Skip to contentSkip to conversation

drm/amdgpu: check if vram is lost v2

PATCH In review Chunming Zhou 5 participants 22 messages May 16, 2017 – May 24, 2017 17 messages not in the archive

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/4 drm/amdgpu: check if vram is lost v2

bakup first 64 byte of gart table as reset magic, check if magic is same after gpu hw reset.
v2: use memcmp instead of manual innovation.

Change-Id I9a73720da4084ea8677c3031dfb62e8157ee5704
---
 drivers/gpu/drm/amd/amdgpu/amdgpu.h        |  2 ++
 drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 20 +++++++++++++++++++-
 2 files changed, 21 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +2 -0
index de08ff0..f9da215 100644
@@ -1502,6 +1502,7 @@ struct amdgpu_ssg {
 #endif
 };
 
+#define AMDGPU_RESET_MAGIC_NUM 64
 struct amdgpu_device {
 	struct device			*dev;
 	struct drm_device		*ddev;
@@ -1705,6 +1706,7 @@ struct amdgpu_device {
 
 	/* record hw reset is performed */
 	bool has_hw_reset;
+	u8				reset_magic[AMDGPU_RESET_MAGIC_NUM];
 
 };
 
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +19 -1
index 0a31fb1..c56ae4a 100644
@@ -1685,6 +1685,17 @@ static int amdgpu_init(struct amdgpu_device *adev)
 	return 0;
 }
 
+static void amdgpu_fill_reset_magic(struct amdgpu_device *adev)
+{
+	memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM);
+}
+
+static bool amdgpu_check_vram_lost(struct amdgpu_device *adev)
+{
+	return !!memcmp(adev->gart.ptr, adev->reset_magic,
+			AMDGPU_RESET_MAGIC_NUM);
+}
+
 static int amdgpu_late_init(struct amdgpu_device *adev)
 {
 	int i = 0, r;
@@ -1715,6 +1726,8 @@ static int amdgpu_late_init(struct amdgpu_device *adev)
 		}
 	}
 
+	amdgpu_fill_reset_magic(adev);
+
 	return 0;
 }
 
@@ -2830,7 +2843,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
 	struct drm_atomic_state *state = NULL;
 	int i, r;
 	int resched;
-	bool need_full_reset;
+	bool need_full_reset, vram_lost = false;
 
 	if (amdgpu_sriov_vf(adev))
 		return amdgpu_sriov_gpu_reset(adev, true);
@@ -2899,12 +2912,17 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
 			r = amdgpu_resume_phase1(adev);
 			if (r)
 				goto out;
+			vram_lost = amdgpu_check_vram_lost(adev);
+			if (vram_lost)
+				DRM_ERROR("VRAM is lost!\n");
 			r = amdgpu_ttm_recover_gart(adev);
 			if (r)
 				goto out;
 			r = amdgpu_resume_phase2(adev);
 			if (r)
 				goto out;
+			if (vram_lost)
+				amdgpu_fill_reset_magic(adev);
 		}
 	}
 out:
Signature · 6 lines
1.9.1

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 1494926750-1081-1-git-send-email-David1.Zhou-5C7GfCeVMHo@public.gmane.org

#2 lore raw

PATCH 2/4 drm/amdgpu: return -ENODEV to user space when vram is lost v2

below ioctl will return -ENODEV:
amdgpu_cs_ioctl
amdgpu_cs_wait_ioctl
amdgpu_cs_wait_fences_ioctl
amdgpu_gem_va_ioctl
amdgpu_info_ioctl

v2: only for map and replace cases in amdgpu_gem_va_ioctl

Change-Id I8970cde3301b7cfeb4263cc0f0e54aece215c98e
---
 drivers/gpu/drm/amd/amdgpu/amdgpu.h        |  4 ++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c     |  9 +++++++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_device.c |  4 +++-
 drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c    |  5 +++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c    | 10 ++++++++++
 5 files changed, 31 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +4 -0
index f9da215..dcd6203 100644
@@ -855,6 +855,7 @@ struct amdgpu_fpriv {
 	struct amdgpu_ctx_mgr	ctx_mgr;
 	spinlock_t		sem_handles_lock;
 	struct idr		sem_handles;
+	u32			vram_lost_counter;
 };
 
 /*
@@ -1607,6 +1608,7 @@ struct amdgpu_device {
 	atomic64_t			num_bytes_moved;
 	atomic64_t			num_evictions;
 	atomic_t			gpu_reset_counter;
+	atomic_t			vram_lost_counter;
 
 	/* data for buffer migration throttling */
 	struct {
@@ -2005,6 +2007,8 @@ static inline void amdgpu_unregister_atpx_handler(void) {}
 extern const struct drm_ioctl_desc amdgpu_ioctls_kms[];
 extern const int amdgpu_max_kms_ioctl;
 
+bool amdgpu_kms_vram_lost(struct amdgpu_device *adev,
+			  struct amdgpu_fpriv *fpriv);
 int amdgpu_driver_load_kms(struct drm_device *dev, unsigned long flags);
 int amdgpu_driver_unload_kms(struct drm_device *dev);
 void amdgpu_driver_lastclose_kms(struct drm_device *dev);
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +9 -0
index b803412..911aa02 100644
@@ -1097,6 +1097,7 @@ static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
 int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
 {
 	struct amdgpu_device *adev = dev->dev_private;
+	struct amdgpu_fpriv *fpriv = filp->driver_priv;
 	union drm_amdgpu_cs *cs = data;
 	struct amdgpu_cs_parser parser = {};
 	bool reserved_buffers = false;
@@ -1104,6 +1105,8 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
 
 	if (!adev->accel_working)
 		return -EBUSY;
+	if (amdgpu_kms_vram_lost(adev, fpriv))
+		return -ENODEV;
 
 	parser.adev = adev;
 	parser.filp = filp;
@@ -1165,12 +1168,15 @@ int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data,
 {
 	union drm_amdgpu_wait_cs *wait = data;
 	struct amdgpu_device *adev = dev->dev_private;
+	struct amdgpu_fpriv *fpriv = filp->driver_priv;
 	unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
 	struct amdgpu_ring *ring = NULL;
 	struct amdgpu_ctx *ctx;
 	struct fence *fence;
 	long r;
 
+	if (amdgpu_kms_vram_lost(adev, fpriv))
+		return -ENODEV;
 	r = amdgpu_cs_get_ring(adev, wait->in.ip_type, wait->in.ip_instance,
 			       wait->in.ring, &ring);
 	if (r)
@@ -1344,12 +1350,15 @@ int amdgpu_cs_wait_fences_ioctl(struct drm_device *dev, void *data,
 				struct drm_file *filp)
 {
 	struct amdgpu_device *adev = dev->dev_private;
+	struct amdgpu_fpriv *fpriv = filp->driver_priv;
 	union drm_amdgpu_wait_fences *wait = data;
 	uint32_t fence_count = wait->in.fence_count;
 	struct drm_amdgpu_fence *fences_user;
 	struct drm_amdgpu_fence *fences;
 	int r;
 
+	if (amdgpu_kms_vram_lost(adev, fpriv))
+		return -ENODEV;
 	/* Get the fences from userspace */
 	fences = kmalloc_array(fence_count, sizeof(struct drm_amdgpu_fence),
 			GFP_KERNEL);
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +3 -1
index c56ae4a..2f0fcf8 100644
@@ -2913,8 +2913,10 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
 			if (r)
 				goto out;
 			vram_lost = amdgpu_check_vram_lost(adev);
-			if (vram_lost)
+			if (vram_lost) {
 				DRM_ERROR("VRAM is lost!\n");
+				atomic_inc(&adev->vram_lost_counter);
+			}
 			r = amdgpu_ttm_recover_gart(adev);
 			if (r)
 				goto out;
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +5 -0
index d8275ef..83bc94c 100644
@@ -802,6 +802,11 @@ int amdgpu_gem_va_ioctl(struct drm_device *dev, void *data,
 			args->operation);
 		return -EINVAL;
 	}
+	if ((args->operation == AMDGPU_VA_OP_MAP) ||
+	    (args->operation == AMDGPU_VA_OP_REPLACE)) {
+		if (amdgpu_kms_vram_lost(adev, fpriv))
+			return -ENODEV;
+	}
 
 	INIT_LIST_HEAD(&list);
 	if ((args->operation != AMDGPU_VA_OP_CLEAR) &&
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +10 -0
index 368829a..a231aa1 100644
@@ -235,6 +235,7 @@ static int amdgpu_firmware_info(struct drm_amdgpu_info_firmware *fw_info,
 static int amdgpu_info_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
 {
 	struct amdgpu_device *adev = dev->dev_private;
+	struct amdgpu_fpriv *fpriv = filp->driver_priv;
 	struct drm_amdgpu_info *info = data;
 	struct amdgpu_mode_info *minfo = &adev->mode_info;
 	void __user *out = (void __user *)(uintptr_t)info->return_pointer;
@@ -247,6 +248,8 @@ static int amdgpu_info_ioctl(struct drm_device *dev, void *data, struct drm_file
 
 	if (!info->return_size || !info->return_pointer)
 		return -EINVAL;
+	if (amdgpu_kms_vram_lost(adev, fpriv))
+		return -ENODEV;
 
 	switch (info->query) {
 	case AMDGPU_INFO_VIRTUAL_RANGE: {
@@ -779,6 +782,12 @@ void amdgpu_driver_lastclose_kms(struct drm_device *dev)
 	vga_switcheroo_process_delayed_switch();
 }
 
+bool amdgpu_kms_vram_lost(struct amdgpu_device *adev,
+			  struct amdgpu_fpriv *fpriv)
+{
+	return fpriv->vram_lost_counter != atomic_read(&adev->vram_lost_counter);
+}
+
 /**
  * amdgpu_driver_open_kms - drm callback for open
  *
@@ -833,6 +842,7 @@ int amdgpu_driver_open_kms(struct drm_device *dev, struct drm_file *file_priv)
 
 	amdgpu_ctx_mgr_init(&fpriv->ctx_mgr);
 
+	fpriv->vram_lost_counter = atomic_read(&adev->vram_lost_counter);
 	file_priv->driver_priv = fpriv;
 
 out_suspend:
Signature · 6 lines
1.9.1

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 1494926750-1081-2-git-send-email-David1.Zhou-5C7GfCeVMHo@public.gmane.org

#4 lore raw

RE: [PATCH 2/4] drm/amdgpu: return -ENODEV to user space when vram is lost v2

Quoted 9 lines

-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Chunming Zhou
Sent: Tuesday, May 16, 2017 5:26 AM
To: [email protected]
Cc: Zhou, David(ChunMing)
Subject: [PATCH 2/4] drm/amdgpu: return -ENODEV to user space when vram is lost v2

Quoted 6 lines

below ioctl will return -ENODEV:
amdgpu_cs_ioctl
amdgpu_cs_wait_ioctl
amdgpu_cs_wait_fences_ioctl
amdgpu_gem_va_ioctl
amdgpu_info_ioctl

Do we want to block the info ioctl? Isn't that where the lost context query is?

