Summary: In order to trigger trace points when a new thread begins/finishes executing, I added functionality to call onSet and onUnset methods on all RequestData objects whenever setContext is called. The main question: will this approach cost us too much overhead?

Reviewed By: djwatson

Differential Revision: D3604948

fbshipit-source-id: 3b704ca0f2b713458427aa49be12f776939057f8

Summary: This code very obviously wrote past the end of the buffer when the length was 1. Furthermore, it was just downright broken for all values. The author obviously meant to type * instead of +. I took the time to verify that the algorithm is actually correct, while I was working on this. My proof is in the test plan.

Reviewed By: yfeldblum, meyering

Differential Revision: D3603255

fbshipit-source-id: 5f2a0011ff5401a70ba03993eab6e53e29d87c1c

Summary: Because Windows already defines a funtion named `GetTickCount` with a different signature.

Reviewed By: yfeldblum

Differential Revision: D3601877

fbshipit-source-id: fa0fb422156a3dc71e149e02a000ccdf3479eea5

Summary: There was an MSVC section, but it's much simpler to just use the glue to solve the issue instead.

Reviewed By: yfeldblum

Differential Revision: D3601184

fbshipit-source-id: f95bd3dfc4b244e1bce21997f44de2970575ad7f

Summary: A `long` on MSVC is only 4 bytes, so the `static_assert` will fail. Remove the static assert and use the `long long` version instead. Do the same for the unsigned version.

Reviewed By: yfeldblum

Differential Revision: D3601064

fbshipit-source-id: 70c5f4bca597ba05c3729f5d15feeea3cc8fde57

Summary: Because we need it for MSVC.

Reviewed By: yfeldblum

Differential Revision: D3601138

fbshipit-source-id: 4d2ba3f9c972717745289686b025d2763f9ef30d

Summary: As the comment in the file says, this has to be included in a weird order due to issues on Windows.

Reviewed By: yfeldblum

Differential Revision: D3601008

fbshipit-source-id: fca4b6d906eb0d1f001c28197987165cd075752d

Summary: As the `sizeof()` is being evaluated in a static context, MSVC doesn't let you reference non-static locals. Solve the issue by getting the type of the local via `decltype` instead.

Reviewed By: yfeldblum

Differential Revision: D3600845

fbshipit-source-id: 825d93ced8f09d9f4bf0dcf02142f47a0ec32605

Summary:
This was causing MSVC to complain due to `DSched` already being declared in `BatonTestHelpers.h` as `::folly::DSched`, but it's then typedef'd again in `BatonTest.cpp`, this time as `::DSched`. MSVC complains about the duplicate definitions even though they resolve to the same type, so kill the one in the source file and keep the one in the header.
The same is the case with `GroupVarint32Decoder` and `GroupVarint64Decoder`, which are both declared in `GroupVarint.h` and then again, this time in an anon namespace, in `GroupVarintTest.cpp`.

Reviewed By: yfeldblum

Differential Revision: D3600948

fbshipit-source-id: 401790aa918289fe4a34dd5af683f8c6ca50d847

Summary:
uniform_int_distribution is implemented differently on this system and
always returns the generated number from our constant random number generator.

Reviewed By: yfeldblum

Differential Revision: D3585418

fbshipit-source-id: 754076599ba77b49c17f03bb0c92264a3bed4ab0

Summary:
Because it doesn't exist in the C++ standard.

This switches them to the explicitly expanded form, which is in the spec.
It also removes a few that were doing absolutely nothing. (MSVC still complained about the division by zero)

Reviewed By: yfeldblum

Differential Revision: D3479260

fbshipit-source-id: 5386e27057beeb4b228f5d6be4e1cf9941cf3176

Summary:
Sure, it's big, but it compiles.

There are a few different changes to make this happen, the first is to deal with quite a few places where MSVC was complaining about the implicit coersion of template and noexcept parameters to bool.
Next, the gating that disabled this for 4.7 is now gone.
MSVC's runtime expected `difference_type` to be defined on our custom allocator, so it now is.
`ReadTSC()` was changed to use MSVC's intrinsic for this.
And finally, the named variadic parameters to macros were switched to normal variadic macro params, as MSVC doesn't support named ones.

Reviewed By: yfeldblum

Differential Revision: D3479761

fbshipit-source-id: 703a5ef942ffc9b63381d13fc3960ac6f2780aa0

Summary: Because they aren't captured, and thus aren't accessible, in the callback used further on in the function.

Reviewed By: yfeldblum

Differential Revision: D3600874

fbshipit-source-id: 699b3d4caa0a310b1ccc7810d670671850f4366b

Summary: We don't have `fork()` to begin with, so disable the test and the includes it required if we're compiling on Windows.

Reviewed By: yfeldblum

Differential Revision: D3600600

fbshipit-source-id: 5399705753b456139105f4ac757190e1ba1765f2

Summary:
[Folly] `allocate_sys_buffer`.

