Skip to contentSkip to conversation

Add CRIU support for amdgpu dmabuf

PATCH Superseded David Francis 3 participants 12 messages Aug 13, 2025 – Aug 22, 2025

Subsystems (4)

v14 of 16 Superseded Latest is v16 →

lore
Share & tools
lore
#0 lore raw

PATCH v14 1/1 Add CRIU support for amdgpu dmabuf

This patch series adds support for CRIU checkpointing of processes that share memory with the amdgpu dmabuf interface.

This v14 adds the coherence and uncached flags to gem create and fixes some formatting issues in amdgpu_drm.h

Accompanying CRIU changes:
https://github.com/checkpoint-restore/criu/pull/2613

#1 lore raw

PATCH v14 1/4 drm/amdgpu: Allow more flags to be set on gem create.

AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE is a flag that specifies that gem memory contains sensitive information and should be cleared to prevent snooping.

The COHERENT and UNCACHED gem create flags enable memory features related to sharing memory across devices.

These should be settable in GEM_CREATE_IOCTL but weren't.

Make a new define AMDGPU_GEM_CREATE_SETTABLE_MASK to track which gem flags can be used with gem create, and add these flags to it.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 10 +---------
 drivers/gpu/drm/amd/amdgpu/amdgpu_gem.h | 14 ++++++++++++++
 2 files changed, 15 insertions(+), 9 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +1 -9
index e3f65977eeee..aefae3a9e6f4 100644
@@ -442,15 +442,7 @@ int amdgpu_gem_create_ioctl(struct drm_device *dev, void *data,
 	int r;
 
 	/* reject invalid gem flags */
-	if (flags & ~(AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
-		      AMDGPU_GEM_CREATE_NO_CPU_ACCESS |
-		      AMDGPU_GEM_CREATE_CPU_GTT_USWC |
-		      AMDGPU_GEM_CREATE_VRAM_CLEARED |
-		      AMDGPU_GEM_CREATE_VM_ALWAYS_VALID |
-		      AMDGPU_GEM_CREATE_EXPLICIT_SYNC |
-		      AMDGPU_GEM_CREATE_ENCRYPTED |
-		      AMDGPU_GEM_CREATE_GFX12_DCC |
-		      AMDGPU_GEM_CREATE_DISCARDABLE))
+	if (flags & ~AMDGPU_GEM_CREATE_SETTABLE_MASK)
 		return -EINVAL;
 
 	/* reject invalid gem domains */
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.h +14 -0
index b51e8f95ee86..b3047d73fe07 100644
@@ -71,4 +71,18 @@ int amdgpu_gem_op_ioctl(struct drm_device *dev, void *data,
 int amdgpu_gem_metadata_ioctl(struct drm_device *dev, void *data,
 				struct drm_file *filp);
 
+#define AMDGPU_GEM_CREATE_SETTABLE_MASK	(AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED | \
+	AMDGPU_GEM_CREATE_NO_CPU_ACCESS | \
+	AMDGPU_GEM_CREATE_CPU_GTT_USWC | \
+	AMDGPU_GEM_CREATE_VRAM_CLEARED | \
+	AMDGPU_GEM_CREATE_VM_ALWAYS_VALID | \
+	AMDGPU_GEM_CREATE_EXPLICIT_SYNC | \
+	AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE | \
+	AMDGPU_GEM_CREATE_ENCRYPTED | \
+	AMDGPU_GEM_CREATE_GFX12_DCC | \
+	AMDGPU_GEM_CREATE_DISCARDABLE | \
+	AMDGPU_GEM_CREATE_COHERENT | \
+	AMDGPU_GEM_CREATE_UNCACHED | \
+	AMDGPU_GEM_CREATE_EXT_COHERENT)
+
 #endif
Signature · 1 line
2.34.1
#2 lore raw

Re: [PATCH v14 1/4] drm/amdgpu: Allow more flags to be set on gem create.

Quoted 8 lines · David Francis

AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE is a flag that specifies that gem memory contains sensitive information and should be cleared to prevent snooping.

The COHERENT and UNCACHED gem create flags enable memory features related to sharing memory across devices.

These should be settable in GEM_CREATE_IOCTL but weren't.

"Should be" is questionable, surely there were reasons for not allowing user space to use them so far.

Signature · 2 lines
Earthling Michel Dänzer       \        GNOME / Xwayland / Mesa developer
https://redhat.com             \               Libre software enthusiast
maintainer
#3 lore raw

Re: [PATCH v14 1/4] drm/amdgpu: Allow more flags to be set on gem create.

In reply to David Francis
Quoted 8 lines · David Francis

AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE is a flag that specifies that gem memory contains sensitive information and should be cleared to prevent snooping.

The COHERENT and UNCACHED gem create flags enable memory features related to sharing memory across devices.

These should be settable in GEM_CREATE_IOCTL but weren't.

That needs further explanation. Something like:

For CRIU we need to re-create KFD BOs through the GEM CREATE IOCTL, so allow those KFD specific flags here as well. This will also aid us in the future and allows to move the KFD components over using the render node for allocations.

Quoted 4 lines

Make a new define AMDGPU_GEM_CREATE_SETTABLE_MASK to track which gem flags can be used with gem create, and add these flags to it.

This describes what the patch does which should be obvious by reading the patch. I would drop it.

Quoted 33 lines
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 10 +---------
 drivers/gpu/drm/amd/amdgpu/amdgpu_gem.h | 14 ++++++++++++++
 2 files changed, 15 insertions(+), 9 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +1 -9
index e3f65977eeee..aefae3a9e6f4 100644
@@ -442,15 +442,7 @@ int amdgpu_gem_create_ioctl(struct drm_device *dev, void *data,
 	int r;
 
 	/* reject invalid gem flags */
-	if (flags & ~(AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
-		      AMDGPU_GEM_CREATE_NO_CPU_ACCESS |
-		      AMDGPU_GEM_CREATE_CPU_GTT_USWC |
-		      AMDGPU_GEM_CREATE_VRAM_CLEARED |
-		      AMDGPU_GEM_CREATE_VM_ALWAYS_VALID |
-		      AMDGPU_GEM_CREATE_EXPLICIT_SYNC |
-		      AMDGPU_GEM_CREATE_ENCRYPTED |
-		      AMDGPU_GEM_CREATE_GFX12_DCC |
-		      AMDGPU_GEM_CREATE_DISCARDABLE))
+	if (flags & ~AMDGPU_GEM_CREATE_SETTABLE_MASK)
 		return -EINVAL;
 
 	/* reject invalid gem domains */
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.h +0 -0
index b51e8f95ee86..b3047d73fe07 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_gem.h int amdgpu_gem_op_ioctl(struct drm_device *dev, void *data, · line 73

Quoted 3 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.h +0 -0
@@ -71,4 +71,18 @@ int amdgpu_gem_op_ioctl(struct drm_device *dev, void *data,
 int amdgpu_gem_metadata_ioctl(struct drm_device *dev, void *data,
 				struct drm_file *filp);
 

A comment explaining what this does would be nice to have.

