RE: [PATCH rfcv1 6/6] intel_iommu: add a framework to check and sync host IOMMU cap/ecap

[Duan, Zhenzhong]

>-----Original Message-----
>From: Eric Auger <eric.auger@redhat.com>
>Subject: Re: [PATCH rfcv1 6/6] intel_iommu: add a framework to check and
>sync host IOMMU cap/ecap
>
>
>
>On 1/18/24 10:30, Duan, Zhenzhong wrote:
>> Hi Eric,
>>
>>> -----Original Message-----
>>> From: Eric Auger <eric.auger@redhat.com>
>>> Subject: Re: [PATCH rfcv1 6/6] intel_iommu: add a framework to check
>and
>>> sync host IOMMU cap/ecap
>>>
>>> Hi Zhenzhong,
>>>
>>> On 1/15/24 11:13, Zhenzhong Duan wrote:
>>>> From: Yi Liu <yi.l.liu@intel.com>
>>>>
>>>> Add a framework to check and synchronize host IOMMU cap/ecap with
>>>> vIOMMU cap/ecap.
>>>>
>>>> Currently only stage-2 translation is supported which is backed by
>>>> shadow page table on host side. So we don't need exact matching of
>>>> each bit of cap/ecap between vIOMMU and host. However, we can still
>>>> utilize this framework to ensure compatibility of host and vIOMMU's
>>>> address width at least, i.e., vIOMMU's aw_bits <= host aw_bits,
>>>> which is missed before.
>>>>
>>>> When stage-1 translation is supported in future, a.k.a. scalable
>>>> modern mode, we need to ensure compatibility of each bits. Some
>>>> bits are user controllable, they should be checked with host side
>>>> to ensure compatibility. Other bits are not, they should be synced
>>>> into vIOMMU cap/ecap for compatibility.
>>>>
>>>> The sequence will be:
>>>>
>>>> vtd_cap_init() initializes iommu->cap/ecap. ---- vtd_cap_init()
>>>> iommu->host_cap/ecap is initialized as iommu->cap/ecap.  ---- vtd_init()
>>>> iommu->host_cap/ecap is checked and updated some bits with host
>>> cap/ecap. ---- vtd_sync_hw_info()
>>>> iommu->cap/ecap is finalized as iommu->host_cap/ecap.  ----
>>> vtd_machine_done_hook()
>>>> iommu->host_cap/ecap is a temporary storage to hold intermediate
>value
>>>> when synthesize host cap/ecap and vIOMMU's initial configured
>cap/ecap.
>>>>
>>>> Signed-off-by: Yi Liu <yi.l.liu@intel.com>
>>>> Signed-off-by: Yi Sun <yi.y.sun@linux.intel.com>
>>>> Signed-off-by: Zhenzhong Duan <zhenzhong.duan@intel.com>
>>>> ---
>>>>  include/hw/i386/intel_iommu.h |  4 ++
>>>>  hw/i386/intel_iommu.c         | 78
>>> +++++++++++++++++++++++++++++++----
>>>>  2 files changed, 75 insertions(+), 7 deletions(-)
>>>>
>>>> diff --git a/include/hw/i386/intel_iommu.h
>>> b/include/hw/i386/intel_iommu.h
>>>> index c65fdde56f..b8abbcce12 100644
>>>> --- a/include/hw/i386/intel_iommu.h
>>>> +++ b/include/hw/i386/intel_iommu.h
>>>> @@ -292,6 +292,9 @@ struct IntelIOMMUState {
>>>>      uint64_t cap;                   /* The value of capability reg */
>>>>      uint64_t ecap;                  /* The value of extended capability 
>>>> reg */
>>>>
>>>> +    uint64_t host_cap;              /* The value of host capability reg */
>>>> +    uint64_t host_ecap;             /* The value of host ext-capability 
>>>> reg */
>>>> +
>>>>      uint32_t context_cache_gen;     /* Should be in [1,MAX] */
>>>>      GHashTable *iotlb;              /* IOTLB */
>>>>
>>>> @@ -314,6 +317,7 @@ struct IntelIOMMUState {
>>>>      bool dma_translation;           /* Whether DMA translation supported
>*/
>>>>      bool pasid;                     /* Whether to support PASID */
>>>>
