On Thu, Aug 15, 2019 at 01:44:39PM +0800, Zhen Lei wrote:
> When (smmu_domain->smmu->features & ARM_SMMU_FEAT_ATS) is true, even if a
> smmu domain does not contain any ats master, the operations of
> arm_smmu_atc_inv_to_cmd() and lock protection in arm_smmu_atc_inv_domain()
> are always executed. This will impact performance, especially in
> multi-core and stress scenarios. For my FIO test scenario, about 8%
> performance reduced.
>
> In fact, we can use a struct member to record how many ats masters that
> the smmu contains. And check that without traverse the list and check all
> masters one by one in the lock protection.
>
> Fixes: 9ce27afc0830 ("iommu/arm-smmu-v3: Add support for PCI ATS")
> Signed-off-by: Zhen Lei <[email protected]>
> ---
> drivers/iommu/arm-smmu-v3.c | 14 +++++++++++++-
> 1 file changed, 13 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/iommu/arm-smmu-v3.c b/drivers/iommu/arm-smmu-v3.c
> index 29056d9bb12aa01..154334d3310c9b8 100644
> --- a/drivers/iommu/arm-smmu-v3.c
> +++ b/drivers/iommu/arm-smmu-v3.c
> @@ -631,6 +631,7 @@ struct arm_smmu_domain {
>
> struct io_pgtable_ops *pgtbl_ops;
> bool non_strict;
> + int nr_ats_masters;
>
> enum arm_smmu_domain_stage stage;
> union {
> @@ -1531,7 +1532,16 @@ static int arm_smmu_atc_inv_domain(struct
> arm_smmu_domain *smmu_domain,
> struct arm_smmu_cmdq_ent cmd;
> struct arm_smmu_master *master;
>
> - if (!(smmu_domain->smmu->features & ARM_SMMU_FEAT_ATS))
> + /*
> + * The protectiom of spinlock(&iommu_domain->devices_lock) is omitted.
> + * Because for a given master, its map/unmap operations should only be
> + * happened after it has been attached and before it has been detached.
> + * So that, if at least one master need to be atc invalidated, the
> + * value of smmu_domain->nr_ats_masters can not be zero.
> + *
> + * This can alleviate performance loss in multi-core scenarios.
> + */
I find this reasoning pretty dubious, since I think you're assuming that
an endpoint cannot issue speculative ATS translation requests once its
ATS capability is enabled. That said, I think it also means we should enable
ATS in the STE *before* enabling it in the endpoint -- the current logic
looks like it's the wrong way round to me (including in detach()).
Anyway, these speculative translations could race with a concurrent unmap()
call and end up with the ATC containing translations for unmapped pages,
which I think we should try to avoid.
Did the RCU approach not work out? You could use an rwlock instead as a
temporary bodge if the performance doesn't hurt too much.
Alternatively... maybe we could change the attach flow to do something
like:
enable_ats_in_ste(master);
enable_ats_at_pcie_endpoint(master);
spin_lock(devices_lock)
add_to_device_list(master);
nr_ats_masters++;
spin_unlock(devices_lock);
invalidate_atc(master);
in which case, the concurrent unmapper will be doing something like:
issue_tlbi();
smp_mb();
if (READ_ONCE(nr_ats_masters)) {
...
}
and I *think* that means that either the unmapper will see the
nr_ats_masters update and perform the invalidation, or they'll miss
the update but the attach will invalidate the ATC /after/ the TLBI
in the command queue.
Also, John's idea of converting this stuff over to my command batching
mechanism should help a lot if we can defer this to sync time using the
gather structure. Maybe an rwlock would be alright for that. Dunno.
Will
