Skip to contentSkip to conversation

Support for nesting IOMMUs in VFIO

PATCH Superseded Will Deacon 4 participants 12 messages Sep 2, 2014 – Sep 12, 2014 7 messages not in the archive

Subsystems (2)

v3 of 5 Superseded Latest is v5 →

lore
Share & tools
lore
#0 lore raw

v3 cover Support for nesting IOMMUs in VFIO

Hello,

This is version three of the patches I originally posted here:

RFCv1: http://permalink.gmane.org/gmane.linux.kernel.iommu/5552
RFCv2: http://permalink.gmane.org/gmane.linux.kernel.iommu/5700

Changes since RFCv2 include:

- Dropped the RFC tag
- Rebased onto 3.17-rc*
- Iterate the support bus_types (currently just PCI) and check that nesting is actually supported

The corresponding arm-smmu changes are included to show how the new domain attribute can be used.

All feedback welcome,

Will

-->8

Will Deacon (3):
iommu: introduce domain attribute for nesting IOMMUs vfio/iommu_type1: add new VFIO_TYPE1_NESTING_IOMMU IOMMU type iommu/arm-smmu: add support for DOMAIN_ATTR_NESTING attribute

 drivers/iommu/arm-smmu.c        | 110 ++++++++++++++++++++++++++++++++--------
 drivers/vfio/vfio_iommu_type1.c |  87 ++++++++++++++++++++++++++++---
 include/linux/iommu.h           |   1 +
 include/uapi/linux/vfio.h       |   2 +
 4 files changed, 173 insertions(+), 27 deletions(-)
Signature · 1 line
2.1.0

A message that is not in the archive (sent off-list or not yet received): 1409651619-27342-1-git-send-email-will.deacon-5wv7dgnIgG8@public.gmane.org

#2 lore raw

PATCH v3 1/3 iommu: introduce domain attribute for nesting IOMMUs

Some IOMMUs, such as the ARM SMMU, support two stages of translation. The idea behind such a scheme is to allow a guest operating system to use the IOMMU for DMA mappings in the first stage of translation, with the hypervisor then installing mappings in the second stage to provide isolation of the DMA to the physical range assigned to that virtual machine.

In order to allow IOMMU domains to be used for second-stage translation, this patch adds a new iommu_attr (IOMMU_ATTR_NESTING) for setting second-stage domains prior to device attach. The attribute can also be queried to see if a domain is actually making use of nesting.

Cc Joerg Roedel <[email protected]>
Cc Alex Williamson <[email protected]>
---
 include/linux/iommu.h | 1 +
 1 file changed, 1 insertion(+)
include/linux/iommu.h +1 -0
index 20f9a527922a..7b02bcc85b9e 100644
@@ -80,6 +80,7 @@ enum iommu_attr {
 	DOMAIN_ATTR_FSL_PAMU_STASH,
 	DOMAIN_ATTR_FSL_PAMU_ENABLE,
 	DOMAIN_ATTR_FSL_PAMUV1,
+	DOMAIN_ATTR_NESTING,	/* two stages of translation */
 	DOMAIN_ATTR_MAX,
 };
 
Signature · 1 line
2.1.0
#3 lore raw

PATCH v3 2/3 vfio/iommu_type1: add new VFIO_TYPE1_NESTING_IOMMU IOMMU type

In reply to a message not in the archive

VFIO allows devices to be safely handed off to userspace by putting them behind an IOMMU configured to ensure DMA and interrupt isolation. This enables userspace KVM clients, such as kvmtool and qemu, to further map the device into a virtual machine.

With IOMMUs such as the ARM SMMU, it is then possible to provide SMMU translation services to the guest operating system, which are nested with the existing translation installed by VFIO. However, enabling this feature means that the IOMMU driver must be informed that the VFIO domain is being created for the purposes of nested translation.

This patch adds a new IOMMU type (VFIO_TYPE1_NESTING_IOMMU) to the VFIO type-1 driver. The new IOMMU type acts identically to the VFIO_TYPE1v2_IOMMU type, but additionally sets the DOMAIN_ATTR_NESTING attribute on its IOMMU domains. Userspace can check whether nesting is actually available using the VFIO_CHECK_EXTENSION ioctl, in a similar manner to checking for cache-coherent DMA.

Cc Joerg Roedel <[email protected]>
Cc Alex Williamson <[email protected]>
---
 drivers/vfio/vfio_iommu_type1.c | 87 +++++++++++++++++++++++++++++++++++++----
 include/uapi/linux/vfio.h       |  2 +
 2 files changed, 82 insertions(+), 7 deletions(-)
drivers/vfio/vfio_iommu_type1.c +80 -7
index 0734fbe5b651..50aa56c7b6a7 100644
@@ -53,11 +53,15 @@ module_param_named(disable_hugepages,
 MODULE_PARM_DESC(disable_hugepages,
 		 "Disable VFIO IOMMU support for IOMMU hugepages.");
 
+/* Feature flags for VFIO Type-1 IOMMUs */
+#define VFIO_IOMMU_FEAT_V2	(1 << 0)
+#define VFIO_IOMMU_FEAT_NESTING	(1 << 1)
+
 struct vfio_iommu {
 	struct list_head	domain_list;
 	struct mutex		lock;
 	struct rb_root		dma_list;
-	bool v2;
+	int			features;
 };
 
 struct vfio_domain {
@@ -441,7 +445,7 @@ static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
 	 * will only return success and a size of zero if there were no
 	 * mappings within the range.
 	 */
-	if (iommu->v2) {
+	if (iommu->features & VFIO_IOMMU_FEAT_V2) {
 		dma = vfio_find_dma(iommu, unmap->iova, 0);
 		if (dma && dma->iova != unmap->iova) {
 			ret = -EINVAL;
@@ -455,7 +459,8 @@ static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
 	}
 
 	while ((dma = vfio_find_dma(iommu, unmap->iova, unmap->size))) {
-		if (!iommu->v2 && unmap->iova > dma->iova)
+		if (!(iommu->features & VFIO_IOMMU_FEAT_V2) &&
+		    unmap->iova > dma->iova)
 			break;
 		unmapped += dma->size;
 		vfio_remove_dma(iommu, dma);
@@ -671,7 +676,7 @@ static int vfio_iommu_type1_attach_group(void *iommu_data,
 	struct vfio_group *group, *g;
 	struct vfio_domain *domain, *d;
 	struct bus_type *bus = NULL;
-	int ret;
+	int ret, attr = 1;
 
 	mutex_lock(&iommu->lock);
 
