1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
|
/*
* SPDX-License-Identifier: MIT
*
* Copyright © 2018 Intel Corporation
*/
#include <linux/mutex.h>
#include "i915_drv.h"
#include "i915_globals.h"
#include "i915_request.h"
#include "i915_scheduler.h"
static struct i915_global_scheduler {
struct i915_global base;
struct kmem_cache *slab_dependencies;
struct kmem_cache *slab_priorities;
} global;
static DEFINE_SPINLOCK(schedule_lock);
static const struct i915_request *
node_to_request(const struct i915_sched_node *node)
{
return container_of(node, const struct i915_request, sched);
}
static inline bool node_started(const struct i915_sched_node *node)
{
return i915_request_started(node_to_request(node));
}
static inline bool node_signaled(const struct i915_sched_node *node)
{
return i915_request_completed(node_to_request(node));
}
static inline struct i915_priolist *to_priolist(struct rb_node *rb)
{
return rb_entry(rb, struct i915_priolist, node);
}
static void assert_priolists(struct intel_engine_execlists * const execlists)
{
struct rb_node *rb;
long last_prio, i;
if (!IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM))
return;
GEM_BUG_ON(rb_first_cached(&execlists->queue) !=
rb_first(&execlists->queue.rb_root));
last_prio = (INT_MAX >> I915_USER_PRIORITY_SHIFT) + 1;
for (rb = rb_first_cached(&execlists->queue); rb; rb = rb_next(rb)) {
const struct i915_priolist *p = to_priolist(rb);
GEM_BUG_ON(p->priority >= last_prio);
last_prio = p->priority;
GEM_BUG_ON(!p->used);
for (i = 0; i < ARRAY_SIZE(p->requests); i++) {
if (list_empty(&p->requests[i]))
continue;
GEM_BUG_ON(!(p->used & BIT(i)));
}
}
}
struct list_head *
i915_sched_lookup_priolist(struct intel_engine_cs *engine, int prio)
{
struct intel_engine_execlists * const execlists = &engine->execlists;
struct i915_priolist *p;
struct rb_node **parent, *rb;
bool first = true;
int idx, i;
lockdep_assert_held(&engine->active.lock);
assert_priolists(execlists);
/* buckets sorted from highest [in slot 0] to lowest priority */
idx = I915_PRIORITY_COUNT - (prio & I915_PRIORITY_MASK) - 1;
prio >>= I915_USER_PRIORITY_SHIFT;
if (unlikely(execlists->no_priolist))
prio = I915_PRIORITY_NORMAL;
find_priolist:
/* most positive priority is scheduled first, equal priorities fifo */
rb = NULL;
parent = &execlists->queue.rb_root.rb_node;
while (*parent) {
rb = *parent;
p = to_priolist(rb);
if (prio > p->priority) {
parent = &rb->rb_left;
} else if (prio < p->priority) {
parent = &rb->rb_right;
first = false;
} else {
goto out;
}
}
if (prio == I915_PRIORITY_NORMAL) {
p = &execlists->default_priolist;
} else {
p = kmem_cache_alloc(global.slab_priorities, GFP_ATOMIC);
/* Convert an allocation failure to a priority bump */
if (unlikely(!p)) {
prio = I915_PRIORITY_NORMAL; /* recurses just once */
/* To maintain ordering with all rendering, after an
* allocation failure we have to disable all scheduling.
* Requests will then be executed in fifo, and schedule
* will ensure that dependencies are emitted in fifo.
* There will be still some reordering with existing
* requests, so if userspace lied about their
* dependencies that reordering may be visible.