Alex

Quoted 190 lines

v2: only for map and replace cases in amdgpu_gem_va_ioctl

Change-Id I8970cde3301b7cfeb4263cc0f0e54aece215c98e
---
 drivers/gpu/drm/amd/amdgpu/amdgpu.h        |  4 ++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c     |  9 +++++++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_device.c |  4 +++-
 drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c    |  5 +++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c    | 10 ++++++++++
 5 files changed, 31 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +4 -0
index f9da215..dcd6203 100644
@@ -855,6 +855,7 @@ struct amdgpu_fpriv {
 	struct amdgpu_ctx_mgr	ctx_mgr;
 	spinlock_t		sem_handles_lock;
 	struct idr		sem_handles;
+	u32			vram_lost_counter;
 };
 
 /*
@@ -1607,6 +1608,7 @@ struct amdgpu_device {
 	atomic64_t			num_bytes_moved;
 	atomic64_t			num_evictions;
 	atomic_t			gpu_reset_counter;
+	atomic_t			vram_lost_counter;
 
 	/* data for buffer migration throttling */
 	struct {
@@ -2005,6 +2007,8 @@ static inline void
 amdgpu_unregister_atpx_handler(void) {}
 extern const struct drm_ioctl_desc amdgpu_ioctls_kms[];
 extern const int amdgpu_max_kms_ioctl;
 
+bool amdgpu_kms_vram_lost(struct amdgpu_device *adev,
+			  struct amdgpu_fpriv *fpriv);
 int amdgpu_driver_load_kms(struct drm_device *dev, unsigned long flags);
 int amdgpu_driver_unload_kms(struct drm_device *dev);

void amdgpu_driver_lastclose_kms(struct drm_device *dev);

drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +1 -0
index b803412..911aa02 100644
@@ -1097,6 +1097,7 @@ static int amdgpu_cs_submit(struct
 amdgpu_cs_parser *p,
 int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file
 *filp)
 {
 	struct amdgpu_device *adev = dev->dev_private;
+	struct amdgpu_fpriv *fpriv = filp->driver_priv;
 	union drm_amdgpu_cs *cs = data;
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +2 -0
 	struct amdgpu_cs_parser parser = {};
 	bool reserved_buffers = false;
@@ -1104,6 +1105,8 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void
 *data, struct drm_file *filp)
 
 	if (!adev->accel_working)
 		return -EBUSY;
+	if (amdgpu_kms_vram_lost(adev, fpriv))
+		return -ENODEV;
 
 	parser.adev = adev;
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +3 -0
 	parser.filp = filp;
@@ -1165,12 +1168,15 @@ int amdgpu_cs_wait_ioctl(struct drm_device
 *dev, void *data,
 {
 	union drm_amdgpu_wait_cs *wait = data;
 	struct amdgpu_device *adev = dev->dev_private;
+	struct amdgpu_fpriv *fpriv = filp->driver_priv;
 	unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
 	struct amdgpu_ring *ring = NULL;
 	struct amdgpu_ctx *ctx;
 	struct fence *fence;
 	long r;
 
+	if (amdgpu_kms_vram_lost(adev, fpriv))
+		return -ENODEV;
 	r = amdgpu_cs_get_ring(adev, wait->in.ip_type, wait-
 >in.ip_instance,
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +3 -0
 			       wait->in.ring, &ring);
 	if (r)
@@ -1344,12 +1350,15 @@ int amdgpu_cs_wait_fences_ioctl(struct
 drm_device *dev, void *data,
 				struct drm_file *filp)
 {
 	struct amdgpu_device *adev = dev->dev_private;
+	struct amdgpu_fpriv *fpriv = filp->driver_priv;
 	union drm_amdgpu_wait_fences *wait = data;
 	uint32_t fence_count = wait->in.fence_count;
 	struct drm_amdgpu_fence *fences_user;
 	struct drm_amdgpu_fence *fences;
 	int r;
 
+	if (amdgpu_kms_vram_lost(adev, fpriv))
+		return -ENODEV;
 	/* Get the fences from userspace */
 	fences = kmalloc_array(fence_count, sizeof(struct

drm_amdgpu_fence),

drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +0 -0
 			GFP_KERNEL);
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +3 -1
index c56ae4a..2f0fcf8 100644
@@ -2913,8 +2913,10 @@ int amdgpu_gpu_reset(struct amdgpu_device
 *adev)
 			if (r)
 				goto out;
 			vram_lost = amdgpu_check_vram_lost(adev);
-			if (vram_lost)
+			if (vram_lost) {
 				DRM_ERROR("VRAM is lost!\n");
+				atomic_inc(&adev->vram_lost_counter);
+			}
 			r = amdgpu_ttm_recover_gart(adev);
 			if (r)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
 				goto out;
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +5 -0
index d8275ef..83bc94c 100644
@@ -802,6 +802,11 @@ int amdgpu_gem_va_ioctl(struct drm_device *dev,
 void *data,
 			args->operation);
 		return -EINVAL;
 	}
+	if ((args->operation == AMDGPU_VA_OP_MAP) ||
+	    (args->operation == AMDGPU_VA_OP_REPLACE)) {
+		if (amdgpu_kms_vram_lost(adev, fpriv))
+			return -ENODEV;
+	}
 
 	INIT_LIST_HEAD(&list);
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +0 -0
 	if ((args->operation != AMDGPU_VA_OP_CLEAR) &&
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +1 -0
index 368829a..a231aa1 100644
@@ -235,6 +235,7 @@ static int amdgpu_firmware_info(struct
 drm_amdgpu_info_firmware *fw_info,
 static int amdgpu_info_ioctl(struct drm_device *dev, void *data, struct
 drm_file *filp)
 {
 	struct amdgpu_device *adev = dev->dev_private;
+	struct amdgpu_fpriv *fpriv = filp->driver_priv;
 	struct drm_amdgpu_info *info = data;
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +0 -0
 	struct amdgpu_mode_info *minfo = &adev->mode_info;
 	void __user *out = (void __user *)(uintptr_t)info->return_pointer;
@@ -247,6 +248,8 @@ static int amdgpu_info_ioctl(struct drm_device *dev,

void *data, struct drm_file

drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +8 -0
 	if (!info->return_size || !info->return_pointer)
 		return -EINVAL;
+	if (amdgpu_kms_vram_lost(adev, fpriv))
+		return -ENODEV;
 
 	switch (info->query) {
 	case AMDGPU_INFO_VIRTUAL_RANGE: {
@@ -779,6 +782,12 @@ void amdgpu_driver_lastclose_kms(struct
 drm_device *dev)
 	vga_switcheroo_process_delayed_switch();
 }
 
+bool amdgpu_kms_vram_lost(struct amdgpu_device *adev,
+			  struct amdgpu_fpriv *fpriv)
+{
+	return fpriv->vram_lost_counter != atomic_read(&adev-
 >vram_lost_counter);
+}
+
 /**

* amdgpu_driver_open_kms - drm callback for open *

drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +1 -0
@@ -833,6 +842,7 @@ int amdgpu_driver_open_kms(struct drm_device
 *dev, struct drm_file *file_priv)
 
 	amdgpu_ctx_mgr_init(&fpriv->ctx_mgr);
 
+	fpriv->vram_lost_counter = atomic_read(&adev-
 >vram_lost_counter);
 	file_priv->driver_priv = fpriv;
Quoted 8 lines

out_suspend:
--
1.9.1

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): BN6PR12MB1652E1AFF691F58C92FC2CD7F7F90-/b2+HYfkarQqUD6E6FAiowdYzm3356FpvxpqHgZTriW3zl9H0oFU5g@public.gmane.org

#6 lore raw

Re: [PATCH 2/4] drm/amdgpu: return -ENODEV to user space when vram is lost v2

Am 23.05.2017 um 17:08 schrieb Deucher, Alexander:

Quoted 9 lines
Quoted 9 lines

-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Chunming Zhou
Sent: Tuesday, May 16, 2017 5:26 AM
To: [email protected]
Cc: Zhou, David(ChunMing)
Subject: [PATCH 2/4] drm/amdgpu: return -ENODEV to user space when vram is lost v2

Quoted 7 lines
Quoted 6 lines

below ioctl will return -ENODEV:
amdgpu_cs_ioctl
amdgpu_cs_wait_ioctl
amdgpu_cs_wait_fences_ioctl
amdgpu_gem_va_ioctl
amdgpu_info_ioctl

Do we want to block the info ioctl? Isn't that where the lost context query is?

No, that's amdgpu_ctx_ioctl.

But I think the conclusion is that we want to move the vram_lost counter to be per CTX and not per device.

Christian.

Quoted 196 lines

Alex

Quoted 193 lines

v2: only for map and replace cases in amdgpu_gem_va_ioctl

Change-Id I8970cde3301b7cfeb4263cc0f0e54aece215c98e
---
  drivers/gpu/drm/amd/amdgpu/amdgpu.h        |  4 ++++
  drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c     |  9 +++++++++
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c |  4 +++-
  drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c    |  5 +++++
  drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c    | 10 ++++++++++
  5 files changed, 31 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +4 -0
index f9da215..dcd6203 100644
@@ -855,6 +855,7 @@ struct amdgpu_fpriv {
  	struct amdgpu_ctx_mgr	ctx_mgr;
  	spinlock_t		sem_handles_lock;
  	struct idr		sem_handles;
+	u32			vram_lost_counter;
  };
 
  /*
@@ -1607,6 +1608,7 @@ struct amdgpu_device {
  	atomic64_t			num_bytes_moved;
  	atomic64_t			num_evictions;
  	atomic_t			gpu_reset_counter;
+	atomic_t			vram_lost_counter;
 
  	/* data for buffer migration throttling */
  	struct {
@@ -2005,6 +2007,8 @@ static inline void
 amdgpu_unregister_atpx_handler(void) {}
  extern const struct drm_ioctl_desc amdgpu_ioctls_kms[];
  extern const int amdgpu_max_kms_ioctl;
 
+bool amdgpu_kms_vram_lost(struct amdgpu_device *adev,
+			  struct amdgpu_fpriv *fpriv);
  int amdgpu_driver_load_kms(struct drm_device *dev, unsigned long flags);
  int amdgpu_driver_unload_kms(struct drm_device *dev);
  void amdgpu_driver_lastclose_kms(struct drm_device *dev);
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +1 -0
index b803412..911aa02 100644
@@ -1097,6 +1097,7 @@ static int amdgpu_cs_submit(struct
 amdgpu_cs_parser *p,
  int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file
 *filp)
  {
  	struct amdgpu_device *adev = dev->dev_private;
+	struct amdgpu_fpriv *fpriv = filp->driver_priv;
  	union drm_amdgpu_cs *cs = data;
  	struct amdgpu_cs_parser parser = {};
  	bool reserved_buffers = false;
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +2 -0
@@ -1104,6 +1105,8 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void
 *data, struct drm_file *filp)
 
  	if (!adev->accel_working)
  		return -EBUSY;
+	if (amdgpu_kms_vram_lost(adev, fpriv))
+		return -ENODEV;
 
  	parser.adev = adev;
  	parser.filp = filp;
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +3 -0
@@ -1165,12 +1168,15 @@ int amdgpu_cs_wait_ioctl(struct drm_device
 *dev, void *data,
  {
  	union drm_amdgpu_wait_cs *wait = data;
  	struct amdgpu_device *adev = dev->dev_private;
+	struct amdgpu_fpriv *fpriv = filp->driver_priv;
  	unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
  	struct amdgpu_ring *ring = NULL;
  	struct amdgpu_ctx *ctx;
  	struct fence *fence;
  	long r;
 
