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 13 static DEFINE_SPINLOCK(schedule_lock); 14 15 static const struct i915_request * 16 node_to_request(const struct i915_sched_node *node) 17 { 18 return container_of(node, const struct i915_request, sched); 19 } 20 21 static inline bool node_signaled(const struct i915_sched_node *node) 22 { 23 return i915_request_completed(node_to_request(node)); 24 } 25 26 void i915_sched_node_init(struct i915_sched_node *node) 27 { 28 INIT_LIST_HEAD(&node->signalers_list); 29 INIT_LIST_HEAD(&node->waiters_list); 30 INIT_LIST_HEAD(&node->link); 31 node->attr.priority = I915_PRIORITY_INVALID; 32 } 33 34 static struct i915_dependency * 35 i915_dependency_alloc(struct drm_i915_private *i915) 36 { 37 return kmem_cache_alloc(i915->dependencies, GFP_KERNEL); 38 } 39 40 static void 41 i915_dependency_free(struct drm_i915_private *i915, 42 struct i915_dependency *dep) 43 { 44 kmem_cache_free(i915->dependencies, dep); 45 } 46 47 bool __i915_sched_node_add_dependency(struct i915_sched_node *node, 48 struct i915_sched_node *signal, 49 struct i915_dependency *dep, 50 unsigned long flags) 51 { 52 bool ret = false; 53 54 spin_lock(&schedule_lock); 55 56 if (!node_signaled(signal)) { 57 INIT_LIST_HEAD(&dep->dfs_link); 58 list_add(&dep->wait_link, &signal->waiters_list); 59 list_add(&dep->signal_link, &node->signalers_list); 60 dep->signaler = signal; 61 dep->flags = flags; 62 63 ret = true; 64 } 65 66 spin_unlock(&schedule_lock); 67 68 return ret; 69 } 70 71 int i915_sched_node_add_dependency(struct drm_i915_private *i915, 72 struct i915_sched_node *node, 73 struct i915_sched_node *signal) 74 { 75 struct i915_dependency *dep; 76 77 dep = i915_dependency_alloc(i915); 78 if (!dep) 79 return -ENOMEM; 80 81 if (!__i915_sched_node_add_dependency(node, signal, dep, 82 I915_DEPENDENCY_ALLOC)) 83 i915_dependency_free(i915, dep); 84 85 return 0; 86 } 87 88 void i915_sched_node_fini(struct drm_i915_private *i915, 89 struct i915_sched_node *node) 90 { 91 struct i915_dependency *dep, *tmp; 92 93 GEM_BUG_ON(!list_empty(&node->link)); 94 95 spin_lock(&schedule_lock); 96 97 /* 98 * Everyone we depended upon (the fences we wait to be signaled) 99 * should retire before us and remove themselves from our list. 100 * However, retirement is run independently on each timeline and 101 * so we may be called out-of-order. 102 */ 103 list_for_each_entry_safe(dep, tmp, &node->signalers_list, signal_link) { 104 GEM_BUG_ON(!node_signaled(dep->signaler)); 105 GEM_BUG_ON(!list_empty(&dep->dfs_link)); 106 107 list_del(&dep->wait_link); 108 if (dep->flags & I915_DEPENDENCY_ALLOC) 109 i915_dependency_free(i915, dep); 110 } 111 112 /* Remove ourselves from everyone who depends upon us */ 113 list_for_each_entry_safe(dep, tmp, &node->waiters_list, wait_link) { 114 GEM_BUG_ON(dep->signaler != node); 115 GEM_BUG_ON(!list_empty(&dep->dfs_link)); 116 117 list_del(&dep->signal_link); 118 if (dep->flags & I915_DEPENDENCY_ALLOC) 119 i915_dependency_free(i915, dep); 120 } 121 122 spin_unlock(&schedule_lock); 123 } 124 125 static inline struct i915_priolist *to_priolist(struct rb_node *rb) 126 { 127 return rb_entry(rb, struct i915_priolist, node); 128 } 129 130 static void assert_priolists(struct intel_engine_execlists * const execlists, 131 long queue_priority) 132 { 133 struct rb_node *rb; 134 long last_prio, i; 135 136 if (!IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM)) 137 return; 138 139 GEM_BUG_ON(rb_first_cached(&execlists->queue) != 140 rb_first(&execlists->queue.rb_root)); 141 142 last_prio = (queue_priority >> I915_USER_PRIORITY_SHIFT) + 1; 143 for (rb = rb_first_cached(&execlists->queue); rb; rb = rb_next(rb)) { 144 const struct i915_priolist *p = to_priolist(rb); 145 146 GEM_BUG_ON(p->priority >= last_prio); 147 last_prio = p->priority; 148 149 GEM_BUG_ON(!p->used); 150 for (i = 0; i < ARRAY_SIZE(p->requests); i++) { 151 if (list_empty(&p->requests[i])) 152 continue; 153 154 GEM_BUG_ON(!