Apart from that looks good to me,
Christian.

Quoted 15 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.h +14 -0
+#define AMDGPU_GEM_CREATE_SETTABLE_MASK	(AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED | \
+	AMDGPU_GEM_CREATE_NO_CPU_ACCESS | \
+	AMDGPU_GEM_CREATE_CPU_GTT_USWC | \
+	AMDGPU_GEM_CREATE_VRAM_CLEARED | \
+	AMDGPU_GEM_CREATE_VM_ALWAYS_VALID | \
+	AMDGPU_GEM_CREATE_EXPLICIT_SYNC | \
+	AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE | \
+	AMDGPU_GEM_CREATE_ENCRYPTED | \
+	AMDGPU_GEM_CREATE_GFX12_DCC | \
+	AMDGPU_GEM_CREATE_DISCARDABLE | \
+	AMDGPU_GEM_CREATE_COHERENT | \
+	AMDGPU_GEM_CREATE_UNCACHED | \
+	AMDGPU_GEM_CREATE_EXT_COHERENT)
+
 #endif
maintainer Reviewed-by on v14 1/4
#4 lore raw

Re: [PATCH v14 1/4] drm/amdgpu: Allow more flags to be set on gem create.

In reply to David Francis
Quoted 7 lines · David Francis

AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE is a flag that specifies that gem memory contains sensitive information and should be cleared to prevent snooping.

The COHERENT and UNCACHED gem create flags enable memory features related to sharing memory across devices.

Quoted 7 lines

These should be settable in GEM_CREATE_IOCTL but weren't.

Make a new define AMDGPU_GEM_CREATE_SETTABLE_MASK to track which gem flags can be used with gem create, and add these flags to it.

Quoted 49 lines
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 10 +---------
 drivers/gpu/drm/amd/amdgpu/amdgpu_gem.h | 14 ++++++++++++++
 2 files changed, 15 insertions(+), 9 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +1 -9
index e3f65977eeee..aefae3a9e6f4 100644
@@ -442,15 +442,7 @@ int amdgpu_gem_create_ioctl(struct drm_device *dev, void *data,
 	int r;
 
 	/* reject invalid gem flags */
-	if (flags & ~(AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
-		      AMDGPU_GEM_CREATE_NO_CPU_ACCESS |
-		      AMDGPU_GEM_CREATE_CPU_GTT_USWC |
-		      AMDGPU_GEM_CREATE_VRAM_CLEARED |
-		      AMDGPU_GEM_CREATE_VM_ALWAYS_VALID |
-		      AMDGPU_GEM_CREATE_EXPLICIT_SYNC |
-		      AMDGPU_GEM_CREATE_ENCRYPTED |
-		      AMDGPU_GEM_CREATE_GFX12_DCC |
-		      AMDGPU_GEM_CREATE_DISCARDABLE))
+	if (flags & ~AMDGPU_GEM_CREATE_SETTABLE_MASK)
 		return -EINVAL;
 
 	/* reject invalid gem domains */
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.h +14 -0
index b51e8f95ee86..b3047d73fe07 100644
@@ -71,4 +71,18 @@ int amdgpu_gem_op_ioctl(struct drm_device *dev, void *data,
 int amdgpu_gem_metadata_ioctl(struct drm_device *dev, void *data,
 				struct drm_file *filp);
 
+#define AMDGPU_GEM_CREATE_SETTABLE_MASK	(AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED | \
+	AMDGPU_GEM_CREATE_NO_CPU_ACCESS | \
+	AMDGPU_GEM_CREATE_CPU_GTT_USWC | \
+	AMDGPU_GEM_CREATE_VRAM_CLEARED | \
+	AMDGPU_GEM_CREATE_VM_ALWAYS_VALID | \
+	AMDGPU_GEM_CREATE_EXPLICIT_SYNC | \
+	AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE | \
+	AMDGPU_GEM_CREATE_ENCRYPTED | \
+	AMDGPU_GEM_CREATE_GFX12_DCC | \
+	AMDGPU_GEM_CREATE_DISCARDABLE | \
+	AMDGPU_GEM_CREATE_COHERENT | \
+	AMDGPU_GEM_CREATE_UNCACHED | \
+	AMDGPU_GEM_CREATE_EXT_COHERENT)
+
 #endif
#5 lore raw

PATCH v14 2/4 drm/amdgpu: Add ioctl to get all gem handles for a process

In reply to David Francis

Add new ioctl DRM_IOCTL_AMDGPU_GEM_LIST_HANDLES.

This ioctl returns a list of bos with their handles, sizes, and flags and domains.

This ioctl is meant to be used during CRIU checkpoint and provide information needed to reconstruct the bos in CRIU restore.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c |  1 +
 drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 79 +++++++++++++++++++++++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_gem.h |  2 +
 include/uapi/drm/amdgpu_drm.h           | 31 ++++++++++
 4 files changed, 113 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +1 -0
index 4ff3a2eaaf55..f19795dddf9d 100644
@@ -3031,6 +3031,7 @@ const struct drm_ioctl_desc amdgpu_ioctls_kms[] = {
 	DRM_IOCTL_DEF_DRV(AMDGPU_USERQ, amdgpu_userq_ioctl, DRM_AUTH|DRM_RENDER_ALLOW),
 	DRM_IOCTL_DEF_DRV(AMDGPU_USERQ_SIGNAL, amdgpu_userq_signal_ioctl, DRM_AUTH|DRM_RENDER_ALLOW),
 	DRM_IOCTL_DEF_DRV(AMDGPU_USERQ_WAIT, amdgpu_userq_wait_ioctl, DRM_AUTH|DRM_RENDER_ALLOW),
+	DRM_IOCTL_DEF_DRV(AMDGPU_GEM_LIST_HANDLES, amdgpu_gem_list_handles_ioctl, DRM_AUTH|DRM_RENDER_ALLOW),
 };
 
 static const struct drm_driver amdgpu_kms_driver = {
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +79 -0

