xref: /freebsd/contrib/jemalloc/ChangeLog (revision 0d66206fff44f864ea8a4b220c3a53b4caa959a0)
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* 5.2.1 (August 5, 2019)
8
9  This release is primarily about Windows.  A critical virtual memory leak is
10  resolved on all Windows platforms.  The regression was present in all releases
11  since 5.0.0.
12
13  Bug fixes:
14  - Fix a severe virtual memory leak on Windows.  This regression was first
15    released in 5.0.0.  (@Ignition, @j0t, @frederik-h, @davidtgoldblatt,
16    @interwq)
17  - Fix size 0 handling in posix_memalign().  This regression was first released
18    in 5.2.0.  (@interwq)
19  - Fix the prof_log unit test which may observe unexpected backtraces from
20    compiler optimizations.  The test was first added in 5.2.0.  (@marxin,
21    @gnzlbg, @interwq)
22  - Fix the declaration of the extent_avail tree.  This regression was first
23    released in 5.1.0.  (@zoulasc)
24  - Fix an incorrect reference in jeprof.  This functionality was first released
25    in 3.0.0.  (@prehistoric-penguin)
26  - Fix an assertion on the deallocation fast-path.  This regression was first
27    released in 5.2.0.  (@yinan1048576)
28  - Fix the TLS_MODEL attribute in headers.  This regression was first released
29    in 5.0.0.  (@zoulasc, @interwq)
30
31  Optimizations and refactors:
32  - Implement opt.retain on Windows and enable by default on 64-bit.  (@interwq,
33    @davidtgoldblatt)
34  - Optimize away a branch on the operator delete[] path.  (@mgrice)
35  - Add format annotation to the format generator function.  (@zoulasc)
36  - Refactor and improve the size class header generation.  (@yinan1048576)
37  - Remove best fit.  (@djwatson)
38  - Avoid blocking on background thread locks for stats.  (@oranagra, @interwq)
39
40* 5.2.0 (April 2, 2019)
41
42  This release includes a few notable improvements, which are summarized below:
43  1) improved fast-path performance from the optimizations by @djwatson; 2)
44  reduced virtual memory fragmentation and metadata usage; and 3) bug fixes on
45  setting the number of background threads.  In addition, peak / spike memory
46  usage is improved with certain allocation patterns.  As usual, the release and
47  prior dev versions have gone through large-scale production testing.
48
49  New features:
50  - Implement oversize_threshold, which uses a dedicated arena for allocations
51    crossing the specified threshold to reduce fragmentation.  (@interwq)
52  - Add extents usage information to stats.  (@tyleretzel)
53  - Log time information for sampled allocations.  (@tyleretzel)
54  - Support 0 size in sdallocx.  (@djwatson)
55  - Output rate for certain counters in malloc_stats.  (@zinoale)
56  - Add configure option --enable-readlinkat, which allows the use of readlinkat
57    over readlink.  (@davidtgoldblatt)
58  - Add configure options --{enable,disable}-{static,shared} to allow not
59    building unwanted libraries.  (@Ericson2314)
60  - Add configure option --disable-libdl to enable fully static builds.
61    (@interwq)
62  - Add mallctl interfaces:
63	+ opt.oversize_threshold (@interwq)
64	+ stats.arenas.<i>.extent_avail (@tyleretzel)
65	+ stats.arenas.<i>.extents.<j>.n{dirty,muzzy,retained} (@tyleretzel)
66	+ stats.arenas.<i>.extents.<j>.{dirty,muzzy,retained}_bytes
67	  (@tyleretzel)
68
69  Portability improvements:
70  - Update MSVC builds.  (@maksqwe, @rustyx)
71  - Workaround a compiler optimizer bug on s390x.  (@rkmisra)
72  - Make use of pthread_set_name_np(3) on FreeBSD.  (@trasz)
73  - Implement malloc_getcpu() to enable percpu_arena for windows.  (@santagada)
74  - Link against -pthread instead of -lpthread.  (@paravoid)
75  - Make background_thread not dependent on libdl.  (@interwq)
76  - Add stringify to fix a linker directive issue on MSVC.  (@daverigby)
77  - Detect and fall back when 8-bit atomics are unavailable.  (@interwq)
78  - Fall back to the default pthread_create if dlsym(3) fails.  (@interwq)
79
80  Optimizations and refactors:
81  - Refactor the TSD module.  (@davidtgoldblatt)
82  - Avoid taking extents_muzzy mutex when muzzy is disabled.  (@interwq)
83  - Avoid taking large_mtx for auto arenas on the tcache flush path.  (@interwq)
84  - Optimize ixalloc by avoiding a size lookup.  (@interwq)
85  - Implement opt.oversize_threshold which uses a dedicated arena for requests
86    crossing the threshold, also eagerly purges the oversize extents.  Default
87    the threshold to 8 MiB.  (@interwq)
88  - Clean compilation with -Wextra.  (@gnzlbg, @jasone)
89  - Refactor the size class module.  (@davidtgoldblatt)
90  - Refactor the stats emitter.  (@tyleretzel)
91  - Optimize pow2_ceil.  (@rkmisra)
92  - Avoid runtime detection of lazy purging on FreeBSD.  (@trasz)
93  - Optimize mmap(2) alignment handling on FreeBSD.  (@trasz)
94  - Improve error handling for THP state initialization.  (@jsteemann)
95  - Rework the malloc() fast path.  (@djwatson)
96  - Rework the free() fast path.  (@djwatson)
97  - Refactor and optimize the tcache fill / flush paths.  (@djwatson)
98  - Optimize sync / lwsync on PowerPC.  (@chmeeedalf)
99  - Bypass extent_dalloc() when retain is enabled.  (@interwq)
100  - Optimize the locking on large deallocation.  (@interwq)
101  - Reduce the number of pages committed from sanity checking in debug build.
102    (@trasz, @interwq)
103  - Deprecate OSSpinLock.  (@interwq)
104  - Lower the default number of background threads to 4 (when the feature
105    is enabled).  (@interwq)
106  - Optimize the trylock spin wait.  (@djwatson)
107  - Use arena index for arena-matching checks.  (@interwq)
108  - Avoid forced decay on thread termination when using background threads.
109    (@interwq)
110  - Disable muzzy decay by default.  (@djwatson, @interwq)
111  - Only initialize libgcc unwinder when profiling is enabled.  (@paravoid,
112    @interwq)
113
114  Bug fixes (all only relevant to jemalloc 5.x):
115  - Fix background thread index issues with max_background_threads.  (@djwatson,
116    @interwq)
117  - Fix stats output for opt.lg_extent_max_active_fit.  (@interwq)
118  - Fix opt.prof_prefix initialization.  (@davidtgoldblatt)
119  - Properly trigger decay on tcache destroy.  (@interwq, @amosbird)
120  - Fix tcache.flush.  (@interwq)
121  - Detect whether explicit extent zero out is necessary with huge pages or
122    custom extent hooks, which may change the purge semantics.  (@interwq)
123  - Fix a side effect caused by extent_max_active_fit combined with decay-based
124    purging, where freed extents can accumulate and not be reused for an
125    extended period of time.  (@interwq, @mpghf)
126  - Fix a missing unlock on extent register error handling.  (@zoulasc)
127
128  Testing:
129  - Simplify the Travis script output.  (@gnzlbg)
130  - Update the test scripts for FreeBSD.  (@devnexen)
131  - Add unit tests for the producer-consumer pattern.  (@interwq)
132  - Add Cirrus-CI config for FreeBSD builds.  (@jasone)
133  - Add size-matching sanity checks on tcache flush.  (@davidtgoldblatt,
134    @interwq)
135
136  Incompatible changes:
137  - Remove --with-lg-page-sizes.  (@davidtgoldblatt)
138
139  Documentation:
140  - Attempt to build docs by default, however skip doc building when xsltproc
141    is missing. (@interwq, @cmuellner)
142
143* 5.1.0 (May 4, 2018)
144
145  This release is primarily about fine-tuning, ranging from several new features
146  to numerous notable performance and portability enhancements.  The release and
147  prior dev versions have been running in multiple large scale applications for
148  months, and the cumulative improvements are substantial in many cases.
149
150  Given the long and successful production runs, this release is likely a good
151  candidate for applications to upgrade, from both jemalloc 5.0 and before.  For
152  performance-critical applications, the newly added TUNING.md provides
153  guidelines on jemalloc tuning.
154
155  New features:
156  - Implement transparent huge page support for internal metadata.  (@interwq)
157  - Add opt.thp to allow enabling / disabling transparent huge pages for all
158    mappings.  (@interwq)
159  - Add maximum background thread count option.  (@djwatson)
160  - Allow prof_active to control opt.lg_prof_interval and prof.gdump.
161    (@interwq)
162  - Allow arena index lookup based on allocation addresses via mallctl.
163    (@lionkov)
164  - Allow disabling initial-exec TLS model.  (@davidtgoldblatt, @KenMacD)
165  - Add opt.lg_extent_max_active_fit to set the max ratio between the size of
166    the active extent selected (to split off from) and the size of the requested
167    allocation.  (@interwq, @davidtgoldblatt)
168  - Add retain_grow_limit to set the max size when growing virtual address
169    space.  (@interwq)
170  - Add mallctl interfaces:
171    + arena.<i>.retain_grow_limit  (@interwq)
172    + arenas.lookup  (@lionkov)
173    + max_background_threads  (@djwatson)
174    + opt.lg_extent_max_active_fit  (@interwq)
175    + opt.max_background_threads  (@djwatson)
176    + opt.metadata_thp  (@interwq)
177    + opt.thp  (@interwq)
178    + stats.metadata_thp  (@interwq)
179
180  Portability improvements:
181  - Support GNU/kFreeBSD configuration.  (@paravoid)
182  - Support m68k, nios2 and SH3 architectures.  (@paravoid)
183  - Fall back to FD_CLOEXEC when O_CLOEXEC is unavailable.  (@zonyitoo)
184  - Fix symbol listing for cross-compiling.  (@tamird)
185  - Fix high bits computation on ARM.  (@davidtgoldblatt, @paravoid)
186  - Disable the CPU_SPINWAIT macro for Power.  (@davidtgoldblatt, @marxin)
187  - Fix MSVC 2015 & 2017 builds.  (@rustyx)
188  - Improve RISC-V support.  (@EdSchouten)
189  - Set name mangling script in strict mode.  (@nicolov)
190  - Avoid MADV_HUGEPAGE on ARM.  (@marxin)
191  - Modify configure to determine return value of strerror_r.
