Re: [PATCH] iommu/vt-d: Fix a bug in intel_iommu_iova_to_phys() for huge page

From: Yonghyun Hwang
Date: Wed Feb 19 2020 - 19:12:56 EST


Hello Baolu,

Yes, pfn_to_dma_pte() doesn't put 0 for level. However, it's hard for
me to get that without comment, which is the reason why I put a dummy
"if" there. Anyway, what about the following code?

static inline unsigned long offset_mask(int level)
{
return BIT_MASK(level_to_offset_bits(level) + VTD_PAGE_SHIFT) - 1;
}

static phys_addr_t intel_iommu_iova_to_phys(struct iommu_domain *domain,
dma_addr_t iova)
{
struct dmar_domain *dmar_domain = to_dmar_domain(domain);
struct dma_pte *pte;
int level = 0;
u64 phys = 0;

if (dmar_domain->flags & DOMAIN_FLAG_LOSE_CHILDREN)
return 0;

pte = pfn_to_dma_pte(dmar_domain, iova >> VTD_PAGE_SHIFT, &level);
if (pte)
phys = dma_pte_addr(pte) + (iova & offset_mask(level));

return phys;
}

On Wed, Feb 19, 2020 at 3:33 PM Lu Baolu <baolu.lu@xxxxxxxxxxxxxxx> wrote:
>
> Hi,
>
> On 2020/2/20 2:51, Yonghyun Hwang wrote:
> > Hello Lu and Moritz,
> >
> > I think it's better to keep "if (level > 1)" because
> > level_to_offset_bits() returns (unsigned int) -1 * LEVEL_STRIDE if level
> > becomes 0.
>
> @level will never be 0 in this case.
>
> Best regards,
> baolu
>
> >
> > static inline unsigned int level_to_offset_bits(int level)
> > {
> > return (level - 1) * LEVEL_STRIDE;
> > }
> >
> > On Tue, Feb 18, 2020 at 8:42 PM Moritz Fischer <mdf@xxxxxxxxxx
> > <mailto:mdf@xxxxxxxxxx>> wrote:
> >
> > Hi Baolu, Yonghyun
> >
> > On Wed, Feb 19, 2020 at 11:15:36AM +0800, Lu Baolu wrote:
> > > Hi Yonghyun,
> > >
> > > Thanks for the patch.
> > >
> > > On 2020/2/19 6:23, Yonghyun Hwang wrote:
> > > > intel_iommu_iova_to_phys() has a bug when it translates an IOVA
> > for a huge
> > > > page onto its corresponding physical address. This commit fixes
> > the bug by
> > > > accomodating the level of page entry for the IOVA and adds
> > IOVA's lower
> > > > address to the physical address. >
> > > > Signed-off-by: Yonghyun Hwang <yonghyun@xxxxxxxxxx
> > <mailto:yonghyun@xxxxxxxxxx>>
> > > > ---
> > > > drivers/iommu/intel-iommu.c | 11 +++++++++--
> > > > 1 file changed, 9 insertions(+), 2 deletions(-)
> > > >
> > > > diff --git a/drivers/iommu/intel-iommu.c
> > b/drivers/iommu/intel-iommu.c
> > > > index 0c8d81f56a30..ed6e69adb578 100644
> > > > --- a/drivers/iommu/intel-iommu.c
> > > > +++ b/drivers/iommu/intel-iommu.c
> > > > @@ -5555,13 +5555,20 @@ static phys_addr_t
> > intel_iommu_iova_to_phys(struct iommu_domain *domain,
> > > > struct dma_pte *pte;
> > > > int level = 0;
> > > > u64 phys = 0;
> > > > + const unsigned long pfn = iova >> VTD_PAGE_SHIFT;
> > >
> > > Why do you need a "const unsigned long" here?
> >
> >
> >
> > It's because pfn_to_dma_pte() takes pfn for its second parameter as
> > "unsigned int" type and pfn is not changed in intel_iommu_iova_to_phys().
> >
> > >
> > > > if (dmar_domain->flags & DOMAIN_FLAG_LOSE_CHILDREN)
> > > > return 0;
> > > > - pte = pfn_to_dma_pte(dmar_domain, iova >> VTD_PAGE_SHIFT,
> > &level);
> > > > - if (pte)
> > > > + pte = pfn_to_dma_pte(dmar_domain, pfn, &level);
> > > > + if (pte) {
> > > > phys = dma_pte_addr(pte);
> > > > + if (level > 1)
> > > > + phys += (pfn &
> > > > + ((1UL <<
> > level_to_offset_bits(level)) - 1))
> > > > + << VTD_PAGE_SHIFT;
> > > > + phys += iova & (VTD_PAGE_SIZE - 1);
> > >
> > > How about
> >
> > >
> > > diff --git a/drivers/iommu/intel-iommu.c
> > b/drivers/iommu/intel-iommu.c
> > > index 9dc37672bf89..bd17c2510bb2 100644
> > > --- a/drivers/iommu/intel-iommu.c
> > > +++ b/drivers/iommu/intel-iommu.c
> > > @@ -5693,8 +5693,14 @@ static phys_addr_t
> > intel_iommu_iova_to_phys(struct
> > > iommu_domain *domain,
> > > u64 phys = 0;
> > >
> > > pte = pfn_to_dma_pte(dmar_domain, iova >> VTD_PAGE_SHIFT,
> > &level);
> > > - if (pte)