@@ -705,6 +710,9 @@ static int vfio_iommu_type1_attach_group(void *iommu_data,
 		goto out_free;
 	}
 
+	if (iommu->features & VFIO_IOMMU_FEAT_NESTING)
+		iommu_domain_set_attr(domain->domain, DOMAIN_ATTR_NESTING, &attr);
+
 	ret = iommu_attach_group(domain->domain, iommu_group);
 	if (ret)
 		goto out_domain;
@@ -815,12 +823,50 @@ done:
 	mutex_unlock(&iommu->lock);
 }
 
+static int vfio_iommus_have_nesting(void)
+{
+	static struct bus_type *bus_types[] = { &pci_bus_type };
+	int i, ret = 0;
+
+	for (i = 0; !ret && i < ARRAY_SIZE(bus_types); ++i) {
+		struct iommu_domain *domain;
+		struct bus_type *bus = bus_types[i];
+		int attr = 1;
+
+		if (!iommu_present(bus))
+			continue;
+
+		domain = iommu_domain_alloc(bus);
+		if (!domain)
+			continue;
+
+		if (!iommu_domain_set_attr(domain, DOMAIN_ATTR_NESTING, &attr))
+			ret = 1;
+
+		iommu_domain_free(domain);
+	}
+
+	return ret;
+}
+
 static void *vfio_iommu_type1_open(unsigned long arg)
 {
 	struct vfio_iommu *iommu;
-
-	if (arg != VFIO_TYPE1_IOMMU && arg != VFIO_TYPE1v2_IOMMU)
+	int features = 0;
+
+	switch (arg) {
+	case VFIO_TYPE1_IOMMU:
+		break;
+	case VFIO_TYPE1_NESTING_IOMMU:
+		if (!vfio_iommus_have_nesting())
+			return ERR_PTR(-EOPNOTSUPP);
+		features |= VFIO_IOMMU_FEAT_NESTING;
+	case VFIO_TYPE1v2_IOMMU:
+		features |= VFIO_IOMMU_FEAT_V2;
+		break;
+	default:
 		return ERR_PTR(-EINVAL);
+	}
 
 	iommu = kzalloc(sizeof(*iommu), GFP_KERNEL);
 	if (!iommu)
@@ -829,7 +875,7 @@ static void *vfio_iommu_type1_open(unsigned long arg)
 	INIT_LIST_HEAD(&iommu->domain_list);
 	iommu->dma_list = RB_ROOT;
 	mutex_init(&iommu->lock);
-	iommu->v2 = (arg == VFIO_TYPE1v2_IOMMU);
+	iommu->features = features;
 
 	return iommu;
 }
@@ -875,6 +921,29 @@ static int vfio_domains_have_iommu_cache(struct vfio_iommu *iommu)
 	return ret;
 }
 
+static int vfio_domains_have_iommu_nesting(struct vfio_iommu *iommu)
+{
+	struct vfio_domain *domain;
+	int ret = 1;
+
+	if (!(iommu->features & VFIO_IOMMU_FEAT_NESTING))
+		return 0;
+
+	mutex_lock(&iommu->lock);
+	list_for_each_entry(domain, &iommu->domain_list, next) {
+		int attr;
+
+		if (iommu_domain_get_attr(domain->domain, DOMAIN_ATTR_NESTING,
+					   &attr) || !attr) {
+			ret = 0;
+			break;
+		}
+	}
+	mutex_unlock(&iommu->lock);
+
+	return ret;
+}
+
 static long vfio_iommu_type1_ioctl(void *iommu_data,
 				   unsigned int cmd, unsigned long arg)
 {
@@ -886,6 +955,10 @@ static long vfio_iommu_type1_ioctl(void *iommu_data,
 		case VFIO_TYPE1_IOMMU:
 		case VFIO_TYPE1v2_IOMMU:
 			return 1;
+		case VFIO_TYPE1_NESTING_IOMMU:
+			if (!iommu)
+				return vfio_iommus_have_nesting();
+			return vfio_domains_have_iommu_nesting(iommu);
 		case VFIO_DMA_CC_IOMMU:
 			if (!iommu)
 				return 0;
include/uapi/linux/vfio.h +2 -0
index 6612974c64bf..ac79f6214fd6 100644
@@ -24,11 +24,13 @@
 #define VFIO_TYPE1_IOMMU		1
 #define VFIO_SPAPR_TCE_IOMMU		2
 #define VFIO_TYPE1v2_IOMMU		3
+
 /*
  * IOMMU enforces DMA cache coherence (ex. PCIe NoSnoop stripping).  This
  * capability is subject to change as groups are added or removed.
  */
 #define VFIO_DMA_CC_IOMMU		4
+#define VFIO_TYPE1_NESTING_IOMMU	5	/* Implies v2 */
 
 /* Check if EEH is supported */
 #define VFIO_EEH			5
Signature · 1 line
2.1.0

A message that is not in the archive (sent off-list or not yet received): 1409651619-27342-3-git-send-email-will.deacon-5wv7dgnIgG8@public.gmane.org

#5 lore raw

Re: [PATCH v3 2/3] vfio/iommu_type1: add new VFIO_TYPE1_NESTING_IOMMU IOMMU type

Quoted 34 lines · Will Deacon

VFIO allows devices to be safely handed off to userspace by putting them behind an IOMMU configured to ensure DMA and interrupt isolation. This enables userspace KVM clients, such as kvmtool and qemu, to further map the device into a virtual machine.

With IOMMUs such as the ARM SMMU, it is then possible to provide SMMU translation services to the guest operating system, which are nested with the existing translation installed by VFIO. However, enabling this feature means that the IOMMU driver must be informed that the VFIO domain is being created for the purposes of nested translation.

This patch adds a new IOMMU type (VFIO_TYPE1_NESTING_IOMMU) to the VFIO type-1 driver. The new IOMMU type acts identically to the VFIO_TYPE1v2_IOMMU type, but additionally sets the DOMAIN_ATTR_NESTING attribute on its IOMMU domains. Userspace can check whether nesting is actually available using the VFIO_CHECK_EXTENSION ioctl, in a similar manner to checking for cache-coherent DMA.

