| 1 | |
| 2 | * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR
|
| 3 | * policies)
|
| 4 |
|
| 5 |
|
| 6 | #include "sched.h"
|
| 7 |
|
| 8 | #include <linux/slab.h>
|
| 9 |
|
| 10 | static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
|
| 11 |
|
| 12 | struct rt_bandwidth def_rt_bandwidth;
|
| 13 |
|
| 14 | static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
|
| 15 | {
|
| 16 | struct rt_bandwidth *rt_b =
|
| 17 | container_of(timer, struct rt_bandwidth, rt_period_timer);
|
| 18 | ktime_t now;
|
| 19 | int overrun;
|
| 20 | int idle = 0;
|
| 21 |
|
| 22 | for (;;) {
|
| 23 | now = hrtimer_cb_get_time(timer);
|
| 24 | overrun = hrtimer_forward(timer, now, rt_b->rt_period);
|
| 25 |
|
| 26 | if (!overrun)
|
| 27 | break;
|
| 28 |
|
| 29 | idle = do_sched_rt_period_timer(rt_b, overrun);
|
| 30 | }
|
| 31 |
|
| 32 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
|
| 33 | }
|
| 34 |
|
| 35 | void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
|
| 36 | {
|
| 37 | rt_b->rt_period = ns_to_ktime(period);
|
| 38 | rt_b->rt_runtime = runtime;
|
| 39 |
|
| 40 | raw_spin_lock_init(&rt_b->rt_runtime_lock);
|
| 41 |
|
| 42 | hrtimer_init(&rt_b->rt_period_timer,
|
| 43 | CLOCK_MONOTONIC, HRTIMER_MODE_REL);
|
| 44 | rt_b->rt_period_timer.function = sched_rt_period_timer;
|
| 45 | }
|
| 46 |
|
| 47 | static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
|
| 48 | {
|
| 49 | if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
|
| 50 | return;
|
| 51 |
|
| 52 | if (hrtimer_active(&rt_b->rt_period_timer))
|
| 53 | return;
|
| 54 |
|
| 55 | raw_spin_lock(&rt_b->rt_runtime_lock);
|
| 56 | start_bandwidth_timer(&rt_b->rt_period_timer, rt_b->rt_period);
|
| 57 | raw_spin_unlock(&rt_b->rt_runtime_lock);
|
| 58 | }
|
| 59 |
|
| 60 | void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
|
| 61 | {
|
| 62 | struct rt_prio_array *array;
|
| 63 | int i;
|
| 64 |
|
| 65 | array = &rt_rq->active;
|
| 66 | for (i = 0; i < MAX_RT_PRIO; i++) {
|
| 67 | INIT_LIST_HEAD(array->queue + i);
|
| 68 | __clear_bit(i, array->bitmap);
|
| 69 | }
|
| 70 |
|
| 71 | __set_bit(MAX_RT_PRIO, array->bitmap);
|
| 72 |
|
| 73 | #if defined CONFIG_SMP
|
| 74 | rt_rq->highest_prio.curr = MAX_RT_PRIO;
|
| 75 | rt_rq->highest_prio.next = MAX_RT_PRIO;
|
| 76 | rt_rq->rt_nr_migratory = 0;
|
| 77 | rt_rq->overloaded = 0;
|
| 78 | plist_head_init(&rt_rq->pushable_tasks);
|
| 79 | #endif
|
| 80 |
|
| 81 | rt_rq->rt_time = 0;
|
| 82 | rt_rq->rt_throttled = 0;
|
| 83 | rt_rq->rt_runtime = 0;
|
| 84 | raw_spin_lock_init(&rt_rq->rt_runtime_lock);
|
| 85 | }
|
| 86 |
|
| 87 | #ifdef CONFIG_RT_GROUP_SCHED
|
| 88 | static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
|
| 89 | {
|
| 90 | hrtimer_cancel(&rt_b->rt_period_timer);
|
| 91 | }
|
| 92 |
|
| 93 | #define rt_entity_is_task(rt_se) (!(rt_se)->my_q)
|
| 94 |
|
| 95 | static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
|
| 96 | {
|
| 97 | #ifdef CONFIG_SCHED_DEBUG
|
| 98 | WARN_ON_ONCE(!rt_entity_is_task(rt_se));
|
| 99 | #endif
|
| 100 | return container_of(rt_se, struct task_struct, rt);
|
| 101 | }
|
| 102 |
|
| 103 | static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
|
| 104 | {
|
| 105 | return rt_rq->rq;
|
| 106 | }
|
| 107 |
|
| 108 | static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
|
| 109 | {
|
| 110 | return rt_se->rt_rq;
|
| 111 | }
|
| 112 |
|
| 113 | void free_rt_sched_group(struct task_group *tg)
|
| 114 | {
|
| 115 | int i;
|
| 116 |
|
| 117 | if (tg->rt_se)
|
| 118 | destroy_rt_bandwidth(&tg->rt_bandwidth);
|
| 119 |
|
| 120 | for_each_possible_cpu(i) {
|
| 121 | if (tg->rt_rq)
|
| 122 | kfree(tg->rt_rq[i]);
|
| 123 | if (tg->rt_se)
|
| 124 | kfree(tg->rt_se[i]);
|
| 125 | }
|
| 126 |
|
| 127 | kfree(tg->rt_rq);
|
| 128 | kfree(tg->rt_se);
|
| 129 | }
|
| 130 |
|
| 131 | void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
|
| 132 | struct sched_rt_entity *rt_se, int cpu,
|
| 133 | struct sched_rt_entity *parent)
|
| 134 | {
|
| 135 | struct rq *rq = cpu_rq(cpu);
|
| 136 |
|
| 137 | rt_rq->highest_prio.curr = MAX_RT_PRIO;
|
| 138 | rt_rq->rt_nr_boosted = 0;
|
| 139 | rt_rq->rq = rq;
|
| 140 | rt_rq->tg = tg;
|
| 141 |
|
| 142 | tg->rt_rq[cpu] = rt_rq;
|
| 143 | tg->rt_se[cpu] = rt_se;
|
| 144 |
|
| 145 | if (!rt_se)
|
| 146 | return;
|
| 147 |
|
| 148 | if (!parent)
|
| 149 | rt_se->rt_rq = &rq->rt;
|
| 150 | else
|
| 151 | rt_se->rt_rq = parent->my_q;
|
| 152 |
|
| 153 | rt_se->my_q = rt_rq;
|
| 154 | rt_se->parent = parent;
|
| 155 | INIT_LIST_HEAD(&rt_se->run_list);
|
| 156 | }
|
| 157 |
|
| 158 | int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
|
| 159 | {
|
| 160 | struct rt_rq *rt_rq;
|
| 161 | struct sched_rt_entity *rt_se;
|
| 162 | int i;
|
| 163 |
|
| 164 | tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
|
| 165 | if (!tg->rt_rq)
|
| 166 | goto err;
|
| 167 | tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
|
| 168 | if (!tg->rt_se)
|
| 169 | goto err;
|
| 170 |
|
| 171 | init_rt_bandwidth(&tg->rt_bandwidth,
|
| 172 | ktime_to_ns(def_rt_bandwidth.rt_period), 0);
|
| 173 |
|
| 174 | for_each_possible_cpu(i) {
|
| 175 | rt_rq = kzalloc_node(sizeof(struct rt_rq),
|
| 176 | GFP_KERNEL, cpu_to_node(i));
|
| 177 | if (!rt_rq)
|
| 178 | goto err;
|
| 179 |
|
| 180 | rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
|
| 181 | GFP_KERNEL, cpu_to_node(i));
|
| 182 | if (!rt_se)
|
| 183 | goto err_free_rq;
|
| 184 |
|
| 185 | init_rt_rq(rt_rq, cpu_rq(i));
|
| 186 | rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
|
| 187 | init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
|
| 188 | }
|
| 189 |
|
| 190 | return 1;
|
| 191 |
|
| 192 | err_free_rq:
|
| 193 | kfree(rt_rq);
|
| 194 | err:
|
| 195 | return 0;
|
| 196 | }
|
| 197 |
|
| 198 | #else
|
| 199 |
|
| 200 | #define rt_entity_is_task(rt_se) (1)
|
| 201 |
|
| 202 | static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
|
| 203 | {
|
| 204 | return container_of(rt_se, struct task_struct, rt);
|
| 205 | }
|
| 206 |
|
| 207 | static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
|
| 208 | {
|
| 209 | return container_of(rt_rq, struct rq, rt);
|
| 210 | }
|
| 211 |
|
| 212 | static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
|
| 213 | {
|
| 214 | struct task_struct *p = rt_task_of(rt_se);
|
| 215 | struct rq *rq = task_rq(p);
|
| 216 |
|
| 217 | return &rq->rt;
|
| 218 | }
|
| 219 |
|
| 220 | void free_rt_sched_group(struct task_group *tg) { }
|
| 221 |
|
| 222 | int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
|
| 223 | {
|
| 224 | return 1;
|
| 225 | }
|
| 226 | #endif
|
| 227 |
|
| 228 | #ifdef CONFIG_SMP
|
| 229 |
|
| 230 | static inline int rt_overloaded(struct rq *rq)
|
| 231 | {
|
| 232 | return atomic_read(&rq->rd->rto_count);
|
| 233 | }
|
| 234 |
|
| 235 | static inline void rt_set_overload(struct rq *rq)
|
| 236 | {
|
| 237 | if (!rq->online)
|
| 238 | return;
|
| 239 |
|
| 240 | cpumask_set_cpu(rq->cpu, rq->rd->rto_mask);
|
| 241 | |
| 242 | * Make sure the mask is visible before we set
|
| 243 | * the overload count. That is checked to determine
|
| 244 | * if we should look at the mask. It would be a shame
|
| 245 | * if we looked at the mask, but the mask was not
|
| 246 | * updated yet.