(p->used & BIT(i))); 155 } 156 } 157 } 158 159 struct list_head * 160 i915_sched_lookup_priolist(struct intel_engine_cs *engine, int prio) 161 { 162 struct intel_engine_execlists * const execlists = &engine->execlists; 163 struct i915_priolist *p; 164 struct rb_node **parent, *rb; 165 bool first = true; 166 int idx, i; 167 168 lockdep_assert_held(&engine->timeline.lock); 169 assert_priolists(execlists, INT_MAX); 170 171 /* buckets sorted from highest [in slot 0] to lowest priority */ 172 idx = I915_PRIORITY_COUNT - (prio & I915_PRIORITY_MASK) - 1; 173 prio >>= I915_USER_PRIORITY_SHIFT; 174 if (unlikely(execlists->no_priolist)) 175 prio = I915_PRIORITY_NORMAL; 176 177 find_priolist: 178 /* most positive priority is scheduled first, equal priorities fifo */ 179 rb = NULL; 180 parent = &execlists->queue.rb_root.rb_node; 181 while (*parent) { 182 rb = *parent; 183 p = to_priolist(rb); 184 if (prio > p->priority) { 185 parent = &rb->rb_left; 186 } else if (prio < p->priority) { 187 parent = &rb->rb_right; 188 first = false; 189 } else { 190 goto out; 191 } 192 } 193 194 if (prio == I915_PRIORITY_NORMAL) { 195 p = &execlists->default_priolist; 196 } else { 197 p = kmem_cache_alloc(engine->i915->priorities, GFP_ATOMIC); 198 /* Convert an allocation failure to a priority bump */ 199 if (unlikely(!p)) { 200 prio = I915_PRIORITY_NORMAL; /* recurses just once */ 201 202 /* To maintain ordering with all rendering, after an 203 * allocation failure we have to disable all scheduling. 204 * Requests will then be executed in fifo, and schedule 205 * will ensure that dependencies are emitted in fifo. 206 * There will be still some reordering with existing 207 * requests, so if userspace lied about their 208 * dependencies that reordering may be visible. 209 */ 210 execlists->no_priolist = true; 211 goto find_priolist; 212 } 213 } 214 215 p->priority = prio; 216 for (i = 0; i < ARRAY_SIZE(p->requests); i++) 217 INIT_LIST_HEAD(&p->requests[i]); 218 rb_link_node(&p->node, rb, parent); 219 rb_insert_color_cached(&p->node, &execlists->queue, first); 220 p->used = 0; 221 222 out: 223 p->used |= BIT(idx); 224 return &p->requests[idx]; 225 } 226 227 static struct intel_engine_cs * 228 sched_lock_engine(struct i915_sched_node *node, struct intel_engine_cs *locked) 229 { 230 struct intel_engine_cs *engine = node_to_request(node)->engine; 231 232 GEM_BUG_ON(!locked); 233 234 if (engine != locked) { 235 spin_unlock(&locked->timeline.lock); 236 spin_lock(&engine->timeline.lock); 237 } 238 239 return engine; 240 } 241 242 static void __i915_schedule(struct i915_request *rq, 243 const struct i915_sched_attr *attr) 244 { 245 struct list_head *uninitialized_var(pl); 246 struct intel_engine_cs *engine, *last; 247 struct i915_dependency *dep, *p; 248 struct i915_dependency stack; 249 const int prio = attr->priority; 250 LIST_HEAD(dfs); 251 252 /* Needed in order to use the temporary link inside i915_dependency */ 253 lockdep_assert_held(&schedule_lock); 254 GEM_BUG_ON(prio == I915_PRIORITY_INVALID); 255 256 if (i915_request_completed(rq)) 257 return; 258 259 if (prio <= READ_ONCE(rq->sched.attr.priority)) 260 return; 261 262 stack.signaler = &rq->sched; 263 list_add(&stack.dfs_link, &dfs); 264 265 /* 266 * Recursively bump all dependent priorities to match the new request. 267 * 268 * A naive approach would be to use recursion: 269 * static void update_priorities(struct i915_sched_node *node, prio) { 270 * list_for_each_entry(dep, &node->signalers_list, signal_link) 271 * update_priorities(dep->signal, prio) 272 * queue_request(node); 273 * } 274 * but that may have unlimited recursion depth and so runs a very 275 * real risk of overunning the kernel stack. Instead, we build 276 * a flat list of all dependencies starting with the current request. 