Summary: All singleton registration has to happen in .cpp. Otherwise singleton may be double registered if inlined.

Reviewed By: yfeldblum

Differential Revision: D2956951

fb-gh-sync-id: bea425c31270d614f4b8a780204694168602fe86
shipit-source-id: bea425c31270d614f4b8a780204694168602fe86

Summary:It's possible to have nested loopUntilNoReady (with different FiberManagers). See unit test for a simple example.
This diff fixes currentFiberManager_ to be a stack.

Reviewed By: meyering

Differential Revision: D2953468

fb-gh-sync-id: 0abdcb7f43c94e7bb0adef8440699dc8e138d21a
shipit-source-id: 0abdcb7f43c94e7bb0adef8440699dc8e138d21a

Summary: EventBase loop should do it as well, but it's possible for LoopCallback to be removed before EventBaseOnDestructionCallback.

Reviewed By: pavlo-fb

Differential Revision: D2951303

fb-gh-sync-id: 43f5a91e7ecffa7ab7feec32fa45e01b28db66a3
shipit-source-id: 43f5a91e7ecffa7ab7feec32fa45e01b28db66a3

Summary:This implements StaticSingletonManager which is then used to create all leaked Meyers singletons.

StaticSingletonManager is a singleton itself, which is created in a separate compilation unit (Singleton.cpp) and so we can be sure that other compilation units will always see a single instance of StaticSingletonManager, even if linked dynamically.

StaticSingletonManager then keeps a dictionary of typeid -> object pointer, which is used to de-duplicate same singleton being re-created from different compilation units (linked dynamically), usually because of code inlining.

override-unit-failures

Reviewed By: yfeldblum

Differential Revision: D2913027

fb-gh-sync-id: 1f5015a79a7a8297ebf5f0fe3fd0cc7eb44f706b
shipit-source-id: 1f5015a79a7a8297ebf5f0fe3fd0cc7eb44f706b

Summary: `mask(numBits)` returns a temporary variable. After its lifetime ends, subbytes becomes a dangling pointer.

Reviewed By: yfeldblum

Differential Revision: D2953699

fb-gh-sync-id: 5d5a35716ecaa1b3d96edb5a459615756848b92f
shipit-source-id: 5d5a35716ecaa1b3d96edb5a459615756848b92f

Summary:[Folly] `Indestructible`.

For when you need a Meyers singleton that will never be destructed, even at program exit.

Good for large data structures with many heap allocations, as long as they have no behavior. No point in ever destructing them.

HT: Proxygen, and Orvid.

Reviewed By: andriigrynenko

Differential Revision: D2947959

fb-gh-sync-id: 7dd1f725edf137ba81fbf4032a0819fd7c627261
shipit-source-id: 7dd1f725edf137ba81fbf4032a0819fd7c627261

folly: symbolizer: slow address->{file+line-number} lookup if `.debug_aranges` is missing (e.g. --strip-debug-non-line)

Summary:Binaries linked with `gold` and `--strip-debug-non-line` don't have an `.debug_aranges` section

We still want to map `address->{file+line-number}` to get nice stack
traces even though this might be slower (linear search all compilation unit entries in `.debug_info`).

Before:
```
$ # link with gold + --strip-debug-non-line
$ folly/experimental/exception_tracer/exception_tracer_test
E0217 15:02:13.694947 1321814 ExceptionTracer.cpp:179] Exception type: std::runtime_error (9 frames)
    @ 000000000040ad2d __cxa_throw
    @ 0000000000409df3 bar()
    @ 0000000000409eab baz()
    @ 0000000000407c77 main
    @ 00007f00dd9860f5 __libc_start_main
    @ 000000000040991b (unknown)
```

After (similar to the output without `--strip-debug-non-line`):
```
E0217 18:37:37.579596 1583124 ExceptionTracer.cpp:179] Exception type: std::runtime_error (9 frames)
    @ 000000000040ad6d __cxa_throw
                       ./folly/experimental/exception_tracer/ExceptionTracerLib.cpp:57
    @ 0000000000409e33 bar()
                       ./folly/experimental/exception_tracer/ExceptionTracerTest.cpp:24
    @ 0000000000409eeb baz()
                       ./folly/experimental/exception_tracer/ExceptionTracerTest.cpp:51
    @ 0000000000407c87 main
                       ./folly/experimental/exception_tracer/ExceptionTracerTest.cpp:96
    @ 00007f1d16ff80f5 __libc_start_main
    @ 000000000040995b (unknown)
```

