xref: /linux/drivers/gpu/drm/i915/i915_scheduler.c (revision 762f99f4f3cb41a775b5157dd761217beba65873)
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_request.h"
11 #include "i915_scheduler.h"
12 
13 static struct kmem_cache *slab_dependencies;
14 static struct kmem_cache *slab_priorities;
15 
16 static DEFINE_SPINLOCK(schedule_lock);
17 
18 static const struct i915_request *
node_to_request(const struct i915_sched_node * node)19 node_to_request(const struct i915_sched_node *node)
20 {
21 	return container_of(node, const struct i915_request, sched);
22 }
23 
node_started(const struct i915_sched_node * node)24 static inline bool node_started(const struct i915_sched_node *node)
25 {
26 	return i915_request_started(node_to_request(node));
27 }
28 
node_signaled(const struct i915_sched_node * node)29 static inline bool node_signaled(const struct i915_sched_node *node)
30 {
31 	return i915_request_completed(node_to_request(node));
32 }
33 
to_priolist(struct rb_node * rb)34 static inline struct i915_priolist *to_priolist(struct rb_node *rb)
35 {
36 	return rb_entry(rb, struct i915_priolist, node);
37 }
38 
assert_priolists(struct i915_sched_engine * const sched_engine)39 static void assert_priolists(struct i915_sched_engine * const sched_engine)
40 {
41 	struct rb_node *rb;
42 	long last_prio;
43 
44 	if (!IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM))
45 		return;
46 
47 	GEM_BUG_ON(rb_first_cached(&sched_engine->queue) !=
48 		   rb_first(&sched_engine->queue.rb_root));
49 
50 	last_prio = INT_MAX;
51 	for (rb = rb_first_cached(&sched_engine->queue); rb; rb = rb_next(rb)) {
52 		const struct i915_priolist *p = to_priolist(rb);
53 
54 		GEM_BUG_ON(p->priority > last_prio);
55 		last_prio = p->priority;
56 	}
57 }
58 
59 struct list_head *
i915_sched_lookup_priolist(struct i915_sched_engine * sched_engine,int prio)60 i915_sched_lookup_priolist(struct i915_sched_engine *sched_engine, int prio)
61 {
62 	struct i915_priolist *p;
63 	struct rb_node **parent, *rb;
64 	bool first = true;
65 
66 	lockdep_assert_held(&sched_engine->lock);
67 	assert_priolists(sched_engine);
68 
69 	if (unlikely(sched_engine->no_priolist))
70 		prio = I915_PRIORITY_NORMAL;
71 
72 find_priolist:
73 	/* most positive priority is scheduled first, equal priorities fifo */
74 	rb = NULL;
75 	parent = &sched_engine->queue.rb_root.rb_node;
76 	while (*parent) {
77 		rb = *parent;
78 		p = to_priolist(rb);
79 		if (prio > p->priority) {
80 			parent = &rb->rb_left;
81 		} else if (prio < p->priority) {
82 			parent = &rb->rb_right;
83 			first = false;
84 		} else {
85 			return &p->requests;
86 		}
87 	}
88 
89 	if (prio == I915_PRIORITY_NORMAL) {
90 		p = &sched_engine->default_priolist;
91 	} else {
92 		p = kmem_cache_alloc(slab_priorities, GFP_ATOMIC);
93 		/* Convert an allocation failure to a priority bump */
94 		if (unlikely(!p)) {
95 			prio = I915_PRIORITY_NORMAL; /* recurses just once */
96 
97 			/* To maintain ordering with all rendering, after an
98 			 * allocation failure we have to disable all scheduling.
99 			 * Requests will then be executed in fifo, and schedule
100 			 * will ensure that dependencies are emitted in fifo.
101 			 * There will be still some reordering with existing
102 			 * requests, so if userspace lied about their
103 			 * dependencies that reordering may be visible.
