[libvirt] [PATCH 1/2] capabilities: Provide info about host IOMMU support

[Michal Privoznik]

On 05/25/2018 12:39 PM, Filip Alac wrote:
> Fixes:
> https://bugzilla.redhat.com/show_bug.cgi?id=967231
> 
> Signed-off-by: Filip Alac <filipalac at gmail.com>
> ---
>  docs/schemas/capability.rng  | 11 +++++++++++
>  src/conf/capabilities.c      |  8 ++++++++
>  src/conf/capabilities.h      |  5 +++++
>  src/libvirt_private.syms     |  1 +
>  src/qemu/qemu_capabilities.c |  4 ++++
>  src/util/virpci.c            | 19 +++++++++++++++++++
>  src/util/virpci.h            |  2 ++
>  7 files changed, 50 insertions(+)
> 
> diff --git a/docs/schemas/capability.rng b/docs/schemas/capability.rng
> index e1ab5c224..7e7d8fbc7 100644
> --- a/docs/schemas/capability.rng
> +++ b/docs/schemas/capability.rng
> @@ -39,6 +39,9 @@
>        <optional>
>          <ref name='power_management'/>
>        </optional>
> +      <optional>
> +        <ref name='iommu_support'/>
> +      </optional>
>        <optional>
>          <ref name='migration'/>
>        </optional>
> @@ -148,6 +151,14 @@
>      </element>
>    </define>
>  
> +  <define name='iommu_support'>
> +    <optional>
> +      <attribute name='support'>
> +        <ref name='virYesNo'/>
> +      </attribute>
> +    </optional>
> +  </define>

This does not make much sense. This defines an optional attribute
'support' but does not say to which element.

> +
>    <define name='migration'>
>      <element name='migration_features'>
>        <optional>
> diff --git a/src/conf/capabilities.c b/src/conf/capabilities.c
> index dd2fc77f9..eb387916f 100644
> --- a/src/conf/capabilities.c
> +++ b/src/conf/capabilities.c
> @@ -1020,6 +1020,8 @@ virCapabilitiesFormatXML(virCapsPtr caps)
>          }
>          virBufferAdjustIndent(&buf, -2);
>          virBufferAddLit(&buf, "</power_management>\n");
> +        virBufferAsprintf(&buf, "<iommu support='%s'/>\n",
> +                          caps->host.iommu  ? "yes" : "no");

This doesn't make much sense. This block starts with:

if (caps->host.powerMgmt) {

So you only print the <iommu/> iff power mgmt is available. These are
orthogonal features and have nothing in common.

>      } else {
>          /* The host does not support any PM feature. */
>          virBufferAddLit(&buf, "<power_management/>\n");
> @@ -1743,3 +1745,9 @@ virCapabilitiesInitCaches(virCapsPtr caps)
>      virBitmapFree(cpus);
>      return ret;
>  }
> +
> +int
> +virCapabilitiesHostInitIOMMU(virCapsPtr caps)
> +{
> +    return caps->host.iommu = virPCIHasIOMMU();
> +}

Whoa, this is very unsafe. I don't quite see why
virCapabilitiesHostInitIOMMU() should fail when host doesn't have IOMMU.
In fact, it's quite the opposite because false is 0 and true is 1.

> diff --git a/src/conf/capabilities.h b/src/conf/capabilities.h
> index f0a06a24d..4d41363a3 100644
> --- a/src/conf/capabilities.h
> +++ b/src/conf/capabilities.h
> @@ -183,6 +183,7 @@ struct _virCapsHost {
>      int nPagesSize;             /* size of pagesSize array */
>      unsigned int *pagesSize;    /* page sizes support on the system */
>      unsigned char host_uuid[VIR_UUID_BUFLEN];
> +    int iommu;

You declare this int even though you use it only to store a boolean.

>  };
>  
>  typedef int (*virDomainDefNamespaceParse)(xmlDocPtr, xmlNodePtr,
> @@ -327,4 +328,8 @@ void virCapsHostCacheBankFree(virCapsHostCacheBankPtr ptr);
>  
>  int virCapabilitiesInitCaches(virCapsPtr caps);
>  
> +int virCapabilitiesInitCaches(virCapsPtr caps);

No need to copy this line ;-)

> +
> +int virCapabilitiesHostInitIOMMU(virCapsPtr caps);
> +
>  #endif /* __VIR_CAPABILITIES_H */
> diff --git a/src/libvirt_private.syms b/src/libvirt_private.syms
> index a97b7fe22..258d02962 100644
> --- a/src/libvirt_private.syms
> +++ b/src/libvirt_private.syms
> @@ -58,6 +58,7 @@ virCapabilitiesFreeMachines;
>  virCapabilitiesFreeNUMAInfo;
>  virCapabilitiesGetCpusForNodemask;
>  virCapabilitiesGetNodeInfo;
> +virCapabilitiesHostInitIOMMU;
>  virCapabilitiesHostSecModelAddBaseLabel;
>  virCapabilitiesInitCaches;
>  virCapabilitiesInitNUMA;

You should expose virPCIHasIOMMU too. After all, you're exposing it in
header file.

> diff --git a/src/qemu/qemu_capabilities.c b/src/qemu/qemu_capabilities.c
> index 8a63db5f4..552d5452d 100644
> --- a/src/qemu/qemu_capabilities.c
> +++ b/src/qemu/qemu_capabilities.c
> @@ -948,6 +948,10 @@ virQEMUCapsInit(virFileCachePtr cache)
>      if (virCapabilitiesInitPages(caps) < 0)
>          VIR_WARN("Failed to get pages info");
>  
> +    /* Add IOMMU info */
> +    if (virCapabilitiesHostInitIOMMU(caps) < 0)
> +        VIR_WARN("Failed to get iommmu info");
> +

In the XML you placed <iommu/> between <power_management/> and
<migration_features/>. Might be worth honouring that order here too.

>      /* Add domain migration transport URIs */
>      virCapabilitiesAddHostMigrateTransport(caps, "tcp");
>      virCapabilitiesAddHostMigrateTransport(caps, "rdma");
> diff --git a/src/util/virpci.c b/src/util/virpci.c
> index 8d0236666..c88b13c97 100644
> --- a/src/util/virpci.c
> +++ b/src/util/virpci.c
> @@ -3288,3 +3288,22 @@ virPCIEDeviceInfoFree(virPCIEDeviceInfoPtr dev)
>      VIR_FREE(dev->link_sta);
>      VIR_FREE(dev);
>  }
> +

See how the rest of functions in the file is separated by two empty
lines? Might be worth preserving that.

> +bool
> +virPCIHasIOMMU(void)
> +{
> +    struct stat sb;
> +
> +    /* We can only check on newer kernels with iommu groups & vfio */
> +    if (stat("/sys/kernel/iommu_groups", &sb) < 0)
> +        return false;
> +
> +    if (!S_ISDIR(sb.st_mode))
> +        return false;
> +
> +    /* Check if folder is empty */
> +    if (sb.st_nlink <= 2)
> +        return false;
> +
> +    return true;
> +}

This is similar to qemuHostdevHostSupportsPassthroughVFIO() which got me
thinking, do we even need this patch? I mean, we already kind of expose
if host is VFIO capable:

/domainCapabilities/hostdev/enum[name='pciBackend'/value/vfio.

Michal