jemalloc Files

This release includes over 390 commits spanning bug fixes, new features, performance optimizations, and portability improvements. Multiple percent of system-level metric improvements were measured in tested production workloads. The release has gone through large-scale production testing at Meta.

New features:

• Support pvalloc. (@Lapenkov: 5b1f2cc5)
• Add double free detection for the debug build. (@izaitsevfb: 36366f3c, @guangli-dai: 42daa1ac, @divanorama: 1897f185)
• Add compile-time option --enable-pageid to enable memory mapping annotation. (@devnexen: 4fc5c4fb)
• Add runtime option prof_bt_max to control the max stack depth for profiling. (@guangli-dai: a0734fd6)
• Add compile-time option --enable-force-getenv to use getenv instead of secure_getenv. (@interwq: 481bbfc9)
• Add compile-time option --disable-dss to disable the usage of sbrk(2). (@Svetlitski: ea5b7bea)
• Add runtime option tcache_ncached_max to control the number of items in each size bin in the thread cache. (@guangli-dai: 8a22d10b)
• Add runtime option calloc_madvise_threshold to determine if kernel or memset is used to zero the allocations for calloc. (@nullptr0-0: 5081c16b)
• Add compile-time option --disable-user-config to disable reading the runtime configurations from /etc/malloc.conf or environment variable MALLOC_CONF. (@roblabla: c17bf8b3)
• Add runtime option disable_large_size_classes to guard the new usable size calculation, which minimizes the memory overhead for large allocations, i.e., >= 4 * PAGE. (@guangli-dai: c067a55c, 8347f104)
• Enable process_madvise usage, add runtime option process_madvise_max_batch to control the max # of regions in each madvise batch. (@interwq: 22440a02, @spredolac: 4246475b)
• Add mallctl interfaces:
  • opt.prof_bt_max (@guangli-dai: a0734fd6)
  • arena.<i>.name to set and get arena names. (@guangli-dai: ba19d2cb)
  • thread.tcache.max to set and get the tcache_max of the current thread. (@guangli-dai: a442d9b8)
  • thread.tcache.ncached_max.write and thread.tcache.ncached_max.read_sizeclass to set and get the ncached_max setup of the current thread. (@guangli-dai: 630f7de9, 6b197fdd)
  • arenas.hugepage to return the hugepage size used, also exported to malloc stats. (@ilvokhin: 90c627ed)
  • approximate_stats.active to return an estimate of the current active bytes, which should not be compared with other stats retrieved. (@guangli-dai: 0988583d)

Bug fixes:

• Prevent potential deadlocks in decaying during reentrancy. (@interwq: 434a68e2)
• Fix segfault in extent coalescing. (@Svetlitski: 12311fe6)
• Add null pointer detections in mallctl calls. (@Svetlitski: dc0a184f, 0288126d)
• Make mallctl arenas.lookup triable without crashing on invalid pointers. (@auxten: 019cccc2, 5bac3849)
• Demote sampled allocations for proper deallocations during arena_reset. (@Svetlitski: 62648c88)
• Fix jemalloc's read(2) and write(2). (@Svetlitski: d2c9ed3d, @lexprfuncall: 9fdc1160)
• Fix the pkg-config metadata file. (@BtbN: ed7e6fe7, ce8ce99a)
• Fix the autogen.sh so that it accepts quoted extra options. (@honggyukim: f6fe6abd)
• Fix rallocx() to set errno to ENOMEM upon OOMing. (@arter97: 38056fea, @interwq: 83b07578)
• Avoid stack overflow for internal variable array usage. (@nullptr0-0: 47c9bcd4, 48f66cf4, @xinydev: 9169e927)
• Fix background thread initialization race. (@puzpuzpuz: 4d0ffa07)
• Guard os_page_id against a NULL address. (@lexprfuncall: 79cc7dcc)
• Handle tcache init failures gracefully. (@lexprfuncall: a056c20d)
• Fix missing release of acquired neighbor edata in extent_try_coalesce_impl. (@spredolac: 675ab079)
• Fix memory leak of old curr_reg on san_bump_grow_locked failure. (@spredolac: 5904a421)
• Fix large alloc nrequests under-counting on cache misses. (@spredolac: 3cc56d32)

