Re: [PATCH v2 1/4] iommu/arm-smmu-v3: Clean up address masking

[Will Deacon <will.deacon-5wv7dgnIgG8@public.gmane.org>]

Hi Robin,

On Tue, Feb 27, 2018 at 01:28:31PM +0000, Robin Murphy wrote:
> On 26/02/18 18:04, Will Deacon wrote:
> >On Thu, Dec 14, 2017 at 04:58:50PM +0000, Robin Murphy wrote:
> >>Before trying to add the SMMUv3.1 support for 52-bit addresses, make
> >>things bearable by cleaning up the various address mask definitions to
> >>use GENMASK_ULL() consistently. The fact that doing so reveals (and
> >>fixes) a latent off-by-one in Q_BASE_ADDR_MASK only goes to show what a
> >>jolly good idea it is...
> >>
> >>Tested-by: Nate Watterson <nwatters-sgV2jX0FEOL9JmXXK+q4OQ@public.gmane.org>
> >>Signed-off-by: Robin Murphy <robin.murphy-5wv7dgnIgG8@public.gmane.org>
> >>---
> >>
> >>v2: Clean up one more now-unnecessary linewrap
> >>
> >>  drivers/iommu/arm-smmu-v3.c | 53 ++++++++++++++++++---------------------------
> >>  1 file changed, 21 insertions(+), 32 deletions(-)
> >
> >Whilst I agree that using GENMASK is better, this patch does mean that the
> >driver is (more) inconsistent with its _MASK terminology in that you can't
> >generally tell whether a definition that ends in _MASK is shifted or not,
> >and this isn't even consistent for fields within the same register.
> 
> The apparently slightly-less-than-obvious internal consistency is that every
> mask used for an *address field* is now in-place, while other types of field
> are still handled as inconsistently as they were before. It should also be
> the case that every x_MASK without a corresponding x_SHIFT defined next to
> it is unshifted.
> 
> Either way it's certainly no *worse* than the current situation where
> address masks sometimes have a nonzero shift, sometimes have zero bits at
> the bottom and a shift of 0, and sometimes have no shift defined at all.
> 
> Thinking about it some more, the address masks should only ever be needed
> when *extracting* an address from a register/structure word, or validating
> them in the context of an address *before* inserting into a field - if we
> can't trust input to be correct then just silently masking off bits probably
> isn't the best idea either way - so IMHO there is plenty of contextual
> disambiguation too.
> 
> >Should we be using GENMASK/BIT for all fields instead and removing all of
> >the _SHIFT definitions?
> 
> I'm all aboard using BIT() consistently for single-bit boolean fields, but
> for multi-bit fields in general we do have to keep an explicit shift defined
> *somewhere* in order to make sensible use of the value, i.e. either:
> 
> 	val = (reg >> 22) & 0x1f;
> 	reg = (val & 0x1f) << 22;
> 
> or:
> 	val = (reg & 0x07c00000) >> 22;
> 	reg = (val << 22) & 0x07c00000;
> 
> [ but ideally not this mess we currently have in some places:
> 
> 	val = (reg & 0x1f << 22) >> 22;
> ]
> 
> Again, I'd gladly clean everything up to at least be self-consistent (and
> line up more with how we did things in SMMUv2) if you think it's worthwhile.
> Although I guess that means I'd get the job of fixing up future stable
> backport conflicts too ;)

I reckon it would be worth the cleanup since you're in the area. I don't
mind keeping the SHITF definitions where they're needed, but using BIT and
GENMASK wherever we can.

Will