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.
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`.
There are several benefits to this:
1. It's cleaner and more reliable to use the builtin to
inform the compiler of assumptions instead of hoping that the
optimizer understands your intentions.
2. `clang` will warn you if any of your assumptions would produce
side-effects (which the compiler will discard). [This blog post](https://fastcompression.blogspot.com/2019/01/compiler-checked-contracts.html)
by Yann Collet highlights that a hazard of using the
`unreachable()`-based method of signaling assumptions is that it
can sometimes result in additional instructions being generated (see
[this Godbolt link](https://godbolt.org/z/lKNMs3) from the blog post
for an example).
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.
Header files are now self-contained, which makes the relationships
between the files clearer, and crucially allows LSP tools like `clangd`
to function correctly in all of our header files. I have verified that
the headers are self-contained (aside from the various Windows shims) by
compiling them as if they were C files – in a follow-up commit I plan to
add this to CI to ensure we don't regress on this front.
Previously, small allocations which were sampled as part of heap
profiling were rounded up to `SC_LARGE_MINCLASS`. This additional memory
usage becomes problematic when the page size is increased, as noted in #2358.
Small allocations are now rounded up to the nearest multiple of `PAGE`
instead, reducing the memory overhead by a factor of 4 in the most
extreme cases.
In #2433, I inadvertently introduced a regression which causes the use of
uninitialized data. Namely, the control path I added for the safety
check in `arena_prof_info_get` neglected to set `prof_info->alloc_tctx`
when the check fails, resulting in `prof_info.alloc_tctx` being
uninitialized [when it is read at the end of
`prof_free`](90176f8a87/include/jemalloc/internal/prof_inlines.h (L272)).
Static analysis flagged this. It's possible to segfault in the
`*_tree_remove` function generated by `rb_gen`, as `nodep` may
still be `NULL` after the initial for loop. I can confirm from reviewing
the fleetwide coredump data that this was in fact being hit in
production, primarily through `tctx_tree_remove`, and much more rarely
through `gctx_tree_remove`.
Decay should not be triggered during reentrant calls (may cause lock order
reversal / deadlocks). Added a delay_trigger flag to the tickers to bypass
decay when rentrancy_level is not zero.
Adding guarded extents, which are regular extents surrounded by guard pages
(mprotected). To reduce syscalls, small guarded extents are cached as a
separate eset in ecache, and decay through the dirty / muzzy / retained pipeline
as usual.
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.
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 pulls out the various abstractions where some stats counter is sometimes an
atomic, sometimes a plain variable, sometimes always protected by a lock,
sometimes protected by reads but not writes, etc. With this change, these cases
are treated consistently, and access patterns tagged.
In the process, we fix a few missed-update bugs (where one caller assumes
"protected-by-a-lock" semantics and another does not).
This is logically at a higher level of the stack; extent should just allocate
things at the page-level; it shouldn't care exactly why the callers wants a
given number of pages.
When deferred initialization was added, initializing required copying
sizeof(extent_hooks_t) bytes after a pointer chase. Today, it's just a single
pointer loaded from the base_t. In subsequent diffs, we'll get rid of even that.
Explicitly define three setters:
- `prof_tctx_reset()`: set `prof_tctx` to `1U`, if we don't know in
advance whether the allocation is large or not;
- `prof_tctx_reset_sampled()`: set `prof_tctx` to `1U`, if we already
know in advance that the allocation is large;
- `prof_info_set()`: set a real `prof_tctx`, and also set other
profiling info e.g. the allocation time.
Code structure wise, the prof level is kept as a thin wrapper, the
large level only provides low level setter APIs, and the arena level
carries out the main logic.
Develop new data structure and code logic for holding profiling
related information stored in the extent that may be needed after the
extent is released, which in particular is the case for the
reallocation code path (e.g. in `rallocx()` and `xallocx()`). The
data structure is a generalization of `prof_tctx_t`: we previously
only copy out the `prof_tctx` before the extent is released, but we
may be in need of additional fields. Currently the only additional
field is the allocation time field, but there may be more fields in
the future.
The restructuring also resolved a bug: `prof_realloc()` mistakenly
passed the new `ptr` to `prof_free_sampled_object()`, but passing in
the `old_ptr` would crash because it's already been released. Now
the essential profiling information is collectively copied out early
and safely passed to `prof_free_sampled_object()` after the extent is
released.
