The setting of the SNP bit in the intel-iommu page tables should not
be dependent on the current capability of the iommu domain. The
current VT-d spec (2.2) indicates the SNP bit is "treated as
reserved[0] by hardware implementations not supporting Snoop Control".
Furthermore, section 3.7.3 indicates:
If the Snoop Control (SC) field in extended capability Register is
reported as 0, snoop behavior for access to the page mapped through
second-level translation is determined by the no-snoop attribute in
the request.
This all seems to indicate that hardware incapable of Snoop Control
will handle the SNP bit as zero regardless of the value stored in
the PTE.
The trouble with the current implementation is that mapping flags
depend on the state of the iommu domain at the time of the mapping,
yet no attempt is made to update existing mappings when the iommu
domain composition changes. This leaves the iommu domain in a state
where some mappings may enforce coherency, others do not, and the user
of the IOMMU API has no ability to later enable the desired flags
atomically with respect to DMA.
If we always honor the IOMMU_CACHE flag then an IOMMU API user who
specifies IOMMU_CACHE for all mappings can assume that the coherency
of the mappings within a domain follow the coherency capability of
the domain itself.
Signed-off-by: Alex Williamson <[email protected]>
---
drivers/iommu/intel-iommu.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/drivers/iommu/intel-iommu.c b/drivers/iommu/intel-iommu.c
index 15e9b57..c46c6a6 100644
--- a/drivers/iommu/intel-iommu.c
+++ b/drivers/iommu/intel-iommu.c
@@ -4084,7 +4084,7 @@ static int intel_iommu_map(struct iommu_domain *domain,
prot |= DMA_PTE_READ;
if (iommu_prot & IOMMU_WRITE)
prot |= DMA_PTE_WRITE;
- if ((iommu_prot & IOMMU_CACHE) && dmar_domain->iommu_snooping)
+ if (iommu_prot & IOMMU_CACHE)
prot |= DMA_PTE_SNP;
max_addr = iova + size;
Alex Williamson wrote on 2013-12-24:
> David,
>
> Any comments on this patch? Thanks,
>
Hi Alex,
There do have some IOMMUs will treat SNP bit in the PTE as reserved (0) and will cause a reserved field violation fault if it is set but hardware not support snoop-control(bit 7 in ECAP_REG is 0). So your patch seems wrong to me.
> Alex
>
> On Tue, 2013-10-29 at 10:21 -0600, Alex Williamson wrote:
>> The setting of the SNP bit in the intel-iommu page tables should not
>> be dependent on the current capability of the iommu domain. The
>> current VT-d spec (2.2) indicates the SNP bit is "treated as
>> reserved[0] by hardware implementations not supporting Snoop Control".
>> Furthermore, section 3.7.3 indicates:
>>
>> If the Snoop Control (SC) field in extended capability Register is
>> reported as 0, snoop behavior for access to the page mapped through
>> second-level translation is determined by the no-snoop attribute in
>> the request.
>> This all seems to indicate that hardware incapable of Snoop Control
>> will handle the SNP bit as zero regardless of the value stored in
>> the PTE.
>>
>> The trouble with the current implementation is that mapping flags
>> depend on the state of the iommu domain at the time of the mapping,
>> yet no attempt is made to update existing mappings when the iommu
>> domain composition changes. This leaves the iommu domain in a state
>> where some mappings may enforce coherency, others do not, and the
>> user of the IOMMU API has no ability to later enable the desired
>> flags atomically with respect to DMA.
>>
>> If we always honor the IOMMU_CACHE flag then an IOMMU API user who
>> specifies IOMMU_CACHE for all mappings can assume that the coherency
>> of the mappings within a domain follow the coherency capability of
>> the domain itself.
