Skip to contentSkip to conversation

drm/amdgpu: fix gpu reset issue

PATCH NAK Roger.He 5 participants 18 messages Apr 21, 2017 – May 4, 2017 13 messages not in the archive

Subsystems (2)
lore
Share & tools
lore

Roger.He

author
#0 lore raw

PATCH 3/3 drm/amdgpu: fix gpu reset issue

Change-Id Ib77d33a09f348ebf2e3a9d7861411f4b951ebf7c
---
 drivers/gpu/drm/amd/amdgpu/amdgpu.h        |  6 ++++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 16 +++++++++++++++-
 drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c    |  4 ++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c  |  7 ++++++-
 4 files changed, 31 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +6 -0
index 71364f5..ab0ffa8 100644
@@ -1526,6 +1526,7 @@ struct amdgpu_device {
 	atomic64_t			num_bytes_moved;
 	atomic64_t			num_evictions;
 	atomic_t			gpu_reset_counter;
+	atomic_t			gpu_reset_state;
 
 	/* data for buffer migration throttling */
 	struct {
@@ -1851,6 +1852,11 @@ amdgpu_get_sdma_instance(struct amdgpu_ring *ring)
 #define amdgpu_psp_check_fw_loading_status(adev, i) (adev)->firmware.funcs->check_fw_loading_status((adev), (i))
 
 /* Common functions */
+static inline int amdgpu_device_is_reset(struct amdgpu_device *adev)
+{
+	return atomic_read(&adev->gpu_reset_state);
+}
+
 int amdgpu_gpu_reset(struct amdgpu_device *adev);
 bool amdgpu_need_backup(struct amdgpu_device *adev);
 void amdgpu_pci_config_reset(struct amdgpu_device *adev);
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +15 -1
index f882496..0fb4716 100644
@@ -1894,6 +1894,7 @@ int amdgpu_device_init(struct amdgpu_device *adev,
 	mutex_init(&adev->grbm_idx_mutex);
 	mutex_init(&adev->mn_lock);
 	hash_init(adev->mn_hash);
+	atomic_set(&adev->gpu_reset_state, 0);
 
 	amdgpu_check_arguments(adev);
 
@@ -2655,6 +2656,18 @@ int amdgpu_sriov_gpu_reset(struct amdgpu_device *adev, bool voluntary)
 }
 
 /**
+ * amdgpu_device_set_reset_state - set gpu reset state
+ *
+ * @adev: amdgpu device pointer
+ * @state: true when start to reset gpu; false: reset done
+ */
+static inline void amdgpu_device_set_reset_state(struct amdgpu_device *adev,
+							bool state)
+{
+	atomic_set(&adev->gpu_reset_state, state);
+}
+
+/**
  * amdgpu_gpu_reset - reset the asic
  *
  * @adev: amdgpu device pointer
@@ -2678,7 +2691,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
 	}
 
 	atomic_inc(&adev->gpu_reset_counter);
-
+	amdgpu_device_set_reset_state(adev, true);
 	/* block TTM */
 	resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
 	/* store modesetting */
@@ -2811,6 +2824,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
 		dev_info(adev->dev, "GPU reset failed\n");
 	}
 
+	amdgpu_device_set_reset_state(adev, false);
 	return r;
 }
 
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +4 -0
index ead00d7..8cc14af 100644
@@ -679,11 +679,15 @@ long amdgpu_drm_ioctl(struct file *filp,
 	struct drm_file *file_priv = filp->private_data;
 	struct drm_device *dev;
 	long ret;
+
 	dev = file_priv->minor->dev;
 	ret = pm_runtime_get_sync(dev->dev);
 	if (ret < 0)
 		return ret;
 
+	while (amdgpu_device_is_reset(dev->dev_private))
+		msleep(100);
+
 	ret = drm_ioctl(filp, cmd, arg);
 
 	pm_runtime_mark_last_busy(dev->dev);
drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c +6 -1
index 2648291..22b8059 100644
@@ -36,10 +36,15 @@
 long amdgpu_kms_compat_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
 {
 	unsigned int nr = DRM_IOCTL_NR(cmd);
+	struct drm_file *file_priv = filp->private_data;
+	struct amdgpu_device *adev = file_priv->minor->dev->dev_private;
 	int ret;
 
-	if (nr < DRM_COMMAND_BASE)
+	if (nr < DRM_COMMAND_BASE) {
+		while (amdgpu_device_is_reset(adev))
+			msleep(100);
 		return drm_compat_ioctl(filp, cmd, arg);
+	}
 
 	ret = amdgpu_drm_ioctl(filp, cmd, arg);
 
Signature · 6 lines
2.7.4

_______________________________________________
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): 1492758490-12631-1-git-send-email-Hongbo.He-5C7GfCeVMHo@public.gmane.org

#2 lore raw
Quoted 2 lines · Roger.He
Change-Id Ib77d33a09f348ebf2e3a9d7861411f4b951ebf7c

Please add more explaination in comment, like "prevent ioctl during gpu reset, otherwise many un-expected behaviour happen."

Regards,
David Zhou

Quoted 110 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu.h        |  6 ++++++
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 16 +++++++++++++++-
  drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c    |  4 ++++
  drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c  |  7 ++++++-
  4 files changed, 31 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +6 -0
index 71364f5..ab0ffa8 100644
@@ -1526,6 +1526,7 @@ struct amdgpu_device {
  	atomic64_t			num_bytes_moved;
  	atomic64_t			num_evictions;
  	atomic_t			gpu_reset_counter;
+	atomic_t			gpu_reset_state;
 
  	/* data for buffer migration throttling */
  	struct {
@@ -1851,6 +1852,11 @@ amdgpu_get_sdma_instance(struct amdgpu_ring *ring)
  #define amdgpu_psp_check_fw_loading_status(adev, i) (adev)->firmware.funcs->check_fw_loading_status((adev), (i))
 
  /* Common functions */
+static inline int amdgpu_device_is_reset(struct amdgpu_device *adev)
+{
+	return atomic_read(&adev->gpu_reset_state);
+}
+
  int amdgpu_gpu_reset(struct amdgpu_device *adev);
  bool amdgpu_need_backup(struct amdgpu_device *adev);
  void amdgpu_pci_config_reset(struct amdgpu_device *adev);
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +15 -1
index f882496..0fb4716 100644
@@ -1894,6 +1894,7 @@ int amdgpu_device_init(struct amdgpu_device *adev,
  	mutex_init(&adev->grbm_idx_mutex);
  	mutex_init(&adev->mn_lock);
  	hash_init(adev->mn_hash);
+	atomic_set(&adev->gpu_reset_state, 0);
 
  	amdgpu_check_arguments(adev);
 
@@ -2655,6 +2656,18 @@ int amdgpu_sriov_gpu_reset(struct amdgpu_device *adev, bool voluntary)
  }
 
  /**
+ * amdgpu_device_set_reset_state - set gpu reset state
+ *
+ * @adev: amdgpu device pointer
+ * @state: true when start to reset gpu; false: reset done
+ */
+static inline void amdgpu_device_set_reset_state(struct amdgpu_device *adev,
+							bool state)
+{
+	atomic_set(&adev->gpu_reset_state, state);
+}
+
+/**
   * amdgpu_gpu_reset - reset the asic
   *
   * @adev: amdgpu device pointer
@@ -2678,7 +2691,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
  	}
 
  	atomic_inc(&adev->gpu_reset_counter);
-
+	amdgpu_device_set_reset_state(adev, true);
  	/* block TTM */
  	resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
  	/* store modesetting */
@@ -2811,6 +2824,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
  		dev_info(adev->dev, "GPU reset failed\n");
  	}
 
+	amdgpu_device_set_reset_state(adev, false);
  	return r;
  }
 
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +4 -0
index ead00d7..8cc14af 100644
@@ -679,11 +679,15 @@ long amdgpu_drm_ioctl(struct file *filp,
  	struct drm_file *file_priv = filp->private_data;
  	struct drm_device *dev;
  	long ret;
+
  	dev = file_priv->minor->dev;
  	ret = pm_runtime_get_sync(dev->dev);
  	if (ret < 0)
  		return ret;
 
+	while (amdgpu_device_is_reset(dev->dev_private))
+		msleep(100);
+
  	ret = drm_ioctl(filp, cmd, arg);
 
  	pm_runtime_mark_last_busy(dev->dev);
drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c +2 -0
index 2648291..22b8059 100644
@@ -36,10 +36,15 @@
  long amdgpu_kms_compat_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
  {
  	unsigned int nr = DRM_IOCTL_NR(cmd);
+	struct drm_file *file_priv = filp->private_data;
+	struct amdgpu_device *adev = file_priv->minor->dev->dev_private;
  	int ret;

drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c line 50

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c +4 -1
 
-	if (nr < DRM_COMMAND_BASE)
+	if (nr < DRM_COMMAND_BASE) {
+		while (amdgpu_device_is_reset(adev))
+			msleep(100);
  		return drm_compat_ioctl(filp, cmd, arg);
+	}
 
  	ret = amdgpu_drm_ioctl(filp, cmd, arg);
 

_______________________________________________
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]

#4 lore raw

First of all, Your patch will impact on SR-IOV case, because "gpu_reset_state" will be referenced by two place:
1) IOCTL; 2)GPU-RESET ,
SRIOV use srio_gpu_reset, but SRIOV use the same IOCTL routines, that way the gpu_reset_state is not consistent for SR-IOV case

I suggest you firstly unify this gpu_reset_state for both SRIOV and bare-metal case

For sriov, we have "adev->gfx.in_reset " and you now have "adev->gpu_reset_state", I suggest you replace all "adev->gfx.in_reset" with "adev->gpu_reset_state",

BR Monk

-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of zhoucm1
Sent: Friday, April 21, 2017 3:18 PM
To: He, Hongbo <[email protected]>; [email protected]
Subject: Re: [PATCH 3/3] drm/amdgpu: fix gpu reset issue

Quoted 2 lines · Roger.He
Change-Id Ib77d33a09f348ebf2e3a9d7861411f4b951ebf7c

Please add more explaination in comment, like "prevent ioctl during gpu reset, otherwise many un-expected behaviour happen."

Regards,
David Zhou

Quoted 113 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu.h        |  6 ++++++
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 16 +++++++++++++++-
  drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c    |  4 ++++
  drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c  |  7 ++++++-
  4 files changed, 31 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +6 -0
index 71364f5..ab0ffa8 100644
@@ -1526,6 +1526,7 @@ struct amdgpu_device {
  	atomic64_t			num_bytes_moved;
  	atomic64_t			num_evictions;
  	atomic_t			gpu_reset_counter;
+	atomic_t			gpu_reset_state;
 
  	/* data for buffer migration throttling */
  	struct {
@@ -1851,6 +1852,11 @@ amdgpu_get_sdma_instance(struct amdgpu_ring *ring)
  #define amdgpu_psp_check_fw_loading_status(adev, i) 
 (adev)->firmware.funcs->check_fw_loading_status((adev), (i))
 
  /* Common functions */
+static inline int amdgpu_device_is_reset(struct amdgpu_device *adev) 
+{
+	return atomic_read(&adev->gpu_reset_state);
+}
+
  int amdgpu_gpu_reset(struct amdgpu_device *adev);
  bool amdgpu_need_backup(struct amdgpu_device *adev);

void amdgpu_pci_config_reset(struct amdgpu_device *adev); diff --git
a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index f882496..0fb4716 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +13 -1
@@ -1894,6 +1894,7 @@ int amdgpu_device_init(struct amdgpu_device *adev,
  	mutex_init(&adev->grbm_idx_mutex);
  	mutex_init(&adev->mn_lock);
  	hash_init(adev->mn_hash);
+	atomic_set(&adev->gpu_reset_state, 0);
 
  	amdgpu_check_arguments(adev);
 
@@ -2655,6 +2656,18 @@ int amdgpu_sriov_gpu_reset(struct amdgpu_device *adev, bool voluntary)
  }
 
  /**
+ * amdgpu_device_set_reset_state - set gpu reset state
+ *
+ * @adev: amdgpu device pointer
+ * @state: true when start to reset gpu; false: reset done  */ static 
+inline void amdgpu_device_set_reset_state(struct amdgpu_device *adev,
+							bool state)
+{
+	atomic_set(&adev->gpu_reset_state, state); }
+
+/**
   * amdgpu_gpu_reset - reset the asic
   *
   * @adev: amdgpu device pointer
@@ -2678,7 +2691,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
  	}
 
  	atomic_inc(&adev->gpu_reset_counter);
-
+	amdgpu_device_set_reset_state(adev, true);
  	/* block TTM */
  	resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
  	/* store modesetting */
@@ -2811,6 +2824,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
  		dev_info(adev->dev, "GPU reset failed\n");
  	}
 
