Since commit d869da91cccb, an unmount of a pNFS SCSI layout-enabled NFS
will dereference a NULL block_device in:

  bl_unregister_scsi+0x16/0xe0 [blocklayoutdriver]
  bl_free_device+0x70/0x80 [blocklayoutdriver]
  bl_free_deviceid_node+0x12/0x30 [blocklayoutdriver]
  nfs4_put_deviceid_node+0x60/0xc0 [nfsv4]
  nfs4_deviceid_purge_client+0x132/0x190 [nfsv4]
  unset_pnfs_layoutdriver+0x59/0x60 [nfsv4]
  nfs4_destroy_server+0x36/0x70 [nfsv4]
  nfs_free_server+0x23/0xe0 [nfs]
  deactivate_locked_super+0x30/0xb0
  cleanup_mnt+0xba/0x150
  task_work_run+0x59/0x90
  syscall_exit_to_user_mode+0x217/0x220
  do_syscall_64+0x8e/0x160

This happens because even though we were able to create the
nfs4_deviceid_node, the lookup for the device was unable to attach the
block device to the pnfs_block_dev.

If we never found a block device to register, we can avoid this case with
the PNFS_BDEV_REGISTERED flag.  Move the deref behind the test for the
flag.

Fixes: d869da91cccb ("nfs/blocklayout: Fix premature PR key unregistration")
Signed-off-by: Benjamin Coddington <bcodding@redhat.com>
---
 fs/nfs/blocklayout/dev.c | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)

diff --git a/fs/nfs/blocklayout/dev.c b/fs/nfs/blocklayout/dev.c
index 6252f4447945..7ae79814f4ff 100644
--- a/fs/nfs/blocklayout/dev.c
+++ b/fs/nfs/blocklayout/dev.c
@@ -16,13 +16,16 @@
 
 static void bl_unregister_scsi(struct pnfs_block_dev *dev)
 {
-	struct block_device *bdev = file_bdev(dev->bdev_file);
-	const struct pr_ops *ops = bdev->bd_disk->fops->pr_ops;
+	struct block_device *bdev;
+	const struct pr_ops *ops;
 	int status;
 
 	if (!test_and_clear_bit(PNFS_BDEV_REGISTERED, &dev->flags))
 		return;
 
+	bdev = file_bdev(dev->bdev_file);
+	ops = bdev->bd_disk->fops->pr_ops;
+
 	status = ops->pr_register(bdev, dev->pr_key, 0, false);
 	if (status)
 		trace_bl_pr_key_unreg_err(bdev, dev->pr_key, status);
-- 
2.47.0

On Mon, Nov 18, 2024 at 05:40:40PM -0500, Benjamin Coddington wrote:
> Since commit d869da91cccb, an unmount of a pNFS SCSI layout-enabled NFS

Please also spell out the commit subject in the commit log body, similar
to to the Fixes tag.

> index 6252f4447945..7ae79814f4ff 100644
> --- a/fs/nfs/blocklayout/dev.c
> +++ b/fs/nfs/blocklayout/dev.c
> @@ -16,13 +16,16 @@
>  
>  static void bl_unregister_scsi(struct pnfs_block_dev *dev)
>  {
> -	struct block_device *bdev = file_bdev(dev->bdev_file);
> -	const struct pr_ops *ops = bdev->bd_disk->fops->pr_ops;
> +	struct block_device *bdev;
> +	const struct pr_ops *ops;
>  	int status;
>  
>  	if (!test_and_clear_bit(PNFS_BDEV_REGISTERED, &dev->flags))
>  		return;
>  
> +	bdev = file_bdev(dev->bdev_file);
> +	ops = bdev->bd_disk->fops->pr_ops;
> +

Hmm.  Just moving the test_and_clear_bit to the caller would
feel cleaner than this to me.

But either way the change looks good:

Reviewed-by: Christoph Hellwig <hch@lst.de>

On 19 Nov 2024, at 7:42, Christoph Hellwig wrote:

> On Mon, Nov 18, 2024 at 05:40:40PM -0500, Benjamin Coddington wrote:
>> Since commit d869da91cccb, an unmount of a pNFS SCSI layout-enabled NFS
>
> Please also spell out the commit subject in the commit log body, similar
> to to the Fixes tag.

Will do.

>> index 6252f4447945..7ae79814f4ff 100644
>> --- a/fs/nfs/blocklayout/dev.c
>> +++ b/fs/nfs/blocklayout/dev.c
>> @@ -16,13 +16,16 @@
>>
>>  static void bl_unregister_scsi(struct pnfs_block_dev *dev)
>>  {
>> -	struct block_device *bdev = file_bdev(dev->bdev_file);
>> -	const struct pr_ops *ops = bdev->bd_disk->fops->pr_ops;
>> +	struct block_device *bdev;
>> +	const struct pr_ops *ops;
>>  	int status;
>>
>>  	if (!test_and_clear_bit(PNFS_BDEV_REGISTERED, &dev->flags))
>>  		return;
>>
>> +	bdev = file_bdev(dev->bdev_file);
>> +	ops = bdev->bd_disk->fops->pr_ops;
>> +
>
> Hmm.  Just moving the test_and_clear_bit to the caller would
> feel cleaner than this to me.

We can do this too - I'll send another version.  I didn't go this way
because I felt the bit test and clear was an important part of the function
that could be lost if we ever had another caller.

> But either way the change looks good:
>
> Reviewed-by: Christoph Hellwig <hch@lst.de>

Thanks for the review.

Ben