| 1 | /* |
| 2 | * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH) |
| 3 | * |
| 4 | * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com> |
| 5 | * |
| 6 | * Interactivity improvements by Mike Galbraith |
| 7 | * (C) 2007 Mike Galbraith <efault@gmx.de> |
| 8 | * |
| 9 | * Various enhancements by Dmitry Adamushko. |
| 10 | * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com> |
| 11 | * |
| 12 | * Group scheduling enhancements by Srivatsa Vaddagiri |
| 13 | * Copyright IBM Corporation, 2007 |
| 14 | * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> |
| 15 | * |
| 16 | * Scaled math optimizations by Thomas Gleixner |
| 17 | * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de> |
| 18 | * |
| 19 | * Adaptive scheduling granularity, math enhancements by Peter Zijlstra |
| 20 | * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com> |
| 21 | */ |
| 22 | |
| 23 | #include <linux/latencytop.h> |
| 24 | |
| 25 | /* |
| 26 | * Targeted preemption latency for CPU-bound tasks: |
| 27 | * (default: 5ms * (1 + ilog(ncpus)), units: nanoseconds) |
| 28 | * |
| 29 | * NOTE: this latency value is not the same as the concept of |
| 30 | * 'timeslice length' - timeslices in CFS are of variable length |
| 31 | * and have no persistent notion like in traditional, time-slice |
| 32 | * based scheduling concepts. |
| 33 | * |
| 34 | * (to see the precise effective timeslice length of your workload, |
| 35 | * run vmstat and monitor the context-switches (cs) field) |
| 36 | */ |
| 37 | unsigned int sysctl_sched_latency = 5000000ULL; |
| 38 | |
| 39 | /* |
| 40 | * Minimal preemption granularity for CPU-bound tasks: |
| 41 | * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds) |
| 42 | */ |
| 43 | unsigned int sysctl_sched_min_granularity = 1000000ULL; |
| 44 | |
| 45 | /* |
| 46 | * is kept at sysctl_sched_latency / sysctl_sched_min_granularity |
| 47 | */ |
| 48 | static unsigned int sched_nr_latency = 5; |
| 49 | |
| 50 | /* |
| 51 | * After fork, child runs first. If set to 0 (default) then |
| 52 | * parent will (try to) run first. |
| 53 | */ |
| 54 | unsigned int sysctl_sched_child_runs_first __read_mostly; |
| 55 | |
| 56 | /* |
| 57 | * sys_sched_yield() compat mode |
| 58 | * |
| 59 | * This option switches the agressive yield implementation of the |
| 60 | * old scheduler back on. |
| 61 | */ |
| 62 | unsigned int __read_mostly sysctl_sched_compat_yield; |
| 63 | |
| 64 | /* |
| 65 | * SCHED_OTHER wake-up granularity. |
| 66 | * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds) |
| 67 | * |
| 68 | * This option delays the preemption effects of decoupled workloads |
| 69 | * and reduces their over-scheduling. Synchronous workloads will still |
| 70 | * have immediate wakeup/sleep latencies. |
| 71 | */ |
| 72 | unsigned int sysctl_sched_wakeup_granularity = 1000000UL; |
| 73 | |
| 74 | const_debug unsigned int sysctl_sched_migration_cost = 500000UL; |
| 75 | |
| 76 | static const struct sched_class fair_sched_class; |
| 77 | |
| 78 | /************************************************************** |
| 79 | * CFS operations on generic schedulable entities: |
| 80 | */ |
| 81 | |
| 82 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 83 | |
| 84 | /* cpu runqueue to which this cfs_rq is attached */ |
| 85 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
| 86 | { |
| 87 | return cfs_rq->rq; |
| 88 | } |
| 89 | |
| 90 | /* An entity is a task if it doesn't "own" a runqueue */ |
| 91 | #define entity_is_task(se) (!se->my_q) |
| 92 | |
| 93 | static inline struct task_struct *task_of(struct sched_entity *se) |
| 94 | { |
| 95 | #ifdef CONFIG_SCHED_DEBUG |
| 96 | WARN_ON_ONCE(!entity_is_task(se)); |
| 97 | #endif |
| 98 | return container_of(se, struct task_struct, se); |
| 99 | } |
| 100 | |
| 101 | /* Walk up scheduling entities hierarchy */ |
| 102 | #define for_each_sched_entity(se) \ |
| 103 | for (; se; se = se->parent) |
| 104 | |
| 105 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) |
| 106 | { |
| 107 | return p->se.cfs_rq; |
| 108 | } |
| 109 | |
| 110 | /* runqueue on which this entity is (to be) queued */ |
| 111 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) |
| 112 | { |
| 113 | return se->cfs_rq; |
| 114 | } |
| 115 | |
| 116 | /* runqueue "owned" by this group */ |
| 117 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) |
| 118 | { |
| 119 | return grp->my_q; |
| 120 | } |
| 121 | |
| 122 | /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on |
| 123 | * another cpu ('this_cpu') |
| 124 | */ |
| 125 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) |
| 126 | { |
| 127 | return cfs_rq->tg->cfs_rq[this_cpu]; |
| 128 | } |
| 129 | |
| 130 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ |
| 131 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ |
| 132 | list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) |
| 133 | |
| 134 | /* Do the two (enqueued) entities belong to the same group ? */ |
| 135 | static inline int |
| 136 | is_same_group(struct sched_entity *se, struct sched_entity *pse) |
| 137 | { |
| 138 | if (se->cfs_rq == pse->cfs_rq) |
| 139 | return 1; |
| 140 | |
| 141 | return 0; |
| 142 | } |
| 143 | |
| 144 | static inline struct sched_entity *parent_entity(struct sched_entity *se) |
| 145 | { |
| 146 | return se->parent; |
| 147 | } |
| 148 | |
| 149 | /* return depth at which a sched entity is present in the hierarchy */ |
| 150 | static inline int depth_se(struct sched_entity *se) |
| 151 | { |
| 152 | int depth = 0; |
| 153 | |
| 154 | for_each_sched_entity(se) |
| 155 | depth++; |
| 156 | |
| 157 | return depth; |
| 158 | } |
| 159 | |
| 160 | static void |
| 161 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) |
| 162 | { |
| 163 | int se_depth, pse_depth; |
| 164 | |
| 165 | /* |
| 166 | * preemption test can be made between sibling entities who are in the |
| 167 | * same cfs_rq i.e who have a common parent. Walk up the hierarchy of |
| 168 | * both tasks until we find their ancestors who are siblings of common |
| 169 | * parent. |
| 170 | */ |
| 171 | |
| 172 | /* First walk up until both entities are at same depth */ |
| 173 | se_depth = depth_se(*se); |
| 174 | pse_depth = depth_se(*pse); |
| 175 | |
| 176 | while (se_depth > pse_depth) { |
| 177 | se_depth--; |
| 178 | *se = parent_entity(*se); |
| 179 | } |
| 180 | |
| 181 | while (pse_depth > se_depth) { |
| 182 | pse_depth--; |
| 183 | *pse = parent_entity(*pse); |
| 184 | } |
| 185 | |
| 186 | while (!is_same_group(*se, *pse)) { |
| 187 | *se = parent_entity(*se); |
| 188 | *pse = parent_entity(*pse); |
| 189 | } |
| 190 | } |
| 191 | |
| 192 | #else /* !CONFIG_FAIR_GROUP_SCHED */ |
| 193 | |
| 194 | static inline struct task_struct *task_of(struct sched_entity *se) |
| 195 | { |
| 196 | return container_of(se, struct task_struct, se); |
| 197 | } |
| 198 | |
| 199 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
| 200 | { |
| 201 | return container_of(cfs_rq, struct rq, cfs); |
| 202 | } |
| 203 | |
| 204 | #define entity_is_task(se) 1 |
| 205 | |
| 206 | #define for_each_sched_entity(se) \ |
| 207 | for (; se; se = NULL) |
| 208 | |
| 209 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) |
| 210 | { |
| 211 | return &task_rq(p)->cfs; |
| 212 | } |
| 213 | |
| 214 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) |
| 215 | { |
| 216 | struct task_struct *p = task_of(se); |
| 217 | struct rq *rq = task_rq(p); |
| 218 | |
| 219 | return &rq->cfs; |
| 220 | } |
| 221 | |
| 222 | /* runqueue "owned" by this group */ |
| 223 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) |
| 224 | { |
| 225 | return NULL; |
| 226 | } |
| 227 | |
| 228 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) |
| 229 | { |
| 230 | return &cpu_rq(this_cpu)->cfs; |
| 231 | } |
| 232 | |
| 233 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ |
| 234 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) |
| 235 | |
| 236 | static inline int |
| 237 | is_same_group(struct sched_entity *se, struct sched_entity *pse) |
| 238 | { |
| 239 | return 1; |
| 240 | } |
| 241 | |
| 242 | static inline struct sched_entity *parent_entity(struct sched_entity *se) |
