xref: /freebsd/contrib/jemalloc/ChangeLog (revision af6a5351a1fdb1130f18be6c782c4d48916eb971)
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* 4.4.0 (December 3, 2016)
8
9  New features:
10  - Add configure support for *-*-linux-android.  (@cferris1000, @jasone)
11  - Add the --disable-syscall configure option, for use on systems that place
12    security-motivated limitations on syscall(2).  (@jasone)
13  - Add support for Debian GNU/kFreeBSD.  (@thesam)
14
15  Optimizations:
16  - Add extent serial numbers and use them where appropriate as a sort key that
17    is higher priority than address, so that the allocation policy prefers older
18    extents.  This tends to improve locality (decrease fragmentation) when
19    memory grows downward.  (@jasone)
20  - Refactor madvise(2) configuration so that MADV_FREE is detected and utilized
21    on Linux 4.5 and newer.  (@jasone)
22  - Mark partially purged arena chunks as non-huge-page.  This improves
23    interaction with Linux's transparent huge page functionality.  (@jasone)
24
25  Bug fixes:
26  - Fix size class computations for edge conditions involving extremely large
27    allocations.  This regression was first released in 4.0.0.  (@jasone,
28    @ingvarha)
29  - Remove overly restrictive assertions related to the cactive statistic.  This
30    regression was first released in 4.1.0.  (@jasone)
31  - Implement a more reliable detection scheme for os_unfair_lock on macOS.
32    (@jszakmeister)
33
34* 4.3.1 (November 7, 2016)
35
36  Bug fixes:
37  - Fix a severe virtual memory leak.  This regression was first released in
38    4.3.0.  (@interwq, @jasone)
39  - Refactor atomic and prng APIs to restore support for 32-bit platforms that
40    use pre-C11 toolchains, e.g. FreeBSD's mips.  (@jasone)
41
42* 4.3.0 (November 4, 2016)
43
44  This is the first release that passes the test suite for multiple Windows
45  configurations, thanks in large part to @glandium setting up continuous
46  integration via AppVeyor (and Travis CI for Linux and OS X).
47
48  New features:
49  - Add "J" (JSON) support to malloc_stats_print().  (@jasone)
50  - Add Cray compiler support.  (@ronawho)
51
52  Optimizations:
53  - Add/use adaptive spinning for bootstrapping and radix tree node
54    initialization.  (@jasone)
55
56  Bug fixes:
57  - Fix large allocation to search starting in the optimal size class heap,
58    which can substantially reduce virtual memory churn and fragmentation.  This
59    regression was first released in 4.0.0.  (@mjp41, @jasone)
60  - Fix stats.arenas.<i>.nthreads accounting.  (@interwq)
61  - Fix and simplify decay-based purging.  (@jasone)
62  - Make DSS (sbrk(2)-related) operations lockless, which resolves potential
63    deadlocks during thread exit.  (@jasone)
64  - Fix over-sized allocation of radix tree leaf nodes.  (@mjp41, @ogaun,
65    @jasone)
66  - Fix over-sized allocation of arena_t (plus associated stats) data
67    structures.  (@jasone, @interwq)
68  - Fix EXTRA_CFLAGS to not affect configuration.  (@jasone)
69  - Fix a Valgrind integration bug.  (@ronawho)
70  - Disallow 0x5a junk filling when running in Valgrind.  (@jasone)
71  - Fix a file descriptor leak on Linux.  This regression was first released in
72    4.2.0.  (@vsarunas, @jasone)
73  - Fix static linking of jemalloc with glibc.  (@djwatson)
74  - Use syscall(2) rather than {open,read,close}(2) during boot on Linux.  This
75    works around other libraries' system call wrappers performing reentrant
76    allocation.  (@kspinka, @Whissi, @jasone)
77  - Fix OS X default zone replacement to work with OS X 10.12.  (@glandium,
78    @jasone)
79  - Fix cached memory management to avoid needless commit/decommit operations
80    during purging, which resolves permanent virtual memory map fragmentation
81    issues on Windows.  (@mjp41, @jasone)
82  - Fix TSD fetches to avoid (recursive) allocation.  This is relevant to
83    non-TLS and Windows configurations.  (@jasone)
84  - Fix malloc_conf overriding to work on Windows.  (@jasone)
85  - Forcibly disable lazy-lock on Windows (was forcibly *enabled*).  (@jasone)
86
87* 4.2.1 (June 8, 2016)
88
89  Bug fixes:
90  - Fix bootstrapping issues for configurations that require allocation during
91    tsd initialization (e.g. --disable-tls).  (@cferris1000, @jasone)
92  - Fix gettimeofday() version of nstime_update().  (@ronawho)
93  - Fix Valgrind regressions in calloc() and chunk_alloc_wrapper().  (@ronawho)
94  - Fix potential VM map fragmentation regression.  (@jasone)
95  - Fix opt_zero-triggered in-place huge reallocation zeroing.  (@jasone)
96  - Fix heap profiling context leaks in reallocation edge cases.  (@jasone)
97
98* 4.2.0 (May 12, 2016)
99
100  New features:
101  - Add the arena.<i>.reset mallctl, which makes it possible to discard all of
102    an arena's allocations in a single operation.  (@jasone)
103  - Add the stats.retained and stats.arenas.<i>.retained statistics.  (@jasone)
104  - Add the --with-version configure option.  (@jasone)
105  - Support --with-lg-page values larger than actual page size.  (@jasone)
106
107  Optimizations:
108  - Use pairing heaps rather than red-black trees for various hot data
109    structures.  (@djwatson, @jasone)
110  - Streamline fast paths of rtree operations.  (@jasone)
111  - Optimize the fast paths of calloc() and [m,d,sd]allocx().  (@jasone)
112  - Decommit unused virtual memory if the OS does not overcommit.  (@jasone)
113  - Specify MAP_NORESERVE on Linux if [heuristic] overcommit is active, in order
114    to avoid unfortunate interactions during fork(2).  (@jasone)
115
116  Bug fixes:
117  - Fix chunk accounting related to triggering gdump profiles.  (@jasone)
118  - Link against librt for clock_gettime(2) if glibc < 2.17.  (@jasone)
119  - Scale leak report summary according to sampling probability.  (@jasone)
120
121* 4.1.1 (May 3, 2016)
122
123  This bugfix release resolves a variety of mostly minor issues, though the
124  bitmap fix is critical for 64-bit Windows.
