Skip to contentSkip to conversation

drm/radeon/kms/evergreen: fix gpu hangs in userspace accel code

PATCH In review Alex Deucher 2 participants 4 messages Sep 3, 2010

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 drm/radeon/kms/evergreen: fix gpu hangs in userspace accel code

These VGT regs need to be programmed via the ring rather than MMIO as on previous asics (r6xx/r7xx).

---
 drivers/gpu/drm/radeon/evergreen.c |   39 +++++++++++++++++++++++++++++++++++-
 drivers/gpu/drm/radeon/r600.c      |    5 +---
 2 files changed, 39 insertions(+), 5 deletions(-)
drivers/gpu/drm/radeon/evergreen.c +38 -1
index 875d891..7b81e5c 100644
@@ -1110,6 +1110,43 @@ static int evergreen_cp_load_microcode(struct radeon_device *rdev)
 	return 0;
 }
 
+static int evergreen_cp_start(struct radeon_device *rdev)
+{
+	int r;
+	uint32_t cp_me;
+
+	r = radeon_ring_lock(rdev, 7);
+	if (r) {
+		DRM_ERROR("radeon: cp failed to lock ring (%d).\n", r);
+		return r;
+	}
+	radeon_ring_write(rdev, PACKET3(PACKET3_ME_INITIALIZE, 5));
+	radeon_ring_write(rdev, 0x1);
+	radeon_ring_write(rdev, 0x0);
+	radeon_ring_write(rdev, rdev->config.evergreen.max_hw_contexts - 1);
+	radeon_ring_write(rdev, PACKET3_ME_INITIALIZE_DEVICE_ID(1));
+	radeon_ring_write(rdev, 0);
+	radeon_ring_write(rdev, 0);
+	radeon_ring_unlock_commit(rdev);
+
+	cp_me = 0xff;
+	WREG32(CP_ME_CNTL, cp_me);
+
+	r = radeon_ring_lock(rdev, 4);
+	if (r) {
+		DRM_ERROR("radeon: cp failed to lock ring (%d).\n", r);
+		return r;
+	}
+	/* init some VGT regs */
+	radeon_ring_write(rdev, PACKET3(PACKET3_SET_CONTEXT_REG, 2));
+	radeon_ring_write(rdev, (VGT_VERTEX_REUSE_BLOCK_CNTL - PACKET3_SET_CONTEXT_REG_START) >> 2);
+	radeon_ring_write(rdev, 0xe);
+	radeon_ring_write(rdev, 0x10);
+	radeon_ring_unlock_commit(rdev);
+
+	return 0;
+}
+
 int evergreen_cp_resume(struct radeon_device *rdev)
 {
 	u32 tmp;
@@ -1165,7 +1202,7 @@ int evergreen_cp_resume(struct radeon_device *rdev)
 	rdev->cp.rptr = RREG32(CP_RB_RPTR);
 	rdev->cp.wptr = RREG32(CP_RB_WPTR);
 
-	r600_cp_start(rdev);
+	evergreen_cp_start(rdev);
 	rdev->cp.ready = true;
 	r = radeon_ring_test(rdev);
 	if (r) {
drivers/gpu/drm/radeon/r600.c +1 -4
index 73e6a1a..f280717 100644
@@ -2120,10 +2120,7 @@ int r600_cp_start(struct radeon_device *rdev)
 	}
 	radeon_ring_write(rdev, PACKET3(PACKET3_ME_INITIALIZE, 5));
 	radeon_ring_write(rdev, 0x1);
-	if (rdev->family >= CHIP_CEDAR) {
-		radeon_ring_write(rdev, 0x0);
-		radeon_ring_write(rdev, rdev->config.evergreen.max_hw_contexts - 1);
-	} else if (rdev->family >= CHIP_RV770) {
+	if (rdev->family >= CHIP_RV770) {
 		radeon_ring_write(rdev, 0x0);
 		radeon_ring_write(rdev, rdev->config.rv770.max_hw_contexts - 1);
 	} else {
Signature · 1 line
1.7.1.1
#1 lore raw
Quoted 3 lines · Alex Deucher

These VGT regs need to be programmed via the ring rather than MMIO as on previous asics (r6xx/r7xx).

