Summary:
To encourage migration to safer executors, add transitional *Unsafe names for the old, mutable, global executors. This is consistent with the use of *Unsafe elsewhere in the async framework.

(Note: this ignores all push blocking failures!)

Reviewed By: yfeldblum

Differential Revision: D26418111

fbshipit-source-id: 414ee238ed4ae43619ce8d671fa5725f0679f67e

Summary:
It looks like the various CMAKE_CXX_FLAGS_* are simply ineffective, and I'm not
sure why, setting CMAKE_CXX_FLAGS does work though, and CMake appears to add
some release/debug flags to it when generating Ninja files.

Reviewed By: fanzeyi

Differential Revision: D27862117

fbshipit-source-id: a89f6182b5bae9a087f8fcfd4c5f9526f91e2adf

Summary: Avoid calling arbitrary user code from within the scope of framework `catch` blocks. Instead, get the current exception in the catch block, save it, and call the user code after.

Differential Revision: D27842318

fbshipit-source-id: de3ebabc5c08332701a619a93b6e01bfccc0ea71

Summary:
Storing the `__PRETTY_FUNCTION__` for logging has an unreasonable cost on (non-strippable) binary size for debug builds, as different template instantiation will produce long, demangled, and non-mergeable strings.

If necessary, all the information is retrievable from the debug info (even if the function is inlined), where it's stored in much more efficient form. So, switch to `__func__`, which is stripped.

Reviewed By: yfeldblum, luciang

Differential Revision: D27825136

fbshipit-source-id: d499084c7e7b24db1cd46aa2b03f4a590c6eddc3

Summary: Doing so is a bug that can take some time to pin down. Fail at the callsite instead to make this easier.

Reviewed By: yfeldblum

Differential Revision: D27829947

fbshipit-source-id: e3628331ffb687718309e130e8e0f2f78c4eb1aa

Summary:
Add support for AsyncSocket connect bind to device

(Note: this ignores all push blocking failures!)

Reviewed By: danobi

Differential Revision: D27784176

fbshipit-source-id: d7efe3ef59b26aa69552d5f8109f89457ee5fe66

Summary: To avoid the marginal but still unnecessary compile-time cost of constructing a lambda. And enclose the full thing in braces so it can be used in an unbraced `if`.

Reviewed By: ot, luciang

Differential Revision: D27817005

fbshipit-source-id: 9a1a6e0389fea5907a3a20cffbd6a029b0a85eb9

Reviewed By: Mizuchi

Differential Revision: D27669414

fbshipit-source-id: 617e0b5f17fa8ce5505f7f16b00516ba7f21b064

Summary: Why is unsafeGetUnlocked() used here? It seems obviously not safe when we're iterating on the vector after. Am I missing something?

Reviewed By: andriigrynenko

Differential Revision: D27748519

fbshipit-source-id: fe8a78408fdf59e83f5261740e4703f876ad9f9d

Summary: Avoid enlarging data sections. The use of `static` causes data sections to be smaller in certain critical cases.

Reviewed By: cmds, luciang

Differential Revision: D27813138

fbshipit-source-id: 020b14927be5ac035d5f66e7cd652a58f8f820d0

Summary:
Add support for AsyncSocket readv operations

(Note: this ignores all push blocking failures!)

Reviewed By: danobi

Differential Revision: D27585275

fbshipit-source-id: d9a410a4c5d60af06c187e0abb5bd11b7054a6b9

Summary: Should address various build issues.

Reviewed By: Orvid

Differential Revision: D27699043

fbshipit-source-id: 7bcb827c3372e6243554e8bf04426c041b2c62a4

Summary:
It's not safe to expect that the executor is alive after a task is added to the queue (this task could be holding the last KeepAlive and when finished - it may unblock the executor shutdown).
The fix is to hold a KeepAlive if some operation has to be done after adding to the queue. However it's safe to avoid this in a case where executor thread count will never drop below 1 (otherwise not having active threads may result in a deadlock) and we already have the maximum number of threads running. This optimization should help avoid grabbing the KeepAlive on the fast path.

The difference from the previous implementation is that we still start with 0 threads. But keep minThreads at at least 1.

Differential Revision: D27751098

fbshipit-source-id: 7b992a73b40984376893c1fe2a4a072002f0a12a

Summary: To differentiate serialization and deserialization errors.

Reviewed By: yfeldblum

Differential Revision: D27743086

fbshipit-source-id: 8eb4fe1a1733e3effd0aa0478f59160381738db2

Summary:
The build was failing at link time with the claim that `__PRETTY_FUNCTION__` was in a discarded section. This occurs only under some versions of gcc with asan enabled.

This approach works by making a copy of `__PRETTY_FUNCTION__` without ever taking a reference to it.

This may not be the ideal fix, but it is the first one that worked upon experimentation.

