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

[Richard Weinberger <richard@nod.at>]

[-- Attachment #1: Type: text/plain, Size: 1466 bytes --]

Ezequiel,

Am 02.11.2014 um 22:52 schrieb Ezequiel Garcia:
> Maybe you can explain a bit better what's this all about?

In short, blk-mq is the future and the current blk interface will be legacy. :-)
Christoph asked me to convert the MTD block drivers to blk-mq.

> Both the commit that introduces blk-mq and the paper on it talk about
> high IOPS devices, multi-core, NUMA systems. I'm not sure this is the
> case for UBI-based devices.
> 
> Probably some numbers would help us decide. Does the patch increases the
> dynamic memory footprint? Is there any performance benefit?

I did a very rough micro benchmark:

root@(none):~# dd if=/dev/ubiblock0_0 of=/dev/null bs=1M
121+1 records in
121+1 records out
127420416 bytes (127 MB) copied, 1.59056 s, 80.1 MB/s

vs.

root@(none):~# dd if=/dev/ubiblock0_0 of=/dev/null bs=1M
121+1 records in
121+1 records out
127420416 bytes (127 MB) copied, 0.916117 s, 139 MB/s

So, yes there is a performance gain.

> I kind of like the negative diffstat, but the code doesn't look cleaner
> or simpler.
> 
> In other words, we need a good reason before we agree on making this
> "zen style" driver more complex.

After reading my patch again I think we could move ubiblock_read_to_sg()
to kapi.c or io.c. It is rather generic and maybe we can tun more UBI users to
scattergather such that less vmalloc()s are needed.

This would also make the diffstat nicer...

Thanks,
//richard


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 836 bytes --]