192    (@davidtgoldblatt, @cferris1000)
193  - Make sure CXXFLAGS is tested with CPP compiler.  (@nehaljwani)
194  - Fix 32-bit build on MSVC.  (@rustyx)
195  - Fix external symbol on MSVC.  (@maksqwe)
196  - Avoid a printf format specifier warning.  (@jasone)
197  - Add configure option --disable-initial-exec-tls which can allow jemalloc to
198    be dynamically loaded after program startup.  (@davidtgoldblatt, @KenMacD)
199  - AArch64: Add ILP32 support.  (@cmuellner)
200  - Add --with-lg-vaddr configure option to support cross compiling.
201    (@cmuellner, @davidtgoldblatt)
202
203  Optimizations and refactors:
204  - Improve active extent fit with extent_max_active_fit.  This considerably
205    reduces fragmentation over time and improves virtual memory and metadata
206    usage.  (@davidtgoldblatt, @interwq)
207  - Eagerly coalesce large extents to reduce fragmentation.  (@interwq)
208  - sdallocx: only read size info when page aligned (i.e. possibly sampled),
209    which speeds up the sized deallocation path significantly.  (@interwq)
210  - Avoid attempting new mappings for in place expansion with retain, since
211    it rarely succeeds in practice and causes high overhead.  (@interwq)
212  - Refactor OOM handling in newImpl.  (@wqfish)
213  - Add internal fine-grained logging functionality for debugging use.
214    (@davidtgoldblatt)
215  - Refactor arena / tcache interactions.  (@davidtgoldblatt)
216  - Refactor extent management with dumpable flag.  (@davidtgoldblatt)
217  - Add runtime detection of lazy purging.  (@interwq)
218  - Use pairing heap instead of red-black tree for extents_avail.  (@djwatson)
219  - Use sysctl on startup in FreeBSD.  (@trasz)
220  - Use thread local prng state instead of atomic.  (@djwatson)
221  - Make decay to always purge one more extent than before, because in
222    practice large extents are usually the ones that cross the decay threshold.
223    Purging the additional extent helps save memory as well as reduce VM
224    fragmentation.  (@interwq)
225  - Fast division by dynamic values.  (@davidtgoldblatt)
226  - Improve the fit for aligned allocation.  (@interwq, @edwinsmith)
227  - Refactor extent_t bitpacking.  (@rkmisra)
228  - Optimize the generated assembly for ticker operations.  (@davidtgoldblatt)
229  - Convert stats printing to use a structured text emitter.  (@davidtgoldblatt)
230  - Remove preserve_lru feature for extents management.  (@djwatson)
231  - Consolidate two memory loads into one on the fast deallocation path.
232    (@davidtgoldblatt, @interwq)
233
234  Bug fixes (most of the issues are only relevant to jemalloc 5.0):
235  - Fix deadlock with multithreaded fork in OS X.  (@davidtgoldblatt)
236  - Validate returned file descriptor before use.  (@zonyitoo)
237  - Fix a few background thread initialization and shutdown issues.  (@interwq)
238  - Fix an extent coalesce + decay race by taking both coalescing extents off
239    the LRU list.  (@interwq)
240  - Fix potentially unbound increase during decay, caused by one thread keep
241    stashing memory to purge while other threads generating new pages.  The
242    number of pages to purge is checked to prevent this.  (@interwq)
243  - Fix a FreeBSD bootstrap assertion.  (@strejda, @interwq)
244  - Handle 32 bit mutex counters.  (@rkmisra)
245  - Fix a indexing bug when creating background threads.  (@davidtgoldblatt,
246    @binliu19)
247  - Fix arguments passed to extent_init.  (@yuleniwo, @interwq)
248  - Fix addresses used for ordering mutexes.  (@rkmisra)
249  - Fix abort_conf processing during bootstrap.  (@interwq)
250  - Fix include path order for out-of-tree builds.  (@cmuellner)
251
252  Incompatible changes:
253  - Remove --disable-thp.  (@interwq)
254  - Remove mallctl interfaces:
255    + config.thp  (@interwq)
256
257  Documentation:
258  - Add TUNING.md.  (@interwq, @davidtgoldblatt, @djwatson)
259
260* 5.0.1 (July 1, 2017)
261
262  This bugfix release fixes several issues, most of which are obscure enough
263  that typical applications are not impacted.
264
265  Bug fixes:
266  - Update decay->nunpurged before purging, in order to avoid potential update
267    races and subsequent incorrect purging volume.  (@interwq)
268  - Only abort on dlsym(3) error if the failure impacts an enabled feature (lazy
269    locking and/or background threads).  This mitigates an initialization
270    failure bug for which we still do not have a clear reproduction test case.
271    (@interwq)
272  - Modify tsd management so that it neither crashes nor leaks if a thread's
273    only allocation activity is to call free() after TLS destructors have been
274    executed.  This behavior was observed when operating with GNU libc, and is
275    unlikely to be an issue with other libc implementations.  (@interwq)
276  - Mask signals during background thread creation.  This prevents signals from
277    being inadvertently delivered to background threads.  (@jasone,
278    @davidtgoldblatt, @interwq)
279  - Avoid inactivity checks within background threads, in order to prevent
280    recursive mutex acquisition.  (@interwq)
281  - Fix extent_grow_retained() to use the specified hooks when the
282    arena.<i>.extent_hooks mallctl is used to override the default hooks.
283    (@interwq)
284  - Add missing reentrancy support for custom extent hooks which allocate.
285    (@interwq)
286  - Post-fork(2), re-initialize the list of tcaches associated with each arena
287    to contain no tcaches except the forking thread's.  (@interwq)
288  - Add missing post-fork(2) mutex reinitialization for extent_grow_mtx.  This
289    fixes potential deadlocks after fork(2).  (@interwq)
290  - Enforce minimum autoconf version (currently 2.68), since 2.63 is known to
291    generate corrupt configure scripts.  (@jasone)
292  - Ensure that the configured page size (--with-lg-page) is no larger than the
293    configured huge page size (--with-lg-hugepage).  (@jasone)
294
295* 5.0.0 (June 13, 2017)
296
297  Unlike all previous jemalloc releases, this release does not use naturally
298  aligned "chunks" for virtual memory management, and instead uses page-aligned
299  "extents".  This change has few externally visible effects, but the internal
300  impacts are... extensive.  Many other internal changes combine to make this
301  the most cohesively designed version of jemalloc so far, with ample
302  opportunity for further enhancements.
303
304  Continuous integration is now an integral aspect of development thanks to the
305  efforts of @davidtgoldblatt, and the dev branch tends to remain reasonably
306  stable on the tested platforms (Linux, FreeBSD, macOS, and Windows).  As a
307  side effect the official release frequency may decrease over time.
308
309  New features:
310  - Implement optional per-CPU arena support; threads choose which arena to use
311    based on current CPU rather than on fixed thread-->arena associations.
312    (@interwq)
313  - Implement two-phase decay of unused dirty pages.  Pages transition from
314    dirty-->muzzy-->clean, where the first phase transition relies on
315    madvise(... MADV_FREE) semantics, and the second phase transition discards
316    pages such that they are replaced with demand-zeroed pages on next access.
317    (@jasone)
318  - Increase decay time resolution from seconds to milliseconds.  (@jasone)
319  - Implement opt-in per CPU background threads, and use them for asynchronous
320    decay-driven unused dirty page purging.  (@interwq)
321  - Add mutex profiling, which collects a variety of statistics useful for
322    diagnosing overhead/contention issues.  (@interwq)
323  - Add C++ new/delete operator bindings.  (@djwatson)
324  - Support manually created arena destruction, such that all data and metadata
325    are discarded.  Add MALLCTL_ARENAS_DESTROYED for accessing merged stats
326    associated with destroyed arenas.  (@jasone)
327  - Add MALLCTL_ARENAS_ALL as a fixed index for use in accessing
328    merged/destroyed arena statistics via mallctl.  (@jasone)
329  - Add opt.abort_conf to optionally abort if invalid configuration options are
330    detected during initialization.  (@interwq)
331  - Add opt.stats_print_opts, so that e.g. JSON output can be selected for the
332    stats dumped during exit if opt.stats_print is true.  (@jasone)
333  - Add --with-version=VERSION for use when embedding jemalloc into another
334    project's git repository.  (@jasone)
335  - Add --disable-thp to support cross compiling.  (@jasone)
336  - Add --with-lg-hugepage to support cross compiling.  (@jasone)
337  - Add mallctl interfaces (various authors):
338    + background_thread
339    + opt.abort_conf
340    + opt.retain
341    + opt.percpu_arena
342    + opt.background_thread
343    + opt.{dirty,muzzy}_decay_ms
344    + opt.stats_print_opts
345    + arena.<i>.initialized
346    + arena.<i>.destroy
347    + arena.<i>.{dirty,muzzy}_decay_ms
348    + arena.<i>.extent_hooks
349    + arenas.{dirty,muzzy}_decay_ms
350    + arenas.bin.<i>.slab_size
351    + arenas.nlextents
352    + arenas.lextent.<i>.size
353    + arenas.create
354    + stats.background_thread.{num_threads,num_runs,run_interval}
355    + stats.mutexes.{ctl,background_thread,prof,reset}.