85 lines not on this page · raw message

drivers/gpu/drm/amd/amdgpu/amdgpu_gem.h +2 -0
index b3047d73fe07..338fd6d6423c 100644
@@ -67,6 +67,8 @@ int amdgpu_gem_va_ioctl(struct drm_device *dev, void *data,
 			  struct drm_file *filp);
 int amdgpu_gem_op_ioctl(struct drm_device *dev, void *data,
 			struct drm_file *filp);
+int amdgpu_gem_list_handles_ioctl(struct drm_device *dev, void *data,
+			      struct drm_file *filp);
 
 int amdgpu_gem_metadata_ioctl(struct drm_device *dev, void *data,
 				struct drm_file *filp);
include/uapi/drm/amdgpu_drm.h +31 -0
index bdedbaccf776..59b423883e91 100644
@@ -57,6 +57,7 @@ extern "C" {
 #define DRM_AMDGPU_USERQ		0x16
 #define DRM_AMDGPU_USERQ_SIGNAL		0x17
 #define DRM_AMDGPU_USERQ_WAIT		0x18
+#define DRM_AMDGPU_GEM_LIST_HANDLES	0x19
 
 #define DRM_IOCTL_AMDGPU_GEM_CREATE	DRM_IOWR(DRM_COMMAND_BASE + DRM_AMDGPU_GEM_CREATE, union drm_amdgpu_gem_create)
 #define DRM_IOCTL_AMDGPU_GEM_MMAP	DRM_IOWR(DRM_COMMAND_BASE + DRM_AMDGPU_GEM_MMAP, union drm_amdgpu_gem_mmap)
@@ -77,6 +78,7 @@ extern "C" {
 #define DRM_IOCTL_AMDGPU_USERQ		DRM_IOWR(DRM_COMMAND_BASE + DRM_AMDGPU_USERQ, union drm_amdgpu_userq)
 #define DRM_IOCTL_AMDGPU_USERQ_SIGNAL	DRM_IOWR(DRM_COMMAND_BASE + DRM_AMDGPU_USERQ_SIGNAL, struct drm_amdgpu_userq_signal)
 #define DRM_IOCTL_AMDGPU_USERQ_WAIT	DRM_IOWR(DRM_COMMAND_BASE + DRM_AMDGPU_USERQ_WAIT, struct drm_amdgpu_userq_wait)
+#define DRM_IOCTL_AMDGPU_GEM_LIST_HANDLES DRM_IOWR(DRM_COMMAND_BASE + DRM_AMDGPU_GEM_LIST_HANDLES, struct drm_amdgpu_gem_list_handles)
 
 /**
  * DOC: memory domains
@@ -811,6 +813,35 @@ struct drm_amdgpu_gem_op {
 	__u64	value;
 };
 
+#define AMDGPU_GEM_LIST_HANDLES_FLAG_IS_IMPORT	(1 << 0)
+
+struct drm_amdgpu_gem_list_handles {
+	/* User pointer to array of drm_amdgpu_gem_bo_info_entry */
+	__u64   entries;
+
+	/* IN: Size of entries buffer. OUT: Number of handles in process (if larger than size of buffer, must retry) */
+	__u32   num_entries;
+
+	__u32 padding;
+};
+
+struct drm_amdgpu_gem_list_handles_entry {
+	/* gem handle of buffer object */
+	__u32 gem_handle;
+
+	/* Currently just one flag: IS_IMPORT */
+	__u32 flags;
+
+	/* Size of bo */
+	__u64 size;
+
+	/* Preferred domains for GEM_CREATE */
+	__u64 preferred_domains;
+
+	/* GEM_CREATE flags for re-creation of buffer */
+	__u64 alloc_flags;
+};
+
 #define AMDGPU_VA_OP_MAP			1
 #define AMDGPU_VA_OP_UNMAP			2
 #define AMDGPU_VA_OP_CLEAR			3
Signature · 1 line
2.34.1
maintainer
#6 lore raw

Re: [PATCH v14 2/4] drm/amdgpu: Add ioctl to get all gem handles for a process

Quoted 153 lines · David Francis

Add new ioctl DRM_IOCTL_AMDGPU_GEM_LIST_HANDLES.

This ioctl returns a list of bos with their handles, sizes, and flags and domains.

This ioctl is meant to be used during CRIU checkpoint and provide information needed to reconstruct the bos in CRIU restore.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c |  1 +
 drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 79 +++++++++++++++++++++++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_gem.h |  2 +
 include/uapi/drm/amdgpu_drm.h           | 31 ++++++++++
 4 files changed, 113 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +1 -0
index 4ff3a2eaaf55..f19795dddf9d 100644
@@ -3031,6 +3031,7 @@ const struct drm_ioctl_desc amdgpu_ioctls_kms[] = {
 	DRM_IOCTL_DEF_DRV(AMDGPU_USERQ, amdgpu_userq_ioctl, DRM_AUTH|DRM_RENDER_ALLOW),
 	DRM_IOCTL_DEF_DRV(AMDGPU_USERQ_SIGNAL, amdgpu_userq_signal_ioctl, DRM_AUTH|DRM_RENDER_ALLOW),
 	DRM_IOCTL_DEF_DRV(AMDGPU_USERQ_WAIT, amdgpu_userq_wait_ioctl, DRM_AUTH|DRM_RENDER_ALLOW),
+	DRM_IOCTL_DEF_DRV(AMDGPU_GEM_LIST_HANDLES, amdgpu_gem_list_handles_ioctl, DRM_AUTH|DRM_RENDER_ALLOW),
 };
 
 static const struct drm_driver amdgpu_kms_driver = {
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +79 -0
index aefae3a9e6f4..d17cc7ba66d4 100644
@@ -1024,6 +1024,85 @@ int amdgpu_gem_op_ioctl(struct drm_device *dev, void *data,
 	return r;
 }
 