+	if (amdgpu_kms_vram_lost(adev, fpriv))
+		return -ENODEV;
  	r = amdgpu_cs_get_ring(adev, wait->in.ip_type, wait-
 > in.ip_instance,

wait->in.ring, &ring);
if (r)

drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +3 -0
@@ -1344,12 +1350,15 @@ int amdgpu_cs_wait_fences_ioctl(struct
 drm_device *dev, void *data,
  				struct drm_file *filp)
  {
  	struct amdgpu_device *adev = dev->dev_private;
+	struct amdgpu_fpriv *fpriv = filp->driver_priv;
  	union drm_amdgpu_wait_fences *wait = data;
  	uint32_t fence_count = wait->in.fence_count;
  	struct drm_amdgpu_fence *fences_user;
  	struct drm_amdgpu_fence *fences;
  	int r;
 
+	if (amdgpu_kms_vram_lost(adev, fpriv))
+		return -ENODEV;
  	/* Get the fences from userspace */
  	fences = kmalloc_array(fence_count, sizeof(struct

drm_amdgpu_fence),
GFP_KERNEL);

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +3 -1
index c56ae4a..2f0fcf8 100644
@@ -2913,8 +2913,10 @@ int amdgpu_gpu_reset(struct amdgpu_device
 *adev)
  			if (r)
  				goto out;
  			vram_lost = amdgpu_check_vram_lost(adev);
-			if (vram_lost)
+			if (vram_lost) {
  				DRM_ERROR("VRAM is lost!\n");
+				atomic_inc(&adev->vram_lost_counter);
+			}
  			r = amdgpu_ttm_recover_gart(adev);
  			if (r)

goto out;

drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +5 -0
index d8275ef..83bc94c 100644
@@ -802,6 +802,11 @@ int amdgpu_gem_va_ioctl(struct drm_device *dev,
 void *data,
  			args->operation);
  		return -EINVAL;
  	}
+	if ((args->operation == AMDGPU_VA_OP_MAP) ||
+	    (args->operation == AMDGPU_VA_OP_REPLACE)) {
+		if (amdgpu_kms_vram_lost(adev, fpriv))
+			return -ENODEV;
+	}
 
  	INIT_LIST_HEAD(&list);

if ((args->operation != AMDGPU_VA_OP_CLEAR) &&

drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +1 -0
index 368829a..a231aa1 100644
@@ -235,6 +235,7 @@ static int amdgpu_firmware_info(struct
 drm_amdgpu_info_firmware *fw_info,
  static int amdgpu_info_ioctl(struct drm_device *dev, void *data, struct
 drm_file *filp)
  {
  	struct amdgpu_device *adev = dev->dev_private;
+	struct amdgpu_fpriv *fpriv = filp->driver_priv;
  	struct drm_amdgpu_info *info = data;
  	struct amdgpu_mode_info *minfo = &adev->mode_info;
  	void __user *out = (void __user *)(uintptr_t)info->return_pointer;
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +0 -0
@@ -247,6 +248,8 @@ static int amdgpu_info_ioctl(struct drm_device *dev,

void *data, struct drm_file

  	if (!info->return_size || !info->return_pointer)
  		return -EINVAL;
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +8 -0
+	if (amdgpu_kms_vram_lost(adev, fpriv))
+		return -ENODEV;
 
  	switch (info->query) {
  	case AMDGPU_INFO_VIRTUAL_RANGE: {
@@ -779,6 +782,12 @@ void amdgpu_driver_lastclose_kms(struct
 drm_device *dev)
  	vga_switcheroo_process_delayed_switch();
  }
 
+bool amdgpu_kms_vram_lost(struct amdgpu_device *adev,
+			  struct amdgpu_fpriv *fpriv)
+{
+	return fpriv->vram_lost_counter != atomic_read(&adev-
 > vram_lost_counter);
+}
+
  /**

* amdgpu_driver_open_kms - drm callback for open *

drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +1 -0
@@ -833,6 +842,7 @@ int amdgpu_driver_open_kms(struct drm_device
 *dev, struct drm_file *file_priv)
 
  	amdgpu_ctx_mgr_init(&fpriv->ctx_mgr);
 
+	fpriv->vram_lost_counter = atomic_read(&adev-
 > vram_lost_counter);
  	file_priv->driver_priv = fpriv;

out_suspend:
--
1.9.1

Quoted 8 lines
Quoted 4 lines

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 69717c0b-b2c1-589a-c466-5d6be9518eda-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

author
#8 lore raw

Re: [PATCH 2/4] drm/amdgpu: return -ENODEV to user space when vram is lost v2

Quoted 16 lines · Christian König

Am 23.05.2017 um 17:08 schrieb Deucher, Alexander:

Quoted 15 lines
Quoted 15 lines

-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Chunming Zhou
Sent: Tuesday, May 16, 2017 5:26 AM
To: [email protected]
Cc: Zhou, David(ChunMing)
Subject: [PATCH 2/4] drm/amdgpu: return -ENODEV to user space when vram is lost v2

below ioctl will return -ENODEV:
amdgpu_cs_ioctl
amdgpu_cs_wait_ioctl
amdgpu_cs_wait_fences_ioctl
amdgpu_gem_va_ioctl
amdgpu_info_ioctl

Quoted 5 lines
Quoted 1 line

Do we want to block the info ioctl? Isn't that where the lost context query is?

No, that's amdgpu_ctx_ioctl.

But I think the conclusion is that we want to move the vram_lost counter to be per CTX and not per device.

Yes, Monk is working on it for virt case, after it, I think we can reuse it.

Regards,
David zhou

Quoted 214 lines

Christian.

Quoted 209 lines

Alex

Quoted 194 lines

v2: only for map and replace cases in amdgpu_gem_va_ioctl

Change-Id I8970cde3301b7cfeb4263cc0f0e54aece215c98e
---
  drivers/gpu/drm/amd/amdgpu/amdgpu.h        |  4 ++++
  drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c     |  9 +++++++++
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c |  4 +++-
  drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c    |  5 +++++
  drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c    | 10 ++++++++++
  5 files changed, 31 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +4 -0
index f9da215..dcd6203 100644
@@ -855,6 +855,7 @@ struct amdgpu_fpriv {
      struct amdgpu_ctx_mgr    ctx_mgr;
      spinlock_t        sem_handles_lock;
      struct idr        sem_handles;
+    u32            vram_lost_counter;
  };
 
  /*
@@ -1607,6 +1608,7 @@ struct amdgpu_device {
      atomic64_t            num_bytes_moved;
      atomic64_t            num_evictions;
      atomic_t            gpu_reset_counter;
+    atomic_t            vram_lost_counter;
 
      /* data for buffer migration throttling */
      struct {
@@ -2005,6 +2007,8 @@ static inline void
 amdgpu_unregister_atpx_handler(void) {}
  extern const struct drm_ioctl_desc amdgpu_ioctls_kms[];
  extern const int amdgpu_max_kms_ioctl;
 
+bool amdgpu_kms_vram_lost(struct amdgpu_device *adev,
+              struct amdgpu_fpriv *fpriv);
  int amdgpu_driver_load_kms(struct drm_device *dev, unsigned long 
 flags);
  int amdgpu_driver_unload_kms(struct drm_device *dev);
  void amdgpu_driver_lastclose_kms(struct drm_device *dev);
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +0 -0
index b803412..911aa02 100644
@@ -1097,6 +1097,7 @@ static int amdgpu_cs_submit(struct
 amdgpu_cs_parser *p,
  int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct 

drm_file
*filp)

  {
      struct amdgpu_device *adev = dev->dev_private;
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +3 -0
+    struct amdgpu_fpriv *fpriv = filp->driver_priv;
      union drm_amdgpu_cs *cs = data;
      struct amdgpu_cs_parser parser = {};
      bool reserved_buffers = false;
@@ -1104,6 +1105,8 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void
 *data, struct drm_file *filp)
 
      if (!adev->accel_working)
          return -EBUSY;
+    if (amdgpu_kms_vram_lost(adev, fpriv))
+        return -ENODEV;
 
      parser.adev = adev;
      parser.filp = filp;
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +3 -0
@@ -1165,12 +1168,15 @@ int amdgpu_cs_wait_ioctl(struct drm_device
 *dev, void *data,
  {
      union drm_amdgpu_wait_cs *wait = data;
      struct amdgpu_device *adev = dev->dev_private;
+    struct amdgpu_fpriv *fpriv = filp->driver_priv;
      unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
      struct amdgpu_ring *ring = NULL;
      struct amdgpu_ctx *ctx;
      struct fence *fence;
      long r;
 
+    if (amdgpu_kms_vram_lost(adev, fpriv))
+        return -ENODEV;
      r = amdgpu_cs_get_ring(adev, wait->in.ip_type, wait-
 > in.ip_instance,
                     wait->in.ring, &ring);
      if (r)
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +3 -0
@@ -1344,12 +1350,15 @@ int amdgpu_cs_wait_fences_ioctl(struct
 drm_device *dev, void *data,
                  struct drm_file *filp)
  {
      struct amdgpu_device *adev = dev->dev_private;
+    struct amdgpu_fpriv *fpriv = filp->driver_priv;
      union drm_amdgpu_wait_fences *wait = data;
      uint32_t fence_count = wait->in.fence_count;
      struct drm_amdgpu_fence *fences_user;
      struct drm_amdgpu_fence *fences;
      int r;
 
+    if (amdgpu_kms_vram_lost(adev, fpriv))
+        return -ENODEV;
      /* Get the fences from userspace */
      fences = kmalloc_array(fence_count, sizeof(struct

drm_amdgpu_fence),

              GFP_KERNEL);
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +3 -1
index c56ae4a..2f0fcf8 100644
@@ -2913,8 +2913,10 @@ int amdgpu_gpu_reset(struct amdgpu_device
 *adev)
              if (r)
                  goto out;
              vram_lost = amdgpu_check_vram_lost(adev);
-            if (vram_lost)
+            if (vram_lost) {
                  DRM_ERROR("VRAM is lost!\n");
+                atomic_inc(&adev->vram_lost_counter);
+            }
              r = amdgpu_ttm_recover_gart(adev);
              if (r)
                  goto out;
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +5 -0
index d8275ef..83bc94c 100644
@@ -802,6 +802,11 @@ int amdgpu_gem_va_ioctl(struct drm_device *dev,
 void *data,
              args->operation);
          return -EINVAL;
      }
+    if ((args->operation == AMDGPU_VA_OP_MAP) ||
+        (args->operation == AMDGPU_VA_OP_REPLACE)) {
+        if (amdgpu_kms_vram_lost(adev, fpriv))
+            return -ENODEV;
+    }
 
