Hello, Ming.

On Wed, Jul 06, 2016 at 09:10:00AM +0800, Ming Lei wrote:

quoted

Then we did some research and find that in kernel version 4.3 brought in
blk_queue_split() function to split the big size bio into several parts,
and some of them are calling the generic_make_request() again, this result
the bio been throttled more than once. so the actual bio sent to device is
less than we expected.

Except for blk_queue_split(), there are other(stacked) drivers which call
generic_make_request() too, such as drbd, dm, md and bcache.

So, blk-throtl already uses REQ_THROTTLED to avoid throttling the same
bio multiple times.  The problem seems that the flag isn't maintained
through clone.

quoted

We have checked the newest kernel of 4.7-rc5, this problem is still exist.

Based on this kind of situation, we propose a fix solution to add a flag bit
in bio to let the splited bio bypass the blk_queue_split(). Below is the patch
we used to fix this problem.

The splitted bio is just a fast-cloned bio(except for discard bio) and not very
special compared with other fast-cloned bio, which is quite common used.

So I guess what you need is to bypass BIO_CLONED bio for this purpose
since all fast-cloned bio shares the same bvec table of the source bio.

Depending on how a device handles a bio, that could allow bios to
bypass throttling entirely, no?  Wouldn't adding REQ_THROTTLED to
REQ_CLONE_MASK work?

Thanks.

-- 
tejun