| 243 | { |
| 244 | return NULL; |
| 245 | } |
| 246 | |
| 247 | static inline void |
| 248 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) |
| 249 | { |
| 250 | } |
| 251 | |
| 252 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
| 253 | |
| 254 | |
| 255 | /************************************************************** |
| 256 | * Scheduling class tree data structure manipulation methods: |
| 257 | */ |
| 258 | |
| 259 | static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime) |
| 260 | { |
| 261 | s64 delta = (s64)(vruntime - min_vruntime); |
| 262 | if (delta > 0) |
| 263 | min_vruntime = vruntime; |
| 264 | |
| 265 | return min_vruntime; |
| 266 | } |
| 267 | |
| 268 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) |
| 269 | { |
| 270 | s64 delta = (s64)(vruntime - min_vruntime); |
| 271 | if (delta < 0) |
| 272 | min_vruntime = vruntime; |
| 273 | |
| 274 | return min_vruntime; |
| 275 | } |
| 276 | |
| 277 | static inline int entity_before(struct sched_entity *a, |
| 278 | struct sched_entity *b) |
| 279 | { |
| 280 | return (s64)(a->vruntime - b->vruntime) < 0; |
| 281 | } |
| 282 | |
| 283 | static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se) |
| 284 | { |
| 285 | return se->vruntime - cfs_rq->min_vruntime; |
| 286 | } |
| 287 | |
| 288 | static void update_min_vruntime(struct cfs_rq *cfs_rq) |
| 289 | { |
| 290 | u64 vruntime = cfs_rq->min_vruntime; |
| 291 | |
| 292 | if (cfs_rq->curr) |
| 293 | vruntime = cfs_rq->curr->vruntime; |
| 294 | |
| 295 | if (cfs_rq->rb_leftmost) { |
| 296 | struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost, |
| 297 | struct sched_entity, |
| 298 | run_node); |
| 299 | |
| 300 | if (!cfs_rq->curr) |
| 301 | vruntime = se->vruntime; |
| 302 | else |
| 303 | vruntime = min_vruntime(vruntime, se->vruntime); |
| 304 | } |
| 305 | |
| 306 | cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); |
| 307 | } |
| 308 | |
| 309 | /* |
| 310 | * Enqueue an entity into the rb-tree: |
| 311 | */ |
| 312 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
| 313 | { |
| 314 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; |
| 315 | struct rb_node *parent = NULL; |
| 316 | struct sched_entity *entry; |
| 317 | s64 key = entity_key(cfs_rq, se); |
| 318 | int leftmost = 1; |
| 319 | |
| 320 | /* |
| 321 | * Find the right place in the rbtree: |
| 322 | */ |
| 323 | while (*link) { |
| 324 | parent = *link; |
| 325 | entry = rb_entry(parent, struct sched_entity, run_node); |
| 326 | /* |
| 327 | * We dont care about collisions. Nodes with |
| 328 | * the same key stay together. |
| 329 | */ |
| 330 | if (key < entity_key(cfs_rq, entry)) { |
| 331 | link = &parent->rb_left; |
| 332 | } else { |
| 333 | link = &parent->rb_right; |
| 334 | leftmost = 0; |
| 335 | } |
| 336 | } |
| 337 | |
| 338 | /* |
| 339 | * Maintain a cache of leftmost tree entries (it is frequently |
| 340 | * used): |
| 341 | */ |
| 342 | if (leftmost) |
| 343 | cfs_rq->rb_leftmost = &se->run_node; |
| 344 | |
| 345 | rb_link_node(&se->run_node, parent, link); |
| 346 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); |
| 347 | } |
| 348 | |
| 349 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
| 350 | { |
| 351 | if (cfs_rq->rb_leftmost == &se->run_node) { |
| 352 | struct rb_node *next_node; |
| 353 | |
| 354 | next_node = rb_next(&se->run_node); |
| 355 | cfs_rq->rb_leftmost = next_node; |
| 356 | } |
| 357 | |
| 358 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); |
| 359 | } |
| 360 | |
| 361 | static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq) |
| 362 | { |
| 363 | struct rb_node *left = cfs_rq->rb_leftmost; |
| 364 | |
| 365 | if (!left) |
| 366 | return NULL; |
| 367 | |
| 368 | return rb_entry(left, struct sched_entity, run_node); |
| 369 | } |
| 370 | |
| 371 | static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
| 372 | { |
| 373 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline); |
| 374 | |
| 375 | if (!last) |
| 376 | return NULL; |
| 377 | |
| 378 | return rb_entry(last, struct sched_entity, run_node); |
| 379 | } |
| 380 | |
| 381 | /************************************************************** |
| 382 | * Scheduling class statistics methods: |
| 383 | */ |
| 384 | |
| 385 | #ifdef CONFIG_SCHED_DEBUG |
| 386 | int sched_nr_latency_handler(struct ctl_table *table, int write, |
| 387 | struct file *filp, void __user *buffer, size_t *lenp, |
| 388 | loff_t *ppos) |
| 389 | { |
| 390 | int ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos); |
| 391 | |
| 392 | if (ret || !write) |
| 393 | return ret; |
| 394 | |
| 395 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, |
| 396 | sysctl_sched_min_granularity); |
| 397 | |
| 398 | return 0; |
| 399 | } |
| 400 | #endif |
| 401 | |
| 402 | /* |
| 403 | * delta /= w |
| 404 | */ |
| 405 | static inline unsigned long |
| 406 | calc_delta_fair(unsigned long delta, struct sched_entity *se) |
| 407 | { |
| 408 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
| 409 | delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load); |
| 410 | |
| 411 | return delta; |
| 412 | } |
| 413 | |
| 414 | /* |
| 415 | * The idea is to set a period in which each task runs once. |
| 416 | * |
| 417 | * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch |
| 418 | * this period because otherwise the slices get too small. |
| 419 | * |
| 420 | * p = (nr <= nl) ? l : l*nr/nl |
| 421 | */ |
| 422 | static u64 __sched_period(unsigned long nr_running) |
| 423 | { |
| 424 | u64 period = sysctl_sched_latency; |
| 425 | unsigned long nr_latency = sched_nr_latency; |
| 426 | |
| 427 | if (unlikely(nr_running > nr_latency)) { |
| 428 | period = sysctl_sched_min_granularity; |
| 429 | period *= nr_running; |
| 430 | } |
| 431 | |
| 432 | return period; |
| 433 | } |
| 434 | |
| 435 | /* |
| 436 | * We calculate the wall-time slice from the period by taking a part |
| 437 | * proportional to the weight. |
| 438 | * |
| 439 | * s = p*P[w/rw] |
| 440 | */ |
| 441 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
| 442 | { |
| 443 | u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq); |
| 444 | |
| 445 | for_each_sched_entity(se) { |
| 446 | struct load_weight *load; |
| 447 | struct load_weight lw; |
| 448 | |
| 449 | cfs_rq = cfs_rq_of(se); |
| 450 | load = &cfs_rq->load; |
| 451 | |
| 452 | if (unlikely(!se->on_rq)) { |
| 453 | lw = cfs_rq->load; |
| 454 | |
| 455 | update_load_add(&lw, se->load.weight); |
| 456 | load = &lw; |
| 457 | } |
| 458 | slice = calc_delta_mine(slice, se->load.weight, load); |
| 459 | } |
| 460 | return slice; |
| 461 | } |
| 462 | |
| 463 | /* |
| 464 | * We calculate the vruntime slice of a to be inserted task |
| 465 | * |
| 466 | * vs = s/w |
| 467 | */ |
| 468 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
| 469 | { |
| 470 | return calc_delta_fair(sched_slice(cfs_rq, se), se); |
| 471 | } |
| 472 | |
| 473 | /* |
| 474 | * Update the current task's runtime statistics. Skip current tasks that |
| 475 | * are not in our scheduling class. |
| 476 | */ |
| 477 | static inline void |
| 478 | __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr, |
| 479 | unsigned long delta_exec) |
| 480 | { |
| 481 | unsigned long delta_exec_weighted; |
| 482 | |
| 483 | schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max)); |
| 484 | |
| 485 | curr->sum_exec_runtime += delta_exec; |
| 486 | schedstat_add(cfs_rq, exec_clock, delta_exec); |
| 487 | delta_exec_weighted = calc_delta_fair(delta_exec, curr); |
| 488 | curr->vruntime += delta_exec_weighted; |
| 489 | update_min_vruntime(cfs_rq); |
| 490 | } |
| 491 | |
| 492 | static void update_curr(struct cfs_rq *cfs_rq) |
| 493 | { |
| 494 | struct sched_entity *curr = cfs_rq->curr; |
| 495 | u64 now = rq_of(cfs_rq)->clock; |
| 496 | unsigned long delta_exec; |
| 497 | |
| 498 | if (unlikely(!curr)) |
| 499 | return; |
| 500 | |
| 501 | /* |
| 502 | * Get the amount of time the current task was running |
| 503 | * since the last time we changed load (this cannot |
| 504 | * overflow on 32 bits): |
| 505 | */ |
| 506 | delta_exec = (unsigned long)(now - curr->exec_start); |
