The bug is subtle but critical: if application performs the following
three actions in sequence: (a) turn `prof_active` off, (b) make at
least one allocation that triggers the malloc slow path via the
`if (unlikely(bytes_until_sample < 0))` path, and (c) turn
`prof_active` back on, then the application would never get another
sample (until a very very long time later).
The fix is to properly reset `bytes_until_sample` rather than
throwing it all the way to `SSIZE_MAX`.
A side minor change is to call `prof_active_get_unlocked()` rather
than directly grabbing the `prof_active` variable - it is the very
reason why we defined the `prof_active_get_unlocked()` function.
The -1 value of low_water indicates if the cache has been depleted and
refilled. Track the status explicitly in the tcache struct.
This allows the fast path to check if (cur_ptr > low_water), instead of >=,
which avoids reaching slow path when the last item is allocated.
With the cache bin metadata switched to pointers, ncached_max is usually
accessed and timed by sizeof(ptr). Store the results in tcache_bin_info for
direct access, and add a helper function for the ncached_max value.
Implement the pointer-based metadata for tcache bins --
- 3 pointers are maintained to represent each bin;
- 2 of the pointers are compressed on 64-bit;
- is_full / is_empty done through pointer comparison;
Comparing to the previous counter based design --
- fast-path speed up ~15% in benchmarks
- direct pointer comparison and de-reference
- no need to access tcache_bin_info in common case
Without buffering `malloc_stats_print` would invoke the write back
call (which could mean an expensive `malloc_write_fd` call) for every
single `printf` (including printing each line break and each leading
tab/space for indentation).
`tcache->prof_accumbytes` should always be cleared after being
transferred to arena; otherwise the allocations would be double
counted, leading to excessive prof dumps.
Refactored core profiling codebase into two logical parts:
(a) `prof_data.c`: core internal data structure managing & dumping;
(b) `prof.c`: mutexes & outward-facing APIs.
Some internal functions had to be exposed out, but there are not
that many of them if the modularization is (hopefully) clean enough.
Prof logging is conceptually seperate from core profiling, so
split it out as a module of its own. There are a few internal
functions that had to be exposed but I think it is a fair trade-off.
W/o retain, split and merge are disallowed on Windows. Avoid doing first-fit
which needs splitting almost always. Instead, try exact fit only and bail out
early.
Refactored core profiling codebase into two logical parts:
(a) `prof_data.c`: core internal data structure managing & dumping;
(b) `prof.c`: mutexes & outward-facing APIs.
Some internal functions had to be exposed out, but there are not
that many of them if the modularization is (hopefully) clean enough.
`prof.c` is growing too long, so trying to modularize it. There are
a few internal functions that had to be exposed but I think it is a
fair trade-off.
extent_register may only fail if the underlying extent and region got stolen /
coalesced before we lock. Avoid doing extent_leak (which purges the region)
since we don't really own the region.
This can only happen on Windows and with opt.retain disabled (which isn't the
default). The solution is suboptimal, however not a common case as retain is
the long term plan for all platforms anyway.
The VirtualAlloc and VirtualFree APIs are different because MEM_DECOMMIT cannot
be used across multiple VirtualAlloc regions. To properly support decommit,
only allow merge / split within the same region -- this is done by tracking the
"is_head" state of extents and not merging cross-region.
Add a new state is_head (only relevant for retain && !maps_coalesce), which is
true for the first extent in each VirtualAlloc region. Determine if two extents
can be merged based on the head state, and use serial numbers for sanity checks.
The original logic can be disastrous if `PROF_DUMP_BUFSIZE` is less
than `slen` -- `prof_dump_buf_end + slen <= PROF_DUMP_BUFSIZE` would
always be `false`, so `memcpy` would always try to copy
`PROF_DUMP_BUFSIZE - prof_dump_buf_end` chars, which can be
dangerous: in the last round of the `while` loop it would not only
illegally read the memory beyond `s` (which might not always be
disastrous), but it would also illegally overwrite the memory beyond
`prof_dump_buf` (which can be pretty disastrous). `slen` probably
has never gone beyond `PROF_DUMP_BUFSIZE` so we were just lucky.
`cbopaque` can now be overriden without overriding `write_cb` in
the first place. (Otherwise there would be no need to have the
`cbopaque` parameter in `malloc_message`.)
Background threads may run for a long time, especially when the # of dirty pages
is high. Avoid blocking stats calls because of this (which may cause latency
spikes).
The new experimental mallctl exposes the arena pactive counter to applications,
which allows fast read w/o going through the mallctl / epoch steps. This is
particularly useful when frequent balancing is required, e.g. when having
multiple manual arenas, and threads are multiplexed to them based on usage.
If the confirm_conf option is set, when the program starts, each of
the four malloc_conf strings will be printed, and each option will
be printed when being set.
GCC-9.1 reports following error when trying to compile file
src/malloc_io.c and with CFLAGS='-Werror' :
src/malloc_io.c: In function ‘malloc_vsnprintf’:
src/malloc_io.c:369:2: error: case label value exceeds maximum value for type [-Werror]
369 | case '?' | 0x80: \
| ^~~~
src/malloc_io.c:581:5: note: in expansion of macro ‘GET_ARG_NUMERIC’
581 | GET_ARG_NUMERIC(val, 'p');
| ^~~~~~~~~~~~~~~
...
