| 1 | /* |
| 2 | * kernel/sched.c |
| 3 | * |
| 4 | * Kernel scheduler and related syscalls |
| 5 | * |
| 6 | * Copyright (C) 1991-2002 Linus Torvalds |
| 7 | * |
| 8 | * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and |
| 9 | * make semaphores SMP safe |
| 10 | * 1998-11-19 Implemented schedule_timeout() and related stuff |
| 11 | * by Andrea Arcangeli |
| 12 | * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar: |
| 13 | * hybrid priority-list and round-robin design with |
| 14 | * an array-switch method of distributing timeslices |
| 15 | * and per-CPU runqueues. Cleanups and useful suggestions |
| 16 | * by Davide Libenzi, preemptible kernel bits by Robert Love. |
| 17 | * 2003-09-03 Interactivity tuning by Con Kolivas. |
| 18 | * 2004-04-02 Scheduler domains code by Nick Piggin |
| 19 | * 2007-04-15 Work begun on replacing all interactivity tuning with a |
| 20 | * fair scheduling design by Con Kolivas. |
| 21 | * 2007-05-05 Load balancing (smp-nice) and other improvements |
| 22 | * by Peter Williams |
| 23 | * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith |
| 24 | * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri |
| 25 | */ |
| 26 | |
| 27 | #include <linux/mm.h> |
| 28 | #include <linux/module.h> |
| 29 | #include <linux/nmi.h> |
| 30 | #include <linux/init.h> |
| 31 | #include <linux/uaccess.h> |
| 32 | #include <linux/highmem.h> |
| 33 | #include <linux/smp_lock.h> |
| 34 | #include <asm/mmu_context.h> |
| 35 | #include <linux/interrupt.h> |
| 36 | #include <linux/capability.h> |
| 37 | #include <linux/completion.h> |
| 38 | #include <linux/kernel_stat.h> |
| 39 | #include <linux/debug_locks.h> |
| 40 | #include <linux/security.h> |
| 41 | #include <linux/notifier.h> |
| 42 | #include <linux/profile.h> |
| 43 | #include <linux/freezer.h> |
| 44 | #include <linux/vmalloc.h> |
| 45 | #include <linux/blkdev.h> |
| 46 | #include <linux/delay.h> |
| 47 | #include <linux/pid_namespace.h> |
| 48 | #include <linux/smp.h> |
| 49 | #include <linux/threads.h> |
| 50 | #include <linux/timer.h> |
| 51 | #include <linux/rcupdate.h> |
| 52 | #include <linux/cpu.h> |
| 53 | #include <linux/cpuset.h> |
| 54 | #include <linux/percpu.h> |
| 55 | #include <linux/kthread.h> |
| 56 | #include <linux/seq_file.h> |
| 57 | #include <linux/sysctl.h> |
| 58 | #include <linux/syscalls.h> |
| 59 | #include <linux/times.h> |
| 60 | #include <linux/tsacct_kern.h> |
| 61 | #include <linux/kprobes.h> |
| 62 | #include <linux/delayacct.h> |
| 63 | #include <linux/reciprocal_div.h> |
| 64 | #include <linux/unistd.h> |
| 65 | #include <linux/pagemap.h> |
| 66 | |
| 67 | #include <asm/tlb.h> |
| 68 | #include <asm/irq_regs.h> |
| 69 | |
| 70 | /* |
| 71 | * Scheduler clock - returns current time in nanosec units. |
| 72 | * This is default implementation. |
| 73 | * Architectures and sub-architectures can override this. |
| 74 | */ |
| 75 | unsigned long long __attribute__((weak)) sched_clock(void) |
| 76 | { |
| 77 | return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ); |
| 78 | } |
| 79 | |
| 80 | /* |
| 81 | * Convert user-nice values [ -20 ... 0 ... 19 ] |
| 82 | * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ], |
| 83 | * and back. |
| 84 | */ |
| 85 | #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20) |
| 86 | #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20) |
| 87 | #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio) |
| 88 | |
| 89 | /* |
| 90 | * 'User priority' is the nice value converted to something we |
| 91 | * can work with better when scaling various scheduler parameters, |
| 92 | * it's a [ 0 ... 39 ] range. |
| 93 | */ |
| 94 | #define USER_PRIO(p) ((p)-MAX_RT_PRIO) |
| 95 | #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio) |
| 96 | #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO)) |
| 97 | |
| 98 | /* |
| 99 | * Some helpers for converting nanosecond timing to jiffy resolution |
| 100 | */ |
| 101 | #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ)) |
| 102 | #define JIFFIES_TO_NS(TIME) ((TIME) * (NSEC_PER_SEC / HZ)) |
| 103 | |
| 104 | #define NICE_0_LOAD SCHED_LOAD_SCALE |
| 105 | #define NICE_0_SHIFT SCHED_LOAD_SHIFT |
| 106 | |
| 107 | /* |
| 108 | * These are the 'tuning knobs' of the scheduler: |
| 109 | * |
| 110 | * default timeslice is 100 msecs (used only for SCHED_RR tasks). |
| 111 | * Timeslices get refilled after they expire. |
| 112 | */ |
| 113 | #define DEF_TIMESLICE (100 * HZ / 1000) |
| 114 | |
| 115 | #ifdef CONFIG_SMP |
| 116 | /* |
| 117 | * Divide a load by a sched group cpu_power : (load / sg->__cpu_power) |
| 118 | * Since cpu_power is a 'constant', we can use a reciprocal divide. |
| 119 | */ |
| 120 | static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load) |
| 121 | { |
| 122 | return reciprocal_divide(load, sg->reciprocal_cpu_power); |
| 123 | } |
| 124 | |
| 125 | /* |
| 126 | * Each time a sched group cpu_power is changed, |
| 127 | * we must compute its reciprocal value |
| 128 | */ |
| 129 | static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val) |
| 130 | { |
| 131 | sg->__cpu_power += val; |
| 132 | sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power); |
| 133 | } |
| 134 | #endif |
| 135 | |
| 136 | static inline int rt_policy(int policy) |
| 137 | { |
| 138 | if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR)) |
| 139 | return 1; |
| 140 | return 0; |
| 141 | } |
| 142 | |
| 143 | static inline int task_has_rt_policy(struct task_struct *p) |
| 144 | { |
| 145 | return rt_policy(p->policy); |
| 146 | } |
| 147 | |
| 148 | /* |
| 149 | * This is the priority-queue data structure of the RT scheduling class: |
| 150 | */ |
| 151 | struct rt_prio_array { |
| 152 | DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */ |
| 153 | struct list_head queue[MAX_RT_PRIO]; |
| 154 | }; |
| 155 | |
| 156 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 157 | |
| 158 | #include <linux/cgroup.h> |
| 159 | |
| 160 | struct cfs_rq; |
| 161 | |
| 162 | /* task group related information */ |
| 163 | struct task_group { |
| 164 | #ifdef CONFIG_FAIR_CGROUP_SCHED |
| 165 | struct cgroup_subsys_state css; |
| 166 | #endif |
| 167 | /* schedulable entities of this group on each cpu */ |
| 168 | struct sched_entity **se; |
| 169 | /* runqueue "owned" by this group on each cpu */ |
| 170 | struct cfs_rq **cfs_rq; |
| 171 | |
| 172 | /* |
| 173 | * shares assigned to a task group governs how much of cpu bandwidth |
| 174 | * is allocated to the group. The more shares a group has, the more is |
| 175 | * the cpu bandwidth allocated to it. |
| 176 | * |
| 177 | * For ex, lets say that there are three task groups, A, B and C which |
| 178 | * have been assigned shares 1000, 2000 and 3000 respectively. Then, |
| 179 | * cpu bandwidth allocated by the scheduler to task groups A, B and C |
| 180 | * should be: |
| 181 | * |
| 182 | * Bw(A) = 1000/(1000+2000+3000) * 100 = 16.66% |
| 183 | * Bw(B) = 2000/(1000+2000+3000) * 100 = 33.33% |
| 184 | * Bw(C) = 3000/(1000+2000+3000) * 100 = 50% |
| 185 | * |
| 186 | * The weight assigned to a task group's schedulable entities on every |
| 187 | * cpu (task_group.se[a_cpu]->load.weight) is derived from the task |
| 188 | * group's shares. For ex: lets say that task group A has been |
| 189 | * assigned shares of 1000 and there are two CPUs in a system. Then, |
| 190 | * |
| 191 | * tg_A->se[0]->load.weight = tg_A->se[1]->load.weight = 1000; |
| 192 | * |
| 193 | * Note: It's not necessary that each of a task's group schedulable |
| 194 | * entity have the same weight on all CPUs. If the group |
| 195 | * has 2 of its tasks on CPU0 and 1 task on CPU1, then a |
| 196 | * better distribution of weight could be: |
| 197 | * |
| 198 | * tg_A->se[0]->load.weight = 2/3 * 2000 = 1333 |
| 199 | * tg_A->se[1]->load.weight = 1/2 * 2000 = 667 |
| 200 | * |
| 201 | * rebalance_shares() is responsible for distributing the shares of a |
| 202 | * task groups like this among the group's schedulable entities across |
| 203 | * cpus. |
| 204 | * |
| 205 | */ |
| 206 | unsigned long shares; |
| 207 | |
| 208 | struct rcu_head rcu; |
| 209 | }; |
| 210 | |
| 211 | /* Default task group's sched entity on each cpu */ |
| 212 | static DEFINE_PER_CPU(struct sched_entity, init_sched_entity); |
| 213 | /* Default task group's cfs_rq on each cpu */ |
| 214 | static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp; |
| 215 | |
| 216 | static struct sched_entity *init_sched_entity_p[NR_CPUS]; |
| 217 | static struct cfs_rq *init_cfs_rq_p[NR_CPUS]; |
| 218 | |
| 219 | /* task_group_mutex serializes add/remove of task groups and also changes to |
| 220 | * a task group's cpu shares. |
| 221 | */ |
| 222 | static DEFINE_MUTEX(task_group_mutex); |
| 223 | |
| 224 | /* doms_cur_mutex serializes access to doms_cur[] array */ |
| 225 | static DEFINE_MUTEX(doms_cur_mutex); |
| 226 | |
| 227 | #ifdef CONFIG_SMP |
| 228 | /* kernel thread that runs rebalance_shares() periodically */ |
| 229 | static struct task_struct *lb_monitor_task; |
| 230 | static int load_balance_monitor(void *unused); |
| 231 | #endif |
| 232 | |
| 233 | static void set_se_shares(struct sched_entity *se, unsigned long shares); |
| 234 | |
| 235 | /* Default task group. |
| 236 | * Every task in system belong to this group at bootup. |
| 237 | */ |
| 238 | struct task_group init_task_group = { |
| 239 | .se = init_sched_entity_p, |
| 240 | .cfs_rq = init_cfs_rq_p, |
| 241 | }; |
| 242 | |
| 243 | #ifdef CONFIG_FAIR_USER_SCHED |
| 244 | # define INIT_TASK_GROUP_LOAD 2*NICE_0_LOAD |
| 245 | #else |
| 246 | # define INIT_TASK_GROUP_LOAD NICE_0_LOAD |
| 247 | #endif |
| 248 | |
| 249 | #define MIN_GROUP_SHARES 2 |
| 250 | |
| 251 | static int init_task_group_load = INIT_TASK_GROUP_LOAD; |
| 252 | |
| 253 | /* return group to which a task belongs */ |
| 254 | static inline struct task_group *task_group(struct task_struct *p) |
| 255 | { |
| 256 | struct task_group *tg; |
| 257 | |
| 258 | #ifdef CONFIG_FAIR_USER_SCHED |
| 259 | tg = p->user->tg; |
| 260 | #elif defined(CONFIG_FAIR_CGROUP_SCHED) |
| 261 | tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id), |
| 262 | struct task_group, css); |
| 263 | #else |
| 264 | tg = &init_task_group; |
| 265 | #endif |
| 266 | return tg; |
| 267 | } |
| 268 | |
| 269 | /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */ |
| 270 | static inline void set_task_cfs_rq(struct task_struct *p, unsigned int cpu) |
| 271 | { |
| 272 | p->se.cfs_rq = task_group(p)->cfs_rq[cpu]; |
| 273 | p->se.parent = task_group(p)->se[cpu]; |
| 274 | } |
| 275 | |
| 276 | static inline void lock_task_group_list(void) |
| 277 | { |
| 278 | mutex_lock(&task_group_mutex); |
| 279 | } |
| 280 | |
| 281 | static inline void unlock_task_group_list(void) |
| 282 | { |
| 283 | mutex_unlock(&task_group_mutex); |
| 284 | } |
| 285 | |
| 286 | static inline void lock_doms_cur(void) |
| 287 | { |
| 288 | mutex_lock(&doms_cur_mutex); |
| 289 | } |
| 290 | |
| 291 | static inline void unlock_doms_cur(void) |
| 292 | { |
| 293 | mutex_unlock(&doms_cur_mutex); |
| 294 | } |
| 295 | |
| 296 | #else |
| 297 | |
| 298 | static inline void set_task_cfs_rq(struct task_struct *p, unsigned int cpu) { } |
| 299 | static inline void lock_task_group_list(void) { } |
| 300 | static inline void unlock_task_group_list(void) { } |
| 301 | static inline void lock_doms_cur(void) { } |
| 302 | static inline void unlock_doms_cur(void) { } |
| 303 | |
| 304 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
| 305 | |
| 306 | /* CFS-related fields in a runqueue */ |
| 307 | struct cfs_rq { |
| 308 | struct load_weight load; |
| 309 | unsigned long nr_running; |
| 310 | |
| 311 | u64 exec_clock; |
| 312 | u64 min_vruntime; |
| 313 | |
| 314 | struct rb_root tasks_timeline; |
| 315 | struct rb_node *rb_leftmost; |
| 316 | struct rb_node *rb_load_balance_curr; |
| 317 | /* 'curr' points to currently running entity on this cfs_rq. |
| 318 | * It is set to NULL otherwise (i.e when none are currently running). |
| 319 | */ |
| 320 | struct sched_entity *curr; |
| 321 | |
| 322 | unsigned long nr_spread_over; |
| 323 | |
| 324 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 325 | struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */ |
| 326 | |
| 327 | /* |
| 328 | * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in |
| 329 | * a hierarchy). Non-leaf lrqs hold other higher schedulable entities |
| 330 | * (like users, containers etc.) |
| 331 | * |
| 332 | * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This |
| 333 | * list is used during load balance. |
| 334 | */ |
| 335 | struct list_head leaf_cfs_rq_list; |
| 336 | struct task_group *tg; /* group that "owns" this runqueue */ |
| 337 | #endif |
| 338 | }; |
| 339 | |
| 340 | /* Real-Time classes' related field in a runqueue: */ |
| 341 | struct rt_rq { |
| 342 | struct rt_prio_array active; |
| 343 | int rt_load_balance_idx; |
| 344 | struct list_head *rt_load_balance_head, *rt_load_balance_curr; |
| 345 | unsigned long rt_nr_running; |
| 346 | unsigned long rt_nr_migratory; |
| 347 | /* highest queued rt task prio */ |
| 348 | int highest_prio; |
| 349 | int overloaded; |
| 350 | }; |
| 351 | |
| 352 | /* |
| 353 | * This is the main, per-CPU runqueue data structure. |
| 354 | * |
| 355 | * Locking rule: those places that want to lock multiple runqueues |
| 356 | * (such as the load balancing or the thread migration code), lock |
| 357 | * acquire operations must be ordered by ascending &runqueue. |
| 358 | */ |
| 359 | struct rq { |
| 360 | /* runqueue lock: */ |
| 361 | spinlock_t lock; |
| 362 | |
| 363 | /* |
| 364 | * nr_running and cpu_load should be in the same cacheline because |
| 365 | * remote CPUs use both these fields when doing load calculation. |
| 366 | */ |
| 367 | unsigned long nr_running; |
| 368 | #define CPU_LOAD_IDX_MAX 5 |
| 369 | unsigned long cpu_load[CPU_LOAD_IDX_MAX]; |
| 370 | unsigned char idle_at_tick; |
| 371 | #ifdef CONFIG_NO_HZ |
| 372 | unsigned char in_nohz_recently; |
| 373 | #endif |
| 374 | /* capture load from *all* tasks on this cpu: */ |
| 375 | struct load_weight load; |
| 376 | unsigned long nr_load_updates; |
| 377 | u64 nr_switches; |
| 378 | |
| 379 | struct cfs_rq cfs; |
| 380 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 381 | /* list of leaf cfs_rq on this cpu: */ |
| 382 | struct list_head leaf_cfs_rq_list; |
| 383 | #endif |
| 384 | struct rt_rq rt; |
| 385 | |
| 386 | /* |
| 387 | * This is part of a global counter where only the total sum |
| 388 | * over all CPUs matters. A task can increase this counter on |
| 389 | * one CPU and if it got migrated afterwards it may decrease |
| 390 | * it on another CPU. Always updated under the runqueue lock: |
| 391 | */ |
| 392 | unsigned long nr_uninterruptible; |
| 393 | |
| 394 | struct task_struct *curr, *idle; |
| 395 | unsigned long next_balance; |
| 396 | struct mm_struct *prev_mm; |
| 397 | |
| 398 | u64 clock, prev_clock_raw; |
| 399 | s64 clock_max_delta; |
| 400 | |
| 401 | unsigned int clock_warps, clock_overflows; |
| 402 | u64 idle_clock; |
| 403 | unsigned int clock_deep_idle_events; |
| 404 | u64 tick_timestamp; |
| 405 | |
| 406 | atomic_t nr_iowait; |
| 407 | |
| 408 | #ifdef CONFIG_SMP |
| 409 | struct sched_domain *sd; |
| 410 | |
| 411 | /* For active balancing */ |
| 412 | int active_balance; |
| 413 | int push_cpu; |
| 414 | /* cpu of this runqueue: */ |
| 415 | int cpu; |
| 416 | |
| 417 | struct task_struct *migration_thread; |
| 418 | struct list_head migration_queue; |
| 419 | #endif |
| 420 | |
| 421 | #ifdef CONFIG_SCHEDSTATS |
| 422 | /* latency stats */ |
| 423 | struct sched_info rq_sched_info; |
| 424 | |
| 425 | /* sys_sched_yield() stats */ |
| 426 | unsigned int yld_exp_empty; |
| 427 | unsigned int yld_act_empty; |
| 428 | unsigned int yld_both_empty; |
| 429 | unsigned int yld_count; |
| 430 | |
| 431 | /* schedule() stats */ |
| 432 | unsigned int sched_switch; |
| 433 | unsigned int sched_count; |
| 434 | unsigned int sched_goidle; |
| 435 | |
| 436 | /* try_to_wake_up() stats */ |
| 437 | unsigned int ttwu_count; |
| 438 | unsigned int ttwu_local; |
| 439 | |
| 440 | /* BKL stats */ |
| 441 | unsigned int bkl_count; |
| 442 | #endif |
| 443 | struct lock_class_key rq_lock_key; |
| 444 | }; |
| 445 | |
| 446 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); |
| 447 | |
| 448 | static inline void check_preempt_curr(struct rq *rq, struct task_struct *p) |
| 449 | { |
| 450 | rq->curr->sched_class->check_preempt_curr(rq, p); |
| 451 | } |
| 452 | |
| 453 | static inline int cpu_of(struct rq *rq) |
| 454 | { |
| 455 | #ifdef CONFIG_SMP |
| 456 | return rq->cpu; |
| 457 | #else |
| 458 | return 0; |
| 459 | #endif |
| 460 | } |
| 461 | |
| 462 | /* |
| 463 | * Update the per-runqueue clock, as finegrained as the platform can give |
| 464 | * us, but without assuming monotonicity, etc.: |
| 465 | */ |
| 466 | static void __update_rq_clock(struct rq *rq) |
| 467 | { |
| 468 | u64 prev_raw = rq->prev_clock_raw; |
| 469 | u64 now = sched_clock(); |
| 470 | s64 delta = now - prev_raw; |
| 471 | u64 clock = rq->clock; |
| 472 | |
| 473 | #ifdef CONFIG_SCHED_DEBUG |
| 474 | WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); |
| 475 | #endif |
| 476 | /* |
| 477 | * Protect against sched_clock() occasionally going backwards: |
| 478 | */ |
| 479 | if (unlikely(delta < 0)) { |
| 480 | clock++; |
| 481 | rq->clock_warps++; |
| 482 | } else { |
| 483 | /* |
| 484 | * Catch too large forward jumps too: |
| 485 | */ |
| 486 | if (unlikely(clock + delta > rq->tick_timestamp + TICK_NSEC)) { |
| 487 | if (clock < rq->tick_timestamp + TICK_NSEC) |
| 488 | clock = rq->tick_timestamp + TICK_NSEC; |
| 489 | else |
| 490 | clock++; |
| 491 | rq->clock_overflows++; |
| 492 | } else { |
| 493 | if (unlikely(delta > rq->clock_max_delta)) |
| 494 | rq->clock_max_delta = delta; |
| 495 | clock += delta; |
| 496 | } |
| 497 | } |
| 498 | |
| 499 | rq->prev_clock_raw = now; |
| 500 | rq->clock = clock; |
| 501 | } |
| 502 | |
| 503 | static void update_rq_clock(struct rq *rq) |
| 504 | { |
| 505 | if (likely(smp_processor_id() == cpu_of(rq))) |
| 506 | __update_rq_clock(rq); |
| 507 | } |
| 508 | |
| 509 | /* |
| 510 | * The domain tree (rq->sd) is protected by RCU's quiescent state transition. |
| 511 | * See detach_destroy_domains: synchronize_sched for details. |
| 512 | * |
| 513 | * The domain tree of any CPU may only be accessed from within |
| 514 | * preempt-disabled sections. |
| 515 | */ |
| 516 | #define for_each_domain(cpu, __sd) \ |
| 517 | for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent) |
| 518 | |
| 519 | #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu))) |
| 520 | #define this_rq() (&__get_cpu_var(runqueues)) |
| 521 | #define task_rq(p) cpu_rq(task_cpu(p)) |
| 522 | #define cpu_curr(cpu) (cpu_rq(cpu)->curr) |
| 523 | |
| 524 | /* |
| 525 | * Tunables that become constants when CONFIG_SCHED_DEBUG is off: |
| 526 | */ |
| 527 | #ifdef CONFIG_SCHED_DEBUG |
| 528 | # define const_debug __read_mostly |
| 529 | #else |
| 530 | # define const_debug static const |
| 531 | #endif |
| 532 | |
| 533 | /* |
| 534 | * Debugging: various feature bits |
| 535 | */ |
| 536 | enum { |
| 537 | SCHED_FEAT_NEW_FAIR_SLEEPERS = 1, |
| 538 | SCHED_FEAT_WAKEUP_PREEMPT = 2, |
| 539 | SCHED_FEAT_START_DEBIT = 4, |
| 540 | SCHED_FEAT_TREE_AVG = 8, |
| 541 | SCHED_FEAT_APPROX_AVG = 16, |
| 542 | }; |
| 543 | |
| 544 | const_debug unsigned int sysctl_sched_features = |
| 545 | SCHED_FEAT_NEW_FAIR_SLEEPERS * 1 | |
| 546 | SCHED_FEAT_WAKEUP_PREEMPT * 1 | |
| 547 | SCHED_FEAT_START_DEBIT * 1 | |
| 548 | SCHED_FEAT_TREE_AVG * 0 | |
| 549 | SCHED_FEAT_APPROX_AVG * 0; |
| 550 | |
| 551 | #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x) |
| 552 | |
| 553 | /* |
| 554 | * Number of tasks to iterate in a single balance run. |
| 555 | * Limited because this is done with IRQs disabled. |
| 556 | */ |
| 557 | const_debug unsigned int sysctl_sched_nr_migrate = 32; |
| 558 | |
| 559 | /* |
| 560 | * For kernel-internal use: high-speed (but slightly incorrect) per-cpu |
| 561 | * clock constructed from sched_clock(): |
| 562 | */ |
| 563 | unsigned long long cpu_clock(int cpu) |
| 564 | { |
| 565 | unsigned long long now; |
| 566 | unsigned long flags; |
| 567 | struct rq *rq; |
| 568 | |
| 569 | local_irq_save(flags); |
| 570 | rq = cpu_rq(cpu); |
| 571 | /* |
| 572 | * Only call sched_clock() if the scheduler has already been |
| 573 | * initialized (some code might call cpu_clock() very early): |
| 574 | */ |
| 575 | if (rq->idle) |
| 576 | update_rq_clock(rq); |
| 577 | now = rq->clock; |
| 578 | local_irq_restore(flags); |
| 579 | |
| 580 | return now; |
| 581 | } |
| 582 | EXPORT_SYMBOL_GPL(cpu_clock); |
| 583 | |
| 584 | #ifndef prepare_arch_switch |
| 585 | # define prepare_arch_switch(next) do { } while (0) |
| 586 | #endif |
| 587 | #ifndef finish_arch_switch |
| 588 | # define finish_arch_switch(prev) do { } while (0) |
| 589 | #endif |
| 590 | |
| 591 | static inline int task_current(struct rq *rq, struct task_struct *p) |
| 592 | { |
| 593 | return rq->curr == p; |
| 594 | } |
| 595 | |
| 596 | #ifndef __ARCH_WANT_UNLOCKED_CTXSW |
| 597 | static inline int task_running(struct rq *rq, struct task_struct *p) |
| 598 | { |
| 599 | return task_current(rq, p); |
| 600 | } |
| 601 | |
| 602 | static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next) |
| 603 | { |
| 604 | } |
| 605 | |
| 606 | static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev) |
| 607 | { |
| 608 | #ifdef CONFIG_DEBUG_SPINLOCK |
| 609 | /* this is a valid case when another task releases the spinlock */ |
| 610 | rq->lock.owner = current; |
| 611 | #endif |
| 612 | /* |
| 613 | * If we are tracking spinlock dependencies then we have to |
| 614 | * fix up the runqueue lock - which gets 'carried over' from |
| 615 | * prev into current: |
| 616 | */ |
| 617 | spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_); |
| 618 | |
| 619 | spin_unlock_irq(&rq->lock); |
| 620 | } |
| 621 | |
| 622 | #else /* __ARCH_WANT_UNLOCKED_CTXSW */ |
| 623 | static inline int task_running(struct rq *rq, struct task_struct *p) |
| 624 | { |
| 625 | #ifdef CONFIG_SMP |
| 626 | return p->oncpu; |
| 627 | #else |
| 628 | return task_current(rq, p); |
| 629 | #endif |
| 630 | } |
| 631 | |
| 632 | static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next) |
| 633 | { |
| 634 | #ifdef CONFIG_SMP |
| 635 | /* |
| 636 | * We can optimise this out completely for !SMP, because the |
| 637 | * SMP rebalancing from interrupt is the only thing that cares |
| 638 | * here. |
| 639 | */ |
| 640 | next->oncpu = 1; |
| 641 | #endif |
| 642 | #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW |
| 643 | spin_unlock_irq(&rq->lock); |
| 644 | #else |
| 645 | spin_unlock(&rq->lock); |
| 646 | #endif |
| 647 | } |
| 648 | |
| 649 | static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev) |
| 650 | { |
| 651 | #ifdef CONFIG_SMP |
| 652 | /* |
| 653 | * After ->oncpu is cleared, the task can be moved to a different CPU. |
| 654 | * We must ensure this doesn't happen until the switch is completely |
| 655 | * finished. |
| 656 | */ |
| 657 | smp_wmb(); |
| 658 | prev->oncpu = 0; |
| 659 | #endif |
| 660 | #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW |
| 661 | local_irq_enable(); |
| 662 | #endif |
| 663 | } |
| 664 | #endif /* __ARCH_WANT_UNLOCKED_CTXSW */ |
| 665 | |
| 666 | /* |
| 667 | * __task_rq_lock - lock the runqueue a given task resides on. |
| 668 | * Must be called interrupts disabled. |
| 669 | */ |
| 670 | static inline struct rq *__task_rq_lock(struct task_struct *p) |
| 671 | __acquires(rq->lock) |
| 672 | { |
| 673 | for (;;) { |
| 674 | struct rq *rq = task_rq(p); |
| 675 | spin_lock(&rq->lock); |
| 676 | if (likely(rq == task_rq(p))) |
| 677 | return rq; |
| 678 | spin_unlock(&rq->lock); |
| 679 | } |
| 680 | } |
| 681 | |
| 682 | /* |
| 683 | * task_rq_lock - lock the runqueue a given task resides on and disable |
| 684 | * interrupts. Note the ordering: we can safely lookup the task_rq without |
| 685 | * explicitly disabling preemption. |
| 686 | */ |
| 687 | static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) |
| 688 | __acquires(rq->lock) |
| 689 | { |
| 690 | struct rq *rq; |
| 691 | |
| 692 | for (;;) { |
| 693 | local_irq_save(*flags); |
| 694 | rq = task_rq(p); |
| 695 | spin_lock(&rq->lock); |
| 696 | if (likely(rq == task_rq(p))) |
| 697 | return rq; |
| 698 | spin_unlock_irqrestore(&rq->lock, *flags); |
| 699 | } |
| 700 | } |
| 701 | |
| 702 | static void __task_rq_unlock(struct rq *rq) |
| 703 | __releases(rq->lock) |
| 704 | { |
| 705 | spin_unlock(&rq->lock); |
| 706 | } |
| 707 | |
| 708 | static inline void task_rq_unlock(struct rq *rq, unsigned long *flags) |
| 709 | __releases(rq->lock) |
| 710 | { |
| 711 | spin_unlock_irqrestore(&rq->lock, *flags); |
| 712 | } |
| 713 | |
| 714 | /* |
| 715 | * this_rq_lock - lock this runqueue and disable interrupts. |
| 716 | */ |
| 717 | static struct rq *this_rq_lock(void) |
| 718 | __acquires(rq->lock) |
| 719 | { |
| 720 | struct rq *rq; |
| 721 | |
| 722 | local_irq_disable(); |
| 723 | rq = this_rq(); |
| 724 | spin_lock(&rq->lock); |
| 725 | |
| 726 | return rq; |
| 727 | } |
| 728 | |
| 729 | /* |
| 730 | * We are going deep-idle (irqs are disabled): |
| 731 | */ |
| 732 | void sched_clock_idle_sleep_event(void) |
| 733 | { |
| 734 | struct rq *rq = cpu_rq(smp_processor_id()); |
| 735 | |
| 736 | spin_lock(&rq->lock); |
| 737 | __update_rq_clock(rq); |
| 738 | spin_unlock(&rq->lock); |
| 739 | rq->clock_deep_idle_events++; |
| 740 | } |
| 741 | EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event); |
| 742 | |
| 743 | /* |
| 744 | * We just idled delta nanoseconds (called with irqs disabled): |
| 745 | */ |
| 746 | void sched_clock_idle_wakeup_event(u64 delta_ns) |
| 747 | { |
| 748 | struct rq *rq = cpu_rq(smp_processor_id()); |
| 749 | u64 now = sched_clock(); |
| 750 | |
| 751 | touch_softlockup_watchdog(); |
| 752 | rq->idle_clock += delta_ns; |
| 753 | /* |
| 754 | * Override the previous timestamp and ignore all |
| 755 | * sched_clock() deltas that occured while we idled, |
| 756 | * and use the PM-provided delta_ns to advance the |
| 757 | * rq clock: |
| 758 | */ |
| 759 | spin_lock(&rq->lock); |
| 760 | rq->prev_clock_raw = now; |
| 761 | rq->clock += delta_ns; |
| 762 | spin_unlock(&rq->lock); |
| 763 | } |
| 764 | EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event); |
| 765 | |
| 766 | /* |
| 767 | * resched_task - mark a task 'to be rescheduled now'. |
| 768 | * |
| 769 | * On UP this means the setting of the need_resched flag, on SMP it |
| 770 | * might also involve a cross-CPU call to trigger the scheduler on |
| 771 | * the target CPU. |
| 772 | */ |
| 773 | #ifdef CONFIG_SMP |
| 774 | |
| 775 | #ifndef tsk_is_polling |
| 776 | #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG) |
| 777 | #endif |
| 778 | |
| 779 | static void resched_task(struct task_struct *p) |
| 780 | { |
| 781 | int cpu; |
| 782 | |
| 783 | assert_spin_locked(&task_rq(p)->lock); |
| 784 | |
| 785 | if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED))) |
| 786 | return; |
| 787 | |
| 788 | set_tsk_thread_flag(p, TIF_NEED_RESCHED); |
| 789 | |
| 790 | cpu = task_cpu(p); |
| 791 | if (cpu == smp_processor_id()) |
| 792 | return; |
| 793 | |
| 794 | /* NEED_RESCHED must be visible before we test polling */ |
| 795 | smp_mb(); |
| 796 | if (!tsk_is_polling(p)) |
| 797 | smp_send_reschedule(cpu); |
| 798 | } |
| 799 | |
| 800 | static void resched_cpu(int cpu) |
| 801 | { |
| 802 | struct rq *rq = cpu_rq(cpu); |
| 803 | unsigned long flags; |
| 804 | |
| 805 | if (!spin_trylock_irqsave(&rq->lock, flags)) |
| 806 | return; |
| 807 | resched_task(cpu_curr(cpu)); |
| 808 | spin_unlock_irqrestore(&rq->lock, flags); |
| 809 | } |
| 810 | #else |
| 811 | static inline void resched_task(struct task_struct *p) |
| 812 | { |
| 813 | assert_spin_locked(&task_rq(p)->lock); |
| 814 | set_tsk_need_resched(p); |
| 815 | } |
| 816 | #endif |
| 817 | |
| 818 | #if BITS_PER_LONG == 32 |
| 819 | # define WMULT_CONST (~0UL) |
| 820 | #else |
| 821 | # define WMULT_CONST (1UL << 32) |
| 822 | #endif |
| 823 | |
| 824 | #define WMULT_SHIFT 32 |
| 825 | |
| 826 | /* |
| 827 | * Shift right and round: |
| 828 | */ |
| 829 | #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y)) |
| 830 | |
| 831 | static unsigned long |
| 832 | calc_delta_mine(unsigned long delta_exec, unsigned long weight, |
| 833 | struct load_weight *lw) |
| 834 | { |
| 835 | u64 tmp; |
| 836 | |
| 837 | if (unlikely(!lw->inv_weight)) |
| 838 | lw->inv_weight = (WMULT_CONST - lw->weight/2) / lw->weight + 1; |
| 839 | |
| 840 | tmp = (u64)delta_exec * weight; |
| 841 | /* |
| 842 | * Check whether we'd overflow the 64-bit multiplication: |
| 843 | */ |
| 844 | if (unlikely(tmp > WMULT_CONST)) |
| 845 | tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight, |
| 846 | WMULT_SHIFT/2); |
| 847 | else |
| 848 | tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT); |
| 849 | |
| 850 | return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX); |
| 851 | } |
| 852 | |
| 853 | static inline unsigned long |
| 854 | calc_delta_fair(unsigned long delta_exec, struct load_weight *lw) |
| 855 | { |
| 856 | return calc_delta_mine(delta_exec, NICE_0_LOAD, lw); |
| 857 | } |
| 858 | |
| 859 | static inline void update_load_add(struct load_weight *lw, unsigned long inc) |
| 860 | { |
| 861 | lw->weight += inc; |
| 862 | } |
| 863 | |
| 864 | static inline void update_load_sub(struct load_weight *lw, unsigned long dec) |
| 865 | { |
| 866 | lw->weight -= dec; |
| 867 | } |
| 868 | |
| 869 | /* |
| 870 | * To aid in avoiding the subversion of "niceness" due to uneven distribution |
| 871 | * of tasks with abnormal "nice" values across CPUs the contribution that |
| 872 | * each task makes to its run queue's load is weighted according to its |
| 873 | * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a |
| 874 | * scaled version of the new time slice allocation that they receive on time |
| 875 | * slice expiry etc. |
| 876 | */ |
| 877 | |
| 878 | #define WEIGHT_IDLEPRIO 2 |
| 879 | #define WMULT_IDLEPRIO (1 << 31) |
| 880 | |
| 881 | /* |
| 882 | * Nice levels are multiplicative, with a gentle 10% change for every |
| 883 | * nice level changed. I.e. when a CPU-bound task goes from nice 0 to |
| 884 | * nice 1, it will get ~10% less CPU time than another CPU-bound task |
| 885 | * that remained on nice 0. |
| 886 | * |
| 887 | * The "10% effect" is relative and cumulative: from _any_ nice level, |
| 888 | * if you go up 1 level, it's -10% CPU usage, if you go down 1 level |
| 889 | * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25. |
| 890 | * If a task goes up by ~10% and another task goes down by ~10% then |
| 891 | * the relative distance between them is ~25%.) |
| 892 | */ |
| 893 | static const int prio_to_weight[40] = { |
| 894 | /* -20 */ 88761, 71755, 56483, 46273, 36291, |
| 895 | /* -15 */ 29154, 23254, 18705, 14949, 11916, |
| 896 | /* -10 */ 9548, 7620, 6100, 4904, 3906, |
| 897 | /* -5 */ 3121, 2501, 1991, 1586, 1277, |
| 898 | /* 0 */ 1024, 820, 655, 526, 423, |
| 899 | /* 5 */ 335, 272, 215, 172, 137, |
| 900 | /* 10 */ 110, 87, 70, 56, 45, |
| 901 | /* 15 */ 36, 29, 23, 18, 15, |
| 902 | }; |
| 903 | |
| 904 | /* |
| 905 | * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated. |
| 906 | * |
| 907 | * In cases where the weight does not change often, we can use the |
| 908 | * precalculated inverse to speed up arithmetics by turning divisions |
| 909 | * into multiplications: |
| 910 | */ |
| 911 | static const u32 prio_to_wmult[40] = { |
| 912 | /* -20 */ 48388, 59856, 76040, 92818, 118348, |
| 913 | /* -15 */ 147320, 184698, 229616, 287308, 360437, |
| 914 | /* -10 */ 449829, 563644, 704093, 875809, 1099582, |
| 915 | /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326, |
| 916 | /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587, |
| 917 | /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126, |
| 918 | /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717, |
| 919 | /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153, |
| 920 | }; |
| 921 | |
| 922 | static void activate_task(struct rq *rq, struct task_struct *p, int wakeup); |
| 923 | |
| 924 | /* |
| 925 | * runqueue iterator, to support SMP load-balancing between different |
| 926 | * scheduling classes, without having to expose their internal data |
