Fixes PR45673
The commit 9180c14fe4 (D76206) resolved only a part of the problem
of concurrent .gcda file creation. It ensured that only one process
creates the file but did not ensure that the process locks the
file first. If not, the process which created the file may clobber
the contents written by a process which locked the file first.
This is the cause of PR45673.
This commit prevents the clobbering by revising the assumption
that a process which creates the file locks the file first.
Regardless of file creation, a process which locked the file first
uses fwrite (new_file==1) and other processes use mmap (new_file==0).
I also tried to keep the creation/first-lock process same by using
mkstemp/link/unlink but the code gets long. This commit is more
simple.
Note: You may be confused with other changes which try to resolve
concurrent file access. My understanding is (may not be correct):
D76206: Resolve race of .gcda file creation (but not lock)
This one: Resolve race of .gcda file creation and lock
D54599: Same as D76206 but abandoned?
D70910: Resolve race of multi-threaded counter flushing
D74953: Resolve counter sharing between parent/children processes
D78477: Revision of D74953
Differential Revision: https://reviews.llvm.org/D79556
In a big-endian .gcda file, the first four bytes are "gcda" instead of "adcg".
All 32-bit values are in big-endian.
With this change, libclang_rt.profile can hopefully produce gcov
compatible output.
D49132 is partially correct. For 64-bit values, the lower 32-bit part comes
before the higher 32-bit part (in a little-endian manner).
For 32-bit values, libgcov reads/writes 32-bit values in native endianness.
gcov 9 (r264462) started to use GCOV_TAG_OBJECT_SUMMARY. Before,
GCOV_TAG_PROGRAM_SUMMARY was used.
libclang_rt.profile should emit just one tag according to the version.
Another bug introduced by rL194499 is that the wrong runcount field was
selected.
Fix the two bugs so that gcov can correctly decode "Runs:" from
libclang_rt.profile produced .gcda files, and llvm-cov gcov can
correctly decode "Runs:" from libgcov produced .gcda files.
Defaulting to -Xclang -coverage-version='407*' makes .gcno/.gcda
compatible with gcov [4.7,8)
In addition, delete clang::CodeGenOptionsBase::CoverageExtraChecksum and GCOVOptions::UseCfgChecksum.
We can infer the information from the version.
With this change, .gcda files produced by `clang --coverage a.o` linked executable can be read by gcov 4.7~7.
We don't need other -Xclang -coverage* options.
There may be a mismatching version warning, though.
(Note, GCC r173147 "split checksum into cfg checksum and line checksum"
made gcov 4.7 incompatible with previous versions.)
rL144865 incorrectly wrote function names for GCOV_TAG_FUNCTION
(this might be part of the reasons the header says
"We emit files in a corrupt version of GCOV's "gcda" file format").
rL176173 and rL177475 realized the problem and introduced -coverage-no-function-names-in-data
to work around the issue. (However, the description is wrong.
libgcov never writes function names, even before GCC 4.2).
In reality, the linker command line has to look like:
clang --coverage -Xclang -coverage-version='407*' -Xclang -coverage-cfg-checksum -Xclang -coverage-no-function-names-in-data
Failing to pass -coverage-no-function-names-in-data can make gcov 4.7~7
either produce wrong results (for one gcov-4.9 program, I see "No executable lines")
or segfault (gcov-7).
(gcov-8 uses an incompatible format.)
This patch deletes -coverage-no-function-names-in-data and the related
function names support from libclang_rt.profile
Summary:
When forking in several threads, the counters were written out in using the same global static variables (see GCDAProfiling.c): that leads to crashes.
So when there is a fork, the counters are resetted in the child process and they will be dumped at exit using the interprocess file locking.
When there is an exec, the counters are written out and in case of failures they're resetted.
Reviewers: jfb, vsk, marco-c, serge-sans-paille
Reviewed By: marco-c, serge-sans-paille
Subscribers: llvm-commits, serge-sans-paille, dmajor, cfe-commits, hiraditya, dexonsmith, #sanitizers, marco-c, sylvestre.ledru
Tags: #sanitizers, #clang, #llvm
Differential Revision: https://reviews.llvm.org/D78477
Now that write data continously into the memory mapping, we don't need
to keep the VMO handle around after it has been mapped. This change also
ensures that the VMO is always closed on error.
Differential Revision: https://reviews.llvm.org/D76963
The intent of the `llvm_gcda_start_file` function is that only
one process create the .gcda file and initialize it to be updated
by other processes later.