Differential Revision: D27786615

fbshipit-source-id: 2845c85bdc8158d4b2da7cd200ce55f982285a21

Summary: Since `[[gnu::weakref]]` can be used effectively in cases where `[[gnu::weak]]` cannot.

Reviewed By: Orvid

Differential Revision: D27672340

fbshipit-source-id: 4882f0176726aece411359c25411f0f4be7fc59f

Summary: Add some functions for looking at the code generated by `AccessSpreader<>::current()` and `AccessSpreader<>::cachedCurrent()`.

Differential Revision: D27747508

fbshipit-source-id: 0bc44938e810ec76ead50b6d4c1ade4aec9f8282

Summary:
Add support for AsyncServerSocket bind to device

(Note: this ignores all push blocking failures!)

Reviewed By: danobi

Differential Revision: D27778929

fbshipit-source-id: fa051be2aa5c1b3df5e0c0f30adc3a58de7b7704

Summary: Since HeaderClientChannel now accepts a transport unique_ptr there's no need to have this deleter exposed outside of HeaderClientChannel.

Reviewed By: iahs

Differential Revision: D27729209

fbshipit-source-id: 064b03afdfe567b6df6437348596f0f6f97f6aaf

Summary:
Pull Request resolved: https://github.com/facebook/folly/pull/1561

As title.

Reviewed By: jdq

Differential Revision: D27736665

fbshipit-source-id: 2f09ea4dad25389a46faccca4a5803ea9e077e0d

Summary:
Create a new class, ConstructorCallback, that when included as a member in
another class (e.g., Foo) allows other functions to register and receive
callbacks every time any of Foo's constructors are called.

Main implementation concern is to be extremely lightweight (e.g., avoid locks)
assuming that most code calling this will be performance optimized (both
for CPU and mem).

Implement as a static array of functions:

1) Static array instead of dynamic (e.g., vector):
  vector occasionally resizes so avoid locking during resize
2) Don't allow removing callback functions:
  simplify the calling of funcitons, again to avoid locks and races
3) Use function callbacks instead of objects with an observer pattern
  (e.g., like AsyncSocket::LifecycleObserver) to ensure that there
  won't be races on shutdown.  Functions are static where classes can
  de-allocate at shutdown and since we don't allow un-registering the
  call back, it's not easy to remove the race for a class.

Reviewed By: simpkins

Differential Revision: D27056139

fbshipit-source-id: 0846e0d55124cfde2e15acbe18e01ca5e327e7df

Summary:
`DigestBuilder` caches the cpu id and invalidates it when it detects contention, but the critical section is small enough that it is possible for two thread to end up on a conflicting slot but not detect it for some time, causing cache-line pingponging in the meantime.

Switch to `cachedCurrent()` instead, which uses the number of accesses to invalidate the cache, and is successfully adopted by other high-concurrency primitives.

Reviewed By: yfeldblum

Differential Revision: D27726518

fbshipit-source-id: 05d70ad9f101df4a8797e44cc3cbe94cbd1cb7b8

Summary:
GmockHelpersTest.cpp is the only one that fails to build with GCC after this change.
Also only 5 tests are failing at this point.

Reviewed By: pixelb

Differential Revision: D27728248

fbshipit-source-id: ded994f466347a8826ba90429a13237c21d3af5e

Summary:
F14 map `iterator` doesn't define `operator==` or `operator!=` to itself.
Those member functions take `const_iterator` as the right-hand side,
relying on the ability to convert implicitly from `iterator` to
`const_iterator`. This interacts poorly with c++20's default equality
operators, triggering clang's `-Wambiguous-reversed-operator`.

This diff replaces the equality member functions with friends that use
the same left-hand and right-hand types.  iterator != const_iterator
will now compile by implicitly converting the left-hand-side and then
calling `operator!=(const_iterator const&, const_iterator const&)`,
rather than calling `iterator::operator!=(const_iterator const&)`.
This resolves the ambiguity in c++20, and should work fine on earlier
versions, although I have not tested it on anything except c++17.

Pull Request resolved: https://github.com/facebook/folly/pull/1544

Test Plan: does it build?

Reviewed By: yfeldblum

Differential Revision: D27479086

Pulled By: shixiao

fbshipit-source-id: 0d42dbdc056cec58a36de3a816883736a761d0f6

Summary: Consistently use `Observer` and `Snapshot` like everywhere else in these files instead of `folly::Observer::{Observer,Snapshot}`.

Reviewed By: yfeldblum

Differential Revision: D27685699

fbshipit-source-id: 7eca35b9d113d8d9cd8ca631be2ad54c827431ea

Summary:
Nothing uses the skipFraction in LockFreeRingBuffer::currentTail, and
floating point math for discrete fractions can be hard to reason
about, so remove this dead code.

