated or initialized. + * + * 2) in an active or transition state -- In both cases, the edata can + * be discovered from the emap, however the state tracked in the rtree + * will prevent other threads from accessing the actual edata. + */ + EMAP_DECLARE_RTREE_CTX; + rtree_leaf_elm_t *elm = rtree_leaf_elm_lookup(tsdn, &emap->rtree, + rtree_ctx, (uintptr_t)edata_base_get(edata), /* dependent */ true, + /* init_missing */ false); + if (elm == NULL) { + return true; + } + rtree_contents_t contents = rtree_leaf_elm_read(tsdn, &emap->rtree, elm, + /* dependent */ true); + if (contents.edata == NULL || + contents.metadata.state == extent_state_active || + edata_state_in_transition(contents.metadata.state)) { + return true; + } + + return false; +} + +JEMALLOC_ALWAYS_INLINE void +extent_assert_can_coalesce(const edata_t *inner, const edata_t *outer) { + assert(edata_arena_ind_get(inner) == edata_arena_ind_get(outer)); + assert(edata_pai_get(inner) == edata_pai_get(outer)); + assert(edata_committed_get(inner) == edata_committed_get(outer)); + assert(edata_state_get(inner) == extent_state_active); + assert(edata_state_get(outer) == extent_state_merging); + assert(!edata_guarded_get(inner) && !edata_guarded_get(outer)); + assert(edata_base_get(inner) == edata_past_get(outer) || + edata_base_get(outer) == edata_past_get(inner)); +} + +JEMALLOC_ALWAYS_INLINE void +extent_assert_can_expand(const edata_t *original, const edata_t *expand) { + assert(edata_arena_ind_get(original) == edata_arena_ind_get(expand)); + assert(edata_pai_get(original) == edata_pai_get(expand)); + assert(edata_state_get(original) == extent_state_active); + assert(edata_state_get(expand) == extent_state_merging); + assert(edata_past_get(original) == edata_base_get(expand)); +} + +JEMALLOC_ALWAYS_INLINE edata_t * +emap_edata_lookup(tsdn_t *tsdn, emap_t *emap, const void *ptr) { + EMAP_DECLARE_RTREE_CTX; + + return rtree_read(tsdn, &emap->rtree, rtree_ctx, (uintptr_t)ptr).edata; +} + +/* Fills in alloc_ctx with the info in the map. */ +JEMALLOC_ALWAYS_INLINE void +emap_alloc_ctx_lookup(tsdn_t *tsdn, emap_t *emap, const void *ptr, + emap_alloc_ctx_t *alloc_ctx) { + EMAP_DECLARE_RTREE_CTX; + + rtree_metadata_t metadata = rtree_metadata_read(tsdn, &emap->rtree, + rtree_ctx, (uintptr_t)ptr); + alloc_ctx->szind = metadata.szind; + alloc_ctx->slab = metadata.slab; +} + +/* The pointer must be mapped. */ +JEMALLOC_ALWAYS_INLINE void +emap_full_alloc_ctx_lookup(tsdn_t *tsdn, emap_t *emap, const void *ptr, + emap_full_alloc_ctx_t *full_alloc_ctx) { + EMAP_DECLARE_RTREE_CTX; + + rtree_contents_t contents = rtree_read(tsdn, &emap->rtree, rtree_ctx, + (uintptr_t)ptr); + full_alloc_ctx->edata = contents.edata; + full_alloc_ctx->szind = contents.metadata.szind; + full_alloc_ctx->slab = contents.metadata.slab; +} + +/* + * The pointer is allowed to not be mapped. + * + * Returns true when the pointer is not present. + */ +JEMALLOC_ALWAYS_INLINE bool +emap_full_alloc_ctx_try_lookup(tsdn_t *tsdn, emap_t *emap, const void *ptr, + emap_full_alloc_ctx_t *full_alloc_ctx) { + EMAP_DECLARE_RTREE_CTX; + + rtree_contents_t contents; + bool err = rtree_read_independent(tsdn, &emap->rtree, rtree_ctx, + (uintptr_t)ptr, &contents); + if (err) { + return true; + } + full_alloc_ctx->edata = contents.edata; + full_alloc_ctx->szind = contents.metadata.szind; + full_alloc_ctx->slab = contents.metadata.slab; + return false; +} + +/* + * Only used on the fastpath of free. Returns true when cannot be fulfilled by + * fast path, e.g. when the metadata key is not cached. + */ +JEMALLOC_ALWAYS_INLINE bool +emap_alloc_ctx_try_lookup_fast(tsd_t *tsd, emap_t *emap, const void *ptr, + emap_alloc_ctx_t *alloc_ctx) { + /* Use the unsafe getter since this may gets called during exit. */ + rtree_ctx_t *rtree_ctx = tsd_rtree_ctxp_get_unsafe(tsd); + + rtree_metadata_t metadata; + bool err = rtree_metadata_try_read_fast(tsd_tsdn(tsd), &emap->rtree, + rtree_ctx, (uintptr_t)ptr, &metadata); + if (err) { + return true; + } + alloc_ctx->szind = metadata.szind; + alloc_ctx->slab = metadata.slab; + return false; +} + +/* + * We want to do batch lookups out of the cache bins, which use + * cache_bin_ptr_array_get to access the i'th element of the bin (since they + * invert usual ordering in deciding what to flush). This lets the emap avoid + * caring about its caller's ordering. + */ +typedef const void *(*emap_ptr_getter)(void *ctx, size_t ind); +/* + * This allows size-checking assertions, which we can only do while we're in the + * process of edata lookups. + */ +typedef void (*emap_metadata_visitor)(void *ctx, emap_full_alloc_ctx_t *alloc_ctx); + +typedef union emap_batch_lookup_result_u emap_batch_lookup_result_t; +union emap_batch_lookup_result_u { + edata_t *edata; + rtree_leaf_elm_t *rtree_leaf; +}; + +JEMALLOC_ALWAYS_INLINE void +emap_edata_lookup_batch(tsd_t *tsd, emap_t *emap, size_t nptrs, + emap_ptr_getter ptr_getter, void *ptr_getter_ctx, + emap_metadata_visitor metadata_visitor, void *metadata_visitor_ctx, + emap_batch_lookup_result_t *result) { + /* Avoids null-checking tsdn in the loop below. */ + util_assume(tsd != NULL); + rtree_ctx_t *rtree_ctx = tsd_rtree_ctxp_get(tsd); + + for (size_t i = 0; i < nptrs; i++) { + const void *ptr = ptr_getter(ptr_getter_ctx, i); + /* + * Reuse the edatas array as a temp buffer, lying a little about + * the types. + */ + result[i].rtree_leaf = rtree_leaf_elm_lookup(tsd_tsdn(tsd), + &emap->rtree, rtree_ctx, (uintptr_t)ptr, + /* dependent */ true, /* init_missing */ false); + } + + for (size_t i = 0; i < nptrs; i++) { + rtree_leaf_elm_t *elm = result[i].rtree_leaf; + rtree_contents_t contents = rtree_leaf_elm_read(tsd_tsdn(tsd), + &emap->rtree, elm, /* dependent */ true); + result[i].edata = contents.edata; + emap_full_alloc_ctx_t alloc_ctx; + /* + * Not all these fields are read in practice by the metadata + * visitor. But the compiler can easily optimize away the ones + * that aren't, so no sense in being incomplete. + */ + alloc_ctx.szind = contents.metadata.szind; + alloc_ctx.slab = contents.metadata.slab; + alloc_ctx.edata = contents.edata; + metadata_visitor(metadata_visitor_ctx, &alloc_ctx); + } +} + +#endif /* JEMALLOC_INTERNAL_EMAP_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/emitter.h b/contrib/jemalloc/include/jemalloc/internal/emitter.h --- a/contrib/jemalloc/include/jemalloc/internal/emitter.h +++ b/contrib/jemalloc/include/jemalloc/internal/emitter.h @@ -6,6 +6,7 @@ typedef enum emitter_output_e emitter_output_t; enum emitter_output_e { emitter_output_json, + emitter_output_json_compact, emitter_output_table }; @@ -21,6 +22,7 @@ enum emitter_type_e { emitter_type_bool, emitter_type_int, + emitter_type_int64, emitter_type_unsigned, emitter_type_uint32, emitter_type_uint64, @@ -66,7 +68,7 @@ struct emitter_s { emitter_output_t output; /* The output information. */ - void (*write_cb)(void *, const char *); + write_cb_t *write_cb; void *cbopaque; int nesting_depth; /* True if we've already emitted a value at the given depth. */ @@ -75,6 +77,12 @@ bool emitted_key; }; +static inline bool +emitter_outputs_json(emitter_t *emitter) { + return emitter->output == emitter_output_json || + emitter->output == emitter_output_json_compact; +} + /* Internal convenience function. Write to the emitter the given string. */ JEMALLOC_FORMAT_PRINTF(2, 3) static inline void @@ -135,13 +143,16 @@ switch (value_type) { case emitter_type_bool: - emitter_printf(emitter, + emitter_printf(emitter, emitter_gen_fmt(fmt, FMT_SIZE, "%s", justify, width), *(const bool *)value ? "true" : "false"); break; case emitter_type_int: EMIT_SIMPLE(int, "%d") break; + case emitter_type_int64: + EMIT_SIMPLE(int64_t, "%" FMTd64) + break; case emitter_type_unsigned: EMIT_SIMPLE(unsigned, "%u") break; @@ -159,7 +170,7 @@ * anywhere near the fmt size. */ assert(str_written < BUF_SIZE); - emitter_printf(emitter, + emitter_printf(emitter, emitter_gen_fmt(fmt, FMT_SIZE, "%s", justify, width), buf); break; case emitter_type_uint32: @@ -196,6 +207,7 @@ emitter_indent(emitter_t *emitter) { int amount = emitter->nesting_depth; const char *indent_str; + assert(emitter->output != emitter_output_json_compact); if (emitter->output == emitter_output_json) { indent_str = "\t"; } else { @@ -209,12 +221,18 @@ static inline void emitter_json_key_prefix(emitter_t *emitter) { + assert(emitter_outputs_json(emitter)); if (emitter->emitted_key) { emitter->emitted_key = false; return; } - emitter_printf(emitter, "%s\n", emitter->item_at_depth ? "," : ""); - emitter_indent(emitter); + if (emitter->item_at_depth) { + emitter_printf(emitter, ","); + } + if (emitter->output != emitter_output_json_compact) { + emitter_printf(emitter, "\n"); + emitter_indent(emitter); + } } /******************************************************************************/ @@ -222,27 +240,28 @@ static inline void emitter_init(emitter_t *emitter, emitter_output_t emitter_output, - void (*write_cb)(void *, const char *), void *cbopaque) { + write_cb_t *write_cb, void *cbopaque) { emitter->output = emitter_output; emitter->write_cb = write_cb; emitter->cbopaque = cbopaque; emitter->item_at_depth = false; - emitter->emitted_key = false; + emitter->emitted_key = false; emitter->nesting_depth = 0; } /******************************************************************************/ /* JSON public API. */ -/* +/* * Emits a key (e.g. as appears in an object). The next json entity emitted will * be the corresponding value. */ static inline void emitter_json_key(emitter_t *emitter, const char *json_key) { - if (emitter->output == emitter_output_json) { + if (emitter_outputs_json(emitter)) { emitter_json_key_prefix(emitter); - emitter_printf(emitter, "\"%s\": ", json_key); + emitter_printf(emitter, "\"%s\":%s", json_key, + emitter->output == emitter_output_json_compact ? "" : " "); emitter->emitted_key = true; } } @@ -250,7 +269,7 @@ static inline void emitter_json_value(emitter_t *emitter, emitter_type_t value_type, const void *value) { - if (emitter->output == emitter_output_json) { + if (emitter_outputs_json(emitter)) { emitter_json_key_prefix(emitter); emitter_print_value(emitter, emitter_justify_none, -1, value_type, value); @@ -268,7 +287,7 @@ static inline void emitter_json_array_begin(emitter_t *emitter) { - if (emitter->output == emitter_output_json) { + if (emitter_outputs_json(emitter)) { emitter_json_key_prefix(emitter); emitter_printf(emitter, "["); emitter_nest_inc(emitter); @@ -284,18 +303,20 @@ static inline void emitter_json_array_end(emitter_t *emitter) { - if (emitter->output == emitter_output_json) { + if (emitter_outputs_json(emitter)) { assert(emitter->nesting_depth > 0); emitter_nest_dec(emitter); - emitter_printf(emitter, "\n"); - emitter_indent(emitter); + if (emitter->output != emitter_output_json_compact) { + emitter_printf(emitter, "\n"); + emitter_indent(emitter); + } emitter_printf(emitter, "]"); } } static inline void emitter_json_object_begin(emitter_t *emitter) { - if (emitter->output == emitter_output_json) { + if (emitter_outputs_json(emitter)) { emitter_json_key_prefix(emitter); emitter_printf(emitter, "{"); emitter_nest_inc(emitter); @@ -311,11 +332,13 @@ static inline void emitter_json_object_end(emitter_t *emitter) { - if (emitter->output == emitter_output_json) { + if (emitter_outputs_json(emitter)) { assert(emitter->nesting_depth > 0); emitter_nest_dec(emitter); - emitter_printf(emitter, "\n"); - emitter_indent(emitter); + if (emitter->output != emitter_output_json_compact) { + emitter_printf(emitter, "\n"); + emitter_indent(emitter); + } emitter_printf(emitter, "}"); } } @@ -420,7 +443,7 @@ emitter_type_t value_type, const void *value, const char *table_note_key, emitter_type_t table_note_value_type, const void *table_note_value) { - if (emitter->output == emitter_output_json) { + if (emitter_outputs_json(emitter)) { emitter_json_key(emitter, json_key); emitter_json_value(emitter, value_type, value); } else { @@ -440,7 +463,7 @@ static inline void emitter_dict_begin(emitter_t *emitter, const char *json_key, const char *table_header) { - if (emitter->output == emitter_output_json) { + if (emitter_outputs_json(emitter)) { emitter_json_key(emitter, json_key); emitter_json_object_begin(emitter); } else { @@ -450,7 +473,7 @@ static inline void emitter_dict_end(emitter_t *emitter) { - if (emitter->output == emitter_output_json) { + if (emitter_outputs_json(emitter)) { emitter_json_object_end(emitter); } else { emitter_table_dict_end(emitter); @@ -459,7 +482,7 @@ static inline void emitter_begin(emitter_t *emitter) { - if (emitter->output == emitter_output_json) { + if (emitter_outputs_json(emitter)) { assert(emitter->nesting_depth == 0); emitter_printf(emitter, "{"); emitter_nest_inc(emitter); @@ -476,10 +499,11 @@ static inline void emitter_end(emitter_t *emitter) { - if (emitter->output == emitter_output_json) { + if (emitter_outputs_json(emitter)) { assert(emitter->nesting_depth == 1); emitter_nest_dec(emitter); - emitter_printf(emitter, "\n}\n"); + emitter_printf(emitter, "%s", emitter->output == + emitter_output_json_compact ? "}" : "\n}\n"); } } diff --git a/contrib/jemalloc/include/jemalloc/internal/eset.h b/contrib/jemalloc/include/jemalloc/internal/eset.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/eset.h @@ -0,0 +1,77 @@ +#ifndef JEMALLOC_INTERNAL_ESET_H +#define JEMALLOC_INTERNAL_ESET_H + +#include "jemalloc/internal/atomic.h" +#include "jemalloc/internal/fb.h" +#include "jemalloc/internal/edata.h" +#include "jemalloc/internal/mutex.h" + +/* + * An eset ("extent set") is a quantized collection of extents, with built-in + * LRU queue. + * + * This class is not thread-safe; synchronization must be done externally if + * there are mutating operations. One exception is the stats counters, which + * may be read without any locking. + */ + +typedef struct eset_bin_s eset_bin_t; +struct eset_bin_s { + edata_heap_t heap; + /* + * We do first-fit across multiple size classes. If we compared against + * the min element in each heap directly, we'd take a cache miss per + * extent we looked at. If we co-locate the edata summaries, we only + * take a miss on the edata we're actually going to return (which is + * inevitable anyways). + */ + edata_cmp_summary_t heap_min; +}; + +typedef struct eset_bin_stats_s eset_bin_stats_t; +struct eset_bin_stats_s { + atomic_zu_t nextents; + atomic_zu_t nbytes; +}; + +typedef struct eset_s eset_t; +struct eset_s { + /* Bitmap for which set bits correspond to non-empty heaps. */ + fb_group_t bitmap[FB_NGROUPS(SC_NPSIZES + 1)]; + + /* Quantized per size class heaps of extents. */ + eset_bin_t bins[SC_NPSIZES + 1]; + + eset_bin_stats_t bin_stats[SC_NPSIZES + 1]; + + /* LRU of all extents in heaps. */ + edata_list_inactive_t lru; + + /* Page sum for all extents in heaps. */ + atomic_zu_t npages; + + /* + * A duplication of the data in the containing ecache. We use this only + * for assertions on the states of the passed-in extents. + */ + extent_state_t state; +}; + +void eset_init(eset_t *eset, extent_state_t state); + +size_t eset_npages_get(eset_t *eset); +/* Get the number of extents in the given page size index. */ +size_t eset_nextents_get(eset_t *eset, pszind_t ind); +/* Get the sum total bytes of the extents in the given page size index. */ +size_t eset_nbytes_get(eset_t *eset, pszind_t ind); + +void eset_insert(eset_t *eset, edata_t *edata); +void eset_remove(eset_t *eset, edata_t *edata); +/* + * Select an extent from this eset of the given size and alignment. Returns + * null if no such item could be found. + */ +edata_t *eset_fit(eset_t *eset, size_t esize, size_t alignment, bool exact_only, + unsigned lg_max_fit); + +#endif /* JEMALLOC_INTERNAL_ESET_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/exp_grow.h b/contrib/jemalloc/include/jemalloc/internal/exp_grow.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/exp_grow.h @@ -0,0 +1,50 @@ +#ifndef JEMALLOC_INTERNAL_EXP_GROW_H +#define JEMALLOC_INTERNAL_EXP_GROW_H + +typedef struct exp_grow_s exp_grow_t; +struct exp_grow_s { + /* + * Next extent size class in a growing series to use when satisfying a + * request via the extent hooks (only if opt_retain). This limits the + * number of disjoint virtual memory ranges so that extent merging can + * be effective even if multiple arenas' extent allocation requests are + * highly interleaved. + * + * retain_grow_limit is the max allowed size ind to expand (unless the + * required size is greater). Default is no limit, and controlled + * through mallctl only. + */ + pszind_t next; + pszind_t limit; +}; + +static inline bool +exp_grow_size_prepare(exp_grow_t *exp_grow, size_t alloc_size_min, + size_t *r_alloc_size, pszind_t *r_skip) { + *r_skip = 0; + *r_alloc_size = sz_pind2sz(exp_grow->next + *r_skip); + while (*r_alloc_size < alloc_size_min) { + (*r_skip)++; + if (exp_grow->next + *r_skip >= + sz_psz2ind(SC_LARGE_MAXCLASS)) { + /* Outside legal range. */ + return true; + } + *r_alloc_size = sz_pind2sz(exp_grow->next + *r_skip); + } + return false; +} + +static inline void +exp_grow_size_commit(exp_grow_t *exp_grow, pszind_t skip) { + if (exp_grow->next + skip + 1 <= exp_grow->limit) { + exp_grow->next += skip + 1; + } else { + exp_grow->next = exp_grow->limit; + } + +} + +void exp_grow_init(exp_grow_t *exp_grow); + +#endif /* JEMALLOC_INTERNAL_EXP_GROW_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/extent.h b/contrib/jemalloc/include/jemalloc/internal/extent.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/extent.h @@ -0,0 +1,137 @@ +#ifndef JEMALLOC_INTERNAL_EXTENT_H +#define JEMALLOC_INTERNAL_EXTENT_H + +#include "jemalloc/internal/ecache.h" +#include "jemalloc/internal/ehooks.h" +#include "jemalloc/internal/ph.h" +#include "jemalloc/internal/rtree.h" + +/* + * This module contains the page-level allocator. It chooses the addresses that + * allocations requested by other modules will inhabit, and updates the global + * metadata to reflect allocation/deallocation/purging decisions. + */ + +/* + * When reuse (and split) an active extent, (1U << opt_lg_extent_max_active_fit) + * is the max ratio between the size of the active extent and the new extent. + */ +#define LG_EXTENT_MAX_ACTIVE_FIT_DEFAULT 6 +extern size_t opt_lg_extent_max_active_fit; + +edata_t *ecache_alloc(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, + ecache_t *ecache, edata_t *expand_edata, size_t size, size_t alignment, + bool zero, bool guarded); +edata_t *ecache_alloc_grow(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, + ecache_t *ecache, edata_t *expand_edata, size_t size, size_t alignment, + bool zero, bool guarded); +void ecache_dalloc(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, + ecache_t *ecache, edata_t *edata); +edata_t *ecache_evict(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, + ecache_t *ecache, size_t npages_min); + +void extent_gdump_add(tsdn_t *tsdn, const edata_t *edata); +void extent_record(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, ecache_t *ecache, + edata_t *edata); +void extent_dalloc_gap(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, + edata_t *edata); +edata_t *extent_alloc_wrapper(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, + void *new_addr, size_t size, size_t alignment, bool zero, bool *commit, + bool growing_retained); +void extent_dalloc_wrapper(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, + edata_t *edata); +void extent_destroy_wrapper(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, + edata_t *edata); +bool extent_commit_wrapper(tsdn_t *tsdn, ehooks_t *ehooks, edata_t *edata, + size_t offset, size_t length); +bool extent_decommit_wrapper(tsdn_t *tsdn, ehooks_t *ehooks, edata_t *edata, + size_t offset, size_t length); +bool extent_purge_lazy_wrapper(tsdn_t *tsdn, ehooks_t *ehooks, edata_t *edata, + size_t offset, size_t length); +bool extent_purge_forced_wrapper(tsdn_t *tsdn, ehooks_t *ehooks, edata_t *edata, + size_t offset, size_t length); +edata_t *extent_split_wrapper(tsdn_t *tsdn, pac_t *pac, + ehooks_t *ehooks, edata_t *edata, size_t size_a, size_t size_b, + bool holding_core_locks); +bool extent_merge_wrapper(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, + edata_t *a, edata_t *b); +bool extent_commit_zero(tsdn_t *tsdn, ehooks_t *ehooks, edata_t *edata, + bool commit, bool zero, bool growing_retained); +size_t extent_sn_next(pac_t *pac); +bool extent_boot(void); + +JEMALLOC_ALWAYS_INLINE bool +extent_neighbor_head_state_mergeable(bool edata_is_head, + bool neighbor_is_head, bool forward) { + /* + * Head states checking: disallow merging if the higher addr extent is a + * head extent. This helps preserve first-fit, and more importantly + * makes sure no merge across arenas. + */ + if (forward) { + if (neighbor_is_head) { + return false; + } + } else { + if (edata_is_head) { + return false; + } + } + return true; +} + +JEMALLOC_ALWAYS_INLINE bool +extent_can_acquire_neighbor(edata_t *edata, rtree_contents_t contents, + extent_pai_t pai, extent_state_t expected_state, bool forward, + bool expanding) { + edata_t *neighbor = contents.edata; + if (neighbor == NULL) { + return false; + } + /* It's not safe to access *neighbor yet; must verify states first. */ + bool neighbor_is_head = contents.metadata.is_head; + if (!extent_neighbor_head_state_mergeable(edata_is_head_get(edata), + neighbor_is_head, forward)) { + return false; + } + extent_state_t neighbor_state = contents.metadata.state; + if (pai == EXTENT_PAI_PAC) { + if (neighbor_state != expected_state) { + return false; + } + /* From this point, it's safe to access *neighbor. */ + if (!expanding && (edata_committed_get(edata) != + edata_committed_get(neighbor))) { + /* + * Some platforms (e.g. Windows) require an explicit + * commit step (and writing to uncommitted memory is not + * allowed). + */ + return false; + } + } else { + if (neighbor_state == extent_state_active) { + return false; + } + /* From this point, it's safe to access *neighbor. */ + } + + assert(edata_pai_get(edata) == pai); + if (edata_pai_get(neighbor) != pai) { + return false; + } + if (opt_retain) { + assert(edata_arena_ind_get(edata) == + edata_arena_ind_get(neighbor)); + } else { + if (edata_arena_ind_get(edata) != + edata_arena_ind_get(neighbor)) { + return false; + } + } + assert(!edata_guarded_get(edata) && !edata_guarded_get(neighbor)); + + return true; +} + +#endif /* JEMALLOC_INTERNAL_EXTENT_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/extent_externs.h b/contrib/jemalloc/include/jemalloc/internal/extent_externs.h deleted file mode 100644 --- a/contrib/jemalloc/include/jemalloc/internal/extent_externs.h +++ /dev/null @@ -1,83 +0,0 @@ -#ifndef JEMALLOC_INTERNAL_EXTENT_EXTERNS_H -#define JEMALLOC_INTERNAL_EXTENT_EXTERNS_H - -#include "jemalloc/internal/mutex.h" -#include "jemalloc/internal/mutex_pool.h" -#include "jemalloc/internal/ph.h" -#include "jemalloc/internal/rtree.h" - -extern size_t opt_lg_extent_max_active_fit; - -extern rtree_t extents_rtree; -extern const extent_hooks_t extent_hooks_default; -extern mutex_pool_t extent_mutex_pool; - -extent_t *extent_alloc(tsdn_t *tsdn, arena_t *arena); -void extent_dalloc(tsdn_t *tsdn, arena_t *arena, extent_t *extent); - -extent_hooks_t *extent_hooks_get(arena_t *arena); -extent_hooks_t *extent_hooks_set(tsd_t *tsd, arena_t *arena, - extent_hooks_t *extent_hooks); - -#ifdef JEMALLOC_JET -size_t extent_size_quantize_floor(size_t size); -size_t extent_size_quantize_ceil(size_t size); -#endif - -ph_proto(, extent_avail_, extent_tree_t, extent_t) -ph_proto(, extent_heap_, extent_heap_t, extent_t) - -bool extents_init(tsdn_t *tsdn, extents_t *extents, extent_state_t state, - bool delay_coalesce); -extent_state_t extents_state_get(const extents_t *extents); -size_t extents_npages_get(extents_t *extents); -/* Get the number of extents in the given page size index. */ -size_t extents_nextents_get(extents_t *extents, pszind_t ind); -/* Get the sum total bytes of the extents in the given page size index. */ -size_t extents_nbytes_get(extents_t *extents, pszind_t ind); -extent_t *extents_alloc(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, extents_t *extents, void *new_addr, - size_t size, size_t pad, size_t alignment, bool slab, szind_t szind, - bool *zero, bool *commit); -void extents_dalloc(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, extents_t *extents, extent_t *extent); -extent_t *extents_evict(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, extents_t *extents, size_t npages_min); -void extents_prefork(tsdn_t *tsdn, extents_t *extents); -void extents_postfork_parent(tsdn_t *tsdn, extents_t *extents); -void extents_postfork_child(tsdn_t *tsdn, extents_t *extents); -extent_t *extent_alloc_wrapper(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, void *new_addr, size_t size, size_t pad, - size_t alignment, bool slab, szind_t szind, bool *zero, bool *commit); -void extent_dalloc_gap(tsdn_t *tsdn, arena_t *arena, extent_t *extent); -void extent_dalloc_wrapper(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, extent_t *extent); -void extent_destroy_wrapper(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, extent_t *extent); -bool extent_commit_wrapper(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, extent_t *extent, size_t offset, - size_t length); -bool extent_decommit_wrapper(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, extent_t *extent, size_t offset, - size_t length); -bool extent_purge_lazy_wrapper(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, extent_t *extent, size_t offset, - size_t length); -bool extent_purge_forced_wrapper(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, extent_t *extent, size_t offset, - size_t length); -extent_t *extent_split_wrapper(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, extent_t *extent, size_t size_a, - szind_t szind_a, bool slab_a, size_t size_b, szind_t szind_b, bool slab_b); -bool extent_merge_wrapper(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, extent_t *a, extent_t *b); - -bool extent_boot(void); - -void extent_util_stats_get(tsdn_t *tsdn, const void *ptr, - size_t *nfree, size_t *nregs, size_t *size); -void extent_util_stats_verbose_get(tsdn_t *tsdn, const void *ptr, - size_t *nfree, size_t *nregs, size_t *size, - size_t *bin_nfree, size_t *bin_nregs, void **slabcur_addr); - -#endif /* JEMALLOC_INTERNAL_EXTENT_EXTERNS_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/extent_inlines.h b/contrib/jemalloc/include/jemalloc/internal/extent_inlines.h deleted file mode 100644 --- a/contrib/jemalloc/include/jemalloc/internal/extent_inlines.h +++ /dev/null @@ -1,501 +0,0 @@ -#ifndef JEMALLOC_INTERNAL_EXTENT_INLINES_H -#define JEMALLOC_INTERNAL_EXTENT_INLINES_H - -#include "jemalloc/internal/mutex.h" -#include "jemalloc/internal/mutex_pool.h" -#include "jemalloc/internal/pages.h" -#include "jemalloc/internal/prng.h" -#include "jemalloc/internal/ql.h" -#include "jemalloc/internal/sc.h" -#include "jemalloc/internal/sz.h" - -static inline void -extent_lock(tsdn_t *tsdn, extent_t *extent) { - assert(extent != NULL); - mutex_pool_lock(tsdn, &extent_mutex_pool, (uintptr_t)extent); -} - -static inline void -extent_unlock(tsdn_t *tsdn, extent_t *extent) { - assert(extent != NULL); - mutex_pool_unlock(tsdn, &extent_mutex_pool, (uintptr_t)extent); -} - -static inline void -extent_lock2(tsdn_t *tsdn, extent_t *extent1, extent_t *extent2) { - assert(extent1 != NULL && extent2 != NULL); - mutex_pool_lock2(tsdn, &extent_mutex_pool, (uintptr_t)extent1, - (uintptr_t)extent2); -} - -static inline void -extent_unlock2(tsdn_t *tsdn, extent_t *extent1, extent_t *extent2) { - assert(extent1 != NULL && extent2 != NULL); - mutex_pool_unlock2(tsdn, &extent_mutex_pool, (uintptr_t)extent1, - (uintptr_t)extent2); -} - -static inline unsigned -extent_arena_ind_get(const extent_t *extent) { - unsigned arena_ind = (unsigned)((extent->e_bits & - EXTENT_BITS_ARENA_MASK) >> EXTENT_BITS_ARENA_SHIFT); - assert(arena_ind < MALLOCX_ARENA_LIMIT); - - return arena_ind; -} - -static inline arena_t * -extent_arena_get(const extent_t *extent) { - unsigned arena_ind = extent_arena_ind_get(extent); - - return (arena_t *)atomic_load_p(&arenas[arena_ind], ATOMIC_ACQUIRE); -} - -static inline szind_t -extent_szind_get_maybe_invalid(const extent_t *extent) { - szind_t szind = (szind_t)((extent->e_bits & EXTENT_BITS_SZIND_MASK) >> - EXTENT_BITS_SZIND_SHIFT); - assert(szind <= SC_NSIZES); - return szind; -} - -static inline szind_t -extent_szind_get(const extent_t *extent) { - szind_t szind = extent_szind_get_maybe_invalid(extent); - assert(szind < SC_NSIZES); /* Never call when "invalid". */ - return szind; -} - -static inline size_t -extent_usize_get(const extent_t *extent) { - return sz_index2size(extent_szind_get(extent)); -} - -static inline unsigned -extent_binshard_get(const extent_t *extent) { - unsigned binshard = (unsigned)((extent->e_bits & - EXTENT_BITS_BINSHARD_MASK) >> EXTENT_BITS_BINSHARD_SHIFT); - assert(binshard < bin_infos[extent_szind_get(extent)].n_shards); - return binshard; -} - -static inline size_t -extent_sn_get(const extent_t *extent) { - return (size_t)((extent->e_bits & EXTENT_BITS_SN_MASK) >> - EXTENT_BITS_SN_SHIFT); -} - -static inline extent_state_t -extent_state_get(const extent_t *extent) { - return (extent_state_t)((extent->e_bits & EXTENT_BITS_STATE_MASK) >> - EXTENT_BITS_STATE_SHIFT); -} - -static inline bool -extent_zeroed_get(const extent_t *extent) { - return (bool)((extent->e_bits & EXTENT_BITS_ZEROED_MASK) >> - EXTENT_BITS_ZEROED_SHIFT); -} - -static inline bool -extent_committed_get(const extent_t *extent) { - return (bool)((extent->e_bits & EXTENT_BITS_COMMITTED_MASK) >> - EXTENT_BITS_COMMITTED_SHIFT); -} - -static inline bool -extent_dumpable_get(const extent_t *extent) { - return (bool)((extent->e_bits & EXTENT_BITS_DUMPABLE_MASK) >> - EXTENT_BITS_DUMPABLE_SHIFT); -} - -static inline bool -extent_slab_get(const extent_t *extent) { - return (bool)((extent->e_bits & EXTENT_BITS_SLAB_MASK) >> - EXTENT_BITS_SLAB_SHIFT); -} - -static inline unsigned -extent_nfree_get(const extent_t *extent) { - assert(extent_slab_get(extent)); - return (unsigned)((extent->e_bits & EXTENT_BITS_NFREE_MASK) >> - EXTENT_BITS_NFREE_SHIFT); -} - -static inline void * -extent_base_get(const extent_t *extent) { - assert(extent->e_addr == PAGE_ADDR2BASE(extent->e_addr) || - !extent_slab_get(extent)); - return PAGE_ADDR2BASE(extent->e_addr); -} - -static inline void * -extent_addr_get(const extent_t *extent) { - assert(extent->e_addr == PAGE_ADDR2BASE(extent->e_addr) || - !extent_slab_get(extent)); - return extent->e_addr; -} - -static inline size_t -extent_size_get(const extent_t *extent) { - return (extent->e_size_esn & EXTENT_SIZE_MASK); -} - -static inline size_t -extent_esn_get(const extent_t *extent) { - return (extent->e_size_esn & EXTENT_ESN_MASK); -} - -static inline size_t -extent_bsize_get(const extent_t *extent) { - return extent->e_bsize; -} - -static inline void * -extent_before_get(const extent_t *extent) { - return (void *)((uintptr_t)extent_base_get(extent) - PAGE); -} - -static inline void * -extent_last_get(const extent_t *extent) { - return (void *)((uintptr_t)extent_base_get(extent) + - extent_size_get(extent) - PAGE); -} - -static inline void * -extent_past_get(const extent_t *extent) { - return (void *)((uintptr_t)extent_base_get(extent) + - extent_size_get(extent)); -} - -static inline arena_slab_data_t * -extent_slab_data_get(extent_t *extent) { - assert(extent_slab_get(extent)); - return &extent->e_slab_data; -} - -static inline const arena_slab_data_t * -extent_slab_data_get_const(const extent_t *extent) { - assert(extent_slab_get(extent)); - return &extent->e_slab_data; -} - -static inline prof_tctx_t * -extent_prof_tctx_get(const extent_t *extent) { - return (prof_tctx_t *)atomic_load_p(&extent->e_prof_tctx, - ATOMIC_ACQUIRE); -} - -static inline nstime_t -extent_prof_alloc_time_get(const extent_t *extent) { - return extent->e_alloc_time; -} - -static inline void -extent_arena_set(extent_t *extent, arena_t *arena) { - unsigned arena_ind = (arena != NULL) ? arena_ind_get(arena) : ((1U << - MALLOCX_ARENA_BITS) - 1); - extent->e_bits = (extent->e_bits & ~EXTENT_BITS_ARENA_MASK) | - ((uint64_t)arena_ind << EXTENT_BITS_ARENA_SHIFT); -} - -static inline void -extent_binshard_set(extent_t *extent, unsigned binshard) { - /* The assertion assumes szind is set already. */ - assert(binshard < bin_infos[extent_szind_get(extent)].n_shards); - extent->e_bits = (extent->e_bits & ~EXTENT_BITS_BINSHARD_MASK) | - ((uint64_t)binshard << EXTENT_BITS_BINSHARD_SHIFT); -} - -static inline void -extent_addr_set(extent_t *extent, void *addr) { - extent->e_addr = addr; -} - -static inline void -extent_addr_randomize(tsdn_t *tsdn, extent_t *extent, size_t alignment) { - assert(extent_base_get(extent) == extent_addr_get(extent)); - - if (alignment < PAGE) { - unsigned lg_range = LG_PAGE - - lg_floor(CACHELINE_CEILING(alignment)); - size_t r; - if (!tsdn_null(tsdn)) { - tsd_t *tsd = tsdn_tsd(tsdn); - r = (size_t)prng_lg_range_u64( - tsd_offset_statep_get(tsd), lg_range); - } else { - r = prng_lg_range_zu( - &extent_arena_get(extent)->offset_state, - lg_range, true); - } - uintptr_t random_offset = ((uintptr_t)r) << (LG_PAGE - - lg_range); - extent->e_addr = (void *)((uintptr_t)extent->e_addr + - random_offset); - assert(ALIGNMENT_ADDR2BASE(extent->e_addr, alignment) == - extent->e_addr); - } -} - -static inline void -extent_size_set(extent_t *extent, size_t size) { - assert((size & ~EXTENT_SIZE_MASK) == 0); - extent->e_size_esn = size | (extent->e_size_esn & ~EXTENT_SIZE_MASK); -} - -static inline void -extent_esn_set(extent_t *extent, size_t esn) { - extent->e_size_esn = (extent->e_size_esn & ~EXTENT_ESN_MASK) | (esn & - EXTENT_ESN_MASK); -} - -static inline void -extent_bsize_set(extent_t *extent, size_t bsize) { - extent->e_bsize = bsize; -} - -static inline void -extent_szind_set(extent_t *extent, szind_t szind) { - assert(szind <= SC_NSIZES); /* SC_NSIZES means "invalid". */ - extent->e_bits = (extent->e_bits & ~EXTENT_BITS_SZIND_MASK) | - ((uint64_t)szind << EXTENT_BITS_SZIND_SHIFT); -} - -static inline void -extent_nfree_set(extent_t *extent, unsigned nfree) { - assert(extent_slab_get(extent)); - extent->e_bits = (extent->e_bits & ~EXTENT_BITS_NFREE_MASK) | - ((uint64_t)nfree << EXTENT_BITS_NFREE_SHIFT); -} - -static inline void -extent_nfree_binshard_set(extent_t *extent, unsigned nfree, unsigned binshard) { - /* The assertion assumes szind is set already. */ - assert(binshard < bin_infos[extent_szind_get(extent)].n_shards); - extent->e_bits = (extent->e_bits & - (~EXTENT_BITS_NFREE_MASK & ~EXTENT_BITS_BINSHARD_MASK)) | - ((uint64_t)binshard << EXTENT_BITS_BINSHARD_SHIFT) | - ((uint64_t)nfree << EXTENT_BITS_NFREE_SHIFT); -} - -static inline void -extent_nfree_inc(extent_t *extent) { - assert(extent_slab_get(extent)); - extent->e_bits += ((uint64_t)1U << EXTENT_BITS_NFREE_SHIFT); -} - -static inline void -extent_nfree_dec(extent_t *extent) { - assert(extent_slab_get(extent)); - extent->e_bits -= ((uint64_t)1U << EXTENT_BITS_NFREE_SHIFT); -} - -static inline void -extent_nfree_sub(extent_t *extent, uint64_t n) { - assert(extent_slab_get(extent)); - extent->e_bits -= (n << EXTENT_BITS_NFREE_SHIFT); -} - -static inline void -extent_sn_set(extent_t *extent, size_t sn) { - extent->e_bits = (extent->e_bits & ~EXTENT_BITS_SN_MASK) | - ((uint64_t)sn << EXTENT_BITS_SN_SHIFT); -} - -static inline void -extent_state_set(extent_t *extent, extent_state_t state) { - extent->e_bits = (extent->e_bits & ~EXTENT_BITS_STATE_MASK) | - ((uint64_t)state << EXTENT_BITS_STATE_SHIFT); -} - -static inline void -extent_zeroed_set(extent_t *extent, bool zeroed) { - extent->e_bits = (extent->e_bits & ~EXTENT_BITS_ZEROED_MASK) | - ((uint64_t)zeroed << EXTENT_BITS_ZEROED_SHIFT); -} - -static inline void -extent_committed_set(extent_t *extent, bool committed) { - extent->e_bits = (extent->e_bits & ~EXTENT_BITS_COMMITTED_MASK) | - ((uint64_t)committed << EXTENT_BITS_COMMITTED_SHIFT); -} - -static inline void -extent_dumpable_set(extent_t *extent, bool dumpable) { - extent->e_bits = (extent->e_bits & ~EXTENT_BITS_DUMPABLE_MASK) | - ((uint64_t)dumpable << EXTENT_BITS_DUMPABLE_SHIFT); -} - -static inline void -extent_slab_set(extent_t *extent, bool slab) { - extent->e_bits = (extent->e_bits & ~EXTENT_BITS_SLAB_MASK) | - ((uint64_t)slab << EXTENT_BITS_SLAB_SHIFT); -} - -static inline void -extent_prof_tctx_set(extent_t *extent, prof_tctx_t *tctx) { - atomic_store_p(&extent->e_prof_tctx, tctx, ATOMIC_RELEASE); -} - -static inline void -extent_prof_alloc_time_set(extent_t *extent, nstime_t t) { - nstime_copy(&extent->e_alloc_time, &t); -} - -static inline bool -extent_is_head_get(extent_t *extent) { - if (maps_coalesce) { - not_reached(); - } - - return (bool)((extent->e_bits & EXTENT_BITS_IS_HEAD_MASK) >> - EXTENT_BITS_IS_HEAD_SHIFT); -} - -static inline void -extent_is_head_set(extent_t *extent, bool is_head) { - if (maps_coalesce) { - not_reached(); - } - - extent->e_bits = (extent->e_bits & ~EXTENT_BITS_IS_HEAD_MASK) | - ((uint64_t)is_head << EXTENT_BITS_IS_HEAD_SHIFT); -} - -static inline void -extent_init(extent_t *extent, arena_t *arena, void *addr, size_t size, - bool slab, szind_t szind, size_t sn, extent_state_t state, bool zeroed, - bool committed, bool dumpable, extent_head_state_t is_head) { - assert(addr == PAGE_ADDR2BASE(addr) || !slab); - - extent_arena_set(extent, arena); - extent_addr_set(extent, addr); - extent_size_set(extent, size); - extent_slab_set(extent, slab); - extent_szind_set(extent, szind); - extent_sn_set(extent, sn); - extent_state_set(extent, state); - extent_zeroed_set(extent, zeroed); - extent_committed_set(extent, committed); - extent_dumpable_set(extent, dumpable); - ql_elm_new(extent, ql_link); - if (!maps_coalesce) { - extent_is_head_set(extent, (is_head == EXTENT_IS_HEAD) ? true : - false); - } - if (config_prof) { - extent_prof_tctx_set(extent, NULL); - } -} - -static inline void -extent_binit(extent_t *extent, void *addr, size_t bsize, size_t sn) { - extent_arena_set(extent, NULL); - extent_addr_set(extent, addr); - extent_bsize_set(extent, bsize); - extent_slab_set(extent, false); - extent_szind_set(extent, SC_NSIZES); - extent_sn_set(extent, sn); - extent_state_set(extent, extent_state_active); - extent_zeroed_set(extent, true); - extent_committed_set(extent, true); - extent_dumpable_set(extent, true); -} - -static inline void -extent_list_init(extent_list_t *list) { - ql_new(list); -} - -static inline extent_t * -extent_list_first(const extent_list_t *list) { - return ql_first(list); -} - -static inline extent_t * -extent_list_last(const extent_list_t *list) { - return ql_last(list, ql_link); -} - -static inline void -extent_list_append(extent_list_t *list, extent_t *extent) { - ql_tail_insert(list, extent, ql_link); -} - -static inline void -extent_list_prepend(extent_list_t *list, extent_t *extent) { - ql_head_insert(list, extent, ql_link); -} - -static inline void -extent_list_replace(extent_list_t *list, extent_t *to_remove, - extent_t *to_insert) { - ql_after_insert(to_remove, to_insert, ql_link); - ql_remove(list, to_remove, ql_link); -} - -static inline void -extent_list_remove(extent_list_t *list, extent_t *extent) { - ql_remove(list, extent, ql_link); -} - -static inline int -extent_sn_comp(const extent_t *a, const extent_t *b) { - size_t a_sn = extent_sn_get(a); - size_t b_sn = extent_sn_get(b); - - return (a_sn > b_sn) - (a_sn < b_sn); -} - -static inline int -extent_esn_comp(const extent_t *a, const extent_t *b) { - size_t a_esn = extent_esn_get(a); - size_t b_esn = extent_esn_get(b); - - return (a_esn > b_esn) - (a_esn < b_esn); -} - -static inline int -extent_ad_comp(const extent_t *a, const extent_t *b) { - uintptr_t a_addr = (uintptr_t)extent_addr_get(a); - uintptr_t b_addr = (uintptr_t)extent_addr_get(b); - - return (a_addr > b_addr) - (a_addr < b_addr); -} - -static inline int -extent_ead_comp(const extent_t *a, const extent_t *b) { - uintptr_t a_eaddr = (uintptr_t)a; - uintptr_t b_eaddr = (uintptr_t)b; - - return (a_eaddr > b_eaddr) - (a_eaddr < b_eaddr); -} - -static inline int -extent_snad_comp(const extent_t *a, const extent_t *b) { - int ret; - - ret = extent_sn_comp(a, b); - if (ret != 0) { - return ret; - } - - ret = extent_ad_comp(a, b); - return ret; -} - -static inline int -extent_esnead_comp(const extent_t *a, const extent_t *b) { - int ret; - - ret = extent_esn_comp(a, b); - if (ret != 0) { - return ret; - } - - ret = extent_ead_comp(a, b); - return ret; -} - -#endif /* JEMALLOC_INTERNAL_EXTENT_INLINES_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/extent_structs.h b/contrib/jemalloc/include/jemalloc/internal/extent_structs.h deleted file mode 100644 --- a/contrib/jemalloc/include/jemalloc/internal/extent_structs.h +++ /dev/null @@ -1,256 +0,0 @@ -#ifndef JEMALLOC_INTERNAL_EXTENT_STRUCTS_H -#define JEMALLOC_INTERNAL_EXTENT_STRUCTS_H - -#include "jemalloc/internal/atomic.h" -#include "jemalloc/internal/bit_util.h" -#include "jemalloc/internal/bitmap.h" -#include "jemalloc/internal/mutex.h" -#include "jemalloc/internal/ql.h" -#include "jemalloc/internal/ph.h" -#include "jemalloc/internal/sc.h" - -typedef enum { - extent_state_active = 0, - extent_state_dirty = 1, - extent_state_muzzy = 2, - extent_state_retained = 3 -} extent_state_t; - -/* Extent (span of pages). Use accessor functions for e_* fields. */ -struct extent_s { - /* - * Bitfield containing several fields: - * - * a: arena_ind - * b: slab - * c: committed - * d: dumpable - * z: zeroed - * t: state - * i: szind - * f: nfree - * s: bin_shard - * n: sn - * - * nnnnnnnn ... nnnnnnss ssssffff ffffffii iiiiiitt zdcbaaaa aaaaaaaa - * - * arena_ind: Arena from which this extent came, or all 1 bits if - * unassociated. - * - * slab: The slab flag indicates whether the extent is used for a slab - * of small regions. This helps differentiate small size classes, - * and it indicates whether interior pointers can be looked up via - * iealloc(). - * - * committed: The committed flag indicates whether physical memory is - * committed to the extent, whether explicitly or implicitly - * as on a system that overcommits and satisfies physical - * memory needs on demand via soft page faults. - * - * dumpable: The dumpable flag indicates whether or not we've set the - * memory in question to be dumpable. Note that this - * interacts somewhat subtly with user-specified extent hooks, - * since we don't know if *they* are fiddling with - * dumpability (in which case, we don't want to undo whatever - * they're doing). To deal with this scenario, we: - * - Make dumpable false only for memory allocated with the - * default hooks. - * - Only allow memory to go from non-dumpable to dumpable, - * and only once. - * - Never make the OS call to allow dumping when the - * dumpable bit is already set. - * These three constraints mean that we will never - * accidentally dump user memory that the user meant to set - * nondumpable with their extent hooks. - * - * - * zeroed: The zeroed flag is used by extent recycling code to track - * whether memory is zero-filled. - * - * state: The state flag is an extent_state_t. - * - * szind: The szind flag indicates usable size class index for - * allocations residing in this extent, regardless of whether the - * extent is a slab. Extent size and usable size often differ - * even for non-slabs, either due to sz_large_pad or promotion of - * sampled small regions. - * - * nfree: Number of free regions in slab. - * - * bin_shard: the shard of the bin from which this extent came. - * - * sn: Serial number (potentially non-unique). - * - * Serial numbers may wrap around if !opt_retain, but as long as - * comparison functions fall back on address comparison for equal - * serial numbers, stable (if imperfect) ordering is maintained. - * - * Serial numbers may not be unique even in the absence of - * wrap-around, e.g. when splitting an extent and assigning the same - * serial number to both resulting adjacent extents. - */ - uint64_t e_bits; -#define MASK(CURRENT_FIELD_WIDTH, CURRENT_FIELD_SHIFT) ((((((uint64_t)0x1U) << (CURRENT_FIELD_WIDTH)) - 1)) << (CURRENT_FIELD_SHIFT)) - -#define EXTENT_BITS_ARENA_WIDTH MALLOCX_ARENA_BITS -#define EXTENT_BITS_ARENA_SHIFT 0 -#define EXTENT_BITS_ARENA_MASK MASK(EXTENT_BITS_ARENA_WIDTH, EXTENT_BITS_ARENA_SHIFT) - -#define EXTENT_BITS_SLAB_WIDTH 1 -#define EXTENT_BITS_SLAB_SHIFT (EXTENT_BITS_ARENA_WIDTH + EXTENT_BITS_ARENA_SHIFT) -#define EXTENT_BITS_SLAB_MASK MASK(EXTENT_BITS_SLAB_WIDTH, EXTENT_BITS_SLAB_SHIFT) - -#define EXTENT_BITS_COMMITTED_WIDTH 1 -#define EXTENT_BITS_COMMITTED_SHIFT (EXTENT_BITS_SLAB_WIDTH + EXTENT_BITS_SLAB_SHIFT) -#define EXTENT_BITS_COMMITTED_MASK MASK(EXTENT_BITS_COMMITTED_WIDTH, EXTENT_BITS_COMMITTED_SHIFT) - -#define EXTENT_BITS_DUMPABLE_WIDTH 1 -#define EXTENT_BITS_DUMPABLE_SHIFT (EXTENT_BITS_COMMITTED_WIDTH + EXTENT_BITS_COMMITTED_SHIFT) -#define EXTENT_BITS_DUMPABLE_MASK MASK(EXTENT_BITS_DUMPABLE_WIDTH, EXTENT_BITS_DUMPABLE_SHIFT) - -#define EXTENT_BITS_ZEROED_WIDTH 1 -#define EXTENT_BITS_ZEROED_SHIFT (EXTENT_BITS_DUMPABLE_WIDTH + EXTENT_BITS_DUMPABLE_SHIFT) -#define EXTENT_BITS_ZEROED_MASK MASK(EXTENT_BITS_ZEROED_WIDTH, EXTENT_BITS_ZEROED_SHIFT) - -#define EXTENT_BITS_STATE_WIDTH 2 -#define EXTENT_BITS_STATE_SHIFT (EXTENT_BITS_ZEROED_WIDTH + EXTENT_BITS_ZEROED_SHIFT) -#define EXTENT_BITS_STATE_MASK MASK(EXTENT_BITS_STATE_WIDTH, EXTENT_BITS_STATE_SHIFT) - -#define EXTENT_BITS_SZIND_WIDTH LG_CEIL(SC_NSIZES) -#define EXTENT_BITS_SZIND_SHIFT (EXTENT_BITS_STATE_WIDTH + EXTENT_BITS_STATE_SHIFT) -#define EXTENT_BITS_SZIND_MASK MASK(EXTENT_BITS_SZIND_WIDTH, EXTENT_BITS_SZIND_SHIFT) - -#define EXTENT_BITS_NFREE_WIDTH (LG_SLAB_MAXREGS + 1) -#define EXTENT_BITS_NFREE_SHIFT (EXTENT_BITS_SZIND_WIDTH + EXTENT_BITS_SZIND_SHIFT) -#define EXTENT_BITS_NFREE_MASK MASK(EXTENT_BITS_NFREE_WIDTH, EXTENT_BITS_NFREE_SHIFT) - -#define EXTENT_BITS_BINSHARD_WIDTH 6 -#define EXTENT_BITS_BINSHARD_SHIFT (EXTENT_BITS_NFREE_WIDTH + EXTENT_BITS_NFREE_SHIFT) -#define EXTENT_BITS_BINSHARD_MASK MASK(EXTENT_BITS_BINSHARD_WIDTH, EXTENT_BITS_BINSHARD_SHIFT) - -#define EXTENT_BITS_IS_HEAD_WIDTH 1 -#define EXTENT_BITS_IS_HEAD_SHIFT (EXTENT_BITS_BINSHARD_WIDTH + EXTENT_BITS_BINSHARD_SHIFT) -#define EXTENT_BITS_IS_HEAD_MASK MASK(EXTENT_BITS_IS_HEAD_WIDTH, EXTENT_BITS_IS_HEAD_SHIFT) - -#define EXTENT_BITS_SN_SHIFT (EXTENT_BITS_IS_HEAD_WIDTH + EXTENT_BITS_IS_HEAD_SHIFT) -#define EXTENT_BITS_SN_MASK (UINT64_MAX << EXTENT_BITS_SN_SHIFT) - - /* Pointer to the extent that this structure is responsible for. */ - void *e_addr; - - union { - /* - * Extent size and serial number associated with the extent - * structure (different than the serial number for the extent at - * e_addr). - * - * ssssssss [...] ssssssss ssssnnnn nnnnnnnn - */ - size_t e_size_esn; - #define EXTENT_SIZE_MASK ((size_t)~(PAGE-1)) - #define EXTENT_ESN_MASK ((size_t)PAGE-1) - /* Base extent size, which may not be a multiple of PAGE. */ - size_t e_bsize; - }; - - /* - * List linkage, used by a variety of lists: - * - bin_t's slabs_full - * - extents_t's LRU - * - stashed dirty extents - * - arena's large allocations - */ - ql_elm(extent_t) ql_link; - - /* - * Linkage for per size class sn/address-ordered heaps, and - * for extent_avail - */ - phn(extent_t) ph_link; - - union { - /* Small region slab metadata. */ - arena_slab_data_t e_slab_data; - - /* Profiling data, used for large objects. */ - struct { - /* Time when this was allocated. */ - nstime_t e_alloc_time; - /* Points to a prof_tctx_t. */ - atomic_p_t e_prof_tctx; - }; - }; -}; -typedef ql_head(extent_t) extent_list_t; -typedef ph(extent_t) extent_tree_t; -typedef ph(extent_t) extent_heap_t; - -/* Quantized collection of extents, with built-in LRU queue. */ -struct extents_s { - malloc_mutex_t mtx; - - /* - * Quantized per size class heaps of extents. - * - * Synchronization: mtx. - */ - extent_heap_t heaps[SC_NPSIZES + 1]; - atomic_zu_t nextents[SC_NPSIZES + 1]; - atomic_zu_t nbytes[SC_NPSIZES + 1]; - - /* - * Bitmap for which set bits correspond to non-empty heaps. - * - * Synchronization: mtx. - */ - bitmap_t bitmap[BITMAP_GROUPS(SC_NPSIZES + 1)]; - - /* - * LRU of all extents in heaps. - * - * Synchronization: mtx. - */ - extent_list_t lru; - - /* - * Page sum for all extents in heaps. - * - * The synchronization here is a little tricky. Modifications to npages - * must hold mtx, but reads need not (though, a reader who sees npages - * without holding the mutex can't assume anything about the rest of the - * state of the extents_t). - */ - atomic_zu_t npages; - - /* All stored extents must be in the same state. */ - extent_state_t state; - - /* - * If true, delay coalescing until eviction; otherwise coalesce during - * deallocation. - */ - bool delay_coalesce; -}; - -/* - * The following two structs are for experimental purposes. See - * experimental_utilization_query_ctl and - * experimental_utilization_batch_query_ctl in src/ctl.c. - */ - -struct extent_util_stats_s { - size_t nfree; - size_t nregs; - size_t size; -}; - -struct extent_util_stats_verbose_s { - void *slabcur_addr; - size_t nfree; - size_t nregs; - size_t size; - size_t bin_nfree; - size_t bin_nregs; -}; - -#endif /* JEMALLOC_INTERNAL_EXTENT_STRUCTS_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/extent_types.h b/contrib/jemalloc/include/jemalloc/internal/extent_types.h deleted file mode 100644 --- a/contrib/jemalloc/include/jemalloc/internal/extent_types.h +++ /dev/null @@ -1,23 +0,0 @@ -#ifndef JEMALLOC_INTERNAL_EXTENT_TYPES_H -#define JEMALLOC_INTERNAL_EXTENT_TYPES_H - -typedef struct extent_s extent_t; -typedef struct extents_s extents_t; - -typedef struct extent_util_stats_s extent_util_stats_t; -typedef struct extent_util_stats_verbose_s extent_util_stats_verbose_t; - -#define EXTENT_HOOKS_INITIALIZER NULL - -/* - * When reuse (and split) an active extent, (1U << opt_lg_extent_max_active_fit) - * is the max ratio between the size of the active extent and the new extent. - */ -#define LG_EXTENT_MAX_ACTIVE_FIT_DEFAULT 6 - -typedef enum { - EXTENT_NOT_HEAD, - EXTENT_IS_HEAD /* Only relevant for Windows && opt.retain. */ -} extent_head_state_t; - -#endif /* JEMALLOC_INTERNAL_EXTENT_TYPES_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/fb.h b/contrib/jemalloc/include/jemalloc/internal/fb.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/fb.h @@ -0,0 +1,373 @@ +#ifndef JEMALLOC_INTERNAL_FB_H +#define JEMALLOC_INTERNAL_FB_H + +/* + * The flat bitmap module. This has a larger API relative to the bitmap module + * (supporting things like backwards searches, and searching for both set and + * unset bits), at the cost of slower operations for very large bitmaps. + * + * Initialized flat bitmaps start at all-zeros (all bits unset). + */ + +typedef unsigned long fb_group_t; +#define FB_GROUP_BITS (ZU(1) << (LG_SIZEOF_LONG + 3)) +#define FB_NGROUPS(nbits) ((nbits) / FB_GROUP_BITS \ + + ((nbits) % FB_GROUP_BITS == 0 ? 0 : 1)) + +static inline void +fb_init(fb_group_t *fb, size_t nbits) { + size_t ngroups = FB_NGROUPS(nbits); + memset(fb, 0, ngroups * sizeof(fb_group_t)); +} + +static inline bool +fb_empty(fb_group_t *fb, size_t nbits) { + size_t ngroups = FB_NGROUPS(nbits); + for (size_t i = 0; i < ngroups; i++) { + if (fb[i] != 0) { + return false; + } + } + return true; +} + +static inline bool +fb_full(fb_group_t *fb, size_t nbits) { + size_t ngroups = FB_NGROUPS(nbits); + size_t trailing_bits = nbits % FB_GROUP_BITS; + size_t limit = (trailing_bits == 0 ? ngroups : ngroups - 1); + for (size_t i = 0; i < limit; i++) { + if (fb[i] != ~(fb_group_t)0) { + return false; + } + } + if (trailing_bits == 0) { + return true; + } + return fb[ngroups - 1] == ((fb_group_t)1 << trailing_bits) - 1; +} + +static inline bool +fb_get(fb_group_t *fb, size_t nbits, size_t bit) { + assert(bit < nbits); + size_t group_ind = bit / FB_GROUP_BITS; + size_t bit_ind = bit % FB_GROUP_BITS; + return (bool)(fb[group_ind] & ((fb_group_t)1 << bit_ind)); +} + +static inline void +fb_set(fb_group_t *fb, size_t nbits, size_t bit) { + assert(bit < nbits); + size_t group_ind = bit / FB_GROUP_BITS; + size_t bit_ind = bit % FB_GROUP_BITS; + fb[group_ind] |= ((fb_group_t)1 << bit_ind); +} + +static inline void +fb_unset(fb_group_t *fb, size_t nbits, size_t bit) { + assert(bit < nbits); + size_t group_ind = bit / FB_GROUP_BITS; + size_t bit_ind = bit % FB_GROUP_BITS; + fb[group_ind] &= ~((fb_group_t)1 << bit_ind); +} + + +/* + * Some implementation details. This visitation function lets us apply a group + * visitor to each group in the bitmap (potentially modifying it). The mask + * indicates which bits are logically part of the visitation. + */ +typedef void (*fb_group_visitor_t)(void *ctx, fb_group_t *fb, fb_group_t mask); +JEMALLOC_ALWAYS_INLINE void +fb_visit_impl(fb_group_t *fb, size_t nbits, fb_group_visitor_t visit, void *ctx, + size_t start, size_t cnt) { + assert(cnt > 0); + assert(start + cnt <= nbits); + size_t group_ind = start / FB_GROUP_BITS; + size_t start_bit_ind = start % FB_GROUP_BITS; + /* + * The first group is special; it's the only one we don't start writing + * to from bit 0. + */ + size_t first_group_cnt = (start_bit_ind + cnt > FB_GROUP_BITS + ? FB_GROUP_BITS - start_bit_ind : cnt); + /* + * We can basically split affected words into: + * - The first group, where we touch only the high bits + * - The last group, where we touch only the low bits + * - The middle, where we set all the bits to the same thing. + * We treat each case individually. The last two could be merged, but + * this can lead to bad codegen for those middle words. + */ + /* First group */ + fb_group_t mask = ((~(fb_group_t)0) + >> (FB_GROUP_BITS - first_group_cnt)) + << start_bit_ind; + visit(ctx, &fb[group_ind], mask); + + cnt -= first_group_cnt; + group_ind++; + /* Middle groups */ + while (cnt > FB_GROUP_BITS) { + visit(ctx, &fb[group_ind], ~(fb_group_t)0); + cnt -= FB_GROUP_BITS; + group_ind++; + } + /* Last group */ + if (cnt != 0) { + mask = (~(fb_group_t)0) >> (FB_GROUP_BITS - cnt); + visit(ctx, &fb[group_ind], mask); + } +} + +JEMALLOC_ALWAYS_INLINE void +fb_assign_visitor(void *ctx, fb_group_t *fb, fb_group_t mask) { + bool val = *(bool *)ctx; + if (val) { + *fb |= mask; + } else { + *fb &= ~mask; + } +} + +/* Sets the cnt bits starting at position start. Must not have a 0 count. */ +static inline void +fb_set_range(fb_group_t *fb, size_t nbits, size_t start, size_t cnt) { + bool val = true; + fb_visit_impl(fb, nbits, &fb_assign_visitor, &val, start, cnt); +} + +/* Unsets the cnt bits starting at position start. Must not have a 0 count. */ +static inline void +fb_unset_range(fb_group_t *fb, size_t nbits, size_t start, size_t cnt) { + bool val = false; + fb_visit_impl(fb, nbits, &fb_assign_visitor, &val, start, cnt); +} + +JEMALLOC_ALWAYS_INLINE void +fb_scount_visitor(void *ctx, fb_group_t *fb, fb_group_t mask) { + size_t *scount = (size_t *)ctx; + *scount += popcount_lu(*fb & mask); +} + +/* Finds the number of set bit in the of length cnt starting at start. */ +JEMALLOC_ALWAYS_INLINE size_t +fb_scount(fb_group_t *fb, size_t nbits, size_t start, size_t cnt) { + size_t scount = 0; + fb_visit_impl(fb, nbits, &fb_scount_visitor, &scount, start, cnt); + return scount; +} + +/* Finds the number of unset bit in the of length cnt starting at start. */ +JEMALLOC_ALWAYS_INLINE size_t +fb_ucount(fb_group_t *fb, size_t nbits, size_t start, size_t cnt) { + size_t scount = fb_scount(fb, nbits, start, cnt); + return cnt - scount; +} + +/* + * An implementation detail; find the first bit at position >= min_bit with the + * value val. + * + * Returns the number of bits in the bitmap if no such bit exists. + */ +JEMALLOC_ALWAYS_INLINE ssize_t +fb_find_impl(fb_group_t *fb, size_t nbits, size_t start, bool val, + bool forward) { + assert(start < nbits); + size_t ngroups = FB_NGROUPS(nbits); + ssize_t group_ind = start / FB_GROUP_BITS; + size_t bit_ind = start % FB_GROUP_BITS; + + fb_group_t maybe_invert = (val ? 0 : (fb_group_t)-1); + + fb_group_t group = fb[group_ind]; + group ^= maybe_invert; + if (forward) { + /* Only keep ones in bits bit_ind and above. */ + group &= ~((1LU << bit_ind) - 1); + } else { + /* + * Only keep ones in bits bit_ind and below. You might more + * naturally express this as (1 << (bit_ind + 1)) - 1, but + * that shifts by an invalid amount if bit_ind is one less than + * FB_GROUP_BITS. + */ + group &= ((2LU << bit_ind) - 1); + } + ssize_t group_ind_bound = forward ? (ssize_t)ngroups : -1; + while (group == 0) { + group_ind += forward ? 1 : -1; + if (group_ind == group_ind_bound) { + return forward ? (ssize_t)nbits : (ssize_t)-1; + } + group = fb[group_ind]; + group ^= maybe_invert; + } + assert(group != 0); + size_t bit = forward ? ffs_lu(group) : fls_lu(group); + size_t pos = group_ind * FB_GROUP_BITS + bit; + /* + * The high bits of a partially filled last group are zeros, so if we're + * looking for zeros we don't want to report an invalid result. + */ + if (forward && !val && pos > nbits) { + return nbits; + } + return pos; +} + +/* + * Find the first set bit in the bitmap with an index >= min_bit. Returns the + * number of bits in the bitmap if no such bit exists. + */ +static inline size_t +fb_ffu(fb_group_t *fb, size_t nbits, size_t min_bit) { + return (size_t)fb_find_impl(fb, nbits, min_bit, /* val */ false, + /* forward */ true); +} + +/* The same, but looks for an unset bit. */ +static inline size_t +fb_ffs(fb_group_t *fb, size_t nbits, size_t min_bit) { + return (size_t)fb_find_impl(fb, nbits, min_bit, /* val */ true, + /* forward */ true); +} + +/* + * Find the last set bit in the bitmap with an index <= max_bit. Returns -1 if + * no such bit exists. + */ +static inline ssize_t +fb_flu(fb_group_t *fb, size_t nbits, size_t max_bit) { + return fb_find_impl(fb, nbits, max_bit, /* val */ false, + /* forward */ false); +} + +static inline ssize_t +fb_fls(fb_group_t *fb, size_t nbits, size_t max_bit) { + return fb_find_impl(fb, nbits, max_bit, /* val */ true, + /* forward */ false); +} + +/* Returns whether or not we found a range. */ +JEMALLOC_ALWAYS_INLINE bool +fb_iter_range_impl(fb_group_t *fb, size_t nbits, size_t start, size_t *r_begin, + size_t *r_len, bool val, bool forward) { + assert(start < nbits); + ssize_t next_range_begin = fb_find_impl(fb, nbits, start, val, forward); + if ((forward && next_range_begin == (ssize_t)nbits) + || (!forward && next_range_begin == (ssize_t)-1)) { + return false; + } + /* Half open range; the set bits are [begin, end). */ + ssize_t next_range_end = fb_find_impl(fb, nbits, next_range_begin, !val, + forward); + if (forward) { + *r_begin = next_range_begin; + *r_len = next_range_end - next_range_begin; + } else { + *r_begin = next_range_end + 1; + *r_len = next_range_begin - next_range_end; + } + return true; +} + +/* + * Used to iterate through ranges of set bits. + * + * Tries to find the next contiguous sequence of set bits with a first index >= + * start. If one exists, puts the earliest bit of the range in *r_begin, its + * length in *r_len, and returns true. Otherwise, returns false (without + * touching *r_begin or *r_end). + */ +static inline bool +fb_srange_iter(fb_group_t *fb, size_t nbits, size_t start, size_t *r_begin, + size_t *r_len) { + return fb_iter_range_impl(fb, nbits, start, r_begin, r_len, + /* val */ true, /* forward */ true); +} + +/* + * The same as fb_srange_iter, but searches backwards from start rather than + * forwards. (The position returned is still the earliest bit in the range). + */ +static inline bool +fb_srange_riter(fb_group_t *fb, size_t nbits, size_t start, size_t *r_begin, + size_t *r_len) { + return fb_iter_range_impl(fb, nbits, start, r_begin, r_len, + /* val */ true, /* forward */ false); +} + +/* Similar to fb_srange_iter, but searches for unset bits. */ +static inline bool +fb_urange_iter(fb_group_t *fb, size_t nbits, size_t start, size_t *r_begin, + size_t *r_len) { + return fb_iter_range_impl(fb, nbits, start, r_begin, r_len, + /* val */ false, /* forward */ true); +} + +/* Similar to fb_srange_riter, but searches for unset bits. */ +static inline bool +fb_urange_riter(fb_group_t *fb, size_t nbits, size_t start, size_t *r_begin, + size_t *r_len) { + return fb_iter_range_impl(fb, nbits, start, r_begin, r_len, + /* val */ false, /* forward */ false); +} + +JEMALLOC_ALWAYS_INLINE size_t +fb_range_longest_impl(fb_group_t *fb, size_t nbits, bool val) { + size_t begin = 0; + size_t longest_len = 0; + size_t len = 0; + while (begin < nbits && fb_iter_range_impl(fb, nbits, begin, &begin, + &len, val, /* forward */ true)) { + if (len > longest_len) { + longest_len = len; + } + begin += len; + } + return longest_len; +} + +static inline size_t +fb_srange_longest(fb_group_t *fb, size_t nbits) { + return fb_range_longest_impl(fb, nbits, /* val */ true); +} + +static inline size_t +fb_urange_longest(fb_group_t *fb, size_t nbits) { + return fb_range_longest_impl(fb, nbits, /* val */ false); +} + +/* + * Initializes each bit of dst with the bitwise-AND of the corresponding bits of + * src1 and src2. All bitmaps must be the same size. + */ +static inline void +fb_bit_and(fb_group_t *dst, fb_group_t *src1, fb_group_t *src2, size_t nbits) { + size_t ngroups = FB_NGROUPS(nbits); + for (size_t i = 0; i < ngroups; i++) { + dst[i] = src1[i] & src2[i]; + } +} + +/* Like fb_bit_and, but with bitwise-OR. */ +static inline void +fb_bit_or(fb_group_t *dst, fb_group_t *src1, fb_group_t *src2, size_t nbits) { + size_t ngroups = FB_NGROUPS(nbits); + for (size_t i = 0; i < ngroups; i++) { + dst[i] = src1[i] | src2[i]; + } +} + +/* Initializes dst bit i to the negation of source bit i. */ +static inline void +fb_bit_not(fb_group_t *dst, fb_group_t *src, size_t nbits) { + size_t ngroups = FB_NGROUPS(nbits); + for (size_t i = 0; i < ngroups; i++) { + dst[i] = ~src[i]; + } +} + +#endif /* JEMALLOC_INTERNAL_FB_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/fxp.h b/contrib/jemalloc/include/jemalloc/internal/fxp.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/fxp.h @@ -0,0 +1,126 @@ +#ifndef JEMALLOC_INTERNAL_FXP_H +#define JEMALLOC_INTERNAL_FXP_H + +/* + * A simple fixed-point math implementation, supporting only unsigned values + * (with overflow being an error). + * + * It's not in general safe to use floating point in core code, because various + * libc implementations we get linked against can assume that malloc won't touch + * floating point state and call it with an unusual calling convention. + */ + +/* + * High 16 bits are the integer part, low 16 are the fractional part. Or + * equivalently, repr == 2**16 * val, where we use "val" to refer to the + * (imaginary) fractional representation of the true value. + * + * We pick a uint32_t here since it's convenient in some places to + * double the representation size (i.e. multiplication and division use + * 64-bit integer types), and a uint64_t is the largest type we're + * certain is available. + */ +typedef uint32_t fxp_t; +#define FXP_INIT_INT(x) ((x) << 16) +#define FXP_INIT_PERCENT(pct) (((pct) << 16) / 100) + +/* + * Amount of precision used in parsing and printing numbers. The integer bound + * is simply because the integer part of the number gets 16 bits, and so is + * bounded by 65536. + * + * We use a lot of precision for the fractional part, even though most of it + * gets rounded off; this lets us get exact values for the important special + * case where the denominator is a small power of 2 (for instance, + * 1/512 == 0.001953125 is exactly representable even with only 16 bits of + * fractional precision). We need to left-shift by 16 before dividing by + * 10**precision, so we pick precision to be floor(log(2**48)) = 14. + */ +#define FXP_INTEGER_PART_DIGITS 5 +#define FXP_FRACTIONAL_PART_DIGITS 14 + +/* + * In addition to the integer and fractional parts of the number, we need to + * include a null character and (possibly) a decimal point. + */ +#define FXP_BUF_SIZE (FXP_INTEGER_PART_DIGITS + FXP_FRACTIONAL_PART_DIGITS + 2) + +static inline fxp_t +fxp_add(fxp_t a, fxp_t b) { + return a + b; +} + +static inline fxp_t +fxp_sub(fxp_t a, fxp_t b) { + assert(a >= b); + return a - b; +} + +static inline fxp_t +fxp_mul(fxp_t a, fxp_t b) { + uint64_t unshifted = (uint64_t)a * (uint64_t)b; + /* + * Unshifted is (a.val * 2**16) * (b.val * 2**16) + * == (a.val * b.val) * 2**32, but we want + * (a.val * b.val) * 2 ** 16. + */ + return (uint32_t)(unshifted >> 16); +} + +static inline fxp_t +fxp_div(fxp_t a, fxp_t b) { + assert(b != 0); + uint64_t unshifted = ((uint64_t)a << 32) / (uint64_t)b; + /* + * Unshifted is (a.val * 2**16) * (2**32) / (b.val * 2**16) + * == (a.val / b.val) * (2 ** 32), which again corresponds to a right + * shift of 16. + */ + return (uint32_t)(unshifted >> 16); +} + +static inline uint32_t +fxp_round_down(fxp_t a) { + return a >> 16; +} + +static inline uint32_t +fxp_round_nearest(fxp_t a) { + uint32_t fractional_part = (a & ((1U << 16) - 1)); + uint32_t increment = (uint32_t)(fractional_part >= (1U << 15)); + return (a >> 16) + increment; +} + +/* + * Approximately computes x * frac, without the size limitations that would be + * imposed by converting u to an fxp_t. + */ +static inline size_t +fxp_mul_frac(size_t x_orig, fxp_t frac) { + assert(frac <= (1U << 16)); + /* + * Work around an over-enthusiastic warning about type limits below (on + * 32-bit platforms, a size_t is always less than 1ULL << 48). + */ + uint64_t x = (uint64_t)x_orig; + /* + * If we can guarantee no overflow, multiply first before shifting, to + * preserve some precision. Otherwise, shift first and then multiply. + * In the latter case, we only lose the low 16 bits of a 48-bit number, + * so we're still accurate to within 1/2**32. + */ + if (x < (1ULL << 48)) { + return (size_t)((x * frac) >> 16); + } else { + return (size_t)((x >> 16) * (uint64_t)frac); + } +} + +/* + * Returns true on error. Otherwise, returns false and updates *ptr to point to + * the first character not parsed (because it wasn't a digit). + */ +bool fxp_parse(fxp_t *a, const char *ptr, char **end); +void fxp_print(fxp_t a, char buf[FXP_BUF_SIZE]); + +#endif /* JEMALLOC_INTERNAL_FXP_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/hash.h b/contrib/jemalloc/include/jemalloc/internal/hash.h --- a/contrib/jemalloc/include/jemalloc/internal/hash.h +++ b/contrib/jemalloc/include/jemalloc/internal/hash.h @@ -104,8 +104,8 @@ uint32_t k1 = 0; switch (len & 3) { - case 3: k1 ^= tail[2] << 16; JEMALLOC_FALLTHROUGH - case 2: k1 ^= tail[1] << 8; JEMALLOC_FALLTHROUGH + case 3: k1 ^= tail[2] << 16; JEMALLOC_FALLTHROUGH; + case 2: k1 ^= tail[1] << 8; JEMALLOC_FALLTHROUGH; case 1: k1 ^= tail[0]; k1 *= c1; k1 = hash_rotl_32(k1, 15); k1 *= c2; h1 ^= k1; } @@ -177,29 +177,29 @@ uint32_t k4 = 0; switch (len & 15) { - case 15: k4 ^= tail[14] << 16; JEMALLOC_FALLTHROUGH - case 14: k4 ^= tail[13] << 8; JEMALLOC_FALLTHROUGH + case 15: k4 ^= tail[14] << 16; JEMALLOC_FALLTHROUGH; + case 14: k4 ^= tail[13] << 8; JEMALLOC_FALLTHROUGH; case 13: k4 ^= tail[12] << 0; k4 *= c4; k4 = hash_rotl_32(k4, 18); k4 *= c1; h4 ^= k4; - JEMALLOC_FALLTHROUGH - case 12: k3 ^= tail[11] << 24; JEMALLOC_FALLTHROUGH - case 11: k3 ^= tail[10] << 16; JEMALLOC_FALLTHROUGH - case 10: k3 ^= tail[ 9] << 8; JEMALLOC_FALLTHROUGH + JEMALLOC_FALLTHROUGH; + case 12: k3 ^= (uint32_t) tail[11] << 24; JEMALLOC_FALLTHROUGH; + case 11: k3 ^= tail[10] << 16; JEMALLOC_FALLTHROUGH; + case 10: k3 ^= tail[ 9] << 8; JEMALLOC_FALLTHROUGH; case 9: k3 ^= tail[ 8] << 0; - k3 *= c3; k3 = hash_rotl_32(k3, 17); k3 *= c4; h3 ^= k3; - JEMALLOC_FALLTHROUGH - case 8: k2 ^= tail[ 7] << 24; JEMALLOC_FALLTHROUGH - case 7: k2 ^= tail[ 6] << 16; JEMALLOC_FALLTHROUGH - case 6: k2 ^= tail[ 5] << 8; JEMALLOC_FALLTHROUGH + k3 *= c3; k3 = hash_rotl_32(k3, 17); k3 *= c4; h3 ^= k3; + JEMALLOC_FALLTHROUGH; + case 8: k2 ^= (uint32_t) tail[ 7] << 24; JEMALLOC_FALLTHROUGH; + case 7: k2 ^= tail[ 6] << 16; JEMALLOC_FALLTHROUGH; + case 6: k2 ^= tail[ 5] << 8; JEMALLOC_FALLTHROUGH; case 5: k2 ^= tail[ 4] << 0; k2 *= c2; k2 = hash_rotl_32(k2, 16); k2 *= c3; h2 ^= k2; - JEMALLOC_FALLTHROUGH - case 4: k1 ^= tail[ 3] << 24; JEMALLOC_FALLTHROUGH - case 3: k1 ^= tail[ 2] << 16; JEMALLOC_FALLTHROUGH - case 2: k1 ^= tail[ 1] << 8; JEMALLOC_FALLTHROUGH + JEMALLOC_FALLTHROUGH; + case 4: k1 ^= (uint32_t) tail[ 3] << 24; JEMALLOC_FALLTHROUGH; + case 3: k1 ^= tail[ 2] << 16; JEMALLOC_FALLTHROUGH; + case 2: k1 ^= tail[ 1] << 8; JEMALLOC_FALLTHROUGH; case 1: k1 ^= tail[ 0] << 0; k1 *= c1; k1 = hash_rotl_32(k1, 15); k1 *= c2; h1 ^= k1; - JEMALLOC_FALLTHROUGH + break; } } @@ -261,24 +261,25 @@ uint64_t k2 = 0; switch (len & 15) { - case 15: k2 ^= ((uint64_t)(tail[14])) << 48; JEMALLOC_FALLTHROUGH - case 14: k2 ^= ((uint64_t)(tail[13])) << 40; JEMALLOC_FALLTHROUGH - case 13: k2 ^= ((uint64_t)(tail[12])) << 32; JEMALLOC_FALLTHROUGH - case 12: k2 ^= ((uint64_t)(tail[11])) << 24; JEMALLOC_FALLTHROUGH - case 11: k2 ^= ((uint64_t)(tail[10])) << 16; JEMALLOC_FALLTHROUGH - case 10: k2 ^= ((uint64_t)(tail[ 9])) << 8; JEMALLOC_FALLTHROUGH + case 15: k2 ^= ((uint64_t)(tail[14])) << 48; JEMALLOC_FALLTHROUGH; + case 14: k2 ^= ((uint64_t)(tail[13])) << 40; JEMALLOC_FALLTHROUGH; + case 13: k2 ^= ((uint64_t)(tail[12])) << 32; JEMALLOC_FALLTHROUGH; + case 12: k2 ^= ((uint64_t)(tail[11])) << 24; JEMALLOC_FALLTHROUGH; + case 11: k2 ^= ((uint64_t)(tail[10])) << 16; JEMALLOC_FALLTHROUGH; + case 10: k2 ^= ((uint64_t)(tail[ 9])) << 8; JEMALLOC_FALLTHROUGH; case 9: k2 ^= ((uint64_t)(tail[ 8])) << 0; k2 *= c2; k2 = hash_rotl_64(k2, 33); k2 *= c1; h2 ^= k2; - JEMALLOC_FALLTHROUGH - case 8: k1 ^= ((uint64_t)(tail[ 7])) << 56; JEMALLOC_FALLTHROUGH - case 7: k1 ^= ((uint64_t)(tail[ 6])) << 48; JEMALLOC_FALLTHROUGH - case 6: k1 ^= ((uint64_t)(tail[ 5])) << 40; JEMALLOC_FALLTHROUGH - case 5: k1 ^= ((uint64_t)(tail[ 4])) << 32; JEMALLOC_FALLTHROUGH - case 4: k1 ^= ((uint64_t)(tail[ 3])) << 24; JEMALLOC_FALLTHROUGH - case 3: k1 ^= ((uint64_t)(tail[ 2])) << 16; JEMALLOC_FALLTHROUGH - case 2: k1 ^= ((uint64_t)(tail[ 1])) << 8; JEMALLOC_FALLTHROUGH + JEMALLOC_FALLTHROUGH; + case 8: k1 ^= ((uint64_t)(tail[ 7])) << 56; JEMALLOC_FALLTHROUGH; + case 7: k1 ^= ((uint64_t)(tail[ 6])) << 48; JEMALLOC_FALLTHROUGH; + case 6: k1 ^= ((uint64_t)(tail[ 5])) << 40; JEMALLOC_FALLTHROUGH; + case 5: k1 ^= ((uint64_t)(tail[ 4])) << 32; JEMALLOC_FALLTHROUGH; + case 4: k1 ^= ((uint64_t)(tail[ 3])) << 24; JEMALLOC_FALLTHROUGH; + case 3: k1 ^= ((uint64_t)(tail[ 2])) << 16; JEMALLOC_FALLTHROUGH; + case 2: k1 ^= ((uint64_t)(tail[ 1])) << 8; JEMALLOC_FALLTHROUGH; case 1: k1 ^= ((uint64_t)(tail[ 0])) << 0; k1 *= c1; k1 = hash_rotl_64(k1, 31); k1 *= c2; h1 ^= k1; + break; } } diff --git a/contrib/jemalloc/include/jemalloc/internal/hpa.h b/contrib/jemalloc/include/jemalloc/internal/hpa.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/hpa.h @@ -0,0 +1,182 @@ +#ifndef JEMALLOC_INTERNAL_HPA_H +#define JEMALLOC_INTERNAL_HPA_H + +#include "jemalloc/internal/exp_grow.h" +#include "jemalloc/internal/hpa_hooks.h" +#include "jemalloc/internal/hpa_opts.h" +#include "jemalloc/internal/pai.h" +#include "jemalloc/internal/psset.h" + +typedef struct hpa_central_s hpa_central_t; +struct hpa_central_s { + /* + * The mutex guarding most of the operations on the central data + * structure. + */ + malloc_mutex_t mtx; + /* + * Guards expansion of eden. We separate this from the regular mutex so + * that cheaper operations can still continue while we're doing the OS + * call. + */ + malloc_mutex_t grow_mtx; + /* + * Either NULL (if empty), or some integer multiple of a + * hugepage-aligned number of hugepages. We carve them off one at a + * time to satisfy new pageslab requests. + * + * Guarded by grow_mtx. + */ + void *eden; + size_t eden_len; + /* Source for metadata. */ + base_t *base; + /* Number of grow operations done on this hpa_central_t. */ + uint64_t age_counter; + + /* The HPA hooks. */ + hpa_hooks_t hooks; +}; + +typedef struct hpa_shard_nonderived_stats_s hpa_shard_nonderived_stats_t; +struct hpa_shard_nonderived_stats_s { + /* + * The number of times we've purged within a hugepage. + * + * Guarded by mtx. + */ + uint64_t npurge_passes; + /* + * The number of individual purge calls we perform (which should always + * be bigger than npurge_passes, since each pass purges at least one + * extent within a hugepage. + * + * Guarded by mtx. + */ + uint64_t npurges; + + /* + * The number of times we've hugified a pageslab. + * + * Guarded by mtx. + */ + uint64_t nhugifies; + /* + * The number of times we've dehugified a pageslab. + * + * Guarded by mtx. + */ + uint64_t ndehugifies; +}; + +/* Completely derived; only used by CTL. */ +typedef struct hpa_shard_stats_s hpa_shard_stats_t; +struct hpa_shard_stats_s { + psset_stats_t psset_stats; + hpa_shard_nonderived_stats_t nonderived_stats; +}; + +typedef struct hpa_shard_s hpa_shard_t; +struct hpa_shard_s { + /* + * pai must be the first member; we cast from a pointer to it to a + * pointer to the hpa_shard_t. + */ + pai_t pai; + + /* The central allocator we get our hugepages from. */ + hpa_central_t *central; + /* Protects most of this shard's state. */ + malloc_mutex_t mtx; + /* + * Guards the shard's access to the central allocator (preventing + * multiple threads operating on this shard from accessing the central + * allocator). + */ + malloc_mutex_t grow_mtx; + /* The base metadata allocator. */ + base_t *base; + + /* + * This edata cache is the one we use when allocating a small extent + * from a pageslab. The pageslab itself comes from the centralized + * allocator, and so will use its edata_cache. + */ + edata_cache_fast_t ecf; + + psset_t psset; + + /* + * How many grow operations have occurred. + * + * Guarded by grow_mtx. + */ + uint64_t age_counter; + + /* The arena ind we're associated with. */ + unsigned ind; + + /* + * Our emap. This is just a cache of the emap pointer in the associated + * hpa_central. + */ + emap_t *emap; + + /* The configuration choices for this hpa shard. */ + hpa_shard_opts_t opts; + + /* + * How many pages have we started but not yet finished purging in this + * hpa shard. + */ + size_t npending_purge; + + /* + * Those stats which are copied directly into the CTL-centric hpa shard + * stats. + */ + hpa_shard_nonderived_stats_t stats; + + /* + * Last time we performed purge on this shard. + */ + nstime_t last_purge; +}; + +/* + * Whether or not the HPA can be used given the current configuration. This is + * is not necessarily a guarantee that it backs its allocations by hugepages, + * just that it can function properly given the system it's running on. + */ +bool hpa_supported(); +bool hpa_central_init(hpa_central_t *central, base_t *base, const hpa_hooks_t *hooks); +bool hpa_shard_init(hpa_shard_t *shard, hpa_central_t *central, emap_t *emap, + base_t *base, edata_cache_t *edata_cache, unsigned ind, + const hpa_shard_opts_t *opts); + +void hpa_shard_stats_accum(hpa_shard_stats_t *dst, hpa_shard_stats_t *src); +void hpa_shard_stats_merge(tsdn_t *tsdn, hpa_shard_t *shard, + hpa_shard_stats_t *dst); + +/* + * Notify the shard that we won't use it for allocations much longer. Due to + * the possibility of races, we don't actually prevent allocations; just flush + * and disable the embedded edata_cache_small. + */ +void hpa_shard_disable(tsdn_t *tsdn, hpa_shard_t *shard); +void hpa_shard_destroy(tsdn_t *tsdn, hpa_shard_t *shard); + +void hpa_shard_set_deferral_allowed(tsdn_t *tsdn, hpa_shard_t *shard, + bool deferral_allowed); +void hpa_shard_do_deferred_work(tsdn_t *tsdn, hpa_shard_t *shard); + +/* + * We share the fork ordering with the PA and arena prefork handling; that's why + * these are 3 and 4 rather than 0 and 1. + */ +void hpa_shard_prefork3(tsdn_t *tsdn, hpa_shard_t *shard); +void hpa_shard_prefork4(tsdn_t *tsdn, hpa_shard_t *shard); +void hpa_shard_postfork_parent(tsdn_t *tsdn, hpa_shard_t *shard); +void hpa_shard_postfork_child(tsdn_t *tsdn, hpa_shard_t *shard); + +#endif /* JEMALLOC_INTERNAL_HPA_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/hpa_hooks.h b/contrib/jemalloc/include/jemalloc/internal/hpa_hooks.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/hpa_hooks.h @@ -0,0 +1,17 @@ +#ifndef JEMALLOC_INTERNAL_HPA_HOOKS_H +#define JEMALLOC_INTERNAL_HPA_HOOKS_H + +typedef struct hpa_hooks_s hpa_hooks_t; +struct hpa_hooks_s { + void *(*map)(size_t size); + void (*unmap)(void *ptr, size_t size); + void (*purge)(void *ptr, size_t size); + void (*hugify)(void *ptr, size_t size); + void (*dehugify)(void *ptr, size_t size); + void (*curtime)(nstime_t *r_time, bool first_reading); + uint64_t (*ms_since)(nstime_t *r_time); +}; + +extern hpa_hooks_t hpa_hooks_default; + +#endif /* JEMALLOC_INTERNAL_HPA_HOOKS_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/hpa_opts.h b/contrib/jemalloc/include/jemalloc/internal/hpa_opts.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/hpa_opts.h @@ -0,0 +1,74 @@ +#ifndef JEMALLOC_INTERNAL_HPA_OPTS_H +#define JEMALLOC_INTERNAL_HPA_OPTS_H + +#include "jemalloc/internal/fxp.h" + +/* + * This file is morally part of hpa.h, but is split out for header-ordering + * reasons. + */ + +typedef struct hpa_shard_opts_s hpa_shard_opts_t; +struct hpa_shard_opts_s { + /* + * The largest size we'll allocate out of the shard. For those + * allocations refused, the caller (in practice, the PA module) will + * fall back to the more general (for now) PAC, which can always handle + * any allocation request. + */ + size_t slab_max_alloc; + + /* + * When the number of active bytes in a hugepage is >= + * hugification_threshold, we force hugify it. + */ + size_t hugification_threshold; + + /* + * The HPA purges whenever the number of pages exceeds dirty_mult * + * active_pages. This may be set to (fxp_t)-1 to disable purging. + */ + fxp_t dirty_mult; + + /* + * Whether or not the PAI methods are allowed to defer work to a + * subsequent hpa_shard_do_deferred_work() call. Practically, this + * corresponds to background threads being enabled. We track this + * ourselves for encapsulation purposes. + */ + bool deferral_allowed; + + /* + * How long a hugepage has to be a hugification candidate before it will + * actually get hugified. + */ + uint64_t hugify_delay_ms; + + /* + * Minimum amount of time between purges. + */ + uint64_t min_purge_interval_ms; +}; + +#define HPA_SHARD_OPTS_DEFAULT { \ + /* slab_max_alloc */ \ + 64 * 1024, \ + /* hugification_threshold */ \ + HUGEPAGE * 95 / 100, \ + /* dirty_mult */ \ + FXP_INIT_PERCENT(25), \ + /* \ + * deferral_allowed \ + * \ + * Really, this is always set by the arena during creation \ + * or by an hpa_shard_set_deferral_allowed call, so the value \ + * we put here doesn't matter. \ + */ \ + false, \ + /* hugify_delay_ms */ \ + 10 * 1000, \ + /* min_purge_interval_ms */ \ + 5 * 1000 \ +} + +#endif /* JEMALLOC_INTERNAL_HPA_OPTS_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/hpdata.h b/contrib/jemalloc/include/jemalloc/internal/hpdata.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/hpdata.h @@ -0,0 +1,413 @@ +#ifndef JEMALLOC_INTERNAL_HPDATA_H +#define JEMALLOC_INTERNAL_HPDATA_H + +#include "jemalloc/internal/fb.h" +#include "jemalloc/internal/ph.h" +#include "jemalloc/internal/ql.h" +#include "jemalloc/internal/typed_list.h" + +/* + * The metadata representation we use for extents in hugepages. While the PAC + * uses the edata_t to represent both active and inactive extents, the HP only + * uses the edata_t for active ones; instead, inactive extent state is tracked + * within hpdata associated with the enclosing hugepage-sized, hugepage-aligned + * region of virtual address space. + * + * An hpdata need not be "truly" backed by a hugepage (which is not necessarily + * an observable property of any given region of address space). It's just + * hugepage-sized and hugepage-aligned; it's *potentially* huge. + */ +typedef struct hpdata_s hpdata_t; +ph_structs(hpdata_age_heap, hpdata_t); +struct hpdata_s { + /* + * We likewise follow the edata convention of mangling names and forcing + * the use of accessors -- this lets us add some consistency checks on + * access. + */ + + /* + * The address of the hugepage in question. This can't be named h_addr, + * since that conflicts with a macro defined in Windows headers. + */ + void *h_address; + /* Its age (measured in psset operations). */ + uint64_t h_age; + /* Whether or not we think the hugepage is mapped that way by the OS. */ + bool h_huge; + + /* + * For some properties, we keep parallel sets of bools; h_foo_allowed + * and h_in_psset_foo_container. This is a decoupling mechanism to + * avoid bothering the hpa (which manages policies) from the psset + * (which is the mechanism used to enforce those policies). This allows + * all the container management logic to live in one place, without the + * HPA needing to know or care how that happens. + */ + + /* + * Whether or not the hpdata is allowed to be used to serve allocations, + * and whether or not the psset is currently tracking it as such. + */ + bool h_alloc_allowed; + bool h_in_psset_alloc_container; + + /* + * The same, but with purging. There's no corresponding + * h_in_psset_purge_container, because the psset (currently) always + * removes hpdatas from their containers during updates (to implement + * LRU for purging). + */ + bool h_purge_allowed; + + /* And with hugifying. */ + bool h_hugify_allowed; + /* When we became a hugification candidate. */ + nstime_t h_time_hugify_allowed; + bool h_in_psset_hugify_container; + + /* Whether or not a purge or hugify is currently happening. */ + bool h_mid_purge; + bool h_mid_hugify; + + /* + * Whether or not the hpdata is being updated in the psset (i.e. if + * there has been a psset_update_begin call issued without a matching + * psset_update_end call). Eventually this will expand to other types + * of updates. + */ + bool h_updating; + + /* Whether or not the hpdata is in a psset. */ + bool h_in_psset; + + union { + /* When nonempty (and also nonfull), used by the psset bins. */ + hpdata_age_heap_link_t age_link; + /* + * When empty (or not corresponding to any hugepage), list + * linkage. + */ + ql_elm(hpdata_t) ql_link_empty; + }; + + /* + * Linkage for the psset to track candidates for purging and hugifying. + */ + ql_elm(hpdata_t) ql_link_purge; + ql_elm(hpdata_t) ql_link_hugify; + + /* The length of the largest contiguous sequence of inactive pages. */ + size_t h_longest_free_range; + + /* Number of active pages. */ + size_t h_nactive; + + /* A bitmap with bits set in the active pages. */ + fb_group_t active_pages[FB_NGROUPS(HUGEPAGE_PAGES)]; + + /* + * Number of dirty or active pages, and a bitmap tracking them. One + * way to think of this is as which pages are dirty from the OS's + * perspective. + */ + size_t h_ntouched; + + /* The touched pages (using the same definition as above). */ + fb_group_t touched_pages[FB_NGROUPS(HUGEPAGE_PAGES)]; +}; + +TYPED_LIST(hpdata_empty_list, hpdata_t, ql_link_empty) +TYPED_LIST(hpdata_purge_list, hpdata_t, ql_link_purge) +TYPED_LIST(hpdata_hugify_list, hpdata_t, ql_link_hugify) + +ph_proto(, hpdata_age_heap, hpdata_t); + +static inline void * +hpdata_addr_get(const hpdata_t *hpdata) { + return hpdata->h_address; +} + +static inline void +hpdata_addr_set(hpdata_t *hpdata, void *addr) { + assert(HUGEPAGE_ADDR2BASE(addr) == addr); + hpdata->h_address = addr; +} + +static inline uint64_t +hpdata_age_get(const hpdata_t *hpdata) { + return hpdata->h_age; +} + +static inline void +hpdata_age_set(hpdata_t *hpdata, uint64_t age) { + hpdata->h_age = age; +} + +static inline bool +hpdata_huge_get(const hpdata_t *hpdata) { + return hpdata->h_huge; +} + +static inline bool +hpdata_alloc_allowed_get(const hpdata_t *hpdata) { + return hpdata->h_alloc_allowed; +} + +static inline void +hpdata_alloc_allowed_set(hpdata_t *hpdata, bool alloc_allowed) { + hpdata->h_alloc_allowed = alloc_allowed; +} + +static inline bool +hpdata_in_psset_alloc_container_get(const hpdata_t *hpdata) { + return hpdata->h_in_psset_alloc_container; +} + +static inline void +hpdata_in_psset_alloc_container_set(hpdata_t *hpdata, bool in_container) { + assert(in_container != hpdata->h_in_psset_alloc_container); + hpdata->h_in_psset_alloc_container = in_container; +} + +static inline bool +hpdata_purge_allowed_get(const hpdata_t *hpdata) { + return hpdata->h_purge_allowed; +} + +static inline void +hpdata_purge_allowed_set(hpdata_t *hpdata, bool purge_allowed) { + assert(purge_allowed == false || !hpdata->h_mid_purge); + hpdata->h_purge_allowed = purge_allowed; +} + +static inline bool +hpdata_hugify_allowed_get(const hpdata_t *hpdata) { + return hpdata->h_hugify_allowed; +} + +static inline void +hpdata_allow_hugify(hpdata_t *hpdata, nstime_t now) { + assert(!hpdata->h_mid_hugify); + hpdata->h_hugify_allowed = true; + hpdata->h_time_hugify_allowed = now; +} + +static inline nstime_t +hpdata_time_hugify_allowed(hpdata_t *hpdata) { + return hpdata->h_time_hugify_allowed; +} + +static inline void +hpdata_disallow_hugify(hpdata_t *hpdata) { + hpdata->h_hugify_allowed = false; +} + +static inline bool +hpdata_in_psset_hugify_container_get(const hpdata_t *hpdata) { + return hpdata->h_in_psset_hugify_container; +} + +static inline void +hpdata_in_psset_hugify_container_set(hpdata_t *hpdata, bool in_container) { + assert(in_container != hpdata->h_in_psset_hugify_container); + hpdata->h_in_psset_hugify_container = in_container; +} + +static inline bool +hpdata_mid_purge_get(const hpdata_t *hpdata) { + return hpdata->h_mid_purge; +} + +static inline void +hpdata_mid_purge_set(hpdata_t *hpdata, bool mid_purge) { + assert(mid_purge != hpdata->h_mid_purge); + hpdata->h_mid_purge = mid_purge; +} + +static inline bool +hpdata_mid_hugify_get(const hpdata_t *hpdata) { + return hpdata->h_mid_hugify; +} + +static inline void +hpdata_mid_hugify_set(hpdata_t *hpdata, bool mid_hugify) { + assert(mid_hugify != hpdata->h_mid_hugify); + hpdata->h_mid_hugify = mid_hugify; +} + +static inline bool +hpdata_changing_state_get(const hpdata_t *hpdata) { + return hpdata->h_mid_purge || hpdata->h_mid_hugify; +} + + +static inline bool +hpdata_updating_get(const hpdata_t *hpdata) { + return hpdata->h_updating; +} + +static inline void +hpdata_updating_set(hpdata_t *hpdata, bool updating) { + assert(updating != hpdata->h_updating); + hpdata->h_updating = updating; +} + +static inline bool +hpdata_in_psset_get(const hpdata_t *hpdata) { + return hpdata->h_in_psset; +} + +static inline void +hpdata_in_psset_set(hpdata_t *hpdata, bool in_psset) { + assert(in_psset != hpdata->h_in_psset); + hpdata->h_in_psset = in_psset; +} + +static inline size_t +hpdata_longest_free_range_get(const hpdata_t *hpdata) { + return hpdata->h_longest_free_range; +} + +static inline void +hpdata_longest_free_range_set(hpdata_t *hpdata, size_t longest_free_range) { + assert(longest_free_range <= HUGEPAGE_PAGES); + hpdata->h_longest_free_range = longest_free_range; +} + +static inline size_t +hpdata_nactive_get(hpdata_t *hpdata) { + return hpdata->h_nactive; +} + +static inline size_t +hpdata_ntouched_get(hpdata_t *hpdata) { + return hpdata->h_ntouched; +} + +static inline size_t +hpdata_ndirty_get(hpdata_t *hpdata) { + return hpdata->h_ntouched - hpdata->h_nactive; +} + +static inline size_t +hpdata_nretained_get(hpdata_t *hpdata) { + return HUGEPAGE_PAGES - hpdata->h_ntouched; +} + +static inline void +hpdata_assert_empty(hpdata_t *hpdata) { + assert(fb_empty(hpdata->active_pages, HUGEPAGE_PAGES)); + assert(hpdata->h_nactive == 0); +} + +/* + * Only used in tests, and in hpdata_assert_consistent, below. Verifies some + * consistency properties of the hpdata (e.g. that cached counts of page stats + * match computed ones). + */ +static inline bool +hpdata_consistent(hpdata_t *hpdata) { + if(fb_urange_longest(hpdata->active_pages, HUGEPAGE_PAGES) + != hpdata_longest_free_range_get(hpdata)) { + return false; + } + if (fb_scount(hpdata->active_pages, HUGEPAGE_PAGES, 0, HUGEPAGE_PAGES) + != hpdata->h_nactive) { + return false; + } + if (fb_scount(hpdata->touched_pages, HUGEPAGE_PAGES, 0, HUGEPAGE_PAGES) + != hpdata->h_ntouched) { + return false; + } + if (hpdata->h_ntouched < hpdata->h_nactive) { + return false; + } + if (hpdata->h_huge && hpdata->h_ntouched != HUGEPAGE_PAGES) { + return false; + } + if (hpdata_changing_state_get(hpdata) + && ((hpdata->h_purge_allowed) || hpdata->h_hugify_allowed)) { + return false; + } + if (hpdata_hugify_allowed_get(hpdata) + != hpdata_in_psset_hugify_container_get(hpdata)) { + return false; + } + return true; +} + +static inline void +hpdata_assert_consistent(hpdata_t *hpdata) { + assert(hpdata_consistent(hpdata)); +} + +static inline bool +hpdata_empty(hpdata_t *hpdata) { + return hpdata->h_nactive == 0; +} + +static inline bool +hpdata_full(hpdata_t *hpdata) { + return hpdata->h_nactive == HUGEPAGE_PAGES; +} + +void hpdata_init(hpdata_t *hpdata, void *addr, uint64_t age); + +/* + * Given an hpdata which can serve an allocation request, pick and reserve an + * offset within that allocation. + */ +void *hpdata_reserve_alloc(hpdata_t *hpdata, size_t sz); +void hpdata_unreserve(hpdata_t *hpdata, void *begin, size_t sz); + +/* + * The hpdata_purge_prepare_t allows grabbing the metadata required to purge + * subranges of a hugepage while holding a lock, drop the lock during the actual + * purging of them, and reacquire it to update the metadata again. + */ +typedef struct hpdata_purge_state_s hpdata_purge_state_t; +struct hpdata_purge_state_s { + size_t npurged; + size_t ndirty_to_purge; + fb_group_t to_purge[FB_NGROUPS(HUGEPAGE_PAGES)]; + size_t next_purge_search_begin; +}; + +/* + * Initializes purge state. The access to hpdata must be externally + * synchronized with other hpdata_* calls. + * + * You can tell whether or not a thread is purging or hugifying a given hpdata + * via hpdata_changing_state_get(hpdata). Racing hugification or purging + * operations aren't allowed. + * + * Once you begin purging, you have to follow through and call hpdata_purge_next + * until you're done, and then end. Allocating out of an hpdata undergoing + * purging is not allowed. + * + * Returns the number of dirty pages that will be purged. + */ +size_t hpdata_purge_begin(hpdata_t *hpdata, hpdata_purge_state_t *purge_state); + +/* + * If there are more extents to purge, sets *r_purge_addr and *r_purge_size to + * true, and returns true. Otherwise, returns false to indicate that we're + * done. + * + * This requires exclusive access to the purge state, but *not* to the hpdata. + * In particular, unreserve calls are allowed while purging (i.e. you can dalloc + * into one part of the hpdata while purging a different part). + */ +bool hpdata_purge_next(hpdata_t *hpdata, hpdata_purge_state_t *purge_state, + void **r_purge_addr, size_t *r_purge_size); +/* + * Updates the hpdata metadata after all purging is done. Needs external + * synchronization. + */ +void hpdata_purge_end(hpdata_t *hpdata, hpdata_purge_state_t *purge_state); + +void hpdata_hugify(hpdata_t *hpdata); +void hpdata_dehugify(hpdata_t *hpdata); + +#endif /* JEMALLOC_INTERNAL_HPDATA_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/inspect.h b/contrib/jemalloc/include/jemalloc/internal/inspect.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/inspect.h @@ -0,0 +1,40 @@ +#ifndef JEMALLOC_INTERNAL_INSPECT_H +#define JEMALLOC_INTERNAL_INSPECT_H + +/* + * This module contains the heap introspection capabilities. For now they are + * exposed purely through mallctl APIs in the experimental namespace, but this + * may change over time. + */ + +/* + * The following two structs are for experimental purposes. See + * experimental_utilization_query_ctl and + * experimental_utilization_batch_query_ctl in src/ctl.c. + */ +typedef struct inspect_extent_util_stats_s inspect_extent_util_stats_t; +struct inspect_extent_util_stats_s { + size_t nfree; + size_t nregs; + size_t size; +}; + +typedef struct inspect_extent_util_stats_verbose_s + inspect_extent_util_stats_verbose_t; + +struct inspect_extent_util_stats_verbose_s { + void *slabcur_addr; + size_t nfree; + size_t nregs; + size_t size; + size_t bin_nfree; + size_t bin_nregs; +}; + +void inspect_extent_util_stats_get(tsdn_t *tsdn, const void *ptr, + size_t *nfree, size_t *nregs, size_t *size); +void inspect_extent_util_stats_verbose_get(tsdn_t *tsdn, const void *ptr, + size_t *nfree, size_t *nregs, size_t *size, + size_t *bin_nfree, size_t *bin_nregs, void **slabcur_addr); + +#endif /* JEMALLOC_INTERNAL_INSPECT_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/jemalloc_internal_decls.h b/contrib/jemalloc/include/jemalloc/internal/jemalloc_internal_decls.h --- a/contrib/jemalloc/include/jemalloc/internal/jemalloc_internal_decls.h +++ b/contrib/jemalloc/include/jemalloc/internal/jemalloc_internal_decls.h @@ -8,6 +8,7 @@ #ifdef _WIN32 # include # include "msvc_compat/windows_extra.h" +# include "msvc_compat/strings.h" # ifdef _WIN64 # if LG_VADDR <= 32 # error Generate the headers using x64 vcargs @@ -34,8 +35,12 @@ # include # endif # include -# ifdef __FreeBSD__ +# if defined(__FreeBSD__) || defined(__DragonFly__) # include +# include +# if defined(__FreeBSD__) +# define cpu_set_t cpuset_t +# endif # endif # include # ifdef JEMALLOC_OS_UNFAIR_LOCK @@ -94,4 +99,13 @@ #endif #include +/* + * The Win32 midl compiler has #define small char; we don't use midl, but + * "small" is a nice identifier to have available when talking about size + * classes. + */ +#ifdef small +# undef small +#endif + #endif /* JEMALLOC_INTERNAL_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/jemalloc_internal_externs.h b/contrib/jemalloc/include/jemalloc/internal/jemalloc_internal_externs.h --- a/contrib/jemalloc/include/jemalloc/internal/jemalloc_internal_externs.h +++ b/contrib/jemalloc/include/jemalloc/internal/jemalloc_internal_externs.h @@ -2,7 +2,10 @@ #define JEMALLOC_INTERNAL_EXTERNS_H #include "jemalloc/internal/atomic.h" +#include "jemalloc/internal/hpa_opts.h" +#include "jemalloc/internal/sec_opts.h" #include "jemalloc/internal/tsd_types.h" +#include "jemalloc/internal/nstime.h" /* TSD checks this to set thread local slow state accordingly. */ extern bool malloc_slow; @@ -10,14 +13,30 @@ /* Run-time options. */ extern bool opt_abort; extern bool opt_abort_conf; +extern bool opt_trust_madvise; extern bool opt_confirm_conf; +extern bool opt_hpa; +extern hpa_shard_opts_t opt_hpa_opts; +extern sec_opts_t opt_hpa_sec_opts; + extern const char *opt_junk; extern bool opt_junk_alloc; extern bool opt_junk_free; +extern void (*junk_free_callback)(void *ptr, size_t size); +extern void (*junk_alloc_callback)(void *ptr, size_t size); extern bool opt_utrace; extern bool opt_xmalloc; +extern bool opt_experimental_infallible_new; extern bool opt_zero; extern unsigned opt_narenas; +extern zero_realloc_action_t opt_zero_realloc_action; +extern malloc_init_t malloc_init_state; +extern const char *zero_realloc_mode_names[]; +extern atomic_zu_t zero_realloc_count; +extern bool opt_cache_oblivious; + +/* Escape free-fastpath when ptr & mask == 0 (for sanitization purpose). */ +extern uintptr_t san_cache_bin_nonfast_mask; /* Number of CPUs. */ extern unsigned ncpus; @@ -41,17 +60,16 @@ void bootstrap_free(void *ptr); void arena_set(unsigned ind, arena_t *arena); unsigned narenas_total_get(void); -arena_t *arena_init(tsdn_t *tsdn, unsigned ind, extent_hooks_t *extent_hooks); -arena_tdata_t *arena_tdata_get_hard(tsd_t *tsd, unsigned ind); +arena_t *arena_init(tsdn_t *tsdn, unsigned ind, const arena_config_t *config); arena_t *arena_choose_hard(tsd_t *tsd, bool internal); -void arena_migrate(tsd_t *tsd, unsigned oldind, unsigned newind); +void arena_migrate(tsd_t *tsd, arena_t *oldarena, arena_t *newarena); void iarena_cleanup(tsd_t *tsd); void arena_cleanup(tsd_t *tsd); -void arenas_tdata_cleanup(tsd_t *tsd); +size_t batch_alloc(void **ptrs, size_t num, size_t size, int flags); void jemalloc_prefork(void); void jemalloc_postfork_parent(void); void jemalloc_postfork_child(void); -bool malloc_initialized(void); void je_sdallocx_noflags(void *ptr, size_t size); +void *malloc_default(size_t size); #endif /* JEMALLOC_INTERNAL_EXTERNS_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/jemalloc_internal_includes.h b/contrib/jemalloc/include/jemalloc/internal/jemalloc_internal_includes.h --- a/contrib/jemalloc/include/jemalloc/internal/jemalloc_internal_includes.h +++ b/contrib/jemalloc/include/jemalloc/internal/jemalloc_internal_includes.h @@ -10,7 +10,7 @@ * structs, externs, and inlines), and included each header file multiple times * in this file, picking out the portion we want on each pass using the * following #defines: - * JEMALLOC_H_TYPES : Preprocessor-defined constants and psuedo-opaque data + * JEMALLOC_H_TYPES : Preprocessor-defined constants and pseudo-opaque data * types. * JEMALLOC_H_STRUCTS : Data structures. * JEMALLOC_H_EXTERNS : Extern data declarations and function prototypes. @@ -40,8 +40,6 @@ /* TYPES */ /******************************************************************************/ -#include "jemalloc/internal/extent_types.h" -#include "jemalloc/internal/base_types.h" #include "jemalloc/internal/arena_types.h" #include "jemalloc/internal/tcache_types.h" #include "jemalloc/internal/prof_types.h" @@ -50,11 +48,8 @@ /* STRUCTS */ /******************************************************************************/ -#include "jemalloc/internal/arena_structs_a.h" -#include "jemalloc/internal/extent_structs.h" -#include "jemalloc/internal/base_structs.h" #include "jemalloc/internal/prof_structs.h" -#include "jemalloc/internal/arena_structs_b.h" +#include "jemalloc/internal/arena_structs.h" #include "jemalloc/internal/tcache_structs.h" #include "jemalloc/internal/background_thread_structs.h" @@ -63,8 +58,6 @@ /******************************************************************************/ #include "jemalloc/internal/jemalloc_internal_externs.h" -#include "jemalloc/internal/extent_externs.h" -#include "jemalloc/internal/base_externs.h" #include "jemalloc/internal/arena_externs.h" #include "jemalloc/internal/large_externs.h" #include "jemalloc/internal/tcache_externs.h" @@ -76,19 +69,16 @@ /******************************************************************************/ #include "jemalloc/internal/jemalloc_internal_inlines_a.h" -#include "jemalloc/internal/base_inlines.h" /* * Include portions of arena code interleaved with tcache code in order to * resolve circular dependencies. */ -#include "jemalloc/internal/prof_inlines_a.h" #include "jemalloc/internal/arena_inlines_a.h" -#include "jemalloc/internal/extent_inlines.h" #include "jemalloc/internal/jemalloc_internal_inlines_b.h" #include "jemalloc/internal/tcache_inlines.h" #include "jemalloc/internal/arena_inlines_b.h" #include "jemalloc/internal/jemalloc_internal_inlines_c.h" -#include "jemalloc/internal/prof_inlines_b.h" +#include "jemalloc/internal/prof_inlines.h" #include "jemalloc/internal/background_thread_inlines.h" #endif /* JEMALLOC_INTERNAL_INCLUDES_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/jemalloc_internal_inlines_a.h b/contrib/jemalloc/include/jemalloc/internal/jemalloc_internal_inlines_a.h --- a/contrib/jemalloc/include/jemalloc/internal/jemalloc_internal_inlines_a.h +++ b/contrib/jemalloc/include/jemalloc/internal/jemalloc_internal_inlines_a.h @@ -56,31 +56,6 @@ } } -static inline arena_tdata_t * -arena_tdata_get(tsd_t *tsd, unsigned ind, bool refresh_if_missing) { - arena_tdata_t *tdata; - arena_tdata_t *arenas_tdata = tsd_arenas_tdata_get(tsd); - - if (unlikely(arenas_tdata == NULL)) { - /* arenas_tdata hasn't been initialized yet. */ - return arena_tdata_get_hard(tsd, ind); - } - if (unlikely(ind >= tsd_narenas_tdata_get(tsd))) { - /* - * ind is invalid, cache is old (too small), or tdata to be - * initialized. - */ - return (refresh_if_missing ? arena_tdata_get_hard(tsd, ind) : - NULL); - } - - tdata = &arenas_tdata[ind]; - if (likely(tdata != NULL) || !refresh_if_missing) { - return tdata; - } - return arena_tdata_get_hard(tsd, ind); -} - static inline arena_t * arena_get(tsdn_t *tsdn, unsigned ind, bool init_if_missing) { arena_t *ret; @@ -90,36 +65,12 @@ ret = (arena_t *)atomic_load_p(&arenas[ind], ATOMIC_ACQUIRE); if (unlikely(ret == NULL)) { if (init_if_missing) { - ret = arena_init(tsdn, ind, - (extent_hooks_t *)&extent_hooks_default); + ret = arena_init(tsdn, ind, &arena_config_default); } } return ret; } -static inline ticker_t * -decay_ticker_get(tsd_t *tsd, unsigned ind) { - arena_tdata_t *tdata; - - tdata = arena_tdata_get(tsd, ind, true); - if (unlikely(tdata == NULL)) { - return NULL; - } - return &tdata->decay_ticker; -} - -JEMALLOC_ALWAYS_INLINE cache_bin_t * -tcache_small_bin_get(tcache_t *tcache, szind_t binind) { - assert(binind < SC_NBINS); - return &tcache->bins_small[binind]; -} - -JEMALLOC_ALWAYS_INLINE cache_bin_t * -tcache_large_bin_get(tcache_t *tcache, szind_t binind) { - assert(binind >= SC_NBINS &&binind < nhbins); - return &tcache->bins_large[binind - SC_NBINS]; -} - JEMALLOC_ALWAYS_INLINE bool tcache_available(tsd_t *tsd) { /* @@ -129,9 +80,9 @@ */ if (likely(tsd_tcache_enabled_get(tsd))) { /* Associated arena == NULL implies tcache init in progress. */ - assert(tsd_tcachep_get(tsd)->arena == NULL || - tcache_small_bin_get(tsd_tcachep_get(tsd), 0)->avail != - NULL); + if (config_debug && tsd_tcache_slowp_get(tsd)->arena != NULL) { + tcache_assert_initialized(tsd_tcachep_get(tsd)); + } return true; } @@ -147,28 +98,25 @@ return tsd_tcachep_get(tsd); } +JEMALLOC_ALWAYS_INLINE tcache_slow_t * +tcache_slow_get(tsd_t *tsd) { + if (!tcache_available(tsd)) { + return NULL; + } + + return tsd_tcache_slowp_get(tsd); +} + static inline void pre_reentrancy(tsd_t *tsd, arena_t *arena) { /* arena is the current context. Reentry from a0 is not allowed. */ assert(arena != arena_get(tsd_tsdn(tsd), 0, false)); - - bool fast = tsd_fast(tsd); - assert(tsd_reentrancy_level_get(tsd) < INT8_MAX); - ++*tsd_reentrancy_levelp_get(tsd); - if (fast) { - /* Prepare slow path for reentrancy. */ - tsd_slow_update(tsd); - assert(tsd_state_get(tsd) == tsd_state_nominal_slow); - } + tsd_pre_reentrancy_raw(tsd); } static inline void post_reentrancy(tsd_t *tsd) { - int8_t *reentrancy_level = tsd_reentrancy_levelp_get(tsd); - assert(*reentrancy_level > 0); - if (--*reentrancy_level == 0) { - tsd_slow_update(tsd); - } + tsd_post_reentrancy_raw(tsd); } #endif /* JEMALLOC_INTERNAL_INLINES_A_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/jemalloc_internal_inlines_b.h b/contrib/jemalloc/include/jemalloc/internal/jemalloc_internal_inlines_b.h --- a/contrib/jemalloc/include/jemalloc/internal/jemalloc_internal_inlines_b.h +++ b/contrib/jemalloc/include/jemalloc/internal/jemalloc_internal_inlines_b.h @@ -1,7 +1,31 @@ #ifndef JEMALLOC_INTERNAL_INLINES_B_H #define JEMALLOC_INTERNAL_INLINES_B_H -#include "jemalloc/internal/rtree.h" +#include "jemalloc/internal/extent.h" + +static inline void +percpu_arena_update(tsd_t *tsd, unsigned cpu) { + assert(have_percpu_arena); + arena_t *oldarena = tsd_arena_get(tsd); + assert(oldarena != NULL); + unsigned oldind = arena_ind_get(oldarena); + + if (oldind != cpu) { + unsigned newind = cpu; + arena_t *newarena = arena_get(tsd_tsdn(tsd), newind, true); + assert(newarena != NULL); + + /* Set new arena/tcache associations. */ + arena_migrate(tsd, oldarena, newarena); + tcache_t *tcache = tcache_get(tsd); + if (tcache != NULL) { + tcache_slow_t *tcache_slow = tsd_tcache_slowp_get(tsd); + tcache_arena_reassociate(tsd_tsdn(tsd), tcache_slow, + tcache, newarena); + } + } +} + /* Choose an arena based on a per-thread value. */ static inline arena_t * @@ -22,18 +46,19 @@ ret = arena_choose_hard(tsd, internal); assert(ret); if (tcache_available(tsd)) { - tcache_t *tcache = tcache_get(tsd); - if (tcache->arena != NULL) { - /* See comments in tcache_data_init().