> > > + if (pte) {
> > > + unsigned long offset_mask;
> > > +
> > > + offset_mask = BIT_MASK(level_to_offset_bits(level) +
> > > + VTD_PAGE_SHIFT) - 1;
> > > phys = dma_pte_addr(pte);
> > > + phys += iova & (bitmask - 1);
> > Did you mean:
> >
> > phys += iova & (offset_mask - 1);
> >
> > > + }
> >
> > At that point why not define a helper:
> >
> > static inline unsigned long offset_mask(int level)
> > {
> > return BIT_MASK(level_to_offset_bits(level) +
> > VTD_PAGE_SHIFT) - 1;
> > }
> >
> > At which point the above would reduce to:
> >
> > if (pte)
> > phys = dma_pte_addr(pte) + iova & offset_mask(level) - 1;
> >
> >
> >
> > Let me introduce the helper function and upstream the patch again. Thank
> > you Lu and Moritz.
> >
> >
> > I might've fatfingered something here, but you get the idea :)
> >
> > Cheers,
> > Moritz
> >
> >
> > On Tue, Feb 18, 2020 at 8:42 PM Moritz Fischer <mdf@xxxxxxxxxx
> > <mailto:mdf@xxxxxxxxxx>> wrote:
> >
> > Hi Baolu, Yonghyun
> >
> > On Wed, Feb 19, 2020 at 11:15:36AM +0800, Lu Baolu wrote:
> > > Hi Yonghyun,
> > >
> > > Thanks for the patch.
> > >
> > > On 2020/2/19 6:23, Yonghyun Hwang wrote:
> > > > intel_iommu_iova_to_phys() has a bug when it translates an IOVA
> > for a huge
> > > > page onto its corresponding physical address. This commit fixes
> > the bug by
> > > > accomodating the level of page entry for the IOVA and adds
> > IOVA's lower
> > > > address to the physical address. >
> > > > Signed-off-by: Yonghyun Hwang <yonghyun@xxxxxxxxxx
> > <mailto:yonghyun@xxxxxxxxxx>>
> > > > ---
> > > > drivers/iommu/intel-iommu.c | 11 +++++++++--
> > > > 1 file changed, 9 insertions(+), 2 deletions(-)
> > > >
> > > > diff --git a/drivers/iommu/intel-iommu.c
> > b/drivers/iommu/intel-iommu.c
> > > > index 0c8d81f56a30..ed6e69adb578 100644
> > > > --- a/drivers/iommu/intel-iommu.c
> > > > +++ b/drivers/iommu/intel-iommu.c
> > > > @@ -5555,13 +5555,20 @@ static phys_addr_t
> > intel_iommu_iova_to_phys(struct iommu_domain *domain,
> > > > struct dma_pte *pte;
> > > > int level = 0;
> > > > u64 phys = 0;
> > > > + const unsigned long pfn = iova >> VTD_PAGE_SHIFT;
> > >
> > > Why do you need a "const unsigned long" here?
> > >
> > > > if (dmar_domain->flags & DOMAIN_FLAG_LOSE_CHILDREN)
> > > > return 0;
> > > > - pte = pfn_to_dma_pte(dmar_domain, iova >> VTD_PAGE_SHIFT,
> > &level);
> > > > - if (pte)
> > > > + pte = pfn_to_dma_pte(dmar_domain, pfn, &level);
> > > > + if (pte) {
> > > > phys = dma_pte_addr(pte);
> > > > + if (level > 1)
> > > > + phys += (pfn &
> > > > + ((1UL <<
> > level_to_offset_bits(level)) - 1))
> > > > + << VTD_PAGE_SHIFT;
> > > > + phys += iova & (VTD_PAGE_SIZE - 1);
> > >
> > > How about
> >
> > >
> > > diff --git a/drivers/iommu/intel-iommu.c
> > b/drivers/iommu/intel-iommu.c
> > > index 9dc37672bf89..bd17c2510bb2 100644
> > > --- a/drivers/iommu/intel-iommu.c
> > > +++ b/drivers/iommu/intel-iommu.c
> > > @@ -5693,8 +5693,14 @@ static phys_addr_t
> > intel_iommu_iova_to_phys(struct
> > > iommu_domain *domain,
> > > u64 phys = 0;
> > >
> > > pte = pfn_to_dma_pte(dmar_domain, iova >> VTD_PAGE_SHIFT,
> > &level);
> > > - if (pte)
> > > + if (pte) {
> > > + unsigned long offset_mask;
> > > +
> > > + offset_mask = BIT_MASK(level_to_offset_bits(level) +
> > > + VTD_PAGE_SHIFT) - 1;
> > > phys = dma_pte_addr(pte);
> > > + phys += iova & (bitmask - 1);
> > Did you mean:
> >
> > phys += iova & (offset_mask - 1);
> >
> > > + }
> >
> > At that point why not define a helper:
> >
> > static inline unsigned long offset_mask(int level)
> > {
> > return BIT_MASK(level_to_offset_bits(level) +
> > VTD_PAGE_SHIFT) - 1;
> > }
> >
> > At which point the above would reduce to:
> >
> > if (pte)
> > phys = dma_pte_addr(pte) + iova & offset_mask(level) - 1;
> >
> > I might've fatfingered something here, but you get the idea :)
> >
> > Cheers,
> > Moritz
> >