For when a `malloc`'d buffer is required, with an associated deleter that calls `free`.

Reviewed By: JonCoens

Differential Revision: D3590516

fbshipit-source-id: 644f4b5d5e8f19dbc8f29efe3e93517fba0ad72f

Summary: Record whether cached certificate was used

Reviewed By: anirudhvr

Differential Revision: D3582807

fbshipit-source-id: 246107ce383ff31718ee7dcccf8bbea459b559a8

Summary:
This makes folly/init/Init.cpp compile on macOS by leveraging
equivalent functionality in the glog library when the folly symbolizer is not
available.  This is true for macOS and also for Windows.  I haven't done
anything to handle Windows in this diff.

Reviewed By: yfeldblum

Differential Revision: D3585509

fbshipit-source-id: 2e0c29520a53826acbf656a7a02659b4e905802f

Summary:
On Linux hosts, the fast path to get the current time in
nanoseconds without doing a syscall to the kernel is available via the
VDSO kernel-runtime interface.

In this diff, I:

1. Expose portability wrappers `folly::chrono::clock_gettime()` and
`folly::chrono::clock_gettime_ns()`
2. Implement a VDSO wrapper on Linux hosts to implement those without
a round-trip to the kernel

Depends On D3418054

Reviewed By: bmaurer

Differential Revision: D3418087

fbshipit-source-id: 3fb99f0dd946f19ba29d0d52a1038dad3556bafd

Summary: First cut at autoconf rules for generating a pkg-config file for Folly. The only interesting part is that we want to handle packages that provide their own .pc files as dependencies rather than just stuffing them into the list of libraries required from packages that don't provide .pc.

Reviewed By: simpkins

Differential Revision: D3556611

fbshipit-source-id: d1c5f89416d1d96d2cdf5ccdc5bd7117fb822a82

Summary: It is an error to attempt to get the size of a zero length array, such as `unsigned char[0]`, which is what you get if `MaxInline` has been passed in as 0. We can work around this by simply defining `InlineStorageType` to be `void*` if the size is exactly 0, which will result in the capacity correctly being stored out of line.

Reviewed By: yfeldblum

Differential Revision: D3572898

fbshipit-source-id: c96bb7cc6a890044bb74b0f6d5238c503552ee25

Summary: iters was only being used on one side of the producer/consumer.

Reviewed By: yfeldblum

Differential Revision: D3588593

fbshipit-source-id: 5e9a8a02bb9addbd6f936b8cc411d58cdf82b6e2

Summary: [Folly] Make the `mprotect` variant of `asymmetricHeavyBarrier` work when `mlock` fails.

Reviewed By: djwatson

Differential Revision: D3585948

fbshipit-source-id: c3a46884434b7f9da9caa9cf203573f9e3ce7444

Summary:
FOLLY_X64 is usually defined, so check whehter it is defined
to 0 or 1.  In addition, memory barriers are not necessarily widely
available on the mobile platforms. What leads me to believe this is
that, at least for older ndk's, atomics are not guaranteed to work for
multicore platforms.

Reviewed By: andriigrynenko

Differential Revision: D3586806

fbshipit-source-id: 3ac8c4b74cac09e41bc3cb65c8adc2732b8b2256

Summary:
macOS doesn't have timedwait functionality and this header file was
gating including the boost headers on that check which caused dependent modules
to fail to compile because `boost::mutex` and `boost::recursive_mutex` were not
known to the compiler.

Reviewed By: yfeldblum

Differential Revision: D3585470

fbshipit-source-id: 7f8d9603e95ce01328103c7c6ac0bc75a35ddf4d

Summary: Because, unfortunately, `pthread_exit` is not marked as noreturn in the pthread library primarily used on Windows, we need to return something to avoid errors.

Reviewed By: yfeldblum

Differential Revision: D3571465

fbshipit-source-id: 4713b2364d6584ae255fb83cd21406fdc33ee299

Summary: membarrier guarantees that there's at most one update to thread-local counter, which collecting thread may not see.

Reviewed By: djwatson

Differential Revision: D3532952

fbshipit-source-id: 6106bfe87c70c5f864573a424662778e20423bbb

Summary:
Since OS X 10.6 it is possible to set the name of the current thread.
This diff adjusts our setThreadName routine to do this, partially fixing
the associated test case.  Even though this doesn't completely cover
all cases it is still a valid improvement: most callers are
threads setting their own name.

I've amended the tests so that they can accomodate systems that cannot
set the names of other threads.

Reviewed By: yfeldblum

Differential Revision: D3576281

fbshipit-source-id: 13caf0dca6496aa2da897631e8d7327a6ee452bb

Summary: When installing the mock, we should make sure to remove singleton from the creation_order list.