*/ - assert(tcache->arena == + tcache_slow_t *tcache_slow = tsd_tcache_slowp_get(tsd); + tcache_t *tcache = tsd_tcachep_get(tsd); + if (tcache_slow->arena != NULL) { + /* See comments in tsd_tcache_data_init().*/ + assert(tcache_slow->arena == arena_get(tsd_tsdn(tsd), 0, false)); - if (tcache->arena != ret) { + if (tcache_slow->arena != ret) { tcache_arena_reassociate(tsd_tsdn(tsd), - tcache, ret); + tcache_slow, tcache, ret); } } else { - tcache_arena_associate(tsd_tsdn(tsd), tcache, - ret); + tcache_arena_associate(tsd_tsdn(tsd), + tcache_slow, tcache, ret); } } } @@ -75,13 +100,4 @@ return (arena_ind_get(arena) < manual_arena_base); } -JEMALLOC_ALWAYS_INLINE extent_t * -iealloc(tsdn_t *tsdn, const void *ptr) { - rtree_ctx_t rtree_ctx_fallback; - rtree_ctx_t *rtree_ctx = tsdn_rtree_ctx(tsdn, &rtree_ctx_fallback); - - return rtree_extent_read(tsdn, &extents_rtree, rtree_ctx, - (uintptr_t)ptr, true); -} - #endif /* JEMALLOC_INTERNAL_INLINES_B_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/jemalloc_internal_inlines_c.h b/contrib/jemalloc/include/jemalloc/internal/jemalloc_internal_inlines_c.h --- a/contrib/jemalloc/include/jemalloc/internal/jemalloc_internal_inlines_c.h +++ b/contrib/jemalloc/include/jemalloc/internal/jemalloc_internal_inlines_c.h @@ -3,7 +3,9 @@ #include "jemalloc/internal/hook.h" #include "jemalloc/internal/jemalloc_internal_types.h" +#include "jemalloc/internal/log.h" #include "jemalloc/internal/sz.h" +#include "jemalloc/internal/thread_event.h" #include "jemalloc/internal/witness.h" /* @@ -101,8 +103,8 @@ } JEMALLOC_ALWAYS_INLINE void -idalloctm(tsdn_t *tsdn, void *ptr, tcache_t *tcache, alloc_ctx_t *alloc_ctx, - bool is_internal, bool slow_path) { +idalloctm(tsdn_t *tsdn, void *ptr, tcache_t *tcache, + emap_alloc_ctx_t *alloc_ctx, bool is_internal, bool slow_path) { assert(ptr != NULL); assert(!is_internal || tcache == NULL); assert(!is_internal || arena_is_auto(iaalloc(tsdn, ptr))); @@ -125,7 +127,7 @@ JEMALLOC_ALWAYS_INLINE void isdalloct(tsdn_t *tsdn, void *ptr, size_t size, tcache_t *tcache, - alloc_ctx_t *alloc_ctx, bool slow_path) { + emap_alloc_ctx_t *alloc_ctx, bool slow_path) { witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), WITNESS_RANK_CORE, 0); arena_sdalloc(tsdn, ptr, size, tcache, alloc_ctx, slow_path); @@ -219,4 +221,120 @@ newsize); } +JEMALLOC_ALWAYS_INLINE void +fastpath_success_finish(tsd_t *tsd, uint64_t allocated_after, + cache_bin_t *bin, void *ret) { + thread_allocated_set(tsd, allocated_after); + if (config_stats) { + bin->tstats.nrequests++; + } + + LOG("core.malloc.exit", "result: %p", ret); +} + +JEMALLOC_ALWAYS_INLINE bool +malloc_initialized(void) { + return (malloc_init_state == malloc_init_initialized); +} + +/* + * malloc() fastpath. Included here so that we can inline it into operator new; + * function call overhead there is non-negligible as a fraction of total CPU in + * allocation-heavy C++ programs. We take the fallback alloc to allow malloc + * (which can return NULL) to differ in its behavior from operator new (which + * can't). It matches the signature of malloc / operator new so that we can + * tail-call the fallback allocator, allowing us to avoid setting up the call + * frame in the common case. + * + * Fastpath assumes size <= SC_LOOKUP_MAXCLASS, and that we hit + * tcache. If either of these is false, we tail-call to the slowpath, + * malloc_default(). Tail-calling is used to avoid any caller-saved + * registers. + * + * fastpath supports ticker and profiling, both of which will also + * tail-call to the slowpath if they fire. + */ +JEMALLOC_ALWAYS_INLINE void * +imalloc_fastpath(size_t size, void *(fallback_alloc)(size_t)) { + LOG("core.malloc.entry", "size: %zu", size); + if (tsd_get_allocates() && unlikely(!malloc_initialized())) { + return fallback_alloc(size); + } + + tsd_t *tsd = tsd_get(false); + if (unlikely((size > SC_LOOKUP_MAXCLASS) || tsd == NULL)) { + return fallback_alloc(size); + } + /* + * The code below till the branch checking the next_event threshold may + * execute before malloc_init(), in which case the threshold is 0 to + * trigger slow path and initialization. + * + * Note that when uninitialized, only the fast-path variants of the sz / + * tsd facilities may be called. + */ + szind_t ind; + /* + * The thread_allocated counter in tsd serves as a general purpose + * accumulator for bytes of allocation to trigger different types of + * events. usize is always needed to advance thread_allocated, though + * it's not always needed in the core allocation logic. + */ + size_t usize; + sz_size2index_usize_fastpath(size, &ind, &usize); + /* Fast path relies on size being a bin. */ + assert(ind < SC_NBINS); + assert((SC_LOOKUP_MAXCLASS < SC_SMALL_MAXCLASS) && + (size <= SC_SMALL_MAXCLASS)); + + uint64_t allocated, threshold; + te_malloc_fastpath_ctx(tsd, &allocated, &threshold); + uint64_t allocated_after = allocated + usize; + /* + * The ind and usize might be uninitialized (or partially) before + * malloc_init(). The assertions check for: 1) full correctness (usize + * & ind) when initialized; and 2) guaranteed slow-path (threshold == 0) + * when !initialized. + */ + if (!malloc_initialized()) { + assert(threshold == 0); + } else { + assert(ind == sz_size2index(size)); + assert(usize > 0 && usize == sz_index2size(ind)); + } + /* + * Check for events and tsd non-nominal (fast_threshold will be set to + * 0) in a single branch. + */ + if (unlikely(allocated_after >= threshold)) { + return fallback_alloc(size); + } + assert(tsd_fast(tsd)); + + tcache_t *tcache = tsd_tcachep_get(tsd); + assert(tcache == tcache_get(tsd)); + cache_bin_t *bin = &tcache->bins[ind]; + bool tcache_success; + void *ret; + + /* + * We split up the code this way so that redundant low-water + * computation doesn't happen on the (more common) case in which we + * don't touch the low water mark. The compiler won't do this + * duplication on its own. + */ + ret = cache_bin_alloc_easy(bin, &tcache_success); + if (tcache_success) { + fastpath_success_finish(tsd, allocated_after, bin, ret); + return ret; + } + ret = cache_bin_alloc(bin, &tcache_success); + if (tcache_success) { + fastpath_success_finish(tsd, allocated_after, bin, ret); + return ret; + } + + return fallback_alloc(size); +} + #endif /* JEMALLOC_INTERNAL_INLINES_C_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/jemalloc_internal_macros.h b/contrib/jemalloc/include/jemalloc/internal/jemalloc_internal_macros.h --- a/contrib/jemalloc/include/jemalloc/internal/jemalloc_internal_macros.h +++ b/contrib/jemalloc/include/jemalloc/internal/jemalloc_internal_macros.h @@ -4,7 +4,11 @@ #ifdef JEMALLOC_DEBUG # define JEMALLOC_ALWAYS_INLINE static inline #else -# define JEMALLOC_ALWAYS_INLINE JEMALLOC_ATTR(always_inline) static inline +# ifdef _MSC_VER +# define JEMALLOC_ALWAYS_INLINE static __forceinline +# else +# define JEMALLOC_ALWAYS_INLINE JEMALLOC_ATTR(always_inline) static inline +# endif #endif #ifdef _MSC_VER # define inline _inline @@ -40,13 +44,6 @@ #define JEMALLOC_VA_ARGS_HEAD(head, ...) head #define JEMALLOC_VA_ARGS_TAIL(head, ...) __VA_ARGS__ -#if (defined(__GNUC__) || defined(__GNUG__)) && !defined(__clang__) \ - && defined(JEMALLOC_HAVE_ATTR) && (__GNUC__ >= 7) -#define JEMALLOC_FALLTHROUGH JEMALLOC_ATTR(fallthrough); -#else -#define JEMALLOC_FALLTHROUGH /* falls through */ -#endif - /* Diagnostic suppression macros */ #if defined(_MSC_VER) && !defined(__clang__) # define JEMALLOC_DIAGNOSTIC_PUSH __pragma(warning(push)) diff --git a/contrib/jemalloc/include/jemalloc/internal/jemalloc_internal_types.h b/contrib/jemalloc/include/jemalloc/internal/jemalloc_internal_types.h --- a/contrib/jemalloc/include/jemalloc/internal/jemalloc_internal_types.h +++ b/contrib/jemalloc/include/jemalloc/internal/jemalloc_internal_types.h @@ -3,15 +3,31 @@ #include "jemalloc/internal/quantum.h" -/* Page size index type. */ -typedef unsigned pszind_t; - -/* Size class index type. */ -typedef unsigned szind_t; - /* Processor / core id type. */ typedef int malloc_cpuid_t; +/* When realloc(non-null-ptr, 0) is called, what happens? */ +enum zero_realloc_action_e { + /* Realloc(ptr, 0) is free(ptr); return malloc(0); */ + zero_realloc_action_alloc = 0, + /* Realloc(ptr, 0) is free(ptr); */ + zero_realloc_action_free = 1, + /* Realloc(ptr, 0) aborts. */ + zero_realloc_action_abort = 2 +}; +typedef enum zero_realloc_action_e zero_realloc_action_t; + +/* Signature of write callback. */ +typedef void (write_cb_t)(void *, const char *); + +enum malloc_init_e { + malloc_init_uninitialized = 3, + malloc_init_a0_initialized = 2, + malloc_init_recursible = 1, + malloc_init_initialized = 0 /* Common case --> jnz. */ +}; +typedef enum malloc_init_e malloc_init_t; + /* * Flags bits: * diff --git a/contrib/jemalloc/include/jemalloc/internal/large_externs.h b/contrib/jemalloc/include/jemalloc/internal/large_externs.h --- a/contrib/jemalloc/include/jemalloc/internal/large_externs.h +++ b/contrib/jemalloc/include/jemalloc/internal/large_externs.h @@ -6,27 +6,19 @@ void *large_malloc(tsdn_t *tsdn, arena_t *arena, size_t usize, bool zero); void *large_palloc(tsdn_t *tsdn, arena_t *arena, size_t usize, size_t alignment, bool zero); -bool large_ralloc_no_move(tsdn_t *tsdn, extent_t *extent, size_t usize_min, +bool large_ralloc_no_move(tsdn_t *tsdn, edata_t *edata, size_t usize_min, size_t usize_max, bool zero); void *large_ralloc(tsdn_t *tsdn, arena_t *arena, void *ptr, size_t usize, size_t alignment, bool zero, tcache_t *tcache, hook_ralloc_args_t *hook_args); -typedef void (large_dalloc_junk_t)(void *, size_t); -extern large_dalloc_junk_t *JET_MUTABLE large_dalloc_junk; - -typedef void (large_dalloc_maybe_junk_t)(void *, size_t); -extern large_dalloc_maybe_junk_t *JET_MUTABLE large_dalloc_maybe_junk; - -void large_dalloc_prep_junked_locked(tsdn_t *tsdn, extent_t *extent); -void large_dalloc_finish(tsdn_t *tsdn, extent_t *extent); -void large_dalloc(tsdn_t *tsdn, extent_t *extent); -size_t large_salloc(tsdn_t *tsdn, const extent_t *extent); -prof_tctx_t *large_prof_tctx_get(tsdn_t *tsdn, const extent_t *extent); -void large_prof_tctx_set(tsdn_t *tsdn, extent_t *extent, prof_tctx_t *tctx); -void large_prof_tctx_reset(tsdn_t *tsdn, extent_t *extent); - -nstime_t large_prof_alloc_time_get(const extent_t *extent); -void large_prof_alloc_time_set(extent_t *extent, nstime_t time); +void large_dalloc_prep_locked(tsdn_t *tsdn, edata_t *edata); +void large_dalloc_finish(tsdn_t *tsdn, edata_t *edata); +void large_dalloc(tsdn_t *tsdn, edata_t *edata); +size_t large_salloc(tsdn_t *tsdn, const edata_t *edata); +void large_prof_info_get(tsd_t *tsd, edata_t *edata, prof_info_t *prof_info, + bool reset_recent); +void large_prof_tctx_reset(edata_t *edata); +void large_prof_info_set(edata_t *edata, prof_tctx_t *tctx, size_t size); #endif /* JEMALLOC_INTERNAL_LARGE_EXTERNS_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/lockedint.h b/contrib/jemalloc/include/jemalloc/internal/lockedint.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/lockedint.h @@ -0,0 +1,204 @@ +#ifndef JEMALLOC_INTERNAL_LOCKEDINT_H +#define JEMALLOC_INTERNAL_LOCKEDINT_H + +/* + * In those architectures that support 64-bit atomics, we use atomic updates for + * our 64-bit values. Otherwise, we use a plain uint64_t and synchronize + * externally. + */ + +typedef struct locked_u64_s locked_u64_t; +#ifdef JEMALLOC_ATOMIC_U64 +struct locked_u64_s { + atomic_u64_t val; +}; +#else +/* Must hold the associated mutex. */ +struct locked_u64_s { + uint64_t val; +}; +#endif + +typedef struct locked_zu_s locked_zu_t; +struct locked_zu_s { + atomic_zu_t val; +}; + +#ifndef JEMALLOC_ATOMIC_U64 +# define LOCKEDINT_MTX_DECLARE(name) malloc_mutex_t name; +# define LOCKEDINT_MTX_INIT(mu, name, rank, rank_mode) \ + malloc_mutex_init(&(mu), name, rank, rank_mode) +# define LOCKEDINT_MTX(mtx) (&(mtx)) +# define LOCKEDINT_MTX_LOCK(tsdn, mu) malloc_mutex_lock(tsdn, &(mu)) +# define LOCKEDINT_MTX_UNLOCK(tsdn, mu) malloc_mutex_unlock(tsdn, &(mu)) +# define LOCKEDINT_MTX_PREFORK(tsdn, mu) malloc_mutex_prefork(tsdn, &(mu)) +# define LOCKEDINT_MTX_POSTFORK_PARENT(tsdn, mu) \ + malloc_mutex_postfork_parent(tsdn, &(mu)) +# define LOCKEDINT_MTX_POSTFORK_CHILD(tsdn, mu) \ + malloc_mutex_postfork_child(tsdn, &(mu)) +#else +# define LOCKEDINT_MTX_DECLARE(name) +# define LOCKEDINT_MTX(mtx) NULL +# define LOCKEDINT_MTX_INIT(mu, name, rank, rank_mode) false +# define LOCKEDINT_MTX_LOCK(tsdn, mu) +# define LOCKEDINT_MTX_UNLOCK(tsdn, mu) +# define LOCKEDINT_MTX_PREFORK(tsdn, mu) +# define LOCKEDINT_MTX_POSTFORK_PARENT(tsdn, mu) +# define LOCKEDINT_MTX_POSTFORK_CHILD(tsdn, mu) +#endif + +#ifdef JEMALLOC_ATOMIC_U64 +# define LOCKEDINT_MTX_ASSERT_INTERNAL(tsdn, mtx) assert((mtx) == NULL) +#else +# define LOCKEDINT_MTX_ASSERT_INTERNAL(tsdn, mtx) \ + malloc_mutex_assert_owner(tsdn, (mtx)) +#endif + +static inline uint64_t +locked_read_u64(tsdn_t *tsdn, malloc_mutex_t *mtx, locked_u64_t *p) { + LOCKEDINT_MTX_ASSERT_INTERNAL(tsdn, mtx); +#ifdef JEMALLOC_ATOMIC_U64 + return atomic_load_u64(&p->val, ATOMIC_RELAXED); +#else + return p->val; +#endif +} + +static inline void +locked_inc_u64(tsdn_t *tsdn, malloc_mutex_t *mtx, locked_u64_t *p, + uint64_t x) { + LOCKEDINT_MTX_ASSERT_INTERNAL(tsdn, mtx); +#ifdef JEMALLOC_ATOMIC_U64 + atomic_fetch_add_u64(&p->val, x, ATOMIC_RELAXED); +#else + p->val += x; +#endif +} + +static inline void +locked_dec_u64(tsdn_t *tsdn, malloc_mutex_t *mtx, locked_u64_t *p, + uint64_t x) { + LOCKEDINT_MTX_ASSERT_INTERNAL(tsdn, mtx); +#ifdef JEMALLOC_ATOMIC_U64 + uint64_t r = atomic_fetch_sub_u64(&p->val, x, ATOMIC_RELAXED); + assert(r - x <= r); +#else + p->val -= x; + assert(p->val + x >= p->val); +#endif +} + +/* Increment and take modulus. Returns whether the modulo made any change. */ +static inline bool +locked_inc_mod_u64(tsdn_t *tsdn, malloc_mutex_t *mtx, locked_u64_t *p, + const uint64_t x, const uint64_t modulus) { + LOCKEDINT_MTX_ASSERT_INTERNAL(tsdn, mtx); + uint64_t before, after; + bool overflow; +#ifdef JEMALLOC_ATOMIC_U64 + before = atomic_load_u64(&p->val, ATOMIC_RELAXED); + do { + after = before + x; + assert(after >= before); + overflow = (after >= modulus); + if (overflow) { + after %= modulus; + } + } while (!atomic_compare_exchange_weak_u64(&p->val, &before, after, + ATOMIC_RELAXED, ATOMIC_RELAXED)); +#else + before = p->val; + after = before + x; + overflow = (after >= modulus); + if (overflow) { + after %= modulus; + } + p->val = after; +#endif + return overflow; +} + +/* + * Non-atomically sets *dst += src. *dst needs external synchronization. + * This lets us avoid the cost of a fetch_add when its unnecessary (note that + * the types here are atomic). + */ +static inline void +locked_inc_u64_unsynchronized(locked_u64_t *dst, uint64_t src) { +#ifdef JEMALLOC_ATOMIC_U64 + uint64_t cur_dst = atomic_load_u64(&dst->val, ATOMIC_RELAXED); + atomic_store_u64(&dst->val, src + cur_dst, ATOMIC_RELAXED); +#else + dst->val += src; +#endif +} + +static inline uint64_t +locked_read_u64_unsynchronized(locked_u64_t *p) { +#ifdef JEMALLOC_ATOMIC_U64 + return atomic_load_u64(&p->val, ATOMIC_RELAXED); +#else + return p->val; +#endif +} + +static inline void +locked_init_u64_unsynchronized(locked_u64_t *p, uint64_t x) { +#ifdef JEMALLOC_ATOMIC_U64 + atomic_store_u64(&p->val, x, ATOMIC_RELAXED); +#else + p->val = x; +#endif +} + +static inline size_t +locked_read_zu(tsdn_t *tsdn, malloc_mutex_t *mtx, locked_zu_t *p) { + LOCKEDINT_MTX_ASSERT_INTERNAL(tsdn, mtx); +#ifdef JEMALLOC_ATOMIC_U64 + return atomic_load_zu(&p->val, ATOMIC_RELAXED); +#else + return atomic_load_zu(&p->val, ATOMIC_RELAXED); +#endif +} + +static inline void +locked_inc_zu(tsdn_t *tsdn, malloc_mutex_t *mtx, locked_zu_t *p, + size_t x) { + LOCKEDINT_MTX_ASSERT_INTERNAL(tsdn, mtx); +#ifdef JEMALLOC_ATOMIC_U64 + atomic_fetch_add_zu(&p->val, x, ATOMIC_RELAXED); +#else + size_t cur = atomic_load_zu(&p->val, ATOMIC_RELAXED); + atomic_store_zu(&p->val, cur + x, ATOMIC_RELAXED); +#endif +} + +static inline void +locked_dec_zu(tsdn_t *tsdn, malloc_mutex_t *mtx, locked_zu_t *p, + size_t x) { + LOCKEDINT_MTX_ASSERT_INTERNAL(tsdn, mtx); +#ifdef JEMALLOC_ATOMIC_U64 + size_t r = atomic_fetch_sub_zu(&p->val, x, ATOMIC_RELAXED); + assert(r - x <= r); +#else + size_t cur = atomic_load_zu(&p->val, ATOMIC_RELAXED); + atomic_store_zu(&p->val, cur - x, ATOMIC_RELAXED); +#endif +} + +/* Like the _u64 variant, needs an externally synchronized *dst. */ +static inline void +locked_inc_zu_unsynchronized(locked_zu_t *dst, size_t src) { + size_t cur_dst = atomic_load_zu(&dst->val, ATOMIC_RELAXED); + atomic_store_zu(&dst->val, src + cur_dst, ATOMIC_RELAXED); +} + +/* + * Unlike the _u64 variant, this is safe to call unconditionally. + */ +static inline size_t +locked_read_atomic_zu(locked_zu_t *p) { + return atomic_load_zu(&p->val, ATOMIC_RELAXED); +} + +#endif /* JEMALLOC_INTERNAL_LOCKEDINT_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/malloc_io.h b/contrib/jemalloc/include/jemalloc/internal/malloc_io.h --- a/contrib/jemalloc/include/jemalloc/internal/malloc_io.h +++ b/contrib/jemalloc/include/jemalloc/internal/malloc_io.h @@ -1,6 +1,8 @@ #ifndef JEMALLOC_INTERNAL_MALLOC_IO_H #define JEMALLOC_INTERNAL_MALLOC_IO_H +#include "jemalloc/internal/jemalloc_internal_types.h" + #ifdef _WIN32 # ifdef _WIN64 # define FMT64_PREFIX "ll" @@ -40,6 +42,7 @@ */ #define MALLOC_PRINTF_BUFSIZE 4096 +write_cb_t wrtmessage; int buferror(int err, char *buf, size_t buflen); uintmax_t malloc_strtoumax(const char *restrict nptr, char **restrict endptr, int base); @@ -57,10 +60,10 @@ * The caller can set write_cb to null to choose to print with the * je_malloc_message hook. */ -void malloc_vcprintf(void (*write_cb)(void *, const char *), void *cbopaque, - const char *format, va_list ap); -void malloc_cprintf(void (*write_cb)(void *, const char *), void *cbopaque, - const char *format, ...) JEMALLOC_FORMAT_PRINTF(3, 4); +void malloc_vcprintf(write_cb_t *write_cb, void *cbopaque, const char *format, + va_list ap); +void malloc_cprintf(write_cb_t *write_cb, void *cbopaque, const char *format, + ...) JEMALLOC_FORMAT_PRINTF(3, 4); void malloc_printf(const char *format, ...) JEMALLOC_FORMAT_PRINTF(1, 2); static inline ssize_t diff --git a/contrib/jemalloc/include/jemalloc/internal/mpsc_queue.h b/contrib/jemalloc/include/jemalloc/internal/mpsc_queue.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/mpsc_queue.h @@ -0,0 +1,134 @@ +#ifndef JEMALLOC_INTERNAL_MPSC_QUEUE_H +#define JEMALLOC_INTERNAL_MPSC_QUEUE_H + +#include "jemalloc/internal/atomic.h" + +/* + * A concurrent implementation of a multi-producer, single-consumer queue. It + * supports three concurrent operations: + * - Push + * - Push batch + * - Pop batch + * + * These operations are all lock-free. + * + * The implementation is the simple two-stack queue built on a Treiber stack. + * It's not terribly efficient, but this isn't expected to go into anywhere with + * hot code. In fact, we don't really even need queue semantics in any + * anticipated use cases; we could get away with just the stack. But this way + * lets us frame the API in terms of the existing list types, which is a nice + * convenience. We can save on cache misses by introducing our own (parallel) + * single-linked list type here, and dropping FIFO semantics, if we need this to + * get faster. Since we're currently providing queue semantics though, we use + * the prev field in the link rather than the next field for Treiber-stack + * linkage, so that we can preserve order for bash-pushed lists (recall that the + * two-stack tricks reverses orders in the lock-free first stack). + */ + +#define mpsc_queue(a_type) \ +struct { \ + atomic_p_t tail; \ +} + +#define mpsc_queue_proto(a_attr, a_prefix, a_queue_type, a_type, \ + a_list_type) \ +/* Initialize a queue. */ \ +a_attr void \ +a_prefix##new(a_queue_type *queue); \ +/* Insert all items in src into the queue, clearing src. */ \ +a_attr void \ +a_prefix##push_batch(a_queue_type *queue, a_list_type *src); \ +/* Insert node into the queue. */ \ +a_attr void \ +a_prefix##push(a_queue_type *queue, a_type *node); \ +/* \ + * Pop all items in the queue into the list at dst. dst should already \ + * be initialized (and may contain existing items, which then remain \ + * in dst). \ + */ \ +a_attr void \ +a_prefix##pop_batch(a_queue_type *queue, a_list_type *dst); + +#define mpsc_queue_gen(a_attr, a_prefix, a_queue_type, a_type, \ + a_list_type, a_link) \ +a_attr void \ +a_prefix##new(a_queue_type *queue) { \ + atomic_store_p(&queue->tail, NULL, ATOMIC_RELAXED); \ +} \ +a_attr void \ +a_prefix##push_batch(a_queue_type *queue, a_list_type *src) { \ + /* \ + * Reuse the ql list next field as the Treiber stack next \ + * field. \ + */ \ + a_type *first = ql_first(src); \ + a_type *last = ql_last(src, a_link); \ + void* cur_tail = atomic_load_p(&queue->tail, ATOMIC_RELAXED); \ + do { \ + /* \ + * Note that this breaks the queue ring structure; \ + * it's not a ring any more! \ + */ \ + first->a_link.qre_prev = cur_tail; \ + /* \ + * Note: the upcoming CAS doesn't need an atomic; every \ + * push only needs to synchronize with the next pop, \ + * which we get from the release sequence rules. \ + */ \ + } while (!atomic_compare_exchange_weak_p(&queue->tail, \ + &cur_tail, last, ATOMIC_RELEASE, ATOMIC_RELAXED)); \ + ql_new(src); \ +} \ +a_attr void \ +a_prefix##push(a_queue_type *queue, a_type *node) { \ + ql_elm_new(node, a_link); \ + a_list_type list; \ + ql_new(&list); \ + ql_head_insert(&list, node, a_link); \ + a_prefix##push_batch(queue, &list); \ +} \ +a_attr void \ +a_prefix##pop_batch(a_queue_type *queue, a_list_type *dst) { \ + a_type *tail = atomic_load_p(&queue->tail, ATOMIC_RELAXED); \ + if (tail == NULL) { \ + /* \ + * In the common special case where there are no \ + * pending elements, bail early without a costly RMW. \ + */ \ + return; \ + } \ + tail = atomic_exchange_p(&queue->tail, NULL, ATOMIC_ACQUIRE); \ + /* \ + * It's a single-consumer queue, so if cur started non-NULL, \ + * it'd better stay non-NULL. \ + */ \ + assert(tail != NULL); \ + /* \ + * We iterate through the stack and both fix up the link \ + * structure (stack insertion broke the list requirement that \ + * the list be circularly linked). It's just as efficient at \ + * this point to make the queue a "real" queue, so do that as \ + * well. \ + * If this ever gets to be a hot spot, we can omit this fixup \ + * and make the queue a bag (i.e. not necessarily ordered), but \ + * that would mean jettisoning the existing list API as the \ + * batch pushing/popping interface. \ + */ \ + a_list_type reversed; \ + ql_new(&reversed); \ + while (tail != NULL) { \ + /* \ + * Pop an item off the stack, prepend it onto the list \ + * (reversing the order). Recall that we use the \ + * list prev field as the Treiber stack next field to \ + * preserve order of batch-pushed items when reversed. \ + */ \ + a_type *next = tail->a_link.qre_prev; \ + ql_elm_new(tail, a_link); \ + ql_head_insert(&reversed, tail, a_link); \ + tail = next; \ + } \ + ql_concat(dst, &reversed, a_link); \ +} + +#endif /* JEMALLOC_INTERNAL_MPSC_QUEUE_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/mutex.h b/contrib/jemalloc/include/jemalloc/internal/mutex.h --- a/contrib/jemalloc/include/jemalloc/internal/mutex.h +++ b/contrib/jemalloc/include/jemalloc/internal/mutex.h @@ -6,6 +6,8 @@ #include "jemalloc/internal/tsd.h" #include "jemalloc/internal/witness.h" +extern int64_t opt_mutex_max_spin; + typedef enum { /* Can only acquire one mutex of a given witness rank at a time. */ malloc_mutex_rank_exclusive, @@ -43,7 +45,7 @@ #else pthread_mutex_t lock; #endif - /* + /* * Hint flag to avoid exclusive cache line contention * during spin waiting */ @@ -67,12 +69,6 @@ #endif }; -/* - * Based on benchmark results, a fixed spin with this amount of retries works - * well for our critical sections. - */ -#define MALLOC_MUTEX_MAX_SPIN 250 - #ifdef _WIN32 # if _WIN32_WINNT >= 0x0600 # define MALLOC_MUTEX_LOCK(m) AcquireSRWLockExclusive(&(m)->lock) @@ -243,22 +239,25 @@ witness_assert_not_owner(tsdn_witness_tsdp_get(tsdn), &mutex->witness); } -/* Copy the prof data from mutex for processing. */ static inline void -malloc_mutex_prof_read(tsdn_t *tsdn, mutex_prof_data_t *data, - malloc_mutex_t *mutex) { - mutex_prof_data_t *source = &mutex->prof_data; - /* Can only read holding the mutex. */ - malloc_mutex_assert_owner(tsdn, mutex); - +malloc_mutex_prof_copy(mutex_prof_data_t *dst, mutex_prof_data_t *source) { /* * Not *really* allowed (we shouldn't be doing non-atomic loads of * atomic data), but the mutex protection makes this safe, and writing * a member-for-member copy is tedious for this situation. */ - *data = *source; + *dst = *source; /* n_wait_thds is not reported (modified w/o locking). */ - atomic_store_u32(&data->n_waiting_thds, 0, ATOMIC_RELAXED); + atomic_store_u32(&dst->n_waiting_thds, 0, ATOMIC_RELAXED); +} + +/* Copy the prof data from mutex for processing. */ +static inline void +malloc_mutex_prof_read(tsdn_t *tsdn, mutex_prof_data_t *data, + malloc_mutex_t *mutex) { + /* Can only read holding the mutex. */ + malloc_mutex_assert_owner(tsdn, mutex); + malloc_mutex_prof_copy(data, &mutex->prof_data); } static inline void @@ -283,4 +282,36 @@ data->n_lock_ops += source->n_lock_ops; } +/* Compare the prof data and update to the maximum. */ +static inline void +malloc_mutex_prof_max_update(tsdn_t *tsdn, mutex_prof_data_t *data, + malloc_mutex_t *mutex) { + mutex_prof_data_t *source = &mutex->prof_data; + /* Can only read holding the mutex. */ + malloc_mutex_assert_owner(tsdn, mutex); + + if (nstime_compare(&source->tot_wait_time, &data->tot_wait_time) > 0) { + nstime_copy(&data->tot_wait_time, &source->tot_wait_time); + } + if (nstime_compare(&source->max_wait_time, &data->max_wait_time) > 0) { + nstime_copy(&data->max_wait_time, &source->max_wait_time); + } + if (source->n_wait_times > data->n_wait_times) { + data->n_wait_times = source->n_wait_times; + } + if (source->n_spin_acquired > data->n_spin_acquired) { + data->n_spin_acquired = source->n_spin_acquired; + } + if (source->max_n_thds > data->max_n_thds) { + data->max_n_thds = source->max_n_thds; + } + if (source->n_owner_switches > data->n_owner_switches) { + data->n_owner_switches = source->n_owner_switches; + } + if (source->n_lock_ops > data->n_lock_ops) { + data->n_lock_ops = source->n_lock_ops; + } + /* n_wait_thds is not reported. */ +} + #endif /* JEMALLOC_INTERNAL_MUTEX_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/mutex_pool.h b/contrib/jemalloc/include/jemalloc/internal/mutex_pool.h deleted file mode 100644 --- a/contrib/jemalloc/include/jemalloc/internal/mutex_pool.h +++ /dev/null @@ -1,94 +0,0 @@ -#ifndef JEMALLOC_INTERNAL_MUTEX_POOL_H -#define JEMALLOC_INTERNAL_MUTEX_POOL_H - -#include "jemalloc/internal/hash.h" -#include "jemalloc/internal/mutex.h" -#include "jemalloc/internal/witness.h" - -/* We do mod reductions by this value, so it should be kept a power of 2. */ -#define MUTEX_POOL_SIZE 256 - -typedef struct mutex_pool_s mutex_pool_t; -struct mutex_pool_s { - malloc_mutex_t mutexes[MUTEX_POOL_SIZE]; -}; - -bool mutex_pool_init(mutex_pool_t *pool, const char *name, witness_rank_t rank); - -/* Internal helper - not meant to be called outside this module. */ -static inline malloc_mutex_t * -mutex_pool_mutex(mutex_pool_t *pool, uintptr_t key) { - size_t hash_result[2]; - hash(&key, sizeof(key), 0xd50dcc1b, hash_result); - return &pool->mutexes[hash_result[0] % MUTEX_POOL_SIZE]; -} - -static inline void -mutex_pool_assert_not_held(tsdn_t *tsdn, mutex_pool_t *pool) { - for (int i = 0; i < MUTEX_POOL_SIZE; i++) { - malloc_mutex_assert_not_owner(tsdn, &pool->mutexes[i]); - } -} - -/* - * Note that a mutex pool doesn't work exactly the way an embdedded mutex would. - * You're not allowed to acquire mutexes in the pool one at a time. You have to - * acquire all the mutexes you'll need in a single function call, and then - * release them all in a single function call. - */ - -static inline void -mutex_pool_lock(tsdn_t *tsdn, mutex_pool_t *pool, uintptr_t key) { - mutex_pool_assert_not_held(tsdn, pool); - - malloc_mutex_t *mutex = mutex_pool_mutex(pool, key); - malloc_mutex_lock(tsdn, mutex); -} - -static inline void -mutex_pool_unlock(tsdn_t *tsdn, mutex_pool_t *pool, uintptr_t key) { - malloc_mutex_t *mutex = mutex_pool_mutex(pool, key); - malloc_mutex_unlock(tsdn, mutex); - - mutex_pool_assert_not_held(tsdn, pool); -} - -static inline void -mutex_pool_lock2(tsdn_t *tsdn, mutex_pool_t *pool, uintptr_t key1, - uintptr_t key2) { - mutex_pool_assert_not_held(tsdn, pool); - - malloc_mutex_t *mutex1 = mutex_pool_mutex(pool, key1); - malloc_mutex_t *mutex2 = mutex_pool_mutex(pool, key2); - if ((uintptr_t)mutex1 < (uintptr_t)mutex2) { - malloc_mutex_lock(tsdn, mutex1); - malloc_mutex_lock(tsdn, mutex2); - } else if ((uintptr_t)mutex1 == (uintptr_t)mutex2) { - malloc_mutex_lock(tsdn, mutex1); - } else { - malloc_mutex_lock(tsdn, mutex2); - malloc_mutex_lock(tsdn, mutex1); - } -} - -static inline void -mutex_pool_unlock2(tsdn_t *tsdn, mutex_pool_t *pool, uintptr_t key1, - uintptr_t key2) { - malloc_mutex_t *mutex1 = mutex_pool_mutex(pool, key1); - malloc_mutex_t *mutex2 = mutex_pool_mutex(pool, key2); - if (mutex1 == mutex2) { - malloc_mutex_unlock(tsdn, mutex1); - } else { - malloc_mutex_unlock(tsdn, mutex1); - malloc_mutex_unlock(tsdn, mutex2); - } - - mutex_pool_assert_not_held(tsdn, pool); -} - -static inline void -mutex_pool_assert_owner(tsdn_t *tsdn, mutex_pool_t *pool, uintptr_t key) { - malloc_mutex_assert_owner(tsdn, mutex_pool_mutex(pool, key)); -} - -#endif /* JEMALLOC_INTERNAL_MUTEX_POOL_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/mutex_prof.h b/contrib/jemalloc/include/jemalloc/internal/mutex_prof.h --- a/contrib/jemalloc/include/jemalloc/internal/mutex_prof.h +++ b/contrib/jemalloc/include/jemalloc/internal/mutex_prof.h @@ -7,8 +7,14 @@ #define MUTEX_PROF_GLOBAL_MUTEXES \ OP(background_thread) \ + OP(max_per_bg_thd) \ OP(ctl) \ - OP(prof) + OP(prof) \ + OP(prof_thds_data) \ + OP(prof_dump) \ + OP(prof_recent_alloc) \ + OP(prof_recent_dump) \ + OP(prof_stats) typedef enum { #define OP(mtx) global_prof_mutex_##mtx, @@ -26,7 +32,10 @@ OP(decay_dirty) \ OP(decay_muzzy) \ OP(base) \ - OP(tcache_list) + OP(tcache_list) \ + OP(hpa_shard) \ + OP(hpa_shard_grow) \ + OP(hpa_sec) typedef enum { #define OP(mtx) arena_prof_mutex_##mtx, diff --git a/contrib/jemalloc/include/jemalloc/internal/nstime.h b/contrib/jemalloc/include/jemalloc/internal/nstime.h --- a/contrib/jemalloc/include/jemalloc/internal/nstime.h +++ b/contrib/jemalloc/include/jemalloc/internal/nstime.h @@ -3,12 +3,23 @@ /* Maximum supported number of seconds (~584 years). */ #define NSTIME_SEC_MAX KQU(18446744072) -#define NSTIME_ZERO_INITIALIZER {0} + +#define NSTIME_MAGIC ((uint32_t)0xb8a9ce37) +#ifdef JEMALLOC_DEBUG +# define NSTIME_ZERO_INITIALIZER {0, NSTIME_MAGIC} +#else +# define NSTIME_ZERO_INITIALIZER {0} +#endif typedef struct { uint64_t ns; +#ifdef JEMALLOC_DEBUG + uint32_t magic; /* Tracks if initialized. */ +#endif } nstime_t; +static const nstime_t nstime_zero = NSTIME_ZERO_INITIALIZER; + void nstime_init(nstime_t *time, uint64_t ns); void nstime_init2(nstime_t *time, uint64_t sec, uint64_t nsec); uint64_t nstime_ns(const nstime_t *time); @@ -24,11 +35,39 @@ void nstime_imultiply(nstime_t *time, uint64_t multiplier); void nstime_idivide(nstime_t *time, uint64_t divisor); uint64_t nstime_divide(const nstime_t *time, const nstime_t *divisor); +uint64_t nstime_ns_since(const nstime_t *past); typedef bool (nstime_monotonic_t)(void); extern nstime_monotonic_t *JET_MUTABLE nstime_monotonic; -typedef bool (nstime_update_t)(nstime_t *); +typedef void (nstime_update_t)(nstime_t *); extern nstime_update_t *JET_MUTABLE nstime_update; +typedef void (nstime_prof_update_t)(nstime_t *); +extern nstime_prof_update_t *JET_MUTABLE nstime_prof_update; + +void nstime_init_update(nstime_t *time); +void nstime_prof_init_update(nstime_t *time); + +enum prof_time_res_e { + prof_time_res_default = 0, + prof_time_res_high = 1 +}; +typedef enum prof_time_res_e prof_time_res_t; + +extern prof_time_res_t opt_prof_time_res; +extern const char *prof_time_res_mode_names[]; + +JEMALLOC_ALWAYS_INLINE void +nstime_init_zero(nstime_t *time) { + nstime_copy(time, &nstime_zero); +} + +JEMALLOC_ALWAYS_INLINE bool +nstime_equals_zero(nstime_t *time) { + int diff = nstime_compare(time, &nstime_zero); + assert(diff >= 0); + return diff == 0; +} + #endif /* JEMALLOC_INTERNAL_NSTIME_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/pa.h b/contrib/jemalloc/include/jemalloc/internal/pa.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/pa.h @@ -0,0 +1,243 @@ +#ifndef JEMALLOC_INTERNAL_PA_H +#define JEMALLOC_INTERNAL_PA_H + +#include "jemalloc/internal/base.h" +#include "jemalloc/internal/decay.h" +#include "jemalloc/internal/ecache.h" +#include "jemalloc/internal/edata_cache.h" +#include "jemalloc/internal/emap.h" +#include "jemalloc/internal/hpa.h" +#include "jemalloc/internal/lockedint.h" +#include "jemalloc/internal/pac.h" +#include "jemalloc/internal/pai.h" +#include "jemalloc/internal/sec.h" + +/* + * The page allocator; responsible for acquiring pages of memory for + * allocations. It picks the implementation of the page allocator interface + * (i.e. a pai_t) to handle a given page-level allocation request. For now, the + * only such implementation is the PAC code ("page allocator classic"), but + * others will be coming soon. + */ + +typedef struct pa_central_s pa_central_t; +struct pa_central_s { + hpa_central_t hpa; +}; + +/* + * The stats for a particular pa_shard. Because of the way the ctl module + * handles stats epoch data collection (it has its own arena_stats, and merges + * the stats from each arena into it), this needs to live in the arena_stats_t; + * hence we define it here and let the pa_shard have a pointer (rather than the + * more natural approach of just embedding it in the pa_shard itself). + * + * We follow the arena_stats_t approach of marking the derived fields. These + * are the ones that are not maintained on their own; instead, their values are + * derived during those stats merges. + */ +typedef struct pa_shard_stats_s pa_shard_stats_t; +struct pa_shard_stats_s { + /* Number of edata_t structs allocated by base, but not being used. */ + size_t edata_avail; /* Derived. */ + /* + * Stats specific to the PAC. For now, these are the only stats that + * exist, but there will eventually be other page allocators. Things + * like edata_avail make sense in a cross-PA sense, but things like + * npurges don't. + */ + pac_stats_t pac_stats; +}; + +/* + * The local allocator handle. Keeps the state necessary to satisfy page-sized + * allocations. + * + * The contents are mostly internal to the PA module. The key exception is that + * arena decay code is allowed to grab pointers to the dirty and muzzy ecaches + * decay_ts, for a couple of queries, passing them back to a PA function, or + * acquiring decay.mtx and looking at decay.purging. The reasoning is that, + * while PA decides what and how to purge, the arena code decides when and where + * (e.g. on what thread). It's allowed to use the presence of another purger to + * decide. + * (The background thread code also touches some other decay internals, but + * that's not fundamental; its' just an artifact of a partial refactoring, and + * its accesses could be straightforwardly moved inside the decay module). + */ +typedef struct pa_shard_s pa_shard_t; +struct pa_shard_s { + /* The central PA this shard is associated with. */ + pa_central_t *central; + + /* + * Number of pages in active extents. + * + * Synchronization: atomic. + */ + atomic_zu_t nactive; + + /* + * Whether or not we should prefer the hugepage allocator. Atomic since + * it may be concurrently modified by a thread setting extent hooks. + * Note that we still may do HPA operations in this arena; if use_hpa is + * changed from true to false, we'll free back to the hugepage allocator + * for those allocations. + */ + atomic_b_t use_hpa; + + /* + * If we never used the HPA to begin with, it wasn't initialized, and so + * we shouldn't try to e.g. acquire its mutexes during fork. This + * tracks that knowledge. + */ + bool ever_used_hpa; + + /* Allocates from a PAC. */ + pac_t pac; + + /* + * We place a small extent cache in front of the HPA, since we intend + * these configurations to use many fewer arenas, and therefore have a + * higher risk of hot locks. + */ + sec_t hpa_sec; + hpa_shard_t hpa_shard; + + /* The source of edata_t objects. */ + edata_cache_t edata_cache; + + unsigned ind; + + malloc_mutex_t *stats_mtx; + pa_shard_stats_t *stats; + + /* The emap this shard is tied to. */ + emap_t *emap; + + /* The base from which we get the ehooks and allocate metadat. */ + base_t *base; +}; + +static inline bool +pa_shard_dont_decay_muzzy(pa_shard_t *shard) { + return ecache_npages_get(&shard->pac.ecache_muzzy) == 0 && + pac_decay_ms_get(&shard->pac, extent_state_muzzy) <= 0; +} + +static inline ehooks_t * +pa_shard_ehooks_get(pa_shard_t *shard) { + return base_ehooks_get(shard->base); +} + +/* Returns true on error. */ +bool pa_central_init(pa_central_t *central, base_t *base, bool hpa, + hpa_hooks_t *hpa_hooks); + +/* Returns true on error. */ +bool pa_shard_init(tsdn_t *tsdn, pa_shard_t *shard, pa_central_t *central, + emap_t *emap, base_t *base, unsigned ind, pa_shard_stats_t *stats, + malloc_mutex_t *stats_mtx, nstime_t *cur_time, size_t oversize_threshold, + ssize_t dirty_decay_ms, ssize_t muzzy_decay_ms); + +/* + * This isn't exposed to users; we allow late enablement of the HPA shard so + * that we can boot without worrying about the HPA, then turn it on in a0. + */ +bool pa_shard_enable_hpa(tsdn_t *tsdn, pa_shard_t *shard, + const hpa_shard_opts_t *hpa_opts, const sec_opts_t *hpa_sec_opts); + +/* + * We stop using the HPA when custom extent hooks are installed, but still + * redirect deallocations to it. + */ +void pa_shard_disable_hpa(tsdn_t *tsdn, pa_shard_t *shard); + +/* + * This does the PA-specific parts of arena reset (i.e. freeing all active + * allocations). + */ +void pa_shard_reset(tsdn_t *tsdn, pa_shard_t *shard); + +/* + * Destroy all the remaining retained extents. Should only be called after + * decaying all active, dirty, and muzzy extents to the retained state, as the + * last step in destroying the shard. + */ +void pa_shard_destroy(tsdn_t *tsdn, pa_shard_t *shard); + +/* Gets an edata for the given allocation. */ +edata_t *pa_alloc(tsdn_t *tsdn, pa_shard_t *shard, size_t size, + size_t alignment, bool slab, szind_t szind, bool zero, bool guarded, + bool *deferred_work_generated); +/* Returns true on error, in which case nothing changed. */ +bool pa_expand(tsdn_t *tsdn, pa_shard_t *shard, edata_t *edata, size_t old_size, + size_t new_size, szind_t szind, bool zero, bool *deferred_work_generated); +/* + * The same. Sets *generated_dirty to true if we produced new dirty pages, and + * false otherwise. + */ +bool pa_shrink(tsdn_t *tsdn, pa_shard_t *shard, edata_t *edata, size_t old_size, + size_t new_size, szind_t szind, bool *deferred_work_generated); +/* + * Frees the given edata back to the pa. Sets *generated_dirty if we produced + * new dirty pages (well, we always set it for now; but this need not be the + * case). + * (We could make generated_dirty the return value of course, but this is more + * consistent with the shrink pathway and our error codes here). + */ +void pa_dalloc(tsdn_t *tsdn, pa_shard_t *shard, edata_t *edata, + bool *deferred_work_generated); +bool pa_decay_ms_set(tsdn_t *tsdn, pa_shard_t *shard, extent_state_t state, + ssize_t decay_ms, pac_purge_eagerness_t eagerness); +ssize_t pa_decay_ms_get(pa_shard_t *shard, extent_state_t state); + +/* + * Do deferred work on this PA shard. + * + * Morally, this should do both PAC decay and the HPA deferred work. For now, + * though, the arena, background thread, and PAC modules are tightly interwoven + * in a way that's tricky to extricate, so we only do the HPA-specific parts. + */ +void pa_shard_set_deferral_allowed(tsdn_t *tsdn, pa_shard_t *shard, + bool deferral_allowed); +void pa_shard_do_deferred_work(tsdn_t *tsdn, pa_shard_t *shard); +void pa_shard_try_deferred_work(tsdn_t *tsdn, pa_shard_t *shard); +uint64_t pa_shard_time_until_deferred_work(tsdn_t *tsdn, pa_shard_t *shard); + +/******************************************************************************/ +/* + * Various bits of "boring" functionality that are still part of this module, + * but that we relegate to pa_extra.c, to keep the core logic in pa.c as + * readable as possible. + */ + +/* + * These fork phases are synchronized with the arena fork phase numbering to + * make it easy to keep straight. That's why there's no prefork1. + */ +void pa_shard_prefork0(tsdn_t *tsdn, pa_shard_t *shard); +void pa_shard_prefork2(tsdn_t *tsdn, pa_shard_t *shard); +void pa_shard_prefork3(tsdn_t *tsdn, pa_shard_t *shard); +void pa_shard_prefork4(tsdn_t *tsdn, pa_shard_t *shard); +void pa_shard_prefork5(tsdn_t *tsdn, pa_shard_t *shard); +void pa_shard_postfork_parent(tsdn_t *tsdn, pa_shard_t *shard); +void pa_shard_postfork_child(tsdn_t *tsdn, pa_shard_t *shard); + +void pa_shard_basic_stats_merge(pa_shard_t *shard, size_t *nactive, + size_t *ndirty, size_t *nmuzzy); + +void pa_shard_stats_merge(tsdn_t *tsdn, pa_shard_t *shard, + pa_shard_stats_t *pa_shard_stats_out, pac_estats_t *estats_out, + hpa_shard_stats_t *hpa_stats_out, sec_stats_t *sec_stats_out, + size_t *resident); + +/* + * Reads the PA-owned mutex stats into the output stats array, at the + * appropriate positions. Morally, these stats should really live in + * pa_shard_stats_t, but the indices are sort of baked into the various mutex + * prof macros. This would be a good thing to do at some point. + */ +void pa_shard_mtx_stats_read(tsdn_t *tsdn, pa_shard_t *shard, + mutex_prof_data_t mutex_prof_data[mutex_prof_num_arena_mutexes]); + +#endif /* JEMALLOC_INTERNAL_PA_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/pac.h b/contrib/jemalloc/include/jemalloc/internal/pac.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/pac.h @@ -0,0 +1,178 @@ +#ifndef JEMALLOC_INTERNAL_PAC_H +#define JEMALLOC_INTERNAL_PAC_H + +#include "jemalloc/internal/exp_grow.h" +#include "jemalloc/internal/pai.h" +#include "san_bump.h" + + +/* + * Page allocator classic; an implementation of the PAI interface that: + * - Can be used for arenas with custom extent hooks. + * - Can always satisfy any allocation request (including highly-fragmentary + * ones). + * - Can use efficient OS-level zeroing primitives for demand-filled pages. + */ + +/* How "eager" decay/purging should be. */ +enum pac_purge_eagerness_e { + PAC_PURGE_ALWAYS, + PAC_PURGE_NEVER, + PAC_PURGE_ON_EPOCH_ADVANCE +}; +typedef enum pac_purge_eagerness_e pac_purge_eagerness_t; + +typedef struct pac_decay_stats_s pac_decay_stats_t; +struct pac_decay_stats_s { + /* Total number of purge sweeps. */ + locked_u64_t npurge; + /* Total number of madvise calls made. */ + locked_u64_t nmadvise; + /* Total number of pages purged. */ + locked_u64_t purged; +}; + +typedef struct pac_estats_s pac_estats_t; +struct pac_estats_s { + /* + * Stats for a given index in the range [0, SC_NPSIZES] in the various + * ecache_ts. + * We track both bytes and # of extents: two extents in the same bucket + * may have different sizes if adjacent size classes differ by more than + * a page, so bytes cannot always be derived from # of extents. + */ + size_t ndirty; + size_t dirty_bytes; + size_t nmuzzy; + size_t muzzy_bytes; + size_t nretained; + size_t retained_bytes; +}; + +typedef struct pac_stats_s pac_stats_t; +struct pac_stats_s { + pac_decay_stats_t decay_dirty; + pac_decay_stats_t decay_muzzy; + + /* + * Number of unused virtual memory bytes currently retained. Retained + * bytes are technically mapped (though always decommitted or purged), + * but they are excluded from the mapped statistic (above). + */ + size_t retained; /* Derived. */ + + /* + * Number of bytes currently mapped, excluding retained memory (and any + * base-allocated memory, which is tracked by the arena stats). + * + * We name this "pac_mapped" to avoid confusion with the arena_stats + * "mapped". + */ + atomic_zu_t pac_mapped; + + /* VM space had to be leaked (undocumented). Normally 0. */ + atomic_zu_t abandoned_vm; +}; + +struct pac_s { + /* + * Must be the first member (we convert it to a PAC given only a + * pointer). The handle to the allocation interface. + */ + pai_t pai; + /* + * Collections of extents that were previously allocated. These are + * used when allocating extents, in an attempt to re-use address space. + * + * Synchronization: internal. + */ + ecache_t ecache_dirty; + ecache_t ecache_muzzy; + ecache_t ecache_retained; + + base_t *base; + emap_t *emap; + edata_cache_t *edata_cache; + + /* The grow info for the retained ecache. */ + exp_grow_t exp_grow; + malloc_mutex_t grow_mtx; + + /* Special allocator for guarded frequently reused extents. */ + san_bump_alloc_t sba; + + /* How large extents should be before getting auto-purged. */ + atomic_zu_t oversize_threshold; + + /* + * Decay-based purging state, responsible for scheduling extent state + * transitions. + * + * Synchronization: via the internal mutex. + */ + decay_t decay_dirty; /* dirty --> muzzy */ + decay_t decay_muzzy; /* muzzy --> retained */ + + malloc_mutex_t *stats_mtx; + pac_stats_t *stats; + + /* Extent serial number generator state. */ + atomic_zu_t extent_sn_next; +}; + +bool pac_init(tsdn_t *tsdn, pac_t *pac, base_t *base, emap_t *emap, + edata_cache_t *edata_cache, nstime_t *cur_time, size_t oversize_threshold, + ssize_t dirty_decay_ms, ssize_t muzzy_decay_ms, pac_stats_t *pac_stats, + malloc_mutex_t *stats_mtx); + +static inline size_t +pac_mapped(pac_t *pac) { + return atomic_load_zu(&pac->stats->pac_mapped, ATOMIC_RELAXED); +} + +static inline ehooks_t * +pac_ehooks_get(pac_t *pac) { + return base_ehooks_get(pac->base); +} + +/* + * All purging functions require holding decay->mtx. This is one of the few + * places external modules are allowed to peek inside pa_shard_t internals. + */ + +/* + * Decays the number of pages currently in the ecache. This might not leave the + * ecache empty if other threads are inserting dirty objects into it + * concurrently with the call. + */ +void pac_decay_all(tsdn_t *tsdn, pac_t *pac, decay_t *decay, + pac_decay_stats_t *decay_stats, ecache_t *ecache, bool fully_decay); +/* + * Updates decay settings for the current time, and conditionally purges in + * response (depending on decay_purge_setting). Returns whether or not the + * epoch advanced. + */ +bool pac_maybe_decay_purge(tsdn_t *tsdn, pac_t *pac, decay_t *decay, + pac_decay_stats_t *decay_stats, ecache_t *ecache, + pac_purge_eagerness_t eagerness); + +/* + * Gets / sets the maximum amount that we'll grow an arena down the + * grow-retained pathways (unless forced to by an allocaction request). + * + * Set new_limit to NULL if it's just a query, or old_limit to NULL if you don't + * care about the previous value. + * + * Returns true on error (if the new limit is not valid). + */ +bool pac_retain_grow_limit_get_set(tsdn_t *tsdn, pac_t *pac, size_t *old_limit, + size_t *new_limit); + +bool pac_decay_ms_set(tsdn_t *tsdn, pac_t *pac, extent_state_t state, + ssize_t decay_ms, pac_purge_eagerness_t eagerness); +ssize_t pac_decay_ms_get(pac_t *pac, extent_state_t state); + +void pac_reset(tsdn_t *tsdn, pac_t *pac); +void pac_destroy(tsdn_t *tsdn, pac_t *pac); + +#endif /* JEMALLOC_INTERNAL_PAC_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/pages.h b/contrib/jemalloc/include/jemalloc/internal/pages.h --- a/contrib/jemalloc/include/jemalloc/internal/pages.h +++ b/contrib/jemalloc/include/jemalloc/internal/pages.h @@ -13,10 +13,27 @@ /* Return the smallest pagesize multiple that is >= s. */ #define PAGE_CEILING(s) \ (((s) + PAGE_MASK) & ~PAGE_MASK) +/* Return the largest pagesize multiple that is <=s. */ +#define PAGE_FLOOR(s) \ + ((s) & ~PAGE_MASK) /* Huge page size. LG_HUGEPAGE is determined by the configure script. */ #define HUGEPAGE ((size_t)(1U << LG_HUGEPAGE)) #define HUGEPAGE_MASK ((size_t)(HUGEPAGE - 1)) + +#if LG_HUGEPAGE != 0 +# define HUGEPAGE_PAGES (HUGEPAGE / PAGE) +#else +/* + * It's convenient to define arrays (or bitmaps) of HUGEPAGE_PAGES lengths. If + * we can't autodetect the hugepage size, it gets treated as 0, in which case + * we'll trigger a compiler error in those arrays. Avoid this case by ensuring + * that this value is at least 1. (We won't ever run in this degraded state; + * hpa_supported() returns false in this case. + */ +# define HUGEPAGE_PAGES 1 +#endif + /* Return the huge page base address for the huge page containing address a. */ #define HUGEPAGE_ADDR2BASE(a) \ ((void *)((uintptr_t)(a) & ~HUGEPAGE_MASK)) @@ -58,6 +75,18 @@ #endif ; +#if defined(JEMALLOC_HAVE_MADVISE_HUGE) || defined(JEMALLOC_HAVE_MEMCNTL) +# define PAGES_CAN_HUGIFY +#endif + +static const bool pages_can_hugify = +#ifdef PAGES_CAN_HUGIFY + true +#else + false +#endif + ; + typedef enum { thp_mode_default = 0, /* Do not change hugepage settings. */ thp_mode_always = 1, /* Always set MADV_HUGEPAGE. */ @@ -84,5 +113,7 @@ bool pages_dodump(void *addr, size_t size); bool pages_boot(void); void pages_set_thp_state (void *ptr, size_t size); +void pages_mark_guards(void *head, void *tail); +void pages_unmark_guards(void *head, void *tail); #endif /* JEMALLOC_INTERNAL_PAGES_EXTERNS_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/pai.h b/contrib/jemalloc/include/jemalloc/internal/pai.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/pai.h @@ -0,0 +1,95 @@ +#ifndef JEMALLOC_INTERNAL_PAI_H +#define JEMALLOC_INTERNAL_PAI_H + +/* An interface for page allocation. */ + +typedef struct pai_s pai_t; +struct pai_s { + /* Returns NULL on failure. */ + edata_t *(*alloc)(tsdn_t *tsdn, pai_t *self, size_t size, + size_t alignment, bool zero, bool guarded, bool frequent_reuse, + bool *deferred_work_generated); + /* + * Returns the number of extents added to the list (which may be fewer + * than requested, in case of OOM). The list should already be + * initialized. The only alignment guarantee is page-alignment, and + * the results are not necessarily zeroed. + */ + size_t (*alloc_batch)(tsdn_t *tsdn, pai_t *self, size_t size, + size_t nallocs, edata_list_active_t *results, + bool *deferred_work_generated); + bool (*expand)(tsdn_t *tsdn, pai_t *self, edata_t *edata, + size_t old_size, size_t new_size, bool zero, + bool *deferred_work_generated); + bool (*shrink)(tsdn_t *tsdn, pai_t *self, edata_t *edata, + size_t old_size, size_t new_size, bool *deferred_work_generated); + void (*dalloc)(tsdn_t *tsdn, pai_t *self, edata_t *edata, + bool *deferred_work_generated); + /* This function empties out list as a side-effect of being called. */ + void (*dalloc_batch)(tsdn_t *tsdn, pai_t *self, + edata_list_active_t *list, bool *deferred_work_generated); + uint64_t (*time_until_deferred_work)(tsdn_t *tsdn, pai_t *self); +}; + +/* + * These are just simple convenience functions to avoid having to reference the + * same pai_t twice on every invocation. + */ + +static inline edata_t * +pai_alloc(tsdn_t *tsdn, pai_t *self, size_t size, size_t alignment, + bool zero, bool guarded, bool frequent_reuse, + bool *deferred_work_generated) { + return self->alloc(tsdn, self, size, alignment, zero, guarded, + frequent_reuse, deferred_work_generated); +} + +static inline size_t +pai_alloc_batch(tsdn_t *tsdn, pai_t *self, size_t size, size_t nallocs, + edata_list_active_t *results, bool *deferred_work_generated) { + return self->alloc_batch(tsdn, self, size, nallocs, results, + deferred_work_generated); +} + +static inline bool +pai_expand(tsdn_t *tsdn, pai_t *self, edata_t *edata, size_t old_size, + size_t new_size, bool zero, bool *deferred_work_generated) { + return self->expand(tsdn, self, edata, old_size, new_size, zero, + deferred_work_generated); +} + +static inline bool +pai_shrink(tsdn_t *tsdn, pai_t *self, edata_t *edata, size_t old_size, + size_t new_size, bool *deferred_work_generated) { + return self->shrink(tsdn, self, edata, old_size, new_size, + deferred_work_generated); +} + +static inline void +pai_dalloc(tsdn_t *tsdn, pai_t *self, edata_t *edata, + bool *deferred_work_generated) { + self->dalloc(tsdn, self, edata, deferred_work_generated); +} + +static inline void +pai_dalloc_batch(tsdn_t *tsdn, pai_t *self, edata_list_active_t *list, + bool *deferred_work_generated) { + self->dalloc_batch(tsdn, self, list, deferred_work_generated); +} + +static inline uint64_t +pai_time_until_deferred_work(tsdn_t *tsdn, pai_t *self) { + return self->time_until_deferred_work(tsdn, self); +} + +/* + * An implementation of batch allocation that simply calls alloc once for + * each item in the list. + */ +size_t pai_alloc_batch_default(tsdn_t *tsdn, pai_t *self, size_t size, + size_t nallocs, edata_list_active_t *results, bool *deferred_work_generated); +/* Ditto, for dalloc. */ +void pai_dalloc_batch_default(tsdn_t *tsdn, pai_t *self, + edata_list_active_t *list, bool *deferred_work_generated); + +#endif /* JEMALLOC_INTERNAL_PAI_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/peak.h b/contrib/jemalloc/include/jemalloc/internal/peak.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/peak.h @@ -0,0 +1,37 @@ +#ifndef JEMALLOC_INTERNAL_PEAK_H +#define JEMALLOC_INTERNAL_PEAK_H + +typedef struct peak_s peak_t; +struct peak_s { + /* The highest recorded peak value, after adjustment (see below). */ + uint64_t cur_max; + /* + * The difference between alloc and dalloc at the last set_zero call; + * this lets us cancel out the appropriate amount of excess. + */ + uint64_t adjustment; +}; + +#define PEAK_INITIALIZER {0, 0} + +static inline uint64_t +peak_max(peak_t *peak) { + return peak->cur_max; +} + +static inline void +peak_update(peak_t *peak, uint64_t alloc, uint64_t dalloc) { + int64_t candidate_max = (int64_t)(alloc - dalloc - peak->adjustment); + if (candidate_max > (int64_t)peak->cur_max) { + peak->cur_max = candidate_max; + } +} + +/* Resets the counter to zero; all peaks are now relative to this point. */ +static inline void +peak_set_zero(peak_t *peak, uint64_t alloc, uint64_t dalloc) { + peak->cur_max = 0; + peak->adjustment = alloc - dalloc; +} + +#endif /* JEMALLOC_INTERNAL_PEAK_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/peak_event.h b/contrib/jemalloc/include/jemalloc/internal/peak_event.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/peak_event.h @@ -0,0 +1,24 @@ +#ifndef JEMALLOC_INTERNAL_PEAK_EVENT_H +#define JEMALLOC_INTERNAL_PEAK_EVENT_H + +/* + * While peak.h contains the simple helper struct that tracks state, this + * contains the allocator tie-ins (and knows about tsd, the event module, etc.). + */ + +/* Update the peak with current tsd state. */ +void peak_event_update(tsd_t *tsd); +/* Set current state to zero. */ +void peak_event_zero(tsd_t *tsd); +uint64_t peak_event_max(tsd_t *tsd); + +/* Manual hooks. */ +/* The activity-triggered hooks. */ +uint64_t peak_alloc_new_event_wait(tsd_t *tsd); +uint64_t peak_alloc_postponed_event_wait(tsd_t *tsd); +void peak_alloc_event_handler(tsd_t *tsd, uint64_t elapsed); +uint64_t peak_dalloc_new_event_wait(tsd_t *tsd); +uint64_t peak_dalloc_postponed_event_wait(tsd_t *tsd); +void peak_dalloc_event_handler(tsd_t *tsd, uint64_t elapsed); + +#endif /* JEMALLOC_INTERNAL_PEAK_EVENT_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/ph.h b/contrib/jemalloc/include/jemalloc/internal/ph.h --- a/contrib/jemalloc/include/jemalloc/internal/ph.h +++ b/contrib/jemalloc/include/jemalloc/internal/ph.h @@ -1,3 +1,6 @@ +#ifndef JEMALLOC_INTERNAL_PH_H +#define JEMALLOC_INTERNAL_PH_H + /* * A Pairing Heap implementation. * @@ -10,382 +13,508 @@ * http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.106.2988&rep=rep1&type=pdf * ******************************************************************************* + * + * We include a non-obvious optimization: + * - First, we introduce a new pop-and-link operation; pop the two most + * recently-inserted items off the aux-list, link them, and push the resulting + * heap. + * - We maintain a count of the number of insertions since the last time we + * merged the aux-list (i.e. via first() or remove_first()). After N inserts, + * we do ffs(N) pop-and-link operations. + * + * One way to think of this is that we're progressively building up a tree in + * the aux-list, rather than a linked-list (think of the series of merges that + * will be performed as the aux-count grows). + * + * There's a couple reasons we benefit from this: + * - Ordinarily, after N insertions, the aux-list is of size N. With our + * strategy, it's of size O(log(N)). So we decrease the worst-case time of + * first() calls, and reduce the average cost of remove_min calls. Since + * these almost always occur while holding a lock, we practically reduce the + * frequency of unusually long hold times. + * - This moves the bulk of the work of merging the aux-list onto the threads + * that are inserting into the heap. In some common scenarios, insertions + * happen in bulk, from a single thread (think tcache flushing; we potentially + * move many slabs from slabs_full to slabs_nonfull). All the nodes in this + * case are in the inserting threads cache, and linking them is very cheap + * (cache misses dominate linking cost). Without this optimization, linking + * happens on the next call to remove_first. Since that remove_first call + * likely happens on a different thread (or at least, after the cache has + * gotten cold if done on the same thread), deferring linking trades cheap + * link operations now for expensive ones later. + * + * The ffs trick keeps amortized insert cost at constant time. Similar + * strategies based on periodically sorting the list after a batch of operations + * perform worse than this in practice, even with various fancy tricks; they + * all took amortized complexity of an insert from O(1) to O(log(n)). */ -#ifndef PH_H_ -#define PH_H_ +typedef int (*ph_cmp_t)(void *, void *); /* Node structure. */ -#define phn(a_type) \ -struct { \ - a_type *phn_prev; \ - a_type *phn_next; \ - a_type *phn_lchild; \ +typedef struct phn_link_s phn_link_t; +struct phn_link_s { + void *prev; + void *next; + void *lchild; +}; + +typedef struct ph_s ph_t; +struct ph_s { + void *root; + /* + * Inserts done since the last aux-list merge. This is not necessarily + * the size of the aux-list, since it's possible that removals have + * happened since, and we don't track whether or not those removals are + * from the aux list. + */ + size_t auxcount; +}; + +JEMALLOC_ALWAYS_INLINE phn_link_t * +phn_link_get(void *phn, size_t offset) { + return (phn_link_t *)(((uintptr_t)phn) + offset); } -/* Root structure. */ -#define ph(a_type) \ -struct { \ - a_type *ph_root; \ +JEMALLOC_ALWAYS_INLINE void +phn_link_init(void *phn, size_t offset) { + phn_link_get(phn, offset)->prev = NULL; + phn_link_get(phn, offset)->next = NULL; + phn_link_get(phn, offset)->lchild = NULL; } -/* Internal utility macros. */ -#define phn_lchild_get(a_type, a_field, a_phn) \ - (a_phn->a_field.phn_lchild) -#define phn_lchild_set(a_type, a_field, a_phn, a_lchild) do { \ - a_phn->a_field.phn_lchild = a_lchild; \ -} while (0) - -#define phn_next_get(a_type, a_field, a_phn) \ - (a_phn->a_field.phn_next) -#define phn_prev_set(a_type, a_field, a_phn, a_prev) do { \ - a_phn->a_field.phn_prev = a_prev; \ -} while (0) - -#define phn_prev_get(a_type, a_field, a_phn) \ - (a_phn->a_field.phn_prev) -#define phn_next_set(a_type, a_field, a_phn, a_next) do { \ - a_phn->a_field.phn_next = a_next; \ -} while (0) - -#define phn_merge_ordered(a_type, a_field, a_phn0, a_phn1, a_cmp) do { \ - a_type *phn0child; \ - \ - assert(a_phn0 != NULL); \ - assert(a_phn1 != NULL); \ - assert(a_cmp(a_phn0, a_phn1) <= 0); \ - \ - phn_prev_set(a_type, a_field, a_phn1, a_phn0); \ - phn0child = phn_lchild_get(a_type, a_field, a_phn0); \ - phn_next_set(a_type, a_field, a_phn1, phn0child); \ - if (phn0child != NULL) { \ - phn_prev_set(a_type, a_field, phn0child, a_phn1); \ - } \ - phn_lchild_set(a_type, a_field, a_phn0, a_phn1); \ -} while (0) - -#define phn_merge(a_type, a_field, a_phn0, a_phn1, a_cmp, r_phn) do { \ - if (a_phn0 == NULL) { \ - r_phn = a_phn1; \ - } else if (a_phn1 == NULL) { \ - r_phn = a_phn0; \ - } else if (a_cmp(a_phn0, a_phn1) < 0) { \ - phn_merge_ordered(a_type, a_field, a_phn0, a_phn1, \ - a_cmp); \ - r_phn = a_phn0; \ - } else { \ - phn_merge_ordered(a_type, a_field, a_phn1, a_phn0, \ - a_cmp); \ - r_phn = a_phn1; \ - } \ -} while (0) +/* Internal utility helpers. */ +JEMALLOC_ALWAYS_INLINE void * +phn_lchild_get(void *phn, size_t offset) { + return phn_link_get(phn, offset)->lchild; +} + +JEMALLOC_ALWAYS_INLINE void +phn_lchild_set(void *phn, void *lchild, size_t offset) { + phn_link_get(phn, offset)->lchild = lchild; +} + +JEMALLOC_ALWAYS_INLINE void * +phn_next_get(void *phn, size_t offset) { + return phn_link_get(phn, offset)->next; +} + +JEMALLOC_ALWAYS_INLINE void +phn_next_set(void *phn, void *next, size_t offset) { + phn_link_get(phn, offset)->next = next; +} + +JEMALLOC_ALWAYS_INLINE void * +phn_prev_get(void *phn, size_t offset) { + return phn_link_get(phn, offset)->prev; +} + +JEMALLOC_ALWAYS_INLINE void +phn_prev_set(void *phn, void *prev, size_t offset) { + phn_link_get(phn, offset)->prev = prev; +} + +JEMALLOC_ALWAYS_INLINE void +phn_merge_ordered(void *phn0, void *phn1, size_t offset, + ph_cmp_t cmp) { + void *phn0child; + + assert(phn0 != NULL); + assert(phn1 != NULL); + assert(cmp(phn0, phn1) <= 0); + + phn_prev_set(phn1, phn0, offset); + phn0child = phn_lchild_get(phn0, offset); + phn_next_set(phn1, phn0child, offset); + if (phn0child != NULL) { + phn_prev_set(phn0child, phn1, offset); + } + phn_lchild_set(phn0, phn1, offset); +} + +JEMALLOC_ALWAYS_INLINE void * +phn_merge(void *phn0, void *phn1, size_t offset, ph_cmp_t cmp) { + void *result; + if (phn0 == NULL) { + result = phn1; + } else if (phn1 == NULL) { + result = phn0; + } else if (cmp(phn0, phn1) < 0) { + phn_merge_ordered(phn0, phn1, offset, cmp); + result = phn0; + } else { + phn_merge_ordered(phn1, phn0, offset, cmp); + result = phn1; + } + return result; +} + +JEMALLOC_ALWAYS_INLINE void * +phn_merge_siblings(void *phn, size_t offset, ph_cmp_t cmp) { + void *head = NULL; + void *tail = NULL; + void *phn0 = phn; + void *phn1 = phn_next_get(phn0, offset); + + /* + * Multipass merge, wherein the first two elements of a FIFO + * are repeatedly merged, and each result is appended to the + * singly linked FIFO, until the FIFO contains only a single + * element. We start with a sibling list but no reference to + * its tail, so we do a single pass over the sibling list to + * populate the FIFO. + */ + if (phn1 != NULL) { + void *phnrest = phn_next_get(phn1, offset); + if (phnrest != NULL) { + phn_prev_set(phnrest, NULL, offset); + } + phn_prev_set(phn0, NULL, offset); + phn_next_set(phn0, NULL, offset); + phn_prev_set(phn1, NULL, offset); + phn_next_set(phn1, NULL, offset); + phn0 = phn_merge(phn0, phn1, offset, cmp); + head = tail = phn0; + phn0 = phnrest; + while (phn0 != NULL) { + phn1 = phn_next_get(phn0, offset); + if (phn1 != NULL) { + phnrest = phn_next_get(phn1, offset); + if (phnrest != NULL) { + phn_prev_set(phnrest, NULL, offset); + } + phn_prev_set(phn0, NULL, offset); + phn_next_set(phn0, NULL, offset); + phn_prev_set(phn1, NULL, offset); + phn_next_set(phn1, NULL, offset); + phn0 = phn_merge(phn0, phn1, offset, cmp); + phn_next_set(tail, phn0, offset); + tail = phn0; + phn0 = phnrest; + } else { + phn_next_set(tail, phn0, offset); + tail = phn0; + phn0 = NULL; + } + } + phn0 = head; + phn1 = phn_next_get(phn0, offset); + if (phn1 != NULL) { + while (true) { + head = phn_next_get(phn1, offset); + assert(phn_prev_get(phn0, offset) == NULL); + phn_next_set(phn0, NULL, offset); + assert(phn_prev_get(phn1, offset) == NULL); + phn_next_set(phn1, NULL, offset); + phn0 = phn_merge(phn0, phn1, offset, cmp); + if (head == NULL) { + break; + } + phn_next_set(tail, phn0, offset); + tail = phn0; + phn0 = head; + phn1 = phn_next_get(phn0, offset); + } + } + } + return phn0; +} + +JEMALLOC_ALWAYS_INLINE void +ph_merge_aux(ph_t *ph, size_t offset, ph_cmp_t cmp) { + ph->auxcount = 0; + void *phn = phn_next_get(ph->root, offset); + if (phn != NULL) { + phn_prev_set(ph->root, NULL, offset); + phn_next_set(ph->root, NULL, offset); + phn_prev_set(phn, NULL, offset); + phn = phn_merge_siblings(phn, offset, cmp); + assert(phn_next_get(phn, offset) == NULL); + ph->root = phn_merge(ph->root, phn, offset, cmp); + } +} + +JEMALLOC_ALWAYS_INLINE void * +ph_merge_children(void *phn, size_t offset, ph_cmp_t cmp) { + void *result; + void *lchild = phn_lchild_get(phn, offset); + if (lchild == NULL) { + result = NULL; + } else { + result = phn_merge_siblings(lchild, offset, cmp); + } + return result; +} + +JEMALLOC_ALWAYS_INLINE void +ph_new(ph_t *ph) { + ph->root = NULL; + ph->auxcount = 0; +} + +JEMALLOC_ALWAYS_INLINE bool +ph_empty(ph_t *ph) { + return ph->root == NULL; +} + +JEMALLOC_ALWAYS_INLINE void * +ph_first(ph_t *ph, size_t offset, ph_cmp_t cmp) { + if (ph->root == NULL) { + return NULL; + } + ph_merge_aux(ph, offset, cmp); + return ph->root; +} + +JEMALLOC_ALWAYS_INLINE void * +ph_any(ph_t *ph, size_t offset) { + if (ph->root == NULL) { + return NULL; + } + void *aux = phn_next_get(ph->root, offset); + if (aux != NULL) { + return aux; + } + return ph->root; +} + +/* Returns true if we should stop trying to merge. */ +JEMALLOC_ALWAYS_INLINE bool +ph_try_aux_merge_pair(ph_t *ph, size_t offset, ph_cmp_t cmp) { + assert(ph->root != NULL); + void *phn0 = phn_next_get(ph->root, offset); + if (phn0 == NULL) { + return true; + } + void *phn1 = phn_next_get(phn0, offset); + if (phn1 == NULL) { + return true; + } + void *next_phn1 = phn_next_get(phn1, offset); + phn_next_set(phn0, NULL, offset); + phn_prev_set(phn0, NULL, offset); + phn_next_set(phn1, NULL, offset); + phn_prev_set(phn1, NULL, offset); + phn0 = phn_merge(phn0, phn1, offset, cmp); + phn_next_set(phn0, next_phn1, offset); + if (next_phn1 != NULL) { + phn_prev_set(next_phn1, phn0, offset); + } + phn_next_set(ph->root, phn0, offset); + phn_prev_set(phn0, ph->root, offset); + return next_phn1 == NULL; +} + +JEMALLOC_ALWAYS_INLINE void +ph_insert(ph_t *ph, void *phn, size_t offset, ph_cmp_t cmp) { + phn_link_init(phn, offset); -#define ph_merge_siblings(a_type, a_field, a_phn, a_cmp, r_phn) do { \ - a_type *head = NULL; \ - a_type *tail = NULL; \ - a_type *phn0 = a_phn; \ - a_type *phn1 = phn_next_get(a_type, a_field, phn0); \ + /* + * Treat the root as an aux list during insertion, and lazily merge + * during a_prefix##remove_first(). For elements that are inserted, + * then removed via a_prefix##remove() before the aux list is ever + * processed, this makes insert/remove constant-time, whereas eager + * merging would make insert O(log n). + */ + if (ph->root == NULL) { + ph->root = phn; + } else { + /* + * As a special case, check to see if we can replace the root. + * This is practically common in some important cases, and lets + * us defer some insertions (hopefully, until the point where + * some of the items in the aux list have been removed, savings + * us from linking them at all). + */ + if (cmp(phn, ph->root) < 0) { + phn_lchild_set(phn, ph->root, offset); + phn_prev_set(ph->root, phn, offset); + ph->root = phn; + ph->auxcount = 0; + return; + } + ph->auxcount++; + phn_next_set(phn, phn_next_get(ph->root, offset), offset); + if (phn_next_get(ph->root, offset) != NULL) { + phn_prev_set(phn_next_get(ph->root, offset), phn, + offset); + } + phn_prev_set(phn, ph->root, offset); + phn_next_set(ph->root, phn, offset); + } + if (ph->auxcount > 1) { + unsigned nmerges = ffs_zu(ph->auxcount - 1); + bool done = false; + for (unsigned i = 0; i < nmerges && !done; i++) { + done = ph_try_aux_merge_pair(ph, offset, cmp); + } + } +} + +JEMALLOC_ALWAYS_INLINE void * +ph_remove_first(ph_t *ph, size_t offset, ph_cmp_t cmp) { + void *ret; + + if (ph->root == NULL) { + return NULL; + } + ph_merge_aux(ph, offset, cmp); + ret = ph->root; + ph->root = ph_merge_children(ph->root, offset, cmp); + + return ret; + +} + +JEMALLOC_ALWAYS_INLINE void +ph_remove(ph_t *ph, void *phn, size_t offset, ph_cmp_t cmp) { + void *replace; + void *parent; + + if (ph->root == phn) { + /* + * We can delete from aux list without merging it, but we need + * to merge if we are dealing with the root node and it has + * children. + */ + if (phn_lchild_get(phn, offset) == NULL) { + ph->root = phn_next_get(phn, offset); + if (ph->root != NULL) { + phn_prev_set(ph->root, NULL, offset); + } + return; + } + ph_merge_aux(ph, offset, cmp); + if (ph->root == phn) { + ph->root = ph_merge_children(ph->root, offset, cmp); + return; + } + } + + /* Get parent (if phn is leftmost child) before mutating. */ + if ((parent = phn_prev_get(phn, offset)) != NULL) { + if (phn_lchild_get(parent, offset) != phn) { + parent = NULL; + } + } + /* Find a possible replacement node, and link to parent. */ + replace = ph_merge_children(phn, offset, cmp); + /* Set next/prev for sibling linked list. */ + if (replace != NULL) { + if (parent != NULL) { + phn_prev_set(replace, parent, offset); + phn_lchild_set(parent, replace, offset); + } else { + phn_prev_set(replace, phn_prev_get(phn, offset), + offset); + if (phn_prev_get(phn, offset) != NULL) { + phn_next_set(phn_prev_get(phn, offset), replace, + offset); + } + } + phn_next_set(replace, phn_next_get(phn, offset), offset); + if (phn_next_get(phn, offset) != NULL) { + phn_prev_set(phn_next_get(phn, offset), replace, + offset); + } + } else { + if (parent != NULL) { + void *next = phn_next_get(phn, offset); + phn_lchild_set(parent, next, offset); + if (next != NULL) { + phn_prev_set(next, parent, offset); + } + } else { + assert(phn_prev_get(phn, offset) != NULL); + phn_next_set( + phn_prev_get(phn, offset), + phn_next_get(phn, offset), offset); + } + if (phn_next_get(phn, offset) != NULL) { + phn_prev_set( + phn_next_get(phn, offset), + phn_prev_get(phn, offset), offset); + } + } +} + +#define ph_structs(a_prefix, a_type) \ +typedef struct { \ + phn_link_t link; \ +} a_prefix##_link_t; \ \ - /* \ - * Multipass merge, wherein the first two elements of a FIFO \ - * are repeatedly merged, and each result is appended to the \ - * singly linked FIFO, until the FIFO contains only a single \ - * element. We start with a sibling list but no reference to \ - * its tail, so we do a single pass over the sibling list to \ - * populate the FIFO. \ - */ \ - if (phn1 != NULL) { \ - a_type *phnrest = phn_next_get(a_type, a_field, phn1); \ - if (phnrest != NULL) { \ - phn_prev_set(a_type, a_field, phnrest, NULL); \ - } \ - phn_prev_set(a_type, a_field, phn0, NULL); \ - phn_next_set(a_type, a_field, phn0, NULL); \ - phn_prev_set(a_type, a_field, phn1, NULL); \ - phn_next_set(a_type, a_field, phn1, NULL); \ - phn_merge(a_type, a_field, phn0, phn1, a_cmp, phn0); \ - head = tail = phn0; \ - phn0 = phnrest; \ - while (phn0 != NULL) { \ - phn1 = phn_next_get(a_type, a_field, phn0); \ - if (phn1 != NULL) { \ - phnrest = phn_next_get(a_type, a_field, \ - phn1); \ - if (phnrest != NULL) { \ - phn_prev_set(a_type, a_field, \ - phnrest, NULL); \ - } \ - phn_prev_set(a_type, a_field, phn0, \ - NULL); \ - phn_next_set(a_type, a_field, phn0, \ - NULL); \ - phn_prev_set(a_type, a_field, phn1, \ - NULL); \ - phn_next_set(a_type, a_field, phn1, \ - NULL); \ - phn_merge(a_type, a_field, phn0, phn1, \ - a_cmp, phn0); \ - phn_next_set(a_type, a_field, tail, \ - phn0); \ - tail = phn0; \ - phn0 = phnrest; \ - } else { \ - phn_next_set(a_type, a_field, tail, \ - phn0); \ - tail = phn0; \ - phn0 = NULL; \ - } \ - } \ - phn0 = head; \ - phn1 = phn_next_get(a_type, a_field, phn0); \ - if (phn1 != NULL) { \ - while (true) { \ - head = phn_next_get(a_type, a_field, \ - phn1); \ - assert(phn_prev_get(a_type, a_field, \ - phn0) == NULL); \ - phn_next_set(a_type, a_field, phn0, \ - NULL); \ - assert(phn_prev_get(a_type, a_field, \ - phn1) == NULL); \ - phn_next_set(a_type, a_field, phn1, \ - NULL); \ - phn_merge(a_type, a_field, phn0, phn1, \ - a_cmp, phn0); \ - if (head == NULL) { \ - break; \ - } \ - phn_next_set(a_type, a_field, tail, \ - phn0); \ - tail = phn0; \ - phn0 = head; \ - phn1 = phn_next_get(a_type, a_field, \ - phn0); \ - } \ - } \ - } \ - r_phn = phn0; \ -} while (0) - -#define ph_merge_aux(a_type, a_field, a_ph, a_cmp) do { \ - a_type *phn = phn_next_get(a_type, a_field, a_ph->ph_root); \ - if (phn != NULL) { \ - phn_prev_set(a_type, a_field, a_ph->ph_root, NULL); \ - phn_next_set(a_type, a_field, a_ph->ph_root, NULL); \ - phn_prev_set(a_type, a_field, phn, NULL); \ - ph_merge_siblings(a_type, a_field, phn, a_cmp, phn); \ - assert(phn_next_get(a_type, a_field, phn) == NULL); \ - phn_merge(a_type, a_field, a_ph->ph_root, phn, a_cmp, \ - a_ph->ph_root); \ - } \ -} while (0) - -#define ph_merge_children(a_type, a_field, a_phn, a_cmp, r_phn) do { \ - a_type *lchild = phn_lchild_get(a_type, a_field, a_phn); \ - if (lchild == NULL) { \ - r_phn = NULL; \ - } else { \ - ph_merge_siblings(a_type, a_field, lchild, a_cmp, \ - r_phn); \ - } \ -} while (0) +typedef struct { \ + ph_t ph; \ +} a_prefix##_t; /* * The ph_proto() macro generates function prototypes that correspond to the * functions generated by an equivalently parameterized call to ph_gen(). */ -#define ph_proto(a_attr, a_prefix, a_ph_type, a_type) \ -a_attr void a_prefix##new(a_ph_type *ph); \ -a_attr bool a_prefix##empty(a_ph_type *ph); \ -a_attr a_type *a_prefix##first(a_ph_type *ph); \ -a_attr a_type *a_prefix##any(a_ph_type *ph); \ -a_attr void a_prefix##insert(a_ph_type *ph, a_type *phn); \ -a_attr a_type *a_prefix##remove_first(a_ph_type *ph); \ -a_attr a_type *a_prefix##remove_any(a_ph_type *ph); \ -a_attr void a_prefix##remove(a_ph_type *ph, a_type *phn); +#define ph_proto(a_attr, a_prefix, a_type) \ + \ +a_attr void a_prefix##_new(a_prefix##_t *ph); \ +a_attr bool a_prefix##_empty(a_prefix##_t *ph); \ +a_attr a_type *a_prefix##_first(a_prefix##_t *ph); \ +a_attr a_type *a_prefix##_any(a_prefix##_t *ph); \ +a_attr void a_prefix##_insert(a_prefix##_t *ph, a_type *phn); \ +a_attr a_type *a_prefix##_remove_first(a_prefix##_t *ph); \ +a_attr void a_prefix##_remove(a_prefix##_t *ph, a_type *phn); \ +a_attr a_type *a_prefix##_remove_any(a_prefix##_t *ph); -/* - * The ph_gen() macro generates a type-specific pairing heap implementation, - * based on the above cpp macros. - */ -#define ph_gen(a_attr, a_prefix, a_ph_type, a_type, a_field, a_cmp) \ +/* The ph_gen() macro generates a type-specific pairing heap implementation. */ +#define ph_gen(a_attr, a_prefix, a_type, a_field, a_cmp) \ +JEMALLOC_ALWAYS_INLINE int \ +a_prefix##_ph_cmp(void *a, void *b) { \ + return a_cmp((a_type *)a, (a_type *)b); \ +} \ + \ a_attr void \ -a_prefix##new(a_ph_type *ph) { \ - memset(ph, 0, sizeof(ph(a_type))); \ +a_prefix##_new(a_prefix##_t *ph) { \ + ph_new(&ph->ph); \ } \ + \ a_attr bool \ -a_prefix##empty(a_ph_type *ph) { \ - return (ph->ph_root == NULL); \ +a_prefix##_empty(a_prefix##_t *ph) { \ + return ph_empty(&ph->ph); \ } \ + \ a_attr a_type * \ -a_prefix##first(a_ph_type *ph) { \ - if (ph->ph_root == NULL) { \ - return NULL; \ - } \ - ph_merge_aux(a_type, a_field, ph, a_cmp); \ - return ph->ph_root; \ +a_prefix##_first(a_prefix##_t *ph) { \ + return ph_first(&ph->ph, offsetof(a_type, a_field), \ + &a_prefix##_ph_cmp); \ } \ + \ a_attr a_type * \ -a_prefix##any(a_ph_type *ph) { \ - if (ph->ph_root == NULL) { \ - return NULL; \ - } \ - a_type *aux = phn_next_get(a_type, a_field, ph->ph_root); \ - if (aux != NULL) { \ - return aux; \ - } \ - return ph->ph_root; \ +a_prefix##_any(a_prefix##_t *ph) { \ + return ph_any(&ph->ph, offsetof(a_type, a_field)); \ } \ -a_attr void \ -a_prefix##insert(a_ph_type *ph, a_type *phn) { \ - memset(&phn->a_field, 0, sizeof(phn(a_type))); \ \ - /* \ - * Treat the root as an aux list during insertion, and lazily \ - * merge during a_prefix##remove_first(). For elements that \ - * are inserted, then removed via a_prefix##remove() before the \ - * aux list is ever processed, this makes insert/remove \ - * constant-time, whereas eager merging would make insert \ - * O(log n). \ - */ \ - if (ph->ph_root == NULL) { \ - ph->ph_root = phn; \ - } else { \ - phn_next_set(a_type, a_field, phn, phn_next_get(a_type, \ - a_field, ph->ph_root)); \ - if (phn_next_get(a_type, a_field, ph->ph_root) != \ - NULL) { \ - phn_prev_set(a_type, a_field, \ - phn_next_get(a_type, a_field, ph->ph_root), \ - phn); \ - } \ - phn_prev_set(a_type, a_field, phn, ph->ph_root); \ - phn_next_set(a_type, a_field, ph->ph_root, phn); \ - } \ +a_attr void \ +a_prefix##_insert(a_prefix##_t *ph, a_type *phn) { \ + ph_insert(&ph->ph, phn, offsetof(a_type, a_field), \ + a_prefix##_ph_cmp); \ } \ -a_attr a_type * \ -a_prefix##remove_first(a_ph_type *ph) { \ - a_type *ret; \ \ - if (ph->ph_root == NULL) { \ - return NULL; \ - } \ - ph_merge_aux(a_type, a_field, ph, a_cmp); \ - \ - ret = ph->ph_root; \ - \ - ph_merge_children(a_type, a_field, ph->ph_root, a_cmp, \ - ph->ph_root); \ +a_attr a_type * \ +a_prefix##_remove_first(a_prefix##_t *ph) { \ + return ph_remove_first(&ph->ph, offsetof(a_type, a_field), \ + a_prefix##_ph_cmp); \ +} \ \ - return ret; \ +a_attr void \ +a_prefix##_remove(a_prefix##_t *ph, a_type *phn) { \ + ph_remove(&ph->ph, phn, offsetof(a_type, a_field), \ + a_prefix##_ph_cmp); \ } \ + \ a_attr a_type * \ -a_prefix##remove_any(a_ph_type *ph) { \ - /* \ - * Remove the most recently inserted aux list element, or the \ - * root if the aux list is empty. This has the effect of \ - * behaving as a LIFO (and insertion/removal is therefore \ - * constant-time) if a_prefix##[remove_]first() are never \ - * called. \ - */ \ - if (ph->ph_root == NULL) { \ - return NULL; \ - } \ - a_type *ret = phn_next_get(a_type, a_field, ph->ph_root); \ +a_prefix##_remove_any(a_prefix##_t *ph) { \ + a_type *ret = a_prefix##_any(ph); \ if (ret != NULL) { \ - a_type *aux = phn_next_get(a_type, a_field, ret); \ - phn_next_set(a_type, a_field, ph->ph_root, aux); \ - if (aux != NULL) { \ - phn_prev_set(a_type, a_field, aux, \ - ph->ph_root); \ - } \ - return ret; \ + a_prefix##_remove(ph, ret); \ } \ - ret = ph->ph_root; \ - ph_merge_children(a_type, a_field, ph->ph_root, a_cmp, \ - ph->ph_root); \ return ret; \ -} \ -a_attr void \ -a_prefix##remove(a_ph_type *ph, a_type *phn) { \ - a_type *replace, *parent; \ - \ - if (ph->ph_root == phn) { \ - /* \ - * We can delete from aux list without merging it, but \ - * we need to merge if we are dealing with the root \ - * node and it has children. \ - */ \ - if (phn_lchild_get(a_type, a_field, phn) == NULL) { \ - ph->ph_root = phn_next_get(a_type, a_field, \ - phn); \ - if (ph->ph_root != NULL) { \ - phn_prev_set(a_type, a_field, \ - ph->ph_root, NULL); \ - } \ - return; \ - } \ - ph_merge_aux(a_type, a_field, ph, a_cmp); \ - if (ph->ph_root == phn) { \ - ph_merge_children(a_type, a_field, ph->ph_root, \ - a_cmp, ph->ph_root); \ - return; \ - } \ - } \ - \ - /* Get parent (if phn is leftmost child) before mutating. */ \ - if ((parent = phn_prev_get(a_type, a_field, phn)) != NULL) { \ - if (phn_lchild_get(a_type, a_field, parent) != phn) { \ - parent = NULL; \ - } \ - } \ - /* Find a possible replacement node, and link to parent. */ \ - ph_merge_children(a_type, a_field, phn, a_cmp, replace); \ - /* Set next/prev for sibling linked list. */ \ - if (replace != NULL) { \ - if (parent != NULL) { \ - phn_prev_set(a_type, a_field, replace, parent); \ - phn_lchild_set(a_type, a_field, parent, \ - replace); \ - } else { \ - phn_prev_set(a_type, a_field, replace, \ - phn_prev_get(a_type, a_field, phn)); \ - if (phn_prev_get(a_type, a_field, phn) != \ - NULL) { \ - phn_next_set(a_type, a_field, \ - phn_prev_get(a_type, a_field, phn), \ - replace); \ - } \ - } \ - phn_next_set(a_type, a_field, replace, \ - phn_next_get(a_type, a_field, phn)); \ - if (phn_next_get(a_type, a_field, phn) != NULL) { \ - phn_prev_set(a_type, a_field, \ - phn_next_get(a_type, a_field, phn), \ - replace); \ - } \ - } else { \ - if (parent != NULL) { \ - a_type *next = phn_next_get(a_type, a_field, \ - phn); \ - phn_lchild_set(a_type, a_field, parent, next); \ - if (next != NULL) { \ - phn_prev_set(a_type, a_field, next, \ - parent); \ - } \ - } else { \ - assert(phn_prev_get(a_type, a_field, phn) != \ - NULL); \ - phn_next_set(a_type, a_field, \ - phn_prev_get(a_type, a_field, phn), \ - phn_next_get(a_type, a_field, phn)); \ - } \ - if (phn_next_get(a_type, a_field, phn) != NULL) { \ - phn_prev_set(a_type, a_field, \ - phn_next_get(a_type, a_field, phn), \ - phn_prev_get(a_type, a_field, phn)); \ - } \ - } \ } -#endif /* PH_H_ */ +#endif /* JEMALLOC_INTERNAL_PH_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/prng.h b/contrib/jemalloc/include/jemalloc/internal/prng.h --- a/contrib/jemalloc/include/jemalloc/internal/prng.h +++ b/contrib/jemalloc/include/jemalloc/internal/prng.h @@ -1,7 +1,6 @@ #ifndef JEMALLOC_INTERNAL_PRNG_H #define JEMALLOC_INTERNAL_PRNG_H -#include "jemalloc/internal/atomic.h" #include "jemalloc/internal/bit_util.h" /* @@ -59,66 +58,38 @@ /* * The prng_lg_range functions give a uniform int in the half-open range [0, - * 2**lg_range). If atomic is true, they do so safely from multiple threads. - * Multithreaded 64-bit prngs aren't supported. + * 2**lg_range). */ JEMALLOC_ALWAYS_INLINE uint32_t -prng_lg_range_u32(atomic_u32_t *state, unsigned lg_range, bool atomic) { - uint32_t ret, state0, state1; - +prng_lg_range_u32(uint32_t *state, unsigned lg_range) { assert(lg_range > 0); assert(lg_range <= 32); - state0 = atomic_load_u32(state, ATOMIC_RELAXED); - - if (atomic) { - do { - state1 = prng_state_next_u32(state0); - } while (!atomic_compare_exchange_weak_u32(state, &state0, - state1, ATOMIC_RELAXED, ATOMIC_RELAXED)); - } else { - state1 = prng_state_next_u32(state0); - atomic_store_u32(state, state1, ATOMIC_RELAXED); - } - ret = state1 >> (32 - lg_range); + *state = prng_state_next_u32(*state); + uint32_t ret = *state >> (32 - lg_range); return ret; } JEMALLOC_ALWAYS_INLINE uint64_t prng_lg_range_u64(uint64_t *state, unsigned lg_range) { - uint64_t ret, state1; - assert(lg_range > 0); assert(lg_range <= 64); - state1 = prng_state_next_u64(*state); - *state = state1; - ret = state1 >> (64 - lg_range); + *state = prng_state_next_u64(*state); + uint64_t ret = *state >> (64 - lg_range); return ret; } JEMALLOC_ALWAYS_INLINE size_t -prng_lg_range_zu(atomic_zu_t *state, unsigned lg_range, bool atomic) { - size_t ret, state0, state1; - +prng_lg_range_zu(size_t *state, unsigned lg_range) { assert(lg_range > 0); assert(lg_range <= ZU(1) << (3 + LG_SIZEOF_PTR)); - state0 = atomic_load_zu(state, ATOMIC_RELAXED); - - if (atomic) { - do { - state1 = prng_state_next_zu(state0); - } while (atomic_compare_exchange_weak_zu(state, &state0, - state1, ATOMIC_RELAXED, ATOMIC_RELAXED)); - } else { - state1 = prng_state_next_zu(state0); - atomic_store_zu(state, state1, ATOMIC_RELAXED); - } - ret = state1 >> ((ZU(1) << (3 + LG_SIZEOF_PTR)) - lg_range); + *state = prng_state_next_zu(*state); + size_t ret = *state >> ((ZU(1) << (3 + LG_SIZEOF_PTR)) - lg_range); return ret; } @@ -129,18 +100,24 @@ */ JEMALLOC_ALWAYS_INLINE uint32_t -prng_range_u32(atomic_u32_t *state, uint32_t range, bool atomic) { - uint32_t ret; - unsigned lg_range; - - assert(range > 1); +prng_range_u32(uint32_t *state, uint32_t range) { + assert(range != 0); + /* + * If range were 1, lg_range would be 0, so the shift in + * prng_lg_range_u32 would be a shift of a 32-bit variable by 32 bits, + * which is UB. Just handle this case as a one-off. + */ + if (range == 1) { + return 0; + } /* Compute the ceiling of lg(range). */ - lg_range = ffs_u32(pow2_ceil_u32(range)) - 1; + unsigned lg_range = ffs_u32(pow2_ceil_u32(range)); /* Generate a result in [0..range) via repeated trial. */ + uint32_t ret; do { - ret = prng_lg_range_u32(state, lg_range, atomic); + ret = prng_lg_range_u32(state, lg_range); } while (ret >= range); return ret; @@ -148,15 +125,18 @@ JEMALLOC_ALWAYS_INLINE uint64_t prng_range_u64(uint64_t *state, uint64_t range) { - uint64_t ret; - unsigned lg_range; + assert(range != 0); - assert(range > 1); + /* See the note in prng_range_u32. */ + if (range == 1) { + return 0; + } /* Compute the ceiling of lg(range). */ - lg_range = ffs_u64(pow2_ceil_u64(range)) - 1; + unsigned lg_range = ffs_u64(pow2_ceil_u64(range)); /* Generate a result in [0..range) via repeated trial. */ + uint64_t ret; do { ret = prng_lg_range_u64(state, lg_range); } while (ret >= range); @@ -165,18 +145,21 @@ } JEMALLOC_ALWAYS_INLINE size_t -prng_range_zu(atomic_zu_t *state, size_t range, bool atomic) { - size_t ret; - unsigned lg_range; +prng_range_zu(size_t *state, size_t range) { + assert(range != 0); - assert(range > 1); + /* See the note in prng_range_u32. */ + if (range == 1) { + return 0; + } /* Compute the ceiling of lg(range). */ - lg_range = ffs_u64(pow2_ceil_u64(range)) - 1; + unsigned lg_range = ffs_u64(pow2_ceil_u64(range)); /* Generate a result in [0..range) via repeated trial. */ + size_t ret; do { - ret = prng_lg_range_zu(state, lg_range, atomic); + ret = prng_lg_range_zu(state, lg_range); } while (ret >= range); return ret; diff --git a/contrib/jemalloc/include/jemalloc/internal/prof_data.h b/contrib/jemalloc/include/jemalloc/internal/prof_data.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/prof_data.h @@ -0,0 +1,37 @@ +#ifndef JEMALLOC_INTERNAL_PROF_DATA_H +#define JEMALLOC_INTERNAL_PROF_DATA_H + +#include "jemalloc/internal/mutex.h" + +extern malloc_mutex_t bt2gctx_mtx; +extern malloc_mutex_t tdatas_mtx; +extern malloc_mutex_t prof_dump_mtx; + +extern malloc_mutex_t *gctx_locks; +extern malloc_mutex_t *tdata_locks; + +extern size_t prof_unbiased_sz[PROF_SC_NSIZES]; +extern size_t prof_shifted_unbiased_cnt[PROF_SC_NSIZES]; + +void prof_bt_hash(const void *key, size_t r_hash[2]); +bool prof_bt_keycomp(const void *k1, const void *k2); + +bool prof_data_init(tsd_t *tsd); +prof_tctx_t *prof_lookup(tsd_t *tsd, prof_bt_t *bt); +char *prof_thread_name_alloc(tsd_t *tsd, const char *thread_name); +int prof_thread_name_set_impl(tsd_t *tsd, const char *thread_name); +void prof_unbias_map_init(); +void prof_dump_impl(tsd_t *tsd, write_cb_t *prof_dump_write, void *cbopaque, + prof_tdata_t *tdata, bool leakcheck); +prof_tdata_t * prof_tdata_init_impl(tsd_t *tsd, uint64_t thr_uid, + uint64_t thr_discrim, char *thread_name, bool active); +void prof_tdata_detach(tsd_t *tsd, prof_tdata_t *tdata); +void prof_reset(tsd_t *tsd, size_t lg_sample); +void prof_tctx_try_destroy(tsd_t *tsd, prof_tctx_t *tctx); + +/* Used in unit tests. */ +size_t prof_tdata_count(void); +size_t prof_bt_count(void); +void prof_cnt_all(prof_cnt_t *cnt_all); + +#endif /* JEMALLOC_INTERNAL_PROF_DATA_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/prof_externs.h b/contrib/jemalloc/include/jemalloc/internal/prof_externs.h --- a/contrib/jemalloc/include/jemalloc/internal/prof_externs.h +++ b/contrib/jemalloc/include/jemalloc/internal/prof_externs.h @@ -2,75 +2,72 @@ #define JEMALLOC_INTERNAL_PROF_EXTERNS_H #include "jemalloc/internal/mutex.h" +#include "jemalloc/internal/prof_hook.h" -extern malloc_mutex_t bt2gctx_mtx; - -extern bool opt_prof; -extern bool opt_prof_active; -extern bool opt_prof_thread_active_init; -extern size_t opt_lg_prof_sample; /* Mean bytes between samples. */ -extern ssize_t opt_lg_prof_interval; /* lg(prof_interval). */ -extern bool opt_prof_gdump; /* High-water memory dumping. */ -extern bool opt_prof_final; /* Final profile dumping. */ -extern bool opt_prof_leak; /* Dump leak summary at exit. */ -extern bool opt_prof_accum; /* Report cumulative bytes. */ -extern bool opt_prof_log; /* Turn logging on at boot. */ -extern char opt_prof_prefix[ +extern bool opt_prof; +extern bool opt_prof_active; +extern bool opt_prof_thread_active_init; +extern size_t opt_lg_prof_sample; /* Mean bytes between samples. */ +extern ssize_t opt_lg_prof_interval; /* lg(prof_interval). */ +extern bool opt_prof_gdump; /* High-water memory dumping. */ +extern bool opt_prof_final; /* Final profile dumping. */ +extern bool opt_prof_leak; /* Dump leak summary at exit. */ +extern bool opt_prof_leak_error; /* Exit with error code if memory leaked */ +extern bool opt_prof_accum; /* Report cumulative bytes. */ +extern bool opt_prof_log; /* Turn logging on at boot. */ +extern char opt_prof_prefix[ /* Minimize memory bloat for non-prof builds. */ #ifdef JEMALLOC_PROF PATH_MAX + #endif 1]; +extern bool opt_prof_unbias; + +/* For recording recent allocations */ +extern ssize_t opt_prof_recent_alloc_max; + +/* Whether to use thread name provided by the system or by mallctl. */ +extern bool opt_prof_sys_thread_name; + +/* Whether to record per size class counts and request size totals. */ +extern bool opt_prof_stats; /* Accessed via prof_active_[gs]et{_unlocked,}(). */ -extern bool prof_active; +extern bool prof_active_state; /* Accessed via prof_gdump_[gs]et{_unlocked,}(). */ -extern bool prof_gdump_val; +extern bool prof_gdump_val; -/* - * Profile dump interval, measured in bytes allocated. Each arena triggers a - * profile dump when it reaches this threshold. The effect is that the - * interval between profile dumps averages prof_interval, though the actual - * interval between dumps will tend to be sporadic, and the interval will be a - * maximum of approximately (prof_interval * narenas). - */ -extern uint64_t prof_interval; +/* Profile dump interval, measured in bytes allocated. */ +extern uint64_t prof_interval; /* * Initialized as opt_lg_prof_sample, and potentially modified during profiling * resets. */ -extern size_t lg_prof_sample; - -void prof_alloc_rollback(tsd_t *tsd, prof_tctx_t *tctx, bool updated); -void prof_malloc_sample_object(tsdn_t *tsdn, const void *ptr, size_t usize, - prof_tctx_t *tctx); -void prof_free_sampled_object(tsd_t *tsd, const void *ptr, size_t usize, - prof_tctx_t *tctx); -void bt_init(prof_bt_t *bt, void **vec); -void prof_backtrace(prof_bt_t *bt); -prof_tctx_t *prof_lookup(tsd_t *tsd, prof_bt_t *bt); -#ifdef JEMALLOC_JET -size_t prof_tdata_count(void); -size_t prof_bt_count(void); -#endif -typedef int (prof_dump_open_t)(bool, const char *); -extern prof_dump_open_t *JET_MUTABLE prof_dump_open; - -typedef bool (prof_dump_header_t)(tsdn_t *, bool, const prof_cnt_t *); -extern prof_dump_header_t *JET_MUTABLE prof_dump_header; -#ifdef JEMALLOC_JET -void prof_cnt_all(uint64_t *curobjs, uint64_t *curbytes, uint64_t *accumobjs, - uint64_t *accumbytes); -#endif -bool prof_accum_init(tsdn_t *tsdn, prof_accum_t *prof_accum); +extern size_t lg_prof_sample; + +extern bool prof_booted; + +void prof_backtrace_hook_set(prof_backtrace_hook_t hook); +prof_backtrace_hook_t prof_backtrace_hook_get(); + +void prof_dump_hook_set(prof_dump_hook_t hook); +prof_dump_hook_t prof_dump_hook_get(); + +/* Functions only accessed in prof_inlines.h */ +prof_tdata_t *prof_tdata_init(tsd_t *tsd); +prof_tdata_t *prof_tdata_reinit(tsd_t *tsd, prof_tdata_t *tdata); + +void prof_alloc_rollback(tsd_t *tsd, prof_tctx_t *tctx); +void prof_malloc_sample_object(tsd_t *tsd, const void *ptr, size_t size, + size_t usize, prof_tctx_t *tctx); +void prof_free_sampled_object(tsd_t *tsd, size_t usize, prof_info_t *prof_info); +prof_tctx_t *prof_tctx_create(tsd_t *tsd); void prof_idump(tsdn_t *tsdn); bool prof_mdump(tsd_t *tsd, const char *filename); void prof_gdump(tsdn_t *tsdn); -prof_tdata_t *prof_tdata_init(tsd_t *tsd); -prof_tdata_t *prof_tdata_reinit(tsd_t *tsd, prof_tdata_t *tdata); -void prof_reset(tsd_t *tsd, size_t lg_sample); + void prof_tdata_cleanup(tsd_t *tsd); bool prof_active_get(tsdn_t *tsdn); bool prof_active_set(tsdn_t *tsdn, bool active); @@ -84,22 +81,15 @@ bool prof_gdump_set(tsdn_t *tsdn, bool active); void prof_boot0(void); void prof_boot1(void); -bool prof_boot2(tsd_t *tsd); +bool prof_boot2(tsd_t *tsd, base_t *base); void prof_prefork0(tsdn_t *tsdn); void prof_prefork1(tsdn_t *tsdn); void prof_postfork_parent(tsdn_t *tsdn); void prof_postfork_child(tsdn_t *tsdn); -void prof_sample_threshold_update(prof_tdata_t *tdata); - -bool prof_log_start(tsdn_t *tsdn, const char *filename); -bool prof_log_stop(tsdn_t *tsdn); -#ifdef JEMALLOC_JET -size_t prof_log_bt_count(void); -size_t prof_log_alloc_count(void); -size_t prof_log_thr_count(void); -bool prof_log_is_logging(void); -bool prof_log_rep_check(void); -void prof_log_dummy_set(bool new_value); -#endif + +/* Only accessed by thread event. */ +uint64_t prof_sample_new_event_wait(tsd_t *tsd); +uint64_t prof_sample_postponed_event_wait(tsd_t *tsd); +void prof_sample_event_handler(tsd_t *tsd, uint64_t elapsed); #endif /* JEMALLOC_INTERNAL_PROF_EXTERNS_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/prof_hook.h b/contrib/jemalloc/include/jemalloc/internal/prof_hook.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/prof_hook.h @@ -0,0 +1,21 @@ +#ifndef JEMALLOC_INTERNAL_PROF_HOOK_H +#define JEMALLOC_INTERNAL_PROF_HOOK_H + +/* + * The hooks types of which are declared in this file are experimental and + * undocumented, thus the typedefs are located in an 'internal' header. + */ + +/* + * A hook to mock out backtrace functionality. This can be handy, since it's + * otherwise difficult to guarantee that two allocations are reported as coming + * from the exact same stack trace in the presence of an optimizing compiler. + */ +typedef void (*prof_backtrace_hook_t)(void **, unsigned *, unsigned); + +/* + * A callback hook that notifies about recently dumped heap profile. + */ +typedef void (*prof_dump_hook_t)(const char *filename); + +#endif /* JEMALLOC_INTERNAL_PROF_HOOK_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/prof_inlines.h b/contrib/jemalloc/include/jemalloc/internal/prof_inlines.