xref: /freebsd/contrib/jemalloc/ChangeLog (revision c43cad87172039ccf38172129c79755ea79e6102)
1Following are change highlights associated with official releases.  Important
2bug fixes are all mentioned, but some internal enhancements are omitted here for
3brevity.  Much more detail can be found in the git revision history:
4
5    https://github.com/jemalloc/jemalloc
6
7* 5.3.0 (May 6, 2022)
8
9  This release contains many speed and space optimizations, from micro
10  optimizations on common paths to rework of internal data structures and
11  locking schemes, and many more too detailed to list below.  Multiple percent
12  of system level metric improvements were measured in tested production
13  workloads.  The release has gone through large-scale production testing.
14
15  New features:
16  - Add the thread.idle mallctl which hints that the calling thread will be
17    idle for a nontrivial period of time.  (@davidtgoldblatt)
18  - Allow small size classes to be the maximum size class to cache in the
19    thread-specific cache, through the opt.[lg_]tcache_max option.  (@interwq,
20    @jordalgo)
21  - Make the behavior of realloc(ptr, 0) configurable with opt.zero_realloc.
22    (@davidtgoldblatt)
23  - Add 'make uninstall' support.  (@sangshuduo, @Lapenkov)
24  - Support C++17 over-aligned allocation.  (@marksantaniello)
25  - Add the thread.peak mallctl for approximate per-thread peak memory tracking.
26    (@davidtgoldblatt)
27  - Add interval-based stats output opt.stats_interval.  (@interwq)
28  - Add prof.prefix to override filename prefixes for dumps.  (@zhxchen17)
29  - Add high resolution timestamp support for profiling.  (@tyroguru)
30  - Add the --collapsed flag to jeprof for flamegraph generation.
31    (@igorwwwwwwwwwwwwwwwwwwww)
32  - Add the --debug-syms-by-id option to jeprof for debug symbols discovery.
33    (@DeannaGelbart)
34  - Add the opt.prof_leak_error option to exit with error code when leak is
35    detected using opt.prof_final.  (@yunxuo)
36  - Add opt.cache_oblivious as an runtime alternative to config.cache_oblivious.
37    (@interwq)
38  - Add mallctl interfaces:
39    + opt.zero_realloc  (@davidtgoldblatt)
40    + opt.cache_oblivious  (@interwq)
41    + opt.prof_leak_error  (@yunxuo)
42    + opt.stats_interval  (@interwq)
43    + opt.stats_interval_opts  (@interwq)
44    + opt.tcache_max  (@interwq)
45    + opt.trust_madvise  (@azat)
46    + prof.prefix  (@zhxchen17)
47    + stats.zero_reallocs  (@davidtgoldblatt)
48    + thread.idle  (@davidtgoldblatt)
49    + thread.peak.{read,reset}  (@davidtgoldblatt)
50
51  Bug fixes:
52  - Fix the synchronization around explicit tcache creation which could cause
53    invalid tcache identifiers.  This regression was first released in 5.0.0.
54    (@yoshinorim, @davidtgoldblatt)
55  - Fix a profiling biasing issue which could cause incorrect heap usage and
56    object counts.  This issue existed in all previous releases with the heap
57    profiling feature.  (@davidtgoldblatt)
58  - Fix the order of stats counter updating on large realloc which could cause
59    failed assertions.  This regression was first released in 5.0.0.  (@azat)
60  - Fix the locking on the arena destroy mallctl, which could cause concurrent
61    arena creations to fail.  This functionality was first introduced in 5.0.0.
62    (@interwq)
63
64  Portability improvements:
65  - Remove nothrow from system function declarations on macOS and FreeBSD.
66    (@davidtgoldblatt, @fredemmott, @leres)
67  - Improve overcommit and page alignment settings on NetBSD.  (@zoulasc)
68  - Improve CPU affinity support on BSD platforms.  (@devnexen)
69  - Improve utrace detection and support.  (@devnexen)
70  - Improve QEMU support with MADV_DONTNEED zeroed pages detection.  (@azat)
71  - Add memcntl support on Solaris / illumos.  (@devnexen)
72  - Improve CPU_SPINWAIT on ARM.  (@AWSjswinney)
73  - Improve TSD cleanup on FreeBSD.  (@Lapenkov)
74  - Disable percpu_arena if the CPU count cannot be reliably detected.  (@azat)
75  - Add malloc_size(3) override support.  (@devnexen)
76  - Add mmap VM_MAKE_TAG support.  (@devnexen)
77  - Add support for MADV_[NO]CORE.  (@devnexen)
78  - Add support for DragonFlyBSD.  (@devnexen)
79  - Fix the QUANTUM setting on MIPS64.  (@brooksdavis)
80  - Add the QUANTUM setting for ARC.  (@vineetgarc)
81  - Add the QUANTUM setting for LoongArch.  (@wangjl-uos)
82  - Add QNX support.  (@jqian-aurora)
83  - Avoid atexit(3) calls unless the relevant profiling features are enabled.
84    (@BusyJay, @laiwei-rice, @interwq)
85  - Fix unknown option detection when using Clang.  (@Lapenkov)
86  - Fix symbol conflict with musl libc.  (@georgthegreat)
87  - Add -Wimplicit-fallthrough checks.  (@nickdesaulniers)
88  - Add __forceinline support on MSVC.  (@santagada)
89  - Improve FreeBSD and Windows CI support.  (@Lapenkov)
90  - Add CI support for PPC64LE architecture.  (@ezeeyahoo)
91
92  Incompatible changes:
93  - Maximum size class allowed in tcache (opt.[lg_]tcache_max) now has an upper
94    bound of 8MiB.  (@interwq)
95
96  Optimizations and refactors (@davidtgoldblatt, @Lapenkov, @interwq):
97  - Optimize the common cases of the thread cache operations.
98  - Optimize internal data structures, including RB tree and pairing heap.
99  - Optimize the internal locking on extent management.
100  - Extract and refactor the internal page allocator and interface modules.
101
102  Documentation:
103  - Fix doc build with --with-install-suffix.  (@lawmurray, @interwq)
104  - Add PROFILING_INTERNALS.md.  (@davidtgoldblatt)
105  - Ensure the proper order of doc building and installation.  (@Mingli-Yu)
106
107* 5.2.1 (August 5, 2019)
108
109  This release is primarily about Windows.  A critical virtual memory leak is
110  resolved on all Windows platforms.  The regression was present in all releases
111  since 5.0.0.
112
113  Bug fixes:
114  - Fix a severe virtual memory leak on Windows.  This regression was first
115    released in 5.0.0.  (@Ignition, @j0t, @frederik-h, @davidtgoldblatt,
116    @interwq)
117  - Fix size 0 handling in posix_memalign().  This regression was first released
118    in 5.2.0.  (@interwq)
119  - Fix the prof_log unit test which may observe unexpected backtraces from
120    compiler optimizations.  The test was first added in 5.2.0.  (@marxin,
121    @gnzlbg, @interwq)
122  - Fix the declaration of the extent_avail tree.  This regression was first
123    released in 5.1.0.  (@zoulasc)
124  - Fix an incorrect reference in jeprof.  This functionality was first released
125    in 3.0.0.  (@prehistoric-penguin)
126  - Fix an assertion on the deallocation fast-path.  This regression was first
127    released in 5.2.0.  (@yinan1048576)
128  - Fix the TLS_MODEL attribute in headers.  This regression was first released
129    in 5.0.0.  (@zoulasc, @interwq)
130
131  Optimizations and refactors:
132  - Implement opt.retain on Windows and enable by default on 64-bit.  (@interwq,
133    @davidtgoldblatt)
134  - Optimize away a branch on the operator delete[] path.  (@mgrice)
135  - Add format annotation to the format generator function.  (@zoulasc)
136  - Refactor and improve the size class header generation.  (@yinan1048576)
137  - Remove best fit.  (@djwatson)
138  - Avoid blocking on background thread locks for stats.  (@oranagra, @interwq)
139
140* 5.2.0 (April 2, 2019)
141
142  This release includes a few notable improvements, which are summarized below:
143  1) improved fast-path performance from the optimizations by @djwatson; 2)
144  reduced virtual memory fragmentation and metadata usage; and 3) bug fixes on
145  setting the number of background threads.  In addition, peak / spike memory
146  usage is improved with certain allocation patterns.  As usual, the release and
147  prior dev versions have gone through large-scale production testing.
148
149  New features:
150  - Implement oversize_threshold, which uses a dedicated arena for allocations
151    crossing the specified threshold to reduce fragmentation.  (@interwq)
152  - Add extents usage information to stats.  (@tyleretzel)
153  - Log time information for sampled allocations.  (@tyleretzel)
154  - Support 0 size in sdallocx.  (@djwatson)
155  - Output rate for certain counters in malloc_stats.  (@zinoale)
156  - Add configure option --enable-readlinkat, which allows the use of readlinkat
157    over readlink.  (@davidtgoldblatt)
158  - Add configure options --{enable,disable}-{static,shared} to allow not
159    building unwanted libraries.  (@Ericson2314)
160  - Add configure option --disable-libdl to enable fully static builds.
161    (@interwq)
162  - Add mallctl interfaces:
163	+ opt.oversize_threshold (@interwq)
164	+ stats.arenas.<i>.extent_avail (@tyleretzel)
165	+ stats.arenas.<i>.extents.<j>.n{dirty,muzzy,retained} (@tyleretzel)
166	+ stats.arenas.<i>.extents.<j>.{dirty,muzzy,retained}_bytes
167	  (@tyleretzel)
168
169  Portability improvements:
170  - Update MSVC builds.  (@maksqwe, @rustyx)
171  - Workaround a compiler optimizer bug on s390x.  (@rkmisra)
172  - Make use of pthread_set_name_np(3) on FreeBSD.  (@trasz)
173  - Implement malloc_getcpu() to enable percpu_arena for windows.  (@santagada)
174  - Link against -pthread instead of -lpthread.  (@paravoid)
175  - Make background_thread not dependent on libdl.  (@interwq)
176  - Add stringify to fix a linker directive issue on MSVC.  (@daverigby)
177  - Detect and fall back when 8-bit atomics are unavailable.  (@interwq)
178  - Fall back to the default pthread_create if dlsym(3) fails.  (@interwq)
179
180  Optimizations and refactors:
181  - Refactor the TSD module.  (@davidtgoldblatt)
182  - Avoid taking extents_muzzy mutex when muzzy is disabled.  (@interwq)
183  - Avoid taking large_mtx for auto arenas on the tcache flush path.  (@interwq)
184  - Optimize ixalloc by avoiding a size lookup.  (@interwq)
185  - Implement opt.oversize_threshold which uses a dedicated arena for requests
186    crossing the threshold, also eagerly purges the oversize extents.  Default
187    the threshold to 8 MiB.  (@interwq)
188  - Clean compilation with -Wextra.  (@gnzlbg, @jasone)
189  - Refactor the size class module.  (@davidtgoldblatt)
190  - Refactor the stats emitter.  (@tyleretzel)
191  - Optimize pow2_ceil.  (@rkmisra)
192  - Avoid runtime detection of lazy purging on FreeBSD.  (@trasz)
193  - Optimize mmap(2) alignment handling on FreeBSD.  (@trasz)
194  - Improve error handling for THP state initialization.  (@jsteemann)
195  - Rework the malloc() fast path.  (@djwatson)
196  - Rework the free() fast path.  (@djwatson)
197  - Refactor and optimize the tcache fill / flush paths.  (@djwatson)
198  - Optimize sync / lwsync on PowerPC.  (@chmeeedalf)
199  - Bypass extent_dalloc() when retain is enabled.  (@interwq)
200  - Optimize the locking on large deallocation.  (@interwq)
201  - Reduce the number of pages committed from sanity checking in debug build.
