Skip to contentSkip to conversation

md/md-cluster: handle REMOVE message earlier

PATCH Applied Heming Zhao 3 participants 3 messages Jul 28, 2025 – Jul 30, 2025 MD RAID

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 md/md-cluster: handle REMOVE message earlier

Commit a1fd37f97808 ("md: Don't wait for MD_RECOVERY_NEEDED for HOT_REMOVE_DISK ioctl") introduced a regression in the md_cluster module. (Failed cases 02r1_Manage_re-add & 02r10_Manage_re-add)

Consider a 2-node cluster:
- node1 set faulty & remove command on a disk.
- node2 must correctly update the array metadata.

Before a1fd37f97808, on node1, the delay between msg:METADATA_UPDATED (triggered by faulty) and msg:REMOVE was sufficient for node2 to reload the disk info (written by node1).
After a1fd37f97808, node1 no longer waits between faulty and remove, causing it to send msg:REMOVE while node2 is still reloading disk info. This often results in node2 failing to remove the faulty disk.

== how to trigger ==

set up a 2-node cluster (node1 & node2) with disks vdc & vdd.

on node1:
mdadm -CR /dev/md0 -l1 -b clustered -n2 /dev/vdc /dev/vdd --assume-clean ssh node2-ip mdadm -A /dev/md0 /dev/vdc /dev/vdd
mdadm --manage /dev/md0 --fail /dev/vdc --remove /dev/vdc

check array status on both nodes with "mdadm -D /dev/md0". node1 output:
Number Major Minor RaidDevice State
- 0 0 0 removed
1 254 48 1 active sync /dev/vdd
node2 output:
Number Major Minor RaidDevice State
- 0 0 0 removed
1 254 48 1 active sync /dev/vdd

0 254 32 - faulty /dev/vdc

Fixes a1fd37f97808 ("md: Don't wait for MD_RECOVERY_NEEDED for HOT_REMOVE_DISK ioctl")
---
 drivers/md/md.c | 9 ++++++---
 1 file changed, 6 insertions(+), 3 deletions(-)
drivers/md/md.c +6 -3
index 0f03b21e66e4..de9f9a345ed3 100644
@@ -9758,8 +9758,8 @@ void md_check_recovery(struct mddev *mddev)
 			 * remove disk.
 			 */
 			rdev_for_each_safe(rdev, tmp, mddev) {
-				if (test_and_clear_bit(ClusterRemove, &rdev->flags) &&
-						rdev->raid_disk < 0)
+				if (rdev->raid_disk < 0 &&
+				    test_and_clear_bit(ClusterRemove, &rdev->flags))
 					md_kick_rdev_from_array(rdev);
 			}
 		}
@@ -10065,8 +10065,11 @@ static void check_sb_changes(struct mddev *mddev, struct md_rdev *rdev)
 
 	/* Check for change of roles in the active devices */
 	rdev_for_each_safe(rdev2, tmp, mddev) {
-		if (test_bit(Faulty, &rdev2->flags))
+		if (test_bit(Faulty, &rdev2->flags)) {
+			if (test_bit(ClusterRemove, &rdev2->flags))
+				set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
 			continue;
+		}
 
 		/* Check if the roles changed */
 		role = le16_to_cpu(sb->dev_roles[rdev2->desc_nr]);
Signature · 1 line
2.43.0
Reviewed-by on v1 1/1
#1 lore raw
Quoted 55 lines · Heming Zhao

Commit a1fd37f97808 ("md: Don't wait for MD_RECOVERY_NEEDED for HOT_REMOVE_DISK ioctl") introduced a regression in the md_cluster module. (Failed cases 02r1_Manage_re-add & 02r10_Manage_re-add)

Consider a 2-node cluster:
- node1 set faulty & remove command on a disk.
- node2 must correctly update the array metadata.

Before a1fd37f97808, on node1, the delay between msg:METADATA_UPDATED (triggered by faulty) and msg:REMOVE was sufficient for node2 to reload the disk info (written by node1).
After a1fd37f97808, node1 no longer waits between faulty and remove, causing it to send msg:REMOVE while node2 is still reloading disk info. This often results in node2 failing to remove the faulty disk.

