Patchwork [5/8] vfio: Add guest side IOMMU support

login
register
mail settings
Submitter Alexey Kardashevskiy
Date Aug. 7, 2013, 8:21 a.m.
Message ID <1375863692-12207-6-git-send-email-aik@ozlabs.ru>
Download mbox | patch
Permalink /patch/265397/
State New
Headers show

Comments

Alexey Kardashevskiy - Aug. 7, 2013, 8:21 a.m.
This patch uses the new IOMMU notifiers to allow VFIO pass through devices
to work with guest side IOMMUs, as long as the host-side VFIO iommu has
sufficient capability and granularity to match the guest side. This works
by tracking all map and unmap operations on the guest IOMMU using the
notifiers, and mirroring them into VFIO.

There are a number of FIXMEs, and the scheme involves rather more notifier
structures than I'd like, but it shuold make for a reasonable proof of
concept.

Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
---
 hw/misc/vfio.c | 149 +++++++++++++++++++++++++++++++++++++++++++++++++--------
 1 file changed, 129 insertions(+), 20 deletions(-)
Alex Williamson - Aug. 12, 2013, 10:07 p.m.
On Wed, 2013-08-07 at 18:21 +1000, Alexey Kardashevskiy wrote:
> This patch uses the new IOMMU notifiers to allow VFIO pass through devices
> to work with guest side IOMMUs, as long as the host-side VFIO iommu has
> sufficient capability and granularity to match the guest side. This works
> by tracking all map and unmap operations on the guest IOMMU using the
> notifiers, and mirroring them into VFIO.
> 
> There are a number of FIXMEs, and the scheme involves rather more notifier
> structures than I'd like, but it shuold make for a reasonable proof of
> concept.
> 
> Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
> ---
>  hw/misc/vfio.c | 149 +++++++++++++++++++++++++++++++++++++++++++++++++--------
>  1 file changed, 129 insertions(+), 20 deletions(-)
> 
> diff --git a/hw/misc/vfio.c b/hw/misc/vfio.c
> index e1ee56e..3855efe 100644
> --- a/hw/misc/vfio.c
> +++ b/hw/misc/vfio.c
> @@ -150,10 +150,18 @@ typedef struct VFIOContainer {
>          };
>          void (*release)(struct VFIOContainer *);
>      } iommu_data;
> +    QLIST_HEAD(, VFIOGuestIOMMU) guest_iommus;
>      QLIST_HEAD(, VFIOGroup) group_list;
>      QLIST_ENTRY(VFIOContainer) next;
>  } VFIOContainer;
>  
> +typedef struct VFIOGuestIOMMU {
> +    VFIOContainer *container;
> +    MemoryRegion *iommu;
> +    Notifier n;
> +    QLIST_ENTRY(VFIOGuestIOMMU) list;

Within vfio the convention is that an ENTRY is next or foo_next and a
HEAD is foo_list.  So perhaps giommu_list and next?

> +} VFIOGuestIOMMU;
> +
>  /* Cache of MSI-X setup plus extra mmap and memory region for split BAR map */
>  typedef struct VFIOMSIXInfo {
>      uint8_t table_bar;
> @@ -1917,19 +1925,70 @@ static int vfio_dma_map(VFIOContainer *container, hwaddr iova,
>  
>  static bool vfio_listener_skipped_section(MemoryRegionSection *section)
>  {
> -    return !memory_region_is_ram(section->mr);
> +    return !memory_region_is_ram(section->mr) &&
> +        !memory_region_is_iommu(section->mr);
>  }
>  
> -static void vfio_listener_region_add(MemoryListener *listener,
> -                                     MemoryRegionSection *section)
> +static void vfio_iommu_map_notify(Notifier *n, void *data)
>  {
> -    VFIOContainer *container = container_of(listener, VFIOContainer,
> -                                            iommu_data.listener);
> -    hwaddr iova, end;
> +    VFIOGuestIOMMU *giommu = container_of(n, VFIOGuestIOMMU, n);
> +    VFIOContainer *container = giommu->container;
> +    IOMMUTLBEntry *iotlb = data;
> +    MemoryRegion *mr;
> +    hwaddr xlat;
> +    hwaddr len = iotlb->addr_mask + 1;
>      void *vaddr;
>      int ret;
>  
> -    assert(!memory_region_is_iommu(section->mr));

Yay!

