Skip to contentSkip to conversation

drm/amdgpu:fix IH overflow on Cz

PATCH In review Defang Bo 2 participants 2 messages Jan 5, 2021

Subsystems (2)
lore
Share & tools
lore

Defang Bo

author
#0 lore raw

PATCH v2 1/1 drm/amdgpu:fix IH overflow on Cz

Similar to commit <b82175750131>("drm/amdgpu: fix IH overflow on Vega10 v2"). When an ring buffer overflow happens the appropriate bit is set in the WPTR register which is also written back to memory. But clearing the bit in the WPTR doesn't trigger another memory writeback.

So what can happen is that we end up processing the buffer overflow over and over again because the bit is never cleared. Resulting in a random system lockup because of an infinite loop in an interrupt handler.

---
Changes since v1:
- Modify the code and correct the wrong fix.
---

---
 drivers/gpu/drm/amd/amdgpu/cz_ih.c | 38 ++++++++++++++++++++++++--------------
 1 file changed, 24 insertions(+), 14 deletions(-)
drivers/gpu/drm/amd/amdgpu/cz_ih.c +24 -14
index 1dca0cabc326..45dd47f45fa2 100644
@@ -190,22 +190,32 @@ static u32 cz_ih_get_wptr(struct amdgpu_device *adev,
 			  struct amdgpu_ih_ring *ih)
 {
 	u32 wptr, tmp;
-
+
 	wptr = le32_to_cpu(*ih->wptr_cpu);
 
-	if (REG_GET_FIELD(wptr, IH_RB_WPTR, RB_OVERFLOW)) {
-		wptr = REG_SET_FIELD(wptr, IH_RB_WPTR, RB_OVERFLOW, 0);
-		/* When a ring buffer overflow happen start parsing interrupt
-		 * from the last not overwritten vector (wptr + 16). Hopefully
-		 * this should allow us to catchup.
-		 */
-		dev_warn(adev->dev, "IH ring buffer overflow (0x%08X, 0x%08X, 0x%08X)\n",
-			wptr, ih->rptr, (wptr + 16) & ih->ptr_mask);
-		ih->rptr = (wptr + 16) & ih->ptr_mask;
-		tmp = RREG32(mmIH_RB_CNTL);
-		tmp = REG_SET_FIELD(tmp, IH_RB_CNTL, WPTR_OVERFLOW_CLEAR, 1);
-		WREG32(mmIH_RB_CNTL, tmp);
-	}
+	if (!REG_GET_FIELD(wptr, IH_RB_WPTR, RB_OVERFLOW))
+		goto out;
+
+	wptr = RREG32(mmIH_RB_CNTL);
+
+	if (!REG_GET_FIELD(wptr, IH_RB_WPTR, RB_OVERFLOW))
+		goto out;
+
+	wptr = REG_SET_FIELD(wptr, IH_RB_WPTR, RB_OVERFLOW, 0);
+
+	/* When a ring buffer overflow happen start parsing interrupt
+	 * from the last not overwritten vector (wptr + 16). Hopefully
+	 * this should allow us to catchup.
+	 */
+	dev_warn(adev->dev, "IH ring buffer overflow (0x%08X, 0x%08X, 0x%08X)\n",
+		wptr, ih->rptr, (wptr + 16) & ih->ptr_mask);
+	ih->rptr = (wptr + 16) & ih->ptr_mask;
+	tmp = RREG32(mmIH_RB_CNTL);
+	tmp = REG_SET_FIELD(tmp, IH_RB_CNTL, WPTR_OVERFLOW_CLEAR, 1);
+	WREG32(mmIH_RB_CNTL, tmp);
+
+
+out:
 	return (wptr & ih->ptr_mask);
 }
 
Signature · 6 lines
2.7.4

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
maintainer
#1 lore raw

Am 05.01.21 um 08:32 schrieb Defang Bo:

Quoted 8 lines

Similar to commit <b82175750131>("drm/amdgpu: fix IH overflow on Vega10 v2"). When an ring buffer overflow happens the appropriate bit is set in the WPTR register which is also written back to memory. But clearing the bit in the WPTR doesn't trigger another memory writeback.

