Summary:
* folly/test/PackedSyncPtrTest.cpp: Avoid an operator new [] vs
operator delete mismatch.

Test Plan:
fbconfig --platform-all=gcc-4.8.1-glibc-2.17 --sanitize=address \
folly/test:packed_sync_ptr_test && fbmake runtests

Reviewed By: tudorb@fb.com

FB internal diff: D1111227

Summary:
- Add autoconf check for cplus_demangle_v3_callback() in libiberty and
avoid is usage when not available. Clang/libc++ on Mac OS X doesn't
support this.

Test Plan:
- fbconfig -r folly && fbmake runtests
- ./configure && make check on Mac OS X

Reviewed By: tudorb@fb.com

FB internal diff: D1108543

Summary: This reverts commit 2c4f8da739ba0294f2c947753a0e30fee291feb5.

Test Plan: Compiles

Reviewed By: tudorb@fb.com

FB internal diff: D1108840

Test Plan: Tested with subsequent thrift diff

Reviewed By: davejwatson@fb.com

FB internal diff: D1108428

Summary:
- Clang's libc++ doesn't provide these header files. Detect libc++ via
the _LIBCPP_VERSION symbol (pulling it in by sourcing some header files
earlier if necessary) and avoid using these files.
- This is another attempt at D1074481.

Test Plan: .

Reviewed By: andrei.alexandrescu@fb.com

FB internal diff: D1074723

Summary:
- This fixes a bug introduced in D1002959 that broke -fb platform
compilation: there was noone to forward-declare these symbols.

Test Plan:
- Built the gcc-4.8.1-glibc-2.17-fb platform
- fbconfig -r folly && fbmake runtests
- configure/make check on Mac OS X

Reviewed By: andrei.alexandrescu@fb.com

FB internal diff: D1074720

Blame Revision: D1002959

Test Plan: fbconfig -r folly/experimental/symbolizer folly/experimental/exception_tracer && fbmake runtests_opt

Reviewed By: wez@fb.com

FB internal diff: D1104426

Summary:
Yes, I know about boost::starts_with, but I think the convenience is worth it.
Also, I've written then equivalent of removePrefix / removeSuffix way too
many times.

Test Plan: test added

Reviewed By: andrei.alexandrescu@fb.com

FB internal diff: D1106425

Summary:
Even though it should be according to the docs; tdep_trace allocates memory on
first use from each thread.

Wrote a slow loop that we can use from the signal handler. The exception tracer
still uses the fast version.

Test Plan: fbconfig -r folly/experimental/symbolizer folly/experimental/exception_tracer && fbmake runtests_opt

Reviewed By: philipp@fb.com

FB internal diff: D1101095

Summary:
Add futexTimedWait to Futex which allows callers to wait on the futex
for a specified max duration.

Test Plan: -- Ran all unitests

Reviewed By: ngbronson@fb.com

FB internal diff: D1090115

Summary:
AtomicStruct<T> acts like std::atomic<T>, but it supports any
trivial or trivially-copyable type up to 8 bytes in size.  I'm not sure
why these types weren't included in the original std::atomic standard,
since this construct removes a lot of boilerplate for some uses of
std::atomic.

Test Plan:
1. unit tests
2. this code is adapted from production use in tao/queues/LifoSem

Reviewed By: davejwatson@fb.com

FB internal diff: D1085106

Summary:
folly::json::serialize by default doesn't check for valid UTF8, and as a result can generate invalid JSON.  There is an option to check for valid UTF8, which throws on an error.

This diff introduces a new option, `skip_invalid`, which replaces invalid chars with U+FFFD. http://en.wikipedia.org/wiki/Specials_(Unicode_block) seems to suggest that this is the correct replacement.

@override-unit-failures

Test Plan: g-unittest

Reviewed By: delong.j@fb.com

FB internal diff: D1102923

Summary:
Adds LZMA2 and LZMA2_VARINT_SIZE compression support for
folly::io::Compression.  This format shows some big wins for
compressing ELF object files and is useful in our modified
ccache client.

Test Plan:
Compression unittests.  Also, tested compressing object files built
in fbcode.  On average, the compression percentage improved from
~16.5% to ~12%.  But we save a lot more as object files get bigger,
which can help make bigger object files fit over fewer memcache
keys.

