Summary:
The priority queue in the manual executor implementation is backwards. This means that scheduled things run in reverse order, and a later thing will block an earlier thing if you advance to a timestamp in between the two.

This diff fixes the problem and adds tests to confirm the fix. These tests fail on the old implementation.

Reviewed By: yfeldblum

Differential Revision: D4739101

fbshipit-source-id: 6e429828460df5b3c656580568a9ae1eb4009527

Summary: Macros are not the best solution anymore.

Reviewed By: yfeldblum

Differential Revision: D4270626

fbshipit-source-id: 8acbd5b778d4effcfb587655b4397737bef0d379

Remove/make private the default ***Holder constructor to allow compile time detection of ***Holder(mutexPtr_) constructs

Summary: Avoid unintended C++ stuff - this is not the same as the "fleeting rvalue" - this change might break the build but I will fix it

Reviewed By: yfeldblum

Differential Revision: D4814104

fbshipit-source-id: 058a0eac44893c573062fcf5665d4fd022ee64a0

Summary:
A (almost) lock-free atomic_shared_ptr, matching the proposed concurrency TS interface.

See notes at top of file

Reviewed By: davidtgoldblatt

Differential Revision: D4716098

fbshipit-source-id: b9ca2443ba9e227ebb6f40807128073c6e14222a

Summary:
`escapeString` is very slow even when there's very little to nothing to escape. This diff adds a fast path to copy sequences of bytes that don't need escaping.
It also optimizes appending escape sequences: `string::push_back` is slow because it has to do a capacity check for every character.

Before:
```
  ============================================================================
  folly/test/JsonOtherTest.cpp                    relative  time/iter  iters/s
  ============================================================================
  jsonSerialize                                              818.55ns    1.22M
  jsonSerializeWithNonAsciiEncoding                            1.35us  738.06K
  jsonSerializeWithUtf8Validation                              1.42us  705.60K
  jsonSerializeAsciiWithUtf8Validation                         3.27us  306.06K
  parseSmallStringWithUtf                                      1.91us  522.38K
  parseNormalString                                            1.51us  660.27K
  parseBigString                                             384.44ns    2.60M
  toJson                                                     480.54ns    2.08M
  ============================================================================
```

After:
```
  ============================================================================
  folly/test/JsonOtherTest.cpp                    relative  time/iter  iters/s
  ============================================================================
  jsonSerialize                                              781.69ns    1.28M
  jsonSerializeWithNonAsciiEncoding                          847.68ns    1.18M
  jsonSerializeWithUtf8Validation                            928.68ns    1.08M
  jsonSerializeAsciiWithUtf8Validation                       199.85ns    5.00M
  parseSmallStringWithUtf                                      1.93us  518.39K
  parseNormalString                                            1.45us  689.11K
  parseBigString                                             378.66ns    2.64M
  toJson                                                     446.38ns    2.24M
  ============================================================================
```

All string escaping benchmarks are slightly faster, and ASCII-only with no escapes is 8x faster.

Reviewed By: luciang, evilmucedin, yfeldblum

Differential Revision: D4793233

fbshipit-source-id: c40d07708bd787799c8c00f9f23a417b862ee9ae

Summary: Infinite retries is something that is often needed for read-modify-write like workflows. The current implementation was creating a nested chain of implicit promises. This manifests as a memory leak after some time. Worse yet, even if it succeeds, it will take a long time to churn thru the chain of promises.

