drm/amdgpu: Convert to common fdinfo format v5

Convert fdinfo format to one documented in drm-usage-stats.rst.

It turned out that the existing implementation was actually completely
nonsense. The calculated percentages indeed represented the usage of the
engine, but with varying time slices.

So 10% usage for application A could mean something completely different
than 10% usage for application B.

Completely nuke that and just use the now standardized nanosecond
interface.

v2: drop the documentation change for now, nuke percentage calculation
v3: only account for each hw_ip, move the time_spend to the ctx mgr.
v4: move general ctx changes into separate patch, rework the fdinfo to
    ctx_mgr interface so that all usages are calculated at once, drop
    some unecessary and dangerous refcount dance.
v5: add one more comment how we calculate the time spend

Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Shashank Sharma <shashank.sharma@amd.com>
Cc: Daniel Vetter <daniel@ffwll.ch>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

1 files changed, 91 insertions, 86 deletions

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
index a61e4c83a545..7dc92ef36b2b 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
@@ -162,17 +162,50 @@ static unsigned int amdgpu_ctx_get_hw_prio(struct amdgpu_ctx *ctx, u32 hw_ip)
 	return hw_prio;
 }
 
+/* Calculate the time spend on the hw */
+static ktime_t amdgpu_ctx_fence_time(struct dma_fence *fence)
+{
+	struct drm_sched_fence *s_fence;
+
+	if (!fence)
+		return ns_to_ktime(0);
+
+	/* When the fence is not even scheduled it can't have spend time */
+	s_fence = to_drm_sched_fence(fence);
+	if (!test_bit(DMA_FENCE_FLAG_TIMESTAMP_BIT, &s_fence->scheduled.flags))
+		return ns_to_ktime(0);
+
+	/* When it is still running account how much already spend */
+	if (!test_bit(DMA_FENCE_FLAG_TIMESTAMP_BIT, &s_fence->finished.flags))
+		return ktime_sub(ktime_get(), s_fence->scheduled.timestamp);
+
+	return ktime_sub(s_fence->finished.timestamp,
+			 s_fence->scheduled.timestamp);
+}
+
+static ktime_t amdgpu_ctx_entity_time(struct amdgpu_ctx *ctx,
+				      struct amdgpu_ctx_entity *centity)
+{
+	ktime_t res = ns_to_ktime(0);
+	uint32_t i;
+
+	spin_lock(&ctx->ring_lock);
+	for (i = 0; i < amdgpu_sched_jobs; i++) {
+		res = ktime_add(res, amdgpu_ctx_fence_time(centity->fences[i]));
+	}
+	spin_unlock(&ctx->ring_lock);
+	return res;
+}
 
 static int amdgpu_ctx_init_entity(struct amdgpu_ctx *ctx, u32 hw_ip,
 				  const u32 ring)
 {
+	struct drm_gpu_scheduler **scheds = NULL, *sched = NULL;
 	struct amdgpu_device *adev = ctx->mgr->adev;
 	struct amdgpu_ctx_entity *entity;
-	struct drm_gpu_scheduler **scheds = NULL, *sched = NULL;
-	unsigned num_scheds = 0;
-	int32_t ctx_prio;
-	unsigned int hw_prio;
 	enum drm_sched_priority drm_prio;
+	unsigned int hw_prio, num_scheds;
+	int32_t ctx_prio;
 	int r;
 
 	entity = kzalloc(struct_size(entity, fences, amdgpu_sched_jobs),
@@ -182,6 +215,7 @@ static int amdgpu_ctx_init_entity(struct amdgpu_ctx *ctx, u32 hw_ip,
 
 	ctx_prio = (ctx->override_priority == AMDGPU_CTX_PRIORITY_UNSET) ?
 			ctx->init_priority : ctx->override_priority;
+	entity->hw_ip = hw_ip;
 	entity->sequence = 1;
 	hw_prio = amdgpu_ctx_get_hw_prio(ctx, hw_ip);
 	drm_prio = amdgpu_ctx_to_drm_sched_prio(ctx_prio);
@@ -220,6 +254,23 @@ error_free_entity:
 	return r;
 }
 
+static ktime_t amdgpu_ctx_fini_entity(struct amdgpu_ctx_entity *entity)
+{
+	ktime_t res = ns_to_ktime(0);
+	int i;
+
+	if (!entity)
+		return res;
+
+	for (i = 0; i < amdgpu_sched_jobs; ++i) {
+		res = ktime_add(res, amdgpu_ctx_fence_time(entity->fences[i]));
+		dma_fence_put(entity->fences[i]);
+	}
+
+	kfree(entity);
+	return res;
+}
+
 static int amdgpu_ctx_init(struct amdgpu_ctx_mgr *mgr, int32_t priority,
 			   struct drm_file *filp, struct amdgpu_ctx *ctx)
 {
@@ -246,20 +297,6 @@ static int amdgpu_ctx_init(struct amdgpu_ctx_mgr *mgr, int32_t priority,
 	return 0;
 }
 
-static void amdgpu_ctx_fini_entity(struct amdgpu_ctx_entity *entity)
-{
-
-	int i;
-
-	if (!entity)
-		return;
-
-	for (i = 0; i < amdgpu_sched_jobs; ++i)
-		dma_fence_put(entity->fences[i]);
-
-	kfree(entity);
-}
-
 static int amdgpu_ctx_get_stable_pstate(struct amdgpu_ctx *ctx,
 					u32 *stable_pstate)
 {
@@ -351,8 +388,10 @@ static void amdgpu_ctx_fini(struct kref *ref)
 
 	for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) {
 		for (j = 0; j < AMDGPU_MAX_ENTITY_NUM; ++j) {
-			amdgpu_ctx_fini_entity(ctx->entities[i][j]);
-			ctx->entities[i][j] = NULL;
+			ktime_t spend;
+
+			spend = amdgpu_ctx_fini_entity(ctx->entities[i][j]);
+			atomic64_add(ktime_to_ns(spend), &mgr->time_spend[i]);
 		}
 	}
 