277 * As we walk the list of dependencies, we add all of its dependencies 278 * to the end of the list (this may include an already visited 279 * request) and continue to walk onwards onto the new dependencies. The 280 * end result is a topological list of requests in reverse order, the 281 * last element in the list is the request we must execute first. 282 */ 283 list_for_each_entry(dep, &dfs, dfs_link) { 284 struct i915_sched_node *node = dep->signaler; 285 286 /* 287 * Within an engine, there can be no cycle, but we may 288 * refer to the same dependency chain multiple times 289 * (redundant dependencies are not eliminated) and across 290 * engines. 291 */ 292 list_for_each_entry(p, &node->signalers_list, signal_link) { 293 GEM_BUG_ON(p == dep); /* no cycles! */ 294 295 if (node_signaled(p->signaler)) 296 continue; 297 298 GEM_BUG_ON(p->signaler->attr.priority < node->attr.priority); 299 if (prio > READ_ONCE(p->signaler->attr.priority)) 300 list_move_tail(&p->dfs_link, &dfs); 301 } 302 } 303 304 /* 305 * If we didn't need to bump any existing priorities, and we haven't 306 * yet submitted this request (i.e. there is no potential race with 307 * execlists_submit_request()), we can set our own priority and skip 308 * acquiring the engine locks. 309 */ 310 if (rq->sched.attr.priority == I915_PRIORITY_INVALID) { 311 GEM_BUG_ON(!list_empty(&rq->sched.link)); 312 rq->sched.attr = *attr; 313 314 if (stack.dfs_link.next == stack.dfs_link.prev) 315 return; 316 317 __list_del_entry(&stack.dfs_link); 318 } 319 320 last = NULL; 321 engine = rq->engine; 322 spin_lock_irq(&engine->timeline.lock); 323 324 /* Fifo and depth-first replacement ensure our deps execute before us */ 325 list_for_each_entry_safe_reverse(dep, p, &dfs, dfs_link) { 326 struct i915_sched_node *node = dep->signaler; 327 328 INIT_LIST_HEAD(&dep->dfs_link); 329 330 engine = sched_lock_engine(node, engine); 331 332 /* Recheck after acquiring the engine->timeline.lock */ 333 if (prio <= node->attr.priority || node_signaled(node)) 334 continue; 335 336 node->attr.priority = prio; 337 if (!list_empty(&node->link)) { 338 if (last != engine) { 339 pl = i915_sched_lookup_priolist(engine, prio); 340 last = engine; 341 } 342 list_move_tail(&node->link, pl); 343 } else { 344 /* 345 * If the request is not in the priolist queue because 346 * it is not yet runnable, then it doesn't contribute 347 * to our preemption decisions. On the other hand, 348 * if the request is on the HW, it too is not in the 349 * queue; but in that case we may still need to reorder 350 * the inflight requests. 351 */ 352 if (!i915_sw_fence_done(&node_to_request(node)->submit)) 353 continue; 354 } 355 356 if (prio <= engine->execlists.queue_priority) 357 continue; 358 359 /* 360 * If we are already the currently executing context, don't 361 * bother evaluating if we should preempt ourselves. 362 */ 363 if (node_to_request(node)->global_seqno && 364 i915_seqno_passed(port_request(engine->execlists.port)->global_seqno, 365 node_to_request(node)->global_seqno)) 366 continue; 367 368 /* Defer (tasklet) submission until after all of our updates. */ 369 engine->execlists.queue_priority = prio; 370 tasklet_hi_schedule(&engine->execlists.tasklet); 371 } 372 373 spin_unlock_irq(&engine->timeline.lock); 374 } 375 376 void i915_schedule(struct i915_request *rq, const struct i915_sched_attr *attr) 377 { 378 spin_lock(&schedule_lock); 379 __i915_schedule(rq, attr); 380 spin_unlock(&schedule_lock); 381 } 382 383 void i915_schedule_bump_priority(struct i915_request *rq, unsigned int bump) 384 { 385 struct i915_sched_attr attr; 386 387 GEM_BUG_ON(bump & ~I915_PRIORITY_MASK); 388 389 if (READ_ONCE(rq->sched.attr.priority) == I915_PRIORITY_INVALID) 390 return; 391 392 spin_lock_bh(&schedule_lock); 393 394 attr = rq->sched.attr; 395 attr.priority |= bump; 396 __i915_schedule(rq, &attr); 397 398 spin_unlock_bh(&schedule_lock); 399 } 400