On Sat, Aug 27 2016 at 11:09am -0400,
Mikulas Patocka [off-list ref] wrote:

On Fri, 26 Aug 2016, Mike Snitzer wrote:

quoted

On Thu, Aug 25 2016 at  4:13pm -0400,
Jens Axboe [off-list ref] wrote:

quoted

On 08/25/2016 12:34 PM, Mikulas Patocka wrote:

quoted

Device mapper can't split the bio in generic_make_request - it frees the
md->queue->bio_split bioset, to save one kernel thread per device. Device
mapper uses its own splitting mechanism.

So what is the final decision? - should device mapper split the big bio or
should bcache not submit big bios?

I think splitting big bios in the device mapper is better - simply because
it is much less code than reworking bcache to split bios internally.

BTW. In the device mapper, we have a layer dm-io, that was created to work
around bio size limitations - it accepts unlimited I/O request and splits
it to several bios. When bio size limitations are gone, we could simplify
dm-io too.

The patch from Ming Lei was applied for 4.8 the other day.

See linux-block commit:
4d70dca4eadf2f block: make sure a big bio is split into at most 256 bvecs
http://git.kernel.dk/cgit/linux-block/commit/?h=for-linus&id=4d70dca4eadf2f95abe389116ac02b8439c2d16c

But this patch won't work for device mapper, blk_bio_segment_split is 
called from blk_queue_split and device mapper doesn't use blk_queue_split 
(it can't because it frees queue->bio_split).

We still need these two patches:
https://www.redhat.com/archives/dm-devel/2016-May/msg00211.html
https://www.redhat.com/archives/dm-devel/2016-May/msg00210.html

I'm left wondering whether it'd be better to revert commit dbba42d8a9e
("dm: eliminate unused "bioset" process for each bio-based DM device")

And also look for other areas of DM that could leverage the block core's
relatively new splitting capabilities.

Otherwise, we'll just be sprinking more BIO_MAX_PAGES-based splitting
code in DM targets because DM is so "special" that it doesn't need the
block core's splitting (even though it clearly would now benefit).

Mike