>>>> +    bool cap_finalized;             /* Whether VTD capability finalized */
>>>>      /*
>>>>       * Protects IOMMU states in general.  Currently it protects the
>>>>       * per-IOMMU IOTLB cache, and context entry cache in
>>> VTDAddressSpace.
>>>> diff --git a/hw/i386/intel_iommu.c b/hw/i386/intel_iommu.c
>>>> index 4c1d058ebd..be03fcbf52 100644
>>>> --- a/hw/i386/intel_iommu.c
>>>> +++ b/hw/i386/intel_iommu.c
>>>> @@ -3819,6 +3819,47 @@ VTDAddressSpace
>>> *vtd_find_add_as(IntelIOMMUState *s, PCIBus *bus,
>>>>      return vtd_dev_as;
>>>>  }
>>>>
>>>> +static bool vtd_sync_hw_info(IntelIOMMUState *s, struct
>>> iommu_hw_info_vtd *vtd,
>>>> +                             Error **errp)
>>>> +{
>>>> +    uint64_t addr_width;
>>>> +
>>>> +    addr_width = (vtd->cap_reg >> 16) & 0x3fULL;
>>>> +    if (s->aw_bits > addr_width) {
>>>> +        error_setg(errp, "User aw-bits: %u > host address width: %lu",
>>>> +                   s->aw_bits, addr_width);
>>>> +        return false;
>>>> +    }
>>>> +
>>>> +    /* TODO: check and sync host cap/ecap into vIOMMU cap/ecap */
>>>> +
>>>> +    return true;
>>>> +}
>>>> +
>>>> +/*
>>>> + * virtual VT-d which wants nested needs to check the host IOMMU
>>>> + * nesting cap info behind the assigned devices. Thus that vIOMMU
>>>> + * could bind guest page table to host.
>>>> + */
>>>> +static bool vtd_check_idev(IntelIOMMUState *s, IOMMUFDDevice
>*idev,
>>>> +                           Error **errp)
>>>> +{
>>>> +    struct iommu_hw_info_vtd vtd;
>>>> +    enum iommu_hw_info_type type =
>>> IOMMU_HW_INFO_TYPE_INTEL_VTD;
>>>> +
>>>> +    if (iommufd_device_get_info(idev, &type, sizeof(vtd), &vtd)) {
>>>> +        error_setg(errp, "Failed to get IOMMU capability!!!");
>>>> +        return false;
>>>> +    }
>>>> +
>>>> +    if (type != IOMMU_HW_INFO_TYPE_INTEL_VTD) {
>>>> +        error_setg(errp, "IOMMU hardware is not compatible!!!");
>>>> +        return false;
>>>> +    }
>>>> +
>>>> +    return vtd_sync_hw_info(s, &vtd, errp);
>>>> +}
>>>> +
>>>>  static int vtd_dev_set_iommu_device(PCIBus *bus, void *opaque,
>int32_t
>>> devfn,
>>>>                                      IOMMUFDDevice *idev, Error **errp)
>>>>  {
>>>> @@ -3837,6 +3878,10 @@ static int
>vtd_dev_set_iommu_device(PCIBus
>>> *bus, void *opaque, int32_t devfn,
>>>>          return 0;
>>>>      }
>>>>
>>>> +    if (!vtd_check_idev(s, idev, errp)) {In
>>> In
>>> [RFC 0/7] VIRTIO-IOMMU/VFIO: Fix host iommu geometry handling for
>>> hotplugged devices
>>> https://lore.kernel.org/all/20240117080414.316890-1-
>>> eric.auger@redhat.com/
>>>
>>> I also attempt to pass host iommu info to the virtio-iommu but with
>>> legacy BE.
>> I think your patch works with iommufd BE too😊 Because iommufd BE
>> also fills bcontainer->iova_ranges in iommufd_cdev_get_info_iova_range().
>correct. I wanted to emphasize that we also have the need to pass host
>iommu info in legacy mode for instance. In this series you introduce an
>object that works with the iommufd backed but I think if we go this way
>we would need another one for the legacy device. So maybe introducing a
>base object derived into 2 ones may be the most appropriate? Maybe,
>given the assumption that we will use iommufd for new use cases this
>legacy object will implement much fewer interfaces but still.

