Thread (13 messages) 13 messages, 3 authors, 2023-05-26

Re: [PATCH 3/3] md/raid10: fix io loss while replacement replace rdev

From: Yu Kuai <hidden>
Date: 2023-05-22 13:03:20
Also in: lkml

Hi,

在 2023/05/22 19:54, linan666@huaweicloud.com 写道:
From: Li Nan <redacted>

When we remove a disk which has replacement, first set rdev to NULL
and then set replacement to rdev, finally set replacement to NULL (see
raid10_remove_disk()). If io is submitted during the same time, it might
read both rdev and replacement as NULL, and io will not be submitted.

   rdev -> NULL
			read rdev
   replacement -> NULL
			read replacement

Fix it by reading replacement first and rdev later, meanwhile, use smp_mb()
to prevent memory reordering.
Looks good, feel free to add:

Reviewed-by: Yu Kuai <redacted>
quoted hunk ↗ jump to hunk
Fixes: 475b0321a4df ("md/raid10: writes should get directed to replacement as well as original.")
Signed-off-by: Li Nan <redacted>
---
  drivers/md/raid10.c | 22 ++++++++++++++++++----
  1 file changed, 18 insertions(+), 4 deletions(-)
diff --git a/drivers/md/raid10.c b/drivers/md/raid10.c
index 70cc87c7ee57..25a5a7b1e95c 100644
--- a/drivers/md/raid10.c
+++ b/drivers/md/raid10.c
@@ -779,8 +779,16 @@ static struct md_rdev *read_balance(struct r10conf *conf,
  		disk = r10_bio->devs[slot].devnum;
  		rdev = rcu_dereference(conf->mirrors[disk].replacement);
  		if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
-		    r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
+		    r10_bio->devs[slot].addr + sectors >
+		    rdev->recovery_offset) {
+			/*
+			 * Read replacement first to prevent reading both rdev
+			 * and replacement as NULL during replacement replace
+			 * rdev.
+			 */
+			smp_mb();
  			rdev = rcu_dereference(conf->mirrors[disk].rdev);
+		    }
  		if (rdev == NULL ||
  		    test_bit(Faulty, &rdev->flags))
  			continue;
@@ -1479,9 +1487,15 @@ static void raid10_write_request(struct mddev *mddev, struct bio *bio,
  
  	for (i = 0;  i < conf->copies; i++) {
  		int d = r10_bio->devs[i].devnum;
-		struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
-		struct md_rdev *rrdev = rcu_dereference(
-			conf->mirrors[d].replacement);
+		struct md_rdev *rdev, *rrdev;
+
+		rrdev = rcu_dereference(conf->mirrors[d].replacement);
+		/*
+		 * Read replacement first to Prevent reading both rdev and
+		 * replacement as NULL during replacement replace rdev.
+		 */
+		smp_mb();
+		rdev = rcu_dereference(conf->mirrors[d].rdev);
  		if (rdev == rrdev)
  			rrdev = NULL;
  		if (rdev && (test_bit(Faulty, &rdev->flags)))
  
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help