blob: fb376b268031a41a02683da0cc886db70da15a77 [file] [log] [blame]
Jason Evans379f8472010-10-24 16:18:29 -07001Following are change highlights associated with official releases. Important
Jason Evansbe09b812015-07-07 09:33:22 -07002bug fixes are all mentioned, but some internal enhancements are omitted here for
3brevity. Much more detail can be found in the git revision history:
Jason Evans379f8472010-10-24 16:18:29 -07004
Jason Evansb9ec5c92014-02-25 16:43:51 -08005 https://github.com/jemalloc/jemalloc
6
Jason Evans38e2c8f2015-09-17 10:05:56 -07007* 4.0.2 (XXX)
8
9 Bug fixes:
10 - Fix ixallocx_prof_sample() to never modify nor create sampled small
11 allocations. xallocx() is in general incapable of moving small allocations,
12 so this fix removes buggy code without loss of generality.
Jason Evans4be9c792015-09-17 10:17:55 -070013 - Fix irallocx_prof_sample() to always allocate large regions, even when
14 alignment is non-zero.
Jason Evans3ca0cf62015-09-17 14:47:39 -070015 - Fix prof_alloc_rollback() to read tdata from thread-specific data rather
16 than dereferencing a potentially invalid tctx.
Jason Evans38e2c8f2015-09-17 10:05:56 -070017
Jason Evans1d7540c2015-09-15 15:26:23 -070018* 4.0.1 (September 15, 2015)
19
20 This is a bugfix release that is somewhat high risk due to the amount of
21 refactoring required to address deep xallocx() problems. As a side effect of
22 these fixes, xallocx() now tries harder to partially fulfill requests for
23 optional extra space. Note that a couple of minor heap profiling
24 optimizations are included, but these are better thought of as performance
25 fixes that were integral to disovering most of the other bugs.
26
27 Optimizations:
28 - Avoid a chunk metadata read in arena_prof_tctx_set(), since it is in the
29 fast path when heap profiling is enabled. Additionally, split a special
30 case out into arena_prof_tctx_reset(), which also avoids chunk metadata
31 reads.
32 - Optimize irallocx_prof() to optimistically update the sampler state. The
33 prior implementation appears to have been a holdover from when
34 rallocx()/xallocx() functionality was combined as rallocm().
Jason Evans5ef33a92015-08-19 14:12:05 -070035
36 Bug fixes:
Jason Evans1d7540c2015-09-15 15:26:23 -070037 - Fix TLS configuration such that it is enabled by default for platforms on
38 which it works correctly.
Jason Evans30949da2015-08-25 16:13:59 -070039 - Fix arenas_cache_cleanup() and arena_get_hard() to handle
40 allocation/deallocation within the application's thread-specific data
41 cleanup functions even after arenas_cache is torn down.
Jason Evans1d7540c2015-09-15 15:26:23 -070042 - Fix xallocx() bugs related to size+extra exceeding HUGE_MAXCLASS.
Mike Hommey6d8075f2015-08-27 20:30:15 -070043 - Fix chunk purge hook calls for in-place huge shrinking reallocation to
44 specify the old chunk size rather than the new chunk size. This bug caused
45 no correctness issues for the default chunk purge function, but was
46 visible to custom functions set via the "arena.<i>.chunk_hooks" mallctl.
Jason Evans1d7540c2015-09-15 15:26:23 -070047 - Fix heap profiling bugs:
48 + Fix heap profiling to distinguish among otherwise identical sample sites
49 with interposed resets (triggered via the "prof.reset" mallctl). This bug
50 could cause data structure corruption that would most likely result in a
51 segfault.
52 + Fix irealloc_prof() to prof_alloc_rollback() on OOM.
53 + Make one call to prof_active_get_unlocked() per allocation event, and use
54 the result throughout the relevant functions that handle an allocation
55 event. Also add a missing check in prof_realloc(). These fixes protect
56 allocation events against concurrent prof_active changes.
57 + Fix ixallocx_prof() to pass usize_max and zero to ixallocx_prof_sample()
58 in the correct order.
59 + Fix prof_realloc() to call prof_free_sampled_object() after calling
60 prof_malloc_sample_object(). Prior to this fix, if tctx and old_tctx were
61 the same, the tctx could have been prematurely destroyed.
62 - Fix portability bugs:
63 + Don't bitshift by negative amounts when encoding/decoding run sizes in
64 chunk header maps. This affected systems with page sizes greater than 8
65 KiB.
66 + Rename index_t to szind_t to avoid an existing type on Solaris.
67 + Add JEMALLOC_CXX_THROW to the memalign() function prototype, in order to
68 match glibc and avoid compilation errors when including both
69 jemalloc/jemalloc.h and malloc.h in C++ code.
70 + Don't assume that /bin/sh is appropriate when running size_classes.sh
71 during configuration.
72 + Consider __sparcv9 a synonym for __sparc64__ when defining LG_QUANTUM.
73 + Link tests to librt if it contains clock_gettime(2).
Jason Evans5ef33a92015-08-19 14:12:05 -070074
Jason Evans9b68f672015-08-17 13:21:08 -070075* 4.0.0 (August 17, 2015)
Jason Evans54673fd2015-02-23 22:28:43 -080076
77 This version contains many speed and space optimizations, both minor and
78 major. The major themes are generalization, unification, and simplification.
79 Although many of these optimizations cause no visible behavior change, their
80 cumulative effect is substantial.
81
82 New features:
83 - Normalize size class spacing to be consistent across the complete size
84 range. By default there are four size classes per size doubling, but this
85 is now configurable via the --with-lg-size-class-group option. Also add the
86 --with-lg-page, --with-lg-page-sizes, --with-lg-quantum, and
87 --with-lg-tiny-min options, which can be used to tweak page and size class
88 settings. Impacts:
89 + Worst case performance for incrementally growing/shrinking reallocation
90 is improved because there are far fewer size classes, and therefore
91 copying happens less often.
