Re: [PATCH rfcv2 09/17] intel_iommu: Flush stage-1 cache in iotlb invalidation

[CLEMENT MATHIEU--DRIF <clement.mathieu--drif@eviden.com>]

Hi Zhenzhong

On 22/05/2024 08:23, Zhenzhong Duan wrote:
> Caution: External email. Do not open attachments or click links, unless this email comes from a known sender and you know the content is safe.
>
>
> According to spec, Page-Selective-within-Domain Invalidation (11b):
>
> 1. IOTLB entries caching second-stage mappings (PGTT=010b) or pass-through
> (PGTT=100b) mappings associated with the specified domain-id and the
> input-address range are invalidated.
> 2. IOTLB entries caching first-stage (PGTT=001b) or nested (PGTT=011b)
> mapping associated with specified domain-id are invalidated.
>
> So per spec definition the Page-Selective-within-Domain Invalidation
> needs to flush first stage and nested cached IOTLB enties as well.
>
> We don't support nested yet and pass-through mapping is never cached,
> so what in iotlb cache are only first-stage and second-stage mappings.
>
> Add a tag pgtt in VTDIOTLBEntry to mark PGTT type of the mapping and
> invalidate entries based on PGTT type.
>
> Signed-off-by: Zhenzhong Duan <zhenzhong.duan@intel.com>
> ---
>   include/hw/i386/intel_iommu.h |  1 +
>   hw/i386/intel_iommu.c         | 20 +++++++++++++++++---
>   2 files changed, 18 insertions(+), 3 deletions(-)
>
> diff --git a/include/hw/i386/intel_iommu.h b/include/hw/i386/intel_iommu.h
> index 011f374883..b0d5b5a5be 100644
> --- a/include/hw/i386/intel_iommu.h
> +++ b/include/hw/i386/intel_iommu.h
> @@ -156,6 +156,7 @@ struct VTDIOTLBEntry {
>       uint64_t pte;
>       uint64_t mask;
>       uint8_t access_flags;
> +    uint8_t pgtt;
>   };
>
>   /* VT-d Source-ID Qualifier types */
> diff --git a/hw/i386/intel_iommu.c b/hw/i386/intel_iommu.c
> index 0801112e2e..0078bad9d4 100644
> --- a/hw/i386/intel_iommu.c
> +++ b/hw/i386/intel_iommu.c
> @@ -287,9 +287,21 @@ static gboolean vtd_hash_remove_by_page(gpointer key, gpointer value,
>       VTDIOTLBPageInvInfo *info = (VTDIOTLBPageInvInfo *)user_data;
>       uint64_t gfn = (info->addr >> VTD_PAGE_SHIFT_4K) & info->mask;
>       uint64_t gfn_tlb = (info->addr & entry->mask) >> VTD_PAGE_SHIFT_4K;
> -    return (entry->domain_id == info->domain_id) &&
> -            (((entry->gfn & info->mask) == gfn) ||
> -             (entry->gfn == gfn_tlb));
> +
> +    if (entry->domain_id != info->domain_id) {
> +        return false;
> +    }
> +
> +    /*
> +     * According to spec, IOTLB entries caching first-stage (PGTT=001b) or
> +     * nested (PGTT=011b) mapping associated with specified domain-id are
> +     * invalidated. Nested isn't supported yet, so only need to check 001b.
> +     */
> +    if (entry->pgtt == VTD_SM_PASID_ENTRY_FLT) {
> +        return true;
> +    }
> +
> +    return (entry->gfn & info->mask) == gfn || entry->gfn == gfn_tlb;
>   }
>
>   /* Reset all the gen of VTDAddressSpace to zero and set the gen of
> @@ -382,6 +394,8 @@ static void vtd_update_iotlb(IntelIOMMUState *s, uint16_t source_id,
>       entry->access_flags = access_flags;
>       entry->mask = vtd_slpt_level_page_mask(level);
>       entry->pasid = pasid;
> +    entry->pgtt = s->scalable_modern ? VTD_SM_PASID_ENTRY_FLT
> +                                     : VTD_SM_PASID_ENTRY_SLT;
What about passing pgtt as a parameter so that the translation type 
detection is done only once (in vtd_do_iommu_translate)?
>
>       key->gfn = gfn;
>       key->sid = source_id;
> --
> 2.34.1
>
#cmd