+	amdgpu_device_set_reset_state(adev, false);
  	return r;
  }
 
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +4 -0
index ead00d7..8cc14af 100644
@@ -679,11 +679,15 @@ long amdgpu_drm_ioctl(struct file *filp,
  	struct drm_file *file_priv = filp->private_data;
  	struct drm_device *dev;
  	long ret;
+
  	dev = file_priv->minor->dev;
  	ret = pm_runtime_get_sync(dev->dev);
  	if (ret < 0)
  		return ret;
 
+	while (amdgpu_device_is_reset(dev->dev_private))
+		msleep(100);
+
  	ret = drm_ioctl(filp, cmd, arg);
 
  	pm_runtime_mark_last_busy(dev->dev);
drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c +2 -0
index 2648291..22b8059 100644
@@ -36,10 +36,15 @@
  long amdgpu_kms_compat_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
  {
  	unsigned int nr = DRM_IOCTL_NR(cmd);
+	struct drm_file *file_priv = filp->private_data;
+	struct amdgpu_device *adev = file_priv->minor->dev->dev_private;
  	int ret;

drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c line 50

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c +4 -1
 
-	if (nr < DRM_COMMAND_BASE)
+	if (nr < DRM_COMMAND_BASE) {
+		while (amdgpu_device_is_reset(adev))
+			msleep(100);
  		return drm_compat_ioctl(filp, cmd, arg);
+	}
 
  	ret = amdgpu_drm_ioctl(filp, cmd, arg);
 

_______________________________________________
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): DM5PR12MB161079A7788AC1D9E4CD5D6284EA0-2J9CzHegvk++jCVTvoAFKAdYzm3356FpvxpqHgZTriW3zl9H0oFU5g@public.gmane.org

#6 lore raw
Quoted 1 line

I suggest you firstly unify this gpu_reset_state for both SRIOV and bare-metal case

Yeah, that sounds like a good idea to me as well.

I suggest to have a function which covers things necessary for both cases, e.g. stopping the scheduler and blocking TTM.

And inside that one we call specific functions for bare metal and SRIOV for the special handling.

Regards,
Christian.

Am 04.05.2017 um 12:42 schrieb Liu, Monk:

Quoted 149 lines

First of all, Your patch will impact on SR-IOV case, because "gpu_reset_state" will be referenced by two place:
1) IOCTL; 2)GPU-RESET ,
SRIOV use srio_gpu_reset, but SRIOV use the same IOCTL routines, that way the gpu_reset_state is not consistent for SR-IOV case

I suggest you firstly unify this gpu_reset_state for both SRIOV and bare-metal case

For sriov, we have "adev->gfx.in_reset " and you now have "adev->gpu_reset_state", I suggest you replace all "adev->gfx.in_reset" with "adev->gpu_reset_state",

BR Monk

-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of zhoucm1
Sent: Friday, April 21, 2017 3:18 PM
To: He, Hongbo <[email protected]>; [email protected]
Subject: Re: [PATCH 3/3] drm/amdgpu: fix gpu reset issue

Quoted 2 lines · Roger.He
Change-Id Ib77d33a09f348ebf2e3a9d7861411f4b951ebf7c

Please add more explaination in comment, like "prevent ioctl during gpu reset, otherwise many un-expected behaviour happen."

Regards,
David Zhou

Quoted 121 lines
---
   drivers/gpu/drm/amd/amdgpu/amdgpu.h        |  6 ++++++
   drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 16 +++++++++++++++-
   drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c    |  4 ++++
   drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c  |  7 ++++++-
   4 files changed, 31 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +6 -0
index 71364f5..ab0ffa8 100644
@@ -1526,6 +1526,7 @@ struct amdgpu_device {
   	atomic64_t			num_bytes_moved;
   	atomic64_t			num_evictions;
   	atomic_t			gpu_reset_counter;
+	atomic_t			gpu_reset_state;
 
   	/* data for buffer migration throttling */
   	struct {
@@ -1851,6 +1852,11 @@ amdgpu_get_sdma_instance(struct amdgpu_ring *ring)
   #define amdgpu_psp_check_fw_loading_status(adev, i)
 (adev)->firmware.funcs->check_fw_loading_status((adev), (i))
 
   /* Common functions */
+static inline int amdgpu_device_is_reset(struct amdgpu_device *adev)
+{
+	return atomic_read(&adev->gpu_reset_state);
+}
+
   int amdgpu_gpu_reset(struct amdgpu_device *adev);
   bool amdgpu_need_backup(struct amdgpu_device *adev);
   void amdgpu_pci_config_reset(struct amdgpu_device *adev); diff --git

a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index f882496..0fb4716 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +13 -1
@@ -1894,6 +1894,7 @@ int amdgpu_device_init(struct amdgpu_device *adev,
   	mutex_init(&adev->grbm_idx_mutex);
   	mutex_init(&adev->mn_lock);
   	hash_init(adev->mn_hash);
+	atomic_set(&adev->gpu_reset_state, 0);
 
   	amdgpu_check_arguments(adev);
 
@@ -2655,6 +2656,18 @@ int amdgpu_sriov_gpu_reset(struct amdgpu_device *adev, bool voluntary)
   }
 
   /**
+ * amdgpu_device_set_reset_state - set gpu reset state
+ *
+ * @adev: amdgpu device pointer
+ * @state: true when start to reset gpu; false: reset done  */ static
+inline void amdgpu_device_set_reset_state(struct amdgpu_device *adev,
+							bool state)
+{
+	atomic_set(&adev->gpu_reset_state, state); }
+
+/**
    * amdgpu_gpu_reset - reset the asic
    *
    * @adev: amdgpu device pointer
@@ -2678,7 +2691,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
   	}
 
   	atomic_inc(&adev->gpu_reset_counter);
-
+	amdgpu_device_set_reset_state(adev, true);
   	/* block TTM */
   	resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
   	/* store modesetting */
@@ -2811,6 +2824,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
   		dev_info(adev->dev, "GPU reset failed\n");
   	}
 
+	amdgpu_device_set_reset_state(adev, false);
   	return r;
   }
 
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +4 -0
index ead00d7..8cc14af 100644
@@ -679,11 +679,15 @@ long amdgpu_drm_ioctl(struct file *filp,
   	struct drm_file *file_priv = filp->private_data;
   	struct drm_device *dev;
   	long ret;
+
   	dev = file_priv->minor->dev;
   	ret = pm_runtime_get_sync(dev->dev);
   	if (ret < 0)
   		return ret;
 
+	while (amdgpu_device_is_reset(dev->dev_private))
+		msleep(100);
+
   	ret = drm_ioctl(filp, cmd, arg);
 
   	pm_runtime_mark_last_busy(dev->dev);
drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c +6 -1
index 2648291..22b8059 100644
@@ -36,10 +36,15 @@
   long amdgpu_kms_compat_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
   {
   	unsigned int nr = DRM_IOCTL_NR(cmd);
+	struct drm_file *file_priv = filp->private_data;
+	struct amdgpu_device *adev = file_priv->minor->dev->dev_private;
   	int ret;
 
-	if (nr < DRM_COMMAND_BASE)
+	if (nr < DRM_COMMAND_BASE) {
+		while (amdgpu_device_is_reset(adev))
+			msleep(100);
   		return drm_compat_ioctl(filp, cmd, arg);
+	}
 
Quoted 10 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c line 50

Quoted 2 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c +0 -0
   	ret = amdgpu_drm_ioctl(filp, cmd, arg);
 

_______________________________________________
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): c88049b2-2440-142e-129a-542c136b8d61-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

#8 lore raw

I still want to keep different police and design for bare-metal and sr-iov on gpu reset e.g. what I designed is per sched reset, not that like current bare-metal's reset all sched, etc...

after we all make this feature stable, we can consider merge most part into one

BR Monk

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: Thursday, May 04, 2017 7:30 PM
To: Liu, Monk <[email protected]>; Zhou, David(ChunMing) <[email protected]>; He, Hongbo <[email protected]>; [email protected]
Subject: Re: [PATCH 3/3] drm/amdgpu: fix gpu reset issue

Quoted 2 lines

I suggest you firstly unify this gpu_reset_state for both SRIOV and bare-metal case

Yeah, that sounds like a good idea to me as well.

I suggest to have a function which covers things necessary for both cases, e.g. stopping the scheduler and blocking TTM.

And inside that one we call specific functions for bare metal and SRIOV for the special handling.

Regards,
Christian.

Am 04.05.2017 um 12:42 schrieb Liu, Monk:

Quoted 153 lines

First of all, Your patch will impact on SR-IOV case, because "gpu_reset_state" will be referenced by two place:
1) IOCTL; 2)GPU-RESET ,
SRIOV use srio_gpu_reset, but SRIOV use the same IOCTL routines, that
way the gpu_reset_state is not consistent for SR-IOV case

I suggest you firstly unify this gpu_reset_state for both SRIOV and bare-metal case

For sriov, we have "adev->gfx.in_reset " and you now have "adev->gpu_reset_state", I suggest you replace all "adev->gfx.in_reset" with "adev->gpu_reset_state",

BR Monk

-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of zhoucm1
Sent: Friday, April 21, 2017 3:18 PM
To: He, Hongbo <[email protected]>; [email protected]
Subject: Re: [PATCH 3/3] drm/amdgpu: fix gpu reset issue

Quoted 2 lines · Roger.He
Change-Id Ib77d33a09f348ebf2e3a9d7861411f4b951ebf7c

Please add more explaination in comment, like "prevent ioctl during gpu reset, otherwise many un-expected behaviour happen."

Regards,
David Zhou

Quoted 121 lines
---
   drivers/gpu/drm/amd/amdgpu/amdgpu.h        |  6 ++++++
   drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 16 +++++++++++++++-
   drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c    |  4 ++++
   drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c  |  7 ++++++-
   4 files changed, 31 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +6 -0
index 71364f5..ab0ffa8 100644
@@ -1526,6 +1526,7 @@ struct amdgpu_device {
   	atomic64_t			num_bytes_moved;
   	atomic64_t			num_evictions;
   	atomic_t			gpu_reset_counter;
+	atomic_t			gpu_reset_state;
 
   	/* data for buffer migration throttling */
   	struct {
@@ -1851,6 +1852,11 @@ amdgpu_get_sdma_instance(struct amdgpu_ring *ring)
   #define amdgpu_psp_check_fw_loading_status(adev, i) 
 (adev)->firmware.funcs->check_fw_loading_status((adev), (i))
 
   /* Common functions */
+static inline int amdgpu_device_is_reset(struct amdgpu_device *adev) 
+{
+	return atomic_read(&adev->gpu_reset_state);
+}
+
   int amdgpu_gpu_reset(struct amdgpu_device *adev);
   bool amdgpu_need_backup(struct amdgpu_device *adev);

void amdgpu_pci_config_reset(struct amdgpu_device *adev); diff
--git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index f882496..0fb4716 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +13 -1
@@ -1894,6 +1894,7 @@ int amdgpu_device_init(struct amdgpu_device *adev,
   	mutex_init(&adev->grbm_idx_mutex);
   	mutex_init(&adev->mn_lock);
   	hash_init(adev->mn_hash);
+	atomic_set(&adev->gpu_reset_state, 0);
 
   	amdgpu_check_arguments(adev);
 
@@ -2655,6 +2656,18 @@ int amdgpu_sriov_gpu_reset(struct amdgpu_device *adev, bool voluntary)
   }
 
   /**
+ * amdgpu_device_set_reset_state - set gpu reset state
+ *
+ * @adev: amdgpu device pointer
+ * @state: true when start to reset gpu; false: reset done  */ 
+static inline void amdgpu_device_set_reset_state(struct amdgpu_device *adev,
+							bool state)
+{
+	atomic_set(&adev->gpu_reset_state, state); }
+
+/**
    * amdgpu_gpu_reset - reset the asic
    *
    * @adev: amdgpu device pointer
@@ -2678,7 +2691,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
   	}
 
   	atomic_inc(&adev->gpu_reset_counter);
-
+	amdgpu_device_set_reset_state(adev, true);
   	/* block TTM */
   	resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
   	/* store modesetting */
@@ -2811,6 +2824,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
   		dev_info(adev->dev, "GPU reset failed\n");
   	}
 
+	amdgpu_device_set_reset_state(adev, false);
   	return r;
   }
 
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +4 -0
index ead00d7..8cc14af 100644
@@ -679,11 +679,15 @@ long amdgpu_drm_ioctl(struct file *filp,
   	struct drm_file *file_priv = filp->private_data;
   	struct drm_device *dev;
   	long ret;
+
   	dev = file_priv->minor->dev;
   	ret = pm_runtime_get_sync(dev->dev);
   	if (ret < 0)
   		return ret;
 