Reviewed By: njormrod@fb.com

FB internal diff: D1092576

Summary:
... even though Google's signal handler does it
(https://code.google.com/p/google-glog/source/browse/trunk/src/signalhandler.cc?r=16#235)

Assuming the existence of an invalid pthread_t is a lesser evil than reading
from another thread's stack, IMO.

Test Plan: folly/experimental/symbolizer/test

Reviewed By: simpkins@fb.com

FB internal diff: D1096620

Summary: Makes clang nervous

@override-unit-failures

Test Plan: fbconfig -r folly && fbmake dbg succeeds

Reviewed By: delong.j@fb.com

FB internal diff: D1094209

Summary:
If a previous handler was found, the code was missing the step of
actually re-raising the handler after restoring the old handler.

Test Plan: Verified that our service actually exits after receiving SIGTERM.

Reviewed By: tudorb@fb.com

FB internal diff: D1096360

Summary:
We want to install the signal handler first and then give users an opportunity
to add their callbacks.

Test Plan: built

Reviewed By: lucian@fb.com

FB internal diff: D1092083

@override-unit-failures

Summary:
(not committing yet, but I want to trigger unittests)

The glog symbolizer that we use has a few bugs (abort()s on certain small
shared libraries) and doesn't allow us to distinguish between template
specializations and function overloads (which, given that our code is more
template-heavy than Google's, has in fact been an issue).

Luckily, we have our own in folly, which doesn't have these problems and also
supports mappings from address to file and line number.

Switch translateFrames (aka the fb303 call that cpprof uses) to our symbolizer.
Also, removed a lot of dead code in common/process.

Test Plan: common/process, tested cpprof by hand

Reviewed By: lucian@fb.com

FB internal diff: D1090907

@override-unit-failures

Summary: SCOPE_SUCCESS may legitimately throw, so changed the `noexcept` attribute to a conditional one. Also added the noexcept promise to the lambdas in the hope the compiler will check code during compilation appropriately. (It currently doesn't.)

Test Plan: Added a unittest that failed, not passes.

Reviewed By: simpkins@fb.com

FB internal diff: D1093328

@override-unit-failures

Summary:
The alignment requirements and details required to avoid false
sharing belong in CacheLocality, so this diff moves them there.

@override-unit-failures

Test Plan: fbmake runtests

Reviewed By: davejwatson@fb.com

FB internal diff: D1086090

Summary:
D1090936 noticed some problems with fbstring iomanip behavior.
The root cause is that os.write(ostream, char*, size_t) is an
UnformattedOutputFunction, so disregards setw(), setfill(), and
left/right alignment.

The use of os.write instead of os << str.data() is intentional:
D367009 switched from the latter to the former so that strings
containing a '\0' are printed properly.

There does not seem to be a public function to write with formatting.
Where needed in libgcc, the function __ostream_insert is used. Since
FBString already uses such 'private' functions, __ostream_insert is an
appropriate solution.

@override-unit-failures

Test Plan:
Added test cases to FBStringTest.cpp to cover iomanip.
fbconfig -r folly && fbmake opt && fbmake runtests_opt

Reviewed By: andrei.alexandrescu@fb.com

FB internal diff: D1091474

Summary:
Also, switch to the simpler unw_backtrace(), which has the nice advantage of
actually doing IP adjustment (-1 in certain frames) correctly, unlike me :)

This is in preparation for the faster backtrace in libunwind 1.1.

Test Plan: folly/experimental/exception_tracer, folly/experimental/symbolizer, admarket/lib/util:memory_tracker_test

Reviewed By: lucian@fb.com

FB internal diff: D1088357

Summary:
CacheLocality reads cache sharing information from sysfs to
determine how CPUs should be grouped to minimize contention, Getcpu
provides fast access to the current CPU via __vdso_getcpu, and
AccessSpreader uses these two to optimally spread accesses among a
predetermined number of stripes.

AccessSpreader<>::current(n) microbenchmarks at 22 nanos, which is
substantially less than the cost of a cache miss.  This means that we
can effectively use it to reduce cache line ping-pong on striped data
structures such as IndexedMemPool or statistics counters.

