Skip to contentSkip to conversation

bcachefs: remove sb lock and flags update on explicit shutdown

PATCH In review Brian Foster 2 participants 4 messages Nov 30, 2023 – Dec 2, 2023

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 bcachefs: remove sb lock and flags update on explicit shutdown

bcachefs grabs s_umount and sets SB_RDONLY when the fs is shutdown via the ioctl() interface. This has a couple issues related to interactions between shutdown and freeze:

1. The flags == FSOP_GOING_FLAGS_DEFAULT case is a deadlock vector because freeze_bdev() calls into freeze_super(), which also acquires s_umount.

2. If an explicit shutdown occurs while the sb is frozen, SB_RDONLY alters the thaw path as if the sb was read-only at freeze time. This effectively leaks the frozen state and leaves the sb frozen indefinitely.

The usage of SB_RDONLY here goes back to the initial bcachefs commit and AFAICT is simply historical behavior. This behavior is unique to bcachefs relative to the handful of other filesystems that support the shutdown ioctl(). Typically, SB_RDONLY is reserved for the proper remount path, which itself is restricted from modifying frozen superblocks in reconfigure_super(). Drop the unnecessary sb lock and flags update bch2_ioc_goingdown() to address both of these issues.

---

This is currently churning through CI here (though I had already pushed an early version that ran clean):

https://evilpiepirate.org/~testdashboard/ci?branch=bfoster

Brian

 fs/bcachefs/fs-ioctl.c | 12 ++----------
 1 file changed, 2 insertions(+), 10 deletions(-)
fs/bcachefs/fs-ioctl.c +2 -10
index 5a39bcb597a3..0baa23c35ef5 100644
@@ -287,34 +287,26 @@ static int bch2_ioc_goingdown(struct bch_fs *c, u32 __user *arg)
 
 	bch_notice(c, "shutdown by ioctl type %u", flags);
 
-	down_write(&c->vfs_sb->s_umount);
-
 	switch (flags) {
 	case FSOP_GOING_FLAGS_DEFAULT:
 		ret = freeze_bdev(c->vfs_sb->s_bdev);
 		if (ret)
-			goto err;
-
+			break;
 		bch2_journal_flush(&c->journal);
-		c->vfs_sb->s_flags |= SB_RDONLY;
 		bch2_fs_emergency_read_only(c);
 		thaw_bdev(c->vfs_sb->s_bdev);
 		break;
-
 	case FSOP_GOING_FLAGS_LOGFLUSH:
 		bch2_journal_flush(&c->journal);
 		fallthrough;
-
 	case FSOP_GOING_FLAGS_NOLOGFLUSH:
-		c->vfs_sb->s_flags |= SB_RDONLY;
 		bch2_fs_emergency_read_only(c);
 		break;
 	default:
 		ret = -EINVAL;
 		break;
 	}
-err:
-	up_write(&c->vfs_sb->s_umount);
+
 	return ret;
 }
 
Signature · 1 line
2.41.0
#1 lore raw
Quoted 13 lines · Brian Foster

bcachefs grabs s_umount and sets SB_RDONLY when the fs is shutdown via the ioctl() interface. This has a couple issues related to interactions between shutdown and freeze:

1. The flags == FSOP_GOING_FLAGS_DEFAULT case is a deadlock vector because freeze_bdev() calls into freeze_super(), which also acquires s_umount.

2. If an explicit shutdown occurs while the sb is frozen, SB_RDONLY alters the thaw path as if the sb was read-only at freeze time. This effectively leaks the frozen state and leaves the sb frozen indefinitely.

Quoted 8 lines

The usage of SB_RDONLY here goes back to the initial bcachefs commit and AFAICT is simply historical behavior. This behavior is unique to bcachefs relative to the handful of other filesystems that support the shutdown ioctl(). Typically, SB_RDONLY is reserved for the proper remount path, which itself is restricted from modifying frozen superblocks in reconfigure_super(). Drop the unnecessary sb lock and flags update bch2_ioc_goingdown() to address both of these issues.

