Hi,
在 2024/02/20 23:30, Xiao Ni 写道:
MD_RECOVERY_WAIT is used by dmraid to delay reshape process by patch
commit 644e2537fdc7 ("dm raid: fix stripe adding reshape deadlock").
Before patch commit f52f5c71f3d4b ("md: fix stopping sync thread")
dmraid stopped sync thread directy by calling md_reap_sync_thread.
After this patch dmraid stops sync thread asynchronously as md does.
This is right. Now the dmraid stop process is like this:
1. raid_postsuspend->md_stop_writes->__md_stop_writes->stop_sync_thread.
stop_sync_thread sets MD_RECOVERY_INTR and wait until MD_RECOVERY_RUNNING
is cleared
2. md_do_sync finds MD_RECOVERY_WAIT is set and return. (This is the
root cause for this deadlock. We hope md_do_sync can set MD_RECOVERY_DONE)
3. md thread calls md_check_recovery (This is the place to reap sync
thread. Because MD_RECOVERY_DONE is not set. md thread can't reap sync
thread)
4. raid_dtr stops/free struct mddev and release dmraid related resources
dmraid only sets MD_RECOVERY_WAIT but doesn't clear it. It needs to clear
this bit when stopping the dmraid before stopping sync thread.
But the deadlock still can happen sometimes even MD_RECOVERY_WAIT is
cleared before stopping sync thread. It's the reason stop_sync_thread only
wakes up task. If the task isn't running, it still needs to wake up sync
thread too.
This deadlock can be reproduced 100% by these commands:
modprobe brd rd_size=34816 rd_nr=5
while [ 1 ]; do
vgcreate test_vg /dev/ram*
lvcreate --type raid5 -L 16M -n test_lv test_vg
lvconvert -y --stripes 4 /dev/test_vg/test_lv
vgremove test_vg -ff
sleep 1
done
Fixes: 644e2537fdc7 ("dm raid: fix stripe adding reshape deadlock")
Fixes: f52f5c71f3d4 ("md: fix stopping sync thread")
Signed-off-by: Xiao Ni <xni@xxxxxxxxxx>
---
drivers/md/dm-raid.c | 2 ++
drivers/md/md.c | 1 +
2 files changed, 3 insertions(+)
diff --git a/drivers/md/dm-raid.c b/drivers/md/dm-raid.c
index eb009d6bb03a..325767c1140f 100644
--- a/drivers/md/dm-raid.c
+++ b/drivers/md/dm-raid.c
@@ -3796,6 +3796,8 @@ static void raid_postsuspend(struct dm_target *ti)
struct raid_set *rs = ti->private;
if (!test_and_set_bit(RT_FLAG_RS_SUSPENDED, &rs->runtime_flags)) {
+ if (test_bit(MD_RECOVERY_WAIT, &rs->md.recovery))
+ clear_bit(MD_RECOVERY_WAIT, &rs->md.recovery);
Notice that 'MD_RECOVERY_WAIT' will never be cleared, hence sync_thread
will never make progress until table reload for dm-raid.
And other than stopping dm-raid, raid_postsuspend() call also be called
by ioctl to suspend dm-raid, hence this change is wrong.
I think we can never clear 'MD_RECOVERY_FROZEN' in this case so that
'MD_RECOVERY_WAIT' can be removed, or use '!MD_RECOVERY_WAIT' as a
condition to register new sync_thread.
Thanks,
Kuai
/* Writes have to be stopped before suspending to avoid deadlocks. */
if (!test_bit(MD_RECOVERY_FROZEN, &rs->md.recovery))
md_stop_writes(&rs->md);
diff --git a/drivers/md/md.c b/drivers/md/md.c
index 2266358d8074..54790261254d 100644
--- a/drivers/md/md.c
+++ b/drivers/md/md.c
@@ -4904,6 +4904,7 @@ static void stop_sync_thread(struct mddev *mddev, bool locked, bool check_seq)
* never happen
*/
md_wakeup_thread_directly(mddev->sync_thread);
+ md_wakeup_thread(mddev->sync_thread);
if (work_pending(&mddev->sync_work))
flush_work(&mddev->sync_work);
