Skip to content

Tags

Tags

v2026.10.26.00

Toggle v2026.10.26.00's commit message
Make signal stack-trace test TSAN-compatible

Summary:
[🤖 Metamate] Make `StackTraceTest.Signal` signal-safe so it can run under TSAN rather than report a runtime skip or be compiled out.

The handler captures only `getStackTraceSafe()` into fixed storage and publishes its count with a statically verified lock-free atomic. Preserve `errno` in the handler and restore the previous signal action after the test. Perform symbolization and assertions after `raise()` returns, checking errno preservation and that the captured trace includes the handler and interrupted caller in order. `Simple` continues comparing all three stack-trace APIs outside signal context.

___

landed-with-radar-review

Differential Revision: D127472550

fbshipit-source-id: da7c1eb3b5bdc5bdf8fd6a12c88492aa6a8747c4

v2026.10.19.00

Toggle v2026.10.19.00's commit message
Cherry-pick: "Eliminate exit-time destructor in TypedTestSuitePState."

Summary:
This is a cherry-pick of commit hash b66cffcd5f5d6ee34a4f96d57bf016d9a22cb023 from googletest upstream.

Github: google/googletest@b66cffc

 The original commit summary is:

`TYPED_TEST_SUITE_P` defines a static `TypedTestSuitePState` variable at namespace scope. Storing registered tests in a `std::map<std::string, CodeLocation>` required dynamic initialization, heap allocations, and an exit-time destructor, triggering `-Wexit-time-destructors` warnings in builds that enforce trivial exit.

Store registered tests in an intrusive sorted singly-linked list of static POD `RegisteredTest` nodes defined by `TYPED_TEST_P`. `TypedTestSuitePState` is now `constexpr`-constructible and trivially destructible without heap allocations.

PiperOrigin-RevId: 998634744
Change-Id: I0c833ffeebdf21b155038b7e95e497637e68564a

### Manual Fix

Upstream replaced the internal `TypedTestSuitePState::AddTestName()` with `AddTest(RegisteredTest*, const char*)`, which takes a caller-owned static `RegisteredTest` node. Four Meta-side copies of the old `TYPED_TEST_P` macro still call `AddTestName()` and fail to compile:

```
error: no member named 'AddTestName' in 'testing::internal::TypedTestSuitePState'
```

- `xplat/follyfex/gtest/GTestHelpers.h`: `CO_TYPED_TEST_P` declares a static `RegisteredTest` and registers it with `AddTest()`, matching the new upstream `TYPED_TEST_P`. Fixes `xplat/follyfex/gtest/tests:gtestAndroid`, the `xplat/assistant/oacr/contacts` warm-start tests, and `xplat/rtc/crucible/tests:callSetupFoundation` / `:callSetupRsys`.
- `fbcode/folly/coro/GtestHelpers.h`, `xplat/folly/coro/GtestHelpers.h` (identical copies): `CO_TYPED_TEST_P_` registers through the new `folly::detail::gtestAddTypedTest()`, which calls `AddTest()` when `TypedTestSuitePState::RegisteredTest` exists and falls back to `AddTestName()` otherwise. folly is open source and its getdeps build uses googletest v1.17.0, which predates this change. Fixes `folly/coro/test:gtest_helpers_test` in fbcode and xplat.
- `fbcode/paxi/common/AsyncTest.h`: `TYPED_ASYNC_TEST_P` gets the same change as follyfex. Fixes `fbcode//paxi/api/test:dummy_paxi_log_test`.

CI only flagged the follyfex and contacts targets; the crucible, folly and paxi breaks came from local builds. Test behavior is unchanged.

#build-all
allow-large-files
allow-many-files

### Upstream comparison
All 4 changed googletest files are byte-identical to upstream -- no Meta patches. The Manual Fix files are Meta-side changes and need a normal review.