      INIT_LIST_HEAD(&list);
      if ((args->operation != AMDGPU_VA_OP_CLEAR) &&
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +0 -0
index 368829a..a231aa1 100644
@@ -235,6 +235,7 @@ static int amdgpu_firmware_info(struct
 drm_amdgpu_info_firmware *fw_info,
  static int amdgpu_info_ioctl(struct drm_device *dev, void *data, 

struct
drm_file *filp)

  {
      struct amdgpu_device *adev = dev->dev_private;
+    struct amdgpu_fpriv *fpriv = filp->driver_priv;
      struct drm_amdgpu_info *info = data;
      struct amdgpu_mode_info *minfo = &adev->mode_info;
      void __user *out = (void __user 

*)(uintptr_t)info->return_pointer;

drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +0 -0
@@ -247,6 +248,8 @@ static int amdgpu_info_ioctl(struct drm_device 
 *dev,

void *data, struct drm_file

      if (!info->return_size || !info->return_pointer)
          return -EINVAL;
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +8 -0
+    if (amdgpu_kms_vram_lost(adev, fpriv))
+        return -ENODEV;
 
      switch (info->query) {
      case AMDGPU_INFO_VIRTUAL_RANGE: {
@@ -779,6 +782,12 @@ void amdgpu_driver_lastclose_kms(struct
 drm_device *dev)
      vga_switcheroo_process_delayed_switch();
  }
 
+bool amdgpu_kms_vram_lost(struct amdgpu_device *adev,
+              struct amdgpu_fpriv *fpriv)
+{
+    return fpriv->vram_lost_counter != atomic_read(&adev-
 > vram_lost_counter);
+}
+
  /**

* amdgpu_driver_open_kms - drm callback for open *

drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +1 -0
@@ -833,6 +842,7 @@ int amdgpu_driver_open_kms(struct drm_device
 *dev, struct drm_file *file_priv)
 
      amdgpu_ctx_mgr_init(&fpriv->ctx_mgr);
 
+    fpriv->vram_lost_counter = atomic_read(&adev-
 > vram_lost_counter);
      file_priv->driver_priv = fpriv;
Quoted 8 lines

out_suspend:

Signature · 6 lines
1.9.1

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

#9 lore raw

PATCH 3/4 drm/amdgpu: skip all jobs of guilty vm

In reply to a message not in the archive
Change-Id Ib305282a21d37d6872afe4c1ce63d53b6517f338
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 15 +++++++++++----
 1 file changed, 11 insertions(+), 4 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +11 -4
index 4a5ffa5..da2f87e1 100644
@@ -167,6 +167,7 @@ static struct fence *amdgpu_job_run(struct amd_sched_job *sched_job)
 {
 	struct fence *fence = NULL;
 	struct amdgpu_job *job;
+	struct amdgpu_fpriv *fpriv = NULL;
 	int r;
 
 	if (!sched_job) {
@@ -178,10 +179,16 @@ static struct fence *amdgpu_job_run(struct amd_sched_job *sched_job)
 	BUG_ON(amdgpu_sync_peek_fence(&job->sync, NULL));
 
 	trace_amdgpu_sched_run_job(job);
-	r = amdgpu_ib_schedule(job->ring, job->num_ibs, job->ibs, job, &fence);
-	if (r)
-		DRM_ERROR("Error scheduling IBs (%d)\n", r);
-
+	if (job->vm)
+		fpriv = container_of(job->vm, struct amdgpu_fpriv, vm);
+	/* skip ib schedule when vram is lost */
+	if (fpriv && amdgpu_kms_vram_lost(job->adev, fpriv))
+		DRM_ERROR("Skip scheduling IBs!\n");
+	else {
+		r = amdgpu_ib_schedule(job->ring, job->num_ibs, job->ibs, job, &fence);
+		if (r)
+			DRM_ERROR("Error scheduling IBs (%d)\n", r);
+	}
 	/* if gpu reset, hw fence will be replaced here */
 	fence_put(job->fence);
 	job->fence = fence_get(fence);
Signature · 6 lines
1.9.1

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
#10 lore raw

PATCH 4/4 drm/amdgpu: reset fpriv vram_lost_counter

In reply to a message not in the archive
Change-Id I8eb6d7f558da05510e429d3bf1d48c8cec6c1977
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 3 +++
 include/uapi/drm/amdgpu_drm.h          | 1 +
 2 files changed, 4 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +3 -0
index bca1fb5..f3e7525 100644
@@ -2547,6 +2547,9 @@ int amdgpu_vm_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
 	case AMDGPU_VM_OP_UNRESERVE_VMID:
 		amdgpu_vm_free_reserved_vmid(adev, &fpriv->vm, AMDGPU_GFXHUB);
 		break;
+	case AMDGPU_VM_OP_RESET:
+		fpriv->vram_lost_counter = atomic_read(&adev->vram_lost_counter);
+		break;
 	default:
 		return -EINVAL;
 	}
include/uapi/drm/amdgpu_drm.h +1 -0
index 81141e1..ed2ef9b 100644
@@ -246,6 +246,7 @@ struct drm_amdgpu_sem_in {
 /* vm ioctl */
 #define AMDGPU_VM_OP_RESERVE_VMID	1
 #define AMDGPU_VM_OP_UNRESERVE_VMID	2
+#define AMDGPU_VM_OP_RESET		3
 
 struct drm_amdgpu_vm_in {
 	/** AMDGPU_VM_OP_* */
Signature · 6 lines
1.9.1

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 1494926750-1081-4-git-send-email-David1.Zhou-5C7GfCeVMHo@public.gmane.org

#12 lore raw

Re: [PATCH 4/4] drm/amdgpu: reset fpriv vram_lost_counter

Quoted 12 lines · Chunming Zhou
Change-Id I8eb6d7f558da05510e429d3bf1d48c8cec6c1977
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 3 +++
 include/uapi/drm/amdgpu_drm.h          | 1 +
 2 files changed, 4 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +0 -0
index bca1fb5..f3e7525 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c int amdgpu_vm_ioctl(struct drm_device *dev, void *data, struct drm_file *filp) · line 2552

Quoted 6 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +3 -0
@@ -2547,6 +2547,9 @@ int amdgpu_vm_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
 	case AMDGPU_VM_OP_UNRESERVE_VMID:
 		amdgpu_vm_free_reserved_vmid(adev, &fpriv->vm, AMDGPU_GFXHUB);
 		break;
+	case AMDGPU_VM_OP_RESET:
+		fpriv->vram_lost_counter = atomic_read(&adev->vram_lost_counter);
+		break;

How do you envision the UMDs using this? I can mostly think of them calling this ioctl when a context is created or destroyed. But that would also allow any other remaining contexts using the same DRM file descriptor to use all ioctls again. So, I think there needs to be a vram_lost_counter in struct amdgpu_ctx instead of in struct amdgpu_fpriv.

But then I'm not sure this ioctl will be useful... I guess the UMD could try re-uploading the contents of crucial BOs (shader code, resource descriptors etc.) for an existing context and then call this ioctl. How about the GPUVM page tables? Will the kernel driver automatically re-generate those as needed, or will the UMD also need to e.g. destroy and re-create the VM mappings for all BOs before calling this ioctl?

It's hard to be sure whether that's workable for the UMD without at least a working prototype...

Signature · 6 lines
Earthling Michel Dänzer               |               http://www.amd.com
Libre software enthusiast             |             Mesa and X developer
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 58988726-543a-535a-3011-860d29b9f2da-otUistvHUpPR7s880joybQ@public.gmane.org

author
#14 lore raw

Re: [PATCH 4/4] drm/amdgpu: reset fpriv vram_lost_counter

Quoted 8 lines · Michel Dänzer
Quoted 7 lines · Chunming Zhou
Change-Id I8eb6d7f558da05510e429d3bf1d48c8cec6c1977
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 3 +++
  include/uapi/drm/amdgpu_drm.h          | 1 +
  2 files changed, 4 insertions(+)
Quoted 16 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c int amdgpu_vm_ioctl(struct drm_device *dev, void *data, struct drm_file *filp) · line 2552

Quoted 11 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +3 -0
index bca1fb5..f3e7525 100644
@@ -2547,6 +2547,9 @@ int amdgpu_vm_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
  	case AMDGPU_VM_OP_UNRESERVE_VMID:
  		amdgpu_vm_free_reserved_vmid(adev, &fpriv->vm, AMDGPU_GFXHUB);
  		break;
+	case AMDGPU_VM_OP_RESET:
+		fpriv->vram_lost_counter = atomic_read(&adev->vram_lost_counter);
+		break;

How do you envision the UMDs using this? I can mostly think of them calling this ioctl when a context is created or destroyed. But that would also allow any other remaining contexts using the same DRM file descriptor to use all ioctls again. So, I think there needs to be a vram_lost_counter in struct amdgpu_ctx instead of in struct amdgpu_fpriv.

struct amdgpu_fpriv for vram_lost_counter is proper place, especially for ioctl return value. if you need to reset ctx one by one, we can mark all contexts of that vm, and then reset by userspace.

Quoted 7 lines

But then I'm not sure this ioctl will be useful... I guess the UMD could try re-uploading the contents of crucial BOs (shader code, resource descriptors etc.) for an existing context and then call this ioctl. How about the GPUVM page tables? Will the kernel driver automatically re-generate those as needed, or will the UMD also need to e.g. destroy and re-create the VM mappings for all BOs before calling this ioctl?

VM page table will recover by vm shadow BOs.

Quoted 3 lines

It's hard to be sure whether that's workable for the UMD without at least a working prototype...

Totally agree, if you can help to do this in userspace, I'd like to support you from kernel side, or Christian.

Thanks,
David Zhou

Quoted 2 lines

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

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

#16 lore raw

Re: [PATCH 4/4] drm/amdgpu: reset fpriv vram_lost_counter

Quoted 28 lines · On 17/05/17 12:04 PM, zhoucm1 wrote:
Quoted 19 lines · Michel Dänzer
Quoted 18 lines · Chunming Zhou
Change-Id I8eb6d7f558da05510e429d3bf1d48c8cec6c1977
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 3 +++
  include/uapi/drm/amdgpu_drm.h          | 1 +
  2 files changed, 4 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +1 -0
index bca1fb5..f3e7525 100644
@@ -2547,6 +2547,9 @@ int amdgpu_vm_ioctl(struct drm_device *dev,
 void *data, struct drm_file *filp)
      case AMDGPU_VM_OP_UNRESERVE_VMID:
          amdgpu_vm_free_reserved_vmid(adev, &fpriv->vm, AMDGPU_GFXHUB);
          break;
+    case AMDGPU_VM_OP_RESET:
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c int amdgpu_vm_ioctl(struct drm_device *dev, · line 2554

Quoted 3 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +2 -0
+        fpriv->vram_lost_counter =
 atomic_read(&adev->vram_lost_counter);
+        break;

How do you envision the UMDs using this? I can mostly think of them calling this ioctl when a context is created or destroyed. But that would also allow any other remaining contexts using the same DRM file descriptor to use all ioctls again. So, I think there needs to be a vram_lost_counter in struct amdgpu_ctx instead of in struct amdgpu_fpriv.

Quoted 4 lines

struct amdgpu_fpriv for vram_lost_counter is proper place, especially for ioctl return value.
if you need to reset ctx one by one, we can mark all contexts of that vm, and then reset by userspace.

