| 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 | */ |
| 20 | |
| 21 | #include <linux/mm.h> |
| 22 | #include <linux/module.h> |
| 23 | #include <linux/nmi.h> |
| 24 | #include <linux/init.h> |
| 25 | #include <asm/uaccess.h> |
| 26 | #include <linux/highmem.h> |
| 27 | #include <linux/smp_lock.h> |
| 28 | #include <asm/mmu_context.h> |
| 29 | #include <linux/interrupt.h> |
| 30 | #include <linux/capability.h> |
| 31 | #include <linux/completion.h> |
| 32 | #include <linux/kernel_stat.h> |
| 33 | #include <linux/security.h> |
| 34 | #include <linux/notifier.h> |
| 35 | #include <linux/profile.h> |
| 36 | #include <linux/suspend.h> |
| 37 | #include <linux/vmalloc.h> |
| 38 | #include <linux/blkdev.h> |
| 39 | #include <linux/delay.h> |
| 40 | #include <linux/smp.h> |
| 41 | #include <linux/threads.h> |
| 42 | #include <linux/timer.h> |
| 43 | #include <linux/rcupdate.h> |
| 44 | #include <linux/cpu.h> |
| 45 | #include <linux/cpuset.h> |
| 46 | #include <linux/percpu.h> |
| 47 | #include <linux/kthread.h> |
| 48 | #include <linux/seq_file.h> |
| 49 | #include <linux/syscalls.h> |
| 50 | #include <linux/times.h> |
| 51 | #include <linux/acct.h> |
| 52 | #include <asm/tlb.h> |
| 53 | |
| 54 | #include <asm/unistd.h> |
| 55 | |
| 56 | /* |
| 57 | * Convert user-nice values [ -20 ... 0 ... 19 ] |
| 58 | * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ], |
| 59 | * and back. |
| 60 | */ |
| 61 | #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20) |
| 62 | #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20) |
| 63 | #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio) |
| 64 | |
| 65 | /* |
| 66 | * 'User priority' is the nice value converted to something we |
| 67 | * can work with better when scaling various scheduler parameters, |
| 68 | * it's a [ 0 ... 39 ] range. |
| 69 | */ |
| 70 | #define USER_PRIO(p) ((p)-MAX_RT_PRIO) |
| 71 | #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio) |
| 72 | #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO)) |
| 73 | |
| 74 | /* |
| 75 | * Some helpers for converting nanosecond timing to jiffy resolution |
| 76 | */ |
| 77 | #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ)) |
| 78 | #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ)) |
| 79 | |
| 80 | /* |
| 81 | * These are the 'tuning knobs' of the scheduler: |
| 82 | * |
| 83 | * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger), |
| 84 | * default timeslice is 100 msecs, maximum timeslice is 800 msecs. |
| 85 | * Timeslices get refilled after they expire. |
| 86 | */ |
| 87 | #define MIN_TIMESLICE max(5 * HZ / 1000, 1) |
| 88 | #define DEF_TIMESLICE (100 * HZ / 1000) |
| 89 | #define ON_RUNQUEUE_WEIGHT 30 |
| 90 | #define CHILD_PENALTY 95 |
| 91 | #define PARENT_PENALTY 100 |
| 92 | #define EXIT_WEIGHT 3 |
| 93 | #define PRIO_BONUS_RATIO 25 |
| 94 | #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100) |
| 95 | #define INTERACTIVE_DELTA 2 |
| 96 | #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS) |
| 97 | #define STARVATION_LIMIT (MAX_SLEEP_AVG) |
| 98 | #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG)) |
| 99 | |
| 100 | /* |
| 101 | * If a task is 'interactive' then we reinsert it in the active |
| 102 | * array after it has expired its current timeslice. (it will not |
| 103 | * continue to run immediately, it will still roundrobin with |
| 104 | * other interactive tasks.) |
| 105 | * |
| 106 | * This part scales the interactivity limit depending on niceness. |
| 107 | * |
| 108 | * We scale it linearly, offset by the INTERACTIVE_DELTA delta. |
| 109 | * Here are a few examples of different nice levels: |
| 110 | * |
| 111 | * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0] |
| 112 | * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0] |
| 113 | * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0] |
| 114 | * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0] |
| 115 | * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0] |
| 116 | * |
| 117 | * (the X axis represents the possible -5 ... 0 ... +5 dynamic |
| 118 | * priority range a task can explore, a value of '1' means the |
| 119 | * task is rated interactive.) |
| 120 | * |
| 121 | * Ie. nice +19 tasks can never get 'interactive' enough to be |
| 122 | * reinserted into the active array. And only heavily CPU-hog nice -20 |
| 123 | * tasks will be expired. Default nice 0 tasks are somewhere between, |
| 124 | * it takes some effort for them to get interactive, but it's not |
| 125 | * too hard. |
| 126 | */ |
| 127 | |
| 128 | #define CURRENT_BONUS(p) \ |
| 129 | (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \ |
| 130 | MAX_SLEEP_AVG) |
| 131 | |
| 132 | #define GRANULARITY (10 * HZ / 1000 ? : 1) |
| 133 | |
| 134 | #ifdef CONFIG_SMP |
| 135 | #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \ |
| 136 | (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \ |
| 137 | num_online_cpus()) |
| 138 | #else |
| 139 | #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \ |
| 140 | (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1))) |
| 141 | #endif |
| 142 | |
| 143 | #define SCALE(v1,v1_max,v2_max) \ |
| 144 | (v1) * (v2_max) / (v1_max) |
| 145 | |
| 146 | #define DELTA(p) \ |
| 147 | (SCALE(TASK_NICE(p), 40, MAX_BONUS) + INTERACTIVE_DELTA) |
| 148 | |
| 149 | #define TASK_INTERACTIVE(p) \ |
| 150 | ((p)->prio <= (p)->static_prio - DELTA(p)) |
| 151 | |
| 152 | #define INTERACTIVE_SLEEP(p) \ |
| 153 | (JIFFIES_TO_NS(MAX_SLEEP_AVG * \ |
| 154 | (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1)) |
| 155 | |
| 156 | #define TASK_PREEMPTS_CURR(p, rq) \ |
| 157 | ((p)->prio < (rq)->curr->prio) |
| 158 | |
| 159 | /* |
| 160 | * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ] |
| 161 | * to time slice values: [800ms ... 100ms ... 5ms] |
| 162 | * |
| 163 | * The higher a thread's priority, the bigger timeslices |
| 164 | * it gets during one round of execution. But even the lowest |
| 165 | * priority thread gets MIN_TIMESLICE worth of execution time. |
| 166 | */ |
| 167 | |
| 168 | #define SCALE_PRIO(x, prio) \ |
| 169 | max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO/2), MIN_TIMESLICE) |
| 170 | |
| 171 | static unsigned int task_timeslice(task_t *p) |
| 172 | { |
| 173 | if (p->static_prio < NICE_TO_PRIO(0)) |
| 174 | return SCALE_PRIO(DEF_TIMESLICE*4, p->static_prio); |
| 175 | else |
| 176 | return SCALE_PRIO(DEF_TIMESLICE, p->static_prio); |
| 177 | } |
| 178 | #define task_hot(p, now, sd) ((long long) ((now) - (p)->last_ran) \ |
| 179 | < (long long) (sd)->cache_hot_time) |
| 180 | |
| 181 | void __put_task_struct_cb(struct rcu_head *rhp) |
| 182 | { |
| 183 | __put_task_struct(container_of(rhp, struct task_struct, rcu)); |
| 184 | } |
| 185 | |
| 186 | EXPORT_SYMBOL_GPL(__put_task_struct_cb); |
| 187 | |
| 188 | /* |
| 189 | * These are the runqueue data structures: |
| 190 | */ |
| 191 | |
| 192 | #define BITMAP_SIZE ((((MAX_PRIO+1+7)/8)+sizeof(long)-1)/sizeof(long)) |
| 193 | |
| 194 | typedef struct runqueue runqueue_t; |
| 195 | |
| 196 | struct prio_array { |
| 197 | unsigned int nr_active; |
| 198 | unsigned long bitmap[BITMAP_SIZE]; |
| 199 | struct list_head queue[MAX_PRIO]; |
| 200 | }; |
| 201 | |
| 202 | /* |
| 203 | * This is the main, per-CPU runqueue data structure. |
| 204 | * |
| 205 | * Locking rule: those places that want to lock multiple runqueues |
| 206 | * (such as the load balancing or the thread migration code), lock |
| 207 | * acquire operations must be ordered by ascending &runqueue. |
| 208 | */ |
| 209 | struct runqueue { |
| 210 | spinlock_t lock; |
| 211 | |
| 212 | /* |
| 213 | * nr_running and cpu_load should be in the same cacheline because |
| 214 | * remote CPUs use both these fields when doing load calculation. |
| 215 | */ |
| 216 | unsigned long nr_running; |
| 217 | #ifdef CONFIG_SMP |
| 218 | unsigned long prio_bias; |
| 219 | unsigned long cpu_load[3]; |
| 220 | #endif |
| 221 | unsigned long long nr_switches; |
| 222 | |
| 223 | /* |
| 224 | * This is part of a global counter where only the total sum |
| 225 | * over all CPUs matters. A task can increase this counter on |
| 226 | * one CPU and if it got migrated afterwards it may decrease |
| 227 | * it on another CPU. Always updated under the runqueue lock: |
| 228 | */ |
| 229 | unsigned long nr_uninterruptible; |
| 230 | |
| 231 | unsigned long expired_timestamp; |
| 232 | unsigned long long timestamp_last_tick; |
| 233 | task_t *curr, *idle; |
| 234 | struct mm_struct *prev_mm; |
| 235 | prio_array_t *active, *expired, arrays[2]; |
| 236 | int best_expired_prio; |
| 237 | atomic_t nr_iowait; |
| 238 | |
| 239 | #ifdef CONFIG_SMP |
| 240 | struct sched_domain *sd; |
| 241 | |
| 242 | /* For active balancing */ |
| 243 | int active_balance; |
| 244 | int push_cpu; |
| 245 | |
| 246 | task_t *migration_thread; |
| 247 | struct list_head migration_queue; |
| 248 | #endif |
| 249 | |
| 250 | #ifdef CONFIG_SCHEDSTATS |
| 251 | /* latency stats */ |
| 252 | struct sched_info rq_sched_info; |
| 253 | |
| 254 | /* sys_sched_yield() stats */ |
| 255 | unsigned long yld_exp_empty; |
| 256 | unsigned long yld_act_empty; |
| 257 | unsigned long yld_both_empty; |
| 258 | unsigned long yld_cnt; |
| 259 | |
| 260 | /* schedule() stats */ |
| 261 | unsigned long sched_switch; |
| 262 | unsigned long sched_cnt; |
| 263 | unsigned long sched_goidle; |
| 264 | |
| 265 | /* try_to_wake_up() stats */ |
| 266 | unsigned long ttwu_cnt; |
| 267 | unsigned long ttwu_local; |
| 268 | #endif |
| 269 | }; |
| 270 | |
| 271 | static DEFINE_PER_CPU(struct runqueue, runqueues); |
| 272 | |
| 273 | /* |
| 274 | * The domain tree (rq->sd) is protected by RCU's quiescent state transition. |
| 275 | * See detach_destroy_domains: synchronize_sched for details. |
| 276 | * |
| 277 | * The domain tree of any CPU may only be accessed from within |
| 278 | * preempt-disabled sections. |
| 279 | */ |
| 280 | #define for_each_domain(cpu, domain) \ |
| 281 | for (domain = rcu_dereference(cpu_rq(cpu)->sd); domain; domain = domain->parent) |
| 282 | |
| 283 | #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu))) |
| 284 | #define this_rq() (&__get_cpu_var(runqueues)) |
| 285 | #define task_rq(p) cpu_rq(task_cpu(p)) |
| 286 | #define cpu_curr(cpu) (cpu_rq(cpu)->curr) |
| 287 | |
| 288 | #ifndef prepare_arch_switch |
| 289 | # define prepare_arch_switch(next) do { } while (0) |
| 290 | #endif |
| 291 | #ifndef finish_arch_switch |
| 292 | # define finish_arch_switch(prev) do { } while (0) |
| 293 | #endif |
| 294 | |
| 295 | #ifndef __ARCH_WANT_UNLOCKED_CTXSW |
| 296 | static inline int task_running(runqueue_t *rq, task_t *p) |
| 297 | { |
| 298 | return rq->curr == p; |
| 299 | } |
| 300 | |
| 301 | static inline void prepare_lock_switch(runqueue_t *rq, task_t *next) |
| 302 | { |
| 303 | } |
| 304 | |
| 305 | static inline void finish_lock_switch(runqueue_t *rq, task_t *prev) |
| 306 | { |
| 307 | #ifdef CONFIG_DEBUG_SPINLOCK |
| 308 | /* this is a valid case when another task releases the spinlock */ |
| 309 | rq->lock.owner = current; |
| 310 | #endif |
| 311 | spin_unlock_irq(&rq->lock); |
| 312 | } |
| 313 | |
| 314 | #else /* __ARCH_WANT_UNLOCKED_CTXSW */ |
| 315 | static inline int task_running(runqueue_t *rq, task_t *p) |
| 316 | { |
| 317 | #ifdef CONFIG_SMP |
| 318 | return p->oncpu; |
| 319 | #else |
| 320 | return rq->curr == p; |
| 321 | #endif |
| 322 | } |
| 323 | |
| 324 | static inline void prepare_lock_switch(runqueue_t *rq, task_t *next) |
| 325 | { |
| 326 | #ifdef CONFIG_SMP |
| 327 | /* |
| 328 | * We can optimise this out completely for !SMP, because the |
| 329 | * SMP rebalancing from interrupt is the only thing that cares |
| 330 | * here. |
| 331 | */ |
| 332 | next->oncpu = 1; |
| 333 | #endif |
| 334 | #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW |
| 335 | spin_unlock_irq(&rq->lock); |
| 336 | #else |
| 337 | spin_unlock(&rq->lock); |
| 338 | #endif |
| 339 | } |
| 340 | |
| 341 | static inline void finish_lock_switch(runqueue_t *rq, task_t *prev) |
| 342 | { |
| 343 | #ifdef CONFIG_SMP |
| 344 | /* |
| 345 | * After ->oncpu is cleared, the task can be moved to a different CPU. |
| 346 | * We must ensure this doesn't happen until the switch is completely |
| 347 | * finished. |
| 348 | */ |
| 349 | smp_wmb(); |
| 350 | prev->oncpu = 0; |
| 351 | #endif |
| 352 | #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW |
| 353 | local_irq_enable(); |
| 354 | #endif |
| 355 | } |
| 356 | #endif /* __ARCH_WANT_UNLOCKED_CTXSW */ |
| 357 | |
| 358 | /* |
| 359 | * task_rq_lock - lock the runqueue a given task resides on and disable |
| 360 | * interrupts. Note the ordering: we can safely lookup the task_rq without |
| 361 | * explicitly disabling preemption. |
| 362 | */ |
| 363 | static inline runqueue_t *task_rq_lock(task_t *p, unsigned long *flags) |
| 364 | __acquires(rq->lock) |
| 365 | { |
| 366 | struct runqueue *rq; |
| 367 | |
| 368 | repeat_lock_task: |
| 369 | local_irq_save(*flags); |
| 370 | rq = task_rq(p); |
| 371 | spin_lock(&rq->lock); |
| 372 | if (unlikely(rq != task_rq(p))) { |
| 373 | spin_unlock_irqrestore(&rq->lock, *flags); |
| 374 | goto repeat_lock_task; |
| 375 | } |
| 376 | return rq; |
| 377 | } |
| 378 | |
| 379 | static inline void task_rq_unlock(runqueue_t *rq, unsigned long *flags) |
| 380 | __releases(rq->lock) |
| 381 | { |
| 382 | spin_unlock_irqrestore(&rq->lock, *flags); |
| 383 | } |
| 384 | |
| 385 | #ifdef CONFIG_SCHEDSTATS |
| 386 | /* |
| 387 | * bump this up when changing the output format or the meaning of an existing |
| 388 | * format, so that tools can adapt (or abort) |
| 389 | */ |
| 390 | #define SCHEDSTAT_VERSION 12 |
| 391 | |
| 392 | static int show_schedstat(struct seq_file *seq, void *v) |
| 393 | { |
| 394 | int cpu; |
| 395 | |
| 396 | seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION); |
| 397 | seq_printf(seq, "timestamp %lu\n", jiffies); |
| 398 | for_each_online_cpu(cpu) { |
| 399 | runqueue_t *rq = cpu_rq(cpu); |
| 400 | #ifdef CONFIG_SMP |
| 401 | struct sched_domain *sd; |
| 402 | int dcnt = 0; |
| 403 | #endif |
| 404 | |
| 405 | /* runqueue-specific stats */ |
| 406 | seq_printf(seq, |
| 407 | "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu", |
| 408 | cpu, rq->yld_both_empty, |
| 409 | rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt, |
| 410 | rq->sched_switch, rq->sched_cnt, rq->sched_goidle, |
| 411 | rq->ttwu_cnt, rq->ttwu_local, |
| 412 | rq->rq_sched_info.cpu_time, |
| 413 | rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt); |
| 414 | |
| 415 | seq_printf(seq, "\n"); |
| 416 | |
| 417 | #ifdef CONFIG_SMP |
| 418 | /* domain-specific stats */ |
| 419 | preempt_disable(); |
| 420 | for_each_domain(cpu, sd) { |
| 421 | enum idle_type itype; |
| 422 | char mask_str[NR_CPUS]; |
| 423 | |
| 424 | cpumask_scnprintf(mask_str, NR_CPUS, sd->span); |
| 425 | seq_printf(seq, "domain%d %s", dcnt++, mask_str); |
| 426 | for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES; |
| 427 | itype++) { |
| 428 | seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu", |
| 429 | sd->lb_cnt[itype], |
| 430 | sd->lb_balanced[itype], |
| 431 | sd->lb_failed[itype], |
| 432 | sd->lb_imbalance[itype], |
| 433 | sd->lb_gained[itype], |
| 434 | sd->lb_hot_gained[itype], |
| 435 | sd->lb_nobusyq[itype], |
| 436 | sd->lb_nobusyg[itype]); |
| 437 | } |
| 438 | seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n", |
| 439 | sd->alb_cnt, sd->alb_failed, sd->alb_pushed, |
| 440 | sd->sbe_cnt, sd->sbe_balanced, sd->sbe_pushed, |
| 441 | sd->sbf_cnt, sd->sbf_balanced, sd->sbf_pushed, |
| 442 | sd->ttwu_wake_remote, sd->ttwu_move_affine, sd->ttwu_move_balance); |
| 443 | } |
| 444 | preempt_enable(); |
| 445 | #endif |
| 446 | } |
| 447 | return 0; |
| 448 | } |
| 449 | |
| 450 | static int schedstat_open(struct inode *inode, struct file *file) |
| 451 | { |
| 452 | unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32); |
| 453 | char *buf = kmalloc(size, GFP_KERNEL); |
| 454 | struct seq_file *m; |
| 455 | int res; |
| 456 | |
| 457 | if (!buf) |
| 458 | return -ENOMEM; |
| 459 | res = single_open(file, show_schedstat, NULL); |
| 460 | if (!res) { |
| 461 | m = file->private_data; |
| 462 | m->buf = buf; |
| 463 | m->size = size; |
| 464 | } else |
| 465 | kfree(buf); |
| 466 | return res; |
| 467 | } |
| 468 | |
| 469 | struct file_operations proc_schedstat_operations = { |
| 470 | .open = schedstat_open, |
| 471 | .read = seq_read, |
| 472 | .llseek = seq_lseek, |
| 473 | .release = single_release, |
| 474 | }; |
| 475 | |
| 476 | # define schedstat_inc(rq, field) do { (rq)->field++; } while (0) |
| 477 | # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0) |
| 478 | #else /* !CONFIG_SCHEDSTATS */ |
| 479 | # define schedstat_inc(rq, field) do { } while (0) |
| 480 | # define schedstat_add(rq, field, amt) do { } while (0) |
| 481 | #endif |
| 482 | |
| 483 | /* |
| 484 | * rq_lock - lock a given runqueue and disable interrupts. |
| 485 | */ |
| 486 | static inline runqueue_t *this_rq_lock(void) |
| 487 | __acquires(rq->lock) |
| 488 | { |
| 489 | runqueue_t *rq; |
| 490 | |
| 491 | local_irq_disable(); |
| 492 | rq = this_rq(); |
| 493 | spin_lock(&rq->lock); |
| 494 | |
| 495 | return rq; |
| 496 | } |
| 497 | |
| 498 | #ifdef CONFIG_SCHEDSTATS |
| 499 | /* |
| 500 | * Called when a process is dequeued from the active array and given |
| 501 | * the cpu. We should note that with the exception of interactive |
| 502 | * tasks, the expired queue will become the active queue after the active |
| 503 | * queue is empty, without explicitly dequeuing and requeuing tasks in the |
| 504 | * expired queue. (Interactive tasks may be requeued directly to the |
| 505 | * active queue, thus delaying tasks in the expired queue from running; |
| 506 | * see scheduler_tick()). |
| 507 | * |
| 508 | * This function is only called from sched_info_arrive(), rather than |
| 509 | * dequeue_task(). Even though a task may be queued and dequeued multiple |
| 510 | * times as it is shuffled about, we're really interested in knowing how |
| 511 | * long it was from the *first* time it was queued to the time that it |
| 512 | * finally hit a cpu. |
| 513 | */ |
| 514 | static inline void sched_info_dequeued(task_t *t) |
| 515 | { |
| 516 | t->sched_info.last_queued = 0; |
| 517 | } |
| 518 | |
| 519 | /* |
| 520 | * Called when a task finally hits the cpu. We can now calculate how |
| 521 | * long it was waiting to run. We also note when it began so that we |
| 522 | * can keep stats on how long its timeslice is. |
| 523 | */ |
| 524 | static void sched_info_arrive(task_t *t) |
| 525 | { |
| 526 | unsigned long now = jiffies, diff = 0; |
| 527 | struct runqueue *rq = task_rq(t); |
| 528 | |
| 529 | if (t->sched_info.last_queued) |
| 530 | diff = now - t->sched_info.last_queued; |
| 531 | sched_info_dequeued(t); |
| 532 | t->sched_info.run_delay += diff; |
| 533 | t->sched_info.last_arrival = now; |
| 534 | t->sched_info.pcnt++; |
| 535 | |
| 536 | if (!rq) |
| 537 | return; |
| 538 | |
| 539 | rq->rq_sched_info.run_delay += diff; |
| 540 | rq->rq_sched_info.pcnt++; |
| 541 | } |
| 542 | |
| 543 | /* |
| 544 | * Called when a process is queued into either the active or expired |
| 545 | * array. The time is noted and later used to determine how long we |
| 546 | * had to wait for us to reach the cpu. Since the expired queue will |
| 547 | * become the active queue after active queue is empty, without dequeuing |
| 548 | * and requeuing any tasks, we are interested in queuing to either. It |
| 549 | * is unusual but not impossible for tasks to be dequeued and immediately |
| 550 | * requeued in the same or another array: this can happen in sched_yield(), |
| 551 | * set_user_nice(), and even load_balance() as it moves tasks from runqueue |
| 552 | * to runqueue. |
| 553 | * |
| 554 | * This function is only called from enqueue_task(), but also only updates |
| 555 | * the timestamp if it is already not set. It's assumed that |
| 556 | * sched_info_dequeued() will clear that stamp when appropriate. |
| 557 | */ |
| 558 | static inline void sched_info_queued(task_t *t) |
| 559 | { |
| 560 | if (!t->sched_info.last_queued) |
| 561 | t->sched_info.last_queued = jiffies; |
| 562 | } |
| 563 | |
| 564 | /* |
| 565 | * Called when a process ceases being the active-running process, either |
| 566 | * voluntarily or involuntarily. Now we can calculate how long we ran. |
| 567 | */ |
| 568 | static inline void sched_info_depart(task_t *t) |
| 569 | { |
| 570 | struct runqueue *rq = task_rq(t); |
| 571 | unsigned long diff = jiffies - t->sched_info.last_arrival; |
| 572 | |
| 573 | t->sched_info.cpu_time += diff; |
| 574 | |
| 575 | if (rq) |
| 576 | rq->rq_sched_info.cpu_time += diff; |
| 577 | } |
| 578 | |
| 579 | /* |
| 580 | * Called when tasks are switched involuntarily due, typically, to expiring |
| 581 | * their time slice. (This may also be called when switching to or from |
| 582 | * the idle task.) We are only called when prev != next. |
| 583 | */ |
| 584 | static inline void sched_info_switch(task_t *prev, task_t *next) |
| 585 | { |
| 586 | struct runqueue *rq = task_rq(prev); |
| 587 | |
| 588 | /* |
| 589 | * prev now departs the cpu. It's not interesting to record |
| 590 | * stats about how efficient we were at scheduling the idle |
| 591 | * process, however. |
| 592 | */ |
| 593 | if (prev != rq->idle) |
| 594 | sched_info_depart(prev); |
| 595 | |
| 596 | if (next != rq->idle) |
| 597 | sched_info_arrive(next); |
| 598 | } |
| 599 | #else |
| 600 | #define sched_info_queued(t) do { } while (0) |
| 601 | #define sched_info_switch(t, next) do { } while (0) |
| 602 | #endif /* CONFIG_SCHEDSTATS */ |
| 603 | |
| 604 | /* |
| 605 | * Adding/removing a task to/from a priority array: |
| 606 | */ |
| 607 | static void dequeue_task(struct task_struct *p, prio_array_t *array) |
| 608 | { |
| 609 | array->nr_active--; |
| 610 | list_del(&p->run_list); |
| 611 | if (list_empty(array->queue + p->prio)) |
| 612 | __clear_bit(p->prio, array->bitmap); |
| 613 | } |
| 614 | |
| 615 | static void enqueue_task(struct task_struct *p, prio_array_t *array) |
| 616 | { |
| 617 | sched_info_queued(p); |
| 618 | list_add_tail(&p->run_list, array->queue + p->prio); |
| 619 | __set_bit(p->prio, array->bitmap); |
| 620 | array->nr_active++; |
| 621 | p->array = array; |
| 622 | } |
| 623 | |
| 624 | /* |
| 625 | * Put task to the end of the run list without the overhead of dequeue |
| 626 | * followed by enqueue. |
| 627 | */ |
| 628 | static void requeue_task(struct task_struct *p, prio_array_t *array) |
| 629 | { |
| 630 | list_move_tail(&p->run_list, array->queue + p->prio); |
| 631 | } |
| 632 | |
| 633 | static inline void enqueue_task_head(struct task_struct *p, prio_array_t *array) |
| 634 | { |
| 635 | list_add(&p->run_list, array->queue + p->prio); |
| 636 | __set_bit(p->prio, array->bitmap); |
| 637 | array->nr_active++; |
| 638 | p->array = array; |
| 639 | } |
| 640 | |
| 641 | /* |
| 642 | * effective_prio - return the priority that is based on the static |
| 643 | * priority but is modified by bonuses/penalties. |
| 644 | * |
| 645 | * We scale the actual sleep average [0 .... MAX_SLEEP_AVG] |
| 646 | * into the -5 ... 0 ... +5 bonus/penalty range. |
| 647 | * |
| 648 | * We use 25% of the full 0...39 priority range so that: |
| 649 | * |
| 650 | * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs. |
| 651 | * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks. |
| 652 | * |
| 653 | * Both properties are important to certain workloads. |
| 654 | */ |
| 655 | static int effective_prio(task_t *p) |
| 656 | { |
| 657 | int bonus, prio; |
| 658 | |
| 659 | if (rt_task(p)) |
| 660 | return p->prio; |
| 661 | |
| 662 | bonus = CURRENT_BONUS(p) - MAX_BONUS / 2; |
| 663 | |
| 664 | prio = p->static_prio - bonus; |
| 665 | if (prio < MAX_RT_PRIO) |
| 666 | prio = MAX_RT_PRIO; |
| 667 | if (prio > MAX_PRIO-1) |
| 668 | prio = MAX_PRIO-1; |
| 669 | return prio; |
| 670 | } |
| 671 | |
| 672 | #ifdef CONFIG_SMP |
| 673 | static inline void inc_prio_bias(runqueue_t *rq, int prio) |
| 674 | { |
| 675 | rq->prio_bias += MAX_PRIO - prio; |
| 676 | } |
| 677 | |
| 678 | static inline void dec_prio_bias(runqueue_t *rq, int prio) |
| 679 | { |
| 680 | rq->prio_bias -= MAX_PRIO - prio; |
| 681 | } |
| 682 | |
| 683 | static inline void inc_nr_running(task_t *p, runqueue_t *rq) |
| 684 | { |
| 685 | rq->nr_running++; |
| 686 | if (rt_task(p)) { |
| 687 | if (p != rq->migration_thread) |
| 688 | /* |
| 689 | * The migration thread does the actual balancing. Do |
| 690 | * not bias by its priority as the ultra high priority |
| 691 | * will skew balancing adversely. |
| 692 | */ |
| 693 | inc_prio_bias(rq, p->prio); |
| 694 | } else |
| 695 | inc_prio_bias(rq, p->static_prio); |
| 696 | } |
| 697 | |
| 698 | static inline void dec_nr_running(task_t *p, runqueue_t *rq) |
| 699 | { |
| 700 | rq->nr_running--; |
| 701 | if (rt_task(p)) { |
| 702 | if (p != rq->migration_thread) |
| 703 | dec_prio_bias(rq, p->prio); |
| 704 | } else |
| 705 | dec_prio_bias(rq, p->static_prio); |
| 706 | } |
| 707 | #else |
| 708 | static inline void inc_prio_bias(runqueue_t *rq, int prio) |
| 709 | { |
| 710 | } |
| 711 | |
| 712 | static inline void dec_prio_bias(runqueue_t *rq, int prio) |
| 713 | { |
| 714 | } |
| 715 | |
| 716 | static inline void inc_nr_running(task_t *p, runqueue_t *rq) |
| 717 | { |
| 718 | rq->nr_running++; |
| 719 | } |
| 720 | |
| 721 | static inline void dec_nr_running(task_t *p, runqueue_t *rq) |
| 722 | { |
| 723 | rq->nr_running--; |
| 724 | } |
| 725 | #endif |
| 726 | |
| 727 | /* |
| 728 | * __activate_task - move a task to the runqueue. |
| 729 | */ |
| 730 | static inline void __activate_task(task_t *p, runqueue_t *rq) |
| 731 | { |
| 732 | enqueue_task(p, rq->active); |
| 733 | inc_nr_running(p, rq); |
| 734 | } |
| 735 | |
| 736 | /* |
| 737 | * __activate_idle_task - move idle task to the _front_ of runqueue. |
| 738 | */ |
| 739 | static inline void __activate_idle_task(task_t *p, runqueue_t *rq) |
| 740 | { |
| 741 | enqueue_task_head(p, rq->active); |
| 742 | inc_nr_running(p, rq); |
| 743 | } |
| 744 | |
| 745 | static int recalc_task_prio(task_t *p, unsigned long long now) |
| 746 | { |
| 747 | /* Caller must always ensure 'now >= p->timestamp' */ |
| 748 | unsigned long long __sleep_time = now - p->timestamp; |
| 749 | unsigned long sleep_time; |
| 750 | |
| 751 | if (unlikely(p->policy == SCHED_BATCH)) |
| 752 | sleep_time = 0; |
| 753 | else { |
| 754 | if (__sleep_time > NS_MAX_SLEEP_AVG) |
| 755 | sleep_time = NS_MAX_SLEEP_AVG; |
| 756 | else |
| 757 | sleep_time = (unsigned long)__sleep_time; |
| 758 | } |
| 759 | |
| 760 | if (likely(sleep_time > 0)) { |
| 761 | /* |
| 762 | * User tasks that sleep a long time are categorised as |
| 763 | * idle and will get just interactive status to stay active & |
| 764 | * prevent them suddenly becoming cpu hogs and starving |
| 765 | * other processes. |
| 766 | */ |