___

landed-with-radar-review

Differential Revision: D125187238

fbshipit-source-id: 9fb208c40ec0be388bb3f5908d3ccf883d52f767

v2026.10.12.00

Toggle v2026.10.12.00's commit message

Unverified

This commit is not signed, but one or more authors requires that any commit attributed to them is signed.
Make cert utils build against OpenSSL 4.0 (#2706)

Summary:
Pull Request resolved: #2706

OpenSSL 4.0 makes ASN1_STRING opaque and const-corrects the X509 accessors (X509_get_subject_name/issuer_name, X509_get_ext, X509_EXTENSION_get_object/get_data).

Switch the affected call sites in OpenSSLUtils and OpenSSLCertUtils to the const-correct API, using ASN1_STRING_get0_data/length instead of direct struct access. The const_casts bridge pre-4.0 getters that still take non-const parameters; they are no-ops there and keep the code compiling against 1.1.1 and 3.x as well.

Note on the const_casts: fbcode builds folly against OpenSSL 1.1.1, where e.g. X509_NAME_get_index_by_NID(X509_NAME *name, ...) is still non-const (openssl/1.1.1 x509.h:816); removing the cast fails the build, so it is required there and a no-op on 3.x/4.0 where the parameter is const. Verified by building with and without the cast.

Upstream issue: #2705

Reviewed By: yfeldblum

Differential Revision: D120544956

fbshipit-source-id: 6f239c7395b6c6f7b9cbaf1750fe7b537c2c2f09

v2026.10.05.00

Toggle v2026.10.05.00's commit message
Add RequestData::clearable() opt-out

Summary:
`clearContextData()` (and overwrite / duplicate `setContextData()`) can remove a `RequestData` from a context while other threads still reference it, thus it is almost impossible to use safely.

Furthermore, some `RequestData` implementations might rely on not being clearable by the business logic, for example to ensure that thread-local state is always cleaned up by pairing each `onSet()` with an `onUnset()` (which is not guaranteed right now if the object is cleared). These objects generally try to achieve this by hiding the request token, but that is just a string that can be hardcoded.

Add a `clearable()` opt-out (default `true`). A `RequestData` that returns `false` cannot be removed from a context that may be shared: `clearContextData()`, a duplicate `setContextData()`, and `overwriteContextData()` with `safe = false` throw `std::invalid_argument`. It can still be replaced or removed with `ShallowCopyRequestContextScopeGuard`, which overwrites it with `safe = true` in a copy of the context that is not shared yet.

The opt-out is meant to be the new default by convention, since it would be a large effort to retroactively fix all the existing implementations.

Reviewed By: uvdn7

Differential Revision: D108613243

fbshipit-source-id: 4ba72901490ad89b402eb5768a7cff0a696b958a

v2026.09.28.00

Toggle v2026.09.28.00's commit message
Define the installed rule root with $HOME

Summary:
Quoted `{FA}` paths do not expand a leading `~`. Define `{FA}` as
`$HOME/folly_agents` in both installer templates, and include that definition in
the loader hash so any harness using the installed loader gets the existing
migration warning.

Drive-by fixes:

- Another failure to expand `~` was in `--preamble-dir=~/...`. The old
  parser rejected space-separated options.  Fix that bug, and instead expressly
  forbid duplicate options (the intent is to prevent overrides).
- Reject relative `TMPDIR` before creating reviewer paths.
- Allow reviewer to retry cheap startup failures.
- `stale_rule_alert.py` now detect divergent definitions of `{FA}`.

___

Differential Revision: D122104077

fbshipit-source-id: 8e0923581babdd99bbcc871e8e7cf34bc730f7af

v2026.09.21.00

Toggle v2026.09.21.00's commit message
Fix Arena::merge returning already-allocated memory from an empty des…

…tination

Summary:
## What goes wrong

```cpp
SysArena src;
auto* live = static_cast<char*>(src.allocate(1));
*live = 'x';

SysArena dst;                // never allocated
dst.merge(std::move(src));   // dst takes ownership of src's blocks

dst.allocate(1);             // returns `live` again: an address already allocated
```

## Why

```cpp
blocks_.splice_after(blocks_.before_begin(), other.blocks_);
```

`blocks_` is a `boost::intrusive::slist` of `Block` headers, each at the head of one malloc'd chunk. `splice_after(prev, x)` moves **every** element of `x` into this list immediately after `prev` and leaves `x` empty, relinking next-pointers only — no allocation, no copy, node addresses unchanged, which is what makes it a legitimate ownership transfer. It is `_after` because a singly-linked list has no back-pointers, so positions are expressed via the predecessor; `before_begin()` is the sentinel root before the first element, so this means **insert at the front**.