> +    DPRINTF("iommu map @ %"HWADDR_PRIx" - %"HWADDR_PRIx"\n",
> +            iotlb->iova, iotlb->iova + iotlb->address_mask);
> +
> +    /* The IOMMU TLB entry we have just covers translation through
> +     * this IOMMU to its immediate target.  We need to translate
> +     * it the rest of the way through to memory. */

/*
 * Multi-line comment this way please
 */

> +    mr = address_space_translate(&address_space_memory,
> +                                 iotlb->translated_addr,
> +                                 &xlat, &len, iotlb->perm & IOMMU_WO);
> +    if (!memory_region_is_ram(mr)) {
> +        DPRINTF("iommu map to non memory area %"HWADDR_PRIx"\n",
> +                xlat);
> +        return;
> +    }
> +    if (len & iotlb->addr_mask) {
> +        DPRINTF("iommu has granularity incompatible with target AS\n");
> +        return;
> +    }
> +
> +    vaddr = memory_region_get_ram_ptr(mr) + xlat;
> +
> +    if (iotlb->perm != IOMMU_NONE) {
> +        ret = vfio_dma_map(container, iotlb->iova,
> +                           iotlb->addr_mask + 1, vaddr,
> +                           !(iotlb->perm & IOMMU_WO) || mr->readonly);
> +        if (ret) {
> +            error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
> +                         "0x%"HWADDR_PRIx", %p) = %d (%m)",
> +                         container, iotlb->iova,
> +                         iotlb->addr_mask + 1, vaddr, ret);
> +        }
> +    } else {
> +        ret = vfio_dma_unmap(container, iotlb->iova, iotlb->addr_mask + 1);
> +        if (ret) {
> +            error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
> +                         "0x%"HWADDR_PRIx") = %d (%m)",
> +                         container, iotlb->iova,
> +                         iotlb->addr_mask + 1, ret);
> +        }
> +    }
> +}
> +
> +static void vfio_listener_region_add(MemoryListener *listener,
> +                                     MemoryRegionSection *section)
> +{
> +    VFIOContainer *container = container_of(listener, VFIOContainer,
> +                                            iommu_data.listener);
> +    hwaddr iova, end;
> +    int ret;
>  
>      if (vfio_listener_skipped_section(section)) {
>          DPRINTF("SKIPPING region_add %"HWADDR_PRIx" - %"PRIx64"\n",
> @@ -1952,19 +2011,51 @@ static void vfio_listener_region_add(MemoryListener *listener,
>          return;
>      }
>  
> -    vaddr = memory_region_get_ram_ptr(section->mr) +
> -            section->offset_within_region +
> -            (iova - section->offset_within_address_space);
> -
> -    DPRINTF("region_add %"HWADDR_PRIx" - %"HWADDR_PRIx" [%p]\n",
> -            iova, end - 1, vaddr);
> -
> -    memory_region_ref(section->mr);
> -    ret = vfio_dma_map(container, iova, end - iova, vaddr, section->readonly);
> -    if (ret) {
> -        error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
> -                     "0x%"HWADDR_PRIx", %p) = %d (%m)",
> -                     container, iova, end - iova, vaddr, ret);
> +    if (memory_region_is_iommu(section->mr)) {
> +        VFIOGuestIOMMU *giommu;
> +
> +        DPRINTF("region_add [iommu] %"HWADDR_PRIx" - %"HWADDR_PRIx"\n",
> +                iova, end - 1);
> +
> +        memory_region_ref(section->mr);
> +        /*
> +         * FIXME: We should do some checking to see if the
> +         * capabilities of the host VFIO IOMMU are adequate to model
> +         * the guest IOMMU
> +         *
> +         * FIXME: This assumes that the guest IOMMU is empty of
> +         * mappings at this point - we should either enforce this, or
> +         * loop through existing mappings to map them into VFIO.
> +         *
> +         * FIXME: For VFIO iommu types which have KVM acceleration to
> +         * avoid bouncing all map/unmaps through qemu this way, this
> +         * would be the right place to wire that up (tell the KVM
> +         * device emulation the VFIO iommu handles to use).
> +         */
> +        giommu = g_malloc0(sizeof(*giommu));
> +        giommu->iommu = section->mr;
> +        giommu->container = container;
> +        giommu->n.notify = vfio_iommu_map_notify;
> +        QLIST_INSERT_HEAD(&container->guest_iommus, giommu, list);
> +        memory_region_register_iommu_notifier(giommu->iommu, &giommu->n);
> +
> +    } else if (memory_region_is_ram(section->mr)) {
> +        void *vaddr;
> +
> +        vaddr = memory_region_get_ram_ptr(section->mr) +
> +                section->offset_within_region +
> +                (iova - section->offset_within_address_space);
> +
> +        DPRINTF("region_add [ram] %"HWADDR_PRIx" - %"HWADDR_PRIx" [%p]\n",
> +                iova, end - 1, vaddr);
> +
> +        memory_region_ref(section->mr);


Where are these unref'd?