+	while (amdgpu_device_is_reset(dev->dev_private))
+		msleep(100);
+
   	ret = drm_ioctl(filp, cmd, arg);
 
   	pm_runtime_mark_last_busy(dev->dev);
drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c +6 -1
index 2648291..22b8059 100644
@@ -36,10 +36,15 @@
   long amdgpu_kms_compat_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
   {
   	unsigned int nr = DRM_IOCTL_NR(cmd);
+	struct drm_file *file_priv = filp->private_data;
+	struct amdgpu_device *adev = file_priv->minor->dev->dev_private;
   	int ret;
 
-	if (nr < DRM_COMMAND_BASE)
+	if (nr < DRM_COMMAND_BASE) {
+		while (amdgpu_device_is_reset(adev))
+			msleep(100);
   		return drm_compat_ioctl(filp, cmd, arg);
+	}
 
Quoted 10 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c line 50

Quoted 2 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c +0 -0
   	ret = amdgpu_drm_ioctl(filp, cmd, arg);
 

_______________________________________________
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
In reply to a message not in the archive
Quoted 13 lines

-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Liu, Monk
Sent: Thursday, May 04, 2017 6:42 AM
To: Zhou, David(ChunMing); He, Hongbo; [email protected]
Subject: RE: [PATCH 3/3] drm/amdgpu: fix gpu reset issue

First of all, Your patch will impact on SR-IOV case, because "gpu_reset_state" will be referenced by two place:
1) IOCTL; 2)GPU-RESET ,
SRIOV use srio_gpu_reset, but SRIOV use the same IOCTL routines, that way the gpu_reset_state is not consistent for SR-IOV case

Quoted 6 lines

I suggest you firstly unify this gpu_reset_state for both SRIOV and bare- metal case

For sriov, we have "adev->gfx.in_reset " and you now have "adev-

Quoted 1 line

gpu_reset_state",

I suggest you replace all "adev->gfx.in_reset" with "adev->gpu_reset_state",

Yes, I agree. Please unify these.

Alex

Quoted 151 lines

BR Monk

-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of zhoucm1
Sent: Friday, April 21, 2017 3:18 PM
To: He, Hongbo <[email protected]>; [email protected]
Subject: Re: [PATCH 3/3] drm/amdgpu: fix gpu reset issue

Quoted 2 lines · Roger.He
Change-Id Ib77d33a09f348ebf2e3a9d7861411f4b951ebf7c

Please add more explaination in comment, like "prevent ioctl during gpu reset, otherwise many un-expected behaviour happen."

Regards,
David Zhou

Quoted 3 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu.h        |  6 ++++++
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 16

+++++++++++++++-

drivers/gpu/drm/amd/amdgpu/amdgpu.h amdgpu_get_sdma_instance(struct

Quoted 18 lines
  drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c    |  4 ++++
  drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c  |  7 ++++++-
  4 files changed, 31 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +1 -0
index 71364f5..ab0ffa8 100644
@@ -1526,6 +1526,7 @@ struct amdgpu_device {
  	atomic64_t			num_bytes_moved;
  	atomic64_t			num_evictions;
  	atomic_t			gpu_reset_counter;
+	atomic_t			gpu_reset_state;
 
  	/* data for buffer migration throttling */
  	struct {
@@ -1851,6 +1852,11 @@ amdgpu_get_sdma_instance(struct

amdgpu_ring *ring)

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c int amdgpu_device_init(struct amdgpu_device

Quoted 18 lines
drivers/gpu/drm/amd/amdgpu/amdgpu.h +5 -0
  #define amdgpu_psp_check_fw_loading_status(adev, i)
 (adev)->firmware.funcs->check_fw_loading_status((adev), (i))
 
  /* Common functions */
+static inline int amdgpu_device_is_reset(struct amdgpu_device *adev)
+{
+	return atomic_read(&adev->gpu_reset_state);
+}
+
  int amdgpu_gpu_reset(struct amdgpu_device *adev);
  bool amdgpu_need_backup(struct amdgpu_device *adev);

void amdgpu_pci_config_reset(struct amdgpu_device *adev); diff --git
a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index f882496..0fb4716 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
@@ -1894,6 +1894,7 @@ int amdgpu_device_init(struct amdgpu_device

*adev,

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c int amdgpu_sriov_gpu_reset(struct

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1 -0
  	mutex_init(&adev->grbm_idx_mutex);
  	mutex_init(&adev->mn_lock);
  	hash_init(adev->mn_hash);
+	atomic_set(&adev->gpu_reset_state, 0);
 
  	amdgpu_check_arguments(adev);
 
@@ -2655,6 +2656,18 @@ int amdgpu_sriov_gpu_reset(struct

amdgpu_device *adev, bool voluntary)

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c int amdgpu_sriov_gpu_reset(struct amdgpu_device *adev, bool voluntary)

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +5 -0
  }
 
  /**
+ * amdgpu_device_set_reset_state - set gpu reset state
+ *
+ * @adev: amdgpu device pointer
+ * @state: true when start to reset gpu; false: reset done  */ static
+inline void amdgpu_device_set_reset_state(struct amdgpu_device

*adev,

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c int amdgpu_gpu_reset(struct amdgpu_device

Quoted 9 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +5 -0
+							bool state)
+{
+	atomic_set(&adev->gpu_reset_state, state); }
+
+/**
   * amdgpu_gpu_reset - reset the asic
   *
   * @adev: amdgpu device pointer
@@ -2678,7 +2691,7 @@ int amdgpu_gpu_reset(struct amdgpu_device

*adev)

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c int amdgpu_gpu_reset(struct amdgpu_device

Quoted 9 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1 -1
  	}
 
  	atomic_inc(&adev->gpu_reset_counter);
-
+	amdgpu_device_set_reset_state(adev, true);
  	/* block TTM */
  	resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
  	/* store modesetting */
@@ -2811,6 +2824,7 @@ int amdgpu_gpu_reset(struct amdgpu_device

*adev)

Quoted 34 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1 -0
  		dev_info(adev->dev, "GPU reset failed\n");
  	}
 
+	amdgpu_device_set_reset_state(adev, false);
  	return r;
  }
 
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +4 -0
index ead00d7..8cc14af 100644
@@ -679,11 +679,15 @@ long amdgpu_drm_ioctl(struct file *filp,
  	struct drm_file *file_priv = filp->private_data;
  	struct drm_device *dev;
  	long ret;
+
  	dev = file_priv->minor->dev;
  	ret = pm_runtime_get_sync(dev->dev);
  	if (ret < 0)
  		return ret;
 
+	while (amdgpu_device_is_reset(dev->dev_private))
+		msleep(100);
+
  	ret = drm_ioctl(filp, cmd, arg);
 
  	pm_runtime_mark_last_busy(dev->dev);
drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c +0 -0
index 2648291..22b8059 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c line 36

Quoted 1 line
drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c +0 -0
@@ -36,10 +36,15 @@
  long amdgpu_kms_compat_ioctl(struct file *filp, unsigned int cmd,

unsigned long arg)

Quoted 13 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c +6 -1
  {
  	unsigned int nr = DRM_IOCTL_NR(cmd);
+	struct drm_file *file_priv = filp->private_data;
+	struct amdgpu_device *adev = file_priv->minor->dev->dev_private;
  	int ret;
 
-	if (nr < DRM_COMMAND_BASE)
+	if (nr < DRM_COMMAND_BASE) {
+		while (amdgpu_device_is_reset(adev))
+			msleep(100);
  		return drm_compat_ioctl(filp, cmd, arg);
+	}
 
Quoted 11 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c line 50

Quoted 2 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c +0 -0
  	ret = amdgpu_drm_ioctl(filp, cmd, arg);
 

_______________________________________________
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

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

NAK, that is exactly what we wanted to avoid.

We used to have an exclusive lock for this and it cause a whole bunch of problems.

Please elaborate why that should be necessary.

Regards,
Christian.

Am 21.04.2017 um 09:08 schrieb Roger.He:

Quoted 112 lines
Change-Id Ib77d33a09f348ebf2e3a9d7861411f4b951ebf7c
---
  drivers/gpu/drm/amd/amdgpu/amdgpu.h        |  6 ++++++
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 16 +++++++++++++++-
  drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c    |  4 ++++
  drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c  |  7 ++++++-
  4 files changed, 31 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +6 -0
index 71364f5..ab0ffa8 100644
@@ -1526,6 +1526,7 @@ struct amdgpu_device {
  	atomic64_t			num_bytes_moved;
  	atomic64_t			num_evictions;
  	atomic_t			gpu_reset_counter;
+	atomic_t			gpu_reset_state;
 
  	/* data for buffer migration throttling */
  	struct {
@@ -1851,6 +1852,11 @@ amdgpu_get_sdma_instance(struct amdgpu_ring *ring)
  #define amdgpu_psp_check_fw_loading_status(adev, i) (adev)->firmware.funcs->check_fw_loading_status((adev), (i))
 
  /* Common functions */
+static inline int amdgpu_device_is_reset(struct amdgpu_device *adev)
+{
+	return atomic_read(&adev->gpu_reset_state);
+}
+
  int amdgpu_gpu_reset(struct amdgpu_device *adev);
  bool amdgpu_need_backup(struct amdgpu_device *adev);
  void amdgpu_pci_config_reset(struct amdgpu_device *adev);
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +15 -1
index f882496..0fb4716 100644
@@ -1894,6 +1894,7 @@ int amdgpu_device_init(struct amdgpu_device *adev,
  	mutex_init(&adev->grbm_idx_mutex);
  	mutex_init(&adev->mn_lock);
  	hash_init(adev->mn_hash);
+	atomic_set(&adev->gpu_reset_state, 0);
 
  	amdgpu_check_arguments(adev);
 
@@ -2655,6 +2656,18 @@ int amdgpu_sriov_gpu_reset(struct amdgpu_device *adev, bool voluntary)
  }
 
  /**
+ * amdgpu_device_set_reset_state - set gpu reset state
+ *
+ * @adev: amdgpu device pointer
+ * @state: true when start to reset gpu; false: reset done
+ */
+static inline void amdgpu_device_set_reset_state(struct amdgpu_device *adev,
+							bool state)
+{
+	atomic_set(&adev->gpu_reset_state, state);
+}
+
+/**
   * amdgpu_gpu_reset - reset the asic
   *
   * @adev: amdgpu device pointer
@@ -2678,7 +2691,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
  	}
 
  	atomic_inc(&adev->gpu_reset_counter);
-
+	amdgpu_device_set_reset_state(adev, true);
  	/* block TTM */
  	resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
  	/* store modesetting */
@@ -2811,6 +2824,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
  		dev_info(adev->dev, "GPU reset failed\n");
  	}
 
+	amdgpu_device_set_reset_state(adev, false);
  	return r;
  }
 
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +4 -0
index ead00d7..8cc14af 100644
@@ -679,11 +679,15 @@ long amdgpu_drm_ioctl(struct file *filp,
  	struct drm_file *file_priv = filp->private_data;
  	struct drm_device *dev;
  	long ret;
+
  	dev = file_priv->minor->dev;
  	ret = pm_runtime_get_sync(dev->dev);
  	if (ret < 0)
  		return ret;
 
+	while (amdgpu_device_is_reset(dev->dev_private))
+		msleep(100);
+
  	ret = drm_ioctl(filp, cmd, arg);
 
  	pm_runtime_mark_last_busy(dev->dev);
drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c +2 -0
index 2648291..22b8059 100644
@@ -36,10 +36,15 @@
  long amdgpu_kms_compat_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
  {
  	unsigned int nr = DRM_IOCTL_NR(cmd);
+	struct drm_file *file_priv = filp->private_data;
+	struct amdgpu_device *adev = file_priv->minor->dev->dev_private;
  	int ret;

drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c line 50

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c +4 -1
 
-	if (nr < DRM_COMMAND_BASE)
+	if (nr < DRM_COMMAND_BASE) {
+		while (amdgpu_device_is_reset(adev))
+			msleep(100);
  		return drm_compat_ioctl(filp, cmd, arg);
+	}
 
  	ret = amdgpu_drm_ioctl(filp, cmd, arg);
 

_______________________________________________
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): ef2c40cd-bb55-5921-9318-e0147f1b4719-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

#12 lore raw

Hi Roger,

Could you try to move 'id->current_gpu_reset_count = atomic_read(&adev->gpu_reset_counter);' in amdgpu_vm_grab_id to no_flush_needed label.

Regards,
David Zhou

Quoted 117 lines · Christian König

NAK, that is exactly what we wanted to avoid.

We used to have an exclusive lock for this and it cause a whole bunch of problems.

Please elaborate why that should be necessary.

