Summary:
The previous approach (D16760775) reserved space for inlined function info in each `SymbolizedFrame`.

```
struct SymbolizedFrame {
  Dwarf::LocationInfo location;
  Dwarf::LocationInfo inlineLocations[Dwarf::kMaxLocationInfoPerFrame];
  ...
```

That increased the size of `SymbolizedFrame` and `FrameArray` and lead to stack overflow in some already deep stacks.
```
template <size_t N>
struct FrameArray {
  FrameArray() {}

  size_t frameCount = 0;
  uintptr_t addresses[N];
  SymbolizedFrame frames[N];
};
```

To avoid allocate more space on stack, changed to use extra frames to store inline calls:
- Usually the callers allocate `FrameArray<100>` frames, but the stack trace is usually smaller than 100
- Use the unused slots to fill in inlined function info:
 -- each function gets at most `kMaxLocationInfoPerFrame` (currently 3) inlined entries
 -- when the available buffer fills up no more inlined functions are filled in.

To find the inline calling stack, we need first need to find the subprogram Debug Info Entry (with tag DW_TAG_subprogram) with the given address, then recursively find all the inline subroutines (with tag DW_TAG_inlined_subroutine) in the call stack. Sadly debug info has no index we can use for jump, and a linear search over debug info entries (and their attributes) is needed during the process, which would cause performance regression.

```
buck run mode/opt folly/experimental/symbolizer/test:dwarf_benchmark -- --benchmark

============================================================================
folly/experimental/symbolizer/test/DwarfBenchmark.cpprelative  time/iter  iters/s
============================================================================
DwarfFindAddressFast                                         4.03us  248.36K
DwarfFindAddressFull                                         4.03us  248.18K
DwarfFindAddressFullWithInline                             293.23us    3.41K
============================================================================
```

Reviewed By: luciang

Differential Revision: D17586385

fbshipit-source-id: 1b84b3f3a576573ce24092b433a501a3bdf76be0

Summary: users may need AsyncUDPSocket to send back datagram to clients.

Reviewed By: yfeldblum

Differential Revision: D18764100

fbshipit-source-id: 7ab7cce5711f47db92f43a68b94568729d5746d0

Summary: Workaround for something that looks like a compiler bug

Reviewed By: kevin-vigor

Differential Revision: D18700404

fbshipit-source-id: 6d9c9a40e9e468ee910c5456bcbfdf02c1b1c59a

Summary:
This is the header which provides `_mm_set1_epi8`
(https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_set1_epi8).

Although this header is transitively found via other headers (which is why this
isn't a build failure), without this explicit include IWYU's heursitics around
template functions wrongly assume that this header should be provided by
downstream includers of this header.

Reviewed By: yfeldblum

Differential Revision: D18753275

fbshipit-source-id: 7d6ee934542867f2d1e84d40397f48ea8146facc

Summary:
localPush invoked in recycleIndex can end up looping more than once if it has to retry the update of head. When that happens, the onRecycle call will be made again which will call destructor on the object again if eagerRecycle is true.

Fixed it by making sure only the 1st pass of the loop will call onRecycle.

(Note: this ignores all push blocking failures!)

Reviewed By: nbronson

Differential Revision: D18752349

fbshipit-source-id: 27dcb5c7840f724bbd39f4dc0176e21095a22284

Summary: For this iterator to satisfy the `input_or_output_iterator` concept in range-v3, it must also satisfy the `semiregular` concept, which requires a default constructor.

Reviewed By: yfeldblum

Differential Revision: D18689298

fbshipit-source-id: a38cde0041ceba1aac79128d428df05ee0f00ab4

Summary:
This is an update to the fbcode_builder codebase to allow setting up the python virtualenv with python dependencies installed. I've included wheel and cython (with a pinned version to 0.28.6 which is the only version that works with thriftpy3 at the moment, due to https://github.com/cython/cython/issues/2985) as standard packages since these are required by some of our top-level dependencies (folly and thrift)

As far as I know, there are no other projects that use PYTHON_VENV at the moment except LogDevice so the impact should be minimal.