Differential Revision: D27717906

fbshipit-source-id: b25e2bea25ab66c59e8c76a80d75140db474c46d

Summary: To facilitate debugging of T88337980

Reviewed By: rdavies

Differential Revision: D27610329

fbshipit-source-id: 6b01af24e2cfdda9afcbef10d2951ac071d2fd10

Summary:
- Observer implementation backed by hazptrs
- We were using TLObserver in a library our team maintains but were seeing global lock contention on process startup due to the TL https://fburl.com/scuba/offcpu/aud1qe8p (more details here https://fb.workplace.com/groups/wfi/permalink/5462384263810144/)
- Also might be a perf win when compared with TLObserver when using getLocalSnapshot (per micro benchmarks / strobelights in D27535141)

Reviewed By: yfeldblum

Differential Revision: D27565756

fbshipit-source-id: f345a34c5fd2c69fde6b98f698f79640016f6eee

Summary: Pull Request resolved: https://github.com/facebook/folly/pull/1556

Reviewed By: Orvid

Differential Revision: D27672765

Pulled By: yfeldblum

fbshipit-source-id: edcf9b300c737799399096a8dbc401b3ab49b88c

Summary: Pull Request resolved: https://github.com/facebook/folly/pull/1557

Differential Revision: D27701865

Pulled By: yfeldblum

fbshipit-source-id: f9ed96a8e9283a30f1f96d15b4b8166ade5d3321

Summary: To get this to link across platforms, requires moving the thread-locals into functions. To avoid curious performance hits, requires removing the field initializers.

Differential Revision: D27699178

fbshipit-source-id: 7617b8e4aedeccae4cd95ffe606d1414dc352705

Reviewed By: yfeldblum

Differential Revision: D27695785

fbshipit-source-id: da8666781e03667a0060e8a7daab4206f60fc71d

Summary: There is no sense to return the iterator that points to the outdated node that just was overwritten by `assign_if_equal` call.

Reviewed By: yfeldblum

Differential Revision: D27678357

fbshipit-source-id: f07ce21df6556819f7eb103a623d3f7a46e72072

Summary: DRY the `ShallowCopyRequestContextScopeGuard` ctor taking multiple items.

Differential Revision: D27698685

fbshipit-source-id: 7f23305855918c4248da11c15e5fab839592d372

Summary:
LockFreeRingBuffer is racy for nontrivial types. I could reproduce
ASAN failures with nontrivial types like std::string, and TSAN
failures with trivial types.

Require triviality by default, and sequence the turns, loads, and
stores with atomic_thread_fence.

Reviewed By: yfeldblum

Differential Revision: D27684743

fbshipit-source-id: e0b32855d3ed5c2edbd7ab72092f2c533071a70f

Summary:
In preparation for making LockFreeRingBuffer trivial-only, remove the
ability to convert types during reads and writes.

Reviewed By: yfeldblum

Differential Revision: D27684584

fbshipit-source-id: 092009b2852d000fa3821756d1ec4bfd2bf002c8

Summary:
`SequentialThreadId` is only ever instantiated with `std::atomic` so just make it no longer a template.

This permits moving the implementation into the source file.

Reviewed By: Mizuchi

Differential Revision: D27669406

fbshipit-source-id: ec2cddbff433df31434548b61111a34ca2da5cef

Summary:
Fix a bug in small_vector reported as a warning via gcc.

```
folly/small_vector.h: In instantiation of 'constexpr const size_t folly::small_vector<short int, 10>::MaxInline':
folly/small_vector.h:1183:56:   required from 'class folly::small_vector<short int, 10>'
bits/stl_stack.h:134:47:   required from 'class std::stack<short int, folly::small_vector<short int, 10> >'
folly/small_vector.h:462:36: error: division 'sizeof (short int*) / sizeof (short int)' does not compute the number of array elements [-Werror=sizeof-pointer-div]
  462 |       constexpr_max(sizeof(Value*) / sizeof(Value), RequestedMaxInline)};
      |                     ~~~~~~~~~~~~~~~^~~~~~~~~~~~~~~
```

Reviewed By: yfeldblum

Differential Revision: D27688663

fbshipit-source-id: 8da5b017c597c542826e26f2c36b0e3b46e87ed9

Summary: Migrate the codebase away from the ad-hoc `folly::uint64ToBufferUnsafe` and to `folly::to_ascii_decimal` which is intended for these cases.

Reviewed By: WillerZ

Differential Revision: D27281875

fbshipit-source-id: 0c98749e4aed9c873853eed2221cf54a89279ff4

Summary:
It's not safe to expect that the executor is alive after a task is added to the queue (this task could be holding the last KeepAlive and when finished - it may unblock the executor shutdown).
The fix is to hold a KeepAlive if some operation has to be done after adding to the queue. However it's safe to avoid this in a case where executor thread count will never drop below 1 (otherwise not having active threads may result in a deadlock) and we already have the maximum number of threads running. This optimization should help avoid grabbing the KeepAlive on the fast path.

Differential Revision: D27584518

fbshipit-source-id: e1242e3f4c40cee4f7e0c6dfca39abe6d17415f1