This is a repeat of a prior unresolved situation. During the full regression
suite, subtest t5562.8 (fetch gzipped empty) appears to have hung on the
NonStop port. The code was in a loop repeatedly calling waitpid here in
git-http-backend:
waitpid + 0x320 (SLr)
$n_EnterPriv + 0x280 (Milli)
cleanup_children + 0x250 (UCr)
cleanup_children_on_exit + 0x30 (UCr)
git_atexit_dispatch + 0xB0 (UCr)
__process_atexit_functions + 0xA0 (DLL zcredll)
CRE_TERMINATOR_ + 0xB50 (DLL zcredll)
exit + 0x2A0 (DLL zcrtldll)
die_webcgi + 0xA0 (UCr)
die_errno + 0x150 (UCr)
write_or_die + 0xC0 (UCr)
end_headers + 0x70 (UCr)
die_webcgi + 0x160 (UCr)
die + 0x120 (UCr)
inflate_request + 0x3D0 (UCr)
run_service + 0x540 (UCr)
service_rpc + 0x1F0 (UCr)
cmd_main + 0x3B0 (UCr)
main + 0x120 (UCr)
_MAIN + 0x80 (UCr)
I do not see any relevant diff between rc0 and rc1 that could account for
this. I also experienced this hang during 2.28.0 testing and in earlier
releases transiently. 2.21.0 had this hang also. I have noticed a number of
reports of this subtest hanging on different platforms. It looks like the
situation is here:
while (waitpid(p->pid, NULL, 0) < 0 && errno == EINTR)
; /* spin waiting for process exit or error */
It is possible that, without kernel threads, the child never terminates (or
is waiting on a deferred signal that is not being delivered). If both git
and git-http-backend are in cleanup_children at the same time, it is
possible that neither will terminate, if git-http-backend has a child that
is waiting on something else - apparent from this trace.
Sincerely,
Randall
-- Brief whoami:
NonStop developer since approximately 211288444200000000
UNIX developer since approximately 421664400
-- In my real life, I talk too much.
