xref: /linux/drivers/gpu/drm/i915/i915_scheduler.c (revision 44f57d78)
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->timeline.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 	struct intel_engine_cs *engine = node_to_request(node)->engine;
154 
155 	GEM_BUG_ON(!locked);
156 
157 	if (engine != locked) {
158 		spin_unlock(&locked->timeline.lock);
159 		memset(cache, 0, sizeof(*cache));
160 		spin_lock(&engine->timeline.lock);
161 	}
162 
163 	return engine;
164 }
165 
166 static bool inflight(const struct i915_request *rq,
167 		     const struct intel_engine_cs *engine)
168 {
169 	const struct i915_request *active;
170 
171 	if (!i915_request_is_active(rq))
172 		return false;
173 
174 	active = port_request(engine->execlists.port);
175 	return active->hw_context == rq->hw_context;
176 }
177 
178 static void __i915_schedule(struct i915_sched_node *node,
179 			    const struct i915_sched_attr *attr)
180 {
181 	struct intel_engine_cs *engine;
182 	struct i915_dependency *dep, *p;
183 	struct i915_dependency stack;
184 	const int prio = attr->priority;
185 	struct sched_cache cache;
186 	LIST_HEAD(dfs);
187 
188 	/* Needed in order to use the temporary link inside i915_dependency */
189 	lockdep_assert_held(&schedule_lock);
190 	GEM_BUG_ON(prio == I915_PRIORITY_INVALID);
191 
192 	if (node_signaled(node))
193 		return;
194 
195 	if (prio <= READ_ONCE(node->attr.priority))
196 		return;
197 
198 	stack.signaler = node;
199 	list_add(&stack.dfs_link, &dfs);
200 
201 	/*
202 	 * Recursively bump all dependent priorities to match the new request.
203 	 *
204 	 * A naive approach would be to use recursion:
205 	 * static void update_priorities(struct i915_sched_node *node, prio) {
206 	 *	list_for_each_entry(dep, &node->signalers_list, signal_link)
207 	 *		update_priorities(dep->signal, prio)
208 	 *	queue_request(node);
209 	 * }
210 	 * but that may have unlimited recursion depth and so runs a very
211 	 * real risk of overunning the kernel stack. Instead, we build
212 	 * a flat list of all dependencies starting with the current request.
213 	 * As we walk the list of dependencies, we add all of its dependencies
214 	 * to the end of the list (this may include an already visited
215 	 * request) and continue to walk onwards onto the new dependencies. The
216 	 * end result is a topological list of requests in reverse order, the
217 	 * last element in the list is the request we must execute first.
218 	 */
219 	list_for_each_entry(dep, &dfs, dfs_link) {
220 		struct i915_sched_node *node = dep->signaler;
221 
222 		/* If we are already flying, we know we have no signalers */
223 		if (node_started(node))
224 			continue;
225 
226 		/*
227 		 * Within an engine, there can be no cycle, but we may
228 		 * refer to the same dependency chain multiple times
229 		 * (redundant dependencies are not eliminated) and across
230 		 * engines.
231 		 */
232 		list_for_each_entry(p, &node->signalers_list, signal_link) {
233 			GEM_BUG_ON(p == dep); /* no cycles! */
234 
235 			if (node_signaled(p->signaler))
236 				continue;
237 
238 			if (prio > READ_ONCE(p->signaler->attr.priority))
239 				list_move_tail(&p->dfs_link, &dfs);
240 		}
241 	}
242 
243 	/*
244 	 * If we didn't need to bump any existing priorities, and we haven't
245 	 * yet submitted this request (i.e. there is no potential race with
246 	 * execlists_submit_request()), we can set our own priority and skip
247 	 * acquiring the engine locks.