<snip>
cc1: all warnings being treated as errors
make: *** [Makefile:388: src/malloc_io.sym.o] Error 1
The warning is reported as by default the type 'char' is 'signed char'
and or-ing 0x80 will turn the case label char negative which will be
beyond the printable ascii range (0 - 127).
The patch fixes this by explicitly casting the 'len' variable as
unsigned char' inside the 'switch' statement so that value of
expression " '?' | 0x80 " falls within the legal values of the
variable 'len'.
Small is added purely for convenience. Large flushes wasn't tracked before and
can be useful in analysis. Large fill simply reports nmalloc, since there is no
batch fill for large currently.
This option saves a few CPU cycles, but potentially adds a lot of
fragmentation - so much so that there are workarounds like
max_active. Instead, let's just drop it entirely. It only made
a difference in one service I tested (.3% cpu regression), while
many services saw a memory win (also small, less than 1% mem P99)
When config_stats is enabled track the size of bin->slabs_nonfull in
the new nonfull_slabs counter in bin_stats_t. This metric should be
useful for establishing an upper ceiling on the savings possible by
meshing.
Summary: sdallocx is checking a flag that will never be set (at least in the provided C++ destructor implementation). This branch will probably only rarely be mispredicted however it removes two instructions in sdallocx and one at the callsite (to zero out flags).
This is discovered and suggested by @jasone in #1468. When custom extent hooks
are in use, we should ensure page alignment on the extent alloc path, instead of
relying on the user hooks to do so.
The analytics tool is put under experimental.utilization namespace in
mallctl. Input is one pointer or an array of pointers and the output
is a list of memory utilization statistics.
Proposed fix for #1444 - ensure that `tls_callback` in the `#pragma comment(linker)`directive gets the same prefix added as it does i the C declaration.
This feature uses an dedicated arena to handle huge requests, which
significantly improves VM fragmentation. In production workload we tested it
often reduces VM size by >30%.
For low arena count settings, the huge threshold feature may trigger an unwanted
bg thd creation. Given that the huge arena does eager purging by default,
bypass bg thd creation when initializing the huge arena.
When custom extent_hooks or transparent huge pages are in use, the purging
semantics may change, which means we may not get zeroed pages on repopulating.
Fixing the issue by manually memset for such cases.
This makes it possible to have multiple set of bins in an arena, which improves
arena scalability because the bins (especially the small ones) are always the
limiting factor in production workload.
A bin shard is picked on allocation; each extent tracks the bin shard id for
deallocation. The shard size will be determined using runtime options.
If there are 3 or more threads spin-waiting on the same mutex,
there will be excessive exclusive cacheline contention because
pthread_trylock() immediately tries to CAS in a new value, instead
of first checking if the lock is locked.
This diff adds a 'locked' hint flag, and we will only spin wait
without trylock()ing while set. I don't know of any other portable
way to get the same behavior as pthread_mutex_lock().
This is pretty easy to test via ttest, e.g.
./ttest1 500 3 10000 1 100
Throughput is nearly 3x as fast.
This blames to the mutex profiling changes, however, we almost never
have 3 or more threads contending in properly configured production
workloads, but still worth fixing.
Refactor tcache_fill, introducing a new function arena_slab_reg_alloc_batch,
which will fill multiple pointers from a slab.
There should be no functional changes here, but allows future optimization
on reg_alloc_batch.
Add unsized and sized deallocation fastpaths. Similar to the malloc()
fastpath, this removes all frame manipulation for the majority of
free() calls. The performance advantages here are less than that
of the malloc() fastpath, but from prod tests seems to still be half
a percent or so of improvement.
Stats and sampling a both supported (sdallocx needs a sampling check,
for rtree lookups slab will only be set for unsampled objects).
We don't support flush, any flush requests go to the slowpath.
We eagerly coalesce large buffers when deallocating, however the previous logic
around this introduced extra lock overhead -- when coalescing we always lock the
neighbors even if they are active, while for active extents nothing can be done.
This commit checks if the neighbor extents are potentially active before
locking, and avoids locking if possible. This speeds up large_dalloc by ~20%.
It also fixes some undesired behavior: we could stop coalescing because a small
buffer was merged, while a large neighbor was ignored on the other side.
When retain is enabled, the default dalloc hook does nothing (since we avoid
munmap). But the overhead preparing the call is high, specifically the extent
de-register and re-register involve locking and extent / rtree modifications.
Bypass the call with retain in this diff.
This diff adds a fastpath that assumes size <= SC_LOOKUP_MAXCLASS, and
that we hit tcache. If either of these is false, we fall back to
the previous codepath (renamed 'malloc_default').
Crucially, we only tail call malloc_default, and with the same kind
and number of arguments, so that both clang and gcc tail-calling
will kick in - therefore malloc() gets treated as a leaf function,
and there are *no* caller-saved registers. Previously malloc() contained
5 caller saved registers on x64, resulting in at least 10 extra
memory-movement instructions.
In microbenchmarks this results in up to ~10% improvement in malloc()
fastpath. In real programs, this is a ~1% CPU and latency improvement
overall.
