Skip to contentSkip to conversation

xfs_repair: coordinate parallel updates to the rt bitmap

PATCH Superseded Darrick J. Wong 2 participants 5 messages Sep 23, 2020 – Sep 24, 2020

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 xfs_repair: coordinate parallel updates to the rt bitmap

From: Darrick J. Wong <[email protected]>

Actually take the rt lock before updating the bitmap from multiple threads. This fixes an infrequent corruption problem when running generic/013 and rtinherit=1 is set on the root dir.

---
 repair/dinode.c |    4 ++++
 1 file changed, 4 insertions(+)
repair/dinode.c +4 -0
index 57013bf149b2..07f3f83aef8c 100644
@@ -184,6 +184,7 @@ process_rt_rec(
 	xfs_rfsblock_t		*tot,
 	int			check_dups)
 {
+	struct aglock		*lock = &ag_locks[(signed)NULLAGNUMBER];
 	xfs_fsblock_t		b, lastb;
 	xfs_rtblock_t		ext;
 	int			state;
@@ -245,6 +246,7 @@ _("data fork in rt ino %" PRIu64 " claims dup rt extent,"
 			continue;
 		}
 
+		pthread_mutex_lock(&lock->lock);
 		state = get_rtbmap(ext);
 		switch (state)  {
 		case XR_E_FREE:
@@ -270,6 +272,7 @@ _("data fork in rt inode %" PRIu64 " found metadata block %" PRIu64 " in rt bmap
 			do_warn(
 _("data fork in rt inode %" PRIu64 " claims used rt block %" PRIu64 "\n"),
 				ino, ext);
+			pthread_mutex_unlock(&lock->lock);
 			return 1;
 		case XR_E_FREE1:
 		default:
@@ -277,6 +280,7 @@ _("data fork in rt inode %" PRIu64 " claims used rt block %" PRIu64 "\n"),
 _("illegal state %d in rt block map %" PRIu64 "\n"),
 				state, b);
 		}
+		pthread_mutex_unlock(&lock->lock);
 	}
 
 	/*
#1 lore raw
Quoted 16 lines · Darrick J. Wong

From: Darrick J. Wong <[email protected]>

Actually take the rt lock before updating the bitmap from multiple threads. This fixes an infrequent corruption problem when running generic/013 and rtinherit=1 is set on the root dir.

---
 repair/dinode.c |    4 ++++
 1 file changed, 4 insertions(+)
repair/dinode.c +0 -0
index 57013bf149b2..07f3f83aef8c 100644

repair/dinode.c process_rt_rec( · line 187

Quoted 4 lines
repair/dinode.c +1 -0
@@ -184,6 +184,7 @@ process_rt_rec(
 	xfs_rfsblock_t		*tot,
 	int			check_dups)
 {
+	struct aglock		*lock = &ag_locks[(signed)NULLAGNUMBER];

Err, what is this weird cast doing here?

The rest looks sane.

#2 lore raw
Quoted 12 lines · Christoph Hellwig
Quoted 11 lines · Darrick J. Wong

From: Darrick J. Wong <[email protected]>

Actually take the rt lock before updating the bitmap from multiple threads. This fixes an infrequent corruption problem when running generic/013 and rtinherit=1 is set on the root dir.

---
 repair/dinode.c |    4 ++++
 1 file changed, 4 insertions(+)
Quoted 11 lines

repair/dinode.c process_rt_rec( · line 187

Quoted 9 lines
repair/dinode.c +1 -0
index 57013bf149b2..07f3f83aef8c 100644
@@ -184,6 +184,7 @@ process_rt_rec(
 	xfs_rfsblock_t		*tot,
 	int			check_dups)
 {
+	struct aglock		*lock = &ag_locks[(signed)NULLAGNUMBER];

Err, what is this weird cast doing here?

Well.... ag_locks is allocated with length ag_locks[agcount + 1], and then the pointer is incremented so that ag_locks[-1] is the rt lock.

NULLAGNUMBER is cast to xfs_agnumber_t, which is uint32_t, so we have to cast it back to signed so that we actually do the pointer subtraction(!)

Yeah, I know, it's nuts...

--D

Quoted 1 line

The rest looks sane.

#3 lore raw
Quoted 6 lines · Darrick J. Wong
Quoted 3 lines

repair/dinode.c process_rt_rec( · line 187

Quoted 1 line
repair/dinode.c +1 -0
+	struct aglock		*lock = &ag_locks[(signed)NULLAGNUMBER];

Err, what is this weird cast doing here?

Well.... ag_locks is allocated with length ag_locks[agcount + 1], and then the pointer is incremented so that ag_locks[-1] is the rt lock.

At least in the for-next branch it isn't:

	ag_locks = calloc(mp->m_sb.sb_agcount, sizeof(struct aglock));

More importantly, I can't even find other uses of ag_locks for the RT subvolume. Is this hidden in one of your series?

Either way I think a separate lock for the RT subvolume would make a whole lot more sense.

#4 lore raw
Quoted 8 lines · Christoph Hellwig
Quoted 6 lines · Darrick J. Wong
Quoted 3 lines

repair/dinode.c process_rt_rec( · line 187

Quoted 1 line
repair/dinode.c +1 -0
+	struct aglock		*lock = &ag_locks[(signed)NULLAGNUMBER];

Err, what is this weird cast doing here?

Well.... ag_locks is allocated with length ag_locks[agcount + 1], and then the pointer is incremented so that ag_locks[-1] is the rt lock.

Quoted 6 lines

At least in the for-next branch it isn't:

	ag_locks = calloc(mp->m_sb.sb_agcount, sizeof(struct aglock));

More importantly, I can't even find other uses of ag_locks for the RT subvolume. Is this hidden in one of your series?

Doh. Yes, it is, in the realtime rmap series. :( :(

Quoted 2 lines

Either way I think a separate lock for the RT subvolume would make a whole lot more sense.

Yes, let's do it that way.

--D