x86/iommu: fix dma pte address size error
PATCH 1/1 x86/iommu: fix dma pte address size error
In Intel Vt-D specs, Chapter 9.3 Page-Table Entry,
The size of ADDR(address) field is 12:51, but the function dma_pte_addr
treats it as 12:63.
--- drivers/iommu/intel-iommu.c | 4 ++-- include/linux/dma_remapping.h | 2 ++ 2 files changed, 4 insertions(+), 2 deletions(-)
drivers/iommu/intel-iommu.c +2 -2
@@ -311,10 +311,10 @@ static inline void dma_set_pte_prot(struct dma_pte *pte, unsigned long prot)static inline u64 dma_pte_addr (struct dma_pte *pte) {#ifdef CONFIG_64BIT -return pte->val &VTD_PAGE_MASK ; +return pte->val &DMA_PTE_MASK ;#else /* Must have a full atomic 64-bit read */ -return __cmpxchg64 (&pte->val,0ULL ,0ULL ) &VTD_PAGE_MASK ; +return __cmpxchg64 (&pte->val,0ULL ,0ULL ) &DMA_PTE_MASK ;#endif }
include/linux/dma_remapping.h +2 -0
@@ -16,6 +16,8 @@#define DMA_PTE_WRITE (2)#define DMA_PTE_LARGE_PAGE (1 << 7)#define DMA_PTE_SNP (1 << 11) +#define DMA_PTE_ADD_LENGTH (40) +#define DMA_PTE_MASK ((((u64)1 << DMA_PTE_ADD_LENGTH) - 1) << VTD_PAGE_SHIFT)#define CONTEXT_TT_MULTI_LEVEL 0#define CONTEXT_TT_DEV_IOTLB 1
Signature · 1 line
1.7.10.4
Hi guys,
I think this patch is necessary. Please take a look at it.
Thanks
Zhen-Hua
Quoted 40 lines · Zhen-Hua
In Intel Vt-D specs, Chapter 9.3 Page-Table Entry,
The size of ADDR(address) field is 12:51, but the functiondma_pte_addr
treats it as 12:63.Signed-off-by Li, Zhen-Hua <[email protected]>--- drivers/iommu/intel-iommu.c | 4 ++-- include/linux/dma_remapping.h | 2 ++ 2 files changed, 4 insertions(+), 2 deletions(-)drivers/iommu/intel-iommu.c +2
-2@@ -311,10 +311,10 @@ static inline void dma_set_pte_prot(struct dma_pte *pte, unsigned long prot)static inline u64 dma_pte_addr (struct dma_pte *pte) {#ifdef CONFIG_64BIT -return pte->val &VTD_PAGE_MASK ; +return pte->val &DMA_PTE_MASK ;#else /* Must have a full atomic 64-bit read */ -return __cmpxchg64 (&pte->val,0ULL ,0ULL ) &VTD_PAGE_MASK ; +return __cmpxchg64 (&pte->val,0ULL ,0ULL ) &DMA_PTE_MASK ;#endif }include/linux/dma_remapping.h +2
-0@@ -16,6 +16,8 @@#define DMA_PTE_WRITE (2)#define DMA_PTE_LARGE_PAGE (1 << 7)#define DMA_PTE_SNP (1 << 11) +#define DMA_PTE_ADD_LENGTH (40) +#define DMA_PTE_MASK ((((u64)1 << DMA_PTE_ADD_LENGTH) - 1) << VTD_PAGE_SHIFT)#define CONTEXT_TT_MULTI_LEVEL 0#define CONTEXT_TT_DEV_IOTLB 1
Quoted 10 lines · Zhen-Hua
In Intel Vt-D specs, Chapter 9.3 Page-Table Entry,
The size of ADDR(address) field is 12:51, but the functiondma_pte_addr
treats it as 12:63.Signed-off-by Li, Zhen-Hua<[email protected]>--- drivers/iommu/intel-iommu.c | 4 ++-- include/linux/dma_remapping.h | 2 ++ 2 files changed, 4 insertions(+), 2 deletions(-)
Is this patching for the sake of spec interpretation?
a dma-pte format (consumed by iommu) has 63,61:52 as available for sw, ignored by hw. 62 is 'transient mapping' bit, which is a _hint_ for selecting iotlbs to flush sooner. finally, the system would have to have a memory map that actually has bit 62 set to be affected.