Because CacheLocality looks at all of the cache levels, it can be used for
different levels of optimization.  AccessSpreader<>::stripeByChip.current()
uses as few stripes as possible to avoid off-chip communication,
AccessSpreader<>::stripeByCore.current() uses as many stripes as necessary
to get the optimal speedup, but no more.

@override-unit-failures

Test Plan: new unit tests

Reviewed By: davejwatson@fb.com

FB internal diff: D1076718

Summary:
Per-thread caching calls malloc, which means it doesn't work for malloc's
profiling (and also it isn't async-signal-safe, despite the documentation).

Test Plan: compiled

Reviewed By: philipp@fb.com

FB internal diff: D1085368

@override-unit-failures

Summary:
Also, C++ify the interface and switch to per-thread caching in libunwind as per
D1081259

Test Plan: folly/experimental/symbolizer/test

Reviewed By: lucian@fb.com

FB internal diff: D1081272

Summary:
If the compare part of cmpxchg fails, it writes the unexpected value
to %rax. At certain optimization levels this was causing me to hit an
incorrectly failed assertion in some thrift code. I also cleaned up the asm
statement to use named operands.

Test Plan: Added new test that fails before this diff.

Reviewed By: delong.j@fb.com

FB internal diff: D1083222
@override-unit-failures

Summary:
Update IOBuf::gather() and RWCursor::gather() to check their arguments
more carefully, and throw std::overflow_errors if the caller attempts to
gather more data than is available.

The comments for IOBuf::gather() claimed that it ensured that maxLength
bytes would always be available when it returned.  However, if the total
chain length was less than maxLength, it would simply coalesce the full
chain and return successfully, even though fewer than maxLength bytes
were available.  This fixes the code to throw a std::overflow_error
rather than coalescing the chain.

Additional, this updates RWCursor::gather() to ensure that it does not
attempt to gather past the end of the IOBuf chain.  Previously it could
gather past the end of the chain, coalescing the head of the chain into
the tail.  This would free the IOBuf head, which was owned by an
external caller.

A new RWCursor::gatherAtMost() API is provided for callers that wish to
gather as much as requested if possible, but still succeed if less than
this is available.

Test Plan:
Updated the unit tests to test calling gather() with more bytes than
actually available.

Reviewed By: davejwatson@fb.com

FB internal diff: D1081995

Summary:
My previous change to re-implement IOBuf's internal storage mechanism
introduced a build failure when compiling with clang.

This fixes the new compilation error in IOBuf.cpp, as well as two
existing build failures in some of the unit tests.

Test Plan: Built the folly/io code with clang.

Reviewed By: andrewjcg@fb.com

FB internal diff: D1082086

Summary:
This removes kFlagExt, and instead implements the internal buffer
optimization using operator new and delete.

IOBuf::createCombined() will allocate a single buffer for the IOBuf,
SharedInfo, and the actual data itself.  Each heap allocated IOBuf now
contains a set of flags indicating if the memory segment indicating when
it can actually be deleted.  The delete operator won't actually free the
storage until both the IOBuf and the data buffer are unused (the
SharedInfo object always becomes unused at the same time as the data
buffer).

This has a couple advantages over the old mechanism:

- Other IOBufs can continue to use and share the data storage space even
after the original IOBuf associated with it is deleted.  With the old
internal buffer mechanism, internal buffers could never be shared.

- This simplifies several parts of the code, since kFlagExt no longer
exists.  Code that previously required special handling for internal
buffers no longer needs to be aware of them.

One downside is that this can result in wasted space in some cases if
the original IOBuf is changed to point at a different buffer, since the
space for the data buffer cannot be freed until the IOBuf itself is also
destroyed.  The old internal buffer mechanism also had this problem,
which we mitigated simply by disallowing internal buffers for larger
than ~200 bytes.  With the new mechanism we currently allocate an
internal buffer for up to 1024 bytes by default, but we also allow
callers to explicitly decide if they want an internal buffer or not.

Test Plan:
Added some new unit tests for the combined buffer behavior.  Also ran
all of the existing IOBuf unit tests, with and without ASAN.  (ASAN
performs additional memory checking, but also changes IOBuf's behavior
slightly as usingJEMalloc() is false with ASAN.)

