Re: [PATCH] iommu/vtd: fix address translation for superpages

[Jan Beulich <jbeulich@xxxxxxxx>]

On 10.05.2023 12:22, Roger Pau Monné wrote:
> On Wed, May 10, 2023 at 12:00:51PM +0200, Jan Beulich wrote:
>> On 10.05.2023 10:27, Roger Pau Monné wrote:
>>> On Tue, May 09, 2023 at 06:06:45PM +0200, Jan Beulich wrote:
>>>> On 09.05.2023 12:41, Roger Pau Monne wrote:
>>>>> When translating an address that falls inside of a superpage in the
>>>>> IOMMU page tables the fetching of the PTE physical address field
>>>>> wasn't using dma_pte_addr(), which caused the returned data to be
>>>>> corrupt as it would contain bits not related to the address field.
>>>>
>>>> I'm afraid I don't understand:
>>>>
>>>>> --- a/xen/drivers/passthrough/vtd/iommu.c
>>>>> +++ b/xen/drivers/passthrough/vtd/iommu.c
>>>>> @@ -359,16 +359,18 @@ static uint64_t addr_to_dma_page_maddr(struct 
>>>>> domain *domain, daddr_t addr,
>>>>>  
>>>>>              if ( !alloc )
>>>>>              {
>>>>> -                pte_maddr = 0;
>>>>>                  if ( !dma_pte_present(*pte) )
>>>>> +                {
>>>>> +                    pte_maddr = 0;
>>>>>                      break;
>>>>> +                }
>>>>>  
>>>>>                  /*
>>>>>                   * When the leaf entry was requested, pass back the full 
>>>>> PTE,
>>>>>                   * with the address adjusted to account for the residual 
>>>>> of
>>>>>                   * the walk.
>>>>>                   */
>>>>> -                pte_maddr = pte->val +
>>>>> +                pte_maddr +=
>>>>>                      (addr & ((1UL << level_to_offset_bits(level)) - 1) &
>>>>>                       PAGE_MASK);
>>>>
>>>> With this change you're now violating what the comment says (plus what
>>>> the comment ahead of the function says). And it says what it says for
>>>> a reason - see intel_iommu_lookup_page(), which I think your change is
>>>> breaking.
>>>
>>> Hm, but the code in intel_iommu_lookup_page() is now wrong as it takes
>>> the bits in DMA_PTE_CONTIG_MASK as part of the physical address when
>>> doing the conversion to mfn?  maddr_to_mfn() doesn't perform a any
>>> masking to remove the bits above PADDR_BITS.
>>
>> Oh, right. But that's a missing dma_pte_addr() in intel_iommu_lookup_page()
>> then. (It would likely be better anyway to switch "uint64_t val" to
>> "struct dma_pte pte" there, to make more visible that it's a PTE we're
>> dealing with.) I indeed overlooked this aspect when doing the earlier
>> change.
> 
> I guess I'm still confused, as the other return value for target == 0
> (when the address is not part of a superpage) does return
> dma_pte_addr(pte).  I think that needs further fixing then.

Hmm, indeed. But I think it's worse than this: addr_to_dma_page_maddr()
also does one too many iterations in that case. All "normal" callers
supply a positive "target". We need to terminate the walk at level 1
also when target == 0.

Jan