Skip to contentSkip to conversation

drm/amdgpu: stop all rings before doing gpu recover

PATCH NAK Monk Liu 5 participants 32 messages Feb 28, 2018 – Mar 1, 2018 23 messages not in the archive

Subsystems (2)
lore
Share & tools
lore

Monk Liu

author
#0 lore raw

PATCH 1/4 drm/amdgpu: stop all rings before doing gpu recover

found recover_vram_from_shadow sometimes get executed in paralle with SDMA scheduler, should stop all schedulers before doing gpu reset/recover

Change-Id Ibaef3e3c015f3cf88f84b2eaf95cda95ae1a64e3
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 40 +++++++++++-------------------
 1 file changed, 15 insertions(+), 25 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +15 -25
index 75d1733..e9d81a8 100644
@@ -2649,22 +2649,23 @@ int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
 
 	/* block TTM */
 	resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
+
 	/* store modesetting */
 	if (amdgpu_device_has_dc_support(adev))
 		state = drm_atomic_helper_suspend(adev->ddev);
 
-	/* block scheduler */
+	/* block all schedulers and reset given job's ring */
 	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
 		struct amdgpu_ring *ring = adev->rings[i];
 
 		if (!ring || !ring->sched.thread)
 			continue;
 
-		/* only focus on the ring hit timeout if &job not NULL */
+		kthread_park(ring->sched.thread);
+
 		if (job && job->ring->idx != i)
 			continue;
 
-		kthread_park(ring->sched.thread);
 		drm_sched_hw_job_reset(&ring->sched, &job->base);
 
 		/* after all hw jobs are reset, hw fence is meaningless, so force_completion */
@@ -2707,33 +2708,22 @@ int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
 			}
 			dma_fence_put(fence);
 		}
+	}
 
-		for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
-			struct amdgpu_ring *ring = adev->rings[i];
-
-			if (!ring || !ring->sched.thread)
-				continue;
+	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
+		struct amdgpu_ring *ring = adev->rings[i];
 
-			/* only focus on the ring hit timeout if &job not NULL */
-			if (job && job->ring->idx != i)
-				continue;
+		if (!ring || !ring->sched.thread)
+			continue;
 
+		/* only need recovery sched of the given job's ring
+		 * or all rings (in the case @job is NULL)
+		 * after above amdgpu_reset accomplished
+		 */
+		if ((!job || job->ring->idx == i) && !r)
 			drm_sched_job_recovery(&ring->sched);
-			kthread_unpark(ring->sched.thread);
-		}
-	} else {
-		for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
-			struct amdgpu_ring *ring = adev->rings[i];
 
-			if (!ring || !ring->sched.thread)
-				continue;
-
-			/* only focus on the ring hit timeout if &job not NULL */
-			if (job && job->ring->idx != i)
-				continue;
-
-			kthread_unpark(adev->rings[i]->sched.thread);
-		}
+		kthread_unpark(ring->sched.thread);
 	}
 
 	if (amdgpu_device_has_dc_support(adev)) {
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): 1519802463-9090-1-git-send-email-Monk.Liu-5C7GfCeVMHo@public.gmane.org

author
#2 lore raw

PATCH 2/4 drm/amdgpu: cleanups for vram lost handling

1)create a routine "handle_vram_lost" to do the vram recovery, and put it into amdgpu_device_reset/reset_sriov, this way no need of the extra paramter to hold the VRAM LOST information and the related macros can be removed.

3)show vram_recover failure if time out, and set TMO equal to lockup_timeout if vram_recover is under SRIOV runtime mode.

4)report error if any ip reset failed for SR-IOV

Change-Id I686e2b6133844c14948c206a2315c064a78c1d9c
---
 drivers/gpu/drm/amd/amdgpu/amdgpu.h        |   4 -
 drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 136 +++++++++++++++--------------
 2 files changed, 72 insertions(+), 68 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +0 -4
index 5bddfc1..abbc3f1 100644
@@ -181,10 +181,6 @@ extern int amdgpu_cik_support;
 #define CIK_CURSOR_WIDTH 128
 #define CIK_CURSOR_HEIGHT 128
 
-/* GPU RESET flags */
-#define AMDGPU_RESET_INFO_VRAM_LOST  (1 << 0)
-#define AMDGPU_RESET_INFO_FULLRESET  (1 << 1)
-
 struct amdgpu_device;
 struct amdgpu_ib;
 struct amdgpu_cs_parser;
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +72 -64

205 lines not on this page · raw message

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): 1519802463-9090-2-git-send-email-Monk.Liu-5C7GfCeVMHo@public.gmane.org

#4 lore raw

Re: [PATCH 2/4] drm/amdgpu: cleanups for vram lost handling

Am 28.02.2018 um 08:21 schrieb Monk Liu:

Quoted 5 lines

1)create a routine "handle_vram_lost" to do the vram recovery, and put it into amdgpu_device_reset/reset_sriov, this way no need of the extra paramter to hold the VRAM LOST information and the related macros can be removed.

Quoted 7 lines

3)show vram_recover failure if time out, and set TMO equal to lockup_timeout if vram_recover is under SRIOV runtime mode.

4)report error if any ip reset failed for SR-IOV

Change-Id I686e2b6133844c14948c206a2315c064a78c1d9c

Looks good to me, but not 100% sure if all of this is technically correct.

Patch is Acked-by: Christian König <[email protected]> for now.

Andrey and/or David should probably take a look as well.

Christian.

Quoted 227 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu.h        |   4 -
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 136 +++++++++++++++--------------
  2 files changed, 72 insertions(+), 68 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +0 -4
index 5bddfc1..abbc3f1 100644
@@ -181,10 +181,6 @@ extern int amdgpu_cik_support;
  #define CIK_CURSOR_WIDTH 128
  #define CIK_CURSOR_HEIGHT 128
 
-/* GPU RESET flags */
-#define AMDGPU_RESET_INFO_VRAM_LOST  (1 << 0)
-#define AMDGPU_RESET_INFO_FULLRESET  (1 << 1)
-
  struct amdgpu_device;
  struct amdgpu_ib;
  struct amdgpu_cs_parser;
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +71 -58
index e9d81a8..39ece7f 100644
@@ -1591,6 +1591,8 @@ static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev)
 
  			r = block->version->funcs->hw_init(adev);
  			DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name, r?"failed":"successed");
+			if (r)
+				return r;
  		}
  	}
 
@@ -1624,6 +1626,8 @@ static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev)
 
  			r = block->version->funcs->hw_init(adev);
  			DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name, r?"failed":"successed");
+			if (r)
+				return r;
  		}
  	}
 
@@ -2471,17 +2475,71 @@ static int amdgpu_device_recover_vram_from_shadow(struct amdgpu_device *adev,
  	return r;
  }
 
+static int amdgpu_handle_vram_lost(struct amdgpu_device *adev)
+{
+	struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
+	struct amdgpu_bo *bo, *tmp;
+	struct dma_fence *fence = NULL, *next = NULL;
+	long r = 1;
+	int i = 0;
+	long tmo;
+
+	if (amdgpu_sriov_runtime(adev))
+		tmo = msecs_to_jiffies(amdgpu_lockup_timeout);
+	else
+		tmo = msecs_to_jiffies(100);
+
+	DRM_INFO("recover vram bo from shadow start\n");
+	mutex_lock(&adev->shadow_list_lock);
+	list_for_each_entry_safe(bo, tmp, &adev->shadow_list, shadow_list) {
+		next = NULL;
+		amdgpu_device_recover_vram_from_shadow(adev, ring, bo, &next);
+		if (fence) {
+			r = dma_fence_wait_timeout(fence, false, tmo);
+			if (r == 0)
+				pr_err("wait fence %p[%d] timeout\n", fence, i);
+			else if (r < 0)
+				pr_err("wait fence %p[%d] interrupted\n", fence, i);
+			if (r < 1) {
+				dma_fence_put(fence);
+				fence = next;
+				break;
+			}
+			i++;
+		}
+
+		dma_fence_put(fence);
+		fence = next;
+	}
+	mutex_unlock(&adev->shadow_list_lock);
+
+	if (fence) {
+		r = dma_fence_wait_timeout(fence, false, tmo);
+		if (r == 0)
+			pr_err("wait fence %p[%d] timeout\n", fence, i);
+		else if (r < 0)
+			pr_err("wait fence %p[%d] interrupted\n", fence, i);
+
+	}
+	dma_fence_put(fence);
+
+	if (r > 0)
+		DRM_INFO("recover vram bo from shadow done\n");
+	else
+		DRM_ERROR("recover vram bo from shadow failed\n");
+
+	return (r > 0?0:1);
+}
+
  /*
   * amdgpu_device_reset - reset ASIC/GPU for bare-metal or passthrough
   *
   * @adev: amdgpu device pointer
- * @reset_flags: output param tells caller the reset result
   *
   * attempt to do soft-reset or full-reset and reinitialize Asic
   * return 0 means successed otherwise failed
  */
-static int amdgpu_device_reset(struct amdgpu_device *adev,
-			       uint64_t* reset_flags)
+static int amdgpu_device_reset(struct amdgpu_device *adev)
  {
  	bool need_full_reset, vram_lost = 0;
  	int r;
@@ -2496,7 +2554,6 @@ static int amdgpu_device_reset(struct amdgpu_device *adev,
  			DRM_INFO("soft reset failed, will fallback to full reset!\n");
  			need_full_reset = true;
  		}
-
  	}
 
  	if (need_full_reset) {
@@ -2545,13 +2602,8 @@ static int amdgpu_device_reset(struct amdgpu_device *adev,
  		}
  	}
 
-	if (reset_flags) {
-		if (vram_lost)
-			(*reset_flags) |= AMDGPU_RESET_INFO_VRAM_LOST;
-
-		if (need_full_reset)
-			(*reset_flags) |= AMDGPU_RESET_INFO_FULLRESET;
-	}
+	if (!r && ((need_full_reset && !(adev->flags & AMD_IS_APU)) || vram_lost))
+		r = amdgpu_handle_vram_lost(adev);
 
  	return r;
  }
@@ -2560,14 +2612,11 @@ static int amdgpu_device_reset(struct amdgpu_device *adev,
   * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf
   *
   * @adev: amdgpu device pointer
- * @reset_flags: output param tells caller the reset result
   *
   * do VF FLR and reinitialize Asic
   * return 0 means successed otherwise failed
  */
-static int amdgpu_device_reset_sriov(struct amdgpu_device *adev,
-				     uint64_t *reset_flags,
-				     bool from_hypervisor)
+static int amdgpu_device_reset_sriov(struct amdgpu_device *adev, bool from_hypervisor)
  {
  	int r;
 
@@ -2588,28 +2637,20 @@ static int amdgpu_device_reset_sriov(struct amdgpu_device *adev,
 
  	/* now we are okay to resume SMC/CP/SDMA */
  	r = amdgpu_device_ip_reinit_late_sriov(adev);
+	amdgpu_virt_release_full_gpu(adev, true);
  	if (r)
  		goto error;
 
  	amdgpu_irq_gpu_reset_resume_helper(adev);
  	r = amdgpu_ib_ring_tests(adev);
-	if (r)
-		dev_err(adev->dev, "[GPU_RESET] ib ring test failed (%d).\n", r);
 
-error:
-	/* release full control of GPU after ib test */
-	amdgpu_virt_release_full_gpu(adev, true);
-
-	if (reset_flags) {
-		if (adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
-			(*reset_flags) |= AMDGPU_RESET_INFO_VRAM_LOST;
-			atomic_inc(&adev->vram_lost_counter);
-		}
-
-		/* VF FLR or hotlink reset is always full-reset */
-		(*reset_flags) |= AMDGPU_RESET_INFO_FULLRESET;
+	if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
+		atomic_inc(&adev->vram_lost_counter);
+		r = amdgpu_handle_vram_lost(adev);
  	}
 
+error:
+
  	return r;
  }
 
@@ -2673,42 +2714,9 @@ int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
  	}
 
  	if (amdgpu_sriov_vf(adev))
-		r = amdgpu_device_reset_sriov(adev, &reset_flags, job ? false : true);
+		r = amdgpu_device_reset_sriov(adev, job ? false : true);
  	else
-		r = amdgpu_device_reset(adev, &reset_flags);
-
-	if (!r) {
-		if (((reset_flags & AMDGPU_RESET_INFO_FULLRESET) && !(adev->flags & AMD_IS_APU)) ||
-			(reset_flags & AMDGPU_RESET_INFO_VRAM_LOST)) {
-			struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
-			struct amdgpu_bo *bo, *tmp;
-			struct dma_fence *fence = NULL, *next = NULL;
-
-			DRM_INFO("recover vram bo from shadow\n");
-			mutex_lock(&adev->shadow_list_lock);
-			list_for_each_entry_safe(bo, tmp, &adev->shadow_list, shadow_list) {
-				next = NULL;
-				amdgpu_device_recover_vram_from_shadow(adev, ring, bo, &next);
-				if (fence) {
-					r = dma_fence_wait(fence, false);
-					if (r) {
-						WARN(r, "recovery from shadow isn't completed\n");
-						break;
-					}
-				}
-
-				dma_fence_put(fence);
-				fence = next;
-			}
-			mutex_unlock(&adev->shadow_list_lock);
-			if (fence) {
-				r = dma_fence_wait(fence, false);

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c int amdgpu_device_gpu_recover(struct amdgpu_device *adev, · line 2722

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1 -6
-				if (r)
-					WARN(r, "recovery from shadow isn't completed\n");
-			}
-			dma_fence_put(fence);
-		}
-	}
+		r = amdgpu_device_reset(adev);
 
  	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
  		struct amdgpu_ring *ring = adev->rings[i];

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

#5 lore raw

Re: [PATCH 2/4] drm/amdgpu: cleanups for vram lost handling

In reply to a message not in the archive
Quoted 171 lines · Monk Liu

1)create a routine "handle_vram_lost" to do the vram recovery, and put it into amdgpu_device_reset/reset_sriov, this way no need of the extra paramter to hold the VRAM LOST information and the related macros can be removed.

3)show vram_recover failure if time out, and set TMO equal to lockup_timeout if vram_recover is under SRIOV runtime mode.

4)report error if any ip reset failed for SR-IOV

Change-Id I686e2b6133844c14948c206a2315c064a78c1d9c
---
  drivers/gpu/drm/amd/amdgpu/amdgpu.h        |   4 -
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 136 +++++++++++++++--------------
  2 files changed, 72 insertions(+), 68 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +0 -4
index 5bddfc1..abbc3f1 100644
@@ -181,10 +181,6 @@ extern int amdgpu_cik_support;
  #define CIK_CURSOR_WIDTH 128
  #define CIK_CURSOR_HEIGHT 128
 
-/* GPU RESET flags */
-#define AMDGPU_RESET_INFO_VRAM_LOST  (1 << 0)
-#define AMDGPU_RESET_INFO_FULLRESET  (1 << 1)
-
  struct amdgpu_device;
  struct amdgpu_ib;
  struct amdgpu_cs_parser;
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +64 -15
index e9d81a8..39ece7f 100644
@@ -1591,6 +1591,8 @@ static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev)
 
  			r = block->version->funcs->hw_init(adev);
  			DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name, r?"failed":"successed");
+			if (r)
+				return r;
  		}
  	}
 
@@ -1624,6 +1626,8 @@ static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev)
 
  			r = block->version->funcs->hw_init(adev);
  			DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name, r?"failed":"successed");
+			if (r)
+				return r;
  		}
  	}
 
@@ -2471,17 +2475,71 @@ static int amdgpu_device_recover_vram_from_shadow(struct amdgpu_device *adev,
  	return r;
  }
 
+static int amdgpu_handle_vram_lost(struct amdgpu_device *adev)
+{
+	struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
+	struct amdgpu_bo *bo, *tmp;
+	struct dma_fence *fence = NULL, *next = NULL;
+	long r = 1;
+	int i = 0;
+	long tmo;
+
+	if (amdgpu_sriov_runtime(adev))
+		tmo = msecs_to_jiffies(amdgpu_lockup_timeout);
+	else
+		tmo = msecs_to_jiffies(100);
+
+	DRM_INFO("recover vram bo from shadow start\n");
+	mutex_lock(&adev->shadow_list_lock);
+	list_for_each_entry_safe(bo, tmp, &adev->shadow_list, shadow_list) {
+		next = NULL;
+		amdgpu_device_recover_vram_from_shadow(adev, ring, bo, &next);
+		if (fence) {
+			r = dma_fence_wait_timeout(fence, false, tmo);
+			if (r == 0)
+				pr_err("wait fence %p[%d] timeout\n", fence, i);
+			else if (r < 0)
+				pr_err("wait fence %p[%d] interrupted\n", fence, i);
+			if (r < 1) {
+				dma_fence_put(fence);
+				fence = next;
+				break;
+			}
+			i++;
+		}
+
+		dma_fence_put(fence);
+		fence = next;
+	}
+	mutex_unlock(&adev->shadow_list_lock);
+
+	if (fence) {
+		r = dma_fence_wait_timeout(fence, false, tmo);
+		if (r == 0)
+			pr_err("wait fence %p[%d] timeout\n", fence, i);
+		else if (r < 0)
+			pr_err("wait fence %p[%d] interrupted\n", fence, i);
+
+	}
+	dma_fence_put(fence);
+
+	if (r > 0)
+		DRM_INFO("recover vram bo from shadow done\n");
+	else
+		DRM_ERROR("recover vram bo from shadow failed\n");
+
+	return (r > 0?0:1);
+}
+
  /*
   * amdgpu_device_reset - reset ASIC/GPU for bare-metal or passthrough
   *
   * @adev: amdgpu device pointer
- * @reset_flags: output param tells caller the reset result
   *
   * attempt to do soft-reset or full-reset and reinitialize Asic
   * return 0 means successed otherwise failed
  */
-static int amdgpu_device_reset(struct amdgpu_device *adev,
-			       uint64_t* reset_flags)
+static int amdgpu_device_reset(struct amdgpu_device *adev)
  {
  	bool need_full_reset, vram_lost = 0;
  	int r;
@@ -2496,7 +2554,6 @@ static int amdgpu_device_reset(struct amdgpu_device *adev,
  			DRM_INFO("soft reset failed, will fallback to full reset!\n");
  			need_full_reset = true;
  		}
-
  	}
 
  	if (need_full_reset) {
@@ -2545,13 +2602,8 @@ static int amdgpu_device_reset(struct amdgpu_device *adev,
  		}
  	}
 
-	if (reset_flags) {
-		if (vram_lost)
-			(*reset_flags) |= AMDGPU_RESET_INFO_VRAM_LOST;
-
-		if (need_full_reset)
-			(*reset_flags) |= AMDGPU_RESET_INFO_FULLRESET;
-	}
+	if (!r && ((need_full_reset && !(adev->flags & AMD_IS_APU)) || vram_lost))
+		r = amdgpu_handle_vram_lost(adev);
 
  	return r;
  }
@@ -2560,14 +2612,11 @@ static int amdgpu_device_reset(struct amdgpu_device *adev,
   * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf
   *
   * @adev: amdgpu device pointer
- * @reset_flags: output param tells caller the reset result
   *
   * do VF FLR and reinitialize Asic
   * return 0 means successed otherwise failed
  */
-static int amdgpu_device_reset_sriov(struct amdgpu_device *adev,
-				     uint64_t *reset_flags,
-				     bool from_hypervisor)
+static int amdgpu_device_reset_sriov(struct amdgpu_device *adev, bool from_hypervisor)
  {
  	int r;
 
@@ -2588,28 +2637,20 @@ static int amdgpu_device_reset_sriov(struct amdgpu_device *adev,
 

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c static int amdgpu_device_reset_sriov(struct amdgpu_device *adev, · line 2597

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1 -2
  	/* now we are okay to resume SMC/CP/SDMA */
  	r = amdgpu_device_ip_reinit_late_sriov(adev);
+	amdgpu_virt_release_full_gpu(adev, true);
  	if (r)
  		goto error;
 
  	amdgpu_irq_gpu_reset_resume_helper(adev);
  	r = amdgpu_ib_ring_tests(adev);
-	if (r)
-		dev_err(adev->dev, "[GPU_RESET] ib ring test failed (%d).\n", r);

This 2 lines deletion seems unintentional

Quoted 11 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -10
-error:
-	/* release full control of GPU after ib test */
-	amdgpu_virt_release_full_gpu(adev, true);
-
-	if (reset_flags) {
-		if (adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
-			(*reset_flags) |= AMDGPU_RESET_INFO_VRAM_LOST;
-			atomic_inc(&adev->vram_lost_counter);
-		}
-

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c static int amdgpu_device_reset_sriov(struct amdgpu_device *adev, · line 2654

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +5 -2
-		/* VF FLR or hotlink reset is always full-reset */
-		(*reset_flags) |= AMDGPU_RESET_INFO_FULLRESET;
+	if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
+		atomic_inc(&adev->vram_lost_counter);
+		r = amdgpu_handle_vram_lost(adev);
  	}
 
+error:
+
  	return r;

Not doing incrementation of VRAM lost counter in case of error anymore is intentional ?