Cc Joerg Roedel <[email protected]>
Cc Alex Williamson <[email protected]>
---
 drivers/vfio/vfio_iommu_type1.c | 87 +++++++++++++++++++++++++++++++++++++----
 include/uapi/linux/vfio.h       |  2 +
 2 files changed, 82 insertions(+), 7 deletions(-)
drivers/vfio/vfio_iommu_type1.c +0 -0
index 0734fbe5b651..50aa56c7b6a7 100644
@@ -53,11 +53,15 @@ module_param_named(disable_hugepages,
 MODULE_PARM_DESC(disable_hugepages,
 		 "Disable VFIO IOMMU support for IOMMU hugepages.");
 

drivers/vfio/vfio_iommu_type1.c module_param_named(disable_hugepages, · line 64

Quoted 10 lines
drivers/vfio/vfio_iommu_type1.c +5 -1
+/* Feature flags for VFIO Type-1 IOMMUs */
+#define VFIO_IOMMU_FEAT_V2	(1 << 0)
+#define VFIO_IOMMU_FEAT_NESTING	(1 << 1)
+
 struct vfio_iommu {
 	struct list_head	domain_list;
 	struct mutex		lock;
 	struct rb_root		dma_list;
-	bool v2;
+	int			features;

Personally I'd rather add a nesting bool than mask out flag bits.

Quoted 32 lines
drivers/vfio/vfio_iommu_type1.c +4 -3
 };
 
 struct vfio_domain {
@@ -441,7 +445,7 @@ static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
 	 * will only return success and a size of zero if there were no
 	 * mappings within the range.
 	 */
-	if (iommu->v2) {
+	if (iommu->features & VFIO_IOMMU_FEAT_V2) {
 		dma = vfio_find_dma(iommu, unmap->iova, 0);
 		if (dma && dma->iova != unmap->iova) {
 			ret = -EINVAL;
@@ -455,7 +459,8 @@ static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
 	}
 
 	while ((dma = vfio_find_dma(iommu, unmap->iova, unmap->size))) {
-		if (!iommu->v2 && unmap->iova > dma->iova)
+		if (!(iommu->features & VFIO_IOMMU_FEAT_V2) &&
+		    unmap->iova > dma->iova)
 			break;
 		unmapped += dma->size;
 		vfio_remove_dma(iommu, dma);
@@ -671,7 +676,7 @@ static int vfio_iommu_type1_attach_group(void *iommu_data,
 	struct vfio_group *group, *g;
 	struct vfio_domain *domain, *d;
 	struct bus_type *bus = NULL;
-	int ret;
+	int ret, attr = 1;
 
 	mutex_lock(&iommu->lock);
 

drivers/vfio/vfio_iommu_type1.c static int vfio_iommu_type1_attach_group(void *iommu_data, · line 715

Quoted 6 lines
drivers/vfio/vfio_iommu_type1.c +3 -0
@@ -705,6 +710,9 @@ static int vfio_iommu_type1_attach_group(void *iommu_data,
 		goto out_free;
 	}
 
+	if (iommu->features & VFIO_IOMMU_FEAT_NESTING)
+		iommu_domain_set_attr(domain->domain, DOMAIN_ATTR_NESTING, &attr);
+

Shouldn't we fail if this doesn't work?

drivers/vfio/vfio_iommu_type1.c done: · line 828

Quoted 10 lines
drivers/vfio/vfio_iommu_type1.c +3 -0
 	ret = iommu_attach_group(domain->domain, iommu_group);
 	if (ret)
 		goto out_domain;
@@ -815,12 +823,50 @@ done:
 	mutex_unlock(&iommu->lock);
 }
 
+static int vfio_iommus_have_nesting(void)
+{
+	static struct bus_type *bus_types[] = { &pci_bus_type };

Hmm, we just got rid of all traces of pci in this file, so I'm not too excited about adding it back. Don't we no longer include pci.h? I'm surprised this doesn't throw a warning. We should try to be independent of CONFIG_PCI here.

Quoted 9 lines
drivers/vfio/vfio_iommu_type1.c +9 -0
+	int i, ret = 0;
+
+	for (i = 0; !ret && i < ARRAY_SIZE(bus_types); ++i) {
+		struct iommu_domain *domain;
+		struct bus_type *bus = bus_types[i];
+		int attr = 1;
+
+		if (!iommu_present(bus))
+			continue;

drivers/vfio/vfio_iommu_type1.c done: · line 847

Quoted 10 lines
drivers/vfio/vfio_iommu_type1.c +10 -0
+
+		domain = iommu_domain_alloc(bus);
+		if (!domain)
+			continue;
+
+		if (!iommu_domain_set_attr(domain, DOMAIN_ATTR_NESTING, &attr))
+			ret = 1;
+
+		iommu_domain_free(domain);
+	}

Can this support Joerg's iommu_capable() interface so we can test w/o creating a domain? We still have the bus_type problem though.

Quoted 67 lines
drivers/vfio/vfio_iommu_type1.c +42 -3
+
+	return ret;
+}
+
 static void *vfio_iommu_type1_open(unsigned long arg)
 {
 	struct vfio_iommu *iommu;
-
-	if (arg != VFIO_TYPE1_IOMMU && arg != VFIO_TYPE1v2_IOMMU)
+	int features = 0;
+
+	switch (arg) {
+	case VFIO_TYPE1_IOMMU:
+		break;
+	case VFIO_TYPE1_NESTING_IOMMU:
+		if (!vfio_iommus_have_nesting())
+			return ERR_PTR(-EOPNOTSUPP);
+		features |= VFIO_IOMMU_FEAT_NESTING;
+	case VFIO_TYPE1v2_IOMMU:
+		features |= VFIO_IOMMU_FEAT_V2;
+		break;
+	default:
 		return ERR_PTR(-EINVAL);
+	}
 
 	iommu = kzalloc(sizeof(*iommu), GFP_KERNEL);
 	if (!iommu)
@@ -829,7 +875,7 @@ static void *vfio_iommu_type1_open(unsigned long arg)
 	INIT_LIST_HEAD(&iommu->domain_list);
 	iommu->dma_list = RB_ROOT;
 	mutex_init(&iommu->lock);
-	iommu->v2 = (arg == VFIO_TYPE1v2_IOMMU);
+	iommu->features = features;
 
 	return iommu;
 }
@@ -875,6 +921,29 @@ static int vfio_domains_have_iommu_cache(struct vfio_iommu *iommu)
 	return ret;
 }
 
+static int vfio_domains_have_iommu_nesting(struct vfio_iommu *iommu)
+{
+	struct vfio_domain *domain;
+	int ret = 1;
+
+	if (!(iommu->features & VFIO_IOMMU_FEAT_NESTING))
+		return 0;
+
+	mutex_lock(&iommu->lock);
+	list_for_each_entry(domain, &iommu->domain_list, next) {
+		int attr;
+
+		if (iommu_domain_get_attr(domain->domain, DOMAIN_ATTR_NESTING,
+					   &attr) || !attr) {
+			ret = 0;
+			break;
+		}
+	}
+	mutex_unlock(&iommu->lock);
+
+	return ret;
+}
+
 static long vfio_iommu_type1_ioctl(void *iommu_data,
 				   unsigned int cmd, unsigned long arg)
 {

drivers/vfio/vfio_iommu_type1.c static long vfio_iommu_type1_ioctl(void *iommu_data, · line 961