|
| 247 |
|
| 248 | wmb();
|
| 249 | atomic_inc(&rq->rd->rto_count);
|
| 250 | }
|
| 251 |
|
| 252 | static inline void rt_clear_overload(struct rq *rq)
|
| 253 | {
|
| 254 | if (!rq->online)
|
| 255 | return;
|
| 256 |
|
| 257 |
|
| 258 | atomic_dec(&rq->rd->rto_count);
|
| 259 | cpumask_clear_cpu(rq->cpu, rq->rd->rto_mask);
|
| 260 | }
|
| 261 |
|
| 262 | static void update_rt_migration(struct rt_rq *rt_rq)
|
| 263 | {
|
| 264 | if (rt_rq->rt_nr_migratory && rt_rq->rt_nr_total > 1) {
|
| 265 | if (!rt_rq->overloaded) {
|
| 266 | rt_set_overload(rq_of_rt_rq(rt_rq));
|
| 267 | rt_rq->overloaded = 1;
|
| 268 | }
|
| 269 | } else if (rt_rq->overloaded) {
|
| 270 | rt_clear_overload(rq_of_rt_rq(rt_rq));
|
| 271 | rt_rq->overloaded = 0;
|
| 272 | }
|
| 273 | }
|
| 274 |
|
| 275 | static void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
|
| 276 | {
|
| 277 | if (!rt_entity_is_task(rt_se))
|
| 278 | return;
|
| 279 |
|
| 280 | rt_rq = &rq_of_rt_rq(rt_rq)->rt;
|
| 281 |
|
| 282 | rt_rq->rt_nr_total++;
|
| 283 | if (rt_se->nr_cpus_allowed > 1)
|
| 284 | rt_rq->rt_nr_migratory++;
|
| 285 |
|
| 286 | update_rt_migration(rt_rq);
|
| 287 | }
|
| 288 |
|
| 289 | static void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
|
| 290 | {
|
| 291 | if (!rt_entity_is_task(rt_se))
|
| 292 | return;
|
| 293 |
|
| 294 | rt_rq = &rq_of_rt_rq(rt_rq)->rt;
|
| 295 |
|
| 296 | rt_rq->rt_nr_total--;
|
| 297 | if (rt_se->nr_cpus_allowed > 1)
|
| 298 | rt_rq->rt_nr_migratory--;
|
| 299 |
|
| 300 | update_rt_migration(rt_rq);
|
| 301 | }
|
| 302 |
|
| 303 | static inline int has_pushable_tasks(struct rq *rq)
|
| 304 | {
|
| 305 | return !plist_head_empty(&rq->rt.pushable_tasks);
|
| 306 | }
|
| 307 |
|
| 308 | static void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
|
| 309 | {
|
| 310 | plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
|
| 311 | plist_node_init(&p->pushable_tasks, p->prio);
|
| 312 | plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks);
|
| 313 |
|
| 314 |
|
| 315 | if (p->prio < rq->rt.highest_prio.next)
|
| 316 | rq->rt.highest_prio.next = p->prio;
|
| 317 | }
|
| 318 |
|
| 319 | static void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
|
| 320 | {
|
| 321 | plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
|
| 322 |
|
| 323 |
|
| 324 | if (has_pushable_tasks(rq)) {
|
| 325 | p = plist_first_entry(&rq->rt.pushable_tasks,
|
| 326 | struct task_struct, pushable_tasks);
|
| 327 | rq->rt.highest_prio.next = p->prio;
|
| 328 | } else
|
| 329 | rq->rt.highest_prio.next = MAX_RT_PRIO;
|
| 330 | }
|
| 331 |
|
| 332 | #else
|
| 333 |
|
| 334 | static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
|
| 335 | {
|
| 336 | }
|
| 337 |
|
| 338 | static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
|
| 339 | {
|
| 340 | }
|
| 341 |
|
| 342 | static inline
|
| 343 | void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
|
| 344 | {
|
| 345 | }
|
| 346 |
|
| 347 | static inline
|
| 348 | void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
|
| 349 | {
|
| 350 | }
|
| 351 |
|
| 352 | #endif
|
| 353 |
|
| 354 | static inline int on_rt_rq(struct sched_rt_entity *rt_se)
|
| 355 | {
|
| 356 | return !list_empty(&rt_se->run_list);
|
| 357 | }
|
| 358 |
|
| 359 | #ifdef CONFIG_RT_GROUP_SCHED
|
| 360 |
|
| 361 | static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
|
| 362 | {
|
| 363 | if (!rt_rq->tg)
|
| 364 | return RUNTIME_INF;
|
| 365 |
|
| 366 | return rt_rq->rt_runtime;
|
| 367 | }
|
| 368 |
|
| 369 | static inline u64 sched_rt_period(struct rt_rq *rt_rq)
|
| 370 | {
|
| 371 | return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
|
| 372 | }
|
| 373 |
|
| 374 | typedef struct task_group *rt_rq_iter_t;
|
| 375 |
|
| 376 | static inline struct task_group *next_task_group(struct task_group *tg)
|
| 377 | {
|
| 378 | do {
|
| 379 | tg = list_entry_rcu(tg->list.next,
|
| 380 | typeof(struct task_group), list);
|
| 381 | } while (&tg->list != &task_groups && task_group_is_autogroup(tg));
|
| 382 |
|
| 383 | if (&tg->list == &task_groups)
|
| 384 | tg = NULL;
|
| 385 |
|
| 386 | return tg;
|
| 387 | }
|
| 388 |
|
| 389 | #define for_each_rt_rq(rt_rq, iter, rq) \
|
| 390 | for (iter = container_of(&task_groups, typeof(*iter), list); \
|
| 391 | (iter = next_task_group(iter)) && \
|
| 392 | (rt_rq = iter->rt_rq[cpu_of(rq)]);)
|
| 393 |
|
| 394 | static inline void list_add_leaf_rt_rq(struct rt_rq *rt_rq)
|
| 395 | {
|
| 396 | list_add_rcu(&rt_rq->leaf_rt_rq_list,
|
| 397 | &rq_of_rt_rq(rt_rq)->leaf_rt_rq_list);
|
| 398 | }
|
| 399 |
|
| 400 | static inline void list_del_leaf_rt_rq(struct rt_rq *rt_rq)
|
| 401 | {
|
| 402 | list_del_rcu(&rt_rq->leaf_rt_rq_list);
|
| 403 | }
|
| 404 |
|
| 405 | #define for_each_leaf_rt_rq(rt_rq, rq) \
|
| 406 | list_for_each_entry_rcu(rt_rq, &rq->leaf_rt_rq_list, leaf_rt_rq_list)
|
| 407 |
|
| 408 | #define for_each_sched_rt_entity(rt_se) \
|
| 409 | for (; rt_se; rt_se = rt_se->parent)
|
| 410 |
|
| 411 | static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
|
| 412 | {
|
| 413 | return rt_se->my_q;
|
| 414 | }
|
| 415 |
|
| 416 | static void enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head);
|
| 417 | static void dequeue_rt_entity(struct sched_rt_entity *rt_se);
|
| 418 |
|
| 419 | static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
|
| 420 | {
|
| 421 | struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
|
| 422 | struct sched_rt_entity *rt_se;
|
| 423 |
|
| 424 | int cpu = cpu_of(rq_of_rt_rq(rt_rq));
|
| 425 |
|
| 426 | rt_se = rt_rq->tg->rt_se[cpu];
|
| 427 |
|
| 428 | if (rt_rq->rt_nr_running) {
|
| 429 | if (rt_se && !on_rt_rq(rt_se))
|
| 430 | enqueue_rt_entity(rt_se, false);
|
| 431 | if (rt_rq->highest_prio.curr < curr->prio)
|
| 432 | resched_task(curr);
|
| 433 | }
|
| 434 | }
|
| 435 |
|
| 436 | static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
|
| 437 | {
|
| 438 | struct sched_rt_entity *rt_se;
|
| 439 | int cpu = cpu_of(rq_of_rt_rq(rt_rq));
|
| 440 |
|
| 441 | rt_se = rt_rq->tg->rt_se[cpu];
|
| 442 |
|
| 443 | if (rt_se && on_rt_rq(rt_se))
|
| 444 | dequeue_rt_entity(rt_se);
|
| 445 | }
|
| 446 |
|
| 447 | static inline int rt_rq_throttled(struct rt_rq *rt_rq)
|
| 448 | {
|
| 449 | return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
|
| 450 | }
|
| 451 |
|
| 452 | static int rt_se_boosted(struct sched_rt_entity *rt_se)
|
| 453 | {
|
| 454 | struct rt_rq *rt_rq = group_rt_rq(rt_se);
|
| 455 | struct task_struct *p;
|
| 456 |
|
| 457 | if (rt_rq)
|
| 458 | return !!rt_rq->rt_nr_boosted;
|
| 459 |
|
| 460 | p = rt_task_of(rt_se);
|
| 461 | return p->prio != p->normal_prio;
|
| 462 | }
|
| 463 |
|
| 464 | #ifdef CONFIG_SMP
|
| 465 | static inline const struct cpumask *sched_rt_period_mask(void)
|
| 466 | {
|
| 467 | return cpu_rq(smp_processor_id())->rd->span;
|
| 468 | }
|
| 469 | #else
|
| 470 | static inline const struct cpumask *sched_rt_period_mask(void)
|
| 471 | {
|
| 472 | return cpu_online_mask;
|
| 473 | }
|
| 474 | #endif
|
| 475 |
|
| 476 | static inline
|
| 477 | struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
|
| 478 | {
|
| 479 | return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
|
| 480 | }
|
| 481 |
|
| 482 | static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
|
| 483 | {
|
| 484 | return &rt_rq->tg->rt_bandwidth;
|
| 485 | }
|
| 486 |
|
| 487 | #else
|
| 488 |
|
| 489 | static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
|
| 490 | {
|
| 491 | return rt_rq->rt_runtime;
|
| 492 | }
|
| 493 |
|
| 494 | static inline u64 sched_rt_period(struct rt_rq *rt_rq)
|
| 495 | {
|
| 496 | return ktime_to_ns(def_rt_bandwidth.rt_period);
|
| 497 | }
|
| 498 |
|
| 499 | typedef struct rt_rq *rt_rq_iter_t;
|
| 500 |
|
| 501 | #define for_each_rt_rq(rt_rq, iter, rq) \
|
| 502 | for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
|
| 503 |
|
| 504 | static inline void list_add_leaf_rt_rq(struct rt_rq *rt_rq)
|
| 505 | {
|
| 506 | }
|
| 507 |
|
| 508 | static inline void list_del_leaf_rt_rq(struct rt_rq *rt_rq)
|
| 509 | {
|
| 510 | }
|
| 511 |
|
| 512 | #define for_each_leaf_rt_rq(rt_rq, rq) \
|
| 513 | for (rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
|
| 514 |
|
| 515 | #define for_each_sched_rt_entity(rt_se) \
|
| 516 | for (; rt_se; rt_se = NULL)
|
| 517 |
|
| 518 | static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
|
| 519 | {
|
| 520 | return NULL;
|
| 521 | }
|
| 522 |
|
| 523 | static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
|
| 524 | {
|
| 525 | if (rt_rq->rt_nr_running)
|
| 526 | resched_task(rq_of_rt_rq(rt_rq)->curr);
|
| 527 | }
|
| 528 |
|
| 529 | static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
|
| 530 | {
|
| 531 | }
|
| 532 |
|
| 533 | static inline int rt_rq_throttled(struct rt_rq *rt_rq)
|
| 534 | {
|
| 535 | return rt_rq->rt_throttled;
|
| 536 | }
|
| 537 |
|
| 538 | static inline const struct cpumask *sched_rt_period_mask(void)
|
| 539 | {
|
| 540 | return cpu_online_mask;
|
| 541 | }
|
| 542 |
|
| 543 | static inline
|
| 544 | struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
|
| 545 | {
|
| 546 | return &cpu_rq(cpu)->rt;
|
| 547 | }
|
| 548 |
|
| 549 | static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
|
| 550 | {
|
| 551 | return &def_rt_bandwidth;
|
| 552 | }
|
| 553 |
|
| 554 | #endif
|
| 555 |
|
| 556 | #ifdef CONFIG_SMP
|
| 557 | |
| 558 | * We ran out of runtime, see if we can borrow some from our neighbours.