Reviewed By: yfeldblum

Differential Revision: D4770335

fbshipit-source-id: 44b8d6df1084de4514b66919a9838cf2322d6dce

Summary:
ubsan flags this problem:

```
buck-out/opt-ubsan/gen/folly/__default_headers__#header-mode-symlink-tree-with-header-map,headers/folly/experimental/StringKeyedCommon.h:31:18: runtime error: null pointer passed as argument 2, which is declared to never be null
third-party-buck/gcc-4.9-glibc-2.20-fb/build/glibc/include/string.h:47:28: note: nonnull attribute specified here
    #0 0x5cb88f in std::pair<std::__detail::_Node_iterator<std::pair<folly::Range<char const*> const, facebook::eden::GitIgnore>, false, true>, bool> folly::StringKeyedUnorderedMap<facebook::eden::GitIgnore, folly::Hash, std::equal_to<folly::Range<char const*> >, std::allocator<std::pair<folly::Range<char const*> const, facebook::eden::GitIgnore> > >::emplace<facebook::eden::GitIgnore>(folly::Range<char const*>, facebook::eden::GitIgnore&&) buck-out/opt-ubsan/gen/folly/__default_headers__#header-mode-symlink-tree-with-header-map,headers/folly/experimental/StringKeyedCommon.h:31
    #1 0x5c6652 in facebook::eden::(anonymous namespace)::IgnoreChecker::isIgnored(facebook::eden::detail::RelativePathBase<folly::Range<char const*> >) eden/fs/inodes/Dirstate.cpp:226
    #2 0x5c6037 in facebook::eden::(anonymous namespace)::IgnoreChecker::isIgnored(facebook::eden::detail::RelativePathBase<folly::Range<char const*> >) eden/fs/inodes/Dirstate.cpp:180
    #3 0x5c5173 in facebook::eden::Dirstate::getStatusForExistingDirectory(facebook::eden::detail::RelativePathBase<folly::Range<char const*> >) const eden/fs/inodes/Dirstate.cpp:372
    #4 0x5c4509 in facebook::eden::Dirstate::getStatus() const eden/fs/inodes/Dirstate.cpp:272
    #5 0x50a448 in verifyExpectedDirstate(facebook::eden::Dirstate const*, std::unordered_map<std::basic_fbstring<char, std::char_traits<char>, std::allocator<char>, std::fbstring_core<char> >, facebook::eden::StatusCode, std::hash<std::basic_fbstring<char, std::char_traits<char>, std::allocator<char>, std::fbstring_core<char> > >, std::equal_to<std::basic_fbstring<char, std::char_traits<char>, std::allocator<char>, std::fbstring_core<char> > >, std::allocator<std::pair<std::basic_fbstring<char, std::char_traits<char>, std::allocator<char>, std::fbstring_core<char> > const, facebook::eden::StatusCode> > >&&) eden/fs/inodes/test/DirstateTest.cpp:48
    #6 0x50c70c in Dirstate_addDirectoriesWithMixOfFiles_Test::TestBody() eden/fs/inodes/test/DirstateTest.cpp:220
    #7 0xb45857 in void testing::internal::HandleExceptionsInMethodIfSupported<testing::Test, void>(testing::Test*, void (testing::Test::*)(), char const*) /home/engshare/third-party2/gtest/1.7.0/src/gtest-1.7.0/./src/gtest.cc:2364
    #8 0xb36784 in testing::Test::Run() /home/engshare/third-party2/gtest/1.7.0/src/gtest-1.7.0/./src/gtest.cc:2437
    #9 0xb36957 in testing::TestInfo::Run() [clone .part.558] /home/engshare/third-party2/gtest/1.7.0/src/gtest-1.7.0/./src/gtest.cc:2612
    #10 0xb36b74 in testing::TestCase::Run() [clone .part.559] /home/engshare/third-party2/gtest/1.7.0/src/gtest-1.7.0/./src/gtest.cc:2587
    #11 0xb3806e in testing::internal::UnitTestImpl::RunAllTests() [clone .part.561] /home/engshare/third-party2/gtest/1.7.0/src/gtest-1.7.0/./src/gtest.cc:4571
    #12 0xb382d9 in testing::UnitTest::Run() /home/engshare/third-party2/gtest/1.7.0/src/gtest-1.7.0/./src/gtest.cc:4519
    #13 0x861ca7 in main third-party-buck/gcc-4.9-glibc-2.20-fb/build/gtest/include/gtest/gtest.h:2326
    #14 0x7f20be0740f5 in __libc_start_main /home/engshare/third-party2/glibc/2.20/src/glibc-2.20/csu/libc-start.c:289
    #15 0x4a552c in _start /home/engshare/third-party2/glibc/2.20/src/glibc-2.20/csu/../sysdeps/x86_64/start.S:122

UndefinedBehaviorSanitizer: undefined-behavior buck-out/opt-ubsan/gen/folly/__default_headers__#header-mode-symlink-tree-with-header-map,headers/folly/experimental/StringKeyedCommon.h:31:18
```

The issue is that `StringPiece` default constructs to `{nullptr, nullptr}` as a valid representation of an empty string.
This is tripping up UBSAN in this case.