104 			 */
105 			sched_engine->no_priolist = true;
106 			goto find_priolist;
107 		}
108 	}
109 
110 	p->priority = prio;
111 	INIT_LIST_HEAD(&p->requests);
112 
113 	rb_link_node(&p->node, rb, parent);
114 	rb_insert_color_cached(&p->node, &sched_engine->queue, first);
115 
116 	return &p->requests;
117 }
118 
__i915_priolist_free(struct i915_priolist * p)119 void __i915_priolist_free(struct i915_priolist *p)
120 {
121 	kmem_cache_free(slab_priorities, p);
122 }
123 
124 struct sched_cache {
125 	struct list_head *priolist;
126 };
127 
128 static struct i915_sched_engine *
lock_sched_engine(struct i915_sched_node * node,struct i915_sched_engine * locked,struct sched_cache * cache)129 lock_sched_engine(struct i915_sched_node *node,
130 		  struct i915_sched_engine *locked,
131 		  struct sched_cache *cache)
132 {
133 	const struct i915_request *rq = node_to_request(node);
134 	struct i915_sched_engine *sched_engine;
135 
136 	GEM_BUG_ON(!locked);
137 
138 	/*
139 	 * Virtual engines complicate acquiring the engine timeline lock,
140 	 * as their rq->engine pointer is not stable until under that
141 	 * engine lock. The simple ploy we use is to take the lock then
142 	 * check that the rq still belongs to the newly locked engine.
143 	 */
144 	while (locked != (sched_engine = READ_ONCE(rq->engine)->sched_engine)) {
145 		spin_unlock(&locked->lock);
146 		memset(cache, 0, sizeof(*cache));
147 		spin_lock(&sched_engine->lock);
148 		locked = sched_engine;
149 	}
150 
151 	GEM_BUG_ON(locked != sched_engine);
152 	return locked;
153 }
154 
__i915_schedule(struct i915_sched_node * node,const struct i915_sched_attr * attr)155 static void __i915_schedule(struct i915_sched_node *node,
156 			    const struct i915_sched_attr *attr)
157 {
158 	const int prio = max(attr->priority, node->attr.priority);
159 	struct i915_sched_engine *sched_engine;
160 	struct i915_dependency *dep, *p;
161 	struct i915_dependency stack;
162 	struct sched_cache cache;
163 	LIST_HEAD(dfs);
164 
165 	/* Needed in order to use the temporary link inside i915_dependency */
166 	lockdep_assert_held(&schedule_lock);
167 	GEM_BUG_ON(prio == I915_PRIORITY_INVALID);
168 
169 	if (node_signaled(node))
170 		return;
171 
172 	stack.signaler = node;
173 	list_add(&stack.dfs_link, &dfs);
174 
175 	/*
176 	 * Recursively bump all dependent priorities to match the new request.
177 	 *
178 	 * A naive approach would be to use recursion:
179 	 * static void update_priorities(struct i915_sched_node *node, prio) {
180 	 *	list_for_each_entry(dep, &node->signalers_list, signal_link)
181 	 *		update_priorities(dep->signal, prio)
182 	 *	queue_request(node);
183 	 * }
184 	 * but that may have unlimited recursion depth and so runs a very
185 	 * real risk of overunning the kernel stack. Instead, we build
186 	 * a flat list of all dependencies starting with the current request.
187 	 * As we walk the list of dependencies, we add all of its dependencies
188 	 * to the end of the list (this may include an already visited
189 	 * request) and continue to walk onwards onto the new dependencies. The
190 	 * end result is a topological list of requests in reverse order, the
191 	 * last element in the list is the request we must execute first.
192 	 */
193 	list_for_each_entry(dep, &dfs, dfs_link) {
194 		struct i915_sched_node *node = dep->signaler;
195 
196 		/* If we are already flying, we know we have no signalers */
197 		if (node_started(node))
198 			continue;
199 
200 		/*
201 		 * Within an engine, there can be no cycle, but we may
202 		 * refer to the same dependency chain multiple times
203 		 * (redundant dependencies are not eliminated) and across
204 		 * engines.