Andrey

Quoted 37 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1 -29
  }
 
@@ -2673,42 +2714,9 @@ int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
  	}
 
  	if (amdgpu_sriov_vf(adev))
-		r = amdgpu_device_reset_sriov(adev, &reset_flags, job ? false : true);
+		r = amdgpu_device_reset_sriov(adev, job ? false : true);
  	else
-		r = amdgpu_device_reset(adev, &reset_flags);
-
-	if (!r) {
-		if (((reset_flags & AMDGPU_RESET_INFO_FULLRESET) && !(adev->flags & AMD_IS_APU)) ||
-			(reset_flags & AMDGPU_RESET_INFO_VRAM_LOST)) {
-			struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
-			struct amdgpu_bo *bo, *tmp;
-			struct dma_fence *fence = NULL, *next = NULL;
-
-			DRM_INFO("recover vram bo from shadow\n");
-			mutex_lock(&adev->shadow_list_lock);
-			list_for_each_entry_safe(bo, tmp, &adev->shadow_list, shadow_list) {
-				next = NULL;
-				amdgpu_device_recover_vram_from_shadow(adev, ring, bo, &next);
-				if (fence) {
-					r = dma_fence_wait(fence, false);
-					if (r) {
-						WARN(r, "recovery from shadow isn't completed\n");
-						break;
-					}
-				}
-
-				dma_fence_put(fence);
-				fence = next;
-			}
-			mutex_unlock(&adev->shadow_list_lock);
-			if (fence) {
-				r = dma_fence_wait(fence, false);

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c int amdgpu_device_gpu_recover(struct amdgpu_device *adev, · line 2722

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1 -6
-				if (r)
-					WARN(r, "recovery from shadow isn't completed\n");
-			}
-			dma_fence_put(fence);
-		}
-	}
+		r = amdgpu_device_reset(adev);
 
  	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
  		struct amdgpu_ring *ring = adev->rings[i];

_______________________________________________
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): 13337cd9-78e9-df36-f2ab-749cf182177b-5C7GfCeVMHo@public.gmane.org

author
#7 lore raw

RE: [PATCH 2/4] drm/amdgpu: cleanups for vram lost handling

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c static int amdgpu_device_reset_sriov(struct amdgpu_device *adev, · line 2597

Quoted 6 lines

goto error;

  	amdgpu_irq_gpu_reset_resume_helper(adev);
  	r = amdgpu_ib_ring_tests(adev);
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -2
-	if (r)
-		dev_err(adev->dev, "[GPU_RESET] ib ring test failed (%d).\n", r);

This 2 lines deletion seems unintentional

No, intentional, because if IB TEST failed the error will be reported by the fault IP

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c static int amdgpu_device_reset_sriov(struct amdgpu_device *adev, · line 2654

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +5 -0
+	if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
+		atomic_inc(&adev->vram_lost_counter);
+		r = amdgpu_handle_vram_lost(adev);
  	}
 
+error:
+
  	return r;

Not doing incrementation of VRAM lost counter in case of error anymore is intentional ?

[ML] yeah, no need to increase the counter because GPU reset is failed so everything is meaningless, driver cannot used anymore

-----Original Message-----
From: Grodzovsky, Andrey
Sent: 2018年2月28日 21:29
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 2/4] drm/amdgpu: cleanups for vram lost handling

Quoted 178 lines · Monk Liu

1)create a routine "handle_vram_lost" to do the vram recovery, and put it into amdgpu_device_reset/reset_sriov, this way no need of the extra paramter to hold the VRAM LOST information and the related macros can be removed.

3)show vram_recover failure if time out, and set TMO equal to lockup_timeout if vram_recover is under SRIOV runtime mode.

4)report error if any ip reset failed for SR-IOV

Change-Id I686e2b6133844c14948c206a2315c064a78c1d9c
---
  drivers/gpu/drm/amd/amdgpu/amdgpu.h        |   4 -
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 136 +++++++++++++++--------------
  2 files changed, 72 insertions(+), 68 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +0 -3
index 5bddfc1..abbc3f1 100644
@@ -181,10 +181,6 @@ extern int amdgpu_cik_support;
  #define CIK_CURSOR_WIDTH 128
  #define CIK_CURSOR_HEIGHT 128
 
-/* GPU RESET flags */
-#define AMDGPU_RESET_INFO_VRAM_LOST  (1 << 0) -#define 
 AMDGPU_RESET_INFO_FULLRESET  (1 << 1)
-
  struct amdgpu_device;
  struct amdgpu_ib;
  struct amdgpu_cs_parser;
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +2 -0
index e9d81a8..39ece7f 100644
@@ -1591,6 +1591,8 @@ static int 
 amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev)
 
  			r = block->version->funcs->hw_init(adev);
  			DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name, 
 r?"failed":"successed");
+			if (r)
+				return r;
  		}
  	}
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +2 -0
@@ -1624,6 +1626,8 @@ static int 
 amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev)
 
  			r = block->version->funcs->hw_init(adev);
  			DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name, 
 r?"failed":"successed");
+			if (r)
+				return r;
  		}
  	}
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +60 -15
@@ -2471,17 +2475,71 @@ static int amdgpu_device_recover_vram_from_shadow(struct amdgpu_device *adev,
  	return r;
  }
 
+static int amdgpu_handle_vram_lost(struct amdgpu_device *adev) {
+	struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
+	struct amdgpu_bo *bo, *tmp;
+	struct dma_fence *fence = NULL, *next = NULL;
+	long r = 1;
+	int i = 0;
+	long tmo;
+
+	if (amdgpu_sriov_runtime(adev))
+		tmo = msecs_to_jiffies(amdgpu_lockup_timeout);
+	else
+		tmo = msecs_to_jiffies(100);
+
+	DRM_INFO("recover vram bo from shadow start\n");
+	mutex_lock(&adev->shadow_list_lock);
+	list_for_each_entry_safe(bo, tmp, &adev->shadow_list, shadow_list) {
+		next = NULL;
+		amdgpu_device_recover_vram_from_shadow(adev, ring, bo, &next);
+		if (fence) {
+			r = dma_fence_wait_timeout(fence, false, tmo);
+			if (r == 0)
+				pr_err("wait fence %p[%d] timeout\n", fence, i);
+			else if (r < 0)
+				pr_err("wait fence %p[%d] interrupted\n", fence, i);
+			if (r < 1) {
+				dma_fence_put(fence);
+				fence = next;
+				break;
+			}
+			i++;
+		}
+
+		dma_fence_put(fence);
+		fence = next;
+	}
+	mutex_unlock(&adev->shadow_list_lock);
+
+	if (fence) {
+		r = dma_fence_wait_timeout(fence, false, tmo);
+		if (r == 0)
+			pr_err("wait fence %p[%d] timeout\n", fence, i);
+		else if (r < 0)
+			pr_err("wait fence %p[%d] interrupted\n", fence, i);
+
+	}
+	dma_fence_put(fence);
+
+	if (r > 0)
+		DRM_INFO("recover vram bo from shadow done\n");
+	else
+		DRM_ERROR("recover vram bo from shadow failed\n");
+
+	return (r > 0?0:1);
+}
+
  /*
   * amdgpu_device_reset - reset ASIC/GPU for bare-metal or passthrough
   *
   * @adev: amdgpu device pointer
- * @reset_flags: output param tells caller the reset result
   *
   * attempt to do soft-reset or full-reset and reinitialize Asic
   * return 0 means successed otherwise failed
  */
-static int amdgpu_device_reset(struct amdgpu_device *adev,
-			       uint64_t* reset_flags)
+static int amdgpu_device_reset(struct amdgpu_device *adev)
  {
  	bool need_full_reset, vram_lost = 0;
  	int r;
@@ -2496,7 +2554,6 @@ static int amdgpu_device_reset(struct amdgpu_device *adev,
  			DRM_INFO("soft reset failed, will fallback to full reset!\n");
  			need_full_reset = true;
  		}
-
  	}
 
  	if (need_full_reset) {
@@ -2545,13 +2602,8 @@ static int amdgpu_device_reset(struct amdgpu_device *adev,
  		}
  	}
 
-	if (reset_flags) {
-		if (vram_lost)
-			(*reset_flags) |= AMDGPU_RESET_INFO_VRAM_LOST;
-
-		if (need_full_reset)
-			(*reset_flags) |= AMDGPU_RESET_INFO_FULLRESET;
-	}
+	if (!r && ((need_full_reset && !(adev->flags & AMD_IS_APU)) || vram_lost))
+		r = amdgpu_handle_vram_lost(adev);
 
  	return r;
  }
@@ -2560,14 +2612,11 @@ static int amdgpu_device_reset(struct amdgpu_device *adev,
   * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf
   *
   * @adev: amdgpu device pointer
- * @reset_flags: output param tells caller the reset result
   *
   * do VF FLR and reinitialize Asic
   * return 0 means successed otherwise failed
  */
-static int amdgpu_device_reset_sriov(struct amdgpu_device *adev,
-				     uint64_t *reset_flags,
-				     bool from_hypervisor)
+static int amdgpu_device_reset_sriov(struct amdgpu_device *adev, bool 
+from_hypervisor)
  {
  	int r;
@@ -2588,28 +2637,20 @@ static int amdgpu_device_reset_sriov(struct 
 amdgpu_device *adev,
 

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c static int amdgpu_device_reset_sriov(struct · line 2598

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1 -2
  	/* now we are okay to resume SMC/CP/SDMA */
  	r = amdgpu_device_ip_reinit_late_sriov(adev);
+	amdgpu_virt_release_full_gpu(adev, true);
  	if (r)
  		goto error;
 
  	amdgpu_irq_gpu_reset_resume_helper(adev);
  	r = amdgpu_ib_ring_tests(adev);
-	if (r)
-		dev_err(adev->dev, "[GPU_RESET] ib ring test failed (%d).\n", r);

This 2 lines deletion seems unintentional

Quoted 11 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -10
-error:
-	/* release full control of GPU after ib test */
-	amdgpu_virt_release_full_gpu(adev, true);
-
-	if (reset_flags) {
-		if (adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
-			(*reset_flags) |= AMDGPU_RESET_INFO_VRAM_LOST;
-			atomic_inc(&adev->vram_lost_counter);
-		}
-

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c static int amdgpu_device_reset_sriov(struct amdgpu_device *adev, · line 2654

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +5 -2
-		/* VF FLR or hotlink reset is always full-reset */
-		(*reset_flags) |= AMDGPU_RESET_INFO_FULLRESET;
+	if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
+		atomic_inc(&adev->vram_lost_counter);
+		r = amdgpu_handle_vram_lost(adev);
  	}
 
+error:
+
  	return r;

Not doing incrementation of VRAM lost counter in case of error anymore is intentional ?

Andrey

Quoted 37 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1 -29
  }
 
@@ -2673,42 +2714,9 @@ int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
  	}
 
  	if (amdgpu_sriov_vf(adev))
-		r = amdgpu_device_reset_sriov(adev, &reset_flags, job ? false : true);
+		r = amdgpu_device_reset_sriov(adev, job ? false : true);
  	else
-		r = amdgpu_device_reset(adev, &reset_flags);
-
-	if (!r) {
-		if (((reset_flags & AMDGPU_RESET_INFO_FULLRESET) && !(adev->flags & AMD_IS_APU)) ||
-			(reset_flags & AMDGPU_RESET_INFO_VRAM_LOST)) {
-			struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
-			struct amdgpu_bo *bo, *tmp;
-			struct dma_fence *fence = NULL, *next = NULL;
-
-			DRM_INFO("recover vram bo from shadow\n");
-			mutex_lock(&adev->shadow_list_lock);
-			list_for_each_entry_safe(bo, tmp, &adev->shadow_list, shadow_list) {
-				next = NULL;
-				amdgpu_device_recover_vram_from_shadow(adev, ring, bo, &next);
-				if (fence) {
-					r = dma_fence_wait(fence, false);
-					if (r) {
-						WARN(r, "recovery from shadow isn't completed\n");
-						break;
-					}
-				}
-
-				dma_fence_put(fence);
-				fence = next;
-			}
-			mutex_unlock(&adev->shadow_list_lock);
-			if (fence) {
-				r = dma_fence_wait(fence, false);

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c int amdgpu_device_gpu_recover(struct amdgpu_device *adev, · line 2722

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1 -6
-				if (r)
-					WARN(r, "recovery from shadow isn't completed\n");
-			}
-			dma_fence_put(fence);
-		}
-	}
+		r = amdgpu_device_reset(adev);
 
  	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
  		struct amdgpu_ring *ring = adev->rings[i];

_______________________________________________
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): BLUPR12MB0449D9513F4798028569ED0884C70-7LeqcoF/[email protected]

#9 lore raw

Re: [PATCH 2/4] drm/amdgpu: cleanups for vram lost handling

Patch look good to me then.

Andrey

Quoted 294 lines · Monk

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c static int amdgpu_device_reset_sriov(struct amdgpu_device *adev, · line 2597

Quoted 6 lines
   		goto error;
   
   	amdgpu_irq_gpu_reset_resume_helper(adev);
   	r = amdgpu_ib_ring_tests(adev);
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -2
-	if (r)
-		dev_err(adev->dev, "[GPU_RESET] ib ring test failed (%d).\n", r);

This 2 lines deletion seems unintentional

No, intentional, because if IB TEST failed the error will be reported by the fault IP

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c static int amdgpu_device_reset_sriov(struct amdgpu_device *adev, · line 2654

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +5 -0
+	if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
+		atomic_inc(&adev->vram_lost_counter);
+		r = amdgpu_handle_vram_lost(adev);
   	}
 
+error:
+
   	return r;

Not doing incrementation of VRAM lost counter in case of error anymore is intentional ?

[ML] yeah, no need to increase the counter because GPU reset is failed so everything is meaningless, driver cannot used anymore

-----Original Message-----
From: Grodzovsky, Andrey
Sent: 2018年2月28日 21:29
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 2/4] drm/amdgpu: cleanups for vram lost handling

Quoted 178 lines · Monk Liu

1)create a routine "handle_vram_lost" to do the vram recovery, and put it into amdgpu_device_reset/reset_sriov, this way no need of the extra paramter to hold the VRAM LOST information and the related macros can be removed.

3)show vram_recover failure if time out, and set TMO equal to lockup_timeout if vram_recover is under SRIOV runtime mode.

