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 >> I915_USER_PRIORITY_SHIFT) + 1;
 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 | __NO_PREEMPTION;
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, or if
215	 * we expect nothing to change as a result of running the
216	 * tasklet, i.e. we have not change the priority queue
217	 * sufficiently to oust the running context.
218	 */
219	if (inflight->hw_context == rq->hw_context)
220		goto unlock;
221
222	engine->execlists.queue_priority_hint = prio;
223	if (need_preempt(prio, rq_prio(inflight)))
224		tasklet_hi_schedule(&engine->execlists.tasklet);
225
226unlock:
227	rcu_read_unlock();
228}
229
230static void __i915_schedule(struct i915_sched_node *node,
231			    const struct i915_sched_attr *attr)
232{
233	struct intel_engine_cs *engine;
234	struct i915_dependency *dep, *p;
235	struct i915_dependency stack;
236	const int prio = attr->priority;
237	struct sched_cache cache;
238	LIST_HEAD(dfs);
239
240	/* Needed in order to use the temporary link inside i915_dependency */
241	lockdep_assert_held(&schedule_lock);
242	GEM_BUG_ON(prio == I915_PRIORITY_INVALID);
243
244	if (prio <= READ_ONCE(node->attr.priority))
245		return;
246
247	if (node_signaled(node))
248		return;
249
250	stack.signaler = node;
251	list_add(&stack.dfs_link, &dfs);
252
253	/*
254	 * Recursively bump all dependent priorities to match the new request.
255	 *
256	 * A naive approach would be to use recursion:
257	 * static void update_priorities(struct i915_sched_node *node, prio) {
258	 *	list_for_each_entry(dep, &node->signalers_list, signal_link)
259	 *		update_priorities(dep->signal, prio)
260	 *	queue_request(node);
261	 * }
262	 * but that may have unlimited recursion depth and so runs a very
263	 * real risk of overunning the kernel stack. Instead, we build
264	 * a flat list of all dependencies starting with the current request.
265	 * As we walk the list of dependencies, we add all of its dependencies
266	 * to the end of the list (this may include an already visited
267	 * request) and continue to walk onwards onto the new dependencies. The
268	 * end result is a topological list of requests in reverse order, the
269	 * last element in the list is the request we must execute first.
270	 */
271	list_for_each_entry(dep, &dfs, dfs_link) {
272		struct i915_sched_node *node = dep->signaler;
273
274		/* If we are already flying, we know we have no signalers */
275		if (node_started(node))
276			continue;
277
278		/*
279		 * Within an engine, there can be no cycle, but we may
280		 * refer to the same dependency chain multiple times
281		 * (redundant dependencies are not eliminated) and across
282		 * engines.
283		 */
284		list_for_each_entry(p, &node->signalers_list, signal_link) {
285			GEM_BUG_ON(p == dep); /* no cycles! */
286
287			if (node_signaled(p->signaler))
288				continue;
289
290			if (prio > READ_ONCE(p->signaler->attr.priority))
291				list_move_tail(&p->dfs_link, &dfs);
292		}
293	}
294
295	/*
296	 * If we didn't need to bump any existing priorities, and we haven't
297	 * yet submitted this request (i.e. there is no potential race with
298	 * execlists_submit_request()), we can set our own priority and skip
299	 * acquiring the engine locks.
