drivers/gpu/drm/i915/i915_scheduler.c

   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;
  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;
  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 }
  62
  63 struct list_head *
  64 i915_sched_lookup_priolist(struct intel_engine_cs *engine, int prio)
  65 {
  66         struct intel_engine_execlists * const execlists = &engine->execlists;
  67         struct i915_priolist *p;
  68         struct rb_node **parent, *rb;
  69         bool first = true;
  70
  71         lockdep_assert_held(&engine->active.lock);
  72         assert_priolists(execlists);
  73
  74         if (unlikely(execlists->no_priolist))
  75                 prio = I915_PRIORITY_NORMAL;
  76
  77 find_priolist:
  78         /* most positive priority is scheduled first, equal priorities fifo */
  79         rb = NULL;
  80         parent = &execlists->queue.rb_root.rb_node;
  81         while (*parent) {
  82                 rb = *parent;
  83                 p = to_priolist(rb);
  84                 if (prio > p->priority) {
  85                         parent = &rb->rb_left;
  86                 } else if (prio < p->priority) {
  87                         parent = &rb->rb_right;
  88                         first = false;
  89                 } else {
  90                         return &p->requests;
  91                 }
  92         }
  93
  94         if (prio == I915_PRIORITY_NORMAL) {
  95                 p = &execlists->default_priolist;
  96         } else {
  97                 p = kmem_cache_alloc(global.slab_priorities, GFP_ATOMIC);
  98                 /* Convert an allocation failure to a priority bump */
  99                 if (unlikely(!p)) {
 100                         prio = I915_PRIORITY_NORMAL; /* recurses just once */
 101
 102                         /* To maintain ordering with all rendering, after an
 103                          * allocation failure we have to disable all scheduling.
 104                          * Requests will then be executed in fifo, and schedule
 105                          * will ensure that dependencies are emitted in fifo.
 106                          * There will be still some reordering with existing
 107                          * requests, so if userspace lied about their
 108                          * dependencies that reordering may be visible.
 109                          */
 110                         execlists->no_priolist = true;
 111                         goto find_priolist;
 112                 }
 113         }
 114
 115         p->priority = prio;
 116         INIT_LIST_HEAD(&p->requests);
 117
 118         rb_link_node(&p->node, rb, parent);
 119         rb_insert_color_cached(&p->node, &execlists->queue, first);
 120
 121         return &p->requests;
 122 }
 123
 124 void __i915_priolist_free(struct i915_priolist *p)
 125 {
 126         kmem_cache_free(global.slab_priorities, p);
 127 }
 128
 129 struct sched_cache {
 130         struct list_head *priolist;
 131 };
 132
 133 static struct intel_engine_cs *
 134 sched_lock_engine(const struct i915_sched_node *node,
 135                   struct intel_engine_cs *locked,
 136                   struct sched_cache *cache)
 137 {
 138         const struct i915_request *rq = node_to_request(node);
 139         struct intel_engine_cs *engine;
 140
 141         GEM_BUG_ON(!locked);
 142
 143         /*
 144          * Virtual engines complicate acquiring the engine timeline lock,
 145          * as their rq->engine pointer is not stable until under that
 146          * engine lock. The simple ploy we use is to take the lock then
 147          * check that the rq still belongs to the newly locked engine.
 148          */
 149         while (locked != (engine = READ_ONCE(rq->engine))) {
 150                 spin_unlock(&locked->active.lock);
 151                 memset(cache, 0, sizeof(*cache));
 152                 spin_lock(&engine->active.lock);
 153                 locked = engine;
 154         }
 155
 156         GEM_BUG_ON(locked != engine);
 157         return locked;
 158 }
 159
 160 static inline int rq_prio(const struct i915_request *rq)
 161 {
 162         return rq->sched.attr.priority;
 163 }
 164
 165 static inline bool need_preempt(int prio, int active)
 166 {
 167         /*
 168          * Allow preemption of low -> normal -> high, but we do
 169          * not allow low priority tasks to preempt other low priority
 170          * tasks under the impression that latency for low priority
 171          * tasks does not matter (as much as background throughput),
 172          * so kiss.
 173          */
 174         return prio >= max(I915_PRIORITY_NORMAL, active);
 175 }
 176
 177 static void kick_submission(struct intel_engine_cs *engine,
 178                             const struct i915_request *rq,
 179                             int prio)
 180 {
 181         const struct i915_request *inflight;
 182
 183         /*
 184          * We only need to kick the tasklet once for the high priority
 185          * new context we add into the queue.
