Summary:
Idle memory in IO threads.   If loop is unused for a period of time, free associated memory, and call epoll again.

Had to add a new list of callbacks that don't make the loop nonblocking (i.e. using runInLoop() instead would use the nonblocking version of epoll).

Could bake this in to EventBase directly, but that seems like the wrong abstraction, since EventBase doesn't actually control the thread - for example, that approach would also free up memory for stack-allocated EventBases where they are used synchronously by clients.

This diff doesn't change IO scheduling at all - current IO work is round robin, so this probably only helps if the whole server is idle (at least until we add smarter scheduling)

Test Plan:
Based on top of D1585087.

fbconfig thrift/perf/cpp; fbmake dbg
_bin/thrift/perf/cpp/ThriftServer
_bin/thrift/perf/cpp/loadgen -num_threads=100 -weight_sendrecv=1 -cpp2 -async

Ran loadgen for a while, watched res memory in top.  Stopped loadgen.  After ~5 sec, res memory was much reduced.

Reviewed By: jsedgwick@fb.com

Subscribers: trunkagent, doug, fugalh, njormrod, folly-diffs@

FB internal diff: D1641057

Tasks: 5002425

Summary:

Test Plan:

Reviewed By: cgheorghe@fb.com

Subscribers:

FB internal diff: D1642334

Blame Revision: D1587625

Summary:
Add an operator()(StringPiece) method to Appender, so it can be used
directly with folly::Formatter objects.  Also add printf() and vprintf()
methods to Appender, for places that need to use existing printf-style
formatting.

This also includes versions of push() and pushAtMost() that accept
ByteRange objects.  Previously we only had push() implementations that
took separate buffer and size arguments.

Test Plan: Added new unit tests to IOBufCursorTest.cpp

Reviewed By: davejwatson@fb.com

Subscribers: trunkagent, doug, net-systems@, exa, njormrod

FB internal diff: D1583684

Summary: Jenkins ci build was broken for a while, so we missed these  in contbuild results

Test Plan: watch ci build for fbthrift

Reviewed By: hans@fb.com

Subscribers: doug, fugalh, njormrod, folly-diffs@

FB internal diff: D1641515

Summary:
Move async socket to folly.

Changes:
* Made an AsyncSocketException type instead of TTransportException: Some of the exceptions didn't fit nicely in to std::exception types (like TIMED_OUT).  There are some wrappers in thrift/lib/cpp/async to convert back to TTransportException, so all existing code still compiles.
* Moved read/write callbacks out of AsyncTransport: filters are going to want to do the read/write stuff separately (see revproxy/tunnel/filters, and discussions in D1483148).

Test Plan:
fbconfig -r thrift; fbmake runtests
contbuild should catch everything else - exception types shouldn't change for existing code

Reviewed By: dcsommer@fb.com

Subscribers: mshneer, folly-diffs@, trunkagent, doug, alandau, bmatheny, njormrod, fugalh, jsedgwick

FB internal diff: D1587625

Summary:
In D1618240 I introduced a race condition in `detachOne()`, where `detached_` is incremented and then tested. If the promise and future are racing, they can both see `detached_ == 2` in the conditional, and then they'll both try to free the Core object. This fixes that.

It also fixes a related problem (which actually showed up more often with the test I wrote), where we transition into the Done state before setting the value, and then `maybeCallback()` observes the state is Done (because we're just reading an atomic, not grabbing the lock, which is intentional), tries to execute the callback, but `folly::Optional` throws an exception because the value hasn't been set yet (at least in debug it does). I should have listened to my gut and kept the state assignment after the transition action in the first place.

Test Plan: New unit test

Reviewed By: jsedgwick@fb.com

Subscribers: trunkagent, net-systems@, fugalh, exa, njormrod, folly-diffs@, mnd

FB internal diff: D1636490

Tasks: 5438209

Blame Revision: D1618240

Summary: Reverts D1633874. Companion to D1636490 which fixes the bug.

