Reviewed By: zertosh

Differential Revision: D9993527

fbshipit-source-id: 52c4c0bfe6b5c73c652dd350f599fb846a003a4f

Summary: (Note: this ignores all push blocking failures!)

Reviewed By: spalamarchuk

Differential Revision: D9994424

fbshipit-source-id: 2055d3b09ce2e1d783624a9902463efade4714d4

Summary:
This is known at compile time, but we still pay the cost since
SettingCore has to look it up through its own reference.
Hint it directly at the callsite instead.

Differential Revision: D9983460

fbshipit-source-id: 1c6ca5e3cbd9e859ba88d641cdca59f0d4d65dfb

Summary: Requires a new header to break the circular dependency

Differential Revision: D9983196

fbshipit-source-id: f15ab3c6128d30906ac25384df92813aa9d76c9a

Summary:
- Change the get() API to return T by value if T is small
- Internally store small T in a global atomic to avoid thread local lookups

Differential Revision: D9982433

fbshipit-source-id: c42d06a75b9ff48307a90f77d0bb72bfa6375d6c

Summary: Makes further changes cleaner

Differential Revision: D9982335

fbshipit-source-id: 19268c8e3fc1b6aee589a8d3d38b0348db8c5a43

Summary: inline the small, fast path only

Differential Revision: D9982097

fbshipit-source-id: f1bcbab2b444d372e178aa548175a5c27be972b6

Summary:
With this change the setting lookup is now inlined at each callsite as:
```
if (UNLIKELY(global_setting_ptr == 0)) {
  global_setting_ptr = &initialize_setting(); // OK to race
}
// use *global_setting_ptr
```

Differential Revision: D9981958

fbshipit-source-id: a04b296d8345d1bb37dc06ce72277e00e745a94a

Summary:
The folly::coro::Task<T> and folly::coro::Future<T> types would previously return an lvalue-reference to the result when you co_await them.

This means that for move-only types that code like `auto x = co_await someTask;` would fail to compile. Instead you would need to write `auto x = std::move(co_await someTask);`.

Awaiting a Task<T> or Future<T> now returns type T instead of T&.

As part of this change we now only allow co_awaiting an rvalue folly::coro::Task<T>, folly::coro::Future<T> as well as folly::coro::Future<T> to indicate that the operation is destructive and a one-time operation.

Reviewed By: andriigrynenko

Differential Revision: D9731525

fbshipit-source-id: bee9e633b57b203a0d048cf3eb0e2fc48b899481

Summary: It was possible for interruptHandler() to be called concurrently with the ThreadWheelTimekeeper shutdown resulting in use-after-free when trying to schedule work on the EventBase.

Reviewed By: yfeldblum

Differential Revision: D9948555

fbshipit-source-id: 4eff3cce0488778559dffc47bfc6eb59a65bb203

Summary: The && overload here caused me to call the wrong function and delay an error as it changed the type of the Future completely. There seems to be little benefit to making this r-value qualified, and is likely a side effect of consistency with the forwarding reference in the non-Try overload.

Reviewed By: yfeldblum

Differential Revision: D9979805

fbshipit-source-id: e8121d5bc2e5fe02d38d6ea9cb83adbe98010df2

Summary: The tests that would have caught this aren't currently in the CMake file, but this is needed before trying to use the intrinsics here.

Reviewed By: yfeldblum

Differential Revision: D9952025

fbshipit-source-id: 40c7034ef7e0877ccb95525fa09ab46a3526af0b

Summary: FOLLY_TLS is not guaranteed to be defined, so add a check

Differential Revision: D9952230

fbshipit-source-id: 1f85d1ec39514b5b3c355339a74d067ecd85b267

Summary: [Folly] Lift some crc magic constants in `folly/hash/detail/Crc32CombineDetail.cpp`.

Differential Revision: D9934404

fbshipit-source-id: e2508a22e300e22f2265383a0a111df53d2738c5

Summary: Fix opt mode error: unused parameter 'data'

Reviewed By: aary

Differential Revision: D9970478

fbshipit-source-id: 8d78e5f12613776486e33e8c685e400cd88f9d03

Summary: This change will allow simplifying DeferredExecutor and implementing both wait() and timed wait() using custom executors.

