1 Following are change highlights associated with official releases. Important 2 bug fixes are all mentioned, but some internal enhancements are omitted here for 3 brevity. Much more detail can be found in the git revision history: 4 5 https://github.com/jemalloc/jemalloc 6 7 * 5.1.0 (May 4th, 2018) 8 9 This release is primarily about fine-tuning, ranging from several new features 10 to numerous notable performance and portability enhancements. The release and 11 prior dev versions have been running in multiple large scale applications for 12 months, and the cumulative improvements are substantial in many cases. 13 14 Given the long and successful production runs, this release is likely a good 15 candidate for applications to upgrade, from both jemalloc 5.0 and before. For 16 performance-critical applications, the newly added TUNING.md provides 17 guidelines on jemalloc tuning. 18 19 New features: 20 - Implement transparent huge page support for internal metadata. (@interwq) 21 - Add opt.thp to allow enabling / disabling transparent huge pages for all 22 mappings. (@interwq) 23 - Add maximum background thread count option. (@djwatson) 24 - Allow prof_active to control opt.lg_prof_interval and prof.gdump. 25 (@interwq) 26 - Allow arena index lookup based on allocation addresses via mallctl. 27 (@lionkov) 28 - Allow disabling initial-exec TLS model. (@davidtgoldblatt, @KenMacD) 29 - Add opt.lg_extent_max_active_fit to set the max ratio between the size of 30 the active extent selected (to split off from) and the size of the requested 31 allocation. (@interwq, @davidtgoldblatt) 32 - Add retain_grow_limit to set the max size when growing virtual address 33 space. (@interwq) 34 - Add mallctl interfaces: 35 + arena.<i>.retain_grow_limit (@interwq) 36 + arenas.lookup (@lionkov) 37 + max_background_threads (@djwatson) 38 + opt.lg_extent_max_active_fit (@interwq) 39 + opt.max_background_threads (@djwatson) 40 + opt.metadata_thp (@interwq) 41 + opt.thp (@interwq) 42 + stats.metadata_thp (@interwq) 43 44 Portability improvements: 45 - Support GNU/kFreeBSD configuration. (@paravoid) 46 - Support m68k, nios2 and SH3 architectures. (@paravoid) 47 - Fall back to FD_CLOEXEC when O_CLOEXEC is unavailable. (@zonyitoo) 48 - Fix symbol listing for cross-compiling. (@tamird) 49 - Fix high bits computation on ARM. (@davidtgoldblatt, @paravoid) 50 - Disable the CPU_SPINWAIT macro for Power. (@davidtgoldblatt, @marxin) 51 - Fix MSVC 2015 & 2017 builds. (@rustyx) 52 - Improve RISC-V support. (@EdSchouten) 53 - Set name mangling script in strict mode. (@nicolov) 54 - Avoid MADV_HUGEPAGE on ARM. (@marxin) 55 - Modify configure to determine return value of strerror_r. 56 (@davidtgoldblatt, @cferris1000) 57 - Make sure CXXFLAGS is tested with CPP compiler. (@nehaljwani) 58 - Fix 32-bit build on MSVC. (@rustyx) 59 - Fix external symbol on MSVC. (@maksqwe) 60 - Avoid a printf format specifier warning. (@jasone) 61 - Add configure option --disable-initial-exec-tls which can allow jemalloc to 62 be dynamically loaded after program startup. (@davidtgoldblatt, @KenMacD) 63 - AArch64: Add ILP32 support. (@cmuellner) 64 - Add --with-lg-vaddr configure option to support cross compiling. 65 (@cmuellner, @davidtgoldblatt) 66 67 Optimizations and refactors: 68 - Improve active extent fit with extent_max_active_fit. This considerably 69 reduces fragmentation over time and improves virtual memory and metadata 70 usage. (@davidtgoldblatt, @interwq) 71 - Eagerly coalesce large extents to reduce fragmentation. (@interwq) 72 - sdallocx: only read size info when page aligned (i.e. possibly sampled), 73 which speeds up the sized deallocation path significantly. (@interwq) 74 - Avoid attempting new mappings for in place expansion with retain, since 75 it rarely succeeds in practice and causes high overhead. (@interwq) 76 - Refactor OOM handling in newImpl. (@wqfish) 77 - Add internal fine-grained logging functionality for debugging use. 78 (@davidtgoldblatt) 79 - Refactor arena / tcache interactions. (@davidtgoldblatt) 80 - Refactor extent management with dumpable flag. (@davidtgoldblatt) 81 - Add runtime detection of lazy purging. (@interwq) 82 - Use pairing heap instead of red-black tree for extents_avail. (@djwatson) 83 - Use sysctl on startup in FreeBSD. (@trasz) 84 - Use thread local prng state instead of atomic. (@djwatson) 85 - Make decay to always purge one more extent than before, because in 86 practice large extents are usually the ones that cross the decay threshold. 87 Purging the additional extent helps save memory as well as reduce VM 88 fragmentation. (@interwq) 89 - Fast division by dynamic values. (@davidtgoldblatt) 90 - Improve the fit for aligned allocation. (@interwq, @edwinsmith) 91 - Refactor extent_t bitpacking. (@rkmisra) 92 - Optimize the generated assembly for ticker operations. (@davidtgoldblatt) 93 - Convert stats printing to use a structured text emitter. (@davidtgoldblatt) 94 - Remove preserve_lru feature for extents management. (@djwatson) 95 - Consolidate two memory loads into one on the fast deallocation path. 96 (@davidtgoldblatt, @interwq) 97 98 Bug fixes (most of the issues are only relevant to jemalloc 5.0): 99 - Fix deadlock with multithreaded fork in OS X. (@davidtgoldblatt) 100 - Validate returned file descriptor before use. (@zonyitoo) 101 - Fix a few background thread initialization and shutdown issues. (@interwq) 102 - Fix an extent coalesce + decay race by taking both coalescing extents off 103 the LRU list. (@interwq) 104 - Fix potentially unbound increase during decay, caused by one thread keep 105 stashing memory to purge while other threads generating new pages. The 106 number of pages to purge is checked to prevent this. (@interwq) 107 - Fix a FreeBSD bootstrap assertion. (@strejda, @interwq) 108 - Handle 32 bit mutex counters. (@rkmisra) 109 - Fix a indexing bug when creating background threads. (@davidtgoldblatt, 110 @binliu19) 111 - Fix arguments passed to extent_init. (@yuleniwo, @interwq) 112 - Fix addresses used for ordering mutexes. (@rkmisra) 113 - Fix abort_conf processing during bootstrap. (@interwq) 114 - Fix include path order for out-of-tree builds. (@cmuellner) 115 116 Incompatible changes: 117 - Remove --disable-thp. (@interwq) 118 - Remove mallctl interfaces: 119 + config.thp (@interwq) 120 121 Documentation: 122 - Add TUNING.md. (@interwq, @davidtgoldblatt, @djwatson) 123 124 * 5.0.1 (July 1, 2017) 125 126 This bugfix release fixes several issues, most of which are obscure enough 127 that typical applications are not impacted. 128 129 Bug fixes: 130 - Update decay->nunpurged before purging, in order to avoid potential update 131 races and subsequent incorrect purging volume. (@interwq) 132 - Only abort on dlsym(3) error if the failure impacts an enabled feature (lazy 133 locking and/or background threads). This mitigates an initialization 134 failure bug for which we still do not have a clear reproduction test case. 135 (@interwq) 136 - Modify tsd management so that it neither crashes nor leaks if a thread's 137 only allocation activity is to call free() after TLS destructors have been 138 executed. This behavior was observed when operating with GNU libc, and is 139 unlikely to be an issue with other libc implementations. (@interwq) 140 - Mask signals during background thread creation. This prevents signals from 141 being inadvertently delivered to background threads. (@jasone, 142 @davidtgoldblatt, @interwq) 143 - Avoid inactivity checks within background threads, in order to prevent 144 recursive mutex acquisition. (@interwq) 145 - Fix extent_grow_retained() to use the specified hooks when the 146 arena.<i>.extent_hooks mallctl is used to override the default hooks. 147 (@interwq) 148 - Add missing reentrancy support for custom extent hooks which allocate. 149 (@interwq) 150 - Post-fork(2), re-initialize the list of tcaches associated with each arena 151 to contain no tcaches except the forking thread's. (@interwq) 152 - Add missing post-fork(2) mutex reinitialization for extent_grow_mtx. This 153 fixes potential deadlocks after fork(2). (@interwq) 154 - Enforce minimum autoconf version (currently 2.68), since 2.63 is known to 155 generate corrupt configure scripts. (@jasone) 156 - Ensure that the configured page size (--with-lg-page) is no larger than the 157 configured huge page size (--with-lg-hugepage). (@jasone) 158 159 * 5.0.0 (June 13, 2017) 160 161 Unlike all previous jemalloc releases, this release does not use naturally 162 aligned "chunks" for virtual memory management, and instead uses page-aligned 163 "extents". This change has few externally visible effects, but the internal 164 impacts are... extensive. Many other internal changes combine to make this 165 the most cohesively designed version of jemalloc so far, with ample 166 opportunity for further enhancements. 