When installing a block mapping, we unconditionally overwrite a non-leaf
PTE if we find one. However, this can cause a problem if the following
sequence of events occur:

  (1) iommu_map called for a 4k (i.e. PAGE_SIZE) mapping at some address
      - We initialise the page table all the way down to a leaf entry
      - No TLB maintenance is required, because we're going from invalid
        to valid.

  (2) iommu_unmap is called on the mapping installed in (1)
      - We walk the page table to the final (leaf) entry and zero it
      - We only changed a valid leaf entry, so we invalidate leaf-only

  (3) iommu_map is called on the same address as (1), but this time for
      a 2MB (i.e. BLOCK_SIZE) mapping)
      - We walk the page table down to the penultimate level, where we
        find a table entry
      - We overwrite the table entry with a block mapping and return
        without any TLB maintenance and without freeing the memory used
        by the now-orphaned table.

This last step can lead to a walk-cache caching the overwritten table
entry, causing unexpected faults when the new mapping is accessed by a
device. One way to fix this would be to collapse the page table when
freeing the last page at a given level, but this would require expensive
iteration on every map call. Instead, this patch detects the case when
we are overwriting a table entry and explicitly unmaps the table first,
which takes care of both freeing and TLB invalidation.

Cc: <stable@vger.kernel.org>
Reported-by: Brian Starkey <brian.starkey@arm.com>
Tested-by: Brian Starkey <brian.starkey@arm.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Joerg Roedel <jroedel@suse.de>