Quoted 7 lines
drivers/vfio/vfio_iommu_type1.c +4 -0
@@ -886,6 +955,10 @@ static long vfio_iommu_type1_ioctl(void *iommu_data,
 		case VFIO_TYPE1_IOMMU:
 		case VFIO_TYPE1v2_IOMMU:
 			return 1;
+		case VFIO_TYPE1_NESTING_IOMMU:
+			if (!iommu)
+				return vfio_iommus_have_nesting();
+			return vfio_domains_have_iommu_nesting(iommu);

I'd just about prefer to statically return 1 for NESTING and then fail to add groups if we can't set the attribute. It sucks, but here we test an arbitrary bus type to determine whether nesting is supported, allow a nesting iommu type to be set regardless of what kind of devices are about to go into it, and require that the user test to see whether the requested IOMMU attribute actually stuck. That's not a very intuitive interface. Thanks,

Alex

Quoted 21 lines
drivers/vfio/vfio_iommu_type1.c +0 -0
 		case VFIO_DMA_CC_IOMMU:
 			if (!iommu)
 				return 0;
include/uapi/linux/vfio.h +2 -0
index 6612974c64bf..ac79f6214fd6 100644
@@ -24,11 +24,13 @@
 #define VFIO_TYPE1_IOMMU		1
 #define VFIO_SPAPR_TCE_IOMMU		2
 #define VFIO_TYPE1v2_IOMMU		3
+
 /*
  * IOMMU enforces DMA cache coherence (ex. PCIe NoSnoop stripping).  This
  * capability is subject to change as groups are added or removed.
  */
 #define VFIO_DMA_CC_IOMMU		4
+#define VFIO_TYPE1_NESTING_IOMMU	5	/* Implies v2 */
 
 /* Check if EEH is supported */
 #define VFIO_EEH			5

A message that is not in the archive (sent off-list or not yet received): 1410462748.2982.342.camel-85EaTFmN5p//[email protected]

#7 lore raw

Re: [PATCH v3 2/3] vfio/iommu_type1: add new VFIO_TYPE1_NESTING_IOMMU IOMMU type

Hi Alex,

Thanks for taking a look.

Quoted 9 lines · Alex Williamson
Quoted 8 lines · Will Deacon
drivers/vfio/vfio_iommu_type1.c +0 -0
index 0734fbe5b651..50aa56c7b6a7 100644
@@ -53,11 +53,15 @@ module_param_named(disable_hugepages,
 MODULE_PARM_DESC(disable_hugepages,
 		 "Disable VFIO IOMMU support for IOMMU hugepages.");
 
Quoted 12 lines

drivers/vfio/vfio_iommu_type1.c module_param_named(disable_hugepages, · line 64

Quoted 10 lines
drivers/vfio/vfio_iommu_type1.c +5 -1
+/* Feature flags for VFIO Type-1 IOMMUs */
+#define VFIO_IOMMU_FEAT_V2	(1 << 0)
+#define VFIO_IOMMU_FEAT_NESTING	(1 << 1)
+
 struct vfio_iommu {
 	struct list_head	domain_list;
 	struct mutex		lock;
 	struct rb_root		dma_list;
-	bool v2;
+	int			features;

Personally I'd rather add a nesting bool than mask out flag bits.

Sure.

Quoted 9 lines

drivers/vfio/vfio_iommu_type1.c static int vfio_iommu_type1_attach_group(void *iommu_data, · line 715

Quoted 7 lines
drivers/vfio/vfio_iommu_type1.c +3 -0
@@ -705,6 +710,9 @@ static int vfio_iommu_type1_attach_group(void *iommu_data,
 		goto out_free;
 	}
 
+	if (iommu->features & VFIO_IOMMU_FEAT_NESTING)
+		iommu_domain_set_attr(domain->domain, DOMAIN_ATTR_NESTING, &attr);
+

Shouldn't we fail if this doesn't work?

The reason I don't fail is because device passthrough can still work with an IOMMU that isn't capable of nesting. The part that won't work is attempting to instantiate a virtual SMMU interface for the guest.

However, it just struck me that userspace knows whether or not it's going to want a virtual SMMU interface. If it does, then failure to set NESTING is fatal. If it doesn't, then it doesn't need to set the flag so there is no issue.

Quoted 7 lines
Quoted 7 lines
drivers/vfio/vfio_iommu_type1.c +0 -0
 	ret = iommu_attach_group(domain->domain, iommu_group);
 	if (ret)
 		goto out_domain;
@@ -815,12 +823,50 @@ done:
 	mutex_unlock(&iommu->lock);
 }
 
Quoted 8 lines

drivers/vfio/vfio_iommu_type1.c done: · line 828

Quoted 3 lines
drivers/vfio/vfio_iommu_type1.c +3 -0
+static int vfio_iommus_have_nesting(void)
+{
+	static struct bus_type *bus_types[] = { &pci_bus_type };

Hmm, we just got rid of all traces of pci in this file, so I'm not too excited about adding it back. Don't we no longer include pci.h? I'm surprised this doesn't throw a warning. We should try to be independent of CONFIG_PCI here.

I don't see any warning here.

Quoted 15 lines

drivers/vfio/vfio_iommu_type1.c done: · line 847

Quoted 12 lines
drivers/vfio/vfio_iommu_type1.c +12 -0
+		if (!iommu_present(bus))
+			continue;
+
+		domain = iommu_domain_alloc(bus);
+		if (!domain)
+			continue;
+
+		if (!iommu_domain_set_attr(domain, DOMAIN_ATTR_NESTING, &attr))
+			ret = 1;
+
+		iommu_domain_free(domain);
+	}

Can this support Joerg's iommu_capable() interface so we can test w/o creating a domain? We still have the bus_type problem though.

[...]

