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

---
v3: reword comments, add reviewed by tag. (Kevin)
v2: use helper function iommu_iotlb_gather_queued() instead of open coding

---
 drivers/iommu/intel/iommu.c | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)

index 59df7e42fd53..2ee270e4d484 100644
@@ -4346,7 +4346,12 @@ 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,
+	 * so there is no need to track page and sync iotlb.
+	 */
+	if (!iommu_iotlb_gather_queued(gather))
+		iommu_iotlb_gather_add_page(domain, gather, iova, size);
 
 	return size;
 }

2.25.1

Hi,

I've stumbled upon this patch trying to figure out how lazy invalidation is implemented in intel IOMMUs. The patch suggests that lazy invalidation is active whenever iotlb_gather.queued is set.

However, the only place in which gather.queued is written seems to be in dma-iommu.c:

-- drivers/iommu/dma-iommu.c
820: iotlb_gather.queued = READ_ONCE(cookie->fq_domain);
2038: iotlb_gather.queued = free_iova && READ_ONCE(cookie->fq_domain);

Both of these depend on fq_domain but fq_domain is always NULL for intel iommus, since iommu/intel/iommu.c reports IOMMU_CAP_DEFERRED_FLUSH:

-- drivers/iommu/dma-iommu.c:708

	if (domain->type == IOMMU_DOMAIN_DMA_FQ &&
	    (!device_iommu_capable(dev, IOMMU_CAP_DEFERRED_FLUSH) || iommu_dma_init_fq(domain)))
		domain->type = IOMMU_DOMAIN_DMA;

(Above line numbers are from Kernel 6.17).

So I'm not sure, this commit does what it should?
Please let me know what I'm missing here.

Best Regards,
Ferdinand Schober

Hi,

I've stumbled upon this patch trying to figure out how lazy invalidation is implemented in intel IOMMUs. The patch suggests that lazy invalidation is active whenever iotlb_gather.queued is set.

However, the only place in which gather.queued is written seems to be in dma-iommu.c:

-- drivers/iommu/dma-iommu.c
820: iotlb_gather.queued = READ_ONCE(cookie->fq_domain);
2038: iotlb_gather.queued = free_iova && READ_ONCE(cookie->fq_domain);

Both of these depend on fq_domain but fq_domain is always NULL for intel iommus, since iommu/intel/iommu.c reports IOMMU_CAP_DEFERRED_FLUSH:

-- drivers/iommu/dma-iommu.c:708

	if (domain->type == IOMMU_DOMAIN_DMA_FQ &&
	    (!device_iommu_capable(dev, IOMMU_CAP_DEFERRED_FLUSH) || iommu_dma_init_fq(domain)))
		domain->type = IOMMU_DOMAIN_DMA;

(Above line numbers are from Kernel 6.17).

So I'm not sure, this commit does what it should?
Please let me know what I'm missing here.

Best Regards,
Ferdinand Schober

IOMMU_CAP_DEFERRED_FLUSH *is* the "I can usefully support flush queues" capability; if that is reported then iommu_dma_init_fq() is called, and if that succeeds then fq_domain will have been set. If it fails, or if the IOMMU doesn't support flush queues in the first place, then we fall back to the regular strict domain type.

Thanks,
Robin.