| 767 | if (p->mm && p->activated != -1 && |
| 768 | sleep_time > INTERACTIVE_SLEEP(p)) { |
| 769 | p->sleep_avg = JIFFIES_TO_NS(MAX_SLEEP_AVG - |
| 770 | DEF_TIMESLICE); |
| 771 | } else { |
| 772 | /* |
| 773 | * The lower the sleep avg a task has the more |
| 774 | * rapidly it will rise with sleep time. |
| 775 | */ |
| 776 | sleep_time *= (MAX_BONUS - CURRENT_BONUS(p)) ? : 1; |
| 777 | |
| 778 | /* |
| 779 | * Tasks waking from uninterruptible sleep are |
| 780 | * limited in their sleep_avg rise as they |
| 781 | * are likely to be waiting on I/O |
| 782 | */ |
| 783 | if (p->activated == -1 && p->mm) { |
| 784 | if (p->sleep_avg >= INTERACTIVE_SLEEP(p)) |
| 785 | sleep_time = 0; |
| 786 | else if (p->sleep_avg + sleep_time >= |
| 787 | INTERACTIVE_SLEEP(p)) { |
| 788 | p->sleep_avg = INTERACTIVE_SLEEP(p); |
| 789 | sleep_time = 0; |
| 790 | } |
| 791 | } |
| 792 | |
| 793 | /* |
| 794 | * This code gives a bonus to interactive tasks. |
| 795 | * |
| 796 | * The boost works by updating the 'average sleep time' |
| 797 | * value here, based on ->timestamp. The more time a |
| 798 | * task spends sleeping, the higher the average gets - |
| 799 | * and the higher the priority boost gets as well. |
| 800 | */ |
| 801 | p->sleep_avg += sleep_time; |
| 802 | |
| 803 | if (p->sleep_avg > NS_MAX_SLEEP_AVG) |
| 804 | p->sleep_avg = NS_MAX_SLEEP_AVG; |
| 805 | } |
| 806 | } |
| 807 | |
| 808 | return effective_prio(p); |
| 809 | } |
| 810 | |
| 811 | /* |
| 812 | * activate_task - move a task to the runqueue and do priority recalculation |
| 813 | * |
| 814 | * Update all the scheduling statistics stuff. (sleep average |
| 815 | * calculation, priority modifiers, etc.) |
| 816 | */ |
| 817 | static void activate_task(task_t *p, runqueue_t *rq, int local) |
| 818 | { |
| 819 | unsigned long long now; |
| 820 | |
| 821 | now = sched_clock(); |
| 822 | #ifdef CONFIG_SMP |
| 823 | if (!local) { |
| 824 | /* Compensate for drifting sched_clock */ |
| 825 | runqueue_t *this_rq = this_rq(); |
| 826 | now = (now - this_rq->timestamp_last_tick) |
| 827 | + rq->timestamp_last_tick; |
| 828 | } |
| 829 | #endif |
| 830 | |
| 831 | if (!rt_task(p)) |
| 832 | p->prio = recalc_task_prio(p, now); |
| 833 | |
| 834 | /* |
| 835 | * This checks to make sure it's not an uninterruptible task |
| 836 | * that is now waking up. |
| 837 | */ |
| 838 | if (!p->activated) { |
| 839 | /* |
| 840 | * Tasks which were woken up by interrupts (ie. hw events) |
| 841 | * are most likely of interactive nature. So we give them |
| 842 | * the credit of extending their sleep time to the period |
| 843 | * of time they spend on the runqueue, waiting for execution |
| 844 | * on a CPU, first time around: |
| 845 | */ |
| 846 | if (in_interrupt()) |
| 847 | p->activated = 2; |
| 848 | else { |
| 849 | /* |
| 850 | * Normal first-time wakeups get a credit too for |
| 851 | * on-runqueue time, but it will be weighted down: |
| 852 | */ |
| 853 | p->activated = 1; |
| 854 | } |
| 855 | } |
| 856 | p->timestamp = now; |
| 857 | |
| 858 | __activate_task(p, rq); |
| 859 | } |
| 860 | |
| 861 | /* |
| 862 | * deactivate_task - remove a task from the runqueue. |
| 863 | */ |
| 864 | static void deactivate_task(struct task_struct *p, runqueue_t *rq) |
| 865 | { |
| 866 | dec_nr_running(p, rq); |
| 867 | dequeue_task(p, p->array); |
| 868 | p->array = NULL; |
| 869 | } |
| 870 | |
| 871 | /* |
| 872 | * resched_task - mark a task 'to be rescheduled now'. |
| 873 | * |
| 874 | * On UP this means the setting of the need_resched flag, on SMP it |
| 875 | * might also involve a cross-CPU call to trigger the scheduler on |
| 876 | * the target CPU. |
| 877 | */ |
| 878 | #ifdef CONFIG_SMP |
| 879 | static void resched_task(task_t *p) |
| 880 | { |
| 881 | int cpu; |
| 882 | |
| 883 | assert_spin_locked(&task_rq(p)->lock); |
| 884 | |
| 885 | if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED))) |
| 886 | return; |
| 887 | |
| 888 | set_tsk_thread_flag(p, TIF_NEED_RESCHED); |
| 889 | |
| 890 | cpu = task_cpu(p); |
| 891 | if (cpu == smp_processor_id()) |
| 892 | return; |
| 893 | |
| 894 | /* NEED_RESCHED must be visible before we test POLLING_NRFLAG */ |
| 895 | smp_mb(); |
| 896 | if (!test_tsk_thread_flag(p, TIF_POLLING_NRFLAG)) |
| 897 | smp_send_reschedule(cpu); |
| 898 | } |
| 899 | #else |
| 900 | static inline void resched_task(task_t *p) |
| 901 | { |
| 902 | assert_spin_locked(&task_rq(p)->lock); |
| 903 | set_tsk_need_resched(p); |
| 904 | } |
| 905 | #endif |
| 906 | |
| 907 | /** |
| 908 | * task_curr - is this task currently executing on a CPU? |
| 909 | * @p: the task in question. |
| 910 | */ |
| 911 | inline int task_curr(const task_t *p) |
| 912 | { |
| 913 | return cpu_curr(task_cpu(p)) == p; |
| 914 | } |
| 915 | |
| 916 | #ifdef CONFIG_SMP |
| 917 | typedef struct { |
| 918 | struct list_head list; |
| 919 | |
| 920 | task_t *task; |
| 921 | int dest_cpu; |
| 922 | |
| 923 | struct completion done; |
| 924 | } migration_req_t; |
| 925 | |
| 926 | /* |
| 927 | * The task's runqueue lock must be held. |
| 928 | * Returns true if you have to wait for migration thread. |
| 929 | */ |
| 930 | static int migrate_task(task_t *p, int dest_cpu, migration_req_t *req) |
| 931 | { |
| 932 | runqueue_t *rq = task_rq(p); |
| 933 | |
| 934 | /* |
| 935 | * If the task is not on a runqueue (and not running), then |
| 936 | * it is sufficient to simply update the task's cpu field. |
| 937 | */ |
| 938 | if (!p->array && !task_running(rq, p)) { |
| 939 | set_task_cpu(p, dest_cpu); |
| 940 | return 0; |
| 941 | } |
| 942 | |
| 943 | init_completion(&req->done); |
| 944 | req->task = p; |
| 945 | req->dest_cpu = dest_cpu; |
| 946 | list_add(&req->list, &rq->migration_queue); |
| 947 | return 1; |
| 948 | } |
| 949 | |
| 950 | /* |
| 951 | * wait_task_inactive - wait for a thread to unschedule. |
| 952 | * |
| 953 | * The caller must ensure that the task *will* unschedule sometime soon, |
| 954 | * else this function might spin for a *long* time. This function can't |
| 955 | * be called with interrupts off, or it may introduce deadlock with |
| 956 | * smp_call_function() if an IPI is sent by the same process we are |
| 957 | * waiting to become inactive. |
| 958 | */ |
| 959 | void wait_task_inactive(task_t *p) |
| 960 | { |
| 961 | unsigned long flags; |
| 962 | runqueue_t *rq; |
| 963 | int preempted; |
| 964 | |
| 965 | repeat: |
| 966 | rq = task_rq_lock(p, &flags); |
| 967 | /* Must be off runqueue entirely, not preempted. */ |
| 968 | if (unlikely(p->array || task_running(rq, p))) { |
| 969 | /* If it's preempted, we yield. It could be a while. */ |
| 970 | preempted = !task_running(rq, p); |
| 971 | task_rq_unlock(rq, &flags); |
| 972 | cpu_relax(); |
| 973 | if (preempted) |
| 974 | yield(); |
| 975 | goto repeat; |
| 976 | } |
| 977 | task_rq_unlock(rq, &flags); |
| 978 | } |
| 979 | |
| 980 | /*** |
| 981 | * kick_process - kick a running thread to enter/exit the kernel |
| 982 | * @p: the to-be-kicked thread |
| 983 | * |
| 984 | * Cause a process which is running on another CPU to enter |
| 985 | * kernel-mode, without any delay. (to get signals handled.) |
| 986 | * |
| 987 | * NOTE: this function doesnt have to take the runqueue lock, |
| 988 | * because all it wants to ensure is that the remote task enters |
| 989 | * the kernel. If the IPI races and the task has been migrated |
| 990 | * to another CPU then no harm is done and the purpose has been |
| 991 | * achieved as well. |
| 992 | */ |
| 993 | void kick_process(task_t *p) |
| 994 | { |
| 995 | int cpu; |
| 996 | |
| 997 | preempt_disable(); |
| 998 | cpu = task_cpu(p); |
| 999 | if ((cpu != smp_processor_id()) && task_curr(p)) |
| 1000 | smp_send_reschedule(cpu); |
| 1001 | preempt_enable(); |
| 1002 | } |
| 1003 | |
| 1004 | /* |
| 1005 | * Return a low guess at the load of a migration-source cpu. |
| 1006 | * |
| 1007 | * We want to under-estimate the load of migration sources, to |
| 1008 | * balance conservatively. |
| 1009 | */ |
| 1010 | static unsigned long __source_load(int cpu, int type, enum idle_type idle) |
| 1011 | { |
| 1012 | runqueue_t *rq = cpu_rq(cpu); |
| 1013 | unsigned long running = rq->nr_running; |
| 1014 | unsigned long source_load, cpu_load = rq->cpu_load[type-1], |
| 1015 | load_now = running * SCHED_LOAD_SCALE; |
| 1016 | |
| 1017 | if (type == 0) |
| 1018 | source_load = load_now; |
| 1019 | else |
| 1020 | source_load = min(cpu_load, load_now); |
| 1021 | |
| 1022 | if (running > 1 || (idle == NOT_IDLE && running)) |
| 1023 | /* |
| 1024 | * If we are busy rebalancing the load is biased by |
| 1025 | * priority to create 'nice' support across cpus. When |
| 1026 | * idle rebalancing we should only bias the source_load if |
| 1027 | * there is more than one task running on that queue to |
| 1028 | * prevent idle rebalance from trying to pull tasks from a |
| 1029 | * queue with only one running task. |
| 1030 | */ |
| 1031 | source_load = source_load * rq->prio_bias / running; |
| 1032 | |
| 1033 | return source_load; |
| 1034 | } |
| 1035 | |
| 1036 | static inline unsigned long source_load(int cpu, int type) |
| 1037 | { |
| 1038 | return __source_load(cpu, type, NOT_IDLE); |
| 1039 | } |
| 1040 | |
| 1041 | /* |
| 1042 | * Return a high guess at the load of a migration-target cpu |
| 1043 | */ |
| 1044 | static inline unsigned long __target_load(int cpu, int type, enum idle_type idle) |
| 1045 | { |
| 1046 | runqueue_t *rq = cpu_rq(cpu); |
| 1047 | unsigned long running = rq->nr_running; |
| 1048 | unsigned long target_load, cpu_load = rq->cpu_load[type-1], |
| 1049 | load_now = running * SCHED_LOAD_SCALE; |
| 1050 | |
| 1051 | if (type == 0) |
| 1052 | target_load = load_now; |
| 1053 | else |
| 1054 | target_load = max(cpu_load, load_now); |
| 1055 | |
| 1056 | if (running > 1 || (idle == NOT_IDLE && running)) |
| 1057 | target_load = target_load * rq->prio_bias / running; |
| 1058 | |
| 1059 | return target_load; |
| 1060 | } |
| 1061 | |
| 1062 | static inline unsigned long target_load(int cpu, int type) |
| 1063 | { |
| 1064 | return __target_load(cpu, type, NOT_IDLE); |
| 1065 | } |
| 1066 | |
| 1067 | /* |
| 1068 | * find_idlest_group finds and returns the least busy CPU group within the |
| 1069 | * domain. |
| 1070 | */ |
| 1071 | static struct sched_group * |
| 1072 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu) |
| 1073 | { |
| 1074 | struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups; |
| 1075 | unsigned long min_load = ULONG_MAX, this_load = 0; |
| 1076 | int load_idx = sd->forkexec_idx; |
| 1077 | int imbalance = 100 + (sd->imbalance_pct-100)/2; |
| 1078 | |
| 1079 | do { |
| 1080 | unsigned long load, avg_load; |
| 1081 | int local_group; |
| 1082 | int i; |
| 1083 | |
| 1084 | /* Skip over this group if it has no CPUs allowed */ |
| 1085 | if (!cpus_intersects(group->cpumask, p->cpus_allowed)) |
| 1086 | goto nextgroup; |
| 1087 | |
| 1088 | local_group = cpu_isset(this_cpu, group->cpumask); |
| 1089 | |
| 1090 | /* Tally up the load of all CPUs in the group */ |
| 1091 | avg_load = 0; |
| 1092 | |
| 1093 | for_each_cpu_mask(i, group->cpumask) { |
| 1094 | /* Bias balancing toward cpus of our domain */ |
| 1095 | if (local_group) |
| 1096 | load = source_load(i, load_idx); |
| 1097 | else |
| 1098 | load = target_load(i, load_idx); |
| 1099 | |
| 1100 | avg_load += load; |
| 1101 | } |
| 1102 | |
| 1103 | /* Adjust by relative CPU power of the group */ |
| 1104 | avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power; |
| 1105 | |
| 1106 | if (local_group) { |
| 1107 | this_load = avg_load; |
| 1108 | this = group; |
| 1109 | } else if (avg_load < min_load) { |
| 1110 | min_load = avg_load; |
| 1111 | idlest = group; |
| 1112 | } |
| 1113 | nextgroup: |
| 1114 | group = group->next; |
| 1115 | } while (group != sd->groups); |
| 1116 | |
| 1117 | if (!idlest || 100*this_load < imbalance*min_load) |
| 1118 | return NULL; |
| 1119 | return idlest; |
| 1120 | } |
| 1121 | |
| 1122 | /* |
| 1123 | * find_idlest_queue - find the idlest runqueue among the cpus in group. |
| 1124 | */ |
| 1125 | static int |
| 1126 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) |
| 1127 | { |
| 1128 | cpumask_t tmp; |
| 1129 | unsigned long load, min_load = ULONG_MAX; |
| 1130 | int idlest = -1; |
| 1131 | int i; |
| 1132 | |
| 1133 | /* Traverse only the allowed CPUs */ |
| 1134 | cpus_and(tmp, group->cpumask, p->cpus_allowed); |
| 1135 | |
| 1136 | for_each_cpu_mask(i, tmp) { |
| 1137 | load = source_load(i, 0); |
| 1138 | |
| 1139 | if (load < min_load || (load == min_load && i == this_cpu)) { |
| 1140 | min_load = load; |
| 1141 | idlest = i; |
| 1142 | } |
| 1143 | } |
| 1144 | |
| 1145 | return idlest; |
| 1146 | } |
| 1147 | |
| 1148 | /* |
| 1149 | * sched_balance_self: balance the current task (running on cpu) in domains |
| 1150 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and |
| 1151 | * SD_BALANCE_EXEC. |
| 1152 | * |
| 1153 | * Balance, ie. select the least loaded group. |
| 1154 | * |
| 1155 | * Returns the target CPU number, or the same CPU if no balancing is needed. |
| 1156 | * |
| 1157 | * preempt must be disabled. |
| 1158 | */ |
| 1159 | static int sched_balance_self(int cpu, int flag) |
| 1160 | { |
| 1161 | struct task_struct *t = current; |
| 1162 | struct sched_domain *tmp, *sd = NULL; |
| 1163 | |
| 1164 | for_each_domain(cpu, tmp) |
| 1165 | if (tmp->flags & flag) |
| 1166 | sd = tmp; |
| 1167 | |
| 1168 | while (sd) { |
| 1169 | cpumask_t span; |
| 1170 | struct sched_group *group; |
| 1171 | int new_cpu; |
| 1172 | int weight; |
| 1173 | |
| 1174 | span = sd->span; |
| 1175 | group = find_idlest_group(sd, t, cpu); |
| 1176 | if (!group) |
| 1177 | goto nextlevel; |
| 1178 | |
| 1179 | new_cpu = find_idlest_cpu(group, t, cpu); |
| 1180 | if (new_cpu == -1 || new_cpu == cpu) |
| 1181 | goto nextlevel; |
| 1182 | |
| 1183 | /* Now try balancing at a lower domain level */ |
| 1184 | cpu = new_cpu; |
| 1185 | nextlevel: |
| 1186 | sd = NULL; |
| 1187 | weight = cpus_weight(span); |
| 1188 | for_each_domain(cpu, tmp) { |
| 1189 | if (weight <= cpus_weight(tmp->span)) |
| 1190 | break; |
| 1191 | if (tmp->flags & flag) |
| 1192 | sd = tmp; |
| 1193 | } |
| 1194 | /* while loop will break here if sd == NULL */ |
| 1195 | } |
| 1196 | |
| 1197 | return cpu; |
| 1198 | } |
| 1199 | |
| 1200 | #endif /* CONFIG_SMP */ |
| 1201 | |
| 1202 | /* |
| 1203 | * wake_idle() will wake a task on an idle cpu if task->cpu is |
| 1204 | * not idle and an idle cpu is available. The span of cpus to |
| 1205 | * search starts with cpus closest then further out as needed, |
| 1206 | * so we always favor a closer, idle cpu. |
| 1207 | * |
| 1208 | * Returns the CPU we should wake onto. |
| 1209 | */ |
| 1210 | #if defined(ARCH_HAS_SCHED_WAKE_IDLE) |
| 1211 | static int wake_idle(int cpu, task_t *p) |
| 1212 | { |
| 1213 | cpumask_t tmp; |
| 1214 | struct sched_domain *sd; |
| 1215 | int i; |
| 1216 | |
| 1217 | if (idle_cpu(cpu)) |
| 1218 | return cpu; |
| 1219 | |
| 1220 | for_each_domain(cpu, sd) { |
| 1221 | if (sd->flags & SD_WAKE_IDLE) { |
| 1222 | cpus_and(tmp, sd->span, p->cpus_allowed); |
| 1223 | for_each_cpu_mask(i, tmp) { |
| 1224 | if (idle_cpu(i)) |
| 1225 | return i; |
| 1226 | } |
| 1227 | } |
| 1228 | else |
| 1229 | break; |
| 1230 | } |
| 1231 | return cpu; |
| 1232 | } |
| 1233 | #else |
| 1234 | static inline int wake_idle(int cpu, task_t *p) |
| 1235 | { |
| 1236 | return cpu; |
| 1237 | } |
| 1238 | #endif |
| 1239 | |
| 1240 | /*** |
| 1241 | * try_to_wake_up - wake up a thread |
| 1242 | * @p: the to-be-woken-up thread |
| 1243 | * @state: the mask of task states that can be woken |
| 1244 | * @sync: do a synchronous wakeup? |
| 1245 | * |
| 1246 | * Put it on the run-queue if it's not already there. The "current" |
| 1247 | * thread is always on the run-queue (except when the actual |
| 1248 | * re-schedule is in progress), and as such you're allowed to do |
| 1249 | * the simpler "current->state = TASK_RUNNING" to mark yourself |
| 1250 | * runnable without the overhead of this. |
| 1251 | * |
| 1252 | * returns failure only if the task is already active. |
| 1253 | */ |
| 1254 | static int try_to_wake_up(task_t *p, unsigned int state, int sync) |
| 1255 | { |
| 1256 | int cpu, this_cpu, success = 0; |
| 1257 | unsigned long flags; |
| 1258 | long old_state; |
| 1259 | runqueue_t *rq; |
| 1260 | #ifdef CONFIG_SMP |
| 1261 | unsigned long load, this_load; |
| 1262 | struct sched_domain *sd, *this_sd = NULL; |
| 1263 | int new_cpu; |
| 1264 | #endif |
| 1265 | |
| 1266 | rq = task_rq_lock(p, &flags); |
| 1267 | old_state = p->state; |
| 1268 | if (!(old_state & state)) |
| 1269 | goto out; |
| 1270 | |
| 1271 | if (p->array) |
| 1272 | goto out_running; |
| 1273 | |
| 1274 | cpu = task_cpu(p); |
| 1275 | this_cpu = smp_processor_id(); |
| 1276 | |
| 1277 | #ifdef CONFIG_SMP |
| 1278 | if (unlikely(task_running(rq, p))) |
| 1279 | goto out_activate; |
| 1280 | |
| 1281 | new_cpu = cpu; |
| 1282 | |
| 1283 | schedstat_inc(rq, ttwu_cnt); |
| 1284 | if (cpu == this_cpu) { |
| 1285 | schedstat_inc(rq, ttwu_local); |
| 1286 | goto out_set_cpu; |
| 1287 | } |
| 1288 | |
| 1289 | for_each_domain(this_cpu, sd) { |
| 1290 | if (cpu_isset(cpu, sd->span)) { |
| 1291 | schedstat_inc(sd, ttwu_wake_remote); |
| 1292 | this_sd = sd; |
| 1293 | break; |
| 1294 | } |
| 1295 | } |
| 1296 | |
| 1297 | if (p->last_waker_cpu != this_cpu) |
| 1298 | goto out_set_cpu; |
| 1299 | |
| 1300 | if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed))) |
| 1301 | goto out_set_cpu; |
| 1302 | |
| 1303 | /* |
| 1304 | * Check for affine wakeup and passive balancing possibilities. |
| 1305 | */ |
| 1306 | if (this_sd) { |
| 1307 | int idx = this_sd->wake_idx; |
| 1308 | unsigned int imbalance; |
| 1309 | |
| 1310 | imbalance = 100 + (this_sd->imbalance_pct - 100) / 2; |
| 1311 | |
| 1312 | load = source_load(cpu, idx); |
| 1313 | this_load = target_load(this_cpu, idx); |
| 1314 | |
| 1315 | new_cpu = this_cpu; /* Wake to this CPU if we can */ |
| 1316 | |
| 1317 | if (this_sd->flags & SD_WAKE_AFFINE) { |
| 1318 | unsigned long tl = this_load; |
| 1319 | /* |
| 1320 | * If sync wakeup then subtract the (maximum possible) |
| 1321 | * effect of the currently running task from the load |
| 1322 | * of the current CPU: |
| 1323 | */ |
| 1324 | if (sync) |
| 1325 | tl -= SCHED_LOAD_SCALE; |
| 1326 | |
| 1327 | if ((tl <= load && |
| 1328 | tl + target_load(cpu, idx) <= SCHED_LOAD_SCALE) || |
| 1329 | 100*(tl + SCHED_LOAD_SCALE) <= imbalance*load) { |
| 1330 | /* |
| 1331 | * This domain has SD_WAKE_AFFINE and |
| 1332 | * p is cache cold in this domain, and |
| 1333 | * there is no bad imbalance. |
| 1334 | */ |
| 1335 | schedstat_inc(this_sd, ttwu_move_affine); |
| 1336 | goto out_set_cpu; |
| 1337 | } |
| 1338 | } |
| 1339 | |
| 1340 | /* |
| 1341 | * Start passive balancing when half the imbalance_pct |
| 1342 | * limit is reached. |
| 1343 | */ |
| 1344 | if (this_sd->flags & SD_WAKE_BALANCE) { |
| 1345 | if (imbalance*this_load <= 100*load) { |
| 1346 | schedstat_inc(this_sd, ttwu_move_balance); |
| 1347 | goto out_set_cpu; |
| 1348 | } |
| 1349 | } |
| 1350 | } |
| 1351 | |
| 1352 | new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */ |
| 1353 | out_set_cpu: |
| 1354 | new_cpu = wake_idle(new_cpu, p); |
| 1355 | if (new_cpu != cpu) { |
| 1356 | set_task_cpu(p, new_cpu); |
| 1357 | task_rq_unlock(rq, &flags); |
| 1358 | /* might preempt at this point */ |
| 1359 | rq = task_rq_lock(p, &flags); |
| 1360 | old_state = p->state; |
| 1361 | if (!(old_state & state)) |
| 1362 | goto out; |
| 1363 | if (p->array) |
| 1364 | goto out_running; |
| 1365 | |
| 1366 | this_cpu = smp_processor_id(); |
| 1367 | cpu = task_cpu(p); |
| 1368 | } |
| 1369 | |
| 1370 | p->last_waker_cpu = this_cpu; |
| 1371 | |
| 1372 | out_activate: |
| 1373 | #endif /* CONFIG_SMP */ |
| 1374 | if (old_state == TASK_UNINTERRUPTIBLE) { |
| 1375 | rq->nr_uninterruptible--; |
| 1376 | /* |
| 1377 | * Tasks on involuntary sleep don't earn |
| 1378 | * sleep_avg beyond just interactive state. |
| 1379 | */ |
| 1380 | p->activated = -1; |
| 1381 | } |
| 1382 | |
| 1383 | /* |
| 1384 | * Tasks that have marked their sleep as noninteractive get |
| 1385 | * woken up without updating their sleep average. (i.e. their |
| 1386 | * sleep is handled in a priority-neutral manner, no priority |
| 1387 | * boost and no penalty.) |
| 1388 | */ |
| 1389 | if (old_state & TASK_NONINTERACTIVE) |
| 1390 | __activate_task(p, rq); |
| 1391 | else |
| 1392 | activate_task(p, rq, cpu == this_cpu); |
| 1393 | /* |
| 1394 | * Sync wakeups (i.e. those types of wakeups where the waker |
| 1395 | * has indicated that it will leave the CPU in short order) |
| 1396 | * don't trigger a preemption, if the woken up task will run on |
| 1397 | * this cpu. (in this case the 'I will reschedule' promise of |
| 1398 | * the waker guarantees that the freshly woken up task is going |
| 1399 | * to be considered on this CPU.) |
| 1400 | */ |
| 1401 | if (!sync || cpu != this_cpu) { |
| 1402 | if (TASK_PREEMPTS_CURR(p, rq)) |
| 1403 | resched_task(rq->curr); |
| 1404 | } |
| 1405 | success = 1; |
| 1406 | |
| 1407 | out_running: |
| 1408 | p->state = TASK_RUNNING; |
| 1409 | out: |
| 1410 | task_rq_unlock(rq, &flags); |
| 1411 | |
| 1412 | return success; |
| 1413 | } |
| 1414 | |
| 1415 | int fastcall wake_up_process(task_t *p) |
| 1416 | { |
| 1417 | return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED | |
| 1418 | TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0); |
| 1419 | } |
| 1420 | |
| 1421 | EXPORT_SYMBOL(wake_up_process); |
| 1422 | |
| 1423 | int fastcall wake_up_state(task_t *p, unsigned int state) |
| 1424 | { |
| 1425 | return try_to_wake_up(p, state, 0); |
| 1426 | } |
| 1427 | |
| 1428 | /* |
| 1429 | * Perform scheduler related setup for a newly forked process p. |
| 1430 | * p is forked by current. |
| 1431 | */ |
| 1432 | void fastcall sched_fork(task_t *p, int clone_flags) |
| 1433 | { |
| 1434 | int cpu = get_cpu(); |
| 1435 | |
| 1436 | #ifdef CONFIG_SMP |
| 1437 | cpu = sched_balance_self(cpu, SD_BALANCE_FORK); |
| 1438 | #endif |
| 1439 | set_task_cpu(p, cpu); |
| 1440 | |
| 1441 | /* |
| 1442 | * We mark the process as running here, but have not actually |
| 1443 | * inserted it onto the runqueue yet. This guarantees that |
| 1444 | * nobody will actually run it, and a signal or other external |
| 1445 | * event cannot wake it up and insert it on the runqueue either. |
| 1446 | */ |
| 1447 | p->state = TASK_RUNNING; |
| 1448 | INIT_LIST_HEAD(&p->run_list); |
| 1449 | p->array = NULL; |
| 1450 | #ifdef CONFIG_SCHEDSTATS |
| 1451 | memset(&p->sched_info, 0, sizeof(p->sched_info)); |
| 1452 | #endif |
| 1453 | #if defined(CONFIG_SMP) |
| 1454 | p->last_waker_cpu = cpu; |
| 1455 | #if defined(__ARCH_WANT_UNLOCKED_CTXSW) |
| 1456 | p->oncpu = 0; |
| 1457 | #endif |
| 1458 | #endif |
| 1459 | #ifdef CONFIG_PREEMPT |
| 1460 | /* Want to start with kernel preemption disabled. */ |
| 1461 | task_thread_info(p)->preempt_count = 1; |
| 1462 | #endif |
| 1463 | /* |
| 1464 | * Share the timeslice between parent and child, thus the |
| 1465 | * total amount of pending timeslices in the system doesn't change, |
| 1466 | * resulting in more scheduling fairness. |
| 1467 | */ |
| 1468 | local_irq_disable(); |
| 1469 | p->time_slice = (current->time_slice + 1) >> 1; |
| 1470 | /* |
| 1471 | * The remainder of the first timeslice might be recovered by |
| 1472 | * the parent if the child exits early enough. |
| 1473 | */ |
| 1474 | p->first_time_slice = 1; |
| 1475 | current->time_slice >>= 1; |
| 1476 | p->timestamp = sched_clock(); |
| 1477 | if (unlikely(!current->time_slice)) { |
| 1478 | /* |
| 1479 | * This case is rare, it happens when the parent has only |
| 1480 | * a single jiffy left from its timeslice. Taking the |
| 1481 | * runqueue lock is not a problem. |
| 1482 | */ |
| 1483 | current->time_slice = 1; |
| 1484 | scheduler_tick(); |
| 1485 | } |
| 1486 | local_irq_enable(); |
| 1487 | put_cpu(); |
| 1488 | } |
| 1489 | |
| 1490 | /* |
| 1491 | * wake_up_new_task - wake up a newly created task for the first time. |
| 1492 | * |
| 1493 | * This function will do some initial scheduler statistics housekeeping |
| 1494 | * that must be done for every newly created context, then puts the task |
| 1495 | * on the runqueue and wakes it. |
| 1496 | */ |
| 1497 | void fastcall wake_up_new_task(task_t *p, unsigned long clone_flags) |
| 1498 | { |
| 1499 | unsigned long flags; |
| 1500 | int this_cpu, cpu; |
| 1501 | runqueue_t *rq, *this_rq; |
| 1502 | |
| 1503 | rq = task_rq_lock(p, &flags); |
| 1504 | BUG_ON(p->state != TASK_RUNNING); |
| 1505 | this_cpu = smp_processor_id(); |
| 1506 | cpu = task_cpu(p); |
| 1507 | |
| 1508 | /* |
| 1509 | * We decrease the sleep average of forking parents |
| 1510 | * and children as well, to keep max-interactive tasks |
| 1511 | * from forking tasks that are max-interactive. The parent |
| 1512 | * (current) is done further down, under its lock. |
| 1513 | */ |
| 1514 | p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) * |
| 1515 | CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS); |
| 1516 | |
| 1517 | p->prio = effective_prio(p); |
| 1518 | |
| 1519 | if (likely(cpu == this_cpu)) { |
| 1520 | if (!(clone_flags & CLONE_VM)) { |
| 1521 | /* |
| 1522 | * The VM isn't cloned, so we're in a good position to |
| 1523 | * do child-runs-first in anticipation of an exec. This |
| 1524 | * usually avoids a lot of COW overhead. |
| 1525 | */ |
| 1526 | if (unlikely(!current->array)) |
| 1527 | __activate_task(p, rq); |
| 1528 | else { |
| 1529 | p->prio = current->prio; |
| 1530 | list_add_tail(&p->run_list, ¤t->run_list); |
| 1531 | p->array = current->array; |
| 1532 | p->array->nr_active++; |
| 1533 | inc_nr_running(p, rq); |
| 1534 | } |
| 1535 | set_need_resched(); |
| 1536 | } else |
| 1537 | /* Run child last */ |
| 1538 | __activate_task(p, rq); |
| 1539 | /* |
| 1540 | * We skip the following code due to cpu == this_cpu |
| 1541 | * |
| 1542 | * task_rq_unlock(rq, &flags); |
| 1543 | * this_rq = task_rq_lock(current, &flags); |
| 1544 | */ |
| 1545 | this_rq = rq; |
| 1546 | } else { |
| 1547 | this_rq = cpu_rq(this_cpu); |
| 1548 | |
| 1549 | /* |
| 1550 | * Not the local CPU - must adjust timestamp. This should |
| 1551 | * get optimised away in the !CONFIG_SMP case. |
| 1552 | */ |
| 1553 | p->timestamp = (p->timestamp - this_rq->timestamp_last_tick) |
| 1554 | + rq->timestamp_last_tick; |
| 1555 | __activate_task(p, rq); |
| 1556 | if (TASK_PREEMPTS_CURR(p, rq)) |
| 1557 | resched_task(rq->curr); |
| 1558 | |
| 1559 | /* |
| 1560 | * Parent and child are on different CPUs, now get the |
| 1561 | * parent runqueue to update the parent's ->sleep_avg: |
| 1562 | */ |
| 1563 | task_rq_unlock(rq, &flags); |
| 1564 | this_rq = task_rq_lock(current, &flags); |
| 1565 | } |
| 1566 | current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) * |
