Skip to contentSkip to conversation

vfio iommu: Add dma limit capability

PATCH Superseded Matthew Rosato 2 participants 5 messages Sep 11, 2020

Subsystems (2)

v1 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 vfio iommu: Add dma limit capability

Commit 492855939bdb ("vfio/type1: Limit DMA mappings per container") added a limit to the number of concurrent DMA requests for a vfio container. However, lazy unmapping in s390 can in fact cause quite a large number of outstanding DMA requests to build up prior to being purged, potentially the entire guest DMA space. This results in unexpected 'VFIO_MAP_DMA
failed: No space left on device' conditions seen in QEMU.

This patch proposes to provide the DMA limit via the VFIO_IOMMU_GET_INFO ioctl as a new capability. A subsequent patchset to QEMU would collect this information and use it in s390 PCI support to tap the guest on the shoulder before overrunning the vfio limit.

Matthew Rosato (1):
vfio iommu: Add dma limit capability

 drivers/vfio/vfio_iommu_type1.c | 17 +++++++++++++++++
 include/uapi/linux/vfio.h       | 16 ++++++++++++++++
 2 files changed, 33 insertions(+)
Signature · 1 line
1.8.3.1
#1 lore raw

PATCH 1/1 vfio iommu: Add dma limit capability

Commit 492855939bdb ("vfio/type1: Limit DMA mappings per container") added the ability to limit the number of memory backed DMA mappings. However on s390x, when lazy mapping is in use, we use a very large number of concurrent mappings. Let's provide the limitation to userspace via the IOMMU info chain so that userspace can take appropriate mitigation.

---
 drivers/vfio/vfio_iommu_type1.c | 17 +++++++++++++++++
 include/uapi/linux/vfio.h       | 16 ++++++++++++++++
 2 files changed, 33 insertions(+)
drivers/vfio/vfio_iommu_type1.c +17 -0
index 5fbf0c1..573c2c9 100644
@@ -2609,6 +2609,20 @@ static int vfio_iommu_migration_build_caps(struct vfio_iommu *iommu,
 	return vfio_info_add_capability(caps, &cap_mig.header, sizeof(cap_mig));
 }
 
+static int vfio_iommu_dma_limit_build_caps(struct vfio_iommu *iommu,
+					   struct vfio_info_cap *caps)
+{
+	struct vfio_iommu_type1_info_dma_limit cap_dma_limit;
+
+	cap_dma_limit.header.id = VFIO_IOMMU_TYPE1_INFO_DMA_LIMIT;
+	cap_dma_limit.header.version = 1;
+
+	cap_dma_limit.max = dma_entry_limit;
+
+	return vfio_info_add_capability(caps, &cap_dma_limit.header,
+					sizeof(cap_dma_limit));
+}
+
 static int vfio_iommu_type1_get_info(struct vfio_iommu *iommu,
 				     unsigned long arg)
 {
@@ -2642,6 +2656,9 @@ static int vfio_iommu_type1_get_info(struct vfio_iommu *iommu,
 	ret = vfio_iommu_migration_build_caps(iommu, &caps);
 
 	if (!ret)
+		ret = vfio_iommu_dma_limit_build_caps(iommu, &caps);
+
+	if (!ret)
 		ret = vfio_iommu_iova_build_caps(iommu, &caps);
 
 	mutex_unlock(&iommu->lock);
include/uapi/linux/vfio.h +16 -0
index 9204705..c91e471 100644
@@ -1039,6 +1039,22 @@ struct vfio_iommu_type1_info_cap_migration {
 	__u64	max_dirty_bitmap_size;		/* in bytes */
 };
 
+/*
+ * The DMA limit capability allows to report the number of simultaneously
+ * outstanding DMA mappings are supported.
+ *
+ * The structures below define version 1 of this capability.
+ *
+ * max: specifies the maximum number of outstanding DMA mappings allowed.
+ */
+#define VFIO_IOMMU_TYPE1_INFO_DMA_LIMIT 3
+
+struct vfio_iommu_type1_info_dma_limit {
+	struct	vfio_info_cap_header header;
+	__u32	max;
+};
+
+
 #define VFIO_IOMMU_GET_INFO _IO(VFIO_TYPE, VFIO_BASE + 12)
 
 /**
Signature · 1 line
1.8.3.1
#2 lore raw
Quoted 20 lines · On Fri, 11 Sep 2020 12:44:03 -0400 Matthew Rosato <[email protected]>…

Commit 492855939bdb ("vfio/type1: Limit DMA mappings per container") added the ability to limit the number of memory backed DMA mappings. However on s390x, when lazy mapping is in use, we use a very large number of concurrent mappings. Let's provide the limitation to userspace via the IOMMU info chain so that userspace can take appropriate mitigation.