205 		 */
206 		list_for_each_entry(p, &node->signalers_list, signal_link) {
207 			GEM_BUG_ON(p == dep); /* no cycles! */
208 
209 			if (node_signaled(p->signaler))
210 				continue;
211 
212 			if (prio > READ_ONCE(p->signaler->attr.priority))
213 				list_move_tail(&p->dfs_link, &dfs);
214 		}
215 	}
216 
217 	/*
218 	 * If we didn't need to bump any existing priorities, and we haven't
219 	 * yet submitted this request (i.e. there is no potential race with
220 	 * execlists_submit_request()), we can set our own priority and skip
221 	 * acquiring the engine locks.
222 	 */
223 	if (node->attr.priority == I915_PRIORITY_INVALID) {
224 		GEM_BUG_ON(!list_empty(&node->link));
225 		node->attr = *attr;
226 
227 		if (stack.dfs_link.next == stack.dfs_link.prev)
228 			return;
229 
230 		__list_del_entry(&stack.dfs_link);
231 	}
232 
233 	memset(&cache, 0, sizeof(cache));
234 	sched_engine = node_to_request(node)->engine->sched_engine;
235 	spin_lock(&sched_engine->lock);
236 
237 	/* Fifo and depth-first replacement ensure our deps execute before us */
238 	sched_engine = lock_sched_engine(node, sched_engine, &cache);
239 	list_for_each_entry_safe_reverse(dep, p, &dfs, dfs_link) {
240 		struct i915_request *from = container_of(dep->signaler,
241 							 struct i915_request,
242 							 sched);
243 		INIT_LIST_HEAD(&dep->dfs_link);
244 
245 		node = dep->signaler;
246 		sched_engine = lock_sched_engine(node, sched_engine, &cache);
247 		lockdep_assert_held(&sched_engine->lock);
248 
249 		/* Recheck after acquiring the engine->timeline.lock */
250 		if (prio <= node->attr.priority || node_signaled(node))
251 			continue;
252 
253 		GEM_BUG_ON(node_to_request(node)->engine->sched_engine !=
254 			   sched_engine);
255 
256 		/* Must be called before changing the nodes priority */
257 		if (sched_engine->bump_inflight_request_prio)
258 			sched_engine->bump_inflight_request_prio(from, prio);
259 
260 		WRITE_ONCE(node->attr.priority, prio);
261 
262 		/*
263 		 * Once the request is ready, it will be placed into the
264 		 * priority lists and then onto the HW runlist. Before the
265 		 * request is ready, it does not contribute to our preemption
266 		 * decisions and we can safely ignore it, as it will, and
267 		 * any preemption required, be dealt with upon submission.
268 		 * See engine->submit_request()
269 		 */
270 		if (list_empty(&node->link))
271 			continue;
272 
273 		if (i915_request_in_priority_queue(node_to_request(node))) {
274 			if (!cache.priolist)
275 				cache.priolist =
276 					i915_sched_lookup_priolist(sched_engine,
277 								   prio);
278 			list_move_tail(&node->link, cache.priolist);
279 		}
280 
281 		/* Defer (tasklet) submission until after all of our updates. */
282 		if (sched_engine->kick_backend)
283 			sched_engine->kick_backend(node_to_request(node), prio);
284 	}
285 
286 	spin_unlock(&sched_engine->lock);
287 }
288 
i915_schedule(struct i915_request * rq,const struct i915_sched_attr * attr)289 void i915_schedule(struct i915_request *rq, const struct i915_sched_attr *attr)
290 {
