atexit(3) can deadlock internally during its own initialization if
jemalloc calls atexit() during jemalloc initialization. Mitigate the
impact by restructuring prof initialization to avoid calling atexit()
unless the registered function will actually dump a final heap profile.
Additionally, disable prof_final by default so that this land mine is
opt-in rather than opt-out.
This resolves#144.
According to the docbook documentation for <funcprototype>, its parent
must be <funcsynopsis>; fix accordingly. Nonetheless, the man page
processor fails badly when this construct is embedded in a <para> (which
is documented to be legal), although the html processor does fine.
This adds support for expanding huge allocations in-place by requesting
memory at a specific address from the chunk allocator.
It's currently only implemented for the chunk recycling path, although
in theory it could also be done by optimistically allocating new chunks.
On Linux, it could attempt an in-place mremap. However, that won't work
in practice since the heap is grown downwards and memory is not unmapped
(in a normal build, at least).
Repeated vector reallocation micro-benchmark:
#include <string.h>
#include <stdlib.h>
int main(void) {
for (size_t i = 0; i < 100; i++) {
void *ptr = NULL;
size_t old_size = 0;
for (size_t size = 4; size < (1 << 30); size *= 2) {
ptr = realloc(ptr, size);
if (!ptr) return 1;
memset(ptr + old_size, 0xff, size - old_size);
old_size = size;
}
free(ptr);
}
}
The glibc allocator fails to do any in-place reallocations on this
benchmark once it passes the M_MMAP_THRESHOLD (default 128k) but it
elides the cost of copies via mremap, which is currently not something
that jemalloc can use.
With this improvement, jemalloc still fails to do any in-place huge
reallocations for the first outer loop, but then succeeds 100% of the
time for the remaining 99 iterations. The time spent doing allocations
and copies drops down to under 5%, with nearly all of it spent doing
purging + faulting (when huge pages are disabled) and the array memset.
An improved mremap API (MREMAP_RETAIN - #138) would be far more general
but this is a portable optimization and would still be useful on Linux
for xallocx.
Numbers with transparent huge pages enabled:
glibc (copies elided via MREMAP_MAYMOVE): 8.471s
jemalloc: 17.816s
jemalloc + no-op madvise: 13.236s
jemalloc + this commit: 6.787s
jemalloc + this commit + no-op madvise: 6.144s
Numbers with transparent huge pages disabled:
glibc (copies elided via MREMAP_MAYMOVE): 15.403s
jemalloc: 39.456s
jemalloc + no-op madvise: 12.768s
jemalloc + this commit: 15.534s
jemalloc + this commit + no-op madvise: 6.354s
Closes#137
Implement/test/fix the opt.prof_thread_active_init,
prof.thread_active_init, and thread.prof.active mallctl's.
Test/fix the thread.prof.name mallctl.
Refactor opt_prof_active to be read-only and move mutable state into the
prof_active variable. Stop leaning on ctl-related locking for
protection.
This adds a new `sdallocx` function to the external API, allowing the
size to be passed by the caller. It avoids some extra reads in the
thread cache fast path. In the case where stats are enabled, this
avoids the work of calculating the size from the pointer.
An assertion validates the size that's passed in, so enabling debugging
will allow users of the API to debug cases where an incorrect size is
passed in.
The performance win for a contrived microbenchmark doing an allocation
and immediately freeing it is ~10%. It may have a different impact on a
real workload.
Closes#28
Rename data structures (prof_thr_cnt_t-->prof_tctx_t,
prof_ctx_t-->prof_gctx_t), and convert to storing a prof_tctx_t for
sampled objects.
Convert PROF_ALLOC_PREP() to prof_alloc_prep(), since precise backtrace
depth within jemalloc functions is no longer an issue (pprof prunes
irrelevant frames).
Implement mallctl's:
- prof.reset implements full sample data reset, and optional change of
sample interval.
- prof.lg_sample reads the current sample interval (opt.lg_prof_sample
was the permanent source of truth prior to prof.reset).
- thread.prof.name provides naming capability for threads within heap
profile dumps.
- thread.prof.active makes it possible to activate/deactivate heap
profiling for individual threads.
Modify the heap dump files to contain per thread heap profile data.