Regards,
Christian.

Am 21.04.2017 um 09:08 schrieb Roger.He:

Quoted 107 lines
Change-Id Ib77d33a09f348ebf2e3a9d7861411f4b951ebf7c
---
  drivers/gpu/drm/amd/amdgpu/amdgpu.h        |  6 ++++++
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 16 +++++++++++++++-
  drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c    |  4 ++++
  drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c  |  7 ++++++-
  4 files changed, 31 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +6 -0
index 71364f5..ab0ffa8 100644
@@ -1526,6 +1526,7 @@ struct amdgpu_device {
      atomic64_t            num_bytes_moved;
      atomic64_t            num_evictions;
      atomic_t            gpu_reset_counter;
+    atomic_t            gpu_reset_state;
        /* data for buffer migration throttling */
      struct {
@@ -1851,6 +1852,11 @@ amdgpu_get_sdma_instance(struct amdgpu_ring 
 *ring)
  #define amdgpu_psp_check_fw_loading_status(adev, i) 
 (adev)->firmware.funcs->check_fw_loading_status((adev), (i))
    /* Common functions */
+static inline int amdgpu_device_is_reset(struct amdgpu_device *adev)
+{
+    return atomic_read(&adev->gpu_reset_state);
+}
+
  int amdgpu_gpu_reset(struct amdgpu_device *adev);
  bool amdgpu_need_backup(struct amdgpu_device *adev);
  void amdgpu_pci_config_reset(struct amdgpu_device *adev);
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1 -0
index f882496..0fb4716 100644
@@ -1894,6 +1894,7 @@ int amdgpu_device_init(struct amdgpu_device *adev,
      mutex_init(&adev->grbm_idx_mutex);
      mutex_init(&adev->mn_lock);
      hash_init(adev->mn_hash);
+    atomic_set(&adev->gpu_reset_state, 0);
        amdgpu_check_arguments(adev);
  @@ -2655,6 +2656,18 @@ int amdgpu_sriov_gpu_reset(struct 
 amdgpu_device *adev, bool voluntary)

}

    /**

+ * amdgpu_device_set_reset_state - set gpu reset state
+ *
+ * @adev: amdgpu device pointer
+ * @state: true when start to reset gpu; false: reset done
+ */
+static inline void amdgpu_device_set_reset_state(struct amdgpu_device *adev,

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +7 -1
+                            bool state)
+{
+    atomic_set(&adev->gpu_reset_state, state);
+}
+
+/**
   * amdgpu_gpu_reset - reset the asic
   *
   * @adev: amdgpu device pointer
@@ -2678,7 +2691,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
      }
        atomic_inc(&adev->gpu_reset_counter);
-
+    amdgpu_device_set_reset_state(adev, true);
      /* block TTM */
      resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
      /* store modesetting */
@@ -2811,6 +2824,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
          dev_info(adev->dev, "GPU reset failed\n");
      }
  +    amdgpu_device_set_reset_state(adev, false);
      return r;
  }
  diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c 

b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
index ead00d7..8cc14af 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +3 -0
@@ -679,11 +679,15 @@ long amdgpu_drm_ioctl(struct file *filp,
      struct drm_file *file_priv = filp->private_data;
      struct drm_device *dev;
      long ret;
+
      dev = file_priv->minor->dev;
      ret = pm_runtime_get_sync(dev->dev);
      if (ret < 0)
          return ret;
  +    while (amdgpu_device_is_reset(dev->dev_private))
+        msleep(100);
+
      ret = drm_ioctl(filp, cmd, arg);
        pm_runtime_mark_last_busy(dev->dev);
drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c +2 -0
index 2648291..22b8059 100644
@@ -36,10 +36,15 @@
  long amdgpu_kms_compat_ioctl(struct file *filp, unsigned int cmd, 
 unsigned long arg)
  {
      unsigned int nr = DRM_IOCTL_NR(cmd);
+    struct drm_file *file_priv = filp->private_data;
+    struct amdgpu_device *adev = file_priv->minor->dev->dev_private;
Quoted 10 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c line 49

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c +4 -0
      int ret;
  -    if (nr < DRM_COMMAND_BASE)
+    if (nr < DRM_COMMAND_BASE) {
+        while (amdgpu_device_is_reset(adev))
+            msleep(100);
          return drm_compat_ioctl(filp, cmd, arg);
+    }
        ret = amdgpu_drm_ioctl(filp, cmd, arg);

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

author NAK
#13 lore raw
In reply to a message not in the archive

Hi Christian:

During GPU reset, IOCTL still submit command to GPU.
That will result in lots of strange problems when we test this feature.

I have no better idea to handle this case. At that time, I have tried with return error directly rather than waiting GPU reset completion. But that doesn't work and it lead to other problems.

Thanks
Roger(Hongbo.He)

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: Friday, April 21, 2017 4:27 PM
To: He, Hongbo; [email protected]
Cc: Zhou, David(ChunMing)
Subject: Re: [PATCH 3/3] drm/amdgpu: fix gpu reset issue

NAK, that is exactly what we wanted to avoid.

We used to have an exclusive lock for this and it cause a whole bunch of problems.

Please elaborate why that should be necessary.

Regards,
Christian.

Am 21.04.2017 um 09:08 schrieb Roger.He:

Quoted 115 lines
Change-Id Ib77d33a09f348ebf2e3a9d7861411f4b951ebf7c
---
  drivers/gpu/drm/amd/amdgpu/amdgpu.h        |  6 ++++++
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 16 +++++++++++++++-
  drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c    |  4 ++++
  drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c  |  7 ++++++-
  4 files changed, 31 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +6 -0
index 71364f5..ab0ffa8 100644
@@ -1526,6 +1526,7 @@ struct amdgpu_device {
  	atomic64_t			num_bytes_moved;
  	atomic64_t			num_evictions;
  	atomic_t			gpu_reset_counter;
+	atomic_t			gpu_reset_state;
 
  	/* data for buffer migration throttling */
  	struct {
@@ -1851,6 +1852,11 @@ amdgpu_get_sdma_instance(struct amdgpu_ring *ring)
  #define amdgpu_psp_check_fw_loading_status(adev, i) 
 (adev)->firmware.funcs->check_fw_loading_status((adev), (i))
 
  /* Common functions */
+static inline int amdgpu_device_is_reset(struct amdgpu_device *adev) 
+{
+	return atomic_read(&adev->gpu_reset_state);
+}
+
  int amdgpu_gpu_reset(struct amdgpu_device *adev);
  bool amdgpu_need_backup(struct amdgpu_device *adev);

void amdgpu_pci_config_reset(struct amdgpu_device *adev); diff --git
a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index f882496..0fb4716 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +13 -1
@@ -1894,6 +1894,7 @@ int amdgpu_device_init(struct amdgpu_device *adev,
  	mutex_init(&adev->grbm_idx_mutex);
  	mutex_init(&adev->mn_lock);
  	hash_init(adev->mn_hash);
+	atomic_set(&adev->gpu_reset_state, 0);
 
  	amdgpu_check_arguments(adev);
 
@@ -2655,6 +2656,18 @@ int amdgpu_sriov_gpu_reset(struct amdgpu_device *adev, bool voluntary)
  }
 
  /**
+ * amdgpu_device_set_reset_state - set gpu reset state
+ *
+ * @adev: amdgpu device pointer
+ * @state: true when start to reset gpu; false: reset done  */ static 
+inline void amdgpu_device_set_reset_state(struct amdgpu_device *adev,
+							bool state)
+{
+	atomic_set(&adev->gpu_reset_state, state); }
+
+/**
   * amdgpu_gpu_reset - reset the asic
   *
   * @adev: amdgpu device pointer
@@ -2678,7 +2691,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
  	}
 
  	atomic_inc(&adev->gpu_reset_counter);
-
+	amdgpu_device_set_reset_state(adev, true);
  	/* block TTM */
  	resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
  	/* store modesetting */
@@ -2811,6 +2824,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
  		dev_info(adev->dev, "GPU reset failed\n");
  	}
 
+	amdgpu_device_set_reset_state(adev, false);
  	return r;
  }
 
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +4 -0
index ead00d7..8cc14af 100644
@@ -679,11 +679,15 @@ long amdgpu_drm_ioctl(struct file *filp,
  	struct drm_file *file_priv = filp->private_data;
  	struct drm_device *dev;
  	long ret;
+
  	dev = file_priv->minor->dev;
  	ret = pm_runtime_get_sync(dev->dev);
  	if (ret < 0)
  		return ret;
 
+	while (amdgpu_device_is_reset(dev->dev_private))
+		msleep(100);
+
  	ret = drm_ioctl(filp, cmd, arg);
 
  	pm_runtime_mark_last_busy(dev->dev);
drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c +2 -0
index 2648291..22b8059 100644
@@ -36,10 +36,15 @@
  long amdgpu_kms_compat_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
  {
  	unsigned int nr = DRM_IOCTL_NR(cmd);
+	struct drm_file *file_priv = filp->private_data;
+	struct amdgpu_device *adev = file_priv->minor->dev->dev_private;
  	int ret;

drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c line 50

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c +4 -1
 
-	if (nr < DRM_COMMAND_BASE)
+	if (nr < DRM_COMMAND_BASE) {
+		while (amdgpu_device_is_reset(adev))
+			msleep(100);
  		return drm_compat_ioctl(filp, cmd, arg);
+	}
 
  	ret = amdgpu_drm_ioctl(filp, cmd, arg);
 

_______________________________________________
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): MWHPR1201MB01272142B19FFEB9EAD281F3FD1A0-3iK1xFAIwjq9imrIu4W8xGrFom/aUZj6nBOFsp37pqbUKgpGm//BTAC/[email protected]

#15 lore raw
Quoted 1 line

During GPU reset, IOCTL still submit command to GPU.

No it won't. We disable the GPU scheduler during GPU reset, so that isn't an issue.

Do you have a concrete problem (e.g. dmesg?) you ran into?

Regards,
Christian.

Am 21.04.2017 um 11:00 schrieb He, Hongbo:

Quoted 151 lines

Hi Christian:

During GPU reset, IOCTL still submit command to GPU.
That will result in lots of strange problems when we test this feature.

I have no better idea to handle this case. At that time, I have tried with return error directly rather than waiting GPU reset completion. But that doesn't work and it lead to other problems.

Thanks
Roger(Hongbo.He)

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: Friday, April 21, 2017 4:27 PM
To: He, Hongbo; [email protected]
Cc: Zhou, David(ChunMing)
Subject: Re: [PATCH 3/3] drm/amdgpu: fix gpu reset issue

NAK, that is exactly what we wanted to avoid.

We used to have an exclusive lock for this and it cause a whole bunch of problems.

Please elaborate why that should be necessary.

Regards,
Christian.

Am 21.04.2017 um 09:08 schrieb Roger.He:

Quoted 123 lines
Change-Id Ib77d33a09f348ebf2e3a9d7861411f4b951ebf7c
---
   drivers/gpu/drm/amd/amdgpu/amdgpu.h        |  6 ++++++
   drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 16 +++++++++++++++-
   drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c    |  4 ++++
   drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c  |  7 ++++++-
   4 files changed, 31 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +6 -0
index 71364f5..ab0ffa8 100644
@@ -1526,6 +1526,7 @@ struct amdgpu_device {
   	atomic64_t			num_bytes_moved;
   	atomic64_t			num_evictions;
   	atomic_t			gpu_reset_counter;
+	atomic_t			gpu_reset_state;
 
   	/* data for buffer migration throttling */
   	struct {
@@ -1851,6 +1852,11 @@ amdgpu_get_sdma_instance(struct amdgpu_ring *ring)
   #define amdgpu_psp_check_fw_loading_status(adev, i)
 (adev)->firmware.funcs->check_fw_loading_status((adev), (i))
 
   /* Common functions */
+static inline int amdgpu_device_is_reset(struct amdgpu_device *adev)
+{
+	return atomic_read(&adev->gpu_reset_state);
+}
+
   int amdgpu_gpu_reset(struct amdgpu_device *adev);
   bool amdgpu_need_backup(struct amdgpu_device *adev);
   void amdgpu_pci_config_reset(struct amdgpu_device *adev); diff --git

a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index f882496..0fb4716 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +13 -1
@@ -1894,6 +1894,7 @@ int amdgpu_device_init(struct amdgpu_device *adev,
   	mutex_init(&adev->grbm_idx_mutex);
   	mutex_init(&adev->mn_lock);
   	hash_init(adev->mn_hash);
+	atomic_set(&adev->gpu_reset_state, 0);
 
   	amdgpu_check_arguments(adev);
 
@@ -2655,6 +2656,18 @@ int amdgpu_sriov_gpu_reset(struct amdgpu_device *adev, bool voluntary)
   }
 
