Skip to contentSkip to conversation

: iommu fix potential overflow in alloc_iommu()

PATCH In review Prarit Bhargava 2 participants 7 messages Jul 21, 2008

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 : iommu fix potential overflow in alloc_iommu()

(This didn't appear on LKML or any of the mirrors ... trying again)

It is possible that alloc_iommu()'s boundary_size overflows as dma_get_seg_boundary can return 0xffffffff. In that case, further usage of boundary_size triggers a BUG_ON() in the iommu code.

arch/x86/kernel/pci-gart_64.c +1 -1
index faf3229..baecb47 100644
@@ -91,7 +91,7 @@ static unsigned long alloc_iommu(struct device *dev, int size)
 
 	base_index = ALIGN(iommu_bus_base & dma_get_seg_boundary(dev),
 			   PAGE_SIZE) >> PAGE_SHIFT;
-	boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
+	boundary_size = ALIGN((unsigned long long)dma_get_seg_boundary(dev) + 1,
 			      PAGE_SIZE) >> PAGE_SHIFT;
 
 	spin_lock_irqsave(&iommu_bitmap_lock, flags);
#1 lore raw

PATCH 1/1 : PCI: GART iommu alignment fixes

pci_alloc_consistent/dma_alloc_coherent is supposed to return size aligned addresses.

Quoted 1 line

From Documentation/DMA-mapping.txt:

"pci_alloc_consistent returns two values: the virtual address which you can use to access it from the CPU and dma_handle which you pass to the card.

The cpu return address and the DMA bus master address are both guaranteed to be aligned to the smallest PAGE_SIZE order which is greater than or equal to the requested size. This invariant exists (for example) to guarantee that if you allocate a chunk which is smaller than or equal to 64 kilobytes, the extent of the buffer you receive will not cross a 64K boundary."

Fix the GART's alloc_iommu code to return a size aligned address. Also fix an incorrect alignment calculation in the iommu-helper code.

arch/x86/kernel/pci-gart_64.c +3 -2
index faf3229..329718e 100644
@@ -96,11 +96,12 @@ static unsigned long alloc_iommu(struct device *dev, int size)
 
 	spin_lock_irqsave(&iommu_bitmap_lock, flags);
 	offset = iommu_area_alloc(iommu_gart_bitmap, iommu_pages, next_bit,
-				  size, base_index, boundary_size, 0);
+				  size, base_index, boundary_size, size - 1);
 	if (offset == -1) {
 		need_flush = 1;
 		offset = iommu_area_alloc(iommu_gart_bitmap, iommu_pages, 0,
-					  size, base_index, boundary_size, 0);
+					  size, base_index, boundary_size,
+					  size - 1);
 	}
 	if (offset != -1) {
 		set_bit_string(iommu_gart_bitmap, offset, size);
lib/iommu-helper.c +1 -1
index a3b8d4c..b970f1b 100644
@@ -16,7 +16,7 @@ again:
 	index = find_next_zero_bit(map, size, start);
 
 	/* Align allocation */
-	index = (index + align_mask) & ~align_mask;
+	index = (index + align_mask + 1) & ~align_mask;
 
 	end = index + nr;
 	if (end >= size)
#2 lore raw

Re: [PATCH]: PCI: GART iommu alignment fixes

Quoted 28 lines · On Mon, 21 Jul 2008 10:15:27 -0400 Prarit Bhargava <[email protected]> wrote:

pci_alloc_consistent/dma_alloc_coherent is supposed to return size aligned addresses.

From Documentation/DMA-mapping.txt:

"pci_alloc_consistent returns two values: the virtual address which you can use to access it from the CPU and dma_handle which you pass to the card.

The cpu return address and the DMA bus master address are both guaranteed to be aligned to the smallest PAGE_SIZE order which is greater than or equal to the requested size. This invariant exists (for example) to guarantee that if you allocate a chunk which is smaller than or equal to 64 kilobytes, the extent of the buffer you receive will not cross a 64K boundary."

