Converting size to usize is what jemalloc has been done by ceiling
size to the closest size class. However, this causes lots of memory
wastes with HPA enabled. This commit changes how usize is calculated so
that the gap between two contiguous usize is no larger than a page.
Specifically, this commit includes the following changes:
1. Adding a build-time config option (--enable-limit-usize-gap) and a
runtime one (limit_usize_gap) to guard the changes.
When build-time
config is enabled, some minor CPU overhead is expected because usize
will be stored and accessed apart from index. When runtime option is
also enabled (it can only be enabled with the build-time config
enabled). a new usize calculation approach wil be employed. This new
calculation will ceil size to the closest multiple of PAGE for all sizes
larger than USIZE_GROW_SLOW_THRESHOLD instead of using the size classes.
Note when the build-time config is enabled, the runtime option is
default on.
2. Prepare tcache for size to grow by PAGE over GROUP*PAGE.
To prepare for the upcoming changes where size class grows by PAGE when
larger than NGROUP * PAGE, disable the tcache when it is larger than 2 *
NGROUP * PAGE. The threshold for tcache is set higher to prevent perf
regression as much as possible while usizes between NGROUP * PAGE and 2 *
NGROUP * PAGE happen to grow by PAGE.
3. Prepare pac and hpa psset for size to grow by PAGE over GROUP*PAGE
For PAC, to avoid having too many bins, arena bins still have the same
layout. This means some extra search is needed for a page-level request that
is not aligned with the orginal size class: it should also search the heap
before the current index since the previous heap might also be able to
have some allocations satisfying it. The same changes apply to HPA's
psset.
This search relies on the enumeration of the heap because not all allocs in
the previous heap are guaranteed to satisfy the request. To balance the
memory and CPU overhead, we currently enumerate at most a fixed number
of nodes before concluding none can satisfy the request during an
enumeration.
4. Add bytes counter to arena large stats.
To prepare for the upcoming usize changes, stats collected by
multiplying alive allocations and the bin size is no longer accurate.
Thus, add separate counters to record the bytes malloced and dalloced.
5. Change structs use when freeing to avoid using index2size for large sizes.
- Change the definition of emap_alloc_ctx_t
- Change the read of both from edata_t.
- Change the assignment and usage of emap_alloc_ctx_t.
- Change other callsites of index2size.
Note for the changes in the data structure, i.e., emap_alloc_ctx_t,
will be used when the build-time config (--enable-limit-usize-gap) is
enabled but they will store the same value as index2size(szind) if the
runtime option (opt_limit_usize_gap) is not enabled.
6. Adapt hpa to the usize changes.
Change the settings in sec to limit is usage for sizes larger than
USIZE_GROW_SLOW_THRESHOLD and modify corresponding tests.
7. Modify usize calculation and corresponding tests.
Change the sz_s2u_compute. Note sz_index2size is not always safe now
while sz_size2index still works as expected.
This lets us easily see what fraction of flush load is being taken up by the
bins, and helps guide future optimization approaches (for example: should we
prefetch during cache bin fills? It depends on how many objects the average fill
pops out of the batch).
This adds a fast-path for threads freeing a small number of allocations to
bins which are not their "home-base" and which encounter lock contention in
attempting to do so. In producer-consumer workflows, such small lock hold times
can cause lock convoying that greatly increases overall bin mutex contention.
This accomplishes two things:
- It avoids a full array scan (and any attendant branch prediction misses, etc.)
while holding the bin lock.
- It allows us to know the number of items that will be flushed before flushing
them, which will (in an upcoming commit) let us know if it's safe to use the
batched flush (in which case we won't acquire the bin mutex).
The main bits of shared code are the edata filtering and the stats flushing
logic, both of which are fairly simple to read and not so painful to duplicate.
The shared code comes at the cost of guarding all the subtle logic with
`if (small)`, which doesn't feel worth it.
When there were many items stashed, it's possible that after flushing stashed,
ncached is already lower than the remain, in which case the flush can simply
return at that point.
1. Pre-generate all default tcache ncached_max in tcache_boot;
2. Add getters returning default ncached_max and ncached_max_set;
3. Refactor tcache init so that it is always init with a given setting.
1. `thread_tcache_ncached_max_read_sizeclass` allows users to get the
ncached_max of the bin with the input sizeclass, passed in through
oldp (will be upper casted if not an exact bin size is given).
