Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 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> |
Randy.Dunlap | c59ede7 | 2006-01-11 12:17:46 -0800 | [diff] [blame] | 30 | #include <linux/capability.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 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> |
akpm@osdl.org | 198e2f1 | 2006-01-12 01:05:30 -0800 | [diff] [blame] | 37 | #include <linux/vmalloc.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 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> |
bibo mao | c6fd91f | 2006-03-26 01:38:20 -0800 | [diff] [blame] | 52 | #include <linux/kprobes.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 53 | #include <asm/tlb.h> |
| 54 | |
| 55 | #include <asm/unistd.h> |
| 56 | |
| 57 | /* |
| 58 | * Convert user-nice values [ -20 ... 0 ... 19 ] |
| 59 | * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ], |
| 60 | * and back. |
| 61 | */ |
| 62 | #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20) |
| 63 | #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20) |
| 64 | #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio) |
| 65 | |
| 66 | /* |
| 67 | * 'User priority' is the nice value converted to something we |
| 68 | * can work with better when scaling various scheduler parameters, |
| 69 | * it's a [ 0 ... 39 ] range. |
| 70 | */ |
| 71 | #define USER_PRIO(p) ((p)-MAX_RT_PRIO) |
| 72 | #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio) |
| 73 | #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO)) |
| 74 | |
| 75 | /* |
| 76 | * Some helpers for converting nanosecond timing to jiffy resolution |
| 77 | */ |
| 78 | #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ)) |
| 79 | #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ)) |
| 80 | |
| 81 | /* |
| 82 | * These are the 'tuning knobs' of the scheduler: |
| 83 | * |
| 84 | * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger), |
| 85 | * default timeslice is 100 msecs, maximum timeslice is 800 msecs. |
| 86 | * Timeslices get refilled after they expire. |
| 87 | */ |
| 88 | #define MIN_TIMESLICE max(5 * HZ / 1000, 1) |
| 89 | #define DEF_TIMESLICE (100 * HZ / 1000) |
| 90 | #define ON_RUNQUEUE_WEIGHT 30 |
| 91 | #define CHILD_PENALTY 95 |
| 92 | #define PARENT_PENALTY 100 |
| 93 | #define EXIT_WEIGHT 3 |
| 94 | #define PRIO_BONUS_RATIO 25 |
| 95 | #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100) |
| 96 | #define INTERACTIVE_DELTA 2 |
| 97 | #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS) |
| 98 | #define STARVATION_LIMIT (MAX_SLEEP_AVG) |
| 99 | #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG)) |
| 100 | |
| 101 | /* |
| 102 | * If a task is 'interactive' then we reinsert it in the active |
| 103 | * array after it has expired its current timeslice. (it will not |
| 104 | * continue to run immediately, it will still roundrobin with |
| 105 | * other interactive tasks.) |
| 106 | * |
| 107 | * This part scales the interactivity limit depending on niceness. |
| 108 | * |
| 109 | * We scale it linearly, offset by the INTERACTIVE_DELTA delta. |
| 110 | * Here are a few examples of different nice levels: |
| 111 | * |
| 112 | * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0] |
| 113 | * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0] |
| 114 | * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0] |
| 115 | * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0] |
| 116 | * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0] |
| 117 | * |