| 1567 | PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS); |
| 1568 | task_rq_unlock(this_rq, &flags); |
| 1569 | } |
| 1570 | |
| 1571 | /* |
| 1572 | * Potentially available exiting-child timeslices are |
| 1573 | * retrieved here - this way the parent does not get |
| 1574 | * penalized for creating too many threads. |
| 1575 | * |
| 1576 | * (this cannot be used to 'generate' timeslices |
| 1577 | * artificially, because any timeslice recovered here |
| 1578 | * was given away by the parent in the first place.) |
| 1579 | */ |
| 1580 | void fastcall sched_exit(task_t *p) |
| 1581 | { |
| 1582 | unsigned long flags; |
| 1583 | runqueue_t *rq; |
| 1584 | |
| 1585 | /* |
| 1586 | * If the child was a (relative-) CPU hog then decrease |
| 1587 | * the sleep_avg of the parent as well. |
| 1588 | */ |
| 1589 | rq = task_rq_lock(p->parent, &flags); |
| 1590 | if (p->first_time_slice && task_cpu(p) == task_cpu(p->parent)) { |
| 1591 | p->parent->time_slice += p->time_slice; |
| 1592 | if (unlikely(p->parent->time_slice > task_timeslice(p))) |
| 1593 | p->parent->time_slice = task_timeslice(p); |
| 1594 | } |
| 1595 | if (p->sleep_avg < p->parent->sleep_avg) |
| 1596 | p->parent->sleep_avg = p->parent->sleep_avg / |
| 1597 | (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg / |
| 1598 | (EXIT_WEIGHT + 1); |
| 1599 | task_rq_unlock(rq, &flags); |
| 1600 | } |
| 1601 | |
| 1602 | /** |
| 1603 | * prepare_task_switch - prepare to switch tasks |
| 1604 | * @rq: the runqueue preparing to switch |
| 1605 | * @next: the task we are going to switch to. |
| 1606 | * |
| 1607 | * This is called with the rq lock held and interrupts off. It must |
| 1608 | * be paired with a subsequent finish_task_switch after the context |
| 1609 | * switch. |
| 1610 | * |
| 1611 | * prepare_task_switch sets up locking and calls architecture specific |
| 1612 | * hooks. |
| 1613 | */ |
| 1614 | static inline void prepare_task_switch(runqueue_t *rq, task_t *next) |
| 1615 | { |
| 1616 | prepare_lock_switch(rq, next); |
| 1617 | prepare_arch_switch(next); |
| 1618 | } |
| 1619 | |
| 1620 | /** |
| 1621 | * finish_task_switch - clean up after a task-switch |
| 1622 | * @rq: runqueue associated with task-switch |
| 1623 | * @prev: the thread we just switched away from. |
| 1624 | * |
| 1625 | * finish_task_switch must be called after the context switch, paired |
| 1626 | * with a prepare_task_switch call before the context switch. |
| 1627 | * finish_task_switch will reconcile locking set up by prepare_task_switch, |
| 1628 | * and do any other architecture-specific cleanup actions. |
| 1629 | * |
| 1630 | * Note that we may have delayed dropping an mm in context_switch(). If |
| 1631 | * so, we finish that here outside of the runqueue lock. (Doing it |
| 1632 | * with the lock held can cause deadlocks; see schedule() for |
| 1633 | * details.) |
| 1634 | */ |
| 1635 | static inline void finish_task_switch(runqueue_t *rq, task_t *prev) |
| 1636 | __releases(rq->lock) |
| 1637 | { |
| 1638 | struct mm_struct *mm = rq->prev_mm; |
| 1639 | unsigned long prev_task_flags; |
| 1640 | |
| 1641 | rq->prev_mm = NULL; |
| 1642 | |
| 1643 | /* |
| 1644 | * A task struct has one reference for the use as "current". |
| 1645 | * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and |
| 1646 | * calls schedule one last time. The schedule call will never return, |
| 1647 | * and the scheduled task must drop that reference. |
| 1648 | * The test for EXIT_ZOMBIE must occur while the runqueue locks are |
| 1649 | * still held, otherwise prev could be scheduled on another cpu, die |
| 1650 | * there before we look at prev->state, and then the reference would |
| 1651 | * be dropped twice. |
| 1652 | * Manfred Spraul <manfred@colorfullife.com> |
| 1653 | */ |
| 1654 | prev_task_flags = prev->flags; |
| 1655 | finish_arch_switch(prev); |
| 1656 | finish_lock_switch(rq, prev); |
| 1657 | if (mm) |
| 1658 | mmdrop(mm); |
| 1659 | if (unlikely(prev_task_flags & PF_DEAD)) |
| 1660 | put_task_struct(prev); |
| 1661 | } |
| 1662 | |
| 1663 | /** |
| 1664 | * schedule_tail - first thing a freshly forked thread must call. |
| 1665 | * @prev: the thread we just switched away from. |
| 1666 | */ |
| 1667 | asmlinkage void schedule_tail(task_t *prev) |
| 1668 | __releases(rq->lock) |
| 1669 | { |
| 1670 | runqueue_t *rq = this_rq(); |
| 1671 | finish_task_switch(rq, prev); |
| 1672 | #ifdef __ARCH_WANT_UNLOCKED_CTXSW |
| 1673 | /* In this case, finish_task_switch does not reenable preemption */ |
| 1674 | preempt_enable(); |
| 1675 | #endif |
| 1676 | if (current->set_child_tid) |
| 1677 | put_user(current->pid, current->set_child_tid); |
| 1678 | } |
| 1679 | |
| 1680 | /* |
| 1681 | * context_switch - switch to the new MM and the new |
| 1682 | * thread's register state. |
| 1683 | */ |
| 1684 | static inline |
| 1685 | task_t * context_switch(runqueue_t *rq, task_t *prev, task_t *next) |
| 1686 | { |
| 1687 | struct mm_struct *mm = next->mm; |
| 1688 | struct mm_struct *oldmm = prev->active_mm; |
| 1689 | |
| 1690 | if (unlikely(!mm)) { |
| 1691 | next->active_mm = oldmm; |
| 1692 | atomic_inc(&oldmm->mm_count); |
| 1693 | enter_lazy_tlb(oldmm, next); |
| 1694 | } else |
| 1695 | switch_mm(oldmm, mm, next); |
| 1696 | |
| 1697 | if (unlikely(!prev->mm)) { |
| 1698 | prev->active_mm = NULL; |
| 1699 | WARN_ON(rq->prev_mm); |
| 1700 | rq->prev_mm = oldmm; |
| 1701 | } |
| 1702 | |
| 1703 | /* Here we just switch the register state and the stack. */ |
| 1704 | switch_to(prev, next, prev); |
| 1705 | |
| 1706 | return prev; |
| 1707 | } |
| 1708 | |
| 1709 | /* |
| 1710 | * nr_running, nr_uninterruptible and nr_context_switches: |
| 1711 | * |
| 1712 | * externally visible scheduler statistics: current number of runnable |
| 1713 | * threads, current number of uninterruptible-sleeping threads, total |
| 1714 | * number of context switches performed since bootup. |
| 1715 | */ |
| 1716 | unsigned long nr_running(void) |
| 1717 | { |
| 1718 | unsigned long i, sum = 0; |
| 1719 | |
| 1720 | for_each_online_cpu(i) |
| 1721 | sum += cpu_rq(i)->nr_running; |
| 1722 | |
| 1723 | return sum; |
| 1724 | } |
| 1725 | |
| 1726 | unsigned long nr_uninterruptible(void) |
| 1727 | { |
| 1728 | unsigned long i, sum = 0; |
| 1729 | |
| 1730 | for_each_cpu(i) |
| 1731 | sum += cpu_rq(i)->nr_uninterruptible; |
| 1732 | |
| 1733 | /* |
| 1734 | * Since we read the counters lockless, it might be slightly |
| 1735 | * inaccurate. Do not allow it to go below zero though: |
| 1736 | */ |
| 1737 | if (unlikely((long)sum < 0)) |
| 1738 | sum = 0; |
| 1739 | |
| 1740 | return sum; |
| 1741 | } |
| 1742 | |
| 1743 | unsigned long long nr_context_switches(void) |
| 1744 | { |
| 1745 | unsigned long long i, sum = 0; |
| 1746 | |
| 1747 | for_each_cpu(i) |
| 1748 | sum += cpu_rq(i)->nr_switches; |
| 1749 | |
| 1750 | return sum; |
| 1751 | } |
| 1752 | |
| 1753 | unsigned long nr_iowait(void) |
| 1754 | { |
| 1755 | unsigned long i, sum = 0; |
| 1756 | |
| 1757 | for_each_cpu(i) |
| 1758 | sum += atomic_read(&cpu_rq(i)->nr_iowait); |
| 1759 | |
| 1760 | return sum; |
| 1761 | } |
| 1762 | |
| 1763 | #ifdef CONFIG_SMP |
| 1764 | |
| 1765 | /* |
| 1766 | * double_rq_lock - safely lock two runqueues |
| 1767 | * |
| 1768 | * Note this does not disable interrupts like task_rq_lock, |
| 1769 | * you need to do so manually before calling. |
| 1770 | */ |
| 1771 | static void double_rq_lock(runqueue_t *rq1, runqueue_t *rq2) |
| 1772 | __acquires(rq1->lock) |
| 1773 | __acquires(rq2->lock) |
| 1774 | { |
| 1775 | if (rq1 == rq2) { |
| 1776 | spin_lock(&rq1->lock); |
| 1777 | __acquire(rq2->lock); /* Fake it out ;) */ |
| 1778 | } else { |
| 1779 | if (rq1 < rq2) { |
| 1780 | spin_lock(&rq1->lock); |
| 1781 | spin_lock(&rq2->lock); |
| 1782 | } else { |
| 1783 | spin_lock(&rq2->lock); |
| 1784 | spin_lock(&rq1->lock); |
| 1785 | } |
| 1786 | } |
| 1787 | } |
| 1788 | |
| 1789 | /* |
| 1790 | * double_rq_unlock - safely unlock two runqueues |
| 1791 | * |
| 1792 | * Note this does not restore interrupts like task_rq_unlock, |
| 1793 | * you need to do so manually after calling. |
| 1794 | */ |
| 1795 | static void double_rq_unlock(runqueue_t *rq1, runqueue_t *rq2) |
| 1796 | __releases(rq1->lock) |
| 1797 | __releases(rq2->lock) |
| 1798 | { |
| 1799 | spin_unlock(&rq1->lock); |
| 1800 | if (rq1 != rq2) |
| 1801 | spin_unlock(&rq2->lock); |
| 1802 | else |
| 1803 | __release(rq2->lock); |
| 1804 | } |
| 1805 | |
| 1806 | /* |
| 1807 | * double_lock_balance - lock the busiest runqueue, this_rq is locked already. |
| 1808 | */ |
| 1809 | static void double_lock_balance(runqueue_t *this_rq, runqueue_t *busiest) |
| 1810 | __releases(this_rq->lock) |
| 1811 | __acquires(busiest->lock) |
| 1812 | __acquires(this_rq->lock) |
| 1813 | { |
| 1814 | if (unlikely(!spin_trylock(&busiest->lock))) { |
| 1815 | if (busiest < this_rq) { |
| 1816 | spin_unlock(&this_rq->lock); |
| 1817 | spin_lock(&busiest->lock); |
| 1818 | spin_lock(&this_rq->lock); |
| 1819 | } else |
| 1820 | spin_lock(&busiest->lock); |
| 1821 | } |
| 1822 | } |
| 1823 | |
| 1824 | /* |
| 1825 | * If dest_cpu is allowed for this process, migrate the task to it. |
| 1826 | * This is accomplished by forcing the cpu_allowed mask to only |
| 1827 | * allow dest_cpu, which will force the cpu onto dest_cpu. Then |
| 1828 | * the cpu_allowed mask is restored. |
| 1829 | */ |
| 1830 | static void sched_migrate_task(task_t *p, int dest_cpu) |
| 1831 | { |
| 1832 | migration_req_t req; |
| 1833 | runqueue_t *rq; |
| 1834 | unsigned long flags; |
| 1835 | |
| 1836 | rq = task_rq_lock(p, &flags); |
| 1837 | if (!cpu_isset(dest_cpu, p->cpus_allowed) |
| 1838 | || unlikely(cpu_is_offline(dest_cpu))) |
| 1839 | goto out; |
| 1840 | |
| 1841 | /* force the process onto the specified CPU */ |
| 1842 | if (migrate_task(p, dest_cpu, &req)) { |
| 1843 | /* Need to wait for migration thread (might exit: take ref). */ |
| 1844 | struct task_struct *mt = rq->migration_thread; |
| 1845 | get_task_struct(mt); |
| 1846 | task_rq_unlock(rq, &flags); |
| 1847 | wake_up_process(mt); |
| 1848 | put_task_struct(mt); |
| 1849 | wait_for_completion(&req.done); |
| 1850 | return; |
| 1851 | } |
| 1852 | out: |
| 1853 | task_rq_unlock(rq, &flags); |
| 1854 | } |
| 1855 | |
| 1856 | /* |
| 1857 | * sched_exec - execve() is a valuable balancing opportunity, because at |
| 1858 | * this point the task has the smallest effective memory and cache footprint. |
| 1859 | */ |
| 1860 | void sched_exec(void) |
| 1861 | { |
| 1862 | int new_cpu, this_cpu = get_cpu(); |
| 1863 | new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC); |
| 1864 | put_cpu(); |
| 1865 | if (new_cpu != this_cpu) |
| 1866 | sched_migrate_task(current, new_cpu); |
| 1867 | } |
| 1868 | |
| 1869 | /* |
| 1870 | * pull_task - move a task from a remote runqueue to the local runqueue. |
| 1871 | * Both runqueues must be locked. |
| 1872 | */ |
| 1873 | static |
| 1874 | void pull_task(runqueue_t *src_rq, prio_array_t *src_array, task_t *p, |
| 1875 | runqueue_t *this_rq, prio_array_t *this_array, int this_cpu) |
| 1876 | { |
| 1877 | dequeue_task(p, src_array); |
| 1878 | dec_nr_running(p, src_rq); |
| 1879 | set_task_cpu(p, this_cpu); |
| 1880 | inc_nr_running(p, this_rq); |
| 1881 | enqueue_task(p, this_array); |
| 1882 | p->timestamp = (p->timestamp - src_rq->timestamp_last_tick) |
| 1883 | + this_rq->timestamp_last_tick; |
| 1884 | /* |
| 1885 | * Note that idle threads have a prio of MAX_PRIO, for this test |
| 1886 | * to be always true for them. |
| 1887 | */ |
| 1888 | if (TASK_PREEMPTS_CURR(p, this_rq)) |
| 1889 | resched_task(this_rq->curr); |
| 1890 | } |
| 1891 | |
| 1892 | /* |
| 1893 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? |
| 1894 | */ |
| 1895 | static |
| 1896 | int can_migrate_task(task_t *p, runqueue_t *rq, int this_cpu, |
| 1897 | struct sched_domain *sd, enum idle_type idle, |
| 1898 | int *all_pinned) |
| 1899 | { |
| 1900 | /* |
| 1901 | * We do not migrate tasks that are: |
| 1902 | * 1) running (obviously), or |
| 1903 | * 2) cannot be migrated to this CPU due to cpus_allowed, or |
| 1904 | * 3) are cache-hot on their current CPU. |
| 1905 | */ |
| 1906 | if (!cpu_isset(this_cpu, p->cpus_allowed)) |
| 1907 | return 0; |
| 1908 | *all_pinned = 0; |
| 1909 | |
| 1910 | if (task_running(rq, p)) |
| 1911 | return 0; |
| 1912 | |
| 1913 | /* |
| 1914 | * Aggressive migration if: |
| 1915 | * 1) task is cache cold, or |
| 1916 | * 2) too many balance attempts have failed. |
| 1917 | */ |
| 1918 | |
| 1919 | if (sd->nr_balance_failed > sd->cache_nice_tries) |
| 1920 | return 1; |
| 1921 | |
| 1922 | if (task_hot(p, rq->timestamp_last_tick, sd)) |
| 1923 | return 0; |
| 1924 | return 1; |
| 1925 | } |
| 1926 | |
| 1927 | /* |
| 1928 | * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq, |
| 1929 | * as part of a balancing operation within "domain". Returns the number of |
| 1930 | * tasks moved. |
| 1931 | * |
| 1932 | * Called with both runqueues locked. |
| 1933 | */ |
| 1934 | static int move_tasks(runqueue_t *this_rq, int this_cpu, runqueue_t *busiest, |
| 1935 | unsigned long max_nr_move, struct sched_domain *sd, |
| 1936 | enum idle_type idle, int *all_pinned) |
| 1937 | { |
| 1938 | prio_array_t *array, *dst_array; |
| 1939 | struct list_head *head, *curr; |
| 1940 | int idx, pulled = 0, pinned = 0; |
| 1941 | task_t *tmp; |
| 1942 | |
| 1943 | if (max_nr_move == 0) |
| 1944 | goto out; |
| 1945 | |
| 1946 | pinned = 1; |
| 1947 | |
| 1948 | /* |
| 1949 | * We first consider expired tasks. Those will likely not be |
| 1950 | * executed in the near future, and they are most likely to |
| 1951 | * be cache-cold, thus switching CPUs has the least effect |
| 1952 | * on them. |
| 1953 | */ |
| 1954 | if (busiest->expired->nr_active) { |
| 1955 | array = busiest->expired; |
| 1956 | dst_array = this_rq->expired; |
| 1957 | } else { |
| 1958 | array = busiest->active; |
| 1959 | dst_array = this_rq->active; |
| 1960 | } |
| 1961 | |
| 1962 | new_array: |
| 1963 | /* Start searching at priority 0: */ |
| 1964 | idx = 0; |
| 1965 | skip_bitmap: |
| 1966 | if (!idx) |
| 1967 | idx = sched_find_first_bit(array->bitmap); |
| 1968 | else |
| 1969 | idx = find_next_bit(array->bitmap, MAX_PRIO, idx); |
| 1970 | if (idx >= MAX_PRIO) { |
| 1971 | if (array == busiest->expired && busiest->active->nr_active) { |
| 1972 | array = busiest->active; |
| 1973 | dst_array = this_rq->active; |
| 1974 | goto new_array; |
| 1975 | } |
| 1976 | goto out; |
| 1977 | } |
| 1978 | |
| 1979 | head = array->queue + idx; |
| 1980 | curr = head->prev; |
| 1981 | skip_queue: |
| 1982 | tmp = list_entry(curr, task_t, run_list); |
| 1983 | |
| 1984 | curr = curr->prev; |
| 1985 | |
| 1986 | if (!can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) { |
| 1987 | if (curr != head) |
| 1988 | goto skip_queue; |
| 1989 | idx++; |
| 1990 | goto skip_bitmap; |
| 1991 | } |
| 1992 | |
| 1993 | #ifdef CONFIG_SCHEDSTATS |
| 1994 | if (task_hot(tmp, busiest->timestamp_last_tick, sd)) |
| 1995 | schedstat_inc(sd, lb_hot_gained[idle]); |
| 1996 | #endif |
| 1997 | |
| 1998 | pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu); |
| 1999 | pulled++; |
| 2000 | |
| 2001 | /* We only want to steal up to the prescribed number of tasks. */ |
| 2002 | if (pulled < max_nr_move) { |
| 2003 | if (curr != head) |
| 2004 | goto skip_queue; |
| 2005 | idx++; |
| 2006 | goto skip_bitmap; |
| 2007 | } |
| 2008 | out: |
| 2009 | /* |
| 2010 | * Right now, this is the only place pull_task() is called, |
| 2011 | * so we can safely collect pull_task() stats here rather than |
| 2012 | * inside pull_task(). |
| 2013 | */ |
| 2014 | schedstat_add(sd, lb_gained[idle], pulled); |
| 2015 | |
| 2016 | if (all_pinned) |
| 2017 | *all_pinned = pinned; |
| 2018 | return pulled; |
| 2019 | } |
| 2020 | |
| 2021 | /* |
| 2022 | * find_busiest_group finds and returns the busiest CPU group within the |
| 2023 | * domain. It calculates and returns the number of tasks which should be |
| 2024 | * moved to restore balance via the imbalance parameter. |
| 2025 | */ |
| 2026 | static struct sched_group * |
| 2027 | find_busiest_group(struct sched_domain *sd, int this_cpu, |
| 2028 | unsigned long *imbalance, enum idle_type idle, int *sd_idle) |
| 2029 | { |
| 2030 | struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups; |
| 2031 | unsigned long max_load, avg_load, total_load, this_load, total_pwr; |
| 2032 | unsigned long max_pull; |
| 2033 | int load_idx; |
| 2034 | |
| 2035 | max_load = this_load = total_load = total_pwr = 0; |
| 2036 | if (idle == NOT_IDLE) |
| 2037 | load_idx = sd->busy_idx; |
| 2038 | else if (idle == NEWLY_IDLE) |
| 2039 | load_idx = sd->newidle_idx; |
| 2040 | else |
| 2041 | load_idx = sd->idle_idx; |
| 2042 | |
| 2043 | do { |
| 2044 | unsigned long load; |
| 2045 | int local_group; |
| 2046 | int i; |
| 2047 | |
| 2048 | local_group = cpu_isset(this_cpu, group->cpumask); |
| 2049 | |
| 2050 | /* Tally up the load of all CPUs in the group */ |
| 2051 | avg_load = 0; |
| 2052 | |
| 2053 | for_each_cpu_mask(i, group->cpumask) { |
| 2054 | if (*sd_idle && !idle_cpu(i)) |
| 2055 | *sd_idle = 0; |
| 2056 | |
| 2057 | /* Bias balancing toward cpus of our domain */ |
| 2058 | if (local_group) |
| 2059 | load = __target_load(i, load_idx, idle); |
| 2060 | else |
| 2061 | load = __source_load(i, load_idx, idle); |
| 2062 | |
| 2063 | avg_load += load; |
| 2064 | } |
| 2065 | |
| 2066 | total_load += avg_load; |
| 2067 | total_pwr += group->cpu_power; |
| 2068 | |
| 2069 | /* Adjust by relative CPU power of the group */ |
| 2070 | avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power; |
| 2071 | |
| 2072 | if (local_group) { |
| 2073 | this_load = avg_load; |
| 2074 | this = group; |
| 2075 | } else if (avg_load > max_load) { |
| 2076 | max_load = avg_load; |
| 2077 | busiest = group; |
| 2078 | } |
| 2079 | group = group->next; |
| 2080 | } while (group != sd->groups); |
| 2081 | |
| 2082 | if (!busiest || this_load >= max_load || max_load <= SCHED_LOAD_SCALE) |
| 2083 | goto out_balanced; |
| 2084 | |
| 2085 | avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr; |
| 2086 | |
| 2087 | if (this_load >= avg_load || |
| 2088 | 100*max_load <= sd->imbalance_pct*this_load) |
| 2089 | goto out_balanced; |
| 2090 | |
| 2091 | /* |
| 2092 | * We're trying to get all the cpus to the average_load, so we don't |
| 2093 | * want to push ourselves above the average load, nor do we wish to |
| 2094 | * reduce the max loaded cpu below the average load, as either of these |
| 2095 | * actions would just result in more rebalancing later, and ping-pong |
| 2096 | * tasks around. Thus we look for the minimum possible imbalance. |
| 2097 | * Negative imbalances (*we* are more loaded than anyone else) will |
| 2098 | * be counted as no imbalance for these purposes -- we can't fix that |
| 2099 | * by pulling tasks to us. Be careful of negative numbers as they'll |
| 2100 | * appear as very large values with unsigned longs. |
| 2101 | */ |
| 2102 | |
| 2103 | /* Don't want to pull so many tasks that a group would go idle */ |
| 2104 | max_pull = min(max_load - avg_load, max_load - SCHED_LOAD_SCALE); |
| 2105 | |
| 2106 | /* How much load to actually move to equalise the imbalance */ |
| 2107 | *imbalance = min(max_pull * busiest->cpu_power, |
| 2108 | (avg_load - this_load) * this->cpu_power) |
| 2109 | / SCHED_LOAD_SCALE; |
| 2110 | |
| 2111 | if (*imbalance < SCHED_LOAD_SCALE) { |
| 2112 | unsigned long pwr_now = 0, pwr_move = 0; |
| 2113 | unsigned long tmp; |
| 2114 | |
| 2115 | if (max_load - this_load >= SCHED_LOAD_SCALE*2) { |
| 2116 | *imbalance = 1; |
| 2117 | return busiest; |
| 2118 | } |
| 2119 | |
| 2120 | /* |
| 2121 | * OK, we don't have enough imbalance to justify moving tasks, |
| 2122 | * however we may be able to increase total CPU power used by |
| 2123 | * moving them. |
| 2124 | */ |
| 2125 | |
| 2126 | pwr_now += busiest->cpu_power*min(SCHED_LOAD_SCALE, max_load); |
| 2127 | pwr_now += this->cpu_power*min(SCHED_LOAD_SCALE, this_load); |
| 2128 | pwr_now /= SCHED_LOAD_SCALE; |
| 2129 | |
| 2130 | /* Amount of load we'd subtract */ |
| 2131 | tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/busiest->cpu_power; |
| 2132 | if (max_load > tmp) |
| 2133 | pwr_move += busiest->cpu_power*min(SCHED_LOAD_SCALE, |
| 2134 | max_load - tmp); |
| 2135 | |
| 2136 | /* Amount of load we'd add */ |
| 2137 | if (max_load*busiest->cpu_power < |
| 2138 | SCHED_LOAD_SCALE*SCHED_LOAD_SCALE) |
| 2139 | tmp = max_load*busiest->cpu_power/this->cpu_power; |
| 2140 | else |
| 2141 | tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/this->cpu_power; |
| 2142 | pwr_move += this->cpu_power*min(SCHED_LOAD_SCALE, this_load + tmp); |
| 2143 | pwr_move /= SCHED_LOAD_SCALE; |
| 2144 | |
| 2145 | /* Move if we gain throughput */ |
| 2146 | if (pwr_move <= pwr_now) |
| 2147 | goto out_balanced; |
| 2148 | |
| 2149 | *imbalance = 1; |
| 2150 | return busiest; |
| 2151 | } |
| 2152 | |
| 2153 | /* Get rid of the scaling factor, rounding down as we divide */ |
| 2154 | *imbalance = *imbalance / SCHED_LOAD_SCALE; |
| 2155 | return busiest; |
| 2156 | |
| 2157 | out_balanced: |
| 2158 | |
| 2159 | *imbalance = 0; |
| 2160 | return NULL; |
| 2161 | } |
| 2162 | |
| 2163 | /* |
| 2164 | * find_busiest_queue - find the busiest runqueue among the cpus in group. |
| 2165 | */ |
| 2166 | static runqueue_t *find_busiest_queue(struct sched_group *group, |
| 2167 | enum idle_type idle) |
| 2168 | { |
| 2169 | unsigned long load, max_load = 0; |
| 2170 | runqueue_t *busiest = NULL; |
| 2171 | int i; |
| 2172 | |
| 2173 | for_each_cpu_mask(i, group->cpumask) { |
| 2174 | load = __source_load(i, 0, idle); |
| 2175 | |
| 2176 | if (load > max_load) { |
| 2177 | max_load = load; |
| 2178 | busiest = cpu_rq(i); |
| 2179 | } |
| 2180 | } |
| 2181 | |
| 2182 | return busiest; |
| 2183 | } |
| 2184 | |
| 2185 | /* |
| 2186 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but |
| 2187 | * so long as it is large enough. |
| 2188 | */ |
| 2189 | #define MAX_PINNED_INTERVAL 512 |
| 2190 | |
| 2191 | /* |
| 2192 | * Check this_cpu to ensure it is balanced within domain. Attempt to move |
| 2193 | * tasks if there is an imbalance. |
| 2194 | * |
| 2195 | * Called with this_rq unlocked. |
| 2196 | */ |
| 2197 | static int load_balance(int this_cpu, runqueue_t *this_rq, |
| 2198 | struct sched_domain *sd, enum idle_type idle) |
| 2199 | { |
| 2200 | struct sched_group *group; |
| 2201 | runqueue_t *busiest; |
| 2202 | unsigned long imbalance; |
| 2203 | int nr_moved, all_pinned = 0; |
| 2204 | int active_balance = 0; |
| 2205 | int sd_idle = 0; |
| 2206 | |
| 2207 | if (idle != NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER) |
| 2208 | sd_idle = 1; |
| 2209 | |
| 2210 | schedstat_inc(sd, lb_cnt[idle]); |
| 2211 | |
| 2212 | group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle); |
| 2213 | if (!group) { |
| 2214 | schedstat_inc(sd, lb_nobusyg[idle]); |
| 2215 | goto out_balanced; |
| 2216 | } |
| 2217 | |
| 2218 | busiest = find_busiest_queue(group, idle); |
| 2219 | if (!busiest) { |
| 2220 | schedstat_inc(sd, lb_nobusyq[idle]); |
| 2221 | goto out_balanced; |
| 2222 | } |
| 2223 | |
| 2224 | BUG_ON(busiest == this_rq); |
| 2225 | |
| 2226 | schedstat_add(sd, lb_imbalance[idle], imbalance); |
| 2227 | |
| 2228 | nr_moved = 0; |
| 2229 | if (busiest->nr_running > 1) { |
| 2230 | /* |
| 2231 | * Attempt to move tasks. If find_busiest_group has found |
| 2232 | * an imbalance but busiest->nr_running <= 1, the group is |
| 2233 | * still unbalanced. nr_moved simply stays zero, so it is |
| 2234 | * correctly treated as an imbalance. |
| 2235 | */ |
| 2236 | double_rq_lock(this_rq, busiest); |
| 2237 | nr_moved = move_tasks(this_rq, this_cpu, busiest, |
| 2238 | imbalance, sd, idle, &all_pinned); |
| 2239 | double_rq_unlock(this_rq, busiest); |
| 2240 | |
| 2241 | /* All tasks on this runqueue were pinned by CPU affinity */ |
| 2242 | if (unlikely(all_pinned)) |
| 2243 | goto out_balanced; |
| 2244 | } |
| 2245 | |
| 2246 | if (!nr_moved) { |
| 2247 | schedstat_inc(sd, lb_failed[idle]); |
| 2248 | sd->nr_balance_failed++; |
| 2249 | |
| 2250 | if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) { |
| 2251 | |
| 2252 | spin_lock(&busiest->lock); |
| 2253 | |
| 2254 | /* don't kick the migration_thread, if the curr |
| 2255 | * task on busiest cpu can't be moved to this_cpu |
| 2256 | */ |
| 2257 | if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) { |
| 2258 | spin_unlock(&busiest->lock); |
| 2259 | all_pinned = 1; |
| 2260 | goto out_one_pinned; |
| 2261 | } |
| 2262 | |
| 2263 | if (!busiest->active_balance) { |
| 2264 | busiest->active_balance = 1; |
| 2265 | busiest->push_cpu = this_cpu; |
| 2266 | active_balance = 1; |
| 2267 | } |
| 2268 | spin_unlock(&busiest->lock); |
| 2269 | if (active_balance) |
| 2270 | wake_up_process(busiest->migration_thread); |
| 2271 | |
| 2272 | /* |
| 2273 | * We've kicked active balancing, reset the failure |
| 2274 | * counter. |
| 2275 | */ |
| 2276 | sd->nr_balance_failed = sd->cache_nice_tries+1; |
| 2277 | } |
| 2278 | } else |
| 2279 | sd->nr_balance_failed = 0; |
| 2280 | |
| 2281 | if (likely(!active_balance)) { |
| 2282 | /* We were unbalanced, so reset the balancing interval */ |
| 2283 | sd->balance_interval = sd->min_interval; |
| 2284 | } else { |
| 2285 | /* |
| 2286 | * If we've begun active balancing, start to back off. This |
| 2287 | * case may not be covered by the all_pinned logic if there |
| 2288 | * is only 1 task on the busy runqueue (because we don't call |
| 2289 | * move_tasks). |
| 2290 | */ |
| 2291 | if (sd->balance_interval < sd->max_interval) |
| 2292 | sd->balance_interval *= 2; |
| 2293 | } |
| 2294 | |
| 2295 | if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER) |
| 2296 | return -1; |
| 2297 | return nr_moved; |
| 2298 | |
| 2299 | out_balanced: |
| 2300 | schedstat_inc(sd, lb_balanced[idle]); |
| 2301 | |
| 2302 | sd->nr_balance_failed = 0; |
| 2303 | |
| 2304 | out_one_pinned: |
| 2305 | /* tune up the balancing interval */ |
| 2306 | if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) || |
| 2307 | (sd->balance_interval < sd->max_interval)) |
| 2308 | sd->balance_interval *= 2; |
| 2309 | |
| 2310 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER) |
| 2311 | return -1; |
| 2312 | return 0; |
| 2313 | } |
| 2314 | |
| 2315 | /* |
| 2316 | * Check this_cpu to ensure it is balanced within domain. Attempt to move |
| 2317 | * tasks if there is an imbalance. |
| 2318 | * |
| 2319 | * Called from schedule when this_rq is about to become idle (NEWLY_IDLE). |
| 2320 | * this_rq is locked. |
| 2321 | */ |
| 2322 | static int load_balance_newidle(int this_cpu, runqueue_t *this_rq, |
| 2323 | struct sched_domain *sd) |
| 2324 | { |
| 2325 | struct sched_group *group; |
| 2326 | runqueue_t *busiest = NULL; |
| 2327 | unsigned long imbalance; |
| 2328 | int nr_moved = 0; |
| 2329 | int sd_idle = 0; |
| 2330 | |
