When purging, large allocations are usually the ones that cross the npages_limit
threshold, simply because they are "large". This means we often leave the large
extent around for a while, which has the downsides of: 1) high RSS and 2) more
chance of them getting fragmented. Given that they are not likely to be reused
very soon (LRU), let's over purge by 1 extent (which is often large and not
reused frequently).
When allocating from dirty extents (which we always prefer if available), large
active extents can get split even if the new allocation is much smaller, in
which case the introduced fragmentation causes high long term damage. This new
option controls the threshold to reuse and split an existing active extent. We
avoid using a large extent for much smaller sizes, in order to reduce
fragmentation. In some workload, adding the threshold improves virtual memory
usage by >10x.
While working on #852, I noticed the prng state is atomic. This is the only
atomic use of prng in all of jemalloc. Instead, use a threadlocal prng
state if possible to avoid unnecessary cache line contention.
Added an upper bound on how many pages we can decay during the current run.
Without this, decay could have unbounded increase in stashed, since other
threads could add new pages into the extents.
This option controls the max size when grow_retained. This is useful when we
have customized extent hooks reserving physical memory (e.g. 1G huge pages).
Without this feature, the default increasing sequence could result in fragmented
and wasted physical memory.
We observed that arena 0 can have much more metadata allocated comparing to
other arenas. Tune the auto mode to only switch to huge page on the 5th block
(instead of 3 previously) for a0.
On x86 Linux, we define our own MADV_FREE if madvise(2) is available, but no
MADV_FREE is detected. This allows the feature to be built in and enabled with
runtime detection.
Quoting from https://github.com/jemalloc/jemalloc/issues/761 :
[...] reading the Power ISA documentation[1], the assembly in [the CPU_SPINWAIT
macro] isn't correct anyway (as @marxin points out): the setting of the
program-priority register is "sticky", and we never undo the lowering.
We could do something similar, but given that we don't have testing here in the
first place, I'm inclined to simply not try. I'll put something up reverting the
problematic commit tomorrow.
[1] Book II, chapter 3 of the 2.07B or 3.0B ISA documents.
There does not seem to be any overlap between usage of
extent_avail and extent_heap, so we can use the same hook.
The only remaining usage of rb trees is in the profiling code,
which has some 'interesting' iteration constraints.
Fixes#888
In userspace ARM on Linux, zero-ing the high bits is the correct way to do this.
This doesn't fix the fact that we currently set LG_VADDR to 48 on ARM, when in
fact larger virtual address sizes are coming soon. We'll cross that bridge when
we come to it.
If we guarantee no malloc activity in extent hooks, it's possible to make
customized hooks working on arena 0. Remove the non-a0 assertion to enable such
use cases.
To avoid the high RSS caused by THP + low usage arena (i.e. THP becomes a
significant percentage), added a new "auto" option which will only start using
THP after a base allocator used up the first THP region. Starting from the
second hugepage (in a single arena), "auto" behaves the same as "always",
i.e. madvise hugepage right away.
This eliminates the need for the arena stats code to "know" about tcaches; all
that it needs is a cache_bin_array_descriptor_t to tell it where to find
cache_bins whose stats it should aggregate.
This is the first step towards breaking up the tcache and arena (since they
interact primarily at the bin level). It should also make a future arena
caching implementation more straightforward.
As part of the metadata_thp support, We now have a separate swtich
(JEMALLOC_HAVE_MADVISE_HUGE) for MADV_HUGEPAGE availability. Use that instead
of JEMALLOC_THP (which doesn't guard pages_huge anymore) in tests.
The external linkage for spin_adaptive was not used, and the inline
declaration of spin_adaptive that was used caused a probem on FreeBSD
where CPU_SPINWAIT is implemented as a call to a static procedure for
x86 architectures.
Currently we have to log by writing something like:
static log_var_t log_a_b_c = LOG_VAR_INIT("a.b.c");
log (log_a_b_c, "msg");
This is sort of annoying. Let's just write:
log("a.b.c", "msg");
Currently, the log macro requires at least one argument after the format string,
because of the way the preprocessor handles varargs macros. We can hide some of
that irritation by pushing the extra arguments into a varargs function.
