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
13static struct kmem_cache *slab_dependencies;
14static struct kmem_cache *slab_priorities;
15
16static DEFINE_SPINLOCK(schedule_lock);
17
18static const struct i915_request *
19node_to_request(const struct i915_sched_node *node)
20{
21 return container_of(node, const struct i915_request, sched);
22}
23
24static inline bool node_started(const struct i915_sched_node *node)
25{
26 return i915_request_started(rq: node_to_request(node));
27}
28
29static inline bool node_signaled(const struct i915_sched_node *node)
30{
31 return i915_request_completed(rq: node_to_request(node));
32}
33
34static inline struct i915_priolist *to_priolist(struct rb_node *rb)
35{
36 return rb_entry(rb, struct i915_priolist, node);
37}
38
39static 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
59struct list_head *
60i915_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
72find_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(list: &p->requests);
112
113 rb_link_node(node: &p->node, parent: rb, rb_link: parent);
114 rb_insert_color_cached(node: &p->node, root: &sched_engine->queue, leftmost: first);
115
116 return &p->requests;
117}
118
119void __i915_priolist_free(struct i915_priolist *p)
120{
121 kmem_cache_free(s: slab_priorities, objp: p);
122}
123
124struct sched_cache {
125 struct list_head *priolist;
126};
127
128static struct i915_sched_engine *
129lock_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(lock: &locked->lock);
146 memset(cache, 0, sizeof(*cache));
147 spin_lock(lock: &sched_engine->lock);
148 locked = sched_engine;
149 }
150
151 GEM_BUG_ON(locked != sched_engine);
152 return locked;
153}
154
155static 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(new: &stack.dfs_link, head: &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(node: p->signaler))
210 continue;
211
212 if (prio > READ_ONCE(p->signaler->attr.priority))
213 list_move_tail(list: &p->dfs_link, head: &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(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(lock: &sched_engine->lock);
236
237 /* Fifo and depth-first replacement ensure our deps execute before us */
238 sched_engine = lock_sched_engine(node, locked: sched_engine, cache: &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(list: &dep->dfs_link);
244
245 node = dep->signaler;
246 sched_engine = lock_sched_engine(node, locked: sched_engine, cache: &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(head: &node->link))
271 continue;
272
273 if (i915_request_in_priority_queue(rq: node_to_request(node))) {
274 if (!cache.priolist)
275 cache.priolist =
276 i915_sched_lookup_priolist(sched_engine,
277 prio);
278 list_move_tail(list: &node->link, head: 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(lock: &sched_engine->lock);
287}
288
289void i915_schedule(struct i915_request *rq, const struct i915_sched_attr *attr)
290{
291 spin_lock_irq(lock: &schedule_lock);
292 __i915_schedule(node: &rq->sched, attr);
293 spin_unlock_irq(lock: &schedule_lock);
294}
295
296void i915_sched_node_init(struct i915_sched_node *node)
297{
298 INIT_LIST_HEAD(list: &node->signalers_list);
299 INIT_LIST_HEAD(list: &node->waiters_list);
300 INIT_LIST_HEAD(list: &node->link);
301
302 i915_sched_node_reinit(node);
303}
304
305void 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
316static struct i915_dependency *
317i915_dependency_alloc(void)
318{
319 return kmem_cache_alloc(slab_dependencies, GFP_KERNEL);
320}
321
322static void
323i915_dependency_free(struct i915_dependency *dep)
324{
325 kmem_cache_free(s: slab_dependencies, objp: dep);
326}
327
328bool __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(lock: &schedule_lock);
336
337 if (!node_signaled(node: signal)) {
338 INIT_LIST_HEAD(list: &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(new: &dep->signal_link, head: &node->signalers_list);
345 list_add_rcu(new: &dep->wait_link, head: &signal->waiters_list);
346
347 /* Propagate the chains */
348 node->flags |= signal->flags;
349 ret = true;
350 }
351
352 spin_unlock_irq(lock: &schedule_lock);
353
354 return ret;
355}
356
357int 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: flags | I915_DEPENDENCY_ALLOC))
369 i915_dependency_free(dep);
370
371 return 0;
372}
373
374void i915_sched_node_fini(struct i915_sched_node *node)
375{
376 struct i915_dependency *dep, *tmp;
377
378 spin_lock_irq(lock: &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(entry: &dep->wait_link);
390 if (dep->flags & I915_DEPENDENCY_ALLOC)
391 i915_dependency_free(dep);
392 }
393 INIT_LIST_HEAD(list: &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(entry: &dep->signal_link);
401 if (dep->flags & I915_DEPENDENCY_ALLOC)
402 i915_dependency_free(dep);
403 }
404 INIT_LIST_HEAD(list: &node->waiters_list);
405
406 spin_unlock_irq(lock: &schedule_lock);
407}
408
409void 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(node: 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(rq: signaler))
430 continue;
431
432 i915_request_show(m, rq: signaler, prefix, indent: indent + 2);
433 }
434 rcu_read_unlock();
435}
436
437static 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(t: &sched_engine->tasklet); /* flush the callback */
443 kfree(objp: sched_engine);
444}
445
446static bool default_disabled(struct i915_sched_engine *sched_engine)
447{
448 return false;
449}
450
451struct i915_sched_engine *
452i915_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(kref: &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(list: &sched_engine->requests);
468 INIT_LIST_HEAD(list: &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
488void i915_scheduler_module_exit(void)
489{
490 kmem_cache_destroy(s: slab_dependencies);
491 kmem_cache_destroy(s: slab_priorities);
492}
493
494int __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
508err_priorities:
509 kmem_cache_destroy(s: slab_dependencies);
510 return -ENOMEM;
511}
512

source code of linux/drivers/gpu/drm/i915/i915_scheduler.c