[PATCH v3 0/3] VIRTIO-IOMMU: Set default granule to host page size

[PATCH v3 0/3] VIRTIO-IOMMU: Set default granule to host page size

[Eric Auger]

We used to set the default granule to 4KB but with VFIO assignment
it makes more sense to use the actual host page size.

Indeed when hotplugging a VFIO device protected by a virtio-iommu
on a 64kB/64kB host/guest config, we current get a qemu crash:

"vfio: DMA mapping failed, unable to continue"

This is due to the hot-attached VFIO device calling
memory_region_iommu_set_page_size_mask() with 64kB granule
whereas the virtio-iommu granule was already frozen to 4KB on
machine init done.

Introduce a new granule property and set this latter to "host"
and introduce a new compat.

Note that the new default will prevent 4kB guest on 64kB host
because the granule will be set to 64kB which would be larger
than the guest page size. In that situation, the virtio-iommu
driver fails on viommu_domain_finalise() with
"granule 0x10000 larger than system page size 0x1000".

The current limitation of global granule in the virtio-iommu
should be removed and turned into per domain granule. But
until we get this upgraded, this new default is probably
better because I don't think anyone is currently interested in
running a 4KB page size guest with virtio-iommu on a 64KB host.
However supporting 64kB guest on 64kB host with virtio-iommu and
VFIO looks a more important feature.

This series can be found at:
https://github.com/eauger/qemu/tree/granule-v1

Previous version:
[PATCH v2] virtio-iommu: Use qemu_real_host_page_mask as default page_size_mask
https://lore.kernel.org/all/20240117132039.332273-1-eric.auger@redhat.com/

Applied on top of
[PATCH v5 0/4] VIRTIO-IOMMU: Introduce an aw-bits option
https://lore.kernel.org/all/20240215084315.863897-1-eric.auger@redhat.com/

History:
v2 -> v3
- introduce a dedicated granule option to handle the compat

Eric Auger (3):
  qdev: Add a granule_mode property
  virtio-iommu: Add a granule property
  virtio-iommu: Change the default granule to the host page size

 include/hw/qdev-properties-system.h |  3 +++
 include/hw/virtio/virtio-iommu.h    | 11 ++++++++++
 hw/core/machine.c                   |  1 +
 hw/core/qdev-properties-system.c    | 15 +++++++++++++
 hw/virtio/virtio-iommu.c            | 34 ++++++++++++++++++++++++++---
 qemu-options.hx                     |  3 +++
 6 files changed, 64 insertions(+), 3 deletions(-)

-- 
2.41.0

[PATCH v3 1/3] qdev: Add a granule_mode property

[Eric Auger]

Introduce a new enum type property allowing to set an
IOMMU granule. Values are 4K, 16K, 64K and host. This
latter indicates the vIOMMU granule will matches the
host page size.

A subsequent patch will add such a property to the
virtio-iommu device.

Signed-off-by: Eric Auger <eric.auger@redhat.com>
---
 include/hw/qdev-properties-system.h |  3 +++
 include/hw/virtio/virtio-iommu.h    | 11 +++++++++++
 hw/core/qdev-properties-system.c    | 15 +++++++++++++++
 hw/virtio/virtio-iommu.c            | 10 ++++++++++
 4 files changed, 39 insertions(+)