167 168 Continuous integration is now an integral aspect of development thanks to the 169 efforts of @davidtgoldblatt, and the dev branch tends to remain reasonably 170 stable on the tested platforms (Linux, FreeBSD, macOS, and Windows). As a 171 side effect the official release frequency may decrease over time. 172 173 New features: 174 - Implement optional per-CPU arena support; threads choose which arena to use 175 based on current CPU rather than on fixed thread-->arena associations. 176 (@interwq) 177 - Implement two-phase decay of unused dirty pages. Pages transition from 178 dirty-->muzzy-->clean, where the first phase transition relies on 179 madvise(... MADV_FREE) semantics, and the second phase transition discards 180 pages such that they are replaced with demand-zeroed pages on next access. 181 (@jasone) 182 - Increase decay time resolution from seconds to milliseconds. (@jasone) 183 - Implement opt-in per CPU background threads, and use them for asynchronous 184 decay-driven unused dirty page purging. (@interwq) 185 - Add mutex profiling, which collects a variety of statistics useful for 186 diagnosing overhead/contention issues. (@interwq) 187 - Add C++ new/delete operator bindings. (@djwatson) 188 - Support manually created arena destruction, such that all data and metadata 189 are discarded. Add MALLCTL_ARENAS_DESTROYED for accessing merged stats 190 associated with destroyed arenas. (@jasone) 191 - Add MALLCTL_ARENAS_ALL as a fixed index for use in accessing 192 merged/destroyed arena statistics via mallctl. (@jasone) 193 - Add opt.abort_conf to optionally abort if invalid configuration options are 194 detected during initialization. (@interwq) 195 - Add opt.stats_print_opts, so that e.g. JSON output can be selected for the 196 stats dumped during exit if opt.stats_print is true. (@jasone) 197 - Add --with-version=VERSION for use when embedding jemalloc into another 198 project's git repository. (@jasone) 199 - Add --disable-thp to support cross compiling. (@jasone) 200 - Add --with-lg-hugepage to support cross compiling. (@jasone) 201 - Add mallctl interfaces (various authors): 202 + background_thread 203 + opt.abort_conf 204 + opt.retain 205 + opt.percpu_arena 206 + opt.background_thread 207 + opt.{dirty,muzzy}_decay_ms 208 + opt.stats_print_opts 209 + arena.<i>.initialized 210 + arena.<i>.destroy 211 + arena.<i>.{dirty,muzzy}_decay_ms 212 + arena.<i>.extent_hooks 213 + arenas.{dirty,muzzy}_decay_ms 214 + arenas.bin.<i>.slab_size 215 + arenas.nlextents 216 + arenas.lextent.<i>.size 217 + arenas.create 218 + stats.background_thread.{num_threads,num_runs,run_interval} 219 + stats.mutexes.{ctl,background_thread,prof,reset}. 220 {num_ops,num_spin_acq,num_wait,max_wait_time,total_wait_time,max_num_thds, 221 num_owner_switch} 222 + stats.arenas.<i>.{dirty,muzzy}_decay_ms 223 + stats.arenas.<i>.uptime 224 + stats.arenas.<i>.{pmuzzy,base,internal,resident} 225 + stats.arenas.<i>.{dirty,muzzy}_{npurge,nmadvise,purged} 226 + stats.arenas.<i>.bins.<j>.{nslabs,reslabs,curslabs} 227 + stats.arenas.<i>.bins.<j>.mutex. 228 {num_ops,num_spin_acq,num_wait,max_wait_time,total_wait_time,max_num_thds, 229 num_owner_switch} 230 + stats.arenas.<i>.lextents.<j>.{nmalloc,ndalloc,nrequests,curlextents} 231 + stats.arenas.i.mutexes.{large,extent_avail,extents_dirty,extents_muzzy, 232 extents_retained,decay_dirty,decay_muzzy,base,tcache_list}. 233 {num_ops,num_spin_acq,num_wait,max_wait_time,total_wait_time,max_num_thds, 234 num_owner_switch} 235 236 Portability improvements: 237 - Improve reentrant allocation support, such that deadlock is less likely if 238 e.g. a system library call in turn allocates memory. (@davidtgoldblatt, 239 @interwq) 240 - Support static linking of jemalloc with glibc. (@djwatson) 241 242 Optimizations and refactors: 243 - Organize virtual memory as "extents" of virtual memory pages, rather than as 244 naturally aligned "chunks", and store all metadata in arbitrarily distant 245 locations. This reduces virtual memory external fragmentation, and will 246 interact better with huge pages (not yet explicitly supported). (@jasone) 247 - Fold large and huge size classes together; only small and large size classes 248 remain. (@jasone) 249 - Unify the allocation paths, and merge most fast-path branching decisions. 250 (@davidtgoldblatt, @interwq) 251 - Embed per thread automatic tcache into thread-specific data, which reduces 252 conditional branches and dereferences. Also reorganize tcache to increase 253 fast-path data locality. (@interwq) 254 - Rewrite atomics to closely model the C11 API, convert various 255 synchronization from mutex-based to atomic, and use the explicit memory 256 ordering control to resolve various hypothetical races without increasing 257 synchronization overhead. (@davidtgoldblatt) 258 - Extensively optimize rtree via various methods: 259 + Add multiple layers of rtree lookup caching, since rtree lookups are now 260 part of fast-path deallocation. (@interwq) 261 + Determine rtree layout at compile time. (@jasone) 262 + Make the tree shallower for common configurations. (@jasone) 263 + Embed the root node in the top-level rtree data structure, thus avoiding 264 one level of indirection. (@jasone) 265 + Further specialize leaf elements as compared to internal node elements, 266 and directly embed extent metadata needed for fast-path deallocation. 267 (@jasone) 268 + Ignore leading always-zero address bits (architecture-specific). 269 (@jasone) 270 - Reorganize headers (ongoing work) to make them hermetic, and disentangle 271 various module dependencies. (@davidtgoldblatt) 272 - Convert various internal data structures such as size class metadata from 273 boot-time-initialized to compile-time-initialized. Propagate resulting data 274 structure simplifications, such as making arena metadata fixed-size. 275 (@jasone) 276 - Simplify size class lookups when constrained to size classes that are 277 multiples of the page size. This speeds lookups, but the primary benefit is 278 complexity reduction in code that was the source of numerous regressions. 279 (@jasone) 280 - Lock individual extents when possible for localized extent operations, 281 rather than relying on a top-level arena lock. (@davidtgoldblatt, @jasone) 282 - Use first fit layout policy instead of best fit, in order to improve 283 packing. (@jasone) 284 - If munmap(2) is not in use, use an exponential series to grow each arena's 285 virtual memory, so that the number of disjoint virtual memory mappings 286 remains low. (@jasone) 287 - Implement per arena base allocators, so that arenas never share any virtual 288 memory pages. (@jasone) 289 - Automatically generate private symbol name mangling macros. (@jasone) 290 291 Incompatible changes: 292 - Replace chunk hooks with an expanded/normalized set of extent hooks. 293 (@jasone) 294 - Remove ratio-based purging. (@jasone) 295 - Remove --disable-tcache. (@jasone) 296 - Remove --disable-tls. (@jasone) 297 - Remove --enable-ivsalloc. (@jasone) 298 - Remove --with-lg-size-class-group. (@jasone) 299 - Remove --with-lg-tiny-min. (@jasone) 300 - Remove --disable-cc-silence. (@jasone) 301 - Remove --enable-code-coverage. (@jasone) 302 - Remove --disable-munmap (replaced by opt.retain). (@jasone) 303 - Remove Valgrind support. (@jasone) 304 - Remove quarantine support. (@jasone) 305 - Remove redzone support. (@jasone) 306 - Remove mallctl interfaces (various authors): 307 + config.munmap 308 + config.tcache 309 + config.tls 310 + config.valgrind 311 + opt.lg_chunk 312 + opt.purge 313 + opt.lg_dirty_mult 314 + opt.decay_time 315 + opt.quarantine 316 + opt.redzone 317 + opt.thp 318 + arena.<i>.lg_dirty_mult 319 + arena.<i>.decay_time 320 + arena.<i>.chunk_hooks 321 + arenas.initialized 322 + arenas.lg_dirty_mult 323 + arenas.decay_time 324 + arenas.bin.<i>.run_size 325 + arenas.nlruns 326 + arenas.lrun.<i>.size 327 + arenas.nhchunks 328 + arenas.hchunk.<i>.size 329 + arenas.extend 330 + stats.cactive 331 + stats.arenas.<i>.lg_dirty_mult 332 + stats.arenas.<i>.decay_time 333 + stats.arenas.<i>.metadata.{mapped,allocated} 334 + stats.arenas.<i>.{npurge,nmadvise,purged} 335 + stats.arenas.<i>.huge.{allocated,nmalloc,ndalloc,nrequests} 336 + stats.arenas.<i>.bins.<j>.{nruns,reruns,curruns} 337 + stats.arenas.<i>.lruns.<j>.{nmalloc,ndalloc,nrequests,curruns} 338 + stats.arenas.<i>.hchunks.<j>.