   /**
+ * amdgpu_device_set_reset_state - set gpu reset state
+ *
+ * @adev: amdgpu device pointer
+ * @state: true when start to reset gpu; false: reset done  */ static
+inline void amdgpu_device_set_reset_state(struct amdgpu_device *adev,
+							bool state)
+{
+	atomic_set(&adev->gpu_reset_state, state); }
+
+/**
    * amdgpu_gpu_reset - reset the asic
    *
    * @adev: amdgpu device pointer
@@ -2678,7 +2691,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
   	}
 
   	atomic_inc(&adev->gpu_reset_counter);
-
+	amdgpu_device_set_reset_state(adev, true);
   	/* block TTM */
   	resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
   	/* store modesetting */
@@ -2811,6 +2824,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
   		dev_info(adev->dev, "GPU reset failed\n");
   	}
 
+	amdgpu_device_set_reset_state(adev, false);
   	return r;
   }
 
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +4 -0
index ead00d7..8cc14af 100644
@@ -679,11 +679,15 @@ long amdgpu_drm_ioctl(struct file *filp,
   	struct drm_file *file_priv = filp->private_data;
   	struct drm_device *dev;
   	long ret;
+
   	dev = file_priv->minor->dev;
   	ret = pm_runtime_get_sync(dev->dev);
   	if (ret < 0)
   		return ret;
 
+	while (amdgpu_device_is_reset(dev->dev_private))
+		msleep(100);
+
   	ret = drm_ioctl(filp, cmd, arg);
 
   	pm_runtime_mark_last_busy(dev->dev);
drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c +6 -1
index 2648291..22b8059 100644
@@ -36,10 +36,15 @@
   long amdgpu_kms_compat_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
   {
   	unsigned int nr = DRM_IOCTL_NR(cmd);
+	struct drm_file *file_priv = filp->private_data;
+	struct amdgpu_device *adev = file_priv->minor->dev->dev_private;
   	int ret;
 
-	if (nr < DRM_COMMAND_BASE)
+	if (nr < DRM_COMMAND_BASE) {
+		while (amdgpu_device_is_reset(adev))
+			msleep(100);
   		return drm_compat_ioctl(filp, cmd, arg);
+	}
 
Quoted 3 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c line 50

Quoted 2 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c +0 -0
   	ret = amdgpu_drm_ioctl(filp, cmd, arg);
 

_______________________________________________
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): 05b9d783-75ab-9924-6881-472540fc77d1-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

author
#17 lore raw
Quoted 1 line

During GPU reset, IOCTL still submit command to GPU.

No it won't. We disable the GPU scheduler during GPU reset, so that isn't an issue.
Could you point out the related code? Thanks.

Thanks
Roger(Hongbo.He)

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: Friday, April 21, 2017 5:16 PM
To: He, Hongbo; [email protected]
Cc: Zhou, David(ChunMing)
Subject: Re: [PATCH 3/3] drm/amdgpu: fix gpu reset issue

Quoted 1 line

During GPU reset, IOCTL still submit command to GPU.

No it won't. We disable the GPU scheduler during GPU reset, so that isn't an issue.

Do you have a concrete problem (e.g. dmesg?) you ran into?

Regards,
Christian.

Am 21.04.2017 um 11:00 schrieb He, Hongbo:

Quoted 151 lines

Hi Christian:

During GPU reset, IOCTL still submit command to GPU.
That will result in lots of strange problems when we test this feature.

I have no better idea to handle this case. At that time, I have tried with return error directly rather than waiting GPU reset completion. But that doesn't work and it lead to other problems.

Thanks
Roger(Hongbo.He)

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: Friday, April 21, 2017 4:27 PM
To: He, Hongbo; [email protected]
Cc: Zhou, David(ChunMing)
Subject: Re: [PATCH 3/3] drm/amdgpu: fix gpu reset issue

NAK, that is exactly what we wanted to avoid.

We used to have an exclusive lock for this and it cause a whole bunch of problems.

Please elaborate why that should be necessary.

Regards,
Christian.

Am 21.04.2017 um 09:08 schrieb Roger.He:

Quoted 123 lines
Change-Id Ib77d33a09f348ebf2e3a9d7861411f4b951ebf7c
---
   drivers/gpu/drm/amd/amdgpu/amdgpu.h        |  6 ++++++
   drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 16 +++++++++++++++-
   drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c    |  4 ++++
   drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c  |  7 ++++++-
   4 files changed, 31 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +6 -0
index 71364f5..ab0ffa8 100644
@@ -1526,6 +1526,7 @@ struct amdgpu_device {
   	atomic64_t			num_bytes_moved;
   	atomic64_t			num_evictions;
   	atomic_t			gpu_reset_counter;
+	atomic_t			gpu_reset_state;
 
   	/* data for buffer migration throttling */
   	struct {
@@ -1851,6 +1852,11 @@ amdgpu_get_sdma_instance(struct amdgpu_ring *ring)
   #define amdgpu_psp_check_fw_loading_status(adev, i) 
 (adev)->firmware.funcs->check_fw_loading_status((adev), (i))
 
   /* Common functions */
+static inline int amdgpu_device_is_reset(struct amdgpu_device *adev) 
+{
+	return atomic_read(&adev->gpu_reset_state);
+}
+
   int amdgpu_gpu_reset(struct amdgpu_device *adev);
   bool amdgpu_need_backup(struct amdgpu_device *adev);

void amdgpu_pci_config_reset(struct amdgpu_device *adev); diff
--git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index f882496..0fb4716 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +13 -1
@@ -1894,6 +1894,7 @@ int amdgpu_device_init(struct amdgpu_device *adev,
   	mutex_init(&adev->grbm_idx_mutex);
   	mutex_init(&adev->mn_lock);
   	hash_init(adev->mn_hash);
+	atomic_set(&adev->gpu_reset_state, 0);
 
   	amdgpu_check_arguments(adev);
 
@@ -2655,6 +2656,18 @@ int amdgpu_sriov_gpu_reset(struct amdgpu_device *adev, bool voluntary)
   }
 
   /**
+ * amdgpu_device_set_reset_state - set gpu reset state
+ *
+ * @adev: amdgpu device pointer
+ * @state: true when start to reset gpu; false: reset done  */ 
+static inline void amdgpu_device_set_reset_state(struct amdgpu_device *adev,
+							bool state)
+{
+	atomic_set(&adev->gpu_reset_state, state); }
+
+/**
    * amdgpu_gpu_reset - reset the asic
    *
    * @adev: amdgpu device pointer
@@ -2678,7 +2691,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
   	}
 
   	atomic_inc(&adev->gpu_reset_counter);
-
+	amdgpu_device_set_reset_state(adev, true);
   	/* block TTM */
   	resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
   	/* store modesetting */
@@ -2811,6 +2824,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
   		dev_info(adev->dev, "GPU reset failed\n");
   	}
 
+	amdgpu_device_set_reset_state(adev, false);
   	return r;
   }
 
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +4 -0
index ead00d7..8cc14af 100644
@@ -679,11 +679,15 @@ long amdgpu_drm_ioctl(struct file *filp,
   	struct drm_file *file_priv = filp->private_data;
   	struct drm_device *dev;
   	long ret;
+
   	dev = file_priv->minor->dev;
   	ret = pm_runtime_get_sync(dev->dev);
   	if (ret < 0)
   		return ret;
 
+	while (amdgpu_device_is_reset(dev->dev_private))
+		msleep(100);
+
   	ret = drm_ioctl(filp, cmd, arg);
 
   	pm_runtime_mark_last_busy(dev->dev);
drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c +6 -1
index 2648291..22b8059 100644
@@ -36,10 +36,15 @@
   long amdgpu_kms_compat_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
   {
   	unsigned int nr = DRM_IOCTL_NR(cmd);
+	struct drm_file *file_priv = filp->private_data;
+	struct amdgpu_device *adev = file_priv->minor->dev->dev_private;
   	int ret;
 
-	if (nr < DRM_COMMAND_BASE)
+	if (nr < DRM_COMMAND_BASE) {
+		while (amdgpu_device_is_reset(adev))
+			msleep(100);
   		return drm_compat_ioctl(filp, cmd, arg);
+	}
 
Quoted 3 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c line 50

Quoted 2 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c +0 -0
   	ret = amdgpu_drm_ioctl(filp, cmd, arg);
 

_______________________________________________
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): MWHPR1201MB0127A593761702A7A179855EFD1A0-3iK1xFAIwjq9imrIu4W8xGrFom/aUZj6nBOFsp37pqbUKgpGm//BTAC/[email protected]

#19 lore raw

See amdgpu_gpu_reset():

Quoted 9 lines
        /* block scheduler */
        for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
                struct amdgpu_ring *ring = adev->rings[i];

                if (!ring)
                        continue;
                kthread_park(ring->sched.thread);
                amd_sched_hw_job_reset(&ring->sched);
        }

and

Quoted 8 lines
                for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
                        struct amdgpu_ring *ring = adev->rings[i];
                        if (!ring)
                                continue;
amd_sched_job_recovery(&ring->sched);
                        kthread_unpark(ring->sched.thread);
                }

Regards,
Christian.

Am 21.04.2017 um 11:54 schrieb He, Hongbo:

Quoted 175 lines
Quoted 1 line

During GPU reset, IOCTL still submit command to GPU.

No it won't. We disable the GPU scheduler during GPU reset, so that isn't an issue.
Could you point out the related code? Thanks.

Thanks
Roger(Hongbo.He)

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: Friday, April 21, 2017 5:16 PM
To: He, Hongbo; [email protected]
Cc: Zhou, David(ChunMing)
Subject: Re: [PATCH 3/3] drm/amdgpu: fix gpu reset issue

Quoted 1 line

During GPU reset, IOCTL still submit command to GPU.

No it won't. We disable the GPU scheduler during GPU reset, so that isn't an issue.

Do you have a concrete problem (e.g. dmesg?) you ran into?

Regards,
Christian.

Am 21.04.2017 um 11:00 schrieb He, Hongbo:

Quoted 151 lines

Hi Christian:

During GPU reset, IOCTL still submit command to GPU.
That will result in lots of strange problems when we test this feature.

I have no better idea to handle this case. At that time, I have tried with return error directly rather than waiting GPU reset completion. But that doesn't work and it lead to other problems.

Thanks
Roger(Hongbo.He)

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: Friday, April 21, 2017 4:27 PM
To: He, Hongbo; [email protected]
Cc: Zhou, David(ChunMing)
Subject: Re: [PATCH 3/3] drm/amdgpu: fix gpu reset issue

NAK, that is exactly what we wanted to avoid.

We used to have an exclusive lock for this and it cause a whole bunch of problems.

Please elaborate why that should be necessary.

Regards,
Christian.

Am 21.04.2017 um 09:08 schrieb Roger.He:

Quoted 123 lines
Change-Id Ib77d33a09f348ebf2e3a9d7861411f4b951ebf7c
---
    drivers/gpu/drm/amd/amdgpu/amdgpu.h        |  6 ++++++
    drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 16 +++++++++++++++-
    drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c    |  4 ++++
    drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c  |  7 ++++++-
    4 files changed, 31 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +6 -0
index 71364f5..ab0ffa8 100644
@@ -1526,6 +1526,7 @@ struct amdgpu_device {
    	atomic64_t			num_bytes_moved;
    	atomic64_t			num_evictions;
    	atomic_t			gpu_reset_counter;
+	atomic_t			gpu_reset_state;
 
    	/* data for buffer migration throttling */
    	struct {
@@ -1851,6 +1852,11 @@ amdgpu_get_sdma_instance(struct amdgpu_ring *ring)
    #define amdgpu_psp_check_fw_loading_status(adev, i)
 (adev)->firmware.funcs->check_fw_loading_status((adev), (i))
 
    /* Common functions */
+static inline int amdgpu_device_is_reset(struct amdgpu_device *adev)
+{
+	return atomic_read(&adev->gpu_reset_state);
+}
+
    int amdgpu_gpu_reset(struct amdgpu_device *adev);
    bool amdgpu_need_backup(struct amdgpu_device *adev);

void amdgpu_pci_config_reset(struct amdgpu_device *adev); diff
--git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index f882496..0fb4716 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +13 -1
@@ -1894,6 +1894,7 @@ int amdgpu_device_init(struct amdgpu_device *adev,
    	mutex_init(&adev->grbm_idx_mutex);
    	mutex_init(&adev->mn_lock);
    	hash_init(adev->mn_hash);
+	atomic_set(&adev->gpu_reset_state, 0);
 
    	amdgpu_check_arguments(adev);
 
@@ -2655,6 +2656,18 @@ int amdgpu_sriov_gpu_reset(struct amdgpu_device *adev, bool voluntary)
    }
 
