The driver can create a bypass domain by passing the
VIRTIO_IOMMU_ATTACH_F_BYPASS flag on the ATTACH request. Bypass domains
perform slightly better than domains with identity mappings since they
skip translation.

Signed-off-by: Jean-Philippe Brucker <jean-philippe@linaro.org>
---
 hw/virtio/virtio-iommu.c | 32 ++++++++++++++++++++++++++++++--
 1 file changed, 30 insertions(+), 2 deletions(-)

diff --git a/hw/virtio/virtio-iommu.c b/hw/virtio/virtio-iommu.c
index ec02029bb6..a112428c65 100644
--- a/hw/virtio/virtio-iommu.c
+++ b/hw/virtio/virtio-iommu.c
@@ -43,6 +43,7 @@
 
 typedef struct VirtIOIOMMUDomain {
     uint32_t id;
+    bool bypass;
     GTree *mappings;
     QLIST_HEAD(, VirtIOIOMMUEndpoint) endpoint_list;
 } VirtIOIOMMUDomain;
@@ -258,12 +259,16 @@ static void virtio_iommu_put_endpoint(gpointer data)
 }
 
 static VirtIOIOMMUDomain *virtio_iommu_get_domain(VirtIOIOMMU *s,
-                                                  uint32_t domain_id)
+                                                  uint32_t domain_id,
+                                                  bool bypass)
 {
     VirtIOIOMMUDomain *domain;
 
     domain = g_tree_lookup(s->domains, GUINT_TO_POINTER(domain_id));
     if (domain) {
+        if (domain->bypass != bypass) {
+            return NULL;
+        }
         return domain;
     }
     domain = g_malloc0(sizeof(*domain));
@@ -271,6 +276,7 @@ static VirtIOIOMMUDomain *virtio_iommu_get_domain(VirtIOIOMMU *s,
     domain->mappings = g_tree_new_full((GCompareDataFunc)interval_cmp,
                                    NULL, (GDestroyNotify)g_free,
                                    (GDestroyNotify)g_free);
+    domain->bypass = bypass;
     g_tree_insert(s->domains, GUINT_TO_POINTER(domain_id), domain);
     QLIST_INIT(&domain->endpoint_list);
     trace_virtio_iommu_get_domain(domain_id);
@@ -334,11 +340,16 @@ static int virtio_iommu_attach(VirtIOIOMMU *s,
 {
     uint32_t domain_id = le32_to_cpu(req->domain);
     uint32_t ep_id = le32_to_cpu(req->endpoint);
+    uint32_t flags = le32_to_cpu(req->flags);
     VirtIOIOMMUDomain *domain;
     VirtIOIOMMUEndpoint *ep;
 
     trace_virtio_iommu_attach(domain_id, ep_id);
 
+    if (flags & ~VIRTIO_IOMMU_ATTACH_F_BYPASS) {
+        return VIRTIO_IOMMU_S_INVAL;
+    }
+
     ep = virtio_iommu_get_endpoint(s, ep_id);
     if (!ep) {
         return VIRTIO_IOMMU_S_NOENT;
@@ -356,7 +367,12 @@ static int virtio_iommu_attach(VirtIOIOMMU *s,
         }
     }
 
-    domain = virtio_iommu_get_domain(s, domain_id);
+    domain = virtio_iommu_get_domain(s, domain_id,
+                                     flags & VIRTIO_IOMMU_ATTACH_F_BYPASS);
+    if (!domain) {
+        /* Incompatible bypass flag */
+        return VIRTIO_IOMMU_S_INVAL;
+    }
     QLIST_INSERT_HEAD(&domain->endpoint_list, ep, next);
 
     ep->domain = domain;
@@ -419,6 +435,10 @@ static int virtio_iommu_map(VirtIOIOMMU *s,
         return VIRTIO_IOMMU_S_NOENT;
     }
 
