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 pool slab_dependencies; 14static struct pool 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(node_to_request(node)); 27} 28 29static inline bool node_signaled(const struct i915_sched_node *node) 30{ 31 return i915_request_completed(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#ifdef __linux__ 93 p = kmem_cache_alloc(slab_priorities, GFP_ATOMIC); 94#else 95 p = pool_get(&slab_priorities, PR_NOWAIT); 96#endif 97 /* Convert an allocation failure to a priority bump */ 98 if (unlikely(!p)) { 99 prio = I915_PRIORITY_NORMAL; /* recurses just once */ 100 101 /* To maintain ordering with all rendering, after an 102 * allocation failure we have to disable all scheduling. 103 * Requests will then be executed in fifo, and schedule 104 * will ensure that dependencies are emitted in fifo. 105 * There will be still some reordering with existing 106 * requests, so if userspace lied about their 107 * dependencies that reordering may be visible. 108 */ 109 sched_engine->no_priolist = true; 110 goto find_priolist; 111 } 112 } 113 114 p->priority = prio; 115 INIT_LIST_HEAD(&p->requests); 116 117 rb_link_node(&p->node, rb, parent); 118 rb_insert_color_cached(&p->node, &sched_engine->queue, first); 119 120 return &p->requests; 121} 122 123void __i915_priolist_free(struct i915_priolist *p) 124{ 125#ifdef __linux__ 126 kmem_cache_free(slab_priorities, p); 127#else 128 pool_put(&slab_priorities, p); 129#endif 130} 131 132struct sched_cache { 133 struct list_head *priolist; 134}; 135 136static struct i915_sched_engine * 137lock_sched_engine(struct i915_sched_node *node, 138 struct i915_sched_engine *locked, 139 struct sched_cache *cache) 140{ 141 const struct i915_request *rq = node_to_request(node); 142 struct i915_sched_engine *sched_engine; 143 144 GEM_BUG_ON(!locked); 145 146 /* 147 * Virtual engines complicate acquiring the engine timeline lock, 148 * as their rq->engine pointer is not stable until under that 149 * engine lock. The simple ploy we use is to take the lock then 150 * check that the rq still belongs to the newly locked engine. 151 */ 152 while (locked != (sched_engine = READ_ONCE(rq->engine)->sched_engine)) { 153 spin_unlock(&locked->lock); 154 memset(cache, 0, sizeof(*cache)); 155 spin_lock(&sched_engine->lock); 156 locked = sched_engine; 157 } 158 159 GEM_BUG_ON(locked != sched_engine); 160 return locked; 161} 162 163static void __i915_schedule(struct i915_sched_node *node, 164 const struct i915_sched_attr *attr) 165{ 166 const int prio = max(attr->priority, node->attr.priority); 167 struct i915_sched_engine *sched_engine; 168 struct i915_dependency *dep, *p; 169 struct i915_dependency stack; 170 struct sched_cache cache; 171 DRM_LIST_HEAD(dfs); 172 173 /* Needed in order to use the temporary link inside i915_dependency */ 174 lockdep_assert_held(&schedule_lock); 175 GEM_BUG_ON(prio == I915_PRIORITY_INVALID); 176 177 if (node_signaled(node)) 178 return; 179 180 stack.signaler = node; 181 list_add(&stack.dfs_link, &dfs); 182 183 /* 184 * Recursively bump all dependent priorities to match the new request. 185 * 186 * A naive approach would be to use recursion: 187 * static void update_priorities(struct i915_sched_node *node, prio) { 188 * list_for_each_entry(dep, &node->signalers_list, signal_link) 189 * update_priorities(dep->signal, prio) 190 * queue_request(node); 191 * } 192 * but that may have unlimited recursion depth and so runs a very 193 * real risk of overunning the kernel stack. Instead, we build 194 * a flat list of all dependencies starting with the current request. 