Passing is_background_thread down the decay path, so that background thread
itself won't attempt inactivity_check. This fixes an issue with background
thread doing trylock on a mutex it already owns.
We use the minimal_initilized tsd (which requires no cleanup) for free()
specifically, if tsd hasn't been initialized yet.
Any other activity will transit the state from minimal to normal. This is to
workaround the case where a thread has no malloc calls in its lifetime until
during thread termination, free() happens after tls destructors.
To avoid complications, avoid invoking pthread_create "internally", instead rely
on thread0 to launch new threads, and also terminating threads when asked.
Avoid holding arenas_lock and background_thread_lock when creating background
threads, because pthread_create may take internal locks, and potentially cause
deadlock with jemalloc internal locks.
Fix management of extent_grow_next to serialize operations that may grow
retained memory. This assures that the sizes of the newly allocated
extents correspond to the size classes in the intended growth sequence.
Fix management of extent_grow_next to skip size classes if a request is
too large to be satisfied by the next size in the growth sequence. This
avoids the potential for an arbitrary number of requests to bypass
triggering extent_grow_next increases.
This resolves#858.
When # of dirty pages move below npages_limit (e.g. they are reused), we should
not lower number of unpurged pages because that would cause the reused pages to
be double counted in the backlog (as a result, decay happen slower than it
should). Instead, set number of unpurged to the greater of current npages and
npages_limit.
Added an assertion: the ceiling # of pages should be greater than npages_limit.
To avoid background threads sleeping forever with idle arenas, we eagerly check
background threads' sleep time after extents_dalloc, and signal the thread if
necessary.
Added opt.background_thread to enable background threads, which handles purging
currently. When enabled, decay ticks will not trigger purging (which will be
left to the background threads). We limit the max number of threads to NCPUs.
When percpu arena is enabled, set CPU affinity for the background threads as
well.
The sleep interval of background threads is dynamic and determined by computing
number of pages to purge in the future (based on backlog).
Instead of embedding a lock bit in rtree leaf elements, we associate extents
with a small set of mutexes. This gets us two things:
- We can use the system mutexes. This (hypothetically) protects us from
priority inversion, and lets us stop doing a backoff/sleep loop, instead
opting for precise wakeups from the mutex.
- Cuts down on the number of mutex acquisitions we have to do (from 4 in the
worst case to two).
We end up simplifying most of the rtree code (which no longer has to deal with
locking or concurrency at all), at the cost of additional complexity in the
extent code: since the mutex protecting the rtree leaf elements is determined by
reading the extent out of those elements, the initial read is racy, so that we
may acquire an out of date mutex. We re-check the extent in the leaf after
acquiring the mutex to protect us from this race.
This lets us specify whether and how mutexes of the same rank are allowed to be
acquired. Currently, we only allow two polices (only a single mutex at a given
rank at a time, and mutexes acquired in ascending order), but we can plausibly
allow more (e.g. the "release uncontended mutexes before blocking").
Support millisecond resolution for decay times. Among other use cases
this makes it possible to specify a short initial dirty-->muzzy decay
phase, followed by a longer muzzy-->clean decay phase.
This resolves#812.
Rather than using a manually maintained list of internal symbols to
drive name mangling, add a compilation phase to automatically extract
the list of internal symbols.
This resolves#677.
Instead, always define function pointers for interceptable functions,
but mark them const unless testing, so that the compiler can optimize
out the pointer dereferences.
Redeclaration causes compilations failures with e.g. gcc 4.2.1 on
FreeBSD. This regression was introduced by
89e2d3c12b (Header refactoring: ctl -
unify and remove from catchall.).
Re-read the leaf element when atomic CAS fails due to a race with
another thread that has locked the leaf element, since
atomic_compare_exchange_strong_p() overwrites the expected value with
the actual value on failure. This regression was introduced by
0ee0e0c155 (Implement compact rtree leaf
element representation.).
This resolves#798.
Refactor rtree_leaf_elm_extent_write() as
rtree_leaf_elm_extent_lock_write(), so that whether the leaf element is
currently acquired is separate from what lock state to write. This
allows for a relaxed atomic read when releasing the lock.
