Currently, Breakpad relies on facilities that are disallowed inside the Linux seccomp sandbox. Specifically, it sets a signal handler to catch faults (currently disallowed), forks a new process, and uses ptrace() (also disallowed) to read the memory of the faulted process.
There are three ways we could do crash dumping of seccomp-sandboxed processes:
- Find a way to permit signal handling safely inside the sandbox (see below).
- Allow the kernel's core dumper to kick in and write a core file.
- This seems risky because this code tends not to be well-tested.
- This will not work if the process is chrooted, so it would not work if the seccomp sandbox is stacked with the SUID sandbox.
- Have an unsandboxed helper process which
ptrace()s the sandboxed process to catch faults.
In case a trusted thread faults with a SIGSEGV, we must make sure that an untrusted thread cannot register a signal handler that will run in the context of the trusted thread.
Here are some mechanisms that could make this safe:
sigaltstack()is per-thread. If we opt not to set a signal stack for trusted threads, and set %esp/%rsp to an invalid address, trusted threads will die safely if they fault.- This means the trusted thread cannot set a signal stack on behalf of the untrusted thread once the latter has switched to seccomp mode. The signal stack would have to be set up when the thread is created and not subsequently changed.
clone()has aCLONE_SIGHANDflag. By omitting this flag, trusted and untrusted threads can have different sets of signal handlers. This means we can opt not to set signal handlers for trusted threads.- Again, per-thread signal handler sets would mean the trusted thread cannot change signal handlers on behalf of untrusted threads.
sigprocmask()/pthread_sigmask(): These can be used to block signal handling in trusted threads.