125
126  Bug fixes:
127  - Fix the linear scan version of bitmap_sfu() to shift by the proper amount
128    even when sizeof(long) is not the same as sizeof(void *), as on 64-bit
129    Windows.  (@jasone)
130  - Fix hashing functions to avoid unaligned memory accesses (and resulting
131    crashes).  This is relevant at least to some ARM-based platforms.
132    (@rkmisra)
133  - Fix fork()-related lock rank ordering reversals.  These reversals were
134    unlikely to cause deadlocks in practice except when heap profiling was
135    enabled and active.  (@jasone)
136  - Fix various chunk leaks in OOM code paths.  (@jasone)
137  - Fix malloc_stats_print() to print opt.narenas correctly.  (@jasone)
138  - Fix MSVC-specific build/test issues.  (@rustyx, @yuslepukhin)
139  - Fix a variety of test failures that were due to test fragility rather than
140    core bugs.  (@jasone)
141
142* 4.1.0 (February 28, 2016)
143
144  This release is primarily about optimizations, but it also incorporates a lot
145  of portability-motivated refactoring and enhancements.  Many people worked on
146  this release, to an extent that even with the omission here of minor changes
147  (see git revision history), and of the people who reported and diagnosed
148  issues, so much of the work was contributed that starting with this release,
149  changes are annotated with author credits to help reflect the collaborative
150  effort involved.
151
152  New features:
153  - Implement decay-based unused dirty page purging, a major optimization with
154    mallctl API impact.  This is an alternative to the existing ratio-based
155    unused dirty page purging, and is intended to eventually become the sole
156    purging mechanism.  New mallctls:
157    + opt.purge
158    + opt.decay_time
159    + arena.<i>.decay
160    + arena.<i>.decay_time
161    + arenas.decay_time
162    + stats.arenas.<i>.decay_time
163    (@jasone, @cevans87)
164  - Add --with-malloc-conf, which makes it possible to embed a default
165    options string during configuration.  This was motivated by the desire to
166    specify --with-malloc-conf=purge:decay , since the default must remain
167    purge:ratio until the 5.0.0 release.  (@jasone)
168  - Add MS Visual Studio 2015 support.  (@rustyx, @yuslepukhin)
169  - Make *allocx() size class overflow behavior defined.  The maximum
170    size class is now less than PTRDIFF_MAX to protect applications against
171    numerical overflow, and all allocation functions are guaranteed to indicate
172    errors rather than potentially crashing if the request size exceeds the
173    maximum size class.  (@jasone)
174  - jeprof:
175    + Add raw heap profile support.  (@jasone)
176    + Add --retain and --exclude for backtrace symbol filtering.  (@jasone)
177
178  Optimizations:
179  - Optimize the fast path to combine various bootstrapping and configuration
180    checks and execute more streamlined code in the common case.  (@interwq)
181  - Use linear scan for small bitmaps (used for small object tracking).  In
182    addition to speeding up bitmap operations on 64-bit systems, this reduces
183    allocator metadata overhead by approximately 0.2%.  (@djwatson)
184  - Separate arena_avail trees, which substantially speeds up run tree
185    operations.  (@djwatson)
186  - Use memoization (boot-time-computed table) for run quantization.  Separate
187    arena_avail trees reduced the importance of this optimization.  (@jasone)
188  - Attempt mmap-based in-place huge reallocation.  This can dramatically speed
189    up incremental huge reallocation.  (@jasone)
190
191  Incompatible changes:
192  - Make opt.narenas unsigned rather than size_t.  (@jasone)
193
194  Bug fixes:
195  - Fix stats.cactive accounting regression.  (@rustyx, @jasone)
196  - Handle unaligned keys in hash().  This caused problems for some ARM systems.
197    (@jasone, @cferris1000)
198  - Refactor arenas array.  In addition to fixing a fork-related deadlock, this
199    makes arena lookups faster and simpler.  (@jasone)
200  - Move retained memory allocation out of the default chunk allocation
201    function, to a location that gets executed even if the application installs
202    a custom chunk allocation function.  This resolves a virtual memory leak.
203    (@buchgr)
204  - Fix a potential tsd cleanup leak.  (@cferris1000, @jasone)
205  - Fix run quantization.  In practice this bug had no impact unless
206    applications requested memory with alignment exceeding one page.
207    (@jasone, @djwatson)
208  - Fix LinuxThreads-specific bootstrapping deadlock.  (Cosmin Paraschiv)
209  - jeprof:
210    + Don't discard curl options if timeout is not defined.  (@djwatson)
211    + Detect failed profile fetches.  (@djwatson)
212  - Fix stats.arenas.<i>.{dss,lg_dirty_mult,decay_time,pactive,pdirty} for
213    --disable-stats case.  (@jasone)
214
215* 4.0.4 (October 24, 2015)
216
217  This bugfix release fixes another xallocx() regression.  No other regressions
218  have come to light in over a month, so this is likely a good starting point
219  for people who prefer to wait for "dot one" releases with all the major issues
220  shaken out.
221
222  Bug fixes:
223  - Fix xallocx(..., MALLOCX_ZERO to zero the last full trailing page of large
224    allocations that have been randomly assigned an offset of 0 when
225    --enable-cache-oblivious configure option is enabled.
226
227* 4.0.3 (September 24, 2015)
228
229  This bugfix release continues the trend of xallocx() and heap profiling fixes.
230
231  Bug fixes:
232  - Fix xallocx(..., MALLOCX_ZERO) to zero all trailing bytes of large
233    allocations when --enable-cache-oblivious configure option is enabled.
234  - Fix xallocx(..., MALLOCX_ZERO) to zero trailing bytes of huge allocations
235    when resizing from/to a size class that is not a multiple of the chunk size.
236  - Fix prof_tctx_dump_iter() to filter out nodes that were created after heap
237    profile dumping started.
238  - Work around a potentially bad thread-specific data initialization
239    interaction with NPTL (glibc's pthreads implementation).
