drm/i915: Coordinate i915_active with its own mutex

Forgo the struct_mutex serialisation for i915_active, and interpose its
own mutex handling for active/retire.

This is a multi-layered sleight-of-hand. First, we had to ensure that no
active/retire callbacks accidentally inverted the mutex ordering rules,
nor assumed that they were themselves serialised by struct_mutex. More
challenging though, is the rule over updating elements of the active
rbtree. Instead of the whole i915_active now being serialised by
struct_mutex, allocations/rotations of the tree are serialised by the
i915_active.mutex and individual nodes are serialised by the caller
using the i915_timeline.mutex (we need to use nested spinlocks to
interact with the dma_fence callback lists).

The pain point here is that instead of a single mutex around execbuf, we
now have to take a mutex for active tracker (one for each vma, context,
etc) and a couple of spinlocks for each fence update. The improvement in
fine grained locking allowing for multiple concurrent clients
(eventually!) should be worth it in typical loads.

v2: Add some comments that barely elucidate anything :(
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191004134015.13204-6-chris@chris-wilson.co.uk

drivers/gpu/drm/i915/display/intel_frontbuffer.c

@@ -257,7 +257,7 @@ intel_frontbuffer_get(struct drm_i915_gem_object *obj)
 	front->obj = obj;
 	kref_init(&front->ref);
 	atomic_set(&front->bits, 0);
-	i915_active_init(i915, &front->write,
+	i915_active_init(&front->write,
 			 frontbuffer_active,
 			 i915_active_may_sleep(frontbuffer_retire));
 

drivers/gpu/drm/i915/display/intel_overlay.c

@@ -1360,8 +1360,7 @@ void intel_overlay_setup(struct drm_i915_private *dev_priv)
 	overlay->contrast = 75;
 	overlay->saturation = 146;
 
-	i915_active_init(dev_priv,
-			 &overlay->last_flip,
+	i915_active_init(&overlay->last_flip,
 			 NULL, intel_overlay_last_flip_retire);
 
 	ret = get_registers(overlay, OVERLAY_NEEDS_PHYSICAL(dev_priv));

drivers/gpu/drm/i915/gem/i915_gem_context.c

@@ -868,20 +868,18 @@ static int context_barrier_task(struct i915_gem_context *ctx,
 				void (*task)(void *data),
 				void *data)
 {
-	struct drm_i915_private *i915 = ctx->i915;
 	struct context_barrier_task *cb;
 	struct i915_gem_engines_iter it;
 	struct intel_context *ce;
 	int err = 0;
 
-	lockdep_assert_held(&i915->drm.struct_mutex);
 	GEM_BUG_ON(!task);
 
 	cb = kmalloc(sizeof(*cb), GFP_KERNEL);
 	if (!cb)
 		return -ENOMEM;
 
-	i915_active_init(i915, &cb->base, NULL, cb_retire);
+	i915_active_init(&cb->base, NULL, cb_retire);
 	err = i915_active_acquire(&cb->base);
 	if (err) {
 		kfree(cb);

drivers/gpu/drm/i915/gem/i915_gem_object_types.h

@@ -8,6 +8,7 @@
 #define __I915_GEM_OBJECT_TYPES_H__
 
 #include <drm/drm_gem.h>
+#include <uapi/drm/i915_drm.h>
 
 #include "i915_active.h"
 #include "i915_selftest.h"

drivers/gpu/drm/i915/gem/i915_gem_pm.c

@@ -16,14 +16,11 @@ static void call_idle_barriers(struct intel_engine_cs *engine)
 	struct llist_node *node, *next;
 
 	llist_for_each_safe(node, next, llist_del_all(&engine->barrier_tasks)) {
-		struct i915_active_request *active =
+		struct dma_fence_cb *cb =
 			container_of((struct list_head *)node,
-				     typeof(*active), link);
+				     typeof(*cb), node);
 
-		INIT_LIST_HEAD(&active->link);
-		RCU_INIT_POINTER(active->request, NULL);
-
-		active->retire(active, NULL);
+		cb->func(NULL, cb);
 	}
 }
 

drivers/gpu/drm/i915/gt/intel_context.c

@@ -240,7 +240,7 @@ intel_context_init(struct intel_context *ce,
 
 	mutex_init(&ce->pin_mutex);
 