248 	 */
249 	if (node->attr.priority == I915_PRIORITY_INVALID) {
250 		GEM_BUG_ON(!list_empty(&node->link));
251 		node->attr = *attr;
252 
253 		if (stack.dfs_link.next == stack.dfs_link.prev)
254 			return;
255 
256 		__list_del_entry(&stack.dfs_link);
257 	}
258 
259 	memset(&cache, 0, sizeof(cache));
260 	engine = node_to_request(node)->engine;
261 	spin_lock(&engine->timeline.lock);
262 
263 	/* Fifo and depth-first replacement ensure our deps execute before us */
264 	list_for_each_entry_safe_reverse(dep, p, &dfs, dfs_link) {
265 		INIT_LIST_HEAD(&dep->dfs_link);
266 
267 		node = dep->signaler;
268 		engine = sched_lock_engine(node, engine, &cache);
269 		lockdep_assert_held(&engine->timeline.lock);
270 
271 		/* Recheck after acquiring the engine->timeline.lock */
272 		if (prio <= node->attr.priority || node_signaled(node))
273 			continue;
274 
275 		node->attr.priority = prio;
276 		if (!list_empty(&node->link)) {
277 			if (!cache.priolist)
278 				cache.priolist =
279 					i915_sched_lookup_priolist(engine,
280 								   prio);
281 			list_move_tail(&node->link, cache.priolist);
282 		} else {
283 			/*
284 			 * If the request is not in the priolist queue because
285 			 * it is not yet runnable, then it doesn't contribute
286 			 * to our preemption decisions. On the other hand,
287 			 * if the request is on the HW, it too is not in the
288 			 * queue; but in that case we may still need to reorder
289 			 * the inflight requests.
290 			 */
291 			if (!i915_sw_fence_done(&node_to_request(node)->submit))
292 				continue;
293 		}
294 
295 		if (prio <= engine->execlists.queue_priority_hint)
296 			continue;
297 
298 		engine->execlists.queue_priority_hint = prio;
299 
300 		/*
301 		 * If we are already the currently executing context, don't
302 		 * bother evaluating if we should preempt ourselves.
303 		 */
304 		if (inflight(node_to_request(node), engine))
305 			continue;
306 
307 		/* Defer (tasklet) submission until after all of our updates. */
308 		tasklet_hi_schedule(&engine->execlists.tasklet);
309 	}
310 
311 	spin_unlock(&engine->timeline.lock);
312 }
313 
314 void i915_schedule(struct i915_request *rq, const struct i915_sched_attr *attr)
315 {
316 	spin_lock_irq(&schedule_lock);
317 	__i915_schedule(&rq->sched, attr);
318 	spin_unlock_irq(&schedule_lock);
319 }
320 
321 static void __bump_priority(struct i915_sched_node *node, unsigned int bump)
322 {
323 	struct i915_sched_attr attr = node->attr;
324 
325 	attr.priority |= bump;
326 	__i915_schedule(node, &attr);
327 }
328 
329 void i915_schedule_bump_priority(struct i915_request *rq, unsigned int bump)
330 {
331 	unsigned long flags;
332 
333 	GEM_BUG_ON(bump & ~I915_PRIORITY_MASK);
334 
335 	if (READ_ONCE(rq->sched.attr.priority) == I915_PRIORITY_INVALID)
336 		return;
337 
338 	spin_lock_irqsave(&schedule_lock, flags);
339 	__bump_priority(&rq->sched, bump);
340 	spin_unlock_irqrestore(&schedule_lock, flags);
341 }
342 
343 void i915_sched_node_init(struct i915_sched_node *node)
344 {
345 	INIT_LIST_HEAD(&node->signalers_list);
346 	INIT_LIST_HEAD(&node->waiters_list);
347 	INIT_LIST_HEAD(&node->link);
348 	node->attr.priority = I915_PRIORITY_INVALID;
349 	node->semaphores = 0;
350 	node->flags = 0;
351 }
352 
353 static struct i915_dependency *
354 i915_dependency_alloc(void)
355 {
356 	return kmem_cache_alloc(global.slab_dependencies, GFP_KERNEL);
357 }
358 
359 static void
360 i915_dependency_free(struct i915_dependency *dep)
361 {
362 	kmem_cache_free(global.slab_dependencies, dep);
363 }
364 
365 bool __i915_sched_node_add_dependency(struct i915_sched_node *node,
366 				      struct i915_sched_node *signal,
367 				      struct i915_dependency *dep,
368 				      unsigned long flags)
369 {
370 	bool ret = false;
371 
372 	spin_lock_irq(&schedule_lock);
373 
374 	if (!node_signaled(signal)) {
375 		INIT_LIST_HEAD(&dep->dfs_link);
376 		list_add(&dep->wait_link, &signal->waiters_list);
377 		list_add(&dep->signal_link, &node->signalers_list);
378 		dep->signaler = signal;
379 		dep->flags = flags;
380 
381 		/* Keep track of whether anyone on this chain has a semaphore */
382 		if (signal->flags & I915_SCHED_HAS_SEMAPHORE_CHAIN &&
383 		    !node_started(signal))
384 			node->flags |= I915_SCHED_HAS_SEMAPHORE_CHAIN;
385 
386 		/*
387 		 * As we do not allow WAIT to preempt inflight requests,
388 		 * once we have executed a request, along with triggering
389 		 * any execution callbacks, we must preserve its ordering
390 		 * within the non-preemptible FIFO.