240
241* 4.0.2 (September 21, 2015)
242
243  This bugfix release addresses a few bugs specific to heap profiling.
244
245  Bug fixes:
246  - Fix ixallocx_prof_sample() to never modify nor create sampled small
247    allocations.  xallocx() is in general incapable of moving small allocations,
248    so this fix removes buggy code without loss of generality.
249  - Fix irallocx_prof_sample() to always allocate large regions, even when
250    alignment is non-zero.
251  - Fix prof_alloc_rollback() to read tdata from thread-specific data rather
252    than dereferencing a potentially invalid tctx.
253
254* 4.0.1 (September 15, 2015)
255
256  This is a bugfix release that is somewhat high risk due to the amount of
257  refactoring required to address deep xallocx() problems.  As a side effect of
258  these fixes, xallocx() now tries harder to partially fulfill requests for
259  optional extra space.  Note that a couple of minor heap profiling
260  optimizations are included, but these are better thought of as performance
261  fixes that were integral to disovering most of the other bugs.
262
263  Optimizations:
264  - Avoid a chunk metadata read in arena_prof_tctx_set(), since it is in the
265    fast path when heap profiling is enabled.  Additionally, split a special
266    case out into arena_prof_tctx_reset(), which also avoids chunk metadata
267    reads.
268  - Optimize irallocx_prof() to optimistically update the sampler state.  The
269    prior implementation appears to have been a holdover from when
270    rallocx()/xallocx() functionality was combined as rallocm().
271
272  Bug fixes:
273  - Fix TLS configuration such that it is enabled by default for platforms on
274    which it works correctly.
275  - Fix arenas_cache_cleanup() and arena_get_hard() to handle
276    allocation/deallocation within the application's thread-specific data
277    cleanup functions even after arenas_cache is torn down.
278  - Fix xallocx() bugs related to size+extra exceeding HUGE_MAXCLASS.
279  - Fix chunk purge hook calls for in-place huge shrinking reallocation to
280    specify the old chunk size rather than the new chunk size.  This bug caused
281    no correctness issues for the default chunk purge function, but was
282    visible to custom functions set via the "arena.<i>.chunk_hooks" mallctl.
283  - Fix heap profiling bugs:
284    + Fix heap profiling to distinguish among otherwise identical sample sites
285      with interposed resets (triggered via the "prof.reset" mallctl).  This bug
286      could cause data structure corruption that would most likely result in a
287      segfault.
288    + Fix irealloc_prof() to prof_alloc_rollback() on OOM.
289    + Make one call to prof_active_get_unlocked() per allocation event, and use
290      the result throughout the relevant functions that handle an allocation
291      event.  Also add a missing check in prof_realloc().  These fixes protect
292      allocation events against concurrent prof_active changes.
293    + Fix ixallocx_prof() to pass usize_max and zero to ixallocx_prof_sample()
294      in the correct order.
295    + Fix prof_realloc() to call prof_free_sampled_object() after calling
296      prof_malloc_sample_object().  Prior to this fix, if tctx and old_tctx were
297      the same, the tctx could have been prematurely destroyed.
298  - Fix portability bugs:
299    + Don't bitshift by negative amounts when encoding/decoding run sizes in
300      chunk header maps.  This affected systems with page sizes greater than 8
301      KiB.
302    + Rename index_t to szind_t to avoid an existing type on Solaris.
303    + Add JEMALLOC_CXX_THROW to the memalign() function prototype, in order to
304      match glibc and avoid compilation errors when including both
305      jemalloc/jemalloc.h and malloc.h in C++ code.
306    + Don't assume that /bin/sh is appropriate when running size_classes.sh
307      during configuration.
308    + Consider __sparcv9 a synonym for __sparc64__ when defining LG_QUANTUM.
309    + Link tests to librt if it contains clock_gettime(2).
310
311* 4.0.0 (August 17, 2015)
312
313  This version contains many speed and space optimizations, both minor and
314  major.  The major themes are generalization, unification, and simplification.
315  Although many of these optimizations cause no visible behavior change, their
316  cumulative effect is substantial.
317
318  New features:
319  - Normalize size class spacing to be consistent across the complete size
320    range.  By default there are four size classes per size doubling, but this
321    is now configurable via the --with-lg-size-class-group option.  Also add the
322    --with-lg-page, --with-lg-page-sizes, --with-lg-quantum, and
323    --with-lg-tiny-min options, which can be used to tweak page and size class
324    settings.  Impacts:
325    + Worst case performance for incrementally growing/shrinking reallocation
326      is improved because there are far fewer size classes, and therefore
327      copying happens less often.
328    + Internal fragmentation is limited to 20% for all but the smallest size
329      classes (those less than four times the quantum).  (1B + 4 KiB)
330      and (1B + 4 MiB) previously suffered nearly 50% internal fragmentation.
331    + Chunk fragmentation tends to be lower because there are fewer distinct run
332      sizes to pack.
333  - Add support for explicit tcaches.  The "tcache.create", "tcache.flush", and
334    "tcache.destroy" mallctls control tcache lifetime and flushing, and the
335    MALLOCX_TCACHE(tc) and MALLOCX_TCACHE_NONE flags to the *allocx() API
336    control which tcache is used for each operation.
337  - Implement per thread heap profiling, as well as the ability to
338    enable/disable heap profiling on a per thread basis.  Add the "prof.reset",
339    "prof.lg_sample", "thread.prof.name", "thread.prof.active",
340    "opt.prof_thread_active_init", "prof.thread_active_init", and
341    "thread.prof.active" mallctls.
342  - Add support for per arena application-specified chunk allocators, configured
343    via the "arena.<i>.chunk_hooks" mallctl.
344  - Refactor huge allocation to be managed by arenas, so that arenas now
345    function as general purpose independent allocators.  This is important in
346    the context of user-specified chunk allocators, aside from the scalability
347    benefits.  Related new statistics:
348    + The "stats.arenas.<i>.huge.allocated", "stats.arenas.<i>.huge.nmalloc",
349      "stats.arenas.<i>.huge.ndalloc", and "stats.arenas.<i>.huge.nrequests"
350      mallctls provide high level per arena huge allocation statistics.
351    + The "arenas.nhchunks", "arenas.hchunk.<i>.size",
352      "stats.arenas.<i>.hchunks.<j>.nmalloc",
353      "stats.arenas.<i>.hchunks.<j>.ndalloc",
354      "stats.arenas.<i>.hchunks.<j>.nrequests", and
355      "stats.arenas.<i>.hchunks.<j>.curhchunks" mallctls provide per size class
356      statistics.