-	i915_active_init(ctx->i915, &ce->active,
+	i915_active_init(&ce->active,
 			 __intel_context_active, __intel_context_retire);
 }
 
@@ -307,7 +307,7 @@ int intel_context_prepare_remote_request(struct intel_context *ce,
 			return err;
 
 		/* Queue this switch after current activity by this context. */
-		err = i915_active_request_set(&tl->last_request, rq);
+		err = i915_active_fence_set(&tl->last_request, rq);
 		mutex_unlock(&tl->mutex);
 		if (err)
 			return err;

drivers/gpu/drm/i915/gt/intel_engine_pool.c

@@ -95,7 +95,7 @@ node_create(struct intel_engine_pool *pool, size_t sz)
 		return ERR_PTR(-ENOMEM);
 
 	node->pool = pool;
-	i915_active_init(engine->i915, &node->active, pool_active, pool_retire);
+	i915_active_init(&node->active, pool_active, pool_retire);
 
 	obj = i915_gem_object_create_internal(engine->i915, sz);
 	if (IS_ERR(obj)) {

drivers/gpu/drm/i915/gt/intel_reset.c

@@ -844,10 +844,10 @@ static bool __intel_gt_unset_wedged(struct intel_gt *gt)
 	 */
 	spin_lock_irqsave(&timelines->lock, flags);
 	list_for_each_entry(tl, &timelines->active_list, link) {
-		struct i915_request *rq;
+		struct dma_fence *fence;
 
-		rq = i915_active_request_get_unlocked(&tl->last_request);
-		if (!rq)
+		fence = i915_active_fence_get(&tl->last_request);
+		if (!fence)
 			continue;
 
 		spin_unlock_irqrestore(&timelines->lock, flags);
@@ -859,8 +859,8 @@ static bool __intel_gt_unset_wedged(struct intel_gt *gt)
 		 * (I915_FENCE_TIMEOUT) so this wait should not be unbounded
 		 * in the worst case.
 		 */
-		dma_fence_default_wait(&rq->fence, false, MAX_SCHEDULE_TIMEOUT);
-		i915_request_put(rq);
+		dma_fence_default_wait(fence, false, MAX_SCHEDULE_TIMEOUT);
+		dma_fence_put(fence);
 
 		/* Restart iteration after droping lock */
 		spin_lock_irqsave(&timelines->lock, flags);

drivers/gpu/drm/i915/gt/intel_timeline.c

@@ -178,8 +178,7 @@ cacheline_alloc(struct intel_timeline_hwsp *hwsp, unsigned int cacheline)
 	cl->hwsp = hwsp;
 	cl->vaddr = page_pack_bits(vaddr, cacheline);
 
-	i915_active_init(hwsp->gt->i915, &cl->active,
-			 __cacheline_active, __cacheline_retire);
+	i915_active_init(&cl->active, __cacheline_active, __cacheline_retire);
 
 	return cl;
 }
@@ -255,7 +254,7 @@ int intel_timeline_init(struct intel_timeline *timeline,
 
 	mutex_init(&timeline->mutex);
 
-	INIT_ACTIVE_REQUEST(&timeline->last_request, &timeline->mutex);
+	INIT_ACTIVE_FENCE(&timeline->last_request, &timeline->mutex);
 	INIT_LIST_HEAD(&timeline->requests);
 
 	i915_syncmap_init(&timeline->sync);
@@ -443,7 +442,7 @@ __intel_timeline_get_seqno(struct intel_timeline *tl,
 	 * free it after the current request is retired, which ensures that
 	 * all writes into the cacheline from previous requests are complete.
 	 */
-	err = i915_active_ref(&tl->hwsp_cacheline->active, tl, rq);
+	err = i915_active_ref(&tl->hwsp_cacheline->active, tl, &rq->fence);
 	if (err)
 		goto err_cacheline;
 

drivers/gpu/drm/i915/gt/intel_timeline_types.h

@@ -58,12 +58,13 @@ struct intel_timeline {
 	 */
 	struct list_head requests;
 