Before this change, if multiple processes are started simultaneously,
some of them may initialize the file because both the first and
second `open` calls may succeed in a race condition and `new_file`
becomes 1 in those processes. This leads incorrect coverage counter
values. This often happens in MPI (Message Passing Interface) programs.
The test program added in this change is a simple reproducer.
This change ensures only one process creates/initializes the file by
using the `O_EXCL` flag.
Differential Revision: https://reviews.llvm.org/D76206
This avoids the test failure that was introduced in rG32bddad where
this function pulls in the rest of InstrProfilingFile.c which is
undesirable in use cases when profile runtime is being used without
the rest of libc.
This also allows additional cleanup by eliminating another variable
from platforms that don't need it.
Differential Revision: https://reviews.llvm.org/D76750
On Fuchsia, we always use the continuous mode with runtime counter
relocation, so there's no need for atexit hook or support for dumping
the profile manually.
Differential Revision: https://reviews.llvm.org/D76556
While the VMO size is always page aligned, we can record the content
size as a property and then use this metadata when writing the profile
to a file.
Differential Revision: https://reviews.llvm.org/D76402
See discussion on PR44792.
This reverts commit 02ce9d8ef5.
It also reverts the follow-up commits
8f46269f0 "[profile] Don't dump counters when forking and don't reset when calling exec** functions"
62c7d8402 "[profile] gcov_mutex must be static"
Summary:
There is no need to write out gcdas when forking because we can just reset the counters in the parent process.
Let say a counter is N before the fork, then fork and this counter is set to 0 in the child process.
In the parent process, the counter is incremented by P and in the child process it's incremented by C.
When dump is ran at exit, parent process will dump N+P for the given counter and the child process will dump 0+C, so when the gcdas are merged the resulting counter will be N+P+C.
About exec** functions, since the current process is replaced by an another one there is no need to reset the counters but just write out the gcdas since the counters are definitely lost.
To avoid to have lists in a bad state, we just lock them during the fork and the flush (if called explicitely) and lock them when an element is added.
Reviewers: marco-c
Reviewed By: marco-c
Subscribers: hiraditya, cfe-commits, #sanitizers, llvm-commits, sylvestre.ledru
Tags: #clang, #sanitizers, #llvm
Differential Revision: https://reviews.llvm.org/D74953
This is an alternative to the continous mode that was implemented in
D68351. This mode relies on padding and the ability to mmap a file over
the existing mapping which is generally only available on POSIX systems
and isn't suitable for other platforms.
This change instead introduces the ability to relocate counters at
runtime using a level of indirection. On every counter access, we add a
bias to the counter address. This bias is stored in a symbol that's
provided by the profile runtime and is initially set to zero, meaning no
relocation. The runtime can mmap the profile into memory at abitrary
location, and set bias to the offset between the original and the new
counter location, at which point every subsequent counter access will be
to the new location, which allows updating profile directly akin to the
continous mode.
The advantage of this implementation is that doesn't require any special
OS support. The disadvantage is the extra overhead due to additional
instructions required for each counter access (overhead both in terms of
binary size and performance) plus duplication of counters (i.e. one copy
in the binary itself and another copy that's mmapped).
Differential Revision: https://reviews.llvm.org/D69740
The executable acquires an advisory record lock (`fcntl(fd, F_SETLKW, *)`) on a profile file.
Merge pool size >= 10 may be beneficial when the concurrency is large.
Also fix a small problem about snprintf. It can cause the filename to be truncated after %m.
Reviewed By: davidxl
Differential Revision: https://reviews.llvm.org/D71970
Summary:
Counters can be flushed in a multi-threaded context for example when the process is forked in different threads (https://github.com/llvm/llvm-project/blob/master/llvm/lib/Transforms/Instrumentation/GCOVProfiling.cpp#L632-L663).
In order to avoid pretty bad things, a critical section is needed around the flush.
We had a lot of crashes in this code in Firefox CI when we switched to clang for linux ccov builds and those crashes disappeared with this patch.
Reviewers: marco-c, froydnj, dmajor, davidxl, vsk
Reviewed By: marco-c, dmajor
Subscribers: ahatanak, froydnj, dmajor, dberris, jfb, #sanitizers, llvm-commits, sylvestre.ledru
Tags: #sanitizers, #llvm
Differential Revision: https://reviews.llvm.org/D70910
When writing out a profile, avoid allocating a page on the stack for the
purpose of writing out zeroes, as some embedded environments do not have
enough stack space to accomodate this.