Differential Revision: D2947965

fb-gh-sync-id: e517bab324b1dcb70cadc9a5211ce794e35c83a5
shipit-source-id: e517bab324b1dcb70cadc9a5211ce794e35c83a5

Summary: MSVC is actually smarter about this than GCC, as MSVC doesn't support VLAs, it tries to eval the length at compile time. GCC on the other hand doesn't try to eval it at compile time, resulting in compiles via CMake telling us that this is a VLA.

Reviewed By: yfeldblum

Differential Revision: D2911929

fb-gh-sync-id: ffaa133bcf4129a3e02f7e875966d3ae6a97be6a
shipit-source-id: ffaa133bcf4129a3e02f7e875966d3ae6a97be6a

Summary:The existing implementation of folly::applyTuple does not support
mutable callables (such as mutable lambdas and other functor objects
that only implement non-const operator()).

This commit adds a few more unit tests and changes the implementation
so that new and existing tests pass.

Reviewed By: yfeldblum

Differential Revision: D2942622

fb-gh-sync-id: 82478f290e9fd2020358ff79ef0a6bcf8a43738c
shipit-source-id: 82478f290e9fd2020358ff79ef0a6bcf8a43738c

Summary:[Folly] Sort the `Makefile` headers a bit.

Just manually moving a few obviously out-of-order entries.

Reviewed By: Orvid

Differential Revision: D2945668

fb-gh-sync-id: 604976d25e5913ef01d591e23737d61dacfbfc86
shipit-source-id: 604976d25e5913ef01d591e23737d61dacfbfc86

Reviewed By: djwatson

Differential Revision: D2945770

fb-gh-sync-id: 9e4ee3b052b0bbb33ef796b8070edd24c6942589
shipit-source-id: 9e4ee3b052b0bbb33ef796b8070edd24c6942589

Summary: std::call_once is a nice API but the current GCC implementation is slower than it needs to be.  Adding a header-only wrapper with an additional atomic bool.

Reviewed By: luciang

Differential Revision: D2938884

fb-gh-sync-id: 5939c94fe62a1523053dcff26c880ecaec9e1150
shipit-source-id: 5939c94fe62a1523053dcff26c880ecaec9e1150

Summary:--Perhaps this should move to Portability.h instead.  There is another instance of this in io/async/AsyncSocket.cpp.--

Added `portability/SysUio.h` to handle the c-macro conversion into a usable value.

Reviewed By: yfeldblum

Differential Revision: D2940778

fb-gh-sync-id: 897e44b430c02e5a7d826f3e8da9e4979b7b898c
shipit-source-id: 897e44b430c02e5a7d826f3e8da9e4979b7b898c

Summary: Minor changes to AsyncSSLSocket for async crypto

Reviewed By: siyengar

Differential Revision: D2516765

fb-gh-sync-id: 354baeb94e6f63e8d5cdf8455ff5ca49a6aa479c
shipit-source-id: 354baeb94e6f63e8d5cdf8455ff5ca49a6aa479c

Summary: `folly::wrapvFull` fail if the passed in `iovec*` has more than 1024 (`IOV_MAX`) elements. In particular, it returns -1 with errno 22 [Invalid argument]. The fix is to limit maximum size of iovec* to IOV_MAX that is passed in to `readv/writev/...` in a single iteration of outer loop.

Reviewed By: yfeldblum

Differential Revision: D2935540

fb-gh-sync-id: 6c0a073ac0b59db3d53fb4269b13ddfcc479efb1
shipit-source-id: 6c0a073ac0b59db3d53fb4269b13ddfcc479efb1

Summary:There doesn't appear to be a way to use std::enable_if on a
destructor, so conditionally select from one of two storage types.