357  - Add the 'util' column to malloc_stats_print() output, which reports the
358    proportion of available regions that are currently in use for each small
359    size class.
360  - Add "alloc" and "free" modes for for junk filling (see the "opt.junk"
361    mallctl), so that it is possible to separately enable junk filling for
362    allocation versus deallocation.
363  - Add the jemalloc-config script, which provides information about how
364    jemalloc was configured, and how to integrate it into application builds.
365  - Add metadata statistics, which are accessible via the "stats.metadata",
366    "stats.arenas.<i>.metadata.mapped", and
367    "stats.arenas.<i>.metadata.allocated" mallctls.
368  - Add the "stats.resident" mallctl, which reports the upper limit of
369    physically resident memory mapped by the allocator.
370  - Add per arena control over unused dirty page purging, via the
371    "arenas.lg_dirty_mult", "arena.<i>.lg_dirty_mult", and
372    "stats.arenas.<i>.lg_dirty_mult" mallctls.
373  - Add the "prof.gdump" mallctl, which makes it possible to toggle the gdump
374    feature on/off during program execution.
375  - Add sdallocx(), which implements sized deallocation.  The primary
376    optimization over dallocx() is the removal of a metadata read, which often
377    suffers an L1 cache miss.
378  - Add missing header includes in jemalloc/jemalloc.h, so that applications
379    only have to #include <jemalloc/jemalloc.h>.
380  - Add support for additional platforms:
381    + Bitrig
382    + Cygwin
383    + DragonFlyBSD
384    + iOS
385    + OpenBSD
386    + OpenRISC/or1k
387
388  Optimizations:
389  - Maintain dirty runs in per arena LRUs rather than in per arena trees of
390    dirty-run-containing chunks.  In practice this change significantly reduces
391    dirty page purging volume.
392  - Integrate whole chunks into the unused dirty page purging machinery.  This
393    reduces the cost of repeated huge allocation/deallocation, because it
394    effectively introduces a cache of chunks.
395  - Split the arena chunk map into two separate arrays, in order to increase
396    cache locality for the frequently accessed bits.
397  - Move small run metadata out of runs, into arena chunk headers.  This reduces
398    run fragmentation, smaller runs reduce external fragmentation for small size
399    classes, and packed (less uniformly aligned) metadata layout improves CPU
400    cache set distribution.
401  - Randomly distribute large allocation base pointer alignment relative to page
402    boundaries in order to more uniformly utilize CPU cache sets.  This can be
403    disabled via the --disable-cache-oblivious configure option, and queried via
404    the "config.cache_oblivious" mallctl.
405  - Micro-optimize the fast paths for the public API functions.
406  - Refactor thread-specific data to reside in a single structure.  This assures
407    that only a single TLS read is necessary per call into the public API.
408  - Implement in-place huge allocation growing and shrinking.
409  - Refactor rtree (radix tree for chunk lookups) to be lock-free, and make
410    additional optimizations that reduce maximum lookup depth to one or two
411    levels.  This resolves what was a concurrency bottleneck for per arena huge
412    allocation, because a global data structure is critical for determining
413    which arenas own which huge allocations.
414
415  Incompatible changes:
416  - Replace --enable-cc-silence with --disable-cc-silence to suppress spurious
417    warnings by default.
418  - Assure that the constness of malloc_usable_size()'s return type matches that
419    of the system implementation.
420  - Change the heap profile dump format to support per thread heap profiling,
421    rename pprof to jeprof, and enhance it with the --thread=<n> option.  As a
422    result, the bundled jeprof must now be used rather than the upstream
423    (gperftools) pprof.
424  - Disable "opt.prof_final" by default, in order to avoid atexit(3), which can
425    internally deadlock on some platforms.
426  - Change the "arenas.nlruns" mallctl type from size_t to unsigned.
427  - Replace the "stats.arenas.<i>.bins.<j>.allocated" mallctl with
428    "stats.arenas.<i>.bins.<j>.curregs".
429  - Ignore MALLOC_CONF in set{uid,gid,cap} binaries.
430  - Ignore MALLOCX_ARENA(a) in dallocx(), in favor of using the
431    MALLOCX_TCACHE(tc) and MALLOCX_TCACHE_NONE flags to control tcache usage.
432
433  Removed features:
434  - Remove the *allocm() API, which is superseded by the *allocx() API.
435  - Remove the --enable-dss options, and make dss non-optional on all platforms
436    which support sbrk(2).
437  - Remove the "arenas.purge" mallctl, which was obsoleted by the
438    "arena.<i>.purge" mallctl in 3.1.0.
439  - Remove the unnecessary "opt.valgrind" mallctl; jemalloc automatically
440    detects whether it is running inside Valgrind.
441  - Remove the "stats.huge.allocated", "stats.huge.nmalloc", and
442    "stats.huge.ndalloc" mallctls.
443  - Remove the --enable-mremap option.
444  - Remove the "stats.chunks.current", "stats.chunks.total", and
445    "stats.chunks.high" mallctls.
446
447  Bug fixes:
448  - Fix the cactive statistic to decrease (rather than increase) when active
449    memory decreases.  This regression was first released in 3.5.0.
450  - Fix OOM handling in memalign() and valloc().  A variant of this bug existed
451    in all releases since 2.0.0, which introduced these functions.
452  - Fix an OOM-related regression in arena_tcache_fill_small(), which could
453    cause cache corruption on OOM.  This regression was present in all releases
454    from 2.2.0 through 3.6.0.
455  - Fix size class overflow handling for malloc(), posix_memalign(), memalign(),
456    calloc(), and realloc() when profiling is enabled.
457  - Fix the "arena.<i>.dss" mallctl to return an error if "primary" or
458    "secondary" precedence is specified, but sbrk(2) is not supported.
459  - Fix fallback lg_floor() implementations to handle extremely large inputs.
460  - Ensure the default purgeable zone is after the default zone on OS X.
461  - Fix latent bugs in atomic_*().
462  - Fix the "arena.<i>.dss" mallctl to handle read-only calls.
463  - Fix tls_model configuration to enable the initial-exec model when possible.