This change is incompatible with the existing pprof, which will require
enhancements to read and process the enriched data.
Refactor huge allocation to be managed by arenas (though the global
red-black tree of huge allocations remains for lookup during
deallocation). This is the logical conclusion of recent changes that 1)
made per arena dss precedence apply to huge allocation, and 2) made it
possible to replace the per arena chunk allocation/deallocation
functions.
Remove the top level huge stats, and replace them with per arena huge
stats.
Normalize function names and types to *dalloc* (some were *dealloc*).
Remove the --enable-mremap option. As jemalloc currently operates, this
is a performace regression for some applications, but planned work to
logarithmically space huge size classes should provide similar amortized
performance. The motivation for this change was that mremap-based huge
reallocation forced leaky abstractions that prevented refactoring.
Add new mallctl endpoints "arena<i>.chunk.alloc" and
"arena<i>.chunk.dealloc" to allow userspace to configure
jemalloc's chunk allocator and deallocator on a per-arena
basis.
Forcefully disable tcache if running inside Valgrind, and remove
Valgrind calls in tcache-specific code.
Restructure Valgrind-related code to move most Valgrind calls out of the
fast path functions.
Take advantage of static knowledge to elide some branches in
JEMALLOC_VALGRIND_REALLOC().
Make dss non-optional on all platforms which support sbrk(2).
Fix the "arena.<i>.dss" mallctl to return an error if "primary" or
"secondary" precedence is specified, but sbrk(2) is not supported.
Extract profiling code from malloc(), imemalign(), calloc(), realloc(),
mallocx(), rallocx(), and xallocx(). This slightly reduces the amount
of code compiled into the fast paths, but the primary benefit is the
combinatorial complexity reduction.
Simplify iralloc[t]() by creating a separate ixalloc() that handles the
no-move cases.
Further simplify [mrxn]allocx() (and by implication [mrn]allocm()) to
make request size overflows due to size class and/or alignment
constraints trigger undefined behavior (detected by debug-only
assertions).
Report ENOMEM rather than EINVAL if an OOM occurs during heap profiling
backtrace creation in imemalign(). This bug impacted posix_memalign()
and aligned_alloc().
*mallctl() always returns EINVAL and does partial result copying when
*oldlenp is to short to hold the requested value, rather than returning
ENOMEM. Therefore remove ENOMEM from the documented set of possible
errors.
Implement the *allocx() API, which is a successor to the *allocm() API.
The *allocx() functions are slightly simpler to use because they have
fewer parameters, they directly return the results of primary interest,
and mallocx()/rallocx() avoid the strict aliasing pitfall that
allocm()/rallocx() share with posix_memalign(). The following code
violates strict aliasing rules:
foo_t *foo;
allocm((void **)&foo, NULL, 42, 0);
whereas the following is safe:
foo_t *foo;
void *p;
allocm(&p, NULL, 42, 0);
foo = (foo_t *)p;
mallocx() does not have this problem:
foo_t *foo = (foo_t *)mallocx(42, 0);
Modify processing of the lg_chunk option so that it clips an
out-of-range input to the edge of the valid range. This makes it
possible to request the minimum possible chunk size without intimate
knowledge of allocator internals.
Submitted by Ian Lepore (see FreeBSD PR bin/174641).
Purge unused dirty pages in an order that first performs clean/dirty run
defragmentation, in order to mitigate available run fragmentation.
Remove the limitation that prevented purging unless at least one chunk
worth of dirty pages had accumulated in an arena. This limitation was
intended to avoid excessive purging for small applications, but the
threshold was arbitrary, and the effect of questionable utility.
Relax opt_lg_dirty_mult from 5 to 3. This compensates for increased
likelihood of allocating clean runs, given the same ratio of clean:dirty
runs, and reduces the potential for repeated purging in pathological
large malloc/free loops that push the active:dirty page ratio just over
the purge threshold.
Add the "arenas.extend" mallctl, so that it is possible to create new
arenas that are outside the set that jemalloc automatically multiplexes
threads onto.
Add the ALLOCM_ARENA() flag for {,r,d}allocm(), so that it is possible
to explicitly allocate from a particular arena.
Add the "opt.dss" mallctl, which controls the default precedence of dss
allocation relative to mmap allocation.
