xref: /freebsd/contrib/jemalloc/ChangeLog (revision 93e779a26c651610ac6e7986d67ecc9ed2cadcbf)
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* 4.0.4 (October 24, 2015)
8
9  This bugfix release fixes another xallocx() regression.  No other regressions
10  have come to light in over a month, so this is likely a good starting point
11  for people who prefer to wait for "dot one" releases with all the major issues
12  shaken out.
13
14  Bug fixes:
15  - Fix xallocx(..., MALLOCX_ZERO to zero the last full trailing page of large
16    allocations that have been randomly assigned an offset of 0 when
17    --enable-cache-oblivious configure option is enabled.
18
19* 4.0.3 (September 24, 2015)
20
21  This bugfix release continues the trend of xallocx() and heap profiling fixes.
22
23  Bug fixes:
24  - Fix xallocx(..., MALLOCX_ZERO) to zero all trailing bytes of large
25    allocations when --enable-cache-oblivious configure option is enabled.
26  - Fix xallocx(..., MALLOCX_ZERO) to zero trailing bytes of huge allocations
27    when resizing from/to a size class that is not a multiple of the chunk size.
28  - Fix prof_tctx_dump_iter() to filter out nodes that were created after heap
29    profile dumping started.
30  - Work around a potentially bad thread-specific data initialization
31    interaction with NPTL (glibc's pthreads implementation).
32
33* 4.0.2 (September 21, 2015)
34
35  This bugfix release addresses a few bugs specific to heap profiling.
36
37  Bug fixes:
38  - Fix ixallocx_prof_sample() to never modify nor create sampled small
39    allocations.  xallocx() is in general incapable of moving small allocations,
40    so this fix removes buggy code without loss of generality.
41  - Fix irallocx_prof_sample() to always allocate large regions, even when
42    alignment is non-zero.
43  - Fix prof_alloc_rollback() to read tdata from thread-specific data rather
44    than dereferencing a potentially invalid tctx.
45
46* 4.0.1 (September 15, 2015)
47
48  This is a bugfix release that is somewhat high risk due to the amount of
49  refactoring required to address deep xallocx() problems.  As a side effect of
50  these fixes, xallocx() now tries harder to partially fulfill requests for
51  optional extra space.  Note that a couple of minor heap profiling
52  optimizations are included, but these are better thought of as performance
53  fixes that were integral to disovering most of the other bugs.
54
55  Optimizations:
56  - Avoid a chunk metadata read in arena_prof_tctx_set(), since it is in the
57    fast path when heap profiling is enabled.  Additionally, split a special
58    case out into arena_prof_tctx_reset(), which also avoids chunk metadata
59    reads.
60  - Optimize irallocx_prof() to optimistically update the sampler state.  The
61    prior implementation appears to have been a holdover from when
62    rallocx()/xallocx() functionality was combined as rallocm().
63
64  Bug fixes:
65  - Fix TLS configuration such that it is enabled by default for platforms on
66    which it works correctly.
67  - Fix arenas_cache_cleanup() and arena_get_hard() to handle
68    allocation/deallocation within the application's thread-specific data
69    cleanup functions even after arenas_cache is torn down.
70  - Fix xallocx() bugs related to size+extra exceeding HUGE_MAXCLASS.
71  - Fix chunk purge hook calls for in-place huge shrinking reallocation to
72    specify the old chunk size rather than the new chunk size.  This bug caused
73    no correctness issues for the default chunk purge function, but was
74    visible to custom functions set via the "arena.<i>.chunk_hooks" mallctl.
75  - Fix heap profiling bugs:
76    + Fix heap profiling to distinguish among otherwise identical sample sites
77      with interposed resets (triggered via the "prof.reset" mallctl).  This bug
78      could cause data structure corruption that would most likely result in a
79      segfault.
80    + Fix irealloc_prof() to prof_alloc_rollback() on OOM.
81    + Make one call to prof_active_get_unlocked() per allocation event, and use
82      the result throughout the relevant functions that handle an allocation
83      event.  Also add a missing check in prof_realloc().  These fixes protect
84      allocation events against concurrent prof_active changes.
85    + Fix ixallocx_prof() to pass usize_max and zero to ixallocx_prof_sample()
86      in the correct order.
87    + Fix prof_realloc() to call prof_free_sampled_object() after calling
88      prof_malloc_sample_object().  Prior to this fix, if tctx and old_tctx were
89      the same, the tctx could have been prematurely destroyed.
90  - Fix portability bugs:
91    + Don't bitshift by negative amounts when encoding/decoding run sizes in
92      chunk header maps.  This affected systems with page sizes greater than 8
93      KiB.
94    + Rename index_t to szind_t to avoid an existing type on Solaris.
95    + Add JEMALLOC_CXX_THROW to the memalign() function prototype, in order to
96      match glibc and avoid compilation errors when including both
97      jemalloc/jemalloc.h and malloc.h in C++ code.
