iommu/vt-d: Clear PRQ overflow only when PRQ is empty
authorLu Baolu <baolu.lu@linux.intel.com>
Tue, 26 Jan 2021 08:07:29 +0000 (16:07 +0800)
committerJoerg Roedel <jroedel@suse.de>
Fri, 29 Jan 2021 08:25:24 +0000 (09:25 +0100)
It is incorrect to always clear PRO when it's set w/o first checking
whether the overflow condition has been cleared. Current code assumes
that if an overflow condition occurs it must have been cleared by earlier
loop. However since the code runs in a threaded context, the overflow
condition could occur even after setting the head to the tail under some
extreme condition. To be sane, we should read both head/tail again when
seeing a pending PRO and only clear PRO after all pending PRs have been
handled.

Suggested-by: Kevin Tian <kevin.tian@intel.com>
Signed-off-by: Lu Baolu <baolu.lu@linux.intel.com>
Link: https://lore.kernel.org/linux-iommu/MWHPR11MB18862D2EA5BD432BF22D99A48CA09@MWHPR11MB1886.namprd11.prod.outlook.com/
Link: https://lore.kernel.org/r/20210126080730.2232859-2-baolu.lu@linux.intel.com
Signed-off-by: Joerg Roedel <jroedel@suse.de>
drivers/iommu/intel/svm.c

index 033b258..d7c98c5 100644 (file)
@@ -1042,8 +1042,17 @@ prq_advance:
         * Clear the page request overflow bit and wake up all threads that
         * are waiting for the completion of this handling.
         */
-       if (readl(iommu->reg + DMAR_PRS_REG) & DMA_PRS_PRO)
-               writel(DMA_PRS_PRO, iommu->reg + DMAR_PRS_REG);
+       if (readl(iommu->reg + DMAR_PRS_REG) & DMA_PRS_PRO) {
+               pr_info_ratelimited("IOMMU: %s: PRQ overflow detected\n",
+                                   iommu->name);
+               head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
+               tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
+               if (head == tail) {
+                       writel(DMA_PRS_PRO, iommu->reg + DMAR_PRS_REG);
+                       pr_info_ratelimited("IOMMU: %s: PRQ overflow cleared",
+                                           iommu->name);
+               }
+       }
 
        if (!completion_done(&iommu->prq_complete))
                complete(&iommu->prq_complete);