Summary: Adds an AdFork::unregisterHandler command, such that we can register and unregister fork handlers.

Reviewed By: yfeldblum

Differential Revision: D6439796

fbshipit-source-id: b710152f8cf98371c330b484cd5fe42de44e74ac

Summary:
UnboundedQueue supports:
- SPSC, MPSC, SCMP, MPMC
- Non-waiting, waiting, and timed consumer operations.
- Producers never wait or fail (unless out-of-memory).
- Memory usage grows and shrinks dynamically

```
/// UnboundedQueue supports a variety of options for unbounded
/// dynamically expanding an shrinking queues, including variations of:
/// - Single vs. multiple producers
/// - Single vs. multiple consumers
/// - Blocking vs. spin-waiting
/// - Non-waiting, timed, and waiting consumer operations.
/// Producer operations never wait or fail (unless out-of-memory).
///
/// Template parameters:
/// - T: element type
/// - SingleProducer: true if there can be only one producer at a
///   time.
/// - SingleConsumer: true if there can be only one consumer at a
///   time.
/// - MayBlock: true if consumers may block, false if they only
///   spins. A performance tuning parameter.
/// - LgSegmentSize (default 8): Log base 2 of number of elements per
///   segment. A performance tuning parameter. See below.
///
/// When to use UnboundedQueue:
/// - If a small bound may lead to deadlock or performance degradation
///   under bursty patterns.
/// - If there is no risk of the queue growing too much.
///
/// When not to use UnboundedQueue:
/// - If there is risk of the queue growing too much and a large bound
///   is acceptable, then use DynamicBoundedQueue.
/// - If the queue must not allocate on enqueue or it must have a
///   small bound, then use fixed-size MPMCQueue or (if non-blocking
///   SPSC) ProducerConsumerQueue.
///
/// Template Aliases:
///   USPSCQueue<T, MayBlock, LgSegmentSize>
///   UMPSCQueue<T, MayBlock, LgSegmentSize>
///   USPMCQueue<T, MayBlock, LgSegmentSize>
///   UMPMCQueue<T, MayBlock, LgSegmentSize>
///
/// Functions:
///   Producer operations never wait or fail (unless OOM)
///     void enqueue(const T&);
///     void enqueue(T&&);
///         Adds an element to the end of the queue.
///
///   Consumer operations:
///     void dequeue(T&);
///         Extracts an element from the front of the queue. Waits
///         until an element is available if needed.
///     bool try_dequeue(T&);
///         Tries to extracts an element from the front of the queue
///         if available. Returns true if successful, false otherwise.
///     bool try_dequeue_until(T&, time_point& deadline);
///         Tries to extracts an element from the front of the queue
///         if available until the specified deadline.  Returns true
///         if successful, false otherwise.
///     bool try_dequeue_for(T&, duration&);
///         Tries to extracts an element from the front of the queue
///         if available for for the specified duration.  Returns true
///         if successful, false otherwise.
///
///   Secondary functions:
///     size_t size();
///         Returns an estimate of the size of the queue.
///     bool empty();
///         Returns true only if the queue was empty during the call.
///     Note: size() and empty() are guaranteed to be accurate only if
```

Reviewed By: djwatson

Differential Revision: D6157613

fbshipit-source-id: db423f86d1d0604d22f6b9c71ea0ed08be32e2a1

Summary:
X509_digest, or the "fingerprint" of the cert, is useful to identify
certificates with the same CN/SANs. This adds an API to get the SHA1 or SHA256
digest of a cert.

Reviewed By: yfeldblum

Differential Revision: D6457917

fbshipit-source-id: 148a5a2270e938b39065e00d7947c2fc57cd7f75

Summary:
AFAICT this is currently the only thing preventing the non-dynamic version of Folly's MPMCQueue from being allocation-free on enqueue and dequeue operations. Allocating this with a constant expression should move the allocation from runtime to link (possibly compile?) time, which will let us use it in allocation sensitive contexts.

Feel free to suggest other reviewers, I couldn't find an existing folly reviewer group.

Reviewed By: yfeldblum, agola11, nbronson

Differential Revision: D6447848

fbshipit-source-id: 86b84b19d62f1e1bcecdb9e757a6dfa90597b084

Summary: As in title, for eg larger stack size

Reviewed By: yfeldblum

Differential Revision: D6466762

fbshipit-source-id: 303e8732c8bfb64fbd8619163779942fe5f8b5a8

Summary:
Add makeSystemError() helper functions that are similar to the existing
throwSystemError() functions but just return the exception rather than throwing
it.  This is helpful for callers using folly::Expected or folly::Future, where
they need to return an exception type rather than throwing it.