This removes the tsd macros (which are used only for tsd_t in real builds). We
break up the circular dependencies involving tsd.
We also move all tsd access through getters and setters. This allows us to
assert that we only touch data when tsd is in a valid state.
We simplify the usages of the x macro trick, removing all the customizability
(get/set, init, cleanup), moving the lifetime logic to tsd_init and tsd_cleanup.
This lets us make initialization order independent of order within tsd_t.
Add the extent_destroy_t extent destruction hook to extent_hooks_t, and
use it during arena destruction. This hook explicitly communicates to
the callee that the extent must be destroyed or tracked for later reuse,
lest it be permanently leaked. Prior to this change, retained extents
could unintentionally be leaked if extent retention was enabled.
This resolves#560.
This reverts commit b0c2a28280. Production
benchmark shows this caused significant regression in both CPU and memory
consumption. Will investigate separately later on.
Control use of munmap(2) via a run-time option rather than a
compile-time option (with the same per platform default). The old
behavior of --disable-munmap can be achieved with
--with-malloc-conf=munmap:false.
This partially resolves#580.
The explicit compiler warning suppression controlled by this option is
universally desirable, so remove the ability to disable suppression.
This partially resolves#580.
This can catch bugs in which one header defines a numeric constant, and another
uses it without including the defining header. Undefined preprocessor symbols
expand to '0', so that this will compile fine, silently doing the math wrong.
Continue to use ivsalloc() when --enable-debug is specified (and add
assertions to guard against 0 size), but stop providing a documented
explicit semantics-changing band-aid to dodge undefined behavior in
sallocx() and malloc_usable_size(). ivsalloc() remains compiled in,
unlike when #211 restored --enable-ivsalloc, and if
JEMALLOC_FORCE_IVSALLOC is defined during compilation, sallocx() and
malloc_usable_size() will still use ivsalloc().
This partially resolves#580.
Some architectures like AArch64 may not have the open syscall because it
was superseded by the openat syscall, so check and use SYS_openat if
SYS_open is not available.
Additionally, Android headers for AArch64 define SYS_open to __NR_open,
even though __NR_open is undefined. Undefine SYS_open in that case so
SYS_openat is used.
Simplify configuration by removing the --disable-tcache option, but
replace the testing for that configuration with
--with-malloc-conf=tcache:false.
Fix the thread.arena and thread.tcache.flush mallctls to work correctly
if tcache is disabled.
This partially resolves#580.
Tracking extents is required by arena_reset. To support this, the extent
linkage was used for tracking 1) large allocations, and 2) full slabs. However
modifying the extent linkage could be an expensive operation as it likely incurs
cache misses. Since we forbid arena_reset on auto arenas, let's bypass the
linkage operations for auto arenas.
All mappings continue to be PAGE-aligned, even if the system page size
is smaller. This change is primarily intended to provide a mechanism
for supporting multiple page sizes with the same binary; smaller page
sizes work better in conjunction with jemalloc's design.
This resolves#467.
Some systems use a native 64 KiB page size, which means that the bitmap
for the smallest size class can be 8192 bits, not just 512 bits as when
the page size is 4 KiB. Linear search in bitmap_{sfu,ffu}() is
unacceptably slow for such large bitmaps.
This reverts commit 7c00f04ff4.
Rather than using a LIFO queue to track available extent_t structures,
use a red-black tree, and always choose the oldest/lowest available
during reuse.
Two levels of rcache is implemented: a direct mapped cache as L1, combined with
a LRU cache as L2. The L1 cache offers low cost on cache hit, but could suffer
collision under circumstances. This is complemented by the L2 LRU cache, which
is slower on cache access (overhead from linear search + reordering), but solves
collison of L1 rather well.
Previously we had a general detection and support of reentrancy, at the cost of
having branches and inc / dec operations on fast paths. To avoid taxing fast
paths, we move the reentrancy operations onto tsd slow state, and only modify
reentrancy level around external calls (that might trigger reentrancy).
Added tsd_state_nominal_slow, which on fast path malloc() incorporates
tcache_enabled check, and on fast path free() bundles both malloc_slow and
tcache_enabled branches.
