| File: | dev/pci/drm/i915/i915_scheduler.c |
| Warning: | line 105, column 4 Value stored to 'first' is never read |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 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 | #ifdef __linux__ |
| 17 | struct kmem_cache *slab_dependencies; |
| 18 | struct kmem_cache *slab_priorities; |
| 19 | #else |
| 20 | struct pool slab_dependencies; |
| 21 | struct pool slab_priorities; |
| 22 | #endif |
| 23 | } global; |
| 24 | |
| 25 | static DEFINE_SPINLOCK(schedule_lock)struct mutex schedule_lock = { ((void *)0), ((((0x9)) > 0x0 && ((0x9)) < 0x9) ? 0x9 : ((0x9))), 0x0 }; |
| 26 | |
| 27 | static const struct i915_request * |
| 28 | node_to_request(const struct i915_sched_node *node) |
| 29 | { |
| 30 | return container_of(node, const struct i915_request, sched)({ const __typeof( ((const struct i915_request *)0)->sched ) *__mptr = (node); (const struct i915_request *)( (char *)__mptr - __builtin_offsetof(const struct i915_request, sched) );}); |
| 31 | } |
| 32 | |
| 33 | static inline bool_Bool node_started(const struct i915_sched_node *node) |
| 34 | { |
| 35 | return i915_request_started(node_to_request(node)); |
| 36 | } |
| 37 | |
| 38 | static inline bool_Bool node_signaled(const struct i915_sched_node *node) |
| 39 | { |
| 40 | return i915_request_completed(node_to_request(node)); |
| 41 | } |
| 42 | |
| 43 | static inline struct i915_priolist *to_priolist(struct rb_node *rb) |
| 44 | { |
| 45 | return rb_entry(rb, struct i915_priolist, node)({ const __typeof( ((struct i915_priolist *)0)->node ) *__mptr = (rb); (struct i915_priolist *)( (char *)__mptr - __builtin_offsetof (struct i915_priolist, node) );}); |
| 46 | } |
| 47 | |
| 48 | static void assert_priolists(struct intel_engine_execlists * const execlists) |
| 49 | { |
| 50 | struct rb_node *rb; |
| 51 | long last_prio, i; |
| 52 | |
| 53 | if (!IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM)0) |
| 54 | return; |
| 55 | |
| 56 | GEM_BUG_ON(rb_first_cached(&execlists->queue) !=((void)0) |
| 57 | rb_first(&execlists->queue.rb_root))((void)0); |
| 58 | |
| 59 | last_prio = INT_MAX0x7fffffff; |
| 60 | for (rb = rb_first_cached(&execlists->queue)linux_root_RB_MINMAX((struct linux_root *)(&(&execlists ->queue)->rb_root), -1); rb; rb = rb_next(rb)linux_root_RB_NEXT((rb))) { |
| 61 | const struct i915_priolist *p = to_priolist(rb); |
| 62 | |
| 63 | GEM_BUG_ON(p->priority > last_prio)((void)0); |
| 64 | last_prio = p->priority; |
| 65 | |
| 66 | GEM_BUG_ON(!p->used)((void)0); |
| 67 | for (i = 0; i < ARRAY_SIZE(p->requests)(sizeof((p->requests)) / sizeof((p->requests)[0])); i++) { |
| 68 | if (list_empty(&p->requests[i])) |
| 69 | continue; |
| 70 | |
| 71 | GEM_BUG_ON(!(p->used & BIT(i)))((void)0); |
| 72 | } |
| 73 | } |
| 74 | } |
| 75 | |
| 76 | struct list_head * |
| 77 | i915_sched_lookup_priolist(struct intel_engine_cs *engine, int prio) |
| 78 | { |
| 79 | struct intel_engine_execlists * const execlists = &engine->execlists; |
| 80 | struct i915_priolist *p; |
| 81 | struct rb_node **parent, *rb; |
| 82 | bool_Bool first = true1; |
| 83 | int idx, i; |
| 84 | |
| 85 | lockdep_assert_held(&engine->active.lock)do { (void)(&engine->active.lock); } while(0); |
| 86 | assert_priolists(execlists); |
| 87 | |
