1/*
  2 * SPDX-License-Identifier: MIT
  3 *
  4 * Copyright © 2018 Intel Corporation
  5 */
  6
  7#include <linux/mutex.h>
  8
  9#include "i915_drv.h"
 10#include "i915_globals.h"
 11#include "i915_request.h"
 12#include "i915_scheduler.h"
 13
 14static struct i915_global_scheduler {
 15	struct i915_global base;
 16	struct kmem_cache *slab_dependencies;
 17	struct kmem_cache *slab_priorities;
 18} global;
 19
 20static DEFINE_SPINLOCK(schedule_lock);
 21
 22static const struct i915_request *
 23node_to_request(const struct i915_sched_node *node)
 24{
 25	return container_of(node, const struct i915_request, sched);
 26}
 27
 28static inline bool node_started(const struct i915_sched_node *node)
 29{
 30	return i915_request_started(node_to_request(node));
 31}
 32
 33static inline bool node_signaled(const struct i915_sched_node *node)
 34{
 35	return i915_request_completed(node_to_request(node));
 36}
 37
 38static inline struct i915_priolist *to_priolist(struct rb_node *rb)
 39{
 40	return rb_entry(rb, struct i915_priolist, node);
 41}
 42
 43static void assert_priolists(struct intel_engine_execlists * const execlists)
 44{
 45	struct rb_node *rb;
 46	long last_prio, i;
 47
 48	if (!IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM))
 49		return;
 50
 51	GEM_BUG_ON(rb_first_cached(&execlists->queue) !=
 52		   rb_first(&execlists->queue.rb_root));
 53
 54	last_prio = INT_MAX;
 55	for (rb = rb_first_cached(&execlists->queue); rb; rb = rb_next(rb)) {
 56		const struct i915_priolist *p = to_priolist(rb);
 57
 58		GEM_BUG_ON(p->priority > last_prio);
 59		last_prio = p->priority;
 60
 61		GEM_BUG_ON(!p->used);
 62		for (i = 0; i < ARRAY_SIZE(p->requests); i++) {
 63			if (list_empty(&p->requests[i]))
 64				continue;
 65
 66			GEM_BUG_ON(!(p->used & BIT(i)));
 67		}
 68	}
 69}
 70
 71struct list_head *
 72i915_sched_lookup_priolist(struct intel_engine_cs *engine, int prio)
 73{
 74	struct intel_engine_execlists * const execlists = &engine->execlists;
 75	struct i915_priolist *p;
 76	struct rb_node **parent, *rb;
 77	bool first = true;
 78	int idx, i;
 79
 80	lockdep_assert_held(&engine->active.lock);
 81	assert_priolists(execlists);
 82
 83	/* buckets sorted from highest [in slot 0] to lowest priority */
 84	idx = I915_PRIORITY_COUNT - (prio & I915_PRIORITY_MASK) - 1;
 85	prio >>= I915_USER_PRIORITY_SHIFT;
 86	if (unlikely(execlists->no_priolist))
 87		prio = I915_PRIORITY_NORMAL;
 88
 89find_priolist:
 90	/* most positive priority is scheduled first, equal priorities fifo */
 91	rb = NULL;
 92	parent = &execlists->queue.rb_root.rb_node;
 93	while (*parent) {
 94		rb = *parent;
 95		p = to_priolist(rb);
 96		if (prio > p->priority) {
 97			parent = &rb->rb_left;
 98		} else if (prio < p->priority) {
 99			parent = &rb->rb_right;
100			first = false;
101		} else {
102			goto out;
103		}
104	}
105
106	if (prio == I915_PRIORITY_NORMAL) {
107		p = &execlists->default_priolist;
108	} else {
109		p = kmem_cache_alloc(global.slab_priorities, GFP_ATOMIC);
110		/* Convert an allocation failure to a priority bump */
111		if (unlikely(!p)) {
112			prio = I915_PRIORITY_NORMAL; /* recurses just once */
113
114			/* To maintain ordering with all rendering, after an
115			 * allocation failure we have to disable all scheduling.
116			 * Requests will then be executed in fifo, and schedule
117			 * will ensure that dependencies are emitted in fifo.
118			 * There will be still some reordering with existing
119			 * requests, so if userspace lied about their
120			 * dependencies that reordering may be visible.
