projects / thirdparty / kernel / stable-queue.git / blame
| Commit | Line | Data |
|---|---|---|
| 1762a8ed GKH |
1 | From 51e9ac77035a3dfcb6fc0a88a0d80b6f99b5edb1 Mon Sep 17 00:00:00 2001 |
| 2 | From: NeilBrown <neilb@suse.de> | |
| 3 | Date: Sat, 7 Aug 2010 21:17:00 +1000 | |
| 4 | Subject: md/raid10: fix deadlock with unaligned read during resync | |
| 5 | ||
| 6 | From: NeilBrown <neilb@suse.de> | |
| 7 | ||
| 8 | commit 51e9ac77035a3dfcb6fc0a88a0d80b6f99b5edb1 upstream. | |
| 9 | ||
| 10 | If the 'bio_split' path in raid10-read is used while | |
| 11 | resync/recovery is happening it is possible to deadlock. | |
| 12 | Fix this be elevating ->nr_waiting for the duration of both | |
| 13 | parts of the split request. | |
| 14 | ||
| 15 | This fixes a bug that has been present since 2.6.22 | |
| 16 | but has only started manifesting recently for unknown reasons. | |
| 17 | It is suitable for and -stable since then. | |
| 18 | ||
| 19 | Reported-by: Justin Bronder <jsbronder@gentoo.org> | |
| 20 | Tested-by: Justin Bronder <jsbronder@gentoo.org> | |
| 21 | Signed-off-by: NeilBrown <neilb@suse.de> | |
| 22 | Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de> | |
| 23 | ||
| 24 | --- | |
| 25 | drivers/md/raid10.c | 18 ++++++++++++++++++ | |
| 26 | 1 file changed, 18 insertions(+) | |
| 27 | ||
| 28 | --- a/drivers/md/raid10.c | |
| 29 | +++ b/drivers/md/raid10.c | |
| 30 | @@ -824,11 +824,29 @@ static int make_request(struct request_q | |
| 31 | */ | |
| 32 | bp = bio_split(bio, | |
| 33 | chunk_sects - (bio->bi_sector & (chunk_sects - 1)) ); | |
| 34 | + | |
| 35 | + /* Each of these 'make_request' calls will call 'wait_barrier'. | |
| 36 | + * If the first succeeds but the second blocks due to the resync | |
| 37 | + * thread raising the barrier, we will deadlock because the | |
| 38 | + * IO to the underlying device will be queued in generic_make_request | |
| 39 | + * and will never complete, so will never reduce nr_pending. | |
| 40 | + * So increment nr_waiting here so no new raise_barriers will | |
| 41 | + * succeed, and so the second wait_barrier cannot block. | |
| 42 | + */ | |
| 43 | + spin_lock_irq(&conf->resync_lock); | |
| 44 | + conf->nr_waiting++; | |
| 45 | + spin_unlock_irq(&conf->resync_lock); | |
| 46 | + | |
| 47 | if (make_request(q, &bp->bio1)) | |
| 48 | generic_make_request(&bp->bio1); | |
| 49 | if (make_request(q, &bp->bio2)) | |
| 50 | generic_make_request(&bp->bio2); | |
| 51 | ||
| 52 | + spin_lock_irq(&conf->resync_lock); | |
| 53 | + conf->nr_waiting--; | |
| 54 | + wake_up(&conf->wait_barrier); | |
| 55 | + spin_unlock_irq(&conf->resync_lock); | |
| 56 | + | |
| 57 | bio_pair_release(bp); | |
| 58 | return 0; | |
| 59 | bad_map: |