+/**
+ * drm_amdgpu_gem_list_handles_ioctl - get information about a process' buffer objects
+ *
+ * @dev: drm device pointer
+ * @data: drm_amdgpu_gem_list_handles
+ * @filp: drm file pointer
+ *
+ * num_entries is set as an input to the size of the entries array.
+ * num_entries is sent back as output as the number of bos in the process.
+ * If that number is larger than the size of the array, the ioctl must
+ * be retried.
+ *
+ * Returns:
+ * 0 for success, -errno for errors.
+ */
+int amdgpu_gem_list_handles_ioctl(struct drm_device *dev, void *data,
+			      struct drm_file *filp)
+{
+	struct drm_amdgpu_gem_list_handles *args = data;
+	struct drm_amdgpu_gem_list_handles_entry *bo_entries;
+	struct drm_gem_object *gobj;
+	int id, ret = 0;
+	int bo_index = 0;
+	int num_bos = 0;
+
+	spin_lock(&filp->table_lock);
+	idr_for_each_entry(&filp->object_idr, gobj, id)
+		num_bos += 1;
+	spin_unlock(&filp->table_lock);
+
+	if (args->num_entries < num_bos) {
+		args->num_entries = num_bos;
+		return 0;
+	}
+
+	if (num_bos == 0) {
+		args->num_entries = 0;
+		return 0;
+	}
+
+	bo_entries = kvcalloc(num_bos, sizeof(*bo_entries), GFP_KERNEL);
+	if (!bo_entries)
+		return -ENOMEM;
+
+	spin_lock(&filp->table_lock);
+	idr_for_each_entry(&filp->object_idr, gobj, id) {
+		struct amdgpu_bo *bo = gem_to_amdgpu_bo(gobj);
+		struct drm_amdgpu_gem_list_handles_entry *bo_entry;
+
+		if (bo_index >= num_bos) {
+			ret = -EAGAIN;
+			break;
+		}
+
+		bo_entry = &bo_entries[bo_index];
+
+		bo_entry->size = amdgpu_bo_size(bo);
+		bo_entry->alloc_flags = bo->flags & AMDGPU_GEM_CREATE_SETTABLE_MASK;
+		bo_entry->preferred_domains = bo->preferred_domains;
+		bo_entry->gem_handle = id;
+
+		if (bo->tbo.base.import_attach)
+			bo_entry->flags |= AMDGPU_GEM_LIST_HANDLES_FLAG_IS_IMPORT;
+
+		bo_index += 1;
+	}
+	spin_unlock(&filp->table_lock);
+
+	args->num_entries = bo_index;
+
+	if (!ret)
+		ret = copy_to_user(u64_to_user_ptr(args->entries), bo_entries, num_bos * sizeof(*bo_entries));
+
+	kvfree(bo_entries);
+
+	return ret;
+}
+
+
 static int amdgpu_gem_align_pitch(struct amdgpu_device *adev,
 				  int width,
 				  int cpp,
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.h +2 -0
index b3047d73fe07..338fd6d6423c 100644
@@ -67,6 +67,8 @@ int amdgpu_gem_va_ioctl(struct drm_device *dev, void *data,
 			  struct drm_file *filp);
 int amdgpu_gem_op_ioctl(struct drm_device *dev, void *data,
 			struct drm_file *filp);
+int amdgpu_gem_list_handles_ioctl(struct drm_device *dev, void *data,
+			      struct drm_file *filp);
 
 int amdgpu_gem_metadata_ioctl(struct drm_device *dev, void *data,
 				struct drm_file *filp);
include/uapi/drm/amdgpu_drm.h +2 -0
index bdedbaccf776..59b423883e91 100644
@@ -57,6 +57,7 @@ extern "C" {
 #define DRM_AMDGPU_USERQ		0x16
 #define DRM_AMDGPU_USERQ_SIGNAL		0x17
 #define DRM_AMDGPU_USERQ_WAIT		0x18
+#define DRM_AMDGPU_GEM_LIST_HANDLES	0x19
 
 #define DRM_IOCTL_AMDGPU_GEM_CREATE	DRM_IOWR(DRM_COMMAND_BASE + DRM_AMDGPU_GEM_CREATE, union drm_amdgpu_gem_create)
 #define DRM_IOCTL_AMDGPU_GEM_MMAP	DRM_IOWR(DRM_COMMAND_BASE + DRM_AMDGPU_GEM_MMAP, union drm_amdgpu_gem_mmap)
@@ -77,6 +78,7 @@ extern "C" {
 #define DRM_IOCTL_AMDGPU_USERQ		DRM_IOWR(DRM_COMMAND_BASE + DRM_AMDGPU_USERQ, union drm_amdgpu_userq)
 #define DRM_IOCTL_AMDGPU_USERQ_SIGNAL	DRM_IOWR(DRM_COMMAND_BASE + DRM_AMDGPU_USERQ_SIGNAL, struct drm_amdgpu_userq_signal)
 #define DRM_IOCTL_AMDGPU_USERQ_WAIT	DRM_IOWR(DRM_COMMAND_BASE + DRM_AMDGPU_USERQ_WAIT, struct drm_amdgpu_userq_wait)
+#define DRM_IOCTL_AMDGPU_GEM_LIST_HANDLES DRM_IOWR(DRM_COMMAND_BASE + DRM_AMDGPU_GEM_LIST_HANDLES, struct drm_amdgpu_gem_list_handles)
 
 /**
  * DOC: memory domains

include/uapi/drm/amdgpu_drm.h struct drm_amdgpu_gem_op { · line 822

Quoted 10 lines
include/uapi/drm/amdgpu_drm.h +7 -0
@@ -811,6 +813,35 @@ struct drm_amdgpu_gem_op {
 	__u64	value;
 };
 
+#define AMDGPU_GEM_LIST_HANDLES_FLAG_IS_IMPORT	(1 << 0)
+
+struct drm_amdgpu_gem_list_handles {
+	/* User pointer to array of drm_amdgpu_gem_bo_info_entry */
+	__u64   entries;
+
+	/* IN: Size of entries buffer. OUT: Number of handles in process (if larger than size of buffer, must retry) */

That line looks to long, please make sure checkpatch.pl doesn't complain.

And drop the "IN", that's also an returned parameter.

Apart from those nit picks feel free to add Reviewed-by: Christian König <[email protected]>.

Regards,
Christian.

Quoted 25 lines
include/uapi/drm/amdgpu_drm.h +22 -0
+	__u32   num_entries;
+
+	__u32 padding;
+};
+
+struct drm_amdgpu_gem_list_handles_entry {
+	/* gem handle of buffer object */
+	__u32 gem_handle;
+
+	/* Currently just one flag: IS_IMPORT */
+	__u32 flags;
+
+	/* Size of bo */
+	__u64 size;
+
+	/* Preferred domains for GEM_CREATE */
+	__u64 preferred_domains;
+
+	/* GEM_CREATE flags for re-creation of buffer */
+	__u64 alloc_flags;
+};
+
 #define AMDGPU_VA_OP_MAP			1
 #define AMDGPU_VA_OP_UNMAP			2
 #define AMDGPU_VA_OP_CLEAR			3
maintainer
#7 lore raw

Re: [PATCH v14 2/4] drm/amdgpu: Add ioctl to get all gem handles for a process

In reply to David Francis
Quoted 175 lines · David Francis

Add new ioctl DRM_IOCTL_AMDGPU_GEM_LIST_HANDLES.