| 507 | if (!delta_exec) |
| 508 | return; |
| 509 | |
| 510 | __update_curr(cfs_rq, curr, delta_exec); |
| 511 | curr->exec_start = now; |
| 512 | |
| 513 | if (entity_is_task(curr)) { |
| 514 | struct task_struct *curtask = task_of(curr); |
| 515 | |
| 516 | trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); |
| 517 | cpuacct_charge(curtask, delta_exec); |
| 518 | account_group_exec_runtime(curtask, delta_exec); |
| 519 | } |
| 520 | } |
| 521 | |
| 522 | static inline void |
| 523 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
| 524 | { |
| 525 | schedstat_set(se->wait_start, rq_of(cfs_rq)->clock); |
| 526 | } |
| 527 | |
| 528 | /* |
| 529 | * Task is being enqueued - update stats: |
| 530 | */ |
| 531 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
| 532 | { |
| 533 | /* |
| 534 | * Are we enqueueing a waiting task? (for current tasks |
| 535 | * a dequeue/enqueue event is a NOP) |
| 536 | */ |
| 537 | if (se != cfs_rq->curr) |
| 538 | update_stats_wait_start(cfs_rq, se); |
| 539 | } |
| 540 | |
| 541 | static void |
| 542 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
| 543 | { |
| 544 | schedstat_set(se->wait_max, max(se->wait_max, |
| 545 | rq_of(cfs_rq)->clock - se->wait_start)); |
| 546 | schedstat_set(se->wait_count, se->wait_count + 1); |
| 547 | schedstat_set(se->wait_sum, se->wait_sum + |
| 548 | rq_of(cfs_rq)->clock - se->wait_start); |
| 549 | #ifdef CONFIG_SCHEDSTATS |
| 550 | if (entity_is_task(se)) { |
| 551 | trace_sched_stat_wait(task_of(se), |
| 552 | rq_of(cfs_rq)->clock - se->wait_start); |
| 553 | } |
| 554 | #endif |
| 555 | schedstat_set(se->wait_start, 0); |
| 556 | } |
| 557 | |
| 558 | static inline void |
| 559 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
| 560 | { |
| 561 | /* |
| 562 | * Mark the end of the wait period if dequeueing a |
| 563 | * waiting task: |
| 564 | */ |
| 565 | if (se != cfs_rq->curr) |
| 566 | update_stats_wait_end(cfs_rq, se); |
| 567 | } |
| 568 | |
| 569 | /* |
| 570 | * We are picking a new current task - update its stats: |
| 571 | */ |
| 572 | static inline void |
| 573 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
| 574 | { |
| 575 | /* |
| 576 | * We are starting a new run period: |
| 577 | */ |
| 578 | se->exec_start = rq_of(cfs_rq)->clock; |
| 579 | } |
| 580 | |
| 581 | /************************************************** |
| 582 | * Scheduling class queueing methods: |
| 583 | */ |
| 584 | |
| 585 | #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED |
| 586 | static void |
| 587 | add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight) |
| 588 | { |
| 589 | cfs_rq->task_weight += weight; |
| 590 | } |
| 591 | #else |
| 592 | static inline void |
| 593 | add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight) |
| 594 | { |
| 595 | } |
| 596 | #endif |
| 597 | |
| 598 | static void |
| 599 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
| 600 | { |
| 601 | update_load_add(&cfs_rq->load, se->load.weight); |
| 602 | if (!parent_entity(se)) |
| 603 | inc_cpu_load(rq_of(cfs_rq), se->load.weight); |
| 604 | if (entity_is_task(se)) { |
| 605 | add_cfs_task_weight(cfs_rq, se->load.weight); |
| 606 | list_add(&se->group_node, &cfs_rq->tasks); |
| 607 | } |
| 608 | cfs_rq->nr_running++; |
| 609 | se->on_rq = 1; |
| 610 | } |
| 611 | |
| 612 | static void |
| 613 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
| 614 | { |
| 615 | update_load_sub(&cfs_rq->load, se->load.weight); |
| 616 | if (!parent_entity(se)) |
| 617 | dec_cpu_load(rq_of(cfs_rq), se->load.weight); |
| 618 | if (entity_is_task(se)) { |
| 619 | add_cfs_task_weight(cfs_rq, -se->load.weight); |
| 620 | list_del_init(&se->group_node); |
| 621 | } |
| 622 | cfs_rq->nr_running--; |
| 623 | se->on_rq = 0; |
| 624 | } |
| 625 | |
| 626 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
| 627 | { |
| 628 | #ifdef CONFIG_SCHEDSTATS |
| 629 | struct task_struct *tsk = NULL; |
| 630 | |
| 631 | if (entity_is_task(se)) |
| 632 | tsk = task_of(se); |
| 633 | |
| 634 | if (se->sleep_start) { |
| 635 | u64 delta = rq_of(cfs_rq)->clock - se->sleep_start; |
| 636 | |
| 637 | if ((s64)delta < 0) |
| 638 | delta = 0; |
| 639 | |
| 640 | if (unlikely(delta > se->sleep_max)) |
| 641 | se->sleep_max = delta; |
| 642 | |
| 643 | se->sleep_start = 0; |
| 644 | se->sum_sleep_runtime += delta; |
| 645 | |
| 646 | if (tsk) { |
| 647 | account_scheduler_latency(tsk, delta >> 10, 1); |
| 648 | trace_sched_stat_sleep(tsk, delta); |
| 649 | } |
| 650 | } |
| 651 | if (se->block_start) { |
| 652 | u64 delta = rq_of(cfs_rq)->clock - se->block_start; |
| 653 | |
| 654 | if ((s64)delta < 0) |
| 655 | delta = 0; |
| 656 | |
| 657 | if (unlikely(delta > se->block_max)) |
| 658 | se->block_max = delta; |
| 659 | |
| 660 | se->block_start = 0; |
| 661 | se->sum_sleep_runtime += delta; |
| 662 | |
| 663 | if (tsk) { |
| 664 | if (tsk->in_iowait) { |
| 665 | se->iowait_sum += delta; |
| 666 | se->iowait_count++; |
| 667 | trace_sched_stat_iowait(tsk, delta); |
| 668 | } |
| 669 | |
| 670 | /* |
| 671 | * Blocking time is in units of nanosecs, so shift by |
| 672 | * 20 to get a milliseconds-range estimation of the |
| 673 | * amount of time that the task spent sleeping: |
| 674 | */ |
| 675 | if (unlikely(prof_on == SLEEP_PROFILING)) { |
| 676 | profile_hits(SLEEP_PROFILING, |
| 677 | (void *)get_wchan(tsk), |
| 678 | delta >> 20); |
| 679 | } |
| 680 | account_scheduler_latency(tsk, delta >> 10, 0); |
| 681 | } |
| 682 | } |
| 683 | #endif |
| 684 | } |
| 685 | |
| 686 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
| 687 | { |
| 688 | #ifdef CONFIG_SCHED_DEBUG |
| 689 | s64 d = se->vruntime - cfs_rq->min_vruntime; |
| 690 | |
| 691 | if (d < 0) |
| 692 | d = -d; |
| 693 | |
| 694 | if (d > 3*sysctl_sched_latency) |
| 695 | schedstat_inc(cfs_rq, nr_spread_over); |
| 696 | #endif |
| 697 | } |
| 698 | |
| 699 | static void |
| 700 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) |
| 701 | { |
| 702 | u64 vruntime = cfs_rq->min_vruntime; |
| 703 | |
| 704 | /* |
| 705 | * The 'current' period is already promised to the current tasks, |
| 706 | * however the extra weight of the new task will slow them down a |
| 707 | * little, place the new task so that it fits in the slot that |
| 708 | * stays open at the end. |
| 709 | */ |
| 710 | if (initial && sched_feat(START_DEBIT)) |
| 711 | vruntime += sched_vslice(cfs_rq, se); |
| 712 | |
| 713 | if (!initial) { |
| 714 | /* sleeps upto a single latency don't count. */ |
| 715 | if (sched_feat(FAIR_SLEEPERS)) { |
| 716 | unsigned long thresh = sysctl_sched_latency; |
| 717 | |
| 718 | /* |
| 719 | * Convert the sleeper threshold into virtual time. |
| 720 | * SCHED_IDLE is a special sub-class. We care about |
| 721 | * fairness only relative to other SCHED_IDLE tasks, |
| 722 | * all of which have the same weight. |
| 723 | */ |
| 724 | if (sched_feat(NORMALIZED_SLEEPER) && |
| 725 | (!entity_is_task(se) || |
| 726 | task_of(se)->policy != SCHED_IDLE)) |
| 727 | thresh = calc_delta_fair(thresh, se); |
| 728 | |
| 729 | /* |
| 730 | * Halve their sleep time's effect, to allow |
| 731 | * for a gentler effect of sleepers: |
| 732 | */ |
| 733 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) |
| 734 | thresh >>= 1; |
| 735 | |
| 736 | vruntime -= thresh; |
| 737 | } |
| 738 | } |
| 739 | |
| 740 | /* ensure we never gain time by being placed backwards. */ |
| 741 | vruntime = max_vruntime(se->vruntime, vruntime); |
| 742 | |
| 743 | se->vruntime = vruntime; |
| 744 | } |
| 745 | |
| 746 | static void |
| 747 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup) |
| 748 | { |
| 749 | /* |
| 750 | * Update run-time statistics of the 'current'. |
| 751 | */ |
| 752 | update_curr(cfs_rq); |
| 753 | account_entity_enqueue(cfs_rq, se); |
| 754 | |
| 755 | if (wakeup) { |
| 756 | place_entity(cfs_rq, se, 0); |
| 757 | enqueue_sleeper(cfs_rq, se); |
| 758 | } |
| 759 | |
| 760 | update_stats_enqueue(cfs_rq, se); |
| 761 | check_spread(cfs_rq, se); |
| 762 | if (se != cfs_rq->curr) |
| 763 | __enqueue_entity(cfs_rq, se); |
| 764 | } |
| 765 | |
| 766 | static void __clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
