blk_validate_limits validation of block size (was Re: [PATCH v2] null_blk: fix validation of block size)

[John Garry]

(trim list)

On 29/06/2024 06:07, Christoph Hellwig wrote:
> On Fri, Jun 28, 2024 at 03:30:00PM +0100, John Garry wrote:
> > On 04/06/2024 05:46, Christoph Hellwig wrote:
> > > > It also looks like a good idea if this check was just done in
> > > > blk_validate_limits() so that each driver doesn't have to do their own
> > > > checks. That block function is kind of recent though.
> > > Yes.  We already discussed this in another thread a few days ago.
> > Has anyone taken this work? I was going to unless someone else wants to. 4
> > or 5 drivers directly reference blk_validate_block_size() now.
>
> I haven't look at it yet, so from my point of view feel free to tackle
> it.

I spent a bit of time on this, and the driver changes are pretty 
straightforward, apart from nbd.

For nbd, we cannot only change to just stop calling 
blk_validate_limits(). This is because the LBS is possibly updated in a
2-stage process:
a. update block size in the driver and validate
b. update queue limits

like:

static int __nbd_set_size(struct nbd_device *nbd, loff_t bytesize,
		loff_t blksize)
{
	...

	if (blk_validate_block_size(blksize))
		return -EINVAL;

	nbd->config->bytesize = bytesize;
	nbd->config->blksize_bits = __ffs(blksize);

	if (!nbd->pid)
		return 0;

	lim = queue_limits_start_update(nbd->disk->queue);
	...
	error = queue_limits_commit_update(nbd->disk->queue, &lim);

So if we stop validating the limits in a., there is a user-visible 
change in behaviour (as we stop rejecting invalid limits from the 
NBD_SET_BLKSIZE ioctl).

We could add a "dryrun" option to queue_limits_commit_update() (and call 
that instead of blk_validate_block_size(), which is effectively the same 
as calling blk_validate_block_size()). Or we can keep
nbd as the only blk_validate_limits() user (outside the block layer).

Any better ideas?