Add the "arena.<i>.dss" mallctl, which makes it possible to set the
default dss precedence on a per arena or global basis.
Add the "arena.<i>.purge" mallctl, which obsoletes "arenas.purge".
Add the "stats.arenas.<i>.dss" mallctl.
Generalize "stats.mapped" documentation to state that all inactive
chunks are omitted, now that it is possible for mmap'ed chunks to be
omitted in addition to DSS chunks.
Add the --enable-mremap option, and disable the use of mremap(2) by
default, for the same reason that freeing chunks via munmap(2) is
disabled by default on Linux: semi-permanent VM map fragmentation.
Fix a memory corruption bug in chunk_alloc_dss() that was due to
claiming newly allocated memory is zeroed.
Reverse order of preference between mmap() and sbrk() to prefer mmap().
Clean up management of 'zero' parameter in chunk_alloc*().
Change the "opt.lg_prof_sample" default from 0 to 19 (1 B to 512 KiB).
Change the "opt.prof_accum" default from true to false.
Add the "opt.prof_final" mallctl, so that "opt.prof_prefix" need not be
abused to disable final profile dumping.
Add the --disable-munmap option, remove the configure test that
attempted to detect the VM allocation quirk known to exist on Linux
x86[_64], and make --disable-munmap implicit on Linux.
Always disable redzone by default, even when --enable-debug is
specified. The memory overhead for redzones can be substantial, which
makes this feature something that should only be opted into.
Implement Valgrind support, as well as the redzone and quarantine
features, which help Valgrind detect memory errors. Redzones are only
implemented for small objects because the changes necessary to support
redzones around large and huge objects are complicated by in-place
reallocation, to the point that it isn't clear that the maintenance
burden is worth the incremental improvement to Valgrind support.
Merge arena_salloc() and arena_salloc_demote().
Refactor i[v]salloc() to expose the 'demote' option.
s/PAGE_SHIFT/LG_PAGE/g and s/PAGE_SIZE/PAGE/g.
Remove remnants of the dynamic-page-shift code.
Rename the "arenas.pagesize" mallctl to "arenas.page".
Remove the "arenas.chunksize" mallctl, which is redundant with
"opt.lg_chunk".
Implement aligned_alloc(), which was added in the C11 standard. The
function is weakly specified to the point that a minimally compliant
implementation would be painful to use (size must be an integral
multiple of alignment!), which in practice makes posix_memalign() a
safer choice.
Remove the lg_tcache_gc_sweep option, because it is no longer
very useful. Prior to the addition of dynamic adjustment of tcache fill
count, it was possible for fill/flush overhead to be a problem, but this
problem no longer occurs.
Add nallocm(), which computes the real allocation size that would result
from the corresponding allocm() call. nallocm() is a functional
superset of OS X's malloc_good_size(), in that it takes alignment
constraints into account.
Program-generate small size class tables for all valid combinations of
LG_TINY_MIN, LG_QUANTUM, and PAGE_SHIFT. Use the appropriate table to generate
all relevant data structures, and remove the distinction between
tiny/quantum/cacheline/subpage bins.
Remove --enable-dynamic-page-shift. This option didn't prove useful in
practice, and it prevented optimizations.
Add Tilera architecture support.
Remove opt.lg_prof_bt_max, and hard code it to 7. The original
intention of this option was to enable faster backtracing by limiting
backtrace depth. However, this makes graphical pprof output very
difficult to interpret. In practice, decreasing sampling frequency is a
better mechanism for limiting profiling overhead.
Remove the opt.lg_prof_tcmax option and hard-code a cache size of 1024.
This setting is something that users just shouldn't have to worry about.
If lock contention actually ends up being a problem, the simple solution
available to the user is to reduce sampling frequency.
When tiny size class support was first added, it was intended to support
truly tiny size classes (even 2 bytes). However, this wasn't very
useful in practice, so the minimum tiny size class has been limited to
sizeof(void *) for a long time now. This is too small to be standards
compliant, but other commonly used malloc implementations do not even
bother using a 16-byte quantum on systems with vector units (SSE2+,
AltiVEC, etc.). As such, it is safe in practice to support an 8-byte
tiny size class on 64-bit systems that support 16-byte types.