>>
>> Signed-off-by: Alex Williamson <[email protected]>
>> ---
>> drivers/iommu/intel-iommu.c | 2 +-
>> 1 file changed, 1 insertion(+), 1 deletion(-)
>> diff --git a/drivers/iommu/intel-iommu.c
>> b/drivers/iommu/intel-iommu.c index 15e9b57..c46c6a6 100644
>> --- a/drivers/iommu/intel-iommu.c
>> +++ b/drivers/iommu/intel-iommu.c
>> @@ -4084,7 +4084,7 @@ static int intel_iommu_map(struct iommu_domain
> *domain,
>> prot |= DMA_PTE_READ;
>> if (iommu_prot & IOMMU_WRITE)
>> prot |= DMA_PTE_WRITE;
>> - if ((iommu_prot & IOMMU_CACHE) && dmar_domain->iommu_snooping)
>> + if (iommu_prot & IOMMU_CACHE)
>> prot |= DMA_PTE_SNP;
>>
>> max_addr = iova + size;
>
>
Best regards,
Yang
????{.n?+???????+%?????ݶ??w??{.n?+????{??G?????{ay?ʇڙ?,j??f???h?????????z_??(?階?ݢj"???m??????G????????????&???~???iO???z??v?^?m????????????I?
On Tue, 2014-01-07 at 00:54 +0000, Zhang, Yang Z wrote:
> Alex Williamson wrote on 2013-12-24:
> > David,
> >
> > Any comments on this patch? Thanks,
> >
>
> Hi Alex,
>
> There do have some IOMMUs will treat SNP bit in the PTE as reserved
> (0) and will cause a reserved field violation fault if it is set but
> hardware not support snoop-control(bit 7 in ECAP_REG is 0). So your
> patch seems wrong to me.
Thanks for the reply Yang. So effectively IOMMU_CACHE (SNP) is unusable
on intel-iommu if we want a domain that can manage any device in the
system, right? We have no way to atomically change the SNP bit across a
domain so the only chance for enabling it is if all the DRHD units
support snoop-control, but then a hot-added DRHD makes that impossible
to predict. Is there any way we can make intel-iommu "do the right
thing"? These are rather low-level implementation details of
intel-iommu that I'd rather not know about as a user of the IOMMU API.
Thanks,
Alex
> > On Tue, 2013-10-29 at 10:21 -0600, Alex Williamson wrote:
> >> The setting of the SNP bit in the intel-iommu page tables should not
> >> be dependent on the current capability of the iommu domain. The
> >> current VT-d spec (2.2) indicates the SNP bit is "treated as
> >> reserved[0] by hardware implementations not supporting Snoop Control".
> >> Furthermore, section 3.7.3 indicates:
> >>
> >> If the Snoop Control (SC) field in extended capability Register is
> >> reported as 0, snoop behavior for access to the page mapped through
> >> second-level translation is determined by the no-snoop attribute in
> >> the request.
> >> This all seems to indicate that hardware incapable of Snoop Control
> >> will handle the SNP bit as zero regardless of the value stored in
> >> the PTE.
> >>
> >> The trouble with the current implementation is that mapping flags
> >> depend on the state of the iommu domain at the time of the mapping,
> >> yet no attempt is made to update existing mappings when the iommu
> >> domain composition changes. This leaves the iommu domain in a state
> >> where some mappings may enforce coherency, others do not, and the
> >> user of the IOMMU API has no ability to later enable the desired
> >> flags atomically with respect to DMA.
> >>
> >> If we always honor the IOMMU_CACHE flag then an IOMMU API user who
> >> specifies IOMMU_CACHE for all mappings can assume that the coherency
> >> of the mappings within a domain follow the coherency capability of
> >> the domain itself.
> >>
> >> Signed-off-by: Alex Williamson <[email protected]>
> >> ---
> >> drivers/iommu/intel-iommu.c | 2 +-
> >> 1 file changed, 1 insertion(+), 1 deletion(-)
> >> diff --git a/drivers/iommu/intel-iommu.c
> >> b/drivers/iommu/intel-iommu.c index 15e9b57..c46c6a6 100644
> >> --- a/drivers/iommu/intel-iommu.c
> >> +++ b/drivers/iommu/intel-iommu.c
> >> @@ -4084,7 +4084,7 @@ static int intel_iommu_map(struct iommu_domain
> > *domain,
> >> prot |= DMA_PTE_READ;
> >> if (iommu_prot & IOMMU_WRITE)
> >> prot |= DMA_PTE_WRITE;
> >> - if ((iommu_prot & IOMMU_CACHE) && dmar_domain->iommu_snooping)
> >> + if (iommu_prot & IOMMU_CACHE)
> >> prot |= DMA_PTE_SNP;
> >>
> >> max_addr = iova + size;
> >
> >
>
>
> Best regards,
> Yang
>
>