4)report error if any ip reset failed for SR-IOV

Change-Id I686e2b6133844c14948c206a2315c064a78c1d9c
---
   drivers/gpu/drm/amd/amdgpu/amdgpu.h        |   4 -
   drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 136 +++++++++++++++--------------
   2 files changed, 72 insertions(+), 68 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +0 -3
index 5bddfc1..abbc3f1 100644
@@ -181,10 +181,6 @@ extern int amdgpu_cik_support;
   #define CIK_CURSOR_WIDTH 128
   #define CIK_CURSOR_HEIGHT 128
 
-/* GPU RESET flags */
-#define AMDGPU_RESET_INFO_VRAM_LOST  (1 << 0) -#define
 AMDGPU_RESET_INFO_FULLRESET  (1 << 1)
-
   struct amdgpu_device;
   struct amdgpu_ib;
   struct amdgpu_cs_parser;
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +2 -0
index e9d81a8..39ece7f 100644
@@ -1591,6 +1591,8 @@ static int
 amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev)
 
   			r = block->version->funcs->hw_init(adev);
   			DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name,
 r?"failed":"successed");
+			if (r)
+				return r;
   		}
   	}
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +2 -0
@@ -1624,6 +1626,8 @@ static int
 amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev)
 
   			r = block->version->funcs->hw_init(adev);
   			DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name,
 r?"failed":"successed");
+			if (r)
+				return r;
   		}
   	}
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +60 -15
@@ -2471,17 +2475,71 @@ static int amdgpu_device_recover_vram_from_shadow(struct amdgpu_device *adev,
   	return r;
   }
 
+static int amdgpu_handle_vram_lost(struct amdgpu_device *adev) {
+	struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
+	struct amdgpu_bo *bo, *tmp;
+	struct dma_fence *fence = NULL, *next = NULL;
+	long r = 1;
+	int i = 0;
+	long tmo;
+
+	if (amdgpu_sriov_runtime(adev))
+		tmo = msecs_to_jiffies(amdgpu_lockup_timeout);
+	else
+		tmo = msecs_to_jiffies(100);
+
+	DRM_INFO("recover vram bo from shadow start\n");
+	mutex_lock(&adev->shadow_list_lock);
+	list_for_each_entry_safe(bo, tmp, &adev->shadow_list, shadow_list) {
+		next = NULL;
+		amdgpu_device_recover_vram_from_shadow(adev, ring, bo, &next);
+		if (fence) {
+			r = dma_fence_wait_timeout(fence, false, tmo);
+			if (r == 0)
+				pr_err("wait fence %p[%d] timeout\n", fence, i);
+			else if (r < 0)
+				pr_err("wait fence %p[%d] interrupted\n", fence, i);
+			if (r < 1) {
+				dma_fence_put(fence);
+				fence = next;
+				break;
+			}
+			i++;
+		}
+
+		dma_fence_put(fence);
+		fence = next;
+	}
+	mutex_unlock(&adev->shadow_list_lock);
+
+	if (fence) {
+		r = dma_fence_wait_timeout(fence, false, tmo);
+		if (r == 0)
+			pr_err("wait fence %p[%d] timeout\n", fence, i);
+		else if (r < 0)
+			pr_err("wait fence %p[%d] interrupted\n", fence, i);
+
+	}
+	dma_fence_put(fence);
+
+	if (r > 0)
+		DRM_INFO("recover vram bo from shadow done\n");
+	else
+		DRM_ERROR("recover vram bo from shadow failed\n");
+
+	return (r > 0?0:1);
+}
+
   /*
    * amdgpu_device_reset - reset ASIC/GPU for bare-metal or passthrough
    *
    * @adev: amdgpu device pointer
- * @reset_flags: output param tells caller the reset result
    *
    * attempt to do soft-reset or full-reset and reinitialize Asic
    * return 0 means successed otherwise failed
   */
-static int amdgpu_device_reset(struct amdgpu_device *adev,
-			       uint64_t* reset_flags)
+static int amdgpu_device_reset(struct amdgpu_device *adev)
   {
   	bool need_full_reset, vram_lost = 0;
   	int r;
@@ -2496,7 +2554,6 @@ static int amdgpu_device_reset(struct amdgpu_device *adev,
   			DRM_INFO("soft reset failed, will fallback to full reset!\n");
   			need_full_reset = true;
   		}
-
   	}
 
   	if (need_full_reset) {
@@ -2545,13 +2602,8 @@ static int amdgpu_device_reset(struct amdgpu_device *adev,
   		}
   	}
 
-	if (reset_flags) {
-		if (vram_lost)
-			(*reset_flags) |= AMDGPU_RESET_INFO_VRAM_LOST;
-
-		if (need_full_reset)
-			(*reset_flags) |= AMDGPU_RESET_INFO_FULLRESET;
-	}
+	if (!r && ((need_full_reset && !(adev->flags & AMD_IS_APU)) || vram_lost))
+		r = amdgpu_handle_vram_lost(adev);
 
   	return r;
   }
@@ -2560,14 +2612,11 @@ static int amdgpu_device_reset(struct amdgpu_device *adev,
    * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf
    *
    * @adev: amdgpu device pointer
- * @reset_flags: output param tells caller the reset result
    *
    * do VF FLR and reinitialize Asic
    * return 0 means successed otherwise failed
   */
-static int amdgpu_device_reset_sriov(struct amdgpu_device *adev,
-				     uint64_t *reset_flags,
-				     bool from_hypervisor)
+static int amdgpu_device_reset_sriov(struct amdgpu_device *adev, bool
+from_hypervisor)
   {
   	int r;
@@ -2588,28 +2637,20 @@ static int amdgpu_device_reset_sriov(struct
 amdgpu_device *adev,
 

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c static int amdgpu_device_reset_sriov(struct · line 2598

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1 -2
   	/* now we are okay to resume SMC/CP/SDMA */
   	r = amdgpu_device_ip_reinit_late_sriov(adev);
+	amdgpu_virt_release_full_gpu(adev, true);
   	if (r)
   		goto error;
 
   	amdgpu_irq_gpu_reset_resume_helper(adev);
   	r = amdgpu_ib_ring_tests(adev);
-	if (r)
-		dev_err(adev->dev, "[GPU_RESET] ib ring test failed (%d).\n", r);

This 2 lines deletion seems unintentional

Quoted 11 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -10
-error:
-	/* release full control of GPU after ib test */
-	amdgpu_virt_release_full_gpu(adev, true);
-
-	if (reset_flags) {
-		if (adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
-			(*reset_flags) |= AMDGPU_RESET_INFO_VRAM_LOST;
-			atomic_inc(&adev->vram_lost_counter);
-		}
-

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c static int amdgpu_device_reset_sriov(struct amdgpu_device *adev, · line 2654

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +5 -2
-		/* VF FLR or hotlink reset is always full-reset */
-		(*reset_flags) |= AMDGPU_RESET_INFO_FULLRESET;
+	if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
+		atomic_inc(&adev->vram_lost_counter);
+		r = amdgpu_handle_vram_lost(adev);
   	}
 
+error:
+
   	return r;

Not doing incrementation of VRAM lost counter in case of error anymore is intentional ?

Andrey

Quoted 45 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +2 -35
   }
 
@@ -2673,42 +2714,9 @@ int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
   	}
 
   	if (amdgpu_sriov_vf(adev))
-		r = amdgpu_device_reset_sriov(adev, &reset_flags, job ? false : true);
+		r = amdgpu_device_reset_sriov(adev, job ? false : true);
   	else
-		r = amdgpu_device_reset(adev, &reset_flags);
-
-	if (!r) {
-		if (((reset_flags & AMDGPU_RESET_INFO_FULLRESET) && !(adev->flags & AMD_IS_APU)) ||
-			(reset_flags & AMDGPU_RESET_INFO_VRAM_LOST)) {
-			struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
-			struct amdgpu_bo *bo, *tmp;
-			struct dma_fence *fence = NULL, *next = NULL;
-
-			DRM_INFO("recover vram bo from shadow\n");
-			mutex_lock(&adev->shadow_list_lock);
-			list_for_each_entry_safe(bo, tmp, &adev->shadow_list, shadow_list) {
-				next = NULL;
-				amdgpu_device_recover_vram_from_shadow(adev, ring, bo, &next);
-				if (fence) {
-					r = dma_fence_wait(fence, false);
-					if (r) {
-						WARN(r, "recovery from shadow isn't completed\n");
-						break;
-					}
-				}
-
-				dma_fence_put(fence);
-				fence = next;
-			}
-			mutex_unlock(&adev->shadow_list_lock);
-			if (fence) {
-				r = dma_fence_wait(fence, false);
-				if (r)
-					WARN(r, "recovery from shadow isn't completed\n");
-			}
-			dma_fence_put(fence);
-		}
-	}
+		r = amdgpu_device_reset(adev);
 
Quoted 2 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c int amdgpu_device_gpu_recover(struct amdgpu_device *adev, · line 2722

Quoted 2 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
   	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
   		struct amdgpu_ring *ring = adev->rings[i];

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

#10 lore raw

Re: [PATCH 2/4] drm/amdgpu: cleanups for vram lost handling

In reply to a message not in the archive
Quoted 55 lines · Monk Liu

1)create a routine "handle_vram_lost" to do the vram recovery, and put it into amdgpu_device_reset/reset_sriov, this way no need of the extra paramter to hold the VRAM LOST information and the related macros can be removed.

3)show vram_recover failure if time out, and set TMO equal to lockup_timeout if vram_recover is under SRIOV runtime mode.

4)report error if any ip reset failed for SR-IOV

Change-Id I686e2b6133844c14948c206a2315c064a78c1d9c
---
 drivers/gpu/drm/amd/amdgpu/amdgpu.h        |   4 -
 drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 136 +++++++++++++++--------------
 2 files changed, 72 insertions(+), 68 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +0 -4
index 5bddfc1..abbc3f1 100644
@@ -181,10 +181,6 @@ extern int amdgpu_cik_support;
 #define CIK_CURSOR_WIDTH 128
 #define CIK_CURSOR_HEIGHT 128
 
-/* GPU RESET flags */
-#define AMDGPU_RESET_INFO_VRAM_LOST  (1 << 0)
-#define AMDGPU_RESET_INFO_FULLRESET  (1 << 1)
-
 struct amdgpu_device;
 struct amdgpu_ib;
 struct amdgpu_cs_parser;
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +4 -0
index e9d81a8..39ece7f 100644
@@ -1591,6 +1591,8 @@ static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev)
 
                        r = block->version->funcs->hw_init(adev);
                        DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name, r?"failed":"successed");
+                       if (r)
+                               return r;
                }
        }
 
@@ -1624,6 +1626,8 @@ static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev)
 
                        r = block->version->funcs->hw_init(adev);
                        DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name, r?"failed":"successed");
+                       if (r)
+                               return r;
                }
        }
 

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c static int amdgpu_device_recover_vram_from_shadow(struct amdgpu_device *adev, · line 2478

Quoted 4 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1 -0
@@ -2471,17 +2475,71 @@ static int amdgpu_device_recover_vram_from_shadow(struct amdgpu_device *adev,
        return r;
 }
 
+static int amdgpu_handle_vram_lost(struct amdgpu_device *adev)

Please rename this function to amdgpu_device_handle_vram_lost() for consistency.with other functions in the file.

Alex

Quoted 193 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +67 -65
+{
+       struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
+       struct amdgpu_bo *bo, *tmp;
+       struct dma_fence *fence = NULL, *next = NULL;
+       long r = 1;
+       int i = 0;
+       long tmo;
+
+       if (amdgpu_sriov_runtime(adev))
+               tmo = msecs_to_jiffies(amdgpu_lockup_timeout);
+       else
+               tmo = msecs_to_jiffies(100);
+
+       DRM_INFO("recover vram bo from shadow start\n");
+       mutex_lock(&adev->shadow_list_lock);
+       list_for_each_entry_safe(bo, tmp, &adev->shadow_list, shadow_list) {
+               next = NULL;
+               amdgpu_device_recover_vram_from_shadow(adev, ring, bo, &next);
+               if (fence) {
+                       r = dma_fence_wait_timeout(fence, false, tmo);
+                       if (r == 0)
+                               pr_err("wait fence %p[%d] timeout\n", fence, i);
+                       else if (r < 0)
+                               pr_err("wait fence %p[%d] interrupted\n", fence, i);
+                       if (r < 1) {
+                               dma_fence_put(fence);
+                               fence = next;
+                               break;
+                       }
+                       i++;
+               }
+
+               dma_fence_put(fence);
+               fence = next;
+       }
+       mutex_unlock(&adev->shadow_list_lock);
+
+       if (fence) {
+               r = dma_fence_wait_timeout(fence, false, tmo);
+               if (r == 0)
+                       pr_err("wait fence %p[%d] timeout\n", fence, i);
+               else if (r < 0)
+                       pr_err("wait fence %p[%d] interrupted\n", fence, i);
+
+       }
+       dma_fence_put(fence);
+
+       if (r > 0)
+               DRM_INFO("recover vram bo from shadow done\n");
+       else
+               DRM_ERROR("recover vram bo from shadow failed\n");
+
+       return (r > 0?0:1);
+}
+
 /*
  * amdgpu_device_reset - reset ASIC/GPU for bare-metal or passthrough
  *
  * @adev: amdgpu device pointer
- * @reset_flags: output param tells caller the reset result
  *
  * attempt to do soft-reset or full-reset and reinitialize Asic
  * return 0 means successed otherwise failed
 */
-static int amdgpu_device_reset(struct amdgpu_device *adev,
-                              uint64_t* reset_flags)
+static int amdgpu_device_reset(struct amdgpu_device *adev)
 {
        bool need_full_reset, vram_lost = 0;
        int r;
@@ -2496,7 +2554,6 @@ static int amdgpu_device_reset(struct amdgpu_device *adev,
                        DRM_INFO("soft reset failed, will fallback to full reset!\n");
                        need_full_reset = true;
                }
-
        }
 
        if (need_full_reset) {
@@ -2545,13 +2602,8 @@ static int amdgpu_device_reset(struct amdgpu_device *adev,
                }
        }
 
-       if (reset_flags) {
-               if (vram_lost)
-                       (*reset_flags) |= AMDGPU_RESET_INFO_VRAM_LOST;
-
-               if (need_full_reset)
-                       (*reset_flags) |= AMDGPU_RESET_INFO_FULLRESET;
-       }
+       if (!r && ((need_full_reset && !(adev->flags & AMD_IS_APU)) || vram_lost))
+               r = amdgpu_handle_vram_lost(adev);
 
        return r;
 }
@@ -2560,14 +2612,11 @@ static int amdgpu_device_reset(struct amdgpu_device *adev,
  * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf
  *
  * @adev: amdgpu device pointer
- * @reset_flags: output param tells caller the reset result
  *
  * do VF FLR and reinitialize Asic
  * return 0 means successed otherwise failed
 */
-static int amdgpu_device_reset_sriov(struct amdgpu_device *adev,
-                                    uint64_t *reset_flags,
-                                    bool from_hypervisor)
+static int amdgpu_device_reset_sriov(struct amdgpu_device *adev, bool from_hypervisor)
 {
        int r;
 
@@ -2588,28 +2637,20 @@ static int amdgpu_device_reset_sriov(struct amdgpu_device *adev,
 
        /* now we are okay to resume SMC/CP/SDMA */
        r = amdgpu_device_ip_reinit_late_sriov(adev);
+       amdgpu_virt_release_full_gpu(adev, true);
        if (r)
                goto error;
 
        amdgpu_irq_gpu_reset_resume_helper(adev);
        r = amdgpu_ib_ring_tests(adev);
-       if (r)
-               dev_err(adev->dev, "[GPU_RESET] ib ring test failed (%d).\n", r);
 
-error:
-       /* release full control of GPU after ib test */
-       amdgpu_virt_release_full_gpu(adev, true);
-
-       if (reset_flags) {
-               if (adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
-                       (*reset_flags) |= AMDGPU_RESET_INFO_VRAM_LOST;
-                       atomic_inc(&adev->vram_lost_counter);
-               }
-
-               /* VF FLR or hotlink reset is always full-reset */
-               (*reset_flags) |= AMDGPU_RESET_INFO_FULLRESET;
+       if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
+               atomic_inc(&adev->vram_lost_counter);
+               r = amdgpu_handle_vram_lost(adev);
        }
 
+error:
+
        return r;
 }
 
@@ -2673,42 +2714,9 @@ int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
        }
 
        if (amdgpu_sriov_vf(adev))
-               r = amdgpu_device_reset_sriov(adev, &reset_flags, job ? false : true);
+               r = amdgpu_device_reset_sriov(adev, job ? false : true);
        else
-               r = amdgpu_device_reset(adev, &reset_flags);
-
-       if (!r) {
-               if (((reset_flags & AMDGPU_RESET_INFO_FULLRESET) && !(adev->flags & AMD_IS_APU)) ||
-                       (reset_flags & AMDGPU_RESET_INFO_VRAM_LOST)) {
-                       struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
-                       struct amdgpu_bo *bo, *tmp;
-                       struct dma_fence *fence = NULL, *next = NULL;
-
-                       DRM_INFO("recover vram bo from shadow\n");
-                       mutex_lock(&adev->shadow_list_lock);
-                       list_for_each_entry_safe(bo, tmp, &adev->shadow_list, shadow_list) {
-                               next = NULL;
-                               amdgpu_device_recover_vram_from_shadow(adev, ring, bo, &next);
-                               if (fence) {
-                                       r = dma_fence_wait(fence, false);
-                                       if (r) {
-                                               WARN(r, "recovery from shadow isn't completed\n");
-                                               break;
-                                       }
-                               }
-
-                               dma_fence_put(fence);
-                               fence = next;
-                       }
-                       mutex_unlock(&adev->shadow_list_lock);
-                       if (fence) {
-                               r = dma_fence_wait(fence, false);
-                               if (r)
-                                       WARN(r, "recovery from shadow isn't completed\n");
-                       }
-                       dma_fence_put(fence);
-               }
-       }
+               r = amdgpu_device_reset(adev);
 
        for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
                struct amdgpu_ring *ring = adev->rings[i];
--

2.7.4

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

author
#11 lore raw

PATCH 3/4 drm/amdgpu: don't return when ring not ready for fill_buffer

In reply to a message not in the archive

because this time SDMA may under GPU RESET so its ring->ready may not true, keep going and GPU scheduler will reschedule this job if it failed.

give a warning on copy_buffer when go through direct_submit while ring->ready is false

Change-Id Ife6cd55e0e843d99900e5bed5418499e88633685
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 6 +-----
 1 file changed, 1 insertion(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +1 -5
index e38e6db..7b75ac9 100644
@@ -1656,6 +1656,7 @@ int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
 	amdgpu_ring_pad_ib(ring, &job->ibs[0]);
 	WARN_ON(job->ibs[0].length_dw > num_dw);
 	if (direct_submit) {
+		WARN_ON(!ring->ready);
 		r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs,
 				       NULL, fence);
 		job->fence = dma_fence_get(*fence);
@@ -1692,11 +1693,6 @@ int amdgpu_fill_buffer(struct amdgpu_bo *bo,
 	struct amdgpu_job *job;
 	int r;
 
-	if (!ring->ready) {
-		DRM_ERROR("Trying to clear memory with ring turned off.\n");
-		return -EINVAL;
-	}
-
 	if (bo->tbo.mem.mem_type == TTM_PL_TT) {
 		r = amdgpu_ttm_alloc_gart(&bo->tbo);
 		if (r)
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): 1519802463-9090-3-git-send-email-Monk.Liu-5C7GfCeVMHo@public.gmane.org

#13 lore raw

Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Am 28.02.2018 um 08:21 schrieb Monk Liu:

Quoted 6 lines

because this time SDMA may under GPU RESET so its ring->ready may not true, keep going and GPU scheduler will reschedule this job if it failed.

give a warning on copy_buffer when go through direct_submit while ring->ready is false

NAK, that test has already saved us quite a bunch of trouble with the fb layer.

Why exactly are you running into issues with that?

Christian.

Quoted 21 lines
Change-Id Ife6cd55e0e843d99900e5bed5418499e88633685
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 6 +-----
  1 file changed, 1 insertion(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +1 -0
index e38e6db..7b75ac9 100644
@@ -1656,6 +1656,7 @@ int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
  	amdgpu_ring_pad_ib(ring, &job->ibs[0]);
  	WARN_ON(job->ibs[0].length_dw > num_dw);
  	if (direct_submit) {
+		WARN_ON(!ring->ready);
  		r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs,
  				       NULL, fence);
  		job->fence = dma_fence_get(*fence);
@@ -1692,11 +1693,6 @@ int amdgpu_fill_buffer(struct amdgpu_bo *bo,
  	struct amdgpu_job *job;

drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c int amdgpu_fill_buffer(struct amdgpu_bo *bo, · line 1698

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +0 -5
  	int r;
 
-	if (!ring->ready) {
-		DRM_ERROR("Trying to clear memory with ring turned off.\n");
-		return -EINVAL;
-	}
-
  	if (bo->tbo.mem.mem_type == TTM_PL_TT) {
  		r = amdgpu_ttm_alloc_gart(&bo->tbo);
  		if (r)

_______________________________________________
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): 01879d0d-edea-680e-c9f2-1005d94f1dfd-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org

author NAK
#15 lore raw

RE: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Because when SDMA was hang by like process A, and meanwhile another process B is already running into the code of fill_buffer() So just let process B continue, don't block it otherwise process B would fail by software reason .

