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