Test Plan:
git reverting code that was git reverted and hasn't changed in the interim
Won't be checked in without the companion bugfix diff (D1636490)

Reviewed By: davejwatson@fb.com

Subscribers: trunkagent, net-systems@, fugalh, exa, njormrod, folly-diffs@

FB internal diff: D1636487

Tasks: 5438209

Blame Revision: D1633874

Summary: EventBaseManager was moved to folly, we shouldn't have deps on apache

Test Plan: fbconfig -r folly/experimental/wangle/concurrency; fbmake runtests

Reviewed By: hans@fb.com

Subscribers: doug, fugalh, njormrod, folly-diffs@

FB internal diff: D1636006

Summary:

Test Plan:

Reviewed By: harishs@fb.com

Subscribers: net-systems@, fugalh, exa, njormrod, folly-diffs@

FB internal diff: D1633874

Tasks: 5438209

Summary: I'm still a little mystified how one makes a Future<const T> (I tried to make one explicitly in a test and failed), but this is clearly the bug @reedriley sees in https://phabricator.fb.com/D1620805#21

Test Plan:
fbconfig titan/cachius/test
fbmake

Reviewed By: reedriley@fb.com

Subscribers: trunkagent, net-systems@, fugalh, exa, njormrod, reedriley, folly-diffs@

FB internal diff: D1630378

Blame Revision: D1620805

Summary:
Changes:
* namespace to folly
* some std::chrono replacesments
* Moves shutdownSocketSet stuff to thriftserver instead of in the socket itself.
* Changes exception types.  I don't see anywhere that uses the TAsync*SERVER*socket exceptions depending on the TTransport type, but I could be wrong? I don't really know what to do about the exception types

unittests still postponed overnight.

Test Plan:
fbconfig -r thrift; fbmake runtests
fbconfig -r proxygen; fbmake runtests

Reviewed By: dcsommer@fb.com

Subscribers: hphp-diffs@, ps, folly-diffs@, anca, trunkagent, jsedgwick, fugalh, doug, alandau, bmatheny, njormrod

FB internal diff: D1579187

Summary:
fix fbthrift build, broken in D1621717

I wonder if it'd be possible to lint against this somehow. The jenkins failures are easy to miss especially on folly diffs with lots of failure noise (in this case it was still postponed)

Test Plan: wait for jenkins

Reviewed By: davejwatson@fb.com

Subscribers: njormrod, folly-diffs@

FB internal diff: D1630833

Summary: This makes it more likely that SingletonVaultDestructor won't be created after atexit calls.

Test Plan: unit test

Reviewed By: chip@fb.com

Subscribers: hphp-diffs@, ps, njormrod, folly-diffs@

FB internal diff: D1629804

Tasks: 5353022

Summary:
This would cause debug builds to do a bad thing (access the variable `this->detached_` within an assert, after `delete this`).

Test Plan: unit tests

Hopefully now that we have a dummy cpp file in `folly/wangle/detail` contbuild
will pick it up and all the dependencies will also run their tests.
Right now, we suspect maybe maelstrom (@wez) and adinvoicer and zookeeper
(@jying) and probably others are seeing this in unit test failures (esp. if
they use asan, which is rightly detecting read after free). Hoping contbuild
will catch them.

Reviewed By: davejwatson@fb.com

Subscribers: net-systems@, fugalh, exa, njormrod, folly-diffs@, wez, dcapel, jying, cgheorghe

FB internal diff: D1630301

Tasks: 5424546, 5435720

Blame Revision: D1618240

Summary: Fixing the race will take some more thought. Probably a condvar. For now, I'm turning waitFor into a spinlock instead.

Test Plan: new unit test that deadlocks before the workaround

Reviewed By: jsedgwick@fb.com

Subscribers: trunkagent, net-systems@, fugalh, exa, njormrod, folly-diffs@

FB internal diff: D1628015

Tasks: 5427828

Summary: If you try to include files in a weird (but not incorrect) order you get compilation errors. No longer!

