Skip to contentSkip to conversation

drm/amdgpu/mes11: set connected queue index in the reset packet

PATCH New Jesse Zhang 1 participant 3 messages Jul 17

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

PATCH 1/3 drm/amdgpu/mes11: set connected queue index in the reset packet

Sync MESAPI__RESET with the firmware definition (the use_connected_queue_index[_p1] flags and the trailing active_vmids / connected_queue_index[_p1] fields) and program the connected queue index from the detect input, so the MES can locate a CP-frozen gfx user queue via its CP_CNTX_STAT fallback.

An older MES firmware ignores the flags (previously reserved bits) and falls back to query_status, so no firmware version gate is required.

---
 drivers/gpu/drm/amd/amdgpu/mes_v11_0.c        | 11 +++++++++++
 drivers/gpu/drm/amd/include/mes_v11_api_def.h | 11 ++++++++++-
 2 files changed, 21 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/mes_v11_0.c +11 -0
index b72da75dd851..3e6d3b5e9af9 100644
@@ -1115,6 +1115,17 @@ static int mes_v11_0_detect_and_reset_hung_queues(struct amdgpu_mes *mes,
 	else
 		mes_reset_queue_pkt.hang_detect_then_reset = 1;
 
+	if (input->use_connected_queue_index) {
+		mes_reset_queue_pkt.use_connected_queue_index = 1;
+		mes_reset_queue_pkt.connected_queue_index =
+			input->connected_queue_index;
+	}
+	if (input->use_connected_queue_index_p1) {
+		mes_reset_queue_pkt.use_connected_queue_index_p1 = 1;
+		mes_reset_queue_pkt.connected_queue_index_p1 =
+			input->connected_queue_index_p1;
+	}
+
 	return mes_v11_0_submit_pkt_and_poll_completion(mes,
 			&mes_reset_queue_pkt, sizeof(mes_reset_queue_pkt),
 			offsetof(union MESAPI__RESET, api_status));
drivers/gpu/drm/amd/include/mes_v11_api_def.h +10 -1
index a8e6ccc7817c..2762b55f3474 100644
@@ -468,7 +468,11 @@ union MESAPI__RESET {
 			uint32_t                hang_detect_only : 1;
 			/* Rest HP and LP kernel queues not managed by MES */
 			uint32_t                reset_legacy_gfx : 1;
-			uint32_t                reserved : 28;
+			/* Fallback to connected queue index when CP_CNTX_STAT fails (gfx pipe 0) */
+			uint32_t                use_connected_queue_index : 1;
+			/* For gfx pipe 1 */
+			uint32_t                use_connected_queue_index_p1 : 1;
+			uint32_t                reserved : 26;
 		};
 
 		uint64_t			gang_context_addr;
@@ -496,6 +500,11 @@ union MESAPI__RESET {
 		uint64_t			wptr_addr_hp;
 
 		struct MES_API_STATUS		api_status;
+		uint32_t			active_vmids;
+		uint64_t			timestamp;
+		uint32_t			gang_context_array_index;
+		uint32_t			connected_queue_index;
+		uint32_t			connected_queue_index_p1;
 	};
 
 	uint32_t	max_dwords_in_api[API_FRAME_SIZE_IN_DWORDS];
Signature · 1 line
2.49.0
#1 lore raw

PATCH 2/3 drm/amdgpu/mes12: set connected queue index in the reset packet

Program the connected queue index into the MES reset packet on gfx12 as well, so the CP_CNTX_STAT fallback can find and reset a CP-frozen gfx user queue without query_status. mes_v12_api_def.h already carries the fields.

---
 drivers/gpu/drm/amd/amdgpu/mes_v12_0.c | 11 +++++++++++
 1 file changed, 11 insertions(+)
drivers/gpu/drm/amd/amdgpu/mes_v12_0.c +11 -0
index 04465804c254..2a38d0c560b1 100644
@@ -1151,6 +1151,17 @@ static int mes_v12_0_detect_and_reset_hung_queues(struct amdgpu_mes *mes,
 	else
 		mes_reset_queue_pkt.hang_detect_then_reset = 1;
 
+	if (input->use_connected_queue_index) {
+		mes_reset_queue_pkt.use_connected_queue_index = 1;
+		mes_reset_queue_pkt.connected_queue_index =
+			input->connected_queue_index;
+	}
+	if (input->use_connected_queue_index_p1) {
+		mes_reset_queue_pkt.use_connected_queue_index_p1 = 1;
+		mes_reset_queue_pkt.connected_queue_index_p1 =
+			input->connected_queue_index_p1;
+	}
+
 	return mes_v12_0_submit_pkt_and_poll_completion(mes, AMDGPU_MES_SCHED_PIPE,
 			&mes_reset_queue_pkt, sizeof(mes_reset_queue_pkt),
 			offsetof(union MESAPI__RESET, api_status));
Signature · 1 line
2.49.0
#2 lore raw

