GitLab

There is a recurring issue where bugs in browser process code introduce
deserialization errors on the main viz browser-to-GPU process message
pipe. The message pipe is closed and the GPU process crashes trying to
reinitialize the viz thread. Without a way to differentiate message pipe
closure due to a deserialization error vs the other side closing it, the
crash only happens when the browser tries to reinitialize the viz thread
after what it thinks is a GPU crash. There is a very little information
available in these crash reports, the message pipe is shared by multiple
interfaces and figuring out what IPC message the error occurred in is a
guessing game.

There is a global error handler that triggers for any mojo
deserialization errors which includes a string containing the interface
name and a message identifier. This CL adds a handler which stores the
error string. If the GPU process crashes due to reinitializing the viz
thread within 5 seconds after the handler fired, it populates a crash
key with that error message. While it's not guaranteed the last
deserialization error was what caused the viz thread to get
reinitialized, deserialization errors are infrequent enough it shouldn't
be an issue.

Bug: 1075495
Change-Id: I3a762fd5e9d50c0c1c301b3179f08ce892ff4912
Reviewed-on: https://chromium-review.googlesource.com/c/chromium/src/+/2356546
Commit-Queue: kylechar <kylechar@chromium.org>
Reviewed-by: Ken Rockot <rockot@google.com>
Cr-Commit-Position: refs/heads/master@{#798645}