Test Plan: new dummy `cpp_binary` target

Reviewed By: davejwatson@fb.com

Subscribers: trunkagent, folly-diffs@, net-systems@, fugalh, exa, njormrod

FB internal diff: D1621083

Summary:
Modeled very closely after Finagle's interrupts. Compare with https://github.com/twitter/util/blob/master/util-core/src/main/scala/com/twitter/util/Promise.scala if you like.
The basic idea is the promise holder can register an interrupt handler, and then interrupts will call that handler. A typical handler would fulfil the promise with an exception (or special value) indicating that it was interrupted (if it was interrupted in time).
Raising an interrupt does not prevent setting a value or callbacks executing or any of that - it is only advisory to the promise holder.

Test Plan: I wrote some unit tests.

Reviewed By: davejwatson@fb.com

Subscribers: folly-diffs@, net-systems@, fugalh, exa, hannesr, njormrod

FB internal diff: D1620805

Tasks: 4618297

Summary:
This is a refactor of the current functionality to use a state machine (inheriting from `FSM`). This will make it easier to extend to do cancellation and Future collapsing. Performance is the same, maybe slightly faster (abt 1%). (I might be a little conservative with the atomics, it might be worth going through and reasoning about whether they all need to be atomic).

The state machine is two states, Waiting (no value), and Done (has a value). Transitioning to Done will execute the callback if it exists and we are active. Otherwise the callback will happen when it is set and active is true.

There is a subjective balancing act in place here, between making a state for every single mutable bit of information (which results in an explosion of states: hasValue X hasCallback X isActive X isCancelled …), and finding a sweet spot of expressivity. This isn't too far from the way Twitter did it, though we don't (yet) have all the states they have (and they don't have the concept of hot/cold futures). But I got there by way of replacing the `mutex_` with the state, after changing all those variables to atomics so they didn't need mutex protection (resulting in only `callback_` and `result_` needing it). I expect the state machine will morph as the rest of the functionality is added, but hopefully it will be easier to understand and keep correct (that's the idea anyway).

Test Plan: Tests still pass (and not by accident, I made several awesome mistakes along the way).

Reviewed By: davejwatson@fb.com

Subscribers: net-systems@, fugalh, exa, njormrod

FB internal diff: D1618240

Tasks: 4618297

Summary:
Just some cleaning hose in `detail::Core`.
- rename `fulfil` to `setResult`
- and `value_` to `result_`
- remove unnecessary helper methods (I can be convinced to keep them if you like them)

Test Plan: mechanical changes, tests still pass

Reviewed By: davejwatson@fb.com

Subscribers: trunkagent, folly-diffs@, net-systems@, fugalh, exa, njormrod

FB internal diff: D1618161

Summary:
Later is now superfluous. Just hold a deactivated future (e.g. one returned by `Future::via`), and then activate (or destruct) it when you're ready.
ThreadGate has proven to be too rigid and too heavyweight to be really useful. And is superfluous now with `Future::via`. e.g.
f.via(east).then(a1).then(a2).via(west).then(b1).then(b2);

This is stage one of removing these: mark them deprecated.
Stage two will be for me to change all existing usages in our code to not use them.
Stage three will be to rip these out altogether.

Test Plan:
Everything still builds and works, but now with deprecation warnings.

fbconfig -r \
common/smc/cpp/tests \
folly/wangle \
langtech/audiens \
messaging/avscan \
messaging/maelstrom \
neteng/traffic_manager \
notifications/nudge/service \
tao/client \
zeus/datashuttle
fbmake runtests

Reviewed By: jsedgwick@fb.com

Subscribers: net-systems@, fugalh, exa, njormrod, folly-diffs@

FB internal diff: D1626412

Tasks: 5409538

Summary: It was deprecated, now is unused.