| 2331 | if (sd->flags & SD_SHARE_CPUPOWER) |
| 2332 | sd_idle = 1; |
| 2333 | |
| 2334 | schedstat_inc(sd, lb_cnt[NEWLY_IDLE]); |
| 2335 | group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE, &sd_idle); |
| 2336 | if (!group) { |
| 2337 | schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]); |
| 2338 | goto out_balanced; |
| 2339 | } |
| 2340 | |
| 2341 | busiest = find_busiest_queue(group, NEWLY_IDLE); |
| 2342 | if (!busiest) { |
| 2343 | schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]); |
| 2344 | goto out_balanced; |
| 2345 | } |
| 2346 | |
| 2347 | BUG_ON(busiest == this_rq); |
| 2348 | |
| 2349 | schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance); |
| 2350 | |
| 2351 | nr_moved = 0; |
| 2352 | if (busiest->nr_running > 1) { |
| 2353 | /* Attempt to move tasks */ |
| 2354 | double_lock_balance(this_rq, busiest); |
| 2355 | nr_moved = move_tasks(this_rq, this_cpu, busiest, |
| 2356 | imbalance, sd, NEWLY_IDLE, NULL); |
| 2357 | spin_unlock(&busiest->lock); |
| 2358 | } |
| 2359 | |
| 2360 | if (!nr_moved) { |
| 2361 | schedstat_inc(sd, lb_failed[NEWLY_IDLE]); |
| 2362 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER) |
| 2363 | return -1; |
| 2364 | } else |
| 2365 | sd->nr_balance_failed = 0; |
| 2366 | |
| 2367 | return nr_moved; |
| 2368 | |
| 2369 | out_balanced: |
| 2370 | schedstat_inc(sd, lb_balanced[NEWLY_IDLE]); |
| 2371 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER) |
| 2372 | return -1; |
| 2373 | sd->nr_balance_failed = 0; |
| 2374 | return 0; |
| 2375 | } |
| 2376 | |
| 2377 | /* |
| 2378 | * idle_balance is called by schedule() if this_cpu is about to become |
| 2379 | * idle. Attempts to pull tasks from other CPUs. |
| 2380 | */ |
| 2381 | static void idle_balance(int this_cpu, runqueue_t *this_rq) |
| 2382 | { |
| 2383 | struct sched_domain *sd; |
| 2384 | |
| 2385 | for_each_domain(this_cpu, sd) { |
| 2386 | if (sd->flags & SD_BALANCE_NEWIDLE) { |
| 2387 | if (load_balance_newidle(this_cpu, this_rq, sd)) { |
| 2388 | /* We've pulled tasks over so stop searching */ |
| 2389 | break; |
| 2390 | } |
| 2391 | } |
| 2392 | } |
| 2393 | } |
| 2394 | |
| 2395 | /* |
| 2396 | * active_load_balance is run by migration threads. It pushes running tasks |
| 2397 | * off the busiest CPU onto idle CPUs. It requires at least 1 task to be |
| 2398 | * running on each physical CPU where possible, and avoids physical / |
| 2399 | * logical imbalances. |
| 2400 | * |
| 2401 | * Called with busiest_rq locked. |
| 2402 | */ |
| 2403 | static void active_load_balance(runqueue_t *busiest_rq, int busiest_cpu) |
| 2404 | { |
| 2405 | struct sched_domain *sd; |
| 2406 | runqueue_t *target_rq; |
| 2407 | int target_cpu = busiest_rq->push_cpu; |
| 2408 | |
| 2409 | if (busiest_rq->nr_running <= 1) |
| 2410 | /* no task to move */ |
| 2411 | return; |
| 2412 | |
| 2413 | target_rq = cpu_rq(target_cpu); |
| 2414 | |
| 2415 | /* |
| 2416 | * This condition is "impossible", if it occurs |
| 2417 | * we need to fix it. Originally reported by |
| 2418 | * Bjorn Helgaas on a 128-cpu setup. |
| 2419 | */ |
| 2420 | BUG_ON(busiest_rq == target_rq); |
| 2421 | |
| 2422 | /* move a task from busiest_rq to target_rq */ |
| 2423 | double_lock_balance(busiest_rq, target_rq); |
| 2424 | |
| 2425 | /* Search for an sd spanning us and the target CPU. */ |
| 2426 | for_each_domain(target_cpu, sd) |
| 2427 | if ((sd->flags & SD_LOAD_BALANCE) && |
| 2428 | cpu_isset(busiest_cpu, sd->span)) |
| 2429 | break; |
| 2430 | |
| 2431 | if (unlikely(sd == NULL)) |
| 2432 | goto out; |
| 2433 | |
| 2434 | schedstat_inc(sd, alb_cnt); |
| 2435 | |
| 2436 | if (move_tasks(target_rq, target_cpu, busiest_rq, 1, sd, SCHED_IDLE, NULL)) |
| 2437 | schedstat_inc(sd, alb_pushed); |
| 2438 | else |
| 2439 | schedstat_inc(sd, alb_failed); |
| 2440 | out: |
| 2441 | spin_unlock(&target_rq->lock); |
| 2442 | } |
| 2443 | |
| 2444 | /* |
| 2445 | * rebalance_tick will get called every timer tick, on every CPU. |
| 2446 | * |
| 2447 | * It checks each scheduling domain to see if it is due to be balanced, |
| 2448 | * and initiates a balancing operation if so. |
| 2449 | * |
| 2450 | * Balancing parameters are set up in arch_init_sched_domains. |
| 2451 | */ |
| 2452 | |
| 2453 | /* Don't have all balancing operations going off at once */ |
| 2454 | #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS) |
| 2455 | |
| 2456 | static void rebalance_tick(int this_cpu, runqueue_t *this_rq, |
| 2457 | enum idle_type idle) |
| 2458 | { |
| 2459 | unsigned long old_load, this_load; |
| 2460 | unsigned long j = jiffies + CPU_OFFSET(this_cpu); |
| 2461 | struct sched_domain *sd; |
| 2462 | int i; |
| 2463 | |
| 2464 | this_load = this_rq->nr_running * SCHED_LOAD_SCALE; |
| 2465 | /* Update our load */ |
| 2466 | for (i = 0; i < 3; i++) { |
| 2467 | unsigned long new_load = this_load; |
| 2468 | int scale = 1 << i; |
| 2469 | old_load = this_rq->cpu_load[i]; |
| 2470 | /* |
| 2471 | * Round up the averaging division if load is increasing. This |
| 2472 | * prevents us from getting stuck on 9 if the load is 10, for |
| 2473 | * example. |
| 2474 | */ |
| 2475 | if (new_load > old_load) |
| 2476 | new_load += scale-1; |
| 2477 | this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) / scale; |
| 2478 | } |
| 2479 | |
| 2480 | for_each_domain(this_cpu, sd) { |
| 2481 | unsigned long interval; |
| 2482 | |
| 2483 | if (!(sd->flags & SD_LOAD_BALANCE)) |
| 2484 | continue; |
| 2485 | |
| 2486 | interval = sd->balance_interval; |
| 2487 | if (idle != SCHED_IDLE) |
| 2488 | interval *= sd->busy_factor; |
| 2489 | |
| 2490 | /* scale ms to jiffies */ |
| 2491 | interval = msecs_to_jiffies(interval); |
| 2492 | if (unlikely(!interval)) |
| 2493 | interval = 1; |
| 2494 | |
| 2495 | if (j - sd->last_balance >= interval) { |
| 2496 | if (load_balance(this_cpu, this_rq, sd, idle)) { |
| 2497 | /* |
| 2498 | * We've pulled tasks over so either we're no |
| 2499 | * longer idle, or one of our SMT siblings is |
| 2500 | * not idle. |
| 2501 | */ |
| 2502 | idle = NOT_IDLE; |
| 2503 | } |
| 2504 | sd->last_balance += interval; |
| 2505 | } |
| 2506 | } |
| 2507 | } |
| 2508 | #else |
| 2509 | /* |
| 2510 | * on UP we do not need to balance between CPUs: |
| 2511 | */ |
| 2512 | static inline void rebalance_tick(int cpu, runqueue_t *rq, enum idle_type idle) |
| 2513 | { |
| 2514 | } |
| 2515 | static inline void idle_balance(int cpu, runqueue_t *rq) |
| 2516 | { |
| 2517 | } |
| 2518 | #endif |
| 2519 | |
| 2520 | static inline int wake_priority_sleeper(runqueue_t *rq) |
| 2521 | { |
| 2522 | int ret = 0; |
| 2523 | #ifdef CONFIG_SCHED_SMT |
| 2524 | spin_lock(&rq->lock); |
| 2525 | /* |
| 2526 | * If an SMT sibling task has been put to sleep for priority |
| 2527 | * reasons reschedule the idle task to see if it can now run. |
| 2528 | */ |
| 2529 | if (rq->nr_running) { |
| 2530 | resched_task(rq->idle); |
| 2531 | ret = 1; |
| 2532 | } |
| 2533 | spin_unlock(&rq->lock); |
| 2534 | #endif |
| 2535 | return ret; |
| 2536 | } |
| 2537 | |
| 2538 | DEFINE_PER_CPU(struct kernel_stat, kstat); |
| 2539 | |
| 2540 | EXPORT_PER_CPU_SYMBOL(kstat); |
| 2541 | |
| 2542 | /* |
| 2543 | * This is called on clock ticks and on context switches. |
| 2544 | * Bank in p->sched_time the ns elapsed since the last tick or switch. |
| 2545 | */ |
| 2546 | static inline void update_cpu_clock(task_t *p, runqueue_t *rq, |
| 2547 | unsigned long long now) |
| 2548 | { |
| 2549 | unsigned long long last = max(p->timestamp, rq->timestamp_last_tick); |
| 2550 | p->sched_time += now - last; |
| 2551 | } |
| 2552 | |
| 2553 | /* |
| 2554 | * Return current->sched_time plus any more ns on the sched_clock |
| 2555 | * that have not yet been banked. |
| 2556 | */ |
| 2557 | unsigned long long current_sched_time(const task_t *tsk) |
| 2558 | { |
| 2559 | unsigned long long ns; |
| 2560 | unsigned long flags; |
| 2561 | local_irq_save(flags); |
| 2562 | ns = max(tsk->timestamp, task_rq(tsk)->timestamp_last_tick); |
| 2563 | ns = tsk->sched_time + (sched_clock() - ns); |
| 2564 | local_irq_restore(flags); |
| 2565 | return ns; |
| 2566 | } |
| 2567 | |
| 2568 | /* |
| 2569 | * We place interactive tasks back into the active array, if possible. |
| 2570 | * |
| 2571 | * To guarantee that this does not starve expired tasks we ignore the |
| 2572 | * interactivity of a task if the first expired task had to wait more |
| 2573 | * than a 'reasonable' amount of time. This deadline timeout is |
| 2574 | * load-dependent, as the frequency of array switched decreases with |
| 2575 | * increasing number of running tasks. We also ignore the interactivity |
| 2576 | * if a better static_prio task has expired: |
| 2577 | */ |
| 2578 | #define EXPIRED_STARVING(rq) \ |
| 2579 | ((STARVATION_LIMIT && ((rq)->expired_timestamp && \ |
| 2580 | (jiffies - (rq)->expired_timestamp >= \ |
| 2581 | STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \ |
| 2582 | ((rq)->curr->static_prio > (rq)->best_expired_prio)) |
| 2583 | |
| 2584 | /* |
| 2585 | * Account user cpu time to a process. |
| 2586 | * @p: the process that the cpu time gets accounted to |
| 2587 | * @hardirq_offset: the offset to subtract from hardirq_count() |
| 2588 | * @cputime: the cpu time spent in user space since the last update |
| 2589 | */ |
| 2590 | void account_user_time(struct task_struct *p, cputime_t cputime) |
| 2591 | { |
| 2592 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; |
| 2593 | cputime64_t tmp; |
| 2594 | |
| 2595 | p->utime = cputime_add(p->utime, cputime); |
| 2596 | |
| 2597 | /* Add user time to cpustat. */ |
| 2598 | tmp = cputime_to_cputime64(cputime); |
| 2599 | if (TASK_NICE(p) > 0) |
| 2600 | cpustat->nice = cputime64_add(cpustat->nice, tmp); |
| 2601 | else |
| 2602 | cpustat->user = cputime64_add(cpustat->user, tmp); |
| 2603 | } |
| 2604 | |
| 2605 | /* |
| 2606 | * Account system cpu time to a process. |
| 2607 | * @p: the process that the cpu time gets accounted to |
| 2608 | * @hardirq_offset: the offset to subtract from hardirq_count() |
| 2609 | * @cputime: the cpu time spent in kernel space since the last update |
| 2610 | */ |
| 2611 | void account_system_time(struct task_struct *p, int hardirq_offset, |
| 2612 | cputime_t cputime) |
| 2613 | { |
| 2614 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; |
| 2615 | runqueue_t *rq = this_rq(); |
| 2616 | cputime64_t tmp; |
| 2617 | |
| 2618 | p->stime = cputime_add(p->stime, cputime); |
| 2619 | |
| 2620 | /* Add system time to cpustat. */ |
| 2621 | tmp = cputime_to_cputime64(cputime); |
| 2622 | if (hardirq_count() - hardirq_offset) |
| 2623 | cpustat->irq = cputime64_add(cpustat->irq, tmp); |
| 2624 | else if (softirq_count()) |
| 2625 | cpustat->softirq = cputime64_add(cpustat->softirq, tmp); |
| 2626 | else if (p != rq->idle) |
| 2627 | cpustat->system = cputime64_add(cpustat->system, tmp); |
| 2628 | else if (atomic_read(&rq->nr_iowait) > 0) |
| 2629 | cpustat->iowait = cputime64_add(cpustat->iowait, tmp); |
| 2630 | else |
| 2631 | cpustat->idle = cputime64_add(cpustat->idle, tmp); |
| 2632 | /* Account for system time used */ |
| 2633 | acct_update_integrals(p); |
| 2634 | } |
| 2635 | |
| 2636 | /* |
| 2637 | * Account for involuntary wait time. |
| 2638 | * @p: the process from which the cpu time has been stolen |
| 2639 | * @steal: the cpu time spent in involuntary wait |
| 2640 | */ |
| 2641 | void account_steal_time(struct task_struct *p, cputime_t steal) |
| 2642 | { |
| 2643 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; |
| 2644 | cputime64_t tmp = cputime_to_cputime64(steal); |
| 2645 | runqueue_t *rq = this_rq(); |
| 2646 | |
| 2647 | if (p == rq->idle) { |
| 2648 | p->stime = cputime_add(p->stime, steal); |
| 2649 | if (atomic_read(&rq->nr_iowait) > 0) |
| 2650 | cpustat->iowait = cputime64_add(cpustat->iowait, tmp); |
| 2651 | else |
| 2652 | cpustat->idle = cputime64_add(cpustat->idle, tmp); |
| 2653 | } else |
| 2654 | cpustat->steal = cputime64_add(cpustat->steal, tmp); |
| 2655 | } |
| 2656 | |
| 2657 | /* |
| 2658 | * This function gets called by the timer code, with HZ frequency. |
| 2659 | * We call it with interrupts disabled. |
| 2660 | * |
| 2661 | * It also gets called by the fork code, when changing the parent's |
| 2662 | * timeslices. |
| 2663 | */ |
| 2664 | void scheduler_tick(void) |
| 2665 | { |
| 2666 | int cpu = smp_processor_id(); |
| 2667 | runqueue_t *rq = this_rq(); |
| 2668 | task_t *p = current; |
| 2669 | unsigned long long now = sched_clock(); |
| 2670 | |
| 2671 | update_cpu_clock(p, rq, now); |
| 2672 | |
| 2673 | rq->timestamp_last_tick = now; |
| 2674 | |
| 2675 | if (p == rq->idle) { |
| 2676 | if (wake_priority_sleeper(rq)) |
| 2677 | goto out; |
| 2678 | rebalance_tick(cpu, rq, SCHED_IDLE); |
| 2679 | return; |
| 2680 | } |
| 2681 | |
| 2682 | /* Task might have expired already, but not scheduled off yet */ |
| 2683 | if (p->array != rq->active) { |
| 2684 | set_tsk_need_resched(p); |
| 2685 | goto out; |
| 2686 | } |
| 2687 | spin_lock(&rq->lock); |
| 2688 | /* |
| 2689 | * The task was running during this tick - update the |
| 2690 | * time slice counter. Note: we do not update a thread's |
| 2691 | * priority until it either goes to sleep or uses up its |
| 2692 | * timeslice. This makes it possible for interactive tasks |
| 2693 | * to use up their timeslices at their highest priority levels. |
| 2694 | */ |
| 2695 | if (rt_task(p)) { |
| 2696 | /* |
| 2697 | * RR tasks need a special form of timeslice management. |
| 2698 | * FIFO tasks have no timeslices. |
| 2699 | */ |
| 2700 | if ((p->policy == SCHED_RR) && !--p->time_slice) { |
| 2701 | p->time_slice = task_timeslice(p); |
| 2702 | p->first_time_slice = 0; |
| 2703 | set_tsk_need_resched(p); |
| 2704 | |
| 2705 | /* put it at the end of the queue: */ |
| 2706 | requeue_task(p, rq->active); |
| 2707 | } |
| 2708 | goto out_unlock; |
| 2709 | } |
| 2710 | if (!--p->time_slice) { |
| 2711 | dequeue_task(p, rq->active); |
| 2712 | set_tsk_need_resched(p); |
| 2713 | p->prio = effective_prio(p); |
| 2714 | p->time_slice = task_timeslice(p); |
| 2715 | p->first_time_slice = 0; |
| 2716 | |
| 2717 | if (!rq->expired_timestamp) |
| 2718 | rq->expired_timestamp = jiffies; |
| 2719 | if (!TASK_INTERACTIVE(p) || EXPIRED_STARVING(rq)) { |
| 2720 | enqueue_task(p, rq->expired); |
| 2721 | if (p->static_prio < rq->best_expired_prio) |
| 2722 | rq->best_expired_prio = p->static_prio; |
| 2723 | } else |
| 2724 | enqueue_task(p, rq->active); |
| 2725 | } else { |
| 2726 | /* |
| 2727 | * Prevent a too long timeslice allowing a task to monopolize |
| 2728 | * the CPU. We do this by splitting up the timeslice into |
| 2729 | * smaller pieces. |
| 2730 | * |
| 2731 | * Note: this does not mean the task's timeslices expire or |
| 2732 | * get lost in any way, they just might be preempted by |
| 2733 | * another task of equal priority. (one with higher |
| 2734 | * priority would have preempted this task already.) We |
| 2735 | * requeue this task to the end of the list on this priority |
| 2736 | * level, which is in essence a round-robin of tasks with |
| 2737 | * equal priority. |
| 2738 | * |
| 2739 | * This only applies to tasks in the interactive |
| 2740 | * delta range with at least TIMESLICE_GRANULARITY to requeue. |
| 2741 | */ |
| 2742 | if (TASK_INTERACTIVE(p) && !((task_timeslice(p) - |
| 2743 | p->time_slice) % TIMESLICE_GRANULARITY(p)) && |
| 2744 | (p->time_slice >= TIMESLICE_GRANULARITY(p)) && |
| 2745 | (p->array == rq->active)) { |
| 2746 | |
| 2747 | requeue_task(p, rq->active); |
| 2748 | set_tsk_need_resched(p); |
| 2749 | } |
| 2750 | } |
| 2751 | out_unlock: |
| 2752 | spin_unlock(&rq->lock); |
| 2753 | out: |
| 2754 | rebalance_tick(cpu, rq, NOT_IDLE); |
| 2755 | } |
| 2756 | |
| 2757 | #ifdef CONFIG_SCHED_SMT |
| 2758 | static inline void wakeup_busy_runqueue(runqueue_t *rq) |
| 2759 | { |
| 2760 | /* If an SMT runqueue is sleeping due to priority reasons wake it up */ |
| 2761 | if (rq->curr == rq->idle && rq->nr_running) |
| 2762 | resched_task(rq->idle); |
| 2763 | } |
| 2764 | |
| 2765 | static void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq) |
| 2766 | { |
| 2767 | struct sched_domain *tmp, *sd = NULL; |
| 2768 | cpumask_t sibling_map; |
| 2769 | int i; |
| 2770 | |
| 2771 | for_each_domain(this_cpu, tmp) |
| 2772 | if (tmp->flags & SD_SHARE_CPUPOWER) |
| 2773 | sd = tmp; |
| 2774 | |
| 2775 | if (!sd) |
| 2776 | return; |
| 2777 | |
| 2778 | /* |
| 2779 | * Unlock the current runqueue because we have to lock in |
| 2780 | * CPU order to avoid deadlocks. Caller knows that we might |
| 2781 | * unlock. We keep IRQs disabled. |
| 2782 | */ |
| 2783 | spin_unlock(&this_rq->lock); |
| 2784 | |
| 2785 | sibling_map = sd->span; |
| 2786 | |
| 2787 | for_each_cpu_mask(i, sibling_map) |
| 2788 | spin_lock(&cpu_rq(i)->lock); |
| 2789 | /* |
| 2790 | * We clear this CPU from the mask. This both simplifies the |
| 2791 | * inner loop and keps this_rq locked when we exit: |
| 2792 | */ |
| 2793 | cpu_clear(this_cpu, sibling_map); |
| 2794 | |
| 2795 | for_each_cpu_mask(i, sibling_map) { |
| 2796 | runqueue_t *smt_rq = cpu_rq(i); |
| 2797 | |
| 2798 | wakeup_busy_runqueue(smt_rq); |
| 2799 | } |
| 2800 | |
| 2801 | for_each_cpu_mask(i, sibling_map) |
| 2802 | spin_unlock(&cpu_rq(i)->lock); |
| 2803 | /* |
| 2804 | * We exit with this_cpu's rq still held and IRQs |
| 2805 | * still disabled: |
| 2806 | */ |
| 2807 | } |
| 2808 | |
| 2809 | /* |
| 2810 | * number of 'lost' timeslices this task wont be able to fully |
| 2811 | * utilize, if another task runs on a sibling. This models the |
| 2812 | * slowdown effect of other tasks running on siblings: |
| 2813 | */ |
| 2814 | static inline unsigned long smt_slice(task_t *p, struct sched_domain *sd) |
| 2815 | { |
| 2816 | return p->time_slice * (100 - sd->per_cpu_gain) / 100; |
| 2817 | } |
| 2818 | |
| 2819 | static int dependent_sleeper(int this_cpu, runqueue_t *this_rq) |
| 2820 | { |
| 2821 | struct sched_domain *tmp, *sd = NULL; |
| 2822 | cpumask_t sibling_map; |
| 2823 | prio_array_t *array; |
| 2824 | int ret = 0, i; |
| 2825 | task_t *p; |
| 2826 | |
| 2827 | for_each_domain(this_cpu, tmp) |
| 2828 | if (tmp->flags & SD_SHARE_CPUPOWER) |
| 2829 | sd = tmp; |
| 2830 | |
| 2831 | if (!sd) |
| 2832 | return 0; |
| 2833 | |
| 2834 | /* |
| 2835 | * The same locking rules and details apply as for |
| 2836 | * wake_sleeping_dependent(): |
| 2837 | */ |
| 2838 | spin_unlock(&this_rq->lock); |
| 2839 | sibling_map = sd->span; |
| 2840 | for_each_cpu_mask(i, sibling_map) |
| 2841 | spin_lock(&cpu_rq(i)->lock); |
| 2842 | cpu_clear(this_cpu, sibling_map); |
| 2843 | |
| 2844 | /* |
| 2845 | * Establish next task to be run - it might have gone away because |
| 2846 | * we released the runqueue lock above: |
| 2847 | */ |
| 2848 | if (!this_rq->nr_running) |
| 2849 | goto out_unlock; |
| 2850 | array = this_rq->active; |
| 2851 | if (!array->nr_active) |
| 2852 | array = this_rq->expired; |
| 2853 | BUG_ON(!array->nr_active); |
| 2854 | |
| 2855 | p = list_entry(array->queue[sched_find_first_bit(array->bitmap)].next, |
| 2856 | task_t, run_list); |
| 2857 | |
| 2858 | for_each_cpu_mask(i, sibling_map) { |
| 2859 | runqueue_t *smt_rq = cpu_rq(i); |
| 2860 | task_t *smt_curr = smt_rq->curr; |
| 2861 | |
| 2862 | /* Kernel threads do not participate in dependent sleeping */ |
| 2863 | if (!p->mm || !smt_curr->mm || rt_task(p)) |
| 2864 | goto check_smt_task; |
| 2865 | |
| 2866 | /* |
| 2867 | * If a user task with lower static priority than the |
| 2868 | * running task on the SMT sibling is trying to schedule, |
| 2869 | * delay it till there is proportionately less timeslice |
| 2870 | * left of the sibling task to prevent a lower priority |
| 2871 | * task from using an unfair proportion of the |
| 2872 | * physical cpu's resources. -ck |
| 2873 | */ |
| 2874 | if (rt_task(smt_curr)) { |
| 2875 | /* |
| 2876 | * With real time tasks we run non-rt tasks only |
| 2877 | * per_cpu_gain% of the time. |
| 2878 | */ |
| 2879 | if ((jiffies % DEF_TIMESLICE) > |
| 2880 | (sd->per_cpu_gain * DEF_TIMESLICE / 100)) |
| 2881 | ret = 1; |
| 2882 | } else |
| 2883 | if (smt_curr->static_prio < p->static_prio && |
| 2884 | !TASK_PREEMPTS_CURR(p, smt_rq) && |
| 2885 | smt_slice(smt_curr, sd) > task_timeslice(p)) |
| 2886 | ret = 1; |
| 2887 | |
| 2888 | check_smt_task: |
| 2889 | if ((!smt_curr->mm && smt_curr != smt_rq->idle) || |
| 2890 | rt_task(smt_curr)) |
| 2891 | continue; |
| 2892 | if (!p->mm) { |
| 2893 | wakeup_busy_runqueue(smt_rq); |
| 2894 | continue; |
| 2895 | } |
| 2896 | |
| 2897 | /* |
| 2898 | * Reschedule a lower priority task on the SMT sibling for |
| 2899 | * it to be put to sleep, or wake it up if it has been put to |
| 2900 | * sleep for priority reasons to see if it should run now. |
| 2901 | */ |
| 2902 | if (rt_task(p)) { |
| 2903 | if ((jiffies % DEF_TIMESLICE) > |
| 2904 | (sd->per_cpu_gain * DEF_TIMESLICE / 100)) |
| 2905 | resched_task(smt_curr); |
| 2906 | } else { |
| 2907 | if (TASK_PREEMPTS_CURR(p, smt_rq) && |
| 2908 | smt_slice(p, sd) > task_timeslice(smt_curr)) |
| 2909 | resched_task(smt_curr); |
| 2910 | else |
| 2911 | wakeup_busy_runqueue(smt_rq); |
| 2912 | } |
| 2913 | } |
| 2914 | out_unlock: |
| 2915 | for_each_cpu_mask(i, sibling_map) |
| 2916 | spin_unlock(&cpu_rq(i)->lock); |
| 2917 | return ret; |
| 2918 | } |
| 2919 | #else |
| 2920 | static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq) |
| 2921 | { |
| 2922 | } |
| 2923 | |
| 2924 | static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq) |
| 2925 | { |
| 2926 | return 0; |
| 2927 | } |
| 2928 | #endif |
| 2929 | |
| 2930 | #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT) |
| 2931 | |
| 2932 | void fastcall add_preempt_count(int val) |
| 2933 | { |
| 2934 | /* |
| 2935 | * Underflow? |
| 2936 | */ |
| 2937 | BUG_ON((preempt_count() < 0)); |
| 2938 | preempt_count() += val; |
| 2939 | /* |
| 2940 | * Spinlock count overflowing soon? |
| 2941 | */ |
| 2942 | BUG_ON((preempt_count() & PREEMPT_MASK) >= PREEMPT_MASK-10); |
| 2943 | } |
| 2944 | EXPORT_SYMBOL(add_preempt_count); |
| 2945 | |
| 2946 | void fastcall sub_preempt_count(int val) |
| 2947 | { |
| 2948 | /* |
| 2949 | * Underflow? |
| 2950 | */ |
| 2951 | BUG_ON(val > preempt_count()); |
| 2952 | /* |
| 2953 | * Is the spinlock portion underflowing? |
| 2954 | */ |
| 2955 | BUG_ON((val < PREEMPT_MASK) && !(preempt_count() & PREEMPT_MASK)); |
| 2956 | preempt_count() -= val; |
| 2957 | } |
| 2958 | EXPORT_SYMBOL(sub_preempt_count); |
| 2959 | |
| 2960 | #endif |
| 2961 | |
| 2962 | /* |
| 2963 | * schedule() is the main scheduler function. |
| 2964 | */ |
| 2965 | asmlinkage void __sched schedule(void) |
| 2966 | { |
| 2967 | long *switch_count; |
| 2968 | task_t *prev, *next; |
| 2969 | runqueue_t *rq; |
| 2970 | prio_array_t *array; |
| 2971 | struct list_head *queue; |
| 2972 | unsigned long long now; |
| 2973 | unsigned long run_time; |
| 2974 | int cpu, idx, new_prio; |
| 2975 | |
| 2976 | /* |
| 2977 | * Test if we are atomic. Since do_exit() needs to call into |
| 2978 | * schedule() atomically, we ignore that path for now. |
| 2979 | * Otherwise, whine if we are scheduling when we should not be. |
| 2980 | */ |
| 2981 | if (likely(!current->exit_state)) { |
| 2982 | if (unlikely(in_atomic())) { |
| 2983 | printk(KERN_ERR "scheduling while atomic: " |
| 2984 | "%s/0x%08x/%d\n", |
| 2985 | current->comm, preempt_count(), current->pid); |
| 2986 | dump_stack(); |
| 2987 | } |
| 2988 | } |
| 2989 | profile_hit(SCHED_PROFILING, __builtin_return_address(0)); |
| 2990 | |
| 2991 | need_resched: |
| 2992 | preempt_disable(); |
| 2993 | prev = current; |
| 2994 | release_kernel_lock(prev); |
| 2995 | need_resched_nonpreemptible: |
| 2996 | rq = this_rq(); |
| 2997 | |
| 2998 | /* |
| 2999 | * The idle thread is not allowed to schedule! |
| 3000 | * Remove this check after it has been exercised a bit. |
| 3001 | */ |
| 3002 | if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) { |
| 3003 | printk(KERN_ERR "bad: scheduling from the idle thread!\n"); |
| 3004 | dump_stack(); |
| 3005 | } |
| 3006 | |
| 3007 | schedstat_inc(rq, sched_cnt); |
| 3008 | now = sched_clock(); |
| 3009 | if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) { |
| 3010 | run_time = now - prev->timestamp; |
| 3011 | if (unlikely((long long)(now - prev->timestamp) < 0)) |
| 3012 | run_time = 0; |
| 3013 | } else |
| 3014 | run_time = NS_MAX_SLEEP_AVG; |
| 3015 | |
| 3016 | /* |
| 3017 | * Tasks charged proportionately less run_time at high sleep_avg to |
| 3018 | * delay them losing their interactive status |
| 3019 | */ |
| 3020 | run_time /= (CURRENT_BONUS(prev) ? : 1); |
| 3021 | |
| 3022 | spin_lock_irq(&rq->lock); |
| 3023 | |
| 3024 | if (unlikely(prev->flags & PF_DEAD)) |
| 3025 | prev->state = EXIT_DEAD; |
| 3026 | |
| 3027 | switch_count = &prev->nivcsw; |
| 3028 | if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { |
| 3029 | switch_count = &prev->nvcsw; |
| 3030 | if (unlikely((prev->state & TASK_INTERRUPTIBLE) && |
| 3031 | unlikely(signal_pending(prev)))) |
| 3032 | prev->state = TASK_RUNNING; |
| 3033 | else { |
| 3034 | if (prev->state == TASK_UNINTERRUPTIBLE) |
| 3035 | rq->nr_uninterruptible++; |
| 3036 | deactivate_task(prev, rq); |
| 3037 | } |
| 3038 | } |
| 3039 | |
| 3040 | cpu = smp_processor_id(); |
| 3041 | if (unlikely(!rq->nr_running)) { |
| 3042 | go_idle: |
| 3043 | idle_balance(cpu, rq); |
| 3044 | if (!rq->nr_running) { |
| 3045 | next = rq->idle; |
| 3046 | rq->expired_timestamp = 0; |
| 3047 | wake_sleeping_dependent(cpu, rq); |
| 3048 | /* |
| 3049 | * wake_sleeping_dependent() might have released |
| 3050 | * the runqueue, so break out if we got new |
| 3051 | * tasks meanwhile: |
| 3052 | */ |
| 3053 | if (!rq->nr_running) |
| 3054 | goto switch_tasks; |
| 3055 | } |
| 3056 | } else { |
| 3057 | if (dependent_sleeper(cpu, rq)) { |
| 3058 | next = rq->idle; |
| 3059 | goto switch_tasks; |
| 3060 | } |
| 3061 | /* |
| 3062 | * dependent_sleeper() releases and reacquires the runqueue |
| 3063 | * lock, hence go into the idle loop if the rq went |
| 3064 | * empty meanwhile: |
| 3065 | */ |
| 3066 | if (unlikely(!rq->nr_running)) |
| 3067 | goto go_idle; |
| 3068 | } |
| 3069 | |
| 3070 | array = rq->active; |
| 3071 | if (unlikely(!array->nr_active)) { |
| 3072 | /* |
| 3073 | * Switch the active and expired arrays. |
| 3074 | */ |
| 3075 | schedstat_inc(rq, sched_switch); |
| 3076 | rq->active = rq->expired; |
| 3077 | rq->expired = array; |
| 3078 | array = rq->active; |
| 3079 | rq->expired_timestamp = 0; |
| 3080 | rq->best_expired_prio = MAX_PRIO; |
| 3081 | } |
| 3082 | |
| 3083 | idx = sched_find_first_bit(array->bitmap); |
| 3084 | queue = array->queue + idx; |
| 3085 | next = list_entry(queue->next, task_t, run_list); |
| 3086 | |
| 3087 | if (!rt_task(next) && next->activated > 0) { |
| 3088 | unsigned long long delta = now - next->timestamp; |
| 3089 | if (unlikely((long long)(now - next->timestamp) < 0)) |
