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 * 4.1.0 (February 28, 2016) 8 9 This release is primarily about optimizations, but it also incorporates a lot 10 of portability-motivated refactoring and enhancements. Many people worked on 11 this release, to an extent that even with the omission here of minor changes 12 (see git revision history), and of the people who reported and diagnosed 13 issues, so much of the work was contributed that starting with this release, 14 changes are annotated with author credits to help reflect the collaborative 15 effort involved. 16 17 New features: 18 - Implement decay-based unused dirty page purging, a major optimization with 19 mallctl API impact. This is an alternative to the existing ratio-based 20 unused dirty page purging, and is intended to eventually become the sole 21 purging mechanism. New mallctls: 22 + opt.purge 23 + opt.decay_time 24 + arena.<i>.decay 25 + arena.<i>.decay_time 26 + arenas.decay_time 27 + stats.arenas.<i>.decay_time 28 (@jasone, @cevans87) 29 - Add --with-malloc-conf, which makes it possible to embed a default 30 options string during configuration. This was motivated by the desire to 31 specify --with-malloc-conf=purge:decay , since the default must remain 32 purge:ratio until the 5.0.0 release. (@jasone) 33 - Add MS Visual Studio 2015 support. (@rustyx, @yuslepukhin) 34 - Make *allocx() size class overflow behavior defined. The maximum 35 size class is now less than PTRDIFF_MAX to protect applications against 36 numerical overflow, and all allocation functions are guaranteed to indicate 37 errors rather than potentially crashing if the request size exceeds the 38 maximum size class. (@jasone) 39 - jeprof: 40 + Add raw heap profile support. (@jasone) 41 + Add --retain and --exclude for backtrace symbol filtering. (@jasone) 42 43 Optimizations: 44 - Optimize the fast path to combine various bootstrapping and configuration 45 checks and execute more streamlined code in the common case. (@interwq) 46 - Use linear scan for small bitmaps (used for small object tracking). In 47 addition to speeding up bitmap operations on 64-bit systems, this reduces 48 allocator metadata overhead by approximately 0.2%. (@djwatson) 49 - Separate arena_avail trees, which substantially speeds up run tree 50 operations. (@djwatson) 51 - Use memoization (boot-time-computed table) for run quantization. Separate 52 arena_avail trees reduced the importance of this optimization. (@jasone) 53 - Attempt mmap-based in-place huge reallocation. This can dramatically speed 54 up incremental huge reallocation. (@jasone) 55 56 Incompatible changes: 57 - Make opt.narenas unsigned rather than size_t. (@jasone) 58 59 Bug fixes: 60 - Fix stats.cactive accounting regression. (@rustyx, @jasone) 61 - Handle unaligned keys in hash(). This caused problems for some ARM systems. 62 (@jasone, Christopher Ferris) 63 - Refactor arenas array. In addition to fixing a fork-related deadlock, this 64 makes arena lookups faster and simpler. (@jasone) 65 - Move retained memory allocation out of the default chunk allocation 66 function, to a location that gets executed even if the application installs 67 a custom chunk allocation function. This resolves a virtual memory leak. 68 (@buchgr) 69 - Fix a potential tsd cleanup leak. (Christopher Ferris, @jasone) 70 - Fix run quantization. In practice this bug had no impact unless 71 applications requested memory with alignment exceeding one page. 72 (@jasone, @djwatson) 73 - Fix LinuxThreads-specific bootstrapping deadlock. (Cosmin Paraschiv) 74 - jeprof: 75 + Don't discard curl options if timeout is not defined. (@djwatson) 76 + Detect failed profile fetches. (@djwatson) 77 - Fix stats.arenas.<i>.{dss,lg_dirty_mult,decay_time,pactive,pdirty} for 78 --disable-stats case. (@jasone) 79 80 * 4.0.4 (October 24, 2015) 81 82 This bugfix release fixes another xallocx() regression. No other regressions 83 have come to light in over a month, so this is likely a good starting point 84 for people who prefer to wait for "dot one" releases with all the major issues 85 shaken out. 86 87 Bug fixes: 88 - Fix xallocx(..., MALLOCX_ZERO to zero the last full trailing page of large 89 allocations that have been randomly assigned an offset of 0 when 90 --enable-cache-oblivious configure option is enabled. 91 92 * 4.0.3 (September 24, 2015) 93 94 This bugfix release continues the trend of xallocx() and heap profiling fixes. 95 96 Bug fixes: 97 - Fix xallocx(..., MALLOCX_ZERO) to zero all trailing bytes of large 98 allocations when --enable-cache-oblivious configure option is enabled. 99 - Fix xallocx(..., MALLOCX_ZERO) to zero trailing bytes of huge allocations 100 when resizing from/to a size class that is not a multiple of the chunk size. 101 - Fix prof_tctx_dump_iter() to filter out nodes that were created after heap 102 profile dumping started. 103 - Work around a potentially bad thread-specific data initialization 104 interaction with NPTL (glibc's pthreads implementation). 105 106 * 4.0.2 (September 21, 2015) 107 108 This bugfix release addresses a few bugs specific to heap profiling. 109 110 Bug fixes: 111 - Fix ixallocx_prof_sample() to never modify nor create sampled small 112 allocations. xallocx() is in general incapable of moving small allocations, 113 so this fix removes buggy code without loss of generality. 