Summary:
- Support empty hazptr_holder, move constructor and assignment operator
- Limit thread caching to the default domain to improve performance of thread caching
- Fix unnecessary calls to stats singleton
- Use the mprotect version of AsymmetricMemoryBarrier for reducing the overhead of bulkReclaim().
- Update read-side benchmark results

Reviewed By: djwatson

Differential Revision: D5292885

fbshipit-source-id: bc5713ac95492a7114e1e467e71d2278e64b165d

Summary:
Remove the "LogLevel::" prefix from the logLevelToString() output.

This makes the output shorter and easier to read, and makes it simpler to write
LogFormatter implementations that want a short human-readable level
description.  Callers can always add a "LogLevel::" prefix on their own if
desired.

Reviewed By: wez

Differential Revision: D5288601

fbshipit-source-id: 046579ebdda76842d5a44c33d68941d73dd81162

Summary: The existing assertion errors do not give a lot of information to track down which thread is incorrectly using the `EventBase`. Print both the current thread name and the name of the thread running the event base loop.

Reviewed By: luciang

Differential Revision: D5289790

fbshipit-source-id: 563c7f68b7f9b7a6e85e22290d7e81afbf89871e

Summary: Make `IndexedMemPool` call `Traits::onRecycle` on the element just before it is marked as deallocated. This mirrors the allocation behavior implemented in D5177462 and simplifies preventing access to recycled elements (the client just needs to check `isAllocated` before accessing the element).

Reviewed By: nbronson

Differential Revision: D5275283

fbshipit-source-id: 58365b5b7b32b07fa56529c476078f241fc20811

Summary: Tests and benchmarks don't belong in the same directory as source code.

Reviewed By: yfeldblum

Differential Revision: D5288080

fbshipit-source-id: 865c6dfbeff751bda49870e5284e168d33aff256

Summary:
Original sys_membarrier call supported an EXPEDITED flag, that was much faster than the current version, at some extra CPU cost.

The basic idea was a normal one would be used in a background thread, while EXPEDITED would be useful inline in foreground threads.

Add the flag here, and use the mprotect hack until sys_membarrier supports both.

Reviewed By: magedm

Differential Revision: D5292680

fbshipit-source-id: 30edf27514976991bbaa8e776a7138eb815ade36

Summary: as title, similar to TTLS

Reviewed By: djwatson

Differential Revision: D5284200

fbshipit-source-id: 7eb95668740b239349c6e73f3b152e6506671072

Summary: Absence of "uint32" type in gFlagAdders map breaks all apps that use gflags of this type (i.e. define a flag with "DEFINE_uint32").

Reviewed By: WillerZ

Differential Revision: D5286720

fbshipit-source-id: c02bd959cb9ea9a47fba1e01429181ba09edf5dd

Summary: Complete the `ThreadName.h` API.

Reviewed By: luciang, Orvid

Differential Revision: D5289160

fbshipit-source-id: a48e61093008039da50b1c568364fa5b8744b401

Summary:
This refactors the logging code with the aim of reducing the amount of assembly
code emitted for each log statement, particularly for `XLOG()` statements.
Ideally it should be possible to put debug `XLOG()` statements throughout your
code without having to worry about the performance overhead.  Therefore we
should attempt to make sure that `XLOG()` statements do not generate a lot of
assembly and hurt icache performance.

This diff does not have any code behavior changes--it just moves code around a
bit.  The high-level summary is:

- Move as much code as possible into the LogStreamProcessor constructors and
  destructor.  Make sure these methods are defined in LogStreamProcessor.cpp to
  avoid having them be emitted inline at each log statement.
- Move some of the XLOG level checking logic into separate non-inline functions
  defined in xlog.cpp
- Pass xlog category information around as a pair of (categoryName,
  isOverridden) parameters.  If isOverridden is true then the categoryName
  parameter should be used as the category name directly.  If isOverridden is
  false, then categoryName is a filename that needs to go through filename to
  category name string processing.  This allows the category name processing to
  be done in non-inlined code.