With this change, when profiling is enabled, we avoid doing redundant rtree
lookups. Also changed dalloc_atx_t to alloc_atx_t, as it's now used on
allocation path as well (to speed up profiling).
This is a biggy. jemalloc_internal.h has been doing multiple jobs for a while
now:
- The source of system-wide definitions.
- The catch-all include file.
- The module header file for jemalloc.c
This commit splits up this functionality. The system-wide definitions
responsibility has moved to jemalloc_preamble.h. The catch-all include file is
now jemalloc_internal_includes.h. The module headers for jemalloc.c are now in
jemalloc_internal_[externs|inlines|types].h, just as they are for the other
modules.
This checks whether or not we're reentrant using thread-local data, and, if we
are, moves certain internal allocations to use arena 0 (which should be properly
initialized after bootstrapping).
The immediate thing this allows is spinning up threads in arena_new, which will
enable spinning up background threads there.
1) Re-organize TSD so that frequently accessed fields are closer to the
beginning and more compact. Assuming 64-bit, the first 2.5 cachelines now
contains everything needed on tcache fast path, expect the tcache struct itself.
2) Re-organize tcache and tbins. Take lg_fill_div out of tbin, and reduce tbin
to 24 bytes (down from 32). Split tbins into tbins_small and tbins_large, and
place tbins_small close to the beginning.
The embedded tcache is initialized upon tsd initialization. The avail arrays
for the tbins will be allocated / deallocated accordingly during init / cleanup.
With this change, the pointer to the auto tcache will always be available, as
long as we have access to the TSD. tcache_available() (called in tcache_get())
is provided to check if we should use tcache.
This will facilitate embedding tcache into tsd, which will require proper
initialization cannot be done via the static initializer. Make tsd->rtree_ctx
to be initialized via rtree_ctx_data_init().
Compact extent_t to 128 bytes on 64-bit systems by moving
arena_slab_data_t's nfree into extent_t's e_bits.
Cacheline-align extent_t structures so that they always cross the
minimum number of cacheline boundaries.
Re-order extent_t fields such that all fields except the slab bitmap
(and overlaid heap profiling context pointer) are in the first
cacheline.
This resolves#461.
Remove tree-structured bitmap support, in order to reduce complexity and
ease maintenance. No bitmaps larger than 512 bits have been necessary
since before 4.0.0, and there is no current plan that would increase
maximum bitmap size. Although tree-structured bitmaps were used on
32-bit platforms prior to this change, the overall benefits were
questionable (higher metadata overhead, higher bitmap modification cost,
marginally lower search cost).
This fixes an extent searching regression on 32-bit systems, caused by
the initial bitmap_ffu() implementation in
c8021d01f6 (Implement bitmap_ffu(), which
finds the first unset bit.), as first used in
5d33233a5e (Use a bitmap in extents_t to
speed up search.).
A fixed max spin count is used -- with benchmark results showing it
solves almost all problems. As the benchmark used was rather intense,
the upper bound could be a little bit high. However it should offer a
good tradeoff between spinning and blocking.
Use tsd_rtree_ctx() rather than tsdn_rtree_ctx() when tcache is
non-NULL, in order to avoid an extra branch (and potentially extra stack
space) in the fast path.
If a single virtual adddress pointer has enough unused bits to pack
{szind_t, extent_t *, bool, bool}, use a single pointer-sized field in
each rtree leaf element, rather than using three separate fields. This
has little impact on access speed (fewer loads/stores, but more bit
twiddling), except that denser representation increases TLB
effectiveness.
Expand and restructure the rtree API such that all common operations can
be achieved with minimal work, regardless of whether the rtree leaf
fields are independent versus packed into a single atomic pointer.
This allows leaf elements to differ in size from internal node elements.
In principle it would be more correct to use a different type for each
level of the tree, but due to implementation details related to atomic
operations, we use casts anyway, thus counteracting the value of
additional type correctness. Furthermore, such a scheme would require
function code generation (via cpp macros), as well as either unwieldy
type names for leaves or type aliases, e.g.
typedef struct rtree_elm_d2_s rtree_leaf_elm_t;
This alternate strategy would be more correct, and with less code
duplication, but probably not worth the complexity.