Let it run and finally process B's job would fail and GPU recover will repeat it again (since it is a kernel job)

Without this solution other process will be greatly harmed by one black sheep that triggering GPU recover

/Monk

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: 2018年2月28日 20:24
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Am 28.02.2018 um 08:21 schrieb Monk Liu:

Quoted 6 lines

because this time SDMA may under GPU RESET so its ring->ready may not true, keep going and GPU scheduler will reschedule this job if it failed.

give a warning on copy_buffer when go through direct_submit while ring->ready is false

NAK, that test has already saved us quite a bunch of trouble with the fb layer.

Why exactly are you running into issues with that?

Christian.

Quoted 24 lines
Change-Id Ife6cd55e0e843d99900e5bed5418499e88633685
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 6 +-----
  1 file changed, 1 insertion(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +1 -0
index e38e6db..7b75ac9 100644
@@ -1656,6 +1656,7 @@ int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
  	amdgpu_ring_pad_ib(ring, &job->ibs[0]);
  	WARN_ON(job->ibs[0].length_dw > num_dw);
  	if (direct_submit) {
+		WARN_ON(!ring->ready);
  		r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs,
  				       NULL, fence);
  		job->fence = dma_fence_get(*fence); @@ -1692,11 +1693,6 @@ int 

amdgpu_fill_buffer(struct amdgpu_bo *bo,
struct amdgpu_job *job;
int r;

drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c int amdgpu_fill_buffer(struct amdgpu_bo *bo, · line 1698

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +0 -5
-	if (!ring->ready) {
-		DRM_ERROR("Trying to clear memory with ring turned off.\n");
-		return -EINVAL;
-	}
-
  	if (bo->tbo.mem.mem_type == TTM_PL_TT) {
  		r = amdgpu_ttm_alloc_gart(&bo->tbo);
  		if (r)

_______________________________________________
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): BLUPR12MB0449BEDBC1FCE72C0021C88184C70-7LeqcoF/[email protected]

author NAK
#17 lore raw

RE: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

What about this for fill_buffer():

 if (ring->ready != true && !adev->in_gpu_reset) {
	return -EINVAL;

}

/Monk

-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Liu, Monk
Sent: 2018年2月28日 20:35
To: Koenig, Christian <[email protected]>; [email protected]
Subject: RE: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Because when SDMA was hang by like process A, and meanwhile another process B is already running into the code of fill_buffer() So just let process B continue, don't block it otherwise process B would fail by software reason .

Let it run and finally process B's job would fail and GPU recover will repeat it again (since it is a kernel job)

Without this solution other process will be greatly harmed by one black sheep that triggering GPU recover

/Monk

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: 2018年2月28日 20:24
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Am 28.02.2018 um 08:21 schrieb Monk Liu:

Quoted 6 lines

because this time SDMA may under GPU RESET so its ring->ready may not true, keep going and GPU scheduler will reschedule this job if it failed.

give a warning on copy_buffer when go through direct_submit while ring->ready is false

NAK, that test has already saved us quite a bunch of trouble with the fb layer.

Why exactly are you running into issues with that?

Christian.

Quoted 24 lines
Change-Id Ife6cd55e0e843d99900e5bed5418499e88633685
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 6 +-----
  1 file changed, 1 insertion(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +1 -0
index e38e6db..7b75ac9 100644
@@ -1656,6 +1656,7 @@ int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
  	amdgpu_ring_pad_ib(ring, &job->ibs[0]);
  	WARN_ON(job->ibs[0].length_dw > num_dw);
  	if (direct_submit) {
+		WARN_ON(!ring->ready);
  		r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs,
  				       NULL, fence);
  		job->fence = dma_fence_get(*fence); @@ -1692,11 +1693,6 @@ int 

amdgpu_fill_buffer(struct amdgpu_bo *bo,
struct amdgpu_job *job;
int r;

drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c int amdgpu_fill_buffer(struct amdgpu_bo *bo, · line 1698

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +0 -5
-	if (!ring->ready) {
-		DRM_ERROR("Trying to clear memory with ring turned off.\n");
-		return -EINVAL;
-	}
-
  	if (bo->tbo.mem.mem_type == TTM_PL_TT) {
  		r = amdgpu_ttm_alloc_gart(&bo->tbo);
  		if (r)

_______________________________________________
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

#18 lore raw

Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

In reply to a message not in the archive

Good point, but in this case we need some other handling.

Please change the test to use ring->ring_obj instead, this way we still bail out if somebody tries to submit commands before the ring is even allocated.

And you also need to fix a couple of other places in amdgpu_ttm.c.

Regards,
Christian.

Am 28.02.2018 um 13:34 schrieb Liu, Monk:

Quoted 53 lines

Because when SDMA was hang by like process A, and meanwhile another process B is already running into the code of fill_buffer() So just let process B continue, don't block it otherwise process B would fail by software reason .

Let it run and finally process B's job would fail and GPU recover will repeat it again (since it is a kernel job)

Without this solution other process will be greatly harmed by one black sheep that triggering GPU recover

/Monk

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: 2018年2月28日 20:24
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Am 28.02.2018 um 08:21 schrieb Monk Liu:

Quoted 6 lines

because this time SDMA may under GPU RESET so its ring->ready may not true, keep going and GPU scheduler will reschedule this job if it failed.

give a warning on copy_buffer when go through direct_submit while ring->ready is false

NAK, that test has already saved us quite a bunch of trouble with the fb layer.

Why exactly are you running into issues with that?

Christian.

Quoted 23 lines
Change-Id Ife6cd55e0e843d99900e5bed5418499e88633685
---
   drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 6 +-----
   1 file changed, 1 insertion(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +1 -0
index e38e6db..7b75ac9 100644
@@ -1656,6 +1656,7 @@ int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
   	amdgpu_ring_pad_ib(ring, &job->ibs[0]);
   	WARN_ON(job->ibs[0].length_dw > num_dw);
   	if (direct_submit) {
+		WARN_ON(!ring->ready);
   		r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs,
   				       NULL, fence);
   		job->fence = dma_fence_get(*fence); @@ -1692,11 +1693,6 @@ int

amdgpu_fill_buffer(struct amdgpu_bo *bo,

   	struct amdgpu_job *job;
   	int r;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c int amdgpu_fill_buffer(struct amdgpu_bo *bo, · line 1698

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +0 -5
-	if (!ring->ready) {
-		DRM_ERROR("Trying to clear memory with ring turned off.\n");
-		return -EINVAL;
-	}
-
   	if (bo->tbo.mem.mem_type == TTM_PL_TT) {
   		r = amdgpu_ttm_alloc_gart(&bo->tbo);
   		if (r)

_______________________________________________
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): 3a8945f9-848e-dd19-373d-5dddc69f76cb-5C7GfCeVMHo@public.gmane.org

author
#20 lore raw

RE: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Quoted 1 line

Please change the test to use ring->ring_obj instead, this way we still bail out if somebody tries to submit commands before the ring is even allocated.

I don't understand how could fill_buffer() get run under the case that ring->ring_obj is not even allocated ... where is such case ?

/Monk

-----Original Message-----
From: Koenig, Christian
Sent: 2018年2月28日 20:46
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Good point, but in this case we need some other handling.

Please change the test to use ring->ring_obj instead, this way we still bail out if somebody tries to submit commands before the ring is even allocated.

And you also need to fix a couple of other places in amdgpu_ttm.c.

Regards,
Christian.

Am 28.02.2018 um 13:34 schrieb Liu, Monk:

Quoted 56 lines

Because when SDMA was hang by like process A, and meanwhile another
process B is already running into the code of fill_buffer() So just let process B continue, don't block it otherwise process B would fail by software reason .

Let it run and finally process B's job would fail and GPU recover will repeat it again (since it is a kernel job)

Without this solution other process will be greatly harmed by one black sheep that triggering GPU recover

/Monk

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: 2018年2月28日 20:24
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Am 28.02.2018 um 08:21 schrieb Monk Liu:

Quoted 6 lines

because this time SDMA may under GPU RESET so its ring->ready may not true, keep going and GPU scheduler will reschedule this job if it failed.

give a warning on copy_buffer when go through direct_submit while ring->ready is false

NAK, that test has already saved us quite a bunch of trouble with the fb layer.

Why exactly are you running into issues with that?

Christian.

Quoted 23 lines
Change-Id Ife6cd55e0e843d99900e5bed5418499e88633685
---
   drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 6 +-----
   1 file changed, 1 insertion(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +1 -0
index e38e6db..7b75ac9 100644
@@ -1656,6 +1656,7 @@ int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
   	amdgpu_ring_pad_ib(ring, &job->ibs[0]);
   	WARN_ON(job->ibs[0].length_dw > num_dw);
   	if (direct_submit) {
+		WARN_ON(!ring->ready);
   		r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs,
   				       NULL, fence);
   		job->fence = dma_fence_get(*fence); @@ -1692,11 +1693,6 @@ int 

amdgpu_fill_buffer(struct amdgpu_bo *bo,

   	struct amdgpu_job *job;
   	int r;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +0 -5
-	if (!ring->ready) {
-		DRM_ERROR("Trying to clear memory with ring turned off.\n");
-		return -EINVAL;
-	}
-
   	if (bo->tbo.mem.mem_type == TTM_PL_TT) {
   		r = amdgpu_ttm_alloc_gart(&bo->tbo);
   		if (r)

_______________________________________________
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): BLUPR12MB04498FBA36C747652758D56E84C60-7LeqcoF/[email protected]

#22 lore raw

Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Ah, crap yeah that won't work since we don't free the ring.

Key point is we need to distinct between the ring doesn't work temporary because we are in a GPU reset and it doesn't work at all because we are missing firmware or stuff like that.

And no, checking the gpu_reset flag is totally racy and can't be done either. How about checking accel_working instead?

Christian.

Am 01.03.2018 um 07:01 schrieb Liu, Monk:

Quoted 78 lines
Quoted 1 line

Please change the test to use ring->ring_obj instead, this way we still bail out if somebody tries to submit commands before the ring is even allocated.

I don't understand how could fill_buffer() get run under the case that ring->ring_obj is not even allocated ... where is such case ?

/Monk

-----Original Message-----
From: Koenig, Christian
Sent: 2018年2月28日 20:46
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Good point, but in this case we need some other handling.

Please change the test to use ring->ring_obj instead, this way we still bail out if somebody tries to submit commands before the ring is even allocated.

And you also need to fix a couple of other places in amdgpu_ttm.c.

Regards,
Christian.

Am 28.02.2018 um 13:34 schrieb Liu, Monk:

Quoted 56 lines

Because when SDMA was hang by like process A, and meanwhile another
process B is already running into the code of fill_buffer() So just let process B continue, don't block it otherwise process B would fail by software reason .

Let it run and finally process B's job would fail and GPU recover will repeat it again (since it is a kernel job)

Without this solution other process will be greatly harmed by one black sheep that triggering GPU recover

/Monk

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: 2018年2月28日 20:24
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Am 28.02.2018 um 08:21 schrieb Monk Liu:

Quoted 6 lines

because this time SDMA may under GPU RESET so its ring->ready may not true, keep going and GPU scheduler will reschedule this job if it failed.

give a warning on copy_buffer when go through direct_submit while ring->ready is false

NAK, that test has already saved us quite a bunch of trouble with the fb layer.

Why exactly are you running into issues with that?

Christian.

Quoted 23 lines
Change-Id Ife6cd55e0e843d99900e5bed5418499e88633685
---
    drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 6 +-----
    1 file changed, 1 insertion(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +1 -0
index e38e6db..7b75ac9 100644
@@ -1656,6 +1656,7 @@ int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
    	amdgpu_ring_pad_ib(ring, &job->ibs[0]);
    	WARN_ON(job->ibs[0].length_dw > num_dw);
    	if (direct_submit) {
+		WARN_ON(!ring->ready);
    		r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs,
    				       NULL, fence);
    		job->fence = dma_fence_get(*fence); @@ -1692,11 +1693,6 @@ int

amdgpu_fill_buffer(struct amdgpu_bo *bo,

    	struct amdgpu_job *job;
    	int r;
Quoted 12 lines
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +0 -5
-	if (!ring->ready) {
-		DRM_ERROR("Trying to clear memory with ring turned off.\n");
-		return -EINVAL;
-	}
-
    	if (bo->tbo.mem.mem_type == TTM_PL_TT) {
    		r = amdgpu_ttm_alloc_gart(&bo->tbo);
    		if (r)

_______________________________________________
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): 7af9b6fb-28e8-4e25-5d4a-5b566a00cbea-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org

author
#24 lore raw

RE: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Accel_working is only for GFX ring, I don't think test it is appropriate for SDMA jobs ....

If we are in gpu rest stage, we can let SDMA job live with the case of ring not ready, can you point out where is the problem ?

In fact we did stress TDR test with this patch and it can really fix couple over kill issues

/Monk

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: 2018年3月1日 16:16
To: Liu, Monk <[email protected]>; Koenig, Christian <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Ah, crap yeah that won't work since we don't free the ring.

Key point is we need to distinct between the ring doesn't work temporary because we are in a GPU reset and it doesn't work at all because we are missing firmware or stuff like that.

And no, checking the gpu_reset flag is totally racy and can't be done either. How about checking accel_working instead?

Christian.

Am 01.03.2018 um 07:01 schrieb Liu, Monk:

Quoted 79 lines
Quoted 1 line

Please change the test to use ring->ring_obj instead, this way we still bail out if somebody tries to submit commands before the ring is even allocated.

I don't understand how could fill_buffer() get run under the case that ring->ring_obj is not even allocated ... where is such case ?

/Monk

-----Original Message-----
From: Koenig, Christian
Sent: 2018年2月28日 20:46
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Good point, but in this case we need some other handling.

Please change the test to use ring->ring_obj instead, this way we still bail out if somebody tries to submit commands before the ring is even allocated.

And you also need to fix a couple of other places in amdgpu_ttm.c.

Regards,
Christian.

Am 28.02.2018 um 13:34 schrieb Liu, Monk:

Quoted 56 lines

Because when SDMA was hang by like process A, and meanwhile another
process B is already running into the code of fill_buffer() So just let process B continue, don't block it otherwise process B would fail by software reason .

Let it run and finally process B's job would fail and GPU recover will repeat it again (since it is a kernel job)

Without this solution other process will be greatly harmed by one black sheep that triggering GPU recover

/Monk

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: 2018年2月28日 20:24
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Am 28.02.2018 um 08:21 schrieb Monk Liu:

Quoted 6 lines

because this time SDMA may under GPU RESET so its ring->ready may not true, keep going and GPU scheduler will reschedule this job if it failed.

give a warning on copy_buffer when go through direct_submit while ring->ready is false

NAK, that test has already saved us quite a bunch of trouble with the fb layer.

Why exactly are you running into issues with that?

Christian.

Quoted 23 lines
Change-Id Ife6cd55e0e843d99900e5bed5418499e88633685
---
    drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 6 +-----
    1 file changed, 1 insertion(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +1 -0
index e38e6db..7b75ac9 100644
@@ -1656,6 +1656,7 @@ int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
    	amdgpu_ring_pad_ib(ring, &job->ibs[0]);
    	WARN_ON(job->ibs[0].length_dw > num_dw);
    	if (direct_submit) {
+		WARN_ON(!ring->ready);
    		r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs,
    				       NULL, fence);
    		job->fence = dma_fence_get(*fence); @@ -1692,11 +1693,6 @@ int 

amdgpu_fill_buffer(struct amdgpu_bo *bo,

    	struct amdgpu_job *job;
    	int r;
Quoted 12 lines
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +0 -5
-	if (!ring->ready) {
-		DRM_ERROR("Trying to clear memory with ring turned off.\n");
-		return -EINVAL;
-	}
-
    	if (bo->tbo.mem.mem_type == TTM_PL_TT) {
    		r = amdgpu_ttm_alloc_gart(&bo->tbo);
    		if (r)

_______________________________________________
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): BLUPR12MB04490BF8390F6270B265969984C60-7LeqcoF/[email protected]

#26 lore raw

Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Quoted 1 line

If we are in gpu rest stage, we can let SDMA job live with the case of ring not ready, can you point out where is the problem ?

amdgpu_evict_flags() and amdgpu_bo_move() needs to know if the SDMA ring is available during suspend/resume.

E.g. suspend works as this:
1. Evict all normal BOs from VRAM using the SDMA.
2. Disable SDMA, GART, PSP etc... e.g. disable the hardware.
3. Unpin all BOs in VRAM used by hardware engines, e.g. the GART table, firmware etc...
4. Evict all remaining BOs from VRAM using CPU copies.

Resume works the other way around:
1. Move all BOs for GART table and firmware back into VRAM using CPU copies.
2. Pin those BOs in VRAM and enable the hardware.
3. Use the SDMA to move back all other BOs into VRAM.

To figure out if we should use the CPU copy or the SDMA the ring->ready flag is used. So removing this check is really a no-go because it would break suspend/resume.

The only other alternative I can see is to add a separate flag in amdgpu_mman if buffer_funcs should be used or not. See all the callers of amdgpu_ttm_set_active_vram_size() as well.

Something like renaming amdgpu_ttm_set_active_vram_size() into amdgpu_ttm_set_buffer_funcs_read() and ignoring any change if gpu_reset is set.

BTW: Changing the active VRAM size in GPU reset is quite a bug you stumbled over here, so this really needs fixing anyway.

Regards,
Christian.

Am 01.03.2018 um 09:23 schrieb Liu, Monk:

Quoted 17 lines

Accel_working is only for GFX ring, I don't think test it is appropriate for SDMA jobs ....

If we are in gpu rest stage, we can let SDMA job live with the case of ring not ready, can you point out where is the problem ?

In fact we did stress TDR test with this patch and it can really fix couple over kill issues

/Monk

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: 2018年3月1日 16:16
To: Liu, Monk <[email protected]>; Koenig, Christian <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Ah, crap yeah that won't work since we don't free the ring.

Quoted 98 lines

Key point is we need to distinct between the ring doesn't work temporary because we are in a GPU reset and it doesn't work at all because we are missing firmware or stuff like that.

And no, checking the gpu_reset flag is totally racy and can't be done either. How about checking accel_working instead?

Christian.

Am 01.03.2018 um 07:01 schrieb Liu, Monk:

Quoted 91 lines
Quoted 1 line

Please change the test to use ring->ring_obj instead, this way we still bail out if somebody tries to submit commands before the ring is even allocated.

I don't understand how could fill_buffer() get run under the case that ring->ring_obj is not even allocated ... where is such case ?

/Monk

-----Original Message-----
From: Koenig, Christian
Sent: 2018年2月28日 20:46
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Good point, but in this case we need some other handling.