464  - Mark malloc_conf as a weak symbol so that the application can override it.
465  - Correctly detect glibc's adaptive pthread mutexes.
466  - Fix the --without-export configure option.
467
468* 3.6.0 (March 31, 2014)
469
470  This version contains a critical bug fix for a regression present in 3.5.0 and
471  3.5.1.
472
473  Bug fixes:
474  - Fix a regression in arena_chunk_alloc() that caused crashes during
475    small/large allocation if chunk allocation failed.  In the absence of this
476    bug, chunk allocation failure would result in allocation failure, e.g.  NULL
477    return from malloc().  This regression was introduced in 3.5.0.
478  - Fix backtracing for gcc intrinsics-based backtracing by specifying
479    -fno-omit-frame-pointer to gcc.  Note that the application (and all the
480    libraries it links to) must also be compiled with this option for
481    backtracing to be reliable.
482  - Use dss allocation precedence for huge allocations as well as small/large
483    allocations.
484  - Fix test assertion failure message formatting.  This bug did not manifest on
485    x86_64 systems because of implementation subtleties in va_list.
486  - Fix inconsequential test failures for hash and SFMT code.
487
488  New features:
489  - Support heap profiling on FreeBSD.  This feature depends on the proc
490    filesystem being mounted during heap profile dumping.
491
492* 3.5.1 (February 25, 2014)
493
494  This version primarily addresses minor bugs in test code.
495
496  Bug fixes:
497  - Configure Solaris/Illumos to use MADV_FREE.
498  - Fix junk filling for mremap(2)-based huge reallocation.  This is only
499    relevant if configuring with the --enable-mremap option specified.
500  - Avoid compilation failure if 'restrict' C99 keyword is not supported by the
501    compiler.
502  - Add a configure test for SSE2 rather than assuming it is usable on i686
503    systems.  This fixes test compilation errors, especially on 32-bit Linux
504    systems.
505  - Fix mallctl argument size mismatches (size_t vs. uint64_t) in the stats unit
506    test.
507  - Fix/remove flawed alignment-related overflow tests.
508  - Prevent compiler optimizations that could change backtraces in the
509    prof_accum unit test.
510
511* 3.5.0 (January 22, 2014)
512
513  This version focuses on refactoring and automated testing, though it also
514  includes some non-trivial heap profiling optimizations not mentioned below.
515
516  New features:
517  - Add the *allocx() API, which is a successor to the experimental *allocm()
518    API.  The *allocx() functions are slightly simpler to use because they have
519    fewer parameters, they directly return the results of primary interest, and
520    mallocx()/rallocx() avoid the strict aliasing pitfall that
521    allocm()/rallocm() share with posix_memalign().  Note that *allocm() is
522    slated for removal in the next non-bugfix release.
523  - Add support for LinuxThreads.
524
525  Bug fixes:
526  - Unless heap profiling is enabled, disable floating point code and don't link
527    with libm.  This, in combination with e.g. EXTRA_CFLAGS=-mno-sse on x64
528    systems, makes it possible to completely disable floating point register
529    use.  Some versions of glibc neglect to save/restore caller-saved floating
530    point registers during dynamic lazy symbol loading, and the symbol loading
531    code uses whatever malloc the application happens to have linked/loaded
532    with, the result being potential floating point register corruption.
533  - Report ENOMEM rather than EINVAL if an OOM occurs during heap profiling
534    backtrace creation in imemalign().  This bug impacted posix_memalign() and
535    aligned_alloc().
536  - Fix a file descriptor leak in a prof_dump_maps() error path.
537  - Fix prof_dump() to close the dump file descriptor for all relevant error
538    paths.
539  - Fix rallocm() to use the arena specified by the ALLOCM_ARENA(s) flag for
540    allocation, not just deallocation.
541  - Fix a data race for large allocation stats counters.
542  - Fix a potential infinite loop during thread exit.  This bug occurred on
543    Solaris, and could affect other platforms with similar pthreads TSD
544    implementations.
545  - Don't junk-fill reallocations unless usable size changes.  This fixes a
546    violation of the *allocx()/*allocm() semantics.
547  - Fix growing large reallocation to junk fill new space.
548  - Fix huge deallocation to junk fill when munmap is disabled.
549  - Change the default private namespace prefix from empty to je_, and change
550    --with-private-namespace-prefix so that it prepends an additional prefix
551    rather than replacing je_.  This reduces the likelihood of applications
552    which statically link jemalloc experiencing symbol name collisions.
553  - Add missing private namespace mangling (relevant when
554    --with-private-namespace is specified).
555  - Add and use JEMALLOC_INLINE_C so that static inline functions are marked as
556    static even for debug builds.
557  - Add a missing mutex unlock in a malloc_init_hard() error path.  In practice
558    this error path is never executed.
559  - Fix numerous bugs in malloc_strotumax() error handling/reporting.  These
560    bugs had no impact except for malformed inputs.
561  - Fix numerous bugs in malloc_snprintf().  These bugs were not exercised by
562    existing calls, so they had no impact.
563
564* 3.4.1 (October 20, 2013)
565
566  Bug fixes:
567  - Fix a race in the "arenas.extend" mallctl that could cause memory corruption
568    of internal data structures and subsequent crashes.
569  - Fix Valgrind integration flaws that caused Valgrind warnings about reads of
570    uninitialized memory in:
571    + arena chunk headers
572    + internal zero-initialized data structures (relevant to tcache and prof
573      code)
574  - Preserve errno during the first allocation.  A readlink(2) call during
575    initialization fails unless /etc/malloc.conf exists, so errno was typically
576    set during the first allocation prior to this fix.
577  - Fix compilation warnings reported by gcc 4.8.1.
578
579* 3.4.0 (June 2, 2013)
580
581  This version is essentially a small bugfix release, but the addition of
582  aarch64 support requires that the minor version be incremented.
583
584  Bug fixes:
585  - Fix race-triggered deadlocks in chunk_record().  These deadlocks were
586    typically triggered by multiple threads concurrently deallocating huge
587    objects.
588
589  New features:
590  - Add support for the aarch64 architecture.
591
592* 3.3.1 (March 6, 2013)
593
594  This version fixes bugs that are typically encountered only when utilizing
595  custom run-time options.
