On Thu, May 23, 2024 at 12:33 AM Logan Gunthorpe <logang@xxxxxxxxxxxx> wrote:
>
>
>
> On 2024-05-22 02:46, Xiao Ni wrote:
> > Hi Logan
> >
> > Thanks for your suggestion. I tried to create signalfd before fork()
> > but it still can't work. And I call sleep(2) before child exits, the
> > problem still can happen sometimes. This is what I tried. If you have
> > time to have a look, it's great. It's not a hurry thing.
>
> Can you send sample prints from this patch? I'm very surprised that with
> the delay on the child processes it can still happen.
There are three changes:
1. initialize signalfd before fork
2. Block SIGCHLD before fork
3. sleep 2 seconds before child exists
4. break when receiving SIGINT signal
>
> What do you do to make it happen? How frequently does it occur?
mdadm -Ss
i=0
while [ 1 ]; do
/usr/sbin/mdadm -CfR /dev/md0 -l 5 -n3 /dev/loop0 /dev/loop1
/dev/loop2 --write-zeroes --auto=yes -v
mdadm --wait /dev/md0
mdadm -Ss
sleep 1
i=$((i+1))
echo $i
done
It's easy to reproduce in my environment. The loop device is 20MB
which is the device size during regression tests.
>
>
> > @@ -185,17 +186,6 @@ static int wait_for_zero_forks(int *zero_pids, int count)
> > if (!wait_count)
> > return 0;
> >
> > - sigemptyset(&sigset);
> > - sigaddset(&sigset, SIGINT);
> > - sigaddset(&sigset, SIGCHLD);
> > - sigprocmask(SIG_BLOCK, &sigset, NULL);
> > -
> > - sfd = signalfd(-1, &sigset, 0);
> > - if (sfd < 0) {
> > - pr_err("Unable to create signalfd: %s\n", strerror(errno));
> > - return 1;
> > - }
>
> Strictly speaking, I don't think it's necessary to move the signalfd
> initialization. Blocking the signals should be enough, then any signals
> can be retrieved at a later time with signalfd. Though, I don't think it
> should hurt to do it this way.
I did a test in a simple c program.
part0: block SIGCHLD and SIGINT
part1: fork 3 child process. It doesn't do anything in the child
process and exits.
part2: read from sfd to check SIGCHLD
part-initsfd: sleep 2 seconds and sfd=signalfd
If I put part-initsfd between part1 and part2, it can happen the same
thing. part2 only sees one SIGCHLD and it's stuck. If part-initsfd is
before part1, it works well. The father process can see the 3 SIGCHLD
signals. So it's the reason it init sfd before fork in my patch.
Best Regards
Xiao
>
> Logan
>