202    (@trasz, @interwq)
203  - Deprecate OSSpinLock.  (@interwq)
204  - Lower the default number of background threads to 4 (when the feature
205    is enabled).  (@interwq)
206  - Optimize the trylock spin wait.  (@djwatson)
207  - Use arena index for arena-matching checks.  (@interwq)
208  - Avoid forced decay on thread termination when using background threads.
209    (@interwq)
210  - Disable muzzy decay by default.  (@djwatson, @interwq)
211  - Only initialize libgcc unwinder when profiling is enabled.  (@paravoid,
212    @interwq)
213
214  Bug fixes (all only relevant to jemalloc 5.x):
215  - Fix background thread index issues with max_background_threads.  (@djwatson,
216    @interwq)
217  - Fix stats output for opt.lg_extent_max_active_fit.  (@interwq)
218  - Fix opt.prof_prefix initialization.  (@davidtgoldblatt)
219  - Properly trigger decay on tcache destroy.  (@interwq, @amosbird)
220  - Fix tcache.flush.  (@interwq)
221  - Detect whether explicit extent zero out is necessary with huge pages or
222    custom extent hooks, which may change the purge semantics.  (@interwq)
223  - Fix a side effect caused by extent_max_active_fit combined with decay-based
224    purging, where freed extents can accumulate and not be reused for an
225    extended period of time.  (@interwq, @mpghf)
226  - Fix a missing unlock on extent register error handling.  (@zoulasc)
227
228  Testing:
229  - Simplify the Travis script output.  (@gnzlbg)
230  - Update the test scripts for FreeBSD.  (@devnexen)
231  - Add unit tests for the producer-consumer pattern.  (@interwq)
232  - Add Cirrus-CI config for FreeBSD builds.  (@jasone)
233  - Add size-matching sanity checks on tcache flush.  (@davidtgoldblatt,
234    @interwq)
235
236  Incompatible changes:
237  - Remove --with-lg-page-sizes.  (@davidtgoldblatt)
238
239  Documentation:
240  - Attempt to build docs by default, however skip doc building when xsltproc
241    is missing. (@interwq, @cmuellner)
242
243* 5.1.0 (May 4, 2018)
244
245  This release is primarily about fine-tuning, ranging from several new features
246  to numerous notable performance and portability enhancements.  The release and
247  prior dev versions have been running in multiple large scale applications for
248  months, and the cumulative improvements are substantial in many cases.
249
250  Given the long and successful production runs, this release is likely a good
251  candidate for applications to upgrade, from both jemalloc 5.0 and before.  For
252  performance-critical applications, the newly added TUNING.md provides
253  guidelines on jemalloc tuning.
254
255  New features:
256  - Implement transparent huge page support for internal metadata.  (@interwq)
257  - Add opt.thp to allow enabling / disabling transparent huge pages for all
258    mappings.  (@interwq)
259  - Add maximum background thread count option.  (@djwatson)
260  - Allow prof_active to control opt.lg_prof_interval and prof.gdump.
261    (@interwq)
262  - Allow arena index lookup based on allocation addresses via mallctl.
263    (@lionkov)
264  - Allow disabling initial-exec TLS model.  (@davidtgoldblatt, @KenMacD)
265  - Add opt.lg_extent_max_active_fit to set the max ratio between the size of
266    the active extent selected (to split off from) and the size of the requested
267    allocation.  (@interwq, @davidtgoldblatt)
268  - Add retain_grow_limit to set the max size when growing virtual address
269    space.  (@interwq)
270  - Add mallctl interfaces:
271    + arena.<i>.retain_grow_limit  (@interwq)
272    + arenas.lookup  (@lionkov)
273    + max_background_threads  (@djwatson)
274    + opt.lg_extent_max_active_fit  (@interwq)
275    + opt.max_background_threads  (@djwatson)
276    + opt.metadata_thp  (@interwq)
277    + opt.thp  (@interwq)
278    + stats.metadata_thp  (@interwq)
279
280  Portability improvements:
281  - Support GNU/kFreeBSD configuration.  (@paravoid)
282  - Support m68k, nios2 and SH3 architectures.  (@paravoid)
283  - Fall back to FD_CLOEXEC when O_CLOEXEC is unavailable.  (@zonyitoo)
284  - Fix symbol listing for cross-compiling.  (@tamird)
285  - Fix high bits computation on ARM.  (@davidtgoldblatt, @paravoid)
286  - Disable the CPU_SPINWAIT macro for Power.  (@davidtgoldblatt, @marxin)
287  - Fix MSVC 2015 & 2017 builds.  (@rustyx)
288  - Improve RISC-V support.  (@EdSchouten)
289  - Set name mangling script in strict mode.  (@nicolov)
290  - Avoid MADV_HUGEPAGE on ARM.  (@marxin)
291  - Modify configure to determine return value of strerror_r.
292    (@davidtgoldblatt, @cferris1000)
293  - Make sure CXXFLAGS is tested with CPP compiler.  (@nehaljwani)
294  - Fix 32-bit build on MSVC.  (@rustyx)
295  - Fix external symbol on MSVC.  (@maksqwe)
296  - Avoid a printf format specifier warning.  (@jasone)
297  - Add configure option --disable-initial-exec-tls which can allow jemalloc to
298    be dynamically loaded after program startup.  (@davidtgoldblatt, @KenMacD)
299  - AArch64: Add ILP32 support.  (@cmuellner)
300  - Add --with-lg-vaddr configure option to support cross compiling.
301    (@cmuellner, @davidtgoldblatt)
302
303  Optimizations and refactors:
304  - Improve active extent fit with extent_max_active_fit.  This considerably
305    reduces fragmentation over time and improves virtual memory and metadata
306    usage.  (@davidtgoldblatt, @interwq)
307  - Eagerly coalesce large extents to reduce fragmentation.  (@interwq)
308  - sdallocx: only read size info when page aligned (i.e. possibly sampled),
309    which speeds up the sized deallocation path significantly.  (@interwq)
310  - Avoid attempting new mappings for in place expansion with retain, since
311    it rarely succeeds in practice and causes high overhead.  (@interwq)
312  - Refactor OOM handling in newImpl.  (@wqfish)
313  - Add internal fine-grained logging functionality for debugging use.
314    (@davidtgoldblatt)
315  - Refactor arena / tcache interactions.  (@davidtgoldblatt)
316  - Refactor extent management with dumpable flag.  (@davidtgoldblatt)
317  - Add runtime detection of lazy purging.  (@interwq)
318  - Use pairing heap instead of red-black tree for extents_avail.  (@djwatson)
319  - Use sysctl on startup in FreeBSD.  (@trasz)
320  - Use thread local prng state instead of atomic.  (@djwatson)
321  - Make decay to always purge one more extent than before, because in
322    practice large extents are usually the ones that cross the decay threshold.
323    Purging the additional extent helps save memory as well as reduce VM
324    fragmentation.  (@interwq)
325  - Fast division by dynamic values.  (@davidtgoldblatt)
326  - Improve the fit for aligned allocation.  (@interwq, @edwinsmith)
327  - Refactor extent_t bitpacking.  (@rkmisra)
328  - Optimize the generated assembly for ticker operations.  (@davidtgoldblatt)
329  - Convert stats printing to use a structured text emitter.  (@davidtgoldblatt)
330  - Remove preserve_lru feature for extents management.  (@djwatson)
331  - Consolidate two memory loads into one on the fast deallocation path.
332    (@davidtgoldblatt, @interwq)
333
334  Bug fixes (most of the issues are only relevant to jemalloc 5.0):
335  - Fix deadlock with multithreaded fork in OS X.  (@davidtgoldblatt)
336  - Validate returned file descriptor before use.  (@zonyitoo)
337  - Fix a few background thread initialization and shutdown issues.  (@interwq)
338  - Fix an extent coalesce + decay race by taking both coalescing extents off
339    the LRU list.  (@interwq)
340  - Fix potentially unbound increase during decay, caused by one thread keep
341    stashing memory to purge while other threads generating new pages.  The
342    number of pages to purge is checked to prevent this.  (@interwq)
343  - Fix a FreeBSD bootstrap assertion.  (@strejda, @interwq)
344  - Handle 32 bit mutex counters.  (@rkmisra)
345  - Fix a indexing bug when creating background threads.  (@davidtgoldblatt,
346    @binliu19)
347  - Fix arguments passed to extent_init.  (@yuleniwo, @interwq)
348  - Fix addresses used for ordering mutexes.  (@rkmisra)
349  - Fix abort_conf processing during bootstrap.  (@interwq)
350  - Fix include path order for out-of-tree builds.  (@cmuellner)
351
352  Incompatible changes:
353  - Remove --disable-thp.  (@interwq)
354  - Remove mallctl interfaces:
355    + config.thp  (@interwq)
356
357  Documentation:
358  - Add TUNING.md.  (@interwq, @davidtgoldblatt, @djwatson)
359
360* 5.0.1 (July 1, 2017)
361
362  This bugfix release fixes several issues, most of which are obscure enough
363  that typical applications are not impacted.
364
365  Bug fixes:
366  - Update decay->nunpurged before purging, in order to avoid potential update
367    races and subsequent incorrect purging volume.  (@interwq)
368  - Only abort on dlsym(3) error if the failure impacts an enabled feature (lazy
369    locking and/or background threads).  This mitigates an initialization
370    failure bug for which we still do not have a clear reproduction test case.