> +        ret = vfio_dma_map(container, iova, end - iova, vaddr, section->readonly);
> +        if (ret) {
> +            error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
> +                         "0x%"HWADDR_PRIx", %p) = %d (%m)",
> +                         container, iova, end - iova, vaddr, ret);
> +        }
>      }
>  }
>  
> @@ -1989,6 +2080,24 @@ static void vfio_listener_region_del(MemoryListener *listener,
>          return;
>      }
>  
> +    if (memory_region_is_iommu(section->mr)) {
> +        VFIOGuestIOMMU *giommu;
> +
> +        QLIST_FOREACH(giommu, &container->guest_iommus, list) {
> +            if (giommu->iommu == section->mr) {
> +                memory_region_unregister_iommu_notifier(&giommu->n);
> +                QLIST_REMOVE(giommu, list);
> +                g_free(giommu);
> +                break;
> +            }
> +        }
> +
> +        /* FIXME: We assume the one big unmap below is adequate to
> +         * remove any individual page mappings in the IOMMU which
> +         * might have been copied into VFIO.  That may not be true for
> +         * all IOMMU types */
> +    }
> +
>      iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
>      end = (section->offset_within_address_space + int128_get64(section->size)) &
>            TARGET_PAGE_MASK;
Alexey Kardashevskiy - Aug. 15, 2013, 6:02 a.m.
On 08/13/2013 08:07 AM, Alex Williamson wrote:
>> +static void vfio_listener_region_add(MemoryListener *listener,
>> +                                     MemoryRegionSection *section)
>> +{
>> +    VFIOContainer *container = container_of(listener, VFIOContainer,
>> +                                            iommu_data.listener);
>> +    hwaddr iova, end;
>> +    int ret;
>>  
>>      if (vfio_listener_skipped_section(section)) {
>>          DPRINTF("SKIPPING region_add %"HWADDR_PRIx" - %"PRIx64"\n",
>> @@ -1952,19 +2011,51 @@ static void vfio_listener_region_add(MemoryListener *listener,
>>          return;
>>      }
>>  
>> -    vaddr = memory_region_get_ram_ptr(section->mr) +
>> -            section->offset_within_region +
>> -            (iova - section->offset_within_address_space);
>> -
>> -    DPRINTF("region_add %"HWADDR_PRIx" - %"HWADDR_PRIx" [%p]\n",
>> -            iova, end - 1, vaddr);
>> -
>> -    memory_region_ref(section->mr);
>> -    ret = vfio_dma_map(container, iova, end - iova, vaddr, section->readonly);
>> -    if (ret) {
>> -        error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
>> -                     "0x%"HWADDR_PRIx", %p) = %d (%m)",
>> -                     container, iova, end - iova, vaddr, ret);
>> +    if (memory_region_is_iommu(section->mr)) {
>> +        VFIOGuestIOMMU *giommu;
>> +
>> +        DPRINTF("region_add [iommu] %"HWADDR_PRIx" - %"HWADDR_PRIx"\n",
>> +                iova, end - 1);
>> +
>> +        memory_region_ref(section->mr);
>> +        /*
>> +         * FIXME: We should do some checking to see if the
>> +         * capabilities of the host VFIO IOMMU are adequate to model
>> +         * the guest IOMMU
>> +         *
>> +         * FIXME: This assumes that the guest IOMMU is empty of
>> +         * mappings at this point - we should either enforce this, or
>> +         * loop through existing mappings to map them into VFIO.
>> +         *
>> +         * FIXME: For VFIO iommu types which have KVM acceleration to
>> +         * avoid bouncing all map/unmaps through qemu this way, this
>> +         * would be the right place to wire that up (tell the KVM
>> +         * device emulation the VFIO iommu handles to use).
>> +         */
>> +        giommu = g_malloc0(sizeof(*giommu));
>> +        giommu->iommu = section->mr;
>> +        giommu->container = container;
>> +        giommu->n.notify = vfio_iommu_map_notify;
>> +        QLIST_INSERT_HEAD(&container->guest_iommus, giommu, list);
>> +        memory_region_register_iommu_notifier(giommu->iommu, &giommu->n);
>> +
>> +    } else if (memory_region_is_ram(section->mr)) {
>> +        void *vaddr;
>> +
>> +        vaddr = memory_region_get_ram_ptr(section->mr) +
>> +                section->offset_within_region +
>> +                (iova - section->offset_within_address_space);
>> +
>> +        DPRINTF("region_add [ram] %"HWADDR_PRIx" - %"HWADDR_PRIx" [%p]\n",
>> +                iova, end - 1, vaddr);
>> +
>> +        memory_region_ref(section->mr);
> 
> 
> Where are these unref'd?