The fix is a trivial nullptr check

Reviewed By: igorsugak

Differential Revision: D4791015

fbshipit-source-id: dec7484b29ecb29c17b8dd6a9b0e8093f07d63cb

Summary:
[Folly] Some constexpr symbols for whether sanitizers are enabled.

Useful where we need to do or not do some particular thing when sanitizers are enabled or disabled. Such cases do happen. It is better to use a C++ `constexpr` symbol rather than a preprocessor symbol in such cases, where possible.

Reviewed By: meyering

Differential Revision: D4805110

fbshipit-source-id: ca9ce074edef73dce3a04693dda102698c9cd29f

Summary:
[Folly] Fix double-`delete` in `Singleton::make_mock`.

In that function, we unconditionally destroy any existing singleton instance of the same singleton. Actually, we must conditionally destroy it - conditionally on one having been created and not yet destroyed.

This problem only rarely appears because it is masked by `delete static_cast<T*>(nullptr)` being a no-op. For example, if we call `make_mock` before actually instantiating the singleton, we call that no-op. The way to make it appear is to instantiate the singleton, and then to call `make_mock` twice in a row. The first call to `make_mock` unconditionally destroys the existing instance (it still should have checked) and the second call does it again, but because the existing instance is not `nullptr`, the second call is a double-delete of a non-`nullptr` instance and crashes. In the simple case, as reproduced in an attached test, the failure is observable with ASAN. In other cases, the double-`free` is tolerated, but the failure may only be observable depending on the singleton object's state - if running the dtor twice fails.

Reviewed By: ivmaykov

Differential Revision: D4798163

fbshipit-source-id: e7b65d030d89225dfdc2fad4c778d3653460806e

Summary: Self-moves are exceedingly rare and need not preserve the state of the object. They must only leave the object in a valid but unspecified state. By removing the branch, we make the common case (non-self move) faster.

Reviewed By: spacedentist, ot

Differential Revision: D4803486

fbshipit-source-id: 3ef2e1e13cd08d9221ecb154bfb3338b16487717

Properly constrain folly::Function's generic conversion constructor and fix its noexcept specification

Summary:
The generic conversion constructor for `folly::Function` was not checking that the source object could successfully be copy/move constructed, leading to some `is_constructible` false positives.

Also, the `noexcept` specification on the `Function` constructor wasn't taking into account that the source object might be copied into the Function, instead of moved. The copy could throw.

Reviewed By: yfeldblum

Differential Revision: D4775037

fbshipit-source-id: f337b41bf9ac431baa9457a501e63c18ca099e57

Summary: Useful for the same reason as the `get_ptr()` path function, but when you want to use a default value.

Reviewed By: luciang, yfeldblum

Differential Revision: D4785728

fbshipit-source-id: 70fd56f9ffa7a9edd6740f6dd712d3a251bf9fb0

Summary:
https://our.intern.facebook.com/intern/sevmanager/view/s/145809/?hc_location=ufi

```
hg backout d4e3fdfd57750c8dcabdb4c01ab3b528c4818624
```

Reviewed By: lbrandy

Differential Revision: D4799568

fbshipit-source-id: 76366040702e68becc7e245d7ce15832f933e10e

Summary: Extending ApplyTuple to support this common functional construct, with good forwarding semantics.

Reviewed By: yfeldblum

Differential Revision: D4787560

fbshipit-source-id: 2c740e448e0cb916abe948b79709d5ecd8ba54bb

Summary:
[Folly] Cut extraneous typename in `StringKeyedCommon.h`.

In this case, `typename` is optional, and, to avoid confusion, should be removed.

"In this case" means in a template definition, with a qualified non-dependent name.

Reviewed By: ericniebler

Differential Revision: D4792087

fbshipit-source-id: dbb9d2c747432a1f9bc8823e72a7668e78a2fbaa

Summary:
This diff adds two tests to futures/test/ViaTest.cpp:
viaDummyExecutorFutureSetValueFirst and
viaDummyExecutorFutureSetCallbackFirst. The latter resulted in a
dead-lock before the fix contained in this diff.