291 	spin_lock_irq(&schedule_lock);
292 	__i915_schedule(&rq->sched, attr);
293 	spin_unlock_irq(&schedule_lock);
294 }
295 
i915_sched_node_init(struct i915_sched_node * node)296 void i915_sched_node_init(struct i915_sched_node *node)
297 {
298 	INIT_LIST_HEAD(&node->signalers_list);
299 	INIT_LIST_HEAD(&node->waiters_list);
300 	INIT_LIST_HEAD(&node->link);
301 
302 	i915_sched_node_reinit(node);
303 }
304 
i915_sched_node_reinit(struct i915_sched_node * node)305 void i915_sched_node_reinit(struct i915_sched_node *node)
306 {
307 	node->attr.priority = I915_PRIORITY_INVALID;
308 	node->semaphores = 0;
309 	node->flags = 0;
310 
311 	GEM_BUG_ON(!list_empty(&node->signalers_list));
312 	GEM_BUG_ON(!list_empty(&node->waiters_list));
313 	GEM_BUG_ON(!list_empty(&node->link));
314 }
315 
316 static struct i915_dependency *
i915_dependency_alloc(void)317 i915_dependency_alloc(void)
318 {
319 	return kmem_cache_alloc(slab_dependencies, GFP_KERNEL);
320 }
321 
322 static void
i915_dependency_free(struct i915_dependency * dep)323 i915_dependency_free(struct i915_dependency *dep)
324 {
325 	kmem_cache_free(slab_dependencies, dep);
326 }
327 
__i915_sched_node_add_dependency(struct i915_sched_node * node,struct i915_sched_node * signal,struct i915_dependency * dep,unsigned long flags)328 bool __i915_sched_node_add_dependency(struct i915_sched_node *node,
329 				      struct i915_sched_node *signal,
330 				      struct i915_dependency *dep,
331 				      unsigned long flags)
332 {
333 	bool ret = false;
334 
335 	spin_lock_irq(&schedule_lock);
336 
337 	if (!node_signaled(signal)) {
338 		INIT_LIST_HEAD(&dep->dfs_link);
339 		dep->signaler = signal;
340 		dep->waiter = node;
341 		dep->flags = flags;
342 
343 		/* All set, now publish. Beware the lockless walkers. */
344 		list_add_rcu(&dep->signal_link, &node->signalers_list);
345 		list_add_rcu(&dep->wait_link, &signal->waiters_list);
346 
347 		/* Propagate the chains */
348 		node->flags |= signal->flags;
349 		ret = true;
350 	}
351 
352 	spin_unlock_irq(&schedule_lock);
353 
354 	return ret;
355 }
356 
i915_sched_node_add_dependency(struct i915_sched_node * node,struct i915_sched_node * signal,unsigned long flags)357 int i915_sched_node_add_dependency(struct i915_sched_node *node,
358 				   struct i915_sched_node *signal,
359 				   unsigned long flags)
360 {
361 	struct i915_dependency *dep;
362 
363 	dep = i915_dependency_alloc();
364 	if (!dep)
365 		return -ENOMEM;
366 
367 	if (!__i915_sched_node_add_dependency(node, signal, dep,
368 					      flags | I915_DEPENDENCY_ALLOC))
369 		i915_dependency_free(dep);
370 
371 	return 0;
372 }
373 
i915_sched_node_fini(struct i915_sched_node * node)374 void i915_sched_node_fini(struct i915_sched_node *node)
375 {
376 	struct i915_dependency *dep, *tmp;
377 
378 	spin_lock_irq(&schedule_lock);
379 
380 	/*
381 	 * Everyone we depended upon (the fences we wait to be signaled)
382 	 * should retire before us and remove themselves from our list.
383 	 * However, retirement is run independently on each timeline and
384 	 * so we may be called out-of-order.
