On 2021-07-08 09:08, Joerg Roedel wrote:

On Wed, Jul 07, 2021 at 01:00:13PM -0700, Doug Anderson wrote:

quoted

a) Nothing is inherently broken with my current approach.

b) My current approach doesn't make anybody terribly upset even if
nobody is totally in love with it.

Well, no, sorry :)

I don't think it is a good idea to allow drivers to opt-out of the
strict-setting. This is a platform or user decision, and the driver
should accept whatever it gets.

So the real question is still why strict is the default setting and how
to change that. Or document for the users that want performance how to
change the setting, so that they can decide.

As I mentioned before, conceptually I think this very much belongs in 
sysfs as a user decision. We essentially have 4 levels of "strictness":

1: DMA domain with bounce pages
2: DMA domain
3: DMA domain with flush queue
4: Identity domain

The "make this device go faster because I trust it" use-case is why we 
have the sysfs interface to switch between 2 and 4, so it's entirely 
logical to have the intermediate option as well for when 3 is "faster" 
enough while still giving a bit more peace of mind than full-on bypass.

Making it a platform-specific decision that's hidden in a driver - 
arm-smmu-qcom can be considered a dumping ground of detailed platform 
knowledge ;) - happens to work as a reasonable compromise for this 
specific case, but I concur that it could be viewed as setting a 
precedent for other cases which we definitely aren't as reasonable.

I've been thinking about the sysfs thing some more, and since it's 
Friday afternoon and I can't concentrate on what I'm supposed to be 
doing anyway, let's see how far I can get by Monday...

Robin.