I'm not following. With vram_lost_counter in amdgpu_fpriv, if any context calls this ioctl, all other contexts using the same file descriptor will also be considered safe again, right? So I'm still not sure how this is supposed to be used by the UMDs. Can you describe your idea for that?

Quoted 4 lines
Quoted 2 lines

It's hard to be sure whether that's workable for the UMD without at least a working prototype...

Totally agree, if you can help to do this in userspace, I'd like to support you from kernel side, or Christian.

I'm busy with other stuff.

Signature · 6 lines
Earthling Michel Dänzer               |               http://www.amd.com
Libre software enthusiast             |             Mesa and X developer
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 29fe2142-7fd1-e23a-49d9-c38dc685db92-otUistvHUpPR7s880joybQ@public.gmane.org

author
#18 lore raw

Re: [PATCH 4/4] drm/amdgpu: reset fpriv vram_lost_counter

Quoted 29 lines · Michel Dänzer
Quoted 28 lines · On 17/05/17 12:04 PM, zhoucm1 wrote:
Quoted 27 lines · Michel Dänzer
Quoted 13 lines · Chunming Zhou
Change-Id I8eb6d7f558da05510e429d3bf1d48c8cec6c1977
---
   drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 3 +++
   include/uapi/drm/amdgpu_drm.h          | 1 +
   2 files changed, 4 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +0 -0
index bca1fb5..f3e7525 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c int amdgpu_vm_ioctl(struct drm_device *dev, · line 2554

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +3 -0
@@ -2547,6 +2547,9 @@ int amdgpu_vm_ioctl(struct drm_device *dev,
 void *data, struct drm_file *filp)
       case AMDGPU_VM_OP_UNRESERVE_VMID:
           amdgpu_vm_free_reserved_vmid(adev, &fpriv->vm, AMDGPU_GFXHUB);
           break;
+    case AMDGPU_VM_OP_RESET:
+        fpriv->vram_lost_counter =
 atomic_read(&adev->vram_lost_counter);
+        break;

How do you envision the UMDs using this? I can mostly think of them calling this ioctl when a context is created or destroyed. But that would also allow any other remaining contexts using the same DRM file descriptor to use all ioctls again. So, I think there needs to be a vram_lost_counter in struct amdgpu_ctx instead of in struct amdgpu_fpriv.

Quoted 7 lines
Quoted 4 lines

struct amdgpu_fpriv for vram_lost_counter is proper place, especially for ioctl return value.
if you need to reset ctx one by one, we can mark all contexts of that vm, and then reset by userspace.

I'm not following. With vram_lost_counter in amdgpu_fpriv, if any context calls this ioctl, all other contexts using the same file descriptor will also be considered safe again, right?

Yes, but it really depends on userspace requirement, if you need to reset ctx one by one, we can mark all contexts of that vm to guilty, and then reset one context by userspace.

Quoted 3 lines

So I'm still not
sure how this is supposed to be used by the UMDs. Can you describe your idea for that?

Correct first, this idea is picked up from Christian. We just one to provide a possibility to handle ENODEV and recover system, rather than just system dead when vram is lost. And how UMDs handle reset? which obviously need to more discussion between kernel and userspace.

Regards,
David Zhou

Quoted 9 lines
Quoted 6 lines
Quoted 4 lines

It's hard to be sure whether that's workable for the UMD without at least a working prototype...

Totally agree, if you can help to do this in userspace, I'd like to support you from kernel side, or Christian.

I'm busy with other stuff.

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

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

#20 lore raw

Re: [PATCH 4/4] drm/amdgpu: reset fpriv vram_lost_counter

Quoted 32 lines · On 17/05/17 01:28 PM, zhoucm1 wrote:
Quoted 31 lines · Michel Dänzer
Quoted 30 lines · On 17/05/17 12:04 PM, zhoucm1 wrote:
Quoted 17 lines · Michel Dänzer
Quoted 16 lines · Chunming Zhou
Change-Id I8eb6d7f558da05510e429d3bf1d48c8cec6c1977
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +2 -0
index bca1fb5..f3e7525 100644
@@ -2547,6 +2547,9 @@ int amdgpu_vm_ioctl(struct drm_device *dev,
 void *data, struct drm_file *filp)
       case AMDGPU_VM_OP_UNRESERVE_VMID:
           amdgpu_vm_free_reserved_vmid(adev, &fpriv->vm,
 AMDGPU_GFXHUB);
           break;
+    case AMDGPU_VM_OP_RESET:
+        fpriv->vram_lost_counter =
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c int amdgpu_vm_ioctl(struct drm_device *dev, · line 2555

Quoted 2 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +1 -0
 atomic_read(&adev->vram_lost_counter);
+        break;

How do you envision the UMDs using this? I can mostly think of them calling this ioctl when a context is created or destroyed. But that would also allow any other remaining contexts using the same DRM file descriptor to use all ioctls again. So, I think there needs to be a vram_lost_counter in struct amdgpu_ctx instead of in struct amdgpu_fpriv.

struct amdgpu_fpriv for vram_lost_counter is proper place, especially for ioctl return value.
if you need to reset ctx one by one, we can mark all contexts of that vm, and then reset by userspace.

Quoted 6 lines
Quoted 3 lines

I'm not following. With vram_lost_counter in amdgpu_fpriv, if any context calls this ioctl, all other contexts using the same file descriptor will also be considered safe again, right?

Yes, but it really depends on userspace requirement, if you need to reset ctx one by one, we can mark all contexts of that vm to guilty, and then reset one context by userspace.

Still not sure what you mean by that.

E.g. what do you mean by "guilty"? I thought that refers to the context which caused a hang. But it seems like you're using it to refer to any context which hasn't reacted yet to VRAM contents being lost.

Also not sure what you mean by "if you need to reset ctx one by one".

Signature · 6 lines
Earthling Michel Dänzer               |               http://www.amd.com
Libre software enthusiast             |             Mesa and X developer
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 7d87bc8e-9c09-ad25-de6e-dfbd8116bf6e-otUistvHUpPR7s880joybQ@public.gmane.org

author
#22 lore raw

Re: [PATCH 4/4] drm/amdgpu: reset fpriv vram_lost_counter

Quoted 36 lines · Michel Dänzer
Quoted 35 lines · On 17/05/17 01:28 PM, zhoucm1 wrote:
Quoted 34 lines · Michel Dänzer
Quoted 26 lines · On 17/05/17 12:04 PM, zhoucm1 wrote:
Quoted 17 lines · Michel Dänzer
Quoted 16 lines · Chunming Zhou
Change-Id I8eb6d7f558da05510e429d3bf1d48c8cec6c1977
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +2 -0
index bca1fb5..f3e7525 100644
@@ -2547,6 +2547,9 @@ int amdgpu_vm_ioctl(struct drm_device *dev,
 void *data, struct drm_file *filp)
        case AMDGPU_VM_OP_UNRESERVE_VMID:
            amdgpu_vm_free_reserved_vmid(adev, &fpriv->vm,
 AMDGPU_GFXHUB);
            break;
+    case AMDGPU_VM_OP_RESET:
+        fpriv->vram_lost_counter =
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c int amdgpu_vm_ioctl(struct drm_device *dev, · line 2555

Quoted 2 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +1 -0
 atomic_read(&adev->vram_lost_counter);
+        break;

How do you envision the UMDs using this? I can mostly think of them calling this ioctl when a context is created or destroyed. But that would also allow any other remaining contexts using the same DRM file descriptor to use all ioctls again. So, I think there needs to be a vram_lost_counter in struct amdgpu_ctx instead of in struct amdgpu_fpriv.

Quoted 4 lines

struct amdgpu_fpriv for vram_lost_counter is proper place, especially for ioctl return value.
if you need to reset ctx one by one, we can mark all contexts of that vm, and then reset by userspace.

I'm not following. With vram_lost_counter in amdgpu_fpriv, if any context calls this ioctl, all other contexts using the same file descriptor will also be considered safe again, right?

Quoted 8 lines
Quoted 3 lines

Yes, but it really depends on userspace requirement, if you need to reset ctx one by one, we can mark all contexts of that vm to guilty, and then reset one context by userspace.

Still not sure what you mean by that.

E.g. what do you mean by "guilty"? I thought that refers to the context which caused a hang. But it seems like you're using it to refer to any context which hasn't reacted yet to VRAM contents being lost.

When vram is lost, we treat all contexts need to reset.

Regards,
David Zhou

Quoted 4 lines

Also not sure what you mean by "if you need to reset ctx one by one".

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

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

#24 lore raw

Re: [PATCH 4/4] drm/amdgpu: reset fpriv vram_lost_counter

Quoted 40 lines · On 17/05/17 04:13 PM, zhoucm1 wrote:
Quoted 39 lines · Michel Dänzer
Quoted 38 lines · On 17/05/17 01:28 PM, zhoucm1 wrote:
Quoted 27 lines · Michel Dänzer
Quoted 26 lines · On 17/05/17 12:04 PM, zhoucm1 wrote:
Quoted 25 lines · Michel Dänzer
Quoted 9 lines · Chunming Zhou
Change-Id I8eb6d7f558da05510e429d3bf1d48c8cec6c1977
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +0 -0
index bca1fb5..f3e7525 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c int amdgpu_vm_ioctl(struct drm_device *dev, · line 2555

Quoted 9 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +3 -0
@@ -2547,6 +2547,9 @@ int amdgpu_vm_ioctl(struct drm_device *dev,
 void *data, struct drm_file *filp)
        case AMDGPU_VM_OP_UNRESERVE_VMID:
            amdgpu_vm_free_reserved_vmid(adev, &fpriv->vm,
 AMDGPU_GFXHUB);
            break;
+    case AMDGPU_VM_OP_RESET:
+        fpriv->vram_lost_counter =
 atomic_read(&adev->vram_lost_counter);
+        break;

How do you envision the UMDs using this? I can mostly think of them calling this ioctl when a context is created or destroyed. But that would also allow any other remaining contexts using the same DRM file descriptor to use all ioctls again. So, I think there needs to be a vram_lost_counter in struct amdgpu_ctx instead of in struct amdgpu_fpriv.

Quoted 7 lines
Quoted 4 lines

struct amdgpu_fpriv for vram_lost_counter is proper place, especially for ioctl return value.
if you need to reset ctx one by one, we can mark all contexts of that vm, and then reset by userspace.

I'm not following. With vram_lost_counter in amdgpu_fpriv, if any context calls this ioctl, all other contexts using the same file descriptor will also be considered safe again, right?

Yes, but it really depends on userspace requirement, if you need to reset ctx one by one, we can mark all contexts of that vm to guilty, and then reset one context by userspace.

Quoted 6 lines
Quoted 5 lines

Still not sure what you mean by that.

E.g. what do you mean by "guilty"? I thought that refers to the context which caused a hang. But it seems like you're using it to refer to any context which hasn't reacted yet to VRAM contents being lost.

When vram is lost, we treat all contexts need to reset.

Essentially, your patches only track VRAM contents being lost per file descriptor, not per context. I'm not sure (rather skeptical) that this is suitable for OpenGL UMDs, since state is usually tracked per context. Marek / Nicolai?