Differential Revision: D3580725

fbshipit-source-id: dfb489de1be860ab639380644eab0b45a07a1450

Summary:
The check was overly restrictive, the conditional packing was originally used to limit it to GCC only (as it used the `__attribute__` directly), but that restriction is no longer needed as `FOLLY_PACK_*` will be defined appropriately for whatever platform we are currently on.

This just switches it to unconditionally use the `FOLLY_PACK_*` macros rather than defining it's own `FB_PACK_*` macros.

See https://github.com/facebook/folly/commit/aafd8ff9151615a8de09682d41d1c29966e4c8fe for where the `FB_PACK_*` macros could originally have been removed.

Reviewed By: yfeldblum

Differential Revision: D3373779

fbshipit-source-id: 9e41b0836f8ba12de4f37aba58c51f841be08b41

Summary: The static analysis doesn't quite understand that we are assigning a NUL to, at most, `patternBuf[MAX_PATH + 1]` so increase the buffer size slightly to make it happy.

Reviewed By: yfeldblum

Differential Revision: D3571923

fbshipit-source-id: c027e2b253dc3e56d13e9a15a0e960a6aa3a0e6c

Summary: Because there are situations where we need to do an atomic write to a plain pointer, rather than an atomic value. Unfortunately, there is no support in the standard library for this, so different compilers implement the primitives differently. This specifically implements an `int64_t` overload for `__sync_fetch_and_add` for use in the atomic hash map test.

Reviewed By: yfeldblum

Differential Revision: D3571830

fbshipit-source-id: c27d8d2a5238bbc9aba6a9e48e4b3412a199288f

Summary:
this fixes TimeTest on macOS.

The problem was that `task_info_count` is an in/out parameter and it wasn't correctly initialized due to a typo.  The lack of initialization meant that the `task_info()` call would fail at runtime.

Reviewed By: yfeldblum

Differential Revision: D3576262

fbshipit-source-id: 8388fa3c5020309d64a0fa7e01fd6cea80200219

Summary:
the RNG max and min methods must be marked constexpr in order to
compile.  The macos compiler doesn't know to propagate the constexpr-ness from
numeric_limits::max on its own, so we have to be explicit (this also matches
the annotations in ThreadLocalPRNG in Random.h)

Reviewed By: yfeldblum

Differential Revision: D3576189

fbshipit-source-id: e4eeb3361d1c48f582dad5a52e35cae133f007a1

Summary:
The compiler on macos is pedantic about the size of the 9UL not
matching the size of the type returned by `oneBasedMsbPos`.  Cast it to
the appropriate size.

Reviewed By: meyering

Differential Revision: D3576183

fbshipit-source-id: 41e9afc78eed2994e34238da119774e9bf6b7cea

Summary:
Getting some build errors when doing `buck test folly/...`. This seems to be fixing it.

  folly/test/FBStringTest.cpp:1290:23: error: unknown type name 'basic_fbstring'; did you mean 'basic_fstream'?
    using u16fbstring = basic_fbstring<char16_t>;
                        ^~~~~~~~~~~~~~
                        basic_fstream

Also the test `testHashChar16` was failing because only part of the string is used when computing hash.

Reviewed By: yfeldblum

Differential Revision: D3575858

fbshipit-source-id: cdd5bdb9653d50beaf0ec82b659d31354b345441

Differential Revision: D3572671

fbshipit-source-id: a80390921b41e47ed2794d48d943a9e4060c7135

Summary: In a separate diff, I made changes that switched include order, and suddently this file failed to compile. This fixes it.

Reviewed By: andriigrynenko

Differential Revision: D3574775

fbshipit-source-id: 02422d803708aa8490123d19aab5a16c10fdc36e

Summary: Gtest-1.7's ASSERT macros don't work with types with explicit conversions to bool, so do the cast for it.

Reviewed By: yfeldblum

Differential Revision: D3570782

fbshipit-source-id: 885a1022675e63afca00219c5b9de3791742207f

Summary: Recent changes to SharedMutex.h broke mcrouter's open source build. See https://travis-ci.org/facebook/mcrouter/builds/140608809. It looks like we're getting bitten by https://sourceware.org/bugzilla/show_bug.cgi?id=16773. We're hitting it now because the extern template is forcing the instantiation of the thread-local static member to be located in SharedMutex.o, but the inline members that reference that thread-local are *not* located in SharedMutex.o. binutils seems to be stepping on its own feet trying to fix up the references at link time. We can fix it by making sure the code that references the thread-local is colocated with the thread-local.

Reviewed By: yfeldblum

Differential Revision: D3498477

fbshipit-source-id: 86ea86812010ff1ef7351e6f8c106bb4291d0234

Summary: Because, if it follows the same pattern as a normal config flag, it simply won't be defined if it isn't available.

Reviewed By: yfeldblum

Differential Revision: D3571862

fbshipit-source-id: 3322ceb0ce349de16b29310b66d1333aa04da516