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/prof_inlines.h @@ -0,0 +1,261 @@ +#ifndef JEMALLOC_INTERNAL_PROF_INLINES_H +#define JEMALLOC_INTERNAL_PROF_INLINES_H + +#include "jemalloc/internal/safety_check.h" +#include "jemalloc/internal/sz.h" +#include "jemalloc/internal/thread_event.h" + +JEMALLOC_ALWAYS_INLINE void +prof_active_assert() { + cassert(config_prof); + /* + * If opt_prof is off, then prof_active must always be off, regardless + * of whether prof_active_mtx is in effect or not. + */ + assert(opt_prof || !prof_active_state); +} + +JEMALLOC_ALWAYS_INLINE bool +prof_active_get_unlocked(void) { + prof_active_assert(); + /* + * Even if opt_prof is true, sampling can be temporarily disabled by + * setting prof_active to false. No locking is used when reading + * prof_active in the fast path, so there are no guarantees regarding + * how long it will take for all threads to notice state changes. + */ + return prof_active_state; +} + +JEMALLOC_ALWAYS_INLINE bool +prof_gdump_get_unlocked(void) { + /* + * No locking is used when reading prof_gdump_val in the fast path, so + * there are no guarantees regarding how long it will take for all + * threads to notice state changes. + */ + return prof_gdump_val; +} + +JEMALLOC_ALWAYS_INLINE prof_tdata_t * +prof_tdata_get(tsd_t *tsd, bool create) { + prof_tdata_t *tdata; + + cassert(config_prof); + + tdata = tsd_prof_tdata_get(tsd); + if (create) { + assert(tsd_reentrancy_level_get(tsd) == 0); + if (unlikely(tdata == NULL)) { + if (tsd_nominal(tsd)) { + tdata = prof_tdata_init(tsd); + tsd_prof_tdata_set(tsd, tdata); + } + } else if (unlikely(tdata->expired)) { + tdata = prof_tdata_reinit(tsd, tdata); + tsd_prof_tdata_set(tsd, tdata); + } + assert(tdata == NULL || tdata->attached); + } + + return tdata; +} + +JEMALLOC_ALWAYS_INLINE void +prof_info_get(tsd_t *tsd, const void *ptr, emap_alloc_ctx_t *alloc_ctx, + prof_info_t *prof_info) { + cassert(config_prof); + assert(ptr != NULL); + assert(prof_info != NULL); + + arena_prof_info_get(tsd, ptr, alloc_ctx, prof_info, false); +} + +JEMALLOC_ALWAYS_INLINE void +prof_info_get_and_reset_recent(tsd_t *tsd, const void *ptr, + emap_alloc_ctx_t *alloc_ctx, prof_info_t *prof_info) { + cassert(config_prof); + assert(ptr != NULL); + assert(prof_info != NULL); + + arena_prof_info_get(tsd, ptr, alloc_ctx, prof_info, true); +} + +JEMALLOC_ALWAYS_INLINE void +prof_tctx_reset(tsd_t *tsd, const void *ptr, emap_alloc_ctx_t *alloc_ctx) { + cassert(config_prof); + assert(ptr != NULL); + + arena_prof_tctx_reset(tsd, ptr, alloc_ctx); +} + +JEMALLOC_ALWAYS_INLINE void +prof_tctx_reset_sampled(tsd_t *tsd, const void *ptr) { + cassert(config_prof); + assert(ptr != NULL); + + arena_prof_tctx_reset_sampled(tsd, ptr); +} + +JEMALLOC_ALWAYS_INLINE void +prof_info_set(tsd_t *tsd, edata_t *edata, prof_tctx_t *tctx, size_t size) { + cassert(config_prof); + assert(edata != NULL); + assert((uintptr_t)tctx > (uintptr_t)1U); + + arena_prof_info_set(tsd, edata, tctx, size); +} + +JEMALLOC_ALWAYS_INLINE bool +prof_sample_should_skip(tsd_t *tsd, bool sample_event) { + cassert(config_prof); + + /* Fastpath: no need to load tdata */ + if (likely(!sample_event)) { + return true; + } + + /* + * sample_event is always obtained from the thread event module, and + * whenever it's true, it means that the thread event module has + * already checked the reentrancy level. + */ + assert(tsd_reentrancy_level_get(tsd) == 0); + + prof_tdata_t *tdata = prof_tdata_get(tsd, true); + if (unlikely(tdata == NULL)) { + return true; + } + + return !tdata->active; +} + +JEMALLOC_ALWAYS_INLINE prof_tctx_t * +prof_alloc_prep(tsd_t *tsd, bool prof_active, bool sample_event) { + prof_tctx_t *ret; + + if (!prof_active || + likely(prof_sample_should_skip(tsd, sample_event))) { + ret = (prof_tctx_t *)(uintptr_t)1U; + } else { + ret = prof_tctx_create(tsd); + } + + return ret; +} + +JEMALLOC_ALWAYS_INLINE void +prof_malloc(tsd_t *tsd, const void *ptr, size_t size, size_t usize, + emap_alloc_ctx_t *alloc_ctx, prof_tctx_t *tctx) { + cassert(config_prof); + assert(ptr != NULL); + assert(usize == isalloc(tsd_tsdn(tsd), ptr)); + + if (unlikely((uintptr_t)tctx > (uintptr_t)1U)) { + prof_malloc_sample_object(tsd, ptr, size, usize, tctx); + } else { + prof_tctx_reset(tsd, ptr, alloc_ctx); + } +} + +JEMALLOC_ALWAYS_INLINE void +prof_realloc(tsd_t *tsd, const void *ptr, size_t size, size_t usize, + prof_tctx_t *tctx, bool prof_active, const void *old_ptr, size_t old_usize, + prof_info_t *old_prof_info, bool sample_event) { + bool sampled, old_sampled, moved; + + cassert(config_prof); + assert(ptr != NULL || (uintptr_t)tctx <= (uintptr_t)1U); + + if (prof_active && ptr != NULL) { + assert(usize == isalloc(tsd_tsdn(tsd), ptr)); + if (prof_sample_should_skip(tsd, sample_event)) { + /* + * Don't sample. The usize passed to prof_alloc_prep() + * was larger than what actually got allocated, so a + * backtrace was captured for this allocation, even + * though its actual usize was insufficient to cross the + * sample threshold. + */ + prof_alloc_rollback(tsd, tctx); + tctx = (prof_tctx_t *)(uintptr_t)1U; + } + } + + sampled = ((uintptr_t)tctx > (uintptr_t)1U); + old_sampled = ((uintptr_t)old_prof_info->alloc_tctx > (uintptr_t)1U); + moved = (ptr != old_ptr); + + if (unlikely(sampled)) { + prof_malloc_sample_object(tsd, ptr, size, usize, tctx); + } else if (moved) { + prof_tctx_reset(tsd, ptr, NULL); + } else if (unlikely(old_sampled)) { + /* + * prof_tctx_reset() would work for the !moved case as well, + * but prof_tctx_reset_sampled() is slightly cheaper, and the + * proper thing to do here in the presence of explicit + * knowledge re: moved state. + */ + prof_tctx_reset_sampled(tsd, ptr); + } else { + prof_info_t prof_info; + prof_info_get(tsd, ptr, NULL, &prof_info); + assert((uintptr_t)prof_info.alloc_tctx == (uintptr_t)1U); + } + + /* + * The prof_free_sampled_object() call must come after the + * prof_malloc_sample_object() call, because tctx and old_tctx may be + * the same, in which case reversing the call order could cause the tctx + * to be prematurely destroyed as a side effect of momentarily zeroed + * counters. + */ + if (unlikely(old_sampled)) { + prof_free_sampled_object(tsd, old_usize, old_prof_info); + } +} + +JEMALLOC_ALWAYS_INLINE size_t +prof_sample_align(size_t orig_align) { + /* + * Enforce page alignment, so that sampled allocations can be identified + * w/o metadata lookup. + */ + assert(opt_prof); + return (opt_cache_oblivious && orig_align < PAGE) ? PAGE : + orig_align; +} + +JEMALLOC_ALWAYS_INLINE bool +prof_sample_aligned(const void *ptr) { + return ((uintptr_t)ptr & PAGE_MASK) == 0; +} + +JEMALLOC_ALWAYS_INLINE bool +prof_sampled(tsd_t *tsd, const void *ptr) { + prof_info_t prof_info; + prof_info_get(tsd, ptr, NULL, &prof_info); + bool sampled = (uintptr_t)prof_info.alloc_tctx > (uintptr_t)1U; + if (sampled) { + assert(prof_sample_aligned(ptr)); + } + return sampled; +} + +JEMALLOC_ALWAYS_INLINE void +prof_free(tsd_t *tsd, const void *ptr, size_t usize, + emap_alloc_ctx_t *alloc_ctx) { + prof_info_t prof_info; + prof_info_get_and_reset_recent(tsd, ptr, alloc_ctx, &prof_info); + + cassert(config_prof); + assert(usize == isalloc(tsd_tsdn(tsd), ptr)); + + if (unlikely((uintptr_t)prof_info.alloc_tctx > (uintptr_t)1U)) { + assert(prof_sample_aligned(ptr)); + prof_free_sampled_object(tsd, usize, &prof_info); + } +} + +#endif /* JEMALLOC_INTERNAL_PROF_INLINES_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/prof_inlines_a.h b/contrib/jemalloc/include/jemalloc/internal/prof_inlines_a.h deleted file mode 100644 --- a/contrib/jemalloc/include/jemalloc/internal/prof_inlines_a.h +++ /dev/null @@ -1,85 +0,0 @@ -#ifndef JEMALLOC_INTERNAL_PROF_INLINES_A_H -#define JEMALLOC_INTERNAL_PROF_INLINES_A_H - -#include "jemalloc/internal/mutex.h" - -static inline bool -prof_accum_add(tsdn_t *tsdn, prof_accum_t *prof_accum, - uint64_t accumbytes) { - cassert(config_prof); - - bool overflow; - uint64_t a0, a1; - - /* - * If the application allocates fast enough (and/or if idump is slow - * enough), extreme overflow here (a1 >= prof_interval * 2) can cause - * idump trigger coalescing. This is an intentional mechanism that - * avoids rate-limiting allocation. - */ -#ifdef JEMALLOC_ATOMIC_U64 - a0 = atomic_load_u64(&prof_accum->accumbytes, ATOMIC_RELAXED); - do { - a1 = a0 + accumbytes; - assert(a1 >= a0); - overflow = (a1 >= prof_interval); - if (overflow) { - a1 %= prof_interval; - } - } while (!atomic_compare_exchange_weak_u64(&prof_accum->accumbytes, &a0, - a1, ATOMIC_RELAXED, ATOMIC_RELAXED)); -#else - malloc_mutex_lock(tsdn, &prof_accum->mtx); - a0 = prof_accum->accumbytes; - a1 = a0 + accumbytes; - overflow = (a1 >= prof_interval); - if (overflow) { - a1 %= prof_interval; - } - prof_accum->accumbytes = a1; - malloc_mutex_unlock(tsdn, &prof_accum->mtx); -#endif - return overflow; -} - -static inline void -prof_accum_cancel(tsdn_t *tsdn, prof_accum_t *prof_accum, - size_t usize) { - cassert(config_prof); - - /* - * Cancel out as much of the excessive prof_accumbytes increase as - * possible without underflowing. Interval-triggered dumps occur - * slightly more often than intended as a result of incomplete - * canceling. - */ - uint64_t a0, a1; -#ifdef JEMALLOC_ATOMIC_U64 - a0 = atomic_load_u64(&prof_accum->accumbytes, ATOMIC_RELAXED); - do { - a1 = (a0 >= SC_LARGE_MINCLASS - usize) - ? a0 - (SC_LARGE_MINCLASS - usize) : 0; - } while (!atomic_compare_exchange_weak_u64(&prof_accum->accumbytes, &a0, - a1, ATOMIC_RELAXED, ATOMIC_RELAXED)); -#else - malloc_mutex_lock(tsdn, &prof_accum->mtx); - a0 = prof_accum->accumbytes; - a1 = (a0 >= SC_LARGE_MINCLASS - usize) - ? a0 - (SC_LARGE_MINCLASS - usize) : 0; - prof_accum->accumbytes = a1; - malloc_mutex_unlock(tsdn, &prof_accum->mtx); -#endif -} - -JEMALLOC_ALWAYS_INLINE bool -prof_active_get_unlocked(void) { - /* - * Even if opt_prof is true, sampling can be temporarily disabled by - * setting prof_active to false. No locking is used when reading - * prof_active in the fast path, so there are no guarantees regarding - * how long it will take for all threads to notice state changes. - */ - return prof_active; -} - -#endif /* JEMALLOC_INTERNAL_PROF_INLINES_A_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/prof_inlines_b.h b/contrib/jemalloc/include/jemalloc/internal/prof_inlines_b.h deleted file mode 100644 --- a/contrib/jemalloc/include/jemalloc/internal/prof_inlines_b.h +++ /dev/null @@ -1,250 +0,0 @@ -#ifndef JEMALLOC_INTERNAL_PROF_INLINES_B_H -#define JEMALLOC_INTERNAL_PROF_INLINES_B_H - -#include "jemalloc/internal/safety_check.h" -#include "jemalloc/internal/sz.h" - -JEMALLOC_ALWAYS_INLINE bool -prof_gdump_get_unlocked(void) { - /* - * No locking is used when reading prof_gdump_val in the fast path, so - * there are no guarantees regarding how long it will take for all - * threads to notice state changes. - */ - return prof_gdump_val; -} - -JEMALLOC_ALWAYS_INLINE prof_tdata_t * -prof_tdata_get(tsd_t *tsd, bool create) { - prof_tdata_t *tdata; - - cassert(config_prof); - - tdata = tsd_prof_tdata_get(tsd); - if (create) { - if (unlikely(tdata == NULL)) { - if (tsd_nominal(tsd)) { - tdata = prof_tdata_init(tsd); - tsd_prof_tdata_set(tsd, tdata); - } - } else if (unlikely(tdata->expired)) { - tdata = prof_tdata_reinit(tsd, tdata); - tsd_prof_tdata_set(tsd, tdata); - } - assert(tdata == NULL || tdata->attached); - } - - return tdata; -} - -JEMALLOC_ALWAYS_INLINE prof_tctx_t * -prof_tctx_get(tsdn_t *tsdn, const void *ptr, alloc_ctx_t *alloc_ctx) { - cassert(config_prof); - assert(ptr != NULL); - - return arena_prof_tctx_get(tsdn, ptr, alloc_ctx); -} - -JEMALLOC_ALWAYS_INLINE void -prof_tctx_set(tsdn_t *tsdn, const void *ptr, size_t usize, - alloc_ctx_t *alloc_ctx, prof_tctx_t *tctx) { - cassert(config_prof); - assert(ptr != NULL); - - arena_prof_tctx_set(tsdn, ptr, usize, alloc_ctx, tctx); -} - -JEMALLOC_ALWAYS_INLINE void -prof_tctx_reset(tsdn_t *tsdn, const void *ptr, prof_tctx_t *tctx) { - cassert(config_prof); - assert(ptr != NULL); - - arena_prof_tctx_reset(tsdn, ptr, tctx); -} - -JEMALLOC_ALWAYS_INLINE nstime_t -prof_alloc_time_get(tsdn_t *tsdn, const void *ptr, alloc_ctx_t *alloc_ctx) { - cassert(config_prof); - assert(ptr != NULL); - - return arena_prof_alloc_time_get(tsdn, ptr, alloc_ctx); -} - -JEMALLOC_ALWAYS_INLINE void -prof_alloc_time_set(tsdn_t *tsdn, const void *ptr, alloc_ctx_t *alloc_ctx, - nstime_t t) { - cassert(config_prof); - assert(ptr != NULL); - - arena_prof_alloc_time_set(tsdn, ptr, alloc_ctx, t); -} - -JEMALLOC_ALWAYS_INLINE bool -prof_sample_check(tsd_t *tsd, size_t usize, bool update) { - ssize_t check = update ? 0 : usize; - - int64_t bytes_until_sample = tsd_bytes_until_sample_get(tsd); - if (update) { - bytes_until_sample -= usize; - if (tsd_nominal(tsd)) { - tsd_bytes_until_sample_set(tsd, bytes_until_sample); - } - } - if (likely(bytes_until_sample >= check)) { - return true; - } - - return false; -} - -JEMALLOC_ALWAYS_INLINE bool -prof_sample_accum_update(tsd_t *tsd, size_t usize, bool update, - prof_tdata_t **tdata_out) { - prof_tdata_t *tdata; - - cassert(config_prof); - - /* Fastpath: no need to load tdata */ - if (likely(prof_sample_check(tsd, usize, update))) { - return true; - } - - bool booted = tsd_prof_tdata_get(tsd); - tdata = prof_tdata_get(tsd, true); - if (unlikely((uintptr_t)tdata <= (uintptr_t)PROF_TDATA_STATE_MAX)) { - tdata = NULL; - } - - if (tdata_out != NULL) { - *tdata_out = tdata; - } - - if (unlikely(tdata == NULL)) { - return true; - } - - /* - * If this was the first creation of tdata, then - * prof_tdata_get() reset bytes_until_sample, so decrement and - * check it again - */ - if (!booted && prof_sample_check(tsd, usize, update)) { - return true; - } - - if (tsd_reentrancy_level_get(tsd) > 0) { - return true; - } - /* Compute new sample threshold. */ - if (update) { - prof_sample_threshold_update(tdata); - } - return !tdata->active; -} - -JEMALLOC_ALWAYS_INLINE prof_tctx_t * -prof_alloc_prep(tsd_t *tsd, size_t usize, bool prof_active, bool update) { - prof_tctx_t *ret; - prof_tdata_t *tdata; - prof_bt_t bt; - - assert(usize == sz_s2u(usize)); - - if (!prof_active || likely(prof_sample_accum_update(tsd, usize, update, - &tdata))) { - ret = (prof_tctx_t *)(uintptr_t)1U; - } else { - bt_init(&bt, tdata->vec); - prof_backtrace(&bt); - ret = prof_lookup(tsd, &bt); - } - - return ret; -} - -JEMALLOC_ALWAYS_INLINE void -prof_malloc(tsdn_t *tsdn, const void *ptr, size_t usize, alloc_ctx_t *alloc_ctx, - prof_tctx_t *tctx) { - cassert(config_prof); - assert(ptr != NULL); - assert(usize == isalloc(tsdn, ptr)); - - if (unlikely((uintptr_t)tctx > (uintptr_t)1U)) { - prof_malloc_sample_object(tsdn, ptr, usize, tctx); - } else { - prof_tctx_set(tsdn, ptr, usize, alloc_ctx, - (prof_tctx_t *)(uintptr_t)1U); - } -} - -JEMALLOC_ALWAYS_INLINE void -prof_realloc(tsd_t *tsd, const void *ptr, size_t usize, prof_tctx_t *tctx, - bool prof_active, bool updated, const void *old_ptr, size_t old_usize, - prof_tctx_t *old_tctx) { - bool sampled, old_sampled, moved; - - cassert(config_prof); - assert(ptr != NULL || (uintptr_t)tctx <= (uintptr_t)1U); - - if (prof_active && !updated && ptr != NULL) { - assert(usize == isalloc(tsd_tsdn(tsd), ptr)); - if (prof_sample_accum_update(tsd, usize, true, NULL)) { - /* - * Don't sample. The usize passed to prof_alloc_prep() - * was larger than what actually got allocated, so a - * backtrace was captured for this allocation, even - * though its actual usize was insufficient to cross the - * sample threshold. - */ - prof_alloc_rollback(tsd, tctx, true); - tctx = (prof_tctx_t *)(uintptr_t)1U; - } - } - - sampled = ((uintptr_t)tctx > (uintptr_t)1U); - old_sampled = ((uintptr_t)old_tctx > (uintptr_t)1U); - moved = (ptr != old_ptr); - - if (unlikely(sampled)) { - prof_malloc_sample_object(tsd_tsdn(tsd), ptr, usize, tctx); - } else if (moved) { - prof_tctx_set(tsd_tsdn(tsd), ptr, usize, NULL, - (prof_tctx_t *)(uintptr_t)1U); - } else if (unlikely(old_sampled)) { - /* - * prof_tctx_set() would work for the !moved case as well, but - * prof_tctx_reset() is slightly cheaper, and the proper thing - * to do here in the presence of explicit knowledge re: moved - * state. - */ - prof_tctx_reset(tsd_tsdn(tsd), ptr, tctx); - } else { - assert((uintptr_t)prof_tctx_get(tsd_tsdn(tsd), ptr, NULL) == - (uintptr_t)1U); - } - - /* - * The prof_free_sampled_object() call must come after the - * prof_malloc_sample_object() call, because tctx and old_tctx may be - * the same, in which case reversing the call order could cause the tctx - * to be prematurely destroyed as a side effect of momentarily zeroed - * counters. - */ - if (unlikely(old_sampled)) { - prof_free_sampled_object(tsd, ptr, old_usize, old_tctx); - } -} - -JEMALLOC_ALWAYS_INLINE void -prof_free(tsd_t *tsd, const void *ptr, size_t usize, alloc_ctx_t *alloc_ctx) { - prof_tctx_t *tctx = prof_tctx_get(tsd_tsdn(tsd), ptr, alloc_ctx); - - cassert(config_prof); - assert(usize == isalloc(tsd_tsdn(tsd), ptr)); - - if (unlikely((uintptr_t)tctx > (uintptr_t)1U)) { - prof_free_sampled_object(tsd, ptr, usize, tctx); - } -} - -#endif /* JEMALLOC_INTERNAL_PROF_INLINES_B_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/prof_log.h b/contrib/jemalloc/include/jemalloc/internal/prof_log.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/prof_log.h @@ -0,0 +1,22 @@ +#ifndef JEMALLOC_INTERNAL_PROF_LOG_H +#define JEMALLOC_INTERNAL_PROF_LOG_H + +#include "jemalloc/internal/mutex.h" + +extern malloc_mutex_t log_mtx; + +void prof_try_log(tsd_t *tsd, size_t usize, prof_info_t *prof_info); +bool prof_log_init(tsd_t *tsdn); + +/* Used in unit tests. */ +size_t prof_log_bt_count(void); +size_t prof_log_alloc_count(void); +size_t prof_log_thr_count(void); +bool prof_log_is_logging(void); +bool prof_log_rep_check(void); +void prof_log_dummy_set(bool new_value); + +bool prof_log_start(tsdn_t *tsdn, const char *filename); +bool prof_log_stop(tsdn_t *tsdn); + +#endif /* JEMALLOC_INTERNAL_PROF_LOG_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/prof_recent.h b/contrib/jemalloc/include/jemalloc/internal/prof_recent.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/prof_recent.h @@ -0,0 +1,23 @@ +#ifndef JEMALLOC_INTERNAL_PROF_RECENT_H +#define JEMALLOC_INTERNAL_PROF_RECENT_H + +extern malloc_mutex_t prof_recent_alloc_mtx; +extern malloc_mutex_t prof_recent_dump_mtx; + +bool prof_recent_alloc_prepare(tsd_t *tsd, prof_tctx_t *tctx); +void prof_recent_alloc(tsd_t *tsd, edata_t *edata, size_t size, size_t usize); +void prof_recent_alloc_reset(tsd_t *tsd, edata_t *edata); +bool prof_recent_init(); +void edata_prof_recent_alloc_init(edata_t *edata); + +/* Used in unit tests. */ +typedef ql_head(prof_recent_t) prof_recent_list_t; +extern prof_recent_list_t prof_recent_alloc_list; +edata_t *prof_recent_alloc_edata_get_no_lock_test(const prof_recent_t *node); +prof_recent_t *edata_prof_recent_alloc_get_no_lock_test(const edata_t *edata); + +ssize_t prof_recent_alloc_max_ctl_read(); +ssize_t prof_recent_alloc_max_ctl_write(tsd_t *tsd, ssize_t max); +void prof_recent_alloc_dump(tsd_t *tsd, write_cb_t *write_cb, void *cbopaque); + +#endif /* JEMALLOC_INTERNAL_PROF_RECENT_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/prof_stats.h b/contrib/jemalloc/include/jemalloc/internal/prof_stats.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/prof_stats.h @@ -0,0 +1,17 @@ +#ifndef JEMALLOC_INTERNAL_PROF_STATS_H +#define JEMALLOC_INTERNAL_PROF_STATS_H + +typedef struct prof_stats_s prof_stats_t; +struct prof_stats_s { + uint64_t req_sum; + uint64_t count; +}; + +extern malloc_mutex_t prof_stats_mtx; + +void prof_stats_inc(tsd_t *tsd, szind_t ind, size_t size); +void prof_stats_dec(tsd_t *tsd, szind_t ind, size_t size); +void prof_stats_get_live(tsd_t *tsd, szind_t ind, prof_stats_t *stats); +void prof_stats_get_accum(tsd_t *tsd, szind_t ind, prof_stats_t *stats); + +#endif /* JEMALLOC_INTERNAL_PROF_STATS_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/prof_structs.h b/contrib/jemalloc/include/jemalloc/internal/prof_structs.h --- a/contrib/jemalloc/include/jemalloc/internal/prof_structs.h +++ b/contrib/jemalloc/include/jemalloc/internal/prof_structs.h @@ -2,6 +2,7 @@ #define JEMALLOC_INTERNAL_PROF_STRUCTS_H #include "jemalloc/internal/ckh.h" +#include "jemalloc/internal/edata.h" #include "jemalloc/internal/mutex.h" #include "jemalloc/internal/prng.h" #include "jemalloc/internal/rb.h" @@ -15,26 +16,22 @@ #ifdef JEMALLOC_PROF_LIBGCC /* Data structure passed to libgcc _Unwind_Backtrace() callback functions. */ typedef struct { - prof_bt_t *bt; + void **vec; + unsigned *len; unsigned max; } prof_unwind_data_t; #endif -struct prof_accum_s { -#ifndef JEMALLOC_ATOMIC_U64 - malloc_mutex_t mtx; - uint64_t accumbytes; -#else - atomic_u64_t accumbytes; -#endif -}; - struct prof_cnt_s { /* Profiling counters. */ uint64_t curobjs; + uint64_t curobjs_shifted_unbiased; uint64_t curbytes; + uint64_t curbytes_unbiased; uint64_t accumobjs; + uint64_t accumobjs_shifted_unbiased; uint64_t accumbytes; + uint64_t accumbytes_unbiased; }; typedef enum { @@ -55,6 +52,12 @@ uint64_t thr_uid; uint64_t thr_discrim; + /* + * Reference count of how many times this tctx object is referenced in + * recent allocation / deallocation records, protected by tdata->lock. + */ + uint64_t recent_count; + /* Profiling counters, protected by tdata->lock. */ prof_cnt_t cnts; @@ -96,6 +99,15 @@ }; typedef rb_tree(prof_tctx_t) prof_tctx_tree_t; +struct prof_info_s { + /* Time when the allocation was made. */ + nstime_t alloc_time; + /* Points to the prof_tctx_t corresponding to the allocation. */ + prof_tctx_t *alloc_tctx; + /* Allocation request size. */ + size_t alloc_size; +}; + struct prof_gctx_s { /* Protects nlimbo, cnt_summed, and tctxs. */ malloc_mutex_t *lock; @@ -167,9 +179,6 @@ */ ckh_t bt2tctx; - /* Sampling state. */ - uint64_t prng_state; - /* State used to avoid dumping while operating on prof internals. */ bool enq; bool enq_idump; @@ -197,4 +206,16 @@ }; typedef rb_tree(prof_tdata_t) prof_tdata_tree_t; +struct prof_recent_s { + nstime_t alloc_time; + nstime_t dalloc_time; + + ql_elm(prof_recent_t) link; + size_t size; + size_t usize; + atomic_p_t alloc_edata; /* NULL means allocation has been freed. */ + prof_tctx_t *alloc_tctx; + prof_tctx_t *dalloc_tctx; +}; + #endif /* JEMALLOC_INTERNAL_PROF_STRUCTS_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/prof_sys.h b/contrib/jemalloc/include/jemalloc/internal/prof_sys.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/prof_sys.h @@ -0,0 +1,30 @@ +#ifndef JEMALLOC_INTERNAL_PROF_SYS_H +#define JEMALLOC_INTERNAL_PROF_SYS_H + +extern malloc_mutex_t prof_dump_filename_mtx; +extern base_t *prof_base; + +void bt_init(prof_bt_t *bt, void **vec); +void prof_backtrace(tsd_t *tsd, prof_bt_t *bt); +void prof_hooks_init(); +void prof_unwind_init(); +void prof_sys_thread_name_fetch(tsd_t *tsd); +int prof_getpid(void); +void prof_get_default_filename(tsdn_t *tsdn, char *filename, uint64_t ind); +bool prof_prefix_set(tsdn_t *tsdn, const char *prefix); +void prof_fdump_impl(tsd_t *tsd); +void prof_idump_impl(tsd_t *tsd); +bool prof_mdump_impl(tsd_t *tsd, const char *filename); +void prof_gdump_impl(tsd_t *tsd); + +/* Used in unit tests. */ +typedef int (prof_sys_thread_name_read_t)(char *buf, size_t limit); +extern prof_sys_thread_name_read_t *JET_MUTABLE prof_sys_thread_name_read; +typedef int (prof_dump_open_file_t)(const char *, int); +extern prof_dump_open_file_t *JET_MUTABLE prof_dump_open_file; +typedef ssize_t (prof_dump_write_file_t)(int, const void *, size_t); +extern prof_dump_write_file_t *JET_MUTABLE prof_dump_write_file; +typedef int (prof_dump_open_maps_t)(); +extern prof_dump_open_maps_t *JET_MUTABLE prof_dump_open_maps; + +#endif /* JEMALLOC_INTERNAL_PROF_SYS_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/prof_types.h b/contrib/jemalloc/include/jemalloc/internal/prof_types.h --- a/contrib/jemalloc/include/jemalloc/internal/prof_types.h +++ b/contrib/jemalloc/include/jemalloc/internal/prof_types.h @@ -2,11 +2,12 @@ #define JEMALLOC_INTERNAL_PROF_TYPES_H typedef struct prof_bt_s prof_bt_t; -typedef struct prof_accum_s prof_accum_t; typedef struct prof_cnt_s prof_cnt_t; typedef struct prof_tctx_s prof_tctx_t; +typedef struct prof_info_s prof_info_t; typedef struct prof_gctx_s prof_gctx_t; typedef struct prof_tdata_s prof_tdata_t; +typedef struct prof_recent_s prof_recent_t; /* Option defaults. */ #ifdef JEMALLOC_PROF @@ -28,7 +29,23 @@ #define PROF_CKH_MINITEMS 64 /* Size of memory buffer to use when writing dump files. */ -#define PROF_DUMP_BUFSIZE 65536 +#ifndef JEMALLOC_PROF +/* Minimize memory bloat for non-prof builds. */ +# define PROF_DUMP_BUFSIZE 1 +#elif defined(JEMALLOC_DEBUG) +/* Use a small buffer size in debug build, mainly to facilitate testing. */ +# define PROF_DUMP_BUFSIZE 16 +#else +# define PROF_DUMP_BUFSIZE 65536 +#endif + +/* Size of size class related tables */ +#ifdef JEMALLOC_PROF +# define PROF_SC_NSIZES SC_NSIZES +#else +/* Minimize memory bloat for non-prof builds. */ +# define PROF_SC_NSIZES 1 +#endif /* Size of stack-allocated buffer used by prof_printf(). */ #define PROF_PRINTF_BUFSIZE 128 @@ -45,12 +62,14 @@ */ #define PROF_NTDATA_LOCKS 256 -/* - * prof_tdata pointers close to NULL are used to encode state information that - * is used for cleaning up during thread shutdown. - */ -#define PROF_TDATA_STATE_REINCARNATED ((prof_tdata_t *)(uintptr_t)1) -#define PROF_TDATA_STATE_PURGATORY ((prof_tdata_t *)(uintptr_t)2) -#define PROF_TDATA_STATE_MAX PROF_TDATA_STATE_PURGATORY +/* Minimize memory bloat for non-prof builds. */ +#ifdef JEMALLOC_PROF +#define PROF_DUMP_FILENAME_LEN (PATH_MAX + 1) +#else +#define PROF_DUMP_FILENAME_LEN 1 +#endif + +/* Default number of recent allocations to record. */ +#define PROF_RECENT_ALLOC_MAX_DEFAULT 0 #endif /* JEMALLOC_INTERNAL_PROF_TYPES_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/psset.h b/contrib/jemalloc/include/jemalloc/internal/psset.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/psset.h @@ -0,0 +1,131 @@ +#ifndef JEMALLOC_INTERNAL_PSSET_H +#define JEMALLOC_INTERNAL_PSSET_H + +#include "jemalloc/internal/hpdata.h" + +/* + * A page-slab set. What the eset is to PAC, the psset is to HPA. It maintains + * a collection of page-slabs (the intent being that they are backed by + * hugepages, or at least could be), and handles allocation and deallocation + * requests. + */ + +/* + * One more than the maximum pszind_t we will serve out of the HPA. + * Practically, we expect only the first few to be actually used. This + * corresponds to a maximum size of of 512MB on systems with 4k pages and + * SC_NGROUP == 4, which is already an unreasonably large maximum. Morally, you + * can think of this as being SC_NPSIZES, but there's no sense in wasting that + * much space in the arena, making bitmaps that much larger, etc. + */ +#define PSSET_NPSIZES 64 + +/* + * We keep two purge lists per page size class; one for hugified hpdatas (at + * index 2*pszind), and one for the non-hugified hpdatas (at index 2*pszind + + * 1). This lets us implement a preference for purging non-hugified hpdatas + * among similarly-dirty ones. + * We reserve the last two indices for empty slabs, in that case purging + * hugified ones (which are definitionally all waste) before non-hugified ones + * (i.e. reversing the order). + */ +#define PSSET_NPURGE_LISTS (2 * PSSET_NPSIZES) + +typedef struct psset_bin_stats_s psset_bin_stats_t; +struct psset_bin_stats_s { + /* How many pageslabs are in this bin? */ + size_t npageslabs; + /* Of them, how many pages are active? */ + size_t nactive; + /* And how many are dirty? */ + size_t ndirty; +}; + +typedef struct psset_stats_s psset_stats_t; +struct psset_stats_s { + /* + * The second index is huge stats; nonfull_slabs[pszind][0] contains + * stats for the non-huge slabs in bucket pszind, while + * nonfull_slabs[pszind][1] contains stats for the huge slabs. + */ + psset_bin_stats_t nonfull_slabs[PSSET_NPSIZES][2]; + + /* + * Full slabs don't live in any edata heap, but we still track their + * stats. + */ + psset_bin_stats_t full_slabs[2]; + + /* Empty slabs are similar. */ + psset_bin_stats_t empty_slabs[2]; +}; + +typedef struct psset_s psset_t; +struct psset_s { + /* + * The pageslabs, quantized by the size class of the largest contiguous + * free run of pages in a pageslab. + */ + hpdata_age_heap_t pageslabs[PSSET_NPSIZES]; + /* Bitmap for which set bits correspond to non-empty heaps. */ + fb_group_t pageslab_bitmap[FB_NGROUPS(PSSET_NPSIZES)]; + /* + * The sum of all bin stats in stats. This lets us quickly answer + * queries for the number of dirty, active, and retained pages in the + * entire set. + */ + psset_bin_stats_t merged_stats; + psset_stats_t stats; + /* + * Slabs with no active allocations, but which are allowed to serve new + * allocations. + */ + hpdata_empty_list_t empty; + /* + * Slabs which are available to be purged, ordered by how much we want + * to purge them (with later indices indicating slabs we want to purge + * more). + */ + hpdata_purge_list_t to_purge[PSSET_NPURGE_LISTS]; + /* Bitmap for which set bits correspond to non-empty purge lists. */ + fb_group_t purge_bitmap[FB_NGROUPS(PSSET_NPURGE_LISTS)]; + /* Slabs which are available to be hugified. */ + hpdata_hugify_list_t to_hugify; +}; + +void psset_init(psset_t *psset); +void psset_stats_accum(psset_stats_t *dst, psset_stats_t *src); + +/* + * Begin or end updating the given pageslab's metadata. While the pageslab is + * being updated, it won't be returned from psset_fit calls. + */ +void psset_update_begin(psset_t *psset, hpdata_t *ps); +void psset_update_end(psset_t *psset, hpdata_t *ps); + +/* Analogous to the eset_fit; pick a hpdata to serve the request. */ +hpdata_t *psset_pick_alloc(psset_t *psset, size_t size); +/* Pick one to purge. */ +hpdata_t *psset_pick_purge(psset_t *psset); +/* Pick one to hugify. */ +hpdata_t *psset_pick_hugify(psset_t *psset); + +void psset_insert(psset_t *psset, hpdata_t *ps); +void psset_remove(psset_t *psset, hpdata_t *ps); + +static inline size_t +psset_npageslabs(psset_t *psset) { + return psset->merged_stats.npageslabs; +} + +static inline size_t +psset_nactive(psset_t *psset) { + return psset->merged_stats.nactive; +} + +static inline size_t +psset_ndirty(psset_t *psset) { + return psset->merged_stats.ndirty; +} + +#endif /* JEMALLOC_INTERNAL_PSSET_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/ql.h b/contrib/jemalloc/include/jemalloc/internal/ql.h --- a/contrib/jemalloc/include/jemalloc/internal/ql.h +++ b/contrib/jemalloc/include/jemalloc/internal/ql.h @@ -3,37 +3,85 @@ #include "jemalloc/internal/qr.h" +/* + * A linked-list implementation. + * + * This is built on top of the ring implementation, but that can be viewed as an + * implementation detail (i.e. trying to advance past the tail of the list + * doesn't wrap around). + * + * You define a struct like so: + * typedef strucy my_s my_t; + * struct my_s { + * int data; + * ql_elm(my_t) my_link; + * }; + * + * // We wobble between "list" and "head" for this type; we're now mostly + * // heading towards "list". + * typedef ql_head(my_t) my_list_t; + * + * You then pass a my_list_t * for a_head arguments, a my_t * for a_elm + * arguments, the token "my_link" for a_field arguments, and the token "my_t" + * for a_type arguments. + */ + /* List definitions. */ #define ql_head(a_type) \ struct { \ a_type *qlh_first; \ } +/* Static initializer for an empty list. */ #define ql_head_initializer(a_head) {NULL} +/* The field definition. */ #define ql_elm(a_type) qr(a_type) -/* List functions. */ +/* A pointer to the first element in the list, or NULL if the list is empty. */ +#define ql_first(a_head) ((a_head)->qlh_first) + +/* Dynamically initializes a list. */ #define ql_new(a_head) do { \ - (a_head)->qlh_first = NULL; \ + ql_first(a_head) = NULL; \ } while (0) -#define ql_elm_new(a_elm, a_field) qr_new((a_elm), a_field) +/* + * Sets dest to be the contents of src (overwriting any elements there), leaving + * src empty. + */ +#define ql_move(a_head_dest, a_head_src) do { \ + ql_first(a_head_dest) = ql_first(a_head_src); \ + ql_new(a_head_src); \ +} while (0) -#define ql_first(a_head) ((a_head)->qlh_first) +/* True if the list is empty, otherwise false. */ +#define ql_empty(a_head) (ql_first(a_head) == NULL) + +/* + * Initializes a ql_elm. Must be called even if the field is about to be + * overwritten. + */ +#define ql_elm_new(a_elm, a_field) qr_new((a_elm), a_field) +/* + * Obtains the last item in the list. + */ #define ql_last(a_head, a_field) \ - ((ql_first(a_head) != NULL) \ - ? qr_prev(ql_first(a_head), a_field) : NULL) + (ql_empty(a_head) ? NULL : qr_prev(ql_first(a_head), a_field)) +/* + * Gets a pointer to the next/prev element in the list. Trying to advance past + * the end or retreat before the beginning of the list returns NULL. + */ #define ql_next(a_head, a_elm, a_field) \ ((ql_last(a_head, a_field) != (a_elm)) \ ? qr_next((a_elm), a_field) : NULL) - #define ql_prev(a_head, a_elm, a_field) \ ((ql_first(a_head) != (a_elm)) ? qr_prev((a_elm), a_field) \ : NULL) +/* Inserts a_elm before a_qlelm in the list. */ #define ql_before_insert(a_head, a_qlelm, a_elm, a_field) do { \ qr_before_insert((a_qlelm), (a_elm), a_field); \ if (ql_first(a_head) == (a_qlelm)) { \ @@ -41,23 +89,41 @@ } \ } while (0) +/* Inserts a_elm after a_qlelm in the list. */ #define ql_after_insert(a_qlelm, a_elm, a_field) \ qr_after_insert((a_qlelm), (a_elm), a_field) +/* Inserts a_elm as the first item in the list. */ #define ql_head_insert(a_head, a_elm, a_field) do { \ - if (ql_first(a_head) != NULL) { \ + if (!ql_empty(a_head)) { \ qr_before_insert(ql_first(a_head), (a_elm), a_field); \ } \ ql_first(a_head) = (a_elm); \ } while (0) +/* Inserts a_elm as the last item in the list. */ #define ql_tail_insert(a_head, a_elm, a_field) do { \ - if (ql_first(a_head) != NULL) { \ + if (!ql_empty(a_head)) { \ qr_before_insert(ql_first(a_head), (a_elm), a_field); \ } \ ql_first(a_head) = qr_next((a_elm), a_field); \ } while (0) +/* + * Given lists a = [a_1, ..., a_n] and [b_1, ..., b_n], results in: + * a = [a1, ..., a_n, b_1, ..., b_n] and b = []. + */ +#define ql_concat(a_head_a, a_head_b, a_field) do { \ + if (ql_empty(a_head_a)) { \ + ql_move(a_head_a, a_head_b); \ + } else if (!ql_empty(a_head_b)) { \ + qr_meld(ql_first(a_head_a), ql_first(a_head_b), \ + a_field); \ + ql_new(a_head_b); \ + } \ +} while (0) + +/* Removes a_elm from the list. */ #define ql_remove(a_head, a_elm, a_field) do { \ if (ql_first(a_head) == (a_elm)) { \ ql_first(a_head) = qr_next(ql_first(a_head), a_field); \ @@ -65,20 +131,63 @@ if (ql_first(a_head) != (a_elm)) { \ qr_remove((a_elm), a_field); \ } else { \ - ql_first(a_head) = NULL; \ + ql_new(a_head); \ } \ } while (0) +/* Removes the first item in the list. */ #define ql_head_remove(a_head, a_type, a_field) do { \ a_type *t = ql_first(a_head); \ ql_remove((a_head), t, a_field); \ } while (0) +/* Removes the last item in the list. */ #define ql_tail_remove(a_head, a_type, a_field) do { \ a_type *t = ql_last(a_head, a_field); \ ql_remove((a_head), t, a_field); \ } while (0) +/* + * Given a = [a_1, a_2, ..., a_n-1, a_n, a_n+1, ...], + * ql_split(a, a_n, b, some_field) results in + * a = [a_1, a_2, ..., a_n-1] + * and replaces b's contents with: + * b = [a_n, a_n+1, ...] + */ +#define ql_split(a_head_a, a_elm, a_head_b, a_field) do { \ + if (ql_first(a_head_a) == (a_elm)) { \ + ql_move(a_head_b, a_head_a); \ + } else { \ + qr_split(ql_first(a_head_a), (a_elm), a_field); \ + ql_first(a_head_b) = (a_elm); \ + } \ +} while (0) + +/* + * An optimized version of: + * a_type *t = ql_first(a_head); + * ql_remove((a_head), t, a_field); + * ql_tail_insert((a_head), t, a_field); + */ +#define ql_rotate(a_head, a_field) do { \ + ql_first(a_head) = qr_next(ql_first(a_head), a_field); \ +} while (0) + +/* + * Helper macro to iterate over each element in a list in order, starting from + * the head (or in reverse order, starting from the tail). The usage is + * (assuming my_t and my_list_t defined as above). + * + * int sum(my_list_t *list) { + * int sum = 0; + * my_t *iter; + * ql_foreach(iter, list, link) { + * sum += iter->data; + * } + * return sum; + * } + */ + #define ql_foreach(a_var, a_head, a_field) \ qr_foreach((a_var), ql_first(a_head), a_field) diff --git a/contrib/jemalloc/include/jemalloc/internal/qr.h b/contrib/jemalloc/include/jemalloc/internal/qr.h --- a/contrib/jemalloc/include/jemalloc/internal/qr.h +++ b/contrib/jemalloc/include/jemalloc/internal/qr.h @@ -1,6 +1,21 @@ #ifndef JEMALLOC_INTERNAL_QR_H #define JEMALLOC_INTERNAL_QR_H +/* + * A ring implementation based on an embedded circular doubly-linked list. + * + * You define your struct like so: + * + * typedef struct my_s my_t; + * struct my_s { + * int data; + * qr(my_t) my_link; + * }; + * + * And then pass a my_t * into macros for a_qr arguments, and the token + * "my_link" into a_field fields. + */ + /* Ring definitions. */ #define qr(a_type) \ struct { \ @@ -8,61 +23,114 @@ a_type *qre_prev; \ } -/* Ring functions. */ +/* + * Initialize a qr link. Every link must be initialized before being used, even + * if that initialization is going to be immediately overwritten (say, by being + * passed into an insertion macro). + */ #define qr_new(a_qr, a_field) do { \ (a_qr)->a_field.qre_next = (a_qr); \ (a_qr)->a_field.qre_prev = (a_qr); \ } while (0) +/* + * Go forwards or backwards in the ring. Note that (the ring being circular), this + * always succeeds -- you just keep looping around and around the ring if you + * chase pointers without end. + */ #define qr_next(a_qr, a_field) ((a_qr)->a_field.qre_next) - #define qr_prev(a_qr, a_field) ((a_qr)->a_field.qre_prev) -#define qr_before_insert(a_qrelm, a_qr, a_field) do { \ - (a_qr)->a_field.qre_prev = (a_qrelm)->a_field.qre_prev; \ - (a_qr)->a_field.qre_next = (a_qrelm); \ - (a_qr)->a_field.qre_prev->a_field.qre_next = (a_qr); \ - (a_qrelm)->a_field.qre_prev = (a_qr); \ +/* + * Given two rings: + * a -> a_1 -> ... -> a_n -- + * ^ | + * |------------------------ + * + * b -> b_1 -> ... -> b_n -- + * ^ | + * |------------------------ + * + * Results in the ring: + * a -> a_1 -> ... -> a_n -> b -> b_1 -> ... -> b_n -- + * ^ | + * |-------------------------------------------------| + * + * a_qr_a can directly be a qr_next() macro, but a_qr_b cannot. + */ +#define qr_meld(a_qr_a, a_qr_b, a_field) do { \ + (a_qr_b)->a_field.qre_prev->a_field.qre_next = \ + (a_qr_a)->a_field.qre_prev; \ + (a_qr_a)->a_field.qre_prev = (a_qr_b)->a_field.qre_prev; \ + (a_qr_b)->a_field.qre_prev = \ + (a_qr_b)->a_field.qre_prev->a_field.qre_next; \ + (a_qr_a)->a_field.qre_prev->a_field.qre_next = (a_qr_a); \ + (a_qr_b)->a_field.qre_prev->a_field.qre_next = (a_qr_b); \ } while (0) -#define qr_after_insert(a_qrelm, a_qr, a_field) do { \ - (a_qr)->a_field.qre_next = (a_qrelm)->a_field.qre_next; \ - (a_qr)->a_field.qre_prev = (a_qrelm); \ - (a_qr)->a_field.qre_next->a_field.qre_prev = (a_qr); \ - (a_qrelm)->a_field.qre_next = (a_qr); \ -} while (0) +/* + * Logically, this is just a meld. The intent, though, is that a_qrelm is a + * single-element ring, so that "before" has a more obvious interpretation than + * meld. + */ +#define qr_before_insert(a_qrelm, a_qr, a_field) \ + qr_meld((a_qrelm), (a_qr), a_field) -#define qr_meld(a_qr_a, a_qr_b, a_type, a_field) do { \ - a_type *t; \ - (a_qr_a)->a_field.qre_prev->a_field.qre_next = (a_qr_b); \ - (a_qr_b)->a_field.qre_prev->a_field.qre_next = (a_qr_a); \ - t = (a_qr_a)->a_field.qre_prev; \ - (a_qr_a)->a_field.qre_prev = (a_qr_b)->a_field.qre_prev; \ - (a_qr_b)->a_field.qre_prev = t; \ -} while (0) +/* Ditto, but inserting after rather than before. */ +#define qr_after_insert(a_qrelm, a_qr, a_field) \ + qr_before_insert(qr_next(a_qrelm, a_field), (a_qr), a_field) /* + * Inverts meld; given the ring: + * a -> a_1 -> ... -> a_n -> b -> b_1 -> ... -> b_n -- + * ^ | + * |-------------------------------------------------| + * + * Results in two rings: + * a -> a_1 -> ... -> a_n -- + * ^ | + * |------------------------ + * + * b -> b_1 -> ... -> b_n -- + * ^ | + * |------------------------ + * * qr_meld() and qr_split() are functionally equivalent, so there's no need to * have two copies of the code. */ -#define qr_split(a_qr_a, a_qr_b, a_type, a_field) \ - qr_meld((a_qr_a), (a_qr_b), a_type, a_field) +#define qr_split(a_qr_a, a_qr_b, a_field) \ + qr_meld((a_qr_a), (a_qr_b), a_field) -#define qr_remove(a_qr, a_field) do { \ - (a_qr)->a_field.qre_prev->a_field.qre_next \ - = (a_qr)->a_field.qre_next; \ - (a_qr)->a_field.qre_next->a_field.qre_prev \ - = (a_qr)->a_field.qre_prev; \ - (a_qr)->a_field.qre_next = (a_qr); \ - (a_qr)->a_field.qre_prev = (a_qr); \ -} while (0) +/* + * Splits off a_qr from the rest of its ring, so that it becomes a + * single-element ring. + */ +#define qr_remove(a_qr, a_field) \ + qr_split(qr_next(a_qr, a_field), (a_qr), a_field) +/* + * Helper macro to iterate over each element in a ring exactly once, starting + * with a_qr. The usage is (assuming my_t defined as above): + * + * int sum(my_t *item) { + * int sum = 0; + * my_t *iter; + * qr_foreach(iter, item, link) { + * sum += iter->data; + * } + * return sum; + * } + */ #define qr_foreach(var, a_qr, a_field) \ for ((var) = (a_qr); \ (var) != NULL; \ (var) = (((var)->a_field.qre_next != (a_qr)) \ ? (var)->a_field.qre_next : NULL)) +/* + * The same (and with the same usage) as qr_foreach, but in the opposite order, + * ending with a_qr. + */ #define qr_reverse_foreach(var, a_qr, a_field) \ for ((var) = ((a_qr) != NULL) ? qr_prev(a_qr, a_field) : NULL; \ (var) != NULL; \ diff --git a/contrib/jemalloc/include/jemalloc/internal/quantum.h b/contrib/jemalloc/include/jemalloc/internal/quantum.h --- a/contrib/jemalloc/include/jemalloc/internal/quantum.h +++ b/contrib/jemalloc/include/jemalloc/internal/quantum.h @@ -30,11 +30,18 @@ # ifdef __hppa__ # define LG_QUANTUM 4 # endif +# ifdef __loongarch__ +# define LG_QUANTUM 4 +# endif # ifdef __m68k__ # define LG_QUANTUM 3 # endif # ifdef __mips__ -# define LG_QUANTUM 3 +# if defined(__mips_n32) || defined(__mips_n64) +# define LG_QUANTUM 4 +# else +# define LG_QUANTUM 3 +# endif # endif # ifdef __nios2__ # define LG_QUANTUM 3 @@ -61,6 +68,9 @@ # ifdef __le32__ # define LG_QUANTUM 4 # endif +# ifdef __arc__ +# define LG_QUANTUM 3 +# endif # ifndef LG_QUANTUM # error "Unknown minimum alignment for architecture; specify via " "--with-lg-quantum" diff --git a/contrib/jemalloc/include/jemalloc/internal/rb.h b/contrib/jemalloc/include/jemalloc/internal/rb.h --- a/contrib/jemalloc/include/jemalloc/internal/rb.h +++ b/contrib/jemalloc/include/jemalloc/internal/rb.h @@ -1,3 +1,6 @@ +#ifndef JEMALLOC_INTERNAL_RB_H +#define JEMALLOC_INTERNAL_RB_H + /*- ******************************************************************************* * @@ -19,13 +22,19 @@ ******************************************************************************* */ -#ifndef RB_H_ -#define RB_H_ - #ifndef __PGI #define RB_COMPACT #endif +/* + * Each node in the RB tree consumes at least 1 byte of space (for the linkage + * if nothing else, so there are a maximum of sizeof(void *) << 3 rb tree nodes + * in any process (and thus, at most sizeof(void *) << 3 nodes in any rb tree). + * The choice of algorithm bounds the depth of a tree to twice the binary log of + * the number of elements in the tree; the following bound follows. + */ +#define RB_MAX_DEPTH (sizeof(void *) << 4) + #ifdef RB_COMPACT /* Node structure. */ #define rb_node(a_type) \ @@ -159,12 +168,22 @@ rbtn_right_set(a_type, a_field, (r_node), (a_node)); \ } while (0) +#define rb_summarized_only_false(...) +#define rb_summarized_only_true(...) __VA_ARGS__ +#define rb_empty_summarize(a_node, a_lchild, a_rchild) false + /* - * The rb_proto() macro generates function prototypes that correspond to the - * functions generated by an equivalently parameterized call to rb_gen(). + * The rb_proto() and rb_summarized_proto() macros generate function prototypes + * that correspond to the functions generated by an equivalently parameterized + * call to rb_gen() or rb_summarized_gen(), respectively. */ #define rb_proto(a_attr, a_prefix, a_rbt_type, a_type) \ + rb_proto_impl(a_attr, a_prefix, a_rbt_type, a_type, false) +#define rb_summarized_proto(a_attr, a_prefix, a_rbt_type, a_type) \ + rb_proto_impl(a_attr, a_prefix, a_rbt_type, a_type, true) +#define rb_proto_impl(a_attr, a_prefix, a_rbt_type, a_type, \ + a_is_summarized) \ a_attr void \ a_prefix##new(a_rbt_type *rbtree); \ a_attr bool \ @@ -195,31 +214,94 @@ a_type *(*cb)(a_rbt_type *, a_type *, void *), void *arg); \ a_attr void \ a_prefix##destroy(a_rbt_type *rbtree, void (*cb)(a_type *, void *), \ - void *arg); + void *arg); \ +/* Extended API */ \ +rb_summarized_only_##a_is_summarized( \ +a_attr void \ +a_prefix##update_summaries(a_rbt_type *rbtree, a_type *node); \ +a_attr bool \ +a_prefix##empty_filtered(a_rbt_type *rbtree, \ + bool (*filter_node)(void *, a_type *), \ + bool (*filter_subtree)(void *, a_type *), \ + void *filter_ctx); \ +a_attr a_type * \ +a_prefix##first_filtered(a_rbt_type *rbtree, \ + bool (*filter_node)(void *, a_type *), \ + bool (*filter_subtree)(void *, a_type *), \ + void *filter_ctx); \ +a_attr a_type * \ +a_prefix##last_filtered(a_rbt_type *rbtree, \ + bool (*filter_node)(void *, a_type *), \ + bool (*filter_subtree)(void *, a_type *), \ + void *filter_ctx); \ +a_attr a_type * \ +a_prefix##next_filtered(a_rbt_type *rbtree, a_type *node, \ + bool (*filter_node)(void *, a_type *), \ + bool (*filter_subtree)(void *, a_type *), \ + void *filter_ctx); \ +a_attr a_type * \ +a_prefix##prev_filtered(a_rbt_type *rbtree, a_type *node, \ + bool (*filter_node)(void *, a_type *), \ + bool (*filter_subtree)(void *, a_type *), \ + void *filter_ctx); \ +a_attr a_type * \ +a_prefix##search_filtered(a_rbt_type *rbtree, const a_type *key, \ + bool (*filter_node)(void *, a_type *), \ + bool (*filter_subtree)(void *, a_type *), \ + void *filter_ctx); \ +a_attr a_type * \ +a_prefix##nsearch_filtered(a_rbt_type *rbtree, const a_type *key, \ + bool (*filter_node)(void *, a_type *), \ + bool (*filter_subtree)(void *, a_type *), \ + void *filter_ctx); \ +a_attr a_type * \ +a_prefix##psearch_filtered(a_rbt_type *rbtree, const a_type *key, \ + bool (*filter_node)(void *, a_type *), \ + bool (*filter_subtree)(void *, a_type *), \ + void *filter_ctx); \ +a_attr a_type * \ +a_prefix##iter_filtered(a_rbt_type *rbtree, a_type *start, \ + a_type *(*cb)(a_rbt_type *, a_type *, void *), void *arg, \ + bool (*filter_node)(void *, a_type *), \ + bool (*filter_subtree)(void *, a_type *), \ + void *filter_ctx); \ +a_attr a_type * \ +a_prefix##reverse_iter_filtered(a_rbt_type *rbtree, a_type *start, \ + a_type *(*cb)(a_rbt_type *, a_type *, void *), void *arg, \ + bool (*filter_node)(void *, a_type *), \ + bool (*filter_subtree)(void *, a_type *), \ + void *filter_ctx); \ +) /* * The rb_gen() macro generates a type-specific red-black tree implementation, * based on the above cpp macros. - * * Arguments: * - * a_attr : Function attribute for generated functions (ex: static). - * a_prefix : Prefix for generated functions (ex: ex_). - * a_rb_type : Type for red-black tree data structure (ex: ex_t). - * a_type : Type for red-black tree node data structure (ex: ex_node_t). - * a_field : Name of red-black tree node linkage (ex: ex_link). - * a_cmp : Node comparison function name, with the following prototype: - * int (a_cmp *)(a_type *a_node, a_type *a_other); - * ^^^^^^ - * or a_key - * Interpretation of comparison function return values: - * -1 : a_node < a_other - * 0 : a_node == a_other - * 1 : a_node > a_other - * In all cases, the a_node or a_key macro argument is the first - * argument to the comparison function, which makes it possible - * to write comparison functions that treat the first argument - * specially. + * a_attr: + * Function attribute for generated functions (ex: static). + * a_prefix: + * Prefix for generated functions (ex: ex_). + * a_rb_type: + * Type for red-black tree data structure (ex: ex_t). + * a_type: + * Type for red-black tree node data structure (ex: ex_node_t). + * a_field: + * Name of red-black tree node linkage (ex: ex_link). + * a_cmp: + * Node comparison function name, with the following prototype: + * + * int a_cmp(a_type *a_node, a_type *a_other); + * ^^^^^^ + * or a_key + * Interpretation of comparison function return values: + * -1 : a_node < a_other + * 0 : a_node == a_other + * 1 : a_node > a_other + * In all cases, the a_node or a_key macro argument is the first argument to + * the comparison function, which makes it possible to write comparison + * functions that treat the first argument specially. a_cmp must be a total + * order on values inserted into the tree -- duplicates are not allowed. * * Assuming the following setup: * @@ -338,8 +420,193 @@ * during iteration. There is no way to stop iteration once it * has begun. * arg : Opaque pointer passed to cb(). + * + * The rb_summarized_gen() macro generates all the functions above, but has an + * expanded interface. In introduces the notion of summarizing subtrees, and of + * filtering searches in the tree according to the information contained in + * those summaries. + * The extra macro argument is: + * a_summarize: + * Tree summarization function name, with the following prototype: + * + * bool a_summarize(a_type *a_node, const a_type *a_left_child, + * const a_type *a_right_child); + * + * This function should update a_node with the summary of the subtree rooted + * there, using the data contained in it and the summaries in a_left_child + * and a_right_child. One or both of them may be NULL. When the tree + * changes due to an insertion or removal, it updates the summaries of all + * nodes whose subtrees have changed (always updating the summaries of + * children before their parents). If the user alters a node in the tree in + * a way that may change its summary, they can call the generated + * update_summaries function to bubble up the summary changes to the root. + * It should return true if the summary changed (or may have changed), and + * false if it didn't (which will allow the implementation to terminate + * "bubbling up" the summaries early). + * As the parameter names indicate, the children are ordered as they are in + * the tree, a_left_child, if it is not NULL, compares less than a_node, + * which in turn compares less than a_right_child (if a_right_child is not + * NULL). + * + * Using the same setup as above but replacing the macro with + * rb_summarized_gen(static, ex_, ex_t, ex_node_t, ex_link, ex_cmp, + * ex_summarize) + * + * Generates all the previous functions, but adds some more: + * + * static void + * ex_update_summaries(ex_t *tree, ex_node_t *node); + * Description: Recompute all summaries of ancestors of node. + * Args: + * tree: Pointer to an initialized red-black tree object. + * node: The element of the tree whose summary may have changed. + * + * For each of ex_empty, ex_first, ex_last, ex_next, ex_prev, ex_search, + * ex_nsearch, ex_psearch, ex_iter, and ex_reverse_iter, an additional function + * is generated as well, with the suffix _filtered (e.g. ex_empty_filtered, + * ex_first_filtered, etc.). These use the concept of a "filter"; a binary + * property some node either satisfies or does not satisfy. Clever use of the + * a_summary argument to rb_summarized_gen can allow efficient computation of + * these predicates across whole subtrees of the tree. + * The extended API functions accept three additional arguments after the + * arguments to the corresponding non-extended equivalent. + * + * ex_fn(..., bool (*filter_node)(void *, ex_node_t *), + * bool (*filter_subtree)(void *, ex_node_t *), void *filter_ctx); + * filter_node : Returns true if the node passes the filter. + * filter_subtree : Returns true if some node in the subtree rooted at + * node passes the filter. + * filter_ctx : A context argument passed to the filters. + * + * For a more concrete example of summarizing and filtering, suppose we're using + * the red-black tree to track a set of integers: + * + * struct ex_node_s { + * rb_node(ex_node_t) ex_link; + * unsigned data; + * }; + * + * Suppose, for some application-specific reason, we want to be able to quickly + * find numbers in the set which are divisible by large powers of 2 (say, for + * aligned allocation purposes). We augment the node with a summary field: + * + * struct ex_node_s { + * rb_node(ex_node_t) ex_link; + * unsigned data; + * unsigned max_subtree_ffs; + * } + * + * and define our summarization function as follows: + * + * bool + * ex_summarize(ex_node_t *node, const ex_node_t *lchild, + * const ex_node_t *rchild) { + * unsigned new_max_subtree_ffs = ffs(node->data); + * if (lchild != NULL && lchild->max_subtree_ffs > new_max_subtree_ffs) { + * new_max_subtree_ffs = lchild->max_subtree_ffs; + * } + * if (rchild != NULL && rchild->max_subtree_ffs > new_max_subtree_ffs) { + * new_max_subtree_ffs = rchild->max_subtree_ffs; + * } + * bool changed = (node->max_subtree_ffs != new_max_subtree_ffs) + * node->max_subtree_ffs = new_max_subtree_ffs; + * // This could be "return true" without any correctness or big-O + * // performance changes; but practically, precisely reporting summary + * // changes reduces the amount of work that has to be done when "bubbling + * // up" summary changes. + * return changed; + * } + * + * We can now implement our filter functions as follows: + * bool + * ex_filter_node(void *filter_ctx, ex_node_t *node) { + * unsigned required_ffs = *(unsigned *)filter_ctx; + * return ffs(node->data) >= required_ffs; + * } + * bool + * ex_filter_subtree(void *filter_ctx, ex_node_t *node) { + * unsigned required_ffs = *(unsigned *)filter_ctx; + * return node->max_subtree_ffs >= required_ffs; + * } + * + * We can now easily search for, e.g., the smallest integer in the set that's + * divisible by 128: + * ex_node_t * + * find_div_128(ex_tree_t *tree) { + * unsigned min_ffs = 7; + * return ex_first_filtered(tree, &ex_filter_node, &ex_filter_subtree, + * &min_ffs); + * } + * + * We could with similar ease: + * - Fnd the next multiple of 128 in the set that's larger than 12345 (with + * ex_nsearch_filtered) + * - Iterate over just those multiples of 64 that are in the set (with + * ex_iter_filtered) + * - Determine if the set contains any multiples of 1024 (with + * ex_empty_filtered). + * + * Some possibly subtle API notes: + * - The node argument to ex_next_filtered and ex_prev_filtered need not pass + * the filter; it will find the next/prev node that passes the filter. + * - ex_search_filtered will fail even for a node in the tree, if that node does + * not pass the filter. ex_psearch_filtered and ex_nsearch_filtered behave + * similarly; they may return a node larger/smaller than the key, even if a + * node equivalent to the key is in the tree (but does not pass the filter). + * - Similarly, if the start argument to a filtered iteration function does not + * pass the filter, the callback won't be invoked on it. + * + * These should make sense after a moment's reflection; each post-condition is + * the same as with the unfiltered version, with the added constraint that the + * returned node must pass the filter. */ #define rb_gen(a_attr, a_prefix, a_rbt_type, a_type, a_field, a_cmp) \ + rb_gen_impl(a_attr, a_prefix, a_rbt_type, a_type, a_field, a_cmp, \ + rb_empty_summarize, false) +#define rb_summarized_gen(a_attr, a_prefix, a_rbt_type, a_type, \ + a_field, a_cmp, a_summarize) \ + rb_gen_impl(a_attr, a_prefix, a_rbt_type, a_type, a_field, a_cmp, \ + a_summarize, true) + +#define rb_gen_impl(a_attr, a_prefix, a_rbt_type, a_type, \ + a_field, a_cmp, a_summarize, a_is_summarized) \ +typedef struct { \ + a_type *node; \ + int cmp; \ +} a_prefix##path_entry_t; \ +static inline void \ +a_prefix##summarize_range(a_prefix##path_entry_t *rfirst, \ + a_prefix##path_entry_t *rlast) { \ + while ((uintptr_t)rlast >= (uintptr_t)rfirst) { \ + a_type *node = rlast->node; \ + /* Avoid a warning when a_summarize is rb_empty_summarize. */ \ + (void)node; \ + bool changed = a_summarize(node, rbtn_left_get(a_type, a_field, \ + node), rbtn_right_get(a_type, a_field, node)); \ + if (!changed) { \ + break; \ + } \ + rlast--; \ + } \ +} \ +/* On the remove pathways, we sometimes swap the node being removed */\ +/* and its first successor; in such cases we need to do two range */\ +/* updates; one from the node to its (former) swapped successor, the */\ +/* next from that successor to the root (with either allowed to */\ +/* bail out early if appropriate. */\ +static inline void \ +a_prefix##summarize_swapped_range(a_prefix##path_entry_t *rfirst, \ + a_prefix##path_entry_t *rlast, a_prefix##path_entry_t *swap_loc) { \ + if (swap_loc == NULL || rlast <= swap_loc) { \ + a_prefix##summarize_range(rfirst, rlast); \ + } else { \ + a_prefix##summarize_range(swap_loc + 1, rlast); \ + (void)a_summarize(swap_loc->node, \ + rbtn_left_get(a_type, a_field, swap_loc->node), \ + rbtn_right_get(a_type, a_field, swap_loc->node)); \ + a_prefix##summarize_range(rfirst, swap_loc - 1); \ + } \ +} \ a_attr void \ a_prefix##new(a_rbt_type *rbtree) { \ rb_new(a_type, a_field, rbtree); \ @@ -465,10 +732,8 @@ } \ a_attr void \ a_prefix##insert(a_rbt_type *rbtree, a_type *node) { \ - struct { \ - a_type *node; \ - int cmp; \ - } path[sizeof(void *) << 4], *pathp; \ + a_prefix##path_entry_t path[RB_MAX_DEPTH]; \ + a_prefix##path_entry_t *pathp; \ rbt_node_new(a_type, a_field, rbtree, node); \ /* Wind. */ \ path->node = rbtree->rbt_root; \ @@ -484,6 +749,13 @@ } \ } \ pathp->node = node; \ + /* A loop invariant we maintain is that all nodes with */\ + /* out-of-date summaries live in path[0], path[1], ..., *pathp. */\ + /* To maintain this, we have to summarize node, since we */\ + /* decrement pathp before the first iteration. */\ + assert(rbtn_left_get(a_type, a_field, node) == NULL); \ + assert(rbtn_right_get(a_type, a_field, node) == NULL); \ + (void)a_summarize(node, NULL, NULL); \ /* Unwind. */ \ for (pathp--; (uintptr_t)pathp >= (uintptr_t)path; pathp--) { \ a_type *cnode = pathp->node; \ @@ -498,9 +770,13 @@ a_type *tnode; \ rbtn_black_set(a_type, a_field, leftleft); \ rbtn_rotate_right(a_type, a_field, cnode, tnode); \ + (void)a_summarize(cnode, \ + rbtn_left_get(a_type, a_field, cnode), \ + rbtn_right_get(a_type, a_field, cnode)); \ cnode = tnode; \ } \ } else { \ + a_prefix##summarize_range(path, pathp); \ return; \ } \ } else { \ @@ -521,13 +797,20 @@ rbtn_rotate_left(a_type, a_field, cnode, tnode); \ rbtn_color_set(a_type, a_field, tnode, tred); \ rbtn_red_set(a_type, a_field, cnode); \ + (void)a_summarize(cnode, \ + rbtn_left_get(a_type, a_field, cnode), \ + rbtn_right_get(a_type, a_field, cnode)); \ cnode = tnode; \ } \ } else { \ + a_prefix##summarize_range(path, pathp); \ return; \ } \ } \ pathp->node = cnode; \ + (void)a_summarize(cnode, \ + rbtn_left_get(a_type, a_field, cnode), \ + rbtn_right_get(a_type, a_field, cnode)); \ } \ /* Set root, and make it black. */ \ rbtree->rbt_root = path->node; \ @@ -535,12 +818,18 @@ } \ a_attr void \ a_prefix##remove(a_rbt_type *rbtree, a_type *node) { \ - struct { \ - a_type *node; \ - int cmp; \ - } *pathp, *nodep, path[sizeof(void *) << 4]; \ + a_prefix##path_entry_t path[RB_MAX_DEPTH]; \ + a_prefix##path_entry_t *pathp; \ + a_prefix##path_entry_t *nodep; \ + a_prefix##path_entry_t *swap_loc; \ + /* This is a "real" sentinel -- NULL means we didn't swap the */\ + /* node to be pruned with one of its successors, and so */\ + /* summarization can terminate early whenever some summary */\ + /* doesn't change. */\ + swap_loc = NULL; \ + /* This is just to silence a compiler warning. */ \ + nodep = NULL; \ /* Wind. */ \ - nodep = NULL; /* Silence compiler warning. */ \ path->node = rbtree->rbt_root; \ for (pathp = path; pathp->node != NULL; pathp++) { \ int cmp = pathp->cmp = a_cmp(node, pathp->node); \ @@ -567,6 +856,7 @@ pathp--; \ if (pathp->node != node) { \ /* Swap node with its successor. */ \ + swap_loc = nodep; \ bool tred = rbtn_red_get(a_type, a_field, pathp->node); \ rbtn_color_set(a_type, a_field, pathp->node, \ rbtn_red_get(a_type, a_field, node)); \ @@ -604,6 +894,9 @@ rbtn_black_set(a_type, a_field, left); \ if (pathp == path) { \ rbtree->rbt_root = left; \ + /* Nothing to summarize -- the subtree rooted at the */\ + /* node's left child hasn't changed, and it's now the */\ + /* root. */\ } else { \ if (pathp[-1].cmp < 0) { \ rbtn_left_set(a_type, a_field, pathp[-1].node, \ @@ -612,6 +905,8 @@ rbtn_right_set(a_type, a_field, pathp[-1].node, \ left); \ } \ + a_prefix##summarize_swapped_range(path, &pathp[-1], \ + swap_loc); \ } \ return; \ } else if (pathp == path) { \ @@ -620,10 +915,15 @@ return; \ } \ } \ + /* We've now established the invariant that the node has no right */\ + /* child (well, morally; we didn't bother nulling it out if we */\ + /* swapped it with its successor), and that the only nodes with */\ + /* out-of-date summaries live in path[0], path[1], ..., pathp[-1].*/\ if (rbtn_red_get(a_type, a_field, pathp->node)) { \ /* Prune red node, which requires no fixup. */ \ assert(pathp[-1].cmp < 0); \ rbtn_left_set(a_type, a_field, pathp[-1].node, NULL); \ + a_prefix##summarize_swapped_range(path, &pathp[-1], swap_loc); \ return; \ } \ /* The node to be pruned is black, so unwind until balance is */\ @@ -657,6 +957,12 @@ rbtn_right_set(a_type, a_field, pathp->node, tnode);\ rbtn_rotate_left(a_type, a_field, pathp->node, \ tnode); \ + (void)a_summarize(pathp->node, \ + rbtn_left_get(a_type, a_field, pathp->node), \ + rbtn_right_get(a_type, a_field, pathp->node)); \ + (void)a_summarize(right, \ + rbtn_left_get(a_type, a_field, right), \ + rbtn_right_get(a_type, a_field, right)); \ } else { \ /* || */\ /* pathp(r) */\ @@ -667,7 +973,12 @@ /* */\ rbtn_rotate_left(a_type, a_field, pathp->node, \ tnode); \ + (void)a_summarize(pathp->node, \ + rbtn_left_get(a_type, a_field, pathp->node), \ + rbtn_right_get(a_type, a_field, pathp->node)); \ } \ + (void)a_summarize(tnode, rbtn_left_get(a_type, a_field, \ + tnode), rbtn_right_get(a_type, a_field, tnode)); \ /* Balance restored, but rotation modified subtree */\ /* root. */\ assert((uintptr_t)pathp > (uintptr_t)path); \ @@ -678,6 +989,8 @@ rbtn_right_set(a_type, a_field, pathp[-1].node, \ tnode); \ } \ + a_prefix##summarize_swapped_range(path, &pathp[-1], \ + swap_loc); \ return; \ } else { \ a_type *right = rbtn_right_get(a_type, a_field, \ @@ -698,6 +1011,15 @@ rbtn_right_set(a_type, a_field, pathp->node, tnode);\ rbtn_rotate_left(a_type, a_field, pathp->node, \ tnode); \ + (void)a_summarize(pathp->node, \ + rbtn_left_get(a_type, a_field, pathp->node), \ + rbtn_right_get(a_type, a_field, pathp->node)); \ + (void)a_summarize(right, \ + rbtn_left_get(a_type, a_field, right), \ + rbtn_right_get(a_type, a_field, right)); \ + (void)a_summarize(tnode, \ + rbtn_left_get(a_type, a_field, tnode), \ + rbtn_right_get(a_type, a_field, tnode)); \ /* Balance restored, but rotation modified */\ /* subtree root, which may actually be the tree */\ /* root. */\ @@ -712,6 +1034,8 @@ rbtn_right_set(a_type, a_field, \ pathp[-1].node, tnode); \ } \ + a_prefix##summarize_swapped_range(path, \ + &pathp[-1], swap_loc); \ } \ return; \ } else { \ @@ -725,6 +1049,12 @@ rbtn_red_set(a_type, a_field, pathp->node); \ rbtn_rotate_left(a_type, a_field, pathp->node, \ tnode); \ + (void)a_summarize(pathp->node, \ + rbtn_left_get(a_type, a_field, pathp->node), \ + rbtn_right_get(a_type, a_field, pathp->node)); \ + (void)a_summarize(tnode, \ + rbtn_left_get(a_type, a_field, tnode), \ + rbtn_right_get(a_type, a_field, tnode)); \ pathp->node = tnode; \ } \ } \ @@ -757,6 +1087,12 @@ tnode); \ rbtn_right_set(a_type, a_field, unode, tnode); \ rbtn_rotate_left(a_type, a_field, unode, tnode); \ + (void)a_summarize(pathp->node, \ + rbtn_left_get(a_type, a_field, pathp->node), \ + rbtn_right_get(a_type, a_field, pathp->node)); \ + (void)a_summarize(unode, \ + rbtn_left_get(a_type, a_field, unode), \ + rbtn_right_get(a_type, a_field, unode)); \ } else { \ /* || */\ /* pathp(b) */\ @@ -771,7 +1107,13 @@ rbtn_rotate_right(a_type, a_field, pathp->node, \ tnode); \ rbtn_black_set(a_type, a_field, tnode); \ + (void)a_summarize(pathp->node, \ + rbtn_left_get(a_type, a_field, pathp->node), \ + rbtn_right_get(a_type, a_field, pathp->node)); \ } \ + (void)a_summarize(tnode, \ + rbtn_left_get(a_type, a_field, tnode), \ + rbtn_right_get(a_type, a_field, tnode)); \ /* Balance restored, but rotation modified subtree */\ /* root, which may actually be the tree root. */\ if (pathp == path) { \ @@ -785,6 +1127,8 @@ rbtn_right_set(a_type, a_field, pathp[-1].node, \ tnode); \ } \ + a_prefix##summarize_swapped_range(path, &pathp[-1], \ + swap_loc); \ } \ return; \ } else if (rbtn_red_get(a_type, a_field, pathp->node)) { \ @@ -803,6 +1147,12 @@ rbtn_black_set(a_type, a_field, leftleft); \ rbtn_rotate_right(a_type, a_field, pathp->node, \ tnode); \ + (void)a_summarize(pathp->node, \ + rbtn_left_get(a_type, a_field, pathp->node), \ + rbtn_right_get(a_type, a_field, pathp->node)); \ + (void)a_summarize(tnode, \ + rbtn_left_get(a_type, a_field, tnode), \ + rbtn_right_get(a_type, a_field, tnode)); \ /* Balance restored, but rotation modified */\ /* subtree root. */\ assert((uintptr_t)pathp > (uintptr_t)path); \ @@ -813,6 +1163,8 @@ rbtn_right_set(a_type, a_field, pathp[-1].node, \ tnode); \ } \ + a_prefix##summarize_swapped_range(path, &pathp[-1], \ + swap_loc); \ return; \ } else { \ /* || */\ @@ -824,6 +1176,8 @@ rbtn_red_set(a_type, a_field, left); \ rbtn_black_set(a_type, a_field, pathp->node); \ /* Balance restored. */ \ + a_prefix##summarize_swapped_range(path, pathp, \ + swap_loc); \ return; \ } \ } else { \ @@ -840,6 +1194,12 @@ rbtn_black_set(a_type, a_field, leftleft); \ rbtn_rotate_right(a_type, a_field, pathp->node, \ tnode); \ + (void)a_summarize(pathp->node, \ + rbtn_left_get(a_type, a_field, pathp->node), \ + rbtn_right_get(a_type, a_field, pathp->node)); \ + (void)a_summarize(tnode, \ + rbtn_left_get(a_type, a_field, tnode), \ + rbtn_right_get(a_type, a_field, tnode)); \ /* Balance restored, but rotation modified */\ /* subtree root, which may actually be the tree */\ /* root. */\ @@ -854,6 +1214,8 @@ rbtn_right_set(a_type, a_field, \ pathp[-1].node, tnode); \ } \ + a_prefix##summarize_swapped_range(path, \ + &pathp[-1], swap_loc); \ } \ return; \ } else { \ @@ -864,6 +1226,9 @@ /* / */\ /* (b) */\ rbtn_red_set(a_type, a_field, left); \ + (void)a_summarize(pathp->node, \ + rbtn_left_get(a_type, a_field, pathp->node), \ + rbtn_right_get(a_type, a_field, pathp->node)); \ } \ } \ } \ @@ -1001,6 +1366,491 @@ void *arg) { \ a_prefix##destroy_recurse(rbtree, rbtree->rbt_root, cb, arg); \ rbtree->rbt_root = NULL; \ -} +} \ +/* BEGIN SUMMARIZED-ONLY IMPLEMENTATION */ \ +rb_summarized_only_##a_is_summarized( \ +static inline a_prefix##path_entry_t * \ +a_prefix##wind(a_rbt_type *rbtree, \ + a_prefix##path_entry_t path[RB_MAX_DEPTH], a_type *node) { \ + a_prefix##path_entry_t *pathp; \ + path->node = rbtree->rbt_root; \ + for (pathp = path; ; pathp++) { \ + assert((size_t)(pathp - path) < RB_MAX_DEPTH); \ + pathp->cmp = a_cmp(node, pathp->node); \ + if (pathp->cmp < 0) { \ + pathp[1].node = rbtn_left_get(a_type, a_field, \ + pathp->node); \ + } else if (pathp->cmp == 0) { \ + return pathp; \ + } else { \ + pathp[1].node = rbtn_right_get(a_type, a_field, \ + pathp->node); \ + } \ + } \ + unreachable(); \ +} \ +a_attr void \ +a_prefix##update_summaries(a_rbt_type *rbtree, a_type *node) { \ + a_prefix##path_entry_t path[RB_MAX_DEPTH]; \ + a_prefix##path_entry_t *pathp = a_prefix##wind(rbtree, path, node); \ + a_prefix##summarize_range(path, pathp); \ +} \ +a_attr bool \ +a_prefix##empty_filtered(a_rbt_type *rbtree, \ + bool (*filter_node)(void *, a_type *), \ + bool (*filter_subtree)(void *, a_type *), \ + void *filter_ctx) { \ + a_type *node = rbtree->rbt_root; \ + return node == NULL || !filter_subtree(filter_ctx, node); \ +} \ +static inline a_type * \ +a_prefix##first_filtered_from_node(a_type *node, \ + bool (*filter_node)(void *, a_type *), \ + bool (*filter_subtree)(void *, a_type *), \ + void *filter_ctx) { \ + assert(node != NULL && filter_subtree(filter_ctx, node)); \ + while (true) { \ + a_type *left = rbtn_left_get(a_type, a_field, node); \ + a_type *right = rbtn_right_get(a_type, a_field, node); \ + if (left != NULL && filter_subtree(filter_ctx, left)) { \ + node = left; \ + } else if (filter_node(filter_ctx, node)) { \ + return node; \ + } else { \ + assert(right != NULL \ + && filter_subtree(filter_ctx, right)); \ + node = right; \ + } \ + } \ + unreachable(); \ +} \ +a_attr a_type * \ +a_prefix##first_filtered(a_rbt_type *rbtree, \ + bool (*filter_node)(void *, a_type *), \ + bool (*filter_subtree)(void *, a_type *), \ + void *filter_ctx) { \ + a_type *node = rbtree->rbt_root; \ + if (node == NULL || !filter_subtree(filter_ctx, node)) { \ + return NULL; \ + } \ + return a_prefix##first_filtered_from_node(node, filter_node, \ + filter_subtree, filter_ctx); \ +} \ +static inline a_type * \ +a_prefix##last_filtered_from_node(a_type *node, \ + bool (*filter_node)(void *, a_type *), \ + bool (*filter_subtree)(void *, a_type *), \ + void *filter_ctx) { \ + assert(node != NULL && filter_subtree(filter_ctx, node)); \ + while (true) { \ + a_type *left = rbtn_left_get(a_type, a_field, node); \ + a_type *right = rbtn_right_get(a_type, a_field, node); \ + if (right != NULL && filter_subtree(filter_ctx, right)) { \ + node = right; \ + } else if (filter_node(filter_ctx, node)) { \ + return node; \ + } else { \ + assert(left != NULL \ + && filter_subtree(filter_ctx, left)); \ + node = left; \ + } \ + } \ + unreachable(); \ +} \ +a_attr a_type * \ +a_prefix##last_filtered(a_rbt_type *rbtree, \ + bool (*filter_node)(void *, a_type *), \ + bool (*filter_subtree)(void *, a_type *), \ + void *filter_ctx) { \ + a_type *node = rbtree->rbt_root; \ + if (node == NULL || !filter_subtree(filter_ctx, node)) { \ + return NULL; \ + } \ + return a_prefix##last_filtered_from_node(node, filter_node, \ + filter_subtree, filter_ctx); \ +} \ +/* Internal implementation function. Search for a node comparing */\ +/* equal to key matching the filter. If such a node is in the tree, */\ +/* return it. Additionally, the caller has the option to ask for */\ +/* bounds on the next / prev node in the tree passing the filter. */\ +/* If nextbound is true, then this function will do one of the */\ +/* following: */\ +/* - Fill in *nextbound_node with the smallest node in the tree */\ +/* greater than key passing the filter, and NULL-out */\ +/* *nextbound_subtree. */\ +/* - Fill in *nextbound_subtree with a parent of that node which is */\ +/* not a parent of the searched-for node, and NULL-out */\ +/* *nextbound_node. */\ +/* - NULL-out both *nextbound_node and *nextbound_subtree, in which */\ +/* case no node greater than key but passing the filter is in the */\ +/* tree. */\ +/* The prevbound case is similar. If the caller knows that key is in */\ +/* the tree and that the subtree rooted at key does not contain a */\ +/* node satisfying the bound being searched for, then they can pass */\ +/* false for include_subtree, in which case we won't bother searching */\ +/* there (risking a cache miss). */\ +/* */\ +/* This API is unfortunately complex; but the logic for filtered */\ +/* searches is very subtle, and otherwise we would have to repeat it */\ +/* multiple times for filtered search, nsearch, psearch, next, and */\ +/* prev. */\ +static inline a_type * \ +a_prefix##search_with_filter_bounds(a_rbt_type *rbtree, \ + const a_type *key, \ + bool (*filter_node)(void *, a_type *), \ + bool (*filter_subtree)(void *, a_type *), \ + void *filter_ctx, \ + bool include_subtree, \ + bool nextbound, a_type **nextbound_node, a_type **nextbound_subtree, \ + bool prevbound, a_type **prevbound_node, a_type **prevbound_subtree) {\ + if (nextbound) { \ + *nextbound_node = NULL; \ + *nextbound_subtree = NULL; \ + } \ + if (prevbound) { \ + *prevbound_node = NULL; \ + *prevbound_subtree = NULL; \ + } \ + a_type *tnode = rbtree->rbt_root; \ + while (tnode != NULL && filter_subtree(filter_ctx, tnode)) { \ + int cmp = a_cmp(key, tnode); \ + a_type *tleft = rbtn_left_get(a_type, a_field, tnode); \ + a_type *tright = rbtn_right_get(a_type, a_field, tnode); \ + if (cmp < 0) { \ + if (nextbound) { \ + if (filter_node(filter_ctx, tnode)) { \ + *nextbound_node = tnode; \ + *nextbound_subtree = NULL; \ + } else if (tright != NULL && filter_subtree( \ + filter_ctx, tright)) { \ + *nextbound_node = NULL; \ + *nextbound_subtree = tright; \ + } \ + } \ + tnode = tleft; \ + } else if (cmp > 0) { \ + if (prevbound) { \ + if (filter_node(filter_ctx, tnode)) { \ + *prevbound_node = tnode; \ + *prevbound_subtree = NULL; \ + } else if (tleft != NULL && filter_subtree( \ + filter_ctx, tleft)) { \ + *prevbound_node = NULL; \ + *prevbound_subtree = tleft; \ + } \ + } \ + tnode = tright; \ + } else { \ + if (filter_node(filter_ctx, tnode)) { \ + return tnode; \ + } \ + if (include_subtree) { \ + if (prevbound && tleft != NULL && filter_subtree( \ + filter_ctx, tleft)) { \ + *prevbound_node = NULL; \ + *prevbound_subtree = tleft; \ + } \ + if (nextbound && tright != NULL && filter_subtree( \ + filter_ctx, tright)) { \ + *nextbound_node = NULL; \ + *nextbound_subtree = tright; \ + } \ + } \ + return NULL; \ + } \ + } \ + return NULL; \ +} \ +a_attr a_type * \ +a_prefix##next_filtered(a_rbt_type *rbtree, a_type *node, \ + bool (*filter_node)(void *, a_type *), \ + bool (*filter_subtree)(void *, a_type *), \ + void *filter_ctx) { \ + a_type *nright = rbtn_right_get(a_type, a_field, node); \ + if (nright != NULL && filter_subtree(filter_ctx, nright)) { \ + return a_prefix##first_filtered_from_node(nright, filter_node, \ + filter_subtree, filter_ctx); \ + } \ + a_type *node_candidate; \ + a_type *subtree_candidate; \ + a_type *search_result = a_prefix##search_with_filter_bounds( \ + rbtree, node, filter_node, filter_subtree, filter_ctx, \ + /* include_subtree */ false, \ + /* nextbound */ true, &node_candidate, &subtree_candidate, \ + /* prevbound */ false, NULL, NULL); \ + assert(node == search_result \ + || !filter_node(filter_ctx, node)); \ + if (node_candidate != NULL) { \ + return node_candidate; \ + } \ + if (subtree_candidate != NULL) { \ + return a_prefix##first_filtered_from_node( \ + subtree_candidate, filter_node, filter_subtree, \ + filter_ctx); \ + } \ + return NULL; \ +} \ +a_attr a_type * \ +a_prefix##prev_filtered(a_rbt_type *rbtree, a_type *node, \ + bool (*filter_node)(void *, a_type *), \ + bool (*filter_subtree)(void *, a_type *), \ + void *filter_ctx) { \ + a_type *nleft = rbtn_left_get(a_type, a_field, node); \ + if (nleft != NULL && filter_subtree(filter_ctx, nleft)) { \ + return a_prefix##last_filtered_from_node(nleft, filter_node, \ + filter_subtree, filter_ctx); \ + } \ + a_type *node_candidate; \ + a_type *subtree_candidate; \ + a_type *search_result = a_prefix##search_with_filter_bounds( \ + rbtree, node, filter_node, filter_subtree, filter_ctx, \ + /* include_subtree */ false, \ + /* nextbound */ false, NULL, NULL, \ + /* prevbound */ true, &node_candidate, &subtree_candidate); \ + assert(node == search_result \ + || !filter_node(filter_ctx, node)); \ + if (node_candidate != NULL) { \ + return node_candidate; \ + } \ + if (subtree_candidate != NULL) { \ + return a_prefix##last_filtered_from_node( \ + subtree_candidate, filter_node, filter_subtree, \ + filter_ctx); \ + } \ + return NULL; \ +} \ +a_attr a_type * \ +a_prefix##search_filtered(a_rbt_type *rbtree, const a_type *key, \ + bool (*filter_node)(void *, a_type *), \ + bool (*filter_subtree)(void *, a_type *), \ + void *filter_ctx) { \ + a_type *result = a_prefix##search_with_filter_bounds(rbtree, key, \ + filter_node, filter_subtree, filter_ctx, \ + /* include_subtree */ false, \ + /* nextbound */ false, NULL, NULL, \ + /* prevbound */ false, NULL, NULL); \ + return result; \ +} \ +a_attr a_type * \ +a_prefix##nsearch_filtered(a_rbt_type *rbtree, const a_type *key, \ + bool (*filter_node)(void *, a_type *), \ + bool (*filter_subtree)(void *, a_type *), \ + void *filter_ctx) { \ + a_type *node_candidate; \ + a_type *subtree_candidate; \ + a_type *result = a_prefix##search_with_filter_bounds(rbtree, key, \ + filter_node, filter_subtree, filter_ctx, \ + /* include_subtree */ true, \ + /* nextbound */ true, &node_candidate, &subtree_candidate, \ + /* prevbound */ false, NULL, NULL); \ + if (result != NULL) { \ + return result; \ + } \ + if (node_candidate != NULL) { \ + return node_candidate; \ + } \ + if (subtree_candidate != NULL) { \ + return a_prefix##first_filtered_from_node( \ + subtree_candidate, filter_node, filter_subtree, \ + filter_ctx); \ + } \ + return NULL; \ +} \ +a_attr a_type * \ +a_prefix##psearch_filtered(a_rbt_type *rbtree, const a_type *key, \ + bool (*filter_node)(void *, a_type *), \ + bool (*filter_subtree)(void *, a_type *), \ + void *filter_ctx) { \ + a_type *node_candidate; \ + a_type *subtree_candidate; \ + a_type *result = a_prefix##search_with_filter_bounds(rbtree, key, \ + filter_node, filter_subtree, filter_ctx, \ + /* include_subtree */ true, \ + /* nextbound */ false, NULL, NULL, \ + /* prevbound */ true, &node_candidate, &subtree_candidate); \ + if (result != NULL) { \ + return result; \ + } \ + if (node_candidate != NULL) { \ + return node_candidate; \ + } \ + if (subtree_candidate != NULL) { \ + return a_prefix##last_filtered_from_node( \ + subtree_candidate, filter_node, filter_subtree, \ + filter_ctx); \ + } \ + return NULL; \ +} \ +a_attr a_type * \ +a_prefix##iter_recurse_filtered(a_rbt_type *rbtree, a_type *node, \ + a_type *(*cb)(a_rbt_type *, a_type *, void *), void *arg, \ + bool (*filter_node)(void *, a_type *), \ + bool (*filter_subtree)(void *, a_type *), \ + void *filter_ctx) { \ + if (node == NULL || !filter_subtree(filter_ctx, node)) { \ + return NULL; \ + } \ + a_type *ret; \ + a_type *left = rbtn_left_get(a_type, a_field, node); \ + a_type *right = rbtn_right_get(a_type, a_field, node); \ + ret = a_prefix##iter_recurse_filtered(rbtree, left, cb, arg, \ + filter_node, filter_subtree, filter_ctx); \ + if (ret != NULL) { \ + return ret; \ + } \ + if (filter_node(filter_ctx, node)) { \ + ret = cb(rbtree, node, arg); \ + } \ + if (ret != NULL) { \ + return ret; \ + } \ + return a_prefix##iter_recurse_filtered(rbtree, right, cb, arg, \ + filter_node, filter_subtree, filter_ctx); \ +} \ +a_attr a_type * \ +a_prefix##iter_start_filtered(a_rbt_type *rbtree, a_type *start, \ + a_type *node, a_type *(*cb)(a_rbt_type *, a_type *, void *), \ + void *arg, bool (*filter_node)(void *, a_type *), \ + bool (*filter_subtree)(void *, a_type *), \ + void *filter_ctx) { \ + if (!filter_subtree(filter_ctx, node)) { \ + return NULL; \ + } \ + int cmp = a_cmp(start, node); \ + a_type *ret; \ + a_type *left = rbtn_left_get(a_type, a_field, node); \ + a_type *right = rbtn_right_get(a_type, a_field, node); \ + if (cmp < 0) { \ + ret = a_prefix##iter_start_filtered(rbtree, start, left, cb, \ + arg, filter_node, filter_subtree, filter_ctx); \ + if (ret != NULL) { \ + return ret; \ + } \ + if (filter_node(filter_ctx, node)) { \ + ret = cb(rbtree, node, arg); \ + if (ret != NULL) { \ + return ret; \ + } \ + } \ + return a_prefix##iter_recurse_filtered(rbtree, right, cb, arg, \ + filter_node, filter_subtree, filter_ctx); \ + } else if (cmp > 0) { \ + return a_prefix##iter_start_filtered(rbtree, start, right, \ + cb, arg, filter_node, filter_subtree, filter_ctx); \ + } else { \ + if (filter_node(filter_ctx, node)) { \ + ret = cb(rbtree, node, arg); \ + if (ret != NULL) { \ + return ret; \ + } \ + } \ + return a_prefix##iter_recurse_filtered(rbtree, right, cb, arg, \ + filter_node, filter_subtree, filter_ctx); \ + } \ +} \ +a_attr a_type * \ +a_prefix##iter_filtered(a_rbt_type *rbtree, a_type *start, \ + a_type *(*cb)(a_rbt_type *, a_type *, void *), void *arg, \ + bool (*filter_node)(void *, a_type *), \ + bool (*filter_subtree)(void *, a_type *), \ + void *filter_ctx) { \ + a_type *ret; \ + if (start != NULL) { \ + ret = a_prefix##iter_start_filtered(rbtree, start, \ + rbtree->rbt_root, cb, arg, filter_node, filter_subtree, \ + filter_ctx); \ + } else { \ + ret = a_prefix##iter_recurse_filtered(rbtree, rbtree->rbt_root, \ + cb, arg, filter_node, filter_subtree, filter_ctx); \ + } \ + return ret; \ +} \ +a_attr a_type * \ +a_prefix##reverse_iter_recurse_filtered(a_rbt_type *rbtree, \ + a_type *node, a_type *(*cb)(a_rbt_type *, a_type *, void *), \ + void *arg, \ + bool (*filter_node)(void *, a_type *), \ + bool (*filter_subtree)(void *, a_type *), \ + void *filter_ctx) { \ + if (node == NULL || !filter_subtree(filter_ctx, node)) { \ + return NULL; \ + } \ + a_type *ret; \ + a_type *left = rbtn_left_get(a_type, a_field, node); \ + a_type *right = rbtn_right_get(a_type, a_field, node); \ + ret = a_prefix##reverse_iter_recurse_filtered(rbtree, right, cb, \ + arg, filter_node, filter_subtree, filter_ctx); \ + if (ret != NULL) { \ + return ret; \ + } \ + if (filter_node(filter_ctx, node)) { \ + ret = cb(rbtree, node, arg); \ + } \ + if (ret != NULL) { \ + return ret; \ + } \ + return a_prefix##reverse_iter_recurse_filtered(rbtree, left, cb, \ + arg, filter_node, filter_subtree, filter_ctx); \ +} \ +a_attr a_type * \ +a_prefix##reverse_iter_start_filtered(a_rbt_type *rbtree, a_type *start,\ + a_type *node, a_type *(*cb)(a_rbt_type *, a_type *, void *), \ + void *arg, bool (*filter_node)(void *, a_type *), \ + bool (*filter_subtree)(void *, a_type *), \ + void *filter_ctx) { \ + if (!filter_subtree(filter_ctx, node)) { \ + return NULL; \ + } \ + int cmp = a_cmp(start, node); \ + a_type *ret; \ + a_type *left = rbtn_left_get(a_type, a_field, node); \ + a_type *right = rbtn_right_get(a_type, a_field, node); \ + if (cmp > 0) { \ + ret = a_prefix##reverse_iter_start_filtered(rbtree, start, \ + right, cb, arg, filter_node, filter_subtree, filter_ctx); \ + if (ret != NULL) { \ + return ret; \ + } \ + if (filter_node(filter_ctx, node)) { \ + ret = cb(rbtree, node, arg); \ + if (ret != NULL) { \ + return ret; \ + } \ + } \ + return a_prefix##reverse_iter_recurse_filtered(rbtree, left, cb,\ + arg, filter_node, filter_subtree, filter_ctx); \ + } else if (cmp < 0) { \ + return a_prefix##reverse_iter_start_filtered(rbtree, start, \ + left, cb, arg, filter_node, filter_subtree, filter_ctx); \ + } else { \ + if (filter_node(filter_ctx, node)) { \ + ret = cb(rbtree, node, arg); \ + if (ret != NULL) { \ + return ret; \ + } \ + } \ + return a_prefix##reverse_iter_recurse_filtered(rbtree, left, cb,\ + arg, filter_node, filter_subtree, filter_ctx); \ + } \ +} \ +a_attr a_type * \ +a_prefix##reverse_iter_filtered(a_rbt_type *rbtree, a_type *start, \ + a_type *(*cb)(a_rbt_type *, a_type *, void *), void *arg, \ + bool (*filter_node)(void *, a_type *), \ + bool (*filter_subtree)(void *, a_type *), \ + void *filter_ctx) { \ + a_type *ret; \ + if (start != NULL) { \ + ret = a_prefix##reverse_iter_start_filtered(rbtree, start, \ + rbtree->rbt_root, cb, arg, filter_node, filter_subtree, \ + filter_ctx); \ + } else { \ + ret = a_prefix##reverse_iter_recurse_filtered(rbtree, \ + rbtree->rbt_root, cb, arg, filter_node, filter_subtree, \ + filter_ctx); \ + } \ + return ret; \ +} \ +) /* end rb_summarized_only */ -#endif /* RB_H_ */ +#endif /* JEMALLOC_INTERNAL_RB_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/rtree.h b/contrib/jemalloc/include/jemalloc/internal/rtree.h --- a/contrib/jemalloc/include/jemalloc/internal/rtree.h +++ b/contrib/jemalloc/include/jemalloc/internal/rtree.h @@ -35,33 +35,52 @@ # define RTREE_LEAF_COMPACT #endif -/* Needed for initialization only. */ -#define RTREE_LEAFKEY_INVALID ((uintptr_t)1) - typedef struct rtree_node_elm_s rtree_node_elm_t; struct rtree_node_elm_s { atomic_p_t child; /* (rtree_{node,leaf}_elm_t *) */ }; +typedef struct rtree_metadata_s rtree_metadata_t; +struct rtree_metadata_s { + szind_t szind; + extent_state_t state; /* Mirrors edata->state. */ + bool is_head; /* Mirrors edata->is_head. */ + bool slab; +}; + +typedef struct rtree_contents_s rtree_contents_t; +struct rtree_contents_s { + edata_t *edata; + rtree_metadata_t metadata; +}; + +#define RTREE_LEAF_STATE_WIDTH EDATA_BITS_STATE_WIDTH +#define RTREE_LEAF_STATE_SHIFT 2 +#define RTREE_LEAF_STATE_MASK MASK(RTREE_LEAF_STATE_WIDTH, RTREE_LEAF_STATE_SHIFT) + struct rtree_leaf_elm_s { #ifdef RTREE_LEAF_COMPACT /* * Single pointer-width field containing all three leaf element fields. * For example, on a 64-bit x64 system with 48 significant virtual - * memory address bits, the index, extent, and slab fields are packed as + * memory address bits, the index, edata, and slab fields are packed as * such: * * x: index - * e: extent + * e: edata + * s: state + * h: is_head * b: slab * - * 00000000 xxxxxxxx eeeeeeee [...] eeeeeeee eeee000b + * 00000000 xxxxxxxx eeeeeeee [...] eeeeeeee e00ssshb */ atomic_p_t le_bits; #else - atomic_p_t le_extent; /* (extent_t *) */ - atomic_u_t le_szind; /* (szind_t) */ - atomic_b_t le_slab; /* (bool) */ + atomic_p_t le_edata; /* (edata_t *) */ + /* + * From high to low bits: szind (8 bits), state (4 bits), is_head, slab + */ + atomic_u_t le_metadata; #endif }; @@ -78,6 +97,7 @@ typedef struct rtree_s rtree_t; struct rtree_s { + base_t *base; malloc_mutex_t init_lock; /* Number of elements based on rtree_levels[0].bits. */ #if RTREE_HEIGHT > 1 @@ -109,42 +129,29 @@ #endif }; -bool rtree_new(rtree_t *rtree, bool zeroed); - -typedef rtree_node_elm_t *(rtree_node_alloc_t)(tsdn_t *, rtree_t *, size_t); -extern rtree_node_alloc_t *JET_MUTABLE rtree_node_alloc; +bool rtree_new(rtree_t *rtree, base_t *base, bool zeroed); -typedef rtree_leaf_elm_t *(rtree_leaf_alloc_t)(tsdn_t *, rtree_t *, size_t); -extern rtree_leaf_alloc_t *JET_MUTABLE rtree_leaf_alloc; - -typedef void (rtree_node_dalloc_t)(tsdn_t *, rtree_t *, rtree_node_elm_t *); -extern rtree_node_dalloc_t *JET_MUTABLE rtree_node_dalloc; - -typedef void (rtree_leaf_dalloc_t)(tsdn_t *, rtree_t *, rtree_leaf_elm_t *); -extern rtree_leaf_dalloc_t *JET_MUTABLE rtree_leaf_dalloc; -#ifdef JEMALLOC_JET -void rtree_delete(tsdn_t *tsdn, rtree_t *rtree); -#endif rtree_leaf_elm_t *rtree_leaf_elm_lookup_hard(tsdn_t *tsdn, rtree_t *rtree, rtree_ctx_t *rtree_ctx, uintptr_t key, bool dependent, bool init_missing); -JEMALLOC_ALWAYS_INLINE uintptr_t -rtree_leafkey(uintptr_t key) { +JEMALLOC_ALWAYS_INLINE unsigned +rtree_leaf_maskbits(void) { unsigned ptrbits = ZU(1) << (LG_SIZEOF_PTR+3); unsigned cumbits = (rtree_levels[RTREE_HEIGHT-1].cumbits - rtree_levels[RTREE_HEIGHT-1].bits); - unsigned maskbits = ptrbits - cumbits; - uintptr_t mask = ~((ZU(1) << maskbits) - 1); + return ptrbits - cumbits; +} + +JEMALLOC_ALWAYS_INLINE uintptr_t +rtree_leafkey(uintptr_t key) { + uintptr_t mask = ~((ZU(1) << rtree_leaf_maskbits()) - 1); return (key & mask); } JEMALLOC_ALWAYS_INLINE size_t rtree_cache_direct_map(uintptr_t key) { - unsigned ptrbits = ZU(1) << (LG_SIZEOF_PTR+3); - unsigned cumbits = (rtree_levels[RTREE_HEIGHT-1].cumbits - - rtree_levels[RTREE_HEIGHT-1].bits); - unsigned maskbits = ptrbits - cumbits; - return (size_t)((key >> maskbits) & (RTREE_CTX_NCACHE - 1)); + return (size_t)((key >> rtree_leaf_maskbits()) & + (RTREE_CTX_NCACHE - 1)); } JEMALLOC_ALWAYS_INLINE uintptr_t @@ -176,151 +183,174 @@ ? ATOMIC_RELAXED : ATOMIC_ACQUIRE); } -JEMALLOC_ALWAYS_INLINE extent_t * -rtree_leaf_elm_bits_extent_get(uintptr_t bits) { +JEMALLOC_ALWAYS_INLINE uintptr_t +rtree_leaf_elm_bits_encode(rtree_contents_t contents) { + assert((uintptr_t)contents.edata % (uintptr_t)EDATA_ALIGNMENT == 0); + uintptr_t edata_bits = (uintptr_t)contents.edata + & (((uintptr_t)1 << LG_VADDR) - 1); + + uintptr_t szind_bits = (uintptr_t)contents.metadata.szind << LG_VADDR; + uintptr_t slab_bits = (uintptr_t)contents.metadata.slab; + uintptr_t is_head_bits = (uintptr_t)contents.metadata.is_head << 1; + uintptr_t state_bits = (uintptr_t)contents.metadata.state << + RTREE_LEAF_STATE_SHIFT; + uintptr_t metadata_bits = szind_bits | state_bits | is_head_bits | + slab_bits; + assert((edata_bits & metadata_bits) == 0); + + return edata_bits | metadata_bits; +} + +JEMALLOC_ALWAYS_INLINE rtree_contents_t +rtree_leaf_elm_bits_decode(uintptr_t bits) { + rtree_contents_t contents; + /* Do the easy things first. */ + contents.metadata.szind = bits >> LG_VADDR; + contents.metadata.slab = (bool)(bits & 1); + contents.metadata.is_head = (bool)(bits & (1 << 1)); + + uintptr_t state_bits = (bits & RTREE_LEAF_STATE_MASK) >> + RTREE_LEAF_STATE_SHIFT; + assert(state_bits <= extent_state_max); + contents.metadata.state = (extent_state_t)state_bits; + + uintptr_t low_bit_mask = ~((uintptr_t)EDATA_ALIGNMENT - 1); # ifdef __aarch64__ /* * aarch64 doesn't sign extend the highest virtual address bit to set - * the higher ones. Instead, the high bits gets zeroed. + * the higher ones. Instead, the high bits get zeroed. */ uintptr_t high_bit_mask = ((uintptr_t)1 << LG_VADDR) - 1; - /* Mask off the slab bit. */ - uintptr_t low_bit_mask = ~(uintptr_t)1; + /* Mask off metadata. */ uintptr_t mask = high_bit_mask & low_bit_mask; - return (extent_t *)(bits & mask); + contents.edata = (edata_t *)(bits & mask); # else - /* Restore sign-extended high bits, mask slab bit. */ - return (extent_t *)((uintptr_t)((intptr_t)(bits << RTREE_NHIB) >> - RTREE_NHIB) & ~((uintptr_t)0x1)); + /* Restore sign-extended high bits, mask metadata bits. */ + contents.edata = (edata_t *)((uintptr_t)((intptr_t)(bits << RTREE_NHIB) + >> RTREE_NHIB) & low_bit_mask); # endif + assert((uintptr_t)contents.edata % (uintptr_t)EDATA_ALIGNMENT == 0); + return contents; } -JEMALLOC_ALWAYS_INLINE szind_t -rtree_leaf_elm_bits_szind_get(uintptr_t bits) { - return (szind_t)(bits >> LG_VADDR); -} - -JEMALLOC_ALWAYS_INLINE bool -rtree_leaf_elm_bits_slab_get(uintptr_t bits) { - return (bool)(bits & (uintptr_t)0x1); -} +# endif /* RTREE_LEAF_COMPACT */ -# endif - -JEMALLOC_ALWAYS_INLINE extent_t * -rtree_leaf_elm_extent_read(tsdn_t *tsdn, rtree_t *rtree, - rtree_leaf_elm_t *elm, bool dependent) { +JEMALLOC_ALWAYS_INLINE rtree_contents_t +rtree_leaf_elm_read(tsdn_t *tsdn, rtree_t *rtree, rtree_leaf_elm_t *elm, + bool dependent) { #ifdef RTREE_LEAF_COMPACT uintptr_t bits = rtree_leaf_elm_bits_read(tsdn, rtree, elm, dependent); - return rtree_leaf_elm_bits_extent_get(bits); + rtree_contents_t contents = rtree_leaf_elm_bits_decode(bits); + return contents; #else - extent_t *extent = (extent_t *)atomic_load_p(&elm->le_extent, dependent + rtree_contents_t contents; + unsigned metadata_bits = atomic_load_u(&elm->le_metadata, dependent ? ATOMIC_RELAXED : ATOMIC_ACQUIRE); - return extent; -#endif -} + contents.metadata.slab = (bool)(metadata_bits & 1); + contents.metadata.is_head = (bool)(metadata_bits & (1 << 1)); -JEMALLOC_ALWAYS_INLINE szind_t -rtree_leaf_elm_szind_read(tsdn_t *tsdn, rtree_t *rtree, - rtree_leaf_elm_t *elm, bool dependent) { -#ifdef RTREE_LEAF_COMPACT - uintptr_t bits = rtree_leaf_elm_bits_read(tsdn, rtree, elm, dependent); - return rtree_leaf_elm_bits_szind_get(bits); -#else - return (szind_t)atomic_load_u(&elm->le_szind, dependent ? ATOMIC_RELAXED - : ATOMIC_ACQUIRE); + uintptr_t state_bits = (metadata_bits & RTREE_LEAF_STATE_MASK) >> + RTREE_LEAF_STATE_SHIFT; + assert(state_bits <= extent_state_max); + contents.metadata.state = (extent_state_t)state_bits; + contents.metadata.szind = metadata_bits >> (RTREE_LEAF_STATE_SHIFT + + RTREE_LEAF_STATE_WIDTH); + + contents.edata = (edata_t *)atomic_load_p(&elm->le_edata, dependent + ? ATOMIC_RELAXED : ATOMIC_ACQUIRE); + + return contents; #endif } -JEMALLOC_ALWAYS_INLINE bool -rtree_leaf_elm_slab_read(tsdn_t *tsdn, rtree_t *rtree, - rtree_leaf_elm_t *elm, bool dependent) { +JEMALLOC_ALWAYS_INLINE void +rtree_contents_encode(rtree_contents_t contents, void **bits, + unsigned *additional) { #ifdef RTREE_LEAF_COMPACT - uintptr_t bits = rtree_leaf_elm_bits_read(tsdn, rtree, elm, dependent); - return rtree_leaf_elm_bits_slab_get(bits); + *bits = (void *)rtree_leaf_elm_bits_encode(contents); #else - return atomic_load_b(&elm->le_slab, dependent ? ATOMIC_RELAXED : - ATOMIC_ACQUIRE); + *additional = (unsigned)contents.metadata.slab + | ((unsigned)contents.metadata.is_head << 1) + | ((unsigned)contents.metadata.state << RTREE_LEAF_STATE_SHIFT) + | ((unsigned)contents.metadata.szind << (RTREE_LEAF_STATE_SHIFT + + RTREE_LEAF_STATE_WIDTH)); + *bits = contents.edata; #endif } -static inline void -rtree_leaf_elm_extent_write(tsdn_t *tsdn, rtree_t *rtree, - rtree_leaf_elm_t *elm, extent_t *extent) { +JEMALLOC_ALWAYS_INLINE void +rtree_leaf_elm_write_commit(tsdn_t *tsdn, rtree_t *rtree, + rtree_leaf_elm_t *elm, void *bits, unsigned additional) { #ifdef RTREE_LEAF_COMPACT - uintptr_t old_bits = rtree_leaf_elm_bits_read(tsdn, rtree, elm, true); - uintptr_t bits = ((uintptr_t)rtree_leaf_elm_bits_szind_get(old_bits) << - LG_VADDR) | ((uintptr_t)extent & (((uintptr_t)0x1 << LG_VADDR) - 1)) - | ((uintptr_t)rtree_leaf_elm_bits_slab_get(old_bits)); - atomic_store_p(&elm->le_bits, (void *)bits, ATOMIC_RELEASE); + atomic_store_p(&elm->le_bits, bits, ATOMIC_RELEASE); #else - atomic_store_p(&elm->le_extent, extent, ATOMIC_RELEASE); + atomic_store_u(&elm->le_metadata, additional, ATOMIC_RELEASE); + /* + * Write edata last, since the element is atomically considered valid + * as soon as the edata field is non-NULL. + */ + atomic_store_p(&elm->le_edata, bits, ATOMIC_RELEASE); #endif } -static inline void -rtree_leaf_elm_szind_write(tsdn_t *tsdn, rtree_t *rtree, - rtree_leaf_elm_t *elm, szind_t szind) { - assert(szind <= SC_NSIZES); +JEMALLOC_ALWAYS_INLINE void +rtree_leaf_elm_write(tsdn_t *tsdn, rtree_t *rtree, + rtree_leaf_elm_t *elm, rtree_contents_t contents) { + assert((uintptr_t)contents.edata % EDATA_ALIGNMENT == 0); + void *bits; + unsigned additional; -#ifdef RTREE_LEAF_COMPACT - uintptr_t old_bits = rtree_leaf_elm_bits_read(tsdn, rtree, elm, - true); - uintptr_t bits = ((uintptr_t)szind << LG_VADDR) | - ((uintptr_t)rtree_leaf_elm_bits_extent_get(old_bits) & - (((uintptr_t)0x1 << LG_VADDR) - 1)) | - ((uintptr_t)rtree_leaf_elm_bits_slab_get(old_bits)); - atomic_store_p(&elm->le_bits, (void *)bits, ATOMIC_RELEASE); -#else - atomic_store_u(&elm->le_szind, szind, ATOMIC_RELEASE); -#endif + rtree_contents_encode(contents, &bits, &additional); + rtree_leaf_elm_write_commit(tsdn, rtree, elm, bits, additional); } -static inline void -rtree_leaf_elm_slab_write(tsdn_t *tsdn, rtree_t *rtree, - rtree_leaf_elm_t *elm, bool slab) { +/* The state field can be updated independently (and more frequently). */ +JEMALLOC_ALWAYS_INLINE void +rtree_leaf_elm_state_update(tsdn_t *tsdn, rtree_t *rtree, + rtree_leaf_elm_t *elm1, rtree_leaf_elm_t *elm2, extent_state_t state) { + assert(elm1 != NULL); #ifdef RTREE_LEAF_COMPACT - uintptr_t old_bits = rtree_leaf_elm_bits_read(tsdn, rtree, elm, - true); - uintptr_t bits = ((uintptr_t)rtree_leaf_elm_bits_szind_get(old_bits) << - LG_VADDR) | ((uintptr_t)rtree_leaf_elm_bits_extent_get(old_bits) & - (((uintptr_t)0x1 << LG_VADDR) - 1)) | ((uintptr_t)slab); - atomic_store_p(&elm->le_bits, (void *)bits, ATOMIC_RELEASE); + uintptr_t bits = rtree_leaf_elm_bits_read(tsdn, rtree, elm1, + /* dependent */ true); + bits &= ~RTREE_LEAF_STATE_MASK; + bits |= state << RTREE_LEAF_STATE_SHIFT; + atomic_store_p(&elm1->le_bits, (void *)bits, ATOMIC_RELEASE); + if (elm2 != NULL) { + atomic_store_p(&elm2->le_bits, (void *)bits, ATOMIC_RELEASE); + } #else - atomic_store_b(&elm->le_slab, slab, ATOMIC_RELEASE); + unsigned bits = atomic_load_u(&elm1->le_metadata, ATOMIC_RELAXED); + bits &= ~RTREE_LEAF_STATE_MASK; + bits |= state << RTREE_LEAF_STATE_SHIFT; + atomic_store_u(&elm1->le_metadata, bits, ATOMIC_RELEASE); + if (elm2 != NULL) { + atomic_store_u(&elm2->le_metadata, bits, ATOMIC_RELEASE); + } #endif } -static inline void -rtree_leaf_elm_write(tsdn_t *tsdn, rtree_t *rtree, - rtree_leaf_elm_t *elm, extent_t *extent, szind_t szind, bool slab) { -#ifdef RTREE_LEAF_COMPACT - uintptr_t bits = ((uintptr_t)szind << LG_VADDR) | - ((uintptr_t)extent & (((uintptr_t)0x1 << LG_VADDR) - 1)) | - ((uintptr_t)slab); - atomic_store_p(&elm->le_bits, (void *)bits, ATOMIC_RELEASE); -#else - rtree_leaf_elm_slab_write(tsdn, rtree, elm, slab); - rtree_leaf_elm_szind_write(tsdn, rtree, elm, szind); - /* - * Write extent last, since the element is atomically considered valid - * as soon as the extent field is non-NULL. - */ - rtree_leaf_elm_extent_write(tsdn, rtree, elm, extent); -#endif -} +/* + * Tries to look up the key in the L1 cache, returning false if there's a hit, or + * true if there's a miss. + * Key is allowed to be NULL; returns true in this case. + */ +JEMALLOC_ALWAYS_INLINE bool +rtree_leaf_elm_lookup_fast(tsdn_t *tsdn, rtree_t *rtree, rtree_ctx_t *rtree_ctx, + uintptr_t key, rtree_leaf_elm_t **elm) { + size_t slot = rtree_cache_direct_map(key); + uintptr_t leafkey = rtree_leafkey(key); + assert(leafkey != RTREE_LEAFKEY_INVALID); -static inline void -rtree_leaf_elm_szind_slab_update(tsdn_t *tsdn, rtree_t *rtree, - rtree_leaf_elm_t *elm, szind_t szind, bool slab) { - assert(!slab || szind < SC_NBINS); + if (unlikely(rtree_ctx->cache[slot].leafkey != leafkey)) { + return true; + } - /* - * The caller implicitly assures that it is the only writer to the szind - * and slab fields, and that the extent field cannot currently change. - */ - rtree_leaf_elm_slab_write(tsdn, rtree, elm, slab); - rtree_leaf_elm_szind_write(tsdn, rtree, elm, szind); + rtree_leaf_elm_t *leaf = rtree_ctx->cache[slot].leaf; + assert(leaf != NULL); + uintptr_t subkey = rtree_subkey(key, RTREE_HEIGHT-1); + *elm = &leaf[subkey]; + + return false; } JEMALLOC_ALWAYS_INLINE rtree_leaf_elm_t * @@ -382,147 +412,143 @@ dependent, init_missing); } +/* + * Returns true on lookup failure. + */ static inline bool -rtree_write(tsdn_t *tsdn, rtree_t *rtree, rtree_ctx_t *rtree_ctx, uintptr_t key, - extent_t *extent, szind_t szind, bool slab) { - /* Use rtree_clear() to set the extent to NULL. */ - assert(extent != NULL); - +rtree_read_independent(tsdn_t *tsdn, rtree_t *rtree, rtree_ctx_t *rtree_ctx, + uintptr_t key, rtree_contents_t *r_contents) { rtree_leaf_elm_t *elm = rtree_leaf_elm_lookup(tsdn, rtree, rtree_ctx, - key, false, true); + key, /* dependent */ false, /* init_missing */ false); if (elm == NULL) { return true; } - - assert(rtree_leaf_elm_extent_read(tsdn, rtree, elm, false) == NULL); - rtree_leaf_elm_write(tsdn, rtree, elm, extent, szind, slab); - + *r_contents = rtree_leaf_elm_read(tsdn, rtree, elm, + /* dependent */ false); return false; } -JEMALLOC_ALWAYS_INLINE rtree_leaf_elm_t * -rtree_read(tsdn_t *tsdn, rtree_t *rtree, rtree_ctx_t *rtree_ctx, uintptr_t key, - bool dependent) { +static inline rtree_contents_t +rtree_read(tsdn_t *tsdn, rtree_t *rtree, rtree_ctx_t *rtree_ctx, + uintptr_t key) { rtree_leaf_elm_t *elm = rtree_leaf_elm_lookup(tsdn, rtree, rtree_ctx, - key, dependent, false); - if (!dependent && elm == NULL) { - return NULL; - } + key, /* dependent */ true, /* init_missing */ false); assert(elm != NULL); - return elm; + return rtree_leaf_elm_read(tsdn, rtree, elm, /* dependent */ true); } -JEMALLOC_ALWAYS_INLINE extent_t * -rtree_extent_read(tsdn_t *tsdn, rtree_t *rtree, rtree_ctx_t *rtree_ctx, - uintptr_t key, bool dependent) { - rtree_leaf_elm_t *elm = rtree_read(tsdn, rtree, rtree_ctx, key, - dependent); - if (!dependent && elm == NULL) { - return NULL; - } - return rtree_leaf_elm_extent_read(tsdn, rtree, elm, dependent); -} - -JEMALLOC_ALWAYS_INLINE szind_t -rtree_szind_read(tsdn_t *tsdn, rtree_t *rtree, rtree_ctx_t *rtree_ctx, - uintptr_t key, bool dependent) { - rtree_leaf_elm_t *elm = rtree_read(tsdn, rtree, rtree_ctx, key, - dependent); - if (!dependent && elm == NULL) { - return SC_NSIZES; - } - return rtree_leaf_elm_szind_read(tsdn, rtree, elm, dependent); +static inline rtree_metadata_t +rtree_metadata_read(tsdn_t *tsdn, rtree_t *rtree, rtree_ctx_t *rtree_ctx, + uintptr_t key) { + rtree_leaf_elm_t *elm = rtree_leaf_elm_lookup(tsdn, rtree, rtree_ctx, + key, /* dependent */ true, /* init_missing */ false); + assert(elm != NULL); + return rtree_leaf_elm_read(tsdn, rtree, elm, + /* dependent */ true).metadata; } /* - * rtree_slab_read() is intentionally omitted because slab is always read in - * conjunction with szind, which makes rtree_szind_slab_read() a better choice. + * Returns true when the request cannot be fulfilled by fastpath. */ - -JEMALLOC_ALWAYS_INLINE bool -rtree_extent_szind_read(tsdn_t *tsdn, rtree_t *rtree, rtree_ctx_t *rtree_ctx, - uintptr_t key, bool dependent, extent_t **r_extent, szind_t *r_szind) { - rtree_leaf_elm_t *elm = rtree_read(tsdn, rtree, rtree_ctx, key, - dependent); - if (!dependent && elm == NULL) { +static inline bool +rtree_metadata_try_read_fast(tsdn_t *tsdn, rtree_t *rtree, rtree_ctx_t *rtree_ctx, + uintptr_t key, rtree_metadata_t *r_rtree_metadata) { + rtree_leaf_elm_t *elm; + /* + * Should check the bool return value (lookup success or not) instead of + * elm == NULL (which will result in an extra branch). This is because + * when the cache lookup succeeds, there will never be a NULL pointer + * returned (which is unknown to the compiler). + */ + if (rtree_leaf_elm_lookup_fast(tsdn, rtree, rtree_ctx, key, &elm)) { return true; } - *r_extent = rtree_leaf_elm_extent_read(tsdn, rtree, elm, dependent); - *r_szind = rtree_leaf_elm_szind_read(tsdn, rtree, elm, dependent); + assert(elm != NULL); + *r_rtree_metadata = rtree_leaf_elm_read(tsdn, rtree, elm, + /* dependent */ true).metadata; return false; } -/* - * Try to read szind_slab from the L1 cache. Returns true on a hit, - * and fills in r_szind and r_slab. Otherwise returns false. - * - * Key is allowed to be NULL in order to save an extra branch on the - * fastpath. returns false in this case. - */ -JEMALLOC_ALWAYS_INLINE bool -rtree_szind_slab_read_fast(tsdn_t *tsdn, rtree_t *rtree, rtree_ctx_t *rtree_ctx, - uintptr_t key, szind_t *r_szind, bool *r_slab) { - rtree_leaf_elm_t *elm; - - size_t slot = rtree_cache_direct_map(key); - uintptr_t leafkey = rtree_leafkey(key); - assert(leafkey != RTREE_LEAFKEY_INVALID); - - if (likely(rtree_ctx->cache[slot].leafkey == leafkey)) { - rtree_leaf_elm_t *leaf = rtree_ctx->cache[slot].leaf; - assert(leaf != NULL); - uintptr_t subkey = rtree_subkey(key, RTREE_HEIGHT-1); - elm = &leaf[subkey]; - -#ifdef RTREE_LEAF_COMPACT - uintptr_t bits = rtree_leaf_elm_bits_read(tsdn, rtree, - elm, true); - *r_szind = rtree_leaf_elm_bits_szind_get(bits); - *r_slab = rtree_leaf_elm_bits_slab_get(bits); -#else - *r_szind = rtree_leaf_elm_szind_read(tsdn, rtree, elm, true); - *r_slab = rtree_leaf_elm_slab_read(tsdn, rtree, elm, true); -#endif - return true; - } else { - return false; +JEMALLOC_ALWAYS_INLINE void +rtree_write_range_impl(tsdn_t *tsdn, rtree_t *rtree, rtree_ctx_t *rtree_ctx, + uintptr_t base, uintptr_t end, rtree_contents_t contents, bool clearing) { + assert((base & PAGE_MASK) == 0 && (end & PAGE_MASK) == 0); + /* + * Only used for emap_(de)register_interior, which implies the + * boundaries have been registered already. Therefore all the lookups + * are dependent w/o init_missing, assuming the range spans across at + * most 2 rtree leaf nodes (each covers 1 GiB of vaddr). + */ + void *bits; + unsigned additional; + rtree_contents_encode(contents, &bits, &additional); + + rtree_leaf_elm_t *elm = NULL; /* Dead store. */ + for (uintptr_t addr = base; addr <= end; addr += PAGE) { + if (addr == base || + (addr & ((ZU(1) << rtree_leaf_maskbits()) - 1)) == 0) { + elm = rtree_leaf_elm_lookup(tsdn, rtree, rtree_ctx, addr, + /* dependent */ true, /* init_missing */ false); + assert(elm != NULL); + } + assert(elm == rtree_leaf_elm_lookup(tsdn, rtree, rtree_ctx, addr, + /* dependent */ true, /* init_missing */ false)); + assert(!clearing || rtree_leaf_elm_read(tsdn, rtree, elm, + /* dependent */ true).edata != NULL); + rtree_leaf_elm_write_commit(tsdn, rtree, elm, bits, additional); + elm++; } } + +JEMALLOC_ALWAYS_INLINE void +rtree_write_range(tsdn_t *tsdn, rtree_t *rtree, rtree_ctx_t *rtree_ctx, + uintptr_t base, uintptr_t end, rtree_contents_t contents) { + rtree_write_range_impl(tsdn, rtree, rtree_ctx, base, end, contents, + /* clearing */ false); +} + JEMALLOC_ALWAYS_INLINE bool -rtree_szind_slab_read(tsdn_t *tsdn, rtree_t *rtree, rtree_ctx_t *rtree_ctx, - uintptr_t key, bool dependent, szind_t *r_szind, bool *r_slab) { - rtree_leaf_elm_t *elm = rtree_read(tsdn, rtree, rtree_ctx, key, - dependent); - if (!dependent && elm == NULL) { +rtree_write(tsdn_t *tsdn, rtree_t *rtree, rtree_ctx_t *rtree_ctx, uintptr_t key, + rtree_contents_t contents) { + rtree_leaf_elm_t *elm = rtree_leaf_elm_lookup(tsdn, rtree, rtree_ctx, + key, /* dependent */ false, /* init_missing */ true); + if (elm == NULL) { return true; } -#ifdef RTREE_LEAF_COMPACT - uintptr_t bits = rtree_leaf_elm_bits_read(tsdn, rtree, elm, dependent); - *r_szind = rtree_leaf_elm_bits_szind_get(bits); - *r_slab = rtree_leaf_elm_bits_slab_get(bits); -#else - *r_szind = rtree_leaf_elm_szind_read(tsdn, rtree, elm, dependent); - *r_slab = rtree_leaf_elm_slab_read(tsdn, rtree, elm, dependent); -#endif - return false; -} -static inline void -rtree_szind_slab_update(tsdn_t *tsdn, rtree_t *rtree, rtree_ctx_t *rtree_ctx, - uintptr_t key, szind_t szind, bool slab) { - assert(!slab || szind < SC_NBINS); + rtree_leaf_elm_write(tsdn, rtree, elm, contents); - rtree_leaf_elm_t *elm = rtree_read(tsdn, rtree, rtree_ctx, key, true); - rtree_leaf_elm_szind_slab_update(tsdn, rtree, elm, szind, slab); + return false; } static inline void rtree_clear(tsdn_t *tsdn, rtree_t *rtree, rtree_ctx_t *rtree_ctx, uintptr_t key) { - rtree_leaf_elm_t *elm = rtree_read(tsdn, rtree, rtree_ctx, key, true); - assert(rtree_leaf_elm_extent_read(tsdn, rtree, elm, false) != - NULL); - rtree_leaf_elm_write(tsdn, rtree, elm, NULL, SC_NSIZES, false); + rtree_leaf_elm_t *elm = rtree_leaf_elm_lookup(tsdn, rtree, rtree_ctx, + key, /* dependent */ true, /* init_missing */ false); + assert(elm != NULL); + assert(rtree_leaf_elm_read(tsdn, rtree, elm, + /* dependent */ true).edata != NULL); + rtree_contents_t contents; + contents.edata = NULL; + contents.metadata.szind = SC_NSIZES; + contents.metadata.slab = false; + contents.metadata.is_head = false; + contents.metadata.state = (extent_state_t)0; + rtree_leaf_elm_write(tsdn, rtree, elm, contents); +} + +static inline void +rtree_clear_range(tsdn_t *tsdn, rtree_t *rtree, rtree_ctx_t *rtree_ctx, + uintptr_t base, uintptr_t end) { + rtree_contents_t contents; + contents.edata = NULL; + contents.metadata.szind = SC_NSIZES; + contents.metadata.slab = false; + contents.metadata.is_head = false; + contents.metadata.state = (extent_state_t)0; + rtree_write_range_impl(tsdn, rtree, rtree_ctx, base, end, contents, + /* clearing */ true); } #endif /* JEMALLOC_INTERNAL_RTREE_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/rtree_tsd.h b/contrib/jemalloc/include/jemalloc/internal/rtree_tsd.h --- a/contrib/jemalloc/include/jemalloc/internal/rtree_tsd.h +++ b/contrib/jemalloc/include/jemalloc/internal/rtree_tsd.h @@ -18,16 +18,28 @@ * cache misses if made overly large, plus the cost of linear search in the LRU * cache. */ -#define RTREE_CTX_LG_NCACHE 4 -#define RTREE_CTX_NCACHE (1 << RTREE_CTX_LG_NCACHE) +#define RTREE_CTX_NCACHE 16 #define RTREE_CTX_NCACHE_L2 8 +/* Needed for initialization only. */ +#define RTREE_LEAFKEY_INVALID ((uintptr_t)1) +#define RTREE_CTX_CACHE_ELM_INVALID {RTREE_LEAFKEY_INVALID, NULL} + +#define RTREE_CTX_INIT_ELM_1 RTREE_CTX_CACHE_ELM_INVALID +#define RTREE_CTX_INIT_ELM_2 RTREE_CTX_INIT_ELM_1, RTREE_CTX_INIT_ELM_1 +#define RTREE_CTX_INIT_ELM_4 RTREE_CTX_INIT_ELM_2, RTREE_CTX_INIT_ELM_2 +#define RTREE_CTX_INIT_ELM_8 RTREE_CTX_INIT_ELM_4, RTREE_CTX_INIT_ELM_4 +#define RTREE_CTX_INIT_ELM_16 RTREE_CTX_INIT_ELM_8, RTREE_CTX_INIT_ELM_8 + +#define _RTREE_CTX_INIT_ELM_DATA(n) RTREE_CTX_INIT_ELM_##n +#define RTREE_CTX_INIT_ELM_DATA(n) _RTREE_CTX_INIT_ELM_DATA(n) + /* - * Zero initializer required for tsd initialization only. Proper initialization - * done via rtree_ctx_data_init(). + * Static initializer (to invalidate the cache entries) is required because the + * free fastpath may access the rtree cache before a full tsd initialization. */ -#define RTREE_CTX_ZERO_INITIALIZER {{{0, 0}}, {{0, 0}}} - +#define RTREE_CTX_INITIALIZER {{RTREE_CTX_INIT_ELM_DATA(RTREE_CTX_NCACHE)}, \ + {RTREE_CTX_INIT_ELM_DATA(RTREE_CTX_NCACHE_L2)}} typedef struct rtree_leaf_elm_s rtree_leaf_elm_t; diff --git a/contrib/jemalloc/include/jemalloc/internal/safety_check.h b/contrib/jemalloc/include/jemalloc/internal/safety_check.h --- a/contrib/jemalloc/include/jemalloc/internal/safety_check.h +++ b/contrib/jemalloc/include/jemalloc/internal/safety_check.h @@ -1,9 +1,14 @@ #ifndef JEMALLOC_INTERNAL_SAFETY_CHECK_H #define JEMALLOC_INTERNAL_SAFETY_CHECK_H +void safety_check_fail_sized_dealloc(bool current_dealloc, const void *ptr, + size_t true_size, size_t input_size); void safety_check_fail(const char *format, ...); + +typedef void (*safety_check_abort_hook_t)(const char *message); + /* Can set to NULL for a default. */ -void safety_check_set_abort(void (*abort_fn)()); +void safety_check_set_abort(safety_check_abort_hook_t abort_fn); JEMALLOC_ALWAYS_INLINE void safety_check_set_redzone(void *ptr, size_t usize, size_t bumped_usize) { diff --git a/contrib/jemalloc/include/jemalloc/internal/san.h b/contrib/jemalloc/include/jemalloc/internal/san.