    /**
+ * amdgpu_device_set_reset_state - set gpu reset state
+ *
+ * @adev: amdgpu device pointer
+ * @state: true when start to reset gpu; false: reset done  */
+static inline void amdgpu_device_set_reset_state(struct amdgpu_device *adev,
+							bool state)
+{
+	atomic_set(&adev->gpu_reset_state, state); }
+
+/**
     * amdgpu_gpu_reset - reset the asic
     *
     * @adev: amdgpu device pointer
@@ -2678,7 +2691,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
    	}
 
    	atomic_inc(&adev->gpu_reset_counter);
-
+	amdgpu_device_set_reset_state(adev, true);
    	/* block TTM */
    	resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
    	/* store modesetting */
@@ -2811,6 +2824,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
    		dev_info(adev->dev, "GPU reset failed\n");
    	}
 
+	amdgpu_device_set_reset_state(adev, false);
    	return r;
    }
 
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +4 -0
index ead00d7..8cc14af 100644
@@ -679,11 +679,15 @@ long amdgpu_drm_ioctl(struct file *filp,
    	struct drm_file *file_priv = filp->private_data;
    	struct drm_device *dev;
    	long ret;
+
    	dev = file_priv->minor->dev;
    	ret = pm_runtime_get_sync(dev->dev);
    	if (ret < 0)
    		return ret;
 
+	while (amdgpu_device_is_reset(dev->dev_private))
+		msleep(100);
+
    	ret = drm_ioctl(filp, cmd, arg);
 
    	pm_runtime_mark_last_busy(dev->dev);
drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c +6 -1
index 2648291..22b8059 100644
@@ -36,10 +36,15 @@
    long amdgpu_kms_compat_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
    {
    	unsigned int nr = DRM_IOCTL_NR(cmd);
+	struct drm_file *file_priv = filp->private_data;
+	struct amdgpu_device *adev = file_priv->minor->dev->dev_private;
    	int ret;
 
-	if (nr < DRM_COMMAND_BASE)
+	if (nr < DRM_COMMAND_BASE) {
+		while (amdgpu_device_is_reset(adev))
+			msleep(100);
    		return drm_compat_ioctl(filp, cmd, arg);
+	}
 
Quoted 6 lines
Quoted 2 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c line 50

Quoted 2 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c +0 -0
    	ret = amdgpu_drm_ioctl(filp, cmd, arg);
 

_______________________________________________
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): eeb3931b-ef67-8645-a3bc-bc7f3db4002e-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

author
#21 lore raw

Thank you!

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: Friday, April 21, 2017 7:47 PM
To: He, Hongbo; [email protected]
Cc: Zhou, David(ChunMing)
Subject: Re: [PATCH 3/3] drm/amdgpu: fix gpu reset issue

See amdgpu_gpu_reset():

Quoted 9 lines
        /* block scheduler */
        for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
                struct amdgpu_ring *ring = adev->rings[i];

                if (!ring)
                        continue;
                kthread_park(ring->sched.thread);
                amd_sched_hw_job_reset(&ring->sched);
        }

and

Quoted 8 lines
                for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
                        struct amdgpu_ring *ring = adev->rings[i];
                        if (!ring)
                                continue;
amd_sched_job_recovery(&ring->sched);
                        kthread_unpark(ring->sched.thread);
                }

Regards,
Christian.

Am 21.04.2017 um 11:54 schrieb He, Hongbo:

Quoted 175 lines
Quoted 1 line

During GPU reset, IOCTL still submit command to GPU.

No it won't. We disable the GPU scheduler during GPU reset, so that isn't an issue.
Could you point out the related code? Thanks.

Thanks
Roger(Hongbo.He)

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: Friday, April 21, 2017 5:16 PM
To: He, Hongbo; [email protected]
Cc: Zhou, David(ChunMing)
Subject: Re: [PATCH 3/3] drm/amdgpu: fix gpu reset issue

Quoted 1 line

During GPU reset, IOCTL still submit command to GPU.

No it won't. We disable the GPU scheduler during GPU reset, so that isn't an issue.

Do you have a concrete problem (e.g. dmesg?) you ran into?

Regards,
Christian.

Am 21.04.2017 um 11:00 schrieb He, Hongbo:

Quoted 151 lines

Hi Christian:

During GPU reset, IOCTL still submit command to GPU.
That will result in lots of strange problems when we test this feature.

I have no better idea to handle this case. At that time, I have tried with return error directly rather than waiting GPU reset completion. But that doesn't work and it lead to other problems.

Thanks
Roger(Hongbo.He)

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: Friday, April 21, 2017 4:27 PM
To: He, Hongbo; [email protected]
Cc: Zhou, David(ChunMing)
Subject: Re: [PATCH 3/3] drm/amdgpu: fix gpu reset issue

NAK, that is exactly what we wanted to avoid.

We used to have an exclusive lock for this and it cause a whole bunch of problems.

Please elaborate why that should be necessary.

Regards,
Christian.

Am 21.04.2017 um 09:08 schrieb Roger.He:

Quoted 123 lines
Change-Id Ib77d33a09f348ebf2e3a9d7861411f4b951ebf7c
---
    drivers/gpu/drm/amd/amdgpu/amdgpu.h        |  6 ++++++
    drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 16 +++++++++++++++-
    drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c    |  4 ++++
    drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c  |  7 ++++++-
    4 files changed, 31 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +6 -0
index 71364f5..ab0ffa8 100644
@@ -1526,6 +1526,7 @@ struct amdgpu_device {
    	atomic64_t			num_bytes_moved;
    	atomic64_t			num_evictions;
    	atomic_t			gpu_reset_counter;
+	atomic_t			gpu_reset_state;
 
    	/* data for buffer migration throttling */
    	struct {
@@ -1851,6 +1852,11 @@ amdgpu_get_sdma_instance(struct amdgpu_ring *ring)
    #define amdgpu_psp_check_fw_loading_status(adev, i) 
 (adev)->firmware.funcs->check_fw_loading_status((adev), (i))
 
    /* Common functions */
+static inline int amdgpu_device_is_reset(struct amdgpu_device 
+*adev) {
+	return atomic_read(&adev->gpu_reset_state);
+}
+
    int amdgpu_gpu_reset(struct amdgpu_device *adev);
    bool amdgpu_need_backup(struct amdgpu_device *adev);

void amdgpu_pci_config_reset(struct amdgpu_device *adev); diff
--git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index f882496..0fb4716 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +13 -1
@@ -1894,6 +1894,7 @@ int amdgpu_device_init(struct amdgpu_device *adev,
    	mutex_init(&adev->grbm_idx_mutex);
    	mutex_init(&adev->mn_lock);
    	hash_init(adev->mn_hash);
+	atomic_set(&adev->gpu_reset_state, 0);
 
    	amdgpu_check_arguments(adev);
 
@@ -2655,6 +2656,18 @@ int amdgpu_sriov_gpu_reset(struct amdgpu_device *adev, bool voluntary)
    }
 
    /**
+ * amdgpu_device_set_reset_state - set gpu reset state
+ *
+ * @adev: amdgpu device pointer
+ * @state: true when start to reset gpu; false: reset done  */ 
+static inline void amdgpu_device_set_reset_state(struct amdgpu_device *adev,
+							bool state)
+{
+	atomic_set(&adev->gpu_reset_state, state); }
+
+/**
     * amdgpu_gpu_reset - reset the asic
     *
     * @adev: amdgpu device pointer
@@ -2678,7 +2691,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
    	}
 
    	atomic_inc(&adev->gpu_reset_counter);
-
+	amdgpu_device_set_reset_state(adev, true);
    	/* block TTM */
    	resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
    	/* store modesetting */
@@ -2811,6 +2824,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
    		dev_info(adev->dev, "GPU reset failed\n");
    	}
 
+	amdgpu_device_set_reset_state(adev, false);
    	return r;
    }
 
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +4 -0
index ead00d7..8cc14af 100644
@@ -679,11 +679,15 @@ long amdgpu_drm_ioctl(struct file *filp,
    	struct drm_file *file_priv = filp->private_data;
    	struct drm_device *dev;
    	long ret;
+
    	dev = file_priv->minor->dev;
    	ret = pm_runtime_get_sync(dev->dev);
    	if (ret < 0)
    		return ret;
 
+	while (amdgpu_device_is_reset(dev->dev_private))
+		msleep(100);
+
    	ret = drm_ioctl(filp, cmd, arg);
 
    	pm_runtime_mark_last_busy(dev->dev);
drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c +6 -1
index 2648291..22b8059 100644
@@ -36,10 +36,15 @@
    long amdgpu_kms_compat_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
    {
    	unsigned int nr = DRM_IOCTL_NR(cmd);
+	struct drm_file *file_priv = filp->private_data;
+	struct amdgpu_device *adev = file_priv->minor->dev->dev_private;
    	int ret;
 
-	if (nr < DRM_COMMAND_BASE)
+	if (nr < DRM_COMMAND_BASE) {
+		while (amdgpu_device_is_reset(adev))
+			msleep(100);
    		return drm_compat_ioctl(filp, cmd, arg);
+	}
 
Quoted 6 lines
Quoted 2 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c line 50

Quoted 2 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c +0 -0
    	ret = amdgpu_drm_ioctl(filp, cmd, arg);
 

_______________________________________________
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

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

Hi Christian,

We possibly must re-consider this patch, the direct issue is VM FAULT after gpu reset. one obvious bug is we could bind gtt to gart during GPU reset, binding gtt to gart needs to fill gart table and tlb flush, which is parallel with gpu reset thread.

Regards,
David Zhou

Quoted 117 lines · Christian König

NAK, that is exactly what we wanted to avoid.

We used to have an exclusive lock for this and it cause a whole bunch of problems.

Please elaborate why that should be necessary.

Regards,
Christian.

Am 21.04.2017 um 09:08 schrieb Roger.He:

Quoted 107 lines
Change-Id Ib77d33a09f348ebf2e3a9d7861411f4b951ebf7c
---
  drivers/gpu/drm/amd/amdgpu/amdgpu.h        |  6 ++++++
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 16 +++++++++++++++-
  drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c    |  4 ++++
  drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c  |  7 ++++++-
  4 files changed, 31 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +6 -0
index 71364f5..ab0ffa8 100644
@@ -1526,6 +1526,7 @@ struct amdgpu_device {
      atomic64_t            num_bytes_moved;
      atomic64_t            num_evictions;
      atomic_t            gpu_reset_counter;
+    atomic_t            gpu_reset_state;
        /* data for buffer migration throttling */
      struct {
@@ -1851,6 +1852,11 @@ amdgpu_get_sdma_instance(struct amdgpu_ring 
 *ring)
  #define amdgpu_psp_check_fw_loading_status(adev, i) 
 (adev)->firmware.funcs->check_fw_loading_status((adev), (i))
    /* Common functions */
+static inline int amdgpu_device_is_reset(struct amdgpu_device *adev)
+{
+    return atomic_read(&adev->gpu_reset_state);
+}
+
  int amdgpu_gpu_reset(struct amdgpu_device *adev);
  bool amdgpu_need_backup(struct amdgpu_device *adev);
  void amdgpu_pci_config_reset(struct amdgpu_device *adev);
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1 -0
index f882496..0fb4716 100644
@@ -1894,6 +1894,7 @@ int amdgpu_device_init(struct amdgpu_device *adev,
      mutex_init(&adev->grbm_idx_mutex);
      mutex_init(&adev->mn_lock);
      hash_init(adev->mn_hash);
+    atomic_set(&adev->gpu_reset_state, 0);
        amdgpu_check_arguments(adev);
  @@ -2655,6 +2656,18 @@ int amdgpu_sriov_gpu_reset(struct 
 amdgpu_device *adev, bool voluntary)

}

    /**

+ * amdgpu_device_set_reset_state - set gpu reset state
+ *
+ * @adev: amdgpu device pointer
+ * @state: true when start to reset gpu; false: reset done
+ */
+static inline void amdgpu_device_set_reset_state(struct amdgpu_device *adev,

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +7 -1
+                            bool state)
+{
+    atomic_set(&adev->gpu_reset_state, state);
+}
+
+/**
   * amdgpu_gpu_reset - reset the asic
   *
   * @adev: amdgpu device pointer
@@ -2678,7 +2691,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
      }
        atomic_inc(&adev->gpu_reset_counter);
-
+    amdgpu_device_set_reset_state(adev, true);
      /* block TTM */
      resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
      /* store modesetting */
@@ -2811,6 +2824,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
          dev_info(adev->dev, "GPU reset failed\n");
      }
  +    amdgpu_device_set_reset_state(adev, false);
      return r;
  }
  diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c 

b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
index ead00d7..8cc14af 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +3 -0
@@ -679,11 +679,15 @@ long amdgpu_drm_ioctl(struct file *filp,
      struct drm_file *file_priv = filp->private_data;
      struct drm_device *dev;
      long ret;
+
      dev = file_priv->minor->dev;
      ret = pm_runtime_get_sync(dev->dev);
      if (ret < 0)
          return ret;
  +    while (amdgpu_device_is_reset(dev->dev_private))
+        msleep(100);
+
      ret = drm_ioctl(filp, cmd, arg);
        pm_runtime_mark_last_busy(dev->dev);
drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c +2 -0
index 2648291..22b8059 100644
@@ -36,10 +36,15 @@
  long amdgpu_kms_compat_ioctl(struct file *filp, unsigned int cmd, 
 unsigned long arg)
  {
      unsigned int nr = DRM_IOCTL_NR(cmd);
+    struct drm_file *file_priv = filp->private_data;
+    struct amdgpu_device *adev = file_priv->minor->dev->dev_private;
Quoted 10 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c line 49

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c +4 -0
      int ret;
  -    if (nr < DRM_COMMAND_BASE)
+    if (nr < DRM_COMMAND_BASE) {
+        while (amdgpu_device_is_reset(adev))
+            msleep(100);
          return drm_compat_ioctl(filp, cmd, arg);
+    }
        ret = amdgpu_drm_ioctl(filp, cmd, arg);

_______________________________________________
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

But that should be harmless because we rebuild the GART table after GPU reset, don't we?