How about this:

enum IOMMU_LEGACY_DEVICE_TYPE {
    IOMMU_LEGACY_VFIO_DEVICE,
    IOMMU_LEGACY_VDPA_DEVICE,
}

typedef struct IOMMULegacyDevice {
    enum IOMMU_LEGACY_DEVICE_TYPE type;

    /* common field */

    union {
        ....
    }

} IOMMULegacyDevice;

typedef struct IOMMUFDDevice {
    IOMMUFDBackend *iommufd;
    uint32_t dev_id;
    uint32_t ioas_id;
} IOMMUFDDevice;

enum IOMMUDEVICE_TYPE {
    IOMMUFD_DEVICE,
    IOMMU_LEGACY_DEVICE,
}

struct IOMMUDevice {
    enum IOMMU_DEVICE_TYPE type;

    /* common field */
    GList *iova_ranges;

    union {
        IOMMULegacyDevice legacy_dev;
        IOMMUFDDevice idev;
    }
}

>>
>>> In my case I want to pass the reserved memory regions which
>>> also model the aw.
>>> So this is a pretty similar use case.
>> Yes.
>>
>>> Why don't we pass the pointer to an opaque iommu_hw_info instead,
>>> through the PCIIOMMUOps?
>> Passing iommu_hw_info is ok for this series, but we want more from
>> IOMMUFDDevice in nesting series. I.e., allocate/free ioas/hwpt,
>> attach/detach from hwpt, get dirty bitmap, etc. It's more flexible to
>> let vIOMMU get what it want itself.
>OK, would be interesting to define the class for this object. Worth to
>be introduced either in the cover letter or in the 1st patch

Not a QOM class because we don't want it showed out through
query-qmp-schema.

