Skip to contentSkip to conversation

Add CRIU support for amdgpu dmabuf

PATCH Superseded David Francis 3 participants 7 messages Aug 7, 2025 – Aug 11, 2025

Subsystems (4)

v11 of 16 Superseded Latest is v16 →

lore
Share & tools
lore
#0 lore raw

PATCH v11 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 v11 fixes some locking and formatting issues identified by Christian

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

#1 lore raw

PATCH v11 1/3 drm/amdgpu: Add ioctl to get all gem handles for a process

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 | 82 +++++++++++++++++++++++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_gem.h |  2 +
 include/uapi/drm/amdgpu_drm.h           | 33 ++++++++++
 4 files changed, 118 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 +82 -0

88 lines not on this page · raw message

drivers/gpu/drm/amd/amdgpu/amdgpu_gem.h +2 -0
index b51e8f95ee86..7cdb6237bb92 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 +33 -0
index bdedbaccf776..218a179818d4 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 +813,37 @@ 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;
+
+	/* Pending how to handle this; provides information needed to remake the buffer on restore */
+	__u32 preferred_domains;
+
+	/* Size of bo */
+	__u64 size;
+
+	/* GEM_CREATE flags for re-creation of buffer */
+	__u64 alloc_flags;
+
+	/* Currently just one flag: IS_IMPORT */
+	__u32 flags;
+
+	__u32 padding;
+};
+
 #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
#2 lore raw

Re: [PATCH v11 1/3] drm/amdgpu: Add ioctl to get all gem handles for a process

Quoted 79 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 | 82 +++++++++++++++++++++++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_gem.h |  2 +
 include/uapi/drm/amdgpu_drm.h           | 33 ++++++++++
 4 files changed, 118 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 +42 -0
index e3f65977eeee..7f55e3b7d8a2 100644
@@ -1032,6 +1032,88 @@ 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_bos is set as an input to the size of the bo_buckets array.
+ * num_bos 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;
+	}
+	args->num_entries = num_bos;
+	if (num_bos == 0)
+		return 0;
+
+	bo_entries = kvcalloc(num_bos, sizeof(*bo_entries), GFP_KERNEL);
+	if (!bo_entries) {
+		ret = -ENOMEM;
+		goto exit_free;

drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c int amdgpu_gem_op_ioctl(struct drm_device *dev, void *data, · line 1086

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +10 -0
+	}
+
+	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 = -EINVAL;
+			break;

spin_unlock() + goto exit_free.

Apart from that looks good to me. Feel free to add my rb, but I think Felix and others might want to take a look as well.

Regards,
Christian.

Quoted 104 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +30 -0
+		}
+
+		bo_entry = &bo_entries[bo_index];
+
+		bo_entry->size = amdgpu_bo_size(bo);
+		bo_entry->alloc_flags = bo->flags & (~AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE);
+		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);
+
+	if (!ret) {
+		ret = copy_to_user((void __user *)args->entries, bo_entries, num_bos * sizeof(*bo_entries));
+		if (ret) {
+			pr_debug("Failed to copy BO information to user\n");
+			ret = -EFAULT;
+		}
+	}
+exit_free:
+	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 b51e8f95ee86..7cdb6237bb92 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 +33 -0
index bdedbaccf776..218a179818d4 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 +813,37 @@ 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;
+
+	/* Pending how to handle this; provides information needed to remake the buffer on restore */
+	__u32 preferred_domains;
+
+	/* Size of bo */
+	__u64 size;
+
+	/* GEM_CREATE flags for re-creation of buffer */
+	__u64 alloc_flags;
+
+	/* Currently just one flag: IS_IMPORT */
+	__u32 flags;
+
+	__u32 padding;
+};
+
 #define AMDGPU_VA_OP_MAP			1
 #define AMDGPU_VA_OP_UNMAP			2
 #define AMDGPU_VA_OP_CLEAR			3
#3 lore raw

Re: [PATCH v11 1/3] drm/amdgpu: Add ioctl to get all gem handles for a process

In reply to David Francis
Quoted 62 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 | 82 +++++++++++++++++++++++++
  drivers/gpu/drm/amd/amdgpu/amdgpu_gem.h |  2 +
  include/uapi/drm/amdgpu_drm.h           | 33 ++++++++++
  4 files changed, 118 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 +25 -0
index e3f65977eeee..7f55e3b7d8a2 100644
@@ -1032,6 +1032,88 @@ 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_bos is set as an input to the size of the bo_buckets array.
+ * num_bos 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;
+

drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c int amdgpu_gem_op_ioctl(struct drm_device *dev, void *data, · line 1069

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +10 -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;
+	}
+	args->num_entries = num_bos;

Looks like authoritative number is only obtained in the second iteration so I think it should be assigned only there.