It is already vfio_listener_region_del(), as the original ref() which this
patch just moves around.
Paolo Bonzini - Aug. 19, 2013, 1:12 p.m.
Il 15/08/2013 08:02, Alexey Kardashevskiy ha scritto:
> On 08/13/2013 08:07 AM, Alex Williamson wrote:
>>> +static void vfio_listener_region_add(MemoryListener *listener,
>>> +                                     MemoryRegionSection *section)
>>> +{
>>> +    VFIOContainer *container = container_of(listener, VFIOContainer,
>>> +                                            iommu_data.listener);
>>> +    hwaddr iova, end;
>>> +    int ret;
>>>  
>>>      if (vfio_listener_skipped_section(section)) {
>>>          DPRINTF("SKIPPING region_add %"HWADDR_PRIx" - %"PRIx64"\n",
>>> @@ -1952,19 +2011,51 @@ static void vfio_listener_region_add(MemoryListener *listener,
>>>          return;
>>>      }
>>>  
>>> -    vaddr = memory_region_get_ram_ptr(section->mr) +
>>> -            section->offset_within_region +
>>> -            (iova - section->offset_within_address_space);
>>> -
>>> -    DPRINTF("region_add %"HWADDR_PRIx" - %"HWADDR_PRIx" [%p]\n",
>>> -            iova, end - 1, vaddr);
>>> -
>>> -    memory_region_ref(section->mr);
>>> -    ret = vfio_dma_map(container, iova, end - iova, vaddr, section->readonly);
>>> -    if (ret) {
>>> -        error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
>>> -                     "0x%"HWADDR_PRIx", %p) = %d (%m)",
>>> -                     container, iova, end - iova, vaddr, ret);
>>> +    if (memory_region_is_iommu(section->mr)) {
>>> +        VFIOGuestIOMMU *giommu;
>>> +
>>> +        DPRINTF("region_add [iommu] %"HWADDR_PRIx" - %"HWADDR_PRIx"\n",
>>> +                iova, end - 1);
>>> +
>>> +        memory_region_ref(section->mr);
>>> +        /*
>>> +         * FIXME: We should do some checking to see if the
>>> +         * capabilities of the host VFIO IOMMU are adequate to model
>>> +         * the guest IOMMU
>>> +         *
>>> +         * FIXME: This assumes that the guest IOMMU is empty of
>>> +         * mappings at this point - we should either enforce this, or
>>> +         * loop through existing mappings to map them into VFIO.
>>> +         *
>>> +         * FIXME: For VFIO iommu types which have KVM acceleration to
>>> +         * avoid bouncing all map/unmaps through qemu this way, this
>>> +         * would be the right place to wire that up (tell the KVM
>>> +         * device emulation the VFIO iommu handles to use).
>>> +         */
>>> +        giommu = g_malloc0(sizeof(*giommu));
>>> +        giommu->iommu = section->mr;
>>> +        giommu->container = container;
>>> +        giommu->n.notify = vfio_iommu_map_notify;
>>> +        QLIST_INSERT_HEAD(&container->guest_iommus, giommu, list);
>>> +        memory_region_register_iommu_notifier(giommu->iommu, &giommu->n);
>>> +
>>> +    } else if (memory_region_is_ram(section->mr)) {

Please change this to an "else", and leave the refs outside the if, just
after checking for vfio_listener_skipped_section.  This way it's clearer
that vfio_listener_region_add matches vfio_listener_region_del.