{nmalloc,ndalloc,nrequests,curhchunks} 339 340 Bug fixes: 341 - Improve interval-based profile dump triggering to dump only one profile when 342 a single allocation's size exceeds the interval. (@jasone) 343 - Use prefixed function names (as controlled by --with-jemalloc-prefix) when 344 pruning backtrace frames in jeprof. (@jasone) 345 346 * 4.5.0 (February 28, 2017) 347 348 This is the first release to benefit from much broader continuous integration 349 testing, thanks to @davidtgoldblatt. Had we had this testing infrastructure 350 in place for prior releases, it would have caught all of the most serious 351 regressions fixed by this release. 352 353 New features: 354 - Add --disable-thp and the opt.thp mallctl to provide opt-out mechanisms for 355 transparent huge page integration. (@jasone) 356 - Update zone allocator integration to work with macOS 10.12. (@glandium) 357 - Restructure *CFLAGS configuration, so that CFLAGS behaves typically, and 358 EXTRA_CFLAGS provides a way to specify e.g. -Werror during building, but not 359 during configuration. (@jasone, @ronawho) 360 361 Bug fixes: 362 - Fix DSS (sbrk(2)-based) allocation. This regression was first released in 363 4.3.0. (@jasone) 364 - Handle race in per size class utilization computation. This functionality 365 was first released in 4.0.0. (@interwq) 366 - Fix lock order reversal during gdump. (@jasone) 367 - Fix/refactor tcache synchronization. This regression was first released in 368 4.0.0. (@jasone) 369 - Fix various JSON-formatted malloc_stats_print() bugs. This functionality 370 was first released in 4.3.0. (@jasone) 371 - Fix huge-aligned allocation. This regression was first released in 4.4.0. 372 (@jasone) 373 - When transparent huge page integration is enabled, detect what state pages 374 start in according to the kernel's current operating mode, and only convert 375 arena chunks to non-huge during purging if that is not their initial state. 376 This functionality was first released in 4.4.0. (@jasone) 377 - Fix lg_chunk clamping for the --enable-cache-oblivious --disable-fill case. 378 This regression was first released in 4.0.0. (@jasone, @428desmo) 379 - Properly detect sparc64 when building for Linux. (@glaubitz) 380 381 * 4.4.0 (December 3, 2016) 382 383 New features: 384 - Add configure support for *-*-linux-android. (@cferris1000, @jasone) 385 - Add the --disable-syscall configure option, for use on systems that place 386 security-motivated limitations on syscall(2). (@jasone) 387 - Add support for Debian GNU/kFreeBSD. (@thesam) 388 389 Optimizations: 390 - Add extent serial numbers and use them where appropriate as a sort key that 391 is higher priority than address, so that the allocation policy prefers older 392 extents. This tends to improve locality (decrease fragmentation) when 393 memory grows downward. (@jasone) 394 - Refactor madvise(2) configuration so that MADV_FREE is detected and utilized 395 on Linux 4.5 and newer. (@jasone) 396 - Mark partially purged arena chunks as non-huge-page. This improves 397 interaction with Linux's transparent huge page functionality. (@jasone) 398 399 Bug fixes: 400 - Fix size class computations for edge conditions involving extremely large 401 allocations. This regression was first released in 4.0.0. (@jasone, 402 @ingvarha) 403 - Remove overly restrictive assertions related to the cactive statistic. This 404 regression was first released in 4.1.0. (@jasone) 405 - Implement a more reliable detection scheme for os_unfair_lock on macOS. 406 (@jszakmeister) 407 408 * 4.3.1 (November 7, 2016) 409 410 Bug fixes: 411 - Fix a severe virtual memory leak. This regression was first released in 412 4.3.0. (@interwq, @jasone) 413 - Refactor atomic and prng APIs to restore support for 32-bit platforms that 414 use pre-C11 toolchains, e.g. FreeBSD's mips. (@jasone) 415 416 * 4.3.0 (November 4, 2016) 417 418 This is the first release that passes the test suite for multiple Windows 419 configurations, thanks in large part to @glandium setting up continuous 420 integration via AppVeyor (and Travis CI for Linux and OS X). 421 422 New features: 423 - Add "J" (JSON) support to malloc_stats_print(). (@jasone) 424 - Add Cray compiler support. (@ronawho) 425 426 Optimizations: 427 - Add/use adaptive spinning for bootstrapping and radix tree node 428 initialization. (@jasone) 429 430 Bug fixes: 431 - Fix large allocation to search starting in the optimal size class heap, 432 which can substantially reduce virtual memory churn and fragmentation. This 433 regression was first released in 4.0.0. (@mjp41, @jasone) 434 - Fix stats.arenas.<i>.nthreads accounting. (@interwq) 435 - Fix and simplify decay-based purging. (@jasone) 436 - Make DSS (sbrk(2)-related) operations lockless, which resolves potential 437 deadlocks during thread exit. (@jasone) 438 - Fix over-sized allocation of radix tree leaf nodes. (@mjp41, @ogaun, 439 @jasone) 440 - Fix over-sized allocation of arena_t (plus associated stats) data 441 structures. (@jasone, @interwq) 442 - Fix EXTRA_CFLAGS to not affect configuration. (@jasone) 443 - Fix a Valgrind integration bug. (@ronawho) 444 - Disallow 0x5a junk filling when running in Valgrind. (@jasone) 445 - Fix a file descriptor leak on Linux. This regression was first released in 446 4.2.0. (@vsarunas, @jasone) 447 - Fix static linking of jemalloc with glibc. (@djwatson) 448 - Use syscall(2) rather than {open,read,close}(2) during boot on Linux. This 449 works around other libraries' system call wrappers performing reentrant 450 allocation. (@kspinka, @Whissi, @jasone) 451 - Fix OS X default zone replacement to work with OS X 10.12. (@glandium, 452 @jasone) 453 - Fix cached memory management to avoid needless commit/decommit operations 454 during purging, which resolves permanent virtual memory map fragmentation 455 issues on Windows. (@mjp41, @jasone) 456 - Fix TSD fetches to avoid (recursive) allocation. This is relevant to 457 non-TLS and Windows configurations. (@jasone) 458 - Fix malloc_conf overriding to work on Windows. (@jasone) 459 - Forcibly disable lazy-lock on Windows (was forcibly *enabled*). (@jasone) 460 461 * 4.2.1 (June 8, 2016) 462 463 Bug fixes: 464 - Fix bootstrapping issues for configurations that require allocation during 465 tsd initialization (e.g. --disable-tls). (@cferris1000, @jasone) 466 - Fix gettimeofday() version of nstime_update(). (@ronawho) 467 - Fix Valgrind regressions in calloc() and chunk_alloc_wrapper(). (@ronawho) 468 - Fix potential VM map fragmentation regression. (@jasone) 469 - Fix opt_zero-triggered in-place huge reallocation zeroing. (@jasone) 470 - Fix heap profiling context leaks in reallocation edge cases. (@jasone) 471 472 * 4.2.0 (May 12, 2016) 473 474 New features: 475 - Add the arena.<i>.reset mallctl, which makes it possible to discard all of 476 an arena's allocations in a single operation. (@jasone) 477 - Add the stats.retained and stats.arenas.<i>.retained statistics. (@jasone) 478 - Add the --with-version configure option. (@jasone) 479 - Support --with-lg-page values larger than actual page size. (@jasone) 480 481 Optimizations: 482 - Use pairing heaps rather than red-black trees for various hot data 483 structures. (@djwatson, @jasone) 484 - Streamline fast paths of rtree operations. (@jasone) 485 - Optimize the fast paths of calloc() and [m,d,sd]allocx(). (@jasone) 486 - Decommit unused virtual memory if the OS does not overcommit. (@jasone) 487 - Specify MAP_NORESERVE on Linux if [heuristic] overcommit is active, in order 488 to avoid unfortunate interactions during fork(2). (@jasone) 489 490 Bug fixes: 491 - Fix chunk accounting related to triggering gdump profiles. (@jasone) 492 - Link against librt for clock_gettime(2) if glibc < 2.17. (@jasone) 493 - Scale leak report summary according to sampling probability. (@jasone) 494 495 * 4.1.1 (May 3, 2016) 496 497 This bugfix release resolves a variety of mostly minor issues, though the 498 bitmap fix is critical for 64-bit Windows. 499 500 Bug fixes: 501 - Fix the linear scan version of bitmap_sfu() to shift by the proper amount 502 even when sizeof(long) is not the same as sizeof(void *), as on 64-bit 503 Windows. (@jasone) 504 - Fix hashing functions to avoid unaligned memory accesses (and resulting 505 crashes). This is relevant at least to some ARM-based platforms. 