385 	 */
386 	list_for_each_entry_safe(dep, tmp, &node->signalers_list, signal_link) {
387 		GEM_BUG_ON(!list_empty(&dep->dfs_link));
388 
389 		list_del_rcu(&dep->wait_link);
390 		if (dep->flags & I915_DEPENDENCY_ALLOC)
391 			i915_dependency_free(dep);
392 	}
393 	INIT_LIST_HEAD(&node->signalers_list);
394 
395 	/* Remove ourselves from everyone who depends upon us */
396 	list_for_each_entry_safe(dep, tmp, &node->waiters_list, wait_link) {
397 		GEM_BUG_ON(dep->signaler != node);
398 		GEM_BUG_ON(!list_empty(&dep->dfs_link));
399 
400 		list_del_rcu(&dep->signal_link);
401 		if (dep->flags & I915_DEPENDENCY_ALLOC)
402 			i915_dependency_free(dep);
403 	}
404 	INIT_LIST_HEAD(&node->waiters_list);
405 
406 	spin_unlock_irq(&schedule_lock);
407 }
408 
i915_request_show_with_schedule(struct drm_printer * m,const struct i915_request * rq,const char * prefix,int indent)409 void i915_request_show_with_schedule(struct drm_printer *m,
410 				     const struct i915_request *rq,
411 				     const char *prefix,
412 				     int indent)
413 {
414 	struct i915_dependency *dep;
415 
416 	i915_request_show(m, rq, prefix, indent);
417 	if (i915_request_completed(rq))
418 		return;
419 
420 	rcu_read_lock();
421 	for_each_signaler(dep, rq) {
422 		const struct i915_request *signaler =
423 			node_to_request(dep->signaler);
424 
425 		/* Dependencies along the same timeline are expected. */
426 		if (signaler->timeline == rq->timeline)
427 			continue;
428 
429 		if (__i915_request_is_complete(signaler))
430 			continue;
431 
432 		i915_request_show(m, signaler, prefix, indent + 2);
433 	}
434 	rcu_read_unlock();
435 }
436 
default_destroy(struct kref * kref)437 static void default_destroy(struct kref *kref)
438 {
439 	struct i915_sched_engine *sched_engine =
440 		container_of(kref, typeof(*sched_engine), ref);
441 
442 	tasklet_kill(&sched_engine->tasklet); /* flush the callback */
443 	kfree(sched_engine);
444 }
445 
default_disabled(struct i915_sched_engine * sched_engine)446 static bool default_disabled(struct i915_sched_engine *sched_engine)
447 {
448 	return false;
449 }
450 
451 struct i915_sched_engine *
i915_sched_engine_create(unsigned int subclass)452 i915_sched_engine_create(unsigned int subclass)
453 {
454 	struct i915_sched_engine *sched_engine;
455 
456 	sched_engine = kzalloc(sizeof(*sched_engine), GFP_KERNEL);
457 	if (!sched_engine)
458 		return NULL;
459 
460 	kref_init(&sched_engine->ref);
461 
462 	sched_engine->queue = RB_ROOT_CACHED;
463 	sched_engine->queue_priority_hint = INT_MIN;
464 	sched_engine->destroy = default_destroy;
465 	sched_engine->disabled = default_disabled;
466 
467 	INIT_LIST_HEAD(&sched_engine->requests);
468 	INIT_LIST_HEAD(&sched_engine->hold);
469 
470 	spin_lock_init(&sched_engine->lock);
471 	lockdep_set_subclass(&sched_engine->lock, subclass);
472 
473 	/*
474 	 * Due to an interesting quirk in lockdep's internal debug tracking,
475 	 * after setting a subclass we must ensure the lock is used. Otherwise,
476 	 * nr_unused_locks is incremented once too often.
477 	 */
478 #ifdef CONFIG_DEBUG_LOCK_ALLOC
479 	local_irq_disable();
480 	lock_map_acquire(&sched_engine->lock.dep_map);
481 	lock_map_release(&sched_engine->lock.dep_map);
482 	local_irq_enable();
483 #endif
484 
485 	return sched_engine;
486 }
487 
i915_scheduler_module_exit(void)488 void i915_scheduler_module_exit(void)
489 {
490 	kmem_cache_destroy(slab_dependencies);
491 	kmem_cache_destroy(slab_priorities);
492 }
493 
i915_scheduler_module_init(void)494 int __init i915_scheduler_module_init(void)
495 {
496 	slab_dependencies = KMEM_CACHE(i915_dependency,
497 					      SLAB_HWCACHE_ALIGN |
498 					      SLAB_TYPESAFE_BY_RCU);
499 	if (!slab_dependencies)
500 		return -ENOMEM;
501 
502 	slab_priorities = KMEM_CACHE(i915_priolist, 0);
503 	if (!slab_priorities)
504 		goto err_priorities;
505 
506 	return 0;
507 
508 err_priorities:
509 	kmem_cache_destroy(slab_priorities);
510 	return -ENOMEM;
511 }
512