Reviewed By: wez

Differential Revision: D5269976

fbshipit-source-id: 7a7877ddfed66cd27ed82f052330b6aa2be4b37b

Summary:
Add new `FB_LOGC()` and `XLOGC()` macros that accept C-style printf format
strings.  (The `FB_LOGF()` and `XLOGF()` macro names are already used for
`folly::format()` style formatting.)

This will make it easier for users to update existing printf-style code to use
this new logging library.

These are in a separate `printf.h` header file that must be explicitly included
to have access to these macros.  The intent is to encourage users to use one of
the other APIs (streaming, append-style, or `folly::format()`) instead of these
printf-like APIs in new code.

Reviewed By: omry

Differential Revision: D5269974

fbshipit-source-id: 56e55f9642bb00806d9b4c762fb6a91778ef6ad3

Summary:
The `XLOG_GET_CATEGORY()` macro was previously written assuming it was only
used inside `XLOG()` statement.  When used inside the main translation unit
being compiled (e.g., a .cpp file and not a header file), the code assumed that
the file-scope category had already been initialized, since a level check had
presumably already been performed.

However, it is useful in some places for external users (outside of the logging
library itself) to call `XLOG_GET_CATEGORY()`.  In these cases a log level
check may not have been performed yet, so the file-scope category may not be
initialized yet.

This diff renames the existing `XLOG_GET_CATEGORY()` macro to
`XLOG_GET_CATEGORY_INTERNAL()` and adds a new `XLOG_GET_CATEGORY()` macro that
is slower (it always looks up the category by name) but always safe to use.

This also adds a new `XLOG_GET_CATEGORY_NAME()` macro, and renames the existing
`XLOG_SET_CATEGORY()` macro to `XLOG_SET_CATEGORY_NAME()` for API consistency.

Reviewed By: wez

Differential Revision: D5269975

fbshipit-source-id: 373805255823855282fa7a8d4a4d232ba06367f6

Summary:
This improves the test that exercises the AsyncFileWriter message discarding
logic.

Previously each writer thread wrote a fixed number of small messages.  This was
fairly slow, and wasn't always guaranteed to reliably produce discards.
Now each writer thread writes larger messages, which produce discards faster.
The test also automatically stops after 10 separate discard events, so that it
finishes faster (typically a few hundred milliseconds, rather than 5+ seconds).

This also updates the test to use XLOG() internally rather than using fprintf()
to print to stderr, now that the XLOG() diffs have landed.

Reviewed By: wez

Differential Revision: D5261059

fbshipit-source-id: 120224706fee36948ef76efbeb579ccc56400c51

Summary:
Add new `FATAL` and `DFATAL` log levels.

Any log message with a level of `FATAL` always crashes the program.  Log
messages with a level of `DFATAL` crash the program in debug build modes.

Before crashing, the code makes sure to flush all LogHandlers, so that the
fatal message is not lost if some of the LogHandlers process messages
asynchronously.  If no LogHandlers were configured, the message is printed to
stderr.

Reviewed By: wez

Differential Revision: D5189497

fbshipit-source-id: c45dbd582fb1c3a962d00effb2967737ef97cc8b

Summary:
Updated to the interface to be in synch with the latest standard proposal in P0233R4 as follows:
- Renamed hazptr_owner as hazptr_holder.
- Combined hazptr_holder member functions set() and clear() as reset().
- Replaced the template parameter A for hazptr_holder member function templates try_protect() and get_protected with atomic<T*>.
- Moved the template parameter T from the class hazptr_holder to its member functions try_protect(), get_protected(), and reset().
- Added a non-template overload of hazptr_holder::reset() with an optional nullptr_t parameter.
- Removed the template parameter T from the free function swap() as hazptr_holder is no longer a template.