*/
execlists->no_priolist = true;
goto find_priolist;
}
}
p->priority = prio;
for (i = 0; i < ARRAY_SIZE(p->requests); i++)
INIT_LIST_HEAD(&p->requests[i]);
rb_link_node(&p->node, rb, parent);
rb_insert_color_cached(&p->node, &execlists->queue, first);
p->used = 0;
out:
p->used |= BIT(idx);
return &p->requests[idx];
}
void __i915_priolist_free(struct i915_priolist *p)
{
kmem_cache_free(global.slab_priorities, p);
}
struct sched_cache {
struct list_head *priolist;
};
static struct intel_engine_cs *
sched_lock_engine(const struct i915_sched_node *node,
struct intel_engine_cs *locked,
struct sched_cache *cache)
{
const struct i915_request *rq = node_to_request(node);
struct intel_engine_cs *engine;
GEM_BUG_ON(!locked);
/*
* Virtual engines complicate acquiring the engine timeline lock,
* as their rq->engine pointer is not stable until under that
* engine lock. The simple ploy we use is to take the lock then
* check that the rq still belongs to the newly locked engine.
*/
while (locked != (engine = READ_ONCE(rq->engine))) {
spin_unlock(&locked->active.lock);
memset(cache, 0, sizeof(*cache));
spin_lock(&engine->active.lock);
locked = engine;
}
GEM_BUG_ON(locked != engine);
return locked;
}
static inline int rq_prio(const struct i915_request *rq)
{
return rq->sched.attr.priority | __NO_PREEMPTION;
}
static inline bool need_preempt(int prio, int active)
{
/*
* Allow preemption of low -> normal -> high, but we do
* not allow low priority tasks to preempt other low priority
* tasks under the impression that latency for low priority
* tasks does not matter (as much as background throughput),
* so kiss.
*/
return prio >= max(I915_PRIORITY_NORMAL, active);
}
static void kick_submission(struct intel_engine_cs *engine,
const struct i915_request *rq,
int prio)
{
const struct i915_request *inflight;
/*
* We only need to kick the tasklet once for the high priority
* new context we add into the queue.
*/
if (prio <= engine->execlists.queue_priority_hint)
return;
rcu_read_lock();
/* Nothing currently active? We're overdue for a submission! */
inflight = execlists_active(&engine->execlists);
if (!inflight)
goto unlock;
engine->execlists.queue_priority_hint = prio;
/*
* If we are already the currently executing context, don't
* bother evaluating if we should preempt ourselves.
*/
if (inflight->context == rq->context)
goto unlock;
if (need_preempt(prio, rq_prio(inflight)))
tasklet_hi_schedule(&engine->execlists.tasklet);
unlock:
rcu_read_unlock();
}
static void __i915_schedule(struct i915_sched_node *node,
const struct i915_sched_attr *attr)
{
const int prio = max(attr->priority, node->attr.priority);
struct intel_engine_cs *engine;
struct i915_dependency *dep, *p;
struct i915_dependency stack;
struct sched_cache cache;
LIST_HEAD(dfs);
/* Needed in order to use the temporary link inside i915_dependency */
lockdep_assert_held(&schedule_lock);
GEM_BUG_ON(prio == I915_PRIORITY_INVALID);
if (node_signaled(node))
return;
stack.signaler = node;
list_add(&stack.dfs_link, &dfs);
/*
* Recursively bump all dependent priorities to match the new request.
*
* A naive approach would be to use recursion:
* static void update_priorities(struct i915_sched_node *node, prio) {
* list_for_each_entry(dep, &node->signalers_list, signal_link)
* update_priorities(dep->signal, prio)
* queue_request(node);
* }
* but that may have unlimited recursion depth and so runs a very
* real risk of overunning the kernel stack. Instead, we build
* a flat list of all dependencies starting with the current request.
* As we walk the list of dependencies, we add all of its dependencies
* to the end of the list (this may include an already visited
* request) and continue to walk onwards onto the new dependencies. The
* end result is a topological list of requests in reverse order, the
* last element in the list is the request we must execute first.
*/
list_for_each_entry(dep, &dfs, dfs_link) {
struct i915_sched_node *node = dep->signaler;
/* If we are already flying, we know we have no signalers */
if (node_started(node))
continue;
/*
* Within an engine, there can be no cycle, but we may
* refer to the same dependency chain multiple times
* (redundant dependencies are not eliminated) and across
* engines.