It is important that the callback function is destroyed after
it is executed, since it may capture objects (like a Promise)
that should be destroyed (so that e.g. a corresponding Future
throws BrokenPromise). When the callback is executed via an
executor, it is possible that the executor doesn't get around to
executing the task. We shouldn't rely on the task being executed
to do necessary clean-up. That clean-up should (also) happen
when the task (with its captured data) is destroyed (in the
spirit of RIAA).

Reviewed By: djwatson

Differential Revision: D4779215

fbshipit-source-id: d029cf8b8f7b55e1b03357749c5fb62d95986ca7

Summary:
Add codec that does automatic format detection for decompression.
The codec returned by `getCodec(CodecType::AUTOMATIC)` can decompress codecs `LZ4_FRAME`, `ZSTD`, `ZLIB`, `GZIP`, and `LZMA2`.
Additionally, the function `getAutomaticCodec()` allows users to pass their own custom codecs, but the 5 default codecs must also be supported.
We force the default codecs on users because having a common subset of formats is in general useful, and because we want to be able to add codecs to this set.

When compiled in debug mode, on construction it checks that the codecs are compatible.
This helps make sure that users don't accidentally add ambiguous codecs, and if we add a new default codec in the future, tests will fail on ambiguity.

Reviewed By: yfeldblum

Differential Revision: D4760065

fbshipit-source-id: 9e65844aba9f1ae3d5c7b86643931782c06ab3eb

Summary:
If the compiler doesn't see thru the call to tryRead which always initializes the val variable, it will emit the following error.
I was getting error when compiling my android project:
error: 'val' may be used uninitialized in this function [-Werror=maybe-uninitialized]
This should not have a negative perf impact as smart compilers which inline call to tryRead can elliminate the initialization.

Reviewed By: djwatson

Differential Revision: D4790864

fbshipit-source-id: f353cfe54ca4d056b6ddfc075d00580c9f2d391e

Summary:
`ZlibCodec::doUncompress()` always allocates at least 4 MiB when the uncompressed size is unkown.
We can get a 10-20% speed boost for small inputs (as well as significant memory savings) if we dynamically choose the buffer size.

Benchmarks before the change:
  ============================================================================
  experimental/terrelln/CompressionBenchmark.cpp  relative  time/iter  iters/s
  ============================================================================
  uncompress(ZLIB_100)                                         2.96us  338.19K
  uncompress(ZLIB_1000)                                        7.22us  138.44K
  uncompress(ZLIB_10000)                                      36.86us   27.13K
  uncompress(ZLIB_100000)                                    299.70us    3.34K
  uncompress(ZLIB_1000000)                                     2.74ms   365.32
  uncompress(ZLIB_10000000)                                   26.91ms    37.16
  uncompress(ZLIB_100000000)                                 273.92ms     3.65
  ============================================================================
After:
  ============================================================================
  experimental/terrelln/CompressionBenchmark.cpp  relative  time/iter  iters/s
  ============================================================================
  uncompress(ZLIB_100)                                         2.34us  426.68K
  uncompress(ZLIB_1000)                                        6.24us  160.38K
  uncompress(ZLIB_10000)                                      35.52us   28.16K
  uncompress(ZLIB_100000)                                    296.30us    3.37K
  uncompress(ZLIB_1000000)                                     2.74ms   365.25
  uncompress(ZLIB_10000000)                                   27.16ms    36.82
  uncompress(ZLIB_100000000)                                 273.72ms     3.65
  ============================================================================

Reviewed By: yfeldblum

Differential Revision: D4783871

fbshipit-source-id: 0fcce191ea520fd2117551db679c86f5ff734893

Summary: It's needed downstream in HHVM.

Reviewed By: yfeldblum

Differential Revision: D4785204

fbshipit-source-id: fe81ed8e23497fffeb44b2a366c457b788bcaada

Reviewed By: yfeldblum

Differential Revision: D4782707

fbshipit-source-id: 535b42b4f2558cadc78037268d6de81a8bb49840

Summary: In some situations it is useful to be able to sort the JSON keys according to some different scheme for readability, so allow the sorting function to be changed by the serializer_opts

Reviewed By: yfeldblum

Differential Revision: D4782077

fbshipit-source-id: 032fa60a38804452bd1c22c67ba897521cb2cd1d

Summary:
D4715918 broke open source builds on Ubuntu 14.04, since it has lz4-r114, but the lz4 frame API was introduced in [r123](https://github.com/lz4/lz4/blob/r123/lz4frame.h).
Put the `LZ4FrameCodec` behind a lz4 version check.
Fixes https://github.com/facebook/fbthrift/issues/209.

