Occasional segfaults deep in glibc when execing?

[amluto]

I'm seeing various abrt reports that generally look like Fedora bug 1263052, which is the first one I've gotten since updating Fedora's package to fish 2.2.0.

My current best guess is that the iothread_drain_all hack isn't working right.

[amluto]

More info. We're blowing up here in exec_job here:

                            bool builtin_io_done = do_builtin_io(outbuff.data(), outbuff.size(), errbuff.data(), errbuff.size());
                            if (! builtin_io_done)
                            {
                                show_stackframe();
                            }

Something terrible is happening that causes show_stackframe() to segfault, in fprintf, but something must have already gone wrong before that. As a wild guess, if write_loop fails due to EPIPE, bad things will happen.

Also, shouldn't show_stackframe use write_loop instead of fwprintf?

[amluto]

I bet I'm right. The open file descriptors were:

0:/dev/pts/5 (deleted)
pos:    0
flags:  0100002
mnt_id: 22

1:/dev/pts/5 (deleted)
pos:    0
flags:  0100002
mnt_id: 22

2:/dev/pts/5 (deleted)
pos:    0
flags:  0100002
mnt_id: 22

(and some more, stripped for brevity)

[ridiculousfish]

What is suspicious about those open file descriptors? That looks OK to me, except for (deleted) which I don't understand (what do those mean?)

[amluto]

I assume it means that the pty master (terminal) is gone.

[amluto]

See #2393 for a possible fix

[amluto]

I'm submitting Fedora updates-testing builds with these patch, and I'll ask the original reporter to test.

[ridiculousfish]

Assigning to me to remind myself to review this

[amluto]

FWIW, this is now in Fedora's updates-testing, so it should get some test coverage over the next few days.

[faho]

Fixed by merging #2393.