114 - Fix irallocx_prof_sample() to always allocate large regions, even when 115 alignment is non-zero. 116 - Fix prof_alloc_rollback() to read tdata from thread-specific data rather 117 than dereferencing a potentially invalid tctx. 118 119 * 4.0.1 (September 15, 2015) 120 121 This is a bugfix release that is somewhat high risk due to the amount of 122 refactoring required to address deep xallocx() problems. As a side effect of 123 these fixes, xallocx() now tries harder to partially fulfill requests for 124 optional extra space. Note that a couple of minor heap profiling 125 optimizations are included, but these are better thought of as performance 126 fixes that were integral to disovering most of the other bugs. 127 128 Optimizations: 129 - Avoid a chunk metadata read in arena_prof_tctx_set(), since it is in the 130 fast path when heap profiling is enabled. Additionally, split a special 131 case out into arena_prof_tctx_reset(), which also avoids chunk metadata 132 reads. 133 - Optimize irallocx_prof() to optimistically update the sampler state. The 134 prior implementation appears to have been a holdover from when 135 rallocx()/xallocx() functionality was combined as rallocm(). 136 137 Bug fixes: 138 - Fix TLS configuration such that it is enabled by default for platforms on 139 which it works correctly. 140 - Fix arenas_cache_cleanup() and arena_get_hard() to handle 141 allocation/deallocation within the application's thread-specific data 142 cleanup functions even after arenas_cache is torn down. 143 - Fix xallocx() bugs related to size+extra exceeding HUGE_MAXCLASS. 144 - Fix chunk purge hook calls for in-place huge shrinking reallocation to 145 specify the old chunk size rather than the new chunk size. This bug caused 146 no correctness issues for the default chunk purge function, but was 147 visible to custom functions set via the "arena.<i>.chunk_hooks" mallctl. 148 - Fix heap profiling bugs: 149 + Fix heap profiling to distinguish among otherwise identical sample sites 150 with interposed resets (triggered via the "prof.reset" mallctl). This bug 151 could cause data structure corruption that would most likely result in a 152 segfault. 153 + Fix irealloc_prof() to prof_alloc_rollback() on OOM. 154 + Make one call to prof_active_get_unlocked() per allocation event, and use 155 the result throughout the relevant functions that handle an allocation 156 event. Also add a missing check in prof_realloc(). These fixes protect 157 allocation events against concurrent prof_active changes. 158 + Fix ixallocx_prof() to pass usize_max and zero to ixallocx_prof_sample() 159 in the correct order. 160 + Fix prof_realloc() to call prof_free_sampled_object() after calling 161 prof_malloc_sample_object(). Prior to this fix, if tctx and old_tctx were 162 the same, the tctx could have been prematurely destroyed. 163 - Fix portability bugs: 164 + Don't bitshift by negative amounts when encoding/decoding run sizes in 165 chunk header maps. This affected systems with page sizes greater than 8 166 KiB. 167 + Rename index_t to szind_t to avoid an existing type on Solaris. 168 + Add JEMALLOC_CXX_THROW to the memalign() function prototype, in order to 169 match glibc and avoid compilation errors when including both 170 jemalloc/jemalloc.h and malloc.h in C++ code. 171 + Don't assume that /bin/sh is appropriate when running size_classes.sh 172 during configuration. 173 + Consider __sparcv9 a synonym for __sparc64__ when defining LG_QUANTUM. 174 + Link tests to librt if it contains clock_gettime(2). 175 176 * 4.0.0 (August 17, 2015) 177 178 This version contains many speed and space optimizations, both minor and 179 major. The major themes are generalization, unification, and simplification. 180 Although many of these optimizations cause no visible behavior change, their 181 cumulative effect is substantial. 182 183 New features: 184 - Normalize size class spacing to be consistent across the complete size 185 range. By default there are four size classes per size doubling, but this 186 is now configurable via the --with-lg-size-class-group option. Also add the 187 --with-lg-page, --with-lg-page-sizes, --with-lg-quantum, and 188 --with-lg-tiny-min options, which can be used to tweak page and size class 189 settings. Impacts: 190 + Worst case performance for incrementally growing/shrinking reallocation 191 is improved because there are far fewer size classes, and therefore 192 copying happens less often. 193 + Internal fragmentation is limited to 20% for all but the smallest size 194 classes (those less than four times the quantum). (1B + 4 KiB) 195 and (1B + 4 MiB) previously suffered nearly 50% internal fragmentation. 196 + Chunk fragmentation tends to be lower because there are fewer distinct run 197 sizes to pack. 198 - Add support for explicit tcaches. The "tcache.create", "tcache.flush", and 199 "tcache.destroy" mallctls control tcache lifetime and flushing, and the 200 MALLOCX_TCACHE(tc) and MALLOCX_TCACHE_NONE flags to the *allocx() API 201 control which tcache is used for each operation. 202 - Implement per thread heap profiling, as well as the ability to 203 enable/disable heap profiling on a per thread basis. Add the "prof.reset", 204 "prof.lg_sample", "thread.prof.name", "thread.prof.active", 205 "opt.prof_thread_active_init", "prof.thread_active_init", and 206 "thread.prof.active" mallctls. 