92 + Internal fragmentation is limited to 20% for all but the smallest size
93 classes (those less than four times the quantum). (1B + 4 KiB)
94 and (1B + 4 MiB) previously suffered nearly 50% internal fragmentation.
95 + Chunk fragmentation tends to be lower because there are fewer distinct run
96 sizes to pack.
97 - Add support for explicit tcaches. The "tcache.create", "tcache.flush", and
98 "tcache.destroy" mallctls control tcache lifetime and flushing, and the
99 MALLOCX_TCACHE(tc) and MALLOCX_TCACHE_NONE flags to the *allocx() API
100 control which tcache is used for each operation.
101 - Implement per thread heap profiling, as well as the ability to
102 enable/disable heap profiling on a per thread basis. Add the "prof.reset",
103 "prof.lg_sample", "thread.prof.name", "thread.prof.active",
104 "opt.prof_thread_active_init", "prof.thread_active_init", and
105 "thread.prof.active" mallctls.
106 - Add support for per arena application-specified chunk allocators, configured
Jason Evansb49a3342015-07-28 11:28:19 -0400107 via the "arena.<i>.chunk_hooks" mallctl.
Jason Evans54673fd2015-02-23 22:28:43 -0800108 - Refactor huge allocation to be managed by arenas, so that arenas now
109 function as general purpose independent allocators. This is important in
110 the context of user-specified chunk allocators, aside from the scalability
111 benefits. Related new statistics:
112 + The "stats.arenas.<i>.huge.allocated", "stats.arenas.<i>.huge.nmalloc",
113 "stats.arenas.<i>.huge.ndalloc", and "stats.arenas.<i>.huge.nrequests"
114 mallctls provide high level per arena huge allocation statistics.
Qinfan Wu89750352015-04-21 16:57:42 -0700115 + The "arenas.nhchunks", "arenas.hchunk.<i>.size",
Jason Evans54673fd2015-02-23 22:28:43 -0800116 "stats.arenas.<i>.hchunks.<j>.nmalloc",
117 "stats.arenas.<i>.hchunks.<j>.ndalloc",
118 "stats.arenas.<i>.hchunks.<j>.nrequests", and
119 "stats.arenas.<i>.hchunks.<j>.curhchunks" mallctls provide per size class
120 statistics.
121 - Add the 'util' column to malloc_stats_print() output, which reports the
122 proportion of available regions that are currently in use for each small
123 size class.
124 - Add "alloc" and "free" modes for for junk filling (see the "opt.junk"
125 mallctl), so that it is possible to separately enable junk filling for
126 allocation versus deallocation.
127 - Add the jemalloc-config script, which provides information about how
128 jemalloc was configured, and how to integrate it into application builds.
129 - Add metadata statistics, which are accessible via the "stats.metadata",
130 "stats.arenas.<i>.metadata.mapped", and
131 "stats.arenas.<i>.metadata.allocated" mallctls.
Jason Evans4acd75a2015-03-23 17:25:57 -0700132 - Add the "stats.resident" mallctl, which reports the upper limit of
133 physically resident memory mapped by the allocator.
Jason Evans562d2662015-03-24 16:36:12 -0700134 - Add per arena control over unused dirty page purging, via the
135 "arenas.lg_dirty_mult", "arena.<i>.lg_dirty_mult", and
136 "stats.arenas.<i>.lg_dirty_mult" mallctls.
Jason Evans54673fd2015-02-23 22:28:43 -0800137 - Add the "prof.gdump" mallctl, which makes it possible to toggle the gdump
138 feature on/off during program execution.
139 - Add sdallocx(), which implements sized deallocation. The primary
140 optimization over dallocx() is the removal of a metadata read, which often
141 suffers an L1 cache miss.
142 - Add missing header includes in jemalloc/jemalloc.h, so that applications
143 only have to #include <jemalloc/jemalloc.h>.
144 - Add support for additional platforms:
145 + Bitrig
146 + Cygwin
147 + DragonFlyBSD
148 + iOS
149 + OpenBSD
150 + OpenRISC/or1k
151
152 Optimizations:
Jason Evans54673fd2015-02-23 22:28:43 -0800153 - Maintain dirty runs in per arena LRUs rather than in per arena trees of
154 dirty-run-containing chunks. In practice this change significantly reduces
155 dirty page purging volume.
156 - Integrate whole chunks into the unused dirty page purging machinery. This
157 reduces the cost of repeated huge allocation/deallocation, because it
158 effectively introduces a cache of chunks.
159 - Split the arena chunk map into two separate arrays, in order to increase
160 cache locality for the frequently accessed bits.
161 - Move small run metadata out of runs, into arena chunk headers. This reduces
162 run fragmentation, smaller runs reduce external fragmentation for small size
163 classes, and packed (less uniformly aligned) metadata layout improves CPU
164 cache set distribution.
Jason Evans8a03cf02015-05-04 09:58:36 -0700165 - Randomly distribute large allocation base pointer alignment relative to page
166 boundaries in order to more uniformly utilize CPU cache sets. This can be
Jason Evansf2bc8522015-07-17 16:38:25 -0700167 disabled via the --disable-cache-oblivious configure option, and queried via
168 the "config.cache_oblivious" mallctl.
Jason Evans54673fd2015-02-23 22:28:43 -0800169 - Micro-optimize the fast paths for the public API functions.
170 - Refactor thread-specific data to reside in a single structure. This assures
171 that only a single TLS read is necessary per call into the public API.
172 - Implement in-place huge allocation growing and shrinking.
173 - Refactor rtree (radix tree for chunk lookups) to be lock-free, and make
174 additional optimizations that reduce maximum lookup depth to one or two
175 levels. This resolves what was a concurrency bottleneck for per arena huge
176 allocation, because a global data structure is critical for determining
177 which arenas own which huge allocations.
178
179 Incompatible changes:
180 - Replace --enable-cc-silence with --disable-cc-silence to suppress spurious
181 warnings by default.
182 - Assure that the constness of malloc_usable_size()'s return type matches that
183 of the system implementation.