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/san.h @@ -0,0 +1,191 @@ +#ifndef JEMALLOC_INTERNAL_GUARD_H +#define JEMALLOC_INTERNAL_GUARD_H + +#include "jemalloc/internal/ehooks.h" +#include "jemalloc/internal/emap.h" + +#define SAN_PAGE_GUARD PAGE +#define SAN_PAGE_GUARDS_SIZE (SAN_PAGE_GUARD * 2) + +#define SAN_GUARD_LARGE_EVERY_N_EXTENTS_DEFAULT 0 +#define SAN_GUARD_SMALL_EVERY_N_EXTENTS_DEFAULT 0 + +#define SAN_LG_UAF_ALIGN_DEFAULT (-1) +#define SAN_CACHE_BIN_NONFAST_MASK_DEFAULT (uintptr_t)(-1) + +static const uintptr_t uaf_detect_junk = (uintptr_t)0x5b5b5b5b5b5b5b5bULL; + +/* 0 means disabled, i.e. never guarded. */ +extern size_t opt_san_guard_large; +extern size_t opt_san_guard_small; +/* -1 means disabled, i.e. never check for use-after-free. */ +extern ssize_t opt_lg_san_uaf_align; + +void san_guard_pages(tsdn_t *tsdn, ehooks_t *ehooks, edata_t *edata, + emap_t *emap, bool left, bool right, bool remap); +void san_unguard_pages(tsdn_t *tsdn, ehooks_t *ehooks, edata_t *edata, + emap_t *emap, bool left, bool right); +/* + * Unguard the extent, but don't modify emap boundaries. Must be called on an + * extent that has been erased from emap and shouldn't be placed back. + */ +void san_unguard_pages_pre_destroy(tsdn_t *tsdn, ehooks_t *ehooks, + edata_t *edata, emap_t *emap); +void san_check_stashed_ptrs(void **ptrs, size_t nstashed, size_t usize); + +void tsd_san_init(tsd_t *tsd); +void san_init(ssize_t lg_san_uaf_align); + +static inline void +san_guard_pages_two_sided(tsdn_t *tsdn, ehooks_t *ehooks, edata_t *edata, + emap_t *emap, bool remap) { + san_guard_pages(tsdn, ehooks, edata, emap, true, true, remap); +} + +static inline void +san_unguard_pages_two_sided(tsdn_t *tsdn, ehooks_t *ehooks, edata_t *edata, + emap_t *emap) { + san_unguard_pages(tsdn, ehooks, edata, emap, true, true); +} + +static inline size_t +san_two_side_unguarded_sz(size_t size) { + assert(size % PAGE == 0); + assert(size >= SAN_PAGE_GUARDS_SIZE); + return size - SAN_PAGE_GUARDS_SIZE; +} + +static inline size_t +san_two_side_guarded_sz(size_t size) { + assert(size % PAGE == 0); + return size + SAN_PAGE_GUARDS_SIZE; +} + +static inline size_t +san_one_side_unguarded_sz(size_t size) { + assert(size % PAGE == 0); + assert(size >= SAN_PAGE_GUARD); + return size - SAN_PAGE_GUARD; +} + +static inline size_t +san_one_side_guarded_sz(size_t size) { + assert(size % PAGE == 0); + return size + SAN_PAGE_GUARD; +} + +static inline bool +san_guard_enabled(void) { + return (opt_san_guard_large != 0 || opt_san_guard_small != 0); +} + +static inline bool +san_large_extent_decide_guard(tsdn_t *tsdn, ehooks_t *ehooks, size_t size, + size_t alignment) { + if (opt_san_guard_large == 0 || ehooks_guard_will_fail(ehooks) || + tsdn_null(tsdn)) { + return false; + } + + tsd_t *tsd = tsdn_tsd(tsdn); + uint64_t n = tsd_san_extents_until_guard_large_get(tsd); + assert(n >= 1); + if (n > 1) { + /* + * Subtract conditionally because the guard may not happen due + * to alignment or size restriction below. + */ + *tsd_san_extents_until_guard_largep_get(tsd) = n - 1; + } + + if (n == 1 && (alignment <= PAGE) && + (san_two_side_guarded_sz(size) <= SC_LARGE_MAXCLASS)) { + *tsd_san_extents_until_guard_largep_get(tsd) = + opt_san_guard_large; + return true; + } else { + assert(tsd_san_extents_until_guard_large_get(tsd) >= 1); + return false; + } +} + +static inline bool +san_slab_extent_decide_guard(tsdn_t *tsdn, ehooks_t *ehooks) { + if (opt_san_guard_small == 0 || ehooks_guard_will_fail(ehooks) || + tsdn_null(tsdn)) { + return false; + } + + tsd_t *tsd = tsdn_tsd(tsdn); + uint64_t n = tsd_san_extents_until_guard_small_get(tsd); + assert(n >= 1); + if (n == 1) { + *tsd_san_extents_until_guard_smallp_get(tsd) = + opt_san_guard_small; + return true; + } else { + *tsd_san_extents_until_guard_smallp_get(tsd) = n - 1; + assert(tsd_san_extents_until_guard_small_get(tsd) >= 1); + return false; + } +} + +static inline void +san_junk_ptr_locations(void *ptr, size_t usize, void **first, void **mid, + void **last) { + size_t ptr_sz = sizeof(void *); + + *first = ptr; + + *mid = (void *)((uintptr_t)ptr + ((usize >> 1) & ~(ptr_sz - 1))); + assert(*first != *mid || usize == ptr_sz); + assert((uintptr_t)*first <= (uintptr_t)*mid); + + /* + * When usize > 32K, the gap between requested_size and usize might be + * greater than 4K -- this means the last write may access an + * likely-untouched page (default settings w/ 4K pages). However by + * default the tcache only goes up to the 32K size class, and is usually + * tuned lower instead of higher, which makes it less of a concern. + */ + *last = (void *)((uintptr_t)ptr + usize - sizeof(uaf_detect_junk)); + assert(*first != *last || usize == ptr_sz); + assert(*mid != *last || usize <= ptr_sz * 2); + assert((uintptr_t)*mid <= (uintptr_t)*last); +} + +static inline bool +san_junk_ptr_should_slow(void) { + /* + * The latter condition (pointer size greater than the min size class) + * is not expected -- fall back to the slow path for simplicity. + */ + return config_debug || (LG_SIZEOF_PTR > SC_LG_TINY_MIN); +} + +static inline void +san_junk_ptr(void *ptr, size_t usize) { + if (san_junk_ptr_should_slow()) { + memset(ptr, (char)uaf_detect_junk, usize); + return; + } + + void *first, *mid, *last; + san_junk_ptr_locations(ptr, usize, &first, &mid, &last); + *(uintptr_t *)first = uaf_detect_junk; + *(uintptr_t *)mid = uaf_detect_junk; + *(uintptr_t *)last = uaf_detect_junk; +} + +static inline bool +san_uaf_detection_enabled(void) { + bool ret = config_uaf_detection && (opt_lg_san_uaf_align != -1); + if (config_uaf_detection && ret) { + assert(san_cache_bin_nonfast_mask == ((uintptr_t)1 << + opt_lg_san_uaf_align) - 1); + } + + return ret; +} + +#endif /* JEMALLOC_INTERNAL_GUARD_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/san_bump.h b/contrib/jemalloc/include/jemalloc/internal/san_bump.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/san_bump.h @@ -0,0 +1,51 @@ +#ifndef JEMALLOC_INTERNAL_SAN_BUMP_H +#define JEMALLOC_INTERNAL_SAN_BUMP_H + +#include "jemalloc/internal/edata.h" +#include "jemalloc/internal/exp_grow.h" +#include "jemalloc/internal/mutex.h" + +#define SBA_RETAINED_ALLOC_SIZE ((size_t)4 << 20) + +extern bool opt_retain; + +typedef struct pac_s pac_t; + +typedef struct san_bump_alloc_s san_bump_alloc_t; +struct san_bump_alloc_s { + malloc_mutex_t mtx; + + edata_t *curr_reg; +}; + +static inline bool +san_bump_enabled() { + /* + * We enable san_bump allocator only when it's possible to break up a + * mapping and unmap a part of it (maps_coalesce). This is needed to + * ensure the arena destruction process can destroy all retained guarded + * extents one by one and to unmap a trailing part of a retained guarded + * region when it's too small to fit a pending allocation. + * opt_retain is required, because this allocator retains a large + * virtual memory mapping and returns smaller parts of it. + */ + return maps_coalesce && opt_retain; +} + +static inline bool +san_bump_alloc_init(san_bump_alloc_t* sba) { + bool err = malloc_mutex_init(&sba->mtx, "sanitizer_bump_allocator", + WITNESS_RANK_SAN_BUMP_ALLOC, malloc_mutex_rank_exclusive); + if (err) { + return true; + } + sba->curr_reg = NULL; + + return false; +} + +edata_t * +san_bump_alloc(tsdn_t *tsdn, san_bump_alloc_t* sba, pac_t *pac, ehooks_t *ehooks, + size_t size, bool zero); + +#endif /* JEMALLOC_INTERNAL_SAN_BUMP_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/sc.h b/contrib/jemalloc/include/jemalloc/internal/sc.h --- a/contrib/jemalloc/include/jemalloc/internal/sc.h +++ b/contrib/jemalloc/include/jemalloc/internal/sc.h @@ -197,30 +197,34 @@ (SC_LG_BASE_MAX - SC_LG_FIRST_REGULAR_BASE + 1) - 1) #define SC_NSIZES (SC_NTINY + SC_NPSEUDO + SC_NREGULAR) -/* The number of size classes that are a multiple of the page size. */ -#define SC_NPSIZES ( \ - /* Start with all the size classes. */ \ - SC_NSIZES \ - /* Subtract out those groups with too small a base. */ \ - - (LG_PAGE - 1 - SC_LG_FIRST_REGULAR_BASE) * SC_NGROUP \ - /* And the pseudo-group. */ \ - - SC_NPSEUDO \ - /* And the tiny group. */ \ - - SC_NTINY \ - /* Sizes where ndelta*delta is not a multiple of the page size. */ \ - - (SC_LG_NGROUP * SC_NGROUP)) /* - * Note that the last line is computed as the sum of the second column in the - * following table: - * lg(base) | count of sizes to exclude - * ------------------------------|----------------------------- - * LG_PAGE - 1 | SC_NGROUP - 1 - * LG_PAGE | SC_NGROUP - 1 - * LG_PAGE + 1 | SC_NGROUP - 2 - * LG_PAGE + 2 | SC_NGROUP - 4 - * ... | ... - * LG_PAGE + (SC_LG_NGROUP - 1) | SC_NGROUP - (SC_NGROUP / 2) + * The number of size classes that are a multiple of the page size. + * + * Here are the first few bases that have a page-sized SC. + * + * lg(base) | base | highest SC | page-multiple SCs + * --------------|------------------------------------------ + * LG_PAGE - 1 | PAGE / 2 | PAGE | 1 + * LG_PAGE | PAGE | 2 * PAGE | 1 + * LG_PAGE + 1 | 2 * PAGE | 4 * PAGE | 2 + * LG_PAGE + 2 | 4 * PAGE | 8 * PAGE | 4 + * + * The number of page-multiple SCs continues to grow in powers of two, up until + * lg_delta == lg_page, which corresponds to setting lg_base to lg_page + + * SC_LG_NGROUP. So, then, the number of size classes that are multiples of the + * page size whose lg_delta is less than the page size are + * is 1 + (2**0 + 2**1 + ... + 2**(lg_ngroup - 1) == 2**lg_ngroup. + * + * For each base with lg_base in [lg_page + lg_ngroup, lg_base_max), there are + * NGROUP page-sized size classes, and when lg_base == lg_base_max, there are + * NGROUP - 1. + * + * This gives us the quantity we seek. */ +#define SC_NPSIZES ( \ + SC_NGROUP \ + + (SC_LG_BASE_MAX - (LG_PAGE + SC_LG_NGROUP)) * SC_NGROUP \ + + SC_NGROUP - 1) /* * We declare a size class is binnable if size < page size * group. Or, in other @@ -242,17 +246,23 @@ # error "Too many small size classes" #endif -/* The largest size class in the lookup table. */ -#define SC_LOOKUP_MAXCLASS ((size_t)1 << 12) +/* The largest size class in the lookup table, and its binary log. */ +#define SC_LG_MAX_LOOKUP 12 +#define SC_LOOKUP_MAXCLASS (1 << SC_LG_MAX_LOOKUP) /* Internal, only used for the definition of SC_SMALL_MAXCLASS. */ -#define SC_SMALL_MAX_BASE ((size_t)1 << (LG_PAGE + SC_LG_NGROUP - 1)) -#define SC_SMALL_MAX_DELTA ((size_t)1 << (LG_PAGE - 1)) +#define SC_SMALL_MAX_BASE (1 << (LG_PAGE + SC_LG_NGROUP - 1)) +#define SC_SMALL_MAX_DELTA (1 << (LG_PAGE - 1)) /* The largest size class allocated out of a slab. */ #define SC_SMALL_MAXCLASS (SC_SMALL_MAX_BASE \ + (SC_NGROUP - 1) * SC_SMALL_MAX_DELTA) +/* The fastpath assumes all lookup-able sizes are small. */ +#if (SC_SMALL_MAXCLASS < SC_LOOKUP_MAXCLASS) +# error "Lookup table sizes must be small" +#endif + /* The smallest size class not allocated out of a slab. */ #define SC_LARGE_MINCLASS ((size_t)1ULL << (LG_PAGE + SC_LG_NGROUP)) #define SC_LG_LARGE_MINCLASS (LG_PAGE + SC_LG_NGROUP) @@ -264,6 +274,19 @@ /* The largest size class supported. */ #define SC_LARGE_MAXCLASS (SC_MAX_BASE + (SC_NGROUP - 1) * SC_MAX_DELTA) +/* Maximum number of regions in one slab. */ +#ifndef CONFIG_LG_SLAB_MAXREGS +# define SC_LG_SLAB_MAXREGS (LG_PAGE - SC_LG_TINY_MIN) +#else +# if CONFIG_LG_SLAB_MAXREGS < (LG_PAGE - SC_LG_TINY_MIN) +# error "Unsupported SC_LG_SLAB_MAXREGS" +# else +# define SC_LG_SLAB_MAXREGS CONFIG_LG_SLAB_MAXREGS +# endif +#endif + +#define SC_SLAB_MAXREGS (1U << SC_LG_SLAB_MAXREGS) + typedef struct sc_s sc_t; struct sc_s { /* Size class index, or -1 if not a valid size class. */ @@ -321,10 +344,11 @@ sc_t sc[SC_NSIZES]; }; +size_t reg_size_compute(int lg_base, int lg_delta, int ndelta); void sc_data_init(sc_data_t *data); /* * Updates slab sizes in [begin, end] to be pgs pages in length, if possible. - * Otherwise, does its best to accomodate the request. + * Otherwise, does its best to accommodate the request. */ void sc_data_update_slab_size(sc_data_t *data, size_t begin, size_t end, int pgs); diff --git a/contrib/jemalloc/include/jemalloc/internal/sec.h b/contrib/jemalloc/include/jemalloc/internal/sec.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/sec.h @@ -0,0 +1,120 @@ +#ifndef JEMALLOC_INTERNAL_SEC_H +#define JEMALLOC_INTERNAL_SEC_H + +#include "jemalloc/internal/atomic.h" +#include "jemalloc/internal/pai.h" + +/* + * Small extent cache. + * + * This includes some utilities to cache small extents. We have a per-pszind + * bin with its own list of extents of that size. We don't try to do any + * coalescing of extents (since it would in general require cross-shard locks or + * knowledge of the underlying PAI implementation). + */ + +/* + * For now, this is just one field; eventually, we'll probably want to get more + * fine-grained data out (like per-size class statistics). + */ +typedef struct sec_stats_s sec_stats_t; +struct sec_stats_s { + /* Sum of bytes_cur across all shards. */ + size_t bytes; +}; + +static inline void +sec_stats_accum(sec_stats_t *dst, sec_stats_t *src) { + dst->bytes += src->bytes; +} + +/* A collections of free extents, all of the same size. */ +typedef struct sec_bin_s sec_bin_t; +struct sec_bin_s { + /* + * When we fail to fulfill an allocation, we do a batch-alloc on the + * underlying allocator to fill extra items, as well. We drop the SEC + * lock while doing so, to allow operations on other bins to succeed. + * That introduces the possibility of other threads also trying to + * allocate out of this bin, failing, and also going to the backing + * allocator. To avoid a thundering herd problem in which lots of + * threads do batch allocs and overfill this bin as a result, we only + * allow one batch allocation at a time for a bin. This bool tracks + * whether or not some thread is already batch allocating. + * + * Eventually, the right answer may be a smarter sharding policy for the + * bins (e.g. a mutex per bin, which would also be more scalable + * generally; the batch-allocating thread could hold it while + * batch-allocating). + */ + bool being_batch_filled; + + /* + * Number of bytes in this particular bin (as opposed to the + * sec_shard_t's bytes_cur. This isn't user visible or reported in + * stats; rather, it allows us to quickly determine the change in the + * centralized counter when flushing. + */ + size_t bytes_cur; + edata_list_active_t freelist; +}; + +typedef struct sec_shard_s sec_shard_t; +struct sec_shard_s { + /* + * We don't keep per-bin mutexes, even though that would allow more + * sharding; this allows global cache-eviction, which in turn allows for + * better balancing across free lists. + */ + malloc_mutex_t mtx; + /* + * A SEC may need to be shut down (i.e. flushed of its contents and + * prevented from further caching). To avoid tricky synchronization + * issues, we just track enabled-status in each shard, guarded by a + * mutex. In practice, this is only ever checked during brief races, + * since the arena-level atomic boolean tracking HPA enabled-ness means + * that we won't go down these pathways very often after custom extent + * hooks are installed. + */ + bool enabled; + sec_bin_t *bins; + /* Number of bytes in all bins in the shard. */ + size_t bytes_cur; + /* The next pszind to flush in the flush-some pathways. */ + pszind_t to_flush_next; +}; + +typedef struct sec_s sec_t; +struct sec_s { + pai_t pai; + pai_t *fallback; + + sec_opts_t opts; + sec_shard_t *shards; + pszind_t npsizes; +}; + +bool sec_init(tsdn_t *tsdn, sec_t *sec, base_t *base, pai_t *fallback, + const sec_opts_t *opts); +void sec_flush(tsdn_t *tsdn, sec_t *sec); +void sec_disable(tsdn_t *tsdn, sec_t *sec); + +/* + * Morally, these two stats methods probably ought to be a single one (and the + * mutex_prof_data ought to live in the sec_stats_t. But splitting them apart + * lets them fit easily into the pa_shard stats framework (which also has this + * split), which simplifies the stats management. + */ +void sec_stats_merge(tsdn_t *tsdn, sec_t *sec, sec_stats_t *stats); +void sec_mutex_stats_read(tsdn_t *tsdn, sec_t *sec, + mutex_prof_data_t *mutex_prof_data); + +/* + * We use the arena lock ordering; these are acquired in phase 2 of forking, but + * should be acquired before the underlying allocator mutexes. + */ +void sec_prefork2(tsdn_t *tsdn, sec_t *sec); +void sec_postfork_parent(tsdn_t *tsdn, sec_t *sec); +void sec_postfork_child(tsdn_t *tsdn, sec_t *sec); + +#endif /* JEMALLOC_INTERNAL_SEC_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/sec_opts.h b/contrib/jemalloc/include/jemalloc/internal/sec_opts.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/sec_opts.h @@ -0,0 +1,59 @@ +#ifndef JEMALLOC_INTERNAL_SEC_OPTS_H +#define JEMALLOC_INTERNAL_SEC_OPTS_H + +/* + * The configuration settings used by an sec_t. Morally, this is part of the + * SEC interface, but we put it here for header-ordering reasons. + */ + +typedef struct sec_opts_s sec_opts_t; +struct sec_opts_s { + /* + * We don't necessarily always use all the shards; requests are + * distributed across shards [0, nshards - 1). + */ + size_t nshards; + /* + * We'll automatically refuse to cache any objects in this sec if + * they're larger than max_alloc bytes, instead forwarding such objects + * directly to the fallback. + */ + size_t max_alloc; + /* + * Exceeding this amount of cached extents in a shard causes us to start + * flushing bins in that shard until we fall below bytes_after_flush. + */ + size_t max_bytes; + /* + * The number of bytes (in all bins) we flush down to when we exceed + * bytes_cur. We want this to be less than bytes_cur, because + * otherwise we could get into situations where a shard undergoing + * net-deallocation keeps bytes_cur very near to max_bytes, so that + * most deallocations get immediately forwarded to the underlying PAI + * implementation, defeating the point of the SEC. + */ + size_t bytes_after_flush; + /* + * When we can't satisfy an allocation out of the SEC because there are + * no available ones cached, we allocate multiple of that size out of + * the fallback allocator. Eventually we might want to do something + * cleverer, but for now we just grab a fixed number. + */ + size_t batch_fill_extra; +}; + +#define SEC_OPTS_DEFAULT { \ + /* nshards */ \ + 4, \ + /* max_alloc */ \ + (32 * 1024) < PAGE ? PAGE : (32 * 1024), \ + /* max_bytes */ \ + 256 * 1024, \ + /* bytes_after_flush */ \ + 128 * 1024, \ + /* batch_fill_extra */ \ + 0 \ +} + + +#endif /* JEMALLOC_INTERNAL_SEC_OPTS_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/slab_data.h b/contrib/jemalloc/include/jemalloc/internal/slab_data.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/slab_data.h @@ -0,0 +1,12 @@ +#ifndef JEMALLOC_INTERNAL_SLAB_DATA_H +#define JEMALLOC_INTERNAL_SLAB_DATA_H + +#include "jemalloc/internal/bitmap.h" + +typedef struct slab_data_s slab_data_t; +struct slab_data_s { + /* Per region allocated/deallocated bitmap. */ + bitmap_t bitmap[BITMAP_GROUPS_MAX]; +}; + +#endif /* JEMALLOC_INTERNAL_SLAB_DATA_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/stats.h b/contrib/jemalloc/include/jemalloc/internal/stats.h --- a/contrib/jemalloc/include/jemalloc/internal/stats.h +++ b/contrib/jemalloc/include/jemalloc/internal/stats.h @@ -11,7 +11,8 @@ OPTION('b', bins, true, false) \ OPTION('l', large, true, false) \ OPTION('x', mutex, true, false) \ - OPTION('e', extents, true, false) + OPTION('e', extents, true, false) \ + OPTION('h', hpa, config_stats, false) enum { #define OPTION(o, v, d, s) stats_print_option_num_##v, @@ -24,8 +25,30 @@ extern bool opt_stats_print; extern char opt_stats_print_opts[stats_print_tot_num_options+1]; +/* Utilities for stats_interval. */ +extern int64_t opt_stats_interval; +extern char opt_stats_interval_opts[stats_print_tot_num_options+1]; + +#define STATS_INTERVAL_DEFAULT -1 +/* + * Batch-increment the counter to reduce synchronization overhead. Each thread + * merges after (interval >> LG_BATCH_SIZE) bytes of allocations; also limit the + * BATCH_MAX for accuracy when the interval is huge (which is expected). + */ +#define STATS_INTERVAL_ACCUM_LG_BATCH_SIZE 6 +#define STATS_INTERVAL_ACCUM_BATCH_MAX (4 << 20) + +/* Only accessed by thread event. */ +uint64_t stats_interval_new_event_wait(tsd_t *tsd); +uint64_t stats_interval_postponed_event_wait(tsd_t *tsd); +void stats_interval_event_handler(tsd_t *tsd, uint64_t elapsed); + /* Implements je_malloc_stats_print. */ -void stats_print(void (*write_cb)(void *, const char *), void *cbopaque, - const char *opts); +void stats_print(write_cb_t *write_cb, void *cbopaque, const char *opts); + +bool stats_boot(void); +void stats_prefork(tsdn_t *tsdn); +void stats_postfork_parent(tsdn_t *tsdn); +void stats_postfork_child(tsdn_t *tsdn); #endif /* JEMALLOC_INTERNAL_STATS_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/sz.h b/contrib/jemalloc/include/jemalloc/internal/sz.h --- a/contrib/jemalloc/include/jemalloc/internal/sz.h +++ b/contrib/jemalloc/include/jemalloc/internal/sz.h @@ -22,6 +22,12 @@ * size that would result from such an allocation. */ +/* Page size index type. */ +typedef unsigned pszind_t; + +/* Size class index type. */ +typedef unsigned szind_t; + /* * sz_pind2sz_tab encodes the same information as could be computed by * sz_pind2sz_compute(). @@ -39,34 +45,62 @@ */ extern uint8_t sz_size2index_tab[]; -static const size_t sz_large_pad = -#ifdef JEMALLOC_CACHE_OBLIVIOUS - PAGE -#else - 0 -#endif - ; +/* + * Padding for large allocations: PAGE when opt_cache_oblivious == true (to + * enable cache index randomization); 0 otherwise. + */ +extern size_t sz_large_pad; -extern void sz_boot(const sc_data_t *sc_data); +extern void sz_boot(const sc_data_t *sc_data, bool cache_oblivious); JEMALLOC_ALWAYS_INLINE pszind_t sz_psz2ind(size_t psz) { + assert(psz > 0); if (unlikely(psz > SC_LARGE_MAXCLASS)) { return SC_NPSIZES; } - pszind_t x = lg_floor((psz<<1)-1); - pszind_t shift = (x < SC_LG_NGROUP + LG_PAGE) ? + /* x is the lg of the first base >= psz. */ + pszind_t x = lg_ceil(psz); + /* + * sc.h introduces a lot of size classes. These size classes are divided + * into different size class groups. There is a very special size class + * group, each size class in or after it is an integer multiple of PAGE. + * We call it first_ps_rg. It means first page size regular group. The + * range of first_ps_rg is (base, base * 2], and base == PAGE * + * SC_NGROUP. off_to_first_ps_rg begins from 1, instead of 0. e.g. + * off_to_first_ps_rg is 1 when psz is (PAGE * SC_NGROUP + 1). + */ + pszind_t off_to_first_ps_rg = (x < SC_LG_NGROUP + LG_PAGE) ? 0 : x - (SC_LG_NGROUP + LG_PAGE); - pszind_t grp = shift << SC_LG_NGROUP; - pszind_t lg_delta = (x < SC_LG_NGROUP + LG_PAGE + 1) ? - LG_PAGE : x - SC_LG_NGROUP - 1; + /* + * Same as sc_s::lg_delta. + * Delta for off_to_first_ps_rg == 1 is PAGE, + * for each increase in offset, it's multiplied by two. + * Therefore, lg_delta = LG_PAGE + (off_to_first_ps_rg - 1). + */ + pszind_t lg_delta = (off_to_first_ps_rg == 0) ? + LG_PAGE : LG_PAGE + (off_to_first_ps_rg - 1); - size_t delta_inverse_mask = ZU(-1) << lg_delta; - pszind_t mod = ((((psz-1) & delta_inverse_mask) >> lg_delta)) & - ((ZU(1) << SC_LG_NGROUP) - 1); + /* + * Let's write psz in binary, e.g. 0011 for 0x3, 0111 for 0x7. + * The leftmost bits whose len is lg_base decide the base of psz. + * The rightmost bits whose len is lg_delta decide (pgz % PAGE). + * The middle bits whose len is SC_LG_NGROUP decide ndelta. + * ndelta is offset to the first size class in the size class group, + * starts from 1. + * If you don't know lg_base, ndelta or lg_delta, see sc.h. + * |xxxxxxxxxxxxxxxxxxxx|------------------------|yyyyyyyyyyyyyyyyyyyyy| + * |<-- len: lg_base -->|<-- len: SC_LG_NGROUP-->|<-- len: lg_delta -->| + * |<-- ndelta -->| + * rg_inner_off = ndelta - 1 + * Why use (psz - 1)? + * To handle case: psz % (1 << lg_delta) == 0. + */ + pszind_t rg_inner_off = (((psz - 1)) >> lg_delta) & (SC_NGROUP - 1); - pszind_t ind = grp + mod; + pszind_t base_ind = off_to_first_ps_rg << SC_LG_NGROUP; + pszind_t ind = base_ind + rg_inner_off; return ind; } @@ -152,10 +186,15 @@ } JEMALLOC_ALWAYS_INLINE szind_t -sz_size2index_lookup(size_t size) { +sz_size2index_lookup_impl(size_t size) { assert(size <= SC_LOOKUP_MAXCLASS); - szind_t ret = (sz_size2index_tab[(size + (ZU(1) << SC_LG_TINY_MIN) - 1) - >> SC_LG_TINY_MIN]); + return sz_size2index_tab[(size + (ZU(1) << SC_LG_TINY_MIN) - 1) + >> SC_LG_TINY_MIN]; +} + +JEMALLOC_ALWAYS_INLINE szind_t +sz_size2index_lookup(size_t size) { + szind_t ret = sz_size2index_lookup_impl(size); assert(ret == sz_size2index_compute(size)); return ret; } @@ -194,9 +233,14 @@ } } +JEMALLOC_ALWAYS_INLINE size_t +sz_index2size_lookup_impl(szind_t index) { + return sz_index2size_tab[index]; +} + JEMALLOC_ALWAYS_INLINE size_t sz_index2size_lookup(szind_t index) { - size_t ret = (size_t)sz_index2size_tab[index]; + size_t ret = sz_index2size_lookup_impl(index); assert(ret == sz_index2size_compute(index)); return ret; } @@ -207,6 +251,12 @@ return sz_index2size_lookup(index); } +JEMALLOC_ALWAYS_INLINE void +sz_size2index_usize_fastpath(size_t size, szind_t *ind, size_t *usize) { + *ind = sz_size2index_lookup_impl(size); + *usize = sz_index2size_lookup_impl(*ind); +} + JEMALLOC_ALWAYS_INLINE size_t sz_s2u_compute(size_t size) { if (unlikely(size > SC_LARGE_MAXCLASS)) { @@ -266,7 +316,7 @@ assert(alignment != 0 && ((alignment - 1) & alignment) == 0); /* Try for a small size class. */ - if (size <= SC_SMALL_MAXCLASS && alignment < PAGE) { + if (size <= SC_SMALL_MAXCLASS && alignment <= PAGE) { /* * Round size up to the nearest multiple of alignment. * @@ -315,4 +365,7 @@ return usize; } +size_t sz_psz_quantize_floor(size_t size); +size_t sz_psz_quantize_ceil(size_t size); + #endif /* JEMALLOC_INTERNAL_SIZE_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/tcache_externs.h b/contrib/jemalloc/include/jemalloc/internal/tcache_externs.h --- a/contrib/jemalloc/include/jemalloc/internal/tcache_externs.h +++ b/contrib/jemalloc/include/jemalloc/internal/tcache_externs.h @@ -1,10 +1,17 @@ #ifndef JEMALLOC_INTERNAL_TCACHE_EXTERNS_H #define JEMALLOC_INTERNAL_TCACHE_EXTERNS_H -extern bool opt_tcache; -extern ssize_t opt_lg_tcache_max; - -extern cache_bin_info_t *tcache_bin_info; +extern bool opt_tcache; +extern size_t opt_tcache_max; +extern ssize_t opt_lg_tcache_nslots_mul; +extern unsigned opt_tcache_nslots_small_min; +extern unsigned opt_tcache_nslots_small_max; +extern unsigned opt_tcache_nslots_large; +extern ssize_t opt_lg_tcache_shift; +extern size_t opt_tcache_gc_incr_bytes; +extern size_t opt_tcache_gc_delay_bytes; +extern unsigned opt_lg_tcache_flush_small_div; +extern unsigned opt_lg_tcache_flush_large_div; /* * Number of tcache bins. There are SC_NBINS small-object bins, plus 0 or more @@ -15,6 +22,8 @@ /* Maximum cached size class. */ extern size_t tcache_maxclass; +extern cache_bin_info_t *tcache_bin_info; + /* * Explicit tcaches, managed via the tcache.{create,flush,destroy} mallctls and * usable via the MALLOCX_TCACHE() flag. The automatic per thread tcaches are @@ -25,24 +34,27 @@ */ extern tcaches_t *tcaches; -size_t tcache_salloc(tsdn_t *tsdn, const void *ptr); -void tcache_event_hard(tsd_t *tsd, tcache_t *tcache); -void *tcache_alloc_small_hard(tsdn_t *tsdn, arena_t *arena, tcache_t *tcache, +size_t tcache_salloc(tsdn_t *tsdn, const void *ptr); +void *tcache_alloc_small_hard(tsdn_t *tsdn, arena_t *arena, tcache_t *tcache, cache_bin_t *tbin, szind_t binind, bool *tcache_success); -void tcache_bin_flush_small(tsd_t *tsd, tcache_t *tcache, cache_bin_t *tbin, + +void tcache_bin_flush_small(tsd_t *tsd, tcache_t *tcache, cache_bin_t *tbin, szind_t binind, unsigned rem); -void tcache_bin_flush_large(tsd_t *tsd, cache_bin_t *tbin, szind_t binind, - unsigned rem, tcache_t *tcache); -void tcache_arena_reassociate(tsdn_t *tsdn, tcache_t *tcache, - arena_t *arena); +void tcache_bin_flush_large(tsd_t *tsd, tcache_t *tcache, cache_bin_t *tbin, + szind_t binind, unsigned rem); +void tcache_bin_flush_stashed(tsd_t *tsd, tcache_t *tcache, cache_bin_t *bin, + szind_t binind, bool is_small); +void tcache_arena_reassociate(tsdn_t *tsdn, tcache_slow_t *tcache_slow, + tcache_t *tcache, arena_t *arena); tcache_t *tcache_create_explicit(tsd_t *tsd); -void tcache_cleanup(tsd_t *tsd); -void tcache_stats_merge(tsdn_t *tsdn, tcache_t *tcache, arena_t *arena); -bool tcaches_create(tsd_t *tsd, unsigned *r_ind); -void tcaches_flush(tsd_t *tsd, unsigned ind); -void tcaches_destroy(tsd_t *tsd, unsigned ind); -bool tcache_boot(tsdn_t *tsdn); -void tcache_arena_associate(tsdn_t *tsdn, tcache_t *tcache, arena_t *arena); +void tcache_cleanup(tsd_t *tsd); +void tcache_stats_merge(tsdn_t *tsdn, tcache_t *tcache, arena_t *arena); +bool tcaches_create(tsd_t *tsd, base_t *base, unsigned *r_ind); +void tcaches_flush(tsd_t *tsd, unsigned ind); +void tcaches_destroy(tsd_t *tsd, unsigned ind); +bool tcache_boot(tsdn_t *tsdn, base_t *base); +void tcache_arena_associate(tsdn_t *tsdn, tcache_slow_t *tcache_slow, + tcache_t *tcache, arena_t *arena); void tcache_prefork(tsdn_t *tsdn); void tcache_postfork_parent(tsdn_t *tsdn); void tcache_postfork_child(tsdn_t *tsdn); @@ -50,4 +62,14 @@ bool tsd_tcache_data_init(tsd_t *tsd); bool tsd_tcache_enabled_data_init(tsd_t *tsd); +void tcache_assert_initialized(tcache_t *tcache); + +/* Only accessed by thread event. */ +uint64_t tcache_gc_new_event_wait(tsd_t *tsd); +uint64_t tcache_gc_postponed_event_wait(tsd_t *tsd); +void tcache_gc_event_handler(tsd_t *tsd, uint64_t elapsed); +uint64_t tcache_gc_dalloc_new_event_wait(tsd_t *tsd); +uint64_t tcache_gc_dalloc_postponed_event_wait(tsd_t *tsd); +void tcache_gc_dalloc_event_handler(tsd_t *tsd, uint64_t elapsed); + #endif /* JEMALLOC_INTERNAL_TCACHE_EXTERNS_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/tcache_inlines.h b/contrib/jemalloc/include/jemalloc/internal/tcache_inlines.h --- a/contrib/jemalloc/include/jemalloc/internal/tcache_inlines.h +++ b/contrib/jemalloc/include/jemalloc/internal/tcache_inlines.h @@ -3,9 +3,9 @@ #include "jemalloc/internal/bin.h" #include "jemalloc/internal/jemalloc_internal_types.h" +#include "jemalloc/internal/san.h" #include "jemalloc/internal/sc.h" #include "jemalloc/internal/sz.h" -#include "jemalloc/internal/ticker.h" #include "jemalloc/internal/util.h" static inline bool @@ -27,28 +27,29 @@ tsd_slow_update(tsd); } -JEMALLOC_ALWAYS_INLINE void -tcache_event(tsd_t *tsd, tcache_t *tcache) { - if (TCACHE_GC_INCR == 0) { - return; +JEMALLOC_ALWAYS_INLINE bool +tcache_small_bin_disabled(szind_t ind, cache_bin_t *bin) { + assert(ind < SC_NBINS); + bool ret = (cache_bin_info_ncached_max(&tcache_bin_info[ind]) == 0); + if (ret && bin != NULL) { + /* small size class but cache bin disabled. */ + assert(ind >= nhbins); + assert((uintptr_t)(*bin->stack_head) == + cache_bin_preceding_junk); } - if (unlikely(ticker_tick(&tcache->gc_ticker))) { - tcache_event_hard(tsd, tcache); - } + return ret; } JEMALLOC_ALWAYS_INLINE void * tcache_alloc_small(tsd_t *tsd, arena_t *arena, tcache_t *tcache, size_t size, szind_t binind, bool zero, bool slow_path) { void *ret; - cache_bin_t *bin; bool tcache_success; - size_t usize JEMALLOC_CC_SILENCE_INIT(0); assert(binind < SC_NBINS); - bin = tcache_small_bin_get(tcache, binind); - ret = cache_bin_alloc_easy(bin, &tcache_success); + cache_bin_t *bin = &tcache->bins[binind]; + ret = cache_bin_alloc(bin, &tcache_success); assert(tcache_success == (ret != NULL)); if (unlikely(!tcache_success)) { bool tcache_hard_success; @@ -56,6 +57,13 @@ if (unlikely(arena == NULL)) { return NULL; } + if (unlikely(tcache_small_bin_disabled(binind, bin))) { + /* stats and zero are handled directly by the arena. */ + return arena_malloc_hard(tsd_tsdn(tsd), arena, size, + binind, zero); + } + tcache_bin_flush_stashed(tsd, tcache, bin, binind, + /* is_small */ true); ret = tcache_alloc_small_hard(tsd_tsdn(tsd), arena, tcache, bin, binind, &tcache_hard_success); @@ -65,38 +73,14 @@ } assert(ret); - /* - * Only compute usize if required. The checks in the following if - * statement are all static. - */ - if (config_prof || (slow_path && config_fill) || unlikely(zero)) { - usize = sz_index2size(binind); + if (unlikely(zero)) { + size_t usize = sz_index2size(binind); assert(tcache_salloc(tsd_tsdn(tsd), ret) == usize); - } - - if (likely(!zero)) { - if (slow_path && config_fill) { - if (unlikely(opt_junk_alloc)) { - arena_alloc_junk_small(ret, &bin_infos[binind], - false); - } else if (unlikely(opt_zero)) { - memset(ret, 0, usize); - } - } - } else { - if (slow_path && config_fill && unlikely(opt_junk_alloc)) { - arena_alloc_junk_small(ret, &bin_infos[binind], true); - } memset(ret, 0, usize); } - if (config_stats) { bin->tstats.nrequests++; } - if (config_prof) { - tcache->prof_accumbytes += usize; - } - tcache_event(tsd, tcache); return ret; } @@ -104,12 +88,11 @@ tcache_alloc_large(tsd_t *tsd, arena_t *arena, tcache_t *tcache, size_t size, szind_t binind, bool zero, bool slow_path) { void *ret; - cache_bin_t *bin; bool tcache_success; - assert(binind >= SC_NBINS &&binind < nhbins); - bin = tcache_large_bin_get(tcache, binind); - ret = cache_bin_alloc_easy(bin, &tcache_success); + assert(binind >= SC_NBINS && binind < nhbins); + cache_bin_t *bin = &tcache->bins[binind]; + ret = cache_bin_alloc(bin, &tcache_success); assert(tcache_success == (ret != NULL)); if (unlikely(!tcache_success)) { /* @@ -120,96 +103,79 @@ if (unlikely(arena == NULL)) { return NULL; } + tcache_bin_flush_stashed(tsd, tcache, bin, binind, + /* is_small */ false); ret = large_malloc(tsd_tsdn(tsd), arena, sz_s2u(size), zero); if (ret == NULL) { return NULL; } } else { - size_t usize JEMALLOC_CC_SILENCE_INIT(0); - - /* Only compute usize on demand */ - if (config_prof || (slow_path && config_fill) || - unlikely(zero)) { - usize = sz_index2size(binind); + if (unlikely(zero)) { + size_t usize = sz_index2size(binind); assert(usize <= tcache_maxclass); - } - - if (likely(!zero)) { - if (slow_path && config_fill) { - if (unlikely(opt_junk_alloc)) { - memset(ret, JEMALLOC_ALLOC_JUNK, - usize); - } else if (unlikely(opt_zero)) { - memset(ret, 0, usize); - } - } - } else { memset(ret, 0, usize); } if (config_stats) { bin->tstats.nrequests++; } - if (config_prof) { - tcache->prof_accumbytes += usize; - } } - tcache_event(tsd, tcache); return ret; } JEMALLOC_ALWAYS_INLINE void tcache_dalloc_small(tsd_t *tsd, tcache_t *tcache, void *ptr, szind_t binind, bool slow_path) { - cache_bin_t *bin; - cache_bin_info_t *bin_info; + assert(tcache_salloc(tsd_tsdn(tsd), ptr) <= SC_SMALL_MAXCLASS); - assert(tcache_salloc(tsd_tsdn(tsd), ptr) - <= SC_SMALL_MAXCLASS); - - if (slow_path && config_fill && unlikely(opt_junk_free)) { - arena_dalloc_junk_small(ptr, &bin_infos[binind]); + cache_bin_t *bin = &tcache->bins[binind]; + /* + * Not marking the branch unlikely because this is past free_fastpath() + * (which handles the most common cases), i.e. at this point it's often + * uncommon cases. + */ + if (cache_bin_nonfast_aligned(ptr)) { + /* Junk unconditionally, even if bin is full. */ + san_junk_ptr(ptr, sz_index2size(binind)); + if (cache_bin_stash(bin, ptr)) { + return; + } + assert(cache_bin_full(bin)); + /* Bin full; fall through into the flush branch. */ } - bin = tcache_small_bin_get(tcache, binind); - bin_info = &tcache_bin_info[binind]; - if (unlikely(!cache_bin_dalloc_easy(bin, bin_info, ptr))) { - tcache_bin_flush_small(tsd, tcache, bin, binind, - (bin_info->ncached_max >> 1)); - bool ret = cache_bin_dalloc_easy(bin, bin_info, ptr); + if (unlikely(!cache_bin_dalloc_easy(bin, ptr))) { + if (unlikely(tcache_small_bin_disabled(binind, bin))) { + arena_dalloc_small(tsd_tsdn(tsd), ptr); + return; + } + cache_bin_sz_t max = cache_bin_info_ncached_max( + &tcache_bin_info[binind]); + unsigned remain = max >> opt_lg_tcache_flush_small_div; + tcache_bin_flush_small(tsd, tcache, bin, binind, remain); + bool ret = cache_bin_dalloc_easy(bin, ptr); assert(ret); } - - tcache_event(tsd, tcache); } JEMALLOC_ALWAYS_INLINE void tcache_dalloc_large(tsd_t *tsd, tcache_t *tcache, void *ptr, szind_t binind, bool slow_path) { - cache_bin_t *bin; - cache_bin_info_t *bin_info; assert(tcache_salloc(tsd_tsdn(tsd), ptr) > SC_SMALL_MAXCLASS); assert(tcache_salloc(tsd_tsdn(tsd), ptr) <= tcache_maxclass); - if (slow_path && config_fill && unlikely(opt_junk_free)) { - large_dalloc_junk(ptr, sz_index2size(binind)); - } - - bin = tcache_large_bin_get(tcache, binind); - bin_info = &tcache_bin_info[binind]; - if (unlikely(bin->ncached == bin_info->ncached_max)) { - tcache_bin_flush_large(tsd, bin, binind, - (bin_info->ncached_max >> 1), tcache); + cache_bin_t *bin = &tcache->bins[binind]; + if (unlikely(!cache_bin_dalloc_easy(bin, ptr))) { + unsigned remain = cache_bin_info_ncached_max( + &tcache_bin_info[binind]) >> opt_lg_tcache_flush_large_div; + tcache_bin_flush_large(tsd, tcache, bin, binind, remain); + bool ret = cache_bin_dalloc_easy(bin, ptr); + assert(ret); } - assert(bin->ncached < bin_info->ncached_max); - bin->ncached++; - *(bin->avail - bin->ncached) = ptr; - - tcache_event(tsd, tcache); } JEMALLOC_ALWAYS_INLINE tcache_t * diff --git a/contrib/jemalloc/include/jemalloc/internal/tcache_structs.h b/contrib/jemalloc/include/jemalloc/internal/tcache_structs.h --- a/contrib/jemalloc/include/jemalloc/internal/tcache_structs.h +++ b/contrib/jemalloc/include/jemalloc/internal/tcache_structs.h @@ -7,36 +7,19 @@ #include "jemalloc/internal/ticker.h" #include "jemalloc/internal/tsd_types.h" -/* Various uses of this struct need it to be a named type. */ -typedef ql_elm(tsd_t) tsd_link_t; +/* + * The tcache state is split into the slow and hot path data. Each has a + * pointer to the other, and the data always comes in pairs. The layout of each + * of them varies in practice; tcache_slow lives in the TSD for the automatic + * tcache, and as part of a dynamic allocation for manual allocations. Keeping + * a pointer to tcache_slow lets us treat these cases uniformly, rather than + * splitting up the tcache [de]allocation code into those paths called with the + * TSD tcache and those called with a manual tcache. + */ -struct tcache_s { - /* - * To minimize our cache-footprint, we put the frequently accessed data - * together at the start of this struct. - */ - - /* Cleared after arena_prof_accum(). */ - uint64_t prof_accumbytes; - /* Drives incremental GC. */ - ticker_t gc_ticker; - /* - * The pointer stacks associated with bins follow as a contiguous array. - * During tcache initialization, the avail pointer in each element of - * tbins is initialized to point to the proper offset within this array. - */ - cache_bin_t bins_small[SC_NBINS]; - - /* - * This data is less hot; we can be a little less careful with our - * footprint here. - */ +struct tcache_slow_s { /* Lets us track all the tcaches in an arena. */ - ql_elm(tcache_t) link; - - /* Logically scoped to tsd, but put here for cache layout reasons. */ - ql_elm(tsd_t) tsd_link; - bool in_hook; + ql_elm(tcache_slow_t) link; /* * The descriptor lets the arena find our cache bins without seeing the @@ -51,12 +34,27 @@ szind_t next_gc_bin; /* For small bins, fill (ncached_max >> lg_fill_div). */ uint8_t lg_fill_div[SC_NBINS]; + /* For small bins, whether has been refilled since last GC. */ + bool bin_refilled[SC_NBINS]; + /* + * For small bins, the number of items we can pretend to flush before + * actually flushing. + */ + uint8_t bin_flush_delay_items[SC_NBINS]; /* - * We put the cache bins for large size classes at the end of the - * struct, since some of them might not get used. This might end up - * letting us avoid touching an extra page if we don't have to. + * The start of the allocation containing the dynamic allocation for + * either the cache bins alone, or the cache bin memory as well as this + * tcache_slow_t and its associated tcache_t. */ - cache_bin_t bins_large[SC_NSIZES-SC_NBINS]; + void *dyn_alloc; + + /* The associated bins. */ + tcache_t *tcache; +}; + +struct tcache_s { + tcache_slow_t *tcache_slow; + cache_bin_t bins[TCACHE_NBINS_MAX]; }; /* Linkage for list of available (previously used) explicit tcache IDs. */ diff --git a/contrib/jemalloc/include/jemalloc/internal/tcache_types.h b/contrib/jemalloc/include/jemalloc/internal/tcache_types.h --- a/contrib/jemalloc/include/jemalloc/internal/tcache_types.h +++ b/contrib/jemalloc/include/jemalloc/internal/tcache_types.h @@ -3,6 +3,7 @@ #include "jemalloc/internal/sc.h" +typedef struct tcache_slow_s tcache_slow_t; typedef struct tcache_s tcache_t; typedef struct tcaches_s tcaches_t; @@ -16,39 +17,9 @@ #define TCACHE_STATE_PURGATORY ((tcache_t *)(uintptr_t)3) #define TCACHE_STATE_MAX TCACHE_STATE_PURGATORY -/* - * Absolute minimum number of cache slots for each small bin. - */ -#define TCACHE_NSLOTS_SMALL_MIN 20 - -/* - * Absolute maximum number of cache slots for each small bin in the thread - * cache. This is an additional constraint beyond that imposed as: twice the - * number of regions per slab for this size class. - * - * This constant must be an even number. - */ -#define TCACHE_NSLOTS_SMALL_MAX 200 - -/* Number of cache slots for large size classes. */ -#define TCACHE_NSLOTS_LARGE 20 - -/* (1U << opt_lg_tcache_max) is used to compute tcache_maxclass. */ -#define LG_TCACHE_MAXCLASS_DEFAULT 15 - -/* - * TCACHE_GC_SWEEP is the approximate number of allocation events between - * full GC sweeps. Integer rounding may cause the actual number to be - * slightly higher, since GC is performed incrementally. - */ -#define TCACHE_GC_SWEEP 8192 - -/* Number of tcache allocation/deallocation events between incremental GCs. */ -#define TCACHE_GC_INCR \ - ((TCACHE_GC_SWEEP / SC_NBINS) + ((TCACHE_GC_SWEEP / SC_NBINS == 0) ? 0 : 1)) - -/* Used in TSD static initializer only. Real init in tcache_data_init(). */ +/* Used in TSD static initializer only. Real init in tsd_tcache_data_init(). */ #define TCACHE_ZERO_INITIALIZER {0} +#define TCACHE_SLOW_ZERO_INITIALIZER {0} /* Used in TSD static initializer only. Will be initialized to opt_tcache. */ #define TCACHE_ENABLED_ZERO_INITIALIZER false @@ -56,4 +27,9 @@ /* Used for explicit tcache only. Means flushed but not destroyed. */ #define TCACHES_ELM_NEED_REINIT ((tcache_t *)(uintptr_t)1) +#define TCACHE_LG_MAXCLASS_LIMIT 23 /* tcache_maxclass = 8M */ +#define TCACHE_MAXCLASS_LIMIT ((size_t)1 << TCACHE_LG_MAXCLASS_LIMIT) +#define TCACHE_NBINS_MAX (SC_NBINS + SC_NGROUP * \ + (TCACHE_LG_MAXCLASS_LIMIT - SC_LG_LARGE_MINCLASS) + 1) + #endif /* JEMALLOC_INTERNAL_TCACHE_TYPES_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/test_hooks.h b/contrib/jemalloc/include/jemalloc/internal/test_hooks.h --- a/contrib/jemalloc/include/jemalloc/internal/test_hooks.h +++ b/contrib/jemalloc/include/jemalloc/internal/test_hooks.h @@ -4,9 +4,21 @@ extern JEMALLOC_EXPORT void (*test_hooks_arena_new_hook)(); extern JEMALLOC_EXPORT void (*test_hooks_libc_hook)(); -#define JEMALLOC_HOOK(fn, hook) ((void)(hook != NULL && (hook(), 0)), fn) +#if defined(JEMALLOC_JET) || defined(JEMALLOC_UNIT_TEST) +# define JEMALLOC_TEST_HOOK(fn, hook) ((void)(hook != NULL && (hook(), 0)), fn) +# define open JEMALLOC_TEST_HOOK(open, test_hooks_libc_hook) +# define read JEMALLOC_TEST_HOOK(read, test_hooks_libc_hook) +# define write JEMALLOC_TEST_HOOK(write, test_hooks_libc_hook) +# define readlink JEMALLOC_TEST_HOOK(readlink, test_hooks_libc_hook) +# define close JEMALLOC_TEST_HOOK(close, test_hooks_libc_hook) +# define creat JEMALLOC_TEST_HOOK(creat, test_hooks_libc_hook) +# define secure_getenv JEMALLOC_TEST_HOOK(secure_getenv, test_hooks_libc_hook) /* Note that this is undef'd and re-define'd in src/prof.c. */ -#define _Unwind_Backtrace JEMALLOC_HOOK(_Unwind_Backtrace, test_hooks_libc_hook) +# define _Unwind_Backtrace JEMALLOC_TEST_HOOK(_Unwind_Backtrace, test_hooks_libc_hook) +#else +# define JEMALLOC_TEST_HOOK(fn, hook) fn +#endif + #endif /* JEMALLOC_INTERNAL_TEST_HOOKS_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/thread_event.h b/contrib/jemalloc/include/jemalloc/internal/thread_event.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/thread_event.h @@ -0,0 +1,301 @@ +#ifndef JEMALLOC_INTERNAL_THREAD_EVENT_H +#define JEMALLOC_INTERNAL_THREAD_EVENT_H + +#include "jemalloc/internal/tsd.h" + +/* "te" is short for "thread_event" */ + +/* + * TE_MIN_START_WAIT should not exceed the minimal allocation usize. + */ +#define TE_MIN_START_WAIT ((uint64_t)1U) +#define TE_MAX_START_WAIT UINT64_MAX + +/* + * Maximum threshold on thread_(de)allocated_next_event_fast, so that there is + * no need to check overflow in malloc fast path. (The allocation size in malloc + * fast path never exceeds SC_LOOKUP_MAXCLASS.) + */ +#define TE_NEXT_EVENT_FAST_MAX (UINT64_MAX - SC_LOOKUP_MAXCLASS + 1U) + +/* + * The max interval helps make sure that malloc stays on the fast path in the + * common case, i.e. thread_allocated < thread_allocated_next_event_fast. When + * thread_allocated is within an event's distance to TE_NEXT_EVENT_FAST_MAX + * above, thread_allocated_next_event_fast is wrapped around and we fall back to + * the medium-fast path. The max interval makes sure that we're not staying on + * the fallback case for too long, even if there's no active event or if all + * active events have long wait times. + */ +#define TE_MAX_INTERVAL ((uint64_t)(4U << 20)) + +/* + * Invalid elapsed time, for situations where elapsed time is not needed. See + * comments in thread_event.c for more info. + */ +#define TE_INVALID_ELAPSED UINT64_MAX + +typedef struct te_ctx_s { + bool is_alloc; + uint64_t *current; + uint64_t *last_event; + uint64_t *next_event; + uint64_t *next_event_fast; +} te_ctx_t; + +void te_assert_invariants_debug(tsd_t *tsd); +void te_event_trigger(tsd_t *tsd, te_ctx_t *ctx); +void te_recompute_fast_threshold(tsd_t *tsd); +void tsd_te_init(tsd_t *tsd); + +/* + * List of all events, in the following format: + * E(event, (condition), is_alloc_event) + */ +#define ITERATE_OVER_ALL_EVENTS \ + E(tcache_gc, (opt_tcache_gc_incr_bytes > 0), true) \ + E(prof_sample, (config_prof && opt_prof), true) \ + E(stats_interval, (opt_stats_interval >= 0), true) \ + E(tcache_gc_dalloc, (opt_tcache_gc_incr_bytes > 0), false) \ + E(peak_alloc, config_stats, true) \ + E(peak_dalloc, config_stats, false) + +#define E(event, condition_unused, is_alloc_event_unused) \ + C(event##_event_wait) + +/* List of all thread event counters. */ +#define ITERATE_OVER_ALL_COUNTERS \ + C(thread_allocated) \ + C(thread_allocated_last_event) \ + ITERATE_OVER_ALL_EVENTS \ + C(prof_sample_last_event) \ + C(stats_interval_last_event) + +/* Getters directly wrap TSD getters. */ +#define C(counter) \ +JEMALLOC_ALWAYS_INLINE uint64_t \ +counter##_get(tsd_t *tsd) { \ + return tsd_##counter##_get(tsd); \ +} + +ITERATE_OVER_ALL_COUNTERS +#undef C + +/* + * Setters call the TSD pointer getters rather than the TSD setters, so that + * the counters can be modified even when TSD state is reincarnated or + * minimal_initialized: if an event is triggered in such cases, we will + * temporarily delay the event and let it be immediately triggered at the next + * allocation call. + */ +#define C(counter) \ +JEMALLOC_ALWAYS_INLINE void \ +counter##_set(tsd_t *tsd, uint64_t v) { \ + *tsd_##counter##p_get(tsd) = v; \ +} + +ITERATE_OVER_ALL_COUNTERS +#undef C + +/* + * For generating _event_wait getter / setter functions for each individual + * event. + */ +#undef E + +/* + * The malloc and free fastpath getters -- use the unsafe getters since tsd may + * be non-nominal, in which case the fast_threshold will be set to 0. This + * allows checking for events and tsd non-nominal in a single branch. + * + * Note that these can only be used on the fastpath. + */ +JEMALLOC_ALWAYS_INLINE void +te_malloc_fastpath_ctx(tsd_t *tsd, uint64_t *allocated, uint64_t *threshold) { + *allocated = *tsd_thread_allocatedp_get_unsafe(tsd); + *threshold = *tsd_thread_allocated_next_event_fastp_get_unsafe(tsd); + assert(*threshold <= TE_NEXT_EVENT_FAST_MAX); +} + +JEMALLOC_ALWAYS_INLINE void +te_free_fastpath_ctx(tsd_t *tsd, uint64_t *deallocated, uint64_t *threshold) { + /* Unsafe getters since this may happen before tsd_init. */ + *deallocated = *tsd_thread_deallocatedp_get_unsafe(tsd); + *threshold = *tsd_thread_deallocated_next_event_fastp_get_unsafe(tsd); + assert(*threshold <= TE_NEXT_EVENT_FAST_MAX); +} + +JEMALLOC_ALWAYS_INLINE bool +te_ctx_is_alloc(te_ctx_t *ctx) { + return ctx->is_alloc; +} + +JEMALLOC_ALWAYS_INLINE uint64_t +te_ctx_current_bytes_get(te_ctx_t *ctx) { + return *ctx->current; +} + +JEMALLOC_ALWAYS_INLINE void +te_ctx_current_bytes_set(te_ctx_t *ctx, uint64_t v) { + *ctx->current = v; +} + +JEMALLOC_ALWAYS_INLINE uint64_t +te_ctx_last_event_get(te_ctx_t *ctx) { + return *ctx->last_event; +} + +JEMALLOC_ALWAYS_INLINE void +te_ctx_last_event_set(te_ctx_t *ctx, uint64_t v) { + *ctx->last_event = v; +} + +/* Below 3 for next_event_fast. */ +JEMALLOC_ALWAYS_INLINE uint64_t +te_ctx_next_event_fast_get(te_ctx_t *ctx) { + uint64_t v = *ctx->next_event_fast; + assert(v <= TE_NEXT_EVENT_FAST_MAX); + return v; +} + +JEMALLOC_ALWAYS_INLINE void +te_ctx_next_event_fast_set(te_ctx_t *ctx, uint64_t v) { + assert(v <= TE_NEXT_EVENT_FAST_MAX); + *ctx->next_event_fast = v; +} + +JEMALLOC_ALWAYS_INLINE void +te_next_event_fast_set_non_nominal(tsd_t *tsd) { + /* + * Set the fast thresholds to zero when tsd is non-nominal. Use the + * unsafe getter as this may get called during tsd init and clean up. + */ + *tsd_thread_allocated_next_event_fastp_get_unsafe(tsd) = 0; + *tsd_thread_deallocated_next_event_fastp_get_unsafe(tsd) = 0; +} + +/* For next_event. Setter also updates the fast threshold. */ +JEMALLOC_ALWAYS_INLINE uint64_t +te_ctx_next_event_get(te_ctx_t *ctx) { + return *ctx->next_event; +} + +JEMALLOC_ALWAYS_INLINE void +te_ctx_next_event_set(tsd_t *tsd, te_ctx_t *ctx, uint64_t v) { + *ctx->next_event = v; + te_recompute_fast_threshold(tsd); +} + +/* + * The function checks in debug mode whether the thread event counters are in + * a consistent state, which forms the invariants before and after each round + * of thread event handling that we can rely on and need to promise. + * The invariants are only temporarily violated in the middle of + * te_event_advance() if an event is triggered (the te_event_trigger() call at + * the end will restore the invariants). + */ +JEMALLOC_ALWAYS_INLINE void +te_assert_invariants(tsd_t *tsd) { + if (config_debug) { + te_assert_invariants_debug(tsd); + } +} + +JEMALLOC_ALWAYS_INLINE void +te_ctx_get(tsd_t *tsd, te_ctx_t *ctx, bool is_alloc) { + ctx->is_alloc = is_alloc; + if (is_alloc) { + ctx->current = tsd_thread_allocatedp_get(tsd); + ctx->last_event = tsd_thread_allocated_last_eventp_get(tsd); + ctx->next_event = tsd_thread_allocated_next_eventp_get(tsd); + ctx->next_event_fast = + tsd_thread_allocated_next_event_fastp_get(tsd); + } else { + ctx->current = tsd_thread_deallocatedp_get(tsd); + ctx->last_event = tsd_thread_deallocated_last_eventp_get(tsd); + ctx->next_event = tsd_thread_deallocated_next_eventp_get(tsd); + ctx->next_event_fast = + tsd_thread_deallocated_next_event_fastp_get(tsd); + } +} + +/* + * The lookahead functionality facilitates events to be able to lookahead, i.e. + * without touching the event counters, to determine whether an event would be + * triggered. The event counters are not advanced until the end of the + * allocation / deallocation calls, so the lookahead can be useful if some + * preparation work for some event must be done early in the allocation / + * deallocation calls. + * + * Currently only the profiling sampling event needs the lookahead + * functionality, so we don't yet define general purpose lookahead functions. + * + * Surplus is a terminology referring to the amount of bytes beyond what's + * needed for triggering an event, which can be a useful quantity to have in + * general when lookahead is being called. + */ + +JEMALLOC_ALWAYS_INLINE bool +te_prof_sample_event_lookahead_surplus(tsd_t *tsd, size_t usize, + size_t *surplus) { + if (surplus != NULL) { + /* + * This is a dead store: the surplus will be overwritten before + * any read. The initialization suppresses compiler warnings. + * Meanwhile, using SIZE_MAX to initialize is good for + * debugging purpose, because a valid surplus value is strictly + * less than usize, which is at most SIZE_MAX. + */ + *surplus = SIZE_MAX; + } + if (unlikely(!tsd_nominal(tsd) || tsd_reentrancy_level_get(tsd) > 0)) { + return false; + } + /* The subtraction is intentionally susceptible to underflow. */ + uint64_t accumbytes = tsd_thread_allocated_get(tsd) + usize - + tsd_thread_allocated_last_event_get(tsd); + uint64_t sample_wait = tsd_prof_sample_event_wait_get(tsd); + if (accumbytes < sample_wait) { + return false; + } + assert(accumbytes - sample_wait < (uint64_t)usize); + if (surplus != NULL) { + *surplus = (size_t)(accumbytes - sample_wait); + } + return true; +} + +JEMALLOC_ALWAYS_INLINE bool +te_prof_sample_event_lookahead(tsd_t *tsd, size_t usize) { + return te_prof_sample_event_lookahead_surplus(tsd, usize, NULL); +} + +JEMALLOC_ALWAYS_INLINE void +te_event_advance(tsd_t *tsd, size_t usize, bool is_alloc) { + te_assert_invariants(tsd); + + te_ctx_t ctx; + te_ctx_get(tsd, &ctx, is_alloc); + + uint64_t bytes_before = te_ctx_current_bytes_get(&ctx); + te_ctx_current_bytes_set(&ctx, bytes_before + usize); + + /* The subtraction is intentionally susceptible to underflow. */ + if (likely(usize < te_ctx_next_event_get(&ctx) - bytes_before)) { + te_assert_invariants(tsd); + } else { + te_event_trigger(tsd, &ctx); + } +} + +JEMALLOC_ALWAYS_INLINE void +thread_dalloc_event(tsd_t *tsd, size_t usize) { + te_event_advance(tsd, usize, false); +} + +JEMALLOC_ALWAYS_INLINE void +thread_alloc_event(tsd_t *tsd, size_t usize) { + te_event_advance(tsd, usize, true); +} + +#endif /* JEMALLOC_INTERNAL_THREAD_EVENT_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/ticker.h b/contrib/jemalloc/include/jemalloc/internal/ticker.h --- a/contrib/jemalloc/include/jemalloc/internal/ticker.h +++ b/contrib/jemalloc/include/jemalloc/internal/ticker.h @@ -1,6 +1,7 @@ #ifndef JEMALLOC_INTERNAL_TICKER_H #define JEMALLOC_INTERNAL_TICKER_H +#include "jemalloc/internal/prng.h" #include "jemalloc/internal/util.h" /** @@ -10,11 +11,11 @@ * have occurred with a call to ticker_ticks), which will return true (and reset * the counter) if the countdown hit zero. */ - -typedef struct { +typedef struct ticker_s ticker_t; +struct ticker_s { int32_t tick; int32_t nticks; -} ticker_t; +}; static inline void ticker_init(ticker_t *ticker, int32_t nticks) { @@ -75,7 +76,7 @@ return ticker_ticks(ticker, 1); } -/* +/* * Try to tick. If ticker would fire, return true, but rely on * slowpath to reset ticker. */ @@ -88,4 +89,87 @@ return false; } +/* + * The ticker_geom_t is much like the ticker_t, except that instead of ticker + * having a constant countdown, it has an approximate one; each tick has + * approximately a 1/nticks chance of triggering the count. + * + * The motivation is in triggering arena decay. With a naive strategy, each + * thread would maintain a ticker per arena, and check if decay is necessary + * each time that the arena's ticker fires. This has two costs: + * - Since under reasonable assumptions both threads and arenas can scale + * linearly with the number of CPUs, maintaining per-arena data in each thread + * scales quadratically with the number of CPUs. + * - These tickers are often a cache miss down tcache flush pathways. + * + * By giving each tick a 1/nticks chance of firing, we still maintain the same + * average number of ticks-until-firing per arena, with only a single ticker's + * worth of metadata. + */ + +/* See ticker.c for an explanation of these constants. */ +#define TICKER_GEOM_NBITS 6 +#define TICKER_GEOM_MUL 61 +extern const uint8_t ticker_geom_table[1 << TICKER_GEOM_NBITS]; + +/* Not actually any different from ticker_t; just for type safety. */ +typedef struct ticker_geom_s ticker_geom_t; +struct ticker_geom_s { + int32_t tick; + int32_t nticks; +}; + +/* + * Just pick the average delay for the first counter. We're more concerned with + * the behavior over long periods of time rather than the exact timing of the + * initial ticks. + */ +#define TICKER_GEOM_INIT(nticks) {nticks, nticks} + +static inline void +ticker_geom_init(ticker_geom_t *ticker, int32_t nticks) { + /* + * Make sure there's no overflow possible. This shouldn't really be a + * problem for reasonable nticks choices, which are all static and + * relatively small. + */ + assert((uint64_t)nticks * (uint64_t)255 / (uint64_t)TICKER_GEOM_MUL + <= (uint64_t)INT32_MAX); + ticker->tick = nticks; + ticker->nticks = nticks; +} + +static inline int32_t +ticker_geom_read(const ticker_geom_t *ticker) { + return ticker->tick; +} + +/* Same deal as above. */ +#if defined(__GNUC__) && !defined(__clang__) \ + && (defined(__x86_64__) || defined(__i386__)) +JEMALLOC_NOINLINE +#endif +static bool +ticker_geom_fixup(ticker_geom_t *ticker, uint64_t *prng_state) { + uint64_t idx = prng_lg_range_u64(prng_state, TICKER_GEOM_NBITS); + ticker->tick = (uint32_t)( + (uint64_t)ticker->nticks * (uint64_t)ticker_geom_table[idx] + / (uint64_t)TICKER_GEOM_MUL); + return true; +} + +static inline bool +ticker_geom_ticks(ticker_geom_t *ticker, uint64_t *prng_state, int32_t nticks) { + ticker->tick -= nticks; + if (unlikely(ticker->tick < 0)) { + return ticker_geom_fixup(ticker, prng_state); + } + return false; +} + +static inline bool +ticker_geom_tick(ticker_geom_t *ticker, uint64_t *prng_state) { + return ticker_geom_ticks(ticker, prng_state, 1); +} + #endif /* JEMALLOC_INTERNAL_TICKER_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/tsd.h b/contrib/jemalloc/include/jemalloc/internal/tsd.h --- a/contrib/jemalloc/include/jemalloc/internal/tsd.h +++ b/contrib/jemalloc/include/jemalloc/internal/tsd.h @@ -1,10 +1,12 @@ #ifndef JEMALLOC_INTERNAL_TSD_H #define JEMALLOC_INTERNAL_TSD_H +#include "jemalloc/internal/activity_callback.h" #include "jemalloc/internal/arena_types.h" #include "jemalloc/internal/assert.h" #include "jemalloc/internal/bin_types.h" #include "jemalloc/internal/jemalloc_internal_externs.h" +#include "jemalloc/internal/peak.h" #include "jemalloc/internal/prof_types.h" #include "jemalloc/internal/ql.h" #include "jemalloc/internal/rtree_tsd.h" @@ -15,39 +17,30 @@ /* * Thread-Specific-Data layout - * --- data accessed on tcache fast path: state, rtree_ctx, stats, prof --- - * s: state - * e: tcache_enabled - * m: thread_allocated (config_stats) - * f: thread_deallocated (config_stats) - * p: prof_tdata (config_prof) - * c: rtree_ctx (rtree cache accessed on deallocation) - * t: tcache - * --- data not accessed on tcache fast path: arena-related fields --- - * d: arenas_tdata_bypass - * r: reentrancy_level - * x: narenas_tdata - * i: iarena - * a: arena - * o: arenas_tdata - * Loading TSD data is on the critical path of basically all malloc operations. - * In particular, tcache and rtree_ctx rely on hot CPU cache to be effective. - * Use a compact layout to reduce cache footprint. - * +--- 64-bit and 64B cacheline; 1B each letter; First byte on the left. ---+ - * |---------------------------- 1st cacheline ----------------------------| - * | sedrxxxx mmmmmmmm ffffffff pppppppp [c * 32 ........ ........ .......] | - * |---------------------------- 2nd cacheline ----------------------------| - * | [c * 64 ........ ........ ........ ........ ........ ........ .......] | - * |---------------------------- 3nd cacheline ----------------------------| - * | [c * 32 ........ ........ .......] iiiiiiii aaaaaaaa oooooooo [t...... | - * +-------------------------------------------------------------------------+ - * Note: the entire tcache is embedded into TSD and spans multiple cachelines. * - * The last 3 members (i, a and o) before tcache isn't really needed on tcache - * fast path. However we have a number of unused tcache bins and witnesses - * (never touched unless config_debug) at the end of tcache, so we place them - * there to avoid breaking the cachelines and possibly paging in an extra page. + * At least some thread-local data gets touched on the fast-path of almost all + * malloc operations. But much of it is only necessary down slow-paths, or + * testing. We want to colocate the fast-path data so that it can live on the + * same cacheline if possible. So we define three tiers of hotness: + * TSD_DATA_FAST: Touched on the alloc/dalloc fast paths. + * TSD_DATA_SLOW: Touched down slow paths. "Slow" here is sort of general; + * there are "semi-slow" paths like "not a sized deallocation, but can still + * live in the tcache". We'll want to keep these closer to the fast-path + * data. + * TSD_DATA_SLOWER: Only touched in test or debug modes, or not touched at all. + * + * An additional concern is that the larger tcache bins won't be used (we have a + * bin per size class, but by default only cache relatively small objects). So + * the earlier bins are in the TSD_DATA_FAST tier, but the later ones are in the + * TSD_DATA_SLOWER tier. + * + * As a result of all this, we put the slow data first, then the fast data, then + * the slower data, while keeping the tcache as the last element of the fast + * data (so that the fast -> slower transition happens midway through the + * tcache). While we don't yet play alignment tricks to guarantee it, this + * increases our odds of getting some cache/page locality on fast paths. */ + #ifdef JEMALLOC_JET typedef void (*test_callback_t)(int *); # define MALLOC_TSD_TEST_DATA_INIT 0x72b65c10 @@ -60,50 +53,112 @@ # define MALLOC_TEST_TSD_INITIALIZER #endif -/* O(name, type, nullable type */ -#define MALLOC_TSD \ +typedef ql_elm(tsd_t) tsd_link_t; + +/* O(name, type, nullable type) */ +#define TSD_DATA_SLOW \ O(tcache_enabled, bool, bool) \ - O(arenas_tdata_bypass, bool, bool) \ O(reentrancy_level, int8_t, int8_t) \ - O(narenas_tdata, uint32_t, uint32_t) \ - O(offset_state, uint64_t, uint64_t) \ - O(thread_allocated, uint64_t, uint64_t) \ - O(thread_deallocated, uint64_t, uint64_t) \ - O(bytes_until_sample, int64_t, int64_t) \ + O(thread_allocated_last_event, uint64_t, uint64_t) \ + O(thread_allocated_next_event, uint64_t, uint64_t) \ + O(thread_deallocated_last_event, uint64_t, uint64_t) \ + O(thread_deallocated_next_event, uint64_t, uint64_t) \ + O(tcache_gc_event_wait, uint64_t, uint64_t) \ + O(tcache_gc_dalloc_event_wait, uint64_t, uint64_t) \ + O(prof_sample_event_wait, uint64_t, uint64_t) \ + O(prof_sample_last_event, uint64_t, uint64_t) \ + O(stats_interval_event_wait, uint64_t, uint64_t) \ + O(stats_interval_last_event, uint64_t, uint64_t) \ + O(peak_alloc_event_wait, uint64_t, uint64_t) \ + O(peak_dalloc_event_wait, uint64_t, uint64_t) \ O(prof_tdata, prof_tdata_t *, prof_tdata_t *) \ - O(rtree_ctx, rtree_ctx_t, rtree_ctx_t) \ + O(prng_state, uint64_t, uint64_t) \ + O(san_extents_until_guard_small, uint64_t, uint64_t) \ + O(san_extents_until_guard_large, uint64_t, uint64_t) \ O(iarena, arena_t *, arena_t *) \ O(arena, arena_t *, arena_t *) \ - O(arenas_tdata, arena_tdata_t *, arena_tdata_t *)\ + O(arena_decay_ticker, ticker_geom_t, ticker_geom_t) \ + O(sec_shard, uint8_t, uint8_t) \ O(binshards, tsd_binshards_t, tsd_binshards_t)\ - O(tcache, tcache_t, tcache_t) \ + O(tsd_link, tsd_link_t, tsd_link_t) \ + O(in_hook, bool, bool) \ + O(peak, peak_t, peak_t) \ + O(activity_callback_thunk, activity_callback_thunk_t, \ + activity_callback_thunk_t) \ + O(tcache_slow, tcache_slow_t, tcache_slow_t) \ + O(rtree_ctx, rtree_ctx_t, rtree_ctx_t) + +#define TSD_DATA_SLOW_INITIALIZER \ + /* tcache_enabled */ TCACHE_ENABLED_ZERO_INITIALIZER, \ + /* reentrancy_level */ 0, \ + /* thread_allocated_last_event */ 0, \ + /* thread_allocated_next_event */ 0, \ + /* thread_deallocated_last_event */ 0, \ + /* thread_deallocated_next_event */ 0, \ + /* tcache_gc_event_wait */ 0, \ + /* tcache_gc_dalloc_event_wait */ 0, \ + /* prof_sample_event_wait */ 0, \ + /* prof_sample_last_event */ 0, \ + /* stats_interval_event_wait */ 0, \ + /* stats_interval_last_event */ 0, \ + /* peak_alloc_event_wait */ 0, \ + /* peak_dalloc_event_wait */ 0, \ + /* prof_tdata */ NULL, \ + /* prng_state */ 0, \ + /* san_extents_until_guard_small */ 0, \ + /* san_extents_until_guard_large */ 0, \ + /* iarena */ NULL, \ + /* arena */ NULL, \ + /* arena_decay_ticker */ \ + TICKER_GEOM_INIT(ARENA_DECAY_NTICKS_PER_UPDATE), \ + /* sec_shard */ (uint8_t)-1, \ + /* binshards */ TSD_BINSHARDS_ZERO_INITIALIZER, \ + /* tsd_link */ {NULL}, \ + /* in_hook */ false, \ + /* peak */ PEAK_INITIALIZER, \ + /* activity_callback_thunk */ \ + ACTIVITY_CALLBACK_THUNK_INITIALIZER, \ + /* tcache_slow */ TCACHE_SLOW_ZERO_INITIALIZER, \ + /* rtree_ctx */ RTREE_CTX_INITIALIZER, + +/* O(name, type, nullable type) */ +#define TSD_DATA_FAST \ + O(thread_allocated, uint64_t, uint64_t) \ + O(thread_allocated_next_event_fast, uint64_t, uint64_t) \ + O(thread_deallocated, uint64_t, uint64_t) \ + O(thread_deallocated_next_event_fast, uint64_t, uint64_t) \ + O(tcache, tcache_t, tcache_t) + +#define TSD_DATA_FAST_INITIALIZER \ + /* thread_allocated */ 0, \ + /* thread_allocated_next_event_fast */ 0, \ + /* thread_deallocated */ 0, \ + /* thread_deallocated_next_event_fast */ 0, \ + /* tcache */ TCACHE_ZERO_INITIALIZER, + +/* O(name, type, nullable type) */ +#define TSD_DATA_SLOWER \ O(witness_tsd, witness_tsd_t, witness_tsdn_t) \ MALLOC_TEST_TSD +#define TSD_DATA_SLOWER_INITIALIZER \ + /* witness */ WITNESS_TSD_INITIALIZER \ + /* test data */ MALLOC_TEST_TSD_INITIALIZER + + #define TSD_INITIALIZER { \ - ATOMIC_INIT(tsd_state_uninitialized), \ - TCACHE_ENABLED_ZERO_INITIALIZER, \ - false, \ - 0, \ - 0, \ - 0, \ - 0, \ - 0, \ - 0, \ - NULL, \ - RTREE_CTX_ZERO_INITIALIZER, \ - NULL, \ - NULL, \ - NULL, \ - TSD_BINSHARDS_ZERO_INITIALIZER, \ - TCACHE_ZERO_INITIALIZER, \ - WITNESS_TSD_INITIALIZER \ - MALLOC_TEST_TSD_INITIALIZER \ + TSD_DATA_SLOW_INITIALIZER \ + /* state */ ATOMIC_INIT(tsd_state_uninitialized), \ + TSD_DATA_FAST_INITIALIZER \ + TSD_DATA_SLOWER_INITIALIZER \ } +#if defined(JEMALLOC_MALLOC_THREAD_CLEANUP) || defined(_WIN32) +void _malloc_tsd_cleanup_register(bool (*f)(void)); +#endif + void *malloc_tsd_malloc(size_t size); void malloc_tsd_dalloc(void *wrapper); -void malloc_tsd_cleanup_register(bool (*f)(void)); tsd_t *malloc_tsd_boot0(void); void malloc_tsd_boot1(void); void tsd_cleanup(void *arg); @@ -189,14 +244,17 @@ * setters below. */ +#define O(n, t, nt) \ + t TSD_MANGLE(n); + + TSD_DATA_SLOW /* * We manually limit the state to just a single byte. Unless the 8-bit * atomics are unavailable (which is rare). */ tsd_state_t state; -#define O(n, t, nt) \ - t TSD_MANGLE(n); -MALLOC_TSD + TSD_DATA_FAST + TSD_DATA_SLOWER #undef O /* AddressSanitizer requires TLS data to be aligned to at least 8 bytes. */ } JEMALLOC_ALIGNED(16); @@ -263,7 +321,9 @@ tsd_##n##p_get_unsafe(tsd_t *tsd) { \ return &tsd->TSD_MANGLE(n); \ } -MALLOC_TSD +TSD_DATA_SLOW +TSD_DATA_FAST +TSD_DATA_SLOWER #undef O /* tsd_foop_get(tsd) returns a pointer to the thread-local instance of foo. */ @@ -282,7 +342,9 @@ state == tsd_state_minimal_initialized); \ return tsd_##n##p_get_unsafe(tsd); \ } -MALLOC_TSD +TSD_DATA_SLOW +TSD_DATA_FAST +TSD_DATA_SLOWER #undef O /* @@ -298,7 +360,9 @@ tsd_t *tsd = tsdn_tsd(tsdn); \ return (nt *)tsd_##n##p_get(tsd); \ } -MALLOC_TSD +TSD_DATA_SLOW +TSD_DATA_FAST +TSD_DATA_SLOWER #undef O /* tsd_foo_get(tsd) returns the value of the thread-local instance of foo. */ @@ -307,7 +371,9 @@ tsd_##n##_get(tsd_t *tsd) { \ return *tsd_##n##p_get(tsd); \ } -MALLOC_TSD +TSD_DATA_SLOW +TSD_DATA_FAST +TSD_DATA_SLOWER #undef O /* tsd_foo_set(tsd, val) updates the thread-local instance of foo to be val. */ @@ -318,7 +384,9 @@ tsd_state_get(tsd) != tsd_state_minimal_initialized); \ *tsd_##n##p_get(tsd) = val; \ } -MALLOC_TSD +TSD_DATA_SLOW +TSD_DATA_FAST +TSD_DATA_SLOWER #undef O JEMALLOC_ALWAYS_INLINE void @@ -383,7 +451,10 @@ static inline bool tsd_nominal(tsd_t *tsd) { - return (tsd_state_get(tsd) <= tsd_state_nominal_max); + bool nominal = tsd_state_get(tsd) <= tsd_state_nominal_max; + assert(nominal || tsd_reentrancy_level_get(tsd) > 0); + + return nominal; } JEMALLOC_ALWAYS_INLINE tsdn_t * @@ -413,4 +484,36 @@ return tsd_rtree_ctx(tsdn_tsd(tsdn)); } +static inline bool +tsd_state_nocleanup(tsd_t *tsd) { + return tsd_state_get(tsd) == tsd_state_reincarnated || + tsd_state_get(tsd) == tsd_state_minimal_initialized; +} + +/* + * These "raw" tsd reentrancy functions don't have any debug checking to make + * sure that we're not touching arena 0. Better is to call pre_reentrancy and + * post_reentrancy if this is possible. + */ +static inline void +tsd_pre_reentrancy_raw(tsd_t *tsd) { + bool fast = tsd_fast(tsd); + assert(tsd_reentrancy_level_get(tsd) < INT8_MAX); + ++*tsd_reentrancy_levelp_get(tsd); + if (fast) { + /* Prepare slow path for reentrancy. */ + tsd_slow_update(tsd); + assert(tsd_state_get(tsd) == tsd_state_nominal_slow); + } +} + +static inline void +tsd_post_reentrancy_raw(tsd_t *tsd) { + int8_t *reentrancy_level = tsd_reentrancy_levelp_get(tsd); + assert(*reentrancy_level > 0); + if (--*reentrancy_level == 0) { + tsd_slow_update(tsd); + } +} + #endif /* JEMALLOC_INTERNAL_TSD_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/tsd_generic.h b/contrib/jemalloc/include/jemalloc/internal/tsd_generic.h --- a/contrib/jemalloc/include/jemalloc/internal/tsd_generic.h +++ b/contrib/jemalloc/include/jemalloc/internal/tsd_generic.h @@ -52,6 +52,9 @@ JEMALLOC_ALWAYS_INLINE void tsd_wrapper_set(tsd_wrapper_t *wrapper) { + if (unlikely(!tsd_booted)) { + return; + } if (pthread_setspecific(tsd_tsd, (void *)wrapper) != 0) { malloc_write(": Error setting TSD\n"); abort(); @@ -60,7 +63,13 @@ JEMALLOC_ALWAYS_INLINE tsd_wrapper_t * tsd_wrapper_get(bool init) { - tsd_wrapper_t *wrapper = (tsd_wrapper_t *)pthread_getspecific(tsd_tsd); + tsd_wrapper_t *wrapper; + + if (unlikely(!tsd_booted)) { + return &tsd_boot_wrapper; + } + + wrapper = (tsd_wrapper_t *)pthread_getspecific(tsd_tsd); if (init && unlikely(wrapper == NULL)) { tsd_init_block_t block; @@ -91,11 +100,21 @@ JEMALLOC_ALWAYS_INLINE bool tsd_boot0(void) { + tsd_wrapper_t *wrapper; + tsd_init_block_t block; + + wrapper = (tsd_wrapper_t *) + tsd_init_check_recursion(&tsd_init_head, &block); + if (wrapper) { + return false; + } + block.data = &tsd_boot_wrapper; if (pthread_key_create(&tsd_tsd, tsd_cleanup_wrapper) != 0) { return true; } - tsd_wrapper_set(&tsd_boot_wrapper); tsd_booted = true; + tsd_wrapper_set(&tsd_boot_wrapper); + tsd_init_finish(&tsd_init_head, &block); return false; } diff --git a/contrib/jemalloc/include/jemalloc/internal/tsd_malloc_thread_cleanup.h b/contrib/jemalloc/include/jemalloc/internal/tsd_malloc_thread_cleanup.h --- a/contrib/jemalloc/include/jemalloc/internal/tsd_malloc_thread_cleanup.h +++ b/contrib/jemalloc/include/jemalloc/internal/tsd_malloc_thread_cleanup.h @@ -21,7 +21,7 @@ JEMALLOC_ALWAYS_INLINE bool tsd_boot0(void) { - malloc_tsd_cleanup_register(&tsd_cleanup_wrapper); + _malloc_tsd_cleanup_register(&tsd_cleanup_wrapper); tsd_booted = true; return false; } diff --git a/contrib/jemalloc/include/jemalloc/internal/tsd_types.h b/contrib/jemalloc/include/jemalloc/internal/tsd_types.h --- a/contrib/jemalloc/include/jemalloc/internal/tsd_types.h +++ b/contrib/jemalloc/include/jemalloc/internal/tsd_types.h @@ -1,7 +1,7 @@ #ifndef JEMALLOC_INTERNAL_TSD_TYPES_H #define JEMALLOC_INTERNAL_TSD_TYPES_H -#define MALLOC_TSD_CLEANUPS_MAX 2 +#define MALLOC_TSD_CLEANUPS_MAX 4 typedef struct tsd_s tsd_t; typedef struct tsdn_s tsdn_t; diff --git a/contrib/jemalloc/include/jemalloc/internal/typed_list.h b/contrib/jemalloc/include/jemalloc/internal/typed_list.h new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/include/jemalloc/internal/typed_list.h @@ -0,0 +1,55 @@ +#ifndef JEMALLOC_INTERNAL_TYPED_LIST_H +#define JEMALLOC_INTERNAL_TYPED_LIST_H + +/* + * This wraps the ql module to implement a list class in a way that's a little + * bit easier to use; it handles ql_elm_new calls and provides type safety. + */ + +#define TYPED_LIST(list_type, el_type, linkage) \ +typedef struct { \ + ql_head(el_type) head; \ +} list_type##_t; \ +static inline void \ +list_type##_init(list_type##_t *list) { \ + ql_new(&list->head); \ +} \ +static inline el_type * \ +list_type##_first(const list_type##_t *list) { \ + return ql_first(&list->head); \ +} \ +static inline el_type * \ +list_type##_last(const list_type##_t *list) { \ + return ql_last(&list->head, linkage); \ +} \ +static inline void \ +list_type##_append(list_type##_t *list, el_type *item) { \ + ql_elm_new(item, linkage); \ + ql_tail_insert(&list->head, item, linkage); \ +} \ +static inline void \ +list_type##_prepend(list_type##_t *list, el_type *item) { \ + ql_elm_new(item, linkage); \ + ql_head_insert(&list->head, item, linkage); \ +} \ +static inline void \ +list_type##_replace(list_type##_t *list, el_type *to_remove, \ + el_type *to_insert) { \ + ql_elm_new(to_insert, linkage); \ + ql_after_insert(to_remove, to_insert, linkage); \ + ql_remove(&list->head, to_remove, linkage); \ +} \ +static inline void \ +list_type##_remove(list_type##_t *list, el_type *item) { \ + ql_remove(&list->head, item, linkage); \ +} \ +static inline bool \ +list_type##_empty(list_type##_t *list) { \ + return ql_empty(&list->head); \ +} \ +static inline void \ +list_type##_concat(list_type##_t *list_a, list_type##_t *list_b) { \ + ql_concat(&list_a->head, &list_b->head, linkage); \ +} + +#endif /* JEMALLOC_INTERNAL_TYPED_LIST_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/util.h b/contrib/jemalloc/include/jemalloc/internal/util.h --- a/contrib/jemalloc/include/jemalloc/internal/util.h +++ b/contrib/jemalloc/include/jemalloc/internal/util.h @@ -62,6 +62,62 @@ #endif } +JEMALLOC_ALWAYS_INLINE void +util_assume(bool b) { + if (!b) { + unreachable(); + } +} + +/* ptr should be valid. */ +JEMALLOC_ALWAYS_INLINE void +util_prefetch_read(void *ptr) { + /* + * This should arguably be a config check; but any version of GCC so old + * that it doesn't support __builtin_prefetch is also too old to build + * jemalloc. + */ +#ifdef __GNUC__ + if (config_debug) { + /* Enforce the "valid ptr" requirement. */ + *(volatile char *)ptr; + } + __builtin_prefetch(ptr, /* read or write */ 0, /* locality hint */ 3); +#else + *(volatile char *)ptr; +#endif +} + +JEMALLOC_ALWAYS_INLINE void +util_prefetch_write(void *ptr) { +#ifdef __GNUC__ + if (config_debug) { + *(volatile char *)ptr; + } + /* + * The only difference from the read variant is that this has a 1 as the + * second argument (the write hint). + */ + __builtin_prefetch(ptr, 1, 3); +#else + *(volatile char *)ptr; +#endif +} + +JEMALLOC_ALWAYS_INLINE void +util_prefetch_read_range(void *ptr, size_t sz) { + for (size_t i = 0; i < sz; i += CACHELINE) { + util_prefetch_read((void *)((uintptr_t)ptr + i)); + } +} + +JEMALLOC_ALWAYS_INLINE void +util_prefetch_write_range(void *ptr, size_t sz) { + for (size_t i = 0; i < sz; i += CACHELINE) { + util_prefetch_write((void *)((uintptr_t)ptr + i)); + } +} + #undef UTIL_INLINE #endif /* JEMALLOC_INTERNAL_UTIL_H */ diff --git a/contrib/jemalloc/include/jemalloc/internal/witness.h b/contrib/jemalloc/include/jemalloc/internal/witness.h --- a/contrib/jemalloc/include/jemalloc/internal/witness.h +++ b/contrib/jemalloc/include/jemalloc/internal/witness.h @@ -7,60 +7,76 @@ /* LOCK RANKS */ /******************************************************************************/ -/* - * Witnesses with rank WITNESS_RANK_OMIT are completely ignored by the witness - * machinery. - */ - -#define WITNESS_RANK_OMIT 0U - -#define WITNESS_RANK_MIN 1U - -#define WITNESS_RANK_INIT 1U -#define WITNESS_RANK_CTL 1U -#define WITNESS_RANK_TCACHES 2U -#define WITNESS_RANK_ARENAS 3U - -#define WITNESS_RANK_BACKGROUND_THREAD_GLOBAL 4U - -#define WITNESS_RANK_PROF_DUMP 5U -#define WITNESS_RANK_PROF_BT2GCTX 6U -#define WITNESS_RANK_PROF_TDATAS 7U -#define WITNESS_RANK_PROF_TDATA 8U -#define WITNESS_RANK_PROF_LOG 9U -#define WITNESS_RANK_PROF_GCTX 10U -#define WITNESS_RANK_BACKGROUND_THREAD 11U - -/* - * Used as an argument to witness_assert_depth_to_rank() in order to validate - * depth excluding non-core locks with lower ranks. Since the rank argument to - * witness_assert_depth_to_rank() is inclusive rather than exclusive, this - * definition can have the same value as the minimally ranked core lock. - */ -#define WITNESS_RANK_CORE 12U - -#define WITNESS_RANK_DECAY 12U -#define WITNESS_RANK_TCACHE_QL 13U -#define WITNESS_RANK_EXTENT_GROW 14U -#define WITNESS_RANK_EXTENTS 15U -#define WITNESS_RANK_EXTENT_AVAIL 16U - -#define WITNESS_RANK_EXTENT_POOL 17U -#define WITNESS_RANK_RTREE 18U -#define WITNESS_RANK_BASE 19U -#define WITNESS_RANK_ARENA_LARGE 20U -#define WITNESS_RANK_HOOK 21U - -#define WITNESS_RANK_LEAF 0xffffffffU -#define WITNESS_RANK_BIN WITNESS_RANK_LEAF -#define WITNESS_RANK_ARENA_STATS WITNESS_RANK_LEAF -#define WITNESS_RANK_DSS WITNESS_RANK_LEAF -#define WITNESS_RANK_PROF_ACTIVE WITNESS_RANK_LEAF -#define WITNESS_RANK_PROF_ACCUM WITNESS_RANK_LEAF -#define WITNESS_RANK_PROF_DUMP_SEQ WITNESS_RANK_LEAF -#define WITNESS_RANK_PROF_GDUMP WITNESS_RANK_LEAF -#define WITNESS_RANK_PROF_NEXT_THR_UID WITNESS_RANK_LEAF -#define WITNESS_RANK_PROF_THREAD_ACTIVE_INIT WITNESS_RANK_LEAF +enum witness_rank_e { + /* + * Order matters within this enum listing -- higher valued locks can + * only be acquired after lower-valued ones. We use the + * auto-incrementing-ness of enum values to enforce this. + */ + + /* + * Witnesses with rank WITNESS_RANK_OMIT are completely ignored by the + * witness machinery. + */ + WITNESS_RANK_OMIT, + WITNESS_RANK_MIN, + WITNESS_RANK_INIT = WITNESS_RANK_MIN, + WITNESS_RANK_CTL, + WITNESS_RANK_TCACHES, + WITNESS_RANK_ARENAS, + WITNESS_RANK_BACKGROUND_THREAD_GLOBAL, + WITNESS_RANK_PROF_DUMP, + WITNESS_RANK_PROF_BT2GCTX, + WITNESS_RANK_PROF_TDATAS, + WITNESS_RANK_PROF_TDATA, + WITNESS_RANK_PROF_LOG, + WITNESS_RANK_PROF_GCTX, + WITNESS_RANK_PROF_RECENT_DUMP, + WITNESS_RANK_BACKGROUND_THREAD, + /* + * Used as an argument to witness_assert_depth_to_rank() in order to + * validate depth excluding non-core locks with lower ranks. Since the + * rank argument to witness_assert_depth_to_rank() is inclusive rather + * than exclusive, this definition can have the same value as the + * minimally ranked core lock. + */ + WITNESS_RANK_CORE, + WITNESS_RANK_DECAY = WITNESS_RANK_CORE, + WITNESS_RANK_TCACHE_QL, + + WITNESS_RANK_SEC_SHARD, + + WITNESS_RANK_EXTENT_GROW, + WITNESS_RANK_HPA_SHARD_GROW = WITNESS_RANK_EXTENT_GROW, + WITNESS_RANK_SAN_BUMP_ALLOC = WITNESS_RANK_EXTENT_GROW, + + WITNESS_RANK_EXTENTS, + WITNESS_RANK_HPA_SHARD = WITNESS_RANK_EXTENTS, + + WITNESS_RANK_HPA_CENTRAL_GROW, + WITNESS_RANK_HPA_CENTRAL, + + WITNESS_RANK_EDATA_CACHE, + + WITNESS_RANK_RTREE, + WITNESS_RANK_BASE, + WITNESS_RANK_ARENA_LARGE, + WITNESS_RANK_HOOK, + + WITNESS_RANK_LEAF=0x1000, + WITNESS_RANK_BIN = WITNESS_RANK_LEAF, + WITNESS_RANK_ARENA_STATS = WITNESS_RANK_LEAF, + WITNESS_RANK_COUNTER_ACCUM = WITNESS_RANK_LEAF, + WITNESS_RANK_DSS = WITNESS_RANK_LEAF, + WITNESS_RANK_PROF_ACTIVE = WITNESS_RANK_LEAF, + WITNESS_RANK_PROF_DUMP_FILENAME = WITNESS_RANK_LEAF, + WITNESS_RANK_PROF_GDUMP = WITNESS_RANK_LEAF, + WITNESS_RANK_PROF_NEXT_THR_UID = WITNESS_RANK_LEAF, + WITNESS_RANK_PROF_RECENT_ALLOC = WITNESS_RANK_LEAF, + WITNESS_RANK_PROF_STATS = WITNESS_RANK_LEAF, + WITNESS_RANK_PROF_THREAD_ACTIVE_INIT = WITNESS_RANK_LEAF, +}; +typedef enum witness_rank_e witness_rank_t; /******************************************************************************/ /* PER-WITNESS DATA */ @@ -72,7 +88,6 @@ #endif typedef struct witness_s witness_t; -typedef unsigned witness_rank_t; typedef ql_head(witness_t) witness_list_t; typedef int witness_comp_t (const witness_t *, void *, const witness_t *, void *); @@ -82,8 +97,8 @@ const char *name; /* - * Witness rank, where 0 is lowest and UINT_MAX is highest. Witnesses - * must be acquired in order of increasing rank. + * Witness rank, where 0 is lowest and WITNESS_RANK_LEAF is highest. + * Witnesses must be acquired in order of increasing rank. */ witness_rank_t rank; @@ -228,26 +243,13 @@ } } -static inline void -witness_assert_depth_to_rank(witness_tsdn_t *witness_tsdn, - witness_rank_t rank_inclusive, unsigned depth) { - witness_tsd_t *witness_tsd; - unsigned d; - witness_list_t *witnesses; - witness_t *w; - - if (!config_debug) { - return; - } +/* Returns depth. Not intended for direct use. */ +static inline unsigned +witness_depth_to_rank(witness_list_t *witnesses, witness_rank_t rank_inclusive) +{ + unsigned d = 0; + witness_t *w = ql_last(witnesses, link); - if (witness_tsdn_null(witness_tsdn)) { - return; - } - witness_tsd = witness_tsdn_tsd(witness_tsdn); - - d = 0; - witnesses = &witness_tsd->witnesses; - w = ql_last(witnesses, link); if (w != NULL) { ql_reverse_foreach(w, witnesses, link) { if (w->rank < rank_inclusive) { @@ -256,6 +258,20 @@ d++; } } + + return d; +} + +static inline void +witness_assert_depth_to_rank(witness_tsdn_t *witness_tsdn, + witness_rank_t rank_inclusive, unsigned depth) { + if (!config_debug || witness_tsdn_null(witness_tsdn)) { + return; + } + + witness_list_t *witnesses = &witness_tsdn_tsd(witness_tsdn)->witnesses; + unsigned d = witness_depth_to_rank(witnesses, rank_inclusive); + if (d != depth) { witness_depth_error(witnesses, rank_inclusive, depth); } @@ -271,6 +287,21 @@ witness_assert_depth(witness_tsdn, 0); } +static inline void +witness_assert_positive_depth_to_rank(witness_tsdn_t *witness_tsdn, + witness_rank_t rank_inclusive) { + if (!config_debug || witness_tsdn_null(witness_tsdn)) { + return; + } + + witness_list_t *witnesses = &witness_tsdn_tsd(witness_tsdn)->witnesses; + unsigned d = witness_depth_to_rank(witnesses, rank_inclusive); + + if (d == 0) { + witness_depth_error(witnesses, rank_inclusive, 1); + } +} + static inline void witness_lock(witness_tsdn_t *witness_tsdn, witness_t *witness) { witness_tsd_t *witness_tsd; diff --git a/contrib/jemalloc/src/arena.c b/contrib/jemalloc/src/arena.c --- a/contrib/jemalloc/src/arena.c +++ b/contrib/jemalloc/src/arena.c @@ -1,11 +1,12 @@ -#define JEMALLOC_ARENA_C_ #include "jemalloc/internal/jemalloc_preamble.h" #include "jemalloc/internal/jemalloc_internal_includes.h" #include "jemalloc/internal/assert.h" -#include "jemalloc/internal/div.h" +#include "jemalloc/internal/decay.h" +#include "jemalloc/internal/ehooks.h" #include "jemalloc/internal/extent_dss.h" #include "jemalloc/internal/extent_mmap.h" +#include "jemalloc/internal/san.h" #include "jemalloc/internal/mutex.h" #include "jemalloc/internal/rtree.h" #include "jemalloc/internal/safety_check.h" @@ -35,34 +36,37 @@ static atomic_zd_t dirty_decay_ms_default; static atomic_zd_t muzzy_decay_ms_default; -const uint64_t h_steps[SMOOTHSTEP_NSTEPS] = { -#define STEP(step, h, x, y) \ - h, - SMOOTHSTEP -#undef STEP -}; +emap_t arena_emap_global; +pa_central_t arena_pa_central_global; -static div_info_t arena_binind_div_info[SC_NBINS]; +div_info_t arena_binind_div_info[SC_NBINS]; size_t opt_oversize_threshold = OVERSIZE_THRESHOLD_DEFAULT; size_t oversize_threshold = OVERSIZE_THRESHOLD_DEFAULT; + +uint32_t arena_bin_offsets[SC_NBINS]; +static unsigned nbins_total; + static unsigned huge_arena_ind; +const arena_config_t arena_config_default = { + /* .extent_hooks = */ (extent_hooks_t *)&ehooks_default_extent_hooks, + /* .metadata_use_hooks = */ true, +}; + /******************************************************************************/ /* * Function prototypes for static functions that are referenced prior to * definition. */ -static void arena_decay_to_limit(tsdn_t *tsdn, arena_t *arena, - arena_decay_t *decay, extents_t *extents, bool all, size_t npages_limit, - size_t npages_decay_max, bool is_background_thread); static bool arena_decay_dirty(tsdn_t *tsdn, arena_t *arena, bool is_background_thread, bool all); -static void arena_dalloc_bin_slab(tsdn_t *tsdn, arena_t *arena, extent_t *slab, - bin_t *bin); -static void arena_bin_lower_slab(tsdn_t *tsdn, arena_t *arena, extent_t *slab, +static void arena_bin_lower_slab(tsdn_t *tsdn, arena_t *arena, edata_t *slab, bin_t *bin); +static void +arena_maybe_do_deferred_work(tsdn_t *tsdn, arena_t *arena, decay_t *decay, + size_t npages_new); /******************************************************************************/ @@ -72,19 +76,17 @@ size_t *nactive, size_t *ndirty, size_t *nmuzzy) { *nthreads += arena_nthreads_get(arena, false); *dss = dss_prec_names[arena_dss_prec_get(arena)]; - *dirty_decay_ms = arena_dirty_decay_ms_get(arena); - *muzzy_decay_ms = arena_muzzy_decay_ms_get(arena); - *nactive += atomic_load_zu(&arena->nactive, ATOMIC_RELAXED); - *ndirty += extents_npages_get(&arena->extents_dirty); - *nmuzzy += extents_npages_get(&arena->extents_muzzy); + *dirty_decay_ms = arena_decay_ms_get(arena, extent_state_dirty); + *muzzy_decay_ms = arena_decay_ms_get(arena, extent_state_muzzy); + pa_shard_basic_stats_merge(&arena->pa_shard, nactive, ndirty, nmuzzy); } void arena_stats_merge(tsdn_t *tsdn, arena_t *arena, unsigned *nthreads, const char **dss, ssize_t *dirty_decay_ms, ssize_t *muzzy_decay_ms, size_t *nactive, size_t *ndirty, size_t *nmuzzy, arena_stats_t *astats, - bin_stats_t *bstats, arena_stats_large_t *lstats, - arena_stats_extents_t *estats) { + bin_stats_data_t *bstats, arena_stats_large_t *lstats, + pac_estats_t *estats, hpa_shard_stats_t *hpastats, sec_stats_t *secstats) { cassert(config_stats); arena_basic_stats_merge(tsdn, arena, nthreads, dss, dirty_decay_ms, @@ -93,122 +95,74 @@ size_t base_allocated, base_resident, base_mapped, metadata_thp; base_stats_get(tsdn, arena->base, &base_allocated, &base_resident, &base_mapped, &metadata_thp); + size_t pac_mapped_sz = pac_mapped(&arena->pa_shard.pac); + astats->mapped += base_mapped + pac_mapped_sz; + astats->resident += base_resident; - arena_stats_lock(tsdn, &arena->stats); - - arena_stats_accum_zu(&astats->mapped, base_mapped - + arena_stats_read_zu(tsdn, &arena->stats, &arena->stats.mapped)); - arena_stats_accum_zu(&astats->retained, - extents_npages_get(&arena->extents_retained) << LG_PAGE); - - atomic_store_zu(&astats->extent_avail, - atomic_load_zu(&arena->extent_avail_cnt, ATOMIC_RELAXED), - ATOMIC_RELAXED); + LOCKEDINT_MTX_LOCK(tsdn, arena->stats.mtx); - arena_stats_accum_u64(&astats->decay_dirty.npurge, - arena_stats_read_u64(tsdn, &arena->stats, - &arena->stats.decay_dirty.npurge)); - arena_stats_accum_u64(&astats->decay_dirty.nmadvise, - arena_stats_read_u64(tsdn, &arena->stats, - &arena->stats.decay_dirty.nmadvise)); - arena_stats_accum_u64(&astats->decay_dirty.purged, - arena_stats_read_u64(tsdn, &arena->stats, - &arena->stats.decay_dirty.purged)); - - arena_stats_accum_u64(&astats->decay_muzzy.npurge, - arena_stats_read_u64(tsdn, &arena->stats, - &arena->stats.decay_muzzy.npurge)); - arena_stats_accum_u64(&astats->decay_muzzy.nmadvise, - arena_stats_read_u64(tsdn, &arena->stats, - &arena->stats.decay_muzzy.nmadvise)); - arena_stats_accum_u64(&astats->decay_muzzy.purged, - arena_stats_read_u64(tsdn, &arena->stats, - &arena->stats.decay_muzzy.purged)); - - arena_stats_accum_zu(&astats->base, base_allocated); - arena_stats_accum_zu(&astats->internal, arena_internal_get(arena)); - arena_stats_accum_zu(&astats->metadata_thp, metadata_thp); - arena_stats_accum_zu(&astats->resident, base_resident + - (((atomic_load_zu(&arena->nactive, ATOMIC_RELAXED) + - extents_npages_get(&arena->extents_dirty) + - extents_npages_get(&arena->extents_muzzy)) << LG_PAGE))); - arena_stats_accum_zu(&astats->abandoned_vm, atomic_load_zu( - &arena->stats.abandoned_vm, ATOMIC_RELAXED)); + astats->base += base_allocated; + atomic_load_add_store_zu(&astats->internal, arena_internal_get(arena)); + astats->metadata_thp += metadata_thp; for (szind_t i = 0; i < SC_NSIZES - SC_NBINS; i++) { - uint64_t nmalloc = arena_stats_read_u64(tsdn, &arena->stats, + uint64_t nmalloc = locked_read_u64(tsdn, + LOCKEDINT_MTX(arena->stats.mtx), &arena->stats.lstats[i].nmalloc); - arena_stats_accum_u64(&lstats[i].nmalloc, nmalloc); - arena_stats_accum_u64(&astats->nmalloc_large, nmalloc); + locked_inc_u64_unsynchronized(&lstats[i].nmalloc, nmalloc); + astats->nmalloc_large += nmalloc; - uint64_t ndalloc = arena_stats_read_u64(tsdn, &arena->stats, + uint64_t ndalloc = locked_read_u64(tsdn, + LOCKEDINT_MTX(arena->stats.mtx), &arena->stats.lstats[i].ndalloc); - arena_stats_accum_u64(&lstats[i].ndalloc, ndalloc); - arena_stats_accum_u64(&astats->ndalloc_large, ndalloc); + locked_inc_u64_unsynchronized(&lstats[i].ndalloc, ndalloc); + astats->ndalloc_large += ndalloc; - uint64_t nrequests = arena_stats_read_u64(tsdn, &arena->stats, + uint64_t nrequests = locked_read_u64(tsdn, + LOCKEDINT_MTX(arena->stats.mtx), &arena->stats.lstats[i].nrequests); - arena_stats_accum_u64(&lstats[i].nrequests, - nmalloc + nrequests); - arena_stats_accum_u64(&astats->nrequests_large, + locked_inc_u64_unsynchronized(&lstats[i].nrequests, nmalloc + nrequests); + astats->nrequests_large += nmalloc + nrequests; /* nfill == nmalloc for large currently. */ - arena_stats_accum_u64(&lstats[i].nfills, nmalloc); - arena_stats_accum_u64(&astats->nfills_large, nmalloc); + locked_inc_u64_unsynchronized(&lstats[i].nfills, nmalloc); + astats->nfills_large += nmalloc; - uint64_t nflush = arena_stats_read_u64(tsdn, &arena->stats, + uint64_t nflush = locked_read_u64(tsdn, + LOCKEDINT_MTX(arena->stats.mtx), &arena->stats.lstats[i].nflushes); - arena_stats_accum_u64(&lstats[i].nflushes, nflush); - arena_stats_accum_u64(&astats->nflushes_large, nflush); + locked_inc_u64_unsynchronized(&lstats[i].nflushes, nflush); + astats->nflushes_large += nflush; assert(nmalloc >= ndalloc); assert(nmalloc - ndalloc <= SIZE_T_MAX); size_t curlextents = (size_t)(nmalloc - ndalloc); lstats[i].curlextents += curlextents; - arena_stats_accum_zu(&astats->allocated_large, - curlextents * sz_index2size(SC_NBINS + i)); - } - - for (pszind_t i = 0; i < SC_NPSIZES; i++) { - size_t dirty, muzzy, retained, dirty_bytes, muzzy_bytes, - retained_bytes; - dirty = extents_nextents_get(&arena->extents_dirty, i); - muzzy = extents_nextents_get(&arena->extents_muzzy, i); - retained = extents_nextents_get(&arena->extents_retained, i); - dirty_bytes = extents_nbytes_get(&arena->extents_dirty, i); - muzzy_bytes = extents_nbytes_get(&arena->extents_muzzy, i); - retained_bytes = - extents_nbytes_get(&arena->extents_retained, i); - - atomic_store_zu(&estats[i].ndirty, dirty, ATOMIC_RELAXED); - atomic_store_zu(&estats[i].nmuzzy, muzzy, ATOMIC_RELAXED); - atomic_store_zu(&estats[i].nretained, retained, ATOMIC_RELAXED); - atomic_store_zu(&estats[i].dirty_bytes, dirty_bytes, - ATOMIC_RELAXED); - atomic_store_zu(&estats[i].muzzy_bytes, muzzy_bytes, - ATOMIC_RELAXED); - atomic_store_zu(&estats[i].retained_bytes, retained_bytes, - ATOMIC_RELAXED); - } - - arena_stats_unlock(tsdn, &arena->stats); - - /* tcache_bytes counts currently cached bytes. */ - atomic_store_zu(&astats->tcache_bytes, 0, ATOMIC_RELAXED); + astats->allocated_large += + curlextents * sz_index2size(SC_NBINS + i); + } + + pa_shard_stats_merge(tsdn, &arena->pa_shard, &astats->pa_shard_stats, + estats, hpastats, secstats, &astats->resident); + + LOCKEDINT_MTX_UNLOCK(tsdn, arena->stats.mtx); + + /* Currently cached bytes and sanitizer-stashed bytes in tcache. */ + astats->tcache_bytes = 0; + astats->tcache_stashed_bytes = 0; malloc_mutex_lock(tsdn, &arena->tcache_ql_mtx); cache_bin_array_descriptor_t *descriptor; ql_foreach(descriptor, &arena->cache_bin_array_descriptor_ql, link) { - szind_t i = 0; - for (; i < SC_NBINS; i++) { - cache_bin_t *tbin = &descriptor->bins_small[i]; - arena_stats_accum_zu(&astats->tcache_bytes, - tbin->ncached * sz_index2size(i)); - } - for (; i < nhbins; i++) { - cache_bin_t *tbin = &descriptor->bins_large[i]; - arena_stats_accum_zu(&astats->tcache_bytes, - tbin->ncached * sz_index2size(i)); + for (szind_t i = 0; i < nhbins; i++) { + cache_bin_t *cache_bin = &descriptor->bins[i]; + cache_bin_sz_t ncached, nstashed; + cache_bin_nitems_get_remote(cache_bin, + &tcache_bin_info[i], &ncached, &nstashed); + + astats->tcache_bytes += ncached * sz_index2size(i); + astats->tcache_stashed_bytes += nstashed * + sz_index2size(i); } } malloc_mutex_prof_read(tsdn, @@ -224,21 +178,11 @@ /* Gather per arena mutex profiling data. */ READ_ARENA_MUTEX_PROF_DATA(large_mtx, arena_prof_mutex_large); - READ_ARENA_MUTEX_PROF_DATA(extent_avail_mtx, - arena_prof_mutex_extent_avail) - READ_ARENA_MUTEX_PROF_DATA(extents_dirty.mtx, - arena_prof_mutex_extents_dirty) - READ_ARENA_MUTEX_PROF_DATA(extents_muzzy.mtx, - arena_prof_mutex_extents_muzzy) - READ_ARENA_MUTEX_PROF_DATA(extents_retained.mtx, - arena_prof_mutex_extents_retained) - READ_ARENA_MUTEX_PROF_DATA(decay_dirty.mtx, - arena_prof_mutex_decay_dirty) - READ_ARENA_MUTEX_PROF_DATA(decay_muzzy.mtx, - arena_prof_mutex_decay_muzzy) READ_ARENA_MUTEX_PROF_DATA(base->mtx, - arena_prof_mutex_base) + arena_prof_mutex_base); #undef READ_ARENA_MUTEX_PROF_DATA + pa_shard_mtx_stats_read(tsdn, &arena->pa_shard, + astats->mutex_prof_data); nstime_copy(&astats->uptime, &arena->create_time); nstime_update(&astats->uptime); @@ -247,55 +191,67 @@ for (szind_t i = 0; i < SC_NBINS; i++) { for (unsigned j = 0; j < bin_infos[i].n_shards; j++) { bin_stats_merge(tsdn, &bstats[i], - &arena->bins[i].bin_shards[j]); + arena_get_bin(arena, i, j)); } } } -void -arena_extents_dirty_dalloc(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, extent_t *extent) { +static void +arena_background_thread_inactivity_check(tsdn_t *tsdn, arena_t *arena, + bool is_background_thread) { + if (!background_thread_enabled() || is_background_thread) { + return; + } + background_thread_info_t *info = + arena_background_thread_info_get(arena); + if (background_thread_indefinite_sleep(info)) { + arena_maybe_do_deferred_work(tsdn, arena, + &arena->pa_shard.pac.decay_dirty, 0); + } +} + +/* + * React to deferred work generated by a PAI function. + */ +void arena_handle_deferred_work(tsdn_t *tsdn, arena_t *arena) { witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), WITNESS_RANK_CORE, 0); - extents_dalloc(tsdn, arena, r_extent_hooks, &arena->extents_dirty, - extent); - if (arena_dirty_decay_ms_get(arena) == 0) { + if (decay_immediately(&arena->pa_shard.pac.decay_dirty)) { arena_decay_dirty(tsdn, arena, false, true); - } else { - arena_background_thread_inactivity_check(tsdn, arena, false); } + arena_background_thread_inactivity_check(tsdn, arena, false); } static void * -arena_slab_reg_alloc(extent_t *slab, const bin_info_t *bin_info) { +arena_slab_reg_alloc(edata_t *slab, const bin_info_t *bin_info) { void *ret; - arena_slab_data_t *slab_data = extent_slab_data_get(slab); + slab_data_t *slab_data = edata_slab_data_get(slab); size_t regind; - assert(extent_nfree_get(slab) > 0); + assert(edata_nfree_get(slab) > 0); assert(!bitmap_full(slab_data->bitmap, &bin_info->bitmap_info)); regind = bitmap_sfu(slab_data->bitmap, &bin_info->bitmap_info); - ret = (void *)((uintptr_t)extent_addr_get(slab) + + ret = (void *)((uintptr_t)edata_addr_get(slab) + (uintptr_t)(bin_info->reg_size * regind)); - extent_nfree_dec(slab); + edata_nfree_dec(slab); return ret; } static void -arena_slab_reg_alloc_batch(extent_t *slab, const bin_info_t *bin_info, +arena_slab_reg_alloc_batch(edata_t *slab, const bin_info_t *bin_info, unsigned cnt, void** ptrs) { - arena_slab_data_t *slab_data = extent_slab_data_get(slab); + slab_data_t *slab_data = edata_slab_data_get(slab); - assert(extent_nfree_get(slab) >= cnt); + assert(edata_nfree_get(slab) >= cnt); assert(!bitmap_full(slab_data->bitmap, &bin_info->bitmap_info)); #if (! defined JEMALLOC_INTERNAL_POPCOUNTL) || (defined BITMAP_USE_TREE) for (unsigned i = 0; i < cnt; i++) { size_t regind = bitmap_sfu(slab_data->bitmap, &bin_info->bitmap_info); - *(ptrs + i) = (void *)((uintptr_t)extent_addr_get(slab) + + *(ptrs + i) = (void *)((uintptr_t)edata_addr_get(slab) + (uintptr_t)(bin_info->reg_size * regind)); } #else @@ -316,7 +272,7 @@ * Load from memory locations only once, outside the * hot loop below. */ - uintptr_t base = (uintptr_t)extent_addr_get(slab); + uintptr_t base = (uintptr_t)edata_addr_get(slab); uintptr_t regsize = (uintptr_t)bin_info->reg_size; while (pop--) { size_t bit = cfs_lu(&g); @@ -328,56 +284,7 @@ slab_data->bitmap[group] = g; } #endif - extent_nfree_sub(slab, cnt); -} - -#ifndef JEMALLOC_JET -static -#endif -size_t -arena_slab_regind(extent_t *slab, szind_t binind, const void *ptr) { - size_t diff, regind; - - /* Freeing a pointer outside the slab can cause assertion failure. */ - assert((uintptr_t)ptr >= (uintptr_t)extent_addr_get(slab)); - assert((uintptr_t)ptr < (uintptr_t)extent_past_get(slab)); - /* Freeing an interior pointer can cause assertion failure. */ - assert(((uintptr_t)ptr - (uintptr_t)extent_addr_get(slab)) % - (uintptr_t)bin_infos[binind].reg_size == 0); - - diff = (size_t)((uintptr_t)ptr - (uintptr_t)extent_addr_get(slab)); - - /* Avoid doing division with a variable divisor. */ - regind = div_compute(&arena_binind_div_info[binind], diff); - - assert(regind < bin_infos[binind].nregs); - - return regind; -} - -static void -arena_slab_reg_dalloc(extent_t *slab, arena_slab_data_t *slab_data, void *ptr) { - szind_t binind = extent_szind_get(slab); - const bin_info_t *bin_info = &bin_infos[binind]; - size_t regind = arena_slab_regind(slab, binind, ptr); - - assert(extent_nfree_get(slab) < bin_info->nregs); - /* Freeing an unallocated pointer can cause assertion failure. */ - assert(bitmap_get(slab_data->bitmap, &bin_info->bitmap_info, regind)); - - bitmap_unset(slab_data->bitmap, &bin_info->bitmap_info, regind); - extent_nfree_inc(slab); -} - -static void -arena_nactive_add(arena_t *arena, size_t add_pages) { - atomic_fetch_add_zu(&arena->nactive, add_pages, ATOMIC_RELAXED); -} - -static void -arena_nactive_sub(arena_t *arena, size_t sub_pages) { - assert(atomic_load_zu(&arena->nactive, ATOMIC_RELAXED) >= sub_pages); - atomic_fetch_sub_zu(&arena->nactive, sub_pages, ATOMIC_RELAXED); + edata_nfree_sub(slab, cnt); } static void @@ -392,7 +299,7 @@ index = sz_size2index(usize); hindex = (index >= SC_NBINS) ? index - SC_NBINS : 0; - arena_stats_add_u64(tsdn, &arena->stats, + locked_inc_u64(tsdn, LOCKEDINT_MTX(arena->stats.mtx), &arena->stats.lstats[hindex].nmalloc, 1); } @@ -408,551 +315,118 @@ index = sz_size2index(usize); hindex = (index >= SC_NBINS) ? index - SC_NBINS : 0; - arena_stats_add_u64(tsdn, &arena->stats, + locked_inc_u64(tsdn, LOCKEDINT_MTX(arena->stats.mtx), &arena->stats.lstats[hindex].ndalloc, 1); } static void arena_large_ralloc_stats_update(tsdn_t *tsdn, arena_t *arena, size_t oldusize, size_t usize) { - arena_large_dalloc_stats_update(tsdn, arena, oldusize); arena_large_malloc_stats_update(tsdn, arena, usize); + arena_large_dalloc_stats_update(tsdn, arena, oldusize); } -static bool -arena_may_have_muzzy(arena_t *arena) { - return (pages_can_purge_lazy && (arena_muzzy_decay_ms_get(arena) != 0)); -} - -extent_t * +edata_t * arena_extent_alloc_large(tsdn_t *tsdn, arena_t *arena, size_t usize, - size_t alignment, bool *zero) { - extent_hooks_t *extent_hooks = EXTENT_HOOKS_INITIALIZER; + size_t alignment, bool zero) { + bool deferred_work_generated = false; + szind_t szind = sz_size2index(usize); + size_t esize = usize + sz_large_pad; - witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), - WITNESS_RANK_CORE, 0); + bool guarded = san_large_extent_decide_guard(tsdn, + arena_get_ehooks(arena), esize, alignment); + edata_t *edata = pa_alloc(tsdn, &arena->pa_shard, esize, alignment, + /* slab */ false, szind, zero, guarded, &deferred_work_generated); + assert(deferred_work_generated == false); - szind_t szind = sz_size2index(usize); - size_t mapped_add; - bool commit = true; - extent_t *extent = extents_alloc(tsdn, arena, &extent_hooks, - &arena->extents_dirty, NULL, usize, sz_large_pad, alignment, false, - szind, zero, &commit); - if (extent == NULL && arena_may_have_muzzy(arena)) { - extent = extents_alloc(tsdn, arena, &extent_hooks, - &arena->extents_muzzy, NULL, usize, sz_large_pad, alignment, - false, szind, zero, &commit); - } - size_t size = usize + sz_large_pad; - if (extent == NULL) { - extent = extent_alloc_wrapper(tsdn, arena, &extent_hooks, NULL, - usize, sz_large_pad, alignment, false, szind, zero, - &commit); + if (edata != NULL) { if (config_stats) { - /* - * extent may be NULL on OOM, but in that case - * mapped_add isn't used below, so there's no need to - * conditionlly set it to 0 here. - */ - mapped_add = size; + LOCKEDINT_MTX_LOCK(tsdn, arena->stats.mtx); + arena_large_malloc_stats_update(tsdn, arena, usize); + LOCKEDINT_MTX_UNLOCK(tsdn, arena->stats.mtx); } - } else if (config_stats) { - mapped_add = 0; } - if (extent != NULL) { - if (config_stats) { - arena_stats_lock(tsdn, &arena->stats); - arena_large_malloc_stats_update(tsdn, arena, usize); - if (mapped_add != 0) { - arena_stats_add_zu(tsdn, &arena->stats, - &arena->stats.mapped, mapped_add); - } - arena_stats_unlock(tsdn, &arena->stats); - } - arena_nactive_add(arena, size >> LG_PAGE); + if (edata != NULL && sz_large_pad != 0) { + arena_cache_oblivious_randomize(tsdn, arena, edata, alignment); } - return extent; + return edata; } void -arena_extent_dalloc_large_prep(tsdn_t *tsdn, arena_t *arena, extent_t *extent) { +arena_extent_dalloc_large_prep(tsdn_t *tsdn, arena_t *arena, edata_t *edata) { if (config_stats) { - arena_stats_lock(tsdn, &arena->stats); + LOCKEDINT_MTX_LOCK(tsdn, arena->stats.mtx); arena_large_dalloc_stats_update(tsdn, arena, - extent_usize_get(extent)); - arena_stats_unlock(tsdn, &arena->stats); + edata_usize_get(edata)); + LOCKEDINT_MTX_UNLOCK(tsdn, arena->stats.mtx); } - arena_nactive_sub(arena, extent_size_get(extent) >> LG_PAGE); } void -arena_extent_ralloc_large_shrink(tsdn_t *tsdn, arena_t *arena, extent_t *extent, +arena_extent_ralloc_large_shrink(tsdn_t *tsdn, arena_t *arena, edata_t *edata, size_t oldusize) { - size_t usize = extent_usize_get(extent); - size_t udiff = oldusize - usize; + size_t usize = edata_usize_get(edata); if (config_stats) { - arena_stats_lock(tsdn, &arena->stats); + LOCKEDINT_MTX_LOCK(tsdn, arena->stats.mtx); arena_large_ralloc_stats_update(tsdn, arena, oldusize, usize); - arena_stats_unlock(tsdn, &arena->stats); + LOCKEDINT_MTX_UNLOCK(tsdn, arena->stats.mtx); } - arena_nactive_sub(arena, udiff >> LG_PAGE); } void -arena_extent_ralloc_large_expand(tsdn_t *tsdn, arena_t *arena, extent_t *extent, +arena_extent_ralloc_large_expand(tsdn_t *tsdn, arena_t *arena, edata_t *edata, size_t oldusize) { - size_t usize = extent_usize_get(extent); - size_t udiff = usize - oldusize; + size_t usize = edata_usize_get(edata); if (config_stats) { - arena_stats_lock(tsdn, &arena->stats); + LOCKEDINT_MTX_LOCK(tsdn, arena->stats.mtx); arena_large_ralloc_stats_update(tsdn, arena, oldusize, usize); - arena_stats_unlock(tsdn, &arena->stats); - } - arena_nactive_add(arena, udiff >> LG_PAGE); -} - -static ssize_t -arena_decay_ms_read(arena_decay_t *decay) { - return atomic_load_zd(&decay->time_ms, ATOMIC_RELAXED); -} - -static void -arena_decay_ms_write(arena_decay_t *decay, ssize_t decay_ms) { - atomic_store_zd(&decay->time_ms, decay_ms, ATOMIC_RELAXED); -} - -static void -arena_decay_deadline_init(arena_decay_t *decay) { - /* - * Generate a new deadline that is uniformly random within the next - * epoch after the current one. - */ - nstime_copy(&decay->deadline, &decay->epoch); - nstime_add(&decay->deadline, &decay->interval); - if (arena_decay_ms_read(decay) > 0) { - nstime_t jitter; - - nstime_init(&jitter, prng_range_u64(&decay->jitter_state, - nstime_ns(&decay->interval))); - nstime_add(&decay->deadline, &jitter); - } -} - -static bool -arena_decay_deadline_reached(const arena_decay_t *decay, const nstime_t *time) { - return (nstime_compare(&decay->deadline, time) <= 0); -} - -static size_t -arena_decay_backlog_npages_limit(const arena_decay_t *decay) { - uint64_t sum; - size_t npages_limit_backlog; - unsigned i; - - /* - * For each element of decay_backlog, multiply by the corresponding - * fixed-point smoothstep decay factor. Sum the products, then divide - * to round down to the nearest whole number of pages. - */ - sum = 0; - for (i = 0; i < SMOOTHSTEP_NSTEPS; i++) { - sum += decay->backlog[i] * h_steps[i]; - } - npages_limit_backlog = (size_t)(sum >> SMOOTHSTEP_BFP); - - return npages_limit_backlog; -} - -static void -arena_decay_backlog_update_last(arena_decay_t *decay, size_t current_npages) { - size_t npages_delta = (current_npages > decay->nunpurged) ? - current_npages - decay->nunpurged : 0; - decay->backlog[SMOOTHSTEP_NSTEPS-1] = npages_delta; - - if (config_debug) { - if (current_npages > decay->ceil_npages) { - decay->ceil_npages = current_npages; - } - size_t npages_limit = arena_decay_backlog_npages_limit(decay); - assert(decay->ceil_npages >= npages_limit); - if (decay->ceil_npages > npages_limit) { - decay->ceil_npages = npages_limit; - } - } -} - -static void -arena_decay_backlog_update(arena_decay_t *decay, uint64_t nadvance_u64, - size_t current_npages) { - if (nadvance_u64 >= SMOOTHSTEP_NSTEPS) { - memset(decay->backlog, 0, (SMOOTHSTEP_NSTEPS-1) * - sizeof(size_t)); - } else { - size_t nadvance_z = (size_t)nadvance_u64; - - assert((uint64_t)nadvance_z == nadvance_u64); - - memmove(decay->backlog, &decay->backlog[nadvance_z], - (SMOOTHSTEP_NSTEPS - nadvance_z) * sizeof(size_t)); - if (nadvance_z > 1) { - memset(&decay->backlog[SMOOTHSTEP_NSTEPS - - nadvance_z], 0, (nadvance_z-1) * sizeof(size_t)); - } - } - - arena_decay_backlog_update_last(decay, current_npages); -} - -static void -arena_decay_try_purge(tsdn_t *tsdn, arena_t *arena, arena_decay_t *decay, - extents_t *extents, size_t current_npages, size_t npages_limit, - bool is_background_thread) { - if (current_npages > npages_limit) { - arena_decay_to_limit(tsdn, arena, decay, extents, false, - npages_limit, current_npages - npages_limit, - is_background_thread); - } -} - -static void -arena_decay_epoch_advance_helper(arena_decay_t *decay, const nstime_t *time, - size_t current_npages) { - assert(arena_decay_deadline_reached(decay, time)); - - nstime_t delta; - nstime_copy(&delta, time); - nstime_subtract(&delta, &decay->epoch); - - uint64_t nadvance_u64 = nstime_divide(&delta, &decay->interval); - assert(nadvance_u64 > 0); - - /* Add nadvance_u64 decay intervals to epoch. */ - nstime_copy(&delta, &decay->interval); - nstime_imultiply(&delta, nadvance_u64); - nstime_add(&decay->epoch, &delta); - - /* Set a new deadline. */ - arena_decay_deadline_init(decay); - - /* Update the backlog. */ - arena_decay_backlog_update(decay, nadvance_u64, current_npages); -} - -static void -arena_decay_epoch_advance(tsdn_t *tsdn, arena_t *arena, arena_decay_t *decay, - extents_t *extents, const nstime_t *time, bool is_background_thread) { - size_t current_npages = extents_npages_get(extents); - arena_decay_epoch_advance_helper(decay, time, current_npages); - - size_t npages_limit = arena_decay_backlog_npages_limit(decay); - /* We may unlock decay->mtx when try_purge(). Finish logging first. */ - decay->nunpurged = (npages_limit > current_npages) ? npages_limit : - current_npages; - - if (!background_thread_enabled() || is_background_thread) { - arena_decay_try_purge(tsdn, arena, decay, extents, - current_npages, npages_limit, is_background_thread); - } -} - -static void -arena_decay_reinit(arena_decay_t *decay, ssize_t decay_ms) { - arena_decay_ms_write(decay, decay_ms); - if (decay_ms > 0) { - nstime_init(&decay->interval, (uint64_t)decay_ms * - KQU(1000000)); - nstime_idivide(&decay->interval, SMOOTHSTEP_NSTEPS); - } - - nstime_init(&decay->epoch, 0); - nstime_update(&decay->epoch); - decay->jitter_state = (uint64_t)(uintptr_t)decay; - arena_decay_deadline_init(decay); - decay->nunpurged = 0; - memset(decay->backlog, 0, SMOOTHSTEP_NSTEPS * sizeof(size_t)); -} - -static bool -arena_decay_init(arena_decay_t *decay, ssize_t decay_ms, - arena_stats_decay_t *stats) { - if (config_debug) { - for (size_t i = 0; i < sizeof(arena_decay_t); i++) { - assert(((char *)decay)[i] == 0); - } - decay->ceil_npages = 0; - } - if (malloc_mutex_init(&decay->mtx, "decay", WITNESS_RANK_DECAY, - malloc_mutex_rank_exclusive)) { - return true; - } - decay->purging = false; - arena_decay_reinit(decay, decay_ms); - /* Memory is zeroed, so there is no need to clear stats. */ - if (config_stats) { - decay->stats = stats; - } - return false; -} - -static bool -arena_decay_ms_valid(ssize_t decay_ms) { - if (decay_ms < -1) { - return false; - } - if (decay_ms == -1 || (uint64_t)decay_ms <= NSTIME_SEC_MAX * - KQU(1000)) { - return true; + LOCKEDINT_MTX_UNLOCK(tsdn, arena->stats.mtx); } - return false; } -static bool -arena_maybe_decay(tsdn_t *tsdn, arena_t *arena, arena_decay_t *decay, - extents_t *extents, bool is_background_thread) { - malloc_mutex_assert_owner(tsdn, &decay->mtx); - - /* Purge all or nothing if the option is disabled. */ - ssize_t decay_ms = arena_decay_ms_read(decay); - if (decay_ms <= 0) { - if (decay_ms == 0) { - arena_decay_to_limit(tsdn, arena, decay, extents, false, - 0, extents_npages_get(extents), - is_background_thread); - } - return false; - } - - nstime_t time; - nstime_init(&time, 0); - nstime_update(&time); - if (unlikely(!nstime_monotonic() && nstime_compare(&decay->epoch, &time) - > 0)) { - /* - * Time went backwards. Move the epoch back in time and - * generate a new deadline, with the expectation that time - * typically flows forward for long enough periods of time that - * epochs complete. Unfortunately, this strategy is susceptible - * to clock jitter triggering premature epoch advances, but - * clock jitter estimation and compensation isn't feasible here - * because calls into this code are event-driven. - */ - nstime_copy(&decay->epoch, &time); - arena_decay_deadline_init(decay); +/* + * In situations where we're not forcing a decay (i.e. because the user + * specifically requested it), should we purge ourselves, or wait for the + * background thread to get to it. + */ +static pac_purge_eagerness_t +arena_decide_unforced_purge_eagerness(bool is_background_thread) { + if (is_background_thread) { + return PAC_PURGE_ALWAYS; + } else if (!is_background_thread && background_thread_enabled()) { + return PAC_PURGE_NEVER; } else { - /* Verify that time does not go backwards. */ - assert(nstime_compare(&decay->epoch, &time) <= 0); + return PAC_PURGE_ON_EPOCH_ADVANCE; } - - /* - * If the deadline has been reached, advance to the current epoch and - * purge to the new limit if necessary. Note that dirty pages created - * during the current epoch are not subject to purge until a future - * epoch, so as a result purging only happens during epoch advances, or - * being triggered by background threads (scheduled event). - */ - bool advance_epoch = arena_decay_deadline_reached(decay, &time); - if (advance_epoch) { - arena_decay_epoch_advance(tsdn, arena, decay, extents, &time, - is_background_thread); - } else if (is_background_thread) { - arena_decay_try_purge(tsdn, arena, decay, extents, - extents_npages_get(extents), - arena_decay_backlog_npages_limit(decay), - is_background_thread); - } - - return advance_epoch; -} - -static ssize_t -arena_decay_ms_get(arena_decay_t *decay) { - return arena_decay_ms_read(decay); -} - -ssize_t -arena_dirty_decay_ms_get(arena_t *arena) { - return arena_decay_ms_get(&arena->decay_dirty); -} - -ssize_t -arena_muzzy_decay_ms_get(arena_t *arena) { - return arena_decay_ms_get(&arena->decay_muzzy); -} - -static bool -arena_decay_ms_set(tsdn_t *tsdn, arena_t *arena, arena_decay_t *decay, - extents_t *extents, ssize_t decay_ms) { - if (!arena_decay_ms_valid(decay_ms)) { - return true; - } - - malloc_mutex_lock(tsdn, &decay->mtx); - /* - * Restart decay backlog from scratch, which may cause many dirty pages - * to be immediately purged. It would conceptually be possible to map - * the old backlog onto the new backlog, but there is no justification - * for such complexity since decay_ms changes are intended to be - * infrequent, either between the {-1, 0, >0} states, or a one-time - * arbitrary change during initial arena configuration. - */ - arena_decay_reinit(decay, decay_ms); - arena_maybe_decay(tsdn, arena, decay, extents, false); - malloc_mutex_unlock(tsdn, &decay->mtx); - - return false; } bool -arena_dirty_decay_ms_set(tsdn_t *tsdn, arena_t *arena, +arena_decay_ms_set(tsdn_t *tsdn, arena_t *arena, extent_state_t state, ssize_t decay_ms) { - return arena_decay_ms_set(tsdn, arena, &arena->decay_dirty, - &arena->extents_dirty, decay_ms); -} - -bool -arena_muzzy_decay_ms_set(tsdn_t *tsdn, arena_t *arena, - ssize_t decay_ms) { - return arena_decay_ms_set(tsdn, arena, &arena->decay_muzzy, - &arena->extents_muzzy, decay_ms); -} - -static size_t -arena_stash_decayed(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, extents_t *extents, size_t npages_limit, - size_t npages_decay_max, extent_list_t *decay_extents) { - witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), - WITNESS_RANK_CORE, 0); - - /* Stash extents according to npages_limit. */ - size_t nstashed = 0; - extent_t *extent; - while (nstashed < npages_decay_max && - (extent = extents_evict(tsdn, arena, r_extent_hooks, extents, - npages_limit)) != NULL) { - extent_list_append(decay_extents, extent); - nstashed += extent_size_get(extent) >> LG_PAGE; - } - return nstashed; + pac_purge_eagerness_t eagerness = arena_decide_unforced_purge_eagerness( + /* is_background_thread */ false); + return pa_decay_ms_set(tsdn, &arena->pa_shard, state, decay_ms, + eagerness); } -static size_t -arena_decay_stashed(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, arena_decay_t *decay, extents_t *extents, - bool all, extent_list_t *decay_extents, bool is_background_thread) { - size_t nmadvise, nunmapped; - size_t npurged; - - if (config_stats) { - nmadvise = 0; - nunmapped = 0; - } - npurged = 0; - - ssize_t muzzy_decay_ms = arena_muzzy_decay_ms_get(arena); - for (extent_t *extent = extent_list_first(decay_extents); extent != - NULL; extent = extent_list_first(decay_extents)) { - if (config_stats) { - nmadvise++; - } - size_t npages = extent_size_get(extent) >> LG_PAGE; - npurged += npages; - extent_list_remove(decay_extents, extent); - switch (extents_state_get(extents)) { - case extent_state_active: - not_reached(); - case extent_state_dirty: - if (!all && muzzy_decay_ms != 0 && - !extent_purge_lazy_wrapper(tsdn, arena, - r_extent_hooks, extent, 0, - extent_size_get(extent))) { - extents_dalloc(tsdn, arena, r_extent_hooks, - &arena->extents_muzzy, extent); - arena_background_thread_inactivity_check(tsdn, - arena, is_background_thread); - break; - } - /* Fall through. */ - case extent_state_muzzy: - extent_dalloc_wrapper(tsdn, arena, r_extent_hooks, - extent); - if (config_stats) { - nunmapped += npages; - } - break; - case extent_state_retained: - default: - not_reached(); - } - } - - if (config_stats) { - arena_stats_lock(tsdn, &arena->stats); - arena_stats_add_u64(tsdn, &arena->stats, &decay->stats->npurge, - 1); - arena_stats_add_u64(tsdn, &arena->stats, - &decay->stats->nmadvise, nmadvise); - arena_stats_add_u64(tsdn, &arena->stats, &decay->stats->purged, - npurged); - arena_stats_sub_zu(tsdn, &arena->stats, &arena->stats.mapped, - nunmapped << LG_PAGE); - arena_stats_unlock(tsdn, &arena->stats); - } - - return npurged; -} - -/* - * npages_limit: Decay at most npages_decay_max pages without violating the - * invariant: (extents_npages_get(extents) >= npages_limit). We need an upper - * bound on number of pages in order to prevent unbounded growth (namely in - * stashed), otherwise unbounded new pages could be added to extents during the - * current decay run, so that the purging thread never finishes. - */ -static void -arena_decay_to_limit(tsdn_t *tsdn, arena_t *arena, arena_decay_t *decay, - extents_t *extents, bool all, size_t npages_limit, size_t npages_decay_max, - bool is_background_thread) { - witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), - WITNESS_RANK_CORE, 1); - malloc_mutex_assert_owner(tsdn, &decay->mtx); - - if (decay->purging) { - return; - } - decay->purging = true; - malloc_mutex_unlock(tsdn, &decay->mtx); - - extent_hooks_t *extent_hooks = extent_hooks_get(arena); - - extent_list_t decay_extents; - extent_list_init(&decay_extents); - - size_t npurge = arena_stash_decayed(tsdn, arena, &extent_hooks, extents, - npages_limit, npages_decay_max, &decay_extents); - if (npurge != 0) { - size_t npurged = arena_decay_stashed(tsdn, arena, - &extent_hooks, decay, extents, all, &decay_extents, - is_background_thread); - assert(npurged == npurge); - } - - malloc_mutex_lock(tsdn, &decay->mtx); - decay->purging = false; +ssize_t +arena_decay_ms_get(arena_t *arena, extent_state_t state) { + return pa_decay_ms_get(&arena->pa_shard, state); } static bool -arena_decay_impl(tsdn_t *tsdn, arena_t *arena, arena_decay_t *decay, - extents_t *extents, bool is_background_thread, bool all) { +arena_decay_impl(tsdn_t *tsdn, arena_t *arena, decay_t *decay, + pac_decay_stats_t *decay_stats, ecache_t *ecache, + bool is_background_thread, bool all) { if (all) { malloc_mutex_lock(tsdn, &decay->mtx); - arena_decay_to_limit(tsdn, arena, decay, extents, all, 0, - extents_npages_get(extents), is_background_thread); + pac_decay_all(tsdn, &arena->pa_shard.pac, decay, decay_stats, + ecache, /* fully_decay */ all); malloc_mutex_unlock(tsdn, &decay->mtx); - return false; } @@ -960,20 +434,20 @@ /* No need to wait if another thread is in progress. */ return true; } - - bool epoch_advanced = arena_maybe_decay(tsdn, arena, decay, extents, - is_background_thread); + pac_purge_eagerness_t eagerness = + arena_decide_unforced_purge_eagerness(is_background_thread); + bool epoch_advanced = pac_maybe_decay_purge(tsdn, &arena->pa_shard.pac, + decay, decay_stats, ecache, eagerness); size_t npages_new; if (epoch_advanced) { /* Backlog is updated on epoch advance. */ - npages_new = decay->backlog[SMOOTHSTEP_NSTEPS-1]; + npages_new = decay_epoch_npages_delta(decay); } malloc_mutex_unlock(tsdn, &decay->mtx); if (have_background_thread && background_thread_enabled() && epoch_advanced && !is_background_thread) { - background_thread_interval_check(tsdn, arena, decay, - npages_new); + arena_maybe_do_deferred_work(tsdn, arena, decay, npages_new); } return false; @@ -982,53 +456,143 @@ static bool arena_decay_dirty(tsdn_t *tsdn, arena_t *arena, bool is_background_thread, bool all) { - return arena_decay_impl(tsdn, arena, &arena->decay_dirty, - &arena->extents_dirty, is_background_thread, all); + return arena_decay_impl(tsdn, arena, &arena->pa_shard.pac.decay_dirty, + &arena->pa_shard.pac.stats->decay_dirty, + &arena->pa_shard.pac.ecache_dirty, is_background_thread, all); } static bool arena_decay_muzzy(tsdn_t *tsdn, arena_t *arena, bool is_background_thread, bool all) { - return arena_decay_impl(tsdn, arena, &arena->decay_muzzy, - &arena->extents_muzzy, is_background_thread, all); + if (pa_shard_dont_decay_muzzy(&arena->pa_shard)) { + return false; + } + return arena_decay_impl(tsdn, arena, &arena->pa_shard.pac.decay_muzzy, + &arena->pa_shard.pac.stats->decay_muzzy, + &arena->pa_shard.pac.ecache_muzzy, is_background_thread, all); } void arena_decay(tsdn_t *tsdn, arena_t *arena, bool is_background_thread, bool all) { + if (all) { + /* + * We should take a purge of "all" to mean "save as much memory + * as possible", including flushing any caches (for situations + * like thread death, or manual purge calls). + */ + sec_flush(tsdn, &arena->pa_shard.hpa_sec); + } if (arena_decay_dirty(tsdn, arena, is_background_thread, all)) { return; } arena_decay_muzzy(tsdn, arena, is_background_thread, all); } +static bool +arena_should_decay_early(tsdn_t *tsdn, arena_t *arena, decay_t *decay, + background_thread_info_t *info, nstime_t *remaining_sleep, + size_t npages_new) { + malloc_mutex_assert_owner(tsdn, &info->mtx); + + if (malloc_mutex_trylock(tsdn, &decay->mtx)) { + return false; + } + + if (!decay_gradually(decay)) { + malloc_mutex_unlock(tsdn, &decay->mtx); + return false; + } + + nstime_init(remaining_sleep, background_thread_wakeup_time_get(info)); + if (nstime_compare(remaining_sleep, &decay->epoch) <= 0) { + malloc_mutex_unlock(tsdn, &decay->mtx); + return false; + } + nstime_subtract(remaining_sleep, &decay->epoch); + if (npages_new > 0) { + uint64_t npurge_new = decay_npages_purge_in(decay, + remaining_sleep, npages_new); + info->npages_to_purge_new += npurge_new; + } + malloc_mutex_unlock(tsdn, &decay->mtx); + return info->npages_to_purge_new > + ARENA_DEFERRED_PURGE_NPAGES_THRESHOLD; +} + +/* + * Check if deferred work needs to be done sooner than planned. + * For decay we might want to wake up earlier because of an influx of dirty + * pages. Rather than waiting for previously estimated time, we proactively + * purge those pages. + * If background thread sleeps indefinitely, always wake up because some + * deferred work has been generated. + */ static void -arena_slab_dalloc(tsdn_t *tsdn, arena_t *arena, extent_t *slab) { - arena_nactive_sub(arena, extent_size_get(slab) >> LG_PAGE); +arena_maybe_do_deferred_work(tsdn_t *tsdn, arena_t *arena, decay_t *decay, + size_t npages_new) { + background_thread_info_t *info = arena_background_thread_info_get( + arena); + if (malloc_mutex_trylock(tsdn, &info->mtx)) { + /* + * Background thread may hold the mutex for a long period of + * time. We'd like to avoid the variance on application + * threads. So keep this non-blocking, and leave the work to a + * future epoch. + */ + return; + } + if (!background_thread_is_started(info)) { + goto label_done; + } + + nstime_t remaining_sleep; + if (background_thread_indefinite_sleep(info)) { + background_thread_wakeup_early(info, NULL); + } else if (arena_should_decay_early(tsdn, arena, decay, info, + &remaining_sleep, npages_new)) { + info->npages_to_purge_new = 0; + background_thread_wakeup_early(info, &remaining_sleep); + } +label_done: + malloc_mutex_unlock(tsdn, &info->mtx); +} - extent_hooks_t *extent_hooks = EXTENT_HOOKS_INITIALIZER; - arena_extents_dirty_dalloc(tsdn, arena, &extent_hooks, slab); +/* Called from background threads. */ +void +arena_do_deferred_work(tsdn_t *tsdn, arena_t *arena) { + arena_decay(tsdn, arena, true, false); + pa_shard_do_deferred_work(tsdn, &arena->pa_shard); +} + +void +arena_slab_dalloc(tsdn_t *tsdn, arena_t *arena, edata_t *slab) { + bool deferred_work_generated = false; + pa_dalloc(tsdn, &arena->pa_shard, slab, &deferred_work_generated); + if (deferred_work_generated) { + arena_handle_deferred_work(tsdn, arena); + } } static void -arena_bin_slabs_nonfull_insert(bin_t *bin, extent_t *slab) { - assert(extent_nfree_get(slab) > 0); - extent_heap_insert(&bin->slabs_nonfull, slab); +arena_bin_slabs_nonfull_insert(bin_t *bin, edata_t *slab) { + assert(edata_nfree_get(slab) > 0); + edata_heap_insert(&bin->slabs_nonfull, slab); if (config_stats) { bin->stats.nonfull_slabs++; } } static void -arena_bin_slabs_nonfull_remove(bin_t *bin, extent_t *slab) { - extent_heap_remove(&bin->slabs_nonfull, slab); +arena_bin_slabs_nonfull_remove(bin_t *bin, edata_t *slab) { + edata_heap_remove(&bin->slabs_nonfull, slab); if (config_stats) { bin->stats.nonfull_slabs--; } } -static extent_t * +static edata_t * arena_bin_slabs_nonfull_tryget(bin_t *bin) { - extent_t *slab = extent_heap_remove_first(&bin->slabs_nonfull); + edata_t *slab = edata_heap_remove_first(&bin->slabs_nonfull); if (slab == NULL) { return NULL; } @@ -1040,30 +604,30 @@ } static void -arena_bin_slabs_full_insert(arena_t *arena, bin_t *bin, extent_t *slab) { - assert(extent_nfree_get(slab) == 0); +arena_bin_slabs_full_insert(arena_t *arena, bin_t *bin, edata_t *slab) { + assert(edata_nfree_get(slab) == 0); /* * Tracking extents is required by arena_reset, which is not allowed - * for auto arenas. Bypass this step to avoid touching the extent + * for auto arenas. Bypass this step to avoid touching the edata * linkage (often results in cache misses) for auto arenas. */ if (arena_is_auto(arena)) { return; } - extent_list_append(&bin->slabs_full, slab); + edata_list_active_append(&bin->slabs_full, slab); } static void -arena_bin_slabs_full_remove(arena_t *arena, bin_t *bin, extent_t *slab) { +arena_bin_slabs_full_remove(arena_t *arena, bin_t *bin, edata_t *slab) { if (arena_is_auto(arena)) { return; } - extent_list_remove(&bin->slabs_full, slab); + edata_list_active_remove(&bin->slabs_full, slab); } static void arena_bin_reset(tsd_t *tsd, arena_t *arena, bin_t *bin) { - extent_t *slab; + edata_t *slab; malloc_mutex_lock(tsd_tsdn(tsd), &bin->lock); if (bin->slabcur != NULL) { @@ -1073,13 +637,13 @@ arena_slab_dalloc(tsd_tsdn(tsd), arena, slab); malloc_mutex_lock(tsd_tsdn(tsd), &bin->lock); } - while ((slab = extent_heap_remove_first(&bin->slabs_nonfull)) != NULL) { + while ((slab = edata_heap_remove_first(&bin->slabs_nonfull)) != NULL) { malloc_mutex_unlock(tsd_tsdn(tsd), &bin->lock); arena_slab_dalloc(tsd_tsdn(tsd), arena, slab); malloc_mutex_lock(tsd_tsdn(tsd), &bin->lock); } - for (slab = extent_list_first(&bin->slabs_full); slab != NULL; - slab = extent_list_first(&bin->slabs_full)) { + for (slab = edata_list_active_first(&bin->slabs_full); slab != NULL; + slab = edata_list_active_first(&bin->slabs_full)) { arena_bin_slabs_full_remove(arena, bin, slab); malloc_mutex_unlock(tsd_tsdn(tsd), &bin->lock); arena_slab_dalloc(tsd_tsdn(tsd), arena, slab); @@ -1111,16 +675,15 @@ /* Large allocations. */ malloc_mutex_lock(tsd_tsdn(tsd), &arena->large_mtx); - for (extent_t *extent = extent_list_first(&arena->large); extent != - NULL; extent = extent_list_first(&arena->large)) { - void *ptr = extent_base_get(extent); + for (edata_t *edata = edata_list_active_first(&arena->large); + edata != NULL; edata = edata_list_active_first(&arena->large)) { + void *ptr = edata_base_get(edata); size_t usize; malloc_mutex_unlock(tsd_tsdn(tsd), &arena->large_mtx); - alloc_ctx_t alloc_ctx; - rtree_ctx_t *rtree_ctx = tsd_rtree_ctx(tsd); - rtree_szind_slab_read(tsd_tsdn(tsd), &extents_rtree, rtree_ctx, - (uintptr_t)ptr, true, &alloc_ctx.szind, &alloc_ctx.slab); + emap_alloc_ctx_t alloc_ctx; + emap_alloc_ctx_lookup(tsd_tsdn(tsd), &arena_emap_global, ptr, + &alloc_ctx); assert(alloc_ctx.szind != SC_NSIZES); if (config_stats || (config_prof && opt_prof)) { @@ -1131,7 +694,7 @@ if (config_prof && opt_prof) { prof_free(tsd, ptr, usize, &alloc_ctx); } - large_dalloc(tsd_tsdn(tsd), extent); + large_dalloc(tsd_tsdn(tsd), edata); malloc_mutex_lock(tsd_tsdn(tsd), &arena->large_mtx); } malloc_mutex_unlock(tsd_tsdn(tsd), &arena->large_mtx); @@ -1139,32 +702,95 @@ /* Bins. */ for (unsigned i = 0; i < SC_NBINS; i++) { for (unsigned j = 0; j < bin_infos[i].n_shards; j++) { - arena_bin_reset(tsd, arena, - &arena->bins[i].bin_shards[j]); + arena_bin_reset(tsd, arena, arena_get_bin(arena, i, j)); } } + pa_shard_reset(tsd_tsdn(tsd), &arena->pa_shard); +} - atomic_store_zu(&arena->nactive, 0, ATOMIC_RELAXED); +static void +arena_prepare_base_deletion_sync_finish(tsd_t *tsd, malloc_mutex_t **mutexes, + unsigned n_mtx) { + for (unsigned i = 0; i < n_mtx; i++) { + malloc_mutex_lock(tsd_tsdn(tsd), mutexes[i]); + malloc_mutex_unlock(tsd_tsdn(tsd), mutexes[i]); + } } +#define ARENA_DESTROY_MAX_DELAYED_MTX 32 static void -arena_destroy_retained(tsdn_t *tsdn, arena_t *arena) { +arena_prepare_base_deletion_sync(tsd_t *tsd, malloc_mutex_t *mtx, + malloc_mutex_t **delayed_mtx, unsigned *n_delayed) { + if (!malloc_mutex_trylock(tsd_tsdn(tsd), mtx)) { + /* No contention. */ + malloc_mutex_unlock(tsd_tsdn(tsd), mtx); + return; + } + unsigned n = *n_delayed; + assert(n < ARENA_DESTROY_MAX_DELAYED_MTX); + /* Add another to the batch. */ + delayed_mtx[n++] = mtx; + + if (n == ARENA_DESTROY_MAX_DELAYED_MTX) { + arena_prepare_base_deletion_sync_finish(tsd, delayed_mtx, n); + n = 0; + } + *n_delayed = n; +} + +static void +arena_prepare_base_deletion(tsd_t *tsd, base_t *base_to_destroy) { /* - * Iterate over the retained extents and destroy them. This gives the - * extent allocator underlying the extent hooks an opportunity to unmap - * all retained memory without having to keep its own metadata - * structures. In practice, virtual memory for dss-allocated extents is - * leaked here, so best practice is to avoid dss for arenas to be - * destroyed, or provide custom extent hooks that track retained - * dss-based extents for later reuse. + * In order to coalesce, emap_try_acquire_edata_neighbor will attempt to + * check neighbor edata's state to determine eligibility. This means + * under certain conditions, the metadata from an arena can be accessed + * w/o holding any locks from that arena. In order to guarantee safe + * memory access, the metadata and the underlying base allocator needs + * to be kept alive, until all pending accesses are done. + * + * 1) with opt_retain, the arena boundary implies the is_head state + * (tracked in the rtree leaf), and the coalesce flow will stop at the + * head state branch. Therefore no cross arena metadata access + * possible. + * + * 2) w/o opt_retain, the arena id needs to be read from the edata_t, + * meaning read only cross-arena metadata access is possible. The + * coalesce attempt will stop at the arena_id mismatch, and is always + * under one of the ecache locks. To allow safe passthrough of such + * metadata accesses, the loop below will iterate through all manual + * arenas' ecache locks. As all the metadata from this base allocator + * have been unlinked from the rtree, after going through all the + * relevant ecache locks, it's safe to say that a) pending accesses are + * all finished, and b) no new access will be generated. */ - extent_hooks_t *extent_hooks = extent_hooks_get(arena); - extent_t *extent; - while ((extent = extents_evict(tsdn, arena, &extent_hooks, - &arena->extents_retained, 0)) != NULL) { - extent_destroy_wrapper(tsdn, arena, &extent_hooks, extent); + if (opt_retain) { + return; + } + unsigned destroy_ind = base_ind_get(base_to_destroy); + assert(destroy_ind >= manual_arena_base); + + tsdn_t *tsdn = tsd_tsdn(tsd); + malloc_mutex_t *delayed_mtx[ARENA_DESTROY_MAX_DELAYED_MTX]; + unsigned n_delayed = 0, total = narenas_total_get(); + for (unsigned i = 0; i < total; i++) { + if (i == destroy_ind) { + continue; + } + arena_t *arena = arena_get(tsdn, i, false); + if (arena == NULL) { + continue; + } + pac_t *pac = &arena->pa_shard.pac; + arena_prepare_base_deletion_sync(tsd, &pac->ecache_dirty.mtx, + delayed_mtx, &n_delayed); + arena_prepare_base_deletion_sync(tsd, &pac->ecache_muzzy.mtx, + delayed_mtx, &n_delayed); + arena_prepare_base_deletion_sync(tsd, &pac->ecache_retained.mtx, + delayed_mtx, &n_delayed); } + arena_prepare_base_deletion_sync_finish(tsd, delayed_mtx, n_delayed); } +#undef ARENA_DESTROY_MAX_DELAYED_MTX void arena_destroy(tsd_t *tsd, arena_t *arena) { @@ -1175,13 +801,10 @@ /* * No allocations have occurred since arena_reset() was called. * Furthermore, the caller (arena_i_destroy_ctl()) purged all cached - * extents, so only retained extents may remain. + * extents, so only retained extents may remain and it's safe to call + * pa_shard_destroy_retained. */ - assert(extents_npages_get(&arena->extents_dirty) == 0); - assert(extents_npages_get(&arena->extents_muzzy) == 0); - - /* Deallocate retained memory. */ - arena_destroy_retained(tsd_tsdn(tsd), arena); + pa_shard_destroy(tsd_tsdn(tsd), &arena->pa_shard); /* * Remove the arena pointer from the arenas array. We rely on the fact @@ -1197,316 +820,370 @@ /* * Destroy the base allocator, which manages all metadata ever mapped by - * this arena. + * this arena. The prepare function will make sure no pending access to + * the metadata in this base anymore. */ + arena_prepare_base_deletion(tsd, arena->base); base_delete(tsd_tsdn(tsd), arena->base); } -static extent_t * -arena_slab_alloc_hard(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, const bin_info_t *bin_info, - szind_t szind) { - extent_t *slab; - bool zero, commit; - - witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), - WITNESS_RANK_CORE, 0); - - zero = false; - commit = true; - slab = extent_alloc_wrapper(tsdn, arena, r_extent_hooks, NULL, - bin_info->slab_size, 0, PAGE, true, szind, &zero, &commit); - - if (config_stats && slab != NULL) { - arena_stats_mapped_add(tsdn, &arena->stats, - bin_info->slab_size); - } - - return slab; -} - -static extent_t * +static edata_t * arena_slab_alloc(tsdn_t *tsdn, arena_t *arena, szind_t binind, unsigned binshard, const bin_info_t *bin_info) { + bool deferred_work_generated = false; witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), WITNESS_RANK_CORE, 0); - extent_hooks_t *extent_hooks = EXTENT_HOOKS_INITIALIZER; - szind_t szind = sz_size2index(bin_info->reg_size); - bool zero = false; - bool commit = true; - extent_t *slab = extents_alloc(tsdn, arena, &extent_hooks, - &arena->extents_dirty, NULL, bin_info->slab_size, 0, PAGE, true, - binind, &zero, &commit); - if (slab == NULL && arena_may_have_muzzy(arena)) { - slab = extents_alloc(tsdn, arena, &extent_hooks, - &arena->extents_muzzy, NULL, bin_info->slab_size, 0, PAGE, - true, binind, &zero, &commit); + bool guarded = san_slab_extent_decide_guard(tsdn, + arena_get_ehooks(arena)); + edata_t *slab = pa_alloc(tsdn, &arena->pa_shard, bin_info->slab_size, + /* alignment */ PAGE, /* slab */ true, /* szind */ binind, + /* zero */ false, guarded, &deferred_work_generated); + + if (deferred_work_generated) { + arena_handle_deferred_work(tsdn, arena); } + if (slab == NULL) { - slab = arena_slab_alloc_hard(tsdn, arena, &extent_hooks, - bin_info, szind); - if (slab == NULL) { - return NULL; - } + return NULL; } - assert(extent_slab_get(slab)); + assert(edata_slab_get(slab)); /* Initialize slab internals. */ - arena_slab_data_t *slab_data = extent_slab_data_get(slab); - extent_nfree_binshard_set(slab, bin_info->nregs, binshard); + slab_data_t *slab_data = edata_slab_data_get(slab); + edata_nfree_binshard_set(slab, bin_info->nregs, binshard); bitmap_init(slab_data->bitmap, &bin_info->bitmap_info, false); - arena_nactive_add(arena, extent_size_get(slab) >> LG_PAGE); - return slab; } -static extent_t * -arena_bin_nonfull_slab_get(tsdn_t *tsdn, arena_t *arena, bin_t *bin, - szind_t binind, unsigned binshard) { - extent_t *slab; - const bin_info_t *bin_info; - - /* Look for a usable slab. */ - slab = arena_bin_slabs_nonfull_tryget(bin); - if (slab != NULL) { - return slab; - } - /* No existing slabs have any space available. */ - - bin_info = &bin_infos[binind]; - - /* Allocate a new slab. */ - malloc_mutex_unlock(tsdn, &bin->lock); - /******************************/ - slab = arena_slab_alloc(tsdn, arena, binind, binshard, bin_info); - /********************************/ - malloc_mutex_lock(tsdn, &bin->lock); - if (slab != NULL) { - if (config_stats) { - bin->stats.nslabs++; - bin->stats.curslabs++; - } - return slab; +/* + * Before attempting the _with_fresh_slab approaches below, the _no_fresh_slab + * variants (i.e. through slabcur and nonfull) must be tried first. + */ +static void +arena_bin_refill_slabcur_with_fresh_slab(tsdn_t *tsdn, arena_t *arena, + bin_t *bin, szind_t binind, edata_t *fresh_slab) { + malloc_mutex_assert_owner(tsdn, &bin->lock); + /* Only called after slabcur and nonfull both failed. */ + assert(bin->slabcur == NULL); + assert(edata_heap_first(&bin->slabs_nonfull) == NULL); + assert(fresh_slab != NULL); + + /* A new slab from arena_slab_alloc() */ + assert(edata_nfree_get(fresh_slab) == bin_infos[binind].nregs); + if (config_stats) { + bin->stats.nslabs++; + bin->stats.curslabs++; } + bin->slabcur = fresh_slab; +} - /* - * arena_slab_alloc() failed, but another thread may have made - * sufficient memory available while this one dropped bin->lock above, - * so search one more time. - */ - slab = arena_bin_slabs_nonfull_tryget(bin); - if (slab != NULL) { - return slab; - } +/* Refill slabcur and then alloc using the fresh slab */ +static void * +arena_bin_malloc_with_fresh_slab(tsdn_t *tsdn, arena_t *arena, bin_t *bin, + szind_t binind, edata_t *fresh_slab) { + malloc_mutex_assert_owner(tsdn, &bin->lock); + arena_bin_refill_slabcur_with_fresh_slab(tsdn, arena, bin, binind, + fresh_slab); - return NULL; + return arena_slab_reg_alloc(bin->slabcur, &bin_infos[binind]); } -/* Re-fill bin->slabcur, then call arena_slab_reg_alloc(). */ -static void * -arena_bin_malloc_hard(tsdn_t *tsdn, arena_t *arena, bin_t *bin, - szind_t binind, unsigned binshard) { - const bin_info_t *bin_info; - extent_t *slab; +static bool +arena_bin_refill_slabcur_no_fresh_slab(tsdn_t *tsdn, arena_t *arena, + bin_t *bin) { + malloc_mutex_assert_owner(tsdn, &bin->lock); + /* Only called after arena_slab_reg_alloc[_batch] failed. */ + assert(bin->slabcur == NULL || edata_nfree_get(bin->slabcur) == 0); - bin_info = &bin_infos[binind]; - if (!arena_is_auto(arena) && bin->slabcur != NULL) { - arena_bin_slabs_full_insert(arena, bin, bin->slabcur); - bin->slabcur = NULL; - } - slab = arena_bin_nonfull_slab_get(tsdn, arena, bin, binind, binshard); if (bin->slabcur != NULL) { - /* - * Another thread updated slabcur while this one ran without the - * bin lock in arena_bin_nonfull_slab_get(). - */ - if (extent_nfree_get(bin->slabcur) > 0) { - void *ret = arena_slab_reg_alloc(bin->slabcur, - bin_info); - if (slab != NULL) { - /* - * arena_slab_alloc() may have allocated slab, - * or it may have been pulled from - * slabs_nonfull. Therefore it is unsafe to - * make any assumptions about how slab has - * previously been used, and - * arena_bin_lower_slab() must be called, as if - * a region were just deallocated from the slab. - */ - if (extent_nfree_get(slab) == bin_info->nregs) { - arena_dalloc_bin_slab(tsdn, arena, slab, - bin); - } else { - arena_bin_lower_slab(tsdn, arena, slab, - bin); - } - } - return ret; - } - arena_bin_slabs_full_insert(arena, bin, bin->slabcur); - bin->slabcur = NULL; - } - - if (slab == NULL) { - return NULL; } - bin->slabcur = slab; - assert(extent_nfree_get(bin->slabcur) > 0); + /* Look for a usable slab. */ + bin->slabcur = arena_bin_slabs_nonfull_tryget(bin); + assert(bin->slabcur == NULL || edata_nfree_get(bin->slabcur) > 0); - return arena_slab_reg_alloc(slab, bin_info); + return (bin->slabcur == NULL); } -/* Choose a bin shard and return the locked bin. */ bin_t * -arena_bin_choose_lock(tsdn_t *tsdn, arena_t *arena, szind_t binind, - unsigned *binshard) { - bin_t *bin; +arena_bin_choose(tsdn_t *tsdn, arena_t *arena, szind_t binind, + unsigned *binshard_p) { + unsigned binshard; if (tsdn_null(tsdn) || tsd_arena_get(tsdn_tsd(tsdn)) == NULL) { - *binshard = 0; + binshard = 0; } else { - *binshard = tsd_binshardsp_get(tsdn_tsd(tsdn))->binshard[binind]; + binshard = tsd_binshardsp_get(tsdn_tsd(tsdn))->binshard[binind]; } - assert(*binshard < bin_infos[binind].n_shards); - bin = &arena->bins[binind].bin_shards[*binshard]; - malloc_mutex_lock(tsdn, &bin->lock); - - return bin; + assert(binshard < bin_infos[binind].n_shards); + if (binshard_p != NULL) { + *binshard_p = binshard; + } + return arena_get_bin(arena, binind, binshard); } void -arena_tcache_fill_small(tsdn_t *tsdn, arena_t *arena, tcache_t *tcache, - cache_bin_t *tbin, szind_t binind, uint64_t prof_accumbytes) { - unsigned i, nfill, cnt; +arena_cache_bin_fill_small(tsdn_t *tsdn, arena_t *arena, + cache_bin_t *cache_bin, cache_bin_info_t *cache_bin_info, szind_t binind, + const unsigned nfill) { + assert(cache_bin_ncached_get_local(cache_bin, cache_bin_info) == 0); + + const bin_info_t *bin_info = &bin_infos[binind]; + + CACHE_BIN_PTR_ARRAY_DECLARE(ptrs, nfill); + cache_bin_init_ptr_array_for_fill(cache_bin, cache_bin_info, &ptrs, + nfill); + /* + * Bin-local resources are used first: 1) bin->slabcur, and 2) nonfull + * slabs. After both are exhausted, new slabs will be allocated through + * arena_slab_alloc(). + * + * Bin lock is only taken / released right before / after the while(...) + * refill loop, with new slab allocation (which has its own locking) + * kept outside of the loop. This setup facilitates flat combining, at + * the cost of the nested loop (through goto label_refill). + * + * To optimize for cases with contention and limited resources + * (e.g. hugepage-backed or non-overcommit arenas), each fill-iteration + * gets one chance of slab_alloc, and a retry of bin local resources + * after the slab allocation (regardless if slab_alloc failed, because + * the bin lock is dropped during the slab allocation). + * + * In other words, new slab allocation is allowed, as long as there was + * progress since the previous slab_alloc. This is tracked with + * made_progress below, initialized to true to jump start the first + * iteration. + * + * In other words (again), the loop will only terminate early (i.e. stop + * with filled < nfill) after going through the three steps: a) bin + * local exhausted, b) unlock and slab_alloc returns null, c) re-lock + * and bin local fails again. + */ + bool made_progress = true; + edata_t *fresh_slab = NULL; + bool alloc_and_retry = false; + unsigned filled = 0; + unsigned binshard; + bin_t *bin = arena_bin_choose(tsdn, arena, binind, &binshard); + +label_refill: + malloc_mutex_lock(tsdn, &bin->lock); + + while (filled < nfill) { + /* Try batch-fill from slabcur first. */ + edata_t *slabcur = bin->slabcur; + if (slabcur != NULL && edata_nfree_get(slabcur) > 0) { + unsigned tofill = nfill - filled; + unsigned nfree = edata_nfree_get(slabcur); + unsigned cnt = tofill < nfree ? tofill : nfree; + + arena_slab_reg_alloc_batch(slabcur, bin_info, cnt, + &ptrs.ptr[filled]); + made_progress = true; + filled += cnt; + continue; + } + /* Next try refilling slabcur from nonfull slabs. */ + if (!arena_bin_refill_slabcur_no_fresh_slab(tsdn, arena, bin)) { + assert(bin->slabcur != NULL); + continue; + } + + /* Then see if a new slab was reserved already. */ + if (fresh_slab != NULL) { + arena_bin_refill_slabcur_with_fresh_slab(tsdn, arena, + bin, binind, fresh_slab); + assert(bin->slabcur != NULL); + fresh_slab = NULL; + continue; + } + + /* Try slab_alloc if made progress (or never did slab_alloc). */ + if (made_progress) { + assert(bin->slabcur == NULL); + assert(fresh_slab == NULL); + alloc_and_retry = true; + /* Alloc a new slab then come back. */ + break; + } + + /* OOM. */ + + assert(fresh_slab == NULL); + assert(!alloc_and_retry); + break; + } /* while (filled < nfill) loop. */ + + if (config_stats && !alloc_and_retry) { + bin->stats.nmalloc += filled; + bin->stats.nrequests += cache_bin->tstats.nrequests; + bin->stats.curregs += filled; + bin->stats.nfills++; + cache_bin->tstats.nrequests = 0; + } + + malloc_mutex_unlock(tsdn, &bin->lock); - assert(tbin->ncached == 0); + if (alloc_and_retry) { + assert(fresh_slab == NULL); + assert(filled < nfill); + assert(made_progress); - if (config_prof && arena_prof_accum(tsdn, arena, prof_accumbytes)) { - prof_idump(tsdn); + fresh_slab = arena_slab_alloc(tsdn, arena, binind, binshard, + bin_info); + /* fresh_slab NULL case handled in the for loop. */ + + alloc_and_retry = false; + made_progress = false; + goto label_refill; + } + assert(filled == nfill || (fresh_slab == NULL && !made_progress)); + + /* Release if allocated but not used. */ + if (fresh_slab != NULL) { + assert(edata_nfree_get(fresh_slab) == bin_info->nregs); + arena_slab_dalloc(tsdn, arena, fresh_slab); + fresh_slab = NULL; } + cache_bin_finish_fill(cache_bin, cache_bin_info, &ptrs, filled); + arena_decay_tick(tsdn, arena); +} + +size_t +arena_fill_small_fresh(tsdn_t *tsdn, arena_t *arena, szind_t binind, + void **ptrs, size_t nfill, bool zero) { + assert(binind < SC_NBINS); + const bin_info_t *bin_info = &bin_infos[binind]; + const size_t nregs = bin_info->nregs; + assert(nregs > 0); + const size_t usize = bin_info->reg_size; + + const bool manual_arena = !arena_is_auto(arena); unsigned binshard; - bin_t *bin = arena_bin_choose_lock(tsdn, arena, binind, &binshard); - - for (i = 0, nfill = (tcache_bin_info[binind].ncached_max >> - tcache->lg_fill_div[binind]); i < nfill; i += cnt) { - extent_t *slab; - if ((slab = bin->slabcur) != NULL && extent_nfree_get(slab) > - 0) { - unsigned tofill = nfill - i; - cnt = tofill < extent_nfree_get(slab) ? - tofill : extent_nfree_get(slab); - arena_slab_reg_alloc_batch( - slab, &bin_infos[binind], cnt, - tbin->avail - nfill + i); - } else { - cnt = 1; - void *ptr = arena_bin_malloc_hard(tsdn, arena, bin, - binind, binshard); - /* - * OOM. tbin->avail isn't yet filled down to its first - * element, so the successful allocations (if any) must - * be moved just before tbin->avail before bailing out. - */ - if (ptr == NULL) { - if (i > 0) { - memmove(tbin->avail - i, - tbin->avail - nfill, - i * sizeof(void *)); - } - break; - } - /* Insert such that low regions get used first. */ - *(tbin->avail - nfill + i) = ptr; + bin_t *bin = arena_bin_choose(tsdn, arena, binind, &binshard); + + size_t nslab = 0; + size_t filled = 0; + edata_t *slab = NULL; + edata_list_active_t fulls; + edata_list_active_init(&fulls); + + while (filled < nfill && (slab = arena_slab_alloc(tsdn, arena, binind, + binshard, bin_info)) != NULL) { + assert((size_t)edata_nfree_get(slab) == nregs); + ++nslab; + size_t batch = nfill - filled; + if (batch > nregs) { + batch = nregs; } - if (config_fill && unlikely(opt_junk_alloc)) { - for (unsigned j = 0; j < cnt; j++) { - void* ptr = *(tbin->avail - nfill + i + j); - arena_alloc_junk_small(ptr, &bin_infos[binind], - true); + assert(batch > 0); + arena_slab_reg_alloc_batch(slab, bin_info, (unsigned)batch, + &ptrs[filled]); + assert(edata_addr_get(slab) == ptrs[filled]); + if (zero) { + memset(ptrs[filled], 0, batch * usize); + } + filled += batch; + if (batch == nregs) { + if (manual_arena) { + edata_list_active_append(&fulls, slab); } + slab = NULL; } } + + malloc_mutex_lock(tsdn, &bin->lock); + /* + * Only the last slab can be non-empty, and the last slab is non-empty + * iff slab != NULL. + */ + if (slab != NULL) { + arena_bin_lower_slab(tsdn, arena, slab, bin); + } + if (manual_arena) { + edata_list_active_concat(&bin->slabs_full, &fulls); + } + assert(edata_list_active_empty(&fulls)); if (config_stats) { - bin->stats.nmalloc += i; - bin->stats.nrequests += tbin->tstats.nrequests; - bin->stats.curregs += i; - bin->stats.nfills++; - tbin->tstats.nrequests = 0; + bin->stats.nslabs += nslab; + bin->stats.curslabs += nslab; + bin->stats.nmalloc += filled; + bin->stats.nrequests += filled; + bin->stats.curregs += filled; } malloc_mutex_unlock(tsdn, &bin->lock); - tbin->ncached = i; + arena_decay_tick(tsdn, arena); + return filled; } -void -arena_alloc_junk_small(void *ptr, const bin_info_t *bin_info, bool zero) { - if (!zero) { - memset(ptr, JEMALLOC_ALLOC_JUNK, bin_info->reg_size); +/* + * Without allocating a new slab, try arena_slab_reg_alloc() and re-fill + * bin->slabcur if necessary. + */ +static void * +arena_bin_malloc_no_fresh_slab(tsdn_t *tsdn, arena_t *arena, bin_t *bin, + szind_t binind) { + malloc_mutex_assert_owner(tsdn, &bin->lock); + if (bin->slabcur == NULL || edata_nfree_get(bin->slabcur) == 0) { + if (arena_bin_refill_slabcur_no_fresh_slab(tsdn, arena, bin)) { + return NULL; + } } -} -static void -arena_dalloc_junk_small_impl(void *ptr, const bin_info_t *bin_info) { - memset(ptr, JEMALLOC_FREE_JUNK, bin_info->reg_size); + assert(bin->slabcur != NULL && edata_nfree_get(bin->slabcur) > 0); + return arena_slab_reg_alloc(bin->slabcur, &bin_infos[binind]); } -arena_dalloc_junk_small_t *JET_MUTABLE arena_dalloc_junk_small = - arena_dalloc_junk_small_impl; static void * arena_malloc_small(tsdn_t *tsdn, arena_t *arena, szind_t binind, bool zero) { - void *ret; - bin_t *bin; - size_t usize; - extent_t *slab; - assert(binind < SC_NBINS); - usize = sz_index2size(binind); + const bin_info_t *bin_info = &bin_infos[binind]; + size_t usize = sz_index2size(binind); unsigned binshard; - bin = arena_bin_choose_lock(tsdn, arena, binind, &binshard); - - if ((slab = bin->slabcur) != NULL && extent_nfree_get(slab) > 0) { - ret = arena_slab_reg_alloc(slab, &bin_infos[binind]); - } else { - ret = arena_bin_malloc_hard(tsdn, arena, bin, binind, binshard); - } + bin_t *bin = arena_bin_choose(tsdn, arena, binind, &binshard); + malloc_mutex_lock(tsdn, &bin->lock); + edata_t *fresh_slab = NULL; + void *ret = arena_bin_malloc_no_fresh_slab(tsdn, arena, bin, binind); if (ret == NULL) { malloc_mutex_unlock(tsdn, &bin->lock); - return NULL; + /******************************/ + fresh_slab = arena_slab_alloc(tsdn, arena, binind, binshard, + bin_info); + /********************************/ + malloc_mutex_lock(tsdn, &bin->lock); + /* Retry since the lock was dropped. */ + ret = arena_bin_malloc_no_fresh_slab(tsdn, arena, bin, binind); + if (ret == NULL) { + if (fresh_slab == NULL) { + /* OOM */ + malloc_mutex_unlock(tsdn, &bin->lock); + return NULL; + } + ret = arena_bin_malloc_with_fresh_slab(tsdn, arena, bin, + binind, fresh_slab); + fresh_slab = NULL; + } } - if (config_stats) { bin->stats.nmalloc++; bin->stats.nrequests++; bin->stats.curregs++; } malloc_mutex_unlock(tsdn, &bin->lock); - if (config_prof && arena_prof_accum(tsdn, arena, usize)) { - prof_idump(tsdn); - } - if (!zero) { - if (config_fill) { - if (unlikely(opt_junk_alloc)) { - arena_alloc_junk_small(ret, - &bin_infos[binind], false); - } else if (unlikely(opt_zero)) { - memset(ret, 0, usize); - } - } - } else { - if (config_fill && unlikely(opt_junk_alloc)) { - arena_alloc_junk_small(ret, &bin_infos[binind], - true); - } + if (fresh_slab != NULL) { + arena_slab_dalloc(tsdn, arena, fresh_slab); + } + if (zero) { memset(ret, 0, usize); } - arena_decay_tick(tsdn, arena); + return ret; } @@ -1533,10 +1210,17 @@ bool zero, tcache_t *tcache) { void *ret; - if (usize <= SC_SMALL_MAXCLASS - && (alignment < PAGE - || (alignment == PAGE && (usize & PAGE_MASK) == 0))) { + if (usize <= SC_SMALL_MAXCLASS) { /* Small; alignment doesn't require special slab placement. */ + + /* usize should be a result of sz_sa2u() */ + assert((usize & (alignment - 1)) == 0); + + /* + * Small usize can't come from an alignment larger than a page. + */ + assert(alignment <= PAGE); + ret = arena_malloc(tsdn, arena, usize, sz_size2index(usize), zero, tcache, true); } else { @@ -1560,33 +1244,22 @@ safety_check_set_redzone(ptr, usize, SC_LARGE_MINCLASS); } - rtree_ctx_t rtree_ctx_fallback; - rtree_ctx_t *rtree_ctx = tsdn_rtree_ctx(tsdn, &rtree_ctx_fallback); - - extent_t *extent = rtree_extent_read(tsdn, &extents_rtree, rtree_ctx, - (uintptr_t)ptr, true); - arena_t *arena = extent_arena_get(extent); + edata_t *edata = emap_edata_lookup(tsdn, &arena_emap_global, ptr); szind_t szind = sz_size2index(usize); - extent_szind_set(extent, szind); - rtree_szind_slab_update(tsdn, &extents_rtree, rtree_ctx, (uintptr_t)ptr, - szind, false); - - prof_accum_cancel(tsdn, &arena->prof_accum, usize); + edata_szind_set(edata, szind); + emap_remap(tsdn, &arena_emap_global, edata, szind, /* slab */ false); assert(isalloc(tsdn, ptr) == usize); } static size_t -arena_prof_demote(tsdn_t *tsdn, extent_t *extent, const void *ptr) { +arena_prof_demote(tsdn_t *tsdn, edata_t *edata, const void *ptr) { cassert(config_prof); assert(ptr != NULL); - extent_szind_set(extent, SC_NBINS); - rtree_ctx_t rtree_ctx_fallback; - rtree_ctx_t *rtree_ctx = tsdn_rtree_ctx(tsdn, &rtree_ctx_fallback); - rtree_szind_slab_update(tsdn, &extents_rtree, rtree_ctx, (uintptr_t)ptr, - SC_NBINS, false); + edata_szind_set(edata, SC_NBINS); + emap_remap(tsdn, &arena_emap_global, edata, SC_NBINS, /* slab */ false); assert(isalloc(tsdn, ptr) == SC_LARGE_MINCLASS); @@ -1599,9 +1272,9 @@ cassert(config_prof); assert(opt_prof); - extent_t *extent = iealloc(tsdn, ptr); - size_t usize = extent_usize_get(extent); - size_t bumped_usize = arena_prof_demote(tsdn, extent, ptr); + edata_t *edata = emap_edata_lookup(tsdn, &arena_emap_global, ptr); + size_t usize = edata_usize_get(edata); + size_t bumped_usize = arena_prof_demote(tsdn, edata, ptr); if (config_opt_safety_checks && usize < SC_LARGE_MINCLASS) { /* * Currently, we only do redzoning for small sampled @@ -1614,17 +1287,17 @@ tcache_dalloc_large(tsdn_tsd(tsdn), tcache, ptr, sz_size2index(bumped_usize), slow_path); } else { - large_dalloc(tsdn, extent); + large_dalloc(tsdn, edata); } } static void -arena_dissociate_bin_slab(arena_t *arena, extent_t *slab, bin_t *bin) { +arena_dissociate_bin_slab(arena_t *arena, edata_t *slab, bin_t *bin) { /* Dissociate slab from bin. */ if (slab == bin->slabcur) { bin->slabcur = NULL; } else { - szind_t binind = extent_szind_get(slab); + szind_t binind = edata_szind_get(slab); const bin_info_t *bin_info = &bin_infos[binind]; /* @@ -1641,24 +1314,9 @@ } static void -arena_dalloc_bin_slab(tsdn_t *tsdn, arena_t *arena, extent_t *slab, - bin_t *bin) { - assert(slab != bin->slabcur); - - malloc_mutex_unlock(tsdn, &bin->lock); - /******************************/ - arena_slab_dalloc(tsdn, arena, slab); - /****************************/ - malloc_mutex_lock(tsdn, &bin->lock); - if (config_stats) { - bin->stats.curslabs--; - } -} - -static void -arena_bin_lower_slab(tsdn_t *tsdn, arena_t *arena, extent_t *slab, +arena_bin_lower_slab(tsdn_t *tsdn, arena_t *arena, edata_t *slab, bin_t *bin) { - assert(extent_nfree_get(slab) > 0); + assert(edata_nfree_get(slab) > 0); /* * Make sure that if bin->slabcur is non-NULL, it refers to the @@ -1666,9 +1324,9 @@ * than proactively keeping it pointing at the oldest/lowest non-full * slab. */ - if (bin->slabcur != NULL && extent_snad_comp(bin->slabcur, slab) > 0) { + if (bin->slabcur != NULL && edata_snad_comp(bin->slabcur, slab) > 0) { /* Switch slabcur. */ - if (extent_nfree_get(bin->slabcur) > 0) { + if (edata_nfree_get(bin->slabcur) > 0) { arena_bin_slabs_nonfull_insert(bin, bin->slabcur); } else { arena_bin_slabs_full_insert(arena, bin, bin->slabcur); @@ -1683,56 +1341,54 @@ } static void -arena_dalloc_bin_locked_impl(tsdn_t *tsdn, arena_t *arena, bin_t *bin, - szind_t binind, extent_t *slab, void *ptr, bool junked) { - arena_slab_data_t *slab_data = extent_slab_data_get(slab); - const bin_info_t *bin_info = &bin_infos[binind]; - - if (!junked && config_fill && unlikely(opt_junk_free)) { - arena_dalloc_junk_small(ptr, bin_info); - } - - arena_slab_reg_dalloc(slab, slab_data, ptr); - unsigned nfree = extent_nfree_get(slab); - if (nfree == bin_info->nregs) { - arena_dissociate_bin_slab(arena, slab, bin); - arena_dalloc_bin_slab(tsdn, arena, slab, bin); - } else if (nfree == 1 && slab != bin->slabcur) { - arena_bin_slabs_full_remove(arena, bin, slab); - arena_bin_lower_slab(tsdn, arena, slab, bin); - } +arena_dalloc_bin_slab_prepare(tsdn_t *tsdn, edata_t *slab, bin_t *bin) { + malloc_mutex_assert_owner(tsdn, &bin->lock); + assert(slab != bin->slabcur); if (config_stats) { - bin->stats.ndalloc++; - bin->stats.curregs--; + bin->stats.curslabs--; } } void -arena_dalloc_bin_junked_locked(tsdn_t *tsdn, arena_t *arena, bin_t *bin, - szind_t binind, extent_t *extent, void *ptr) { - arena_dalloc_bin_locked_impl(tsdn, arena, bin, binind, extent, ptr, - true); +arena_dalloc_bin_locked_handle_newly_empty(tsdn_t *tsdn, arena_t *arena, + edata_t *slab, bin_t *bin) { + arena_dissociate_bin_slab(arena, slab, bin); + arena_dalloc_bin_slab_prepare(tsdn, slab, bin); +} + +void +arena_dalloc_bin_locked_handle_newly_nonempty(tsdn_t *tsdn, arena_t *arena, + edata_t *slab, bin_t *bin) { + arena_bin_slabs_full_remove(arena, bin, slab); + arena_bin_lower_slab(tsdn, arena, slab, bin); } static void -arena_dalloc_bin(tsdn_t *tsdn, arena_t *arena, extent_t *extent, void *ptr) { - szind_t binind = extent_szind_get(extent); - unsigned binshard = extent_binshard_get(extent); - bin_t *bin = &arena->bins[binind].bin_shards[binshard]; +arena_dalloc_bin(tsdn_t *tsdn, arena_t *arena, edata_t *edata, void *ptr) { + szind_t binind = edata_szind_get(edata); + unsigned binshard = edata_binshard_get(edata); + bin_t *bin = arena_get_bin(arena, binind, binshard); malloc_mutex_lock(tsdn, &bin->lock); - arena_dalloc_bin_locked_impl(tsdn, arena, bin, binind, extent, ptr, - false); + arena_dalloc_bin_locked_info_t info; + arena_dalloc_bin_locked_begin(&info, binind); + bool ret = arena_dalloc_bin_locked_step(tsdn, arena, bin, + &info, binind, edata, ptr); + arena_dalloc_bin_locked_finish(tsdn, arena, bin, &info); malloc_mutex_unlock(tsdn, &bin->lock); + + if (ret) { + arena_slab_dalloc(tsdn, arena, edata); + } } void arena_dalloc_small(tsdn_t *tsdn, void *ptr) { - extent_t *extent = iealloc(tsdn, ptr); - arena_t *arena = extent_arena_get(extent); + edata_t *edata = emap_edata_lookup(tsdn, &arena_emap_global, ptr); + arena_t *arena = arena_get_from_edata(edata); - arena_dalloc_bin(tsdn, arena, extent, ptr); + arena_dalloc_bin(tsdn, arena, edata, ptr); arena_decay_tick(tsdn, arena); } @@ -1743,7 +1399,7 @@ /* Calls with non-zero extra had to clamp extra. */ assert(extra == 0 || size + extra <= SC_LARGE_MAXCLASS); - extent_t *extent = iealloc(tsdn, ptr); + edata_t *edata = emap_edata_lookup(tsdn, &arena_emap_global, ptr); if (unlikely(size > SC_LARGE_MAXCLASS)) { ret = true; goto done; @@ -1766,18 +1422,19 @@ goto done; } - arena_decay_tick(tsdn, extent_arena_get(extent)); + arena_t *arena = arena_get_from_edata(edata); + arena_decay_tick(tsdn, arena); ret = false; } else if (oldsize >= SC_LARGE_MINCLASS && usize_max >= SC_LARGE_MINCLASS) { - ret = large_ralloc_no_move(tsdn, extent, usize_min, usize_max, + ret = large_ralloc_no_move(tsdn, edata, usize_min, usize_max, zero); } else { ret = true; } done: - assert(extent == iealloc(tsdn, ptr)); - *newsize = extent_usize_get(extent); + assert(edata == emap_edata_lookup(tsdn, &arena_emap_global, ptr)); + *newsize = edata_usize_get(edata); return ret; } @@ -1800,7 +1457,7 @@ arena_ralloc(tsdn_t *tsdn, arena_t *arena, void *ptr, size_t oldsize, size_t size, size_t alignment, bool zero, tcache_t *tcache, hook_ralloc_args_t *hook_args) { - size_t usize = sz_s2u(size); + size_t usize = alignment == 0 ? sz_s2u(size) : sz_sa2u(size, alignment); if (unlikely(usize == 0 || size > SC_LARGE_MAXCLASS)) { return NULL; } @@ -1850,6 +1507,29 @@ return ret; } +ehooks_t * +arena_get_ehooks(arena_t *arena) { + return base_ehooks_get(arena->base); +} + +extent_hooks_t * +arena_set_extent_hooks(tsd_t *tsd, arena_t *arena, + extent_hooks_t *extent_hooks) { + background_thread_info_t *info; + if (have_background_thread) { + info = arena_background_thread_info_get(arena); + malloc_mutex_lock(tsd_tsdn(tsd), &info->mtx); + } + /* No using the HPA now that we have the custom hooks. */ + pa_shard_disable_hpa(tsd_tsdn(tsd), &arena->pa_shard); + extent_hooks_t *ret = base_extent_hooks_set(arena->base, extent_hooks); + if (have_background_thread) { + malloc_mutex_unlock(tsd_tsdn(tsd), &info->mtx); + } + + return ret; +} + dss_prec_t arena_dss_prec_get(arena_t *arena) { return (dss_prec_t)atomic_load_u(&arena->dss_prec, ATOMIC_ACQUIRE); @@ -1871,7 +1551,7 @@ bool arena_dirty_decay_ms_default_set(ssize_t decay_ms) { - if (!arena_decay_ms_valid(decay_ms)) { + if (!decay_ms_valid(decay_ms)) { return true; } atomic_store_zd(&dirty_decay_ms_default, decay_ms, ATOMIC_RELAXED); @@ -1885,7 +1565,7 @@ bool arena_muzzy_decay_ms_default_set(ssize_t decay_ms) { - if (!arena_decay_ms_valid(decay_ms)) { + if (!decay_ms_valid(decay_ms)) { return true; } atomic_store_zd(&muzzy_decay_ms_default, decay_ms, ATOMIC_RELAXED); @@ -1896,26 +1576,8 @@ arena_retain_grow_limit_get_set(tsd_t *tsd, arena_t *arena, size_t *old_limit, size_t *new_limit) { assert(opt_retain); - - pszind_t new_ind JEMALLOC_CC_SILENCE_INIT(0); - if (new_limit != NULL) { - size_t limit = *new_limit; - /* Grow no more than the new limit. */ - if ((new_ind = sz_psz2ind(limit + 1) - 1) >= SC_NPSIZES) { - return true; - } - } - - malloc_mutex_lock(tsd_tsdn(tsd), &arena->extent_grow_mtx); - if (old_limit != NULL) { - *old_limit = sz_pind2sz(arena->retain_grow_limit); - } - if (new_limit != NULL) { - arena->retain_grow_limit = new_ind; - } - malloc_mutex_unlock(tsd_tsdn(tsd), &arena->extent_grow_mtx); - - return false; + return pac_retain_grow_limit_get_set(tsd_tsdn(tsd), + &arena->pa_shard.pac, old_limit, new_limit); } unsigned @@ -1933,13 +1595,8 @@ atomic_fetch_sub_u(&arena->nthreads[internal], 1, ATOMIC_RELAXED); } -size_t -arena_extent_sn_next(arena_t *arena) { - return atomic_fetch_add_zu(&arena->extent_sn_next, 1, ATOMIC_RELAXED); -} - arena_t * -arena_new(tsdn_t *tsdn, unsigned ind, extent_hooks_t *extent_hooks) { +arena_new(tsdn_t *tsdn, unsigned ind, const arena_config_t *config) { arena_t *arena; base_t *base; unsigned i; @@ -1947,16 +1604,13 @@ if (ind == 0) { base = b0get(); } else { - base = base_new(tsdn, ind, extent_hooks); + base = base_new(tsdn, ind, config->extent_hooks, + config->metadata_use_hooks); if (base == NULL) { return NULL; } } - unsigned nbins_total = 0; - for (i = 0; i < SC_NBINS; i++) { - nbins_total += bin_infos[i].n_shards; - } size_t arena_size = sizeof(arena_t) + sizeof(bin_t) * nbins_total; arena = (arena_t *)base_alloc(tsdn, base, arena_size, CACHELINE); if (arena == NULL) { @@ -1980,110 +1634,56 @@ } } - if (config_prof) { - if (prof_accum_init(tsdn, &arena->prof_accum)) { - goto label_error; - } - } - - if (config_cache_oblivious) { - /* - * A nondeterministic seed based on the address of arena reduces - * the likelihood of lockstep non-uniform cache index - * utilization among identical concurrent processes, but at the - * cost of test repeatability. For debug builds, instead use a - * deterministic seed. - */ - atomic_store_zu(&arena->offset_state, config_debug ? ind : - (size_t)(uintptr_t)arena, ATOMIC_RELAXED); - } - - atomic_store_zu(&arena->extent_sn_next, 0, ATOMIC_RELAXED); - atomic_store_u(&arena->dss_prec, (unsigned)extent_dss_prec_get(), ATOMIC_RELAXED); - atomic_store_zu(&arena->nactive, 0, ATOMIC_RELAXED); - - extent_list_init(&arena->large); + edata_list_active_init(&arena->large); if (malloc_mutex_init(&arena->large_mtx, "arena_large", WITNESS_RANK_ARENA_LARGE, malloc_mutex_rank_exclusive)) { goto label_error; } - /* - * Delay coalescing for dirty extents despite the disruptive effect on - * memory layout for best-fit extent allocation, since cached extents - * are likely to be reused soon after deallocation, and the cost of - * merging/splitting extents is non-trivial. - */ - if (extents_init(tsdn, &arena->extents_dirty, extent_state_dirty, - true)) { - goto label_error; - } - /* - * Coalesce muzzy extents immediately, because operations on them are in - * the critical path much less often than for dirty extents. - */ - if (extents_init(tsdn, &arena->extents_muzzy, extent_state_muzzy, - false)) { - goto label_error; - } - /* - * Coalesce retained extents immediately, in part because they will - * never be evicted (and therefore there's no opportunity for delayed - * coalescing), but also because operations on retained extents are not - * in the critical path. - */ - if (extents_init(tsdn, &arena->extents_retained, extent_state_retained, - false)) { - goto label_error; - } - - if (arena_decay_init(&arena->decay_dirty, - arena_dirty_decay_ms_default_get(), &arena->stats.decay_dirty)) { - goto label_error; - } - if (arena_decay_init(&arena->decay_muzzy, - arena_muzzy_decay_ms_default_get(), &arena->stats.decay_muzzy)) { - goto label_error; - } - - arena->extent_grow_next = sz_psz2ind(HUGEPAGE); - arena->retain_grow_limit = sz_psz2ind(SC_LARGE_MAXCLASS); - if (malloc_mutex_init(&arena->extent_grow_mtx, "extent_grow", - WITNESS_RANK_EXTENT_GROW, malloc_mutex_rank_exclusive)) { - goto label_error; - } - - extent_avail_new(&arena->extent_avail); - if (malloc_mutex_init(&arena->extent_avail_mtx, "extent_avail", - WITNESS_RANK_EXTENT_AVAIL, malloc_mutex_rank_exclusive)) { + nstime_t cur_time; + nstime_init_update(&cur_time); + if (pa_shard_init(tsdn, &arena->pa_shard, &arena_pa_central_global, + &arena_emap_global, base, ind, &arena->stats.pa_shard_stats, + LOCKEDINT_MTX(arena->stats.mtx), &cur_time, oversize_threshold, + arena_dirty_decay_ms_default_get(), + arena_muzzy_decay_ms_default_get())) { goto label_error; } /* Initialize bins. */ - uintptr_t bin_addr = (uintptr_t)arena + sizeof(arena_t); atomic_store_u(&arena->binshard_next, 0, ATOMIC_RELEASE); - for (i = 0; i < SC_NBINS; i++) { - unsigned nshards = bin_infos[i].n_shards; - arena->bins[i].bin_shards = (bin_t *)bin_addr; - bin_addr += nshards * sizeof(bin_t); - for (unsigned j = 0; j < nshards; j++) { - bool err = bin_init(&arena->bins[i].bin_shards[j]); - if (err) { - goto label_error; - } + for (i = 0; i < nbins_total; i++) { + bool err = bin_init(&arena->bins[i]); + if (err) { + goto label_error; } } - assert(bin_addr == (uintptr_t)arena + arena_size); arena->base = base; /* Set arena before creating background threads. */ arena_set(ind, arena); + arena->ind = ind; - nstime_init(&arena->create_time, 0); - nstime_update(&arena->create_time); + nstime_init_update(&arena->create_time); + + /* + * We turn on the HPA if set to. There are two exceptions: + * - Custom extent hooks (we should only return memory allocated from + * them in that case). + * - Arena 0 initialization. In this case, we're mid-bootstrapping, and + * so arena_hpa_global is not yet initialized. + */ + if (opt_hpa && ehooks_are_default(base_ehooks_get(base)) && ind != 0) { + hpa_shard_opts_t hpa_shard_opts = opt_hpa_opts; + hpa_shard_opts.deferral_allowed = background_thread_enabled(); + if (pa_shard_enable_hpa(tsdn, &arena->pa_shard, + &hpa_shard_opts, &opt_hpa_sec_opts)) { + goto label_error; + } + } /* We don't support reentrancy for arena 0 bootstrapping. */ if (ind != 0) { @@ -2129,10 +1729,12 @@ * expected for huge allocations. */ if (arena_dirty_decay_ms_default_get() > 0) { - arena_dirty_decay_ms_set(tsd_tsdn(tsd), huge_arena, 0); + arena_decay_ms_set(tsd_tsdn(tsd), huge_arena, + extent_state_dirty, 0); } if (arena_muzzy_decay_ms_default_get() > 0) { - arena_muzzy_decay_ms_set(tsd_tsdn(tsd), huge_arena, 0); + arena_decay_ms_set(tsd_tsdn(tsd), huge_arena, + extent_state_muzzy, 0); } } @@ -2167,8 +1769,8 @@ return (arena_ind == huge_arena_ind); } -void -arena_boot(sc_data_t *sc_data) { +bool +arena_boot(sc_data_t *sc_data, base_t *base, bool hpa) { arena_dirty_decay_ms_default_set(opt_dirty_decay_ms); arena_muzzy_decay_ms_default_set(opt_muzzy_decay_ms); for (unsigned i = 0; i < SC_NBINS; i++) { @@ -2176,12 +1778,20 @@ div_init(&arena_binind_div_info[i], (1U << sc->lg_base) + (sc->ndelta << sc->lg_delta)); } + + uint32_t cur_offset = (uint32_t)offsetof(arena_t, bins); + for (szind_t i = 0; i < SC_NBINS; i++) { + arena_bin_offsets[i] = cur_offset; + nbins_total += bin_infos[i].n_shards; + cur_offset += (uint32_t)(bin_infos[i].n_shards * sizeof(bin_t)); + } + return pa_central_init(&arena_pa_central_global, base, hpa, + &hpa_hooks_default); } void arena_prefork0(tsdn_t *tsdn, arena_t *arena) { - malloc_mutex_prefork(tsdn, &arena->decay_dirty.mtx); - malloc_mutex_prefork(tsdn, &arena->decay_muzzy.mtx); + pa_shard_prefork0(tsdn, &arena->pa_shard); } void @@ -2193,59 +1803,50 @@ void arena_prefork2(tsdn_t *tsdn, arena_t *arena) { - malloc_mutex_prefork(tsdn, &arena->extent_grow_mtx); + pa_shard_prefork2(tsdn, &arena->pa_shard); } void arena_prefork3(tsdn_t *tsdn, arena_t *arena) { - extents_prefork(tsdn, &arena->extents_dirty); - extents_prefork(tsdn, &arena->extents_muzzy); - extents_prefork(tsdn, &arena->extents_retained); + pa_shard_prefork3(tsdn, &arena->pa_shard); } void arena_prefork4(tsdn_t *tsdn, arena_t *arena) { - malloc_mutex_prefork(tsdn, &arena->extent_avail_mtx); + pa_shard_prefork4(tsdn, &arena->pa_shard); } void arena_prefork5(tsdn_t *tsdn, arena_t *arena) { - base_prefork(tsdn, arena->base); + pa_shard_prefork5(tsdn, &arena->pa_shard); } void arena_prefork6(tsdn_t *tsdn, arena_t *arena) { - malloc_mutex_prefork(tsdn, &arena->large_mtx); + base_prefork(tsdn, arena->base); } void arena_prefork7(tsdn_t *tsdn, arena_t *arena) { - for (unsigned i = 0; i < SC_NBINS; i++) { - for (unsigned j = 0; j < bin_infos[i].n_shards; j++) { - bin_prefork(tsdn, &arena->bins[i].bin_shards[j]); - } + malloc_mutex_prefork(tsdn, &arena->large_mtx); +} + +void +arena_prefork8(tsdn_t *tsdn, arena_t *arena) { + for (unsigned i = 0; i < nbins_total; i++) { + bin_prefork(tsdn, &arena->bins[i]); } } void arena_postfork_parent(tsdn_t *tsdn, arena_t *arena) { - unsigned i; - - for (i = 0; i < SC_NBINS; i++) { - for (unsigned j = 0; j < bin_infos[i].n_shards; j++) { - bin_postfork_parent(tsdn, - &arena->bins[i].bin_shards[j]); - } + for (unsigned i = 0; i < nbins_total; i++) { + bin_postfork_parent(tsdn, &arena->bins[i]); } + malloc_mutex_postfork_parent(tsdn, &arena->large_mtx); base_postfork_parent(tsdn, arena->base); - malloc_mutex_postfork_parent(tsdn, &arena->extent_avail_mtx); - extents_postfork_parent(tsdn, &arena->extents_dirty); - extents_postfork_parent(tsdn, &arena->extents_muzzy); - extents_postfork_parent(tsdn, &arena->extents_retained); - malloc_mutex_postfork_parent(tsdn, &arena->extent_grow_mtx); - malloc_mutex_postfork_parent(tsdn, &arena->decay_dirty.mtx); - malloc_mutex_postfork_parent(tsdn, &arena->decay_muzzy.mtx); + pa_shard_postfork_parent(tsdn, &arena->pa_shard); if (config_stats) { malloc_mutex_postfork_parent(tsdn, &arena->tcache_ql_mtx); } @@ -2253,8 +1854,6 @@ void arena_postfork_child(tsdn_t *tsdn, arena_t *arena) { - unsigned i; - atomic_store_u(&arena->nthreads[0], 0, ATOMIC_RELAXED); atomic_store_u(&arena->nthreads[1], 0, ATOMIC_RELAXED); if (tsd_arena_get(tsdn_tsd(tsdn)) == arena) { @@ -2266,32 +1865,26 @@ if (config_stats) { ql_new(&arena->tcache_ql); ql_new(&arena->cache_bin_array_descriptor_ql); - tcache_t *tcache = tcache_get(tsdn_tsd(tsdn)); - if (tcache != NULL && tcache->arena == arena) { - ql_elm_new(tcache, link); - ql_tail_insert(&arena->tcache_ql, tcache, link); + tcache_slow_t *tcache_slow = tcache_slow_get(tsdn_tsd(tsdn)); + if (tcache_slow != NULL && tcache_slow->arena == arena) { + tcache_t *tcache = tcache_slow->tcache; + ql_elm_new(tcache_slow, link); + ql_tail_insert(&arena->tcache_ql, tcache_slow, link); cache_bin_array_descriptor_init( - &tcache->cache_bin_array_descriptor, - tcache->bins_small, tcache->bins_large); + &tcache_slow->cache_bin_array_descriptor, + tcache->bins); ql_tail_insert(&arena->cache_bin_array_descriptor_ql, - &tcache->cache_bin_array_descriptor, link); + &tcache_slow->cache_bin_array_descriptor, link); } } - for (i = 0; i < SC_NBINS; i++) { - for (unsigned j = 0; j < bin_infos[i].n_shards; j++) { - bin_postfork_child(tsdn, &arena->bins[i].bin_shards[j]); - } + for (unsigned i = 0; i < nbins_total; i++) { + bin_postfork_child(tsdn, &arena->bins[i]); } + malloc_mutex_postfork_child(tsdn, &arena->large_mtx); base_postfork_child(tsdn, arena->base); - malloc_mutex_postfork_child(tsdn, &arena->extent_avail_mtx); - extents_postfork_child(tsdn, &arena->extents_dirty); - extents_postfork_child(tsdn, &arena->extents_muzzy); - extents_postfork_child(tsdn, &arena->extents_retained); - malloc_mutex_postfork_child(tsdn, &arena->extent_grow_mtx); - malloc_mutex_postfork_child(tsdn, &arena->decay_dirty.mtx); - malloc_mutex_postfork_child(tsdn, &arena->decay_muzzy.mtx); + pa_shard_postfork_child(tsdn, &arena->pa_shard); if (config_stats) { malloc_mutex_postfork_child(tsdn, &arena->tcache_ql_mtx); } diff --git a/contrib/jemalloc/src/background_thread.c b/contrib/jemalloc/src/background_thread.c --- a/contrib/jemalloc/src/background_thread.c +++ b/contrib/jemalloc/src/background_thread.c @@ -1,4 +1,3 @@ -#define JEMALLOC_BACKGROUND_THREAD_C_ #include "jemalloc/internal/jemalloc_preamble.h" #include "jemalloc/internal/jemalloc_internal_includes.h" @@ -54,8 +53,9 @@ bool background_thread_create(tsd_t *tsd, unsigned arena_ind) NOT_REACHED bool background_threads_enable(tsd_t *tsd) NOT_REACHED bool background_threads_disable(tsd_t *tsd) NOT_REACHED -void background_thread_interval_check(tsdn_t *tsdn, arena_t *arena, - arena_decay_t *decay, size_t npages_new) NOT_REACHED +bool background_thread_is_started(background_thread_info_t *info) NOT_REACHED +void background_thread_wakeup_early(background_thread_info_t *info, + nstime_t *remaining_sleep) NOT_REACHED void background_thread_prefork0(tsdn_t *tsdn) NOT_REACHED void background_thread_prefork1(tsdn_t *tsdn) NOT_REACHED void background_thread_postfork_parent(tsdn_t *tsdn) NOT_REACHED @@ -74,7 +74,7 @@ info->npages_to_purge_new = 0; if (config_stats) { info->tot_n_runs = 0; - nstime_init(&info->tot_sleep_time, 0); + nstime_init_zero(&info->tot_sleep_time); } } @@ -82,136 +82,40 @@ set_current_thread_affinity(int cpu) { #if defined(JEMALLOC_HAVE_SCHED_SETAFFINITY) cpu_set_t cpuset; +#else +# ifndef __NetBSD__ + cpuset_t cpuset; +# else + cpuset_t *cpuset; +# endif +#endif + +#ifndef __NetBSD__ CPU_ZERO(&cpuset); CPU_SET(cpu, &cpuset); - int ret = sched_setaffinity(0, sizeof(cpu_set_t), &cpuset); +#else + cpuset = cpuset_create(); +#endif - return (ret != 0); +#if defined(JEMALLOC_HAVE_SCHED_SETAFFINITY) + return (sched_setaffinity(0, sizeof(cpu_set_t), &cpuset) != 0); #else - return false; +# ifndef __NetBSD__ + int ret = pthread_setaffinity_np(pthread_self(), sizeof(cpuset_t), + &cpuset); +# else + int ret = pthread_setaffinity_np(pthread_self(), cpuset_size(cpuset), + cpuset); + cpuset_destroy(cpuset); +# endif + return ret != 0; #endif } -/* Threshold for determining when to wake up the background thread. */ -#define BACKGROUND_THREAD_NPAGES_THRESHOLD UINT64_C(1024) #define BILLION UINT64_C(1000000000) /* Minimal sleep interval 100 ms. */ #define BACKGROUND_THREAD_MIN_INTERVAL_NS (BILLION / 10) -static inline size_t -decay_npurge_after_interval(arena_decay_t *decay, size_t interval) { - size_t i; - uint64_t sum = 0; - for (i = 0; i < interval; i++) { - sum += decay->backlog[i] * h_steps[i]; - } - for (; i < SMOOTHSTEP_NSTEPS; i++) { - sum += decay->backlog[i] * (h_steps[i] - h_steps[i - interval]); - } - - return (size_t)(sum >> SMOOTHSTEP_BFP); -} - -static uint64_t -arena_decay_compute_purge_interval_impl(tsdn_t *tsdn, arena_decay_t *decay, - extents_t *extents) { - if (malloc_mutex_trylock(tsdn, &decay->mtx)) { - /* Use minimal interval if decay is contended. */ - return BACKGROUND_THREAD_MIN_INTERVAL_NS; - } - - uint64_t interval; - ssize_t decay_time = atomic_load_zd(&decay->time_ms, ATOMIC_RELAXED); - if (decay_time <= 0) { - /* Purging is eagerly done or disabled currently. */ - interval = BACKGROUND_THREAD_INDEFINITE_SLEEP; - goto label_done; - } - - uint64_t decay_interval_ns = nstime_ns(&decay->interval); - assert(decay_interval_ns > 0); - size_t npages = extents_npages_get(extents); - if (npages == 0) { - unsigned i; - for (i = 0; i < SMOOTHSTEP_NSTEPS; i++) { - if (decay->backlog[i] > 0) { - break; - } - } - if (i == SMOOTHSTEP_NSTEPS) { - /* No dirty pages recorded. Sleep indefinitely. */ - interval = BACKGROUND_THREAD_INDEFINITE_SLEEP; - goto label_done; - } - } - if (npages <= BACKGROUND_THREAD_NPAGES_THRESHOLD) { - /* Use max interval. */ - interval = decay_interval_ns * SMOOTHSTEP_NSTEPS; - goto label_done; - } - - size_t lb = BACKGROUND_THREAD_MIN_INTERVAL_NS / decay_interval_ns; - size_t ub = SMOOTHSTEP_NSTEPS; - /* Minimal 2 intervals to ensure reaching next epoch deadline. */ - lb = (lb < 2) ? 2 : lb; - if ((decay_interval_ns * ub <= BACKGROUND_THREAD_MIN_INTERVAL_NS) || - (lb + 2 > ub)) { - interval = BACKGROUND_THREAD_MIN_INTERVAL_NS; - goto label_done; - } - - assert(lb + 2 <= ub); - size_t npurge_lb, npurge_ub; - npurge_lb = decay_npurge_after_interval(decay, lb); - if (npurge_lb > BACKGROUND_THREAD_NPAGES_THRESHOLD) { - interval = decay_interval_ns * lb; - goto label_done; - } - npurge_ub = decay_npurge_after_interval(decay, ub); - if (npurge_ub < BACKGROUND_THREAD_NPAGES_THRESHOLD) { - interval = decay_interval_ns * ub; - goto label_done; - } - - unsigned n_search = 0; - size_t target, npurge; - while ((npurge_lb + BACKGROUND_THREAD_NPAGES_THRESHOLD < npurge_ub) - && (lb + 2 < ub)) { - target = (lb + ub) / 2; - npurge = decay_npurge_after_interval(decay, target); - if (npurge > BACKGROUND_THREAD_NPAGES_THRESHOLD) { - ub = target; - npurge_ub = npurge; - } else { - lb = target; - npurge_lb = npurge; - } - assert(n_search++ < lg_floor(SMOOTHSTEP_NSTEPS) + 1); - } - interval = decay_interval_ns * (ub + lb) / 2; -label_done: - interval = (interval < BACKGROUND_THREAD_MIN_INTERVAL_NS) ? - BACKGROUND_THREAD_MIN_INTERVAL_NS : interval; - malloc_mutex_unlock(tsdn, &decay->mtx); - - return interval; -} - -/* Compute purge interval for background threads. */ -static uint64_t -arena_decay_compute_purge_interval(tsdn_t *tsdn, arena_t *arena) { - uint64_t i1, i2; - i1 = arena_decay_compute_purge_interval_impl(tsdn, &arena->decay_dirty, - &arena->extents_dirty); - if (i1 == BACKGROUND_THREAD_MIN_INTERVAL_NS) { - return i1; - } - i2 = arena_decay_compute_purge_interval_impl(tsdn, &arena->decay_muzzy, - &arena->extents_muzzy); - - return i1 < i2 ? i1 : i2; -} - static void background_thread_sleep(tsdn_t *tsdn, background_thread_info_t *info, uint64_t interval) { @@ -228,7 +132,8 @@ int ret; if (interval == BACKGROUND_THREAD_INDEFINITE_SLEEP) { - assert(background_thread_indefinite_sleep(info)); + background_thread_wakeup_time_set(tsdn, info, + BACKGROUND_THREAD_INDEFINITE_SLEEP); ret = pthread_cond_wait(&info->cond, &info->mtx.lock); assert(ret == 0); } else { @@ -236,8 +141,7 @@ interval <= BACKGROUND_THREAD_INDEFINITE_SLEEP); /* We need malloc clock (can be different from tv). */ nstime_t next_wakeup; - nstime_init(&next_wakeup, 0); - nstime_update(&next_wakeup); + nstime_init_update(&next_wakeup); nstime_iadd(&next_wakeup, interval); assert(nstime_ns(&next_wakeup) < BACKGROUND_THREAD_INDEFINITE_SLEEP); @@ -254,8 +158,6 @@ assert(!background_thread_indefinite_sleep(info)); ret = pthread_cond_timedwait(&info->cond, &info->mtx.lock, &ts); assert(ret == ETIMEDOUT || ret == 0); - background_thread_wakeup_time_set(tsdn, info, - BACKGROUND_THREAD_INDEFINITE_SLEEP); } if (config_stats) { gettimeofday(&tv, NULL); @@ -283,28 +185,48 @@ } static inline void -background_work_sleep_once(tsdn_t *tsdn, background_thread_info_t *info, unsigned ind) { - uint64_t min_interval = BACKGROUND_THREAD_INDEFINITE_SLEEP; +background_work_sleep_once(tsdn_t *tsdn, background_thread_info_t *info, + unsigned ind) { + uint64_t ns_until_deferred = BACKGROUND_THREAD_DEFERRED_MAX; unsigned narenas = narenas_total_get(); + bool slept_indefinitely = background_thread_indefinite_sleep(info); for (unsigned i = ind; i < narenas; i += max_background_threads) { arena_t *arena = arena_get(tsdn, i, false); if (!arena) { continue; } - arena_decay(tsdn, arena, true, false); - if (min_interval == BACKGROUND_THREAD_MIN_INTERVAL_NS) { + /* + * If thread was woken up from the indefinite sleep, don't + * do the work instantly, but rather check when the deferred + * work that caused this thread to wake up is scheduled for. + */ + if (!slept_indefinitely) { + arena_do_deferred_work(tsdn, arena); + } + if (ns_until_deferred <= BACKGROUND_THREAD_MIN_INTERVAL_NS) { /* Min interval will be used. */ continue; } - uint64_t interval = arena_decay_compute_purge_interval(tsdn, - arena); - assert(interval >= BACKGROUND_THREAD_MIN_INTERVAL_NS); - if (min_interval > interval) { - min_interval = interval; + uint64_t ns_arena_deferred = pa_shard_time_until_deferred_work( + tsdn, &arena->pa_shard); + if (ns_arena_deferred < ns_until_deferred) { + ns_until_deferred = ns_arena_deferred; } } - background_thread_sleep(tsdn, info, min_interval); + + uint64_t sleep_ns; + if (ns_until_deferred == BACKGROUND_THREAD_DEFERRED_MAX) { + sleep_ns = BACKGROUND_THREAD_INDEFINITE_SLEEP; + } else { + sleep_ns = + (ns_until_deferred < BACKGROUND_THREAD_MIN_INTERVAL_NS) + ? BACKGROUND_THREAD_MIN_INTERVAL_NS + : ns_until_deferred; + + } + + background_thread_sleep(tsdn, info, sleep_ns); } static bool @@ -508,7 +430,7 @@ assert(thread_ind < max_background_threads); #ifdef JEMALLOC_HAVE_PTHREAD_SETNAME_NP pthread_setname_np(pthread_self(), "jemalloc_bg_thd"); -#elif defined(__FreeBSD__) +#elif defined(__FreeBSD__) || defined(__DragonFly__) pthread_set_name_np(pthread_self(), "jemalloc_bg_thd"); #endif if (opt_percpu_arena != percpu_arena_disabled) { @@ -608,16 +530,16 @@ malloc_mutex_assert_owner(tsd_tsdn(tsd), &background_thread_lock); VARIABLE_ARRAY(bool, marked, max_background_threads); - unsigned i, nmarked; - for (i = 0; i < max_background_threads; i++) { + unsigned nmarked; + for (unsigned i = 0; i < max_background_threads; i++) { marked[i] = false; } nmarked = 0; /* Thread 0 is required and created at the end. */ marked[0] = true; /* Mark the threads we need to create for thread 0. */ - unsigned n = narenas_total_get(); - for (i = 1; i < n; i++) { + unsigned narenas = narenas_total_get(); + for (unsigned i = 1; i < narenas; i++) { if (marked[i % max_background_threads] || arena_get(tsd_tsdn(tsd), i, false) == NULL) { continue; @@ -634,7 +556,18 @@ } } - return background_thread_create_locked(tsd, 0); + bool err = background_thread_create_locked(tsd, 0); + if (err) { + return true; + } + for (unsigned i = 0; i < narenas; i++) { + arena_t *arena = arena_get(tsd_tsdn(tsd), i, false); + if (arena != NULL) { + pa_shard_set_deferral_allowed(tsd_tsdn(tsd), + &arena->pa_shard, true); + } + } + return false; } bool @@ -648,92 +581,36 @@ return true; } assert(n_background_threads == 0); + unsigned narenas = narenas_total_get(); + for (unsigned i = 0; i < narenas; i++) { + arena_t *arena = arena_get(tsd_tsdn(tsd), i, false); + if (arena != NULL) { + pa_shard_set_deferral_allowed(tsd_tsdn(tsd), + &arena->pa_shard, false); + } + } return false; } -/* Check if we need to signal the background thread early. */ +bool +background_thread_is_started(background_thread_info_t *info) { + return info->state == background_thread_started; +} + void -background_thread_interval_check(tsdn_t *tsdn, arena_t *arena, - arena_decay_t *decay, size_t npages_new) { - background_thread_info_t *info = arena_background_thread_info_get( - arena); - if (malloc_mutex_trylock(tsdn, &info->mtx)) { - /* - * Background thread may hold the mutex for a long period of - * time. We'd like to avoid the variance on application - * threads. So keep this non-blocking, and leave the work to a - * future epoch. - */ +background_thread_wakeup_early(background_thread_info_t *info, + nstime_t *remaining_sleep) { + /* + * This is an optimization to increase batching. At this point + * we know that background thread wakes up soon, so the time to cache + * the just freed memory is bounded and low. + */ + if (remaining_sleep != NULL && nstime_ns(remaining_sleep) < + BACKGROUND_THREAD_MIN_INTERVAL_NS) { return; } - - if (info->state != background_thread_started) { - goto label_done; - } - if (malloc_mutex_trylock(tsdn, &decay->mtx)) { - goto label_done; - } - - ssize_t decay_time = atomic_load_zd(&decay->time_ms, ATOMIC_RELAXED); - if (decay_time <= 0) { - /* Purging is eagerly done or disabled currently. */ - goto label_done_unlock2; - } - uint64_t decay_interval_ns = nstime_ns(&decay->interval); - assert(decay_interval_ns > 0); - - nstime_t diff; - nstime_init(&diff, background_thread_wakeup_time_get(info)); - if (nstime_compare(&diff, &decay->epoch) <= 0) { - goto label_done_unlock2; - } - nstime_subtract(&diff, &decay->epoch); - if (nstime_ns(&diff) < BACKGROUND_THREAD_MIN_INTERVAL_NS) { - goto label_done_unlock2; - } - - if (npages_new > 0) { - size_t n_epoch = (size_t)(nstime_ns(&diff) / decay_interval_ns); - /* - * Compute how many new pages we would need to purge by the next - * wakeup, which is used to determine if we should signal the - * background thread. - */ - uint64_t npurge_new; - if (n_epoch >= SMOOTHSTEP_NSTEPS) { - npurge_new = npages_new; - } else { - uint64_t h_steps_max = h_steps[SMOOTHSTEP_NSTEPS - 1]; - assert(h_steps_max >= - h_steps[SMOOTHSTEP_NSTEPS - 1 - n_epoch]); - npurge_new = npages_new * (h_steps_max - - h_steps[SMOOTHSTEP_NSTEPS - 1 - n_epoch]); - npurge_new >>= SMOOTHSTEP_BFP; - } - info->npages_to_purge_new += npurge_new; - } - - bool should_signal; - if (info->npages_to_purge_new > BACKGROUND_THREAD_NPAGES_THRESHOLD) { - should_signal = true; - } else if (unlikely(background_thread_indefinite_sleep(info)) && - (extents_npages_get(&arena->extents_dirty) > 0 || - extents_npages_get(&arena->extents_muzzy) > 0 || - info->npages_to_purge_new > 0)) { - should_signal = true; - } else { - should_signal = false; - } - - if (should_signal) { - info->npages_to_purge_new = 0; - pthread_cond_signal(&info->cond); - } -label_done_unlock2: - malloc_mutex_unlock(tsdn, &decay->mtx); -label_done: - malloc_mutex_unlock(tsdn, &info->mtx); + pthread_cond_signal(&info->cond); } void @@ -794,9 +671,11 @@ return true; } - stats->num_threads = n_background_threads; + nstime_init_zero(&stats->run_interval); + memset(&stats->max_counter_per_bg_thd, 0, sizeof(mutex_prof_data_t)); + uint64_t num_runs = 0; - nstime_init(&stats->run_interval, 0); + stats->num_threads = n_background_threads; for (unsigned i = 0; i < max_background_threads; i++) { background_thread_info_t *info = &background_thread_info[i]; if (malloc_mutex_trylock(tsdn, &info->mtx)) { @@ -809,6 +688,8 @@ if (info->state != background_thread_stopped) { num_runs += info->tot_n_runs; nstime_add(&stats->run_interval, &info->tot_sleep_time); + malloc_mutex_prof_max_update(tsdn, + &stats->max_counter_per_bg_thd, &info->mtx); } malloc_mutex_unlock(tsdn, &info->mtx); } @@ -892,7 +773,7 @@ } bool -background_thread_boot1(tsdn_t *tsdn) { +background_thread_boot1(tsdn_t *tsdn, base_t *base) { #ifdef JEMALLOC_BACKGROUND_THREAD assert(have_background_thread); assert(narenas_total_get() > 0); @@ -911,7 +792,7 @@ } background_thread_info = (background_thread_info_t *)base_alloc(tsdn, - b0get(), opt_max_background_threads * + base, opt_max_background_threads * sizeof(background_thread_info_t), CACHELINE); if (background_thread_info == NULL) { return true; diff --git a/contrib/jemalloc/src/base.c b/contrib/jemalloc/src/base.c --- a/contrib/jemalloc/src/base.c +++ b/contrib/jemalloc/src/base.c @@ -1,4 +1,3 @@ -#define JEMALLOC_BASE_C_ #include "jemalloc/internal/jemalloc_preamble.h" #include "jemalloc/internal/jemalloc_internal_includes.h" @@ -7,6 +6,15 @@ #include "jemalloc/internal/mutex.h" #include "jemalloc/internal/sz.h" +/* + * In auto mode, arenas switch to huge pages for the base allocator on the + * second base block. a0 switches to thp on the 5th block (after 20 megabytes + * of metadata), since more metadata (e.g. rtree nodes) come from a0's base. + */ + +#define BASE_AUTO_THP_THRESHOLD 2 +#define BASE_AUTO_THP_THRESHOLD_A0 5 + /******************************************************************************/ /* Data. */ @@ -29,7 +37,7 @@ } static void * -base_map(tsdn_t *tsdn, extent_hooks_t *extent_hooks, unsigned ind, size_t size) { +base_map(tsdn_t *tsdn, ehooks_t *ehooks, unsigned ind, size_t size) { void *addr; bool zero = true; bool commit = true; @@ -37,22 +45,21 @@ /* Use huge page sizes and alignment regardless of opt_metadata_thp. */ assert(size == HUGEPAGE_CEILING(size)); size_t alignment = HUGEPAGE; - if (extent_hooks == &extent_hooks_default) { + if (ehooks_are_default(ehooks)) { addr = extent_alloc_mmap(NULL, size, alignment, &zero, &commit); + if (have_madvise_huge && addr) { + pages_set_thp_state(addr, size); + } } else { - /* No arena context as we are creating new arenas. */ - tsd_t *tsd = tsdn_null(tsdn) ? tsd_fetch() : tsdn_tsd(tsdn); - pre_reentrancy(tsd, NULL); - addr = extent_hooks->alloc(extent_hooks, NULL, size, alignment, - &zero, &commit, ind); - post_reentrancy(tsd); + addr = ehooks_alloc(tsdn, ehooks, NULL, size, alignment, &zero, + &commit); } return addr; } static void -base_unmap(tsdn_t *tsdn, extent_hooks_t *extent_hooks, unsigned ind, void *addr, +base_unmap(tsdn_t *tsdn, ehooks_t *ehooks, unsigned ind, void *addr, size_t size) { /* * Cascade through dalloc, decommit, purge_forced, and purge_lazy, @@ -64,7 +71,7 @@ * may in fact want the end state of all associated virtual memory to be * in some consistent-but-allocated state. */ - if (extent_hooks == &extent_hooks_default) { + if (ehooks_are_default(ehooks)) { if (!extent_dalloc_mmap(addr, size)) { goto label_done; } @@ -80,31 +87,19 @@ /* Nothing worked. This should never happen. */ not_reached(); } else { - tsd_t *tsd = tsdn_null(tsdn) ? tsd_fetch() : tsdn_tsd(tsdn); - pre_reentrancy(tsd, NULL); - if (extent_hooks->dalloc != NULL && - !extent_hooks->dalloc(extent_hooks, addr, size, true, - ind)) { - goto label_post_reentrancy; + if (!ehooks_dalloc(tsdn, ehooks, addr, size, true)) { + goto label_done; } - if (extent_hooks->decommit != NULL && - !extent_hooks->decommit(extent_hooks, addr, size, 0, size, - ind)) { - goto label_post_reentrancy; + if (!ehooks_decommit(tsdn, ehooks, addr, size, 0, size)) { + goto label_done; } - if (extent_hooks->purge_forced != NULL && - !extent_hooks->purge_forced(extent_hooks, addr, size, 0, - size, ind)) { - goto label_post_reentrancy; + if (!ehooks_purge_forced(tsdn, ehooks, addr, size, 0, size)) { + goto label_done; } - if (extent_hooks->purge_lazy != NULL && - !extent_hooks->purge_lazy(extent_hooks, addr, size, 0, size, - ind)) { - goto label_post_reentrancy; + if (!ehooks_purge_lazy(tsdn, ehooks, addr, size, 0, size)) { + goto label_done; } /* Nothing worked. That's the application's problem. */ - label_post_reentrancy: - post_reentrancy(tsd); } label_done: if (metadata_thp_madvise()) { @@ -116,14 +111,14 @@ } static void -base_extent_init(size_t *extent_sn_next, extent_t *extent, void *addr, +base_edata_init(size_t *extent_sn_next, edata_t *edata, void *addr, size_t size) { size_t sn; sn = *extent_sn_next; (*extent_sn_next)++; - extent_binit(extent, addr, size, sn); + edata_binit(edata, addr, size, sn); } static size_t @@ -169,7 +164,7 @@ pages_huge(block, block->size); if (config_stats) { base->n_thp += HUGEPAGE_CEILING(block->size - - extent_bsize_get(&block->extent)) >> LG_HUGEPAGE; + edata_bsize_get(&block->edata)) >> LG_HUGEPAGE; } block = block->next; assert(block == NULL || (base_ind_get(base) == 0)); @@ -177,34 +172,34 @@ } static void * -base_extent_bump_alloc_helper(extent_t *extent, size_t *gap_size, size_t size, +base_extent_bump_alloc_helper(edata_t *edata, size_t *gap_size, size_t size, size_t alignment) { void *ret; assert(alignment == ALIGNMENT_CEILING(alignment, QUANTUM)); assert(size == ALIGNMENT_CEILING(size, alignment)); - *gap_size = ALIGNMENT_CEILING((uintptr_t)extent_addr_get(extent), - alignment) - (uintptr_t)extent_addr_get(extent); - ret = (void *)((uintptr_t)extent_addr_get(extent) + *gap_size); - assert(extent_bsize_get(extent) >= *gap_size + size); - extent_binit(extent, (void *)((uintptr_t)extent_addr_get(extent) + - *gap_size + size), extent_bsize_get(extent) - *gap_size - size, - extent_sn_get(extent)); + *gap_size = ALIGNMENT_CEILING((uintptr_t)edata_addr_get(edata), + alignment) - (uintptr_t)edata_addr_get(edata); + ret = (void *)((uintptr_t)edata_addr_get(edata) + *gap_size); + assert(edata_bsize_get(edata) >= *gap_size + size); + edata_binit(edata, (void *)((uintptr_t)edata_addr_get(edata) + + *gap_size + size), edata_bsize_get(edata) - *gap_size - size, + edata_sn_get(edata)); return ret; } static void -base_extent_bump_alloc_post(base_t *base, extent_t *extent, size_t gap_size, +base_extent_bump_alloc_post(base_t *base, edata_t *edata, size_t gap_size, void *addr, size_t size) { - if (extent_bsize_get(extent) > 0) { + if (edata_bsize_get(edata) > 0) { /* * Compute the index for the largest size class that does not * exceed extent's size. */ szind_t index_floor = - sz_size2index(extent_bsize_get(extent) + 1) - 1; - extent_heap_insert(&base->avail[index_floor], extent); + sz_size2index(edata_bsize_get(edata) + 1) - 1; + edata_heap_insert(&base->avail[index_floor], edata); } if (config_stats) { @@ -229,13 +224,13 @@ } static void * -base_extent_bump_alloc(base_t *base, extent_t *extent, size_t size, +base_extent_bump_alloc(base_t *base, edata_t *edata, size_t size, size_t alignment) { void *ret; size_t gap_size; - ret = base_extent_bump_alloc_helper(extent, &gap_size, size, alignment); - base_extent_bump_alloc_post(base, extent, gap_size, ret, size); + ret = base_extent_bump_alloc_helper(edata, &gap_size, size, alignment); + base_extent_bump_alloc_post(base, edata, gap_size, ret, size); return ret; } @@ -245,8 +240,8 @@ * On success a pointer to the initialized base_block_t header is returned. */ static base_block_t * -base_block_alloc(tsdn_t *tsdn, base_t *base, extent_hooks_t *extent_hooks, - unsigned ind, pszind_t *pind_last, size_t *extent_sn_next, size_t size, +base_block_alloc(tsdn_t *tsdn, base_t *base, ehooks_t *ehooks, unsigned ind, + pszind_t *pind_last, size_t *extent_sn_next, size_t size, size_t alignment) { alignment = ALIGNMENT_CEILING(alignment, QUANTUM); size_t usize = ALIGNMENT_CEILING(size, alignment); @@ -267,7 +262,7 @@ size_t next_block_size = HUGEPAGE_CEILING(sz_pind2sz(pind_next)); size_t block_size = (min_block_size > next_block_size) ? min_block_size : next_block_size; - base_block_t *block = (base_block_t *)base_map(tsdn, extent_hooks, ind, + base_block_t *block = (base_block_t *)base_map(tsdn, ehooks, ind, block_size); if (block == NULL) { return NULL; @@ -295,7 +290,7 @@ block->size = block_size; block->next = NULL; assert(block_size >= header_size); - base_extent_init(extent_sn_next, &block->extent, + base_edata_init(extent_sn_next, &block->edata, (void *)((uintptr_t)block + header_size), block_size - header_size); return block; } @@ -304,17 +299,17 @@ * Allocate an extent that is at least as large as specified size, with * specified alignment. */ -static extent_t * +static edata_t * base_extent_alloc(tsdn_t *tsdn, base_t *base, size_t size, size_t alignment) { malloc_mutex_assert_owner(tsdn, &base->mtx); - extent_hooks_t *extent_hooks = base_extent_hooks_get(base); + ehooks_t *ehooks = base_ehooks_get_for_metadata(base); /* * Drop mutex during base_block_alloc(), because an extent hook will be * called. */ malloc_mutex_unlock(tsdn, &base->mtx); - base_block_t *block = base_block_alloc(tsdn, base, extent_hooks, + base_block_t *block = base_block_alloc(tsdn, base, ehooks, base_ind_get(base), &base->pind_last, &base->extent_sn_next, size, alignment); malloc_mutex_lock(tsdn, &base->mtx); @@ -338,7 +333,7 @@ assert(base->resident <= base->mapped); assert(base->n_thp << LG_HUGEPAGE <= base->mapped); } - return &block->extent; + return &block->edata; } base_t * @@ -347,10 +342,22 @@ } base_t * -base_new(tsdn_t *tsdn, unsigned ind, extent_hooks_t *extent_hooks) { +base_new(tsdn_t *tsdn, unsigned ind, const extent_hooks_t *extent_hooks, + bool metadata_use_hooks) { pszind_t pind_last = 0; size_t extent_sn_next = 0; - base_block_t *block = base_block_alloc(tsdn, NULL, extent_hooks, ind, + + /* + * The base will contain the ehooks eventually, but it itself is + * allocated using them. So we use some stack ehooks to bootstrap its + * memory, and then initialize the ehooks within the base_t. + */ + ehooks_t fake_ehooks; + ehooks_init(&fake_ehooks, metadata_use_hooks ? + (extent_hooks_t *)extent_hooks : + (extent_hooks_t *)&ehooks_default_extent_hooks, ind); + + base_block_t *block = base_block_alloc(tsdn, NULL, &fake_ehooks, ind, &pind_last, &extent_sn_next, sizeof(base_t), QUANTUM); if (block == NULL) { return NULL; @@ -359,13 +366,15 @@ size_t gap_size; size_t base_alignment = CACHELINE; size_t base_size = ALIGNMENT_CEILING(sizeof(base_t), base_alignment); - base_t *base = (base_t *)base_extent_bump_alloc_helper(&block->extent, + base_t *base = (base_t *)base_extent_bump_alloc_helper(&block->edata, &gap_size, base_size, base_alignment); - base->ind = ind; - atomic_store_p(&base->extent_hooks, extent_hooks, ATOMIC_RELAXED); + ehooks_init(&base->ehooks, (extent_hooks_t *)extent_hooks, ind); + ehooks_init(&base->ehooks_base, metadata_use_hooks ? + (extent_hooks_t *)extent_hooks : + (extent_hooks_t *)&ehooks_default_extent_hooks, ind); if (malloc_mutex_init(&base->mtx, "base", WITNESS_RANK_BASE, malloc_mutex_rank_exclusive)) { - base_unmap(tsdn, extent_hooks, ind, block, block->size); + base_unmap(tsdn, &fake_ehooks, ind, block, block->size); return NULL; } base->pind_last = pind_last; @@ -373,7 +382,7 @@ base->blocks = block; base->auto_thp_switched = false; for (szind_t i = 0; i < SC_NSIZES; i++) { - extent_heap_new(&base->avail[i]); + edata_heap_new(&base->avail[i]); } if (config_stats) { base->allocated = sizeof(base_block_t); @@ -386,7 +395,7 @@ assert(base->resident <= base->mapped); assert(base->n_thp << LG_HUGEPAGE <= base->mapped); } - base_extent_bump_alloc_post(base, &block->extent, gap_size, base, + base_extent_bump_alloc_post(base, &block->edata, gap_size, base, base_size); return base; @@ -394,26 +403,31 @@ void base_delete(tsdn_t *tsdn, base_t *base) { - extent_hooks_t *extent_hooks = base_extent_hooks_get(base); + ehooks_t *ehooks = base_ehooks_get_for_metadata(base); base_block_t *next = base->blocks; do { base_block_t *block = next; next = block->next; - base_unmap(tsdn, extent_hooks, base_ind_get(base), block, + base_unmap(tsdn, ehooks, base_ind_get(base), block, block->size); } while (next != NULL); } -extent_hooks_t * -base_extent_hooks_get(base_t *base) { - return (extent_hooks_t *)atomic_load_p(&base->extent_hooks, - ATOMIC_ACQUIRE); +ehooks_t * +base_ehooks_get(base_t *base) { + return &base->ehooks; +} + +ehooks_t * +base_ehooks_get_for_metadata(base_t *base) { + return &base->ehooks_base; } extent_hooks_t * base_extent_hooks_set(base_t *base, extent_hooks_t *extent_hooks) { - extent_hooks_t *old_extent_hooks = base_extent_hooks_get(base); - atomic_store_p(&base->extent_hooks, extent_hooks, ATOMIC_RELEASE); + extent_hooks_t *old_extent_hooks = + ehooks_get_extent_hooks_ptr(&base->ehooks); + ehooks_init(&base->ehooks, extent_hooks, ehooks_ind_get(&base->ehooks)); return old_extent_hooks; } @@ -424,28 +438,28 @@ size_t usize = ALIGNMENT_CEILING(size, alignment); size_t asize = usize + alignment - QUANTUM; - extent_t *extent = NULL; + edata_t *edata = NULL; malloc_mutex_lock(tsdn, &base->mtx); for (szind_t i = sz_size2index(asize); i < SC_NSIZES; i++) { - extent = extent_heap_remove_first(&base->avail[i]); - if (extent != NULL) { + edata = edata_heap_remove_first(&base->avail[i]); + if (edata != NULL) { /* Use existing space. */ break; } } - if (extent == NULL) { + if (edata == NULL) { /* Try to allocate more space. */ - extent = base_extent_alloc(tsdn, base, usize, alignment); + edata = base_extent_alloc(tsdn, base, usize, alignment); } void *ret; - if (extent == NULL) { + if (edata == NULL) { ret = NULL; goto label_return; } - ret = base_extent_bump_alloc(base, extent, usize, alignment); + ret = base_extent_bump_alloc(base, edata, usize, alignment); if (esn != NULL) { - *esn = extent_sn_get(extent); + *esn = (size_t)edata_sn_get(edata); } label_return: malloc_mutex_unlock(tsdn, &base->mtx); @@ -465,16 +479,16 @@ return base_alloc_impl(tsdn, base, size, alignment, NULL); } -extent_t * -base_alloc_extent(tsdn_t *tsdn, base_t *base) { +edata_t * +base_alloc_edata(tsdn_t *tsdn, base_t *base) { size_t esn; - extent_t *extent = base_alloc_impl(tsdn, base, sizeof(extent_t), - CACHELINE, &esn); - if (extent == NULL) { + edata_t *edata = base_alloc_impl(tsdn, base, sizeof(edata_t), + EDATA_ALIGNMENT, &esn); + if (edata == NULL) { return NULL; } - extent_esn_set(extent, esn); - return extent; + edata_esn_set(edata, esn); + return edata; } void @@ -509,6 +523,7 @@ bool base_boot(tsdn_t *tsdn) { - b0 = base_new(tsdn, 0, (extent_hooks_t *)&extent_hooks_default); + b0 = base_new(tsdn, 0, (extent_hooks_t *)&ehooks_default_extent_hooks, + /* metadata_use_hooks */ true); return (b0 == NULL); } diff --git a/contrib/jemalloc/src/bin.c b/contrib/jemalloc/src/bin.c --- a/contrib/jemalloc/src/bin.c +++ b/contrib/jemalloc/src/bin.c @@ -6,26 +6,6 @@ #include "jemalloc/internal/sc.h" #include "jemalloc/internal/witness.h" -bin_info_t bin_infos[SC_NBINS]; - -static void -bin_infos_init(sc_data_t *sc_data, unsigned bin_shard_sizes[SC_NBINS], - bin_info_t bin_infos[SC_NBINS]) { - for (unsigned i = 0; i < SC_NBINS; i++) { - bin_info_t *bin_info = &bin_infos[i]; - sc_t *sc = &sc_data->sc[i]; - bin_info->reg_size = ((size_t)1U << sc->lg_base) - + ((size_t)sc->ndelta << sc->lg_delta); - bin_info->slab_size = (sc->pgs << LG_PAGE); - bin_info->nregs = - (uint32_t)(bin_info->slab_size / bin_info->reg_size); - bin_info->n_shards = bin_shard_sizes[i]; - bitmap_info_t bitmap_info = BITMAP_INFO_INITIALIZER( - bin_info->nregs); - bin_info->bitmap_info = bitmap_info; - } -} - bool bin_update_shard_size(unsigned bin_shard_sizes[SC_NBINS], size_t start_size, size_t end_size, size_t nshards) { @@ -58,12 +38,6 @@ } } -void -bin_boot(sc_data_t *sc_data, unsigned bin_shard_sizes[SC_NBINS]) { - assert(sc_data->initialized); - bin_infos_init(sc_data, bin_shard_sizes, bin_infos); -} - bool bin_init(bin_t *bin) { if (malloc_mutex_init(&bin->lock, "bin", WITNESS_RANK_BIN, @@ -71,8 +45,8 @@ return true; } bin->slabcur = NULL; - extent_heap_new(&bin->slabs_nonfull); - extent_list_init(&bin->slabs_full); + edata_heap_new(&bin->slabs_nonfull); + edata_list_active_init(&bin->slabs_full); if (config_stats) { memset(&bin->stats, 0, sizeof(bin_stats_t)); } diff --git a/contrib/jemalloc/src/bin_info.c b/contrib/jemalloc/src/bin_info.c new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/src/bin_info.c @@ -0,0 +1,30 @@ +#include "jemalloc/internal/jemalloc_preamble.h" +#include "jemalloc/internal/jemalloc_internal_includes.h" + +#include "jemalloc/internal/bin_info.h" + +bin_info_t bin_infos[SC_NBINS]; + +static void +bin_infos_init(sc_data_t *sc_data, unsigned bin_shard_sizes[SC_NBINS], + bin_info_t infos[SC_NBINS]) { + for (unsigned i = 0; i < SC_NBINS; i++) { + bin_info_t *bin_info = &infos[i]; + sc_t *sc = &sc_data->sc[i]; + bin_info->reg_size = ((size_t)1U << sc->lg_base) + + ((size_t)sc->ndelta << sc->lg_delta); + bin_info->slab_size = (sc->pgs << LG_PAGE); + bin_info->nregs = + (uint32_t)(bin_info->slab_size / bin_info->reg_size); + bin_info->n_shards = bin_shard_sizes[i]; + bitmap_info_t bitmap_info = BITMAP_INFO_INITIALIZER( + bin_info->nregs); + bin_info->bitmap_info = bitmap_info; + } +} + +void +bin_info_boot(sc_data_t *sc_data, unsigned bin_shard_sizes[SC_NBINS]) { + assert(sc_data->initialized); + bin_infos_init(sc_data, bin_shard_sizes, bin_infos); +} diff --git a/contrib/jemalloc/src/bitmap.c b/contrib/jemalloc/src/bitmap.c --- a/contrib/jemalloc/src/bitmap.c +++ b/contrib/jemalloc/src/bitmap.c @@ -1,4 +1,3 @@ -#define JEMALLOC_BITMAP_C_ #include "jemalloc/internal/jemalloc_preamble.h" #include "jemalloc/internal/jemalloc_internal_includes.h" diff --git