Paolo
Alexey Kardashevskiy - Aug. 21, 2013, 5:31 a.m.
On 08/19/2013 11:12 PM, Paolo Bonzini wrote:
> Il 15/08/2013 08:02, Alexey Kardashevskiy ha scritto:
>> On 08/13/2013 08:07 AM, Alex Williamson wrote:
>>>> +static void vfio_listener_region_add(MemoryListener *listener,
>>>> +                                     MemoryRegionSection *section)
>>>> +{
>>>> +    VFIOContainer *container = container_of(listener, VFIOContainer,
>>>> +                                            iommu_data.listener);
>>>> +    hwaddr iova, end;
>>>> +    int ret;
>>>>  
>>>>      if (vfio_listener_skipped_section(section)) {
>>>>          DPRINTF("SKIPPING region_add %"HWADDR_PRIx" - %"PRIx64"\n",
>>>> @@ -1952,19 +2011,51 @@ static void vfio_listener_region_add(MemoryListener *listener,
>>>>          return;
>>>>      }
>>>>  
>>>> -    vaddr = memory_region_get_ram_ptr(section->mr) +
>>>> -            section->offset_within_region +
>>>> -            (iova - section->offset_within_address_space);
>>>> -
>>>> -    DPRINTF("region_add %"HWADDR_PRIx" - %"HWADDR_PRIx" [%p]\n",
>>>> -            iova, end - 1, vaddr);
>>>> -
>>>> -    memory_region_ref(section->mr);
>>>> -    ret = vfio_dma_map(container, iova, end - iova, vaddr, section->readonly);
>>>> -    if (ret) {
>>>> -        error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
>>>> -                     "0x%"HWADDR_PRIx", %p) = %d (%m)",
>>>> -                     container, iova, end - iova, vaddr, ret);
>>>> +    if (memory_region_is_iommu(section->mr)) {
>>>> +        VFIOGuestIOMMU *giommu;
>>>> +
>>>> +        DPRINTF("region_add [iommu] %"HWADDR_PRIx" - %"HWADDR_PRIx"\n",
>>>> +                iova, end - 1);
>>>> +
>>>> +        memory_region_ref(section->mr);
>>>> +        /*
>>>> +         * FIXME: We should do some checking to see if the
>>>> +         * capabilities of the host VFIO IOMMU are adequate to model
>>>> +         * the guest IOMMU
>>>> +         *
>>>> +         * FIXME: This assumes that the guest IOMMU is empty of
>>>> +         * mappings at this point - we should either enforce this, or
>>>> +         * loop through existing mappings to map them into VFIO.
>>>> +         *
>>>> +         * FIXME: For VFIO iommu types which have KVM acceleration to
>>>> +         * avoid bouncing all map/unmaps through qemu this way, this
>>>> +         * would be the right place to wire that up (tell the KVM
>>>> +         * device emulation the VFIO iommu handles to use).
>>>> +         */
>>>> +        giommu = g_malloc0(sizeof(*giommu));
>>>> +        giommu->iommu = section->mr;
>>>> +        giommu->container = container;
>>>> +        giommu->n.notify = vfio_iommu_map_notify;
>>>> +        QLIST_INSERT_HEAD(&container->guest_iommus, giommu, list);
>>>> +        memory_region_register_iommu_notifier(giommu->iommu, &giommu->n);
>>>> +
>>>> +    } else if (memory_region_is_ram(section->mr)) {
> 
> Please change this to an "else", and leave the refs outside the if, just
> after checking for vfio_listener_skipped_section.

Why right after vfio_listener_skipped_section? The whole block looks now as:

===
static void vfio_listener_region_add(MemoryListener *listener,
                                     MemoryRegionSection *section)
{
    VFIOContainer *container = container_of(listener, VFIOContainer,
                                            iommu_data.listener);
    hwaddr iova, end;
    void *vaddr;
    int ret;

    if (vfio_listener_skipped_section(section)) {
        DPRINTF("SKIPPING region_add %"HWADDR_PRIx" - %"PRIx64"\n",
                section->offset_within_address_space,
                section->offset_within_address_space + section->size - 1);
        return;
    }

    if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) !=
                 (section->offset_within_region & ~TARGET_PAGE_MASK))) {
        error_report("%s received unaligned region", __func__);
        return;
    }

    iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
    end = (section->offset_within_address_space +
int128_get64(section->size)) &
          TARGET_PAGE_MASK;

    if (iova >= end) {
        return;
    }

    memory_region_ref(section->mr);
===

You want me to move memory_region_ref() earlier to add a reference even if
two last checks fail?