Quoted 9 lines

drivers/vfio/vfio_iommu_type1.c static long vfio_iommu_type1_ioctl(void *iommu_data, · line 961

Quoted 8 lines
drivers/vfio/vfio_iommu_type1.c +4 -0
@@ -886,6 +955,10 @@ static long vfio_iommu_type1_ioctl(void *iommu_data,
 		case VFIO_TYPE1_IOMMU:
 		case VFIO_TYPE1v2_IOMMU:
 			return 1;
+		case VFIO_TYPE1_NESTING_IOMMU:
+			if (!iommu)
+				return vfio_iommus_have_nesting();
+			return vfio_domains_have_iommu_nesting(iommu);
Quoted 7 lines

I'd just about prefer to statically return 1 for NESTING and then fail to add groups if we can't set the attribute. It sucks, but here we test an arbitrary bus type to determine whether nesting is supported, allow a nesting iommu type to be set regardless of what kind of devices are about to go into it, and require that the user test to see whether the requested IOMMU attribute actually stuck. That's not a very intuitive interface. Thanks,

Yes, I totally agree that the interface is clunky. Returning 1 for NESTING solves both the pci bus problem and the iommu_capable issue, so that sounds like a great simplification of the code.

That means userspace:

- Can always open a VFIO_TYPE1_NESTING_IOMMU
- VFIO_CHECK_EXTENSION will always return 1
- When adding groups (e.g. SET_IOMMU), the iommu_domain_set_attr can fail, which will cause the corresponding ioctl to fail

If nesting is not required because there is no virtual SMMU interface, then the usual TYPE1 IOMMU can be used.

Sound good?

Will

A message that is not in the archive (sent off-list or not yet received): [email protected]

#9 lore raw

Re: [PATCH v3 2/3] vfio/iommu_type1: add new VFIO_TYPE1_NESTING_IOMMU IOMMU type

Quoted 112 lines · Will Deacon

Hi Alex,

Thanks for taking a look.

Quoted 9 lines · Alex Williamson
Quoted 8 lines · Will Deacon
drivers/vfio/vfio_iommu_type1.c +0 -0
index 0734fbe5b651..50aa56c7b6a7 100644
@@ -53,11 +53,15 @@ module_param_named(disable_hugepages,
 MODULE_PARM_DESC(disable_hugepages,
 		 "Disable VFIO IOMMU support for IOMMU hugepages.");
 
Quoted 12 lines

drivers/vfio/vfio_iommu_type1.c module_param_named(disable_hugepages, · line 64

Quoted 10 lines
drivers/vfio/vfio_iommu_type1.c +5 -1
+/* Feature flags for VFIO Type-1 IOMMUs */
+#define VFIO_IOMMU_FEAT_V2	(1 << 0)
+#define VFIO_IOMMU_FEAT_NESTING	(1 << 1)
+
 struct vfio_iommu {
 	struct list_head	domain_list;
 	struct mutex		lock;
 	struct rb_root		dma_list;
-	bool v2;
+	int			features;

Personally I'd rather add a nesting bool than mask out flag bits.

Sure.

Quoted 9 lines

drivers/vfio/vfio_iommu_type1.c static int vfio_iommu_type1_attach_group(void *iommu_data, · line 715

Quoted 7 lines
drivers/vfio/vfio_iommu_type1.c +3 -0
@@ -705,6 +710,9 @@ static int vfio_iommu_type1_attach_group(void *iommu_data,
 		goto out_free;
 	}
 
+	if (iommu->features & VFIO_IOMMU_FEAT_NESTING)
+		iommu_domain_set_attr(domain->domain, DOMAIN_ATTR_NESTING, &attr);
+

Shouldn't we fail if this doesn't work?

The reason I don't fail is because device passthrough can still work with an IOMMU that isn't capable of nesting. The part that won't work is attempting to instantiate a virtual SMMU interface for the guest.

However, it just struck me that userspace knows whether or not it's going to want a virtual SMMU interface. If it does, then failure to set NESTING is fatal. If it doesn't, then it doesn't need to set the flag so there is no issue.

Quoted 7 lines
Quoted 7 lines
drivers/vfio/vfio_iommu_type1.c +0 -0
 	ret = iommu_attach_group(domain->domain, iommu_group);
 	if (ret)
 		goto out_domain;
@@ -815,12 +823,50 @@ done:
 	mutex_unlock(&iommu->lock);
 }
 
Quoted 8 lines

drivers/vfio/vfio_iommu_type1.c done: · line 828

Quoted 3 lines
drivers/vfio/vfio_iommu_type1.c +3 -0
+static int vfio_iommus_have_nesting(void)
+{
+	static struct bus_type *bus_types[] = { &pci_bus_type };

Hmm, we just got rid of all traces of pci in this file, so I'm not too excited about adding it back. Don't we no longer include pci.h? I'm surprised this doesn't throw a warning. We should try to be independent of CONFIG_PCI here.

I don't see any warning here.

Quoted 15 lines

drivers/vfio/vfio_iommu_type1.c done: · line 847

Quoted 12 lines
drivers/vfio/vfio_iommu_type1.c +12 -0
+		if (!iommu_present(bus))
+			continue;
+
+		domain = iommu_domain_alloc(bus);
+		if (!domain)
+			continue;
+
+		if (!iommu_domain_set_attr(domain, DOMAIN_ATTR_NESTING, &attr))
+			ret = 1;
+
+		iommu_domain_free(domain);
+	}

Can this support Joerg's iommu_capable() interface so we can test w/o creating a domain? We still have the bus_type problem though.

[...]

Quoted 9 lines

drivers/vfio/vfio_iommu_type1.c static long vfio_iommu_type1_ioctl(void *iommu_data, · line 961

Quoted 8 lines
drivers/vfio/vfio_iommu_type1.c +4 -0
@@ -886,6 +955,10 @@ static long vfio_iommu_type1_ioctl(void *iommu_data,
 		case VFIO_TYPE1_IOMMU:
 		case VFIO_TYPE1v2_IOMMU:
 			return 1;
+		case VFIO_TYPE1_NESTING_IOMMU:
+			if (!iommu)
+				return vfio_iommus_have_nesting();
+			return vfio_domains_have_iommu_nesting(iommu);
Quoted 7 lines

I'd just about prefer to statically return 1 for NESTING and then fail to add groups if we can't set the attribute. It sucks, but here we test an arbitrary bus type to determine whether nesting is supported, allow a nesting iommu type to be set regardless of what kind of devices are about to go into it, and require that the user test to see whether the requested IOMMU attribute actually stuck. That's not a very intuitive interface. Thanks,

Yes, I totally agree that the interface is clunky. Returning 1 for NESTING solves both the pci bus problem and the iommu_capable issue, so that sounds like a great simplification of the code.