Signature · 6 lines
Earthling Michel Dänzer               |               http://www.amd.com
Libre software enthusiast             |             Mesa and X developer
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 31db7a30-dd98-5cb2-4125-187d3d0e2a49-otUistvHUpPR7s880joybQ@public.gmane.org

#26 lore raw

Re: [PATCH 4/4] drm/amdgpu: reset fpriv vram_lost_counter

Am 17.05.2017 um 10:01 schrieb Michel Dänzer:

Quoted 43 lines · On 17/05/17 04:13 PM, zhoucm1 wrote:
Quoted 42 lines · Michel Dänzer
Quoted 41 lines · On 17/05/17 01:28 PM, zhoucm1 wrote:
Quoted 32 lines · Michel Dänzer
Quoted 31 lines · On 17/05/17 12:04 PM, zhoucm1 wrote:
Quoted 30 lines · Michel Dänzer
Quoted 17 lines · Chunming Zhou
Quoted 16 lines
Change-Id I8eb6d7f558da05510e429d3bf1d48c8cec6c1977
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +2 -0
index bca1fb5..f3e7525 100644
@@ -2547,6 +2547,9 @@ int amdgpu_vm_ioctl(struct drm_device *dev,
 void *data, struct drm_file *filp)
         case AMDGPU_VM_OP_UNRESERVE_VMID:
             amdgpu_vm_free_reserved_vmid(adev, &fpriv->vm,
 AMDGPU_GFXHUB);
             break;
+    case AMDGPU_VM_OP_RESET:
+        fpriv->vram_lost_counter =
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c int amdgpu_vm_ioctl(struct drm_device *dev, · line 2555

Quoted 2 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +1 -0
 atomic_read(&adev->vram_lost_counter);
+        break;

How do you envision the UMDs using this? I can mostly think of them calling this ioctl when a context is created or destroyed. But that would also allow any other remaining contexts using the same DRM file descriptor to use all ioctls again. So, I think there needs to be a vram_lost_counter in struct amdgpu_ctx instead of in struct amdgpu_fpriv.

struct amdgpu_fpriv for vram_lost_counter is proper place, especially for ioctl return value.
if you need to reset ctx one by one, we can mark all contexts of that vm, and then reset by userspace.

Quoted 6 lines
Quoted 3 lines

I'm not following. With vram_lost_counter in amdgpu_fpriv, if any context calls this ioctl, all other contexts using the same file descriptor will also be considered safe again, right?

Yes, but it really depends on userspace requirement, if you need to reset ctx one by one, we can mark all contexts of that vm to guilty, and then reset one context by userspace.

Still not sure what you mean by that.

Quoted 8 lines
Quoted 4 lines
Quoted 3 lines

E.g. what do you mean by "guilty"? I thought that refers to the context which caused a hang. But it seems like you're using it to refer to any context which hasn't reacted yet to VRAM contents being lost.

When vram is lost, we treat all contexts need to reset.

Essentially, your patches only track VRAM contents being lost per file descriptor, not per context. I'm not sure (rather skeptical) that this is suitable for OpenGL UMDs, since state is usually tracked per context. Marek / Nicolai?

Oh, yeah that's a good point.

The problem with tracking it per context is that Vulkan also wants the ENODEV on the amdgpu_gem_va_ioct() and amdgpu_info_ioctl() which are context less.

But thinking more about this blocking those two doesn't make much sense. The VM content can be restored and why should be disallow reading GPU info?

Christian.

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 7a302ebe-1de1-734f-fb21-aadcc7904d37-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

author
#28 lore raw

Re: [PATCH 4/4] drm/amdgpu: reset fpriv vram_lost_counter

Quoted 53 lines · Christian König

Am 17.05.2017 um 10:01 schrieb Michel Dänzer:

Quoted 43 lines · On 17/05/17 04:13 PM, zhoucm1 wrote:
Quoted 42 lines · Michel Dänzer
Quoted 41 lines · On 17/05/17 01:28 PM, zhoucm1 wrote:
Quoted 32 lines · Michel Dänzer
Quoted 31 lines · On 17/05/17 12:04 PM, zhoucm1 wrote:
Quoted 30 lines · Michel Dänzer
Quoted 17 lines · Chunming Zhou
Quoted 16 lines
Change-Id I8eb6d7f558da05510e429d3bf1d48c8cec6c1977
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +2 -0
index bca1fb5..f3e7525 100644
@@ -2547,6 +2547,9 @@ int amdgpu_vm_ioctl(struct drm_device *dev,
 void *data, struct drm_file *filp)
         case AMDGPU_VM_OP_UNRESERVE_VMID:
             amdgpu_vm_free_reserved_vmid(adev, &fpriv->vm,
 AMDGPU_GFXHUB);
             break;
+    case AMDGPU_VM_OP_RESET:
+        fpriv->vram_lost_counter =
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c int amdgpu_vm_ioctl(struct drm_device *dev, · line 2555

Quoted 2 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +1 -0
 atomic_read(&adev->vram_lost_counter);
+        break;

How do you envision the UMDs using this? I can mostly think of them calling this ioctl when a context is created or destroyed. But that would also allow any other remaining contexts using the same DRM file descriptor to use all ioctls again. So, I think there needs to be a vram_lost_counter in struct amdgpu_ctx instead of in struct amdgpu_fpriv.

struct amdgpu_fpriv for vram_lost_counter is proper place, especially for ioctl return value.
if you need to reset ctx one by one, we can mark all contexts of that vm, and then reset by userspace.

Quoted 6 lines
Quoted 3 lines

I'm not following. With vram_lost_counter in amdgpu_fpriv, if any context calls this ioctl, all other contexts using the same file descriptor will also be considered safe again, right?

Yes, but it really depends on userspace requirement, if you need to reset ctx one by one, we can mark all contexts of that vm to guilty, and then reset one context by userspace.

Still not sure what you mean by that.

Quoted 8 lines
Quoted 4 lines
Quoted 3 lines

E.g. what do you mean by "guilty"? I thought that refers to the context which caused a hang. But it seems like you're using it to refer to any context which hasn't reacted yet to VRAM contents being lost.

When vram is lost, we treat all contexts need to reset.

Essentially, your patches only track VRAM contents being lost per file descriptor, not per context. I'm not sure (rather skeptical) that this is suitable for OpenGL UMDs, since state is usually tracked per context. Marek / Nicolai?

Quoted 5 lines

Oh, yeah that's a good point.

The problem with tracking it per context is that Vulkan also wants the ENODEV on the amdgpu_gem_va_ioct() and amdgpu_info_ioctl() which are context less.

But thinking more about this blocking those two doesn't make much sense. The VM content can be restored and why should be disallow reading GPU info?

I can re-paste the Vulkan APIs requiring ENODEV:
"

The Vulkan APIs listed below could return VK_ERROR_DEVICE_LOST according to the spec.

I tries to provide a list of u/k interfaces that could be called for each vk API.

vkCreateDevice

-amdgpu_device_initialize.

-amdgpu_query_gpu_info

vkQueueSubmit

-amdgpu_cs_submit

vkWaitForFences

amdgpu_cs_wait_fences

vkGetEventStatus

vkQueueWaitIdle

vkDeviceWaitIdle

vkGetQueryPoolResults**

amdgpu_cs_query_Fence_status

vkQueueBindSparse**

amdgpu_bo_va_op

amdgpu_bo_va_op_raw

vkCreateSwapchainKHR**

vkAcquireNextImageKHR**

vkQueuePresentKHR

Not related with u/k interface.**

**

Besides those listed above, I think amdgpu_cs_signal_Sem/amdgpu_cs_wait_sem should respond to gpu reset as well."

Quoted 3 lines

Christian.

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

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

#30 lore raw

Re: [PATCH 4/4] drm/amdgpu: reset fpriv vram_lost_counter

Quoted 88 lines · On 17/05/17 05:46 PM, zhoucm1 wrote:
Quoted 63 lines · Christian König

Am 17.05.2017 um 10:01 schrieb Michel Dänzer:

Quoted 54 lines · On 17/05/17 04:13 PM, zhoucm1 wrote:
Quoted 53 lines · Michel Dänzer
Quoted 52 lines · On 17/05/17 01:28 PM, zhoucm1 wrote:
Quoted 38 lines · Michel Dänzer
Quoted 37 lines · On 17/05/17 12:04 PM, zhoucm1 wrote:
Quoted 36 lines · Michel Dänzer
Quoted 21 lines · Chunming Zhou
Quoted 9 lines
Change-Id I8eb6d7f558da05510e429d3bf1d48c8cec6c1977
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +0 -0
index bca1fb5..f3e7525 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c int amdgpu_vm_ioctl(struct drm_device *dev, · line 2555

Quoted 9 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +3 -0
@@ -2547,6 +2547,9 @@ int amdgpu_vm_ioctl(struct drm_device *dev,
 void *data, struct drm_file *filp)
         case AMDGPU_VM_OP_UNRESERVE_VMID:
             amdgpu_vm_free_reserved_vmid(adev, &fpriv->vm,
 AMDGPU_GFXHUB);
             break;
+    case AMDGPU_VM_OP_RESET:
+        fpriv->vram_lost_counter =
 atomic_read(&adev->vram_lost_counter);
+        break;

How do you envision the UMDs using this? I can mostly think of them

Quoted 8 lines

calling this ioctl when a context is created or destroyed. But that
would also allow any other remaining contexts using the same DRM file
descriptor to use all ioctls again. So, I think there needs to be a
vram_lost_counter in struct amdgpu_ctx instead of in struct amdgpu_fpriv.

struct amdgpu_fpriv for vram_lost_counter is proper place, especially
for ioctl return value.
if you need to reset ctx one by one, we can mark all contexts of that
vm, and then reset by userspace.

Quoted 7 lines
Quoted 3 lines

I'm not following. With vram_lost_counter in amdgpu_fpriv, if any context calls this ioctl, all other contexts using the same file descriptor will also be considered safe again, right?

Yes, but it really depends on userspace requirement, if you need to reset ctx one by one, we can mark all contexts of that vm to guilty, and
then reset one context by userspace.

Still not sure what you mean by that.

E.g. what do you mean by "guilty"? I thought that refers to the context
which caused a hang. But it seems like you're using it to refer to any context which hasn't reacted yet to VRAM contents being lost.

Quoted 5 lines
Quoted 1 line

When vram is lost, we treat all contexts need to reset.

Essentially, your patches only track VRAM contents being lost per file descriptor, not per context. I'm not sure (rather skeptical) that this is suitable for OpenGL UMDs, since state is usually tracked per context. Marek / Nicolai?

Oh, yeah that's a good point.

Quoted 7 lines

The problem with tracking it per context is that Vulkan also wants the ENODEV on the amdgpu_gem_va_ioct() and amdgpu_info_ioctl() which are context less.

But thinking more about this blocking those two doesn't make much sense. The VM content can be restored and why should be disallow reading GPU info?

I can re-paste the Vulkan APIs requiring ENODEV:
"

The Vulkan APIs listed below could return VK_ERROR_DEVICE_LOST according to the spec.