371    (@interwq)
372  - Modify tsd management so that it neither crashes nor leaks if a thread's
373    only allocation activity is to call free() after TLS destructors have been
374    executed.  This behavior was observed when operating with GNU libc, and is
375    unlikely to be an issue with other libc implementations.  (@interwq)
376  - Mask signals during background thread creation.  This prevents signals from
377    being inadvertently delivered to background threads.  (@jasone,
378    @davidtgoldblatt, @interwq)
379  - Avoid inactivity checks within background threads, in order to prevent
380    recursive mutex acquisition.  (@interwq)
381  - Fix extent_grow_retained() to use the specified hooks when the
382    arena.<i>.extent_hooks mallctl is used to override the default hooks.
383    (@interwq)
384  - Add missing reentrancy support for custom extent hooks which allocate.
385    (@interwq)
386  - Post-fork(2), re-initialize the list of tcaches associated with each arena
387    to contain no tcaches except the forking thread's.  (@interwq)
388  - Add missing post-fork(2) mutex reinitialization for extent_grow_mtx.  This
389    fixes potential deadlocks after fork(2).  (@interwq)
390  - Enforce minimum autoconf version (currently 2.68), since 2.63 is known to
391    generate corrupt configure scripts.  (@jasone)
392  - Ensure that the configured page size (--with-lg-page) is no larger than the
393    configured huge page size (--with-lg-hugepage).  (@jasone)
394
395* 5.0.0 (June 13, 2017)
396
397  Unlike all previous jemalloc releases, this release does not use naturally
398  aligned "chunks" for virtual memory management, and instead uses page-aligned
399  "extents".  This change has few externally visible effects, but the internal
400  impacts are... extensive.  Many other internal changes combine to make this
401  the most cohesively designed version of jemalloc so far, with ample
402  opportunity for further enhancements.
403
404  Continuous integration is now an integral aspect of development thanks to the
405  efforts of @davidtgoldblatt, and the dev branch tends to remain reasonably
406  stable on the tested platforms (Linux, FreeBSD, macOS, and Windows).  As a
407  side effect the official release frequency may decrease over time.
408
409  New features:
410  - Implement optional per-CPU arena support; threads choose which arena to use
411    based on current CPU rather than on fixed thread-->arena associations.
412    (@interwq)
413  - Implement two-phase decay of unused dirty pages.  Pages transition from
414    dirty-->muzzy-->clean, where the first phase transition relies on
415    madvise(... MADV_FREE) semantics, and the second phase transition discards
416    pages such that they are replaced with demand-zeroed pages on next access.
417    (@jasone)
418  - Increase decay time resolution from seconds to milliseconds.  (@jasone)
419  - Implement opt-in per CPU background threads, and use them for asynchronous
420    decay-driven unused dirty page purging.  (@interwq)
421  - Add mutex profiling, which collects a variety of statistics useful for
422    diagnosing overhead/contention issues.  (@interwq)
423  - Add C++ new/delete operator bindings.  (@djwatson)
424  - Support manually created arena destruction, such that all data and metadata
425    are discarded.  Add MALLCTL_ARENAS_DESTROYED for accessing merged stats
426    associated with destroyed arenas.  (@jasone)
427  - Add MALLCTL_ARENAS_ALL as a fixed index for use in accessing
428    merged/destroyed arena statistics via mallctl.  (@jasone)
429  - Add opt.abort_conf to optionally abort if invalid configuration options are
430    detected during initialization.  (@interwq)
431  - Add opt.stats_print_opts, so that e.g. JSON output can be selected for the
432    stats dumped during exit if opt.stats_print is true.  (@jasone)
433  - Add --with-version=VERSION for use when embedding jemalloc into another
434    project's git repository.  (@jasone)
435  - Add --disable-thp to support cross compiling.  (@jasone)
436  - Add --with-lg-hugepage to support cross compiling.  (@jasone)
437  - Add mallctl interfaces (various authors):
438    + background_thread
439    + opt.abort_conf
440    + opt.retain
441    + opt.percpu_arena
442    + opt.background_thread
443    + opt.{dirty,muzzy}_decay_ms
444    + opt.stats_print_opts
445    + arena.<i>.initialized
446    + arena.<i>.destroy
447    + arena.<i>.{dirty,muzzy}_decay_ms
448    + arena.<i>.extent_hooks
449    + arenas.{dirty,muzzy}_decay_ms
450    + arenas.bin.<i>.slab_size
451    + arenas.nlextents
452    + arenas.lextent.<i>.size
453    + arenas.create
454    + stats.background_thread.{num_threads,num_runs,run_interval}
455    + stats.mutexes.{ctl,background_thread,prof,reset}.
456      {num_ops,num_spin_acq,num_wait,max_wait_time,total_wait_time,max_num_thds,
457      num_owner_switch}
458    + stats.arenas.<i>.{dirty,muzzy}_decay_ms
459    + stats.arenas.<i>.uptime
460    + stats.arenas.<i>.{pmuzzy,base,internal,resident}
461    + stats.arenas.<i>.{dirty,muzzy}_{npurge,nmadvise,purged}
462    + stats.arenas.<i>.bins.<j>.{nslabs,reslabs,curslabs}
463    + stats.arenas.<i>.bins.<j>.mutex.
464      {num_ops,num_spin_acq,num_wait,max_wait_time,total_wait_time,max_num_thds,
465      num_owner_switch}
466    + stats.arenas.<i>.lextents.<j>.{nmalloc,ndalloc,nrequests,curlextents}
467    + stats.arenas.i.mutexes.{large,extent_avail,extents_dirty,extents_muzzy,
468      extents_retained,decay_dirty,decay_muzzy,base,tcache_list}.
469      {num_ops,num_spin_acq,num_wait,max_wait_time,total_wait_time,max_num_thds,
470      num_owner_switch}
471
472  Portability improvements:
473  - Improve reentrant allocation support, such that deadlock is less likely if
474    e.g. a system library call in turn allocates memory.  (@davidtgoldblatt,
475    @interwq)
476  - Support static linking of jemalloc with glibc.  (@djwatson)
477
478  Optimizations and refactors:
479  - Organize virtual memory as "extents" of virtual memory pages, rather than as
480    naturally aligned "chunks", and store all metadata in arbitrarily distant
481    locations.  This reduces virtual memory external fragmentation, and will
482    interact better with huge pages (not yet explicitly supported).  (@jasone)
483  - Fold large and huge size classes together; only small and large size classes
484    remain.  (@jasone)
485  - Unify the allocation paths, and merge most fast-path branching decisions.
486    (@davidtgoldblatt, @interwq)
487  - Embed per thread automatic tcache into thread-specific data, which reduces
488    conditional branches and dereferences.  Also reorganize tcache to increase
489    fast-path data locality.  (@interwq)
490  - Rewrite atomics to closely model the C11 API, convert various
491    synchronization from mutex-based to atomic, and use the explicit memory
492    ordering control to resolve various hypothetical races without increasing
493    synchronization overhead.  (@davidtgoldblatt)
494  - Extensively optimize rtree via various methods:
495    + Add multiple layers of rtree lookup caching, since rtree lookups are now
496      part of fast-path deallocation.  (@interwq)
497    + Determine rtree layout at compile time.  (@jasone)
498    + Make the tree shallower for common configurations.  (@jasone)
499    + Embed the root node in the top-level rtree data structure, thus avoiding
500      one level of indirection.  (@jasone)
501    + Further specialize leaf elements as compared to internal node elements,
502      and directly embed extent metadata needed for fast-path deallocation.
503      (@jasone)
504    + Ignore leading always-zero address bits (architecture-specific).
505      (@jasone)
506  - Reorganize headers (ongoing work) to make them hermetic, and disentangle
507    various module dependencies.  (@davidtgoldblatt)
508  - Convert various internal data structures such as size class metadata from
509    boot-time-initialized to compile-time-initialized.  Propagate resulting data
510    structure simplifications, such as making arena metadata fixed-size.
511    (@jasone)
512  - Simplify size class lookups when constrained to size classes that are
513    multiples of the page size.  This speeds lookups, but the primary benefit is
514    complexity reduction in code that was the source of numerous regressions.
515    (@jasone)
516  - Lock individual extents when possible for localized extent operations,
517    rather than relying on a top-level arena lock.  (@davidtgoldblatt, @jasone)
518  - Use first fit layout policy instead of best fit, in order to improve
519    packing.  (@jasone)
520  - If munmap(2) is not in use, use an exponential series to grow each arena's
521    virtual memory, so that the number of disjoint virtual memory mappings
522    remains low.  (@jasone)
523  - Implement per arena base allocators, so that arenas never share any virtual
524    memory pages.  (@jasone)
525  - Automatically generate private symbol name mangling macros.  (@jasone)
526
527  Incompatible changes:
528  - Replace chunk hooks with an expanded/normalized set of extent hooks.
529    (@jasone)
530  - Remove ratio-based purging.  (@jasone)
531  - Remove --disable-tcache.  (@jasone)
532  - Remove --disable-tls.  (@jasone)
533  - Remove --enable-ivsalloc.  (@jasone)
534  - Remove --with-lg-size-class-group.  (@jasone)
535  - Remove --with-lg-tiny-min.  (@jasone)
536  - Remove --disable-cc-silence.  (@jasone)
537  - Remove --enable-code-coverage.  (@jasone)
538  - Remove --disable-munmap (replaced by opt.retain).  (@jasone)
539  - Remove Valgrind support.  (@jasone)
540  - Remove quarantine support.  (@jasone)
541  - Remove redzone support.  (@jasone)
542  - Remove mallctl interfaces (various authors):
543    + config.munmap
544    + config.tcache
545    + config.tls
546    + config.valgrind
547    + opt.lg_chunk
548    + opt.purge
549    + opt.lg_dirty_mult
550    + opt.decay_time
551    + opt.quarantine
552    + opt.redzone
553    + opt.thp
554    + arena.<i>.lg_dirty_mult
555    + arena.<i>.decay_time
556    + arena.<i>.chunk_hooks
557    + arenas.initialized
558    + arenas.lg_dirty_mult
559    + arenas.decay_time
560    + arenas.bin.<i>.run_size
561    + arenas.nlruns
562    + arenas.lrun.<i>.size
563    + arenas.nhchunks
564    + arenas.hchunk.<i>.size
565    + arenas.extend
566    + stats.cactive
567    + stats.arenas.<i>.lg_dirty_mult
568    + stats.arenas.<i>.decay_time
569    + stats.arenas.<i>.metadata.{mapped,allocated}
570    + stats.arenas.<i>.{npurge,nmadvise,purged}
571    + stats.arenas.<i>.huge.{allocated,nmalloc,ndalloc,nrequests}
572    + stats.arenas.<i>.bins.<j>.{nruns,reruns,curruns}
573    + stats.arenas.<i>.lruns.<j>.{nmalloc,ndalloc,nrequests,curruns}
574    + stats.arenas.<i>.hchunks.<j>.{nmalloc,ndalloc,nrequests,curhchunks}
575
576  Bug fixes:
577  - Improve interval-based profile dump triggering to dump only one profile when
578    a single allocation's size exceeds the interval.  (@jasone)
579  - Use prefixed function names (as controlled by --with-jemalloc-prefix) when
580    pruning backtrace frames in jeprof.  (@jasone)
581
582* 4.5.0 (February 28, 2017)
583
584  This is the first release to benefit from much broader continuous integration
585  testing, thanks to @davidtgoldblatt.  Had we had this testing infrastructure
586  in place for prior releases, it would have caught all of the most serious
587  regressions fixed by this release.