a/contrib/jemalloc/src/buf_writer.c b/contrib/jemalloc/src/buf_writer.c new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/src/buf_writer.c @@ -0,0 +1,144 @@ +#include "jemalloc/internal/jemalloc_preamble.h" +#include "jemalloc/internal/jemalloc_internal_includes.h" + +#include "jemalloc/internal/buf_writer.h" +#include "jemalloc/internal/malloc_io.h" + +static void * +buf_writer_allocate_internal_buf(tsdn_t *tsdn, size_t buf_len) { +#ifdef JEMALLOC_JET + if (buf_len > SC_LARGE_MAXCLASS) { + return NULL; + } +#else + assert(buf_len <= SC_LARGE_MAXCLASS); +#endif + return iallocztm(tsdn, buf_len, sz_size2index(buf_len), false, NULL, + true, arena_get(tsdn, 0, false), true); +} + +static void +buf_writer_free_internal_buf(tsdn_t *tsdn, void *buf) { + if (buf != NULL) { + idalloctm(tsdn, buf, NULL, NULL, true, true); + } +} + +static void +buf_writer_assert(buf_writer_t *buf_writer) { + assert(buf_writer != NULL); + assert(buf_writer->write_cb != NULL); + if (buf_writer->buf != NULL) { + assert(buf_writer->buf_size > 0); + } else { + assert(buf_writer->buf_size == 0); + assert(buf_writer->internal_buf); + } + assert(buf_writer->buf_end <= buf_writer->buf_size); +} + +bool +buf_writer_init(tsdn_t *tsdn, buf_writer_t *buf_writer, write_cb_t *write_cb, + void *cbopaque, char *buf, size_t buf_len) { + if (write_cb != NULL) { + buf_writer->write_cb = write_cb; + } else { + buf_writer->write_cb = je_malloc_message != NULL ? + je_malloc_message : wrtmessage; + } + buf_writer->cbopaque = cbopaque; + assert(buf_len >= 2); + if (buf != NULL) { + buf_writer->buf = buf; + buf_writer->internal_buf = false; + } else { + buf_writer->buf = buf_writer_allocate_internal_buf(tsdn, + buf_len); + buf_writer->internal_buf = true; + } + if (buf_writer->buf != NULL) { + buf_writer->buf_size = buf_len - 1; /* Allowing for '\0'. */ + } else { + buf_writer->buf_size = 0; + } + buf_writer->buf_end = 0; + buf_writer_assert(buf_writer); + return buf_writer->buf == NULL; +} + +void +buf_writer_flush(buf_writer_t *buf_writer) { + buf_writer_assert(buf_writer); + if (buf_writer->buf == NULL) { + return; + } + buf_writer->buf[buf_writer->buf_end] = '\0'; + buf_writer->write_cb(buf_writer->cbopaque, buf_writer->buf); + buf_writer->buf_end = 0; + buf_writer_assert(buf_writer); +} + +void +buf_writer_cb(void *buf_writer_arg, const char *s) { + buf_writer_t *buf_writer = (buf_writer_t *)buf_writer_arg; + buf_writer_assert(buf_writer); + if (buf_writer->buf == NULL) { + buf_writer->write_cb(buf_writer->cbopaque, s); + return; + } + size_t i, slen, n; + for (i = 0, slen = strlen(s); i < slen; i += n) { + if (buf_writer->buf_end == buf_writer->buf_size) { + buf_writer_flush(buf_writer); + } + size_t s_remain = slen - i; + size_t buf_remain = buf_writer->buf_size - buf_writer->buf_end; + n = s_remain < buf_remain ? s_remain : buf_remain; + memcpy(buf_writer->buf + buf_writer->buf_end, s + i, n); + buf_writer->buf_end += n; + buf_writer_assert(buf_writer); + } + assert(i == slen); +} + +void +buf_writer_terminate(tsdn_t *tsdn, buf_writer_t *buf_writer) { + buf_writer_assert(buf_writer); + buf_writer_flush(buf_writer); + if (buf_writer->internal_buf) { + buf_writer_free_internal_buf(tsdn, buf_writer->buf); + } +} + +void +buf_writer_pipe(buf_writer_t *buf_writer, read_cb_t *read_cb, + void *read_cbopaque) { + /* + * A tiny local buffer in case the buffered writer failed to allocate + * at init. + */ + static char backup_buf[16]; + static buf_writer_t backup_buf_writer; + + buf_writer_assert(buf_writer); + assert(read_cb != NULL); + if (buf_writer->buf == NULL) { + buf_writer_init(TSDN_NULL, &backup_buf_writer, + buf_writer->write_cb, buf_writer->cbopaque, backup_buf, + sizeof(backup_buf)); + buf_writer = &backup_buf_writer; + } + assert(buf_writer->buf != NULL); + ssize_t nread = 0; + do { + buf_writer->buf_end += nread; + buf_writer_assert(buf_writer); + if (buf_writer->buf_end == buf_writer->buf_size) { + buf_writer_flush(buf_writer); + } + nread = read_cb(read_cbopaque, + buf_writer->buf + buf_writer->buf_end, + buf_writer->buf_size - buf_writer->buf_end); + } while (nread > 0); + buf_writer_flush(buf_writer); +} diff --git a/contrib/jemalloc/src/cache_bin.c b/contrib/jemalloc/src/cache_bin.c new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/src/cache_bin.c @@ -0,0 +1,99 @@ +#include "jemalloc/internal/jemalloc_preamble.h" +#include "jemalloc/internal/jemalloc_internal_includes.h" + +#include "jemalloc/internal/bit_util.h" +#include "jemalloc/internal/cache_bin.h" +#include "jemalloc/internal/safety_check.h" + +void +cache_bin_info_init(cache_bin_info_t *info, + cache_bin_sz_t ncached_max) { + assert(ncached_max <= CACHE_BIN_NCACHED_MAX); + size_t stack_size = (size_t)ncached_max * sizeof(void *); + assert(stack_size < ((size_t)1 << (sizeof(cache_bin_sz_t) * 8))); + info->ncached_max = (cache_bin_sz_t)ncached_max; +} + +void +cache_bin_info_compute_alloc(cache_bin_info_t *infos, szind_t ninfos, + size_t *size, size_t *alignment) { + /* For the total bin stack region (per tcache), reserve 2 more slots so + * that + * 1) the empty position can be safely read on the fast path before + * checking "is_empty"; and + * 2) the cur_ptr can go beyond the empty position by 1 step safely on + * the fast path (i.e. no overflow). + */ + *size = sizeof(void *) * 2; + for (szind_t i = 0; i < ninfos; i++) { + assert(infos[i].ncached_max > 0); + *size += infos[i].ncached_max * sizeof(void *); + } + + /* + * Align to at least PAGE, to minimize the # of TLBs needed by the + * smaller sizes; also helps if the larger sizes don't get used at all. + */ + *alignment = PAGE; +} + +void +cache_bin_preincrement(cache_bin_info_t *infos, szind_t ninfos, void *alloc, + size_t *cur_offset) { + if (config_debug) { + size_t computed_size; + size_t computed_alignment; + + /* Pointer should be as aligned as we asked for. */ + cache_bin_info_compute_alloc(infos, ninfos, &computed_size, + &computed_alignment); + assert(((uintptr_t)alloc & (computed_alignment - 1)) == 0); + } + + *(uintptr_t *)((uintptr_t)alloc + *cur_offset) = + cache_bin_preceding_junk; + *cur_offset += sizeof(void *); +} + +void +cache_bin_postincrement(cache_bin_info_t *infos, szind_t ninfos, void *alloc, + size_t *cur_offset) { + *(uintptr_t *)((uintptr_t)alloc + *cur_offset) = + cache_bin_trailing_junk; + *cur_offset += sizeof(void *); +} + +void +cache_bin_init(cache_bin_t *bin, cache_bin_info_t *info, void *alloc, + size_t *cur_offset) { + /* + * The full_position points to the lowest available space. Allocations + * will access the slots toward higher addresses (for the benefit of + * adjacent prefetch). + */ + void *stack_cur = (void *)((uintptr_t)alloc + *cur_offset); + void *full_position = stack_cur; + uint16_t bin_stack_size = info->ncached_max * sizeof(void *); + + *cur_offset += bin_stack_size; + void *empty_position = (void *)((uintptr_t)alloc + *cur_offset); + + /* Init to the empty position. */ + bin->stack_head = (void **)empty_position; + bin->low_bits_low_water = (uint16_t)(uintptr_t)bin->stack_head; + bin->low_bits_full = (uint16_t)(uintptr_t)full_position; + bin->low_bits_empty = (uint16_t)(uintptr_t)empty_position; + cache_bin_sz_t free_spots = cache_bin_diff(bin, + bin->low_bits_full, (uint16_t)(uintptr_t)bin->stack_head, + /* racy */ false); + assert(free_spots == bin_stack_size); + assert(cache_bin_ncached_get_local(bin, info) == 0); + assert(cache_bin_empty_position_get(bin) == empty_position); + + assert(bin_stack_size > 0 || empty_position == full_position); +} + +bool +cache_bin_still_zero_initialized(cache_bin_t *bin) { + return bin->stack_head == NULL; +} diff --git a/contrib/jemalloc/src/ckh.c b/contrib/jemalloc/src/ckh.c --- a/contrib/jemalloc/src/ckh.c +++ b/contrib/jemalloc/src/ckh.c @@ -34,7 +34,6 @@ * respectively. * ******************************************************************************/ -#define JEMALLOC_CKH_C_ #include "jemalloc/internal/jemalloc_preamble.h" #include "jemalloc/internal/ckh.h" @@ -357,14 +356,14 @@ } bool -ckh_new(tsd_t *tsd, ckh_t *ckh, size_t minitems, ckh_hash_t *hash, +ckh_new(tsd_t *tsd, ckh_t *ckh, size_t minitems, ckh_hash_t *ckh_hash, ckh_keycomp_t *keycomp) { bool ret; size_t mincells, usize; unsigned lg_mincells; assert(minitems > 0); - assert(hash != NULL); + assert(ckh_hash != NULL); assert(keycomp != NULL); #ifdef CKH_COUNT @@ -393,7 +392,7 @@ } ckh->lg_minbuckets = lg_mincells - LG_CKH_BUCKET_CELLS; ckh->lg_curbuckets = lg_mincells - LG_CKH_BUCKET_CELLS; - ckh->hash = hash; + ckh->hash = ckh_hash; ckh->keycomp = keycomp; usize = sz_sa2u(sizeof(ckhc_t) << lg_mincells, CACHELINE); diff --git a/contrib/jemalloc/src/counter.c b/contrib/jemalloc/src/counter.c new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/src/counter.c @@ -0,0 +1,30 @@ +#include "jemalloc/internal/jemalloc_preamble.h" +#include "jemalloc/internal/jemalloc_internal_includes.h" + +#include "jemalloc/internal/counter.h" + +bool +counter_accum_init(counter_accum_t *counter, uint64_t interval) { + if (LOCKEDINT_MTX_INIT(counter->mtx, "counter_accum", + WITNESS_RANK_COUNTER_ACCUM, malloc_mutex_rank_exclusive)) { + return true; + } + locked_init_u64_unsynchronized(&counter->accumbytes, 0); + counter->interval = interval; + return false; +} + +void +counter_prefork(tsdn_t *tsdn, counter_accum_t *counter) { + LOCKEDINT_MTX_PREFORK(tsdn, counter->mtx); +} + +void +counter_postfork_parent(tsdn_t *tsdn, counter_accum_t *counter) { + LOCKEDINT_MTX_POSTFORK_PARENT(tsdn, counter->mtx); +} + +void +counter_postfork_child(tsdn_t *tsdn, counter_accum_t *counter) { + LOCKEDINT_MTX_POSTFORK_CHILD(tsdn, counter->mtx); +} diff --git a/contrib/jemalloc/src/ctl.c b/contrib/jemalloc/src/ctl.c --- a/contrib/jemalloc/src/ctl.c +++ b/contrib/jemalloc/src/ctl.c @@ -1,4 +1,3 @@ -#define JEMALLOC_CTL_C_ #include "jemalloc/internal/jemalloc_preamble.h" #include "jemalloc/internal/jemalloc_internal_includes.h" @@ -6,8 +5,16 @@ #include "jemalloc/internal/ctl.h" #include "jemalloc/internal/extent_dss.h" #include "jemalloc/internal/extent_mmap.h" +#include "jemalloc/internal/inspect.h" #include "jemalloc/internal/mutex.h" #include "jemalloc/internal/nstime.h" +#include "jemalloc/internal/peak_event.h" +#include "jemalloc/internal/prof_data.h" +#include "jemalloc/internal/prof_log.h" +#include "jemalloc/internal/prof_recent.h" +#include "jemalloc/internal/prof_stats.h" +#include "jemalloc/internal/prof_sys.h" +#include "jemalloc/internal/safety_check.h" #include "jemalloc/internal/sc.h" #include "jemalloc/internal/util.h" @@ -60,6 +67,8 @@ CTL_PROTO(max_background_threads) CTL_PROTO(thread_tcache_enabled) CTL_PROTO(thread_tcache_flush) +CTL_PROTO(thread_peak_read) +CTL_PROTO(thread_peak_reset) CTL_PROTO(thread_prof_name) CTL_PROTO(thread_prof_active) CTL_PROTO(thread_arena) @@ -67,6 +76,7 @@ CTL_PROTO(thread_allocatedp) CTL_PROTO(thread_deallocated) CTL_PROTO(thread_deallocatedp) +CTL_PROTO(thread_idle) CTL_PROTO(config_cache_oblivious) CTL_PROTO(config_debug) CTL_PROTO(config_fill) @@ -81,7 +91,20 @@ CTL_PROTO(config_xmalloc) CTL_PROTO(opt_abort) CTL_PROTO(opt_abort_conf) +CTL_PROTO(opt_cache_oblivious) +CTL_PROTO(opt_trust_madvise) CTL_PROTO(opt_confirm_conf) +CTL_PROTO(opt_hpa) +CTL_PROTO(opt_hpa_slab_max_alloc) +CTL_PROTO(opt_hpa_hugification_threshold) +CTL_PROTO(opt_hpa_hugify_delay_ms) +CTL_PROTO(opt_hpa_min_purge_interval_ms) +CTL_PROTO(opt_hpa_dirty_mult) +CTL_PROTO(opt_hpa_sec_nshards) +CTL_PROTO(opt_hpa_sec_max_alloc) +CTL_PROTO(opt_hpa_sec_max_bytes) +CTL_PROTO(opt_hpa_sec_bytes_after_flush) +CTL_PROTO(opt_hpa_sec_batch_fill_extra) CTL_PROTO(opt_metadata_thp) CTL_PROTO(opt_retain) CTL_PROTO(opt_dss) @@ -89,19 +112,31 @@ CTL_PROTO(opt_percpu_arena) CTL_PROTO(opt_oversize_threshold) CTL_PROTO(opt_background_thread) +CTL_PROTO(opt_mutex_max_spin) CTL_PROTO(opt_max_background_threads) CTL_PROTO(opt_dirty_decay_ms) CTL_PROTO(opt_muzzy_decay_ms) CTL_PROTO(opt_stats_print) CTL_PROTO(opt_stats_print_opts) +CTL_PROTO(opt_stats_interval) +CTL_PROTO(opt_stats_interval_opts) CTL_PROTO(opt_junk) CTL_PROTO(opt_zero) CTL_PROTO(opt_utrace) CTL_PROTO(opt_xmalloc) +CTL_PROTO(opt_experimental_infallible_new) CTL_PROTO(opt_tcache) +CTL_PROTO(opt_tcache_max) +CTL_PROTO(opt_tcache_nslots_small_min) +CTL_PROTO(opt_tcache_nslots_small_max) +CTL_PROTO(opt_tcache_nslots_large) +CTL_PROTO(opt_lg_tcache_nslots_mul) +CTL_PROTO(opt_tcache_gc_incr_bytes) +CTL_PROTO(opt_tcache_gc_delay_bytes) +CTL_PROTO(opt_lg_tcache_flush_small_div) +CTL_PROTO(opt_lg_tcache_flush_large_div) CTL_PROTO(opt_thp) CTL_PROTO(opt_lg_extent_max_active_fit) -CTL_PROTO(opt_lg_tcache_max) CTL_PROTO(opt_prof) CTL_PROTO(opt_prof_prefix) CTL_PROTO(opt_prof_active) @@ -111,7 +146,14 @@ CTL_PROTO(opt_prof_gdump) CTL_PROTO(opt_prof_final) CTL_PROTO(opt_prof_leak) +CTL_PROTO(opt_prof_leak_error) CTL_PROTO(opt_prof_accum) +CTL_PROTO(opt_prof_recent_alloc_max) +CTL_PROTO(opt_prof_stats) +CTL_PROTO(opt_prof_sys_thread_name) +CTL_PROTO(opt_prof_time_res) +CTL_PROTO(opt_lg_san_uaf_align) +CTL_PROTO(opt_zero_realloc) CTL_PROTO(tcache_create) CTL_PROTO(tcache_flush) CTL_PROTO(tcache_destroy) @@ -121,6 +163,7 @@ CTL_PROTO(arena_i_reset) CTL_PROTO(arena_i_destroy) CTL_PROTO(arena_i_dss) +CTL_PROTO(arena_i_oversize_threshold) CTL_PROTO(arena_i_dirty_decay_ms) CTL_PROTO(arena_i_muzzy_decay_ms) CTL_PROTO(arena_i_extent_hooks) @@ -148,11 +191,18 @@ CTL_PROTO(prof_active) CTL_PROTO(prof_dump) CTL_PROTO(prof_gdump) +CTL_PROTO(prof_prefix) CTL_PROTO(prof_reset) CTL_PROTO(prof_interval) CTL_PROTO(lg_prof_sample) CTL_PROTO(prof_log_start) CTL_PROTO(prof_log_stop) +CTL_PROTO(prof_stats_bins_i_live) +CTL_PROTO(prof_stats_bins_i_accum) +INDEX_PROTO(prof_stats_bins_i) +CTL_PROTO(prof_stats_lextents_i_live) +CTL_PROTO(prof_stats_lextents_i_accum) +INDEX_PROTO(prof_stats_lextents_i) CTL_PROTO(stats_arenas_i_small_allocated) CTL_PROTO(stats_arenas_i_small_nmalloc) CTL_PROTO(stats_arenas_i_small_ndalloc) @@ -188,6 +238,39 @@ CTL_PROTO(stats_arenas_i_extents_j_muzzy_bytes) CTL_PROTO(stats_arenas_i_extents_j_retained_bytes) INDEX_PROTO(stats_arenas_i_extents_j) +CTL_PROTO(stats_arenas_i_hpa_shard_npurge_passes) +CTL_PROTO(stats_arenas_i_hpa_shard_npurges) +CTL_PROTO(stats_arenas_i_hpa_shard_nhugifies) +CTL_PROTO(stats_arenas_i_hpa_shard_ndehugifies) + +/* We have a set of stats for full slabs. */ +CTL_PROTO(stats_arenas_i_hpa_shard_full_slabs_npageslabs_nonhuge) +CTL_PROTO(stats_arenas_i_hpa_shard_full_slabs_npageslabs_huge) +CTL_PROTO(stats_arenas_i_hpa_shard_full_slabs_nactive_nonhuge) +CTL_PROTO(stats_arenas_i_hpa_shard_full_slabs_nactive_huge) +CTL_PROTO(stats_arenas_i_hpa_shard_full_slabs_ndirty_nonhuge) +CTL_PROTO(stats_arenas_i_hpa_shard_full_slabs_ndirty_huge) + +/* A parallel set for the empty slabs. */ +CTL_PROTO(stats_arenas_i_hpa_shard_empty_slabs_npageslabs_nonhuge) +CTL_PROTO(stats_arenas_i_hpa_shard_empty_slabs_npageslabs_huge) +CTL_PROTO(stats_arenas_i_hpa_shard_empty_slabs_nactive_nonhuge) +CTL_PROTO(stats_arenas_i_hpa_shard_empty_slabs_nactive_huge) +CTL_PROTO(stats_arenas_i_hpa_shard_empty_slabs_ndirty_nonhuge) +CTL_PROTO(stats_arenas_i_hpa_shard_empty_slabs_ndirty_huge) + +/* + * And one for the slabs that are neither empty nor full, but indexed by how + * full they are. + */ +CTL_PROTO(stats_arenas_i_hpa_shard_nonfull_slabs_j_npageslabs_nonhuge) +CTL_PROTO(stats_arenas_i_hpa_shard_nonfull_slabs_j_npageslabs_huge) +CTL_PROTO(stats_arenas_i_hpa_shard_nonfull_slabs_j_nactive_nonhuge) +CTL_PROTO(stats_arenas_i_hpa_shard_nonfull_slabs_j_nactive_huge) +CTL_PROTO(stats_arenas_i_hpa_shard_nonfull_slabs_j_ndirty_nonhuge) +CTL_PROTO(stats_arenas_i_hpa_shard_nonfull_slabs_j_ndirty_huge) + +INDEX_PROTO(stats_arenas_i_hpa_shard_nonfull_slabs_j) CTL_PROTO(stats_arenas_i_nthreads) CTL_PROTO(stats_arenas_i_uptime) CTL_PROTO(stats_arenas_i_dss) @@ -209,8 +292,10 @@ CTL_PROTO(stats_arenas_i_internal) CTL_PROTO(stats_arenas_i_metadata_thp) CTL_PROTO(stats_arenas_i_tcache_bytes) +CTL_PROTO(stats_arenas_i_tcache_stashed_bytes) CTL_PROTO(stats_arenas_i_resident) CTL_PROTO(stats_arenas_i_abandoned_vm) +CTL_PROTO(stats_arenas_i_hpa_sec_bytes) INDEX_PROTO(stats_arenas_i) CTL_PROTO(stats_allocated) CTL_PROTO(stats_active) @@ -222,12 +307,21 @@ CTL_PROTO(stats_resident) CTL_PROTO(stats_mapped) CTL_PROTO(stats_retained) +CTL_PROTO(stats_zero_reallocs) CTL_PROTO(experimental_hooks_install) CTL_PROTO(experimental_hooks_remove) +CTL_PROTO(experimental_hooks_prof_backtrace) +CTL_PROTO(experimental_hooks_prof_dump) +CTL_PROTO(experimental_hooks_safety_check_abort) +CTL_PROTO(experimental_thread_activity_callback) CTL_PROTO(experimental_utilization_query) CTL_PROTO(experimental_utilization_batch_query) CTL_PROTO(experimental_arenas_i_pactivep) INDEX_PROTO(experimental_arenas_i) +CTL_PROTO(experimental_prof_recent_alloc_max) +CTL_PROTO(experimental_prof_recent_alloc_dump) +CTL_PROTO(experimental_batch_alloc) +CTL_PROTO(experimental_arenas_create_ext) #define MUTEX_STATS_CTL_PROTO_GEN(n) \ CTL_PROTO(stats_##n##_num_ops) \ @@ -275,6 +369,11 @@ {NAME("flush"), CTL(thread_tcache_flush)} }; +static const ctl_named_node_t thread_peak_node[] = { + {NAME("read"), CTL(thread_peak_read)}, + {NAME("reset"), CTL(thread_peak_reset)}, +}; + static const ctl_named_node_t thread_prof_node[] = { {NAME("name"), CTL(thread_prof_name)}, {NAME("active"), CTL(thread_prof_active)} @@ -287,7 +386,9 @@ {NAME("deallocated"), CTL(thread_deallocated)}, {NAME("deallocatedp"), CTL(thread_deallocatedp)}, {NAME("tcache"), CHILD(named, thread_tcache)}, - {NAME("prof"), CHILD(named, thread_prof)} + {NAME("peak"), CHILD(named, thread_peak)}, + {NAME("prof"), CHILD(named, thread_prof)}, + {NAME("idle"), CTL(thread_idle)} }; static const ctl_named_node_t config_node[] = { @@ -308,27 +409,60 @@ static const ctl_named_node_t opt_node[] = { {NAME("abort"), CTL(opt_abort)}, {NAME("abort_conf"), CTL(opt_abort_conf)}, + {NAME("cache_oblivious"), CTL(opt_cache_oblivious)}, + {NAME("trust_madvise"), CTL(opt_trust_madvise)}, {NAME("confirm_conf"), CTL(opt_confirm_conf)}, + {NAME("hpa"), CTL(opt_hpa)}, + {NAME("hpa_slab_max_alloc"), CTL(opt_hpa_slab_max_alloc)}, + {NAME("hpa_hugification_threshold"), + CTL(opt_hpa_hugification_threshold)}, + {NAME("hpa_hugify_delay_ms"), CTL(opt_hpa_hugify_delay_ms)}, + {NAME("hpa_min_purge_interval_ms"), CTL(opt_hpa_min_purge_interval_ms)}, + {NAME("hpa_dirty_mult"), CTL(opt_hpa_dirty_mult)}, + {NAME("hpa_sec_nshards"), CTL(opt_hpa_sec_nshards)}, + {NAME("hpa_sec_max_alloc"), CTL(opt_hpa_sec_max_alloc)}, + {NAME("hpa_sec_max_bytes"), CTL(opt_hpa_sec_max_bytes)}, + {NAME("hpa_sec_bytes_after_flush"), + CTL(opt_hpa_sec_bytes_after_flush)}, + {NAME("hpa_sec_batch_fill_extra"), + CTL(opt_hpa_sec_batch_fill_extra)}, {NAME("metadata_thp"), CTL(opt_metadata_thp)}, {NAME("retain"), CTL(opt_retain)}, {NAME("dss"), CTL(opt_dss)}, {NAME("narenas"), CTL(opt_narenas)}, {NAME("percpu_arena"), CTL(opt_percpu_arena)}, {NAME("oversize_threshold"), CTL(opt_oversize_threshold)}, + {NAME("mutex_max_spin"), CTL(opt_mutex_max_spin)}, {NAME("background_thread"), CTL(opt_background_thread)}, {NAME("max_background_threads"), CTL(opt_max_background_threads)}, {NAME("dirty_decay_ms"), CTL(opt_dirty_decay_ms)}, {NAME("muzzy_decay_ms"), CTL(opt_muzzy_decay_ms)}, {NAME("stats_print"), CTL(opt_stats_print)}, {NAME("stats_print_opts"), CTL(opt_stats_print_opts)}, + {NAME("stats_interval"), CTL(opt_stats_interval)}, + {NAME("stats_interval_opts"), CTL(opt_stats_interval_opts)}, {NAME("junk"), CTL(opt_junk)}, {NAME("zero"), CTL(opt_zero)}, {NAME("utrace"), CTL(opt_utrace)}, {NAME("xmalloc"), CTL(opt_xmalloc)}, + {NAME("experimental_infallible_new"), + CTL(opt_experimental_infallible_new)}, {NAME("tcache"), CTL(opt_tcache)}, + {NAME("tcache_max"), CTL(opt_tcache_max)}, + {NAME("tcache_nslots_small_min"), + CTL(opt_tcache_nslots_small_min)}, + {NAME("tcache_nslots_small_max"), + CTL(opt_tcache_nslots_small_max)}, + {NAME("tcache_nslots_large"), CTL(opt_tcache_nslots_large)}, + {NAME("lg_tcache_nslots_mul"), CTL(opt_lg_tcache_nslots_mul)}, + {NAME("tcache_gc_incr_bytes"), CTL(opt_tcache_gc_incr_bytes)}, + {NAME("tcache_gc_delay_bytes"), CTL(opt_tcache_gc_delay_bytes)}, + {NAME("lg_tcache_flush_small_div"), + CTL(opt_lg_tcache_flush_small_div)}, + {NAME("lg_tcache_flush_large_div"), + CTL(opt_lg_tcache_flush_large_div)}, {NAME("thp"), CTL(opt_thp)}, {NAME("lg_extent_max_active_fit"), CTL(opt_lg_extent_max_active_fit)}, - {NAME("lg_tcache_max"), CTL(opt_lg_tcache_max)}, {NAME("prof"), CTL(opt_prof)}, {NAME("prof_prefix"), CTL(opt_prof_prefix)}, {NAME("prof_active"), CTL(opt_prof_active)}, @@ -338,7 +472,14 @@ {NAME("prof_gdump"), CTL(opt_prof_gdump)}, {NAME("prof_final"), CTL(opt_prof_final)}, {NAME("prof_leak"), CTL(opt_prof_leak)}, - {NAME("prof_accum"), CTL(opt_prof_accum)} + {NAME("prof_leak_error"), CTL(opt_prof_leak_error)}, + {NAME("prof_accum"), CTL(opt_prof_accum)}, + {NAME("prof_recent_alloc_max"), CTL(opt_prof_recent_alloc_max)}, + {NAME("prof_stats"), CTL(opt_prof_stats)}, + {NAME("prof_sys_thread_name"), CTL(opt_prof_sys_thread_name)}, + {NAME("prof_time_resolution"), CTL(opt_prof_time_res)}, + {NAME("lg_san_uaf_align"), CTL(opt_lg_san_uaf_align)}, + {NAME("zero_realloc"), CTL(opt_zero_realloc)} }; static const ctl_named_node_t tcache_node[] = { @@ -354,6 +495,11 @@ {NAME("reset"), CTL(arena_i_reset)}, {NAME("destroy"), CTL(arena_i_destroy)}, {NAME("dss"), CTL(arena_i_dss)}, + /* + * Undocumented for now, since we anticipate an arena API in flux after + * we cut the last 5-series release. + */ + {NAME("oversize_threshold"), CTL(arena_i_oversize_threshold)}, {NAME("dirty_decay_ms"), CTL(arena_i_dirty_decay_ms)}, {NAME("muzzy_decay_ms"), CTL(arena_i_muzzy_decay_ms)}, {NAME("extent_hooks"), CTL(arena_i_extent_hooks)}, @@ -408,17 +554,51 @@ {NAME("lookup"), CTL(arenas_lookup)} }; +static const ctl_named_node_t prof_stats_bins_i_node[] = { + {NAME("live"), CTL(prof_stats_bins_i_live)}, + {NAME("accum"), CTL(prof_stats_bins_i_accum)} +}; + +static const ctl_named_node_t super_prof_stats_bins_i_node[] = { + {NAME(""), CHILD(named, prof_stats_bins_i)} +}; + +static const ctl_indexed_node_t prof_stats_bins_node[] = { + {INDEX(prof_stats_bins_i)} +}; + +static const ctl_named_node_t prof_stats_lextents_i_node[] = { + {NAME("live"), CTL(prof_stats_lextents_i_live)}, + {NAME("accum"), CTL(prof_stats_lextents_i_accum)} +}; + +static const ctl_named_node_t super_prof_stats_lextents_i_node[] = { + {NAME(""), CHILD(named, prof_stats_lextents_i)} +}; + +static const ctl_indexed_node_t prof_stats_lextents_node[] = { + {INDEX(prof_stats_lextents_i)} +}; + +static const ctl_named_node_t prof_stats_node[] = { + {NAME("bins"), CHILD(indexed, prof_stats_bins)}, + {NAME("lextents"), CHILD(indexed, prof_stats_lextents)}, +}; + static const ctl_named_node_t prof_node[] = { {NAME("thread_active_init"), CTL(prof_thread_active_init)}, {NAME("active"), CTL(prof_active)}, {NAME("dump"), CTL(prof_dump)}, {NAME("gdump"), CTL(prof_gdump)}, + {NAME("prefix"), CTL(prof_prefix)}, {NAME("reset"), CTL(prof_reset)}, {NAME("interval"), CTL(prof_interval)}, {NAME("lg_sample"), CTL(lg_prof_sample)}, {NAME("log_start"), CTL(prof_log_start)}, - {NAME("log_stop"), CTL(prof_log_stop)} + {NAME("log_stop"), CTL(prof_log_stop)}, + {NAME("stats"), CHILD(named, prof_stats)} }; + static const ctl_named_node_t stats_arenas_i_small_node[] = { {NAME("allocated"), CTL(stats_arenas_i_small_allocated)}, {NAME("nmalloc"), CTL(stats_arenas_i_small_nmalloc)}, @@ -521,6 +701,75 @@ #undef OP }; +static const ctl_named_node_t stats_arenas_i_hpa_shard_full_slabs_node[] = { + {NAME("npageslabs_nonhuge"), + CTL(stats_arenas_i_hpa_shard_full_slabs_npageslabs_nonhuge)}, + {NAME("npageslabs_huge"), + CTL(stats_arenas_i_hpa_shard_full_slabs_npageslabs_huge)}, + {NAME("nactive_nonhuge"), + CTL(stats_arenas_i_hpa_shard_full_slabs_nactive_nonhuge)}, + {NAME("nactive_huge"), + CTL(stats_arenas_i_hpa_shard_full_slabs_nactive_huge)}, + {NAME("ndirty_nonhuge"), + CTL(stats_arenas_i_hpa_shard_full_slabs_ndirty_nonhuge)}, + {NAME("ndirty_huge"), + CTL(stats_arenas_i_hpa_shard_full_slabs_ndirty_huge)} +}; + +static const ctl_named_node_t stats_arenas_i_hpa_shard_empty_slabs_node[] = { + {NAME("npageslabs_nonhuge"), + CTL(stats_arenas_i_hpa_shard_empty_slabs_npageslabs_nonhuge)}, + {NAME("npageslabs_huge"), + CTL(stats_arenas_i_hpa_shard_empty_slabs_npageslabs_huge)}, + {NAME("nactive_nonhuge"), + CTL(stats_arenas_i_hpa_shard_empty_slabs_nactive_nonhuge)}, + {NAME("nactive_huge"), + CTL(stats_arenas_i_hpa_shard_empty_slabs_nactive_huge)}, + {NAME("ndirty_nonhuge"), + CTL(stats_arenas_i_hpa_shard_empty_slabs_ndirty_nonhuge)}, + {NAME("ndirty_huge"), + CTL(stats_arenas_i_hpa_shard_empty_slabs_ndirty_huge)} +}; + +static const ctl_named_node_t stats_arenas_i_hpa_shard_nonfull_slabs_j_node[] = { + {NAME("npageslabs_nonhuge"), + CTL(stats_arenas_i_hpa_shard_nonfull_slabs_j_npageslabs_nonhuge)}, + {NAME("npageslabs_huge"), + CTL(stats_arenas_i_hpa_shard_nonfull_slabs_j_npageslabs_huge)}, + {NAME("nactive_nonhuge"), + CTL(stats_arenas_i_hpa_shard_nonfull_slabs_j_nactive_nonhuge)}, + {NAME("nactive_huge"), + CTL(stats_arenas_i_hpa_shard_nonfull_slabs_j_nactive_huge)}, + {NAME("ndirty_nonhuge"), + CTL(stats_arenas_i_hpa_shard_nonfull_slabs_j_ndirty_nonhuge)}, + {NAME("ndirty_huge"), + CTL(stats_arenas_i_hpa_shard_nonfull_slabs_j_ndirty_huge)} +}; + +static const ctl_named_node_t super_stats_arenas_i_hpa_shard_nonfull_slabs_j_node[] = { + {NAME(""), + CHILD(named, stats_arenas_i_hpa_shard_nonfull_slabs_j)} +}; + +static const ctl_indexed_node_t stats_arenas_i_hpa_shard_nonfull_slabs_node[] = +{ + {INDEX(stats_arenas_i_hpa_shard_nonfull_slabs_j)} +}; + +static const ctl_named_node_t stats_arenas_i_hpa_shard_node[] = { + {NAME("full_slabs"), CHILD(named, + stats_arenas_i_hpa_shard_full_slabs)}, + {NAME("empty_slabs"), CHILD(named, + stats_arenas_i_hpa_shard_empty_slabs)}, + {NAME("nonfull_slabs"), CHILD(indexed, + stats_arenas_i_hpa_shard_nonfull_slabs)}, + + {NAME("npurge_passes"), CTL(stats_arenas_i_hpa_shard_npurge_passes)}, + {NAME("npurges"), CTL(stats_arenas_i_hpa_shard_npurges)}, + {NAME("nhugifies"), CTL(stats_arenas_i_hpa_shard_nhugifies)}, + {NAME("ndehugifies"), CTL(stats_arenas_i_hpa_shard_ndehugifies)} +}; + static const ctl_named_node_t stats_arenas_i_node[] = { {NAME("nthreads"), CTL(stats_arenas_i_nthreads)}, {NAME("uptime"), CTL(stats_arenas_i_uptime)}, @@ -543,14 +792,18 @@ {NAME("internal"), CTL(stats_arenas_i_internal)}, {NAME("metadata_thp"), CTL(stats_arenas_i_metadata_thp)}, {NAME("tcache_bytes"), CTL(stats_arenas_i_tcache_bytes)}, + {NAME("tcache_stashed_bytes"), + CTL(stats_arenas_i_tcache_stashed_bytes)}, {NAME("resident"), CTL(stats_arenas_i_resident)}, {NAME("abandoned_vm"), CTL(stats_arenas_i_abandoned_vm)}, + {NAME("hpa_sec_bytes"), CTL(stats_arenas_i_hpa_sec_bytes)}, {NAME("small"), CHILD(named, stats_arenas_i_small)}, {NAME("large"), CHILD(named, stats_arenas_i_large)}, {NAME("bins"), CHILD(indexed, stats_arenas_i_bins)}, {NAME("lextents"), CHILD(indexed, stats_arenas_i_lextents)}, {NAME("extents"), CHILD(indexed, stats_arenas_i_extents)}, - {NAME("mutexes"), CHILD(named, stats_arenas_i_mutexes)} + {NAME("mutexes"), CHILD(named, stats_arenas_i_mutexes)}, + {NAME("hpa_shard"), CHILD(named, stats_arenas_i_hpa_shard)} }; static const ctl_named_node_t super_stats_arenas_i_node[] = { {NAME(""), CHILD(named, stats_arenas_i)} @@ -589,12 +842,21 @@ {NAME("background_thread"), CHILD(named, stats_background_thread)}, {NAME("mutexes"), CHILD(named, stats_mutexes)}, - {NAME("arenas"), CHILD(indexed, stats_arenas)} + {NAME("arenas"), CHILD(indexed, stats_arenas)}, + {NAME("zero_reallocs"), CTL(stats_zero_reallocs)}, }; static const ctl_named_node_t experimental_hooks_node[] = { {NAME("install"), CTL(experimental_hooks_install)}, - {NAME("remove"), CTL(experimental_hooks_remove)} + {NAME("remove"), CTL(experimental_hooks_remove)}, + {NAME("prof_backtrace"), CTL(experimental_hooks_prof_backtrace)}, + {NAME("prof_dump"), CTL(experimental_hooks_prof_dump)}, + {NAME("safety_check_abort"), CTL(experimental_hooks_safety_check_abort)}, +}; + +static const ctl_named_node_t experimental_thread_node[] = { + {NAME("activity_callback"), + CTL(experimental_thread_activity_callback)} }; static const ctl_named_node_t experimental_utilization_node[] = { @@ -613,10 +875,19 @@ {INDEX(experimental_arenas_i)} }; +static const ctl_named_node_t experimental_prof_recent_node[] = { + {NAME("alloc_max"), CTL(experimental_prof_recent_alloc_max)}, + {NAME("alloc_dump"), CTL(experimental_prof_recent_alloc_dump)}, +}; + static const ctl_named_node_t experimental_node[] = { {NAME("hooks"), CHILD(named, experimental_hooks)}, {NAME("utilization"), CHILD(named, experimental_utilization)}, - {NAME("arenas"), CHILD(indexed, experimental_arenas)} + {NAME("arenas"), CHILD(indexed, experimental_arenas)}, + {NAME("arenas_create_ext"), CTL(experimental_arenas_create_ext)}, + {NAME("prof_recent"), CHILD(named, experimental_prof_recent)}, + {NAME("batch_alloc"), CTL(experimental_batch_alloc)}, + {NAME("thread"), CHILD(named, experimental_thread)} }; static const ctl_named_node_t root_node[] = { @@ -650,28 +921,13 @@ * synchronized by the ctl mutex. */ static void -ctl_accum_arena_stats_u64(arena_stats_u64_t *dst, arena_stats_u64_t *src) { -#ifdef JEMALLOC_ATOMIC_U64 - uint64_t cur_dst = atomic_load_u64(dst, ATOMIC_RELAXED); - uint64_t cur_src = atomic_load_u64(src, ATOMIC_RELAXED); - atomic_store_u64(dst, cur_dst + cur_src, ATOMIC_RELAXED); -#else - *dst += *src; -#endif -} - -/* Likewise: with ctl mutex synchronization, reading is simple. */ -static uint64_t -ctl_arena_stats_read_u64(arena_stats_u64_t *p) { -#ifdef JEMALLOC_ATOMIC_U64 - return atomic_load_u64(p, ATOMIC_RELAXED); -#else - return *p; -#endif +ctl_accum_locked_u64(locked_u64_t *dst, locked_u64_t *src) { + locked_inc_u64_unsynchronized(dst, + locked_read_u64_unsynchronized(src)); } static void -accum_atomic_zu(atomic_zu_t *dst, atomic_zu_t *src) { +ctl_accum_atomic_zu(atomic_zu_t *dst, atomic_zu_t *src) { size_t cur_dst = atomic_load_zu(dst, ATOMIC_RELAXED); size_t cur_src = atomic_load_zu(src, ATOMIC_RELAXED); atomic_store_zu(dst, cur_dst + cur_src, ATOMIC_RELAXED); @@ -783,11 +1039,15 @@ ctl_arena->astats->nfills_small = 0; ctl_arena->astats->nflushes_small = 0; memset(ctl_arena->astats->bstats, 0, SC_NBINS * - sizeof(bin_stats_t)); + sizeof(bin_stats_data_t)); memset(ctl_arena->astats->lstats, 0, (SC_NSIZES - SC_NBINS) * sizeof(arena_stats_large_t)); memset(ctl_arena->astats->estats, 0, SC_NPSIZES * - sizeof(arena_stats_extents_t)); + sizeof(pac_estats_t)); + memset(&ctl_arena->astats->hpastats, 0, + sizeof(hpa_shard_stats_t)); + memset(&ctl_arena->astats->secstats, 0, + sizeof(sec_stats_t)); } } @@ -801,22 +1061,19 @@ &ctl_arena->muzzy_decay_ms, &ctl_arena->pactive, &ctl_arena->pdirty, &ctl_arena->pmuzzy, &ctl_arena->astats->astats, ctl_arena->astats->bstats, - ctl_arena->astats->lstats, ctl_arena->astats->estats); + ctl_arena->astats->lstats, ctl_arena->astats->estats, + &ctl_arena->astats->hpastats, &ctl_arena->astats->secstats); for (i = 0; i < SC_NBINS; i++) { - ctl_arena->astats->allocated_small += - ctl_arena->astats->bstats[i].curregs * + bin_stats_t *bstats = + &ctl_arena->astats->bstats[i].stats_data; + ctl_arena->astats->allocated_small += bstats->curregs * sz_index2size(i); - ctl_arena->astats->nmalloc_small += - ctl_arena->astats->bstats[i].nmalloc; - ctl_arena->astats->ndalloc_small += - ctl_arena->astats->bstats[i].ndalloc; - ctl_arena->astats->nrequests_small += - ctl_arena->astats->bstats[i].nrequests; - ctl_arena->astats->nfills_small += - ctl_arena->astats->bstats[i].nfills; - ctl_arena->astats->nflushes_small += - ctl_arena->astats->bstats[i].nflushes; + ctl_arena->astats->nmalloc_small += bstats->nmalloc; + ctl_arena->astats->ndalloc_small += bstats->ndalloc; + ctl_arena->astats->nrequests_small += bstats->nrequests; + ctl_arena->astats->nfills_small += bstats->nfills; + ctl_arena->astats->nflushes_small += bstats->nflushes; } } else { arena_basic_stats_merge(tsdn, arena, &ctl_arena->nthreads, @@ -848,27 +1105,32 @@ ctl_arena_stats_t *astats = ctl_arena->astats; if (!destroyed) { - accum_atomic_zu(&sdstats->astats.mapped, - &astats->astats.mapped); - accum_atomic_zu(&sdstats->astats.retained, - &astats->astats.retained); - accum_atomic_zu(&sdstats->astats.extent_avail, - &astats->astats.extent_avail); - } - - ctl_accum_arena_stats_u64(&sdstats->astats.decay_dirty.npurge, - &astats->astats.decay_dirty.npurge); - ctl_accum_arena_stats_u64(&sdstats->astats.decay_dirty.nmadvise, - &astats->astats.decay_dirty.nmadvise); - ctl_accum_arena_stats_u64(&sdstats->astats.decay_dirty.purged, - &astats->astats.decay_dirty.purged); - - ctl_accum_arena_stats_u64(&sdstats->astats.decay_muzzy.npurge, - &astats->astats.decay_muzzy.npurge); - ctl_accum_arena_stats_u64(&sdstats->astats.decay_muzzy.nmadvise, - &astats->astats.decay_muzzy.nmadvise); - ctl_accum_arena_stats_u64(&sdstats->astats.decay_muzzy.purged, - &astats->astats.decay_muzzy.purged); + sdstats->astats.mapped += astats->astats.mapped; + sdstats->astats.pa_shard_stats.pac_stats.retained + += astats->astats.pa_shard_stats.pac_stats.retained; + sdstats->astats.pa_shard_stats.edata_avail + += astats->astats.pa_shard_stats.edata_avail; + } + + ctl_accum_locked_u64( + &sdstats->astats.pa_shard_stats.pac_stats.decay_dirty.npurge, + &astats->astats.pa_shard_stats.pac_stats.decay_dirty.npurge); + ctl_accum_locked_u64( + &sdstats->astats.pa_shard_stats.pac_stats.decay_dirty.nmadvise, + &astats->astats.pa_shard_stats.pac_stats.decay_dirty.nmadvise); + ctl_accum_locked_u64( + &sdstats->astats.pa_shard_stats.pac_stats.decay_dirty.purged, + &astats->astats.pa_shard_stats.pac_stats.decay_dirty.purged); + + ctl_accum_locked_u64( + &sdstats->astats.pa_shard_stats.pac_stats.decay_muzzy.npurge, + &astats->astats.pa_shard_stats.pac_stats.decay_muzzy.npurge); + ctl_accum_locked_u64( + &sdstats->astats.pa_shard_stats.pac_stats.decay_muzzy.nmadvise, + &astats->astats.pa_shard_stats.pac_stats.decay_muzzy.nmadvise); + ctl_accum_locked_u64( + &sdstats->astats.pa_shard_stats.pac_stats.decay_muzzy.purged, + &astats->astats.pa_shard_stats.pac_stats.decay_muzzy.purged); #define OP(mtx) malloc_mutex_prof_merge( \ &(sdstats->astats.mutex_prof_data[ \ @@ -878,14 +1140,11 @@ MUTEX_PROF_ARENA_MUTEXES #undef OP if (!destroyed) { - accum_atomic_zu(&sdstats->astats.base, - &astats->astats.base); - accum_atomic_zu(&sdstats->astats.internal, + sdstats->astats.base += astats->astats.base; + sdstats->astats.resident += astats->astats.resident; + sdstats->astats.metadata_thp += astats->astats.metadata_thp; + ctl_accum_atomic_zu(&sdstats->astats.internal, &astats->astats.internal); - accum_atomic_zu(&sdstats->astats.resident, - &astats->astats.resident); - accum_atomic_zu(&sdstats->astats.metadata_thp, - &astats->astats.metadata_thp); } else { assert(atomic_load_zu( &astats->astats.internal, ATOMIC_RELAXED) == 0); @@ -903,23 +1162,23 @@ sdstats->nflushes_small += astats->nflushes_small; if (!destroyed) { - accum_atomic_zu(&sdstats->astats.allocated_large, - &astats->astats.allocated_large); + sdstats->astats.allocated_large += + astats->astats.allocated_large; } else { - assert(atomic_load_zu(&astats->astats.allocated_large, - ATOMIC_RELAXED) == 0); + assert(astats->astats.allocated_large == 0); } - ctl_accum_arena_stats_u64(&sdstats->astats.nmalloc_large, - &astats->astats.nmalloc_large); - ctl_accum_arena_stats_u64(&sdstats->astats.ndalloc_large, - &astats->astats.ndalloc_large); - ctl_accum_arena_stats_u64(&sdstats->astats.nrequests_large, - &astats->astats.nrequests_large); - accum_atomic_zu(&sdstats->astats.abandoned_vm, - &astats->astats.abandoned_vm); - - accum_atomic_zu(&sdstats->astats.tcache_bytes, - &astats->astats.tcache_bytes); + sdstats->astats.nmalloc_large += astats->astats.nmalloc_large; + sdstats->astats.ndalloc_large += astats->astats.ndalloc_large; + sdstats->astats.nrequests_large + += astats->astats.nrequests_large; + sdstats->astats.nflushes_large += astats->astats.nflushes_large; + ctl_accum_atomic_zu( + &sdstats->astats.pa_shard_stats.pac_stats.abandoned_vm, + &astats->astats.pa_shard_stats.pac_stats.abandoned_vm); + + sdstats->astats.tcache_bytes += astats->astats.tcache_bytes; + sdstats->astats.tcache_stashed_bytes += + astats->astats.tcache_stashed_bytes; if (ctl_arena->arena_ind == 0) { sdstats->astats.uptime = astats->astats.uptime; @@ -927,29 +1186,26 @@ /* Merge bin stats. */ for (i = 0; i < SC_NBINS; i++) { - sdstats->bstats[i].nmalloc += astats->bstats[i].nmalloc; - sdstats->bstats[i].ndalloc += astats->bstats[i].ndalloc; - sdstats->bstats[i].nrequests += - astats->bstats[i].nrequests; + bin_stats_t *bstats = &astats->bstats[i].stats_data; + bin_stats_t *merged = &sdstats->bstats[i].stats_data; + merged->nmalloc += bstats->nmalloc; + merged->ndalloc += bstats->ndalloc; + merged->nrequests += bstats->nrequests; if (!destroyed) { - sdstats->bstats[i].curregs += - astats->bstats[i].curregs; + merged->curregs += bstats->curregs; } else { - assert(astats->bstats[i].curregs == 0); + assert(bstats->curregs == 0); } - sdstats->bstats[i].nfills += astats->bstats[i].nfills; - sdstats->bstats[i].nflushes += - astats->bstats[i].nflushes; - sdstats->bstats[i].nslabs += astats->bstats[i].nslabs; - sdstats->bstats[i].reslabs += astats->bstats[i].reslabs; + merged->nfills += bstats->nfills; + merged->nflushes += bstats->nflushes; + merged->nslabs += bstats->nslabs; + merged->reslabs += bstats->reslabs; if (!destroyed) { - sdstats->bstats[i].curslabs += - astats->bstats[i].curslabs; - sdstats->bstats[i].nonfull_slabs += - astats->bstats[i].nonfull_slabs; + merged->curslabs += bstats->curslabs; + merged->nonfull_slabs += bstats->nonfull_slabs; } else { - assert(astats->bstats[i].curslabs == 0); - assert(astats->bstats[i].nonfull_slabs == 0); + assert(bstats->curslabs == 0); + assert(bstats->nonfull_slabs == 0); } malloc_mutex_prof_merge(&sdstats->bstats[i].mutex_data, &astats->bstats[i].mutex_data); @@ -957,11 +1213,11 @@ /* Merge stats for large allocations. */ for (i = 0; i < SC_NSIZES - SC_NBINS; i++) { - ctl_accum_arena_stats_u64(&sdstats->lstats[i].nmalloc, + ctl_accum_locked_u64(&sdstats->lstats[i].nmalloc, &astats->lstats[i].nmalloc); - ctl_accum_arena_stats_u64(&sdstats->lstats[i].ndalloc, + ctl_accum_locked_u64(&sdstats->lstats[i].ndalloc, &astats->lstats[i].ndalloc); - ctl_accum_arena_stats_u64(&sdstats->lstats[i].nrequests, + ctl_accum_locked_u64(&sdstats->lstats[i].nrequests, &astats->lstats[i].nrequests); if (!destroyed) { sdstats->lstats[i].curlextents += @@ -973,19 +1229,21 @@ /* Merge extents stats. */ for (i = 0; i < SC_NPSIZES; i++) { - accum_atomic_zu(&sdstats->estats[i].ndirty, - &astats->estats[i].ndirty); - accum_atomic_zu(&sdstats->estats[i].nmuzzy, - &astats->estats[i].nmuzzy); - accum_atomic_zu(&sdstats->estats[i].nretained, - &astats->estats[i].nretained); - accum_atomic_zu(&sdstats->estats[i].dirty_bytes, - &astats->estats[i].dirty_bytes); - accum_atomic_zu(&sdstats->estats[i].muzzy_bytes, - &astats->estats[i].muzzy_bytes); - accum_atomic_zu(&sdstats->estats[i].retained_bytes, - &astats->estats[i].retained_bytes); + sdstats->estats[i].ndirty += astats->estats[i].ndirty; + sdstats->estats[i].nmuzzy += astats->estats[i].nmuzzy; + sdstats->estats[i].nretained + += astats->estats[i].nretained; + sdstats->estats[i].dirty_bytes + += astats->estats[i].dirty_bytes; + sdstats->estats[i].muzzy_bytes + += astats->estats[i].muzzy_bytes; + sdstats->estats[i].retained_bytes + += astats->estats[i].retained_bytes; } + + /* Merge HPA stats. */ + hpa_shard_stats_accum(&sdstats->hpastats, &astats->hpastats); + sec_stats_accum(&sdstats->secstats, &astats->secstats); } } @@ -1001,7 +1259,7 @@ } static unsigned -ctl_arena_init(tsd_t *tsd, extent_hooks_t *extent_hooks) { +ctl_arena_init(tsd_t *tsd, const arena_config_t *config) { unsigned arena_ind; ctl_arena_t *ctl_arena; @@ -1019,7 +1277,7 @@ } /* Initialize new arena. */ - if (arena_init(tsd_tsdn(tsd), arena_ind, extent_hooks) == NULL) { + if (arena_init(tsd_tsdn(tsd), arena_ind, config) == NULL) { return UINT_MAX; } @@ -1036,8 +1294,11 @@ if (!have_background_thread || background_thread_stats_read(tsdn, stats)) { memset(stats, 0, sizeof(background_thread_stats_t)); - nstime_init(&stats->run_interval, 0); + nstime_init_zero(&stats->run_interval); } + malloc_mutex_prof_copy( + &ctl_stats->mutex_prof_data[global_prof_mutex_max_per_bg_thd], + &stats->max_counter_per_bg_thd); } static void @@ -1069,21 +1330,17 @@ if (config_stats) { ctl_stats->allocated = ctl_sarena->astats->allocated_small + - atomic_load_zu(&ctl_sarena->astats->astats.allocated_large, - ATOMIC_RELAXED); + ctl_sarena->astats->astats.allocated_large; ctl_stats->active = (ctl_sarena->pactive << LG_PAGE); - ctl_stats->metadata = atomic_load_zu( - &ctl_sarena->astats->astats.base, ATOMIC_RELAXED) + + ctl_stats->metadata = ctl_sarena->astats->astats.base + atomic_load_zu(&ctl_sarena->astats->astats.internal, ATOMIC_RELAXED); - ctl_stats->metadata_thp = atomic_load_zu( - &ctl_sarena->astats->astats.metadata_thp, ATOMIC_RELAXED); - ctl_stats->resident = atomic_load_zu( - &ctl_sarena->astats->astats.resident, ATOMIC_RELAXED); - ctl_stats->mapped = atomic_load_zu( - &ctl_sarena->astats->astats.mapped, ATOMIC_RELAXED); - ctl_stats->retained = atomic_load_zu( - &ctl_sarena->astats->astats.retained, ATOMIC_RELAXED); + ctl_stats->resident = ctl_sarena->astats->astats.resident; + ctl_stats->metadata_thp = + ctl_sarena->astats->astats.metadata_thp; + ctl_stats->mapped = ctl_sarena->astats->astats.mapped; + ctl_stats->retained = ctl_sarena->astats->astats + .pa_shard_stats.pac_stats.retained; ctl_background_thread_stats_read(tsdn); @@ -1093,8 +1350,20 @@ malloc_mutex_unlock(tsdn, &mtx); if (config_prof && opt_prof) { - READ_GLOBAL_MUTEX_PROF_DATA(global_prof_mutex_prof, - bt2gctx_mtx); + READ_GLOBAL_MUTEX_PROF_DATA( + global_prof_mutex_prof, bt2gctx_mtx); + READ_GLOBAL_MUTEX_PROF_DATA( + global_prof_mutex_prof_thds_data, tdatas_mtx); + READ_GLOBAL_MUTEX_PROF_DATA( + global_prof_mutex_prof_dump, prof_dump_mtx); + READ_GLOBAL_MUTEX_PROF_DATA( + global_prof_mutex_prof_recent_alloc, + prof_recent_alloc_mtx); + READ_GLOBAL_MUTEX_PROF_DATA( + global_prof_mutex_prof_recent_dump, + prof_recent_dump_mtx); + READ_GLOBAL_MUTEX_PROF_DATA( + global_prof_mutex_prof_stats, prof_stats_mtx); } if (have_background_thread) { READ_GLOBAL_MUTEX_PROF_DATA( @@ -1191,8 +1460,9 @@ } static int -ctl_lookup(tsdn_t *tsdn, const char *name, ctl_node_t const **nodesp, - size_t *mibp, size_t *depthp) { +ctl_lookup(tsdn_t *tsdn, const ctl_named_node_t *starting_node, + const char *name, const ctl_named_node_t **ending_nodep, size_t *mibp, + size_t *depthp) { int ret; const char *elm, *tdot, *dot; size_t elen, i, j; @@ -1206,7 +1476,7 @@ ret = ENOENT; goto label_return; } - node = super_root_node; + node = starting_node; for (i = 0; i < *depthp; i++) { assert(node); assert(node->nchildren > 0); @@ -1220,10 +1490,6 @@ if (strlen(child->name) == elen && strncmp(elm, child->name, elen) == 0) { node = child; - if (nodesp != NULL) { - nodesp[i] = - (const ctl_node_t *)node; - } mibp[i] = j; break; } @@ -1250,13 +1516,11 @@ goto label_return; } - if (nodesp != NULL) { - nodesp[i] = (const ctl_node_t *)node; - } mibp[i] = (size_t)index; } - if (node->ctl != NULL) { + /* Reached the end? */ + if (node->ctl != NULL || *dot == '\0') { /* Terminal node. */ if (*dot != '\0') { /* @@ -1272,16 +1536,14 @@ } /* Update elm. */ - if (*dot == '\0') { - /* No more elements. */ - ret = ENOENT; - goto label_return; - } elm = &dot[1]; dot = ((tdot = strchr(elm, '.')) != NULL) ? tdot : strchr(elm, '\0'); elen = (size_t)((uintptr_t)dot - (uintptr_t)elm); } + if (ending_nodep != NULL) { + *ending_nodep = node; + } ret = 0; label_return: @@ -1293,7 +1555,6 @@ void *newp, size_t newlen) { int ret; size_t depth; - ctl_node_t const *nodes[CTL_MAX_DEPTH]; size_t mib[CTL_MAX_DEPTH]; const ctl_named_node_t *node; @@ -1303,12 +1564,12 @@ } depth = CTL_MAX_DEPTH; - ret = ctl_lookup(tsd_tsdn(tsd), name, nodes, mib, &depth); + ret = ctl_lookup(tsd_tsdn(tsd), super_root_node, name, &node, mib, + &depth); if (ret != 0) { goto label_return; } - node = ctl_named_node(nodes[depth-1]); if (node != NULL && node->ctl) { ret = node->ctl(tsd, mib, depth, oldp, oldlenp, newp, newlen); } else { @@ -1329,26 +1590,19 @@ goto label_return; } - ret = ctl_lookup(tsd_tsdn(tsd), name, NULL, mibp, miblenp); + ret = ctl_lookup(tsd_tsdn(tsd), super_root_node, name, NULL, mibp, + miblenp); label_return: return(ret); } -int -ctl_bymib(tsd_t *tsd, const size_t *mib, size_t miblen, void *oldp, - size_t *oldlenp, void *newp, size_t newlen) { +static int +ctl_lookupbymib(tsdn_t *tsdn, const ctl_named_node_t **ending_nodep, + const size_t *mib, size_t miblen) { int ret; - const ctl_named_node_t *node; - size_t i; - - if (!ctl_initialized && ctl_init(tsd)) { - ret = EAGAIN; - goto label_return; - } - /* Iterate down the tree. */ - node = super_root_node; - for (i = 0; i < miblen; i++) { + const ctl_named_node_t *node = super_root_node; + for (size_t i = 0; i < miblen; i++) { assert(node); assert(node->nchildren > 0); if (ctl_named_node(node->children) != NULL) { @@ -1363,13 +1617,36 @@ /* Indexed element. */ inode = ctl_indexed_node(node->children); - node = inode->index(tsd_tsdn(tsd), mib, miblen, mib[i]); + node = inode->index(tsdn, mib, miblen, mib[i]); if (node == NULL) { ret = ENOENT; goto label_return; } } } + assert(ending_nodep != NULL); + *ending_nodep = node; + ret = 0; + +label_return: + return(ret); +} + +int +ctl_bymib(tsd_t *tsd, const size_t *mib, size_t miblen, void *oldp, + size_t *oldlenp, void *newp, size_t newlen) { + int ret; + const ctl_named_node_t *node; + + if (!ctl_initialized && ctl_init(tsd)) { + ret = EAGAIN; + goto label_return; + } + + ret = ctl_lookupbymib(tsd_tsdn(tsd), &node, mib, miblen); + if (ret != 0) { + goto label_return; + } /* Call the ctl function. */ if (node && node->ctl) { @@ -1383,6 +1660,81 @@ return(ret); } +int +ctl_mibnametomib(tsd_t *tsd, size_t *mib, size_t miblen, const char *name, + size_t *miblenp) { + int ret; + const ctl_named_node_t *node; + + if (!ctl_initialized && ctl_init(tsd)) { + ret = EAGAIN; + goto label_return; + } + + ret = ctl_lookupbymib(tsd_tsdn(tsd), &node, mib, miblen); + if (ret != 0) { + goto label_return; + } + if (node == NULL || node->ctl != NULL) { + ret = ENOENT; + goto label_return; + } + + assert(miblenp != NULL); + assert(*miblenp >= miblen); + *miblenp -= miblen; + ret = ctl_lookup(tsd_tsdn(tsd), node, name, NULL, mib + miblen, + miblenp); + *miblenp += miblen; +label_return: + return(ret); +} + +int +ctl_bymibname(tsd_t *tsd, size_t *mib, size_t miblen, const char *name, + size_t *miblenp, void *oldp, size_t *oldlenp, void *newp, size_t newlen) { + int ret; + const ctl_named_node_t *node; + + if (!ctl_initialized && ctl_init(tsd)) { + ret = EAGAIN; + goto label_return; + } + + ret = ctl_lookupbymib(tsd_tsdn(tsd), &node, mib, miblen); + if (ret != 0) { + goto label_return; + } + if (node == NULL || node->ctl != NULL) { + ret = ENOENT; + goto label_return; + } + + assert(miblenp != NULL); + assert(*miblenp >= miblen); + *miblenp -= miblen; + /* + * The same node supplies the starting node and stores the ending node. + */ + ret = ctl_lookup(tsd_tsdn(tsd), node, name, &node, mib + miblen, + miblenp); + *miblenp += miblen; + if (ret != 0) { + goto label_return; + } + + if (node != NULL && node->ctl) { + ret = node->ctl(tsd, mib, *miblenp, oldp, oldlenp, newp, + newlen); + } else { + /* The name refers to a partial path through the ctl tree. */ + ret = ENOENT; + } + +label_return: + return(ret); +} + bool ctl_boot(void) { if (malloc_mutex_init(&ctl_mtx, "ctl", WITNESS_RANK_CTL, @@ -1410,6 +1762,11 @@ malloc_mutex_postfork_child(tsdn, &ctl_mtx); } +void +ctl_mtx_assert_held(tsdn_t *tsdn) { + malloc_mutex_assert_owner(tsdn, &ctl_mtx); +} + /******************************************************************************/ /* *_ctl() functions. */ @@ -1427,6 +1784,7 @@ } \ } while (0) +/* Can read or write, but not both. */ #define READ_XOR_WRITE() do { \ if ((oldp != NULL && oldlenp != NULL) && (newp != NULL || \ newlen != 0)) { \ @@ -1435,12 +1793,31 @@ } \ } while (0) +/* Can neither read nor write. */ +#define NEITHER_READ_NOR_WRITE() do { \ + if (oldp != NULL || oldlenp != NULL || newp != NULL || \ + newlen != 0) { \ + ret = EPERM; \ + goto label_return; \ + } \ +} while (0) + +/* Verify that the space provided is enough. */ +#define VERIFY_READ(t) do { \ + if (oldp == NULL || oldlenp == NULL || *oldlenp != sizeof(t)) { \ + *oldlenp = 0; \ + ret = EINVAL; \ + goto label_return; \ + } \ +} while (0) + #define READ(v, t) do { \ if (oldp != NULL && oldlenp != NULL) { \ if (*oldlenp != sizeof(t)) { \ size_t copylen = (sizeof(t) <= *oldlenp) \ ? sizeof(t) : *oldlenp; \ memcpy(oldp, (void *)&(v), copylen); \ + *oldlenp = copylen; \ ret = EINVAL; \ goto label_return; \ } \ @@ -1458,6 +1835,14 @@ } \ } while (0) +#define ASSURED_WRITE(v, t) do { \ + if (newp == NULL || newlen != sizeof(t)) { \ + ret = EINVAL; \ + goto label_return; \ + } \ + (v) = *(t *)newp; \ +} while (0) + #define MIB_UNSIGNED(v, i) do { \ if (mib[i] > UINT_MAX) { \ ret = EFAULT; \ @@ -1497,8 +1882,8 @@ #define CTL_RO_CGEN(c, n, v, t) \ static int \ -n##_ctl(tsd_t *tsd, const size_t *mib, size_t miblen, \ - void *oldp, size_t *oldlenp, void *newp, size_t newlen) { \ +n##_ctl(tsd_t *tsd, const size_t *mib, size_t miblen, \ + void *oldp, size_t *oldlenp, void *newp, size_t newlen) { \ int ret; \ t oldval; \ \ @@ -1540,8 +1925,8 @@ */ #define CTL_RO_NL_CGEN(c, n, v, t) \ static int \ -n##_ctl(tsd_t *tsd, const size_t *mib, size_t miblen, \ - void *oldp, size_t *oldlenp, void *newp, size_t newlen) { \ +n##_ctl(tsd_t *tsd, const size_t *mib, size_t miblen, \ + void *oldp, size_t *oldlenp, void *newp, size_t newlen) { \ int ret; \ t oldval; \ \ @@ -1559,8 +1944,8 @@ #define CTL_RO_NL_GEN(n, v, t) \ static int \ -n##_ctl(tsd_t *tsd, const size_t *mib, size_t miblen, \ - void *oldp, size_t *oldlenp, void *newp, size_t newlen) { \ +n##_ctl(tsd_t *tsd, const size_t *mib, size_t miblen, \ + void *oldp, size_t *oldlenp, void *newp, size_t newlen) { \ int ret; \ t oldval; \ \ @@ -1573,29 +1958,10 @@ return ret; \ } -#define CTL_TSD_RO_NL_CGEN(c, n, m, t) \ -static int \ -n##_ctl(tsd_t *tsd, const size_t *mib, size_t miblen, void *oldp, \ - size_t *oldlenp, void *newp, size_t newlen) { \ - int ret; \ - t oldval; \ - \ - if (!(c)) { \ - return ENOENT; \ - } \ - READONLY(); \ - oldval = (m(tsd)); \ - READ(oldval, t); \ - \ - ret = 0; \ -label_return: \ - return ret; \ -} - #define CTL_RO_CONFIG_GEN(n, t) \ static int \ -n##_ctl(tsd_t *tsd, const size_t *mib, size_t miblen, \ - void *oldp, size_t *oldlenp, void *newp, size_t newlen) { \ +n##_ctl(tsd_t *tsd, const size_t *mib, size_t miblen, \ + void *oldp, size_t *oldlenp, void *newp, size_t newlen) { \ int ret; \ t oldval; \ \ @@ -1761,7 +2127,34 @@ CTL_RO_NL_GEN(opt_abort, opt_abort, bool) CTL_RO_NL_GEN(opt_abort_conf, opt_abort_conf, bool) +CTL_RO_NL_GEN(opt_cache_oblivious, opt_cache_oblivious, bool) +CTL_RO_NL_GEN(opt_trust_madvise, opt_trust_madvise, bool) CTL_RO_NL_GEN(opt_confirm_conf, opt_confirm_conf, bool) + +/* HPA options. */ +CTL_RO_NL_GEN(opt_hpa, opt_hpa, bool) +CTL_RO_NL_GEN(opt_hpa_hugification_threshold, + opt_hpa_opts.hugification_threshold, size_t) +CTL_RO_NL_GEN(opt_hpa_hugify_delay_ms, opt_hpa_opts.hugify_delay_ms, uint64_t) +CTL_RO_NL_GEN(opt_hpa_min_purge_interval_ms, opt_hpa_opts.min_purge_interval_ms, + uint64_t) + +/* + * This will have to change before we publicly document this option; fxp_t and + * its representation are internal implementation details. + */ +CTL_RO_NL_GEN(opt_hpa_dirty_mult, opt_hpa_opts.dirty_mult, fxp_t) +CTL_RO_NL_GEN(opt_hpa_slab_max_alloc, opt_hpa_opts.slab_max_alloc, size_t) + +/* HPA SEC options */ +CTL_RO_NL_GEN(opt_hpa_sec_nshards, opt_hpa_sec_opts.nshards, size_t) +CTL_RO_NL_GEN(opt_hpa_sec_max_alloc, opt_hpa_sec_opts.max_alloc, size_t) +CTL_RO_NL_GEN(opt_hpa_sec_max_bytes, opt_hpa_sec_opts.max_bytes, size_t) +CTL_RO_NL_GEN(opt_hpa_sec_bytes_after_flush, opt_hpa_sec_opts.bytes_after_flush, + size_t) +CTL_RO_NL_GEN(opt_hpa_sec_batch_fill_extra, opt_hpa_sec_opts.batch_fill_extra, + size_t) + CTL_RO_NL_GEN(opt_metadata_thp, metadata_thp_mode_names[opt_metadata_thp], const char *) CTL_RO_NL_GEN(opt_retain, opt_retain, bool) @@ -1769,6 +2162,7 @@ CTL_RO_NL_GEN(opt_narenas, opt_narenas, unsigned) CTL_RO_NL_GEN(opt_percpu_arena, percpu_arena_mode_names[opt_percpu_arena], const char *) +CTL_RO_NL_GEN(opt_mutex_max_spin, opt_mutex_max_spin, int64_t) CTL_RO_NL_GEN(opt_oversize_threshold, opt_oversize_threshold, size_t) CTL_RO_NL_GEN(opt_background_thread, opt_background_thread, bool) CTL_RO_NL_GEN(opt_max_background_threads, opt_max_background_threads, size_t) @@ -1776,15 +2170,31 @@ CTL_RO_NL_GEN(opt_muzzy_decay_ms, opt_muzzy_decay_ms, ssize_t) CTL_RO_NL_GEN(opt_stats_print, opt_stats_print, bool) CTL_RO_NL_GEN(opt_stats_print_opts, opt_stats_print_opts, const char *) +CTL_RO_NL_GEN(opt_stats_interval, opt_stats_interval, int64_t) +CTL_RO_NL_GEN(opt_stats_interval_opts, opt_stats_interval_opts, const char *) CTL_RO_NL_CGEN(config_fill, opt_junk, opt_junk, const char *) CTL_RO_NL_CGEN(config_fill, opt_zero, opt_zero, bool) CTL_RO_NL_CGEN(config_utrace, opt_utrace, opt_utrace, bool) CTL_RO_NL_CGEN(config_xmalloc, opt_xmalloc, opt_xmalloc, bool) +CTL_RO_NL_CGEN(config_enable_cxx, opt_experimental_infallible_new, + opt_experimental_infallible_new, bool) CTL_RO_NL_GEN(opt_tcache, opt_tcache, bool) +CTL_RO_NL_GEN(opt_tcache_max, opt_tcache_max, size_t) +CTL_RO_NL_GEN(opt_tcache_nslots_small_min, opt_tcache_nslots_small_min, + unsigned) +CTL_RO_NL_GEN(opt_tcache_nslots_small_max, opt_tcache_nslots_small_max, + unsigned) +CTL_RO_NL_GEN(opt_tcache_nslots_large, opt_tcache_nslots_large, unsigned) +CTL_RO_NL_GEN(opt_lg_tcache_nslots_mul, opt_lg_tcache_nslots_mul, ssize_t) +CTL_RO_NL_GEN(opt_tcache_gc_incr_bytes, opt_tcache_gc_incr_bytes, size_t) +CTL_RO_NL_GEN(opt_tcache_gc_delay_bytes, opt_tcache_gc_delay_bytes, size_t) +CTL_RO_NL_GEN(opt_lg_tcache_flush_small_div, opt_lg_tcache_flush_small_div, + unsigned) +CTL_RO_NL_GEN(opt_lg_tcache_flush_large_div, opt_lg_tcache_flush_large_div, + unsigned) CTL_RO_NL_GEN(opt_thp, thp_mode_names[opt_thp], const char *) CTL_RO_NL_GEN(opt_lg_extent_max_active_fit, opt_lg_extent_max_active_fit, size_t) -CTL_RO_NL_GEN(opt_lg_tcache_max, opt_lg_tcache_max, ssize_t) CTL_RO_NL_CGEN(config_prof, opt_prof, opt_prof, bool) CTL_RO_NL_CGEN(config_prof, opt_prof_prefix, opt_prof_prefix, const char *) CTL_RO_NL_CGEN(config_prof, opt_prof_active, opt_prof_active, bool) @@ -1796,6 +2206,18 @@ CTL_RO_NL_CGEN(config_prof, opt_prof_gdump, opt_prof_gdump, bool) CTL_RO_NL_CGEN(config_prof, opt_prof_final, opt_prof_final, bool) CTL_RO_NL_CGEN(config_prof, opt_prof_leak, opt_prof_leak, bool) +CTL_RO_NL_CGEN(config_prof, opt_prof_leak_error, opt_prof_leak_error, bool) +CTL_RO_NL_CGEN(config_prof, opt_prof_recent_alloc_max, + opt_prof_recent_alloc_max, ssize_t) +CTL_RO_NL_CGEN(config_prof, opt_prof_stats, opt_prof_stats, bool) +CTL_RO_NL_CGEN(config_prof, opt_prof_sys_thread_name, opt_prof_sys_thread_name, + bool) +CTL_RO_NL_CGEN(config_prof, opt_prof_time_res, + prof_time_res_mode_names[opt_prof_time_res], const char *) +CTL_RO_NL_CGEN(config_uaf_detection, opt_lg_san_uaf_align, + opt_lg_san_uaf_align, ssize_t) +CTL_RO_NL_GEN(opt_zero_realloc, + zero_realloc_mode_names[opt_zero_realloc_action], const char *) /******************************************************************************/ @@ -1843,10 +2265,11 @@ goto label_return; } /* Set new arena/tcache associations. */ - arena_migrate(tsd, oldind, newind); + arena_migrate(tsd, oldarena, newarena); if (tcache_available(tsd)) { tcache_arena_reassociate(tsd_tsdn(tsd), - tsd_tcachep_get(tsd), newarena); + tsd_tcache_slowp_get(tsd), tsd_tcachep_get(tsd), + newarena); } } @@ -1855,14 +2278,10 @@ return ret; } -CTL_TSD_RO_NL_CGEN(config_stats, thread_allocated, tsd_thread_allocated_get, - uint64_t) -CTL_TSD_RO_NL_CGEN(config_stats, thread_allocatedp, tsd_thread_allocatedp_get, - uint64_t *) -CTL_TSD_RO_NL_CGEN(config_stats, thread_deallocated, tsd_thread_deallocated_get, - uint64_t) -CTL_TSD_RO_NL_CGEN(config_stats, thread_deallocatedp, - tsd_thread_deallocatedp_get, uint64_t *) +CTL_RO_NL_GEN(thread_allocated, tsd_thread_allocated_get(tsd), uint64_t) +CTL_RO_NL_GEN(thread_allocatedp, tsd_thread_allocatedp_get(tsd), uint64_t *) +CTL_RO_NL_GEN(thread_deallocated, tsd_thread_deallocated_get(tsd), uint64_t) +CTL_RO_NL_GEN(thread_deallocatedp, tsd_thread_deallocatedp_get(tsd), uint64_t *) static int thread_tcache_enabled_ctl(tsd_t *tsd, const size_t *mib, @@ -1897,8 +2316,7 @@ goto label_return; } - READONLY(); - WRITEONLY(); + NEITHER_READ_NOR_WRITE(); tcache_flush(tsd); @@ -1907,13 +2325,45 @@ return ret; } +static int +thread_peak_read_ctl(tsd_t *tsd, const size_t *mib, + size_t miblen, void *oldp, size_t *oldlenp, void *newp, + size_t newlen) { + int ret; + if (!config_stats) { + return ENOENT; + } + READONLY(); + peak_event_update(tsd); + uint64_t result = peak_event_max(tsd); + READ(result, uint64_t); + ret = 0; +label_return: + return ret; +} + +static int +thread_peak_reset_ctl(tsd_t *tsd, const size_t *mib, + size_t miblen, void *oldp, size_t *oldlenp, void *newp, + size_t newlen) { + int ret; + if (!config_stats) { + return ENOENT; + } + NEITHER_READ_NOR_WRITE(); + peak_event_zero(tsd); + ret = 0; +label_return: + return ret; +} + static int thread_prof_name_ctl(tsd_t *tsd, const size_t *mib, size_t miblen, void *oldp, size_t *oldlenp, void *newp, size_t newlen) { int ret; - if (!config_prof) { + if (!config_prof || !opt_prof) { return ENOENT; } @@ -1950,8 +2400,12 @@ return ENOENT; } - oldval = prof_thread_active_get(tsd); + oldval = opt_prof ? prof_thread_active_get(tsd) : false; if (newp != NULL) { + if (!opt_prof) { + ret = ENOENT; + goto label_return; + } if (newlen != sizeof(bool)) { ret = EINVAL; goto label_return; @@ -1968,6 +2422,39 @@ return ret; } +static int +thread_idle_ctl(tsd_t *tsd, const size_t *mib, + size_t miblen, void *oldp, size_t *oldlenp, void *newp, + size_t newlen) { + int ret; + + NEITHER_READ_NOR_WRITE(); + + if (tcache_available(tsd)) { + tcache_flush(tsd); + } + /* + * This heuristic is perhaps not the most well-considered. But it + * matches the only idling policy we have experience with in the status + * quo. Over time we should investigate more principled approaches. + */ + if (opt_narenas > ncpus * 2) { + arena_t *arena = arena_choose(tsd, NULL); + if (arena != NULL) { + arena_decay(tsd_tsdn(tsd), arena, false, true); + } + /* + * The missing arena case is not actually an error; a thread + * might be idle before it associates itself to one. This is + * unusual, but not wrong. + */ + } + + ret = 0; +label_return: + return ret; +} + /******************************************************************************/ static int @@ -1977,7 +2464,8 @@ unsigned tcache_ind; READONLY(); - if (tcaches_create(tsd, &tcache_ind)) { + VERIFY_READ(unsigned); + if (tcaches_create(tsd, b0get(), &tcache_ind)) { ret = EFAULT; goto label_return; } @@ -1995,12 +2483,7 @@ unsigned tcache_ind; WRITEONLY(); - tcache_ind = UINT_MAX; - WRITE(tcache_ind, unsigned); - if (tcache_ind == UINT_MAX) { - ret = EFAULT; - goto label_return; - } + ASSURED_WRITE(tcache_ind, unsigned); tcaches_flush(tsd, tcache_ind); ret = 0; @@ -2015,12 +2498,7 @@ unsigned tcache_ind; WRITEONLY(); - tcache_ind = UINT_MAX; - WRITE(tcache_ind, unsigned); - if (tcache_ind == UINT_MAX) { - ret = EFAULT; - goto label_return; - } + ASSURED_WRITE(tcache_ind, unsigned); tcaches_destroy(tsd, tcache_ind); ret = 0; @@ -2105,8 +2583,7 @@ int ret; unsigned arena_ind; - READONLY(); - WRITEONLY(); + NEITHER_READ_NOR_WRITE(); MIB_UNSIGNED(arena_ind, 1); arena_i_decay(tsd_tsdn(tsd), arena_ind, false); @@ -2121,8 +2598,7 @@ int ret; unsigned arena_ind; - READONLY(); - WRITEONLY(); + NEITHER_READ_NOR_WRITE(); MIB_UNSIGNED(arena_ind, 1); arena_i_decay(tsd_tsdn(tsd), arena_ind, true); @@ -2137,8 +2613,7 @@ arena_t **arena) { int ret; - READONLY(); - WRITEONLY(); + NEITHER_READ_NOR_WRITE(); MIB_UNSIGNED(*arena_ind, 1); *arena = arena_get(tsd_tsdn(tsd), *arena_ind, false); @@ -2211,6 +2686,8 @@ arena_t *arena; ctl_arena_t *ctl_darena, *ctl_arena; + malloc_mutex_lock(tsd_tsdn(tsd), &ctl_mtx); + ret = arena_i_reset_destroy_helper(tsd, mib, miblen, oldp, oldlenp, newp, newlen, &arena_ind, &arena); if (ret != 0) { @@ -2241,6 +2718,8 @@ assert(ret == 0); label_return: + malloc_mutex_unlock(tsd_tsdn(tsd), &ctl_mtx); + return ret; } @@ -2306,8 +2785,40 @@ } static int -arena_i_decay_ms_ctl_impl(tsd_t *tsd, const size_t *mib, size_t miblen, - void *oldp, size_t *oldlenp, void *newp, size_t newlen, bool dirty) { +arena_i_oversize_threshold_ctl(tsd_t *tsd, const size_t *mib, size_t miblen, + void *oldp, size_t *oldlenp, void *newp, size_t newlen) { + int ret; + + unsigned arena_ind; + MIB_UNSIGNED(arena_ind, 1); + + arena_t *arena = arena_get(tsd_tsdn(tsd), arena_ind, false); + if (arena == NULL) { + ret = EFAULT; + goto label_return; + } + + if (oldp != NULL && oldlenp != NULL) { + size_t oldval = atomic_load_zu( + &arena->pa_shard.pac.oversize_threshold, ATOMIC_RELAXED); + READ(oldval, size_t); + } + if (newp != NULL) { + if (newlen != sizeof(size_t)) { + ret = EINVAL; + goto label_return; + } + atomic_store_zu(&arena->pa_shard.pac.oversize_threshold, + *(size_t *)newp, ATOMIC_RELAXED); + } + ret = 0; +label_return: + return ret; +} + +static int +arena_i_decay_ms_ctl_impl(tsd_t *tsd, const size_t *mib, size_t miblen, + void *oldp, size_t *oldlenp, void *newp, size_t newlen, bool dirty) { int ret; unsigned arena_ind; arena_t *arena; @@ -2318,10 +2829,10 @@ ret = EFAULT; goto label_return; } + extent_state_t state = dirty ? extent_state_dirty : extent_state_muzzy; if (oldp != NULL && oldlenp != NULL) { - size_t oldval = dirty ? arena_dirty_decay_ms_get(arena) : - arena_muzzy_decay_ms_get(arena); + size_t oldval = arena_decay_ms_get(arena, state); READ(oldval, ssize_t); } if (newp != NULL) { @@ -2340,9 +2851,9 @@ goto label_return; } } - if (dirty ? arena_dirty_decay_ms_set(tsd_tsdn(tsd), arena, - *(ssize_t *)newp) : arena_muzzy_decay_ms_set(tsd_tsdn(tsd), - arena, *(ssize_t *)newp)) { + + if (arena_decay_ms_set(tsd_tsdn(tsd), arena, state, + *(ssize_t *)newp)) { ret = EFAULT; goto label_return; } @@ -2385,15 +2896,18 @@ goto label_return; } old_extent_hooks = - (extent_hooks_t *)&extent_hooks_default; + (extent_hooks_t *)&ehooks_default_extent_hooks; READ(old_extent_hooks, extent_hooks_t *); if (newp != NULL) { /* Initialize a new arena as a side effect. */ extent_hooks_t *new_extent_hooks JEMALLOC_CC_SILENCE_INIT(NULL); WRITE(new_extent_hooks, extent_hooks_t *); + arena_config_t config = arena_config_default; + config.extent_hooks = new_extent_hooks; + arena = arena_init(tsd_tsdn(tsd), arena_ind, - new_extent_hooks); + &config); if (arena == NULL) { ret = EFAULT; goto label_return; @@ -2404,11 +2918,13 @@ extent_hooks_t *new_extent_hooks JEMALLOC_CC_SILENCE_INIT(NULL); WRITE(new_extent_hooks, extent_hooks_t *); - old_extent_hooks = extent_hooks_set(tsd, arena, - new_extent_hooks); + old_extent_hooks = arena_set_extent_hooks(tsd, + arena, new_extent_hooks); READ(old_extent_hooks, extent_hooks_t *); } else { - old_extent_hooks = extent_hooks_get(arena); + old_extent_hooks = + ehooks_get_extent_hooks_ptr( + arena_get_ehooks(arena)); READ(old_extent_hooks, extent_hooks_t *); } } @@ -2493,10 +3009,6 @@ malloc_mutex_lock(tsd_tsdn(tsd), &ctl_mtx); READONLY(); - if (*oldlenp != sizeof(unsigned)) { - ret = EINVAL; - goto label_return; - } narenas = ctl_arenas->narenas; READ(narenas, unsigned); @@ -2582,14 +3094,14 @@ arenas_create_ctl(tsd_t *tsd, const size_t *mib, size_t miblen, void *oldp, size_t *oldlenp, void *newp, size_t newlen) { int ret; - extent_hooks_t *extent_hooks; unsigned arena_ind; malloc_mutex_lock(tsd_tsdn(tsd), &ctl_mtx); - extent_hooks = (extent_hooks_t *)&extent_hooks_default; - WRITE(extent_hooks, extent_hooks_t *); - if ((arena_ind = ctl_arena_init(tsd, extent_hooks)) == UINT_MAX) { + VERIFY_READ(unsigned); + arena_config_t config = arena_config_default; + WRITE(config.extent_hooks, extent_hooks_t *); + if ((arena_ind = ctl_arena_init(tsd, &config)) == UINT_MAX) { ret = EAGAIN; goto label_return; } @@ -2601,6 +3113,30 @@ return ret; } +static int +experimental_arenas_create_ext_ctl(tsd_t *tsd, + const size_t *mib, size_t miblen, + void *oldp, size_t *oldlenp, void *newp, size_t newlen) { + int ret; + unsigned arena_ind; + + malloc_mutex_lock(tsd_tsdn(tsd), &ctl_mtx); + + arena_config_t config = arena_config_default; + VERIFY_READ(unsigned); + WRITE(config, arena_config_t); + + if ((arena_ind = ctl_arena_init(tsd, &config)) == UINT_MAX) { + ret = EAGAIN; + goto label_return; + } + READ(arena_ind, unsigned); + ret = 0; +label_return: + malloc_mutex_unlock(tsd_tsdn(tsd), &ctl_mtx); + return ret; +} + static int arenas_lookup_ctl(tsd_t *tsd, const size_t *mib, size_t miblen, void *oldp, size_t *oldlenp, void *newp, @@ -2608,20 +3144,22 @@ int ret; unsigned arena_ind; void *ptr; - extent_t *extent; + edata_t *edata; arena_t *arena; ptr = NULL; ret = EINVAL; malloc_mutex_lock(tsd_tsdn(tsd), &ctl_mtx); WRITE(ptr, void *); - extent = iealloc(tsd_tsdn(tsd), ptr); - if (extent == NULL) + edata = emap_edata_lookup(tsd_tsdn(tsd), &arena_emap_global, ptr); + if (edata == NULL) { goto label_return; + } - arena = extent_arena_get(extent); - if (arena == NULL) + arena = arena_get_from_edata(edata); + if (arena == NULL) { goto label_return; + } arena_ind = arena_ind_get(arena); READ(arena_ind, unsigned); @@ -2646,6 +3184,10 @@ } if (newp != NULL) { + if (!opt_prof) { + ret = ENOENT; + goto label_return; + } if (newlen != sizeof(bool)) { ret = EINVAL; goto label_return; @@ -2653,7 +3195,8 @@ oldval = prof_thread_active_init_set(tsd_tsdn(tsd), *(bool *)newp); } else { - oldval = prof_thread_active_init_get(tsd_tsdn(tsd)); + oldval = opt_prof ? prof_thread_active_init_get(tsd_tsdn(tsd)) : + false; } READ(oldval, bool); @@ -2669,7 +3212,8 @@ bool oldval; if (!config_prof) { - return ENOENT; + ret = ENOENT; + goto label_return; } if (newp != NULL) { @@ -2677,9 +3221,20 @@ ret = EINVAL; goto label_return; } - oldval = prof_active_set(tsd_tsdn(tsd), *(bool *)newp); + bool val = *(bool *)newp; + if (!opt_prof) { + if (val) { + ret = ENOENT; + goto label_return; + } else { + /* No change needed (already off). */ + oldval = false; + } + } else { + oldval = prof_active_set(tsd_tsdn(tsd), val); + } } else { - oldval = prof_active_get(tsd_tsdn(tsd)); + oldval = opt_prof ? prof_active_get(tsd_tsdn(tsd)) : false; } READ(oldval, bool); @@ -2694,7 +3249,7 @@ int ret; const char *filename = NULL; - if (!config_prof) { + if (!config_prof || !opt_prof) { return ENOENT; } @@ -2722,13 +3277,17 @@ } if (newp != NULL) { + if (!opt_prof) { + ret = ENOENT; + goto label_return; + } if (newlen != sizeof(bool)) { ret = EINVAL; goto label_return; } oldval = prof_gdump_set(tsd_tsdn(tsd), *(bool *)newp); } else { - oldval = prof_gdump_get(tsd_tsdn(tsd)); + oldval = opt_prof ? prof_gdump_get(tsd_tsdn(tsd)) : false; } READ(oldval, bool); @@ -2737,13 +3296,33 @@ return ret; } +static int +prof_prefix_ctl(tsd_t *tsd, const size_t *mib, size_t miblen, + void *oldp, size_t *oldlenp, void *newp, size_t newlen) { + int ret; + const char *prefix = NULL; + + if (!config_prof || !opt_prof) { + return ENOENT; + } + + malloc_mutex_lock(tsd_tsdn(tsd), &ctl_mtx); + WRITEONLY(); + WRITE(prefix, const char *); + + ret = prof_prefix_set(tsd_tsdn(tsd), prefix) ? EFAULT : 0; +label_return: + malloc_mutex_unlock(tsd_tsdn(tsd), &ctl_mtx); + return ret; +} + static int prof_reset_ctl(tsd_t *tsd, const size_t *mib, size_t miblen, void *oldp, size_t *oldlenp, void *newp, size_t newlen) { int ret; size_t lg_sample = lg_prof_sample; - if (!config_prof) { + if (!config_prof || !opt_prof) { return ENOENT; } @@ -2770,7 +3349,7 @@ const char *filename = NULL; - if (!config_prof) { + if (!config_prof || !opt_prof) { return ENOENT; } @@ -2790,7 +3369,7 @@ static int prof_log_stop_ctl(tsd_t *tsd, const size_t *mib, size_t miblen, void *oldp, size_t *oldlenp, void *newp, size_t newlen) { - if (!config_prof) { + if (!config_prof || !opt_prof) { return ENOENT; } @@ -2801,6 +3380,87 @@ return 0; } +static int +experimental_hooks_prof_backtrace_ctl(tsd_t *tsd, const size_t *mib, + size_t miblen, void *oldp, size_t *oldlenp, void *newp, size_t newlen) { + int ret; + + if (oldp == NULL && newp == NULL) { + ret = EINVAL; + goto label_return; + } + if (oldp != NULL) { + prof_backtrace_hook_t old_hook = + prof_backtrace_hook_get(); + READ(old_hook, prof_backtrace_hook_t); + } + if (newp != NULL) { + if (!opt_prof) { + ret = ENOENT; + goto label_return; + } + prof_backtrace_hook_t new_hook JEMALLOC_CC_SILENCE_INIT(NULL); + WRITE(new_hook, prof_backtrace_hook_t); + if (new_hook == NULL) { + ret = EINVAL; + goto label_return; + } + prof_backtrace_hook_set(new_hook); + } + ret = 0; +label_return: + return ret; +} + +static int +experimental_hooks_prof_dump_ctl(tsd_t *tsd, const size_t *mib, + size_t miblen, void *oldp, size_t *oldlenp, void *newp, size_t newlen) { + int ret; + + if (oldp == NULL && newp == NULL) { + ret = EINVAL; + goto label_return; + } + if (oldp != NULL) { + prof_dump_hook_t old_hook = + prof_dump_hook_get(); + READ(old_hook, prof_dump_hook_t); + } + if (newp != NULL) { + if (!opt_prof) { + ret = ENOENT; + goto label_return; + } + prof_dump_hook_t new_hook JEMALLOC_CC_SILENCE_INIT(NULL); + WRITE(new_hook, prof_dump_hook_t); + prof_dump_hook_set(new_hook); + } + ret = 0; +label_return: + return ret; +} + +/* For integration test purpose only. No plan to move out of experimental. */ +static int +experimental_hooks_safety_check_abort_ctl(tsd_t *tsd, const size_t *mib, + size_t miblen, void *oldp, size_t *oldlenp, void *newp, size_t newlen) { + int ret; + + WRITEONLY(); + if (newp != NULL) { + if (newlen != sizeof(safety_check_abort_hook_t)) { + ret = EINVAL; + goto label_return; + } + safety_check_abort_hook_t hook JEMALLOC_CC_SILENCE_INIT(NULL); + WRITE(hook, safety_check_abort_hook_t); + safety_check_set_abort(hook); + } + ret = 0; +label_return: + return ret; +} + /******************************************************************************/ CTL_RO_CGEN(config_stats, stats_allocated, ctl_stats->allocated, size_t) @@ -2818,6 +3478,9 @@ CTL_RO_CGEN(config_stats, stats_background_thread_run_interval, nstime_ns(&ctl_stats->background_thread.run_interval), uint64_t) +CTL_RO_CGEN(config_stats, stats_zero_reallocs, + atomic_load_zu(&zero_realloc_count, ATOMIC_RELAXED), size_t) + CTL_RO_GEN(stats_arenas_i_dss, arenas_i(mib[2])->dss, const char *) CTL_RO_GEN(stats_arenas_i_dirty_decay_ms, arenas_i(mib[2])->dirty_decay_ms, ssize_t) @@ -2830,55 +3493,61 @@ CTL_RO_GEN(stats_arenas_i_pdirty, arenas_i(mib[2])->pdirty, size_t) CTL_RO_GEN(stats_arenas_i_pmuzzy, arenas_i(mib[2])->pmuzzy, size_t) CTL_RO_CGEN(config_stats, stats_arenas_i_mapped, - atomic_load_zu(&arenas_i(mib[2])->astats->astats.mapped, ATOMIC_RELAXED), - size_t) + arenas_i(mib[2])->astats->astats.mapped, size_t) CTL_RO_CGEN(config_stats, stats_arenas_i_retained, - atomic_load_zu(&arenas_i(mib[2])->astats->astats.retained, ATOMIC_RELAXED), - size_t) + arenas_i(mib[2])->astats->astats.pa_shard_stats.pac_stats.retained, size_t) CTL_RO_CGEN(config_stats, stats_arenas_i_extent_avail, - atomic_load_zu(&arenas_i(mib[2])->astats->astats.extent_avail, - ATOMIC_RELAXED), - size_t) + arenas_i(mib[2])->astats->astats.pa_shard_stats.edata_avail, size_t) CTL_RO_CGEN(config_stats, stats_arenas_i_dirty_npurge, - ctl_arena_stats_read_u64( - &arenas_i(mib[2])->astats->astats.decay_dirty.npurge), uint64_t) + locked_read_u64_unsynchronized( + &arenas_i(mib[2])->astats->astats.pa_shard_stats.pac_stats.decay_dirty.npurge), + uint64_t) CTL_RO_CGEN(config_stats, stats_arenas_i_dirty_nmadvise, - ctl_arena_stats_read_u64( - &arenas_i(mib[2])->astats->astats.decay_dirty.nmadvise), uint64_t) + locked_read_u64_unsynchronized( + &arenas_i(mib[2])->astats->astats.pa_shard_stats.pac_stats.decay_dirty.nmadvise), + uint64_t) CTL_RO_CGEN(config_stats, stats_arenas_i_dirty_purged, - ctl_arena_stats_read_u64( - &arenas_i(mib[2])->astats->astats.decay_dirty.purged), uint64_t) + locked_read_u64_unsynchronized( + &arenas_i(mib[2])->astats->astats.pa_shard_stats.pac_stats.decay_dirty.purged), + uint64_t) CTL_RO_CGEN(config_stats, stats_arenas_i_muzzy_npurge, - ctl_arena_stats_read_u64( - &arenas_i(mib[2])->astats->astats.decay_muzzy.npurge), uint64_t) + locked_read_u64_unsynchronized( + &arenas_i(mib[2])->astats->astats.pa_shard_stats.pac_stats.decay_muzzy.npurge), + uint64_t) CTL_RO_CGEN(config_stats, stats_arenas_i_muzzy_nmadvise, - ctl_arena_stats_read_u64( - &arenas_i(mib[2])->astats->astats.decay_muzzy.nmadvise), uint64_t) + locked_read_u64_unsynchronized( + &arenas_i(mib[2])->astats->astats.pa_shard_stats.pac_stats.decay_muzzy.nmadvise), + uint64_t) CTL_RO_CGEN(config_stats, stats_arenas_i_muzzy_purged, - ctl_arena_stats_read_u64( - &arenas_i(mib[2])->astats->astats.decay_muzzy.purged), uint64_t) + locked_read_u64_unsynchronized( + &arenas_i(mib[2])->astats->astats.pa_shard_stats.pac_stats.decay_muzzy.purged), + uint64_t) CTL_RO_CGEN(config_stats, stats_arenas_i_base, - atomic_load_zu(&arenas_i(mib[2])->astats->astats.base, ATOMIC_RELAXED), + arenas_i(mib[2])->astats->astats.base, size_t) CTL_RO_CGEN(config_stats, stats_arenas_i_internal, atomic_load_zu(&arenas_i(mib[2])->astats->astats.internal, ATOMIC_RELAXED), size_t) CTL_RO_CGEN(config_stats, stats_arenas_i_metadata_thp, - atomic_load_zu(&arenas_i(mib[2])->astats->astats.metadata_thp, - ATOMIC_RELAXED), size_t) + arenas_i(mib[2])->astats->astats.metadata_thp, size_t) CTL_RO_CGEN(config_stats, stats_arenas_i_tcache_bytes, - atomic_load_zu(&arenas_i(mib[2])->astats->astats.tcache_bytes, - ATOMIC_RELAXED), size_t) + arenas_i(mib[2])->astats->astats.tcache_bytes, size_t) +CTL_RO_CGEN(config_stats, stats_arenas_i_tcache_stashed_bytes, + arenas_i(mib[2])->astats->astats.tcache_stashed_bytes, size_t) CTL_RO_CGEN(config_stats, stats_arenas_i_resident, - atomic_load_zu(&arenas_i(mib[2])->astats->astats.resident, ATOMIC_RELAXED), + arenas_i(mib[2])->astats->astats.resident, size_t) CTL_RO_CGEN(config_stats, stats_arenas_i_abandoned_vm, - atomic_load_zu(&arenas_i(mib[2])->astats->astats.abandoned_vm, + atomic_load_zu( + &arenas_i(mib[2])->astats->astats.pa_shard_stats.pac_stats.abandoned_vm, ATOMIC_RELAXED), size_t) +CTL_RO_CGEN(config_stats, stats_arenas_i_hpa_sec_bytes, + arenas_i(mib[2])->astats->secstats.bytes, size_t) + CTL_RO_CGEN(config_stats, stats_arenas_i_small_allocated, arenas_i(mib[2])->astats->allocated_small, size_t) CTL_RO_CGEN(config_stats, stats_arenas_i_small_nmalloc, @@ -2892,27 +3561,21 @@ CTL_RO_CGEN(config_stats, stats_arenas_i_small_nflushes, arenas_i(mib[2])->astats->nflushes_small, uint64_t) CTL_RO_CGEN(config_stats, stats_arenas_i_large_allocated, - atomic_load_zu(&arenas_i(mib[2])->astats->astats.allocated_large, - ATOMIC_RELAXED), size_t) + arenas_i(mib[2])->astats->astats.allocated_large, size_t) CTL_RO_CGEN(config_stats, stats_arenas_i_large_nmalloc, - ctl_arena_stats_read_u64( - &arenas_i(mib[2])->astats->astats.nmalloc_large), uint64_t) + arenas_i(mib[2])->astats->astats.nmalloc_large, uint64_t) CTL_RO_CGEN(config_stats, stats_arenas_i_large_ndalloc, - ctl_arena_stats_read_u64( - &arenas_i(mib[2])->astats->astats.ndalloc_large), uint64_t) + arenas_i(mib[2])->astats->astats.ndalloc_large, uint64_t) CTL_RO_CGEN(config_stats, stats_arenas_i_large_nrequests, - ctl_arena_stats_read_u64( - &arenas_i(mib[2])->astats->astats.nrequests_large), uint64_t) + arenas_i(mib[2])->astats->astats.nrequests_large, uint64_t) /* * Note: "nmalloc_large" here instead of "nfills" in the read. This is * intentional (large has no batch fill). */ CTL_RO_CGEN(config_stats, stats_arenas_i_large_nfills, - ctl_arena_stats_read_u64( - &arenas_i(mib[2])->astats->astats.nmalloc_large), uint64_t) + arenas_i(mib[2])->astats->astats.nmalloc_large, uint64_t) CTL_RO_CGEN(config_stats, stats_arenas_i_large_nflushes, - ctl_arena_stats_read_u64( - &arenas_i(mib[2])->astats->astats.nflushes_large), uint64_t) + arenas_i(mib[2])->astats->astats.nflushes_large, uint64_t) /* Lock profiling related APIs below. */ #define RO_MUTEX_CTL_GEN(n, l) \ @@ -2972,9 +3635,13 @@ } if (config_prof && opt_prof) { MUTEX_PROF_RESET(bt2gctx_mtx); + MUTEX_PROF_RESET(tdatas_mtx); + MUTEX_PROF_RESET(prof_dump_mtx); + MUTEX_PROF_RESET(prof_recent_alloc_mtx); + MUTEX_PROF_RESET(prof_recent_dump_mtx); + MUTEX_PROF_RESET(prof_stats_mtx); } - /* Per arena mutexes. */ unsigned n = narenas_total_get(); @@ -2984,18 +3651,18 @@ continue; } MUTEX_PROF_RESET(arena->large_mtx); - MUTEX_PROF_RESET(arena->extent_avail_mtx); - MUTEX_PROF_RESET(arena->extents_dirty.mtx); - MUTEX_PROF_RESET(arena->extents_muzzy.mtx); - MUTEX_PROF_RESET(arena->extents_retained.mtx); - MUTEX_PROF_RESET(arena->decay_dirty.mtx); - MUTEX_PROF_RESET(arena->decay_muzzy.mtx); + MUTEX_PROF_RESET(arena->pa_shard.edata_cache.mtx); + MUTEX_PROF_RESET(arena->pa_shard.pac.ecache_dirty.mtx); + MUTEX_PROF_RESET(arena->pa_shard.pac.ecache_muzzy.mtx); + MUTEX_PROF_RESET(arena->pa_shard.pac.ecache_retained.mtx); + MUTEX_PROF_RESET(arena->pa_shard.pac.decay_dirty.mtx); + MUTEX_PROF_RESET(arena->pa_shard.pac.decay_muzzy.mtx); MUTEX_PROF_RESET(arena->tcache_ql_mtx); MUTEX_PROF_RESET(arena->base->mtx); - for (szind_t i = 0; i < SC_NBINS; i++) { - for (unsigned j = 0; j < bin_infos[i].n_shards; j++) { - bin_t *bin = &arena->bins[i].bin_shards[j]; + for (szind_t j = 0; j < SC_NBINS; j++) { + for (unsigned k = 0; k < bin_infos[j].n_shards; k++) { + bin_t *bin = arena_get_bin(arena, j, k); MUTEX_PROF_RESET(bin->lock); } } @@ -3005,25 +3672,25 @@ } CTL_RO_CGEN(config_stats, stats_arenas_i_bins_j_nmalloc, - arenas_i(mib[2])->astats->bstats[mib[4]].nmalloc, uint64_t) + arenas_i(mib[2])->astats->bstats[mib[4]].stats_data.nmalloc, uint64_t) CTL_RO_CGEN(config_stats, stats_arenas_i_bins_j_ndalloc, - arenas_i(mib[2])->astats->bstats[mib[4]].ndalloc, uint64_t) + arenas_i(mib[2])->astats->bstats[mib[4]].stats_data.ndalloc, uint64_t) CTL_RO_CGEN(config_stats, stats_arenas_i_bins_j_nrequests, - arenas_i(mib[2])->astats->bstats[mib[4]].nrequests, uint64_t) + arenas_i(mib[2])->astats->bstats[mib[4]].stats_data.nrequests, uint64_t) CTL_RO_CGEN(config_stats, stats_arenas_i_bins_j_curregs, - arenas_i(mib[2])->astats->bstats[mib[4]].curregs, size_t) + arenas_i(mib[2])->astats->bstats[mib[4]].stats_data.curregs, size_t) CTL_RO_CGEN(config_stats, stats_arenas_i_bins_j_nfills, - arenas_i(mib[2])->astats->bstats[mib[4]].nfills, uint64_t) + arenas_i(mib[2])->astats->bstats[mib[4]].stats_data.nfills, uint64_t) CTL_RO_CGEN(config_stats, stats_arenas_i_bins_j_nflushes, - arenas_i(mib[2])->astats->bstats[mib[4]].nflushes, uint64_t) + arenas_i(mib[2])->astats->bstats[mib[4]].stats_data.nflushes, uint64_t) CTL_RO_CGEN(config_stats, stats_arenas_i_bins_j_nslabs, - arenas_i(mib[2])->astats->bstats[mib[4]].nslabs, uint64_t) + arenas_i(mib[2])->astats->bstats[mib[4]].stats_data.nslabs, uint64_t) CTL_RO_CGEN(config_stats, stats_arenas_i_bins_j_nreslabs, - arenas_i(mib[2])->astats->bstats[mib[4]].reslabs, uint64_t) + arenas_i(mib[2])->astats->bstats[mib[4]].stats_data.reslabs, uint64_t) CTL_RO_CGEN(config_stats, stats_arenas_i_bins_j_curslabs, - arenas_i(mib[2])->astats->bstats[mib[4]].curslabs, size_t) + arenas_i(mib[2])->astats->bstats[mib[4]].stats_data.curslabs, size_t) CTL_RO_CGEN(config_stats, stats_arenas_i_bins_j_nonfull_slabs, - arenas_i(mib[2])->astats->bstats[mib[4]].nonfull_slabs, size_t) + arenas_i(mib[2])->astats->bstats[mib[4]].stats_data.nonfull_slabs, size_t) static const ctl_named_node_t * stats_arenas_i_bins_j_index(tsdn_t *tsdn, const size_t *mib, @@ -3035,13 +3702,13 @@ } CTL_RO_CGEN(config_stats, stats_arenas_i_lextents_j_nmalloc, - ctl_arena_stats_read_u64( + locked_read_u64_unsynchronized( &arenas_i(mib[2])->astats->lstats[mib[4]].nmalloc), uint64_t) CTL_RO_CGEN(config_stats, stats_arenas_i_lextents_j_ndalloc, - ctl_arena_stats_read_u64( + locked_read_u64_unsynchronized( &arenas_i(mib[2])->astats->lstats[mib[4]].ndalloc), uint64_t) CTL_RO_CGEN(config_stats, stats_arenas_i_lextents_j_nrequests, - ctl_arena_stats_read_u64( + locked_read_u64_unsynchronized( &arenas_i(mib[2])->astats->lstats[mib[4]].nrequests), uint64_t) CTL_RO_CGEN(config_stats, stats_arenas_i_lextents_j_curlextents, arenas_i(mib[2])->astats->lstats[mib[4]].curlextents, size_t) @@ -3056,29 +3723,17 @@ } CTL_RO_CGEN(config_stats, stats_arenas_i_extents_j_ndirty, - atomic_load_zu( - &arenas_i(mib[2])->astats->estats[mib[4]].ndirty, - ATOMIC_RELAXED), size_t); + arenas_i(mib[2])->astats->estats[mib[4]].ndirty, size_t); CTL_RO_CGEN(config_stats, stats_arenas_i_extents_j_nmuzzy, - atomic_load_zu( - &arenas_i(mib[2])->astats->estats[mib[4]].nmuzzy, - ATOMIC_RELAXED), size_t); + arenas_i(mib[2])->astats->estats[mib[4]].nmuzzy, size_t); CTL_RO_CGEN(config_stats, stats_arenas_i_extents_j_nretained, - atomic_load_zu( - &arenas_i(mib[2])->astats->estats[mib[4]].nretained, - ATOMIC_RELAXED), size_t); + arenas_i(mib[2])->astats->estats[mib[4]].nretained, size_t); CTL_RO_CGEN(config_stats, stats_arenas_i_extents_j_dirty_bytes, - atomic_load_zu( - &arenas_i(mib[2])->astats->estats[mib[4]].dirty_bytes, - ATOMIC_RELAXED), size_t); + arenas_i(mib[2])->astats->estats[mib[4]].dirty_bytes, size_t); CTL_RO_CGEN(config_stats, stats_arenas_i_extents_j_muzzy_bytes, - atomic_load_zu( - &arenas_i(mib[2])->astats->estats[mib[4]].muzzy_bytes, - ATOMIC_RELAXED), size_t); + arenas_i(mib[2])->astats->estats[mib[4]].muzzy_bytes, size_t); CTL_RO_CGEN(config_stats, stats_arenas_i_extents_j_retained_bytes, - atomic_load_zu( - &arenas_i(mib[2])->astats->estats[mib[4]].retained_bytes, - ATOMIC_RELAXED), size_t); + arenas_i(mib[2])->astats->estats[mib[4]].retained_bytes, size_t); static const ctl_named_node_t * stats_arenas_i_extents_j_index(tsdn_t *tsdn, const size_t *mib, @@ -3089,6 +3744,82 @@ return super_stats_arenas_i_extents_j_node; } +CTL_RO_CGEN(config_stats, stats_arenas_i_hpa_shard_npurge_passes, + arenas_i(mib[2])->astats->hpastats.nonderived_stats.npurge_passes, uint64_t); +CTL_RO_CGEN(config_stats, stats_arenas_i_hpa_shard_npurges, + arenas_i(mib[2])->astats->hpastats.nonderived_stats.npurges, uint64_t); +CTL_RO_CGEN(config_stats, stats_arenas_i_hpa_shard_nhugifies, + arenas_i(mib[2])->astats->hpastats.nonderived_stats.nhugifies, uint64_t); +CTL_RO_CGEN(config_stats, stats_arenas_i_hpa_shard_ndehugifies, + arenas_i(mib[2])->astats->hpastats.nonderived_stats.ndehugifies, uint64_t); + +/* Full, nonhuge */ +CTL_RO_CGEN(config_stats, stats_arenas_i_hpa_shard_full_slabs_npageslabs_nonhuge, + arenas_i(mib[2])->astats->hpastats.psset_stats.full_slabs[0].npageslabs, + size_t); +CTL_RO_CGEN(config_stats, stats_arenas_i_hpa_shard_full_slabs_nactive_nonhuge, + arenas_i(mib[2])->astats->hpastats.psset_stats.full_slabs[0].nactive, size_t); +CTL_RO_CGEN(config_stats, stats_arenas_i_hpa_shard_full_slabs_ndirty_nonhuge, + arenas_i(mib[2])->astats->hpastats.psset_stats.full_slabs[0].ndirty, size_t); + +/* Full, huge */ +CTL_RO_CGEN(config_stats, stats_arenas_i_hpa_shard_full_slabs_npageslabs_huge, + arenas_i(mib[2])->astats->hpastats.psset_stats.full_slabs[1].npageslabs, + size_t); +CTL_RO_CGEN(config_stats, stats_arenas_i_hpa_shard_full_slabs_nactive_huge, + arenas_i(mib[2])->astats->hpastats.psset_stats.full_slabs[1].nactive, size_t); +CTL_RO_CGEN(config_stats, stats_arenas_i_hpa_shard_full_slabs_ndirty_huge, + arenas_i(mib[2])->astats->hpastats.psset_stats.full_slabs[1].ndirty, size_t); + +/* Empty, nonhuge */ +CTL_RO_CGEN(config_stats, stats_arenas_i_hpa_shard_empty_slabs_npageslabs_nonhuge, + arenas_i(mib[2])->astats->hpastats.psset_stats.empty_slabs[0].npageslabs, + size_t); +CTL_RO_CGEN(config_stats, stats_arenas_i_hpa_shard_empty_slabs_nactive_nonhuge, + arenas_i(mib[2])->astats->hpastats.psset_stats.empty_slabs[0].nactive, size_t); +CTL_RO_CGEN(config_stats, stats_arenas_i_hpa_shard_empty_slabs_ndirty_nonhuge, + arenas_i(mib[2])->astats->hpastats.psset_stats.empty_slabs[0].ndirty, size_t); + +/* Empty, huge */ +CTL_RO_CGEN(config_stats, stats_arenas_i_hpa_shard_empty_slabs_npageslabs_huge, + arenas_i(mib[2])->astats->hpastats.psset_stats.empty_slabs[1].npageslabs, + size_t); +CTL_RO_CGEN(config_stats, stats_arenas_i_hpa_shard_empty_slabs_nactive_huge, + arenas_i(mib[2])->astats->hpastats.psset_stats.empty_slabs[1].nactive, size_t); +CTL_RO_CGEN(config_stats, stats_arenas_i_hpa_shard_empty_slabs_ndirty_huge, + arenas_i(mib[2])->astats->hpastats.psset_stats.empty_slabs[1].ndirty, size_t); + +/* Nonfull, nonhuge */ +CTL_RO_CGEN(config_stats, stats_arenas_i_hpa_shard_nonfull_slabs_j_npageslabs_nonhuge, + arenas_i(mib[2])->astats->hpastats.psset_stats.nonfull_slabs[mib[5]][0].npageslabs, + size_t); +CTL_RO_CGEN(config_stats, stats_arenas_i_hpa_shard_nonfull_slabs_j_nactive_nonhuge, + arenas_i(mib[2])->astats->hpastats.psset_stats.nonfull_slabs[mib[5]][0].nactive, + size_t); +CTL_RO_CGEN(config_stats, stats_arenas_i_hpa_shard_nonfull_slabs_j_ndirty_nonhuge, + arenas_i(mib[2])->astats->hpastats.psset_stats.nonfull_slabs[mib[5]][0].ndirty, + size_t); + +/* Nonfull, huge */ +CTL_RO_CGEN(config_stats, stats_arenas_i_hpa_shard_nonfull_slabs_j_npageslabs_huge, + arenas_i(mib[2])->astats->hpastats.psset_stats.nonfull_slabs[mib[5]][1].npageslabs, + size_t); +CTL_RO_CGEN(config_stats, stats_arenas_i_hpa_shard_nonfull_slabs_j_nactive_huge, + arenas_i(mib[2])->astats->hpastats.psset_stats.nonfull_slabs[mib[5]][1].nactive, + size_t); +CTL_RO_CGEN(config_stats, stats_arenas_i_hpa_shard_nonfull_slabs_j_ndirty_huge, + arenas_i(mib[2])->astats->hpastats.psset_stats.nonfull_slabs[mib[5]][1].ndirty, + size_t); + +static const ctl_named_node_t * +stats_arenas_i_hpa_shard_nonfull_slabs_j_index(tsdn_t *tsdn, const size_t *mib, + size_t miblen, size_t j) { + if (j >= PSSET_NPSIZES) { + return NULL; + } + return super_stats_arenas_i_hpa_shard_nonfull_slabs_j_node; +} + static bool ctl_arenas_i_verify(size_t i) { size_t a = arenas_i2a_impl(i, true, true); @@ -3161,6 +3892,32 @@ return ret; } +static int +experimental_thread_activity_callback_ctl(tsd_t *tsd, const size_t *mib, + size_t miblen, void *oldp, size_t *oldlenp, void *newp, size_t newlen) { + int ret; + + if (!config_stats) { + return ENOENT; + } + + activity_callback_thunk_t t_old = tsd_activity_callback_thunk_get(tsd); + READ(t_old, activity_callback_thunk_t); + + if (newp != NULL) { + /* + * This initialization is unnecessary. If it's omitted, though, + * clang gets confused and warns on the subsequent use of t_new. + */ + activity_callback_thunk_t t_new = {NULL, NULL}; + WRITE(t_new, activity_callback_thunk_t); + tsd_activity_callback_thunk_set(tsd, t_new); + } + ret = 0; +label_return: + return ret; +} + /* * Output six memory utilization entries for an input pointer, the first one of * type (void *) and the remaining five of type size_t, describing the following @@ -3178,7 +3935,8 @@ * otherwise their values are undefined. * * This API is mainly intended for small class allocations, where extents are - * used as slab. + * used as slab. Note that if the bin the extent belongs to is completely + * full, "(a)" will be NULL. * * In case of large class allocations, "(a)" will be NULL, and "(e)" and "(f)" * will be zero (if stats are enabled; otherwise undefined). The other three @@ -3232,11 +3990,11 @@ size_t miblen, void *oldp, size_t *oldlenp, void *newp, size_t newlen) { int ret; - assert(sizeof(extent_util_stats_verbose_t) + assert(sizeof(inspect_extent_util_stats_verbose_t) == sizeof(void *) + sizeof(size_t) * 5); if (oldp == NULL || oldlenp == NULL - || *oldlenp != sizeof(extent_util_stats_verbose_t) + || *oldlenp != sizeof(inspect_extent_util_stats_verbose_t) || newp == NULL) { ret = EINVAL; goto label_return; @@ -3244,9 +4002,9 @@ void *ptr = NULL; WRITE(ptr, void *); - extent_util_stats_verbose_t *util_stats - = (extent_util_stats_verbose_t *)oldp; - extent_util_stats_verbose_get(tsd_tsdn(tsd), ptr, + inspect_extent_util_stats_verbose_t *util_stats + = (inspect_extent_util_stats_verbose_t *)oldp; + inspect_extent_util_stats_verbose_get(tsd_tsdn(tsd), ptr, &util_stats->nfree, &util_stats->nregs, &util_stats->size, &util_stats->bin_nfree, &util_stats->bin_nregs, &util_stats->slabcur_addr); @@ -3357,21 +4115,22 @@ size_t miblen, void *oldp, size_t *oldlenp, void *newp, size_t newlen) { int ret; - assert(sizeof(extent_util_stats_t) == sizeof(size_t) * 3); + assert(sizeof(inspect_extent_util_stats_t) == sizeof(size_t) * 3); const size_t len = newlen / sizeof(const void *); if (oldp == NULL || oldlenp == NULL || newp == NULL || newlen == 0 || newlen != len * sizeof(const void *) - || *oldlenp != len * sizeof(extent_util_stats_t)) { + || *oldlenp != len * sizeof(inspect_extent_util_stats_t)) { ret = EINVAL; goto label_return; } void **ptrs = (void **)newp; - extent_util_stats_t *util_stats = (extent_util_stats_t *)oldp; + inspect_extent_util_stats_t *util_stats = + (inspect_extent_util_stats_t *)oldp; size_t i; for (i = 0; i < len; ++i) { - extent_util_stats_get(tsd_tsdn(tsd), ptrs[i], + inspect_extent_util_stats_get(tsd_tsdn(tsd), ptrs[i], &util_stats[i].nfree, &util_stats[i].nregs, &util_stats[i].size); } @@ -3420,7 +4179,7 @@ #if defined(JEMALLOC_GCC_ATOMIC_ATOMICS) || \ defined(JEMALLOC_GCC_SYNC_ATOMICS) || defined(_MSC_VER) /* Expose the underlying counter for fast read. */ - pactivep = (size_t *)&(arena->nactive.repr); + pactivep = (size_t *)&(arena->pa_shard.nactive.repr); READ(pactivep, size_t *); ret = 0; #else @@ -3433,3 +4192,223 @@ malloc_mutex_unlock(tsd_tsdn(tsd), &ctl_mtx); return ret; } + +static int +experimental_prof_recent_alloc_max_ctl(tsd_t *tsd, const size_t *mib, + size_t miblen, void *oldp, size_t *oldlenp, void *newp, size_t newlen) { + int ret; + + if (!(config_prof && opt_prof)) { + ret = ENOENT; + goto label_return; + } + + ssize_t old_max; + if (newp != NULL) { + if (newlen != sizeof(ssize_t)) { + ret = EINVAL; + goto label_return; + } + ssize_t max = *(ssize_t *)newp; + if (max < -1) { + ret = EINVAL; + goto label_return; + } + old_max = prof_recent_alloc_max_ctl_write(tsd, max); + } else { + old_max = prof_recent_alloc_max_ctl_read(); + } + READ(old_max, ssize_t); + + ret = 0; + +label_return: + return ret; +} + +typedef struct write_cb_packet_s write_cb_packet_t; +struct write_cb_packet_s { + write_cb_t *write_cb; + void *cbopaque; +}; + +static int +experimental_prof_recent_alloc_dump_ctl(tsd_t *tsd, const size_t *mib, + size_t miblen, void *oldp, size_t *oldlenp, void *newp, size_t newlen) { + int ret; + + if (!(config_prof && opt_prof)) { + ret = ENOENT; + goto label_return; + } + + assert(sizeof(write_cb_packet_t) == sizeof(void *) * 2); + + WRITEONLY(); + write_cb_packet_t write_cb_packet; + ASSURED_WRITE(write_cb_packet, write_cb_packet_t); + + prof_recent_alloc_dump(tsd, write_cb_packet.write_cb, + write_cb_packet.cbopaque); + + ret = 0; + +label_return: + return ret; +} + +typedef struct batch_alloc_packet_s batch_alloc_packet_t; +struct batch_alloc_packet_s { + void **ptrs; + size_t num; + size_t size; + int flags; +}; + +static int +experimental_batch_alloc_ctl(tsd_t *tsd, const size_t *mib, + size_t miblen, void *oldp, size_t *oldlenp, void *newp, size_t newlen) { + int ret; + + VERIFY_READ(size_t); + + batch_alloc_packet_t batch_alloc_packet; + ASSURED_WRITE(batch_alloc_packet, batch_alloc_packet_t); + size_t filled = batch_alloc(batch_alloc_packet.ptrs, + batch_alloc_packet.num, batch_alloc_packet.size, + batch_alloc_packet.flags); + READ(filled, size_t); + + ret = 0; + +label_return: + return ret; +} + +static int +prof_stats_bins_i_live_ctl(tsd_t *tsd, const size_t *mib, size_t miblen, + void *oldp, size_t *oldlenp, void *newp, size_t newlen) { + int ret; + unsigned binind; + prof_stats_t stats; + + if (!(config_prof && opt_prof && opt_prof_stats)) { + ret = ENOENT; + goto label_return; + } + + READONLY(); + MIB_UNSIGNED(binind, 3); + if (binind >= SC_NBINS) { + ret = EINVAL; + goto label_return; + } + prof_stats_get_live(tsd, (szind_t)binind, &stats); + READ(stats, prof_stats_t); + + ret = 0; +label_return: + return ret; +} + +static int +prof_stats_bins_i_accum_ctl(tsd_t *tsd, const size_t *mib, size_t miblen, + void *oldp, size_t *oldlenp, void *newp, size_t newlen) { + int ret; + unsigned binind; + prof_stats_t stats; + + if (!(config_prof && opt_prof && opt_prof_stats)) { + ret = ENOENT; + goto label_return; + } + + READONLY(); + MIB_UNSIGNED(binind, 3); + if (binind >= SC_NBINS) { + ret = EINVAL; + goto label_return; + } + prof_stats_get_accum(tsd, (szind_t)binind, &stats); + READ(stats, prof_stats_t); + + ret = 0; +label_return: + return ret; +} + +static const ctl_named_node_t * +prof_stats_bins_i_index(tsdn_t *tsdn, const size_t *mib, size_t miblen, + size_t i) { + if (!(config_prof && opt_prof && opt_prof_stats)) { + return NULL; + } + if (i >= SC_NBINS) { + return NULL; + } + return super_prof_stats_bins_i_node; +} + +static int +prof_stats_lextents_i_live_ctl(tsd_t *tsd, const size_t *mib, size_t miblen, + void *oldp, size_t *oldlenp, void *newp, size_t newlen) { + int ret; + unsigned lextent_ind; + prof_stats_t stats; + + if (!(config_prof && opt_prof && opt_prof_stats)) { + ret = ENOENT; + goto label_return; + } + + READONLY(); + MIB_UNSIGNED(lextent_ind, 3); + if (lextent_ind >= SC_NSIZES - SC_NBINS) { + ret = EINVAL; + goto label_return; + } + prof_stats_get_live(tsd, (szind_t)(lextent_ind + SC_NBINS), &stats); + READ(stats, prof_stats_t); + + ret = 0; +label_return: + return ret; +} + +static int +prof_stats_lextents_i_accum_ctl(tsd_t *tsd, const size_t *mib, size_t miblen, + void *oldp, size_t *oldlenp, void *newp, size_t newlen) { + int ret; + unsigned lextent_ind; + prof_stats_t stats; + + if (!(config_prof && opt_prof && opt_prof_stats)) { + ret = ENOENT; + goto label_return; + } + + READONLY(); + MIB_UNSIGNED(lextent_ind, 3); + if (lextent_ind >= SC_NSIZES - SC_NBINS) { + ret = EINVAL; + goto label_return; + } + prof_stats_get_accum(tsd, (szind_t)(lextent_ind + SC_NBINS), &stats); + READ(stats, prof_stats_t); + + ret = 0; +label_return: + return ret; +} + +static const ctl_named_node_t * +prof_stats_lextents_i_index(tsdn_t *tsdn, const size_t *mib, size_t miblen, + size_t i) { + if (!