This ioctl returns a list of bos with their handles, sizes, and flags and domains.

This ioctl is meant to be used during CRIU checkpoint and provide information needed to reconstruct the bos in CRIU restore.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c |  1 +
 drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 79 +++++++++++++++++++++++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_gem.h |  2 +
 include/uapi/drm/amdgpu_drm.h           | 31 ++++++++++
 4 files changed, 113 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +1 -0
index 4ff3a2eaaf55..f19795dddf9d 100644
@@ -3031,6 +3031,7 @@ const struct drm_ioctl_desc amdgpu_ioctls_kms[] = {
 	DRM_IOCTL_DEF_DRV(AMDGPU_USERQ, amdgpu_userq_ioctl, DRM_AUTH|DRM_RENDER_ALLOW),
 	DRM_IOCTL_DEF_DRV(AMDGPU_USERQ_SIGNAL, amdgpu_userq_signal_ioctl, DRM_AUTH|DRM_RENDER_ALLOW),
 	DRM_IOCTL_DEF_DRV(AMDGPU_USERQ_WAIT, amdgpu_userq_wait_ioctl, DRM_AUTH|DRM_RENDER_ALLOW),
+	DRM_IOCTL_DEF_DRV(AMDGPU_GEM_LIST_HANDLES, amdgpu_gem_list_handles_ioctl, DRM_AUTH|DRM_RENDER_ALLOW),
 };
 
 static const struct drm_driver amdgpu_kms_driver = {
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +79 -0
index aefae3a9e6f4..d17cc7ba66d4 100644
@@ -1024,6 +1024,85 @@ int amdgpu_gem_op_ioctl(struct drm_device *dev, void *data,
 	return r;
 }
 
+/**
+ * drm_amdgpu_gem_list_handles_ioctl - get information about a process' buffer objects
+ *
+ * @dev: drm device pointer
+ * @data: drm_amdgpu_gem_list_handles
+ * @filp: drm file pointer
+ *
+ * num_entries is set as an input to the size of the entries array.
+ * num_entries is sent back as output as the number of bos in the process.
+ * If that number is larger than the size of the array, the ioctl must
+ * be retried.
+ *
+ * Returns:
+ * 0 for success, -errno for errors.
+ */
+int amdgpu_gem_list_handles_ioctl(struct drm_device *dev, void *data,
+			      struct drm_file *filp)
+{
+	struct drm_amdgpu_gem_list_handles *args = data;
+	struct drm_amdgpu_gem_list_handles_entry *bo_entries;
+	struct drm_gem_object *gobj;
+	int id, ret = 0;
+	int bo_index = 0;
+	int num_bos = 0;
+
+	spin_lock(&filp->table_lock);
+	idr_for_each_entry(&filp->object_idr, gobj, id)
+		num_bos += 1;
+	spin_unlock(&filp->table_lock);
+
+	if (args->num_entries < num_bos) {
+		args->num_entries = num_bos;
+		return 0;
+	}
+
+	if (num_bos == 0) {
+		args->num_entries = 0;
+		return 0;
+	}
+
+	bo_entries = kvcalloc(num_bos, sizeof(*bo_entries), GFP_KERNEL);
+	if (!bo_entries)
+		return -ENOMEM;
+
+	spin_lock(&filp->table_lock);
+	idr_for_each_entry(&filp->object_idr, gobj, id) {
+		struct amdgpu_bo *bo = gem_to_amdgpu_bo(gobj);
+		struct drm_amdgpu_gem_list_handles_entry *bo_entry;
+
+		if (bo_index >= num_bos) {
+			ret = -EAGAIN;
+			break;
+		}
+
+		bo_entry = &bo_entries[bo_index];
+
+		bo_entry->size = amdgpu_bo_size(bo);
+		bo_entry->alloc_flags = bo->flags & AMDGPU_GEM_CREATE_SETTABLE_MASK;
+		bo_entry->preferred_domains = bo->preferred_domains;
+		bo_entry->gem_handle = id;
+
+		if (bo->tbo.base.import_attach)
+			bo_entry->flags |= AMDGPU_GEM_LIST_HANDLES_FLAG_IS_IMPORT;
+
+		bo_index += 1;
+	}
+	spin_unlock(&filp->table_lock);
+
+	args->num_entries = bo_index;
+
+	if (!ret)
+		ret = copy_to_user(u64_to_user_ptr(args->entries), bo_entries, num_bos * sizeof(*bo_entries));
+
+	kvfree(bo_entries);
+
+	return ret;
+}
+
+
 static int amdgpu_gem_align_pitch(struct amdgpu_device *adev,
 				  int width,
 				  int cpp,
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.h +2 -0
index b3047d73fe07..338fd6d6423c 100644
@@ -67,6 +67,8 @@ int amdgpu_gem_va_ioctl(struct drm_device *dev, void *data,
 			  struct drm_file *filp);
 int amdgpu_gem_op_ioctl(struct drm_device *dev, void *data,
 			struct drm_file *filp);
+int amdgpu_gem_list_handles_ioctl(struct drm_device *dev, void *data,
+			      struct drm_file *filp);
 
 int amdgpu_gem_metadata_ioctl(struct drm_device *dev, void *data,
 				struct drm_file *filp);
include/uapi/drm/amdgpu_drm.h +19 -0
index bdedbaccf776..59b423883e91 100644
@@ -57,6 +57,7 @@ extern "C" {
 #define DRM_AMDGPU_USERQ		0x16
 #define DRM_AMDGPU_USERQ_SIGNAL		0x17
 #define DRM_AMDGPU_USERQ_WAIT		0x18
+#define DRM_AMDGPU_GEM_LIST_HANDLES	0x19
 