+    if (domain->bypass) {
+        return VIRTIO_IOMMU_S_INVAL;
+    }
+
     interval = g_malloc0(sizeof(*interval));
 
     interval->low = virt_start;
@@ -464,6 +484,11 @@ static int virtio_iommu_unmap(VirtIOIOMMU *s,
     if (!domain) {
         return VIRTIO_IOMMU_S_NOENT;
     }
+
+    if (domain->bypass) {
+        return VIRTIO_IOMMU_S_INVAL;
+    }
+
     interval.low = virt_start;
     interval.high = virt_end;
 
@@ -780,6 +805,9 @@ static IOMMUTLBEntry virtio_iommu_translate(IOMMUMemoryRegion *mr, hwaddr addr,
             entry.perm = flag;
         }
         goto unlock;
+    } else if (ep->domain->bypass) {
+        entry.perm = flag;
+        goto unlock;
     }
 
     found = g_tree_lookup_extended(ep->domain->mappings, (gpointer)(&interval),
-- 
2.34.1

Hi Jean,

On 1/27/22 3:29 PM, Jean-Philippe Brucker wrote:
> The driver can create a bypass domain by passing the
> VIRTIO_IOMMU_ATTACH_F_BYPASS flag on the ATTACH request. Bypass domains
> perform slightly better than domains with identity mappings since they
> skip translation.
>
> Signed-off-by: Jean-Philippe Brucker <jean-philippe@linaro.org>
> ---
>  hw/virtio/virtio-iommu.c | 32 ++++++++++++++++++++++++++++++--
>  1 file changed, 30 insertions(+), 2 deletions(-)
>
> diff --git a/hw/virtio/virtio-iommu.c b/hw/virtio/virtio-iommu.c
> index ec02029bb6..a112428c65 100644
> --- a/hw/virtio/virtio-iommu.c
> +++ b/hw/virtio/virtio-iommu.c
> @@ -43,6 +43,7 @@
>  
>  typedef struct VirtIOIOMMUDomain {
>      uint32_t id;
> +    bool bypass;
I am afraid this will break the migration if you don't change
vmstate_domain.

See static const VMStateDescription vmstate_domain.
Also you need to migrate the new bypass field.

Logically we should handle this with a vmstate subsection I think to
handle migration of older devices. However I doubt the device has been
used in production environment supporting migration so my guess is we
may skip that burden and just add the missing field. Adding Juan, Dave &
Peter for advices.