207 - Add support for per arena application-specified chunk allocators, configured 208 via the "arena.<i>.chunk_hooks" mallctl. 209 - Refactor huge allocation to be managed by arenas, so that arenas now 210 function as general purpose independent allocators. This is important in 211 the context of user-specified chunk allocators, aside from the scalability 212 benefits. Related new statistics: 213 + The "stats.arenas.<i>.huge.allocated", "stats.arenas.<i>.huge.nmalloc", 214 "stats.arenas.<i>.huge.ndalloc", and "stats.arenas.<i>.huge.nrequests" 215 mallctls provide high level per arena huge allocation statistics. 216 + The "arenas.nhchunks", "arenas.hchunk.<i>.size", 217 "stats.arenas.<i>.hchunks.<j>.nmalloc", 218 "stats.arenas.<i>.hchunks.<j>.ndalloc", 219 "stats.arenas.<i>.hchunks.<j>.nrequests", and 220 "stats.arenas.<i>.hchunks.<j>.curhchunks" mallctls provide per size class 221 statistics. 222 - Add the 'util' column to malloc_stats_print() output, which reports the 223 proportion of available regions that are currently in use for each small 224 size class. 225 - Add "alloc" and "free" modes for for junk filling (see the "opt.junk" 226 mallctl), so that it is possible to separately enable junk filling for 227 allocation versus deallocation. 228 - Add the jemalloc-config script, which provides information about how 229 jemalloc was configured, and how to integrate it into application builds. 230 - Add metadata statistics, which are accessible via the "stats.metadata", 231 "stats.arenas.<i>.metadata.mapped", and 232 "stats.arenas.<i>.metadata.allocated" mallctls. 233 - Add the "stats.resident" mallctl, which reports the upper limit of 234 physically resident memory mapped by the allocator. 235 - Add per arena control over unused dirty page purging, via the 236 "arenas.lg_dirty_mult", "arena.<i>.lg_dirty_mult", and 237 "stats.arenas.<i>.lg_dirty_mult" mallctls. 238 - Add the "prof.gdump" mallctl, which makes it possible to toggle the gdump 239 feature on/off during program execution. 240 - Add sdallocx(), which implements sized deallocation. The primary 241 optimization over dallocx() is the removal of a metadata read, which often 242 suffers an L1 cache miss. 243 - Add missing header includes in jemalloc/jemalloc.h, so that applications 244 only have to #include <jemalloc/jemalloc.h>. 245 - Add support for additional platforms: 246 + Bitrig 247 + Cygwin 248 + DragonFlyBSD 249 + iOS 250 + OpenBSD 251 + OpenRISC/or1k 252 253 Optimizations: 254 - Maintain dirty runs in per arena LRUs rather than in per arena trees of 255 dirty-run-containing chunks. In practice this change significantly reduces 256 dirty page purging volume. 257 - Integrate whole chunks into the unused dirty page purging machinery. This 258 reduces the cost of repeated huge allocation/deallocation, because it 259 effectively introduces a cache of chunks. 260 - Split the arena chunk map into two separate arrays, in order to increase 261 cache locality for the frequently accessed bits. 262 - Move small run metadata out of runs, into arena chunk headers. This reduces 263 run fragmentation, smaller runs reduce external fragmentation for small size 264 classes, and packed (less uniformly aligned) metadata layout improves CPU 265 cache set distribution. 266 - Randomly distribute large allocation base pointer alignment relative to page 267 boundaries in order to more uniformly utilize CPU cache sets. This can be 268 disabled via the --disable-cache-oblivious configure option, and queried via 269 the "config.cache_oblivious" mallctl. 270 - Micro-optimize the fast paths for the public API functions. 271 - Refactor thread-specific data to reside in a single structure. This assures 272 that only a single TLS read is necessary per call into the public API. 273 - Implement in-place huge allocation growing and shrinking. 274 - Refactor rtree (radix tree for chunk lookups) to be lock-free, and make 275 additional optimizations that reduce maximum lookup depth to one or two 276 levels. This resolves what was a concurrency bottleneck for per arena huge 277 allocation, because a global data structure is critical for determining 278 which arenas own which huge allocations. 279 280 Incompatible changes: 281 - Replace --enable-cc-silence with --disable-cc-silence to suppress spurious 282 warnings by default. 283 - Assure that the constness of malloc_usable_size()'s return type matches that 284 of the system implementation. 285 - Change the heap profile dump format to support per thread heap profiling, 286 rename pprof to jeprof, and enhance it with the --thread=<n> option. As a 287 result, the bundled jeprof must now be used rather than the upstream 288 (gperftools) pprof. 289 - Disable "opt.prof_final" by default, in order to avoid atexit(3), which can 290 internally deadlock on some platforms. 291 - Change the "arenas.nlruns" mallctl type from size_t to unsigned. 292 - Replace the "stats.arenas.<i>.bins.<j>.allocated" mallctl with 293 "stats.arenas.<i>.bins.<j>.curregs". 294 - Ignore MALLOC_CONF in set{uid,gid,cap} binaries. 295 - Ignore MALLOCX_ARENA(a) in dallocx(), in favor of using the 296 MALLOCX_TCACHE(tc) and MALLOCX_TCACHE_NONE flags to control tcache usage. 