| 767 | { |
| 768 | if (!se || cfs_rq->last == se) |
| 769 | cfs_rq->last = NULL; |
| 770 | |
| 771 | if (!se || cfs_rq->next == se) |
| 772 | cfs_rq->next = NULL; |
| 773 | } |
| 774 | |
| 775 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
| 776 | { |
| 777 | for_each_sched_entity(se) |
| 778 | __clear_buddies(cfs_rq_of(se), se); |
| 779 | } |
| 780 | |
| 781 | static void |
| 782 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep) |
| 783 | { |
| 784 | /* |
| 785 | * Update run-time statistics of the 'current'. |
| 786 | */ |
| 787 | update_curr(cfs_rq); |
| 788 | |
| 789 | update_stats_dequeue(cfs_rq, se); |
| 790 | if (sleep) { |
| 791 | #ifdef CONFIG_SCHEDSTATS |
| 792 | if (entity_is_task(se)) { |
| 793 | struct task_struct *tsk = task_of(se); |
| 794 | |
| 795 | if (tsk->state & TASK_INTERRUPTIBLE) |
| 796 | se->sleep_start = rq_of(cfs_rq)->clock; |
| 797 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
| 798 | se->block_start = rq_of(cfs_rq)->clock; |
| 799 | } |
| 800 | #endif |
| 801 | } |
| 802 | |
| 803 | clear_buddies(cfs_rq, se); |
| 804 | |
| 805 | if (se != cfs_rq->curr) |
| 806 | __dequeue_entity(cfs_rq, se); |
| 807 | account_entity_dequeue(cfs_rq, se); |
| 808 | update_min_vruntime(cfs_rq); |
| 809 | } |
| 810 | |
| 811 | /* |
| 812 | * Preempt the current task with a newly woken task if needed: |
| 813 | */ |
| 814 | static void |
| 815 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
| 816 | { |
| 817 | unsigned long ideal_runtime, delta_exec; |
| 818 | |
| 819 | ideal_runtime = sched_slice(cfs_rq, curr); |
| 820 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
| 821 | if (delta_exec > ideal_runtime) { |
| 822 | resched_task(rq_of(cfs_rq)->curr); |
| 823 | /* |
| 824 | * The current task ran long enough, ensure it doesn't get |
| 825 | * re-elected due to buddy favours. |
| 826 | */ |
| 827 | clear_buddies(cfs_rq, curr); |
| 828 | } |
| 829 | } |
| 830 | |
| 831 | static void |
| 832 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
| 833 | { |
| 834 | /* 'current' is not kept within the tree. */ |
| 835 | if (se->on_rq) { |
| 836 | /* |
| 837 | * Any task has to be enqueued before it get to execute on |
| 838 | * a CPU. So account for the time it spent waiting on the |
| 839 | * runqueue. |
| 840 | */ |
| 841 | update_stats_wait_end(cfs_rq, se); |
| 842 | __dequeue_entity(cfs_rq, se); |
| 843 | } |
| 844 | |
| 845 | update_stats_curr_start(cfs_rq, se); |
| 846 | cfs_rq->curr = se; |
| 847 | #ifdef CONFIG_SCHEDSTATS |
| 848 | /* |
| 849 | * Track our maximum slice length, if the CPU's load is at |
| 850 | * least twice that of our own weight (i.e. dont track it |
| 851 | * when there are only lesser-weight tasks around): |
| 852 | */ |
| 853 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
| 854 | se->slice_max = max(se->slice_max, |
| 855 | se->sum_exec_runtime - se->prev_sum_exec_runtime); |
| 856 | } |
| 857 | #endif |
| 858 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
| 859 | } |
| 860 | |
| 861 | static int |
| 862 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); |
| 863 | |
| 864 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
| 865 | { |
| 866 | struct sched_entity *se = __pick_next_entity(cfs_rq); |
| 867 | |
| 868 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, se) < 1) |
| 869 | return cfs_rq->next; |
| 870 | |
| 871 | if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, se) < 1) |
| 872 | return cfs_rq->last; |
| 873 | |
| 874 | return se; |
| 875 | } |
| 876 | |
| 877 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
| 878 | { |
| 879 | /* |
| 880 | * If still on the runqueue then deactivate_task() |
| 881 | * was not called and update_curr() has to be done: |
| 882 | */ |
| 883 | if (prev->on_rq) |
| 884 | update_curr(cfs_rq); |
| 885 | |
| 886 | check_spread(cfs_rq, prev); |
| 887 | if (prev->on_rq) { |
| 888 | update_stats_wait_start(cfs_rq, prev); |
| 889 | /* Put 'current' back into the tree. */ |
| 890 | __enqueue_entity(cfs_rq, prev); |
| 891 | } |
| 892 | cfs_rq->curr = NULL; |
| 893 | } |
| 894 | |
| 895 | static void |
| 896 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) |
| 897 | { |
| 898 | /* |
| 899 | * Update run-time statistics of the 'current'. |
| 900 | */ |
| 901 | update_curr(cfs_rq); |
| 902 | |
| 903 | #ifdef CONFIG_SCHED_HRTICK |
| 904 | /* |
| 905 | * queued ticks are scheduled to match the slice, so don't bother |
| 906 | * validating it and just reschedule. |
| 907 | */ |
| 908 | if (queued) { |
| 909 | resched_task(rq_of(cfs_rq)->curr); |
| 910 | return; |
| 911 | } |
| 912 | /* |
| 913 | * don't let the period tick interfere with the hrtick preemption |
| 914 | */ |
| 915 | if (!sched_feat(DOUBLE_TICK) && |
| 916 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) |
| 917 | return; |
| 918 | #endif |
| 919 | |
| 920 | if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT)) |
| 921 | check_preempt_tick(cfs_rq, curr); |
| 922 | } |
| 923 | |
| 924 | /************************************************** |
| 925 | * CFS operations on tasks: |
| 926 | */ |
| 927 | |
| 928 | #ifdef CONFIG_SCHED_HRTICK |
| 929 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) |
| 930 | { |
| 931 | struct sched_entity *se = &p->se; |
| 932 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
| 933 | |
| 934 | WARN_ON(task_rq(p) != rq); |
| 935 | |
| 936 | if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) { |
| 937 | u64 slice = sched_slice(cfs_rq, se); |
| 938 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; |
| 939 | s64 delta = slice - ran; |
| 940 | |
| 941 | if (delta < 0) { |
| 942 | if (rq->curr == p) |
| 943 | resched_task(p); |
| 944 | return; |
| 945 | } |
| 946 | |
| 947 | /* |
| 948 | * Don't schedule slices shorter than 10000ns, that just |
| 949 | * doesn't make sense. Rely on vruntime for fairness. |
| 950 | */ |
| 951 | if (rq->curr != p) |
| 952 | delta = max_t(s64, 10000LL, delta); |
| 953 | |
| 954 | hrtick_start(rq, delta); |
| 955 | } |
| 956 | } |
| 957 | |
| 958 | /* |
| 959 | * called from enqueue/dequeue and updates the hrtick when the |
| 960 | * current task is from our class and nr_running is low enough |
| 961 | * to matter. |
| 962 | */ |
| 963 | static void hrtick_update(struct rq *rq) |
| 964 | { |
| 965 | struct task_struct *curr = rq->curr; |
| 966 | |
| 967 | if (curr->sched_class != &fair_sched_class) |
| 968 | return; |
| 969 | |
| 970 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) |
| 971 | hrtick_start_fair(rq, curr); |
| 972 | } |
| 973 | #else /* !CONFIG_SCHED_HRTICK */ |
| 974 | static inline void |
| 975 | hrtick_start_fair(struct rq *rq, struct task_struct *p) |
| 976 | { |
| 977 | } |
| 978 | |
| 979 | static inline void hrtick_update(struct rq *rq) |
| 980 | { |
| 981 | } |
| 982 | #endif |
| 983 | |
| 984 | /* |
| 985 | * The enqueue_task method is called before nr_running is |
| 986 | * increased. Here we update the fair scheduling stats and |
| 987 | * then put the task into the rbtree: |
| 988 | */ |
| 989 | static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup) |
| 990 | { |
| 991 | struct cfs_rq *cfs_rq; |
| 992 | struct sched_entity *se = &p->se; |
| 993 | |
| 994 | for_each_sched_entity(se) { |
| 995 | if (se->on_rq) |
| 996 | break; |
| 997 | cfs_rq = cfs_rq_of(se); |
| 998 | enqueue_entity(cfs_rq, se, wakeup); |
| 999 | wakeup = 1; |
| 1000 | } |
| 1001 | |
| 1002 | hrtick_update(rq); |
| 1003 | } |
| 1004 | |
| 1005 | /* |
| 1006 | * The dequeue_task method is called before nr_running is |
| 1007 | * decreased. We remove the task from the rbtree and |
| 1008 | * update the fair scheduling stats: |
| 1009 | */ |
| 1010 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep) |
| 1011 | { |
| 1012 | struct cfs_rq *cfs_rq; |
| 1013 | struct sched_entity *se = &p->se; |
| 1014 | |
| 1015 | for_each_sched_entity(se) { |
| 1016 | cfs_rq = cfs_rq_of(se); |
| 1017 | dequeue_entity(cfs_rq, se, sleep); |
| 1018 | /* Don't dequeue parent if it has other entities besides us */ |
| 1019 | if (cfs_rq->load.weight) |
| 1020 | break; |
| 1021 | sleep = 1; |
| 1022 | } |
| 1023 | |