98    + Don't assume that /bin/sh is appropriate when running size_classes.sh
99      during configuration.
100    + Consider __sparcv9 a synonym for __sparc64__ when defining LG_QUANTUM.
101    + Link tests to librt if it contains clock_gettime(2).
102
103* 4.0.0 (August 17, 2015)
104
105  This version contains many speed and space optimizations, both minor and
106  major.  The major themes are generalization, unification, and simplification.
107  Although many of these optimizations cause no visible behavior change, their
108  cumulative effect is substantial.
109
110  New features:
111  - Normalize size class spacing to be consistent across the complete size
112    range.  By default there are four size classes per size doubling, but this
113    is now configurable via the --with-lg-size-class-group option.  Also add the
114    --with-lg-page, --with-lg-page-sizes, --with-lg-quantum, and
115    --with-lg-tiny-min options, which can be used to tweak page and size class
116    settings.  Impacts:
117    + Worst case performance for incrementally growing/shrinking reallocation
118      is improved because there are far fewer size classes, and therefore
119      copying happens less often.
120    + Internal fragmentation is limited to 20% for all but the smallest size
121      classes (those less than four times the quantum).  (1B + 4 KiB)
122      and (1B + 4 MiB) previously suffered nearly 50% internal fragmentation.
123    + Chunk fragmentation tends to be lower because there are fewer distinct run
124      sizes to pack.
125  - Add support for explicit tcaches.  The "tcache.create", "tcache.flush", and
126    "tcache.destroy" mallctls control tcache lifetime and flushing, and the
127    MALLOCX_TCACHE(tc) and MALLOCX_TCACHE_NONE flags to the *allocx() API
128    control which tcache is used for each operation.
129  - Implement per thread heap profiling, as well as the ability to
130    enable/disable heap profiling on a per thread basis.  Add the "prof.reset",
131    "prof.lg_sample", "thread.prof.name", "thread.prof.active",
132    "opt.prof_thread_active_init", "prof.thread_active_init", and
133    "thread.prof.active" mallctls.
134  - Add support for per arena application-specified chunk allocators, configured
135    via the "arena.<i>.chunk_hooks" mallctl.
136  - Refactor huge allocation to be managed by arenas, so that arenas now
137    function as general purpose independent allocators.  This is important in
138    the context of user-specified chunk allocators, aside from the scalability
139    benefits.  Related new statistics:
140    + The "stats.arenas.<i>.huge.allocated", "stats.arenas.<i>.huge.nmalloc",
141      "stats.arenas.<i>.huge.ndalloc", and "stats.arenas.<i>.huge.nrequests"
142      mallctls provide high level per arena huge allocation statistics.
143    + The "arenas.nhchunks", "arenas.hchunk.<i>.size",
144      "stats.arenas.<i>.hchunks.<j>.nmalloc",
145      "stats.arenas.<i>.hchunks.<j>.ndalloc",
146      "stats.arenas.<i>.hchunks.<j>.nrequests", and
147      "stats.arenas.<i>.hchunks.<j>.curhchunks" mallctls provide per size class
148      statistics.
149  - Add the 'util' column to malloc_stats_print() output, which reports the
150    proportion of available regions that are currently in use for each small
151    size class.
152  - Add "alloc" and "free" modes for for junk filling (see the "opt.junk"
153    mallctl), so that it is possible to separately enable junk filling for
154    allocation versus deallocation.
155  - Add the jemalloc-config script, which provides information about how
156    jemalloc was configured, and how to integrate it into application builds.
157  - Add metadata statistics, which are accessible via the "stats.metadata",
158    "stats.arenas.<i>.metadata.mapped", and
159    "stats.arenas.<i>.metadata.allocated" mallctls.
160  - Add the "stats.resident" mallctl, which reports the upper limit of
161    physically resident memory mapped by the allocator.
162  - Add per arena control over unused dirty page purging, via the
163    "arenas.lg_dirty_mult", "arena.<i>.lg_dirty_mult", and
164    "stats.arenas.<i>.lg_dirty_mult" mallctls.
165  - Add the "prof.gdump" mallctl, which makes it possible to toggle the gdump
166    feature on/off during program execution.
167  - Add sdallocx(), which implements sized deallocation.  The primary
168    optimization over dallocx() is the removal of a metadata read, which often
169    suffers an L1 cache miss.
170  - Add missing header includes in jemalloc/jemalloc.h, so that applications
171    only have to #include <jemalloc/jemalloc.h>.
172  - Add support for additional platforms:
173    + Bitrig
174    + Cygwin
175    + DragonFlyBSD
176    + iOS
177    + OpenBSD
178    + OpenRISC/or1k
179
180  Optimizations:
181  - Maintain dirty runs in per arena LRUs rather than in per arena trees of
182    dirty-run-containing chunks.  In practice this change significantly reduces
183    dirty page purging volume.
