Bug Fix drivers/pci/intel-iommu.c: convert paddr to correct DMA pfn

When calling domain_pfn_mapping(), paddr should be converted to DMA pfn correctly instead just shift by VTD_PAGE_SHIFT.

This issue causes kernel panic on PAGE_SIZE>VTD_PAGE_SIZE platforms e.g. ia64 platforms.

---

 drivers/pci/intel-iommu.c |    3 ++-
 1 files changed, 2 insertions(+), 1 deletion(-)

index bec29ed..54ee63d 100644
@@ -2558,6 +2564,7 @@ static dma_addr_t __intel_map_single(struct device *hwdev, phys_addr_t paddr,
 	int prot = 0;
 	int ret;
 	struct intel_iommu *iommu;
+	unsigned long paddr_pfn = paddr >> PAGE_SHIFT;
 
 	BUG_ON(dir = DMA_NONE);
 

@@ -2592,7 +2599,7 @@ static dma_addr_t __intel_map_single(struct device *hwdev, phys_addr_t paddr,
 	 * is not a big problem
 	 */
 	ret = domain_pfn_mapping(domain, mm_to_dma_pfn(iova->pfn_lo),
-				 paddr >> VTD_PAGE_SHIFT, size, prot);
+				 mm_to_dma_pfn(paddr_pfn), size, prot);
 	if (ret)
 		goto error;
 

The iova->pfn_lo should be converted to VTD page address before it's passed to iommu_flush_dev_iotlb().

This issue may cause DMA failure on PAGE_SIZE>VTD_PAGE_SIZE platforms e.g. ia64 platforms.

---

 drivers/pci/intel-iommu.c |    2 +-
 1 files changed, 1 insertion(+), 1 deletion(-)

index bec29ed..54ee63d 100644
@@ -2647,7 +2654,7 @@ static void flush_unmaps(void)
 			mask = (iova->pfn_hi - iova->pfn_lo + 1) << PAGE_SHIFT;
 			mask = ilog2(mask >> VTD_PAGE_SHIFT);
 			iommu_flush_dev_iotlb(deferred_flush[i].domain[j],
-					iova->pfn_lo << PAGE_SHIFT, mask);
+					iova->pfn_lo << VTD_PAGE_SHIFT, mask);
 			__free_iova(&deferred_flush[i].domain[j]->iovad, iova);
 		}
 		deferred_flush[i].next = 0;

Using the word 'alignment' somewhere in the above commit would have confused me a lot less -- I got most of the way through composing a mail asking how this patch wasn't a no-op, before I spotted it.

An alternative would be to align the size to DMA pages? Would that make the sg case easier?

(Thanks for fixing this, btw)

dwmw2

I think this one is wrong. We allocate IOVA space in MM pages, not DMA pages -- so the pfn_lo is an MM page already, and shifting by PAGE_SHIFT is the right thing to do.

(I could be tempted to allocate it in DMA pages instead, but that's a different story).

David Woodhouse                            Open Source Technology Centre
[email protected]                              Intel Corporation

When compiling drm_cache.c on ia64, gcc reports undefined drm_clflush_ipi_handler() error. The problem is because drm_cache_ipi_handler is a typo.

One straightforward fix is changing drm_cache_ipi_handler to drm_clflush_ipi_handler.

But since drm_clflush_ipi_handler() does nothing on non-x86 platforms and ipi handling is costly, it would be better to just do nothing on non-x86 platforms and remove drm_clflush_ipi_handler() for those platforms.

---
The issue was introduced by commit c9c97b8c75019814d8c007059bc827bb475be917 in linux-next.

 drm_cache.c |    7 -------
 1 files changed, 7 deletions(-)

index 3a5575e..8054a5c 100644
@@ -62,10 +62,6 @@ drm_clflush_ipi_handler(void *null)
 {
 	wbinvd();
 }
-#elif !defined(__powerpc__)
-static void drm_cache_ipi_handler(void *dummy)
-{
-}
 #endif
 void
 drm_clflush_pages(struct page *pages[], unsigned long num_pages)

@@ -94,9 +90,6 @@ drm_clflush_pages(struct page *pages[], unsigned long num_pages)
 				   (unsigned long)page_virtual + PAGE_SIZE);
 		kunmap_atomic(page_virtual, KM_USER0);
 	}
-#else
-	if (on_each_cpu(drm_clflush_ipi_handler, NULL, 1) != 0)
-		printk(KERN_ERR "Timed out waiting for drm cache flush\n");
 #endif
 }
 EXPORT_SYMBOL(drm_clflush_pages);