| 88 | /* buckets sorted from highest [in slot 0] to lowest priority */ |
| 89 | idx = I915_PRIORITY_COUNT(1UL << (0)) - (prio & I915_PRIORITY_MASK((1UL << (0)) - 1)) - 1; |
| 90 | prio >>= I915_USER_PRIORITY_SHIFT0; |
| 91 | if (unlikely(execlists->no_priolist)__builtin_expect(!!(execlists->no_priolist), 0)) |
| 92 | prio = I915_PRIORITY_NORMAL; |
| 93 | |
| 94 | find_priolist: |
| 95 | /* most positive priority is scheduled first, equal priorities fifo */ |
| 96 | rb = NULL((void *)0); |
| 97 | parent = &execlists->queue.rb_root.rb_node; |
| 98 | while (*parent) { |
| 99 | rb = *parent; |
| 100 | p = to_priolist(rb); |
| 101 | if (prio > p->priority) { |
| 102 | parent = &rb->rb_left__entry.rbe_left; |
| 103 | } else if (prio < p->priority) { |
| 104 | parent = &rb->rb_right__entry.rbe_right; |
| 105 | first = false0; |
Value stored to 'first' is never read | |
| 106 | } else { |
| 107 | goto out; |
| 108 | } |
| 109 | } |
| 110 | |
| 111 | if (prio == I915_PRIORITY_NORMAL) { |
| 112 | p = &execlists->default_priolist; |
| 113 | } else { |
| 114 | #ifdef __linux__ |
| 115 | p = kmem_cache_alloc(global.slab_priorities, GFP_ATOMIC0x0002); |
| 116 | #else |
| 117 | p = pool_get(&global.slab_priorities, PR_NOWAIT0x0002); |
| 118 | #endif |
| 119 | /* Convert an allocation failure to a priority bump */ |
| 120 | if (unlikely(!p)__builtin_expect(!!(!p), 0)) { |
| 121 | prio = I915_PRIORITY_NORMAL; /* recurses just once */ |
| 122 | |
| 123 | /* To maintain ordering with all rendering, after an |
| 124 | * allocation failure we have to disable all scheduling. |
| 125 | * Requests will then be executed in fifo, and schedule |
| 126 | * will ensure that dependencies are emitted in fifo. |
| 127 | * There will be still some reordering with existing |
| 128 | * requests, so if userspace lied about their |
| 129 | * dependencies that reordering may be visible. |
| 130 | */ |
| 131 | execlists->no_priolist = true1; |
| 132 | goto find_priolist; |
| 133 | } |
| 134 | } |
| 135 | |
| 136 | p->priority = prio; |
| 137 | for (i = 0; i < ARRAY_SIZE(p->requests)(sizeof((p->requests)) / sizeof((p->requests)[0])); i++) |
| 138 | INIT_LIST_HEAD(&p->requests[i]); |
| 139 | rb_link_node(&p->node, rb, parent); |
| 140 | rb_insert_color_cached(&p->node, &execlists->queue, first)linux_root_RB_INSERT_COLOR((struct linux_root *)(&(&execlists ->queue)->rb_root), (&p->node)); |
| 141 | p->used = 0; |
| 142 | |
| 143 | out: |
| 144 | p->used |= BIT(idx)(1UL << (idx)); |
| 145 | return &p->requests[idx]; |
| 146 | } |
| 147 | |
| 148 | void __i915_priolist_free(struct i915_priolist *p) |
| 149 | { |
| 150 | #ifdef __linux__ |
| 151 | kmem_cache_free(global.slab_priorities, p); |
| 152 | #else |
| 153 | pool_put(&global.slab_priorities, p); |
| 154 | #endif |
| 155 | } |
| 156 | |
| 157 | struct sched_cache { |
| 158 | struct list_head *priolist; |
| 159 | }; |
| 160 | |
| 161 | static struct intel_engine_cs * |
| 162 | sched_lock_engine(const struct i915_sched_node *node, |
| 163 | struct intel_engine_cs *locked, |
| 164 | struct sched_cache *cache) |
| 165 | { |
| 166 | const struct i915_request *rq = node_to_request(node); |
| 167 | struct intel_engine_cs *engine; |
| 168 | |
| 169 | GEM_BUG_ON(!locked)((void)0); |
| 170 | |
| 171 | /* |
| 172 | * Virtual engines complicate acquiring the engine timeline lock, |