| 3090 | delta = 0; |
| 3091 | |
| 3092 | if (next->activated == 1) |
| 3093 | delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128; |
| 3094 | |
| 3095 | array = next->array; |
| 3096 | new_prio = recalc_task_prio(next, next->timestamp + delta); |
| 3097 | |
| 3098 | if (unlikely(next->prio != new_prio)) { |
| 3099 | dequeue_task(next, array); |
| 3100 | next->prio = new_prio; |
| 3101 | enqueue_task(next, array); |
| 3102 | } else |
| 3103 | requeue_task(next, array); |
| 3104 | } |
| 3105 | next->activated = 0; |
| 3106 | switch_tasks: |
| 3107 | if (next == rq->idle) |
| 3108 | schedstat_inc(rq, sched_goidle); |
| 3109 | prefetch(next); |
| 3110 | prefetch_stack(next); |
| 3111 | clear_tsk_need_resched(prev); |
| 3112 | rcu_qsctr_inc(task_cpu(prev)); |
| 3113 | |
| 3114 | update_cpu_clock(prev, rq, now); |
| 3115 | |
| 3116 | prev->sleep_avg -= run_time; |
| 3117 | if ((long)prev->sleep_avg <= 0) |
| 3118 | prev->sleep_avg = 0; |
| 3119 | prev->timestamp = prev->last_ran = now; |
| 3120 | |
| 3121 | sched_info_switch(prev, next); |
| 3122 | if (likely(prev != next)) { |
| 3123 | next->timestamp = now; |
| 3124 | rq->nr_switches++; |
| 3125 | rq->curr = next; |
| 3126 | ++*switch_count; |
| 3127 | |
| 3128 | prepare_task_switch(rq, next); |
| 3129 | prev = context_switch(rq, prev, next); |
| 3130 | barrier(); |
| 3131 | /* |
| 3132 | * this_rq must be evaluated again because prev may have moved |
| 3133 | * CPUs since it called schedule(), thus the 'rq' on its stack |
| 3134 | * frame will be invalid. |
| 3135 | */ |
| 3136 | finish_task_switch(this_rq(), prev); |
| 3137 | } else |
| 3138 | spin_unlock_irq(&rq->lock); |
| 3139 | |
| 3140 | prev = current; |
| 3141 | if (unlikely(reacquire_kernel_lock(prev) < 0)) |
| 3142 | goto need_resched_nonpreemptible; |
| 3143 | preempt_enable_no_resched(); |
| 3144 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) |
| 3145 | goto need_resched; |
| 3146 | } |
| 3147 | |
| 3148 | EXPORT_SYMBOL(schedule); |
| 3149 | |
| 3150 | #ifdef CONFIG_PREEMPT |
| 3151 | /* |
| 3152 | * this is is the entry point to schedule() from in-kernel preemption |
| 3153 | * off of preempt_enable. Kernel preemptions off return from interrupt |
| 3154 | * occur there and call schedule directly. |
| 3155 | */ |
| 3156 | asmlinkage void __sched preempt_schedule(void) |
| 3157 | { |
| 3158 | struct thread_info *ti = current_thread_info(); |
| 3159 | #ifdef CONFIG_PREEMPT_BKL |
| 3160 | struct task_struct *task = current; |
| 3161 | int saved_lock_depth; |
| 3162 | #endif |
| 3163 | /* |
| 3164 | * If there is a non-zero preempt_count or interrupts are disabled, |
| 3165 | * we do not want to preempt the current task. Just return.. |
| 3166 | */ |
| 3167 | if (unlikely(ti->preempt_count || irqs_disabled())) |
| 3168 | return; |
| 3169 | |
| 3170 | need_resched: |
| 3171 | add_preempt_count(PREEMPT_ACTIVE); |
| 3172 | /* |
| 3173 | * We keep the big kernel semaphore locked, but we |
| 3174 | * clear ->lock_depth so that schedule() doesnt |
| 3175 | * auto-release the semaphore: |
| 3176 | */ |
| 3177 | #ifdef CONFIG_PREEMPT_BKL |
| 3178 | saved_lock_depth = task->lock_depth; |
| 3179 | task->lock_depth = -1; |
| 3180 | #endif |
| 3181 | schedule(); |
| 3182 | #ifdef CONFIG_PREEMPT_BKL |
| 3183 | task->lock_depth = saved_lock_depth; |
| 3184 | #endif |
| 3185 | sub_preempt_count(PREEMPT_ACTIVE); |
| 3186 | |
| 3187 | /* we could miss a preemption opportunity between schedule and now */ |
| 3188 | barrier(); |
| 3189 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) |
| 3190 | goto need_resched; |
| 3191 | } |
| 3192 | |
| 3193 | EXPORT_SYMBOL(preempt_schedule); |
| 3194 | |
| 3195 | /* |
| 3196 | * this is is the entry point to schedule() from kernel preemption |
| 3197 | * off of irq context. |
| 3198 | * Note, that this is called and return with irqs disabled. This will |
| 3199 | * protect us against recursive calling from irq. |
| 3200 | */ |
| 3201 | asmlinkage void __sched preempt_schedule_irq(void) |
| 3202 | { |
| 3203 | struct thread_info *ti = current_thread_info(); |
| 3204 | #ifdef CONFIG_PREEMPT_BKL |
| 3205 | struct task_struct *task = current; |
| 3206 | int saved_lock_depth; |
| 3207 | #endif |
| 3208 | /* Catch callers which need to be fixed*/ |
| 3209 | BUG_ON(ti->preempt_count || !irqs_disabled()); |
| 3210 | |
| 3211 | need_resched: |
| 3212 | add_preempt_count(PREEMPT_ACTIVE); |
| 3213 | /* |
| 3214 | * We keep the big kernel semaphore locked, but we |
| 3215 | * clear ->lock_depth so that schedule() doesnt |
| 3216 | * auto-release the semaphore: |
| 3217 | */ |
| 3218 | #ifdef CONFIG_PREEMPT_BKL |
| 3219 | saved_lock_depth = task->lock_depth; |
| 3220 | task->lock_depth = -1; |
| 3221 | #endif |
| 3222 | local_irq_enable(); |
| 3223 | schedule(); |
| 3224 | local_irq_disable(); |
| 3225 | #ifdef CONFIG_PREEMPT_BKL |
| 3226 | task->lock_depth = saved_lock_depth; |
| 3227 | #endif |
| 3228 | sub_preempt_count(PREEMPT_ACTIVE); |
| 3229 | |
| 3230 | /* we could miss a preemption opportunity between schedule and now */ |
| 3231 | barrier(); |
| 3232 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) |
| 3233 | goto need_resched; |
| 3234 | } |
| 3235 | |
| 3236 | #endif /* CONFIG_PREEMPT */ |
| 3237 | |
| 3238 | int default_wake_function(wait_queue_t *curr, unsigned mode, int sync, |
| 3239 | void *key) |
| 3240 | { |
| 3241 | task_t *p = curr->private; |
| 3242 | return try_to_wake_up(p, mode, sync); |
| 3243 | } |
| 3244 | |
| 3245 | EXPORT_SYMBOL(default_wake_function); |
| 3246 | |
| 3247 | /* |
| 3248 | * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just |
| 3249 | * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve |
| 3250 | * number) then we wake all the non-exclusive tasks and one exclusive task. |
| 3251 | * |
| 3252 | * There are circumstances in which we can try to wake a task which has already |
| 3253 | * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns |
| 3254 | * zero in this (rare) case, and we handle it by continuing to scan the queue. |
| 3255 | */ |
| 3256 | static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, |
| 3257 | int nr_exclusive, int sync, void *key) |
| 3258 | { |
| 3259 | struct list_head *tmp, *next; |
| 3260 | |
| 3261 | list_for_each_safe(tmp, next, &q->task_list) { |
| 3262 | wait_queue_t *curr; |
| 3263 | unsigned flags; |
| 3264 | curr = list_entry(tmp, wait_queue_t, task_list); |
| 3265 | flags = curr->flags; |
| 3266 | if (curr->func(curr, mode, sync, key) && |
| 3267 | (flags & WQ_FLAG_EXCLUSIVE) && |
| 3268 | !--nr_exclusive) |
| 3269 | break; |
| 3270 | } |
| 3271 | } |
| 3272 | |
| 3273 | /** |
| 3274 | * __wake_up - wake up threads blocked on a waitqueue. |
| 3275 | * @q: the waitqueue |
| 3276 | * @mode: which threads |
| 3277 | * @nr_exclusive: how many wake-one or wake-many threads to wake up |
| 3278 | * @key: is directly passed to the wakeup function |
| 3279 | */ |
| 3280 | void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode, |
| 3281 | int nr_exclusive, void *key) |
| 3282 | { |
| 3283 | unsigned long flags; |
| 3284 | |
| 3285 | spin_lock_irqsave(&q->lock, flags); |
| 3286 | __wake_up_common(q, mode, nr_exclusive, 0, key); |
| 3287 | spin_unlock_irqrestore(&q->lock, flags); |
| 3288 | } |
| 3289 | |
| 3290 | EXPORT_SYMBOL(__wake_up); |
| 3291 | |
| 3292 | /* |
| 3293 | * Same as __wake_up but called with the spinlock in wait_queue_head_t held. |
| 3294 | */ |
| 3295 | void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode) |
| 3296 | { |
| 3297 | __wake_up_common(q, mode, 1, 0, NULL); |
| 3298 | } |
| 3299 | |
| 3300 | /** |
| 3301 | * __wake_up_sync - wake up threads blocked on a waitqueue. |
| 3302 | * @q: the waitqueue |
| 3303 | * @mode: which threads |
| 3304 | * @nr_exclusive: how many wake-one or wake-many threads to wake up |
| 3305 | * |
| 3306 | * The sync wakeup differs that the waker knows that it will schedule |
| 3307 | * away soon, so while the target thread will be woken up, it will not |
| 3308 | * be migrated to another CPU - ie. the two threads are 'synchronized' |
| 3309 | * with each other. This can prevent needless bouncing between CPUs. |
| 3310 | * |
| 3311 | * On UP it can prevent extra preemption. |
| 3312 | */ |
| 3313 | void fastcall |
| 3314 | __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) |
| 3315 | { |
| 3316 | unsigned long flags; |
| 3317 | int sync = 1; |
| 3318 | |
| 3319 | if (unlikely(!q)) |
| 3320 | return; |
| 3321 | |
| 3322 | if (unlikely(!nr_exclusive)) |
| 3323 | sync = 0; |
| 3324 | |
| 3325 | spin_lock_irqsave(&q->lock, flags); |
| 3326 | __wake_up_common(q, mode, nr_exclusive, sync, NULL); |
| 3327 | spin_unlock_irqrestore(&q->lock, flags); |
| 3328 | } |
| 3329 | EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ |
| 3330 | |
| 3331 | void fastcall complete(struct completion *x) |
| 3332 | { |
| 3333 | unsigned long flags; |
| 3334 | |
| 3335 | spin_lock_irqsave(&x->wait.lock, flags); |
| 3336 | x->done++; |
| 3337 | __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE, |
| 3338 | 1, 0, NULL); |
| 3339 | spin_unlock_irqrestore(&x->wait.lock, flags); |
| 3340 | } |
| 3341 | EXPORT_SYMBOL(complete); |
| 3342 | |
| 3343 | void fastcall complete_all(struct completion *x) |
| 3344 | { |
| 3345 | unsigned long flags; |
| 3346 | |
| 3347 | spin_lock_irqsave(&x->wait.lock, flags); |
| 3348 | x->done += UINT_MAX/2; |
| 3349 | __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE, |
| 3350 | 0, 0, NULL); |
| 3351 | spin_unlock_irqrestore(&x->wait.lock, flags); |
| 3352 | } |
| 3353 | EXPORT_SYMBOL(complete_all); |
| 3354 | |
| 3355 | void fastcall __sched wait_for_completion(struct completion *x) |
| 3356 | { |
| 3357 | might_sleep(); |
| 3358 | spin_lock_irq(&x->wait.lock); |
| 3359 | if (!x->done) { |
| 3360 | DECLARE_WAITQUEUE(wait, current); |
| 3361 | |
| 3362 | wait.flags |= WQ_FLAG_EXCLUSIVE; |
| 3363 | __add_wait_queue_tail(&x->wait, &wait); |
| 3364 | do { |
| 3365 | __set_current_state(TASK_UNINTERRUPTIBLE); |
| 3366 | spin_unlock_irq(&x->wait.lock); |
| 3367 | schedule(); |
| 3368 | spin_lock_irq(&x->wait.lock); |
| 3369 | } while (!x->done); |
| 3370 | __remove_wait_queue(&x->wait, &wait); |
| 3371 | } |
| 3372 | x->done--; |
| 3373 | spin_unlock_irq(&x->wait.lock); |
| 3374 | } |
| 3375 | EXPORT_SYMBOL(wait_for_completion); |
| 3376 | |
| 3377 | unsigned long fastcall __sched |
| 3378 | wait_for_completion_timeout(struct completion *x, unsigned long timeout) |
| 3379 | { |
| 3380 | might_sleep(); |
| 3381 | |
| 3382 | spin_lock_irq(&x->wait.lock); |
| 3383 | if (!x->done) { |
| 3384 | DECLARE_WAITQUEUE(wait, current); |
| 3385 | |
| 3386 | wait.flags |= WQ_FLAG_EXCLUSIVE; |
| 3387 | __add_wait_queue_tail(&x->wait, &wait); |
| 3388 | do { |
| 3389 | __set_current_state(TASK_UNINTERRUPTIBLE); |
| 3390 | spin_unlock_irq(&x->wait.lock); |
| 3391 | timeout = schedule_timeout(timeout); |
| 3392 | spin_lock_irq(&x->wait.lock); |
| 3393 | if (!timeout) { |
| 3394 | __remove_wait_queue(&x->wait, &wait); |
| 3395 | goto out; |
| 3396 | } |
| 3397 | } while (!x->done); |
| 3398 | __remove_wait_queue(&x->wait, &wait); |
| 3399 | } |
| 3400 | x->done--; |
| 3401 | out: |
| 3402 | spin_unlock_irq(&x->wait.lock); |
| 3403 | return timeout; |
| 3404 | } |
| 3405 | EXPORT_SYMBOL(wait_for_completion_timeout); |
| 3406 | |
| 3407 | int fastcall __sched wait_for_completion_interruptible(struct completion *x) |
| 3408 | { |
| 3409 | int ret = 0; |
| 3410 | |
| 3411 | might_sleep(); |
| 3412 | |
| 3413 | spin_lock_irq(&x->wait.lock); |
| 3414 | if (!x->done) { |
| 3415 | DECLARE_WAITQUEUE(wait, current); |
| 3416 | |
| 3417 | wait.flags |= WQ_FLAG_EXCLUSIVE; |
| 3418 | __add_wait_queue_tail(&x->wait, &wait); |
| 3419 | do { |
| 3420 | if (signal_pending(current)) { |
| 3421 | ret = -ERESTARTSYS; |
| 3422 | __remove_wait_queue(&x->wait, &wait); |
| 3423 | goto out; |
| 3424 | } |
| 3425 | __set_current_state(TASK_INTERRUPTIBLE); |
| 3426 | spin_unlock_irq(&x->wait.lock); |
| 3427 | schedule(); |
| 3428 | spin_lock_irq(&x->wait.lock); |
| 3429 | } while (!x->done); |
| 3430 | __remove_wait_queue(&x->wait, &wait); |
| 3431 | } |
| 3432 | x->done--; |
| 3433 | out: |
| 3434 | spin_unlock_irq(&x->wait.lock); |
| 3435 | |
| 3436 | return ret; |
| 3437 | } |
| 3438 | EXPORT_SYMBOL(wait_for_completion_interruptible); |
| 3439 | |
| 3440 | unsigned long fastcall __sched |
| 3441 | wait_for_completion_interruptible_timeout(struct completion *x, |
| 3442 | unsigned long timeout) |
| 3443 | { |
| 3444 | might_sleep(); |
| 3445 | |
| 3446 | spin_lock_irq(&x->wait.lock); |
| 3447 | if (!x->done) { |
| 3448 | DECLARE_WAITQUEUE(wait, current); |
| 3449 | |
| 3450 | wait.flags |= WQ_FLAG_EXCLUSIVE; |
| 3451 | __add_wait_queue_tail(&x->wait, &wait); |
| 3452 | do { |
| 3453 | if (signal_pending(current)) { |
| 3454 | timeout = -ERESTARTSYS; |
| 3455 | __remove_wait_queue(&x->wait, &wait); |
| 3456 | goto out; |
| 3457 | } |
| 3458 | __set_current_state(TASK_INTERRUPTIBLE); |
| 3459 | spin_unlock_irq(&x->wait.lock); |
| 3460 | timeout = schedule_timeout(timeout); |
| 3461 | spin_lock_irq(&x->wait.lock); |
| 3462 | if (!timeout) { |
| 3463 | __remove_wait_queue(&x->wait, &wait); |
| 3464 | goto out; |
| 3465 | } |
| 3466 | } while (!x->done); |
| 3467 | __remove_wait_queue(&x->wait, &wait); |
| 3468 | } |
| 3469 | x->done--; |
| 3470 | out: |
| 3471 | spin_unlock_irq(&x->wait.lock); |
| 3472 | return timeout; |
| 3473 | } |
| 3474 | EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); |
| 3475 | |
| 3476 | |
| 3477 | #define SLEEP_ON_VAR \ |
| 3478 | unsigned long flags; \ |
| 3479 | wait_queue_t wait; \ |
| 3480 | init_waitqueue_entry(&wait, current); |
| 3481 | |
| 3482 | #define SLEEP_ON_HEAD \ |
| 3483 | spin_lock_irqsave(&q->lock,flags); \ |
| 3484 | __add_wait_queue(q, &wait); \ |
| 3485 | spin_unlock(&q->lock); |
| 3486 | |
| 3487 | #define SLEEP_ON_TAIL \ |
| 3488 | spin_lock_irq(&q->lock); \ |
| 3489 | __remove_wait_queue(q, &wait); \ |
| 3490 | spin_unlock_irqrestore(&q->lock, flags); |
| 3491 | |
| 3492 | void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q) |
| 3493 | { |
| 3494 | SLEEP_ON_VAR |
| 3495 | |
| 3496 | current->state = TASK_INTERRUPTIBLE; |
| 3497 | |
| 3498 | SLEEP_ON_HEAD |
| 3499 | schedule(); |
| 3500 | SLEEP_ON_TAIL |
| 3501 | } |
| 3502 | |
| 3503 | EXPORT_SYMBOL(interruptible_sleep_on); |
| 3504 | |
| 3505 | long fastcall __sched |
| 3506 | interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) |
| 3507 | { |
| 3508 | SLEEP_ON_VAR |
| 3509 | |
| 3510 | current->state = TASK_INTERRUPTIBLE; |
| 3511 | |
| 3512 | SLEEP_ON_HEAD |
| 3513 | timeout = schedule_timeout(timeout); |
| 3514 | SLEEP_ON_TAIL |
| 3515 | |
| 3516 | return timeout; |
| 3517 | } |
| 3518 | |
| 3519 | EXPORT_SYMBOL(interruptible_sleep_on_timeout); |
| 3520 | |
| 3521 | void fastcall __sched sleep_on(wait_queue_head_t *q) |
| 3522 | { |
| 3523 | SLEEP_ON_VAR |
| 3524 | |
| 3525 | current->state = TASK_UNINTERRUPTIBLE; |
| 3526 | |
| 3527 | SLEEP_ON_HEAD |
| 3528 | schedule(); |
| 3529 | SLEEP_ON_TAIL |
| 3530 | } |
| 3531 | |
| 3532 | EXPORT_SYMBOL(sleep_on); |
| 3533 | |
| 3534 | long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) |
| 3535 | { |
| 3536 | SLEEP_ON_VAR |
| 3537 | |
| 3538 | current->state = TASK_UNINTERRUPTIBLE; |
| 3539 | |
| 3540 | SLEEP_ON_HEAD |
| 3541 | timeout = schedule_timeout(timeout); |
| 3542 | SLEEP_ON_TAIL |
| 3543 | |
| 3544 | return timeout; |
| 3545 | } |
| 3546 | |
| 3547 | EXPORT_SYMBOL(sleep_on_timeout); |
| 3548 | |
| 3549 | void set_user_nice(task_t *p, long nice) |
| 3550 | { |
| 3551 | unsigned long flags; |
| 3552 | prio_array_t *array; |
| 3553 | runqueue_t *rq; |
| 3554 | int old_prio, new_prio, delta; |
| 3555 | |
| 3556 | if (TASK_NICE(p) == nice || nice < -20 || nice > 19) |
| 3557 | return; |
| 3558 | /* |
| 3559 | * We have to be careful, if called from sys_setpriority(), |
| 3560 | * the task might be in the middle of scheduling on another CPU. |
| 3561 | */ |
| 3562 | rq = task_rq_lock(p, &flags); |
| 3563 | /* |
| 3564 | * The RT priorities are set via sched_setscheduler(), but we still |
| 3565 | * allow the 'normal' nice value to be set - but as expected |
| 3566 | * it wont have any effect on scheduling until the task is |
| 3567 | * not SCHED_NORMAL/SCHED_BATCH: |
| 3568 | */ |
| 3569 | if (rt_task(p)) { |
| 3570 | p->static_prio = NICE_TO_PRIO(nice); |
| 3571 | goto out_unlock; |
| 3572 | } |
| 3573 | array = p->array; |
| 3574 | if (array) { |
| 3575 | dequeue_task(p, array); |
| 3576 | dec_prio_bias(rq, p->static_prio); |
| 3577 | } |
| 3578 | |
| 3579 | old_prio = p->prio; |
| 3580 | new_prio = NICE_TO_PRIO(nice); |
| 3581 | delta = new_prio - old_prio; |
| 3582 | p->static_prio = NICE_TO_PRIO(nice); |
| 3583 | p->prio += delta; |
| 3584 | |
| 3585 | if (array) { |
| 3586 | enqueue_task(p, array); |
| 3587 | inc_prio_bias(rq, p->static_prio); |
| 3588 | /* |
| 3589 | * If the task increased its priority or is running and |
| 3590 | * lowered its priority, then reschedule its CPU: |
| 3591 | */ |
| 3592 | if (delta < 0 || (delta > 0 && task_running(rq, p))) |
| 3593 | resched_task(rq->curr); |
| 3594 | } |
| 3595 | out_unlock: |
| 3596 | task_rq_unlock(rq, &flags); |
| 3597 | } |
| 3598 | |
| 3599 | EXPORT_SYMBOL(set_user_nice); |
| 3600 | |
| 3601 | /* |
| 3602 | * can_nice - check if a task can reduce its nice value |
| 3603 | * @p: task |
| 3604 | * @nice: nice value |
| 3605 | */ |
| 3606 | int can_nice(const task_t *p, const int nice) |
| 3607 | { |
| 3608 | /* convert nice value [19,-20] to rlimit style value [1,40] */ |
| 3609 | int nice_rlim = 20 - nice; |
| 3610 | return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur || |
| 3611 | capable(CAP_SYS_NICE)); |
| 3612 | } |
| 3613 | |
| 3614 | #ifdef __ARCH_WANT_SYS_NICE |
| 3615 | |
| 3616 | /* |
| 3617 | * sys_nice - change the priority of the current process. |
| 3618 | * @increment: priority increment |
| 3619 | * |
| 3620 | * sys_setpriority is a more generic, but much slower function that |
| 3621 | * does similar things. |
| 3622 | */ |
| 3623 | asmlinkage long sys_nice(int increment) |
| 3624 | { |
| 3625 | int retval; |
| 3626 | long nice; |
| 3627 | |
| 3628 | /* |
| 3629 | * Setpriority might change our priority at the same moment. |
| 3630 | * We don't have to worry. Conceptually one call occurs first |
| 3631 | * and we have a single winner. |
| 3632 | */ |
| 3633 | if (increment < -40) |
| 3634 | increment = -40; |
| 3635 | if (increment > 40) |
| 3636 | increment = 40; |
| 3637 | |
| 3638 | nice = PRIO_TO_NICE(current->static_prio) + increment; |
| 3639 | if (nice < -20) |
| 3640 | nice = -20; |
| 3641 | if (nice > 19) |
| 3642 | nice = 19; |
| 3643 | |
| 3644 | if (increment < 0 && !can_nice(current, nice)) |
| 3645 | return -EPERM; |
| 3646 | |
| 3647 | retval = security_task_setnice(current, nice); |
| 3648 | if (retval) |
| 3649 | return retval; |
| 3650 | |
| 3651 | set_user_nice(current, nice); |
| 3652 | return 0; |
| 3653 | } |
| 3654 | |
| 3655 | #endif |
| 3656 | |
| 3657 | /** |
| 3658 | * task_prio - return the priority value of a given task. |
| 3659 | * @p: the task in question. |
| 3660 | * |
| 3661 | * This is the priority value as seen by users in /proc. |
| 3662 | * RT tasks are offset by -200. Normal tasks are centered |
| 3663 | * around 0, value goes from -16 to +15. |
| 3664 | */ |
| 3665 | int task_prio(const task_t *p) |
| 3666 | { |
| 3667 | return p->prio - MAX_RT_PRIO; |
| 3668 | } |
| 3669 | |
| 3670 | /** |
| 3671 | * task_nice - return the nice value of a given task. |
| 3672 | * @p: the task in question. |
| 3673 | */ |
| 3674 | int task_nice(const task_t *p) |
| 3675 | { |
| 3676 | return TASK_NICE(p); |
| 3677 | } |
| 3678 | EXPORT_SYMBOL_GPL(task_nice); |
| 3679 | |
| 3680 | /** |
| 3681 | * idle_cpu - is a given cpu idle currently? |
| 3682 | * @cpu: the processor in question. |
| 3683 | */ |
| 3684 | int idle_cpu(int cpu) |
| 3685 | { |
| 3686 | return cpu_curr(cpu) == cpu_rq(cpu)->idle; |
| 3687 | } |
| 3688 | |
| 3689 | /** |
| 3690 | * idle_task - return the idle task for a given cpu. |
| 3691 | * @cpu: the processor in question. |
| 3692 | */ |
| 3693 | task_t *idle_task(int cpu) |
| 3694 | { |
| 3695 | return cpu_rq(cpu)->idle; |
| 3696 | } |
| 3697 | |
| 3698 | /** |
| 3699 | * find_process_by_pid - find a process with a matching PID value. |
| 3700 | * @pid: the pid in question. |
| 3701 | */ |
| 3702 | static inline task_t *find_process_by_pid(pid_t pid) |
| 3703 | { |
| 3704 | return pid ? find_task_by_pid(pid) : current; |
| 3705 | } |
| 3706 | |
| 3707 | /* Actually do priority change: must hold rq lock. */ |
| 3708 | static void __setscheduler(struct task_struct *p, int policy, int prio) |
| 3709 | { |
| 3710 | BUG_ON(p->array); |
| 3711 | p->policy = policy; |
| 3712 | p->rt_priority = prio; |
| 3713 | if (policy != SCHED_NORMAL && policy != SCHED_BATCH) { |
| 3714 | p->prio = MAX_RT_PRIO-1 - p->rt_priority; |
| 3715 | } else { |
| 3716 | p->prio = p->static_prio; |
| 3717 | /* |
| 3718 | * SCHED_BATCH tasks are treated as perpetual CPU hogs: |
| 3719 | */ |
| 3720 | if (policy == SCHED_BATCH) |
| 3721 | p->sleep_avg = 0; |
| 3722 | } |
| 3723 | } |
| 3724 | |
| 3725 | /** |
| 3726 | * sched_setscheduler - change the scheduling policy and/or RT priority of |
| 3727 | * a thread. |
| 3728 | * @p: the task in question. |
| 3729 | * @policy: new policy. |
| 3730 | * @param: structure containing the new RT priority. |
| 3731 | */ |
| 3732 | int sched_setscheduler(struct task_struct *p, int policy, |
| 3733 | struct sched_param *param) |
| 3734 | { |
| 3735 | int retval; |
| 3736 | int oldprio, oldpolicy = -1; |
| 3737 | prio_array_t *array; |
| 3738 | unsigned long flags; |
| 3739 | runqueue_t *rq; |
| 3740 | |
| 3741 | recheck: |
| 3742 | /* double check policy once rq lock held */ |
| 3743 | if (policy < 0) |
| 3744 | policy = oldpolicy = p->policy; |
| 3745 | else if (policy != SCHED_FIFO && policy != SCHED_RR && |
| 3746 | policy != SCHED_NORMAL && policy != SCHED_BATCH) |
| 3747 | return -EINVAL; |
| 3748 | /* |
| 3749 | * Valid priorities for SCHED_FIFO and SCHED_RR are |
| 3750 | * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and |
| 3751 | * SCHED_BATCH is 0. |
| 3752 | */ |
| 3753 | if (param->sched_priority < 0 || |
| 3754 | (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || |
| 3755 | (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) |
| 3756 | return -EINVAL; |
| 3757 | if ((policy == SCHED_NORMAL || policy == SCHED_BATCH) |
| 3758 | != (param->sched_priority == 0)) |
| 3759 | return -EINVAL; |
| 3760 | |
| 3761 | /* |
| 3762 | * Allow unprivileged RT tasks to decrease priority: |
| 3763 | */ |
| 3764 | if (!capable(CAP_SYS_NICE)) { |
| 3765 | /* |
| 3766 | * can't change policy, except between SCHED_NORMAL |
| 3767 | * and SCHED_BATCH: |
| 3768 | */ |
| 3769 | if (((policy != SCHED_NORMAL && p->policy != SCHED_BATCH) && |
| 3770 | (policy != SCHED_BATCH && p->policy != SCHED_NORMAL)) && |
| 3771 | !p->signal->rlim[RLIMIT_RTPRIO].rlim_cur) |
| 3772 | return -EPERM; |
| 3773 | /* can't increase priority */ |
| 3774 | if ((policy != SCHED_NORMAL && policy != SCHED_BATCH) && |
| 3775 | param->sched_priority > p->rt_priority && |
| 3776 | param->sched_priority > |
| 3777 | p->signal->rlim[RLIMIT_RTPRIO].rlim_cur) |
| 3778 | return -EPERM; |
| 3779 | /* can't change other user's priorities */ |
| 3780 | if ((current->euid != p->euid) && |
| 3781 | (current->euid != p->uid)) |
| 3782 | return -EPERM; |
| 3783 | } |
| 3784 | |
| 3785 | retval = security_task_setscheduler(p, policy, param); |
| 3786 | if (retval) |
| 3787 | return retval; |
| 3788 | /* |
| 3789 | * To be able to change p->policy safely, the apropriate |
| 3790 | * runqueue lock must be held. |
| 3791 | */ |
| 3792 | rq = task_rq_lock(p, &flags); |
| 3793 | /* recheck policy now with rq lock held */ |
| 3794 | if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { |
| 3795 | policy = oldpolicy = -1; |
| 3796 | task_rq_unlock(rq, &flags); |
| 3797 | goto recheck; |
| 3798 | } |
| 3799 | array = p->array; |
| 3800 | if (array) |
| 3801 | deactivate_task(p, rq); |
| 3802 | oldprio = p->prio; |
| 3803 | __setscheduler(p, policy, param->sched_priority); |
| 3804 | if (array) { |
| 3805 | __activate_task(p, rq); |
| 3806 | /* |
| 3807 | * Reschedule if we are currently running on this runqueue and |
| 3808 | * our priority decreased, or if we are not currently running on |
| 3809 | * this runqueue and our priority is higher than the current's |
| 3810 | */ |
| 3811 | if (task_running(rq, p)) { |
| 3812 | if (p->prio > oldprio) |
| 3813 | resched_task(rq->curr); |
| 3814 | } else if (TASK_PREEMPTS_CURR(p, rq)) |
| 3815 | resched_task(rq->curr); |
| 3816 | } |
| 3817 | task_rq_unlock(rq, &flags); |
| 3818 | return 0; |
| 3819 | } |
| 3820 | EXPORT_SYMBOL_GPL(sched_setscheduler); |
| 3821 | |
| 3822 | static int |
| 3823 | do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) |
| 3824 | { |
| 3825 | int retval; |
| 3826 | struct sched_param lparam; |
| 3827 | struct task_struct *p; |
| 3828 | |
| 3829 | if (!param || pid < 0) |
| 3830 | return -EINVAL; |
| 3831 | if (copy_from_user(&lparam, param, sizeof(struct sched_param))) |
| 3832 | return -EFAULT; |
| 3833 | read_lock_irq(&tasklist_lock); |
| 3834 | p = find_process_by_pid(pid); |
| 3835 | if (!p) { |
| 3836 | read_unlock_irq(&tasklist_lock); |
| 3837 | return -ESRCH; |
| 3838 | } |
| 3839 | retval = sched_setscheduler(p, policy, &lparam); |
| 3840 | read_unlock_irq(&tasklist_lock); |
| 3841 | return retval; |
| 3842 | } |
| 3843 | |
| 3844 | /** |
| 3845 | * sys_sched_setscheduler - set/change the scheduler policy and RT priority |
| 3846 | * @pid: the pid in question. |
| 3847 | * @policy: new policy. |
| 3848 | * @param: structure containing the new RT priority. |
| 3849 | */ |
| 3850 | asmlinkage long sys_sched_setscheduler(pid_t pid, int policy, |
| 3851 | struct sched_param __user *param) |
| 3852 | { |
| 3853 | return do_sched_setscheduler(pid, policy, param); |
| 3854 | } |
| 3855 | |
| 3856 | /** |
| 3857 | * sys_sched_setparam - set/change the RT priority of a thread |
| 3858 | * @pid: the pid in question. |
| 3859 | * @param: structure containing the new RT priority. |
| 3860 | */ |
| 3861 | asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param) |
| 3862 | { |
| 3863 | return do_sched_setscheduler(pid, -1, param); |
| 3864 | } |
| 3865 | |
| 3866 | /** |
| 3867 | * sys_sched_getscheduler - get the policy (scheduling class) of a thread |
| 3868 | * @pid: the pid in question. |
| 3869 | */ |
| 3870 | asmlinkage long sys_sched_getscheduler(pid_t pid) |
| 3871 | { |
| 3872 | int retval = -EINVAL; |
| 3873 | task_t *p; |
| 3874 | |
| 3875 | if (pid < 0) |
| 3876 | goto out_nounlock; |
| 3877 | |
| 3878 | retval = -ESRCH; |
| 3879 | read_lock(&tasklist_lock); |
| 3880 | p = find_process_by_pid(pid); |
| 3881 | if (p) { |
| 3882 | retval = security_task_getscheduler(p); |
| 3883 | if (!retval) |
| 3884 | retval = p->policy; |
| 3885 | } |