300	 */
301	if (node->attr.priority == I915_PRIORITY_INVALID) {
302		GEM_BUG_ON(!list_empty(&node->link));
303		node->attr = *attr;
304
305		if (stack.dfs_link.next == stack.dfs_link.prev)
306			return;
307
308		__list_del_entry(&stack.dfs_link);
309	}
310
311	memset(&cache, 0, sizeof(cache));
312	engine = node_to_request(node)->engine;
313	spin_lock(&engine->active.lock);
314
315	/* Fifo and depth-first replacement ensure our deps execute before us */
316	engine = sched_lock_engine(node, engine, &cache);
317	list_for_each_entry_safe_reverse(dep, p, &dfs, dfs_link) {
318		INIT_LIST_HEAD(&dep->dfs_link);
319
320		node = dep->signaler;
321		engine = sched_lock_engine(node, engine, &cache);
322		lockdep_assert_held(&engine->active.lock);
323
324		/* Recheck after acquiring the engine->timeline.lock */
325		if (prio <= node->attr.priority || node_signaled(node))
326			continue;
327
328		GEM_BUG_ON(node_to_request(node)->engine != engine);
329
330		node->attr.priority = prio;
331
332		if (list_empty(&node->link)) {
333			/*
334			 * If the request is not in the priolist queue because
335			 * it is not yet runnable, then it doesn't contribute
336			 * to our preemption decisions. On the other hand,
337			 * if the request is on the HW, it too is not in the
338			 * queue; but in that case we may still need to reorder
339			 * the inflight requests.
340			 */
341			continue;
342		}
343
344		if (!intel_engine_is_virtual(engine) &&
345		    !i915_request_is_active(node_to_request(node))) {
346			if (!cache.priolist)
347				cache.priolist =
348					i915_sched_lookup_priolist(engine,
349								   prio);
350			list_move_tail(&node->link, cache.priolist);
351		}
352
353		/* Defer (tasklet) submission until after all of our updates. */
354		kick_submission(engine, node_to_request(node), prio);
355	}
356
357	spin_unlock(&engine->active.lock);
358}
359
360void i915_schedule(struct i915_request *rq, const struct i915_sched_attr *attr)
361{
362	spin_lock_irq(&schedule_lock);
363	__i915_schedule(&rq->sched, attr);
364	spin_unlock_irq(&schedule_lock);
365}
366
367static void __bump_priority(struct i915_sched_node *node, unsigned int bump)
368{
369	struct i915_sched_attr attr = node->attr;
370
371	attr.priority |= bump;
372	__i915_schedule(node, &attr);
373}
374
375void i915_schedule_bump_priority(struct i915_request *rq, unsigned int bump)
376{
377	unsigned long flags;
378
379	GEM_BUG_ON(bump & ~I915_PRIORITY_MASK);
380	if (READ_ONCE(rq->sched.attr.priority) & bump)
381		return;
382
383	spin_lock_irqsave(&schedule_lock, flags);
384	__bump_priority(&rq->sched, bump);
385	spin_unlock_irqrestore(&schedule_lock, flags);
386}
387
388void i915_sched_node_init(struct i915_sched_node *node)
389{
390	INIT_LIST_HEAD(&node->signalers_list);
391	INIT_LIST_HEAD(&node->waiters_list);
392	INIT_LIST_HEAD(&node->link);
393	node->attr.priority = I915_PRIORITY_INVALID;
394	node->semaphores = 0;
395	node->flags = 0;
396}
397
398static struct i915_dependency *
399i915_dependency_alloc(void)
400{
401	return kmem_cache_alloc(global.slab_dependencies, GFP_KERNEL);
402}
403
404static void
405i915_dependency_free(struct i915_dependency *dep)
406{
407	kmem_cache_free(global.slab_dependencies, dep);
408}
409
410bool __i915_sched_node_add_dependency(struct i915_sched_node *node,
411				      struct i915_sched_node *signal,
412				      struct i915_dependency *dep,
413				      unsigned long flags)
414{
415	bool ret = false;
416
417	spin_lock_irq(&schedule_lock);
418
419	if (!node_signaled(signal)) {
420		INIT_LIST_HEAD(&dep->dfs_link);
421		list_add(&dep->wait_link, &signal->waiters_list);
422		list_add(&dep->signal_link, &node->signalers_list);
423		dep->signaler = signal;
424		dep->waiter = node;
425		dep->flags = flags;
426
427		/* Keep track of whether anyone on this chain has a semaphore */
428		if (signal->flags & I915_SCHED_HAS_SEMAPHORE_CHAIN &&
429		    !node_started(signal))
430			node->flags |= I915_SCHED_HAS_SEMAPHORE_CHAIN;
431
432		/*
433		 * As we do not allow WAIT to preempt inflight requests,
434		 * once we have executed a request, along with triggering
435		 * any execution callbacks, we must preserve its ordering
436		 * within the non-preemptible FIFO.