596
597  Bug fixes:
598  - Fix a locking order bug that could cause deadlock during fork if heap
599    profiling were enabled.
600  - Fix a chunk recycling bug that could cause the allocator to lose track of
601    whether a chunk was zeroed.  On FreeBSD, NetBSD, and OS X, it could cause
602    corruption if allocating via sbrk(2) (unlikely unless running with the
603    "dss:primary" option specified).  This was completely harmless on Linux
604    unless using mlockall(2) (and unlikely even then, unless the
605    --disable-munmap configure option or the "dss:primary" option was
606    specified).  This regression was introduced in 3.1.0 by the
607    mlockall(2)/madvise(2) interaction fix.
608  - Fix TLS-related memory corruption that could occur during thread exit if the
609    thread never allocated memory.  Only the quarantine and prof facilities were
610    susceptible.
611  - Fix two quarantine bugs:
612    + Internal reallocation of the quarantined object array leaked the old
613      array.
614    + Reallocation failure for internal reallocation of the quarantined object
615      array (very unlikely) resulted in memory corruption.
616  - Fix Valgrind integration to annotate all internally allocated memory in a
617    way that keeps Valgrind happy about internal data structure access.
618  - Fix building for s390 systems.
619
620* 3.3.0 (January 23, 2013)
621
622  This version includes a few minor performance improvements in addition to the
623  listed new features and bug fixes.
624
625  New features:
626  - Add clipping support to lg_chunk option processing.
627  - Add the --enable-ivsalloc option.
628  - Add the --without-export option.
629  - Add the --disable-zone-allocator option.
630
631  Bug fixes:
632  - Fix "arenas.extend" mallctl to output the number of arenas.
633  - Fix chunk_recycle() to unconditionally inform Valgrind that returned memory
634    is undefined.
635  - Fix build break on FreeBSD related to alloca.h.
636
637* 3.2.0 (November 9, 2012)
638
639  In addition to a couple of bug fixes, this version modifies page run
640  allocation and dirty page purging algorithms in order to better control
641  page-level virtual memory fragmentation.
642
643  Incompatible changes:
644  - Change the "opt.lg_dirty_mult" default from 5 to 3 (32:1 to 8:1).
645
646  Bug fixes:
647  - Fix dss/mmap allocation precedence code to use recyclable mmap memory only
648    after primary dss allocation fails.
649  - Fix deadlock in the "arenas.purge" mallctl.  This regression was introduced
650    in 3.1.0 by the addition of the "arena.<i>.purge" mallctl.
651
652* 3.1.0 (October 16, 2012)
653
654  New features:
655  - Auto-detect whether running inside Valgrind, thus removing the need to
656    manually specify MALLOC_CONF=valgrind:true.
657  - Add the "arenas.extend" mallctl, which allows applications to create
658    manually managed arenas.
659  - Add the ALLOCM_ARENA() flag for {,r,d}allocm().
660  - Add the "opt.dss", "arena.<i>.dss", and "stats.arenas.<i>.dss" mallctls,
661    which provide control over dss/mmap precedence.
662  - Add the "arena.<i>.purge" mallctl, which obsoletes "arenas.purge".
663  - Define LG_QUANTUM for hppa.
664
665  Incompatible changes:
666  - Disable tcache by default if running inside Valgrind, in order to avoid
667    making unallocated objects appear reachable to Valgrind.
668  - Drop const from malloc_usable_size() argument on Linux.
669
670  Bug fixes:
671  - Fix heap profiling crash if sampled object is freed via realloc(p, 0).
672  - Remove const from __*_hook variable declarations, so that glibc can modify
673    them during process forking.
674  - Fix mlockall(2)/madvise(2) interaction.
675  - Fix fork(2)-related deadlocks.
676  - Fix error return value for "thread.tcache.enabled" mallctl.
677
678* 3.0.0 (May 11, 2012)
679
680  Although this version adds some major new features, the primary focus is on
681  internal code cleanup that facilitates maintainability and portability, most
682  of which is not reflected in the ChangeLog.  This is the first release to
683  incorporate substantial contributions from numerous other developers, and the
684  result is a more broadly useful allocator (see the git revision history for
685  contribution details).  Note that the license has been unified, thanks to
686  Facebook granting a license under the same terms as the other copyright
687  holders (see COPYING).
688
689  New features:
690  - Implement Valgrind support, redzones, and quarantine.
691  - Add support for additional platforms:
692    + FreeBSD
693    + Mac OS X Lion
694    + MinGW
695    + Windows (no support yet for replacing the system malloc)
696  - Add support for additional architectures:
697    + MIPS
698    + SH4
699    + Tilera
700  - Add support for cross compiling.
701  - Add nallocm(), which rounds a request size up to the nearest size class
702    without actually allocating.
703  - Implement aligned_alloc() (blame C11).
704  - Add the "thread.tcache.enabled" mallctl.
705  - Add the "opt.prof_final" mallctl.
706  - Update pprof (from gperftools 2.0).
707  - Add the --with-mangling option.
708  - Add the --disable-experimental option.
709  - Add the --disable-munmap option, and make it the default on Linux.
710  - Add the --enable-mremap option, which disables use of mremap(2) by default.
711
712  Incompatible changes:
713  - Enable stats by default.
714  - Enable fill by default.
715  - Disable lazy locking by default.
716  - Rename the "tcache.flush" mallctl to "thread.tcache.flush".
717  - Rename the "arenas.pagesize" mallctl to "arenas.page".
718  - Change the "opt.lg_prof_sample" default from 0 to 19 (1 B to 512 KiB).
719  - Change the "opt.prof_accum" default from true to false.
720
721  Removed features:
722  - Remove the swap feature, including the "config.swap", "swap.avail",
723    "swap.prezeroed", "swap.nfds", and "swap.fds" mallctls.
724  - Remove highruns statistics, including the
725    "stats.arenas.<i>.bins.<j>.highruns" and
726    "stats.arenas.<i>.lruns.<j>.highruns" mallctls.
727  - As part of small size class refactoring, remove the "opt.lg_[qc]space_max",
728    "arenas.cacheline", "arenas.subpage", "arenas.[tqcs]space_{min,max}", and
729    "arenas.[tqcs]bins" mallctls.
730  - Remove the "arenas.chunksize" mallctl.