184 - Change the heap profile dump format to support per thread heap profiling,
Jason Evans70417202015-05-01 12:31:12 -0700185 rename pprof to jeprof, and enhance it with the --thread=<n> option. As a
186 result, the bundled jeprof must now be used rather than the upstream
187 (gperftools) pprof.
Jason Evans54673fd2015-02-23 22:28:43 -0800188 - Disable "opt.prof_final" by default, in order to avoid atexit(3), which can
189 internally deadlock on some platforms.
190 - Change the "arenas.nlruns" mallctl type from size_t to unsigned.
191 - Replace the "stats.arenas.<i>.bins.<j>.allocated" mallctl with
192 "stats.arenas.<i>.bins.<j>.curregs".
193 - Ignore MALLOC_CONF in set{uid,gid,cap} binaries.
194 - Ignore MALLOCX_ARENA(a) in dallocx(), in favor of using the
195 MALLOCX_TCACHE(tc) and MALLOCX_TCACHE_NONE flags to control tcache usage.
196
197 Removed features:
198 - Remove the *allocm() API, which is superseded by the *allocx() API.
199 - Remove the --enable-dss options, and make dss non-optional on all platforms
200 which support sbrk(2).
201 - Remove the "arenas.purge" mallctl, which was obsoleted by the
202 "arena.<i>.purge" mallctl in 3.1.0.
203 - Remove the unnecessary "opt.valgrind" mallctl; jemalloc automatically
204 detects whether it is running inside Valgrind.
205 - Remove the "stats.huge.allocated", "stats.huge.nmalloc", and
206 "stats.huge.ndalloc" mallctls.
207 - Remove the --enable-mremap option.
Jason Evans54673fd2015-02-23 22:28:43 -0800208 - Remove the "stats.chunks.current", "stats.chunks.total", and
209 "stats.chunks.high" mallctls.
210
211 Bug fixes:
212 - Fix the cactive statistic to decrease (rather than increase) when active
213 memory decreases. This regression was first released in 3.5.0.
214 - Fix OOM handling in memalign() and valloc(). A variant of this bug existed
215 in all releases since 2.0.0, which introduced these functions.
Jason Evans32dca112015-07-09 11:34:13 -0700216 - Fix an OOM-related regression in arena_tcache_fill_small(), which could
217 cause cache corruption on OOM. This regression was present in all releases
218 from 2.2.0 through 3.6.0.
Jason Evans241abc62015-06-23 18:47:07 -0700219 - Fix size class overflow handling for malloc(), posix_memalign(), memalign(),
220 calloc(), and realloc() when profiling is enabled.
Jason Evans54673fd2015-02-23 22:28:43 -0800221 - Fix the "arena.<i>.dss" mallctl to return an error if "primary" or
222 "secondary" precedence is specified, but sbrk(2) is not supported.
223 - Fix fallback lg_floor() implementations to handle extremely large inputs.
224 - Ensure the default purgeable zone is after the default zone on OS X.
225 - Fix latent bugs in atomic_*().
226 - Fix the "arena.<i>.dss" mallctl to handle read-only calls.
227 - Fix tls_model configuration to enable the initial-exec model when possible.
228 - Mark malloc_conf as a weak symbol so that the application can override it.
229 - Correctly detect glibc's adaptive pthread mutexes.
230 - Fix the --without-export configure option.
231
Jason Evansff536312014-03-31 09:23:10 -0700232* 3.6.0 (March 31, 2014)
233
234 This version contains a critical bug fix for a regression present in 3.5.0 and
235 3.5.1.
236
237 Bug fixes:
238 - Fix a regression in arena_chunk_alloc() that caused crashes during
239 small/large allocation if chunk allocation failed. In the absence of this
240 bug, chunk allocation failure would result in allocation failure, e.g. NULL
241 return from malloc(). This regression was introduced in 3.5.0.
242 - Fix backtracing for gcc intrinsics-based backtracing by specifying
243 -fno-omit-frame-pointer to gcc. Note that the application (and all the
244 libraries it links to) must also be compiled with this option for
245 backtracing to be reliable.
246 - Use dss allocation precedence for huge allocations as well as small/large
247 allocations.
Jason Evans54673fd2015-02-23 22:28:43 -0800248 - Fix test assertion failure message formatting. This bug did not manifest on
Jason Evansff536312014-03-31 09:23:10 -0700249 x86_64 systems because of implementation subtleties in va_list.
250 - Fix inconsequential test failures for hash and SFMT code.
251
252 New features:
253 - Support heap profiling on FreeBSD. This feature depends on the proc
254 filesystem being mounted during heap profile dumping.
255
Jason Evansb9ec5c92014-02-25 16:43:51 -0800256* 3.5.1 (February 25, 2014)
257
258 This version primarily addresses minor bugs in test code.
259
260 Bug fixes:
261 - Configure Solaris/Illumos to use MADV_FREE.
262 - Fix junk filling for mremap(2)-based huge reallocation. This is only
263 relevant if configuring with the --enable-mremap option specified.
264 - Avoid compilation failure if 'restrict' C99 keyword is not supported by the
265 compiler.
266 - Add a configure test for SSE2 rather than assuming it is usable on i686
267 systems. This fixes test compilation errors, especially on 32-bit Linux
268 systems.
269 - Fix mallctl argument size mismatches (size_t vs. uint64_t) in the stats unit
270 test.
271 - Fix/remove flawed alignment-related overflow tests.
272 - Prevent compiler optimizations that could change backtraces in the
273 prof_accum unit test.
Jason Evans379f8472010-10-24 16:18:29 -0700274
Jason Evans798a4812014-01-22 11:09:50 -0800275* 3.5.0 (January 22, 2014)
Jason Evans86abd0d2013-11-30 15:25:42 -0800276
Jason Evans264dfd32014-01-17 17:01:23 -0800277 This version focuses on refactoring and automated testing, though it also
278 includes some non-trivial heap profiling optimizations not mentioned below.