Test Plan:
fbgs
check contbuild

Reviewed By: hannesr@fb.com

Subscribers: trunkagent, net-systems@, fugalh, exa, njormrod, folly-diffs@

FB internal diff: D1624221

Summary: for great contbuild

Test Plan: these projects still build. contbuild should start picking them up, based on discussion with @robbert

Reviewed By: robbert@fb.com

Subscribers: trunkagent, net-systems@, fugalh, exa, mwa, jgehring, fuegen, njormrod, folly-diffs@, robbert

FB internal diff: D1624206

Tasks: 5399744

Summary:
Like the magic macros. As it says in the comment, this can lead to nicer code.
Also added/moved a couple examples to the top of `test/FSM.cpp`.

Test Plan: new tests

Reviewed By: davejwatson@fb.com

Subscribers: trunkagent, net-systems@, fugalh, exa, njormrod

FB internal diff: D1618184

Summary: as above

Test Plan:

Reviewed By: davejwatson@fb.com

Subscribers: fugalh, njormrod

FB internal diff: D1589439

Summary:
This is the last of the changes to make folly and thrift portable. The old
thrift SpinLock implementation is exactly we
want here so port this into folly. If someone later makes efficeint
versions of MicroSpinLock for platforms other than x64 they can easily
be plugged in without changing the calling code.

Test Plan: compiled on multiple platforms

Reviewed By: alandau@fb.com

Subscribers: folly-diffs@, ldbrandy, trunkagent, alandau, bmatheny, njormrod, mshneer, shikong, benyluo, kmdent, fma, ranjeeth, subodh

FB internal diff: D1621717

Tasks: 5183325

Summary: mcrouter now uses folly::Singleton.

Test Plan: visual

Reviewed By: andrii@fb.com

Subscribers: njormrod

FB internal diff: D1622647

Test Plan: yes

Reviewed By: delong.j@fb.com

Summary:
pthread_atfork isn't defined in the Android NDK (and therefore is not actually supported even though it's in bionic/libc) so don't call it there.

Test Plan: compiled on Linux, Android

Reviewed By: meyering@fb.com

Subscribers: njormrod, shikong, kmdent, fma

FB internal diff: D1620584

Tasks: 5183325

Summary: Buffer size calculation had fallen out of sync with the actual test code. Good thing we have ASAN!

Test Plan: Run the test with address santizer.

Reviewed By: mpawlowski@fb.com

Subscribers: mpawlowski, trunkagent, njormrod

FB internal diff: D1619472

Tasks: 5325399

Summary:
I can see how {<,>}{,=} comparisons might be ambiguous, but equality
should never be, so let's allow it.

Test Plan: tests

Reviewed By: tjackson@fb.com

Subscribers: trunkagent, njormrod

FB internal diff: D1618088

Summary:
Move shutdownsocketset to folly, since it is a dep of the asyncsockets

Previoulsy tried moving it in to the server directly: D1583629, but had issues - close(fd) is called before the error callback, so we can't remove the fd before the close, which is essential to it working properly.

Just move it to folly instead.

Test Plan: fbconfig -r thrift/lib/cpp thrift/lib/cpp2; fbmake runtests

Reviewed By: dcsommer@fb.com

Subscribers: mshneer, trunkagent, fugalh, jsedgwick, doug, alandau, bmatheny, njormrod

FB internal diff: D1594950

Summary:
Conditional changes required to compile on iOS and Android.

Test Plan: compiled on all platforms

Reviewed By: meyering@fb.com

Subscribers: kmccray, trunkagent, shilin, njormrod, ranjeeth, subodh, bmatheny

FB internal diff: D1618028

Tasks: 5183325

Summary:
This is a fix for situations where you know that something is keeping singleton alive (by getting a weak_ptr and locking it), and request a singleton instance via get() method (if e.g. it's hard to pass a pointer to singleton instance from a method which locked it). If shutdown was scheduled - an exception was previously thrown even though the object was not destroyed yet.