Reviewed By: davidtgoldblatt

Differential Revision: D5283863

fbshipit-source-id: 2bc1a09f4f844aa72d9b7dff9c450540bbe09972

Summary:
When
- only one function is scheduled in FunctionScheduler; and
- the function is running while cancelFunctionAndWait() is being called;
FunctionScheduler.cancelFunctionAndWait() will hang forever.  The root cause is that the condition in cancelFunctionAndWait() is incorrect:

```
runningCondvar_.wait(l, [currentFunction, this]() {
  return currentFunction != currentFunction_;
});
```

because currentFunction will not be changed if only one function is in the scheduler.

The fix here is to
- clear currentFunction as nullptr.  This also makes the internal behaviors of cancelFunction() and cancelFunctionAndWait() consistent.
- introduces additional variable to indicate the state of cancelling current function.  After running the function, the background thread will detect cancellation of current function and clear the variable.
- cancelFunctionAndWait() condition variable will wait for the variable to be cleared.

Similarly, cancelAllFunctionsAndWait() also suffers from the same issue.

Unit tests are added to reproduce the issue.

Reviewed By: yfeldblum

Differential Revision: D5271664

fbshipit-source-id: acb223304d3eab23129907ce9ff5e57e55f1e909

Summary:
Add a method to flush all LogHandler objects.

This will be necessary to implement FB_LOG(FATAL), as we will want to flush all
outstanding messages before crashing.

Reviewed By: wez

Differential Revision: D5189501

fbshipit-source-id: faf260b8e71e5dfed4a3b1c1aee32f072bd7b764

Summary:
Add a flush() call to the LogHandler interface.  This is needed to implement
`FB_LOG(FATAL)` so that we can flush all LogHandlers before aborting the
program.

Reviewed By: wez

Differential Revision: D5189499

fbshipit-source-id: 75fa4d7e75ea26de5b7383bf7e8d073fb37e9309

Summary:
Add a flag to the LogWriter so we can ensure that particular messages are never
discarded, even when the LogWriter is throttling messages.

This functionality will be necessary to implement `FB_LOG(FATAL)` to that we
can avoid discarding the reason for crashing.

Reviewed By: wez

Differential Revision: D5189498

fbshipit-source-id: dc4322ea5ba449a341cdbdc32afb0ed466019801

Summary:
This adds a small example program which demonstrates using the logging library.

This gives a very basic example of how the library is intended to be used, and
can also be used to play around with controlling the log levels from the
command line argument.

Reviewed By: wez

Differential Revision: D5083104

fbshipit-source-id: ab09c6c88db33065f6e39f35b28014f2a6153cef

Summary: It is not used anywhere, not even in tests, so kill it.

Reviewed By: yfeldblum

Differential Revision: D5280800

fbshipit-source-id: 7e6a308bf09198548b77dcc1bfacc0ee95eb4887

Summary: This reverts commit 07639e155421f31a6cc7ed16cba2034750e44325

Differential Revision: D5278412

fbshipit-source-id: 4b4d60f1f472ebe2e68e8eea1a31f42bc55c60ed

Summary: The only reason these were split was because of the use of `folly::format`

Reviewed By: yfeldblum

Differential Revision: D5278412

fbshipit-source-id: 07639e155421f31a6cc7ed16cba2034750e44325

Summary: Because an `int` plays much much nicer with code already written for Windows than `void*` does. It takes quite a bit of hackery to make it possible with the pthread implementation we support using, but it is possible and is worth the effort.

Reviewed By: yfeldblum

Differential Revision: D5276968

fbshipit-source-id: 4cd0d3120a4f19976e76853ef4b337e96d5005a8

Summary: On Windows it has include order requirements.

Reviewed By: yfeldblum

Differential Revision: D5277151

fbshipit-source-id: a6cf923d78e5944ac34e885e32a7cf072f904720

Reviewed By: yfeldblum

Differential Revision: D5228760

fbshipit-source-id: bcad4237997e611e06c93b476c5b673dc93c0621

Summary: It was broken by D5139886 which added a call to `std::__throw_bad_alloc` without first including the correct portability header.

