drm/amdkfd: Update logic for CU occupancy calculations

Currently, the code uses the IH_VMID_X_LUT register to map
a queue's vmid to the corresponding PASID. This logic is racy
since CP can update the VMID-PASID mapping anytime especially
when there are more processes than number of vmids. Update the
logic to calculate CU occupancy by matching doorbell offset of
the queue with valid wave counts against the process's queues.

Signed-off-by: Mukul Joshi <mukul.joshi@amd.com>
Reviewed-by: Harish Kasiviswanathan <Harish.Kasiviswanathan@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

3 files changed, 36 insertions, 1 deletions

diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
index 71b465f8d83e..29578550b478 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
@@ -3540,6 +3540,26 @@ int debug_refresh_runlist(struct device_queue_manager *dqm)
 	return debug_map_and_unlock(dqm);
 }
 
+bool kfd_dqm_is_queue_in_process(struct device_queue_manager *dqm,
+				 struct qcm_process_device *qpd,
+				 int doorbell_off)
+{
+	struct queue *q;
+	bool r = false;
+
+	dqm_lock(dqm);
+
+	list_for_each_entry(q, &qpd->queues_list, list) {
+		if (q->properties.doorbell_off == doorbell_off) {
+			r = true;
+			goto out;
+		}
+	}
+
+out:
+	dqm_unlock(dqm);
+	return r;
+}
 #if defined(CONFIG_DEBUG_FS)
 
 static void seq_reg_dump(struct seq_file *m,
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h
index 08b40826ad1e..80be2036abea 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h
@@ -324,6 +324,9 @@ void set_queue_snapshot_entry(struct queue *q,
 int debug_lock_and_unmap(struct device_queue_manager *dqm);
 int debug_map_and_unlock(struct device_queue_manager *dqm);
 int debug_refresh_runlist(struct device_queue_manager *dqm);
+bool kfd_dqm_is_queue_in_process(struct device_queue_manager *dqm,
+				 struct qcm_process_device *qpd,
+				 int doorbell_off);
 
 static inline unsigned int get_sh_mem_bases_32(struct kfd_process_device *pdd)
 {
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
index a902950cc060..d73841268c9b 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
@@ -270,6 +270,10 @@ static int kfd_get_cu_occupancy(struct attribute *attr, char *buffer)
 	struct kfd_node *dev = NULL;
 	struct kfd_process *proc = NULL;
 	struct kfd_process_device *pdd = NULL;
+	int i;
+	struct kfd_cu_occupancy cu_occupancy[AMDGPU_MAX_QUEUES];
+
+	memset(cu_occupancy, 0x0, sizeof(cu_occupancy));
 
 	pdd = container_of(attr, struct kfd_process_device, attr_cu_occupancy);
 	dev = pdd->dev;
@@ -287,9 +291,17 @@ static int kfd_get_cu_occupancy(struct attribute *attr, char *buffer)
 	/* Collect wave count from device if it supports */
 	wave_cnt = 0;
 	max_waves_per_cu = 0;
-	dev->kfd2kgd->get_cu_occupancy(dev->adev, proc->pasid, &wave_cnt,
+
+	dev->kfd2kgd->get_cu_occupancy(dev->adev, cu_occupancy,
 			&max_waves_per_cu, 0);
 
+	for (i = 0; i < AMDGPU_MAX_QUEUES; i++) {
+		if (cu_occupancy[i].wave_cnt != 0 &&
+		    kfd_dqm_is_queue_in_process(dev->dqm, &pdd->qpd,
+						cu_occupancy[i].doorbell_off))
+			wave_cnt += cu_occupancy[i].wave_cnt;
+	}
+
 	/* Translate wave count to number of compute units */
 	cu_cnt = (wave_cnt + (max_waves_per_cu - 1)) / max_waves_per_cu;
 	return snprintf(buffer, PAGE_SIZE, "%d\n", cu_cnt);