588
589  New features:
590  - Add --disable-thp and the opt.thp mallctl to provide opt-out mechanisms for
591    transparent huge page integration.  (@jasone)
592  - Update zone allocator integration to work with macOS 10.12.  (@glandium)
593  - Restructure *CFLAGS configuration, so that CFLAGS behaves typically, and
594    EXTRA_CFLAGS provides a way to specify e.g. -Werror during building, but not
595    during configuration.  (@jasone, @ronawho)
596
597  Bug fixes:
598  - Fix DSS (sbrk(2)-based) allocation.  This regression was first released in
599    4.3.0.  (@jasone)
600  - Handle race in per size class utilization computation.  This functionality
601    was first released in 4.0.0.  (@interwq)
602  - Fix lock order reversal during gdump.  (@jasone)
603  - Fix/refactor tcache synchronization.  This regression was first released in
604    4.0.0.  (@jasone)
605  - Fix various JSON-formatted malloc_stats_print() bugs.  This functionality
606    was first released in 4.3.0.  (@jasone)
607  - Fix huge-aligned allocation.  This regression was first released in 4.4.0.
608    (@jasone)
609  - When transparent huge page integration is enabled, detect what state pages
610    start in according to the kernel's current operating mode, and only convert
611    arena chunks to non-huge during purging if that is not their initial state.
612    This functionality was first released in 4.4.0.  (@jasone)
613  - Fix lg_chunk clamping for the --enable-cache-oblivious --disable-fill case.
614    This regression was first released in 4.0.0.  (@jasone, @428desmo)
615  - Properly detect sparc64 when building for Linux.  (@glaubitz)
616
617* 4.4.0 (December 3, 2016)
618
619  New features:
620  - Add configure support for *-*-linux-android.  (@cferris1000, @jasone)
621  - Add the --disable-syscall configure option, for use on systems that place
622    security-motivated limitations on syscall(2).  (@jasone)
623  - Add support for Debian GNU/kFreeBSD.  (@thesam)
624
625  Optimizations:
626  - Add extent serial numbers and use them where appropriate as a sort key that
627    is higher priority than address, so that the allocation policy prefers older
628    extents.  This tends to improve locality (decrease fragmentation) when
629    memory grows downward.  (@jasone)
630  - Refactor madvise(2) configuration so that MADV_FREE is detected and utilized
631    on Linux 4.5 and newer.  (@jasone)
632  - Mark partially purged arena chunks as non-huge-page.  This improves
633    interaction with Linux's transparent huge page functionality.  (@jasone)
634
635  Bug fixes:
636  - Fix size class computations for edge conditions involving extremely large
637    allocations.  This regression was first released in 4.0.0.  (@jasone,
638    @ingvarha)
639  - Remove overly restrictive assertions related to the cactive statistic.  This
640    regression was first released in 4.1.0.  (@jasone)
641  - Implement a more reliable detection scheme for os_unfair_lock on macOS.
642    (@jszakmeister)
643
644* 4.3.1 (November 7, 2016)
645
646  Bug fixes:
647  - Fix a severe virtual memory leak.  This regression was first released in
648    4.3.0.  (@interwq, @jasone)
649  - Refactor atomic and prng APIs to restore support for 32-bit platforms that
650    use pre-C11 toolchains, e.g. FreeBSD's mips.  (@jasone)
651
652* 4.3.0 (November 4, 2016)
653
654  This is the first release that passes the test suite for multiple Windows
655  configurations, thanks in large part to @glandium setting up continuous
656  integration via AppVeyor (and Travis CI for Linux and OS X).
657
658  New features:
659  - Add "J" (JSON) support to malloc_stats_print().  (@jasone)
660  - Add Cray compiler support.  (@ronawho)
661
662  Optimizations:
663  - Add/use adaptive spinning for bootstrapping and radix tree node
664    initialization.  (@jasone)
665
666  Bug fixes:
667  - Fix large allocation to search starting in the optimal size class heap,
668    which can substantially reduce virtual memory churn and fragmentation.  This
669    regression was first released in 4.0.0.  (@mjp41, @jasone)
670  - Fix stats.arenas.<i>.nthreads accounting.  (@interwq)
671  - Fix and simplify decay-based purging.  (@jasone)
672  - Make DSS (sbrk(2)-related) operations lockless, which resolves potential
673    deadlocks during thread exit.  (@jasone)
674  - Fix over-sized allocation of radix tree leaf nodes.  (@mjp41, @ogaun,
675    @jasone)
676  - Fix over-sized allocation of arena_t (plus associated stats) data
677    structures.  (@jasone, @interwq)
678  - Fix EXTRA_CFLAGS to not affect configuration.  (@jasone)
679  - Fix a Valgrind integration bug.  (@ronawho)
680  - Disallow 0x5a junk filling when running in Valgrind.  (@jasone)
681  - Fix a file descriptor leak on Linux.  This regression was first released in
682    4.2.0.  (@vsarunas, @jasone)
683  - Fix static linking of jemalloc with glibc.  (@djwatson)
684  - Use syscall(2) rather than {open,read,close}(2) during boot on Linux.  This
685    works around other libraries' system call wrappers performing reentrant
686    allocation.  (@kspinka, @Whissi, @jasone)
687  - Fix OS X default zone replacement to work with OS X 10.12.  (@glandium,
688    @jasone)
689  - Fix cached memory management to avoid needless commit/decommit operations
690    during purging, which resolves permanent virtual memory map fragmentation
691    issues on Windows.  (@mjp41, @jasone)
692  - Fix TSD fetches to avoid (recursive) allocation.  This is relevant to
693    non-TLS and Windows configurations.  (@jasone)
694  - Fix malloc_conf overriding to work on Windows.  (@jasone)
695  - Forcibly disable lazy-lock on Windows (was forcibly *enabled*).  (@jasone)
696
697* 4.2.1 (June 8, 2016)
698
699  Bug fixes:
700  - Fix bootstrapping issues for configurations that require allocation during
701    tsd initialization (e.g. --disable-tls).  (@cferris1000, @jasone)
702  - Fix gettimeofday() version of nstime_update().  (@ronawho)
703  - Fix Valgrind regressions in calloc() and chunk_alloc_wrapper().  (@ronawho)
704  - Fix potential VM map fragmentation regression.  (@jasone)
705  - Fix opt_zero-triggered in-place huge reallocation zeroing.  (@jasone)
706  - Fix heap profiling context leaks in reallocation edge cases.  (@jasone)
707
708* 4.2.0 (May 12, 2016)
709
710  New features:
711  - Add the arena.<i>.reset mallctl, which makes it possible to discard all of
712    an arena's allocations in a single operation.  (@jasone)
713  - Add the stats.retained and stats.arenas.<i>.retained statistics.  (@jasone)
714  - Add the --with-version configure option.  (@jasone)
715  - Support --with-lg-page values larger than actual page size.  (@jasone)
716
717  Optimizations:
718  - Use pairing heaps rather than red-black trees for various hot data
719    structures.  (@djwatson, @jasone)
720  - Streamline fast paths of rtree operations.  (@jasone)
721  - Optimize the fast paths of calloc() and [m,d,sd]allocx().  (@jasone)
722  - Decommit unused virtual memory if the OS does not overcommit.  (@jasone)
723  - Specify MAP_NORESERVE on Linux if [heuristic] overcommit is active, in order
724    to avoid unfortunate interactions during fork(2).  (@jasone)
725
726  Bug fixes:
727  - Fix chunk accounting related to triggering gdump profiles.  (@jasone)
728  - Link against librt for clock_gettime(2) if glibc < 2.17.  (@jasone)
729  - Scale leak report summary according to sampling probability.  (@jasone)
730
731* 4.1.1 (May 3, 2016)
732
733  This bugfix release resolves a variety of mostly minor issues, though the
734  bitmap fix is critical for 64-bit Windows.
735
736  Bug fixes:
737  - Fix the linear scan version of bitmap_sfu() to shift by the proper amount
738    even when sizeof(long) is not the same as sizeof(void *), as on 64-bit
739    Windows.  (@jasone)
740  - Fix hashing functions to avoid unaligned memory accesses (and resulting
741    crashes).  This is relevant at least to some ARM-based platforms.
742    (@rkmisra)
743  - Fix fork()-related lock rank ordering reversals.  These reversals were
744    unlikely to cause deadlocks in practice except when heap profiling was
745    enabled and active.  (@jasone)
746  - Fix various chunk leaks in OOM code paths.  (@jasone)
747  - Fix malloc_stats_print() to print opt.narenas correctly.  (@jasone)
748  - Fix MSVC-specific build/test issues.  (@rustyx, @yuslepukhin)
749  - Fix a variety of test failures that were due to test fragility rather than
750    core bugs.  (@jasone)
751
752* 4.1.0 (February 28, 2016)
753
754  This release is primarily about optimizations, but it also incorporates a lot
755  of portability-motivated refactoring and enhancements.  Many people worked on
756  this release, to an extent that even with the omission here of minor changes
757  (see git revision history), and of the people who reported and diagnosed
758  issues, so much of the work was contributed that starting with this release,
759  changes are annotated with author credits to help reflect the collaborative
760  effort involved.