195 * As we walk the list of dependencies, we add all of its dependencies 196 * to the end of the list (this may include an already visited 197 * request) and continue to walk onwards onto the new dependencies. The 198 * end result is a topological list of requests in reverse order, the 199 * last element in the list is the request we must execute first. 200 */ 201 list_for_each_entry(dep, &dfs, dfs_link) { 202 struct i915_sched_node *node = dep->signaler; 203 204 /* If we are already flying, we know we have no signalers */ 205 if (node_started(node)) 206 continue; 207 208 /* 209 * Within an engine, there can be no cycle, but we may 210 * refer to the same dependency chain multiple times 211 * (redundant dependencies are not eliminated) and across 212 * engines. 213 */ 214 list_for_each_entry(p, &node->signalers_list, signal_link) { 215 GEM_BUG_ON(p == dep); /* no cycles! */ 216 217 if (node_signaled(p->signaler)) 218 continue; 219 220 if (prio > READ_ONCE(p->signaler->attr.priority)) 221 list_move_tail(&p->dfs_link, &dfs); 222 } 223 } 224 225 /* 226 * If we didn't need to bump any existing priorities, and we haven't 227 * yet submitted this request (i.e. there is no potential race with 228 * execlists_submit_request()), we can set our own priority and skip 229 * acquiring the engine locks. 230 */ 231 if (node->attr.priority == I915_PRIORITY_INVALID) { 232 GEM_BUG_ON(!list_empty(&node->link)); 233 node->attr = *attr; 234 235 if (stack.dfs_link.next == stack.dfs_link.prev) 236 return; 237 238 __list_del_entry(&stack.dfs_link); 239 } 240 241 memset(&cache, 0, sizeof(cache)); 242 sched_engine = node_to_request(node)->engine->sched_engine; 243 spin_lock(&sched_engine->lock); 244 245 /* Fifo and depth-first replacement ensure our deps execute before us */ 246 sched_engine = lock_sched_engine(node, sched_engine, &cache); 247 list_for_each_entry_safe_reverse(dep, p, &dfs, dfs_link) { 248 struct i915_request *from = container_of(dep->signaler, 249 struct i915_request, 250 sched); 251 INIT_LIST_HEAD(&dep->dfs_link); 252 253 node = dep->signaler; 254 sched_engine = lock_sched_engine(node, sched_engine, &cache); 255 lockdep_assert_held(&sched_engine->lock); 256 257 /* Recheck after acquiring the engine->timeline.lock */ 258 if (prio <= node->attr.priority || node_signaled(node)) 259 continue; 260 261 GEM_BUG_ON(node_to_request(node)->engine->sched_engine != 262 sched_engine); 263 264 /* Must be called before changing the nodes priority */ 265 if (sched_engine->bump_inflight_request_prio) 266 sched_engine->bump_inflight_request_prio(from, prio); 267 268 WRITE_ONCE(node->attr.priority, prio); 269 270 /* 271 * Once the request is ready, it will be placed into the 272 * priority lists and then onto the HW runlist. Before the 273 * request is ready, it does not contribute to our preemption 274 * decisions and we can safely ignore it, as it will, and 275 * any preemption required, be dealt with upon submission. 