356      {num_ops,num_spin_acq,num_wait,max_wait_time,total_wait_time,max_num_thds,
357      num_owner_switch}
358    + stats.arenas.<i>.{dirty,muzzy}_decay_ms
359    + stats.arenas.<i>.uptime
360    + stats.arenas.<i>.{pmuzzy,base,internal,resident}
361    + stats.arenas.<i>.{dirty,muzzy}_{npurge,nmadvise,purged}
362    + stats.arenas.<i>.bins.<j>.{nslabs,reslabs,curslabs}
363    + stats.arenas.<i>.bins.<j>.mutex.
364      {num_ops,num_spin_acq,num_wait,max_wait_time,total_wait_time,max_num_thds,
365      num_owner_switch}
366    + stats.arenas.<i>.lextents.<j>.{nmalloc,ndalloc,nrequests,curlextents}
367    + stats.arenas.i.mutexes.{large,extent_avail,extents_dirty,extents_muzzy,
368      extents_retained,decay_dirty,decay_muzzy,base,tcache_list}.
369      {num_ops,num_spin_acq,num_wait,max_wait_time,total_wait_time,max_num_thds,
370      num_owner_switch}
371
372  Portability improvements:
373  - Improve reentrant allocation support, such that deadlock is less likely if
374    e.g. a system library call in turn allocates memory.  (@davidtgoldblatt,
375    @interwq)
376  - Support static linking of jemalloc with glibc.  (@djwatson)
377
378  Optimizations and refactors:
379  - Organize virtual memory as "extents" of virtual memory pages, rather than as
380    naturally aligned "chunks", and store all metadata in arbitrarily distant
381    locations.  This reduces virtual memory external fragmentation, and will
382    interact better with huge pages (not yet explicitly supported).  (@jasone)
383  - Fold large and huge size classes together; only small and large size classes
384    remain.  (@jasone)
385  - Unify the allocation paths, and merge most fast-path branching decisions.
386    (@davidtgoldblatt, @interwq)
387  - Embed per thread automatic tcache into thread-specific data, which reduces
388    conditional branches and dereferences.  Also reorganize tcache to increase
389    fast-path data locality.  (@interwq)
390  - Rewrite atomics to closely model the C11 API, convert various
391    synchronization from mutex-based to atomic, and use the explicit memory
392    ordering control to resolve various hypothetical races without increasing
393    synchronization overhead.  (@davidtgoldblatt)
394  - Extensively optimize rtree via various methods:
395    + Add multiple layers of rtree lookup caching, since rtree lookups are now
396      part of fast-path deallocation.  (@interwq)
397    + Determine rtree layout at compile time.  (@jasone)
398    + Make the tree shallower for common configurations.  (@jasone)
399    + Embed the root node in the top-level rtree data structure, thus avoiding
400      one level of indirection.  (@jasone)
401    + Further specialize leaf elements as compared to internal node elements,
402      and directly embed extent metadata needed for fast-path deallocation.
403      (@jasone)
404    + Ignore leading always-zero address bits (architecture-specific).
405      (@jasone)
406  - Reorganize headers (ongoing work) to make them hermetic, and disentangle
407    various module dependencies.  (@davidtgoldblatt)
408  - Convert various internal data structures such as size class metadata from
409    boot-time-initialized to compile-time-initialized.  Propagate resulting data
410    structure simplifications, such as making arena metadata fixed-size.
411    (@jasone)
412  - Simplify size class lookups when constrained to size classes that are
413    multiples of the page size.  This speeds lookups, but the primary benefit is
414    complexity reduction in code that was the source of numerous regressions.
415    (@jasone)
416  - Lock individual extents when possible for localized extent operations,
417    rather than relying on a top-level arena lock.  (@davidtgoldblatt, @jasone)
418  - Use first fit layout policy instead of best fit, in order to improve
419    packing.  (@jasone)
420  - If munmap(2) is not in use, use an exponential series to grow each arena's
421    virtual memory, so that the number of disjoint virtual memory mappings
422    remains low.  (@jasone)
423  - Implement per arena base allocators, so that arenas never share any virtual
424    memory pages.  (@jasone)
425  - Automatically generate private symbol name mangling macros.  (@jasone)
426
427  Incompatible changes:
428  - Replace chunk hooks with an expanded/normalized set of extent hooks.
429    (@jasone)
430  - Remove ratio-based purging.  (@jasone)
431  - Remove --disable-tcache.  (@jasone)
432  - Remove --disable-tls.  (@jasone)
433  - Remove --enable-ivsalloc.  (@jasone)
434  - Remove --with-lg-size-class-group.  (@jasone)
435  - Remove --with-lg-tiny-min.  (@jasone)
436  - Remove --disable-cc-silence.  (@jasone)
437  - Remove --enable-code-coverage.  (@jasone)
438  - Remove --disable-munmap (replaced by opt.retain).  (@jasone)
439  - Remove Valgrind support.  (@jasone)
440  - Remove quarantine support.  (@jasone)
441  - Remove redzone support.  (@jasone)
442  - Remove mallctl interfaces (various authors):
443    + config.munmap
444    + config.tcache
445    + config.tls
446    + config.valgrind
447    + opt.lg_chunk
448    + opt.purge
449    + opt.lg_dirty_mult
450    + opt.decay_time
451    + opt.quarantine
452    + opt.redzone
453    + opt.thp
454    + arena.<i>.lg_dirty_mult
455    + arena.<i>.decay_time
456    + arena.<i>.chunk_hooks
457    + arenas.initialized
458    + arenas.lg_dirty_mult
459    + arenas.decay_time
460    + arenas.bin.<i>.run_size
461    + arenas.nlruns
462    + arenas.lrun.<i>.size
463    + arenas.nhchunks
464    + arenas.hchunk.<i>.size
465    + arenas.extend
466    + stats.cactive
467    + stats.arenas.<i>.lg_dirty_mult
468    + stats.arenas.<i>.decay_time
469    + stats.arenas.<i>.metadata.{mapped,allocated}
470    + stats.arenas.<i>.{npurge,nmadvise,purged}
471    + stats.arenas.<i>.huge.{allocated,nmalloc,ndalloc,nrequests}
472    + stats.arenas.<i>.bins.<j>.{nruns,reruns,curruns}
473    + stats.arenas.<i>.lruns.<j>.{nmalloc,ndalloc,nrequests,curruns}
474    + stats.arenas.<i>.hchunks.<j>.{nmalloc,ndalloc,nrequests,curhchunks}
475
476  Bug fixes:
477  - Improve interval-based profile dump triggering to dump only one profile when
478    a single allocation's size exceeds the interval.  (@jasone)
479  - Use prefixed function names (as controlled by --with-jemalloc-prefix) when
480    pruning backtrace frames in jeprof.  (@jasone)
481
482* 4.5.0 (February 28, 2017)
483
484  This is the first release to benefit from much broader continuous integration
485  testing, thanks to @davidtgoldblatt.  Had we had this testing infrastructure
486  in place for prior releases, it would have caught all of the most serious
487  regressions fixed by this release.
488
489  New features:
490  - Add --disable-thp and the opt.thp mallctl to provide opt-out mechanisms for
491    transparent huge page integration.  (@jasone)
492  - Update zone allocator integration to work with macOS 10.12.  (@glandium)
493  - Restructure *CFLAGS configuration, so that CFLAGS behaves typically, and
494    EXTRA_CFLAGS provides a way to specify e.g. -Werror during building, but not
495    during configuration.  (@jasone, @ronawho)
496
497  Bug fixes:
498  - Fix DSS (sbrk(2)-based) allocation.  This regression was first released in
499    4.3.0.  (@jasone)
500  - Handle race in per size class utilization computation.  This functionality
501    was first released in 4.0.0.  (@interwq)
502  - Fix lock order reversal during gdump.  (@jasone)
503  - Fix/refactor tcache synchronization.  This regression was first released in
504    4.0.0.  (@jasone)
505  - Fix various JSON-formatted malloc_stats_print() bugs.  This functionality
506    was first released in 4.3.0.  (@jasone)
507  - Fix huge-aligned allocation.  This regression was first released in 4.4.0.
508    (@jasone)
509  - When transparent huge page integration is enabled, detect what state pages
510    start in according to the kernel's current operating mode, and only convert
511    arena chunks to non-huge during purging if that is not their initial state.
512    This functionality was first released in 4.4.0.  (@jasone)
513  - Fix lg_chunk clamping for the --enable-cache-oblivious --disable-fill case.
514    This regression was first released in 4.0.0.  (@jasone, @428desmo)
515  - Properly detect sparc64 when building for Linux.  (@glaubitz)
516
517* 4.4.0 (December 3, 2016)
518
519  New features:
520  - Add configure support for *-*-linux-android.  (@cferris1000, @jasone)
521  - Add the --disable-syscall configure option, for use on systems that place
522    security-motivated limitations on syscall(2).  (@jasone)
523  - Add support for Debian GNU/kFreeBSD.  (@thesam)
524
525  Optimizations:
526  - Add extent serial numbers and use them where appropriate as a sort key that
527    is higher priority than address, so that the allocation policy prefers older
528    extents.  This tends to improve locality (decrease fragmentation) when
529    memory grows downward.  (@jasone)
530  - Refactor madvise(2) configuration so that MADV_FREE is detected and utilized
531    on Linux 4.5 and newer.  (@jasone)
532  - Mark partially purged arena chunks as non-huge-page.  This improves
533    interaction with Linux's transparent huge page functionality.  (@jasone)
534
535  Bug fixes:
536  - Fix size class computations for edge conditions involving extremely large
537    allocations.  This regression was first released in 4.0.0.  (@jasone,
538    @ingvarha)
539  - Remove overly restrictive assertions related to the cactive statistic.  This
540    regression was first released in 4.1.0.  (@jasone)
541  - Implement a more reliable detection scheme for os_unfair_lock on macOS.