506 (@rkmisra) 507 - Fix fork()-related lock rank ordering reversals. These reversals were 508 unlikely to cause deadlocks in practice except when heap profiling was 509 enabled and active. (@jasone) 510 - Fix various chunk leaks in OOM code paths. (@jasone) 511 - Fix malloc_stats_print() to print opt.narenas correctly. (@jasone) 512 - Fix MSVC-specific build/test issues. (@rustyx, @yuslepukhin) 513 - Fix a variety of test failures that were due to test fragility rather than 514 core bugs. (@jasone) 515 516 * 4.1.0 (February 28, 2016) 517 518 This release is primarily about optimizations, but it also incorporates a lot 519 of portability-motivated refactoring and enhancements. Many people worked on 520 this release, to an extent that even with the omission here of minor changes 521 (see git revision history), and of the people who reported and diagnosed 522 issues, so much of the work was contributed that starting with this release, 523 changes are annotated with author credits to help reflect the collaborative 524 effort involved. 525 526 New features: 527 - Implement decay-based unused dirty page purging, a major optimization with 528 mallctl API impact. This is an alternative to the existing ratio-based 529 unused dirty page purging, and is intended to eventually become the sole 530 purging mechanism. New mallctls: 531 + opt.purge 532 + opt.decay_time 533 + arena.<i>.decay 534 + arena.<i>.decay_time 535 + arenas.decay_time 536 + stats.arenas.<i>.decay_time 537 (@jasone, @cevans87) 538 - Add --with-malloc-conf, which makes it possible to embed a default 539 options string during configuration. This was motivated by the desire to 540 specify --with-malloc-conf=purge:decay , since the default must remain 541 purge:ratio until the 5.0.0 release. (@jasone) 542 - Add MS Visual Studio 2015 support. (@rustyx, @yuslepukhin) 543 - Make *allocx() size class overflow behavior defined. The maximum 544 size class is now less than PTRDIFF_MAX to protect applications against 545 numerical overflow, and all allocation functions are guaranteed to indicate 546 errors rather than potentially crashing if the request size exceeds the 547 maximum size class. (@jasone) 548 - jeprof: 549 + Add raw heap profile support. (@jasone) 550 + Add --retain and --exclude for backtrace symbol filtering. (@jasone) 551 552 Optimizations: 553 - Optimize the fast path to combine various bootstrapping and configuration 554 checks and execute more streamlined code in the common case. (@interwq) 555 - Use linear scan for small bitmaps (used for small object tracking). In 556 addition to speeding up bitmap operations on 64-bit systems, this reduces 557 allocator metadata overhead by approximately 0.2%. (@djwatson) 558 - Separate arena_avail trees, which substantially speeds up run tree 559 operations. (@djwatson) 560 - Use memoization (boot-time-computed table) for run quantization. Separate 561 arena_avail trees reduced the importance of this optimization. (@jasone) 562 - Attempt mmap-based in-place huge reallocation. This can dramatically speed 563 up incremental huge reallocation. (@jasone) 564 565 Incompatible changes: 566 - Make opt.narenas unsigned rather than size_t. (@jasone) 567 568 Bug fixes: 569 - Fix stats.cactive accounting regression. (@rustyx, @jasone) 570 - Handle unaligned keys in hash(). This caused problems for some ARM systems. 571 (@jasone, @cferris1000) 572 - Refactor arenas array. In addition to fixing a fork-related deadlock, this 573 makes arena lookups faster and simpler. (@jasone) 574 - Move retained memory allocation out of the default chunk allocation 575 function, to a location that gets executed even if the application installs 576 a custom chunk allocation function. This resolves a virtual memory leak. 577 (@buchgr) 578 - Fix a potential tsd cleanup leak. (@cferris1000, @jasone) 579 - Fix run quantization. In practice this bug had no impact unless 580 applications requested memory with alignment exceeding one page. 581 (@jasone, @djwatson) 582 - Fix LinuxThreads-specific bootstrapping deadlock. (Cosmin Paraschiv) 583 - jeprof: 584 + Don't discard curl options if timeout is not defined. (@djwatson) 585 + Detect failed profile fetches. (@djwatson) 586 - Fix stats.arenas.<i>.{dss,lg_dirty_mult,decay_time,pactive,pdirty} for 587 --disable-stats case. (@jasone) 588 589 * 4.0.4 (October 24, 2015) 590 591 This bugfix release fixes another xallocx() regression. No other regressions 592 have come to light in over a month, so this is likely a good starting point 593 for people who prefer to wait for "dot one" releases with all the major issues 594 shaken out. 595 596 Bug fixes: 597 - Fix xallocx(..., MALLOCX_ZERO to zero the last full trailing page of large 598 allocations that have been randomly assigned an offset of 0 when 599 --enable-cache-oblivious configure option is enabled. 600 601 * 4.0.3 (September 24, 2015) 602 603 This bugfix release continues the trend of xallocx() and heap profiling fixes. 604 605 Bug fixes: 606 - Fix xallocx(..., MALLOCX_ZERO) to zero all trailing bytes of large 607 allocations when --enable-cache-oblivious configure option is enabled. 608 - Fix xallocx(..., MALLOCX_ZERO) to zero trailing bytes of huge allocations 609 when resizing from/to a size class that is not a multiple of the chunk size. 610 - Fix prof_tctx_dump_iter() to filter out nodes that were created after heap 611 profile dumping started. 612 - Work around a potentially bad thread-specific data initialization 613 interaction with NPTL (glibc's pthreads implementation). 614 615 * 4.0.2 (September 21, 2015) 616 617 This bugfix release addresses a few bugs specific to heap profiling. 618 619 Bug fixes: 620 - Fix ixallocx_prof_sample() to never modify nor create sampled small 621 allocations. xallocx() is in general incapable of moving small allocations, 622 so this fix removes buggy code without loss of generality. 623 - Fix irallocx_prof_sample() to always allocate large regions, even when 624 alignment is non-zero. 625 - Fix prof_alloc_rollback() to read tdata from thread-specific data rather 626 than dereferencing a potentially invalid tctx. 627 628 * 4.0.1 (September 15, 2015) 629 630 This is a bugfix release that is somewhat high risk due to the amount of 631 refactoring required to address deep xallocx() problems. As a side effect of 632 these fixes, xallocx() now tries harder to partially fulfill requests for 633 optional extra space. Note that a couple of minor heap profiling 634 optimizations are included, but these are better thought of as performance 635 fixes that were integral to discovering most of the other bugs. 636 637 Optimizations: 638 - Avoid a chunk metadata read in arena_prof_tctx_set(), since it is in the 639 fast path when heap profiling is enabled. Additionally, split a special 640 case out into arena_prof_tctx_reset(), which also avoids chunk metadata 641 reads. 642 - Optimize irallocx_prof() to optimistically update the sampler state. The 643 prior implementation appears to have been a holdover from when 644 rallocx()/xallocx() functionality was combined as rallocm(). 645 646 Bug fixes: 647 - Fix TLS configuration such that it is enabled by default for platforms on 648 which it works correctly. 649 - Fix arenas_cache_cleanup() and arena_get_hard() to handle 650 allocation/deallocation within the application's thread-specific data 651 cleanup functions even after arenas_cache is torn down. 652 - Fix xallocx() bugs related to size+extra exceeding HUGE_MAXCLASS. 653 - Fix chunk purge hook calls for in-place huge shrinking reallocation to 654 specify the old chunk size rather than the new chunk size. This bug caused 655 no correctness issues for the default chunk purge function, but was 656 visible to custom functions set via the "arena.<i>.chunk_hooks" mallctl. 657 - Fix heap profiling bugs: 658 + Fix heap profiling to distinguish among otherwise identical sample sites 659 with interposed resets (triggered via the "prof.reset" mallctl). This bug 660 could cause data structure corruption that would most likely result in a 661 segfault. 662 + Fix irealloc_prof() to prof_alloc_rollback() on OOM. 663 + Make one call to prof_active_get_unlocked() per allocation event, and use 664 the result throughout the relevant functions that handle an allocation 665 event. Also add a missing check in prof_realloc(). These fixes protect 666 allocation events against concurrent prof_active changes. 667 + Fix ixallocx_prof() to pass usize_max and zero to ixallocx_prof_sample() 668 in the correct order. 669 + Fix prof_realloc() to call prof_free_sampled_object() after calling 670 prof_malloc_sample_object(). Prior to this fix, if tctx and old_tctx were 671 the same, the tctx could have been prematurely destroyed. 672 - Fix portability bugs: 673 + Don't bitshift by negative amounts when encoding/decoding run sizes in 674 chunk header maps. This affected systems with page sizes greater than 8 675 KiB. 676 + Rename index_t to szind_t to avoid an existing type on Solaris. 