Reviewed By: lucaspmelo

Differential Revision: D18758383

fbshipit-source-id: 264941311c5e3a19dc4ef2bb78c9a1baa34dfd8c

Summary: As titled

Differential Revision: D18726457

fbshipit-source-id: 717a6cda4a78e1b246f20328df1f01b78560cac5

Summary: [Folly] More direct overload resolution control in `folly::coro::concat`. Just name the type directly rather than using SFINAE, when the SFINAE would just name the type directly in a more roundabout way.

Differential Revision: D18746633

fbshipit-source-id: 6da909834b511426b1cf7bbda0fd7643b5761f14

Summary:
This diff addresses a collection of issues with our thrift-py3 build story in open source.
- Cython 0.29 bug https://github.com/cython/cython/issues/2985 prevents us from loading the generated binary python extensions. I've tested 0.28.6 and it works perfectly.
- Adds ssl to the extensions (since we need SSLContext in get_client, we must have access to the module)
- No need to special case folly_pic. There is really no easy way to get the python bindings to work without using shared library libfolly.so. So, In our build (logdevice) we will build folly as a shared library, this also reduces the complexity everywhere.
- We also need iobuf as extension for thrift to work, added that to setup.py
- Refactored some of the CMake code that had hardcoded file names to use globbing to future-proof this as much as possible.
- Moved `python/lang/cast.pxd` to match the python module name `folly.cast`, so now it lives in `python/cast.pxd`. This simplifies the globbing as well.
- Moved `setup.py` to live in the `/python` directory.
- Added `*.py` in setup.py to pickup any raw python code (including `__init__.py`)

Reviewed By: yfeldblum

Differential Revision: D18685009

fbshipit-source-id: 749b30942a3e5e9e314b5248cc0aa90c6ac1581f

Summary: This is another step of moving the python extensions for folly that are currently defined in thrift's codebase. Another step for getting us closer on unblocking LogDevice to use the python3 thrift client in open-source.

Reviewed By: yfeldblum, vitaut

Differential Revision: D16221352

fbshipit-source-id: bc24fce073d67397f20f28f79c68512444704402

Reviewed By: farnz

Differential Revision: D18726052

fbshipit-source-id: 4f18430342ab6fd4fc82cc7d03e21f3e50e0ce25

Summary: [Folly] Avoid linear recursion in `folly::coro::concat`.

Reviewed By: kirkshoop

Differential Revision: D18746601

fbshipit-source-id: a383715bfc47ae8c0d4baad8799e98699c7cc14f

Summary: [Folly] Remove remnants of replacing `std::string` with the implementation of `folly::fbstring`.

Reviewed By: luciang

Differential Revision: D18738367

fbshipit-source-id: b57d38100c9da38c6351145979ac6d2f677ba817

Summary: [Folly] No unnecessary SFINAE in `ElfFile::at` - just use an explicit `static_assert` directly, rather than relying on overload resolution failure for the same goal.

Reviewed By: luciang

Differential Revision: D18746254

fbshipit-source-id: 4bb209bf3f563f2425e4532f6553e66856cb9bbb

Summary: [Folly] `try_call_once`, a nothrow variant of `call_once`. Takes a `noexcept` function which returns `bool`; the function returning `false` indicates failure and that the `once_flag` is not to be marked as as set, while the function returning `true` indicates success and that the `once_flag` is to be marked as set.

Reviewed By: andriigrynenko, luciang

Differential Revision: D18742590

fbshipit-source-id: 3b3a6fab9a328f7540cc6fee1452a0f1367b6e0f

Summary: [Folly] Let `test_once` be `noexcept` since it is required not to fail.

Reviewed By: andriigrynenko, luciang

Differential Revision: D18742588

fbshipit-source-id: fda78057d10fca1521ee6335f7c7758d05c43b88

Summary: [Folly] Hide some test helper functions to permit some tests to be linked together into the same binary, if desired.

Differential Revision: D18746529

fbshipit-source-id: 9829eb1227cb348f65734ceb670c85508577763d

Summary:
[Folly] Mark most `ElfFile` members as `noexcept`.

Since they may be used in signal handlers, they are required not to throw exceptions. Mark it so.

Reviewed By: luciang

Differential Revision: D18743292

fbshipit-source-id: 95fb2a20511ca4cc7c0e832c1b20b8324ca7d0af

Summary: The cargo builder will be used to verify if an opensource Rust project passes Cargo build, test and (optionally) documentation build.