diff --git a/include/hw/qdev-properties-system.h b/include/hw/qdev-properties-system.h
index 06c359c190..626be87dd3 100644
--- a/include/hw/qdev-properties-system.h
+++ b/include/hw/qdev-properties-system.h
@@ -8,6 +8,7 @@ extern const PropertyInfo qdev_prop_macaddr;
 extern const PropertyInfo qdev_prop_reserved_region;
 extern const PropertyInfo qdev_prop_multifd_compression;
 extern const PropertyInfo qdev_prop_mig_mode;
+extern const PropertyInfo qdev_prop_granule_mode;
 extern const PropertyInfo qdev_prop_losttickpolicy;
 extern const PropertyInfo qdev_prop_blockdev_on_error;
 extern const PropertyInfo qdev_prop_bios_chs_trans;
@@ -47,6 +48,8 @@ extern const PropertyInfo qdev_prop_iothread_vq_mapping_list;
 #define DEFINE_PROP_MIG_MODE(_n, _s, _f, _d) \
     DEFINE_PROP_SIGNED(_n, _s, _f, _d, qdev_prop_mig_mode, \
                        MigMode)
+#define DEFINE_PROP_GRANULE_MODE(_n, _s, _f, _d) \
+    DEFINE_PROP_SIGNED(_n, _s, _f, _d, qdev_prop_granule_mode, GranuleMode)
 #define DEFINE_PROP_LOSTTICKPOLICY(_n, _s, _f, _d) \
     DEFINE_PROP_SIGNED(_n, _s, _f, _d, qdev_prop_losttickpolicy, \
                         LostTickPolicy)
diff --git a/include/hw/virtio/virtio-iommu.h b/include/hw/virtio/virtio-iommu.h
index 5fbe4677c2..a82c4fa471 100644
--- a/include/hw/virtio/virtio-iommu.h
+++ b/include/hw/virtio/virtio-iommu.h
@@ -31,6 +31,16 @@ OBJECT_DECLARE_SIMPLE_TYPE(VirtIOIOMMU, VIRTIO_IOMMU)
 
 #define TYPE_VIRTIO_IOMMU_MEMORY_REGION "virtio-iommu-memory-region"
 
+typedef enum GranuleMode {
+    GRANULE_MODE_4K,
+    GRANULE_MODE_16K,
+    GRANULE_MODE_64K,
+    GRANULE_MODE_HOST,
+    GRANULE_MODE__MAX,
+} GranuleMode;
+
+extern const QEnumLookup GranuleMode_lookup;
+
 typedef struct IOMMUDevice {
     void         *viommu;
     PCIBus       *bus;
@@ -67,6 +77,7 @@ struct VirtIOIOMMU {
     Notifier machine_done;
     bool granule_frozen;
     uint8_t aw_bits;
+    GranuleMode granule_mode;
 };
 
 #endif
diff --git a/hw/core/qdev-properties-system.c b/hw/core/qdev-properties-system.c
index 1a396521d5..3a0b36a7a7 100644
--- a/hw/core/qdev-properties-system.c
+++ b/hw/core/qdev-properties-system.c
@@ -34,6 +34,7 @@
 #include "net/net.h"
 #include "hw/pci/pci.h"
 #include "hw/pci/pcie.h"
+#include "hw/virtio/virtio-iommu.h"
 #include "hw/i386/x86.h"
 #include "util/block-helpers.h"
 
@@ -679,6 +680,20 @@ const PropertyInfo qdev_prop_mig_mode = {
     .set_default_value = qdev_propinfo_set_default_value_enum,
 };
 
+/* --- GranuleMode --- */
+
+QEMU_BUILD_BUG_ON(sizeof(GranuleMode) != sizeof(int));
+
+const PropertyInfo qdev_prop_granule_mode = {
+    .name = "GranuleMode",
+    .description = "granule_mode values, "
+                   "4K, 16K, 64K, host",
+    .enum_table = &GranuleMode_lookup,
+    .get = qdev_propinfo_get_enum,
+    .set = qdev_propinfo_set_enum,
+    .set_default_value = qdev_propinfo_set_default_value_enum,
+};
+
 /* --- Reserved Region --- */
 
 /*
diff --git a/hw/virtio/virtio-iommu.c b/hw/virtio/virtio-iommu.c
index 2ec5ef3cd1..2ce5839c60 100644
--- a/hw/virtio/virtio-iommu.c
+++ b/hw/virtio/virtio-iommu.c
@@ -44,6 +44,16 @@
 #define VIOMMU_DEFAULT_QUEUE_SIZE 256
 #define VIOMMU_PROBE_SIZE 512
 
+const QEnumLookup GranuleMode_lookup = {
+    .array = (const char *const[]) {
+        [GRANULE_MODE_4K]  = "4K",
+        [GRANULE_MODE_16K] = "16K",
+        [GRANULE_MODE_64K] = "64K",
+        [GRANULE_MODE_HOST] = "host",
+    },
+    .size = GRANULE_MODE__MAX
+};
+
 typedef struct VirtIOIOMMUDomain {
     uint32_t id;
     bool bypass;
-- 
2.41.0

[PATCH v3 2/3] virtio-iommu: Add a granule property

[Eric Auger]

This allows to choose which granule will be used by
default by the virtio-iommu. Current default is 4K.