| 118 | * (the X axis represents the possible -5 ... 0 ... +5 dynamic |
| 119 | * priority range a task can explore, a value of '1' means the |
| 120 | * task is rated interactive.) |
| 121 | * |
| 122 | * Ie. nice +19 tasks can never get 'interactive' enough to be |
| 123 | * reinserted into the active array. And only heavily CPU-hog nice -20 |
| 124 | * tasks will be expired. Default nice 0 tasks are somewhere between, |
| 125 | * it takes some effort for them to get interactive, but it's not |
| 126 | * too hard. |
| 127 | */ |
| 128 | |
| 129 | #define CURRENT_BONUS(p) \ |
| 130 | (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \ |
| 131 | MAX_SLEEP_AVG) |
| 132 | |
| 133 | #define GRANULARITY (10 * HZ / 1000 ? : 1) |
| 134 | |
| 135 | #ifdef CONFIG_SMP |
| 136 | #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \ |
| 137 | (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \ |
| 138 | num_online_cpus()) |
| 139 | #else |
| 140 | #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \ |
| 141 | (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1))) |
| 142 | #endif |
| 143 | |
| 144 | #define SCALE(v1,v1_max,v2_max) \ |
| 145 | (v1) * (v2_max) / (v1_max) |
| 146 | |
| 147 | #define DELTA(p) \ |
Martin Andersson | 013d386 | 2006-03-27 01:15:18 -0800 | [diff] [blame] | 148 | (SCALE(TASK_NICE(p) + 20, 40, MAX_BONUS) - 20 * MAX_BONUS / 40 + \ |
| 149 | INTERACTIVE_DELTA) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 150 | |
| 151 | #define TASK_INTERACTIVE(p) \ |
| 152 | ((p)->prio <= (p)->static_prio - DELTA(p)) |
| 153 | |
| 154 | #define INTERACTIVE_SLEEP(p) \ |
| 155 | (JIFFIES_TO_NS(MAX_SLEEP_AVG * \ |
| 156 | (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1)) |
| 157 | |
| 158 | #define TASK_PREEMPTS_CURR(p, rq) \ |
| 159 | ((p)->prio < (rq)->curr->prio) |
| 160 | |
| 161 | /* |
| 162 | * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ] |
| 163 | * to time slice values: [800ms ... 100ms ... 5ms] |
| 164 | * |
| 165 | * The higher a thread's priority, the bigger timeslices |
| 166 | * it gets during one round of execution. But even the lowest |
| 167 | * priority thread gets MIN_TIMESLICE worth of execution time. |
| 168 | */ |
| 169 | |
| 170 | #define SCALE_PRIO(x, prio) \ |
| 171 | max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO/2), MIN_TIMESLICE) |
| 172 | |
Ingo Molnar | 48c08d3 | 2005-06-25 14:57:22 -0700 | [diff] [blame] | 173 | static unsigned int task_timeslice(task_t *p) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 174 | { |
| 175 | if (p->static_prio < NICE_TO_PRIO(0)) |
| 176 | return SCALE_PRIO(DEF_TIMESLICE*4, p->static_prio); |
| 177 | else |
| 178 | return SCALE_PRIO(DEF_TIMESLICE, p->static_prio); |
| 179 | } |
| 180 | #define task_hot(p, now, sd) ((long long) ((now) - (p)->last_ran) \ |
| 181 | < (long long) (sd)->cache_hot_time) |
| 182 | |
| 183 | /* |
| 184 | * These are the runqueue data structures: |
| 185 | */ |
| 186 | |
| 187 | #define BITMAP_SIZE ((((MAX_PRIO+1+7)/8)+sizeof(long)-1)/sizeof(long)) |
| 188 | |
| 189 | typedef struct runqueue runqueue_t; |
| 190 | |
| 191 | struct prio_array { |
| 192 | unsigned int nr_active; |
| 193 | unsigned long bitmap[BITMAP_SIZE]; |
| 194 | struct list_head queue[MAX_PRIO]; |
| 195 | }; |
| 196 | |
| 197 | /* |
| 198 | * This is the main, per-CPU runqueue data structure. |
| 199 | * |
| 200 | * Locking rule: those places that want to lock multiple runqueues |
| 201 | * (such as the load balancing or the thread migration code), lock |
| 202 | * acquire operations must be ordered by ascending &runqueue. |