Fix the GART's alloc_iommu code to return a size aligned address. Also fix an incorrect alignment calculation in the iommu-helper code.

arch/x86/kernel/pci-gart_64.c +0 -0
index faf3229..329718e 100644
@@ -96,11 +96,12 @@ static unsigned long alloc_iommu(struct device *dev, int size)
 
 	spin_lock_irqsave(&iommu_bitmap_lock, flags);

arch/x86/kernel/pci-gart_64.c static unsigned long alloc_iommu(struct device *dev, int size) · line 105

Quoted 10 lines
arch/x86/kernel/pci-gart_64.c +3 -2
 	offset = iommu_area_alloc(iommu_gart_bitmap, iommu_pages, next_bit,
-				  size, base_index, boundary_size, 0);
+				  size, base_index, boundary_size, size - 1);
 	if (offset == -1) {
 		need_flush = 1;
 		offset = iommu_area_alloc(iommu_gart_bitmap, iommu_pages, 0,
-					  size, base_index, boundary_size, 0);
+					  size, base_index, boundary_size,
+					  size - 1);
 	}

This affects map_sg and map_single unnecessarily.

lib/iommu-helper.c again: · line 19

Quoted 12 lines
arch/x86/kernel/pci-gart_64.c +0 -0
 	if (offset != -1) {
 		set_bit_string(iommu_gart_bitmap, offset, size);
lib/iommu-helper.c +1 -1
index a3b8d4c..b970f1b 100644
@@ -16,7 +16,7 @@ again:
 	index = find_next_zero_bit(map, size, start);
 
 	/* Align allocation */
-	index = (index + align_mask) & ~align_mask;
+	index = (index + align_mask + 1) & ~align_mask;

Doesn't look right. What we do here is __ALIGN_MASK().

#3 lore raw
In reply to Prarit Bhargava
Quoted 5 lines · On Mon, 21 Jul 2008 10:15:22 -0400 Prarit Bhargava <[email protected]> wrote:

(This didn't appear on LKML or any of the mirrors ... trying again)

It is possible that alloc_iommu()'s boundary_size overflows as dma_get_seg_boundary can return 0xffffffff. In that case, further usage of boundary_size triggers a BUG_ON() in the iommu code.

Did you actually hit this? pci-gart_64.c is used only by X86_64.

#4 lore raw
Quoted 4 lines · FUJITA Tomonori

On Mon, 21 Jul 2008 10:15:22 -0400
Prarit Bhargava <[email protected]> wrote:

Quoted 9 lines
Quoted 6 lines

(This didn't appear on LKML or any of the mirrors ... trying again)

It is possible that alloc_iommu()'s boundary_size overflows as dma_get_seg_boundary can return 0xffffffff. In that case, further usage of boundary_size triggers a BUG_ON() in the iommu code.

Did you actually hit this? pci-gart_64.c is used only by X86_64.

I hit this by declaring a device struct and not declaring a value for dev->dma_parms->segment_boundary_mask.

I was attempting to alloc out of the IOMMU and the code then dies on the bugcheck in iommu_is_span_boundary() because boundary_size = 0.

P.

#5 lore raw
Quoted 14 lines · On Mon, 21 Jul 2008 11:21:39 -0400 Prarit Bhargava <[email protected]> wrote:
Quoted 11 lines · FUJITA Tomonori
Quoted 6 lines · On Mon, 21 Jul 2008 10:15:22 -0400 Prarit Bhargava <[email protected]> wrote:

(This didn't appear on LKML or any of the mirrors ... trying again)

It is possible that alloc_iommu()'s boundary_size overflows as dma_get_seg_boundary can return 0xffffffff. In that case, further usage of boundary_size triggers a BUG_ON() in the iommu code.

Quoted 5 lines
Quoted 2 lines

Did you actually hit this? pci-gart_64.c is used only by X86_64.

I hit this by declaring a device struct and not declaring a value for dev->dma_parms->segment_boundary_mask.

What do you mean? You set dev->dma_params but does't set dev->dma_parms->segment_boundary_mask? If so, you need to fix your code. If you set dev->dma_parms, dma_parms needs to be initialized properly.

#6 lore raw
Quoted 21 lines · FUJITA Tomonori
Quoted 17 lines · On Mon, 21 Jul 2008 11:21:39 -0400 Prarit Bhargava <[email protected]> wrote:
Quoted 15 lines · FUJITA Tomonori
Quoted 7 lines · On Mon, 21 Jul 2008 10:15:22 -0400 Prarit Bhargava <[email protected]> wrote:

(This didn't appear on LKML or any of the mirrors ... trying again)

It is possible that alloc_iommu()'s boundary_size overflows as dma_get_seg_boundary can return 0xffffffff. In that case, further usage of boundary_size triggers a BUG_ON() in the iommu code.

Did you actually hit this? pci-gart_64.c is used only by X86_64.

Quoted 8 lines
Quoted 2 lines

I hit this by declaring a device struct and not declaring a value for dev->dma_parms->segment_boundary_mask.

What do you mean? You set dev->dma_params but does't set dev->dma_parms->segment_boundary_mask? If so, you need to fix your code. If you set dev->dma_parms, dma_parms needs to be initialized properly.

No, I didn't set dev->dma_params.

P.