Re: [PATCH 1/1] pNFS: Serialize SCSI PR registration to avoid reservation conflicts

[Christoph Hellwig <hch@infradead.org>]

On Sun, Mar 01, 2026 at 04:51:23PM -0800, Dai Ngo wrote:
> This problem can be reproduced by running 'fio' test with this
> workload:

I wish we could wire this up somewhere.  Not sure what the right
place for these kinds of nfs tests are, though.

>  2 files changed, 7 insertions(+), 1 deletion(-)
> 
> diff --git a/fs/nfs/blocklayout/blocklayout.h b/fs/nfs/blocklayout/blocklayout.h
> index 6da40ca19570..535db8b0e89c 100644
> --- a/fs/nfs/blocklayout/blocklayout.h
> +++ b/fs/nfs/blocklayout/blocklayout.h
> @@ -117,6 +117,7 @@ struct pnfs_block_dev {
>  
>  	bool (*map)(struct pnfs_block_dev *dev, u64 offset,
>  			struct pnfs_block_dev_map *map);
> +	struct mutex			pbd_mutex;

Can you keep this up with the non-function pointer fields?  I guess
pbd_registration_lock might be a more descriptive name, and comment
explaining what the lock protects also never hurts.

> diff --git a/fs/nfs/blocklayout/dev.c b/fs/nfs/blocklayout/dev.c
> index cc6327d97a91..45630781f1a8 100644
> --- a/fs/nfs/blocklayout/dev.c
> +++ b/fs/nfs/blocklayout/dev.c
> @@ -33,10 +33,14 @@ static bool bl_register_scsi(struct pnfs_block_dev *dev)
>  	const struct pr_ops *ops = bdev->bd_disk->fops->pr_ops;
>  	int status;
>  
> -	if (test_and_set_bit(PNFS_BDEV_REGISTERED, &dev->flags))
> +	mutex_lock(&dev->pbd_mutex);
> +	if (test_and_set_bit(PNFS_BDEV_REGISTERED, &dev->flags)) {
> +		mutex_unlock(&dev->pbd_mutex);
>  		return true;
> +	}

This seems to only lock the registration side, and not the
unregistration side, which is a bit odd.  If you fully protect
register/unregister we also don't need atomic bitops for
PNFS_BDEV_REGISTERED and have a more consistent locking scheme.