| 1024 | hrtick_update(rq); |
| 1025 | } |
| 1026 | |
| 1027 | /* |
| 1028 | * sched_yield() support is very simple - we dequeue and enqueue. |
| 1029 | * |
| 1030 | * If compat_yield is turned on then we requeue to the end of the tree. |
| 1031 | */ |
| 1032 | static void yield_task_fair(struct rq *rq) |
| 1033 | { |
| 1034 | struct task_struct *curr = rq->curr; |
| 1035 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
| 1036 | struct sched_entity *rightmost, *se = &curr->se; |
| 1037 | |
| 1038 | /* |
| 1039 | * Are we the only task in the tree? |
| 1040 | */ |
| 1041 | if (unlikely(cfs_rq->nr_running == 1)) |
| 1042 | return; |
| 1043 | |
| 1044 | clear_buddies(cfs_rq, se); |
| 1045 | |
| 1046 | if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) { |
| 1047 | update_rq_clock(rq); |
| 1048 | /* |
| 1049 | * Update run-time statistics of the 'current'. |
| 1050 | */ |
| 1051 | update_curr(cfs_rq); |
| 1052 | |
| 1053 | return; |
| 1054 | } |
| 1055 | /* |
| 1056 | * Find the rightmost entry in the rbtree: |
| 1057 | */ |
| 1058 | rightmost = __pick_last_entity(cfs_rq); |
| 1059 | /* |
| 1060 | * Already in the rightmost position? |
| 1061 | */ |
| 1062 | if (unlikely(!rightmost || entity_before(rightmost, se))) |
| 1063 | return; |
| 1064 | |
| 1065 | /* |
| 1066 | * Minimally necessary key value to be last in the tree: |
| 1067 | * Upon rescheduling, sched_class::put_prev_task() will place |
| 1068 | * 'current' within the tree based on its new key value. |
| 1069 | */ |
| 1070 | se->vruntime = rightmost->vruntime + 1; |
| 1071 | } |
| 1072 | |
| 1073 | #ifdef CONFIG_SMP |
| 1074 | |
| 1075 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 1076 | /* |
| 1077 | * effective_load() calculates the load change as seen from the root_task_group |
| 1078 | * |
| 1079 | * Adding load to a group doesn't make a group heavier, but can cause movement |
| 1080 | * of group shares between cpus. Assuming the shares were perfectly aligned one |
| 1081 | * can calculate the shift in shares. |
| 1082 | * |
| 1083 | * The problem is that perfectly aligning the shares is rather expensive, hence |
| 1084 | * we try to avoid doing that too often - see update_shares(), which ratelimits |
| 1085 | * this change. |
| 1086 | * |
| 1087 | * We compensate this by not only taking the current delta into account, but |
| 1088 | * also considering the delta between when the shares were last adjusted and |
| 1089 | * now. |
| 1090 | * |
| 1091 | * We still saw a performance dip, some tracing learned us that between |
| 1092 | * cgroup:/ and cgroup:/foo balancing the number of affine wakeups increased |
| 1093 | * significantly. Therefore try to bias the error in direction of failing |
| 1094 | * the affine wakeup. |
| 1095 | * |
| 1096 | */ |
| 1097 | static long effective_load(struct task_group *tg, int cpu, |
| 1098 | long wl, long wg) |
| 1099 | { |
| 1100 | struct sched_entity *se = tg->se[cpu]; |
| 1101 | |
| 1102 | if (!tg->parent) |
| 1103 | return wl; |
| 1104 | |
| 1105 | /* |
| 1106 | * By not taking the decrease of shares on the other cpu into |
| 1107 | * account our error leans towards reducing the affine wakeups. |
| 1108 | */ |
| 1109 | if (!wl && sched_feat(ASYM_EFF_LOAD)) |
| 1110 | return wl; |
| 1111 | |
| 1112 | for_each_sched_entity(se) { |
| 1113 | long S, rw, s, a, b; |
| 1114 | long more_w; |
| 1115 | |
| 1116 | /* |
| 1117 | * Instead of using this increment, also add the difference |
| 1118 | * between when the shares were last updated and now. |
| 1119 | */ |
| 1120 | more_w = se->my_q->load.weight - se->my_q->rq_weight; |
| 1121 | wl += more_w; |
| 1122 | wg += more_w; |
| 1123 | |
| 1124 | S = se->my_q->tg->shares; |
| 1125 | s = se->my_q->shares; |
| 1126 | rw = se->my_q->rq_weight; |
| 1127 | |
| 1128 | a = S*(rw + wl); |
| 1129 | b = S*rw + s*wg; |
| 1130 | |
| 1131 | wl = s*(a-b); |
| 1132 | |
| 1133 | if (likely(b)) |
| 1134 | wl /= b; |
| 1135 | |
| 1136 | /* |
| 1137 | * Assume the group is already running and will |
| 1138 | * thus already be accounted for in the weight. |
| 1139 | * |
| 1140 | * That is, moving shares between CPUs, does not |
| 1141 | * alter the group weight. |
| 1142 | */ |
| 1143 | wg = 0; |
| 1144 | } |
| 1145 | |
| 1146 | return wl; |
| 1147 | } |
| 1148 | |
| 1149 | #else |
| 1150 | |
| 1151 | static inline unsigned long effective_load(struct task_group *tg, int cpu, |
| 1152 | unsigned long wl, unsigned long wg) |
| 1153 | { |
| 1154 | return wl; |
| 1155 | } |
| 1156 | |
| 1157 | #endif |
| 1158 | |
| 1159 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync) |
| 1160 | { |
| 1161 | struct task_struct *curr = current; |
| 1162 | unsigned long this_load, load; |
| 1163 | int idx, this_cpu, prev_cpu; |
| 1164 | unsigned long tl_per_task; |
| 1165 | unsigned int imbalance; |
| 1166 | struct task_group *tg; |
| 1167 | unsigned long weight; |
| 1168 | int balanced; |
| 1169 | |
| 1170 | idx = sd->wake_idx; |
| 1171 | this_cpu = smp_processor_id(); |
| 1172 | prev_cpu = task_cpu(p); |
| 1173 | load = source_load(prev_cpu, idx); |
| 1174 | this_load = target_load(this_cpu, idx); |
| 1175 | |
| 1176 | if (sync) { |
| 1177 | if (sched_feat(SYNC_LESS) && |
| 1178 | (curr->se.avg_overlap > sysctl_sched_migration_cost || |
| 1179 | p->se.avg_overlap > sysctl_sched_migration_cost)) |
| 1180 | sync = 0; |
| 1181 | } else { |
| 1182 | if (sched_feat(SYNC_MORE) && |
| 1183 | (curr->se.avg_overlap < sysctl_sched_migration_cost && |
| 1184 | p->se.avg_overlap < sysctl_sched_migration_cost)) |
| 1185 | sync = 1; |
| 1186 | } |
| 1187 | |
| 1188 | /* |
| 1189 | * If sync wakeup then subtract the (maximum possible) |
| 1190 | * effect of the currently running task from the load |
| 1191 | * of the current CPU: |
| 1192 | */ |
| 1193 | if (sync) { |
| 1194 | tg = task_group(current); |
| 1195 | weight = current->se.load.weight; |
| 1196 | |
| 1197 | this_load += effective_load(tg, this_cpu, -weight, -weight); |
| 1198 | load += effective_load(tg, prev_cpu, 0, -weight); |
| 1199 | } |
| 1200 | |
| 1201 | tg = task_group(p); |
| 1202 | weight = p->se.load.weight; |
| 1203 | |
| 1204 | imbalance = 100 + (sd->imbalance_pct - 100) / 2; |
| 1205 | |
| 1206 | /* |
| 1207 | * In low-load situations, where prev_cpu is idle and this_cpu is idle |
| 1208 | * due to the sync cause above having dropped this_load to 0, we'll |
| 1209 | * always have an imbalance, but there's really nothing you can do |
| 1210 | * about that, so that's good too. |
| 1211 | * |
| 1212 | * Otherwise check if either cpus are near enough in load to allow this |
| 1213 | * task to be woken on this_cpu. |
| 1214 | */ |
| 1215 | balanced = !this_load || |
| 1216 | 100*(this_load + effective_load(tg, this_cpu, weight, weight)) <= |
| 1217 | imbalance*(load + effective_load(tg, prev_cpu, 0, weight)); |
| 1218 | |
| 1219 | /* |
| 1220 | * If the currently running task will sleep within |
| 1221 | * a reasonable amount of time then attract this newly |
| 1222 | * woken task: |
| 1223 | */ |
| 1224 | if (sync && balanced) |
| 1225 | return 1; |
| 1226 | |
| 1227 | schedstat_inc(p, se.nr_wakeups_affine_attempts); |
| 1228 | tl_per_task = cpu_avg_load_per_task(this_cpu); |
| 1229 | |
| 1230 | if (balanced || |
| 1231 | (this_load <= load && |
| 1232 | this_load + target_load(prev_cpu, idx) <= tl_per_task)) { |
| 1233 | /* |
| 1234 | * This domain has SD_WAKE_AFFINE and |
| 1235 | * p is cache cold in this domain, and |
| 1236 | * there is no bad imbalance. |
| 1237 | */ |
| 1238 | schedstat_inc(sd, ttwu_move_affine); |
| 1239 | schedstat_inc(p, se.nr_wakeups_affine); |
| 1240 | |
| 1241 | return 1; |
| 1242 | } |
| 1243 | return 0; |
| 1244 | } |
| 1245 | |
| 1246 | /* |
| 1247 | * find_idlest_group finds and returns the least busy CPU group within the |
| 1248 | * domain. |
| 1249 | */ |
| 1250 | static struct sched_group * |
| 1251 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, |
| 1252 | int this_cpu, int load_idx) |
| 1253 | { |
| 1254 | struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups; |
| 1255 | unsigned long min_load = ULONG_MAX, this_load = 0; |
| 1256 | int imbalance = 100 + (sd->imbalance_pct-100)/2; |