| 927 | * structures to the load-balancing proper: |
| 928 | */ |
| 929 | struct rq_iterator { |
| 930 | void *arg; |
| 931 | struct task_struct *(*start)(void *); |
| 932 | struct task_struct *(*next)(void *); |
| 933 | }; |
| 934 | |
| 935 | #ifdef CONFIG_SMP |
| 936 | static unsigned long |
| 937 | balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, |
| 938 | unsigned long max_load_move, struct sched_domain *sd, |
| 939 | enum cpu_idle_type idle, int *all_pinned, |
| 940 | int *this_best_prio, struct rq_iterator *iterator); |
| 941 | |
| 942 | static int |
| 943 | iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest, |
| 944 | struct sched_domain *sd, enum cpu_idle_type idle, |
| 945 | struct rq_iterator *iterator); |
| 946 | #endif |
| 947 | |
| 948 | #ifdef CONFIG_CGROUP_CPUACCT |
| 949 | static void cpuacct_charge(struct task_struct *tsk, u64 cputime); |
| 950 | #else |
| 951 | static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {} |
| 952 | #endif |
| 953 | |
| 954 | static inline void inc_cpu_load(struct rq *rq, unsigned long load) |
| 955 | { |
| 956 | update_load_add(&rq->load, load); |
| 957 | } |
| 958 | |
| 959 | static inline void dec_cpu_load(struct rq *rq, unsigned long load) |
| 960 | { |
| 961 | update_load_sub(&rq->load, load); |
| 962 | } |
| 963 | |
| 964 | #ifdef CONFIG_SMP |
| 965 | static unsigned long source_load(int cpu, int type); |
| 966 | static unsigned long target_load(int cpu, int type); |
| 967 | static unsigned long cpu_avg_load_per_task(int cpu); |
| 968 | static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd); |
| 969 | #endif /* CONFIG_SMP */ |
| 970 | |
| 971 | #include "sched_stats.h" |
| 972 | #include "sched_idletask.c" |
| 973 | #include "sched_fair.c" |
| 974 | #include "sched_rt.c" |
| 975 | #ifdef CONFIG_SCHED_DEBUG |
| 976 | # include "sched_debug.c" |
| 977 | #endif |
| 978 | |
| 979 | #define sched_class_highest (&rt_sched_class) |
| 980 | |
| 981 | static void inc_nr_running(struct task_struct *p, struct rq *rq) |
| 982 | { |
| 983 | rq->nr_running++; |
| 984 | } |
| 985 | |
| 986 | static void dec_nr_running(struct task_struct *p, struct rq *rq) |
| 987 | { |
| 988 | rq->nr_running--; |
| 989 | } |
| 990 | |
| 991 | static void set_load_weight(struct task_struct *p) |
| 992 | { |
| 993 | if (task_has_rt_policy(p)) { |
| 994 | p->se.load.weight = prio_to_weight[0] * 2; |
| 995 | p->se.load.inv_weight = prio_to_wmult[0] >> 1; |
| 996 | return; |
| 997 | } |
| 998 | |
| 999 | /* |
| 1000 | * SCHED_IDLE tasks get minimal weight: |
| 1001 | */ |
| 1002 | if (p->policy == SCHED_IDLE) { |
| 1003 | p->se.load.weight = WEIGHT_IDLEPRIO; |
| 1004 | p->se.load.inv_weight = WMULT_IDLEPRIO; |
| 1005 | return; |
| 1006 | } |
| 1007 | |
| 1008 | p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO]; |
| 1009 | p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO]; |
| 1010 | } |
| 1011 | |
| 1012 | static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup) |
| 1013 | { |
| 1014 | sched_info_queued(p); |
| 1015 | p->sched_class->enqueue_task(rq, p, wakeup); |
| 1016 | p->se.on_rq = 1; |
| 1017 | } |
| 1018 | |
| 1019 | static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep) |
| 1020 | { |
| 1021 | p->sched_class->dequeue_task(rq, p, sleep); |
| 1022 | p->se.on_rq = 0; |
| 1023 | } |
| 1024 | |
| 1025 | /* |
| 1026 | * __normal_prio - return the priority that is based on the static prio |
| 1027 | */ |
| 1028 | static inline int __normal_prio(struct task_struct *p) |
| 1029 | { |
| 1030 | return p->static_prio; |
| 1031 | } |
| 1032 | |
| 1033 | /* |
| 1034 | * Calculate the expected normal priority: i.e. priority |
| 1035 | * without taking RT-inheritance into account. Might be |
| 1036 | * boosted by interactivity modifiers. Changes upon fork, |
| 1037 | * setprio syscalls, and whenever the interactivity |
| 1038 | * estimator recalculates. |
| 1039 | */ |
| 1040 | static inline int normal_prio(struct task_struct *p) |
| 1041 | { |
| 1042 | int prio; |
| 1043 | |
| 1044 | if (task_has_rt_policy(p)) |
| 1045 | prio = MAX_RT_PRIO-1 - p->rt_priority; |
| 1046 | else |
| 1047 | prio = __normal_prio(p); |
| 1048 | return prio; |
| 1049 | } |
| 1050 | |
| 1051 | /* |
| 1052 | * Calculate the current priority, i.e. the priority |
| 1053 | * taken into account by the scheduler. This value might |
| 1054 | * be boosted by RT tasks, or might be boosted by |
| 1055 | * interactivity modifiers. Will be RT if the task got |
| 1056 | * RT-boosted. If not then it returns p->normal_prio. |
| 1057 | */ |
| 1058 | static int effective_prio(struct task_struct *p) |
| 1059 | { |
| 1060 | p->normal_prio = normal_prio(p); |
| 1061 | /* |
| 1062 | * If we are RT tasks or we were boosted to RT priority, |
| 1063 | * keep the priority unchanged. Otherwise, update priority |
| 1064 | * to the normal priority: |
| 1065 | */ |
| 1066 | if (!rt_prio(p->prio)) |
| 1067 | return p->normal_prio; |
| 1068 | return p->prio; |
| 1069 | } |
| 1070 | |
| 1071 | /* |
| 1072 | * activate_task - move a task to the runqueue. |
| 1073 | */ |
| 1074 | static void activate_task(struct rq *rq, struct task_struct *p, int wakeup) |
| 1075 | { |
| 1076 | if (p->state == TASK_UNINTERRUPTIBLE) |
| 1077 | rq->nr_uninterruptible--; |
| 1078 | |
| 1079 | enqueue_task(rq, p, wakeup); |
| 1080 | inc_nr_running(p, rq); |
| 1081 | } |
| 1082 | |
| 1083 | /* |
| 1084 | * deactivate_task - remove a task from the runqueue. |
| 1085 | */ |
| 1086 | static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep) |
| 1087 | { |
| 1088 | if (p->state == TASK_UNINTERRUPTIBLE) |
| 1089 | rq->nr_uninterruptible++; |
| 1090 | |
| 1091 | dequeue_task(rq, p, sleep); |
| 1092 | dec_nr_running(p, rq); |
| 1093 | } |
| 1094 | |
| 1095 | /** |
| 1096 | * task_curr - is this task currently executing on a CPU? |
| 1097 | * @p: the task in question. |
| 1098 | */ |
| 1099 | inline int task_curr(const struct task_struct *p) |
| 1100 | { |
| 1101 | return cpu_curr(task_cpu(p)) == p; |
| 1102 | } |
| 1103 | |
| 1104 | /* Used instead of source_load when we know the type == 0 */ |
| 1105 | unsigned long weighted_cpuload(const int cpu) |
| 1106 | { |
| 1107 | return cpu_rq(cpu)->load.weight; |
| 1108 | } |
| 1109 | |
| 1110 | static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu) |
| 1111 | { |
| 1112 | set_task_cfs_rq(p, cpu); |
| 1113 | #ifdef CONFIG_SMP |
| 1114 | /* |
| 1115 | * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be |
| 1116 | * successfuly executed on another CPU. We must ensure that updates of |
| 1117 | * per-task data have been completed by this moment. |
| 1118 | */ |
| 1119 | smp_wmb(); |
| 1120 | task_thread_info(p)->cpu = cpu; |
| 1121 | #endif |
| 1122 | } |
| 1123 | |
| 1124 | #ifdef CONFIG_SMP |
| 1125 | |
| 1126 | /* |
| 1127 | * Is this task likely cache-hot: |
| 1128 | */ |
| 1129 | static int |
| 1130 | task_hot(struct task_struct *p, u64 now, struct sched_domain *sd) |
| 1131 | { |
| 1132 | s64 delta; |
| 1133 | |
| 1134 | if (p->sched_class != &fair_sched_class) |
| 1135 | return 0; |
| 1136 | |
| 1137 | if (sysctl_sched_migration_cost == -1) |
| 1138 | return 1; |
| 1139 | if (sysctl_sched_migration_cost == 0) |
| 1140 | return 0; |
| 1141 | |
| 1142 | delta = now - p->se.exec_start; |
| 1143 | |
| 1144 | return delta < (s64)sysctl_sched_migration_cost; |
| 1145 | } |
| 1146 | |
| 1147 | |
| 1148 | void set_task_cpu(struct task_struct *p, unsigned int new_cpu) |
| 1149 | { |
| 1150 | int old_cpu = task_cpu(p); |
| 1151 | struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu); |
| 1152 | struct cfs_rq *old_cfsrq = task_cfs_rq(p), |
| 1153 | *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu); |
| 1154 | u64 clock_offset; |
| 1155 | |
| 1156 | clock_offset = old_rq->clock - new_rq->clock; |
| 1157 | |
| 1158 | #ifdef CONFIG_SCHEDSTATS |
| 1159 | if (p->se.wait_start) |
| 1160 | p->se.wait_start -= clock_offset; |
| 1161 | if (p->se.sleep_start) |
| 1162 | p->se.sleep_start -= clock_offset; |
| 1163 | if (p->se.block_start) |
| 1164 | p->se.block_start -= clock_offset; |
| 1165 | if (old_cpu != new_cpu) { |
| 1166 | schedstat_inc(p, se.nr_migrations); |
| 1167 | if (task_hot(p, old_rq->clock, NULL)) |
| 1168 | schedstat_inc(p, se.nr_forced2_migrations); |
| 1169 | } |
| 1170 | #endif |
| 1171 | p->se.vruntime -= old_cfsrq->min_vruntime - |
| 1172 | new_cfsrq->min_vruntime; |
| 1173 | |
| 1174 | __set_task_cpu(p, new_cpu); |
| 1175 | } |
| 1176 | |
| 1177 | struct migration_req { |
| 1178 | struct list_head list; |
| 1179 | |
| 1180 | struct task_struct *task; |
| 1181 | int dest_cpu; |
| 1182 | |
| 1183 | struct completion done; |
| 1184 | }; |
| 1185 | |
| 1186 | /* |
| 1187 | * The task's runqueue lock must be held. |
| 1188 | * Returns true if you have to wait for migration thread. |
| 1189 | */ |
| 1190 | static int |
| 1191 | migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req) |
| 1192 | { |
| 1193 | struct rq *rq = task_rq(p); |
| 1194 | |
| 1195 | /* |
| 1196 | * If the task is not on a runqueue (and not running), then |
| 1197 | * it is sufficient to simply update the task's cpu field. |
| 1198 | */ |
| 1199 | if (!p->se.on_rq && !task_running(rq, p)) { |
| 1200 | set_task_cpu(p, dest_cpu); |
| 1201 | return 0; |
| 1202 | } |
| 1203 | |
| 1204 | init_completion(&req->done); |
| 1205 | req->task = p; |
| 1206 | req->dest_cpu = dest_cpu; |
| 1207 | list_add(&req->list, &rq->migration_queue); |
| 1208 | |
| 1209 | return 1; |
| 1210 | } |
| 1211 | |
| 1212 | /* |
| 1213 | * wait_task_inactive - wait for a thread to unschedule. |
| 1214 | * |
| 1215 | * The caller must ensure that the task *will* unschedule sometime soon, |
| 1216 | * else this function might spin for a *long* time. This function can't |
| 1217 | * be called with interrupts off, or it may introduce deadlock with |
| 1218 | * smp_call_function() if an IPI is sent by the same process we are |
| 1219 | * waiting to become inactive. |
| 1220 | */ |
| 1221 | void wait_task_inactive(struct task_struct *p) |
| 1222 | { |
| 1223 | unsigned long flags; |
| 1224 | int running, on_rq; |
| 1225 | struct rq *rq; |
| 1226 | |
| 1227 | for (;;) { |
| 1228 | /* |
| 1229 | * We do the initial early heuristics without holding |
| 1230 | * any task-queue locks at all. We'll only try to get |
| 1231 | * the runqueue lock when things look like they will |
| 1232 | * work out! |
| 1233 | */ |
| 1234 | rq = task_rq(p); |
| 1235 | |
| 1236 | /* |
| 1237 | * If the task is actively running on another CPU |
| 1238 | * still, just relax and busy-wait without holding |
| 1239 | * any locks. |
| 1240 | * |
| 1241 | * NOTE! Since we don't hold any locks, it's not |
| 1242 | * even sure that "rq" stays as the right runqueue! |
| 1243 | * But we don't care, since "task_running()" will |
| 1244 | * return false if the runqueue has changed and p |
| 1245 | * is actually now running somewhere else! |
| 1246 | */ |
| 1247 | while (task_running(rq, p)) |
| 1248 | cpu_relax(); |
| 1249 | |
| 1250 | /* |
| 1251 | * Ok, time to look more closely! We need the rq |
| 1252 | * lock now, to be *sure*. If we're wrong, we'll |
| 1253 | * just go back and repeat. |
| 1254 | */ |
| 1255 | rq = task_rq_lock(p, &flags); |
| 1256 | running = task_running(rq, p); |
| 1257 | on_rq = p->se.on_rq; |
| 1258 | task_rq_unlock(rq, &flags); |
| 1259 | |
| 1260 | /* |
| 1261 | * Was it really running after all now that we |
| 1262 | * checked with the proper locks actually held? |
| 1263 | * |
| 1264 | * Oops. Go back and try again.. |
| 1265 | */ |
| 1266 | if (unlikely(running)) { |
| 1267 | cpu_relax(); |
| 1268 | continue; |
| 1269 | } |
| 1270 | |
| 1271 | /* |
| 1272 | * It's not enough that it's not actively running, |
| 1273 | * it must be off the runqueue _entirely_, and not |
| 1274 | * preempted! |
| 1275 | * |
| 1276 | * So if it wa still runnable (but just not actively |
| 1277 | * running right now), it's preempted, and we should |
| 1278 | * yield - it could be a while. |
| 1279 | */ |
| 1280 | if (unlikely(on_rq)) { |
| 1281 | schedule_timeout_uninterruptible(1); |
| 1282 | continue; |
| 1283 | } |
| 1284 | |
| 1285 | /* |
| 1286 | * Ahh, all good. It wasn't running, and it wasn't |
| 1287 | * runnable, which means that it will never become |
| 1288 | * running in the future either. We're all done! |
| 1289 | */ |
| 1290 | break; |
| 1291 | } |
| 1292 | } |
| 1293 | |
| 1294 | /*** |
| 1295 | * kick_process - kick a running thread to enter/exit the kernel |
| 1296 | * @p: the to-be-kicked thread |
| 1297 | * |
| 1298 | * Cause a process which is running on another CPU to enter |
| 1299 | * kernel-mode, without any delay. (to get signals handled.) |
| 1300 | * |
| 1301 | * NOTE: this function doesnt have to take the runqueue lock, |
| 1302 | * because all it wants to ensure is that the remote task enters |
| 1303 | * the kernel. If the IPI races and the task has been migrated |
| 1304 | * to another CPU then no harm is done and the purpose has been |
| 1305 | * achieved as well. |
| 1306 | */ |
| 1307 | void kick_process(struct task_struct *p) |
| 1308 | { |
| 1309 | int cpu; |
| 1310 | |
| 1311 | preempt_disable(); |
| 1312 | cpu = task_cpu(p); |
| 1313 | if ((cpu != smp_processor_id()) && task_curr(p)) |
| 1314 | smp_send_reschedule(cpu); |
| 1315 | preempt_enable(); |
| 1316 | } |
| 1317 | |
| 1318 | /* |
| 1319 | * Return a low guess at the load of a migration-source cpu weighted |
| 1320 | * according to the scheduling class and "nice" value. |
| 1321 | * |
| 1322 | * We want to under-estimate the load of migration sources, to |
| 1323 | * balance conservatively. |
| 1324 | */ |
| 1325 | static unsigned long source_load(int cpu, int type) |
| 1326 | { |
| 1327 | struct rq *rq = cpu_rq(cpu); |
| 1328 | unsigned long total = weighted_cpuload(cpu); |
| 1329 | |
| 1330 | if (type == 0) |
| 1331 | return total; |
| 1332 | |
| 1333 | return min(rq->cpu_load[type-1], total); |
| 1334 | } |
| 1335 | |
| 1336 | /* |
| 1337 | * Return a high guess at the load of a migration-target cpu weighted |
| 1338 | * according to the scheduling class and "nice" value. |
| 1339 | */ |
| 1340 | static unsigned long target_load(int cpu, int type) |
| 1341 | { |
| 1342 | struct rq *rq = cpu_rq(cpu); |
| 1343 | unsigned long total = weighted_cpuload(cpu); |
| 1344 | |
| 1345 | if (type == 0) |
| 1346 | return total; |
| 1347 | |
| 1348 | return max(rq->cpu_load[type-1], total); |
| 1349 | } |
| 1350 | |
| 1351 | /* |
| 1352 | * Return the average load per task on the cpu's run queue |
| 1353 | */ |
| 1354 | static unsigned long cpu_avg_load_per_task(int cpu) |
| 1355 | { |
| 1356 | struct rq *rq = cpu_rq(cpu); |
| 1357 | unsigned long total = weighted_cpuload(cpu); |
| 1358 | unsigned long n = rq->nr_running; |
| 1359 | |
| 1360 | return n ? total / n : SCHED_LOAD_SCALE; |
| 1361 | } |
| 1362 | |
| 1363 | /* |
| 1364 | * find_idlest_group finds and returns the least busy CPU group within the |
| 1365 | * domain. |
| 1366 | */ |
| 1367 | static struct sched_group * |
| 1368 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu) |
| 1369 | { |
| 1370 | struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups; |
| 1371 | unsigned long min_load = ULONG_MAX, this_load = 0; |
| 1372 | int load_idx = sd->forkexec_idx; |
| 1373 | int imbalance = 100 + (sd->imbalance_pct-100)/2; |
| 1374 | |
| 1375 | do { |
| 1376 | unsigned long load, avg_load; |
| 1377 | int local_group; |
| 1378 | int i; |
| 1379 | |
| 1380 | /* Skip over this group if it has no CPUs allowed */ |
| 1381 | if (!cpus_intersects(group->cpumask, p->cpus_allowed)) |
| 1382 | continue; |
| 1383 | |
| 1384 | local_group = cpu_isset(this_cpu, group->cpumask); |
| 1385 | |
| 1386 | /* Tally up the load of all CPUs in the group */ |
| 1387 | avg_load = 0; |
| 1388 | |
| 1389 | for_each_cpu_mask(i, group->cpumask) { |
| 1390 | /* Bias balancing toward cpus of our domain */ |
| 1391 | if (local_group) |
| 1392 | load = source_load(i, load_idx); |
| 1393 | else |
| 1394 | load = target_load(i, load_idx); |
| 1395 | |
| 1396 | avg_load += load; |
| 1397 | } |
| 1398 | |
| 1399 | /* Adjust by relative CPU power of the group */ |
| 1400 | avg_load = sg_div_cpu_power(group, |
| 1401 | avg_load * SCHED_LOAD_SCALE); |
| 1402 | |
| 1403 | if (local_group) { |
| 1404 | this_load = avg_load; |
| 1405 | this = group; |
| 1406 | } else if (avg_load < min_load) { |
| 1407 | min_load = avg_load; |
| 1408 | idlest = group; |
| 1409 | } |
| 1410 | } while (group = group->next, group != sd->groups); |
| 1411 | |
| 1412 | if (!idlest || 100*this_load < imbalance*min_load) |
| 1413 | return NULL; |
| 1414 | return idlest; |
| 1415 | } |
| 1416 | |
| 1417 | /* |
| 1418 | * find_idlest_cpu - find the idlest cpu among the cpus in group. |
| 1419 | */ |
| 1420 | static int |
| 1421 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) |
| 1422 | { |
| 1423 | cpumask_t tmp; |
| 1424 | unsigned long load, min_load = ULONG_MAX; |
| 1425 | int idlest = -1; |
| 1426 | int i; |
| 1427 | |
| 1428 | /* Traverse only the allowed CPUs */ |
| 1429 | cpus_and(tmp, group->cpumask, p->cpus_allowed); |
| 1430 | |
| 1431 | for_each_cpu_mask(i, tmp) { |
| 1432 | load = weighted_cpuload(i); |
| 1433 | |
| 1434 | if (load < min_load || (load == min_load && i == this_cpu)) { |
| 1435 | min_load = load; |
| 1436 | idlest = i; |
| 1437 | } |
| 1438 | } |
| 1439 | |
| 1440 | return idlest; |
| 1441 | } |
| 1442 | |
| 1443 | /* |
| 1444 | * sched_balance_self: balance the current task (running on cpu) in domains |
| 1445 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and |
| 1446 | * SD_BALANCE_EXEC. |
| 1447 | * |
| 1448 | * Balance, ie. select the least loaded group. |
| 1449 | * |
| 1450 | * Returns the target CPU number, or the same CPU if no balancing is needed. |
| 1451 | * |
| 1452 | * preempt must be disabled. |
| 1453 | */ |
| 1454 | static int sched_balance_self(int cpu, int flag) |
| 1455 | { |
| 1456 | struct task_struct *t = current; |
| 1457 | struct sched_domain *tmp, *sd = NULL; |
| 1458 | |
| 1459 | for_each_domain(cpu, tmp) { |
| 1460 | /* |
| 1461 | * If power savings logic is enabled for a domain, stop there. |
| 1462 | */ |
| 1463 | if (tmp->flags & SD_POWERSAVINGS_BALANCE) |
| 1464 | break; |
| 1465 | if (tmp->flags & flag) |
| 1466 | sd = tmp; |
| 1467 | } |
| 1468 | |
| 1469 | while (sd) { |
| 1470 | cpumask_t span; |
| 1471 | struct sched_group *group; |
| 1472 | int new_cpu, weight; |
| 1473 | |
| 1474 | if (!(sd->flags & flag)) { |
| 1475 | sd = sd->child; |
| 1476 | continue; |
| 1477 | } |
| 1478 | |
| 1479 | span = sd->span; |
| 1480 | group = find_idlest_group(sd, t, cpu); |
| 1481 | if (!group) { |
| 1482 | sd = sd->child; |
| 1483 | continue; |
| 1484 | } |
| 1485 | |
| 1486 | new_cpu = find_idlest_cpu(group, t, cpu); |
| 1487 | if (new_cpu == -1 || new_cpu == cpu) { |
| 1488 | /* Now try balancing at a lower domain level of cpu */ |
| 1489 | sd = sd->child; |
| 1490 | continue; |
| 1491 | } |
| 1492 | |
| 1493 | /* Now try balancing at a lower domain level of new_cpu */ |
| 1494 | cpu = new_cpu; |
| 1495 | sd = NULL; |
| 1496 | weight = cpus_weight(span); |
| 1497 | for_each_domain(cpu, tmp) { |
| 1498 | if (weight <= cpus_weight(tmp->span)) |
| 1499 | break; |
| 1500 | if (tmp->flags & flag) |
| 1501 | sd = tmp; |
| 1502 | } |
| 1503 | /* while loop will break here if sd == NULL */ |
| 1504 | } |
| 1505 | |
| 1506 | return cpu; |
| 1507 | } |
| 1508 | |
| 1509 | #endif /* CONFIG_SMP */ |
| 1510 | |
| 1511 | /*** |
| 1512 | * try_to_wake_up - wake up a thread |
| 1513 | * @p: the to-be-woken-up thread |
| 1514 | * @state: the mask of task states that can be woken |
| 1515 | * @sync: do a synchronous wakeup? |
| 1516 | * |
| 1517 | * Put it on the run-queue if it's not already there. The "current" |
| 1518 | * thread is always on the run-queue (except when the actual |
| 1519 | * re-schedule is in progress), and as such you're allowed to do |
| 1520 | * the simpler "current->state = TASK_RUNNING" to mark yourself |
| 1521 | * runnable without the overhead of this. |
| 1522 | * |
| 1523 | * returns failure only if the task is already active. |
| 1524 | */ |
| 1525 | static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync) |
| 1526 | { |
| 1527 | int cpu, orig_cpu, this_cpu, success = 0; |
| 1528 | unsigned long flags; |
| 1529 | long old_state; |
| 1530 | struct rq *rq; |
| 1531 | #ifdef CONFIG_SMP |
| 1532 | int new_cpu; |
| 1533 | #endif |
| 1534 | |
| 1535 | rq = task_rq_lock(p, &flags); |
| 1536 | old_state = p->state; |
| 1537 | if (!(old_state & state)) |
| 1538 | goto out; |
| 1539 | |
| 1540 | if (p->se.on_rq) |
| 1541 | goto out_running; |
| 1542 | |
| 1543 | cpu = task_cpu(p); |
| 1544 | orig_cpu = cpu; |
| 1545 | this_cpu = smp_processor_id(); |
| 1546 | |
| 1547 | #ifdef CONFIG_SMP |
| 1548 | if (unlikely(task_running(rq, p))) |
| 1549 | goto out_activate; |
| 1550 | |
| 1551 | new_cpu = p->sched_class->select_task_rq(p, sync); |
| 1552 | if (new_cpu != cpu) { |
| 1553 | set_task_cpu(p, new_cpu); |
| 1554 | task_rq_unlock(rq, &flags); |
| 1555 | /* might preempt at this point */ |
| 1556 | rq = task_rq_lock(p, &flags); |
| 1557 | old_state = p->state; |
| 1558 | if (!(old_state & state)) |
| 1559 | goto out; |
| 1560 | if (p->se.on_rq) |
| 1561 | goto out_running; |
| 1562 | |
| 1563 | this_cpu = smp_processor_id(); |
| 1564 | cpu = task_cpu(p); |
| 1565 | } |
| 1566 | |
| 1567 | #ifdef CONFIG_SCHEDSTATS |
| 1568 | schedstat_inc(rq, ttwu_count); |
| 1569 | if (cpu == this_cpu) |
| 1570 | schedstat_inc(rq, ttwu_local); |
| 1571 | else { |
| 1572 | struct sched_domain *sd; |
| 1573 | for_each_domain(this_cpu, sd) { |
| 1574 | if (cpu_isset(cpu, sd->span)) { |
| 1575 | schedstat_inc(sd, ttwu_wake_remote); |
| 1576 | break; |
| 1577 | } |
| 1578 | } |
| 1579 | } |
| 1580 | |
| 1581 | #endif |
| 1582 | |
| 1583 | |
| 1584 | out_activate: |
| 1585 | #endif /* CONFIG_SMP */ |
| 1586 | schedstat_inc(p, se.nr_wakeups); |
| 1587 | if (sync) |
| 1588 | schedstat_inc(p, se.nr_wakeups_sync); |
| 1589 | if (orig_cpu != cpu) |
| 1590 | schedstat_inc(p, se.nr_wakeups_migrate); |
| 1591 | if (cpu == this_cpu) |
| 1592 | schedstat_inc(p, se.nr_wakeups_local); |
| 1593 | else |
| 1594 | schedstat_inc(p, se.nr_wakeups_remote); |
| 1595 | update_rq_clock(rq); |
| 1596 | activate_task(rq, p, 1); |
| 1597 | check_preempt_curr(rq, p); |
| 1598 | success = 1; |
| 1599 | |
| 1600 | out_running: |
| 1601 | p->state = TASK_RUNNING; |
| 1602 | wakeup_balance_rt(rq, p); |
| 1603 | out: |
| 1604 | task_rq_unlock(rq, &flags); |
| 1605 | |
| 1606 | return success; |
| 1607 | } |
| 1608 | |
| 1609 | int fastcall wake_up_process(struct task_struct *p) |
| 1610 | { |
| 1611 | return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED | |
| 1612 | TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0); |
| 1613 | } |
| 1614 | EXPORT_SYMBOL(wake_up_process); |
| 1615 | |
| 1616 | int fastcall wake_up_state(struct task_struct *p, unsigned int state) |
| 1617 | { |
| 1618 | return try_to_wake_up(p, state, 0); |
| 1619 | } |
| 1620 | |
| 1621 | /* |
| 1622 | * Perform scheduler related setup for a newly forked process p. |
| 1623 | * p is forked by current. |
| 1624 | * |
| 1625 | * __sched_fork() is basic setup used by init_idle() too: |
| 1626 | */ |
| 1627 | static void __sched_fork(struct task_struct *p) |
| 1628 | { |
| 1629 | p->se.exec_start = 0; |
| 1630 | p->se.sum_exec_runtime = 0; |
| 1631 | p->se.prev_sum_exec_runtime = 0; |
| 1632 | |
| 1633 | #ifdef CONFIG_SCHEDSTATS |
| 1634 | p->se.wait_start = 0; |
| 1635 | p->se.sum_sleep_runtime = 0; |
| 1636 | p->se.sleep_start = 0; |
| 1637 | p->se.block_start = 0; |
| 1638 | p->se.sleep_max = 0; |
| 1639 | p->se.block_max = 0; |
| 1640 | p->se.exec_max = 0; |
| 1641 | p->se.slice_max = 0; |
| 1642 | p->se.wait_max = 0; |
| 1643 | #endif |
| 1644 | |
| 1645 | INIT_LIST_HEAD(&p->run_list); |
| 1646 | p->se.on_rq = 0; |
| 1647 | |
| 1648 | #ifdef CONFIG_PREEMPT_NOTIFIERS |
| 1649 | INIT_HLIST_HEAD(&p->preempt_notifiers); |
| 1650 | #endif |
| 1651 | |
| 1652 | /* |
| 1653 | * We mark the process as running here, but have not actually |
| 1654 | * inserted it onto the runqueue yet. This guarantees that |
| 1655 | * nobody will actually run it, and a signal or other external |
| 1656 | * event cannot wake it up and insert it on the runqueue either. |
| 1657 | */ |
| 1658 | p->state = TASK_RUNNING; |
| 1659 | } |
| 1660 | |
| 1661 | /* |
| 1662 | * fork()/clone()-time setup: |
| 1663 | */ |
| 1664 | void sched_fork(struct task_struct *p, int clone_flags) |
| 1665 | { |
| 1666 | int cpu = get_cpu(); |
| 1667 | |
| 1668 | __sched_fork(p); |
| 1669 | |
| 1670 | #ifdef CONFIG_SMP |
| 1671 | cpu = sched_balance_self(cpu, SD_BALANCE_FORK); |
| 1672 | #endif |
| 1673 | set_task_cpu(p, cpu); |
| 1674 | |
| 1675 | /* |
| 1676 | * Make sure we do not leak PI boosting priority to the child: |
| 1677 | */ |
| 1678 | p->prio = current->normal_prio; |
| 1679 | if (!rt_prio(p->prio)) |
| 1680 | p->sched_class = &fair_sched_class; |
| 1681 | |
| 1682 | #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) |
| 1683 | if (likely(sched_info_on())) |
| 1684 | memset(&p->sched_info, 0, sizeof(p->sched_info)); |
| 1685 | #endif |
| 1686 | #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) |
| 1687 | p->oncpu = 0; |
| 1688 | #endif |
| 1689 | #ifdef CONFIG_PREEMPT |
| 1690 | /* Want to start with kernel preemption disabled. */ |
| 1691 | task_thread_info(p)->preempt_count = 1; |
| 1692 | #endif |
| 1693 | put_cpu(); |
| 1694 | } |
| 1695 | |
| 1696 | /* |
| 1697 | * wake_up_new_task - wake up a newly created task for the first time. |
| 1698 | * |
| 1699 | * This function will do some initial scheduler statistics housekeeping |
| 1700 | * that must be done for every newly created context, then puts the task |
| 1701 | * on the runqueue and wakes it. |
| 1702 | */ |
| 1703 | void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags) |
| 1704 | { |
| 1705 | unsigned long flags; |
| 1706 | struct rq *rq; |
| 1707 | |
| 1708 | rq = task_rq_lock(p, &flags); |
| 1709 | BUG_ON(p->state != TASK_RUNNING); |
| 1710 | update_rq_clock(rq); |
| 1711 | |
| 1712 | p->prio = effective_prio(p); |
| 1713 | |
| 1714 | if (!p->sched_class->task_new || !current->se.on_rq) { |
| 1715 | activate_task(rq, p, 0); |
| 1716 | } else { |
| 1717 | /* |
| 1718 | * Let the scheduling class do new task startup |
| 1719 | * management (if any): |
| 1720 | */ |
| 1721 | p->sched_class->task_new(rq, p); |
| 1722 | inc_nr_running(p, rq); |
| 1723 | } |
| 1724 | check_preempt_curr(rq, p); |
| 1725 | wakeup_balance_rt(rq, p); |
| 1726 | task_rq_unlock(rq, &flags); |
| 1727 | } |
| 1728 | |
| 1729 | #ifdef CONFIG_PREEMPT_NOTIFIERS |
| 1730 | |
| 1731 | /** |
| 1732 | * preempt_notifier_register - tell me when current is being being preempted & rescheduled |
| 1733 | * @notifier: notifier struct to register |
| 1734 | */ |
| 1735 | void preempt_notifier_register(struct preempt_notifier *notifier) |
| 1736 | { |
| 1737 | hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); |
| 1738 | } |
| 1739 | EXPORT_SYMBOL_GPL(preempt_notifier_register); |
| 1740 | |
| 1741 | /** |
| 1742 | * preempt_notifier_unregister - no longer interested in preemption notifications |
| 1743 | * @notifier: notifier struct to unregister |
| 1744 | * |
| 1745 | * This is safe to call from within a preemption notifier. |
| 1746 | */ |
| 1747 | void preempt_notifier_unregister(struct preempt_notifier *notifier) |
| 1748 | { |
| 1749 | hlist_del(¬ifier->link); |
| 1750 | } |
| 1751 | EXPORT_SYMBOL_GPL(preempt_notifier_unregister); |
| 1752 | |
| 1753 | static void fire_sched_in_preempt_notifiers(struct task_struct *curr) |
| 1754 | { |
| 1755 | struct preempt_notifier *notifier; |
| 1756 | struct hlist_node *node; |
| 1757 | |
| 1758 | hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) |
| 1759 | notifier->ops->sched_in(notifier, raw_smp_processor_id()); |
| 1760 | } |
| 1761 | |
| 1762 | static void |
| 1763 | fire_sched_out_preempt_notifiers(struct task_struct *curr, |
| 1764 | struct task_struct *next) |
| 1765 | { |
| 1766 | struct preempt_notifier *notifier; |
| 1767 | struct hlist_node *node; |
| 1768 | |
| 1769 | hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) |
| 1770 | notifier->ops->sched_out(notifier, next); |
| 1771 | } |
| 1772 | |
| 1773 | #else |
| 1774 | |
| 1775 | static void fire_sched_in_preempt_notifiers(struct task_struct *curr) |
| 1776 | { |
| 1777 | } |
| 1778 | |
| 1779 | static void |
| 1780 | fire_sched_out_preempt_notifiers(struct task_struct *curr, |
| 1781 | struct task_struct *next) |
| 1782 | { |
| 1783 | } |
| 1784 | |
| 1785 | #endif |
| 1786 | |
| 1787 | /** |
| 1788 | * prepare_task_switch - prepare to switch tasks |
| 1789 | * @rq: the runqueue preparing to switch |
| 1790 | * @prev: the current task that is being switched out |
| 1791 | * @next: the task we are going to switch to. |
| 1792 | * |
| 1793 | * This is called with the rq lock held and interrupts off. It must |
| 1794 | * be paired with a subsequent finish_task_switch after the context |
| 1795 | * switch. |
| 1796 | * |
| 1797 | * prepare_task_switch sets up locking and calls architecture specific |
| 1798 | * hooks. |
| 1799 | */ |
| 1800 | static inline void |
| 1801 | prepare_task_switch(struct rq *rq, struct task_struct *prev, |
| 1802 | struct task_struct *next) |
| 1803 | { |
| 1804 | fire_sched_out_preempt_notifiers(prev, next); |
| 1805 | prepare_lock_switch(rq, next); |
| 1806 | prepare_arch_switch(next); |
| 1807 | } |
| 1808 | |
| 1809 | /** |
| 1810 | * finish_task_switch - clean up after a task-switch |
| 1811 | * @rq: runqueue associated with task-switch |
| 1812 | * @prev: the thread we just switched away from. |
| 1813 | * |
| 1814 | * finish_task_switch must be called after the context switch, paired |
| 1815 | * with a prepare_task_switch call before the context switch. |
| 1816 | * finish_task_switch will reconcile locking set up by prepare_task_switch, |
| 1817 | * and do any other architecture-specific cleanup actions. |
| 1818 | * |
| 1819 | * Note that we may have delayed dropping an mm in context_switch(). If |
| 1820 | * so, we finish that here outside of the runqueue lock. (Doing it |
| 1821 | * with the lock held can cause deadlocks; see schedule() for |
| 1822 | * details.) |
| 1823 | */ |
| 1824 | static void finish_task_switch(struct rq *rq, struct task_struct *prev) |
| 1825 | __releases(rq->lock) |
| 1826 | { |
| 1827 | struct mm_struct *mm = rq->prev_mm; |
| 1828 | long prev_state; |
| 1829 | |
| 1830 | rq->prev_mm = NULL; |
| 1831 | |
| 1832 | /* |
| 1833 | * A task struct has one reference for the use as "current". |
| 1834 | * If a task dies, then it sets TASK_DEAD in tsk->state and calls |
| 1835 | * schedule one last time. The schedule call will never return, and |
| 1836 | * the scheduled task must drop that reference. |
| 1837 | * The test for TASK_DEAD must occur while the runqueue locks are |
| 1838 | * still held, otherwise prev could be scheduled on another cpu, die |
| 1839 | * there before we look at prev->state, and then the reference would |
| 1840 | * be dropped twice. |
| 1841 | * Manfred Spraul <manfred@colorfullife.com> |
| 1842 | */ |
| 1843 | prev_state = prev->state; |
| 1844 | finish_arch_switch(prev); |
| 1845 | finish_lock_switch(rq, prev); |
| 1846 | schedule_tail_balance_rt(rq); |
| 1847 | |
| 1848 | fire_sched_in_preempt_notifiers(current); |
| 1849 | if (mm) |
| 1850 | mmdrop(mm); |
| 1851 | if (unlikely(prev_state == TASK_DEAD)) { |
| 1852 | /* |
| 1853 | * Remove function-return probe instances associated with this |
| 1854 | * task and put them back on the free list. |
| 1855 | */ |
| 1856 | kprobe_flush_task(prev); |
| 1857 | put_task_struct(prev); |
| 1858 | } |
| 1859 | } |
| 1860 | |
| 1861 | /** |
| 1862 | * schedule_tail - first thing a freshly forked thread must call. |
| 1863 | * @prev: the thread we just switched away from. |
| 1864 | */ |
| 1865 | asmlinkage void schedule_tail(struct task_struct *prev) |
| 1866 | __releases(rq->lock) |
| 1867 | { |
| 1868 | struct rq *rq = this_rq(); |
| 1869 | |
| 1870 | finish_task_switch(rq, prev); |
| 1871 | #ifdef __ARCH_WANT_UNLOCKED_CTXSW |
| 1872 | /* In this case, finish_task_switch does not reenable preemption */ |
| 1873 | preempt_enable(); |
| 1874 | #endif |
| 1875 | if (current->set_child_tid) |
| 1876 | put_user(task_pid_vnr(current), current->set_child_tid); |
| 1877 | } |
| 1878 | |
| 1879 | /* |
| 1880 | * context_switch - switch to the new MM and the new |