This also includes comments about the fact that this code is using the wrong
error category.  The C++ standard indicates that std::generic_category() for
POSIX errno values.  I am not fixing this as part of this diff, since this
change has the potential to break existing users, and would need a fair amount
of testing first.

Reviewed By: yfeldblum

Differential Revision: D6456771

fbshipit-source-id: 4724b51b8d4a7f513ae70ea1b0c4f0516cfc205f

Summary:
Add an explicit `std::move()` around the return value of
`logConfigToDynamic()`.  This explicit move is required pre-C++14,
but is not required after DR 1579:
http://www.open-std.org/jtc1/sc22/wg21/docs/cwg_defects.html#1579

Newer versions of g++ and clang no longer require the move.  However gcc
versions earlier than 5.0 do require it.

Reviewed By: yfeldblum

Differential Revision: D6466447

fbshipit-source-id: 1b15934490d4966e9b3e5612e0e2ecbb43c979ca

Summary:
Replace the initLoggingGlogStyle() function with a more generic initLogging()
function that accepts a log config string to be parsed with parseLogConfig().

Reviewed By: bolinfest, yfeldblum

Differential Revision: D6342086

fbshipit-source-id: fb1bffd11f190b70e03e2ccbf2b30be08d655242

Reviewed By: yfeldblum

Differential Revision: D6453972

fbshipit-source-id: fa083c0b782188f5f45a2a354045aee03e528829

Summary:
Implement conversions between std::chrono::duration and POSIX-style time
structures even when neither the numerator nor the denominator of the duration
ratio are 1.

Both of these are done by first converting to an intermediate type where the
numerator is 1, and then using the conversion routines for that case.

Reviewed By: yfeldblum

Differential Revision: D6366647

fbshipit-source-id: 8f9495fb4101cac6d8b4cf0353a679107007b298

Summary: As title

Reviewed By: yfeldblum

Differential Revision: D6440723

fbshipit-source-id: 3168d7bb616ae0ff3fe42f7584c5a255c4953875

Summary:
Fix bug in non-jemalloc build.

(Note: this ignores all push blocking failures!)

Reviewed By: igorsugak

Differential Revision: D6451342

fbshipit-source-id: 44725d4e0c685a59325fa3d08877d0f62120c5c2

Summary:
[Folly] Let `SaturatingSemaphore::try_wait` be non-`const` and add `ready`.

For internal API consistency.

Reviewed By: magedm

Differential Revision: D6450089

fbshipit-source-id: 65b9b723672521710a69719b192eb2922a27b778

Summary: Add preprocessor checks to guard jemalloc headers if any sanitizer is enabled.

Reviewed By: yfeldblum

Differential Revision: D6441578

fbshipit-source-id: 85364fca5af33c95cd05f232d1e3f9cbac9a4120

Summary:
[Folly] Some fixes for custom conversions of enums.

Of note, `to` was defined for enum -> all conversions, including enum -> string conversions, but we actually want enum -> string conversions to be done via ADL-discovered toAppend.

Reviewed By: ot

Differential Revision: D6411250

fbshipit-source-id: 852b64309e6adf1c68e5153635cb29632e2d86d4

Summary:
SaturatingSemaphore is a flag that allows:
- multiple concurrent posters
- multiple concurrent waiters
- idempotent posting
- non-destructive waiting
- blocking and spinning
- pre-block spin time control

```
/// SaturatingSemaphore is a flag that allows concurrent posting by
/// multiple posters and concurrent non-destructive waiting by
/// multiple waiters.
///
/// A SaturatingSemaphore allows one or more waiter threads to check,
/// spin, or block, indefinitely or with timeout, for a flag to be set
/// by one or more poster threads. By setting the flag, posters
/// announce to waiters (that may be already waiting or will check
/// the flag in the future) that some condition is true. Posts to an
/// already set flag are idempotent.
```

Reviewed By: djwatson

Differential Revision: D6379704

fbshipit-source-id: 59aed76caa2d159639e75425a778a9c63f18f375

Summary: Most of the time we want to guard some code when any of the sanitizers are enabled not just one.

Reviewed By: yfeldblum

Differential Revision: D6429720

fbshipit-source-id: 7157a0a13f2b2acabd36a19c61b9fdbf38ba96b8

Summary:
Add a method for merging the settings from two LogConfig objects.
This allows LogConfig objects to be merged before applying them to the
LoggerDB.  The effects are the same as two sequential LoggerDB::updateConfig()
calls, but without having to apply the intermediate state to the LoggerDB.