`allocate()` only ever advances its cursor to the next block, toward the tail. So the invariant is: *any block behind `currentBlock_` is unreachable and can never be handed out again.* Front-splicing is deliberately putting the merged blocks out of reach.

- **Destination has already allocated** — `currentBlock_` is anchored on one of its own blocks, which sit *after* the spliced-in ones, so it can only ever advance into its own reusable blocks. `last_` is not updated either, correctly: inserting at the front does not change the tail.
- **Destination has never allocated** — a fresh `Arena` sets `currentBlock_ = blocks_.last()`, and for an empty `cache_last` slist `last_` **is** the root node, so `currentBlock_ == before_begin()`. That is the one position in the list *before* the merged blocks. `splice_after` moves `last_` to the merged tail (its `cache_last` bookkeeping fires only when `prev` is the tail, i.e. only when the list was empty) and nothing rewinds `currentBlock_`. `canReuseExistingBlock()` then sees `currentBlock_ != blocks_.last()`, so `allocate()` advances the cursor to the next block — the first merged one — and hands it out.

https://pxl.cl/ddJKx

## Fix

Park `currentBlock_` on the tail when the destination was empty, restoring the invariant that the cursor is at or after every merged block. The next allocation then falls through to `allocateSlow()` and appends a fresh block.

## Impact

Latent today. The only in-tree caller is `ThreadCachedArena::zombify`, whose destination arena only accumulates merged blocks and is never allocated from. It bites as soon as anything merges into an arena it then allocates from.

Reviewed By: vitaut

Differential Revision: D120221054

fbshipit-source-id: 80382d2e1b42daed5cad8afb3509ef72c11ed711

v2026.09.14.00

Toggle v2026.09.14.00's commit message
Remove unnecessary map lookup in RequestContext::State::insertNewData()

Summary: The lookup is always false, as verified by the `DCHECK(inserted);` below.

Reviewed By: IosifSpulber

Differential Revision: D116023688

fbshipit-source-id: cb011414350bc1c74814334fae44da3618e62cdd

v2026.09.07.00

Toggle v2026.09.07.00's commit message
AsyncSocket: clear errMessageCallback_ before delayed zero copy drain

Summary:
Fixes a similar problem to D118665976. The delayed zero copy drain may cause AsyncSocket to outlive errMessageCallback_, so it isn't safe to call from within the the delayed drain callback.

Clear it when scheduling the drain.

Reviewed By: dmm-fb

Differential Revision: D118666023

fbshipit-source-id: 4b47200d03a9e3d1aff0ad2a41e6de202d256fa8

v2026.08.31.00

Toggle v2026.08.31.00's commit message
folly/lang: guard the -Wcast-function-type-mismatch pragma

Summary:
`reinterpret_function_cast` suppresses `-Wcast-function-type-mismatch` with an
unconditional `FOLLY_CLANG_DISABLE_WARNING`. That warning group did not exist
before clang 19, and clang diagnoses an unknown group in a
`#pragma clang diagnostic ignored` as `-Wunknown-warning-option` -- a warning
normally, an error under `-Werror`. Any clang older than 19 that reaches
`Cast.h` in a `-Werror` build fails to compile it.

Nothing hits this today because `Cast.h` has a narrow include closure. It is a
latent trap for whoever widens that closure, so guard the pragma with
`FOLLY_HAS_WARNING`, which is `__has_warning` on clang and 0 everywhere else.

___

Differential Revision: D118025808

fbshipit-source-id: db6ba0d798e3d6431b05a99d951dc1d49990d41f