-	/* Contains an RCU guarded pointer to the last request. No reference is
+	/*
+	 * Contains an RCU guarded pointer to the last request. No reference is
 	 * held to the request, users must carefully acquire a reference to
-	 * the request using i915_active_request_get_request_rcu(), or hold the
-	 * struct_mutex.
+	 * the request using i915_active_fence_get(), or manage the RCU
+	 * protection themselves (cf the i915_active_fence API).
 	 */
-	struct i915_active_request last_request;
+	struct i915_active_fence last_request;
 
 	/**
 	 * We track the most recent seqno that we wait on in every context so

drivers/gpu/drm/i915/gt/selftest_context.c

@@ -47,24 +47,20 @@ static int context_sync(struct intel_context *ce)
 
 	mutex_lock(&tl->mutex);
 	do {
-		struct i915_request *rq;
+		struct dma_fence *fence;
 		long timeout;
 
-		rcu_read_lock();
-		rq = rcu_dereference(tl->last_request.request);
-		if (rq)
-			rq = i915_request_get_rcu(rq);
-		rcu_read_unlock();
-		if (!rq)
+		fence = i915_active_fence_get(&tl->last_request);
+		if (!fence)
 			break;
 
-		timeout = i915_request_wait(rq, 0, HZ / 10);
+		timeout = dma_fence_wait_timeout(fence, false, HZ / 10);
 		if (timeout < 0)
 			err = timeout;
 		else
-			i915_request_retire_upto(rq);
+			i915_request_retire_upto(to_request(fence));
 
-		i915_request_put(rq);
+		dma_fence_put(fence);
 	} while (!err);
 	mutex_unlock(&tl->mutex);
 

drivers/gpu/drm/i915/gt/selftest_lrc.c

@@ -1172,9 +1172,13 @@ static struct i915_request *dummy_request(struct intel_engine_cs *engine)
 	if (!rq)
 		return NULL;
 
-	INIT_LIST_HEAD(&rq->active_list);
 	rq->engine = engine;
 
+	spin_lock_init(&rq->lock);
+	INIT_LIST_HEAD(&rq->fence.cb_list);
+	rq->fence.lock = &rq->lock;
+	rq->fence.ops = &i915_fence_ops;
+
 	i915_sched_node_init(&rq->sched);
 
 	/* mark this request as permanently incomplete */
@@ -1267,8 +1271,8 @@ static int live_suppress_wait_preempt(void *arg)
 				}
 
 				/* Disable NEWCLIENT promotion */
-				__i915_active_request_set(&i915_request_timeline(rq[i])->last_request,
-							  dummy);
+				__i915_active_fence_set(&i915_request_timeline(rq[i])->last_request,
+							&dummy->fence);
 				i915_request_add(rq[i]);
 			}
 

drivers/gpu/drm/i915/gt/selftests/mock_timeline.c

@@ -15,7 +15,7 @@ void mock_timeline_init(struct intel_timeline *timeline, u64 context)
 
 	mutex_init(&timeline->mutex);
 
-	INIT_ACTIVE_REQUEST(&timeline->last_request, &timeline->mutex);
+	INIT_ACTIVE_FENCE(&timeline->last_request, &timeline->mutex);
 	INIT_LIST_HEAD(&timeline->requests);
 
 	i915_syncmap_init(&timeline->sync);

drivers/gpu/drm/i915/gvt/scheduler.c

@@ -385,11 +385,8 @@ intel_gvt_workload_req_alloc(struct intel_vgpu_workload *workload)
 {
 	struct intel_vgpu *vgpu = workload->vgpu;
 	struct intel_vgpu_submission *s = &vgpu->submission;
-	struct drm_i915_private *dev_priv = vgpu->gvt->dev_priv;
 	struct i915_request *rq;
 
-	lockdep_assert_held(&dev_priv->drm.struct_mutex);
-
 	if (workload->req)
 		return 0;
 

drivers/gpu/drm/i915/i915_active_types.h

@@ -17,17 +17,9 @@
 
 #include "i915_utils.h"
 