Instead, use a small, fixed-size zero buffer that can be written
repeatedly.
For a synthetic file with >100,000 functions, I did not measure a
significant difference in profile write times. We are removing a
page-length zero-fill `memset()` in favor of several smaller buffered
`fwrite()` calls: in practice, I am not sure there is much of a
difference. The performance impact is only expected to affect the
continuous sync mode (%c) -- zero padding is less than 8 bytes in all
other cases.
rdar://57810014
Differential Revision: https://reviews.llvm.org/D71323
Summary:
Counters can be flushed in a multi-threaded context for example when the process is forked in different threads (https://github.com/llvm/llvm-project/blob/master/llvm/lib/Transforms/Instrumentation/GCOVProfiling.cpp#L632-L663).
In order to avoid pretty bad things, a critical section is needed around the flush.
We had a lot of crashes in this code in Firefox CI when we switched to clang for linux ccov builds and those crashes disappeared with this patch.
Reviewers: marco-c, froydnj, dmajor, davidxl
Reviewed By: marco-c, dmajor
Subscribers: froydnj, dmajor, dberris, jfb, #sanitizers, llvm-commits, sylvestre.ledru
Tags: #sanitizers, #llvm
Differential Revision: https://reviews.llvm.org/D70910
See PR43425:
https://bugs.llvm.org/show_bug.cgi?id=43425
When writing profile data on Windows we were opening profile file with
exclusive read/write access.
In case we are trying to write to the file from multiple processes
simultaneously, subsequent calls to CreateFileA would return
INVALID_HANDLE_VALUE.
To fix this, I changed to open without exclusive access and then take a
lock.
Patch by Michael Holman!
Differential revision: https://reviews.llvm.org/D70330
This header fragment is useful on its own for any consumer that wants
to use custom instruction profile runtime with the LLVM instrumentation.
The concrete use case is in Fuchsia's kernel where we want to use
instruction profile instrumentation, but we cannot use the compiler-rt
runtime because it's not designed for use in the kernel environment.
This change allows installing this header as part of compiler-rt.
Differential Revision: https://reviews.llvm.org/D64532
This header fragment is useful on its own for any consumer that wants
to use custom instruction profile runtime with the LLVM instrumentation.
The concrete use case is in Fuchsia's kernel where we want to use
instruction profile instrumentation, but we cannot use the compiler-rt
runtime because it's not designed for use in the kernel environment.
This change allows installing this header as part of compiler-rt.
Differential Revision: https://reviews.llvm.org/D64532
Continuous mode is not yet supported on Fuchsia/Windows, however an
error should not be reported unless the user attempted to actually
enable continuous mode.
Make it possible to use online profile merging ("%m" mode) with
continuous sync ("%c" mode).
To implement this, the merged profile is locked in the runtime
initialization step and either a) filled out for the first time or b)
checked for compatibility. Then, the profile can simply be mmap()'d with
MAP_SHARED set. With the mmap() in place, counter updates from every
process which uses an image are mapped onto the same set of physical
pages assigned by the filesystem cache. After the mmap() is set up, the
profile is unlocked.
Differential Revision: https://reviews.llvm.org/D69586
Split out the logic to get the size of a merged profile and to do a
compatibility check. This can be shared with both the continuous+merging
mode implementation, as well as the runtime-allocated counters
implementation planned for Fuchsia.
Lifted out of D69586.
Differential Revision: https://reviews.llvm.org/D70135
A few years back a similar change was made for getenv since neither function is supported on the PS4 platform.
Recently, commit d889d1e added a call to setenv in compiler-rt which was causing linking errors because the symbol was not found. This fixes that issue by putting in a shim similar to how we previously dealt with the lack of getenv.
Differential Revision: https://reviews.llvm.org/D70033
This commit added use of a Windows API in InstrProfilingPort.h.
When _MSC_VER is defined (for MSVC), windows.h is already included
earlier in the same header (for atomics), but MinGW, the gcc
atomics builtins are used instead. Therefore explicitly include
windows.h here, where the API is used.
