Summary: There have been several ASAN bugs cropping up because the lifetime of an `HHWheelTimer` is not being manager properly. I think that people are too comfortable passing around a raw `HHWheelTimer*` even when it is difficult to prove correctness. The recommended solution used to be to create a `DestructorGuard` every time it was needed. There is enough friction there that not everyone is doing that like they should. We should make resource management easier---as easy as using raw pointers, in fact.

I've fixed the broken copy semantics of `DestructorGuard` and added the operators that allow it to be used as a smart pointer. I added the `IntrusivePtr` helper that can manage an arbitrary derived class of `DelayedDestructionBase`. Now, clients can all safely pass around an `IntrusivePtr` everywhere they used to use a raw pointer. They will get automatic resource management for free.

If you are not convinced that `DestructorGuard` should be changed, then note that the existing behavior is dangerously buggy. Consider the following innocent code that accidentally uses the implicitly-defined copy constructor:
  auto d = DestructorGuard(p);
This results in undefined behavior, because `p` can be accessed after it is destroyed! The bug happens because the default copy constructor copies the raw pointer but doesn't increment the count.

In a separate diff, I'll change all clients who pass around a raw `HHWheelTimer*` to use an `IntrusivePtr<HHWheelTimer>` instead. Then we can even entertain a long-term plan of switching from intrusive pointers to the standard `shared_ptr`.

Reviewed By: djwatson

Differential Revision: D2627941

fb-gh-sync-id: 58a934d64540d0bbab334adc4f23d31d507692da

Summary: My colleague tuomaspelkonen discovered a weird UNSYNCHRONIZED issue a few weeks ago and we ever since stopped using it. Now I finally have some time to root cause it. It turns out UNSYNCHRONIZED unlock the mutex then lock the mutex again, because it copy constructs LockedPtr for overriding the name within the scope, and copy construct locks the mutex again. A one character fix here is to take a reference of LockedPtr instead of copy construct it. However since this is my first time look at the code here, please advise if this is horribly wrong or propose better fix. Also added a test to reproduce the issue without the fix as well as verify the fix.

Reviewed By: yfeldblum

Differential Revision: D2633028

fb-gh-sync-id: a9e8d39b08d4d1265979f8bdaae83619566d10a0

Summary: hg backout -r c9f7b5f3185a
Revert my change (which broke down the cont build t9048692)

Reviewed By: djwatson

Differential Revision: D2640797

fb-gh-sync-id: 51f196ac5a3560fde4dc8fe7bb6ef278d74136e5

Summary: Wait uses baton & callback running baton.post
when the original future is ready. However wrapping baton.post
with a funciton call (preparing a new value) adds the following race: baton.wait wakes up
before that function has call actually finished.
The explanation is the following: to prepare the value of the new future it's necessary
1. baton.post() 2. set the value (move constructor, memory operations, ...)
and this code is executed in a separate thread.
The main idea of this fix is to avoid creating a new future
(whose value is calculated using that 2-step procedure)
and set a callback instead. This callback will be executed when the future is ready and actually
it either will be the last step of promise.setValue or it will run immediately if the future
we are waiting for already contains a value.

Reviewed By: fugalh

Differential Revision: D2636409

fb-gh-sync-id: df3e9bbcc56a5fac5834ffecc63f1bcb94ace02c

Summary: Revert D2632752

Reviewed By: afrind

Differential Revision: D2638974

fb-gh-sync-id: 43d8135421510db7840a99f3c197119a0fd26c09

Summary: Using Clang 3.7, this minimal test program:

  #include <folly/FBString.h>
  struct S { folly::basic_fbstring<char> x; };
  int main(int argc, char *argv[]) {
    S s {};
    return 0;
  }

... breaks with the following error output:

  FBStringTest.cpp:5:8: error: chosen constructor is explicit in copy-initialization
    S s {};
         ^
  ./folly/FBString.h:1009:12: note: constructor declared here
    explicit basic_fbstring(const A& = A()) noexcept {
             ^
  FBStringTest.cpp:3:40: note: in implicit initialization of field 'x' with omitted initializer
  struct S { folly::basic_fbstring<char> x; };

... because this `basic_fbstring` is used in a struct and the struct is being default-constructed.