|
| 559 |
|
| 560 | static int do_balance_runtime(struct rt_rq *rt_rq)
|
| 561 | {
|
| 562 | struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
|
| 563 | struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
|
| 564 | int i, weight, more = 0;
|
| 565 | u64 rt_period;
|
| 566 |
|
| 567 | weight = cpumask_weight(rd->span);
|
| 568 |
|
| 569 | raw_spin_lock(&rt_b->rt_runtime_lock);
|
| 570 | rt_period = ktime_to_ns(rt_b->rt_period);
|
| 571 | for_each_cpu(i, rd->span) {
|
| 572 | struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
|
| 573 | s64 diff;
|
| 574 |
|
| 575 | if (iter == rt_rq)
|
| 576 | continue;
|
| 577 |
|
| 578 | raw_spin_lock(&iter->rt_runtime_lock);
|
| 579 | |
| 580 | * Either all rqs have inf runtime and there's nothing to steal
|
| 581 | * or __disable_runtime() below sets a specific rq to inf to
|
| 582 | * indicate its been disabled and disalow stealing.
|
| 583 |
|
| 584 | if (iter->rt_runtime == RUNTIME_INF)
|
| 585 | goto next;
|
| 586 |
|
| 587 | |
| 588 | * From runqueues with spare time, take 1/n part of their
|
| 589 | * spare time, but no more than our period.
|
| 590 |
|
| 591 | diff = iter->rt_runtime - iter->rt_time;
|
| 592 | if (diff > 0) {
|
| 593 | diff = div_u64((u64)diff, weight);
|
| 594 | if (rt_rq->rt_runtime + diff > rt_period)
|
| 595 | diff = rt_period - rt_rq->rt_runtime;
|
| 596 | iter->rt_runtime -= diff;
|
| 597 | rt_rq->rt_runtime += diff;
|
| 598 | more = 1;
|
| 599 | if (rt_rq->rt_runtime == rt_period) {
|
| 600 | raw_spin_unlock(&iter->rt_runtime_lock);
|
| 601 | break;
|
| 602 | }
|
| 603 | }
|
| 604 | next:
|
| 605 | raw_spin_unlock(&iter->rt_runtime_lock);
|
| 606 | }
|
| 607 | raw_spin_unlock(&rt_b->rt_runtime_lock);
|
| 608 |
|
| 609 | return more;
|
| 610 | }
|
| 611 |
|
| 612 | |
| 613 | * Ensure this RQ takes back all the runtime it lend to its neighbours.
|
| 614 |
|
| 615 | static void __disable_runtime(struct rq *rq)
|
| 616 | {
|
| 617 | struct root_domain *rd = rq->rd;
|
| 618 | rt_rq_iter_t iter;
|
| 619 | struct rt_rq *rt_rq;
|
| 620 |
|
| 621 | if (unlikely(!scheduler_running))
|
| 622 | return;
|
| 623 |
|
| 624 | for_each_rt_rq(rt_rq, iter, rq) {
|
| 625 | struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
|
| 626 | s64 want;
|
| 627 | int i;
|
| 628 |
|
| 629 | raw_spin_lock(&rt_b->rt_runtime_lock);
|
| 630 | raw_spin_lock(&rt_rq->rt_runtime_lock);
|
| 631 | |
| 632 | * Either we're all inf and nobody needs to borrow, or we're
|
| 633 | * already disabled and thus have nothing to do, or we have
|
| 634 | * exactly the right amount of runtime to take out.
|
| 635 |
|
| 636 | if (rt_rq->rt_runtime == RUNTIME_INF ||
|
| 637 | rt_rq->rt_runtime == rt_b->rt_runtime)
|
| 638 | goto balanced;
|
| 639 | raw_spin_unlock(&rt_rq->rt_runtime_lock);
|
| 640 |
|
| 641 | |
| 642 | * Calculate the difference between what we started out with
|
| 643 | * and what we current have, that's the amount of runtime
|
| 644 | * we lend and now have to reclaim.
|
| 645 |
|
| 646 | want = rt_b->rt_runtime - rt_rq->rt_runtime;
|
| 647 |
|
| 648 | |
| 649 | * Greedy reclaim, take back as much as we can.
|
| 650 |
|
| 651 | for_each_cpu(i, rd->span) {
|
| 652 | struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
|
| 653 | s64 diff;
|
| 654 |
|
| 655 | |
| 656 | * Can't reclaim from ourselves or disabled runqueues.
|
| 657 |
|
| 658 | if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
|
| 659 | continue;
|
| 660 |
|
| 661 | raw_spin_lock(&iter->rt_runtime_lock);
|
| 662 | if (want > 0) {
|
| 663 | diff = min_t(s64, iter->rt_runtime, want);
|
| 664 | iter->rt_runtime -= diff;
|
| 665 | want -= diff;
|
| 666 | } else {
|
| 667 | iter->rt_runtime -= want;
|
| 668 | want -= want;
|
| 669 | }
|
| 670 | raw_spin_unlock(&iter->rt_runtime_lock);
|
| 671 |
|
| 672 | if (!want)
|
| 673 | break;
|
| 674 | }
|
| 675 |
|
| 676 | raw_spin_lock(&rt_rq->rt_runtime_lock);
|
| 677 | |
| 678 | * We cannot be left wanting - that would mean some runtime
|
| 679 | * leaked out of the system.
|
| 680 |
|
| 681 | BUG_ON(want);
|
| 682 | balanced:
|
| 683 | |
| 684 | * Disable all the borrow logic by pretending we have inf
|
| 685 | * runtime - in which case borrowing doesn't make sense.
|
| 686 |
|
| 687 | rt_rq->rt_runtime = RUNTIME_INF;
|
| 688 | raw_spin_unlock(&rt_rq->rt_runtime_lock);
|
| 689 | raw_spin_unlock(&rt_b->rt_runtime_lock);
|
| 690 | }
|
| 691 | }
|
| 692 |
|
| 693 | static void disable_runtime(struct rq *rq)
|
| 694 | {
|
| 695 | unsigned long flags;
|
| 696 |
|
| 697 | raw_spin_lock_irqsave(&rq->lock, flags);
|
| 698 | __disable_runtime(rq);
|
| 699 | raw_spin_unlock_irqrestore(&rq->lock, flags);
|
| 700 | }
|
| 701 |
|
| 702 | static void __enable_runtime(struct rq *rq)
|
| 703 | {
|
| 704 | rt_rq_iter_t iter;
|
| 705 | struct rt_rq *rt_rq;
|
| 706 |
|
| 707 | if (unlikely(!scheduler_running))
|
| 708 | return;
|
| 709 |
|
| 710 | |
| 711 | * Reset each runqueue's bandwidth settings
|
| 712 |
|
| 713 | for_each_rt_rq(rt_rq, iter, rq) {
|
| 714 | struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
|
| 715 |
|
| 716 | raw_spin_lock(&rt_b->rt_runtime_lock);
|
| 717 | raw_spin_lock(&rt_rq->rt_runtime_lock);
|
| 718 | rt_rq->rt_runtime = rt_b->rt_runtime;
|
| 719 | rt_rq->rt_time = 0;
|
| 720 | rt_rq->rt_throttled = 0;
|
| 721 | raw_spin_unlock(&rt_rq->rt_runtime_lock);
|
| 722 | raw_spin_unlock(&rt_b->rt_runtime_lock);
|
| 723 | }
|
| 724 | }
|
| 725 |
|
| 726 | static void enable_runtime(struct rq *rq)
|
| 727 | {
|
| 728 | unsigned long flags;
|
| 729 |
|
| 730 | raw_spin_lock_irqsave(&rq->lock, flags);
|
| 731 | __enable_runtime(rq);
|
| 732 | raw_spin_unlock_irqrestore(&rq->lock, flags);
|
| 733 | }
|
| 734 |
|
| 735 | int update_runtime(struct notifier_block *nfb, unsigned long action, void *hcpu)
|
| 736 | {
|
| 737 | int cpu = (int)(long)hcpu;
|
| 738 |
|
| 739 | switch (action) {
|
| 740 | case CPU_DOWN_PREPARE:
|
| 741 | case CPU_DOWN_PREPARE_FROZEN:
|
| 742 | disable_runtime(cpu_rq(cpu));
|
| 743 | return NOTIFY_OK;
|
| 744 |
|
| 745 | case CPU_DOWN_FAILED:
|
| 746 | case CPU_DOWN_FAILED_FROZEN:
|
| 747 | case CPU_ONLINE:
|
| 748 | case CPU_ONLINE_FROZEN:
|
| 749 | enable_runtime(cpu_rq(cpu));
|
| 750 | return NOTIFY_OK;
|
| 751 |
|
| 752 | default:
|
| 753 | return NOTIFY_DONE;
|
| 754 | }
|
| 755 | }
|
| 756 |
|
| 757 | static int balance_runtime(struct rt_rq *rt_rq)
|
| 758 | {
|
| 759 | int more = 0;
|
| 760 |
|
| 761 | if (!sched_feat(RT_RUNTIME_SHARE))
|
| 762 | return more;
|
| 763 |
|
| 764 | if (rt_rq->rt_time > rt_rq->rt_runtime) {
|
| 765 | raw_spin_unlock(&rt_rq->rt_runtime_lock);
|
| 766 | more = do_balance_runtime(rt_rq);
|
| 767 | raw_spin_lock(&rt_rq->rt_runtime_lock);
|
| 768 | }
|
| 769 |
|
| 770 | return more;
|
| 771 | }
|
| 772 | #else
|
| 773 | static inline int balance_runtime(struct rt_rq *rt_rq)
|
| 774 | {
|
| 775 | return 0;
|
| 776 | }
|
| 777 | #endif
|
| 778 |
|
| 779 | static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
|
| 780 | {
|
| 781 | int i, idle = 1, throttled = 0;
|
| 782 | const struct cpumask *span;
|
| 783 |
|
| 784 | span = sched_rt_period_mask();
|
| 785 | for_each_cpu(i, span) {
|
| 786 | int enqueue = 0;
|
| 787 | struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
|
| 788 | struct rq *rq = rq_of_rt_rq(rt_rq);
|
| 789 |
|
| 790 | raw_spin_lock(&rq->lock);
|
| 791 | if (rt_rq->rt_time) {
|
| 792 | u64 runtime;
|
| 793 |
|
| 794 | raw_spin_lock(&rt_rq->rt_runtime_lock);
|
| 795 | if (rt_rq->rt_throttled)
|
| 796 | balance_runtime(rt_rq);
|
| 797 | runtime = rt_rq->rt_runtime;
|
| 798 | rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
|
| 799 | if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
|
| 800 | rt_rq->rt_throttled = 0;
|
| 801 | enqueue = 1;
|
| 802 |
|
| 803 | |
| 804 | * Force a clock update if the CPU was idle,
|
| 805 | * lest wakeup -> unthrottle time accumulate.
