Reviewed By: yfeldblum

Differential Revision: D3691541

fbshipit-source-id: 9488c1487ebd0500a23300292215456ca3220f03

Summary:
Deleter helper object, which may release multiple ReadMostlyMainPtrs at once.
This allows underlying ref count implementation to perform expensive synchronization operations (e.g. asymmetric heavy barrier only once).

Reviewed By: yfeldblum

Differential Revision: D3691524

fbshipit-source-id: 3ac593b0d813345daba3a81ff4e2eb71b4db292e

Summary:
If one of the assertions failed, it would result in any threads that are still running accessing already-disposed data, so wait for the threads to exit, even when an exception is thrown.

This also slightly cleans up the handling of threads a bit further down, because the mess it was doing was completely unnecessary.

Reviewed By: yfeldblum

Differential Revision: D3692438

fbshipit-source-id: 9082ba248b2cf1062e46c97faf0bc062312ee9ae

Summary:
See the comment on `msvcReturnInvalidParams` for more info on the issue.
This also adds a use of it to TestUtilTest.

Reviewed By: yfeldblum

Differential Revision: D3691526

fbshipit-source-id: f0f596e9b4c830e1d31de01926162636757329e8

Summary: Also switch `pipe` to return a socket pair instead, as libevent can't poll against a pipe on Windows. And lastly, fix the file descriptor for sockets leaking when close is called.

Reviewed By: yfeldblum

Differential Revision: D3691255

fbshipit-source-id: c684242d255ac5158a4c805d432d345dac1b3bd8

Summary: Pah! This definitely wasn't causing the lock tests to segfault. Nope, not at all <_>.

Reviewed By: yfeldblum

Differential Revision: D3691120

fbshipit-source-id: 6601db637f22c7b0bc326907a2a7976f6df7c159

Summary:
Also handle invalid file descriptors correctly and switch away from `WSASendMsg` for the implementation of `sendmsg` due to limitations imposed by WinSock.
This also fixes up a few places that were explicitly referencing the global version of some socket functions, which was bypassing the socket portability layer.

Reviewed By: yfeldblum

Differential Revision: D3692358

fbshipit-source-id: 05f394dcc9bac0591df7581345071ba2501b7695

Summary: Just calling `SetEnvironmentVariableA` wasn't updating `_environ`, which meant that calls to `getenv` weren't reflecting the changes made via `setenv`. The correct way to implement it is using `_putenv_s`, but there's one problem with that: passing an empty string as the value to `_putenv_s` results in the environment variable being unset. To make it possible to set the environment variable to an empty string, we shall dive head-first into the implementation details of the CRT and emerge victorious by blatently ignoring the documentation of `getenv` and modifying the string it returns to terminate it early.

Reviewed By: yfeldblum

Differential Revision: D3691007

fbshipit-source-id: 350c2ec72ec90b9178a9a45b2c2ed2659b788e37

Summary: Windows is weird and calls it USERNAME instead, so just use one that everything agrees on: PATH

Reviewed By: yfeldblum

Differential Revision: D3691072

fbshipit-source-id: 579c6484736ef47e130049c29bef8b59c66a4482

Summary: Because the comment was a lie; sockets are blocking by default, not non-blocking.

Reviewed By: yfeldblum

Differential Revision: D3691145

fbshipit-source-id: 5d3c62b3573205fe416d77fe4b5b9fbd593ffd93

Summary:
The getTFOsucceeded doesn't really work right now
as intended. Currently we check right after sendmsg,
however we can only know whether or not the server
has acked the data after 1-RTT.

This removes getTFOSucceeded. Will fix in another diff.

Differential Revision: D3679235

fbshipit-source-id: 6b2bb01d3743ea7e68ad3cc9a26be6806f17ffbe

Summary:
[Folly] `constexpr_abs`.

Is `constexpr`.

Works over integral and floating types.

If given an integral type, the return type is the usigned version of that integral type, thereby avoiding the undefined behavior of `std::abs(std::numeric_limits<int>::min())`.

Reviewed By: simpkins

Differential Revision: D3654072

fbshipit-source-id: 24fefc0c3b055f78ba3e07472c38fb9c550e0f31

Removing noexcept specifications in constructors for Synchronized that call contextualLock() and contextualRLock()

Summary:
Most mutex lock() functions do not have a noexcept guarantee, saying that the
constructor for Synchronized based on whether the underlying constructor for
the type stored is not enough.  Although this will *rarely* cause bugs, it
probably can be fixed

Reviewed By: simpkins

Differential Revision: D3682974

fbshipit-source-id: ec0bb701d0af41ffc79128fe8db7935a5f19bc70

Summary:
We are definitely not prepared to handle
SIGPIPEs, so add MSG_NOSIGNAL to sendmsg
of AsyncSSLSocket.

This is a problem which exists in openssl
as well which calls send with flags = 0.
We recently made a change to move the send
into our control, so we can now supply the
flag

Reviewed By: yfeldblum

Differential Revision: D3686679

fbshipit-source-id: ff8fe662e62923c25876bdfd516352639505dca6

Summary:
Update the documentation for Synchronized to remove references to the
various SYNCHRONIZED macros, and document the lock() and withLock() APIs
instead.

Reviewed By: yfeldblum

Differential Revision: D3618871

fbshipit-source-id: e970f9a23e45831d62232eea2ba133a55e5a4a49

Summary: It calls syscall, but does not include the header.

Reviewed By: knekritz

Differential Revision: D3586636

fbshipit-source-id: 2ef05fc4b88f236a84b79afa708e98fe21529685

Summary:
Depends on D3648195

This test example is intended to demonstrate the methodology for using DeterministicSchedule support for auxiliary data and global invariants.
The main goal is fine-grained invariant checking, ideally after every shared update.
The secondary goals are:
- Minimize intrusion in the original code. In this proposed methodology, it is adding a friend.
- Minimize duplication of original tested code. Unfortunately, depending on the original code, it seems that significant duplication may be unavoidable if we don't want to change the original code.

This diff is primarily about the methodology for testing already developed code. I plan to apply what we agree on through this diff to the dynamic MPMCQueue code (D3462592).

A future goal to keep in mind is creating a methodology for developing new code with hooks for DSched aux. data and invariant checking integrated in it, in order to minimize or eliminate duplication of tested code.  In past projects, I used non-standard source code (basically algorithm code) that is automatically translatable through scripts and macros to input to a DSched-like tool as well as to compilable code. The main challenge for such a methodology is to allow the original source code to be standard readable C++ code.

Reviewed By: djwatson

Differential Revision: D3675447

fbshipit-source-id: aae2c9f0550af88dc3a5dcbe53318a75a86b6e2b

Summary: This switches the places in Folly that need to explicitly reference the socket portability implementation to do exactly that.

Reviewed By: yfeldblum

Differential Revision: D3299984

fbshipit-source-id: 57cd8ebe66c9055aba66581a8c0fcf6c125d96f9

Summary:
Add some tests to test the other levels of the timer wheel.

They are slow by necessity.

Reviewed By: yfeldblum

Differential Revision: D3637132

fbshipit-source-id: badf8d37d726dbeb5a8220d50c60b8efdaee7989

Summary: This feature doesn't make sense when wheeltimer doesn't constantly tick - we always have to get the current clock time.  On the plus side, we'll only be grabbing the clock on timer schedule or timeout, never for individual ticks.

Reviewed By: yfeldblum

Differential Revision: D3637088

fbshipit-source-id: ed8fe52419259332a14b6dc1d357979dcf258a20

Summary: It's perfectly possible that the default crypto context simply hasn't been created yet, so try to create it if the initial acquisition fails.

Reviewed By: yfeldblum

Differential Revision: D3673138

fbshipit-source-id: 122955df04055ff4f99513b182375d4388dd0305

Summary: As-of C++11, destructors are assumed to be `noexcept` unless explicitly marked otherwise. `Watchdog`'s destructor throws, so switch it to a `LOG(FATAL)` instead.

Reviewed By: andriigrynenko

Differential Revision: D3672621

fbshipit-source-id: 5224ecf85d101462e02e12da257e033bab4db1a1

Adding policies for all the upgrade and downgrade mutex transitions that are going to be supported by Synchronized

Summary:
This diff contains the lock policies that will be used by Synchronized to
implement mutex transitions by means of heterogenous RAII mutex wrappers

Reviewed By: yfeldblum

Differential Revision: D3665020

fbshipit-source-id: a5509dfd58a1dd6cd60a7d3afe929d0da860926d

Summary:
This diff adds support for upgradable mutexes to the LockTraits
abstraction used by folly::Synchronized

Reviewed By: yfeldblum

Differential Revision: D3645453

fbshipit-source-id: 30f16eb3fbebc687a4136256f1103962c0e4c465

Summary: Because the former is standardized, and the latter is a syntax error under MSVC.

Reviewed By: yfeldblum

Differential Revision: D3672667

fbshipit-source-id: c0149b11367e36ba5574625240c41a167348199f

Summary: RequestContext update.