| 203 | */ |
| 204 | struct runqueue { |
| 205 | spinlock_t lock; |
| 206 | |
| 207 | /* |
| 208 | * nr_running and cpu_load should be in the same cacheline because |
| 209 | * remote CPUs use both these fields when doing load calculation. |
| 210 | */ |
| 211 | unsigned long nr_running; |
| 212 | #ifdef CONFIG_SMP |
Nick Piggin | 7897986 | 2005-06-25 14:57:13 -0700 | [diff] [blame] | 213 | unsigned long cpu_load[3]; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 214 | #endif |
| 215 | unsigned long long nr_switches; |
| 216 | |
| 217 | /* |
| 218 | * This is part of a global counter where only the total sum |
| 219 | * over all CPUs matters. A task can increase this counter on |
| 220 | * one CPU and if it got migrated afterwards it may decrease |
| 221 | * it on another CPU. Always updated under the runqueue lock: |
| 222 | */ |
| 223 | unsigned long nr_uninterruptible; |
| 224 | |
| 225 | unsigned long expired_timestamp; |
| 226 | unsigned long long timestamp_last_tick; |
| 227 | task_t *curr, *idle; |
| 228 | struct mm_struct *prev_mm; |
| 229 | prio_array_t *active, *expired, arrays[2]; |
| 230 | int best_expired_prio; |
| 231 | atomic_t nr_iowait; |
| 232 | |
| 233 | #ifdef CONFIG_SMP |
| 234 | struct sched_domain *sd; |
| 235 | |
| 236 | /* For active balancing */ |
| 237 | int active_balance; |
| 238 | int push_cpu; |
| 239 | |
| 240 | task_t *migration_thread; |
| 241 | struct list_head migration_queue; |
Anton Blanchard | e9028b0 | 2006-03-23 02:59:20 -0800 | [diff] [blame] | 242 | int cpu; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 243 | #endif |
| 244 | |
| 245 | #ifdef CONFIG_SCHEDSTATS |
| 246 | /* latency stats */ |
| 247 | struct sched_info rq_sched_info; |
| 248 | |
| 249 | /* sys_sched_yield() stats */ |
| 250 | unsigned long yld_exp_empty; |
| 251 | unsigned long yld_act_empty; |
| 252 | unsigned long yld_both_empty; |
| 253 | unsigned long yld_cnt; |
| 254 | |
| 255 | /* schedule() stats */ |
| 256 | unsigned long sched_switch; |
| 257 | unsigned long sched_cnt; |
| 258 | unsigned long sched_goidle; |
| 259 | |
| 260 | /* try_to_wake_up() stats */ |
| 261 | unsigned long ttwu_cnt; |
| 262 | unsigned long ttwu_local; |
| 263 | #endif |
| 264 | }; |
| 265 | |
| 266 | static DEFINE_PER_CPU(struct runqueue, runqueues); |
| 267 | |
Nick Piggin | 674311d | 2005-06-25 14:57:27 -0700 | [diff] [blame] | 268 | /* |
| 269 | * The domain tree (rq->sd) is protected by RCU's quiescent state transition. |
Dinakar Guniguntala | 1a20ff2 | 2005-06-25 14:57:33 -0700 | [diff] [blame] | 270 | * See detach_destroy_domains: synchronize_sched for details. |
Nick Piggin | 674311d | 2005-06-25 14:57:27 -0700 | [diff] [blame] | 271 | * |
| 272 | * The domain tree of any CPU may only be accessed from within |
| 273 | * preempt-disabled sections. |
| 274 | */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 275 | #define for_each_domain(cpu, domain) \ |
Nick Piggin | 674311d | 2005-06-25 14:57:27 -0700 | [diff] [blame] | 276 | for (domain = rcu_dereference(cpu_rq(cpu)->sd); domain; domain = domain->parent) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 277 | |
| 278 | #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu))) |
| 279 | #define this_rq() (&__get_cpu_var(runqueues)) |
| 280 | #define task_rq(p) cpu_rq(task_cpu(p)) |
| 281 | #define cpu_curr(cpu) (cpu_rq(cpu)->curr) |
| 282 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 283 | #ifndef prepare_arch_switch |
