Re: [PATCH v5 07/16] block: introduce dma map backed bio type

On 8/4/26 17:24, Christoph Hellwig wrote:
> On Sat, Aug 01, 2026 at 04:46:19PM +0100, Pavel Begunkov wrote:
>> Premapped buffers don't require a generic bio_vec since these have
>> already been dma mapped. Repurpose the bi_io_vec space to strore dmabuf
>> maps as they are mutually exclusive.
>>
>> Suggested-by: Keith Busch <[email protected]>
>> Signed-off-by: Pavel Begunkov <[email protected]>
>> ---
...>>   
>> +static inline int bio_split_io_at_dmabuf(struct bio *bio,
>> +		const struct queue_limits *lim, unsigned *segs,
>> +		unsigned max_bytes, unsigned len_align_mask,
>> +		unsigned start_align_mask)
>> +{
>> +	unsigned bytes = min(bio->bi_iter.bi_size, max_bytes);
>> +	unsigned seg_shift = bio->bi_dmabuf_map->seg_shift;
>> +	unsigned offset = bio->bi_iter.bi_offset & ((1U << seg_shift) - 1);
>> +
>> +	if ((bio->bi_iter.bi_offset & start_align_mask) ||
>> +	    (bio->bi_iter.bi_size & len_align_mask))
>> +		return -EINVAL;
>> +
>> +	/* single contiguous range into the dma-buf */
>> +	*segs = 1;
>> +
>> +	bytes = min(bytes, ((unsigned)lim->max_segments << seg_shift) - offset);
> 
> I guess ->seg_shift is some sort of encoding of a max
> segment size?

I'll add a comment.

We rather need the minimum segment size, you can always split large
ones. I calculated it even stricter as the least common multiple
pow2 to avoid divs here.

If the device supports N segments and we know that each segment is
at least M bytes, then we should be able to issue IO of size N*M
of full segments. The line above truncates the bio size using
that + offset adjustments.

I guess it might be more straightforward to calculate the worst
case number of segments and then adjust the splitting size, but
since we don't return the number of segments it's more
computations. E.g.

seg_size = 1U << seg_shift;
nsegs = (bio->bi_iter.bi_size + offset + seg_size - 1) / seg_size;

if (nsegs > lim->max_segments) {
	nsegs = lim->max_segments; // unused after the block
	bytes = seg_size * nsegs - offset;
}

-- 
Pavel Begunkov