297 298 Removed features: 299 - Remove the *allocm() API, which is superseded by the *allocx() API. 300 - Remove the --enable-dss options, and make dss non-optional on all platforms 301 which support sbrk(2). 302 - Remove the "arenas.purge" mallctl, which was obsoleted by the 303 "arena.<i>.purge" mallctl in 3.1.0. 304 - Remove the unnecessary "opt.valgrind" mallctl; jemalloc automatically 305 detects whether it is running inside Valgrind. 306 - Remove the "stats.huge.allocated", "stats.huge.nmalloc", and 307 "stats.huge.ndalloc" mallctls. 308 - Remove the --enable-mremap option. 309 - Remove the "stats.chunks.current", "stats.chunks.total", and 310 "stats.chunks.high" mallctls. 311 312 Bug fixes: 313 - Fix the cactive statistic to decrease (rather than increase) when active 314 memory decreases. This regression was first released in 3.5.0. 315 - Fix OOM handling in memalign() and valloc(). A variant of this bug existed 316 in all releases since 2.0.0, which introduced these functions. 317 - Fix an OOM-related regression in arena_tcache_fill_small(), which could 318 cause cache corruption on OOM. This regression was present in all releases 319 from 2.2.0 through 3.6.0. 320 - Fix size class overflow handling for malloc(), posix_memalign(), memalign(), 321 calloc(), and realloc() when profiling is enabled. 322 - Fix the "arena.<i>.dss" mallctl to return an error if "primary" or 323 "secondary" precedence is specified, but sbrk(2) is not supported. 324 - Fix fallback lg_floor() implementations to handle extremely large inputs. 325 - Ensure the default purgeable zone is after the default zone on OS X. 326 - Fix latent bugs in atomic_*(). 327 - Fix the "arena.<i>.dss" mallctl to handle read-only calls. 328 - Fix tls_model configuration to enable the initial-exec model when possible. 329 - Mark malloc_conf as a weak symbol so that the application can override it. 330 - Correctly detect glibc's adaptive pthread mutexes. 331 - Fix the --without-export configure option. 332 333 * 3.6.0 (March 31, 2014) 334 335 This version contains a critical bug fix for a regression present in 3.5.0 and 336 3.5.1. 337 338 Bug fixes: 339 - Fix a regression in arena_chunk_alloc() that caused crashes during 340 small/large allocation if chunk allocation failed. In the absence of this 341 bug, chunk allocation failure would result in allocation failure, e.g. NULL 342 return from malloc(). This regression was introduced in 3.5.0. 343 - Fix backtracing for gcc intrinsics-based backtracing by specifying 344 -fno-omit-frame-pointer to gcc. Note that the application (and all the 345 libraries it links to) must also be compiled with this option for 346 backtracing to be reliable. 347 - Use dss allocation precedence for huge allocations as well as small/large 348 allocations. 349 - Fix test assertion failure message formatting. This bug did not manifest on 350 x86_64 systems because of implementation subtleties in va_list. 351 - Fix inconsequential test failures for hash and SFMT code. 352 353 New features: 354 - Support heap profiling on FreeBSD. This feature depends on the proc 355 filesystem being mounted during heap profile dumping. 356 357 * 3.5.1 (February 25, 2014) 358 359 This version primarily addresses minor bugs in test code. 360 361 Bug fixes: 362 - Configure Solaris/Illumos to use MADV_FREE. 363 - Fix junk filling for mremap(2)-based huge reallocation. This is only 364 relevant if configuring with the --enable-mremap option specified. 365 - Avoid compilation failure if 'restrict' C99 keyword is not supported by the 366 compiler. 367 - Add a configure test for SSE2 rather than assuming it is usable on i686 368 systems. This fixes test compilation errors, especially on 32-bit Linux 369 systems. 370 - Fix mallctl argument size mismatches (size_t vs. uint64_t) in the stats unit 371 test. 372 - Fix/remove flawed alignment-related overflow tests. 373 - Prevent compiler optimizations that could change backtraces in the 374 prof_accum unit test. 375 376 * 3.5.0 (January 22, 2014) 377 378 This version focuses on refactoring and automated testing, though it also 379 includes some non-trivial heap profiling optimizations not mentioned below. 380 381 New features: 382 - Add the *allocx() API, which is a successor to the experimental *allocm() 383 API. The *allocx() functions are slightly simpler to use because they have 384 fewer parameters, they directly return the results of primary interest, and 385 mallocx()/rallocx() avoid the strict aliasing pitfall that 386 allocm()/rallocm() share with posix_memalign(). Note that *allocm() is 387 slated for removal in the next non-bugfix release. 388 - Add support for LinuxThreads. 389 390 Bug fixes: 391 - Unless heap profiling is enabled, disable floating point code and don't link 392 with libm. This, in combination with e.g. EXTRA_CFLAGS=-mno-sse on x64 393 systems, makes it possible to completely disable floating point register 394 use. Some versions of glibc neglect to save/restore caller-saved floating 395 point registers during dynamic lazy symbol loading, and the symbol loading 396 code uses whatever malloc the application happens to have linked/loaded 397 with, the result being potential floating point register corruption. 398 - Report ENOMEM rather than EINVAL if an OOM occurs during heap profiling 399 backtrace creation in imemalign(). This bug impacted posix_memalign() and 400 aligned_alloc(). 