[PATCH v3 3/6] drm/amdgpu: save ring content before resetting the device

[Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>]

Otherwise the content might not be relevant.

When a coredump is generated the rings with outstanding fences
are saved and then printed to the final devcoredump from the
worker thread.
Since this requires memory allocation, the ring capture might
be missing from the generated devcoredump.

Signed-off-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
---
 .../gpu/drm/amd/amdgpu/amdgpu_dev_coredump.c  | 82 +++++++++++++++----
 .../gpu/drm/amd/amdgpu/amdgpu_dev_coredump.h  | 12 +++
 2 files changed, 78 insertions(+), 16 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_dev_coredump.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_dev_coredump.c
index 0c7fc3800f17..58b2e764dd7c 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_dev_coredump.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_dev_coredump.c
@@ -207,7 +207,9 @@ amdgpu_devcoredump_format(char *buffer, size_t count, struct amdgpu_coredump_inf
 	struct drm_print_iterator iter;
 	struct amdgpu_vm_fault_info *fault_info;
 	struct amdgpu_ip_block *ip_block;
-	int ver;
+	struct amdgpu_ring *ring;
+	int ver, i, j;
+	u32 ring_idx, off;
 
 	iter.data = buffer;
 	iter.offset = 0;
@@ -297,23 +299,25 @@ amdgpu_devcoredump_format(char *buffer, size_t count, struct amdgpu_coredump_inf
 
 	/* Add ring buffer information */
 	drm_printf(&p, "Ring buffer information\n");
-	for (int i = 0; i < coredump->adev->num_rings; i++) {
-		int j = 0;
-		struct amdgpu_ring *ring = coredump->adev->rings[i];
+	if (coredump->num_rings) {
+		for (i = 0; i < coredump->num_rings; i++) {
+			ring_idx = coredump->rings[i].ring_index;
+			ring = coredump->adev->rings[ring_idx];
+			off = coredump->rings[i].offset;
 
-		drm_printf(&p, "ring name: %s\n", ring->name);
-		drm_printf(&p, "Rptr: 0x%llx Wptr: 0x%llx RB mask: %x\n",
-			   amdgpu_ring_get_rptr(ring),
-			   amdgpu_ring_get_wptr(ring),
-			   ring->buf_mask);
-		drm_printf(&p, "Ring size in dwords: %d\n",
-			   ring->ring_size / 4);
-		drm_printf(&p, "Ring contents\n");
-		drm_printf(&p, "Offset \t Value\n");
+			drm_printf(&p, "ring name: %s\n", ring->name);
+			drm_printf(&p, "Rptr: 0x%llx Wptr: 0x%llx RB mask: %x\n",
+				   coredump->rings[i].rptr,
+				   coredump->rings[i].wptr,
+				   ring->buf_mask);
+			drm_printf(&p, "Ring size in dwords: %d\n",
+				ring->ring_size / 4);
+			drm_printf(&p, "Ring contents\n");
+			drm_printf(&p, "Offset \t Value\n");
 
-		while (j < ring->ring_size) {
-			drm_printf(&p, "0x%x \t 0x%x\n", j, ring->ring[j / 4]);
-			j += 4;
+			for (j = 0; j < ring->ring_size; j += 4)
+				drm_printf(&p, "0x%x \t 0x%x\n", j,
+					   coredump->rings_dw[off + j / 4]);
 		}
 	}
 
@@ -353,6 +357,8 @@ static void amdgpu_devcoredump_free(void *data)
 	struct amdgpu_coredump_info *coredump = data;
 
 	kvfree(coredump->formatted);
+	kvfree(coredump->rings);
+	kvfree(coredump->rings_dw);
 	kvfree(data);
 }
 
@@ -390,6 +396,9 @@ void amdgpu_coredump(struct amdgpu_device *adev, bool skip_vram_check,
 	struct drm_device *dev = adev_to_drm(adev);
 	struct amdgpu_coredump_info *coredump;
 	struct drm_sched_job *s_job;
+	u64 total_ring_size, ring_count;
+	struct amdgpu_ring *ring;
+	int i, off, idx;
 
 	/* No need to generate a new coredump if there's one in progress already. */
 	if (work_pending(&adev->coredump_work))
@@ -417,6 +426,47 @@ void amdgpu_coredump(struct amdgpu_device *adev, bool skip_vram_check,
 		coredump->ring = to_amdgpu_ring(s_job->sched);
 	}
 
+	/* Dump ring content if memory allocation succeeds. */
+	ring_count = 0;
+	total_ring_size = 0;
+	for (i = 0; i < adev->num_rings; i++) {
+		ring = adev->rings[i];
+
+		/* Only dump rings with unsignalled fences. */
+		if (atomic_read(&ring->fence_drv.last_seq) == ring->fence_drv.sync_seq &&
+		    coredump->ring != ring)
+			continue;
+
+		total_ring_size += ring->ring_size;
+		ring_count++;
+	}
+	coredump->rings_dw = kzalloc(total_ring_size, GFP_NOWAIT);
+	coredump->rings = kcalloc(ring_count, sizeof(struct amdgpu_coredump_ring), GFP_NOWAIT);
+	if (coredump->rings && coredump->rings_dw) {
+		for (i = 0, off = 0, idx = 0; i < adev->num_rings; i++) {
+			ring = adev->rings[i];
+
+			if (atomic_read(&ring->fence_drv.last_seq) == ring->fence_drv.sync_seq &&
+			    coredump->ring != ring)
+				continue;
+
+			coredump->rings[idx].ring_index = ring->idx;
+			coredump->rings[idx].rptr = amdgpu_ring_get_rptr(ring);
+			coredump->rings[idx].wptr = amdgpu_ring_get_wptr(ring);
+			coredump->rings[idx].offset = off;
+
+			memcpy(&coredump->rings_dw[off], ring->ring, ring->ring_size);
+			off += ring->ring_size;
+			idx++;
+		}
+		coredump->num_rings = idx;
+	} else {
+		kvfree(coredump->rings_dw);
+		kvfree(coredump->rings);
+		coredump->rings_dw = NULL;
+		coredump->rings = NULL;
+	}
+
 	coredump->adev = adev;
 
 	ktime_get_ts64(&coredump->reset_time);
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_dev_coredump.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_dev_coredump.h
index b3582d0b4ca4..5d6c58abf589 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_dev_coredump.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_dev_coredump.h
@@ -31,6 +31,13 @@
 
 #define AMDGPU_COREDUMP_VERSION "1"
 
+struct amdgpu_coredump_ring {
+	u64				rptr;
+	u64				wptr;
+	u32				ring_index;
+	u32				offset;
+};
+
 struct amdgpu_coredump_info {
 	struct amdgpu_device            *adev;
 	struct amdgpu_task_info         reset_task_info;
@@ -39,6 +46,11 @@ struct amdgpu_coredump_info {
 	bool                            skip_vram_check;
 	bool                            reset_vram_lost;
 	struct amdgpu_ring              *ring;
+
+	struct amdgpu_coredump_ring	*rings;
+	u32				*rings_dw;
+	u32				num_rings;
+
 	/* Readable form of coredevdump, generate once to speed up
 	 * reading it (see drm_coredump_printer's documentation).
 	 */
-- 
2.43.0