So we need someway for userspace to know this has happened, either a version bump or a change to the info ioctl again.

Dave.

Quoted 79 lines
---
 drivers/gpu/drm/radeon/evergreen.c |   39 +++++++++++++++++++++++++++++++++++-
 drivers/gpu/drm/radeon/r600.c      |    5 +---
 2 files changed, 39 insertions(+), 5 deletions(-)
drivers/gpu/drm/radeon/evergreen.c +38 -1
index 875d891..7b81e5c 100644
@@ -1110,6 +1110,43 @@ static int evergreen_cp_load_microcode(struct radeon_device *rdev)
 	return 0;
 }
 
+static int evergreen_cp_start(struct radeon_device *rdev)
+{
+	int r;
+	uint32_t cp_me;
+
+	r = radeon_ring_lock(rdev, 7);
+	if (r) {
+		DRM_ERROR("radeon: cp failed to lock ring (%d).\n", r);
+		return r;
+	}
+	radeon_ring_write(rdev, PACKET3(PACKET3_ME_INITIALIZE, 5));
+	radeon_ring_write(rdev, 0x1);
+	radeon_ring_write(rdev, 0x0);
+	radeon_ring_write(rdev, rdev->config.evergreen.max_hw_contexts - 1);
+	radeon_ring_write(rdev, PACKET3_ME_INITIALIZE_DEVICE_ID(1));
+	radeon_ring_write(rdev, 0);
+	radeon_ring_write(rdev, 0);
+	radeon_ring_unlock_commit(rdev);
+
+	cp_me = 0xff;
+	WREG32(CP_ME_CNTL, cp_me);
+
+	r = radeon_ring_lock(rdev, 4);
+	if (r) {
+		DRM_ERROR("radeon: cp failed to lock ring (%d).\n", r);
+		return r;
+	}
+	/* init some VGT regs */
+	radeon_ring_write(rdev, PACKET3(PACKET3_SET_CONTEXT_REG, 2));
+	radeon_ring_write(rdev, (VGT_VERTEX_REUSE_BLOCK_CNTL - PACKET3_SET_CONTEXT_REG_START) >> 2);
+	radeon_ring_write(rdev, 0xe);
+	radeon_ring_write(rdev, 0x10);
+	radeon_ring_unlock_commit(rdev);
+
+	return 0;
+}
+
 int evergreen_cp_resume(struct radeon_device *rdev)
 {
 	u32 tmp;
@@ -1165,7 +1202,7 @@ int evergreen_cp_resume(struct radeon_device *rdev)
 	rdev->cp.rptr = RREG32(CP_RB_RPTR);
 	rdev->cp.wptr = RREG32(CP_RB_WPTR);
 
-	r600_cp_start(rdev);
+	evergreen_cp_start(rdev);
 	rdev->cp.ready = true;
 	r = radeon_ring_test(rdev);
 	if (r) {
drivers/gpu/drm/radeon/r600.c +1 -4
index 73e6a1a..f280717 100644
@@ -2120,10 +2120,7 @@ int r600_cp_start(struct radeon_device *rdev)
 	}
 	radeon_ring_write(rdev, PACKET3(PACKET3_ME_INITIALIZE, 5));
 	radeon_ring_write(rdev, 0x1);
-	if (rdev->family >= CHIP_CEDAR) {
-		radeon_ring_write(rdev, 0x0);
-		radeon_ring_write(rdev, rdev->config.evergreen.max_hw_contexts - 1);
-	} else if (rdev->family >= CHIP_RV770) {
+	if (rdev->family >= CHIP_RV770) {
 		radeon_ring_write(rdev, 0x0);
 		radeon_ring_write(rdev, rdev->config.rv770.max_hw_contexts - 1);
 	} else {
#2 lore raw
Quoted 7 lines · Dave Airlie
Quoted 3 lines · Alex Deucher

These VGT regs need to be programmed via the ring rather than MMIO as on previous asics (r6xx/r7xx).

So we need someway for userspace to know this has happened, either a version bump or a change to the info ioctl again.

We can do a version bump for 2.6.36. Not sure what the best option is for 2.6.35.