 186          */
 187         if (prio <= engine->execlists.queue_priority_hint)
 188                 return;
 189
 190         rcu_read_lock();
 191
 192         /* Nothing currently active? We're overdue for a submission! */
 193         inflight = execlists_active(&engine->execlists);
 194         if (!inflight)
 195                 goto unlock;
 196
 197         /*
 198          * If we are already the currently executing context, don't
 199          * bother evaluating if we should preempt ourselves.
 200          */
 201         if (inflight->context == rq->context)
 202                 goto unlock;
 203
 204         ENGINE_TRACE(engine,
 205                      "bumping queue-priority-hint:%d for rq:%llx:%lld, inflight:%llx:%lld prio %d\n",
 206                      prio,
 207                      rq->fence.context, rq->fence.seqno,
 208                      inflight->fence.context, inflight->fence.seqno,
 209                      inflight->sched.attr.priority);
 210
 211         engine->execlists.queue_priority_hint = prio;
 212         if (need_preempt(prio, rq_prio(inflight)))
 213                 tasklet_hi_schedule(&engine->execlists.tasklet);
 214
 215 unlock:
 216         rcu_read_unlock();
 217 }
 218
 219 static void __i915_schedule(struct i915_sched_node *node,
 220                             const struct i915_sched_attr *attr)
 221 {
 222         const int prio = max(attr->priority, node->attr.priority);
 223         struct intel_engine_cs *engine;
 224         struct i915_dependency *dep, *p;
 225         struct i915_dependency stack;
 226         struct sched_cache cache;
 227         LIST_HEAD(dfs);
 228
 229         /* Needed in order to use the temporary link inside i915_dependency */
 230         lockdep_assert_held(&schedule_lock);
 231         GEM_BUG_ON(prio == I915_PRIORITY_INVALID);
 232
 233         if (node_signaled(node))
 234                 return;
 235
 236         stack.signaler = node;
 237         list_add(&stack.dfs_link, &dfs);
 238
 239         /*
 240          * Recursively bump all dependent priorities to match the new request.
 241          *
 242          * A naive approach would be to use recursion:
 243          * static void update_priorities(struct i915_sched_node *node, prio) {
 244          *      list_for_each_entry(dep, &node->signalers_list, signal_link)
 245          *              update_priorities(dep->signal, prio)
 246          *      queue_request(node);
 247          * }
 248          * but that may have unlimited recursion depth and so runs a very
 249          * real risk of overunning the kernel stack. Instead, we build
 250          * a flat list of all dependencies starting with the current request.
 251          * As we walk the list of dependencies, we add all of its dependencies
 252          * to the end of the list (this may include an already visited
 253          * request) and continue to walk onwards onto the new dependencies. The
 254          * end result is a topological list of requests in reverse order, the
 255          * last element in the list is the request we must execute first.
 256          */
 257         list_for_each_entry(dep, &dfs, dfs_link) {
 258                 struct i915_sched_node *node = dep->signaler;
 259
 260                 /* If we are already flying, we know we have no signalers */
 261                 if (node_started(node))
 262                         continue;
 263
 264                 /*
 265                  * Within an engine, there can be no cycle, but we may
 266                  * refer to the same dependency chain multiple times
 267                  * (redundant dependencies are not eliminated) and across
 268                  * engines.
 269                  */
 270                 list_for_each_entry(p, &node->signalers_list, signal_link) {
 271                         GEM_BUG_ON(p == dep); /* no cycles! */
 272
 273                         if (node_signaled(p->signaler))
 274                                 continue;
 275
 276                         if (prio > READ_ONCE(p->signaler->attr.priority))
 277                                 list_move_tail(&p->dfs_link, &dfs);
 278                 }
 279         }
 280
 281         /*
 282          * If we didn't need to bump any existing priorities, and we haven't
 283          * yet submitted this request (i.e. there is no potential race with
 284          * execlists_submit_request()), we can set our own priority and skip
 285          * acquiring the engine locks.