Reviewed By: yfeldblum

Differential Revision: D2923902

fb-gh-sync-id: 2def8d1031d70379fd84e8eb555dad9d2b4996f2
shipit-source-id: 2def8d1031d70379fd84e8eb555dad9d2b4996f2

Summary: If getVia was called on a future modified using via, getVia could deadlock if the original future was updated to a new executor and there was no callback chained after the call to via. In effect: f.via(executor).getVia(executor); deadlocks. This can be a problem if the code is hidden in a library and the precise semantics are unclear. This diff adds a test that exposes the problem and a fix by forcing waitVia to add a callback that will satisfy the new exector, ensuring that drive() has a callback to trigger once the future is satisfied.

Reviewed By: andriigrynenko

Differential Revision: D2906858

fb-gh-sync-id: a3105079530f15d7a7d39a9381c4078665b721a7
shipit-source-id: a3105079530f15d7a7d39a9381c4078665b721a7

Summary: This also fixes a bug where exception was not re-thrown for functions returning void.

Reviewed By: spalamarchuk

Differential Revision: D2936887

fb-gh-sync-id: 9828dec131203528c27eae874aba147168f40d0d
shipit-source-id: 9828dec131203528c27eae874aba147168f40d0d

Summary:Adding padding for Android was wrong, instead, we should have
used `std::aligned_storage`.

Reviewed By: fugalh

Differential Revision: D2934566

fb-gh-sync-id: f541b89309be70791fced48f63b6b5aecc49bfbb
shipit-source-id: f541b89309be70791fced48f63b6b5aecc49bfbb

Summary:Json.PrintTo test if failing with heap-buffer-overflow asan abort. The problem here comes up
when values of `std::string` and `std::ostrstring::str()` are compared.  `std::ostrstring::str()` returns
*non null* terminated array of `char`s. When compared with `std::string` a read from memory after
the end that array is made. Fixing the test by replacing `std::ostrstream` with `std::ostringstream`,
that does append `\0`.

Reviewed By: yfeldblum

Differential Revision: D2934352

fb-gh-sync-id: c6a5d765c9951716b8a14715702cf3d940c6d723
shipit-source-id: c6a5d765c9951716b8a14715702cf3d940c6d723

Summary:This depends on D2897095.

The main goal here is to start moving onRequestSuccess and onRequestFailure calls higher in the stack.

Reviewed By: mmcduff

Differential Revision: D2899959

fb-gh-sync-id: 4e074c2d734f88f5be56000095b892c6d47c0dcc
shipit-source-id: 4e074c2d734f88f5be56000095b892c6d47c0dcc

Summary: Update copyright for 2016

Reviewed By: igorsugak

Differential Revision: D2828047

fb-gh-sync-id: 8638188aeb694c3ca2a615cc2bc3bc9d251e388d
shipit-source-id: 8638188aeb694c3ca2a615cc2bc3bc9d251e388d

Summary: This adds a folly::Singleton->folly::ThreadLocal dependency to make sure folly::ThreadLocal is always loaded first. Otherwise PthreadKeyUnregister singleton is created after folly::Singleton storage, even though it has higher priority.

Reviewed By: andrewjcg

Differential Revision: D2931170

fb-gh-sync-id: 36392d76e98201f2b4416b4bbef451d6c0e8c69d
shipit-source-id: 36392d76e98201f2b4416b4bbef451d6c0e8c69d

Summary:I hit a landmine where rate() returned 0 for my tests where fewer than 60 items were added per minute.  This was because it was truncating it.  And yet, countRate() was working.
It doesn't make sense for a rate (value accumulated per time period) to be integral.
Folly: I changed rate() to return a double by default for folly, as was done by avg() and countRate().  Looked like a typo to me.
Common wrapper: I changed getRate() to allow you to override and make it double, as was done by getAvg().  Defaulting to int (which is usually the value type) is a bad call IMO but it's a riskier change to change it to double, and I want to be consistent with getAvg().

Reviewed By: tracelog

Differential Revision: D2921061

fb-gh-sync-id: 00875f2ab7963ef3ba2db475aedaf6ebd413b38f
shipit-source-id: 00875f2ab7963ef3ba2db475aedaf6ebd413b38f

Summary: Didn't check this with -Wshadow when I first wrote it.