Reviewed By: davejwatson@fb.com

FB internal diff: D1072336

@override-unit-failures

Summary: Because the global cache is slow and contends on locks.

Test Plan: testinproduction

Reviewed By: philipp@fb.com

FB internal diff: D1081259

Summary: It's convenient to get back the fd, similar to unique_ptr::release().

Test Plan: unittest

Reviewed By: tudorb@fb.com

FB internal diff: D1080426

Test Plan: test added

Reviewed By: lucian@fb.com

FB internal diff: D1076170

@override-unit-failures

Summary: To be used in the new fatal signal handler.

Test Plan: test added

Reviewed By: lucian@fb.com

FB internal diff: D1076169

@override-unit-failures

Summary: To be used from the (new) fatal signal handler.

Test Plan: test added

Reviewed By: lucian@fb.com

FB internal diff: D1076168

@override-unit-failures

Summary:
Improve moveToFbString()'s code which determines if we actually have a
buffer that was allocated with malloc().

The old code simply checked if flags_ == kFlagExt.  This check is rather
fragile, and relies on very specific behavior from the existing
construction methods.  It also unnecessarily forced reallocation in some
cases.

This updates the code to have more specific checks for the flags and
fields that actually matter.  In particular, even if we have an external
buffer, if sharedInfo->freeFn is set then it is not managed by malloc().
The old check happened to work because takeOwnership() was the only
function that set set freeFn, and it also set kFlagFreeSharedInfo.

This also improves the code so that it won't unnecessarily reallocate
the buffer if kFlagMaybeShared is set but the buffer isn't really
shared.  The code will also no longer unnecessarily reallocates the
buffer just because kFlagFreeSharedInfo is set.

Test Plan:
Updated the moveToFbString() test to also test with buffers created
using takeOwnership() and wrapBuffer().

Reviewed By: davejwatson@fb.com

FB internal diff: D1072304

@override-unit-failures

Summary:
The IOBuf::reserveSlow() code assumed that external buffers were always
allocated with malloc.  This would cause problems when if reserve() was
ever used on a buffer created with IOBuf::takeOwnership().

This changes the code to now check if a user-specified free function has
been supplied.  If so, then it does not try to use realloc()/rallocm(),
and it now frees the old buffer using the specified free function rather
than free().

User-supplied buffers also have a separately allocated SharedInfo
object, which must be freed when we no longer need it.

Test Plan:
Added additional unit tests to check calling reserve() on IOBufs created
with IOBuf::takeOwnership().

Reviewed By: davejwatson@fb.com

FB internal diff: D1072292

Summary:
A dumb&slow implementation. I just needed something that outputs keys in a sorted order.
No easy to use API to discourage the use due to perf implications.

Test Plan: unit test

Reviewed By: delong.j@fb.com

FB internal diff: D1073757

Summary: Rename the 'capacity' argument so g++ will not complain when -Wshadow is enabled ('capacity' is also a method on MPMCQueue)

Test Plan: compiles, unit tests pass

Reviewed By: ngbronson@fb.com

FB internal diff: D1076220

Test Plan: unittest

Reviewed By: tudorb@fb.com

FB internal diff: D1071715

Summary:
The IOBuf.StringOperations test contained two buggy tests.  According to
the comments, they were intended to test a string spanning multiple
buffers.  They created a chain of two buffers, but used Appender to
write the string, which only writes in the final buffer in the chain.

Additionally, the test didn't request enough space in the final buffer
for the string data, and didn't allow Appender to grow the buffer.  (The
code also didn't check the return value from pushAtMost() to see if it
actually wrote all of the string data.)  The test only passed because
IOBuf would normally allocate more capacity than the caller requested.

I fixed the tests to request sufficient capacity for the string data,
and use push() rather than pushAtMost() to ensure that all data was
written correctly.  I also added two new tests which actually test with
strings spanning multiple buffers.

Test Plan:
Ran this test with some changes to IOBuf that cause it to only allocate
the requested capacity in some cases, and verified that the
StringOperations test pass now.

(The changes I was testing with use goodMallocSize() to determine the
allocation size even for small internal buffers.  When not using
jemalloc, goodMallocSize() returns the original size, so it only
allocates the capacity requested by the caller.)

Reviewed By: pgriess@fb.com

FB internal diff: D1072283