| 173 | * as their rq->engine pointer is not stable until under that |
| 174 | * engine lock. The simple ploy we use is to take the lock then |
| 175 | * check that the rq still belongs to the newly locked engine. |
| 176 | */ |
| 177 | while (locked != (engine = READ_ONCE(rq->engine)({ typeof(rq->engine) __tmp = *(volatile typeof(rq->engine ) *)&(rq->engine); membar_datadep_consumer(); __tmp; } ))) { |
| 178 | spin_unlock(&locked->active.lock)mtx_leave(&locked->active.lock); |
| 179 | memset(cache, 0, sizeof(*cache))__builtin_memset((cache), (0), (sizeof(*cache))); |
| 180 | spin_lock(&engine->active.lock)mtx_enter(&engine->active.lock); |
| 181 | locked = engine; |
| 182 | } |
| 183 | |
| 184 | GEM_BUG_ON(locked != engine)((void)0); |
| 185 | return locked; |
| 186 | } |
| 187 | |
| 188 | static inline int rq_prio(const struct i915_request *rq) |
| 189 | { |
| 190 | return rq->sched.attr.priority; |
| 191 | } |
| 192 | |
| 193 | static inline bool_Bool need_preempt(int prio, int active) |
| 194 | { |
| 195 | /* |
| 196 | * Allow preemption of low -> normal -> high, but we do |
| 197 | * not allow low priority tasks to preempt other low priority |
| 198 | * tasks under the impression that latency for low priority |
| 199 | * tasks does not matter (as much as background throughput), |
| 200 | * so kiss. |
| 201 | */ |
| 202 | return prio >= max(I915_PRIORITY_NORMAL, active)(((I915_PRIORITY_NORMAL)>(active))?(I915_PRIORITY_NORMAL): (active)); |
| 203 | } |
| 204 | |
| 205 | static void kick_submission(struct intel_engine_cs *engine, |
| 206 | const struct i915_request *rq, |
| 207 | int prio) |
| 208 | { |
| 209 | const struct i915_request *inflight; |
| 210 | |
| 211 | /* |
| 212 | * We only need to kick the tasklet once for the high priority |
| 213 | * new context we add into the queue. |
| 214 | */ |
| 215 | if (prio <= engine->execlists.queue_priority_hint) |
| 216 | return; |
| 217 | |
| 218 | rcu_read_lock(); |
| 219 | |
| 220 | /* Nothing currently active? We're overdue for a submission! */ |
| 221 | inflight = execlists_active(&engine->execlists); |
| 222 | if (!inflight) |
| 223 | goto unlock; |
| 224 | |
| 225 | /* |
| 226 | * If we are already the currently executing context, don't |
| 227 | * bother evaluating if we should preempt ourselves. |
| 228 | */ |
| 229 | if (inflight->context == rq->context) |
| 230 | goto unlock; |
| 231 | |
| 232 | ENGINE_TRACE(engine,do { const struct intel_engine_cs *e__ __attribute__((__unused__ )) = (engine); do { } while (0); } while (0) |
| 233 | "bumping queue-priority-hint:%d for rq:%llx:%lld, inflight:%llx:%lld prio %d\n",do { const struct intel_engine_cs *e__ __attribute__((__unused__ )) = (engine); do { } while (0); } while (0) |
| 234 | prio,do { const struct intel_engine_cs *e__ __attribute__((__unused__ )) = (engine); do { } while (0); } while (0) |
| 235 | rq->fence.context, rq->fence.seqno,do { const struct intel_engine_cs *e__ __attribute__((__unused__ )) = (engine); do { } while (0); } while (0) |
| 236 | inflight->fence.context, inflight->fence.seqno,do { const struct intel_engine_cs *e__ __attribute__((__unused__ )) = (engine); do { } while (0); } while (0) |
| 237 | inflight->sched.attr.priority)do { const struct intel_engine_cs *e__ __attribute__((__unused__ )) = (engine); do { } while (0); } while (0); |