Nick Piggin | 4866cde | 2005-06-25 14:57:23 -0700 | [diff] [blame] | 284 | # define prepare_arch_switch(next) do { } while (0) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 285 | #endif |
Nick Piggin | 4866cde | 2005-06-25 14:57:23 -0700 | [diff] [blame] | 286 | #ifndef finish_arch_switch |
| 287 | # define finish_arch_switch(prev) do { } while (0) |
| 288 | #endif |
| 289 | |
| 290 | #ifndef __ARCH_WANT_UNLOCKED_CTXSW |
| 291 | static inline int task_running(runqueue_t *rq, task_t *p) |
| 292 | { |
| 293 | return rq->curr == p; |
| 294 | } |
| 295 | |
| 296 | static inline void prepare_lock_switch(runqueue_t *rq, task_t *next) |
| 297 | { |
| 298 | } |
| 299 | |
| 300 | static inline void finish_lock_switch(runqueue_t *rq, task_t *prev) |
| 301 | { |
Ingo Molnar | da04c03 | 2005-09-13 11:17:59 +0200 | [diff] [blame] | 302 | #ifdef CONFIG_DEBUG_SPINLOCK |
| 303 | /* this is a valid case when another task releases the spinlock */ |
| 304 | rq->lock.owner = current; |
| 305 | #endif |
Nick Piggin | 4866cde | 2005-06-25 14:57:23 -0700 | [diff] [blame] | 306 | spin_unlock_irq(&rq->lock); |
| 307 | } |
| 308 | |
| 309 | #else /* __ARCH_WANT_UNLOCKED_CTXSW */ |
| 310 | static inline int task_running(runqueue_t *rq, task_t *p) |
| 311 | { |
| 312 | #ifdef CONFIG_SMP |
| 313 | return p->oncpu; |
| 314 | #else |
| 315 | return rq->curr == p; |
| 316 | #endif |
| 317 | } |
| 318 | |
| 319 | static inline void prepare_lock_switch(runqueue_t *rq, task_t *next) |
| 320 | { |
| 321 | #ifdef CONFIG_SMP |
| 322 | /* |
| 323 | * We can optimise this out completely for !SMP, because the |
| 324 | * SMP rebalancing from interrupt is the only thing that cares |
| 325 | * here. |
| 326 | */ |
| 327 | next->oncpu = 1; |
| 328 | #endif |
| 329 | #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW |
| 330 | spin_unlock_irq(&rq->lock); |
| 331 | #else |
| 332 | spin_unlock(&rq->lock); |
| 333 | #endif |
| 334 | } |
| 335 | |
| 336 | static inline void finish_lock_switch(runqueue_t *rq, task_t *prev) |
| 337 | { |
| 338 | #ifdef CONFIG_SMP |
| 339 | /* |
| 340 | * After ->oncpu is cleared, the task can be moved to a different CPU. |
| 341 | * We must ensure this doesn't happen until the switch is completely |
| 342 | * finished. |
| 343 | */ |
| 344 | smp_wmb(); |
| 345 | prev->oncpu = 0; |
| 346 | #endif |
| 347 | #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW |
| 348 | local_irq_enable(); |
| 349 | #endif |
| 350 | } |
| 351 | #endif /* __ARCH_WANT_UNLOCKED_CTXSW */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 352 | |
| 353 | /* |
| 354 | * task_rq_lock - lock the runqueue a given task resides on and disable |
| 355 | * interrupts. Note the ordering: we can safely lookup the task_rq without |
| 356 | * explicitly disabling preemption. |
| 357 | */ |
| 358 | static inline runqueue_t *task_rq_lock(task_t *p, unsigned long *flags) |
| 359 | __acquires(rq->lock) |
| 360 | { |
| 361 | struct runqueue *rq; |
| 362 | |
| 363 | repeat_lock_task: |
| 364 | local_irq_save(*flags); |
| 365 | rq = task_rq(p); |
| 366 | spin_lock(&rq->lock); |
| 367 | if (unlikely(rq != task_rq(p))) { |
| 368 | spin_unlock_irqrestore(&rq->lock, *flags); |
| 369 | goto repeat_lock_task; |
| 370 | } |
| 371 | return rq; |
| 372 | } |
| 373 | |
| 374 | static inline void task_rq_unlock(runqueue_t *rq, unsigned long *flags) |
| 375 | __releases(rq->lock) |
| 376 | { |
| 377 | spin_unlock_irqrestore(&rq->lock, *flags); |
| 378 | } |
| 379 | |
| 380 | #ifdef CONFIG_SCHEDSTATS |
| 381 | /* |
| 382 | * bump this up when changing the output format or the meaning of an existing |