Thanks

Eric

>      GTree *mappings;
>      QLIST_HEAD(, VirtIOIOMMUEndpoint) endpoint_list;
>  } VirtIOIOMMUDomain;
> @@ -258,12 +259,16 @@ static void virtio_iommu_put_endpoint(gpointer data)
>  }
>  
>  static VirtIOIOMMUDomain *virtio_iommu_get_domain(VirtIOIOMMU *s,
> -                                                  uint32_t domain_id)
> +                                                  uint32_t domain_id,
> +                                                  bool bypass)
>  {
>      VirtIOIOMMUDomain *domain;
>  
>      domain = g_tree_lookup(s->domains, GUINT_TO_POINTER(domain_id));
>      if (domain) {
> +        if (domain->bypass != bypass) {
> +            return NULL;
> +        }
>          return domain;
>      }
>      domain = g_malloc0(sizeof(*domain));
> @@ -271,6 +276,7 @@ static VirtIOIOMMUDomain *virtio_iommu_get_domain(VirtIOIOMMU *s,
>      domain->mappings = g_tree_new_full((GCompareDataFunc)interval_cmp,
>                                     NULL, (GDestroyNotify)g_free,
>                                     (GDestroyNotify)g_free);
> +    domain->bypass = bypass;
>      g_tree_insert(s->domains, GUINT_TO_POINTER(domain_id), domain);
>      QLIST_INIT(&domain->endpoint_list);
>      trace_virtio_iommu_get_domain(domain_id);
> @@ -334,11 +340,16 @@ static int virtio_iommu_attach(VirtIOIOMMU *s,
>  {
>      uint32_t domain_id = le32_to_cpu(req->domain);
>      uint32_t ep_id = le32_to_cpu(req->endpoint);
> +    uint32_t flags = le32_to_cpu(req->flags);
>      VirtIOIOMMUDomain *domain;
>      VirtIOIOMMUEndpoint *ep;
>  
>      trace_virtio_iommu_attach(domain_id, ep_id);
>  
> +    if (flags & ~VIRTIO_IOMMU_ATTACH_F_BYPASS) {
> +        return VIRTIO_IOMMU_S_INVAL;
> +    }
> +
>      ep = virtio_iommu_get_endpoint(s, ep_id);
>      if (!ep) {
>          return VIRTIO_IOMMU_S_NOENT;
> @@ -356,7 +367,12 @@ static int virtio_iommu_attach(VirtIOIOMMU *s,
>          }
>      }
>  
> -    domain = virtio_iommu_get_domain(s, domain_id);
> +    domain = virtio_iommu_get_domain(s, domain_id,
> +                                     flags & VIRTIO_IOMMU_ATTACH_F_BYPASS);
> +    if (!domain) {
> +        /* Incompatible bypass flag */
> +        return VIRTIO_IOMMU_S_INVAL;
> +    }
>      QLIST_INSERT_HEAD(&domain->endpoint_list, ep, next);
>  
>      ep->domain = domain;
> @@ -419,6 +435,10 @@ static int virtio_iommu_map(VirtIOIOMMU *s,
>          return VIRTIO_IOMMU_S_NOENT;
>      }
>  
> +    if (domain->bypass) {
> +        return VIRTIO_IOMMU_S_INVAL;
> +    }
> +
>      interval = g_malloc0(sizeof(*interval));
>  
>      interval->low = virt_start;
> @@ -464,6 +484,11 @@ static int virtio_iommu_unmap(VirtIOIOMMU *s,
>      if (!domain) {
>          return VIRTIO_IOMMU_S_NOENT;
>      }
> +
> +    if (domain->bypass) {
> +        return VIRTIO_IOMMU_S_INVAL;
> +    }
> +
>      interval.low = virt_start;
>      interval.high = virt_end;
>  
> @@ -780,6 +805,9 @@ static IOMMUTLBEntry virtio_iommu_translate(IOMMUMemoryRegion *mr, hwaddr addr,
>              entry.perm = flag;
>          }
>          goto unlock;
> +    } else if (ep->domain->bypass) {
> +        entry.perm = flag;
> +        goto unlock;
>      }
>  
>      found = g_tree_lookup_extended(ep->domain->mappings, (gpointer)(&interval),

* Eric Auger (eric.auger@redhat.com) wrote:
> Hi Jean,
> 
> On 1/27/22 3:29 PM, Jean-Philippe Brucker wrote:
> > The driver can create a bypass domain by passing the
> > VIRTIO_IOMMU_ATTACH_F_BYPASS flag on the ATTACH request. Bypass domains
> > perform slightly better than domains with identity mappings since they
> > skip translation.
> >
> > Signed-off-by: Jean-Philippe Brucker <jean-philippe@linaro.org>
> > ---
> >  hw/virtio/virtio-iommu.c | 32 ++++++++++++++++++++++++++++++--
> >  1 file changed, 30 insertions(+), 2 deletions(-)
> >
> > diff --git a/hw/virtio/virtio-iommu.c b/hw/virtio/virtio-iommu.c
> > index ec02029bb6..a112428c65 100644
> > --- a/hw/virtio/virtio-iommu.c
> > +++ b/hw/virtio/virtio-iommu.c
> > @@ -43,6 +43,7 @@
> >  
> >  typedef struct VirtIOIOMMUDomain {
> >      uint32_t id;
> > +    bool bypass;
> I am afraid this will break the migration if you don't change
> vmstate_domain.
> 
> See static const VMStateDescription vmstate_domain.
> Also you need to migrate the new bypass field.
> 
> Logically we should handle this with a vmstate subsection I think to
> handle migration of older devices. However I doubt the device has been
> used in production environment supporting migration so my guess is we
> may skip that burden and just add the missing field. Adding Juan, Dave &
> Peter for advices.