---
 drivers/vfio/vfio_iommu_type1.c | 17 +++++++++++++++++
 include/uapi/linux/vfio.h       | 16 ++++++++++++++++
 2 files changed, 33 insertions(+)
drivers/vfio/vfio_iommu_type1.c +0 -0
index 5fbf0c1..573c2c9 100644
@@ -2609,6 +2609,20 @@ static int vfio_iommu_migration_build_caps(struct vfio_iommu *iommu,
 	return vfio_info_add_capability(caps, &cap_mig.header, sizeof(cap_mig));
 }

drivers/vfio/vfio_iommu_type1.c static int vfio_iommu_migration_build_caps(struct vfio_iommu *iommu, · line 2620

Quoted 10 lines
drivers/vfio/vfio_iommu_type1.c +9 -0
 
+static int vfio_iommu_dma_limit_build_caps(struct vfio_iommu *iommu,
+					   struct vfio_info_cap *caps)
+{
+	struct vfio_iommu_type1_info_dma_limit cap_dma_limit;
+
+	cap_dma_limit.header.id = VFIO_IOMMU_TYPE1_INFO_DMA_LIMIT;
+	cap_dma_limit.header.version = 1;
+
+	cap_dma_limit.max = dma_entry_limit;

I think you want to report iommu->dma_avail, which might change the naming and semantics of the capability a bit. dma_entry_limit is a writable module param, so the current value might not be relevant to this container at the time that it's read. When a container is opened we set iommu->dma_avail to the current dma_entry_limit, therefore later modifications of dma_entry_limit are only relevant to subsequent containers.

It seems like there are additional benefits to reporting available dma entries as well, for example on mapping failure userspace could reevaluate, perhaps even validate usage counts between kernel and user. Thanks,

Alex

Quoted 45 lines
drivers/vfio/vfio_iommu_type1.c +8 -0
+
+	return vfio_info_add_capability(caps, &cap_dma_limit.header,
+					sizeof(cap_dma_limit));
+}
+
 static int vfio_iommu_type1_get_info(struct vfio_iommu *iommu,
 				     unsigned long arg)
 {
@@ -2642,6 +2656,9 @@ static int vfio_iommu_type1_get_info(struct vfio_iommu *iommu,
 	ret = vfio_iommu_migration_build_caps(iommu, &caps);
 
 	if (!ret)
+		ret = vfio_iommu_dma_limit_build_caps(iommu, &caps);
+
+	if (!ret)
 		ret = vfio_iommu_iova_build_caps(iommu, &caps);
 
 	mutex_unlock(&iommu->lock);
include/uapi/linux/vfio.h +16 -0
index 9204705..c91e471 100644
@@ -1039,6 +1039,22 @@ struct vfio_iommu_type1_info_cap_migration {
 	__u64	max_dirty_bitmap_size;		/* in bytes */
 };
 
+/*
+ * The DMA limit capability allows to report the number of simultaneously
+ * outstanding DMA mappings are supported.
+ *
+ * The structures below define version 1 of this capability.
+ *
+ * max: specifies the maximum number of outstanding DMA mappings allowed.
+ */
+#define VFIO_IOMMU_TYPE1_INFO_DMA_LIMIT 3
+
+struct vfio_iommu_type1_info_dma_limit {
+	struct	vfio_info_cap_header header;
+	__u32	max;
+};
+
+
 #define VFIO_IOMMU_GET_INFO _IO(VFIO_TYPE, VFIO_BASE + 12)
 
 /**
#3 lore raw
Quoted 43 lines · Alex Williamson
Quoted 20 lines · On Fri, 11 Sep 2020 12:44:03 -0400 Matthew Rosato <[email protected]>…

Commit 492855939bdb ("vfio/type1: Limit DMA mappings per container") added the ability to limit the number of memory backed DMA mappings. However on s390x, when lazy mapping is in use, we use a very large number of concurrent mappings. Let's provide the limitation to userspace via the IOMMU info chain so that userspace can take appropriate mitigation.

