Re: [PATCH] iommu/arm-smmu: Clear global and context bank fault status registers

[Andreas Herrmann <andreas.herrmann-bsGFqQB8/DxBDgjK7y7TUQ@public.gmane.org>]

On Mon, Sep 30, 2013 at 02:30:06PM -0400, Will Deacon wrote:
> On Mon, Sep 30, 2013 at 06:17:16PM +0100, Andreas Herrmann wrote:
> > On Mon, Sep 30, 2013 at 12:06:15PM -0400, Will Deacon wrote:
> > > On Mon, Sep 30, 2013 at 02:56:21PM +0100, Andreas Herrmann wrote:
> > > > 
> > > > After reset these registers have unknown values.
> > > > This might cause problems when evaluating SMMU_GFSR and/or SMMU_CB_FSR
> > > > in handlers for combined interrupts.
> > > > 
> > > > Signed-off-by: Andreas Herrmann <andreas.herrmann-bsGFqQB8/DxBDgjK7y7TUQ@public.gmane.org>
> > > > ---
> > > >  drivers/iommu/arm-smmu.c |   27 ++++++++++++++++++++-------
> > > >  1 file changed, 20 insertions(+), 7 deletions(-)
> > > > 
> > > > diff --git a/drivers/iommu/arm-smmu.c b/drivers/iommu/arm-smmu.c
> > > > index 579b6f8..cbbf597 100644
> > > > --- a/drivers/iommu/arm-smmu.c
> > > > +++ b/drivers/iommu/arm-smmu.c
> > > > @@ -631,6 +631,12 @@ static irqreturn_t arm_smmu_global_fault(int irq, void *dev)
> > > >  	return IRQ_HANDLED;
> > > >  }
> > > >  
> > > > +static void arm_smmu_clear_cb_fsr(struct arm_smmu_device *smmu, u8 cbndx)
> > > > +{
> > > > +	void __iomem *cb_base = ARM_SMMU_CB_BASE(smmu) + ARM_SMMU_CB(smmu, cbndx);
> > > > +	writel_relaxed(FSR_FAULT, cb_base + ARM_SMMU_CB_FSR);
> > > > +}
> > > 
> > > Hmm, why not just stick this in arm_smmu_init_context_bank...
> > 
> > Because we should clear the FSR before we call request_irq.
> > Otherwise we might handle interrupts although the context bank is not
> > enabled.
> > 
> > Moving request_irq after arm_smmu_init_context_bank is not optimal
> > either. (We should have configured the context interrupt before
> > translation is enabled. Otherwise it's possible to miss a fault.)
> 
> How would you miss a fault?

> If the device can start issuing transactions before the SMMU has set
> up the mapping, then there's a race in the caller code which we
> shouldn't attempt to resolve here.

Broken device, broken driver code (maybe violating dma-api)
whatsoever. All that's needed is a device that is already doing DMA
when we enable SMMU handling for its transactions.
Yes, normally this should not happen. But if it happens we get a fault
and better should handle it.

I think, it's somehow logical to have fault handling set up before
fault reporting is switched on for a context bank.


Andreas