677 + Add JEMALLOC_CXX_THROW to the memalign() function prototype, in order to 678 match glibc and avoid compilation errors when including both 679 jemalloc/jemalloc.h and malloc.h in C++ code. 680 + Don't assume that /bin/sh is appropriate when running size_classes.sh 681 during configuration. 682 + Consider __sparcv9 a synonym for __sparc64__ when defining LG_QUANTUM. 683 + Link tests to librt if it contains clock_gettime(2). 684 685 * 4.0.0 (August 17, 2015) 686 687 This version contains many speed and space optimizations, both minor and 688 major. The major themes are generalization, unification, and simplification. 689 Although many of these optimizations cause no visible behavior change, their 690 cumulative effect is substantial. 691 692 New features: 693 - Normalize size class spacing to be consistent across the complete size 694 range. By default there are four size classes per size doubling, but this 695 is now configurable via the --with-lg-size-class-group option. Also add the 696 --with-lg-page, --with-lg-page-sizes, --with-lg-quantum, and 697 --with-lg-tiny-min options, which can be used to tweak page and size class 698 settings. Impacts: 699 + Worst case performance for incrementally growing/shrinking reallocation 700 is improved because there are far fewer size classes, and therefore 701 copying happens less often. 702 + Internal fragmentation is limited to 20% for all but the smallest size 703 classes (those less than four times the quantum). (1B + 4 KiB) 704 and (1B + 4 MiB) previously suffered nearly 50% internal fragmentation. 705 + Chunk fragmentation tends to be lower because there are fewer distinct run 706 sizes to pack. 707 - Add support for explicit tcaches. The "tcache.create", "tcache.flush", and 708 "tcache.destroy" mallctls control tcache lifetime and flushing, and the 709 MALLOCX_TCACHE(tc) and MALLOCX_TCACHE_NONE flags to the *allocx() API 710 control which tcache is used for each operation. 711 - Implement per thread heap profiling, as well as the ability to 712 enable/disable heap profiling on a per thread basis. Add the "prof.reset", 713 "prof.lg_sample", "thread.prof.name", "thread.prof.active", 714 "opt.prof_thread_active_init", "prof.thread_active_init", and 715 "thread.prof.active" mallctls. 716 - Add support for per arena application-specified chunk allocators, configured 717 via the "arena.<i>.chunk_hooks" mallctl. 718 - Refactor huge allocation to be managed by arenas, so that arenas now 719 function as general purpose independent allocators. This is important in 720 the context of user-specified chunk allocators, aside from the scalability 721 benefits. Related new statistics: 722 + The "stats.arenas.<i>.huge.allocated", "stats.arenas.<i>.huge.nmalloc", 723 "stats.arenas.<i>.huge.ndalloc", and "stats.arenas.<i>.huge.nrequests" 724 mallctls provide high level per arena huge allocation statistics. 725 + The "arenas.nhchunks", "arenas.hchunk.<i>.size", 726 "stats.arenas.<i>.hchunks.<j>.nmalloc", 727 "stats.arenas.<i>.hchunks.<j>.ndalloc", 728 "stats.arenas.<i>.hchunks.<j>.nrequests", and 729 "stats.arenas.<i>.hchunks.<j>.curhchunks" mallctls provide per size class 730 statistics. 731 - Add the 'util' column to malloc_stats_print() output, which reports the 732 proportion of available regions that are currently in use for each small 733 size class. 734 - Add "alloc" and "free" modes for for junk filling (see the "opt.junk" 735 mallctl), so that it is possible to separately enable junk filling for 736 allocation versus deallocation. 737 - Add the jemalloc-config script, which provides information about how 738 jemalloc was configured, and how to integrate it into application builds. 739 - Add metadata statistics, which are accessible via the "stats.metadata", 740 "stats.arenas.<i>.metadata.mapped", and 741 "stats.arenas.<i>.metadata.allocated" mallctls. 742 - Add the "stats.resident" mallctl, which reports the upper limit of 743 physically resident memory mapped by the allocator. 744 - Add per arena control over unused dirty page purging, via the 745 "arenas.lg_dirty_mult", "arena.<i>.lg_dirty_mult", and 746 "stats.arenas.<i>.lg_dirty_mult" mallctls. 747 - Add the "prof.gdump" mallctl, which makes it possible to toggle the gdump 748 feature on/off during program execution. 749 - Add sdallocx(), which implements sized deallocation. The primary 750 optimization over dallocx() is the removal of a metadata read, which often 751 suffers an L1 cache miss. 752 - Add missing header includes in jemalloc/jemalloc.h, so that applications 753 only have to #include <jemalloc/jemalloc.h>. 754 - Add support for additional platforms: 755 + Bitrig 756 + Cygwin 757 + DragonFlyBSD 758 + iOS 759 + OpenBSD 760 + OpenRISC/or1k 761 762 Optimizations: 763 - Maintain dirty runs in per arena LRUs rather than in per arena trees of 764 dirty-run-containing chunks. In practice this change significantly reduces 765 dirty page purging volume. 766 - Integrate whole chunks into the unused dirty page purging machinery. This 767 reduces the cost of repeated huge allocation/deallocation, because it 768 effectively introduces a cache of chunks. 769 - Split the arena chunk map into two separate arrays, in order to increase 770 cache locality for the frequently accessed bits. 771 - Move small run metadata out of runs, into arena chunk headers. This reduces 772 run fragmentation, smaller runs reduce external fragmentation for small size 773 classes, and packed (less uniformly aligned) metadata layout improves CPU 774 cache set distribution. 775 - Randomly distribute large allocation base pointer alignment relative to page 776 boundaries in order to more uniformly utilize CPU cache sets. This can be 777 disabled via the --disable-cache-oblivious configure option, and queried via 778 the "config.cache_oblivious" mallctl. 779 - Micro-optimize the fast paths for the public API functions. 780 - Refactor thread-specific data to reside in a single structure. This assures 781 that only a single TLS read is necessary per call into the public API. 782 - Implement in-place huge allocation growing and shrinking. 783 - Refactor rtree (radix tree for chunk lookups) to be lock-free, and make 784 additional optimizations that reduce maximum lookup depth to one or two 785 levels. This resolves what was a concurrency bottleneck for per arena huge 786 allocation, because a global data structure is critical for determining 787 which arenas own which huge allocations. 788 789 Incompatible changes: 790 - Replace --enable-cc-silence with --disable-cc-silence to suppress spurious 791 warnings by default. 792 - Assure that the constness of malloc_usable_size()'s return type matches that 793 of the system implementation. 794 - Change the heap profile dump format to support per thread heap profiling, 795 rename pprof to jeprof, and enhance it with the --thread=<n> option. As a 796 result, the bundled jeprof must now be used rather than the upstream 797 (gperftools) pprof. 798 - Disable "opt.prof_final" by default, in order to avoid atexit(3), which can 799 internally deadlock on some platforms. 800 - Change the "arenas.nlruns" mallctl type from size_t to unsigned. 801 - Replace the "stats.arenas.<i>.bins.<j>.allocated" mallctl with 802 "stats.arenas.<i>.bins.<j>.curregs". 803 - Ignore MALLOC_CONF in set{uid,gid,cap} binaries. 804 - Ignore MALLOCX_ARENA(a) in dallocx(), in favor of using the 805 MALLOCX_TCACHE(tc) and MALLOCX_TCACHE_NONE flags to control tcache usage. 806 807 Removed features: 808 - Remove the *allocm() API, which is superseded by the *allocx() API. 809 - Remove the --enable-dss options, and make dss non-optional on all platforms 810 which support sbrk(2). 811 - Remove the "arenas.purge" mallctl, which was obsoleted by the 812 "arena.<i>.purge" mallctl in 3.1.0. 813 - Remove the unnecessary "opt.valgrind" mallctl; jemalloc automatically 814 detects whether it is running inside Valgrind. 815 - Remove the "stats.huge.allocated", "stats.huge.nmalloc", and 816 "stats.huge.ndalloc" mallctls. 817 - Remove the --enable-mremap option. 818 - Remove the "stats.chunks.current", "stats.chunks.total", and 819 "stats.chunks.high" mallctls. 820 821 Bug fixes: 822 - Fix the cactive statistic to decrease (rather than increase) when active 823 memory decreases. This regression was first released in 3.5.0. 824 - Fix OOM handling in memalign() and valloc(). A variant of this bug existed 825 in all releases since 2.0.0, which introduced these functions. 826 - Fix an OOM-related regression in arena_tcache_fill_small(), which could 827 cause cache corruption on OOM. This regression was present in all releases 828 from 2.2.0 through 3.6.0. 