Reviewed By: ranjeeth

Differential Revision: D2926047

fb-gh-sync-id: 8391986dee2d6b5698491bd2995f039468ec684d
shipit-source-id: 8391986dee2d6b5698491bd2995f039468ec684d

Summary: It should be UL instead of ULL.

Reviewed By: siyengar

Differential Revision: D2926005

fb-gh-sync-id: d2387eb601c50d9820907c831277aa6f8e938cbb
shipit-source-id: d2387eb601c50d9820907c831277aa6f8e938cbb

Summary: Remove boost regex from `stripLeftMargin`. We can shrink some binaries by not including it in the core folly library.

Reviewed By: yfeldblum

Differential Revision: D2922415

fb-gh-sync-id: cee89164c650706f0e5c07eed3d40500831918cd
shipit-source-id: cee89164c650706f0e5c07eed3d40500831918cd

Summary: We want onThreadExit to be canceled as late as possible. Otherwise accessAllThreads() may crash on shutdown (because ThreadEntry is destroyed, but not removed from the global list).

Reviewed By: igorsugak

Differential Revision: D2923435

fb-gh-sync-id: 3a8eef05844bb547c06db33cb96734df1169852a
shipit-source-id: 3a8eef05844bb547c06db33cb96734df1169852a

Summary:
When compiled for -m32, this test fails because of the
difference between longs.

Reviewed By: yfeldblum

Differential Revision: D2924626

fb-gh-sync-id: 8ee863fa2b6df8519bd52d0d3144697a297780bb
shipit-source-id: 8ee863fa2b6df8519bd52d0d3144697a297780bb

Summary: Avoid allocation and logging inside the throw handler, and some other minor tweaks.

Reviewed By: bort, luciang

Differential Revision: D2921073

fb-gh-sync-id: 4491ab2f85a4251e59ba9394bba5abef0bdf7a10
shipit-source-id: 4491ab2f85a4251e59ba9394bba5abef0bdf7a10

Summary:
Previously we could be reading from thread-local FiberManagerMap while it was modified.
This is now fixed by keeping a per-thread list of EventBases which need to be removed from local maps. On the fast-path no action is taken, since list will be empty.

This is second try, since D2853921 got reverted.
The new implementation is simpler and does not rely on AtomicLinkedList.

Reviewed By: yfeldblum

Differential Revision: D2908018

fb-gh-sync-id: 4d7aed974c19761f7e2732ddbf8694af57c69bd6
shipit-source-id: 4d7aed974c19761f7e2732ddbf8694af57c69bd6

Summary:
[Folly] `stripLeftMargin`.

So you can do:

    TEST(MyClass, doSomethingWithString) {
      //  Multiline string literal is indented properly inside the test case
      //  instead of being aligned all the way to the left, which would make
      //  the test case read poorly.
      auto input = folly::stripLeftMargin(R"TEXT(
        first line
        second line
        third line
      )TEXT");
      auto expected = //...
      auto actual = MyClass::doSomethingWithString(input);
      EXPECT_EQ(expected, actual);
    }

Reviewed By: markisaa

Differential Revision: D2909736

fb-gh-sync-id: ebb07da05e1a788535064cfcd9e07f617a007800
shipit-source-id: ebb07da05e1a788535064cfcd9e07f617a007800

Summary:
HHWheelTimer's auto-pointers are kind of funny.  You can do something like this:

```
HHWheelTimer::UniquePtr p = ...;
// create a SharedPtr from UniquePtr
HHWheelTimer::SharedPtr s(p);
// create another SharedPtr from raw ptr
HHWheelTimer::SharedPtr s(p.get());
// No problem.

If you do this:

HHWheelTimer::SharedPtr p = ....;
// this leaks
```

Why?  Because SharedPtr is only have of std::shared_ptr.  It's the refcounting half.  But when the last SharedPtr goes out of scope, it **does not** invoke HHWheelTimer::destroy().

So code like this is possible/expected:

```
MySweetObj::MySweetObj(HHWheelTimer::SharedPtr s) {
  s_ = s;
  s_.scheduleTimeout(a, b);
}

{
  HHWheelTimer::UniquePtr p = ...;

  obj = MySweetObj(p)

  // But what if I decide to kill p:
  p.reset();
}
```

Since MySweetObj takes a SharedPtr and holds it, it can reasonbly expect that it can schedule timeouts on it, and the HHWheelTimer will not be deleted until it releases the SharedPtr.  This is true, but the above code would hit the assert that count_ == 0.