So, for intel-iommu, I don't see a bug occurring. Did you actually have one with previous definition, and if so, could you provide that information ?
Cheers,
- Don
Quoted 30 lines
drivers/iommu/intel-iommu.c +2
-2@@ -311,10 +311,10 @@ static inline void dma_set_pte_prot(struct dma_pte *pte, unsigned long prot)static inline u64 dma_pte_addr (struct dma_pte *pte) {#ifdef CONFIG_64BIT -return pte->val&VTD_PAGE_MASK ; +return pte->val&DMA_PTE_MASK ;#else /* Must have a full atomic 64-bit read */ -return __cmpxchg64 (&pte->val,0ULL ,0ULL )&VTD_PAGE_MASK ; +return __cmpxchg64 (&pte->val,0ULL ,0ULL )&DMA_PTE_MASK ;#endif }include/linux/dma_remapping.h +2
-0@@ -16,6 +16,8 @@#define DMA_PTE_WRITE (2)#define DMA_PTE_LARGE_PAGE (1<< 7)#define DMA_PTE_SNP (1<< 11) +#define DMA_PTE_ADD_LENGTH (40) +#define DMA_PTE_MASK ((((u64)1<< DMA_PTE_ADD_LENGTH) - 1)<< VTD_PAGE_SHIFT)#define CONTEXT_TT_MULTI_LEVEL 0#define CONTEXT_TT_DEV_IOTLB 1
Hi Don,
This patch is not only for the sake of spec interpretation. Till now I did not see any bugs , but it does not meant no bugs will appear in the future. The address returned by dma_pte_addr is used in many places.
Thanks
Zhenhua
Quoted 60 lines · Don Dutile
Quoted 10 lines · Zhen-Hua
In Intel Vt-D specs, Chapter 9.3 Page-Table Entry,
The size of ADDR(address) field is 12:51, but the functiondma_pte_addr
treats it as 12:63.Signed-off-by Li, Zhen-Hua<[email protected]>--- drivers/iommu/intel-iommu.c | 4 ++-- include/linux/dma_remapping.h | 2 ++ 2 files changed, 4 insertions(+), 2 deletions(-)Is this patching for the sake of spec interpretation?
a dma-pte format (consumed by iommu) has 63,61:52 as available for sw, ignored by hw. 62 is 'transient mapping' bit, which is a _hint_ for selecting iotlbs to flush sooner. finally, the system would have to have a memory map that actually has bit 62 set to be affected.So, for intel-iommu, I don't see a bug occurring. Did you actually have one with previous definition, and if so, could you provide that information ?
Cheers,
- DonQuoted 31 lines
drivers/iommu/intel-iommu.c +2
-2@@ -311,10 +311,10 @@ static inline void dma_set_pte_prot(struct dma_pte *pte, unsigned long prot)static inline u64 dma_pte_addr (struct dma_pte *pte) {#ifdef CONFIG_64BIT -return pte->val&VTD_PAGE_MASK ; +return pte->val&DMA_PTE_MASK ;#else /* Must have a full atomic 64-bit read */ -return __cmpxchg64 (&pte->val,0ULL ,0ULL )&VTD_PAGE_MASK ; +return __cmpxchg64 (&pte->val,0ULL ,0ULL )&DMA_PTE_MASK ;#endif }include/linux/dma_remapping.h +2
-0@@ -16,6 +16,8 @@#define DMA_PTE_WRITE (2)#define DMA_PTE_LARGE_PAGE (1<< 7)#define DMA_PTE_SNP (1<< 11) +#define DMA_PTE_ADD_LENGTH (40) +#define DMA_PTE_MASK ((((u64)1<< DMA_PTE_ADD_LENGTH) - 1)<< VTD_PAGE_SHIFT)#define CONTEXT_TT_MULTI_LEVEL 0Quoted 1 line
#define CONTEXT_TT_DEV_IOTLB 1.