Please change the test to use ring->ring_obj instead, this way we still bail out if somebody tries to submit commands before the ring is even allocated.

And you also need to fix a couple of other places in amdgpu_ttm.c.

Regards,
Christian.

Am 28.02.2018 um 13:34 schrieb Liu, Monk:

Quoted 56 lines

Because when SDMA was hang by like process A, and meanwhile another
process B is already running into the code of fill_buffer() So just let process B continue, don't block it otherwise process B would fail by software reason .

Let it run and finally process B's job would fail and GPU recover will repeat it again (since it is a kernel job)

Without this solution other process will be greatly harmed by one black sheep that triggering GPU recover

/Monk

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: 2018年2月28日 20:24
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Am 28.02.2018 um 08:21 schrieb Monk Liu:

Quoted 6 lines

because this time SDMA may under GPU RESET so its ring->ready may not true, keep going and GPU scheduler will reschedule this job if it failed.

give a warning on copy_buffer when go through direct_submit while ring->ready is false

NAK, that test has already saved us quite a bunch of trouble with the fb layer.

Why exactly are you running into issues with that?

Christian.

Quoted 23 lines
Change-Id Ife6cd55e0e843d99900e5bed5418499e88633685
---
     drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 6 +-----
     1 file changed, 1 insertion(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +1 -0
index e38e6db..7b75ac9 100644
@@ -1656,6 +1656,7 @@ int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
     	amdgpu_ring_pad_ib(ring, &job->ibs[0]);
     	WARN_ON(job->ibs[0].length_dw > num_dw);
     	if (direct_submit) {
+		WARN_ON(!ring->ready);
     		r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs,
     				       NULL, fence);
     		job->fence = dma_fence_get(*fence); @@ -1692,11 +1693,6 @@ int

amdgpu_fill_buffer(struct amdgpu_bo *bo,

     	struct amdgpu_job *job;
     	int r;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +0 -5
-	if (!ring->ready) {
-		DRM_ERROR("Trying to clear memory with ring turned off.\n");
-		return -EINVAL;
-	}
-
     	if (bo->tbo.mem.mem_type == TTM_PL_TT) {
     		r = amdgpu_ttm_alloc_gart(&bo->tbo);
     		if (r)

_______________________________________________
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): 6dedfc7f-cf69-676f-463d-be52cda1b1bb-5C7GfCeVMHo@public.gmane.org

author
#28 lore raw

RE: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Quoted 1 line

amdgpu_evict_flags() and amdgpu_bo_move() needs to know if the SDMA ring is available during suspend/resume.

As long as we are in gpu reset, we don't return error when ring not ready, cuz eventually it either success or failed and rescheduled by scheduler since it is kernel job What's your concern ?

/Monk

-----Original Message-----
From: Koenig, Christian
Sent: 2018年3月1日 16:41
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Quoted 1 line

If we are in gpu rest stage, we can let SDMA job live with the case of ring not ready, can you point out where is the problem ?

amdgpu_evict_flags() and amdgpu_bo_move() needs to know if the SDMA ring is available during suspend/resume.

E.g. suspend works as this:
1. Evict all normal BOs from VRAM using the SDMA.
2. Disable SDMA, GART, PSP etc... e.g. disable the hardware.
3. Unpin all BOs in VRAM used by hardware engines, e.g. the GART table, firmware etc...
4. Evict all remaining BOs from VRAM using CPU copies.

Resume works the other way around:
1. Move all BOs for GART table and firmware back into VRAM using CPU copies.
2. Pin those BOs in VRAM and enable the hardware.
3. Use the SDMA to move back all other BOs into VRAM.

To figure out if we should use the CPU copy or the SDMA the ring->ready flag is used. So removing this check is really a no-go because it would break suspend/resume.

The only other alternative I can see is to add a separate flag in amdgpu_mman if buffer_funcs should be used or not. See all the callers of amdgpu_ttm_set_active_vram_size() as well.

Something like renaming amdgpu_ttm_set_active_vram_size() into amdgpu_ttm_set_buffer_funcs_read() and ignoring any change if gpu_reset is set.

BTW: Changing the active VRAM size in GPU reset is quite a bug you stumbled over here, so this really needs fixing anyway.

Regards,
Christian.

Am 01.03.2018 um 09:23 schrieb Liu, Monk:

Quoted 20 lines

Accel_working is only for GFX ring, I don't think test it is appropriate for SDMA jobs ....

If we are in gpu rest stage, we can let SDMA job live with the case of ring not ready, can you point out where is the problem ?

In fact we did stress TDR test with this patch and it can really fix couple over kill issues

/Monk

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: 2018年3月1日 16:16
To: Liu, Monk <[email protected]>; Koenig, Christian <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Ah, crap yeah that won't work since we don't free the ring.

Quoted 98 lines

Key point is we need to distinct between the ring doesn't work temporary because we are in a GPU reset and it doesn't work at all because we are missing firmware or stuff like that.

And no, checking the gpu_reset flag is totally racy and can't be done either. How about checking accel_working instead?

Christian.

Am 01.03.2018 um 07:01 schrieb Liu, Monk:

Quoted 91 lines
Quoted 1 line

Please change the test to use ring->ring_obj instead, this way we still bail out if somebody tries to submit commands before the ring is even allocated.

I don't understand how could fill_buffer() get run under the case that ring->ring_obj is not even allocated ... where is such case ?

/Monk

-----Original Message-----
From: Koenig, Christian
Sent: 2018年2月28日 20:46
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Good point, but in this case we need some other handling.

Please change the test to use ring->ring_obj instead, this way we still bail out if somebody tries to submit commands before the ring is even allocated.

And you also need to fix a couple of other places in amdgpu_ttm.c.

Regards,
Christian.

Am 28.02.2018 um 13:34 schrieb Liu, Monk:

Quoted 56 lines

Because when SDMA was hang by like process A, and meanwhile another
process B is already running into the code of fill_buffer() So just let process B continue, don't block it otherwise process B would fail by software reason .

Let it run and finally process B's job would fail and GPU recover will repeat it again (since it is a kernel job)

Without this solution other process will be greatly harmed by one black sheep that triggering GPU recover

/Monk

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: 2018年2月28日 20:24
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Am 28.02.2018 um 08:21 schrieb Monk Liu:

Quoted 6 lines

because this time SDMA may under GPU RESET so its ring->ready may not true, keep going and GPU scheduler will reschedule this job if it failed.

give a warning on copy_buffer when go through direct_submit while ring->ready is false

NAK, that test has already saved us quite a bunch of trouble with the fb layer.

Why exactly are you running into issues with that?

Christian.

Quoted 23 lines
Change-Id Ife6cd55e0e843d99900e5bed5418499e88633685
---
     drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 6 +-----
     1 file changed, 1 insertion(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +1 -0
index e38e6db..7b75ac9 100644
@@ -1656,6 +1656,7 @@ int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
     	amdgpu_ring_pad_ib(ring, &job->ibs[0]);
     	WARN_ON(job->ibs[0].length_dw > num_dw);
     	if (direct_submit) {
+		WARN_ON(!ring->ready);
     		r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs,
     				       NULL, fence);
     		job->fence = dma_fence_get(*fence); @@ -1692,11 +1693,6 @@ 

int amdgpu_fill_buffer(struct amdgpu_bo *bo,

     	struct amdgpu_job *job;
     	int r;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +0 -5
-	if (!ring->ready) {
-		DRM_ERROR("Trying to clear memory with ring turned off.\n");
-		return -EINVAL;
-	}
-
     	if (bo->tbo.mem.mem_type == TTM_PL_TT) {
     		r = amdgpu_ttm_alloc_gart(&bo->tbo);
     		if (r)

_______________________________________________
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): BLUPR12MB04491577FB4B6D6A5BCC3B8784C60-7LeqcoF/[email protected]

author
#30 lore raw

RE: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Okay I missed the part of suspend/resume/s3

For SRIOV we don't run into that logic, looks I need to keep it in private branch till A better solution raised

/Monk

-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Liu, Monk
Sent: 2018年3月1日 16:52
To: Koenig, Christian <[email protected]>; [email protected]
Subject: RE: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Quoted 1 line

amdgpu_evict_flags() and amdgpu_bo_move() needs to know if the SDMA ring is available during suspend/resume.

As long as we are in gpu reset, we don't return error when ring not ready, cuz eventually it either success or failed and rescheduled by scheduler since it is kernel job What's your concern ?

/Monk

-----Original Message-----
From: Koenig, Christian
Sent: 2018年3月1日 16:41
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Quoted 1 line

If we are in gpu rest stage, we can let SDMA job live with the case of ring not ready, can you point out where is the problem ?

amdgpu_evict_flags() and amdgpu_bo_move() needs to know if the SDMA ring is available during suspend/resume.

E.g. suspend works as this:
1. Evict all normal BOs from VRAM using the SDMA.
2. Disable SDMA, GART, PSP etc... e.g. disable the hardware.
3. Unpin all BOs in VRAM used by hardware engines, e.g. the GART table, firmware etc...
4. Evict all remaining BOs from VRAM using CPU copies.

Resume works the other way around:
1. Move all BOs for GART table and firmware back into VRAM using CPU copies.
2. Pin those BOs in VRAM and enable the hardware.
3. Use the SDMA to move back all other BOs into VRAM.

To figure out if we should use the CPU copy or the SDMA the ring->ready flag is used. So removing this check is really a no-go because it would break suspend/resume.

The only other alternative I can see is to add a separate flag in amdgpu_mman if buffer_funcs should be used or not. See all the callers of amdgpu_ttm_set_active_vram_size() as well.

Something like renaming amdgpu_ttm_set_active_vram_size() into amdgpu_ttm_set_buffer_funcs_read() and ignoring any change if gpu_reset is set.

BTW: Changing the active VRAM size in GPU reset is quite a bug you stumbled over here, so this really needs fixing anyway.

Regards,
Christian.

Am 01.03.2018 um 09:23 schrieb Liu, Monk:

Quoted 20 lines

Accel_working is only for GFX ring, I don't think test it is appropriate for SDMA jobs ....

If we are in gpu rest stage, we can let SDMA job live with the case of ring not ready, can you point out where is the problem ?

In fact we did stress TDR test with this patch and it can really fix couple over kill issues

/Monk

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: 2018年3月1日 16:16
To: Liu, Monk <[email protected]>; Koenig, Christian <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Ah, crap yeah that won't work since we don't free the ring.

Quoted 98 lines

Key point is we need to distinct between the ring doesn't work temporary because we are in a GPU reset and it doesn't work at all because we are missing firmware or stuff like that.

And no, checking the gpu_reset flag is totally racy and can't be done either. How about checking accel_working instead?

Christian.

Am 01.03.2018 um 07:01 schrieb Liu, Monk:

Quoted 91 lines
Quoted 1 line

Please change the test to use ring->ring_obj instead, this way we still bail out if somebody tries to submit commands before the ring is even allocated.

I don't understand how could fill_buffer() get run under the case that ring->ring_obj is not even allocated ... where is such case ?

/Monk

-----Original Message-----
From: Koenig, Christian
Sent: 2018年2月28日 20:46
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Good point, but in this case we need some other handling.

Please change the test to use ring->ring_obj instead, this way we still bail out if somebody tries to submit commands before the ring is even allocated.

And you also need to fix a couple of other places in amdgpu_ttm.c.

Regards,
Christian.

Am 28.02.2018 um 13:34 schrieb Liu, Monk:

Quoted 56 lines

Because when SDMA was hang by like process A, and meanwhile another
process B is already running into the code of fill_buffer() So just let process B continue, don't block it otherwise process B would fail by software reason .

Let it run and finally process B's job would fail and GPU recover will repeat it again (since it is a kernel job)

Without this solution other process will be greatly harmed by one black sheep that triggering GPU recover

/Monk

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: 2018年2月28日 20:24
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Am 28.02.2018 um 08:21 schrieb Monk Liu:

Quoted 6 lines

because this time SDMA may under GPU RESET so its ring->ready may not true, keep going and GPU scheduler will reschedule this job if it failed.

give a warning on copy_buffer when go through direct_submit while ring->ready is false

NAK, that test has already saved us quite a bunch of trouble with the fb layer.

Why exactly are you running into issues with that?

Christian.

Quoted 23 lines
Change-Id Ife6cd55e0e843d99900e5bed5418499e88633685
---
     drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 6 +-----
     1 file changed, 1 insertion(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +1 -0
index e38e6db..7b75ac9 100644
@@ -1656,6 +1656,7 @@ int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
     	amdgpu_ring_pad_ib(ring, &job->ibs[0]);
     	WARN_ON(job->ibs[0].length_dw > num_dw);
     	if (direct_submit) {
+		WARN_ON(!ring->ready);
     		r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs,
     				       NULL, fence);
     		job->fence = dma_fence_get(*fence); @@ -1692,11 +1693,6 @@ 

int amdgpu_fill_buffer(struct amdgpu_bo *bo,

     	struct amdgpu_job *job;
     	int r;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +0 -5
-	if (!ring->ready) {
-		DRM_ERROR("Trying to clear memory with ring turned off.\n");
-		return -EINVAL;
-	}
-
     	if (bo->tbo.mem.mem_type == TTM_PL_TT) {
     		r = amdgpu_ttm_alloc_gart(&bo->tbo);
     		if (r)

_______________________________________________
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): BLUPR12MB044933DC340A46EADC1F7A2784C60-7LeqcoF/[email protected]

#32 lore raw

Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Even for bare-metal changing the VRAM size in the memory manager because we disable/enable the SDMA during GPU reset is quite a bug.

Going to work on this, a correct fix should actually be really trivial.

Regards,
Christian.

Am 01.03.2018 um 10:25 schrieb Liu, Monk:

Quoted 170 lines

Okay I missed the part of suspend/resume/s3

For SRIOV we don't run into that logic, looks I need to keep it in private branch till A better solution raised

/Monk

-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Liu, Monk
Sent: 2018年3月1日 16:52
To: Koenig, Christian <[email protected]>; [email protected]
Subject: RE: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Quoted 1 line

amdgpu_evict_flags() and amdgpu_bo_move() needs to know if the SDMA ring is available during suspend/resume.

As long as we are in gpu reset, we don't return error when ring not ready, cuz eventually it either success or failed and rescheduled by scheduler since it is kernel job What's your concern ?

/Monk

-----Original Message-----
From: Koenig, Christian
Sent: 2018年3月1日 16:41
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Quoted 1 line

If we are in gpu rest stage, we can let SDMA job live with the case of ring not ready, can you point out where is the problem ?

amdgpu_evict_flags() and amdgpu_bo_move() needs to know if the SDMA ring is available during suspend/resume.

E.g. suspend works as this:
1. Evict all normal BOs from VRAM using the SDMA.
2. Disable SDMA, GART, PSP etc... e.g. disable the hardware.
3. Unpin all BOs in VRAM used by hardware engines, e.g. the GART table, firmware etc...
4. Evict all remaining BOs from VRAM using CPU copies.

Resume works the other way around:
1. Move all BOs for GART table and firmware back into VRAM using CPU copies.
2. Pin those BOs in VRAM and enable the hardware.
3. Use the SDMA to move back all other BOs into VRAM.

To figure out if we should use the CPU copy or the SDMA the ring->ready flag is used. So removing this check is really a no-go because it would break suspend/resume.

The only other alternative I can see is to add a separate flag in amdgpu_mman if buffer_funcs should be used or not. See all the callers of amdgpu_ttm_set_active_vram_size() as well.

Something like renaming amdgpu_ttm_set_active_vram_size() into amdgpu_ttm_set_buffer_funcs_read() and ignoring any change if gpu_reset is set.

BTW: Changing the active VRAM size in GPU reset is quite a bug you stumbled over here, so this really needs fixing anyway.

Regards,
Christian.

Am 01.03.2018 um 09:23 schrieb Liu, Monk:

Quoted 20 lines

Accel_working is only for GFX ring, I don't think test it is appropriate for SDMA jobs ....

If we are in gpu rest stage, we can let SDMA job live with the case of ring not ready, can you point out where is the problem ?

In fact we did stress TDR test with this patch and it can really fix couple over kill issues

/Monk

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: 2018年3月1日 16:16
To: Liu, Monk <[email protected]>; Koenig, Christian <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Ah, crap yeah that won't work since we don't free the ring.

Quoted 98 lines

Key point is we need to distinct between the ring doesn't work temporary because we are in a GPU reset and it doesn't work at all because we are missing firmware or stuff like that.

And no, checking the gpu_reset flag is totally racy and can't be done either. How about checking accel_working instead?

Christian.

Am 01.03.2018 um 07:01 schrieb Liu, Monk:

Quoted 91 lines
Quoted 1 line

Please change the test to use ring->ring_obj instead, this way we still bail out if somebody tries to submit commands before the ring is even allocated.

I don't understand how could fill_buffer() get run under the case that ring->ring_obj is not even allocated ... where is such case ?

/Monk

-----Original Message-----
From: Koenig, Christian
Sent: 2018年2月28日 20:46
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Good point, but in this case we need some other handling.

Please change the test to use ring->ring_obj instead, this way we still bail out if somebody tries to submit commands before the ring is even allocated.

And you also need to fix a couple of other places in amdgpu_ttm.c.

Regards,
Christian.

Am 28.02.2018 um 13:34 schrieb Liu, Monk:

Quoted 56 lines

Because when SDMA was hang by like process A, and meanwhile another
process B is already running into the code of fill_buffer() So just let process B continue, don't block it otherwise process B would fail by software reason .

Let it run and finally process B's job would fail and GPU recover will repeat it again (since it is a kernel job)

Without this solution other process will be greatly harmed by one black sheep that triggering GPU recover

/Monk

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: 2018年2月28日 20:24
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Am 28.02.2018 um 08:21 schrieb Monk Liu:

Quoted 6 lines

because this time SDMA may under GPU RESET so its ring->ready may not true, keep going and GPU scheduler will reschedule this job if it failed.

give a warning on copy_buffer when go through direct_submit while ring->ready is false

NAK, that test has already saved us quite a bunch of trouble with the fb layer.

Why exactly are you running into issues with that?

Christian.