Reviewed By: yfeldblum

Differential Revision: D4780830

fbshipit-source-id: 19492a7e6bdd128e610c36b5778274e19eff9548

Summary: Passing a temporary default value to `get_ref_default()` returns a dangling reference when it is used.

Reviewed By: lbrandy, yfeldblum

Differential Revision: D4768769

fbshipit-source-id: 4c58a17dc7662ad553cf88a8544dae20016d2f6f

Summary:
Fix a correctness bug in "SWMRList.h".

Thanks to Maged for the tip! The old code omitted setting a removed node's
"next" pointer to `nullptr`, which meant that if the writer removed node
A and then node B = A->next while the reader was looking at B, then the reader
might happily keep chasing pointers B->next, B->next->next,... without ever
noticing that it was now on a "dead branch" of the linked list.

(And then there's a bit of a trick: you really do have to remove the node
first and *then* set its "next" pointer to null, because if you do the null
assignment first, you'll truncate the list, which means that some readers
will see a truncated list and return the wrong answer. I've added comments
to try to explain this to future-me.)

Style nit: Avoid doing multiple atomic operations on the same line of code.

----

Modernize the parameter-passing conventions in SWMRList.h.

Pass `T`s by const reference, except in `add` where we're likely to want to
make a copy of the `T` anyway. In that case it would be more "STL-correct"
to supply two different overloads `add(T&&)` and `add(const T&)`, but this
is just an example so it makes sense to keep things simple.

----

Fix an undefined behavior in SWMRList example.

Searching an empty SWMRList<int> always invokes undefined behavior by
comparing an uninitialized `T elem` against `val` on the final line of
the `contains` function. Besides, this patch allows SWMRList to work
with `T`s that aren't default-constructible or even copy-constructible.

----

Closes https://github.com/facebook/folly/pull/566

Reviewed By: djwatson

Differential Revision: D4772359

Pulled By: yfeldblum

fbshipit-source-id: 8f96573530800675cb56006aa91e7a5c5c1fb85d

Summary:
Two changes:
- Add a lock holder interface to enable users to move exclusive access and implicitly release the held lock.
- Change an assert to a static_assert to check at compile time if a provided function fits in the Function structure without dynamic allocation.

Reviewed By: djwatson

Differential Revision: D4770229

fbshipit-source-id: 89408164c08d7660231a6ca4e37637dd688356cd

Summary:
Add a `tryRead()`, and endian variants, which try to read into an arithmetic type, and if there isn't enough data they return false.
One use case is to quickly check if an IOBuf starts with a certain prefix, benchmarks show that using `tryReadLE()` is 6x faster than using `pullAtMost()` and `memcmp()`.

Reviewed By: yfeldblum

Differential Revision: D4767855

fbshipit-source-id: feb8c61092772933d4b8496b27d464559ff8b827

Summary: `get_ref_default()` won't accept temporary default values after D4768769.

Reviewed By: yfeldblum

Differential Revision: D4769043

fbshipit-source-id: 46b039e13b80721922f69036c6cc02f5642a26f9

Summary:
The `LZMA2` codec doesn't allow decompression without passing the uncompressed size.
Every other codec that supports streaming decompression without the uncompressed size by streaming into an IOBuf chain.
Benchmarks show that reducing the `defaultBufferLength` to 256 KiB doesn't slow down decompression of large files, but it speeds up decompression of small files (< 1000 bytes).

Reviewed By: yfeldblum

Differential Revision: D4751571

fbshipit-source-id: 39dbe6754a1ecdc2b7ba3107e9face926d4c98ca

Summary:
It is ~10% faster to call `ZSTD_decompress()` than use the
streaming API. The streaming API does some extra `memcpy`s that we
can avoid. We are working on improving the speed of the streaming
API in the case where all the data can be processed in one shot,
but that won't be available in the stable ZSTD api for a few versions.