279
280 New features:
281 - Add the *allocx() API, which is a successor to the experimental *allocm()
282 API. The *allocx() functions are slightly simpler to use because they have
283 fewer parameters, they directly return the results of primary interest, and
284 mallocx()/rallocx() avoid the strict aliasing pitfall that
Jason Evans9c8baec2014-01-22 13:08:47 -0800285 allocm()/rallocm() share with posix_memalign(). Note that *allocm() is
Jason Evans264dfd32014-01-17 17:01:23 -0800286 slated for removal in the next non-bugfix release.
287 - Add support for LinuxThreads.
288
Jason Evans86abd0d2013-11-30 15:25:42 -0800289 Bug fixes:
Jason Evansd37d5ad2013-12-05 23:01:50 -0800290 - Unless heap profiling is enabled, disable floating point code and don't link
291 with libm. This, in combination with e.g. EXTRA_CFLAGS=-mno-sse on x64
292 systems, makes it possible to completely disable floating point register
293 use. Some versions of glibc neglect to save/restore caller-saved floating
294 point registers during dynamic lazy symbol loading, and the symbol loading
295 code uses whatever malloc the application happens to have linked/loaded
296 with, the result being potential floating point register corruption.
Jason Evans264dfd32014-01-17 17:01:23 -0800297 - Report ENOMEM rather than EINVAL if an OOM occurs during heap profiling
298 backtrace creation in imemalign(). This bug impacted posix_memalign() and
299 aligned_alloc().
300 - Fix a file descriptor leak in a prof_dump_maps() error path.
301 - Fix prof_dump() to close the dump file descriptor for all relevant error
302 paths.
303 - Fix rallocm() to use the arena specified by the ALLOCM_ARENA(s) flag for
304 allocation, not just deallocation.
305 - Fix a data race for large allocation stats counters.
306 - Fix a potential infinite loop during thread exit. This bug occurred on
307 Solaris, and could affect other platforms with similar pthreads TSD
308 implementations.
309 - Don't junk-fill reallocations unless usable size changes. This fixes a
310 violation of the *allocx()/*allocm() semantics.
311 - Fix growing large reallocation to junk fill new space.
312 - Fix huge deallocation to junk fill when munmap is disabled.
Jason Evans86abd0d2013-11-30 15:25:42 -0800313 - Change the default private namespace prefix from empty to je_, and change
314 --with-private-namespace-prefix so that it prepends an additional prefix
315 rather than replacing je_. This reduces the likelihood of applications
316 which statically link jemalloc experiencing symbol name collisions.
Jason Evans264dfd32014-01-17 17:01:23 -0800317 - Add missing private namespace mangling (relevant when
318 --with-private-namespace is specified).
319 - Add and use JEMALLOC_INLINE_C so that static inline functions are marked as
320 static even for debug builds.
321 - Add a missing mutex unlock in a malloc_init_hard() error path. In practice
322 this error path is never executed.
323 - Fix numerous bugs in malloc_strotumax() error handling/reporting. These
324 bugs had no impact except for malformed inputs.
325 - Fix numerous bugs in malloc_snprintf(). These bugs were not exercised by
326 existing calls, so they had no impact.
Jason Evans86abd0d2013-11-30 15:25:42 -0800327
Jason Evans0f7ba3f2013-10-20 19:40:09 -0700328* 3.4.1 (October 20, 2013)
Jason Evansff08ef72013-10-19 21:41:10 -0700329
330 Bug fixes:
Jason Evans7b651802013-10-20 14:09:54 -0700331 - Fix a race in the "arenas.extend" mallctl that could cause memory corruption
332 of internal data structures and subsequent crashes.
Jason Evansdda90f52013-10-19 23:48:40 -0700333 - Fix Valgrind integration flaws that caused Valgrind warnings about reads of
334 uninitialized memory in:
335 + arena chunk headers
336 + internal zero-initialized data structures (relevant to tcache and prof
337 code)
Jason Evansff08ef72013-10-19 21:41:10 -0700338 - Preserve errno during the first allocation. A readlink(2) call during
339 initialization fails unless /etc/malloc.conf exists, so errno was typically
340 set during the first allocation prior to this fix.
341 - Fix compilation warnings reported by gcc 4.8.1.
342
Jason Evans765cc2b2013-06-02 20:58:00 -0700343* 3.4.0 (June 2, 2013)
344
345 This version is essentially a small bugfix release, but the addition of
346 aarch64 support requires that the minor version be incremented.
347
348 Bug fixes:
349 - Fix race-triggered deadlocks in chunk_record(). These deadlocks were
350 typically triggered by multiple threads concurrently deallocating huge
351 objects.
352
353 New features:
354 - Add support for the aarch64 architecture.
355
Jason Evans22988352013-03-06 11:11:17 -0800356* 3.3.1 (March 6, 2013)
357
358 This version fixes bugs that are typically encountered only when utilizing
359 custom run-time options.
Jason Evansbbe29d32013-01-30 15:03:11 -0800360
361 Bug fixes:
Jason Evans88c222c2013-02-06 11:59:30 -0800362 - Fix a locking order bug that could cause deadlock during fork if heap
363 profiling were enabled.
Jason Evansa7a28c32013-01-31 16:53:58 -0800364 - Fix a chunk recycling bug that could cause the allocator to lose track of
Jason Evans765cc2b2013-06-02 20:58:00 -0700365 whether a chunk was zeroed. On FreeBSD, NetBSD, and OS X, it could cause
Jason Evansa7a28c32013-01-31 16:53:58 -0800366 corruption if allocating via sbrk(2) (unlikely unless running with the
367 "dss:primary" option specified). This was completely harmless on Linux
368 unless using mlockall(2) (and unlikely even then, unless the
369 --disable-munmap configure option or the "dss:primary" option was
370 specified). This regression was introduced in 3.1.0 by the
371 mlockall(2)/madvise(2) interaction fix.
Jason Evansbbe29d32013-01-30 15:03:11 -0800372 - Fix TLS-related memory corruption that could occur during thread exit if the
373 thread never allocated memory. Only the quarantine and prof facilities were
374 susceptible.