| 1257 | |
| 1258 | do { |
| 1259 | unsigned long load, avg_load; |
| 1260 | int local_group; |
| 1261 | int i; |
| 1262 | |
| 1263 | /* Skip over this group if it has no CPUs allowed */ |
| 1264 | if (!cpumask_intersects(sched_group_cpus(group), |
| 1265 | &p->cpus_allowed)) |
| 1266 | continue; |
| 1267 | |
| 1268 | local_group = cpumask_test_cpu(this_cpu, |
| 1269 | sched_group_cpus(group)); |
| 1270 | |
| 1271 | /* Tally up the load of all CPUs in the group */ |
| 1272 | avg_load = 0; |
| 1273 | |
| 1274 | for_each_cpu(i, sched_group_cpus(group)) { |
| 1275 | /* Bias balancing toward cpus of our domain */ |
| 1276 | if (local_group) |
| 1277 | load = source_load(i, load_idx); |
| 1278 | else |
| 1279 | load = target_load(i, load_idx); |
| 1280 | |
| 1281 | avg_load += load; |
| 1282 | } |
| 1283 | |
| 1284 | /* Adjust by relative CPU power of the group */ |
| 1285 | avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power; |
| 1286 | |
| 1287 | if (local_group) { |
| 1288 | this_load = avg_load; |
| 1289 | this = group; |
| 1290 | } else if (avg_load < min_load) { |
| 1291 | min_load = avg_load; |
| 1292 | idlest = group; |
| 1293 | } |
| 1294 | } while (group = group->next, group != sd->groups); |
| 1295 | |
| 1296 | if (!idlest || 100*this_load < imbalance*min_load) |
| 1297 | return NULL; |
| 1298 | return idlest; |
| 1299 | } |
| 1300 | |
| 1301 | /* |
| 1302 | * find_idlest_cpu - find the idlest cpu among the cpus in group. |
| 1303 | */ |
| 1304 | static int |
| 1305 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) |
| 1306 | { |
| 1307 | unsigned long load, min_load = ULONG_MAX; |
| 1308 | int idlest = -1; |
| 1309 | int i; |
| 1310 | |
| 1311 | /* Traverse only the allowed CPUs */ |
| 1312 | for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) { |
| 1313 | load = weighted_cpuload(i); |
| 1314 | |
| 1315 | if (load < min_load || (load == min_load && i == this_cpu)) { |
| 1316 | min_load = load; |
| 1317 | idlest = i; |
| 1318 | } |
| 1319 | } |
| 1320 | |
| 1321 | return idlest; |
| 1322 | } |
| 1323 | |
| 1324 | /* |
| 1325 | * sched_balance_self: balance the current task (running on cpu) in domains |
| 1326 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and |
| 1327 | * SD_BALANCE_EXEC. |
| 1328 | * |
| 1329 | * Balance, ie. select the least loaded group. |
| 1330 | * |
| 1331 | * Returns the target CPU number, or the same CPU if no balancing is needed. |
| 1332 | * |
| 1333 | * preempt must be disabled. |
| 1334 | */ |
| 1335 | static int select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags) |
| 1336 | { |
| 1337 | struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL; |
| 1338 | int cpu = smp_processor_id(); |
| 1339 | int prev_cpu = task_cpu(p); |
| 1340 | int new_cpu = cpu; |
| 1341 | int want_affine = 0; |
| 1342 | int want_sd = 1; |
| 1343 | int sync = wake_flags & WF_SYNC; |
| 1344 | |
| 1345 | if (sd_flag & SD_BALANCE_WAKE) { |
| 1346 | if (sched_feat(AFFINE_WAKEUPS)) |
| 1347 | want_affine = 1; |
| 1348 | new_cpu = prev_cpu; |
| 1349 | } |
| 1350 | |
| 1351 | rcu_read_lock(); |
| 1352 | for_each_domain(cpu, tmp) { |
| 1353 | /* |
| 1354 | * If power savings logic is enabled for a domain, see if we |
| 1355 | * are not overloaded, if so, don't balance wider. |
| 1356 | */ |
| 1357 | if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) { |
| 1358 | unsigned long power = 0; |
| 1359 | unsigned long nr_running = 0; |
| 1360 | unsigned long capacity; |
| 1361 | int i; |
| 1362 | |
| 1363 | for_each_cpu(i, sched_domain_span(tmp)) { |
| 1364 | power += power_of(i); |
| 1365 | nr_running += cpu_rq(i)->cfs.nr_running; |
| 1366 | } |
| 1367 | |
| 1368 | capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE); |
| 1369 | |
| 1370 | if (tmp->flags & SD_POWERSAVINGS_BALANCE) |
| 1371 | nr_running /= 2; |
| 1372 | |
| 1373 | if (nr_running < capacity) |
| 1374 | want_sd = 0; |
| 1375 | } |
| 1376 | |
| 1377 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && |
| 1378 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { |
| 1379 | |
| 1380 | affine_sd = tmp; |
| 1381 | want_affine = 0; |
| 1382 | } |
| 1383 | |
| 1384 | if (!want_sd && !want_affine) |
| 1385 | break; |
| 1386 | |
| 1387 | if (!(tmp->flags & sd_flag)) |
| 1388 | continue; |
| 1389 | |
| 1390 | if (want_sd) |
| 1391 | sd = tmp; |
| 1392 | } |
| 1393 | |
| 1394 | if (sched_feat(LB_SHARES_UPDATE)) { |
| 1395 | /* |
| 1396 | * Pick the largest domain to update shares over |
| 1397 | */ |
| 1398 | tmp = sd; |
| 1399 | if (affine_sd && (!tmp || |
| 1400 | cpumask_weight(sched_domain_span(affine_sd)) > |
| 1401 | cpumask_weight(sched_domain_span(sd)))) |
| 1402 | tmp = affine_sd; |
| 1403 | |
| 1404 | if (tmp) |
| 1405 | update_shares(tmp); |
| 1406 | } |
| 1407 | |
| 1408 | if (affine_sd && wake_affine(affine_sd, p, sync)) { |
| 1409 | new_cpu = cpu; |
| 1410 | goto out; |
| 1411 | } |
| 1412 | |
| 1413 | while (sd) { |
| 1414 | int load_idx = sd->forkexec_idx; |
| 1415 | struct sched_group *group; |
| 1416 | int weight; |
| 1417 | |
| 1418 | if (!(sd->flags & sd_flag)) { |
| 1419 | sd = sd->child; |
| 1420 | continue; |
| 1421 | } |
| 1422 | |
| 1423 | if (sd_flag & SD_BALANCE_WAKE) |
| 1424 | load_idx = sd->wake_idx; |
| 1425 | |
| 1426 | group = find_idlest_group(sd, p, cpu, load_idx); |
| 1427 | if (!group) { |
| 1428 | sd = sd->child; |
| 1429 | continue; |
| 1430 | } |
| 1431 | |
| 1432 | new_cpu = find_idlest_cpu(group, p, cpu); |
| 1433 | if (new_cpu == -1 || new_cpu == cpu) { |
| 1434 | /* Now try balancing at a lower domain level of cpu */ |
| 1435 | sd = sd->child; |
| 1436 | continue; |
| 1437 | } |
| 1438 | |
| 1439 | /* Now try balancing at a lower domain level of new_cpu */ |
| 1440 | cpu = new_cpu; |
| 1441 | weight = cpumask_weight(sched_domain_span(sd)); |
| 1442 | sd = NULL; |
| 1443 | for_each_domain(cpu, tmp) { |
| 1444 | if (weight <= cpumask_weight(sched_domain_span(tmp))) |
| 1445 | break; |
| 1446 | if (tmp->flags & sd_flag) |
| 1447 | sd = tmp; |
| 1448 | } |
| 1449 | /* while loop will break here if sd == NULL */ |
| 1450 | } |
| 1451 | |
| 1452 | out: |
| 1453 | rcu_read_unlock(); |
| 1454 | return new_cpu; |
| 1455 | } |
| 1456 | #endif /* CONFIG_SMP */ |
| 1457 | |
| 1458 | /* |
| 1459 | * Adaptive granularity |
| 1460 | * |
| 1461 | * se->avg_wakeup gives the average time a task runs until it does a wakeup, |
| 1462 | * with the limit of wakeup_gran -- when it never does a wakeup. |
| 1463 | * |
| 1464 | * So the smaller avg_wakeup is the faster we want this task to preempt, |
| 1465 | * but we don't want to treat the preemptee unfairly and therefore allow it |
| 1466 | * to run for at least the amount of time we'd like to run. |
| 1467 | * |
| 1468 | * NOTE: we use 2*avg_wakeup to increase the probability of actually doing one |
| 1469 | * |
| 1470 | * NOTE: we use *nr_running to scale with load, this nicely matches the |
| 1471 | * degrading latency on load. |
| 1472 | */ |
| 1473 | static unsigned long |
| 1474 | adaptive_gran(struct sched_entity *curr, struct sched_entity *se) |
| 1475 | { |
| 1476 | u64 this_run = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
| 1477 | u64 expected_wakeup = 2*se->avg_wakeup * cfs_rq_of(se)->nr_running; |
| 1478 | u64 gran = 0; |
| 1479 | |
| 1480 | if (this_run < expected_wakeup) |
| 1481 | gran = expected_wakeup - this_run; |
| 1482 | |
| 1483 | return min_t(s64, gran, sysctl_sched_wakeup_granularity); |
| 1484 | } |
| 1485 | |
| 1486 | static unsigned long |
| 1487 | wakeup_gran(struct sched_entity *curr, struct sched_entity *se) |
| 1488 | { |
| 1489 | unsigned long gran = sysctl_sched_wakeup_granularity; |
| 1490 | |
| 1491 | if (cfs_rq_of(curr)->curr && sched_feat(ADAPTIVE_GRAN)) |
| 1492 | gran = adaptive_gran(curr, se); |
| 1493 | |
| 1494 | /* |
| 1495 | * Since its curr running now, convert the gran from real-time |
| 1496 | * to virtual-time in his units. |
| 1497 | */ |
| 1498 | if (sched_feat(ASYM_GRAN)) { |
| 1499 | /* |
| 1500 | * By using 'se' instead of 'curr' we penalize light tasks, so |
| 1501 | * they get preempted easier. That is, if 'se' < 'curr' then |
| 1502 | * the resulting gran will be larger, therefore penalizing the |
