drm/amdkfd: implement per queue sdma reset for gfx 9.4+

jokim-amd · alexdeucher · commit bac38ca8c475 · 2025-03-05T10:47:26.000-05:00
To reset hung SDMA queues on GFX 9.4+ for the GFX9 family, a soft reset
must be issued through SMU.  Since soft resets will reset an entire SDMA
engine, use a common KGD call to do the reset as the KGD will handle
avoiding a reset of in flight GFX and paging queues on that engine.

In addition, create a common call for all reset types to simplify
the handling of module parameter settings that block gpu resets.

Signed-off-by: Jonathan Kim &lt;jonathan.kim@amd.com&gt;
Reviewed-by: Harish Kasiviswanathan &lt;harish.kasiviswanathan@amd.com&gt;
Signed-off-by: Alex Deucher &lt;alexander.deucher@amd.com&gt;
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_aldebaran.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_aldebaran.c
@@ -193,4 +193,5 @@ const struct kfd2kgd_calls aldebaran_kfd2kgd = {
 	.program_trap_handler_settings = kgd_gfx_v9_program_trap_handler_settings,
 	.hqd_get_pq_addr = kgd_gfx_v9_hqd_get_pq_addr,
 	.hqd_reset = kgd_gfx_v9_hqd_reset,
+	.hqd_sdma_get_doorbell = kgd_gfx_v9_hqd_sdma_get_doorbell
 };
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_arcturus.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_arcturus.c
@@ -419,5 +419,6 @@ const struct kfd2kgd_calls arcturus_kfd2kgd = {
 	.get_cu_occupancy = kgd_gfx_v9_get_cu_occupancy,
 	.program_trap_handler_settings = kgd_gfx_v9_program_trap_handler_settings,
 	.hqd_get_pq_addr = kgd_gfx_v9_hqd_get_pq_addr,
-	.hqd_reset = kgd_gfx_v9_hqd_reset
+	.hqd_reset = kgd_gfx_v9_hqd_reset,
+	.hqd_sdma_get_doorbell = kgd_gfx_v9_hqd_sdma_get_doorbell
 };
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gc_9_4_3.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gc_9_4_3.c
@@ -509,6 +509,17 @@ static uint32_t kgd_gfx_v9_4_3_clear_address_watch(struct amdgpu_device *adev,
 	return 0;
 }
 
+static uint32_t kgd_gfx_v9_4_3_hqd_sdma_get_doorbell(struct amdgpu_device *adev,
+						     int engine, int queue)
+{
+	uint32_t reg_offset = get_sdma_rlc_reg_offset(adev, engine, queue);
+	uint32_t status = RREG32(regSDMA_RLC0_CONTEXT_STATUS + reg_offset);
+	uint32_t doorbell_off = RREG32(regSDMA_RLC0_DOORBELL_OFFSET + reg_offset);
+	bool is_active = !!REG_GET_FIELD(status, SDMA_RLC0_CONTEXT_STATUS, SELECTED);
+
+	return is_active ? doorbell_off >> 2 : 0;
+}
+
 const struct kfd2kgd_calls gc_9_4_3_kfd2kgd = {
 	.program_sh_mem_settings = kgd_gfx_v9_program_sh_mem_settings,
 	.set_pasid_vmid_mapping = kgd_gfx_v9_4_3_set_pasid_vmid_mapping,
@@ -543,5 +554,6 @@ const struct kfd2kgd_calls gc_9_4_3_kfd2kgd = {
 	.set_address_watch = kgd_gfx_v9_4_3_set_address_watch,
 	.clear_address_watch = kgd_gfx_v9_4_3_clear_address_watch,
 	.hqd_get_pq_addr = kgd_gfx_v9_hqd_get_pq_addr,
-	.hqd_reset = kgd_gfx_v9_hqd_reset
+	.hqd_reset = kgd_gfx_v9_hqd_reset,
+	.hqd_sdma_get_doorbell = kgd_gfx_v9_4_3_hqd_sdma_get_doorbell
 };
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c
@@ -1084,6 +1084,12 @@ uint64_t kgd_gfx_v10_hqd_reset(struct amdgpu_device *adev,
 	return 0;
 }
 