437		 */
438		BUILD_BUG_ON(__NO_PREEMPTION & ~I915_PRIORITY_MASK);
439		if (flags & I915_DEPENDENCY_EXTERNAL)
440			__bump_priority(signal, __NO_PREEMPTION);
441
442		ret = true;
443	}
444
445	spin_unlock_irq(&schedule_lock);
446
447	return ret;
448}
449
450int i915_sched_node_add_dependency(struct i915_sched_node *node,
451				   struct i915_sched_node *signal)
452{
453	struct i915_dependency *dep;
454
455	dep = i915_dependency_alloc();
456	if (!dep)
457		return -ENOMEM;
458
459	if (!__i915_sched_node_add_dependency(node, signal, dep,
460					      I915_DEPENDENCY_EXTERNAL |
461					      I915_DEPENDENCY_ALLOC))
462		i915_dependency_free(dep);
463
464	return 0;
465}
466
467void i915_sched_node_fini(struct i915_sched_node *node)
468{
469	struct i915_dependency *dep, *tmp;
470
471	spin_lock_irq(&schedule_lock);
472
473	/*
474	 * Everyone we depended upon (the fences we wait to be signaled)
475	 * should retire before us and remove themselves from our list.
476	 * However, retirement is run independently on each timeline and
477	 * so we may be called out-of-order.
478	 */
479	list_for_each_entry_safe(dep, tmp, &node->signalers_list, signal_link) {
480		GEM_BUG_ON(!node_signaled(dep->signaler));
481		GEM_BUG_ON(!list_empty(&dep->dfs_link));
482
483		list_del(&dep->wait_link);
484		if (dep->flags & I915_DEPENDENCY_ALLOC)
485			i915_dependency_free(dep);
486	}
487
488	/* Remove ourselves from everyone who depends upon us */
489	list_for_each_entry_safe(dep, tmp, &node->waiters_list, wait_link) {
490		GEM_BUG_ON(dep->signaler != node);
491		GEM_BUG_ON(!list_empty(&dep->dfs_link));
492
493		list_del(&dep->signal_link);
494		if (dep->flags & I915_DEPENDENCY_ALLOC)
495			i915_dependency_free(dep);
496	}
497
498	spin_unlock_irq(&schedule_lock);
499}
500
501static void i915_global_scheduler_shrink(void)
502{
503	kmem_cache_shrink(global.slab_dependencies);
504	kmem_cache_shrink(global.slab_priorities);
505}
506
507static void i915_global_scheduler_exit(void)
508{
509	kmem_cache_destroy(global.slab_dependencies);
510	kmem_cache_destroy(global.slab_priorities);
511}
512
513static struct i915_global_scheduler global = { {
514	.shrink = i915_global_scheduler_shrink,
515	.exit = i915_global_scheduler_exit,
516} };
517
518int __init i915_global_scheduler_init(void)
519{
520	global.slab_dependencies = KMEM_CACHE(i915_dependency,
521					      SLAB_HWCACHE_ALIGN);
522	if (!global.slab_dependencies)
523		return -ENOMEM;
524
525	global.slab_priorities = KMEM_CACHE(i915_priolist,
526					    SLAB_HWCACHE_ALIGN);
527	if (!global.slab_priorities)
528		goto err_priorities;
529
530	i915_global_register(&global.base);
531	return 0;
532
533err_priorities:
534	kmem_cache_destroy(global.slab_priorities);
535	return -ENOMEM;
536}