| 238 | |
| 239 | engine->execlists.queue_priority_hint = prio; |
| 240 | if (need_preempt(prio, rq_prio(inflight))) |
| 241 | tasklet_hi_schedule(&engine->execlists.tasklet); |
| 242 | |
| 243 | unlock: |
| 244 | rcu_read_unlock(); |
| 245 | } |
| 246 | |
| 247 | static void __i915_schedule(struct i915_sched_node *node, |
| 248 | const struct i915_sched_attr *attr) |
| 249 | { |
| 250 | const int prio = max(attr->priority, node->attr.priority)(((attr->priority)>(node->attr.priority))?(attr-> priority):(node->attr.priority)); |
| 251 | struct intel_engine_cs *engine; |
| 252 | struct i915_dependency *dep, *p; |
| 253 | struct i915_dependency stack; |
| 254 | struct sched_cache cache; |
| 255 | DRM_LIST_HEAD(dfs)struct list_head dfs = { &(dfs), &(dfs) }; |
| 256 | |
| 257 | /* Needed in order to use the temporary link inside i915_dependency */ |
| 258 | lockdep_assert_held(&schedule_lock)do { (void)(&schedule_lock); } while(0); |
| 259 | GEM_BUG_ON(prio == I915_PRIORITY_INVALID)((void)0); |
| 260 | |
| 261 | if (node_signaled(node)) |
| 262 | return; |
| 263 | |
| 264 | stack.signaler = node; |
| 265 | list_add(&stack.dfs_link, &dfs); |
| 266 | |
| 267 | /* |
| 268 | * Recursively bump all dependent priorities to match the new request. |
| 269 | * |
| 270 | * A naive approach would be to use recursion: |
| 271 | * static void update_priorities(struct i915_sched_node *node, prio) { |
| 272 | * list_for_each_entry(dep, &node->signalers_list, signal_link) |
| 273 | * update_priorities(dep->signal, prio) |
| 274 | * queue_request(node); |
| 275 | * } |
| 276 | * but that may have unlimited recursion depth and so runs a very |
| 277 | * real risk of overunning the kernel stack. Instead, we build |
| 278 | * a flat list of all dependencies starting with the current request. |
| 279 | * As we walk the list of dependencies, we add all of its dependencies |
| 280 | * to the end of the list (this may include an already visited |
| 281 | * request) and continue to walk onwards onto the new dependencies. The |
| 282 | * end result is a topological list of requests in reverse order, the |
| 283 | * last element in the list is the request we must execute first. |
| 284 | */ |
| 285 | list_for_each_entry(dep, &dfs, dfs_link)for (dep = ({ const __typeof( ((__typeof(*dep) *)0)->dfs_link ) *__mptr = ((&dfs)->next); (__typeof(*dep) *)( (char *)__mptr - __builtin_offsetof(__typeof(*dep), dfs_link) );}) ; &dep->dfs_link != (&dfs); dep = ({ const __typeof ( ((__typeof(*dep) *)0)->dfs_link ) *__mptr = (dep->dfs_link .next); (__typeof(*dep) *)( (char *)__mptr - __builtin_offsetof (__typeof(*dep), dfs_link) );})) { |
| 286 | struct i915_sched_node *node = dep->signaler; |
| 287 | |
| 288 | /* If we are already flying, we know we have no signalers */ |
| 289 | if (node_started(node)) |
| 290 | continue; |
| 291 | |
| 292 | /* |
| 293 | * Within an engine, there can be no cycle, but we may |
| 294 | * refer to the same dependency chain multiple times |
| 295 | * (redundant dependencies are not eliminated) and across |
| 296 | * engines. |
| 297 | */ |