I tries to provide a list of u/k interfaces that could be called for each vk API.

vkCreateDevice

(quoted diff) +0 -2
-          amdgpu_device_initialize.
-          amdgpu_query_gpu_info

[...]

Quoted 4 lines

vkQueueBindSparse**

amdgpu_bo_va_op
amdgpu_bo_va_op_raw

So these *can* return VK_ERROR_DEVICE_LOST, but do they *have to* do that after a reset? :)

Signature · 6 lines
Earthling Michel Dänzer               |               http://www.amd.com
Libre software enthusiast             |             Mesa and X developer
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
#31 lore raw

Re: [PATCH 4/4] drm/amdgpu: reset fpriv vram_lost_counter

In reply to a message not in the archive

Am 17.05.2017 um 10:46 schrieb zhoucm1:

Quoted 61 lines · Christian König
Quoted 58 lines

Am 17.05.2017 um 10:01 schrieb Michel Dänzer:

Quoted 44 lines · On 17/05/17 04:13 PM, zhoucm1 wrote:
Quoted 43 lines · Michel Dänzer
Quoted 42 lines · On 17/05/17 01:28 PM, zhoucm1 wrote:
Quoted 33 lines · Michel Dänzer
Quoted 32 lines · On 17/05/17 12:04 PM, zhoucm1 wrote:
Quoted 31 lines · Michel Dänzer
Quoted 18 lines · Chunming Zhou
Quoted 17 lines
Change-Id I8eb6d7f558da05510e429d3bf1d48c8cec6c1977
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +2 -0
index bca1fb5..f3e7525 100644
@@ -2547,6 +2547,9 @@ int amdgpu_vm_ioctl(struct drm_device 
 *dev,
 void *data, struct drm_file *filp)
         case AMDGPU_VM_OP_UNRESERVE_VMID:
             amdgpu_vm_free_reserved_vmid(adev, &fpriv->vm,
 AMDGPU_GFXHUB);
             break;
+    case AMDGPU_VM_OP_RESET:
+        fpriv->vram_lost_counter =
Quoted 8 lines
Quoted 2 lines

atomic_read(&adev->vram_lost_counter);
+ break;

How do you envision the UMDs using this? I can mostly think of them calling this ioctl when a context is created or destroyed. But that would also allow any other remaining contexts using the same DRM file descriptor to use all ioctls again. So, I think there needs to be a vram_lost_counter in struct amdgpu_ctx instead of in struct amdgpu_fpriv.

struct amdgpu_fpriv for vram_lost_counter is proper place, especially for ioctl return value.
if you need to reset ctx one by one, we can mark all contexts of that vm, and then reset by userspace.

Quoted 6 lines
Quoted 3 lines

I'm not following. With vram_lost_counter in amdgpu_fpriv, if any context calls this ioctl, all other contexts using the same file descriptor will also be considered safe again, right?

Yes, but it really depends on userspace requirement, if you need to reset ctx one by one, we can mark all contexts of that vm to guilty, and then reset one context by userspace.

Still not sure what you mean by that.

Quoted 8 lines
Quoted 4 lines
Quoted 3 lines

E.g. what do you mean by "guilty"? I thought that refers to the context which caused a hang. But it seems like you're using it to refer to any context which hasn't reacted yet to VRAM contents being lost.

When vram is lost, we treat all contexts need to reset.

Essentially, your patches only track VRAM contents being lost per file descriptor, not per context. I'm not sure (rather skeptical) that this is suitable for OpenGL UMDs, since state is usually tracked per context. Marek / Nicolai?

Oh, yeah that's a good point.

The problem with tracking it per context is that Vulkan also wants the ENODEV on the amdgpu_gem_va_ioct() and amdgpu_info_ioctl() which are context less.

Quoted 8 lines
Quoted 1 line

But thinking more about this blocking those two doesn't make much sense. The VM content can be restored and why should be disallow reading GPU info?

I can re-paste the Vulkan APIs requiring ENODEV:
"

The Vulkan APIs listed below could return VK_ERROR_DEVICE_LOST according to the spec.

I tries to provide a list of u/k interfaces that could be called for each vk API.

Well those are the Vulkan requirements, but that doesn't necessary mean we must follow that on the kernel side. Keep in mind that Vulkan can't made any requirements towards the kernel driver.

IIRC we already have a query Vulkan can use to figure out if a GPU reset happened or not. So they could use that instead.

Regards,
Christian.

Quoted 47 lines

vkCreateDevice

(quoted diff) +0 -2
-amdgpu_device_initialize.
 
-amdgpu_query_gpu_info

vkQueueSubmit

-amdgpu_cs_submit

vkWaitForFences

amdgpu_cs_wait_fences

vkGetEventStatus

vkQueueWaitIdle

vkDeviceWaitIdle

vkGetQueryPoolResults**

amdgpu_cs_query_Fence_status

vkQueueBindSparse**

amdgpu_bo_va_op

amdgpu_bo_va_op_raw

vkCreateSwapchainKHR**

vkAcquireNextImageKHR**

vkQueuePresentKHR

Not related with u/k interface.**

**

Besides those listed above, I think amdgpu_cs_signal_Sem/amdgpu_cs_wait_sem should respond to gpu reset as well."

Quoted 3 lines

Christian.

Quoted 4 lines

_______________________________________________
amd-gfx mailing list [email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 46582a1e-e019-34ac-1913-ed4a2a992e4c-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

#33 lore raw

Re: [PATCH 4/4] drm/amdgpu: reset fpriv vram_lost_counter

David,

We already have a query that returns whether a device is lost. It's called amdgpu_cs_query_reset_state. That should return whether a hang was caused by a certain context or whether the hang happened but the context is innocent. You can extend it to accept no context, in which case it will return either NO_RESET (everything is OK) or UNKNOWN_RESET (= when a hang happened but the caller didn't specify the context).

Marek

Quoted 172 lines · Christian König

Am 17.05.2017 um 10:46 schrieb zhoucm1:

On 2017年05月17日 16:40, Christian König wrote:

Am 17.05.2017 um 10:01 schrieb Michel Dänzer:

On 17/05/17 04:13 PM, zhoucm1 wrote:

On 2017年05月17日 14:57, Michel Dänzer wrote:

On 17/05/17 01:28 PM, zhoucm1 wrote:

On 2017年05月17日 11:15, Michel Dänzer wrote:

On 17/05/17 12:04 PM, zhoucm1 wrote:

On 2017年05月17日 09:18, Michel Dänzer wrote:

On 16/05/17 06:25 PM, Chunming Zhou wrote:

Change-Id I8eb6d7f558da05510e429d3bf1d48c8cec6c1977
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +3 -0
index bca1fb5..f3e7525 100644
@@ -2547,6 +2547,9 @@ int amdgpu_vm_ioctl(struct drm_device *dev,
 void *data, struct drm_file *filp)
         case AMDGPU_VM_OP_UNRESERVE_VMID:
             amdgpu_vm_free_reserved_vmid(adev, &fpriv->vm,
 AMDGPU_GFXHUB);
             break;
+    case AMDGPU_VM_OP_RESET:
+        fpriv->vram_lost_counter =
 atomic_read(&adev->vram_lost_counter);
+        break;

How do you envision the UMDs using this? I can mostly think of them calling this ioctl when a context is created or destroyed. But that would also allow any other remaining contexts using the same DRM file descriptor to use all ioctls again. So, I think there needs to be a vram_lost_counter in struct amdgpu_ctx instead of in struct amdgpu_fpriv.

struct amdgpu_fpriv for vram_lost_counter is proper place, especially for ioctl return value.
if you need to reset ctx one by one, we can mark all contexts of that vm, and then reset by userspace.

I'm not following. With vram_lost_counter in amdgpu_fpriv, if any context calls this ioctl, all other contexts using the same file descriptor will also be considered safe again, right?

Yes, but it really depends on userspace requirement, if you need to reset ctx one by one, we can mark all contexts of that vm to guilty, and then reset one context by userspace.

Still not sure what you mean by that.

E.g. what do you mean by "guilty"? I thought that refers to the context which caused a hang. But it seems like you're using it to refer to any context which hasn't reacted yet to VRAM contents being lost.

When vram is lost, we treat all contexts need to reset.

Essentially, your patches only track VRAM contents being lost per file descriptor, not per context. I'm not sure (rather skeptical) that this is suitable for OpenGL UMDs, since state is usually tracked per context. Marek / Nicolai?

Oh, yeah that's a good point.

The problem with tracking it per context is that Vulkan also wants the ENODEV on the amdgpu_gem_va_ioct() and amdgpu_info_ioctl() which are context less.

But thinking more about this blocking those two doesn't make much sense. The VM content can be restored and why should be disallow reading GPU info?

I can re-paste the Vulkan APIs requiring ENODEV:
"

The Vulkan APIs listed below could return VK_ERROR_DEVICE_LOST according to the spec.

I tries to provide a list of u/k interfaces that could be called for each vk API.

Well those are the Vulkan requirements, but that doesn't necessary mean we must follow that on the kernel side. Keep in mind that Vulkan can't made any requirements towards the kernel driver.

IIRC we already have a query Vulkan can use to figure out if a GPU reset happened or not. So they could use that instead.

Regards,
Christian.

vkCreateDevice

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +0 -2
-          amdgpu_device_initialize.
 
-          amdgpu_query_gpu_info

vkQueueSubmit

- amdgpu_cs_submit

vkWaitForFences

amdgpu_cs_wait_fences

vkGetEventStatus

vkQueueWaitIdle

vkDeviceWaitIdle

vkGetQueryPoolResults

amdgpu_cs_query_Fence_status

vkQueueBindSparse

amdgpu_bo_va_op

amdgpu_bo_va_op_raw

vkCreateSwapchainKHR

vkAcquireNextImageKHR

vkQueuePresentKHR

Not related with u/k interface.

Besides those listed above, I think amdgpu_cs_signal_Sem/amdgpu_cs_wait_sem should respond to gpu reset as well."

Christian.

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

Quoted 5 lines

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): CAAxE2A7sRZXx3MnRSO76DW=61X06nfVs2AHe_a-r+K+46tfJPw-JsoAwUIsXosN+BqQ9rBEUg@public.gmane.org

#35 lore raw

RE: [PATCH 4/4] drm/amdgpu: reset fpriv vram_lost_counter

Yes, agree, it's easy to extend it as your like.
For innocent context, which would need us to find out the context of guilty job, which is TODO.

Regards,
David Zhou

-----Original Message-----
From: Marek Olšák [mailto:[email protected]]
Sent: Wednesday, May 17, 2017 5:50 PM
To: Christian König <[email protected]>
Cc: Zhou, David(ChunMing) <[email protected]>; Michel Dänzer <[email protected]>; amd-gfx mailing list <[email protected]>
Subject: Re: [PATCH 4/4] drm/amdgpu: reset fpriv vram_lost_counter

David,

We already have a query that returns whether a device is lost. It's called amdgpu_cs_query_reset_state. That should return whether a hang was caused by a certain context or whether the hang happened but the context is innocent. You can extend it to accept no context, in which case it will return either NO_RESET (everything is OK) or UNKNOWN_RESET (= when a hang happened but the caller didn't specify the context).