276 * See engine->submit_request() 277 */ 278 if (list_empty(&node->link)) 279 continue; 280 281 if (i915_request_in_priority_queue(node_to_request(node))) { 282 if (!cache.priolist) 283 cache.priolist = 284 i915_sched_lookup_priolist(sched_engine, 285 prio); 286 list_move_tail(&node->link, cache.priolist); 287 } 288 289 /* Defer (tasklet) submission until after all of our updates. */ 290 if (sched_engine->kick_backend) 291 sched_engine->kick_backend(node_to_request(node), prio); 292 } 293 294 spin_unlock(&sched_engine->lock); 295} 296 297void i915_schedule(struct i915_request *rq, const struct i915_sched_attr *attr) 298{ 299 spin_lock_irq(&schedule_lock); 300 __i915_schedule(&rq->sched, attr); 301 spin_unlock_irq(&schedule_lock); 302} 303 304void i915_sched_node_init(struct i915_sched_node *node) 305{ 306 INIT_LIST_HEAD(&node->signalers_list); 307 INIT_LIST_HEAD(&node->waiters_list); 308 INIT_LIST_HEAD(&node->link); 309 310 i915_sched_node_reinit(node); 311} 312 313void i915_sched_node_reinit(struct i915_sched_node *node) 314{ 315 node->attr.priority = I915_PRIORITY_INVALID; 316 node->semaphores = 0; 317 node->flags = 0; 318 319 GEM_BUG_ON(!list_empty(&node->signalers_list)); 320 GEM_BUG_ON(!list_empty(&node->waiters_list)); 321 GEM_BUG_ON(!list_empty(&node->link)); 322} 323 324static struct i915_dependency * 325i915_dependency_alloc(void) 326{ 327#ifdef __linux__ 328 return kmem_cache_alloc(slab_dependencies, GFP_KERNEL); 329#else 330 return pool_get(&slab_dependencies, PR_WAITOK); 331#endif 332} 333 334static void 335i915_dependency_free(struct i915_dependency *dep) 336{ 337#ifdef __linux__ 338 kmem_cache_free(slab_dependencies, dep); 339#else 340 pool_put(&slab_dependencies, dep); 341#endif 342} 343 344bool __i915_sched_node_add_dependency(struct i915_sched_node *node, 345 struct i915_sched_node *signal, 346 struct i915_dependency *dep, 347 unsigned long flags) 348{ 349 bool ret = false; 350 351 spin_lock_irq(&schedule_lock); 352 353 if (!node_signaled(signal)) { 354 INIT_LIST_HEAD(&dep->dfs_link); 355 dep->signaler = signal; 356 dep->waiter = node; 357 dep->flags = flags; 358 359 /* All set, now publish. Beware the lockless walkers. */ 360 list_add_rcu(&dep->signal_link, &node->signalers_list); 361 list_add_rcu(&dep->wait_link, &signal->waiters_list); 362 363 /* Propagate the chains */ 364 node->flags |= signal->flags; 365 ret = true; 366 } 367 368 spin_unlock_irq(&schedule_lock); 369 370 return ret; 371} 372 373int i915_sched_node_add_dependency(struct i915_sched_node *node, 374 struct i915_sched_node *signal, 375 unsigned long flags) 376{ 377 struct i915_dependency *dep; 378 379 dep = i915_dependency_alloc(); 380 if (!dep) 381 return -ENOMEM; 382 383 if (!__i915_sched_node_add_dependency(node, signal, dep, 384 flags | I915_DEPENDENCY_ALLOC)) 385 i915_dependency_free(dep); 386 387 return 0; 388} 389 390void i915_sched_node_fini(struct i915_sched_node *node) 391{ 392 struct i915_dependency *dep, *tmp; 393 394 spin_lock_irq(&schedule_lock); 395 396 /* 397 * Everyone we depended upon (the fences we wait to be signaled) 398 * should retire before us and remove themselves from our list. 399 * However, retirement is run independently on each timeline and 400 * so we may be called out-of-order. 