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 
14 static struct i915_global_scheduler {
15 	struct i915_global base;
16 	struct kmem_cache *slab_dependencies;
17 	struct kmem_cache *slab_priorities;
18 } global;
19 
20 static DEFINE_SPINLOCK(schedule_lock);
21 
22 static const struct i915_request *
23 node_to_request(const struct i915_sched_node *node)
24 {
25 	return container_of(node, const struct i915_request, sched);
26 }
27 
28 static inline bool node_started(const struct i915_sched_node *node)
29 {
30 	return i915_request_started(node_to_request(node));
31 }
32 
33 static inline bool node_signaled(const struct i915_sched_node *node)
34 {
35 	return i915_request_completed(node_to_request(node));
36 }
37 
38 static inline struct i915_priolist *to_priolist(struct rb_node *rb)
39 {
40 	return rb_entry(rb, struct i915_priolist, node);
41 }
42 
43 static 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 
71 struct list_head *
72 i915_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 
89 find_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 
134 out:
135 	p->used |= BIT(idx);
136 	return &p->requests[idx];
137 }
138 
139 void __i915_priolist_free(struct i915_priolist *p)
140 {
141 	kmem_cache_free(global.slab_priorities, p);
142 }
143 
144 struct sched_cache {
145 	struct list_head *priolist;
146 };
147 
148 static struct intel_engine_cs *
149 sched_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 
175 static inline int rq_prio(const struct i915_request *rq)
176 {
177 	return rq->sched.attr.priority | __NO_PREEMPTION;
178 }
179 
180 static 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 
192 static 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->hw_context == rq->hw_context)
217 		goto unlock;
218 
219 	engine->execlists.queue_priority_hint = prio;
220 	if (need_preempt(prio, rq_prio(inflight)))
221 		tasklet_hi_schedule(&engine->execlists.tasklet);
222 
223 unlock:
224 	rcu_read_unlock();
225 }
226 
227 static void __i915_schedule(struct i915_sched_node *node,
228 			    const struct i915_sched_attr *attr)
229 {
230 	struct intel_engine_cs *engine;
231 	struct i915_dependency *dep, *p;
232 	struct i915_dependency stack;
233 	const int prio = attr->priority;
234 	struct sched_cache cache;
235 	LIST_HEAD(dfs);
236 
237 	/* Needed in order to use the temporary link inside i915_dependency */
238 	lockdep_assert_held(&schedule_lock);
239 	GEM_BUG_ON(prio == I915_PRIORITY_INVALID);
240 
241 	if (prio <= READ_ONCE(node->attr.priority))
242 		return;
243 
244 	if (node_signaled(node))
245 		return;
246 
247 	stack.signaler = node;
248 	list_add(&stack.dfs_link, &dfs);
249 
250 	/*
251 	 * Recursively bump all dependent priorities to match the new request.
252 	 *
253 	 * A naive approach would be to use recursion:
254 	 * static void update_priorities(struct i915_sched_node *node, prio) {
255 	 *	list_for_each_entry(dep, &node->signalers_list, signal_link)
256 	 *		update_priorities(dep->signal, prio)
257 	 *	queue_request(node);
258 	 * }
259 	 * but that may have unlimited recursion depth and so runs a very
260 	 * real risk of overunning the kernel stack. Instead, we build
261 	 * a flat list of all dependencies starting with the current request.
262 	 * As we walk the list of dependencies, we add all of its dependencies
263 	 * to the end of the list (this may include an already visited
264 	 * request) and continue to walk onwards onto the new dependencies. The
265 	 * end result is a topological list of requests in reverse order, the
266 	 * last element in the list is the request we must execute first.
267 	 */
268 	list_for_each_entry(dep, &dfs, dfs_link) {
269 		struct i915_sched_node *node = dep->signaler;
270 
271 		/* If we are already flying, we know we have no signalers */
272 		if (node_started(node))
273 			continue;
274 
275 		/*
276 		 * Within an engine, there can be no cycle, but we may
277 		 * refer to the same dependency chain multiple times
278 		 * (redundant dependencies are not eliminated) and across
279 		 * engines.
280 		 */
281 		list_for_each_entry(p, &node->signalers_list, signal_link) {
282 			GEM_BUG_ON(p == dep); /* no cycles! */
283 
284 			if (node_signaled(p->signaler))
285 				continue;
286 
287 			if (prio > READ_ONCE(p->signaler->attr.priority))
288 				list_move_tail(&p->dfs_link, &dfs);
289 		}
290 	}
291 
292 	/*
293 	 * If we didn't need to bump any existing priorities, and we haven't
294 	 * yet submitted this request (i.e. there is no potential race with
295 	 * execlists_submit_request()), we can set our own priority and skip
296 	 * acquiring the engine locks.