== how to trigger ==

set up a 2-node cluster (node1 & node2) with disks vdc & vdd.

on node1:
mdadm -CR /dev/md0 -l1 -b clustered -n2 /dev/vdc /dev/vdd --assume-clean
ssh node2-ip mdadm -A /dev/md0 /dev/vdc /dev/vdd mdadm --manage /dev/md0 --fail /dev/vdc --remove /dev/vdc

check array status on both nodes with "mdadm -D /dev/md0". node1 output:
Number Major Minor RaidDevice State
- 0 0 0 removed
1 254 48 1 active sync /dev/vdd
node2 output:
Number Major Minor RaidDevice State
- 0 0 0 removed
1 254 48 1 active sync /dev/vdd

0 254 32 - faulty /dev/vdc

Fixes a1fd37f97808 ("md: Don't wait for MD_RECOVERY_NEEDED for HOT_REMOVE_DISK ioctl")
---
 drivers/md/md.c | 9 ++++++---
 1 file changed, 6 insertions(+), 3 deletions(-)
drivers/md/md.c +0 -1
index 0f03b21e66e4..de9f9a345ed3 100644
@@ -9758,8 +9758,8 @@ void md_check_recovery(struct mddev 
 *mddev)
 			 * remove disk.
 			 */
 			rdev_for_each_safe(rdev, tmp, mddev) {
-				if (test_and_clear_bit(ClusterRemove, 

&rdev->flags) &&

drivers/md/md.c void md_check_recovery(struct mddev *mddev) · line 9762

Quoted 1 line
drivers/md/md.c +0 -1
-						rdev->raid_disk < 0)

ClusterRemove and MD_RECOVERY_NEEDED are set in removed rdev->flags on node 2 by process_remove_disk(). After commit a1fd37f97808, node 2 could receive REMOVE in short time after receive of METADATA_UPDATED, MD_RECOVERY_NEEDED will be cleared by md_check_recovery() but in that time rdev->raid_disk could be unchanged. Then MD_RECOVERY_NEEDED will be cleared in the end of md_check_recovery() no more chance to meet the condition.

Quoted 20 lines
drivers/md/md.c +2 -0
+				if (rdev->raid_disk < 0 &&
+				    test_and_clear_bit(ClusterRemove, 

&rdev->flags))

drivers/md/md.c +4 -1
 					md_kick_rdev_from_array(rdev);
 			}
 		}
@@ -10065,8 +10065,11 @@ static void check_sb_changes(struct 
 mddev *mddev, struct md_rdev *rdev)
 
 	/* Check for change of roles in the active devices */
 	rdev_for_each_safe(rdev2, tmp, mddev) {
-		if (test_bit(Faulty, &rdev2->flags))
+		if (test_bit(Faulty, &rdev2->flags)) {
+			if (test_bit(ClusterRemove, &rdev2->flags))
+				set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
 			continue;
+		}
 
 		/* Check if the roles changed */

role = le16_to_cpu(sb->dev_roles[rdev2->desc_nr]);

Applied
#2 lore raw
In reply to Heming Zhao
Quoted 34 lines · Heming Zhao

Commit a1fd37f97808 ("md: Don't wait for MD_RECOVERY_NEEDED for HOT_REMOVE_DISK ioctl") introduced a regression in the md_cluster module. (Failed cases 02r1_Manage_re-add & 02r10_Manage_re-add)

Consider a 2-node cluster:
- node1 set faulty & remove command on a disk.
- node2 must correctly update the array metadata.

Before a1fd37f97808, on node1, the delay between msg:METADATA_UPDATED (triggered by faulty) and msg:REMOVE was sufficient for node2 to reload the disk info (written by node1).
After a1fd37f97808, node1 no longer waits between faulty and remove, causing it to send msg:REMOVE while node2 is still reloading disk info. This often results in node2 failing to remove the faulty disk.

== how to trigger ==

set up a 2-node cluster (node1 & node2) with disks vdc & vdd.

on node1:
mdadm -CR /dev/md0 -l1 -b clustered -n2 /dev/vdc /dev/vdd --assume-clean ssh node2-ip mdadm -A /dev/md0 /dev/vdc /dev/vdd
mdadm --manage /dev/md0 --fail /dev/vdc --remove /dev/vdc

check array status on both nodes with "mdadm -D /dev/md0". node1 output:
Number Major Minor RaidDevice State
- 0 0 0 removed
1 254 48 1 active sync /dev/vdd
node2 output:
Number Major Minor RaidDevice State
- 0 0 0 removed
1 254 48 1 active sync /dev/vdd

Quoted 7 lines

0 254 32 - faulty /dev/vdc

Fixes a1fd37f97808 ("md: Don't wait for MD_RECOVERY_NEEDED for HOT_REMOVE_DISK ioctl")
---
  drivers/md/md.c | 9 ++++++---
  1 file changed, 6 insertions(+), 3 deletions(-)

Applied to md-6.17
Thanks