Quoted 23 lines
Change-Id Ife6cd55e0e843d99900e5bed5418499e88633685
---
      drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 6 +-----
      1 file changed, 1 insertion(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +1 -0
index e38e6db..7b75ac9 100644
@@ -1656,6 +1656,7 @@ int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
      	amdgpu_ring_pad_ib(ring, &job->ibs[0]);
      	WARN_ON(job->ibs[0].length_dw > num_dw);
      	if (direct_submit) {
+		WARN_ON(!ring->ready);
      		r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs,
      				       NULL, fence);
      		job->fence = dma_fence_get(*fence); @@ -1692,11 +1693,6 @@

int amdgpu_fill_buffer(struct amdgpu_bo *bo,

      	struct amdgpu_job *job;
      	int r;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +0 -5
-	if (!ring->ready) {
-		DRM_ERROR("Trying to clear memory with ring turned off.\n");
-		return -EINVAL;
-	}
-
      	if (bo->tbo.mem.mem_type == TTM_PL_TT) {
      		r = amdgpu_ttm_alloc_gart(&bo->tbo);
      		if (r)

_______________________________________________
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

author
#33 lore raw

RE: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

In reply to a message not in the archive

Hi Christian

During suspend/resume, which stage will set sdma's ring->ready to false ?? I can only find it in amdgpu_ring_fini() But I saw that it is only invoked in unload_kms(), looks device_suspnd() will not set ring->ready to false

/Monk

-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Liu, Monk
Sent: 2018年3月1日 16:52
To: Koenig, Christian <[email protected]>; [email protected]
Subject: RE: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Quoted 1 line

amdgpu_evict_flags() and amdgpu_bo_move() needs to know if the SDMA ring is available during suspend/resume.

As long as we are in gpu reset, we don't return error when ring not ready, cuz eventually it either success or failed and rescheduled by scheduler since it is kernel job What's your concern ?

/Monk

-----Original Message-----
From: Koenig, Christian
Sent: 2018年3月1日 16:41
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Quoted 1 line

If we are in gpu rest stage, we can let SDMA job live with the case of ring not ready, can you point out where is the problem ?

amdgpu_evict_flags() and amdgpu_bo_move() needs to know if the SDMA ring is available during suspend/resume.

E.g. suspend works as this:
1. Evict all normal BOs from VRAM using the SDMA.
2. Disable SDMA, GART, PSP etc... e.g. disable the hardware.
3. Unpin all BOs in VRAM used by hardware engines, e.g. the GART table, firmware etc...
4. Evict all remaining BOs from VRAM using CPU copies.

Resume works the other way around:
1. Move all BOs for GART table and firmware back into VRAM using CPU copies.
2. Pin those BOs in VRAM and enable the hardware.
3. Use the SDMA to move back all other BOs into VRAM.

To figure out if we should use the CPU copy or the SDMA the ring->ready flag is used. So removing this check is really a no-go because it would break suspend/resume.

The only other alternative I can see is to add a separate flag in amdgpu_mman if buffer_funcs should be used or not. See all the callers of amdgpu_ttm_set_active_vram_size() as well.

Something like renaming amdgpu_ttm_set_active_vram_size() into amdgpu_ttm_set_buffer_funcs_read() and ignoring any change if gpu_reset is set.

BTW: Changing the active VRAM size in GPU reset is quite a bug you stumbled over here, so this really needs fixing anyway.

Regards,
Christian.

Am 01.03.2018 um 09:23 schrieb Liu, Monk:

Quoted 20 lines

Accel_working is only for GFX ring, I don't think test it is appropriate for SDMA jobs ....

If we are in gpu rest stage, we can let SDMA job live with the case of ring not ready, can you point out where is the problem ?

In fact we did stress TDR test with this patch and it can really fix couple over kill issues

/Monk

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: 2018年3月1日 16:16
To: Liu, Monk <[email protected]>; Koenig, Christian <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Ah, crap yeah that won't work since we don't free the ring.

Quoted 98 lines

Key point is we need to distinct between the ring doesn't work temporary because we are in a GPU reset and it doesn't work at all because we are missing firmware or stuff like that.

And no, checking the gpu_reset flag is totally racy and can't be done either. How about checking accel_working instead?

Christian.

Am 01.03.2018 um 07:01 schrieb Liu, Monk:

Quoted 91 lines
Quoted 1 line

Please change the test to use ring->ring_obj instead, this way we still bail out if somebody tries to submit commands before the ring is even allocated.

I don't understand how could fill_buffer() get run under the case that ring->ring_obj is not even allocated ... where is such case ?

/Monk

-----Original Message-----
From: Koenig, Christian
Sent: 2018年2月28日 20:46
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Good point, but in this case we need some other handling.

Please change the test to use ring->ring_obj instead, this way we still bail out if somebody tries to submit commands before the ring is even allocated.

And you also need to fix a couple of other places in amdgpu_ttm.c.

Regards,
Christian.

Am 28.02.2018 um 13:34 schrieb Liu, Monk:

Quoted 56 lines

Because when SDMA was hang by like process A, and meanwhile another
process B is already running into the code of fill_buffer() So just let process B continue, don't block it otherwise process B would fail by software reason .

Let it run and finally process B's job would fail and GPU recover will repeat it again (since it is a kernel job)

Without this solution other process will be greatly harmed by one black sheep that triggering GPU recover

/Monk

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: 2018年2月28日 20:24
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Am 28.02.2018 um 08:21 schrieb Monk Liu:

Quoted 6 lines

because this time SDMA may under GPU RESET so its ring->ready may not true, keep going and GPU scheduler will reschedule this job if it failed.

give a warning on copy_buffer when go through direct_submit while ring->ready is false

NAK, that test has already saved us quite a bunch of trouble with the fb layer.

Why exactly are you running into issues with that?

Christian.

Quoted 23 lines
Change-Id Ife6cd55e0e843d99900e5bed5418499e88633685
---
     drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 6 +-----
     1 file changed, 1 insertion(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +1 -0
index e38e6db..7b75ac9 100644
@@ -1656,6 +1656,7 @@ int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
     	amdgpu_ring_pad_ib(ring, &job->ibs[0]);
     	WARN_ON(job->ibs[0].length_dw > num_dw);
     	if (direct_submit) {
+		WARN_ON(!ring->ready);
     		r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs,
     				       NULL, fence);
     		job->fence = dma_fence_get(*fence); @@ -1692,11 +1693,6 @@ 

int amdgpu_fill_buffer(struct amdgpu_bo *bo,

     	struct amdgpu_job *job;
     	int r;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +0 -5
-	if (!ring->ready) {
-		DRM_ERROR("Trying to clear memory with ring turned off.\n");
-		return -EINVAL;
-	}
-
     	if (bo->tbo.mem.mem_type == TTM_PL_TT) {
     		r = amdgpu_ttm_alloc_gart(&bo->tbo);
     		if (r)

_______________________________________________
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): BLUPR12MB044905DB87968358166BDEF484C60-7LeqcoF/[email protected]

#35 lore raw

Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Hi Monk,

look at si_dma_stop() and/or other SDMA generation code.

Regards,
Christian.

Am 01.03.2018 um 10:37 schrieb Liu, Monk:

Quoted 170 lines

Hi Christian

During suspend/resume, which stage will set sdma's ring->ready to false ?? I can only find it in amdgpu_ring_fini() But I saw that it is only invoked in unload_kms(), looks device_suspnd() will not set ring->ready to false

/Monk

-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Liu, Monk
Sent: 2018年3月1日 16:52
To: Koenig, Christian <[email protected]>; [email protected]
Subject: RE: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Quoted 1 line

amdgpu_evict_flags() and amdgpu_bo_move() needs to know if the SDMA ring is available during suspend/resume.

As long as we are in gpu reset, we don't return error when ring not ready, cuz eventually it either success or failed and rescheduled by scheduler since it is kernel job What's your concern ?

/Monk

-----Original Message-----
From: Koenig, Christian
Sent: 2018年3月1日 16:41
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Quoted 1 line

If we are in gpu rest stage, we can let SDMA job live with the case of ring not ready, can you point out where is the problem ?

amdgpu_evict_flags() and amdgpu_bo_move() needs to know if the SDMA ring is available during suspend/resume.

E.g. suspend works as this:
1. Evict all normal BOs from VRAM using the SDMA.
2. Disable SDMA, GART, PSP etc... e.g. disable the hardware.
3. Unpin all BOs in VRAM used by hardware engines, e.g. the GART table, firmware etc...
4. Evict all remaining BOs from VRAM using CPU copies.

Resume works the other way around:
1. Move all BOs for GART table and firmware back into VRAM using CPU copies.
2. Pin those BOs in VRAM and enable the hardware.
3. Use the SDMA to move back all other BOs into VRAM.

To figure out if we should use the CPU copy or the SDMA the ring->ready flag is used. So removing this check is really a no-go because it would break suspend/resume.

The only other alternative I can see is to add a separate flag in amdgpu_mman if buffer_funcs should be used or not. See all the callers of amdgpu_ttm_set_active_vram_size() as well.

Something like renaming amdgpu_ttm_set_active_vram_size() into amdgpu_ttm_set_buffer_funcs_read() and ignoring any change if gpu_reset is set.

BTW: Changing the active VRAM size in GPU reset is quite a bug you stumbled over here, so this really needs fixing anyway.

Regards,
Christian.

Am 01.03.2018 um 09:23 schrieb Liu, Monk:

Quoted 20 lines

Accel_working is only for GFX ring, I don't think test it is appropriate for SDMA jobs ....

If we are in gpu rest stage, we can let SDMA job live with the case of ring not ready, can you point out where is the problem ?

In fact we did stress TDR test with this patch and it can really fix couple over kill issues

/Monk

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: 2018年3月1日 16:16
To: Liu, Monk <[email protected]>; Koenig, Christian <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Ah, crap yeah that won't work since we don't free the ring.

Quoted 98 lines

Key point is we need to distinct between the ring doesn't work temporary because we are in a GPU reset and it doesn't work at all because we are missing firmware or stuff like that.

And no, checking the gpu_reset flag is totally racy and can't be done either. How about checking accel_working instead?

Christian.

Am 01.03.2018 um 07:01 schrieb Liu, Monk:

Quoted 91 lines
Quoted 1 line

Please change the test to use ring->ring_obj instead, this way we still bail out if somebody tries to submit commands before the ring is even allocated.

I don't understand how could fill_buffer() get run under the case that ring->ring_obj is not even allocated ... where is such case ?

/Monk

-----Original Message-----
From: Koenig, Christian
Sent: 2018年2月28日 20:46
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Good point, but in this case we need some other handling.

Please change the test to use ring->ring_obj instead, this way we still bail out if somebody tries to submit commands before the ring is even allocated.

And you also need to fix a couple of other places in amdgpu_ttm.c.

Regards,
Christian.

Am 28.02.2018 um 13:34 schrieb Liu, Monk:

Quoted 56 lines

Because when SDMA was hang by like process A, and meanwhile another
process B is already running into the code of fill_buffer() So just let process B continue, don't block it otherwise process B would fail by software reason .

Let it run and finally process B's job would fail and GPU recover will repeat it again (since it is a kernel job)

Without this solution other process will be greatly harmed by one black sheep that triggering GPU recover

/Monk

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: 2018年2月28日 20:24
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Am 28.02.2018 um 08:21 schrieb Monk Liu:

Quoted 6 lines

because this time SDMA may under GPU RESET so its ring->ready may not true, keep going and GPU scheduler will reschedule this job if it failed.

give a warning on copy_buffer when go through direct_submit while ring->ready is false

NAK, that test has already saved us quite a bunch of trouble with the fb layer.

Why exactly are you running into issues with that?

Christian.

Quoted 23 lines
Change-Id Ife6cd55e0e843d99900e5bed5418499e88633685
---
      drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 6 +-----
      1 file changed, 1 insertion(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +1 -0
index e38e6db..7b75ac9 100644
@@ -1656,6 +1656,7 @@ int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
      	amdgpu_ring_pad_ib(ring, &job->ibs[0]);
      	WARN_ON(job->ibs[0].length_dw > num_dw);
      	if (direct_submit) {
+		WARN_ON(!ring->ready);
      		r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs,
      				       NULL, fence);
      		job->fence = dma_fence_get(*fence); @@ -1692,11 +1693,6 @@

int amdgpu_fill_buffer(struct amdgpu_bo *bo,

      	struct amdgpu_job *job;
      	int r;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +0 -5
-	if (!ring->ready) {
-		DRM_ERROR("Trying to clear memory with ring turned off.\n");
-		return -EINVAL;
-	}
-
      	if (bo->tbo.mem.mem_type == TTM_PL_TT) {
      		r = amdgpu_ttm_alloc_gart(&bo->tbo);
      		if (r)

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

Quoted 8 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): a6812ba2-1c34-c6a4-d65a-09f924ea0940-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org

author
#37 lore raw

RE: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Oops, looks only si will call this dma_stop() during hw_fini, I checked for sdma4.0/3.0/2.4, no such logic

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: 2018年3月1日 17:39
To: Liu, Monk <[email protected]>; Koenig, Christian <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Hi Monk,

look at si_dma_stop() and/or other SDMA generation code.

Regards,
Christian.

Am 01.03.2018 um 10:37 schrieb Liu, Monk:

Quoted 176 lines

Hi Christian

During suspend/resume, which stage will set sdma's ring->ready to false ?? I can only find it in amdgpu_ring_fini() But I saw that it is only invoked in unload_kms(), looks device_suspnd() will not set ring->ready to false

/Monk

-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Liu, Monk
Sent: 2018年3月1日 16:52
To: Koenig, Christian <[email protected]>; [email protected]
Subject: RE: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Quoted 1 line

amdgpu_evict_flags() and amdgpu_bo_move() needs to know if the SDMA ring is available during suspend/resume.

As long as we are in gpu reset, we don't return error when ring not ready, cuz eventually it either success or failed and rescheduled by scheduler since it is kernel job What's your concern ?

/Monk

-----Original Message-----
From: Koenig, Christian
Sent: 2018年3月1日 16:41
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Quoted 1 line

If we are in gpu rest stage, we can let SDMA job live with the case of ring not ready, can you point out where is the problem ?

amdgpu_evict_flags() and amdgpu_bo_move() needs to know if the SDMA ring is available during suspend/resume.

E.g. suspend works as this:
1. Evict all normal BOs from VRAM using the SDMA.
2. Disable SDMA, GART, PSP etc... e.g. disable the hardware.
3. Unpin all BOs in VRAM used by hardware engines, e.g. the GART table, firmware etc...
4. Evict all remaining BOs from VRAM using CPU copies.

Resume works the other way around:
1. Move all BOs for GART table and firmware back into VRAM using CPU copies.
2. Pin those BOs in VRAM and enable the hardware.
3. Use the SDMA to move back all other BOs into VRAM.

To figure out if we should use the CPU copy or the SDMA the ring->ready flag is used. So removing this check is really a no-go because it would break suspend/resume.

The only other alternative I can see is to add a separate flag in amdgpu_mman if buffer_funcs should be used or not. See all the callers of amdgpu_ttm_set_active_vram_size() as well.

Something like renaming amdgpu_ttm_set_active_vram_size() into amdgpu_ttm_set_buffer_funcs_read() and ignoring any change if gpu_reset is set.

BTW: Changing the active VRAM size in GPU reset is quite a bug you stumbled over here, so this really needs fixing anyway.

Regards,
Christian.

Am 01.03.2018 um 09:23 schrieb Liu, Monk:

Quoted 20 lines

Accel_working is only for GFX ring, I don't think test it is appropriate for SDMA jobs ....

If we are in gpu rest stage, we can let SDMA job live with the case of ring not ready, can you point out where is the problem ?

In fact we did stress TDR test with this patch and it can really fix couple over kill issues

/Monk

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: 2018年3月1日 16:16
To: Liu, Monk <[email protected]>; Koenig, Christian <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Ah, crap yeah that won't work since we don't free the ring.

Quoted 98 lines

Key point is we need to distinct between the ring doesn't work temporary because we are in a GPU reset and it doesn't work at all because we are missing firmware or stuff like that.

And no, checking the gpu_reset flag is totally racy and can't be done either. How about checking accel_working instead?

Christian.

Am 01.03.2018 um 07:01 schrieb Liu, Monk:

Quoted 91 lines
Quoted 1 line

Please change the test to use ring->ring_obj instead, this way we still bail out if somebody tries to submit commands before the ring is even allocated.

I don't understand how could fill_buffer() get run under the case that ring->ring_obj is not even allocated ... where is such case ?

/Monk

-----Original Message-----
From: Koenig, Christian
Sent: 2018年2月28日 20:46
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Good point, but in this case we need some other handling.

Please change the test to use ring->ring_obj instead, this way we still bail out if somebody tries to submit commands before the ring is even allocated.

And you also need to fix a couple of other places in amdgpu_ttm.c.

Regards,
Christian.

Am 28.02.2018 um 13:34 schrieb Liu, Monk:

Quoted 56 lines

Because when SDMA was hang by like process A, and meanwhile another
process B is already running into the code of fill_buffer() So just let process B continue, don't block it otherwise process B would fail by software reason .

Let it run and finally process B's job would fail and GPU recover will repeat it again (since it is a kernel job)

Without this solution other process will be greatly harmed by one black sheep that triggering GPU recover

/Monk

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: 2018年2月28日 20:24
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Am 28.02.2018 um 08:21 schrieb Monk Liu:

Quoted 6 lines

because this time SDMA may under GPU RESET so its ring->ready may not true, keep going and GPU scheduler will reschedule this job if it failed.

give a warning on copy_buffer when go through direct_submit while ring->ready is false

NAK, that test has already saved us quite a bunch of trouble with the fb layer.

Why exactly are you running into issues with that?

Christian.

Quoted 23 lines
Change-Id Ife6cd55e0e843d99900e5bed5418499e88633685
---
      drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 6 +-----
      1 file changed, 1 insertion(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +1 -0
index e38e6db..7b75ac9 100644
@@ -1656,6 +1656,7 @@ int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
      	amdgpu_ring_pad_ib(ring, &job->ibs[0]);
      	WARN_ON(job->ibs[0].length_dw > num_dw);
      	if (direct_submit) {
+		WARN_ON(!ring->ready);
      		r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs,
      				       NULL, fence);
      		job->fence = dma_fence_get(*fence); @@ -1692,11 +1693,6 @@ 

int amdgpu_fill_buffer(struct amdgpu_bo *bo,

      	struct amdgpu_job *job;
      	int r;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +0 -5
-	if (!ring->ready) {
-		DRM_ERROR("Trying to clear memory with ring turned off.\n");
-		return -EINVAL;
-	}
-
      	if (bo->tbo.mem.mem_type == TTM_PL_TT) {
      		r = amdgpu_ttm_alloc_gart(&bo->tbo);
      		if (r)

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

Quoted 8 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): BLUPR12MB044923819BB9561C57EACA8584C60-7LeqcoF/[email protected]

#39 lore raw

Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Look for amdgpu_ttm_set_active_vram_size(), setting ring->ready to false is directly beneath it.

Christian.

Am 01.03.2018 um 10:41 schrieb Liu, Monk:

Quoted 10 lines

Oops, looks only si will call this dma_stop() during hw_fini, I checked for sdma4.0/3.0/2.4, no such logic

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: 2018年3月1日 17:39
To: Liu, Monk <[email protected]>; Koenig, Christian <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Quoted 192 lines

Hi Monk,

look at si_dma_stop() and/or other SDMA generation code.

Regards,
Christian.

