[PATCH v20 03/12] block: add copy offload support

[Nitesh Shetty]

On 01/06/24 08:16AM, Christoph Hellwig wrote:
>> +/* Keeps track of all outstanding copy IO */
>> +struct blkdev_copy_io {
>> +	atomic_t refcount;
>> +	ssize_t copied;
>> +	int status;
>> +	struct task_struct *waiter;
>> +	void (*endio)(void *private, int status, ssize_t copied);
>> +	void *private;
>> +};
>> +
>> +/* Keeps track of single outstanding copy offload IO */
>> +struct blkdev_copy_offload_io {
>> +	struct blkdev_copy_io *cio;
>> +	loff_t offset;
>> +};
>
>The structure names confuse me, and the comments make things even worse.
>
>AFAICT:
>
>blkdev_copy_io is a per-call structure, I'd name it blkdev_copy_ctx.
>blkdev_copy_offload_io is per-bio pair, and something like blkdev_copy_chunk
Acked, your suggestion for structure name looks better.

>might be a better idea.  Or we could just try to kill it entirely and add
>a field to struct bio in the union currently holding the integrity
>information.
We will explore this.

>I'm also quite confused what kind of offset this offset field is.  The
>type and name suggest it is an offset in a file, which for a block device
>based helper is pretty odd to start with.  blkdev_copy_offload
>initializes it to len - rem, so it kind is an offset, but relative
>to the operation and not to a file. blkdev_copy_offload_src_endio then
>uses to set the ->copied field, but based on a min which means
>->copied can only be decreased.  Something is really off there.
>
Offset in this context, is with respect to the operation.
Overall idea was to handle partial copy, where in some of the split copy IO fails.
In this case we want to return minimum bytes copied.
We can try to store the offset in a temporary variable similar to
pos_out, pos_in instead of current (len - rem), to avoid the confusion.
We will update this in next version.

>Taking about types and units: blkdev_copy_offload obviously can only
>work in terms of LBAs.  Any reason to not make it work in terms of
>512-byte block layer sectors instead of in bytes?
>
Just that number of places where we need to sector shift were
comparatively more. We will update this to 512-byte sectors in next
version.

>> +	if ((pos_in & align) || (pos_out & align) || (len & align) || !len ||
>> +	    len >= BLK_COPY_MAX_BYTES)
>> +		return -EINVAL;
>
>This can be cleaned up an optimized a bit:
>
>	if (!len || len >= BLK_COPY_MAX_BYTES)
>		return -EINVAL;
>	if ((pos_in | pos_out | len) & align)
>		return -EINVAL;
>	
Acked.

>> + *
>> + * For synchronous operation returns the length of bytes copied or error
>> + * For asynchronous operation returns -EIOCBQUEUED or error
>> + *
>> + * Description:
>> + *	Copy source offset to destination offset within block device, using
>> + *	device's native copy offload feature.
>> + *	We perform copy operation using 2 bio's.
>> + *	1. We take a plug and send a REQ_OP_COPY_DST bio along with destination
>> + *	sector and length. Once this bio reaches request layer, we form a
>> + *	request and wait for dst bio to arrive.
>> + *	2. We issue REQ_OP_COPY_SRC bio along with source sector, length.
>> + *	Once this bio reaches request layer and find a request with previously
>> + *	sent destination info we merge the source bio and return.
>> + *	3. Release the plug and request is sent to driver
>> + *	This design works only for drivers with request queue.
>
>The wording with all the We here is a bit odd.  Much of this also seem
>superfluous or at least misplaced in the kernel doc comment as it doesn't
>document the API, but just what is done in the code below.
>
Since we were doing IO in unconventional way, we felt would be better to
document this, for easy followup.
We will remove this in next version and document just API.

>> +	cio = kzalloc(sizeof(*cio), gfp);
>> +	if (!cio)
>> +		return -ENOMEM;
>> +	atomic_set(&cio->refcount, 1);
>> +	cio->waiter = current;
>> +	cio->endio = endio;
>> +	cio->private = private;
>
>For the main use this could be allocated on-stack.  Is there any good
>reason to not let callers that really want an async version to implement
>the async behavior themselves using suitable helpers?
>
We cannot do on-stack allocation of cio as we use it in endio handler.
cio will be used to track partial IO completion as well.
Callers requiring async implementation would need to manage all this
bookkeeping themselves, leading to duplication of code. We felt it is
better to do it here onetime.
Do you see it any differently ?

>> +		src_bio = blk_next_bio(dst_bio, bdev, 0, REQ_OP_COPY_SRC, gfp);
>
>Please switch to use bio_chain_and_submit, which is a easier to
>understand API.  I'm trying to phase out blk_next_bio in favour of
>bio_chain_and_submit over the next few merge windows.
>
Acked

>> +		if (!src_bio)
>> +			goto err_free_dst_bio;
>> +		src_bio->bi_iter.bi_size = chunk;
>> +		src_bio->bi_iter.bi_sector = pos_in >> SECTOR_SHIFT;
>> +		src_bio->bi_end_io = blkdev_copy_offload_src_endio;
>> +		src_bio->bi_private = offload_io;
>> +
>> +		atomic_inc(&cio->refcount);
>> +		submit_bio(src_bio);
>> +		blk_finish_plug(&plug);
>
>plugs should be hold over all  I/Os, submitted from the same caller,
>which is the point of them.
>
Acked

Thank You,
Nitesh Shetty