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, 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->active.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         const struct i915_request *rq = node_to_request(node);
 154         struct intel_engine_cs *engine;
 155
 156         GEM_BUG_ON(!locked);
 157
 158         /*
 159          * Virtual engines complicate acquiring the engine timeline lock,
 160          * as their rq->engine pointer is not stable until under that
 161          * engine lock. The simple ploy we use is to take the lock then
 162          * check that the rq still belongs to the newly locked engine.
 163          */
 164         while (locked != (engine = READ_ONCE(rq->engine))) {
 165                 spin_unlock(&locked->active.lock);
 166                 memset(cache, 0, sizeof(*cache));
 167                 spin_lock(&engine->active.lock);
 168                 locked = engine;
 169         }
 170
 171         GEM_BUG_ON(locked != engine);
 172         return locked;
 173 }
 174
 175 static inline int rq_prio(const struct i915_request *rq)
 176 {
 177         return rq->sched.attr.priority | __NO_PREEMPTION;
 178 }
 179
 180 static void kick_submission(struct intel_engine_cs *engine, int prio)
 181 {
 182         const struct i915_request *inflight = *engine->execlists.active;
 183
 184         /*
 185          * If we are already the currently executing context, don't
 186          * bother evaluating if we should preempt ourselves, or if
 187          * we expect nothing to change as a result of running the
 188          * tasklet, i.e. we have not change the priority queue
 189          * sufficiently to oust the running context.
 190          */
 191         if (!inflight || !i915_scheduler_need_preempt(prio, rq_prio(inflight)))
 192                 return;
 193
 194         tasklet_hi_schedule(&engine->execlists.tasklet);
 195 }
 196
 197 static void __i915_schedule(struct i915_sched_node *node,
 198                             const struct i915_sched_attr *attr)
 199 {
 200         struct intel_engine_cs *engine;
 201         struct i915_dependency *dep, *p;
 202         struct i915_dependency stack;
 203         const int prio = attr->priority;
 204         struct sched_cache cache;
 205         LIST_HEAD(dfs);
 206
 207         /* Needed in order to use the temporary link inside i915_dependency */
 208         lockdep_assert_held(&schedule_lock);
 209         GEM_BUG_ON(prio == I915_PRIORITY_INVALID);
 210
 211         if (prio <= READ_ONCE(node->attr.priority))
 212                 return;
 213
 214         if (node_signaled(node))
 215                 return;
 216
 217         stack.signaler = node;
 218         list_add(&stack.dfs_link, &dfs);
 219
 220         /*
 221          * Recursively bump all dependent priorities to match the new request.
 222          *
 223          * A naive approach would be to use recursion:
 224          * static void update_priorities(struct i915_sched_node *node, prio) {
 225          *      list_for_each_entry(dep, &node->signalers_list, signal_link)
 226          *              update_priorities(dep->signal, prio)
 227          *      queue_request(node);
 228          * }
 229          * but that may have unlimited recursion depth and so runs a very
 230          * real risk of overunning the kernel stack. Instead, we build
 231          * a flat list of all dependencies starting with the current request.
 232          * As we walk the list of dependencies, we add all of its dependencies
 233          * to the end of the list (this may include an already visited
 234          * request) and continue to walk onwards onto the new dependencies. The
 235          * end result is a topological list of requests in reverse order, the
 236          * last element in the list is the request we must execute first.
 237          */
 238         list_for_each_entry(dep, &dfs, dfs_link) {
 239                 struct i915_sched_node *node = dep->signaler;
 240
 241                 /* If we are already flying, we know we have no signalers */
 242                 if (node_started(node))
 243                         continue;
 244
 245                 /*
 246                  * Within an engine, there can be no cycle, but we may
 247                  * refer to the same dependency chain multiple times
 248                  * (redundant dependencies are not eliminated) and across
 249                  * engines.
 250                  */
 251                 list_for_each_entry(p, &node->signalers_list, signal_link) {
 252                         GEM_BUG_ON(p == dep); /* no cycles! */
 253
 254                         if (node_signaled(p->signaler))
 255                                 continue;
 256
 257                         if (prio > READ_ONCE(p->signaler->attr.priority))
 258                                 list_move_tail(&p->dfs_link, &dfs);
 259                 }
 260         }
 261
 262         /*
 263          * If we didn't need to bump any existing priorities, and we haven't
 264          * yet submitted this request (i.e. there is no potential race with
 265          * execlists_submit_request()), we can set our own priority and skip
 266          * acquiring the engine locks.