Am 01.03.2018 um 10:37 schrieb Liu, Monk:

Quoted 184 lines

Hi Christian

During suspend/resume, which stage will set sdma's ring->ready to false ?? I can only find it in amdgpu_ring_fini() But I saw that it is only invoked in unload_kms(), looks device_suspnd() will not set ring->ready to false

/Monk

-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Liu, Monk
Sent: 2018年3月1日 16:52
To: Koenig, Christian <[email protected]>; [email protected]
Subject: RE: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Quoted 1 line

amdgpu_evict_flags() and amdgpu_bo_move() needs to know if the SDMA ring is available during suspend/resume.

As long as we are in gpu reset, we don't return error when ring not ready, cuz eventually it either success or failed and rescheduled by scheduler since it is kernel job What's your concern ?

/Monk

-----Original Message-----
From: Koenig, Christian
Sent: 2018年3月1日 16:41
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Quoted 1 line

If we are in gpu rest stage, we can let SDMA job live with the case of ring not ready, can you point out where is the problem ?

amdgpu_evict_flags() and amdgpu_bo_move() needs to know if the SDMA ring is available during suspend/resume.

E.g. suspend works as this:
1. Evict all normal BOs from VRAM using the SDMA.
2. Disable SDMA, GART, PSP etc... e.g. disable the hardware.
3. Unpin all BOs in VRAM used by hardware engines, e.g. the GART table, firmware etc...
4. Evict all remaining BOs from VRAM using CPU copies.

Resume works the other way around:
1. Move all BOs for GART table and firmware back into VRAM using CPU copies.
2. Pin those BOs in VRAM and enable the hardware.
3. Use the SDMA to move back all other BOs into VRAM.

To figure out if we should use the CPU copy or the SDMA the ring->ready flag is used. So removing this check is really a no-go because it would break suspend/resume.

The only other alternative I can see is to add a separate flag in amdgpu_mman if buffer_funcs should be used or not. See all the callers of amdgpu_ttm_set_active_vram_size() as well.

Something like renaming amdgpu_ttm_set_active_vram_size() into amdgpu_ttm_set_buffer_funcs_read() and ignoring any change if gpu_reset is set.

BTW: Changing the active VRAM size in GPU reset is quite a bug you stumbled over here, so this really needs fixing anyway.

Regards,
Christian.

Am 01.03.2018 um 09:23 schrieb Liu, Monk:

Quoted 20 lines

Accel_working is only for GFX ring, I don't think test it is appropriate for SDMA jobs ....

If we are in gpu rest stage, we can let SDMA job live with the case of ring not ready, can you point out where is the problem ?

In fact we did stress TDR test with this patch and it can really fix couple over kill issues

/Monk

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: 2018年3月1日 16:16
To: Liu, Monk <[email protected]>; Koenig, Christian <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Ah, crap yeah that won't work since we don't free the ring.

Quoted 98 lines

Key point is we need to distinct between the ring doesn't work temporary because we are in a GPU reset and it doesn't work at all because we are missing firmware or stuff like that.

And no, checking the gpu_reset flag is totally racy and can't be done either. How about checking accel_working instead?

Christian.

Am 01.03.2018 um 07:01 schrieb Liu, Monk:

Quoted 91 lines
Quoted 1 line

Please change the test to use ring->ring_obj instead, this way we still bail out if somebody tries to submit commands before the ring is even allocated.

I don't understand how could fill_buffer() get run under the case that ring->ring_obj is not even allocated ... where is such case ?

/Monk

-----Original Message-----
From: Koenig, Christian
Sent: 2018年2月28日 20:46
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Good point, but in this case we need some other handling.

Please change the test to use ring->ring_obj instead, this way we still bail out if somebody tries to submit commands before the ring is even allocated.

And you also need to fix a couple of other places in amdgpu_ttm.c.

Regards,
Christian.

Am 28.02.2018 um 13:34 schrieb Liu, Monk:

Quoted 56 lines

Because when SDMA was hang by like process A, and meanwhile another
process B is already running into the code of fill_buffer() So just let process B continue, don't block it otherwise process B would fail by software reason .

Let it run and finally process B's job would fail and GPU recover will repeat it again (since it is a kernel job)

Without this solution other process will be greatly harmed by one black sheep that triggering GPU recover

/Monk

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: 2018年2月28日 20:24
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Am 28.02.2018 um 08:21 schrieb Monk Liu:

Quoted 6 lines

because this time SDMA may under GPU RESET so its ring->ready may not true, keep going and GPU scheduler will reschedule this job if it failed.

give a warning on copy_buffer when go through direct_submit while ring->ready is false

NAK, that test has already saved us quite a bunch of trouble with the fb layer.

Why exactly are you running into issues with that?

Christian.

Quoted 23 lines
Change-Id Ife6cd55e0e843d99900e5bed5418499e88633685
---
       drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 6 +-----
       1 file changed, 1 insertion(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +1 -0
index e38e6db..7b75ac9 100644
@@ -1656,6 +1656,7 @@ int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
       	amdgpu_ring_pad_ib(ring, &job->ibs[0]);
       	WARN_ON(job->ibs[0].length_dw > num_dw);
       	if (direct_submit) {
+		WARN_ON(!ring->ready);
       		r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs,
       				       NULL, fence);
       		job->fence = dma_fence_get(*fence); @@ -1692,11 +1693,6 @@

int amdgpu_fill_buffer(struct amdgpu_bo *bo,

       	struct amdgpu_job *job;
       	int r;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +0 -5
-	if (!ring->ready) {
-		DRM_ERROR("Trying to clear memory with ring turned off.\n");
-		return -EINVAL;
-	}
-
       	if (bo->tbo.mem.mem_type == TTM_PL_TT) {
       		r = amdgpu_ttm_alloc_gart(&bo->tbo);
       		if (r)

_______________________________________________
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

_______________________________________________
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): 26eb2e4c-33dd-2160-5b83-a7ff9e3a2558-5C7GfCeVMHo@public.gmane.org

author
#41 lore raw

RE: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Oh yeah, I grep "ring->ready = false" and lost others ...

-----Original Message-----
From: Koenig, Christian
Sent: 2018年3月1日 17:51
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Look for amdgpu_ttm_set_active_vram_size(), setting ring->ready to false is directly beneath it.

Christian.

Am 01.03.2018 um 10:41 schrieb Liu, Monk:

Quoted 13 lines

Oops, looks only si will call this dma_stop() during hw_fini, I checked for sdma4.0/3.0/2.4, no such logic

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: 2018年3月1日 17:39
To: Liu, Monk <[email protected]>; Koenig, Christian <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Quoted 192 lines

Hi Monk,

look at si_dma_stop() and/or other SDMA generation code.

Regards,
Christian.

Am 01.03.2018 um 10:37 schrieb Liu, Monk:

Quoted 184 lines

Hi Christian

During suspend/resume, which stage will set sdma's ring->ready to false ?? I can only find it in amdgpu_ring_fini() But I saw that it is only invoked in unload_kms(), looks device_suspnd() will not set ring->ready to false

/Monk

-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Liu, Monk
Sent: 2018年3月1日 16:52
To: Koenig, Christian <[email protected]>; [email protected]
Subject: RE: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Quoted 1 line

amdgpu_evict_flags() and amdgpu_bo_move() needs to know if the SDMA ring is available during suspend/resume.

As long as we are in gpu reset, we don't return error when ring not ready, cuz eventually it either success or failed and rescheduled by scheduler since it is kernel job What's your concern ?

/Monk

-----Original Message-----
From: Koenig, Christian
Sent: 2018年3月1日 16:41
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Quoted 1 line

If we are in gpu rest stage, we can let SDMA job live with the case of ring not ready, can you point out where is the problem ?

amdgpu_evict_flags() and amdgpu_bo_move() needs to know if the SDMA ring is available during suspend/resume.

E.g. suspend works as this:
1. Evict all normal BOs from VRAM using the SDMA.
2. Disable SDMA, GART, PSP etc... e.g. disable the hardware.
3. Unpin all BOs in VRAM used by hardware engines, e.g. the GART table, firmware etc...
4. Evict all remaining BOs from VRAM using CPU copies.

Resume works the other way around:
1. Move all BOs for GART table and firmware back into VRAM using CPU copies.
2. Pin those BOs in VRAM and enable the hardware.
3. Use the SDMA to move back all other BOs into VRAM.

To figure out if we should use the CPU copy or the SDMA the ring->ready flag is used. So removing this check is really a no-go because it would break suspend/resume.

The only other alternative I can see is to add a separate flag in amdgpu_mman if buffer_funcs should be used or not. See all the callers of amdgpu_ttm_set_active_vram_size() as well.

Something like renaming amdgpu_ttm_set_active_vram_size() into amdgpu_ttm_set_buffer_funcs_read() and ignoring any change if gpu_reset is set.

BTW: Changing the active VRAM size in GPU reset is quite a bug you stumbled over here, so this really needs fixing anyway.

Regards,
Christian.

Am 01.03.2018 um 09:23 schrieb Liu, Monk:

Quoted 20 lines

Accel_working is only for GFX ring, I don't think test it is appropriate for SDMA jobs ....

If we are in gpu rest stage, we can let SDMA job live with the case of ring not ready, can you point out where is the problem ?

In fact we did stress TDR test with this patch and it can really fix couple over kill issues

/Monk

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: 2018年3月1日 16:16
To: Liu, Monk <[email protected]>; Koenig, Christian <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Ah, crap yeah that won't work since we don't free the ring.

Quoted 98 lines

Key point is we need to distinct between the ring doesn't work temporary because we are in a GPU reset and it doesn't work at all because we are missing firmware or stuff like that.

And no, checking the gpu_reset flag is totally racy and can't be done either. How about checking accel_working instead?

Christian.

Am 01.03.2018 um 07:01 schrieb Liu, Monk:

Quoted 91 lines
Quoted 1 line

Please change the test to use ring->ring_obj instead, this way we still bail out if somebody tries to submit commands before the ring is even allocated.

I don't understand how could fill_buffer() get run under the case that ring->ring_obj is not even allocated ... where is such case ?

/Monk

-----Original Message-----
From: Koenig, Christian
Sent: 2018年2月28日 20:46
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Good point, but in this case we need some other handling.

Please change the test to use ring->ring_obj instead, this way we still bail out if somebody tries to submit commands before the ring is even allocated.

And you also need to fix a couple of other places in amdgpu_ttm.c.

Regards,
Christian.

Am 28.02.2018 um 13:34 schrieb Liu, Monk:

Quoted 56 lines

Because when SDMA was hang by like process A, and meanwhile
another process B is already running into the code of fill_buffer() So just let process B continue, don't block it otherwise process B would fail by software reason .

Let it run and finally process B's job would fail and GPU recover will repeat it again (since it is a kernel job)

Without this solution other process will be greatly harmed by one black sheep that triggering GPU recover

/Monk

-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: 2018年2月28日 20:24
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 3/4] drm/amdgpu: don't return when ring not ready for fill_buffer

Am 28.02.2018 um 08:21 schrieb Monk Liu:

Quoted 6 lines

because this time SDMA may under GPU RESET so its ring->ready may not true, keep going and GPU scheduler will reschedule this job if it failed.

give a warning on copy_buffer when go through direct_submit while ring->ready is false

NAK, that test has already saved us quite a bunch of trouble with the fb layer.

Why exactly are you running into issues with that?

Christian.

Quoted 23 lines
Change-Id Ife6cd55e0e843d99900e5bed5418499e88633685
---
       drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 6 +-----
       1 file changed, 1 insertion(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +1 -0
index e38e6db..7b75ac9 100644
@@ -1656,6 +1656,7 @@ int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
       	amdgpu_ring_pad_ib(ring, &job->ibs[0]);
       	WARN_ON(job->ibs[0].length_dw > num_dw);
       	if (direct_submit) {
+		WARN_ON(!ring->ready);
       		r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs,
       				       NULL, fence);
       		job->fence = dma_fence_get(*fence); @@ -1692,11 +1693,6 

@@ int amdgpu_fill_buffer(struct amdgpu_bo *bo,

       	struct amdgpu_job *job;
       	int r;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +0 -5
-	if (!ring->ready) {
-		DRM_ERROR("Trying to clear memory with ring turned off.\n");
-		return -EINVAL;
-	}
-
       	if (bo->tbo.mem.mem_type == TTM_PL_TT) {
       		r = amdgpu_ttm_alloc_gart(&bo->tbo);
       		if (r)

_______________________________________________
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

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

author
#42 lore raw

PATCH 4/4 drm/amdgpu: try again kiq access if not in IRQ

In reply to a message not in the archive

sometimes GPU is switched to other VFs and won't swich back soon, so the kiq reg access will not signal within a short period, instead of returning TMO error we in fact should sleep 1ms and try again later (when not in interrupt context).

And since there is retry scheme provided so the MAX_KIQ_REG_WAIT tmo shouldn't set to a long time, set it to 10ms is enough.

if gpu already in reset state, don't retry the KIQ reg access otherwise it would always hang because KIQ was already die.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c | 15 +++++++++++++--
 1 file changed, 13 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c +13 -2
index b832651..7081eb6 100644
@@ -22,7 +22,7 @@
  */
 
 #include "amdgpu.h"
-#define MAX_KIQ_REG_WAIT	100000000 /* in usecs */
+#define MAX_KIQ_REG_WAIT	10000 /* in usecs, 10ms */
 
 uint64_t amdgpu_csa_vaddr(struct amdgpu_device *adev)
 {
@@ -152,9 +152,14 @@ uint32_t amdgpu_virt_kiq_rreg(struct amdgpu_device *adev, uint32_t reg)
 	amdgpu_ring_commit(ring);
 	spin_unlock_irqrestore(&kiq->ring_lock, flags);
 
+retry_read:
 	r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
 	if (r < 1) {
 		DRM_ERROR("wait for kiq fence error: %ld\n", r);
+		if (!in_interrupt() && !adev->in_gpu_reset) {
+			schedule();
+			goto retry_read;
+		}
 		return ~0;
 	}
 	val = adev->wb.wb[adev->virt.reg_val_offs];
@@ -179,9 +184,15 @@ void amdgpu_virt_kiq_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
 	amdgpu_ring_commit(ring);
 	spin_unlock_irqrestore(&kiq->ring_lock, flags);
 
+retry_write:
 	r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
-	if (r < 1)
+	if (r < 1) {
 		DRM_ERROR("wait for kiq fence error: %ld\n", r);
+		if (!in_interrupt() && !adev->in_gpu_reset) {
+			schedule();
+			goto retry_write;
+		}
+	}
 }
 
 /**
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): 1519802463-9090-4-git-send-email-Monk.Liu-5C7GfCeVMHo@public.gmane.org

author
#44 lore raw

RE: [PATCH 4/4] drm/amdgpu: try again kiq access if not in IRQ

Drop it, will send V2 patch

-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Monk Liu
Sent: 2018年2月28日 15:21
To: [email protected]
Cc: Liu, Monk <[email protected]>
Subject: [PATCH 4/4] drm/amdgpu: try again kiq access if not in IRQ

sometimes GPU is switched to other VFs and won't swich back soon, so the kiq reg access will not signal within a short period, instead of returning TMO error we in fact should sleep 1ms and try again later (when not in interrupt context).

And since there is retry scheme provided so the MAX_KIQ_REG_WAIT tmo shouldn't set to a long time, set it to 10ms is enough.

if gpu already in reset state, don't retry the KIQ reg access otherwise it would always hang because KIQ was already die.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c | 15 +++++++++++++--
 1 file changed, 13 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c +1 -1
index b832651..7081eb6 100644
@@ -22,7 +22,7 @@
  */
 
 #include "amdgpu.h"
-#define MAX_KIQ_REG_WAIT	100000000 /* in usecs */
+#define MAX_KIQ_REG_WAIT	10000 /* in usecs, 10ms */
 
 uint64_t amdgpu_csa_vaddr(struct amdgpu_device *adev)  { @@ -152,9 +152,14 @@ uint32_t amdgpu_virt_kiq_rreg(struct amdgpu_device *adev, uint32_t reg)
 	amdgpu_ring_commit(ring);

spin_unlock_irqrestore(&kiq->ring_lock, flags);

+retry_read:
r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
if (r < 1) {
DRM_ERROR("wait for kiq fence error: %ld\n", r);
+ if (!in_interrupt() && !adev->in_gpu_reset) {
+ schedule();
+ goto retry_read;
+ }
return ~0;
}
val = adev->wb.wb[adev->virt.reg_val_offs];
@@ -179,9 +184,15 @@ void amdgpu_virt_kiq_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
amdgpu_ring_commit(ring);
spin_unlock_irqrestore(&kiq->ring_lock, flags);

+retry_write:
r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
- if (r < 1)
+ if (r < 1) {
DRM_ERROR("wait for kiq fence error: %ld\n", r);
+ if (!in_interrupt() && !adev->in_gpu_reset) {
+ schedule();
+ goto retry_write;
+ }
+ }
}

/**
--
2.7.4

_______________________________________________
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

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

Andrey please give this set a good testing as well.

Am 28.02.2018 um 08:21 schrieb Monk Liu:

Quoted 6 lines

found recover_vram_from_shadow sometimes get executed in paralle with SDMA scheduler, should stop all schedulers before doing gpu reset/recover

Change-Id Ibaef3e3c015f3cf88f84b2eaf95cda95ae1a64e3

For now this patch is Reviewed-by: Christian König <[email protected]>.

Regards,
Christian.

Quoted 70 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 40 +++++++++++-------------------
  1 file changed, 15 insertions(+), 25 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +14 -19
index 75d1733..e9d81a8 100644
@@ -2649,22 +2649,23 @@ int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
 
  	/* block TTM */
  	resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
+
  	/* store modesetting */
  	if (amdgpu_device_has_dc_support(adev))
  		state = drm_atomic_helper_suspend(adev->ddev);
 
-	/* block scheduler */
+	/* block all schedulers and reset given job's ring */
  	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
  		struct amdgpu_ring *ring = adev->rings[i];
 
  		if (!ring || !ring->sched.thread)
  			continue;
 
-		/* only focus on the ring hit timeout if &job not NULL */
+		kthread_park(ring->sched.thread);
+
  		if (job && job->ring->idx != i)
  			continue;
 
-		kthread_park(ring->sched.thread);
  		drm_sched_hw_job_reset(&ring->sched, &job->base);
 
  		/* after all hw jobs are reset, hw fence is meaningless, so force_completion */
@@ -2707,33 +2708,22 @@ int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
  			}
  			dma_fence_put(fence);
  		}
+	}
 
-		for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
-			struct amdgpu_ring *ring = adev->rings[i];
-
-			if (!ring || !ring->sched.thread)
-				continue;
+	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
+		struct amdgpu_ring *ring = adev->rings[i];
 
-			/* only focus on the ring hit timeout if &job not NULL */
-			if (job && job->ring->idx != i)
-				continue;
+		if (!ring || !ring->sched.thread)
+			continue;
 
+		/* only need recovery sched of the given job's ring
+		 * or all rings (in the case @job is NULL)
+		 * after above amdgpu_reset accomplished
+		 */
+		if ((!job || job->ring->idx == i) && !r)
  			drm_sched_job_recovery(&ring->sched);
-			kthread_unpark(ring->sched.thread);
-		}
-	} else {
-		for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
-			struct amdgpu_ring *ring = adev->rings[i];
 
-			if (!ring || !ring->sched.thread)
-				continue;
-

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c int amdgpu_device_gpu_recover(struct amdgpu_device *adev, · line 2729

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1 -6
-			/* only focus on the ring hit timeout if &job not NULL */
-			if (job && job->ring->idx != i)
-				continue;
-
-			kthread_unpark(adev->rings[i]->sched.thread);
-		}
+		kthread_unpark(ring->sched.thread);
  	}
 
  	if (amdgpu_device_has_dc_support(adev)) {

_______________________________________________
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): 9e575c74-c6ce-76ef-a09c-1dec5a4807a3-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org

#47 lore raw

Will do once Monk sends V2 for  [PATCH 4/4] drm/amdgpu: try again kiq access if not in IRQ

Andrey

Quoted 79 lines · Christian König

Andrey please give this set a good testing as well.