 286          */
 287         if (node->attr.priority == I915_PRIORITY_INVALID) {
 288                 GEM_BUG_ON(!list_empty(&node->link));
 289                 node->attr = *attr;
 290
 291                 if (stack.dfs_link.next == stack.dfs_link.prev)
 292                         return;
 293
 294                 __list_del_entry(&stack.dfs_link);
 295         }
 296
 297         memset(&cache, 0, sizeof(cache));
 298         engine = node_to_request(node)->engine;
 299         spin_lock(&engine->active.lock);
 300
 301         /* Fifo and depth-first replacement ensure our deps execute before us */
 302         engine = sched_lock_engine(node, engine, &cache);
 303         list_for_each_entry_safe_reverse(dep, p, &dfs, dfs_link) {
 304                 INIT_LIST_HEAD(&dep->dfs_link);
 305
 306                 node = dep->signaler;
 307                 engine = sched_lock_engine(node, engine, &cache);
 308                 lockdep_assert_held(&engine->active.lock);
 309
 310                 /* Recheck after acquiring the engine->timeline.lock */
 311                 if (prio <= node->attr.priority || node_signaled(node))
 312                         continue;
 313
 314                 GEM_BUG_ON(node_to_request(node)->engine != engine);
 315
 316                 WRITE_ONCE(node->attr.priority, prio);
 317
 318                 /*
 319                  * Once the request is ready, it will be placed into the
 320                  * priority lists and then onto the HW runlist. Before the
 321                  * request is ready, it does not contribute to our preemption
 322                  * decisions and we can safely ignore it, as it will, and
 323                  * any preemption required, be dealt with upon submission.
 324                  * See engine->submit_request()
 325                  */
 326                 if (list_empty(&node->link))
 327                         continue;
 328
 329                 if (i915_request_in_priority_queue(node_to_request(node))) {
 330                         if (!cache.priolist)
 331                                 cache.priolist =
 332                                         i915_sched_lookup_priolist(engine,
 333                                                                    prio);
 334                         list_move_tail(&node->link, cache.priolist);
 335                 }
 336
 337                 /* Defer (tasklet) submission until after all of our updates. */
 338                 kick_submission(engine, node_to_request(node), prio);
 339         }
 340
 341         spin_unlock(&engine->active.lock);
 342 }
 343
 344 void i915_schedule(struct i915_request *rq, const struct i915_sched_attr *attr)
 345 {
 346         spin_lock_irq(&schedule_lock);
 347         __i915_schedule(&rq->sched, attr);
 348         spin_unlock_irq(&schedule_lock);
 349 }
 350
 351 void i915_sched_node_init(struct i915_sched_node *node)
 352 {
 353         INIT_LIST_HEAD(&node->signalers_list);
 354         INIT_LIST_HEAD(&node->waiters_list);
 355         INIT_LIST_HEAD(&node->link);
 356
 357         i915_sched_node_reinit(node);
 358 }
 359
 360 void i915_sched_node_reinit(struct i915_sched_node *node)
 361 {
 362         node->attr.priority = I915_PRIORITY_INVALID;
 363         node->semaphores = 0;
 364         node->flags = 0;
 365
 366         GEM_BUG_ON(!list_empty(&node->signalers_list));
 367         GEM_BUG_ON(!list_empty(&node->waiters_list));
 368         GEM_BUG_ON(!list_empty(&node->link));
 369 }
 370
 371 static struct i915_dependency *
 372 i915_dependency_alloc(void)
 373 {
 374         return kmem_cache_alloc(global.slab_dependencies, GFP_KERNEL);
 375 }
 376
 377 static void
 378 i915_dependency_free(struct i915_dependency *dep)
 379 {
 380         kmem_cache_free(global.slab_dependencies, dep);
 381 }
 382
 383 bool __i915_sched_node_add_dependency(struct i915_sched_node *node,
 384                                       struct i915_sched_node *signal,
 385                                       struct i915_dependency *dep,
 386                                       unsigned long flags)
 387 {
 388         bool ret = false;
 389
 390         spin_lock_irq(&schedule_lock);
 391
 392         if (!node_signaled(signal)) {
 393                 INIT_LIST_HEAD(&dep->dfs_link);
 394                 dep->signaler = signal;
 395                 dep->waiter = node;
 396                 dep->flags = flags;
 397
 398                 /* All set, now publish. Beware the lockless walkers. */
 399                 list_add_rcu(&dep->signal_link, &node->signalers_list);
 400                 list_add_rcu(&dep->wait_link, &signal->waiters_list);
 401
 402                 /* Propagate the chains */
 403                 node->flags |= signal->flags;
 404                 ret = true;
 405         }
 406
 407         spin_unlock_irq(&schedule_lock);
 408
 409         return ret;
 410 }
 411
 412 int i915_sched_node_add_dependency(struct i915_sched_node *node,
 413                                    struct i915_sched_node *signal,
 414                                    unsigned long flags)
 415 {
 416         struct i915_dependency *dep;
 417
 418         dep = i915_dependency_alloc();
 419         if (!dep)
 420                 return -ENOMEM;
 421
 422         if (!__i915_sched_node_add_dependency(node, signal, dep,
 423                                               flags | I915_DEPENDENCY_ALLOC))
 424                 i915_dependency_free(dep);
 425
 426         return 0;
 427 }
 428
 429 void i915_sched_node_fini(struct i915_sched_node *node)
 430 {
 431         struct i915_dependency *dep, *tmp;
 432
 433         spin_lock_irq(&schedule_lock);
 434
 435         /*
 436          * Everyone we depended upon (the fences we wait to be signaled)
 437          * should retire before us and remove themselves from our list.