 #define DRM_IOCTL_AMDGPU_GEM_CREATE	DRM_IOWR(DRM_COMMAND_BASE + DRM_AMDGPU_GEM_CREATE, union drm_amdgpu_gem_create)
 #define DRM_IOCTL_AMDGPU_GEM_MMAP	DRM_IOWR(DRM_COMMAND_BASE + DRM_AMDGPU_GEM_MMAP, union drm_amdgpu_gem_mmap)
@@ -77,6 +78,7 @@ extern "C" {
 #define DRM_IOCTL_AMDGPU_USERQ		DRM_IOWR(DRM_COMMAND_BASE + DRM_AMDGPU_USERQ, union drm_amdgpu_userq)
 #define DRM_IOCTL_AMDGPU_USERQ_SIGNAL	DRM_IOWR(DRM_COMMAND_BASE + DRM_AMDGPU_USERQ_SIGNAL, struct drm_amdgpu_userq_signal)
 #define DRM_IOCTL_AMDGPU_USERQ_WAIT	DRM_IOWR(DRM_COMMAND_BASE + DRM_AMDGPU_USERQ_WAIT, struct drm_amdgpu_userq_wait)
+#define DRM_IOCTL_AMDGPU_GEM_LIST_HANDLES DRM_IOWR(DRM_COMMAND_BASE + DRM_AMDGPU_GEM_LIST_HANDLES, struct drm_amdgpu_gem_list_handles)
 
 /**
  * DOC: memory domains
@@ -811,6 +813,35 @@ struct drm_amdgpu_gem_op {
 	__u64	value;
 };
 
+#define AMDGPU_GEM_LIST_HANDLES_FLAG_IS_IMPORT	(1 << 0)
+
+struct drm_amdgpu_gem_list_handles {
+	/* User pointer to array of drm_amdgpu_gem_bo_info_entry */
+	__u64   entries;
+
+	/* IN: Size of entries buffer. OUT: Number of handles in process (if larger than size of buffer, must retry) */
+	__u32   num_entries;
+
+	__u32 padding;
+};
+
+struct drm_amdgpu_gem_list_handles_entry {
+	/* gem handle of buffer object */
+	__u32 gem_handle;
 
 
+
+	/* Currently just one flag: IS_IMPORT */

include/uapi/drm/amdgpu_drm.h struct drm_amdgpu_gem_op { · line 844

Quoted 10 lines
include/uapi/drm/amdgpu_drm.h +10 -0
+	__u32 flags;
+
+	/* Size of bo */
+	__u64 size;
+
+	/* Preferred domains for GEM_CREATE */
+	__u64 preferred_domains;
+
+	/* GEM_CREATE flags for re-creation of buffer */
+	__u64 alloc_flags;

I've just realized that we already have all this in the struct drm_amdgpu_gem_create_in and that the alignment is missing!

Please either add the alignment here, or even better use the pre-defined structure.

Regards,
Christian.

Quoted 5 lines
include/uapi/drm/amdgpu_drm.h +2 -0
+};
+
 #define AMDGPU_VA_OP_MAP			1

#define AMDGPU_VA_OP_UNMAP 2
#define AMDGPU_VA_OP_CLEAR 3

#8 lore raw

PATCH v14 3/4 drm/amdgpu: Add mapping info option for GEM_OP ioctl

In reply to David Francis

Add new GEM_OP_IOCTL option GET_MAPPING_INFO, which returns a list of mappings associated with a given bo, along with their positions and offsets.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 87 +++++++++++++++++++++++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h  |  5 ++
 include/uapi/drm/amdgpu_drm.h           | 21 +++++-
 3 files changed, 112 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +87 -0
index d17cc7ba66d4..f66f6e1f1c52 100644
@@ -948,6 +948,89 @@ int amdgpu_gem_va_ioctl(struct drm_device *dev, void *data,
 	return r;
 }
 
+/**
+ * amdgpu_gem_list_mappings - get information about a buffer's mappings
+ *
+ * @gobj: gem object
+ * @args: gem_op arguments
+ * @fpriv: drm file pointer
+ *
+ * num_entries is set as an input to the size of the user-allocated array of
+ * drm_amdgpu_gem_vm_entry stored at args->value.
+ * num_entries is sent back as output as the number of mappings the bo has.
+ * If that number is larger than the size of the array, the ioctl must
+ * be retried.
+ *
+ * Returns:
+ * 0 for success, -errno for errors.
+ */
+static int amdgpu_gem_list_mappings(struct drm_gem_object *gobj, struct amdgpu_fpriv *fpriv,
+					  struct drm_amdgpu_gem_op *args)
+{
+	struct amdgpu_vm *avm = &fpriv->vm;
+	struct amdgpu_bo *bo = gem_to_amdgpu_bo(gobj);
+	struct amdgpu_bo_va *bo_va = amdgpu_vm_bo_find(avm, bo);
+	struct drm_amdgpu_gem_vm_entry *vm_entries;
+	struct amdgpu_bo_va_mapping *mapping;
+	struct drm_exec exec;
+	int num_mappings = 0;
+	int ret;
+
+	if (args->padding)
+		return -EINVAL;
+
+	vm_entries = kvcalloc(args->num_entries, sizeof(*vm_entries), GFP_KERNEL);
+	if (!vm_entries)
+		return -ENOMEM;
+
+	drm_exec_init(&exec, DRM_EXEC_INTERRUPTIBLE_WAIT |
+			  DRM_EXEC_IGNORE_DUPLICATES, 0);
+	drm_exec_until_all_locked(&exec) {
+		if (gobj) {
+			ret = drm_exec_lock_obj(&exec, gobj);
+			drm_exec_retry_on_contention(&exec);
+			if (ret)
+				goto unlock_exec;
+		}
+
+		ret = amdgpu_vm_lock_pd(&fpriv->vm, &exec, 0);
+		drm_exec_retry_on_contention(&exec);
+		if (ret)
+			goto unlock_exec;
+	}
+
+	amdgpu_vm_bo_va_for_each_valid_mapping(bo_va, mapping) {
+		if (num_mappings < args->num_entries) {
+			vm_entries[num_mappings].start = mapping->start;
+			vm_entries[num_mappings].last = mapping->last;
+			vm_entries[num_mappings].offset = mapping->offset;
+			vm_entries[num_mappings].flags = mapping->flags;
+		}
+		num_mappings += 1;
+	}
+
+	amdgpu_vm_bo_va_for_each_invalid_mapping(bo_va, mapping) {
+		if (num_mappings < args->num_entries) {
+			vm_entries[num_mappings].start = mapping->start;
+			vm_entries[num_mappings].last = mapping->last;
+			vm_entries[num_mappings].offset = mapping->offset;
+			vm_entries[num_mappings].flags = mapping->flags;
+		}
+		num_mappings += 1;
+	}
+
+	if (num_mappings > 0 && num_mappings <= args->num_entries)
+		ret = copy_to_user(u64_to_user_ptr(args->value), vm_entries, num_mappings * sizeof(*vm_entries));
+
+	args->num_entries = num_mappings;
+
+unlock_exec:
+	drm_exec_fini(&exec);
+	kvfree(vm_entries);
+
+	return ret;
+}
+
 int amdgpu_gem_op_ioctl(struct drm_device *dev, void *data,
 			struct drm_file *filp)
 {
@@ -1014,6 +1097,10 @@ int amdgpu_gem_op_ioctl(struct drm_device *dev, void *data,
 