829 - Fix size class overflow handling for malloc(), posix_memalign(), memalign(), 830 calloc(), and realloc() when profiling is enabled. 831 - Fix the "arena.<i>.dss" mallctl to return an error if "primary" or 832 "secondary" precedence is specified, but sbrk(2) is not supported. 833 - Fix fallback lg_floor() implementations to handle extremely large inputs. 834 - Ensure the default purgeable zone is after the default zone on OS X. 835 - Fix latent bugs in atomic_*(). 836 - Fix the "arena.<i>.dss" mallctl to handle read-only calls. 837 - Fix tls_model configuration to enable the initial-exec model when possible. 838 - Mark malloc_conf as a weak symbol so that the application can override it. 839 - Correctly detect glibc's adaptive pthread mutexes. 840 - Fix the --without-export configure option. 841 842 * 3.6.0 (March 31, 2014) 843 844 This version contains a critical bug fix for a regression present in 3.5.0 and 845 3.5.1. 846 847 Bug fixes: 848 - Fix a regression in arena_chunk_alloc() that caused crashes during 849 small/large allocation if chunk allocation failed. In the absence of this 850 bug, chunk allocation failure would result in allocation failure, e.g. NULL 851 return from malloc(). This regression was introduced in 3.5.0. 852 - Fix backtracing for gcc intrinsics-based backtracing by specifying 853 -fno-omit-frame-pointer to gcc. Note that the application (and all the 854 libraries it links to) must also be compiled with this option for 855 backtracing to be reliable. 856 - Use dss allocation precedence for huge allocations as well as small/large 857 allocations. 858 - Fix test assertion failure message formatting. This bug did not manifest on 859 x86_64 systems because of implementation subtleties in va_list. 860 - Fix inconsequential test failures for hash and SFMT code. 861 862 New features: 863 - Support heap profiling on FreeBSD. This feature depends on the proc 864 filesystem being mounted during heap profile dumping. 865 866 * 3.5.1 (February 25, 2014) 867 868 This version primarily addresses minor bugs in test code. 869 870 Bug fixes: 871 - Configure Solaris/Illumos to use MADV_FREE. 872 - Fix junk filling for mremap(2)-based huge reallocation. This is only 873 relevant if configuring with the --enable-mremap option specified. 874 - Avoid compilation failure if 'restrict' C99 keyword is not supported by the 875 compiler. 876 - Add a configure test for SSE2 rather than assuming it is usable on i686 877 systems. This fixes test compilation errors, especially on 32-bit Linux 878 systems. 879 - Fix mallctl argument size mismatches (size_t vs. uint64_t) in the stats unit 880 test. 881 - Fix/remove flawed alignment-related overflow tests. 882 - Prevent compiler optimizations that could change backtraces in the 883 prof_accum unit test. 884 885 * 3.5.0 (January 22, 2014) 886 887 This version focuses on refactoring and automated testing, though it also 888 includes some non-trivial heap profiling optimizations not mentioned below. 889 890 New features: 891 - Add the *allocx() API, which is a successor to the experimental *allocm() 892 API. The *allocx() functions are slightly simpler to use because they have 893 fewer parameters, they directly return the results of primary interest, and 894 mallocx()/rallocx() avoid the strict aliasing pitfall that 895 allocm()/rallocm() share with posix_memalign(). Note that *allocm() is 896 slated for removal in the next non-bugfix release. 897 - Add support for LinuxThreads. 898 899 Bug fixes: 900 - Unless heap profiling is enabled, disable floating point code and don't link 901 with libm. This, in combination with e.g. EXTRA_CFLAGS=-mno-sse on x64 902 systems, makes it possible to completely disable floating point register 903 use. Some versions of glibc neglect to save/restore caller-saved floating 904 point registers during dynamic lazy symbol loading, and the symbol loading 905 code uses whatever malloc the application happens to have linked/loaded 906 with, the result being potential floating point register corruption. 907 - Report ENOMEM rather than EINVAL if an OOM occurs during heap profiling 908 backtrace creation in imemalign(). This bug impacted posix_memalign() and 909 aligned_alloc(). 910 - Fix a file descriptor leak in a prof_dump_maps() error path. 911 - Fix prof_dump() to close the dump file descriptor for all relevant error 912 paths. 913 - Fix rallocm() to use the arena specified by the ALLOCM_ARENA(s) flag for 914 allocation, not just deallocation. 915 - Fix a data race for large allocation stats counters. 916 - Fix a potential infinite loop during thread exit. This bug occurred on 917 Solaris, and could affect other platforms with similar pthreads TSD 918 implementations. 919 - Don't junk-fill reallocations unless usable size changes. This fixes a 920 violation of the *allocx()/*allocm() semantics. 921 - Fix growing large reallocation to junk fill new space. 922 - Fix huge deallocation to junk fill when munmap is disabled. 923 - Change the default private namespace prefix from empty to je_, and change 924 --with-private-namespace-prefix so that it prepends an additional prefix 925 rather than replacing je_. This reduces the likelihood of applications 926 which statically link jemalloc experiencing symbol name collisions. 927 - Add missing private namespace mangling (relevant when 928 --with-private-namespace is specified). 929 - Add and use JEMALLOC_INLINE_C so that static inline functions are marked as 930 static even for debug builds. 931 - Add a missing mutex unlock in a malloc_init_hard() error path. In practice 932 this error path is never executed. 933 - Fix numerous bugs in malloc_strotumax() error handling/reporting. These 934 bugs had no impact except for malformed inputs. 935 - Fix numerous bugs in malloc_snprintf(). These bugs were not exercised by 936 existing calls, so they had no impact. 937 938 * 3.4.1 (October 20, 2013) 939 940 Bug fixes: 941 - Fix a race in the "arenas.extend" mallctl that could cause memory corruption 942 of internal data structures and subsequent crashes. 943 - Fix Valgrind integration flaws that caused Valgrind warnings about reads of 944 uninitialized memory in: 945 + arena chunk headers 946 + internal zero-initialized data structures (relevant to tcache and prof 947 code) 948 - Preserve errno during the first allocation. A readlink(2) call during 949 initialization fails unless /etc/malloc.conf exists, so errno was typically 950 set during the first allocation prior to this fix. 951 - Fix compilation warnings reported by gcc 4.8.1. 952 953 * 3.4.0 (June 2, 2013) 954 955 This version is essentially a small bugfix release, but the addition of 956 aarch64 support requires that the minor version be incremented. 957 958 Bug fixes: 959 - Fix race-triggered deadlocks in chunk_record(). These deadlocks were 960 typically triggered by multiple threads concurrently deallocating huge 961 objects. 962 963 New features: 964 - Add support for the aarch64 architecture. 965 966 * 3.3.1 (March 6, 2013) 967 968 This version fixes bugs that are typically encountered only when utilizing 969 custom run-time options. 970 971 Bug fixes: 972 - Fix a locking order bug that could cause deadlock during fork if heap 973 profiling were enabled. 974 - Fix a chunk recycling bug that could cause the allocator to lose track of 975 whether a chunk was zeroed. On FreeBSD, NetBSD, and OS X, it could cause 976 corruption if allocating via sbrk(2) (unlikely unless running with the 977 "dss:primary" option specified). This was completely harmless on Linux 978 unless using mlockall(2) (and unlikely even then, unless the 979 --disable-munmap configure option or the "dss:primary" option was 980 specified). This regression was introduced in 3.1.0 by the 981 mlockall(2)/madvise(2) interaction fix. 982 - Fix TLS-related memory corruption that could occur during thread exit if the 983 thread never allocated memory. Only the quarantine and prof facilities were 984 susceptible. 985 - Fix two quarantine bugs: 986 + Internal reallocation of the quarantined object array leaked the old 987 array. 988 + Reallocation failure for internal reallocation of the quarantined object 989 array (very unlikely) resulted in memory corruption. 990 - Fix Valgrind integration to annotate all internally allocated memory in a 991 way that keeps Valgrind happy about internal data structure access. 992 - Fix building for s390 systems. 993 994 * 3.3.0 (January 23, 2013) 995 996 This version includes a few minor performance improvements in addition to the 997 listed new features and bug fixes. 998 999 New features: 1000 - Add clipping support to lg_chunk option processing. 1001 - Add the --enable-ivsalloc option. 