(config_prof && opt_prof && opt_prof_stats)) { + return NULL; + } + if (i >= SC_NSIZES - SC_NBINS) { + return NULL; + } + return super_prof_stats_lextents_i_node; +} diff --git a/contrib/jemalloc/src/decay.c b/contrib/jemalloc/src/decay.c new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/src/decay.c @@ -0,0 +1,295 @@ +#include "jemalloc/internal/jemalloc_preamble.h" +#include "jemalloc/internal/jemalloc_internal_includes.h" + +#include "jemalloc/internal/decay.h" + +static const uint64_t h_steps[SMOOTHSTEP_NSTEPS] = { +#define STEP(step, h, x, y) \ + h, + SMOOTHSTEP +#undef STEP +}; + +/* + * Generate a new deadline that is uniformly random within the next epoch after + * the current one. + */ +void +decay_deadline_init(decay_t *decay) { + nstime_copy(&decay->deadline, &decay->epoch); + nstime_add(&decay->deadline, &decay->interval); + if (decay_ms_read(decay) > 0) { + nstime_t jitter; + + nstime_init(&jitter, prng_range_u64(&decay->jitter_state, + nstime_ns(&decay->interval))); + nstime_add(&decay->deadline, &jitter); + } +} + +void +decay_reinit(decay_t *decay, nstime_t *cur_time, ssize_t decay_ms) { + atomic_store_zd(&decay->time_ms, decay_ms, ATOMIC_RELAXED); + if (decay_ms > 0) { + nstime_init(&decay->interval, (uint64_t)decay_ms * + KQU(1000000)); + nstime_idivide(&decay->interval, SMOOTHSTEP_NSTEPS); + } + + nstime_copy(&decay->epoch, cur_time); + decay->jitter_state = (uint64_t)(uintptr_t)decay; + decay_deadline_init(decay); + decay->nunpurged = 0; + memset(decay->backlog, 0, SMOOTHSTEP_NSTEPS * sizeof(size_t)); +} + +bool +decay_init(decay_t *decay, nstime_t *cur_time, ssize_t decay_ms) { + if (config_debug) { + for (size_t i = 0; i < sizeof(decay_t); i++) { + assert(((char *)decay)[i] == 0); + } + decay->ceil_npages = 0; + } + if (malloc_mutex_init(&decay->mtx, "decay", WITNESS_RANK_DECAY, + malloc_mutex_rank_exclusive)) { + return true; + } + decay->purging = false; + decay_reinit(decay, cur_time, decay_ms); + return false; +} + +bool +decay_ms_valid(ssize_t decay_ms) { + if (decay_ms < -1) { + return false; + } + if (decay_ms == -1 || (uint64_t)decay_ms <= NSTIME_SEC_MAX * + KQU(1000)) { + return true; + } + return false; +} + +static void +decay_maybe_update_time(decay_t *decay, nstime_t *new_time) { + if (unlikely(!nstime_monotonic() && nstime_compare(&decay->epoch, + new_time) > 0)) { + /* + * Time went backwards. Move the epoch back in time and + * generate a new deadline, with the expectation that time + * typically flows forward for long enough periods of time that + * epochs complete. Unfortunately, this strategy is susceptible + * to clock jitter triggering premature epoch advances, but + * clock jitter estimation and compensation isn't feasible here + * because calls into this code are event-driven. + */ + nstime_copy(&decay->epoch, new_time); + decay_deadline_init(decay); + } else { + /* Verify that time does not go backwards. */ + assert(nstime_compare(&decay->epoch, new_time) <= 0); + } +} + +static size_t +decay_backlog_npages_limit(const decay_t *decay) { + /* + * For each element of decay_backlog, multiply by the corresponding + * fixed-point smoothstep decay factor. Sum the products, then divide + * to round down to the nearest whole number of pages. + */ + uint64_t sum = 0; + for (unsigned i = 0; i < SMOOTHSTEP_NSTEPS; i++) { + sum += decay->backlog[i] * h_steps[i]; + } + size_t npages_limit_backlog = (size_t)(sum >> SMOOTHSTEP_BFP); + + return npages_limit_backlog; +} + +/* + * Update backlog, assuming that 'nadvance_u64' time intervals have passed. + * Trailing 'nadvance_u64' records should be erased and 'current_npages' is + * placed as the newest record. + */ +static void +decay_backlog_update(decay_t *decay, uint64_t nadvance_u64, + size_t current_npages) { + if (nadvance_u64 >= SMOOTHSTEP_NSTEPS) { + memset(decay->backlog, 0, (SMOOTHSTEP_NSTEPS-1) * + sizeof(size_t)); + } else { + size_t nadvance_z = (size_t)nadvance_u64; + + assert((uint64_t)nadvance_z == nadvance_u64); + + memmove(decay->backlog, &decay->backlog[nadvance_z], + (SMOOTHSTEP_NSTEPS - nadvance_z) * sizeof(size_t)); + if (nadvance_z > 1) { + memset(&decay->backlog[SMOOTHSTEP_NSTEPS - + nadvance_z], 0, (nadvance_z-1) * sizeof(size_t)); + } + } + + size_t npages_delta = (current_npages > decay->nunpurged) ? + current_npages - decay->nunpurged : 0; + decay->backlog[SMOOTHSTEP_NSTEPS-1] = npages_delta; + + if (config_debug) { + if (current_npages > decay->ceil_npages) { + decay->ceil_npages = current_npages; + } + size_t npages_limit = decay_backlog_npages_limit(decay); + assert(decay->ceil_npages >= npages_limit); + if (decay->ceil_npages > npages_limit) { + decay->ceil_npages = npages_limit; + } + } +} + +static inline bool +decay_deadline_reached(const decay_t *decay, const nstime_t *time) { + return (nstime_compare(&decay->deadline, time) <= 0); +} + +uint64_t +decay_npages_purge_in(decay_t *decay, nstime_t *time, size_t npages_new) { + uint64_t decay_interval_ns = decay_epoch_duration_ns(decay); + size_t n_epoch = (size_t)(nstime_ns(time) / decay_interval_ns); + + uint64_t npages_purge; + if (n_epoch >= SMOOTHSTEP_NSTEPS) { + npages_purge = npages_new; + } else { + uint64_t h_steps_max = h_steps[SMOOTHSTEP_NSTEPS - 1]; + assert(h_steps_max >= + h_steps[SMOOTHSTEP_NSTEPS - 1 - n_epoch]); + npages_purge = npages_new * (h_steps_max - + h_steps[SMOOTHSTEP_NSTEPS - 1 - n_epoch]); + npages_purge >>= SMOOTHSTEP_BFP; + } + return npages_purge; +} + +bool +decay_maybe_advance_epoch(decay_t *decay, nstime_t *new_time, + size_t npages_current) { + /* Handle possible non-monotonicity of time. */ + decay_maybe_update_time(decay, new_time); + + if (!decay_deadline_reached(decay, new_time)) { + return false; + } + nstime_t delta; + nstime_copy(&delta, new_time); + nstime_subtract(&delta, &decay->epoch); + + uint64_t nadvance_u64 = nstime_divide(&delta, &decay->interval); + assert(nadvance_u64 > 0); + + /* Add nadvance_u64 decay intervals to epoch. */ + nstime_copy(&delta, &decay->interval); + nstime_imultiply(&delta, nadvance_u64); + nstime_add(&decay->epoch, &delta); + + /* Set a new deadline. */ + decay_deadline_init(decay); + + /* Update the backlog. */ + decay_backlog_update(decay, nadvance_u64, npages_current); + + decay->npages_limit = decay_backlog_npages_limit(decay); + decay->nunpurged = (decay->npages_limit > npages_current) ? + decay->npages_limit : npages_current; + + return true; +} + +/* + * Calculate how many pages should be purged after 'interval'. + * + * First, calculate how many pages should remain at the moment, then subtract + * the number of pages that should remain after 'interval'. The difference is + * how many pages should be purged until then. + * + * The number of pages that should remain at a specific moment is calculated + * like this: pages(now) = sum(backlog[i] * h_steps[i]). After 'interval' + * passes, backlog would shift 'interval' positions to the left and sigmoid + * curve would be applied starting with backlog[interval]. + * + * The implementation doesn't directly map to the description, but it's + * essentially the same calculation, optimized to avoid iterating over + * [interval..SMOOTHSTEP_NSTEPS) twice. + */ +static inline size_t +decay_npurge_after_interval(decay_t *decay, size_t interval) { + size_t i; + uint64_t sum = 0; + for (i = 0; i < interval; i++) { + sum += decay->backlog[i] * h_steps[i]; + } + for (; i < SMOOTHSTEP_NSTEPS; i++) { + sum += decay->backlog[i] * + (h_steps[i] - h_steps[i - interval]); + } + + return (size_t)(sum >> SMOOTHSTEP_BFP); +} + +uint64_t decay_ns_until_purge(decay_t *decay, size_t npages_current, + uint64_t npages_threshold) { + if (!decay_gradually(decay)) { + return DECAY_UNBOUNDED_TIME_TO_PURGE; + } + uint64_t decay_interval_ns = decay_epoch_duration_ns(decay); + assert(decay_interval_ns > 0); + if (npages_current == 0) { + unsigned i; + for (i = 0; i < SMOOTHSTEP_NSTEPS; i++) { + if (decay->backlog[i] > 0) { + break; + } + } + if (i == SMOOTHSTEP_NSTEPS) { + /* No dirty pages recorded. Sleep indefinitely. */ + return DECAY_UNBOUNDED_TIME_TO_PURGE; + } + } + if (npages_current <= npages_threshold) { + /* Use max interval. */ + return decay_interval_ns * SMOOTHSTEP_NSTEPS; + } + + /* Minimal 2 intervals to ensure reaching next epoch deadline. */ + size_t lb = 2; + size_t ub = SMOOTHSTEP_NSTEPS; + + size_t npurge_lb, npurge_ub; + npurge_lb = decay_npurge_after_interval(decay, lb); + if (npurge_lb > npages_threshold) { + return decay_interval_ns * lb; + } + npurge_ub = decay_npurge_after_interval(decay, ub); + if (npurge_ub < npages_threshold) { + return decay_interval_ns * ub; + } + + unsigned n_search = 0; + size_t target, npurge; + while ((npurge_lb + npages_threshold < npurge_ub) && (lb + 2 < ub)) { + target = (lb + ub) / 2; + npurge = decay_npurge_after_interval(decay, target); + if (npurge > npages_threshold) { + ub = target; + npurge_ub = npurge; + } else { + lb = target; + npurge_lb = npurge; + } + assert(n_search < lg_floor(SMOOTHSTEP_NSTEPS) + 1); + ++n_search; + } + return decay_interval_ns * (ub + lb) / 2; +} diff --git a/contrib/jemalloc/src/ecache.c b/contrib/jemalloc/src/ecache.c new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/src/ecache.c @@ -0,0 +1,35 @@ +#include "jemalloc/internal/jemalloc_preamble.h" +#include "jemalloc/internal/jemalloc_internal_includes.h" + +#include "jemalloc/internal/san.h" + +bool +ecache_init(tsdn_t *tsdn, ecache_t *ecache, extent_state_t state, unsigned ind, + bool delay_coalesce) { + if (malloc_mutex_init(&ecache->mtx, "extents", WITNESS_RANK_EXTENTS, + malloc_mutex_rank_exclusive)) { + return true; + } + ecache->state = state; + ecache->ind = ind; + ecache->delay_coalesce = delay_coalesce; + eset_init(&ecache->eset, state); + eset_init(&ecache->guarded_eset, state); + + return false; +} + +void +ecache_prefork(tsdn_t *tsdn, ecache_t *ecache) { + malloc_mutex_prefork(tsdn, &ecache->mtx); +} + +void +ecache_postfork_parent(tsdn_t *tsdn, ecache_t *ecache) { + malloc_mutex_postfork_parent(tsdn, &ecache->mtx); +} + +void +ecache_postfork_child(tsdn_t *tsdn, ecache_t *ecache) { + malloc_mutex_postfork_child(tsdn, &ecache->mtx); +} diff --git a/contrib/jemalloc/src/edata.c b/contrib/jemalloc/src/edata.c new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/src/edata.c @@ -0,0 +1,6 @@ +#include "jemalloc/internal/jemalloc_preamble.h" +#include "jemalloc/internal/jemalloc_internal_includes.h" + +ph_gen(, edata_avail, edata_t, avail_link, + edata_esnead_comp) +ph_gen(, edata_heap, edata_t, heap_link, edata_snad_comp) diff --git a/contrib/jemalloc/src/edata_cache.c b/contrib/jemalloc/src/edata_cache.c new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/src/edata_cache.c @@ -0,0 +1,154 @@ +#include "jemalloc/internal/jemalloc_preamble.h" +#include "jemalloc/internal/jemalloc_internal_includes.h" + +bool +edata_cache_init(edata_cache_t *edata_cache, base_t *base) { + edata_avail_new(&edata_cache->avail); + /* + * This is not strictly necessary, since the edata_cache_t is only + * created inside an arena, which is zeroed on creation. But this is + * handy as a safety measure. + */ + atomic_store_zu(&edata_cache->count, 0, ATOMIC_RELAXED); + if (malloc_mutex_init(&edata_cache->mtx, "edata_cache", + WITNESS_RANK_EDATA_CACHE, malloc_mutex_rank_exclusive)) { + return true; + } + edata_cache->base = base; + return false; +} + +edata_t * +edata_cache_get(tsdn_t *tsdn, edata_cache_t *edata_cache) { + malloc_mutex_lock(tsdn, &edata_cache->mtx); + edata_t *edata = edata_avail_first(&edata_cache->avail); + if (edata == NULL) { + malloc_mutex_unlock(tsdn, &edata_cache->mtx); + return base_alloc_edata(tsdn, edata_cache->base); + } + edata_avail_remove(&edata_cache->avail, edata); + atomic_load_sub_store_zu(&edata_cache->count, 1); + malloc_mutex_unlock(tsdn, &edata_cache->mtx); + return edata; +} + +void +edata_cache_put(tsdn_t *tsdn, edata_cache_t *edata_cache, edata_t *edata) { + malloc_mutex_lock(tsdn, &edata_cache->mtx); + edata_avail_insert(&edata_cache->avail, edata); + atomic_load_add_store_zu(&edata_cache->count, 1); + malloc_mutex_unlock(tsdn, &edata_cache->mtx); +} + +void +edata_cache_prefork(tsdn_t *tsdn, edata_cache_t *edata_cache) { + malloc_mutex_prefork(tsdn, &edata_cache->mtx); +} + +void +edata_cache_postfork_parent(tsdn_t *tsdn, edata_cache_t *edata_cache) { + malloc_mutex_postfork_parent(tsdn, &edata_cache->mtx); +} + +void +edata_cache_postfork_child(tsdn_t *tsdn, edata_cache_t *edata_cache) { + malloc_mutex_postfork_child(tsdn, &edata_cache->mtx); +} + +void +edata_cache_fast_init(edata_cache_fast_t *ecs, edata_cache_t *fallback) { + edata_list_inactive_init(&ecs->list); + ecs->fallback = fallback; + ecs->disabled = false; +} + +static void +edata_cache_fast_try_fill_from_fallback(tsdn_t *tsdn, + edata_cache_fast_t *ecs) { + edata_t *edata; + malloc_mutex_lock(tsdn, &ecs->fallback->mtx); + for (int i = 0; i < EDATA_CACHE_FAST_FILL; i++) { + edata = edata_avail_remove_first(&ecs->fallback->avail); + if (edata == NULL) { + break; + } + edata_list_inactive_append(&ecs->list, edata); + atomic_load_sub_store_zu(&ecs->fallback->count, 1); + } + malloc_mutex_unlock(tsdn, &ecs->fallback->mtx); +} + +edata_t * +edata_cache_fast_get(tsdn_t *tsdn, edata_cache_fast_t *ecs) { + witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), + WITNESS_RANK_EDATA_CACHE, 0); + + if (ecs->disabled) { + assert(edata_list_inactive_first(&ecs->list) == NULL); + return edata_cache_get(tsdn, ecs->fallback); + } + + edata_t *edata = edata_list_inactive_first(&ecs->list); + if (edata != NULL) { + edata_list_inactive_remove(&ecs->list, edata); + return edata; + } + /* Slow path; requires synchronization. */ + edata_cache_fast_try_fill_from_fallback(tsdn, ecs); + edata = edata_list_inactive_first(&ecs->list); + if (edata != NULL) { + edata_list_inactive_remove(&ecs->list, edata); + } else { + /* + * Slowest path (fallback was also empty); allocate something + * new. + */ + edata = base_alloc_edata(tsdn, ecs->fallback->base); + } + return edata; +} + +static void +edata_cache_fast_flush_all(tsdn_t *tsdn, edata_cache_fast_t *ecs) { + /* + * You could imagine smarter cache management policies (like + * only flushing down to some threshold in anticipation of + * future get requests). But just flushing everything provides + * a good opportunity to defrag too, and lets us share code between the + * flush and disable pathways. + */ + edata_t *edata; + size_t nflushed = 0; + malloc_mutex_lock(tsdn, &ecs->fallback->mtx); + while ((edata = edata_list_inactive_first(&ecs->list)) != NULL) { + edata_list_inactive_remove(&ecs->list, edata); + edata_avail_insert(&ecs->fallback->avail, edata); + nflushed++; + } + atomic_load_add_store_zu(&ecs->fallback->count, nflushed); + malloc_mutex_unlock(tsdn, &ecs->fallback->mtx); +} + +void +edata_cache_fast_put(tsdn_t *tsdn, edata_cache_fast_t *ecs, edata_t *edata) { + witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), + WITNESS_RANK_EDATA_CACHE, 0); + + if (ecs->disabled) { + assert(edata_list_inactive_first(&ecs->list) == NULL); + edata_cache_put(tsdn, ecs->fallback, edata); + return; + } + + /* + * Prepend rather than append, to do LIFO ordering in the hopes of some + * cache locality. + */ + edata_list_inactive_prepend(&ecs->list, edata); +} + +void +edata_cache_fast_disable(tsdn_t *tsdn, edata_cache_fast_t *ecs) { + edata_cache_fast_flush_all(tsdn, ecs); + ecs->disabled = true; +} diff --git a/contrib/jemalloc/src/ehooks.c b/contrib/jemalloc/src/ehooks.c new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/src/ehooks.c @@ -0,0 +1,275 @@ +#include "jemalloc/internal/jemalloc_preamble.h" +#include "jemalloc/internal/jemalloc_internal_includes.h" + +#include "jemalloc/internal/ehooks.h" +#include "jemalloc/internal/extent_mmap.h" + +void +ehooks_init(ehooks_t *ehooks, extent_hooks_t *extent_hooks, unsigned ind) { + /* All other hooks are optional; this one is not. */ + assert(extent_hooks->alloc != NULL); + ehooks->ind = ind; + ehooks_set_extent_hooks_ptr(ehooks, extent_hooks); +} + +/* + * If the caller specifies (!*zero), it is still possible to receive zeroed + * memory, in which case *zero is toggled to true. arena_extent_alloc() takes + * advantage of this to avoid demanding zeroed extents, but taking advantage of + * them if they are returned. + */ +static void * +extent_alloc_core(tsdn_t *tsdn, arena_t *arena, void *new_addr, size_t size, + size_t alignment, bool *zero, bool *commit, dss_prec_t dss_prec) { + void *ret; + + assert(size != 0); + assert(alignment != 0); + + /* "primary" dss. */ + if (have_dss && dss_prec == dss_prec_primary && (ret = + extent_alloc_dss(tsdn, arena, new_addr, size, alignment, zero, + commit)) != NULL) { + return ret; + } + /* mmap. */ + if ((ret = extent_alloc_mmap(new_addr, size, alignment, zero, commit)) + != NULL) { + return ret; + } + /* "secondary" dss. */ + if (have_dss && dss_prec == dss_prec_secondary && (ret = + extent_alloc_dss(tsdn, arena, new_addr, size, alignment, zero, + commit)) != NULL) { + return ret; + } + + /* All strategies for allocation failed. */ + return NULL; +} + +void * +ehooks_default_alloc_impl(tsdn_t *tsdn, void *new_addr, size_t size, + size_t alignment, bool *zero, bool *commit, unsigned arena_ind) { + arena_t *arena = arena_get(tsdn, arena_ind, false); + /* NULL arena indicates arena_create. */ + assert(arena != NULL || alignment == HUGEPAGE); + dss_prec_t dss = (arena == NULL) ? dss_prec_disabled : + (dss_prec_t)atomic_load_u(&arena->dss_prec, ATOMIC_RELAXED); + void *ret = extent_alloc_core(tsdn, arena, new_addr, size, alignment, + zero, commit, dss); + if (have_madvise_huge && ret) { + pages_set_thp_state(ret, size); + } + return ret; +} + +static void * +ehooks_default_alloc(extent_hooks_t *extent_hooks, void *new_addr, size_t size, + size_t alignment, bool *zero, bool *commit, unsigned arena_ind) { + return ehooks_default_alloc_impl(tsdn_fetch(), new_addr, size, + ALIGNMENT_CEILING(alignment, PAGE), zero, commit, arena_ind); +} + +bool +ehooks_default_dalloc_impl(void *addr, size_t size) { + if (!have_dss || !extent_in_dss(addr)) { + return extent_dalloc_mmap(addr, size); + } + return true; +} + +static bool +ehooks_default_dalloc(extent_hooks_t *extent_hooks, void *addr, size_t size, + bool committed, unsigned arena_ind) { + return ehooks_default_dalloc_impl(addr, size); +} + +void +ehooks_default_destroy_impl(void *addr, size_t size) { + if (!have_dss || !extent_in_dss(addr)) { + pages_unmap(addr, size); + } +} + +static void +ehooks_default_destroy(extent_hooks_t *extent_hooks, void *addr, size_t size, + bool committed, unsigned arena_ind) { + ehooks_default_destroy_impl(addr, size); +} + +bool +ehooks_default_commit_impl(void *addr, size_t offset, size_t length) { + return pages_commit((void *)((uintptr_t)addr + (uintptr_t)offset), + length); +} + +static bool +ehooks_default_commit(extent_hooks_t *extent_hooks, void *addr, size_t size, + size_t offset, size_t length, unsigned arena_ind) { + return ehooks_default_commit_impl(addr, offset, length); +} + +bool +ehooks_default_decommit_impl(void *addr, size_t offset, size_t length) { + return pages_decommit((void *)((uintptr_t)addr + (uintptr_t)offset), + length); +} + +static bool +ehooks_default_decommit(extent_hooks_t *extent_hooks, void *addr, size_t size, + size_t offset, size_t length, unsigned arena_ind) { + return ehooks_default_decommit_impl(addr, offset, length); +} + +#ifdef PAGES_CAN_PURGE_LAZY +bool +ehooks_default_purge_lazy_impl(void *addr, size_t offset, size_t length) { + return pages_purge_lazy((void *)((uintptr_t)addr + (uintptr_t)offset), + length); +} + +static bool +ehooks_default_purge_lazy(extent_hooks_t *extent_hooks, void *addr, size_t size, + size_t offset, size_t length, unsigned arena_ind) { + assert(addr != NULL); + assert((offset & PAGE_MASK) == 0); + assert(length != 0); + assert((length & PAGE_MASK) == 0); + return ehooks_default_purge_lazy_impl(addr, offset, length); +} +#endif + +#ifdef PAGES_CAN_PURGE_FORCED +bool +ehooks_default_purge_forced_impl(void *addr, size_t offset, size_t length) { + return pages_purge_forced((void *)((uintptr_t)addr + + (uintptr_t)offset), length); +} + +static bool +ehooks_default_purge_forced(extent_hooks_t *extent_hooks, void *addr, + size_t size, size_t offset, size_t length, unsigned arena_ind) { + assert(addr != NULL); + assert((offset & PAGE_MASK) == 0); + assert(length != 0); + assert((length & PAGE_MASK) == 0); + return ehooks_default_purge_forced_impl(addr, offset, length); +} +#endif + +bool +ehooks_default_split_impl() { + if (!maps_coalesce) { + /* + * Without retain, only whole regions can be purged (required by + * MEM_RELEASE on Windows) -- therefore disallow splitting. See + * comments in extent_head_no_merge(). + */ + return !opt_retain; + } + + return false; +} + +static bool +ehooks_default_split(extent_hooks_t *extent_hooks, void *addr, size_t size, + size_t size_a, size_t size_b, bool committed, unsigned arena_ind) { + return ehooks_default_split_impl(); +} + +bool +ehooks_default_merge_impl(tsdn_t *tsdn, void *addr_a, void *addr_b) { + assert(addr_a < addr_b); + /* + * For non-DSS cases -- + * a) W/o maps_coalesce, merge is not always allowed (Windows): + * 1) w/o retain, never merge (first branch below). + * 2) with retain, only merge extents from the same VirtualAlloc + * region (in which case MEM_DECOMMIT is utilized for purging). + * + * b) With maps_coalesce, it's always possible to merge. + * 1) w/o retain, always allow merge (only about dirty / muzzy). + * 2) with retain, to preserve the SN / first-fit, merge is still + * disallowed if b is a head extent, i.e. no merging across + * different mmap regions. + * + * a2) and b2) are implemented in emap_try_acquire_edata_neighbor, and + * sanity checked in the second branch below. + */ + if (!maps_coalesce && !opt_retain) { + return true; + } + if (config_debug) { + edata_t *a = emap_edata_lookup(tsdn, &arena_emap_global, + addr_a); + bool head_a = edata_is_head_get(a); + edata_t *b = emap_edata_lookup(tsdn, &arena_emap_global, + addr_b); + bool head_b = edata_is_head_get(b); + emap_assert_mapped(tsdn, &arena_emap_global, a); + emap_assert_mapped(tsdn, &arena_emap_global, b); + assert(extent_neighbor_head_state_mergeable(head_a, head_b, + /* forward */ true)); + } + if (have_dss && !extent_dss_mergeable(addr_a, addr_b)) { + return true; + } + + return false; +} + +bool +ehooks_default_merge(extent_hooks_t *extent_hooks, void *addr_a, size_t size_a, + void *addr_b, size_t size_b, bool committed, unsigned arena_ind) { + tsdn_t *tsdn = tsdn_fetch(); + + return ehooks_default_merge_impl(tsdn, addr_a, addr_b); +} + +void +ehooks_default_zero_impl(void *addr, size_t size) { + /* + * By default, we try to zero out memory using OS-provided demand-zeroed + * pages. If the user has specifically requested hugepages, though, we + * don't want to purge in the middle of a hugepage (which would break it + * up), so we act conservatively and use memset. + */ + bool needs_memset = true; + if (opt_thp != thp_mode_always) { + needs_memset = pages_purge_forced(addr, size); + } + if (needs_memset) { + memset(addr, 0, size); + } +} + +void +ehooks_default_guard_impl(void *guard1, void *guard2) { + pages_mark_guards(guard1, guard2); +} + +void +ehooks_default_unguard_impl(void *guard1, void *guard2) { + pages_unmark_guards(guard1, guard2); +} + +const extent_hooks_t ehooks_default_extent_hooks = { + ehooks_default_alloc, + ehooks_default_dalloc, + ehooks_default_destroy, + ehooks_default_commit, + ehooks_default_decommit, +#ifdef PAGES_CAN_PURGE_LAZY + ehooks_default_purge_lazy, +#else + NULL, +#endif +#ifdef PAGES_CAN_PURGE_FORCED + ehooks_default_purge_forced, +#else + NULL, +#endif + ehooks_default_split, + ehooks_default_merge +}; diff --git a/contrib/jemalloc/src/emap.c b/contrib/jemalloc/src/emap.c new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/src/emap.c @@ -0,0 +1,386 @@ +#include "jemalloc/internal/jemalloc_preamble.h" +#include "jemalloc/internal/jemalloc_internal_includes.h" + +#include "jemalloc/internal/emap.h" + +enum emap_lock_result_e { + emap_lock_result_success, + emap_lock_result_failure, + emap_lock_result_no_extent +}; +typedef enum emap_lock_result_e emap_lock_result_t; + +bool +emap_init(emap_t *emap, base_t *base, bool zeroed) { + return rtree_new(&emap->rtree, base, zeroed); +} + +void +emap_update_edata_state(tsdn_t *tsdn, emap_t *emap, edata_t *edata, + extent_state_t state) { + witness_assert_positive_depth_to_rank(tsdn_witness_tsdp_get(tsdn), + WITNESS_RANK_CORE); + + edata_state_set(edata, state); + + EMAP_DECLARE_RTREE_CTX; + rtree_leaf_elm_t *elm1 = rtree_leaf_elm_lookup(tsdn, &emap->rtree, + rtree_ctx, (uintptr_t)edata_base_get(edata), /* dependent */ true, + /* init_missing */ false); + assert(elm1 != NULL); + rtree_leaf_elm_t *elm2 = edata_size_get(edata) == PAGE ? NULL : + rtree_leaf_elm_lookup(tsdn, &emap->rtree, rtree_ctx, + (uintptr_t)edata_last_get(edata), /* dependent */ true, + /* init_missing */ false); + + rtree_leaf_elm_state_update(tsdn, &emap->rtree, elm1, elm2, state); + + emap_assert_mapped(tsdn, emap, edata); +} + +static inline edata_t * +emap_try_acquire_edata_neighbor_impl(tsdn_t *tsdn, emap_t *emap, edata_t *edata, + extent_pai_t pai, extent_state_t expected_state, bool forward, + bool expanding) { + witness_assert_positive_depth_to_rank(tsdn_witness_tsdp_get(tsdn), + WITNESS_RANK_CORE); + assert(!edata_guarded_get(edata)); + assert(!expanding || forward); + assert(!edata_state_in_transition(expected_state)); + assert(expected_state == extent_state_dirty || + expected_state == extent_state_muzzy || + expected_state == extent_state_retained); + + void *neighbor_addr = forward ? edata_past_get(edata) : + edata_before_get(edata); + /* + * This is subtle; the rtree code asserts that its input pointer is + * non-NULL, and this is a useful thing to check. But it's possible + * that edata corresponds to an address of (void *)PAGE (in practice, + * this has only been observed on FreeBSD when address-space + * randomization is on, but it could in principle happen anywhere). In + * this case, edata_before_get(edata) is NULL, triggering the assert. + */ + if (neighbor_addr == NULL) { + return NULL; + } + + EMAP_DECLARE_RTREE_CTX; + rtree_leaf_elm_t *elm = rtree_leaf_elm_lookup(tsdn, &emap->rtree, + rtree_ctx, (uintptr_t)neighbor_addr, /* dependent*/ false, + /* init_missing */ false); + if (elm == NULL) { + return NULL; + } + + rtree_contents_t neighbor_contents = rtree_leaf_elm_read(tsdn, + &emap->rtree, elm, /* dependent */ true); + if (!extent_can_acquire_neighbor(edata, neighbor_contents, pai, + expected_state, forward, expanding)) { + return NULL; + } + + /* From this point, the neighbor edata can be safely acquired. */ + edata_t *neighbor = neighbor_contents.edata; + assert(edata_state_get(neighbor) == expected_state); + emap_update_edata_state(tsdn, emap, neighbor, extent_state_merging); + if (expanding) { + extent_assert_can_expand(edata, neighbor); + } else { + extent_assert_can_coalesce(edata, neighbor); + } + + return neighbor; +} + +edata_t * +emap_try_acquire_edata_neighbor(tsdn_t *tsdn, emap_t *emap, edata_t *edata, + extent_pai_t pai, extent_state_t expected_state, bool forward) { + return emap_try_acquire_edata_neighbor_impl(tsdn, emap, edata, pai, + expected_state, forward, /* expand */ false); +} + +edata_t * +emap_try_acquire_edata_neighbor_expand(tsdn_t *tsdn, emap_t *emap, + edata_t *edata, extent_pai_t pai, extent_state_t expected_state) { + /* Try expanding forward. */ + return emap_try_acquire_edata_neighbor_impl(tsdn, emap, edata, pai, + expected_state, /* forward */ true, /* expand */ true); +} + +void +emap_release_edata(tsdn_t *tsdn, emap_t *emap, edata_t *edata, + extent_state_t new_state) { + assert(emap_edata_in_transition(tsdn, emap, edata)); + assert(emap_edata_is_acquired(tsdn, emap, edata)); + + emap_update_edata_state(tsdn, emap, edata, new_state); +} + +static bool +emap_rtree_leaf_elms_lookup(tsdn_t *tsdn, emap_t *emap, rtree_ctx_t *rtree_ctx, + const edata_t *edata, bool dependent, bool init_missing, + rtree_leaf_elm_t **r_elm_a, rtree_leaf_elm_t **r_elm_b) { + *r_elm_a = rtree_leaf_elm_lookup(tsdn, &emap->rtree, rtree_ctx, + (uintptr_t)edata_base_get(edata), dependent, init_missing); + if (!dependent && *r_elm_a == NULL) { + return true; + } + assert(*r_elm_a != NULL); + + *r_elm_b = rtree_leaf_elm_lookup(tsdn, &emap->rtree, rtree_ctx, + (uintptr_t)edata_last_get(edata), dependent, init_missing); + if (!dependent && *r_elm_b == NULL) { + return true; + } + assert(*r_elm_b != NULL); + + return false; +} + +static void +emap_rtree_write_acquired(tsdn_t *tsdn, emap_t *emap, rtree_leaf_elm_t *elm_a, + rtree_leaf_elm_t *elm_b, edata_t *edata, szind_t szind, bool slab) { + rtree_contents_t contents; + contents.edata = edata; + contents.metadata.szind = szind; + contents.metadata.slab = slab; + contents.metadata.is_head = (edata == NULL) ? false : + edata_is_head_get(edata); + contents.metadata.state = (edata == NULL) ? 0 : edata_state_get(edata); + rtree_leaf_elm_write(tsdn, &emap->rtree, elm_a, contents); + if (elm_b != NULL) { + rtree_leaf_elm_write(tsdn, &emap->rtree, elm_b, contents); + } +} + +bool +emap_register_boundary(tsdn_t *tsdn, emap_t *emap, edata_t *edata, + szind_t szind, bool slab) { + assert(edata_state_get(edata) == extent_state_active); + EMAP_DECLARE_RTREE_CTX; + + rtree_leaf_elm_t *elm_a, *elm_b; + bool err = emap_rtree_leaf_elms_lookup(tsdn, emap, rtree_ctx, edata, + false, true, &elm_a, &elm_b); + if (err) { + return true; + } + assert(rtree_leaf_elm_read(tsdn, &emap->rtree, elm_a, + /* dependent */ false).edata == NULL); + assert(rtree_leaf_elm_read(tsdn, &emap->rtree, elm_b, + /* dependent */ false).edata == NULL); + emap_rtree_write_acquired(tsdn, emap, elm_a, elm_b, edata, szind, slab); + return false; +} + +/* Invoked *after* emap_register_boundary. */ +void +emap_register_interior(tsdn_t *tsdn, emap_t *emap, edata_t *edata, + szind_t szind) { + EMAP_DECLARE_RTREE_CTX; + + assert(edata_slab_get(edata)); + assert(edata_state_get(edata) == extent_state_active); + + if (config_debug) { + /* Making sure the boundary is registered already. */ + rtree_leaf_elm_t *elm_a, *elm_b; + bool err = emap_rtree_leaf_elms_lookup(tsdn, emap, rtree_ctx, + edata, /* dependent */ true, /* init_missing */ false, + &elm_a, &elm_b); + assert(!err); + rtree_contents_t contents_a, contents_b; + contents_a = rtree_leaf_elm_read(tsdn, &emap->rtree, elm_a, + /* dependent */ true); + contents_b = rtree_leaf_elm_read(tsdn, &emap->rtree, elm_b, + /* dependent */ true); + assert(contents_a.edata == edata && contents_b.edata == edata); + assert(contents_a.metadata.slab && contents_b.metadata.slab); + } + + rtree_contents_t contents; + contents.edata = edata; + contents.metadata.szind = szind; + contents.metadata.slab = true; + contents.metadata.state = extent_state_active; + contents.metadata.is_head = false; /* Not allowed to access. */ + + assert(edata_size_get(edata) > (2 << LG_PAGE)); + rtree_write_range(tsdn, &emap->rtree, rtree_ctx, + (uintptr_t)edata_base_get(edata) + PAGE, + (uintptr_t)edata_last_get(edata) - PAGE, contents); +} + +void +emap_deregister_boundary(tsdn_t *tsdn, emap_t *emap, edata_t *edata) { + /* + * The edata must be either in an acquired state, or protected by state + * based locks. + */ + if (!emap_edata_is_acquired(tsdn, emap, edata)) { + witness_assert_positive_depth_to_rank( + tsdn_witness_tsdp_get(tsdn), WITNESS_RANK_CORE); + } + + EMAP_DECLARE_RTREE_CTX; + rtree_leaf_elm_t *elm_a, *elm_b; + + emap_rtree_leaf_elms_lookup(tsdn, emap, rtree_ctx, edata, + true, false, &elm_a, &elm_b); + emap_rtree_write_acquired(tsdn, emap, elm_a, elm_b, NULL, SC_NSIZES, + false); +} + +void +emap_deregister_interior(tsdn_t *tsdn, emap_t *emap, edata_t *edata) { + EMAP_DECLARE_RTREE_CTX; + + assert(edata_slab_get(edata)); + if (edata_size_get(edata) > (2 << LG_PAGE)) { + rtree_clear_range(tsdn, &emap->rtree, rtree_ctx, + (uintptr_t)edata_base_get(edata) + PAGE, + (uintptr_t)edata_last_get(edata) - PAGE); + } +} + +void +emap_remap(tsdn_t *tsdn, emap_t *emap, edata_t *edata, szind_t szind, + bool slab) { + EMAP_DECLARE_RTREE_CTX; + + if (szind != SC_NSIZES) { + rtree_contents_t contents; + contents.edata = edata; + contents.metadata.szind = szind; + contents.metadata.slab = slab; + contents.metadata.is_head = edata_is_head_get(edata); + contents.metadata.state = edata_state_get(edata); + + rtree_write(tsdn, &emap->rtree, rtree_ctx, + (uintptr_t)edata_addr_get(edata), contents); + /* + * Recall that this is called only for active->inactive and + * inactive->active transitions (since only active extents have + * meaningful values for szind and slab). Active, non-slab + * extents only need to handle lookups at their head (on + * deallocation), so we don't bother filling in the end + * boundary. + * + * For slab extents, we do the end-mapping change. This still + * leaves the interior unmodified; an emap_register_interior + * call is coming in those cases, though. + */ + if (slab && edata_size_get(edata) > PAGE) { + uintptr_t key = (uintptr_t)edata_past_get(edata) + - (uintptr_t)PAGE; + rtree_write(tsdn, &emap->rtree, rtree_ctx, key, + contents); + } + } +} + +bool +emap_split_prepare(tsdn_t *tsdn, emap_t *emap, emap_prepare_t *prepare, + edata_t *edata, size_t size_a, edata_t *trail, size_t size_b) { + EMAP_DECLARE_RTREE_CTX; + + /* + * We use incorrect constants for things like arena ind, zero, ranged, + * and commit state, and head status. This is a fake edata_t, used to + * facilitate a lookup. + */ + edata_t lead = {0}; + edata_init(&lead, 0U, edata_addr_get(edata), size_a, false, 0, 0, + extent_state_active, false, false, EXTENT_PAI_PAC, EXTENT_NOT_HEAD); + + emap_rtree_leaf_elms_lookup(tsdn, emap, rtree_ctx, &lead, false, true, + &prepare->lead_elm_a, &prepare->lead_elm_b); + emap_rtree_leaf_elms_lookup(tsdn, emap, rtree_ctx, trail, false, true, + &prepare->trail_elm_a, &prepare->trail_elm_b); + + if (prepare->lead_elm_a == NULL || prepare->lead_elm_b == NULL + || prepare->trail_elm_a == NULL || prepare->trail_elm_b == NULL) { + return true; + } + return false; +} + +void +emap_split_commit(tsdn_t *tsdn, emap_t *emap, emap_prepare_t *prepare, + edata_t *lead, size_t size_a, edata_t *trail, size_t size_b) { + /* + * We should think about not writing to the lead leaf element. We can + * get into situations where a racing realloc-like call can disagree + * with a size lookup request. I think it's fine to declare that these + * situations are race bugs, but there's an argument to be made that for + * things like xallocx, a size lookup call should return either the old + * size or the new size, but not anything else. + */ + emap_rtree_write_acquired(tsdn, emap, prepare->lead_elm_a, + prepare->lead_elm_b, lead, SC_NSIZES, /* slab */ false); + emap_rtree_write_acquired(tsdn, emap, prepare->trail_elm_a, + prepare->trail_elm_b, trail, SC_NSIZES, /* slab */ false); +} + +void +emap_merge_prepare(tsdn_t *tsdn, emap_t *emap, emap_prepare_t *prepare, + edata_t *lead, edata_t *trail) { + EMAP_DECLARE_RTREE_CTX; + emap_rtree_leaf_elms_lookup(tsdn, emap, rtree_ctx, lead, true, false, + &prepare->lead_elm_a, &prepare->lead_elm_b); + emap_rtree_leaf_elms_lookup(tsdn, emap, rtree_ctx, trail, true, false, + &prepare->trail_elm_a, &prepare->trail_elm_b); +} + +void +emap_merge_commit(tsdn_t *tsdn, emap_t *emap, emap_prepare_t *prepare, + edata_t *lead, edata_t *trail) { + rtree_contents_t clear_contents; + clear_contents.edata = NULL; + clear_contents.metadata.szind = SC_NSIZES; + clear_contents.metadata.slab = false; + clear_contents.metadata.is_head = false; + clear_contents.metadata.state = (extent_state_t)0; + + if (prepare->lead_elm_b != NULL) { + rtree_leaf_elm_write(tsdn, &emap->rtree, + prepare->lead_elm_b, clear_contents); + } + + rtree_leaf_elm_t *merged_b; + if (prepare->trail_elm_b != NULL) { + rtree_leaf_elm_write(tsdn, &emap->rtree, + prepare->trail_elm_a, clear_contents); + merged_b = prepare->trail_elm_b; + } else { + merged_b = prepare->trail_elm_a; + } + + emap_rtree_write_acquired(tsdn, emap, prepare->lead_elm_a, merged_b, + lead, SC_NSIZES, false); +} + +void +emap_do_assert_mapped(tsdn_t *tsdn, emap_t *emap, edata_t *edata) { + EMAP_DECLARE_RTREE_CTX; + + rtree_contents_t contents = rtree_read(tsdn, &emap->rtree, rtree_ctx, + (uintptr_t)edata_base_get(edata)); + assert(contents.edata == edata); + assert(contents.metadata.is_head == edata_is_head_get(edata)); + assert(contents.metadata.state == edata_state_get(edata)); +} + +void +emap_do_assert_not_mapped(tsdn_t *tsdn, emap_t *emap, edata_t *edata) { + emap_full_alloc_ctx_t context1 = {0}; + emap_full_alloc_ctx_try_lookup(tsdn, emap, edata_base_get(edata), + &context1); + assert(context1.edata == NULL); + + emap_full_alloc_ctx_t context2 = {0}; + emap_full_alloc_ctx_try_lookup(tsdn, emap, edata_last_get(edata), + &context2); + assert(context2.edata == NULL); +} diff --git a/contrib/jemalloc/src/eset.c b/contrib/jemalloc/src/eset.c new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/src/eset.c @@ -0,0 +1,282 @@ +#include "jemalloc/internal/jemalloc_preamble.h" +#include "jemalloc/internal/jemalloc_internal_includes.h" + +#include "jemalloc/internal/eset.h" + +#define ESET_NPSIZES (SC_NPSIZES + 1) + +static void +eset_bin_init(eset_bin_t *bin) { + edata_heap_new(&bin->heap); + /* + * heap_min doesn't need initialization; it gets filled in when the bin + * goes from non-empty to empty. + */ +} + +static void +eset_bin_stats_init(eset_bin_stats_t *bin_stats) { + atomic_store_zu(&bin_stats->nextents, 0, ATOMIC_RELAXED); + atomic_store_zu(&bin_stats->nbytes, 0, ATOMIC_RELAXED); +} + +void +eset_init(eset_t *eset, extent_state_t state) { + for (unsigned i = 0; i < ESET_NPSIZES; i++) { + eset_bin_init(&eset->bins[i]); + eset_bin_stats_init(&eset->bin_stats[i]); + } + fb_init(eset->bitmap, ESET_NPSIZES); + edata_list_inactive_init(&eset->lru); + eset->state = state; +} + +size_t +eset_npages_get(eset_t *eset) { + return atomic_load_zu(&eset->npages, ATOMIC_RELAXED); +} + +size_t +eset_nextents_get(eset_t *eset, pszind_t pind) { + return atomic_load_zu(&eset->bin_stats[pind].nextents, ATOMIC_RELAXED); +} + +size_t +eset_nbytes_get(eset_t *eset, pszind_t pind) { + return atomic_load_zu(&eset->bin_stats[pind].nbytes, ATOMIC_RELAXED); +} + +static void +eset_stats_add(eset_t *eset, pszind_t pind, size_t sz) { + size_t cur = atomic_load_zu(&eset->bin_stats[pind].nextents, + ATOMIC_RELAXED); + atomic_store_zu(&eset->bin_stats[pind].nextents, cur + 1, + ATOMIC_RELAXED); + cur = atomic_load_zu(&eset->bin_stats[pind].nbytes, ATOMIC_RELAXED); + atomic_store_zu(&eset->bin_stats[pind].nbytes, cur + sz, + ATOMIC_RELAXED); +} + +static void +eset_stats_sub(eset_t *eset, pszind_t pind, size_t sz) { + size_t cur = atomic_load_zu(&eset->bin_stats[pind].nextents, + ATOMIC_RELAXED); + atomic_store_zu(&eset->bin_stats[pind].nextents, cur - 1, + ATOMIC_RELAXED); + cur = atomic_load_zu(&eset->bin_stats[pind].nbytes, ATOMIC_RELAXED); + atomic_store_zu(&eset->bin_stats[pind].nbytes, cur - sz, + ATOMIC_RELAXED); +} + +void +eset_insert(eset_t *eset, edata_t *edata) { + assert(edata_state_get(edata) == eset->state); + + size_t size = edata_size_get(edata); + size_t psz = sz_psz_quantize_floor(size); + pszind_t pind = sz_psz2ind(psz); + + edata_cmp_summary_t edata_cmp_summary = edata_cmp_summary_get(edata); + if (edata_heap_empty(&eset->bins[pind].heap)) { + fb_set(eset->bitmap, ESET_NPSIZES, (size_t)pind); + /* Only element is automatically the min element. */ + eset->bins[pind].heap_min = edata_cmp_summary; + } else { + /* + * There's already a min element; update the summary if we're + * about to insert a lower one. + */ + if (edata_cmp_summary_comp(edata_cmp_summary, + eset->bins[pind].heap_min) < 0) { + eset->bins[pind].heap_min = edata_cmp_summary; + } + } + edata_heap_insert(&eset->bins[pind].heap, edata); + + if (config_stats) { + eset_stats_add(eset, pind, size); + } + + edata_list_inactive_append(&eset->lru, edata); + size_t npages = size >> LG_PAGE; + /* + * All modifications to npages hold the mutex (as asserted above), so we + * don't need an atomic fetch-add; we can get by with a load followed by + * a store. + */ + size_t cur_eset_npages = + atomic_load_zu(&eset->npages, ATOMIC_RELAXED); + atomic_store_zu(&eset->npages, cur_eset_npages + npages, + ATOMIC_RELAXED); +} + +void +eset_remove(eset_t *eset, edata_t *edata) { + assert(edata_state_get(edata) == eset->state || + edata_state_in_transition(edata_state_get(edata))); + + size_t size = edata_size_get(edata); + size_t psz = sz_psz_quantize_floor(size); + pszind_t pind = sz_psz2ind(psz); + if (config_stats) { + eset_stats_sub(eset, pind, size); + } + + edata_cmp_summary_t edata_cmp_summary = edata_cmp_summary_get(edata); + edata_heap_remove(&eset->bins[pind].heap, edata); + if (edata_heap_empty(&eset->bins[pind].heap)) { + fb_unset(eset->bitmap, ESET_NPSIZES, (size_t)pind); + } else { + /* + * This is a little weird; we compare if the summaries are + * equal, rather than if the edata we removed was the heap + * minimum. The reason why is that getting the heap minimum + * can cause a pairing heap merge operation. We can avoid this + * if we only update the min if it's changed, in which case the + * summaries of the removed element and the min element should + * compare equal. + */ + if (edata_cmp_summary_comp(edata_cmp_summary, + eset->bins[pind].heap_min) == 0) { + eset->bins[pind].heap_min = edata_cmp_summary_get( + edata_heap_first(&eset->bins[pind].heap)); + } + } + edata_list_inactive_remove(&eset->lru, edata); + size_t npages = size >> LG_PAGE; + /* + * As in eset_insert, we hold eset->mtx and so don't need atomic + * operations for updating eset->npages. + */ + size_t cur_extents_npages = + atomic_load_zu(&eset->npages, ATOMIC_RELAXED); + assert(cur_extents_npages >= npages); + atomic_store_zu(&eset->npages, + cur_extents_npages - (size >> LG_PAGE), ATOMIC_RELAXED); +} + +/* + * Find an extent with size [min_size, max_size) to satisfy the alignment + * requirement. For each size, try only the first extent in the heap. + */ +static edata_t * +eset_fit_alignment(eset_t *eset, size_t min_size, size_t max_size, + size_t alignment) { + pszind_t pind = sz_psz2ind(sz_psz_quantize_ceil(min_size)); + pszind_t pind_max = sz_psz2ind(sz_psz_quantize_ceil(max_size)); + + for (pszind_t i = + (pszind_t)fb_ffs(eset->bitmap, ESET_NPSIZES, (size_t)pind); + i < pind_max; + i = (pszind_t)fb_ffs(eset->bitmap, ESET_NPSIZES, (size_t)i + 1)) { + assert(i < SC_NPSIZES); + assert(!edata_heap_empty(&eset->bins[i].heap)); + edata_t *edata = edata_heap_first(&eset->bins[i].heap); + uintptr_t base = (uintptr_t)edata_base_get(edata); + size_t candidate_size = edata_size_get(edata); + assert(candidate_size >= min_size); + + uintptr_t next_align = ALIGNMENT_CEILING((uintptr_t)base, + PAGE_CEILING(alignment)); + if (base > next_align || base + candidate_size <= next_align) { + /* Overflow or not crossing the next alignment. */ + continue; + } + + size_t leadsize = next_align - base; + if (candidate_size - leadsize >= min_size) { + return edata; + } + } + + return NULL; +} + +/* + * Do first-fit extent selection, i.e. select the oldest/lowest extent that is + * large enough. + * + * lg_max_fit is the (log of the) maximum ratio between the requested size and + * the returned size that we'll allow. This can reduce fragmentation by + * avoiding reusing and splitting large extents for smaller sizes. In practice, + * it's set to opt_lg_extent_max_active_fit for the dirty eset and SC_PTR_BITS + * for others. + */ +static edata_t * +eset_first_fit(eset_t *eset, size_t size, bool exact_only, + unsigned lg_max_fit) { + edata_t *ret = NULL; + edata_cmp_summary_t ret_summ JEMALLOC_CC_SILENCE_INIT({0}); + + pszind_t pind = sz_psz2ind(sz_psz_quantize_ceil(size)); + + if (exact_only) { + return edata_heap_empty(&eset->bins[pind].heap) ? NULL : + edata_heap_first(&eset->bins[pind].heap); + } + + for (pszind_t i = + (pszind_t)fb_ffs(eset->bitmap, ESET_NPSIZES, (size_t)pind); + i < ESET_NPSIZES; + i = (pszind_t)fb_ffs(eset->bitmap, ESET_NPSIZES, (size_t)i + 1)) { + assert(!edata_heap_empty(&eset->bins[i].heap)); + if (lg_max_fit == SC_PTR_BITS) { + /* + * We'll shift by this below, and shifting out all the + * bits is undefined. Decreasing is safe, since the + * page size is larger than 1 byte. + */ + lg_max_fit = SC_PTR_BITS - 1; + } + if ((sz_pind2sz(i) >> lg_max_fit) > size) { + break; + } + if (ret == NULL || edata_cmp_summary_comp( + eset->bins[i].heap_min, ret_summ) < 0) { + /* + * We grab the edata as early as possible, even though + * we might change it later. Practically, a large + * portion of eset_fit calls succeed at the first valid + * index, so this doesn't cost much, and we get the + * effect of prefetching the edata as early as possible. + */ + edata_t *edata = edata_heap_first(&eset->bins[i].heap); + assert(edata_size_get(edata) >= size); + assert(ret == NULL || edata_snad_comp(edata, ret) < 0); + assert(ret == NULL || edata_cmp_summary_comp( + eset->bins[i].heap_min, + edata_cmp_summary_get(edata)) == 0); + ret = edata; + ret_summ = eset->bins[i].heap_min; + } + if (i == SC_NPSIZES) { + break; + } + assert(i < SC_NPSIZES); + } + + return ret; +} + +edata_t * +eset_fit(eset_t *eset, size_t esize, size_t alignment, bool exact_only, + unsigned lg_max_fit) { + size_t max_size = esize + PAGE_CEILING(alignment) - PAGE; + /* Beware size_t wrap-around. */ + if (max_size < esize) { + return NULL; + } + + edata_t *edata = eset_first_fit(eset, max_size, exact_only, lg_max_fit); + + if (alignment > PAGE && edata == NULL) { + /* + * max_size guarantees the alignment requirement but is rather + * pessimistic. Next we try to satisfy the aligned allocation + * with sizes in [esize, max_size). + */ + edata = eset_fit_alignment(eset, esize, max_size, alignment); + } + + return edata; +} diff --git a/contrib/jemalloc/src/exp_grow.c b/contrib/jemalloc/src/exp_grow.c new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/src/exp_grow.c @@ -0,0 +1,8 @@ +#include "jemalloc/internal/jemalloc_preamble.h" +#include "jemalloc/internal/jemalloc_internal_includes.h" + +void +exp_grow_init(exp_grow_t *exp_grow) { + exp_grow->next = sz_psz2ind(HUGEPAGE); + exp_grow->limit = sz_psz2ind(SC_LARGE_MAXCLASS); +} diff --git a/contrib/jemalloc/src/extent.c b/contrib/jemalloc/src/extent.c --- a/contrib/jemalloc/src/extent.c +++ b/contrib/jemalloc/src/extent.c @@ -1,93 +1,28 @@ -#define JEMALLOC_EXTENT_C_ #include "jemalloc/internal/jemalloc_preamble.h" #include "jemalloc/internal/jemalloc_internal_includes.h" #include "jemalloc/internal/assert.h" +#include "jemalloc/internal/emap.h" #include "jemalloc/internal/extent_dss.h" #include "jemalloc/internal/extent_mmap.h" #include "jemalloc/internal/ph.h" -#include "jemalloc/internal/rtree.h" #include "jemalloc/internal/mutex.h" -#include "jemalloc/internal/mutex_pool.h" /******************************************************************************/ /* Data. */ -rtree_t extents_rtree; -/* Keyed by the address of the extent_t being protected. */ -mutex_pool_t extent_mutex_pool; - size_t opt_lg_extent_max_active_fit = LG_EXTENT_MAX_ACTIVE_FIT_DEFAULT; -static const bitmap_info_t extents_bitmap_info = - BITMAP_INFO_INITIALIZER(SC_NPSIZES+1); - -static void *extent_alloc_default(extent_hooks_t *extent_hooks, void *new_addr, - size_t size, size_t alignment, bool *zero, bool *commit, - unsigned arena_ind); -static bool extent_dalloc_default(extent_hooks_t *extent_hooks, void *addr, - size_t size, bool committed, unsigned arena_ind); -static void extent_destroy_default(extent_hooks_t *extent_hooks, void *addr, - size_t size, bool committed, unsigned arena_ind); -static bool extent_commit_default(extent_hooks_t *extent_hooks, void *addr, - size_t size, size_t offset, size_t length, unsigned arena_ind); -static bool extent_commit_impl(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, extent_t *extent, size_t offset, - size_t length, bool growing_retained); -static bool extent_decommit_default(extent_hooks_t *extent_hooks, - void *addr, size_t size, size_t offset, size_t length, unsigned arena_ind); -#ifdef PAGES_CAN_PURGE_LAZY -static bool extent_purge_lazy_default(extent_hooks_t *extent_hooks, void *addr, - size_t size, size_t offset, size_t length, unsigned arena_ind); -#endif -static bool extent_purge_lazy_impl(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, extent_t *extent, size_t offset, - size_t length, bool growing_retained); -#ifdef PAGES_CAN_PURGE_FORCED -static bool extent_purge_forced_default(extent_hooks_t *extent_hooks, - void *addr, size_t size, size_t offset, size_t length, unsigned arena_ind); -#endif -static bool extent_purge_forced_impl(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, extent_t *extent, size_t offset, - size_t length, bool growing_retained); -static bool extent_split_default(extent_hooks_t *extent_hooks, void *addr, - size_t size, size_t size_a, size_t size_b, bool committed, - unsigned arena_ind); -static extent_t *extent_split_impl(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, extent_t *extent, size_t size_a, - szind_t szind_a, bool slab_a, size_t size_b, szind_t szind_b, bool slab_b, - bool growing_retained); -static bool extent_merge_default(extent_hooks_t *extent_hooks, void *addr_a, - size_t size_a, void *addr_b, size_t size_b, bool committed, - unsigned arena_ind); -static bool extent_merge_impl(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, extent_t *a, extent_t *b, - bool growing_retained); - -const extent_hooks_t extent_hooks_default = { - extent_alloc_default, - extent_dalloc_default, - extent_destroy_default, - extent_commit_default, - extent_decommit_default -#ifdef PAGES_CAN_PURGE_LAZY - , - extent_purge_lazy_default -#else - , - NULL -#endif -#ifdef PAGES_CAN_PURGE_FORCED - , - extent_purge_forced_default -#else - , - NULL -#endif - , - extent_split_default, - extent_merge_default -}; +static bool extent_commit_impl(tsdn_t *tsdn, ehooks_t *ehooks, edata_t *edata, + size_t offset, size_t length, bool growing_retained); +static bool extent_purge_lazy_impl(tsdn_t *tsdn, ehooks_t *ehooks, + edata_t *edata, size_t offset, size_t length, bool growing_retained); +static bool extent_purge_forced_impl(tsdn_t *tsdn, ehooks_t *ehooks, + edata_t *edata, size_t offset, size_t length, bool growing_retained); +static edata_t *extent_split_impl(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, + edata_t *edata, size_t size_a, size_t size_b, bool holding_core_locks); +static bool extent_merge_impl(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, + edata_t *a, edata_t *b, bool holding_core_locks); /* Used exclusively for gdump triggering. */ static atomic_zu_t curpages; @@ -99,503 +34,158 @@ * definition. */ -static void extent_deregister(tsdn_t *tsdn, extent_t *extent); -static extent_t *extent_recycle(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, extents_t *extents, void *new_addr, - size_t usize, size_t pad, size_t alignment, bool slab, szind_t szind, - bool *zero, bool *commit, bool growing_retained); -static extent_t *extent_try_coalesce(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, rtree_ctx_t *rtree_ctx, extents_t *extents, - extent_t *extent, bool *coalesced, bool growing_retained); -static void extent_record(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, extents_t *extents, extent_t *extent, - bool growing_retained); +static void extent_deregister(tsdn_t *tsdn, pac_t *pac, edata_t *edata); +static edata_t *extent_recycle(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, + ecache_t *ecache, edata_t *expand_edata, size_t usize, size_t alignment, + bool zero, bool *commit, bool growing_retained, bool guarded); +static edata_t *extent_try_coalesce(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, + ecache_t *ecache, edata_t *edata, bool *coalesced); +static edata_t *extent_alloc_retained(tsdn_t *tsdn, pac_t *pac, + ehooks_t *ehooks, edata_t *expand_edata, size_t size, size_t alignment, + bool zero, bool *commit, bool guarded); /******************************************************************************/ -#define ATTR_NONE /* does nothing */ - -ph_gen(ATTR_NONE, extent_avail_, extent_tree_t, extent_t, ph_link, - extent_esnead_comp) - -#undef ATTR_NONE - -typedef enum { - lock_result_success, - lock_result_failure, - lock_result_no_extent -} lock_result_t; - -static lock_result_t -extent_rtree_leaf_elm_try_lock(tsdn_t *tsdn, rtree_leaf_elm_t *elm, - extent_t **result, bool inactive_only) { - extent_t *extent1 = rtree_leaf_elm_extent_read(tsdn, &extents_rtree, - elm, true); - - /* Slab implies active extents and should be skipped. */ - if (extent1 == NULL || (inactive_only && rtree_leaf_elm_slab_read(tsdn, - &extents_rtree, elm, true))) { - return lock_result_no_extent; - } - - /* - * It's possible that the extent changed out from under us, and with it - * the leaf->extent mapping. We have to recheck while holding the lock. - */ - extent_lock(tsdn, extent1); - extent_t *extent2 = rtree_leaf_elm_extent_read(tsdn, - &extents_rtree, elm, true); - - if (extent1 == extent2) { - *result = extent1; - return lock_result_success; - } else { - extent_unlock(tsdn, extent1); - return lock_result_failure; - } -} - -/* - * Returns a pool-locked extent_t * if there's one associated with the given - * address, and NULL otherwise. - */ -static extent_t * -extent_lock_from_addr(tsdn_t *tsdn, rtree_ctx_t *rtree_ctx, void *addr, - bool inactive_only) { - extent_t *ret = NULL; - rtree_leaf_elm_t *elm = rtree_leaf_elm_lookup(tsdn, &extents_rtree, - rtree_ctx, (uintptr_t)addr, false, false); - if (elm == NULL) { - return NULL; - } - lock_result_t lock_result; - do { - lock_result = extent_rtree_leaf_elm_try_lock(tsdn, elm, &ret, - inactive_only); - } while (lock_result == lock_result_failure); - return ret; -} - -extent_t * -extent_alloc(tsdn_t *tsdn, arena_t *arena) { - malloc_mutex_lock(tsdn, &arena->extent_avail_mtx); - extent_t *extent = extent_avail_first(&arena->extent_avail); - if (extent == NULL) { - malloc_mutex_unlock(tsdn, &arena->extent_avail_mtx); - return base_alloc_extent(tsdn, arena->base); - } - extent_avail_remove(&arena->extent_avail, extent); - atomic_fetch_sub_zu(&arena->extent_avail_cnt, 1, ATOMIC_RELAXED); - malloc_mutex_unlock(tsdn, &arena->extent_avail_mtx); - return extent; -} - -void -extent_dalloc(tsdn_t *tsdn, arena_t *arena, extent_t *extent) { - malloc_mutex_lock(tsdn, &arena->extent_avail_mtx); - extent_avail_insert(&arena->extent_avail, extent); - atomic_fetch_add_zu(&arena->extent_avail_cnt, 1, ATOMIC_RELAXED); - malloc_mutex_unlock(tsdn, &arena->extent_avail_mtx); -} - -extent_hooks_t * -extent_hooks_get(arena_t *arena) { - return base_extent_hooks_get(arena->base); -} - -extent_hooks_t * -extent_hooks_set(tsd_t *tsd, arena_t *arena, extent_hooks_t *extent_hooks) { - background_thread_info_t *info; - if (have_background_thread) { - info = arena_background_thread_info_get(arena); - malloc_mutex_lock(tsd_tsdn(tsd), &info->mtx); - } - extent_hooks_t *ret = base_extent_hooks_set(arena->base, extent_hooks); - if (have_background_thread) { - malloc_mutex_unlock(tsd_tsdn(tsd), &info->mtx); - } - - return ret; -} - -static void -extent_hooks_assure_initialized(arena_t *arena, - extent_hooks_t **r_extent_hooks) { - if (*r_extent_hooks == EXTENT_HOOKS_INITIALIZER) { - *r_extent_hooks = extent_hooks_get(arena); - } -} - -#ifndef JEMALLOC_JET -static -#endif size_t -extent_size_quantize_floor(size_t size) { - size_t ret; - pszind_t pind; - - assert(size > 0); - assert((size & PAGE_MASK) == 0); - - pind = sz_psz2ind(size - sz_large_pad + 1); - if (pind == 0) { - /* - * Avoid underflow. This short-circuit would also do the right - * thing for all sizes in the range for which there are - * PAGE-spaced size classes, but it's simplest to just handle - * the one case that would cause erroneous results. - */ - return size; - } - ret = sz_pind2sz(pind - 1) + sz_large_pad; - assert(ret <= size); - return ret; +extent_sn_next(pac_t *pac) { + return atomic_fetch_add_zu(&pac->extent_sn_next, 1, ATOMIC_RELAXED); } -#ifndef JEMALLOC_JET -static -#endif -size_t -extent_size_quantize_ceil(size_t size) { - size_t ret; - - assert(size > 0); - assert(size - sz_large_pad <= SC_LARGE_MAXCLASS); - assert((size & PAGE_MASK) == 0); - - ret = extent_size_quantize_floor(size); - if (ret < size) { - /* - * Skip a quantization that may have an adequately large extent, - * because under-sized extents may be mixed in. This only - * happens when an unusual size is requested, i.e. for aligned - * allocation, and is just one of several places where linear - * search would potentially find sufficiently aligned available - * memory somewhere lower. - */ - ret = sz_pind2sz(sz_psz2ind(ret - sz_large_pad + 1)) + - sz_large_pad; - } - return ret; +static inline bool +extent_may_force_decay(pac_t *pac) { + return !(pac_decay_ms_get(pac, extent_state_dirty) == -1 + || pac_decay_ms_get(pac, extent_state_muzzy) == -1); } -/* Generate pairing heap functions. */ -ph_gen(, extent_heap_, extent_heap_t, extent_t, ph_link, extent_snad_comp) +static bool +extent_try_delayed_coalesce(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, + ecache_t *ecache, edata_t *edata) { + emap_update_edata_state(tsdn, pac->emap, edata, extent_state_active); -bool -extents_init(tsdn_t *tsdn, extents_t *extents, extent_state_t state, - bool delay_coalesce) { - if (malloc_mutex_init(&extents->mtx, "extents", WITNESS_RANK_EXTENTS, - malloc_mutex_rank_exclusive)) { + bool coalesced; + edata = extent_try_coalesce(tsdn, pac, ehooks, ecache, + edata, &coalesced); + emap_update_edata_state(tsdn, pac->emap, edata, ecache->state); + + if (!coalesced) { return true; } - for (unsigned i = 0; i < SC_NPSIZES + 1; i++) { - extent_heap_new(&extents->heaps[i]); - } - bitmap_init(extents->bitmap, &extents_bitmap_info, true); - extent_list_init(&extents->lru); - atomic_store_zu(&extents->npages, 0, ATOMIC_RELAXED); - extents->state = state; - extents->delay_coalesce = delay_coalesce; + eset_insert(&ecache->eset, edata); return false; } -extent_state_t -extents_state_get(const extents_t *extents) { - return extents->state; -} - -size_t -extents_npages_get(extents_t *extents) { - return atomic_load_zu(&extents->npages, ATOMIC_RELAXED); -} - -size_t -extents_nextents_get(extents_t *extents, pszind_t pind) { - return atomic_load_zu(&extents->nextents[pind], ATOMIC_RELAXED); -} - -size_t -extents_nbytes_get(extents_t *extents, pszind_t pind) { - return atomic_load_zu(&extents->nbytes[pind], ATOMIC_RELAXED); -} - -static void -extents_stats_add(extents_t *extent, pszind_t pind, size_t sz) { - size_t cur = atomic_load_zu(&extent->nextents[pind], ATOMIC_RELAXED); - atomic_store_zu(&extent->nextents[pind], cur + 1, ATOMIC_RELAXED); - cur = atomic_load_zu(&extent->nbytes[pind], ATOMIC_RELAXED); - atomic_store_zu(&extent->nbytes[pind], cur + sz, ATOMIC_RELAXED); -} - -static void -extents_stats_sub(extents_t *extent, pszind_t pind, size_t sz) { - size_t cur = atomic_load_zu(&extent->nextents[pind], ATOMIC_RELAXED); - atomic_store_zu(&extent->nextents[pind], cur - 1, ATOMIC_RELAXED); - cur = atomic_load_zu(&extent->nbytes[pind], ATOMIC_RELAXED); - atomic_store_zu(&extent->nbytes[pind], cur - sz, ATOMIC_RELAXED); -} - -static void -extents_insert_locked(tsdn_t *tsdn, extents_t *extents, extent_t *extent) { - malloc_mutex_assert_owner(tsdn, &extents->mtx); - assert(extent_state_get(extent) == extents->state); - - size_t size = extent_size_get(extent); - size_t psz = extent_size_quantize_floor(size); - pszind_t pind = sz_psz2ind(psz); - if (extent_heap_empty(&extents->heaps[pind])) { - bitmap_unset(extents->bitmap, &extents_bitmap_info, - (size_t)pind); - } - extent_heap_insert(&extents->heaps[pind], extent); - - if (config_stats) { - extents_stats_add(extents, pind, size); - } - - extent_list_append(&extents->lru, extent); - size_t npages = size >> LG_PAGE; - /* - * All modifications to npages hold the mutex (as asserted above), so we - * don't need an atomic fetch-add; we can get by with a load followed by - * a store. - */ - size_t cur_extents_npages = - atomic_load_zu(&extents->npages, ATOMIC_RELAXED); - atomic_store_zu(&extents->npages, cur_extents_npages + npages, - ATOMIC_RELAXED); -} - -static void -extents_remove_locked(tsdn_t *tsdn, extents_t *extents, extent_t *extent) { - malloc_mutex_assert_owner(tsdn, &extents->mtx); - assert(extent_state_get(extent) == extents->state); - - size_t size = extent_size_get(extent); - size_t psz = extent_size_quantize_floor(size); - pszind_t pind = sz_psz2ind(psz); - extent_heap_remove(&extents->heaps[pind], extent); - - if (config_stats) { - extents_stats_sub(extents, pind, size); - } - - if (extent_heap_empty(&extents->heaps[pind])) { - bitmap_set(extents->bitmap, &extents_bitmap_info, - (size_t)pind); - } - extent_list_remove(&extents->lru, extent); - size_t npages = size >> LG_PAGE; - /* - * As in extents_insert_locked, we hold extents->mtx and so don't need - * atomic operations for updating extents->npages. - */ - size_t cur_extents_npages = - atomic_load_zu(&extents->npages, ATOMIC_RELAXED); - assert(cur_extents_npages >= npages); - atomic_store_zu(&extents->npages, - cur_extents_npages - (size >> LG_PAGE), ATOMIC_RELAXED); -} - -/* - * Find an extent with size [min_size, max_size) to satisfy the alignment - * requirement. For each size, try only the first extent in the heap. - */ -static extent_t * -extents_fit_alignment(extents_t *extents, size_t min_size, size_t max_size, - size_t alignment) { - pszind_t pind = sz_psz2ind(extent_size_quantize_ceil(min_size)); - pszind_t pind_max = sz_psz2ind(extent_size_quantize_ceil(max_size)); - - for (pszind_t i = (pszind_t)bitmap_ffu(extents->bitmap, - &extents_bitmap_info, (size_t)pind); i < pind_max; i = - (pszind_t)bitmap_ffu(extents->bitmap, &extents_bitmap_info, - (size_t)i+1)) { - assert(i < SC_NPSIZES); - assert(!extent_heap_empty(&extents->heaps[i])); - extent_t *extent = extent_heap_first(&extents->heaps[i]); - uintptr_t base = (uintptr_t)extent_base_get(extent); - size_t candidate_size = extent_size_get(extent); - assert(candidate_size >= min_size); - - uintptr_t next_align = ALIGNMENT_CEILING((uintptr_t)base, - PAGE_CEILING(alignment)); - if (base > next_align || base + candidate_size <= next_align) { - /* Overflow or not crossing the next alignment. */ - continue; - } - - size_t leadsize = next_align - base; - if (candidate_size - leadsize >= min_size) { - return extent; - } - } +edata_t * +ecache_alloc(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, ecache_t *ecache, + edata_t *expand_edata, size_t size, size_t alignment, bool zero, + bool guarded) { + assert(size != 0); + assert(alignment != 0); + witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), + WITNESS_RANK_CORE, 0); - return NULL; + bool commit = true; + edata_t *edata = extent_recycle(tsdn, pac, ehooks, ecache, expand_edata, + size, alignment, zero, &commit, false, guarded); + assert(edata == NULL || edata_pai_get(edata) == EXTENT_PAI_PAC); + assert(edata == NULL || edata_guarded_get(edata) == guarded); + return edata; } -/* - * Do first-fit extent selection, i.e. select the oldest/lowest extent that is - * large enough. - */ -static extent_t * -extents_first_fit_locked(tsdn_t *tsdn, arena_t *arena, extents_t *extents, - size_t size) { - extent_t *ret = NULL; - - pszind_t pind = sz_psz2ind(extent_size_quantize_ceil(size)); - - if (!maps_coalesce && !opt_retain) { - /* - * No split / merge allowed (Windows w/o retain). Try exact fit - * only. - */ - return extent_heap_empty(&extents->heaps[pind]) ? NULL : - extent_heap_first(&extents->heaps[pind]); - } +edata_t * +ecache_alloc_grow(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, ecache_t *ecache, + edata_t *expand_edata, size_t size, size_t alignment, bool zero, + bool guarded) { + assert(size != 0); + assert(alignment != 0); + witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), + WITNESS_RANK_CORE, 0); - for (pszind_t i = (pszind_t)bitmap_ffu(extents->bitmap, - &extents_bitmap_info, (size_t)pind); - i < SC_NPSIZES + 1; - i = (pszind_t)bitmap_ffu(extents->bitmap, &extents_bitmap_info, - (size_t)i+1)) { - assert(!extent_heap_empty(&extents->heaps[i])); - extent_t *extent = extent_heap_first(&extents->heaps[i]); - assert(extent_size_get(extent) >= size); - /* - * In order to reduce fragmentation, avoid reusing and splitting - * large extents for much smaller sizes. - * - * Only do check for dirty extents (delay_coalesce). - */ - if (extents->delay_coalesce && - (sz_pind2sz(i) >> opt_lg_extent_max_active_fit) > size) { - break; - } - if (ret == NULL || extent_snad_comp(extent, ret) < 0) { - ret = extent; + bool commit = true; + edata_t *edata = extent_alloc_retained(tsdn, pac, ehooks, expand_edata, + size, alignment, zero, &commit, guarded); + if (edata == NULL) { + if (opt_retain && expand_edata != NULL) { + /* + * When retain is enabled and trying to expand, we do + * not attempt extent_alloc_wrapper which does mmap that + * is very unlikely to succeed (unless it happens to be + * at the end). + */ + return NULL; } - if (i == SC_NPSIZES) { - break; + if (guarded) { + /* + * Means no cached guarded extents available (and no + * grow_retained was attempted). The pac_alloc flow + * will alloc regular extents to make new guarded ones. + */ + return NULL; } - assert(i < SC_NPSIZES); - } - - return ret; -} - -/* - * Do first-fit extent selection, where the selection policy choice is - * based on extents->delay_coalesce. - */ -static extent_t * -extents_fit_locked(tsdn_t *tsdn, arena_t *arena, extents_t *extents, - size_t esize, size_t alignment) { - malloc_mutex_assert_owner(tsdn, &extents->mtx); - - size_t max_size = esize + PAGE_CEILING(alignment) - PAGE; - /* Beware size_t wrap-around. */ - if (max_size < esize) { - return NULL; - } - - extent_t *extent = - extents_first_fit_locked(tsdn, arena, extents, max_size); - - if (alignment > PAGE && extent == NULL) { - /* - * max_size guarantees the alignment requirement but is rather - * pessimistic. Next we try to satisfy the aligned allocation - * with sizes in [esize, max_size). - */ - extent = extents_fit_alignment(extents, esize, max_size, - alignment); - } - - return extent; -} - -static bool -extent_try_delayed_coalesce(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, rtree_ctx_t *rtree_ctx, extents_t *extents, - extent_t *extent) { - extent_state_set(extent, extent_state_active); - bool coalesced; - extent = extent_try_coalesce(tsdn, arena, r_extent_hooks, rtree_ctx, - extents, extent, &coalesced, false); - extent_state_set(extent, extents_state_get(extents)); - - if (!coalesced) { - return true; + void *new_addr = (expand_edata == NULL) ? NULL : + edata_past_get(expand_edata); + edata = extent_alloc_wrapper(tsdn, pac, ehooks, new_addr, + size, alignment, zero, &commit, + /* growing_retained */ false); } - extents_insert_locked(tsdn, extents, extent); - return false; -} -extent_t * -extents_alloc(tsdn_t *tsdn, arena_t *arena, extent_hooks_t **r_extent_hooks, - extents_t *extents, void *new_addr, size_t size, size_t pad, - size_t alignment, bool slab, szind_t szind, bool *zero, bool *commit) { - assert(size + pad != 0); - assert(alignment != 0); - witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), - WITNESS_RANK_CORE, 0); - - extent_t *extent = extent_recycle(tsdn, arena, r_extent_hooks, extents, - new_addr, size, pad, alignment, slab, szind, zero, commit, false); - assert(extent == NULL || extent_dumpable_get(extent)); - return extent; + assert(edata == NULL || edata_pai_get(edata) == EXTENT_PAI_PAC); + return edata; } void -extents_dalloc(tsdn_t *tsdn, arena_t *arena, extent_hooks_t **r_extent_hooks, - extents_t *extents, extent_t *extent) { - assert(extent_base_get(extent) != NULL); - assert(extent_size_get(extent) != 0); - assert(extent_dumpable_get(extent)); +ecache_dalloc(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, ecache_t *ecache, + edata_t *edata) { + assert(edata_base_get(edata) != NULL); + assert(edata_size_get(edata) != 0); + assert(edata_pai_get(edata) == EXTENT_PAI_PAC); witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), WITNESS_RANK_CORE, 0); - extent_addr_set(extent, extent_base_get(extent)); - extent_zeroed_set(extent, false); + edata_addr_set(edata, edata_base_get(edata)); + edata_zeroed_set(edata, false); - extent_record(tsdn, arena, r_extent_hooks, extents, extent, false); + extent_record(tsdn, pac, ehooks, ecache, edata); } -extent_t * -extents_evict(tsdn_t *tsdn, arena_t *arena, extent_hooks_t **r_extent_hooks, - extents_t *extents, size_t npages_min) { - rtree_ctx_t rtree_ctx_fallback; - rtree_ctx_t *rtree_ctx = tsdn_rtree_ctx(tsdn, &rtree_ctx_fallback); - - malloc_mutex_lock(tsdn, &extents->mtx); +edata_t * +ecache_evict(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, + ecache_t *ecache, size_t npages_min) { + malloc_mutex_lock(tsdn, &ecache->mtx); /* * Get the LRU coalesced extent, if any. If coalescing was delayed, * the loop will iterate until the LRU extent is fully coalesced. */ - extent_t *extent; + edata_t *edata; while (true) { /* Get the LRU extent, if any. */ - extent = extent_list_first(&extents->lru); - if (extent == NULL) { - goto label_return; + eset_t *eset = &ecache->eset; + edata = edata_list_inactive_first(&eset->lru); + if (edata == NULL) { + /* + * Next check if there are guarded extents. They are + * more expensive to purge (since they are not + * mergeable), thus in favor of caching them longer. + */ + eset = &ecache->guarded_eset; + edata = edata_list_inactive_first(&eset->lru); + if (edata == NULL) { + goto label_return; + } } /* Check the eviction limit. */ - size_t extents_npages = atomic_load_zu(&extents->npages, - ATOMIC_RELAXED); + size_t extents_npages = ecache_npages_get(ecache); if (extents_npages <= npages_min) { - extent = NULL; + edata = NULL; goto label_return; } - extents_remove_locked(tsdn, extents, extent); - if (!extents->delay_coalesce) { + eset_remove(eset, edata); + if (!ecache->delay_coalesce || edata_guarded_get(edata)) { break; } /* Try to coalesce. */ - if (extent_try_delayed_coalesce(tsdn, arena, r_extent_hooks, - rtree_ctx, extents, extent)) { + if (extent_try_delayed_coalesce(tsdn, pac, ehooks, ecache, + edata)) { break; } /* @@ -608,23 +198,24 @@ * Either mark the extent active or deregister it to protect against * concurrent operations. */ - switch (extents_state_get(extents)) { + switch (ecache->state) { case extent_state_active: not_reached(); case extent_state_dirty: case extent_state_muzzy: - extent_state_set(extent, extent_state_active); + emap_update_edata_state(tsdn, pac->emap, edata, + extent_state_active); break; case extent_state_retained: - extent_deregister(tsdn, extent); + extent_deregister(tsdn, pac, edata); break; default: not_reached(); } label_return: - malloc_mutex_unlock(tsdn, &extents->mtx); - return extent; + malloc_mutex_unlock(tsdn, &ecache->mtx); + return edata; } /* @@ -632,123 +223,73 @@ * indicates OOM), e.g. when trying to split an existing extent. */ static void -extents_abandon_vm(tsdn_t *tsdn, arena_t *arena, extent_hooks_t **r_extent_hooks, - extents_t *extents, extent_t *extent, bool growing_retained) { - size_t sz = extent_size_get(extent); +extents_abandon_vm(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, ecache_t *ecache, + edata_t *edata, bool growing_retained) { + size_t sz = edata_size_get(edata); if (config_stats) { - arena_stats_accum_zu(&arena->stats.abandoned_vm, sz); + atomic_fetch_add_zu(&pac->stats->abandoned_vm, sz, + ATOMIC_RELAXED); } /* * Leak extent after making sure its pages have already been purged, so * that this is only a virtual memory leak. */ - if (extents_state_get(extents) == extent_state_dirty) { - if (extent_purge_lazy_impl(tsdn, arena, r_extent_hooks, - extent, 0, sz, growing_retained)) { - extent_purge_forced_impl(tsdn, arena, r_extent_hooks, - extent, 0, extent_size_get(extent), - growing_retained); + if (ecache->state == extent_state_dirty) { + if (extent_purge_lazy_impl(tsdn, ehooks, edata, 0, sz, + growing_retained)) { + extent_purge_forced_impl(tsdn, ehooks, edata, 0, + edata_size_get(edata), growing_retained); } } - extent_dalloc(tsdn, arena, extent); -} - -void -extents_prefork(tsdn_t *tsdn, extents_t *extents) { - malloc_mutex_prefork(tsdn, &extents->mtx); -} - -void -extents_postfork_parent(tsdn_t *tsdn, extents_t *extents) { - malloc_mutex_postfork_parent(tsdn, &extents->mtx); -} - -void -extents_postfork_child(tsdn_t *tsdn, extents_t *extents) { - malloc_mutex_postfork_child(tsdn, &extents->mtx); + edata_cache_put(tsdn, pac->edata_cache, edata); } static void -extent_deactivate_locked(tsdn_t *tsdn, arena_t *arena, extents_t *extents, - extent_t *extent) { - assert(extent_arena_get(extent) == arena); - assert(extent_state_get(extent) == extent_state_active); +extent_deactivate_locked_impl(tsdn_t *tsdn, pac_t *pac, ecache_t *ecache, + edata_t *edata) { + malloc_mutex_assert_owner(tsdn, &ecache->mtx); + assert(edata_arena_ind_get(edata) == ecache_ind_get(ecache)); - extent_state_set(extent, extents_state_get(extents)); - extents_insert_locked(tsdn, extents, extent); + emap_update_edata_state(tsdn, pac->emap, edata, ecache->state); + eset_t *eset = edata_guarded_get(edata) ? &ecache->guarded_eset : + &ecache->eset; + eset_insert(eset, edata); } static void -extent_deactivate(tsdn_t *tsdn, arena_t *arena, extents_t *extents, - extent_t *extent) { - malloc_mutex_lock(tsdn, &extents->mtx); - extent_deactivate_locked(tsdn, arena, extents, extent); - malloc_mutex_unlock(tsdn, &extents->mtx); +extent_deactivate_locked(tsdn_t *tsdn, pac_t *pac, ecache_t *ecache, + edata_t *edata) { + assert(edata_state_get(edata) == extent_state_active); + extent_deactivate_locked_impl(tsdn, pac, ecache, edata); } static void -extent_activate_locked(tsdn_t *tsdn, arena_t *arena, extents_t *extents, - extent_t *extent) { - assert(extent_arena_get(extent) == arena); - assert(extent_state_get(extent) == extents_state_get(extents)); - - extents_remove_locked(tsdn, extents, extent); - extent_state_set(extent, extent_state_active); -} - -static bool -extent_rtree_leaf_elms_lookup(tsdn_t *tsdn, rtree_ctx_t *rtree_ctx, - const extent_t *extent, bool dependent, bool init_missing, - rtree_leaf_elm_t **r_elm_a, rtree_leaf_elm_t **r_elm_b) { - *r_elm_a = rtree_leaf_elm_lookup(tsdn, &extents_rtree, rtree_ctx, - (uintptr_t)extent_base_get(extent), dependent, init_missing); - if (!dependent && *r_elm_a == NULL) { - return true; - } - assert(*r_elm_a != NULL); - - *r_elm_b = rtree_leaf_elm_lookup(tsdn, &extents_rtree, rtree_ctx, - (uintptr_t)extent_last_get(extent), dependent, init_missing); - if (!dependent && *r_elm_b == NULL) { - return true; - } - assert(*r_elm_b != NULL); - - return false; +extent_deactivate_check_state_locked(tsdn_t *tsdn, pac_t *pac, ecache_t *ecache, + edata_t *edata, extent_state_t expected_state) { + assert(edata_state_get(edata) == expected_state); + extent_deactivate_locked_impl(tsdn, pac, ecache, edata); } static void -extent_rtree_write_acquired(tsdn_t *tsdn, rtree_leaf_elm_t *elm_a, - rtree_leaf_elm_t *elm_b, extent_t *extent, szind_t szind, bool slab) { - rtree_leaf_elm_write(tsdn, &extents_rtree, elm_a, extent, szind, slab); - if (elm_b != NULL) { - rtree_leaf_elm_write(tsdn, &extents_rtree, elm_b, extent, szind, - slab); - } -} +extent_activate_locked(tsdn_t *tsdn, pac_t *pac, ecache_t *ecache, eset_t *eset, + edata_t *edata) { + assert(edata_arena_ind_get(edata) == ecache_ind_get(ecache)); + assert(edata_state_get(edata) == ecache->state || + edata_state_get(edata) == extent_state_merging); -static void -extent_interior_register(tsdn_t *tsdn, rtree_ctx_t *rtree_ctx, extent_t *extent, - szind_t szind) { - assert(extent_slab_get(extent)); - - /* Register interior. */ - for (size_t i = 1; i < (extent_size_get(extent) >> LG_PAGE) - 1; i++) { - rtree_write(tsdn, &extents_rtree, rtree_ctx, - (uintptr_t)extent_base_get(extent) + (uintptr_t)(i << - LG_PAGE), extent, szind, true); - } + eset_remove(eset, edata); + emap_update_edata_state(tsdn, pac->emap, edata, extent_state_active); } -static void -extent_gdump_add(tsdn_t *tsdn, const extent_t *extent) { +void +extent_gdump_add(tsdn_t *tsdn, const edata_t *edata) { cassert(config_prof); /* prof_gdump() requirement. */ witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), WITNESS_RANK_CORE, 0); - if (opt_prof && extent_state_get(extent) == extent_state_active) { - size_t nadd = extent_size_get(extent) >> LG_PAGE; + if (opt_prof && edata_state_get(edata) == extent_state_active) { + size_t nadd = edata_size_get(edata) >> LG_PAGE; size_t cur = atomic_fetch_add_zu(&curpages, nadd, ATOMIC_RELAXED) + nadd; size_t high = atomic_load_zu(&highpages, ATOMIC_RELAXED); @@ -767,232 +308,184 @@ } static void -extent_gdump_sub(tsdn_t *tsdn, const extent_t *extent) { +extent_gdump_sub(tsdn_t *tsdn, const edata_t *edata) { cassert(config_prof); - if (opt_prof && extent_state_get(extent) == extent_state_active) { - size_t nsub = extent_size_get(extent) >> LG_PAGE; + if (opt_prof && edata_state_get(edata) == extent_state_active) { + size_t nsub = edata_size_get(edata) >> LG_PAGE; assert(atomic_load_zu(&curpages, ATOMIC_RELAXED) >= nsub); atomic_fetch_sub_zu(&curpages, nsub, ATOMIC_RELAXED); } } static bool -extent_register_impl(tsdn_t *tsdn, extent_t *extent, bool gdump_add) { - rtree_ctx_t rtree_ctx_fallback; - rtree_ctx_t *rtree_ctx = tsdn_rtree_ctx(tsdn, &rtree_ctx_fallback); - rtree_leaf_elm_t *elm_a, *elm_b; - +extent_register_impl(tsdn_t *tsdn, pac_t *pac, edata_t *edata, bool gdump_add) { + assert(edata_state_get(edata) == extent_state_active); /* - * We need to hold the lock to protect against a concurrent coalesce - * operation that sees us in a partial state. + * No locking needed, as the edata must be in active state, which + * prevents other threads from accessing the edata. */ - extent_lock(tsdn, extent); - - if (extent_rtree_leaf_elms_lookup(tsdn, rtree_ctx, extent, false, true, - &elm_a, &elm_b)) { - extent_unlock(tsdn, extent); + if (emap_register_boundary(tsdn, pac->emap, edata, SC_NSIZES, + /* slab */ false)) { return true; } - szind_t szind = extent_szind_get_maybe_invalid(extent); - bool slab = extent_slab_get(extent); - extent_rtree_write_acquired(tsdn, elm_a, elm_b, extent, szind, slab); - if (slab) { - extent_interior_register(tsdn, rtree_ctx, extent, szind); - } - - extent_unlock(tsdn, extent); - if (config_prof && gdump_add) { - extent_gdump_add(tsdn, extent); + extent_gdump_add(tsdn, edata); } return false; } static bool -extent_register(tsdn_t *tsdn, extent_t *extent) { - return extent_register_impl(tsdn, extent, true); +extent_register(tsdn_t *tsdn, pac_t *pac, edata_t *edata) { + return extent_register_impl(tsdn, pac, edata, true); } static bool -extent_register_no_gdump_add(tsdn_t *tsdn, extent_t *extent) { - return extent_register_impl(tsdn, extent, false); +extent_register_no_gdump_add(tsdn_t *tsdn, pac_t *pac, edata_t *edata) { + return extent_register_impl(tsdn, pac, edata, false); } static void -extent_reregister(tsdn_t *tsdn, extent_t *extent) { - bool err = extent_register(tsdn, extent); +extent_reregister(tsdn_t *tsdn, pac_t *pac, edata_t *edata) { + bool err = extent_register(tsdn, pac, edata); assert(!err); } -/* - * Removes all pointers to the given extent from the global rtree indices for - * its interior. This is relevant for slab extents, for which we need to do - * metadata lookups at places other than the head of the extent. We deregister - * on the interior, then, when an extent moves from being an active slab to an - * inactive state. - */ -static void -extent_interior_deregister(tsdn_t *tsdn, rtree_ctx_t *rtree_ctx, - extent_t *extent) { - size_t i; - - assert(extent_slab_get(extent)); - - for (i = 1; i < (extent_size_get(extent) >> LG_PAGE) - 1; i++) { - rtree_clear(tsdn, &extents_rtree, rtree_ctx, - (uintptr_t)extent_base_get(extent) + (uintptr_t)(i << - LG_PAGE)); - } -} - /* * Removes all pointers to the given extent from the global rtree. */ static void -extent_deregister_impl(tsdn_t *tsdn, extent_t *extent, bool gdump) { - rtree_ctx_t rtree_ctx_fallback; - rtree_ctx_t *rtree_ctx = tsdn_rtree_ctx(tsdn, &rtree_ctx_fallback); - rtree_leaf_elm_t *elm_a, *elm_b; - extent_rtree_leaf_elms_lookup(tsdn, rtree_ctx, extent, true, false, - &elm_a, &elm_b); - - extent_lock(tsdn, extent); - - extent_rtree_write_acquired(tsdn, elm_a, elm_b, NULL, SC_NSIZES, false); - if (extent_slab_get(extent)) { - extent_interior_deregister(tsdn, rtree_ctx, extent); - extent_slab_set(extent, false); - } - - extent_unlock(tsdn, extent); +extent_deregister_impl(tsdn_t *tsdn, pac_t *pac, edata_t *edata, + bool gdump) { + emap_deregister_boundary(tsdn, pac->emap, edata); if (config_prof && gdump) { - extent_gdump_sub(tsdn, extent); + extent_gdump_sub(tsdn, edata); } } static void -extent_deregister(tsdn_t *tsdn, extent_t *extent) { - extent_deregister_impl(tsdn, extent, true); +extent_deregister(tsdn_t *tsdn, pac_t *pac, edata_t *edata) { + extent_deregister_impl(tsdn, pac, edata, true); } static void -extent_deregister_no_gdump_sub(tsdn_t *tsdn, extent_t *extent) { - extent_deregister_impl(tsdn, extent, false); +extent_deregister_no_gdump_sub(tsdn_t *tsdn, pac_t *pac, + edata_t *edata) { + extent_deregister_impl(tsdn, pac, edata, false); } /* - * Tries to find and remove an extent from extents that can be used for the + * Tries to find and remove an extent from ecache that can be used for the * given allocation request. */ -static extent_t * -extent_recycle_extract(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, rtree_ctx_t *rtree_ctx, extents_t *extents, - void *new_addr, size_t size, size_t pad, size_t alignment, bool slab, - bool growing_retained) { - witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), - WITNESS_RANK_CORE, growing_retained ? 1 : 0); +static edata_t * +extent_recycle_extract(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, + ecache_t *ecache, edata_t *expand_edata, size_t size, size_t alignment, + bool guarded) { + malloc_mutex_assert_owner(tsdn, &ecache->mtx); assert(alignment > 0); - if (config_debug && new_addr != NULL) { + if (config_debug && expand_edata != NULL) { /* - * Non-NULL new_addr has two use cases: - * - * 1) Recycle a known-extant extent, e.g. during purging. - * 2) Perform in-place expanding reallocation. - * - * Regardless of use case, new_addr must either refer to a - * non-existing extent, or to the base of an extant extent, - * since only active slabs support interior lookups (which of - * course cannot be recycled). + * Non-NULL expand_edata indicates in-place expanding realloc. + * new_addr must either refer to a non-existing extent, or to + * the base of an extant extent, since only active slabs support + * interior lookups (which of course cannot be recycled). */ + void *new_addr = edata_past_get(expand_edata); assert(PAGE_ADDR2BASE(new_addr) == new_addr); - assert(pad == 0); assert(alignment <= PAGE); } - size_t esize = size + pad; - malloc_mutex_lock(tsdn, &extents->mtx); - extent_hooks_assure_initialized(arena, r_extent_hooks); - extent_t *extent; - if (new_addr != NULL) { - extent = extent_lock_from_addr(tsdn, rtree_ctx, new_addr, - false); - if (extent != NULL) { - /* - * We might null-out extent to report an error, but we - * still need to unlock the associated mutex after. - */ - extent_t *unlock_extent = extent; - assert(extent_base_get(extent) == new_addr); - if (extent_arena_get(extent) != arena || - extent_size_get(extent) < esize || - extent_state_get(extent) != - extents_state_get(extents)) { - extent = NULL; + edata_t *edata; + eset_t *eset = guarded ? &ecache->guarded_eset : &ecache->eset; + if (expand_edata != NULL) { + edata = emap_try_acquire_edata_neighbor_expand(tsdn, pac->emap, + expand_edata, EXTENT_PAI_PAC, ecache->state); + if (edata != NULL) { + extent_assert_can_expand(expand_edata, edata); + if (edata_size_get(edata) < size) { + emap_release_edata(tsdn, pac->emap, edata, + ecache->state); + edata = NULL; } - extent_unlock(tsdn, unlock_extent); } } else { - extent = extents_fit_locked(tsdn, arena, extents, esize, - alignment); + /* + * A large extent might be broken up from its original size to + * some small size to satisfy a small request. When that small + * request is freed, though, it won't merge back with the larger + * extent if delayed coalescing is on. The large extent can + * then no longer satify a request for its original size. To + * limit this effect, when delayed coalescing is enabled, we + * put a cap on how big an extent we can split for a request. + */ + unsigned lg_max_fit = ecache->delay_coalesce + ? (unsigned)opt_lg_extent_max_active_fit : SC_PTR_BITS; + + /* + * If split and merge are not allowed (Windows w/o retain), try + * exact fit only. + * + * For simplicity purposes, splitting guarded extents is not + * supported. Hence, we do only exact fit for guarded + * allocations. + */ + bool exact_only = (!maps_coalesce && !opt_retain) || guarded; + edata = eset_fit(eset, size, alignment, exact_only, + lg_max_fit); } - if (extent == NULL) { - malloc_mutex_unlock(tsdn, &extents->mtx); + if (edata == NULL) { return NULL; } + assert(!guarded || edata_guarded_get(edata)); + extent_activate_locked(tsdn, pac, ecache, eset, edata); - extent_activate_locked(tsdn, arena, extents, extent); - malloc_mutex_unlock(tsdn, &extents->mtx); - - return extent; + return edata; } /* * Given an allocation request and an extent guaranteed to be able to satisfy - * it, this splits off lead and trail extents, leaving extent pointing to an + * it, this splits off lead and trail extents, leaving edata pointing to an * extent satisfying the allocation. - * This function doesn't put lead or trail into any extents_t; it's the caller's + * This function doesn't put lead or trail into any ecache; it's the caller's * job to ensure that they can be reused. */ typedef enum { /* - * Split successfully. lead, extent, and trail, are modified to extents + * Split successfully. lead, edata, and trail, are modified to extents * describing the ranges before, in, and after the given allocation. */ extent_split_interior_ok, /* * The extent can't satisfy the given allocation request. None of the - * input extent_t *s are touched. + * input edata_t *s are touched. */ extent_split_interior_cant_alloc, /* * In a potentially invalid state. Must leak (if *to_leak is non-NULL), * and salvage what's still salvageable (if *to_salvage is non-NULL). - * None of lead, extent, or trail are valid. + * None of lead, edata, or trail are valid. */ extent_split_interior_error } extent_split_interior_result_t; static extent_split_interior_result_t -extent_split_interior(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, rtree_ctx_t *rtree_ctx, +extent_split_interior(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, /* The result of splitting, in case of success. */ - extent_t **extent, extent_t **lead, extent_t **trail, + edata_t **edata, edata_t **lead, edata_t **trail, /* The mess to clean up, in case of error. */ - extent_t **to_leak, extent_t **to_salvage, - void *new_addr, size_t size, size_t pad, size_t alignment, bool slab, - szind_t szind, bool growing_retained) { - size_t esize = size + pad; - size_t leadsize = ALIGNMENT_CEILING((uintptr_t)extent_base_get(*extent), - PAGE_CEILING(alignment)) - (uintptr_t)extent_base_get(*extent); - assert(new_addr == NULL || leadsize == 0); - if (extent_size_get(*extent) < leadsize + esize) { + edata_t **to_leak, edata_t **to_salvage, + edata_t *expand_edata, size_t size, size_t alignment) { + size_t leadsize = ALIGNMENT_CEILING((uintptr_t)edata_base_get(*edata), + PAGE_CEILING(alignment)) - (uintptr_t)edata_base_get(*edata); + assert(expand_edata == NULL || leadsize == 0); + if (edata_size_get(*edata) < leadsize + size) { return extent_split_interior_cant_alloc; } - size_t trailsize = extent_size_get(*extent) - leadsize - esize; + size_t trailsize = edata_size_get(*edata) - leadsize - size; *lead = NULL; *trail = NULL; @@ -1001,11 +494,11 @@ /* Split the lead. */ if (leadsize != 0) { - *lead = *extent; - *extent = extent_split_impl(tsdn, arena, r_extent_hooks, - *lead, leadsize, SC_NSIZES, false, esize + trailsize, szind, - slab, growing_retained); - if (*extent == NULL) { + assert(!edata_guarded_get(*edata)); + *lead = *edata; + *edata = extent_split_impl(tsdn, pac, ehooks, *lead, leadsize, + size + trailsize, /* holding_core_locks*/ true); + if (*edata == NULL) { *to_leak = *lead; *lead = NULL; return extent_split_interior_error; @@ -1014,36 +507,18 @@ /* Split the trail. */ if (trailsize != 0) { - *trail = extent_split_impl(tsdn, arena, r_extent_hooks, *extent, - esize, szind, slab, trailsize, SC_NSIZES, false, - growing_retained); + assert(!edata_guarded_get(*edata)); + *trail = extent_split_impl(tsdn, pac, ehooks, *edata, size, + trailsize, /* holding_core_locks */ true); if (*trail == NULL) { - *to_leak = *extent; + *to_leak = *edata; *to_salvage = *lead; *lead = NULL; - *extent = NULL; + *edata = NULL; return extent_split_interior_error; } } - if (leadsize == 0 && trailsize == 0) { - /* - * Splitting causes szind to be set as a side effect, but no - * splitting occurred. - */ - extent_szind_set(*extent, szind); - if (szind != SC_NSIZES) { - rtree_szind_slab_update(tsdn, &extents_rtree, rtree_ctx, - (uintptr_t)extent_addr_get(*extent), szind, slab); - if (slab && extent_size_get(*extent) > PAGE) { - rtree_szind_slab_update(tsdn, &extents_rtree, - rtree_ctx, - (uintptr_t)extent_past_get(*extent) - - (uintptr_t)PAGE, szind, slab); - } - } - } - return extent_split_interior_ok; } @@ -1051,42 +526,43 @@ * This fulfills the indicated allocation request out of the given extent (which * the caller should have ensured was big enough). If there's any unused space * before or after the resulting allocation, that space is given its own extent - * and put back into extents. + * and put back into ecache. */ -static extent_t * -extent_recycle_split(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, rtree_ctx_t *rtree_ctx, extents_t *extents, - void *new_addr, size_t size, size_t pad, size_t alignment, bool slab, - szind_t szind, extent_t *extent, bool growing_retained) { - extent_t *lead; - extent_t *trail; - extent_t *to_leak; - extent_t *to_salvage; +static edata_t * +extent_recycle_split(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, + ecache_t *ecache, edata_t *expand_edata, size_t size, size_t alignment, + edata_t *edata, bool growing_retained) { + assert(!edata_guarded_get(edata) || size == edata_size_get(edata)); + malloc_mutex_assert_owner(tsdn, &ecache->mtx); + + edata_t *lead; + edata_t *trail; + edata_t *to_leak JEMALLOC_CC_SILENCE_INIT(NULL); + edata_t *to_salvage JEMALLOC_CC_SILENCE_INIT(NULL); extent_split_interior_result_t result = extent_split_interior( - tsdn, arena, r_extent_hooks, rtree_ctx, &extent, &lead, &trail, - &to_leak, &to_salvage, new_addr, size, pad, alignment, slab, szind, - growing_retained); + tsdn, pac, ehooks, &edata, &lead, &trail, &to_leak, &to_salvage, + expand_edata, size, alignment); if (!maps_coalesce && result != extent_split_interior_ok && !opt_retain) { /* * Split isn't supported (implies Windows w/o retain). Avoid - * leaking the extents. + * leaking the extent. */ assert(to_leak != NULL && lead == NULL && trail == NULL); - extent_deactivate(tsdn, arena, extents, to_leak); + extent_deactivate_locked(tsdn, pac, ecache, to_leak); return NULL; } if (result == extent_split_interior_ok) { if (lead != NULL) { - extent_deactivate(tsdn, arena, extents, lead); + extent_deactivate_locked(tsdn, pac, ecache, lead); } if (trail != NULL) { - extent_deactivate(tsdn, arena, extents, trail); + extent_deactivate_locked(tsdn, pac, ecache, trail); } - return extent; + return edata; } else { /* * We should have picked an extent that was large enough to @@ -1094,294 +570,144 @@ */ assert(result == extent_split_interior_error); if (to_salvage != NULL) { - extent_deregister(tsdn, to_salvage); + extent_deregister(tsdn, pac, to_salvage); } if (to_leak != NULL) { - void *leak = extent_base_get(to_leak); - extent_deregister_no_gdump_sub(tsdn, to_leak); - extents_abandon_vm(tsdn, arena, r_extent_hooks, extents, - to_leak, growing_retained); - assert(extent_lock_from_addr(tsdn, rtree_ctx, leak, - false) == NULL); + extent_deregister_no_gdump_sub(tsdn, pac, to_leak); + /* + * May go down the purge path (which assume no ecache + * locks). Only happens with OOM caused split failures. + */ + malloc_mutex_unlock(tsdn, &ecache->mtx); + extents_abandon_vm(tsdn, pac, ehooks, ecache, to_leak, + growing_retained); + malloc_mutex_lock(tsdn, &ecache->mtx); } return NULL; } unreachable(); } -static bool -extent_need_manual_zero(arena_t *arena) { - /* - * Need to manually zero the extent on repopulating if either; 1) non - * default extent hooks installed (in which case the purge semantics may - * change); or 2) transparent huge pages enabled. - */ - return (!arena_has_default_hooks(arena) || - (opt_thp == thp_mode_always)); -} - /* * Tries to satisfy the given allocation request by reusing one of the extents - * in the given extents_t. + * in the given ecache_t. */ -static extent_t * -extent_recycle(tsdn_t *tsdn, arena_t *arena, extent_hooks_t **r_extent_hooks, - extents_t *extents, void *new_addr, size_t size, size_t pad, - size_t alignment, bool slab, szind_t szind, bool *zero, bool *commit, - bool growing_retained) { +static edata_t * +extent_recycle(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, ecache_t *ecache, + edata_t *expand_edata, size_t size, size_t alignment, bool zero, + bool *commit, bool growing_retained, bool guarded) { witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), WITNESS_RANK_CORE, growing_retained ? 1 : 0); - assert(new_addr == NULL || !slab); - assert(pad == 0 || !slab); - assert(!*zero || !slab); + assert(!guarded || expand_edata == NULL); + assert(!guarded || alignment <= PAGE); - rtree_ctx_t rtree_ctx_fallback; - rtree_ctx_t *rtree_ctx = tsdn_rtree_ctx(tsdn, &rtree_ctx_fallback); + malloc_mutex_lock(tsdn, &ecache->mtx); - extent_t *extent = extent_recycle_extract(tsdn, arena, r_extent_hooks, - rtree_ctx, extents, new_addr, size, pad, alignment, slab, - growing_retained); - if (extent == NULL) { + edata_t *edata = extent_recycle_extract(tsdn, pac, ehooks, ecache, + expand_edata, size, alignment, guarded); + if (edata == NULL) { + malloc_mutex_unlock(tsdn, &ecache->mtx); return NULL; } - extent = extent_recycle_split(tsdn, arena, r_extent_hooks, rtree_ctx, - extents, new_addr, size, pad, alignment, slab, szind, extent, - growing_retained); - if (extent == NULL) { + edata = extent_recycle_split(tsdn, pac, ehooks, ecache, expand_edata, + size, alignment, edata, growing_retained); + malloc_mutex_unlock(tsdn, &ecache->mtx); + if (edata == NULL) { return NULL; } - if (*commit && !extent_committed_get(extent)) { - if (extent_commit_impl(tsdn, arena, r_extent_hooks, extent, - 0, extent_size_get(extent), growing_retained)) { - extent_record(tsdn, arena, r_extent_hooks, extents, - extent, growing_retained); - return NULL; - } - if (!extent_need_manual_zero(arena)) { - extent_zeroed_set(extent, true); - } + assert(edata_state_get(edata) == extent_state_active); + if (extent_commit_zero(tsdn, ehooks, edata, *commit, zero, + growing_retained)) { + extent_record(tsdn, pac, ehooks, ecache, edata); + return NULL; } - - if (extent_committed_get(extent)) { + if (edata_committed_get(edata)) { + /* + * This reverses the purpose of this variable - previously it + * was treated as an input parameter, now it turns into an + * output parameter, reporting if the edata has actually been + * committed. + */ *commit = true; } - if (extent_zeroed_get(extent)) { - *zero = true; - } - - if (pad != 0) { - extent_addr_randomize(tsdn, extent, alignment); - } - assert(extent_state_get(extent) == extent_state_active); - if (slab) { - extent_slab_set(extent, slab); - extent_interior_register(tsdn, rtree_ctx, extent, szind); - } - - if (*zero) { - void *addr = extent_base_get(extent); - if (!extent_zeroed_get(extent)) { - size_t size = extent_size_get(extent); - if (extent_need_manual_zero(arena) || - pages_purge_forced(addr, size)) { - memset(addr, 0, size); - } - } else if (config_debug) { - size_t *p = (size_t *)(uintptr_t)addr; - /* Check the first page only. */ - for (size_t i = 0; i < PAGE / sizeof(size_t); i++) { - assert(p[i] == 0); - } - } - } - return extent; + return edata; } /* - * If the caller specifies (!*zero), it is still possible to receive zeroed - * memory, in which case *zero is toggled to true. arena_extent_alloc() takes - * advantage of this to avoid demanding zeroed extents, but taking advantage of - * them if they are returned. + * If virtual memory is retained, create increasingly larger extents from which + * to split requested extents in order to limit the total number of disjoint + * virtual memory ranges retained by each shard. */ -static void * -extent_alloc_core(tsdn_t *tsdn, arena_t *arena, void *new_addr, size_t size, - size_t alignment, bool *zero, bool *commit, dss_prec_t dss_prec) { - void *ret; +static edata_t * +extent_grow_retained(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, + size_t size, size_t alignment, bool zero, bool *commit) { + malloc_mutex_assert_owner(tsdn, &pac->grow_mtx); - assert(size != 0); - assert(alignment != 0); - - /* "primary" dss. */ - if (have_dss && dss_prec == dss_prec_primary && (ret = - extent_alloc_dss(tsdn, arena, new_addr, size, alignment, zero, - commit)) != NULL) { - return ret; - } - /* mmap. */ - if ((ret = extent_alloc_mmap(new_addr, size, alignment, zero, commit)) - != NULL) { - return ret; - } - /* "secondary" dss. */ - if (have_dss && dss_prec == dss_prec_secondary && (ret = - extent_alloc_dss(tsdn, arena, new_addr, size, alignment, zero, - commit)) != NULL) { - return ret; - } - - /* All strategies for allocation failed. */ - return NULL; -} - -static void * -extent_alloc_default_impl(tsdn_t *tsdn, arena_t *arena, void *new_addr, - size_t size, size_t alignment, bool *zero, bool *commit) { - void *ret = extent_alloc_core(tsdn, arena, new_addr, size, alignment, zero, - commit, (dss_prec_t)atomic_load_u(&arena->dss_prec, - ATOMIC_RELAXED)); - if (have_madvise_huge && ret) { - pages_set_thp_state(ret, size); - } - return ret; -} - -static void * -extent_alloc_default(extent_hooks_t *extent_hooks, void *new_addr, size_t size, - size_t alignment, bool *zero, bool *commit, unsigned arena_ind) { - tsdn_t *tsdn; - arena_t *arena; - - tsdn = tsdn_fetch(); - arena = arena_get(tsdn, arena_ind, false); - /* - * The arena we're allocating on behalf of must have been initialized - * already. - */ - assert(arena != NULL); - - return extent_alloc_default_impl(tsdn, arena, new_addr, size, - ALIGNMENT_CEILING(alignment, PAGE), zero, commit); -} - -static void -extent_hook_pre_reentrancy(tsdn_t *tsdn, arena_t *arena) { - tsd_t *tsd = tsdn_null(tsdn) ? tsd_fetch() : tsdn_tsd(tsdn); - if (arena == arena_get(tsd_tsdn(tsd), 0, false)) { - /* - * The only legitimate case of customized extent hooks for a0 is - * hooks with no allocation activities. One such example is to - * place metadata on pre-allocated resources such as huge pages. - * In that case, rely on reentrancy_level checks to catch - * infinite recursions. - */ - pre_reentrancy(tsd, NULL); - } else { - pre_reentrancy(tsd, arena); - } -} - -static void -extent_hook_post_reentrancy(tsdn_t *tsdn) { - tsd_t *tsd = tsdn_null(tsdn) ? tsd_fetch() : tsdn_tsd(tsdn); - post_reentrancy(tsd); -} - -/* - * If virtual memory is retained, create increasingly larger extents from which - * to split requested extents in order to limit the total number of disjoint - * virtual memory ranges retained by each arena. - */ -static extent_t * -extent_grow_retained(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, size_t size, size_t pad, size_t alignment, - bool slab, szind_t szind, bool *zero, bool *commit) { - malloc_mutex_assert_owner(tsdn, &arena->extent_grow_mtx); - assert(pad == 0 || !slab); - assert(!