contention. The intent is to dynamically adjust to load imbalances, which
can cause severe contention.
Use pthread mutexes where possible instead of libc "spinlocks" (they aren't
actually spin locks). Conceptually, this change is meant only to support
the dynamic load balancing code by enabling the use of spin locks, but it
has the added apparent benefit of substantially improving performance due to
reduced context switches when there is moderate arena lock contention.
Proper tuning parameter configuration for this change is a finicky business,
and it is very much machine-dependent. One seemingly promising solution
would be to run a tuning program during operating system installation that
computes appropriate settings for load balancing. (The pthreads adaptive
spin locks should probably be similarly tuned.)
vector of slots for lazily freed objects. For each deallocation, before
doing the hard work of locking the arena and deallocating, try several times
to randomly insert the object into the vector using atomic operations.
This approach is particularly effective at reducing contention for
multi-threaded applications that use the producer-consumer model, wherein
one producer thread allocates objects, then multiple consumer threads
deallocate those objects.
allocations. [1]
Fix calculation of the number of arenas when 'n' is specified via
MALLOC_OPTIONS.
Clean up various style inconsistencies.
Obtained from: [1] NetBSD
avoid downcasting issues. In particular, this change fixes
posix_memalign(3) for alignments greater than 2^31 on LP64 systems.
Make sure that NDEBUG is always set to be compatible with MALLOC_DEBUG. [1]
Reported by: [1] Lee Hyo geol <hyogeollee@gmail.com>
trees that track all non-full runs for each bin. Use the red-black
trees to be able to guarantee that each new allocation is placed in the
lowest address available in any non-full run. This change completes the
transition to allocating from low addresses in order to reduce the
retention of sparsely used chunks.
If the run in current use by a bin becomes empty, deallocate the run
rather than retaining it for later use. The previous behavior had the
tendency to spread empty runs across multiple chunks, thus preventing
the release of chunks that were completely unused.
Generalize base_chunk_alloc() (and rename it to base_pages_alloc()) to
handle allocation sizes larger than the chunk size, so that it is
possible to support chunk sizes that are smaller than an arena object.
Reduce the minimum chunk size from 64kB to 8kB.
Optimize tracking of addresses for deleted chunks.
Fix a statistics bug for huge allocations.
rounding and overflow. Carefully document what the various overflow
tests actually detect.
The bugs mostly canceled out, such that the worst possible failure
cases resulted in non-fatal over-allocations.
than binary buddies, the alignment guarantees are weaker, which requires
a more complex aligned allocation algorithm, similar to that used for
alignment greater than the chunk size.
Reported by: matteo
chunks. This allows runs to be any multiple of the page size. The
primary advantage is that large objects are no longer constrained to be
2^n pages, which can dramatically decrease internal fragmentation for
large objects. This also allows the sizes for runs that back small
objects to be more finely tuned.
Free runs are searched for linearly using the chunk page map (with the
help of some heuristic optimizations). This changes the allocation
policy from "first best fit" to "first fit". A prototype red-black tree
implementation for tracking free runs that implemented "first best fit"
did not cause a measurable speed or memory usage difference for
realistic chunk sizes (though of course it is possible to construct
benchmarks that favor one allocation policy over another).
Refine the handling of fullness constraints for small runs to be more
tunable.
Restructure the per chunk page map to contain only two fields per entry,
rather than four. Also, increase each entry from 4 to 8 bytes, since it
allows for 32-bit integers, without increasing the number of chunk
header pages.
Relax the maximum chunk size constraint. This is of no practical
interest; it is merely fallout from the chunk page map restructuring.
Revamp statistics gathering and reporting to be faster, clearer and more
informative. Statistics gathering is fast enough now to have little
to no impact on application speed, but it still requires approximately
two extra pages of memory per arena (per process). This memory overhead
may be acceptable for most systems, but we still need to leave
statistics gathering disabled by default in RELENG branches.