Signed-off-by: Eric Auger <eric.auger@redhat.com>
---
 hw/virtio/virtio-iommu.c | 24 +++++++++++++++++++++---
 qemu-options.hx          |  3 +++
 2 files changed, 24 insertions(+), 3 deletions(-)

diff --git a/hw/virtio/virtio-iommu.c b/hw/virtio/virtio-iommu.c
index 2ce5839c60..1eded09134 100644
--- a/hw/virtio/virtio-iommu.c
+++ b/hw/virtio/virtio-iommu.c
@@ -1125,8 +1125,8 @@ static int virtio_iommu_notify_flag_changed(IOMMUMemoryRegion *iommu_mr,
 }
 
 /*
- * The default mask (TARGET_PAGE_MASK) is the smallest supported guest granule,
- * for example 0xfffffffffffff000. When an assigned device has page size
+ * The default mask depends on the "granule" property. For example, with
+ * 4K granule, it is ~0xFFF. When an assigned device has page size
  * restrictions due to the hardware IOMMU configuration, apply this restriction
  * to the mask.
  */
@@ -1323,7 +1323,23 @@ static void virtio_iommu_device_realize(DeviceState *dev, Error **errp)
      * in vfio realize
      */
     s->config.bypass = s->boot_bypass;
-    s->config.page_size_mask = qemu_target_page_mask();
+
+    switch (s->granule_mode) {
+    case GRANULE_MODE_4K:
+        s->config.page_size_mask = ~0xFFF;
+        break;
+    case GRANULE_MODE_16K:
+        s->config.page_size_mask = ~0x3FFF;
+        break;
+    case GRANULE_MODE_64K:
+        s->config.page_size_mask = ~0xFFFF;
+        break;
+    case GRANULE_MODE_HOST:
+        s->config.page_size_mask = qemu_real_host_page_mask();
+        break;
+    default:
+        error_setg(errp, "Unsupported granule mode");
+    }
     if (s->aw_bits < 32 || s->aw_bits > 64) {
         error_setg(errp, "aw-bits must be within [32,64]");
     }
@@ -1537,6 +1553,8 @@ static Property virtio_iommu_properties[] = {
                      TYPE_PCI_BUS, PCIBus *),
     DEFINE_PROP_BOOL("boot-bypass", VirtIOIOMMU, boot_bypass, true),
     DEFINE_PROP_UINT8("aw-bits", VirtIOIOMMU, aw_bits, 0),
+    DEFINE_PROP_GRANULE_MODE("granule", VirtIOIOMMU, granule_mode,
+                             GRANULE_MODE_4K),
     DEFINE_PROP_END_OF_LIST(),
 };
 
diff --git a/qemu-options.hx b/qemu-options.hx
index a98bc7bd60..6ae27c2cff 100644
--- a/qemu-options.hx
+++ b/qemu-options.hx
@@ -1179,6 +1179,9 @@ SRST
     ``aw-bits=val`` (val between 32 and 64, default depends on machine)
         This decides the address width of IOVA address space. It defaults
         to 39 bits on q35 machines and 48 bits on ARM virt machines.
+    ``granule=val`` (possible values are 4K, 16K, 64K and host)
+        This decides the default granule to be be exposed by the
+        virtio-iommu. If host, the granule is the same as the host page size.
 
 ERST
 
-- 
2.41.0

[PATCH v3 3/3] virtio-iommu: Change the default granule to the host page size

[Eric Auger]

We used to set the default granule to 4KB but with VFIO assignment
it makes more sense to use the actual host page size.

Indeed when hotplugging a VFIO device protected by a virtio-iommu
on a 64kB/64kB host/guest config, we current get a qemu crash:

"vfio: DMA mapping failed, unable to continue"

This is due to the hot-attached VFIO device calling
memory_region_iommu_set_page_size_mask() with 64kB granule
whereas the virtio-iommu granule was already frozen to 4KB on
machine init done.