|
| 806 |
|
| 807 | if (rt_rq->rt_nr_running && rq->curr == rq->idle)
|
| 808 | rq->skip_clock_update = -1;
|
| 809 | }
|
| 810 | if (rt_rq->rt_time || rt_rq->rt_nr_running)
|
| 811 | idle = 0;
|
| 812 | raw_spin_unlock(&rt_rq->rt_runtime_lock);
|
| 813 | } else if (rt_rq->rt_nr_running) {
|
| 814 | idle = 0;
|
| 815 | if (!rt_rq_throttled(rt_rq))
|
| 816 | enqueue = 1;
|
| 817 | }
|
| 818 | if (rt_rq->rt_throttled)
|
| 819 | throttled = 1;
|
| 820 |
|
| 821 | if (enqueue)
|
| 822 | sched_rt_rq_enqueue(rt_rq);
|
| 823 | raw_spin_unlock(&rq->lock);
|
| 824 | }
|
| 825 |
|
| 826 | if (!throttled && (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF))
|
| 827 | return 1;
|
| 828 |
|
| 829 | return idle;
|
| 830 | }
|
| 831 |
|
| 832 | static inline int rt_se_prio(struct sched_rt_entity *rt_se)
|
| 833 | {
|
| 834 | #ifdef CONFIG_RT_GROUP_SCHED
|
| 835 | struct rt_rq *rt_rq = group_rt_rq(rt_se);
|
| 836 |
|
| 837 | if (rt_rq)
|
| 838 | return rt_rq->highest_prio.curr;
|
| 839 | #endif
|
| 840 |
|
| 841 | return rt_task_of(rt_se)->prio;
|
| 842 | }
|
| 843 |
|
| 844 | static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
|
| 845 | {
|
| 846 | u64 runtime = sched_rt_runtime(rt_rq);
|
| 847 |
|
| 848 | if (rt_rq->rt_throttled)
|
| 849 | return rt_rq_throttled(rt_rq);
|
| 850 |
|
| 851 | if (runtime >= sched_rt_period(rt_rq))
|
| 852 | return 0;
|
| 853 |
|
| 854 | balance_runtime(rt_rq);
|
| 855 | runtime = sched_rt_runtime(rt_rq);
|
| 856 | if (runtime == RUNTIME_INF)
|
| 857 | return 0;
|
| 858 |
|
| 859 | if (rt_rq->rt_time > runtime) {
|
| 860 | struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
|
| 861 |
|
| 862 | |
| 863 | * Don't actually throttle groups that have no runtime assigned
|
| 864 | * but accrue some time due to boosting.
|
| 865 |
|
| 866 | if (likely(rt_b->rt_runtime)) {
|
| 867 | static bool once = false;
|
| 868 |
|
| 869 | rt_rq->rt_throttled = 1;
|
| 870 |
|
| 871 | if (!once) {
|
| 872 | once = true;
|
| 873 | printk_sched("sched: RT throttling activated\n");
|
| 874 | }
|
| 875 | } else {
|
| 876 | |
| 877 | * In case we did anyway, make it go away,
|
| 878 | * replenishment is a joke, since it will replenish us
|
| 879 | * with exactly 0 ns.
|
| 880 |
|
| 881 | rt_rq->rt_time = 0;
|
| 882 | }
|
| 883 |
|
| 884 | if (rt_rq_throttled(rt_rq)) {
|
| 885 | sched_rt_rq_dequeue(rt_rq);
|
| 886 | return 1;
|
| 887 | }
|
| 888 | }
|
| 889 |
|
| 890 | return 0;
|
| 891 | }
|
| 892 |
|
| 893 | |
| 894 | * Update the current task's runtime statistics. Skip current tasks that
|
| 895 | * are not in our scheduling class.
|
| 896 |
|
| 897 | static void update_curr_rt(struct rq *rq)
|
| 898 | {
|
| 899 | struct task_struct *curr = rq->curr;
|
| 900 | struct sched_rt_entity *rt_se = &curr->rt;
|
| 901 | struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
|
| 902 | u64 delta_exec;
|
| 903 |
|
| 904 | if (curr->sched_class != &rt_sched_class)
|
| 905 | return;
|
| 906 |
|
| 907 | delta_exec = rq->clock_task - curr->se.exec_start;
|
| 908 | if (unlikely((s64)delta_exec < 0))
|
| 909 | delta_exec = 0;
|
| 910 |
|
| 911 | schedstat_set(curr->se.statistics.exec_max,
|
| 912 | max(curr->se.statistics.exec_max, delta_exec));
|
| 913 |
|
| 914 | curr->se.sum_exec_runtime += delta_exec;
|
| 915 | account_group_exec_runtime(curr, delta_exec);
|
| 916 |
|
| 917 | curr->se.exec_start = rq->clock_task;
|
| 918 | cpuacct_charge(curr, delta_exec);
|
| 919 |
|
| 920 | sched_rt_avg_update(rq, delta_exec);
|
| 921 |
|
| 922 | if (!rt_bandwidth_enabled())
|
| 923 | return;
|
| 924 |
|
| 925 | for_each_sched_rt_entity(rt_se) {
|
| 926 | rt_rq = rt_rq_of_se(rt_se);
|
| 927 |
|
| 928 | if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
|
| 929 | raw_spin_lock(&rt_rq->rt_runtime_lock);
|
| 930 | rt_rq->rt_time += delta_exec;
|
| 931 | if (sched_rt_runtime_exceeded(rt_rq))
|
| 932 | resched_task(curr);
|
| 933 | raw_spin_unlock(&rt_rq->rt_runtime_lock);
|
| 934 | }
|
| 935 | }
|
| 936 | }
|
| 937 |
|
| 938 | #if defined CONFIG_SMP
|
| 939 |
|
| 940 | static void
|
| 941 | inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
|
| 942 | {
|
| 943 | struct rq *rq = rq_of_rt_rq(rt_rq);
|
| 944 |
|
| 945 | if (rq->online && prio < prev_prio)
|
| 946 | cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
|
| 947 | }
|
| 948 |
|
| 949 | static void
|
| 950 | dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
|
| 951 | {
|
| 952 | struct rq *rq = rq_of_rt_rq(rt_rq);
|
| 953 |
|
| 954 | if (rq->online && rt_rq->highest_prio.curr != prev_prio)
|
| 955 | cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
|
| 956 | }
|
| 957 |
|
| 958 | #else
|
| 959 |
|
| 960 | static inline
|
| 961 | void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
|
| 962 | static inline
|
| 963 | void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
|
| 964 |
|
| 965 | #endif
|
| 966 |
|
| 967 | #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
|
| 968 | static void
|
| 969 | inc_rt_prio(struct rt_rq *rt_rq, int prio)
|
| 970 | {
|
| 971 | int prev_prio = rt_rq->highest_prio.curr;
|
| 972 |
|
| 973 | if (prio < prev_prio)
|
| 974 | rt_rq->highest_prio.curr = prio;
|
| 975 |
|
| 976 | inc_rt_prio_smp(rt_rq, prio, prev_prio);
|
| 977 | }
|
| 978 |
|
| 979 | static void
|
| 980 | dec_rt_prio(struct rt_rq *rt_rq, int prio)
|
| 981 | {
|
| 982 | int prev_prio = rt_rq->highest_prio.curr;
|
| 983 |
|
| 984 | if (rt_rq->rt_nr_running) {
|
| 985 |
|
| 986 | WARN_ON(prio < prev_prio);
|
| 987 |
|
| 988 | |
| 989 | * This may have been our highest task, and therefore
|
| 990 | * we may have some recomputation to do
|
| 991 |
|
| 992 | if (prio == prev_prio) {
|
| 993 | struct rt_prio_array *array = &rt_rq->active;
|
| 994 |
|
| 995 | rt_rq->highest_prio.curr =
|
| 996 | sched_find_first_bit(array->bitmap);
|
| 997 | }
|
| 998 |
|
| 999 | } else
|
| 1000 | rt_rq->highest_prio.curr = MAX_RT_PRIO;
|
| 1001 |
|
| 1002 | dec_rt_prio_smp(rt_rq, prio, prev_prio);
|
| 1003 | }
|
| 1004 |
|
| 1005 | #else
|
| 1006 |
|
| 1007 | static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
|
| 1008 | static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
|
| 1009 |
|
| 1010 | #endif
|
| 1011 |
|
| 1012 | #ifdef CONFIG_RT_GROUP_SCHED
|
| 1013 |
|
| 1014 | static void
|
| 1015 | inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
|
| 1016 | {
|
| 1017 | if (rt_se_boosted(rt_se))
|
| 1018 | rt_rq->rt_nr_boosted++;
|
| 1019 |
|
| 1020 | if (rt_rq->tg)
|
| 1021 | start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
|
| 1022 | }
|
| 1023 |
|
| 1024 | static void
|
| 1025 | dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
|
| 1026 | {
|
| 1027 | if (rt_se_boosted(rt_se))
|
| 1028 | rt_rq->rt_nr_boosted--;
|
| 1029 |
|
| 1030 | WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
|
| 1031 | }
|
| 1032 |
|
| 1033 | #else
|
| 1034 |
|
| 1035 | static void
|
| 1036 | inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
|
| 1037 | {
|
| 1038 | start_rt_bandwidth(&def_rt_bandwidth);
|
| 1039 | }
|
| 1040 |
|
| 1041 | static inline
|
| 1042 | void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
|
| 1043 |
|
| 1044 | #endif
|
| 1045 |
|
| 1046 | static inline
|
| 1047 | void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
|
| 1048 | {
|
| 1049 | int prio = rt_se_prio(rt_se);
|
| 1050 |
|
| 1051 | WARN_ON(!rt_prio(prio));
|
| 1052 | rt_rq->rt_nr_running++;
|
| 1053 |
|
| 1054 | inc_rt_prio(rt_rq, prio);
|
| 1055 | inc_rt_migration(rt_se, rt_rq);
|
| 1056 | inc_rt_group(rt_se, rt_rq);
|
| 1057 | }
|
| 1058 |
|
| 1059 | static inline
|
| 1060 | void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
|
| 1061 | {
|
| 1062 | WARN_ON(!rt_prio(rt_se_prio(rt_se)));
|
| 1063 | WARN_ON(!rt_rq->rt_nr_running);
|
| 1064 | rt_rq->rt_nr_running--;
|
| 1065 |
|
| 1066 | dec_rt_prio(rt_rq, rt_se_prio(rt_se));
|
| 1067 | dec_rt_migration(rt_se, rt_rq);
|
| 1068 | dec_rt_group(rt_se, rt_rq);
|
| 1069 | }
|
| 1070 |
|
| 1071 | static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head)
|
| 1072 | {
|
| 1073 | struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
|
| 1074 | struct rt_prio_array *array = &rt_rq->active;
|
| 1075 | struct rt_rq *group_rq = group_rt_rq(rt_se);
|
| 1076 | struct list_head *queue = array->queue + rt_se_prio(rt_se);
|
| 1077 |
|
| 1078 | |
| 1079 | * Don't enqueue the group if its throttled, or when empty.