| 3886 | read_unlock(&tasklist_lock); |
| 3887 | |
| 3888 | out_nounlock: |
| 3889 | return retval; |
| 3890 | } |
| 3891 | |
| 3892 | /** |
| 3893 | * sys_sched_getscheduler - get the RT priority of a thread |
| 3894 | * @pid: the pid in question. |
| 3895 | * @param: structure containing the RT priority. |
| 3896 | */ |
| 3897 | asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param) |
| 3898 | { |
| 3899 | struct sched_param lp; |
| 3900 | int retval = -EINVAL; |
| 3901 | task_t *p; |
| 3902 | |
| 3903 | if (!param || pid < 0) |
| 3904 | goto out_nounlock; |
| 3905 | |
| 3906 | read_lock(&tasklist_lock); |
| 3907 | p = find_process_by_pid(pid); |
| 3908 | retval = -ESRCH; |
| 3909 | if (!p) |
| 3910 | goto out_unlock; |
| 3911 | |
| 3912 | retval = security_task_getscheduler(p); |
| 3913 | if (retval) |
| 3914 | goto out_unlock; |
| 3915 | |
| 3916 | lp.sched_priority = p->rt_priority; |
| 3917 | read_unlock(&tasklist_lock); |
| 3918 | |
| 3919 | /* |
| 3920 | * This one might sleep, we cannot do it with a spinlock held ... |
| 3921 | */ |
| 3922 | retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; |
| 3923 | |
| 3924 | out_nounlock: |
| 3925 | return retval; |
| 3926 | |
| 3927 | out_unlock: |
| 3928 | read_unlock(&tasklist_lock); |
| 3929 | return retval; |
| 3930 | } |
| 3931 | |
| 3932 | long sched_setaffinity(pid_t pid, cpumask_t new_mask) |
| 3933 | { |
| 3934 | task_t *p; |
| 3935 | int retval; |
| 3936 | cpumask_t cpus_allowed; |
| 3937 | |
| 3938 | lock_cpu_hotplug(); |
| 3939 | read_lock(&tasklist_lock); |
| 3940 | |
| 3941 | p = find_process_by_pid(pid); |
| 3942 | if (!p) { |
| 3943 | read_unlock(&tasklist_lock); |
| 3944 | unlock_cpu_hotplug(); |
| 3945 | return -ESRCH; |
| 3946 | } |
| 3947 | |
| 3948 | /* |
| 3949 | * It is not safe to call set_cpus_allowed with the |
| 3950 | * tasklist_lock held. We will bump the task_struct's |
| 3951 | * usage count and then drop tasklist_lock. |
| 3952 | */ |
| 3953 | get_task_struct(p); |
| 3954 | read_unlock(&tasklist_lock); |
| 3955 | |
| 3956 | retval = -EPERM; |
| 3957 | if ((current->euid != p->euid) && (current->euid != p->uid) && |
| 3958 | !capable(CAP_SYS_NICE)) |
| 3959 | goto out_unlock; |
| 3960 | |
| 3961 | cpus_allowed = cpuset_cpus_allowed(p); |
| 3962 | cpus_and(new_mask, new_mask, cpus_allowed); |
| 3963 | retval = set_cpus_allowed(p, new_mask); |
| 3964 | |
| 3965 | out_unlock: |
| 3966 | put_task_struct(p); |
| 3967 | unlock_cpu_hotplug(); |
| 3968 | return retval; |
| 3969 | } |
| 3970 | |
| 3971 | static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, |
| 3972 | cpumask_t *new_mask) |
| 3973 | { |
| 3974 | if (len < sizeof(cpumask_t)) { |
| 3975 | memset(new_mask, 0, sizeof(cpumask_t)); |
| 3976 | } else if (len > sizeof(cpumask_t)) { |
| 3977 | len = sizeof(cpumask_t); |
| 3978 | } |
| 3979 | return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; |
| 3980 | } |
| 3981 | |
| 3982 | /** |
| 3983 | * sys_sched_setaffinity - set the cpu affinity of a process |
| 3984 | * @pid: pid of the process |
| 3985 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr |
| 3986 | * @user_mask_ptr: user-space pointer to the new cpu mask |
| 3987 | */ |
| 3988 | asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len, |
| 3989 | unsigned long __user *user_mask_ptr) |
| 3990 | { |
| 3991 | cpumask_t new_mask; |
| 3992 | int retval; |
| 3993 | |
| 3994 | retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask); |
| 3995 | if (retval) |
| 3996 | return retval; |
| 3997 | |
| 3998 | return sched_setaffinity(pid, new_mask); |
| 3999 | } |
| 4000 | |
| 4001 | /* |
| 4002 | * Represents all cpu's present in the system |
| 4003 | * In systems capable of hotplug, this map could dynamically grow |
| 4004 | * as new cpu's are detected in the system via any platform specific |
| 4005 | * method, such as ACPI for e.g. |
| 4006 | */ |
| 4007 | |
| 4008 | cpumask_t cpu_present_map __read_mostly; |
| 4009 | EXPORT_SYMBOL(cpu_present_map); |
| 4010 | |
| 4011 | #ifndef CONFIG_SMP |
| 4012 | cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL; |
| 4013 | cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL; |
| 4014 | #endif |
| 4015 | |
| 4016 | long sched_getaffinity(pid_t pid, cpumask_t *mask) |
| 4017 | { |
| 4018 | int retval; |
| 4019 | task_t *p; |
| 4020 | |
| 4021 | lock_cpu_hotplug(); |
| 4022 | read_lock(&tasklist_lock); |
| 4023 | |
| 4024 | retval = -ESRCH; |
| 4025 | p = find_process_by_pid(pid); |
| 4026 | if (!p) |
| 4027 | goto out_unlock; |
| 4028 | |
| 4029 | retval = 0; |
| 4030 | cpus_and(*mask, p->cpus_allowed, cpu_possible_map); |
| 4031 | |
| 4032 | out_unlock: |
| 4033 | read_unlock(&tasklist_lock); |
| 4034 | unlock_cpu_hotplug(); |
| 4035 | if (retval) |
| 4036 | return retval; |
| 4037 | |
| 4038 | return 0; |
| 4039 | } |
| 4040 | |
| 4041 | /** |
| 4042 | * sys_sched_getaffinity - get the cpu affinity of a process |
| 4043 | * @pid: pid of the process |
| 4044 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr |
| 4045 | * @user_mask_ptr: user-space pointer to hold the current cpu mask |
| 4046 | */ |
| 4047 | asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len, |
| 4048 | unsigned long __user *user_mask_ptr) |
| 4049 | { |
| 4050 | int ret; |
| 4051 | cpumask_t mask; |
| 4052 | |
| 4053 | if (len < sizeof(cpumask_t)) |
| 4054 | return -EINVAL; |
| 4055 | |
| 4056 | ret = sched_getaffinity(pid, &mask); |
| 4057 | if (ret < 0) |
| 4058 | return ret; |
| 4059 | |
| 4060 | if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t))) |
| 4061 | return -EFAULT; |
| 4062 | |
| 4063 | return sizeof(cpumask_t); |
| 4064 | } |
| 4065 | |
| 4066 | /** |
| 4067 | * sys_sched_yield - yield the current processor to other threads. |
| 4068 | * |
| 4069 | * this function yields the current CPU by moving the calling thread |
| 4070 | * to the expired array. If there are no other threads running on this |
| 4071 | * CPU then this function will return. |
| 4072 | */ |
| 4073 | asmlinkage long sys_sched_yield(void) |
| 4074 | { |
| 4075 | runqueue_t *rq = this_rq_lock(); |
| 4076 | prio_array_t *array = current->array; |
| 4077 | prio_array_t *target = rq->expired; |
| 4078 | |
| 4079 | schedstat_inc(rq, yld_cnt); |
| 4080 | /* |
| 4081 | * We implement yielding by moving the task into the expired |
| 4082 | * queue. |
| 4083 | * |
| 4084 | * (special rule: RT tasks will just roundrobin in the active |
| 4085 | * array.) |
| 4086 | */ |
| 4087 | if (rt_task(current)) |
| 4088 | target = rq->active; |
| 4089 | |
| 4090 | if (array->nr_active == 1) { |
| 4091 | schedstat_inc(rq, yld_act_empty); |
| 4092 | if (!rq->expired->nr_active) |
| 4093 | schedstat_inc(rq, yld_both_empty); |
| 4094 | } else if (!rq->expired->nr_active) |
| 4095 | schedstat_inc(rq, yld_exp_empty); |
| 4096 | |
| 4097 | if (array != target) { |
| 4098 | dequeue_task(current, array); |
| 4099 | enqueue_task(current, target); |
| 4100 | } else |
| 4101 | /* |
| 4102 | * requeue_task is cheaper so perform that if possible. |
| 4103 | */ |
| 4104 | requeue_task(current, array); |
| 4105 | |
| 4106 | /* |
| 4107 | * Since we are going to call schedule() anyway, there's |
| 4108 | * no need to preempt or enable interrupts: |
| 4109 | */ |
| 4110 | __release(rq->lock); |
| 4111 | _raw_spin_unlock(&rq->lock); |
| 4112 | preempt_enable_no_resched(); |
| 4113 | |
| 4114 | schedule(); |
| 4115 | |
| 4116 | return 0; |
| 4117 | } |
| 4118 | |
| 4119 | static inline void __cond_resched(void) |
| 4120 | { |
| 4121 | /* |
| 4122 | * The BKS might be reacquired before we have dropped |
| 4123 | * PREEMPT_ACTIVE, which could trigger a second |
| 4124 | * cond_resched() call. |
| 4125 | */ |
| 4126 | if (unlikely(preempt_count())) |
| 4127 | return; |
| 4128 | do { |
| 4129 | add_preempt_count(PREEMPT_ACTIVE); |
| 4130 | schedule(); |
| 4131 | sub_preempt_count(PREEMPT_ACTIVE); |
| 4132 | } while (need_resched()); |
| 4133 | } |
| 4134 | |
| 4135 | int __sched cond_resched(void) |
| 4136 | { |
| 4137 | if (need_resched()) { |
| 4138 | __cond_resched(); |
| 4139 | return 1; |
| 4140 | } |
| 4141 | return 0; |
| 4142 | } |
| 4143 | |
| 4144 | EXPORT_SYMBOL(cond_resched); |
| 4145 | |
| 4146 | /* |
| 4147 | * cond_resched_lock() - if a reschedule is pending, drop the given lock, |
| 4148 | * call schedule, and on return reacquire the lock. |
| 4149 | * |
| 4150 | * This works OK both with and without CONFIG_PREEMPT. We do strange low-level |
| 4151 | * operations here to prevent schedule() from being called twice (once via |
| 4152 | * spin_unlock(), once by hand). |
| 4153 | */ |
| 4154 | int cond_resched_lock(spinlock_t *lock) |
| 4155 | { |
| 4156 | int ret = 0; |
| 4157 | |
| 4158 | if (need_lockbreak(lock)) { |
| 4159 | spin_unlock(lock); |
| 4160 | cpu_relax(); |
| 4161 | ret = 1; |
| 4162 | spin_lock(lock); |
| 4163 | } |
| 4164 | if (need_resched()) { |
| 4165 | _raw_spin_unlock(lock); |
| 4166 | preempt_enable_no_resched(); |
| 4167 | __cond_resched(); |
| 4168 | ret = 1; |
| 4169 | spin_lock(lock); |
| 4170 | } |
| 4171 | return ret; |
| 4172 | } |
| 4173 | |
| 4174 | EXPORT_SYMBOL(cond_resched_lock); |
| 4175 | |
| 4176 | int __sched cond_resched_softirq(void) |
| 4177 | { |
| 4178 | BUG_ON(!in_softirq()); |
| 4179 | |
| 4180 | if (need_resched()) { |
| 4181 | __local_bh_enable(); |
| 4182 | __cond_resched(); |
| 4183 | local_bh_disable(); |
| 4184 | return 1; |
| 4185 | } |
| 4186 | return 0; |
| 4187 | } |
| 4188 | |
| 4189 | EXPORT_SYMBOL(cond_resched_softirq); |
| 4190 | |
| 4191 | |
| 4192 | /** |
| 4193 | * yield - yield the current processor to other threads. |
| 4194 | * |
| 4195 | * this is a shortcut for kernel-space yielding - it marks the |
| 4196 | * thread runnable and calls sys_sched_yield(). |
| 4197 | */ |
| 4198 | void __sched yield(void) |
| 4199 | { |
| 4200 | set_current_state(TASK_RUNNING); |
| 4201 | sys_sched_yield(); |
| 4202 | } |
| 4203 | |
| 4204 | EXPORT_SYMBOL(yield); |
| 4205 | |
| 4206 | /* |
| 4207 | * This task is about to go to sleep on IO. Increment rq->nr_iowait so |
| 4208 | * that process accounting knows that this is a task in IO wait state. |
| 4209 | * |
| 4210 | * But don't do that if it is a deliberate, throttling IO wait (this task |
| 4211 | * has set its backing_dev_info: the queue against which it should throttle) |
| 4212 | */ |
| 4213 | void __sched io_schedule(void) |
| 4214 | { |
| 4215 | struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id()); |
| 4216 | |
| 4217 | atomic_inc(&rq->nr_iowait); |
| 4218 | schedule(); |
| 4219 | atomic_dec(&rq->nr_iowait); |
| 4220 | } |
| 4221 | |
| 4222 | EXPORT_SYMBOL(io_schedule); |
| 4223 | |
| 4224 | long __sched io_schedule_timeout(long timeout) |
| 4225 | { |
| 4226 | struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id()); |
| 4227 | long ret; |
| 4228 | |
| 4229 | atomic_inc(&rq->nr_iowait); |
| 4230 | ret = schedule_timeout(timeout); |
| 4231 | atomic_dec(&rq->nr_iowait); |
| 4232 | return ret; |
| 4233 | } |
| 4234 | |
| 4235 | /** |
| 4236 | * sys_sched_get_priority_max - return maximum RT priority. |
| 4237 | * @policy: scheduling class. |
| 4238 | * |
| 4239 | * this syscall returns the maximum rt_priority that can be used |
| 4240 | * by a given scheduling class. |
| 4241 | */ |
| 4242 | asmlinkage long sys_sched_get_priority_max(int policy) |
| 4243 | { |
| 4244 | int ret = -EINVAL; |
| 4245 | |
| 4246 | switch (policy) { |
| 4247 | case SCHED_FIFO: |
| 4248 | case SCHED_RR: |
| 4249 | ret = MAX_USER_RT_PRIO-1; |
| 4250 | break; |
| 4251 | case SCHED_NORMAL: |
| 4252 | case SCHED_BATCH: |
| 4253 | ret = 0; |
| 4254 | break; |
| 4255 | } |
| 4256 | return ret; |
| 4257 | } |
| 4258 | |
| 4259 | /** |
| 4260 | * sys_sched_get_priority_min - return minimum RT priority. |
| 4261 | * @policy: scheduling class. |
| 4262 | * |
| 4263 | * this syscall returns the minimum rt_priority that can be used |
| 4264 | * by a given scheduling class. |
| 4265 | */ |
| 4266 | asmlinkage long sys_sched_get_priority_min(int policy) |
| 4267 | { |
| 4268 | int ret = -EINVAL; |
| 4269 | |
| 4270 | switch (policy) { |
| 4271 | case SCHED_FIFO: |
| 4272 | case SCHED_RR: |
| 4273 | ret = 1; |
| 4274 | break; |
| 4275 | case SCHED_NORMAL: |
| 4276 | case SCHED_BATCH: |
| 4277 | ret = 0; |
| 4278 | } |
| 4279 | return ret; |
| 4280 | } |
| 4281 | |
| 4282 | /** |
| 4283 | * sys_sched_rr_get_interval - return the default timeslice of a process. |
| 4284 | * @pid: pid of the process. |
| 4285 | * @interval: userspace pointer to the timeslice value. |
| 4286 | * |
| 4287 | * this syscall writes the default timeslice value of a given process |
| 4288 | * into the user-space timespec buffer. A value of '0' means infinity. |
| 4289 | */ |
| 4290 | asmlinkage |
| 4291 | long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval) |
| 4292 | { |
| 4293 | int retval = -EINVAL; |
| 4294 | struct timespec t; |
| 4295 | task_t *p; |
| 4296 | |
| 4297 | if (pid < 0) |
| 4298 | goto out_nounlock; |
| 4299 | |
| 4300 | retval = -ESRCH; |
| 4301 | read_lock(&tasklist_lock); |
| 4302 | p = find_process_by_pid(pid); |
| 4303 | if (!p) |
| 4304 | goto out_unlock; |
| 4305 | |
| 4306 | retval = security_task_getscheduler(p); |
| 4307 | if (retval) |
| 4308 | goto out_unlock; |
| 4309 | |
| 4310 | jiffies_to_timespec(p->policy & SCHED_FIFO ? |
| 4311 | 0 : task_timeslice(p), &t); |
| 4312 | read_unlock(&tasklist_lock); |
| 4313 | retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; |
| 4314 | out_nounlock: |
| 4315 | return retval; |
| 4316 | out_unlock: |
| 4317 | read_unlock(&tasklist_lock); |
| 4318 | return retval; |
| 4319 | } |
| 4320 | |
| 4321 | static inline struct task_struct *eldest_child(struct task_struct *p) |
| 4322 | { |
| 4323 | if (list_empty(&p->children)) return NULL; |
| 4324 | return list_entry(p->children.next,struct task_struct,sibling); |
| 4325 | } |
| 4326 | |
| 4327 | static inline struct task_struct *older_sibling(struct task_struct *p) |
| 4328 | { |
| 4329 | if (p->sibling.prev==&p->parent->children) return NULL; |
| 4330 | return list_entry(p->sibling.prev,struct task_struct,sibling); |
| 4331 | } |
| 4332 | |
| 4333 | static inline struct task_struct *younger_sibling(struct task_struct *p) |
| 4334 | { |
| 4335 | if (p->sibling.next==&p->parent->children) return NULL; |
| 4336 | return list_entry(p->sibling.next,struct task_struct,sibling); |
| 4337 | } |
| 4338 | |
| 4339 | static void show_task(task_t *p) |
| 4340 | { |
| 4341 | task_t *relative; |
| 4342 | unsigned state; |
| 4343 | unsigned long free = 0; |
| 4344 | static const char *stat_nam[] = { "R", "S", "D", "T", "t", "Z", "X" }; |
| 4345 | |
| 4346 | printk("%-13.13s ", p->comm); |
| 4347 | state = p->state ? __ffs(p->state) + 1 : 0; |
| 4348 | if (state < ARRAY_SIZE(stat_nam)) |
| 4349 | printk(stat_nam[state]); |
| 4350 | else |
| 4351 | printk("?"); |
| 4352 | #if (BITS_PER_LONG == 32) |
| 4353 | if (state == TASK_RUNNING) |
| 4354 | printk(" running "); |
| 4355 | else |
| 4356 | printk(" %08lX ", thread_saved_pc(p)); |
| 4357 | #else |
| 4358 | if (state == TASK_RUNNING) |
| 4359 | printk(" running task "); |
| 4360 | else |
| 4361 | printk(" %016lx ", thread_saved_pc(p)); |
| 4362 | #endif |
| 4363 | #ifdef CONFIG_DEBUG_STACK_USAGE |
| 4364 | { |
| 4365 | unsigned long *n = end_of_stack(p); |
| 4366 | while (!*n) |
| 4367 | n++; |
| 4368 | free = (unsigned long)n - (unsigned long)end_of_stack(p); |
| 4369 | } |
| 4370 | #endif |
| 4371 | printk("%5lu %5d %6d ", free, p->pid, p->parent->pid); |
| 4372 | if ((relative = eldest_child(p))) |
| 4373 | printk("%5d ", relative->pid); |
| 4374 | else |
| 4375 | printk(" "); |
| 4376 | if ((relative = younger_sibling(p))) |
| 4377 | printk("%7d", relative->pid); |
| 4378 | else |
| 4379 | printk(" "); |
| 4380 | if ((relative = older_sibling(p))) |
| 4381 | printk(" %5d", relative->pid); |
| 4382 | else |
| 4383 | printk(" "); |
| 4384 | if (!p->mm) |
| 4385 | printk(" (L-TLB)\n"); |
| 4386 | else |
| 4387 | printk(" (NOTLB)\n"); |
| 4388 | |
| 4389 | if (state != TASK_RUNNING) |
| 4390 | show_stack(p, NULL); |
| 4391 | } |
| 4392 | |
| 4393 | void show_state(void) |
| 4394 | { |
| 4395 | task_t *g, *p; |
| 4396 | |
| 4397 | #if (BITS_PER_LONG == 32) |
| 4398 | printk("\n" |
| 4399 | " sibling\n"); |
| 4400 | printk(" task PC pid father child younger older\n"); |
| 4401 | #else |
| 4402 | printk("\n" |
| 4403 | " sibling\n"); |
| 4404 | printk(" task PC pid father child younger older\n"); |
| 4405 | #endif |
| 4406 | read_lock(&tasklist_lock); |
| 4407 | do_each_thread(g, p) { |
| 4408 | /* |
| 4409 | * reset the NMI-timeout, listing all files on a slow |
| 4410 | * console might take alot of time: |
| 4411 | */ |
| 4412 | touch_nmi_watchdog(); |
| 4413 | show_task(p); |
| 4414 | } while_each_thread(g, p); |
| 4415 | |
| 4416 | read_unlock(&tasklist_lock); |
| 4417 | mutex_debug_show_all_locks(); |
| 4418 | } |
| 4419 | |
| 4420 | /** |
| 4421 | * init_idle - set up an idle thread for a given CPU |
| 4422 | * @idle: task in question |
| 4423 | * @cpu: cpu the idle task belongs to |
| 4424 | * |
| 4425 | * NOTE: this function does not set the idle thread's NEED_RESCHED |
| 4426 | * flag, to make booting more robust. |
| 4427 | */ |
| 4428 | void __devinit init_idle(task_t *idle, int cpu) |
| 4429 | { |
| 4430 | runqueue_t *rq = cpu_rq(cpu); |
| 4431 | unsigned long flags; |
| 4432 | |
| 4433 | idle->sleep_avg = 0; |
| 4434 | idle->array = NULL; |
| 4435 | idle->prio = MAX_PRIO; |
| 4436 | idle->state = TASK_RUNNING; |
| 4437 | idle->cpus_allowed = cpumask_of_cpu(cpu); |
| 4438 | set_task_cpu(idle, cpu); |
| 4439 | |
| 4440 | spin_lock_irqsave(&rq->lock, flags); |
| 4441 | rq->curr = rq->idle = idle; |
| 4442 | #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) |
| 4443 | idle->oncpu = 1; |
| 4444 | #endif |
| 4445 | spin_unlock_irqrestore(&rq->lock, flags); |
| 4446 | |
| 4447 | /* Set the preempt count _outside_ the spinlocks! */ |
| 4448 | #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL) |
| 4449 | task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0); |
| 4450 | #else |
| 4451 | task_thread_info(idle)->preempt_count = 0; |
| 4452 | #endif |
| 4453 | } |
| 4454 | |
| 4455 | /* |
| 4456 | * In a system that switches off the HZ timer nohz_cpu_mask |
| 4457 | * indicates which cpus entered this state. This is used |
| 4458 | * in the rcu update to wait only for active cpus. For system |
| 4459 | * which do not switch off the HZ timer nohz_cpu_mask should |
| 4460 | * always be CPU_MASK_NONE. |
| 4461 | */ |
| 4462 | cpumask_t nohz_cpu_mask = CPU_MASK_NONE; |
| 4463 | |
| 4464 | #ifdef CONFIG_SMP |
| 4465 | /* |
| 4466 | * This is how migration works: |
| 4467 | * |
| 4468 | * 1) we queue a migration_req_t structure in the source CPU's |
| 4469 | * runqueue and wake up that CPU's migration thread. |
| 4470 | * 2) we down() the locked semaphore => thread blocks. |
| 4471 | * 3) migration thread wakes up (implicitly it forces the migrated |
| 4472 | * thread off the CPU) |
| 4473 | * 4) it gets the migration request and checks whether the migrated |
| 4474 | * task is still in the wrong runqueue. |
| 4475 | * 5) if it's in the wrong runqueue then the migration thread removes |
| 4476 | * it and puts it into the right queue. |
| 4477 | * 6) migration thread up()s the semaphore. |
| 4478 | * 7) we wake up and the migration is done. |
| 4479 | */ |
| 4480 | |
| 4481 | /* |
| 4482 | * Change a given task's CPU affinity. Migrate the thread to a |
| 4483 | * proper CPU and schedule it away if the CPU it's executing on |
| 4484 | * is removed from the allowed bitmask. |
| 4485 | * |
| 4486 | * NOTE: the caller must have a valid reference to the task, the |
| 4487 | * task must not exit() & deallocate itself prematurely. The |
| 4488 | * call is not atomic; no spinlocks may be held. |
| 4489 | */ |
| 4490 | int set_cpus_allowed(task_t *p, cpumask_t new_mask) |
| 4491 | { |
| 4492 | unsigned long flags; |
| 4493 | int ret = 0; |
| 4494 | migration_req_t req; |
| 4495 | runqueue_t *rq; |
| 4496 | |
| 4497 | rq = task_rq_lock(p, &flags); |
| 4498 | if (!cpus_intersects(new_mask, cpu_online_map)) { |
| 4499 | ret = -EINVAL; |
| 4500 | goto out; |
| 4501 | } |
| 4502 | |
| 4503 | p->cpus_allowed = new_mask; |
| 4504 | /* Can the task run on the task's current CPU? If so, we're done */ |
| 4505 | if (cpu_isset(task_cpu(p), new_mask)) |
| 4506 | goto out; |
| 4507 | |
| 4508 | if (migrate_task(p, any_online_cpu(new_mask), &req)) { |
| 4509 | /* Need help from migration thread: drop lock and wait. */ |
| 4510 | task_rq_unlock(rq, &flags); |
| 4511 | wake_up_process(rq->migration_thread); |
| 4512 | wait_for_completion(&req.done); |
| 4513 | tlb_migrate_finish(p->mm); |
| 4514 | return 0; |
| 4515 | } |
| 4516 | out: |
| 4517 | task_rq_unlock(rq, &flags); |
| 4518 | return ret; |
| 4519 | } |
| 4520 | |
| 4521 | EXPORT_SYMBOL_GPL(set_cpus_allowed); |
| 4522 | |
| 4523 | /* |
| 4524 | * Move (not current) task off this cpu, onto dest cpu. We're doing |
| 4525 | * this because either it can't run here any more (set_cpus_allowed() |
| 4526 | * away from this CPU, or CPU going down), or because we're |
| 4527 | * attempting to rebalance this task on exec (sched_exec). |
| 4528 | * |
| 4529 | * So we race with normal scheduler movements, but that's OK, as long |
| 4530 | * as the task is no longer on this CPU. |
| 4531 | */ |
| 4532 | static void __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) |
| 4533 | { |
| 4534 | runqueue_t *rq_dest, *rq_src; |
| 4535 | |
| 4536 | if (unlikely(cpu_is_offline(dest_cpu))) |
| 4537 | return; |
| 4538 | |
| 4539 | rq_src = cpu_rq(src_cpu); |
| 4540 | rq_dest = cpu_rq(dest_cpu); |
| 4541 | |
| 4542 | double_rq_lock(rq_src, rq_dest); |
| 4543 | /* Already moved. */ |
| 4544 | if (task_cpu(p) != src_cpu) |
| 4545 | goto out; |
| 4546 | /* Affinity changed (again). */ |
| 4547 | if (!cpu_isset(dest_cpu, p->cpus_allowed)) |
| 4548 | goto out; |
| 4549 | |
| 4550 | set_task_cpu(p, dest_cpu); |
| 4551 | if (p->array) { |
| 4552 | /* |
| 4553 | * Sync timestamp with rq_dest's before activating. |
| 4554 | * The same thing could be achieved by doing this step |
| 4555 | * afterwards, and pretending it was a local activate. |
| 4556 | * This way is cleaner and logically correct. |
| 4557 | */ |
| 4558 | p->timestamp = p->timestamp - rq_src->timestamp_last_tick |
| 4559 | + rq_dest->timestamp_last_tick; |
| 4560 | deactivate_task(p, rq_src); |
| 4561 | activate_task(p, rq_dest, 0); |
| 4562 | if (TASK_PREEMPTS_CURR(p, rq_dest)) |
| 4563 | resched_task(rq_dest->curr); |
| 4564 | } |
| 4565 | |
| 4566 | out: |
| 4567 | double_rq_unlock(rq_src, rq_dest); |
| 4568 | } |
| 4569 | |
| 4570 | /* |
| 4571 | * migration_thread - this is a highprio system thread that performs |
| 4572 | * thread migration by bumping thread off CPU then 'pushing' onto |
| 4573 | * another runqueue. |
| 4574 | */ |
| 4575 | static int migration_thread(void *data) |
| 4576 | { |
| 4577 | runqueue_t *rq; |
| 4578 | int cpu = (long)data; |
| 4579 | |
| 4580 | rq = cpu_rq(cpu); |
| 4581 | BUG_ON(rq->migration_thread != current); |
| 4582 | |
| 4583 | set_current_state(TASK_INTERRUPTIBLE); |
| 4584 | while (!kthread_should_stop()) { |
| 4585 | struct list_head *head; |
| 4586 | migration_req_t *req; |
| 4587 | |
| 4588 | try_to_freeze(); |
| 4589 | |
| 4590 | spin_lock_irq(&rq->lock); |
| 4591 | |
| 4592 | if (cpu_is_offline(cpu)) { |
| 4593 | spin_unlock_irq(&rq->lock); |
| 4594 | goto wait_to_die; |
| 4595 | } |
| 4596 | |
| 4597 | if (rq->active_balance) { |
| 4598 | active_load_balance(rq, cpu); |
| 4599 | rq->active_balance = 0; |
| 4600 | } |
| 4601 | |
| 4602 | head = &rq->migration_queue; |
| 4603 | |
| 4604 | if (list_empty(head)) { |
| 4605 | spin_unlock_irq(&rq->lock); |
| 4606 | schedule(); |
| 4607 | set_current_state(TASK_INTERRUPTIBLE); |
| 4608 | continue; |
| 4609 | } |
| 4610 | req = list_entry(head->next, migration_req_t, list); |
| 4611 | list_del_init(head->next); |
| 4612 | |
| 4613 | spin_unlock(&rq->lock); |
| 4614 | __migrate_task(req->task, cpu, req->dest_cpu); |
| 4615 | local_irq_enable(); |
| 4616 | |
| 4617 | complete(&req->done); |
| 4618 | } |
| 4619 | __set_current_state(TASK_RUNNING); |
| 4620 | return 0; |
| 4621 | |
| 4622 | wait_to_die: |
| 4623 | /* Wait for kthread_stop */ |
| 4624 | set_current_state(TASK_INTERRUPTIBLE); |
| 4625 | while (!kthread_should_stop()) { |
| 4626 | schedule(); |
| 4627 | set_current_state(TASK_INTERRUPTIBLE); |
| 4628 | } |
| 4629 | __set_current_state(TASK_RUNNING); |
| 4630 | return 0; |
| 4631 | } |
| 4632 | |
| 4633 | #ifdef CONFIG_HOTPLUG_CPU |
| 4634 | /* Figure out where task on dead CPU should go, use force if neccessary. */ |
| 4635 | static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *tsk) |
| 4636 | { |
| 4637 | int dest_cpu; |
| 4638 | cpumask_t mask; |
| 4639 | |
| 4640 | /* On same node? */ |
| 4641 | mask = node_to_cpumask(cpu_to_node(dead_cpu)); |
| 4642 | cpus_and(mask, mask, tsk->cpus_allowed); |
| 4643 | dest_cpu = any_online_cpu(mask); |
| 4644 | |
| 4645 | /* On any allowed CPU? */ |
| 4646 | if (dest_cpu == NR_CPUS) |
| 4647 | dest_cpu = any_online_cpu(tsk->cpus_allowed); |
| 4648 | |
| 4649 | /* No more Mr. Nice Guy. */ |
| 4650 | if (dest_cpu == NR_CPUS) { |
| 4651 | cpus_setall(tsk->cpus_allowed); |
| 4652 | dest_cpu = any_online_cpu(tsk->cpus_allowed); |
| 4653 | |
| 4654 | /* |
| 4655 | * Don't tell them about moving exiting tasks or |
| 4656 | * kernel threads (both mm NULL), since they never |
| 4657 | * leave kernel. |
| 4658 | */ |
| 4659 | if (tsk->mm && printk_ratelimit()) |
| 4660 | printk(KERN_INFO "process %d (%s) no " |
| 4661 | "longer affine to cpu%d\n", |
| 4662 | tsk->pid, tsk->comm, dead_cpu); |
| 4663 | } |
| 4664 | __migrate_task(tsk, dead_cpu, dest_cpu); |
| 4665 | } |
| 4666 | |
| 4667 | /* |
| 4668 | * While a dead CPU has no uninterruptible tasks queued at this point, |
| 4669 | * it might still have a nonzero ->nr_uninterruptible counter, because |
| 4670 | * for performance reasons the counter is not stricly tracking tasks to |
| 4671 | * their home CPUs. So we just add the counter to another CPU's counter, |
| 4672 | * to keep the global sum constant after CPU-down: |
| 4673 | */ |
| 4674 | static void migrate_nr_uninterruptible(runqueue_t *rq_src) |
| 4675 | { |
| 4676 | runqueue_t *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL)); |
| 4677 | unsigned long flags; |
| 4678 | |
| 4679 | local_irq_save(flags); |
| 4680 | double_rq_lock(rq_src, rq_dest); |
| 4681 | rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible; |
| 4682 | rq_src->nr_uninterruptible = 0; |
| 4683 | double_rq_unlock(rq_src, rq_dest); |
| 4684 | local_irq_restore(flags); |
| 4685 | } |
| 4686 | |
| 4687 | /* Run through task list and migrate tasks from the dead cpu. */ |
| 4688 | static void migrate_live_tasks(int src_cpu) |
| 4689 | { |
| 4690 | struct task_struct *tsk, *t; |
| 4691 | |
| 4692 | write_lock_irq(&tasklist_lock); |
| 4693 | |
| 4694 | do_each_thread(t, tsk) { |
