On 2023-08-23 00:04, Nicolin Chen wrote:
> Hi Robin,
>
> On Tue, Aug 22, 2023 at 09:32:26AM -0700, Nicolin Chen wrote:
>> On Tue, Aug 22, 2023 at 10:30:35AM +0100, Robin Murphy wrote:
>>
>>>> diff --git a/drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3.c b/drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3.c
>>>> index d6c647e1eb01..3f0db30932bd 100644
>>>> --- a/drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3.c
>>>> +++ b/drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3.c
>>>> @@ -1897,7 +1897,14 @@ static void __arm_smmu_tlb_inv_range(struct arm_smmu_cmdq_ent *cmd,
>>>> if (!size)
>>>> return;
>>>>
>>>> - if (smmu->features & ARM_SMMU_FEAT_RANGE_INV) {
>>>> + if (!(smmu->features & ARM_SMMU_FEAT_RANGE_INV)) {
>>>> + /*
>>>> + * When the size reaches a threshold, replace per-granule TLBI
>>>> + * commands with one single per-asid or per-vmid TLBI command.
>>>> + */
>>>> + if (size >= granule * smmu_domain->max_tlbi_ops)
>>>> + return arm_smmu_tlb_inv_domain(smmu_domain);
>>>
>>> This looks like it's at the wrong level - we should have figured this
>>> out before we got as far as low-level command-building. I'd have thought
>>> it would be a case of short-circuiting directly from
>>> arm_smmu_tlb_inv_range_domain() to arm_smmu_tlb_inv_context().
>>
>> OK, I could do that. We would have copies of this same routine
>> though. Also, the shortcut applies to !ARM_SMMU_FEAT_RANGE_INV
>> cases only, so this function feels convenient to me.
>
> I was trying to say that we would need the same piece in both
> arm_smmu_tlb_inv_range_domain() and arm_smmu_tlb_inv_range_asid(),
> though the latter one only needs to call arm_smmu_tlb_inv_asid().
Its not like arm_smmu_tlb_inv_range_asid() doesn't already massively
overlap with arm_smmu_tlb_inv_range_domain() anyway, so a little further
duplication hardly seems like it would hurt. Checking
ARM_SMMU_FEAT_RANGE_INV should be cheap (otherwise we'd really want to
split __arm_smmu_tlb_inv_range() into separate RIL vs. non-RIL versions
to avoid having it in the loop), and it makes the intent clear. What I
just really don't like is a flow where we construct a specific command,
then call the low-level function to issue it, only that function then
actually jumps back out into another high-level function which
constructs a *different* command. This code is already a maze of twisty
little passages...
> Also, arm_smmu_tlb_inv_context() does a full range ATC invalidation
> instead of a given range like what arm_smmu_tlb_inv_range_domain()
> currently does. So, it might be a bit overkill.
>
> Combining all your comments, we'd have something like this:
TBH I'd be inclined to refactor a bit harder, maybe break out some
VMID-based helpers for orthogonality, and aim for a flow like:
if (over threshold)
tlb_inv_domain()
else if (stage 1)
tlb_inv_range_asid()
else
tlb_inv_range_vmid()
atc_inv_range()
or possibly if you prefer:
if (stage 1) {
if (over threshold)
tlb_inv_asid()
else
tlb_inv_range_asid()
} else {
if (over threshold)
tlb_inv_vmid()
else
tlb_inv_range_vmid()
}
atc_inv_range()
where the latter maybe trades more verbosity for less duplication
overall - I'd probably have to try both to see which looks nicer in the
end. And obviously if there's any chance of inventing a clear and
consistent naming scheme in the process, that would be lovely.
Thanks,
Robin.
> -------------------------------------------------------------------
> diff --git a/drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3.c b/drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3.c
> index 7614739ea2c1..2967a6634c7c 100644
> --- a/drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3.c
> +++ b/drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3.c
> @@ -1937,12 +1937,22 @@ static void arm_smmu_tlb_inv_range_domain(unsigned long iova, size_t size,
> size_t granule, bool leaf,
> struct arm_smmu_domain *smmu_domain)
> {
> + struct io_pgtable_cfg *cfg =
> + &io_pgtable_ops_to_pgtable(smmu_domain->pgtbl_ops)->cfg;
> struct arm_smmu_cmdq_ent cmd = {
> .tlbi = {
> .leaf = leaf,
> },
> };
>
> + /*
> + * If the given size is too large that would end up with too many TLBI
> + * commands in CMDQ, short circuit directly to a full invalidation
> + */
> + if (!(smmu_domain->smmu->features & ARM_SMMU_FEAT_RANGE_INV) &&
> + size >= granule * (1UL << cfg->bits_per_level))
> + return arm_smmu_tlb_inv_context(smmu_domain);
> +
> if (smmu_domain->stage == ARM_SMMU_DOMAIN_S1) {
> cmd.opcode = smmu_domain->smmu->features & ARM_SMMU_FEAT_E2H ?
> CMDQ_OP_TLBI_EL2_VA : CMDQ_OP_TLBI_NH_VA;
> @@ -1964,6 +1974,8 @@ void arm_smmu_tlb_inv_range_asid(unsigned long iova, size_t size, int asid,
> size_t granule, bool leaf,
> struct arm_smmu_domain *smmu_domain)
> {
> + struct io_pgtable_cfg *cfg =
> + &io_pgtable_ops_to_pgtable(smmu_domain->pgtbl_ops)->cfg;
> struct arm_smmu_cmdq_ent cmd = {
> .opcode = smmu_domain->smmu->features & ARM_SMMU_FEAT_E2H ?
> CMDQ_OP_TLBI_EL2_VA : CMDQ_OP_TLBI_NH_VA,
> @@ -1973,6 +1985,14 @@ void arm_smmu_tlb_inv_range_asid(unsigned long iova, size_t size, int asid,
> },
> };
>
> + /*
> + * If the given size is too large that would end up with too many TLBI
> + * commands in CMDQ, short circuit directly to a full invalidation
> + */
> + if (!(smmu_domain->smmu->features & ARM_SMMU_FEAT_RANGE_INV) &&
> + size >= granule * (1UL << cfg->bits_per_level))
> + return arm_smmu_tlb_inv_asid(smmu_domain->smmu, asid);
> +
> __arm_smmu_tlb_inv_range(&cmd, iova, size, granule, smmu_domain);
> }
>
> -------------------------------------------------------------------
>
> You're sure that you prefer this, right?
>
> Thanks
> Nicolin