401 - Fix a file descriptor leak in a prof_dump_maps() error path. 402 - Fix prof_dump() to close the dump file descriptor for all relevant error 403 paths. 404 - Fix rallocm() to use the arena specified by the ALLOCM_ARENA(s) flag for 405 allocation, not just deallocation. 406 - Fix a data race for large allocation stats counters. 407 - Fix a potential infinite loop during thread exit. This bug occurred on 408 Solaris, and could affect other platforms with similar pthreads TSD 409 implementations. 410 - Don't junk-fill reallocations unless usable size changes. This fixes a 411 violation of the *allocx()/*allocm() semantics. 412 - Fix growing large reallocation to junk fill new space. 413 - Fix huge deallocation to junk fill when munmap is disabled. 414 - Change the default private namespace prefix from empty to je_, and change 415 --with-private-namespace-prefix so that it prepends an additional prefix 416 rather than replacing je_. This reduces the likelihood of applications 417 which statically link jemalloc experiencing symbol name collisions. 418 - Add missing private namespace mangling (relevant when 419 --with-private-namespace is specified). 420 - Add and use JEMALLOC_INLINE_C so that static inline functions are marked as 421 static even for debug builds. 422 - Add a missing mutex unlock in a malloc_init_hard() error path. In practice 423 this error path is never executed. 424 - Fix numerous bugs in malloc_strotumax() error handling/reporting. These 425 bugs had no impact except for malformed inputs. 426 - Fix numerous bugs in malloc_snprintf(). These bugs were not exercised by 427 existing calls, so they had no impact. 428 429 * 3.4.1 (October 20, 2013) 430 431 Bug fixes: 432 - Fix a race in the "arenas.extend" mallctl that could cause memory corruption 433 of internal data structures and subsequent crashes. 434 - Fix Valgrind integration flaws that caused Valgrind warnings about reads of 435 uninitialized memory in: 436 + arena chunk headers 437 + internal zero-initialized data structures (relevant to tcache and prof 438 code) 439 - Preserve errno during the first allocation. A readlink(2) call during 440 initialization fails unless /etc/malloc.conf exists, so errno was typically 441 set during the first allocation prior to this fix. 442 - Fix compilation warnings reported by gcc 4.8.1. 443 444 * 3.4.0 (June 2, 2013) 445 446 This version is essentially a small bugfix release, but the addition of 447 aarch64 support requires that the minor version be incremented. 448 449 Bug fixes: 450 - Fix race-triggered deadlocks in chunk_record(). These deadlocks were 451 typically triggered by multiple threads concurrently deallocating huge 452 objects. 453 454 New features: 455 - Add support for the aarch64 architecture. 456 457 * 3.3.1 (March 6, 2013) 458 459 This version fixes bugs that are typically encountered only when utilizing 460 custom run-time options. 461 462 Bug fixes: 463 - Fix a locking order bug that could cause deadlock during fork if heap 464 profiling were enabled. 465 - Fix a chunk recycling bug that could cause the allocator to lose track of 466 whether a chunk was zeroed. On FreeBSD, NetBSD, and OS X, it could cause 467 corruption if allocating via sbrk(2) (unlikely unless running with the 468 "dss:primary" option specified). This was completely harmless on Linux 469 unless using mlockall(2) (and unlikely even then, unless the 470 --disable-munmap configure option or the "dss:primary" option was 471 specified). This regression was introduced in 3.1.0 by the 472 mlockall(2)/madvise(2) interaction fix. 473 - Fix TLS-related memory corruption that could occur during thread exit if the 474 thread never allocated memory. Only the quarantine and prof facilities were 475 susceptible. 476 - Fix two quarantine bugs: 477 + Internal reallocation of the quarantined object array leaked the old 478 array. 479 + Reallocation failure for internal reallocation of the quarantined object 480 array (very unlikely) resulted in memory corruption. 481 - Fix Valgrind integration to annotate all internally allocated memory in a 482 way that keeps Valgrind happy about internal data structure access. 483 - Fix building for s390 systems. 484 485 * 3.3.0 (January 23, 2013) 486 487 This version includes a few minor performance improvements in addition to the 488 listed new features and bug fixes. 489 490 New features: 491 - Add clipping support to lg_chunk option processing. 492 - Add the --enable-ivsalloc option. 493 - Add the --without-export option. 494 - Add the --disable-zone-allocator option. 495 496 Bug fixes: 497 - Fix "arenas.extend" mallctl to output the number of arenas. 498 - Fix chunk_recycle() to unconditionally inform Valgrind that returned memory 499 is undefined. 500 - Fix build break on FreeBSD related to alloca.h. 501 502 * 3.2.0 (November 9, 2012) 503 504 In addition to a couple of bug fixes, this version modifies page run 505 allocation and dirty page purging algorithms in order to better control 506 page-level virtual memory fragmentation. 507 508 Incompatible changes: 509 - Change the "opt.lg_dirty_mult" default from 5 to 3 (32:1 to 8:1). 510 511 Bug fixes: 512 - Fix dss/mmap allocation precedence code to use recyclable mmap memory only 513 after primary dss allocation fails. 514 - Fix deadlock in the "arenas.purge" mallctl. This regression was introduced 515 in 3.1.0 by the addition of the "arena.