542    (@jszakmeister)
543
544* 4.3.1 (November 7, 2016)
545
546  Bug fixes:
547  - Fix a severe virtual memory leak.  This regression was first released in
548    4.3.0.  (@interwq, @jasone)
549  - Refactor atomic and prng APIs to restore support for 32-bit platforms that
550    use pre-C11 toolchains, e.g. FreeBSD's mips.  (@jasone)
551
552* 4.3.0 (November 4, 2016)
553
554  This is the first release that passes the test suite for multiple Windows
555  configurations, thanks in large part to @glandium setting up continuous
556  integration via AppVeyor (and Travis CI for Linux and OS X).
557
558  New features:
559  - Add "J" (JSON) support to malloc_stats_print().  (@jasone)
560  - Add Cray compiler support.  (@ronawho)
561
562  Optimizations:
563  - Add/use adaptive spinning for bootstrapping and radix tree node
564    initialization.  (@jasone)
565
566  Bug fixes:
567  - Fix large allocation to search starting in the optimal size class heap,
568    which can substantially reduce virtual memory churn and fragmentation.  This
569    regression was first released in 4.0.0.  (@mjp41, @jasone)
570  - Fix stats.arenas.<i>.nthreads accounting.  (@interwq)
571  - Fix and simplify decay-based purging.  (@jasone)
572  - Make DSS (sbrk(2)-related) operations lockless, which resolves potential
573    deadlocks during thread exit.  (@jasone)
574  - Fix over-sized allocation of radix tree leaf nodes.  (@mjp41, @ogaun,
575    @jasone)
576  - Fix over-sized allocation of arena_t (plus associated stats) data
577    structures.  (@jasone, @interwq)
578  - Fix EXTRA_CFLAGS to not affect configuration.  (@jasone)
579  - Fix a Valgrind integration bug.  (@ronawho)
580  - Disallow 0x5a junk filling when running in Valgrind.  (@jasone)
581  - Fix a file descriptor leak on Linux.  This regression was first released in
582    4.2.0.  (@vsarunas, @jasone)
583  - Fix static linking of jemalloc with glibc.  (@djwatson)
584  - Use syscall(2) rather than {open,read,close}(2) during boot on Linux.  This
585    works around other libraries' system call wrappers performing reentrant
586    allocation.  (@kspinka, @Whissi, @jasone)
587  - Fix OS X default zone replacement to work with OS X 10.12.  (@glandium,
588    @jasone)
589  - Fix cached memory management to avoid needless commit/decommit operations
590    during purging, which resolves permanent virtual memory map fragmentation
591    issues on Windows.  (@mjp41, @jasone)
592  - Fix TSD fetches to avoid (recursive) allocation.  This is relevant to
593    non-TLS and Windows configurations.  (@jasone)
594  - Fix malloc_conf overriding to work on Windows.  (@jasone)
595  - Forcibly disable lazy-lock on Windows (was forcibly *enabled*).  (@jasone)
596
597* 4.2.1 (June 8, 2016)
598
599  Bug fixes:
600  - Fix bootstrapping issues for configurations that require allocation during
601    tsd initialization (e.g. --disable-tls).  (@cferris1000, @jasone)
602  - Fix gettimeofday() version of nstime_update().  (@ronawho)
603  - Fix Valgrind regressions in calloc() and chunk_alloc_wrapper().  (@ronawho)
604  - Fix potential VM map fragmentation regression.  (@jasone)
605  - Fix opt_zero-triggered in-place huge reallocation zeroing.  (@jasone)
606  - Fix heap profiling context leaks in reallocation edge cases.  (@jasone)
607
608* 4.2.0 (May 12, 2016)
609
610  New features:
611  - Add the arena.<i>.reset mallctl, which makes it possible to discard all of
612    an arena's allocations in a single operation.  (@jasone)
613  - Add the stats.retained and stats.arenas.<i>.retained statistics.  (@jasone)
614  - Add the --with-version configure option.  (@jasone)
615  - Support --with-lg-page values larger than actual page size.  (@jasone)
616
617  Optimizations:
618  - Use pairing heaps rather than red-black trees for various hot data
619    structures.  (@djwatson, @jasone)
620  - Streamline fast paths of rtree operations.  (@jasone)
621  - Optimize the fast paths of calloc() and [m,d,sd]allocx().  (@jasone)
622  - Decommit unused virtual memory if the OS does not overcommit.  (@jasone)
623  - Specify MAP_NORESERVE on Linux if [heuristic] overcommit is active, in order
624    to avoid unfortunate interactions during fork(2).  (@jasone)
625
626  Bug fixes:
627  - Fix chunk accounting related to triggering gdump profiles.  (@jasone)
628  - Link against librt for clock_gettime(2) if glibc < 2.17.  (@jasone)
629  - Scale leak report summary according to sampling probability.  (@jasone)
630
631* 4.1.1 (May 3, 2016)
632
633  This bugfix release resolves a variety of mostly minor issues, though the
634  bitmap fix is critical for 64-bit Windows.
635
636  Bug fixes:
637  - Fix the linear scan version of bitmap_sfu() to shift by the proper amount
638    even when sizeof(long) is not the same as sizeof(void *), as on 64-bit
639    Windows.  (@jasone)
640  - Fix hashing functions to avoid unaligned memory accesses (and resulting
641    crashes).  This is relevant at least to some ARM-based platforms.
642    (@rkmisra)
643  - Fix fork()-related lock rank ordering reversals.  These reversals were
644    unlikely to cause deadlocks in practice except when heap profiling was
645    enabled and active.  (@jasone)
646  - Fix various chunk leaks in OOM code paths.  (@jasone)
647  - Fix malloc_stats_print() to print opt.narenas correctly.  (@jasone)
648  - Fix MSVC-specific build/test issues.  (@rustyx, @yuslepukhin)
649  - Fix a variety of test failures that were due to test fragility rather than
650    core bugs.  (@jasone)
651
652* 4.1.0 (February 28, 2016)
653
654  This release is primarily about optimizations, but it also incorporates a lot
655  of portability-motivated refactoring and enhancements.  Many people worked on
656  this release, to an extent that even with the omission here of minor changes
657  (see git revision history), and of the people who reported and diagnosed
658  issues, so much of the work was contributed that starting with this release,
659  changes are annotated with author credits to help reflect the collaborative
660  effort involved.
661
662  New features:
663  - Implement decay-based unused dirty page purging, a major optimization with
664    mallctl API impact.  This is an alternative to the existing ratio-based
665    unused dirty page purging, and is intended to eventually become the sole
666    purging mechanism.  New mallctls:
667    + opt.purge
668    + opt.decay_time
669    + arena.<i>.decay
670    + arena.<i>.decay_time
671    + arenas.decay_time
672    + stats.arenas.<i>.decay_time
673    (@jasone, @cevans87)
674  - Add --with-malloc-conf, which makes it possible to embed a default
675    options string during configuration.  This was motivated by the desire to
676    specify --with-malloc-conf=purge:decay , since the default must remain
677    purge:ratio until the 5.0.0 release.  (@jasone)
678  - Add MS Visual Studio 2015 support.  (@rustyx, @yuslepukhin)
679  - Make *allocx() size class overflow behavior defined.  The maximum
680    size class is now less than PTRDIFF_MAX to protect applications against
681    numerical overflow, and all allocation functions are guaranteed to indicate
682    errors rather than potentially crashing if the request size exceeds the
683    maximum size class.  (@jasone)
684  - jeprof:
685    + Add raw heap profile support.  (@jasone)
686    + Add --retain and --exclude for backtrace symbol filtering.  (@jasone)
687
688  Optimizations:
689  - Optimize the fast path to combine various bootstrapping and configuration
690    checks and execute more streamlined code in the common case.  (@interwq)
691  - Use linear scan for small bitmaps (used for small object tracking).  In
692    addition to speeding up bitmap operations on 64-bit systems, this reduces
693    allocator metadata overhead by approximately 0.2%.  (@djwatson)
694  - Separate arena_avail trees, which substantially speeds up run tree
695    operations.  (@djwatson)
696  - Use memoization (boot-time-computed table) for run quantization.  Separate
697    arena_avail trees reduced the importance of this optimization.  (@jasone)
698  - Attempt mmap-based in-place huge reallocation.  This can dramatically speed
699    up incremental huge reallocation.  (@jasone)
700
701  Incompatible changes:
702  - Make opt.narenas unsigned rather than size_t.  (@jasone)
703
704  Bug fixes:
705  - Fix stats.cactive accounting regression.  (@rustyx, @jasone)
706  - Handle unaligned keys in hash().  This caused problems for some ARM systems.
707    (@jasone, @cferris1000)
708  - Refactor arenas array.  In addition to fixing a fork-related deadlock, this
709    makes arena lookups faster and simpler.  (@jasone)
710  - Move retained memory allocation out of the default chunk allocation
711    function, to a location that gets executed even if the application installs
712    a custom chunk allocation function.  This resolves a virtual memory leak.
713    (@buchgr)
714  - Fix a potential tsd cleanup leak.  (@cferris1000, @jasone)
715  - Fix run quantization.  In practice this bug had no impact unless
716    applications requested memory with alignment exceeding one page.
717    (@jasone, @djwatson)
718  - Fix LinuxThreads-specific bootstrapping deadlock.  (Cosmin Paraschiv)
719  - jeprof:
720    + Don't discard curl options if timeout is not defined.  (@djwatson)
721    + Detect failed profile fetches.  (@djwatson)
722  - Fix stats.arenas.<i>.{dss,lg_dirty_mult,decay_time,pactive,pdirty} for
723    --disable-stats case.  (@jasone)
724
725* 4.0.4 (October 24, 2015)
726
727  This bugfix release fixes another xallocx() regression.  No other regressions
728  have come to light in over a month, so this is likely a good starting point
729  for people who prefer to wait for "dot one" releases with all the major issues
730  shaken out.