297 	 */
298 	if (node->attr.priority == I915_PRIORITY_INVALID) {
299 		GEM_BUG_ON(!list_empty(&node->link));
300 		node->attr = *attr;
301 
302 		if (stack.dfs_link.next == stack.dfs_link.prev)
303 			return;
304 
305 		__list_del_entry(&stack.dfs_link);
306 	}
307 
308 	memset(&cache, 0, sizeof(cache));
309 	engine = node_to_request(node)->engine;
310 	spin_lock(&engine->active.lock);
311 
312 	/* Fifo and depth-first replacement ensure our deps execute before us */
313 	engine = sched_lock_engine(node, engine, &cache);
314 	list_for_each_entry_safe_reverse(dep, p, &dfs, dfs_link) {
315 		INIT_LIST_HEAD(&dep->dfs_link);
316 
317 		node = dep->signaler;
318 		engine = sched_lock_engine(node, engine, &cache);
319 		lockdep_assert_held(&engine->active.lock);
320 
321 		/* Recheck after acquiring the engine->timeline.lock */
322 		if (prio <= node->attr.priority || node_signaled(node))
323 			continue;
324 
325 		GEM_BUG_ON(node_to_request(node)->engine != engine);
326 
327 		node->attr.priority = prio;
328 
329 		if (list_empty(&node->link)) {
330 			/*
331 			 * If the request is not in the priolist queue because
332 			 * it is not yet runnable, then it doesn't contribute
333 			 * to our preemption decisions. On the other hand,
334 			 * if the request is on the HW, it too is not in the
335 			 * queue; but in that case we may still need to reorder
336 			 * the inflight requests.
337 			 */
338 			continue;
339 		}
340 
341 		if (!intel_engine_is_virtual(engine) &&
342 		    !i915_request_is_active(node_to_request(node))) {
343 			if (!cache.priolist)
344 				cache.priolist =
345 					i915_sched_lookup_priolist(engine,
346 								   prio);
347 			list_move_tail(&node->link, cache.priolist);
348 		}
349 
350 		/* Defer (tasklet) submission until after all of our updates. */
351 		kick_submission(engine, node_to_request(node), prio);
352 	}
353 
354 	spin_unlock(&engine->active.lock);
355 }
356 
357 void i915_schedule(struct i915_request *rq, const struct i915_sched_attr *attr)
358 {
359 	spin_lock_irq(&schedule_lock);
360 	__i915_schedule(&rq->sched, attr);
361 	spin_unlock_irq(&schedule_lock);
362 }
363 
364 static void __bump_priority(struct i915_sched_node *node, unsigned int bump)
365 {
366 	struct i915_sched_attr attr = node->attr;
367 
368 	attr.priority |= bump;
369 	__i915_schedule(node, &attr);
370 }
371 
372 void i915_schedule_bump_priority(struct i915_request *rq, unsigned int bump)
373 {
374 	unsigned long flags;
375 
376 	GEM_BUG_ON(bump & ~I915_PRIORITY_MASK);
377 	if (READ_ONCE(rq->sched.attr.priority) & bump)
378 		return;
379 
380 	spin_lock_irqsave(&schedule_lock, flags);
381 	__bump_priority(&rq->sched, bump);
382 	spin_unlock_irqrestore(&schedule_lock, flags);
383 }
384 
385 void i915_sched_node_init(struct i915_sched_node *node)
386 {
387 	INIT_LIST_HEAD(&node->signalers_list);
388 	INIT_LIST_HEAD(&node->waiters_list);
389 	INIT_LIST_HEAD(&node->link);
390 	node->attr.priority = I915_PRIORITY_INVALID;
391 	node->semaphores = 0;
392 	node->flags = 0;
393 }
394 
395 static struct i915_dependency *
396 i915_dependency_alloc(void)
397 {
398 	return kmem_cache_alloc(global.slab_dependencies, GFP_KERNEL);
399 }
400 
401 static void
402 i915_dependency_free(struct i915_dependency *dep)
403 {
404 	kmem_cache_free(global.slab_dependencies, dep);
405 }
406 
407 bool __i915_sched_node_add_dependency(struct i915_sched_node *node,
408 				      struct i915_sched_node *signal,
409 				      struct i915_dependency *dep,
410 				      unsigned long flags)
411 {
412 	bool ret = false;
413 
414 	spin_lock_irq(&schedule_lock);
415 
416 	if (!node_signaled(signal)) {
417 		INIT_LIST_HEAD(&dep->dfs_link);
418 		list_add(&dep->wait_link, &signal->waiters_list);
419 		list_add(&dep->signal_link, &node->signalers_list);
420 		dep->signaler = signal;
421 		dep->waiter = node;
422 		dep->flags = flags;
423 
424 		/* Keep track of whether anyone on this chain has a semaphore */
425 		if (signal->flags & I915_SCHED_HAS_SEMAPHORE_CHAIN &&
426 		    !node_started(signal))
427 			node->flags |= I915_SCHED_HAS_SEMAPHORE_CHAIN;
428 
429 		/*
430 		 * As we do not allow WAIT to preempt inflight requests,
431 		 * once we have executed a request, along with triggering
432 		 * any execution callbacks, we must preserve its ordering
433 		 * within the non-preemptible FIFO.