<i>.purge" mallctl. 516 517 * 3.1.0 (October 16, 2012) 518 519 New features: 520 - Auto-detect whether running inside Valgrind, thus removing the need to 521 manually specify MALLOC_CONF=valgrind:true. 522 - Add the "arenas.extend" mallctl, which allows applications to create 523 manually managed arenas. 524 - Add the ALLOCM_ARENA() flag for {,r,d}allocm(). 525 - Add the "opt.dss", "arena.<i>.dss", and "stats.arenas.<i>.dss" mallctls, 526 which provide control over dss/mmap precedence. 527 - Add the "arena.<i>.purge" mallctl, which obsoletes "arenas.purge". 528 - Define LG_QUANTUM for hppa. 529 530 Incompatible changes: 531 - Disable tcache by default if running inside Valgrind, in order to avoid 532 making unallocated objects appear reachable to Valgrind. 533 - Drop const from malloc_usable_size() argument on Linux. 534 535 Bug fixes: 536 - Fix heap profiling crash if sampled object is freed via realloc(p, 0). 537 - Remove const from __*_hook variable declarations, so that glibc can modify 538 them during process forking. 539 - Fix mlockall(2)/madvise(2) interaction. 540 - Fix fork(2)-related deadlocks. 541 - Fix error return value for "thread.tcache.enabled" mallctl. 542 543 * 3.0.0 (May 11, 2012) 544 545 Although this version adds some major new features, the primary focus is on 546 internal code cleanup that facilitates maintainability and portability, most 547 of which is not reflected in the ChangeLog. This is the first release to 548 incorporate substantial contributions from numerous other developers, and the 549 result is a more broadly useful allocator (see the git revision history for 550 contribution details). Note that the license has been unified, thanks to 551 Facebook granting a license under the same terms as the other copyright 552 holders (see COPYING). 553 554 New features: 555 - Implement Valgrind support, redzones, and quarantine. 556 - Add support for additional platforms: 557 + FreeBSD 558 + Mac OS X Lion 559 + MinGW 560 + Windows (no support yet for replacing the system malloc) 561 - Add support for additional architectures: 562 + MIPS 563 + SH4 564 + Tilera 565 - Add support for cross compiling. 566 - Add nallocm(), which rounds a request size up to the nearest size class 567 without actually allocating. 568 - Implement aligned_alloc() (blame C11). 569 - Add the "thread.tcache.enabled" mallctl. 570 - Add the "opt.prof_final" mallctl. 571 - Update pprof (from gperftools 2.0). 572 - Add the --with-mangling option. 573 - Add the --disable-experimental option. 574 - Add the --disable-munmap option, and make it the default on Linux. 575 - Add the --enable-mremap option, which disables use of mremap(2) by default. 576 577 Incompatible changes: 578 - Enable stats by default. 579 - Enable fill by default. 580 - Disable lazy locking by default. 581 - Rename the "tcache.flush" mallctl to "thread.tcache.flush". 582 - Rename the "arenas.pagesize" mallctl to "arenas.page". 583 - Change the "opt.lg_prof_sample" default from 0 to 19 (1 B to 512 KiB). 584 - Change the "opt.prof_accum" default from true to false. 585 586 Removed features: 587 - Remove the swap feature, including the "config.swap", "swap.avail", 588 "swap.prezeroed", "swap.nfds", and "swap.fds" mallctls. 589 - Remove highruns statistics, including the 590 "stats.arenas.<i>.bins.<j>.highruns" and 591 "stats.arenas.<i>.lruns.<j>.highruns" mallctls. 592 - As part of small size class refactoring, remove the "opt.lg_[qc]space_max", 593 "arenas.cacheline", "arenas.subpage", "arenas.[tqcs]space_{min,max}", and 594 "arenas.[tqcs]bins" mallctls. 595 - Remove the "arenas.chunksize" mallctl. 596 - Remove the "opt.lg_prof_tcmax" option. 597 - Remove the "opt.lg_prof_bt_max" option. 598 - Remove the "opt.lg_tcache_gc_sweep" option. 599 - Remove the --disable-tiny option, including the "config.tiny" mallctl. 600 - Remove the --enable-dynamic-page-shift configure option. 601 - Remove the --enable-sysv configure option. 602 603 Bug fixes: 604 - Fix a statistics-related bug in the "thread.arena" mallctl that could cause 605 invalid statistics and crashes. 606 - Work around TLS deallocation via free() on Linux. This bug could cause 607 write-after-free memory corruption. 608 - Fix a potential deadlock that could occur during interval- and 609 growth-triggered heap profile dumps. 610 - Fix large calloc() zeroing bugs due to dropping chunk map unzeroed flags. 611 - Fix chunk_alloc_dss() to stop claiming memory is zeroed. This bug could 612 cause memory corruption and crashes with --enable-dss specified. 613 - Fix fork-related bugs that could cause deadlock in children between fork 614 and exec. 615 - Fix malloc_stats_print() to honor 'b' and 'l' in the opts parameter. 616 - Fix realloc(p, 0) to act like free(p). 617 - Do not enforce minimum alignment in memalign(). 618 - Check for NULL pointer in malloc_usable_size(). 619 - Fix an off-by-one heap profile statistics bug that could be observed in 620 interval- and growth-triggered heap profiles. 621 - Fix the "epoch" mallctl to update cached stats even if the passed in epoch 622 is 0. 623 - Fix bin->runcur management to fix a layout policy bug. This bug did not 624 affect correctness. 625 - Fix a bug in choose_arena_hard() that potentially caused more arenas to be 626 initialized than necessary. 627 - Add missing "opt.lg_tcache_max" mallctl implementation. 