731
732  Bug fixes:
733  - Fix xallocx(..., MALLOCX_ZERO to zero the last full trailing page of large
734    allocations that have been randomly assigned an offset of 0 when
735    --enable-cache-oblivious configure option is enabled.
736
737* 4.0.3 (September 24, 2015)
738
739  This bugfix release continues the trend of xallocx() and heap profiling fixes.
740
741  Bug fixes:
742  - Fix xallocx(..., MALLOCX_ZERO) to zero all trailing bytes of large
743    allocations when --enable-cache-oblivious configure option is enabled.
744  - Fix xallocx(..., MALLOCX_ZERO) to zero trailing bytes of huge allocations
745    when resizing from/to a size class that is not a multiple of the chunk size.
746  - Fix prof_tctx_dump_iter() to filter out nodes that were created after heap
747    profile dumping started.
748  - Work around a potentially bad thread-specific data initialization
749    interaction with NPTL (glibc's pthreads implementation).
750
751* 4.0.2 (September 21, 2015)
752
753  This bugfix release addresses a few bugs specific to heap profiling.
754
755  Bug fixes:
756  - Fix ixallocx_prof_sample() to never modify nor create sampled small
757    allocations.  xallocx() is in general incapable of moving small allocations,
758    so this fix removes buggy code without loss of generality.
759  - Fix irallocx_prof_sample() to always allocate large regions, even when
760    alignment is non-zero.
761  - Fix prof_alloc_rollback() to read tdata from thread-specific data rather
762    than dereferencing a potentially invalid tctx.
763
764* 4.0.1 (September 15, 2015)
765
766  This is a bugfix release that is somewhat high risk due to the amount of
767  refactoring required to address deep xallocx() problems.  As a side effect of
768  these fixes, xallocx() now tries harder to partially fulfill requests for
769  optional extra space.  Note that a couple of minor heap profiling
770  optimizations are included, but these are better thought of as performance
771  fixes that were integral to discovering most of the other bugs.
772
773  Optimizations:
774  - Avoid a chunk metadata read in arena_prof_tctx_set(), since it is in the
775    fast path when heap profiling is enabled.  Additionally, split a special
776    case out into arena_prof_tctx_reset(), which also avoids chunk metadata
777    reads.
778  - Optimize irallocx_prof() to optimistically update the sampler state.  The
779    prior implementation appears to have been a holdover from when
780    rallocx()/xallocx() functionality was combined as rallocm().
781
782  Bug fixes:
783  - Fix TLS configuration such that it is enabled by default for platforms on
784    which it works correctly.
785  - Fix arenas_cache_cleanup() and arena_get_hard() to handle
786    allocation/deallocation within the application's thread-specific data
787    cleanup functions even after arenas_cache is torn down.
788  - Fix xallocx() bugs related to size+extra exceeding HUGE_MAXCLASS.
789  - Fix chunk purge hook calls for in-place huge shrinking reallocation to
790    specify the old chunk size rather than the new chunk size.  This bug caused
791    no correctness issues for the default chunk purge function, but was
792    visible to custom functions set via the "arena.<i>.chunk_hooks" mallctl.
793  - Fix heap profiling bugs:
794    + Fix heap profiling to distinguish among otherwise identical sample sites
795      with interposed resets (triggered via the "prof.reset" mallctl).  This bug
796      could cause data structure corruption that would most likely result in a
797      segfault.
798    + Fix irealloc_prof() to prof_alloc_rollback() on OOM.
799    + Make one call to prof_active_get_unlocked() per allocation event, and use
800      the result throughout the relevant functions that handle an allocation
801      event.  Also add a missing check in prof_realloc().  These fixes protect
802      allocation events against concurrent prof_active changes.
803    + Fix ixallocx_prof() to pass usize_max and zero to ixallocx_prof_sample()
804      in the correct order.
805    + Fix prof_realloc() to call prof_free_sampled_object() after calling
806      prof_malloc_sample_object().  Prior to this fix, if tctx and old_tctx were
807      the same, the tctx could have been prematurely destroyed.
808  - Fix portability bugs:
809    + Don't bitshift by negative amounts when encoding/decoding run sizes in
810      chunk header maps.  This affected systems with page sizes greater than 8
811      KiB.
812    + Rename index_t to szind_t to avoid an existing type on Solaris.
813    + Add JEMALLOC_CXX_THROW to the memalign() function prototype, in order to
814      match glibc and avoid compilation errors when including both
815      jemalloc/jemalloc.h and malloc.h in C++ code.
816    + Don't assume that /bin/sh is appropriate when running size_classes.sh
817      during configuration.
818    + Consider __sparcv9 a synonym for __sparc64__ when defining LG_QUANTUM.
819    + Link tests to librt if it contains clock_gettime(2).
820
821* 4.0.0 (August 17, 2015)
822
823  This version contains many speed and space optimizations, both minor and
824  major.  The major themes are generalization, unification, and simplification.
825  Although many of these optimizations cause no visible behavior change, their
826  cumulative effect is substantial.
827
828  New features:
829  - Normalize size class spacing to be consistent across the complete size
830    range.  By default there are four size classes per size doubling, but this
831    is now configurable via the --with-lg-size-class-group option.  Also add the
832    --with-lg-page, --with-lg-page-sizes, --with-lg-quantum, and
833    --with-lg-tiny-min options, which can be used to tweak page and size class
834    settings.  Impacts:
835    + Worst case performance for incrementally growing/shrinking reallocation
836      is improved because there are far fewer size classes, and therefore
837      copying happens less often.
838    + Internal fragmentation is limited to 20% for all but the smallest size
839      classes (those less than four times the quantum).  (1B + 4 KiB)
840      and (1B + 4 MiB) previously suffered nearly 50% internal fragmentation.
841    + Chunk fragmentation tends to be lower because there are fewer distinct run
842      sizes to pack.
843  - Add support for explicit tcaches.  The "tcache.create", "tcache.flush", and
844    "tcache.destroy" mallctls control tcache lifetime and flushing, and the
845    MALLOCX_TCACHE(tc) and MALLOCX_TCACHE_NONE flags to the *allocx() API
846    control which tcache is used for each operation.
847  - Implement per thread heap profiling, as well as the ability to
848    enable/disable heap profiling on a per thread basis.  Add the "prof.reset",
849    "prof.lg_sample", "thread.prof.name", "thread.prof.active",
850    "opt.prof_thread_active_init", "prof.thread_active_init", and
851    "thread.prof.active" mallctls.
852  - Add support for per arena application-specified chunk allocators, configured
853    via the "arena.<i>.chunk_hooks" mallctl.
854  - Refactor huge allocation to be managed by arenas, so that arenas now
855    function as general purpose independent allocators.  This is important in
856    the context of user-specified chunk allocators, aside from the scalability
857    benefits.  Related new statistics:
858    + The "stats.arenas.<i>.huge.allocated", "stats.arenas.<i>.huge.nmalloc",
859      "stats.arenas.<i>.huge.ndalloc", and "stats.arenas.<i>.huge.nrequests"
860      mallctls provide high level per arena huge allocation statistics.
861    + The "arenas.nhchunks", "arenas.hchunk.<i>.size",
862      "stats.arenas.<i>.hchunks.<j>.nmalloc",
863      "stats.arenas.<i>.hchunks.<j>.ndalloc",
864      "stats.arenas.<i>.hchunks.<j>.nrequests", and
865      "stats.arenas.<i>.hchunks.<j>.curhchunks" mallctls provide per size class
866      statistics.
867  - Add the 'util' column to malloc_stats_print() output, which reports the
868    proportion of available regions that are currently in use for each small
869    size class.
870  - Add "alloc" and "free" modes for for junk filling (see the "opt.junk"
871    mallctl), so that it is possible to separately enable junk filling for
872    allocation versus deallocation.
873  - Add the jemalloc-config script, which provides information about how
874    jemalloc was configured, and how to integrate it into application builds.
875  - Add metadata statistics, which are accessible via the "stats.metadata",
876    "stats.arenas.<i>.metadata.mapped", and
877    "stats.arenas.<i>.metadata.allocated" mallctls.
878  - Add the "stats.resident" mallctl, which reports the upper limit of
879    physically resident memory mapped by the allocator.
880  - Add per arena control over unused dirty page purging, via the
881    "arenas.lg_dirty_mult", "arena.<i>.lg_dirty_mult", and
882    "stats.arenas.<i>.lg_dirty_mult" mallctls.
883  - Add the "prof.gdump" mallctl, which makes it possible to toggle the gdump
884    feature on/off during program execution.
885  - Add sdallocx(), which implements sized deallocation.  The primary
886    optimization over dallocx() is the removal of a metadata read, which often
887    suffers an L1 cache miss.
888  - Add missing header includes in jemalloc/jemalloc.h, so that applications
889    only have to #include <jemalloc/jemalloc.h>.
890  - Add support for additional platforms:
891    + Bitrig
892    + Cygwin
893    + DragonFlyBSD
894    + iOS
895    + OpenBSD
896    + OpenRISC/or1k
897
898  Optimizations:
899  - Maintain dirty runs in per arena LRUs rather than in per arena trees of
900    dirty-run-containing chunks.  In practice this change significantly reduces
901    dirty page purging volume.
902  - Integrate whole chunks into the unused dirty page purging machinery.  This
903    reduces the cost of repeated huge allocation/deallocation, because it
904    effectively introduces a cache of chunks.
905  - Split the arena chunk map into two separate arrays, in order to increase
906    cache locality for the frequently accessed bits.
907  - Move small run metadata out of runs, into arena chunk headers.  This reduces
908    run fragmentation, smaller runs reduce external fragmentation for small size
909    classes, and packed (less uniformly aligned) metadata layout improves CPU
910    cache set distribution.
911  - Randomly distribute large allocation base pointer alignment relative to page
912    boundaries in order to more uniformly utilize CPU cache sets.  This can be
913    disabled via the --disable-cache-oblivious configure option, and queried via
914    the "config.cache_oblivious" mallctl.
