Re: [PATCH v2] io_uring: Avoid polling configuration errors

[Christoph Hellwig <hch@infradead.org>]

On Thu, Jul 11, 2024 at 04:24:30PM +0800, hexue wrote:
> +	if (!ctx->check_poll_queue) {
> +		struct block_device *bdev;
> +		struct request_queue *q;
> +		struct inode *inode = req->file->f_inode;
> +
> +		if (inode->i_rdev) {
> +			bdev = blkdev_get_no_open(inode->i_rdev);
> +			q = bdev->bd_queue;
> +			if (!test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
> +				pr_warn("the device does't configured with poll queues\n");
> +		}
> +		ctx->check_poll_queue = true;
> +	}

This is wrong for multiple reasons.  One is that we can't simply poke
into block device internals like this in a higher layer like io_uring.
Second blkdev_get_no_open is in no way available for use outside the
block layer.  The fact that the even exist as separate helpers that
aren't entirely hidden is a decade old layering violation in blk-cgroup.

If you want to advertize this properly we'll need a flag in struct
file or something similar.