Skip to content

F
folly
Switch branch/tag
  1. 19 Oct, 2016 12 commits
  3. 18 Oct, 2016 3 commits
    • Andrew Gallagher's avatar
      folly: fixes for use with `-fvisibility-inlines-hidden` · 1a48bcd9
      Andrew Gallagher authored 
      Summary:
- Annotate non-`const` static locals in inline functions which don't require
  a single copy be used globally at runtime.
- Move implmentation from header file to source file (to avoid multiple
  copies at runtime).
- Mark a non-stateful static local as `const` to it easy to ignore when
  searching for problematic static locals.

Reviewed By: yfeldblum

Differential Revision: D4010101

fbshipit-source-id: 3be94a5dc5b7029a26e11b2145c0d41968979a5c
      1a48bcd9
    • Phil Willoughby's avatar
      Fix for ExceptionWrapperTest · b74eb888
      Phil Willoughby authored 
      Summary:
Some platforms do not output anything when `std::terminate` is called; this
fixes the test on those platforms.

Reviewed By: yfeldblum, jsedgwick

Differential Revision: D4001052

fbshipit-source-id: 671fbca9d6d22e372189b6554440c268c0cff60b
      b74eb888
    • Marcelo Juchem's avatar
      Getting fibers to build with boost 1.61 · 0b894c40
      Marcelo Juchem authored 
      Summary: This diff adapts the fibers code to the modified `fcontext` API from Boost 1.61

Reviewed By: Orvid

Differential Revision: D4035769

fbshipit-source-id: 54a97294d33c2901af78b8dec95baaefa085c4f5
      0b894c40
  5. 15 Oct, 2016 1 commit
    • Philip Pronin's avatar
      fall back to .debug_info scan in fatal signal handler · cb3a7e45
      Philip Pronin authored 
      Summary:
We've found clang might be generating incomplete `.debug_aranges`,
while falling back to linear `.debug_info` scan is too expensive and shouldn't
be used by default, we can afford doing that in fatal signal handler.

Also optimize `exception_tracer::printExceptionInfo()` to avoid `LocationInfo`
resolution if `NO_FILE_AND_LINE` is used.

Reviewed By: luciang, ot

Differential Revision: D4020989

fbshipit-source-id: 84172208736b224c19206da48bcb3b5c0b2c67d0
      cb3a7e45
  7. 14 Oct, 2016 3 commits
    • Nicholas Ormrod's avatar
      Backed out changeset 09c1712854b3 · fc838f2b
      Nicholas Ormrod authored 
      Summary: Removing COW from fbstring had adverse memory consequences when sync'd with libgcc. Revert this diff to keep folly and libgcc in sync.

Reviewed By: yfeldblum, luciang

Differential Revision: D4019604

fbshipit-source-id: 80bd31c220098bfab37f0effc90f67876432369d
      fc838f2b
    • Shubhanshu Agrawal's avatar
      BatchDispatcher exception handling · 2c0ece4a
      Shubhanshu Agrawal authored 
      Summary: error handling for batch dispatcher, for the case where dispatch function throws.

Reviewed By: yfeldblum

Differential Revision: D4016246

fbshipit-source-id: ac61b873a425128be7839a119591fca10608210a
      2c0ece4a
    • Matthieu Martin's avatar
      noexcept CollectVariadicContext destructor · 954899b8
      Matthieu Martin authored 
      Summary: To be explicit about our assumption from D4013691

Reviewed By: andriigrynenko

Differential Revision: D4015816

fbshipit-source-id: 24f984ac9959d0c4c0aeac34b3dce8dfcf2c1354
      954899b8
  9. 13 Oct, 2016 5 commits
  11. 12 Oct, 2016 7 commits
    • Zonr Chang's avatar
      Access SSL cipher info with a const reference. · cef27cd0
      Zonr Chang authored 
      Summary:
BoringSSL returns `const SSL_CIPHER*` from sk_SSL_CIPHER_value().
Closes https://github.com/facebook/folly/pull/492

Reviewed By: yfeldblum

Differential Revision: D4009712

Pulled By: Orvid

fbshipit-source-id: cca2e8cb6aab0a459dca55b8f23ef57fa5c8251f
      cef27cd0
    • Zonr Chang's avatar
      Fix call to SSL_SESSION_up_ref(). · b6dfac67
      Zonr Chang authored 
      Summary:
`SSL_SESSION_up_ref()` in both BoringSSL and OpenSSL 1.1.0 takes a
`SSL_SESSION*`.
Closes https://github.com/facebook/folly/pull/493

