Re: [RFC PATCH] pidfs: don't report autoreaped tasks readable before exit info exists

[fujo]

Hi Oleg,

thanks, and sorry for the unclear report.  You are right that the early
wakeup is not specific to CLONE_AUTOREAP: exit_notify() always calls
do_notify_pidfd() for an exiting non-delay_group_leader, so pollers are
woken and pidfd_poll() returns EPOLLIN as soon as exit_state is set.

What I observed is the window between that first wakeup and
release_task() -> pidfs_exit() publishing PIDFS_ATTR_BIT_EXIT: poll()
returns EPOLLIN to userspace at that point (it does not wait for the
later wake_up_all() after the unhash), so a parent that follows the
"poll and then PIDFD_GET_INFO" sequence from the series description can
read a mask without PIDFD_INFO_EXIT.  The second wakeup only helps
pollers that are still asleep when it happens.

Test-case details, as requested:

- clone3(CLONE_PIDFD | CLONE_AUTOREAP, exit_signal = 0, pidfd = &pidfd)
- mode 0: the child creates one thread which exits after 500us and then
  exits itself with 42; mode 1: the child exits immediately
- parent: poll(pidfd, POLLIN, 5000), then
  ioctl(pidfd, PIDFD_GET_INFO, {.mask = PIDFD_INFO_EXIT}) and counts
  results whose mask lacks PIDFD_INFO_EXIT
- x86_64 mainline 58161ead29e5 (v7.3-rc7), defconfig + KASAN + DEBUG_INFO
  + DEBUG_KMEMLEAK, QEMU, 2G:
    mode 0, SMP 2: 18/300
    mode 0, SMP 1: 78/300
    mode 1, SMP 2:  8/300
- the in-tree pidfd_autoreap_test global.autoreap_multithreaded case
  fails the same way: its poll() gets POLLIN and PIDFD_GET_INFO returns
  mask 0x7 (PID|CREDS|CGROUPID) without EXIT.

Your POLLHUP remark resolves it: PIDFD_INFO_EXIT is published before the
task linkage is removed, so POLLHUP is the right condition to wait for.
I changed the autoreap tests to pfd.events = POLLHUP and to check
POLLHUP in revents; on the *unpatched* kernel they now pass ten
consecutive runs (19/19 each), and a POLLHUP variant of the reproducer
gets 300/300 with PIDFD_INFO_EXIT.  So the kernel behaves as designed
and the tests were wrong.

I will withdraw the kernel patch and send a selftests-only v2.  Do you
think the CLONE_AUTOREAP cover letter or some KAPI comment should also
say explicitly that POLLHUP, not POLLIN, guarantees that
PIDFD_INFO_EXIT is available?  The "poll the pidfd and retrieve the exit
status via PIDFD_GET_INFO" wording is what led me there.

Thanks,
fujo