731  - Remove the "opt.lg_prof_tcmax" option.
732  - Remove the "opt.lg_prof_bt_max" option.
733  - Remove the "opt.lg_tcache_gc_sweep" option.
734  - Remove the --disable-tiny option, including the "config.tiny" mallctl.
735  - Remove the --enable-dynamic-page-shift configure option.
736  - Remove the --enable-sysv configure option.
737
738  Bug fixes:
739  - Fix a statistics-related bug in the "thread.arena" mallctl that could cause
740    invalid statistics and crashes.
741  - Work around TLS deallocation via free() on Linux.  This bug could cause
742    write-after-free memory corruption.
743  - Fix a potential deadlock that could occur during interval- and
744    growth-triggered heap profile dumps.
745  - Fix large calloc() zeroing bugs due to dropping chunk map unzeroed flags.
746  - Fix chunk_alloc_dss() to stop claiming memory is zeroed.  This bug could
747    cause memory corruption and crashes with --enable-dss specified.
748  - Fix fork-related bugs that could cause deadlock in children between fork
749    and exec.
750  - Fix malloc_stats_print() to honor 'b' and 'l' in the opts parameter.
751  - Fix realloc(p, 0) to act like free(p).
752  - Do not enforce minimum alignment in memalign().
753  - Check for NULL pointer in malloc_usable_size().
754  - Fix an off-by-one heap profile statistics bug that could be observed in
755    interval- and growth-triggered heap profiles.
756  - Fix the "epoch" mallctl to update cached stats even if the passed in epoch
757    is 0.
758  - Fix bin->runcur management to fix a layout policy bug.  This bug did not
759    affect correctness.
760  - Fix a bug in choose_arena_hard() that potentially caused more arenas to be
761    initialized than necessary.
762  - Add missing "opt.lg_tcache_max" mallctl implementation.
763  - Use glibc allocator hooks to make mixed allocator usage less likely.
764  - Fix build issues for --disable-tcache.
765  - Don't mangle pthread_create() when --with-private-namespace is specified.
766
767* 2.2.5 (November 14, 2011)
768
769  Bug fixes:
770  - Fix huge_ralloc() race when using mremap(2).  This is a serious bug that
771    could cause memory corruption and/or crashes.
772  - Fix huge_ralloc() to maintain chunk statistics.
773  - Fix malloc_stats_print(..., "a") output.
774
775* 2.2.4 (November 5, 2011)
776
777  Bug fixes:
778  - Initialize arenas_tsd before using it.  This bug existed for 2.2.[0-3], as
779    well as for --disable-tls builds in earlier releases.
780  - Do not assume a 4 KiB page size in test/rallocm.c.
781
782* 2.2.3 (August 31, 2011)
783
784  This version fixes numerous bugs related to heap profiling.
785
786  Bug fixes:
787  - Fix a prof-related race condition.  This bug could cause memory corruption,
788    but only occurred in non-default configurations (prof_accum:false).
789  - Fix off-by-one backtracing issues (make sure that prof_alloc_prep() is
790    excluded from backtraces).
791  - Fix a prof-related bug in realloc() (only triggered by OOM errors).
792  - Fix prof-related bugs in allocm() and rallocm().
793  - Fix prof_tdata_cleanup() for --disable-tls builds.
794  - Fix a relative include path, to fix objdir builds.
795
796* 2.2.2 (July 30, 2011)
797
798  Bug fixes:
799  - Fix a build error for --disable-tcache.
800  - Fix assertions in arena_purge() (for real this time).
801  - Add the --with-private-namespace option.  This is a workaround for symbol
802    conflicts that can inadvertently arise when using static libraries.
803
804* 2.2.1 (March 30, 2011)
805
806  Bug fixes:
807  - Implement atomic operations for x86/x64.  This fixes compilation failures
808    for versions of gcc that are still in wide use.
809  - Fix an assertion in arena_purge().
810
811* 2.2.0 (March 22, 2011)
812
813  This version incorporates several improvements to algorithms and data
814  structures that tend to reduce fragmentation and increase speed.
815
816  New features:
817  - Add the "stats.cactive" mallctl.
818  - Update pprof (from google-perftools 1.7).
819  - Improve backtracing-related configuration logic, and add the
820    --disable-prof-libgcc option.
821
822  Bug fixes:
823  - Change default symbol visibility from "internal", to "hidden", which
824    decreases the overhead of library-internal function calls.
825  - Fix symbol visibility so that it is also set on OS X.
826  - Fix a build dependency regression caused by the introduction of the .pic.o
827    suffix for PIC object files.
828  - Add missing checks for mutex initialization failures.
829  - Don't use libgcc-based backtracing except on x64, where it is known to work.
830  - Fix deadlocks on OS X that were due to memory allocation in
831    pthread_mutex_lock().
832  - Heap profiling-specific fixes:
833    + Fix memory corruption due to integer overflow in small region index
834      computation, when using a small enough sample interval that profiling
835      context pointers are stored in small run headers.
836    + Fix a bootstrap ordering bug that only occurred with TLS disabled.
837    + Fix a rallocm() rsize bug.
838    + Fix error detection bugs for aligned memory allocation.
839
840* 2.1.3 (March 14, 2011)
841
842  Bug fixes:
843  - Fix a cpp logic regression (due to the "thread.{de,}allocatedp" mallctl fix
844    for OS X in 2.1.2).
845  - Fix a "thread.arena" mallctl bug.
846  - Fix a thread cache stats merging bug.
847
848* 2.1.2 (March 2, 2011)
849
850  Bug fixes:
851  - Fix "thread.{de,}allocatedp" mallctl for OS X.
852  - Add missing jemalloc.a to build system.
853
854* 2.1.1 (January 31, 2011)
855
856  Bug fixes:
857  - Fix aligned huge reallocation (affected allocm()).
858  - Fix the ALLOCM_LG_ALIGN macro definition.
859  - Fix a heap dumping deadlock.
860  - Fix a "thread.arena" mallctl bug.
861
862* 2.1.0 (December 3, 2010)
863
864  This version incorporates some optimizations that can't quite be considered
865  bug fixes.
866
867  New features:
868  - Use Linux's mremap(2) for huge object reallocation when possible.
869  - Avoid locking in mallctl*() when possible.