434 		 */
435 		BUILD_BUG_ON(__NO_PREEMPTION & ~I915_PRIORITY_MASK);
436 		if (flags & I915_DEPENDENCY_EXTERNAL)
437 			__bump_priority(signal, __NO_PREEMPTION);
438 
439 		ret = true;
440 	}
441 
442 	spin_unlock_irq(&schedule_lock);
443 
444 	return ret;
445 }
446 
447 int i915_sched_node_add_dependency(struct i915_sched_node *node,
448 				   struct i915_sched_node *signal)
449 {
450 	struct i915_dependency *dep;
451 
452 	dep = i915_dependency_alloc();
453 	if (!dep)
454 		return -ENOMEM;
455 
456 	if (!__i915_sched_node_add_dependency(node, signal, dep,
457 					      I915_DEPENDENCY_EXTERNAL |
458 					      I915_DEPENDENCY_ALLOC))
459 		i915_dependency_free(dep);
460 
461 	return 0;
462 }
463 
464 void i915_sched_node_fini(struct i915_sched_node *node)
465 {
466 	struct i915_dependency *dep, *tmp;
467 
468 	spin_lock_irq(&schedule_lock);
469 
470 	/*
471 	 * Everyone we depended upon (the fences we wait to be signaled)
472 	 * should retire before us and remove themselves from our list.
473 	 * However, retirement is run independently on each timeline and
474 	 * so we may be called out-of-order.
475 	 */
476 	list_for_each_entry_safe(dep, tmp, &node->signalers_list, signal_link) {
477 		GEM_BUG_ON(!list_empty(&dep->dfs_link));
478 
479 		list_del(&dep->wait_link);
480 		if (dep->flags & I915_DEPENDENCY_ALLOC)
481 			i915_dependency_free(dep);
482 	}
483 
484 	/* Remove ourselves from everyone who depends upon us */
485 	list_for_each_entry_safe(dep, tmp, &node->waiters_list, wait_link) {
486 		GEM_BUG_ON(dep->signaler != node);
487 		GEM_BUG_ON(!list_empty(&dep->dfs_link));
488 
489 		list_del(&dep->signal_link);
490 		if (dep->flags & I915_DEPENDENCY_ALLOC)
491 			i915_dependency_free(dep);
492 	}
493 
494 	spin_unlock_irq(&schedule_lock);
495 }
496 
497 static void i915_global_scheduler_shrink(void)
498 {
499 	kmem_cache_shrink(global.slab_dependencies);
500 	kmem_cache_shrink(global.slab_priorities);
501 }
502 
503 static void i915_global_scheduler_exit(void)
504 {
505 	kmem_cache_destroy(global.slab_dependencies);
506 	kmem_cache_destroy(global.slab_priorities);
507 }
508 
509 static struct i915_global_scheduler global = { {
510 	.shrink = i915_global_scheduler_shrink,
511 	.exit = i915_global_scheduler_exit,
512 } };
513 
514 int __init i915_global_scheduler_init(void)
515 {
516 	global.slab_dependencies = KMEM_CACHE(i915_dependency,
517 					      SLAB_HWCACHE_ALIGN);
518 	if (!global.slab_dependencies)
519 		return -ENOMEM;
520 
521 	global.slab_priorities = KMEM_CACHE(i915_priolist,
522 					    SLAB_HWCACHE_ALIGN);
523 	if (!global.slab_priorities)
524 		goto err_priorities;
525 
526 	i915_global_register(&global.base);
527 	return 0;
528 
529 err_priorities:
530 	kmem_cache_destroy(global.slab_priorities);
531 	return -ENOMEM;
532 }
533