Child hangs if parent crashes or exits in google_breakpad::ExceptionHandler::SignalHandler

Question

This happens if parent crashes after cloning child process, but before sending the unblocking byte with SendContinueSignalToChild(). In this case pipe file handle remains opened and child stays infinitely blocked on read(...) within WaitForContinueSignal(). After the crash, child is adopted by init process.

Steps to reproduce:

l. Simulate parent crash in google_breakpad::ExceptionHandler::GenerateDump(CrashContext *context):

...
const pid_t child = sys_clone(
ThreadEntry, stack, CLONE_FILES | CLONE_FS | CLONE_UNTRACED, &thread_arg, NULL, NULL, NULL);

int r, status;
// Allow the child to ptrace us
sys_prctl(PR_SET_PTRACER, child, 0, 0, 0);


int *ptr = 0;  
*ptr = 42;     // <------- Crash here


SendContinueSignalToChild();
...

2. Send one of the handled signal to the parent (e.g. SIGSEGV), so that the above GenerateDump(...) method is envoked.
3. Observe that parent exits but child still exists, blocked on WaitForContinueSignal().

Output for the above steps:

dmytro@db:~$ ./test &
[1] 25050
dmytro@db:~$ Test: started

dmytro@db:~$ ps aflxw | grep test
0  1000 25050 18923  20   0  40712  2680 -      R    pts/37     0:13          |   |           \_ ./test
0  1000 25054 18923  20   0   6136   856 pipe_w S+   pts/37     0:00          |   |           \_ grep --color=auto test
dmytro@db:~$ kill -11 25050

[1]+  Segmentation fault      (core dumped) ./test
dmytro@db:~$ ps aflxw | grep test
0  1000 25058 18923  20   0   6136   852 pipe_w S+   pts/37     0:00          |   |           \_ grep --color=auto test
1  1000 25055  1687  20   0  40732   356 pipe_w S    pts/37     0:00          \_ ./test

1687 is the init pid.

In the real world the crash happens in a thread parallel to the one that handles signal. NOTE: the issue can also happen because of normal program termination (i.e. exit(0) is called in a parallel thread).

Tested on Linux 3.3.8-2.2., mips and i686 platforms.

So, my 2 questions:

1. Is it the expected behavior for the breakpad library to keep child alive? My expectation is that child should exit immediately after parent crashes/exits.
2. If it is not expected behavior, what is the best solution to finish client after parent crash/exit?

Thanks in advance!

Answer 1

Any clue on possible solution?

This can also happen during shutdown crash, if the crashed thread is not main, and the parent process exits from main() in exactly this time slot, so apparently it's not that unlikely to happen as it seems at a first glance.

At this moment, I think this is happening because of CLONE_FILES flag of clone() function. This leads to the situation where read() on pipe in child is not returning EOF if parent process quits.

I have not yet done the examination if we can safely get rid of this flag in clone() call.