Use the printf macros from inttypes.h to sidestep -Wformat issues:
/var/lib/buildbot/sanitizer-buildbot6/sanitizer-x86_64-linux-android/build/llvm-project/compiler-rt/lib/profile/InstrProfilingFile.c:425:14: error: format specifies type 'long long' but the argument has type 'off_t' (aka 'long') [-Werror,-Wformat]
CurrentFileOffset, PageSize);
^~~~~~~~~~~~~~~~~
/var/lib/buildbot/sanitizer-buildbot6/sanitizer-x86_64-linux-android/build/llvm-project/compiler-rt/lib/profile/InstrProfilingPort.h:114:50: note: expanded from macro 'PROF_ERR'
fprintf(stderr, "LLVM Profile Error: " Format, __VA_ARGS__);
~~~~~~ ^~~~~~~~~~~
/var/lib/buildbot/sanitizer-buildbot6/sanitizer-x86_64-linux-android/build/llvm-project/compiler-rt/lib/profile/InstrProfilingFile.c:461:41: error: format specifies type 'unsigned long long' but the argument has type 'uint64_t' (aka 'unsigned long') [-Werror,-Wformat]
strerror(errno), CountersBegin, PageAlignedCountersLength, Fileno,
^~~~~~~~~~~~~~~~~~~~~~~~~
/var/lib/buildbot/sanitizer-buildbot6/sanitizer-x86_64-linux-android/build/llvm-project/compiler-rt/lib/profile/InstrProfilingPort.h:114:50: note: expanded from macro 'PROF_ERR'
fprintf(stderr, "LLVM Profile Error: " Format, __VA_ARGS__);
~~~~~~ ^~~~~~~~~~~
/var/lib/buildbot/sanitizer-buildbot6/sanitizer-x86_64-linux-android/build/llvm-project/compiler-rt/lib/profile/InstrProfilingFile.c:462:9: error: format specifies type 'unsigned long long' but the argument has type 'uint64_t' (aka 'unsigned long') [-Werror,-Wformat]
FileOffsetToCounters);
^~~~~~~~~~~~~~~~~~~~
/var/lib/buildbot/sanitizer-buildbot6/sanitizer-x86_64-linux-android/build/llvm-project/compiler-rt/lib/profile/InstrProfilingPort.h:114:50: note: expanded from macro 'PROF_ERR'
fprintf(stderr, "LLVM Profile Error: " Format, __VA_ARGS__);
VLAs in C appear to not work on Windows, so use COMPILER_RT_ALLOCA:
C:\b\slave\sanitizer-windows\llvm-project\compiler-rt\lib\profile\InstrProfilingWriter.c(264): error C2057: expected constant expression
C:\b\slave\sanitizer-windows\llvm-project\compiler-rt\lib\profile\InstrProfilingWriter.c(264): error C2466: cannot allocate an array of constant size 0
C:\b\slave\sanitizer-windows\llvm-project\compiler-rt\lib\profile\InstrProfilingWriter.c(264): error C2133: 'Zeroes': unknown size
Add support for continuously syncing profile counter updates to a file.
The motivation for this is that programs do not always exit cleanly. On
iOS, for example, programs are usually killed via a signal from the OS.
Running atexit() handlers after catching a signal is unreliable, so some
method for progressively writing out profile data is necessary.
The approach taken here is to mmap() the `__llvm_prf_cnts` section onto
a raw profile. To do this, the linker must page-align the counter and
data sections, and the runtime must ensure that counters are mapped to a
page-aligned offset within a raw profile.
Continuous mode is (for the moment) incompatible with the online merging
mode. This limitation is lifted in https://reviews.llvm.org/D69586.
Continuous mode is also (for the moment) incompatible with value
profiling, as I'm not sure whether there is interest in this and the
implementation may be tricky.
As I have not been able to test extensively on non-Darwin platforms,
only Darwin support is included for the moment. However, continuous mode
may "just work" without modification on Linux and some UNIX-likes. AIUI
the default value for the GNU linker's `--section-alignment` flag is set
to the page size on many systems. This appears to be true for LLD as
well, as its `no_nmagic` option is on by default. Continuous mode will
not "just work" on Fuchsia or Windows, as it's not possible to mmap() a
section on these platforms. There is a proposal to add a layer of
indirection to the profile instrumentation to support these platforms.
rdar://54210980
Differential Revision: https://reviews.llvm.org/D68351
When the %m filename pattern is used, the filename is unique to each
image, so the cached value is wrong.
It struck me that the full filename isn't something that's recomputed
often, so perhaps it doesn't need to be cached at all. David Li pointed
out we can go further and just hide lprofCurFilename. This may regress
workflows that depend on using the set-filename API to change filenames
across all loaded DSOs, but this is expected to be very rare.
rdar://55137071
Differential Revision: https://reviews.llvm.org/D69137
llvm-svn: 375301
This reverts commit r371584. It introduced a dependency from compiler-rt
to llvm/include/ADT, which is problematic for multiple reasons.