| 1881 | * thread's register state. |
| 1882 | */ |
| 1883 | static inline void |
| 1884 | context_switch(struct rq *rq, struct task_struct *prev, |
| 1885 | struct task_struct *next) |
| 1886 | { |
| 1887 | struct mm_struct *mm, *oldmm; |
| 1888 | |
| 1889 | prepare_task_switch(rq, prev, next); |
| 1890 | mm = next->mm; |
| 1891 | oldmm = prev->active_mm; |
| 1892 | /* |
| 1893 | * For paravirt, this is coupled with an exit in switch_to to |
| 1894 | * combine the page table reload and the switch backend into |
| 1895 | * one hypercall. |
| 1896 | */ |
| 1897 | arch_enter_lazy_cpu_mode(); |
| 1898 | |
| 1899 | if (unlikely(!mm)) { |
| 1900 | next->active_mm = oldmm; |
| 1901 | atomic_inc(&oldmm->mm_count); |
| 1902 | enter_lazy_tlb(oldmm, next); |
| 1903 | } else |
| 1904 | switch_mm(oldmm, mm, next); |
| 1905 | |
| 1906 | if (unlikely(!prev->mm)) { |
| 1907 | prev->active_mm = NULL; |
| 1908 | rq->prev_mm = oldmm; |
| 1909 | } |
| 1910 | /* |
| 1911 | * Since the runqueue lock will be released by the next |
| 1912 | * task (which is an invalid locking op but in the case |
| 1913 | * of the scheduler it's an obvious special-case), so we |
| 1914 | * do an early lockdep release here: |
| 1915 | */ |
| 1916 | #ifndef __ARCH_WANT_UNLOCKED_CTXSW |
| 1917 | spin_release(&rq->lock.dep_map, 1, _THIS_IP_); |
| 1918 | #endif |
| 1919 | |
| 1920 | /* Here we just switch the register state and the stack. */ |
| 1921 | switch_to(prev, next, prev); |
| 1922 | |
| 1923 | barrier(); |
| 1924 | /* |
| 1925 | * this_rq must be evaluated again because prev may have moved |
| 1926 | * CPUs since it called schedule(), thus the 'rq' on its stack |
| 1927 | * frame will be invalid. |
| 1928 | */ |
| 1929 | finish_task_switch(this_rq(), prev); |
| 1930 | } |
| 1931 | |
| 1932 | /* |
| 1933 | * nr_running, nr_uninterruptible and nr_context_switches: |
| 1934 | * |
| 1935 | * externally visible scheduler statistics: current number of runnable |
| 1936 | * threads, current number of uninterruptible-sleeping threads, total |
| 1937 | * number of context switches performed since bootup. |
| 1938 | */ |
| 1939 | unsigned long nr_running(void) |
| 1940 | { |
| 1941 | unsigned long i, sum = 0; |
| 1942 | |
| 1943 | for_each_online_cpu(i) |
| 1944 | sum += cpu_rq(i)->nr_running; |
| 1945 | |
| 1946 | return sum; |
| 1947 | } |
| 1948 | |
| 1949 | unsigned long nr_uninterruptible(void) |
| 1950 | { |
| 1951 | unsigned long i, sum = 0; |
| 1952 | |
| 1953 | for_each_possible_cpu(i) |
| 1954 | sum += cpu_rq(i)->nr_uninterruptible; |
| 1955 | |
| 1956 | /* |
| 1957 | * Since we read the counters lockless, it might be slightly |
| 1958 | * inaccurate. Do not allow it to go below zero though: |
| 1959 | */ |
| 1960 | if (unlikely((long)sum < 0)) |
| 1961 | sum = 0; |
| 1962 | |
| 1963 | return sum; |
| 1964 | } |
| 1965 | |
| 1966 | unsigned long long nr_context_switches(void) |
| 1967 | { |
| 1968 | int i; |
| 1969 | unsigned long long sum = 0; |
| 1970 | |
| 1971 | for_each_possible_cpu(i) |
| 1972 | sum += cpu_rq(i)->nr_switches; |
| 1973 | |
| 1974 | return sum; |
| 1975 | } |
| 1976 | |
| 1977 | unsigned long nr_iowait(void) |
| 1978 | { |
| 1979 | unsigned long i, sum = 0; |
| 1980 | |
| 1981 | for_each_possible_cpu(i) |
| 1982 | sum += atomic_read(&cpu_rq(i)->nr_iowait); |
| 1983 | |
| 1984 | return sum; |
| 1985 | } |
| 1986 | |
| 1987 | unsigned long nr_active(void) |
| 1988 | { |
| 1989 | unsigned long i, running = 0, uninterruptible = 0; |
| 1990 | |
| 1991 | for_each_online_cpu(i) { |
| 1992 | running += cpu_rq(i)->nr_running; |
| 1993 | uninterruptible += cpu_rq(i)->nr_uninterruptible; |
| 1994 | } |
| 1995 | |
| 1996 | if (unlikely((long)uninterruptible < 0)) |
| 1997 | uninterruptible = 0; |
| 1998 | |
| 1999 | return running + uninterruptible; |
| 2000 | } |
| 2001 | |
| 2002 | /* |
| 2003 | * Update rq->cpu_load[] statistics. This function is usually called every |
| 2004 | * scheduler tick (TICK_NSEC). |
| 2005 | */ |
| 2006 | static void update_cpu_load(struct rq *this_rq) |
| 2007 | { |
| 2008 | unsigned long this_load = this_rq->load.weight; |
| 2009 | int i, scale; |
| 2010 | |
| 2011 | this_rq->nr_load_updates++; |
| 2012 | |
| 2013 | /* Update our load: */ |
| 2014 | for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) { |
| 2015 | unsigned long old_load, new_load; |
| 2016 | |
| 2017 | /* scale is effectively 1 << i now, and >> i divides by scale */ |
| 2018 | |
| 2019 | old_load = this_rq->cpu_load[i]; |
| 2020 | new_load = this_load; |
| 2021 | /* |
| 2022 | * Round up the averaging division if load is increasing. This |
| 2023 | * prevents us from getting stuck on 9 if the load is 10, for |
| 2024 | * example. |
| 2025 | */ |
| 2026 | if (new_load > old_load) |
| 2027 | new_load += scale-1; |
| 2028 | this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i; |
| 2029 | } |
| 2030 | } |
| 2031 | |
| 2032 | #ifdef CONFIG_SMP |
| 2033 | |
| 2034 | /* |
| 2035 | * double_rq_lock - safely lock two runqueues |
| 2036 | * |
| 2037 | * Note this does not disable interrupts like task_rq_lock, |
| 2038 | * you need to do so manually before calling. |
| 2039 | */ |
| 2040 | static void double_rq_lock(struct rq *rq1, struct rq *rq2) |
| 2041 | __acquires(rq1->lock) |
| 2042 | __acquires(rq2->lock) |
| 2043 | { |
| 2044 | BUG_ON(!irqs_disabled()); |
| 2045 | if (rq1 == rq2) { |
| 2046 | spin_lock(&rq1->lock); |
| 2047 | __acquire(rq2->lock); /* Fake it out ;) */ |
| 2048 | } else { |
| 2049 | if (rq1 < rq2) { |
| 2050 | spin_lock(&rq1->lock); |
| 2051 | spin_lock(&rq2->lock); |
| 2052 | } else { |
| 2053 | spin_lock(&rq2->lock); |
| 2054 | spin_lock(&rq1->lock); |
| 2055 | } |
| 2056 | } |
| 2057 | update_rq_clock(rq1); |
| 2058 | update_rq_clock(rq2); |
| 2059 | } |
| 2060 | |
| 2061 | /* |
| 2062 | * double_rq_unlock - safely unlock two runqueues |
| 2063 | * |
| 2064 | * Note this does not restore interrupts like task_rq_unlock, |
| 2065 | * you need to do so manually after calling. |
| 2066 | */ |
| 2067 | static void double_rq_unlock(struct rq *rq1, struct rq *rq2) |
| 2068 | __releases(rq1->lock) |
| 2069 | __releases(rq2->lock) |
| 2070 | { |
| 2071 | spin_unlock(&rq1->lock); |
| 2072 | if (rq1 != rq2) |
| 2073 | spin_unlock(&rq2->lock); |
| 2074 | else |
| 2075 | __release(rq2->lock); |
| 2076 | } |
| 2077 | |
| 2078 | /* |
| 2079 | * double_lock_balance - lock the busiest runqueue, this_rq is locked already. |
| 2080 | */ |
| 2081 | static int double_lock_balance(struct rq *this_rq, struct rq *busiest) |
| 2082 | __releases(this_rq->lock) |
| 2083 | __acquires(busiest->lock) |
| 2084 | __acquires(this_rq->lock) |
| 2085 | { |
| 2086 | int ret = 0; |
| 2087 | |
| 2088 | if (unlikely(!irqs_disabled())) { |
| 2089 | /* printk() doesn't work good under rq->lock */ |
| 2090 | spin_unlock(&this_rq->lock); |
| 2091 | BUG_ON(1); |
| 2092 | } |
| 2093 | if (unlikely(!spin_trylock(&busiest->lock))) { |
| 2094 | if (busiest < this_rq) { |
| 2095 | spin_unlock(&this_rq->lock); |
| 2096 | spin_lock(&busiest->lock); |
| 2097 | spin_lock(&this_rq->lock); |
| 2098 | ret = 1; |
| 2099 | } else |
| 2100 | spin_lock(&busiest->lock); |
| 2101 | } |
| 2102 | return ret; |
| 2103 | } |
| 2104 | |
| 2105 | /* |
| 2106 | * If dest_cpu is allowed for this process, migrate the task to it. |
| 2107 | * This is accomplished by forcing the cpu_allowed mask to only |
| 2108 | * allow dest_cpu, which will force the cpu onto dest_cpu. Then |
| 2109 | * the cpu_allowed mask is restored. |
| 2110 | */ |
| 2111 | static void sched_migrate_task(struct task_struct *p, int dest_cpu) |
| 2112 | { |
| 2113 | struct migration_req req; |
| 2114 | unsigned long flags; |
| 2115 | struct rq *rq; |
| 2116 | |
| 2117 | rq = task_rq_lock(p, &flags); |
| 2118 | if (!cpu_isset(dest_cpu, p->cpus_allowed) |
| 2119 | || unlikely(cpu_is_offline(dest_cpu))) |
| 2120 | goto out; |
| 2121 | |
| 2122 | /* force the process onto the specified CPU */ |
| 2123 | if (migrate_task(p, dest_cpu, &req)) { |
| 2124 | /* Need to wait for migration thread (might exit: take ref). */ |
| 2125 | struct task_struct *mt = rq->migration_thread; |
| 2126 | |
| 2127 | get_task_struct(mt); |
| 2128 | task_rq_unlock(rq, &flags); |
| 2129 | wake_up_process(mt); |
| 2130 | put_task_struct(mt); |
| 2131 | wait_for_completion(&req.done); |
| 2132 | |
| 2133 | return; |
| 2134 | } |
| 2135 | out: |
| 2136 | task_rq_unlock(rq, &flags); |
| 2137 | } |
| 2138 | |
| 2139 | /* |
| 2140 | * sched_exec - execve() is a valuable balancing opportunity, because at |
| 2141 | * this point the task has the smallest effective memory and cache footprint. |
| 2142 | */ |
| 2143 | void sched_exec(void) |
| 2144 | { |
| 2145 | int new_cpu, this_cpu = get_cpu(); |
| 2146 | new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC); |
| 2147 | put_cpu(); |
| 2148 | if (new_cpu != this_cpu) |
| 2149 | sched_migrate_task(current, new_cpu); |
| 2150 | } |
| 2151 | |
| 2152 | /* |
| 2153 | * pull_task - move a task from a remote runqueue to the local runqueue. |
| 2154 | * Both runqueues must be locked. |
| 2155 | */ |
| 2156 | static void pull_task(struct rq *src_rq, struct task_struct *p, |
| 2157 | struct rq *this_rq, int this_cpu) |
| 2158 | { |
| 2159 | deactivate_task(src_rq, p, 0); |
| 2160 | set_task_cpu(p, this_cpu); |
| 2161 | activate_task(this_rq, p, 0); |
| 2162 | /* |
| 2163 | * Note that idle threads have a prio of MAX_PRIO, for this test |
| 2164 | * to be always true for them. |
| 2165 | */ |
| 2166 | check_preempt_curr(this_rq, p); |
| 2167 | } |
| 2168 | |
| 2169 | /* |
| 2170 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? |
| 2171 | */ |
| 2172 | static |
| 2173 | int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu, |
| 2174 | struct sched_domain *sd, enum cpu_idle_type idle, |
| 2175 | int *all_pinned) |
| 2176 | { |
| 2177 | /* |
| 2178 | * We do not migrate tasks that are: |
| 2179 | * 1) running (obviously), or |
| 2180 | * 2) cannot be migrated to this CPU due to cpus_allowed, or |
| 2181 | * 3) are cache-hot on their current CPU. |
| 2182 | */ |
| 2183 | if (!cpu_isset(this_cpu, p->cpus_allowed)) { |
| 2184 | schedstat_inc(p, se.nr_failed_migrations_affine); |
| 2185 | return 0; |
| 2186 | } |
| 2187 | *all_pinned = 0; |
| 2188 | |
| 2189 | if (task_running(rq, p)) { |
| 2190 | schedstat_inc(p, se.nr_failed_migrations_running); |
| 2191 | return 0; |
| 2192 | } |
| 2193 | |
| 2194 | /* |
| 2195 | * Aggressive migration if: |
| 2196 | * 1) task is cache cold, or |
| 2197 | * 2) too many balance attempts have failed. |
| 2198 | */ |
| 2199 | |
| 2200 | if (!task_hot(p, rq->clock, sd) || |
| 2201 | sd->nr_balance_failed > sd->cache_nice_tries) { |
| 2202 | #ifdef CONFIG_SCHEDSTATS |
| 2203 | if (task_hot(p, rq->clock, sd)) { |
| 2204 | schedstat_inc(sd, lb_hot_gained[idle]); |
| 2205 | schedstat_inc(p, se.nr_forced_migrations); |
| 2206 | } |
| 2207 | #endif |
| 2208 | return 1; |
| 2209 | } |
| 2210 | |
| 2211 | if (task_hot(p, rq->clock, sd)) { |
| 2212 | schedstat_inc(p, se.nr_failed_migrations_hot); |
| 2213 | return 0; |
| 2214 | } |
| 2215 | return 1; |
| 2216 | } |
| 2217 | |
| 2218 | static unsigned long |
| 2219 | balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, |
| 2220 | unsigned long max_load_move, struct sched_domain *sd, |
| 2221 | enum cpu_idle_type idle, int *all_pinned, |
| 2222 | int *this_best_prio, struct rq_iterator *iterator) |
| 2223 | { |
| 2224 | int loops = 0, pulled = 0, pinned = 0, skip_for_load; |
| 2225 | struct task_struct *p; |
| 2226 | long rem_load_move = max_load_move; |
| 2227 | |
| 2228 | if (max_load_move == 0) |
| 2229 | goto out; |
| 2230 | |
| 2231 | pinned = 1; |
| 2232 | |
| 2233 | /* |
| 2234 | * Start the load-balancing iterator: |
| 2235 | */ |
| 2236 | p = iterator->start(iterator->arg); |
| 2237 | next: |
| 2238 | if (!p || loops++ > sysctl_sched_nr_migrate) |
| 2239 | goto out; |
| 2240 | /* |
| 2241 | * To help distribute high priority tasks across CPUs we don't |
| 2242 | * skip a task if it will be the highest priority task (i.e. smallest |
| 2243 | * prio value) on its new queue regardless of its load weight |
| 2244 | */ |
| 2245 | skip_for_load = (p->se.load.weight >> 1) > rem_load_move + |
| 2246 | SCHED_LOAD_SCALE_FUZZ; |
| 2247 | if ((skip_for_load && p->prio >= *this_best_prio) || |
| 2248 | !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) { |
| 2249 | p = iterator->next(iterator->arg); |
| 2250 | goto next; |
| 2251 | } |
| 2252 | |
| 2253 | pull_task(busiest, p, this_rq, this_cpu); |
| 2254 | pulled++; |
| 2255 | rem_load_move -= p->se.load.weight; |
| 2256 | |
| 2257 | /* |
| 2258 | * We only want to steal up to the prescribed amount of weighted load. |
| 2259 | */ |
| 2260 | if (rem_load_move > 0) { |
| 2261 | if (p->prio < *this_best_prio) |
| 2262 | *this_best_prio = p->prio; |
| 2263 | p = iterator->next(iterator->arg); |
| 2264 | goto next; |
| 2265 | } |
| 2266 | out: |
| 2267 | /* |
| 2268 | * Right now, this is one of only two places pull_task() is called, |
| 2269 | * so we can safely collect pull_task() stats here rather than |
| 2270 | * inside pull_task(). |
| 2271 | */ |
| 2272 | schedstat_add(sd, lb_gained[idle], pulled); |
| 2273 | |
| 2274 | if (all_pinned) |
| 2275 | *all_pinned = pinned; |
| 2276 | |
| 2277 | return max_load_move - rem_load_move; |
| 2278 | } |
| 2279 | |
| 2280 | /* |
| 2281 | * move_tasks tries to move up to max_load_move weighted load from busiest to |
| 2282 | * this_rq, as part of a balancing operation within domain "sd". |
| 2283 | * Returns 1 if successful and 0 otherwise. |
| 2284 | * |
| 2285 | * Called with both runqueues locked. |
| 2286 | */ |
| 2287 | static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, |
| 2288 | unsigned long max_load_move, |
| 2289 | struct sched_domain *sd, enum cpu_idle_type idle, |
| 2290 | int *all_pinned) |
| 2291 | { |
| 2292 | const struct sched_class *class = sched_class_highest; |
| 2293 | unsigned long total_load_moved = 0; |
| 2294 | int this_best_prio = this_rq->curr->prio; |
| 2295 | |
| 2296 | do { |
| 2297 | total_load_moved += |
| 2298 | class->load_balance(this_rq, this_cpu, busiest, |
| 2299 | max_load_move - total_load_moved, |
| 2300 | sd, idle, all_pinned, &this_best_prio); |
| 2301 | class = class->next; |
| 2302 | } while (class && max_load_move > total_load_moved); |
| 2303 | |
| 2304 | return total_load_moved > 0; |
| 2305 | } |
| 2306 | |
| 2307 | static int |
| 2308 | iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest, |
| 2309 | struct sched_domain *sd, enum cpu_idle_type idle, |
| 2310 | struct rq_iterator *iterator) |
| 2311 | { |
| 2312 | struct task_struct *p = iterator->start(iterator->arg); |
| 2313 | int pinned = 0; |
| 2314 | |
| 2315 | while (p) { |
| 2316 | if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) { |
| 2317 | pull_task(busiest, p, this_rq, this_cpu); |
| 2318 | /* |
| 2319 | * Right now, this is only the second place pull_task() |
| 2320 | * is called, so we can safely collect pull_task() |
| 2321 | * stats here rather than inside pull_task(). |
| 2322 | */ |
| 2323 | schedstat_inc(sd, lb_gained[idle]); |
| 2324 | |
| 2325 | return 1; |
| 2326 | } |
| 2327 | p = iterator->next(iterator->arg); |
| 2328 | } |
| 2329 | |
| 2330 | return 0; |
| 2331 | } |
| 2332 | |
| 2333 | /* |
| 2334 | * move_one_task tries to move exactly one task from busiest to this_rq, as |
| 2335 | * part of active balancing operations within "domain". |
| 2336 | * Returns 1 if successful and 0 otherwise. |
| 2337 | * |
| 2338 | * Called with both runqueues locked. |
| 2339 | */ |
| 2340 | static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest, |
| 2341 | struct sched_domain *sd, enum cpu_idle_type idle) |
| 2342 | { |
| 2343 | const struct sched_class *class; |
| 2344 | |
| 2345 | for (class = sched_class_highest; class; class = class->next) |
| 2346 | if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle)) |
| 2347 | return 1; |
| 2348 | |
| 2349 | return 0; |
| 2350 | } |
| 2351 | |
| 2352 | /* |
| 2353 | * find_busiest_group finds and returns the busiest CPU group within the |
| 2354 | * domain. It calculates and returns the amount of weighted load which |
| 2355 | * should be moved to restore balance via the imbalance parameter. |
| 2356 | */ |
| 2357 | static struct sched_group * |
| 2358 | find_busiest_group(struct sched_domain *sd, int this_cpu, |
| 2359 | unsigned long *imbalance, enum cpu_idle_type idle, |
| 2360 | int *sd_idle, cpumask_t *cpus, int *balance) |
| 2361 | { |
| 2362 | struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups; |
| 2363 | unsigned long max_load, avg_load, total_load, this_load, total_pwr; |
| 2364 | unsigned long max_pull; |
| 2365 | unsigned long busiest_load_per_task, busiest_nr_running; |
| 2366 | unsigned long this_load_per_task, this_nr_running; |
| 2367 | int load_idx, group_imb = 0; |
| 2368 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) |
| 2369 | int power_savings_balance = 1; |
| 2370 | unsigned long leader_nr_running = 0, min_load_per_task = 0; |
| 2371 | unsigned long min_nr_running = ULONG_MAX; |
| 2372 | struct sched_group *group_min = NULL, *group_leader = NULL; |
| 2373 | #endif |
| 2374 | |
| 2375 | max_load = this_load = total_load = total_pwr = 0; |
| 2376 | busiest_load_per_task = busiest_nr_running = 0; |
| 2377 | this_load_per_task = this_nr_running = 0; |
| 2378 | if (idle == CPU_NOT_IDLE) |
| 2379 | load_idx = sd->busy_idx; |
| 2380 | else if (idle == CPU_NEWLY_IDLE) |
| 2381 | load_idx = sd->newidle_idx; |
| 2382 | else |
| 2383 | load_idx = sd->idle_idx; |
| 2384 | |
| 2385 | do { |
| 2386 | unsigned long load, group_capacity, max_cpu_load, min_cpu_load; |
| 2387 | int local_group; |
| 2388 | int i; |
| 2389 | int __group_imb = 0; |
| 2390 | unsigned int balance_cpu = -1, first_idle_cpu = 0; |
| 2391 | unsigned long sum_nr_running, sum_weighted_load; |
| 2392 | |
| 2393 | local_group = cpu_isset(this_cpu, group->cpumask); |
| 2394 | |
| 2395 | if (local_group) |
| 2396 | balance_cpu = first_cpu(group->cpumask); |
| 2397 | |
| 2398 | /* Tally up the load of all CPUs in the group */ |
| 2399 | sum_weighted_load = sum_nr_running = avg_load = 0; |
| 2400 | max_cpu_load = 0; |
| 2401 | min_cpu_load = ~0UL; |
| 2402 | |
| 2403 | for_each_cpu_mask(i, group->cpumask) { |
| 2404 | struct rq *rq; |
| 2405 | |
| 2406 | if (!cpu_isset(i, *cpus)) |
| 2407 | continue; |
| 2408 | |
| 2409 | rq = cpu_rq(i); |
| 2410 | |
| 2411 | if (*sd_idle && rq->nr_running) |
| 2412 | *sd_idle = 0; |
| 2413 | |
| 2414 | /* Bias balancing toward cpus of our domain */ |
| 2415 | if (local_group) { |
| 2416 | if (idle_cpu(i) && !first_idle_cpu) { |
| 2417 | first_idle_cpu = 1; |
| 2418 | balance_cpu = i; |
| 2419 | } |
| 2420 | |
| 2421 | load = target_load(i, load_idx); |
| 2422 | } else { |
| 2423 | load = source_load(i, load_idx); |
| 2424 | if (load > max_cpu_load) |
| 2425 | max_cpu_load = load; |
| 2426 | if (min_cpu_load > load) |
| 2427 | min_cpu_load = load; |
| 2428 | } |
| 2429 | |
| 2430 | avg_load += load; |
| 2431 | sum_nr_running += rq->nr_running; |
| 2432 | sum_weighted_load += weighted_cpuload(i); |
| 2433 | } |
| 2434 | |
| 2435 | /* |
| 2436 | * First idle cpu or the first cpu(busiest) in this sched group |
| 2437 | * is eligible for doing load balancing at this and above |
| 2438 | * domains. In the newly idle case, we will allow all the cpu's |
| 2439 | * to do the newly idle load balance. |
| 2440 | */ |
| 2441 | if (idle != CPU_NEWLY_IDLE && local_group && |
| 2442 | balance_cpu != this_cpu && balance) { |
| 2443 | *balance = 0; |
| 2444 | goto ret; |
| 2445 | } |
| 2446 | |
| 2447 | total_load += avg_load; |
| 2448 | total_pwr += group->__cpu_power; |
| 2449 | |
| 2450 | /* Adjust by relative CPU power of the group */ |
| 2451 | avg_load = sg_div_cpu_power(group, |
| 2452 | avg_load * SCHED_LOAD_SCALE); |
| 2453 | |
| 2454 | if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE) |
| 2455 | __group_imb = 1; |
| 2456 | |
| 2457 | group_capacity = group->__cpu_power / SCHED_LOAD_SCALE; |
| 2458 | |
| 2459 | if (local_group) { |
| 2460 | this_load = avg_load; |
| 2461 | this = group; |
| 2462 | this_nr_running = sum_nr_running; |
| 2463 | this_load_per_task = sum_weighted_load; |
| 2464 | } else if (avg_load > max_load && |
| 2465 | (sum_nr_running > group_capacity || __group_imb)) { |
| 2466 | max_load = avg_load; |
| 2467 | busiest = group; |
| 2468 | busiest_nr_running = sum_nr_running; |
| 2469 | busiest_load_per_task = sum_weighted_load; |
| 2470 | group_imb = __group_imb; |
| 2471 | } |
| 2472 | |
| 2473 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) |
| 2474 | /* |
| 2475 | * Busy processors will not participate in power savings |
| 2476 | * balance. |
| 2477 | */ |
| 2478 | if (idle == CPU_NOT_IDLE || |
| 2479 | !(sd->flags & SD_POWERSAVINGS_BALANCE)) |
| 2480 | goto group_next; |
| 2481 | |
| 2482 | /* |
| 2483 | * If the local group is idle or completely loaded |
| 2484 | * no need to do power savings balance at this domain |
| 2485 | */ |
| 2486 | if (local_group && (this_nr_running >= group_capacity || |
| 2487 | !this_nr_running)) |
| 2488 | power_savings_balance = 0; |
| 2489 | |
| 2490 | /* |
| 2491 | * If a group is already running at full capacity or idle, |
| 2492 | * don't include that group in power savings calculations |
| 2493 | */ |
| 2494 | if (!power_savings_balance || sum_nr_running >= group_capacity |
| 2495 | || !sum_nr_running) |
| 2496 | goto group_next; |
| 2497 | |
| 2498 | /* |
| 2499 | * Calculate the group which has the least non-idle load. |
| 2500 | * This is the group from where we need to pick up the load |
| 2501 | * for saving power |
| 2502 | */ |
| 2503 | if ((sum_nr_running < min_nr_running) || |
| 2504 | (sum_nr_running == min_nr_running && |
| 2505 | first_cpu(group->cpumask) < |
| 2506 | first_cpu(group_min->cpumask))) { |
| 2507 | group_min = group; |
| 2508 | min_nr_running = sum_nr_running; |
| 2509 | min_load_per_task = sum_weighted_load / |
| 2510 | sum_nr_running; |
| 2511 | } |
| 2512 | |
| 2513 | /* |
| 2514 | * Calculate the group which is almost near its |
| 2515 | * capacity but still has some space to pick up some load |
| 2516 | * from other group and save more power |
| 2517 | */ |
| 2518 | if (sum_nr_running <= group_capacity - 1) { |
| 2519 | if (sum_nr_running > leader_nr_running || |
| 2520 | (sum_nr_running == leader_nr_running && |
| 2521 | first_cpu(group->cpumask) > |
| 2522 | first_cpu(group_leader->cpumask))) { |
| 2523 | group_leader = group; |
| 2524 | leader_nr_running = sum_nr_running; |
| 2525 | } |
| 2526 | } |
| 2527 | group_next: |
| 2528 | #endif |
| 2529 | group = group->next; |
| 2530 | } while (group != sd->groups); |
| 2531 | |
| 2532 | if (!busiest || this_load >= max_load || busiest_nr_running == 0) |
| 2533 | goto out_balanced; |
| 2534 | |
| 2535 | avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr; |
| 2536 | |
| 2537 | if (this_load >= avg_load || |
| 2538 | 100*max_load <= sd->imbalance_pct*this_load) |
| 2539 | goto out_balanced; |
| 2540 | |
| 2541 | busiest_load_per_task /= busiest_nr_running; |
| 2542 | if (group_imb) |
| 2543 | busiest_load_per_task = min(busiest_load_per_task, avg_load); |
| 2544 | |
| 2545 | /* |
| 2546 | * We're trying to get all the cpus to the average_load, so we don't |
| 2547 | * want to push ourselves above the average load, nor do we wish to |
| 2548 | * reduce the max loaded cpu below the average load, as either of these |
| 2549 | * actions would just result in more rebalancing later, and ping-pong |
| 2550 | * tasks around. Thus we look for the minimum possible imbalance. |
| 2551 | * Negative imbalances (*we* are more loaded than anyone else) will |
| 2552 | * be counted as no imbalance for these purposes -- we can't fix that |
| 2553 | * by pulling tasks to us. Be careful of negative numbers as they'll |
| 2554 | * appear as very large values with unsigned longs. |
| 2555 | */ |
| 2556 | if (max_load <= busiest_load_per_task) |
| 2557 | goto out_balanced; |
| 2558 | |
| 2559 | /* |
| 2560 | * In the presence of smp nice balancing, certain scenarios can have |
| 2561 | * max load less than avg load(as we skip the groups at or below |
| 2562 | * its cpu_power, while calculating max_load..) |
| 2563 | */ |
| 2564 | if (max_load < avg_load) { |
| 2565 | *imbalance = 0; |
| 2566 | goto small_imbalance; |
| 2567 | } |
| 2568 | |
| 2569 | /* Don't want to pull so many tasks that a group would go idle */ |
| 2570 | max_pull = min(max_load - avg_load, max_load - busiest_load_per_task); |
| 2571 | |
| 2572 | /* How much load to actually move to equalise the imbalance */ |
| 2573 | *imbalance = min(max_pull * busiest->__cpu_power, |
| 2574 | (avg_load - this_load) * this->__cpu_power) |
| 2575 | / SCHED_LOAD_SCALE; |
| 2576 | |
| 2577 | /* |
| 2578 | * if *imbalance is less than the average load per runnable task |
| 2579 | * there is no gaurantee that any tasks will be moved so we'll have |
| 2580 | * a think about bumping its value to force at least one task to be |
| 2581 | * moved |
| 2582 | */ |
| 2583 | if (*imbalance < busiest_load_per_task) { |
| 2584 | unsigned long tmp, pwr_now, pwr_move; |
| 2585 | unsigned int imbn; |
| 2586 | |
| 2587 | small_imbalance: |
| 2588 | pwr_move = pwr_now = 0; |
| 2589 | imbn = 2; |
| 2590 | if (this_nr_running) { |
| 2591 | this_load_per_task /= this_nr_running; |
| 2592 | if (busiest_load_per_task > this_load_per_task) |
| 2593 | imbn = 1; |
| 2594 | } else |
| 2595 | this_load_per_task = SCHED_LOAD_SCALE; |
| 2596 | |
| 2597 | if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >= |
| 2598 | busiest_load_per_task * imbn) { |
| 2599 | *imbalance = busiest_load_per_task; |
| 2600 | return busiest; |
| 2601 | } |
| 2602 | |
| 2603 | /* |
| 2604 | * OK, we don't have enough imbalance to justify moving tasks, |
| 2605 | * however we may be able to increase total CPU power used by |
| 2606 | * moving them. |
| 2607 | */ |
| 2608 | |
| 2609 | pwr_now += busiest->__cpu_power * |
| 2610 | min(busiest_load_per_task, max_load); |
| 2611 | pwr_now += this->__cpu_power * |
| 2612 | min(this_load_per_task, this_load); |
| 2613 | pwr_now /= SCHED_LOAD_SCALE; |
| 2614 | |
| 2615 | /* Amount of load we'd subtract */ |
| 2616 | tmp = sg_div_cpu_power(busiest, |
| 2617 | busiest_load_per_task * SCHED_LOAD_SCALE); |
| 2618 | if (max_load > tmp) |
| 2619 | pwr_move += busiest->__cpu_power * |
| 2620 | min(busiest_load_per_task, max_load - tmp); |
| 2621 | |
| 2622 | /* Amount of load we'd add */ |
| 2623 | if (max_load * busiest->__cpu_power < |
| 2624 | busiest_load_per_task * SCHED_LOAD_SCALE) |
| 2625 | tmp = sg_div_cpu_power(this, |
| 2626 | max_load * busiest->__cpu_power); |
| 2627 | else |
| 2628 | tmp = sg_div_cpu_power(this, |
| 2629 | busiest_load_per_task * SCHED_LOAD_SCALE); |
| 2630 | pwr_move += this->__cpu_power * |
| 2631 | min(this_load_per_task, this_load + tmp); |
| 2632 | pwr_move /= SCHED_LOAD_SCALE; |
| 2633 | |
| 2634 | /* Move if we gain throughput */ |
| 2635 | if (pwr_move > pwr_now) |
| 2636 | *imbalance = busiest_load_per_task; |
| 2637 | } |
| 2638 | |
| 2639 | return busiest; |
| 2640 | |
| 2641 | out_balanced: |
| 2642 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) |
| 2643 | if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE)) |
| 2644 | goto ret; |
| 2645 | |
| 2646 | if (this == group_leader && group_leader != group_min) { |
| 2647 | *imbalance = min_load_per_task; |
| 2648 | return group_min; |
| 2649 | } |
| 2650 | #endif |
| 2651 | ret: |
| 2652 | *imbalance = 0; |
| 2653 | return NULL; |
| 2654 | } |
| 2655 | |
| 2656 | /* |
| 2657 | * find_busiest_queue - find the busiest runqueue among the cpus in group. |
| 2658 | */ |
| 2659 | static struct rq * |
| 2660 | find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle, |
| 2661 | unsigned long imbalance, cpumask_t *cpus) |
| 2662 | { |
| 2663 | struct rq *busiest = NULL, *rq; |
| 2664 | unsigned long max_load = 0; |
| 2665 | int i; |
| 2666 | |
| 2667 | for_each_cpu_mask(i, group->cpumask) { |
| 2668 | unsigned long wl; |
| 2669 | |
| 2670 | if (!cpu_isset(i, *cpus)) |
| 2671 | continue; |
| 2672 | |
| 2673 | rq = cpu_rq(i); |
| 2674 | wl = weighted_cpuload(i); |
| 2675 | |
| 2676 | if (rq->nr_running == 1 && wl > imbalance) |
| 2677 | continue; |
| 2678 | |
| 2679 | if (wl > max_load) { |
| 2680 | max_load = wl; |
| 2681 | busiest = rq; |
| 2682 | } |
| 2683 | } |
| 2684 | |
| 2685 | return busiest; |
| 2686 | } |
| 2687 | |
| 2688 | /* |
| 2689 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but |
| 2690 | * so long as it is large enough. |
| 2691 | */ |
| 2692 | #define MAX_PINNED_INTERVAL 512 |
| 2693 | |
| 2694 | /* |
| 2695 | * Check this_cpu to ensure it is balanced within domain. Attempt to move |
| 2696 | * tasks if there is an imbalance. |
| 2697 | */ |
| 2698 | static int load_balance(int this_cpu, struct rq *this_rq, |
| 2699 | struct sched_domain *sd, enum cpu_idle_type idle, |
| 2700 | int *balance) |
| 2701 | { |
| 2702 | int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0; |
| 2703 | struct sched_group *group; |
| 2704 | unsigned long imbalance; |
| 2705 | struct rq *busiest; |
| 2706 | cpumask_t cpus = CPU_MASK_ALL; |
| 2707 | unsigned long flags; |
| 2708 | |
| 2709 | /* |
| 2710 | * When power savings policy is enabled for the parent domain, idle |
| 2711 | * sibling can pick up load irrespective of busy siblings. In this case, |
| 2712 | * let the state of idle sibling percolate up as CPU_IDLE, instead of |
| 2713 | * portraying it as CPU_NOT_IDLE. |
| 2714 | */ |
| 2715 | if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER && |
| 2716 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) |
| 2717 | sd_idle = 1; |
| 2718 | |
| 2719 | schedstat_inc(sd, lb_count[idle]); |
| 2720 | |
| 2721 | redo: |
| 2722 | group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle, |
| 2723 | &cpus, balance); |
| 2724 | |
| 2725 | if (*balance == 0) |
| 2726 | goto out_balanced; |
| 2727 | |
| 2728 | if (!group) { |
| 2729 | schedstat_inc(sd, lb_nobusyg[idle]); |
| 2730 | goto out_balanced; |
| 2731 | } |
| 2732 | |
| 2733 | busiest = find_busiest_queue(group, idle, imbalance, &cpus); |
| 2734 | if (!busiest) { |
| 2735 | schedstat_inc(sd, lb_nobusyq[idle]); |
| 2736 | goto out_balanced; |
| 2737 | } |
| 2738 | |
| 2739 | BUG_ON(busiest == this_rq); |
| 2740 | |
| 2741 | schedstat_add(sd, lb_imbalance[idle], imbalance); |
| 2742 | |
| 2743 | ld_moved = 0; |
| 2744 | if (busiest->nr_running > 1) { |
| 2745 | /* |
| 2746 | * Attempt to move tasks. If find_busiest_group has found |
| 2747 | * an imbalance but busiest->nr_running <= 1, the group is |
| 2748 | * still unbalanced. ld_moved simply stays zero, so it is |
| 2749 | * correctly treated as an imbalance. |
| 2750 | */ |
| 2751 | local_irq_save(flags); |
| 2752 | double_rq_lock(this_rq, busiest); |
| 2753 | ld_moved = move_tasks(this_rq, this_cpu, busiest, |
| 2754 | imbalance, sd, idle, &all_pinned); |
| 2755 | double_rq_unlock(this_rq, busiest); |
| 2756 | local_irq_restore(flags); |
| 2757 | |
| 2758 | /* |
| 2759 | * some other cpu did the load balance for us. |
| 2760 | */ |
| 2761 | if (ld_moved && this_cpu != smp_processor_id()) |
| 2762 | resched_cpu(this_cpu); |
| 2763 | |
| 2764 | /* All tasks on this runqueue were pinned by CPU affinity */ |
| 2765 | if (unlikely(all_pinned)) { |
| 2766 | cpu_clear(cpu_of(busiest), cpus); |
| 2767 | if (!cpus_empty(cpus)) |
| 2768 | goto redo; |
| 2769 | goto out_balanced; |
| 2770 | } |
| 2771 | } |
| 2772 | |
| 2773 | if (!ld_moved) { |
| 2774 | schedstat_inc(sd, lb_failed[idle]); |
| 2775 | sd->nr_balance_failed++; |
| 2776 | |
| 2777 | if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) { |
| 2778 | |
| 2779 | spin_lock_irqsave(&busiest->lock, flags); |
| 2780 | |
| 2781 | /* don't kick the migration_thread, if the curr |
| 2782 | * task on busiest cpu can't be moved to this_cpu |
| 2783 | */ |
| 2784 | if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) { |
| 2785 | spin_unlock_irqrestore(&busiest->lock, flags); |
| 2786 | all_pinned = 1; |
| 2787 | goto out_one_pinned; |
| 2788 | } |
| 2789 | |
| 2790 | if (!busiest->active_balance) { |
| 2791 | busiest->active_balance = 1; |
| 2792 | busiest->push_cpu = this_cpu; |
| 2793 | active_balance = 1; |
| 2794 | } |
| 2795 | spin_unlock_irqrestore(&busiest->lock, flags); |
| 2796 | if (active_balance) |
| 2797 | wake_up_process(busiest->migration_thread); |
| 2798 | |
| 2799 | /* |