Thanks
Zhenzhong

>
>Eric
>>
>>>
>>>
>>>> +        return -1;
>>>> +    }
>>>> +
>>>>      vtd_iommu_lock(s);
>>>>
>>>>      vtd_idev = g_hash_table_lookup(s->vtd_iommufd_dev, &key);
>>>> @@ -4071,10 +4116,11 @@ static void vtd_init(IntelIOMMUState *s)
>>>>  {
>>>>      X86IOMMUState *x86_iommu = X86_IOMMU_DEVICE(s);
>>>>
>>>> -    memset(s->csr, 0, DMAR_REG_SIZE);
>>>> -    memset(s->wmask, 0, DMAR_REG_SIZE);
>>>> -    memset(s->w1cmask, 0, DMAR_REG_SIZE);
>>>> -    memset(s->womask, 0, DMAR_REG_SIZE);
>>>> +    /* CAP/ECAP are initialized in machine create done stage */
>>>> +    memset(s->csr + DMAR_GCMD_REG, 0, DMAR_REG_SIZE -
>>> DMAR_GCMD_REG);
>>>> +    memset(s->wmask + DMAR_GCMD_REG, 0, DMAR_REG_SIZE -
>>> DMAR_GCMD_REG);
>>>> +    memset(s->w1cmask + DMAR_GCMD_REG, 0, DMAR_REG_SIZE -
>>> DMAR_GCMD_REG);
>>>> +    memset(s->womask + DMAR_GCMD_REG, 0, DMAR_REG_SIZE -
>>> DMAR_GCMD_REG);
>>> This change is not documented in the commit msg.
>>> Sorry I don't get why this is needed?
>> I'll doc it. Above we have one line to explain when cap/ecap are initialized.
>> vtd_init() is called in qemu init and guest reset. In qemu init,
>> Cap/ecap is finalized, after that we don't want cap/ecap to be changed.
>> So we bypass change to cap/ecap here.
>>
>>>>      s->root = 0;
>>>>      s->root_scalable = false;
>>>> @@ -4110,13 +4156,16 @@ static void vtd_init(IntelIOMMUState *s)
>>>>          vtd_spte_rsvd_large[3] &= ~VTD_SPTE_SNP;
>>>>      }
>>>>
>>>> -    vtd_cap_init(s);
>>>> +    if (!s->cap_finalized) {
>>>> +        vtd_cap_init(s);
>>>> +        s->host_cap = s->cap;
>>>> +        s->host_ecap = s->ecap;
>>>> +    }
>>>> +
>>>>      vtd_reset_caches(s);
>>>>
>>>>      /* Define registers with default values and bit semantics */
>>>>      vtd_define_long(s, DMAR_VER_REG, 0x10UL, 0, 0);
>>>> -    vtd_define_quad(s, DMAR_CAP_REG, s->cap, 0, 0);
>>>> -    vtd_define_quad(s, DMAR_ECAP_REG, s->ecap, 0, 0);
>>>>      vtd_define_long(s, DMAR_GCMD_REG, 0, 0xff800000UL, 0);
>>>>      vtd_define_long_wo(s, DMAR_GCMD_REG, 0xff800000UL);
>>>>      vtd_define_long(s, DMAR_GSTS_REG, 0, 0, 0);
>>>> @@ -4241,6 +4290,12 @@ static bool
>>> vtd_decide_config(IntelIOMMUState *s, Error **errp)
>>>>      return true;
>>>>  }
>>>>
>>>> +static void vtd_setup_capability_reg(IntelIOMMUState *s)
>>>> +{
>>>> +    vtd_define_quad(s, DMAR_CAP_REG, s->cap, 0, 0);
>>>> +    vtd_define_quad(s, DMAR_ECAP_REG, s->ecap, 0, 0);
>>>> +}
>>>> +
>>>>  static int vtd_machine_done_notify_one(Object *child, void *unused)
>>>>  {
>>>>      IntelIOMMUState *iommu =
>>> INTEL_IOMMU_DEVICE(x86_iommu_get_default());
>>>> @@ -4259,6 +4314,14 @@ static int
>>> vtd_machine_done_notify_one(Object *child, void *unused)
>>>>  static void vtd_machine_done_hook(Notifier *notifier, void *unused)
>>>>  {
>>>> +    IntelIOMMUState *iommu =
>>> INTEL_IOMMU_DEVICE(x86_iommu_get_default());
>>>> +
>>>> +    iommu->cap = iommu->host_cap;
>>>> +    iommu->ecap = iommu->host_ecap;
>>>> +    iommu->cap_finalized = true;
>>> I don't think you can change the defaults like this without taking care
>>> of compats (migration).
>> Will bump vtd_vmstate .version_id works?
>>
>> Thanks
>> Zhenzhong
>>
>>> Thanks
>>>
>>> Eric
>>>> +
>>>> +    vtd_setup_capability_reg(iommu);
>>>> +
>>>>      object_child_foreach_recursive(object_get_root(),
>>>>                                     vtd_machine_done_notify_one, NULL);
>>>>  }
>>>> @@ -4292,6 +4355,7 @@ static void vtd_realize(DeviceState *dev,
>Error
>>> **errp)
>>>>      QLIST_INIT(&s->vtd_as_with_notifiers);
>>>>      qemu_mutex_init(&s->iommu_lock);
>>>> +    s->cap_finalized = false;
>>>>      memory_region_init_io(&s->csrmem, OBJECT(s), &vtd_mem_ops, s,
>>>>                            "intel_iommu", DMAR_REG_SIZE);
>>>>      memory_region_add_subregion(get_system_memory(),