Set the granule property to "host" and introduce a new compat.

Note that the new default will prevent 4kB guest on 64kB host
because the granule will be set to 64kB which would be larger
than the guest page size. In that situation, the virtio-iommu
driver fails on viommu_domain_finalise() with
"granule 0x10000 larger than system page size 0x1000".

The current limitation of global granule in the virtio-iommu
should be removed and turned into per domain granule. But
until we get this upgraded, this new default is probably
better because I don't think anyone is currently interested in
running a 4KB page size guest with virtio-iommu on a 64KB host.
However supporting 64kB guest on 64kB host with virtio-iommu and
VFIO looks a more important feature.

Signed-off-by: Eric Auger <eric.auger@redhat.com>
---
 hw/core/machine.c        | 1 +
 hw/virtio/virtio-iommu.c | 2 +-
 2 files changed, 2 insertions(+), 1 deletion(-)

diff --git a/hw/core/machine.c b/hw/core/machine.c
index 70ac96954c..38851df4b8 100644
--- a/hw/core/machine.c
+++ b/hw/core/machine.c
@@ -35,6 +35,7 @@
 
 GlobalProperty hw_compat_8_2[] = {
     { TYPE_VIRTIO_IOMMU_PCI, "aw-bits", "64" },
+    { TYPE_VIRTIO_IOMMU_PCI, "granule", "4K" },
 };
 const size_t hw_compat_8_2_len = G_N_ELEMENTS(hw_compat_8_2);
 
diff --git a/hw/virtio/virtio-iommu.c b/hw/virtio/virtio-iommu.c
index 1eded09134..0a6e2e6c9f 100644
--- a/hw/virtio/virtio-iommu.c
+++ b/hw/virtio/virtio-iommu.c
@@ -1554,7 +1554,7 @@ static Property virtio_iommu_properties[] = {
     DEFINE_PROP_BOOL("boot-bypass", VirtIOIOMMU, boot_bypass, true),
     DEFINE_PROP_UINT8("aw-bits", VirtIOIOMMU, aw_bits, 0),
     DEFINE_PROP_GRANULE_MODE("granule", VirtIOIOMMU, granule_mode,
-                             GRANULE_MODE_4K),
+                             GRANULE_MODE_HOST),
     DEFINE_PROP_END_OF_LIST(),
 };
 
-- 
2.41.0

Re: [PATCH v3 1/3] qdev: Add a granule_mode property

[Richard Henderson]