|
| 1080 | * The latter is a consequence of the former when a child group
|
| 1081 | * get throttled and the current group doesn't have any other
|
| 1082 | * active members.
|
| 1083 |
|
| 1084 | if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running))
|
| 1085 | return;
|
| 1086 |
|
| 1087 | if (!rt_rq->rt_nr_running)
|
| 1088 | list_add_leaf_rt_rq(rt_rq);
|
| 1089 |
|
| 1090 | if (head)
|
| 1091 | list_add(&rt_se->run_list, queue);
|
| 1092 | else
|
| 1093 | list_add_tail(&rt_se->run_list, queue);
|
| 1094 | __set_bit(rt_se_prio(rt_se), array->bitmap);
|
| 1095 |
|
| 1096 | inc_rt_tasks(rt_se, rt_rq);
|
| 1097 | }
|
| 1098 |
|
| 1099 | static void __dequeue_rt_entity(struct sched_rt_entity *rt_se)
|
| 1100 | {
|
| 1101 | struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
|
| 1102 | struct rt_prio_array *array = &rt_rq->active;
|
| 1103 |
|
| 1104 | list_del_init(&rt_se->run_list);
|
| 1105 | if (list_empty(array->queue + rt_se_prio(rt_se)))
|
| 1106 | __clear_bit(rt_se_prio(rt_se), array->bitmap);
|
| 1107 |
|
| 1108 | dec_rt_tasks(rt_se, rt_rq);
|
| 1109 | if (!rt_rq->rt_nr_running)
|
| 1110 | list_del_leaf_rt_rq(rt_rq);
|
| 1111 | }
|
| 1112 |
|
| 1113 | |
| 1114 | * Because the prio of an upper entry depends on the lower
|
| 1115 | * entries, we must remove entries top - down.
|
| 1116 |
|
| 1117 | static void dequeue_rt_stack(struct sched_rt_entity *rt_se)
|
| 1118 | {
|
| 1119 | struct sched_rt_entity *back = NULL;
|
| 1120 |
|
| 1121 | for_each_sched_rt_entity(rt_se) {
|
| 1122 | rt_se->back = back;
|
| 1123 | back = rt_se;
|
| 1124 | }
|
| 1125 |
|
| 1126 | for (rt_se = back; rt_se; rt_se = rt_se->back) {
|
| 1127 | if (on_rt_rq(rt_se))
|
| 1128 | __dequeue_rt_entity(rt_se);
|
| 1129 | }
|
| 1130 | }
|
| 1131 |
|
| 1132 | static void enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head)
|
| 1133 | {
|
| 1134 | dequeue_rt_stack(rt_se);
|
| 1135 | for_each_sched_rt_entity(rt_se)
|
| 1136 | __enqueue_rt_entity(rt_se, head);
|
| 1137 | }
|
| 1138 |
|
| 1139 | static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
|
| 1140 | {
|
| 1141 | dequeue_rt_stack(rt_se);
|
| 1142 |
|
| 1143 | for_each_sched_rt_entity(rt_se) {
|
| 1144 | struct rt_rq *rt_rq = group_rt_rq(rt_se);
|
| 1145 |
|
| 1146 | if (rt_rq && rt_rq->rt_nr_running)
|
| 1147 | __enqueue_rt_entity(rt_se, false);
|
| 1148 | }
|
| 1149 | }
|
| 1150 |
|
| 1151 | |
| 1152 | * Adding/removing a task to/from a priority array:
|
| 1153 |
|
| 1154 | static void
|
| 1155 | enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags)
|
| 1156 | {
|
| 1157 | struct sched_rt_entity *rt_se = &p->rt;
|
| 1158 |
|
| 1159 | if (flags & ENQUEUE_WAKEUP)
|
| 1160 | rt_se->timeout = 0;
|
| 1161 |
|
| 1162 | enqueue_rt_entity(rt_se, flags & ENQUEUE_HEAD);
|
| 1163 |
|
| 1164 | if (!task_current(rq, p) && p->rt.nr_cpus_allowed > 1)
|
| 1165 | enqueue_pushable_task(rq, p);
|
| 1166 |
|
| 1167 | inc_nr_running(rq);
|
| 1168 | }
|
| 1169 |
|
| 1170 | static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags)
|
| 1171 | {
|
| 1172 | struct sched_rt_entity *rt_se = &p->rt;
|
| 1173 |
|
| 1174 | update_curr_rt(rq);
|
| 1175 | dequeue_rt_entity(rt_se);
|
| 1176 |
|
| 1177 | dequeue_pushable_task(rq, p);
|
| 1178 |
|
| 1179 | dec_nr_running(rq);
|
| 1180 | }
|
| 1181 |
|
| 1182 | |
| 1183 | * Put task to the head or the end of the run list without the overhead of
|
| 1184 | * dequeue followed by enqueue.
|
| 1185 |
|
| 1186 | static void
|
| 1187 | requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
|
| 1188 | {
|
| 1189 | if (on_rt_rq(rt_se)) {
|
| 1190 | struct rt_prio_array *array = &rt_rq->active;
|
| 1191 | struct list_head *queue = array->queue + rt_se_prio(rt_se);
|
| 1192 |
|
| 1193 | if (head)
|
| 1194 | list_move(&rt_se->run_list, queue);
|
| 1195 | else
|
| 1196 | list_move_tail(&rt_se->run_list, queue);
|
| 1197 | }
|
| 1198 | }
|
| 1199 |
|
| 1200 | static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
|
| 1201 | {
|
| 1202 | struct sched_rt_entity *rt_se = &p->rt;
|
| 1203 | struct rt_rq *rt_rq;
|
| 1204 |
|
| 1205 | for_each_sched_rt_entity(rt_se) {
|
| 1206 | rt_rq = rt_rq_of_se(rt_se);
|
| 1207 | requeue_rt_entity(rt_rq, rt_se, head);
|
| 1208 | }
|
| 1209 | }
|
| 1210 |
|
| 1211 | static void yield_task_rt(struct rq *rq)
|
| 1212 | {
|
| 1213 | requeue_task_rt(rq, rq->curr, 0);
|
| 1214 | }
|
| 1215 |
|
| 1216 | #ifdef CONFIG_SMP
|
| 1217 | static int find_lowest_rq(struct task_struct *task);
|
| 1218 |
|
| 1219 | static int
|
| 1220 | select_task_rq_rt(struct task_struct *p, int sd_flag, int flags)
|
| 1221 | {
|
| 1222 | struct task_struct *curr;
|
| 1223 | struct rq *rq;
|
| 1224 | int cpu;
|
| 1225 |
|
| 1226 | cpu = task_cpu(p);
|
| 1227 |
|
| 1228 | if (p->rt.nr_cpus_allowed == 1)
|
| 1229 | goto out;
|
| 1230 |
|
| 1231 |
|
| 1232 | if (sd_flag != SD_BALANCE_WAKE && sd_flag != SD_BALANCE_FORK)
|
| 1233 | goto out;
|
| 1234 |
|
| 1235 | rq = cpu_rq(cpu);
|
| 1236 |
|
| 1237 | rcu_read_lock();
|
| 1238 | curr = ACCESS_ONCE(rq->curr);
|
| 1239 |
|
| 1240 | |
| 1241 | * If the current task on @p's runqueue is an RT task, then
|
| 1242 | * try to see if we can wake this RT task up on another
|
| 1243 | * runqueue. Otherwise simply start this RT task
|
| 1244 | * on its current runqueue.
|
| 1245 | *
|
| 1246 | * We want to avoid overloading runqueues. If the woken
|
| 1247 | * task is a higher priority, then it will stay on this CPU
|
| 1248 | * and the lower prio task should be moved to another CPU.
|
| 1249 | * Even though this will probably make the lower prio task
|
| 1250 | * lose its cache, we do not want to bounce a higher task
|
| 1251 | * around just because it gave up its CPU, perhaps for a
|
| 1252 | * lock?
|
| 1253 | *
|
| 1254 | * For equal prio tasks, we just let the scheduler sort it out.
|
| 1255 | *
|
| 1256 | * Otherwise, just let it ride on the affined RQ and the
|
| 1257 | * post-schedule router will push the preempted task away
|
| 1258 | *
|
| 1259 | * This test is optimistic, if we get it wrong the load-balancer
|
| 1260 | * will have to sort it out.
|
| 1261 |
|
| 1262 | if (curr && unlikely(rt_task(curr)) &&
|
| 1263 | (curr->rt.nr_cpus_allowed < 2 ||
|
| 1264 | curr->prio <= p->prio) &&
|
| 1265 | (p->rt.nr_cpus_allowed > 1)) {
|
| 1266 | int target = find_lowest_rq(p);
|
| 1267 |
|
| 1268 | if (target != -1)
|
| 1269 | cpu = target;
|
| 1270 | }
|
| 1271 | rcu_read_unlock();
|
| 1272 |
|
| 1273 | out:
|
| 1274 | return cpu;
|
| 1275 | }
|
| 1276 |
|
| 1277 | static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
|
| 1278 | {
|
| 1279 | if (rq->curr->rt.nr_cpus_allowed == 1)
|
| 1280 | return;
|
| 1281 |
|
| 1282 | if (p->rt.nr_cpus_allowed != 1
|
| 1283 | && cpupri_find(&rq->rd->cpupri, p, NULL))
|
| 1284 | return;
|
| 1285 |
|
| 1286 | if (!cpupri_find(&rq->rd->cpupri, rq->curr, NULL))
|
| 1287 | return;
|
| 1288 |
|
| 1289 | |
| 1290 | * There appears to be other cpus that can accept
|
| 1291 | * current and none to run 'p', so lets reschedule
|
| 1292 | * to try and push current away:
|
| 1293 |
|
| 1294 | requeue_task_rt(rq, p, 1);
|
| 1295 | resched_task(rq->curr);
|
| 1296 | }
|
| 1297 |
|
| 1298 | #endif
|
| 1299 |
|
| 1300 | |
| 1301 | * Preempt the current task with a newly woken task if needed:
|
| 1302 |
|
| 1303 | static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags)
|
| 1304 | {
|
| 1305 | if (p->prio < rq->curr->prio) {
|
| 1306 | resched_task(rq->curr);
|
| 1307 | return;
|
| 1308 | }
|
| 1309 |
|
| 1310 | #ifdef CONFIG_SMP
|
| 1311 | |
| 1312 | * If:
|
| 1313 | *
|
| 1314 | * - the newly woken task is of equal priority to the current task
|
| 1315 | * - the newly woken task is non-migratable while current is migratable
|
| 1316 | * - current will be preempted on the next reschedule
|
| 1317 | *
|
| 1318 | * we should check to see if current can readily move to a different
|
| 1319 | * cpu. If so, we will reschedule to allow the push logic to try
|
| 1320 | * to move current somewhere else, making room for our non-migratable
|
| 1321 | * task.