 267          */
 268         if (node->attr.priority == I915_PRIORITY_INVALID) {
 269                 GEM_BUG_ON(!list_empty(&node->link));
 270                 node->attr = *attr;
 271
 272                 if (stack.dfs_link.next == stack.dfs_link.prev)
 273                         return;
 274
 275                 __list_del_entry(&stack.dfs_link);
 276         }
 277
 278         memset(&cache, 0, sizeof(cache));
 279         engine = node_to_request(node)->engine;
 280         spin_lock(&engine->active.lock);
 281
 282         /* Fifo and depth-first replacement ensure our deps execute before us */
 283         engine = sched_lock_engine(node, engine, &cache);
 284         list_for_each_entry_safe_reverse(dep, p, &dfs, dfs_link) {
 285                 INIT_LIST_HEAD(&dep->dfs_link);
 286
 287                 node = dep->signaler;
 288                 engine = sched_lock_engine(node, engine, &cache);
 289                 lockdep_assert_held(&engine->active.lock);
 290
 291                 /* Recheck after acquiring the engine->timeline.lock */
 292                 if (prio <= node->attr.priority || node_signaled(node))
 293                         continue;
 294
 295                 GEM_BUG_ON(node_to_request(node)->engine != engine);
 296
 297                 node->attr.priority = prio;
 298
 299                 if (list_empty(&node->link)) {
 300                         /*
 301                          * If the request is not in the priolist queue because
 302                          * it is not yet runnable, then it doesn't contribute
 303                          * to our preemption decisions. On the other hand,
 304                          * if the request is on the HW, it too is not in the
 305                          * queue; but in that case we may still need to reorder
 306                          * the inflight requests.
 307                          */
 308                         continue;
 309                 }
 310
 311                 if (!intel_engine_is_virtual(engine) &&
 312                     !i915_request_is_active(node_to_request(node))) {
 313                         if (!cache.priolist)
 314                                 cache.priolist =
 315                                         i915_sched_lookup_priolist(engine,
 316                                                                    prio);
 317                         list_move_tail(&node->link, cache.priolist);
 318                 }
 319
 320                 if (prio <= engine->execlists.queue_priority_hint)
 321                         continue;
 322
 323                 engine->execlists.queue_priority_hint = prio;
 324
 325                 /* Defer (tasklet) submission until after all of our updates. */
 326                 kick_submission(engine, prio);
 327         }
 328
 329         spin_unlock(&engine->active.lock);
 330 }
 331
 332 void i915_schedule(struct i915_request *rq, const struct i915_sched_attr *attr)
 333 {
 334         spin_lock_irq(&schedule_lock);
 335         __i915_schedule(&rq->sched, attr);
 336         spin_unlock_irq(&schedule_lock);
 337 }
 338
 339 static void __bump_priority(struct i915_sched_node *node, unsigned int bump)
 340 {
 341         struct i915_sched_attr attr = node->attr;
 342
 343         attr.priority |= bump;
 344         __i915_schedule(node, &attr);
 345 }
 346
 347 void i915_schedule_bump_priority(struct i915_request *rq, unsigned int bump)
 348 {
 349         unsigned long flags;
 350
 351         GEM_BUG_ON(bump & ~I915_PRIORITY_MASK);
 352
 353         if (READ_ONCE(rq->sched.attr.priority) == I915_PRIORITY_INVALID)
 354                 return;
 355
 356         spin_lock_irqsave(&schedule_lock, flags);
 357         __bump_priority(&rq->sched, bump);
 358         spin_unlock_irqrestore(&schedule_lock, flags);
 359 }
 360
 361 void i915_sched_node_init(struct i915_sched_node *node)
 362 {
 363         INIT_LIST_HEAD(&node->signalers_list);
 364         INIT_LIST_HEAD(&node->waiters_list);
 365         INIT_LIST_HEAD(&node->link);
 366         node->attr.priority = I915_PRIORITY_INVALID;
 367         node->semaphores = 0;
 368         node->flags = 0;
 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                 list_add(&dep->wait_link, &signal->waiters_list);
 395                 list_add(&dep->signal_link, &node->signalers_list);
 396                 dep->signaler = signal;
 397                 dep->waiter = node;
 398                 dep->flags = flags;
 399
 400                 /* Keep track of whether anyone on this chain has a semaphore */
 401                 if (signal->flags & I915_SCHED_HAS_SEMAPHORE_CHAIN &&
 402                     !node_started(signal))
 403                         node->flags |= I915_SCHED_HAS_SEMAPHORE_CHAIN;
 404
 405                 /*
 406                  * As we do not allow WAIT to preempt inflight requests,
 407                  * once we have executed a request, along with triggering
 408                  * any execution callbacks, we must preserve its ordering
 409                  * within the non-preemptible FIFO.