Reviewed By: markbt

Differential Revision: D18636934

fbshipit-source-id: e982e6a017eb32913e2994e7457c8add2e9d6b95

Summary: The `accumulate` algorithm will behave similarly to `std::accumulate`, but instead returns a `Task` of the accumulated result from the values in the `AsyncGenerator`

Reviewed By: kirkshoop

Differential Revision: D18714284

fbshipit-source-id: 66625cd1347d767467d98a421b78b6f45d70260c

Summary: [Folly] Check prev thread-id in `EventBase::loopBody` to enforce that it is not called concurrently.

Reviewed By: andriigrynenko

Differential Revision: D18722163

fbshipit-source-id: 01e18758e4d517011697e025519c84e29ca96636

Summary: The `concat` algorithm will concatenate a series of streams known at compile time into a single stream of the same type.

Reviewed By: kirkshoop

Differential Revision: D18695001

fbshipit-source-id: 2da7f53cbc835333ffd9551ceae1aa5db085936d

Summary: This hadn't been updated since we moved to openssl 1.1

Reviewed By: knekritz

Differential Revision: D18668447

fbshipit-source-id: 26d58ef759fccf7e97c4d127b747b093077f6e59

Summary:
[Folly] Use `relaxed` order for `EventBase::loopThread_` consistently. There is only one location which uses a different memory order.

Atomic operations on this location do not guard any other locations, so non-`relaxed` memory orders are not required.

Reviewed By: andriigrynenko

Differential Revision: D18722614

fbshipit-source-id: b73b61e4845c34c58e9f4adbeb368c815b375d38

Summary: `folly::Function` has this explicit bool conversion operator and it makes sense to also have it for `SharedProxy`.

Reviewed By: yfeldblum, ericniebler

Differential Revision: D18504667

fbshipit-source-id: 32fede326c7a3eebd340dea9d118714a0cdfa5c0

Summary: collectAny had this overload, it was missed when collectAnySemiFuture was added.

Reviewed By: mpark

Differential Revision: D18689251

fbshipit-source-id: 1fccd6eca9d4afde46a173842641f14aff78aada

Summary: Make sure that fiber stack overflow is immediatelly discoverable from the crash stack trace, not just from the error log.

Reviewed By: spalamarchuk

Differential Revision: D18702739

fbshipit-source-id: c0a05081e67b12c94b20a929b2d10396bc1b6949

Summary: [Folly] Make `Function` shared proxy empty when constructed with an empty `Function`, just like it would be when constructed with `nullptr`.

Reviewed By: ericniebler, vitaut

Differential Revision: D18518858

fbshipit-source-id: fdad5b4cd30a1fda5465dae80054c5c74837faa9

Summary: [Folly] Let ElfFile open functions return their error messages v.s. taking pointers to where to place them. Overall, making these functions a bit more C++.

Reviewed By: luciang

Differential Revision: D18668222

fbshipit-source-id: 395c20c5d0862c2b1dcc7f6f9db56b7958c5fc90

Summary:
This diff fixes `CMakeLists.txt` to enable building `openr` tests using CMake:
1. It adds `add_openr_test` CMake function that adds executable target, registers it as test, links it with bunch of libraries like GTest and GMock, etc...
2. There is no `openr/tests/OpenrModuleTestBase.cpp` anywhere in the source tree, so this commit replaces it with `openr/common/Flags.cpp`.

Reviewed By: jstrizich

Differential Revision: D18584028

fbshipit-source-id: 07d854ef98d0d2509889a08ad042a371101a2825

Summary: Moving some `iobuf` helpers from the `thrift::py3` namespace into `folly` since this now lives in folly. Also fixing CMake to build iobuf in folly not in thrift.

Reviewed By: vitaut

Differential Revision: D16221272

fbshipit-source-id: a7bdde1b60cd6d7de06bf2040506a87c3153c37f

Summary:
Remove the Facebook-specific dependencies from the Eden manifest: these
dependencies are now available directly in the Eden repository.

Reviewed By: chadaustin

Differential Revision: D18588009

fbshipit-source-id: 590c74e38e4e05939e6955839af3ebb959f9251b

Summary:
Remove the standalone fb-mercurial-rust target that was an internal-only
dependency for the Eden build.  This build step is now done entirely in the
Eden build.