| 298 | list_for_each_entry(p, &node->signalers_list, signal_link)for (p = ({ const __typeof( ((__typeof(*p) *)0)->signal_link ) *__mptr = ((&node->signalers_list)->next); (__typeof (*p) *)( (char *)__mptr - __builtin_offsetof(__typeof(*p), signal_link ) );}); &p->signal_link != (&node->signalers_list ); p = ({ const __typeof( ((__typeof(*p) *)0)->signal_link ) *__mptr = (p->signal_link.next); (__typeof(*p) *)( (char *)__mptr - __builtin_offsetof(__typeof(*p), signal_link) );} )) { |
| 299 | GEM_BUG_ON(p == dep)((void)0); /* no cycles! */ |
| 300 | |
| 301 | if (node_signaled(p->signaler)) |
| 302 | continue; |
| 303 | |
| 304 | if (prio > READ_ONCE(p->signaler->attr.priority)({ typeof(p->signaler->attr.priority) __tmp = *(volatile typeof(p->signaler->attr.priority) *)&(p->signaler ->attr.priority); membar_datadep_consumer(); __tmp; })) |
| 305 | list_move_tail(&p->dfs_link, &dfs); |
| 306 | } |
| 307 | } |
| 308 | |
| 309 | /* |
| 310 | * If we didn't need to bump any existing priorities, and we haven't |
| 311 | * yet submitted this request (i.e. there is no potential race with |
| 312 | * execlists_submit_request()), we can set our own priority and skip |
| 313 | * acquiring the engine locks. |
| 314 | */ |
| 315 | if (node->attr.priority == I915_PRIORITY_INVALID((-0x7fffffff-1) | (u8)((1UL << (0)) - 1))) { |
| 316 | GEM_BUG_ON(!list_empty(&node->link))((void)0); |
| 317 | node->attr = *attr; |
| 318 | |
| 319 | if (stack.dfs_link.next == stack.dfs_link.prev) |
| 320 | return; |
| 321 | |
| 322 | __list_del_entry(&stack.dfs_link)list_del(&stack.dfs_link); |
| 323 | } |
| 324 | |
| 325 | memset(&cache, 0, sizeof(cache))__builtin_memset((&cache), (0), (sizeof(cache))); |
| 326 | engine = node_to_request(node)->engine; |
| 327 | spin_lock(&engine->active.lock)mtx_enter(&engine->active.lock); |
| 328 | |
| 329 | /* Fifo and depth-first replacement ensure our deps execute before us */ |
| 330 | engine = sched_lock_engine(node, engine, &cache); |
| 331 | list_for_each_entry_safe_reverse(dep, p, &dfs, dfs_link)for (dep = ({ const __typeof( ((__typeof(*dep) *)0)->dfs_link ) *__mptr = ((&dfs)->prev); (__typeof(*dep) *)( (char *)__mptr - __builtin_offsetof(__typeof(*dep), dfs_link) );}) , p = ({ const __typeof( ((__typeof(*dep) *)0)->dfs_link ) *__mptr = ((dep)->dfs_link.prev); (__typeof(*dep) *)( (char *)__mptr - __builtin_offsetof(__typeof(*dep), dfs_link) );}) ; &(dep)->dfs_link != (&dfs); dep = p, p = ({ const __typeof( ((__typeof(*p) *)0)->dfs_link ) *__mptr = (p-> dfs_link.prev); (__typeof(*p) *)( (char *)__mptr - __builtin_offsetof (__typeof(*p), dfs_link) );})) { |
| 332 | INIT_LIST_HEAD(&dep->dfs_link); |
| 333 | |
| 334 | node = dep->signaler; |
| 335 | engine = sched_lock_engine(node, engine, &cache); |
| 336 | lockdep_assert_held(&engine->active.lock)do { (void)(&engine->active.lock); } while(0); |
| 337 | |
| 338 | /* Recheck after acquiring the engine->timeline.lock */ |
| 339 | if (prio <= node->attr.priority || node_signaled(node)) |
| 340 | continue; |
| 341 | |
| 342 | GEM_BUG_ON(node_to_request(node)->engine != engine)((void)0); |
| 343 | |
| 344 | WRITE_ONCE(node->attr.priority, prio)({ typeof(node->attr.priority) __tmp = (prio); *(volatile typeof (node->attr.priority) *)&(node->attr.priority) = __tmp ; __tmp; }); |
| 345 | |
| 346 | /* |
| 347 | * Once the request is ready, it will be placed into the |
| 348 | * priority lists and then onto the HW runlist. Before the |