| 2800 | * We've kicked active balancing, reset the failure |
| 2801 | * counter. |
| 2802 | */ |
| 2803 | sd->nr_balance_failed = sd->cache_nice_tries+1; |
| 2804 | } |
| 2805 | } else |
| 2806 | sd->nr_balance_failed = 0; |
| 2807 | |
| 2808 | if (likely(!active_balance)) { |
| 2809 | /* We were unbalanced, so reset the balancing interval */ |
| 2810 | sd->balance_interval = sd->min_interval; |
| 2811 | } else { |
| 2812 | /* |
| 2813 | * If we've begun active balancing, start to back off. This |
| 2814 | * case may not be covered by the all_pinned logic if there |
| 2815 | * is only 1 task on the busy runqueue (because we don't call |
| 2816 | * move_tasks). |
| 2817 | */ |
| 2818 | if (sd->balance_interval < sd->max_interval) |
| 2819 | sd->balance_interval *= 2; |
| 2820 | } |
| 2821 | |
| 2822 | if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER && |
| 2823 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) |
| 2824 | return -1; |
| 2825 | return ld_moved; |
| 2826 | |
| 2827 | out_balanced: |
| 2828 | schedstat_inc(sd, lb_balanced[idle]); |
| 2829 | |
| 2830 | sd->nr_balance_failed = 0; |
| 2831 | |
| 2832 | out_one_pinned: |
| 2833 | /* tune up the balancing interval */ |
| 2834 | if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) || |
| 2835 | (sd->balance_interval < sd->max_interval)) |
| 2836 | sd->balance_interval *= 2; |
| 2837 | |
| 2838 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && |
| 2839 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) |
| 2840 | return -1; |
| 2841 | return 0; |
| 2842 | } |
| 2843 | |
| 2844 | /* |
| 2845 | * Check this_cpu to ensure it is balanced within domain. Attempt to move |
| 2846 | * tasks if there is an imbalance. |
| 2847 | * |
| 2848 | * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE). |
| 2849 | * this_rq is locked. |
| 2850 | */ |
| 2851 | static int |
| 2852 | load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd) |
| 2853 | { |
| 2854 | struct sched_group *group; |
| 2855 | struct rq *busiest = NULL; |
| 2856 | unsigned long imbalance; |
| 2857 | int ld_moved = 0; |
| 2858 | int sd_idle = 0; |
| 2859 | int all_pinned = 0; |
| 2860 | cpumask_t cpus = CPU_MASK_ALL; |
| 2861 | |
| 2862 | /* |
| 2863 | * When power savings policy is enabled for the parent domain, idle |
| 2864 | * sibling can pick up load irrespective of busy siblings. In this case, |
| 2865 | * let the state of idle sibling percolate up as IDLE, instead of |
| 2866 | * portraying it as CPU_NOT_IDLE. |
| 2867 | */ |
| 2868 | if (sd->flags & SD_SHARE_CPUPOWER && |
| 2869 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) |
| 2870 | sd_idle = 1; |
| 2871 | |
| 2872 | schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]); |
| 2873 | redo: |
| 2874 | group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE, |
| 2875 | &sd_idle, &cpus, NULL); |
| 2876 | if (!group) { |
| 2877 | schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]); |
| 2878 | goto out_balanced; |
| 2879 | } |
| 2880 | |
| 2881 | busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, |
| 2882 | &cpus); |
| 2883 | if (!busiest) { |
| 2884 | schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]); |
| 2885 | goto out_balanced; |
| 2886 | } |
| 2887 | |
| 2888 | BUG_ON(busiest == this_rq); |
| 2889 | |
| 2890 | schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance); |
| 2891 | |
| 2892 | ld_moved = 0; |
| 2893 | if (busiest->nr_running > 1) { |
| 2894 | /* Attempt to move tasks */ |
| 2895 | double_lock_balance(this_rq, busiest); |
| 2896 | /* this_rq->clock is already updated */ |
| 2897 | update_rq_clock(busiest); |
| 2898 | ld_moved = move_tasks(this_rq, this_cpu, busiest, |
| 2899 | imbalance, sd, CPU_NEWLY_IDLE, |
| 2900 | &all_pinned); |
| 2901 | spin_unlock(&busiest->lock); |
| 2902 | |
| 2903 | if (unlikely(all_pinned)) { |
| 2904 | cpu_clear(cpu_of(busiest), cpus); |
| 2905 | if (!cpus_empty(cpus)) |
| 2906 | goto redo; |
| 2907 | } |
| 2908 | } |
| 2909 | |
| 2910 | if (!ld_moved) { |
| 2911 | schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]); |
| 2912 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && |
| 2913 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) |
| 2914 | return -1; |
| 2915 | } else |
| 2916 | sd->nr_balance_failed = 0; |
| 2917 | |
| 2918 | return ld_moved; |
| 2919 | |
| 2920 | out_balanced: |
| 2921 | schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]); |
| 2922 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && |
| 2923 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) |
| 2924 | return -1; |
| 2925 | sd->nr_balance_failed = 0; |
| 2926 | |
| 2927 | return 0; |
| 2928 | } |
| 2929 | |
| 2930 | /* |
| 2931 | * idle_balance is called by schedule() if this_cpu is about to become |
| 2932 | * idle. Attempts to pull tasks from other CPUs. |
| 2933 | */ |
| 2934 | static void idle_balance(int this_cpu, struct rq *this_rq) |
| 2935 | { |
| 2936 | struct sched_domain *sd; |
| 2937 | int pulled_task = -1; |
| 2938 | unsigned long next_balance = jiffies + HZ; |
| 2939 | |
| 2940 | for_each_domain(this_cpu, sd) { |
| 2941 | unsigned long interval; |
| 2942 | |
| 2943 | if (!(sd->flags & SD_LOAD_BALANCE)) |
| 2944 | continue; |
| 2945 | |
| 2946 | if (sd->flags & SD_BALANCE_NEWIDLE) |
| 2947 | /* If we've pulled tasks over stop searching: */ |
| 2948 | pulled_task = load_balance_newidle(this_cpu, |
| 2949 | this_rq, sd); |
| 2950 | |
| 2951 | interval = msecs_to_jiffies(sd->balance_interval); |
| 2952 | if (time_after(next_balance, sd->last_balance + interval)) |
| 2953 | next_balance = sd->last_balance + interval; |
| 2954 | if (pulled_task) |
| 2955 | break; |
| 2956 | } |
| 2957 | if (pulled_task || time_after(jiffies, this_rq->next_balance)) { |
| 2958 | /* |
| 2959 | * We are going idle. next_balance may be set based on |
| 2960 | * a busy processor. So reset next_balance. |
| 2961 | */ |
| 2962 | this_rq->next_balance = next_balance; |
| 2963 | } |
| 2964 | } |
| 2965 | |
| 2966 | /* |
| 2967 | * active_load_balance is run by migration threads. It pushes running tasks |
| 2968 | * off the busiest CPU onto idle CPUs. It requires at least 1 task to be |
| 2969 | * running on each physical CPU where possible, and avoids physical / |
| 2970 | * logical imbalances. |
| 2971 | * |
| 2972 | * Called with busiest_rq locked. |
| 2973 | */ |
| 2974 | static void active_load_balance(struct rq *busiest_rq, int busiest_cpu) |
| 2975 | { |
| 2976 | int target_cpu = busiest_rq->push_cpu; |
| 2977 | struct sched_domain *sd; |
| 2978 | struct rq *target_rq; |
| 2979 | |
| 2980 | /* Is there any task to move? */ |
| 2981 | if (busiest_rq->nr_running <= 1) |
| 2982 | return; |
| 2983 | |
| 2984 | target_rq = cpu_rq(target_cpu); |
| 2985 | |
| 2986 | /* |
| 2987 | * This condition is "impossible", if it occurs |
| 2988 | * we need to fix it. Originally reported by |
| 2989 | * Bjorn Helgaas on a 128-cpu setup. |
| 2990 | */ |
| 2991 | BUG_ON(busiest_rq == target_rq); |
| 2992 | |
| 2993 | /* move a task from busiest_rq to target_rq */ |
| 2994 | double_lock_balance(busiest_rq, target_rq); |
| 2995 | update_rq_clock(busiest_rq); |
| 2996 | update_rq_clock(target_rq); |
| 2997 | |
| 2998 | /* Search for an sd spanning us and the target CPU. */ |
| 2999 | for_each_domain(target_cpu, sd) { |
| 3000 | if ((sd->flags & SD_LOAD_BALANCE) && |
| 3001 | cpu_isset(busiest_cpu, sd->span)) |
| 3002 | break; |
| 3003 | } |
| 3004 | |
| 3005 | if (likely(sd)) { |
| 3006 | schedstat_inc(sd, alb_count); |
| 3007 | |
| 3008 | if (move_one_task(target_rq, target_cpu, busiest_rq, |
| 3009 | sd, CPU_IDLE)) |
| 3010 | schedstat_inc(sd, alb_pushed); |
| 3011 | else |
| 3012 | schedstat_inc(sd, alb_failed); |
| 3013 | } |
| 3014 | spin_unlock(&target_rq->lock); |
| 3015 | } |
| 3016 | |
| 3017 | #ifdef CONFIG_NO_HZ |
| 3018 | static struct { |
| 3019 | atomic_t load_balancer; |
| 3020 | cpumask_t cpu_mask; |
| 3021 | } nohz ____cacheline_aligned = { |
| 3022 | .load_balancer = ATOMIC_INIT(-1), |
| 3023 | .cpu_mask = CPU_MASK_NONE, |
| 3024 | }; |
| 3025 | |
| 3026 | /* |
| 3027 | * This routine will try to nominate the ilb (idle load balancing) |
| 3028 | * owner among the cpus whose ticks are stopped. ilb owner will do the idle |
| 3029 | * load balancing on behalf of all those cpus. If all the cpus in the system |
| 3030 | * go into this tickless mode, then there will be no ilb owner (as there is |
| 3031 | * no need for one) and all the cpus will sleep till the next wakeup event |
| 3032 | * arrives... |
| 3033 | * |
| 3034 | * For the ilb owner, tick is not stopped. And this tick will be used |
| 3035 | * for idle load balancing. ilb owner will still be part of |
| 3036 | * nohz.cpu_mask.. |
| 3037 | * |
| 3038 | * While stopping the tick, this cpu will become the ilb owner if there |
| 3039 | * is no other owner. And will be the owner till that cpu becomes busy |
| 3040 | * or if all cpus in the system stop their ticks at which point |
| 3041 | * there is no need for ilb owner. |
| 3042 | * |
| 3043 | * When the ilb owner becomes busy, it nominates another owner, during the |
| 3044 | * next busy scheduler_tick() |
| 3045 | */ |
| 3046 | int select_nohz_load_balancer(int stop_tick) |
| 3047 | { |
| 3048 | int cpu = smp_processor_id(); |
| 3049 | |
| 3050 | if (stop_tick) { |
| 3051 | cpu_set(cpu, nohz.cpu_mask); |
| 3052 | cpu_rq(cpu)->in_nohz_recently = 1; |
| 3053 | |
| 3054 | /* |
| 3055 | * If we are going offline and still the leader, give up! |
| 3056 | */ |
| 3057 | if (cpu_is_offline(cpu) && |
| 3058 | atomic_read(&nohz.load_balancer) == cpu) { |
| 3059 | if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu) |
| 3060 | BUG(); |
| 3061 | return 0; |
| 3062 | } |
| 3063 | |
| 3064 | /* time for ilb owner also to sleep */ |
| 3065 | if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) { |
| 3066 | if (atomic_read(&nohz.load_balancer) == cpu) |
| 3067 | atomic_set(&nohz.load_balancer, -1); |
| 3068 | return 0; |
| 3069 | } |
| 3070 | |
| 3071 | if (atomic_read(&nohz.load_balancer) == -1) { |
| 3072 | /* make me the ilb owner */ |
| 3073 | if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1) |
| 3074 | return 1; |
| 3075 | } else if (atomic_read(&nohz.load_balancer) == cpu) |
| 3076 | return 1; |
| 3077 | } else { |
| 3078 | if (!cpu_isset(cpu, nohz.cpu_mask)) |
| 3079 | return 0; |
| 3080 | |
| 3081 | cpu_clear(cpu, nohz.cpu_mask); |
| 3082 | |
| 3083 | if (atomic_read(&nohz.load_balancer) == cpu) |
| 3084 | if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu) |
| 3085 | BUG(); |
| 3086 | } |
| 3087 | return 0; |
| 3088 | } |
| 3089 | #endif |
| 3090 | |
| 3091 | static DEFINE_SPINLOCK(balancing); |
| 3092 | |
| 3093 | /* |
| 3094 | * It checks each scheduling domain to see if it is due to be balanced, |
| 3095 | * and initiates a balancing operation if so. |
| 3096 | * |
| 3097 | * Balancing parameters are set up in arch_init_sched_domains. |
| 3098 | */ |
| 3099 | static void rebalance_domains(int cpu, enum cpu_idle_type idle) |
| 3100 | { |
| 3101 | int balance = 1; |
| 3102 | struct rq *rq = cpu_rq(cpu); |
| 3103 | unsigned long interval; |
| 3104 | struct sched_domain *sd; |
| 3105 | /* Earliest time when we have to do rebalance again */ |
| 3106 | unsigned long next_balance = jiffies + 60*HZ; |
| 3107 | int update_next_balance = 0; |
| 3108 | |
| 3109 | for_each_domain(cpu, sd) { |
| 3110 | if (!(sd->flags & SD_LOAD_BALANCE)) |
| 3111 | continue; |
| 3112 | |
| 3113 | interval = sd->balance_interval; |
| 3114 | if (idle != CPU_IDLE) |
| 3115 | interval *= sd->busy_factor; |
| 3116 | |
| 3117 | /* scale ms to jiffies */ |
| 3118 | interval = msecs_to_jiffies(interval); |
| 3119 | if (unlikely(!interval)) |
| 3120 | interval = 1; |
| 3121 | if (interval > HZ*NR_CPUS/10) |
| 3122 | interval = HZ*NR_CPUS/10; |
| 3123 | |
| 3124 | |
| 3125 | if (sd->flags & SD_SERIALIZE) { |
| 3126 | if (!spin_trylock(&balancing)) |
| 3127 | goto out; |
| 3128 | } |
| 3129 | |
| 3130 | if (time_after_eq(jiffies, sd->last_balance + interval)) { |
| 3131 | if (load_balance(cpu, rq, sd, idle, &balance)) { |
| 3132 | /* |
| 3133 | * We've pulled tasks over so either we're no |
| 3134 | * longer idle, or one of our SMT siblings is |
| 3135 | * not idle. |
| 3136 | */ |
| 3137 | idle = CPU_NOT_IDLE; |
| 3138 | } |
| 3139 | sd->last_balance = jiffies; |
| 3140 | } |
| 3141 | if (sd->flags & SD_SERIALIZE) |
| 3142 | spin_unlock(&balancing); |
| 3143 | out: |
| 3144 | if (time_after(next_balance, sd->last_balance + interval)) { |
| 3145 | next_balance = sd->last_balance + interval; |
| 3146 | update_next_balance = 1; |
| 3147 | } |
| 3148 | |
| 3149 | /* |
| 3150 | * Stop the load balance at this level. There is another |
| 3151 | * CPU in our sched group which is doing load balancing more |
| 3152 | * actively. |
| 3153 | */ |
| 3154 | if (!balance) |
| 3155 | break; |
| 3156 | } |
| 3157 | |
| 3158 | /* |
| 3159 | * next_balance will be updated only when there is a need. |
| 3160 | * When the cpu is attached to null domain for ex, it will not be |
| 3161 | * updated. |
| 3162 | */ |
| 3163 | if (likely(update_next_balance)) |
| 3164 | rq->next_balance = next_balance; |
| 3165 | } |
| 3166 | |
| 3167 | /* |
| 3168 | * run_rebalance_domains is triggered when needed from the scheduler tick. |
| 3169 | * In CONFIG_NO_HZ case, the idle load balance owner will do the |
| 3170 | * rebalancing for all the cpus for whom scheduler ticks are stopped. |
| 3171 | */ |
| 3172 | static void run_rebalance_domains(struct softirq_action *h) |
| 3173 | { |
| 3174 | int this_cpu = smp_processor_id(); |
| 3175 | struct rq *this_rq = cpu_rq(this_cpu); |
| 3176 | enum cpu_idle_type idle = this_rq->idle_at_tick ? |
| 3177 | CPU_IDLE : CPU_NOT_IDLE; |
| 3178 | |
| 3179 | rebalance_domains(this_cpu, idle); |
| 3180 | |
| 3181 | #ifdef CONFIG_NO_HZ |
| 3182 | /* |
| 3183 | * If this cpu is the owner for idle load balancing, then do the |
| 3184 | * balancing on behalf of the other idle cpus whose ticks are |
| 3185 | * stopped. |
| 3186 | */ |
| 3187 | if (this_rq->idle_at_tick && |
| 3188 | atomic_read(&nohz.load_balancer) == this_cpu) { |
| 3189 | cpumask_t cpus = nohz.cpu_mask; |
| 3190 | struct rq *rq; |
| 3191 | int balance_cpu; |
| 3192 | |
| 3193 | cpu_clear(this_cpu, cpus); |
| 3194 | for_each_cpu_mask(balance_cpu, cpus) { |
| 3195 | /* |
| 3196 | * If this cpu gets work to do, stop the load balancing |
| 3197 | * work being done for other cpus. Next load |
| 3198 | * balancing owner will pick it up. |
| 3199 | */ |
| 3200 | if (need_resched()) |
| 3201 | break; |
| 3202 | |
| 3203 | rebalance_domains(balance_cpu, CPU_IDLE); |
| 3204 | |
| 3205 | rq = cpu_rq(balance_cpu); |
| 3206 | if (time_after(this_rq->next_balance, rq->next_balance)) |
| 3207 | this_rq->next_balance = rq->next_balance; |
| 3208 | } |
| 3209 | } |
| 3210 | #endif |
| 3211 | } |
| 3212 | |
| 3213 | /* |
| 3214 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. |
| 3215 | * |
| 3216 | * In case of CONFIG_NO_HZ, this is the place where we nominate a new |
| 3217 | * idle load balancing owner or decide to stop the periodic load balancing, |
| 3218 | * if the whole system is idle. |
| 3219 | */ |
| 3220 | static inline void trigger_load_balance(struct rq *rq, int cpu) |
| 3221 | { |
| 3222 | #ifdef CONFIG_NO_HZ |
| 3223 | /* |
| 3224 | * If we were in the nohz mode recently and busy at the current |
| 3225 | * scheduler tick, then check if we need to nominate new idle |
| 3226 | * load balancer. |
| 3227 | */ |
| 3228 | if (rq->in_nohz_recently && !rq->idle_at_tick) { |
| 3229 | rq->in_nohz_recently = 0; |
| 3230 | |
| 3231 | if (atomic_read(&nohz.load_balancer) == cpu) { |
| 3232 | cpu_clear(cpu, nohz.cpu_mask); |
| 3233 | atomic_set(&nohz.load_balancer, -1); |
| 3234 | } |
| 3235 | |
| 3236 | if (atomic_read(&nohz.load_balancer) == -1) { |
| 3237 | /* |
| 3238 | * simple selection for now: Nominate the |
| 3239 | * first cpu in the nohz list to be the next |
| 3240 | * ilb owner. |
| 3241 | * |
| 3242 | * TBD: Traverse the sched domains and nominate |
| 3243 | * the nearest cpu in the nohz.cpu_mask. |
| 3244 | */ |
| 3245 | int ilb = first_cpu(nohz.cpu_mask); |
| 3246 | |
| 3247 | if (ilb != NR_CPUS) |
| 3248 | resched_cpu(ilb); |
| 3249 | } |
| 3250 | } |
| 3251 | |
| 3252 | /* |
| 3253 | * If this cpu is idle and doing idle load balancing for all the |
| 3254 | * cpus with ticks stopped, is it time for that to stop? |
| 3255 | */ |
| 3256 | if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu && |
| 3257 | cpus_weight(nohz.cpu_mask) == num_online_cpus()) { |
| 3258 | resched_cpu(cpu); |
| 3259 | return; |
| 3260 | } |
| 3261 | |
| 3262 | /* |
| 3263 | * If this cpu is idle and the idle load balancing is done by |
| 3264 | * someone else, then no need raise the SCHED_SOFTIRQ |
| 3265 | */ |
| 3266 | if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu && |
| 3267 | cpu_isset(cpu, nohz.cpu_mask)) |
| 3268 | return; |
| 3269 | #endif |
| 3270 | if (time_after_eq(jiffies, rq->next_balance)) |
| 3271 | raise_softirq(SCHED_SOFTIRQ); |
| 3272 | } |
| 3273 | |
| 3274 | #else /* CONFIG_SMP */ |
| 3275 | |
| 3276 | /* |
| 3277 | * on UP we do not need to balance between CPUs: |
| 3278 | */ |
| 3279 | static inline void idle_balance(int cpu, struct rq *rq) |
| 3280 | { |
| 3281 | } |
| 3282 | |
| 3283 | #endif |
| 3284 | |
| 3285 | DEFINE_PER_CPU(struct kernel_stat, kstat); |
| 3286 | |
| 3287 | EXPORT_PER_CPU_SYMBOL(kstat); |
| 3288 | |
| 3289 | /* |
| 3290 | * Return p->sum_exec_runtime plus any more ns on the sched_clock |
| 3291 | * that have not yet been banked in case the task is currently running. |
| 3292 | */ |
| 3293 | unsigned long long task_sched_runtime(struct task_struct *p) |
| 3294 | { |
| 3295 | unsigned long flags; |
| 3296 | u64 ns, delta_exec; |
| 3297 | struct rq *rq; |
| 3298 | |
| 3299 | rq = task_rq_lock(p, &flags); |
| 3300 | ns = p->se.sum_exec_runtime; |
| 3301 | if (task_current(rq, p)) { |
| 3302 | update_rq_clock(rq); |
| 3303 | delta_exec = rq->clock - p->se.exec_start; |
| 3304 | if ((s64)delta_exec > 0) |
| 3305 | ns += delta_exec; |
| 3306 | } |
| 3307 | task_rq_unlock(rq, &flags); |
| 3308 | |
| 3309 | return ns; |
| 3310 | } |
| 3311 | |
| 3312 | /* |
| 3313 | * Account user cpu time to a process. |
| 3314 | * @p: the process that the cpu time gets accounted to |
| 3315 | * @cputime: the cpu time spent in user space since the last update |
| 3316 | */ |
| 3317 | void account_user_time(struct task_struct *p, cputime_t cputime) |
| 3318 | { |
| 3319 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; |
| 3320 | cputime64_t tmp; |
| 3321 | |
| 3322 | p->utime = cputime_add(p->utime, cputime); |
| 3323 | |
| 3324 | /* Add user time to cpustat. */ |
| 3325 | tmp = cputime_to_cputime64(cputime); |
| 3326 | if (TASK_NICE(p) > 0) |
| 3327 | cpustat->nice = cputime64_add(cpustat->nice, tmp); |
| 3328 | else |
| 3329 | cpustat->user = cputime64_add(cpustat->user, tmp); |
| 3330 | } |
| 3331 | |
| 3332 | /* |
| 3333 | * Account guest cpu time to a process. |
| 3334 | * @p: the process that the cpu time gets accounted to |
| 3335 | * @cputime: the cpu time spent in virtual machine since the last update |
| 3336 | */ |
| 3337 | static void account_guest_time(struct task_struct *p, cputime_t cputime) |
| 3338 | { |
| 3339 | cputime64_t tmp; |
| 3340 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; |
| 3341 | |
| 3342 | tmp = cputime_to_cputime64(cputime); |
| 3343 | |
| 3344 | p->utime = cputime_add(p->utime, cputime); |
| 3345 | p->gtime = cputime_add(p->gtime, cputime); |
| 3346 | |
| 3347 | cpustat->user = cputime64_add(cpustat->user, tmp); |
| 3348 | cpustat->guest = cputime64_add(cpustat->guest, tmp); |
| 3349 | } |
| 3350 | |
| 3351 | /* |
| 3352 | * Account scaled user cpu time to a process. |
| 3353 | * @p: the process that the cpu time gets accounted to |
| 3354 | * @cputime: the cpu time spent in user space since the last update |
| 3355 | */ |
| 3356 | void account_user_time_scaled(struct task_struct *p, cputime_t cputime) |
| 3357 | { |
| 3358 | p->utimescaled = cputime_add(p->utimescaled, cputime); |
| 3359 | } |
| 3360 | |
| 3361 | /* |
| 3362 | * Account system cpu time to a process. |
| 3363 | * @p: the process that the cpu time gets accounted to |
| 3364 | * @hardirq_offset: the offset to subtract from hardirq_count() |
| 3365 | * @cputime: the cpu time spent in kernel space since the last update |
| 3366 | */ |
| 3367 | void account_system_time(struct task_struct *p, int hardirq_offset, |
| 3368 | cputime_t cputime) |
| 3369 | { |
| 3370 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; |
| 3371 | struct rq *rq = this_rq(); |
| 3372 | cputime64_t tmp; |
| 3373 | |
| 3374 | if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) |
| 3375 | return account_guest_time(p, cputime); |
| 3376 | |
| 3377 | p->stime = cputime_add(p->stime, cputime); |
| 3378 | |
| 3379 | /* Add system time to cpustat. */ |
| 3380 | tmp = cputime_to_cputime64(cputime); |
| 3381 | if (hardirq_count() - hardirq_offset) |
| 3382 | cpustat->irq = cputime64_add(cpustat->irq, tmp); |
| 3383 | else if (softirq_count()) |
| 3384 | cpustat->softirq = cputime64_add(cpustat->softirq, tmp); |
| 3385 | else if (p != rq->idle) |
| 3386 | cpustat->system = cputime64_add(cpustat->system, tmp); |
| 3387 | else if (atomic_read(&rq->nr_iowait) > 0) |
| 3388 | cpustat->iowait = cputime64_add(cpustat->iowait, tmp); |
| 3389 | else |
| 3390 | cpustat->idle = cputime64_add(cpustat->idle, tmp); |
| 3391 | /* Account for system time used */ |
| 3392 | acct_update_integrals(p); |
| 3393 | } |
| 3394 | |
| 3395 | /* |
| 3396 | * Account scaled system cpu time to a process. |
| 3397 | * @p: the process that the cpu time gets accounted to |
| 3398 | * @hardirq_offset: the offset to subtract from hardirq_count() |
| 3399 | * @cputime: the cpu time spent in kernel space since the last update |
| 3400 | */ |
| 3401 | void account_system_time_scaled(struct task_struct *p, cputime_t cputime) |
| 3402 | { |
| 3403 | p->stimescaled = cputime_add(p->stimescaled, cputime); |
| 3404 | } |
| 3405 | |
| 3406 | /* |
| 3407 | * Account for involuntary wait time. |
| 3408 | * @p: the process from which the cpu time has been stolen |
| 3409 | * @steal: the cpu time spent in involuntary wait |
| 3410 | */ |
| 3411 | void account_steal_time(struct task_struct *p, cputime_t steal) |
| 3412 | { |
| 3413 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; |
| 3414 | cputime64_t tmp = cputime_to_cputime64(steal); |
| 3415 | struct rq *rq = this_rq(); |
| 3416 | |
| 3417 | if (p == rq->idle) { |
| 3418 | p->stime = cputime_add(p->stime, steal); |
| 3419 | if (atomic_read(&rq->nr_iowait) > 0) |
| 3420 | cpustat->iowait = cputime64_add(cpustat->iowait, tmp); |
| 3421 | else |
| 3422 | cpustat->idle = cputime64_add(cpustat->idle, tmp); |
| 3423 | } else |
| 3424 | cpustat->steal = cputime64_add(cpustat->steal, tmp); |
| 3425 | } |
| 3426 | |
| 3427 | /* |
| 3428 | * This function gets called by the timer code, with HZ frequency. |
| 3429 | * We call it with interrupts disabled. |
| 3430 | * |
| 3431 | * It also gets called by the fork code, when changing the parent's |
| 3432 | * timeslices. |
| 3433 | */ |
| 3434 | void scheduler_tick(void) |
| 3435 | { |
| 3436 | int cpu = smp_processor_id(); |
| 3437 | struct rq *rq = cpu_rq(cpu); |
| 3438 | struct task_struct *curr = rq->curr; |
| 3439 | u64 next_tick = rq->tick_timestamp + TICK_NSEC; |
| 3440 | |
| 3441 | spin_lock(&rq->lock); |
| 3442 | __update_rq_clock(rq); |
| 3443 | /* |
| 3444 | * Let rq->clock advance by at least TICK_NSEC: |
| 3445 | */ |
| 3446 | if (unlikely(rq->clock < next_tick)) |
| 3447 | rq->clock = next_tick; |
| 3448 | rq->tick_timestamp = rq->clock; |
| 3449 | update_cpu_load(rq); |
| 3450 | if (curr != rq->idle) /* FIXME: needed? */ |
| 3451 | curr->sched_class->task_tick(rq, curr); |
| 3452 | spin_unlock(&rq->lock); |
| 3453 | |
| 3454 | #ifdef CONFIG_SMP |
| 3455 | rq->idle_at_tick = idle_cpu(cpu); |
| 3456 | trigger_load_balance(rq, cpu); |
| 3457 | #endif |
| 3458 | } |
| 3459 | |
| 3460 | #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT) |
| 3461 | |
| 3462 | void fastcall add_preempt_count(int val) |
| 3463 | { |
| 3464 | /* |
| 3465 | * Underflow? |
| 3466 | */ |
| 3467 | if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) |
| 3468 | return; |
| 3469 | preempt_count() += val; |
| 3470 | /* |
| 3471 | * Spinlock count overflowing soon? |
| 3472 | */ |
| 3473 | DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= |
| 3474 | PREEMPT_MASK - 10); |
| 3475 | } |
| 3476 | EXPORT_SYMBOL(add_preempt_count); |
| 3477 | |
| 3478 | void fastcall sub_preempt_count(int val) |
| 3479 | { |
| 3480 | /* |
| 3481 | * Underflow? |
| 3482 | */ |
| 3483 | if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) |
| 3484 | return; |
| 3485 | /* |
| 3486 | * Is the spinlock portion underflowing? |
| 3487 | */ |
| 3488 | if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && |
| 3489 | !(preempt_count() & PREEMPT_MASK))) |
| 3490 | return; |
| 3491 | |
| 3492 | preempt_count() -= val; |
| 3493 | } |
| 3494 | EXPORT_SYMBOL(sub_preempt_count); |
| 3495 | |
| 3496 | #endif |
| 3497 | |
| 3498 | /* |
| 3499 | * Print scheduling while atomic bug: |
| 3500 | */ |
| 3501 | static noinline void __schedule_bug(struct task_struct *prev) |
| 3502 | { |
| 3503 | struct pt_regs *regs = get_irq_regs(); |
| 3504 | |
| 3505 | printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", |
| 3506 | prev->comm, prev->pid, preempt_count()); |
| 3507 | |
| 3508 | debug_show_held_locks(prev); |
| 3509 | if (irqs_disabled()) |
| 3510 | print_irqtrace_events(prev); |
| 3511 | |
| 3512 | if (regs) |
| 3513 | show_regs(regs); |
| 3514 | else |
| 3515 | dump_stack(); |
| 3516 | } |
| 3517 | |
| 3518 | /* |
| 3519 | * Various schedule()-time debugging checks and statistics: |
| 3520 | */ |
| 3521 | static inline void schedule_debug(struct task_struct *prev) |
| 3522 | { |
| 3523 | /* |
| 3524 | * Test if we are atomic. Since do_exit() needs to call into |
| 3525 | * schedule() atomically, we ignore that path for now. |
| 3526 | * Otherwise, whine if we are scheduling when we should not be. |
| 3527 | */ |
| 3528 | if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state)) |
| 3529 | __schedule_bug(prev); |
| 3530 | |
| 3531 | profile_hit(SCHED_PROFILING, __builtin_return_address(0)); |
| 3532 | |
| 3533 | schedstat_inc(this_rq(), sched_count); |
| 3534 | #ifdef CONFIG_SCHEDSTATS |
| 3535 | if (unlikely(prev->lock_depth >= 0)) { |
| 3536 | schedstat_inc(this_rq(), bkl_count); |
| 3537 | schedstat_inc(prev, sched_info.bkl_count); |
| 3538 | } |
| 3539 | #endif |
| 3540 | } |
| 3541 | |
| 3542 | /* |
| 3543 | * Pick up the highest-prio task: |
| 3544 | */ |
| 3545 | static inline struct task_struct * |
| 3546 | pick_next_task(struct rq *rq, struct task_struct *prev) |
| 3547 | { |
| 3548 | const struct sched_class *class; |
| 3549 | struct task_struct *p; |
| 3550 | |
| 3551 | /* |
| 3552 | * Optimization: we know that if all tasks are in |
| 3553 | * the fair class we can call that function directly: |
| 3554 | */ |
| 3555 | if (likely(rq->nr_running == rq->cfs.nr_running)) { |
| 3556 | p = fair_sched_class.pick_next_task(rq); |
| 3557 | if (likely(p)) |
| 3558 | return p; |
| 3559 | } |
| 3560 | |
| 3561 | class = sched_class_highest; |
| 3562 | for ( ; ; ) { |
| 3563 | p = class->pick_next_task(rq); |
| 3564 | if (p) |
| 3565 | return p; |
| 3566 | /* |
| 3567 | * Will never be NULL as the idle class always |
| 3568 | * returns a non-NULL p: |
| 3569 | */ |
| 3570 | class = class->next; |
| 3571 | } |
| 3572 | } |
| 3573 | |
| 3574 | /* |
| 3575 | * schedule() is the main scheduler function. |
| 3576 | */ |
| 3577 | asmlinkage void __sched schedule(void) |
| 3578 | { |
| 3579 | struct task_struct *prev, *next; |
| 3580 | long *switch_count; |
| 3581 | struct rq *rq; |
| 3582 | int cpu; |
| 3583 | |
| 3584 | need_resched: |
| 3585 | preempt_disable(); |
| 3586 | cpu = smp_processor_id(); |
| 3587 | rq = cpu_rq(cpu); |
| 3588 | rcu_qsctr_inc(cpu); |
| 3589 | prev = rq->curr; |
| 3590 | switch_count = &prev->nivcsw; |
| 3591 | |
| 3592 | release_kernel_lock(prev); |
| 3593 | need_resched_nonpreemptible: |
| 3594 | |
| 3595 | schedule_debug(prev); |
| 3596 | |
| 3597 | /* |
| 3598 | * Do the rq-clock update outside the rq lock: |
| 3599 | */ |
| 3600 | local_irq_disable(); |
| 3601 | __update_rq_clock(rq); |
| 3602 | spin_lock(&rq->lock); |
| 3603 | clear_tsk_need_resched(prev); |
| 3604 | |
| 3605 | if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { |
| 3606 | if (unlikely((prev->state & TASK_INTERRUPTIBLE) && |
| 3607 | unlikely(signal_pending(prev)))) { |
| 3608 | prev->state = TASK_RUNNING; |
| 3609 | } else { |
| 3610 | deactivate_task(rq, prev, 1); |
| 3611 | } |
| 3612 | switch_count = &prev->nvcsw; |
| 3613 | } |
| 3614 | |
| 3615 | schedule_balance_rt(rq, prev); |
| 3616 | |
| 3617 | if (unlikely(!rq->nr_running)) |
| 3618 | idle_balance(cpu, rq); |
| 3619 | |
| 3620 | prev->sched_class->put_prev_task(rq, prev); |
| 3621 | next = pick_next_task(rq, prev); |
| 3622 | |
| 3623 | sched_info_switch(prev, next); |
| 3624 | |
| 3625 | if (likely(prev != next)) { |
| 3626 | rq->nr_switches++; |
| 3627 | rq->curr = next; |
| 3628 | ++*switch_count; |
| 3629 | |
| 3630 | context_switch(rq, prev, next); /* unlocks the rq */ |
| 3631 | } else |
| 3632 | spin_unlock_irq(&rq->lock); |
| 3633 | |
| 3634 | if (unlikely(reacquire_kernel_lock(current) < 0)) { |
| 3635 | cpu = smp_processor_id(); |
| 3636 | rq = cpu_rq(cpu); |
| 3637 | goto need_resched_nonpreemptible; |
| 3638 | } |
| 3639 | preempt_enable_no_resched(); |
| 3640 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) |
| 3641 | goto need_resched; |
| 3642 | } |
| 3643 | EXPORT_SYMBOL(schedule); |
| 3644 | |
| 3645 | #ifdef CONFIG_PREEMPT |
| 3646 | /* |
| 3647 | * this is the entry point to schedule() from in-kernel preemption |
| 3648 | * off of preempt_enable. Kernel preemptions off return from interrupt |
| 3649 | * occur there and call schedule directly. |
| 3650 | */ |
| 3651 | asmlinkage void __sched preempt_schedule(void) |
| 3652 | { |
| 3653 | struct thread_info *ti = current_thread_info(); |
| 3654 | #ifdef CONFIG_PREEMPT_BKL |
| 3655 | struct task_struct *task = current; |
| 3656 | int saved_lock_depth; |
| 3657 | #endif |
| 3658 | /* |
| 3659 | * If there is a non-zero preempt_count or interrupts are disabled, |
| 3660 | * we do not want to preempt the current task. Just return.. |
| 3661 | */ |
| 3662 | if (likely(ti->preempt_count || irqs_disabled())) |
| 3663 | return; |
| 3664 | |
| 3665 | do { |
| 3666 | add_preempt_count(PREEMPT_ACTIVE); |
| 3667 | |
| 3668 | /* |
| 3669 | * We keep the big kernel semaphore locked, but we |
| 3670 | * clear ->lock_depth so that schedule() doesnt |
| 3671 | * auto-release the semaphore: |
| 3672 | */ |
| 3673 | #ifdef CONFIG_PREEMPT_BKL |
| 3674 | saved_lock_depth = task->lock_depth; |
| 3675 | task->lock_depth = -1; |
| 3676 | #endif |
| 3677 | schedule(); |
| 3678 | #ifdef CONFIG_PREEMPT_BKL |
| 3679 | task->lock_depth = saved_lock_depth; |
| 3680 | #endif |
| 3681 | sub_preempt_count(PREEMPT_ACTIVE); |
| 3682 | |
| 3683 | /* |
| 3684 | * Check again in case we missed a preemption opportunity |
| 3685 | * between schedule and now. |
| 3686 | */ |
| 3687 | barrier(); |
| 3688 | } while (unlikely(test_thread_flag(TIF_NEED_RESCHED))); |
| 3689 | } |
| 3690 | EXPORT_SYMBOL(preempt_schedule); |
| 3691 | |
| 3692 | /* |
| 3693 | * this is the entry point to schedule() from kernel preemption |
| 3694 | * off of irq context. |
| 3695 | * Note, that this is called and return with irqs disabled. This will |
| 3696 | * protect us against recursive calling from irq. |
| 3697 | */ |
| 3698 | asmlinkage void __sched preempt_schedule_irq(void) |
| 3699 | { |
| 3700 | struct thread_info *ti = current_thread_info(); |
| 3701 | #ifdef CONFIG_PREEMPT_BKL |
| 3702 | struct task_struct *task = current; |
| 3703 | int saved_lock_depth; |
| 3704 | #endif |
| 3705 | /* Catch callers which need to be fixed */ |
| 3706 | BUG_ON(ti->preempt_count || !irqs_disabled()); |
| 3707 | |
| 3708 | do { |
| 3709 | add_preempt_count(PREEMPT_ACTIVE); |
| 3710 | |
| 3711 | /* |
| 3712 | * We keep the big kernel semaphore locked, but we |
| 3713 | * clear ->lock_depth so that schedule() doesnt |
| 3714 | * auto-release the semaphore: |
| 3715 | */ |
| 3716 | #ifdef CONFIG_PREEMPT_BKL |
| 3717 | saved_lock_depth = task->lock_depth; |
