| Message ID | 20261105105816.3012283-1-peter.maydell@linaro.org |
|---|---|
| State | New |
| Headers | show |
| Series | hw/i386/amd_iommu: Mask address when calculating large PTE page size | expand |
nit: For consistency, please use "amd_iommu" as the prefix instead of the full "hw/i386/amd_iommu" path. On 11/5/26 5:58 AM, Peter Maydell wrote: > The amd_iommu handling of the "large PTE page size" (where Next Level bits > are 0b111) case is not correct. We calculate the page size using the > address in the descriptor (as per table 14 of the spec), but we don't > explicitly pick out the Page > > The amd_iommu handling of the "large PTE page size" (where Next Level > bits are 0b111) case is not correct. We calculate the page size > using the address in the descriptor (as per table 14 of the spec), > but we don't explicitly pick out the Page Address field, instead > passing the whole PTE word to cto64(). This means that a PTE which > incorrectly sets all the reserved high bits can cause cto64() to > return 63 or 64, and we will try to shift by an out of range amount, > which is undefined behaviour. > > Mask the PTE with AMDVI_DEV_PT_ROOT_MASK to get the page address > (matching how we do it to get the address for the next page table > level in NEXT_PTE_ADDR). > Fixes: a74bb3110a5b ("amd_iommu: Add helpers to walk AMD v1 Page Table format") > Cc: qemu-stable@nongnu.org > Resolves: https://gitlab.com/qemu-project/qemu/-/work_items/3618 > Reported-by: huntr bubble <https://gitlab.com/bubblehuntr> > Signed-off-by: Peter Maydell <peter.maydell@linaro.org> Reviewed-by: Alejandro Jimenez <alejandro.j.jimenez@oracle.com> > --- > Side note: we don't seem to enforce the spec's requirement that > the page size must be larger than the default for the level > and smaller than the default for the next level up, but that's > a different bug. ACK, will work on a follow up to address those types of issues. Thank you, Alejandro > --- > hw/i386/amd_iommu.h | 3 ++- > 1 file changed, 2 insertions(+), 1 deletion(-) > > diff --git a/hw/i386/amd_iommu.h b/hw/i386/amd_iommu.h > index 948d8611b15..008e882724c 100644 > --- a/hw/i386/amd_iommu.h > +++ b/hw/i386/amd_iommu.h > @@ -216,7 +216,8 @@ > /* > * Take a PTE value with mode=0x07 and return the page size it encodes. > */ > -#define PTE_LARGE_PAGE_SIZE(pte) (1ULL << (1 + cto64(((pte) | 0xfffULL)))) > +#define PTE_LARGE_PAGE_SIZE(pte) \ > + (1ULL << (1 + cto64((((pte) & AMDVI_DEV_PT_ROOT_MASK) | 0xfffULL)))) > > /* Return number of PTEs to use for a given page size (expected power of 2) */ > #define PAGE_SIZE_PTE_COUNT(pgsz) (1ULL << ((ctz64(pgsz) - 12) % 9)) > -- > 2.43.0 >
On 11/5/2026 4:28 PM, Peter Maydell wrote: > The amd_iommu handling of the "large PTE page size" (where Next Level > bits are 0b111) case is not correct. We calculate the page size > using the address in the descriptor (as per table 14 of the spec), > but we don't explicitly pick out the Page Address field, instead > passing the whole PTE word to cto64(). This means that a PTE which > incorrectly sets all the reserved high bits can cause cto64() to > return 63 or 64, and we will try to shift by an out of range amount, > which is undefined behaviour. > > Mask the PTE with AMDVI_DEV_PT_ROOT_MASK to get the page address > (matching how we do it to get the address for the next page table > level in NEXT_PTE_ADDR). > > Cc: qemu-stable@nongnu.org > Resolves: https://gitlab.com/qemu-project/qemu/-/work_items/3618 > Reported-by: huntr bubble <https://gitlab.com/bubblehuntr> > Signed-off-by: Peter Maydell <peter.maydell@linaro.org> > --- > Side note: we don't seem to enforce the spec's requirement that > the page size must be larger than the default for the level > and smaller than the default for the next level up, but that's > a different bug. > --- > hw/i386/amd_iommu.h | 3 ++- > 1 file changed, 2 insertions(+), 1 deletion(-) > > diff --git a/hw/i386/amd_iommu.h b/hw/i386/amd_iommu.h > index 948d8611b15..008e882724c 100644 > --- a/hw/i386/amd_iommu.h > +++ b/hw/i386/amd_iommu.h > @@ -216,7 +216,8 @@ > /* > * Take a PTE value with mode=0x07 and return the page size it encodes. > */ > -#define PTE_LARGE_PAGE_SIZE(pte) (1ULL << (1 + cto64(((pte) | 0xfffULL)))) > +#define PTE_LARGE_PAGE_SIZE(pte) \ > + (1ULL << (1 + cto64((((pte) & AMDVI_DEV_PT_ROOT_MASK) | 0xfffULL)))) > > /* Return number of PTEs to use for a given page size (expected power of 2) */ > #define PAGE_SIZE_PTE_COUNT(pgsz) (1ULL << ((ctz64(pgsz) - 12) % 9)) Reviewed-by: Sairaj kodilkar <sarunkod@amd.com> Thanks Sairaj
diff --git a/hw/i386/amd_iommu.h b/hw/i386/amd_iommu.h index 948d8611b15..008e882724c 100644 --- a/hw/i386/amd_iommu.h +++ b/hw/i386/amd_iommu.h @@ -216,7 +216,8 @@ /* * Take a PTE value with mode=0x07 and return the page size it encodes. */ -#define PTE_LARGE_PAGE_SIZE(pte) (1ULL << (1 + cto64(((pte) | 0xfffULL)))) +#define PTE_LARGE_PAGE_SIZE(pte) \ + (1ULL << (1 + cto64((((pte) & AMDVI_DEV_PT_ROOT_MASK) | 0xfffULL)))) /* Return number of PTEs to use for a given page size (expected power of 2) */ #define PAGE_SIZE_PTE_COUNT(pgsz) (1ULL << ((ctz64(pgsz) - 12) % 9))
The amd_iommu handling of the "large PTE page size" (where Next Level bits are 0b111) case is not correct. We calculate the page size using the address in the descriptor (as per table 14 of the spec), but we don't explicitly pick out the Page Address field, instead passing the whole PTE word to cto64(). This means that a PTE which incorrectly sets all the reserved high bits can cause cto64() to return 63 or 64, and we will try to shift by an out of range amount, which is undefined behaviour. Mask the PTE with AMDVI_DEV_PT_ROOT_MASK to get the page address (matching how we do it to get the address for the next page table level in NEXT_PTE_ADDR). Cc: qemu-stable@nongnu.org Resolves: https://gitlab.com/qemu-project/qemu/-/work_items/3618 Reported-by: huntr bubble <https://gitlab.com/bubblehuntr> Signed-off-by: Peter Maydell <peter.maydell@linaro.org> --- Side note: we don't seem to enforce the spec's requirement that the page size must be larger than the default for the level and smaller than the default for the next level up, but that's a different bug. --- hw/i386/amd_iommu.h | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)