On 2/21/24 10:58, Eric Auger wrote:
> Introduce a new enum type property allowing to set an
> IOMMU granule. Values are 4K, 16K, 64K and host. This
> latter indicates the vIOMMU granule will matches the
> host page size.
> 
> A subsequent patch will add such a property to the
> virtio-iommu device.
> 
> Signed-off-by: Eric Auger <eric.auger@redhat.com>
> ---
>   include/hw/qdev-properties-system.h |  3 +++
>   include/hw/virtio/virtio-iommu.h    | 11 +++++++++++
>   hw/core/qdev-properties-system.c    | 15 +++++++++++++++
>   hw/virtio/virtio-iommu.c            | 10 ++++++++++
>   4 files changed, 39 insertions(+)
> 
> diff --git a/include/hw/qdev-properties-system.h b/include/hw/qdev-properties-system.h
> index 06c359c190..626be87dd3 100644
> --- a/include/hw/qdev-properties-system.h
> +++ b/include/hw/qdev-properties-system.h
> @@ -8,6 +8,7 @@ extern const PropertyInfo qdev_prop_macaddr;
>   extern const PropertyInfo qdev_prop_reserved_region;
>   extern const PropertyInfo qdev_prop_multifd_compression;
>   extern const PropertyInfo qdev_prop_mig_mode;
> +extern const PropertyInfo qdev_prop_granule_mode;
>   extern const PropertyInfo qdev_prop_losttickpolicy;
>   extern const PropertyInfo qdev_prop_blockdev_on_error;
>   extern const PropertyInfo qdev_prop_bios_chs_trans;
> @@ -47,6 +48,8 @@ extern const PropertyInfo qdev_prop_iothread_vq_mapping_list;
>   #define DEFINE_PROP_MIG_MODE(_n, _s, _f, _d) \
>       DEFINE_PROP_SIGNED(_n, _s, _f, _d, qdev_prop_mig_mode, \
>                          MigMode)
> +#define DEFINE_PROP_GRANULE_MODE(_n, _s, _f, _d) \
> +    DEFINE_PROP_SIGNED(_n, _s, _f, _d, qdev_prop_granule_mode, GranuleMode)
>   #define DEFINE_PROP_LOSTTICKPOLICY(_n, _s, _f, _d) \
>       DEFINE_PROP_SIGNED(_n, _s, _f, _d, qdev_prop_losttickpolicy, \
>                           LostTickPolicy)
> diff --git a/include/hw/virtio/virtio-iommu.h b/include/hw/virtio/virtio-iommu.h
> index 5fbe4677c2..a82c4fa471 100644
> --- a/include/hw/virtio/virtio-iommu.h
> +++ b/include/hw/virtio/virtio-iommu.h
> @@ -31,6 +31,16 @@ OBJECT_DECLARE_SIMPLE_TYPE(VirtIOIOMMU, VIRTIO_IOMMU)
>   
>   #define TYPE_VIRTIO_IOMMU_MEMORY_REGION "virtio-iommu-memory-region"
>   
> +typedef enum GranuleMode {
> +    GRANULE_MODE_4K,
> +    GRANULE_MODE_16K,
> +    GRANULE_MODE_64K,
> +    GRANULE_MODE_HOST,
> +    GRANULE_MODE__MAX,
> +} GranuleMode;
> +
> +extern const QEnumLookup GranuleMode_lookup;
> +
>   typedef struct IOMMUDevice {
>       void         *viommu;
>       PCIBus       *bus;
> @@ -67,6 +77,7 @@ struct VirtIOIOMMU {
>       Notifier machine_done;
>       bool granule_frozen;
>       uint8_t aw_bits;
> +    GranuleMode granule_mode;
>   };
>   
>   #endif
> diff --git a/hw/core/qdev-properties-system.c b/hw/core/qdev-properties-system.c
> index 1a396521d5..3a0b36a7a7 100644
> --- a/hw/core/qdev-properties-system.c
> +++ b/hw/core/qdev-properties-system.c
> @@ -34,6 +34,7 @@
>   #include "net/net.h"
>   #include "hw/pci/pci.h"
>   #include "hw/pci/pcie.h"
> +#include "hw/virtio/virtio-iommu.h"
>   #include "hw/i386/x86.h"
>   #include "util/block-helpers.h"
>   
> @@ -679,6 +680,20 @@ const PropertyInfo qdev_prop_mig_mode = {
>       .set_default_value = qdev_propinfo_set_default_value_enum,
>   };
>   
> +/* --- GranuleMode --- */
> +
> +QEMU_BUILD_BUG_ON(sizeof(GranuleMode) != sizeof(int));
> +
> +const PropertyInfo qdev_prop_granule_mode = {
> +    .name = "GranuleMode",
> +    .description = "granule_mode values, "
> +                   "4K, 16K, 64K, host",
> +    .enum_table = &GranuleMode_lookup,
> +    .get = qdev_propinfo_get_enum,
> +    .set = qdev_propinfo_set_enum,
> +    .set_default_value = qdev_propinfo_set_default_value_enum,
> +};
> +
>   /* --- Reserved Region --- */
>   
>   /*
> diff --git a/hw/virtio/virtio-iommu.c b/hw/virtio/virtio-iommu.c
> index 2ec5ef3cd1..2ce5839c60 100644
> --- a/hw/virtio/virtio-iommu.c
> +++ b/hw/virtio/virtio-iommu.c
> @@ -44,6 +44,16 @@
>   #define VIOMMU_DEFAULT_QUEUE_SIZE 256
>   #define VIOMMU_PROBE_SIZE 512
>   
> +const QEnumLookup GranuleMode_lookup = {
> +    .array = (const char *const[]) {
> +        [GRANULE_MODE_4K]  = "4K",
> +        [GRANULE_MODE_16K] = "16K",
> +        [GRANULE_MODE_64K] = "64K",
> +        [GRANULE_MODE_HOST] = "host",
> +    },
> +    .size = GRANULE_MODE__MAX
> +};

Does it ever make sense to use anything other than host?
Especially since 4k fails on a 64k host, as you report.
Why would we want to be able to select a granule that doesn't work?