184  - Integrate whole chunks into the unused dirty page purging machinery.  This
185    reduces the cost of repeated huge allocation/deallocation, because it
186    effectively introduces a cache of chunks.
187  - Split the arena chunk map into two separate arrays, in order to increase
188    cache locality for the frequently accessed bits.
189  - Move small run metadata out of runs, into arena chunk headers.  This reduces
190    run fragmentation, smaller runs reduce external fragmentation for small size
191    classes, and packed (less uniformly aligned) metadata layout improves CPU
192    cache set distribution.
193  - Randomly distribute large allocation base pointer alignment relative to page
194    boundaries in order to more uniformly utilize CPU cache sets.  This can be
195    disabled via the --disable-cache-oblivious configure option, and queried via
196    the "config.cache_oblivious" mallctl.
197  - Micro-optimize the fast paths for the public API functions.
198  - Refactor thread-specific data to reside in a single structure.  This assures
199    that only a single TLS read is necessary per call into the public API.
200  - Implement in-place huge allocation growing and shrinking.
201  - Refactor rtree (radix tree for chunk lookups) to be lock-free, and make
202    additional optimizations that reduce maximum lookup depth to one or two
203    levels.  This resolves what was a concurrency bottleneck for per arena huge
204    allocation, because a global data structure is critical for determining
205    which arenas own which huge allocations.
206
207  Incompatible changes:
208  - Replace --enable-cc-silence with --disable-cc-silence to suppress spurious
209    warnings by default.
210  - Assure that the constness of malloc_usable_size()'s return type matches that
211    of the system implementation.
212  - Change the heap profile dump format to support per thread heap profiling,
213    rename pprof to jeprof, and enhance it with the --thread=<n> option.  As a
214    result, the bundled jeprof must now be used rather than the upstream
215    (gperftools) pprof.
216  - Disable "opt.prof_final" by default, in order to avoid atexit(3), which can
217    internally deadlock on some platforms.
218  - Change the "arenas.nlruns" mallctl type from size_t to unsigned.
219  - Replace the "stats.arenas.<i>.bins.<j>.allocated" mallctl with
220    "stats.arenas.<i>.bins.<j>.curregs".
221  - Ignore MALLOC_CONF in set{uid,gid,cap} binaries.
222  - Ignore MALLOCX_ARENA(a) in dallocx(), in favor of using the
223    MALLOCX_TCACHE(tc) and MALLOCX_TCACHE_NONE flags to control tcache usage.
224
225  Removed features:
226  - Remove the *allocm() API, which is superseded by the *allocx() API.
227  - Remove the --enable-dss options, and make dss non-optional on all platforms
228    which support sbrk(2).
229  - Remove the "arenas.purge" mallctl, which was obsoleted by the
230    "arena.<i>.purge" mallctl in 3.1.0.
231  - Remove the unnecessary "opt.valgrind" mallctl; jemalloc automatically
232    detects whether it is running inside Valgrind.
233  - Remove the "stats.huge.allocated", "stats.huge.nmalloc", and
234    "stats.huge.ndalloc" mallctls.
235  - Remove the --enable-mremap option.
236  - Remove the "stats.chunks.current", "stats.chunks.total", and
237    "stats.chunks.high" mallctls.
238
239  Bug fixes:
240  - Fix the cactive statistic to decrease (rather than increase) when active
241    memory decreases.  This regression was first released in 3.5.0.
242  - Fix OOM handling in memalign() and valloc().  A variant of this bug existed
243    in all releases since 2.0.0, which introduced these functions.
244  - Fix an OOM-related regression in arena_tcache_fill_small(), which could
245    cause cache corruption on OOM.  This regression was present in all releases
246    from 2.2.0 through 3.6.0.
247  - Fix size class overflow handling for malloc(), posix_memalign(), memalign(),
248    calloc(), and realloc() when profiling is enabled.
249  - Fix the "arena.<i>.dss" mallctl to return an error if "primary" or
250    "secondary" precedence is specified, but sbrk(2) is not supported.
251  - Fix fallback lg_floor() implementations to handle extremely large inputs.
252  - Ensure the default purgeable zone is after the default zone on OS X.
253  - Fix latent bugs in atomic_*().
254  - Fix the "arena.<i>.dss" mallctl to handle read-only calls.
255  - Fix tls_model configuration to enable the initial-exec model when possible.
256  - Mark malloc_conf as a weak symbol so that the application can override it.
257  - Correctly detect glibc's adaptive pthread mutexes.
258  - Fix the --without-export configure option.
259
260* 3.6.0 (March 31, 2014)
261
262  This version contains a critical bug fix for a regression present in 3.5.0 and
263  3.5.1.
264
265  Bug fixes:
266  - Fix a regression in arena_chunk_alloc() that caused crashes during
267    small/large allocation if chunk allocation failed.  In the absence of this
268    bug, chunk allocation failure would result in allocation failure, e.g.  NULL
269    return from malloc().  This regression was introduced in 3.5.0.
