Linux Crash Dumping

Official builds of Chrome support crash dumping and reporting using the Google crash servers. This is a guide to how this works.

Breakpad

Breakpad is an open source library which we use for crash reporting across all three platforms (Linux, Mac and Windows). For Linux, a substantial amount of work was required to support cross-process dumping. At the time of writing this code is currently forked from the upstream breakpad repo. While this situation remains, the forked code lives in third_party/breakpad/linux. The upstream repo is mirrored in third_party/breakpad/breakpad.

The code currently supports i386 only. Getting x86-64 to work should only be a minor amount of work.

Minidumps

Breakpad deals in a file format called ‘minidumps’. This is a Microsoft format and thus is defined by in-memory structures which are dumped, raw, to disk. The main header file for this file format is third_party/breakpad/breakpad/src/google_breakpad/common/minidump_format.h.

At the top level, the minidump file format is a list of key-value pairs. Many of the keys are defined by the minidump format and contain cross-platform representations of stacks, threads etc. For Linux we also define a number of custom keys containing /proc/cpuinfo, lsb-release etc. These are defined in third_party/breakpad/breakpad/linux/minidump_format_linux.h.

Catching exceptions

Exceptional conditions (such as invalid memory references, floating point exceptions, etc) are signaled by synchronous signals to the thread which caused them. Synchronous signals are always run on the thread which triggered them as opposed to asynchronous signals which can be handled by any thread in a thread-group which hasn't masked that signal.

All the signals that we wish to catch are synchronous except SIGABRT, and we can always arrange to send SIGABRT to a specific thread. Thus, we find the crashing thread by looking at the current thread in the signal handler.

The signal handlers run on a pre-allocated stack in case the crash was triggered by a stack overflow.

Once we have started handling the signal, we have to assume that the address space is compromised. In order not to fall prey to this and crash (again) in the crash handler, we observe some rules:

  1. We don‘t enter the dynamic linker. This, observably, can trigger crashes in the crash handler. Unfortunately, entering the dynamic linker is very easy and can be triggered by calling a function from a shared library who’s resolution hasn‘t been cached yet. Since we can’t know which functions have been cached we avoid calling any of these functions with one exception: memcpy. Since the compiler can emit calls to memcpy we can't really avoid it.
  2. We don't allocate memory via malloc as the heap may be corrupt. Instead we use a custom allocator (in breadpad/linux/memory.h) which gets clean pages directly from the kernel.

In order to avoid calling into libc we have a couple of header files which wrap the system calls (linux_syscall_support.h) and reimplement a tiny subset of libc (linux_libc_support.h).

Self dumping

The simple case occurs when the browser process crashes. Here we catch the signal and clone a new process to perform the dumping. We have to use a new process because a process cannot ptrace itself.

The dumping process then ptrace attaches to all the threads in the crashed process and writes out a minidump to /tmp. This is generic breakpad code.

Then we reach the Chrome specific parts in chrome/app/breakpad_linux.cc. Here we construct another temporary file and write a MIME wrapping of the crash dump ready for uploading. We then fork off wget to upload the file. Based on Debian popcorn, wget is very commonly installed (much more so than libcurl) and wget handles the HTTPS gubbins for us.

Renderer dumping

In the case of a crash in the renderer, we don't want the renderer handling the crash dumping itself. In the future we will sandbox the renderer and allowing it the authority to crash dump itself is too much.

Thus, we split the crash dumping in two parts: the gathering of information which is done in process and the external dumping which is done out of process. In the case above, the latter half was done in a cloned child. In this case, the browser process handles it.

When renderers are forked off, they have a UNIX DGRAM socket in file descriptor 4. The signal handler then calls into Chrome specific code (chrome/renderer/render_crash_handler_linux.cc) when it would otherwise clone. The Chrome specific code sends a datagram to the socket which contains:

  • Information which is only available to the signal handler (such as the ucontext structure).
  • A file descriptor to a pipe which it then blocks on reading from.
  • A CREDENTIALS structure giving its PID.

The kernel enforces that the renderer isn‘t lying in the CREDENTIALS structure so it can’t ask the browser to crash dump another process.

The browser then performs the ptrace and minidump writing which would otherwise be performed in the cloned process and does the MIME wrapping the uploading as normal.

Once the browser has finished getting information from the crashed renderer via ptrace, it writes a byte to the file descriptor which was passed from the renderer. The renderer than wakes up (because it was blocking on reading from the other end) and rethrows the signal to itself. It then appears to crash ‘normally’ and other parts of the browser notice the abnormal termination and display the sad tab.

How to test Breakpad support in Chromium

  • Build Chromium as normal.

  • Run the browser with the environment variable CHROME_HEADLESS=1. This enables crash dumping but prevents crash dumps from being uploaded and deleted.

    env CHROME_HEADLESS=1 ./out/Debug/chrome-wrapper
    
  • Visit the special URL chrome://crash to trigger a crash in the renderer process.

  • A crash dump file should appear in the directory ~/.config/chromium/Crash Reports.