>  This way it's clearer
> that vfio_listener_region_add matches vfio_listener_region_del.

Yes, true, reworked. Thanks for comments!
Paolo Bonzini - Aug. 21, 2013, 8:55 a.m.
Il 21/08/2013 07:31, Alexey Kardashevskiy ha scritto:
>     memory_region_ref(section->mr);
> ===
> 
> You want me to move memory_region_ref() earlier to add a reference even if
> two last checks fail?

No, of course not---I was looking at _del source code not _add.

Paolo

Patch

diff --git a/hw/misc/vfio.c b/hw/misc/vfio.c
index e1ee56e..3855efe 100644
--- a/hw/misc/vfio.c
+++ b/hw/misc/vfio.c
@@ -150,10 +150,18 @@  typedef struct VFIOContainer {
         };
         void (*release)(struct VFIOContainer *);
     } iommu_data;
+    QLIST_HEAD(, VFIOGuestIOMMU) guest_iommus;
     QLIST_HEAD(, VFIOGroup) group_list;
     QLIST_ENTRY(VFIOContainer) next;
 } VFIOContainer;
 
+typedef struct VFIOGuestIOMMU {
+    VFIOContainer *container;
+    MemoryRegion *iommu;
+    Notifier n;
+    QLIST_ENTRY(VFIOGuestIOMMU) list;
+} VFIOGuestIOMMU;
+
 /* Cache of MSI-X setup plus extra mmap and memory region for split BAR map */
 typedef struct VFIOMSIXInfo {
     uint8_t table_bar;
@@ -1917,19 +1925,70 @@  static int vfio_dma_map(VFIOContainer *container, hwaddr iova,
 
 static bool vfio_listener_skipped_section(MemoryRegionSection *section)
 {
-    return !memory_region_is_ram(section->mr);
+    return !memory_region_is_ram(section->mr) &&
+        !memory_region_is_iommu(section->mr);
 }
 
-static void vfio_listener_region_add(MemoryListener *listener,
-                                     MemoryRegionSection *section)
+static void vfio_iommu_map_notify(Notifier *n, void *data)
 {
-    VFIOContainer *container = container_of(listener, VFIOContainer,
-                                            iommu_data.listener);
-    hwaddr iova, end;
+    VFIOGuestIOMMU *giommu = container_of(n, VFIOGuestIOMMU, n);
+    VFIOContainer *container = giommu->container;
+    IOMMUTLBEntry *iotlb = data;
+    MemoryRegion *mr;
+    hwaddr xlat;
+    hwaddr len = iotlb->addr_mask + 1;
     void *vaddr;
     int ret;
 
-    assert(!memory_region_is_iommu(section->mr));
+    DPRINTF("iommu map @ %"HWADDR_PRIx" - %"HWADDR_PRIx"\n",
+            iotlb->iova, iotlb->iova + iotlb->address_mask);
+
+    /* The IOMMU TLB entry we have just covers translation through
+     * this IOMMU to its immediate target.  We need to translate
+     * it the rest of the way through to memory. */
+    mr = address_space_translate(&address_space_memory,
+                                 iotlb->translated_addr,
+                                 &xlat, &len, iotlb->perm & IOMMU_WO);
+    if (!memory_region_is_ram(mr)) {
+        DPRINTF("iommu map to non memory area %"HWADDR_PRIx"\n",
+                xlat);
+        return;
+    }
+    if (len & iotlb->addr_mask) {
+        DPRINTF("iommu has granularity incompatible with target AS\n");
+        return;
+    }
+
+    vaddr = memory_region_get_ram_ptr(mr) + xlat;
+
+    if (iotlb->perm != IOMMU_NONE) {
+        ret = vfio_dma_map(container, iotlb->iova,
+                           iotlb->addr_mask + 1, vaddr,
+                           !(iotlb->perm & IOMMU_WO) || mr->readonly);
+        if (ret) {
+            error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
+                         "0x%"HWADDR_PRIx", %p) = %d (%m)",
+                         container, iotlb->iova,
+                         iotlb->addr_mask + 1, vaddr, ret);
+        }
+    } else {
+        ret = vfio_dma_unmap(container, iotlb->iova, iotlb->addr_mask + 1);
+        if (ret) {
+            error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
+                         "0x%"HWADDR_PRIx") = %d (%m)",
+                         container, iotlb->iova,
+                         iotlb->addr_mask + 1, ret);
+        }
+    }
+}
+
+static void vfio_listener_region_add(MemoryListener *listener,
+                                     MemoryRegionSection *section)
+{
+    VFIOContainer *container = container_of(listener, VFIOContainer,
+                                            iommu_data.listener);
+    hwaddr iova, end;
+    int ret;
 