|
| 1322 |
|
| 1323 | if (p->prio == rq->curr->prio && !test_tsk_need_resched(rq->curr))
|
| 1324 | check_preempt_equal_prio(rq, p);
|
| 1325 | #endif
|
| 1326 | }
|
| 1327 |
|
| 1328 | static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
|
| 1329 | struct rt_rq *rt_rq)
|
| 1330 | {
|
| 1331 | struct rt_prio_array *array = &rt_rq->active;
|
| 1332 | struct sched_rt_entity *next = NULL;
|
| 1333 | struct list_head *queue;
|
| 1334 | int idx;
|
| 1335 |
|
| 1336 | idx = sched_find_first_bit(array->bitmap);
|
| 1337 | BUG_ON(idx >= MAX_RT_PRIO);
|
| 1338 |
|
| 1339 | queue = array->queue + idx;
|
| 1340 | next = list_entry(queue->next, struct sched_rt_entity, run_list);
|
| 1341 |
|
| 1342 | return next;
|
| 1343 | }
|
| 1344 |
|
| 1345 | static struct task_struct *_pick_next_task_rt(struct rq *rq)
|
| 1346 | {
|
| 1347 | struct sched_rt_entity *rt_se;
|
| 1348 | struct task_struct *p;
|
| 1349 | struct rt_rq *rt_rq;
|
| 1350 |
|
| 1351 | rt_rq = &rq->rt;
|
| 1352 |
|
| 1353 | if (!rt_rq->rt_nr_running)
|
| 1354 | return NULL;
|
| 1355 |
|
| 1356 | if (rt_rq_throttled(rt_rq))
|
| 1357 | return NULL;
|
| 1358 |
|
| 1359 | do {
|
| 1360 | rt_se = pick_next_rt_entity(rq, rt_rq);
|
| 1361 | BUG_ON(!rt_se);
|
| 1362 | rt_rq = group_rt_rq(rt_se);
|
| 1363 | } while (rt_rq);
|
| 1364 |
|
| 1365 | p = rt_task_of(rt_se);
|
| 1366 | p->se.exec_start = rq->clock_task;
|
| 1367 |
|
| 1368 | return p;
|
| 1369 | }
|
| 1370 |
|
| 1371 | static struct task_struct *pick_next_task_rt(struct rq *rq)
|
| 1372 | {
|
| 1373 | struct task_struct *p = _pick_next_task_rt(rq);
|
| 1374 |
|
| 1375 |
|
| 1376 | if (p)
|
| 1377 | dequeue_pushable_task(rq, p);
|
| 1378 |
|
| 1379 | #ifdef CONFIG_SMP
|
| 1380 | |
| 1381 | * We detect this state here so that we can avoid taking the RQ
|
| 1382 | * lock again later if there is no need to push
|
| 1383 |
|
| 1384 | rq->post_schedule = has_pushable_tasks(rq);
|
| 1385 | #endif
|
| 1386 |
|
| 1387 | return p;
|
| 1388 | }
|
| 1389 |
|
| 1390 | static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
|
| 1391 | {
|
| 1392 | update_curr_rt(rq);
|
| 1393 |
|
| 1394 | |
| 1395 | * The previous task needs to be made eligible for pushing
|
| 1396 | * if it is still active
|
| 1397 |
|
| 1398 | if (on_rt_rq(&p->rt) && p->rt.nr_cpus_allowed > 1)
|
| 1399 | enqueue_pushable_task(rq, p);
|
| 1400 | }
|
| 1401 |
|
| 1402 | #ifdef CONFIG_SMP
|
| 1403 |
|
| 1404 |
|
| 1405 | #define RT_MAX_TRIES 3
|
| 1406 |
|
| 1407 | static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
|
| 1408 | {
|
| 1409 | if (!task_running(rq, p) &&
|
| 1410 | (cpu < 0 || cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) &&
|
| 1411 | (p->rt.nr_cpus_allowed > 1))
|
| 1412 | return 1;
|
| 1413 | return 0;
|
| 1414 | }
|
| 1415 |
|
| 1416 |
|
| 1417 | static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu)
|
| 1418 | {
|
| 1419 | struct task_struct *next = NULL;
|
| 1420 | struct sched_rt_entity *rt_se;
|
| 1421 | struct rt_prio_array *array;
|
| 1422 | struct rt_rq *rt_rq;
|
| 1423 | int idx;
|
| 1424 |
|
| 1425 | for_each_leaf_rt_rq(rt_rq, rq) {
|
| 1426 | array = &rt_rq->active;
|
| 1427 | idx = sched_find_first_bit(array->bitmap);
|
| 1428 | next_idx:
|
| 1429 | if (idx >= MAX_RT_PRIO)
|
| 1430 | continue;
|
| 1431 | if (next && next->prio <= idx)
|
| 1432 | continue;
|
| 1433 | list_for_each_entry(rt_se, array->queue + idx, run_list) {
|
| 1434 | struct task_struct *p;
|
| 1435 |
|
| 1436 | if (!rt_entity_is_task(rt_se))
|
| 1437 | continue;
|
| 1438 |
|
| 1439 | p = rt_task_of(rt_se);
|
| 1440 | if (pick_rt_task(rq, p, cpu)) {
|
| 1441 | next = p;
|
| 1442 | break;
|
| 1443 | }
|
| 1444 | }
|
| 1445 | if (!next) {
|
| 1446 | idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
|
| 1447 | goto next_idx;
|
| 1448 | }
|
| 1449 | }
|
| 1450 |
|
| 1451 | return next;
|
| 1452 | }
|
| 1453 |
|
| 1454 | static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
|
| 1455 |
|
| 1456 | static int find_lowest_rq(struct task_struct *task)
|
| 1457 | {
|
| 1458 | struct sched_domain *sd;
|
| 1459 | struct cpumask *lowest_mask = __get_cpu_var(local_cpu_mask);
|
| 1460 | int this_cpu = smp_processor_id();
|
| 1461 | int cpu = task_cpu(task);
|
| 1462 |
|
| 1463 |
|
| 1464 | if (unlikely(!lowest_mask))
|
| 1465 | return -1;
|
| 1466 |
|
| 1467 | if (task->rt.nr_cpus_allowed == 1)
|
| 1468 | return -1;
|
| 1469 |
|
| 1470 | if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
|
| 1471 | return -1;
|
| 1472 |
|
| 1473 | |
| 1474 | * At this point we have built a mask of cpus representing the
|
| 1475 | * lowest priority tasks in the system. Now we want to elect
|
| 1476 | * the best one based on our affinity and topology.
|
| 1477 | *
|
| 1478 | * We prioritize the last cpu that the task executed on since
|
| 1479 | * it is most likely cache-hot in that location.
|
| 1480 |
|
| 1481 | if (cpumask_test_cpu(cpu, lowest_mask))
|
| 1482 | return cpu;
|
| 1483 |
|
| 1484 | |
| 1485 | * Otherwise, we consult the sched_domains span maps to figure
|
| 1486 | * out which cpu is logically closest to our hot cache data.
|
| 1487 |
|
| 1488 | if (!cpumask_test_cpu(this_cpu, lowest_mask))
|
| 1489 | this_cpu = -1;
|
| 1490 |
|
| 1491 | rcu_read_lock();
|
| 1492 | for_each_domain(cpu, sd) {
|
| 1493 | if (sd->flags & SD_WAKE_AFFINE) {
|
| 1494 | int best_cpu;
|
| 1495 |
|
| 1496 | |
| 1497 | * "this_cpu" is cheaper to preempt than a
|
| 1498 | * remote processor.
|
| 1499 |
|
| 1500 | if (this_cpu != -1 &&
|
| 1501 | cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
|
| 1502 | rcu_read_unlock();
|
| 1503 | return this_cpu;
|
| 1504 | }
|
| 1505 |
|
| 1506 | best_cpu = cpumask_first_and(lowest_mask,
|
| 1507 | sched_domain_span(sd));
|
| 1508 | if (best_cpu < nr_cpu_ids) {
|
| 1509 | rcu_read_unlock();
|
| 1510 | return best_cpu;
|
| 1511 | }
|
| 1512 | }
|
| 1513 | }
|
| 1514 | rcu_read_unlock();
|
| 1515 |
|
| 1516 | |
| 1517 | * And finally, if there were no matches within the domains
|
| 1518 | * just give the caller *something* to work with from the compatible
|
| 1519 | * locations.
|
| 1520 |
|
| 1521 | if (this_cpu != -1)
|
| 1522 | return this_cpu;
|
| 1523 |
|
| 1524 | cpu = cpumask_any(lowest_mask);
|
| 1525 | if (cpu < nr_cpu_ids)
|
| 1526 | return cpu;
|
| 1527 | return -1;
|
| 1528 | }
|
| 1529 |
|
| 1530 |
|
| 1531 | static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
|
| 1532 | {
|
| 1533 | struct rq *lowest_rq = NULL;
|
| 1534 | int tries;
|
| 1535 | int cpu;
|
| 1536 |
|
| 1537 | for (tries = 0; tries < RT_MAX_TRIES; tries++) {
|
| 1538 | cpu = find_lowest_rq(task);
|
| 1539 |
|
| 1540 | if ((cpu == -1) || (cpu == rq->cpu))
|
| 1541 | break;
|
| 1542 |
|
| 1543 | lowest_rq = cpu_rq(cpu);
|
| 1544 |
|
| 1545 |
|
| 1546 | if (double_lock_balance(rq, lowest_rq)) {
|
| 1547 | |
| 1548 | * We had to unlock the run queue. In
|
| 1549 | * the mean time, task could have
|
| 1550 | * migrated already or had its affinity changed.
|
| 1551 | * Also make sure that it wasn't scheduled on its rq.
