Re: [syzbot] [btrfs?] KASAN: slab-use-after-free Read in btrfs_open_devices

[Christian Brauner]

On Mon, Aug 07, 2023 at 08:24:36PM -0700, syzbot wrote:
> syzbot has bisected this issue to:
> 
> commit 066d64b26a21a5b5c500a30f27f3e4b1959aac9e
> Author: Christoph Hellwig <hch@xxxxxx>
> Date:   Wed Aug 2 15:41:23 2023 +0000
> 
>     btrfs: open block devices after superblock creation
> 
> bisection log:  https://syzkaller.appspot.com/x/bisect.txt?x=15493371a80000
> start commit:   f7dc24b34138 Add linux-next specific files for 20230807
> git tree:       linux-next
> final oops:     https://syzkaller.appspot.com/x/report.txt?x=17493371a80000
> console output: https://syzkaller.appspot.com/x/log.txt?x=13493371a80000
> kernel config:  https://syzkaller.appspot.com/x/.config?x=d7847c9dca13d6c5
> dashboard link: https://syzkaller.appspot.com/bug?extid=26860029a4d562566231
> syz repro:      https://syzkaller.appspot.com/x/repro.syz?x=179704c9a80000
> C reproducer:   https://syzkaller.appspot.com/x/repro.c?x=17868ba9a80000
> 
> Reported-by: syzbot+26860029a4d562566231@xxxxxxxxxxxxxxxxxxxxxxxxx
> Fixes: 066d64b26a21 ("btrfs: open block devices after superblock creation")
> 
> For information about bisection process see: https://goo.gl/tpsmEJ#bisection

I think the issue might be that before your patch the lifetime of:
@device was aligned with @device->s_fs_info but now that you're dropping
the uuid mutex after btrfs_scan_one_device() that isn't true anymore. So
it feels like:

P1                                       P2
lock_uuid_mutex;
device = btrfs_scan_one_device();
fs_devices = device->fs_devices;
unlock_uuid_mutex;
                                         // earlier mount that gets cleaned up
                                         lock_uuid_mutex; 
					 btrfs_close_devices(fs_devices);
                                         unlock_uuid_mutex;

lock_uuid_mutex;
btrfs_open_devices(fs_devices); // UAF
unlock_uuid_mutex;

But I'm not entirely sure.