270  - Fix backtracing for gcc intrinsics-based backtracing by specifying
271    -fno-omit-frame-pointer to gcc.  Note that the application (and all the
272    libraries it links to) must also be compiled with this option for
273    backtracing to be reliable.
274  - Use dss allocation precedence for huge allocations as well as small/large
275    allocations.
276  - Fix test assertion failure message formatting.  This bug did not manifest on
277    x86_64 systems because of implementation subtleties in va_list.
278  - Fix inconsequential test failures for hash and SFMT code.
279
280  New features:
281  - Support heap profiling on FreeBSD.  This feature depends on the proc
282    filesystem being mounted during heap profile dumping.
283
284* 3.5.1 (February 25, 2014)
285
286  This version primarily addresses minor bugs in test code.
287
288  Bug fixes:
289  - Configure Solaris/Illumos to use MADV_FREE.
290  - Fix junk filling for mremap(2)-based huge reallocation.  This is only
291    relevant if configuring with the --enable-mremap option specified.
292  - Avoid compilation failure if 'restrict' C99 keyword is not supported by the
293    compiler.
294  - Add a configure test for SSE2 rather than assuming it is usable on i686
295    systems.  This fixes test compilation errors, especially on 32-bit Linux
296    systems.
297  - Fix mallctl argument size mismatches (size_t vs. uint64_t) in the stats unit
298    test.
299  - Fix/remove flawed alignment-related overflow tests.
300  - Prevent compiler optimizations that could change backtraces in the
301    prof_accum unit test.
302
303* 3.5.0 (January 22, 2014)
304
305  This version focuses on refactoring and automated testing, though it also
306  includes some non-trivial heap profiling optimizations not mentioned below.
307
308  New features:
309  - Add the *allocx() API, which is a successor to the experimental *allocm()
310    API.  The *allocx() functions are slightly simpler to use because they have
311    fewer parameters, they directly return the results of primary interest, and
312    mallocx()/rallocx() avoid the strict aliasing pitfall that
313    allocm()/rallocm() share with posix_memalign().  Note that *allocm() is
314    slated for removal in the next non-bugfix release.
315  - Add support for LinuxThreads.
316
317  Bug fixes:
318  - Unless heap profiling is enabled, disable floating point code and don't link
319    with libm.  This, in combination with e.g. EXTRA_CFLAGS=-mno-sse on x64
320    systems, makes it possible to completely disable floating point register
321    use.  Some versions of glibc neglect to save/restore caller-saved floating
322    point registers during dynamic lazy symbol loading, and the symbol loading
323    code uses whatever malloc the application happens to have linked/loaded
324    with, the result being potential floating point register corruption.
325  - Report ENOMEM rather than EINVAL if an OOM occurs during heap profiling
326    backtrace creation in imemalign().  This bug impacted posix_memalign() and
327    aligned_alloc().
328  - Fix a file descriptor leak in a prof_dump_maps() error path.
329  - Fix prof_dump() to close the dump file descriptor for all relevant error
330    paths.
331  - Fix rallocm() to use the arena specified by the ALLOCM_ARENA(s) flag for
332    allocation, not just deallocation.
333  - Fix a data race for large allocation stats counters.
334  - Fix a potential infinite loop during thread exit.  This bug occurred on
335    Solaris, and could affect other platforms with similar pthreads TSD
336    implementations.
337  - Don't junk-fill reallocations unless usable size changes.  This fixes a
338    violation of the *allocx()/*allocm() semantics.
339  - Fix growing large reallocation to junk fill new space.
340  - Fix huge deallocation to junk fill when munmap is disabled.
341  - Change the default private namespace prefix from empty to je_, and change
342    --with-private-namespace-prefix so that it prepends an additional prefix
343    rather than replacing je_.  This reduces the likelihood of applications
344    which statically link jemalloc experiencing symbol name collisions.
345  - Add missing private namespace mangling (relevant when
346    --with-private-namespace is specified).
347  - Add and use JEMALLOC_INLINE_C so that static inline functions are marked as
348    static even for debug builds.
349  - Add a missing mutex unlock in a malloc_init_hard() error path.  In practice
350    this error path is never executed.
351  - Fix numerous bugs in malloc_strotumax() error handling/reporting.  These
352    bugs had no impact except for malformed inputs.
353  - Fix numerous bugs in malloc_snprintf().  These bugs were not exercised by
354    existing calls, so they had no impact.
355
356* 3.4.1 (October 20, 2013)
357
358  Bug fixes:
359  - Fix a race in the "arenas.extend" mallctl that could cause memory corruption
360    of internal data structures and subsequent crashes.
361  - Fix Valgrind integration flaws that caused Valgrind warnings about reads of
362    uninitialized memory in:
363    + arena chunk headers
364    + internal zero-initialized data structures (relevant to tcache and prof
365      code)
366  - Preserve errno during the first allocation.  A readlink(2) call during
367    initialization fails unless /etc/malloc.conf exists, so errno was typically
368    set during the first allocation prior to this fix.