Jason Evansd0e942e2013-01-31 14:42:41 -0800375 - Fix two quarantine bugs:
376 + Internal reallocation of the quarantined object array leaked the old
377 array.
378 + Reallocation failure for internal reallocation of the quarantined object
379 array (very unlikely) resulted in memory corruption.
Jason Evans06912752013-01-31 17:02:53 -0800380 - Fix Valgrind integration to annotate all internally allocated memory in a
381 way that keeps Valgrind happy about internal data structure access.
Jason Evans22988352013-03-06 11:11:17 -0800382 - Fix building for s390 systems.
Jason Evansbbe29d32013-01-30 15:03:11 -0800383
Jason Evansb5681fb2013-01-22 22:45:09 -0800384* 3.3.0 (January 23, 2013)
385
386 This version includes a few minor performance improvements in addition to the
387 listed new features and bug fixes.
388
389 New features:
390 - Add clipping support to lg_chunk option processing.
391 - Add the --enable-ivsalloc option.
392 - Add the --without-export option.
393 - Add the --disable-zone-allocator option.
Jason Evans6eb84fb2012-11-29 22:13:04 -0800394
395 Bug fixes:
396 - Fix "arenas.extend" mallctl to output the number of arenas.
Jason Evansa33488d2013-10-03 14:38:39 -0700397 - Fix chunk_recycle() to unconditionally inform Valgrind that returned memory
Jason Evans12711852012-12-12 10:12:18 -0800398 is undefined.
Jason Evansb5681fb2013-01-22 22:45:09 -0800399 - Fix build break on FreeBSD related to alloca.h.
Jason Evans6eb84fb2012-11-29 22:13:04 -0800400
Jason Evans556ddc72012-11-07 15:16:29 -0800401* 3.2.0 (November 9, 2012)
402
403 In addition to a couple of bug fixes, this version modifies page run
404 allocation and dirty page purging algorithms in order to better control
405 page-level virtual memory fragmentation.
Jason Evans12efefb2012-10-16 22:06:56 -0700406
Jason Evanse3d13062012-10-30 15:42:37 -0700407 Incompatible changes:
Jason Evans556ddc72012-11-07 15:16:29 -0800408 - Change the "opt.lg_dirty_mult" default from 5 to 3 (32:1 to 8:1).
Jason Evanse3d13062012-10-30 15:42:37 -0700409
Jason Evans12efefb2012-10-16 22:06:56 -0700410 Bug fixes:
411 - Fix dss/mmap allocation precedence code to use recyclable mmap memory only
412 after primary dss allocation fails.
Jason Evanse3d13062012-10-30 15:42:37 -0700413 - Fix deadlock in the "arenas.purge" mallctl. This regression was introduced
414 in 3.1.0 by the addition of the "arena.<i>.purge" mallctl.
Jason Evans12efefb2012-10-16 22:06:56 -0700415
Jason Evans2b592b02012-10-16 10:12:40 -0700416* 3.1.0 (October 16, 2012)
Jason Evans3860eac2012-05-15 13:53:21 -0700417
Jason Evans781fe752012-05-15 14:48:14 -0700418 New features:
419 - Auto-detect whether running inside Valgrind, thus removing the need to
420 manually specify MALLOC_CONF=valgrind:true.
Jason Evans2b592b02012-10-16 10:12:40 -0700421 - Add the "arenas.extend" mallctl, which allows applications to create
422 manually managed arenas.
423 - Add the ALLOCM_ARENA() flag for {,r,d}allocm().
424 - Add the "opt.dss", "arena.<i>.dss", and "stats.arenas.<i>.dss" mallctls,
425 which provide control over dss/mmap precedence.
426 - Add the "arena.<i>.purge" mallctl, which obsoletes "arenas.purge".
427 - Define LG_QUANTUM for hppa.
Jason Evans781fe752012-05-15 14:48:14 -0700428
Jason Evans174b70e2012-05-15 23:31:53 -0700429 Incompatible changes:
430 - Disable tcache by default if running inside Valgrind, in order to avoid
431 making unallocated objects appear reachable to Valgrind.
Jason Evans2b592b02012-10-16 10:12:40 -0700432 - Drop const from malloc_usable_size() argument on Linux.
Jason Evans174b70e2012-05-15 23:31:53 -0700433
Jason Evans3860eac2012-05-15 13:53:21 -0700434 Bug fixes:
435 - Fix heap profiling crash if sampled object is freed via realloc(p, 0).
Jason Evans5c710ce2012-05-23 16:09:22 -0700436 - Remove const from __*_hook variable declarations, so that glibc can modify
437 them during process forking.
Jason Evans2b592b02012-10-16 10:12:40 -0700438 - Fix mlockall(2)/madvise(2) interaction.
439 - Fix fork(2)-related deadlocks.
440 - Fix error return value for "thread.tcache.enabled" mallctl.
Jason Evans3860eac2012-05-15 13:53:21 -0700441
Jason Evanscbb71ca2012-05-11 17:00:20 -0700442* 3.0.0 (May 11, 2012)
Jason Evans9ef7f5d2012-04-16 18:16:48 -0700443
444 Although this version adds some major new features, the primary focus is on
445 internal code cleanup that facilitates maintainability and portability, most
446 of which is not reflected in the ChangeLog. This is the first release to
447 incorporate substantial contributions from numerous other developers, and the
448 result is a more broadly useful allocator (see the git revision history for
449 contribution details). Note that the license has been unified, thanks to
450 Facebook granting a license under the same terms as the other copyright
451 holders (see COPYING).
452
453 New features:
454 - Implement Valgrind support, redzones, and quarantine.
Jason Evans079687b2012-04-23 12:49:23 -0700455 - Add support for additional platforms:
Jason Evans9ef7f5d2012-04-16 18:16:48 -0700456 + FreeBSD
457 + Mac OS X Lion
Jason Evans40f514f2012-04-22 16:21:06 -0700458 + MinGW
Jason Evanscbb71ca2012-05-11 17:00:20 -0700459 + Windows (no support yet for replacing the system malloc)
Jason Evans9ef7f5d2012-04-16 18:16:48 -0700460 - Add support for additional architectures:
461 + MIPS
462 + SH4
463 + Tilera
464 - Add support for cross compiling.