Reviewed By: mibpl

Differential Revision: D3604937

fbshipit-source-id: 07b315fe22a6dba40f0bab53c8092f35c172ae85

Summary: The portability implementation of `mmap` for Windows doesn't currently support anonymous shared allocations, as they are non-trivial to manage, and the places this is being used doesn't actually need the memory to be allocated as shared, so allocate it as private instead.

Reviewed By: yfeldblum

Differential Revision: D3671495

fbshipit-source-id: e74d4fd925363fef16c1bf8445da1ff32bf8266d

Summary: The tests aren't dependent on it being pthread, so use standardized stuff instead.

Reviewed By: yfeldblum

Differential Revision: D3665698

fbshipit-source-id: ad91facb75a9c5d7a90bfa294cc98bb07629de1b

Summary: Without `WIN32_LEAN_AND_MEAN` defined, including `Windows.h` includes a massive number of extra headers. There are situations that need things that are defined in those headers, so keep folly working even when it isn't defined.

Reviewed By: yfeldblum

Differential Revision: D3671350

fbshipit-source-id: 0de986e17cdd4be386be8a9eeb0075b4af44024d

Summary: To make it possible for things to actually compile, this adds `socketpair` and the `sockaddr_un` struct, and also removes the `FOLLY_HAVE_UNIX_SOCKETS` define, as it is no longer needed.

Reviewed By: yfeldblum

Differential Revision: D3671158

fbshipit-source-id: 7333470f2e85c24edb935be5e1b94b4edc6e7267

Summary: The comment tells the truth: StaticMeta doesn't mix well with MSVC's TLS implementation.

Reviewed By: yfeldblum

Differential Revision: D3671274

fbshipit-source-id: 293eb32699ad5cee59965f24f985ec7099c90136

Summary: Because the lint error was annoying me.

Reviewed By: yfeldblum

Differential Revision: D3671252

fbshipit-source-id: 2962db7616077f531e64c6f2b26ad708c7adf6ed

Summary:
This is done by using the intrinsic directly instead.
This also marks the input parameter as `volatile`, to make it more clear that that's how it is handled by `_InterlockedExchangeAdd64`.

Reviewed By: yfeldblum

Differential Revision: D3671320

fbshipit-source-id: 6010085ec3b6952a3eb1e952965ec6ad87566db2

Summary: Because it contains `wchar_t`'s on Windows.

Reviewed By: yfeldblum

Differential Revision: D3672098

fbshipit-source-id: b49db642dedf48bf7379cf339130c93f52a22f9a

Summary: Which is the case when compiling under MSVC.

Reviewed By: pixelb

Differential Revision: D3671357

fbshipit-source-id: b7e51479cae44933c3f1caa21d45775a32f294a3

Summary:
Because we're constructing the value with a placement new, which has some of C++'s most unhelpful behavior ever put into the spec: If `p` is `nullptr` and we are not compiling in C++14 mode, where the restriction was changed, then the placement new will do absolutely nothing at all. By adding the assumption that `p` is not `nullptr`, we'll trip a segfault in release mode rather than failing silently.
Note that MSVC would generate the nullptr check regardless of which mode it's in, so this assume forces the removal of the check.

Reviewed By: yfeldblum

Differential Revision: D3651116

fbshipit-source-id: ee15a38f85ce4e3cb3186fda0b7bcca39acda27a

Summary: melaniesubbiah introduced onSet / onUnset callbacks on RequestData in D3604948; we need unset() to be called when an RC is overriden with RequestContext::create() so that things work as expected. Also, change the order of calling onSet / onUnset - from RequestData perspective, it shouldn't look like there are two contexts set at the same time

Reviewed By: palmtenor

Differential Revision: D3667017

fbshipit-source-id: b9bfb858fe65ffb11de8e6d6f13b8f4cf6266bc9

Summary: There are a few places that need this, so make it so.

Reviewed By: yfeldblum

Differential Revision: D3667234

fbshipit-source-id: 84c16e1b036d9c2afddf03aa84c43210c5672f0a

Summary: Well, make it compile anyways. There's no simple equivelant for Windows, so just return 0 instead.

Reviewed By: yfeldblum

Differential Revision: D3667474

fbshipit-source-id: 02224c6666dfcfdec237bfbbd4714170407a952a

Summary: Because they simply didn't work before.

Reviewed By: yfeldblum

Differential Revision: D3666894

fbshipit-source-id: 7a34073b2bbc43c489ad963c28fac24dcbb187f2