369  - Fix compilation warnings reported by gcc 4.8.1.
370
371* 3.4.0 (June 2, 2013)
372
373  This version is essentially a small bugfix release, but the addition of
374  aarch64 support requires that the minor version be incremented.
375
376  Bug fixes:
377  - Fix race-triggered deadlocks in chunk_record().  These deadlocks were
378    typically triggered by multiple threads concurrently deallocating huge
379    objects.
380
381  New features:
382  - Add support for the aarch64 architecture.
383
384* 3.3.1 (March 6, 2013)
385
386  This version fixes bugs that are typically encountered only when utilizing
387  custom run-time options.
388
389  Bug fixes:
390  - Fix a locking order bug that could cause deadlock during fork if heap
391    profiling were enabled.
392  - Fix a chunk recycling bug that could cause the allocator to lose track of
393    whether a chunk was zeroed.  On FreeBSD, NetBSD, and OS X, it could cause
394    corruption if allocating via sbrk(2) (unlikely unless running with the
395    "dss:primary" option specified).  This was completely harmless on Linux
396    unless using mlockall(2) (and unlikely even then, unless the
397    --disable-munmap configure option or the "dss:primary" option was
398    specified).  This regression was introduced in 3.1.0 by the
399    mlockall(2)/madvise(2) interaction fix.
400  - Fix TLS-related memory corruption that could occur during thread exit if the
401    thread never allocated memory.  Only the quarantine and prof facilities were
402    susceptible.
403  - Fix two quarantine bugs:
404    + Internal reallocation of the quarantined object array leaked the old
405      array.
406    + Reallocation failure for internal reallocation of the quarantined object
407      array (very unlikely) resulted in memory corruption.
408  - Fix Valgrind integration to annotate all internally allocated memory in a
409    way that keeps Valgrind happy about internal data structure access.
410  - Fix building for s390 systems.
411
412* 3.3.0 (January 23, 2013)
413
414  This version includes a few minor performance improvements in addition to the
415  listed new features and bug fixes.
416
417  New features:
418  - Add clipping support to lg_chunk option processing.
419  - Add the --enable-ivsalloc option.
420  - Add the --without-export option.
421  - Add the --disable-zone-allocator option.
422
423  Bug fixes:
424  - Fix "arenas.extend" mallctl to output the number of arenas.
425  - Fix chunk_recycle() to unconditionally inform Valgrind that returned memory
426    is undefined.
427  - Fix build break on FreeBSD related to alloca.h.
428
429* 3.2.0 (November 9, 2012)
430
431  In addition to a couple of bug fixes, this version modifies page run
432  allocation and dirty page purging algorithms in order to better control
433  page-level virtual memory fragmentation.
434
435  Incompatible changes:
436  - Change the "opt.lg_dirty_mult" default from 5 to 3 (32:1 to 8:1).
437
438  Bug fixes:
439  - Fix dss/mmap allocation precedence code to use recyclable mmap memory only
440    after primary dss allocation fails.
441  - Fix deadlock in the "arenas.purge" mallctl.  This regression was introduced
442    in 3.1.0 by the addition of the "arena.<i>.purge" mallctl.
443
444* 3.1.0 (October 16, 2012)
445
446  New features:
447  - Auto-detect whether running inside Valgrind, thus removing the need to
448    manually specify MALLOC_CONF=valgrind:true.
449  - Add the "arenas.extend" mallctl, which allows applications to create
450    manually managed arenas.
451  - Add the ALLOCM_ARENA() flag for {,r,d}allocm().
452  - Add the "opt.dss", "arena.<i>.dss", and "stats.arenas.<i>.dss" mallctls,
453    which provide control over dss/mmap precedence.
454  - Add the "arena.<i>.purge" mallctl, which obsoletes "arenas.purge".
455  - Define LG_QUANTUM for hppa.
456
457  Incompatible changes:
458  - Disable tcache by default if running inside Valgrind, in order to avoid
459    making unallocated objects appear reachable to Valgrind.
460  - Drop const from malloc_usable_size() argument on Linux.
461
462  Bug fixes:
463  - Fix heap profiling crash if sampled object is freed via realloc(p, 0).
464  - Remove const from __*_hook variable declarations, so that glibc can modify
465    them during process forking.
466  - Fix mlockall(2)/madvise(2) interaction.
467  - Fix fork(2)-related deadlocks.
468  - Fix error return value for "thread.tcache.enabled" mallctl.
469
470* 3.0.0 (May 11, 2012)
471
472  Although this version adds some major new features, the primary focus is on
473  internal code cleanup that facilitates maintainability and portability, most
474  of which is not reflected in the ChangeLog.  This is the first release to
475  incorporate substantial contributions from numerous other developers, and the
476  result is a more broadly useful allocator (see the git revision history for
477  contribution details).  Note that the license has been unified, thanks to
478  Facebook granting a license under the same terms as the other copyright
479  holders (see COPYING).