1002 - Add the --without-export option. 1003 - Add the --disable-zone-allocator option. 1004 1005 Bug fixes: 1006 - Fix "arenas.extend" mallctl to output the number of arenas. 1007 - Fix chunk_recycle() to unconditionally inform Valgrind that returned memory 1008 is undefined. 1009 - Fix build break on FreeBSD related to alloca.h. 1010 1011 * 3.2.0 (November 9, 2012) 1012 1013 In addition to a couple of bug fixes, this version modifies page run 1014 allocation and dirty page purging algorithms in order to better control 1015 page-level virtual memory fragmentation. 1016 1017 Incompatible changes: 1018 - Change the "opt.lg_dirty_mult" default from 5 to 3 (32:1 to 8:1). 1019 1020 Bug fixes: 1021 - Fix dss/mmap allocation precedence code to use recyclable mmap memory only 1022 after primary dss allocation fails. 1023 - Fix deadlock in the "arenas.purge" mallctl. This regression was introduced 1024 in 3.1.0 by the addition of the "arena.<i>.purge" mallctl. 1025 1026 * 3.1.0 (October 16, 2012) 1027 1028 New features: 1029 - Auto-detect whether running inside Valgrind, thus removing the need to 1030 manually specify MALLOC_CONF=valgrind:true. 1031 - Add the "arenas.extend" mallctl, which allows applications to create 1032 manually managed arenas. 1033 - Add the ALLOCM_ARENA() flag for {,r,d}allocm(). 1034 - Add the "opt.dss", "arena.<i>.dss", and "stats.arenas.<i>.dss" mallctls, 1035 which provide control over dss/mmap precedence. 1036 - Add the "arena.<i>.purge" mallctl, which obsoletes "arenas.purge". 1037 - Define LG_QUANTUM for hppa. 1038 1039 Incompatible changes: 1040 - Disable tcache by default if running inside Valgrind, in order to avoid 1041 making unallocated objects appear reachable to Valgrind. 1042 - Drop const from malloc_usable_size() argument on Linux. 1043 1044 Bug fixes: 1045 - Fix heap profiling crash if sampled object is freed via realloc(p, 0). 1046 - Remove const from __*_hook variable declarations, so that glibc can modify 1047 them during process forking. 1048 - Fix mlockall(2)/madvise(2) interaction. 1049 - Fix fork(2)-related deadlocks. 1050 - Fix error return value for "thread.tcache.enabled" mallctl. 1051 1052 * 3.0.0 (May 11, 2012) 1053 1054 Although this version adds some major new features, the primary focus is on 1055 internal code cleanup that facilitates maintainability and portability, most 1056 of which is not reflected in the ChangeLog. This is the first release to 1057 incorporate substantial contributions from numerous other developers, and the 1058 result is a more broadly useful allocator (see the git revision history for 1059 contribution details). Note that the license has been unified, thanks to 1060 Facebook granting a license under the same terms as the other copyright 1061 holders (see COPYING). 1062 1063 New features: 1064 - Implement Valgrind support, redzones, and quarantine. 1065 - Add support for additional platforms: 1066 + FreeBSD 1067 + Mac OS X Lion 1068 + MinGW 1069 + Windows (no support yet for replacing the system malloc) 1070 - Add support for additional architectures: 1071 + MIPS 1072 + SH4 1073 + Tilera 1074 - Add support for cross compiling. 1075 - Add nallocm(), which rounds a request size up to the nearest size class 1076 without actually allocating. 1077 - Implement aligned_alloc() (blame C11). 1078 - Add the "thread.tcache.enabled" mallctl. 1079 - Add the "opt.prof_final" mallctl. 1080 - Update pprof (from gperftools 2.0). 1081 - Add the --with-mangling option. 1082 - Add the --disable-experimental option. 1083 - Add the --disable-munmap option, and make it the default on Linux. 1084 - Add the --enable-mremap option, which disables use of mremap(2) by default. 1085 1086 Incompatible changes: 1087 - Enable stats by default. 1088 - Enable fill by default. 1089 - Disable lazy locking by default. 1090 - Rename the "tcache.flush" mallctl to "thread.tcache.flush". 1091 - Rename the "arenas.pagesize" mallctl to "arenas.page". 1092 - Change the "opt.lg_prof_sample" default from 0 to 19 (1 B to 512 KiB). 1093 - Change the "opt.prof_accum" default from true to false. 1094 1095 Removed features: 1096 - Remove the swap feature, including the "config.swap", "swap.avail", 1097 "swap.prezeroed", "swap.nfds", and "swap.fds" mallctls. 1098 - Remove highruns statistics, including the 1099 "stats.arenas.<i>.bins.<j>.highruns" and 1100 "stats.arenas.<i>.lruns.<j>.highruns" mallctls. 1101 - As part of small size class refactoring, remove the "opt.lg_[qc]space_max", 1102 "arenas.cacheline", "arenas.subpage", "arenas.[tqcs]space_{min,max}", and 1103 "arenas.[tqcs]bins" mallctls. 1104 - Remove the "arenas.chunksize" mallctl. 1105 - Remove the "opt.lg_prof_tcmax" option. 1106 - Remove the "opt.lg_prof_bt_max" option. 1107 - Remove the "opt.lg_tcache_gc_sweep" option. 1108 - Remove the --disable-tiny option, including the "config.tiny" mallctl. 1109 - Remove the --enable-dynamic-page-shift configure option. 1110 - Remove the --enable-sysv configure option. 1111 1112 Bug fixes: 1113 - Fix a statistics-related bug in the "thread.arena" mallctl that could cause 1114 invalid statistics and crashes. 1115 - Work around TLS deallocation via free() on Linux. This bug could cause 1116 write-after-free memory corruption. 1117 - Fix a potential deadlock that could occur during interval- and 1118 growth-triggered heap profile dumps. 1119 - Fix large calloc() zeroing bugs due to dropping chunk map unzeroed flags. 1120 - Fix chunk_alloc_dss() to stop claiming memory is zeroed. This bug could 1121 cause memory corruption and crashes with --enable-dss specified. 1122 - Fix fork-related bugs that could cause deadlock in children between fork 1123 and exec. 1124 - Fix malloc_stats_print() to honor 'b' and 'l' in the opts parameter. 1125 - Fix realloc(p, 0) to act like free(p). 1126 - Do not enforce minimum alignment in memalign(). 1127 - Check for NULL pointer in malloc_usable_size(). 1128 - Fix an off-by-one heap profile statistics bug that could be observed in 1129 interval- and growth-triggered heap profiles. 1130 - Fix the "epoch" mallctl to update cached stats even if the passed in epoch 1131 is 0. 1132 - Fix bin->runcur management to fix a layout policy bug. This bug did not 1133 affect correctness. 1134 - Fix a bug in choose_arena_hard() that potentially caused more arenas to be 1135 initialized than necessary. 1136 - Add missing "opt.lg_tcache_max" mallctl implementation. 1137 - Use glibc allocator hooks to make mixed allocator usage less likely. 1138 - Fix build issues for --disable-tcache. 1139 - Don't mangle pthread_create() when --with-private-namespace is specified. 1140 1141 * 2.2.5 (November 14, 2011) 1142 1143 Bug fixes: 1144 - Fix huge_ralloc() race when using mremap(2). This is a serious bug that 1145 could cause memory corruption and/or crashes. 1146 - Fix huge_ralloc() to maintain chunk statistics. 1147 - Fix malloc_stats_print(..., "a") output. 1148 1149 * 2.2.4 (November 5, 2011) 1150 1151 Bug fixes: 1152 - Initialize arenas_tsd before using it. This bug existed for 2.2.[0-3], as 1153 well as for --disable-tls builds in earlier releases. 1154 - Do not assume a 4 KiB page size in test/rallocm.c. 1155 1156 * 2.2.3 (August 31, 2011) 1157 1158 This version fixes numerous bugs related to heap profiling. 1159 1160 Bug fixes: 1161 - Fix a prof-related race condition. This bug could cause memory corruption, 1162 but only occurred in non-default configurations (prof_accum:false). 1163 - Fix off-by-one backtracing issues (make sure that prof_alloc_prep() is 1164 excluded from backtraces). 1165 - Fix a prof-related bug in realloc() (only triggered by OOM errors). 1166 - Fix prof-related bugs in allocm() and rallocm(). 1167 - Fix prof_tdata_cleanup() for --disable-tls builds. 1168 - Fix a relative include path, to fix objdir builds. 1169 1170 * 2.2.2 (July 30, 2011) 1171 1172 Bug fixes: 1173 - Fix a build error for --disable-tcache. 1174 - Fix assertions in arena_purge() (for real this time). 1175 - Add the --with-private-namespace option. This is a workaround for symbol 1176 conflicts that can inadvertently arise when using static libraries. 1177 1178 * 2.2.1 (March 30, 2011) 1179 1180 Bug fixes: 1181 - Implement atomic operations for x86/x64. This fixes compilation failures 1182 for versions of gcc that are still in wide use. 1183 - Fix an assertion in arena_purge(). 1184 1185 * 2.2.0 (March 22, 2011) 1186 1187 This version incorporates several improvements to algorithms and data 1188 structures that tend to reduce fragmentation and increase speed. 1189 1190 New features: 1191 - Add the "stats.cactive" mallctl. 1192 - Update pprof (from google-perftools 1.7). 