     if (vfio_listener_skipped_section(section)) {
         DPRINTF("SKIPPING region_add %"HWADDR_PRIx" - %"PRIx64"\n",
@@ -1952,19 +2011,51 @@  static void vfio_listener_region_add(MemoryListener *listener,
         return;
     }
 
-    vaddr = memory_region_get_ram_ptr(section->mr) +
-            section->offset_within_region +
-            (iova - section->offset_within_address_space);
-
-    DPRINTF("region_add %"HWADDR_PRIx" - %"HWADDR_PRIx" [%p]\n",
-            iova, end - 1, vaddr);
-
-    memory_region_ref(section->mr);
-    ret = vfio_dma_map(container, iova, end - iova, vaddr, section->readonly);
-    if (ret) {
-        error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
-                     "0x%"HWADDR_PRIx", %p) = %d (%m)",
-                     container, iova, end - iova, vaddr, ret);
+    if (memory_region_is_iommu(section->mr)) {
+        VFIOGuestIOMMU *giommu;
+
+        DPRINTF("region_add [iommu] %"HWADDR_PRIx" - %"HWADDR_PRIx"\n",
+                iova, end - 1);
+
+        memory_region_ref(section->mr);
+        /*
+         * FIXME: We should do some checking to see if the
+         * capabilities of the host VFIO IOMMU are adequate to model
+         * the guest IOMMU
+         *
+         * FIXME: This assumes that the guest IOMMU is empty of
+         * mappings at this point - we should either enforce this, or
+         * loop through existing mappings to map them into VFIO.
+         *
+         * FIXME: For VFIO iommu types which have KVM acceleration to
+         * avoid bouncing all map/unmaps through qemu this way, this
+         * would be the right place to wire that up (tell the KVM
+         * device emulation the VFIO iommu handles to use).
+         */
+        giommu = g_malloc0(sizeof(*giommu));
+        giommu->iommu = section->mr;
+        giommu->container = container;
+        giommu->n.notify = vfio_iommu_map_notify;
+        QLIST_INSERT_HEAD(&container->guest_iommus, giommu, list);
+        memory_region_register_iommu_notifier(giommu->iommu, &giommu->n);
+
+    } else if (memory_region_is_ram(section->mr)) {
+        void *vaddr;
+
+        vaddr = memory_region_get_ram_ptr(section->mr) +
+                section->offset_within_region +
+                (iova - section->offset_within_address_space);
+
+        DPRINTF("region_add [ram] %"HWADDR_PRIx" - %"HWADDR_PRIx" [%p]\n",
+                iova, end - 1, vaddr);
+
+        memory_region_ref(section->mr);
+        ret = vfio_dma_map(container, iova, end - iova, vaddr, section->readonly);
+        if (ret) {
+            error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
+                         "0x%"HWADDR_PRIx", %p) = %d (%m)",
+                         container, iova, end - iova, vaddr, ret);
+        }
     }
 }
 
@@ -1989,6 +2080,24 @@  static void vfio_listener_region_del(MemoryListener *listener,
         return;
     }
 
+    if (memory_region_is_iommu(section->mr)) {
+        VFIOGuestIOMMU *giommu;
+
+        QLIST_FOREACH(giommu, &container->guest_iommus, list) {
+            if (giommu->iommu == section->mr) {
+                memory_region_unregister_iommu_notifier(&giommu->n);
+                QLIST_REMOVE(giommu, list);
+                g_free(giommu);
+                break;
+            }
+        }
+
+        /* FIXME: We assume the one big unmap below is adequate to
+         * remove any individual page mappings in the IOMMU which
+         * might have been copied into VFIO.  That may not be true for
+         * all IOMMU types */
+    }
+
     iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
     end = (section->offset_within_address_space + int128_get64(section->size)) &
           TARGET_PAGE_MASK;