-struct drm_i915_private;
-struct i915_active_request;
-struct i915_request;
-
-typedef void (*i915_active_retire_fn)(struct i915_active_request *,
-				      struct i915_request *);
-
-struct i915_active_request {
-	struct i915_request __rcu *request;
-	struct list_head link;
-	i915_active_retire_fn retire;
+struct i915_active_fence {
+	struct dma_fence __rcu *fence;
+	struct dma_fence_cb cb;
 #if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM)
 	/*
 	 * Incorporeal!
@@ -53,20 +45,17 @@ struct active_node;
 #define i915_active_may_sleep(fn) ptr_pack_bits(&(fn), I915_ACTIVE_MAY_SLEEP, 2)
 
 struct i915_active {
-	struct drm_i915_private *i915;
+	atomic_t count;
+	struct mutex mutex;
 
 	struct active_node *cache;
 	struct rb_root tree;
-	struct mutex mutex;
-	atomic_t count;
 
 	/* Preallocated "exclusive" node */
-	struct dma_fence __rcu *excl;
-	struct dma_fence_cb excl_cb;
+	struct i915_active_fence excl;
 
 	unsigned long flags;
 #define I915_ACTIVE_RETIRE_SLEEPS BIT(0)
-#define I915_ACTIVE_GRAB_BIT 1
 
 	int (*active)(struct i915_active *ref);
 	void (*retire)(struct i915_active *ref);

drivers/gpu/drm/i915/i915_gem.c

@@ -892,28 +892,38 @@ wait_for_timelines(struct intel_gt *gt, unsigned int wait, long timeout)
 
 	spin_lock_irqsave(&timelines->lock, flags);
 	list_for_each_entry(tl, &timelines->active_list, link) {
-		struct i915_request *rq;
+		struct dma_fence *fence;
 
-		rq = i915_active_request_get_unlocked(&tl->last_request);
-		if (!rq)
+		fence = i915_active_fence_get(&tl->last_request);
+		if (!fence)
 			continue;
 
 		spin_unlock_irqrestore(&timelines->lock, flags);
 
+		if (!dma_fence_is_i915(fence)) {
+			timeout = dma_fence_wait_timeout(fence,
+							 flags & I915_WAIT_INTERRUPTIBLE,
+							 timeout);
+		} else {
+			struct i915_request *rq = to_request(fence);
+
 			/*
 			 * "Race-to-idle".
 			 *
-		 * Switching to the kernel context is often used a synchronous
-		 * step prior to idling, e.g. in suspend for flushing all
-		 * current operations to memory before sleeping. These we
-		 * want to complete as quickly as possible to avoid prolonged
-		 * stalls, so allow the gpu to boost to maximum clocks.
+			 * Switching to the kernel context is often used as
+			 * a synchronous step prior to idling, e.g. in suspend
+			 * for flushing all current operations to memory before
+			 * sleeping. These we want to complete as quickly as
+			 * possible to avoid prolonged stalls, so allow the gpu
+			 * to boost to maximum clocks.
 			 */
-		if (wait & I915_WAIT_FOR_IDLE_BOOST)
+			if (flags & I915_WAIT_FOR_IDLE_BOOST)
 				gen6_rps_boost(rq);
 
-		timeout = i915_request_wait(rq, wait, timeout);
-		i915_request_put(rq);
+			timeout = i915_request_wait(rq, flags, timeout);
+		}
+
+		dma_fence_put(fence);
 		if (timeout < 0)
 			return timeout;
 

drivers/gpu/drm/i915/i915_gem_gtt.c

@@ -1861,7 +1861,6 @@ static const struct i915_vma_ops pd_vma_ops = {
 
 static struct i915_vma *pd_vma_create(struct gen6_ppgtt *ppgtt, int size)
 {
-	struct drm_i915_private *i915 = ppgtt->base.vm.i915;
 	struct i915_ggtt *ggtt = ppgtt->base.vm.gt->ggtt;
 	struct i915_vma *vma;
 
@@ -1872,7 +1871,7 @@ static struct i915_vma *pd_vma_create(struct gen6_ppgtt *ppgtt, int size)
 	if (!vma)
 		return ERR_PTR(-ENOMEM);
 
-	i915_active_init(i915, &vma->active, NULL, NULL);
+	i915_active_init(&vma->active, NULL, NULL);
 
 	mutex_init(&vma->pages_mutex);
 	vma->vm = i915_vm_get(&ggtt->vm);

drivers/gpu/drm/i915/i915_gpu_error.c

@@ -1299,7 +1299,7 @@ capture_vma(struct capture_vma *next,
 	if (!c)
 		return next;
 