| 4695 | if (tsk == current) |
| 4696 | continue; |
| 4697 | |
| 4698 | if (task_cpu(tsk) == src_cpu) |
| 4699 | move_task_off_dead_cpu(src_cpu, tsk); |
| 4700 | } while_each_thread(t, tsk); |
| 4701 | |
| 4702 | write_unlock_irq(&tasklist_lock); |
| 4703 | } |
| 4704 | |
| 4705 | /* Schedules idle task to be the next runnable task on current CPU. |
| 4706 | * It does so by boosting its priority to highest possible and adding it to |
| 4707 | * the _front_ of runqueue. Used by CPU offline code. |
| 4708 | */ |
| 4709 | void sched_idle_next(void) |
| 4710 | { |
| 4711 | int cpu = smp_processor_id(); |
| 4712 | runqueue_t *rq = this_rq(); |
| 4713 | struct task_struct *p = rq->idle; |
| 4714 | unsigned long flags; |
| 4715 | |
| 4716 | /* cpu has to be offline */ |
| 4717 | BUG_ON(cpu_online(cpu)); |
| 4718 | |
| 4719 | /* Strictly not necessary since rest of the CPUs are stopped by now |
| 4720 | * and interrupts disabled on current cpu. |
| 4721 | */ |
| 4722 | spin_lock_irqsave(&rq->lock, flags); |
| 4723 | |
| 4724 | __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1); |
| 4725 | /* Add idle task to _front_ of it's priority queue */ |
| 4726 | __activate_idle_task(p, rq); |
| 4727 | |
| 4728 | spin_unlock_irqrestore(&rq->lock, flags); |
| 4729 | } |
| 4730 | |
| 4731 | /* Ensures that the idle task is using init_mm right before its cpu goes |
| 4732 | * offline. |
| 4733 | */ |
| 4734 | void idle_task_exit(void) |
| 4735 | { |
| 4736 | struct mm_struct *mm = current->active_mm; |
| 4737 | |
| 4738 | BUG_ON(cpu_online(smp_processor_id())); |
| 4739 | |
| 4740 | if (mm != &init_mm) |
| 4741 | switch_mm(mm, &init_mm, current); |
| 4742 | mmdrop(mm); |
| 4743 | } |
| 4744 | |
| 4745 | static void migrate_dead(unsigned int dead_cpu, task_t *tsk) |
| 4746 | { |
| 4747 | struct runqueue *rq = cpu_rq(dead_cpu); |
| 4748 | |
| 4749 | /* Must be exiting, otherwise would be on tasklist. */ |
| 4750 | BUG_ON(tsk->exit_state != EXIT_ZOMBIE && tsk->exit_state != EXIT_DEAD); |
| 4751 | |
| 4752 | /* Cannot have done final schedule yet: would have vanished. */ |
| 4753 | BUG_ON(tsk->flags & PF_DEAD); |
| 4754 | |
| 4755 | get_task_struct(tsk); |
| 4756 | |
| 4757 | /* |
| 4758 | * Drop lock around migration; if someone else moves it, |
| 4759 | * that's OK. No task can be added to this CPU, so iteration is |
| 4760 | * fine. |
| 4761 | */ |
| 4762 | spin_unlock_irq(&rq->lock); |
| 4763 | move_task_off_dead_cpu(dead_cpu, tsk); |
| 4764 | spin_lock_irq(&rq->lock); |
| 4765 | |
| 4766 | put_task_struct(tsk); |
| 4767 | } |
| 4768 | |
| 4769 | /* release_task() removes task from tasklist, so we won't find dead tasks. */ |
| 4770 | static void migrate_dead_tasks(unsigned int dead_cpu) |
| 4771 | { |
| 4772 | unsigned arr, i; |
| 4773 | struct runqueue *rq = cpu_rq(dead_cpu); |
| 4774 | |
| 4775 | for (arr = 0; arr < 2; arr++) { |
| 4776 | for (i = 0; i < MAX_PRIO; i++) { |
| 4777 | struct list_head *list = &rq->arrays[arr].queue[i]; |
| 4778 | while (!list_empty(list)) |
| 4779 | migrate_dead(dead_cpu, |
| 4780 | list_entry(list->next, task_t, |
| 4781 | run_list)); |
| 4782 | } |
| 4783 | } |
| 4784 | } |
| 4785 | #endif /* CONFIG_HOTPLUG_CPU */ |
| 4786 | |
| 4787 | /* |
| 4788 | * migration_call - callback that gets triggered when a CPU is added. |
| 4789 | * Here we can start up the necessary migration thread for the new CPU. |
| 4790 | */ |
| 4791 | static int migration_call(struct notifier_block *nfb, unsigned long action, |
| 4792 | void *hcpu) |
| 4793 | { |
| 4794 | int cpu = (long)hcpu; |
| 4795 | struct task_struct *p; |
| 4796 | struct runqueue *rq; |
| 4797 | unsigned long flags; |
| 4798 | |
| 4799 | switch (action) { |
| 4800 | case CPU_UP_PREPARE: |
| 4801 | p = kthread_create(migration_thread, hcpu, "migration/%d",cpu); |
| 4802 | if (IS_ERR(p)) |
| 4803 | return NOTIFY_BAD; |
| 4804 | p->flags |= PF_NOFREEZE; |
| 4805 | kthread_bind(p, cpu); |
| 4806 | /* Must be high prio: stop_machine expects to yield to it. */ |
| 4807 | rq = task_rq_lock(p, &flags); |
| 4808 | __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1); |
| 4809 | task_rq_unlock(rq, &flags); |
| 4810 | cpu_rq(cpu)->migration_thread = p; |
| 4811 | break; |
| 4812 | case CPU_ONLINE: |
| 4813 | /* Strictly unneccessary, as first user will wake it. */ |
| 4814 | wake_up_process(cpu_rq(cpu)->migration_thread); |
| 4815 | break; |
| 4816 | #ifdef CONFIG_HOTPLUG_CPU |
| 4817 | case CPU_UP_CANCELED: |
| 4818 | /* Unbind it from offline cpu so it can run. Fall thru. */ |
| 4819 | kthread_bind(cpu_rq(cpu)->migration_thread, |
| 4820 | any_online_cpu(cpu_online_map)); |
| 4821 | kthread_stop(cpu_rq(cpu)->migration_thread); |
| 4822 | cpu_rq(cpu)->migration_thread = NULL; |
| 4823 | break; |
| 4824 | case CPU_DEAD: |
| 4825 | migrate_live_tasks(cpu); |
| 4826 | rq = cpu_rq(cpu); |
| 4827 | kthread_stop(rq->migration_thread); |
| 4828 | rq->migration_thread = NULL; |
| 4829 | /* Idle task back to normal (off runqueue, low prio) */ |
| 4830 | rq = task_rq_lock(rq->idle, &flags); |
| 4831 | deactivate_task(rq->idle, rq); |
| 4832 | rq->idle->static_prio = MAX_PRIO; |
| 4833 | __setscheduler(rq->idle, SCHED_NORMAL, 0); |
| 4834 | migrate_dead_tasks(cpu); |
| 4835 | task_rq_unlock(rq, &flags); |
| 4836 | migrate_nr_uninterruptible(rq); |
| 4837 | BUG_ON(rq->nr_running != 0); |
| 4838 | |
| 4839 | /* No need to migrate the tasks: it was best-effort if |
| 4840 | * they didn't do lock_cpu_hotplug(). Just wake up |
| 4841 | * the requestors. */ |
| 4842 | spin_lock_irq(&rq->lock); |
| 4843 | while (!list_empty(&rq->migration_queue)) { |
| 4844 | migration_req_t *req; |
| 4845 | req = list_entry(rq->migration_queue.next, |
| 4846 | migration_req_t, list); |
| 4847 | list_del_init(&req->list); |
| 4848 | complete(&req->done); |
| 4849 | } |
| 4850 | spin_unlock_irq(&rq->lock); |
| 4851 | break; |
| 4852 | #endif |
| 4853 | } |
| 4854 | return NOTIFY_OK; |
| 4855 | } |
| 4856 | |
| 4857 | /* Register at highest priority so that task migration (migrate_all_tasks) |
| 4858 | * happens before everything else. |
| 4859 | */ |
| 4860 | static struct notifier_block __devinitdata migration_notifier = { |
| 4861 | .notifier_call = migration_call, |
| 4862 | .priority = 10 |
| 4863 | }; |
| 4864 | |
| 4865 | int __init migration_init(void) |
| 4866 | { |
| 4867 | void *cpu = (void *)(long)smp_processor_id(); |
| 4868 | /* Start one for boot CPU. */ |
| 4869 | migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); |
| 4870 | migration_call(&migration_notifier, CPU_ONLINE, cpu); |
| 4871 | register_cpu_notifier(&migration_notifier); |
| 4872 | return 0; |
| 4873 | } |
| 4874 | #endif |
| 4875 | |
| 4876 | #ifdef CONFIG_SMP |
| 4877 | #undef SCHED_DOMAIN_DEBUG |
| 4878 | #ifdef SCHED_DOMAIN_DEBUG |
| 4879 | static void sched_domain_debug(struct sched_domain *sd, int cpu) |
| 4880 | { |
| 4881 | int level = 0; |
| 4882 | |
| 4883 | if (!sd) { |
| 4884 | printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); |
| 4885 | return; |
| 4886 | } |
| 4887 | |
| 4888 | printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); |
| 4889 | |
| 4890 | do { |
| 4891 | int i; |
| 4892 | char str[NR_CPUS]; |
| 4893 | struct sched_group *group = sd->groups; |
| 4894 | cpumask_t groupmask; |
| 4895 | |
| 4896 | cpumask_scnprintf(str, NR_CPUS, sd->span); |
| 4897 | cpus_clear(groupmask); |
| 4898 | |
| 4899 | printk(KERN_DEBUG); |
| 4900 | for (i = 0; i < level + 1; i++) |
| 4901 | printk(" "); |
| 4902 | printk("domain %d: ", level); |
| 4903 | |
| 4904 | if (!(sd->flags & SD_LOAD_BALANCE)) { |
| 4905 | printk("does not load-balance\n"); |
| 4906 | if (sd->parent) |
| 4907 | printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent"); |
| 4908 | break; |
| 4909 | } |
| 4910 | |
| 4911 | printk("span %s\n", str); |
| 4912 | |
| 4913 | if (!cpu_isset(cpu, sd->span)) |
| 4914 | printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu); |
| 4915 | if (!cpu_isset(cpu, group->cpumask)) |
| 4916 | printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu); |
| 4917 | |
| 4918 | printk(KERN_DEBUG); |
| 4919 | for (i = 0; i < level + 2; i++) |
| 4920 | printk(" "); |
| 4921 | printk("groups:"); |
| 4922 | do { |
| 4923 | if (!group) { |
| 4924 | printk("\n"); |
| 4925 | printk(KERN_ERR "ERROR: group is NULL\n"); |
| 4926 | break; |
| 4927 | } |
| 4928 | |
| 4929 | if (!group->cpu_power) { |
| 4930 | printk("\n"); |
| 4931 | printk(KERN_ERR "ERROR: domain->cpu_power not set\n"); |
| 4932 | } |
| 4933 | |
| 4934 | if (!cpus_weight(group->cpumask)) { |
| 4935 | printk("\n"); |
| 4936 | printk(KERN_ERR "ERROR: empty group\n"); |
| 4937 | } |
| 4938 | |
| 4939 | if (cpus_intersects(groupmask, group->cpumask)) { |
| 4940 | printk("\n"); |
| 4941 | printk(KERN_ERR "ERROR: repeated CPUs\n"); |
| 4942 | } |
| 4943 | |
| 4944 | cpus_or(groupmask, groupmask, group->cpumask); |
| 4945 | |
| 4946 | cpumask_scnprintf(str, NR_CPUS, group->cpumask); |
| 4947 | printk(" %s", str); |
| 4948 | |
| 4949 | group = group->next; |
| 4950 | } while (group != sd->groups); |
| 4951 | printk("\n"); |
| 4952 | |
| 4953 | if (!cpus_equal(sd->span, groupmask)) |
| 4954 | printk(KERN_ERR "ERROR: groups don't span domain->span\n"); |
| 4955 | |
| 4956 | level++; |
| 4957 | sd = sd->parent; |
| 4958 | |
| 4959 | if (sd) { |
| 4960 | if (!cpus_subset(groupmask, sd->span)) |
| 4961 | printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n"); |
| 4962 | } |
| 4963 | |
| 4964 | } while (sd); |
| 4965 | } |
| 4966 | #else |
| 4967 | #define sched_domain_debug(sd, cpu) {} |
| 4968 | #endif |
| 4969 | |
| 4970 | static int sd_degenerate(struct sched_domain *sd) |
| 4971 | { |
| 4972 | if (cpus_weight(sd->span) == 1) |
| 4973 | return 1; |
| 4974 | |
| 4975 | /* Following flags need at least 2 groups */ |
| 4976 | if (sd->flags & (SD_LOAD_BALANCE | |
| 4977 | SD_BALANCE_NEWIDLE | |
| 4978 | SD_BALANCE_FORK | |
| 4979 | SD_BALANCE_EXEC)) { |
| 4980 | if (sd->groups != sd->groups->next) |
| 4981 | return 0; |
| 4982 | } |
| 4983 | |
| 4984 | /* Following flags don't use groups */ |
| 4985 | if (sd->flags & (SD_WAKE_IDLE | |
| 4986 | SD_WAKE_AFFINE | |
| 4987 | SD_WAKE_BALANCE)) |
| 4988 | return 0; |
| 4989 | |
| 4990 | return 1; |
| 4991 | } |
| 4992 | |
| 4993 | static int sd_parent_degenerate(struct sched_domain *sd, |
| 4994 | struct sched_domain *parent) |
| 4995 | { |
| 4996 | unsigned long cflags = sd->flags, pflags = parent->flags; |
| 4997 | |
| 4998 | if (sd_degenerate(parent)) |
| 4999 | return 1; |
| 5000 | |
| 5001 | if (!cpus_equal(sd->span, parent->span)) |
| 5002 | return 0; |
| 5003 | |
| 5004 | /* Does parent contain flags not in child? */ |
| 5005 | /* WAKE_BALANCE is a subset of WAKE_AFFINE */ |
| 5006 | if (cflags & SD_WAKE_AFFINE) |
| 5007 | pflags &= ~SD_WAKE_BALANCE; |
| 5008 | /* Flags needing groups don't count if only 1 group in parent */ |
| 5009 | if (parent->groups == parent->groups->next) { |
| 5010 | pflags &= ~(SD_LOAD_BALANCE | |
| 5011 | SD_BALANCE_NEWIDLE | |
| 5012 | SD_BALANCE_FORK | |
| 5013 | SD_BALANCE_EXEC); |
| 5014 | } |
| 5015 | if (~cflags & pflags) |
| 5016 | return 0; |
| 5017 | |
| 5018 | return 1; |
| 5019 | } |
| 5020 | |
| 5021 | /* |
| 5022 | * Attach the domain 'sd' to 'cpu' as its base domain. Callers must |
| 5023 | * hold the hotplug lock. |
| 5024 | */ |
| 5025 | static void cpu_attach_domain(struct sched_domain *sd, int cpu) |
| 5026 | { |
| 5027 | runqueue_t *rq = cpu_rq(cpu); |
| 5028 | struct sched_domain *tmp; |
| 5029 | |
| 5030 | /* Remove the sched domains which do not contribute to scheduling. */ |
| 5031 | for (tmp = sd; tmp; tmp = tmp->parent) { |
| 5032 | struct sched_domain *parent = tmp->parent; |
| 5033 | if (!parent) |
| 5034 | break; |
| 5035 | if (sd_parent_degenerate(tmp, parent)) |
| 5036 | tmp->parent = parent->parent; |
| 5037 | } |
| 5038 | |
| 5039 | if (sd && sd_degenerate(sd)) |
| 5040 | sd = sd->parent; |
| 5041 | |
| 5042 | sched_domain_debug(sd, cpu); |
| 5043 | |
| 5044 | rcu_assign_pointer(rq->sd, sd); |
| 5045 | } |
| 5046 | |
| 5047 | /* cpus with isolated domains */ |
| 5048 | static cpumask_t __devinitdata cpu_isolated_map = CPU_MASK_NONE; |
| 5049 | |
| 5050 | /* Setup the mask of cpus configured for isolated domains */ |
| 5051 | static int __init isolated_cpu_setup(char *str) |
| 5052 | { |
| 5053 | int ints[NR_CPUS], i; |
| 5054 | |
| 5055 | str = get_options(str, ARRAY_SIZE(ints), ints); |
| 5056 | cpus_clear(cpu_isolated_map); |
| 5057 | for (i = 1; i <= ints[0]; i++) |
| 5058 | if (ints[i] < NR_CPUS) |
| 5059 | cpu_set(ints[i], cpu_isolated_map); |
| 5060 | return 1; |
| 5061 | } |
| 5062 | |
| 5063 | __setup ("isolcpus=", isolated_cpu_setup); |
| 5064 | |
| 5065 | /* |
| 5066 | * init_sched_build_groups takes an array of groups, the cpumask we wish |
| 5067 | * to span, and a pointer to a function which identifies what group a CPU |
| 5068 | * belongs to. The return value of group_fn must be a valid index into the |
| 5069 | * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we |
| 5070 | * keep track of groups covered with a cpumask_t). |
| 5071 | * |
| 5072 | * init_sched_build_groups will build a circular linked list of the groups |
| 5073 | * covered by the given span, and will set each group's ->cpumask correctly, |
| 5074 | * and ->cpu_power to 0. |
| 5075 | */ |
| 5076 | static void init_sched_build_groups(struct sched_group groups[], cpumask_t span, |
| 5077 | int (*group_fn)(int cpu)) |
| 5078 | { |
| 5079 | struct sched_group *first = NULL, *last = NULL; |
| 5080 | cpumask_t covered = CPU_MASK_NONE; |
| 5081 | int i; |
| 5082 | |
| 5083 | for_each_cpu_mask(i, span) { |
| 5084 | int group = group_fn(i); |
| 5085 | struct sched_group *sg = &groups[group]; |
| 5086 | int j; |
| 5087 | |
| 5088 | if (cpu_isset(i, covered)) |
| 5089 | continue; |
| 5090 | |
| 5091 | sg->cpumask = CPU_MASK_NONE; |
| 5092 | sg->cpu_power = 0; |
| 5093 | |
| 5094 | for_each_cpu_mask(j, span) { |
| 5095 | if (group_fn(j) != group) |
| 5096 | continue; |
| 5097 | |
| 5098 | cpu_set(j, covered); |
| 5099 | cpu_set(j, sg->cpumask); |
| 5100 | } |
| 5101 | if (!first) |
| 5102 | first = sg; |
| 5103 | if (last) |
| 5104 | last->next = sg; |
| 5105 | last = sg; |
| 5106 | } |
| 5107 | last->next = first; |
| 5108 | } |
| 5109 | |
| 5110 | #define SD_NODES_PER_DOMAIN 16 |
| 5111 | |
| 5112 | /* |
| 5113 | * Self-tuning task migration cost measurement between source and target CPUs. |
| 5114 | * |
| 5115 | * This is done by measuring the cost of manipulating buffers of varying |
| 5116 | * sizes. For a given buffer-size here are the steps that are taken: |
| 5117 | * |
| 5118 | * 1) the source CPU reads+dirties a shared buffer |
| 5119 | * 2) the target CPU reads+dirties the same shared buffer |
| 5120 | * |
| 5121 | * We measure how long they take, in the following 4 scenarios: |
| 5122 | * |
| 5123 | * - source: CPU1, target: CPU2 | cost1 |
| 5124 | * - source: CPU2, target: CPU1 | cost2 |
| 5125 | * - source: CPU1, target: CPU1 | cost3 |
| 5126 | * - source: CPU2, target: CPU2 | cost4 |
| 5127 | * |
| 5128 | * We then calculate the cost3+cost4-cost1-cost2 difference - this is |
| 5129 | * the cost of migration. |
| 5130 | * |
| 5131 | * We then start off from a small buffer-size and iterate up to larger |
| 5132 | * buffer sizes, in 5% steps - measuring each buffer-size separately, and |
| 5133 | * doing a maximum search for the cost. (The maximum cost for a migration |
| 5134 | * normally occurs when the working set size is around the effective cache |
| 5135 | * size.) |
| 5136 | */ |
| 5137 | #define SEARCH_SCOPE 2 |
| 5138 | #define MIN_CACHE_SIZE (64*1024U) |
| 5139 | #define DEFAULT_CACHE_SIZE (5*1024*1024U) |
| 5140 | #define ITERATIONS 2 |
| 5141 | #define SIZE_THRESH 130 |
| 5142 | #define COST_THRESH 130 |
| 5143 | |
| 5144 | /* |
| 5145 | * The migration cost is a function of 'domain distance'. Domain |
| 5146 | * distance is the number of steps a CPU has to iterate down its |
| 5147 | * domain tree to share a domain with the other CPU. The farther |
| 5148 | * two CPUs are from each other, the larger the distance gets. |
| 5149 | * |
| 5150 | * Note that we use the distance only to cache measurement results, |
| 5151 | * the distance value is not used numerically otherwise. When two |
| 5152 | * CPUs have the same distance it is assumed that the migration |
| 5153 | * cost is the same. (this is a simplification but quite practical) |
| 5154 | */ |
| 5155 | #define MAX_DOMAIN_DISTANCE 32 |
| 5156 | |
| 5157 | static unsigned long long migration_cost[MAX_DOMAIN_DISTANCE] = |
| 5158 | { [ 0 ... MAX_DOMAIN_DISTANCE-1 ] = -1LL }; |
| 5159 | |
| 5160 | /* |
| 5161 | * Allow override of migration cost - in units of microseconds. |
| 5162 | * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost |
| 5163 | * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs: |
| 5164 | */ |
| 5165 | static int __init migration_cost_setup(char *str) |
| 5166 | { |
| 5167 | int ints[MAX_DOMAIN_DISTANCE+1], i; |
| 5168 | |
| 5169 | str = get_options(str, ARRAY_SIZE(ints), ints); |
| 5170 | |
| 5171 | printk("#ints: %d\n", ints[0]); |
| 5172 | for (i = 1; i <= ints[0]; i++) { |
| 5173 | migration_cost[i-1] = (unsigned long long)ints[i]*1000; |
| 5174 | printk("migration_cost[%d]: %Ld\n", i-1, migration_cost[i-1]); |
| 5175 | } |
| 5176 | return 1; |
| 5177 | } |
| 5178 | |
| 5179 | __setup ("migration_cost=", migration_cost_setup); |
| 5180 | |
| 5181 | /* |
| 5182 | * Global multiplier (divisor) for migration-cutoff values, |
| 5183 | * in percentiles. E.g. use a value of 150 to get 1.5 times |
| 5184 | * longer cache-hot cutoff times. |
| 5185 | * |
| 5186 | * (We scale it from 100 to 128 to long long handling easier.) |
| 5187 | */ |
| 5188 | |
| 5189 | #define MIGRATION_FACTOR_SCALE 128 |
| 5190 | |
| 5191 | static unsigned int migration_factor = MIGRATION_FACTOR_SCALE; |
| 5192 | |
| 5193 | static int __init setup_migration_factor(char *str) |
| 5194 | { |
| 5195 | get_option(&str, &migration_factor); |
| 5196 | migration_factor = migration_factor * MIGRATION_FACTOR_SCALE / 100; |
| 5197 | return 1; |
| 5198 | } |
| 5199 | |
| 5200 | __setup("migration_factor=", setup_migration_factor); |
| 5201 | |
| 5202 | /* |
| 5203 | * Estimated distance of two CPUs, measured via the number of domains |
| 5204 | * we have to pass for the two CPUs to be in the same span: |
| 5205 | */ |
| 5206 | static unsigned long domain_distance(int cpu1, int cpu2) |
| 5207 | { |
| 5208 | unsigned long distance = 0; |
| 5209 | struct sched_domain *sd; |
| 5210 | |
| 5211 | for_each_domain(cpu1, sd) { |
| 5212 | WARN_ON(!cpu_isset(cpu1, sd->span)); |
| 5213 | if (cpu_isset(cpu2, sd->span)) |
| 5214 | return distance; |
| 5215 | distance++; |
| 5216 | } |
| 5217 | if (distance >= MAX_DOMAIN_DISTANCE) { |
| 5218 | WARN_ON(1); |
| 5219 | distance = MAX_DOMAIN_DISTANCE-1; |
| 5220 | } |
| 5221 | |
| 5222 | return distance; |
| 5223 | } |
| 5224 | |
| 5225 | static unsigned int migration_debug; |
| 5226 | |
| 5227 | static int __init setup_migration_debug(char *str) |
| 5228 | { |
| 5229 | get_option(&str, &migration_debug); |
| 5230 | return 1; |
| 5231 | } |
| 5232 | |
| 5233 | __setup("migration_debug=", setup_migration_debug); |
| 5234 | |
| 5235 | /* |
| 5236 | * Maximum cache-size that the scheduler should try to measure. |
| 5237 | * Architectures with larger caches should tune this up during |
| 5238 | * bootup. Gets used in the domain-setup code (i.e. during SMP |
| 5239 | * bootup). |
| 5240 | */ |
| 5241 | unsigned int max_cache_size; |
| 5242 | |
| 5243 | static int __init setup_max_cache_size(char *str) |
| 5244 | { |
| 5245 | get_option(&str, &max_cache_size); |
| 5246 | return 1; |
| 5247 | } |
| 5248 | |
| 5249 | __setup("max_cache_size=", setup_max_cache_size); |
| 5250 | |
| 5251 | /* |
| 5252 | * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This |
| 5253 | * is the operation that is timed, so we try to generate unpredictable |
| 5254 | * cachemisses that still end up filling the L2 cache: |
| 5255 | */ |
| 5256 | static void touch_cache(void *__cache, unsigned long __size) |
| 5257 | { |
| 5258 | unsigned long size = __size/sizeof(long), chunk1 = size/3, |
| 5259 | chunk2 = 2*size/3; |
| 5260 | unsigned long *cache = __cache; |
| 5261 | int i; |
| 5262 | |
| 5263 | for (i = 0; i < size/6; i += 8) { |
| 5264 | switch (i % 6) { |
| 5265 | case 0: cache[i]++; |
| 5266 | case 1: cache[size-1-i]++; |
| 5267 | case 2: cache[chunk1-i]++; |
| 5268 | case 3: cache[chunk1+i]++; |
| 5269 | case 4: cache[chunk2-i]++; |
| 5270 | case 5: cache[chunk2+i]++; |
| 5271 | } |
| 5272 | } |
| 5273 | } |
| 5274 | |
| 5275 | /* |
| 5276 | * Measure the cache-cost of one task migration. Returns in units of nsec. |
| 5277 | */ |
| 5278 | static unsigned long long measure_one(void *cache, unsigned long size, |
| 5279 | int source, int target) |
| 5280 | { |
| 5281 | cpumask_t mask, saved_mask; |
| 5282 | unsigned long long t0, t1, t2, t3, cost; |
| 5283 | |
| 5284 | saved_mask = current->cpus_allowed; |
| 5285 | |
| 5286 | /* |
| 5287 | * Flush source caches to RAM and invalidate them: |
| 5288 | */ |
| 5289 | sched_cacheflush(); |
| 5290 | |
| 5291 | /* |
| 5292 | * Migrate to the source CPU: |
| 5293 | */ |
| 5294 | mask = cpumask_of_cpu(source); |
| 5295 | set_cpus_allowed(current, mask); |
| 5296 | WARN_ON(smp_processor_id() != source); |
| 5297 | |
| 5298 | /* |
| 5299 | * Dirty the working set: |
| 5300 | */ |
| 5301 | t0 = sched_clock(); |
| 5302 | touch_cache(cache, size); |
| 5303 | t1 = sched_clock(); |
| 5304 | |
| 5305 | /* |
| 5306 | * Migrate to the target CPU, dirty the L2 cache and access |
| 5307 | * the shared buffer. (which represents the working set |
| 5308 | * of a migrated task.) |
| 5309 | */ |
| 5310 | mask = cpumask_of_cpu(target); |
| 5311 | set_cpus_allowed(current, mask); |
| 5312 | WARN_ON(smp_processor_id() != target); |
| 5313 | |
| 5314 | t2 = sched_clock(); |
| 5315 | touch_cache(cache, size); |
| 5316 | t3 = sched_clock(); |
| 5317 | |
| 5318 | cost = t1-t0 + t3-t2; |
| 5319 | |
| 5320 | if (migration_debug >= 2) |
| 5321 | printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n", |
| 5322 | source, target, t1-t0, t1-t0, t3-t2, cost); |
| 5323 | /* |
| 5324 | * Flush target caches to RAM and invalidate them: |
| 5325 | */ |
| 5326 | sched_cacheflush(); |
| 5327 | |
| 5328 | set_cpus_allowed(current, saved_mask); |
| 5329 | |
| 5330 | return cost; |
| 5331 | } |
| 5332 | |
| 5333 | /* |
| 5334 | * Measure a series of task migrations and return the average |
| 5335 | * result. Since this code runs early during bootup the system |
| 5336 | * is 'undisturbed' and the average latency makes sense. |
| 5337 | * |
| 5338 | * The algorithm in essence auto-detects the relevant cache-size, |
| 5339 | * so it will properly detect different cachesizes for different |
| 5340 | * cache-hierarchies, depending on how the CPUs are connected. |
| 5341 | * |
| 5342 | * Architectures can prime the upper limit of the search range via |
| 5343 | * max_cache_size, otherwise the search range defaults to 20MB...64K. |
| 5344 | */ |
| 5345 | static unsigned long long |
| 5346 | measure_cost(int cpu1, int cpu2, void *cache, unsigned int size) |
| 5347 | { |
| 5348 | unsigned long long cost1, cost2; |
| 5349 | int i; |
| 5350 | |
| 5351 | /* |
| 5352 | * Measure the migration cost of 'size' bytes, over an |
| 5353 | * average of 10 runs: |
| 5354 | * |
| 5355 | * (We perturb the cache size by a small (0..4k) |
| 5356 | * value to compensate size/alignment related artifacts. |
| 5357 | * We also subtract the cost of the operation done on |
| 5358 | * the same CPU.) |
| 5359 | */ |
| 5360 | cost1 = 0; |
| 5361 | |
| 5362 | /* |
| 5363 | * dry run, to make sure we start off cache-cold on cpu1, |
| 5364 | * and to get any vmalloc pagefaults in advance: |
| 5365 | */ |
| 5366 | measure_one(cache, size, cpu1, cpu2); |
| 5367 | for (i = 0; i < ITERATIONS; i++) |
| 5368 | cost1 += measure_one(cache, size - i*1024, cpu1, cpu2); |
| 5369 | |
| 5370 | measure_one(cache, size, cpu2, cpu1); |
| 5371 | for (i = 0; i < ITERATIONS; i++) |
| 5372 | cost1 += measure_one(cache, size - i*1024, cpu2, cpu1); |
| 5373 | |
| 5374 | /* |
| 5375 | * (We measure the non-migrating [cached] cost on both |
| 5376 | * cpu1 and cpu2, to handle CPUs with different speeds) |
| 5377 | */ |
| 5378 | cost2 = 0; |
| 5379 | |
| 5380 | measure_one(cache, size, cpu1, cpu1); |
| 5381 | for (i = 0; i < ITERATIONS; i++) |
| 5382 | cost2 += measure_one(cache, size - i*1024, cpu1, cpu1); |
| 5383 | |
| 5384 | measure_one(cache, size, cpu2, cpu2); |
| 5385 | for (i = 0; i < ITERATIONS; i++) |
| 5386 | cost2 += measure_one(cache, size - i*1024, cpu2, cpu2); |
| 5387 | |
| 5388 | /* |
| 5389 | * Get the per-iteration migration cost: |
| 5390 | */ |
| 5391 | do_div(cost1, 2*ITERATIONS); |
| 5392 | do_div(cost2, 2*ITERATIONS); |
| 5393 | |
| 5394 | return cost1 - cost2; |
| 5395 | } |
| 5396 | |
| 5397 | static unsigned long long measure_migration_cost(int cpu1, int cpu2) |
| 5398 | { |
| 5399 | unsigned long long max_cost = 0, fluct = 0, avg_fluct = 0; |
| 5400 | unsigned int max_size, size, size_found = 0; |
| 5401 | long long cost = 0, prev_cost; |
| 5402 | void *cache; |
| 5403 | |
| 5404 | /* |
| 5405 | * Search from max_cache_size*5 down to 64K - the real relevant |
| 5406 | * cachesize has to lie somewhere inbetween. |
| 5407 | */ |
| 5408 | if (max_cache_size) { |
| 5409 | max_size = max(max_cache_size * SEARCH_SCOPE, MIN_CACHE_SIZE); |
| 5410 | size = max(max_cache_size / SEARCH_SCOPE, MIN_CACHE_SIZE); |
| 5411 | } else { |
| 5412 | /* |
| 5413 | * Since we have no estimation about the relevant |
| 5414 | * search range |
| 5415 | */ |
| 5416 | max_size = DEFAULT_CACHE_SIZE * SEARCH_SCOPE; |
| 5417 | size = MIN_CACHE_SIZE; |
| 5418 | } |
| 5419 | |
| 5420 | if (!cpu_online(cpu1) || !cpu_online(cpu2)) { |
| 5421 | printk("cpu %d and %d not both online!\n", cpu1, cpu2); |
| 5422 | return 0; |
| 5423 | } |
| 5424 | |
| 5425 | /* |
| 5426 | * Allocate the working set: |
| 5427 | */ |
| 5428 | cache = vmalloc(max_size); |
| 5429 | if (!cache) { |
| 5430 | printk("could not vmalloc %d bytes for cache!\n", 2*max_size); |
| 5431 | return 1000000; // return 1 msec on very small boxen |
| 5432 | } |
| 5433 | |
| 5434 | while (size <= max_size) { |
| 5435 | prev_cost = cost; |
| 5436 | cost = measure_cost(cpu1, cpu2, cache, size); |
| 5437 | |
| 5438 | /* |
| 5439 | * Update the max: |
| 5440 | */ |
| 5441 | if (cost > 0) { |
| 5442 | if (max_cost < cost) { |
| 5443 | max_cost = cost; |
| 5444 | size_found = size; |
| 5445 | } |
| 5446 | } |
| 5447 | /* |
| 5448 | * Calculate average fluctuation, we use this to prevent |
| 5449 | * noise from triggering an early break out of the loop: |
| 5450 | */ |
| 5451 | fluct = abs(cost - prev_cost); |
| 5452 | avg_fluct = (avg_fluct + fluct)/2; |
| 5453 | |
| 5454 | if (migration_debug) |
| 5455 | printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): (%8Ld %8Ld)\n", |
| 5456 | cpu1, cpu2, size, |
| 5457 | (long)cost / 1000000, |
