Hi Song,

Song Liu [off-list ref] 於 2023年8月28日 週一 下午1:21寫道：

Hi AceLan,

Thanks for running the experiments.

On Fri, Aug 25, 2023 at 9:32 PM AceLan Kao [off-list ref] wrote:
[...]

quoted

quoted

Could you please run the follow two experiments?

1. Confirm 12a6caf273240a triggers this. Specifically:
   git checkout 12a6caf273240a => repros
   git checkout 12a6caf273240a~1 => cannot repro

Yes, I'm pretty sure about this, that's my bisect result and I just
confirmed it again.
I also tried reverting 12a6caf273240a and the issue is gone.

The log doesn't match my guess. Specifically:

[  420.068142] systemd-shutdown[1]: Stopping MD /dev/md123 (9:123).
[  420.074718] md_open:md123 openers++ = 1 by systemd-shutdow
[  420.080787] systemd-shutdown[1]: Failed to sync MD block device
/dev/md123, ignoring: Input/output error
[  420.090831] md: md123 stopped.
[  420.094465] systemd-shutdown[1]: Stopping MD /dev/md122 (9:122).
[  420.101045] systemd-shutdown[1]: Could not stop MD /dev/md122:
Device or resource busy

For a successful stop on md123, we reach the pr_info() in md_open().
For a failed stop on md122, the kernel returns -EBUSY before that
pr_info() in md_open(). There are some changes in md_open() in
the past few release, so I am not quite sure we are looking at the
same code.

Therefore, could you please help clarify:

1. Which base kernel are you using?

From the log, you are using 6.5-rc7-706a74159504. However,
I think we cannot cleanly revert 12a6caf273240a on top of
6.5-rc7-706a74159504. Did you manually fix some issue in the
revert? If so, could you please share the revert commit?

Yes, I'm basing 6.5.0-rc7 706a74159504 to apply your patch
706a74159504 Linux 6.5-rc7

Attached how I revert the commit.

2. If you are not using 6.5-rc7-706a74159504 as base kernel, which
one are you using?

Thanks,
Song

quoted

quoted

2. Try with the following change (add debug messages), which hopefully
   shows which command is holding a reference on mddev->openers.

Thanks,
Song

diff --git i/drivers/md/md.c w/drivers/md/md.c
index 78be7811a89f..3e9b718b32c1 100644
--- i/drivers/md/md.c
+++ w/drivers/md/md.c

@@ -7574,11 +7574,15 @@ static int md_ioctl(struct block_device *bdev,

blk_mode_t mode,
                if (mddev->pers && atomic_read(&mddev->openers) > 1) {
                        mutex_unlock(&mddev->open_mutex);
                        err = -EBUSY;
+                       pr_warn("%s return -EBUSY for %s with
mddev->openers = %d\n",
+                               __func__, mdname(mddev),
atomic_read(&mddev->openers));
                        goto out;
                }
                if (test_and_set_bit(MD_CLOSING, &mddev->flags)) {
                        mutex_unlock(&mddev->open_mutex);
                        err = -EBUSY;
+                       pr_warn("%s return -EBUSY for %s with
MD_CLOSING bit set\n",
+                               __func__, mdname(mddev));
                        goto out;
                }
                did_set_md_closing = true;

@@ -7789,6 +7793,8 @@ static int md_open(struct gendisk *disk, blk_mode_t mode)
                goto out_unlock;

        atomic_inc(&mddev->openers);
+       pr_info("%s:%s openers++ = %d by %s\n", __func__, mdname(mddev),
+               atomic_read(&mddev->openers), current->comm);
        mutex_unlock(&mddev->open_mutex);

        disk_check_media_change(disk);

@@ -7807,6 +7813,8 @@ static void md_release(struct gendisk *disk)

        BUG_ON(!mddev);
        atomic_dec(&mddev->openers);
+       pr_info("%s:%s openers-- = %d by %s\n", __func__, mdname(mddev),
+               atomic_read(&mddev->openers), current->comm);
        mddev_put(mddev);
 }

It's pretty strange that I can't reproduce the issue after applied the patch.

I tried to figure out which part affect the issue and found when I
comment out the pr_info() In md_release(), the issue could be
reproduced.

--
Chia-Lin Kao(AceLan)
http://blog.acelan.idv.tw/
E-Mail: acelan.kaoATcanonical.com (s/AT/@/)

-- 
Chia-Lin Kao(AceLan)
http://blog.acelan.idv.tw/
E-Mail: acelan.kaoATcanonical.com (s/AT/@/)