761
762  New features:
763  - Implement decay-based unused dirty page purging, a major optimization with
764    mallctl API impact.  This is an alternative to the existing ratio-based
765    unused dirty page purging, and is intended to eventually become the sole
766    purging mechanism.  New mallctls:
767    + opt.purge
768    + opt.decay_time
769    + arena.<i>.decay
770    + arena.<i>.decay_time
771    + arenas.decay_time
772    + stats.arenas.<i>.decay_time
773    (@jasone, @cevans87)
774  - Add --with-malloc-conf, which makes it possible to embed a default
775    options string during configuration.  This was motivated by the desire to
776    specify --with-malloc-conf=purge:decay , since the default must remain
777    purge:ratio until the 5.0.0 release.  (@jasone)
778  - Add MS Visual Studio 2015 support.  (@rustyx, @yuslepukhin)
779  - Make *allocx() size class overflow behavior defined.  The maximum
780    size class is now less than PTRDIFF_MAX to protect applications against
781    numerical overflow, and all allocation functions are guaranteed to indicate
782    errors rather than potentially crashing if the request size exceeds the
783    maximum size class.  (@jasone)
784  - jeprof:
785    + Add raw heap profile support.  (@jasone)
786    + Add --retain and --exclude for backtrace symbol filtering.  (@jasone)
787
788  Optimizations:
789  - Optimize the fast path to combine various bootstrapping and configuration
790    checks and execute more streamlined code in the common case.  (@interwq)
791  - Use linear scan for small bitmaps (used for small object tracking).  In
792    addition to speeding up bitmap operations on 64-bit systems, this reduces
793    allocator metadata overhead by approximately 0.2%.  (@djwatson)
794  - Separate arena_avail trees, which substantially speeds up run tree
795    operations.  (@djwatson)
796  - Use memoization (boot-time-computed table) for run quantization.  Separate
797    arena_avail trees reduced the importance of this optimization.  (@jasone)
798  - Attempt mmap-based in-place huge reallocation.  This can dramatically speed
799    up incremental huge reallocation.  (@jasone)
800
801  Incompatible changes:
802  - Make opt.narenas unsigned rather than size_t.  (@jasone)
803
804  Bug fixes:
805  - Fix stats.cactive accounting regression.  (@rustyx, @jasone)
806  - Handle unaligned keys in hash().  This caused problems for some ARM systems.
807    (@jasone, @cferris1000)
808  - Refactor arenas array.  In addition to fixing a fork-related deadlock, this
809    makes arena lookups faster and simpler.  (@jasone)
810  - Move retained memory allocation out of the default chunk allocation
811    function, to a location that gets executed even if the application installs
812    a custom chunk allocation function.  This resolves a virtual memory leak.
813    (@buchgr)
814  - Fix a potential tsd cleanup leak.  (@cferris1000, @jasone)
815  - Fix run quantization.  In practice this bug had no impact unless
816    applications requested memory with alignment exceeding one page.
817    (@jasone, @djwatson)
818  - Fix LinuxThreads-specific bootstrapping deadlock.  (Cosmin Paraschiv)
819  - jeprof:
820    + Don't discard curl options if timeout is not defined.  (@djwatson)
821    + Detect failed profile fetches.  (@djwatson)
822  - Fix stats.arenas.<i>.{dss,lg_dirty_mult,decay_time,pactive,pdirty} for
823    --disable-stats case.  (@jasone)
824
825* 4.0.4 (October 24, 2015)
826
827  This bugfix release fixes another xallocx() regression.  No other regressions
828  have come to light in over a month, so this is likely a good starting point
829  for people who prefer to wait for "dot one" releases with all the major issues
830  shaken out.
831
832  Bug fixes:
833  - Fix xallocx(..., MALLOCX_ZERO to zero the last full trailing page of large
834    allocations that have been randomly assigned an offset of 0 when
835    --enable-cache-oblivious configure option is enabled.
836
837* 4.0.3 (September 24, 2015)
838
839  This bugfix release continues the trend of xallocx() and heap profiling fixes.
840
841  Bug fixes:
842  - Fix xallocx(..., MALLOCX_ZERO) to zero all trailing bytes of large
843    allocations when --enable-cache-oblivious configure option is enabled.
844  - Fix xallocx(..., MALLOCX_ZERO) to zero trailing bytes of huge allocations
845    when resizing from/to a size class that is not a multiple of the chunk size.
846  - Fix prof_tctx_dump_iter() to filter out nodes that were created after heap
847    profile dumping started.
848  - Work around a potentially bad thread-specific data initialization
849    interaction with NPTL (glibc's pthreads implementation).
850
851* 4.0.2 (September 21, 2015)
852
853  This bugfix release addresses a few bugs specific to heap profiling.
854
855  Bug fixes:
856  - Fix ixallocx_prof_sample() to never modify nor create sampled small
857    allocations.  xallocx() is in general incapable of moving small allocations,
858    so this fix removes buggy code without loss of generality.
859  - Fix irallocx_prof_sample() to always allocate large regions, even when
860    alignment is non-zero.
861  - Fix prof_alloc_rollback() to read tdata from thread-specific data rather
862    than dereferencing a potentially invalid tctx.
863
864* 4.0.1 (September 15, 2015)
865
866  This is a bugfix release that is somewhat high risk due to the amount of
867  refactoring required to address deep xallocx() problems.  As a side effect of
868  these fixes, xallocx() now tries harder to partially fulfill requests for
869  optional extra space.  Note that a couple of minor heap profiling
870  optimizations are included, but these are better thought of as performance
871  fixes that were integral to discovering most of the other bugs.
872
873  Optimizations:
874  - Avoid a chunk metadata read in arena_prof_tctx_set(), since it is in the
875    fast path when heap profiling is enabled.  Additionally, split a special
876    case out into arena_prof_tctx_reset(), which also avoids chunk metadata
877    reads.
878  - Optimize irallocx_prof() to optimistically update the sampler state.  The
879    prior implementation appears to have been a holdover from when
880    rallocx()/xallocx() functionality was combined as rallocm().
881
882  Bug fixes:
883  - Fix TLS configuration such that it is enabled by default for platforms on
884    which it works correctly.
885  - Fix arenas_cache_cleanup() and arena_get_hard() to handle
886    allocation/deallocation within the application's thread-specific data
887    cleanup functions even after arenas_cache is torn down.
888  - Fix xallocx() bugs related to size+extra exceeding HUGE_MAXCLASS.
889  - Fix chunk purge hook calls for in-place huge shrinking reallocation to
890    specify the old chunk size rather than the new chunk size.  This bug caused
891    no correctness issues for the default chunk purge function, but was
892    visible to custom functions set via the "arena.<i>.chunk_hooks" mallctl.
893  - Fix heap profiling bugs:
894    + Fix heap profiling to distinguish among otherwise identical sample sites
895      with interposed resets (triggered via the "prof.reset" mallctl).  This bug
896      could cause data structure corruption that would most likely result in a
897      segfault.
898    + Fix irealloc_prof() to prof_alloc_rollback() on OOM.
899    + Make one call to prof_active_get_unlocked() per allocation event, and use
900      the result throughout the relevant functions that handle an allocation
901      event.  Also add a missing check in prof_realloc().  These fixes protect
902      allocation events against concurrent prof_active changes.
903    + Fix ixallocx_prof() to pass usize_max and zero to ixallocx_prof_sample()
904      in the correct order.
905    + Fix prof_realloc() to call prof_free_sampled_object() after calling
906      prof_malloc_sample_object().  Prior to this fix, if tctx and old_tctx were
907      the same, the tctx could have been prematurely destroyed.
908  - Fix portability bugs:
909    + Don't bitshift by negative amounts when encoding/decoding run sizes in
910      chunk header maps.  This affected systems with page sizes greater than 8
911      KiB.
912    + Rename index_t to szind_t to avoid an existing type on Solaris.
913    + Add JEMALLOC_CXX_THROW to the memalign() function prototype, in order to
914      match glibc and avoid compilation errors when including both
915      jemalloc/jemalloc.h and malloc.h in C++ code.
916    + Don't assume that /bin/sh is appropriate when running size_classes.sh
917      during configuration.
918    + Consider __sparcv9 a synonym for __sparc64__ when defining LG_QUANTUM.
919    + Link tests to librt if it contains clock_gettime(2).
920
921* 4.0.0 (August 17, 2015)
922
923  This version contains many speed and space optimizations, both minor and
924  major.  The major themes are generalization, unification, and simplification.
925  Although many of these optimizations cause no visible behavior change, their
926  cumulative effect is substantial.
927
928  New features:
929  - Normalize size class spacing to be consistent across the complete size
930    range.  By default there are four size classes per size doubling, but this
931    is now configurable via the --with-lg-size-class-group option.  Also add the
932    --with-lg-page, --with-lg-page-sizes, --with-lg-quantum, and
933    --with-lg-tiny-min options, which can be used to tweak page and size class
934    settings.  Impacts:
935    + Worst case performance for incrementally growing/shrinking reallocation
936      is improved because there are far fewer size classes, and therefore
937      copying happens less often.
938    + Internal fragmentation is limited to 20% for all but the smallest size
939      classes (those less than four times the quantum).  (1B + 4 KiB)
940      and (1B + 4 MiB) previously suffered nearly 50% internal fragmentation.
941    + Chunk fragmentation tends to be lower because there are fewer distinct run
942      sizes to pack.
943  - Add support for explicit tcaches.  The "tcache.create", "tcache.flush", and
944    "tcache.destroy" mallctls control tcache lifetime and flushing, and the
945    MALLOCX_TCACHE(tc) and MALLOCX_TCACHE_NONE flags to the *allocx() API
946    control which tcache is used for each operation.
947  - Implement per thread heap profiling, as well as the ability to
948    enable/disable heap profiling on a per thread basis.  Add the "prof.reset",
949    "prof.lg_sample", "thread.prof.name", "thread.prof.active",
950    "opt.prof_thread_active_init", "prof.thread_active_init", and
951    "thread.prof.active" mallctls.
952  - Add support for per arena application-specified chunk allocators, configured
953    via the "arena.<i>.chunk_hooks" mallctl.
954  - Refactor huge allocation to be managed by arenas, so that arenas now
955    function as general purpose independent allocators.  This is important in
956    the context of user-specified chunk allocators, aside from the scalability
957    benefits.  Related new statistics:
958    + The "stats.arenas.<i>.huge.allocated", "stats.arenas.<i>.huge.nmalloc",
959      "stats.arenas.<i>.huge.ndalloc", and "stats.arenas.<i>.huge.nrequests"
960      mallctls provide high level per arena huge allocation statistics.
961    + The "arenas.nhchunks", "arenas.hchunk.<i>.size",
962      "stats.arenas.<i>.hchunks.<j>.nmalloc",
963      "stats.arenas.<i>.hchunks.<j>.ndalloc",
964      "stats.arenas.<i>.hchunks.<j>.nrequests", and
965      "stats.arenas.<i>.hchunks.<j>.curhchunks" mallctls provide per size class
966      statistics.
967  - Add the 'util' column to malloc_stats_print() output, which reports the
968    proportion of available regions that are currently in use for each small
969    size class.
970  - Add "alloc" and "free" modes for for junk filling (see the "opt.junk"
971    mallctl), so that it is possible to separately enable junk filling for
972    allocation versus deallocation.