| 1503 | * lighter, if otoh 'se' > 'curr' then the resulting gran will |
| 1504 | * be smaller, again penalizing the lighter task. |
| 1505 | * |
| 1506 | * This is especially important for buddies when the leftmost |
| 1507 | * task is higher priority than the buddy. |
| 1508 | */ |
| 1509 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
| 1510 | gran = calc_delta_fair(gran, se); |
| 1511 | } else { |
| 1512 | if (unlikely(curr->load.weight != NICE_0_LOAD)) |
| 1513 | gran = calc_delta_fair(gran, curr); |
| 1514 | } |
| 1515 | |
| 1516 | return gran; |
| 1517 | } |
| 1518 | |
| 1519 | /* |
| 1520 | * Should 'se' preempt 'curr'. |
| 1521 | * |
| 1522 | * |s1 |
| 1523 | * |s2 |
| 1524 | * |s3 |
| 1525 | * g |
| 1526 | * |<--->|c |
| 1527 | * |
| 1528 | * w(c, s1) = -1 |
| 1529 | * w(c, s2) = 0 |
| 1530 | * w(c, s3) = 1 |
| 1531 | * |
| 1532 | */ |
| 1533 | static int |
| 1534 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) |
| 1535 | { |
| 1536 | s64 gran, vdiff = curr->vruntime - se->vruntime; |
| 1537 | |
| 1538 | if (vdiff <= 0) |
| 1539 | return -1; |
| 1540 | |
| 1541 | gran = wakeup_gran(curr, se); |
| 1542 | if (vdiff > gran) |
| 1543 | return 1; |
| 1544 | |
| 1545 | return 0; |
| 1546 | } |
| 1547 | |
| 1548 | static void set_last_buddy(struct sched_entity *se) |
| 1549 | { |
| 1550 | if (likely(task_of(se)->policy != SCHED_IDLE)) { |
| 1551 | for_each_sched_entity(se) |
| 1552 | cfs_rq_of(se)->last = se; |
| 1553 | } |
| 1554 | } |
| 1555 | |
| 1556 | static void set_next_buddy(struct sched_entity *se) |
| 1557 | { |
| 1558 | if (likely(task_of(se)->policy != SCHED_IDLE)) { |
| 1559 | for_each_sched_entity(se) |
| 1560 | cfs_rq_of(se)->next = se; |
| 1561 | } |
| 1562 | } |
| 1563 | |
| 1564 | /* |
| 1565 | * Preempt the current task with a newly woken task if needed: |
| 1566 | */ |
| 1567 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
| 1568 | { |
| 1569 | struct task_struct *curr = rq->curr; |
| 1570 | struct sched_entity *se = &curr->se, *pse = &p->se; |
| 1571 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
| 1572 | int sync = wake_flags & WF_SYNC; |
| 1573 | |
| 1574 | update_curr(cfs_rq); |
| 1575 | |
| 1576 | if (unlikely(rt_prio(p->prio))) { |
| 1577 | resched_task(curr); |
| 1578 | return; |
| 1579 | } |
| 1580 | |
| 1581 | if (unlikely(p->sched_class != &fair_sched_class)) |
| 1582 | return; |
| 1583 | |
| 1584 | if (unlikely(se == pse)) |
| 1585 | return; |
| 1586 | |
| 1587 | /* |
| 1588 | * Only set the backward buddy when the current task is still on the |
| 1589 | * rq. This can happen when a wakeup gets interleaved with schedule on |
| 1590 | * the ->pre_schedule() or idle_balance() point, either of which can |
| 1591 | * drop the rq lock. |
| 1592 | * |
| 1593 | * Also, during early boot the idle thread is in the fair class, for |
| 1594 | * obvious reasons its a bad idea to schedule back to the idle thread. |
| 1595 | */ |
| 1596 | if (sched_feat(LAST_BUDDY) && likely(se->on_rq && curr != rq->idle)) |
| 1597 | set_last_buddy(se); |
| 1598 | if (sched_feat(NEXT_BUDDY) && !(wake_flags & WF_FORK)) |
| 1599 | set_next_buddy(pse); |
| 1600 | |
| 1601 | /* |
| 1602 | * We can come here with TIF_NEED_RESCHED already set from new task |
| 1603 | * wake up path. |
| 1604 | */ |
| 1605 | if (test_tsk_need_resched(curr)) |
| 1606 | return; |
| 1607 | |
| 1608 | /* |
| 1609 | * Batch and idle tasks do not preempt (their preemption is driven by |
| 1610 | * the tick): |
| 1611 | */ |
| 1612 | if (unlikely(p->policy != SCHED_NORMAL)) |
| 1613 | return; |
| 1614 | |
| 1615 | /* Idle tasks are by definition preempted by everybody. */ |
| 1616 | if (unlikely(curr->policy == SCHED_IDLE)) { |
| 1617 | resched_task(curr); |
| 1618 | return; |
| 1619 | } |
| 1620 | |
| 1621 | if ((sched_feat(WAKEUP_SYNC) && sync) || |
| 1622 | (sched_feat(WAKEUP_OVERLAP) && |
| 1623 | (se->avg_overlap < sysctl_sched_migration_cost && |
| 1624 | pse->avg_overlap < sysctl_sched_migration_cost))) { |
| 1625 | resched_task(curr); |
| 1626 | return; |
| 1627 | } |
| 1628 | |
| 1629 | if (sched_feat(WAKEUP_RUNNING)) { |
| 1630 | if (pse->avg_running < se->avg_running) { |
| 1631 | set_next_buddy(pse); |
| 1632 | resched_task(curr); |
| 1633 | return; |
| 1634 | } |
| 1635 | } |
| 1636 | |
| 1637 | if (!sched_feat(WAKEUP_PREEMPT)) |
| 1638 | return; |
| 1639 | |
| 1640 | find_matching_se(&se, &pse); |
| 1641 | |
| 1642 | BUG_ON(!pse); |
| 1643 | |
| 1644 | if (wakeup_preempt_entity(se, pse) == 1) |
| 1645 | resched_task(curr); |
| 1646 | } |
| 1647 | |
| 1648 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
| 1649 | { |
| 1650 | struct task_struct *p; |
| 1651 | struct cfs_rq *cfs_rq = &rq->cfs; |
| 1652 | struct sched_entity *se; |
| 1653 | |
| 1654 | if (unlikely(!cfs_rq->nr_running)) |
| 1655 | return NULL; |
| 1656 | |
| 1657 | do { |
| 1658 | se = pick_next_entity(cfs_rq); |
| 1659 | /* |
| 1660 | * If se was a buddy, clear it so that it will have to earn |
| 1661 | * the favour again. |
| 1662 | * |
| 1663 | * If se was not a buddy, clear the buddies because neither |
| 1664 | * was elegible to run, let them earn it again. |
| 1665 | * |
| 1666 | * IOW. unconditionally clear buddies. |
| 1667 | */ |
| 1668 | __clear_buddies(cfs_rq, NULL); |
| 1669 | set_next_entity(cfs_rq, se); |
| 1670 | cfs_rq = group_cfs_rq(se); |
| 1671 | } while (cfs_rq); |
| 1672 | |
| 1673 | p = task_of(se); |
| 1674 | hrtick_start_fair(rq, p); |
| 1675 | |
| 1676 | return p; |
| 1677 | } |
| 1678 | |
| 1679 | /* |
| 1680 | * Account for a descheduled task: |
| 1681 | */ |
| 1682 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
| 1683 | { |
| 1684 | struct sched_entity *se = &prev->se; |
| 1685 | struct cfs_rq *cfs_rq; |
| 1686 | |
| 1687 | for_each_sched_entity(se) { |
| 1688 | cfs_rq = cfs_rq_of(se); |
| 1689 | put_prev_entity(cfs_rq, se); |
| 1690 | } |
| 1691 | } |
| 1692 | |
| 1693 | #ifdef CONFIG_SMP |
| 1694 | /************************************************** |
| 1695 | * Fair scheduling class load-balancing methods: |
| 1696 | */ |
| 1697 | |
| 1698 | /* |
| 1699 | * Load-balancing iterator. Note: while the runqueue stays locked |
| 1700 | * during the whole iteration, the current task might be |
| 1701 | * dequeued so the iterator has to be dequeue-safe. Here we |
| 1702 | * achieve that by always pre-iterating before returning |
| 1703 | * the current task: |
| 1704 | */ |
| 1705 | static struct task_struct * |
| 1706 | __load_balance_iterator(struct cfs_rq *cfs_rq, struct list_head *next) |
| 1707 | { |
| 1708 | struct task_struct *p = NULL; |
| 1709 | struct sched_entity *se; |
| 1710 | |
| 1711 | if (next == &cfs_rq->tasks) |
| 1712 | return NULL; |
| 1713 | |
| 1714 | se = list_entry(next, struct sched_entity, group_node); |
| 1715 | p = task_of(se); |
| 1716 | cfs_rq->balance_iterator = next->next; |
| 1717 | |
| 1718 | return p; |
| 1719 | } |
| 1720 | |
| 1721 | static struct task_struct *load_balance_start_fair(void *arg) |
| 1722 | { |
| 1723 | struct cfs_rq *cfs_rq = arg; |
| 1724 | |
| 1725 | return __load_balance_iterator(cfs_rq, cfs_rq->tasks.next); |
| 1726 | } |
| 1727 | |
| 1728 | static struct task_struct *load_balance_next_fair(void *arg) |
| 1729 | { |
| 1730 | struct cfs_rq *cfs_rq = arg; |
| 1731 | |
| 1732 | return __load_balance_iterator(cfs_rq, cfs_rq->balance_iterator); |
| 1733 | } |
| 1734 | |
| 1735 | static unsigned long |
| 1736 | __load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, |
| 1737 | unsigned long max_load_move, struct sched_domain *sd, |
| 1738 | enum cpu_idle_type idle, int *all_pinned, int *this_best_prio, |
| 1739 | struct cfs_rq *cfs_rq) |
| 1740 | { |
| 1741 | struct rq_iterator cfs_rq_iterator; |
| 1742 | |
| 1743 | cfs_rq_iterator.start = load_balance_start_fair; |
| 1744 | cfs_rq_iterator.next = load_balance_next_fair; |
| 1745 | cfs_rq_iterator.arg = cfs_rq; |
| 1746 | |
| 1747 | return balance_tasks(this_rq, this_cpu, busiest, |
| 1748 | max_load_move, sd, idle, all_pinned, |
| 1749 | this_best_prio, &cfs_rq_iterator); |
| 1750 | } |
| 1751 | |
| 1752 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 1753 | static unsigned long |