This patch splits the (nevertheless-correct) one-default-arg constructor into two, to deconfuse clang and have it select the correct constructor to avoid this issue.

Reviewed By: matbd

Differential Revision: D2632953

fb-gh-sync-id: 2c75ae85732678c31543f5cccc73d58317982f07

Summary: The check was originally added to track down cpu pinning issue. In reality the check did get triggered, so change to soft error to calm the crazy nag bot. what's wrong with you.
We need to see the soft error log to find out what the error is.

Reviewed By: mzlee

Differential Revision: D2632752

fb-gh-sync-id: 87c11b186f97f0eb4a6c5ac13a1117b280198673

Summary: Without this tag type, it's impossible to use `Synchronized` with types like:

```
struct A {
  A(int, const char*);
  A(const A&) = delete;
  A& operator=(const A&) = delete;
  A(A&&) = delete;
  A& operator=(A&&) = delete;
};
```

In-place construction solves this problem. Usage:

```
Synchronized a(construct_in_place, 5, "c");
```

Reviewed By: nbronson

Differential Revision: D2610071

fb-gh-sync-id: 251fe8f8f6a2d7484dda64cf04dcacb998145230

Summary: This is the solution for our `HHWheelTimer` crashes, as suggested in D2617966#25.

djwatson also mentioned that maybe there should be some logging in the destructor if there are outstanding callbacks. I couldn't think of anything that would add to the `assert` that already exists in `destroy`. I'm open to suggestions though.

Reviewed By: djwatson

Differential Revision: D2628154

fb-gh-sync-id: f3db6e9384517c9bf3cbb60af8c1e711703a07fa

Summary: Define `__CLANG_PREREQ` macro to check version of clang.

Reviewed By: yfeldblum

Differential Revision: D2630325

fb-gh-sync-id: 3d666e554e8ddfc2c1fecd439aaf93f015829025

Summary: Callbacks sometimes outlive the `HHWheelTimer` that they reference. Then the `Callback` tries to reference the dead `HHWheelTimer` and it could either misbehave or crash. This was caught reliably by ASAN tests.

Since `HHWheelTimer` already supports intrusive ref counting, the solution is to acquire a reference within the `Callback`.

Reviewed By: djwatson

Differential Revision: D2617966

fb-gh-sync-id: 02be9ffc5851c269d5933288a17ad394b33ac2dd

Summary: Do not cancel TimeoutHandler timeout if the timeout has already run.

Reviewed By: spalamarchuk

Differential Revision: D2622280

fb-gh-sync-id: 27d83b44ab225e2859695f4e5f21cef934824a35

Summary: This name's meaning is very obscure, this method actually sets the
queue size of each acceptor and if all acceptor's queues are full the
connection will be dropped. Change it to a more meaningful name.

Reviewed By: alandau

Differential Revision: D2613681

fb-gh-sync-id: baa374cdf0a87c460df3dd5687e3d755b55f4b4f

Summary: Start server timeout after socket write succeeds in mcrouter.  Add neceessary Fibers logic to enable this behavior.

Reviewed By: pavlo-fb

Differential Revision: D2613344

fb-gh-sync-id: 1bc0fbe8b325a3e91cd010f89104b83ebf183679

Summary: If SharedPromise::getFuture() is invoked after a call to setInterruptHandler,
then the interrupt handler isn't set on the newly created promise. This diff
fixes that.

Reviewed By: yfeldblum

Differential Revision: D2610289

fb-gh-sync-id: bf8fce9e881b83ccac17d13c6788ec2afd0b0153

Summary: Allows AsyncTransportWrapper's to supply the underlying application
protocol being used, for example using NPN for SSL or some other
generic protocol indication mechanism.