Reviewed By: LeeHowes

Differential Revision: D9890644

fbshipit-source-id: ad8472213e7022ccc23b4a9c48f8b9c4b9c70dfa

Summary: This will make it easy to use coro primitives from fibers code.

Reviewed By: lewissbaker

Differential Revision: D9932012

fbshipit-source-id: cd0524ccfee7bee9405f29ddaa3533084b18b38b

Summary: [Folly] Fix move-of-const in `ThreadWheelTimekeeper`.

Reviewed By: capickett

Differential Revision: D9950653

fbshipit-source-id: 76d45cdac1026c41de73489936c4afdac3f57fd7

Summary: - Adding hasValue() function

Reviewed By: yfeldblum

Differential Revision: D9926823

fbshipit-source-id: 2316eb67cd6a906850ad0dec913988f116f840d6

Summary: [Folly] Just use `constexpr` keyword in `folly/lang/Bits.h` - since they are now template functions, `constexpr` is okay even when all instantiations would be non-`constexpr`.

Reviewed By: Orvid

Differential Revision: D9934124

fbshipit-source-id: cdf4fc87262a058af6f63e28bb2c390e44f2cad0

Summary:
`hash_range(c.begin(), c.end()` combines hashes of individual elements of a
container in an ordered manner. This diff provides the equivalent for
unordered containers.

Unlike `hash_range`, `commutative_hash_combine_range` defaults to `folly::Hash`
as its hasher; it mixes the individual hash if the `hasher` is not deemed
avalanching.

It uses a commutative accumulator described in this paper:
https://www.preprints.org/manuscript/201710.0192/v1/download
In the experiments in the paper, the symmetric polynomial yielded a better
spread of hash values and lower collision rates than `+` or `xor`.

Reviewed By: yfeldblum

Differential Revision: D9688687

fbshipit-source-id: c812b25975a53a868d98f78645146cb8bdbb5c32

Summary:
`IOThreadPoolExecutor.pickThread()` has some sketchy behavior where, if the
threadpool has no threads, it returns `*thisThread_` unconditionally. My read
is that this can happen only in two cases.

First, `getEventBase` could be called from outside the threadpool while the
threadpool has no threads (either because it was stopped or because somebody
set it to 0 threads.) In this case, `*thisThread_` will be nullptr and
`getEventBase` will try to immediately dereference it, (hopefully) crashing.

Second, `getEventBase` could be called from inside the threadpool, but from a
thread which is shutting down. This is also somewhat dodgy, because by the time
we get to this case (`me && !contains(threadList_.get(), me)`), stopThreads
will have already called terminateLoopSoon on the evb. While it would still be
possible to use the evb safely (in the sense that one could write a program
which does not hit undefined behavior), I think it'd be strongly dependent on
the way IOTPE and EventBase are currently implemented and therefore probably a
bad idea.

I don't have the bandwidth to really dig into this now and figure out what the
correct behavior is, or how to fix people who might be relying on the broken
one. But I figure I can at least leave a comment behind that this is something
maybe worth looking into.

Reviewed By: meyering

Differential Revision: D9754314

fbshipit-source-id: 6851d49e634707cacc803a95763eda023967b932

Summary: Because reasons :(

Reviewed By: yfeldblum

Differential Revision: D9923447

fbshipit-source-id: 7a3a9ab86764dc5c7f5b70f2f728aec8b67bdf1f

Summary:
This diff adds spurious rehash at places where the caller should
not be assuming reference (unless guaranteed) or iterator stability,
allowing ASAN to detect these problems.  Under ASAN it also junk-fills
deleted value_type instances that can't be immediately deallocated.

Reviewed By: yfeldblum

Differential Revision: D9826273

fbshipit-source-id: 2d543473b17ffd7debb52082a3e72842121ece1a

Summary:
An allocator that uses a jemalloc arena backed by 2MB huge pages.

This uses the jemalloc extent hooks to let jemalloc hand out memory backed by huge pages.
The huge pages are allocated when init(int nr_pages) is called, and does not grow (although extending it to grow in the future should be fairly straightforward).

madvise is used to mark the pages as huge. It does not guarantee success, and the only way to tell is by looking at /proc/<pid>/smaps and the total huge pages in /proc/<pid>/status.