480
481  New features:
482  - Implement Valgrind support, redzones, and quarantine.
483  - Add support for additional platforms:
484    + FreeBSD
485    + Mac OS X Lion
486    + MinGW
487    + Windows (no support yet for replacing the system malloc)
488  - Add support for additional architectures:
489    + MIPS
490    + SH4
491    + Tilera
492  - Add support for cross compiling.
493  - Add nallocm(), which rounds a request size up to the nearest size class
494    without actually allocating.
495  - Implement aligned_alloc() (blame C11).
496  - Add the "thread.tcache.enabled" mallctl.
497  - Add the "opt.prof_final" mallctl.
498  - Update pprof (from gperftools 2.0).
499  - Add the --with-mangling option.
500  - Add the --disable-experimental option.
501  - Add the --disable-munmap option, and make it the default on Linux.
502  - Add the --enable-mremap option, which disables use of mremap(2) by default.
503
504  Incompatible changes:
505  - Enable stats by default.
506  - Enable fill by default.
507  - Disable lazy locking by default.
508  - Rename the "tcache.flush" mallctl to "thread.tcache.flush".
509  - Rename the "arenas.pagesize" mallctl to "arenas.page".
510  - Change the "opt.lg_prof_sample" default from 0 to 19 (1 B to 512 KiB).
511  - Change the "opt.prof_accum" default from true to false.
512
513  Removed features:
514  - Remove the swap feature, including the "config.swap", "swap.avail",
515    "swap.prezeroed", "swap.nfds", and "swap.fds" mallctls.
516  - Remove highruns statistics, including the
517    "stats.arenas.<i>.bins.<j>.highruns" and
518    "stats.arenas.<i>.lruns.<j>.highruns" mallctls.
519  - As part of small size class refactoring, remove the "opt.lg_[qc]space_max",
520    "arenas.cacheline", "arenas.subpage", "arenas.[tqcs]space_{min,max}", and
521    "arenas.[tqcs]bins" mallctls.
522  - Remove the "arenas.chunksize" mallctl.
523  - Remove the "opt.lg_prof_tcmax" option.
524  - Remove the "opt.lg_prof_bt_max" option.
525  - Remove the "opt.lg_tcache_gc_sweep" option.
526  - Remove the --disable-tiny option, including the "config.tiny" mallctl.
527  - Remove the --enable-dynamic-page-shift configure option.
528  - Remove the --enable-sysv configure option.
529
530  Bug fixes:
531  - Fix a statistics-related bug in the "thread.arena" mallctl that could cause
532    invalid statistics and crashes.
533  - Work around TLS deallocation via free() on Linux.  This bug could cause
534    write-after-free memory corruption.
535  - Fix a potential deadlock that could occur during interval- and
536    growth-triggered heap profile dumps.
537  - Fix large calloc() zeroing bugs due to dropping chunk map unzeroed flags.
538  - Fix chunk_alloc_dss() to stop claiming memory is zeroed.  This bug could
539    cause memory corruption and crashes with --enable-dss specified.
540  - Fix fork-related bugs that could cause deadlock in children between fork
541    and exec.
542  - Fix malloc_stats_print() to honor 'b' and 'l' in the opts parameter.
543  - Fix realloc(p, 0) to act like free(p).
544  - Do not enforce minimum alignment in memalign().
545  - Check for NULL pointer in malloc_usable_size().
546  - Fix an off-by-one heap profile statistics bug that could be observed in
547    interval- and growth-triggered heap profiles.
548  - Fix the "epoch" mallctl to update cached stats even if the passed in epoch
549    is 0.
550  - Fix bin->runcur management to fix a layout policy bug.  This bug did not
551    affect correctness.
552  - Fix a bug in choose_arena_hard() that potentially caused more arenas to be
553    initialized than necessary.
554  - Add missing "opt.lg_tcache_max" mallctl implementation.
555  - Use glibc allocator hooks to make mixed allocator usage less likely.
556  - Fix build issues for --disable-tcache.
557  - Don't mangle pthread_create() when --with-private-namespace is specified.
558
559* 2.2.5 (November 14, 2011)
560
561  Bug fixes:
562  - Fix huge_ralloc() race when using mremap(2).  This is a serious bug that
563    could cause memory corruption and/or crashes.
564  - Fix huge_ralloc() to maintain chunk statistics.
565  - Fix malloc_stats_print(..., "a") output.
566
567* 2.2.4 (November 5, 2011)
568
569  Bug fixes:
570  - Initialize arenas_tsd before using it.  This bug existed for 2.2.[0-3], as
571    well as for --disable-tls builds in earlier releases.
572  - Do not assume a 4 KiB page size in test/rallocm.c.