@@ -689,6 +728,9 @@ uint64_t amdgpu_ctx_add_fence(struct amdgpu_ctx *ctx,
 	centity->sequence++;
 	spin_unlock(&ctx->ring_lock);
 
+	atomic64_add(ktime_to_ns(amdgpu_ctx_fence_time(other)),
+		     &ctx->mgr->time_spend[centity->hw_ip]);
+
 	dma_fence_put(other);
 	return seq;
 }
@@ -795,9 +837,14 @@ int amdgpu_ctx_wait_prev_fence(struct amdgpu_ctx *ctx,
 void amdgpu_ctx_mgr_init(struct amdgpu_ctx_mgr *mgr,
 			 struct amdgpu_device *adev)
 {
+	unsigned int i;
+
 	mgr->adev = adev;
 	mutex_init(&mgr->lock);
 	idr_init(&mgr->ctx_handles);
+
+	for (i = 0; i < AMDGPU_HW_IP_NUM; ++i)
+		atomic64_set(&mgr->time_spend[i], 0);
 }
 
 long amdgpu_ctx_mgr_entity_flush(struct amdgpu_ctx_mgr *mgr, long timeout)
@@ -873,80 +920,38 @@ void amdgpu_ctx_mgr_fini(struct amdgpu_ctx_mgr *mgr)
 	mutex_destroy(&mgr->lock);
 }
 
-static void amdgpu_ctx_fence_time(struct amdgpu_ctx *ctx,
-		struct amdgpu_ctx_entity *centity, ktime_t *total, ktime_t *max)
-{
-	ktime_t now, t1;
-	uint32_t i;
-
-	*total = *max = 0;
-
-	now = ktime_get();
-	for (i = 0; i < amdgpu_sched_jobs; i++) {
-		struct dma_fence *fence;
-		struct drm_sched_fence *s_fence;
-
-		spin_lock(&ctx->ring_lock);
-		fence = dma_fence_get(centity->fences[i]);
-		spin_unlock(&ctx->ring_lock);
-		if (!fence)
-			continue;
-		s_fence = to_drm_sched_fence(fence);
-		if (!dma_fence_is_signaled(&s_fence->scheduled)) {
-			dma_fence_put(fence);
-			continue;
-		}
-		t1 = s_fence->scheduled.timestamp;
-		if (!ktime_before(t1, now)) {
-			dma_fence_put(fence);
-			continue;
-		}
-		if (dma_fence_is_signaled(&s_fence->finished) &&
-			s_fence->finished.timestamp < now)
-			*total += ktime_sub(s_fence->finished.timestamp, t1);
-		else
-			*total += ktime_sub(now, t1);
-		t1 = ktime_sub(now, t1);
-		dma_fence_put(fence);
-		*max = max(t1, *max);
-	}
-}
-
-ktime_t amdgpu_ctx_mgr_fence_usage(struct amdgpu_ctx_mgr *mgr, uint32_t hwip,
-		uint32_t idx, uint64_t *elapsed)
+void amdgpu_ctx_mgr_usage(struct amdgpu_ctx_mgr *mgr,
+			  ktime_t usage[AMDGPU_HW_IP_NUM])
 {
-	struct idr *idp;
 	struct amdgpu_ctx *ctx;
+	unsigned int hw_ip, i;
 	uint32_t id;
-	struct amdgpu_ctx_entity *centity;
-	ktime_t total = 0, max = 0;
 
-	if (idx >= AMDGPU_MAX_ENTITY_NUM)
-		return 0;
-	idp = &mgr->ctx_handles;
+	/*
+	 * This is a little bit racy because it can be that a ctx or a fence are
+	 * destroyed just in the moment we try to account them. But that is ok
+	 * since exactly that case is explicitely allowed by the interface.
+	 */
 	mutex_lock(&mgr->lock);
-	idr_for_each_entry(idp, ctx, id) {
-		ktime_t ttotal, tmax;
+	for (hw_ip = 0; hw_ip < AMDGPU_HW_IP_NUM; ++hw_ip) {
+		uint64_t ns = atomic64_read(&mgr->time_spend[hw_ip]);
 
-		if (!ctx->entities[hwip][idx])
-			continue;
+		usage[hw_ip] = ns_to_ktime(ns);
+	}
 
-		centity = ctx->entities[hwip][idx];
-		amdgpu_ctx_fence_time(ctx, centity, &ttotal, &tmax);
+	idr_for_each_entry(&mgr->ctx_handles, ctx, id) {
+		for (hw_ip = 0; hw_ip < AMDGPU_HW_IP_NUM; ++hw_ip) {
+			for (i = 0; i < amdgpu_ctx_num_entities[hw_ip]; ++i) {
+				struct amdgpu_ctx_entity *centity;
+				ktime_t spend;
 
-		/* Harmonic mean approximation diverges for very small
-		 * values. If ratio < 0.01% ignore
-		 */
-		if (AMDGPU_CTX_FENCE_USAGE_MIN_RATIO(tmax, ttotal))
-			continue;
-
-		total = ktime_add(total, ttotal);
-		max = ktime_after(tmax, max) ? tmax : max;
+				centity = ctx->entities[hw_ip][i];
+				if (!centity)
+					continue;
+				spend = amdgpu_ctx_entity_time(ctx, centity);
+				usage[hw_ip] = ktime_add(usage[hw_ip], spend);
+			}
+		}
 	}
-
 	mutex_unlock(&mgr->lock);
-	if (elapsed)
-		*elapsed = max;
-
-	return total;
 }