The HugePageAllocator can be used as template parameter for stl and folly collections.

Reviewed By: yfeldblum

Differential Revision: D9297627

fbshipit-source-id: 734ca1b19972a8b5d7a12ee60cba1ffa3da50dc2

Reviewed By: ngoyal, mingtaoy

Differential Revision: D9862034

fbshipit-source-id: ce1f00fe12c52e755ff67775e980ff3ff4227bbf

Summary: No semantic changes. Reordering, and removal of an unnecessary fwd decl

Reviewed By: nbronson

Differential Revision: D9851048

fbshipit-source-id: dd306435ad2f77e42f74dc29b19ec5d6fe09ae2a

Summary: It's no longer needed, and was a pain to support properly.

Reviewed By: yfeldblum

Differential Revision: D9846778

fbshipit-source-id: b727a3459efed4e6df3db50f8d7fe0963c3608f8

Summary: Handle the dlopen static initialization

Reviewed By: yfeldblum

Differential Revision: D9847032

fbshipit-source-id: 089628ef88cf72d189a8904a3f2231f1affc764c

Summary: It's no longer needed.

Reviewed By: yfeldblum

Differential Revision: D9847073

fbshipit-source-id: 072f4e5fc207ab199689e42c4f492d4fcb67aa8e

Summary:
The last bits of pthread needed for thrift.
Also define PTHREAD_MUTEX_NORMAL properly.

Reviewed By: yfeldblum

Differential Revision: D9846774

fbshipit-source-id: 124013c6df07f9faa214b77fa8ebe8b37d8f460e

Summary: [Folly] Fix `rcu_domain` move-assignment pointer copy which seemingly should be a pointer exchange.

Reviewed By: magedm

Differential Revision: D9846892

fbshipit-source-id: ef3a113780af7456d66dacd376b3bf43670fcb9e

Summary: Documentation has drifted out of date with updated plan. This brings the core parts into line.

Reviewed By: yfeldblum

Differential Revision: D9833211

fbshipit-source-id: b372f9cfc0c787feab86490beeb1dcaa290e8566

Summary:
[Folly] Tweak `ThreadLocal` inlining.

Shrinks overly-large generated code for `ThreadLocal::get()`.

Reviewed By: luciang

Differential Revision: D9732837

fbshipit-source-id: 71eb10d9759e1fa17248e0feed75f610948afe21

Summary: [Folly] Tweak `ThreadLocal` fast-path functions. Mostly code golf.

Differential Revision: D9732836

fbshipit-source-id: fc48a7c067c6961413e6aa94106ae26f87d3d603

Summary: Let's see what breaks.

Reviewed By: yfeldblum

Differential Revision: D9569029

fbshipit-source-id: 2edbafc018ec70d8640f2e28d35735089c9eb69d

Summary: This only exists for perf comparisons, and has been unused for years.

Reviewed By: yfeldblum

Differential Revision: D9846361

fbshipit-source-id: ad28ff12fe16deb0ceee50ce6c098f234e8f9692

Summary: Make comments consistent with the updated recommendations where it makes sense.

Reviewed By: yfeldblum

Differential Revision: D9833280

fbshipit-source-id: 8de3372b013a322e57cd2d54a512d75bf963d1a5

Summary:
lock bts and lock btr with the varying lengths turn out to be faster than
a single CAS instruction

fetch_set uses lock bts and fetch_reset uses lock btr when possible.
std::atomic::fetch_or and std::atomic::fetch_and with a mask otherwise

Reviewed By: yfeldblum

Differential Revision: D9652759

fbshipit-source-id: e6d10fdb119a10e15e32131c7389f1d4203a0d91

Make rcu_default_domain_ folly::Indestructible<rcu_domain<RcuTag>> to avoid asserts on _dl_open/_dl_init __static_initialization_and_destruction_0

Summary: Make rcu_default_domain_ folly::Indestructible<rcu_domain<RcuTag>> to avoid asserts on _dl_open/_dl_init __static_initialization_and_destruction_0

Reviewed By: yfeldblum

Differential Revision: D9845303

fbshipit-source-id: e858e86a6f7885bf6a3ad8d5e957ee947a5ce2b3