[RESEND] md: adding a new flag MD_DELETING

The mddev data structure is freed in mddev_delayed_delete(), which is
schedualed after the array is deconfigured completely when stopping. So
there is a race window between md_open() and do_md_stop(), which leads
to /dev/mdX can still be opened by userspace even it's not accessible
any more. As a result, a DeviceDisappeared event will not be able to be
monitored by mdadm in monitor mode. This patch tries to fix it by adding
this new flag MD_DELETING.

Signed-off-by: Lidong Zhong <redacted>
---
 drivers/md/md.c | 4 +++-
 drivers/md/md.h | 2 ++
 2 files changed, 5 insertions(+), 1 deletion(-)

diff --git a/drivers/md/md.c b/drivers/md/md.c
index 49f897fbb89b..e8fa100a0777 100644
--- a/drivers/md/md.c
+++ b/drivers/md/md.c

@@ -6456,6 +6456,7 @@ static int do_md_stop(struct mddev *mddev, int mode,
 		md_clean(mddev);
 		if (mddev->hold_active == UNTIL_STOP)
 			mddev->hold_active = 0;
+		set_bit(MD_DELETING, &mddev->flags);
 	}
 	md_new_event(mddev);
 	sysfs_notify_dirent_safe(mddev->sysfs_state);

@@ -7843,7 +7844,8 @@ static int md_open(struct block_device *bdev, fmode_t mode)
 	if ((err = mutex_lock_interruptible(&mddev->open_mutex)))
 		goto out;
 
-	if (test_bit(MD_CLOSING, &mddev->flags)) {
+	if (test_bit(MD_CLOSING, &mddev->flags) ||
+	    (test_bit(MD_DELETING, &mddev->flags) && mddev->pers == NULL)) {
 		mutex_unlock(&mddev->open_mutex);
 		err = -ENODEV;
 		goto out;

diff --git a/drivers/md/md.h b/drivers/md/md.h
index fb7eab58cfd5..b70f8885da7a 100644
--- a/drivers/md/md.h
+++ b/drivers/md/md.h

@@ -262,6 +262,8 @@ enum mddev_flags {
 	MD_BROKEN,              /* This is used in RAID-0/LINEAR only, to stop
 				 * I/O in case an array member is gone/failed.
 				 */
+	MD_DELETING,		/* If set, we are deleting the array, do not open
+				 * it then */
 };
 
 enum mddev_sb_flags {

-- 
2.26.2

Hi Lidong,

On Sat, Jun 5, 2021 at 8:29 AM Lidong Zhong [off-list ref] wrote:

The mddev data structure is freed in mddev_delayed_delete(), which is
schedualed after the array is deconfigured completely when stopping. So
there is a race window between md_open() and do_md_stop(), which leads
to /dev/mdX can still be opened by userspace even it's not accessible
any more. As a result, a DeviceDisappeared event will not be able to be
monitored by mdadm in monitor mode. This patch tries to fix it by adding
this new flag MD_DELETING.

Signed-off-by: Lidong Zhong <redacted>

The patch still contains a lot of bad format strings. How did you send
it? I guess
it is not from git-send-email?

Thanks,
Song

Dear Lidong,


Am 05.06.21 um 17:29 schrieb Lidong Zhong:

I’d use imperative mood in the commit message summary [1]: md:

Add new flag MD_DELETING

The mddev data structure is freed in mddev_delayed_delete(), which is
schedualed after the array is deconfigured completely when stopping. So

scheduled

Kind regards,

Paul


[1]: https://chris.beams.io/posts/git-commit/

On 6/9/21 1:29 PM, Song Liu wrote:

Hi Lidong,

On Sat, Jun 5, 2021 at 8:29 AM Lidong Zhong [off-list ref] wrote:

quoted

The mddev data structure is freed in mddev_delayed_delete(), which is
schedualed after the array is deconfigured completely when stopping. So
there is a race window between md_open() and do_md_stop(), which leads
to /dev/mdX can still be opened by userspace even it's not accessible
any more. As a result, a DeviceDisappeared event will not be able to be
monitored by mdadm in monitor mode. This patch tries to fix it by adding
this new flag MD_DELETING.

Signed-off-by: Lidong Zhong <redacted>

The patch still contains a lot of bad format strings. How did you send
it? I guess
it is not from git-send-email?

Hi Song,

Sorry it still doesn't work, but I did send it from git-send-email. I'll
send the patch after I figure out the reason.

Regards,
Lidong