This simplifies conversion of legacy code to folly::Singleton.

Test Plan:
fbconfig -r gatehouse/usersets/tests
fbmake runtests

Reviewed By: chip@fb.com

Subscribers: trunkagent, njormrod

FB internal diff: D1619311

Summary:
Remove global initialization of ThreadLocal in Random
See https://gcc.gnu.org/onlinedocs/gcc-4.8.3/gcc/Thread-Local.html

Static initializers and thread locals don't mix well, and we're seeing
some crashes which we think might be related to this.

Test Plan: Unit tests

Reviewed By: seanc@fb.com

Subscribers: trunkagent, seanc, njormrod

FB internal diff: D1617455

Summary: universal references and perfect forwarding to the rescue

Test Plan: added demonstrative unit tests

Reviewed By: hans@fb.com

Subscribers: trunkagent, fugalh, njormrod, bmatheny

FB internal diff: D1592032

Tasks: 5283342

Summary:
the one in concurrent/ is a bit more generic, so I kept that as Executor and renamed the existing one ScheduledExecutor

because Hans is surfing I took the liberty of renaming Action->Func as an alias for std::function<void()>, because I think it's more reflective
also kept the version of add() that doesn't force rvalue-reference as it's more user friendly and probably not less performant in common cases (insert reference to "want speed? pass by value" here)

Test Plan: compiled some major relevant bits, will let contbuild show me anything I missed

Reviewed By: hans@fb.com

Subscribers: trunkagent, rushix, fbcode-common-diffs@, fugalh, msk, njormrod

FB internal diff: D1591237

Tasks: 5279196

Summary:
Right now destroyInstances is scheduled when SingletonVault is requested. This may change singleton destruction order (folly::Singleton-managed vs unmanaged singletons) when new singleton is registered with folly::Singleton (no matter if it is even used).
This diff changes it to be more stable.

Test Plan: servicerouter unittests

Reviewed By: chip@fb.com

Subscribers: njormrod, mshneer

FB internal diff: D1615213

Tasks: 5353022

Summary:
While working on other code I noticed folly stopped compiling
with clang. Fix that.

Test Plan: fbmake runtests (clang and gcc)

Reviewed By: lucian@fb.com

Subscribers: mathieubaudet, njormrod, bmatheny

FB internal diff: D1618189

Summary: For great speed

Test Plan:
Before
============================================================================
folly/wangle/test/Benchmark.cpp                 relative  time/iter  iters/s
============================================================================
constantFuture                                             241.19ns    4.15M
promiseAndFuture                                 100.32%   240.42ns    4.16M
withThen                                          44.63%   540.47ns    1.85M
----------------------------------------------------------------------------
oneThen                                                    519.20ns    1.93M
twoThens                                          62.83%   826.41ns    1.21M
fourThens                                         36.80%     1.41us  708.75K
hundredThens                                       1.79%    29.05us   34.42K
----------------------------------------------------------------------------
no_contention                                                4.82ms   207.27
contention                                        62.91%     7.67ms   130.39
============================================================================

After
============================================================================
folly/wangle/test/Benchmark.cpp                 relative  time/iter  iters/s
============================================================================
constantFuture                                             159.79ns    6.26M
promiseAndFuture                                 101.23%   157.84ns    6.34M
withThen                                          41.78%   382.47ns    2.61M
----------------------------------------------------------------------------
oneThen                                                    358.23ns    2.79M
twoThens                                          63.07%   568.00ns    1.76M
fourThens                                         36.89%   971.07ns    1.03M
hundredThens                                       1.76%    20.34us   49.17K
----------------------------------------------------------------------------
no_contention                                                3.75ms   266.75
contention                                        59.83%     6.27ms   159.59
============================================================================

That's a 150% speedup.

Reviewed By: davejwatson@fb.com

Subscribers: net-systems@, fugalh, exa, njormrod

FB internal diff: D1617363

Tasks: 5278220