[dm-devel] [RFC PATCH v3 1/2] block: add simple copy support

[Keith Busch]

On Fri, Dec 11, 2020 at 07:21:38PM +0530, SelvaKumar S wrote:
> +int blk_copy_emulate(struct block_device *bdev, struct blk_copy_payload *payload,
> +		gfp_t gfp_mask)
> +{
> +	struct request_queue *q = bdev_get_queue(bdev);
> +	struct bio *bio;
> +	void *buf = NULL;
> +	int i, nr_srcs, max_range_len, ret, cur_dest, cur_size;
> +
> +	nr_srcs = payload->copy_range;
> +	max_range_len = q->limits.max_copy_range_sectors << SECTOR_SHIFT;

The default value for this limit is 0, and this is the function for when
the device doesn't support copy. Are we expecting drivers to set this
value to something else for that case?

> +	cur_dest = payload->dest;
> +	buf = kvmalloc(max_range_len, GFP_ATOMIC);
> +	if (!buf)
> +		return -ENOMEM;
> +
> +	for (i = 0; i < nr_srcs; i++) {
> +		bio = bio_alloc(gfp_mask, 1);
> +		bio->bi_iter.bi_sector = payload->range[i].src;
> +		bio->bi_opf = REQ_OP_READ;
> +		bio_set_dev(bio, bdev);
> +
> +		cur_size = payload->range[i].len << SECTOR_SHIFT;
> +		ret = bio_add_page(bio, virt_to_page(buf), cur_size,
> +						   offset_in_page(payload));

'buf' is vmalloc'ed, so we don't necessarily have congituous pages. I
think you need to allocate the bio with bio_map_kern() or something like
that instead with that kind of memory.

> +		if (ret != cur_size) {
> +			ret = -ENOMEM;
> +			goto out;
> +		}
> +
> +		ret = submit_bio_wait(bio);
> +		bio_put(bio);
> +		if (ret)
> +			goto out;
> +
> +		bio = bio_alloc(gfp_mask, 1);
> +		bio_set_dev(bio, bdev);
> +		bio->bi_opf = REQ_OP_WRITE;
> +		bio->bi_iter.bi_sector = cur_dest;
> +		ret = bio_add_page(bio, virt_to_page(buf), cur_size,
> +						   offset_in_page(payload));
> +		if (ret != cur_size) {
> +			ret = -ENOMEM;
> +			goto out;
> +		}
> +
> +		ret = submit_bio_wait(bio);
> +		bio_put(bio);
> +		if (ret)
> +			goto out;
> +
> +		cur_dest += payload->range[i].len;
> +	}

I think this would be a faster implementation if the reads were
asynchronous with a payload buffer allocated specific to that read, and
the callback can enqueue the write part. This would allow you to
accumulate all the read data and write it in a single call.