 		amdgpu_bo_unreserve(robj);
 		break;
+	case AMDGPU_GEM_OP_GET_MAPPING_INFO:
+		amdgpu_bo_unreserve(robj);
+		r = amdgpu_gem_list_mappings(gobj, filp->driver_priv, args);
+		break;
 	default:
 		amdgpu_bo_unreserve(robj);
 		r = -EINVAL;
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +5 -0
index f9549f6b3d1f..5a63ae490b0e 100644
@@ -668,4 +668,9 @@ void amdgpu_vm_tlb_fence_create(struct amdgpu_device *adev,
 				 struct amdgpu_vm *vm,
 				 struct dma_fence **fence);
 
+#define amdgpu_vm_bo_va_for_each_valid_mapping(bo_va, mapping) \
+		list_for_each_entry(mapping, &bo_va->valids, list)
+#define amdgpu_vm_bo_va_for_each_invalid_mapping(bo_va, mapping) \
+		list_for_each_entry(mapping, &bo_va->invalids, list)
+
 #endif
include/uapi/drm/amdgpu_drm.h +20 -1
index 59b423883e91..bc1e18c06ce4 100644
@@ -802,6 +802,21 @@ union drm_amdgpu_wait_fences {
 
 #define AMDGPU_GEM_OP_GET_GEM_CREATE_INFO	0
 #define AMDGPU_GEM_OP_SET_PLACEMENT		1
+#define AMDGPU_GEM_OP_GET_MAPPING_INFO		2
+
+struct drm_amdgpu_gem_vm_entry {
+	/* Start of mapping (in number of pages) */
+	__u64 start;
+
+	/* End of mapping (in number of pages) */
+	__u64 last;
+
+	/* Mapping offset */
+	__u64 offset;
+
+	/* flags needed to recreate mapping */
+	__u64 flags;
+};
 
 /* Sets or returns a value associated with a buffer. */
 struct drm_amdgpu_gem_op {
@@ -809,8 +824,12 @@ struct drm_amdgpu_gem_op {
 	__u32	handle;
 	/** AMDGPU_GEM_OP_* */
 	__u32	op;
-	/** Input or return value */
+	/** Input or return value. For MAPPING_INFO op: pointer to array of struct drm_amdgpu_gem_vm_entry */
 	__u64	value;
+	/** For MAPPING_INFO op: number of mappings (in/out) */
+	__u32	num_entries;
+
+	__u32	padding;
 };
 
 #define AMDGPU_GEM_LIST_HANDLES_FLAG_IS_IMPORT	(1 << 0)
Signature · 1 line
2.34.1
maintainer
#9 lore raw

Re: [PATCH v14 3/4] drm/amdgpu: Add mapping info option for GEM_OP ioctl

Quoted 65 lines · David Francis

Add new GEM_OP_IOCTL option GET_MAPPING_INFO, which returns a list of mappings associated with a given bo, along with their positions and offsets.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 87 +++++++++++++++++++++++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h  |  5 ++
 include/uapi/drm/amdgpu_drm.h           | 21 +++++-
 3 files changed, 112 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +46 -0
index d17cc7ba66d4..f66f6e1f1c52 100644
@@ -948,6 +948,89 @@ int amdgpu_gem_va_ioctl(struct drm_device *dev, void *data,
 	return r;
 }
 
+/**
+ * amdgpu_gem_list_mappings - get information about a buffer's mappings
+ *
+ * @gobj: gem object
+ * @args: gem_op arguments
+ * @fpriv: drm file pointer
+ *
+ * num_entries is set as an input to the size of the user-allocated array of
+ * drm_amdgpu_gem_vm_entry stored at args->value.
+ * num_entries is sent back as output as the number of mappings the bo has.
+ * If that number is larger than the size of the array, the ioctl must
+ * be retried.
+ *
+ * Returns:
+ * 0 for success, -errno for errors.
+ */
+static int amdgpu_gem_list_mappings(struct drm_gem_object *gobj, struct amdgpu_fpriv *fpriv,
+					  struct drm_amdgpu_gem_op *args)
+{
+	struct amdgpu_vm *avm = &fpriv->vm;
+	struct amdgpu_bo *bo = gem_to_amdgpu_bo(gobj);
+	struct amdgpu_bo_va *bo_va = amdgpu_vm_bo_find(avm, bo);
+	struct drm_amdgpu_gem_vm_entry *vm_entries;
+	struct amdgpu_bo_va_mapping *mapping;
+	struct drm_exec exec;
+	int num_mappings = 0;
+	int ret;
+
+	if (args->padding)
+		return -EINVAL;
+
+	vm_entries = kvcalloc(args->num_entries, sizeof(*vm_entries), GFP_KERNEL);
+	if (!vm_entries)
+		return -ENOMEM;
+
+	drm_exec_init(&exec, DRM_EXEC_INTERRUPTIBLE_WAIT |
+			  DRM_EXEC_IGNORE_DUPLICATES, 0);
+	drm_exec_until_all_locked(&exec) {
+		if (gobj) {
+			ret = drm_exec_lock_obj(&exec, gobj);
+			drm_exec_retry_on_contention(&exec);
+			if (ret)
+				goto unlock_exec;
+		}
+
+		ret = amdgpu_vm_lock_pd(&fpriv->vm, &exec, 0);

drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c int amdgpu_gem_va_ioctl(struct drm_device *dev, void *data, · line 1006

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +9 -0
+		drm_exec_retry_on_contention(&exec);
+		if (ret)
+			goto unlock_exec;
+	}
+
+	amdgpu_vm_bo_va_for_each_valid_mapping(bo_va, mapping) {
+		if (num_mappings < args->num_entries) {
 
+			vm_entries[num_mappings].start = mapping->start;
+			vm_entries[num_mappings].last = mapping->last;

Those two needs to be multiplied by AMDGPU_GPU_PAGE_SIZE.

drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c int amdgpu_gem_va_ioctl(struct drm_device *dev, void *data, · line 1015

Quoted 11 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +10 -0
+			vm_entries[num_mappings].offset = mapping->offset;
+			vm_entries[num_mappings].flags = mapping->flags;
+		}
+		num_mappings += 1;
+	}
+
+	amdgpu_vm_bo_va_for_each_invalid_mapping(bo_va, mapping) {
+		if (num_mappings < args->num_entries) {
 
+			vm_entries[num_mappings].start = mapping->start;
+			vm_entries[num_mappings].last = mapping->last;

Here as well.