PATCH 3/3 drm/amdgpu/mes: pass connected queue index to hung queue detection

In reply to Jesse Zhang

The MES query_status hang detection needs a live CP and so cannot find a CP-frozen gfx user queue. The MES firmware has a CP_CNTX_STAT fallback (find_hung_queue_by_cp_ctx_stat) that locates a hung gfx queue from its pipe/hqd slot without query_status, but the driver never fed it the slot.

Extend mes_detect_and_reset_queue_input and amdgpu_mes_detect_and_reset_hung_queues() with a connected pipe/queue hint (pipe 0 -> connected_queue_index, pipe 1 -> _p1); callers that do not know the slot pass -1 to keep the existing behaviour.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c |  3 ++-
 drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c | 19 +++++++++++++++++--
 drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h |  9 ++++++++-
 3 files changed, 27 insertions(+), 4 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +2 -1
index 9d3b40c385c9..f5ae16f7b4ed 100644
@@ -2299,7 +2299,8 @@ int amdgpu_gfx_reset_mes_compute(struct amdgpu_device *adev,
 		 */
 		goto fence_reset;
 	r = amdgpu_mes_detect_and_reset_hung_queues(adev, AMDGPU_RING_TYPE_COMPUTE,
-						    true, &num_hung, adev->gfx.mec.mes_hung_db_array, 0);
+						    true, &num_hung, adev->gfx.mec.mes_hung_db_array, 0,
+						    -1, 0);
 	if (r)
 		goto out;
 	if (hung_queue_count)
drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c +17 -2
index 1e3a23c81552..40e966a0d676 100644
@@ -513,9 +513,11 @@ int amdgpu_mes_detect_and_reset_hung_queues(struct amdgpu_device *adev,
 					    bool detect_only,
 					    unsigned int *hung_db_num,
 					    u32 *hung_db_array,
-					    uint32_t xcc_id)
+					    uint32_t xcc_id,
+					    int connected_pipe,
+					    int connected_queue_index)
 {
-	struct mes_detect_and_reset_queue_input input;
+	struct mes_detect_and_reset_queue_input input = {0};
 	u32 *db_array = adev->mes.hung_queue_db_array_cpu_addr[xcc_id];
 	int hqd_info_offset = adev->mes.hung_queue_hqd_info_offset, r, i;
 
@@ -534,6 +536,19 @@ int amdgpu_mes_detect_and_reset_hung_queues(struct amdgpu_device *adev,
 	input.detect_only = detect_only;
 	input.xcc_id = xcc_id;
 
+	/*
+	 * For a CP-frozen gfx queue the MES query_status hang detection can't
+	 * see it; hand the MES the exact pipe/queue slot so its CP_CNTX_STAT
+	 * fallback can find and reset it.
+	 */
+	if (connected_pipe == 0) {
+		input.use_connected_queue_index = true;
+		input.connected_queue_index = connected_queue_index;
+	} else if (connected_pipe == 1) {
+		input.use_connected_queue_index_p1 = true;
+		input.connected_queue_index_p1 = connected_queue_index;
+	}
+
 	r = adev->mes.funcs->detect_and_reset_hung_queues(&adev->mes,
 							  &input);
 
drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h +8 -1
index c67db2d6e122..a5d0e649d113 100644
@@ -350,6 +350,11 @@ struct mes_detect_and_reset_queue_input {
 	u32	queue_type;
 	bool		detect_only;
 	u32	xcc_id;
+	/* gfx-only: locate a CP-frozen queue the MES query_status path can't see */
+	bool		use_connected_queue_index;
+	bool		use_connected_queue_index_p1;
+	u32	connected_queue_index;
+	u32	connected_queue_index_p1;
 };
 
 struct mes_inv_tlbs_pasid_input {
@@ -544,7 +549,9 @@ int amdgpu_mes_detect_and_reset_hung_queues(struct amdgpu_device *adev,
 					    bool detect_only,
 					    unsigned int *hung_db_num,
 					    u32 *hung_db_array,
-					    uint32_t xcc_id);
+					    uint32_t xcc_id,
+					    int connected_pipe,
+					    int connected_queue_index);
 
 uint32_t amdgpu_mes_rreg(struct amdgpu_device *adev, uint32_t reg,
 			 uint32_t xcc_id);
Signature · 1 line
2.49.0