I'm not sure about users of this; if no one has used it then yeh; you
could bump up the version_id to make it a bit clearer.

Dave

> Thanks
> 
> Eric
> 
> >      GTree *mappings;
> >      QLIST_HEAD(, VirtIOIOMMUEndpoint) endpoint_list;
> >  } VirtIOIOMMUDomain;
> > @@ -258,12 +259,16 @@ static void virtio_iommu_put_endpoint(gpointer data)
> >  }
> >  
> >  static VirtIOIOMMUDomain *virtio_iommu_get_domain(VirtIOIOMMU *s,
> > -                                                  uint32_t domain_id)
> > +                                                  uint32_t domain_id,
> > +                                                  bool bypass)
> >  {
> >      VirtIOIOMMUDomain *domain;
> >  
> >      domain = g_tree_lookup(s->domains, GUINT_TO_POINTER(domain_id));
> >      if (domain) {
> > +        if (domain->bypass != bypass) {
> > +            return NULL;
> > +        }
> >          return domain;
> >      }
> >      domain = g_malloc0(sizeof(*domain));
> > @@ -271,6 +276,7 @@ static VirtIOIOMMUDomain *virtio_iommu_get_domain(VirtIOIOMMU *s,
> >      domain->mappings = g_tree_new_full((GCompareDataFunc)interval_cmp,
> >                                     NULL, (GDestroyNotify)g_free,
> >                                     (GDestroyNotify)g_free);
> > +    domain->bypass = bypass;
> >      g_tree_insert(s->domains, GUINT_TO_POINTER(domain_id), domain);
> >      QLIST_INIT(&domain->endpoint_list);
> >      trace_virtio_iommu_get_domain(domain_id);
> > @@ -334,11 +340,16 @@ static int virtio_iommu_attach(VirtIOIOMMU *s,
> >  {
> >      uint32_t domain_id = le32_to_cpu(req->domain);
> >      uint32_t ep_id = le32_to_cpu(req->endpoint);
> > +    uint32_t flags = le32_to_cpu(req->flags);
> >      VirtIOIOMMUDomain *domain;
> >      VirtIOIOMMUEndpoint *ep;
> >  
> >      trace_virtio_iommu_attach(domain_id, ep_id);
> >  
> > +    if (flags & ~VIRTIO_IOMMU_ATTACH_F_BYPASS) {
> > +        return VIRTIO_IOMMU_S_INVAL;
> > +    }
> > +
> >      ep = virtio_iommu_get_endpoint(s, ep_id);
> >      if (!ep) {
> >          return VIRTIO_IOMMU_S_NOENT;
> > @@ -356,7 +367,12 @@ static int virtio_iommu_attach(VirtIOIOMMU *s,
> >          }
> >      }
> >  
> > -    domain = virtio_iommu_get_domain(s, domain_id);
> > +    domain = virtio_iommu_get_domain(s, domain_id,
> > +                                     flags & VIRTIO_IOMMU_ATTACH_F_BYPASS);
> > +    if (!domain) {
> > +        /* Incompatible bypass flag */
> > +        return VIRTIO_IOMMU_S_INVAL;
> > +    }
> >      QLIST_INSERT_HEAD(&domain->endpoint_list, ep, next);
> >  
> >      ep->domain = domain;
> > @@ -419,6 +435,10 @@ static int virtio_iommu_map(VirtIOIOMMU *s,
> >          return VIRTIO_IOMMU_S_NOENT;
> >      }
> >  
> > +    if (domain->bypass) {
> > +        return VIRTIO_IOMMU_S_INVAL;
> > +    }
> > +
> >      interval = g_malloc0(sizeof(*interval));
> >  
> >      interval->low = virt_start;
> > @@ -464,6 +484,11 @@ static int virtio_iommu_unmap(VirtIOIOMMU *s,
> >      if (!domain) {
> >          return VIRTIO_IOMMU_S_NOENT;
> >      }
> > +
> > +    if (domain->bypass) {
> > +        return VIRTIO_IOMMU_S_INVAL;
> > +    }
> > +
> >      interval.low = virt_start;
> >      interval.high = virt_end;
> >  
> > @@ -780,6 +805,9 @@ static IOMMUTLBEntry virtio_iommu_translate(IOMMUMemoryRegion *mr, hwaddr addr,
> >              entry.perm = flag;
> >          }
> >          goto unlock;
> > +    } else if (ep->domain->bypass) {
> > +        entry.perm = flag;
> > +        goto unlock;
> >      }
> >  
> >      found = g_tree_lookup_extended(ep->domain->mappings, (gpointer)(&interval),
> 
-- 
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK

Hi Dave,

On 1/31/22 2:07 PM, Dr. David Alan Gilbert wrote:
> * Eric Auger (eric.auger@redhat.com) wrote:
>> Hi Jean,
>>
>> On 1/27/22 3:29 PM, Jean-Philippe Brucker wrote:
>>> The driver can create a bypass domain by passing the
>>> VIRTIO_IOMMU_ATTACH_F_BYPASS flag on the ATTACH request. Bypass domains
>>> perform slightly better than domains with identity mappings since they
>>> skip translation.
>>>
>>> Signed-off-by: Jean-Philippe Brucker <jean-philippe@linaro.org>
>>> ---
>>>  hw/virtio/virtio-iommu.c | 32 ++++++++++++++++++++++++++++++--
>>>  1 file changed, 30 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/hw/virtio/virtio-iommu.c b/hw/virtio/virtio-iommu.c
>>> index ec02029bb6..a112428c65 100644
>>> --- a/hw/virtio/virtio-iommu.c
>>> +++ b/hw/virtio/virtio-iommu.c
>>> @@ -43,6 +43,7 @@
>>>  
>>>  typedef struct VirtIOIOMMUDomain {
>>>      uint32_t id;
>>> +    bool bypass;
>> I am afraid this will break the migration if you don't change
>> vmstate_domain.
>>
>> See static const VMStateDescription vmstate_domain.
>> Also you need to migrate the new bypass field.
>>
>> Logically we should handle this with a vmstate subsection I think to
>> handle migration of older devices. However I doubt the device has been
>> used in production environment supporting migration so my guess is we
>> may skip that burden and just add the missing field. Adding Juan, Dave &
>> Peter for advices.
> I'm not sure about users of this; if no one has used it then yeh; you
> could bump up the version_id to make it a bit clearer.

Thank you for your input. Yes to me it sounds OK to only bump the
version_id while adding the new field.