628 - Use glibc allocator hooks to make mixed allocator usage less likely. 629 - Fix build issues for --disable-tcache. 630 - Don't mangle pthread_create() when --with-private-namespace is specified. 631 632 * 2.2.5 (November 14, 2011) 633 634 Bug fixes: 635 - Fix huge_ralloc() race when using mremap(2). This is a serious bug that 636 could cause memory corruption and/or crashes. 637 - Fix huge_ralloc() to maintain chunk statistics. 638 - Fix malloc_stats_print(..., "a") output. 639 640 * 2.2.4 (November 5, 2011) 641 642 Bug fixes: 643 - Initialize arenas_tsd before using it. This bug existed for 2.2.[0-3], as 644 well as for --disable-tls builds in earlier releases. 645 - Do not assume a 4 KiB page size in test/rallocm.c. 646 647 * 2.2.3 (August 31, 2011) 648 649 This version fixes numerous bugs related to heap profiling. 650 651 Bug fixes: 652 - Fix a prof-related race condition. This bug could cause memory corruption, 653 but only occurred in non-default configurations (prof_accum:false). 654 - Fix off-by-one backtracing issues (make sure that prof_alloc_prep() is 655 excluded from backtraces). 656 - Fix a prof-related bug in realloc() (only triggered by OOM errors). 657 - Fix prof-related bugs in allocm() and rallocm(). 658 - Fix prof_tdata_cleanup() for --disable-tls builds. 659 - Fix a relative include path, to fix objdir builds. 660 661 * 2.2.2 (July 30, 2011) 662 663 Bug fixes: 664 - Fix a build error for --disable-tcache. 665 - Fix assertions in arena_purge() (for real this time). 666 - Add the --with-private-namespace option. This is a workaround for symbol 667 conflicts that can inadvertently arise when using static libraries. 668 669 * 2.2.1 (March 30, 2011) 670 671 Bug fixes: 672 - Implement atomic operations for x86/x64. This fixes compilation failures 673 for versions of gcc that are still in wide use. 674 - Fix an assertion in arena_purge(). 675 676 * 2.2.0 (March 22, 2011) 677 678 This version incorporates several improvements to algorithms and data 679 structures that tend to reduce fragmentation and increase speed. 680 681 New features: 682 - Add the "stats.cactive" mallctl. 683 - Update pprof (from google-perftools 1.7). 684 - Improve backtracing-related configuration logic, and add the 685 --disable-prof-libgcc option. 686 687 Bug fixes: 688 - Change default symbol visibility from "internal", to "hidden", which 689 decreases the overhead of library-internal function calls. 690 - Fix symbol visibility so that it is also set on OS X. 691 - Fix a build dependency regression caused by the introduction of the .pic.o 692 suffix for PIC object files. 693 - Add missing checks for mutex initialization failures. 694 - Don't use libgcc-based backtracing except on x64, where it is known to work. 695 - Fix deadlocks on OS X that were due to memory allocation in 696 pthread_mutex_lock(). 697 - Heap profiling-specific fixes: 698 + Fix memory corruption due to integer overflow in small region index 699 computation, when using a small enough sample interval that profiling 700 context pointers are stored in small run headers. 701 + Fix a bootstrap ordering bug that only occurred with TLS disabled. 702 + Fix a rallocm() rsize bug. 703 + Fix error detection bugs for aligned memory allocation. 704 705 * 2.1.3 (March 14, 2011) 706 707 Bug fixes: 708 - Fix a cpp logic regression (due to the "thread.{de,}allocatedp" mallctl fix 709 for OS X in 2.1.2). 710 - Fix a "thread.arena" mallctl bug. 711 - Fix a thread cache stats merging bug. 712 713 * 2.1.2 (March 2, 2011) 714 715 Bug fixes: 716 - Fix "thread.{de,}allocatedp" mallctl for OS X. 717 - Add missing jemalloc.a to build system. 718 719 * 2.1.1 (January 31, 2011) 720 721 Bug fixes: 722 - Fix aligned huge reallocation (affected allocm()). 723 - Fix the ALLOCM_LG_ALIGN macro definition. 724 - Fix a heap dumping deadlock. 725 - Fix a "thread.arena" mallctl bug. 726 727 * 2.1.0 (December 3, 2010) 728 729 This version incorporates some optimizations that can't quite be considered 730 bug fixes. 731 732 New features: 733 - Use Linux's mremap(2) for huge object reallocation when possible. 734 - Avoid locking in mallctl*() when possible. 735 - Add the "thread.[de]allocatedp" mallctl's. 736 - Convert the manual page source from roff to DocBook, and generate both roff 737 and HTML manuals. 738 739 Bug fixes: 740 - Fix a crash due to incorrect bootstrap ordering. This only impacted 741 --enable-debug --enable-dss configurations. 742 - Fix a minor statistics bug for mallctl("swap.avail", ...). 743 744 * 2.0.1 (October 29, 2010) 745 746 Bug fixes: 747 - Fix a race condition in heap profiling that could cause undefined behavior 748 if "opt.prof_accum" were disabled. 749 - Add missing mutex unlocks for some OOM error paths in the heap profiling 750 code. 751 - Fix a compilation error for non-C99 builds. 752 753 * 2.0.0 (October 24, 2010) 754 755 This version focuses on the experimental *allocm() API, and on improved 756 run-time configuration/introspection. Nonetheless, numerous performance 757 improvements are also included. 