Reviewed By: yfeldblum

Differential Revision: D5273016

fbshipit-source-id: cac1772f41ae5942a29cefa69fb07d9cf94b5aab

Summary: I broke it when I moved the stats tests around.

Reviewed By: yfeldblum

Differential Revision: D5258526

fbshipit-source-id: 70982ffaaed89a548e943e4b011b013c89808666

Summary: It's not safe because of how it manipulates the linked list.  The intended usage is to declare it locally on the stack

Reviewed By: yfeldblum

Differential Revision: D5267507

fbshipit-source-id: 07a55eaee2c6a266bb233432f6a3383236594195

Summary: They were incorrectly comparing against string literals rather than the actual demangled names. MSVC includes `class/struct` as part of the mangled name, so they also appear in the demangled name, in contrast to GCC/Clang, which don't.

Reviewed By: yfeldblum

Differential Revision: D5271087

fbshipit-source-id: 41fa0cebe098b1b935e1f8b1af5882c412cf6254

Summary:
Improve documentation and use the correct pragma for warning
on the unsupported path.

Reviewed By: ot

Differential Revision: D5264582

fbshipit-source-id: 710ee46fef6d8f37f665f4bb6f7e4c5dc0b27436

Summary: This reverts commit cd94c711f0390d249b6736ab292721a477022b02

Differential Revision: D4567233

fbshipit-source-id: a10aff6b4c231666aec9f40ab65e881c3cd870dc

Summary:
get_fiber_manager_map_vevb() and get_fiber_manager_map_evb() don't
provide very useful feedback when called with an empty map (just the former) or
with a program that doesn't have debug symbols (both)

Reviewed By: andriigrynenko

Differential Revision: D5260114

fbshipit-source-id: c44b4e279e5c88dc08507b969339a4befc23d79b

Summary: Previous version failed to compile with "-Werror=shadow" flag, this commit fixes the problem.

Reviewed By: WillerZ

Differential Revision: D5255814

fbshipit-source-id: c1252474ed5415b47759022bcbabc78c1639e10a

Summary: MSVC, GCC, and Clang define the relative comparison operators by default on `enum class` values, however under MSVC defining them explicitly causes the compiler to fail to resolve which operator to use, so just remove them entirely, as they aren't needed.

Reviewed By: simpkins

Differential Revision: D5259835

fbshipit-source-id: 1b91db5de0fc47320daabfdd7132ae910887ff0f

Summary: There are a couple of places where the caught exception was being given a name that was never used, generating a warning under MSVC.

Reviewed By: simpkins

Differential Revision: D5260131

fbshipit-source-id: f82c6bd1266f6a4c916594ec3ac94edc9a2e48fe

Summary: Glog doesn't try to, so there's no real reason to try and be smart. Just do the same thing as every other platform for now until we have reason to do otherwise.

Reviewed By: simpkins

Differential Revision: D5259435

fbshipit-source-id: dcdf55f51f2e13945afd351cb45e9c1a22f56218

Summary: As MSVC correctly warns via C4266, the overload sets of functions in base classes are not part of the overload set in derived classes where you've declared an overload (or override) unless you've explicitly imported it.

Reviewed By: simpkins

Differential Revision: D5260056

fbshipit-source-id: bbbeeea3c13201a3a6eba6e62cfa5a49a9470d43

Summary: It is needed on Windows for the `localtime_r` function.

Reviewed By: simpkins

Differential Revision: D5259944

fbshipit-source-id: e380ecf42252ecabec168cf87668ace3ff4c0a9f

Summary:
The standard does not define that `std::ostream` should have a default constructor, so this was invalid and was an error on Windows.
This explicitly initializes the base class by passing a `nullptr`.

Reviewed By: simpkins

Differential Revision: D5259690

fbshipit-source-id: b8914d73ff2682e1a4447b8338860259778b2247