Re: [PATCH v3 12/17] intel_iommu: Add support for PASID-based device IOTLB invalidation

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

On 17/12/2024 03:13, Jason Wang 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.
>
>
> On Mon, Dec 16, 2024 at 4:22 PM Duan, Zhenzhong
> <zhenzhong.duan@intel.com> wrote:
>>
>>
>>> -----Original Message-----
>>> From: Jason Wang <jasowang@redhat.com>
>>> Sent: Sunday, September 29, 2024 9:59 AM
>>> Subject: Re: [PATCH v3 12/17] intel_iommu: Add support for PASID-based device
>>> IOTLB invalidation
>>>
>>> On Fri, Sep 27, 2024 at 3:18 PM Duan, Zhenzhong
>>> <zhenzhong.duan@intel.com> wrote:
>>>>
>>>>
>>>>> -----Original Message-----
>>>>> From: Jason Wang <jasowang@redhat.com>
>>>>> Subject: Re: [PATCH v3 12/17] intel_iommu: Add support for PASID-based
>>>>> device IOTLB invalidation
>>>>>
>>>>> On Wed, Sep 11, 2024 at 1:27 PM Zhenzhong Duan
>>>>> <zhenzhong.duan@intel.com> wrote:
>>>>>> From: Clément Mathieu--Drif <clement.mathieu--drif@eviden.com>
>>>>>>
>>>>>> Signed-off-by: Clément Mathieu--Drif <clement.mathieu--drif@eviden.com>
>>>>>> Signed-off-by: Zhenzhong Duan <zhenzhong.duan@intel.com>
>>>>>> ---
>>>>>>   hw/i386/intel_iommu_internal.h | 11 ++++++++
>>>>>>   hw/i386/intel_iommu.c          | 50
>>>>> ++++++++++++++++++++++++++++++++++
>>>>>>   2 files changed, 61 insertions(+)
>>>>>>
>>>>>> diff --git a/hw/i386/intel_iommu_internal.h
>>>>> b/hw/i386/intel_iommu_internal.h
>>>>>> index 4f2c3a9350..52bdbf3bc5 100644
>>>>>> --- a/hw/i386/intel_iommu_internal.h
>>>>>> +++ b/hw/i386/intel_iommu_internal.h
>>>>>> @@ -375,6 +375,7 @@ typedef union VTDInvDesc VTDInvDesc;
>>>>>>   #define VTD_INV_DESC_WAIT               0x5 /* Invalidation Wait Descriptor
>>>>> */
>>>>>>   #define VTD_INV_DESC_PIOTLB             0x6 /* PASID-IOTLB Invalidate Desc
>>>>> */
>>>>>>   #define VTD_INV_DESC_PC                 0x7 /* PASID-cache Invalidate Desc */
>>>>>> +#define VTD_INV_DESC_DEV_PIOTLB         0x8 /* PASID-based-DIOTLB
>>>>> inv_desc*/
>>>>>>   #define VTD_INV_DESC_NONE               0   /* Not an Invalidate Descriptor
>>>>> */
>>>>>>   /* Masks for Invalidation Wait Descriptor*/
>>>>>> @@ -413,6 +414,16 @@ typedef union VTDInvDesc VTDInvDesc;
>>>>>>   #define VTD_INV_DESC_DEVICE_IOTLB_RSVD_HI 0xffeULL
>>>>>>   #define VTD_INV_DESC_DEVICE_IOTLB_RSVD_LO 0xffff0000ffe0fff8
>>>>>>
>>>>>> +/* Mask for PASID Device IOTLB Invalidate Descriptor */
>>>>>> +#define VTD_INV_DESC_PASID_DEVICE_IOTLB_ADDR(val) ((val) & \
>>>>>> +                                                   0xfffffffffffff000ULL)
>>>>>> +#define VTD_INV_DESC_PASID_DEVICE_IOTLB_SIZE(val) ((val >> 11) & 0x1)
>>>>>> +#define VTD_INV_DESC_PASID_DEVICE_IOTLB_GLOBAL(val) ((val) & 0x1)
>>>>>> +#define VTD_INV_DESC_PASID_DEVICE_IOTLB_SID(val) (((val) >> 16) &
>>>>> 0xffffULL)
>>>>>> +#define VTD_INV_DESC_PASID_DEVICE_IOTLB_PASID(val) ((val >> 32) &
>>>>> 0xfffffULL)
>>>>>> +#define VTD_INV_DESC_PASID_DEVICE_IOTLB_RSVD_HI 0x7feULL
>>>>>> +#define VTD_INV_DESC_PASID_DEVICE_IOTLB_RSVD_LO
>>>>> 0xfff000000000f000ULL
>>>>>> +
>>>>>>   /* Rsvd field masks for spte */
>>>>>>   #define VTD_SPTE_SNP 0x800ULL
>>>>>>
>>>>>> diff --git a/hw/i386/intel_iommu.c b/hw/i386/intel_iommu.c
>>>>>> index d28c862598..4cf56924e1 100644
>>>>>> --- a/hw/i386/intel_iommu.c