+uint32_t kgd_gfx_v10_hqd_sdma_get_doorbell(struct amdgpu_device *adev,
+					   int engine, int queue)
+{
+	return 0;
+}
+
 const struct kfd2kgd_calls gfx_v10_kfd2kgd = {
 	.program_sh_mem_settings = kgd_program_sh_mem_settings,
 	.set_pasid_vmid_mapping = kgd_set_pasid_vmid_mapping,
@@ -1112,5 +1118,6 @@ const struct kfd2kgd_calls gfx_v10_kfd2kgd = {
 	.build_grace_period_packet_info = kgd_gfx_v10_build_grace_period_packet_info,
 	.program_trap_handler_settings = program_trap_handler_settings,
 	.hqd_get_pq_addr = kgd_gfx_v10_hqd_get_pq_addr,
-	.hqd_reset = kgd_gfx_v10_hqd_reset
+	.hqd_reset = kgd_gfx_v10_hqd_reset,
+	.hqd_sdma_get_doorbell = kgd_gfx_v10_hqd_sdma_get_doorbell
 };
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.h
@@ -65,3 +65,5 @@ uint64_t kgd_gfx_v10_hqd_reset(struct amdgpu_device *adev,
 			      uint32_t queue_id,
 			      uint32_t inst,
 			      unsigned int utimeout);
+uint32_t kgd_gfx_v10_hqd_sdma_get_doorbell(struct amdgpu_device *adev,
+					   int engine, int queue);
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c
@@ -682,5 +682,6 @@ const struct kfd2kgd_calls gfx_v10_3_kfd2kgd = {
 	.set_address_watch = kgd_gfx_v10_set_address_watch,
 	.clear_address_watch = kgd_gfx_v10_clear_address_watch,
 	.hqd_get_pq_addr = kgd_gfx_v10_hqd_get_pq_addr,
-	.hqd_reset = kgd_gfx_v10_hqd_reset
+	.hqd_reset = kgd_gfx_v10_hqd_reset,
+	.hqd_sdma_get_doorbell = kgd_gfx_v10_hqd_sdma_get_doorbell
 };
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v11.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v11.c
@@ -800,6 +800,12 @@ static uint64_t kgd_gfx_v11_hqd_reset(struct amdgpu_device *adev,
 	return 0;
 }
 
+static uint32_t kgd_gfx_v11_hqd_sdma_get_doorbell(struct amdgpu_device *adev,
+						  int engine, int queue)
+{
+	return 0;
+}
+
 const struct kfd2kgd_calls gfx_v11_kfd2kgd = {
 	.program_sh_mem_settings = program_sh_mem_settings_v11,
 	.set_pasid_vmid_mapping = set_pasid_vmid_mapping_v11,
@@ -824,5 +830,6 @@ const struct kfd2kgd_calls gfx_v11_kfd2kgd = {
 	.set_address_watch = kgd_gfx_v11_set_address_watch,
 	.clear_address_watch = kgd_gfx_v11_clear_address_watch,
 	.hqd_get_pq_addr = kgd_gfx_v11_hqd_get_pq_addr,
-	.hqd_reset = kgd_gfx_v11_hqd_reset
+	.hqd_reset = kgd_gfx_v11_hqd_reset,
+	.hqd_sdma_get_doorbell = kgd_gfx_v11_hqd_sdma_get_doorbell
 };
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v12.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v12.c
@@ -361,6 +361,12 @@ static uint32_t kgd_gfx_v12_clear_address_watch(struct amdgpu_device *adev,
 	return 0;
 }
 
+static uint32_t kgd_gfx_v12_hqd_sdma_get_doorbell(struct amdgpu_device *adev,
+						 int engine, int queue)
+{
+	return 0;
+}
+
 const struct kfd2kgd_calls gfx_v12_kfd2kgd = {
 	.init_interrupts = init_interrupts_v12,
 	.hqd_dump = hqd_dump_v12,
@@ -374,4 +380,5 @@ const struct kfd2kgd_calls gfx_v12_kfd2kgd = {
 	.set_wave_launch_mode = kgd_gfx_v12_set_wave_launch_mode,
 	.set_address_watch = kgd_gfx_v12_set_address_watch,
 	.clear_address_watch = kgd_gfx_v12_clear_address_watch,
+	.hqd_sdma_get_doorbell = kgd_gfx_v12_hqd_sdma_get_doorbell
 };
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c
@@ -1131,9 +1131,6 @@ uint64_t kgd_gfx_v9_hqd_get_pq_addr(struct amdgpu_device *adev,
 	uint32_t low, high;
 	uint64_t queue_addr = 0;
 
-	if (!amdgpu_gpu_recovery)
-		return 0;
-
 	kgd_gfx_v9_acquire_queue(adev, pipe_id, queue_id, inst);
 	amdgpu_gfx_rlc_enter_safe_mode(adev, inst);
 
@@ -1182,9 +1179,6 @@ uint64_t kgd_gfx_v9_hqd_reset(struct amdgpu_device *adev,
 	uint32_t low, high, pipe_reset_data = 0;
 	uint64_t queue_addr = 0;
 