121			 */
122			execlists->no_priolist = true;
123			goto find_priolist;
124		}
125	}
126
127	p->priority = prio;
128	for (i = 0; i < ARRAY_SIZE(p->requests); i++)
129		INIT_LIST_HEAD(&p->requests[i]);
130	rb_link_node(&p->node, rb, parent);
131	rb_insert_color_cached(&p->node, &execlists->queue, first);
132	p->used = 0;
133
134out:
135	p->used |= BIT(idx);
136	return &p->requests[idx];
137}
138
139void __i915_priolist_free(struct i915_priolist *p)
140{
141	kmem_cache_free(global.slab_priorities, p);
142}
143
144struct sched_cache {
145	struct list_head *priolist;
146};
147
148static struct intel_engine_cs *
149sched_lock_engine(const struct i915_sched_node *node,
150		  struct intel_engine_cs *locked,
151		  struct sched_cache *cache)
152{
153	const struct i915_request *rq = node_to_request(node);
154	struct intel_engine_cs *engine;
155
156	GEM_BUG_ON(!locked);
157
158	/*
159	 * Virtual engines complicate acquiring the engine timeline lock,
160	 * as their rq->engine pointer is not stable until under that
161	 * engine lock. The simple ploy we use is to take the lock then
162	 * check that the rq still belongs to the newly locked engine.
163	 */
164	while (locked != (engine = READ_ONCE(rq->engine))) {
165		spin_unlock(&locked->active.lock);
166		memset(cache, 0, sizeof(*cache));
167		spin_lock(&engine->active.lock);
168		locked = engine;
169	}
170
171	GEM_BUG_ON(locked != engine);
172	return locked;
173}
174
175static inline int rq_prio(const struct i915_request *rq)
176{
177	return rq->sched.attr.priority;
178}
179
180static inline bool need_preempt(int prio, int active)
181{
182	/*
183	 * Allow preemption of low -> normal -> high, but we do
184	 * not allow low priority tasks to preempt other low priority
185	 * tasks under the impression that latency for low priority
186	 * tasks does not matter (as much as background throughput),
187	 * so kiss.
188	 */
189	return prio >= max(I915_PRIORITY_NORMAL, active);
190}
191
192static void kick_submission(struct intel_engine_cs *engine,
193			    const struct i915_request *rq,
194			    int prio)
195{
196	const struct i915_request *inflight;
197
198	/*
199	 * We only need to kick the tasklet once for the high priority
200	 * new context we add into the queue.
201	 */
202	if (prio <= engine->execlists.queue_priority_hint)
203		return;
204
205	rcu_read_lock();
206
207	/* Nothing currently active? We're overdue for a submission! */
208	inflight = execlists_active(&engine->execlists);
209	if (!inflight)
210		goto unlock;
211
212	/*
213	 * If we are already the currently executing context, don't
214	 * bother evaluating if we should preempt ourselves.
215	 */
216	if (inflight->context == rq->context)
217		goto unlock;
218
219	ENGINE_TRACE(engine,
220		     "bumping queue-priority-hint:%d for rq:%llx:%lld, inflight:%llx:%lld prio %d\n",
221		     prio,
222		     rq->fence.context, rq->fence.seqno,
223		     inflight->fence.context, inflight->fence.seqno,
224		     inflight->sched.attr.priority);
225
226	engine->execlists.queue_priority_hint = prio;
227	if (need_preempt(prio, rq_prio(inflight)))
228		tasklet_hi_schedule(&engine->execlists.tasklet);
229
230unlock:
231	rcu_read_unlock();
232}
233
234static void __i915_schedule(struct i915_sched_node *node,
235			    const struct i915_sched_attr *attr)
236{
237	const int prio = max(attr->priority, node->attr.priority);
238	struct intel_engine_cs *engine;
239	struct i915_dependency *dep, *p;
240	struct i915_dependency stack;
241	struct sched_cache cache;
242	LIST_HEAD(dfs);
243
244	/* Needed in order to use the temporary link inside i915_dependency */
245	lockdep_assert_held(&schedule_lock);
246	GEM_BUG_ON(prio == I915_PRIORITY_INVALID);
247
248	if (node_signaled(node))
249		return;
250
251	stack.signaler = node;
252	list_add(&stack.dfs_link, &dfs);
253
254	/*
255	 * Recursively bump all dependent priorities to match the new request.
256	 *
257	 * A naive approach would be to use recursion:
258	 * static void update_priorities(struct i915_sched_node *node, prio) {
259	 *	list_for_each_entry(dep, &node->signalers_list, signal_link)
260	 *		update_priorities(dep->signal, prio)
261	 *	queue_request(node);
262	 * }
263	 * but that may have unlimited recursion depth and so runs a very
264	 * real risk of overunning the kernel stack. Instead, we build
265	 * a flat list of all dependencies starting with the current request.
266	 * As we walk the list of dependencies, we add all of its dependencies
267	 * to the end of the list (this may include an already visited
268	 * request) and continue to walk onwards onto the new dependencies. The
269	 * end result is a topological list of requests in reverse order, the
270	 * last element in the list is the request we must execute first.