| 349 | * request is ready, it does not contribute to our preemption |
| 350 | * decisions and we can safely ignore it, as it will, and |
| 351 | * any preemption required, be dealt with upon submission. |
| 352 | * See engine->submit_request() |
| 353 | */ |
| 354 | if (list_empty(&node->link)) |
| 355 | continue; |
| 356 | |
| 357 | if (i915_request_in_priority_queue(node_to_request(node))) { |
| 358 | if (!cache.priolist) |
| 359 | cache.priolist = |
| 360 | i915_sched_lookup_priolist(engine, |
| 361 | prio); |
| 362 | list_move_tail(&node->link, cache.priolist); |
| 363 | } |
| 364 | |
| 365 | /* Defer (tasklet) submission until after all of our updates. */ |
| 366 | kick_submission(engine, node_to_request(node), prio); |
| 367 | } |
| 368 | |
| 369 | spin_unlock(&engine->active.lock)mtx_leave(&engine->active.lock); |
| 370 | } |
| 371 | |
| 372 | void i915_schedule(struct i915_request *rq, const struct i915_sched_attr *attr) |
| 373 | { |
| 374 | spin_lock_irq(&schedule_lock)mtx_enter(&schedule_lock); |
| 375 | __i915_schedule(&rq->sched, attr); |
| 376 | spin_unlock_irq(&schedule_lock)mtx_leave(&schedule_lock); |
| 377 | } |
| 378 | |
| 379 | static void __bump_priority(struct i915_sched_node *node, unsigned int bump) |
| 380 | { |
| 381 | struct i915_sched_attr attr = node->attr; |
| 382 | |
| 383 | if (attr.priority & bump) |
| 384 | return; |
| 385 | |
| 386 | attr.priority |= bump; |
| 387 | __i915_schedule(node, &attr); |
| 388 | } |
| 389 | |
| 390 | void i915_schedule_bump_priority(struct i915_request *rq, unsigned int bump) |
| 391 | { |
| 392 | unsigned long flags; |
| 393 | |
| 394 | GEM_BUG_ON(bump & ~I915_PRIORITY_MASK)((void)0); |
| 395 | if (READ_ONCE(rq->sched.attr.priority)({ typeof(rq->sched.attr.priority) __tmp = *(volatile typeof (rq->sched.attr.priority) *)&(rq->sched.attr.priority ); membar_datadep_consumer(); __tmp; }) & bump) |
| 396 | return; |
| 397 | |
| 398 | spin_lock_irqsave(&schedule_lock, flags)do { flags = 0; mtx_enter(&schedule_lock); } while (0); |
| 399 | __bump_priority(&rq->sched, bump); |
| 400 | spin_unlock_irqrestore(&schedule_lock, flags)do { (void)(flags); mtx_leave(&schedule_lock); } while (0 ); |
| 401 | } |
| 402 | |
| 403 | void i915_sched_node_init(struct i915_sched_node *node) |
| 404 | { |
| 405 | INIT_LIST_HEAD(&node->signalers_list); |
| 406 | INIT_LIST_HEAD(&node->waiters_list); |
| 407 | INIT_LIST_HEAD(&node->link); |
| 408 | |
| 409 | i915_sched_node_reinit(node); |
| 410 | } |
| 411 | |
| 412 | void i915_sched_node_reinit(struct i915_sched_node *node) |
| 413 | { |
| 414 | node->attr.priority = I915_PRIORITY_INVALID((-0x7fffffff-1) | (u8)((1UL << (0)) - 1)); |
| 415 | node->semaphores = 0; |
| 416 | node->flags = 0; |
| 417 | |
| 418 | GEM_BUG_ON(!list_empty(&node->signalers_list))((void)0); |
| 419 | GEM_BUG_ON(!list_empty(&node->waiters_list))((void)0); |
| 420 | GEM_BUG_ON(!list_empty(&node->link))((void)0); |
| 421 | } |
| 422 | |
| 423 | static struct i915_dependency * |
| 424 | i915_dependency_alloc(void) |
| 425 | { |
| 426 | #ifdef __linux__ |
| 427 | return kmem_cache_alloc(global.slab_dependencies, GFP_KERNEL(0x0001 | 0x0004)); |
| 428 | #else |
| 429 | return pool_get(&global.slab_dependencies, PR_WAITOK0x0001); |
| 430 | #endif |
| 431 | } |