Reviewed By: yfeldblum

Differential Revision: D4731058

fbshipit-source-id: 39026c499c0f5002466097b5afe7e30f850e0ae8

Summary:
The LZ4 Frame codec encodes data using the LZ4 frame format.
One advantage of the LZ4 frame format is that it has 4 magic bytes in the header, so users can transparently determine compression type.
It also allows the user to interop with the lz4 command line tool.

Reviewed By: yfeldblum

Differential Revision: D4715918

fbshipit-source-id: 689833fef526b1cfe98685179e7b494380d49cba

Summary:
Extract the environment-as-STL-map logic out of `test::EnvVarSaver` and into a
separate class `experimental::EnvironmentState` so that other code which needs
to manipulate the environment can do so more easily.

Add routines to set the process environment from the state of an
`EnvironmentState`'s map, and to extract the environment in the forms required
by `Subprocess` and the UNIX routines `execve` etc.

Reviewed By: yfeldblum

Differential Revision: D4713307

fbshipit-source-id: 6b1380dd29b9ba41c97b886814dd3eee91fc1c0f

Summary:
Looks like some tests for the `io` directory were left behind when they were moved out of experimental. Re-enabled them in the new location, except for the compression test because it takes a lot longer. It's still built and can be executed manually.

Also added some preprocessor guards around compression codecs that might not be compiled in (and therefore fail during test SetUp).

Depends on #547 because the path to gtest appears in the new Makefile.am's `CPPFLAGS`.
Closes https://github.com/facebook/folly/pull/550

Reviewed By: yfeldblum

Differential Revision: D4631244

Pulled By: Orvid

fbshipit-source-id: 496b2c272e4f7293822f159776a02f7dd0c9e04d

Reviewed By: yfeldblum, ot

Differential Revision: D4742262

fbshipit-source-id: 7137bb8e1cb39c74e8ba726e76e9a8a6f3e1984c

Randomize the Unix socket name in AsyncSocketTest.SendMessageAncillaryData test to avoid collisions.

Summary:
Our test framework reports frequent failures of AsyncSocketTest.SendMessageAncillaryData. According to the logs the socket fails to bind:

    folly/io/async/test/AsyncSocketTest2.cpp:3098: Failure
    Expected: (bind(lfd, (struct sockaddr*)&addr, sizeof(addr))) != (-1), actual: -1 vs -1
    Bind failed: 98

This diff adds the socket name randomization to avoid name collisions between tests running concurrently on the same test box.

Reviewed By: yfeldblum

Differential Revision: D4758942

fbshipit-source-id: 6066dbc18222a4521c40b2ff218cb7dab8bd789d

Summary:
Change every instance of EXPECT_EQ(false, ... to the simpler/shorter EXPECT_FALSE(...
Do the same for ASSERT_EQ, for consistency, and for true/TRUE.

Reviewed By: Orvid

Differential Revision: D4751397

fbshipit-source-id: 045293dbf8be1167f0cb1b85a2f9cd59a6a30f73

Summary:
Flat combining template that takes the following template parameters:
 T         Concurrent data structure using FC interface
 Mutex Mutex type (default std::mutex)
 Atom  Atomic template (default std::atomic)
 Req    Optional request structure to hold custom info (default dummy type bool)

Flat combining (FC) was introduced in the SPAA 2010 paper Flat Combining and the Synchronization-Parallelism Tradeoff, by Danny Hendler, Itai Incze, Nir Shavit, and Moran Tzafrir.
http://mcg.cs.tau.ac.il/projects/projects/flat-combining

Reviewed By: djwatson

Differential Revision: D4602402

fbshipit-source-id: 38327f752a3e92bb01e5496c321d8c87c818087a

Summary: The "l" in "Haskell" is not a singleton. Also, friends don't let friends rely on operator precedence.

Reviewed By: yfeldblum

Differential Revision: D4756653

fbshipit-source-id: 7b5360bfc1b26fac06062e4adbfae3a1e477c2dc

Summary: D4719112 changed the signature of the constructor but didn't update this callsite...

Reviewed By: djwatson

Differential Revision: D4754745

fbshipit-source-id: ae538a3ffb64cdc211f23220ea5e0d7bb7eff9b3

Summary: On Android, we should also use a relaxed singleton

Reviewed By: andriigrynenko

Differential Revision: D4754227

fbshipit-source-id: 157844e618d216b768bbb07f36c380ba2f4c9ab8