Re: [PATCH 4/5] iommu/arm-smmu: Make way to add Qcom's smmu-500 errata handling

[Vivek Gautam]

Hi Robin,


On 8/14/2018 10:29 PM, Robin Murphy wrote:
> On 14/08/18 11:55, Vivek Gautam wrote:
> > Cleanup to re-use some of the stuff
> >
> > Signed-off-by: Vivek Gautam <vivek.gautam@xxxxxxxxxxxxxx>
> > ---
> > Â drivers/iommu/arm-smmu.c | 32 +++++++++++++++++++++++++-------
> > Â 1 file changed, 25 insertions(+), 7 deletions(-)
>
> I think the overall diffstat would be an awful lot smaller if the 
> erratum workaround just has its own readl_poll_timeout() as it does in 
> the vendor kernel. The burst-polling loop is for minimising latency in 
> high-throughput situations, and if you're in a workaround which has to 
> lock *every* register write and issue two firmware calls around each 
> sync I think you're already well out of that game.

Sorry for the delayed response. I was on vacation.
I will fix this in my next version by adding the separate 
read_poll_timeout() for the erratum WA.

> > diff --git a/drivers/iommu/arm-smmu.c b/drivers/iommu/arm-smmu.c
> > index 32e86df80428..75c146751c87 100644
> > --- a/drivers/iommu/arm-smmu.c
> > +++ b/drivers/iommu/arm-smmu.c
> > @@ -391,21 +391,31 @@ static void __arm_smmu_free_bitmap(unsigned 
> > long *map, int idx)
> > ÂÂÂÂÂ clear_bit(idx, map);
> > Â }
> > Â -/* Wait for any pending TLB invalidations to complete */
> > -static void __arm_smmu_tlb_sync(struct arm_smmu_device *smmu,
> > -ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ void __iomem *sync, void __iomem *status)
> > +static int __arm_smmu_tlb_sync_wait(void __iomem *status)
> > Â {
> > ÂÂÂÂÂ unsigned int spin_cnt, delay;
> > Â -ÂÂÂ writel_relaxed(0, sync);
> > ÂÂÂÂÂ for (delay = 1; delay < TLB_LOOP_TIMEOUT; delay *= 2) {
> > ÂÂÂÂÂÂÂÂÂ for (spin_cnt = TLB_SPIN_COUNT; spin_cnt > 0; spin_cnt--) {
> > ÂÂÂÂÂÂÂÂÂÂÂÂÂ if (!(readl_relaxed(status) & sTLBGSTATUS_GSACTIVE))
> > -ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ return;
> > +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ return 0;
> > ÂÂÂÂÂÂÂÂÂÂÂÂÂ cpu_relax();
> > ÂÂÂÂÂÂÂÂÂ }
> > ÂÂÂÂÂÂÂÂÂ udelay(delay);
> > ÂÂÂÂÂ }
> > +
> > +ÂÂÂ return -EBUSY;
> > +}
> > +
> > +/* Wait for any pending TLB invalidations to complete */
> > +static void __arm_smmu_tlb_sync(struct arm_smmu_device *smmu,
> > +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ void __iomem *sync, void __iomem *status)
> > +{
> > +ÂÂÂ writel_relaxed(0, sync);
> > +
> > +ÂÂÂ if (!__arm_smmu_tlb_sync_wait(status))
> > +ÂÂÂÂÂÂÂ return;
> > +
> > ÂÂÂÂÂ dev_err_ratelimited(smmu->dev,
> > ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ "TLB sync timed out -- SMMU may be deadlocked\n");
> > Â }
> > @@ -461,8 +471,9 @@ static void arm_smmu_tlb_inv_context_s2(void 
> > *cookie)
> > ÂÂÂÂÂ arm_smmu_tlb_sync_global(smmu);
> > Â }
> > Â -static void arm_smmu_tlb_inv_range_nosync(unsigned long iova, 
> > size_t size,
> > -ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ size_t granule, bool leaf, void *cookie)
> > +static void __arm_smmu_tlb_inv_range_nosync(unsigned long iova, 
> > size_t size,
> > +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ size_t granule, bool leaf,
> > +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ void *cookie)
> > Â {
> > ÂÂÂÂÂ struct arm_smmu_domain *smmu_domain = cookie;
> > ÂÂÂÂÂ struct arm_smmu_cfg *cfg = &smmu_domain->cfg;
> > @@ -498,6 +509,13 @@ static void 
> > arm_smmu_tlb_inv_range_nosync(unsigned long iova, size_t size,
> > ÂÂÂÂÂ }
> > Â }
> > Â +static void arm_smmu_tlb_inv_range_nosync(unsigned long iova, 
> > size_t size,
> > +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ size_t granule, bool leaf,
> > +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ void *cookie)
> > +{
> > +ÂÂÂ __arm_smmu_tlb_inv_range_nosync(iova, size, granule, leaf, cookie);
> > +}
> > +
>
> AFAICS even after patch #5 this does absolutely nothing except make 
> the code needlessly harder to read :(

Sure, I will rather call arm_smmu_tlb_inv_range_nosync() from
qcom_errata_tlb_inv_range_nosync() then make this change.
Thanks for the review.

Best regards
Vivek

> Robin.
>
> > Â /*
> > ÂÂ * On MMU-401 at least, the cost of firing off multiple TLBIVMIDs 
> > appears
> > ÂÂ * almost negligible, but the benefit of getting the first one in 
> > as far ahead