|
| 1552 |
|
| 1553 | if (unlikely(task_rq(task) != rq ||
|
| 1554 | !cpumask_test_cpu(lowest_rq->cpu,
|
| 1555 | tsk_cpus_allowed(task)) ||
|
| 1556 | task_running(rq, task) ||
|
| 1557 | !task->on_rq)) {
|
| 1558 |
|
| 1559 | raw_spin_unlock(&lowest_rq->lock);
|
| 1560 | lowest_rq = NULL;
|
| 1561 | break;
|
| 1562 | }
|
| 1563 | }
|
| 1564 |
|
| 1565 |
|
| 1566 | if (lowest_rq->rt.highest_prio.curr > task->prio)
|
| 1567 | break;
|
| 1568 |
|
| 1569 |
|
| 1570 | double_unlock_balance(rq, lowest_rq);
|
| 1571 | lowest_rq = NULL;
|
| 1572 | }
|
| 1573 |
|
| 1574 | return lowest_rq;
|
| 1575 | }
|
| 1576 |
|
| 1577 | static struct task_struct *pick_next_pushable_task(struct rq *rq)
|
| 1578 | {
|
| 1579 | struct task_struct *p;
|
| 1580 |
|
| 1581 | if (!has_pushable_tasks(rq))
|
| 1582 | return NULL;
|
| 1583 |
|
| 1584 | p = plist_first_entry(&rq->rt.pushable_tasks,
|
| 1585 | struct task_struct, pushable_tasks);
|
| 1586 |
|
| 1587 | BUG_ON(rq->cpu != task_cpu(p));
|
| 1588 | BUG_ON(task_current(rq, p));
|
| 1589 | BUG_ON(p->rt.nr_cpus_allowed <= 1);
|
| 1590 |
|
| 1591 | BUG_ON(!p->on_rq);
|
| 1592 | BUG_ON(!rt_task(p));
|
| 1593 |
|
| 1594 | return p;
|
| 1595 | }
|
| 1596 |
|
| 1597 | |
| 1598 | * If the current CPU has more than one RT task, see if the non
|
| 1599 | * running task can migrate over to a CPU that is running a task
|
| 1600 | * of lesser priority.
|
| 1601 |
|
| 1602 | static int push_rt_task(struct rq *rq)
|
| 1603 | {
|
| 1604 | struct task_struct *next_task;
|
| 1605 | struct rq *lowest_rq;
|
| 1606 | int ret = 0;
|
| 1607 |
|
| 1608 | if (!rq->rt.overloaded)
|
| 1609 | return 0;
|
| 1610 |
|
| 1611 | next_task = pick_next_pushable_task(rq);
|
| 1612 | if (!next_task)
|
| 1613 | return 0;
|
| 1614 |
|
| 1615 | #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
|
| 1616 | if (unlikely(task_running(rq, next_task)))
|
| 1617 | return 0;
|
| 1618 | #endif
|
| 1619 |
|
| 1620 | retry:
|
| 1621 | if (unlikely(next_task == rq->curr)) {
|
| 1622 | WARN_ON(1);
|
| 1623 | return 0;
|
| 1624 | }
|
| 1625 |
|
| 1626 | |
| 1627 | * It's possible that the next_task slipped in of
|
| 1628 | * higher priority than current. If that's the case
|
| 1629 | * just reschedule current.
|
| 1630 |
|
| 1631 | if (unlikely(next_task->prio < rq->curr->prio)) {
|
| 1632 | resched_task(rq->curr);
|
| 1633 | return 0;
|
| 1634 | }
|
| 1635 |
|
| 1636 |
|
| 1637 | get_task_struct(next_task);
|
| 1638 |
|
| 1639 |
|
| 1640 | lowest_rq = find_lock_lowest_rq(next_task, rq);
|
| 1641 | if (!lowest_rq) {
|
| 1642 | struct task_struct *task;
|
| 1643 | |
| 1644 | * find_lock_lowest_rq releases rq->lock
|
| 1645 | * so it is possible that next_task has migrated.
|
| 1646 | *
|
| 1647 | * We need to make sure that the task is still on the same
|
| 1648 | * run-queue and is also still the next task eligible for
|
| 1649 | * pushing.
|
| 1650 |
|
| 1651 | task = pick_next_pushable_task(rq);
|
| 1652 | if (task_cpu(next_task) == rq->cpu && task == next_task) {
|
| 1653 | |
| 1654 | * The task hasn't migrated, and is still the next
|
| 1655 | * eligible task, but we failed to find a run-queue
|
| 1656 | * to push it to. Do not retry in this case, since
|
| 1657 | * other cpus will pull from us when ready.
|
| 1658 |
|
| 1659 | goto out;
|
| 1660 | }
|
| 1661 |
|
| 1662 | if (!task)
|
| 1663 |
|
| 1664 | goto out;
|
| 1665 |
|
| 1666 | |
| 1667 | * Something has shifted, try again.
|
| 1668 |
|
| 1669 | put_task_struct(next_task);
|
| 1670 | next_task = task;
|
| 1671 | goto retry;
|
| 1672 | }
|
| 1673 |
|
| 1674 | deactivate_task(rq, next_task, 0);
|
| 1675 | set_task_cpu(next_task, lowest_rq->cpu);
|
| 1676 | activate_task(lowest_rq, next_task, 0);
|
| 1677 | ret = 1;
|
| 1678 |
|
| 1679 | resched_task(lowest_rq->curr);
|
| 1680 |
|
| 1681 | double_unlock_balance(rq, lowest_rq);
|
| 1682 |
|
| 1683 | out:
|
| 1684 | put_task_struct(next_task);
|
| 1685 |
|
| 1686 | return ret;
|
| 1687 | }
|
| 1688 |
|
| 1689 | static void push_rt_tasks(struct rq *rq)
|
| 1690 | {
|
| 1691 |
|
| 1692 | while (push_rt_task(rq))
|
| 1693 | ;
|
| 1694 | }
|
| 1695 |
|
| 1696 | static int pull_rt_task(struct rq *this_rq)
|
| 1697 | {
|
| 1698 | int this_cpu = this_rq->cpu, ret = 0, cpu;
|
| 1699 | struct task_struct *p;
|
| 1700 | struct rq *src_rq;
|
| 1701 |
|
| 1702 | if (likely(!rt_overloaded(this_rq)))
|
| 1703 | return 0;
|
| 1704 |
|
| 1705 | for_each_cpu(cpu, this_rq->rd->rto_mask) {
|
| 1706 | if (this_cpu == cpu)
|
| 1707 | continue;
|
| 1708 |
|
| 1709 | src_rq = cpu_rq(cpu);
|
| 1710 |
|
| 1711 | |
| 1712 | * Don't bother taking the src_rq->lock if the next highest
|
| 1713 | * task is known to be lower-priority than our current task.
|
| 1714 | * This may look racy, but if this value is about to go
|
| 1715 | * logically higher, the src_rq will push this task away.
|
| 1716 | * And if its going logically lower, we do not care
|
| 1717 |
|
| 1718 | if (src_rq->rt.highest_prio.next >=
|
| 1719 | this_rq->rt.highest_prio.curr)
|
| 1720 | continue;
|
| 1721 |
|
| 1722 | |
| 1723 | * We can potentially drop this_rq's lock in
|
| 1724 | * double_lock_balance, and another CPU could
|
| 1725 | * alter this_rq
|
| 1726 |
|
| 1727 | double_lock_balance(this_rq, src_rq);
|
| 1728 |
|
| 1729 | |
| 1730 | * Are there still pullable RT tasks?
|
| 1731 |
|
| 1732 | if (src_rq->rt.rt_nr_running <= 1)
|
| 1733 | goto skip;
|
| 1734 |
|
| 1735 | p = pick_next_highest_task_rt(src_rq, this_cpu);
|
| 1736 |
|
| 1737 | |
| 1738 | * Do we have an RT task that preempts
|
| 1739 | * the to-be-scheduled task?
|
| 1740 |
|
| 1741 | if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
|
| 1742 | WARN_ON(p == src_rq->curr);
|
| 1743 | WARN_ON(!p->on_rq);
|
| 1744 |
|
| 1745 | |
| 1746 | * There's a chance that p is higher in priority
|
| 1747 | * than what's currently running on its cpu.
|
| 1748 | * This is just that p is wakeing up and hasn't
|
| 1749 | * had a chance to schedule. We only pull
|
| 1750 | * p if it is lower in priority than the
|
| 1751 | * current task on the run queue
|
| 1752 |
|
| 1753 | if (p->prio < src_rq->curr->prio)
|
| 1754 | goto skip;
|
| 1755 |
|
| 1756 | ret = 1;
|
| 1757 |
|
| 1758 | deactivate_task(src_rq, p, 0);
|
| 1759 | set_task_cpu(p, this_cpu);
|
| 1760 | activate_task(this_rq, p, 0);
|
| 1761 | |
| 1762 | * We continue with the search, just in
|
| 1763 | * case there's an even higher prio task
|
| 1764 | * in another runqueue. (low likelihood
|
| 1765 | * but possible)
|
| 1766 |
|
| 1767 | }
|
| 1768 | skip:
|
| 1769 | double_unlock_balance(this_rq, src_rq);
|
| 1770 | }
|
| 1771 |
|
| 1772 | return ret;
|
| 1773 | }
|
| 1774 |
|
| 1775 | static void pre_schedule_rt(struct rq *rq, struct task_struct *prev)
|
| 1776 | {
|
| 1777 |
|
| 1778 | if (rq->rt.highest_prio.curr > prev->prio)
|
| 1779 | pull_rt_task(rq);
|
| 1780 | }
|
| 1781 |
|
| 1782 | static void post_schedule_rt(struct rq *rq)
|
| 1783 | {
|
| 1784 | push_rt_tasks(rq);
|
| 1785 | }
|
| 1786 |
|
| 1787 | |
| 1788 | * If we are not running and we are not going to reschedule soon, we should
|
| 1789 | * try to push tasks away now
|
| 1790 |
|
| 1791 | static void task_woken_rt(struct rq *rq, struct task_struct *p)
|
| 1792 | {
|
| 1793 | if (!task_running(rq, p) &&
|
| 1794 | !test_tsk_need_resched(rq->curr) &&
|
| 1795 | has_pushable_tasks(rq) &&
|
| 1796 | p->rt.nr_cpus_allowed > 1 &&
|
| 1797 | rt_task(rq->curr) &&
|
| 1798 | (rq->curr->rt.nr_cpus_allowed < 2 ||
|
| 1799 | rq->curr->prio <= p->prio))
|
| 1800 | push_rt_tasks(rq);
|
| 1801 | }
|
| 1802 |
|
| 1803 | static void set_cpus_allowed_rt(struct task_struct *p,
|
| 1804 | const struct cpumask *new_mask)
|
| 1805 | {
|
| 1806 | int weight = cpumask_weight(new_mask);
|
| 1807 |
|
| 1808 | BUG_ON(!rt_task(p));
|
| 1809 |
|
| 1810 | |
| 1811 | * Update the migration status of the RQ if we have an RT task
|
| 1812 | * which is running AND changing its weight value.
|
| 1813 |
|
| 1814 | if (p->on_rq && (weight != p->rt.nr_cpus_allowed)) {
|
| 1815 | struct rq *rq = task_rq(p);
|
| 1816 |
|
| 1817 | if (!task_current(rq, p)) {
|
| 1818 | |
| 1819 | * Make sure we dequeue this task from the pushable list
|
| 1820 | * before going further. It will either remain off of
|
| 1821 | * the list because we are no longer pushable, or it
|
| 1822 | * will be requeued.