Rather than storing usize only for large (and prof-promoted)
allocations, store the size class index for allocations that reside
within the extent, such that the size class index is valid for all
extents that contain extant allocations, and invalid otherwise (mainly
to make debugging simpler).
Split decay-based purging into two phases, the first of which uses lazy
purging to convert dirty pages to "muzzy", and the second of which uses
forced purging, decommit, or unmapping to convert pages to clean or
destroy them altogether. Not all operating systems support lazy
purging, yet the application may provide extent hooks that implement
lazy purging, so care must be taken to dynamically omit the first phase
when necessary.
The mallctl interfaces change as follows:
- opt.decay_time --> opt.{dirty,muzzy}_decay_time
- arena.<i>.decay_time --> arena.<i>.{dirty,muzzy}_decay_time
- arenas.decay_time --> arenas.{dirty,muzzy}_decay_time
- stats.arenas.<i>.pdirty --> stats.arenas.<i>.p{dirty,muzzy}
- stats.arenas.<i>.{npurge,nmadvise,purged} -->
stats.arenas.<i>.{dirty,muzzy}_{npurge,nmadvise,purged}
This resolves#521.
Refactor most of the decay-related functions to take as parameters the
decay_t and associated extents_t structures to operate on. This
prepares for supporting both lazy and forced purging on different decay
schedules.
These were all size_ts, so we have atomics support for them on all platforms, so
the conversion is straightforward.
Left non-atomic is curlextents, which AFAICT is not used atomically anywhere.
I expect this to be the trickiest conversion we will see, since we want atomics
on 64-bit platforms, but are also always able to piggyback on some sort of
external synchronization on non-64 bit platforms.
In the process, I changed the implementation of rtree_elm_acquire so that it
won't even try to CAS if its initial read (getting the extent + lock bit)
indicates that the CAS is doomed to fail. This can significantly improve
performance under contention.
The new feature, opt.percpu_arena, determines thread-arena association
dynamically based CPU id. Three modes are supported: "percpu", "phycpu"
and disabled.
"percpu" uses the current core id (with help from sched_getcpu())
directly as the arena index, while "phycpu" will assign threads on the
same physical CPU to the same arena. In other words, "percpu" means # of
arenas == # of CPUs, while "phycpu" has # of arenas == 1/2 * (# of
CPUs). Note that no runtime check on whether hyper threading is enabled
is added yet.
When enabled, threads will be migrated between arenas when a CPU change
is detected. In the current design, to reduce overhead from reading CPU
id, each arena tracks the thread accessed most recently. When a new
thread comes in, we will read CPU id and update arena if necessary.
When witness is enabled, lock rank order needs to be preserved during
prefork, not only for each arena, but also across arenas. This change
breaks arena_prefork into further stages to ensure valid rank order
across arenas. Also changed test/unit/fork to use a manual arena to
catch this case.
In the process, we can do some strength reduction, changing the fetch-adds and
fetch-subs to be simple loads followed by stores, since the modifications all
occur while holding the mutex.
The C11 atomics backport removed this #define, which degraded atomic 64-bit
reads to require a lock even on platforms that support them. This commit fixes
that.
This fixes tcache_flush for manual tcaches, which wasn't able to find
the correct arena it associated with. Also changed the decay test to
cover this case (by using manually created arenas).
These functions select the easiest-to-remove element in the heap, which
is either the most recently inserted aux list element or the root. If
no calls are made to first() or remove_first(), the behavior (and time
complexity) is the same as for a LIFO queue.
Rather than purging uncoalesced extents, perform just enough incremental
coalescing to purge only fully coalesced extents. In the absence of
cached extent reuse, the immediate versus delayed incremental purging
algorithms result in the same purge order.
This resolves#655.
In the C11 atomics backport, we couldn't use not_reached() in
atomic_enum_to_builtin (in atomic_gcc_atomic.h), since atomic.h was hermetic and
assert.h wasn't; there was a dependency issue. assert.h is hermetic now, so we
can include it.
This is the first header refactoring diff, #533. It splits the assert and util
components into separate, hermetic, header files. In the process, it splits out
two of the large sub-components of util (the stdio.h replacement, and bit
manipulation routines) into their own components (malloc_io.h and bit_util.h).