Apart from that I'm fine with having a lock or locks to prevent concurrent things from happening while we do the reset.

But we should NOT do the mistake again and try to prevent everything is one big global lock. That was rather bad idea and using an atomic isn't better in any way.

Regards,
Christian.

Am 04.05.2017 um 12:31 schrieb zhoucm1:

Quoted 137 lines

Hi Christian,

We possibly must re-consider this patch, the direct issue is VM FAULT after gpu reset. one obvious bug is we could bind gtt to gart during GPU reset, binding gtt to gart needs to fill gart table and tlb flush, which is parallel with gpu reset thread.

Regards,
David Zhou

Quoted 118 lines · Christian König

NAK, that is exactly what we wanted to avoid.

We used to have an exclusive lock for this and it cause a whole bunch of problems.

Please elaborate why that should be necessary.

Regards,
Christian.

Am 21.04.2017 um 09:08 schrieb Roger.He:

Quoted 108 lines
Change-Id Ib77d33a09f348ebf2e3a9d7861411f4b951ebf7c
---
  drivers/gpu/drm/amd/amdgpu/amdgpu.h        |  6 ++++++
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 16 +++++++++++++++-
  drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c    |  4 ++++
  drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c  |  7 ++++++-
  4 files changed, 31 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +6 -0
index 71364f5..ab0ffa8 100644
@@ -1526,6 +1526,7 @@ struct amdgpu_device {
      atomic64_t            num_bytes_moved;
      atomic64_t            num_evictions;
      atomic_t            gpu_reset_counter;
+    atomic_t            gpu_reset_state;
        /* data for buffer migration throttling */
      struct {
@@ -1851,6 +1852,11 @@ amdgpu_get_sdma_instance(struct amdgpu_ring 
 *ring)
  #define amdgpu_psp_check_fw_loading_status(adev, i) 
 (adev)->firmware.funcs->check_fw_loading_status((adev), (i))
    /* Common functions */
+static inline int amdgpu_device_is_reset(struct amdgpu_device *adev)
+{
+    return atomic_read(&adev->gpu_reset_state);
+}
+
  int amdgpu_gpu_reset(struct amdgpu_device *adev);
  bool amdgpu_need_backup(struct amdgpu_device *adev);
  void amdgpu_pci_config_reset(struct amdgpu_device *adev);
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1 -0
index f882496..0fb4716 100644
@@ -1894,6 +1894,7 @@ int amdgpu_device_init(struct amdgpu_device 
 *adev,
      mutex_init(&adev->grbm_idx_mutex);
      mutex_init(&adev->mn_lock);
      hash_init(adev->mn_hash);
+    atomic_set(&adev->gpu_reset_state, 0);
        amdgpu_check_arguments(adev);
  @@ -2655,6 +2656,18 @@ int amdgpu_sriov_gpu_reset(struct 

amdgpu_device *adev, bool voluntary)
}

    /**

+ * amdgpu_device_set_reset_state - set gpu reset state
+ *
+ * @adev: amdgpu device pointer
+ * @state: true when start to reset gpu; false: reset done
+ */
+static inline void amdgpu_device_set_reset_state(struct amdgpu_device *adev,

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +7 -1
+                            bool state)
+{
+    atomic_set(&adev->gpu_reset_state, state);
+}
+
+/**
   * amdgpu_gpu_reset - reset the asic
   *
   * @adev: amdgpu device pointer
@@ -2678,7 +2691,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
      }
        atomic_inc(&adev->gpu_reset_counter);
-
+    amdgpu_device_set_reset_state(adev, true);
      /* block TTM */
      resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
      /* store modesetting */
@@ -2811,6 +2824,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
          dev_info(adev->dev, "GPU reset failed\n");
      }
  +    amdgpu_device_set_reset_state(adev, false);
      return r;
  }
  diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c 

b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
index ead00d7..8cc14af 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +3 -0
@@ -679,11 +679,15 @@ long amdgpu_drm_ioctl(struct file *filp,
      struct drm_file *file_priv = filp->private_data;
      struct drm_device *dev;
      long ret;
+
      dev = file_priv->minor->dev;
      ret = pm_runtime_get_sync(dev->dev);
      if (ret < 0)
          return ret;
  +    while (amdgpu_device_is_reset(dev->dev_private))
+        msleep(100);
+
      ret = drm_ioctl(filp, cmd, arg);
        pm_runtime_mark_last_busy(dev->dev);
drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c +2 -0
index 2648291..22b8059 100644
@@ -36,10 +36,15 @@
  long amdgpu_kms_compat_ioctl(struct file *filp, unsigned int cmd, 
 unsigned long arg)
  {
      unsigned int nr = DRM_IOCTL_NR(cmd);
+    struct drm_file *file_priv = filp->private_data;
+    struct amdgpu_device *adev = file_priv->minor->dev->dev_private;
Quoted 10 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c line 49

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c +4 -0
      int ret;
  -    if (nr < DRM_COMMAND_BASE)
+    if (nr < DRM_COMMAND_BASE) {
+        while (amdgpu_device_is_reset(adev))
+            msleep(100);
          return drm_compat_ioctl(filp, cmd, arg);
+    }
        ret = amdgpu_drm_ioctl(filp, cmd, arg);
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): f2c6a892-5929-8644-ca34-09f89e6359a4-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

#26 lore raw

Christian

Why not use rw lock ? that way only gpu reset will make those iOCTL blocked, and in usual case the performance is still the same

BR Monk

-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Christian König
Sent: Thursday, May 04, 2017 7:21 PM
To: Zhou, David(ChunMing) <[email protected]>; He, Hongbo <[email protected]>; [email protected]
Subject: Re: [PATCH 3/3] drm/amdgpu: fix gpu reset issue

But that should be harmless because we rebuild the GART table after GPU reset, don't we?

Apart from that I'm fine with having a lock or locks to prevent concurrent things from happening while we do the reset.

But we should NOT do the mistake again and try to prevent everything is one big global lock. That was rather bad idea and using an atomic isn't better in any way.

Regards,
Christian.

Am 04.05.2017 um 12:31 schrieb zhoucm1:

Quoted 142 lines

Hi Christian,

We possibly must re-consider this patch, the direct issue is VM FAULT after gpu reset. one obvious bug is we could bind gtt to gart during GPU reset, binding gtt to gart needs to fill gart table and tlb flush, which is parallel with gpu reset thread.

Regards,
David Zhou

Quoted 120 lines · Christian König

NAK, that is exactly what we wanted to avoid.

We used to have an exclusive lock for this and it cause a whole bunch of problems.

Please elaborate why that should be necessary.

Regards,
Christian.

Am 21.04.2017 um 09:08 schrieb Roger.He:

Quoted 109 lines
Change-Id Ib77d33a09f348ebf2e3a9d7861411f4b951ebf7c
---
  drivers/gpu/drm/amd/amdgpu/amdgpu.h        |  6 ++++++
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 16 +++++++++++++++-
  drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c    |  4 ++++
  drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c  |  7 ++++++-
  4 files changed, 31 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +6 -0
index 71364f5..ab0ffa8 100644
@@ -1526,6 +1526,7 @@ struct amdgpu_device {
      atomic64_t            num_bytes_moved;
      atomic64_t            num_evictions;
      atomic_t            gpu_reset_counter;
+    atomic_t            gpu_reset_state;
        /* data for buffer migration throttling */
      struct {
@@ -1851,6 +1852,11 @@ amdgpu_get_sdma_instance(struct amdgpu_ring
 *ring)
  #define amdgpu_psp_check_fw_loading_status(adev, i) 
 (adev)->firmware.funcs->check_fw_loading_status((adev), (i))
    /* Common functions */
+static inline int amdgpu_device_is_reset(struct amdgpu_device 
+*adev) {
+    return atomic_read(&adev->gpu_reset_state);
+}
+
  int amdgpu_gpu_reset(struct amdgpu_device *adev);
  bool amdgpu_need_backup(struct amdgpu_device *adev);

void amdgpu_pci_config_reset(struct amdgpu_device *adev); diff
--git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index f882496..0fb4716 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1 -0
@@ -1894,6 +1894,7 @@ int amdgpu_device_init(struct amdgpu_device 
 *adev,
      mutex_init(&adev->grbm_idx_mutex);
      mutex_init(&adev->mn_lock);
      hash_init(adev->mn_hash);
+    atomic_set(&adev->gpu_reset_state, 0);
        amdgpu_check_arguments(adev);
  @@ -2655,6 +2656,18 @@ int amdgpu_sriov_gpu_reset(struct 

amdgpu_device *adev, bool voluntary)
}

    /**

+ * amdgpu_device_set_reset_state - set gpu reset state
+ *
+ * @adev: amdgpu device pointer
+ * @state: true when start to reset gpu; false: reset done */ +static inline void amdgpu_device_set_reset_state(struct amdgpu_device *adev,

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +5 -1
+                            bool state) {
+    atomic_set(&adev->gpu_reset_state, state); }
+
+/**
   * amdgpu_gpu_reset - reset the asic
   *
   * @adev: amdgpu device pointer
@@ -2678,7 +2691,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
      }
        atomic_inc(&adev->gpu_reset_counter);
-
+    amdgpu_device_set_reset_state(adev, true);
      /* block TTM */
      resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
      /* store modesetting */
@@ -2811,6 +2824,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
          dev_info(adev->dev, "GPU reset failed\n");
      }
  +    amdgpu_device_set_reset_state(adev, false);
      return r;
  }
  diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c

b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
index ead00d7..8cc14af 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +3 -0
@@ -679,11 +679,15 @@ long amdgpu_drm_ioctl(struct file *filp,
      struct drm_file *file_priv = filp->private_data;
      struct drm_device *dev;
      long ret;
+
      dev = file_priv->minor->dev;
      ret = pm_runtime_get_sync(dev->dev);
      if (ret < 0)
          return ret;
  +    while (amdgpu_device_is_reset(dev->dev_private))
+        msleep(100);
+
      ret = drm_ioctl(filp, cmd, arg);
        pm_runtime_mark_last_busy(dev->dev);
drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c +3 -0
index 2648291..22b8059 100644
@@ -36,10 +36,15 @@
  long amdgpu_kms_compat_ioctl(struct file *filp, unsigned int cmd, 
 unsigned long arg)
  {
      unsigned int nr = DRM_IOCTL_NR(cmd);
+    struct drm_file *file_priv = filp->private_data;
+    struct amdgpu_device *adev = 
+ file_priv->minor->dev->dev_private;
Quoted 10 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c line 50

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c +4 -0
      int ret;
  -    if (nr < DRM_COMMAND_BASE)
+    if (nr < DRM_COMMAND_BASE) {
+        while (amdgpu_device_is_reset(adev))
+            msleep(100);
          return drm_compat_ioctl(filp, cmd, arg);
+    }
        ret = amdgpu_drm_ioctl(filp, cmd, arg);
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): DM5PR12MB1610FEFA4B2654FF67B2F1B684EA0-2J9CzHegvk++jCVTvoAFKAdYzm3356FpvxpqHgZTriW3zl9H0oFU5g@public.gmane.org

#28 lore raw

That's what we used to have with the exclusive lock.

The issue is not what kind of lock or locking or not locking at all, but rather where to place it.