At least the uapi doc does not suggest ("OUT: Number of handles in process (if larger than size of buffer, must retry)" that the count can be larger than the actual number of handles, and that handles must also be tested for being non-zero. If that is the intention it works as is but needs to be documented in the uapi.

drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c int amdgpu_gem_op_ioctl(struct drm_device *dev, void *data, · line 1076

Quoted 7 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +7 -0
+	if (num_bos == 0)
+		return 0;
+
+	bo_entries = kvcalloc(num_bos, sizeof(*bo_entries), GFP_KERNEL);
+	if (!bo_entries) {
+		ret = -ENOMEM;
+		goto exit_free;

return -ENOMEM is enough here.

Quoted 6 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +6 -0
+	}
+
+	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;

drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c int amdgpu_gem_op_ioctl(struct drm_device *dev, void *data, · line 1092

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +10 -0
+
+		if (bo_index >= num_bos) {
+			ret = -EINVAL;
+			break;
+		}
+
+		bo_entry = &bo_entries[bo_index];
+
+		bo_entry->size = amdgpu_bo_size(bo);
+		bo_entry->alloc_flags = bo->flags & (~AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE);

Why is this flag special and why it is okay to not restore it?

drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c int amdgpu_gem_op_ioctl(struct drm_device *dev, void *data, · line 1097

Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +5 -0
+		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;

Imported ones are only informative (since they will need a completely different path to restore)? But it is required to list them? Do you have the restore side implemented for reference?

drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c int amdgpu_gem_op_ioctl(struct drm_device *dev, void *data, · line 1104

Quoted 7 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +7 -0
+
+		bo_index += 1;
+	}
+	spin_unlock(&filp->table_lock);
+
+	if (!ret) {
+		ret = copy_to_user((void __user *)args->entries, bo_entries, num_bos * sizeof(*bo_entries));

u64_to_user_ptr()

drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c int amdgpu_gem_op_ioctl(struct drm_device *dev, void *data, · line 1106

Quoted 2 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +2 -0
+		if (ret) {
+			pr_debug("Failed to copy BO information to user\n");

I would be tempted to remove this pr_debug since -EFAULT is pretty clear anyway. In which case this whole if (ret) block seems superflous.

Quoted 58 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +10 -0
+			ret = -EFAULT;
+		}
+	}
+exit_free:
+	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 b51e8f95ee86..7cdb6237bb92 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 +10 -0
index bdedbaccf776..218a179818d4 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 +813,37 @@ 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;

include/uapi/drm/amdgpu_drm.h struct drm_amdgpu_gem_op { · line 833

Quoted 10 lines
include/uapi/drm/amdgpu_drm.h +10 -0
+
+	__u32 padding;
+};
+
+struct drm_amdgpu_gem_list_handles_entry {
+	/* gem handle of buffer object */
+	__u32 gem_handle;
+
+	/* Pending how to handle this; provides information needed to remake the buffer on restore */
+	__u32 preferred_domains;

Hm, just as an observation, the width for these is not very well defined in the uapi. Yes flags are just a few, but for example drm_amdgpu_gem_create_in has them as u64 and also value via AMDGPU_GEM_OP_SET_PLACEMENT. So maybe future proofing to __u64 wouldn't harm here. It would also remove the need to pad the structure.

Regards,

Tvrtko

Quoted 16 lines
include/uapi/drm/amdgpu_drm.h +13 -0
+
+	/* Size of bo */
+	__u64 size;
+
+	/* GEM_CREATE flags for re-creation of buffer */
+	__u64 alloc_flags;
+
+	/* Currently just one flag: IS_IMPORT */
+	__u32 flags;
+
+	__u32 padding;
+};
+
  #define AMDGPU_VA_OP_MAP			1
  #define AMDGPU_VA_OP_UNMAP			2
  #define AMDGPU_VA_OP_CLEAR			3
#4 lore raw