870  - Add the "thread.[de]allocatedp" mallctl's.
871  - Convert the manual page source from roff to DocBook, and generate both roff
872    and HTML manuals.
873
874  Bug fixes:
875  - Fix a crash due to incorrect bootstrap ordering.  This only impacted
876    --enable-debug --enable-dss configurations.
877  - Fix a minor statistics bug for mallctl("swap.avail", ...).
878
879* 2.0.1 (October 29, 2010)
880
881  Bug fixes:
882  - Fix a race condition in heap profiling that could cause undefined behavior
883    if "opt.prof_accum" were disabled.
884  - Add missing mutex unlocks for some OOM error paths in the heap profiling
885    code.
886  - Fix a compilation error for non-C99 builds.
887
888* 2.0.0 (October 24, 2010)
889
890  This version focuses on the experimental *allocm() API, and on improved
891  run-time configuration/introspection.  Nonetheless, numerous performance
892  improvements are also included.
893
894  New features:
895  - Implement the experimental {,r,s,d}allocm() API, which provides a superset
896    of the functionality available via malloc(), calloc(), posix_memalign(),
897    realloc(), malloc_usable_size(), and free().  These functions can be used to
898    allocate/reallocate aligned zeroed memory, ask for optional extra memory
899    during reallocation, prevent object movement during reallocation, etc.
900  - Replace JEMALLOC_OPTIONS/JEMALLOC_PROF_PREFIX with MALLOC_CONF, which is
901    more human-readable, and more flexible.  For example:
902      JEMALLOC_OPTIONS=AJP
903    is now:
904      MALLOC_CONF=abort:true,fill:true,stats_print:true
905  - Port to Apple OS X.  Sponsored by Mozilla.
906  - Make it possible for the application to control thread-->arena mappings via
907    the "thread.arena" mallctl.
908  - Add compile-time support for all TLS-related functionality via pthreads TSD.
909    This is mainly of interest for OS X, which does not support TLS, but has a
910    TSD implementation with similar performance.
911  - Override memalign() and valloc() if they are provided by the system.
912  - Add the "arenas.purge" mallctl, which can be used to synchronously purge all
913    dirty unused pages.
914  - Make cumulative heap profiling data optional, so that it is possible to
915    limit the amount of memory consumed by heap profiling data structures.
916  - Add per thread allocation counters that can be accessed via the
917    "thread.allocated" and "thread.deallocated" mallctls.
918
919  Incompatible changes:
920  - Remove JEMALLOC_OPTIONS and malloc_options (see MALLOC_CONF above).
921  - Increase default backtrace depth from 4 to 128 for heap profiling.
922  - Disable interval-based profile dumps by default.
923
924  Bug fixes:
925  - Remove bad assertions in fork handler functions.  These assertions could
926    cause aborts for some combinations of configure settings.
927  - Fix strerror_r() usage to deal with non-standard semantics in GNU libc.
928  - Fix leak context reporting.  This bug tended to cause the number of contexts
929    to be underreported (though the reported number of objects and bytes were
930    correct).
931  - Fix a realloc() bug for large in-place growing reallocation.  This bug could
932    cause memory corruption, but it was hard to trigger.
933  - Fix an allocation bug for small allocations that could be triggered if
934    multiple threads raced to create a new run of backing pages.
935  - Enhance the heap profiler to trigger samples based on usable size, rather
936    than request size.
937  - Fix a heap profiling bug due to sometimes losing track of requested object
938    size for sampled objects.
939
940* 1.0.3 (August 12, 2010)
941
942  Bug fixes:
943  - Fix the libunwind-based implementation of stack backtracing (used for heap
944    profiling).  This bug could cause zero-length backtraces to be reported.
945  - Add a missing mutex unlock in library initialization code.  If multiple
946    threads raced to initialize malloc, some of them could end up permanently
947    blocked.
948
949* 1.0.2 (May 11, 2010)
950
951  Bug fixes:
952  - Fix junk filling of large objects, which could cause memory corruption.
953  - Add MAP_NORESERVE support for chunk mapping, because otherwise virtual
954    memory limits could cause swap file configuration to fail.  Contributed by
955    Jordan DeLong.
956
957* 1.0.1 (April 14, 2010)
958
959  Bug fixes:
960  - Fix compilation when --enable-fill is specified.
961  - Fix threads-related profiling bugs that affected accuracy and caused memory
962    to be leaked during thread exit.
963  - Fix dirty page purging race conditions that could cause crashes.
964  - Fix crash in tcache flushing code during thread destruction.
965
966* 1.0.0 (April 11, 2010)
967
968  This release focuses on speed and run-time introspection.  Numerous
969  algorithmic improvements make this release substantially faster than its
970  predecessors.
971
972  New features:
973  - Implement autoconf-based configuration system.
974  - Add mallctl*(), for the purposes of introspection and run-time
975    configuration.
976  - Make it possible for the application to manually flush a thread's cache, via
977    the "tcache.flush" mallctl.
978  - Base maximum dirty page count on proportion of active memory.
979  - Compute various additional run-time statistics, including per size class
980    statistics for large objects.
981  - Expose malloc_stats_print(), which can be called repeatedly by the
982    application.
983  - Simplify the malloc_message() signature to only take one string argument,
984    and incorporate an opaque data pointer argument for use by the application
985    in combination with malloc_stats_print().
986  - Add support for allocation backed by one or more swap files, and allow the
987    application to disable over-commit if swap files are in use.
988  - Implement allocation profiling and leak checking.
989
990  Removed features:
991  - Remove the dynamic arena rebalancing code, since thread-specific caching
992    reduces its utility.
993
994  Bug fixes:
995  - Modify chunk allocation to work when address space layout randomization
996    (ASLR) is in use.
997  - Fix thread cleanup bugs related to TLS destruction.
998  - Handle 0-size allocation requests in posix_memalign().
999  - Fix a chunk leak.  The leaked chunks were never touched, so this impacted
1000    virtual memory usage, but not physical memory usage.
1001
1002* linux_2008082[78]a (August 27/28, 2008)
1003
1004  These snapshot releases are the simple result of incorporating Linux-specific
1005  support into the FreeBSD malloc sources.
1006
1007--------------------------------------------------------------------------------
1008vim:filetype=text:textwidth=80
1009