That means userspace:

- Can always open a VFIO_TYPE1_NESTING_IOMMU
- VFIO_CHECK_EXTENSION will always return 1
- When adding groups (e.g. SET_IOMMU), the iommu_domain_set_attr can fail, which will cause the corresponding ioctl to fail

Quoted 4 lines

If nesting is not required because there is no virtual SMMU interface, then the usual TYPE1 IOMMU can be used.

Sound good?

It's not 100% satisfying, but I can't think of anything better. VFIO_TYPE1_NESTING_IOMMU is then just a VFIO level indication of support with no indication of underlying hardware support. The user doesn't know until they add a group if it's really possible. Not ideal from a user perspective, but VFIO has always left itself the option of disallowing groups from being added to a container. Thanks,

Alex

#10 lore raw

PATCH v3 3/3 iommu/arm-smmu: add support for DOMAIN_ATTR_NESTING attribute

In reply to a message not in the archive

When domains are set with the DOMAIN_ATTR_NESTING flag, we must ensure that we allocate them to stage-2 context banks if the hardware permits it.

This patch adds support for the attribute to the ARM SMMU driver, with the actual stage being determined depending on the features supported by the hardware.

---
 drivers/iommu/arm-smmu.c | 110 ++++++++++++++++++++++++++++++++++++++---------
 1 file changed, 90 insertions(+), 20 deletions(-)
drivers/iommu/arm-smmu.c +90 -20
index d2d8cdaf4f0b..c745296b170f 100644
@@ -404,9 +404,16 @@ struct arm_smmu_cfg {
 #define ARM_SMMU_CB_ASID(cfg)		((cfg)->cbndx)
 #define ARM_SMMU_CB_VMID(cfg)		((cfg)->cbndx + 1)
 
+enum arm_smmu_domain_stage {
+	ARM_SMMU_DOMAIN_S1 = 0,
+	ARM_SMMU_DOMAIN_S2,
+	ARM_SMMU_DOMAIN_NESTED,
+};
+
 struct arm_smmu_domain {
 	struct arm_smmu_device		*smmu;
 	struct arm_smmu_cfg		cfg;
+	enum arm_smmu_domain_stage	stage;
 	spinlock_t			lock;
 };
 
@@ -899,19 +906,46 @@ static int arm_smmu_init_domain_context(struct iommu_domain *domain,
 	if (smmu_domain->smmu)
 		goto out_unlock;
 
-	if (smmu->features & ARM_SMMU_FEAT_TRANS_NESTED) {
+	/*
+	 * Mapping the requested stage onto what we support is surprisingly
+	 * complicated, mainly because the spec allows S1+S2 SMMUs without
+	 * support for nested translation. That means we end up with the
+	 * following table:
+	 *
+	 * Requested        Supported        Actual
+	 *     S1               N              S1
+	 *     S1             S1+S2            S1
+	 *     S1               S2             S2
+	 *     S1               S1             S1
+	 *     N                N              N
+	 *     N              S1+S2            S2
+	 *     N                S2             S2
+	 *     N                S1             S1
+	 *
+	 * Note that you can't actually request stage-2 mappings.
+	 */
+	if (!(smmu->features & ARM_SMMU_FEAT_TRANS_S1))
+		smmu_domain->stage = ARM_SMMU_DOMAIN_S2;
+	if (!(smmu->features & ARM_SMMU_FEAT_TRANS_S2))
+		smmu_domain->stage = ARM_SMMU_DOMAIN_S1;
+
+	switch (smmu_domain->stage) {
+	case ARM_SMMU_DOMAIN_S1:
+		cfg->cbar = CBAR_TYPE_S1_TRANS_S2_BYPASS;
+		start = smmu->num_s2_context_banks;
+		break;
+	case ARM_SMMU_DOMAIN_NESTED:
 		/*
 		 * We will likely want to change this if/when KVM gets
 		 * involved.
 		 */
-		cfg->cbar = CBAR_TYPE_S1_TRANS_S2_BYPASS;
-		start = smmu->num_s2_context_banks;
-	} else if (smmu->features & ARM_SMMU_FEAT_TRANS_S1) {
-		cfg->cbar = CBAR_TYPE_S1_TRANS_S2_BYPASS;
-		start = smmu->num_s2_context_banks;
-	} else {
+	case ARM_SMMU_DOMAIN_S2:
 		cfg->cbar = CBAR_TYPE_S2_TRANS;
 		start = 0;
+		break;
+	default:
+		ret = -EINVAL;
+		goto out_unlock;
 	}
 
 	ret = __arm_smmu_alloc_bitmap(smmu->context_map, start,
@@ -1637,20 +1671,56 @@ static void arm_smmu_remove_device(struct device *dev)
 	iommu_group_remove_device(dev);
 }
 
+static int arm_smmu_domain_get_attr(struct iommu_domain *domain,
+				    enum iommu_attr attr, void *data)
+{
+	struct arm_smmu_domain *smmu_domain = domain->priv;
+
+	switch (attr) {
+	case DOMAIN_ATTR_NESTING:
+		*(int *)data = (smmu_domain->stage == ARM_SMMU_DOMAIN_NESTED);
+		return 0;
+	default:
+		return -ENODEV;
+	}
+}
+
+static int arm_smmu_domain_set_attr(struct iommu_domain *domain,
+				    enum iommu_attr attr, void *data)
+{
+	struct arm_smmu_domain *smmu_domain = domain->priv;
+
+	switch (attr) {
+	case DOMAIN_ATTR_NESTING:
+		if (smmu_domain->smmu)
+			return -EPERM;
+		if (*(int *)data)
+			smmu_domain->stage = ARM_SMMU_DOMAIN_NESTED;
+		else
+			smmu_domain->stage = ARM_SMMU_DOMAIN_S1;
+
+		return 0;
+	default:
+		return -ENODEV;
+	}
+}
+
 static const struct iommu_ops arm_smmu_ops = {
-	.domain_init	= arm_smmu_domain_init,
-	.domain_destroy	= arm_smmu_domain_destroy,
-	.attach_dev	= arm_smmu_attach_dev,
-	.detach_dev	= arm_smmu_detach_dev,
-	.map		= arm_smmu_map,
-	.unmap		= arm_smmu_unmap,
-	.iova_to_phys	= arm_smmu_iova_to_phys,
-	.domain_has_cap	= arm_smmu_domain_has_cap,
-	.add_device	= arm_smmu_add_device,
-	.remove_device	= arm_smmu_remove_device,
-	.pgsize_bitmap	= (SECTION_SIZE |
-			   ARM_SMMU_PTE_CONT_SIZE |
-			   PAGE_SIZE),
+	.domain_init		= arm_smmu_domain_init,
+	.domain_destroy		= arm_smmu_domain_destroy,
+	.attach_dev		= arm_smmu_attach_dev,
+	.detach_dev		= arm_smmu_detach_dev,
+	.map			= arm_smmu_map,
+	.unmap			= arm_smmu_unmap,
+	.iova_to_phys		= arm_smmu_iova_to_phys,
+	.domain_has_cap		= arm_smmu_domain_has_cap,
+	.add_device		= arm_smmu_add_device,
+	.remove_device		= arm_smmu_remove_device,
+	.domain_get_attr	= arm_smmu_domain_get_attr,
+	.domain_set_attr	= arm_smmu_domain_set_attr,
+	.pgsize_bitmap		= (SECTION_SIZE |
+				   ARM_SMMU_PTE_CONT_SIZE |
+				   PAGE_SIZE),
 };
 