There are a few older hosts with 8k pages, e.g. Alpha, MIPS, Sparc.


r~

Re: [PATCH v3 1/3] qdev: Add a granule_mode property

[Eric Auger]

Hi Richard,

On 2/21/24 22:58, Richard Henderson wrote:
> On 2/21/24 10:58, Eric Auger wrote:
>> Introduce a new enum type property allowing to set an
>> IOMMU granule. Values are 4K, 16K, 64K and host. This
>> latter indicates the vIOMMU granule will matches the
>> host page size.
>>
>> A subsequent patch will add such a property to the
>> virtio-iommu device.
>>
>> Signed-off-by: Eric Auger <eric.auger@redhat.com>
>> ---
>>   include/hw/qdev-properties-system.h |  3 +++
>>   include/hw/virtio/virtio-iommu.h    | 11 +++++++++++
>>   hw/core/qdev-properties-system.c    | 15 +++++++++++++++
>>   hw/virtio/virtio-iommu.c            | 10 ++++++++++
>>   4 files changed, 39 insertions(+)
>>
>> diff --git a/include/hw/qdev-properties-system.h
>> b/include/hw/qdev-properties-system.h
>> index 06c359c190..626be87dd3 100644
>> --- a/include/hw/qdev-properties-system.h
>> +++ b/include/hw/qdev-properties-system.h
>> @@ -8,6 +8,7 @@ extern const PropertyInfo qdev_prop_macaddr;
>>   extern const PropertyInfo qdev_prop_reserved_region;
>>   extern const PropertyInfo qdev_prop_multifd_compression;
>>   extern const PropertyInfo qdev_prop_mig_mode;
>> +extern const PropertyInfo qdev_prop_granule_mode;
>>   extern const PropertyInfo qdev_prop_losttickpolicy;
>>   extern const PropertyInfo qdev_prop_blockdev_on_error;
>>   extern const PropertyInfo qdev_prop_bios_chs_trans;
>> @@ -47,6 +48,8 @@ extern const PropertyInfo
>> qdev_prop_iothread_vq_mapping_list;
>>   #define DEFINE_PROP_MIG_MODE(_n, _s, _f, _d) \
>>       DEFINE_PROP_SIGNED(_n, _s, _f, _d, qdev_prop_mig_mode, \
>>                          MigMode)
>> +#define DEFINE_PROP_GRANULE_MODE(_n, _s, _f, _d) \
>> +    DEFINE_PROP_SIGNED(_n, _s, _f, _d, qdev_prop_granule_mode,
>> GranuleMode)
>>   #define DEFINE_PROP_LOSTTICKPOLICY(_n, _s, _f, _d) \
>>       DEFINE_PROP_SIGNED(_n, _s, _f, _d, qdev_prop_losttickpolicy, \
>>                           LostTickPolicy)
>> diff --git a/include/hw/virtio/virtio-iommu.h
>> b/include/hw/virtio/virtio-iommu.h
>> index 5fbe4677c2..a82c4fa471 100644
>> --- a/include/hw/virtio/virtio-iommu.h
>> +++ b/include/hw/virtio/virtio-iommu.h
>> @@ -31,6 +31,16 @@ OBJECT_DECLARE_SIMPLE_TYPE(VirtIOIOMMU, VIRTIO_IOMMU)
>>     #define TYPE_VIRTIO_IOMMU_MEMORY_REGION "virtio-iommu-memory-region"
>>   +typedef enum GranuleMode {
>> +    GRANULE_MODE_4K,
>> +    GRANULE_MODE_16K,
>> +    GRANULE_MODE_64K,
>> +    GRANULE_MODE_HOST,
>> +    GRANULE_MODE__MAX,
>> +} GranuleMode;
>> +
>> +extern const QEnumLookup GranuleMode_lookup;
>> +
>>   typedef struct IOMMUDevice {
>>       void         *viommu;
>>       PCIBus       *bus;
>> @@ -67,6 +77,7 @@ struct VirtIOIOMMU {
>>       Notifier machine_done;
>>       bool granule_frozen;