Reviewed By: bolinfest

Differential Revision: D6342085

fbshipit-source-id: 0f8a1b7d8d195a80bc74342444dd3152d331fcb6

Summary:
Add a static_assert() to check that the FLAG_INHERIT bit does not conflict with
valid LogLevel values.

Reviewed By: bolinfest

Differential Revision: D6341242

fbshipit-source-id: cec284623c8a612a3c54f1b593d169310790616e

Summary:
The logging code used to support a LogLevel::ERROR value on non-Windows
platforms.  I removed it in D5288600 to make all platforms consistent, but I
forgot to remove this comment describing it.

Reviewed By: bolinfest

Differential Revision: D6341244

fbshipit-source-id: 054427d342066f354c859b12611b907dc2d4fa35

Summary: Add methods for applying config changes from a LogConfig object to the LoggerDB.

Reviewed By: bolinfest

Differential Revision: D6200564

fbshipit-source-id: a25eb99e84b2885bf6853e2222db0d7432a6c37b

Summary: Add a method to get the current LogConfig state from the LoggerDB.

Reviewed By: bolinfest

Differential Revision: D6200596

fbshipit-source-id: 3bc57d498a5d25d19099d861376d71ea9f7e4039

Summary:
Remove the logic that clamps the log level to DFATAL in debug builds.

This behavior logically makes some sense, but results in subtle and potentially
confusing behavior changes.  It seems confusing that after calling
`setLevel(LogLevel::MAX_LEVEL)` on a log category, calling `getLevel()` on that
object would return `DFATAL` rather than `MAX_LEVEL`.  We also weren't clamping
the level to DFATAL consistently: `setLevel()` would clamp the value, but on
construction the initial level was still set to MAX_LEVEL rather than DFATAL.
This resulted in some issues when implementing `LoggerDB::getConfig()` since we
could not consistently detect if a log category was using the default level
settings.  Rather than fix this inconsistency it seems better to simply remove
this clamping behavior.

This means that it is possible for users to disable DFATAL log messages even in
debug builds if they really want to.  Previously this was only allowed in
release builds.  This protection doesn't really seem all that
valuable--presumably most developers won't really want to do this, and if they
really do request this configuration it doesn't seem all that bad to honor it.

Reviewed By: bolinfest, yfeldblum

Differential Revision: D6200569

fbshipit-source-id: 83ef8e2e4d3b61bc5b105038cbe3132979e9ac67

Summary:
Add a method to LogHandler to return its current configuration.  This will
make it possible to query the LoggerDB for its current configuration state.

Reviewed By: bolinfest

Differential Revision: D6200563

fbshipit-source-id: 2b8b9752bbeb26c8aac28d1a73b7e2312fd198c8

Summary:
Update the LoggerDB to track a list of LogHandlers by name, and
LogHandlerFactories by handler type.

This will be needed to support updating the LoggerDB configuration from a
LogConfig object.

Reviewed By: bolinfest

Differential Revision: D6200562

fbshipit-source-id: e365b4e0df65aa5aaa34e118eb3cee9c9c45cb05

Summary:
[Folly] Synchronize coupled caches in `folly::threadlocal_detail::StaticMeta`.

The caches should be set together, and only together, because they are coupled. This prevents bugs where one function that sets one cache but not the other cache is inlined into the caller in one module, and another function that reads both caches is inlined into the caller in another module.

Reviewed By: djwatson

Differential Revision: D6435175

fbshipit-source-id: 846c4972b40e525f2c04da6e6609c2ad54f674c0

Summary:
Add a new LogHandlerFactory interface for creating LogHandler objects from a
LogHandlerConfig.

Also add an initial FileHandlerFactory implementation capable of creating
LogHandler objects that write to a file descriptor.

Reviewed By: bolinfest

Differential Revision: D6200567

fbshipit-source-id: 14b86fc14ad475223aa4b57d45c40638b48c7594

Summary:
Just testing that ElfFile returns expected error code and message for
non-ELF files.

Depends on D6410210

Reviewed By: yfeldblum

Differential Revision: D6418365

fbshipit-source-id: aaab3b9f3ca1a12d384ae98a1772b7f640115192

Summary:
We need to keep LockedPtr returned by rlock/wlock alive for the entire duration
of the loop. Else we are working on a snapshot of the data structure, which
might not be what we want.

Reviewed By: aary, ot, luciang

Differential Revision: D6426613

fbshipit-source-id: 294c0a7b9d756f3116d5cdb3a7b3678f693e0754

Summary:
Replace all assert() checks in the folly logging code with
FOLLY_SAFE_DCHECK().