Reviewed By: yfeldblum

Differential Revision: D4009706

Pulled By: Orvid

fbshipit-source-id: 68ea201821e7a6d5835a79b34187344ba251e239
      b6dfac67
    • Philip Pronin's avatar
      ZSTDCodec should properly handle non-shortcut levels · 339ce0bf
      Philip Pronin authored 
      Summary:
Currently only shortcut levels are properly handled, others result in
using level 1.

Reviewed By: yfeldblum

Differential Revision: D4008123

fbshipit-source-id: 37845eeec139007738f99e72ecfb969c6a2e5652
      339ce0bf
    • Phil Willoughby's avatar
      Print why we called std::terminate · fec7b1bc
      Phil Willoughby authored 
      Summary:
When terminating because we were told to rethrow a nonexistent exception, we
now print that to `std::cerr` for the benefit of users on platforms which do not dump a stack trace on
`std::terminate`

Reviewed By: yfeldblum

Differential Revision: D4001087

fbshipit-source-id: 5b834cfc75bf42bec06a49ac9db55789a83df180
      fec7b1bc
    • Giuseppe Ottaviano's avatar
      Add support in enumerate() for iterators with exotic pointers · 1d2f2dcf
      Giuseppe Ottaviano authored 
      Summary: I thought we wouldn't need this but it turns out Thrift Frozen2 iterators return proxies (it's proxies all the way down). Thanks to ericniebler for trick.

Reviewed By: yfeldblum

Differential Revision: D4005700

fbshipit-source-id: 1911996afa075c1d819a3aaea2ee924bc2ae2f20
      1d2f2dcf
    • Christopher Dykes's avatar
      Fix build on Mac OSX Sierra · a34e06a1
      Christopher Dykes authored 
      Summary:
There are two changes here.
The first is to eliminate `detail::DEFAULT_CLOCK_ID` from `Benchmark.[cpp|h]` as Sierra defines `clockid_t` as an enum type, which means that calling `clock_gettime(detail::DEFAULT_CLOCK_ID` would fail, because the enums are incompatible. As this was being used as a default, but is not actually changable anywhere, I just got rid of `detail::DEFAULT_CLOCK_ID` entirely.
The second is to move `portability/BitsFunctexcept.cpp` into `libfollybase_la_SOURCES`, because it's needed for generating the fingerprint tables.

Reviewed By: yfeldblum

Differential Revision: D4004843

fbshipit-source-id: b2a9c33f8e516d8eb3cdc5ab093f4946ac9ed37e
      a34e06a1
    • Lucian Grijincu's avatar
      folly: signal-handler: dynamic cache size, based on number of dynamic-loaded ELF files · 74ea0a31
      Lucian Grijincu authored 
      Summary:
I added kFatalSignalHandlerCacheSize in {D3984649}, and chose a big
number to satisfy most use-cases. But the size increase is non-trivial.

Memory usage:

- before: 5.1 MB in folly::symbolizer::SignalSafeElfCache::SignalSafeElfCache
- after: 80.0 MB in folly::symbolizer::SignalSafeElfCache::SignalSafeElfCache

Switch to a dynamic approach and everyone pays for what they use.

https://github.com/bminor/glibc/blob/ee19f1de0d0da24114be554fdf94243c0ec6b86c/elf/rtld-debugger-interface.txt#L10-L20
```
The r_debug structure contains (amongst others) the following fields:

  struct link_map *r_map:
    A linked list of loaded objects.

  enum { RT_CONSISTENT, RT_ADD, RT_DELETE } r_state:
    The current state of the r_map list.  RT_CONSISTENT means that r_map
    is not currently being modified and may safely be inspected.  RT_ADD
    means that an object is being added to r_map, and that the list is
    not guaranteed to be consistent.  Likewise RT_DELETE means that an
    object is being removed from the list.
```

https://github.com/bminor/glibc/blob/ee19f1de0d0da24114be554fdf94243c0ec6b86c/elf/rtld.c#L303-L307
```
  /* Call the OS-dependent function to set up life so we can do things like
     file access.  It will call `dl_main' (below) to do all the real work
     of the dynamic linker, and then unwind our frame and run the user
     entry point on the same stack we entered on.  */
  start_addr = _dl_sysdep_start (arg, &dl_main);
```