| 3718 | task->lock_depth = -1; |
| 3719 | #endif |
| 3720 | local_irq_enable(); |
| 3721 | schedule(); |
| 3722 | local_irq_disable(); |
| 3723 | #ifdef CONFIG_PREEMPT_BKL |
| 3724 | task->lock_depth = saved_lock_depth; |
| 3725 | #endif |
| 3726 | sub_preempt_count(PREEMPT_ACTIVE); |
| 3727 | |
| 3728 | /* |
| 3729 | * Check again in case we missed a preemption opportunity |
| 3730 | * between schedule and now. |
| 3731 | */ |
| 3732 | barrier(); |
| 3733 | } while (unlikely(test_thread_flag(TIF_NEED_RESCHED))); |
| 3734 | } |
| 3735 | |
| 3736 | #endif /* CONFIG_PREEMPT */ |
| 3737 | |
| 3738 | int default_wake_function(wait_queue_t *curr, unsigned mode, int sync, |
| 3739 | void *key) |
| 3740 | { |
| 3741 | return try_to_wake_up(curr->private, mode, sync); |
| 3742 | } |
| 3743 | EXPORT_SYMBOL(default_wake_function); |
| 3744 | |
| 3745 | /* |
| 3746 | * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just |
| 3747 | * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve |
| 3748 | * number) then we wake all the non-exclusive tasks and one exclusive task. |
| 3749 | * |
| 3750 | * There are circumstances in which we can try to wake a task which has already |
| 3751 | * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns |
| 3752 | * zero in this (rare) case, and we handle it by continuing to scan the queue. |
| 3753 | */ |
| 3754 | static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, |
| 3755 | int nr_exclusive, int sync, void *key) |
| 3756 | { |
| 3757 | wait_queue_t *curr, *next; |
| 3758 | |
| 3759 | list_for_each_entry_safe(curr, next, &q->task_list, task_list) { |
| 3760 | unsigned flags = curr->flags; |
| 3761 | |
| 3762 | if (curr->func(curr, mode, sync, key) && |
| 3763 | (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive) |
| 3764 | break; |
| 3765 | } |
| 3766 | } |
| 3767 | |
| 3768 | /** |
| 3769 | * __wake_up - wake up threads blocked on a waitqueue. |
| 3770 | * @q: the waitqueue |
| 3771 | * @mode: which threads |
| 3772 | * @nr_exclusive: how many wake-one or wake-many threads to wake up |
| 3773 | * @key: is directly passed to the wakeup function |
| 3774 | */ |
| 3775 | void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode, |
| 3776 | int nr_exclusive, void *key) |
| 3777 | { |
| 3778 | unsigned long flags; |
| 3779 | |
| 3780 | spin_lock_irqsave(&q->lock, flags); |
| 3781 | __wake_up_common(q, mode, nr_exclusive, 0, key); |
| 3782 | spin_unlock_irqrestore(&q->lock, flags); |
| 3783 | } |
| 3784 | EXPORT_SYMBOL(__wake_up); |
| 3785 | |
| 3786 | /* |
| 3787 | * Same as __wake_up but called with the spinlock in wait_queue_head_t held. |
| 3788 | */ |
| 3789 | void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode) |
| 3790 | { |
| 3791 | __wake_up_common(q, mode, 1, 0, NULL); |
| 3792 | } |
| 3793 | |
| 3794 | /** |
| 3795 | * __wake_up_sync - wake up threads blocked on a waitqueue. |
| 3796 | * @q: the waitqueue |
| 3797 | * @mode: which threads |
| 3798 | * @nr_exclusive: how many wake-one or wake-many threads to wake up |
| 3799 | * |
| 3800 | * The sync wakeup differs that the waker knows that it will schedule |
| 3801 | * away soon, so while the target thread will be woken up, it will not |
| 3802 | * be migrated to another CPU - ie. the two threads are 'synchronized' |
| 3803 | * with each other. This can prevent needless bouncing between CPUs. |
| 3804 | * |
| 3805 | * On UP it can prevent extra preemption. |
| 3806 | */ |
| 3807 | void fastcall |
| 3808 | __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) |
| 3809 | { |
| 3810 | unsigned long flags; |
| 3811 | int sync = 1; |
| 3812 | |
| 3813 | if (unlikely(!q)) |
| 3814 | return; |
| 3815 | |
| 3816 | if (unlikely(!nr_exclusive)) |
| 3817 | sync = 0; |
| 3818 | |
| 3819 | spin_lock_irqsave(&q->lock, flags); |
| 3820 | __wake_up_common(q, mode, nr_exclusive, sync, NULL); |
| 3821 | spin_unlock_irqrestore(&q->lock, flags); |
| 3822 | } |
| 3823 | EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ |
| 3824 | |
| 3825 | void complete(struct completion *x) |
| 3826 | { |
| 3827 | unsigned long flags; |
| 3828 | |
| 3829 | spin_lock_irqsave(&x->wait.lock, flags); |
| 3830 | x->done++; |
| 3831 | __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE, |
| 3832 | 1, 0, NULL); |
| 3833 | spin_unlock_irqrestore(&x->wait.lock, flags); |
| 3834 | } |
| 3835 | EXPORT_SYMBOL(complete); |
| 3836 | |
| 3837 | void complete_all(struct completion *x) |
| 3838 | { |
| 3839 | unsigned long flags; |
| 3840 | |
| 3841 | spin_lock_irqsave(&x->wait.lock, flags); |
| 3842 | x->done += UINT_MAX/2; |
| 3843 | __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE, |
| 3844 | 0, 0, NULL); |
| 3845 | spin_unlock_irqrestore(&x->wait.lock, flags); |
| 3846 | } |
| 3847 | EXPORT_SYMBOL(complete_all); |
| 3848 | |
| 3849 | static inline long __sched |
| 3850 | do_wait_for_common(struct completion *x, long timeout, int state) |
| 3851 | { |
| 3852 | if (!x->done) { |
| 3853 | DECLARE_WAITQUEUE(wait, current); |
| 3854 | |
| 3855 | wait.flags |= WQ_FLAG_EXCLUSIVE; |
| 3856 | __add_wait_queue_tail(&x->wait, &wait); |
| 3857 | do { |
| 3858 | if (state == TASK_INTERRUPTIBLE && |
| 3859 | signal_pending(current)) { |
| 3860 | __remove_wait_queue(&x->wait, &wait); |
| 3861 | return -ERESTARTSYS; |
| 3862 | } |
| 3863 | __set_current_state(state); |
| 3864 | spin_unlock_irq(&x->wait.lock); |
| 3865 | timeout = schedule_timeout(timeout); |
| 3866 | spin_lock_irq(&x->wait.lock); |
| 3867 | if (!timeout) { |
| 3868 | __remove_wait_queue(&x->wait, &wait); |
| 3869 | return timeout; |
| 3870 | } |
| 3871 | } while (!x->done); |
| 3872 | __remove_wait_queue(&x->wait, &wait); |
| 3873 | } |
| 3874 | x->done--; |
| 3875 | return timeout; |
| 3876 | } |
| 3877 | |
| 3878 | static long __sched |
| 3879 | wait_for_common(struct completion *x, long timeout, int state) |
| 3880 | { |
| 3881 | might_sleep(); |
| 3882 | |
| 3883 | spin_lock_irq(&x->wait.lock); |
| 3884 | timeout = do_wait_for_common(x, timeout, state); |
| 3885 | spin_unlock_irq(&x->wait.lock); |
| 3886 | return timeout; |
| 3887 | } |
| 3888 | |
| 3889 | void __sched wait_for_completion(struct completion *x) |
| 3890 | { |
| 3891 | wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); |
| 3892 | } |
| 3893 | EXPORT_SYMBOL(wait_for_completion); |
| 3894 | |
| 3895 | unsigned long __sched |
| 3896 | wait_for_completion_timeout(struct completion *x, unsigned long timeout) |
| 3897 | { |
| 3898 | return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE); |
| 3899 | } |
| 3900 | EXPORT_SYMBOL(wait_for_completion_timeout); |
| 3901 | |
| 3902 | int __sched wait_for_completion_interruptible(struct completion *x) |
| 3903 | { |
| 3904 | long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE); |
| 3905 | if (t == -ERESTARTSYS) |
| 3906 | return t; |
| 3907 | return 0; |
| 3908 | } |
| 3909 | EXPORT_SYMBOL(wait_for_completion_interruptible); |
| 3910 | |
| 3911 | unsigned long __sched |
| 3912 | wait_for_completion_interruptible_timeout(struct completion *x, |
| 3913 | unsigned long timeout) |
| 3914 | { |
| 3915 | return wait_for_common(x, timeout, TASK_INTERRUPTIBLE); |
| 3916 | } |
| 3917 | EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); |
| 3918 | |
| 3919 | static long __sched |
| 3920 | sleep_on_common(wait_queue_head_t *q, int state, long timeout) |
| 3921 | { |
| 3922 | unsigned long flags; |
| 3923 | wait_queue_t wait; |
| 3924 | |
| 3925 | init_waitqueue_entry(&wait, current); |
| 3926 | |
| 3927 | __set_current_state(state); |
| 3928 | |
| 3929 | spin_lock_irqsave(&q->lock, flags); |
| 3930 | __add_wait_queue(q, &wait); |
| 3931 | spin_unlock(&q->lock); |
| 3932 | timeout = schedule_timeout(timeout); |
| 3933 | spin_lock_irq(&q->lock); |
| 3934 | __remove_wait_queue(q, &wait); |
| 3935 | spin_unlock_irqrestore(&q->lock, flags); |
| 3936 | |
| 3937 | return timeout; |
| 3938 | } |
| 3939 | |
| 3940 | void __sched interruptible_sleep_on(wait_queue_head_t *q) |
| 3941 | { |
| 3942 | sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); |
| 3943 | } |
| 3944 | EXPORT_SYMBOL(interruptible_sleep_on); |
| 3945 | |
| 3946 | long __sched |
| 3947 | interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) |
| 3948 | { |
| 3949 | return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); |
| 3950 | } |
| 3951 | EXPORT_SYMBOL(interruptible_sleep_on_timeout); |
| 3952 | |
| 3953 | void __sched sleep_on(wait_queue_head_t *q) |
| 3954 | { |
| 3955 | sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); |
| 3956 | } |
| 3957 | EXPORT_SYMBOL(sleep_on); |
| 3958 | |
| 3959 | long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) |
| 3960 | { |
| 3961 | return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); |
| 3962 | } |
| 3963 | EXPORT_SYMBOL(sleep_on_timeout); |
| 3964 | |
| 3965 | #ifdef CONFIG_RT_MUTEXES |
| 3966 | |
| 3967 | /* |
| 3968 | * rt_mutex_setprio - set the current priority of a task |
| 3969 | * @p: task |
| 3970 | * @prio: prio value (kernel-internal form) |
| 3971 | * |
| 3972 | * This function changes the 'effective' priority of a task. It does |
| 3973 | * not touch ->normal_prio like __setscheduler(). |
| 3974 | * |
| 3975 | * Used by the rt_mutex code to implement priority inheritance logic. |
| 3976 | */ |
| 3977 | void rt_mutex_setprio(struct task_struct *p, int prio) |
| 3978 | { |
| 3979 | unsigned long flags; |
| 3980 | int oldprio, on_rq, running; |
| 3981 | struct rq *rq; |
| 3982 | |
| 3983 | BUG_ON(prio < 0 || prio > MAX_PRIO); |
| 3984 | |
| 3985 | rq = task_rq_lock(p, &flags); |
| 3986 | update_rq_clock(rq); |
| 3987 | |
| 3988 | oldprio = p->prio; |
| 3989 | on_rq = p->se.on_rq; |
| 3990 | running = task_current(rq, p); |
| 3991 | if (on_rq) { |
| 3992 | dequeue_task(rq, p, 0); |
| 3993 | if (running) |
| 3994 | p->sched_class->put_prev_task(rq, p); |
| 3995 | } |
| 3996 | |
| 3997 | if (rt_prio(prio)) |
| 3998 | p->sched_class = &rt_sched_class; |
| 3999 | else |
| 4000 | p->sched_class = &fair_sched_class; |
| 4001 | |
| 4002 | p->prio = prio; |
| 4003 | |
| 4004 | if (on_rq) { |
| 4005 | if (running) |
| 4006 | p->sched_class->set_curr_task(rq); |
| 4007 | enqueue_task(rq, p, 0); |
| 4008 | /* |
| 4009 | * Reschedule if we are currently running on this runqueue and |
| 4010 | * our priority decreased, or if we are not currently running on |
| 4011 | * this runqueue and our priority is higher than the current's |
| 4012 | */ |
| 4013 | if (running) { |
| 4014 | if (p->prio > oldprio) |
| 4015 | resched_task(rq->curr); |
| 4016 | } else { |
| 4017 | check_preempt_curr(rq, p); |
| 4018 | } |
| 4019 | } |
| 4020 | task_rq_unlock(rq, &flags); |
| 4021 | } |
| 4022 | |
| 4023 | #endif |
| 4024 | |
| 4025 | void set_user_nice(struct task_struct *p, long nice) |
| 4026 | { |
| 4027 | int old_prio, delta, on_rq; |
| 4028 | unsigned long flags; |
| 4029 | struct rq *rq; |
| 4030 | |
| 4031 | if (TASK_NICE(p) == nice || nice < -20 || nice > 19) |
| 4032 | return; |
| 4033 | /* |
| 4034 | * We have to be careful, if called from sys_setpriority(), |
| 4035 | * the task might be in the middle of scheduling on another CPU. |
| 4036 | */ |
| 4037 | rq = task_rq_lock(p, &flags); |
| 4038 | update_rq_clock(rq); |
| 4039 | /* |
| 4040 | * The RT priorities are set via sched_setscheduler(), but we still |
| 4041 | * allow the 'normal' nice value to be set - but as expected |
| 4042 | * it wont have any effect on scheduling until the task is |
| 4043 | * SCHED_FIFO/SCHED_RR: |
| 4044 | */ |
| 4045 | if (task_has_rt_policy(p)) { |
| 4046 | p->static_prio = NICE_TO_PRIO(nice); |
| 4047 | goto out_unlock; |
| 4048 | } |
| 4049 | on_rq = p->se.on_rq; |
| 4050 | if (on_rq) |
| 4051 | dequeue_task(rq, p, 0); |
| 4052 | |
| 4053 | p->static_prio = NICE_TO_PRIO(nice); |
| 4054 | set_load_weight(p); |
| 4055 | old_prio = p->prio; |
| 4056 | p->prio = effective_prio(p); |
| 4057 | delta = p->prio - old_prio; |
| 4058 | |
| 4059 | if (on_rq) { |
| 4060 | enqueue_task(rq, p, 0); |
| 4061 | /* |
| 4062 | * If the task increased its priority or is running and |
| 4063 | * lowered its priority, then reschedule its CPU: |
| 4064 | */ |
| 4065 | if (delta < 0 || (delta > 0 && task_running(rq, p))) |
| 4066 | resched_task(rq->curr); |
| 4067 | } |
| 4068 | out_unlock: |
| 4069 | task_rq_unlock(rq, &flags); |
| 4070 | } |
| 4071 | EXPORT_SYMBOL(set_user_nice); |
| 4072 | |
| 4073 | /* |
| 4074 | * can_nice - check if a task can reduce its nice value |
| 4075 | * @p: task |
| 4076 | * @nice: nice value |
| 4077 | */ |
| 4078 | int can_nice(const struct task_struct *p, const int nice) |
| 4079 | { |
| 4080 | /* convert nice value [19,-20] to rlimit style value [1,40] */ |
| 4081 | int nice_rlim = 20 - nice; |
| 4082 | |
| 4083 | return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur || |
| 4084 | capable(CAP_SYS_NICE)); |
| 4085 | } |
| 4086 | |
| 4087 | #ifdef __ARCH_WANT_SYS_NICE |
| 4088 | |
| 4089 | /* |
| 4090 | * sys_nice - change the priority of the current process. |
| 4091 | * @increment: priority increment |
| 4092 | * |
| 4093 | * sys_setpriority is a more generic, but much slower function that |
| 4094 | * does similar things. |
| 4095 | */ |
| 4096 | asmlinkage long sys_nice(int increment) |
| 4097 | { |
| 4098 | long nice, retval; |
| 4099 | |
| 4100 | /* |
| 4101 | * Setpriority might change our priority at the same moment. |
| 4102 | * We don't have to worry. Conceptually one call occurs first |
| 4103 | * and we have a single winner. |
| 4104 | */ |
| 4105 | if (increment < -40) |
| 4106 | increment = -40; |
| 4107 | if (increment > 40) |
| 4108 | increment = 40; |
| 4109 | |
| 4110 | nice = PRIO_TO_NICE(current->static_prio) + increment; |
| 4111 | if (nice < -20) |
| 4112 | nice = -20; |
| 4113 | if (nice > 19) |
| 4114 | nice = 19; |
| 4115 | |
| 4116 | if (increment < 0 && !can_nice(current, nice)) |
| 4117 | return -EPERM; |
| 4118 | |
| 4119 | retval = security_task_setnice(current, nice); |
| 4120 | if (retval) |
| 4121 | return retval; |
| 4122 | |
| 4123 | set_user_nice(current, nice); |
| 4124 | return 0; |
| 4125 | } |
| 4126 | |
| 4127 | #endif |
| 4128 | |
| 4129 | /** |
| 4130 | * task_prio - return the priority value of a given task. |
| 4131 | * @p: the task in question. |
| 4132 | * |
| 4133 | * This is the priority value as seen by users in /proc. |
| 4134 | * RT tasks are offset by -200. Normal tasks are centered |
| 4135 | * around 0, value goes from -16 to +15. |
| 4136 | */ |
| 4137 | int task_prio(const struct task_struct *p) |
| 4138 | { |
| 4139 | return p->prio - MAX_RT_PRIO; |
| 4140 | } |
| 4141 | |
| 4142 | /** |
| 4143 | * task_nice - return the nice value of a given task. |
| 4144 | * @p: the task in question. |
| 4145 | */ |
| 4146 | int task_nice(const struct task_struct *p) |
| 4147 | { |
| 4148 | return TASK_NICE(p); |
| 4149 | } |
| 4150 | EXPORT_SYMBOL_GPL(task_nice); |
| 4151 | |
| 4152 | /** |
| 4153 | * idle_cpu - is a given cpu idle currently? |
| 4154 | * @cpu: the processor in question. |
| 4155 | */ |
| 4156 | int idle_cpu(int cpu) |
| 4157 | { |
| 4158 | return cpu_curr(cpu) == cpu_rq(cpu)->idle; |
| 4159 | } |
| 4160 | |
| 4161 | /** |
| 4162 | * idle_task - return the idle task for a given cpu. |
| 4163 | * @cpu: the processor in question. |
| 4164 | */ |
| 4165 | struct task_struct *idle_task(int cpu) |
| 4166 | { |
| 4167 | return cpu_rq(cpu)->idle; |
| 4168 | } |
| 4169 | |
| 4170 | /** |
| 4171 | * find_process_by_pid - find a process with a matching PID value. |
| 4172 | * @pid: the pid in question. |
| 4173 | */ |
| 4174 | static struct task_struct *find_process_by_pid(pid_t pid) |
| 4175 | { |
| 4176 | return pid ? find_task_by_vpid(pid) : current; |
| 4177 | } |
| 4178 | |
| 4179 | /* Actually do priority change: must hold rq lock. */ |
| 4180 | static void |
| 4181 | __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio) |
| 4182 | { |
| 4183 | BUG_ON(p->se.on_rq); |
| 4184 | |
| 4185 | p->policy = policy; |
| 4186 | switch (p->policy) { |
| 4187 | case SCHED_NORMAL: |
| 4188 | case SCHED_BATCH: |
| 4189 | case SCHED_IDLE: |
| 4190 | p->sched_class = &fair_sched_class; |
| 4191 | break; |
| 4192 | case SCHED_FIFO: |
| 4193 | case SCHED_RR: |
| 4194 | p->sched_class = &rt_sched_class; |
| 4195 | break; |
| 4196 | } |
| 4197 | |
| 4198 | p->rt_priority = prio; |
| 4199 | p->normal_prio = normal_prio(p); |
| 4200 | /* we are holding p->pi_lock already */ |
| 4201 | p->prio = rt_mutex_getprio(p); |
| 4202 | set_load_weight(p); |
| 4203 | } |
| 4204 | |
| 4205 | /** |
| 4206 | * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. |
| 4207 | * @p: the task in question. |
| 4208 | * @policy: new policy. |
| 4209 | * @param: structure containing the new RT priority. |
| 4210 | * |
| 4211 | * NOTE that the task may be already dead. |
| 4212 | */ |
| 4213 | int sched_setscheduler(struct task_struct *p, int policy, |
| 4214 | struct sched_param *param) |
| 4215 | { |
| 4216 | int retval, oldprio, oldpolicy = -1, on_rq, running; |
| 4217 | unsigned long flags; |
| 4218 | struct rq *rq; |
| 4219 | |
| 4220 | /* may grab non-irq protected spin_locks */ |
| 4221 | BUG_ON(in_interrupt()); |
| 4222 | recheck: |
| 4223 | /* double check policy once rq lock held */ |
| 4224 | if (policy < 0) |
| 4225 | policy = oldpolicy = p->policy; |
| 4226 | else if (policy != SCHED_FIFO && policy != SCHED_RR && |
| 4227 | policy != SCHED_NORMAL && policy != SCHED_BATCH && |
| 4228 | policy != SCHED_IDLE) |
| 4229 | return -EINVAL; |
| 4230 | /* |
| 4231 | * Valid priorities for SCHED_FIFO and SCHED_RR are |
| 4232 | * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, |
| 4233 | * SCHED_BATCH and SCHED_IDLE is 0. |
| 4234 | */ |
| 4235 | if (param->sched_priority < 0 || |
| 4236 | (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || |
| 4237 | (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) |
| 4238 | return -EINVAL; |
| 4239 | if (rt_policy(policy) != (param->sched_priority != 0)) |
| 4240 | return -EINVAL; |
| 4241 | |
| 4242 | /* |
| 4243 | * Allow unprivileged RT tasks to decrease priority: |
| 4244 | */ |
| 4245 | if (!capable(CAP_SYS_NICE)) { |
| 4246 | if (rt_policy(policy)) { |
| 4247 | unsigned long rlim_rtprio; |
| 4248 | |
| 4249 | if (!lock_task_sighand(p, &flags)) |
| 4250 | return -ESRCH; |
| 4251 | rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur; |
| 4252 | unlock_task_sighand(p, &flags); |
| 4253 | |
| 4254 | /* can't set/change the rt policy */ |
| 4255 | if (policy != p->policy && !rlim_rtprio) |
| 4256 | return -EPERM; |
| 4257 | |
| 4258 | /* can't increase priority */ |
| 4259 | if (param->sched_priority > p->rt_priority && |
| 4260 | param->sched_priority > rlim_rtprio) |
| 4261 | return -EPERM; |
| 4262 | } |
| 4263 | /* |
| 4264 | * Like positive nice levels, dont allow tasks to |
| 4265 | * move out of SCHED_IDLE either: |
| 4266 | */ |
| 4267 | if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) |
| 4268 | return -EPERM; |
| 4269 | |
| 4270 | /* can't change other user's priorities */ |
| 4271 | if ((current->euid != p->euid) && |
| 4272 | (current->euid != p->uid)) |
| 4273 | return -EPERM; |
| 4274 | } |
| 4275 | |
| 4276 | retval = security_task_setscheduler(p, policy, param); |
| 4277 | if (retval) |
| 4278 | return retval; |
| 4279 | /* |
| 4280 | * make sure no PI-waiters arrive (or leave) while we are |
| 4281 | * changing the priority of the task: |
| 4282 | */ |
| 4283 | spin_lock_irqsave(&p->pi_lock, flags); |
| 4284 | /* |
| 4285 | * To be able to change p->policy safely, the apropriate |
| 4286 | * runqueue lock must be held. |
| 4287 | */ |
| 4288 | rq = __task_rq_lock(p); |
| 4289 | /* recheck policy now with rq lock held */ |
| 4290 | if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { |
| 4291 | policy = oldpolicy = -1; |
| 4292 | __task_rq_unlock(rq); |
| 4293 | spin_unlock_irqrestore(&p->pi_lock, flags); |
| 4294 | goto recheck; |
| 4295 | } |
| 4296 | update_rq_clock(rq); |
| 4297 | on_rq = p->se.on_rq; |
| 4298 | running = task_current(rq, p); |
| 4299 | if (on_rq) { |
| 4300 | deactivate_task(rq, p, 0); |
| 4301 | if (running) |
| 4302 | p->sched_class->put_prev_task(rq, p); |
| 4303 | } |
| 4304 | |
| 4305 | oldprio = p->prio; |
| 4306 | __setscheduler(rq, p, policy, param->sched_priority); |
| 4307 | |
| 4308 | if (on_rq) { |
| 4309 | if (running) |
| 4310 | p->sched_class->set_curr_task(rq); |
| 4311 | activate_task(rq, p, 0); |
| 4312 | /* |
| 4313 | * Reschedule if we are currently running on this runqueue and |
| 4314 | * our priority decreased, or if we are not currently running on |
| 4315 | * this runqueue and our priority is higher than the current's |
| 4316 | */ |
| 4317 | if (running) { |
| 4318 | if (p->prio > oldprio) |
| 4319 | resched_task(rq->curr); |
| 4320 | } else { |
| 4321 | check_preempt_curr(rq, p); |
| 4322 | } |
| 4323 | } |
| 4324 | __task_rq_unlock(rq); |
| 4325 | spin_unlock_irqrestore(&p->pi_lock, flags); |
| 4326 | |
| 4327 | rt_mutex_adjust_pi(p); |
| 4328 | |
| 4329 | return 0; |
| 4330 | } |
| 4331 | EXPORT_SYMBOL_GPL(sched_setscheduler); |
| 4332 | |
| 4333 | static int |
| 4334 | do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) |
| 4335 | { |
| 4336 | struct sched_param lparam; |
| 4337 | struct task_struct *p; |
| 4338 | int retval; |
| 4339 | |
| 4340 | if (!param || pid < 0) |
| 4341 | return -EINVAL; |
| 4342 | if (copy_from_user(&lparam, param, sizeof(struct sched_param))) |
| 4343 | return -EFAULT; |
| 4344 | |
| 4345 | rcu_read_lock(); |
| 4346 | retval = -ESRCH; |
| 4347 | p = find_process_by_pid(pid); |
| 4348 | if (p != NULL) |
| 4349 | retval = sched_setscheduler(p, policy, &lparam); |
| 4350 | rcu_read_unlock(); |
| 4351 | |
| 4352 | return retval; |
| 4353 | } |
| 4354 | |
| 4355 | /** |
| 4356 | * sys_sched_setscheduler - set/change the scheduler policy and RT priority |
| 4357 | * @pid: the pid in question. |
| 4358 | * @policy: new policy. |
| 4359 | * @param: structure containing the new RT priority. |
| 4360 | */ |
| 4361 | asmlinkage long |
| 4362 | sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) |
| 4363 | { |
| 4364 | /* negative values for policy are not valid */ |
| 4365 | if (policy < 0) |
| 4366 | return -EINVAL; |
| 4367 | |
| 4368 | return do_sched_setscheduler(pid, policy, param); |
| 4369 | } |
| 4370 | |
| 4371 | /** |
| 4372 | * sys_sched_setparam - set/change the RT priority of a thread |
| 4373 | * @pid: the pid in question. |
| 4374 | * @param: structure containing the new RT priority. |
| 4375 | */ |
| 4376 | asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param) |
| 4377 | { |
| 4378 | return do_sched_setscheduler(pid, -1, param); |
| 4379 | } |
| 4380 | |
| 4381 | /** |
| 4382 | * sys_sched_getscheduler - get the policy (scheduling class) of a thread |
| 4383 | * @pid: the pid in question. |
| 4384 | */ |
| 4385 | asmlinkage long sys_sched_getscheduler(pid_t pid) |
| 4386 | { |
| 4387 | struct task_struct *p; |
| 4388 | int retval; |
| 4389 | |
| 4390 | if (pid < 0) |
| 4391 | return -EINVAL; |
| 4392 | |
| 4393 | retval = -ESRCH; |
| 4394 | read_lock(&tasklist_lock); |
| 4395 | p = find_process_by_pid(pid); |
| 4396 | if (p) { |
| 4397 | retval = security_task_getscheduler(p); |
| 4398 | if (!retval) |
| 4399 | retval = p->policy; |
| 4400 | } |
| 4401 | read_unlock(&tasklist_lock); |
| 4402 | return retval; |
| 4403 | } |
| 4404 | |
| 4405 | /** |
| 4406 | * sys_sched_getscheduler - get the RT priority of a thread |
| 4407 | * @pid: the pid in question. |
| 4408 | * @param: structure containing the RT priority. |
| 4409 | */ |
| 4410 | asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param) |
| 4411 | { |
| 4412 | struct sched_param lp; |
| 4413 | struct task_struct *p; |
| 4414 | int retval; |
| 4415 | |
| 4416 | if (!param || pid < 0) |
| 4417 | return -EINVAL; |
| 4418 | |
| 4419 | read_lock(&tasklist_lock); |
| 4420 | p = find_process_by_pid(pid); |
| 4421 | retval = -ESRCH; |
| 4422 | if (!p) |
| 4423 | goto out_unlock; |
| 4424 | |
| 4425 | retval = security_task_getscheduler(p); |
| 4426 | if (retval) |
| 4427 | goto out_unlock; |
| 4428 | |
| 4429 | lp.sched_priority = p->rt_priority; |
| 4430 | read_unlock(&tasklist_lock); |
| 4431 | |
| 4432 | /* |
| 4433 | * This one might sleep, we cannot do it with a spinlock held ... |
| 4434 | */ |
| 4435 | retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; |
| 4436 | |
| 4437 | return retval; |
| 4438 | |
| 4439 | out_unlock: |
| 4440 | read_unlock(&tasklist_lock); |
| 4441 | return retval; |
| 4442 | } |
| 4443 | |
| 4444 | long sched_setaffinity(pid_t pid, cpumask_t new_mask) |
| 4445 | { |
| 4446 | cpumask_t cpus_allowed; |
| 4447 | struct task_struct *p; |
| 4448 | int retval; |
| 4449 | |
| 4450 | get_online_cpus(); |
| 4451 | read_lock(&tasklist_lock); |
| 4452 | |
| 4453 | p = find_process_by_pid(pid); |
| 4454 | if (!p) { |
| 4455 | read_unlock(&tasklist_lock); |
| 4456 | put_online_cpus(); |
| 4457 | return -ESRCH; |
| 4458 | } |
| 4459 | |
| 4460 | /* |
| 4461 | * It is not safe to call set_cpus_allowed with the |
| 4462 | * tasklist_lock held. We will bump the task_struct's |
| 4463 | * usage count and then drop tasklist_lock. |
| 4464 | */ |
| 4465 | get_task_struct(p); |
| 4466 | read_unlock(&tasklist_lock); |
| 4467 | |
| 4468 | retval = -EPERM; |
| 4469 | if ((current->euid != p->euid) && (current->euid != p->uid) && |
| 4470 | !capable(CAP_SYS_NICE)) |
| 4471 | goto out_unlock; |
| 4472 | |
| 4473 | retval = security_task_setscheduler(p, 0, NULL); |
| 4474 | if (retval) |
| 4475 | goto out_unlock; |
| 4476 | |
| 4477 | cpus_allowed = cpuset_cpus_allowed(p); |
| 4478 | cpus_and(new_mask, new_mask, cpus_allowed); |
| 4479 | again: |
| 4480 | retval = set_cpus_allowed(p, new_mask); |
| 4481 | |
| 4482 | if (!retval) { |
| 4483 | cpus_allowed = cpuset_cpus_allowed(p); |
| 4484 | if (!cpus_subset(new_mask, cpus_allowed)) { |
| 4485 | /* |
| 4486 | * We must have raced with a concurrent cpuset |
| 4487 | * update. Just reset the cpus_allowed to the |
| 4488 | * cpuset's cpus_allowed |
| 4489 | */ |
| 4490 | new_mask = cpus_allowed; |
| 4491 | goto again; |
| 4492 | } |
| 4493 | } |
| 4494 | out_unlock: |
| 4495 | put_task_struct(p); |
| 4496 | put_online_cpus(); |
| 4497 | return retval; |
| 4498 | } |
| 4499 | |
| 4500 | static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, |
| 4501 | cpumask_t *new_mask) |
| 4502 | { |
| 4503 | if (len < sizeof(cpumask_t)) { |
| 4504 | memset(new_mask, 0, sizeof(cpumask_t)); |
| 4505 | } else if (len > sizeof(cpumask_t)) { |
| 4506 | len = sizeof(cpumask_t); |
| 4507 | } |
| 4508 | return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; |
| 4509 | } |
| 4510 | |
| 4511 | /** |
| 4512 | * sys_sched_setaffinity - set the cpu affinity of a process |
| 4513 | * @pid: pid of the process |
| 4514 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr |
| 4515 | * @user_mask_ptr: user-space pointer to the new cpu mask |
| 4516 | */ |
| 4517 | asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len, |
| 4518 | unsigned long __user *user_mask_ptr) |
| 4519 | { |
| 4520 | cpumask_t new_mask; |
| 4521 | int retval; |
| 4522 | |
| 4523 | retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask); |
| 4524 | if (retval) |
| 4525 | return retval; |
| 4526 | |
| 4527 | return sched_setaffinity(pid, new_mask); |
| 4528 | } |
| 4529 | |
| 4530 | /* |
| 4531 | * Represents all cpu's present in the system |
| 4532 | * In systems capable of hotplug, this map could dynamically grow |
| 4533 | * as new cpu's are detected in the system via any platform specific |
| 4534 | * method, such as ACPI for e.g. |
| 4535 | */ |
| 4536 | |
| 4537 | cpumask_t cpu_present_map __read_mostly; |
| 4538 | EXPORT_SYMBOL(cpu_present_map); |
| 4539 | |
| 4540 | #ifndef CONFIG_SMP |
| 4541 | cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL; |
| 4542 | EXPORT_SYMBOL(cpu_online_map); |
| 4543 | |
| 4544 | cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL; |
| 4545 | EXPORT_SYMBOL(cpu_possible_map); |
| 4546 | #endif |
| 4547 | |
| 4548 | long sched_getaffinity(pid_t pid, cpumask_t *mask) |
| 4549 | { |
| 4550 | struct task_struct *p; |
| 4551 | int retval; |
| 4552 | |
| 4553 | get_online_cpus(); |
| 4554 | read_lock(&tasklist_lock); |
| 4555 | |
| 4556 | retval = -ESRCH; |
| 4557 | p = find_process_by_pid(pid); |
| 4558 | if (!p) |
| 4559 | goto out_unlock; |
| 4560 | |
| 4561 | retval = security_task_getscheduler(p); |
| 4562 | if (retval) |
| 4563 | goto out_unlock; |
| 4564 | |
| 4565 | cpus_and(*mask, p->cpus_allowed, cpu_online_map); |
| 4566 | |
| 4567 | out_unlock: |
| 4568 | read_unlock(&tasklist_lock); |
| 4569 | put_online_cpus(); |
| 4570 | |
| 4571 | return retval; |
| 4572 | } |
| 4573 | |
| 4574 | /** |
| 4575 | * sys_sched_getaffinity - get the cpu affinity of a process |
| 4576 | * @pid: pid of the process |
| 4577 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr |
| 4578 | * @user_mask_ptr: user-space pointer to hold the current cpu mask |
| 4579 | */ |
| 4580 | asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len, |
| 4581 | unsigned long __user *user_mask_ptr) |
| 4582 | { |
| 4583 | int ret; |
| 4584 | cpumask_t mask; |
| 4585 | |
| 4586 | if (len < sizeof(cpumask_t)) |
| 4587 | return -EINVAL; |
| 4588 | |
| 4589 | ret = sched_getaffinity(pid, &mask); |
| 4590 | if (ret < 0) |
| 4591 | return ret; |
| 4592 | |
| 4593 | if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t))) |
| 4594 | return -EFAULT; |
| 4595 | |
| 4596 | return sizeof(cpumask_t); |
| 4597 | } |
| 4598 | |
| 4599 | /** |
| 4600 | * sys_sched_yield - yield the current processor to other threads. |
| 4601 | * |
| 4602 | * This function yields the current CPU to other tasks. If there are no |
| 4603 | * other threads running on this CPU then this function will return. |
| 4604 | */ |
| 4605 | asmlinkage long sys_sched_yield(void) |
| 4606 | { |
| 4607 | struct rq *rq = this_rq_lock(); |
| 4608 | |
| 4609 | schedstat_inc(rq, yld_count); |
| 4610 | current->sched_class->yield_task(rq); |
| 4611 | |
| 4612 | /* |
| 4613 | * Since we are going to call schedule() anyway, there's |
| 4614 | * no need to preempt or enable interrupts: |
| 4615 | */ |
| 4616 | __release(rq->lock); |
| 4617 | spin_release(&rq->lock.dep_map, 1, _THIS_IP_); |
| 4618 | _raw_spin_unlock(&rq->lock); |
| 4619 | preempt_enable_no_resched(); |
| 4620 | |
| 4621 | schedule(); |
| 4622 | |
| 4623 | return 0; |
| 4624 | } |
| 4625 | |
| 4626 | static void __cond_resched(void) |
| 4627 | { |
| 4628 | #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP |
| 4629 | __might_sleep(__FILE__, __LINE__); |
| 4630 | #endif |
| 4631 | /* |
| 4632 | * The BKS might be reacquired before we have dropped |
| 4633 | * PREEMPT_ACTIVE, which could trigger a second |
| 4634 | * cond_resched() call. |
| 4635 | */ |
| 4636 | do { |
| 4637 | add_preempt_count(PREEMPT_ACTIVE); |
| 4638 | schedule(); |
| 4639 | sub_preempt_count(PREEMPT_ACTIVE); |
| 4640 | } while (need_resched()); |
| 4641 | } |
| 4642 | |
| 4643 | int __sched cond_resched(void) |
| 4644 | { |
| 4645 | if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) && |
| 4646 | system_state == SYSTEM_RUNNING) { |
| 4647 | __cond_resched(); |
| 4648 | return 1; |
| 4649 | } |
| 4650 | return 0; |
| 4651 | } |
| 4652 | EXPORT_SYMBOL(cond_resched); |
| 4653 | |
| 4654 | /* |
| 4655 | * cond_resched_lock() - if a reschedule is pending, drop the given lock, |
| 4656 | * call schedule, and on return reacquire the lock. |
| 4657 | * |
| 4658 | * This works OK both with and without CONFIG_PREEMPT. We do strange low-level |
| 4659 | * operations here to prevent schedule() from being called twice (once via |
| 4660 | * spin_unlock(), once by hand). |
| 4661 | */ |
| 4662 | int cond_resched_lock(spinlock_t *lock) |
| 4663 | { |
| 4664 | int ret = 0; |
| 4665 | |
| 4666 | if (need_lockbreak(lock)) { |
| 4667 | spin_unlock(lock); |
| 4668 | cpu_relax(); |
| 4669 | ret = 1; |
| 4670 | spin_lock(lock); |
| 4671 | } |
| 4672 | if (need_resched() && system_state == SYSTEM_RUNNING) { |
| 4673 | spin_release(&lock->dep_map, 1, _THIS_IP_); |
| 4674 | _raw_spin_unlock(lock); |
| 4675 | preempt_enable_no_resched(); |
| 4676 | __cond_resched(); |
| 4677 | ret = 1; |
| 4678 | spin_lock(lock); |
| 4679 | } |
| 4680 | return ret; |
| 4681 | } |
| 4682 | EXPORT_SYMBOL(cond_resched_lock); |
| 4683 | |
| 4684 | int __sched cond_resched_softirq(void) |
| 4685 | { |
| 4686 | BUG_ON(!in_softirq()); |
| 4687 | |
| 4688 | if (need_resched() && system_state == SYSTEM_RUNNING) { |
| 4689 | local_bh_enable(); |