915  - Micro-optimize the fast paths for the public API functions.
916  - Refactor thread-specific data to reside in a single structure.  This assures
917    that only a single TLS read is necessary per call into the public API.
918  - Implement in-place huge allocation growing and shrinking.
919  - Refactor rtree (radix tree for chunk lookups) to be lock-free, and make
920    additional optimizations that reduce maximum lookup depth to one or two
921    levels.  This resolves what was a concurrency bottleneck for per arena huge
922    allocation, because a global data structure is critical for determining
923    which arenas own which huge allocations.
924
925  Incompatible changes:
926  - Replace --enable-cc-silence with --disable-cc-silence to suppress spurious
927    warnings by default.
928  - Assure that the constness of malloc_usable_size()'s return type matches that
929    of the system implementation.
930  - Change the heap profile dump format to support per thread heap profiling,
931    rename pprof to jeprof, and enhance it with the --thread=<n> option.  As a
932    result, the bundled jeprof must now be used rather than the upstream
933    (gperftools) pprof.
934  - Disable "opt.prof_final" by default, in order to avoid atexit(3), which can
935    internally deadlock on some platforms.
936  - Change the "arenas.nlruns" mallctl type from size_t to unsigned.
937  - Replace the "stats.arenas.<i>.bins.<j>.allocated" mallctl with
938    "stats.arenas.<i>.bins.<j>.curregs".
939  - Ignore MALLOC_CONF in set{uid,gid,cap} binaries.
940  - Ignore MALLOCX_ARENA(a) in dallocx(), in favor of using the
941    MALLOCX_TCACHE(tc) and MALLOCX_TCACHE_NONE flags to control tcache usage.
942
943  Removed features:
944  - Remove the *allocm() API, which is superseded by the *allocx() API.
945  - Remove the --enable-dss options, and make dss non-optional on all platforms
946    which support sbrk(2).
947  - Remove the "arenas.purge" mallctl, which was obsoleted by the
948    "arena.<i>.purge" mallctl in 3.1.0.
949  - Remove the unnecessary "opt.valgrind" mallctl; jemalloc automatically
950    detects whether it is running inside Valgrind.
951  - Remove the "stats.huge.allocated", "stats.huge.nmalloc", and
952    "stats.huge.ndalloc" mallctls.
953  - Remove the --enable-mremap option.
954  - Remove the "stats.chunks.current", "stats.chunks.total", and
955    "stats.chunks.high" mallctls.
956
957  Bug fixes:
958  - Fix the cactive statistic to decrease (rather than increase) when active
959    memory decreases.  This regression was first released in 3.5.0.
960  - Fix OOM handling in memalign() and valloc().  A variant of this bug existed
961    in all releases since 2.0.0, which introduced these functions.
962  - Fix an OOM-related regression in arena_tcache_fill_small(), which could
963    cause cache corruption on OOM.  This regression was present in all releases
964    from 2.2.0 through 3.6.0.
965  - Fix size class overflow handling for malloc(), posix_memalign(), memalign(),
966    calloc(), and realloc() when profiling is enabled.
967  - Fix the "arena.<i>.dss" mallctl to return an error if "primary" or
968    "secondary" precedence is specified, but sbrk(2) is not supported.
969  - Fix fallback lg_floor() implementations to handle extremely large inputs.
970  - Ensure the default purgeable zone is after the default zone on OS X.
971  - Fix latent bugs in atomic_*().
972  - Fix the "arena.<i>.dss" mallctl to handle read-only calls.
973  - Fix tls_model configuration to enable the initial-exec model when possible.
974  - Mark malloc_conf as a weak symbol so that the application can override it.
975  - Correctly detect glibc's adaptive pthread mutexes.
976  - Fix the --without-export configure option.
977
978* 3.6.0 (March 31, 2014)
979
980  This version contains a critical bug fix for a regression present in 3.5.0 and
981  3.5.1.
982
983  Bug fixes:
984  - Fix a regression in arena_chunk_alloc() that caused crashes during
985    small/large allocation if chunk allocation failed.  In the absence of this
986    bug, chunk allocation failure would result in allocation failure, e.g.  NULL
987    return from malloc().  This regression was introduced in 3.5.0.
988  - Fix backtracing for gcc intrinsics-based backtracing by specifying
989    -fno-omit-frame-pointer to gcc.  Note that the application (and all the
990    libraries it links to) must also be compiled with this option for
991    backtracing to be reliable.
992  - Use dss allocation precedence for huge allocations as well as small/large
993    allocations.
994  - Fix test assertion failure message formatting.  This bug did not manifest on
995    x86_64 systems because of implementation subtleties in va_list.
996  - Fix inconsequential test failures for hash and SFMT code.
997
998  New features:
999  - Support heap profiling on FreeBSD.  This feature depends on the proc
1000    filesystem being mounted during heap profile dumping.
1001
1002* 3.5.1 (February 25, 2014)
1003
1004  This version primarily addresses minor bugs in test code.
1005
1006  Bug fixes:
1007  - Configure Solaris/Illumos to use MADV_FREE.
1008  - Fix junk filling for mremap(2)-based huge reallocation.  This is only
1009    relevant if configuring with the --enable-mremap option specified.
1010  - Avoid compilation failure if 'restrict' C99 keyword is not supported by the
1011    compiler.
1012  - Add a configure test for SSE2 rather than assuming it is usable on i686
1013    systems.  This fixes test compilation errors, especially on 32-bit Linux
1014    systems.
1015  - Fix mallctl argument size mismatches (size_t vs. uint64_t) in the stats unit
1016    test.
1017  - Fix/remove flawed alignment-related overflow tests.
1018  - Prevent compiler optimizations that could change backtraces in the
1019    prof_accum unit test.
1020
1021* 3.5.0 (January 22, 2014)
1022
1023  This version focuses on refactoring and automated testing, though it also
1024  includes some non-trivial heap profiling optimizations not mentioned below.
1025
1026  New features:
1027  - Add the *allocx() API, which is a successor to the experimental *allocm()
1028    API.  The *allocx() functions are slightly simpler to use because they have
1029    fewer parameters, they directly return the results of primary interest, and
1030    mallocx()/rallocx() avoid the strict aliasing pitfall that
1031    allocm()/rallocm() share with posix_memalign().  Note that *allocm() is
1032    slated for removal in the next non-bugfix release.
1033  - Add support for LinuxThreads.
1034
1035  Bug fixes:
1036  - Unless heap profiling is enabled, disable floating point code and don't link
1037    with libm.  This, in combination with e.g. EXTRA_CFLAGS=-mno-sse on x64
1038    systems, makes it possible to completely disable floating point register
1039    use.  Some versions of glibc neglect to save/restore caller-saved floating
1040    point registers during dynamic lazy symbol loading, and the symbol loading
1041    code uses whatever malloc the application happens to have linked/loaded
1042    with, the result being potential floating point register corruption.
1043  - Report ENOMEM rather than EINVAL if an OOM occurs during heap profiling
1044    backtrace creation in imemalign().  This bug impacted posix_memalign() and
1045    aligned_alloc().
1046  - Fix a file descriptor leak in a prof_dump_maps() error path.
1047  - Fix prof_dump() to close the dump file descriptor for all relevant error
1048    paths.
1049  - Fix rallocm() to use the arena specified by the ALLOCM_ARENA(s) flag for
1050    allocation, not just deallocation.
1051  - Fix a data race for large allocation stats counters.
1052  - Fix a potential infinite loop during thread exit.  This bug occurred on
1053    Solaris, and could affect other platforms with similar pthreads TSD
1054    implementations.
1055  - Don't junk-fill reallocations unless usable size changes.  This fixes a
1056    violation of the *allocx()/*allocm() semantics.
1057  - Fix growing large reallocation to junk fill new space.
1058  - Fix huge deallocation to junk fill when munmap is disabled.
1059  - Change the default private namespace prefix from empty to je_, and change
1060    --with-private-namespace-prefix so that it prepends an additional prefix
1061    rather than replacing je_.  This reduces the likelihood of applications
1062    which statically link jemalloc experiencing symbol name collisions.
1063  - Add missing private namespace mangling (relevant when
1064    --with-private-namespace is specified).
1065  - Add and use JEMALLOC_INLINE_C so that static inline functions are marked as
1066    static even for debug builds.
1067  - Add a missing mutex unlock in a malloc_init_hard() error path.  In practice
1068    this error path is never executed.
1069  - Fix numerous bugs in malloc_strotumax() error handling/reporting.  These
1070    bugs had no impact except for malformed inputs.
1071  - Fix numerous bugs in malloc_snprintf().  These bugs were not exercised by
1072    existing calls, so they had no impact.
1073
1074* 3.4.1 (October 20, 2013)
1075
1076  Bug fixes:
1077  - Fix a race in the "arenas.extend" mallctl that could cause memory corruption
1078    of internal data structures and subsequent crashes.
1079  - Fix Valgrind integration flaws that caused Valgrind warnings about reads of
1080    uninitialized memory in:
1081    + arena chunk headers
1082    + internal zero-initialized data structures (relevant to tcache and prof
1083      code)
1084  - Preserve errno during the first allocation.  A readlink(2) call during
1085    initialization fails unless /etc/malloc.conf exists, so errno was typically
1086    set during the first allocation prior to this fix.
1087  - Fix compilation warnings reported by gcc 4.8.1.
1088
1089* 3.4.0 (June 2, 2013)
1090
1091  This version is essentially a small bugfix release, but the addition of
1092  aarch64 support requires that the minor version be incremented.
1093
1094  Bug fixes:
1095  - Fix race-triggered deadlocks in chunk_record().  These deadlocks were
1096    typically triggered by multiple threads concurrently deallocating huge
1097    objects.
1098
1099  New features:
1100  - Add support for the aarch64 architecture.