-	if (!amdgpu_gpu_recovery)
-		return 0;
-
 	kgd_gfx_v9_acquire_queue(adev, pipe_id, queue_id, inst);
 	amdgpu_gfx_rlc_enter_safe_mode(adev, inst);
 
@@ -1229,6 +1223,13 @@ uint64_t kgd_gfx_v9_hqd_reset(struct amdgpu_device *adev,
 	return queue_addr;
 }
 
+uint32_t kgd_gfx_v9_hqd_sdma_get_doorbell(struct amdgpu_device *adev,
+					  int engine, int queue)
+
+{
+	return 0;
+}
+
 const struct kfd2kgd_calls gfx_v9_kfd2kgd = {
 	.program_sh_mem_settings = kgd_gfx_v9_program_sh_mem_settings,
 	.set_pasid_vmid_mapping = kgd_gfx_v9_set_pasid_vmid_mapping,
@@ -1258,5 +1259,6 @@ const struct kfd2kgd_calls gfx_v9_kfd2kgd = {
 	.get_cu_occupancy = kgd_gfx_v9_get_cu_occupancy,
 	.program_trap_handler_settings = kgd_gfx_v9_program_trap_handler_settings,
 	.hqd_get_pq_addr = kgd_gfx_v9_hqd_get_pq_addr,
-	.hqd_reset = kgd_gfx_v9_hqd_reset
+	.hqd_reset = kgd_gfx_v9_hqd_reset,
+	.hqd_sdma_get_doorbell = kgd_gfx_v9_hqd_sdma_get_doorbell
 };
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h
@@ -111,3 +111,5 @@ uint64_t kgd_gfx_v9_hqd_reset(struct amdgpu_device *adev,
 			      uint32_t queue_id,
 			      uint32_t inst,
 			      unsigned int utimeout);
+uint32_t kgd_gfx_v9_hqd_sdma_get_doorbell(struct amdgpu_device *adev,
+					  int engine, int queue);
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
@@ -36,6 +36,7 @@
 #include "kfd_kernel_queue.h"
 #include "amdgpu_amdkfd.h"
 #include "amdgpu_reset.h"
+#include "amdgpu_sdma.h"
 #include "mes_v11_api_def.h"
 #include "kfd_debug.h"
 
@@ -67,6 +68,8 @@ static int allocate_hqd(struct device_queue_manager *dqm, struct queue *q);
 static int allocate_sdma_queue(struct device_queue_manager *dqm,
 				struct queue *q, const uint32_t *restore_sdma_id);
 
+static int reset_queues_on_hws_hang(struct device_queue_manager *dqm, bool is_sdma);
+
 static inline
 enum KFD_MQD_TYPE get_mqd_type_from_queue_type(enum kfd_queue_type type)
 {
@@ -2205,8 +2208,7 @@ static struct queue *find_queue_by_address(struct device_queue_manager *dqm, uin
 	return NULL;
 }
 
-/* only for compute queue */
-static int reset_queues_on_hws_hang(struct device_queue_manager *dqm)
+static int reset_hung_queues(struct device_queue_manager *dqm)
 {
 	int r = 0, reset_count = 0, i;
 
@@ -2259,6 +2261,104 @@ static int reset_queues_on_hws_hang(struct device_queue_manager *dqm)
 	return r;
 }
 