Portability improvements:

• Fix the build in C99. (@abaelhe: 56ddbea2)
• Add pthread_setaffinity_np detection for non Linux/BSD platforms. (@devnexen: 4c95c953)
• Make VARIABLE_ARRAY compatible with compilers not supporting VLA, i.e., Visual Studio C compiler in C11 or C17 modes. (@madscientist: be65438f)
• Fix the build on Linux using musl library. (@marv: aba1645f, 45249cf5)
• Reduce the memory overhead in small allocation sampling for systems with larger page sizes, e.g., ARM. (@Svetlitski: 5a858c64)
• Add C23's free_sized and free_aligned_sized. (@Svetlitski: cdb2c0e0)
• Enable heap profiling on MacOS. (@nullptr0-0: 4b555c11)
• Fix incorrect printing on 32bit. (@sundb: 630434bb)
• Make JEMALLOC_CXX_THROW compatible with C++ versions newer than C++17. (@r-barnes, @guangli-dai: 21bcc0a8)
• Fix mmap tag conflicts on MacOS. (@kdrag0n: c893fcd1)
• Fix monotonic timer assumption for win32. (@burtonli: 8dc97b11)
• Fix VM over-reservation on systems with larger pages, e.g., aarch64. (@interwq: cd05b19f)
• Remove unreachable() macro conditionally to prevent definition conflicts for C23+. (@appujee: d8486b26, 4b88bddb)
• Fix dlsym failure observed on FreeBSD. (@rhelmot: 86bbabac)
• Change the default page size to 64KB on aarch64 Linux. (@lexprfuncall: 9442300c)
• Update config.guess and config.sub to the latest version. (@lexprfuncall: c51949ea)
• Determine the page size on Android from NDK header files. (@lexprfuncall: c51abba1)
• Improve the portability of grep patterns in configure.ac. (@lexprfuncall: 365747bc)
• Add compile-time option --with-cxx-stdlib to specify the C++ standard library. (@yuxuanchen1997: a10ef3e1)

Optimizations and refactors:

• Enable tcache for deallocation-only threads. (@interwq: 143e9c4a)
• Inline to accelerate operator delete. (@guangli-dai: e8f9f138)
• Optimize pairing heap's performance. (@deadalnix: 5266152d, be6da4f6, 543e2d61, 10d71315, 92aa52c0, @Svetlitski: 36ca0c1b)
• Inline the storage for thread name in the profiling data. (@interwq: ce0b7ab6, e62aa478)
• Optimize a hot function edata_cmp_summary_comp to accelerate it. (@Svetlitski: 6841110b, @guangli-dai: 0181aaa4)
• Allocate thread cache using the base allocator, which enables thread cache to use thp when metadata_thp is turned on. (@interwq: 72cfdce7)
• Allow oversize arena not to purge immediately when background threads are enabled, although the default decay time is 0 to be back compatible. (@interwq: d1313313)
• Optimize thread-local storage implementation on Windows. (@mcfi: 9e123a83, 3a0d9cda)
• Optimize fast path to allow static size class computation. (@interwq: 323ed2e3)
• Redesign tcache GC to regulate the frequency and make it locality-aware. The new design is default on, guarded by option experimental_tcache_gc. (@nullptr0-0: 0c88be9e, e2c9f3a9, 14d5dc13, @deadalnix: 5afff2e4)
• Reduce the arena switching overhead by avoiding forced purging when background thread is enabled. (@interwq: a3910b98)
• Improve the reuse efficiency by limiting the maximum coalesced size for large extents. (@jiebinn: 3c14707b)
• Refactor thread events to allow registration of users' thread events and remove prof_threshold as the built-in event. (@spredolac: e6864c60, 015b0179, 34ace916)

Documentation:

• Update Windows building instructions. (@Lapenkov: 37139328)
• Add vcpkg installation instructions. (@LilyWangLL: c0c9783e)
• Update profiling internals with an example. (@jordalgo: b04e7666)