Quoted 22 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +18 -0
+			vm_entries[num_mappings].offset = mapping->offset;
+			vm_entries[num_mappings].flags = mapping->flags;
+		}
+		num_mappings += 1;
+	}
+
+	if (num_mappings > 0 && num_mappings <= args->num_entries)
+		ret = copy_to_user(u64_to_user_ptr(args->value), vm_entries, num_mappings * sizeof(*vm_entries));
+
+	args->num_entries = num_mappings;
+
+unlock_exec:
+	drm_exec_fini(&exec);
+	kvfree(vm_entries);
+
+	return ret;
+}
+
 int amdgpu_gem_op_ioctl(struct drm_device *dev, void *data,
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +0 -0
 			struct drm_file *filp)
 {

drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c int amdgpu_gem_op_ioctl(struct drm_device *dev, void *data, · line 1101

Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +2 -0
@@ -1014,6 +1097,10 @@ int amdgpu_gem_op_ioctl(struct drm_device *dev, void *data,
 
 		amdgpu_bo_unreserve(robj);
 		break;
+	case AMDGPU_GEM_OP_GET_MAPPING_INFO:
+		amdgpu_bo_unreserve(robj);

Yeah that is not really 100% clean.

It would be better if we use drm_exec() in the amdgpu_gem_op_ioctl().

Quoted 25 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +2 -0
+		r = amdgpu_gem_list_mappings(gobj, filp->driver_priv, args);
+		break;
 	default:
 		amdgpu_bo_unreserve(robj);
 		r = -EINVAL;
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +5 -0
index f9549f6b3d1f..5a63ae490b0e 100644
@@ -668,4 +668,9 @@ void amdgpu_vm_tlb_fence_create(struct amdgpu_device *adev,
 				 struct amdgpu_vm *vm,
 				 struct dma_fence **fence);
 
+#define amdgpu_vm_bo_va_for_each_valid_mapping(bo_va, mapping) \
+		list_for_each_entry(mapping, &bo_va->valids, list)
+#define amdgpu_vm_bo_va_for_each_invalid_mapping(bo_va, mapping) \
+		list_for_each_entry(mapping, &bo_va->invalids, list)
+
 #endif
include/uapi/drm/amdgpu_drm.h +0 -0
index 59b423883e91..bc1e18c06ce4 100644
@@ -802,6 +802,21 @@ union drm_amdgpu_wait_fences {
 

include/uapi/drm/amdgpu_drm.h union drm_amdgpu_wait_fences { · line 812

Quoted 10 lines
include/uapi/drm/amdgpu_drm.h +8 -0
 #define AMDGPU_GEM_OP_GET_GEM_CREATE_INFO	0
 #define AMDGPU_GEM_OP_SET_PLACEMENT		1
+#define AMDGPU_GEM_OP_GET_MAPPING_INFO		2
+
+struct drm_amdgpu_gem_vm_entry {
+	/* Start of mapping (in number of pages) */
+	__u64 start;
+
+	/* End of mapping (in number of pages) */
+	__u64 last;

Size please, not the end. To match the UAPI to create mappings.

Quoted 24 lines
include/uapi/drm/amdgpu_drm.h +12 -1
+
+	/* Mapping offset */
+	__u64 offset;
+
+	/* flags needed to recreate mapping */
+	__u64 flags;
+};
 
 /* Sets or returns a value associated with a buffer. */
 struct drm_amdgpu_gem_op {
@@ -809,8 +824,12 @@ struct drm_amdgpu_gem_op {
 	__u32	handle;
 	/** AMDGPU_GEM_OP_* */
 	__u32	op;
-	/** Input or return value */
+	/** Input or return value. For MAPPING_INFO op: pointer to array of struct drm_amdgpu_gem_vm_entry */
 	__u64	value;
+	/** For MAPPING_INFO op: number of mappings (in/out) */
+	__u32	num_entries;
+
+	__u32	padding;
 };
 
 #define AMDGPU_GEM_LIST_HANDLES_FLAG_IS_IMPORT	(1 << 0)
#10 lore raw

PATCH v14 4/4 drm/amdgpu: Allow kfd CRIU with no buffer objects

In reply to David Francis

The kfd CRIU checkpoint ioctl would return an error if trying to checkpoint a process with no kfd buffer objects.

This is a normal case and should not be an error.

---
 drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +2 -2
index 828a9ceef1e7..f7f34b710d3e 100644
@@ -2566,8 +2566,8 @@ static int criu_restore(struct file *filep,
 	pr_debug("CRIU restore (num_devices:%u num_bos:%u num_objects:%u priv_data_size:%llu)\n",
 		 args->num_devices, args->num_bos, args->num_objects, args->priv_data_size);
 
-	if (!args->bos || !args->devices || !args->priv_data || !args->priv_data_size ||
-	    !args->num_devices || !args->num_bos)
+	if ((args->num_bos > 0 && !args->bos) || !args->devices || !args->priv_data ||
+	    !args->priv_data_size || !args->num_devices)
 		return -EINVAL;
 
 	mutex_lock(&p->mutex);
Signature · 1 line
2.34.1
maintainer
#11 lore raw
In reply to David Francis

Going over the whole thing once more I've found another design issue.

How do you re-create the CPU mappings of buffers?

As far as I can see the buffers end up on a different VMA and we don't have any way of restoring that at the moment.

See here as well https://www.kernel.org/doc/html/v4.15/gpu/drm-mm.html#vma-offset-manager

Regards,
Christian.

Quoted 10 lines · David Francis

This patch series adds support for CRIU checkpointing of processes that share memory with the amdgpu dmabuf interface.

This v14 adds the coherence and uncached flags to gem create and fixes some formatting issues in amdgpu_drm.h

Accompanying CRIU changes:
https://github.com/checkpoint-restore/criu/pull/2613