Nice, I noticed a deadlock issue recently with s_umount here but didn't have time to fully debug it - applied!

#2 lore raw
Quoted 23 lines · Kent Overstreet
Quoted 13 lines · Brian Foster

bcachefs grabs s_umount and sets SB_RDONLY when the fs is shutdown via the ioctl() interface. This has a couple issues related to interactions between shutdown and freeze:

1. The flags == FSOP_GOING_FLAGS_DEFAULT case is a deadlock vector because freeze_bdev() calls into freeze_super(), which also acquires s_umount.

2. If an explicit shutdown occurs while the sb is frozen, SB_RDONLY alters the thaw path as if the sb was read-only at freeze time. This effectively leaks the frozen state and leaves the sb frozen indefinitely.

Quoted 8 lines

The usage of SB_RDONLY here goes back to the initial bcachefs commit and AFAICT is simply historical behavior. This behavior is unique to bcachefs relative to the handful of other filesystems that support the shutdown ioctl(). Typically, SB_RDONLY is reserved for the proper remount path, which itself is restricted from modifying frozen superblocks in reconfigure_super(). Drop the unnecessary sb lock and flags update bch2_ioc_goingdown() to address both of these issues.

Quoted 3 lines

Nice, I noticed a deadlock issue recently with s_umount here but didn't have time to fully debug it - applied!

Thanks!

Hmm.. got a reference to that observed deadlock handy? I'd be a little cautious associating here because 1. I dont think any test tooling currently uses the default shutdown mode for the ioctl and 2. the non-ioctl emergency shutdown paths dont set SB_RDONLY, so shouldn't trigger the freeze issue.

So IOW while this path was broken in the ways described in the commit log, it still seems rather unlikely for a normal use or test case to trigger this. It would be good to know if there's still a lingering issue to resolve..

Brian

#3 lore raw
Quoted 28 lines · Brian Foster
Quoted 23 lines · Kent Overstreet
Quoted 13 lines · Brian Foster

bcachefs grabs s_umount and sets SB_RDONLY when the fs is shutdown via the ioctl() interface. This has a couple issues related to interactions between shutdown and freeze:

1. The flags == FSOP_GOING_FLAGS_DEFAULT case is a deadlock vector because freeze_bdev() calls into freeze_super(), which also acquires s_umount.

2. If an explicit shutdown occurs while the sb is frozen, SB_RDONLY alters the thaw path as if the sb was read-only at freeze time. This effectively leaks the frozen state and leaves the sb frozen indefinitely.

Quoted 8 lines

The usage of SB_RDONLY here goes back to the initial bcachefs commit and AFAICT is simply historical behavior. This behavior is unique to bcachefs relative to the handful of other filesystems that support the shutdown ioctl(). Typically, SB_RDONLY is reserved for the proper remount path, which itself is restricted from modifying frozen superblocks in reconfigure_super(). Drop the unnecessary sb lock and flags update bch2_ioc_goingdown() to address both of these issues.

Quoted 3 lines

Nice, I noticed a deadlock issue recently with s_umount here but didn't have time to fully debug it - applied!

Quoted 7 lines

Thanks!

Hmm.. got a reference to that observed deadlock handy? I'd be a little cautious associating here because 1. I dont think any test tooling currently uses the default shutdown mode for the ioctl and 2. the non-ioctl emergency shutdown paths dont set SB_RDONLY, so shouldn't trigger the freeze issue.

No, only observed it once and I was looking at other stuff at the time, alas.

Quoted 4 lines

So IOW while this path was broken in the ways described in the commit log, it still seems rather unlikely for a normal use or test case to trigger this. It would be good to know if there's still a lingering issue to resolve..

Touching a VFS lock here at all was always a bit suspect; I have no idea where I originally cribbed that from. I generally try to keep a clearer separation between VFS locks/state and filesystem locks/state...