fix races in CrashGenerator::CreateChildCrash

The current CreateChildCrash logic is racy when it comes to creating a
crash dump for two reasons:

The main thread that calls kill() on a different thread is guaranteed
the signal will be *queued* when it returns, but not *delivered*.  If
the kernel doesn't automatically schedule the receiving thread, but
instead lets the main thread run to the exit() call, then the signal
never triggers a coredump and the whole process simply exits.

The main thread is using kill() to try to deliver a signal to a
specific thread, but that function is for sending signals to a
process.  That means the kernel is free to deliver the signal to
any thread in the process and not just the one requested.  This
manifests itself as the pr_pid in the coredump not being the one
expected.  Instead, we must use tkill() with the tid (which we
already took care of gathering) to deliver to a specific thread.

These are a lot easier to see on a UMP system as contention is heavier.

BUG=chromium:207918
TEST=`dumper_unittest` still passes, and doesn't flake out in a UMP system
TEST=`linux_client_unittest` still passes
R=benchan@chromium.org

Review URL: https://breakpad.appspot.com/1304005

git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@1299 4c0a9323-5329-0410-9bdc-e9ce6186880e

0 files changed, 0 insertions, 0 deletions