-	if (!i915_active_trygrab(&vma->active)) {
+	if (!i915_active_acquire_if_busy(&vma->active)) {
 		kfree(c);
 		return next;
 	}
@@ -1439,7 +1439,7 @@ gem_record_rings(struct i915_gpu_state *error, struct compress *compress)
 			*this->slot =
 				i915_error_object_create(i915, vma, compress);
 
-			i915_active_ungrab(&vma->active);
+			i915_active_release(&vma->active);
 			i915_vma_put(vma);
 
 			capture = this->next;

drivers/gpu/drm/i915/i915_request.c

@@ -218,8 +218,6 @@ static void remove_from_engine(struct i915_request *rq)
 
 static bool i915_request_retire(struct i915_request *rq)
 {
-	struct i915_active_request *active, *next;
-
 	if (!i915_request_completed(rq))
 		return false;
 
@@ -244,35 +242,6 @@ static bool i915_request_retire(struct i915_request *rq)
 				  &i915_request_timeline(rq)->requests));
 	rq->ring->head = rq->postfix;
 
-	/*
-	 * Walk through the active list, calling retire on each. This allows
-	 * objects to track their GPU activity and mark themselves as idle
-	 * when their *last* active request is completed (updating state
-	 * tracking lists for eviction, active references for GEM, etc).
-	 *
-	 * As the ->retire() may free the node, we decouple it first and
-	 * pass along the auxiliary information (to avoid dereferencing
-	 * the node after the callback).
-	 */
-	list_for_each_entry_safe(active, next, &rq->active_list, link) {
-		/*
-		 * In microbenchmarks or focusing upon time inside the kernel,
-		 * we may spend an inordinate amount of time simply handling
-		 * the retirement of requests and processing their callbacks.
-		 * Of which, this loop itself is particularly hot due to the
-		 * cache misses when jumping around the list of
-		 * i915_active_request.  So we try to keep this loop as
-		 * streamlined as possible and also prefetch the next
-		 * i915_active_request to try and hide the likely cache miss.
-		 */
-		prefetchw(next);
-
-		INIT_LIST_HEAD(&active->link);
-		RCU_INIT_POINTER(active->request, NULL);
-
-		active->retire(active, rq);
-	}
-
 	local_irq_disable();
 
 	/*
@@ -704,7 +673,6 @@ __i915_request_create(struct intel_context *ce, gfp_t gfp)
 	rq->flags = 0;
 	rq->execution_mask = ALL_ENGINES;
 
-	INIT_LIST_HEAD(&rq->active_list);
 	INIT_LIST_HEAD(&rq->execute_cb);
 
 	/*
@@ -743,7 +711,6 @@ __i915_request_create(struct intel_context *ce, gfp_t gfp)
 	ce->ring->emit = rq->head;
 
 	/* Make sure we didn't add ourselves to external state before freeing */
-	GEM_BUG_ON(!list_empty(&rq->active_list));
 	GEM_BUG_ON(!list_empty(&rq->sched.signalers_list));
 	GEM_BUG_ON(!list_empty(&rq->sched.waiters_list));
 
@@ -1174,8 +1141,8 @@ __i915_request_add_to_timeline(struct i915_request *rq)
 	 * precludes optimising to use semaphores serialisation of a single
 	 * timeline across engines.
 	 */
-	prev = rcu_dereference_protected(timeline->last_request.request,
-					 lockdep_is_held(&timeline->mutex));
+	prev = to_request(__i915_active_fence_set(&timeline->last_request,
+						  &rq->fence));
 	if (prev && !i915_request_completed(prev)) {
 		if (is_power_of_2(prev->engine->mask | rq->engine->mask))
 			i915_sw_fence_await_sw_fence(&rq->submit,
@@ -1200,7 +1167,6 @@ __i915_request_add_to_timeline(struct i915_request *rq)
 	 * us, the timeline will hold its seqno which is later than ours.
 	 */
 	GEM_BUG_ON(timeline->seqno != rq->fence.seqno);
-	__i915_active_request_set(&timeline->last_request, rq);
 
 	return prev;
 }

drivers/gpu/drm/i915/i915_request.h

@@ -211,7 +211,6 @@ struct i915_request {
 	 * on the active_list (of their final request).
 	 */
 	struct i915_capture_list *capture_list;
-	struct list_head active_list;
 