2. `thread_tcache_ncached_max_write` takes in a char array
representing the settings for bins in the tcache.
When using metadata_thp, allocate tcache bin stacks from base0, which means they
will be placed on huge pages along with other metadata, instead of mixed with
other regular allocations.
In order to do so, modified the base allocator to support limited reuse: freed
tcached stacks (from thread termination) will be returned to base0 and made
available for reuse, but no merging will be attempted since they were bump
allocated out of base blocks. These reused base extents are managed using
separately allocated base edata_t -- they are cached in base->edata_avail when
the extent is all allocated.
One tricky part is, stats updating must be skipped for such reused extents
(since they were accounted for already, and there is no purging for base). This
requires tracking the "if is reused" state explicitly and bypass the stats
updates when allocating from them.
1. add tcache_max and nhbins into tcache_t so that they are per-tcache,
with one auto tcache per thread, it's also per-thread;
2. add mallctl for each thread to set its own tcache_max (of its auto tcache);
3. store the maximum number of items in each bin instead of using a global storage;
4. add tests for the modifications above.
5. Rename `nhbins` and `tcache_maxclass` to `global_do_not_change_nhbins` and `global_do_not_change_tcache_maxclass`.
Following from PR #2481, we replace all integer-to-pointer casts [which
hide pointer provenance information (and thus inhibit
optimizations)](https://clang.llvm.org/extra/clang-tidy/checks/performance/no-int-to-ptr.html)
with equivalent operations that preserve this information. I have
enabled the corresponding clang-tidy check in our static analysis CI so
that we do not get bitten by this again in the future.
On deallocation, sampled pointers (specially aligned) get junked and stashed
into tcache (to prevent immediate reuse). The expected behavior is to have
read-after-free corrupted and stopped by the junk-filling, while
write-after-free is checked when flushing the stashed pointers.
This saves us a cache miss when lookup up the arena bin offset in a remote
arena during tcache flush. All arenas share the base offset, and so we don't
need to look it up repeatedly for each arena. Secondarily, it shaves 288 bytes
off the arena on, e.g., x86-64.
The items we pick to flush matter a lot, but the order in which they get flushed
doesn't; just use forward scans. This simplifies the accessing code, both in
terms of the C and the generated assembly (i.e. this speeds up the flush
pathways).
By carefully force-inlining the division constants and the operation sum count,
we can eliminate redundant operations in the arena-level dalloc function. Do
so.
This frontloads more of the miss latency. It also moves it to a pathway where
we have not yet acquired any locks, so that it should (hopefully) reduce hold
times.
In practice, many rtree_leaf_elm accesses are cache misses. By restructuring,
we can make it more likely that these misses occur without blocking us from
starting later lookups, taking more of those misses in parallel.
This fixes an incorrect debug-mode assert:
- T1 starts an arena stats update and reads stack_head from another thread's
cache bin, when that cache bin has 1 item in it.
- T2 allocates from that cache bin. The cache_bin's stack_head now points to a
NULL pointer, since the cache bin is empty.
- T1 Re-reads the cache_bin's stack_head to perform an assertion check (since it
previously saw that the bin was empty, whatever stack_head points to should be
non-NULL).
Previously all the small size classes were cached. However this has downsides
-- particularly when page size is greater than 4K (e.g. iOS), which will result
in much higher SMALL_MAXCLASS.
This change allows tcache_max to be set to lower values, to better control
resources taken by tcache.
For locality reasons, tcache bins are integrated in TSD. Allowing all size
classes to be cached has little benefit, but takes up much thread local storage.
In addition, it complicates the layout which we try hard to optimize.
Without a lock held continuously between checking tcaches_past and incrementing
it, it's possible for two threads to go down manual creation path
simultaneously. If the number of tcaches is one less than the maximum, it's
possible for both to create a tcache and increment tcaches_past, with the second
thread returning a value larger than TCACHES_MAX.
In making these settings configurable, 634afc4124
unintentially changed a tuning parameter (reducing the "goal" max by a factor of
4). This commit undoes that change.
guangli-dai