Reviewed By: ranjeeth

Differential Revision: D2614179

fb-gh-sync-id: 2079782bb7d44f898fb14c7df15077209b022424

Summary: Our use case need to have the IndexType Configurable, but in this diff D2583752, lint error shows
  Code from folly::detail is logically private, please avoid use outside of folly.
In order to fix this lint error, I switch the IndexType and Allocator in the AtomicUnorderedInsertMap template.

Reviewed By: nbronson

Differential Revision: D2610921

fb-gh-sync-id: ae81b41e7c8c971f26c61b8c67dabeadff379584

Summary: Allow underlying transport to be accessible from AsyncTransportWrapper.
There are some code paths where we need access to the real transport from
the AsyncTransportWrapper. This allows us to retrieve the underlying transport
and have clients like HTTPSession use it.

Reviewed By: afrind

Differential Revision: D2609200

fb-gh-sync-id: 2b317d1825a005bb64468f83c64bc3f1c9bdfe2c

Summary: .. by including what you use.

These are a handful of headers that are depended on via viral inclusion. Include them directly.

Reviewed By: yfeldblum

Differential Revision: D2601172

fb-gh-sync-id: 215e87263325d085fbcb651f83f429f47d14fc1b

Summary: Prefer `std::move` to `move` for clarity's sake.

Reviewed By: yfeldblum

Differential Revision: D2601138

fb-gh-sync-id: 64a9e58a83da9c8416a0be5c45a292c6b1342664

Summary: This disallows implicitly constructing a StringPiece from a literal nullptr at compile time
(without this change, nullptr would cause a segfault in strlen()).

Reviewed By: meyering, andriigrynenko

Differential Revision: D2603597

fb-gh-sync-id: cafbc45945bacc72a7c89310b99aa62d19a3ff9f

Summary: Fix warnings triggered by `-Wcomment`.

Reviewed By: ajtulloch

Differential Revision: D2603992

fb-gh-sync-id: aae721f7c236d7d8b8bff2c077a481fc2affcf71

Summary: New API - write an entry and advance a cursor to the just-written slot.

Reviewed By: ritzau, yfeldblum

Differential Revision: D2599388

fb-gh-sync-id: f3ebfae97de0341f01ffc80ab10221313d02087c

Summary: Reclaim unnecessary fibers from the free pool periodically.  Turn this behavior
on in mcrouter/proxy.cpp

Reviewed By: pavlo-fb

Differential Revision: D2584825

fb-gh-sync-id: eabc58eefe6fd38972c9e23ca3cbe79eb8316a3e

Summary: Convert helper classes to using statements now that GCC has caught up
with c++11.

Reviewed By: simpkins

Differential Revision: D2561364

fb-gh-sync-id: 712591549aba9450d159468dc3b26a987ffe9b82

Summary: try_get() API is confusing right now, because it returns nullptr in release, but crashes in debug build. Users end up handling nullptr returns, yet their code crashes in debug builds for no reason. See https://www.facebook.com/groups/fbthrift/permalink/1222120054481561/ for an example.

If we want to keep the crashing API we should probably name it differently and make it crash both in debug and release.

Reviewed By: dhruvbird

Differential Revision: D2587818

fb-gh-sync-id: 5834bfa08eb5d9bc6db1c5edf4a048a5b1d3212c

Summary: In my use case, 1.5 billion keys with loadFactor 0.8, the linear probing performs really bad.

Reviewed By: nbronson

Differential Revision: D2579243

fb-gh-sync-id: 5081356de55f770823a4afad55bf7e2114b4e313

Summary: startWork_ is used only if enableTimeMeasurement is set. It gives ~0.5% performance win for mcrouter.

Reviewed By: yfeldblum

Differential Revision: D2590176

fb-gh-sync-id: 07f2189ebdec751cd0d91d191d8f595780d2808a

Summary: SYNCHRONIZED warns with -Wshadow due to `for (auto&
FB_ARG_1(__VA_ARGS__) =`. This diff just suppresses that warning.