Rename NO_MALLOC_EXTRAS to MALLOC_PRODUCTION in order to make its intent
clearer (i.e. it should be defined in RELENG branches).
avoid substantial potential bloat for static binaries that do not
otherwise use any printf(3)-family functions. [1]
Rearrange arena_run_t so that the region bitmask can be minimally sized
according to constraints related to each bin's size class. Previously,
the region bitmask was the same size for all run headers, which wasted
a measurable amount of memory.
Rather than making runs for small objects as large as possible, make
runs as small as possible such that header overhead stays below a
certain bound. There are two exceptions that override the header
overhead bound:
1) If the bound is impossible to honor, it is relaxed on a
per-size-class basis. Since there is one bit of header
overhead per object (plus a constant), it is impossible to
achieve a header overhead less than or equal to 1/(# of bits
per object). For the current setting of maximum 0.5% header
overhead, this relaxation comes into play for {2, 4, 8,
16}-byte objects, for which header overhead is (on 64-bit
systems) {7.1, 4.3, 2.2, 1.2}%, respectively.
2) There is still a cap on small run size, still set to 64kB.
This comes into play for {1024, 2048}-byte objects, for which
header overhead is {1.6, 3.1}%, respectively.
In practice, this reduces the run sizes, which makes worst case
low-water memory usage due to fragmentation less bad. It also reduces
worst case high-water run fragmentation due to non-full runs, but this
is only a constant improvement (most important to small short-lived
processes).
Reduce the default chunk size from 2MB to 1MB. Benchmarks indicate that
the external fragmentation reduction makes 1MB the new sweet spot (as
small as possible without adversely affecting performance).
Reported by: [1] kientzle
This has no impact unless USE_BRK is defined (32-bit platforms), in
which case user allocations are allocated via mmap() if at all possible,
in order to avoid the possibility of unreclaimable chunks in the data
segment.
Fix an obscure bug in base_alloc() that could have allowed undefined
behavior if an application were to use sbrk() in conjunction with a
USE_BRK-enabled malloc.
chunk per arena, rather than immediately deallocating all unused chunks.
This fixes a potential performance issue when allocating/deallocating
an object of size (4kB..1MB] in a loop.
Reported by: davidxu
(size_t)(num * size) == 0
but both num and size are nonzero.
Reported by: Ilja van Sprundel
Approved by: jasone
Security: Integer overflow; calloc was allocating 1 byte in
response to a request for a multiple of 2^32 (or 2^64)
bytes instead of returning NULL.
well as avoiding a switch statement. This change has no significant impact
to performance when branch prediction is successful at predicting the sizes
of objects passed to free(), but in the case that the object sizes are
semi-random, this change has the potential to prevent many branch prediction
misses, thus improving performance substantially.
Take advantage of alignment guarantees in ipalloc(), and pad object sizes to
something less than a power of two when possible. This has the potential
to substantially reduce internal fragmentation for objects allocated via
posix_memalign().
Avoid an unnecessary pow2_ceil() call in arena_ralloc().
Submitted by: djam8193ah@hotmail.com
and instead creating a small allocation for each malloc(0) call. The
optional SysV compatibility behavior remains unchanged.
Add a couple of assertions.
Fix a couple of typos in error message strings.
4kB pages), in order to avoid dangerous rounding error when calculating
fullness limits during run promotion/demotion.
Convert a structure bitfield to a normal field in areana_run_t. This should
have been changed along with the other fields in revision 1.120.
bounds. [1]
Modify logic for utilizing the data segment, such that it is possible to
create huge allocations there.
Shrink the data segment when deallocating a chunk, if it is at the end of
the data segment.
Rename chunk_size to csize in huge_malloc(), in order to avoid masking a
static variable of the same name. [1]
Reported by: Paul Allen <nospam@ugcs.caltech.edu>
races. This isn't currently necessary for libpthread or libthr, but
without it external threads libraries like the linuxthreads port are
not safe to use.
Reported by: ganbold@micom.mng.net
have to be calculated once per allocator operation.
Make nil const.
Update various comments.
Remove/avoid division where possible.