 410                  */
 411                 BUILD_BUG_ON(__NO_PREEMPTION & ~I915_PRIORITY_MASK);
 412                 if (flags & I915_DEPENDENCY_EXTERNAL)
 413                         __bump_priority(signal, __NO_PREEMPTION);
 414
 415                 ret = true;
 416         }
 417
 418         spin_unlock_irq(&schedule_lock);
 419
 420         return ret;
 421 }
 422
 423 int i915_sched_node_add_dependency(struct i915_sched_node *node,
 424                                    struct i915_sched_node *signal)
 425 {
 426         struct i915_dependency *dep;
 427
 428         dep = i915_dependency_alloc();
 429         if (!dep)
 430                 return -ENOMEM;
 431
 432         if (!__i915_sched_node_add_dependency(node, signal, dep,
 433                                               I915_DEPENDENCY_EXTERNAL |
 434                                               I915_DEPENDENCY_ALLOC))
 435                 i915_dependency_free(dep);
 436
 437         return 0;
 438 }
 439
 440 void i915_sched_node_fini(struct i915_sched_node *node)
 441 {
 442         struct i915_dependency *dep, *tmp;
 443
 444         spin_lock_irq(&schedule_lock);
 445
 446         /*
 447          * Everyone we depended upon (the fences we wait to be signaled)
 448          * should retire before us and remove themselves from our list.
 449          * However, retirement is run independently on each timeline and
 450          * so we may be called out-of-order.
 451          */
 452         list_for_each_entry_safe(dep, tmp, &node->signalers_list, signal_link) {
 453                 GEM_BUG_ON(!node_signaled(dep->signaler));
 454                 GEM_BUG_ON(!list_empty(&dep->dfs_link));
 455
 456                 list_del(&dep->wait_link);
 457                 if (dep->flags & I915_DEPENDENCY_ALLOC)
 458                         i915_dependency_free(dep);
 459         }
 460
 461         /* Remove ourselves from everyone who depends upon us */
 462         list_for_each_entry_safe(dep, tmp, &node->waiters_list, wait_link) {
 463                 GEM_BUG_ON(dep->signaler != node);
 464                 GEM_BUG_ON(!list_empty(&dep->dfs_link));
 465
 466                 list_del(&dep->signal_link);
 467                 if (dep->flags & I915_DEPENDENCY_ALLOC)
 468                         i915_dependency_free(dep);
 469         }
 470
 471         spin_unlock_irq(&schedule_lock);
 472 }
 473
 474 static void i915_global_scheduler_shrink(void)
 475 {
 476         kmem_cache_shrink(global.slab_dependencies);
 477         kmem_cache_shrink(global.slab_priorities);
 478 }
 479
 480 static void i915_global_scheduler_exit(void)
 481 {
 482         kmem_cache_destroy(global.slab_dependencies);
 483         kmem_cache_destroy(global.slab_priorities);
 484 }
 485
 486 static struct i915_global_scheduler global = { {
 487         .shrink = i915_global_scheduler_shrink,
 488         .exit = i915_global_scheduler_exit,
 489 } };
 490
 491 int __init i915_global_scheduler_init(void)
 492 {
 493         global.slab_dependencies = KMEM_CACHE(i915_dependency,
 494                                               SLAB_HWCACHE_ALIGN);
 495         if (!global.slab_dependencies)
 496                 return -ENOMEM;
 497
 498         global.slab_priorities = KMEM_CACHE(i915_priolist,
 499                                             SLAB_HWCACHE_ALIGN);
 500         if (!global.slab_priorities)
 501                 goto err_priorities;
 502
 503         i915_global_register(&global.base);
 504         return 0;
 505
 506 err_priorities:
 507         kmem_cache_destroy(global.slab_priorities);
 508         return -ENOMEM;
 509 }