| 432 | |
| 433 | static void |
| 434 | i915_dependency_free(struct i915_dependency *dep) |
| 435 | { |
| 436 | #ifdef __linux__ |
| 437 | kmem_cache_free(global.slab_dependencies, dep); |
| 438 | #else |
| 439 | pool_put(&global.slab_dependencies, dep); |
| 440 | #endif |
| 441 | } |
| 442 | |
| 443 | bool_Bool __i915_sched_node_add_dependency(struct i915_sched_node *node, |
| 444 | struct i915_sched_node *signal, |
| 445 | struct i915_dependency *dep, |
| 446 | unsigned long flags) |
| 447 | { |
| 448 | bool_Bool ret = false0; |
| 449 | |
| 450 | spin_lock_irq(&schedule_lock)mtx_enter(&schedule_lock); |
| 451 | |
| 452 | if (!node_signaled(signal)) { |
| 453 | INIT_LIST_HEAD(&dep->dfs_link); |
| 454 | dep->signaler = signal; |
| 455 | dep->waiter = node; |
| 456 | dep->flags = flags; |
| 457 | |
| 458 | /* All set, now publish. Beware the lockless walkers. */ |
| 459 | list_add_rcu(&dep->signal_link, &node->signalers_list)list_add(&dep->signal_link, &node->signalers_list ); |
| 460 | list_add_rcu(&dep->wait_link, &signal->waiters_list)list_add(&dep->wait_link, &signal->waiters_list ); |
| 461 | |
| 462 | /* Propagate the chains */ |
| 463 | node->flags |= signal->flags; |
| 464 | ret = true1; |
| 465 | } |
| 466 | |
| 467 | spin_unlock_irq(&schedule_lock)mtx_leave(&schedule_lock); |
| 468 | |
| 469 | return ret; |
| 470 | } |
| 471 | |
| 472 | int i915_sched_node_add_dependency(struct i915_sched_node *node, |
| 473 | struct i915_sched_node *signal, |
| 474 | unsigned long flags) |
| 475 | { |
| 476 | struct i915_dependency *dep; |
| 477 | |
| 478 | dep = i915_dependency_alloc(); |
| 479 | if (!dep) |
| 480 | return -ENOMEM12; |
| 481 | |
| 482 | local_bh_disable(); |
| 483 | |
| 484 | if (!__i915_sched_node_add_dependency(node, signal, dep, |
| 485 | flags | I915_DEPENDENCY_ALLOC(1UL << (0)))) |
| 486 | i915_dependency_free(dep); |
| 487 | |
| 488 | local_bh_enable(); /* kick submission tasklet */ |
| 489 | |
| 490 | return 0; |
| 491 | } |
| 492 | |
| 493 | void i915_sched_node_fini(struct i915_sched_node *node) |
| 494 | { |
| 495 | struct i915_dependency *dep, *tmp; |
| 496 | |
| 497 | spin_lock_irq(&schedule_lock)mtx_enter(&schedule_lock); |
| 498 | |
| 499 | /* |
| 500 | * Everyone we depended upon (the fences we wait to be signaled) |
| 501 | * should retire before us and remove themselves from our list. |
| 502 | * However, retirement is run independently on each timeline and |
| 503 | * so we may be called out-of-order. |
| 504 | */ |
| 505 | list_for_each_entry_safe(dep, tmp, &node->signalers_list, signal_link)for (dep = ({ const __typeof( ((__typeof(*dep) *)0)->signal_link ) *__mptr = ((&node->signalers_list)->next); (__typeof (*dep) *)( (char *)__mptr - __builtin_offsetof(__typeof(*dep) , signal_link) );}), tmp = ({ const __typeof( ((__typeof(*dep ) *)0)->signal_link ) *__mptr = (dep->signal_link.next) ; (__typeof(*dep) *)( (char *)__mptr - __builtin_offsetof(__typeof (*dep), signal_link) );}); &dep->signal_link != (& node->signalers_list); dep = tmp, tmp = ({ const __typeof( ((__typeof(*tmp) *)0)->signal_link ) *__mptr = (tmp->signal_link .next); (__typeof(*tmp) *)( (char *)__mptr - __builtin_offsetof (__typeof(*tmp), signal_link) );})) { |
| 506 | GEM_BUG_ON(!list_empty(&dep->dfs_link))((void)0); |
| 507 | |