Marek

Quoted 173 lines · Christian König

Am 17.05.2017 um 10:46 schrieb zhoucm1:

On 2017年05月17日 16:40, Christian König wrote:

Am 17.05.2017 um 10:01 schrieb Michel Dänzer:

On 17/05/17 04:13 PM, zhoucm1 wrote:

On 2017年05月17日 14:57, Michel Dänzer wrote:

On 17/05/17 01:28 PM, zhoucm1 wrote:

On 2017年05月17日 11:15, Michel Dänzer wrote:

On 17/05/17 12:04 PM, zhoucm1 wrote:

On 2017年05月17日 09:18, Michel Dänzer wrote:

On 16/05/17 06:25 PM, Chunming Zhou wrote:

Change-Id I8eb6d7f558da05510e429d3bf1d48c8cec6c1977
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +3 -0
index bca1fb5..f3e7525 100644
@@ -2547,6 +2547,9 @@ int amdgpu_vm_ioctl(struct drm_device *dev, void 
 *data, struct drm_file *filp)
         case AMDGPU_VM_OP_UNRESERVE_VMID:
             amdgpu_vm_free_reserved_vmid(adev, &fpriv->vm, 
 AMDGPU_GFXHUB);
             break;
+    case AMDGPU_VM_OP_RESET:
+        fpriv->vram_lost_counter =
 atomic_read(&adev->vram_lost_counter);
+        break;

How do you envision the UMDs using this? I can mostly think of them calling this ioctl when a context is created or destroyed. But that would also allow any other remaining contexts using the same DRM file descriptor to use all ioctls again. So, I think there needs to be a vram_lost_counter in struct amdgpu_ctx instead of in struct amdgpu_fpriv.

struct amdgpu_fpriv for vram_lost_counter is proper place, especially for ioctl return value.
if you need to reset ctx one by one, we can mark all contexts of that vm, and then reset by userspace.

I'm not following. With vram_lost_counter in amdgpu_fpriv, if any context calls this ioctl, all other contexts using the same file descriptor will also be considered safe again, right?

Yes, but it really depends on userspace requirement, if you need to reset ctx one by one, we can mark all contexts of that vm to guilty, and then reset one context by userspace.

Still not sure what you mean by that.

E.g. what do you mean by "guilty"? I thought that refers to the context which caused a hang. But it seems like you're using it to refer to any context which hasn't reacted yet to VRAM contents being lost.

When vram is lost, we treat all contexts need to reset.

Essentially, your patches only track VRAM contents being lost per file descriptor, not per context. I'm not sure (rather skeptical) that this is suitable for OpenGL UMDs, since state is usually tracked per context. Marek / Nicolai?

Oh, yeah that's a good point.

The problem with tracking it per context is that Vulkan also wants the ENODEV on the amdgpu_gem_va_ioct() and amdgpu_info_ioctl() which are context less.

But thinking more about this blocking those two doesn't make much sense. The VM content can be restored and why should be disallow reading GPU info?

I can re-paste the Vulkan APIs requiring ENODEV:
"

The Vulkan APIs listed below could return VK_ERROR_DEVICE_LOST according to the spec.

I tries to provide a list of u/k interfaces that could be called for each vk API.

Well those are the Vulkan requirements, but that doesn't necessary mean we must follow that on the kernel side. Keep in mind that Vulkan can't made any requirements towards the kernel driver.

IIRC we already have a query Vulkan can use to figure out if a GPU reset happened or not. So they could use that instead.

Regards,
Christian.

vkCreateDevice

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +0 -2
-          amdgpu_device_initialize.
 
-          amdgpu_query_gpu_info

vkQueueSubmit

- amdgpu_cs_submit

vkWaitForFences

amdgpu_cs_wait_fences

vkGetEventStatus

vkQueueWaitIdle

vkDeviceWaitIdle

vkGetQueryPoolResults

amdgpu_cs_query_Fence_status

vkQueueBindSparse

amdgpu_bo_va_op

amdgpu_bo_va_op_raw

vkCreateSwapchainKHR

vkAcquireNextImageKHR

vkQueuePresentKHR

Not related with u/k interface.

Besides those listed above, I think
amdgpu_cs_signal_Sem/amdgpu_cs_wait_sem
should respond to gpu reset as well."

Christian.

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

Quoted 5 lines

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

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

Am 16.05.2017 um 11:25 schrieb Chunming Zhou:

Quoted 6 lines

bakup first 64 byte of gart table as reset magic, check if magic is same after gpu hw reset.
v2: use memcmp instead of manual innovation.

Change-Id I9a73720da4084ea8677c3031dfb62e8157ee5704

Patch #1-#3 are Reviewed-by: Christian König <[email protected]>

Patch #4:

You need to add the new enum on line 591 or otherwise you will get an "unsupported operation" error.

Line 604 should be changed as well or otherwise we need a BO for this operation.

A libdrm test case to just call this IOCTL would probably be a good idea.

Additional to that I would ping Marek (Mesa) and Michel (DDX) for their opinion on this. Could be that this is completely superfluous and the UMDs needs something else.

Regards,
Christian.

Quoted 73 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu.h        |  2 ++
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 20 +++++++++++++++++++-
  2 files changed, 21 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +2 -0
index de08ff0..f9da215 100644
@@ -1502,6 +1502,7 @@ struct amdgpu_ssg {
  #endif
  };
 
+#define AMDGPU_RESET_MAGIC_NUM 64
  struct amdgpu_device {
  	struct device			*dev;
  	struct drm_device		*ddev;
@@ -1705,6 +1706,7 @@ struct amdgpu_device {
 
  	/* record hw reset is performed */
  	bool has_hw_reset;
+	u8				reset_magic[AMDGPU_RESET_MAGIC_NUM];
 
  };
 
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +17 -1
index 0a31fb1..c56ae4a 100644
@@ -1685,6 +1685,17 @@ static int amdgpu_init(struct amdgpu_device *adev)
  	return 0;
  }
 
+static void amdgpu_fill_reset_magic(struct amdgpu_device *adev)
+{
+	memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM);
+}
+
+static bool amdgpu_check_vram_lost(struct amdgpu_device *adev)
+{
+	return !!memcmp(adev->gart.ptr, adev->reset_magic,
+			AMDGPU_RESET_MAGIC_NUM);
+}
+
  static int amdgpu_late_init(struct amdgpu_device *adev)
  {
  	int i = 0, r;
@@ -1715,6 +1726,8 @@ static int amdgpu_late_init(struct amdgpu_device *adev)
  		}
  	}
 
+	amdgpu_fill_reset_magic(adev);
+
  	return 0;
  }
 
@@ -2830,7 +2843,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
  	struct drm_atomic_state *state = NULL;
  	int i, r;
  	int resched;
-	bool need_full_reset;
+	bool need_full_reset, vram_lost = false;
 
  	if (amdgpu_sriov_vf(adev))
  		return amdgpu_sriov_gpu_reset(adev, true);
@@ -2899,12 +2912,17 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
  			r = amdgpu_resume_phase1(adev);
  			if (r)
  				goto out;
+			vram_lost = amdgpu_check_vram_lost(adev);
+			if (vram_lost)
+				DRM_ERROR("VRAM is lost!\n");
  			r = amdgpu_ttm_recover_gart(adev);

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c int amdgpu_gpu_reset(struct amdgpu_device *adev) · line 2928

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +2 -0
  			if (r)
  				goto out;
  			r = amdgpu_resume_phase2(adev);
  			if (r)
  				goto out;
+			if (vram_lost)
+				amdgpu_fill_reset_magic(adev);
  		}
  	}
  out:

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 0c1d89c5-65ea-cdad-100a-80d0377b865c-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

author
#38 lore raw
Quoted 9 lines · Christian König

Am 16.05.2017 um 11:25 schrieb Chunming Zhou:

Quoted 6 lines

bakup first 64 byte of gart table as reset magic, check if magic is same after gpu hw reset.
v2: use memcmp instead of manual innovation.

Change-Id I9a73720da4084ea8677c3031dfb62e8157ee5704

Patch #1-#3 are Reviewed-by: Christian König <[email protected]>

pushed.

Quoted 6 lines

Patch #4:

You need to add the new enum on line 591 or otherwise you will get an "unsupported operation" error.

Line 604 should be changed as well or otherwise we need a BO for this operation.

are you sure you are talking this patch#4? I cannot address what you said.

Quoted 4 lines

A libdrm test case to just call this IOCTL would probably be a good idea.

Additional to that I would ping Marek (Mesa) and Michel (DDX) for their opinion on this. Could be that this is completely superfluous and the UMDs needs something else.

Michel seems have different opinion/concern, maybe we need more discussions before we make new interfaces.

Thanks,
David Zhou

Quoted 74 lines

Regards,
Christian.

Quoted 70 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu.h        |  2 ++
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 20 +++++++++++++++++++-
  2 files changed, 21 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +1 -0
index de08ff0..f9da215 100644
@@ -1502,6 +1502,7 @@ struct amdgpu_ssg {
  #endif
  };
  +#define AMDGPU_RESET_MAGIC_NUM 64
  struct amdgpu_device {
      struct device            *dev;
      struct drm_device        *ddev;
@@ -1705,6 +1706,7 @@ struct amdgpu_device {
        /* record hw reset is performed */
      bool has_hw_reset;
+    u8                reset_magic[AMDGPU_RESET_MAGIC_NUM];
    };
  diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c 

b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index 0a31fb1..c56ae4a 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +15 -1
@@ -1685,6 +1685,17 @@ static int amdgpu_init(struct amdgpu_device 
 *adev)
      return 0;
  }
  +static void amdgpu_fill_reset_magic(struct amdgpu_device *adev)
+{
+    memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM);
+}
+
+static bool amdgpu_check_vram_lost(struct amdgpu_device *adev)
+{
+    return !!memcmp(adev->gart.ptr, adev->reset_magic,
+            AMDGPU_RESET_MAGIC_NUM);
+}
+
  static int amdgpu_late_init(struct amdgpu_device *adev)
  {
      int i = 0, r;
@@ -1715,6 +1726,8 @@ static int amdgpu_late_init(struct 
 amdgpu_device *adev)
          }
      }
  +    amdgpu_fill_reset_magic(adev);
+
      return 0;
  }
  @@ -2830,7 +2843,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
      struct drm_atomic_state *state = NULL;
      int i, r;
      int resched;
-    bool need_full_reset;
+    bool need_full_reset, vram_lost = false;
        if (amdgpu_sriov_vf(adev))
          return amdgpu_sriov_gpu_reset(adev, true);
@@ -2899,12 +2912,17 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
              r = amdgpu_resume_phase1(adev);
              if (r)
                  goto out;
+            vram_lost = amdgpu_check_vram_lost(adev);
+            if (vram_lost)
+                DRM_ERROR("VRAM is lost!\n");
              r = amdgpu_ttm_recover_gart(adev);
              if (r)
                  goto out;
Quoted 10 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c int amdgpu_gpu_reset(struct amdgpu_device *adev) · line 2928

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +2 -0
              r = amdgpu_resume_phase2(adev);
              if (r)
                  goto out;
+            if (vram_lost)
+                amdgpu_fill_reset_magic(adev);
          }
      }
  out:

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx