iommu/vt-d: Avoid superfluous IOTLB tracking in lazy mode

Intel IOMMU driver implements IOTLB flush queue with domain selective or PASID selective invalidations. In this case there's no need to track IOVA page range and sync IOTLBs, which may cause significant performance hit.

This patch adds a check to avoid IOVA gather page and IOTLB sync for the lazy path.

The performance difference on Sapphire Rapids 100Gb NIC is improved by the following (as measured by iperf send):

w/o this fix~48 Gbits/s. with this fix ~54 Gbits/s

Cc <[email protected]>

Tested-by Sanjay Kumar <[email protected]>

Signed-off-by Sanjay Kumar <[email protected]>

Signed-off-by Jacob Pan <[email protected]>

From: Baolu Lu <[email protected]>
Sent: Saturday, February 4, 2023 2:32 PM

Quoted 3 lines · Jacob Pan

Intel IOMMU driver implements IOTLB flush queue with domain selective or PASID selective invalidations. In this case there's no need to track IOVA page range and sync IOTLBs, which may cause significant

performance

Quoted 1 line

hit.

[Add cc Robin]

If I understand this patch correctly, this might be caused by below helper:

/**
* iommu_iotlb_gather_add_page - Gather for page-based TLB invalidation
* @domain: IOMMU domain to be invalidated
* @gather: TLB gather data
* @iova: start of page to invalidate
* @size: size of page to invalidate
*
* Helper for IOMMU drivers to build invalidation commands based on
individual
* pages, or with page size/table level hints which cannot be gathered
if they
* differ.
*/
static inline void iommu_iotlb_gather_add_page(struct iommu_domain *domain,
struct
iommu_iotlb_gather *gather,
unsigned long iova,
size_t size)
{
/*
* If the new page is disjoint from the current range or is
mapped at

          * a different granularity, then sync the TLB so that the gather
          * structure can be rewritten.
          */
         if ((gather->pgsize && gather->pgsize != size) ||
             iommu_iotlb_gather_is_disjoint(gather, iova, size))
                 iommu_iotlb_sync(domain, gather);

         gather->pgsize = size;
         iommu_iotlb_gather_add_range(gather, iova, size);
}

As the comments for iommu_iotlb_gather_is_disjoint() says,

"...For many IOMMUs, flushing the IOMMU in this case is better
than merging the two, which might lead to unnecessary invalidations. ..."

So, perhaps the right fix for this performance issue is to add

	if (!gather->queued)

in iommu_iotlb_gather_add_page() or iommu_iotlb_gather_is_disjoint()? It should benefit other arch's as well.

Quoted 55 lines

From: Baolu Lu <[email protected]>
Sent: Saturday, February 4, 2023 2:32 PM

Quoted 3 lines · Jacob Pan

Intel IOMMU driver implements IOTLB flush queue with domain selective or PASID selective invalidations. In this case there's no need to track IOVA page range and sync IOTLBs, which may cause significant

performance

Quoted 1 line

hit.

[Add cc Robin]

If I understand this patch correctly, this might be caused by below helper:

/**
* iommu_iotlb_gather_add_page - Gather for page-based TLB invalidation
* @domain: IOMMU domain to be invalidated
* @gather: TLB gather data
* @iova: start of page to invalidate
* @size: size of page to invalidate
*
* Helper for IOMMU drivers to build invalidation commands based on
individual
* pages, or with page size/table level hints which cannot be gathered
if they
* differ.
*/
static inline void iommu_iotlb_gather_add_page(struct iommu_domain *domain,
struct
iommu_iotlb_gather *gather,
unsigned long iova,
size_t size)
{
/*
* If the new page is disjoint from the current range or is
mapped at

           * a different granularity, then sync the TLB so that the gather
           * structure can be rewritten.
           */
          if ((gather->pgsize && gather->pgsize != size) ||
              iommu_iotlb_gather_is_disjoint(gather, iova, size))
                  iommu_iotlb_sync(domain, gather);

          gather->pgsize = size;
          iommu_iotlb_gather_add_range(gather, iova, size);
}

As the comments for iommu_iotlb_gather_is_disjoint() says,

"...For many IOMMUs, flushing the IOMMU in this case is better
than merging the two, which might lead to unnecessary invalidations. ..."

Quoted 7 lines

So, perhaps the right fix for this performance issue is to add

	if (!gather->queued)

in iommu_iotlb_gather_add_page() or iommu_iotlb_gather_is_disjoint()? It should benefit other arch's as well.

There are only two callers of this helper: intel and arm-smmu-v3.

Looks other drivers just implements direct flush via io_pgtable_tlb_add_page().

and their unmap callback typically does:

if (!iommu_iotlb_gather_queued(gather))

	io_pgtable_tlb_add_page();

from this angle it's same policy as Jacob's does, i.e. if it's already queued then no need to further call optimization for direct flush.

From: Baolu Lu <[email protected]>
Sent: Tuesday, February 7, 2023 2:46 PM

Quoted 10 lines · Kevin

Quoted 10 lines

From: Baolu Lu <[email protected]>
Sent: Saturday, February 4, 2023 2:32 PM

Quoted 3 lines · Jacob Pan

Intel IOMMU driver implements IOTLB flush queue with domain selective or PASID selective invalidations. In this case there's no need to track IOVA page range and sync IOTLBs, which may cause significant

performance

Quoted 1 line

hit.

Quoted 7 lines

Quoted 7 lines

[Add cc Robin]

If I understand this patch correctly, this might be caused by below helper:

/**
* iommu_iotlb_gather_add_page - Gather for page-based TLB

invalidation

Quoted 42 lines

Quoted 42 lines

* @domain: IOMMU domain to be invalidated
* @gather: TLB gather data
* @iova: start of page to invalidate
* @size: size of page to invalidate
*
* Helper for IOMMU drivers to build invalidation commands based on
individual
* pages, or with page size/table level hints which cannot be gathered
if they
* differ.
*/
static inline void iommu_iotlb_gather_add_page(struct iommu_domain *domain,
struct
iommu_iotlb_gather *gather,
unsigned long iova,
size_t size)
{
/*
* If the new page is disjoint from the current range or is
mapped at

           * a different granularity, then sync the TLB so that the gather
           * structure can be rewritten.
           */
          if ((gather->pgsize && gather->pgsize != size) ||
              iommu_iotlb_gather_is_disjoint(gather, iova, size))
                  iommu_iotlb_sync(domain, gather);

          gather->pgsize = size;
          iommu_iotlb_gather_add_range(gather, iova, size);
}

As the comments for iommu_iotlb_gather_is_disjoint() says,

"...For many IOMMUs, flushing the IOMMU in this case is better
than merging the two, which might lead to unnecessary invalidations. ..."

So, perhaps the right fix for this performance issue is to add

	if (!gather->queued)

Quoted 7 lines

Quoted 3 lines

in iommu_iotlb_gather_add_page() or iommu_iotlb_gather_is_disjoint()? It should benefit other arch's as well.

There are only two callers of this helper: intel and arm-smmu-v3.

Looks other drivers just implements direct flush via

io_pgtable_tlb_add_page().

Quoted 3 lines

and their unmap callback typically does:

Quoted 5 lines

if (!iommu_iotlb_gather_queued(gather))

	io_pgtable_tlb_add_page();

from this angle it's same policy as Jacob's does, i.e. if it's already queued then no need to further call optimization for direct flush.

Perhaps we can use iommu_iotlb_gather_queued() to replace direct gather->queued check in this patch as well?

Quoted 4 lines · Jacob Pan

Intel IOMMU driver implements IOTLB flush queue with domain selective or PASID selective invalidations. In this case there's no need to track IOVA page range and sync IOTLBs, which may cause significant performance hit.

[Add cc Robin]

If I understand this patch correctly, this might be caused by below helper:

/**
 * iommu_iotlb_gather_add_page - Gather for page-based TLB invalidation
 * @domain: IOMMU domain to be invalidated
 * @gather: TLB gather data
 * @iova: start of page to invalidate
 * @size: size of page to invalidate
 *
 * Helper for IOMMU drivers to build invalidation commands based on
individual
 * pages, or with page size/table level hints which cannot be gathered
if they
 * differ.
 */
static inline void iommu_iotlb_gather_add_page(struct iommu_domain *domain,
                                               struct
iommu_iotlb_gather *gather,
                                               unsigned long iova,
size_t size)
{
        /*
         * If the new page is disjoint from the current range or is
mapped at
         * a different granularity, then sync the TLB so that the gather
         * structure can be rewritten.
         */
        if ((gather->pgsize && gather->pgsize != size) || iommu_iotlb_gather_is_disjoint(gather, iova, size)) iommu_iotlb_sync(domain, gather);

        gather->pgsize = size; iommu_iotlb_gather_add_range(gather, iova, size);
}

As the comments for iommu_iotlb_gather_is_disjoint() says,

"...For many IOMMUs, flushing the IOMMU in this case is better
 than merging the two, which might lead to unnecessary invalidations. ..."

So, perhaps the right fix for this performance issue is to add

    if (!gather->queued)

in iommu_iotlb_gather_add_page() or iommu_iotlb_gather_is_disjoint()? It should benefit other arch's as well.

Quoted 5 lines

This patch adds a check to avoid IOVA gather page and IOTLB sync for the lazy path.

The performance difference on Sapphire Rapids 100Gb NIC is improved by the following (as measured by iperf send):

Which test case have you done? Post the real data if you have any.

Quoted 1 line

Quoted 3 lines

w/o this fix~48 Gbits/s. with this fix ~54 Gbits/s

Cc <[email protected]>

Again, add a Fixes tag so that people know how far this fix should be back ported.

Quoted 45 lines · Baolu Lu

Quoted 4 lines · Jacob Pan

Intel IOMMU driver implements IOTLB flush queue with domain selective or PASID selective invalidations. In this case there's no need to track IOVA page range and sync IOTLBs, which may cause significant performance hit.

[Add cc Robin]

If I understand this patch correctly, this might be caused by below helper:

/**
  * iommu_iotlb_gather_add_page - Gather for page-based TLB invalidation
  * @domain: IOMMU domain to be invalidated
  * @gather: TLB gather data
  * @iova: start of page to invalidate
  * @size: size of page to invalidate
  *
  * Helper for IOMMU drivers to build invalidation commands based on individual
  * pages, or with page size/table level hints which cannot be gathered if they
  * differ.
  */
static inline void iommu_iotlb_gather_add_page(struct iommu_domain *domain, struct iommu_iotlb_gather *gather, unsigned long iova, size_t size) {
         /*
          * If the new page is disjoint from the current range or is mapped at
          * a different granularity, then sync the TLB so that the gather
          * structure can be rewritten.
          */
         if ((gather->pgsize && gather->pgsize != size) || iommu_iotlb_gather_is_disjoint(gather, iova, size)) iommu_iotlb_sync(domain, gather);

         gather->pgsize = size; iommu_iotlb_gather_add_range(gather, iova, size); }

As the comments for iommu_iotlb_gather_is_disjoint() says,

"...For many IOMMUs, flushing the IOMMU in this case is better than merging the two, which might lead to unnecessary invalidations. ..."

Quoted 6 lines

So, perhaps the right fix for this performance issue is to add

     if (!gather->queued)

in iommu_iotlb_gather_add_page() or iommu_iotlb_gather_is_disjoint()? It should benefit other arch's as well.

The iotlb_gather helpers are really just that - little tools to help drivers with various common iotlb_gather accounting patterns. The decision whether to bother with that accounting at all should really come beforehand, and whether a driver supports flush queues is orthogonal to whether it uses any particular gather helper(s) or not, so I think the patch as-is is correct.

Quoted 10 lines

Quoted 6 lines

This patch adds a check to avoid IOVA gather page and IOTLB sync for the lazy path.

The performance difference on Sapphire Rapids 100Gb NIC is improved by the following (as measured by iperf send):

Which test case have you done? Post the real data if you have any.

Quoted 1 line

Quoted 6 lines

Quoted 3 lines

w/o this fix~48 Gbits/s. with this fix ~54 Gbits/s

Cc <[email protected]>

Again, add a Fixes tag so that people know how far this fix should be back ported.

Note that the overall issue probably dates back to the initial iommu-dma conversion, but if you think it's important enough to go back beyond 5.15 when gather->queued was introduced, that'll need a different fix.

Intel IOMMU driver implements IOTLB flush queue with domain selective or PASID selective invalidations. In this case there's no need to track IOVA page range and sync IOTLBs, which may cause significant performance hit.

This patch adds a check to avoid IOVA gather page and IOTLB sync for the lazy path.

The performance difference on Sapphire Rapids 100Gb NIC is improved by the following (as measured by iperf send):

w/o this fix~48 Gbits/s. with this fix ~54 Gbits/s

Cc <[email protected]>

Tested-by Sanjay Kumar <[email protected]>

Signed-off-by Sanjay Kumar <[email protected]>

Signed-off-by Jacob Pan <[email protected]>

---
  drivers/iommu/intel/iommu.c | 8 +++++++-
  1 file changed, 7 insertions(+), 1 deletion(-)

drivers/iommu/intel/iommu.c +7 -1

index b4878c7ac008..705a1c66691a 100644
@@ -4352,7 +4352,13 @@ static size_t intel_iommu_unmap(struct iommu_domain *domain,
  	if (dmar_domain->max_addr == iova + size)
  		dmar_domain->max_addr = iova;
 
-	iommu_iotlb_gather_add_page(domain, gather, iova, size);
+	/*
+	 * We do not use page-selective IOTLB invalidation in flush queue,
+	 * There is no need to track page and sync iotlb. Domain-selective or
+	 * PASID-selective validation are used in the flush queue.
+	 */
+	if (!gather->queued)
+		iommu_iotlb_gather_add_page(domain, gather, iova, size);
 
  	return size;
  }