465 - Add nallocm(), which rounds a request size up to the nearest size class
466 without actually allocating.
467 - Implement aligned_alloc() (blame C11).
Jason Evans9ef7f5d2012-04-16 18:16:48 -0700468 - Add the "thread.tcache.enabled" mallctl.
Jason Evans0b25fe72012-04-17 16:39:33 -0700469 - Add the "opt.prof_final" mallctl.
Jason Evans25a000e2012-04-17 15:49:30 -0700470 - Update pprof (from gperftools 2.0).
Jason Evanscbb71ca2012-05-11 17:00:20 -0700471 - Add the --with-mangling option.
472 - Add the --disable-experimental option.
473 - Add the --disable-munmap option, and make it the default on Linux.
474 - Add the --enable-mremap option, which disables use of mremap(2) by default.
Jason Evans9ef7f5d2012-04-16 18:16:48 -0700475
476 Incompatible changes:
477 - Enable stats by default.
478 - Enable fill by default.
479 - Disable lazy locking by default.
480 - Rename the "tcache.flush" mallctl to "thread.tcache.flush".
481 - Rename the "arenas.pagesize" mallctl to "arenas.page".
Jason Evans0b25fe72012-04-17 16:39:33 -0700482 - Change the "opt.lg_prof_sample" default from 0 to 19 (1 B to 512 KiB).
483 - Change the "opt.prof_accum" default from true to false.
Jason Evans9ef7f5d2012-04-16 18:16:48 -0700484
485 Removed features:
486 - Remove the swap feature, including the "config.swap", "swap.avail",
487 "swap.prezeroed", "swap.nfds", and "swap.fds" mallctls.
488 - Remove highruns statistics, including the
489 "stats.arenas.<i>.bins.<j>.highruns" and
490 "stats.arenas.<i>.lruns.<j>.highruns" mallctls.
491 - As part of small size class refactoring, remove the "opt.lg_[qc]space_max",
492 "arenas.cacheline", "arenas.subpage", "arenas.[tqcs]space_{min,max}", and
493 "arenas.[tqcs]bins" mallctls.
494 - Remove the "arenas.chunksize" mallctl.
495 - Remove the "opt.lg_prof_tcmax" option.
496 - Remove the "opt.lg_prof_bt_max" option.
497 - Remove the "opt.lg_tcache_gc_sweep" option.
498 - Remove the --disable-tiny option, including the "config.tiny" mallctl.
499 - Remove the --enable-dynamic-page-shift configure option.
500 - Remove the --enable-sysv configure option.
501
502 Bug fixes:
Jason Evans9ef7f5d2012-04-16 18:16:48 -0700503 - Fix a statistics-related bug in the "thread.arena" mallctl that could cause
504 invalid statistics and crashes.
Jason Evans079687b2012-04-23 12:49:23 -0700505 - Work around TLS deallocation via free() on Linux. This bug could cause
Jason Evans9ef7f5d2012-04-16 18:16:48 -0700506 write-after-free memory corruption.
Jason Evans52386b22012-04-22 16:00:11 -0700507 - Fix a potential deadlock that could occur during interval- and
508 growth-triggered heap profile dumps.
Jason Evansd8ceef62012-05-10 20:59:39 -0700509 - Fix large calloc() zeroing bugs due to dropping chunk map unzeroed flags.
Jason Evans8f0e0eb2012-04-21 13:33:48 -0700510 - Fix chunk_alloc_dss() to stop claiming memory is zeroed. This bug could
511 cause memory corruption and crashes with --enable-dss specified.
Jason Evans52386b22012-04-22 16:00:11 -0700512 - Fix fork-related bugs that could cause deadlock in children between fork
513 and exec.
Jason Evans9ef7f5d2012-04-16 18:16:48 -0700514 - Fix malloc_stats_print() to honor 'b' and 'l' in the opts parameter.
515 - Fix realloc(p, 0) to act like free(p).
516 - Do not enforce minimum alignment in memalign().
517 - Check for NULL pointer in malloc_usable_size().
Jason Evans52386b22012-04-22 16:00:11 -0700518 - Fix an off-by-one heap profile statistics bug that could be observed in
519 interval- and growth-triggered heap profiles.
Jason Evans6b9ed672012-04-25 13:12:46 -0700520 - Fix the "epoch" mallctl to update cached stats even if the passed in epoch
521 is 0.
Jason Evans9ef7f5d2012-04-16 18:16:48 -0700522 - Fix bin->runcur management to fix a layout policy bug. This bug did not
523 affect correctness.
524 - Fix a bug in choose_arena_hard() that potentially caused more arenas to be
525 initialized than necessary.
526 - Add missing "opt.lg_tcache_max" mallctl implementation.
527 - Use glibc allocator hooks to make mixed allocator usage less likely.
528 - Fix build issues for --disable-tcache.
Jason Evans918d6e22012-04-20 13:42:21 -0700529 - Don't mangle pthread_create() when --with-private-namespace is specified.
Jason Evans9ef7f5d2012-04-16 18:16:48 -0700530
Jason Evansb3bd8852011-11-14 17:12:45 -0800531* 2.2.5 (November 14, 2011)
532
533 Bug fixes:
534 - Fix huge_ralloc() race when using mremap(2). This is a serious bug that
535 could cause memory corruption and/or crashes.
536 - Fix huge_ralloc() to maintain chunk statistics.
537 - Fix malloc_stats_print(..., "a") output.
538
Jason Evansca9ee1a2011-11-05 21:46:23 -0700539* 2.2.4 (November 5, 2011)
540
541 Bug fixes:
542 - Initialize arenas_tsd before using it. This bug existed for 2.2.[0-3], as
543 well as for --disable-tls builds in earlier releases.