Instead, relase the check that count_ == 0 only if there are no extra outstanding SharedPtrs.

Reviewed By: viswanathgs, chadparry

Differential Revision: D2908729

fb-gh-sync-id: 9abd4a7d692fe952c5514dbb8d85dfbad95a3cac
shipit-source-id: 9abd4a7d692fe952c5514dbb8d85dfbad95a3cac

Summary:
FBStringTest.cpp was refactored, but the TestUtil and
TemporaryFile was not properly removed.

Reviewed By: yfeldblum

Differential Revision: D2913045

fb-gh-sync-id: 710c3f5b808acb634dfcd65219484ddc257ed52c

Summary:
[Folly] Fix Build: `conditionally_existent_test` in `folly/test/Makefile.am`.

HT: https://github.com/shindo.

Reviewed By: meyering

Differential Revision: D2910782

fb-gh-sync-id: 9593f026e7c58b8644abc5996da65005bcc2d095

Summary:
folly::Future (more precisely folly::detail::Core) can store
callback functors (typically lambdas) up to a certain size
(8*sizeof(void*)) inside the main object. Only bigger functors
are stored in the std::function object in folly::detail::Core,
which will put them on the heap.

The behaviour of folly::Future is slightly different depending
on whether the callback can be embedded in the main object
or not. Small functors will be destructed after the callback
is executed. Functors too big to fit in the lambda space in
the Core object will be deleted when Core is deleted.

Some code assumes that functor objects are deleted as soon
as the callback has been executed. The implementations of
folly::Future::collect and variants do so. If you switch
off this optimisation temporarily (which can be done easily
by setting lambdaBufSize in folly/futures/detail/Core.h to
0), a number of unit tests fail.

Given that the lambda buffer is reasonably large, most
functors can be stored in the Core object. The different
behaviour for very large lambdas may not have been observed.

This diff fixes the inconsitent behaviour.

Firstly, it changes the unit test Future:finish to explicitly
check that the callback functor is deleted after the callback
has been executed. This should be tested, because this
behaviour is assumed in other parts of the futures
implementation (e.g. Future::collect et al.).

Secondly, it adds another unit test Future:finishBigLambda,
similar to Future:finish. The lambda captures much more data
to make sure the lambda won't be stored in the Core object,
but in the std::function object inside Core. The test verifies
that the behaviour is the same, i.e. the callback functor
is destructed after the callback has been executed.

Thirdly, it fixes Core and makes sure that functors of any
size are destructued after the callback has been called. The
new unit test fails without this.

Reviewed By: fugalh

Differential Revision: D2883772

fb-gh-sync-id: 21a410f6592b3e090772a7b55bef6729d8739922

Reviewed By: philippv

Differential Revision: D2900534

fb-gh-sync-id: c711a160d541d937ada24f2b524016dbceb40ee2

Summary: Due to different way of padding, CoreTest failed on android. Loosen the check to make it pass.

Reviewed By: mzlee

Differential Revision: D2907987

fb-gh-sync-id: 81d715725d6908ec7b259b0a67789a91ed63df71

Summary: Some small changes to port folly/futures onto android

Reviewed By: mzlee

Differential Revision: D2787564

fb-gh-sync-id: 8c161942c24c2b7b0484339eaa51c5a356f17de5

Summary:
[Folly] `ConditionallyExistent<typename>`.

For when we need extra member fields in dbg builds, but want to save the space in opt builds. Or other situations along similar lines, but with perhaps another the statically-known condition.

Conditional compilation can have some nasty side-effects. Best to avoid it. Let the compiler see everything, rather than having the preprocessor ruin our day. Let the optimizer remove code that will never be called, after the compiler has seen it. Let us have a single AST in our source file, not one AST for dbg and one for opt, or other statically-known conditions.

Reviewed By: andriigrynenko

Differential Revision: D2879397

fb-gh-sync-id: d631141a984eebd46674f27a40a97f670eb33f54