dl_main: https://github.com/bminor/glibc/blob/ee19f1de0d0da24114be554fdf94243c0ec6b86c/elf/rtld.c#L1192-L1199
```
  /* Initialize the data structures for the search paths for shared
     objects.  */
  _dl_init_paths (library_path);

  /* Initialize _r_debug.  */
  struct r_debug *r = _dl_debug_initialize (GL(dl_rtld_map).l_addr,
					    LM_ID_BASE);
  r->r_state = RT_CONSISTENT;
...

  /* We start adding objects.  */
  r->r_state = RT_ADD;
  _dl_debug_state ();
  LIBC_PROBE (init_start, 2, LM_ID_BASE, r);

...

  /* Notify the debugger all new objects are now ready to go.  We must re-get
     the address since by now the variable might be in another object.  */
  r = _dl_debug_initialize (0, LM_ID_BASE);
  r->r_state = RT_CONSISTENT;
  _dl_debug_state ();
  LIBC_PROBE (init_complete, 2, LM_ID_BASE, r);

```

Reviewed By: bixue2010

Differential Revision: D3996974

fbshipit-source-id: e24d72e3cc0339e4cf1acdd2f4c9a7ebfcfdf739
      74ea0a31
  13. 11 Oct, 2016 3 commits
    • Jim Meyering's avatar
      folly/ConcurrentSkipList.h: avoid shadowing warnings (trivial) · 70b158bb
      Jim Meyering authored 
      Summary:
Compiling with gcc's proposed -Wshadow-compatible-local option
exposed some fix-worthy warnings. Rename to avoid the shadowing.

Reviewed By: evilmucedin

Differential Revision: D3999026

fbshipit-source-id: 26cb3033ba8c5538cc9217993f2fda6aef954a8f
      70b158bb
    • Shubhanshu Agrawal's avatar
      BatchDispatcher add() interface · fb77b407
      Shubhanshu Agrawal authored 
      Summary: converting add(value&&) to add(value) to make it easier for callsites to use.

Reviewed By: A5he

Differential Revision: D4000416

fbshipit-source-id: c66f4d93d1223286bbc786d17f65ed02a63cc794
      fb77b407
    • Louis Boval's avatar
      Disabling failing test for fbandroid · 40e6161e
      Louis Boval authored 
      Summary: this particular test fails reliably for fbandroid. Modifying the code in fbcode, assuming it will be synced promptly.

Reviewed By: edelron

Differential Revision: D4000722

fbshipit-source-id: fca5e691ea5c972ddd245acee1c17720730539d4
      40e6161e
  15. 10 Oct, 2016 3 commits
    • Nicholas Ormrod's avatar
      Remove memcpy-UB in fbstring · 7c2ef80c
      Nicholas Ormrod authored 
      Summary:
The bug report in t13764686 shows that `folly::StringPiece().str()` causes undefined behavior by passing nullptr into memcpy: the default-initialized StringPiece has two null pointers, which will call `string(char*, size)` with nullptr and 0. This gets forwarded to `podCopy` in fbstring, which calls memcpy and has undefined behavior.

The call pattern `string(nullptr, 0)` is not illegal (note: that syntax is ambiguous, as the literal `0` can be converted to a pointer). Prevent this legal call pattern from causing undefined behavior in fbstring.

This diff puts asserts in podCopy to catch errors, and ensures that current podCopy callsites ensure that the pointer is non-null.

Most podCopy callsites do not need to guarded, since the size is guaranteed to be greater than 0 (which will hard crash instead). The interesting callsite is in `initMedium`, because there is a mode (disableSSO) which will allocate empty strings on the heap, so we need to add a guard there.

Reviewed By: luciang

Differential Revision: D3996440

fbshipit-source-id: b311a311973d1d969542245c72035c5b38da58e3
      7c2ef80c
    • Maged Michael's avatar
      Update hazard pointers prototype · 328c1a56
      Maged Michael authored 
      Summary:
Interface:
- Got rid of std::function reclamation functions and added a Deleter template parameter.
- Got rid of the flush() member functions of hazptr_domain
- Added a lock-free non-bool member function to get a protected pointer.
Implementation:
- Implemented the interface changes.
- Changed the order of accesses in reading the shared list of objects vs reading the hazard pointers. I think the previous order would have allowed recently protected objects to be reclaimed incorrectly.
Updated the examples and tests accordingly.