544 - Do not assume a 4 KiB page size in test/rallocm.c.
545
Jason Evansc67e4fd2011-08-31 15:19:13 -0700546* 2.2.3 (August 31, 2011)
547
548 This version fixes numerous bugs related to heap profiling.
549
550 Bug fixes:
551 - Fix a prof-related race condition. This bug could cause memory corruption,
552 but only occurred in non-default configurations (prof_accum:false).
553 - Fix off-by-one backtracing issues (make sure that prof_alloc_prep() is
554 excluded from backtraces).
555 - Fix a prof-related bug in realloc() (only triggered by OOM errors).
556 - Fix prof-related bugs in allocm() and rallocm().
557 - Fix prof_tdata_cleanup() for --disable-tls builds.
558 - Fix a relative include path, to fix objdir builds.
559
Jason Evans4c484812011-07-30 16:59:13 -0700560* 2.2.2 (July 30, 2011)
561
562 Bug fixes:
563 - Fix a build error for --disable-tcache.
564 - Fix assertions in arena_purge() (for real this time).
565 - Add the --with-private-namespace option. This is a workaround for symbol
566 conflicts that can inadvertently arise when using static libraries.
567
Jason Evans7d9ebea2011-03-30 15:01:08 -0700568* 2.2.1 (March 30, 2011)
569
570 Bug fixes:
571 - Implement atomic operations for x86/x64. This fixes compilation failures
572 for versions of gcc that are still in wide use.
573 - Fix an assertion in arena_purge().
574
Jason Evans4bcd9872011-03-22 15:30:22 -0700575* 2.2.0 (March 22, 2011)
576
577 This version incorporates several improvements to algorithms and data
578 structures that tend to reduce fragmentation and increase speed.
579
580 New features:
581 - Add the "stats.cactive" mallctl.
582 - Update pprof (from google-perftools 1.7).
583 - Improve backtracing-related configuration logic, and add the
584 --disable-prof-libgcc option.
585
586 Bug fixes:
587 - Change default symbol visibility from "internal", to "hidden", which
588 decreases the overhead of library-internal function calls.
589 - Fix symbol visibility so that it is also set on OS X.
590 - Fix a build dependency regression caused by the introduction of the .pic.o
591 suffix for PIC object files.
592 - Add missing checks for mutex initialization failures.
593 - Don't use libgcc-based backtracing except on x64, where it is known to work.
594 - Fix deadlocks on OS X that were due to memory allocation in
595 pthread_mutex_lock().
596 - Heap profiling-specific fixes:
597 + Fix memory corruption due to integer overflow in small region index
598 computation, when using a small enough sample interval that profiling
599 context pointers are stored in small run headers.
600 + Fix a bootstrap ordering bug that only occurred with TLS disabled.
601 + Fix a rallocm() rsize bug.
602 + Fix error detection bugs for aligned memory allocation.
603
Jason Evans0e4d0d12011-03-14 16:41:03 -0700604* 2.1.3 (March 14, 2011)
605
606 Bug fixes:
607 - Fix a cpp logic regression (due to the "thread.{de,}allocatedp" mallctl fix
608 for OS X in 2.1.2).
609 - Fix a "thread.arena" mallctl bug.
610 - Fix a thread cache stats merging bug.
611
je6e56e5e2011-03-02 11:23:41 -0800612* 2.1.2 (March 2, 2011)
613
614 Bug fixes:
615 - Fix "thread.{de,}allocatedp" mallctl for OS X.
616 - Add missing jemalloc.a to build system.
617
Jason Evans63692862011-02-07 22:48:35 -0800618* 2.1.1 (January 31, 2011)
Jason Evansada55b22011-01-31 20:08:56 -0800619
Jason Evans63692862011-02-07 22:48:35 -0800620 Bug fixes:
Jason Evansada55b22011-01-31 20:08:56 -0800621 - Fix aligned huge reallocation (affected allocm()).
622 - Fix the ALLOCM_LG_ALIGN macro definition.
623 - Fix a heap dumping deadlock.
624 - Fix a "thread.arena" mallctl bug.
625
Jason Evans63692862011-02-07 22:48:35 -0800626* 2.1.0 (December 3, 2010)
Jason Evans0e8d3d22010-12-03 17:02:16 -0800627
628 This version incorporates some optimizations that can't quite be considered
629 bug fixes.
630
631 New features:
632 - Use Linux's mremap(2) for huge object reallocation when possible.
633 - Avoid locking in mallctl*() when possible.
634 - Add the "thread.[de]allocatedp" mallctl's.
635 - Convert the manual page source from roff to DocBook, and generate both roff
636 and HTML manuals.
637
638 Bug fixes:
639 - Fix a crash due to incorrect bootstrap ordering. This only impacted
640 --enable-debug --enable-dss configurations.
641 - Fix a minor statistics bug for mallctl("swap.avail", ...).
642
Jason Evans63692862011-02-07 22:48:35 -0800643* 2.0.1 (October 29, 2010)
Jason Evans53806fe2010-10-29 20:16:39 -0700644
645 Bug fixes:
646 - Fix a race condition in heap profiling that could cause undefined behavior
Jason Evansada55b22011-01-31 20:08:56 -0800647 if "opt.prof_accum" were disabled.
Jason Evans53806fe2010-10-29 20:16:39 -0700648 - Add missing mutex unlocks for some OOM error paths in the heap profiling
649 code.
650 - Fix a compilation error for non-C99 builds.
651
Jason Evans63692862011-02-07 22:48:35 -0800652* 2.0.0 (October 24, 2010)
Jason Evans379f8472010-10-24 16:18:29 -0700653
654 This version focuses on the experimental *allocm() API, and on improved
655 run-time configuration/introspection. Nonetheless, numerous performance
656 improvements are also included.
657
658 New features:
Jason Evansb059a532010-10-24 16:54:40 -0700659 - Implement the experimental {,r,s,d}allocm() API, which provides a superset
660 of the functionality available via malloc(), calloc(), posix_memalign(),
661 realloc(), malloc_usable_size(), and free(). These functions can be used to
662 allocate/reallocate aligned zeroed memory, ask for optional extra memory
663 during reallocation, prevent object movement during reallocation, etc.