271	 */
272	list_for_each_entry(dep, &dfs, dfs_link) {
273		struct i915_sched_node *node = dep->signaler;
274
275		/* If we are already flying, we know we have no signalers */
276		if (node_started(node))
277			continue;
278
279		/*
280		 * Within an engine, there can be no cycle, but we may
281		 * refer to the same dependency chain multiple times
282		 * (redundant dependencies are not eliminated) and across
283		 * engines.
284		 */
285		list_for_each_entry(p, &node->signalers_list, signal_link) {
286			GEM_BUG_ON(p == dep); /* no cycles! */
287
288			if (node_signaled(p->signaler))
289				continue;
290
291			if (prio > READ_ONCE(p->signaler->attr.priority))
292				list_move_tail(&p->dfs_link, &dfs);
293		}
294	}
295
296	/*
297	 * If we didn't need to bump any existing priorities, and we haven't
298	 * yet submitted this request (i.e. there is no potential race with
299	 * execlists_submit_request()), we can set our own priority and skip
300	 * acquiring the engine locks.
301	 */
302	if (node->attr.priority == I915_PRIORITY_INVALID) {
303		GEM_BUG_ON(!list_empty(&node->link));
304		node->attr = *attr;
305
306		if (stack.dfs_link.next == stack.dfs_link.prev)
307			return;
308
309		__list_del_entry(&stack.dfs_link);
310	}
311
312	memset(&cache, 0, sizeof(cache));
313	engine = node_to_request(node)->engine;
314	spin_lock(&engine->active.lock);
315
316	/* Fifo and depth-first replacement ensure our deps execute before us */
317	engine = sched_lock_engine(node, engine, &cache);
318	list_for_each_entry_safe_reverse(dep, p, &dfs, dfs_link) {
319		INIT_LIST_HEAD(&dep->dfs_link);
320
321		node = dep->signaler;
322		engine = sched_lock_engine(node, engine, &cache);
323		lockdep_assert_held(&engine->active.lock);
324
325		/* Recheck after acquiring the engine->timeline.lock */
326		if (prio <= node->attr.priority || node_signaled(node))
327			continue;
328
329		GEM_BUG_ON(node_to_request(node)->engine != engine);
330
331		WRITE_ONCE(node->attr.priority, prio);
332
333		/*
334		 * Once the request is ready, it will be placed into the
335		 * priority lists and then onto the HW runlist. Before the
336		 * request is ready, it does not contribute to our preemption
337		 * decisions and we can safely ignore it, as it will, and
338		 * any preemption required, be dealt with upon submission.
339		 * See engine->submit_request()
340		 */
341		if (list_empty(&node->link))
342			continue;
343
344		if (i915_request_in_priority_queue(node_to_request(node))) {
345			if (!cache.priolist)
346				cache.priolist =
347					i915_sched_lookup_priolist(engine,
348								   prio);
349			list_move_tail(&node->link, cache.priolist);
350		}
351
352		/* Defer (tasklet) submission until after all of our updates. */
353		kick_submission(engine, node_to_request(node), prio);
354	}
355
356	spin_unlock(&engine->active.lock);
357}
358
359void i915_schedule(struct i915_request *rq, const struct i915_sched_attr *attr)
360{
361	spin_lock_irq(&schedule_lock);
362	__i915_schedule(&rq->sched, attr);
363	spin_unlock_irq(&schedule_lock);
364}
365
366static void __bump_priority(struct i915_sched_node *node, unsigned int bump)
367{
368	struct i915_sched_attr attr = node->attr;
369
370	if (attr.priority & bump)
371		return;
372
373	attr.priority |= bump;
374	__i915_schedule(node, &attr);
375}
376
377void i915_schedule_bump_priority(struct i915_request *rq, unsigned int bump)
378{
379	unsigned long flags;
380
381	GEM_BUG_ON(bump & ~I915_PRIORITY_MASK);
382	if (READ_ONCE(rq->sched.attr.priority) & bump)
383		return;
384
385	spin_lock_irqsave(&schedule_lock, flags);
386	__bump_priority(&rq->sched, bump);
387	spin_unlock_irqrestore(&schedule_lock, flags);
388}
389
390void i915_sched_node_init(struct i915_sched_node *node)
391{
392	INIT_LIST_HEAD(&node->signalers_list);
393	INIT_LIST_HEAD(&node->waiters_list);
394	INIT_LIST_HEAD(&node->link);
395
396	i915_sched_node_reinit(node);
397}
398
399void i915_sched_node_reinit(struct i915_sched_node *node)