>>       uint8_t aw_bits;
>> +    GranuleMode granule_mode;
>>   };
>>     #endif
>> diff --git a/hw/core/qdev-properties-system.c
>> b/hw/core/qdev-properties-system.c
>> index 1a396521d5..3a0b36a7a7 100644
>> --- a/hw/core/qdev-properties-system.c
>> +++ b/hw/core/qdev-properties-system.c
>> @@ -34,6 +34,7 @@
>>   #include "net/net.h"
>>   #include "hw/pci/pci.h"
>>   #include "hw/pci/pcie.h"
>> +#include "hw/virtio/virtio-iommu.h"
>>   #include "hw/i386/x86.h"
>>   #include "util/block-helpers.h"
>>   @@ -679,6 +680,20 @@ const PropertyInfo qdev_prop_mig_mode = {
>>       .set_default_value = qdev_propinfo_set_default_value_enum,
>>   };
>>   +/* --- GranuleMode --- */
>> +
>> +QEMU_BUILD_BUG_ON(sizeof(GranuleMode) != sizeof(int));
>> +
>> +const PropertyInfo qdev_prop_granule_mode = {
>> +    .name = "GranuleMode",
>> +    .description = "granule_mode values, "
>> +                   "4K, 16K, 64K, host",
>> +    .enum_table = &GranuleMode_lookup,
>> +    .get = qdev_propinfo_get_enum,
>> +    .set = qdev_propinfo_set_enum,
>> +    .set_default_value = qdev_propinfo_set_default_value_enum,
>> +};
>> +
>>   /* --- Reserved Region --- */
>>     /*
>> diff --git a/hw/virtio/virtio-iommu.c b/hw/virtio/virtio-iommu.c
>> index 2ec5ef3cd1..2ce5839c60 100644
>> --- a/hw/virtio/virtio-iommu.c
>> +++ b/hw/virtio/virtio-iommu.c
>> @@ -44,6 +44,16 @@
>>   #define VIOMMU_DEFAULT_QUEUE_SIZE 256
>>   #define VIOMMU_PROBE_SIZE 512
>>   +const QEnumLookup GranuleMode_lookup = {
>> +    .array = (const char *const[]) {
>> +        [GRANULE_MODE_4K]  = "4K",
>> +        [GRANULE_MODE_16K] = "16K",
>> +        [GRANULE_MODE_64K] = "64K",
>> +        [GRANULE_MODE_HOST] = "host",
>> +    },
>> +    .size = GRANULE_MODE__MAX
>> +};
>
> Does it ever make sense to use anything other than host?
> Especially since 4k fails on a 64k host, as you report.
Let me reuse Jean-Philippe's support matrix:
Before this series, 4KB granule was always used by the virtio-iommu.
The support matrix was:

 Host | Guest | virtio-net | IGB passthrough
  4k  | 4k    | Y          | Y
  64k | 64k   | Y          | N
  64k | 4k    | Y          | N
  4k  | 64k   | Y          | Y

with this series, host becomes the default. It fixes the 64KB/64KB config by default. It allows 64KB/4KB (virtio only) with manual 4K granule setting. If we don't have this 4K granule we cannot support that use case.

 Host | Guest | virtio-net | IGB passthrough
  4k  | 4k    | Y          | Y
  64k | 64k   | Y          | Y
  64k | 4k    | 4K         | N
  4k  | 64k   | Y          | Y

Besides I need 4KB to handle the compat for older machine types.

> Why would we want to be able to select a granule that doesn't work?

> There are a few older hosts with 8k pages, e.g. Alpha, MIPS, Sparc.
OK. those are not integrated with virtio-iommu but I can definitively add 8K

Eric
>
>
> r~
>