973  - Add the jemalloc-config script, which provides information about how
974    jemalloc was configured, and how to integrate it into application builds.
975  - Add metadata statistics, which are accessible via the "stats.metadata",
976    "stats.arenas.<i>.metadata.mapped", and
977    "stats.arenas.<i>.metadata.allocated" mallctls.
978  - Add the "stats.resident" mallctl, which reports the upper limit of
979    physically resident memory mapped by the allocator.
980  - Add per arena control over unused dirty page purging, via the
981    "arenas.lg_dirty_mult", "arena.<i>.lg_dirty_mult", and
982    "stats.arenas.<i>.lg_dirty_mult" mallctls.
983  - Add the "prof.gdump" mallctl, which makes it possible to toggle the gdump
984    feature on/off during program execution.
985  - Add sdallocx(), which implements sized deallocation.  The primary
986    optimization over dallocx() is the removal of a metadata read, which often
987    suffers an L1 cache miss.
988  - Add missing header includes in jemalloc/jemalloc.h, so that applications
989    only have to #include <jemalloc/jemalloc.h>.
990  - Add support for additional platforms:
991    + Bitrig
992    + Cygwin
993    + DragonFlyBSD
994    + iOS
995    + OpenBSD
996    + OpenRISC/or1k
997
998  Optimizations:
999  - Maintain dirty runs in per arena LRUs rather than in per arena trees of
1000    dirty-run-containing chunks.  In practice this change significantly reduces
1001    dirty page purging volume.
1002  - Integrate whole chunks into the unused dirty page purging machinery.  This
1003    reduces the cost of repeated huge allocation/deallocation, because it
1004    effectively introduces a cache of chunks.
1005  - Split the arena chunk map into two separate arrays, in order to increase
1006    cache locality for the frequently accessed bits.
1007  - Move small run metadata out of runs, into arena chunk headers.  This reduces
1008    run fragmentation, smaller runs reduce external fragmentation for small size
1009    classes, and packed (less uniformly aligned) metadata layout improves CPU
1010    cache set distribution.
1011  - Randomly distribute large allocation base pointer alignment relative to page
1012    boundaries in order to more uniformly utilize CPU cache sets.  This can be
1013    disabled via the --disable-cache-oblivious configure option, and queried via
1014    the "config.cache_oblivious" mallctl.
1015  - Micro-optimize the fast paths for the public API functions.
1016  - Refactor thread-specific data to reside in a single structure.  This assures
1017    that only a single TLS read is necessary per call into the public API.
1018  - Implement in-place huge allocation growing and shrinking.
1019  - Refactor rtree (radix tree for chunk lookups) to be lock-free, and make
1020    additional optimizations that reduce maximum lookup depth to one or two
1021    levels.  This resolves what was a concurrency bottleneck for per arena huge
1022    allocation, because a global data structure is critical for determining
1023    which arenas own which huge allocations.
1024
1025  Incompatible changes:
1026  - Replace --enable-cc-silence with --disable-cc-silence to suppress spurious
1027    warnings by default.
1028  - Assure that the constness of malloc_usable_size()'s return type matches that
1029    of the system implementation.
1030  - Change the heap profile dump format to support per thread heap profiling,
1031    rename pprof to jeprof, and enhance it with the --thread=<n> option.  As a
1032    result, the bundled jeprof must now be used rather than the upstream
1033    (gperftools) pprof.
1034  - Disable "opt.prof_final" by default, in order to avoid atexit(3), which can
1035    internally deadlock on some platforms.
1036  - Change the "arenas.nlruns" mallctl type from size_t to unsigned.
1037  - Replace the "stats.arenas.<i>.bins.<j>.allocated" mallctl with
1038    "stats.arenas.<i>.bins.<j>.curregs".
1039  - Ignore MALLOC_CONF in set{uid,gid,cap} binaries.
1040  - Ignore MALLOCX_ARENA(a) in dallocx(), in favor of using the
1041    MALLOCX_TCACHE(tc) and MALLOCX_TCACHE_NONE flags to control tcache usage.
1042
1043  Removed features:
1044  - Remove the *allocm() API, which is superseded by the *allocx() API.
1045  - Remove the --enable-dss options, and make dss non-optional on all platforms
1046    which support sbrk(2).
1047  - Remove the "arenas.purge" mallctl, which was obsoleted by the
1048    "arena.<i>.purge" mallctl in 3.1.0.
1049  - Remove the unnecessary "opt.valgrind" mallctl; jemalloc automatically
1050    detects whether it is running inside Valgrind.
1051  - Remove the "stats.huge.allocated", "stats.huge.nmalloc", and
1052    "stats.huge.ndalloc" mallctls.
1053  - Remove the --enable-mremap option.
1054  - Remove the "stats.chunks.current", "stats.chunks.total", and
1055    "stats.chunks.high" mallctls.
1056
1057  Bug fixes:
1058  - Fix the cactive statistic to decrease (rather than increase) when active
1059    memory decreases.  This regression was first released in 3.5.0.
1060  - Fix OOM handling in memalign() and valloc().  A variant of this bug existed
1061    in all releases since 2.0.0, which introduced these functions.
1062  - Fix an OOM-related regression in arena_tcache_fill_small(), which could
1063    cause cache corruption on OOM.  This regression was present in all releases
1064    from 2.2.0 through 3.6.0.
1065  - Fix size class overflow handling for malloc(), posix_memalign(), memalign(),
1066    calloc(), and realloc() when profiling is enabled.
1067  - Fix the "arena.<i>.dss" mallctl to return an error if "primary" or
1068    "secondary" precedence is specified, but sbrk(2) is not supported.
1069  - Fix fallback lg_floor() implementations to handle extremely large inputs.
1070  - Ensure the default purgeable zone is after the default zone on OS X.
1071  - Fix latent bugs in atomic_*().
1072  - Fix the "arena.<i>.dss" mallctl to handle read-only calls.
1073  - Fix tls_model configuration to enable the initial-exec model when possible.
1074  - Mark malloc_conf as a weak symbol so that the application can override it.
1075  - Correctly detect glibc's adaptive pthread mutexes.
1076  - Fix the --without-export configure option.
1077
1078* 3.6.0 (March 31, 2014)
1079
1080  This version contains a critical bug fix for a regression present in 3.5.0 and
1081  3.5.1.
1082
1083  Bug fixes:
1084  - Fix a regression in arena_chunk_alloc() that caused crashes during
1085    small/large allocation if chunk allocation failed.  In the absence of this
1086    bug, chunk allocation failure would result in allocation failure, e.g.  NULL
1087    return from malloc().  This regression was introduced in 3.5.0.
1088  - Fix backtracing for gcc intrinsics-based backtracing by specifying
1089    -fno-omit-frame-pointer to gcc.  Note that the application (and all the
1090    libraries it links to) must also be compiled with this option for
1091    backtracing to be reliable.
1092  - Use dss allocation precedence for huge allocations as well as small/large
1093    allocations.
1094  - Fix test assertion failure message formatting.  This bug did not manifest on
1095    x86_64 systems because of implementation subtleties in va_list.
1096  - Fix inconsequential test failures for hash and SFMT code.
1097
1098  New features:
1099  - Support heap profiling on FreeBSD.  This feature depends on the proc
1100    filesystem being mounted during heap profile dumping.
1101
1102* 3.5.1 (February 25, 2014)
1103
1104  This version primarily addresses minor bugs in test code.
1105
1106  Bug fixes:
1107  - Configure Solaris/Illumos to use MADV_FREE.
1108  - Fix junk filling for mremap(2)-based huge reallocation.  This is only
1109    relevant if configuring with the --enable-mremap option specified.
1110  - Avoid compilation failure if 'restrict' C99 keyword is not supported by the
1111    compiler.
1112  - Add a configure test for SSE2 rather than assuming it is usable on i686
1113    systems.  This fixes test compilation errors, especially on 32-bit Linux
1114    systems.
1115  - Fix mallctl argument size mismatches (size_t vs. uint64_t) in the stats unit
1116    test.
1117  - Fix/remove flawed alignment-related overflow tests.
1118  - Prevent compiler optimizations that could change backtraces in the
1119    prof_accum unit test.
1120
1121* 3.5.0 (January 22, 2014)
1122
1123  This version focuses on refactoring and automated testing, though it also
1124  includes some non-trivial heap profiling optimizations not mentioned below.
1125
1126  New features:
1127  - Add the *allocx() API, which is a successor to the experimental *allocm()
1128    API.  The *allocx() functions are slightly simpler to use because they have
1129    fewer parameters, they directly return the results of primary interest, and
1130    mallocx()/rallocx() avoid the strict aliasing pitfall that
1131    allocm()/rallocm() share with posix_memalign().  Note that *allocm() is
1132    slated for removal in the next non-bugfix release.
1133  - Add support for LinuxThreads.
1134
1135  Bug fixes:
1136  - Unless heap profiling is enabled, disable floating point code and don't link
1137    with libm.  This, in combination with e.g. EXTRA_CFLAGS=-mno-sse on x64
1138    systems, makes it possible to completely disable floating point register
1139    use.  Some versions of glibc neglect to save/restore caller-saved floating
1140    point registers during dynamic lazy symbol loading, and the symbol loading
1141    code uses whatever malloc the application happens to have linked/loaded
1142    with, the result being potential floating point register corruption.
1143  - Report ENOMEM rather than EINVAL if an OOM occurs during heap profiling
1144    backtrace creation in imemalign().  This bug impacted posix_memalign() and
1145    aligned_alloc().
1146  - Fix a file descriptor leak in a prof_dump_maps() error path.
1147  - Fix prof_dump() to close the dump file descriptor for all relevant error
1148    paths.
1149  - Fix rallocm() to use the arena specified by the ALLOCM_ARENA(s) flag for
1150    allocation, not just deallocation.
1151  - Fix a data race for large allocation stats counters.
1152  - Fix a potential infinite loop during thread exit.  This bug occurred on
1153    Solaris, and could affect other platforms with similar pthreads TSD
1154    implementations.
1155  - Don't junk-fill reallocations unless usable size changes.  This fixes a
1156    violation of the *allocx()/*allocm() semantics.
1157  - Fix growing large reallocation to junk fill new space.
1158  - Fix huge deallocation to junk fill when munmap is disabled.
1159  - Change the default private namespace prefix from empty to je_, and change
1160    --with-private-namespace-prefix so that it prepends an additional prefix
1161    rather than replacing je_.  This reduces the likelihood of applications
1162    which statically link jemalloc experiencing symbol name collisions.
1163  - Add missing private namespace mangling (relevant when
1164    --with-private-namespace is specified).
1165  - Add and use JEMALLOC_INLINE_C so that static inline functions are marked as
1166    static even for debug builds.