| 1754 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, |
| 1755 | unsigned long max_load_move, |
| 1756 | struct sched_domain *sd, enum cpu_idle_type idle, |
| 1757 | int *all_pinned, int *this_best_prio) |
| 1758 | { |
| 1759 | long rem_load_move = max_load_move; |
| 1760 | int busiest_cpu = cpu_of(busiest); |
| 1761 | struct task_group *tg; |
| 1762 | |
| 1763 | rcu_read_lock(); |
| 1764 | update_h_load(busiest_cpu); |
| 1765 | |
| 1766 | list_for_each_entry_rcu(tg, &task_groups, list) { |
| 1767 | struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu]; |
| 1768 | unsigned long busiest_h_load = busiest_cfs_rq->h_load; |
| 1769 | unsigned long busiest_weight = busiest_cfs_rq->load.weight; |
| 1770 | u64 rem_load, moved_load; |
| 1771 | |
| 1772 | /* |
| 1773 | * empty group |
| 1774 | */ |
| 1775 | if (!busiest_cfs_rq->task_weight) |
| 1776 | continue; |
| 1777 | |
| 1778 | rem_load = (u64)rem_load_move * busiest_weight; |
| 1779 | rem_load = div_u64(rem_load, busiest_h_load + 1); |
| 1780 | |
| 1781 | moved_load = __load_balance_fair(this_rq, this_cpu, busiest, |
| 1782 | rem_load, sd, idle, all_pinned, this_best_prio, |
| 1783 | tg->cfs_rq[busiest_cpu]); |
| 1784 | |
| 1785 | if (!moved_load) |
| 1786 | continue; |
| 1787 | |
| 1788 | moved_load *= busiest_h_load; |
| 1789 | moved_load = div_u64(moved_load, busiest_weight + 1); |
| 1790 | |
| 1791 | rem_load_move -= moved_load; |
| 1792 | if (rem_load_move < 0) |
| 1793 | break; |
| 1794 | } |
| 1795 | rcu_read_unlock(); |
| 1796 | |
| 1797 | return max_load_move - rem_load_move; |
| 1798 | } |
| 1799 | #else |
| 1800 | static unsigned long |
| 1801 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, |
| 1802 | unsigned long max_load_move, |
| 1803 | struct sched_domain *sd, enum cpu_idle_type idle, |
| 1804 | int *all_pinned, int *this_best_prio) |
| 1805 | { |
| 1806 | return __load_balance_fair(this_rq, this_cpu, busiest, |
| 1807 | max_load_move, sd, idle, all_pinned, |
| 1808 | this_best_prio, &busiest->cfs); |
| 1809 | } |
| 1810 | #endif |
| 1811 | |
| 1812 | static int |
| 1813 | move_one_task_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, |
| 1814 | struct sched_domain *sd, enum cpu_idle_type idle) |
| 1815 | { |
| 1816 | struct cfs_rq *busy_cfs_rq; |
| 1817 | struct rq_iterator cfs_rq_iterator; |
| 1818 | |
| 1819 | cfs_rq_iterator.start = load_balance_start_fair; |
| 1820 | cfs_rq_iterator.next = load_balance_next_fair; |
| 1821 | |
| 1822 | for_each_leaf_cfs_rq(busiest, busy_cfs_rq) { |
| 1823 | /* |
| 1824 | * pass busy_cfs_rq argument into |
| 1825 | * load_balance_[start|next]_fair iterators |
| 1826 | */ |
| 1827 | cfs_rq_iterator.arg = busy_cfs_rq; |
| 1828 | if (iter_move_one_task(this_rq, this_cpu, busiest, sd, idle, |
| 1829 | &cfs_rq_iterator)) |
| 1830 | return 1; |
| 1831 | } |
| 1832 | |
| 1833 | return 0; |
| 1834 | } |
| 1835 | #endif /* CONFIG_SMP */ |
| 1836 | |
| 1837 | /* |
| 1838 | * scheduler tick hitting a task of our scheduling class: |
| 1839 | */ |
| 1840 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
| 1841 | { |
| 1842 | struct cfs_rq *cfs_rq; |
| 1843 | struct sched_entity *se = &curr->se; |
| 1844 | |
| 1845 | for_each_sched_entity(se) { |
| 1846 | cfs_rq = cfs_rq_of(se); |
| 1847 | entity_tick(cfs_rq, se, queued); |
| 1848 | } |
| 1849 | } |
| 1850 | |
| 1851 | /* |
| 1852 | * Share the fairness runtime between parent and child, thus the |
| 1853 | * total amount of pressure for CPU stays equal - new tasks |
| 1854 | * get a chance to run but frequent forkers are not allowed to |
| 1855 | * monopolize the CPU. Note: the parent runqueue is locked, |
| 1856 | * the child is not running yet. |
| 1857 | */ |
| 1858 | static void task_new_fair(struct rq *rq, struct task_struct *p) |
| 1859 | { |
| 1860 | struct cfs_rq *cfs_rq = task_cfs_rq(p); |
| 1861 | struct sched_entity *se = &p->se, *curr = cfs_rq->curr; |
| 1862 | int this_cpu = smp_processor_id(); |
| 1863 | |
| 1864 | sched_info_queued(p); |
| 1865 | |
| 1866 | update_curr(cfs_rq); |
| 1867 | if (curr) |
| 1868 | se->vruntime = curr->vruntime; |
| 1869 | place_entity(cfs_rq, se, 1); |
| 1870 | |
| 1871 | /* 'curr' will be NULL if the child belongs to a different group */ |
| 1872 | if (sysctl_sched_child_runs_first && this_cpu == task_cpu(p) && |
| 1873 | curr && entity_before(curr, se)) { |
| 1874 | /* |
| 1875 | * Upon rescheduling, sched_class::put_prev_task() will place |
| 1876 | * 'current' within the tree based on its new key value. |
| 1877 | */ |
| 1878 | swap(curr->vruntime, se->vruntime); |
| 1879 | resched_task(rq->curr); |
| 1880 | } |
| 1881 | |
| 1882 | enqueue_task_fair(rq, p, 0); |
| 1883 | } |
| 1884 | |
| 1885 | /* |
| 1886 | * Priority of the task has changed. Check to see if we preempt |
| 1887 | * the current task. |
| 1888 | */ |
| 1889 | static void prio_changed_fair(struct rq *rq, struct task_struct *p, |
| 1890 | int oldprio, int running) |
| 1891 | { |
| 1892 | /* |
| 1893 | * Reschedule if we are currently running on this runqueue and |
| 1894 | * our priority decreased, or if we are not currently running on |
| 1895 | * this runqueue and our priority is higher than the current's |
| 1896 | */ |
| 1897 | if (running) { |
| 1898 | if (p->prio > oldprio) |
| 1899 | resched_task(rq->curr); |
| 1900 | } else |
| 1901 | check_preempt_curr(rq, p, 0); |
| 1902 | } |
| 1903 | |
| 1904 | /* |
| 1905 | * We switched to the sched_fair class. |
| 1906 | */ |
| 1907 | static void switched_to_fair(struct rq *rq, struct task_struct *p, |
| 1908 | int running) |
| 1909 | { |
| 1910 | /* |
| 1911 | * We were most likely switched from sched_rt, so |
| 1912 | * kick off the schedule if running, otherwise just see |
| 1913 | * if we can still preempt the current task. |
| 1914 | */ |
| 1915 | if (running) |
| 1916 | resched_task(rq->curr); |
| 1917 | else |
| 1918 | check_preempt_curr(rq, p, 0); |
| 1919 | } |
| 1920 | |
| 1921 | /* Account for a task changing its policy or group. |
| 1922 | * |
| 1923 | * This routine is mostly called to set cfs_rq->curr field when a task |
| 1924 | * migrates between groups/classes. |
| 1925 | */ |
| 1926 | static void set_curr_task_fair(struct rq *rq) |
| 1927 | { |
| 1928 | struct sched_entity *se = &rq->curr->se; |
| 1929 | |
| 1930 | for_each_sched_entity(se) |
| 1931 | set_next_entity(cfs_rq_of(se), se); |
| 1932 | } |
| 1933 | |
| 1934 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 1935 | static void moved_group_fair(struct task_struct *p) |
| 1936 | { |
| 1937 | struct cfs_rq *cfs_rq = task_cfs_rq(p); |
| 1938 | |
| 1939 | update_curr(cfs_rq); |
| 1940 | place_entity(cfs_rq, &p->se, 1); |
| 1941 | } |
| 1942 | #endif |
| 1943 | |
| 1944 | /* |
| 1945 | * All the scheduling class methods: |
| 1946 | */ |
| 1947 | static const struct sched_class fair_sched_class = { |
| 1948 | .next = &idle_sched_class, |
| 1949 | .enqueue_task = enqueue_task_fair, |
| 1950 | .dequeue_task = dequeue_task_fair, |
| 1951 | .yield_task = yield_task_fair, |
| 1952 | |
| 1953 | .check_preempt_curr = check_preempt_wakeup, |
| 1954 | |
| 1955 | .pick_next_task = pick_next_task_fair, |
| 1956 | .put_prev_task = put_prev_task_fair, |
| 1957 | |
| 1958 | #ifdef CONFIG_SMP |
| 1959 | .select_task_rq = select_task_rq_fair, |
| 1960 | |
| 1961 | .load_balance = load_balance_fair, |
| 1962 | .move_one_task = move_one_task_fair, |
| 1963 | #endif |
| 1964 | |
| 1965 | .set_curr_task = set_curr_task_fair, |
| 1966 | .task_tick = task_tick_fair, |
| 1967 | .task_new = task_new_fair, |
| 1968 | |
| 1969 | .prio_changed = prio_changed_fair, |
| 1970 | .switched_to = switched_to_fair, |
| 1971 | |
| 1972 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 1973 | .moved_group = moved_group_fair, |
| 1974 | #endif |
| 1975 | }; |
| 1976 | |
| 1977 | #ifdef CONFIG_SCHED_DEBUG |
| 1978 | static void print_cfs_stats(struct seq_file *m, int cpu) |
| 1979 | { |
| 1980 | struct cfs_rq *cfs_rq; |
| 1981 | |
| 1982 | rcu_read_lock(); |
| 1983 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
| 1984 | print_cfs_rq(m, cpu, cfs_rq); |
| 1985 | rcu_read_unlock(); |
| 1986 | } |
| 1987 | #endif |