 static void arm_smmu_device_reset(struct arm_smmu_device *smmu)
Signature · 1 line
2.1.0

A message that is not in the archive (sent off-list or not yet received): 1409651619-27342-4-git-send-email-will.deacon-5wv7dgnIgG8@public.gmane.org

#12 lore raw

RE: [PATCH v3 3/3] iommu/arm-smmu: add support for DOMAIN_ATTR_NESTING attribute

Hi Will,
We would still need a mechanism to distinguish stage1 mapping from stage2 mapping i.e. for nested translation we should be able to specify whether the mapping corresponds to stage1 or stage 2 translation. Also, correspondingly we would require different context banks for stage 1 and stage 2 translation.

Regards
Varun

Quoted 176 lines

-----Original Message-----
From: iommu-bounces-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA@public.gmane.org [mailto:iommu- bounces-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA@public.gmane.org] On Behalf Of Will Deacon
Sent: Tuesday, September 02, 2014 3:24 PM
To: iommu-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA@public.gmane.org
Cc: Will Deacon
Subject: [PATCH v3 3/3] iommu/arm-smmu: add support for
DOMAIN_ATTR_NESTING attribute

When domains are set with the DOMAIN_ATTR_NESTING flag, we must ensure that we allocate them to stage-2 context banks if the hardware permits it.

This patch adds support for the attribute to the ARM SMMU driver, with the actual stage being determined depending on the features supported by the hardware.

---
 drivers/iommu/arm-smmu.c | 110
++++++++++++++++++++++++++++++++++++++---------
 1 file changed, 90 insertions(+), 20 deletions(-)
drivers/iommu/arm-smmu.c index renamed from drivers/iommu/arm-smmu.c +0 -0

d2d8cdaf4f0b..c745296b170f 100644

drivers/iommu/arm-smmu.c +41 -7
@@ -404,9 +404,16 @@ struct arm_smmu_cfg {
 #define ARM_SMMU_CB_ASID(cfg)		((cfg)->cbndx)
 #define ARM_SMMU_CB_VMID(cfg)		((cfg)->cbndx + 1)
 
+enum arm_smmu_domain_stage {
+	ARM_SMMU_DOMAIN_S1 = 0,
+	ARM_SMMU_DOMAIN_S2,
+	ARM_SMMU_DOMAIN_NESTED,
+};
+
 struct arm_smmu_domain {
 	struct arm_smmu_device		*smmu;
 	struct arm_smmu_cfg		cfg;
+	enum arm_smmu_domain_stage	stage;
 	spinlock_t			lock;
 };
 
@@ -899,19 +906,46 @@ static int arm_smmu_init_domain_context(struct
 iommu_domain *domain,
 	if (smmu_domain->smmu)
 		goto out_unlock;
 
-	if (smmu->features & ARM_SMMU_FEAT_TRANS_NESTED) {
+	/*
+	 * Mapping the requested stage onto what we support is surprisingly
+	 * complicated, mainly because the spec allows S1+S2 SMMUs without
+	 * support for nested translation. That means we end up with the
+	 * following table:
+	 *
+	 * Requested        Supported        Actual
+	 *     S1               N              S1
+	 *     S1             S1+S2            S1
+	 *     S1               S2             S2
+	 *     S1               S1             S1
+	 *     N                N              N
+	 *     N              S1+S2            S2
+	 *     N                S2             S2
+	 *     N                S1             S1
+	 *
+	 * Note that you can't actually request stage-2 mappings.
+	 */
+	if (!(smmu->features & ARM_SMMU_FEAT_TRANS_S1))
+		smmu_domain->stage = ARM_SMMU_DOMAIN_S2;
+	if (!(smmu->features & ARM_SMMU_FEAT_TRANS_S2))
+		smmu_domain->stage = ARM_SMMU_DOMAIN_S1;
+
+	switch (smmu_domain->stage) {
+	case ARM_SMMU_DOMAIN_S1:
+		cfg->cbar = CBAR_TYPE_S1_TRANS_S2_BYPASS;
+		start = smmu->num_s2_context_banks;
+		break;
+	case ARM_SMMU_DOMAIN_NESTED:
 		/*
 		 * We will likely want to change this if/when KVM gets
 		 * involved.
 		 */
-		cfg->cbar = CBAR_TYPE_S1_TRANS_S2_BYPASS;
-		start = smmu->num_s2_context_banks;
-	} else if (smmu->features & ARM_SMMU_FEAT_TRANS_S1) {
-		cfg->cbar = CBAR_TYPE_S1_TRANS_S2_BYPASS;
-		start = smmu->num_s2_context_banks;
-	} else {
+	case ARM_SMMU_DOMAIN_S2:
 		cfg->cbar = CBAR_TYPE_S2_TRANS;
 		start = 0;
+		break;
+	default:
+		ret = -EINVAL;
+		goto out_unlock;
 	}
 

ret = __arm_smmu_alloc_bitmap(smmu->context_map, start, @@ -
1637,20 +1671,56 @@ static void arm_smmu_remove_device(struct device *dev)

drivers/iommu/arm-smmu.c +7 -0
 	iommu_group_remove_device(dev);
 }
 
+static int arm_smmu_domain_get_attr(struct iommu_domain *domain,
+				    enum iommu_attr attr, void *data) {
+	struct arm_smmu_domain *smmu_domain = domain->priv;
+
+	switch (attr) {
+	case DOMAIN_ATTR_NESTING:
+		*(int *)data = (smmu_domain->stage ==

