Re: [PATCH 2/2 v2] UBI: Block: Add blk-mq support

[Jens Axboe <axboe@fb.com>]

On 01/13/2015 04:17 PM, Richard Weinberger wrote:
> Am 13.01.2015 um 23:54 schrieb Jens Axboe:
>>>> blk_rq_map_sg returns the number of entries actually mapped, which
>>>> might be smaller than the number passed in due to merging.
>>>
>>> Yep, but the ubi_sql has a fixed number of scatterlist entries, UBI_MAX_SG_COUNT.
>>> And I limit it also to that using: blk_queue_max_segments(dev->rq, UBI_MAX_SG_COUNT);
>>>
>>> Is there another reason why I should use the return value of blk_rq_map_sg()?
>>> Please also note that the UBI block driver is read-only.
>>
>> It can return less than what you asked for, if segments are coalesced.
>> Read/write, doesn't matter. You should always use the returned value as
>> the indication of how many segments to access in pdu->usgl.sg for data
>> transfer.
> 
> Sorry, I don't fully understand.
> 
> Currently the driver does:
> to_read = blk_rq_bytes(req);
> Then it fills pdu->usgl.sg up to to_read bytes
> and calls blk_mq_end_request().
> 
> If I understand you correctly it can happen that blk_rq_bytes() returns
> more bytes than blk_rq_map_sg() allocated, right?

No, the number of bytes will be the same, no magic is involved :-)

But lets say the initial request has 4 bios, with each 2 pages, for a
total of 8 segments. Lets further assume that the pages in each bio are
contiguous, so that blk_rq_map_sg() will map this to 4 sg elements, each
2xpages long.

Now, this may already be handled just fine, and you don't need to
update/store the actual sg count. I just looked at the source, and I'm
assuming it'll do the right thing (ubi_read_sg() will bump the active sg
element, when that size has been consumed), but I don't have
ubi_read_sg() in my tree to verify.

-- 
Jens Axboe