[PATCH 01/30] block: also call ->open for incremental partition opens

[Al Viro]

On Thu, Jun 08, 2023 at 01:02:29PM +0200, Christoph Hellwig wrote:

> --- a/block/bdev.c
> +++ b/block/bdev.c
> @@ -683,9 +683,6 @@ static int blkdev_get_part(struct block_device *part, fmode_t mode)
>  	struct gendisk *disk = part->bd_disk;
>  	int ret;
>  
> -	if (atomic_read(&part->bd_openers))
> -		goto done;
> -
>  	ret = blkdev_get_whole(bdev_whole(part), mode);
>  	if (ret)
>  		return ret;
> @@ -694,9 +691,10 @@ static int blkdev_get_part(struct block_device *part, fmode_t mode)
>  	if (!bdev_nr_sectors(part))
>  		goto out_blkdev_put;
>  
> -	disk->open_partitions++;
> -	set_init_blocksize(part);
> -done:
> +	if (!atomic_read(&part->bd_openers)) {
> +		disk->open_partitions++;
> +		set_init_blocksize(part);
> +	}

[with apologies for very late (and tangential) reply]

That got me curious about the ->bd_openers - do we need it atomic?
Most of the users (and all places that do modifications) are
under ->open_mutex; the only exceptions are
	* early sync logics in blkdev_put(); it's explicitly racy -
see the comment there.
	* callers of disk_openers() in loop and nbd (the ones in
zram are under ->open_mutex).  There's driver-private exclusion
around those, but in any case - READ_ONCE() is no worse than
atomic_read() in those cases.

Is there something subtle I'm missing here?