So what can happen is that we end up processing the buffer overflow over and over again because the bit is never cleared. Resulting in a random system lockup because of an infinite loop in an interrupt handler.

Really good point. I haven't had time to look into other generations since fixing this for Vega.

One major typo below which needs to be fixed.

Quoted 29 lines

---
Changes since v1:
- Modify the code and correct the wrong fix.
---

---
  drivers/gpu/drm/amd/amdgpu/cz_ih.c | 38 ++++++++++++++++++++++++--------------
  1 file changed, 24 insertions(+), 14 deletions(-)
drivers/gpu/drm/amd/amdgpu/cz_ih.c +1 -8
index 1dca0cabc326..45dd47f45fa2 100644
@@ -190,22 +190,32 @@ static u32 cz_ih_get_wptr(struct amdgpu_device *adev,
  			  struct amdgpu_ih_ring *ih)
  {
  	u32 wptr, tmp;
-
+
  	wptr = le32_to_cpu(*ih->wptr_cpu);
 
-	if (REG_GET_FIELD(wptr, IH_RB_WPTR, RB_OVERFLOW)) {
-		wptr = REG_SET_FIELD(wptr, IH_RB_WPTR, RB_OVERFLOW, 0);
-		/* When a ring buffer overflow happen start parsing interrupt
-		 * from the last not overwritten vector (wptr + 16). Hopefully
-		 * this should allow us to catchup.
-		 */
-		dev_warn(adev->dev, "IH ring buffer overflow (0x%08X, 0x%08X, 0x%08X)\n",

drivers/gpu/drm/amd/amdgpu/cz_ih.c static u32 cz_ih_get_wptr(struct amdgpu_device *adev, · line 199

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/cz_ih.c +4 -6
-			wptr, ih->rptr, (wptr + 16) & ih->ptr_mask);
-		ih->rptr = (wptr + 16) & ih->ptr_mask;
-		tmp = RREG32(mmIH_RB_CNTL);
-		tmp = REG_SET_FIELD(tmp, IH_RB_CNTL, WPTR_OVERFLOW_CLEAR, 1);
-		WREG32(mmIH_RB_CNTL, tmp);
-	}
+	if (!REG_GET_FIELD(wptr, IH_RB_WPTR, RB_OVERFLOW))
+		goto out;
+
+	wptr = RREG32(mmIH_RB_CNTL);

That's the wrong register, you need to read the write pointer and not the control register.

Same problem in all other patches.

Regards,
Christian.

Quoted 12 lines
drivers/gpu/drm/amd/amdgpu/cz_ih.c +12 -0
+
+	if (!REG_GET_FIELD(wptr, IH_RB_WPTR, RB_OVERFLOW))
+		goto out;
+
+	wptr = REG_SET_FIELD(wptr, IH_RB_WPTR, RB_OVERFLOW, 0);
+
+	/* When a ring buffer overflow happen start parsing interrupt
+	 * from the last not overwritten vector (wptr + 16). Hopefully
+	 * this should allow us to catchup.
+	 */
+	dev_warn(adev->dev, "IH ring buffer overflow (0x%08X, 0x%08X, 0x%08X)\n",
+		wptr, ih->rptr, (wptr + 16) & ih->ptr_mask);

drivers/gpu/drm/amd/amdgpu/cz_ih.c static u32 cz_ih_get_wptr(struct amdgpu_device *adev, · line 221

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/cz_ih.c +7 -0
+	ih->rptr = (wptr + 16) & ih->ptr_mask;
+	tmp = RREG32(mmIH_RB_CNTL);
+	tmp = REG_SET_FIELD(tmp, IH_RB_CNTL, WPTR_OVERFLOW_CLEAR, 1);
+	WREG32(mmIH_RB_CNTL, tmp);
+
+
+out:
  	return (wptr & ih->ptr_mask);
  }
 

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