1101
1102* 3.3.1 (March 6, 2013)
1103
1104  This version fixes bugs that are typically encountered only when utilizing
1105  custom run-time options.
1106
1107  Bug fixes:
1108  - Fix a locking order bug that could cause deadlock during fork if heap
1109    profiling were enabled.
1110  - Fix a chunk recycling bug that could cause the allocator to lose track of
1111    whether a chunk was zeroed.  On FreeBSD, NetBSD, and OS X, it could cause
1112    corruption if allocating via sbrk(2) (unlikely unless running with the
1113    "dss:primary" option specified).  This was completely harmless on Linux
1114    unless using mlockall(2) (and unlikely even then, unless the
1115    --disable-munmap configure option or the "dss:primary" option was
1116    specified).  This regression was introduced in 3.1.0 by the
1117    mlockall(2)/madvise(2) interaction fix.
1118  - Fix TLS-related memory corruption that could occur during thread exit if the
1119    thread never allocated memory.  Only the quarantine and prof facilities were
1120    susceptible.
1121  - Fix two quarantine bugs:
1122    + Internal reallocation of the quarantined object array leaked the old
1123      array.
1124    + Reallocation failure for internal reallocation of the quarantined object
1125      array (very unlikely) resulted in memory corruption.
1126  - Fix Valgrind integration to annotate all internally allocated memory in a
1127    way that keeps Valgrind happy about internal data structure access.
1128  - Fix building for s390 systems.
1129
1130* 3.3.0 (January 23, 2013)
1131
1132  This version includes a few minor performance improvements in addition to the
1133  listed new features and bug fixes.
1134
1135  New features:
1136  - Add clipping support to lg_chunk option processing.
1137  - Add the --enable-ivsalloc option.
1138  - Add the --without-export option.
1139  - Add the --disable-zone-allocator option.
1140
1141  Bug fixes:
1142  - Fix "arenas.extend" mallctl to output the number of arenas.
1143  - Fix chunk_recycle() to unconditionally inform Valgrind that returned memory
1144    is undefined.
1145  - Fix build break on FreeBSD related to alloca.h.
1146
1147* 3.2.0 (November 9, 2012)
1148
1149  In addition to a couple of bug fixes, this version modifies page run
1150  allocation and dirty page purging algorithms in order to better control
1151  page-level virtual memory fragmentation.
1152
1153  Incompatible changes:
1154  - Change the "opt.lg_dirty_mult" default from 5 to 3 (32:1 to 8:1).
1155
1156  Bug fixes:
1157  - Fix dss/mmap allocation precedence code to use recyclable mmap memory only
1158    after primary dss allocation fails.
1159  - Fix deadlock in the "arenas.purge" mallctl.  This regression was introduced
1160    in 3.1.0 by the addition of the "arena.<i>.purge" mallctl.
1161
1162* 3.1.0 (October 16, 2012)
1163
1164  New features:
1165  - Auto-detect whether running inside Valgrind, thus removing the need to
1166    manually specify MALLOC_CONF=valgrind:true.
1167  - Add the "arenas.extend" mallctl, which allows applications to create
1168    manually managed arenas.
1169  - Add the ALLOCM_ARENA() flag for {,r,d}allocm().
1170  - Add the "opt.dss", "arena.<i>.dss", and "stats.arenas.<i>.dss" mallctls,
1171    which provide control over dss/mmap precedence.
1172  - Add the "arena.<i>.purge" mallctl, which obsoletes "arenas.purge".
1173  - Define LG_QUANTUM for hppa.
1174
1175  Incompatible changes:
1176  - Disable tcache by default if running inside Valgrind, in order to avoid
1177    making unallocated objects appear reachable to Valgrind.
1178  - Drop const from malloc_usable_size() argument on Linux.
1179
1180  Bug fixes:
1181  - Fix heap profiling crash if sampled object is freed via realloc(p, 0).
1182  - Remove const from __*_hook variable declarations, so that glibc can modify
1183    them during process forking.
1184  - Fix mlockall(2)/madvise(2) interaction.
1185  - Fix fork(2)-related deadlocks.
1186  - Fix error return value for "thread.tcache.enabled" mallctl.
1187
1188* 3.0.0 (May 11, 2012)
1189
1190  Although this version adds some major new features, the primary focus is on
1191  internal code cleanup that facilitates maintainability and portability, most
1192  of which is not reflected in the ChangeLog.  This is the first release to
1193  incorporate substantial contributions from numerous other developers, and the
1194  result is a more broadly useful allocator (see the git revision history for
1195  contribution details).  Note that the license has been unified, thanks to
1196  Facebook granting a license under the same terms as the other copyright
1197  holders (see COPYING).
1198
1199  New features:
1200  - Implement Valgrind support, redzones, and quarantine.
1201  - Add support for additional platforms:
1202    + FreeBSD
1203    + Mac OS X Lion
1204    + MinGW
1205    + Windows (no support yet for replacing the system malloc)
1206  - Add support for additional architectures:
1207    + MIPS
1208    + SH4
1209    + Tilera
1210  - Add support for cross compiling.
1211  - Add nallocm(), which rounds a request size up to the nearest size class
1212    without actually allocating.
1213  - Implement aligned_alloc() (blame C11).
1214  - Add the "thread.tcache.enabled" mallctl.
1215  - Add the "opt.prof_final" mallctl.
1216  - Update pprof (from gperftools 2.0).
1217  - Add the --with-mangling option.
1218  - Add the --disable-experimental option.
1219  - Add the --disable-munmap option, and make it the default on Linux.
1220  - Add the --enable-mremap option, which disables use of mremap(2) by default.
1221
1222  Incompatible changes:
1223  - Enable stats by default.
1224  - Enable fill by default.
1225  - Disable lazy locking by default.
1226  - Rename the "tcache.flush" mallctl to "thread.tcache.flush".
1227  - Rename the "arenas.pagesize" mallctl to "arenas.page".
1228  - Change the "opt.lg_prof_sample" default from 0 to 19 (1 B to 512 KiB).
1229  - Change the "opt.prof_accum" default from true to false.
1230
1231  Removed features:
1232  - Remove the swap feature, including the "config.swap", "swap.avail",
1233    "swap.prezeroed", "swap.nfds", and "swap.fds" mallctls.
1234  - Remove highruns statistics, including the
1235    "stats.arenas.<i>.bins.<j>.highruns" and
1236    "stats.arenas.<i>.lruns.<j>.highruns" mallctls.
1237  - As part of small size class refactoring, remove the "opt.lg_[qc]space_max",
1238    "arenas.cacheline", "arenas.subpage", "arenas.[tqcs]space_{min,max}", and
1239    "arenas.[tqcs]bins" mallctls.
1240  - Remove the "arenas.chunksize" mallctl.
1241  - Remove the "opt.lg_prof_tcmax" option.
1242  - Remove the "opt.lg_prof_bt_max" option.
1243  - Remove the "opt.lg_tcache_gc_sweep" option.
1244  - Remove the --disable-tiny option, including the "config.tiny" mallctl.
1245  - Remove the --enable-dynamic-page-shift configure option.
1246  - Remove the --enable-sysv configure option.
1247
1248  Bug fixes:
1249  - Fix a statistics-related bug in the "thread.arena" mallctl that could cause
1250    invalid statistics and crashes.
1251  - Work around TLS deallocation via free() on Linux.  This bug could cause
1252    write-after-free memory corruption.
1253  - Fix a potential deadlock that could occur during interval- and
1254    growth-triggered heap profile dumps.
1255  - Fix large calloc() zeroing bugs due to dropping chunk map unzeroed flags.
1256  - Fix chunk_alloc_dss() to stop claiming memory is zeroed.  This bug could
1257    cause memory corruption and crashes with --enable-dss specified.
1258  - Fix fork-related bugs that could cause deadlock in children between fork
1259    and exec.
1260  - Fix malloc_stats_print() to honor 'b' and 'l' in the opts parameter.
1261  - Fix realloc(p, 0) to act like free(p).
1262  - Do not enforce minimum alignment in memalign().
1263  - Check for NULL pointer in malloc_usable_size().
1264  - Fix an off-by-one heap profile statistics bug that could be observed in
1265    interval- and growth-triggered heap profiles.
1266  - Fix the "epoch" mallctl to update cached stats even if the passed in epoch
1267    is 0.
1268  - Fix bin->runcur management to fix a layout policy bug.  This bug did not
1269    affect correctness.
1270  - Fix a bug in choose_arena_hard() that potentially caused more arenas to be
1271    initialized than necessary.
1272  - Add missing "opt.lg_tcache_max" mallctl implementation.
1273  - Use glibc allocator hooks to make mixed allocator usage less likely.
1274  - Fix build issues for --disable-tcache.
1275  - Don't mangle pthread_create() when --with-private-namespace is specified.
1276
1277* 2.2.5 (November 14, 2011)
1278
1279  Bug fixes:
1280  - Fix huge_ralloc() race when using mremap(2).  This is a serious bug that
1281    could cause memory corruption and/or crashes.
1282  - Fix huge_ralloc() to maintain chunk statistics.
1283  - Fix malloc_stats_print(..., "a") output.
1284
1285* 2.2.4 (November 5, 2011)
1286
1287  Bug fixes:
1288  - Initialize arenas_tsd before using it.  This bug existed for 2.2.[0-3], as
1289    well as for --disable-tls builds in earlier releases.
1290  - Do not assume a 4 KiB page size in test/rallocm.c.
1291
1292* 2.2.3 (August 31, 2011)
1293
1294  This version fixes numerous bugs related to heap profiling.
1295
1296  Bug fixes:
1297  - Fix a prof-related race condition.  This bug could cause memory corruption,
1298    but only occurred in non-default configurations (prof_accum:false).
1299  - Fix off-by-one backtracing issues (make sure that prof_alloc_prep() is
1300    excluded from backtraces).