758 759 New features: 760 - Implement the experimental {,r,s,d}allocm() API, which provides a superset 761 of the functionality available via malloc(), calloc(), posix_memalign(), 762 realloc(), malloc_usable_size(), and free(). These functions can be used to 763 allocate/reallocate aligned zeroed memory, ask for optional extra memory 764 during reallocation, prevent object movement during reallocation, etc. 765 - Replace JEMALLOC_OPTIONS/JEMALLOC_PROF_PREFIX with MALLOC_CONF, which is 766 more human-readable, and more flexible. For example: 767 JEMALLOC_OPTIONS=AJP 768 is now: 769 MALLOC_CONF=abort:true,fill:true,stats_print:true 770 - Port to Apple OS X. Sponsored by Mozilla. 771 - Make it possible for the application to control thread-->arena mappings via 772 the "thread.arena" mallctl. 773 - Add compile-time support for all TLS-related functionality via pthreads TSD. 774 This is mainly of interest for OS X, which does not support TLS, but has a 775 TSD implementation with similar performance. 776 - Override memalign() and valloc() if they are provided by the system. 777 - Add the "arenas.purge" mallctl, which can be used to synchronously purge all 778 dirty unused pages. 779 - Make cumulative heap profiling data optional, so that it is possible to 780 limit the amount of memory consumed by heap profiling data structures. 781 - Add per thread allocation counters that can be accessed via the 782 "thread.allocated" and "thread.deallocated" mallctls. 783 784 Incompatible changes: 785 - Remove JEMALLOC_OPTIONS and malloc_options (see MALLOC_CONF above). 786 - Increase default backtrace depth from 4 to 128 for heap profiling. 787 - Disable interval-based profile dumps by default. 788 789 Bug fixes: 790 - Remove bad assertions in fork handler functions. These assertions could 791 cause aborts for some combinations of configure settings. 792 - Fix strerror_r() usage to deal with non-standard semantics in GNU libc. 793 - Fix leak context reporting. This bug tended to cause the number of contexts 794 to be underreported (though the reported number of objects and bytes were 795 correct). 796 - Fix a realloc() bug for large in-place growing reallocation. This bug could 797 cause memory corruption, but it was hard to trigger. 798 - Fix an allocation bug for small allocations that could be triggered if 799 multiple threads raced to create a new run of backing pages. 800 - Enhance the heap profiler to trigger samples based on usable size, rather 801 than request size. 802 - Fix a heap profiling bug due to sometimes losing track of requested object 803 size for sampled objects. 804 805 * 1.0.3 (August 12, 2010) 806 807 Bug fixes: 808 - Fix the libunwind-based implementation of stack backtracing (used for heap 809 profiling). This bug could cause zero-length backtraces to be reported. 810 - Add a missing mutex unlock in library initialization code. If multiple 811 threads raced to initialize malloc, some of them could end up permanently 812 blocked. 813 814 * 1.0.2 (May 11, 2010) 815 816 Bug fixes: 817 - Fix junk filling of large objects, which could cause memory corruption. 818 - Add MAP_NORESERVE support for chunk mapping, because otherwise virtual 819 memory limits could cause swap file configuration to fail. Contributed by 820 Jordan DeLong. 821 822 * 1.0.1 (April 14, 2010) 823 824 Bug fixes: 825 - Fix compilation when --enable-fill is specified. 826 - Fix threads-related profiling bugs that affected accuracy and caused memory 827 to be leaked during thread exit. 828 - Fix dirty page purging race conditions that could cause crashes. 829 - Fix crash in tcache flushing code during thread destruction. 830 831 * 1.0.0 (April 11, 2010) 832 833 This release focuses on speed and run-time introspection. Numerous 834 algorithmic improvements make this release substantially faster than its 835 predecessors. 836 837 New features: 838 - Implement autoconf-based configuration system. 839 - Add mallctl*(), for the purposes of introspection and run-time 840 configuration. 841 - Make it possible for the application to manually flush a thread's cache, via 842 the "tcache.flush" mallctl. 843 - Base maximum dirty page count on proportion of active memory. 844 - Compute various additional run-time statistics, including per size class 845 statistics for large objects. 846 - Expose malloc_stats_print(), which can be called repeatedly by the 847 application. 848 - Simplify the malloc_message() signature to only take one string argument, 849 and incorporate an opaque data pointer argument for use by the application 850 in combination with malloc_stats_print(). 851 - Add support for allocation backed by one or more swap files, and allow the 852 application to disable over-commit if swap files are in use. 853 - Implement allocation profiling and leak checking. 854 855 Removed features: 856 - Remove the dynamic arena rebalancing code, since thread-specific caching 857 reduces its utility. 858 859 Bug fixes: 860 - Modify chunk allocation to work when address space layout randomization 861 (ASLR) is in use. 862 - Fix thread cleanup bugs related to TLS destruction. 863 - Handle 0-size allocation requests in posix_memalign(). 864 - Fix a chunk leak. The leaked chunks were never touched, so this impacted 865 virtual memory usage, but not physical memory usage. 866 867 * linux_2008082[78]a (August 27/28, 2008) 868 869 These snapshot releases are the simple result of incorporating Linux-specific 870 support into the FreeBSD malloc sources. 871 872 -------------------------------------------------------------------------------- 873 vim:filetype=text:textwidth=80 874