Reviewed By: djwatson

Differential Revision: D2587348

fb-gh-sync-id: 3a2e39fb6ce28da014950ca94e4b62ea80deb65f

Summary: The implicit conversion loses integer precision: 'size_type' (aka 'unsigned
long') to 'int'.

Reviewed By: mzlee

Differential Revision: D2585883

fb-gh-sync-id: 1fc7c84b66c8f19cc36b798dd198730764e19b28

Summary: Further race is presented in HHWheelTimer and EventBase, since changing them to track callback life cycle is more involving, I am hacking around it to prove the concept. With the *fix*, no reproduce of the segmentation fault anymore.

This is for proving the concept, code will be cleaned up if we agree this is reasonable direction to pursue.

Reviewed By: fugalh

Differential Revision: D2577715

fb-gh-sync-id: e0bb0317268e6f02a54fc70454e63959fba70c10

Summary: Expose pending messages in accept queue in AsyncServerSocket.

Set default accept message queue size to 1024.

Reviewed By: djwatson

Differential Revision: D2525161

fb-gh-sync-id: a69ea0ee77729e4a8300bde3e3c07840f2d5d3cb

Reviewed By: yangliu

Differential Revision: D2578043

fb-gh-sync-id: cec8f21219655a495b99d0b6b99f0925615bc068

Summary: This is pretty much the reverse operation from the fromBinary() constructor.

Reviewed By: yfeldblum

Differential Revision: D2578680

fb-gh-sync-id: d8c4e53fe8bc0f5373ebb0b4f7ee498659c1b003

Summary: [Folly] Style nits for `folly/test/FingerprintBenchmark.cpp`.

Reviewed By: Gownta

Differential Revision: D2579463

fb-gh-sync-id: 135bc3ef43708cb024f78c59672811159d2163e0

Summary: Renaming `EnqueuedForAccept` and `DequeuedByAccept` to
`EnqueuedForAcceptor` and `DequeuedByAcceptor` respectively to make it clear
that the connection is queued for the Acceptor not to call `accept` on.

Reviewed By: yfeldblum

Differential Revision: D2554578

fb-gh-sync-id: 072a9ff122658827d1e89f2bef79ad565dad7974

Summary: Tested over weekend with D2571554. There was no EXC_RESOURCE/CPU crash when read errrors were returned on ENOTCONN. EAGAIN seems innocent as the spin detector was disabled on testing group.

Patch folly. I can add #ifdef __APPLE__ around the errno checking code but this should be good practice for non apple code as well.

I will remove debugging code in fbios master in another diff.

Reviewed By: djwatson

Differential Revision: D2580819

fb-gh-sync-id: 9162a3deba01af8b07cd2b336d7da3da040c67a9

Summary: AtomicUnorderedInsertMap's constructor takes a maxLoadFactor argument,
which is given its default argument of 0.8f in all of our code.
The clipping function that was supposed to prevent load factors
greater than one was inverted, resulting in all of our code using a
maxLoadFactor of 1 instead of 0.8.  This diff fixes the computation,
as well as changing all of the use sites so that the actual memory
allocated by existing clients does not change.

Reviewed By: yfeldblum

Differential Revision: D2575238

fb-gh-sync-id: bb9dd8de53114236b259631d175d62af098391cf

Summary: Add a couple benchmarks measuring allocation/deallocation latency.  One benchmark measures an allocate-deallocate repeated pattern, the other measures the allocation of a large chunk of fibers.

We'll use these benchmarks later on after we modify the allocation/deallocation behavior of fibers.

Reviewed By: pavlo-fb

Differential Revision: D2573064

fb-gh-sync-id: 414eb93d6223e42c187c03214a47dfb533491bab

Summary: Some libraries like Haskell's `Data.Aeson` can produce arbitrarily
big numbers with arbitrary precision.
The json standard doesn't specify the ranges for numeric types.
For interoperability, we should allow the user to parse the numbers with
some loss of precision.

Reviewed By: luciang

Differential Revision: D2565140

fb-gh-sync-id: b1a9a46e298bf13cc89d7e79ce28705e9e251a7f