Re: [PATCH] block: add larger order folio instead of pages for passthrough I/O

[Jens Axboe <axboe@kernel.dk>]

On 11/14/24 9:17 AM, Christoph Hellwig wrote:
>> @@ -313,21 +314,35 @@ static int bio_map_user_iov(struct request *rq, struct iov_iter *iter,
>>  		if (unlikely(offs & queue_dma_alignment(rq->q)))
>>  			j = 0;
>>  		else {
>> -			for (j = 0; j < npages; j++) {
>> +			for (j = 0; j < npages; j += num_pages) {
>>  				struct page *page = pages[j];
>> -				unsigned int n = PAGE_SIZE - offs;
>> +				struct folio *folio = page_folio(page);
>>  				bool same_page = false;
>>  
>> -				if (n > bytes)
>> -					n = bytes;
>>  
>> -				if (!bio_add_hw_page(rq->q, bio, page, n, offs,
>> -						     max_sectors, &same_page))
>> +				folio_offset = ((size_t)folio_page_idx(folio,
>> +						page) << PAGE_SHIFT) + offs;
> 
> I'm not sure if Jens want to rush something like this in for 6.13, but if
> we're aiming for the next merge window I actually have a 3/4 done series
> that rips out bio_add_hw_page and all the passthrough special casing by
> simply running the 'do we need to split the bio' helper on the free-form
> bio and return an error if we do.  That means all this code will go away,
> and you'll automatically get all the work done for the normal path for
> passthrough as well.

I'd rather it simmer a bit first, so I'd say we have time since 6.13 is
coming up really soon.

-- 
Jens Axboe