Eric
>
> Dave
>
>> Thanks
>>
>> Eric
>>
>>>      GTree *mappings;
>>>      QLIST_HEAD(, VirtIOIOMMUEndpoint) endpoint_list;
>>>  } VirtIOIOMMUDomain;
>>> @@ -258,12 +259,16 @@ static void virtio_iommu_put_endpoint(gpointer data)
>>>  }
>>>  
>>>  static VirtIOIOMMUDomain *virtio_iommu_get_domain(VirtIOIOMMU *s,
>>> -                                                  uint32_t domain_id)
>>> +                                                  uint32_t domain_id,
>>> +                                                  bool bypass)
>>>  {
>>>      VirtIOIOMMUDomain *domain;
>>>  
>>>      domain = g_tree_lookup(s->domains, GUINT_TO_POINTER(domain_id));
>>>      if (domain) {
>>> +        if (domain->bypass != bypass) {
>>> +            return NULL;
>>> +        }
>>>          return domain;
>>>      }
>>>      domain = g_malloc0(sizeof(*domain));
>>> @@ -271,6 +276,7 @@ static VirtIOIOMMUDomain *virtio_iommu_get_domain(VirtIOIOMMU *s,
>>>      domain->mappings = g_tree_new_full((GCompareDataFunc)interval_cmp,
>>>                                     NULL, (GDestroyNotify)g_free,
>>>                                     (GDestroyNotify)g_free);
>>> +    domain->bypass = bypass;
>>>      g_tree_insert(s->domains, GUINT_TO_POINTER(domain_id), domain);
>>>      QLIST_INIT(&domain->endpoint_list);
>>>      trace_virtio_iommu_get_domain(domain_id);
>>> @@ -334,11 +340,16 @@ static int virtio_iommu_attach(VirtIOIOMMU *s,
>>>  {
>>>      uint32_t domain_id = le32_to_cpu(req->domain);
>>>      uint32_t ep_id = le32_to_cpu(req->endpoint);
>>> +    uint32_t flags = le32_to_cpu(req->flags);
>>>      VirtIOIOMMUDomain *domain;
>>>      VirtIOIOMMUEndpoint *ep;
>>>  
>>>      trace_virtio_iommu_attach(domain_id, ep_id);
>>>  
>>> +    if (flags & ~VIRTIO_IOMMU_ATTACH_F_BYPASS) {
>>> +        return VIRTIO_IOMMU_S_INVAL;
>>> +    }
>>> +
>>>      ep = virtio_iommu_get_endpoint(s, ep_id);
>>>      if (!ep) {
>>>          return VIRTIO_IOMMU_S_NOENT;
>>> @@ -356,7 +367,12 @@ static int virtio_iommu_attach(VirtIOIOMMU *s,
>>>          }
>>>      }
>>>  
>>> -    domain = virtio_iommu_get_domain(s, domain_id);
>>> +    domain = virtio_iommu_get_domain(s, domain_id,
>>> +                                     flags & VIRTIO_IOMMU_ATTACH_F_BYPASS);
>>> +    if (!domain) {
>>> +        /* Incompatible bypass flag */
>>> +        return VIRTIO_IOMMU_S_INVAL;
>>> +    }
>>>      QLIST_INSERT_HEAD(&domain->endpoint_list, ep, next);
>>>  
>>>      ep->domain = domain;
>>> @@ -419,6 +435,10 @@ static int virtio_iommu_map(VirtIOIOMMU *s,
>>>          return VIRTIO_IOMMU_S_NOENT;
>>>      }
>>>  
>>> +    if (domain->bypass) {
>>> +        return VIRTIO_IOMMU_S_INVAL;
>>> +    }
>>> +
>>>      interval = g_malloc0(sizeof(*interval));
>>>  
>>>      interval->low = virt_start;
>>> @@ -464,6 +484,11 @@ static int virtio_iommu_unmap(VirtIOIOMMU *s,
>>>      if (!domain) {
>>>          return VIRTIO_IOMMU_S_NOENT;
>>>      }
>>> +
>>> +    if (domain->bypass) {
>>> +        return VIRTIO_IOMMU_S_INVAL;
>>> +    }
>>> +
>>>      interval.low = virt_start;
>>>      interval.high = virt_end;
>>>  
>>> @@ -780,6 +805,9 @@ static IOMMUTLBEntry virtio_iommu_translate(IOMMUMemoryRegion *mr, hwaddr addr,
>>>              entry.perm = flag;
>>>          }
>>>          goto unlock;
>>> +    } else if (ep->domain->bypass) {
>>> +        entry.perm = flag;
>>> +        goto unlock;
>>>      }
>>>  
>>>      found = g_tree_lookup_extended(ep->domain->mappings, (gpointer)(&interval),