Reviewed By: yfeldblum

Differential Revision: D6422885

fbshipit-source-id: 5cb12dd59f2fe5d346f15b9d317abede8654a879

Summary:
_mm_cvtsi64_si128 and _mm_cvtsi128_si64 intrinsics are available on
x86_64 only, so do not enable SSE instructions when compiling for 32 bits.

Reviewed By: yfeldblum

Differential Revision: D6422271

fbshipit-source-id: a3c5cf3f2ee4b47d5a1edcf93cfff3d1f3fe907e

Summary:
[Folly] Remove some custom test `main()`'s in `folly/experimental/symbolizer/`.

In one of the two cases, we need to do some extra work to give `ElfFile` knowledge of the program path.

Reviewed By: ot, Orvid, philippv, luciang

Differential Revision: D6420517

fbshipit-source-id: 11c694a8342abb6fff05c368f3cda2e6287ccb4b

Summary:
This seems to fix issue in the attached task.

`context_` is set in `scheduleTimeout()` and never gets reset. So when an AsyncTimeout object is held across requests by objects like `IOThreadPool0` in I/O Thread Pool used by wangle/acceptor/ConnectionManager, RequestContext object created for the last request is leaked until `IOThreadPool0` handles another request and overrides `context_`. In the issue described in attached task, unit test has single request, next request never comes in, RequestContext doesn't end until test service stops, logging feature relying on RequestContext dtor doesn't get called in time, and thus unit test fails.

Reviewed By: yfeldblum

Differential Revision: D6402268

fbshipit-source-id: 200c6d358dfa6d7d9aa68ab05f6f1c7f4117b0ec

Summary:
Use FOLLY_SAFE_DCHECK() rather than assert() to avoid compiler errors about
this variable being unused in builds where assert() statements are disabled.

Reviewed By: bolinfest, yfeldblum

Differential Revision: D6421978

fbshipit-source-id: ab1019fa1fa420856e1bf3bc39c441662c4efd75

Summary: Now that we're using C++14 more broadly, we should use it to improve generator performance. This speeds up a microbenchmark >5x by removing type erasure.

Reviewed By: philippv

Differential Revision: D6398730

fbshipit-source-id: 5809058a3b5ff0e66fd4b1e8954698944e1a7d09

Summary:
When loading cert/key pair, order matters:
(a) Wrong key will fail to load if a cert is loaded;
(b) Wrong cert will succeed to load even if a private key is loaded.

So this diff adds:
(1) SSLContext::checkPrivateKey() -- must call for case (b).
(2) SSLContext::loadCertKeyPairFromBufferPEM() -- use this if one loads both cert and key. Guaranteed to throw if cert/key mismatch.

Reviewed By: yfeldblum

Differential Revision: D6416280

fbshipit-source-id: 8ae370883d46e9b5afb69c506c09fbf7ba82b1b9

Summary:
ElfFile in init() assumes file is long enough to contain complete
ElfHeader, which is 64 bytes long. This is not true for valid non-ELF files like
short scripts. They shouldn't cause crash.

Reviewed By: myreg

Differential Revision: D6410210

fbshipit-source-id: 28fd017d8de17c431d7d006a1655ade8a95994bd

Summary: Use newest hazptr hotness in concurrenthashmap.  Shaves ~10% off of the single-thread find performance.

Reviewed By: magedm

Differential Revision: D6259947

fbshipit-source-id: 7ecf99d38fdf8e311fca3313137e0fca5af3f165

Summary: This would save some function calls and reduce code size if we know we will use jemalloc.

Reviewed By: yfeldblum

Differential Revision: D6411810

fbshipit-source-id: 9db471678e473945f5f00354c0a1285ecdb91de6

Summary:
Move `folly::hasher` closer to feature parity with `std::hash`.
This is in order to replace some instances of `folly::hash::hash_combine(...)` with `folly::Hash()(...)` (`std::hash` is the identity for integers, which makes it an unsafe default for more sophisticated hash data structures, including open-addressing hash tables).

The implementation is similar to `libstdc++`'s implementation, in that we handle separately the `0` case, because `0` and `-0` have different binary representations but are equal according to `operator==`, and hash the bytes otherwise. It is probably a little faster than `libstdc++`'s implementation, that delegates a out-of-line Murmur hash routine for arbitrary buffers, while this uses a minimal inlineable machine word hashing routine.

Reviewed By: yfeldblum

Differential Revision: D6410713

fbshipit-source-id: 86d9e4ed8da04fffe283949825852e539ec7d5cf