[PATCH v2 for-5.8-rc 1/6] nvme: fix possible deadlock when I/O is blocked

[Christoph Hellwig]

On Tue, Jul 07, 2020 at 12:57:35PM +0200, Anthony Iliopoulos wrote:
> The revalidation is still required there in order to update the blockdev
> size when multipath is enabled. With this revert, the behavior regressed
> back to before fab7772bfbcf, so the size isn't updated under multipath
> when the device is in use (e.g. mounted), it will only be effectively
> updated next time it is mounted as a side-effect of blkdev_open on first
> opener. This prevents online resizing of filesystems.
> 
> > And if you look at revalidate, it only calls fops->revalidate_disk
> > (which is a noop in the mpath device) and checks the size change.
> 
> The mpath gendisk doesn't have a revalidate_disk fop indeed, but call to
> check_disk_size_change was needed to update its size. Otherwise, the
> size change isn't reflected externally to userspace before next mount.
> cb224c3af4df doesn't fix this as it skips revalidate and thus also skips
> updating the size.
> 
> The issue is that under CONFIG_NVME_MULTIPATH=y the ns->disk becomes a
> hidden gendisk (see nvme_set_disk_name) and thus doesn't have a backing
> bdev, while ns->head->disk becomes the "visible" one to userspace (with
> a backing bdev), and therefore needs its bdev->bd_inode->i_size updated
> when resized.
> 
> The following patch seems to be working, although I'm not sufficiently
> familiar with the code to tell if this won't cause any other deadlocks
> with some confidence:

This looks sensible to me.  Sagi?