*zero || !slab); - - size_t esize = size + pad; - size_t alloc_size_min = esize + PAGE_CEILING(alignment) - PAGE; + size_t alloc_size_min = size + PAGE_CEILING(alignment) - PAGE; /* Beware size_t wrap-around. */ - if (alloc_size_min < esize) { + if (alloc_size_min < size) { goto label_err; } /* * Find the next extent size in the series that would be large enough to * satisfy this request. */ - pszind_t egn_skip = 0; - size_t alloc_size = sz_pind2sz(arena->extent_grow_next + egn_skip); - while (alloc_size < alloc_size_min) { - egn_skip++; - if (arena->extent_grow_next + egn_skip >= - sz_psz2ind(SC_LARGE_MAXCLASS)) { - /* Outside legal range. */ - goto label_err; - } - alloc_size = sz_pind2sz(arena->extent_grow_next + egn_skip); + size_t alloc_size; + pszind_t exp_grow_skip; + bool err = exp_grow_size_prepare(&pac->exp_grow, alloc_size_min, + &alloc_size, &exp_grow_skip); + if (err) { + goto label_err; } - extent_t *extent = extent_alloc(tsdn, arena); - if (extent == NULL) { + edata_t *edata = edata_cache_get(tsdn, pac->edata_cache); + if (edata == NULL) { goto label_err; } bool zeroed = false; bool committed = false; - void *ptr; - if (*r_extent_hooks == &extent_hooks_default) { - ptr = extent_alloc_default_impl(tsdn, arena, NULL, - alloc_size, PAGE, &zeroed, &committed); - } else { - extent_hook_pre_reentrancy(tsdn, arena); - ptr = (*r_extent_hooks)->alloc(*r_extent_hooks, NULL, - alloc_size, PAGE, &zeroed, &committed, - arena_ind_get(arena)); - extent_hook_post_reentrancy(tsdn); - } + void *ptr = ehooks_alloc(tsdn, ehooks, NULL, alloc_size, PAGE, &zeroed, + &committed); - extent_init(extent, arena, ptr, alloc_size, false, SC_NSIZES, - arena_extent_sn_next(arena), extent_state_active, zeroed, - committed, true, EXTENT_IS_HEAD); if (ptr == NULL) { - extent_dalloc(tsdn, arena, extent); + edata_cache_put(tsdn, pac->edata_cache, edata); goto label_err; } - if (extent_register_no_gdump_add(tsdn, extent)) { - extent_dalloc(tsdn, arena, extent); + edata_init(edata, ecache_ind_get(&pac->ecache_retained), ptr, + alloc_size, false, SC_NSIZES, extent_sn_next(pac), + extent_state_active, zeroed, committed, EXTENT_PAI_PAC, + EXTENT_IS_HEAD); + + if (extent_register_no_gdump_add(tsdn, pac, edata)) { + edata_cache_put(tsdn, pac->edata_cache, edata); goto label_err; } - if (extent_zeroed_get(extent) && extent_committed_get(extent)) { - *zero = true; - } - if (extent_committed_get(extent)) { + if (edata_committed_get(edata)) { *commit = true; } - rtree_ctx_t rtree_ctx_fallback; - rtree_ctx_t *rtree_ctx = tsdn_rtree_ctx(tsdn, &rtree_ctx_fallback); + edata_t *lead; + edata_t *trail; + edata_t *to_leak JEMALLOC_CC_SILENCE_INIT(NULL); + edata_t *to_salvage JEMALLOC_CC_SILENCE_INIT(NULL); - extent_t *lead; - extent_t *trail; - extent_t *to_leak; - extent_t *to_salvage; - extent_split_interior_result_t result = extent_split_interior( - tsdn, arena, r_extent_hooks, rtree_ctx, &extent, &lead, &trail, - &to_leak, &to_salvage, NULL, size, pad, alignment, slab, szind, - true); + extent_split_interior_result_t result = extent_split_interior(tsdn, + pac, ehooks, &edata, &lead, &trail, &to_leak, &to_salvage, NULL, + size, alignment); if (result == extent_split_interior_ok) { if (lead != NULL) { - extent_record(tsdn, arena, r_extent_hooks, - &arena->extents_retained, lead, true); + extent_record(tsdn, pac, ehooks, &pac->ecache_retained, + lead); } if (trail != NULL) { - extent_record(tsdn, arena, r_extent_hooks, - &arena->extents_retained, trail, true); + extent_record(tsdn, pac, ehooks, &pac->ecache_retained, + trail); } } else { /* @@ -1393,26 +719,32 @@ if (config_prof) { extent_gdump_add(tsdn, to_salvage); } - extent_record(tsdn, arena, r_extent_hooks, - &arena->extents_retained, to_salvage, true); + extent_record(tsdn, pac, ehooks, &pac->ecache_retained, + to_salvage); } if (to_leak != NULL) { - extent_deregister_no_gdump_sub(tsdn, to_leak); - extents_abandon_vm(tsdn, arena, r_extent_hooks, - &arena->extents_retained, to_leak, true); + extent_deregister_no_gdump_sub(tsdn, pac, to_leak); + extents_abandon_vm(tsdn, pac, ehooks, + &pac->ecache_retained, to_leak, true); } goto label_err; } - if (*commit && !extent_committed_get(extent)) { - if (extent_commit_impl(tsdn, arena, r_extent_hooks, extent, 0, - extent_size_get(extent), true)) { - extent_record(tsdn, arena, r_extent_hooks, - &arena->extents_retained, extent, true); + if (*commit && !edata_committed_get(edata)) { + if (extent_commit_impl(tsdn, ehooks, edata, 0, + edata_size_get(edata), true)) { + extent_record(tsdn, pac, ehooks, + &pac->ecache_retained, edata); goto label_err; } - if (!extent_need_manual_zero(arena)) { - extent_zeroed_set(extent, true); + /* A successful commit should return zeroed memory. */ + if (config_debug) { + void *addr = edata_addr_get(edata); + size_t *p = (size_t *)(uintptr_t)addr; + /* Check the first page only. */ + for (size_t i = 0; i < PAGE / sizeof(size_t); i++) { + assert(p[i] == 0); + } } } @@ -1420,187 +752,74 @@ * Increment extent_grow_next if doing so wouldn't exceed the allowed * range. */ - if (arena->extent_grow_next + egn_skip + 1 <= - arena->retain_grow_limit) { - arena->extent_grow_next += egn_skip + 1; - } else { - arena->extent_grow_next = arena->retain_grow_limit; - } /* All opportunities for failure are past. */ - malloc_mutex_unlock(tsdn, &arena->extent_grow_mtx); + exp_grow_size_commit(&pac->exp_grow, exp_grow_skip); + malloc_mutex_unlock(tsdn, &pac->grow_mtx); if (config_prof) { /* Adjust gdump stats now that extent is final size. */ - extent_gdump_add(tsdn, extent); + extent_gdump_add(tsdn, edata); } - if (pad != 0) { - extent_addr_randomize(tsdn, extent, alignment); - } - if (slab) { - rtree_ctx_t rtree_ctx_fallback; - rtree_ctx_t *rtree_ctx = tsdn_rtree_ctx(tsdn, - &rtree_ctx_fallback); - - extent_slab_set(extent, true); - extent_interior_register(tsdn, rtree_ctx, extent, szind); + if (zero && !edata_zeroed_get(edata)) { + ehooks_zero(tsdn, ehooks, edata_base_get(edata), + edata_size_get(edata)); } - if (*zero && !extent_zeroed_get(extent)) { - void *addr = extent_base_get(extent); - size_t size = extent_size_get(extent); - if (extent_need_manual_zero(arena) || - pages_purge_forced(addr, size)) { - memset(addr, 0, size); - } - } - - return extent; + return edata; label_err: - malloc_mutex_unlock(tsdn, &arena->extent_grow_mtx); + malloc_mutex_unlock(tsdn, &pac->grow_mtx); return NULL; } -static extent_t * -extent_alloc_retained(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, void *new_addr, size_t size, size_t pad, - size_t alignment, bool slab, szind_t szind, bool *zero, bool *commit) { +static edata_t * +extent_alloc_retained(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, + edata_t *expand_edata, size_t size, size_t alignment, bool zero, + bool *commit, bool guarded) { assert(size != 0); assert(alignment != 0); - malloc_mutex_lock(tsdn, &arena->extent_grow_mtx); + malloc_mutex_lock(tsdn, &pac->grow_mtx); - extent_t *extent = extent_recycle(tsdn, arena, r_extent_hooks, - &arena->extents_retained, new_addr, size, pad, alignment, slab, - szind, zero, commit, true); - if (extent != NULL) { - malloc_mutex_unlock(tsdn, &arena->extent_grow_mtx); + edata_t *edata = extent_recycle(tsdn, pac, ehooks, + &pac->ecache_retained, expand_edata, size, alignment, zero, commit, + /* growing_retained */ true, guarded); + if (edata != NULL) { + malloc_mutex_unlock(tsdn, &pac->grow_mtx); if (config_prof) { - extent_gdump_add(tsdn, extent); + extent_gdump_add(tsdn, edata); } - } else if (opt_retain && new_addr == NULL) { - extent = extent_grow_retained(tsdn, arena, r_extent_hooks, size, - pad, alignment, slab, szind, zero, commit); - /* extent_grow_retained() always releases extent_grow_mtx. */ - } else { - malloc_mutex_unlock(tsdn, &arena->extent_grow_mtx); - } - malloc_mutex_assert_not_owner(tsdn, &arena->extent_grow_mtx); - - return extent; -} - -static extent_t * -extent_alloc_wrapper_hard(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, void *new_addr, size_t size, size_t pad, - size_t alignment, bool slab, szind_t szind, bool *zero, bool *commit) { - size_t esize = size + pad; - extent_t *extent = extent_alloc(tsdn, arena); - if (extent == NULL) { - return NULL; - } - void *addr; - size_t palignment = ALIGNMENT_CEILING(alignment, PAGE); - if (*r_extent_hooks == &extent_hooks_default) { - /* Call directly to propagate tsdn. */ - addr = extent_alloc_default_impl(tsdn, arena, new_addr, esize, - palignment, zero, commit); + } else if (opt_retain && expand_edata == NULL && !guarded) { + edata = extent_grow_retained(tsdn, pac, ehooks, size, + alignment, zero, commit); + /* extent_grow_retained() always releases pac->grow_mtx. */ } else { - extent_hook_pre_reentrancy(tsdn, arena); - addr = (*r_extent_hooks)->alloc(*r_extent_hooks, new_addr, - esize, palignment, zero, commit, arena_ind_get(arena)); - extent_hook_post_reentrancy(tsdn); - } - if (addr == NULL) { - extent_dalloc(tsdn, arena, extent); - return NULL; - } - extent_init(extent, arena, addr, esize, slab, szind, - arena_extent_sn_next(arena), extent_state_active, *zero, *commit, - true, EXTENT_NOT_HEAD); - if (pad != 0) { - extent_addr_randomize(tsdn, extent, alignment); - } - if (extent_register(tsdn, extent)) { - extent_dalloc(tsdn, arena, extent); - return NULL; - } - - return extent; -} - -extent_t * -extent_alloc_wrapper(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, void *new_addr, size_t size, size_t pad, - size_t alignment, bool slab, szind_t szind, bool *zero, bool *commit) { - witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), - WITNESS_RANK_CORE, 0); - - extent_hooks_assure_initialized(arena, r_extent_hooks); - - extent_t *extent = extent_alloc_retained(tsdn, arena, r_extent_hooks, - new_addr, size, pad, alignment, slab, szind, zero, commit); - if (extent == NULL) { - if (opt_retain && new_addr != NULL) { - /* - * When retain is enabled and new_addr is set, we do not - * attempt extent_alloc_wrapper_hard which does mmap - * that is very unlikely to succeed (unless it happens - * to be at the end). - */ - return NULL; - } - extent = extent_alloc_wrapper_hard(tsdn, arena, r_extent_hooks, - new_addr, size, pad, alignment, slab, szind, zero, commit); + malloc_mutex_unlock(tsdn, &pac->grow_mtx); } + malloc_mutex_assert_not_owner(tsdn, &pac->grow_mtx); - assert(extent == NULL || extent_dumpable_get(extent)); - return extent; + return edata; } static bool -extent_can_coalesce(arena_t *arena, extents_t *extents, const extent_t *inner, - const extent_t *outer) { - assert(extent_arena_get(inner) == arena); - if (extent_arena_get(outer) != arena) { - return false; - } - - assert(extent_state_get(inner) == extent_state_active); - if (extent_state_get(outer) != extents->state) { - return false; - } - - if (extent_committed_get(inner) != extent_committed_get(outer)) { - return false; - } - - return true; -} - -static bool -extent_coalesce(tsdn_t *tsdn, arena_t *arena, extent_hooks_t **r_extent_hooks, - extents_t *extents, extent_t *inner, extent_t *outer, bool forward, - bool growing_retained) { - assert(extent_can_coalesce(arena, extents, inner, outer)); - - extent_activate_locked(tsdn, arena, extents, outer); - - malloc_mutex_unlock(tsdn, &extents->mtx); - bool err = extent_merge_impl(tsdn, arena, r_extent_hooks, - forward ? inner : outer, forward ? outer : inner, growing_retained); - malloc_mutex_lock(tsdn, &extents->mtx); - +extent_coalesce(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, ecache_t *ecache, + edata_t *inner, edata_t *outer, bool forward) { + extent_assert_can_coalesce(inner, outer); + eset_remove(&ecache->eset, outer); + + bool err = extent_merge_impl(tsdn, pac, ehooks, + forward ? inner : outer, forward ? outer : inner, + /* holding_core_locks */ true); if (err) { - extent_deactivate_locked(tsdn, arena, extents, outer); + extent_deactivate_check_state_locked(tsdn, pac, ecache, outer, + extent_state_merging); } return err; } -static extent_t * -extent_try_coalesce_impl(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, rtree_ctx_t *rtree_ctx, extents_t *extents, - extent_t *extent, bool *coalesced, bool growing_retained, - bool inactive_only) { +static edata_t * +extent_try_coalesce_impl(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, + ecache_t *ecache, edata_t *edata, bool *coalesced) { + assert(!edata_guarded_get(edata)); /* * We avoid checking / locking inactive neighbors for large size * classes, since they are eagerly coalesced on deallocation which can @@ -1615,470 +834,333 @@ again = false; /* Try to coalesce forward. */ - extent_t *next = extent_lock_from_addr(tsdn, rtree_ctx, - extent_past_get(extent), inactive_only); + edata_t *next = emap_try_acquire_edata_neighbor(tsdn, pac->emap, + edata, EXTENT_PAI_PAC, ecache->state, /* forward */ true); if (next != NULL) { - /* - * extents->mtx only protects against races for - * like-state extents, so call extent_can_coalesce() - * before releasing next's pool lock. - */ - bool can_coalesce = extent_can_coalesce(arena, extents, - extent, next); - - extent_unlock(tsdn, next); - - if (can_coalesce && !extent_coalesce(tsdn, arena, - r_extent_hooks, extents, extent, next, true, - growing_retained)) { - if (extents->delay_coalesce) { + if (!extent_coalesce(tsdn, pac, ehooks, ecache, edata, + next, true)) { + if (ecache->delay_coalesce) { /* Do minimal coalescing. */ *coalesced = true; - return extent; + return edata; } again = true; } } /* Try to coalesce backward. */ - extent_t *prev = NULL; - if (extent_before_get(extent) != NULL) { - prev = extent_lock_from_addr(tsdn, rtree_ctx, - extent_before_get(extent), inactive_only); - } + edata_t *prev = emap_try_acquire_edata_neighbor(tsdn, pac->emap, + edata, EXTENT_PAI_PAC, ecache->state, /* forward */ false); if (prev != NULL) { - bool can_coalesce = extent_can_coalesce(arena, extents, - extent, prev); - extent_unlock(tsdn, prev); - - if (can_coalesce && !extent_coalesce(tsdn, arena, - r_extent_hooks, extents, extent, prev, false, - growing_retained)) { - extent = prev; - if (extents->delay_coalesce) { + if (!extent_coalesce(tsdn, pac, ehooks, ecache, edata, + prev, false)) { + edata = prev; + if (ecache->delay_coalesce) { /* Do minimal coalescing. */ *coalesced = true; - return extent; + return edata; } again = true; } } } while (again); - if (extents->delay_coalesce) { + if (ecache->delay_coalesce) { *coalesced = false; } - return extent; + return edata; +} + +static edata_t * +extent_try_coalesce(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, + ecache_t *ecache, edata_t *edata, bool *coalesced) { + return extent_try_coalesce_impl(tsdn, pac, ehooks, ecache, edata, + coalesced); } -static extent_t * -extent_try_coalesce(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, rtree_ctx_t *rtree_ctx, extents_t *extents, - extent_t *extent, bool *coalesced, bool growing_retained) { - return extent_try_coalesce_impl(tsdn, arena, r_extent_hooks, rtree_ctx, - extents, extent, coalesced, growing_retained, false); +static edata_t * +extent_try_coalesce_large(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, + ecache_t *ecache, edata_t *edata, bool *coalesced) { + return extent_try_coalesce_impl(tsdn, pac, ehooks, ecache, edata, + coalesced); } -static extent_t * -extent_try_coalesce_large(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, rtree_ctx_t *rtree_ctx, extents_t *extents, - extent_t *extent, bool *coalesced, bool growing_retained) { - return extent_try_coalesce_impl(tsdn, arena, r_extent_hooks, rtree_ctx, - extents, extent, coalesced, growing_retained, true); +/* Purge a single extent to retained / unmapped directly. */ +static void +extent_maximally_purge(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, + edata_t *edata) { + size_t extent_size = edata_size_get(edata); + extent_dalloc_wrapper(tsdn, pac, ehooks, edata); + if (config_stats) { + /* Update stats accordingly. */ + LOCKEDINT_MTX_LOCK(tsdn, *pac->stats_mtx); + locked_inc_u64(tsdn, + LOCKEDINT_MTX(*pac->stats_mtx), + &pac->stats->decay_dirty.nmadvise, 1); + locked_inc_u64(tsdn, + LOCKEDINT_MTX(*pac->stats_mtx), + &pac->stats->decay_dirty.purged, + extent_size >> LG_PAGE); + LOCKEDINT_MTX_UNLOCK(tsdn, *pac->stats_mtx); + atomic_fetch_sub_zu(&pac->stats->pac_mapped, extent_size, + ATOMIC_RELAXED); + } } /* * Does the metadata management portions of putting an unused extent into the - * given extents_t (coalesces, deregisters slab interiors, the heap operations). + * given ecache_t (coalesces and inserts into the eset). */ -static void -extent_record(tsdn_t *tsdn, arena_t *arena, extent_hooks_t **r_extent_hooks, - extents_t *extents, extent_t *extent, bool growing_retained) { - rtree_ctx_t rtree_ctx_fallback; - rtree_ctx_t *rtree_ctx = tsdn_rtree_ctx(tsdn, &rtree_ctx_fallback); - - assert((extents_state_get(extents) != extent_state_dirty && - extents_state_get(extents) != extent_state_muzzy) || - !extent_zeroed_get(extent)); - - malloc_mutex_lock(tsdn, &extents->mtx); - extent_hooks_assure_initialized(arena, r_extent_hooks); - - extent_szind_set(extent, SC_NSIZES); - if (extent_slab_get(extent)) { - extent_interior_deregister(tsdn, rtree_ctx, extent); - extent_slab_set(extent, false); - } +void +extent_record(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, ecache_t *ecache, + edata_t *edata) { + assert((ecache->state != extent_state_dirty && + ecache->state != extent_state_muzzy) || + !edata_zeroed_get(edata)); + + malloc_mutex_lock(tsdn, &ecache->mtx); - assert(rtree_extent_read(tsdn, &extents_rtree, rtree_ctx, - (uintptr_t)extent_base_get(extent), true) == extent); + emap_assert_mapped(tsdn, pac->emap, edata); - if (!extents->delay_coalesce) { - extent = extent_try_coalesce(tsdn, arena, r_extent_hooks, - rtree_ctx, extents, extent, NULL, growing_retained); - } else if (extent_size_get(extent) >= SC_LARGE_MINCLASS) { - assert(extents == &arena->extents_dirty); + if (edata_guarded_get(edata)) { + goto label_skip_coalesce; + } + if (!ecache->delay_coalesce) { + edata = extent_try_coalesce(tsdn, pac, ehooks, ecache, edata, + NULL); + } else if (edata_size_get(edata) >= SC_LARGE_MINCLASS) { + assert(ecache == &pac->ecache_dirty); /* Always coalesce large extents eagerly. */ bool coalesced; do { - assert(extent_state_get(extent) == extent_state_active); - extent = extent_try_coalesce_large(tsdn, arena, - r_extent_hooks, rtree_ctx, extents, extent, - &coalesced, growing_retained); + assert(edata_state_get(edata) == extent_state_active); + edata = extent_try_coalesce_large(tsdn, pac, ehooks, + ecache, edata, &coalesced); } while (coalesced); - if (extent_size_get(extent) >= oversize_threshold) { + if (edata_size_get(edata) >= + atomic_load_zu(&pac->oversize_threshold, ATOMIC_RELAXED) + && extent_may_force_decay(pac)) { /* Shortcut to purge the oversize extent eagerly. */ - malloc_mutex_unlock(tsdn, &extents->mtx); - arena_decay_extent(tsdn, arena, r_extent_hooks, extent); + malloc_mutex_unlock(tsdn, &ecache->mtx); + extent_maximally_purge(tsdn, pac, ehooks, edata); return; } } - extent_deactivate_locked(tsdn, arena, extents, extent); +label_skip_coalesce: + extent_deactivate_locked(tsdn, pac, ecache, edata); - malloc_mutex_unlock(tsdn, &extents->mtx); + malloc_mutex_unlock(tsdn, &ecache->mtx); } void -extent_dalloc_gap(tsdn_t *tsdn, arena_t *arena, extent_t *extent) { - extent_hooks_t *extent_hooks = EXTENT_HOOKS_INITIALIZER; - +extent_dalloc_gap(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, + edata_t *edata) { witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), WITNESS_RANK_CORE, 0); - if (extent_register(tsdn, extent)) { - extent_dalloc(tsdn, arena, extent); + if (extent_register(tsdn, pac, edata)) { + edata_cache_put(tsdn, pac->edata_cache, edata); return; } - extent_dalloc_wrapper(tsdn, arena, &extent_hooks, extent); + extent_dalloc_wrapper(tsdn, pac, ehooks, edata); } static bool -extent_may_dalloc(void) { - /* With retain enabled, the default dalloc always fails. */ - return !opt_retain; -} - -static bool -extent_dalloc_default_impl(void *addr, size_t size) { - if (!have_dss || !extent_in_dss(addr)) { - return extent_dalloc_mmap(addr, size); - } - return true; -} - -static bool -extent_dalloc_default(extent_hooks_t *extent_hooks, void *addr, size_t size, - bool committed, unsigned arena_ind) { - return extent_dalloc_default_impl(addr, size); -} - -static bool -extent_dalloc_wrapper_try(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, extent_t *extent) { +extent_dalloc_wrapper_try(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, + edata_t *edata) { bool err; - assert(extent_base_get(extent) != NULL); - assert(extent_size_get(extent) != 0); + assert(edata_base_get(edata) != NULL); + assert(edata_size_get(edata) != 0); witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), WITNESS_RANK_CORE, 0); - extent_addr_set(extent, extent_base_get(extent)); + edata_addr_set(edata, edata_base_get(edata)); - extent_hooks_assure_initialized(arena, r_extent_hooks); /* Try to deallocate. */ - if (*r_extent_hooks == &extent_hooks_default) { - /* Call directly to propagate tsdn. */ - err = extent_dalloc_default_impl(extent_base_get(extent), - extent_size_get(extent)); - } else { - extent_hook_pre_reentrancy(tsdn, arena); - err = ((*r_extent_hooks)->dalloc == NULL || - (*r_extent_hooks)->dalloc(*r_extent_hooks, - extent_base_get(extent), extent_size_get(extent), - extent_committed_get(extent), arena_ind_get(arena))); - extent_hook_post_reentrancy(tsdn); - } + err = ehooks_dalloc(tsdn, ehooks, edata_base_get(edata), + edata_size_get(edata), edata_committed_get(edata)); if (!err) { - extent_dalloc(tsdn, arena, extent); + edata_cache_put(tsdn, pac->edata_cache, edata); } return err; } +edata_t * +extent_alloc_wrapper(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, + void *new_addr, size_t size, size_t alignment, bool zero, bool *commit, + bool growing_retained) { + witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), + WITNESS_RANK_CORE, growing_retained ? 1 : 0); + + edata_t *edata = edata_cache_get(tsdn, pac->edata_cache); + if (edata == NULL) { + return NULL; + } + size_t palignment = ALIGNMENT_CEILING(alignment, PAGE); + void *addr = ehooks_alloc(tsdn, ehooks, new_addr, size, palignment, + &zero, commit); + if (addr == NULL) { + edata_cache_put(tsdn, pac->edata_cache, edata); + return NULL; + } + edata_init(edata, ecache_ind_get(&pac->ecache_dirty), addr, + size, /* slab */ false, SC_NSIZES, extent_sn_next(pac), + extent_state_active, zero, *commit, EXTENT_PAI_PAC, + opt_retain ? EXTENT_IS_HEAD : EXTENT_NOT_HEAD); + /* + * Retained memory is not counted towards gdump. Only if an extent is + * allocated as a separate mapping, i.e. growing_retained is false, then + * gdump should be updated. + */ + bool gdump_add = !growing_retained; + if (extent_register_impl(tsdn, pac, edata, gdump_add)) { + edata_cache_put(tsdn, pac->edata_cache, edata); + return NULL; + } + + return edata; +} + void -extent_dalloc_wrapper(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, extent_t *extent) { - assert(extent_dumpable_get(extent)); +extent_dalloc_wrapper(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, + edata_t *edata) { + assert(edata_pai_get(edata) == EXTENT_PAI_PAC); witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), WITNESS_RANK_CORE, 0); /* Avoid calling the default extent_dalloc unless have to. */ - if (*r_extent_hooks != &extent_hooks_default || extent_may_dalloc()) { + if (!ehooks_dalloc_will_fail(ehooks)) { + /* Remove guard pages for dalloc / unmap. */ + if (edata_guarded_get(edata)) { + assert(ehooks_are_default(ehooks)); + san_unguard_pages_two_sided(tsdn, ehooks, edata, + pac->emap); + } /* * Deregister first to avoid a race with other allocating * threads, and reregister if deallocation fails. */ - extent_deregister(tsdn, extent); - if (!extent_dalloc_wrapper_try(tsdn, arena, r_extent_hooks, - extent)) { + extent_deregister(tsdn, pac, edata); + if (!extent_dalloc_wrapper_try(tsdn, pac, ehooks, edata)) { return; } - extent_reregister(tsdn, extent); + extent_reregister(tsdn, pac, edata); } - if (*r_extent_hooks != &extent_hooks_default) { - extent_hook_pre_reentrancy(tsdn, arena); - } /* Try to decommit; purge if that fails. */ bool zeroed; - if (!extent_committed_get(extent)) { + if (!edata_committed_get(edata)) { zeroed = true; - } else if (!extent_decommit_wrapper(tsdn, arena, r_extent_hooks, extent, - 0, extent_size_get(extent))) { + } else if (!extent_decommit_wrapper(tsdn, ehooks, edata, 0, + edata_size_get(edata))) { zeroed = true; - } else if ((*r_extent_hooks)->purge_forced != NULL && - !(*r_extent_hooks)->purge_forced(*r_extent_hooks, - extent_base_get(extent), extent_size_get(extent), 0, - extent_size_get(extent), arena_ind_get(arena))) { + } else if (!ehooks_purge_forced(tsdn, ehooks, edata_base_get(edata), + edata_size_get(edata), 0, edata_size_get(edata))) { zeroed = true; - } else if (extent_state_get(extent) == extent_state_muzzy || - ((*r_extent_hooks)->purge_lazy != NULL && - !(*r_extent_hooks)->purge_lazy(*r_extent_hooks, - extent_base_get(extent), extent_size_get(extent), 0, - extent_size_get(extent), arena_ind_get(arena)))) { + } else if (edata_state_get(edata) == extent_state_muzzy || + !ehooks_purge_lazy(tsdn, ehooks, edata_base_get(edata), + edata_size_get(edata), 0, edata_size_get(edata))) { zeroed = false; } else { zeroed = false; } - if (*r_extent_hooks != &extent_hooks_default) { - extent_hook_post_reentrancy(tsdn); - } - extent_zeroed_set(extent, zeroed); + edata_zeroed_set(edata, zeroed); if (config_prof) { - extent_gdump_sub(tsdn, extent); - } - - extent_record(tsdn, arena, r_extent_hooks, &arena->extents_retained, - extent, false); -} - -static void -extent_destroy_default_impl(void *addr, size_t size) { - if (!have_dss || !extent_in_dss(addr)) { - pages_unmap(addr, size); + extent_gdump_sub(tsdn, edata); } -} -static void -extent_destroy_default(extent_hooks_t *extent_hooks, void *addr, size_t size, - bool committed, unsigned arena_ind) { - extent_destroy_default_impl(addr, size); + extent_record(tsdn, pac, ehooks, &pac->ecache_retained, edata); } void -extent_destroy_wrapper(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, extent_t *extent) { - assert(extent_base_get(extent) != NULL); - assert(extent_size_get(extent) != 0); +extent_destroy_wrapper(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, + edata_t *edata) { + assert(edata_base_get(edata) != NULL); + assert(edata_size_get(edata) != 0); + extent_state_t state = edata_state_get(edata); + assert(state == extent_state_retained || state == extent_state_active); + assert(emap_edata_is_acquired(tsdn, pac->emap, edata)); witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), WITNESS_RANK_CORE, 0); - /* Deregister first to avoid a race with other allocating threads. */ - extent_deregister(tsdn, extent); - - extent_addr_set(extent, extent_base_get(extent)); - - extent_hooks_assure_initialized(arena, r_extent_hooks); - /* Try to destroy; silently fail otherwise. */ - if (*r_extent_hooks == &extent_hooks_default) { - /* Call directly to propagate tsdn. */ - extent_destroy_default_impl(extent_base_get(extent), - extent_size_get(extent)); - } else if ((*r_extent_hooks)->destroy != NULL) { - extent_hook_pre_reentrancy(tsdn, arena); - (*r_extent_hooks)->destroy(*r_extent_hooks, - extent_base_get(extent), extent_size_get(extent), - extent_committed_get(extent), arena_ind_get(arena)); - extent_hook_post_reentrancy(tsdn); + if (edata_guarded_get(edata)) { + assert(opt_retain); + san_unguard_pages_pre_destroy(tsdn, ehooks, edata, pac->emap); } + edata_addr_set(edata, edata_base_get(edata)); - extent_dalloc(tsdn, arena, extent); -} + /* Try to destroy; silently fail otherwise. */ + ehooks_destroy(tsdn, ehooks, edata_base_get(edata), + edata_size_get(edata), edata_committed_get(edata)); -static bool -extent_commit_default(extent_hooks_t *extent_hooks, void *addr, size_t size, - size_t offset, size_t length, unsigned arena_ind) { - return pages_commit((void *)((uintptr_t)addr + (uintptr_t)offset), - length); + edata_cache_put(tsdn, pac->edata_cache, edata); } static bool -extent_commit_impl(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, extent_t *extent, size_t offset, - size_t length, bool growing_retained) { +extent_commit_impl(tsdn_t *tsdn, ehooks_t *ehooks, edata_t *edata, + size_t offset, size_t length, bool growing_retained) { witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), WITNESS_RANK_CORE, growing_retained ? 1 : 0); - - extent_hooks_assure_initialized(arena, r_extent_hooks); - if (*r_extent_hooks != &extent_hooks_default) { - extent_hook_pre_reentrancy(tsdn, arena); - } - bool err = ((*r_extent_hooks)->commit == NULL || - (*r_extent_hooks)->commit(*r_extent_hooks, extent_base_get(extent), - extent_size_get(extent), offset, length, arena_ind_get(arena))); - if (*r_extent_hooks != &extent_hooks_default) { - extent_hook_post_reentrancy(tsdn); - } - extent_committed_set(extent, extent_committed_get(extent) || !err); + bool err = ehooks_commit(tsdn, ehooks, edata_base_get(edata), + edata_size_get(edata), offset, length); + edata_committed_set(edata, edata_committed_get(edata) || !err); return err; } bool -extent_commit_wrapper(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, extent_t *extent, size_t offset, - size_t length) { - return extent_commit_impl(tsdn, arena, r_extent_hooks, extent, offset, - length, false); -} - -static bool -extent_decommit_default(extent_hooks_t *extent_hooks, void *addr, size_t size, - size_t offset, size_t length, unsigned arena_ind) { - return pages_decommit((void *)((uintptr_t)addr + (uintptr_t)offset), - length); +extent_commit_wrapper(tsdn_t *tsdn, ehooks_t *ehooks, edata_t *edata, + size_t offset, size_t length) { + return extent_commit_impl(tsdn, ehooks, edata, offset, length, + /* growing_retained */ false); } bool -extent_decommit_wrapper(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, extent_t *extent, size_t offset, - size_t length) { +extent_decommit_wrapper(tsdn_t *tsdn, ehooks_t *ehooks, edata_t *edata, + size_t offset, size_t length) { witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), WITNESS_RANK_CORE, 0); - - extent_hooks_assure_initialized(arena, r_extent_hooks); - - if (*r_extent_hooks != &extent_hooks_default) { - extent_hook_pre_reentrancy(tsdn, arena); - } - bool err = ((*r_extent_hooks)->decommit == NULL || - (*r_extent_hooks)->decommit(*r_extent_hooks, - extent_base_get(extent), extent_size_get(extent), offset, length, - arena_ind_get(arena))); - if (*r_extent_hooks != &extent_hooks_default) { - extent_hook_post_reentrancy(tsdn); - } - extent_committed_set(extent, extent_committed_get(extent) && err); + bool err = ehooks_decommit(tsdn, ehooks, edata_base_get(edata), + edata_size_get(edata), offset, length); + edata_committed_set(edata, edata_committed_get(edata) && err); return err; } -#ifdef PAGES_CAN_PURGE_LAZY static bool -extent_purge_lazy_default(extent_hooks_t *extent_hooks, void *addr, size_t size, - size_t offset, size_t length, unsigned arena_ind) { - assert(addr != NULL); - assert((offset & PAGE_MASK) == 0); - assert(length != 0); - assert((length & PAGE_MASK) == 0); - - return pages_purge_lazy((void *)((uintptr_t)addr + (uintptr_t)offset), - length); -} -#endif - -static bool -extent_purge_lazy_impl(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, extent_t *extent, size_t offset, - size_t length, bool growing_retained) { +extent_purge_lazy_impl(tsdn_t *tsdn, ehooks_t *ehooks, edata_t *edata, + size_t offset, size_t length, bool growing_retained) { witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), WITNESS_RANK_CORE, growing_retained ? 1 : 0); - - extent_hooks_assure_initialized(arena, r_extent_hooks); - - if ((*r_extent_hooks)->purge_lazy == NULL) { - return true; - } - if (*r_extent_hooks != &extent_hooks_default) { - extent_hook_pre_reentrancy(tsdn, arena); - } - bool err = (*r_extent_hooks)->purge_lazy(*r_extent_hooks, - extent_base_get(extent), extent_size_get(extent), offset, length, - arena_ind_get(arena)); - if (*r_extent_hooks != &extent_hooks_default) { - extent_hook_post_reentrancy(tsdn); - } - + bool err = ehooks_purge_lazy(tsdn, ehooks, edata_base_get(edata), + edata_size_get(edata), offset, length); return err; } bool -extent_purge_lazy_wrapper(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, extent_t *extent, size_t offset, - size_t length) { - return extent_purge_lazy_impl(tsdn, arena, r_extent_hooks, extent, - offset, length, false); -} - -#ifdef PAGES_CAN_PURGE_FORCED -static bool -extent_purge_forced_default(extent_hooks_t *extent_hooks, void *addr, - size_t size, size_t offset, size_t length, unsigned arena_ind) { - assert(addr != NULL); - assert((offset & PAGE_MASK) == 0); - assert(length != 0); - assert((length & PAGE_MASK) == 0); - - return pages_purge_forced((void *)((uintptr_t)addr + - (uintptr_t)offset), length); +extent_purge_lazy_wrapper(tsdn_t *tsdn, ehooks_t *ehooks, edata_t *edata, + size_t offset, size_t length) { + return extent_purge_lazy_impl(tsdn, ehooks, edata, offset, + length, false); } -#endif static bool -extent_purge_forced_impl(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, extent_t *extent, size_t offset, - size_t length, bool growing_retained) { +extent_purge_forced_impl(tsdn_t *tsdn, ehooks_t *ehooks, edata_t *edata, + size_t offset, size_t length, bool growing_retained) { witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), WITNESS_RANK_CORE, growing_retained ? 1 : 0); - - extent_hooks_assure_initialized(arena, r_extent_hooks); - - if ((*r_extent_hooks)->purge_forced == NULL) { - return true; - } - if (*r_extent_hooks != &extent_hooks_default) { - extent_hook_pre_reentrancy(tsdn, arena); - } - bool err = (*r_extent_hooks)->purge_forced(*r_extent_hooks, - extent_base_get(extent), extent_size_get(extent), offset, length, - arena_ind_get(arena)); - if (*r_extent_hooks != &extent_hooks_default) { - extent_hook_post_reentrancy(tsdn); - } + bool err = ehooks_purge_forced(tsdn, ehooks, edata_base_get(edata), + edata_size_get(edata), offset, length); return err; } bool -extent_purge_forced_wrapper(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, extent_t *extent, size_t offset, - size_t length) { - return extent_purge_forced_impl(tsdn, arena, r_extent_hooks, extent, - offset, length, false); -} - -static bool -extent_split_default(extent_hooks_t *extent_hooks, void *addr, size_t size, - size_t size_a, size_t size_b, bool committed, unsigned arena_ind) { - if (!maps_coalesce) { - /* - * Without retain, only whole regions can be purged (required by - * MEM_RELEASE on Windows) -- therefore disallow splitting. See - * comments in extent_head_no_merge(). - */ - return !opt_retain; - } - - return false; +extent_purge_forced_wrapper(tsdn_t *tsdn, ehooks_t *ehooks, edata_t *edata, + size_t offset, size_t length) { + return extent_purge_forced_impl(tsdn, ehooks, edata, offset, length, + false); } /* @@ -2088,183 +1170,95 @@ * with the trail (the higher addressed portion). This makes 'extent' the lead, * and returns the trail (except in case of error). */ -static extent_t * -extent_split_impl(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, extent_t *extent, size_t size_a, - szind_t szind_a, bool slab_a, size_t size_b, szind_t szind_b, bool slab_b, - bool growing_retained) { - assert(extent_size_get(extent) == size_a + size_b); - witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), - WITNESS_RANK_CORE, growing_retained ? 1 : 0); - - extent_hooks_assure_initialized(arena, r_extent_hooks); +static edata_t * +extent_split_impl(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, + edata_t *edata, size_t size_a, size_t size_b, bool holding_core_locks) { + assert(edata_size_get(edata) == size_a + size_b); + /* Only the shrink path may split w/o holding core locks. */ + if (holding_core_locks) { + witness_assert_positive_depth_to_rank( + tsdn_witness_tsdp_get(tsdn), WITNESS_RANK_CORE); + } else { + witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), + WITNESS_RANK_CORE, 0); + } - if ((*r_extent_hooks)->split == NULL) { + if (ehooks_split_will_fail(ehooks)) { return NULL; } - extent_t *trail = extent_alloc(tsdn, arena); + edata_t *trail = edata_cache_get(tsdn, pac->edata_cache); if (trail == NULL) { goto label_error_a; } - extent_init(trail, arena, (void *)((uintptr_t)extent_base_get(extent) + - size_a), size_b, slab_b, szind_b, extent_sn_get(extent), - extent_state_get(extent), extent_zeroed_get(extent), - extent_committed_get(extent), extent_dumpable_get(extent), - EXTENT_NOT_HEAD); - - rtree_ctx_t rtree_ctx_fallback; - rtree_ctx_t *rtree_ctx = tsdn_rtree_ctx(tsdn, &rtree_ctx_fallback); - rtree_leaf_elm_t *lead_elm_a, *lead_elm_b; - { - extent_t lead; - - extent_init(&lead, arena, extent_addr_get(extent), size_a, - slab_a, szind_a, extent_sn_get(extent), - extent_state_get(extent), extent_zeroed_get(extent), - extent_committed_get(extent), extent_dumpable_get(extent), - EXTENT_NOT_HEAD); - - extent_rtree_leaf_elms_lookup(tsdn, rtree_ctx, &lead, false, - true, &lead_elm_a, &lead_elm_b); - } - rtree_leaf_elm_t *trail_elm_a, *trail_elm_b; - extent_rtree_leaf_elms_lookup(tsdn, rtree_ctx, trail, false, true, - &trail_elm_a, &trail_elm_b); - - if (lead_elm_a == NULL || lead_elm_b == NULL || trail_elm_a == NULL - || trail_elm_b == NULL) { + edata_init(trail, edata_arena_ind_get(edata), + (void *)((uintptr_t)edata_base_get(edata) + size_a), size_b, + /* slab */ false, SC_NSIZES, edata_sn_get(edata), + edata_state_get(edata), edata_zeroed_get(edata), + edata_committed_get(edata), EXTENT_PAI_PAC, EXTENT_NOT_HEAD); + emap_prepare_t prepare; + bool err = emap_split_prepare(tsdn, pac->emap, &prepare, edata, + size_a, trail, size_b); + if (err) { goto label_error_b; } - extent_lock2(tsdn, extent, trail); + /* + * No need to acquire trail or edata, because: 1) trail was new (just + * allocated); and 2) edata is either an active allocation (the shrink + * path), or in an acquired state (extracted from the ecache on the + * extent_recycle_split path). + */ + assert(emap_edata_is_acquired(tsdn, pac->emap, edata)); + assert(emap_edata_is_acquired(tsdn, pac->emap, trail)); + + err = ehooks_split(tsdn, ehooks, edata_base_get(edata), size_a + size_b, + size_a, size_b, edata_committed_get(edata)); - if (*r_extent_hooks != &extent_hooks_default) { - extent_hook_pre_reentrancy(tsdn, arena); - } - bool err = (*r_extent_hooks)->split(*r_extent_hooks, extent_base_get(extent), - size_a + size_b, size_a, size_b, extent_committed_get(extent), - arena_ind_get(arena)); - if (*r_extent_hooks != &extent_hooks_default) { - extent_hook_post_reentrancy(tsdn); - } if (err) { - goto label_error_c; + goto label_error_b; } - extent_size_set(extent, size_a); - extent_szind_set(extent, szind_a); - - extent_rtree_write_acquired(tsdn, lead_elm_a, lead_elm_b, extent, - szind_a, slab_a); - extent_rtree_write_acquired(tsdn, trail_elm_a, trail_elm_b, trail, - szind_b, slab_b); - - extent_unlock2(tsdn, extent, trail); + edata_size_set(edata, size_a); + emap_split_commit(tsdn, pac->emap, &prepare, edata, size_a, trail, + size_b); return trail; -label_error_c: - extent_unlock2(tsdn, extent, trail); label_error_b: - extent_dalloc(tsdn, arena, trail); + edata_cache_put(tsdn, pac->edata_cache, trail); label_error_a: return NULL; } -extent_t * -extent_split_wrapper(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, extent_t *extent, size_t size_a, - szind_t szind_a, bool slab_a, size_t size_b, szind_t szind_b, bool slab_b) { - return extent_split_impl(tsdn, arena, r_extent_hooks, extent, size_a, - szind_a, slab_a, size_b, szind_b, slab_b, false); -} - -static bool -extent_merge_default_impl(void *addr_a, void *addr_b) { - if (!maps_coalesce && !opt_retain) { - return true; - } - if (have_dss && !extent_dss_mergeable(addr_a, addr_b)) { - return true; - } - - return false; -} - -/* - * Returns true if the given extents can't be merged because of their head bit - * settings. Assumes the second extent has the higher address. - */ -static bool -extent_head_no_merge(extent_t *a, extent_t *b) { - assert(extent_base_get(a) < extent_base_get(b)); - /* - * When coalesce is not always allowed (Windows), only merge extents - * from the same VirtualAlloc region under opt.retain (in which case - * MEM_DECOMMIT is utilized for purging). - */ - if (maps_coalesce) { - return false; - } - if (!opt_retain) { - return true; - } - /* If b is a head extent, disallow the cross-region merge. */ - if (extent_is_head_get(b)) { - /* - * Additionally, sn should not overflow with retain; sanity - * check that different regions have unique sn. - */ - assert(extent_sn_comp(a, b) != 0); - return true; - } - assert(extent_sn_comp(a, b) == 0); - - return false; +edata_t * +extent_split_wrapper(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, edata_t *edata, + size_t size_a, size_t size_b, bool holding_core_locks) { + return extent_split_impl(tsdn, pac, ehooks, edata, size_a, size_b, + holding_core_locks); } static bool -extent_merge_default(extent_hooks_t *extent_hooks, void *addr_a, size_t size_a, - void *addr_b, size_t size_b, bool committed, unsigned arena_ind) { - if (!maps_coalesce) { - tsdn_t *tsdn = tsdn_fetch(); - extent_t *a = iealloc(tsdn, addr_a); - extent_t *b = iealloc(tsdn, addr_b); - if (extent_head_no_merge(a, b)) { - return true; - } +extent_merge_impl(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, edata_t *a, + edata_t *b, bool holding_core_locks) { + /* Only the expanding path may merge w/o holding ecache locks. */ + if (holding_core_locks) { + witness_assert_positive_depth_to_rank( + tsdn_witness_tsdp_get(tsdn), WITNESS_RANK_CORE); + } else { + witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), + WITNESS_RANK_CORE, 0); } - return extent_merge_default_impl(addr_a, addr_b); -} - -static bool -extent_merge_impl(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, extent_t *a, extent_t *b, - bool growing_retained) { - witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), - WITNESS_RANK_CORE, growing_retained ? 1 : 0); - assert(extent_base_get(a) < extent_base_get(b)); - - extent_hooks_assure_initialized(arena, r_extent_hooks); - if ((*r_extent_hooks)->merge == NULL || extent_head_no_merge(a, b)) { - return true; - } + assert(edata_base_get(a) < edata_base_get(b)); + assert(edata_arena_ind_get(a) == edata_arena_ind_get(b)); + assert(edata_arena_ind_get(a) == ehooks_ind_get(ehooks)); + emap_assert_mapped(tsdn, pac->emap, a); + emap_assert_mapped(tsdn, pac->emap, b); - bool err; - if (*r_extent_hooks == &extent_hooks_default) { - /* Call directly to propagate tsdn. */ - err = extent_merge_default_impl(extent_base_get(a), - extent_base_get(b)); - } else { - extent_hook_pre_reentrancy(tsdn, arena); - err = (*r_extent_hooks)->merge(*r_extent_hooks, - extent_base_get(a), extent_size_get(a), extent_base_get(b), - extent_size_get(b), extent_committed_get(a), - arena_ind_get(arena)); - extent_hook_post_reentrancy(tsdn); - } + bool err = ehooks_merge(tsdn, ehooks, edata_base_get(a), + edata_size_get(a), edata_base_get(b), edata_size_get(b), + edata_committed_get(a)); if (err) { return true; @@ -2275,132 +1269,58 @@ * owned, so the following code uses decomposed helper functions rather * than extent_{,de}register() to do things in the right order. */ - rtree_ctx_t rtree_ctx_fallback; - rtree_ctx_t *rtree_ctx = tsdn_rtree_ctx(tsdn, &rtree_ctx_fallback); - rtree_leaf_elm_t *a_elm_a, *a_elm_b, *b_elm_a, *b_elm_b; - extent_rtree_leaf_elms_lookup(tsdn, rtree_ctx, a, true, false, &a_elm_a, - &a_elm_b); - extent_rtree_leaf_elms_lookup(tsdn, rtree_ctx, b, true, false, &b_elm_a, - &b_elm_b); - - extent_lock2(tsdn, a, b); - - if (a_elm_b != NULL) { - rtree_leaf_elm_write(tsdn, &extents_rtree, a_elm_b, NULL, - SC_NSIZES, false); - } - if (b_elm_b != NULL) { - rtree_leaf_elm_write(tsdn, &extents_rtree, b_elm_a, NULL, - SC_NSIZES, false); - } else { - b_elm_b = b_elm_a; - } + emap_prepare_t prepare; + emap_merge_prepare(tsdn, pac->emap, &prepare, a, b); - extent_size_set(a, extent_size_get(a) + extent_size_get(b)); - extent_szind_set(a, SC_NSIZES); - extent_sn_set(a, (extent_sn_get(a) < extent_sn_get(b)) ? - extent_sn_get(a) : extent_sn_get(b)); - extent_zeroed_set(a, extent_zeroed_get(a) && extent_zeroed_get(b)); + assert(edata_state_get(a) == extent_state_active || + edata_state_get(a) == extent_state_merging); + edata_state_set(a, extent_state_active); + edata_size_set(a, edata_size_get(a) + edata_size_get(b)); + edata_sn_set(a, (edata_sn_get(a) < edata_sn_get(b)) ? + edata_sn_get(a) : edata_sn_get(b)); + edata_zeroed_set(a, edata_zeroed_get(a) && edata_zeroed_get(b)); - extent_rtree_write_acquired(tsdn, a_elm_a, b_elm_b, a, SC_NSIZES, - false); + emap_merge_commit(tsdn, pac->emap, &prepare, a, b); - extent_unlock2(tsdn, a, b); - - extent_dalloc(tsdn, extent_arena_get(b), b); + edata_cache_put(tsdn, pac->edata_cache, b); return false; } bool -extent_merge_wrapper(tsdn_t *tsdn, arena_t *arena, - extent_hooks_t **r_extent_hooks, extent_t *a, extent_t *b) { - return extent_merge_impl(tsdn, arena, r_extent_hooks, a, b, false); +extent_merge_wrapper(tsdn_t *tsdn, pac_t *pac, ehooks_t *ehooks, + edata_t *a, edata_t *b) { + return extent_merge_impl(tsdn, pac, ehooks, a, b, + /* holding_core_locks */ false); } bool -extent_boot(void) { - if (rtree_new(&extents_rtree, true)) { - return true; - } +extent_commit_zero(tsdn_t *tsdn, ehooks_t *ehooks, edata_t *edata, + bool commit, bool zero, bool growing_retained) { + witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), + WITNESS_RANK_CORE, growing_retained ? 1 : 0); - if (mutex_pool_init(&extent_mutex_pool, "extent_mutex_pool", - WITNESS_RANK_EXTENT_POOL)) { - return true; + if (commit && !edata_committed_get(edata)) { + if (extent_commit_impl(tsdn, ehooks, edata, 0, + edata_size_get(edata), growing_retained)) { + return true; + } } - - if (have_dss) { - extent_dss_boot(); + if (zero && !edata_zeroed_get(edata)) { + void *addr = edata_base_get(edata); + size_t size = edata_size_get(edata); + ehooks_zero(tsdn, ehooks, addr, size); } - return false; } -void -extent_util_stats_get(tsdn_t *tsdn, const void *ptr, - size_t *nfree, size_t *nregs, size_t *size) { - assert(ptr != NULL && nfree != NULL && nregs != NULL && size != NULL); - - const extent_t *extent = iealloc(tsdn, ptr); - if (unlikely(extent == NULL)) { - *nfree = *nregs = *size = 0; - return; - } - - *size = extent_size_get(extent); - if (!extent_slab_get(extent)) { - *nfree = 0; - *nregs = 1; - } else { - *nfree = extent_nfree_get(extent); - *nregs = bin_infos[extent_szind_get(extent)].nregs; - assert(*nfree <= *nregs); - assert(*nfree * extent_usize_get(extent) <= *size); - } -} - -void -extent_util_stats_verbose_get(tsdn_t *tsdn, const void *ptr, - size_t *nfree, size_t *nregs, size_t *size, - size_t *bin_nfree, size_t *bin_nregs, void **slabcur_addr) { - assert(ptr != NULL && nfree != NULL && nregs != NULL && size != NULL - && bin_nfree != NULL && bin_nregs != NULL && slabcur_addr != NULL); - - const extent_t *extent = iealloc(tsdn, ptr); - if (unlikely(extent == NULL)) { - *nfree = *nregs = *size = *bin_nfree = *bin_nregs = 0; - *slabcur_addr = NULL; - return; - } +bool +extent_boot(void) { + assert(sizeof(slab_data_t) >= sizeof(e_prof_info_t)); - *size = extent_size_get(extent); - if (!extent_slab_get(extent)) { - *nfree = *bin_nfree = *bin_nregs = 0; - *nregs = 1; - *slabcur_addr = NULL; - return; + if (have_dss) { + extent_dss_boot(); } - *nfree = extent_nfree_get(extent); - const szind_t szind = extent_szind_get(extent); - *nregs = bin_infos[szind].nregs; - assert(*nfree <= *nregs); - assert(*nfree * extent_usize_get(extent) <= *size); - - const arena_t *arena = extent_arena_get(extent); - assert(arena != NULL); - const unsigned binshard = extent_binshard_get(extent); - bin_t *bin = &arena->bins[szind].bin_shards[binshard]; - - malloc_mutex_lock(tsdn, &bin->lock); - if (config_stats) { - *bin_nregs = *nregs * bin->stats.curslabs; - assert(*bin_nregs >= bin->stats.curregs); - *bin_nfree = *bin_nregs - bin->stats.curregs; - } else { - *bin_nfree = *bin_nregs = 0; - } - *slabcur_addr = extent_addr_get(bin->slabcur); - assert(*slabcur_addr != NULL); - malloc_mutex_unlock(tsdn, &bin->lock); + return false; } diff --git a/contrib/jemalloc/src/extent_dss.c b/contrib/jemalloc/src/extent_dss.c --- a/contrib/jemalloc/src/extent_dss.c +++ b/contrib/jemalloc/src/extent_dss.c @@ -1,4 +1,3 @@ -#define JEMALLOC_EXTENT_DSS_C_ #include "jemalloc/internal/jemalloc_preamble.h" #include "jemalloc/internal/jemalloc_internal_includes.h" @@ -109,7 +108,7 @@ void * extent_alloc_dss(tsdn_t *tsdn, arena_t *arena, void *new_addr, size_t size, size_t alignment, bool *zero, bool *commit) { - extent_t *gap; + edata_t *gap; cassert(have_dss); assert(size > 0); @@ -123,7 +122,7 @@ return NULL; } - gap = extent_alloc(tsdn, arena); + gap = edata_cache_get(tsdn, &arena->pa_shard.edata_cache); if (gap == NULL) { return NULL; } @@ -141,6 +140,8 @@ goto label_oom; } + bool head_state = opt_retain ? EXTENT_IS_HEAD : + EXTENT_NOT_HEAD; /* * Compute how much page-aligned gap space (if any) is * necessary to satisfy alignment. This space can be @@ -153,11 +154,12 @@ size_t gap_size_page = (uintptr_t)ret - (uintptr_t)gap_addr_page; if (gap_size_page != 0) { - extent_init(gap, arena, gap_addr_page, - gap_size_page, false, SC_NSIZES, - arena_extent_sn_next(arena), - extent_state_active, false, true, true, - EXTENT_NOT_HEAD); + edata_init(gap, arena_ind_get(arena), + gap_addr_page, gap_size_page, false, + SC_NSIZES, extent_sn_next( + &arena->pa_shard.pac), + extent_state_active, false, true, + EXTENT_PAI_PAC, head_state); } /* * Compute the address just past the end of the desired @@ -186,25 +188,29 @@ extent_dss_extending_finish(); if (gap_size_page != 0) { - extent_dalloc_gap(tsdn, arena, gap); + ehooks_t *ehooks = arena_get_ehooks( + arena); + extent_dalloc_gap(tsdn, + &arena->pa_shard.pac, ehooks, gap); } else { - extent_dalloc(tsdn, arena, gap); + edata_cache_put(tsdn, + &arena->pa_shard.edata_cache, gap); } if (!*commit) { *commit = pages_decommit(ret, size); } if (*zero && *commit) { - extent_hooks_t *extent_hooks = - EXTENT_HOOKS_INITIALIZER; - extent_t extent; + edata_t edata = {0}; + ehooks_t *ehooks = arena_get_ehooks( + arena); - extent_init(&extent, arena, ret, size, + edata_init(&edata, + arena_ind_get(arena), ret, size, size, false, SC_NSIZES, extent_state_active, false, true, - true, EXTENT_NOT_HEAD); + EXTENT_PAI_PAC, head_state); if (extent_purge_forced_wrapper(tsdn, - arena, &extent_hooks, &extent, 0, - size)) { + ehooks, &edata, 0, size)) { memset(ret, 0, size); } } @@ -224,7 +230,7 @@ } label_oom: extent_dss_extending_finish(); - extent_dalloc(tsdn, arena, gap); + edata_cache_put(tsdn, &arena->pa_shard.edata_cache, gap); return NULL; } diff --git a/contrib/jemalloc/src/extent_mmap.c b/contrib/jemalloc/src/extent_mmap.c --- a/contrib/jemalloc/src/extent_mmap.c +++ b/contrib/jemalloc/src/extent_mmap.c @@ -1,4 +1,3 @@ -#define JEMALLOC_EXTENT_MMAP_C_ #include "jemalloc/internal/jemalloc_preamble.h" #include "jemalloc/internal/jemalloc_internal_includes.h" diff --git a/contrib/jemalloc/src/fxp.c b/contrib/jemalloc/src/fxp.c new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/src/fxp.c @@ -0,0 +1,124 @@ +#include "jemalloc/internal/jemalloc_preamble.h" +#include "jemalloc/internal/jemalloc_internal_includes.h" + +#include "jemalloc/internal/fxp.h" + +static bool +fxp_isdigit(char c) { + return '0' <= c && c <= '9'; +} + +bool +fxp_parse(fxp_t *result, const char *str, char **end) { + /* + * Using malloc_strtoumax in this method isn't as handy as you might + * expect (I tried). In the fractional part, significant leading zeros + * mean that you still need to do your own parsing, now with trickier + * math. In the integer part, the casting (uintmax_t to uint32_t) + * forces more reasoning about bounds than just checking for overflow as + * we parse. + */ + uint32_t integer_part = 0; + + const char *cur = str; + + /* The string must start with a digit or a decimal point. */ + if (*cur != '.' && !fxp_isdigit(*cur)) { + return true; + } + + while ('0' <= *cur && *cur <= '9') { + integer_part *= 10; + integer_part += *cur - '0'; + if (integer_part >= (1U << 16)) { + return true; + } + cur++; + } + + /* + * We've parsed all digits at the beginning of the string, without + * overflow. Either we're done, or there's a fractional part. + */ + if (*cur != '.') { + *result = (integer_part << 16); + if (end != NULL) { + *end = (char *)cur; + } + return false; + } + + /* There's a fractional part. */ + cur++; + if (!fxp_isdigit(*cur)) { + /* Shouldn't end on the decimal point. */ + return true; + } + + /* + * We use a lot of precision for the fractional part, even though we'll + * discard most of it; this lets us get exact values for the important + * special case where the denominator is a small power of 2 (for + * instance, 1/512 == 0.001953125 is exactly representable even with + * only 16 bits of fractional precision). We need to left-shift by 16 + * before dividing so we pick the number of digits to be + * floor(log(2**48)) = 14. + */ + uint64_t fractional_part = 0; + uint64_t frac_div = 1; + for (int i = 0; i < FXP_FRACTIONAL_PART_DIGITS; i++) { + fractional_part *= 10; + frac_div *= 10; + if (fxp_isdigit(*cur)) { + fractional_part += *cur - '0'; + cur++; + } + } + /* + * We only parse the first maxdigits characters, but we can still ignore + * any digits after that. + */ + while (fxp_isdigit(*cur)) { + cur++; + } + + assert(fractional_part < frac_div); + uint32_t fractional_repr = (uint32_t)( + (fractional_part << 16) / frac_div); + + /* Success! */ + *result = (integer_part << 16) + fractional_repr; + if (end != NULL) { + *end = (char *)cur; + } + return false; +} + +void +fxp_print(fxp_t a, char buf[FXP_BUF_SIZE]) { + uint32_t integer_part = fxp_round_down(a); + uint32_t fractional_part = (a & ((1U << 16) - 1)); + + int leading_fraction_zeros = 0; + uint64_t fraction_digits = fractional_part; + for (int i = 0; i < FXP_FRACTIONAL_PART_DIGITS; i++) { + if (fraction_digits < (1U << 16) + && fraction_digits * 10 >= (1U << 16)) { + leading_fraction_zeros = i; + } + fraction_digits *= 10; + } + fraction_digits >>= 16; + while (fraction_digits > 0 && fraction_digits % 10 == 0) { + fraction_digits /= 10; + } + + size_t printed = malloc_snprintf(buf, FXP_BUF_SIZE, "%"FMTu32".", + integer_part); + for (int i = 0; i < leading_fraction_zeros; i++) { + buf[printed] = '0'; + printed++; + } + malloc_snprintf(&buf[printed], FXP_BUF_SIZE - printed, "%"FMTu64, + fraction_digits); +} diff --git a/contrib/jemalloc/src/hash.c b/contrib/jemalloc/src/hash.c deleted file mode 100644 --- a/contrib/jemalloc/src/hash.c +++ /dev/null @@ -1,3 +0,0 @@ -#define JEMALLOC_HASH_C_ -#include "jemalloc/internal/jemalloc_preamble.h" -#include "jemalloc/internal/jemalloc_internal_includes.h" diff --git a/contrib/jemalloc/src/hook.c b/contrib/jemalloc/src/hook.c --- a/contrib/jemalloc/src/hook.c +++ b/contrib/jemalloc/src/hook.c @@ -130,9 +130,9 @@ */ static bool in_hook_global = true; tsdn_t *tsdn = tsdn_fetch(); - tcache_t *tcache = tsdn_tcachep_get(tsdn); - if (tcache != NULL) { - return &tcache->in_hook; + bool *in_hook = tsdn_in_hookp_get(tsdn); + if (in_hook!= NULL) { + return in_hook; } return &in_hook_global; } diff --git a/contrib/jemalloc/src/hpa.c b/contrib/jemalloc/src/hpa.c new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/src/hpa.c @@ -0,0 +1,1044 @@ +#include "jemalloc/internal/jemalloc_preamble.h" +#include "jemalloc/internal/jemalloc_internal_includes.h" + +#include "jemalloc/internal/hpa.h" + +#include "jemalloc/internal/fb.h" +#include "jemalloc/internal/witness.h" + +#define HPA_EDEN_SIZE (128 * HUGEPAGE) + +static edata_t *hpa_alloc(tsdn_t *tsdn, pai_t *self, size_t size, + size_t alignment, bool zero, bool guarded, bool frequent_reuse, + bool *deferred_work_generated); +static size_t hpa_alloc_batch(tsdn_t *tsdn, pai_t *self, size_t size, + size_t nallocs, edata_list_active_t *results, bool *deferred_work_generated); +static bool hpa_expand(tsdn_t *tsdn, pai_t *self, edata_t *edata, + size_t old_size, size_t new_size, bool zero, bool *deferred_work_generated); +static bool hpa_shrink(tsdn_t *tsdn, pai_t *self, edata_t *edata, + size_t old_size, size_t new_size, bool *deferred_work_generated); +static void hpa_dalloc(tsdn_t *tsdn, pai_t *self, edata_t *edata, + bool *deferred_work_generated); +static void hpa_dalloc_batch(tsdn_t *tsdn, pai_t *self, + edata_list_active_t *list, bool *deferred_work_generated); +static uint64_t hpa_time_until_deferred_work(tsdn_t *tsdn, pai_t *self); + +bool +hpa_supported() { +#ifdef _WIN32 + /* + * At least until the API and implementation is somewhat settled, we + * don't want to try to debug the VM subsystem on the hardest-to-test + * platform. + */ + return false; +#endif + if (!pages_can_hugify) { + return false; + } + /* + * We fundamentally rely on a address-space-hungry growth strategy for + * hugepages. + */ + if (LG_SIZEOF_PTR != 3) { + return false; + } + /* + * If we couldn't detect the value of HUGEPAGE, HUGEPAGE_PAGES becomes + * this sentinel value -- see the comment in pages.h. + */ + if (HUGEPAGE_PAGES == 1) { + return false; + } + return true; +} + +static void +hpa_do_consistency_checks(hpa_shard_t *shard) { + assert(shard->base != NULL); +} + +bool +hpa_central_init(hpa_central_t *central, base_t *base, const hpa_hooks_t *hooks) { + /* malloc_conf processing should have filtered out these cases. */ + assert(hpa_supported()); + bool err; + err = malloc_mutex_init(¢ral->grow_mtx, "hpa_central_grow", + WITNESS_RANK_HPA_CENTRAL_GROW, malloc_mutex_rank_exclusive); + if (err) { + return true; + } + err = malloc_mutex_init(¢ral->mtx, "hpa_central", + WITNESS_RANK_HPA_CENTRAL, malloc_mutex_rank_exclusive); + if (err) { + return true; + } + central->base = base; + central->eden = NULL; + central->eden_len = 0; + central->age_counter = 0; + central->hooks = *hooks; + return false; +} + +static hpdata_t * +hpa_alloc_ps(tsdn_t *tsdn, hpa_central_t *central) { + return (hpdata_t *)base_alloc(tsdn, central->base, sizeof(hpdata_t), + CACHELINE); +} + +hpdata_t * +hpa_central_extract(tsdn_t *tsdn, hpa_central_t *central, size_t size, + bool *oom) { + /* Don't yet support big allocations; these should get filtered out. */ + assert(size <= HUGEPAGE); + /* + * Should only try to extract from the central allocator if the local + * shard is exhausted. We should hold the grow_mtx on that shard. + */ + witness_assert_positive_depth_to_rank( + tsdn_witness_tsdp_get(tsdn), WITNESS_RANK_HPA_SHARD_GROW); + + malloc_mutex_lock(tsdn, ¢ral->grow_mtx); + *oom = false; + + hpdata_t *ps = NULL; + + /* Is eden a perfect fit? */ + if (central->eden != NULL && central->eden_len == HUGEPAGE) { + ps = hpa_alloc_ps(tsdn, central); + if (ps == NULL) { + *oom = true; + malloc_mutex_unlock(tsdn, ¢ral->grow_mtx); + return NULL; + } + hpdata_init(ps, central->eden, central->age_counter++); + central->eden = NULL; + central->eden_len = 0; + malloc_mutex_unlock(tsdn, ¢ral->grow_mtx); + return ps; + } + + /* + * We're about to try to allocate from eden by splitting. If eden is + * NULL, we have to allocate it too. Otherwise, we just have to + * allocate an edata_t for the new psset. + */ + if (central->eden == NULL) { + /* + * During development, we're primarily concerned with systems + * with overcommit. Eventually, we should be more careful here. + */ + bool commit = true; + /* Allocate address space, bailing if we fail. */ + void *new_eden = pages_map(NULL, HPA_EDEN_SIZE, HUGEPAGE, + &commit); + if (new_eden == NULL) { + *oom = true; + malloc_mutex_unlock(tsdn, ¢ral->grow_mtx); + return NULL; + } + ps = hpa_alloc_ps(tsdn, central); + if (ps == NULL) { + pages_unmap(new_eden, HPA_EDEN_SIZE); + *oom = true; + malloc_mutex_unlock(tsdn, ¢ral->grow_mtx); + return NULL; + } + central->eden = new_eden; + central->eden_len = HPA_EDEN_SIZE; + } else { + /* Eden is already nonempty; only need an edata for ps. */ + ps = hpa_alloc_ps(tsdn, central); + if (ps == NULL) { + *oom = true; + malloc_mutex_unlock(tsdn, ¢ral->grow_mtx); + return NULL; + } + } + assert(ps != NULL); + assert(central->eden != NULL); + assert(central->eden_len > HUGEPAGE); + assert(central->eden_len % HUGEPAGE == 0); + assert(HUGEPAGE_ADDR2BASE(central->eden) == central->eden); + + hpdata_init(ps, central->eden, central->age_counter++); + + char *eden_char = (char *)central->eden; + eden_char += HUGEPAGE; + central->eden = (void *)eden_char; + central->eden_len -= HUGEPAGE; + + malloc_mutex_unlock(tsdn, ¢ral->grow_mtx); + + return ps; +} + +bool +hpa_shard_init(hpa_shard_t *shard, hpa_central_t *central, emap_t *emap, + base_t *base, edata_cache_t *edata_cache, unsigned ind, + const hpa_shard_opts_t *opts) { + /* malloc_conf processing should have filtered out these cases. */ + assert(hpa_supported()); + bool err; + err = malloc_mutex_init(&shard->grow_mtx, "hpa_shard_grow", + WITNESS_RANK_HPA_SHARD_GROW, malloc_mutex_rank_exclusive); + if (err) { + return true; + } + err = malloc_mutex_init(&shard->mtx, "hpa_shard", + WITNESS_RANK_HPA_SHARD, malloc_mutex_rank_exclusive); + if (err) { + return true; + } + + assert(edata_cache != NULL); + shard->central = central; + shard->base = base; + edata_cache_fast_init(&shard->ecf, edata_cache); + psset_init(&shard->psset); + shard->age_counter = 0; + shard->ind = ind; + shard->emap = emap; + + shard->opts = *opts; + + shard->npending_purge = 0; + nstime_init_zero(&shard->last_purge); + + shard->stats.npurge_passes = 0; + shard->stats.npurges = 0; + shard->stats.nhugifies = 0; + shard->stats.ndehugifies = 0; + + /* + * Fill these in last, so that if an hpa_shard gets used despite + * initialization failing, we'll at least crash instead of just + * operating on corrupted data. + */ + shard->pai.alloc = &hpa_alloc; + shard->pai.alloc_batch = &hpa_alloc_batch; + shard->pai.expand = &hpa_expand; + shard->pai.shrink = &hpa_shrink; + shard->pai.dalloc = &hpa_dalloc; + shard->pai.dalloc_batch = &hpa_dalloc_batch; + shard->pai.time_until_deferred_work = &hpa_time_until_deferred_work; + + hpa_do_consistency_checks(shard); + + return false; +} + +/* + * Note that the stats functions here follow the usual stats naming conventions; + * "merge" obtains the stats from some live object of instance, while "accum" + * only combines the stats from one stats objet to another. Hence the lack of + * locking here. + */ +static void +hpa_shard_nonderived_stats_accum(hpa_shard_nonderived_stats_t *dst, + hpa_shard_nonderived_stats_t *src) { + dst->npurge_passes += src->npurge_passes; + dst->npurges += src->npurges; + dst->nhugifies += src->nhugifies; + dst->ndehugifies += src->ndehugifies; +} + +void +hpa_shard_stats_accum(hpa_shard_stats_t *dst, hpa_shard_stats_t *src) { + psset_stats_accum(&dst->psset_stats, &src->psset_stats); + hpa_shard_nonderived_stats_accum(&dst->nonderived_stats, + &src->nonderived_stats); +} + +void +hpa_shard_stats_merge(tsdn_t *tsdn, hpa_shard_t *shard, + hpa_shard_stats_t *dst) { + hpa_do_consistency_checks(shard); + + malloc_mutex_lock(tsdn, &shard->grow_mtx); + malloc_mutex_lock(tsdn, &shard->mtx); + psset_stats_accum(&dst->psset_stats, &shard->psset.stats); + hpa_shard_nonderived_stats_accum(&dst->nonderived_stats, &shard->stats); + malloc_mutex_unlock(tsdn, &shard->mtx); + malloc_mutex_unlock(tsdn, &shard->grow_mtx); +} + +static bool +hpa_good_hugification_candidate(hpa_shard_t *shard, hpdata_t *ps) { + /* + * Note that this needs to be >= rather than just >, because of the + * important special case in which the hugification threshold is exactly + * HUGEPAGE. + */ + return hpdata_nactive_get(ps) * PAGE + >= shard->opts.hugification_threshold; +} + +static size_t +hpa_adjusted_ndirty(tsdn_t *tsdn, hpa_shard_t *shard) { + malloc_mutex_assert_owner(tsdn, &shard->mtx); + return psset_ndirty(&shard->psset) - shard->npending_purge; +} + +static size_t +hpa_ndirty_max(tsdn_t *tsdn, hpa_shard_t *shard) { + malloc_mutex_assert_owner(tsdn, &shard->mtx); + if (shard->opts.dirty_mult == (fxp_t)-1) { + return (size_t)-1; + } + return fxp_mul_frac(psset_nactive(&shard->psset), + shard->opts.dirty_mult); +} + +static bool +hpa_hugify_blocked_by_ndirty(tsdn_t *tsdn, hpa_shard_t *shard) { + malloc_mutex_assert_owner(tsdn, &shard->mtx); + hpdata_t *to_hugify = psset_pick_hugify(&shard->psset); + if (to_hugify == NULL) { + return false; + } + return hpa_adjusted_ndirty(tsdn, shard) + + hpdata_nretained_get(to_hugify) > hpa_ndirty_max(tsdn, shard); +} + +static bool +hpa_should_purge(tsdn_t *tsdn, hpa_shard_t *shard) { + malloc_mutex_assert_owner(tsdn, &shard->mtx); + if (hpa_adjusted_ndirty(tsdn, shard) > hpa_ndirty_max(tsdn, shard)) { + return true; + } + if (hpa_hugify_blocked_by_ndirty(tsdn, shard)) { + return true; + } + return false; +} + +static void +hpa_update_purge_hugify_eligibility(tsdn_t *tsdn, hpa_shard_t *shard, + hpdata_t *ps) { + malloc_mutex_assert_owner(tsdn, &shard->mtx); + if (hpdata_changing_state_get(ps)) { + hpdata_purge_allowed_set(ps, false); + hpdata_disallow_hugify(ps); + return; + } + /* + * Hugepages are distinctly costly to purge, so try to avoid it unless + * they're *particularly* full of dirty pages. Eventually, we should + * use a smarter / more dynamic heuristic for situations where we have + * to manually hugify. + * + * In situations where we don't manually hugify, this problem is + * reduced. The "bad" situation we're trying to avoid is one's that's + * common in some Linux configurations (where both enabled and defrag + * are set to madvise) that can lead to long latency spikes on the first + * access after a hugification. The ideal policy in such configurations + * is probably time-based for both purging and hugifying; only hugify a + * hugepage if it's met the criteria for some extended period of time, + * and only dehugify it if it's failed to meet the criteria for an + * extended period of time. When background threads are on, we should + * try to take this hit on one of them, as well. + * + * I think the ideal setting is THP always enabled, and defrag set to + * deferred; in that case we don't need any explicit calls on the + * allocator's end at all; we just try to pack allocations in a + * hugepage-friendly manner and let the OS hugify in the background. + */ + hpdata_purge_allowed_set(ps, hpdata_ndirty_get(ps) > 0); + if (hpa_good_hugification_candidate(shard, ps) + && !hpdata_huge_get(ps)) { + nstime_t now; + shard->central->hooks.curtime(&now, /* first_reading */ true); + hpdata_allow_hugify(ps, now); + } + /* + * Once a hugepage has become eligible for hugification, we don't mark + * it as ineligible just because it stops meeting the criteria (this + * could lead to situations where a hugepage that spends most of its + * time meeting the criteria never quite getting hugified if there are + * intervening deallocations). The idea is that the hugification delay + * will allow them to get purged, reseting their "hugify-allowed" bit. + * If they don't get purged, then the hugification isn't hurting and + * might help. As an exception, we don't hugify hugepages that are now + * empty; it definitely doesn't help there until the hugepage gets + * reused, which is likely not for a while. + */ + if (hpdata_nactive_get(ps) == 0) { + hpdata_disallow_hugify(ps); + } +} + +static bool +hpa_shard_has_deferred_work(tsdn_t *tsdn, hpa_shard_t *shard) { + malloc_mutex_assert_owner(tsdn, &shard->mtx); + hpdata_t *to_hugify = psset_pick_hugify(&shard->psset); + return to_hugify != NULL || hpa_should_purge(tsdn, shard); +} + +/* Returns whether or not we purged anything. */ +static bool +hpa_try_purge(tsdn_t *tsdn, hpa_shard_t *shard) { + malloc_mutex_assert_owner(tsdn, &shard->mtx); + + hpdata_t *to_purge = psset_pick_purge(&shard->psset); + if (to_purge == NULL) { + return false; + } + assert(hpdata_purge_allowed_get(to_purge)); + assert(!hpdata_changing_state_get(to_purge)); + + /* + * Don't let anyone else purge or hugify this page while + * we're purging it (allocations and deallocations are + * OK). + */ + psset_update_begin(&shard->psset, to_purge); + assert(hpdata_alloc_allowed_get(to_purge)); + hpdata_mid_purge_set(to_purge, true); + hpdata_purge_allowed_set(to_purge, false); + hpdata_disallow_hugify(to_purge); + /* + * Unlike with hugification (where concurrent + * allocations are allowed), concurrent allocation out + * of a hugepage being purged is unsafe; we might hand + * out an extent for an allocation and then purge it + * (clearing out user data). + */ + hpdata_alloc_allowed_set(to_purge, false); + psset_update_end(&shard->psset, to_purge); + + /* Gather all the metadata we'll need during the purge. */ + bool dehugify = hpdata_huge_get(to_purge); + hpdata_purge_state_t purge_state; + size_t num_to_purge = hpdata_purge_begin(to_purge, &purge_state); + + shard->npending_purge += num_to_purge; + + malloc_mutex_unlock(tsdn, &shard->mtx); + + /* Actually do the purging, now that the lock is dropped. */ + if (dehugify) { + shard->central->hooks.dehugify(hpdata_addr_get(to_purge), + HUGEPAGE); + } + size_t total_purged = 0; + uint64_t purges_this_pass = 0; + void *purge_addr; + size_t purge_size; + while (hpdata_purge_next(to_purge, &purge_state, &purge_addr, + &purge_size)) { + total_purged += purge_size; + assert(total_purged <= HUGEPAGE); + purges_this_pass++; + shard->central->hooks.purge(purge_addr, purge_size); + } + + malloc_mutex_lock(tsdn, &shard->mtx); + /* The shard updates */ + shard->npending_purge -= num_to_purge; + shard->stats.npurge_passes++; + shard->stats.npurges += purges_this_pass; + shard->central->hooks.curtime(&shard->last_purge, + /* first_reading */ false); + if (dehugify) { + shard->stats.ndehugifies++; + } + + /* The hpdata updates. */ + psset_update_begin(&shard->psset, to_purge); + if (dehugify) { + hpdata_dehugify(to_purge); + } + hpdata_purge_end(to_purge, &purge_state); + hpdata_mid_purge_set(to_purge, false); + + hpdata_alloc_allowed_set(to_purge, true); + hpa_update_purge_hugify_eligibility(tsdn, shard, to_purge); + + psset_update_end(&shard->psset, to_purge); + + return true; +} + +/* Returns whether or not we hugified anything. */ +static bool +hpa_try_hugify(tsdn_t *tsdn, hpa_shard_t *shard) { + malloc_mutex_assert_owner(tsdn, &shard->mtx); + + if (hpa_hugify_blocked_by_ndirty(tsdn, shard)) { + return false; + } + + hpdata_t *to_hugify = psset_pick_hugify(&shard->psset); + if (to_hugify == NULL) { + return false; + } + assert(hpdata_hugify_allowed_get(to_hugify)); + assert(!hpdata_changing_state_get(to_hugify)); + + /* Make sure that it's been hugifiable for long enough. */ + nstime_t time_hugify_allowed = hpdata_time_hugify_allowed(to_hugify); + uint64_t millis = shard->central->hooks.ms_since(&time_hugify_allowed); + if (millis < shard->opts.hugify_delay_ms) { + return false; + } + + /* + * Don't let anyone else purge or hugify this page while + * we're hugifying it (allocations and deallocations are + * OK). + */ + psset_update_begin(&shard->psset, to_hugify); + hpdata_mid_hugify_set(to_hugify, true); + hpdata_purge_allowed_set(to_hugify, false); + hpdata_disallow_hugify(to_hugify); + assert(hpdata_alloc_allowed_get(to_hugify)); + psset_update_end(&shard->psset, to_hugify); + + malloc_mutex_unlock(tsdn, &shard->mtx); + + shard->central->hooks.hugify(hpdata_addr_get(to_hugify), HUGEPAGE); + + malloc_mutex_lock(tsdn, &shard->mtx); + shard->stats.nhugifies++; + + psset_update_begin(&shard->psset, to_hugify); + hpdata_hugify(to_hugify); + hpdata_mid_hugify_set(to_hugify, false); + hpa_update_purge_hugify_eligibility(tsdn, shard, to_hugify); + psset_update_end(&shard->psset, to_hugify); + + return true; +} + +/* + * Execution of deferred work is forced if it's triggered by an explicit + * hpa_shard_do_deferred_work() call. + */ +static void +hpa_shard_maybe_do_deferred_work(tsdn_t *tsdn, hpa_shard_t *shard, + bool forced) { + malloc_mutex_assert_owner(tsdn, &shard->mtx); + if (!forced && shard->opts.deferral_allowed) { + return; + } + /* + * If we're on a background thread, do work so long as there's work to + * be done. Otherwise, bound latency to not be *too* bad by doing at + * most a small fixed number of operations. + */ + bool hugified = false; + bool purged = false; + size_t max_ops = (forced ? (size_t)-1 : 16); + size_t nops = 0; + do { + /* + * Always purge before hugifying, to make sure we get some + * ability to hit our quiescence targets. + */ + purged = false; + while (hpa_should_purge(tsdn, shard) && nops < max_ops) { + purged = hpa_try_purge(tsdn, shard); + if (purged) { + nops++; + } + } + hugified = hpa_try_hugify(tsdn, shard); + if (hugified) { + nops++; + } + malloc_mutex_assert_owner(tsdn, &shard->mtx); + malloc_mutex_assert_owner(tsdn, &shard->mtx); + } while ((hugified || purged) && nops < max_ops); +} + +static edata_t * +hpa_try_alloc_one_no_grow(tsdn_t *tsdn, hpa_shard_t *shard, size_t size, + bool *oom) { + bool err; + edata_t *edata = edata_cache_fast_get(tsdn, &shard->ecf); + if (edata == NULL) { + *oom = true; + return NULL; + } + + hpdata_t *ps = psset_pick_alloc(&shard->psset, size); + if (ps == NULL) { + edata_cache_fast_put(tsdn, &shard->ecf, edata); + return NULL; + } + + psset_update_begin(&shard->psset, ps); + + if (hpdata_empty(ps)) { + /* + * If the pageslab used to be empty, treat it as though it's + * brand new for fragmentation-avoidance purposes; what we're + * trying to approximate is the age of the allocations *in* that + * pageslab, and the allocations in the new pageslab are + * definitionally the youngest in this hpa shard. + */ + hpdata_age_set(ps, shard->age_counter++); + } + + void *addr = hpdata_reserve_alloc(ps, size); + edata_init(edata, shard->ind, addr, size, /* slab */ false, + SC_NSIZES, /* sn */ hpdata_age_get(ps), extent_state_active, + /* zeroed */ false, /* committed */ true, EXTENT_PAI_HPA, + EXTENT_NOT_HEAD); + edata_ps_set(edata, ps); + + /* + * This could theoretically be moved outside of the critical section, + * but that introduces the potential for a race. Without the lock, the + * (initially nonempty, since this is the reuse pathway) pageslab we + * allocated out of could become otherwise empty while the lock is + * dropped. This would force us to deal with a pageslab eviction down + * the error pathway, which is a pain. + */ + err = emap_register_boundary(tsdn, shard->emap, edata, + SC_NSIZES, /* slab */ false); + if (err) { + hpdata_unreserve(ps, edata_addr_get(edata), + edata_size_get(edata)); + /* + * We should arguably reset dirty state here, but this would + * require some sort of prepare + commit functionality that's a + * little much to deal with for now. + * + * We don't have a do_deferred_work down this pathway, on the + * principle that we didn't *really* affect shard state (we + * tweaked the stats, but our tweaks weren't really accurate). + */ + psset_update_end(&shard->psset, ps); + edata_cache_fast_put(tsdn, &shard->ecf, edata); + *oom = true; + return NULL; + } + + hpa_update_purge_hugify_eligibility(tsdn, shard, ps); + psset_update_end(&shard->psset, ps); + return edata; +} + +static size_t +hpa_try_alloc_batch_no_grow(tsdn_t *tsdn, hpa_shard_t *shard, size_t size, + bool *oom, size_t nallocs, edata_list_active_t *results, + bool *deferred_work_generated) { + malloc_mutex_lock(tsdn, &shard->mtx); + size_t nsuccess = 0; + for (; nsuccess < nallocs; nsuccess++) { + edata_t *edata = hpa_try_alloc_one_no_grow(tsdn, shard, size, + oom); + if (edata == NULL) { + break; + } + edata_list_active_append(results, edata); + } + + hpa_shard_maybe_do_deferred_work(tsdn, shard, /* forced */ false); + *deferred_work_generated = hpa_shard_has_deferred_work(tsdn, shard); + malloc_mutex_unlock(tsdn, &shard->mtx); + return nsuccess; +} + +static size_t +hpa_alloc_batch_psset(tsdn_t *tsdn, hpa_shard_t *shard, size_t size, + size_t nallocs, edata_list_active_t *results, + bool *deferred_work_generated) { + assert(size <= shard->opts.slab_max_alloc); + bool oom = false; + + size_t nsuccess = hpa_try_alloc_batch_no_grow(tsdn, shard, size, &oom, + nallocs, results, deferred_work_generated); + + if (nsuccess == nallocs || oom) { + return nsuccess; + } + + /* + * We didn't OOM, but weren't able to fill everything requested of us; + * try to grow. + */ + malloc_mutex_lock(tsdn, &shard->grow_mtx); + /* + * Check for grow races; maybe some earlier thread expanded the psset + * in between when we dropped the main mutex and grabbed the grow mutex. + */ + nsuccess += hpa_try_alloc_batch_no_grow(tsdn, shard, size, &oom, + nallocs - nsuccess, results, deferred_work_generated); + if (nsuccess == nallocs || oom) { + malloc_mutex_unlock(tsdn, &shard->grow_mtx); + return nsuccess; + } + + /* + * Note that we don't hold shard->mtx here (while growing); + * deallocations (and allocations of smaller sizes) may still succeed + * while we're doing this potentially expensive system call. + */ + hpdata_t *ps = hpa_central_extract(tsdn, shard->central, size, &oom); + if (ps == NULL) { + malloc_mutex_unlock(tsdn, &shard->grow_mtx); + return nsuccess; + } + + /* + * We got the pageslab; allocate from it. This does an unlock followed + * by a lock on the same mutex, and holds the grow mutex while doing + * deferred work, but this is an uncommon path; the simplicity is worth + * it. + */ + malloc_mutex_lock(tsdn, &shard->mtx); + psset_insert(&shard->psset, ps); + malloc_mutex_unlock(tsdn, &shard->mtx); + + nsuccess += hpa_try_alloc_batch_no_grow(tsdn, shard, size, &oom, + nallocs - nsuccess, results, deferred_work_generated); + /* + * Drop grow_mtx before doing deferred work; other threads blocked on it + * should be allowed to proceed while we're working. + */ + malloc_mutex_unlock(tsdn, &shard->grow_mtx); + + return nsuccess; +} + +static hpa_shard_t * +hpa_from_pai(pai_t *self) { + assert(self->alloc = &hpa_alloc); + assert(self->expand = &hpa_expand); + assert(self->shrink = &hpa_shrink); + assert(self->dalloc = &hpa_dalloc); + return (hpa_shard_t *)self; +} + +static size_t +hpa_alloc_batch(tsdn_t *tsdn, pai_t *self, size_t size, size_t nallocs, + edata_list_active_t *results, bool *deferred_work_generated) { + assert(nallocs > 0); + assert((size & PAGE_MASK) == 0); + witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), + WITNESS_RANK_CORE, 0); + hpa_shard_t *shard = hpa_from_pai(self); + + if (size > shard->opts.slab_max_alloc) { + return 0; + } + + size_t nsuccess = hpa_alloc_batch_psset(tsdn, shard, size, nallocs, + results, deferred_work_generated); + + witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), + WITNESS_RANK_CORE, 0); + + /* + * Guard the sanity checks with config_debug because the loop cannot be + * proven non-circular by the compiler, even if everything within the + * loop is optimized away. + */ + if (config_debug) { + edata_t *edata; + ql_foreach(edata, &results->head, ql_link_active) { + emap_assert_mapped(tsdn, shard->emap, edata); + assert(edata_pai_get(edata) == EXTENT_PAI_HPA); + assert(edata_state_get(edata) == extent_state_active); + assert(edata_arena_ind_get(edata) == shard->ind); + assert(edata_szind_get_maybe_invalid(edata) == + SC_NSIZES); + assert(!edata_slab_get(edata)); + assert(edata_committed_get(edata)); + assert(edata_base_get(edata) == edata_addr_get(edata)); + assert(edata_base_get(edata) != NULL); + } + } + return nsuccess; +} + +static edata_t * +hpa_alloc(tsdn_t *tsdn, pai_t *self, size_t size, size_t alignment, bool zero, + bool guarded, bool frequent_reuse, bool *deferred_work_generated) { + assert((size & PAGE_MASK) == 0); + assert(!guarded); + witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), + WITNESS_RANK_CORE, 0); + + /* We don't handle alignment or zeroing for now. */ + if (alignment > PAGE || zero) { + return NULL; + } + /* + * An alloc with alignment == PAGE and zero == false is equivalent to a + * batch alloc of 1. Just do that, so we can share code. + */ + edata_list_active_t results; + edata_list_active_init(&results); + size_t nallocs = hpa_alloc_batch(tsdn, self, size, /* nallocs */ 1, + &results, deferred_work_generated); + assert(nallocs == 0 || nallocs == 1); + edata_t *edata = edata_list_active_first(&results); + return edata; +} + +static bool +hpa_expand(tsdn_t *tsdn, pai_t *self, edata_t *edata, size_t old_size, + size_t new_size, bool zero, bool *deferred_work_generated) { + /* Expand not yet supported. */ + return true; +} + +static bool +hpa_shrink(tsdn_t *tsdn, pai_t *self, edata_t *edata, + size_t old_size, size_t new_size, bool *deferred_work_generated) { + /* Shrink not yet supported. */ + return true; +} + +static void +hpa_dalloc_prepare_unlocked(tsdn_t *tsdn, hpa_shard_t *shard, edata_t *edata) { + malloc_mutex_assert_not_owner(tsdn, &shard->mtx); + + assert(edata_pai_get(edata) == EXTENT_PAI_HPA); + assert(edata_state_get(edata) == extent_state_active); + assert(edata_arena_ind_get(edata) == shard->ind); + assert(edata_szind_get_maybe_invalid(edata) == SC_NSIZES); + assert(edata_committed_get(edata)); + assert(edata_base_get(edata) != NULL); + + /* + * Another thread shouldn't be trying to touch the metadata of an + * allocation being freed. The one exception is a merge attempt from a + * lower-addressed PAC extent; in this case we have a nominal race on + * the edata metadata bits, but in practice the fact that the PAI bits + * are different will prevent any further access. The race is bad, but + * benign in practice, and the long term plan is to track enough state + * in the rtree to prevent these merge attempts in the first place. + */ + edata_addr_set(edata, edata_base_get(edata)); + edata_zeroed_set(edata, false); + emap_deregister_boundary(tsdn, shard->emap, edata); +} + +static void +hpa_dalloc_locked(tsdn_t *tsdn, hpa_shard_t *shard, edata_t *edata) { + malloc_mutex_assert_owner(tsdn, &shard->mtx); + + /* + * Release the metadata early, to avoid having to remember to do it + * while we're also doing tricky purging logic. First, we need to grab + * a few bits of metadata from it. + * + * Note that the shard mutex protects ps's metadata too; it wouldn't be + * correct to try to read most information out of it without the lock. + */ + hpdata_t *ps = edata_ps_get(edata); + /* Currently, all edatas come from pageslabs. */ + assert(ps != NULL); + void *unreserve_addr = edata_addr_get(edata); + size_t unreserve_size = edata_size_get(edata); + edata_cache_fast_put(tsdn, &shard->ecf, edata); + + psset_update_begin(&shard->psset, ps); + hpdata_unreserve(ps, unreserve_addr, unreserve_size); + hpa_update_purge_hugify_eligibility(tsdn, shard, ps); + psset_update_end(&shard->psset, ps); +} + +static void +hpa_dalloc_batch(tsdn_t *tsdn, pai_t *self, edata_list_active_t *list, + bool *deferred_work_generated) { + hpa_shard_t *shard = hpa_from_pai(self); + + edata_t *edata; + ql_foreach(edata, &list->head, ql_link_active) { + hpa_dalloc_prepare_unlocked(tsdn, shard, edata); + } + + malloc_mutex_lock(tsdn, &shard->mtx); + /* Now, remove from the list. */ + while ((edata = edata_list_active_first(list)) != NULL) { + edata_list_active_remove(list, edata); + hpa_dalloc_locked(tsdn, shard, edata); + } + hpa_shard_maybe_do_deferred_work(tsdn, shard, /* forced */ false); + *deferred_work_generated = + hpa_shard_has_deferred_work(tsdn, shard); + + malloc_mutex_unlock(tsdn, &shard->mtx); +} + +static void +hpa_dalloc(tsdn_t *tsdn, pai_t *self, edata_t *edata, + bool *deferred_work_generated) { + assert(!edata_guarded_get(edata)); + /* Just a dalloc_batch of size 1; this lets us share logic. */ + edata_list_active_t dalloc_list; + edata_list_active_init(&dalloc_list); + edata_list_active_append(&dalloc_list, edata); + hpa_dalloc_batch(tsdn, self, &dalloc_list, deferred_work_generated); +} + +/* + * Calculate time until either purging or hugification ought to happen. + * Called by background threads. + */ +static uint64_t +hpa_time_until_deferred_work(tsdn_t *tsdn, pai_t *self) { + hpa_shard_t *shard = hpa_from_pai(self); + uint64_t time_ns = BACKGROUND_THREAD_DEFERRED_MAX; + + malloc_mutex_lock(tsdn, &shard->mtx); + + hpdata_t *to_hugify = psset_pick_hugify(&shard->psset); + if (to_hugify != NULL) { + nstime_t time_hugify_allowed = + hpdata_time_hugify_allowed(to_hugify); + uint64_t since_hugify_allowed_ms = + shard->central->hooks.ms_since(&time_hugify_allowed); + /* + * If not enough time has passed since hugification was allowed, + * sleep for the rest. + */ + if (since_hugify_allowed_ms < shard->opts.hugify_delay_ms) { + time_ns = shard->opts.hugify_delay_ms - + since_hugify_allowed_ms; + time_ns *= 1000 * 1000; + } else { + malloc_mutex_unlock(tsdn, &shard->mtx); + return BACKGROUND_THREAD_DEFERRED_MIN; + } + } + + if (hpa_should_purge(tsdn, shard)) { + /* + * If we haven't purged before, no need to check interval + * between purges. Simply purge as soon as possible. + */ + if (shard->stats.npurge_passes == 0) { + malloc_mutex_unlock(tsdn, &shard->mtx); + return BACKGROUND_THREAD_DEFERRED_MIN; + } + uint64_t since_last_purge_ms = shard->central->hooks.ms_since( + &shard->last_purge); + + if (since_last_purge_ms < shard->opts.min_purge_interval_ms) { + uint64_t until_purge_ns; + until_purge_ns = shard->opts.min_purge_interval_ms - + since_last_purge_ms; + until_purge_ns *= 1000 * 1000; + + if (until_purge_ns < time_ns) { + time_ns = until_purge_ns; + } + } else { + time_ns = BACKGROUND_THREAD_DEFERRED_MIN; + } + } + malloc_mutex_unlock(tsdn, &shard->mtx); + return time_ns; +} + +void +hpa_shard_disable(tsdn_t *tsdn, hpa_shard_t *shard) { + hpa_do_consistency_checks(shard); + + malloc_mutex_lock(tsdn, &shard->mtx); + edata_cache_fast_disable(tsdn, &shard->ecf); + malloc_mutex_unlock(tsdn, &shard->mtx); +} + +static void +hpa_shard_assert_stats_empty(psset_bin_stats_t *bin_stats) { + assert(bin_stats->npageslabs == 0); + assert(bin_stats->nactive == 0); +} + +static void +hpa_assert_empty(tsdn_t *tsdn, hpa_shard_t *shard, psset_t *psset) { + malloc_mutex_assert_owner(tsdn, &shard->mtx); + for (int huge = 0; huge <= 1; huge++) { + hpa_shard_assert_stats_empty(&psset->stats.full_slabs[huge]); + for (pszind_t i = 0; i < PSSET_NPSIZES; i++) { + hpa_shard_assert_stats_empty( + &psset->stats.nonfull_slabs[i][huge]); + } + } +} + +void +hpa_shard_destroy(tsdn_t *tsdn, hpa_shard_t *shard) { + hpa_do_consistency_checks(shard); + /* + * By the time we're here, the arena code should have dalloc'd all the + * active extents, which means we should have eventually evicted + * everything from the psset, so it shouldn't be able to serve even a + * 1-page allocation. + */ + if (config_debug) { + malloc_mutex_lock(tsdn, &shard->mtx); + hpa_assert_empty(tsdn, shard, &shard->psset); + malloc_mutex_unlock(tsdn, &shard->mtx); + } + hpdata_t *ps; + while ((ps = psset_pick_alloc(&shard->psset, PAGE)) != NULL) { + /* There should be no allocations anywhere. */ + assert(hpdata_empty(ps)); + psset_remove(&shard->psset, ps); + shard->central->hooks.unmap(hpdata_addr_get(ps), HUGEPAGE); + } +} + +void +hpa_shard_set_deferral_allowed(tsdn_t *tsdn, hpa_shard_t *shard, + bool deferral_allowed) { + hpa_do_consistency_checks(shard); + + malloc_mutex_lock(tsdn, &shard->mtx); + bool deferral_previously_allowed = shard->opts.deferral_allowed; + shard->opts.deferral_allowed = deferral_allowed; + if (deferral_previously_allowed && !deferral_allowed) { + hpa_shard_maybe_do_deferred_work(tsdn, shard, + /* forced */ true); + } + malloc_mutex_unlock(tsdn, &shard->mtx); +} + +void +hpa_shard_do_deferred_work(tsdn_t *tsdn, hpa_shard_t *shard) { + hpa_do_consistency_checks(shard); + + malloc_mutex_lock(tsdn, &shard->mtx); + hpa_shard_maybe_do_deferred_work(tsdn, shard, /* forced */ true); + malloc_mutex_unlock(tsdn, &shard->mtx); +} + +void +hpa_shard_prefork3(tsdn_t *tsdn, hpa_shard_t *shard) { + hpa_do_consistency_checks(shard); + + malloc_mutex_prefork(tsdn, &shard->grow_mtx); +} + +void +hpa_shard_prefork4(tsdn_t *tsdn, hpa_shard_t *shard) { + hpa_do_consistency_checks(shard); + + malloc_mutex_prefork(tsdn, &shard->mtx); +} + +void +hpa_shard_postfork_parent(tsdn_t *tsdn, hpa_shard_t *shard) { + hpa_do_consistency_checks(shard); + + malloc_mutex_postfork_parent(tsdn, &shard->grow_mtx); + malloc_mutex_postfork_parent(tsdn, &shard->mtx); +} + +void +hpa_shard_postfork_child(tsdn_t *tsdn, hpa_shard_t *shard) { + hpa_do_consistency_checks(shard); + + malloc_mutex_postfork_child(tsdn, &shard->grow_mtx); + malloc_mutex_postfork_child(tsdn, &shard->mtx); +} diff --git a/contrib/jemalloc/src/hpa_hooks.c b/contrib/jemalloc/src/hpa_hooks.c new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/src/hpa_hooks.c @@ -0,0 +1,63 @@ +#include "jemalloc/internal/jemalloc_preamble.h" +#include "jemalloc/internal/jemalloc_internal_includes.h" + +#include "jemalloc/internal/hpa_hooks.h" + +static void *hpa_hooks_map(size_t size); +static void hpa_hooks_unmap(void *ptr, size_t size); +static void hpa_hooks_purge(void *ptr, size_t size); +static void hpa_hooks_hugify(void *ptr, size_t size); +static void hpa_hooks_dehugify(void *ptr, size_t size); +static void hpa_hooks_curtime(nstime_t *r_nstime, bool first_reading); +static uint64_t hpa_hooks_ms_since(nstime_t *past_nstime); + +hpa_hooks_t hpa_hooks_default = { + &hpa_hooks_map, + &hpa_hooks_unmap, + &hpa_hooks_purge, + &hpa_hooks_hugify, + &hpa_hooks_dehugify, + &hpa_hooks_curtime, + &hpa_hooks_ms_since +}; + +static void * +hpa_hooks_map(size_t size) { + bool commit = true; + return pages_map(NULL, size, HUGEPAGE, &commit); +} + +static void +hpa_hooks_unmap(void *ptr, size_t size) { + pages_unmap(ptr, size); +} + +static void +hpa_hooks_purge(void *ptr, size_t size) { + pages_purge_forced(ptr, size); +} + +static void +hpa_hooks_hugify(void *ptr, size_t size) { + bool err = pages_huge(ptr, size); + (void)err; +} + +static void +hpa_hooks_dehugify(void *ptr, size_t size) { + bool err = pages_nohuge(ptr, size); + (void)err; +} + +static void +hpa_hooks_curtime(nstime_t *r_nstime, bool first_reading) { + if (first_reading) { + nstime_init_zero(r_nstime); + } + nstime_update(r_nstime); +} + +static uint64_t +hpa_hooks_ms_since(nstime_t *past_nstime) { + return nstime_ns_since(past_nstime) / 1000 / 1000; +} diff --git a/contrib/jemalloc/src/hpdata.c b/contrib/jemalloc/src/hpdata.c new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/src/hpdata.c @@ -0,0 +1,325 @@ +#include "jemalloc/internal/jemalloc_preamble.h" +#include "jemalloc/internal/jemalloc_internal_includes.h" + +#include "jemalloc/internal/hpdata.h" + +static int +hpdata_age_comp(const hpdata_t *a, const hpdata_t *b) { + uint64_t a_age = hpdata_age_get(a); + uint64_t b_age = hpdata_age_get(b); + /* + * hpdata ages are operation counts in the psset; no two should be the + * same. + */ + assert(a_age != b_age); + return (a_age > b_age) - (a_age < b_age); +} + +ph_gen(, hpdata_age_heap, hpdata_t, age_link, hpdata_age_comp) + +void +hpdata_init(hpdata_t *hpdata, void *addr, uint64_t age) { + hpdata_addr_set(hpdata, addr); + hpdata_age_set(hpdata, age); + hpdata->h_huge = false; + hpdata->h_alloc_allowed = true; + hpdata->h_in_psset_alloc_container = false; + hpdata->h_purge_allowed = false; + hpdata->h_hugify_allowed = false; + hpdata->h_in_psset_hugify_container = false; + hpdata->h_mid_purge = false; + hpdata->h_mid_hugify = false; + hpdata->h_updating = false; + hpdata->h_in_psset = false; + hpdata_longest_free_range_set(hpdata, HUGEPAGE_PAGES); + hpdata->h_nactive = 0; + fb_init(hpdata->active_pages, HUGEPAGE_PAGES); + hpdata->h_ntouched = 0; + fb_init(hpdata->touched_pages, HUGEPAGE_PAGES); + + hpdata_assert_consistent(hpdata); +} + +void * +hpdata_reserve_alloc(hpdata_t *hpdata, size_t sz) { + hpdata_assert_consistent(hpdata); + /* + * This is a metadata change; the hpdata should therefore either not be + * in the psset, or should have explicitly marked itself as being + * mid-update. + */ + assert(!hpdata->h_in_psset || hpdata->h_updating); + assert(hpdata->h_alloc_allowed); + assert((sz & PAGE_MASK) == 0); + size_t npages = sz >> LG_PAGE; + assert(npages <= hpdata_longest_free_range_get(hpdata)); + + size_t result; + + size_t start = 0; + /* + * These are dead stores, but the compiler will issue warnings on them + * since it can't tell statically that found is always true below. + */ + size_t begin = 0; + size_t len = 0; + + size_t largest_unchosen_range = 0; + while (true) { + bool found = fb_urange_iter(hpdata->active_pages, + HUGEPAGE_PAGES, start, &begin, &len); + /* + * A precondition to this function is that hpdata must be able + * to serve the allocation. + */ + assert(found); + assert(len <= hpdata_longest_free_range_get(hpdata)); + if (len >= npages) { + /* + * We use first-fit within the page slabs; this gives + * bounded worst-case fragmentation within a slab. It's + * not necessarily right; we could experiment with + * various other options. + */ + break; + } + if (len > largest_unchosen_range) { + largest_unchosen_range = len; + } + start = begin + len; + } + /* We found a range; remember it. */ + result = begin; + fb_set_range(hpdata->active_pages, HUGEPAGE_PAGES, begin, npages); + hpdata->h_nactive += npages; + + /* + * We might be about to dirty some memory for the first time; update our + * count if so. + */ + size_t new_dirty = fb_ucount(hpdata->touched_pages, HUGEPAGE_PAGES, + result, npages); + fb_set_range(hpdata->touched_pages, HUGEPAGE_PAGES, result, npages); + hpdata->h_ntouched += new_dirty; + + /* + * If we allocated out of a range that was the longest in the hpdata, it + * might be the only one of that size and we'll have to adjust the + * metadata. + */ + if (len == hpdata_longest_free_range_get(hpdata)) { + start = begin + npages; + while (start < HUGEPAGE_PAGES) { + bool found = fb_urange_iter(hpdata->active_pages, + HUGEPAGE_PAGES, start, &begin, &len); + if (!found) { + break; + } + assert(len <= hpdata_longest_free_range_get(hpdata)); + if (len == hpdata_longest_free_range_get(hpdata)) { + largest_unchosen_range = len; + break; + } + if (len > largest_unchosen_range) { + largest_unchosen_range = len; + } + start = begin + len; + } + hpdata_longest_free_range_set(hpdata, largest_unchosen_range); + } + + hpdata_assert_consistent(hpdata); + return (void *)( + (uintptr_t)hpdata_addr_get(hpdata) + (result << LG_PAGE)); +} + +void +hpdata_unreserve(hpdata_t *hpdata, void *addr, size_t sz) { + hpdata_assert_consistent(hpdata); + /* See the comment in reserve. */ + assert(!hpdata->h_in_psset || hpdata->h_updating); + assert(((uintptr_t)addr & PAGE_MASK) == 0); + assert((sz & PAGE_MASK) == 0); + size_t begin = ((uintptr_t)addr - (uintptr_t)hpdata_addr_get(hpdata)) + >> LG_PAGE; + assert(begin < HUGEPAGE_PAGES); + size_t npages = sz >> LG_PAGE; + size_t old_longest_range = hpdata_longest_free_range_get(hpdata); + + fb_unset_range(hpdata->active_pages, HUGEPAGE_PAGES, begin, npages); + /* We might have just created a new, larger range. */ + size_t new_begin = (fb_fls(hpdata->active_pages, HUGEPAGE_PAGES, + begin) + 1); + size_t new_end = fb_ffs(hpdata->active_pages, HUGEPAGE_PAGES, + begin + npages - 1); + size_t new_range_len = new_end - new_begin; + + if (new_range_len > old_longest_range) { + hpdata_longest_free_range_set(hpdata, new_range_len); + } + + hpdata->h_nactive -= npages; + + hpdata_assert_consistent(hpdata); +} + +size_t +hpdata_purge_begin(hpdata_t *hpdata, hpdata_purge_state_t *purge_state) { + hpdata_assert_consistent(hpdata); + /* + * See the comment below; we might purge any inactive extent, so it's + * unsafe for any other thread to turn any inactive extent active while + * we're operating on it. + */ + assert(!hpdata_alloc_allowed_get(hpdata)); + + purge_state->npurged = 0; + purge_state->next_purge_search_begin = 0; + + /* + * Initialize to_purge. + * + * It's possible to end up in situations where two dirty extents are + * separated by a retained extent: + * - 1 page allocated. + * - 1 page allocated. + * - 1 pages allocated. + * + * If the middle page is freed and purged, and then the first and third + * pages are freed, and then another purge pass happens, the hpdata + * looks like this: + * - 1 page dirty. + * - 1 page retained. + * - 1 page dirty. + * + * But it's safe to do a single 3-page purge. + * + * We do this by first computing the dirty pages, and then filling in + * any gaps by extending each range in the dirty bitmap to extend until + * the next active page. This purges more pages, but the expensive part + * of purging is the TLB shootdowns, rather than the kernel state + * tracking; doing a little bit more of the latter is fine if it saves + * us from doing some of the former. + */ + + /* + * The dirty pages are those that are touched but not active. Note that + * in a normal-ish case, HUGEPAGE_PAGES is something like 512 and the + * fb_group_t is 64 bits, so this is 64 bytes, spread across 8 + * fb_group_ts. + */ + fb_group_t dirty_pages[FB_NGROUPS(HUGEPAGE_PAGES)]; + fb_init(dirty_pages, HUGEPAGE_PAGES); + fb_bit_not(dirty_pages, hpdata->active_pages, HUGEPAGE_PAGES); + fb_bit_and(dirty_pages, dirty_pages, hpdata->touched_pages, + HUGEPAGE_PAGES); + + fb_init(purge_state->to_purge, HUGEPAGE_PAGES); + size_t next_bit = 0; + while (next_bit < HUGEPAGE_PAGES) { + size_t next_dirty = fb_ffs(dirty_pages, HUGEPAGE_PAGES, + next_bit); + /* Recall that fb_ffs returns nbits if no set bit is found. */ + if (next_dirty == HUGEPAGE_PAGES) { + break; + } + size_t next_active = fb_ffs(hpdata->active_pages, + HUGEPAGE_PAGES, next_dirty); + /* + * Don't purge past the end of the dirty extent, into retained + * pages. This helps the kernel a tiny bit, but honestly it's + * mostly helpful for testing (where we tend to write test cases + * that think in terms of the dirty ranges). + */ + ssize_t last_dirty = fb_fls(dirty_pages, HUGEPAGE_PAGES, + next_active - 1); + assert(last_dirty >= 0); + assert((size_t)last_dirty >= next_dirty); + assert((size_t)last_dirty - next_dirty + 1 <= HUGEPAGE_PAGES); + + fb_set_range(purge_state->to_purge, HUGEPAGE_PAGES, next_dirty, + last_dirty - next_dirty + 1); + next_bit = next_active + 1; + } + + /* We should purge, at least, everything dirty. */ + size_t ndirty = hpdata->h_ntouched - hpdata->h_nactive; + purge_state->ndirty_to_purge = ndirty; + assert(ndirty <= fb_scount( + purge_state->to_purge, HUGEPAGE_PAGES, 0, HUGEPAGE_PAGES)); + assert(ndirty == fb_scount(dirty_pages, HUGEPAGE_PAGES, 0, + HUGEPAGE_PAGES)); + + hpdata_assert_consistent(hpdata); + + return ndirty; +} + +bool +hpdata_purge_next(hpdata_t *hpdata, hpdata_purge_state_t *purge_state, + void **r_purge_addr, size_t *r_purge_size) { + /* + * Note that we don't have a consistency check here; we're accessing + * hpdata without synchronization, and therefore have no right to expect + * a consistent state. + */ + assert(!hpdata_alloc_allowed_get(hpdata)); + + if (purge_state->next_purge_search_begin == HUGEPAGE_PAGES) { + return false; + } + size_t purge_begin; + size_t purge_len; + bool found_range = fb_srange_iter(purge_state->to_purge, HUGEPAGE_PAGES, + purge_state->next_purge_search_begin, &purge_begin, &purge_len); + if (!found_range) { + return false; + } + + *r_purge_addr = (void *)( + (uintptr_t)hpdata_addr_get(hpdata) + purge_begin * PAGE); + *r_purge_size = purge_len * PAGE; + + purge_state->next_purge_search_begin = purge_begin + purge_len; + purge_state->npurged += purge_len; + assert(purge_state->npurged <= HUGEPAGE_PAGES); + + return true; +} + +void +hpdata_purge_end(hpdata_t *hpdata, hpdata_purge_state_t *purge_state) { + assert(!hpdata_alloc_allowed_get(hpdata)); + hpdata_assert_consistent(hpdata); + /* See the comment in reserve. */ + assert(!hpdata->h_in_psset || hpdata->h_updating); + + assert(purge_state->npurged == fb_scount(purge_state->to_purge, + HUGEPAGE_PAGES, 0, HUGEPAGE_PAGES)); + assert(purge_state->npurged >= purge_state->ndirty_to_purge); + + fb_bit_not(purge_state->to_purge, purge_state->to_purge, + HUGEPAGE_PAGES); + fb_bit_and(hpdata->touched_pages, hpdata->touched_pages, + purge_state->to_purge, HUGEPAGE_PAGES); + assert(hpdata->h_ntouched >= purge_state->ndirty_to_purge); + hpdata->h_ntouched -= purge_state->ndirty_to_purge; + + hpdata_assert_consistent(hpdata); +} + +void +hpdata_hugify(hpdata_t *hpdata) { + hpdata_assert_consistent(hpdata); + hpdata->h_huge = true; + fb_set_range(hpdata->touched_pages, HUGEPAGE_PAGES, 0, HUGEPAGE_PAGES); + hpdata->h_ntouched = HUGEPAGE_PAGES; + hpdata_assert_consistent(hpdata); +} + +void +hpdata_dehugify(hpdata_t *hpdata) { + hpdata_assert_consistent(hpdata); + hpdata->h_huge = false; + hpdata_assert_consistent(hpdata); +} diff --git a/contrib/jemalloc/src/inspect.c b/contrib/jemalloc/src/inspect.c new file mode 100644 --- /dev/null +++ b/contrib/jemalloc/src/inspect.c @@ -0,0 +1,77 @@ +#include "jemalloc/internal/jemalloc_preamble.h" +#include "jemalloc/internal/jemalloc_internal_includes.h" + +void +inspect_extent_util_stats_get(tsdn_t *tsdn, const void *ptr, size_t *nfree, + size_t *nregs, size_t *size) { + assert(ptr != NULL && nfree != NULL && nregs != NULL && size != NULL); + + const edata_t *edata = emap_edata_lookup(tsdn, &arena_emap_global, ptr); + if (unlikely(edata == NULL)) { + *nfree = *nregs = *size = 0; + return; + } + + *size = edata_size_get(edata); + if (!edata_slab_get(edata)) { + *nfree = 0; + *nregs = 1; + } else { + *nfree = edata_nfree_get(edata); + *nregs = bin_infos[edata_szind_get(edata)].nregs; + assert(*nfree <= *nregs); + assert(*nfree * edata_usize_get(edata) <= *size); + } +} + +void +inspect_extent_util_stats_verbose_get(tsdn_t *tsdn, const void *ptr, + size_t *nfree, size_t *nregs, size_t *size, size_t *bin_nfree, + size_t *bin_nregs, void **slabcur_addr) { + assert(ptr != NULL && nfree != NULL && nregs != NULL && size != NULL + && bin_nfree != NULL && bin_nregs != NULL && slabcur_addr != NULL); + + const edata_t *edata = emap_edata_lookup(tsdn, &arena_emap_global, ptr); + if (unlikely(edata == NULL)) { + *nfree = *nregs = *size = *bin_nfree = *bin_nregs = 0; + *slabcur_addr = NULL; + return; + } + + *size = edata_size_get(edata); + if (!edata_slab_get(edata)) { + *nfree = *bin_nfree = *bin_nregs = 0; + *nregs = 1; + *slabcur_addr = NULL; + return; + } + + *nfree = edata_nfree_get(edata); + const szind_t szind = edata_szind_get(edata); + *nregs = bin_infos[szind].nregs; + assert(*nfree <= *nregs); + assert(*nfree * edata_usize_get(edata) <= *size); + + arena_t *arena = (arena_t *)atomic_load_p( + &arenas[edata_arena_ind_get(edata)], ATOMIC_RELAXED); + assert(arena != NULL); + const unsigned binshard = edata_binshard_get(edata); + bin_t *bin = arena_get_bin(arena, szind, binshard); + + malloc_mutex_lock(tsdn, &bin->lock); + if (config_stats) { + *bin_nregs = *nregs * bin->stats.curslabs; + assert(*bin_nregs >= bin->stats.curregs); + *bin_nfree = *bin_nregs - bin->stats.curregs; + } else { + *bin_nfree = *bin_nregs = 0; + } + edata_t *slab; + if (bin->slabcur != NULL) { + slab = bin->slabcur; + } else { + slab = edata_heap_first(&bin->slabs_nonfull); + } + *slabcur_addr = slab != NULL ? edata_addr_get(slab) : NULL; + malloc_mutex_unlock(tsdn, &bin->lock); +} diff --git a/contrib/jemalloc/src/jemalloc.c b/contrib/jemalloc/src/jemalloc.c --- a/contrib/jemalloc/src/jemalloc.c +++ b/contrib/jemalloc/src/jemalloc.c @@ -4,20 +4,26 @@ #include "jemalloc/internal/assert.h" #include "jemalloc/internal/atomic.h" +#include "jemalloc/internal/buf_writer.h" #include "jemalloc/internal/ctl.h" +#include "jemalloc/internal/emap.h" #include "jemalloc/internal/extent_dss.h" #include "jemalloc/internal/extent_mmap.h" +#include "jemalloc/internal/fxp.h" +#include "jemalloc/internal/san.h" #include "jemalloc/internal/hook.h" #include "jemalloc/internal/jemalloc_internal_types.h" #include "jemalloc/internal/log.h" #include "jemalloc/internal/malloc_io.h" #include "jemalloc/internal/mutex.h" +#include "jemalloc/internal/nstime.h" #include "jemalloc/internal/rtree.h" #include "jemalloc/internal/safety_check.h" #include "jemalloc/internal/sc.h" #include "jemalloc/internal/spin.h" #include "jemalloc/internal/sz.h" #include "jemalloc/internal/ticker.h" +#include "jemalloc/internal/thread_event.h" #include "jemalloc/internal/util.h" /******************************************************************************/ @@ -33,6 +39,29 @@ JEMALLOC_ATTR(weak) #endif ; +/* + * The usual rule is that the closer to runtime you are, the higher priority + * your configuration settings are (so the jemalloc config options get lower + * priority than the per-binary setting, which gets lower priority than the /etc + * setting, which gets lower priority than the environment settings). + * + * But it's a fairly common use case in some testing environments for a user to + * be able to control the binary, but nothing else (e.g. a performancy canary + * uses the production OS and environment variables, but can run any binary in + * those circumstances). For these use cases, it's handy to have an in-binary + * mechanism for overriding environment variable settings, with the idea that if + * the results are positive they get promoted to the official settings, and + * moved from the binary to the environment variable. + * + * We don't actually want this to be widespread, so we'll give it a silly name + * and not mention it in headers or documentation. + */ +const char *je_malloc_conf_2_conf_harder +#ifndef _WIN32 + JEMALLOC_ATTR(weak) +#endif + ; + bool opt_abort = #ifdef JEMALLOC_DEBUG true @@ -70,16 +99,73 @@ false #endif ; +bool opt_trust_madvise = +#ifdef JEMALLOC_PURGE_MADVISE_DONTNEED_ZEROS + false +#else + true +#endif + ; + +bool opt_cache_oblivious = +#ifdef JEMALLOC_CACHE_OBLIVIOUS + true +#else + false +#endif + ; + +zero_realloc_action_t opt_zero_realloc_action = +#ifdef JEMALLOC_ZERO_REALLOC_DEFAULT_FREE + zero_realloc_action_free +#else + zero_realloc_action_alloc +#endif + ; + +atomic_zu_t zero_realloc_count = ATOMIC_INIT(0); + +const char *zero_realloc_mode_names[] = { + "alloc", + "free", + "abort", +}; + +/* + * These are the documented values for junk fill debugging facilities -- see the + * man page. + */ +static const uint8_t junk_alloc_byte = 0xa5; +static const uint8_t junk_free_byte = 0x5a; + +static void default_junk_alloc(void *ptr, size_t usize) { + memset(ptr, junk_alloc_byte, usize); +} + +static void default_junk_free(void *ptr, size_t usize) { + memset(ptr, junk_free_byte, usize); +} + +void (*junk_alloc_callback)(void *ptr, size_t size) = &default_junk_alloc; +void (*junk_free_callback)(void *ptr, size_t size) = &default_junk_free; bool opt_utrace = false; bool opt_xmalloc = false; +bool opt_experimental_infallible_new = false; bool opt_zero = false; unsigned opt_narenas = 0; +fxp_t opt_narenas_ratio = FXP_INIT_INT(4); unsigned ncpus; /* Protects arenas initialization. */ malloc_mutex_t arenas_lock; + +/* The global hpa, and whether it's on. */ +bool opt_hpa = false; +hpa_shard_opts_t opt_hpa_opts = HPA_SHARD_OPTS_DEFAULT; +sec_opts_t opt_hpa_sec_opts = SEC_OPTS_DEFAULT; + /* * Arenas that are used to service external requests. Not all elements of the * arenas array are necessarily used; arenas are created lazily as needed. @@ -98,13 +184,7 @@ unsigned narenas_auto; unsigned manual_arena_base; -typedef enum { - malloc_init_uninitialized = 3, - malloc_init_a0_initialized = 2, - malloc_init_recursible = 1, - malloc_init_initialized = 0 /* Common case --> jnz. */ -} malloc_init_t; -static malloc_init_t malloc_init_state = malloc_init_uninitialized; +malloc_init_t malloc_init_state = malloc_init_uninitialized; /* False should be the common case. Set to true to trigger initialization. */ bool malloc_slow = true; @@ -184,7 +264,7 @@ ut.p = (a); \ ut.s = (b); \ ut.r = (c); \ - utrace(&ut, sizeof(ut)); \ + UTRACE_CALL(&ut, sizeof(ut)); \ errno = utrace_serrno; \ } \ } while (0) @@ -209,11 +289,6 @@ * Begin miscellaneous support functions. */ -bool -malloc_initialized(void) { - return (malloc_init_state == malloc_init_initialized); -} - JEMALLOC_ALWAYS_INLINE bool malloc_init_a0(void) { if (unlikely(malloc_init_state == malloc_init_uninitialized)) { @@ -261,7 +336,7 @@ } /* - * FreeBSD's libc uses the bootstrap_*() functions in bootstrap-senstive + * FreeBSD's libc uses the bootstrap_*() functions in bootstrap-sensitive * situations that cannot tolerate TLS variable access (TLS allocation and very * early internal data structure initialization). */ @@ -319,7 +394,7 @@ /* Create a new arena and insert it into the arenas array at index ind. */ static arena_t * -arena_init_locked(tsdn_t *tsdn, unsigned ind, extent_hooks_t *extent_hooks) { +arena_init_locked(tsdn_t *tsdn, unsigned ind, const arena_config_t *config) { arena_t *arena; assert(ind <= narenas_total_get()); @@ -341,7 +416,7 @@ } /* Actually initialize the arena. */ - arena = arena_new(tsdn, ind, extent_hooks); + arena = arena_new(tsdn, ind, config); return arena; } @@ -365,11 +440,11 @@ } arena_t * -arena_init(tsdn_t *tsdn, unsigned ind, extent_hooks_t *extent_hooks) { +arena_init(tsdn_t *tsdn, unsigned ind, const arena_config_t *config) { arena_t *arena; malloc_mutex_lock(tsdn, &arenas_lock); - arena = arena_init_locked(tsdn, ind, extent_hooks); + arena = arena_init_locked(tsdn, ind, config); malloc_mutex_unlock(tsdn, &arenas_lock); arena_new_create_background_thread(tsdn, ind); @@ -398,14 +473,19 @@ } void -arena_migrate(tsd_t *tsd, unsigned oldind, unsigned newind) { - arena_t *oldarena, *newarena; +arena_migrate(tsd_t *tsd, arena_t *oldarena, arena_t *newarena) { + assert(oldarena != NULL); + assert(newarena != NULL); - oldarena = arena_get(tsd_tsdn(tsd), oldind, false); - newarena = arena_get(tsd_tsdn(tsd), newind, false); arena_nthreads_dec(oldarena, false); arena_nthreads_inc(newarena, false); tsd_arena_set(tsd, newarena); + + if (arena_nthreads_get(oldarena, false) == 0) { + /* Purge if the old arena has no associated threads anymore. */ + arena_decay(tsd_tsdn(tsd), oldarena, + /* is_background_thread */ false, /* all */ true); + } } static void @@ -422,82 +502,6 @@ } } -arena_tdata_t * -arena_tdata_get_hard(tsd_t *tsd, unsigned ind) { - arena_tdata_t *tdata, *arenas_tdata_old; - arena_tdata_t *arenas_tdata = tsd_arenas_tdata_get(tsd); - unsigned narenas_tdata_old, i; - unsigned narenas_tdata = tsd_narenas_tdata_get(tsd); - unsigned narenas_actual = narenas_total_get(); - - /* - * Dissociate old tdata array (and set up for deallocation upon return) - * if it's too small. - */ - if (arenas_tdata != NULL && narenas_tdata < narenas_actual) { - arenas_tdata_old = arenas_tdata; - narenas_tdata_old = narenas_tdata; - arenas_tdata = NULL; - narenas_tdata = 0; - tsd_arenas_tdata_set(tsd, arenas_tdata); - tsd_narenas_tdata_set(tsd, narenas_tdata); - } else { - arenas_tdata_old = NULL; - narenas_tdata_old = 0; - } - - /* Allocate tdata array if it's missing. */ - if (arenas_tdata == NULL) { - bool *arenas_tdata_bypassp = tsd_arenas_tdata_bypassp_get(tsd); - narenas_tdata = (ind < narenas_actual) ? narenas_actual : ind+1; - - if (tsd_nominal(tsd) && !*arenas_tdata_bypassp) { - *arenas_tdata_bypassp = true; - arenas_tdata = (arena_tdata_t *)a0malloc( - sizeof(arena_tdata_t) * narenas_tdata); - *arenas_tdata_bypassp = false; - } - if (arenas_tdata == NULL) { - tdata = NULL; - goto label_return; - } - assert(tsd_nominal(tsd) && !*arenas_tdata_bypassp); - tsd_arenas_tdata_set(tsd, arenas_tdata); - tsd_narenas_tdata_set(t