Alex

Quoted 85 lines

Dave.

Quoted 79 lines

---
 drivers/gpu/drm/radeon/evergreen.c |   39 +++++++++++++++++++++++++++++++++++- drivers/gpu/drm/radeon/r600.c      |    5 +---

 2 files changed, 39 insertions(+), 5 deletions(-)
drivers/gpu/drm/radeon/evergreen.c +38 -1
index 875d891..7b81e5c 100644
@@ -1110,6 +1110,43 @@ static int evergreen_cp_load_microcode(struct radeon_device *rdev)
       return 0;
  }
 
+static int evergreen_cp_start(struct radeon_device *rdev)
+{
+     int r;
+     uint32_t cp_me;
+
+     r = radeon_ring_lock(rdev, 7);
+     if (r) {
+             DRM_ERROR("radeon: cp failed to lock ring (%d).\n", r);
+             return r;
+     }
+     radeon_ring_write(rdev, PACKET3(PACKET3_ME_INITIALIZE, 5));
+     radeon_ring_write(rdev, 0x1);
+     radeon_ring_write(rdev, 0x0);
+     radeon_ring_write(rdev, rdev->config.evergreen.max_hw_contexts - 1);
+     radeon_ring_write(rdev, PACKET3_ME_INITIALIZE_DEVICE_ID(1));
+     radeon_ring_write(rdev, 0);
+     radeon_ring_write(rdev, 0);
+     radeon_ring_unlock_commit(rdev);
+
+     cp_me = 0xff;
+     WREG32(CP_ME_CNTL, cp_me);
+
+     r = radeon_ring_lock(rdev, 4);
+     if (r) {
+             DRM_ERROR("radeon: cp failed to lock ring (%d).\n", r);
+             return r;
+     }
+     /* init some VGT regs */
+     radeon_ring_write(rdev, PACKET3(PACKET3_SET_CONTEXT_REG, 2));
+     radeon_ring_write(rdev, (VGT_VERTEX_REUSE_BLOCK_CNTL - PACKET3_SET_CONTEXT_REG_START) >> 2);
+     radeon_ring_write(rdev, 0xe);
+     radeon_ring_write(rdev, 0x10);
+     radeon_ring_unlock_commit(rdev);
+
+     return 0;
+}
+
  int evergreen_cp_resume(struct radeon_device *rdev)
  {
       u32 tmp;
@@ -1165,7 +1202,7 @@ int evergreen_cp_resume(struct radeon_device *rdev)
       rdev->cp.rptr = RREG32(CP_RB_RPTR);
       rdev->cp.wptr = RREG32(CP_RB_WPTR);
 
-     r600_cp_start(rdev);
+     evergreen_cp_start(rdev);
       rdev->cp.ready = true;
       r = radeon_ring_test(rdev);
       if (r) {
drivers/gpu/drm/radeon/r600.c +1 -4
index 73e6a1a..f280717 100644
@@ -2120,10 +2120,7 @@ int r600_cp_start(struct radeon_device *rdev)
       }
       radeon_ring_write(rdev, PACKET3(PACKET3_ME_INITIALIZE, 5));
       radeon_ring_write(rdev, 0x1);
-     if (rdev->family >= CHIP_CEDAR) {
-             radeon_ring_write(rdev, 0x0);
-             radeon_ring_write(rdev, rdev->config.evergreen.max_hw_contexts - 1);
-     } else if (rdev->family >= CHIP_RV770) {
+     if (rdev->family >= CHIP_RV770) {
               radeon_ring_write(rdev, 0x0);
               radeon_ring_write(rdev, rdev->config.rv770.max_hw_contexts - 1);
       } else {
#3 lore raw
Quoted 11 lines · Alex Deucher
Quoted 7 lines · Dave Airlie
Quoted 3 lines · Alex Deucher

These VGT regs need to be programmed via the ring rather than MMIO as on previous asics (r6xx/r7xx).

So we need someway for userspace to know this has happened, either a version bump or a change to the info ioctl again.

We can do a version bump for 2.6.36. Not sure what the best option is for 2.6.35.

Yeah if you wanted to fix 2.6.35, you'd need to add yet another info ioctl.

Dave.