On Wed, Feb 02, 2022 at 02:21:37PM +0100, Eric Auger wrote:
> >>> diff --git a/hw/virtio/virtio-iommu.c b/hw/virtio/virtio-iommu.c
> >>> index ec02029bb6..a112428c65 100644
> >>> --- a/hw/virtio/virtio-iommu.c
> >>> +++ b/hw/virtio/virtio-iommu.c
> >>> @@ -43,6 +43,7 @@
> >>>  
> >>>  typedef struct VirtIOIOMMUDomain {
> >>>      uint32_t id;
> >>> +    bool bypass;
> >> I am afraid this will break the migration if you don't change
> >> vmstate_domain.
> >>
> >> See static const VMStateDescription vmstate_domain.
> >> Also you need to migrate the new bypass field.
> >>
> >> Logically we should handle this with a vmstate subsection I think to
> >> handle migration of older devices. However I doubt the device has been
> >> used in production environment supporting migration so my guess is we
> >> may skip that burden and just add the missing field. Adding Juan, Dave &
> >> Peter for advices.
> > I'm not sure about users of this; if no one has used it then yeh; you
> > could bump up the version_id to make it a bit clearer.
> 
> Thank you for your input. Yes to me it sounds OK to only bump the
> version_id while adding the new field.

Ok. Just to make sure we're on the same page, this means we don't support
migration from new->old or old->new instances, since the migration stream
doesn't carry a version ID for the virtio-iommu-device and domain
vmstates, as far as I understand. I also believe backward-incompatible
changes are fine this time around, though I don't have much visibility in
what's being used.

Thanks,
Jean

* Jean-Philippe Brucker (jean-philippe@linaro.org) wrote:
> On Wed, Feb 02, 2022 at 02:21:37PM +0100, Eric Auger wrote:
> > >>> diff --git a/hw/virtio/virtio-iommu.c b/hw/virtio/virtio-iommu.c
> > >>> index ec02029bb6..a112428c65 100644
> > >>> --- a/hw/virtio/virtio-iommu.c
> > >>> +++ b/hw/virtio/virtio-iommu.c
> > >>> @@ -43,6 +43,7 @@
> > >>>  
> > >>>  typedef struct VirtIOIOMMUDomain {
> > >>>      uint32_t id;
> > >>> +    bool bypass;
> > >> I am afraid this will break the migration if you don't change
> > >> vmstate_domain.
> > >>
> > >> See static const VMStateDescription vmstate_domain.
> > >> Also you need to migrate the new bypass field.
> > >>
> > >> Logically we should handle this with a vmstate subsection I think to
> > >> handle migration of older devices. However I doubt the device has been
> > >> used in production environment supporting migration so my guess is we
> > >> may skip that burden and just add the missing field. Adding Juan, Dave &
> > >> Peter for advices.
> > > I'm not sure about users of this; if no one has used it then yeh; you
> > > could bump up the version_id to make it a bit clearer.
> > 
> > Thank you for your input. Yes to me it sounds OK to only bump the
> > version_id while adding the new field.
> 
> Ok. Just to make sure we're on the same page, this means we don't support
> migration from new->old or old->new instances, since the migration stream
> doesn't carry a version ID for the virtio-iommu-device and domain
> vmstates, as far as I understand. I also believe backward-incompatible
> changes are fine this time around, though I don't have much visibility in
> what's being used.

I think the stream only has it for top level devices; I've not dug into
this device.

Dave

> Thanks,
> Jean
> 
-- 
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK

Hi Jean,

On 2/8/22 2:09 PM, Dr. David Alan Gilbert wrote:
> * Jean-Philippe Brucker (jean-philippe@linaro.org) wrote:
>> On Wed, Feb 02, 2022 at 02:21:37PM +0100, Eric Auger wrote:
>>>>>> diff --git a/hw/virtio/virtio-iommu.c b/hw/virtio/virtio-iommu.c
>>>>>> index ec02029bb6..a112428c65 100644
>>>>>> --- a/hw/virtio/virtio-iommu.c
>>>>>> +++ b/hw/virtio/virtio-iommu.c
>>>>>> @@ -43,6 +43,7 @@
>>>>>>  
>>>>>>  typedef struct VirtIOIOMMUDomain {
>>>>>>      uint32_t id;
>>>>>> +    bool bypass;
>>>>> I am afraid this will break the migration if you don't change
>>>>> vmstate_domain.
>>>>>
>>>>> See static const VMStateDescription vmstate_domain.
>>>>> Also you need to migrate the new bypass field.
>>>>>
>>>>> Logically we should handle this with a vmstate subsection I think to
>>>>> handle migration of older devices. However I doubt the device has been
>>>>> used in production environment supporting migration so my guess is we
>>>>> may skip that burden and just add the missing field. Adding Juan, Dave &
>>>>> Peter for advices.
>>>> I'm not sure about users of this; if no one has used it then yeh; you
>>>> could bump up the version_id to make it a bit clearer.
>>> Thank you for your input. Yes to me it sounds OK to only bump the
>>> version_id while adding the new field.
>> Ok. Just to make sure we're on the same page, this means we don't support
>> migration from new->old or old->new instances, since the migration stream
>> doesn't carry a version ID for the virtio-iommu-device and domain
>> vmstates, as far as I understand. I also believe backward-incompatible
>> changes are fine this time around, though I don't have much visibility in
>> what's being used.
> I think the stream only has it for top level devices; I've not dug into
> this device.
Not sure I get what you meant:

vmstate_virtio_iommu has a version_id. Also vmstate_domain has one.

Thanks

Eric
>
> Dave
>
>> Thanks,
>> Jean
>>

On Tue, Feb 08, 2022 at 02:29:21PM +0100, Eric Auger wrote:
> Hi Jean,
> 
> On 2/8/22 2:09 PM, Dr. David Alan Gilbert wrote:
> > * Jean-Philippe Brucker (jean-philippe@linaro.org) wrote:
> >> On Wed, Feb 02, 2022 at 02:21:37PM +0100, Eric Auger wrote:
> >>>>>> diff --git a/hw/virtio/virtio-iommu.c b/hw/virtio/virtio-iommu.c
> >>>>>> index ec02029bb6..a112428c65 100644
> >>>>>> --- a/hw/virtio/virtio-iommu.c
> >>>>>> +++ b/hw/virtio/virtio-iommu.c
> >>>>>> @@ -43,6 +43,7 @@
> >>>>>>  
> >>>>>>  typedef struct VirtIOIOMMUDomain {
> >>>>>>      uint32_t id;
> >>>>>> +    bool bypass;
> >>>>> I am afraid this will break the migration if you don't change
> >>>>> vmstate_domain.
> >>>>>
> >>>>> See static const VMStateDescription vmstate_domain.
> >>>>> Also you need to migrate the new bypass field.
> >>>>>
> >>>>> Logically we should handle this with a vmstate subsection I think to
> >>>>> handle migration of older devices. However I doubt the device has been
> >>>>> used in production environment supporting migration so my guess is we
> >>>>> may skip that burden and just add the missing field. Adding Juan, Dave &
> >>>>> Peter for advices.
> >>>> I'm not sure about users of this; if no one has used it then yeh; you
> >>>> could bump up the version_id to make it a bit clearer.
> >>> Thank you for your input. Yes to me it sounds OK to only bump the
> >>> version_id while adding the new field.
> >> Ok. Just to make sure we're on the same page, this means we don't support
> >> migration from new->old or old->new instances, since the migration stream
> >> doesn't carry a version ID for the virtio-iommu-device and domain
> >> vmstates, as far as I understand. I also believe backward-incompatible
> >> changes are fine this time around, though I don't have much visibility in
> >> what's being used.
> > I think the stream only has it for top level devices; I've not dug into
> > this device.
> Not sure I get what you meant:
> 
> vmstate_virtio_iommu has a version_id. Also vmstate_domain has one.

These version numbers are not sent as part of the stream, unless I missed
it. On the incoming side, virtio_load() calls vmstate_load_state() for
vmstate_virtio_iommu_device and only looks at the internal version number
(dc->vmsd->version_id), it doesn't check version numbers from the stream.
So if the sender is old and the receiver is new, the stream doesn't have a
bypass field but the receiver will assume it does, and will consider the
stream corrupted.

Since we're not concerned about compatibility for the moment, I think we
could just add the bypass field without bumping the version number, the
behavior will be the same.

Thanks,
Jean