This is mostly to break up cyclic dependencies, but it also breaks off a good
chunk of the catch-all-ness of util, which is nice.
Convert the nrequests field to be partially derived, and the curlextents
to be fully derived, in order to reduce the number of stats updates
needed during common operations.
This change affects ndalloc stats during arena reset, because it is no
longer possible to cancel out ndalloc effects (curlextents would become
negative).
This introduces a backport of C11 atomics. It has four implementations; ranked
in order of preference, they are:
- GCC/Clang __atomic builtins
- GCC/Clang __sync builtins
- MSVC _Interlocked builtins
- C11 atomics, from <stdatomic.h>
The primary advantages are:
- Close adherence to the standard API gives us a defined memory model.
- Type safety: atomic objects are now separate types from non-atomic ones, so
that it's impossible to mix up atomic and non-atomic updates (which is
undefined behavior that compilers are starting to take advantage of).
- Efficiency: we can specify ordering for operations, avoiding fences and
atomic operations on strongly ordered architectures (example:
`atomic_write_u32(ptr, val);` involves a CAS loop, whereas
`atomic_store(ptr, val, ATOMIC_RELEASE);` is a plain store.
This diff leaves in the current atomics API (implementing them in terms of the
backport). This lets us transition uses over piecemeal.
Testing:
This is by nature hard to test. I've manually tested the first three options on
Linux on gcc by futzing with the #defines manually, on freebsd with gcc and
clang, on MSVC, and on OS X with clang. All of these were x86 machines though,
and we don't have any test infrastructure set up for non-x86 platforms.
In the long term, we'll transition to C99-style inline semantics. In the
short-term, this will allow both styles to coexist without breaking one another.
Remove obsolete unit test scaffolding for extent quantization. Remove
redundant assertions. Add an assertion to
extents_first_best_fit_locked() that should help prevent aligned
allocation regressions.
We don't touch witness at all when config_debug == false. Let's only pay the
memory cost in malloc_mutex_s when needed. Note that when !config_debug, we keep
the field in a union so that we don't have to do #ifdefs in multiple places.
Extent splitting and coalescing is a major component of large allocation
overhead, and disabling coalescing of cached extents provides a simple
and effective hysteresis mechanism. Once two-phase purging is
implemented, it will probably make sense to leave coalescing disabled
for the first phase, but coalesce during the second phase.
This avoids a gcc diagnostic note:
note: The ABI for passing parameters with 64-byte alignment has
changed in GCC 4.6
This note related to the cacheline alignment of rtree_ctx_t, which was
introduced by 4a346f5593 (Replace rtree
path cache with LRU cache.).
Fix rtree_subkey() to use uintptr_t rather than unsigned for key
bitmasking. This regression was introduced by
4a346f5593 (Replace rtree path cache with
LRU cache.).
Rather than dynamically building a table to aid per level computations,
define a constant table at compile time. Omit both high and low
insignificant bits. Use one to three tree levels, depending on the
number of significant bits.
Rework rtree_ctx_t to encapsulate an rtree leaf LRU lookup cache rather
than a single-path element lookup cache. The replacement is logically
much simpler, as well as slightly faster in the fast path case and less
prone to degraded performance during non-trivial sequences of lookups.
Refactor arena and extent locking protocols such that arena and
extent locks are never held when calling into the extent_*_wrapper()
API. This requires extra care during purging since the arena lock no
longer protects the inner purging logic. It also requires extra care to
protect extents from being merged with adjacent extents.
Convert extent_t's 'active' flag to an enumerated 'state', so that
retained extents are explicitly marked as such, rather than depending on
ring linkage state.
Refactor the extent collections (and their synchronization) for cached
and retained extents into extents_t. Incorporate LRU functionality to
support purging. Incorporate page count accounting, which replaces
arena->ndirty and arena->stats.retained.
Assert that no core locks are held when entering any internal
[de]allocation functions. This is in addition to existing assertions
that no locks are held when entering external [de]allocation functions.
Audit and document synchronization protocols for all arena_t fields.
This fixes a potential deadlock due to recursive allocation during
gdump, in a similar fashion to b49c649bc1
(Fix lock order reversal during gdump.), but with a necessarily much
broader code impact.