For the one division operation that remains in the critical path, add a
switch statement that has a case for each small size class, and do division
with a constant divisor in each case. This allows the compiler to generate
optimized code that does not use hardware division [1].
Obtained from: peter [1]
* Avoid choosing an arena until it's certain that an arena is needed
for allocation.
* Convert division/multiplication to bitshifting where possible.
* Avoid accessing TLS variables in single-threaded code.
* Reduce the amount of pointer dereferencing.
* Move lock acquisition in critical paths to only protect the the code
that requires synchronization, and completely remove locking where
possible.
determine its value at run time according to other relevant values. This
avoids the creation of runs that are incompletely utilized, as long as
pagesize isn't too large (>32kB, given the current RUN_MIN_REGS_2POW
setting).
Increase the size of several structure bitfields in arena_run_t in order
to avoid integer overflow in the case that a run's header does not overlap
with the space that is usable as application allocation regions. Given
the tiny_min_2pow change, this fix has no additional impact unless
pagesize is >32kB.
Reported by: kris
internally used chunk to start at the beginning of the heap, rather
than at a chunk-aligned address. This reduces mapped memory somewhat
for 32-bit architectures.
Add the arena_run_link_t type and use it wherever a run object is only
used as a ring 'header'. This saves approximately 40 kB of memory per
arena.
Remove an obsolete (no longer used) code path from base_alloc(), which
supported the internal allocation of objects larger than the chunk
size.
Enhance chunk_dealloc() to cache chunk addresses for all deallocated
chunks. This has no impact for most programs, but has the potential
to reduce VM map fragmentation for programs that use huge
allocations.
that no linear searching is necessary if we resort to allocating from a
run that is known to be mostly full. There are pathological edge cases
that could have caused severely degraded performance, and this change
fixes that.
close enough to each other that reallocation would allocate a new region
of the same size. This improves the performance of repeated incremental
reallocations by up to three orders of magnitude. [1]
Fix arena_new() to properly constrain run size if a small chunk size was
specified during runtime configuration.
Suggested by: se [1]
allocation patterns that involve a relatively even mixture of many
different size classes.
Reduce the chunk size from 16 MB to 2 MB. Since chunks are now carved up
using an address-ordered first best fit policy, VM map fragmentation is
much less likely, which makes smaller chunks not as much of a risk. This
reduces the virtual memory size of most applications.
Remove redzones, since program buffer overruns are no longer as likely to
corrupt malloc data structures.
Remove the C MALLOC_OPTIONS flag, and add H and S.
Remove the block of code that tries to use delayed regions in LIFO order,
since from a policy perspective, it conflicts with LRU caching of newly
coalesced regions in arena_undelay(). There are numerous policy
alternatives, and it isn't readily obvious which (if any) is superior;
this change at least has the virtue of being consistent with policy.
fit regions are available, use the delayed regions in LIFO order, in order
to increase locality of reference. We might expect this to cause delayed
regions to be removed from the delay ring buffer more often (since we're
now re-using more recently buffered regions), but numerous tests indicate
that the overall impact on memory usage tends to be good (reduced
fragmentation).
Re-work arena_frag_reg_alloc() so that when large free regions are
exhausted, it uses small regions in a way that favors contiguous allocation
of sequentially allocated small regions. Use arena_frag_reg_alloc() in
this capacity, rather than directly attempting over-fitting of small
requests when no large regions are available.
Remove the bin overfit statistic, since it is no longer relevant due to
the arena_frag_reg_alloc() changes.
Do not specify arena_frag_reg_alloc() as an inline function. It is too
large to benefit much from being inlined, and it is also called in two
places, only one of which is in the critical path (the other call bloated
arena_reg_alloc()).
Call arena_coalesce() for a region before caching it with
arena_mru_cache().
Add assertions that detect the attempted caching of adjacent free regions,
so that we notice this problem when it is first created, rather than in
arena_coalesce(), when it's too late to know how the problem arose.
Reported by: Hans Blancke