| 508 | list_del_rcu(&dep->wait_link)list_del(&dep->wait_link); |
| 509 | if (dep->flags & I915_DEPENDENCY_ALLOC(1UL << (0))) |
| 510 | i915_dependency_free(dep); |
| 511 | } |
| 512 | INIT_LIST_HEAD(&node->signalers_list); |
| 513 | |
| 514 | /* Remove ourselves from everyone who depends upon us */ |
| 515 | list_for_each_entry_safe(dep, tmp, &node->waiters_list, wait_link)for (dep = ({ const __typeof( ((__typeof(*dep) *)0)->wait_link ) *__mptr = ((&node->waiters_list)->next); (__typeof (*dep) *)( (char *)__mptr - __builtin_offsetof(__typeof(*dep) , wait_link) );}), tmp = ({ const __typeof( ((__typeof(*dep) * )0)->wait_link ) *__mptr = (dep->wait_link.next); (__typeof (*dep) *)( (char *)__mptr - __builtin_offsetof(__typeof(*dep) , wait_link) );}); &dep->wait_link != (&node->waiters_list ); dep = tmp, tmp = ({ const __typeof( ((__typeof(*tmp) *)0)-> wait_link ) *__mptr = (tmp->wait_link.next); (__typeof(*tmp ) *)( (char *)__mptr - __builtin_offsetof(__typeof(*tmp), wait_link ) );})) { |
| 516 | GEM_BUG_ON(dep->signaler != node)((void)0); |
| 517 | GEM_BUG_ON(!list_empty(&dep->dfs_link))((void)0); |
| 518 | |
| 519 | list_del_rcu(&dep->signal_link)list_del(&dep->signal_link); |
| 520 | if (dep->flags & I915_DEPENDENCY_ALLOC(1UL << (0))) |
| 521 | i915_dependency_free(dep); |
| 522 | } |
| 523 | INIT_LIST_HEAD(&node->waiters_list); |
| 524 | |
| 525 | spin_unlock_irq(&schedule_lock)mtx_leave(&schedule_lock); |
| 526 | } |
| 527 | |
| 528 | static void i915_global_scheduler_shrink(void) |
| 529 | { |
| 530 | #ifdef notyet |
| 531 | kmem_cache_shrink(global.slab_dependencies); |
| 532 | kmem_cache_shrink(global.slab_priorities); |
| 533 | #endif |
| 534 | } |
| 535 | |
| 536 | static void i915_global_scheduler_exit(void) |
| 537 | { |
| 538 | #ifdef __linux__ |
| 539 | kmem_cache_destroy(global.slab_dependencies); |
| 540 | kmem_cache_destroy(global.slab_priorities); |
| 541 | #else |
| 542 | pool_destroy(&global.slab_dependencies); |
| 543 | pool_destroy(&global.slab_priorities); |
| 544 | #endif |
| 545 | } |
| 546 | |
| 547 | static struct i915_global_scheduler global = { { |
| 548 | .shrink = i915_global_scheduler_shrink, |
| 549 | .exit = i915_global_scheduler_exit, |
| 550 | } }; |
| 551 | |
| 552 | int __init i915_global_scheduler_init(void) |
| 553 | { |
| 554 | #ifdef __linux__ |
| 555 | global.slab_dependencies = KMEM_CACHE(i915_dependency, |
| 556 | SLAB_HWCACHE_ALIGN | |
| 557 | SLAB_TYPESAFE_BY_RCU); |
| 558 | if (!global.slab_dependencies) |
| 559 | return -ENOMEM12; |
| 560 | |
| 561 | global.slab_priorities = KMEM_CACHE(i915_priolist, |
| 562 | SLAB_HWCACHE_ALIGN); |
| 563 | if (!global.slab_priorities) |
| 564 | goto err_priorities; |
| 565 | |
| 566 | i915_global_register(&global.base); |
| 567 | return 0; |
| 568 | |
| 569 | err_priorities: |
| 570 | kmem_cache_destroy(global.slab_priorities); |
| 571 | return -ENOMEM12; |
| 572 | #else |
| 573 | pool_init(&global.slab_dependencies, sizeof(struct i915_dependency), |
| 574 | CACHELINESIZE64, IPL_TTY0x9, 0, "gsdep", NULL((void *)0)); |
| 575 | pool_init(&global.slab_priorities, sizeof(struct i915_priolist), |
| 576 | CACHELINESIZE64, IPL_TTY0x9, 0, "gspri", NULL((void *)0)); |
| 577 | |
| 578 | i915_global_register(&global.base); |
| 579 | return 0; |
| 580 | #endif |
| 581 | } |