Hi Xiao,
I don't have time to dig into this myself, but my guess would be that
the signal for one of the children come too quickly, before the
sigprocmask() call in wait_for_zero_forks().
Seems like SIGCHLD should be blocked before the first call to
write_zeroes_fork(). I'm really not sure why I put in a block to SIGINT
and then a block to SIGCHLD after the processes started. I suspect
adding SIGCHLD to the sigprocmask in add_disks() and just removing the
sigprocmask in write_zeroes_fork() might fix the issue.
Thanks,
Logan
On 2024-05-21 01:05, Xiao Ni wrote:
> Hi Logan
>
> I'm trying to fix errors of mdadm regression failures. There is a
> failure in 00raid5-zero sometimes. I added some logs:
>
> In function write_zeroes_fork:
> if (fallocate(fd, FALLOC_FL_ZERO_RANGE | FALLOC_FL_KEEP_SIZE,
> offset_bytes, sz)) {
> pr_err("zeroing %s failed: %s\n", dv->devname,
> strerror(errno));
> ret = 1;
> break;
> } else
> printf("zeroing good\n");
>
> In function wait_for_zero_forks:
> if (fdsi.ssi_signo == SIGINT) {
> printf("\n");
> pr_info("Interrupting zeroing processes,
> please wait...\n");
> interrupted = true;
> break;
> } else if (fdsi.ssi_signo == SIGCHLD) {
> printf("one child finishes, wait count %d\n",
> wait_count);
> if (!--wait_count)
> break;
> }
>
> 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
> done
>
> zeroing good
> zeroing good
> zeroing good
> one child finishes, wait count 3
> one child finishes, wait count 2
>
> It looks like the farther process misses one child signal.
>
> root 174247 0.0 0.0 3628 2552 pts/0 S+ 02:52 0:00 |
> \_ /usr/sbin/mdadm -CfR /dev/md0 -l 5 -n3 /dev/loop0
> /dev/loop1 /dev/loop2 --write-zeroes --auto=yes -v
> root 174248 0.0 0.0 0 0 pts/0 Z+ 02:52 0:00 |
> \_ [mdadm] <defunct>
> root 174249 0.0 0.0 0 0 pts/0 Z+ 02:52 0:00 |
> \_ [mdadm] <defunct>
> root 174250 0.0 0.0 0 0 pts/0 Z+ 02:52 0:00 |
> \_ [mdadm] <defunct>
>
> ]# cat /proc/174247/stack
> [<0>] signalfd_dequeue+0x14d/0x170
> [<0>] signalfd_read_iter+0x7b/0xd0
> [<0>] vfs_read+0x201/0x330
> [<0>] ksys_read+0x5f/0xe0
> [<0>] do_syscall_64+0x7b/0x160
> [<0>] entry_SYSCALL_64_after_hwframe+0x76/0x7e
>
> Any ideas for this?
>
> Best Regards
> Xiao
>