| 4690 | __cond_resched(); |
| 4691 | local_bh_disable(); |
| 4692 | return 1; |
| 4693 | } |
| 4694 | return 0; |
| 4695 | } |
| 4696 | EXPORT_SYMBOL(cond_resched_softirq); |
| 4697 | |
| 4698 | /** |
| 4699 | * yield - yield the current processor to other threads. |
| 4700 | * |
| 4701 | * This is a shortcut for kernel-space yielding - it marks the |
| 4702 | * thread runnable and calls sys_sched_yield(). |
| 4703 | */ |
| 4704 | void __sched yield(void) |
| 4705 | { |
| 4706 | set_current_state(TASK_RUNNING); |
| 4707 | sys_sched_yield(); |
| 4708 | } |
| 4709 | EXPORT_SYMBOL(yield); |
| 4710 | |
| 4711 | /* |
| 4712 | * This task is about to go to sleep on IO. Increment rq->nr_iowait so |
| 4713 | * that process accounting knows that this is a task in IO wait state. |
| 4714 | * |
| 4715 | * But don't do that if it is a deliberate, throttling IO wait (this task |
| 4716 | * has set its backing_dev_info: the queue against which it should throttle) |
| 4717 | */ |
| 4718 | void __sched io_schedule(void) |
| 4719 | { |
| 4720 | struct rq *rq = &__raw_get_cpu_var(runqueues); |
| 4721 | |
| 4722 | delayacct_blkio_start(); |
| 4723 | atomic_inc(&rq->nr_iowait); |
| 4724 | schedule(); |
| 4725 | atomic_dec(&rq->nr_iowait); |
| 4726 | delayacct_blkio_end(); |
| 4727 | } |
| 4728 | EXPORT_SYMBOL(io_schedule); |
| 4729 | |
| 4730 | long __sched io_schedule_timeout(long timeout) |
| 4731 | { |
| 4732 | struct rq *rq = &__raw_get_cpu_var(runqueues); |
| 4733 | long ret; |
| 4734 | |
| 4735 | delayacct_blkio_start(); |
| 4736 | atomic_inc(&rq->nr_iowait); |
| 4737 | ret = schedule_timeout(timeout); |
| 4738 | atomic_dec(&rq->nr_iowait); |
| 4739 | delayacct_blkio_end(); |
| 4740 | return ret; |
| 4741 | } |
| 4742 | |
| 4743 | /** |
| 4744 | * sys_sched_get_priority_max - return maximum RT priority. |
| 4745 | * @policy: scheduling class. |
| 4746 | * |
| 4747 | * this syscall returns the maximum rt_priority that can be used |
| 4748 | * by a given scheduling class. |
| 4749 | */ |
| 4750 | asmlinkage long sys_sched_get_priority_max(int policy) |
| 4751 | { |
| 4752 | int ret = -EINVAL; |
| 4753 | |
| 4754 | switch (policy) { |
| 4755 | case SCHED_FIFO: |
| 4756 | case SCHED_RR: |
| 4757 | ret = MAX_USER_RT_PRIO-1; |
| 4758 | break; |
| 4759 | case SCHED_NORMAL: |
| 4760 | case SCHED_BATCH: |
| 4761 | case SCHED_IDLE: |
| 4762 | ret = 0; |
| 4763 | break; |
| 4764 | } |
| 4765 | return ret; |
| 4766 | } |
| 4767 | |
| 4768 | /** |
| 4769 | * sys_sched_get_priority_min - return minimum RT priority. |
| 4770 | * @policy: scheduling class. |
| 4771 | * |
| 4772 | * this syscall returns the minimum rt_priority that can be used |
| 4773 | * by a given scheduling class. |
| 4774 | */ |
| 4775 | asmlinkage long sys_sched_get_priority_min(int policy) |
| 4776 | { |
| 4777 | int ret = -EINVAL; |
| 4778 | |
| 4779 | switch (policy) { |
| 4780 | case SCHED_FIFO: |
| 4781 | case SCHED_RR: |
| 4782 | ret = 1; |
| 4783 | break; |
| 4784 | case SCHED_NORMAL: |
| 4785 | case SCHED_BATCH: |
| 4786 | case SCHED_IDLE: |
| 4787 | ret = 0; |
| 4788 | } |
| 4789 | return ret; |
| 4790 | } |
| 4791 | |
| 4792 | /** |
| 4793 | * sys_sched_rr_get_interval - return the default timeslice of a process. |
| 4794 | * @pid: pid of the process. |
| 4795 | * @interval: userspace pointer to the timeslice value. |
| 4796 | * |
| 4797 | * this syscall writes the default timeslice value of a given process |
| 4798 | * into the user-space timespec buffer. A value of '0' means infinity. |
| 4799 | */ |
| 4800 | asmlinkage |
| 4801 | long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval) |
| 4802 | { |
| 4803 | struct task_struct *p; |
| 4804 | unsigned int time_slice; |
| 4805 | int retval; |
| 4806 | struct timespec t; |
| 4807 | |
| 4808 | if (pid < 0) |
| 4809 | return -EINVAL; |
| 4810 | |
| 4811 | retval = -ESRCH; |
| 4812 | read_lock(&tasklist_lock); |
| 4813 | p = find_process_by_pid(pid); |
| 4814 | if (!p) |
| 4815 | goto out_unlock; |
| 4816 | |
| 4817 | retval = security_task_getscheduler(p); |
| 4818 | if (retval) |
| 4819 | goto out_unlock; |
| 4820 | |
| 4821 | /* |
| 4822 | * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER |
| 4823 | * tasks that are on an otherwise idle runqueue: |
| 4824 | */ |
| 4825 | time_slice = 0; |
| 4826 | if (p->policy == SCHED_RR) { |
| 4827 | time_slice = DEF_TIMESLICE; |
| 4828 | } else { |
| 4829 | struct sched_entity *se = &p->se; |
| 4830 | unsigned long flags; |
| 4831 | struct rq *rq; |
| 4832 | |
| 4833 | rq = task_rq_lock(p, &flags); |
| 4834 | if (rq->cfs.load.weight) |
| 4835 | time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se)); |
| 4836 | task_rq_unlock(rq, &flags); |
| 4837 | } |
| 4838 | read_unlock(&tasklist_lock); |
| 4839 | jiffies_to_timespec(time_slice, &t); |
| 4840 | retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; |
| 4841 | return retval; |
| 4842 | |
| 4843 | out_unlock: |
| 4844 | read_unlock(&tasklist_lock); |
| 4845 | return retval; |
| 4846 | } |
| 4847 | |
| 4848 | static const char stat_nam[] = "RSDTtZX"; |
| 4849 | |
| 4850 | void sched_show_task(struct task_struct *p) |
| 4851 | { |
| 4852 | unsigned long free = 0; |
| 4853 | unsigned state; |
| 4854 | |
| 4855 | state = p->state ? __ffs(p->state) + 1 : 0; |
| 4856 | printk(KERN_INFO "%-13.13s %c", p->comm, |
| 4857 | state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); |
| 4858 | #if BITS_PER_LONG == 32 |
| 4859 | if (state == TASK_RUNNING) |
| 4860 | printk(KERN_CONT " running "); |
| 4861 | else |
| 4862 | printk(KERN_CONT " %08lx ", thread_saved_pc(p)); |
| 4863 | #else |
| 4864 | if (state == TASK_RUNNING) |
| 4865 | printk(KERN_CONT " running task "); |
| 4866 | else |
| 4867 | printk(KERN_CONT " %016lx ", thread_saved_pc(p)); |
| 4868 | #endif |
| 4869 | #ifdef CONFIG_DEBUG_STACK_USAGE |
| 4870 | { |
| 4871 | unsigned long *n = end_of_stack(p); |
| 4872 | while (!*n) |
| 4873 | n++; |
| 4874 | free = (unsigned long)n - (unsigned long)end_of_stack(p); |
| 4875 | } |
| 4876 | #endif |
| 4877 | printk(KERN_CONT "%5lu %5d %6d\n", free, |
| 4878 | task_pid_nr(p), task_pid_nr(p->real_parent)); |
| 4879 | |
| 4880 | if (state != TASK_RUNNING) |
| 4881 | show_stack(p, NULL); |
| 4882 | } |
| 4883 | |
| 4884 | void show_state_filter(unsigned long state_filter) |
| 4885 | { |
| 4886 | struct task_struct *g, *p; |
| 4887 | |
| 4888 | #if BITS_PER_LONG == 32 |
| 4889 | printk(KERN_INFO |
| 4890 | " task PC stack pid father\n"); |
| 4891 | #else |
| 4892 | printk(KERN_INFO |
| 4893 | " task PC stack pid father\n"); |
| 4894 | #endif |
| 4895 | read_lock(&tasklist_lock); |
| 4896 | do_each_thread(g, p) { |
| 4897 | /* |
| 4898 | * reset the NMI-timeout, listing all files on a slow |
| 4899 | * console might take alot of time: |
| 4900 | */ |
| 4901 | touch_nmi_watchdog(); |
| 4902 | if (!state_filter || (p->state & state_filter)) |
| 4903 | sched_show_task(p); |
| 4904 | } while_each_thread(g, p); |
| 4905 | |
| 4906 | touch_all_softlockup_watchdogs(); |
| 4907 | |
| 4908 | #ifdef CONFIG_SCHED_DEBUG |
| 4909 | sysrq_sched_debug_show(); |
| 4910 | #endif |
| 4911 | read_unlock(&tasklist_lock); |
| 4912 | /* |
| 4913 | * Only show locks if all tasks are dumped: |
| 4914 | */ |
| 4915 | if (state_filter == -1) |
| 4916 | debug_show_all_locks(); |
| 4917 | } |
| 4918 | |
| 4919 | void __cpuinit init_idle_bootup_task(struct task_struct *idle) |
| 4920 | { |
| 4921 | idle->sched_class = &idle_sched_class; |
| 4922 | } |
| 4923 | |
| 4924 | /** |
| 4925 | * init_idle - set up an idle thread for a given CPU |
| 4926 | * @idle: task in question |
| 4927 | * @cpu: cpu the idle task belongs to |
| 4928 | * |
| 4929 | * NOTE: this function does not set the idle thread's NEED_RESCHED |
| 4930 | * flag, to make booting more robust. |
| 4931 | */ |
| 4932 | void __cpuinit init_idle(struct task_struct *idle, int cpu) |
| 4933 | { |
| 4934 | struct rq *rq = cpu_rq(cpu); |
| 4935 | unsigned long flags; |
| 4936 | |
| 4937 | __sched_fork(idle); |
| 4938 | idle->se.exec_start = sched_clock(); |
| 4939 | |
| 4940 | idle->prio = idle->normal_prio = MAX_PRIO; |
| 4941 | idle->cpus_allowed = cpumask_of_cpu(cpu); |
| 4942 | __set_task_cpu(idle, cpu); |
| 4943 | |
| 4944 | spin_lock_irqsave(&rq->lock, flags); |
| 4945 | rq->curr = rq->idle = idle; |
| 4946 | #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) |
| 4947 | idle->oncpu = 1; |
| 4948 | #endif |
| 4949 | spin_unlock_irqrestore(&rq->lock, flags); |
| 4950 | |
| 4951 | /* Set the preempt count _outside_ the spinlocks! */ |
| 4952 | #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL) |
| 4953 | task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0); |
| 4954 | #else |
| 4955 | task_thread_info(idle)->preempt_count = 0; |
| 4956 | #endif |
| 4957 | /* |
| 4958 | * The idle tasks have their own, simple scheduling class: |
| 4959 | */ |
| 4960 | idle->sched_class = &idle_sched_class; |
| 4961 | } |
| 4962 | |
| 4963 | /* |
| 4964 | * In a system that switches off the HZ timer nohz_cpu_mask |
| 4965 | * indicates which cpus entered this state. This is used |
| 4966 | * in the rcu update to wait only for active cpus. For system |
| 4967 | * which do not switch off the HZ timer nohz_cpu_mask should |
| 4968 | * always be CPU_MASK_NONE. |
| 4969 | */ |
| 4970 | cpumask_t nohz_cpu_mask = CPU_MASK_NONE; |
| 4971 | |
| 4972 | /* |
| 4973 | * Increase the granularity value when there are more CPUs, |
| 4974 | * because with more CPUs the 'effective latency' as visible |
| 4975 | * to users decreases. But the relationship is not linear, |
| 4976 | * so pick a second-best guess by going with the log2 of the |
| 4977 | * number of CPUs. |
| 4978 | * |
| 4979 | * This idea comes from the SD scheduler of Con Kolivas: |
| 4980 | */ |
| 4981 | static inline void sched_init_granularity(void) |
| 4982 | { |
| 4983 | unsigned int factor = 1 + ilog2(num_online_cpus()); |
| 4984 | const unsigned long limit = 200000000; |
| 4985 | |
| 4986 | sysctl_sched_min_granularity *= factor; |
| 4987 | if (sysctl_sched_min_granularity > limit) |
| 4988 | sysctl_sched_min_granularity = limit; |
| 4989 | |
| 4990 | sysctl_sched_latency *= factor; |
| 4991 | if (sysctl_sched_latency > limit) |
| 4992 | sysctl_sched_latency = limit; |
| 4993 | |
| 4994 | sysctl_sched_wakeup_granularity *= factor; |
| 4995 | sysctl_sched_batch_wakeup_granularity *= factor; |
| 4996 | } |
| 4997 | |
| 4998 | #ifdef CONFIG_SMP |
| 4999 | /* |
| 5000 | * This is how migration works: |
| 5001 | * |
| 5002 | * 1) we queue a struct migration_req structure in the source CPU's |
| 5003 | * runqueue and wake up that CPU's migration thread. |
| 5004 | * 2) we down() the locked semaphore => thread blocks. |
| 5005 | * 3) migration thread wakes up (implicitly it forces the migrated |
| 5006 | * thread off the CPU) |
| 5007 | * 4) it gets the migration request and checks whether the migrated |
| 5008 | * task is still in the wrong runqueue. |
| 5009 | * 5) if it's in the wrong runqueue then the migration thread removes |
| 5010 | * it and puts it into the right queue. |
| 5011 | * 6) migration thread up()s the semaphore. |
| 5012 | * 7) we wake up and the migration is done. |
| 5013 | */ |
| 5014 | |
| 5015 | /* |
| 5016 | * Change a given task's CPU affinity. Migrate the thread to a |
| 5017 | * proper CPU and schedule it away if the CPU it's executing on |
| 5018 | * is removed from the allowed bitmask. |
| 5019 | * |
| 5020 | * NOTE: the caller must have a valid reference to the task, the |
| 5021 | * task must not exit() & deallocate itself prematurely. The |
| 5022 | * call is not atomic; no spinlocks may be held. |
| 5023 | */ |
| 5024 | int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask) |
| 5025 | { |
| 5026 | struct migration_req req; |
| 5027 | unsigned long flags; |
| 5028 | struct rq *rq; |
| 5029 | int ret = 0; |
| 5030 | |
| 5031 | rq = task_rq_lock(p, &flags); |
| 5032 | if (!cpus_intersects(new_mask, cpu_online_map)) { |
| 5033 | ret = -EINVAL; |
| 5034 | goto out; |
| 5035 | } |
| 5036 | |
| 5037 | if (p->sched_class->set_cpus_allowed) |
| 5038 | p->sched_class->set_cpus_allowed(p, &new_mask); |
| 5039 | else { |
| 5040 | p->cpus_allowed = new_mask; |
| 5041 | p->nr_cpus_allowed = cpus_weight(new_mask); |
| 5042 | } |
| 5043 | |
| 5044 | /* Can the task run on the task's current CPU? If so, we're done */ |
| 5045 | if (cpu_isset(task_cpu(p), new_mask)) |
| 5046 | goto out; |
| 5047 | |
| 5048 | if (migrate_task(p, any_online_cpu(new_mask), &req)) { |
| 5049 | /* Need help from migration thread: drop lock and wait. */ |
| 5050 | task_rq_unlock(rq, &flags); |
| 5051 | wake_up_process(rq->migration_thread); |
| 5052 | wait_for_completion(&req.done); |
| 5053 | tlb_migrate_finish(p->mm); |
| 5054 | return 0; |
| 5055 | } |
| 5056 | out: |
| 5057 | task_rq_unlock(rq, &flags); |
| 5058 | |
| 5059 | return ret; |
| 5060 | } |
| 5061 | EXPORT_SYMBOL_GPL(set_cpus_allowed); |
| 5062 | |
| 5063 | /* |
| 5064 | * Move (not current) task off this cpu, onto dest cpu. We're doing |
| 5065 | * this because either it can't run here any more (set_cpus_allowed() |
| 5066 | * away from this CPU, or CPU going down), or because we're |
| 5067 | * attempting to rebalance this task on exec (sched_exec). |
| 5068 | * |
| 5069 | * So we race with normal scheduler movements, but that's OK, as long |
| 5070 | * as the task is no longer on this CPU. |
| 5071 | * |
| 5072 | * Returns non-zero if task was successfully migrated. |
| 5073 | */ |
| 5074 | static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) |
| 5075 | { |
| 5076 | struct rq *rq_dest, *rq_src; |
| 5077 | int ret = 0, on_rq; |
| 5078 | |
| 5079 | if (unlikely(cpu_is_offline(dest_cpu))) |
| 5080 | return ret; |
| 5081 | |
| 5082 | rq_src = cpu_rq(src_cpu); |
| 5083 | rq_dest = cpu_rq(dest_cpu); |
| 5084 | |
| 5085 | double_rq_lock(rq_src, rq_dest); |
| 5086 | /* Already moved. */ |
| 5087 | if (task_cpu(p) != src_cpu) |
| 5088 | goto out; |
| 5089 | /* Affinity changed (again). */ |
| 5090 | if (!cpu_isset(dest_cpu, p->cpus_allowed)) |
| 5091 | goto out; |
| 5092 | |
| 5093 | on_rq = p->se.on_rq; |
| 5094 | if (on_rq) |
| 5095 | deactivate_task(rq_src, p, 0); |
| 5096 | |
| 5097 | set_task_cpu(p, dest_cpu); |
| 5098 | if (on_rq) { |
| 5099 | activate_task(rq_dest, p, 0); |
| 5100 | check_preempt_curr(rq_dest, p); |
| 5101 | } |
| 5102 | ret = 1; |
| 5103 | out: |
| 5104 | double_rq_unlock(rq_src, rq_dest); |
| 5105 | return ret; |
| 5106 | } |
| 5107 | |
| 5108 | /* |
| 5109 | * migration_thread - this is a highprio system thread that performs |
| 5110 | * thread migration by bumping thread off CPU then 'pushing' onto |
| 5111 | * another runqueue. |
| 5112 | */ |
| 5113 | static int migration_thread(void *data) |
| 5114 | { |
| 5115 | int cpu = (long)data; |
| 5116 | struct rq *rq; |
| 5117 | |
| 5118 | rq = cpu_rq(cpu); |
| 5119 | BUG_ON(rq->migration_thread != current); |
| 5120 | |
| 5121 | set_current_state(TASK_INTERRUPTIBLE); |
| 5122 | while (!kthread_should_stop()) { |
| 5123 | struct migration_req *req; |
| 5124 | struct list_head *head; |
| 5125 | |
| 5126 | spin_lock_irq(&rq->lock); |
| 5127 | |
| 5128 | if (cpu_is_offline(cpu)) { |
| 5129 | spin_unlock_irq(&rq->lock); |
| 5130 | goto wait_to_die; |
| 5131 | } |
| 5132 | |
| 5133 | if (rq->active_balance) { |
| 5134 | active_load_balance(rq, cpu); |
| 5135 | rq->active_balance = 0; |
| 5136 | } |
| 5137 | |
| 5138 | head = &rq->migration_queue; |
| 5139 | |
| 5140 | if (list_empty(head)) { |
| 5141 | spin_unlock_irq(&rq->lock); |
| 5142 | schedule(); |
| 5143 | set_current_state(TASK_INTERRUPTIBLE); |
| 5144 | continue; |
| 5145 | } |
| 5146 | req = list_entry(head->next, struct migration_req, list); |
| 5147 | list_del_init(head->next); |
| 5148 | |
| 5149 | spin_unlock(&rq->lock); |
| 5150 | __migrate_task(req->task, cpu, req->dest_cpu); |
| 5151 | local_irq_enable(); |
| 5152 | |
| 5153 | complete(&req->done); |
| 5154 | } |
| 5155 | __set_current_state(TASK_RUNNING); |
| 5156 | return 0; |
| 5157 | |
| 5158 | wait_to_die: |
| 5159 | /* Wait for kthread_stop */ |
| 5160 | set_current_state(TASK_INTERRUPTIBLE); |
| 5161 | while (!kthread_should_stop()) { |
| 5162 | schedule(); |
| 5163 | set_current_state(TASK_INTERRUPTIBLE); |
| 5164 | } |
| 5165 | __set_current_state(TASK_RUNNING); |
| 5166 | return 0; |
| 5167 | } |
| 5168 | |
| 5169 | #ifdef CONFIG_HOTPLUG_CPU |
| 5170 | |
| 5171 | static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu) |
| 5172 | { |
| 5173 | int ret; |
| 5174 | |
| 5175 | local_irq_disable(); |
| 5176 | ret = __migrate_task(p, src_cpu, dest_cpu); |
| 5177 | local_irq_enable(); |
| 5178 | return ret; |
| 5179 | } |
| 5180 | |
| 5181 | /* |
| 5182 | * Figure out where task on dead CPU should go, use force if necessary. |
| 5183 | * NOTE: interrupts should be disabled by the caller |
| 5184 | */ |
| 5185 | static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p) |
| 5186 | { |
| 5187 | unsigned long flags; |
| 5188 | cpumask_t mask; |
| 5189 | struct rq *rq; |
| 5190 | int dest_cpu; |
| 5191 | |
| 5192 | do { |
| 5193 | /* On same node? */ |
| 5194 | mask = node_to_cpumask(cpu_to_node(dead_cpu)); |
| 5195 | cpus_and(mask, mask, p->cpus_allowed); |
| 5196 | dest_cpu = any_online_cpu(mask); |
| 5197 | |
| 5198 | /* On any allowed CPU? */ |
| 5199 | if (dest_cpu == NR_CPUS) |
| 5200 | dest_cpu = any_online_cpu(p->cpus_allowed); |
| 5201 | |
| 5202 | /* No more Mr. Nice Guy. */ |
| 5203 | if (dest_cpu == NR_CPUS) { |
| 5204 | cpumask_t cpus_allowed = cpuset_cpus_allowed_locked(p); |
| 5205 | /* |
| 5206 | * Try to stay on the same cpuset, where the |
| 5207 | * current cpuset may be a subset of all cpus. |
| 5208 | * The cpuset_cpus_allowed_locked() variant of |
| 5209 | * cpuset_cpus_allowed() will not block. It must be |
| 5210 | * called within calls to cpuset_lock/cpuset_unlock. |
| 5211 | */ |
| 5212 | rq = task_rq_lock(p, &flags); |
| 5213 | p->cpus_allowed = cpus_allowed; |
| 5214 | dest_cpu = any_online_cpu(p->cpus_allowed); |
| 5215 | task_rq_unlock(rq, &flags); |
| 5216 | |
| 5217 | /* |
| 5218 | * Don't tell them about moving exiting tasks or |
| 5219 | * kernel threads (both mm NULL), since they never |
| 5220 | * leave kernel. |
| 5221 | */ |
| 5222 | if (p->mm && printk_ratelimit()) { |
| 5223 | printk(KERN_INFO "process %d (%s) no " |
| 5224 | "longer affine to cpu%d\n", |
| 5225 | task_pid_nr(p), p->comm, dead_cpu); |
| 5226 | } |
| 5227 | } |
| 5228 | } while (!__migrate_task_irq(p, dead_cpu, dest_cpu)); |
| 5229 | } |
| 5230 | |
| 5231 | /* |
| 5232 | * While a dead CPU has no uninterruptible tasks queued at this point, |
| 5233 | * it might still have a nonzero ->nr_uninterruptible counter, because |
| 5234 | * for performance reasons the counter is not stricly tracking tasks to |
| 5235 | * their home CPUs. So we just add the counter to another CPU's counter, |
| 5236 | * to keep the global sum constant after CPU-down: |
| 5237 | */ |
| 5238 | static void migrate_nr_uninterruptible(struct rq *rq_src) |
| 5239 | { |
| 5240 | struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL)); |
| 5241 | unsigned long flags; |
| 5242 | |
| 5243 | local_irq_save(flags); |
| 5244 | double_rq_lock(rq_src, rq_dest); |
| 5245 | rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible; |
| 5246 | rq_src->nr_uninterruptible = 0; |
| 5247 | double_rq_unlock(rq_src, rq_dest); |
| 5248 | local_irq_restore(flags); |
| 5249 | } |
| 5250 | |
| 5251 | /* Run through task list and migrate tasks from the dead cpu. */ |
| 5252 | static void migrate_live_tasks(int src_cpu) |
| 5253 | { |
| 5254 | struct task_struct *p, *t; |
| 5255 | |
| 5256 | read_lock(&tasklist_lock); |
| 5257 | |
| 5258 | do_each_thread(t, p) { |
| 5259 | if (p == current) |
| 5260 | continue; |
| 5261 | |
| 5262 | if (task_cpu(p) == src_cpu) |
| 5263 | move_task_off_dead_cpu(src_cpu, p); |
| 5264 | } while_each_thread(t, p); |
| 5265 | |
| 5266 | read_unlock(&tasklist_lock); |
| 5267 | } |
| 5268 | |
| 5269 | /* |
| 5270 | * Schedules idle task to be the next runnable task on current CPU. |
| 5271 | * It does so by boosting its priority to highest possible. |
| 5272 | * Used by CPU offline code. |
| 5273 | */ |
| 5274 | void sched_idle_next(void) |
| 5275 | { |
| 5276 | int this_cpu = smp_processor_id(); |
| 5277 | struct rq *rq = cpu_rq(this_cpu); |
| 5278 | struct task_struct *p = rq->idle; |
| 5279 | unsigned long flags; |
| 5280 | |
| 5281 | /* cpu has to be offline */ |
| 5282 | BUG_ON(cpu_online(this_cpu)); |
| 5283 | |
| 5284 | /* |
| 5285 | * Strictly not necessary since rest of the CPUs are stopped by now |
| 5286 | * and interrupts disabled on the current cpu. |
| 5287 | */ |
| 5288 | spin_lock_irqsave(&rq->lock, flags); |
| 5289 | |
| 5290 | __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1); |
| 5291 | |
| 5292 | update_rq_clock(rq); |
| 5293 | activate_task(rq, p, 0); |
| 5294 | |
| 5295 | spin_unlock_irqrestore(&rq->lock, flags); |
| 5296 | } |
| 5297 | |
| 5298 | /* |
| 5299 | * Ensures that the idle task is using init_mm right before its cpu goes |
| 5300 | * offline. |
| 5301 | */ |
| 5302 | void idle_task_exit(void) |
| 5303 | { |
| 5304 | struct mm_struct *mm = current->active_mm; |
| 5305 | |
| 5306 | BUG_ON(cpu_online(smp_processor_id())); |
| 5307 | |
| 5308 | if (mm != &init_mm) |
| 5309 | switch_mm(mm, &init_mm, current); |
| 5310 | mmdrop(mm); |
| 5311 | } |
| 5312 | |
| 5313 | /* called under rq->lock with disabled interrupts */ |
| 5314 | static void migrate_dead(unsigned int dead_cpu, struct task_struct *p) |
| 5315 | { |
| 5316 | struct rq *rq = cpu_rq(dead_cpu); |
| 5317 | |
| 5318 | /* Must be exiting, otherwise would be on tasklist. */ |
| 5319 | BUG_ON(!p->exit_state); |
| 5320 | |
| 5321 | /* Cannot have done final schedule yet: would have vanished. */ |
| 5322 | BUG_ON(p->state == TASK_DEAD); |
| 5323 | |
| 5324 | get_task_struct(p); |
| 5325 | |
| 5326 | /* |
| 5327 | * Drop lock around migration; if someone else moves it, |
| 5328 | * that's OK. No task can be added to this CPU, so iteration is |
| 5329 | * fine. |
| 5330 | */ |
| 5331 | spin_unlock_irq(&rq->lock); |
| 5332 | move_task_off_dead_cpu(dead_cpu, p); |
| 5333 | spin_lock_irq(&rq->lock); |
| 5334 | |
| 5335 | put_task_struct(p); |
| 5336 | } |
| 5337 | |
| 5338 | /* release_task() removes task from tasklist, so we won't find dead tasks. */ |
| 5339 | static void migrate_dead_tasks(unsigned int dead_cpu) |
| 5340 | { |
| 5341 | struct rq *rq = cpu_rq(dead_cpu); |
| 5342 | struct task_struct *next; |
| 5343 | |
| 5344 | for ( ; ; ) { |
| 5345 | if (!rq->nr_running) |
| 5346 | break; |
| 5347 | update_rq_clock(rq); |
| 5348 | next = pick_next_task(rq, rq->curr); |
| 5349 | if (!next) |
| 5350 | break; |
| 5351 | migrate_dead(dead_cpu, next); |
| 5352 | |
| 5353 | } |
| 5354 | } |
| 5355 | #endif /* CONFIG_HOTPLUG_CPU */ |
| 5356 | |
| 5357 | #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) |
| 5358 | |
| 5359 | static struct ctl_table sd_ctl_dir[] = { |
| 5360 | { |
| 5361 | .procname = "sched_domain", |
| 5362 | .mode = 0555, |
| 5363 | }, |
| 5364 | {0, }, |
| 5365 | }; |
| 5366 | |
| 5367 | static struct ctl_table sd_ctl_root[] = { |
| 5368 | { |
| 5369 | .ctl_name = CTL_KERN, |
| 5370 | .procname = "kernel", |
| 5371 | .mode = 0555, |
| 5372 | .child = sd_ctl_dir, |
| 5373 | }, |
| 5374 | {0, }, |
| 5375 | }; |
| 5376 | |
| 5377 | static struct ctl_table *sd_alloc_ctl_entry(int n) |
| 5378 | { |
| 5379 | struct ctl_table *entry = |
| 5380 | kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); |
| 5381 | |
| 5382 | return entry; |
| 5383 | } |
| 5384 | |
| 5385 | static void sd_free_ctl_entry(struct ctl_table **tablep) |
| 5386 | { |
| 5387 | struct ctl_table *entry; |
| 5388 | |
| 5389 | /* |
| 5390 | * In the intermediate directories, both the child directory and |
| 5391 | * procname are dynamically allocated and could fail but the mode |
| 5392 | * will always be set. In the lowest directory the names are |
| 5393 | * static strings and all have proc handlers. |
| 5394 | */ |
| 5395 | for (entry = *tablep; entry->mode; entry++) { |
| 5396 | if (entry->child) |
| 5397 | sd_free_ctl_entry(&entry->child); |
| 5398 | if (entry->proc_handler == NULL) |
| 5399 | kfree(entry->procname); |
| 5400 | } |
| 5401 | |
| 5402 | kfree(*tablep); |
| 5403 | *tablep = NULL; |
| 5404 | } |
| 5405 | |
| 5406 | static void |
| 5407 | set_table_entry(struct ctl_table *entry, |
| 5408 | const char *procname, void *data, int maxlen, |
| 5409 | mode_t mode, proc_handler *proc_handler) |
| 5410 | { |
| 5411 | entry->procname = procname; |
| 5412 | entry->data = data; |
| 5413 | entry->maxlen = maxlen; |
| 5414 | entry->mode = mode; |
| 5415 | entry->proc_handler = proc_handler; |
| 5416 | } |
| 5417 | |
| 5418 | static struct ctl_table * |
| 5419 | sd_alloc_ctl_domain_table(struct sched_domain *sd) |
| 5420 | { |
| 5421 | struct ctl_table *table = sd_alloc_ctl_entry(12); |
| 5422 | |
| 5423 | if (table == NULL) |
| 5424 | return NULL; |
| 5425 | |
| 5426 | set_table_entry(&table[0], "min_interval", &sd->min_interval, |
| 5427 | sizeof(long), 0644, proc_doulongvec_minmax); |
| 5428 | set_table_entry(&table[1], "max_interval", &sd->max_interval, |
| 5429 | sizeof(long), 0644, proc_doulongvec_minmax); |
| 5430 | set_table_entry(&table[2], "busy_idx", &sd->busy_idx, |
| 5431 | sizeof(int), 0644, proc_dointvec_minmax); |
| 5432 | set_table_entry(&table[3], "idle_idx", &sd->idle_idx, |
| 5433 | sizeof(int), 0644, proc_dointvec_minmax); |
| 5434 | set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, |
| 5435 | sizeof(int), 0644, proc_dointvec_minmax); |
| 5436 | set_table_entry(&table[5], "wake_idx", &sd->wake_idx, |
| 5437 | sizeof(int), 0644, proc_dointvec_minmax); |
| 5438 | set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, |
| 5439 | sizeof(int), 0644, proc_dointvec_minmax); |
| 5440 | set_table_entry(&table[7], "busy_factor", &sd->busy_factor, |
| 5441 | sizeof(int), 0644, proc_dointvec_minmax); |
| 5442 | set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, |
| 5443 | sizeof(int), 0644, proc_dointvec_minmax); |
| 5444 | set_table_entry(&table[9], "cache_nice_tries", |
| 5445 | &sd->cache_nice_tries, |
| 5446 | sizeof(int), 0644, proc_dointvec_minmax); |
| 5447 | set_table_entry(&table[10], "flags", &sd->flags, |
| 5448 | sizeof(int), 0644, proc_dointvec_minmax); |
| 5449 | /* &table[11] is terminator */ |
| 5450 | |
| 5451 | return table; |
| 5452 | } |
| 5453 | |
| 5454 | static ctl_table *sd_alloc_ctl_cpu_table(int cpu) |
| 5455 | { |
| 5456 | struct ctl_table *entry, *table; |
| 5457 | struct sched_domain *sd; |
| 5458 | int domain_num = 0, i; |
| 5459 | char buf[32]; |
| 5460 | |
| 5461 | for_each_domain(cpu, sd) |
| 5462 | domain_num++; |
| 5463 | entry = table = sd_alloc_ctl_entry(domain_num + 1); |
| 5464 | if (table == NULL) |
| 5465 | return NULL; |
| 5466 | |
| 5467 | i = 0; |
| 5468 | for_each_domain(cpu, sd) { |
| 5469 | snprintf(buf, 32, "domain%d", i); |
| 5470 | entry->procname = kstrdup(buf, GFP_KERNEL); |
| 5471 | entry->mode = 0555; |
| 5472 | entry->child = sd_alloc_ctl_domain_table(sd); |
| 5473 | entry++; |
| 5474 | i++; |
| 5475 | } |
| 5476 | return table; |
| 5477 | } |
| 5478 | |
| 5479 | static struct ctl_table_header *sd_sysctl_header; |
| 5480 | static void register_sched_domain_sysctl(void) |
| 5481 | { |
| 5482 | int i, cpu_num = num_online_cpus(); |
| 5483 | struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); |
| 5484 | char buf[32]; |
| 5485 | |
| 5486 | WARN_ON(sd_ctl_dir[0].child); |
| 5487 | sd_ctl_dir[0].child = entry; |
| 5488 | |
| 5489 | if (entry == NULL) |
| 5490 | return; |
| 5491 | |
| 5492 | for_each_online_cpu(i) { |
| 5493 | snprintf(buf, 32, "cpu%d", i); |
| 5494 | entry->procname = kstrdup(buf, GFP_KERNEL); |
| 5495 | entry->mode = 0555; |
| 5496 | entry->child = sd_alloc_ctl_cpu_table(i); |
| 5497 | entry++; |
| 5498 | } |
| 5499 | |
| 5500 | WARN_ON(sd_sysctl_header); |
| 5501 | sd_sysctl_header = register_sysctl_table(sd_ctl_root); |
| 5502 | } |
| 5503 | |
| 5504 | /* may be called multiple times per register */ |
| 5505 | static void unregister_sched_domain_sysctl(void) |
| 5506 | { |
| 5507 | if (sd_sysctl_header) |
| 5508 | unregister_sysctl_table(sd_sysctl_header); |
| 5509 | sd_sysctl_header = NULL; |
| 5510 | if (sd_ctl_dir[0].child) |
| 5511 | sd_free_ctl_entry(&sd_ctl_dir[0].child); |
| 5512 | } |
| 5513 | #else |
| 5514 | static void register_sched_domain_sysctl(void) |
| 5515 | { |
| 5516 | } |
| 5517 | static void unregister_sched_domain_sysctl(void) |
| 5518 | { |
| 5519 | } |
| 5520 | #endif |
| 5521 | |
| 5522 | /* |
| 5523 | * migration_call - callback that gets triggered when a CPU is added. |
| 5524 | * Here we can start up the necessary migration thread for the new CPU. |
| 5525 | */ |
| 5526 | static int __cpuinit |
| 5527 | migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) |
| 5528 | { |
| 5529 | struct task_struct *p; |
| 5530 | int cpu = (long)hcpu; |
| 5531 | unsigned long flags; |
| 5532 | struct rq *rq; |
| 5533 | |
| 5534 | switch (action) { |
| 5535 | |
| 5536 | case CPU_UP_PREPARE: |
| 5537 | case CPU_UP_PREPARE_FROZEN: |
| 5538 | p = kthread_create(migration_thread, hcpu, "migration/%d", cpu); |
| 5539 | if (IS_ERR(p)) |
| 5540 | return NOTIFY_BAD; |
| 5541 | kthread_bind(p, cpu); |
| 5542 | /* Must be high prio: stop_machine expects to yield to it. */ |
| 5543 | rq = task_rq_lock(p, &flags); |
| 5544 | __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1); |
| 5545 | task_rq_unlock(rq, &flags); |
| 5546 | cpu_rq(cpu)->migration_thread = p; |
| 5547 | break; |
| 5548 | |
| 5549 | case CPU_ONLINE: |
| 5550 | case CPU_ONLINE_FROZEN: |
| 5551 | /* Strictly unnecessary, as first user will wake it. */ |
| 5552 | wake_up_process(cpu_rq(cpu)->migration_thread); |
| 5553 | break; |
| 5554 | |
| 5555 | #ifdef CONFIG_HOTPLUG_CPU |
| 5556 | case CPU_UP_CANCELED: |
| 5557 | case CPU_UP_CANCELED_FROZEN: |
| 5558 | if (!cpu_rq(cpu)->migration_thread) |
| 5559 | break; |
| 5560 | /* Unbind it from offline cpu so it can run. Fall thru. */ |
| 5561 | kthread_bind(cpu_rq(cpu)->migration_thread, |
| 5562 | any_online_cpu(cpu_online_map)); |
| 5563 | kthread_stop(cpu_rq(cpu)->migration_thread); |
| 5564 | cpu_rq(cpu)->migration_thread = NULL; |
| 5565 | break; |
| 5566 | |
| 5567 | case CPU_DEAD: |
| 5568 | case CPU_DEAD_FROZEN: |
| 5569 | cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */ |
| 5570 | migrate_live_tasks(cpu); |
| 5571 | rq = cpu_rq(cpu); |
| 5572 | kthread_stop(rq->migration_thread); |
| 5573 | rq->migration_thread = NULL; |
| 5574 | /* Idle task back to normal (off runqueue, low prio) */ |
| 5575 | spin_lock_irq(&rq->lock); |
| 5576 | update_rq_clock(rq); |
| 5577 | deactivate_task(rq, rq->idle, 0); |
| 5578 | rq->idle->static_prio = MAX_PRIO; |
| 5579 | __setscheduler(rq, rq->idle, SCHED_NORMAL, 0); |
| 5580 | rq->idle->sched_class = &idle_sched_class; |
| 5581 | migrate_dead_tasks(cpu); |
| 5582 | spin_unlock_irq(&rq->lock); |
| 5583 | cpuset_unlock(); |
| 5584 | migrate_nr_uninterruptible(rq); |
| 5585 | BUG_ON(rq->nr_running != 0); |
| 5586 | |
| 5587 | /* |
| 5588 | * No need to migrate the tasks: it was best-effort if |
| 5589 | * they didn't take sched_hotcpu_mutex. Just wake up |
| 5590 | * the requestors. |
| 5591 | */ |
| 5592 | spin_lock_irq(&rq->lock); |
| 5593 | while (!list_empty(&rq->migration_queue)) { |
| 5594 | struct migration_req *req; |
| 5595 | |
| 5596 | req = list_entry(rq->migration_queue.next, |
| 5597 | struct migration_req, list); |
| 5598 | list_del_init(&req->list); |
| 5599 | complete(&req->done); |
| 5600 | } |
| 5601 | spin_unlock_irq(&rq->lock); |
| 5602 | break; |
| 5603 | #endif |
| 5604 | } |
| 5605 | return NOTIFY_OK; |
| 5606 | } |
| 5607 | |
| 5608 | /* Register at highest priority so that task migration (migrate_all_tasks) |
| 5609 | * happens before everything else. |
| 5610 | */ |
| 5611 | static struct notifier_block __cpuinitdata migration_notifier = { |
| 5612 | .notifier_call = migration_call, |
| 5613 | .priority = 10 |
| 5614 | }; |
| 5615 | |
| 5616 | void __init migration_init(void) |
| 5617 | { |
| 5618 | void *cpu = (void *)(long)smp_processor_id(); |
| 5619 | int err; |
| 5620 | |
| 5621 | /* Start one for the boot CPU: */ |
| 5622 | err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); |
| 5623 | BUG_ON(err == NOTIFY_BAD); |
| 5624 | migration_call(&migration_notifier, CPU_ONLINE, cpu); |
| 5625 | register_cpu_notifier(&migration_notifier); |
| 5626 | } |