*/
list_for_each_entry(p, &node->signalers_list, signal_link) {
GEM_BUG_ON(p == dep); /* no cycles! */
if (node_signaled(p->signaler))
continue;
if (prio > READ_ONCE(p->signaler->attr.priority))
list_move_tail(&p->dfs_link, &dfs);
}
}
/*
* If we didn't need to bump any existing priorities, and we haven't
* yet submitted this request (i.e. there is no potential race with
* execlists_submit_request()), we can set our own priority and skip
* acquiring the engine locks.
*/
if (node->attr.priority == I915_PRIORITY_INVALID) {
GEM_BUG_ON(!list_empty(&node->link));
node->attr = *attr;
if (stack.dfs_link.next == stack.dfs_link.prev)
return;
__list_del_entry(&stack.dfs_link);
}
memset(&cache, 0, sizeof(cache));
engine = node_to_request(node)->engine;
spin_lock(&engine->active.lock);
/* Fifo and depth-first replacement ensure our deps execute before us */
engine = sched_lock_engine(node, engine, &cache);
list_for_each_entry_safe_reverse(dep, p, &dfs, dfs_link) {
INIT_LIST_HEAD(&dep->dfs_link);
node = dep->signaler;
engine = sched_lock_engine(node, engine, &cache);
lockdep_assert_held(&engine->active.lock);
/* Recheck after acquiring the engine->timeline.lock */
if (prio <= node->attr.priority || node_signaled(node))
continue;
GEM_BUG_ON(node_to_request(node)->engine != engine);
WRITE_ONCE(node->attr.priority, prio);
/*
* Once the request is ready, it will be placed into the
* priority lists and then onto the HW runlist. Before the
* request is ready, it does not contribute to our preemption
* decisions and we can safely ignore it, as it will, and
* any preemption required, be dealt with upon submission.
* See engine->submit_request()
*/
if (list_empty(&node->link))
continue;
if (i915_request_in_priority_queue(node_to_request(node))) {
if (!cache.priolist)
cache.priolist =
i915_sched_lookup_priolist(engine,
prio);
list_move_tail(&node->link, cache.priolist);
}
/* Defer (tasklet) submission until after all of our updates. */
kick_submission(engine, node_to_request(node), prio);
}
spin_unlock(&engine->active.lock);
}
void i915_schedule(struct i915_request *rq, const struct i915_sched_attr *attr)
{
spin_lock_irq(&schedule_lock);
__i915_schedule(&rq->sched, attr);
spin_unlock_irq(&schedule_lock);
}
static void __bump_priority(struct i915_sched_node *node, unsigned int bump)
{
struct i915_sched_attr attr = node->attr;
if (attr.priority & bump)
return;
attr.priority |= bump;
__i915_schedule(node, &attr);
}
void i915_schedule_bump_priority(struct i915_request *rq, unsigned int bump)
{
unsigned long flags;
GEM_BUG_ON(bump & ~I915_PRIORITY_MASK);
if (READ_ONCE(rq->sched.attr.priority) & bump)
return;
spin_lock_irqsave(&schedule_lock, flags);
__bump_priority(&rq->sched, bump);
spin_unlock_irqrestore(&schedule_lock, flags);
}
void i915_sched_node_init(struct i915_sched_node *node)
{
INIT_LIST_HEAD(&node->signalers_list);
INIT_LIST_HEAD(&node->waiters_list);
INIT_LIST_HEAD(&node->link);
i915_sched_node_reinit(node);
}
void i915_sched_node_reinit(struct i915_sched_node *node)
{
node->attr.priority = I915_PRIORITY_INVALID;
node->semaphores = 0;
node->flags = 0;
GEM_BUG_ON(!list_empty(&node->signalers_list));
GEM_BUG_ON(!list_empty(&node->waiters_list));
GEM_BUG_ON(!list_empty(&node->link));
}
static struct i915_dependency *
i915_dependency_alloc(void)
{
return kmem_cache_alloc(global.slab_dependencies, GFP_KERNEL);
}
static void
i915_dependency_free(struct i915_dependency *dep)
{
kmem_cache_free(global.slab_dependencies, dep);
}
bool __i915_sched_node_add_dependency(struct i915_sched_node *node,
struct i915_sched_node *signal,
struct i915_dependency *dep,
unsigned long flags)
{
bool ret = false;
spin_lock_irq(&schedule_lock);
if (!node_signaled(signal)) {
INIT_LIST_HEAD(&dep->dfs_link);
dep->signaler = signal;
dep->waiter = node;
dep->flags = flags;
/* Keep track of whether anyone on this chain has a semaphore */
if (signal->flags & I915_SCHED_HAS_SEMAPHORE_CHAIN &&
!node_started(signal))
node->flags |= I915_SCHED_HAS_SEMAPHORE_CHAIN;
/* All set, now publish. Beware the lockless walkers. */
list_add_rcu(&dep->signal_link, &node->signalers_list);
list_add_rcu(&dep->wait_link, &signal->waiters_list);
/*
* As we do not allow WAIT to preempt inflight requests,
* once we have executed a request, along with triggering
* any execution callbacks, we must preserve its ordering
* within the non-preemptible FIFO.