573
574* 2.2.3 (August 31, 2011)
575
576  This version fixes numerous bugs related to heap profiling.
577
578  Bug fixes:
579  - Fix a prof-related race condition.  This bug could cause memory corruption,
580    but only occurred in non-default configurations (prof_accum:false).
581  - Fix off-by-one backtracing issues (make sure that prof_alloc_prep() is
582    excluded from backtraces).
583  - Fix a prof-related bug in realloc() (only triggered by OOM errors).
584  - Fix prof-related bugs in allocm() and rallocm().
585  - Fix prof_tdata_cleanup() for --disable-tls builds.
586  - Fix a relative include path, to fix objdir builds.
587
588* 2.2.2 (July 30, 2011)
589
590  Bug fixes:
591  - Fix a build error for --disable-tcache.
592  - Fix assertions in arena_purge() (for real this time).
593  - Add the --with-private-namespace option.  This is a workaround for symbol
594    conflicts that can inadvertently arise when using static libraries.
595
596* 2.2.1 (March 30, 2011)
597
598  Bug fixes:
599  - Implement atomic operations for x86/x64.  This fixes compilation failures
600    for versions of gcc that are still in wide use.
601  - Fix an assertion in arena_purge().
602
603* 2.2.0 (March 22, 2011)
604
605  This version incorporates several improvements to algorithms and data
606  structures that tend to reduce fragmentation and increase speed.
607
608  New features:
609  - Add the "stats.cactive" mallctl.
610  - Update pprof (from google-perftools 1.7).
611  - Improve backtracing-related configuration logic, and add the
612    --disable-prof-libgcc option.
613
614  Bug fixes:
615  - Change default symbol visibility from "internal", to "hidden", which
616    decreases the overhead of library-internal function calls.
617  - Fix symbol visibility so that it is also set on OS X.
618  - Fix a build dependency regression caused by the introduction of the .pic.o
619    suffix for PIC object files.
620  - Add missing checks for mutex initialization failures.
621  - Don't use libgcc-based backtracing except on x64, where it is known to work.
622  - Fix deadlocks on OS X that were due to memory allocation in
623    pthread_mutex_lock().
624  - Heap profiling-specific fixes:
625    + Fix memory corruption due to integer overflow in small region index
626      computation, when using a small enough sample interval that profiling
627      context pointers are stored in small run headers.
628    + Fix a bootstrap ordering bug that only occurred with TLS disabled.
629    + Fix a rallocm() rsize bug.
630    + Fix error detection bugs for aligned memory allocation.
631
632* 2.1.3 (March 14, 2011)
633
634  Bug fixes:
635  - Fix a cpp logic regression (due to the "thread.{de,}allocatedp" mallctl fix
636    for OS X in 2.1.2).
637  - Fix a "thread.arena" mallctl bug.
638  - Fix a thread cache stats merging bug.
639
640* 2.1.2 (March 2, 2011)
641
642  Bug fixes:
643  - Fix "thread.{de,}allocatedp" mallctl for OS X.
644  - Add missing jemalloc.a to build system.
645
646* 2.1.1 (January 31, 2011)
647
648  Bug fixes:
649  - Fix aligned huge reallocation (affected allocm()).
650  - Fix the ALLOCM_LG_ALIGN macro definition.
651  - Fix a heap dumping deadlock.
652  - Fix a "thread.arena" mallctl bug.
653
654* 2.1.0 (December 3, 2010)
655
656  This version incorporates some optimizations that can't quite be considered
657  bug fixes.
658
659  New features:
660  - Use Linux's mremap(2) for huge object reallocation when possible.
661  - Avoid locking in mallctl*() when possible.
662  - Add the "thread.[de]allocatedp" mallctl's.
663  - Convert the manual page source from roff to DocBook, and generate both roff
664    and HTML manuals.
665
666  Bug fixes:
667  - Fix a crash due to incorrect bootstrap ordering.  This only impacted
668    --enable-debug --enable-dss configurations.
669  - Fix a minor statistics bug for mallctl("swap.avail", ...).
670
671* 2.0.1 (October 29, 2010)
672
673  Bug fixes:
674  - Fix a race condition in heap profiling that could cause undefined behavior
675    if "opt.prof_accum" were disabled.
676  - Add missing mutex unlocks for some OOM error paths in the heap profiling
677    code.
678  - Fix a compilation error for non-C99 builds.
679
680* 2.0.0 (October 24, 2010)
681
682  This version focuses on the experimental *allocm() API, and on improved
683  run-time configuration/introspection.  Nonetheless, numerous performance
684  improvements are also included.
685
686  New features:
687  - Implement the experimental {,r,s,d}allocm() API, which provides a superset
688    of the functionality available via malloc(), calloc(), posix_memalign(),
689    realloc(), malloc_usable_size(), and free().  These functions can be used to
690    allocate/reallocate aligned zeroed memory, ask for optional extra memory
691    during reallocation, prevent object movement during reallocation, etc.