| 5458 | ((long)cost / 100000) % 10, |
| 5459 | (long)max_cost / 1000000, |
| 5460 | ((long)max_cost / 100000) % 10, |
| 5461 | domain_distance(cpu1, cpu2), |
| 5462 | cost, avg_fluct); |
| 5463 | |
| 5464 | /* |
| 5465 | * If we iterated at least 20% past the previous maximum, |
| 5466 | * and the cost has dropped by more than 20% already, |
| 5467 | * (taking fluctuations into account) then we assume to |
| 5468 | * have found the maximum and break out of the loop early: |
| 5469 | */ |
| 5470 | if (size_found && (size*100 > size_found*SIZE_THRESH)) |
| 5471 | if (cost+avg_fluct <= 0 || |
| 5472 | max_cost*100 > (cost+avg_fluct)*COST_THRESH) { |
| 5473 | |
| 5474 | if (migration_debug) |
| 5475 | printk("-> found max.\n"); |
| 5476 | break; |
| 5477 | } |
| 5478 | /* |
| 5479 | * Increase the cachesize in 5% steps: |
| 5480 | */ |
| 5481 | size = size * 20 / 19; |
| 5482 | } |
| 5483 | |
| 5484 | if (migration_debug) |
| 5485 | printk("[%d][%d] working set size found: %d, cost: %Ld\n", |
| 5486 | cpu1, cpu2, size_found, max_cost); |
| 5487 | |
| 5488 | vfree(cache); |
| 5489 | |
| 5490 | /* |
| 5491 | * A task is considered 'cache cold' if at least 2 times |
| 5492 | * the worst-case cost of migration has passed. |
| 5493 | * |
| 5494 | * (this limit is only listened to if the load-balancing |
| 5495 | * situation is 'nice' - if there is a large imbalance we |
| 5496 | * ignore it for the sake of CPU utilization and |
| 5497 | * processing fairness.) |
| 5498 | */ |
| 5499 | return 2 * max_cost * migration_factor / MIGRATION_FACTOR_SCALE; |
| 5500 | } |
| 5501 | |
| 5502 | static void calibrate_migration_costs(const cpumask_t *cpu_map) |
| 5503 | { |
| 5504 | int cpu1 = -1, cpu2 = -1, cpu, orig_cpu = raw_smp_processor_id(); |
| 5505 | unsigned long j0, j1, distance, max_distance = 0; |
| 5506 | struct sched_domain *sd; |
| 5507 | |
| 5508 | j0 = jiffies; |
| 5509 | |
| 5510 | /* |
| 5511 | * First pass - calculate the cacheflush times: |
| 5512 | */ |
| 5513 | for_each_cpu_mask(cpu1, *cpu_map) { |
| 5514 | for_each_cpu_mask(cpu2, *cpu_map) { |
| 5515 | if (cpu1 == cpu2) |
| 5516 | continue; |
| 5517 | distance = domain_distance(cpu1, cpu2); |
| 5518 | max_distance = max(max_distance, distance); |
| 5519 | /* |
| 5520 | * No result cached yet? |
| 5521 | */ |
| 5522 | if (migration_cost[distance] == -1LL) |
| 5523 | migration_cost[distance] = |
| 5524 | measure_migration_cost(cpu1, cpu2); |
| 5525 | } |
| 5526 | } |
| 5527 | /* |
| 5528 | * Second pass - update the sched domain hierarchy with |
| 5529 | * the new cache-hot-time estimations: |
| 5530 | */ |
| 5531 | for_each_cpu_mask(cpu, *cpu_map) { |
| 5532 | distance = 0; |
| 5533 | for_each_domain(cpu, sd) { |
| 5534 | sd->cache_hot_time = migration_cost[distance]; |
| 5535 | distance++; |
| 5536 | } |
| 5537 | } |
| 5538 | /* |
| 5539 | * Print the matrix: |
| 5540 | */ |
| 5541 | if (migration_debug) |
| 5542 | printk("migration: max_cache_size: %d, cpu: %d MHz:\n", |
| 5543 | max_cache_size, |
| 5544 | #ifdef CONFIG_X86 |
| 5545 | cpu_khz/1000 |
| 5546 | #else |
| 5547 | -1 |
| 5548 | #endif |
| 5549 | ); |
| 5550 | printk("migration_cost="); |
| 5551 | for (distance = 0; distance <= max_distance; distance++) { |
| 5552 | if (distance) |
| 5553 | printk(","); |
| 5554 | printk("%ld", (long)migration_cost[distance] / 1000); |
| 5555 | } |
| 5556 | printk("\n"); |
| 5557 | j1 = jiffies; |
| 5558 | if (migration_debug) |
| 5559 | printk("migration: %ld seconds\n", (j1-j0)/HZ); |
| 5560 | |
| 5561 | /* |
| 5562 | * Move back to the original CPU. NUMA-Q gets confused |
| 5563 | * if we migrate to another quad during bootup. |
| 5564 | */ |
| 5565 | if (raw_smp_processor_id() != orig_cpu) { |
| 5566 | cpumask_t mask = cpumask_of_cpu(orig_cpu), |
| 5567 | saved_mask = current->cpus_allowed; |
| 5568 | |
| 5569 | set_cpus_allowed(current, mask); |
| 5570 | set_cpus_allowed(current, saved_mask); |
| 5571 | } |
| 5572 | } |
| 5573 | |
| 5574 | #ifdef CONFIG_NUMA |
| 5575 | |
| 5576 | /** |
| 5577 | * find_next_best_node - find the next node to include in a sched_domain |
| 5578 | * @node: node whose sched_domain we're building |
| 5579 | * @used_nodes: nodes already in the sched_domain |
| 5580 | * |
| 5581 | * Find the next node to include in a given scheduling domain. Simply |
| 5582 | * finds the closest node not already in the @used_nodes map. |
| 5583 | * |
| 5584 | * Should use nodemask_t. |
| 5585 | */ |
| 5586 | static int find_next_best_node(int node, unsigned long *used_nodes) |
| 5587 | { |
| 5588 | int i, n, val, min_val, best_node = 0; |
| 5589 | |
| 5590 | min_val = INT_MAX; |
| 5591 | |
| 5592 | for (i = 0; i < MAX_NUMNODES; i++) { |
| 5593 | /* Start at @node */ |
| 5594 | n = (node + i) % MAX_NUMNODES; |
| 5595 | |
| 5596 | if (!nr_cpus_node(n)) |
| 5597 | continue; |
| 5598 | |
| 5599 | /* Skip already used nodes */ |
| 5600 | if (test_bit(n, used_nodes)) |
| 5601 | continue; |
| 5602 | |
| 5603 | /* Simple min distance search */ |
| 5604 | val = node_distance(node, n); |
| 5605 | |
| 5606 | if (val < min_val) { |
| 5607 | min_val = val; |
| 5608 | best_node = n; |
| 5609 | } |
| 5610 | } |
| 5611 | |
| 5612 | set_bit(best_node, used_nodes); |
| 5613 | return best_node; |
| 5614 | } |
| 5615 | |
| 5616 | /** |
| 5617 | * sched_domain_node_span - get a cpumask for a node's sched_domain |
| 5618 | * @node: node whose cpumask we're constructing |
| 5619 | * @size: number of nodes to include in this span |
| 5620 | * |
| 5621 | * Given a node, construct a good cpumask for its sched_domain to span. It |
| 5622 | * should be one that prevents unnecessary balancing, but also spreads tasks |
| 5623 | * out optimally. |
| 5624 | */ |
| 5625 | static cpumask_t sched_domain_node_span(int node) |
| 5626 | { |
| 5627 | int i; |
| 5628 | cpumask_t span, nodemask; |
| 5629 | DECLARE_BITMAP(used_nodes, MAX_NUMNODES); |
| 5630 | |
| 5631 | cpus_clear(span); |
| 5632 | bitmap_zero(used_nodes, MAX_NUMNODES); |
| 5633 | |
| 5634 | nodemask = node_to_cpumask(node); |
| 5635 | cpus_or(span, span, nodemask); |
| 5636 | set_bit(node, used_nodes); |
| 5637 | |
| 5638 | for (i = 1; i < SD_NODES_PER_DOMAIN; i++) { |
| 5639 | int next_node = find_next_best_node(node, used_nodes); |
| 5640 | nodemask = node_to_cpumask(next_node); |
| 5641 | cpus_or(span, span, nodemask); |
| 5642 | } |
| 5643 | |
| 5644 | return span; |
| 5645 | } |
| 5646 | #endif |
| 5647 | |
| 5648 | /* |
| 5649 | * At the moment, CONFIG_SCHED_SMT is never defined, but leave it in so we |
| 5650 | * can switch it on easily if needed. |
| 5651 | */ |
| 5652 | #ifdef CONFIG_SCHED_SMT |
| 5653 | static DEFINE_PER_CPU(struct sched_domain, cpu_domains); |
| 5654 | static struct sched_group sched_group_cpus[NR_CPUS]; |
| 5655 | static int cpu_to_cpu_group(int cpu) |
| 5656 | { |
| 5657 | return cpu; |
| 5658 | } |
| 5659 | #endif |
| 5660 | |
| 5661 | static DEFINE_PER_CPU(struct sched_domain, phys_domains); |
| 5662 | static struct sched_group sched_group_phys[NR_CPUS]; |
| 5663 | static int cpu_to_phys_group(int cpu) |
| 5664 | { |
| 5665 | #ifdef CONFIG_SCHED_SMT |
| 5666 | return first_cpu(cpu_sibling_map[cpu]); |
| 5667 | #else |
| 5668 | return cpu; |
| 5669 | #endif |
| 5670 | } |
| 5671 | |
| 5672 | #ifdef CONFIG_NUMA |
| 5673 | /* |
| 5674 | * The init_sched_build_groups can't handle what we want to do with node |
| 5675 | * groups, so roll our own. Now each node has its own list of groups which |
| 5676 | * gets dynamically allocated. |
| 5677 | */ |
| 5678 | static DEFINE_PER_CPU(struct sched_domain, node_domains); |
| 5679 | static struct sched_group **sched_group_nodes_bycpu[NR_CPUS]; |
| 5680 | |
| 5681 | static DEFINE_PER_CPU(struct sched_domain, allnodes_domains); |
| 5682 | static struct sched_group *sched_group_allnodes_bycpu[NR_CPUS]; |
| 5683 | |
| 5684 | static int cpu_to_allnodes_group(int cpu) |
| 5685 | { |
| 5686 | return cpu_to_node(cpu); |
| 5687 | } |
| 5688 | #endif |
| 5689 | |
| 5690 | /* |
| 5691 | * Build sched domains for a given set of cpus and attach the sched domains |
| 5692 | * to the individual cpus |
| 5693 | */ |
| 5694 | void build_sched_domains(const cpumask_t *cpu_map) |
| 5695 | { |
| 5696 | int i; |
| 5697 | #ifdef CONFIG_NUMA |
| 5698 | struct sched_group **sched_group_nodes = NULL; |
| 5699 | struct sched_group *sched_group_allnodes = NULL; |
| 5700 | |
| 5701 | /* |
| 5702 | * Allocate the per-node list of sched groups |
| 5703 | */ |
| 5704 | sched_group_nodes = kmalloc(sizeof(struct sched_group*)*MAX_NUMNODES, |
| 5705 | GFP_ATOMIC); |
| 5706 | if (!sched_group_nodes) { |
| 5707 | printk(KERN_WARNING "Can not alloc sched group node list\n"); |
| 5708 | return; |
| 5709 | } |
| 5710 | sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes; |
| 5711 | #endif |
| 5712 | |
| 5713 | /* |
| 5714 | * Set up domains for cpus specified by the cpu_map. |
| 5715 | */ |
| 5716 | for_each_cpu_mask(i, *cpu_map) { |
| 5717 | int group; |
| 5718 | struct sched_domain *sd = NULL, *p; |
| 5719 | cpumask_t nodemask = node_to_cpumask(cpu_to_node(i)); |
| 5720 | |
| 5721 | cpus_and(nodemask, nodemask, *cpu_map); |
| 5722 | |
| 5723 | #ifdef CONFIG_NUMA |
| 5724 | if (cpus_weight(*cpu_map) |
| 5725 | > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) { |
| 5726 | if (!sched_group_allnodes) { |
| 5727 | sched_group_allnodes |
| 5728 | = kmalloc(sizeof(struct sched_group) |
| 5729 | * MAX_NUMNODES, |
| 5730 | GFP_KERNEL); |
| 5731 | if (!sched_group_allnodes) { |
| 5732 | printk(KERN_WARNING |
| 5733 | "Can not alloc allnodes sched group\n"); |
| 5734 | break; |
| 5735 | } |
| 5736 | sched_group_allnodes_bycpu[i] |
| 5737 | = sched_group_allnodes; |
| 5738 | } |
| 5739 | sd = &per_cpu(allnodes_domains, i); |
| 5740 | *sd = SD_ALLNODES_INIT; |
| 5741 | sd->span = *cpu_map; |
| 5742 | group = cpu_to_allnodes_group(i); |
| 5743 | sd->groups = &sched_group_allnodes[group]; |
| 5744 | p = sd; |
| 5745 | } else |
| 5746 | p = NULL; |
| 5747 | |
| 5748 | sd = &per_cpu(node_domains, i); |
| 5749 | *sd = SD_NODE_INIT; |
| 5750 | sd->span = sched_domain_node_span(cpu_to_node(i)); |
| 5751 | sd->parent = p; |
| 5752 | cpus_and(sd->span, sd->span, *cpu_map); |
| 5753 | #endif |
| 5754 | |
| 5755 | p = sd; |
| 5756 | sd = &per_cpu(phys_domains, i); |
| 5757 | group = cpu_to_phys_group(i); |
| 5758 | *sd = SD_CPU_INIT; |
| 5759 | sd->span = nodemask; |
| 5760 | sd->parent = p; |
| 5761 | sd->groups = &sched_group_phys[group]; |
| 5762 | |
| 5763 | #ifdef CONFIG_SCHED_SMT |
| 5764 | p = sd; |
| 5765 | sd = &per_cpu(cpu_domains, i); |
| 5766 | group = cpu_to_cpu_group(i); |
| 5767 | *sd = SD_SIBLING_INIT; |
| 5768 | sd->span = cpu_sibling_map[i]; |
| 5769 | cpus_and(sd->span, sd->span, *cpu_map); |
| 5770 | sd->parent = p; |
| 5771 | sd->groups = &sched_group_cpus[group]; |
| 5772 | #endif |
| 5773 | } |
| 5774 | |
| 5775 | #ifdef CONFIG_SCHED_SMT |
| 5776 | /* Set up CPU (sibling) groups */ |
| 5777 | for_each_cpu_mask(i, *cpu_map) { |
| 5778 | cpumask_t this_sibling_map = cpu_sibling_map[i]; |
| 5779 | cpus_and(this_sibling_map, this_sibling_map, *cpu_map); |
| 5780 | if (i != first_cpu(this_sibling_map)) |
| 5781 | continue; |
| 5782 | |
| 5783 | init_sched_build_groups(sched_group_cpus, this_sibling_map, |
| 5784 | &cpu_to_cpu_group); |
| 5785 | } |
| 5786 | #endif |
| 5787 | |
| 5788 | /* Set up physical groups */ |
| 5789 | for (i = 0; i < MAX_NUMNODES; i++) { |
| 5790 | cpumask_t nodemask = node_to_cpumask(i); |
| 5791 | |
| 5792 | cpus_and(nodemask, nodemask, *cpu_map); |
| 5793 | if (cpus_empty(nodemask)) |
| 5794 | continue; |
| 5795 | |
| 5796 | init_sched_build_groups(sched_group_phys, nodemask, |
| 5797 | &cpu_to_phys_group); |
| 5798 | } |
| 5799 | |
| 5800 | #ifdef CONFIG_NUMA |
| 5801 | /* Set up node groups */ |
| 5802 | if (sched_group_allnodes) |
| 5803 | init_sched_build_groups(sched_group_allnodes, *cpu_map, |
| 5804 | &cpu_to_allnodes_group); |
| 5805 | |
| 5806 | for (i = 0; i < MAX_NUMNODES; i++) { |
| 5807 | /* Set up node groups */ |
| 5808 | struct sched_group *sg, *prev; |
| 5809 | cpumask_t nodemask = node_to_cpumask(i); |
| 5810 | cpumask_t domainspan; |
| 5811 | cpumask_t covered = CPU_MASK_NONE; |
| 5812 | int j; |
| 5813 | |
| 5814 | cpus_and(nodemask, nodemask, *cpu_map); |
| 5815 | if (cpus_empty(nodemask)) { |
| 5816 | sched_group_nodes[i] = NULL; |
| 5817 | continue; |
| 5818 | } |
| 5819 | |
| 5820 | domainspan = sched_domain_node_span(i); |
| 5821 | cpus_and(domainspan, domainspan, *cpu_map); |
| 5822 | |
| 5823 | sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL); |
| 5824 | sched_group_nodes[i] = sg; |
| 5825 | for_each_cpu_mask(j, nodemask) { |
| 5826 | struct sched_domain *sd; |
| 5827 | sd = &per_cpu(node_domains, j); |
| 5828 | sd->groups = sg; |
| 5829 | if (sd->groups == NULL) { |
| 5830 | /* Turn off balancing if we have no groups */ |
| 5831 | sd->flags = 0; |
| 5832 | } |
| 5833 | } |
| 5834 | if (!sg) { |
| 5835 | printk(KERN_WARNING |
| 5836 | "Can not alloc domain group for node %d\n", i); |
| 5837 | continue; |
| 5838 | } |
| 5839 | sg->cpu_power = 0; |
| 5840 | sg->cpumask = nodemask; |
| 5841 | cpus_or(covered, covered, nodemask); |
| 5842 | prev = sg; |
| 5843 | |
| 5844 | for (j = 0; j < MAX_NUMNODES; j++) { |
| 5845 | cpumask_t tmp, notcovered; |
| 5846 | int n = (i + j) % MAX_NUMNODES; |
| 5847 | |
| 5848 | cpus_complement(notcovered, covered); |
| 5849 | cpus_and(tmp, notcovered, *cpu_map); |
| 5850 | cpus_and(tmp, tmp, domainspan); |
| 5851 | if (cpus_empty(tmp)) |
| 5852 | break; |
| 5853 | |
| 5854 | nodemask = node_to_cpumask(n); |
| 5855 | cpus_and(tmp, tmp, nodemask); |
| 5856 | if (cpus_empty(tmp)) |
| 5857 | continue; |
| 5858 | |
| 5859 | sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL); |
| 5860 | if (!sg) { |
| 5861 | printk(KERN_WARNING |
| 5862 | "Can not alloc domain group for node %d\n", j); |
| 5863 | break; |
| 5864 | } |
| 5865 | sg->cpu_power = 0; |
| 5866 | sg->cpumask = tmp; |
| 5867 | cpus_or(covered, covered, tmp); |
| 5868 | prev->next = sg; |
| 5869 | prev = sg; |
| 5870 | } |
| 5871 | prev->next = sched_group_nodes[i]; |
| 5872 | } |
| 5873 | #endif |
| 5874 | |
| 5875 | /* Calculate CPU power for physical packages and nodes */ |
| 5876 | for_each_cpu_mask(i, *cpu_map) { |
| 5877 | int power; |
| 5878 | struct sched_domain *sd; |
| 5879 | #ifdef CONFIG_SCHED_SMT |
| 5880 | sd = &per_cpu(cpu_domains, i); |
| 5881 | power = SCHED_LOAD_SCALE; |
| 5882 | sd->groups->cpu_power = power; |
| 5883 | #endif |
| 5884 | |
| 5885 | sd = &per_cpu(phys_domains, i); |
| 5886 | power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE * |
| 5887 | (cpus_weight(sd->groups->cpumask)-1) / 10; |
| 5888 | sd->groups->cpu_power = power; |
| 5889 | |
| 5890 | #ifdef CONFIG_NUMA |
| 5891 | sd = &per_cpu(allnodes_domains, i); |
| 5892 | if (sd->groups) { |
| 5893 | power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE * |
| 5894 | (cpus_weight(sd->groups->cpumask)-1) / 10; |
| 5895 | sd->groups->cpu_power = power; |
| 5896 | } |
| 5897 | #endif |
| 5898 | } |
| 5899 | |
| 5900 | #ifdef CONFIG_NUMA |
| 5901 | for (i = 0; i < MAX_NUMNODES; i++) { |
| 5902 | struct sched_group *sg = sched_group_nodes[i]; |
| 5903 | int j; |
| 5904 | |
| 5905 | if (sg == NULL) |
| 5906 | continue; |
| 5907 | next_sg: |
| 5908 | for_each_cpu_mask(j, sg->cpumask) { |
| 5909 | struct sched_domain *sd; |
| 5910 | int power; |
| 5911 | |
| 5912 | sd = &per_cpu(phys_domains, j); |
| 5913 | if (j != first_cpu(sd->groups->cpumask)) { |
| 5914 | /* |
| 5915 | * Only add "power" once for each |
| 5916 | * physical package. |
| 5917 | */ |
| 5918 | continue; |
| 5919 | } |
| 5920 | power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE * |
| 5921 | (cpus_weight(sd->groups->cpumask)-1) / 10; |
| 5922 | |
| 5923 | sg->cpu_power += power; |
| 5924 | } |
| 5925 | sg = sg->next; |
| 5926 | if (sg != sched_group_nodes[i]) |
| 5927 | goto next_sg; |
| 5928 | } |
| 5929 | #endif |
| 5930 | |
| 5931 | /* Attach the domains */ |
| 5932 | for_each_cpu_mask(i, *cpu_map) { |
| 5933 | struct sched_domain *sd; |
| 5934 | #ifdef CONFIG_SCHED_SMT |
| 5935 | sd = &per_cpu(cpu_domains, i); |
| 5936 | #else |
| 5937 | sd = &per_cpu(phys_domains, i); |
| 5938 | #endif |
| 5939 | cpu_attach_domain(sd, i); |
| 5940 | } |
| 5941 | /* |
| 5942 | * Tune cache-hot values: |
| 5943 | */ |
| 5944 | calibrate_migration_costs(cpu_map); |
| 5945 | } |
| 5946 | /* |
| 5947 | * Set up scheduler domains and groups. Callers must hold the hotplug lock. |
| 5948 | */ |
| 5949 | static void arch_init_sched_domains(const cpumask_t *cpu_map) |
| 5950 | { |
| 5951 | cpumask_t cpu_default_map; |
| 5952 | |
| 5953 | /* |
| 5954 | * Setup mask for cpus without special case scheduling requirements. |
| 5955 | * For now this just excludes isolated cpus, but could be used to |
| 5956 | * exclude other special cases in the future. |
| 5957 | */ |
| 5958 | cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map); |
| 5959 | |
| 5960 | build_sched_domains(&cpu_default_map); |
| 5961 | } |
| 5962 | |
| 5963 | static void arch_destroy_sched_domains(const cpumask_t *cpu_map) |
| 5964 | { |
| 5965 | #ifdef CONFIG_NUMA |
| 5966 | int i; |
| 5967 | int cpu; |
| 5968 | |
| 5969 | for_each_cpu_mask(cpu, *cpu_map) { |
| 5970 | struct sched_group *sched_group_allnodes |
| 5971 | = sched_group_allnodes_bycpu[cpu]; |
| 5972 | struct sched_group **sched_group_nodes |
| 5973 | = sched_group_nodes_bycpu[cpu]; |
| 5974 | |
| 5975 | if (sched_group_allnodes) { |
| 5976 | kfree(sched_group_allnodes); |
| 5977 | sched_group_allnodes_bycpu[cpu] = NULL; |
| 5978 | } |
| 5979 | |
| 5980 | if (!sched_group_nodes) |
| 5981 | continue; |
| 5982 | |
| 5983 | for (i = 0; i < MAX_NUMNODES; i++) { |
| 5984 | cpumask_t nodemask = node_to_cpumask(i); |
| 5985 | struct sched_group *oldsg, *sg = sched_group_nodes[i]; |
| 5986 | |
| 5987 | cpus_and(nodemask, nodemask, *cpu_map); |
| 5988 | if (cpus_empty(nodemask)) |
| 5989 | continue; |
| 5990 | |
| 5991 | if (sg == NULL) |
| 5992 | continue; |
| 5993 | sg = sg->next; |
| 5994 | next_sg: |
| 5995 | oldsg = sg; |
| 5996 | sg = sg->next; |
| 5997 | kfree(oldsg); |
| 5998 | if (oldsg != sched_group_nodes[i]) |
| 5999 | goto next_sg; |
| 6000 | } |
| 6001 | kfree(sched_group_nodes); |
| 6002 | sched_group_nodes_bycpu[cpu] = NULL; |
| 6003 | } |
| 6004 | #endif |
| 6005 | } |
| 6006 | |
| 6007 | /* |
| 6008 | * Detach sched domains from a group of cpus specified in cpu_map |
| 6009 | * These cpus will now be attached to the NULL domain |
| 6010 | */ |
| 6011 | static void detach_destroy_domains(const cpumask_t *cpu_map) |
| 6012 | { |
| 6013 | int i; |
| 6014 | |
| 6015 | for_each_cpu_mask(i, *cpu_map) |
| 6016 | cpu_attach_domain(NULL, i); |
| 6017 | synchronize_sched(); |
| 6018 | arch_destroy_sched_domains(cpu_map); |
| 6019 | } |
| 6020 | |
| 6021 | /* |
| 6022 | * Partition sched domains as specified by the cpumasks below. |
| 6023 | * This attaches all cpus from the cpumasks to the NULL domain, |
| 6024 | * waits for a RCU quiescent period, recalculates sched |
| 6025 | * domain information and then attaches them back to the |
| 6026 | * correct sched domains |
| 6027 | * Call with hotplug lock held |
| 6028 | */ |
| 6029 | void partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2) |
| 6030 | { |
| 6031 | cpumask_t change_map; |
| 6032 | |
| 6033 | cpus_and(*partition1, *partition1, cpu_online_map); |
| 6034 | cpus_and(*partition2, *partition2, cpu_online_map); |
| 6035 | cpus_or(change_map, *partition1, *partition2); |
| 6036 | |
| 6037 | /* Detach sched domains from all of the affected cpus */ |
| 6038 | detach_destroy_domains(&change_map); |
| 6039 | if (!cpus_empty(*partition1)) |
| 6040 | build_sched_domains(partition1); |
| 6041 | if (!cpus_empty(*partition2)) |
| 6042 | build_sched_domains(partition2); |
| 6043 | } |
| 6044 | |
| 6045 | #ifdef CONFIG_HOTPLUG_CPU |
| 6046 | /* |
| 6047 | * Force a reinitialization of the sched domains hierarchy. The domains |
| 6048 | * and groups cannot be updated in place without racing with the balancing |
| 6049 | * code, so we temporarily attach all running cpus to the NULL domain |
| 6050 | * which will prevent rebalancing while the sched domains are recalculated. |
| 6051 | */ |
| 6052 | static int update_sched_domains(struct notifier_block *nfb, |
| 6053 | unsigned long action, void *hcpu) |
| 6054 | { |
| 6055 | switch (action) { |
| 6056 | case CPU_UP_PREPARE: |
| 6057 | case CPU_DOWN_PREPARE: |
| 6058 | detach_destroy_domains(&cpu_online_map); |
| 6059 | return NOTIFY_OK; |
| 6060 | |
| 6061 | case CPU_UP_CANCELED: |
| 6062 | case CPU_DOWN_FAILED: |
| 6063 | case CPU_ONLINE: |
| 6064 | case CPU_DEAD: |
| 6065 | /* |
| 6066 | * Fall through and re-initialise the domains. |
| 6067 | */ |
| 6068 | break; |
| 6069 | default: |
| 6070 | return NOTIFY_DONE; |
| 6071 | } |
| 6072 | |
| 6073 | /* The hotplug lock is already held by cpu_up/cpu_down */ |
| 6074 | arch_init_sched_domains(&cpu_online_map); |
| 6075 | |
| 6076 | return NOTIFY_OK; |
| 6077 | } |
| 6078 | #endif |
| 6079 | |
| 6080 | void __init sched_init_smp(void) |
| 6081 | { |
| 6082 | lock_cpu_hotplug(); |
| 6083 | arch_init_sched_domains(&cpu_online_map); |
| 6084 | unlock_cpu_hotplug(); |
| 6085 | /* XXX: Theoretical race here - CPU may be hotplugged now */ |
| 6086 | hotcpu_notifier(update_sched_domains, 0); |
| 6087 | } |
| 6088 | #else |
| 6089 | void __init sched_init_smp(void) |
| 6090 | { |
| 6091 | } |
| 6092 | #endif /* CONFIG_SMP */ |
| 6093 | |
| 6094 | int in_sched_functions(unsigned long addr) |
| 6095 | { |
| 6096 | /* Linker adds these: start and end of __sched functions */ |
| 6097 | extern char __sched_text_start[], __sched_text_end[]; |
| 6098 | return in_lock_functions(addr) || |
| 6099 | (addr >= (unsigned long)__sched_text_start |
| 6100 | && addr < (unsigned long)__sched_text_end); |
| 6101 | } |
| 6102 | |
| 6103 | void __init sched_init(void) |
| 6104 | { |
| 6105 | runqueue_t *rq; |
| 6106 | int i, j, k; |
| 6107 | |
| 6108 | for (i = 0; i < NR_CPUS; i++) { |
| 6109 | prio_array_t *array; |
| 6110 | |
| 6111 | rq = cpu_rq(i); |
| 6112 | spin_lock_init(&rq->lock); |
| 6113 | rq->nr_running = 0; |
| 6114 | rq->active = rq->arrays; |
| 6115 | rq->expired = rq->arrays + 1; |
| 6116 | rq->best_expired_prio = MAX_PRIO; |
| 6117 | |
| 6118 | #ifdef CONFIG_SMP |
| 6119 | rq->sd = NULL; |
| 6120 | for (j = 1; j < 3; j++) |
| 6121 | rq->cpu_load[j] = 0; |
| 6122 | rq->active_balance = 0; |
| 6123 | rq->push_cpu = 0; |
| 6124 | rq->migration_thread = NULL; |
| 6125 | INIT_LIST_HEAD(&rq->migration_queue); |
| 6126 | #endif |
| 6127 | atomic_set(&rq->nr_iowait, 0); |
| 6128 | |
| 6129 | for (j = 0; j < 2; j++) { |
| 6130 | array = rq->arrays + j; |
| 6131 | for (k = 0; k < MAX_PRIO; k++) { |
| 6132 | INIT_LIST_HEAD(array->queue + k); |
| 6133 | __clear_bit(k, array->bitmap); |
| 6134 | } |
| 6135 | // delimiter for bitsearch |
| 6136 | __set_bit(MAX_PRIO, array->bitmap); |
| 6137 | } |
| 6138 | } |
| 6139 | |
| 6140 | /* |
| 6141 | * The boot idle thread does lazy MMU switching as well: |
| 6142 | */ |
| 6143 | atomic_inc(&init_mm.mm_count); |
| 6144 | enter_lazy_tlb(&init_mm, current); |
| 6145 | |
| 6146 | /* |
| 6147 | * Make us the idle thread. Technically, schedule() should not be |
| 6148 | * called from this thread, however somewhere below it might be, |
| 6149 | * but because we are the idle thread, we just pick up running again |
| 6150 | * when this runqueue becomes "idle". |
| 6151 | */ |
| 6152 | init_idle(current, smp_processor_id()); |
| 6153 | } |
| 6154 | |
| 6155 | #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP |
| 6156 | void __might_sleep(char *file, int line) |
| 6157 | { |
| 6158 | #if defined(in_atomic) |
| 6159 | static unsigned long prev_jiffy; /* ratelimiting */ |
| 6160 | |
| 6161 | if ((in_atomic() || irqs_disabled()) && |
| 6162 | system_state == SYSTEM_RUNNING && !oops_in_progress) { |
| 6163 | if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) |
| 6164 | return; |
| 6165 | prev_jiffy = jiffies; |
| 6166 | printk(KERN_ERR "Debug: sleeping function called from invalid" |
| 6167 | " context at %s:%d\n", file, line); |
| 6168 | printk("in_atomic():%d, irqs_disabled():%d\n", |
| 6169 | in_atomic(), irqs_disabled()); |
| 6170 | dump_stack(); |
| 6171 | } |
| 6172 | #endif |
| 6173 | } |
| 6174 | EXPORT_SYMBOL(__might_sleep); |
| 6175 | #endif |
| 6176 | |
| 6177 | #ifdef CONFIG_MAGIC_SYSRQ |
| 6178 | void normalize_rt_tasks(void) |
| 6179 | { |
| 6180 | struct task_struct *p; |
| 6181 | prio_array_t *array; |
| 6182 | unsigned long flags; |
| 6183 | runqueue_t *rq; |
| 6184 | |
| 6185 | read_lock_irq(&tasklist_lock); |
| 6186 | for_each_process (p) { |
| 6187 | if (!rt_task(p)) |
| 6188 | continue; |
| 6189 | |
| 6190 | rq = task_rq_lock(p, &flags); |
| 6191 | |
| 6192 | array = p->array; |
| 6193 | if (array) |
| 6194 | deactivate_task(p, task_rq(p)); |
| 6195 | __setscheduler(p, SCHED_NORMAL, 0); |
| 6196 | if (array) { |
| 6197 | __activate_task(p, task_rq(p)); |
| 6198 | resched_task(rq->curr); |
| 6199 | } |
| 6200 | |
| 6201 | task_rq_unlock(rq, &flags); |
| 6202 | } |
| 6203 | read_unlock_irq(&tasklist_lock); |
| 6204 | } |
| 6205 | |
| 6206 | #endif /* CONFIG_MAGIC_SYSRQ */ |
| 6207 | |
| 6208 | #ifdef CONFIG_IA64 |
| 6209 | /* |
| 6210 | * These functions are only useful for the IA64 MCA handling. |
| 6211 | * |
| 6212 | * They can only be called when the whole system has been |
| 6213 | * stopped - every CPU needs to be quiescent, and no scheduling |
| 6214 | * activity can take place. Using them for anything else would |
| 6215 | * be a serious bug, and as a result, they aren't even visible |
| 6216 | * under any other configuration. |
| 6217 | */ |
| 6218 | |
| 6219 | /** |
| 6220 | * curr_task - return the current task for a given cpu. |
| 6221 | * @cpu: the processor in question. |
| 6222 | * |
| 6223 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! |
| 6224 | */ |
| 6225 | task_t *curr_task(int cpu) |
| 6226 | { |
| 6227 | return cpu_curr(cpu); |
| 6228 | } |
| 6229 | |
| 6230 | /** |
| 6231 | * set_curr_task - set the current task for a given cpu. |
| 6232 | * @cpu: the processor in question. |
| 6233 | * @p: the task pointer to set. |
| 6234 | * |
| 6235 | * Description: This function must only be used when non-maskable interrupts |
| 6236 | * are serviced on a separate stack. It allows the architecture to switch the |
| 6237 | * notion of the current task on a cpu in a non-blocking manner. This function |
| 6238 | * must be called with all CPU's synchronized, and interrupts disabled, the |
| 6239 | * and caller must save the original value of the current task (see |
| 6240 | * curr_task() above) and restore that value before reenabling interrupts and |
| 6241 | * re-starting the system. |
| 6242 | * |
| 6243 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! |
| 6244 | */ |
| 6245 | void set_curr_task(int cpu, task_t *p) |
| 6246 | { |
| 6247 | cpu_curr(cpu) = p; |
| 6248 | } |
| 6249 | |
| 6250 | #endif |