Am 28.02.2018 um 08:21 schrieb Monk Liu:

Quoted 6 lines

found recover_vram_from_shadow sometimes get executed in paralle with SDMA scheduler, should stop all schedulers before doing gpu reset/recover

Change-Id Ibaef3e3c015f3cf88f84b2eaf95cda95ae1a64e3

For now this patch is Reviewed-by: Christian König <[email protected]>.

Regards,
Christian.

Quoted 64 lines

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 40
+++++++++++-------------------

  1 file changed, 15 insertions(+), 25 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1 -0
index 75d1733..e9d81a8 100644
@@ -2649,22 +2649,23 @@ int amdgpu_device_gpu_recover(struct 
 amdgpu_device *adev,
         /* block TTM */
       resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
+
       /* store modesetting */
       if (amdgpu_device_has_dc_support(adev))
           state = drm_atomic_helper_suspend(adev->ddev);

  -    /* block scheduler */
+    /* block all schedulers and reset given job's ring */
      for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
          struct amdgpu_ring *ring = adev->rings[i];
            if (!ring || !ring->sched.thread)
              continue;
  -        /* only focus on the ring hit timeout if &job not NULL */
+        kthread_park(ring->sched.thread);
+
          if (job && job->ring->idx != i)
              continue;
  -        kthread_park(ring->sched.thread); drm_sched_hw_job_reset(&ring->sched, &job->base);
            /* after all hw jobs are reset, hw fence is meaningless, so force_completion */

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +3 -4
@@ -2707,33 +2708,22 @@ int amdgpu_device_gpu_recover(struct 
 amdgpu_device *adev,
               }
               dma_fence_put(fence);
           }
+    }
   -        for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
-            struct amdgpu_ring *ring = adev->rings[i];
-
-            if (!ring || !ring->sched.thread)
-                continue;
+    for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
+        struct amdgpu_ring *ring = adev->rings[i];

  -            /* only focus on the ring hit timeout if &job not NULL */ -            if (job && job->ring->idx != i)
-                continue;
+        if (!ring || !ring->sched.thread)
+            continue;
  +        /* only need recovery sched of the given job's ring +         * or all rings (in the case @job is NULL)
+         * after above amdgpu_reset accomplished
+         */
+        if ((!job || job->ring->idx == i) && !r) drm_sched_job_recovery(&ring->sched);
-            kthread_unpark(ring->sched.thread);
-        }
-    } else {
-        for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
-            struct amdgpu_ring *ring = adev->rings[i];
  -            if (!ring || !ring->sched.thread)
-                continue;
-
-            /* only focus on the ring hit timeout if &job not NULL */

Quoted 9 lines
Quoted 8 lines

-            if (job && job->ring->idx != i)
-                continue;
-
-            kthread_unpark(adev->rings[i]->sched.thread); -        }
+        kthread_unpark(ring->sched.thread);
      }
        if (amdgpu_device_has_dc_support(adev)) {

_______________________________________________
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): e4756f8b-d8f9-9849-aad4-a23193e367f6-5C7GfCeVMHo@public.gmane.org

author
#49 lore raw

Already sent

-----Original Message-----
From: Grodzovsky, Andrey
Sent: 2018年2月28日 21:31
To: Koenig, Christian <[email protected]>; Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 1/4] drm/amdgpu: stop all rings before doing gpu recover

Will do once Monk sends V2 for  [PATCH 4/4] drm/amdgpu: try again kiq access if not in IRQ

Andrey

Quoted 83 lines · Christian König

Andrey please give this set a good testing as well.

Am 28.02.2018 um 08:21 schrieb Monk Liu:

Quoted 6 lines

found recover_vram_from_shadow sometimes get executed in paralle with SDMA scheduler, should stop all schedulers before doing gpu reset/recover

Change-Id Ibaef3e3c015f3cf88f84b2eaf95cda95ae1a64e3

For now this patch is Reviewed-by: Christian König <[email protected]>.

Regards,
Christian.

Quoted 67 lines

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 40
+++++++++++-------------------

  1 file changed, 15 insertions(+), 25 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1 -0
index 75d1733..e9d81a8 100644
@@ -2649,22 +2649,23 @@ int amdgpu_device_gpu_recover(struct 
 amdgpu_device *adev,
         /* block TTM */
       resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
+
       /* store modesetting */
       if (amdgpu_device_has_dc_support(adev))
           state = drm_atomic_helper_suspend(adev->ddev);

  -    /* block scheduler */
+    /* block all schedulers and reset given job's ring */ for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
          struct amdgpu_ring *ring = adev->rings[i];
            if (!ring || !ring->sched.thread)
              continue;
  -        /* only focus on the ring hit timeout if &job not NULL */ +        kthread_park(ring->sched.thread);
+
          if (job && job->ring->idx != i)
              continue;
  -        kthread_park(ring->sched.thread); drm_sched_hw_job_reset(&ring->sched, &job->base); /* after all hw jobs are reset, hw fence is meaningless, so force_completion */ @@ -2707,33 +2708,22 @@ int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
              }
              dma_fence_put(fence);
          }
+    }
  -        for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
-            struct amdgpu_ring *ring = adev->rings[i];
-
-            if (!ring || !ring->sched.thread)
-                continue;
+    for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
+        struct amdgpu_ring *ring = adev->rings[i];
  -            /* only focus on the ring hit timeout if &job not NULL */
-            if (job && job->ring->idx != i)
-                continue;
+        if (!ring || !ring->sched.thread)
+            continue;
  +        /* only need recovery sched of the given job's ring +         * or all rings (in the case @job is NULL)
+         * after above amdgpu_reset accomplished
+         */
+        if ((!job || job->ring->idx == i) && !r) drm_sched_job_recovery(&ring->sched);
-            kthread_unpark(ring->sched.thread);
-        }
-    } else {
-        for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
-            struct amdgpu_ring *ring = adev->rings[i];
  -            if (!ring || !ring->sched.thread)
-                continue;
-
-            /* only focus on the ring hit timeout if &job not NULL */

Quoted 9 lines
Quoted 8 lines

-            if (job && job->ring->idx != i)
-                continue;
-
-            kthread_unpark(adev->rings[i]->sched.thread); -        }
+        kthread_unpark(ring->sched.thread);
      }
        if (amdgpu_device_has_dc_support(adev)) {

_______________________________________________
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): BLUPR12MB04491C15749E9DCAB3DEF1A684C70-7LeqcoF/[email protected]

#51 lore raw

Found it.

Thanks.

Quoted 96 lines · Monk

Already sent

-----Original Message-----
From: Grodzovsky, Andrey
Sent: 2018年2月28日 21:31
To: Koenig, Christian <[email protected]>; Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 1/4] drm/amdgpu: stop all rings before doing gpu recover

Will do once Monk sends V2 for  [PATCH 4/4] drm/amdgpu: try again kiq access if not in IRQ

Andrey

Quoted 82 lines · Christian König

Andrey please give this set a good testing as well.

Am 28.02.2018 um 08:21 schrieb Monk Liu:

Quoted 6 lines

found recover_vram_from_shadow sometimes get executed in paralle with SDMA scheduler, should stop all schedulers before doing gpu reset/recover

Change-Id Ibaef3e3c015f3cf88f84b2eaf95cda95ae1a64e3

For now this patch is Reviewed-by: Christian König <[email protected]>.

Regards,
Christian.

Quoted 67 lines
---
   drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 40
+++++++++++-------------------
   1 file changed, 15 insertions(+), 25 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +4 -0
index 75d1733..e9d81a8 100644
@@ -2649,22 +2649,23 @@ int amdgpu_device_gpu_recover(struct
 amdgpu_device *adev,
         /* block TTM */
       resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
+
       /* store modesetting */
       if (amdgpu_device_has_dc_support(adev))
           state = drm_atomic_helper_suspend(adev->ddev);
   -    /* block scheduler */
+    /* block all schedulers and reset given job's ring */
       for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
           struct amdgpu_ring *ring = adev->rings[i];
             if (!ring || !ring->sched.thread)
               continue;
   -        /* only focus on the ring hit timeout if &job not NULL */
+        kthread_park(ring->sched.thread);
+
           if (job && job->ring->idx != i)
               continue;
   -        kthread_park(ring->sched.thread);
           drm_sched_hw_job_reset(&ring->sched, &job->base);
             /* after all hw jobs are reset, hw fence is meaningless,

so force_completion */ @@ -2707,33 +2708,22 @@ int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
              }
              dma_fence_put(fence);
          }
+    }
  -        for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
-            struct amdgpu_ring *ring = adev->rings[i];
-
-            if (!ring || !ring->sched.thread)
-                continue;
+    for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
+        struct amdgpu_ring *ring = adev->rings[i];
  -            /* only focus on the ring hit timeout if &job not NULL
*/
-            if (job && job->ring->idx != i)
-                continue;
+        if (!ring || !ring->sched.thread)
+            continue;
  +        /* only need recovery sched of the given job's ring
+         * or all rings (in the case @job is NULL)
+         * after above amdgpu_reset accomplished
+         */
+        if ((!job || job->ring->idx == i) && !r)
              drm_sched_job_recovery(&ring->sched);
-            kthread_unpark(ring->sched.thread);
-        }
-    } else {
-        for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
-            struct amdgpu_ring *ring = adev->rings[i];
  -            if (!ring || !ring->sched.thread)
-                continue;
-
-            /* only focus on the ring hit timeout if &job not NULL */

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines

-            if (job && job->ring->idx != i)
-                continue;
-
-            kthread_unpark(adev->rings[i]->sched.thread); -        }
+        kthread_unpark(ring->sched.thread);
      }
        if (amdgpu_device_has_dc_support(adev)) {

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

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

No new issues found with those patches, testing GPU reset using libdrm deadlock detection test on Ellsmire.

The patches are Tested-By: Andrey Grodzovsky <[email protected]>

P.S

Noticed existing issues (before Monk's patches)

Multiple [drm:amdgpu_cs_ioctl [amdgpu]] *ERROR* Failed to initialize parser !

And occasional unlock imbalance forn amdgpu_cs_ioctl

DEBUG_LOCKS_WARN_ON(depth <= 0)
[   93.069011 <    0.000017>] WARNING: CPU: 3 PID: 2215 at kernel/locking/lockdep.c:3682 lock_release+0x2e8/0x360

On CZ full reset hangs the system.

Gonna take a look at those issues.

Thanks,

Andrey

From

Quoted 96 lines · Monk

Already sent

-----Original Message-----
From: Grodzovsky, Andrey
Sent: 2018年2月28日 21:31
To: Koenig, Christian <[email protected]>; Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 1/4] drm/amdgpu: stop all rings before doing gpu recover

Will do once Monk sends V2 for  [PATCH 4/4] drm/amdgpu: try again kiq access if not in IRQ

Andrey

Quoted 82 lines · Christian König

Andrey please give this set a good testing as well.

Am 28.02.2018 um 08:21 schrieb Monk Liu:

Quoted 6 lines

found recover_vram_from_shadow sometimes get executed in paralle with SDMA scheduler, should stop all schedulers before doing gpu reset/recover

Change-Id Ibaef3e3c015f3cf88f84b2eaf95cda95ae1a64e3

For now this patch is Reviewed-by: Christian König <[email protected]>.

Regards,
Christian.

Quoted 67 lines
---
   drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 40
+++++++++++-------------------
   1 file changed, 15 insertions(+), 25 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +4 -0
index 75d1733..e9d81a8 100644
@@ -2649,22 +2649,23 @@ int amdgpu_device_gpu_recover(struct
 amdgpu_device *adev,
         /* block TTM */
       resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
+
       /* store modesetting */
       if (amdgpu_device_has_dc_support(adev))
           state = drm_atomic_helper_suspend(adev->ddev);
   -    /* block scheduler */
+    /* block all schedulers and reset given job's ring */
       for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
           struct amdgpu_ring *ring = adev->rings[i];
             if (!ring || !ring->sched.thread)
               continue;
   -        /* only focus on the ring hit timeout if &job not NULL */
+        kthread_park(ring->sched.thread);
+
           if (job && job->ring->idx != i)
               continue;
   -        kthread_park(ring->sched.thread);
           drm_sched_hw_job_reset(&ring->sched, &job->base);
             /* after all hw jobs are reset, hw fence is meaningless,

so force_completion */ @@ -2707,33 +2708,22 @@ int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
              }
              dma_fence_put(fence);
          }
+    }
  -        for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
-            struct amdgpu_ring *ring = adev->rings[i];
-
-            if (!ring || !ring->sched.thread)
-                continue;
+    for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
+        struct amdgpu_ring *ring = adev->rings[i];
  -            /* only focus on the ring hit timeout if &job not NULL
*/
-            if (job && job->ring->idx != i)
-                continue;
+        if (!ring || !ring->sched.thread)
+            continue;
  +        /* only need recovery sched of the given job's ring
+         * or all rings (in the case @job is NULL)
+         * after above amdgpu_reset accomplished
+         */
+        if ((!job || job->ring->idx == i) && !r)
              drm_sched_job_recovery(&ring->sched);
-            kthread_unpark(ring->sched.thread);
-        }
-    } else {
-        for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
-            struct amdgpu_ring *ring = adev->rings[i];
  -            if (!ring || !ring->sched.thread)
-                continue;
-
-            /* only focus on the ring hit timeout if &job not NULL */

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines

-            if (job && job->ring->idx != i)
-                continue;
-
-            kthread_unpark(adev->rings[i]->sched.thread); -        }
+        kthread_unpark(ring->sched.thread);
      }
        if (amdgpu_device_has_dc_support(adev)) {

_______________________________________________
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): a54b5a4f-a370-87ec-7bac-33e6036107f9-5C7GfCeVMHo@public.gmane.org

#54 lore raw

After looking more

"Failed to initialize parser !" is expected since the reason is jobs from guilty context (context causing the GPU hang) are canceled.

Locking imbalance is actually gone with Monk's patches.

Thanks,

Andrey

Quoted 122 lines · Andrey Grodzovsky

No new issues found with those patches, testing GPU reset using libdrm deadlock detection test on Ellsmire.

The patches are Tested-By: Andrey Grodzovsky <[email protected]>

P.S

Noticed existing issues (before Monk's patches)

Multiple [drm:amdgpu_cs_ioctl [amdgpu]] *ERROR* Failed to initialize parser !

And occasional unlock imbalance forn amdgpu_cs_ioctl

DEBUG_LOCKS_WARN_ON(depth <= 0)
[   93.069011 <    0.000017>] WARNING: CPU: 3 PID: 2215 at kernel/locking/lockdep.c:3682 lock_release+0x2e8/0x360

On CZ full reset hangs the system.

Gonna take a look at those issues.

Thanks,

Andrey

From

Quoted 96 lines · Monk

Already sent

-----Original Message-----
From: Grodzovsky, Andrey
Sent: 2018年2月28日 21:31
To: Koenig, Christian <[email protected]>; Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH 1/4] drm/amdgpu: stop all rings before doing gpu recover

Will do once Monk sends V2 for  [PATCH 4/4] drm/amdgpu: try again kiq access if not in IRQ

Andrey

Quoted 82 lines · Christian König

Andrey please give this set a good testing as well.

Am 28.02.2018 um 08:21 schrieb Monk Liu:

Quoted 6 lines

found recover_vram_from_shadow sometimes get executed in paralle with SDMA scheduler, should stop all schedulers before doing gpu reset/recover

Change-Id Ibaef3e3c015f3cf88f84b2eaf95cda95ae1a64e3

For now this patch is Reviewed-by: Christian König <[email protected]>.

Regards,
Christian.

Quoted 67 lines

---
   drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 40
+++++++++++-------------------

   1 file changed, 15 insertions(+), 25 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1 -0
index 75d1733..e9d81a8 100644
@@ -2649,22 +2649,23 @@ int amdgpu_device_gpu_recover(struct
 amdgpu_device *adev,
          /* block TTM */
        resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
+
        /* store modesetting */
        if (amdgpu_device_has_dc_support(adev))
            state = drm_atomic_helper_suspend(adev->ddev);

   -    /* block scheduler */
+    /* block all schedulers and reset given job's ring */ for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
           struct amdgpu_ring *ring = adev->rings[i];
             if (!ring || !ring->sched.thread)
               continue;
   -        /* only focus on the ring hit timeout if &job not NULL */ +        kthread_park(ring->sched.thread);
+
           if (job && job->ring->idx != i)
               continue;
   -        kthread_park(ring->sched.thread); drm_sched_hw_job_reset(&ring->sched, &job->base); /* after all hw jobs are reset, hw fence is meaningless, so force_completion */ @@ -2707,33 +2708,22 @@ int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
               }
               dma_fence_put(fence);
           }
+    }
   -        for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
-            struct amdgpu_ring *ring = adev->rings[i];
-
-            if (!ring || !ring->sched.thread)
-                continue;
+    for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
+        struct amdgpu_ring *ring = adev->rings[i];
   -            /* only focus on the ring hit timeout if &job not NULL */
-            if (job && job->ring->idx != i)
-                continue;
+        if (!ring || !ring->sched.thread)
+            continue;
   +        /* only need recovery sched of the given job's ring +         * or all rings (in the case @job is NULL)
+         * after above amdgpu_reset accomplished
+         */
+        if ((!job || job->ring->idx == i) && !r) drm_sched_job_recovery(&ring->sched);
-            kthread_unpark(ring->sched.thread);
-        }
-    } else {
-        for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
-            struct amdgpu_ring *ring = adev->rings[i];
   -            if (!ring || !ring->sched.thread)
-                continue;
-
-            /* only focus on the ring hit timeout if &job not NULL */

Quoted 9 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines

-            if (job && job->ring->idx != i) -                continue;
-
- kthread_unpark(adev->rings[i]->sched.thread); -        }
+        kthread_unpark(ring->sched.thread); }
         if (amdgpu_device_has_dc_support(adev)) {

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