Reviewed By: fanzeyi

Differential Revision: D18623943

fbshipit-source-id: c62a1155ddd1c0a6b2270c472176ba25194c6145

Summary:
Re-implementation of fiber gdb integration.  All previous functionality still works as well.

Commands:
* `fiber`: switch to fiber by address, alias, or assigned short id
* `fiber apply`: run arbitrary gdb commands for each specified fiber (or "all")
* `fiber deactivate`: switch back to normal thread behavior
* `fiber name`: assign or clear an alias to a particular fiber
* `info fibers`: print out each fiber manager and fiber, with ids and current frame info

Parameters:
* `fiber info-frame-skip-words`: the "uninteresting" function matches to skip frames of when showing fiber info
* `fiber manager-print-limit`: limit of the number of fibers displayed when printing a fiber manager

Fibers are assigned ids of the form "<MANAGER_ID>.<FIBER_ID>".  When specifying fibers for `info fibers` or `fiber apply`, you can specify any number of space-separated fiber ids, and additionally you can specify a manager id by itself to refer to all fibers in that manager.

When printing info with `info fibers`, every stopped fiber will be in `folly::fibers::FiberImpl::deactivate()`, which isn't particularly useful.  To give more-useful information, it will skip past all frames whose function names include the strings "folly::fibers" or "wait" and print the first non-matching frame's information instead.  The `fiber info-frame-skip-words` parameter defines the strings to match on, specified as a space-separated list of words.

Due to gdb limitations, the selecting a fiber "masks" the currently selected thread, and to see the current thread again you will need to run `fiber deactivate`, same as before.

Reviewed By: andriigrynenko

Differential Revision: D18342802

fbshipit-source-id: 5b2204bb09eca4c8f977c6c17b0348e39465dff6

Summary:
- Added sai to getdeps
- added SAI to FBOSS as a dependency
- Included path of SAI headers to FBOSS's cmake file.

Reviewed By: shri-khare

Differential Revision: D18629127

fbshipit-source-id: 119a6b7f2b64bd84414d9e16a903bc8df48a35e7

Summary:
The mutable global executor is problematic: it defaults to Inline, and it is constructed off of an error-prone weak_ptr and it returns a shared_ptr. A KeepAlive-based global immutable executor is a cleaner default.

This change adds immutable global executors, and moves the default executors into separate singletons for cleaner interaction between the mutable executor and the underlying default immutable ones, including the default inline global cpu executor.

Reviewed By: andriigrynenko

Differential Revision: D18513433

fbshipit-source-id: 0ad825c34cc7ba935f57ff81adb8cff3bf001a45

Summary:
With THP set to madvise, page faults on these pages will block until a
huge page is found to service it. However, if memory becomes fragmented
before these pages are touched, then we end up blocking for kcompactd to
make a page available. As this increase pressure, oomd comes in and kills
us :(. So, preemptively touch these pages to get them backed as early as
possible to prevent stalling due to no available huge pages.

Reviewed By: hnaz, interwq

Differential Revision: D18579722

fbshipit-source-id: 5f54676a128207026cb16b9cdc03a055cbb1f9b9

Summary: This diff disables the purging of huge pages and hence they will never be freed and will be counted towards the RSS of the process. This was done to ensure that we don't block on madvise call to get the huge pages and increase memory pressure on the system.

Reviewed By: gdankel

Differential Revision: D18607196

fbshipit-source-id: 79e483a715290fdfbdd1963433ec3918561a0e0e

Summary:
Clang 7 seems to have trouble with `FOLLY_CREATE_MEMBER_INVOKER()` on Windows targets if the invoker's name is the same as the invocable name. It's spewing `-Wmicrosoft-explicit-constructor-call` warnings and the `static_assert`s in LockTraits.h are failing. As a workaround, we can rename the LockTraits.h invokers to use an `_invoker` suffix.

See https://gcc.godbolt.org/z/j4qRd9 for a demonstration of the issue.

Reviewed By: vitaut

Differential Revision: D18626335

fbshipit-source-id: c9147daaa10b02b0692c32963f9ede8cc645e329