Hi,
在 2024/01/18 2:19, Mikulas Patocka 写道:
stop_sync_thread sets MD_RECOVERY_INTR and then waits for
MD_RECOVERY_RUNNING to be cleared. However, md_do_sync will not clear
MD_RECOVERY_RUNNING when exiting, it will set MD_RECOVERY_DONE instead.
So, we must wait for MD_RECOVERY_DONE to be set as well.
You are missing how sync_thread() is stopped. md_do_sync() will set
MD_RECOVERY_DONE first, and md_check_recovery() is responsible for
clearing MD_RECOVERY_RUNNING.
This patch fixes a deadlock in the LVM2 test shell/integrity-caching.sh.
It's still not clear about the root cause. Any way, this patch is really
wrong, it breaks the foundation design of sync_thread.
Thanks,
Kuai
sysrq: Show Blocked State
task:lvm state:D stack:0 pid:11422 tgid:11422 ppid:1374 flags:0x00004002
Call Trace:
<TASK>
__schedule+0x228/0x570
schedule+0x29/0xa0
schedule_timeout+0x6a/0xd0
? timer_shutdown_sync+0x10/0x10
stop_sync_thread+0x141/0x180 [md_mod]
? housekeeping_test_cpu+0x30/0x30
__md_stop_writes+0x10/0xd0 [md_mod]
md_stop+0x9/0x20 [md_mod]
raid_dtr+0x1e/0x60 [dm_raid]
dm_table_destroy+0x53/0x110 [dm_mod]
__dm_destroy+0x10b/0x1e0 [dm_mod]
? table_clear+0xa0/0xa0 [dm_mod]
dev_remove+0xd4/0x110 [dm_mod]
ctl_ioctl+0x2e1/0x570 [dm_mod]
dm_ctl_ioctl+0x5/0x10 [dm_mod]
__x64_sys_ioctl+0x85/0xa0
do_syscall_64+0x5d/0x1a0
entry_SYSCALL_64_after_hwframe+0x46/0x4e
Signed-off-by: Mikulas Patocka <mpatocka@xxxxxxxxxx>
Cc: stable@xxxxxxxxxxxxxxx # v6.7
Fixes: 130443d60b1b ("md: refactor idle/frozen_sync_thread() to fix deadlock")
---
drivers/md/md.c | 4 +++-
1 file changed, 3 insertions(+), 1 deletion(-)
Index: linux-2.6/drivers/md/md.c
===================================================================
--- linux-2.6.orig/drivers/md/md.c
+++ linux-2.6/drivers/md/md.c
@@ -4881,7 +4881,8 @@ static void stop_sync_thread(struct mdde
if (check_seq)
sync_seq = atomic_read(&mddev->sync_seq);
- if (!test_bit(MD_RECOVERY_RUNNING, &mddev->recovery)) {
+ if (!test_bit(MD_RECOVERY_RUNNING, &mddev->recovery) ||
+ test_bit(MD_RECOVERY_DONE, &mddev->recovery)) {
if (!locked)
mddev_unlock(mddev);
return;
@@ -4901,6 +4902,7 @@ retry:
if (!wait_event_timeout(resync_wait,
!test_bit(MD_RECOVERY_RUNNING, &mddev->recovery) ||
+ test_bit(MD_RECOVERY_DONE, &mddev->recovery) ||
(check_seq && sync_seq != atomic_read(&mddev->sync_seq)),
HZ / 10))
goto retry;
.