PATCH v11 2/3 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 | 96 +++++++++++++++++++++++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h  |  5 ++
 include/uapi/drm/amdgpu_drm.h           | 18 +++++
 3 files changed, 119 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +96 -0
index 7f55e3b7d8a2..29ee1df37661 100644
@@ -956,6 +956,98 @@ 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_bucket 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_bucket *vm_buckets;
+	struct amdgpu_bo_va_mapping *mapping;
+	struct drm_exec exec;
+	int num_mappings = 0;
+	int ret;
+
+	vm_buckets = kvcalloc(args->num_entries, sizeof(*vm_buckets), GFP_KERNEL);
+	if (!vm_buckets) {
+		ret = -ENOMEM;
+		goto free_vms;
+	}
+
+	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, 2);
+		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_buckets[num_mappings].start = mapping->start;
+			vm_buckets[num_mappings].last = mapping->last;
+			vm_buckets[num_mappings].offset = mapping->offset;
+			vm_buckets[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_buckets[num_mappings].start = mapping->start;
+			vm_buckets[num_mappings].last = mapping->last;
+			vm_buckets[num_mappings].offset = mapping->offset;
+			vm_buckets[num_mappings].flags = mapping->flags;
+		}
+		num_mappings += 1;
+	}
+
+	drm_exec_fini(&exec);
+
+	if (num_mappings > 0 && num_mappings <= args->num_entries) {
+		ret = copy_to_user((void __user *)args->value, vm_buckets, num_mappings * sizeof(*vm_buckets));
+		if (ret) {
+			pr_debug("Failed to copy BO information to user\n");
+			ret = -EFAULT;
+		}
+	}
+	args->num_entries = num_mappings;
+
+	kvfree(vm_buckets);
+
+	return ret;
+unlock_exec:
+	drm_exec_fini(&exec);
+free_vms:
+	kvfree(vm_buckets);
+
+	return ret;
+}
+
 int amdgpu_gem_op_ioctl(struct drm_device *dev, void *data,
 			struct drm_file *filp)
 {
@@ -1022,6 +1114,10 @@ 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 +18 -0
index 218a179818d4..6b857f528823 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_bucket {
+	/* 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 {
@@ -811,6 +826,9 @@ struct drm_amdgpu_gem_op {
 	__u32	op;
 	/** Input or return value */
 	__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
#5 lore raw

Re: [PATCH v11 2/3] drm/amdgpu: Add mapping info option for GEM_OP ioctl

Quoted 41 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 | 96 +++++++++++++++++++++++++
  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h  |  5 ++
  include/uapi/drm/amdgpu_drm.h           | 18 +++++
  3 files changed, 119 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +22 -0
index 7f55e3b7d8a2..29ee1df37661 100644
@@ -956,6 +956,98 @@ 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_bucket 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);

drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c int amdgpu_gem_va_ioctl(struct drm_device *dev, void *data, · line 990

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +10 -0
+	struct drm_amdgpu_gem_vm_bucket *vm_buckets;
+	struct amdgpu_bo_va_mapping *mapping;
+	struct drm_exec exec;
+	int num_mappings = 0;
+	int ret;
+
+	vm_buckets = kvcalloc(args->num_entries, sizeof(*vm_buckets), GFP_KERNEL);
+	if (!vm_buckets) {
+		ret = -ENOMEM;
+		goto free_vms;

return -ENOMEM would be enough.

drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c int amdgpu_gem_va_ioctl(struct drm_device *dev, void *data, · line 1003

Quoted 13 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +13 -0
+	}
+
+	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, 2);

What are the two reserved fence slots for?

Quoted 23 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +23 -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_buckets[num_mappings].start = mapping->start;
+			vm_buckets[num_mappings].last = mapping->last;
+			vm_buckets[num_mappings].offset = mapping->offset;
+			vm_buckets[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_buckets[num_mappings].start = mapping->start;
+			vm_buckets[num_mappings].last = mapping->last;
+			vm_buckets[num_mappings].offset = mapping->offset;
+			vm_buckets[num_mappings].flags = mapping->flags;
+		}
+		num_mappings += 1;

drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c int amdgpu_gem_va_ioctl(struct drm_device *dev, void *data, · line 1036

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +10 -0
+	}
+
+	drm_exec_fini(&exec);
+
+	if (num_mappings > 0 && num_mappings <= args->num_entries) {
+		ret = copy_to_user((void __user *)args->value, vm_buckets, num_mappings * sizeof(*vm_buckets));
+		if (ret) {
+			pr_debug("Failed to copy BO information to user\n");
+			ret = -EFAULT;
+		}

Same comment as in the first patch - u64_to_user_pointer and I'd consider dropping the pr_debug path.

drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c int amdgpu_gem_va_ioctl(struct drm_device *dev, void *data, · line 1048

Quoted 12 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +12 -0
+	}
+	args->num_entries = num_mappings;
+
+	kvfree(vm_buckets);
+
+	return ret;
+unlock_exec:
+	drm_exec_fini(&exec);
+free_vms:
+	kvfree(vm_buckets);
+
+	return ret;

I think the success/failure exit paths could be consolidated:

	amdgpu_vm_bo_va_for_each_invalid_mapping(bo_va, mapping) {
		...
	}

	args->num_entries = num_mappings;

	if (num_mappings > 0 && num_mappings <= args->num_entries)
		ret = copy_to_user(...);

unlock_exec:

	drm_exec_fini(&exec);
	kvfree(vm_buckets);

	return ret;

Regards,

Tvrtko

Quoted 66 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +6 -0
+}
+
  int amdgpu_gem_op_ioctl(struct drm_device *dev, void *data,
  			struct drm_file *filp)
  {
@@ -1022,6 +1114,10 @@ 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 +18 -0
index 218a179818d4..6b857f528823 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_bucket {
+	/* 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 {
@@ -811,6 +826,9 @@ struct drm_amdgpu_gem_op {
  	__u32	op;
  	/** Input or return value */
  	__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)
#6 lore raw

PATCH v11 3/3 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