Reviewed By: davidtgoldblatt

Differential Revision: D3981284

fbshipit-source-id: 35ff60da3aea1f67c58d82437dda58f6d8b07bf5
      328c1a56
    • Phil Willoughby's avatar
      Make exception_wrapper::throwException() strict · a62c9841
      Phil Willoughby authored 
      Summary:
The definition of `throwException()` is changed to add the `[[noreturn]]`
attribute for consistency with `std::rethrow_exception`. If the
`exception_wrapper` contains an exception, that is rethrown; if not, then
`std::terminate` is called.

An alternative design which threw an exception if the `exception_wrapper` did not contain an exception was rejected because of the ambiguity it introduced as to whether the exception thrown was wrapped or not.

Benchmarks before:

```

============================================================================
folly/test/ExceptionWrapperBenchmark.cpp        relative  time/iter  iters/s
============================================================================
exception_ptr_create_and_test                                1.45us  689.01K
exception_wrapper_create_and_test               4337.80%    33.46ns   29.89M
----------------------------------------------------------------------------
exception_ptr_create_and_test_concurrent                   342.99us    2.92K
exception_wrapper_create_and_test_concurrent     101.41%   338.21us    2.96K
----------------------------------------------------------------------------
exception_ptr_create_and_throw                               3.05us  327.89K
exception_wrapper_create_and_throw               140.22%     2.17us  459.77K
exception_wrapper_create_and_cast               8956.80%    34.05ns   29.37M
----------------------------------------------------------------------------
exception_ptr_create_and_throw_concurrent                  372.68us    2.68K
exception_wrapper_create_and_throw_concurrent    102.54%   363.44us    2.75K
exception_wrapper_create_and_cast_concurrent     110.93%   335.97us    2.98K
============================================================================
```

and after:

```

============================================================================
folly/test/ExceptionWrapperBenchmark.cpp        relative  time/iter  iters/s
============================================================================
exception_ptr_create_and_test                                1.46us  684.09K
exception_wrapper_create_and_test               4368.84%    33.46ns   29.89M
----------------------------------------------------------------------------
exception_ptr_create_and_test_concurrent                   341.20us    2.93K
exception_wrapper_create_and_test_concurrent      99.88%   341.60us    2.93K
----------------------------------------------------------------------------
exception_ptr_create_and_throw                               3.08us  324.93K
exception_wrapper_create_and_throw               106.93%     2.88us  347.46K
exception_wrapper_create_and_cast               9532.05%    32.29ns   30.97M
----------------------------------------------------------------------------
exception_ptr_create_and_throw_concurrent                  363.73us    2.75K
exception_wrapper_create_and_throw_concurrent    101.04%   360.00us    2.78K
exception_wrapper_create_and_cast_concurrent     104.50%   348.07us    2.87K
============================================================================
```

Reviewed By: yfeldblum

Differential Revision: D3923939

fbshipit-source-id: 6fc58ac571e59e4122f1fbd194c678e3a3841057
      a62c9841
  17. 09 Oct, 2016 1 commit
    • Jim Meyering's avatar
      folly/Foreach.h: allow FOR_EACH to be nested with no shadowing · ba22e52f
      Jim Meyering authored 
      Summary:
Prior to this change, any nested use of FOR_EACH would
induce a shadowed declaration for each of the two local
state variables it declares.

This makes the names of those variables __LINE__-dependent,
so that there is no shadowing, as long as each nested use
is on a different line.

This also adds a new test that (prior to this change) would fail to
compile with an option like -Werror=shadow-compatible-local.

Since this change relies on cpp token concatenation, I have included
The fix defines a new helper macro, _FE_ANON, to derive each new variable
name. I wondered whether to do this for every other FOR_* macro here, but
since so far, I have encountered more than 10 cases of nested FOR_EACH
uses in a large corpus, but no nesting of any other FOR_* macro, I am
content to do it only for this one.

Reviewed By: yfeldblum

Differential Revision: D3992956

fbshipit-source-id: f26fba89bc661bb9d22747dec0acdcf8c648fb83
      ba22e52f
  19. 08 Oct, 2016 2 commits