400{
401	node->attr.priority = I915_PRIORITY_INVALID;
402	node->semaphores = 0;
403	node->flags = 0;
404
405	GEM_BUG_ON(!list_empty(&node->signalers_list));
406	GEM_BUG_ON(!list_empty(&node->waiters_list));
407	GEM_BUG_ON(!list_empty(&node->link));
408}
409
410static struct i915_dependency *
411i915_dependency_alloc(void)
412{
413	return kmem_cache_alloc(global.slab_dependencies, GFP_KERNEL);
414}
415
416static void
417i915_dependency_free(struct i915_dependency *dep)
418{
419	kmem_cache_free(global.slab_dependencies, dep);
420}
421
422bool __i915_sched_node_add_dependency(struct i915_sched_node *node,
423				      struct i915_sched_node *signal,
424				      struct i915_dependency *dep,
425				      unsigned long flags)
426{
427	bool ret = false;
428
429	spin_lock_irq(&schedule_lock);
430
431	if (!node_signaled(signal)) {
432		INIT_LIST_HEAD(&dep->dfs_link);
433		dep->signaler = signal;
434		dep->waiter = node;
435		dep->flags = flags;
436
437		/* All set, now publish. Beware the lockless walkers. */
438		list_add_rcu(&dep->signal_link, &node->signalers_list);
439		list_add_rcu(&dep->wait_link, &signal->waiters_list);
440
441		/* Propagate the chains */
442		node->flags |= signal->flags;
443		ret = true;
444	}
445
446	spin_unlock_irq(&schedule_lock);
447
448	return ret;
449}
450
451int i915_sched_node_add_dependency(struct i915_sched_node *node,
452				   struct i915_sched_node *signal,
453				   unsigned long flags)
454{
455	struct i915_dependency *dep;
456
457	dep = i915_dependency_alloc();
458	if (!dep)
459		return -ENOMEM;
460
461	local_bh_disable();
462
463	if (!__i915_sched_node_add_dependency(node, signal, dep,
464					      flags | I915_DEPENDENCY_ALLOC))
465		i915_dependency_free(dep);
466
467	local_bh_enable(); /* kick submission tasklet */
468
469	return 0;
470}
471
472void i915_sched_node_fini(struct i915_sched_node *node)
473{
474	struct i915_dependency *dep, *tmp;
475
476	spin_lock_irq(&schedule_lock);
477
478	/*
479	 * Everyone we depended upon (the fences we wait to be signaled)
480	 * should retire before us and remove themselves from our list.
481	 * However, retirement is run independently on each timeline and
482	 * so we may be called out-of-order.
483	 */
484	list_for_each_entry_safe(dep, tmp, &node->signalers_list, signal_link) {
485		GEM_BUG_ON(!list_empty(&dep->dfs_link));
486
487		list_del_rcu(&dep->wait_link);
488		if (dep->flags & I915_DEPENDENCY_ALLOC)
489			i915_dependency_free(dep);
490	}
491	INIT_LIST_HEAD(&node->signalers_list);
492
493	/* Remove ourselves from everyone who depends upon us */
494	list_for_each_entry_safe(dep, tmp, &node->waiters_list, wait_link) {
495		GEM_BUG_ON(dep->signaler != node);
496		GEM_BUG_ON(!list_empty(&dep->dfs_link));
497
498		list_del_rcu(&dep->signal_link);
499		if (dep->flags & I915_DEPENDENCY_ALLOC)
500			i915_dependency_free(dep);
501	}
502	INIT_LIST_HEAD(&node->waiters_list);
503
504	spin_unlock_irq(&schedule_lock);
505}
506
507static void i915_global_scheduler_shrink(void)
508{
509	kmem_cache_shrink(global.slab_dependencies);
510	kmem_cache_shrink(global.slab_priorities);
511}
512
513static void i915_global_scheduler_exit(void)
514{
515	kmem_cache_destroy(global.slab_dependencies);
516	kmem_cache_destroy(global.slab_priorities);
517}
518
519static struct i915_global_scheduler global = { {
520	.shrink = i915_global_scheduler_shrink,
521	.exit = i915_global_scheduler_exit,
522} };
523
524int __init i915_global_scheduler_init(void)
525{
526	global.slab_dependencies = KMEM_CACHE(i915_dependency,
527					      SLAB_HWCACHE_ALIGN |
528					      SLAB_TYPESAFE_BY_RCU);
529	if (!global.slab_dependencies)
530		return -ENOMEM;
531
532	global.slab_priorities = KMEM_CACHE(i915_priolist,
533					    SLAB_HWCACHE_ALIGN);
534	if (!global.slab_priorities)
535		goto err_priorities;
536
537	i915_global_register(&global.base);
538	return 0;
539
540err_priorities:
541	kmem_cache_destroy(global.slab_priorities);
542	return -ENOMEM;
543}