On Tue, Jul 27, 2021 at 05:01:14PM -0400, Josef Bacik wrote:

We got the following lockdep splat while running xfstests (specifically
btrfs/003 and btrfs/020 in a row) with the new rc.  This was uncovered
by 87579e9b7d8d ("loop: use worker per cgroup instead of kworker") which
converted loop to using workqueues, which comes with lockdep
annotations that don't exist with kworkers.  The lockdep splat is as
follows

======================================================
WARNING: possible circular locking dependency detected
5.14.0-rc2-custom+ #34 Not tainted
------------------------------------------------------
losetup/156417 is trying to acquire lock:
ffff9c7645b02d38 ((wq_completion)loop0){+.+.}-{0:0}, at: flush_workqueue+0x84/0x600

but task is already holding lock:
ffff9c7647395468 (&lo->lo_mutex){+.+.}-{3:3}, at: __loop_clr_fd+0x41/0x650 [loop]

which lock already depends on the new lock.

the existing dependency chain (in reverse order) is:

-> #5 (&lo->lo_mutex){+.+.}-{3:3}:
       __mutex_lock+0xba/0x7c0
       lo_open+0x28/0x60 [loop]
       blkdev_get_whole+0x28/0xf0
       blkdev_get_by_dev.part.0+0x168/0x3c0
       blkdev_open+0xd2/0xe0
       do_dentry_open+0x163/0x3a0
       path_openat+0x74d/0xa40
       do_filp_open+0x9c/0x140
       do_sys_openat2+0xb1/0x170
       __x64_sys_openat+0x54/0x90
       do_syscall_64+0x3b/0x90
       entry_SYSCALL_64_after_hwframe+0x44/0xae

-> #4 (&disk->open_mutex){+.+.}-{3:3}:
       __mutex_lock+0xba/0x7c0
       blkdev_get_by_dev.part.0+0xd1/0x3c0
       blkdev_get_by_path+0xc0/0xd0
       btrfs_scan_one_device+0x52/0x1f0 [btrfs]
       btrfs_control_ioctl+0xac/0x170 [btrfs]
       __x64_sys_ioctl+0x83/0xb0
       do_syscall_64+0x3b/0x90
       entry_SYSCALL_64_after_hwframe+0x44/0xae

-> #3 (uuid_mutex){+.+.}-{3:3}:
       __mutex_lock+0xba/0x7c0
       btrfs_rm_device+0x48/0x6a0 [btrfs]
       btrfs_ioctl+0x2d1c/0x3110 [btrfs]
       __x64_sys_ioctl+0x83/0xb0
       do_syscall_64+0x3b/0x90
       entry_SYSCALL_64_after_hwframe+0x44/0xae

-> #2 (sb_writers#11){.+.+}-{0:0}:
       lo_write_bvec+0x112/0x290 [loop]
       loop_process_work+0x25f/0xcb0 [loop]
       process_one_work+0x28f/0x5d0
       worker_thread+0x55/0x3c0
       kthread+0x140/0x170
       ret_from_fork+0x22/0x30

-> #1 ((work_completion)(&lo->rootcg_work)){+.+.}-{0:0}:
       process_one_work+0x266/0x5d0
       worker_thread+0x55/0x3c0
       kthread+0x140/0x170
       ret_from_fork+0x22/0x30

-> #0 ((wq_completion)loop0){+.+.}-{0:0}:
       __lock_acquire+0x1130/0x1dc0
       lock_acquire+0xf5/0x320
       flush_workqueue+0xae/0x600
       drain_workqueue+0xa0/0x110
       destroy_workqueue+0x36/0x250
       __loop_clr_fd+0x9a/0x650 [loop]
       lo_ioctl+0x29d/0x780 [loop]
       block_ioctl+0x3f/0x50
       __x64_sys_ioctl+0x83/0xb0
       do_syscall_64+0x3b/0x90
       entry_SYSCALL_64_after_hwframe+0x44/0xae

other info that might help us debug this:
Chain exists of:
  (wq_completion)loop0 --> &disk->open_mutex --> &lo->lo_mutex
 Possible unsafe locking scenario:
       CPU0                    CPU1
       ----                    ----
  lock(&lo->lo_mutex);
                               lock(&disk->open_mutex);
                               lock(&lo->lo_mutex);
  lock((wq_completion)loop0);

 *** DEADLOCK ***
1 lock held by losetup/156417:
 #0: ffff9c7647395468 (&lo->lo_mutex){+.+.}-{3:3}, at: __loop_clr_fd+0x41/0x650 [loop]

stack backtrace:
CPU: 8 PID: 156417 Comm: losetup Not tainted 5.14.0-rc2-custom+ #34
Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 0.0.0 02/06/2015
Call Trace:
 dump_stack_lvl+0x57/0x72
 check_noncircular+0x10a/0x120
 __lock_acquire+0x1130/0x1dc0
 lock_acquire+0xf5/0x320
 ? flush_workqueue+0x84/0x600
 flush_workqueue+0xae/0x600
 ? flush_workqueue+0x84/0x600
 drain_workqueue+0xa0/0x110
 destroy_workqueue+0x36/0x250
 __loop_clr_fd+0x9a/0x650 [loop]
 lo_ioctl+0x29d/0x780 [loop]
 ? __lock_acquire+0x3a0/0x1dc0
 ? update_dl_rq_load_avg+0x152/0x360
 ? lock_is_held_type+0xa5/0x120
 ? find_held_lock.constprop.0+0x2b/0x80
 block_ioctl+0x3f/0x50
 __x64_sys_ioctl+0x83/0xb0
 do_syscall_64+0x3b/0x90
 entry_SYSCALL_64_after_hwframe+0x44/0xae
RIP: 0033:0x7f645884de6b

Usually the uuid_mutex exists to protect the fs_devices that map
together all of the devices that match a specific uuid.  In rm_device
we're messing with the uuid of a device, so it makes sense to protect
that here.

However in doing that it pulls in a whole host of lockdep dependencies,
as we call mnt_may_write() on the sb before we grab the uuid_mutex, thus
we end up with the dependency chain under the uuid_mutex being added
under the normal sb write dependency chain, which causes problems with
loop devices.

We don't need the uuid mutex here however.  If we call
btrfs_scan_one_device() before we scratch the super block we will find
the fs_devices and not find the device itself and return EBUSY because
the fs_devices is open.  If we call it after the scratch happens it will
not appear to be a valid btrfs file system.

This is a bit hand wavy but the critical part of the correctness proof,
and it's not explaining it enough IMO. The important piece happens in
device_list_add, the fs_devices lookup and EBUSY, but all that is now
excluded completely by the uuid_mutex from running in parallel with any
part of rm_device.

This means that the state of the device is seen complete by each (scan,
rm device). Without the uuid mutex the scaning can find the signature,
then try to lookup the device in the list, while in parallel the rm
device changes the signature or manipulates the list. But not everything
is covered by the device list mutex so there are combinations of both
tasks with some in-progress state.  Also count in the RCU protection.

From high level it is what you say about ordering scan/scratch, but
otherwise I'm not convinced that the change is not subtly breaking
something.