| 5627 | #endif |
| 5628 | |
| 5629 | #ifdef CONFIG_SMP |
| 5630 | |
| 5631 | /* Number of possible processor ids */ |
| 5632 | int nr_cpu_ids __read_mostly = NR_CPUS; |
| 5633 | EXPORT_SYMBOL(nr_cpu_ids); |
| 5634 | |
| 5635 | #ifdef CONFIG_SCHED_DEBUG |
| 5636 | |
| 5637 | static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level) |
| 5638 | { |
| 5639 | struct sched_group *group = sd->groups; |
| 5640 | cpumask_t groupmask; |
| 5641 | char str[NR_CPUS]; |
| 5642 | |
| 5643 | cpumask_scnprintf(str, NR_CPUS, sd->span); |
| 5644 | cpus_clear(groupmask); |
| 5645 | |
| 5646 | printk(KERN_DEBUG "%*s domain %d: ", level, "", level); |
| 5647 | |
| 5648 | if (!(sd->flags & SD_LOAD_BALANCE)) { |
| 5649 | printk("does not load-balance\n"); |
| 5650 | if (sd->parent) |
| 5651 | printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" |
| 5652 | " has parent"); |
| 5653 | return -1; |
| 5654 | } |
| 5655 | |
| 5656 | printk(KERN_CONT "span %s\n", str); |
| 5657 | |
| 5658 | if (!cpu_isset(cpu, sd->span)) { |
| 5659 | printk(KERN_ERR "ERROR: domain->span does not contain " |
| 5660 | "CPU%d\n", cpu); |
| 5661 | } |
| 5662 | if (!cpu_isset(cpu, group->cpumask)) { |
| 5663 | printk(KERN_ERR "ERROR: domain->groups does not contain" |
| 5664 | " CPU%d\n", cpu); |
| 5665 | } |
| 5666 | |
| 5667 | printk(KERN_DEBUG "%*s groups:", level + 1, ""); |
| 5668 | do { |
| 5669 | if (!group) { |
| 5670 | printk("\n"); |
| 5671 | printk(KERN_ERR "ERROR: group is NULL\n"); |
| 5672 | break; |
| 5673 | } |
| 5674 | |
| 5675 | if (!group->__cpu_power) { |
| 5676 | printk(KERN_CONT "\n"); |
| 5677 | printk(KERN_ERR "ERROR: domain->cpu_power not " |
| 5678 | "set\n"); |
| 5679 | break; |
| 5680 | } |
| 5681 | |
| 5682 | if (!cpus_weight(group->cpumask)) { |
| 5683 | printk(KERN_CONT "\n"); |
| 5684 | printk(KERN_ERR "ERROR: empty group\n"); |
| 5685 | break; |
| 5686 | } |
| 5687 | |
| 5688 | if (cpus_intersects(groupmask, group->cpumask)) { |
| 5689 | printk(KERN_CONT "\n"); |
| 5690 | printk(KERN_ERR "ERROR: repeated CPUs\n"); |
| 5691 | break; |
| 5692 | } |
| 5693 | |
| 5694 | cpus_or(groupmask, groupmask, group->cpumask); |
| 5695 | |
| 5696 | cpumask_scnprintf(str, NR_CPUS, group->cpumask); |
| 5697 | printk(KERN_CONT " %s", str); |
| 5698 | |
| 5699 | group = group->next; |
| 5700 | } while (group != sd->groups); |
| 5701 | printk(KERN_CONT "\n"); |
| 5702 | |
| 5703 | if (!cpus_equal(sd->span, groupmask)) |
| 5704 | printk(KERN_ERR "ERROR: groups don't span domain->span\n"); |
| 5705 | |
| 5706 | if (sd->parent && !cpus_subset(groupmask, sd->parent->span)) |
| 5707 | printk(KERN_ERR "ERROR: parent span is not a superset " |
| 5708 | "of domain->span\n"); |
| 5709 | return 0; |
| 5710 | } |
| 5711 | |
| 5712 | static void sched_domain_debug(struct sched_domain *sd, int cpu) |
| 5713 | { |
| 5714 | int level = 0; |
| 5715 | |
| 5716 | if (!sd) { |
| 5717 | printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); |
| 5718 | return; |
| 5719 | } |
| 5720 | |
| 5721 | printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); |
| 5722 | |
| 5723 | for (;;) { |
| 5724 | if (sched_domain_debug_one(sd, cpu, level)) |
| 5725 | break; |
| 5726 | level++; |
| 5727 | sd = sd->parent; |
| 5728 | if (!sd) |
| 5729 | break; |
| 5730 | } |
| 5731 | } |
| 5732 | #else |
| 5733 | # define sched_domain_debug(sd, cpu) do { } while (0) |
| 5734 | #endif |
| 5735 | |
| 5736 | static int sd_degenerate(struct sched_domain *sd) |
| 5737 | { |
| 5738 | if (cpus_weight(sd->span) == 1) |
| 5739 | return 1; |
| 5740 | |
| 5741 | /* Following flags need at least 2 groups */ |
| 5742 | if (sd->flags & (SD_LOAD_BALANCE | |
| 5743 | SD_BALANCE_NEWIDLE | |
| 5744 | SD_BALANCE_FORK | |
| 5745 | SD_BALANCE_EXEC | |
| 5746 | SD_SHARE_CPUPOWER | |
| 5747 | SD_SHARE_PKG_RESOURCES)) { |
| 5748 | if (sd->groups != sd->groups->next) |
| 5749 | return 0; |
| 5750 | } |
| 5751 | |
| 5752 | /* Following flags don't use groups */ |
| 5753 | if (sd->flags & (SD_WAKE_IDLE | |
| 5754 | SD_WAKE_AFFINE | |
| 5755 | SD_WAKE_BALANCE)) |
| 5756 | return 0; |
| 5757 | |
| 5758 | return 1; |
| 5759 | } |
| 5760 | |
| 5761 | static int |
| 5762 | sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) |
| 5763 | { |
| 5764 | unsigned long cflags = sd->flags, pflags = parent->flags; |
| 5765 | |
| 5766 | if (sd_degenerate(parent)) |
| 5767 | return 1; |
| 5768 | |
| 5769 | if (!cpus_equal(sd->span, parent->span)) |
| 5770 | return 0; |
| 5771 | |
| 5772 | /* Does parent contain flags not in child? */ |
| 5773 | /* WAKE_BALANCE is a subset of WAKE_AFFINE */ |
| 5774 | if (cflags & SD_WAKE_AFFINE) |
| 5775 | pflags &= ~SD_WAKE_BALANCE; |
| 5776 | /* Flags needing groups don't count if only 1 group in parent */ |
| 5777 | if (parent->groups == parent->groups->next) { |
| 5778 | pflags &= ~(SD_LOAD_BALANCE | |
| 5779 | SD_BALANCE_NEWIDLE | |
| 5780 | SD_BALANCE_FORK | |
| 5781 | SD_BALANCE_EXEC | |
| 5782 | SD_SHARE_CPUPOWER | |
| 5783 | SD_SHARE_PKG_RESOURCES); |
| 5784 | } |
| 5785 | if (~cflags & pflags) |
| 5786 | return 0; |
| 5787 | |
| 5788 | return 1; |
| 5789 | } |
| 5790 | |
| 5791 | /* |
| 5792 | * Attach the domain 'sd' to 'cpu' as its base domain. Callers must |
| 5793 | * hold the hotplug lock. |
| 5794 | */ |
| 5795 | static void cpu_attach_domain(struct sched_domain *sd, int cpu) |
| 5796 | { |
| 5797 | struct rq *rq = cpu_rq(cpu); |
| 5798 | struct sched_domain *tmp; |
| 5799 | |
| 5800 | /* Remove the sched domains which do not contribute to scheduling. */ |
| 5801 | for (tmp = sd; tmp; tmp = tmp->parent) { |
| 5802 | struct sched_domain *parent = tmp->parent; |
| 5803 | if (!parent) |
| 5804 | break; |
| 5805 | if (sd_parent_degenerate(tmp, parent)) { |
| 5806 | tmp->parent = parent->parent; |
| 5807 | if (parent->parent) |
| 5808 | parent->parent->child = tmp; |
| 5809 | } |
| 5810 | } |
| 5811 | |
| 5812 | if (sd && sd_degenerate(sd)) { |
| 5813 | sd = sd->parent; |
| 5814 | if (sd) |
| 5815 | sd->child = NULL; |
| 5816 | } |
| 5817 | |
| 5818 | sched_domain_debug(sd, cpu); |
| 5819 | |
| 5820 | rcu_assign_pointer(rq->sd, sd); |
| 5821 | } |
| 5822 | |
| 5823 | /* cpus with isolated domains */ |
| 5824 | static cpumask_t cpu_isolated_map = CPU_MASK_NONE; |
| 5825 | |
| 5826 | /* Setup the mask of cpus configured for isolated domains */ |
| 5827 | static int __init isolated_cpu_setup(char *str) |
| 5828 | { |
| 5829 | int ints[NR_CPUS], i; |
| 5830 | |
| 5831 | str = get_options(str, ARRAY_SIZE(ints), ints); |
| 5832 | cpus_clear(cpu_isolated_map); |
| 5833 | for (i = 1; i <= ints[0]; i++) |
| 5834 | if (ints[i] < NR_CPUS) |
| 5835 | cpu_set(ints[i], cpu_isolated_map); |
| 5836 | return 1; |
| 5837 | } |
| 5838 | |
| 5839 | __setup("isolcpus=", isolated_cpu_setup); |
| 5840 | |
| 5841 | /* |
| 5842 | * init_sched_build_groups takes the cpumask we wish to span, and a pointer |
| 5843 | * to a function which identifies what group(along with sched group) a CPU |
| 5844 | * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS |
| 5845 | * (due to the fact that we keep track of groups covered with a cpumask_t). |
| 5846 | * |
| 5847 | * init_sched_build_groups will build a circular linked list of the groups |
| 5848 | * covered by the given span, and will set each group's ->cpumask correctly, |
| 5849 | * and ->cpu_power to 0. |
| 5850 | */ |
| 5851 | static void |
| 5852 | init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map, |
| 5853 | int (*group_fn)(int cpu, const cpumask_t *cpu_map, |
| 5854 | struct sched_group **sg)) |
| 5855 | { |
| 5856 | struct sched_group *first = NULL, *last = NULL; |
| 5857 | cpumask_t covered = CPU_MASK_NONE; |
| 5858 | int i; |
| 5859 | |
| 5860 | for_each_cpu_mask(i, span) { |
| 5861 | struct sched_group *sg; |
| 5862 | int group = group_fn(i, cpu_map, &sg); |
| 5863 | int j; |
| 5864 | |
| 5865 | if (cpu_isset(i, covered)) |
| 5866 | continue; |
| 5867 | |
| 5868 | sg->cpumask = CPU_MASK_NONE; |
| 5869 | sg->__cpu_power = 0; |
| 5870 | |
| 5871 | for_each_cpu_mask(j, span) { |
| 5872 | if (group_fn(j, cpu_map, NULL) != group) |
| 5873 | continue; |
| 5874 | |
| 5875 | cpu_set(j, covered); |
| 5876 | cpu_set(j, sg->cpumask); |
| 5877 | } |
| 5878 | if (!first) |
| 5879 | first = sg; |
| 5880 | if (last) |
| 5881 | last->next = sg; |
| 5882 | last = sg; |
| 5883 | } |
| 5884 | last->next = first; |
| 5885 | } |
| 5886 | |
| 5887 | #define SD_NODES_PER_DOMAIN 16 |
| 5888 | |
| 5889 | #ifdef CONFIG_NUMA |
| 5890 | |
| 5891 | /** |
| 5892 | * find_next_best_node - find the next node to include in a sched_domain |
| 5893 | * @node: node whose sched_domain we're building |
| 5894 | * @used_nodes: nodes already in the sched_domain |
| 5895 | * |
| 5896 | * Find the next node to include in a given scheduling domain. Simply |
| 5897 | * finds the closest node not already in the @used_nodes map. |
| 5898 | * |
| 5899 | * Should use nodemask_t. |
| 5900 | */ |
| 5901 | static int find_next_best_node(int node, unsigned long *used_nodes) |
| 5902 | { |
| 5903 | int i, n, val, min_val, best_node = 0; |
| 5904 | |
| 5905 | min_val = INT_MAX; |
| 5906 | |
| 5907 | for (i = 0; i < MAX_NUMNODES; i++) { |
| 5908 | /* Start at @node */ |
| 5909 | n = (node + i) % MAX_NUMNODES; |
| 5910 | |
| 5911 | if (!nr_cpus_node(n)) |
| 5912 | continue; |
| 5913 | |
| 5914 | /* Skip already used nodes */ |
| 5915 | if (test_bit(n, used_nodes)) |
| 5916 | continue; |
| 5917 | |
| 5918 | /* Simple min distance search */ |
| 5919 | val = node_distance(node, n); |
| 5920 | |
| 5921 | if (val < min_val) { |
| 5922 | min_val = val; |
| 5923 | best_node = n; |
| 5924 | } |
| 5925 | } |
| 5926 | |
| 5927 | set_bit(best_node, used_nodes); |
| 5928 | return best_node; |
| 5929 | } |
| 5930 | |
| 5931 | /** |
| 5932 | * sched_domain_node_span - get a cpumask for a node's sched_domain |
| 5933 | * @node: node whose cpumask we're constructing |
| 5934 | * @size: number of nodes to include in this span |
| 5935 | * |
| 5936 | * Given a node, construct a good cpumask for its sched_domain to span. It |
| 5937 | * should be one that prevents unnecessary balancing, but also spreads tasks |
| 5938 | * out optimally. |
| 5939 | */ |
| 5940 | static cpumask_t sched_domain_node_span(int node) |
| 5941 | { |
| 5942 | DECLARE_BITMAP(used_nodes, MAX_NUMNODES); |
| 5943 | cpumask_t span, nodemask; |
| 5944 | int i; |
| 5945 | |
| 5946 | cpus_clear(span); |
| 5947 | bitmap_zero(used_nodes, MAX_NUMNODES); |
| 5948 | |
| 5949 | nodemask = node_to_cpumask(node); |
| 5950 | cpus_or(span, span, nodemask); |
| 5951 | set_bit(node, used_nodes); |
| 5952 | |
| 5953 | for (i = 1; i < SD_NODES_PER_DOMAIN; i++) { |
| 5954 | int next_node = find_next_best_node(node, used_nodes); |
| 5955 | |
| 5956 | nodemask = node_to_cpumask(next_node); |
| 5957 | cpus_or(span, span, nodemask); |
| 5958 | } |
| 5959 | |
| 5960 | return span; |
| 5961 | } |
| 5962 | #endif |
| 5963 | |
| 5964 | int sched_smt_power_savings = 0, sched_mc_power_savings = 0; |
| 5965 | |
| 5966 | /* |
| 5967 | * SMT sched-domains: |
| 5968 | */ |
| 5969 | #ifdef CONFIG_SCHED_SMT |
| 5970 | static DEFINE_PER_CPU(struct sched_domain, cpu_domains); |
| 5971 | static DEFINE_PER_CPU(struct sched_group, sched_group_cpus); |
| 5972 | |
| 5973 | static int |
| 5974 | cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg) |
| 5975 | { |
| 5976 | if (sg) |
| 5977 | *sg = &per_cpu(sched_group_cpus, cpu); |
| 5978 | return cpu; |
| 5979 | } |
| 5980 | #endif |
| 5981 | |
| 5982 | /* |
| 5983 | * multi-core sched-domains: |
| 5984 | */ |
| 5985 | #ifdef CONFIG_SCHED_MC |
| 5986 | static DEFINE_PER_CPU(struct sched_domain, core_domains); |
| 5987 | static DEFINE_PER_CPU(struct sched_group, sched_group_core); |
| 5988 | #endif |
| 5989 | |
| 5990 | #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT) |
| 5991 | static int |
| 5992 | cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg) |
| 5993 | { |
| 5994 | int group; |
| 5995 | cpumask_t mask = per_cpu(cpu_sibling_map, cpu); |
| 5996 | cpus_and(mask, mask, *cpu_map); |
| 5997 | group = first_cpu(mask); |
| 5998 | if (sg) |
| 5999 | *sg = &per_cpu(sched_group_core, group); |
| 6000 | return group; |
| 6001 | } |
| 6002 | #elif defined(CONFIG_SCHED_MC) |
| 6003 | static int |
| 6004 | cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg) |
| 6005 | { |
| 6006 | if (sg) |
| 6007 | *sg = &per_cpu(sched_group_core, cpu); |
| 6008 | return cpu; |
| 6009 | } |
| 6010 | #endif |
| 6011 | |
| 6012 | static DEFINE_PER_CPU(struct sched_domain, phys_domains); |
| 6013 | static DEFINE_PER_CPU(struct sched_group, sched_group_phys); |
| 6014 | |
| 6015 | static int |
| 6016 | cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg) |
| 6017 | { |
| 6018 | int group; |
| 6019 | #ifdef CONFIG_SCHED_MC |
| 6020 | cpumask_t mask = cpu_coregroup_map(cpu); |
| 6021 | cpus_and(mask, mask, *cpu_map); |
| 6022 | group = first_cpu(mask); |
| 6023 | #elif defined(CONFIG_SCHED_SMT) |
| 6024 | cpumask_t mask = per_cpu(cpu_sibling_map, cpu); |
| 6025 | cpus_and(mask, mask, *cpu_map); |
| 6026 | group = first_cpu(mask); |
| 6027 | #else |
| 6028 | group = cpu; |
| 6029 | #endif |
| 6030 | if (sg) |
| 6031 | *sg = &per_cpu(sched_group_phys, group); |
| 6032 | return group; |
| 6033 | } |
| 6034 | |
| 6035 | #ifdef CONFIG_NUMA |
| 6036 | /* |
| 6037 | * The init_sched_build_groups can't handle what we want to do with node |
| 6038 | * groups, so roll our own. Now each node has its own list of groups which |
| 6039 | * gets dynamically allocated. |
| 6040 | */ |
| 6041 | static DEFINE_PER_CPU(struct sched_domain, node_domains); |
| 6042 | static struct sched_group **sched_group_nodes_bycpu[NR_CPUS]; |
| 6043 | |
| 6044 | static DEFINE_PER_CPU(struct sched_domain, allnodes_domains); |
| 6045 | static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes); |
| 6046 | |
| 6047 | static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map, |
| 6048 | struct sched_group **sg) |
| 6049 | { |
| 6050 | cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu)); |
| 6051 | int group; |
| 6052 | |
| 6053 | cpus_and(nodemask, nodemask, *cpu_map); |
| 6054 | group = first_cpu(nodemask); |
| 6055 | |
| 6056 | if (sg) |
| 6057 | *sg = &per_cpu(sched_group_allnodes, group); |
| 6058 | return group; |
| 6059 | } |
| 6060 | |
| 6061 | static void init_numa_sched_groups_power(struct sched_group *group_head) |
| 6062 | { |
| 6063 | struct sched_group *sg = group_head; |
| 6064 | int j; |
| 6065 | |
| 6066 | if (!sg) |
| 6067 | return; |
| 6068 | do { |
| 6069 | for_each_cpu_mask(j, sg->cpumask) { |
| 6070 | struct sched_domain *sd; |
| 6071 | |
| 6072 | sd = &per_cpu(phys_domains, j); |
| 6073 | if (j != first_cpu(sd->groups->cpumask)) { |
| 6074 | /* |
| 6075 | * Only add "power" once for each |
| 6076 | * physical package. |
| 6077 | */ |
| 6078 | continue; |
| 6079 | } |
| 6080 | |
| 6081 | sg_inc_cpu_power(sg, sd->groups->__cpu_power); |
| 6082 | } |
| 6083 | sg = sg->next; |
| 6084 | } while (sg != group_head); |
| 6085 | } |
| 6086 | #endif |
| 6087 | |
| 6088 | #ifdef CONFIG_NUMA |
| 6089 | /* Free memory allocated for various sched_group structures */ |
| 6090 | static void free_sched_groups(const cpumask_t *cpu_map) |
| 6091 | { |
| 6092 | int cpu, i; |
| 6093 | |
| 6094 | for_each_cpu_mask(cpu, *cpu_map) { |
| 6095 | struct sched_group **sched_group_nodes |
| 6096 | = sched_group_nodes_bycpu[cpu]; |
| 6097 | |
| 6098 | if (!sched_group_nodes) |
| 6099 | continue; |
| 6100 | |
| 6101 | for (i = 0; i < MAX_NUMNODES; i++) { |
| 6102 | cpumask_t nodemask = node_to_cpumask(i); |
| 6103 | struct sched_group *oldsg, *sg = sched_group_nodes[i]; |
| 6104 | |
| 6105 | cpus_and(nodemask, nodemask, *cpu_map); |
| 6106 | if (cpus_empty(nodemask)) |
| 6107 | continue; |
| 6108 | |
| 6109 | if (sg == NULL) |
| 6110 | continue; |
| 6111 | sg = sg->next; |
| 6112 | next_sg: |
| 6113 | oldsg = sg; |
| 6114 | sg = sg->next; |
| 6115 | kfree(oldsg); |
| 6116 | if (oldsg != sched_group_nodes[i]) |
| 6117 | goto next_sg; |
| 6118 | } |
| 6119 | kfree(sched_group_nodes); |
| 6120 | sched_group_nodes_bycpu[cpu] = NULL; |
| 6121 | } |
| 6122 | } |
| 6123 | #else |
| 6124 | static void free_sched_groups(const cpumask_t *cpu_map) |
| 6125 | { |
| 6126 | } |
| 6127 | #endif |
| 6128 | |
| 6129 | /* |
| 6130 | * Initialize sched groups cpu_power. |
| 6131 | * |
| 6132 | * cpu_power indicates the capacity of sched group, which is used while |
| 6133 | * distributing the load between different sched groups in a sched domain. |
| 6134 | * Typically cpu_power for all the groups in a sched domain will be same unless |
| 6135 | * there are asymmetries in the topology. If there are asymmetries, group |
| 6136 | * having more cpu_power will pickup more load compared to the group having |
| 6137 | * less cpu_power. |
| 6138 | * |
| 6139 | * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents |
| 6140 | * the maximum number of tasks a group can handle in the presence of other idle |
| 6141 | * or lightly loaded groups in the same sched domain. |
| 6142 | */ |
| 6143 | static void init_sched_groups_power(int cpu, struct sched_domain *sd) |
| 6144 | { |
| 6145 | struct sched_domain *child; |
| 6146 | struct sched_group *group; |
| 6147 | |
| 6148 | WARN_ON(!sd || !sd->groups); |
| 6149 | |
| 6150 | if (cpu != first_cpu(sd->groups->cpumask)) |
| 6151 | return; |
| 6152 | |
| 6153 | child = sd->child; |
| 6154 | |
| 6155 | sd->groups->__cpu_power = 0; |
| 6156 | |
| 6157 | /* |
| 6158 | * For perf policy, if the groups in child domain share resources |
| 6159 | * (for example cores sharing some portions of the cache hierarchy |
| 6160 | * or SMT), then set this domain groups cpu_power such that each group |
| 6161 | * can handle only one task, when there are other idle groups in the |
| 6162 | * same sched domain. |
| 6163 | */ |
| 6164 | if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) && |
| 6165 | (child->flags & |
| 6166 | (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) { |
| 6167 | sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE); |
| 6168 | return; |
| 6169 | } |
| 6170 | |
| 6171 | /* |
| 6172 | * add cpu_power of each child group to this groups cpu_power |
| 6173 | */ |
| 6174 | group = child->groups; |
| 6175 | do { |
| 6176 | sg_inc_cpu_power(sd->groups, group->__cpu_power); |
| 6177 | group = group->next; |
| 6178 | } while (group != child->groups); |
| 6179 | } |
| 6180 | |
| 6181 | /* |
| 6182 | * Build sched domains for a given set of cpus and attach the sched domains |
| 6183 | * to the individual cpus |
| 6184 | */ |
| 6185 | static int build_sched_domains(const cpumask_t *cpu_map) |
| 6186 | { |
| 6187 | int i; |
| 6188 | #ifdef CONFIG_NUMA |
| 6189 | struct sched_group **sched_group_nodes = NULL; |
| 6190 | int sd_allnodes = 0; |
| 6191 | |
| 6192 | /* |
| 6193 | * Allocate the per-node list of sched groups |
| 6194 | */ |
| 6195 | sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *), |
| 6196 | GFP_KERNEL); |
| 6197 | if (!sched_group_nodes) { |
| 6198 | printk(KERN_WARNING "Can not alloc sched group node list\n"); |
| 6199 | return -ENOMEM; |
| 6200 | } |
| 6201 | sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes; |
| 6202 | #endif |
| 6203 | |
| 6204 | /* |
| 6205 | * Set up domains for cpus specified by the cpu_map. |
| 6206 | */ |
| 6207 | for_each_cpu_mask(i, *cpu_map) { |
| 6208 | struct sched_domain *sd = NULL, *p; |
| 6209 | cpumask_t nodemask = node_to_cpumask(cpu_to_node(i)); |
| 6210 | |
| 6211 | cpus_and(nodemask, nodemask, *cpu_map); |
| 6212 | |
| 6213 | #ifdef CONFIG_NUMA |
| 6214 | if (cpus_weight(*cpu_map) > |
| 6215 | SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) { |
| 6216 | sd = &per_cpu(allnodes_domains, i); |
| 6217 | *sd = SD_ALLNODES_INIT; |
| 6218 | sd->span = *cpu_map; |
| 6219 | cpu_to_allnodes_group(i, cpu_map, &sd->groups); |
| 6220 | p = sd; |
| 6221 | sd_allnodes = 1; |
| 6222 | } else |
| 6223 | p = NULL; |
| 6224 | |
| 6225 | sd = &per_cpu(node_domains, i); |
| 6226 | *sd = SD_NODE_INIT; |
| 6227 | sd->span = sched_domain_node_span(cpu_to_node(i)); |
| 6228 | sd->parent = p; |
| 6229 | if (p) |
| 6230 | p->child = sd; |
| 6231 | cpus_and(sd->span, sd->span, *cpu_map); |
| 6232 | #endif |
| 6233 | |
| 6234 | p = sd; |
| 6235 | sd = &per_cpu(phys_domains, i); |
| 6236 | *sd = SD_CPU_INIT; |
| 6237 | sd->span = nodemask; |
| 6238 | sd->parent = p; |
| 6239 | if (p) |
| 6240 | p->child = sd; |
| 6241 | cpu_to_phys_group(i, cpu_map, &sd->groups); |
| 6242 | |
| 6243 | #ifdef CONFIG_SCHED_MC |
| 6244 | p = sd; |
| 6245 | sd = &per_cpu(core_domains, i); |
| 6246 | *sd = SD_MC_INIT; |
| 6247 | sd->span = cpu_coregroup_map(i); |
| 6248 | cpus_and(sd->span, sd->span, *cpu_map); |
| 6249 | sd->parent = p; |
| 6250 | p->child = sd; |
| 6251 | cpu_to_core_group(i, cpu_map, &sd->groups); |
| 6252 | #endif |
| 6253 | |
| 6254 | #ifdef CONFIG_SCHED_SMT |
| 6255 | p = sd; |
| 6256 | sd = &per_cpu(cpu_domains, i); |
| 6257 | *sd = SD_SIBLING_INIT; |
| 6258 | sd->span = per_cpu(cpu_sibling_map, i); |
| 6259 | cpus_and(sd->span, sd->span, *cpu_map); |
| 6260 | sd->parent = p; |
| 6261 | p->child = sd; |
| 6262 | cpu_to_cpu_group(i, cpu_map, &sd->groups); |
| 6263 | #endif |
| 6264 | } |
| 6265 | |
| 6266 | #ifdef CONFIG_SCHED_SMT |
| 6267 | /* Set up CPU (sibling) groups */ |
| 6268 | for_each_cpu_mask(i, *cpu_map) { |
| 6269 | cpumask_t this_sibling_map = per_cpu(cpu_sibling_map, i); |
| 6270 | cpus_and(this_sibling_map, this_sibling_map, *cpu_map); |
| 6271 | if (i != first_cpu(this_sibling_map)) |
| 6272 | continue; |
| 6273 | |
| 6274 | init_sched_build_groups(this_sibling_map, cpu_map, |
| 6275 | &cpu_to_cpu_group); |
| 6276 | } |
| 6277 | #endif |
| 6278 | |
| 6279 | #ifdef CONFIG_SCHED_MC |
| 6280 | /* Set up multi-core groups */ |
| 6281 | for_each_cpu_mask(i, *cpu_map) { |
| 6282 | cpumask_t this_core_map = cpu_coregroup_map(i); |
| 6283 | cpus_and(this_core_map, this_core_map, *cpu_map); |
| 6284 | if (i != first_cpu(this_core_map)) |
| 6285 | continue; |
| 6286 | init_sched_build_groups(this_core_map, cpu_map, |
| 6287 | &cpu_to_core_group); |
| 6288 | } |
| 6289 | #endif |
| 6290 | |
| 6291 | /* Set up physical groups */ |
| 6292 | for (i = 0; i < MAX_NUMNODES; i++) { |
| 6293 | cpumask_t nodemask = node_to_cpumask(i); |
| 6294 | |
| 6295 | cpus_and(nodemask, nodemask, *cpu_map); |
| 6296 | if (cpus_empty(nodemask)) |
| 6297 | continue; |
| 6298 | |
| 6299 | init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group); |
| 6300 | } |
| 6301 | |
| 6302 | #ifdef CONFIG_NUMA |
| 6303 | /* Set up node groups */ |
| 6304 | if (sd_allnodes) |
| 6305 | init_sched_build_groups(*cpu_map, cpu_map, |
| 6306 | &cpu_to_allnodes_group); |
| 6307 | |
| 6308 | for (i = 0; i < MAX_NUMNODES; i++) { |
| 6309 | /* Set up node groups */ |
| 6310 | struct sched_group *sg, *prev; |
| 6311 | cpumask_t nodemask = node_to_cpumask(i); |
| 6312 | cpumask_t domainspan; |
| 6313 | cpumask_t covered = CPU_MASK_NONE; |
| 6314 | int j; |
| 6315 | |
| 6316 | cpus_and(nodemask, nodemask, *cpu_map); |
| 6317 | if (cpus_empty(nodemask)) { |
| 6318 | sched_group_nodes[i] = NULL; |
| 6319 | continue; |
| 6320 | } |
| 6321 | |
| 6322 | domainspan = sched_domain_node_span(i); |
| 6323 | cpus_and(domainspan, domainspan, *cpu_map); |
| 6324 | |
| 6325 | sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i); |
| 6326 | if (!sg) { |
| 6327 | printk(KERN_WARNING "Can not alloc domain group for " |
| 6328 | "node %d\n", i); |
| 6329 | goto error; |
| 6330 | } |
| 6331 | sched_group_nodes[i] = sg; |
| 6332 | for_each_cpu_mask(j, nodemask) { |
| 6333 | struct sched_domain *sd; |
| 6334 | |
| 6335 | sd = &per_cpu(node_domains, j); |
| 6336 | sd->groups = sg; |
| 6337 | } |
| 6338 | sg->__cpu_power = 0; |
| 6339 | sg->cpumask = nodemask; |
| 6340 | sg->next = sg; |
| 6341 | cpus_or(covered, covered, nodemask); |
| 6342 | prev = sg; |
| 6343 | |
| 6344 | for (j = 0; j < MAX_NUMNODES; j++) { |
| 6345 | cpumask_t tmp, notcovered; |
| 6346 | int n = (i + j) % MAX_NUMNODES; |
| 6347 | |
| 6348 | cpus_complement(notcovered, covered); |
| 6349 | cpus_and(tmp, notcovered, *cpu_map); |
| 6350 | cpus_and(tmp, tmp, domainspan); |
| 6351 | if (cpus_empty(tmp)) |
| 6352 | break; |
| 6353 | |
| 6354 | nodemask = node_to_cpumask(n); |
| 6355 | cpus_and(tmp, tmp, nodemask); |
| 6356 | if (cpus_empty(tmp)) |
| 6357 | continue; |
| 6358 | |
| 6359 | sg = kmalloc_node(sizeof(struct sched_group), |
| 6360 | GFP_KERNEL, i); |
| 6361 | if (!sg) { |
| 6362 | printk(KERN_WARNING |
| 6363 | "Can not alloc domain group for node %d\n", j); |
| 6364 | goto error; |
| 6365 | } |
| 6366 | sg->__cpu_power = 0; |
| 6367 | sg->cpumask = tmp; |
| 6368 | sg->next = prev->next; |
| 6369 | cpus_or(covered, covered, tmp); |
| 6370 | prev->next = sg; |
| 6371 | prev = sg; |
| 6372 | } |
| 6373 | } |
| 6374 | #endif |
| 6375 | |
| 6376 | /* Calculate CPU power for physical packages and nodes */ |
| 6377 | #ifdef CONFIG_SCHED_SMT |
| 6378 | for_each_cpu_mask(i, *cpu_map) { |
| 6379 | struct sched_domain *sd = &per_cpu(cpu_domains, i); |
| 6380 | |
| 6381 | init_sched_groups_power(i, sd); |
| 6382 | } |
| 6383 | #endif |
| 6384 | #ifdef CONFIG_SCHED_MC |
| 6385 | for_each_cpu_mask(i, *cpu_map) { |
| 6386 | struct sched_domain *sd = &per_cpu(core_domains, i); |
| 6387 | |
| 6388 | init_sched_groups_power(i, sd); |
| 6389 | } |
| 6390 | #endif |
| 6391 | |
| 6392 | for_each_cpu_mask(i, *cpu_map) { |
| 6393 | struct sched_domain *sd = &per_cpu(phys_domains, i); |
| 6394 | |
| 6395 | init_sched_groups_power(i, sd); |
| 6396 | } |
| 6397 | |
| 6398 | #ifdef CONFIG_NUMA |
| 6399 | for (i = 0; i < MAX_NUMNODES; i++) |
| 6400 | init_numa_sched_groups_power(sched_group_nodes[i]); |
| 6401 | |
| 6402 | if (sd_allnodes) { |
| 6403 | struct sched_group *sg; |
| 6404 | |
| 6405 | cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg); |
| 6406 | init_numa_sched_groups_power(sg); |
| 6407 | } |
| 6408 | #endif |
| 6409 | |
| 6410 | /* Attach the domains */ |
| 6411 | for_each_cpu_mask(i, *cpu_map) { |
| 6412 | struct sched_domain *sd; |
| 6413 | #ifdef CONFIG_SCHED_SMT |
| 6414 | sd = &per_cpu(cpu_domains, i); |
| 6415 | #elif defined(CONFIG_SCHED_MC) |
| 6416 | sd = &per_cpu(core_domains, i); |
| 6417 | #else |
| 6418 | sd = &per_cpu(phys_domains, i); |
| 6419 | #endif |
| 6420 | cpu_attach_domain(sd, i); |
| 6421 | } |
| 6422 | |
| 6423 | return 0; |
| 6424 | |
| 6425 | #ifdef CONFIG_NUMA |
| 6426 | error: |
| 6427 | free_sched_groups(cpu_map); |
| 6428 | return -ENOMEM; |
| 6429 | #endif |
| 6430 | } |
| 6431 | |
| 6432 | static cpumask_t *doms_cur; /* current sched domains */ |
| 6433 | static int ndoms_cur; /* number of sched domains in 'doms_cur' */ |
| 6434 | |
| 6435 | /* |
| 6436 | * Special case: If a kmalloc of a doms_cur partition (array of |
| 6437 | * cpumask_t) fails, then fallback to a single sched domain, |
| 6438 | * as determined by the single cpumask_t fallback_doms. |
| 6439 | */ |
| 6440 | static cpumask_t fallback_doms; |
| 6441 | |
| 6442 | /* |
| 6443 | * Set up scheduler domains and groups. Callers must hold the hotplug lock. |
| 6444 | * For now this just excludes isolated cpus, but could be used to |
| 6445 | * exclude other special cases in the future. |
| 6446 | */ |
| 6447 | static int arch_init_sched_domains(const cpumask_t *cpu_map) |
| 6448 | { |
| 6449 | int err; |
| 6450 | |
| 6451 | ndoms_cur = 1; |
| 6452 | doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL); |
| 6453 | if (!doms_cur) |
| 6454 | doms_cur = &fallback_doms; |
| 6455 | cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map); |
| 6456 | err = build_sched_domains(doms_cur); |
| 6457 | register_sched_domain_sysctl(); |
| 6458 | |
| 6459 | return err; |
| 6460 | } |
| 6461 | |
| 6462 | static void arch_destroy_sched_domains(const cpumask_t *cpu_map) |
| 6463 | { |
| 6464 | free_sched_groups(cpu_map); |
| 6465 | } |
| 6466 | |
| 6467 | /* |
| 6468 | * Detach sched domains from a group of cpus specified in cpu_map |
| 6469 | * These cpus will now be attached to the NULL domain |
| 6470 | */ |
| 6471 | static void detach_destroy_domains(const cpumask_t *cpu_map) |
| 6472 | { |
| 6473 | int i; |
| 6474 | |
| 6475 | unregister_sched_domain_sysctl(); |
| 6476 | |
| 6477 | for_each_cpu_mask(i, *cpu_map) |
| 6478 | cpu_attach_domain(NULL, i); |
| 6479 | synchronize_sched(); |
| 6480 | arch_destroy_sched_domains(cpu_map); |
| 6481 | } |
| 6482 | |
| 6483 | /* |
| 6484 | * Partition sched domains as specified by the 'ndoms_new' |
| 6485 | * cpumasks in the array doms_new[] of cpumasks. This compares |
| 6486 | * doms_new[] to the current sched domain partitioning, doms_cur[]. |
| 6487 | * It destroys each deleted domain and builds each new domain. |
| 6488 | * |
| 6489 | * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'. |
| 6490 | * The masks don't intersect (don't overlap.) We should setup one |
| 6491 | * sched domain for each mask. CPUs not in any of the cpumasks will |
| 6492 | * not be load balanced. If the same cpumask appears both in the |
| 6493 | * current 'doms_cur' domains and in the new 'doms_new', we can leave |
| 6494 | * it as it is. |
| 6495 | * |
| 6496 | * The passed in 'doms_new' should be kmalloc'd. This routine takes |
| 6497 | * ownership of it and will kfree it when done with it. If the caller |
| 6498 | * failed the kmalloc call, then it can pass in doms_new == NULL, |
| 6499 | * and partition_sched_domains() will fallback to the single partition |
| 6500 | * 'fallback_doms'. |
| 6501 | * |
| 6502 | * Call with hotplug lock held |
| 6503 | */ |
| 6504 | void partition_sched_domains(int ndoms_new, cpumask_t *doms_new) |
| 6505 | { |
| 6506 | int i, j; |
| 6507 | |
| 6508 | lock_doms_cur(); |
| 6509 | |
| 6510 | /* always unregister in case we don't destroy any domains */ |
| 6511 | unregister_sched_domain_sysctl(); |
| 6512 | |
| 6513 | if (doms_new == NULL) { |
| 6514 | ndoms_new = 1; |
| 6515 | doms_new = &fallback_doms; |
| 6516 | cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map); |
| 6517 | } |
| 6518 | |
| 6519 | /* Destroy deleted domains */ |
| 6520 | for (i = 0; i < ndoms_cur; i++) { |
| 6521 | for (j = 0; j < ndoms_new; j++) { |
| 6522 | if (cpus_equal(doms_cur[i], doms_new[j])) |
| 6523 | goto match1; |
| 6524 | } |
| 6525 | /* no match - a current sched domain not in new doms_new[] */ |
| 6526 | detach_destroy_domains(doms_cur + i); |
| 6527 | match1: |
| 6528 | ; |
| 6529 | } |
| 6530 | |
| 6531 | /* Build new domains */ |
| 6532 | for (i = 0; i < ndoms_new; i++) { |
| 6533 | for (j = 0; j < ndoms_cur; j++) { |
| 6534 | if (cpus_equal(doms_new[i], doms_cur[j])) |
| 6535 | goto match2; |
| 6536 | } |
| 6537 | /* no match - add a new doms_new */ |
| 6538 | build_sched_domains(doms_new + i); |
| 6539 | match2: |
| 6540 | ; |
| 6541 | } |
| 6542 | |
| 6543 | /* Remember the new sched domains */ |
| 6544 | if (doms_cur != &fallback_doms) |
| 6545 | kfree(doms_cur); |
| 6546 | doms_cur = doms_new; |
| 6547 | ndoms_cur = ndoms_new; |
| 6548 | |
| 6549 | register_sched_domain_sysctl(); |
| 6550 | |
| 6551 | unlock_doms_cur(); |
| 6552 | } |
| 6553 | |
| 6554 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) |
| 6555 | static int arch_reinit_sched_domains(void) |
| 6556 | { |
| 6557 | int err; |
| 6558 | |
| 6559 | get_online_cpus(); |
| 6560 | detach_destroy_domains(&cpu_online_map); |
| 6561 | err = arch_init_sched_domains(&cpu_online_map); |
| 6562 | put_online_cpus(); |
| 6563 | |
| 6564 | return err; |
| 6565 | } |
| 6566 | |
| 6567 | static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt) |
| 6568 | { |
| 6569 | int ret; |
| 6570 | |
| 6571 | if (buf[0] != '0' && buf[0] != '1') |
| 6572 | return -EINVAL; |
| 6573 | |
| 6574 | if (smt) |
| 6575 | sched_smt_power_savings = (buf[0] == '1'); |
| 6576 | else |
| 6577 | sched_mc_power_savings = (buf[0] == '1'); |
| 6578 | |
| 6579 | ret = arch_reinit_sched_domains(); |
| 6580 | |
| 6581 | return ret ? ret : count; |
| 6582 | } |
| 6583 | |
| 6584 | #ifdef CONFIG_SCHED_MC |
| 6585 | static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page) |
| 6586 | { |
| 6587 | return sprintf(page, "%u\n", sched_mc_power_savings); |
| 6588 | } |
| 6589 | static ssize_t sched_mc_power_savings_store(struct sys_device *dev, |
| 6590 | const char *buf, size_t count) |
| 6591 | { |
| 6592 | return sched_power_savings_store(buf, count, 0); |
| 6593 | } |
| 6594 | static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show, |
| 6595 | sched_mc_power_savings_store); |
| 6596 | #endif |
| 6597 | |