 438          * However, retirement is run independently on each timeline and
 439          * so we may be called out-of-order.
 440          */
 441         list_for_each_entry_safe(dep, tmp, &node->signalers_list, signal_link) {
 442                 GEM_BUG_ON(!list_empty(&dep->dfs_link));
 443
 444                 list_del_rcu(&dep->wait_link);
 445                 if (dep->flags & I915_DEPENDENCY_ALLOC)
 446                         i915_dependency_free(dep);
 447         }
 448         INIT_LIST_HEAD(&node->signalers_list);
 449
 450         /* Remove ourselves from everyone who depends upon us */
 451         list_for_each_entry_safe(dep, tmp, &node->waiters_list, wait_link) {
 452                 GEM_BUG_ON(dep->signaler != node);
 453                 GEM_BUG_ON(!list_empty(&dep->dfs_link));
 454
 455                 list_del_rcu(&dep->signal_link);
 456                 if (dep->flags & I915_DEPENDENCY_ALLOC)
 457                         i915_dependency_free(dep);
 458         }
 459         INIT_LIST_HEAD(&node->waiters_list);
 460
 461         spin_unlock_irq(&schedule_lock);
 462 }
 463
 464 void i915_request_show_with_schedule(struct drm_printer *m,
 465                                      const struct i915_request *rq,
 466                                      const char *prefix,
 467                                      int indent)
 468 {
 469         struct i915_dependency *dep;
 470
 471         i915_request_show(m, rq, prefix, indent);
 472         if (i915_request_completed(rq))
 473                 return;
 474
 475         rcu_read_lock();
 476         for_each_signaler(dep, rq) {
 477                 const struct i915_request *signaler =
 478                         node_to_request(dep->signaler);
 479
 480                 /* Dependencies along the same timeline are expected. */
 481                 if (signaler->timeline == rq->timeline)
 482                         continue;
 483
 484                 if (__i915_request_is_complete(signaler))
 485                         continue;
 486
 487                 i915_request_show(m, signaler, prefix, indent + 2);
 488         }
 489         rcu_read_unlock();
 490 }
 491
 492 static void i915_global_scheduler_shrink(void)
 493 {
 494         kmem_cache_shrink(global.slab_dependencies);
 495         kmem_cache_shrink(global.slab_priorities);
 496 }
 497
 498 static void i915_global_scheduler_exit(void)
 499 {
 500         kmem_cache_destroy(global.slab_dependencies);
 501         kmem_cache_destroy(global.slab_priorities);
 502 }
 503
 504 static struct i915_global_scheduler global = { {
 505         .shrink = i915_global_scheduler_shrink,
 506         .exit = i915_global_scheduler_exit,
 507 } };
 508
 509 int __init i915_global_scheduler_init(void)
 510 {
 511         global.slab_dependencies = KMEM_CACHE(i915_dependency,
 512                                               SLAB_HWCACHE_ALIGN |
 513                                               SLAB_TYPESAFE_BY_RCU);
 514         if (!global.slab_dependencies)
 515                 return -ENOMEM;
 516
 517         global.slab_priorities = KMEM_CACHE(i915_priolist, 0);
 518         if (!global.slab_priorities)
 519                 goto err_priorities;
 520
 521         i915_global_register(&global.base);
 522         return 0;
 523
 524 err_priorities:
 525         kmem_cache_destroy(global.slab_priorities);
 526         return -ENOMEM;
 527 }