401 */ 402 list_for_each_entry_safe(dep, tmp, &node->signalers_list, signal_link) { 403 GEM_BUG_ON(!list_empty(&dep->dfs_link)); 404 405 list_del_rcu(&dep->wait_link); 406 if (dep->flags & I915_DEPENDENCY_ALLOC) 407 i915_dependency_free(dep); 408 } 409 INIT_LIST_HEAD(&node->signalers_list); 410 411 /* Remove ourselves from everyone who depends upon us */ 412 list_for_each_entry_safe(dep, tmp, &node->waiters_list, wait_link) { 413 GEM_BUG_ON(dep->signaler != node); 414 GEM_BUG_ON(!list_empty(&dep->dfs_link)); 415 416 list_del_rcu(&dep->signal_link); 417 if (dep->flags & I915_DEPENDENCY_ALLOC) 418 i915_dependency_free(dep); 419 } 420 INIT_LIST_HEAD(&node->waiters_list); 421 422 spin_unlock_irq(&schedule_lock); 423} 424 425void i915_request_show_with_schedule(struct drm_printer *m, 426 const struct i915_request *rq, 427 const char *prefix, 428 int indent) 429{ 430 struct i915_dependency *dep; 431 432 i915_request_show(m, rq, prefix, indent); 433 if (i915_request_completed(rq)) 434 return; 435 436 rcu_read_lock(); 437 for_each_signaler(dep, rq) { 438 const struct i915_request *signaler = 439 node_to_request(dep->signaler); 440 441 /* Dependencies along the same timeline are expected. */ 442 if (signaler->timeline == rq->timeline) 443 continue; 444 445 if (__i915_request_is_complete(signaler)) 446 continue; 447 448 i915_request_show(m, signaler, prefix, indent + 2); 449 } 450 rcu_read_unlock(); 451} 452 453static void default_destroy(struct kref *kref) 454{ 455 struct i915_sched_engine *sched_engine = 456 container_of(kref, typeof(*sched_engine), ref); 457 458 tasklet_kill(&sched_engine->tasklet); /* flush the callback */ 459 kfree(sched_engine); 460} 461 462static bool default_disabled(struct i915_sched_engine *sched_engine) 463{ 464 return false; 465} 466 467struct i915_sched_engine * 468i915_sched_engine_create(unsigned int subclass) 469{ 470 struct i915_sched_engine *sched_engine; 471 472 sched_engine = kzalloc(sizeof(*sched_engine), GFP_KERNEL); 473 if (!sched_engine) 474 return NULL; 475 476 kref_init(&sched_engine->ref); 477 478 sched_engine->queue = RB_ROOT_CACHED; 479 sched_engine->queue_priority_hint = INT_MIN; 480 sched_engine->destroy = default_destroy; 481 sched_engine->disabled = default_disabled; 482 483 INIT_LIST_HEAD(&sched_engine->requests); 484 INIT_LIST_HEAD(&sched_engine->hold); 485 486 mtx_init(&sched_engine->lock, IPL_TTY); 487 lockdep_set_subclass(&sched_engine->lock, subclass); 488 489 /* 490 * Due to an interesting quirk in lockdep's internal debug tracking, 491 * after setting a subclass we must ensure the lock is used. Otherwise, 492 * nr_unused_locks is incremented once too often. 493 */ 494#ifdef CONFIG_DEBUG_LOCK_ALLOC 495 local_irq_disable(); 496 lock_map_acquire(&sched_engine->lock.dep_map); 497 lock_map_release(&sched_engine->lock.dep_map); 498 local_irq_enable(); 499#endif 500 501 return sched_engine; 502} 503 504void i915_scheduler_module_exit(void) 505{ 506#ifdef __linux__ 507 kmem_cache_destroy(slab_dependencies); 508 kmem_cache_destroy(slab_priorities); 509#else 510 pool_destroy(&slab_dependencies); 511 pool_destroy(&slab_priorities); 512#endif 513} 514 515int __init i915_scheduler_module_init(void) 516{ 517#ifdef __linux__ 518 slab_dependencies = KMEM_CACHE(i915_dependency, 519 SLAB_HWCACHE_ALIGN | 520 SLAB_TYPESAFE_BY_RCU); 521 if (!slab_dependencies) 522 return -ENOMEM; 523 524 slab_priorities = KMEM_CACHE(i915_priolist, 0); 525 if (!slab_priorities) 526 goto err_priorities; 527 528 return 0; 529 530err_priorities: 531 kmem_cache_destroy(slab_priorities); 532 return -ENOMEM; 533#else 534 pool_init(&slab_dependencies, sizeof(struct i915_dependency), 535 CACHELINESIZE, IPL_TTY, 0, "gsdep", NULL); 536 pool_init(&slab_priorities, sizeof(struct i915_priolist), 537 CACHELINESIZE, IPL_TTY, 0, "gspri", NULL); 538 539 return 0; 540#endif 541} 542