1167  - Add a missing mutex unlock in a malloc_init_hard() error path.  In practice
1168    this error path is never executed.
1169  - Fix numerous bugs in malloc_strotumax() error handling/reporting.  These
1170    bugs had no impact except for malformed inputs.
1171  - Fix numerous bugs in malloc_snprintf().  These bugs were not exercised by
1172    existing calls, so they had no impact.
1173
1174* 3.4.1 (October 20, 2013)
1175
1176  Bug fixes:
1177  - Fix a race in the "arenas.extend" mallctl that could cause memory corruption
1178    of internal data structures and subsequent crashes.
1179  - Fix Valgrind integration flaws that caused Valgrind warnings about reads of
1180    uninitialized memory in:
1181    + arena chunk headers
1182    + internal zero-initialized data structures (relevant to tcache and prof
1183      code)
1184  - Preserve errno during the first allocation.  A readlink(2) call during
1185    initialization fails unless /etc/malloc.conf exists, so errno was typically
1186    set during the first allocation prior to this fix.
1187  - Fix compilation warnings reported by gcc 4.8.1.
1188
1189* 3.4.0 (June 2, 2013)
1190
1191  This version is essentially a small bugfix release, but the addition of
1192  aarch64 support requires that the minor version be incremented.
1193
1194  Bug fixes:
1195  - Fix race-triggered deadlocks in chunk_record().  These deadlocks were
1196    typically triggered by multiple threads concurrently deallocating huge
1197    objects.
1198
1199  New features:
1200  - Add support for the aarch64 architecture.
1201
1202* 3.3.1 (March 6, 2013)
1203
1204  This version fixes bugs that are typically encountered only when utilizing
1205  custom run-time options.
1206
1207  Bug fixes:
1208  - Fix a locking order bug that could cause deadlock during fork if heap
1209    profiling were enabled.
1210  - Fix a chunk recycling bug that could cause the allocator to lose track of
1211    whether a chunk was zeroed.  On FreeBSD, NetBSD, and OS X, it could cause
1212    corruption if allocating via sbrk(2) (unlikely unless running with the
1213    "dss:primary" option specified).  This was completely harmless on Linux
1214    unless using mlockall(2) (and unlikely even then, unless the
1215    --disable-munmap configure option or the "dss:primary" option was
1216    specified).  This regression was introduced in 3.1.0 by the
1217    mlockall(2)/madvise(2) interaction fix.
1218  - Fix TLS-related memory corruption that could occur during thread exit if the
1219    thread never allocated memory.  Only the quarantine and prof facilities were
1220    susceptible.
1221  - Fix two quarantine bugs:
1222    + Internal reallocation of the quarantined object array leaked the old
1223      array.
1224    + Reallocation failure for internal reallocation of the quarantined object
1225      array (very unlikely) resulted in memory corruption.
1226  - Fix Valgrind integration to annotate all internally allocated memory in a
1227    way that keeps Valgrind happy about internal data structure access.
1228  - Fix building for s390 systems.
1229
1230* 3.3.0 (January 23, 2013)
1231
1232  This version includes a few minor performance improvements in addition to the
1233  listed new features and bug fixes.
1234
1235  New features:
1236  - Add clipping support to lg_chunk option processing.
1237  - Add the --enable-ivsalloc option.
1238  - Add the --without-export option.
1239  - Add the --disable-zone-allocator option.
1240
1241  Bug fixes:
1242  - Fix "arenas.extend" mallctl to output the number of arenas.
1243  - Fix chunk_recycle() to unconditionally inform Valgrind that returned memory
1244    is undefined.
1245  - Fix build break on FreeBSD related to alloca.h.
1246
1247* 3.2.0 (November 9, 2012)
1248
1249  In addition to a couple of bug fixes, this version modifies page run
1250  allocation and dirty page purging algorithms in order to better control
1251  page-level virtual memory fragmentation.
1252
1253  Incompatible changes:
1254  - Change the "opt.lg_dirty_mult" default from 5 to 3 (32:1 to 8:1).
1255
1256  Bug fixes:
1257  - Fix dss/mmap allocation precedence code to use recyclable mmap memory only
1258    after primary dss allocation fails.
1259  - Fix deadlock in the "arenas.purge" mallctl.  This regression was introduced
1260    in 3.1.0 by the addition of the "arena.<i>.purge" mallctl.
1261
1262* 3.1.0 (October 16, 2012)
1263
1264  New features:
1265  - Auto-detect whether running inside Valgrind, thus removing the need to
1266    manually specify MALLOC_CONF=valgrind:true.
1267  - Add the "arenas.extend" mallctl, which allows applications to create
1268    manually managed arenas.
1269  - Add the ALLOCM_ARENA() flag for {,r,d}allocm().
1270  - Add the "opt.dss", "arena.<i>.dss", and "stats.arenas.<i>.dss" mallctls,
1271    which provide control over dss/mmap precedence.
1272  - Add the "arena.<i>.purge" mallctl, which obsoletes "arenas.purge".
1273  - Define LG_QUANTUM for hppa.
1274
1275  Incompatible changes:
1276  - Disable tcache by default if running inside Valgrind, in order to avoid
1277    making unallocated objects appear reachable to Valgrind.
1278  - Drop const from malloc_usable_size() argument on Linux.
1279
1280  Bug fixes:
1281  - Fix heap profiling crash if sampled object is freed via realloc(p, 0).
1282  - Remove const from __*_hook variable declarations, so that glibc can modify
1283    them during process forking.
1284  - Fix mlockall(2)/madvise(2) interaction.
1285  - Fix fork(2)-related deadlocks.
1286  - Fix error return value for "thread.tcache.enabled" mallctl.
1287
1288* 3.0.0 (May 11, 2012)
1289
1290  Although this version adds some major new features, the primary focus is on
1291  internal code cleanup that facilitates maintainability and portability, most
1292  of which is not reflected in the ChangeLog.  This is the first release to
1293  incorporate substantial contributions from numerous other developers, and the
1294  result is a more broadly useful allocator (see the git revision history for
1295  contribution details).  Note that the license has been unified, thanks to
1296  Facebook granting a license under the same terms as the other copyright
1297  holders (see COPYING).
1298
1299  New features:
1300  - Implement Valgrind support, redzones, and quarantine.
1301  - Add support for additional platforms:
1302    + FreeBSD
1303    + Mac OS X Lion
1304    + MinGW
1305    + Windows (no support yet for replacing the system malloc)
1306  - Add support for additional architectures:
1307    + MIPS
1308    + SH4
1309    + Tilera
1310  - Add support for cross compiling.
1311  - Add nallocm(), which rounds a request size up to the nearest size class
1312    without actually allocating.
1313  - Implement aligned_alloc() (blame C11).
1314  - Add the "thread.tcache.enabled" mallctl.
1315  - Add the "opt.prof_final" mallctl.
1316  - Update pprof (from gperftools 2.0).
1317  - Add the --with-mangling option.
1318  - Add the --disable-experimental option.
1319  - Add the --disable-munmap option, and make it the default on Linux.
1320  - Add the --enable-mremap option, which disables use of mremap(2) by default.
1321
1322  Incompatible changes:
1323  - Enable stats by default.
1324  - Enable fill by default.
1325  - Disable lazy locking by default.
1326  - Rename the "tcache.flush" mallctl to "thread.tcache.flush".
1327  - Rename the "arenas.pagesize" mallctl to "arenas.page".
1328  - Change the "opt.lg_prof_sample" default from 0 to 19 (1 B to 512 KiB).
1329  - Change the "opt.prof_accum" default from true to false.
1330
1331  Removed features:
1332  - Remove the swap feature, including the "config.swap", "swap.avail",
1333    "swap.prezeroed", "swap.nfds", and "swap.fds" mallctls.
1334  - Remove highruns statistics, including the
1335    "stats.arenas.<i>.bins.<j>.highruns" and
1336    "stats.arenas.<i>.lruns.<j>.highruns" mallctls.
1337  - As part of small size class refactoring, remove the "opt.lg_[qc]space_max",
1338    "arenas.cacheline", "arenas.subpage", "arenas.[tqcs]space_{min,max}", and
1339    "arenas.[tqcs]bins" mallctls.
1340  - Remove the "arenas.chunksize" mallctl.
1341  - Remove the "opt.lg_prof_tcmax" option.
1342  - Remove the "opt.lg_prof_bt_max" option.
1343  - Remove the "opt.lg_tcache_gc_sweep" option.
1344  - Remove the --disable-tiny option, including the "config.tiny" mallctl.
1345  - Remove the --enable-dynamic-page-shift configure option.
1346  - Remove the --enable-sysv configure option.
1347
1348  Bug fixes:
1349  - Fix a statistics-related bug in the "thread.arena" mallctl that could cause
1350    invalid statistics and crashes.
1351  - Work around TLS deallocation via free() on Linux.  This bug could cause
1352    write-after-free memory corruption.
1353  - Fix a potential deadlock that could occur during interval- and
1354    growth-triggered heap profile dumps.
1355  - Fix large calloc() zeroing bugs due to dropping chunk map unzeroed flags.
1356  - Fix chunk_alloc_dss() to stop claiming memory is zeroed.  This bug could
1357    cause memory corruption and crashes with --enable-dss specified.
1358  - Fix fork-related bugs that could cause deadlock in children between fork
1359    and exec.
1360  - Fix malloc_stats_print() to honor 'b' and 'l' in the opts parameter.
1361  - Fix realloc(p, 0) to act like free(p).
1362  - Do not enforce minimum alignment in memalign().
1363  - Check for NULL pointer in malloc_usable_size().
1364  - Fix an off-by-one heap profile statistics bug that could be observed in
1365    interval- and growth-triggered heap profiles.
1366  - Fix the "epoch" mallctl to update cached stats even if the passed in epoch
1367    is 0.
1368  - Fix bin->runcur management to fix a layout policy bug.  This bug did not
1369    affect correctness.
1370  - Fix a bug in choose_arena_hard() that potentially caused more arenas to be
1371    initialized than necessary.
1372  - Add missing "opt.lg_tcache_max" mallctl implementation.
1373  - Use glibc allocator hooks to make mixed allocator usage less likely.
1374  - Fix build issues for --disable-tcache.
1375  - Don't mangle pthread_create() when --with-private-namespace is specified.
1376
1377* 2.2.5 (November 14, 2011)
1378
1379  Bug fixes:
1380  - Fix huge_ralloc() race when using mremap(2).  This is a serious bug that
1381    could cause memory corruption and/or crashes.
1382  - Fix huge_ralloc() to maintain chunk statistics.
1383  - Fix malloc_stats_print(..., "a") output.