1193 - Improve backtracing-related configuration logic, and add the 1194 --disable-prof-libgcc option. 1195 1196 Bug fixes: 1197 - Change default symbol visibility from "internal", to "hidden", which 1198 decreases the overhead of library-internal function calls. 1199 - Fix symbol visibility so that it is also set on OS X. 1200 - Fix a build dependency regression caused by the introduction of the .pic.o 1201 suffix for PIC object files. 1202 - Add missing checks for mutex initialization failures. 1203 - Don't use libgcc-based backtracing except on x64, where it is known to work. 1204 - Fix deadlocks on OS X that were due to memory allocation in 1205 pthread_mutex_lock(). 1206 - Heap profiling-specific fixes: 1207 + Fix memory corruption due to integer overflow in small region index 1208 computation, when using a small enough sample interval that profiling 1209 context pointers are stored in small run headers. 1210 + Fix a bootstrap ordering bug that only occurred with TLS disabled. 1211 + Fix a rallocm() rsize bug. 1212 + Fix error detection bugs for aligned memory allocation. 1213 1214 * 2.1.3 (March 14, 2011) 1215 1216 Bug fixes: 1217 - Fix a cpp logic regression (due to the "thread.{de,}allocatedp" mallctl fix 1218 for OS X in 2.1.2). 1219 - Fix a "thread.arena" mallctl bug. 1220 - Fix a thread cache stats merging bug. 1221 1222 * 2.1.2 (March 2, 2011) 1223 1224 Bug fixes: 1225 - Fix "thread.{de,}allocatedp" mallctl for OS X. 1226 - Add missing jemalloc.a to build system. 1227 1228 * 2.1.1 (January 31, 2011) 1229 1230 Bug fixes: 1231 - Fix aligned huge reallocation (affected allocm()). 1232 - Fix the ALLOCM_LG_ALIGN macro definition. 1233 - Fix a heap dumping deadlock. 1234 - Fix a "thread.arena" mallctl bug. 1235 1236 * 2.1.0 (December 3, 2010) 1237 1238 This version incorporates some optimizations that can't quite be considered 1239 bug fixes. 1240 1241 New features: 1242 - Use Linux's mremap(2) for huge object reallocation when possible. 1243 - Avoid locking in mallctl*() when possible. 1244 - Add the "thread.[de]allocatedp" mallctl's. 1245 - Convert the manual page source from roff to DocBook, and generate both roff 1246 and HTML manuals. 1247 1248 Bug fixes: 1249 - Fix a crash due to incorrect bootstrap ordering. This only impacted 1250 --enable-debug --enable-dss configurations. 1251 - Fix a minor statistics bug for mallctl("swap.avail", ...). 1252 1253 * 2.0.1 (October 29, 2010) 1254 1255 Bug fixes: 1256 - Fix a race condition in heap profiling that could cause undefined behavior 1257 if "opt.prof_accum" were disabled. 1258 - Add missing mutex unlocks for some OOM error paths in the heap profiling 1259 code. 1260 - Fix a compilation error for non-C99 builds. 1261 1262 * 2.0.0 (October 24, 2010) 1263 1264 This version focuses on the experimental *allocm() API, and on improved 1265 run-time configuration/introspection. Nonetheless, numerous performance 1266 improvements are also included. 1267 1268 New features: 1269 - Implement the experimental {,r,s,d}allocm() API, which provides a superset 1270 of the functionality available via malloc(), calloc(), posix_memalign(), 1271 realloc(), malloc_usable_size(), and free(). These functions can be used to 1272 allocate/reallocate aligned zeroed memory, ask for optional extra memory 1273 during reallocation, prevent object movement during reallocation, etc. 1274 - Replace JEMALLOC_OPTIONS/JEMALLOC_PROF_PREFIX with MALLOC_CONF, which is 1275 more human-readable, and more flexible. For example: 1276 JEMALLOC_OPTIONS=AJP 1277 is now: 1278 MALLOC_CONF=abort:true,fill:true,stats_print:true 1279 - Port to Apple OS X. Sponsored by Mozilla. 1280 - Make it possible for the application to control thread-->arena mappings via 1281 the "thread.arena" mallctl. 1282 - Add compile-time support for all TLS-related functionality via pthreads TSD. 1283 This is mainly of interest for OS X, which does not support TLS, but has a 1284 TSD implementation with similar performance. 1285 - Override memalign() and valloc() if they are provided by the system. 1286 - Add the "arenas.purge" mallctl, which can be used to synchronously purge all 1287 dirty unused pages. 1288 - Make cumulative heap profiling data optional, so that it is possible to 1289 limit the amount of memory consumed by heap profiling data structures. 1290 - Add per thread allocation counters that can be accessed via the 1291 "thread.allocated" and "thread.deallocated" mallctls. 1292 1293 Incompatible changes: 1294 - Remove JEMALLOC_OPTIONS and malloc_options (see MALLOC_CONF above). 1295 - Increase default backtrace depth from 4 to 128 for heap profiling. 1296 - Disable interval-based profile dumps by default. 1297 1298 Bug fixes: 1299 - Remove bad assertions in fork handler functions. These assertions could 1300 cause aborts for some combinations of configure settings. 1301 - Fix strerror_r() usage to deal with non-standard semantics in GNU libc. 1302 - Fix leak context reporting. This bug tended to cause the number of contexts 1303 to be underreported (though the reported number of objects and bytes were 1304 correct). 1305 - Fix a realloc() bug for large in-place growing reallocation. This bug could 1306 cause memory corruption, but it was hard to trigger. 1307 - Fix an allocation bug for small allocations that could be triggered if 1308 multiple threads raced to create a new run of backing pages. 1309 - Enhance the heap profiler to trigger samples based on usable size, rather 1310 than request size. 1311 - Fix a heap profiling bug due to sometimes losing track of requested object 1312 size for sampled objects. 1313 1314 * 1.0.3 (August 12, 2010) 1315 1316 Bug fixes: 1317 - Fix the libunwind-based implementation of stack backtracing (used for heap 1318 profiling). This bug could cause zero-length backtraces to be reported. 1319 - Add a missing mutex unlock in library initialization code. If multiple 1320 threads raced to initialize malloc, some of them could end up permanently 1321 blocked. 1322 1323 * 1.0.2 (May 11, 2010) 1324 1325 Bug fixes: 1326 - Fix junk filling of large objects, which could cause memory corruption. 1327 - Add MAP_NORESERVE support for chunk mapping, because otherwise virtual 1328 memory limits could cause swap file configuration to fail. Contributed by 1329 Jordan DeLong. 1330 1331 * 1.0.1 (April 14, 2010) 1332 1333 Bug fixes: 1334 - Fix compilation when --enable-fill is specified. 1335 - Fix threads-related profiling bugs that affected accuracy and caused memory 1336 to be leaked during thread exit. 1337 - Fix dirty page purging race conditions that could cause crashes. 1338 - Fix crash in tcache flushing code during thread destruction. 1339 1340 * 1.0.0 (April 11, 2010) 1341 1342 This release focuses on speed and run-time introspection. Numerous 1343 algorithmic improvements make this release substantially faster than its 1344 predecessors. 1345 1346 New features: 1347 - Implement autoconf-based configuration system. 1348 - Add mallctl*(), for the purposes of introspection and run-time 1349 configuration. 1350 - Make it possible for the application to manually flush a thread's cache, via 1351 the "tcache.flush" mallctl. 1352 - Base maximum dirty page count on proportion of active memory. 1353 - Compute various additional run-time statistics, including per size class 1354 statistics for large objects. 1355 - Expose malloc_stats_print(), which can be called repeatedly by the 1356 application. 1357 - Simplify the malloc_message() signature to only take one string argument, 1358 and incorporate an opaque data pointer argument for use by the application 1359 in combination with malloc_stats_print(). 1360 - Add support for allocation backed by one or more swap files, and allow the 1361 application to disable over-commit if swap files are in use. 1362 - Implement allocation profiling and leak checking. 1363 1364 Removed features: 1365 - Remove the dynamic arena rebalancing code, since thread-specific caching 1366 reduces its utility. 1367 1368 Bug fixes: 1369 - Modify chunk allocation to work when address space layout randomization 1370 (ASLR) is in use. 1371 - Fix thread cleanup bugs related to TLS destruction. 1372 - Handle 0-size allocation requests in posix_memalign(). 1373 - Fix a chunk leak. The leaked chunks were never touched, so this impacted 1374 virtual memory usage, but not physical memory usage. 1375 1376 * linux_2008082[78]a (August 27/28, 2008) 1377 1378 These snapshot releases are the simple result of incorporating Linux-specific 1379 support into the FreeBSD malloc sources. 1380 1381 -------------------------------------------------------------------------------- 1382 vim:filetype=text:textwidth=80 1383