 	/** Time at which this request was emitted, in jiffies. */
 	unsigned long emitted_jiffies;

drivers/gpu/drm/i915/i915_vma.c

@@ -120,8 +120,7 @@ vma_create(struct drm_i915_gem_object *obj,
 	vma->size = obj->base.size;
 	vma->display_alignment = I915_GTT_MIN_ALIGNMENT;
 
-	i915_active_init(vm->i915, &vma->active,
-			 __i915_vma_active, __i915_vma_retire);
+	i915_active_init(&vma->active, __i915_vma_active, __i915_vma_retire);
 
 	/* Declare ourselves safe for use inside shrinkers */
 	if (IS_ENABLED(CONFIG_LOCKDEP)) {
@@ -1148,6 +1147,7 @@ int __i915_vma_unbind(struct i915_vma *vma)
 	if (ret)
 		return ret;
 
+	GEM_BUG_ON(i915_vma_is_active(vma));
 	if (i915_vma_is_pinned(vma)) {
 		vma_print_allocator(vma, "is pinned");
 		return -EBUSY;

drivers/gpu/drm/i915/selftests/i915_active.c

@@ -68,7 +68,7 @@ static struct live_active *__live_alloc(struct drm_i915_private *i915)
 		return NULL;
 
 	kref_init(&active->ref);
-	i915_active_init(i915, &active->base, __live_active, __live_retire);
+	i915_active_init(&active->base, __live_active, __live_retire);
 
 	return active;
 }
@@ -146,19 +146,13 @@ static int live_active_wait(void *arg)
 {
 	struct drm_i915_private *i915 = arg;
 	struct live_active *active;
-	intel_wakeref_t wakeref;
 	int err = 0;
 
 	/* Check that we get a callback when requests retire upon waiting */
 
-	mutex_lock(&i915->drm.struct_mutex);
-	wakeref = intel_runtime_pm_get(&i915->runtime_pm);
-
 	active = __live_active_setup(i915);
-	if (IS_ERR(active)) {
-		err = PTR_ERR(active);
-		goto err;
-	}
+	if (IS_ERR(active))
+		return PTR_ERR(active);
 
 	i915_active_wait(&active->base);
 	if (!READ_ONCE(active->retired)) {
@@ -168,11 +162,9 @@ static int live_active_wait(void *arg)
 
 	__live_put(active);
 
+	mutex_lock(&i915->drm.struct_mutex);
 	if (igt_flush_test(i915, I915_WAIT_LOCKED))
 		err = -EIO;
-
-err:
-	intel_runtime_pm_put(&i915->runtime_pm, wakeref);
 	mutex_unlock(&i915->drm.struct_mutex);
 
 	return err;
@@ -182,23 +174,19 @@ static int live_active_retire(void *arg)
 {
 	struct drm_i915_private *i915 = arg;
 	struct live_active *active;
-	intel_wakeref_t wakeref;
 	int err = 0;
 
 	/* Check that we get a callback when requests are indirectly retired */
 
-	mutex_lock(&i915->drm.struct_mutex);
-	wakeref = intel_runtime_pm_get(&i915->runtime_pm);
-
 	active = __live_active_setup(i915);
-	if (IS_ERR(active)) {
-		err = PTR_ERR(active);
-		goto err;
-	}
+	if (IS_ERR(active))
+		return PTR_ERR(active);
 
 	/* waits for & retires all requests */
+	mutex_lock(&i915->drm.struct_mutex);
 	if (igt_flush_test(i915, I915_WAIT_LOCKED))
 		err = -EIO;
+	mutex_unlock(&i915->drm.struct_mutex);
 
 	if (!READ_ONCE(active->retired)) {
 		pr_err("i915_active not retired after flushing!\n");
@@ -207,10 +195,6 @@ static int live_active_retire(void *arg)
 
 	__live_put(active);
 
-err:
-	intel_runtime_pm_put(&i915->runtime_pm, wakeref);
-	mutex_unlock(&i915->drm.struct_mutex);
-
 	return err;
 }