1384
1385* 2.2.4 (November 5, 2011)
1386
1387  Bug fixes:
1388  - Initialize arenas_tsd before using it.  This bug existed for 2.2.[0-3], as
1389    well as for --disable-tls builds in earlier releases.
1390  - Do not assume a 4 KiB page size in test/rallocm.c.
1391
1392* 2.2.3 (August 31, 2011)
1393
1394  This version fixes numerous bugs related to heap profiling.
1395
1396  Bug fixes:
1397  - Fix a prof-related race condition.  This bug could cause memory corruption,
1398    but only occurred in non-default configurations (prof_accum:false).
1399  - Fix off-by-one backtracing issues (make sure that prof_alloc_prep() is
1400    excluded from backtraces).
1401  - Fix a prof-related bug in realloc() (only triggered by OOM errors).
1402  - Fix prof-related bugs in allocm() and rallocm().
1403  - Fix prof_tdata_cleanup() for --disable-tls builds.
1404  - Fix a relative include path, to fix objdir builds.
1405
1406* 2.2.2 (July 30, 2011)
1407
1408  Bug fixes:
1409  - Fix a build error for --disable-tcache.
1410  - Fix assertions in arena_purge() (for real this time).
1411  - Add the --with-private-namespace option.  This is a workaround for symbol
1412    conflicts that can inadvertently arise when using static libraries.
1413
1414* 2.2.1 (March 30, 2011)
1415
1416  Bug fixes:
1417  - Implement atomic operations for x86/x64.  This fixes compilation failures
1418    for versions of gcc that are still in wide use.
1419  - Fix an assertion in arena_purge().
1420
1421* 2.2.0 (March 22, 2011)
1422
1423  This version incorporates several improvements to algorithms and data
1424  structures that tend to reduce fragmentation and increase speed.
1425
1426  New features:
1427  - Add the "stats.cactive" mallctl.
1428  - Update pprof (from google-perftools 1.7).
1429  - Improve backtracing-related configuration logic, and add the
1430    --disable-prof-libgcc option.
1431
1432  Bug fixes:
1433  - Change default symbol visibility from "internal", to "hidden", which
1434    decreases the overhead of library-internal function calls.
1435  - Fix symbol visibility so that it is also set on OS X.
1436  - Fix a build dependency regression caused by the introduction of the .pic.o
1437    suffix for PIC object files.
1438  - Add missing checks for mutex initialization failures.
1439  - Don't use libgcc-based backtracing except on x64, where it is known to work.
1440  - Fix deadlocks on OS X that were due to memory allocation in
1441    pthread_mutex_lock().
1442  - Heap profiling-specific fixes:
1443    + Fix memory corruption due to integer overflow in small region index
1444      computation, when using a small enough sample interval that profiling
1445      context pointers are stored in small run headers.
1446    + Fix a bootstrap ordering bug that only occurred with TLS disabled.
1447    + Fix a rallocm() rsize bug.
1448    + Fix error detection bugs for aligned memory allocation.
1449
1450* 2.1.3 (March 14, 2011)
1451
1452  Bug fixes:
1453  - Fix a cpp logic regression (due to the "thread.{de,}allocatedp" mallctl fix
1454    for OS X in 2.1.2).
1455  - Fix a "thread.arena" mallctl bug.
1456  - Fix a thread cache stats merging bug.
1457
1458* 2.1.2 (March 2, 2011)
1459
1460  Bug fixes:
1461  - Fix "thread.{de,}allocatedp" mallctl for OS X.
1462  - Add missing jemalloc.a to build system.
1463
1464* 2.1.1 (January 31, 2011)
1465
1466  Bug fixes:
1467  - Fix aligned huge reallocation (affected allocm()).
1468  - Fix the ALLOCM_LG_ALIGN macro definition.
1469  - Fix a heap dumping deadlock.
1470  - Fix a "thread.arena" mallctl bug.
1471
1472* 2.1.0 (December 3, 2010)
1473
1474  This version incorporates some optimizations that can't quite be considered
1475  bug fixes.
1476
1477  New features:
1478  - Use Linux's mremap(2) for huge object reallocation when possible.
1479  - Avoid locking in mallctl*() when possible.
1480  - Add the "thread.[de]allocatedp" mallctl's.
1481  - Convert the manual page source from roff to DocBook, and generate both roff
1482    and HTML manuals.
1483
1484  Bug fixes:
1485  - Fix a crash due to incorrect bootstrap ordering.  This only impacted
1486    --enable-debug --enable-dss configurations.
1487  - Fix a minor statistics bug for mallctl("swap.avail", ...).
1488
1489* 2.0.1 (October 29, 2010)
1490
1491  Bug fixes:
1492  - Fix a race condition in heap profiling that could cause undefined behavior
1493    if "opt.prof_accum" were disabled.
1494  - Add missing mutex unlocks for some OOM error paths in the heap profiling
1495    code.
1496  - Fix a compilation error for non-C99 builds.
1497
1498* 2.0.0 (October 24, 2010)
1499
1500  This version focuses on the experimental *allocm() API, and on improved
1501  run-time configuration/introspection.  Nonetheless, numerous performance
1502  improvements are also included.
1503
1504  New features:
1505  - Implement the experimental {,r,s,d}allocm() API, which provides a superset
1506    of the functionality available via malloc(), calloc(), posix_memalign(),
1507    realloc(), malloc_usable_size(), and free().  These functions can be used to
1508    allocate/reallocate aligned zeroed memory, ask for optional extra memory
1509    during reallocation, prevent object movement during reallocation, etc.
1510  - Replace JEMALLOC_OPTIONS/JEMALLOC_PROF_PREFIX with MALLOC_CONF, which is
1511    more human-readable, and more flexible.  For example:
1512      JEMALLOC_OPTIONS=AJP
1513    is now:
1514      MALLOC_CONF=abort:true,fill:true,stats_print:true
1515  - Port to Apple OS X.  Sponsored by Mozilla.
1516  - Make it possible for the application to control thread-->arena mappings via
1517    the "thread.arena" mallctl.
1518  - Add compile-time support for all TLS-related functionality via pthreads TSD.
1519    This is mainly of interest for OS X, which does not support TLS, but has a
1520    TSD implementation with similar performance.
1521  - Override memalign() and valloc() if they are provided by the system.
1522  - Add the "arenas.purge" mallctl, which can be used to synchronously purge all
1523    dirty unused pages.
1524  - Make cumulative heap profiling data optional, so that it is possible to
1525    limit the amount of memory consumed by heap profiling data structures.
1526  - Add per thread allocation counters that can be accessed via the
1527    "thread.allocated" and "thread.deallocated" mallctls.
1528
1529  Incompatible changes:
1530  - Remove JEMALLOC_OPTIONS and malloc_options (see MALLOC_CONF above).
1531  - Increase default backtrace depth from 4 to 128 for heap profiling.
1532  - Disable interval-based profile dumps by default.
1533
1534  Bug fixes:
1535  - Remove bad assertions in fork handler functions.  These assertions could
1536    cause aborts for some combinations of configure settings.
1537  - Fix strerror_r() usage to deal with non-standard semantics in GNU libc.
1538  - Fix leak context reporting.  This bug tended to cause the number of contexts
1539    to be underreported (though the reported number of objects and bytes were
1540    correct).
1541  - Fix a realloc() bug for large in-place growing reallocation.  This bug could
1542    cause memory corruption, but it was hard to trigger.
1543  - Fix an allocation bug for small allocations that could be triggered if
1544    multiple threads raced to create a new run of backing pages.
1545  - Enhance the heap profiler to trigger samples based on usable size, rather
1546    than request size.
1547  - Fix a heap profiling bug due to sometimes losing track of requested object
1548    size for sampled objects.
1549
1550* 1.0.3 (August 12, 2010)
1551
1552  Bug fixes:
1553  - Fix the libunwind-based implementation of stack backtracing (used for heap
1554    profiling).  This bug could cause zero-length backtraces to be reported.
1555  - Add a missing mutex unlock in library initialization code.  If multiple
1556    threads raced to initialize malloc, some of them could end up permanently
1557    blocked.
1558
1559* 1.0.2 (May 11, 2010)
1560
1561  Bug fixes:
1562  - Fix junk filling of large objects, which could cause memory corruption.
1563  - Add MAP_NORESERVE support for chunk mapping, because otherwise virtual
1564    memory limits could cause swap file configuration to fail.  Contributed by
1565    Jordan DeLong.
1566
1567* 1.0.1 (April 14, 2010)
1568
1569  Bug fixes:
1570  - Fix compilation when --enable-fill is specified.
1571  - Fix threads-related profiling bugs that affected accuracy and caused memory
1572    to be leaked during thread exit.
1573  - Fix dirty page purging race conditions that could cause crashes.
1574  - Fix crash in tcache flushing code during thread destruction.
1575
1576* 1.0.0 (April 11, 2010)
1577
1578  This release focuses on speed and run-time introspection.  Numerous
1579  algorithmic improvements make this release substantially faster than its
1580  predecessors.
1581
1582  New features:
1583  - Implement autoconf-based configuration system.
1584  - Add mallctl*(), for the purposes of introspection and run-time
1585    configuration.
1586  - Make it possible for the application to manually flush a thread's cache, via
1587    the "tcache.flush" mallctl.
1588  - Base maximum dirty page count on proportion of active memory.
1589  - Compute various additional run-time statistics, including per size class
1590    statistics for large objects.
1591  - Expose malloc_stats_print(), which can be called repeatedly by the
1592    application.
1593  - Simplify the malloc_message() signature to only take one string argument,
1594    and incorporate an opaque data pointer argument for use by the application
1595    in combination with malloc_stats_print().
1596  - Add support for allocation backed by one or more swap files, and allow the
1597    application to disable over-commit if swap files are in use.
1598  - Implement allocation profiling and leak checking.
1599
1600  Removed features:
1601  - Remove the dynamic arena rebalancing code, since thread-specific caching
1602    reduces its utility.
1603
1604  Bug fixes:
1605  - Modify chunk allocation to work when address space layout randomization
1606    (ASLR) is in use.
1607  - Fix thread cleanup bugs related to TLS destruction.
1608  - Handle 0-size allocation requests in posix_memalign().
1609  - Fix a chunk leak.  The leaked chunks were never touched, so this impacted
1610    virtual memory usage, but not physical memory usage.
1611
1612* linux_2008082[78]a (August 27/28, 2008)
1613
1614  These snapshot releases are the simple result of incorporating Linux-specific
1615  support into the FreeBSD malloc sources.
1616
1617--------------------------------------------------------------------------------
1618vim:filetype=text:textwidth=80
1619