Is coredump_wait broken in 6.4-rc5?

[Matthew Wilcox <willy@infradead.org>]

Have you seen anything like this?  It could be one of my patches that
broke it:

00135 fio[39217]: segfault at 560494fe654c ip 00007fe9fb6b3e67 sp 00007ffc8f08b6d0 error 4 in libc.so.6[7fe9fb645000+155000] likely on CPU 0 (core 0, socket 0)
00135 Code: 1f 40 00 48 83 ec 08 48 8b 4f 08 48 89 c8 48 83 e0 f8 48 3b 04 07 0f 85 a9 00 00 00 f3 0f 6f 47 10 48 8b 57 18 66 48 0f 7e c0 <48> 3b 78 18 75 7b 48 3b 7a 10 75 75 48 8b 77 10 48 89 50 18 66 0f
00363 INFO: task fsstress:39216 blocked for more than 120 seconds.
00363       Not tainted 6.4.0-rc5-00028-gf824d48d50fc #391
00363 "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
00363 task:fsstress        state:D stack:11496 pid:39216 ppid:39214  flags:0x00000002
00363 Call Trace:
[snip]
00363  wait_for_completion_state+0x11c/0x1e0
00363  do_coredump+0x265/0x15b0
00363  get_signal+0x904/0x9c0
00363  arch_do_signal_or_restart+0x1b/0x250
00363  exit_to_user_mode_prepare+0x99/0x100
00363  syscall_exit_to_user_mode+0x21/0x40

# cat /proc/39216/stack
[<0>] do_coredump+0x265/0x15b0
[<0>] get_signal+0x904/0x9c0
[<0>] arch_do_signal_or_restart+0x1b/0x250
[<0>] exit_to_user_mode_prepare+0x99/0x100
[<0>] syscall_exit_to_user_mode+0x21/0x40
[<0>] do_syscall_64+0x40/0x80
[<0>] entry_SYSCALL_64_after_hwframe+0x63/0xcd

# cat /proc/39214/stack
[<0>] do_wait+0x161/0x2f0
[<0>] kernel_wait4+0x8a/0x110
[<0>] __do_sys_wait4+0x67/0x80
[<0>] __x64_sys_wait4+0x17/0x20
[<0>] do_syscall_64+0x34/0x80
[<0>] entry_SYSCALL_64_after_hwframe+0x63/0xcd

Those are the only two pids running fsstress.

# ps -aux | grep fsstress
root       39214  0.0  0.0   2580  1664 ?        S    15:11   0:00 ./ltp/fsstress -p 1 -n999999999 -f setattr 0 -ffsync 0 -fsync 0 -ffdatasync 0 -f setattr 1 -d /mnt/scratch/fsstress.39009
root       39216  0.3  0.3   8024  6792 ?        D    15:11   0:02 ./ltp/fsstress -p 1 -n999999999 -f setattr 0 -ffsync 0 -fsync 0 -ffdatasync 0 -f setattr 1 -d /mnt/scratch/fsstress.39009

Obviously fio shouldn't've segfaulted in the first place, but sometimes
when it does, fsstress hangs during coredump, and that also seems like
it shouldn't happen.