*/
BUILD_BUG_ON(__NO_PREEMPTION & ~I915_PRIORITY_MASK);
if (flags & I915_DEPENDENCY_EXTERNAL)
__bump_priority(signal, __NO_PREEMPTION);
ret = true;
}
spin_unlock_irq(&schedule_lock);
return ret;
}
int i915_sched_node_add_dependency(struct i915_sched_node *node,
struct i915_sched_node *signal)
{
struct i915_dependency *dep;
dep = i915_dependency_alloc();
if (!dep)
return -ENOMEM;
if (!__i915_sched_node_add_dependency(node, signal, dep,
I915_DEPENDENCY_EXTERNAL |
I915_DEPENDENCY_ALLOC))
i915_dependency_free(dep);
return 0;
}
void i915_sched_node_fini(struct i915_sched_node *node)
{
struct i915_dependency *dep, *tmp;
spin_lock_irq(&schedule_lock);
/*
* Everyone we depended upon (the fences we wait to be signaled)
* should retire before us and remove themselves from our list.
* However, retirement is run independently on each timeline and
* so we may be called out-of-order.
*/
list_for_each_entry_safe(dep, tmp, &node->signalers_list, signal_link) {
GEM_BUG_ON(!list_empty(&dep->dfs_link));
list_del_rcu(&dep->wait_link);
if (dep->flags & I915_DEPENDENCY_ALLOC)
i915_dependency_free(dep);
}
INIT_LIST_HEAD(&node->signalers_list);
/* Remove ourselves from everyone who depends upon us */
list_for_each_entry_safe(dep, tmp, &node->waiters_list, wait_link) {
GEM_BUG_ON(dep->signaler != node);
GEM_BUG_ON(!list_empty(&dep->dfs_link));
list_del_rcu(&dep->signal_link);
if (dep->flags & I915_DEPENDENCY_ALLOC)
i915_dependency_free(dep);
}
INIT_LIST_HEAD(&node->waiters_list);
spin_unlock_irq(&schedule_lock);
}
static void i915_global_scheduler_shrink(void)
{
kmem_cache_shrink(global.slab_dependencies);
kmem_cache_shrink(global.slab_priorities);
}
static void i915_global_scheduler_exit(void)
{
kmem_cache_destroy(global.slab_dependencies);
kmem_cache_destroy(global.slab_priorities);
}
static struct i915_global_scheduler global = { {
.shrink = i915_global_scheduler_shrink,
.exit = i915_global_scheduler_exit,
} };
int __init i915_global_scheduler_init(void)
{
global.slab_dependencies = KMEM_CACHE(i915_dependency,
SLAB_HWCACHE_ALIGN |
SLAB_TYPESAFE_BY_RCU);
if (!global.slab_dependencies)
return -ENOMEM;
global.slab_priorities = KMEM_CACHE(i915_priolist,
SLAB_HWCACHE_ALIGN);
if (!global.slab_priorities)
goto err_priorities;
i915_global_register(&global.base);
return 0;
err_priorities:
kmem_cache_destroy(global.slab_priorities);
return -ENOMEM;
}
|