664 - Replace JEMALLOC_OPTIONS/JEMALLOC_PROF_PREFIX with MALLOC_CONF, which is
665 more human-readable, and more flexible. For example:
666 JEMALLOC_OPTIONS=AJP
667 is now:
668 MALLOC_CONF=abort:true,fill:true,stats_print:true
669 - Port to Apple OS X. Sponsored by Mozilla.
670 - Make it possible for the application to control thread-->arena mappings via
671 the "thread.arena" mallctl.
672 - Add compile-time support for all TLS-related functionality via pthreads TSD.
673 This is mainly of interest for OS X, which does not support TLS, but has a
674 TSD implementation with similar performance.
675 - Override memalign() and valloc() if they are provided by the system.
676 - Add the "arenas.purge" mallctl, which can be used to synchronously purge all
677 dirty unused pages.
678 - Make cumulative heap profiling data optional, so that it is possible to
679 limit the amount of memory consumed by heap profiling data structures.
680 - Add per thread allocation counters that can be accessed via the
681 "thread.allocated" and "thread.deallocated" mallctls.
Jason Evans379f8472010-10-24 16:18:29 -0700682
683 Incompatible changes:
Jason Evansb059a532010-10-24 16:54:40 -0700684 - Remove JEMALLOC_OPTIONS and malloc_options (see MALLOC_CONF above).
685 - Increase default backtrace depth from 4 to 128 for heap profiling.
686 - Disable interval-based profile dumps by default.
Jason Evans379f8472010-10-24 16:18:29 -0700687
688 Bug fixes:
689 - Remove bad assertions in fork handler functions. These assertions could
690 cause aborts for some combinations of configure settings.
691 - Fix strerror_r() usage to deal with non-standard semantics in GNU libc.
692 - Fix leak context reporting. This bug tended to cause the number of contexts
693 to be underreported (though the reported number of objects and bytes were
694 correct).
695 - Fix a realloc() bug for large in-place growing reallocation. This bug could
696 cause memory corruption, but it was hard to trigger.
697 - Fix an allocation bug for small allocations that could be triggered if
698 multiple threads raced to create a new run of backing pages.
699 - Enhance the heap profiler to trigger samples based on usable size, rather
700 than request size.
701 - Fix a heap profiling bug due to sometimes losing track of requested object
702 size for sampled objects.
703
Jason Evans63692862011-02-07 22:48:35 -0800704* 1.0.3 (August 12, 2010)
Jason Evans379f8472010-10-24 16:18:29 -0700705
706 Bug fixes:
707 - Fix the libunwind-based implementation of stack backtracing (used for heap
708 profiling). This bug could cause zero-length backtraces to be reported.
709 - Add a missing mutex unlock in library initialization code. If multiple
710 threads raced to initialize malloc, some of them could end up permanently
711 blocked.
712
Jason Evans63692862011-02-07 22:48:35 -0800713* 1.0.2 (May 11, 2010)
Jason Evans379f8472010-10-24 16:18:29 -0700714
715 Bug fixes:
716 - Fix junk filling of large objects, which could cause memory corruption.
717 - Add MAP_NORESERVE support for chunk mapping, because otherwise virtual
718 memory limits could cause swap file configuration to fail. Contributed by
719 Jordan DeLong.
720
Jason Evans63692862011-02-07 22:48:35 -0800721* 1.0.1 (April 14, 2010)
Jason Evans379f8472010-10-24 16:18:29 -0700722
723 Bug fixes:
724 - Fix compilation when --enable-fill is specified.
725 - Fix threads-related profiling bugs that affected accuracy and caused memory
726 to be leaked during thread exit.
727 - Fix dirty page purging race conditions that could cause crashes.
728 - Fix crash in tcache flushing code during thread destruction.
729
Jason Evans63692862011-02-07 22:48:35 -0800730* 1.0.0 (April 11, 2010)
Jason Evans379f8472010-10-24 16:18:29 -0700731
732 This release focuses on speed and run-time introspection. Numerous
733 algorithmic improvements make this release substantially faster than its
734 predecessors.
735
736 New features:
737 - Implement autoconf-based configuration system.
738 - Add mallctl*(), for the purposes of introspection and run-time
739 configuration.
740 - Make it possible for the application to manually flush a thread's cache, via
741 the "tcache.flush" mallctl.
742 - Base maximum dirty page count on proportion of active memory.
charsyamad6800f2015-07-04 01:06:06 +0900743 - Compute various additional run-time statistics, including per size class
Jason Evans379f8472010-10-24 16:18:29 -0700744 statistics for large objects.
745 - Expose malloc_stats_print(), which can be called repeatedly by the
746 application.
747 - Simplify the malloc_message() signature to only take one string argument,
748 and incorporate an opaque data pointer argument for use by the application
749 in combination with malloc_stats_print().
750 - Add support for allocation backed by one or more swap files, and allow the
751 application to disable over-commit if swap files are in use.
752 - Implement allocation profiling and leak checking.
753
754 Removed features:
755 - Remove the dynamic arena rebalancing code, since thread-specific caching
756 reduces its utility.
757
758 Bug fixes:
759 - Modify chunk allocation to work when address space layout randomization
760 (ASLR) is in use.
761 - Fix thread cleanup bugs related to TLS destruction.
762 - Handle 0-size allocation requests in posix_memalign().
763 - Fix a chunk leak. The leaked chunks were never touched, so this impacted
764 virtual memory usage, but not physical memory usage.
765
Jason Evans63692862011-02-07 22:48:35 -0800766* linux_2008082[78]a (August 27/28, 2008)
Jason Evans379f8472010-10-24 16:18:29 -0700767
768 These snapshot releases are the simple result of incorporating Linux-specific
769 support into the FreeBSD malloc sources.
770
771--------------------------------------------------------------------------------
772vim:filetype=text:textwidth=80