1301  - Fix a prof-related bug in realloc() (only triggered by OOM errors).
1302  - Fix prof-related bugs in allocm() and rallocm().
1303  - Fix prof_tdata_cleanup() for --disable-tls builds.
1304  - Fix a relative include path, to fix objdir builds.
1305
1306* 2.2.2 (July 30, 2011)
1307
1308  Bug fixes:
1309  - Fix a build error for --disable-tcache.
1310  - Fix assertions in arena_purge() (for real this time).
1311  - Add the --with-private-namespace option.  This is a workaround for symbol
1312    conflicts that can inadvertently arise when using static libraries.
1313
1314* 2.2.1 (March 30, 2011)
1315
1316  Bug fixes:
1317  - Implement atomic operations for x86/x64.  This fixes compilation failures
1318    for versions of gcc that are still in wide use.
1319  - Fix an assertion in arena_purge().
1320
1321* 2.2.0 (March 22, 2011)
1322
1323  This version incorporates several improvements to algorithms and data
1324  structures that tend to reduce fragmentation and increase speed.
1325
1326  New features:
1327  - Add the "stats.cactive" mallctl.
1328  - Update pprof (from google-perftools 1.7).
1329  - Improve backtracing-related configuration logic, and add the
1330    --disable-prof-libgcc option.
1331
1332  Bug fixes:
1333  - Change default symbol visibility from "internal", to "hidden", which
1334    decreases the overhead of library-internal function calls.
1335  - Fix symbol visibility so that it is also set on OS X.
1336  - Fix a build dependency regression caused by the introduction of the .pic.o
1337    suffix for PIC object files.
1338  - Add missing checks for mutex initialization failures.
1339  - Don't use libgcc-based backtracing except on x64, where it is known to work.
1340  - Fix deadlocks on OS X that were due to memory allocation in
1341    pthread_mutex_lock().
1342  - Heap profiling-specific fixes:
1343    + Fix memory corruption due to integer overflow in small region index
1344      computation, when using a small enough sample interval that profiling
1345      context pointers are stored in small run headers.
1346    + Fix a bootstrap ordering bug that only occurred with TLS disabled.
1347    + Fix a rallocm() rsize bug.
1348    + Fix error detection bugs for aligned memory allocation.
1349
1350* 2.1.3 (March 14, 2011)
1351
1352  Bug fixes:
1353  - Fix a cpp logic regression (due to the "thread.{de,}allocatedp" mallctl fix
1354    for OS X in 2.1.2).
1355  - Fix a "thread.arena" mallctl bug.
1356  - Fix a thread cache stats merging bug.
1357
1358* 2.1.2 (March 2, 2011)
1359
1360  Bug fixes:
1361  - Fix "thread.{de,}allocatedp" mallctl for OS X.
1362  - Add missing jemalloc.a to build system.
1363
1364* 2.1.1 (January 31, 2011)
1365
1366  Bug fixes:
1367  - Fix aligned huge reallocation (affected allocm()).
1368  - Fix the ALLOCM_LG_ALIGN macro definition.
1369  - Fix a heap dumping deadlock.
1370  - Fix a "thread.arena" mallctl bug.
1371
1372* 2.1.0 (December 3, 2010)
1373
1374  This version incorporates some optimizations that can't quite be considered
1375  bug fixes.
1376
1377  New features:
1378  - Use Linux's mremap(2) for huge object reallocation when possible.
1379  - Avoid locking in mallctl*() when possible.
1380  - Add the "thread.[de]allocatedp" mallctl's.
1381  - Convert the manual page source from roff to DocBook, and generate both roff
1382    and HTML manuals.
1383
1384  Bug fixes:
1385  - Fix a crash due to incorrect bootstrap ordering.  This only impacted
1386    --enable-debug --enable-dss configurations.
1387  - Fix a minor statistics bug for mallctl("swap.avail", ...).
1388
1389* 2.0.1 (October 29, 2010)
1390
1391  Bug fixes:
1392  - Fix a race condition in heap profiling that could cause undefined behavior
1393    if "opt.prof_accum" were disabled.
1394  - Add missing mutex unlocks for some OOM error paths in the heap profiling
1395    code.
1396  - Fix a compilation error for non-C99 builds.
1397
1398* 2.0.0 (October 24, 2010)
1399
1400  This version focuses on the experimental *allocm() API, and on improved
1401  run-time configuration/introspection.  Nonetheless, numerous performance
1402  improvements are also included.
1403
1404  New features:
1405  - Implement the experimental {,r,s,d}allocm() API, which provides a superset
1406    of the functionality available via malloc(), calloc(), posix_memalign(),
1407    realloc(), malloc_usable_size(), and free().  These functions can be used to
1408    allocate/reallocate aligned zeroed memory, ask for optional extra memory
1409    during reallocation, prevent object movement during reallocation, etc.
1410  - Replace JEMALLOC_OPTIONS/JEMALLOC_PROF_PREFIX with MALLOC_CONF, which is
1411    more human-readable, and more flexible.  For example:
1412      JEMALLOC_OPTIONS=AJP
1413    is now:
1414      MALLOC_CONF=abort:true,fill:true,stats_print:true
1415  - Port to Apple OS X.  Sponsored by Mozilla.
1416  - Make it possible for the application to control thread-->arena mappings via
1417    the "thread.arena" mallctl.
1418  - Add compile-time support for all TLS-related functionality via pthreads TSD.
1419    This is mainly of interest for OS X, which does not support TLS, but has a
1420    TSD implementation with similar performance.
1421  - Override memalign() and valloc() if they are provided by the system.
1422  - Add the "arenas.purge" mallctl, which can be used to synchronously purge all
1423    dirty unused pages.
1424  - Make cumulative heap profiling data optional, so that it is possible to
1425    limit the amount of memory consumed by heap profiling data structures.
1426  - Add per thread allocation counters that can be accessed via the
1427    "thread.allocated" and "thread.deallocated" mallctls.
1428
1429  Incompatible changes:
1430  - Remove JEMALLOC_OPTIONS and malloc_options (see MALLOC_CONF above).
1431  - Increase default backtrace depth from 4 to 128 for heap profiling.
1432  - Disable interval-based profile dumps by default.
1433
1434  Bug fixes:
1435  - Remove bad assertions in fork handler functions.  These assertions could
1436    cause aborts for some combinations of configure settings.
1437  - Fix strerror_r() usage to deal with non-standard semantics in GNU libc.
1438  - Fix leak context reporting.  This bug tended to cause the number of contexts
1439    to be underreported (though the reported number of objects and bytes were
1440    correct).
1441  - Fix a realloc() bug for large in-place growing reallocation.  This bug could
1442    cause memory corruption, but it was hard to trigger.
1443  - Fix an allocation bug for small allocations that could be triggered if
1444    multiple threads raced to create a new run of backing pages.
1445  - Enhance the heap profiler to trigger samples based on usable size, rather
1446    than request size.
1447  - Fix a heap profiling bug due to sometimes losing track of requested object
1448    size for sampled objects.
1449
1450* 1.0.3 (August 12, 2010)
1451
1452  Bug fixes:
1453  - Fix the libunwind-based implementation of stack backtracing (used for heap
1454    profiling).  This bug could cause zero-length backtraces to be reported.
1455  - Add a missing mutex unlock in library initialization code.  If multiple
1456    threads raced to initialize malloc, some of them could end up permanently
1457    blocked.
1458
1459* 1.0.2 (May 11, 2010)
1460
1461  Bug fixes:
1462  - Fix junk filling of large objects, which could cause memory corruption.
1463  - Add MAP_NORESERVE support for chunk mapping, because otherwise virtual
1464    memory limits could cause swap file configuration to fail.  Contributed by
1465    Jordan DeLong.
1466
1467* 1.0.1 (April 14, 2010)
1468
1469  Bug fixes:
1470  - Fix compilation when --enable-fill is specified.
1471  - Fix threads-related profiling bugs that affected accuracy and caused memory
1472    to be leaked during thread exit.
1473  - Fix dirty page purging race conditions that could cause crashes.
1474  - Fix crash in tcache flushing code during thread destruction.
1475
1476* 1.0.0 (April 11, 2010)
1477
1478  This release focuses on speed and run-time introspection.  Numerous
1479  algorithmic improvements make this release substantially faster than its
1480  predecessors.
1481
1482  New features:
1483  - Implement autoconf-based configuration system.
1484  - Add mallctl*(), for the purposes of introspection and run-time
1485    configuration.
1486  - Make it possible for the application to manually flush a thread's cache, via
1487    the "tcache.flush" mallctl.
1488  - Base maximum dirty page count on proportion of active memory.
1489  - Compute various additional run-time statistics, including per size class
1490    statistics for large objects.
1491  - Expose malloc_stats_print(), which can be called repeatedly by the
1492    application.
1493  - Simplify the malloc_message() signature to only take one string argument,
1494    and incorporate an opaque data pointer argument for use by the application
1495    in combination with malloc_stats_print().
1496  - Add support for allocation backed by one or more swap files, and allow the
1497    application to disable over-commit if swap files are in use.
1498  - Implement allocation profiling and leak checking.
1499
1500  Removed features:
1501  - Remove the dynamic arena rebalancing code, since thread-specific caching
1502    reduces its utility.
1503
1504  Bug fixes:
1505  - Modify chunk allocation to work when address space layout randomization
1506    (ASLR) is in use.
1507  - Fix thread cleanup bugs related to TLS destruction.
1508  - Handle 0-size allocation requests in posix_memalign().
1509  - Fix a chunk leak.  The leaked chunks were never touched, so this impacted
1510    virtual memory usage, but not physical memory usage.
1511
1512* linux_2008082[78]a (August 27/28, 2008)
1513
1514  These snapshot releases are the simple result of incorporating Linux-specific
1515  support into the FreeBSD malloc sources.
1516
1517--------------------------------------------------------------------------------
1518vim:filetype=text:textwidth=80
1519