|
| 1823 |
|
| 1824 | if (p->rt.nr_cpus_allowed > 1)
|
| 1825 | dequeue_pushable_task(rq, p);
|
| 1826 |
|
| 1827 | |
| 1828 | * Requeue if our weight is changing and still > 1
|
| 1829 |
|
| 1830 | if (weight > 1)
|
| 1831 | enqueue_pushable_task(rq, p);
|
| 1832 |
|
| 1833 | }
|
| 1834 |
|
| 1835 | if ((p->rt.nr_cpus_allowed <= 1) && (weight > 1)) {
|
| 1836 | rq->rt.rt_nr_migratory++;
|
| 1837 | } else if ((p->rt.nr_cpus_allowed > 1) && (weight <= 1)) {
|
| 1838 | BUG_ON(!rq->rt.rt_nr_migratory);
|
| 1839 | rq->rt.rt_nr_migratory--;
|
| 1840 | }
|
| 1841 |
|
| 1842 | update_rt_migration(&rq->rt);
|
| 1843 | }
|
| 1844 | }
|
| 1845 |
|
| 1846 |
|
| 1847 | static void rq_online_rt(struct rq *rq)
|
| 1848 | {
|
| 1849 | if (rq->rt.overloaded)
|
| 1850 | rt_set_overload(rq);
|
| 1851 |
|
| 1852 | __enable_runtime(rq);
|
| 1853 |
|
| 1854 | cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
|
| 1855 | }
|
| 1856 |
|
| 1857 |
|
| 1858 | static void rq_offline_rt(struct rq *rq)
|
| 1859 | {
|
| 1860 | if (rq->rt.overloaded)
|
| 1861 | rt_clear_overload(rq);
|
| 1862 |
|
| 1863 | __disable_runtime(rq);
|
| 1864 |
|
| 1865 | cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
|
| 1866 | }
|
| 1867 |
|
| 1868 | |
| 1869 | * When switch from the rt queue, we bring ourselves to a position
|
| 1870 | * that we might want to pull RT tasks from other runqueues.
|
| 1871 |
|
| 1872 | static void switched_from_rt(struct rq *rq, struct task_struct *p)
|
| 1873 | {
|
| 1874 | |
| 1875 | * If there are other RT tasks then we will reschedule
|
| 1876 | * and the scheduling of the other RT tasks will handle
|
| 1877 | * the balancing. But if we are the last RT task
|
| 1878 | * we may need to handle the pulling of RT tasks
|
| 1879 | * now.
|
| 1880 |
|
| 1881 | if (p->on_rq && !rq->rt.rt_nr_running)
|
| 1882 | pull_rt_task(rq);
|
| 1883 | }
|
| 1884 |
|
| 1885 | void init_sched_rt_class(void)
|
| 1886 | {
|
| 1887 | unsigned int i;
|
| 1888 |
|
| 1889 | for_each_possible_cpu(i) {
|
| 1890 | zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
|
| 1891 | GFP_KERNEL, cpu_to_node(i));
|
| 1892 | }
|
| 1893 | }
|
| 1894 | #endif
|
| 1895 |
|
| 1896 | |
| 1897 | * When switching a task to RT, we may overload the runqueue
|
| 1898 | * with RT tasks. In this case we try to push them off to
|
| 1899 | * other runqueues.
|
| 1900 |
|
| 1901 | static void switched_to_rt(struct rq *rq, struct task_struct *p)
|
| 1902 | {
|
| 1903 | int check_resched = 1;
|
| 1904 |
|
| 1905 | |
| 1906 | * If we are already running, then there's nothing
|
| 1907 | * that needs to be done. But if we are not running
|
| 1908 | * we may need to preempt the current running task.
|
| 1909 | * If that current running task is also an RT task
|
| 1910 | * then see if we can move to another run queue.
|
| 1911 |
|
| 1912 | if (p->on_rq && rq->curr != p) {
|
| 1913 | #ifdef CONFIG_SMP
|
| 1914 | if (rq->rt.overloaded && push_rt_task(rq) &&
|
| 1915 |
|
| 1916 | rq != task_rq(p))
|
| 1917 | check_resched = 0;
|
| 1918 | #endif
|
| 1919 | if (check_resched && p->prio < rq->curr->prio)
|
| 1920 | resched_task(rq->curr);
|
| 1921 | }
|
| 1922 | }
|
| 1923 |
|
| 1924 | |
| 1925 | * Priority of the task has changed. This may cause
|
| 1926 | * us to initiate a push or pull.
|
| 1927 |
|
| 1928 | static void
|
| 1929 | prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio)
|
| 1930 | {
|
| 1931 | if (!p->on_rq)
|
| 1932 | return;
|
| 1933 |
|
| 1934 | if (rq->curr == p) {
|
| 1935 | #ifdef CONFIG_SMP
|
| 1936 | |
| 1937 | * If our priority decreases while running, we
|
| 1938 | * may need to pull tasks to this runqueue.
|
| 1939 |
|
| 1940 | if (oldprio < p->prio)
|
| 1941 | pull_rt_task(rq);
|
| 1942 | |
| 1943 | * If there's a higher priority task waiting to run
|
| 1944 | * then reschedule. Note, the above pull_rt_task
|
| 1945 | * can release the rq lock and p could migrate.
|
| 1946 | * Only reschedule if p is still on the same runqueue.
|
| 1947 |
|
| 1948 | if (p->prio > rq->rt.highest_prio.curr && rq->curr == p)
|
| 1949 | resched_task(p);
|
| 1950 | #else
|
| 1951 |
|
| 1952 | if (oldprio < p->prio)
|
| 1953 | resched_task(p);
|
| 1954 | #endif
|
| 1955 | } else {
|
| 1956 | |
| 1957 | * This task is not running, but if it is
|
| 1958 | * greater than the current running task
|
| 1959 | * then reschedule.
|
| 1960 |
|
| 1961 | if (p->prio < rq->curr->prio)
|
| 1962 | resched_task(rq->curr);
|
| 1963 | }
|
| 1964 | }
|
| 1965 |
|
| 1966 | static void watchdog(struct rq *rq, struct task_struct *p)
|
| 1967 | {
|
| 1968 | unsigned long soft, hard;
|
| 1969 |
|
| 1970 |
|
| 1971 | soft = task_rlimit(p, RLIMIT_RTTIME);
|
| 1972 | hard = task_rlimit_max(p, RLIMIT_RTTIME);
|
| 1973 |
|
| 1974 | if (soft != RLIM_INFINITY) {
|
| 1975 | unsigned long next;
|
| 1976 |
|
| 1977 | p->rt.timeout++;
|
| 1978 | next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
|
| 1979 | if (p->rt.timeout > next)
|
| 1980 | p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
|
| 1981 | }
|
| 1982 | }
|
| 1983 |
|
| 1984 | static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
|
| 1985 | {
|
| 1986 | update_curr_rt(rq);
|
| 1987 |
|
| 1988 | watchdog(rq, p);
|
| 1989 |
|
| 1990 | |
| 1991 | * RR tasks need a special form of timeslice management.
|
| 1992 | * FIFO tasks have no timeslices.
|
| 1993 |
|
| 1994 | if (p->policy != SCHED_RR)
|
| 1995 | return;
|
| 1996 |
|
| 1997 | if (--p->rt.time_slice)
|
| 1998 | return;
|
| 1999 |
|
| 2000 | p->rt.time_slice = RR_TIMESLICE;
|
| 2001 |
|
| 2002 | |
| 2003 | * Requeue to the end of queue if we are not the only element
|
| 2004 | * on the queue:
|
| 2005 |
|
| 2006 | if (p->rt.run_list.prev != p->rt.run_list.next) {
|
| 2007 | requeue_task_rt(rq, p, 0);
|
| 2008 | set_tsk_need_resched(p);
|
| 2009 | }
|
| 2010 | }
|
| 2011 |
|
| 2012 | static void set_curr_task_rt(struct rq *rq)
|
| 2013 | {
|
| 2014 | struct task_struct *p = rq->curr;
|
| 2015 |
|
| 2016 | p->se.exec_start = rq->clock_task;
|
| 2017 |
|
| 2018 |
|
| 2019 | dequeue_pushable_task(rq, p);
|
| 2020 | }
|
| 2021 |
|
| 2022 | static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task)
|
| 2023 | {
|
| 2024 | |
| 2025 | * Time slice is 0 for SCHED_FIFO tasks
|
| 2026 |
|
| 2027 | if (task->policy == SCHED_RR)
|
| 2028 | return RR_TIMESLICE;
|
| 2029 | else
|
| 2030 | return 0;
|
| 2031 | }
|
| 2032 |
|
| 2033 | const struct sched_class rt_sched_class = {
|
| 2034 | .next = &fair_sched_class,
|
| 2035 | .enqueue_task = enqueue_task_rt,
|
| 2036 | .dequeue_task = dequeue_task_rt,
|
| 2037 | .yield_task = yield_task_rt,
|
| 2038 |
|
| 2039 | .check_preempt_curr = check_preempt_curr_rt,
|
| 2040 |
|
| 2041 | .pick_next_task = pick_next_task_rt,
|
| 2042 | .put_prev_task = put_prev_task_rt,
|
| 2043 |
|
| 2044 | #ifdef CONFIG_SMP
|
| 2045 | .select_task_rq = select_task_rq_rt,
|
| 2046 |
|
| 2047 | .set_cpus_allowed = set_cpus_allowed_rt,
|
| 2048 | .rq_online = rq_online_rt,
|
| 2049 | .rq_offline = rq_offline_rt,
|
| 2050 | .pre_schedule = pre_schedule_rt,
|
| 2051 | .post_schedule = post_schedule_rt,
|
| 2052 | .task_woken = task_woken_rt,
|
| 2053 | .switched_from = switched_from_rt,
|
| 2054 | #endif
|
| 2055 |
|
| 2056 | .set_curr_task = set_curr_task_rt,
|
| 2057 | .task_tick = task_tick_rt,
|
| 2058 |
|
| 2059 | .get_rr_interval = get_rr_interval_rt,
|
| 2060 |
|
| 2061 | .prio_changed = prio_changed_rt,
|
| 2062 | .switched_to = switched_to_rt,
|
| 2063 | };
|
| 2064 |
|
| 2065 | #ifdef CONFIG_SCHED_DEBUG
|
| 2066 | extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
|
| 2067 |
|
| 2068 | void print_rt_stats(struct seq_file *m, int cpu)
|
| 2069 | {
|
| 2070 | rt_rq_iter_t iter;
|
| 2071 | struct rt_rq *rt_rq;
|
| 2072 |
|
| 2073 | rcu_read_lock();
|
| 2074 | for_each_rt_rq(rt_rq, iter, cpu_rq(cpu))
|
| 2075 | print_rt_rq(m, cpu, rt_rq);
|
| 2076 | rcu_read_unlock();
|
| 2077 | }
|
| 2078 | #endif
|