On Wed, Nov 08, 2017 at 10:11:02PM +0000, Mark Brown wrote:

On Wed, Nov 08, 2017 at 09:39:11PM +0000, Levin, Alexander (Sasha Levin) wrote:

...

...

On Wed, Nov 08, 2017 at 08:49:55PM +0000, Levin, Alexander (Sasha Levin) wrote: This seems like an optimization not a bug fix...

...

Hm, is it? I read it as "DMA is not being used at all even though we thought we're using it".

Yes, that's how I read it too.

...

Yes, the impact is "just" performance, but doesn't it result in quite a significat impact?

Only about double according to the initial commit adding DMA support which is frankly a bit disappointing although yeah, it's a big win. My worry is that if there's a problem with DMA on some device for which a fix wasn't backported (or where we're using a fallback) this could expose problems if we start using it. If you look at the history of the driver there's some quirks were added later on for example, and I didn't check the DMA controller drivers or anything and obviously can't see any out of tree code users may have.

*Probably* it doesn't break anything but since it's not fixing anything and the risk is data corruption I'd be much more comfortable with a more thorough risk analysis.

I'm considering these commits to be on the safer side because they're much older than the ones Greg usually grabs. There were no upstream fixes to this commit for 10 months now, and given the code changes upstream in that subsystem, this patch seemed to be safe to backport.