Re: [RFC PATCH v1] iommu/arm-smmu: Allow domains to choose a context bank

[Jordan Crouse <jcrouse@codeaurora.org>]

On Tue, Feb 18, 2020 at 10:19:53AM -0800, Rob Clark wrote:
> On Tue, Jan 28, 2020 at 2:34 PM Jordan Crouse <jcrouse@codeaurora.org> wrote:
> >
> > Domains which are being set up for split pagetables usually want to be
> > on a specific context bank for hardware reasons. Force the context
> > bank for domains with the split-pagetable quirk to context bank 0.
> > If context bank 0 is taken, move that context bank to another unused
> > bank and rewrite the stream matching registers accordingly.
> 
> Is the only reason for dealing with the case that bank 0 is already in
> use, due to the DMA domain that gets setup before driver probes?

Right. On Adreno GPUs only one context bank at a time is accessible from the
GPU through an aperture which defaults to context bank 0 and as you might
expect, the aperture controls are protected by the secure world on AC enabled
targets.

Some of the newer targets have a SCM call to switch the aperture but for all the
currently merged platforms we are forced to use context bank 0.

> I'm kinda thinking that we need to invent a way to unwind/detatch the
> DMA domain, and unhook the iommu-dmaops, since this seems to already
> be already causing problems with dma-bufs imported from other drivers
> (who expect that dma_map_*(), with the importing device's dev ptr,
> will do something sane.

That could work, assuming that we could guarantee that our new replacement
domain got the context bank we wanted.

Jordan

> >
> > This is be used by [1] and [2] to leave context bank 0 open so that
> > the Adreno GPU can program it.
> >
> > [1] https://lists.linuxfoundation.org/pipermail/iommu/2020-January/041438.html
> > [2] https://lists.linuxfoundation.org/pipermail/iommu/2020-January/041444.html
> >
> > Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
> > ---
> >
> >  drivers/iommu/arm-smmu.c | 63 +++++++++++++++++++++++++++++++++++++++++++++---
> >  1 file changed, 59 insertions(+), 4 deletions(-)
> >
> > diff --git a/drivers/iommu/arm-smmu.c b/drivers/iommu/arm-smmu.c
> > index 85a6773..799a254 100644
> > --- a/drivers/iommu/arm-smmu.c
> > +++ b/drivers/iommu/arm-smmu.c
> > @@ -254,6 +254,43 @@ static int __arm_smmu_alloc_bitmap(unsigned long *map, int start, int end)
> >         return idx;
> >  }
> >
> > +static void arm_smmu_write_s2cr(struct arm_smmu_device *smmu, int idx);
> > +
> > +static int __arm_smmu_alloc_cb(struct arm_smmu_device *smmu, int start,
> > +               int target)
> > +{
> > +       int new, i;
> > +
> > +       /* Allocate a new context bank id */
> > +       new = __arm_smmu_alloc_bitmap(smmu->context_map, start,
> > +               smmu->num_context_banks);
> > +
> > +       if (new < 0)
> > +               return new;
> > +
> > +       /* If no target is set or we actually got the bank index we wanted */
> > +       if (target == -1 || new == target)
> > +               return new;
> > +
> > +       /* Copy the context configuration to the new index */
> > +       memcpy(&smmu->cbs[new], &smmu->cbs[target], sizeof(*smmu->cbs));
> > +       smmu->cbs[new].cfg->cbndx = new;
> > +
> > +       /* FIXME: Do we need locking here? */
> > +       for (i = 0; i < smmu->num_mapping_groups; i++) {
> > +               if (smmu->s2crs[i].cbndx == target) {
> > +                       smmu->s2crs[i].cbndx = new;
> > +                       arm_smmu_write_s2cr(smmu, i);
> > +               }
> > +       }
> > +
> > +       /*
> > +        * FIXME: Does getting here imply that 'target' is already set in the
> > +        * context_map?
> > +        */
> > +       return target;
> > +}
> > +
> >  static void __arm_smmu_free_bitmap(unsigned long *map, int idx)
> >  {
> >         clear_bit(idx, map);
> > @@ -770,6 +807,7 @@ static int arm_smmu_init_domain_context(struct iommu_domain *domain,
> >         struct arm_smmu_domain *smmu_domain = to_smmu_domain(domain);
> >         struct arm_smmu_cfg *cfg = &smmu_domain->cfg;
> >         unsigned long quirks = 0;
> > +       int forcecb = -1;
> >
> >         mutex_lock(&smmu_domain->init_mutex);
> >         if (smmu_domain->smmu)
> > @@ -844,8 +882,25 @@ static int arm_smmu_init_domain_context(struct iommu_domain *domain,
> >                          * SEP_UPSTREAM so we don't need to reduce the size of
> >                          * the ias to account for the sign extension bit
> >                          */
> > -                       if (smmu_domain->split_pagetables)
> > -                               quirks |= IO_PGTABLE_QUIRK_ARM_TTBR1;
> > +                       if (smmu_domain->split_pagetables) {
> > +                               /*
> > +                                * If split pagetables are enabled, assume that
> > +                                * the user's intent is to use per-instance
> > +                                * pagetables which, at least on a QCOM target,
> > +                                * means that this domain should be on context
> > +                                * bank 0.
> > +                                */
> > +
> > +                               /*
> > +                                * If we can't force to context bank 0 then
> > +                                * don't bother enabling split pagetables which
> > +                                * then would not allow aux domains
> > +                                */
> > +                               if (start == 0) {
> > +                                       forcecb = 0;
> > +                                       quirks |= IO_PGTABLE_QUIRK_ARM_TTBR1;
> > +                               }
> > +                       }
> >                 } else if (cfg->fmt == ARM_SMMU_CTX_FMT_AARCH32_L) {
> >                         fmt = ARM_32_LPAE_S1;
> >                         ias = min(ias, 32UL);
> > @@ -883,8 +938,8 @@ static int arm_smmu_init_domain_context(struct iommu_domain *domain,
> >                 ret = -EINVAL;
> >                 goto out_unlock;
> >         }
> > -       ret = __arm_smmu_alloc_bitmap(smmu->context_map, start,
> > -                                     smmu->num_context_banks);
> > +
> > +       ret = __arm_smmu_alloc_cb(smmu, start, forcecb);
> >         if (ret < 0)
> >                 goto out_unlock;
> >
> > --
> > 2.7.4

-- 
The Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
a Linux Foundation Collaborative Project