ARM_SMMU_DOMAIN_NESTED);

drivers/iommu/arm-smmu.c +16 -0
+		return 0;
+	default:
+		return -ENODEV;
+	}
+}
+
+static int arm_smmu_domain_set_attr(struct iommu_domain *domain,
+				    enum iommu_attr attr, void *data) {
+	struct arm_smmu_domain *smmu_domain = domain->priv;
+
+	switch (attr) {
+	case DOMAIN_ATTR_NESTING:
+		if (smmu_domain->smmu)
+			return -EPERM;
+		if (*(int *)data)
+			smmu_domain->stage =

ARM_SMMU_DOMAIN_NESTED;

drivers/iommu/arm-smmu.c +24 -14
+		else
+			smmu_domain->stage = ARM_SMMU_DOMAIN_S1;
+
+		return 0;
+	default:
+		return -ENODEV;
+	}
+}
+
 static const struct iommu_ops arm_smmu_ops = {
-	.domain_init	= arm_smmu_domain_init,
-	.domain_destroy	= arm_smmu_domain_destroy,
-	.attach_dev	= arm_smmu_attach_dev,
-	.detach_dev	= arm_smmu_detach_dev,
-	.map		= arm_smmu_map,
-	.unmap		= arm_smmu_unmap,
-	.iova_to_phys	= arm_smmu_iova_to_phys,
-	.domain_has_cap	= arm_smmu_domain_has_cap,
-	.add_device	= arm_smmu_add_device,
-	.remove_device	= arm_smmu_remove_device,
-	.pgsize_bitmap	= (SECTION_SIZE |
-			   ARM_SMMU_PTE_CONT_SIZE |
-			   PAGE_SIZE),
+	.domain_init		= arm_smmu_domain_init,
+	.domain_destroy		= arm_smmu_domain_destroy,
+	.attach_dev		= arm_smmu_attach_dev,
+	.detach_dev		= arm_smmu_detach_dev,
+	.map			= arm_smmu_map,
+	.unmap			= arm_smmu_unmap,
+	.iova_to_phys		= arm_smmu_iova_to_phys,
+	.domain_has_cap		= arm_smmu_domain_has_cap,
+	.add_device		= arm_smmu_add_device,
+	.remove_device		= arm_smmu_remove_device,
+	.domain_get_attr	= arm_smmu_domain_get_attr,
+	.domain_set_attr	= arm_smmu_domain_set_attr,
+	.pgsize_bitmap		= (SECTION_SIZE |
+				   ARM_SMMU_PTE_CONT_SIZE |
+				   PAGE_SIZE),
 };
 
 static void arm_smmu_device_reset(struct arm_smmu_device *smmu)
--

2.1.0

_______________________________________________
iommu mailing list iommu-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA@public.gmane.org
https://lists.linuxfoundation.org/mailman/listinfo/iommu

A message that is not in the archive (sent off-list or not yet received): 6c60062d962340bbaeccb07c82eaae73-AZ66ij2kwaacCcN9WK45f+O6mTEJWrR4XA4E9RH9d+qIuWR1G4zioA@public.gmane.org

#14 lore raw

Re: [PATCH v3 3/3] iommu/arm-smmu: add support for DOMAIN_ATTR_NESTING attribute

Quoted 5 lines · Varun Sethi

We would still need a mechanism to distinguish stage1 mapping from stage2 mapping i.e. for nested translation we should be able to specify whether the mapping corresponds to stage1 or stage 2 translation. Also, correspondingly we would require different context banks for stage 1 and stage 2 translation.

Please take a look at my iommu/pci branch where I have much more support for nested translation with the ARM SMMU driver. This series is purely about getting the stage-2 part configured -- you need a virtual SMMU interface for the stage-1 component.

Will

#15 lore raw
In reply to a message not in the archive

Hi Will,

Quoted 7 lines · Will Deacon

This is version three of the patches I originally posted here:

RFCv1: http://permalink.gmane.org/gmane.linux.kernel.iommu/5552
RFCv2: http://permalink.gmane.org/gmane.linux.kernel.iommu/5700

Changes since RFCv2 include:

Quoted 7 lines

- Dropped the RFC tag
- Rebased onto 3.17-rc*
- Iterate the support bus_types (currently just PCI) and check that nesting is actually supported

The corresponding arm-smmu changes are included to show how the new domain attribute can be used.

How is this implemented in the SMMU, can there be only one stage2 mapping, so that a device can be passed through to a KVM guest or can there be multiple stage2 mappings in parallel (like with PRI/PASID on PCI) to support multiple translation contexts on one device?

Joerg

A message that is not in the archive (sent off-list or not yet received): [email protected]

#17 lore raw
Quoted 1 line · Joerg Roedel

Hi Will,

Hi Joerg,

Quoted 13 lines · Will Deacon
Quoted 12 lines

This is version three of the patches I originally posted here:

RFCv1: http://permalink.gmane.org/gmane.linux.kernel.iommu/5552
RFCv2: http://permalink.gmane.org/gmane.linux.kernel.iommu/5700

Changes since RFCv2 include:

- Dropped the RFC tag
- Rebased onto 3.17-rc*
- Iterate the support bus_types (currently just PCI) and check that nesting is actually supported

Quoted 7 lines
Quoted 2 lines

The corresponding arm-smmu changes are included to show how the new domain attribute can be used.

How is this implemented in the SMMU, can there be only one stage2 mapping, so that a device can be passed through to a KVM guest or can there be multiple stage2 mappings in parallel (like with PRI/PASID on PCI) to support multiple translation contexts on one device?

The restriction is that a stage-1 context (i.e. set of guest page tables) can be associated with at most 1 stage-2 context. If a device is capable of emitting multiple Stream IDs and the guest assigns those to different stage-1 contexts, then I guess we could use different stage-2 contexts for them, but it's not something I'd considered.

Note that the stage-2 context is used to walk the stage-1 page tables, so you always needs the guest memory to be mapped there.

Will

#18 lore raw
In reply to a message not in the archive

Hi Alex,

Quoted 14 lines · Will Deacon

Hello,

This is version three of the patches I originally posted here:

RFCv1: http://permalink.gmane.org/gmane.linux.kernel.iommu/5552
RFCv2: http://permalink.gmane.org/gmane.linux.kernel.iommu/5700

Changes since RFCv2 include:

- Dropped the RFC tag
- Rebased onto 3.17-rc*
- Iterate the support bus_types (currently just PCI) and check that nesting is actually supported

Quoted 4 lines

The corresponding arm-smmu changes are included to show how the new domain attribute can be used.

All feedback welcome,

Do you have any comments on this series? I've got a fair amount of work building on top of it, so it would good to either fix it or see about getting it merged.

Cheers,

Will