>>>>>> +++ b/hw/i386/intel_iommu.c
>>>>>> @@ -3017,6 +3017,49 @@ static void
>>>>> do_invalidate_device_tlb(VTDAddressSpace *vtd_dev_as,
>>>>>>       memory_region_notify_iommu(&vtd_dev_as->iommu, 0, event);
>>>>>>   }
>>>>>>
>>>>>> +static bool vtd_process_device_piotlb_desc(IntelIOMMUState *s,
>>>>>> +                                           VTDInvDesc *inv_desc)
>>>>>> +{
>>>>>> +    uint16_t sid;
>>>>>> +    VTDAddressSpace *vtd_dev_as;
>>>>>> +    bool size;
>>>>>> +    bool global;
>>>>>> +    hwaddr addr;
>>>>>> +    uint32_t pasid;
>>>>>> +
>>>>>> +    if ((inv_desc->hi & VTD_INV_DESC_PASID_DEVICE_IOTLB_RSVD_HI) ||
>>>>>> +         (inv_desc->lo & VTD_INV_DESC_PASID_DEVICE_IOTLB_RSVD_LO)) {
>>>>>> +        error_report_once("%s: invalid pasid-based dev iotlb inv desc:"
>>>>>> +                          "hi=%"PRIx64 "(reserved nonzero)",
>>>>>> +                          __func__, inv_desc->hi);
>>>>>> +        return false;
>>>>>> +    }
>>>>>> +
>>>>>> +    global = VTD_INV_DESC_PASID_DEVICE_IOTLB_GLOBAL(inv_desc->hi);
>>>>>> +    size = VTD_INV_DESC_PASID_DEVICE_IOTLB_SIZE(inv_desc->hi);
>>>>>> +    addr = VTD_INV_DESC_PASID_DEVICE_IOTLB_ADDR(inv_desc->hi);
>>>>>> +    sid = VTD_INV_DESC_PASID_DEVICE_IOTLB_SID(inv_desc->lo);
>>>>>> +    if (global) {
>>>>>> +        QLIST_FOREACH(vtd_dev_as, &s->vtd_as_with_notifiers, next) {
>>>>>> +            if ((vtd_dev_as->pasid != PCI_NO_PASID) &&
>>>>>> +                (PCI_BUILD_BDF(pci_bus_num(vtd_dev_as->bus),
>>>>>> +                                           vtd_dev_as->devfn) == sid)) {
>>>>>> +                do_invalidate_device_tlb(vtd_dev_as, size, addr);
>>>>>> +            }
>>>>>> +        }
>>>>>> +    } else {
>>>>>> +        pasid = VTD_INV_DESC_PASID_DEVICE_IOTLB_PASID(inv_desc->lo);
>>>>>> +        vtd_dev_as = vtd_get_as_by_sid_and_pasid(s, sid, pasid);
>>>>>> +        if (!vtd_dev_as) {
>>>>>> +            return true;
>>>>>> +        }
>>>>>> +
>>>>>> +        do_invalidate_device_tlb(vtd_dev_as, size, addr);
>>>>> Question:
>>>>>
>>>>> I wonder if current vhost (which has a device IOTLB abstraction via
>>>>> virtio-pci) can work with this (PASID based IOTLB invalidation)
>>>> Currently, it depends on if caching-mode is on. If it's off, vhost works. E.g.:
>>>>
>>>> -device intel-iommu,caching-mode=off,dma-drain=on,device-iotlb=on,x-
>>> scalable-mode=on
>>>> -netdev tap,id=tap0,vhost=on,script=/etc/qemu-ifup
>>>> -device virtio-net-pci,netdev=tap0,bus=root0,iommu_platform=on,ats=on
>>>>
>>>> It doesn't work currently when caching-mode is on.
>>>> Reason is linux kernel has an optimization to send only piotlb invalidation,
>>>> no device-piotlb invalidation is sent. But I heard from Yi the optimization
>>>> will be dropped, then it will work too when caching-mode is on.
>>> Great, if possible please copy me when sending those fixes.
>> FYI, I just found the optimization had already been dropped since April 2024 by commit https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=304b3bde24b58515a75fd198beb52ca57df6275f
>>
>> After updating guest kernel to a new version containing above commit,
>> vhost works irrespective the value of caching-mode.
>>
>> Thanks
>> Zhenzhong
> Great. Thanks for the updating.

Good news,
Thanks Zhenzhong!