[PATCH v2 2/2] iommu/arm-smmu: Do not access non-existing SMR registers

Hi Olav,

On Wed, Aug 06, 2014 at 05:44:38PM +0100, Olav Haugan wrote:

On 8/6/2014 3:19 AM, Will Deacon wrote:

quoted

On Mon, Aug 04, 2014 at 07:01:03PM +0100, Olav Haugan wrote:

quoted

The SMR registers do not exist when stream matching is not
supported by the hardware. Avoid writing to this register if not needed.

Signed-off-by: Olav Haugan <redacted>
---
 drivers/iommu/arm-smmu.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/drivers/iommu/arm-smmu.c b/drivers/iommu/arm-smmu.c
index c16431f..1f3a5b3 100644
--- a/drivers/iommu/arm-smmu.c
+++ b/drivers/iommu/arm-smmu.c

@@ -1731,7 +1731,10 @@ static void arm_smmu_device_reset(struct arm_smmu_device *smmu)
 
 	/* Mark all SMRn as invalid and all S2CRn as bypass */
 	for (i = 0; i < smmu->num_mapping_groups; ++i) {
-		writel_relaxed(~SMR_VALID, gr0_base + ARM_SMMU_GR0_SMR(i));
+		if (smmu->features & ARM_SMMU_FEAT_STREAM_MATCH) {
+			writel_relaxed(~SMR_VALID,
+					gr0_base + ARM_SMMU_GR0_SMR(i));
+		}
 		writel_relaxed(S2CR_TYPE_BYPASS,
 			gr0_base + ARM_SMMU_GR0_S2CR(i));

smmu->num_mapping_groups should be zero for an SMMU that doesn't include
the SMR registers, so I don't think this change is needed. Are you seeing
problems with real hardware?

Yes, you are correct. However, since that is the case we wouldn't be
setting the S2CR registers to bypass then? Seems like
num_mappings_groups should be initialized regardless whether stream
matching or stream indexing is used. SMMU_IDR0.NUMSMRG specifies the
number of stream mapping register groups (Section 2.5.2 of the ARM
SMMUv1-v2 spec). So with stream indexing support this register should
still tell us how many S2CR registers exist?

Hmm, I'm checking this with the architects because the TRMs aren't exactly
clear. The NUMSMRG works for stream-indexing (i.e. reports the number of
S2CRs), then all we have to do is change the above ~SMR_VALID to 0x0, since
those registers will be SBZP.

As far as I can tell there are no other register telling us how many S2CR
registers exist. That also brings up another point that there is no check
in the code to ensure we are not trying to program more than the available
S2CR registers when we use stream indexing.

On an SMMU using stream-indexing, if the StreamID goes off the end of the
S2CRs that's a fairly serious hardware configuration issue which I don't
think Linux is in a position to handle. I agree that a warning wouldn't
hurt on device add/attach though.

Will