Having a rw lock called something like "in_reset" is ok to me, but we shouldn't take it in the IOCTL path.

Instead we should block for example the GART TLB flush while being inside a reset etc...

Otherwise we run into the same problems (lock inversions for example) as we had before.

Regards,
Christian.

Am 04.05.2017 um 13:28 schrieb Liu, Monk:

Quoted 168 lines

Christian

Why not use rw lock ? that way only gpu reset will make those iOCTL blocked, and in usual case the performance is still the same

BR Monk

-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Christian König
Sent: Thursday, May 04, 2017 7:21 PM
To: Zhou, David(ChunMing) <[email protected]>; He, Hongbo <[email protected]>; [email protected]
Subject: Re: [PATCH 3/3] drm/amdgpu: fix gpu reset issue

But that should be harmless because we rebuild the GART table after GPU reset, don't we?

Apart from that I'm fine with having a lock or locks to prevent concurrent things from happening while we do the reset.

But we should NOT do the mistake again and try to prevent everything is one big global lock. That was rather bad idea and using an atomic isn't better in any way.

Regards,
Christian.

Am 04.05.2017 um 12:31 schrieb zhoucm1:

Quoted 140 lines

Hi Christian,

We possibly must re-consider this patch, the direct issue is VM FAULT after gpu reset. one obvious bug is we could bind gtt to gart during GPU reset, binding gtt to gart needs to fill gart table and tlb flush, which is parallel with gpu reset thread.

Regards,
David Zhou

Quoted 120 lines · Christian König

NAK, that is exactly what we wanted to avoid.

We used to have an exclusive lock for this and it cause a whole bunch of problems.

Please elaborate why that should be necessary.

Regards,
Christian.

Am 21.04.2017 um 09:08 schrieb Roger.He:

Quoted 109 lines
Change-Id Ib77d33a09f348ebf2e3a9d7861411f4b951ebf7c
---
   drivers/gpu/drm/amd/amdgpu/amdgpu.h        |  6 ++++++
   drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 16 +++++++++++++++-
   drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c    |  4 ++++
   drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c  |  7 ++++++-
   4 files changed, 31 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +6 -0
index 71364f5..ab0ffa8 100644
@@ -1526,6 +1526,7 @@ struct amdgpu_device {
       atomic64_t            num_bytes_moved;
       atomic64_t            num_evictions;
       atomic_t            gpu_reset_counter;
+    atomic_t            gpu_reset_state;
         /* data for buffer migration throttling */
       struct {
@@ -1851,6 +1852,11 @@ amdgpu_get_sdma_instance(struct amdgpu_ring
 *ring)
   #define amdgpu_psp_check_fw_loading_status(adev, i)
 (adev)->firmware.funcs->check_fw_loading_status((adev), (i))
     /* Common functions */
+static inline int amdgpu_device_is_reset(struct amdgpu_device
+*adev) {
+    return atomic_read(&adev->gpu_reset_state);
+}
+
   int amdgpu_gpu_reset(struct amdgpu_device *adev);
   bool amdgpu_need_backup(struct amdgpu_device *adev);

void amdgpu_pci_config_reset(struct amdgpu_device *adev); diff
--git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index f882496..0fb4716 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1 -0
@@ -1894,6 +1894,7 @@ int amdgpu_device_init(struct amdgpu_device
 *adev,
       mutex_init(&adev->grbm_idx_mutex);
       mutex_init(&adev->mn_lock);
       hash_init(adev->mn_hash);
+    atomic_set(&adev->gpu_reset_state, 0);
         amdgpu_check_arguments(adev);
   @@ -2655,6 +2656,18 @@ int amdgpu_sriov_gpu_reset(struct

amdgpu_device *adev, bool voluntary)
}

     /**

+ * amdgpu_device_set_reset_state - set gpu reset state
+ *
+ * @adev: amdgpu device pointer
+ * @state: true when start to reset gpu; false: reset done */ +static inline void amdgpu_device_set_reset_state(struct amdgpu_device *adev,

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +5 -1
+                            bool state) {
+    atomic_set(&adev->gpu_reset_state, state); }
+
+/**
    * amdgpu_gpu_reset - reset the asic
    *
    * @adev: amdgpu device pointer
@@ -2678,7 +2691,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
       }
         atomic_inc(&adev->gpu_reset_counter);
-
+    amdgpu_device_set_reset_state(adev, true);
       /* block TTM */
       resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
       /* store modesetting */
@@ -2811,6 +2824,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
           dev_info(adev->dev, "GPU reset failed\n");
       }
   +    amdgpu_device_set_reset_state(adev, false);
       return r;
   }
   diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c

b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
index ead00d7..8cc14af 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +3 -0
@@ -679,11 +679,15 @@ long amdgpu_drm_ioctl(struct file *filp,
       struct drm_file *file_priv = filp->private_data;
       struct drm_device *dev;
       long ret;
+
       dev = file_priv->minor->dev;
       ret = pm_runtime_get_sync(dev->dev);
       if (ret < 0)
           return ret;
   +    while (amdgpu_device_is_reset(dev->dev_private))
+        msleep(100);
+
       ret = drm_ioctl(filp, cmd, arg);
         pm_runtime_mark_last_busy(dev->dev);
drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c +3 -0
index 2648291..22b8059 100644
@@ -36,10 +36,15 @@
   long amdgpu_kms_compat_ioctl(struct file *filp, unsigned int cmd,
 unsigned long arg)
   {
       unsigned int nr = DRM_IOCTL_NR(cmd);
+    struct drm_file *file_priv = filp->private_data;
+    struct amdgpu_device *adev =
+ file_priv->minor->dev->dev_private;
Quoted 9 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c line 50

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c +4 -0
       int ret;
   -    if (nr < DRM_COMMAND_BASE)
+    if (nr < DRM_COMMAND_BASE) {
+        while (amdgpu_device_is_reset(adev))
+            msleep(100);
           return drm_compat_ioctl(filp, cmd, arg);
+    }
         ret = amdgpu_drm_ioctl(filp, cmd, arg);
Quoted 4 lines

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

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): 6321c998-f337-39af-3d64-070341e5fbe4-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

#30 lore raw

Agreed.
We can put rw lock in the place that accessing hardware or pde/pte,
But I'm not sure this is safe enough, maybe one IOCTL access hw at top and bottom, and need consistence access, If gpu reset happens maybe the consistence is broken .... just my worry

But we can do from your approach to try first

BR Monk

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: Thursday, May 04, 2017 7:33 PM
To: Liu, Monk <[email protected]>; Zhou, David(ChunMing) <[email protected]>; He, Hongbo <[email protected]>; [email protected]
Subject: Re: [PATCH 3/3] drm/amdgpu: fix gpu reset issue

That's what we used to have with the exclusive lock.

The issue is not what kind of lock or locking or not locking at all, but rather where to place it.

Having a rw lock called something like "in_reset" is ok to me, but we shouldn't take it in the IOCTL path.

Instead we should block for example the GART TLB flush while being inside a reset etc...

Otherwise we run into the same problems (lock inversions for example) as we had before.

Regards,
Christian.

Am 04.05.2017 um 13:28 schrieb Liu, Monk:

Quoted 170 lines

Christian

Why not use rw lock ? that way only gpu reset will make those iOCTL blocked, and in usual case the performance is still the same

BR Monk

-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Christian König
Sent: Thursday, May 04, 2017 7:21 PM
To: Zhou, David(ChunMing) <[email protected]>; He, Hongbo <[email protected]>; [email protected]
Subject: Re: [PATCH 3/3] drm/amdgpu: fix gpu reset issue

But that should be harmless because we rebuild the GART table after GPU reset, don't we?

Apart from that I'm fine with having a lock or locks to prevent concurrent things from happening while we do the reset.

But we should NOT do the mistake again and try to prevent everything is one big global lock. That was rather bad idea and using an atomic isn't better in any way.

Regards,
Christian.

Am 04.05.2017 um 12:31 schrieb zhoucm1:

Quoted 140 lines

Hi Christian,

We possibly must re-consider this patch, the direct issue is VM FAULT after gpu reset. one obvious bug is we could bind gtt to gart during GPU reset, binding gtt to gart needs to fill gart table and tlb flush, which is parallel with gpu reset thread.

Regards,
David Zhou

Quoted 120 lines · Christian König

NAK, that is exactly what we wanted to avoid.

We used to have an exclusive lock for this and it cause a whole bunch of problems.

Please elaborate why that should be necessary.

Regards,
Christian.

Am 21.04.2017 um 09:08 schrieb Roger.He:

Quoted 109 lines
Change-Id Ib77d33a09f348ebf2e3a9d7861411f4b951ebf7c
---
   drivers/gpu/drm/amd/amdgpu/amdgpu.h        |  6 ++++++
   drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 16 +++++++++++++++-
   drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c    |  4 ++++
   drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c  |  7 ++++++-
   4 files changed, 31 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +6 -0
index 71364f5..ab0ffa8 100644
@@ -1526,6 +1526,7 @@ struct amdgpu_device {
       atomic64_t            num_bytes_moved;
       atomic64_t            num_evictions;
       atomic_t            gpu_reset_counter;
+    atomic_t            gpu_reset_state;
         /* data for buffer migration throttling */
       struct {
@@ -1851,6 +1852,11 @@ amdgpu_get_sdma_instance(struct amdgpu_ring
 *ring)
   #define amdgpu_psp_check_fw_loading_status(adev, i) 
 (adev)->firmware.funcs->check_fw_loading_status((adev), (i))
     /* Common functions */
+static inline int amdgpu_device_is_reset(struct amdgpu_device
+*adev) {
+    return atomic_read(&adev->gpu_reset_state);
+}
+
   int amdgpu_gpu_reset(struct amdgpu_device *adev);
   bool amdgpu_need_backup(struct amdgpu_device *adev);

void amdgpu_pci_config_reset(struct amdgpu_device *adev); diff
--git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index f882496..0fb4716 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1 -0
@@ -1894,6 +1894,7 @@ int amdgpu_device_init(struct amdgpu_device 
 *adev,
       mutex_init(&adev->grbm_idx_mutex);
       mutex_init(&adev->mn_lock);
       hash_init(adev->mn_hash);
+    atomic_set(&adev->gpu_reset_state, 0);
         amdgpu_check_arguments(adev);
   @@ -2655,6 +2656,18 @@ int amdgpu_sriov_gpu_reset(struct 

amdgpu_device *adev, bool voluntary)
}

     /**

+ * amdgpu_device_set_reset_state - set gpu reset state
+ *
+ * @adev: amdgpu device pointer
+ * @state: true when start to reset gpu; false: reset done */ +static inline void amdgpu_device_set_reset_state(struct amdgpu_device *adev,

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +5 -1
+                            bool state) {
+    atomic_set(&adev->gpu_reset_state, state); }
+
+/**
    * amdgpu_gpu_reset - reset the asic
    *
    * @adev: amdgpu device pointer
@@ -2678,7 +2691,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
       }
         atomic_inc(&adev->gpu_reset_counter);
-
+    amdgpu_device_set_reset_state(adev, true);
       /* block TTM */
       resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
       /* store modesetting */
@@ -2811,6 +2824,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
           dev_info(adev->dev, "GPU reset failed\n");
       }
   +    amdgpu_device_set_reset_state(adev, false);
       return r;
   }
   diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c

b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
index ead00d7..8cc14af 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +3 -0
@@ -679,11 +679,15 @@ long amdgpu_drm_ioctl(struct file *filp,
       struct drm_file *file_priv = filp->private_data;
       struct drm_device *dev;
       long ret;
+
       dev = file_priv->minor->dev;
       ret = pm_runtime_get_sync(dev->dev);
       if (ret < 0)
           return ret;
   +    while (amdgpu_device_is_reset(dev->dev_private))
+        msleep(100);
+
       ret = drm_ioctl(filp, cmd, arg);
         pm_runtime_mark_last_busy(dev->dev);
drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c +3 -0
index 2648291..22b8059 100644
@@ -36,10 +36,15 @@
   long amdgpu_kms_compat_ioctl(struct file *filp, unsigned int 
 cmd, unsigned long arg)
   {
       unsigned int nr = DRM_IOCTL_NR(cmd);
+    struct drm_file *file_priv = filp->private_data;
+    struct amdgpu_device *adev =
+ file_priv->minor->dev->dev_private;
Quoted 9 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c line 50

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ioc32.c +4 -0
       int ret;
   -    if (nr < DRM_COMMAND_BASE)
+    if (nr < DRM_COMMAND_BASE) {
+        while (amdgpu_device_is_reset(adev))
+            msleep(100);
           return drm_compat_ioctl(filp, cmd, arg);
+    }
         ret = amdgpu_drm_ioctl(filp, cmd, arg);
Quoted 4 lines

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

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