Re: [PATCH] Revert "iommu/io-pgtable-arm: Optimise non-coherent unmap"

[Will Deacon]

On Fri, Sep 06, 2024 at 04:25:19PM +0100, Robin Murphy wrote:
> On 06/09/2024 11:56 am, Will Deacon wrote:
> > On Thu, Sep 05, 2024 at 05:27:28PM +0100, Robin Murphy wrote:
> > > On 05/09/2024 4:53 pm, Will Deacon wrote:
> > > > On Thu, Sep 05, 2024 at 05:49:56AM -0700, Rob Clark wrote:
> > > > > From: Rob Clark <robdclark@xxxxxxxxxxxx>
> > > > > 
> > > > > This reverts commit 85b715a334583488ad7fbd3001fe6fd617b7d4c0.
> > > > > 
> > > > > It was causing gpu smmu faults on x1e80100.
> > > > > 
> > > > > I _think_ what is causing this is the change in ordering of
> > > > > __arm_lpae_clear_pte() (dma_sync_single_for_device() on the pgtable
> > > > > memory) and io_pgtable_tlb_flush_walk().  I'm not entirely sure how
> > > > > this patch is supposed to work correctly in the face of other
> > > > > concurrent translations (to buffers unrelated to the one being
> > > > > unmapped(), because after the io_pgtable_tlb_flush_walk() we can have
> > > > > stale data read back into the tlb.
> > > > > 
> > > > > Signed-off-by: Rob Clark <robdclark@xxxxxxxxxxxx>
> > > > > ---
> > > > >    drivers/iommu/io-pgtable-arm.c | 31 ++++++++++++++-----------------
> > > > >    1 file changed, 14 insertions(+), 17 deletions(-)
> > > > 
> > > > Please can you try the diff below, instead?
> > > 
> > > Given that the GPU driver's .tlb_add_page is a no-op, I can't see this
> > > making a difference. In fact, given that msm_iommu_pagetable_unmap() still
> > > does a brute-force iommu_flush_iotlb_all() after io-pgtable returns, and in
> > > fact only recently made .tlb_flush_walk start doing anything either for the
> > > sake of the map path, I'm now really wondering how this patch has had any
> > > effect at all... :/
> > 
> > Hmm, yup. Looks like Rob has come back to say the problem lies elsewhere
> > anyway.
> > 
> > One thing below though...
> > 
> > > > 
> > > > Will
> > > > 
> > > > --->8
> > > > 
> > > > diff --git a/drivers/iommu/io-pgtable-arm.c b/drivers/iommu/io-pgtable-arm.c
> > > > index 0e67f1721a3d..0a32e9499e2c 100644
> > > > --- a/drivers/iommu/io-pgtable-arm.c
> > > > +++ b/drivers/iommu/io-pgtable-arm.c
> > > > @@ -672,7 +672,7 @@ static size_t __arm_lpae_unmap(struct arm_lpae_io_pgtable *data,
> > > >                   /* Clear the remaining entries */
> > > >                   __arm_lpae_clear_pte(ptep, &iop->cfg, i);
> > > > -               if (gather && !iommu_iotlb_gather_queued(gather))
> > > > +               if (!iommu_iotlb_gather_queued(gather))
> > > 
> > > Note that this would reintroduce the latent issue which was present
> > > originally, wherein iommu_iotlb_gather_queued(NULL) is false, but if we
> > > actually allow a NULL gather to be passed to io_pgtable_tlb_add_page() it
> > > may end up being dereferenced (e.g. in arm-smmu-v3).
> > 
> > I think there is still something to fix here. arm_lpae_init_pte() can
> > pass a NULL gather to __arm_lpae_unmap() and I don't think skipping the
> > invalidation is correct in that case. Either the drivers need to handle
> > that or we shouldn't be passing NULL.
> > 
> > What do you think?
> 
> The subtlety there is that in that case it's always a non-leaf PTE, so all
> that goes back to the driver is io_pgtable_tlb_flush_walk() and the gather
> is never used.

Beautiful...

Will