| 6598 | #ifdef CONFIG_SCHED_SMT |
| 6599 | static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page) |
| 6600 | { |
| 6601 | return sprintf(page, "%u\n", sched_smt_power_savings); |
| 6602 | } |
| 6603 | static ssize_t sched_smt_power_savings_store(struct sys_device *dev, |
| 6604 | const char *buf, size_t count) |
| 6605 | { |
| 6606 | return sched_power_savings_store(buf, count, 1); |
| 6607 | } |
| 6608 | static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show, |
| 6609 | sched_smt_power_savings_store); |
| 6610 | #endif |
| 6611 | |
| 6612 | int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls) |
| 6613 | { |
| 6614 | int err = 0; |
| 6615 | |
| 6616 | #ifdef CONFIG_SCHED_SMT |
| 6617 | if (smt_capable()) |
| 6618 | err = sysfs_create_file(&cls->kset.kobj, |
| 6619 | &attr_sched_smt_power_savings.attr); |
| 6620 | #endif |
| 6621 | #ifdef CONFIG_SCHED_MC |
| 6622 | if (!err && mc_capable()) |
| 6623 | err = sysfs_create_file(&cls->kset.kobj, |
| 6624 | &attr_sched_mc_power_savings.attr); |
| 6625 | #endif |
| 6626 | return err; |
| 6627 | } |
| 6628 | #endif |
| 6629 | |
| 6630 | /* |
| 6631 | * Force a reinitialization of the sched domains hierarchy. The domains |
| 6632 | * and groups cannot be updated in place without racing with the balancing |
| 6633 | * code, so we temporarily attach all running cpus to the NULL domain |
| 6634 | * which will prevent rebalancing while the sched domains are recalculated. |
| 6635 | */ |
| 6636 | static int update_sched_domains(struct notifier_block *nfb, |
| 6637 | unsigned long action, void *hcpu) |
| 6638 | { |
| 6639 | switch (action) { |
| 6640 | case CPU_UP_PREPARE: |
| 6641 | case CPU_UP_PREPARE_FROZEN: |
| 6642 | case CPU_DOWN_PREPARE: |
| 6643 | case CPU_DOWN_PREPARE_FROZEN: |
| 6644 | detach_destroy_domains(&cpu_online_map); |
| 6645 | return NOTIFY_OK; |
| 6646 | |
| 6647 | case CPU_UP_CANCELED: |
| 6648 | case CPU_UP_CANCELED_FROZEN: |
| 6649 | case CPU_DOWN_FAILED: |
| 6650 | case CPU_DOWN_FAILED_FROZEN: |
| 6651 | case CPU_ONLINE: |
| 6652 | case CPU_ONLINE_FROZEN: |
| 6653 | case CPU_DEAD: |
| 6654 | case CPU_DEAD_FROZEN: |
| 6655 | /* |
| 6656 | * Fall through and re-initialise the domains. |
| 6657 | */ |
| 6658 | break; |
| 6659 | default: |
| 6660 | return NOTIFY_DONE; |
| 6661 | } |
| 6662 | |
| 6663 | /* The hotplug lock is already held by cpu_up/cpu_down */ |
| 6664 | arch_init_sched_domains(&cpu_online_map); |
| 6665 | |
| 6666 | return NOTIFY_OK; |
| 6667 | } |
| 6668 | |
| 6669 | void __init sched_init_smp(void) |
| 6670 | { |
| 6671 | cpumask_t non_isolated_cpus; |
| 6672 | |
| 6673 | get_online_cpus(); |
| 6674 | arch_init_sched_domains(&cpu_online_map); |
| 6675 | cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map); |
| 6676 | if (cpus_empty(non_isolated_cpus)) |
| 6677 | cpu_set(smp_processor_id(), non_isolated_cpus); |
| 6678 | put_online_cpus(); |
| 6679 | /* XXX: Theoretical race here - CPU may be hotplugged now */ |
| 6680 | hotcpu_notifier(update_sched_domains, 0); |
| 6681 | |
| 6682 | /* Move init over to a non-isolated CPU */ |
| 6683 | if (set_cpus_allowed(current, non_isolated_cpus) < 0) |
| 6684 | BUG(); |
| 6685 | sched_init_granularity(); |
| 6686 | |
| 6687 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 6688 | if (nr_cpu_ids == 1) |
| 6689 | return; |
| 6690 | |
| 6691 | lb_monitor_task = kthread_create(load_balance_monitor, NULL, |
| 6692 | "group_balance"); |
| 6693 | if (!IS_ERR(lb_monitor_task)) { |
| 6694 | lb_monitor_task->flags |= PF_NOFREEZE; |
| 6695 | wake_up_process(lb_monitor_task); |
| 6696 | } else { |
| 6697 | printk(KERN_ERR "Could not create load balance monitor thread" |
| 6698 | "(error = %ld) \n", PTR_ERR(lb_monitor_task)); |
| 6699 | } |
| 6700 | #endif |
| 6701 | } |
| 6702 | #else |
| 6703 | void __init sched_init_smp(void) |
| 6704 | { |
| 6705 | sched_init_granularity(); |
| 6706 | } |
| 6707 | #endif /* CONFIG_SMP */ |
| 6708 | |
| 6709 | int in_sched_functions(unsigned long addr) |
| 6710 | { |
| 6711 | return in_lock_functions(addr) || |
| 6712 | (addr >= (unsigned long)__sched_text_start |
| 6713 | && addr < (unsigned long)__sched_text_end); |
| 6714 | } |
| 6715 | |
| 6716 | static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq) |
| 6717 | { |
| 6718 | cfs_rq->tasks_timeline = RB_ROOT; |
| 6719 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 6720 | cfs_rq->rq = rq; |
| 6721 | #endif |
| 6722 | cfs_rq->min_vruntime = (u64)(-(1LL << 20)); |
| 6723 | } |
| 6724 | |
| 6725 | void __init sched_init(void) |
| 6726 | { |
| 6727 | int highest_cpu = 0; |
| 6728 | int i, j; |
| 6729 | |
| 6730 | for_each_possible_cpu(i) { |
| 6731 | struct rt_prio_array *array; |
| 6732 | struct rq *rq; |
| 6733 | |
| 6734 | rq = cpu_rq(i); |
| 6735 | spin_lock_init(&rq->lock); |
| 6736 | lockdep_set_class(&rq->lock, &rq->rq_lock_key); |
| 6737 | rq->nr_running = 0; |
| 6738 | rq->clock = 1; |
| 6739 | init_cfs_rq(&rq->cfs, rq); |
| 6740 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 6741 | INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); |
| 6742 | { |
| 6743 | struct cfs_rq *cfs_rq = &per_cpu(init_cfs_rq, i); |
| 6744 | struct sched_entity *se = |
| 6745 | &per_cpu(init_sched_entity, i); |
| 6746 | |
| 6747 | init_cfs_rq_p[i] = cfs_rq; |
| 6748 | init_cfs_rq(cfs_rq, rq); |
| 6749 | cfs_rq->tg = &init_task_group; |
| 6750 | list_add(&cfs_rq->leaf_cfs_rq_list, |
| 6751 | &rq->leaf_cfs_rq_list); |
| 6752 | |
| 6753 | init_sched_entity_p[i] = se; |
| 6754 | se->cfs_rq = &rq->cfs; |
| 6755 | se->my_q = cfs_rq; |
| 6756 | se->load.weight = init_task_group_load; |
| 6757 | se->load.inv_weight = |
| 6758 | div64_64(1ULL<<32, init_task_group_load); |
| 6759 | se->parent = NULL; |
| 6760 | } |
| 6761 | init_task_group.shares = init_task_group_load; |
| 6762 | #endif |
| 6763 | |
| 6764 | for (j = 0; j < CPU_LOAD_IDX_MAX; j++) |
| 6765 | rq->cpu_load[j] = 0; |
| 6766 | #ifdef CONFIG_SMP |
| 6767 | rq->sd = NULL; |
| 6768 | rq->active_balance = 0; |
| 6769 | rq->next_balance = jiffies; |
| 6770 | rq->push_cpu = 0; |
| 6771 | rq->cpu = i; |
| 6772 | rq->migration_thread = NULL; |
| 6773 | INIT_LIST_HEAD(&rq->migration_queue); |
| 6774 | rq->rt.highest_prio = MAX_RT_PRIO; |
| 6775 | rq->rt.overloaded = 0; |
| 6776 | #endif |
| 6777 | atomic_set(&rq->nr_iowait, 0); |
| 6778 | |
| 6779 | array = &rq->rt.active; |
| 6780 | for (j = 0; j < MAX_RT_PRIO; j++) { |
| 6781 | INIT_LIST_HEAD(array->queue + j); |
| 6782 | __clear_bit(j, array->bitmap); |
| 6783 | } |
| 6784 | highest_cpu = i; |
| 6785 | /* delimiter for bitsearch: */ |
| 6786 | __set_bit(MAX_RT_PRIO, array->bitmap); |
| 6787 | } |
| 6788 | |
| 6789 | set_load_weight(&init_task); |
| 6790 | |
| 6791 | #ifdef CONFIG_PREEMPT_NOTIFIERS |
| 6792 | INIT_HLIST_HEAD(&init_task.preempt_notifiers); |
| 6793 | #endif |
| 6794 | |
| 6795 | #ifdef CONFIG_SMP |
| 6796 | nr_cpu_ids = highest_cpu + 1; |
| 6797 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL); |
| 6798 | #endif |
| 6799 | |
| 6800 | #ifdef CONFIG_RT_MUTEXES |
| 6801 | plist_head_init(&init_task.pi_waiters, &init_task.pi_lock); |
| 6802 | #endif |
| 6803 | |
| 6804 | /* |
| 6805 | * The boot idle thread does lazy MMU switching as well: |
| 6806 | */ |
| 6807 | atomic_inc(&init_mm.mm_count); |
| 6808 | enter_lazy_tlb(&init_mm, current); |
| 6809 | |
| 6810 | /* |
| 6811 | * Make us the idle thread. Technically, schedule() should not be |
| 6812 | * called from this thread, however somewhere below it might be, |
| 6813 | * but because we are the idle thread, we just pick up running again |
| 6814 | * when this runqueue becomes "idle". |
| 6815 | */ |
| 6816 | init_idle(current, smp_processor_id()); |
| 6817 | /* |
| 6818 | * During early bootup we pretend to be a normal task: |
| 6819 | */ |
| 6820 | current->sched_class = &fair_sched_class; |
| 6821 | } |
| 6822 | |
| 6823 | #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP |
| 6824 | void __might_sleep(char *file, int line) |
| 6825 | { |
| 6826 | #ifdef in_atomic |
| 6827 | static unsigned long prev_jiffy; /* ratelimiting */ |
| 6828 | |
| 6829 | if ((in_atomic() || irqs_disabled()) && |
| 6830 | system_state == SYSTEM_RUNNING && !oops_in_progress) { |
| 6831 | if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) |
| 6832 | return; |
| 6833 | prev_jiffy = jiffies; |
| 6834 | printk(KERN_ERR "BUG: sleeping function called from invalid" |
| 6835 | " context at %s:%d\n", file, line); |
| 6836 | printk("in_atomic():%d, irqs_disabled():%d\n", |
| 6837 | in_atomic(), irqs_disabled()); |
| 6838 | debug_show_held_locks(current); |
| 6839 | if (irqs_disabled()) |
| 6840 | print_irqtrace_events(current); |
| 6841 | dump_stack(); |
| 6842 | } |
| 6843 | #endif |
| 6844 | } |
| 6845 | EXPORT_SYMBOL(__might_sleep); |
| 6846 | #endif |
| 6847 | |
| 6848 | #ifdef CONFIG_MAGIC_SYSRQ |
| 6849 | static void normalize_task(struct rq *rq, struct task_struct *p) |
| 6850 | { |
| 6851 | int on_rq; |
| 6852 | update_rq_clock(rq); |
| 6853 | on_rq = p->se.on_rq; |
| 6854 | if (on_rq) |
| 6855 | deactivate_task(rq, p, 0); |
| 6856 | __setscheduler(rq, p, SCHED_NORMAL, 0); |
| 6857 | if (on_rq) { |
| 6858 | activate_task(rq, p, 0); |
| 6859 | resched_task(rq->curr); |
| 6860 | } |
| 6861 | } |
| 6862 | |
| 6863 | void normalize_rt_tasks(void) |
| 6864 | { |
| 6865 | struct task_struct *g, *p; |
| 6866 | unsigned long flags; |
| 6867 | struct rq *rq; |
| 6868 | |
| 6869 | read_lock_irq(&tasklist_lock); |
| 6870 | do_each_thread(g, p) { |
| 6871 | /* |
| 6872 | * Only normalize user tasks: |
| 6873 | */ |
| 6874 | if (!p->mm) |
| 6875 | continue; |
| 6876 | |
| 6877 | p->se.exec_start = 0; |
| 6878 | #ifdef CONFIG_SCHEDSTATS |
| 6879 | p->se.wait_start = 0; |
| 6880 | p->se.sleep_start = 0; |
| 6881 | p->se.block_start = 0; |
| 6882 | #endif |
| 6883 | task_rq(p)->clock = 0; |
| 6884 | |
| 6885 | if (!rt_task(p)) { |
| 6886 | /* |
| 6887 | * Renice negative nice level userspace |
| 6888 | * tasks back to 0: |
| 6889 | */ |
| 6890 | if (TASK_NICE(p) < 0 && p->mm) |
| 6891 | set_user_nice(p, 0); |
| 6892 | continue; |
| 6893 | } |
| 6894 | |
| 6895 | spin_lock_irqsave(&p->pi_lock, flags); |
| 6896 | rq = __task_rq_lock(p); |
| 6897 | |
| 6898 | normalize_task(rq, p); |
| 6899 | |
| 6900 | __task_rq_unlock(rq); |
| 6901 | spin_unlock_irqrestore(&p->pi_lock, flags); |
| 6902 | } while_each_thread(g, p); |
| 6903 | |
| 6904 | read_unlock_irq(&tasklist_lock); |
| 6905 | } |
| 6906 | |
| 6907 | #endif /* CONFIG_MAGIC_SYSRQ */ |
| 6908 | |
| 6909 | #ifdef CONFIG_IA64 |
| 6910 | /* |
| 6911 | * These functions are only useful for the IA64 MCA handling. |
| 6912 | * |
| 6913 | * They can only be called when the whole system has been |
| 6914 | * stopped - every CPU needs to be quiescent, and no scheduling |
| 6915 | * activity can take place. Using them for anything else would |
| 6916 | * be a serious bug, and as a result, they aren't even visible |
| 6917 | * under any other configuration. |
| 6918 | */ |
| 6919 | |
| 6920 | /** |
| 6921 | * curr_task - return the current task for a given cpu. |
| 6922 | * @cpu: the processor in question. |
| 6923 | * |
| 6924 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! |
| 6925 | */ |
| 6926 | struct task_struct *curr_task(int cpu) |
| 6927 | { |
| 6928 | return cpu_curr(cpu); |
| 6929 | } |
| 6930 | |
| 6931 | /** |
| 6932 | * set_curr_task - set the current task for a given cpu. |
| 6933 | * @cpu: the processor in question. |
| 6934 | * @p: the task pointer to set. |
| 6935 | * |
| 6936 | * Description: This function must only be used when non-maskable interrupts |
| 6937 | * are serviced on a separate stack. It allows the architecture to switch the |
| 6938 | * notion of the current task on a cpu in a non-blocking manner. This function |
| 6939 | * must be called with all CPU's synchronized, and interrupts disabled, the |
| 6940 | * and caller must save the original value of the current task (see |
| 6941 | * curr_task() above) and restore that value before reenabling interrupts and |
| 6942 | * re-starting the system. |
| 6943 | * |
| 6944 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! |
| 6945 | */ |
| 6946 | void set_curr_task(int cpu, struct task_struct *p) |
| 6947 | { |
| 6948 | cpu_curr(cpu) = p; |
| 6949 | } |
| 6950 | |
| 6951 | #endif |
| 6952 | |
| 6953 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 6954 | |
| 6955 | #ifdef CONFIG_SMP |
| 6956 | /* |
| 6957 | * distribute shares of all task groups among their schedulable entities, |
| 6958 | * to reflect load distrbution across cpus. |
| 6959 | */ |
| 6960 | static int rebalance_shares(struct sched_domain *sd, int this_cpu) |
| 6961 | { |
| 6962 | struct cfs_rq *cfs_rq; |
| 6963 | struct rq *rq = cpu_rq(this_cpu); |
| 6964 | cpumask_t sdspan = sd->span; |
| 6965 | int balanced = 1; |
| 6966 | |
| 6967 | /* Walk thr' all the task groups that we have */ |
| 6968 | for_each_leaf_cfs_rq(rq, cfs_rq) { |
| 6969 | int i; |
| 6970 | unsigned long total_load = 0, total_shares; |
| 6971 | struct task_group *tg = cfs_rq->tg; |
| 6972 | |
| 6973 | /* Gather total task load of this group across cpus */ |
| 6974 | for_each_cpu_mask(i, sdspan) |
| 6975 | total_load += tg->cfs_rq[i]->load.weight; |
| 6976 | |
| 6977 | /* Nothing to do if this group has no load */ |
| 6978 | if (!total_load) |
| 6979 | continue; |
| 6980 | |
| 6981 | /* |
| 6982 | * tg->shares represents the number of cpu shares the task group |
| 6983 | * is eligible to hold on a single cpu. On N cpus, it is |
| 6984 | * eligible to hold (N * tg->shares) number of cpu shares. |
| 6985 | */ |
| 6986 | total_shares = tg->shares * cpus_weight(sdspan); |
| 6987 | |
| 6988 | /* |
| 6989 | * redistribute total_shares across cpus as per the task load |
| 6990 | * distribution. |
| 6991 | */ |
| 6992 | for_each_cpu_mask(i, sdspan) { |
| 6993 | unsigned long local_load, local_shares; |
| 6994 | |
| 6995 | local_load = tg->cfs_rq[i]->load.weight; |
| 6996 | local_shares = (local_load * total_shares) / total_load; |
| 6997 | if (!local_shares) |
| 6998 | local_shares = MIN_GROUP_SHARES; |
| 6999 | if (local_shares == tg->se[i]->load.weight) |
| 7000 | continue; |
| 7001 | |
| 7002 | spin_lock_irq(&cpu_rq(i)->lock); |
| 7003 | set_se_shares(tg->se[i], local_shares); |
| 7004 | spin_unlock_irq(&cpu_rq(i)->lock); |
| 7005 | balanced = 0; |
| 7006 | } |
| 7007 | } |
| 7008 | |
| 7009 | return balanced; |
| 7010 | } |
| 7011 | |
| 7012 | /* |
| 7013 | * How frequently should we rebalance_shares() across cpus? |
| 7014 | * |
| 7015 | * The more frequently we rebalance shares, the more accurate is the fairness |
| 7016 | * of cpu bandwidth distribution between task groups. However higher frequency |
| 7017 | * also implies increased scheduling overhead. |
| 7018 | * |
| 7019 | * sysctl_sched_min_bal_int_shares represents the minimum interval between |
| 7020 | * consecutive calls to rebalance_shares() in the same sched domain. |
| 7021 | * |
| 7022 | * sysctl_sched_max_bal_int_shares represents the maximum interval between |
| 7023 | * consecutive calls to rebalance_shares() in the same sched domain. |
| 7024 | * |
| 7025 | * These settings allows for the appropriate tradeoff between accuracy of |
| 7026 | * fairness and the associated overhead. |
| 7027 | * |
| 7028 | */ |
| 7029 | |
| 7030 | /* default: 8ms, units: milliseconds */ |
| 7031 | const_debug unsigned int sysctl_sched_min_bal_int_shares = 8; |
| 7032 | |
| 7033 | /* default: 128ms, units: milliseconds */ |
| 7034 | const_debug unsigned int sysctl_sched_max_bal_int_shares = 128; |
| 7035 | |
| 7036 | /* kernel thread that runs rebalance_shares() periodically */ |
| 7037 | static int load_balance_monitor(void *unused) |
| 7038 | { |
| 7039 | unsigned int timeout = sysctl_sched_min_bal_int_shares; |
| 7040 | struct sched_param schedparm; |
| 7041 | int ret; |
| 7042 | |
| 7043 | /* |
| 7044 | * We don't want this thread's execution to be limited by the shares |
| 7045 | * assigned to default group (init_task_group). Hence make it run |
| 7046 | * as a SCHED_RR RT task at the lowest priority. |
| 7047 | */ |
| 7048 | schedparm.sched_priority = 1; |
| 7049 | ret = sched_setscheduler(current, SCHED_RR, &schedparm); |
| 7050 | if (ret) |
| 7051 | printk(KERN_ERR "Couldn't set SCHED_RR policy for load balance" |
| 7052 | " monitor thread (error = %d) \n", ret); |
| 7053 | |
| 7054 | while (!kthread_should_stop()) { |
| 7055 | int i, cpu, balanced = 1; |
| 7056 | |
| 7057 | /* Prevent cpus going down or coming up */ |
| 7058 | get_online_cpus(); |
| 7059 | /* lockout changes to doms_cur[] array */ |
| 7060 | lock_doms_cur(); |
| 7061 | /* |
| 7062 | * Enter a rcu read-side critical section to safely walk rq->sd |
| 7063 | * chain on various cpus and to walk task group list |
| 7064 | * (rq->leaf_cfs_rq_list) in rebalance_shares(). |
| 7065 | */ |
| 7066 | rcu_read_lock(); |
| 7067 | |
| 7068 | for (i = 0; i < ndoms_cur; i++) { |
| 7069 | cpumask_t cpumap = doms_cur[i]; |
| 7070 | struct sched_domain *sd = NULL, *sd_prev = NULL; |
| 7071 | |
| 7072 | cpu = first_cpu(cpumap); |
| 7073 | |
| 7074 | /* Find the highest domain at which to balance shares */ |
| 7075 | for_each_domain(cpu, sd) { |
| 7076 | if (!(sd->flags & SD_LOAD_BALANCE)) |
| 7077 | continue; |
| 7078 | sd_prev = sd; |
| 7079 | } |
| 7080 | |
| 7081 | sd = sd_prev; |
| 7082 | /* sd == NULL? No load balance reqd in this domain */ |
| 7083 | if (!sd) |
| 7084 | continue; |
| 7085 | |
| 7086 | balanced &= rebalance_shares(sd, cpu); |
| 7087 | } |
| 7088 | |
| 7089 | rcu_read_unlock(); |
| 7090 | |
| 7091 | unlock_doms_cur(); |
| 7092 | put_online_cpus(); |
| 7093 | |
| 7094 | if (!balanced) |
| 7095 | timeout = sysctl_sched_min_bal_int_shares; |
| 7096 | else if (timeout < sysctl_sched_max_bal_int_shares) |
| 7097 | timeout *= 2; |
| 7098 | |
| 7099 | msleep_interruptible(timeout); |
| 7100 | } |
| 7101 | |
| 7102 | return 0; |
| 7103 | } |
| 7104 | #endif /* CONFIG_SMP */ |
| 7105 | |
| 7106 | /* allocate runqueue etc for a new task group */ |
| 7107 | struct task_group *sched_create_group(void) |
| 7108 | { |
| 7109 | struct task_group *tg; |
| 7110 | struct cfs_rq *cfs_rq; |
| 7111 | struct sched_entity *se; |
| 7112 | struct rq *rq; |
| 7113 | int i; |
| 7114 | |
| 7115 | tg = kzalloc(sizeof(*tg), GFP_KERNEL); |
| 7116 | if (!tg) |
| 7117 | return ERR_PTR(-ENOMEM); |
| 7118 | |
| 7119 | tg->cfs_rq = kzalloc(sizeof(cfs_rq) * NR_CPUS, GFP_KERNEL); |
| 7120 | if (!tg->cfs_rq) |
| 7121 | goto err; |
| 7122 | tg->se = kzalloc(sizeof(se) * NR_CPUS, GFP_KERNEL); |
| 7123 | if (!tg->se) |
| 7124 | goto err; |
| 7125 | |
| 7126 | for_each_possible_cpu(i) { |
| 7127 | rq = cpu_rq(i); |
| 7128 | |
| 7129 | cfs_rq = kmalloc_node(sizeof(struct cfs_rq), GFP_KERNEL, |
| 7130 | cpu_to_node(i)); |
| 7131 | if (!cfs_rq) |
| 7132 | goto err; |
| 7133 | |
| 7134 | se = kmalloc_node(sizeof(struct sched_entity), GFP_KERNEL, |
| 7135 | cpu_to_node(i)); |
| 7136 | if (!se) |
| 7137 | goto err; |
| 7138 | |
| 7139 | memset(cfs_rq, 0, sizeof(struct cfs_rq)); |
| 7140 | memset(se, 0, sizeof(struct sched_entity)); |
| 7141 | |
| 7142 | tg->cfs_rq[i] = cfs_rq; |
| 7143 | init_cfs_rq(cfs_rq, rq); |
| 7144 | cfs_rq->tg = tg; |
| 7145 | |
| 7146 | tg->se[i] = se; |
| 7147 | se->cfs_rq = &rq->cfs; |
| 7148 | se->my_q = cfs_rq; |
| 7149 | se->load.weight = NICE_0_LOAD; |
| 7150 | se->load.inv_weight = div64_64(1ULL<<32, NICE_0_LOAD); |
| 7151 | se->parent = NULL; |
| 7152 | } |
| 7153 | |
| 7154 | tg->shares = NICE_0_LOAD; |
| 7155 | |
| 7156 | lock_task_group_list(); |
| 7157 | for_each_possible_cpu(i) { |
| 7158 | rq = cpu_rq(i); |
| 7159 | cfs_rq = tg->cfs_rq[i]; |
| 7160 | list_add_rcu(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list); |
| 7161 | } |
| 7162 | unlock_task_group_list(); |
| 7163 | |
| 7164 | return tg; |
| 7165 | |
| 7166 | err: |
| 7167 | for_each_possible_cpu(i) { |
| 7168 | if (tg->cfs_rq) |
| 7169 | kfree(tg->cfs_rq[i]); |
| 7170 | if (tg->se) |
| 7171 | kfree(tg->se[i]); |
| 7172 | } |
| 7173 | kfree(tg->cfs_rq); |
| 7174 | kfree(tg->se); |
| 7175 | kfree(tg); |
| 7176 | |
| 7177 | return ERR_PTR(-ENOMEM); |
| 7178 | } |
| 7179 | |
| 7180 | /* rcu callback to free various structures associated with a task group */ |
| 7181 | static void free_sched_group(struct rcu_head *rhp) |
| 7182 | { |
| 7183 | struct task_group *tg = container_of(rhp, struct task_group, rcu); |
| 7184 | struct cfs_rq *cfs_rq; |
| 7185 | struct sched_entity *se; |
| 7186 | int i; |
| 7187 | |
| 7188 | /* now it should be safe to free those cfs_rqs */ |
| 7189 | for_each_possible_cpu(i) { |
| 7190 | cfs_rq = tg->cfs_rq[i]; |
| 7191 | kfree(cfs_rq); |
| 7192 | |
| 7193 | se = tg->se[i]; |
| 7194 | kfree(se); |
| 7195 | } |
| 7196 | |
| 7197 | kfree(tg->cfs_rq); |
| 7198 | kfree(tg->se); |
| 7199 | kfree(tg); |
| 7200 | } |
| 7201 | |
| 7202 | /* Destroy runqueue etc associated with a task group */ |
| 7203 | void sched_destroy_group(struct task_group *tg) |
| 7204 | { |
| 7205 | struct cfs_rq *cfs_rq = NULL; |
| 7206 | int i; |
| 7207 | |
| 7208 | lock_task_group_list(); |
| 7209 | for_each_possible_cpu(i) { |
| 7210 | cfs_rq = tg->cfs_rq[i]; |
| 7211 | list_del_rcu(&cfs_rq->leaf_cfs_rq_list); |
| 7212 | } |
| 7213 | unlock_task_group_list(); |
| 7214 | |
| 7215 | BUG_ON(!cfs_rq); |
| 7216 | |
| 7217 | /* wait for possible concurrent references to cfs_rqs complete */ |
| 7218 | call_rcu(&tg->rcu, free_sched_group); |
| 7219 | } |
| 7220 | |
| 7221 | /* change task's runqueue when it moves between groups. |
| 7222 | * The caller of this function should have put the task in its new group |
| 7223 | * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to |
| 7224 | * reflect its new group. |
| 7225 | */ |
| 7226 | void sched_move_task(struct task_struct *tsk) |
| 7227 | { |
| 7228 | int on_rq, running; |
| 7229 | unsigned long flags; |
| 7230 | struct rq *rq; |
| 7231 | |
| 7232 | rq = task_rq_lock(tsk, &flags); |
| 7233 | |
| 7234 | if (tsk->sched_class != &fair_sched_class) { |
| 7235 | set_task_cfs_rq(tsk, task_cpu(tsk)); |
| 7236 | goto done; |
| 7237 | } |
| 7238 | |
| 7239 | update_rq_clock(rq); |
| 7240 | |
| 7241 | running = task_current(rq, tsk); |
| 7242 | on_rq = tsk->se.on_rq; |
| 7243 | |
| 7244 | if (on_rq) { |
| 7245 | dequeue_task(rq, tsk, 0); |
| 7246 | if (unlikely(running)) |
| 7247 | tsk->sched_class->put_prev_task(rq, tsk); |
| 7248 | } |
| 7249 | |
| 7250 | set_task_cfs_rq(tsk, task_cpu(tsk)); |
| 7251 | |
| 7252 | if (on_rq) { |
| 7253 | if (unlikely(running)) |
| 7254 | tsk->sched_class->set_curr_task(rq); |
| 7255 | enqueue_task(rq, tsk, 0); |
| 7256 | } |
| 7257 | |
| 7258 | done: |
| 7259 | task_rq_unlock(rq, &flags); |
| 7260 | } |
| 7261 | |
| 7262 | /* rq->lock to be locked by caller */ |
| 7263 | static void set_se_shares(struct sched_entity *se, unsigned long shares) |
| 7264 | { |
| 7265 | struct cfs_rq *cfs_rq = se->cfs_rq; |
| 7266 | struct rq *rq = cfs_rq->rq; |
| 7267 | int on_rq; |
| 7268 | |
| 7269 | if (!shares) |
| 7270 | shares = MIN_GROUP_SHARES; |
| 7271 | |
| 7272 | on_rq = se->on_rq; |
| 7273 | if (on_rq) { |
| 7274 | dequeue_entity(cfs_rq, se, 0); |
| 7275 | dec_cpu_load(rq, se->load.weight); |
| 7276 | } |
| 7277 | |
| 7278 | se->load.weight = shares; |
| 7279 | se->load.inv_weight = div64_64((1ULL<<32), shares); |
| 7280 | |
| 7281 | if (on_rq) { |
| 7282 | enqueue_entity(cfs_rq, se, 0); |
| 7283 | inc_cpu_load(rq, se->load.weight); |
| 7284 | } |
| 7285 | } |
| 7286 | |
| 7287 | int sched_group_set_shares(struct task_group *tg, unsigned long shares) |
| 7288 | { |
| 7289 | int i; |
| 7290 | struct cfs_rq *cfs_rq; |
| 7291 | struct rq *rq; |
| 7292 | |
| 7293 | lock_task_group_list(); |
| 7294 | if (tg->shares == shares) |
| 7295 | goto done; |
| 7296 | |
| 7297 | if (shares < MIN_GROUP_SHARES) |
| 7298 | shares = MIN_GROUP_SHARES; |
| 7299 | |
| 7300 | /* |
| 7301 | * Prevent any load balance activity (rebalance_shares, |
| 7302 | * load_balance_fair) from referring to this group first, |
| 7303 | * by taking it off the rq->leaf_cfs_rq_list on each cpu. |
| 7304 | */ |
| 7305 | for_each_possible_cpu(i) { |
| 7306 | cfs_rq = tg->cfs_rq[i]; |
| 7307 | list_del_rcu(&cfs_rq->leaf_cfs_rq_list); |
| 7308 | } |
| 7309 | |
| 7310 | /* wait for any ongoing reference to this group to finish */ |
| 7311 | synchronize_sched(); |
| 7312 | |
| 7313 | /* |
| 7314 | * Now we are free to modify the group's share on each cpu |
| 7315 | * w/o tripping rebalance_share or load_balance_fair. |
| 7316 | */ |
| 7317 | tg->shares = shares; |
| 7318 | for_each_possible_cpu(i) { |
| 7319 | spin_lock_irq(&cpu_rq(i)->lock); |
| 7320 | set_se_shares(tg->se[i], shares); |
| 7321 | spin_unlock_irq(&cpu_rq(i)->lock); |
| 7322 | } |
| 7323 | |
| 7324 | /* |
| 7325 | * Enable load balance activity on this group, by inserting it back on |
| 7326 | * each cpu's rq->leaf_cfs_rq_list. |
| 7327 | */ |
| 7328 | for_each_possible_cpu(i) { |
| 7329 | rq = cpu_rq(i); |
| 7330 | cfs_rq = tg->cfs_rq[i]; |
| 7331 | list_add_rcu(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list); |
| 7332 | } |
| 7333 | done: |
| 7334 | unlock_task_group_list(); |
| 7335 | return 0; |
| 7336 | } |
| 7337 | |
| 7338 | unsigned long sched_group_shares(struct task_group *tg) |
| 7339 | { |
| 7340 | return tg->shares; |
| 7341 | } |
| 7342 | |
| 7343 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
| 7344 | |
| 7345 | #ifdef CONFIG_FAIR_CGROUP_SCHED |
| 7346 | |
| 7347 | /* return corresponding task_group object of a cgroup */ |
| 7348 | static inline struct task_group *cgroup_tg(struct cgroup *cgrp) |
| 7349 | { |
| 7350 | return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id), |
| 7351 | struct task_group, css); |
| 7352 | } |
| 7353 | |
| 7354 | static struct cgroup_subsys_state * |
| 7355 | cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp) |
| 7356 | { |
| 7357 | struct task_group *tg; |
| 7358 | |
| 7359 | if (!cgrp->parent) { |
| 7360 | /* This is early initialization for the top cgroup */ |
| 7361 | init_task_group.css.cgroup = cgrp; |
| 7362 | return &init_task_group.css; |
| 7363 | } |
| 7364 | |
| 7365 | /* we support only 1-level deep hierarchical scheduler atm */ |
| 7366 | if (cgrp->parent->parent) |
| 7367 | return ERR_PTR(-EINVAL); |
| 7368 | |
| 7369 | tg = sched_create_group(); |
| 7370 | if (IS_ERR(tg)) |
| 7371 | return ERR_PTR(-ENOMEM); |
| 7372 | |
| 7373 | /* Bind the cgroup to task_group object we just created */ |
| 7374 | tg->css.cgroup = cgrp; |
| 7375 | |
| 7376 | return &tg->css; |
| 7377 | } |
| 7378 | |
| 7379 | static void |
| 7380 | cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp) |
| 7381 | { |
| 7382 | struct task_group *tg = cgroup_tg(cgrp); |
| 7383 | |
| 7384 | sched_destroy_group(tg); |
| 7385 | } |
| 7386 | |
| 7387 | static int |
| 7388 | cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp, |
| 7389 | struct task_struct *tsk) |
| 7390 | { |
| 7391 | /* We don't support RT-tasks being in separate groups */ |
| 7392 | if (tsk->sched_class != &fair_sched_class) |
| 7393 | return -EINVAL; |
| 7394 | |
| 7395 | return 0; |
| 7396 | } |
| 7397 | |
| 7398 | static void |
| 7399 | cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp, |
| 7400 | struct cgroup *old_cont, struct task_struct *tsk) |
| 7401 | { |
| 7402 | sched_move_task(tsk); |
| 7403 | } |
| 7404 | |
| 7405 | static int cpu_shares_write_uint(struct cgroup *cgrp, struct cftype *cftype, |
| 7406 | u64 shareval) |
| 7407 | { |
| 7408 | return sched_group_set_shares(cgroup_tg(cgrp), shareval); |
| 7409 | } |
| 7410 | |
| 7411 | static u64 cpu_shares_read_uint(struct cgroup *cgrp, struct cftype *cft) |
| 7412 | { |
| 7413 | struct task_group *tg = cgroup_tg(cgrp); |
| 7414 | |
| 7415 | return (u64) tg->shares; |
| 7416 | } |
| 7417 | |
| 7418 | static struct cftype cpu_files[] = { |
| 7419 | { |
| 7420 | .name = "shares", |
| 7421 | .read_uint = cpu_shares_read_uint, |
| 7422 | .write_uint = cpu_shares_write_uint, |
| 7423 | }, |
| 7424 | }; |
| 7425 | |
| 7426 | static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont) |
| 7427 | { |
| 7428 | return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files)); |
| 7429 | } |
| 7430 | |
| 7431 | struct cgroup_subsys cpu_cgroup_subsys = { |
| 7432 | .name = "cpu", |
| 7433 | .create = cpu_cgroup_create, |
| 7434 | .destroy = cpu_cgroup_destroy, |
| 7435 | .can_attach = cpu_cgroup_can_attach, |
| 7436 | .attach = cpu_cgroup_attach, |
| 7437 | .populate = cpu_cgroup_populate, |
| 7438 | .subsys_id = cpu_cgroup_subsys_id, |
| 7439 | .early_init = 1, |
| 7440 | }; |
| 7441 | |
| 7442 | #endif /* CONFIG_FAIR_CGROUP_SCHED */ |
| 7443 | |
| 7444 | #ifdef CONFIG_CGROUP_CPUACCT |
| 7445 | |
| 7446 | /* |
| 7447 | * CPU accounting code for task groups. |
| 7448 | * |
| 7449 | * Based on the work by Paul Menage (menage@google.com) and Balbir Singh |
| 7450 | * (balbir@in.ibm.com). |
| 7451 | */ |
| 7452 | |
| 7453 | /* track cpu usage of a group of tasks */ |
| 7454 | struct cpuacct { |
| 7455 | struct cgroup_subsys_state css; |
| 7456 | /* cpuusage holds pointer to a u64-type object on every cpu */ |
| 7457 | u64 *cpuusage; |
| 7458 | }; |
| 7459 | |
| 7460 | struct cgroup_subsys cpuacct_subsys; |
| 7461 | |
| 7462 | /* return cpu accounting group corresponding to this container */ |
| 7463 | static inline struct cpuacct *cgroup_ca(struct cgroup *cont) |
| 7464 | { |
| 7465 | return container_of(cgroup_subsys_state(cont, cpuacct_subsys_id), |
| 7466 | struct cpuacct, css); |
| 7467 | } |
| 7468 | |
| 7469 | /* return cpu accounting group to which this task belongs */ |
| 7470 | static inline struct cpuacct *task_ca(struct task_struct *tsk) |
| 7471 | { |
| 7472 | return container_of(task_subsys_state(tsk, cpuacct_subsys_id), |
| 7473 | struct cpuacct, css); |
| 7474 | } |
| 7475 | |
| 7476 | /* create a new cpu accounting group */ |
| 7477 | static struct cgroup_subsys_state *cpuacct_create( |
| 7478 | struct cgroup_subsys *ss, struct cgroup *cont) |
| 7479 | { |
| 7480 | struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL); |
| 7481 | |
| 7482 | if (!ca) |
| 7483 | return ERR_PTR(-ENOMEM); |
| 7484 | |
| 7485 | ca->cpuusage = alloc_percpu(u64); |
| 7486 | if (!ca->cpuusage) { |
| 7487 | kfree(ca); |
| 7488 | return ERR_PTR(-ENOMEM); |
| 7489 | } |
| 7490 | |
| 7491 | return &ca->css; |
| 7492 | } |
| 7493 | |
| 7494 | /* destroy an existing cpu accounting group */ |
| 7495 | static void |
| 7496 | cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cont) |
| 7497 | { |
| 7498 | struct cpuacct *ca = cgroup_ca(cont); |
| 7499 | |
| 7500 | free_percpu(ca->cpuusage); |
| 7501 | kfree(ca); |
| 7502 | } |
| 7503 | |
| 7504 | /* return total cpu usage (in nanoseconds) of a group */ |
| 7505 | static u64 cpuusage_read(struct cgroup *cont, struct cftype *cft) |
| 7506 | { |
| 7507 | struct cpuacct *ca = cgroup_ca(cont); |
| 7508 | u64 totalcpuusage = 0; |
| 7509 | int i; |
| 7510 | |
| 7511 | for_each_possible_cpu(i) { |
| 7512 | u64 *cpuusage = percpu_ptr(ca->cpuusage, i); |
| 7513 | |
| 7514 | /* |
| 7515 | * Take rq->lock to make 64-bit addition safe on 32-bit |
| 7516 | * platforms. |
| 7517 | */ |
| 7518 | spin_lock_irq(&cpu_rq(i)->lock); |
| 7519 | totalcpuusage += *cpuusage; |
| 7520 | spin_unlock_irq(&cpu_rq(i)->lock); |
| 7521 | } |
| 7522 | |
| 7523 | return totalcpuusage; |
| 7524 | } |
| 7525 | |
| 7526 | static struct cftype files[] = { |
| 7527 | { |
| 7528 | .name = "usage", |
| 7529 | .read_uint = cpuusage_read, |
| 7530 | }, |
| 7531 | }; |
| 7532 | |
| 7533 | static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cont) |
| 7534 | { |
| 7535 | return cgroup_add_files(cont, ss, files, ARRAY_SIZE(files)); |
| 7536 | } |
| 7537 | |
| 7538 | /* |
| 7539 | * charge this task's execution time to its accounting group. |
| 7540 | * |
| 7541 | * called with rq->lock held. |
| 7542 | */ |
| 7543 | static void cpuacct_charge(struct task_struct *tsk, u64 cputime) |
| 7544 | { |
| 7545 | struct cpuacct *ca; |
| 7546 | |
| 7547 | if (!cpuacct_subsys.active) |
| 7548 | return; |
| 7549 | |
| 7550 | ca = task_ca(tsk); |
| 7551 | if (ca) { |
| 7552 | u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk)); |
| 7553 | |
| 7554 | *cpuusage += cputime; |
| 7555 | } |
| 7556 | } |
| 7557 | |
| 7558 | struct cgroup_subsys cpuacct_subsys = { |
| 7559 | .name = "cpuacct", |
| 7560 | .create = cpuacct_create, |
| 7561 | .destroy = cpuacct_destroy, |
| 7562 | .populate = cpuacct_populate, |
| 7563 | .subsys_id = cpuacct_subsys_id, |
| 7564 | }; |
| 7565 | #endif /* CONFIG_CGROUP_CPUACCT */ |