391 		 */
392 		BUILD_BUG_ON(__NO_PREEMPTION & ~I915_PRIORITY_MASK);
393 		if (flags & I915_DEPENDENCY_EXTERNAL)
394 			__bump_priority(signal, __NO_PREEMPTION);
395 
396 		ret = true;
397 	}
398 
399 	spin_unlock_irq(&schedule_lock);
400 
401 	return ret;
402 }
403 
404 int i915_sched_node_add_dependency(struct i915_sched_node *node,
405 				   struct i915_sched_node *signal)
406 {
407 	struct i915_dependency *dep;
408 
409 	dep = i915_dependency_alloc();
410 	if (!dep)
411 		return -ENOMEM;
412 
413 	if (!__i915_sched_node_add_dependency(node, signal, dep,
414 					      I915_DEPENDENCY_EXTERNAL |
415 					      I915_DEPENDENCY_ALLOC))
416 		i915_dependency_free(dep);
417 
418 	return 0;
419 }
420 
421 void i915_sched_node_fini(struct i915_sched_node *node)
422 {
423 	struct i915_dependency *dep, *tmp;
424 
425 	GEM_BUG_ON(!list_empty(&node->link));
426 
427 	spin_lock_irq(&schedule_lock);
428 
429 	/*
430 	 * Everyone we depended upon (the fences we wait to be signaled)
431 	 * should retire before us and remove themselves from our list.
432 	 * However, retirement is run independently on each timeline and
433 	 * so we may be called out-of-order.
434 	 */
435 	list_for_each_entry_safe(dep, tmp, &node->signalers_list, signal_link) {
436 		GEM_BUG_ON(!node_signaled(dep->signaler));
437 		GEM_BUG_ON(!list_empty(&dep->dfs_link));
438 
439 		list_del(&dep->wait_link);
440 		if (dep->flags & I915_DEPENDENCY_ALLOC)
441 			i915_dependency_free(dep);
442 	}
443 
444 	/* Remove ourselves from everyone who depends upon us */
445 	list_for_each_entry_safe(dep, tmp, &node->waiters_list, wait_link) {
446 		GEM_BUG_ON(dep->signaler != node);
447 		GEM_BUG_ON(!list_empty(&dep->dfs_link));
448 
449 		list_del(&dep->signal_link);
450 		if (dep->flags & I915_DEPENDENCY_ALLOC)
451 			i915_dependency_free(dep);
452 	}
453 
454 	spin_unlock_irq(&schedule_lock);
455 }
456 
457 static void i915_global_scheduler_shrink(void)
458 {
459 	kmem_cache_shrink(global.slab_dependencies);
460 	kmem_cache_shrink(global.slab_priorities);
461 }
462 
463 static void i915_global_scheduler_exit(void)
464 {
465 	kmem_cache_destroy(global.slab_dependencies);
466 	kmem_cache_destroy(global.slab_priorities);
467 }
468 
469 static struct i915_global_scheduler global = { {
470 	.shrink = i915_global_scheduler_shrink,
471 	.exit = i915_global_scheduler_exit,
472 } };
473 
474 int __init i915_global_scheduler_init(void)
475 {
476 	global.slab_dependencies = KMEM_CACHE(i915_dependency,
477 					      SLAB_HWCACHE_ALIGN);
478 	if (!global.slab_dependencies)
479 		return -ENOMEM;
480 
481 	global.slab_priorities = KMEM_CACHE(i915_priolist,
482 					    SLAB_HWCACHE_ALIGN);
483 	if (!global.slab_priorities)
484 		goto err_priorities;
485 
486 	i915_global_register(&global.base);
487 	return 0;
488 
489 err_priorities:
490 	kmem_cache_destroy(global.slab_priorities);
491 	return -ENOMEM;
492 }
493