One is that it is a novel dependency edge, which needs cross-compliation
machinery for llvm/include/ADT (yes, it is true that right now
compiler-rt included only header-only libraries, however, if we allow
compiler-rt to depend on anything from ADT, other libraries will
eventually get used).
Secondly, depending on ADT from compiler-rt exposes ADT symbols from
compiler-rt, which would cause ODR violations when Clang is built with
the profile library.
llvm-svn: 371598
This patch contains the basic functionality for reporting potentially
incorrect usage of __builtin_expect() by comparing the developer's
annotation against a collected PGO profile. A more detailed proposal and
discussion appears on the CFE-dev mailing list
(http://lists.llvm.org/pipermail/cfe-dev/2019-July/062971.html) and a
prototype of the initial frontend changes appear here in D65300
We revised the work in D65300 by moving the misexpect check into the
LLVM backend, and adding support for IR and sampling based profiles, in
addition to frontend instrumentation.
We add new misexpect metadata tags to those instructions directly
influenced by the llvm.expect intrinsic (branch, switch, and select)
when lowering the intrinsics. The misexpect metadata contains
information about the expected target of the intrinsic so that we can
check against the correct PGO counter when emitting diagnostics, and the
compiler's values for the LikelyBranchWeight and UnlikelyBranchWeight.
We use these branch weight values to determine when to emit the
diagnostic to the user.
A future patch should address the comment at the top of
LowerExpectIntrisic.cpp to hoist the LikelyBranchWeight and
UnlikelyBranchWeight values into a shared space that can be accessed
outside of the LowerExpectIntrinsic pass. Once that is done, the
misexpect metadata can be updated to be smaller.
In the long term, it is possible to reconstruct portions of the
misexpect metadata from the existing profile data. However, we have
avoided this to keep the code simple, and because some kind of metadata
tag will be required to identify which branch/switch/select instructions
are influenced by the use of llvm.expect
Patch By: paulkirth
Differential Revision: https://reviews.llvm.org/D66324
llvm-svn: 371584
We want to avoid doing expensive work during atexit since the process
might be terminated before we can publish the VMO and write out the
symbolizer markup, so move the VMO creation to the initialization
phase and only write data during the atexit phase.
Differential Revision: https://reviews.llvm.org/D66323
llvm-svn: 369180
Summary:
In Android, attempting to mkdir() or even stat() top-level directories
like /data causes noisy selinux denials. During whole-system coverage
instrumentation, this causes a deluge of noisy messages that drown out
legitimate selinux denials, that should be audited and fixed.
To avoid this, skip creating any directory in GCOV_PREFIX (thereby
assuming that it exists).
- Android platform ensures that the GCOV_PREFIX used in Android is
created and read/writable by all processes.
- This only affects the Android platform (by checking against
__ANDROID_API_FUTURE__) and for apps built with Clang coverage, the
runtime will still create any non-existent parent directories for the
coverage files.
Reviewers: srhines, davidxl
Subscribers: krytarowski, #sanitizers, danalbert, llvm-commits
Tags: #sanitizers, #llvm
Differential Revision: https://reviews.llvm.org/D65245
llvm-svn: 367064
Summary:
User code can open a file on its own and pass it to the runtime, rather than
specifying a name and having the runtime open the file. This supports the use
case where a process cannot open a file on its own but can receive a file
descriptor from another process.
Relanding https://reviews.llvm.org/D62541. The original revision unlocked
the file before calling flush, this revision fixes that.
Reviewers: Dor1s, davidxl
Reviewed By: Dor1s
Subscribers: #sanitizers, llvm-commits
Tags: #sanitizers, #llvm
Differential Revision: https://reviews.llvm.org/D63581
llvm-svn: 364231
This caused instrumented Clang to become crashy. See llvm-commits thread
for repro steps.
This also reverts follow-up r362716 which added test cases.
> Author: Sajjad Mirza
>
> Differential Revision: http://reviews.llvm.org/D62541
llvm-svn: 363134
- Several "warning: extra ';' [-Wpedantic]"
- One "C++ style comments are not allowed in ISO C90 [enabled by default]"
in a file that uses C style comments everywhere but in one place
llvm-svn: 360430
Douglas Yung