---
  drivers/vfio/vfio_iommu_type1.c | 17 +++++++++++++++++
  include/uapi/linux/vfio.h       | 16 ++++++++++++++++
  2 files changed, 33 insertions(+)
drivers/vfio/vfio_iommu_type1.c +0 -0
index 5fbf0c1..573c2c9 100644
@@ -2609,6 +2609,20 @@ static int vfio_iommu_migration_build_caps(struct vfio_iommu *iommu,
  	return vfio_info_add_capability(caps, &cap_mig.header, sizeof(cap_mig));
  }

drivers/vfio/vfio_iommu_type1.c static int vfio_iommu_migration_build_caps(struct vfio_iommu *iommu, · line 2620

Quoted 10 lines
drivers/vfio/vfio_iommu_type1.c +9 -0
 
+static int vfio_iommu_dma_limit_build_caps(struct vfio_iommu *iommu,
+					   struct vfio_info_cap *caps)
+{
+	struct vfio_iommu_type1_info_dma_limit cap_dma_limit;
+
+	cap_dma_limit.header.id = VFIO_IOMMU_TYPE1_INFO_DMA_LIMIT;
+	cap_dma_limit.header.version = 1;
+
+	cap_dma_limit.max = dma_entry_limit;

I think you want to report iommu->dma_avail, which might change the naming and semantics of the capability a bit. dma_entry_limit is a writable module param, so the current value might not be relevant to this container at the time that it's read. When a container is opened we set iommu->dma_avail to the current dma_entry_limit, therefore later modifications of dma_entry_limit are only relevant to subsequent containers.

Quoted 3 lines

It seems like there are additional benefits to reporting available dma entries as well, for example on mapping failure userspace could reevaluate, perhaps even validate usage counts between kernel and user.

Hmm, both good points. I'll re-work to something that presents the current dma_avail for the container instead. Thanks!

Quoted 50 lines

Thanks,

Alex

Quoted 45 lines
drivers/vfio/vfio_iommu_type1.c +8 -0
+
+	return vfio_info_add_capability(caps, &cap_dma_limit.header,
+					sizeof(cap_dma_limit));
+}
+
  static int vfio_iommu_type1_get_info(struct vfio_iommu *iommu,
  				     unsigned long arg)
  {
@@ -2642,6 +2656,9 @@ static int vfio_iommu_type1_get_info(struct vfio_iommu *iommu,
  	ret = vfio_iommu_migration_build_caps(iommu, &caps);
 
  	if (!ret)
+		ret = vfio_iommu_dma_limit_build_caps(iommu, &caps);
+
+	if (!ret)
  		ret = vfio_iommu_iova_build_caps(iommu, &caps);
 
  	mutex_unlock(&iommu->lock);
include/uapi/linux/vfio.h +16 -0
index 9204705..c91e471 100644
@@ -1039,6 +1039,22 @@ struct vfio_iommu_type1_info_cap_migration {
  	__u64	max_dirty_bitmap_size;		/* in bytes */
  };
 
+/*
+ * The DMA limit capability allows to report the number of simultaneously
+ * outstanding DMA mappings are supported.
+ *
+ * The structures below define version 1 of this capability.
+ *
+ * max: specifies the maximum number of outstanding DMA mappings allowed.
+ */
+#define VFIO_IOMMU_TYPE1_INFO_DMA_LIMIT 3
+
+struct vfio_iommu_type1_info_dma_limit {
+	struct	vfio_info_cap_header header;
+	__u32	max;
+};
+
+
  #define VFIO_IOMMU_GET_INFO _IO(VFIO_TYPE, VFIO_BASE + 12)
 
  /**
#4 lore raw
In reply to Matthew Rosato
Quoted 7 lines · Matthew Rosato

Commit 492855939bdb ("vfio/type1: Limit DMA mappings per container") added a limit to the number of concurrent DMA requests for a vfio container. However, lazy unmapping in s390 can in fact cause quite a large number of outstanding DMA requests to build up prior to being purged, potentially the entire guest DMA space. This results in unexpected 'VFIO_MAP_DMA
failed: No space left on device' conditions seen in QEMU.

Quoted 5 lines

This patch proposes to provide the DMA limit via the VFIO_IOMMU_GET_INFO ioctl as a new capability. A subsequent patchset to QEMU would collect this information and use it in s390 PCI support to tap the guest on the shoulder before overrunning the vfio limit.

Link to the QEMU patchset:
https://lists.gnu.org/archive/html/qemu-devel/2020-09/msg04322.html

Quoted 7 lines

Matthew Rosato (1):
vfio iommu: Add dma limit capability

  drivers/vfio/vfio_iommu_type1.c | 17 +++++++++++++++++
  include/uapi/linux/vfio.h       | 16 ++++++++++++++++
  2 files changed, 33 insertions(+)