diff options
| author | wittman <wittman@chromium.org> | 2015-10-15 15:07:35 -0700 |
|---|---|---|
| committer | Commit bot <commit-bot@chromium.org> | 2015-10-15 22:08:30 +0000 |
| commit | 4c7da618ef084822c587c62a77bfce0bbbe25608 (patch) | |
| tree | 4352804294731c1e351646f63383519304fc789f | |
| parent | dbde4d0b8dffa1f06ab264f36dda051386772b0b (diff) | |
| download | chromium_src-4c7da618ef084822c587c62a77bfce0bbbe25608.zip chromium_src-4c7da618ef084822c587c62a77bfce0bbbe25608.tar.gz chromium_src-4c7da618ef084822c587c62a77bfce0bbbe25608.tar.bz2 | |
Stack sampling profiler: work around invalid BoringSSL assembly
BoringSSL's custom x86_64 assembly code doesn't follow the Microsoft x64
calling convention, but still has unwind information. This gives wrong
results unwinding the stack in some cases, generating instruction
pointers that don't point within known modules.
This change works around this case in the profiler code until the
BoringSSL code can be fixed.
BUG=542919
Review URL: https://codereview.chromium.org/1404403002
Cr-Commit-Position: refs/heads/master@{#354375}
| -rw-r--r-- | base/profiler/win32_stack_frame_unwinder.cc | 8 |
1 files changed, 7 insertions, 1 deletions
diff --git a/base/profiler/win32_stack_frame_unwinder.cc b/base/profiler/win32_stack_frame_unwinder.cc index ea589bb..2f09e8e 100644 --- a/base/profiler/win32_stack_frame_unwinder.cc +++ b/base/profiler/win32_stack_frame_unwinder.cc @@ -205,7 +205,13 @@ bool Win32StackFrameUnwinder::TryUnwind(CONTEXT* context) { } else { // We're not at the end of the stack. This frame is untrustworthy and we // can't safely unwind from here. - if (unwind_info_present_for_all_frames_) { + if (!image_base) { + // A null image_base means that the the last unwind produced an invalid + // instruction pointer. This has been observed where unwind information + // was present for a function but was inconsistent with the actual + // function code, in particular in BoringSSL. See + // https://crbug.com/542919. + } else if (unwind_info_present_for_all_frames_) { // Unwind information was present for all previous frames, so we can // be confident this is case 2. Record the module to be blacklisted. LeafUnwindBlacklist::GetInstance()->BlacklistModule( |