692  - Replace JEMALLOC_OPTIONS/JEMALLOC_PROF_PREFIX with MALLOC_CONF, which is
693    more human-readable, and more flexible.  For example:
694      JEMALLOC_OPTIONS=AJP
695    is now:
696      MALLOC_CONF=abort:true,fill:true,stats_print:true
697  - Port to Apple OS X.  Sponsored by Mozilla.
698  - Make it possible for the application to control thread-->arena mappings via
699    the "thread.arena" mallctl.
700  - Add compile-time support for all TLS-related functionality via pthreads TSD.
701    This is mainly of interest for OS X, which does not support TLS, but has a
702    TSD implementation with similar performance.
703  - Override memalign() and valloc() if they are provided by the system.
704  - Add the "arenas.purge" mallctl, which can be used to synchronously purge all
705    dirty unused pages.
706  - Make cumulative heap profiling data optional, so that it is possible to
707    limit the amount of memory consumed by heap profiling data structures.
708  - Add per thread allocation counters that can be accessed via the
709    "thread.allocated" and "thread.deallocated" mallctls.
710
711  Incompatible changes:
712  - Remove JEMALLOC_OPTIONS and malloc_options (see MALLOC_CONF above).
713  - Increase default backtrace depth from 4 to 128 for heap profiling.
714  - Disable interval-based profile dumps by default.
715
716  Bug fixes:
717  - Remove bad assertions in fork handler functions.  These assertions could
718    cause aborts for some combinations of configure settings.
719  - Fix strerror_r() usage to deal with non-standard semantics in GNU libc.
720  - Fix leak context reporting.  This bug tended to cause the number of contexts
721    to be underreported (though the reported number of objects and bytes were
722    correct).
723  - Fix a realloc() bug for large in-place growing reallocation.  This bug could
724    cause memory corruption, but it was hard to trigger.
725  - Fix an allocation bug for small allocations that could be triggered if
726    multiple threads raced to create a new run of backing pages.
727  - Enhance the heap profiler to trigger samples based on usable size, rather
728    than request size.
729  - Fix a heap profiling bug due to sometimes losing track of requested object
730    size for sampled objects.
731
732* 1.0.3 (August 12, 2010)
733
734  Bug fixes:
735  - Fix the libunwind-based implementation of stack backtracing (used for heap
736    profiling).  This bug could cause zero-length backtraces to be reported.
737  - Add a missing mutex unlock in library initialization code.  If multiple
738    threads raced to initialize malloc, some of them could end up permanently
739    blocked.
740
741* 1.0.2 (May 11, 2010)
742
743  Bug fixes:
744  - Fix junk filling of large objects, which could cause memory corruption.
745  - Add MAP_NORESERVE support for chunk mapping, because otherwise virtual
746    memory limits could cause swap file configuration to fail.  Contributed by
747    Jordan DeLong.
748
749* 1.0.1 (April 14, 2010)
750
751  Bug fixes:
752  - Fix compilation when --enable-fill is specified.
753  - Fix threads-related profiling bugs that affected accuracy and caused memory
754    to be leaked during thread exit.
755  - Fix dirty page purging race conditions that could cause crashes.
756  - Fix crash in tcache flushing code during thread destruction.
757
758* 1.0.0 (April 11, 2010)
759
760  This release focuses on speed and run-time introspection.  Numerous
761  algorithmic improvements make this release substantially faster than its
762  predecessors.
763
764  New features:
765  - Implement autoconf-based configuration system.
766  - Add mallctl*(), for the purposes of introspection and run-time
767    configuration.
768  - Make it possible for the application to manually flush a thread's cache, via
769    the "tcache.flush" mallctl.
770  - Base maximum dirty page count on proportion of active memory.
771  - Compute various additional run-time statistics, including per size class
772    statistics for large objects.
773  - Expose malloc_stats_print(), which can be called repeatedly by the
774    application.
775  - Simplify the malloc_message() signature to only take one string argument,
776    and incorporate an opaque data pointer argument for use by the application
777    in combination with malloc_stats_print().
778  - Add support for allocation backed by one or more swap files, and allow the
779    application to disable over-commit if swap files are in use.
780  - Implement allocation profiling and leak checking.
781
782  Removed features:
783  - Remove the dynamic arena rebalancing code, since thread-specific caching
784    reduces its utility.
785
786  Bug fixes:
787  - Modify chunk allocation to work when address space layout randomization
788    (ASLR) is in use.
789  - Fix thread cleanup bugs related to TLS destruction.
790  - Handle 0-size allocation requests in posix_memalign().
791  - Fix a chunk leak.  The leaked chunks were never touched, so this impacted
792    virtual memory usage, but not physical memory usage.
793
794* linux_2008082[78]a (August 27/28, 2008)
795
796  These snapshot releases are the simple result of incorporating Linux-specific
797  support into the FreeBSD malloc sources.
798
799--------------------------------------------------------------------------------
800vim:filetype=text:textwidth=80
801