On Fri, Apr 27, 2018 at 01:13:47PM +0300, Dmitry Osipenko wrote:
> On 27.04.2018 12:34, Thierry Reding wrote:
> > On Mon, Apr 09, 2018 at 10:28:31PM +0300, Dmitry Osipenko wrote:
> > [...]
> >> diff --git a/drivers/memory/tegra/mc.c b/drivers/memory/tegra/mc.c
> > [...]
> >> +#define MC_GART_ERROR_REQ 0x30
> >> +#define MC_DECERR_EMEM_OTHERS_STATUS 0x58
> >> +#define MC_SECURITY_VIOLATION_STATUS 0x74
> > [...]
> >> diff --git a/drivers/memory/tegra/mc.h b/drivers/memory/tegra/mc.h
> > [...]
> >> @@ -21,19 +21,30 @@
> >> #define MC_INT_INVALID_SMMU_PAGE (1 << 10)
> >> #define MC_INT_ARBITRATION_EMEM (1 << 9)
> >> #define MC_INT_SECURITY_VIOLATION (1 << 8)
> >> +#define MC_INT_INVALID_GART_PAGE (1 << 7)
> >> #define MC_INT_DECERR_EMEM (1 << 6)
> >>
> >> static inline u32 mc_readl(struct tegra_mc *mc, unsigned long offset)
> >> {
> >> + if (mc->regs2 && offset >= 0x24)
> >> + return readl(mc->regs2 + offset - 0x3c);
> >
> > I'm still not sure how this is supposed to work. If we pass in
> > MC_GART_ERROR_REQ as offset into mc_readl(), then the condition above
> > will be true (0x30 >= 0x24) but then the new offset will be computed
> > and we end up with:
> >
> > return readl(mc->regs2 + 0x30 - 0x3c);
> >
> > which means we'll be adding a negative offset (or rather a very large
> > offset because it will wrap around).
>
> Indeed! Thank you for pointing at it again, now I see the issue. It probably
> works because actual registers mapping is aligned to page(?) size and adding the
> large offset with wraparound is equal to subtraction.
>
> That register belongs to the GART and we can't simply move interrupt handling to
> the GART driver because status register is within the MC in device tree. We can
> omit reading of MC_GART_ERROR_REQ and simply report GART page fault for the
> starter and then reorganize drivers by making MC driver MFD and GART its
> sub-device, what do you think?
Sounds like a good idea. Can you send a fix on top of this that I can
squash into this when applying?
As for integrating GART with MC, I'd prefer something that doesn't use
MFD but rather does something similar to what we have for the SMMU. I
think that's simpler to do and has less boilerplate. I think it's also
warranted because the MC and GART are very tightly coupled, so an MFD
would be slightly over-engineered, in my opinion.
Thierry
Attachment:
signature.asc
Description: PGP signature
