From: Jason Gunthorpe <jgg@nvidia.com>
Date: Fri, 27 Feb 2026 15:27:02 +0000 (-0400)
Subject: iommupt: Optimize the gather processing for DMA-FQ mode
X-Git-Url: http://git.ipfire.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=199036ae01321651fe0e4488f9e19a28af4c5f1d;p=thirdparty%2Flinux.git

iommupt: Optimize the gather processing for DMA-FQ mode

In PT_FEAT_FLUSH_RANGE mode the gather was accumulated but never flushed
and then the accumulated range was discarded by the dma-iommu code in
DMA-FQ mode. This is basically optimal.

However for PT_FEAT_FLUSH_RANGE_NO_GAPS the page table would push flushes
that are redundant with the flush all generated by the DMA-FQ mode.

Disable all range accumulation in the gather, and iommu_pt triggered
flushing when in iommu_iotlb_gather_queued() indicates it is in DMA-FQ
mode.

Reported-by: Robin Murphy <robin.murphy@arm.com>
Closes: https://lore.kernel.org/r/794b6121-b66b-4819-b291-9761ed21cd83@arm.com
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Reviewed-by: Lu Baolu <baolu.lu@linux.intel.com>
Reviewed-by: Samiullah Khawaja <skhawaja@google.com>
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
---

diff --git a/drivers/iommu/generic_pt/iommu_pt.h b/drivers/iommu/generic_pt/iommu_pt.h
index 3e33fe64feab2..9c08bb594e417 100644
--- a/drivers/iommu/generic_pt/iommu_pt.h
+++ b/drivers/iommu/generic_pt/iommu_pt.h
@@ -51,16 +51,27 @@ static void gather_range_pages(struct iommu_iotlb_gather *iotlb_gather,
 		iommu_pages_stop_incoherent_list(free_list,
 						 iommu_table->iommu_device);
 
-	if (pt_feature(common, PT_FEAT_FLUSH_RANGE_NO_GAPS) &&
-	    iommu_iotlb_gather_is_disjoint(iotlb_gather, iova, len)) {
-		iommu_iotlb_sync(&iommu_table->domain, iotlb_gather);
-		/*
-		 * Note that the sync frees the gather's free list, so we must
-		 * not have any pages on that list that are covered by iova/len
-		 */
+	/*
+	 * If running in DMA-FQ mode then the unmap will be followed by an IOTLB
+	 * flush all so we need to optimize by never flushing the IOTLB here.
+	 *
+	 * For NO_GAPS the user gets to pick if flushing all or doing micro
+	 * flushes is better for their work load by choosing DMA vs DMA-FQ
+	 * operation. Drivers should also see shadow_on_flush.
+	 */
+	if (!iommu_iotlb_gather_queued(iotlb_gather)) {
+		if (pt_feature(common, PT_FEAT_FLUSH_RANGE_NO_GAPS) &&
+		    iommu_iotlb_gather_is_disjoint(iotlb_gather, iova, len)) {
+			iommu_iotlb_sync(&iommu_table->domain, iotlb_gather);
+			/*
+			 * Note that the sync frees the gather's free list, so
+			 * we must not have any pages on that list that are
+			 * covered by iova/len
+			 */
+		}
+		iommu_iotlb_gather_add_range(iotlb_gather, iova, len);
 	}
 
-	iommu_iotlb_gather_add_range(iotlb_gather, iova, len);
 	iommu_pages_list_splice(free_list, &iotlb_gather->freelist);
 }