+static bool sdma_has_hang(struct device_queue_manager *dqm)
+{
+	int engine_start = dqm->dev->node_id * get_num_all_sdma_engines(dqm);
+	int engine_end = engine_start + get_num_all_sdma_engines(dqm);
+	int num_queues_per_eng =  dqm->dev->kfd->device_info.num_sdma_queues_per_engine;
+	int i, j;
+
+	for (i = engine_start; i < engine_end; i++) {
+		for (j = 0; j < num_queues_per_eng; j++) {
+			if (!dqm->dev->kfd2kgd->hqd_sdma_get_doorbell(dqm->dev->adev, i, j))
+				continue;
+
+			return true;
+		}
+	}
+
+	return false;
+}
+
+static bool set_sdma_queue_as_reset(struct device_queue_manager *dqm,
+				    uint32_t doorbell_off)
+{
+	struct device_process_node *cur;
+	struct qcm_process_device *qpd;
+	struct queue *q;
+
+	list_for_each_entry(cur, &dqm->queues, list) {
+		qpd = cur->qpd;
+		list_for_each_entry(q, &qpd->queues_list, list) {
+			if ((q->properties.type == KFD_QUEUE_TYPE_SDMA ||
+			     q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) &&
+			     q->properties.doorbell_off == doorbell_off) {
+				set_queue_as_reset(dqm, q, qpd);
+				return true;
+			}
+		}
+	}
+
+	return false;
+}
+
+static int reset_hung_queues_sdma(struct device_queue_manager *dqm)
+{
+	int engine_start = dqm->dev->node_id * get_num_all_sdma_engines(dqm);
+	int engine_end = engine_start + get_num_all_sdma_engines(dqm);
+	int num_queues_per_eng =  dqm->dev->kfd->device_info.num_sdma_queues_per_engine;
+	int r = 0, i, j;
+
+	if (dqm->is_hws_hang)
+		return -EIO;
+
+	/* Scan for hung HW queues and reset engine. */
+	dqm->detect_hang_count = 0;
+	for (i = engine_start; i < engine_end; i++) {
+		for (j = 0; j < num_queues_per_eng; j++) {
+			uint32_t doorbell_off =
+				dqm->dev->kfd2kgd->hqd_sdma_get_doorbell(dqm->dev->adev, i, j);
+
+			if (!doorbell_off)
+				continue;
+
+			/* Reset engine and check. */
+			if (amdgpu_sdma_reset_engine(dqm->dev->adev, i, false) ||
+			    dqm->dev->kfd2kgd->hqd_sdma_get_doorbell(dqm->dev->adev, i, j) ||
+			    !set_sdma_queue_as_reset(dqm, doorbell_off)) {
+				r = -ENOTRECOVERABLE;
+				goto reset_fail;
+			}
+
+			/* Should only expect one queue active per engine */
+			dqm->detect_hang_count++;
+			break;
+		}
+	}
+
+	/* Signal process reset */
+	if (dqm->detect_hang_count)
+		kfd_signal_reset_event(dqm->dev);
+	else
+		r = -ENOTRECOVERABLE;
+
+reset_fail:
+	dqm->detect_hang_count = 0;
+
+	return r;
+}
+
+static int reset_queues_on_hws_hang(struct device_queue_manager *dqm, bool is_sdma)
+{
+	while (halt_if_hws_hang)
+		schedule();
+
+	if (!amdgpu_gpu_recovery)
+		return -ENOTRECOVERABLE;
+
+	return is_sdma ? reset_hung_queues_sdma(dqm) : reset_hung_queues(dqm);
+}
+
 /* dqm->lock mutex has to be locked before calling this function */
 static int unmap_queues_cpsch(struct device_queue_manager *dqm,
 				enum kfd_unmap_queues_filter filter,
@@ -2309,16 +2409,13 @@ static int unmap_queues_cpsch(struct device_queue_manager *dqm,
 	 * check those fields
 	 */
 	mqd_mgr = dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ];
-	if (mqd_mgr->check_preemption_failed(mqd_mgr, dqm->packet_mgr.priv_queue->queue->mqd)) {
-		while (halt_if_hws_hang)
-			schedule();
-		if (reset_queues_on_hws_hang(dqm)) {
-			dqm->is_hws_hang = true;
-			kfd_hws_hang(dqm);
-			retval = -ETIME;
-			goto out;
-		}
-	}
+	if (mqd_mgr->check_preemption_failed(mqd_mgr, dqm->packet_mgr.priv_queue->queue->mqd) &&
+	    reset_queues_on_hws_hang(dqm, false))
+		goto reset_fail;
+
+	/* Check for SDMA hang and attempt SDMA reset */
+	if (sdma_has_hang(dqm) && reset_queues_on_hws_hang(dqm, true))
+		goto reset_fail;
 
 	/* We need to reset the grace period value for this device */
 	if (grace_period != USE_DEFAULT_GRACE_PERIOD) {
@@ -2329,10 +2426,15 @@ static int unmap_queues_cpsch(struct device_queue_manager *dqm,
 
 	pm_release_ib(&dqm->packet_mgr);
 	dqm->active_runlist = false;
-
 out:
 	up_read(&dqm->dev->adev->reset_domain->sem);
 	return retval;
+
+reset_fail:
+	dqm->is_hws_hang = true;
+	kfd_hws_hang(dqm);
+	up_read(&dqm->dev->adev->reset_domain->sem);
+	return -ETIME;
 }
 
 /* only for compute queue */
diff --git a/drivers/gpu/drm/amd/include/kgd_kfd_interface.h b/drivers/gpu/drm/amd/include/kgd_kfd_interface.h
@@ -330,6 +330,8 @@ struct kfd2kgd_calls {
 	uint64_t (*hqd_reset)(struct amdgpu_device *adev,
 			      uint32_t pipe_id, uint32_t queue_id,
 			      uint32_t inst, unsigned int utimeout);
+	uint32_t (*hqd_sdma_get_doorbell)(struct amdgpu_device *adev,
+					  int engine, int queue);
 };
 
 #endif	/* KGD_KFD_INTERFACE_H_INCLUDED */