| 383 | * format, so that tools can adapt (or abort) |
| 384 | */ |
Nick Piggin | 68767a0 | 2005-06-25 14:57:20 -0700 | [diff] [blame] | 385 | #define SCHEDSTAT_VERSION 12 |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 386 | |
| 387 | static int show_schedstat(struct seq_file *seq, void *v) |
| 388 | { |
| 389 | int cpu; |
| 390 | |
| 391 | seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION); |
| 392 | seq_printf(seq, "timestamp %lu\n", jiffies); |
| 393 | for_each_online_cpu(cpu) { |
| 394 | runqueue_t *rq = cpu_rq(cpu); |
| 395 | #ifdef CONFIG_SMP |
| 396 | struct sched_domain *sd; |
| 397 | int dcnt = 0; |
| 398 | #endif |
| 399 | |
| 400 | /* runqueue-specific stats */ |
| 401 | seq_printf(seq, |
| 402 | "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu", |
| 403 | cpu, rq->yld_both_empty, |
| 404 | rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt, |
| 405 | rq->sched_switch, rq->sched_cnt, rq->sched_goidle, |
| 406 | rq->ttwu_cnt, rq->ttwu_local, |
| 407 | rq->rq_sched_info.cpu_time, |
| 408 | rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt); |
| 409 | |
| 410 | seq_printf(seq, "\n"); |
| 411 | |
| 412 | #ifdef CONFIG_SMP |
| 413 | /* domain-specific stats */ |
Nick Piggin | 674311d | 2005-06-25 14:57:27 -0700 | [diff] [blame] | 414 | preempt_disable(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 415 | for_each_domain(cpu, sd) { |
| 416 | enum idle_type itype; |
| 417 | char mask_str[NR_CPUS]; |
| 418 | |
| 419 | cpumask_scnprintf(mask_str, NR_CPUS, sd->span); |
| 420 | seq_printf(seq, "domain%d %s", dcnt++, mask_str); |
| 421 | for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES; |
| 422 | itype++) { |
| 423 | seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu", |
| 424 | sd->lb_cnt[itype], |
| 425 | sd->lb_balanced[itype], |
| 426 | sd->lb_failed[itype], |
| 427 | sd->lb_imbalance[itype], |
| 428 | sd->lb_gained[itype], |
| 429 | sd->lb_hot_gained[itype], |
| 430 | sd->lb_nobusyq[itype], |
| 431 | sd->lb_nobusyg[itype]); |
| 432 | } |
Nick Piggin | 68767a0 | 2005-06-25 14:57:20 -0700 | [diff] [blame] | 433 | seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n", |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 434 | sd->alb_cnt, sd->alb_failed, sd->alb_pushed, |
Nick Piggin | 68767a0 | 2005-06-25 14:57:20 -0700 | [diff] [blame] | 435 | sd->sbe_cnt, sd->sbe_balanced, sd->sbe_pushed, |
| 436 | sd->sbf_cnt, sd->sbf_balanced, sd->sbf_pushed, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 437 | sd->ttwu_wake_remote, sd->ttwu_move_affine, sd->ttwu_move_balance); |
| 438 | } |
Nick Piggin | 674311d | 2005-06-25 14:57:27 -0700 | [diff] [blame] | 439 | preempt_enable(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 440 | #endif |
| 441 | } |
| 442 | return 0; |
| 443 | } |
| 444 | |
| 445 | static int schedstat_open(struct inode *inode, struct file *file) |
| 446 | { |
| 447 | unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32); |
| 448 | char *buf = kmalloc(size, GFP_KERNEL); |
| 449 | struct seq_file *m; |
| 450 | int res; |
| 451 | |
| 452 | if (!buf) |
| 453 | return -ENOMEM; |
| 454 | res = single_open(file, show_schedstat, NULL); |
| 455 | if (!res) { |
| 456 | m = file->private_data; |
| 457 | m->buf = buf; |
| 458 | m->size = size; |
| 459 | } else |
| 460 | kfree(buf); |
| 461 | return res; |
| 462 | } |
| 463 | |
| 464 | struct file_operations proc_schedstat_operations = { |
| 465 | .open = schedstat_open, |
| 466 | .read = seq_read, |
| 467 | .llseek = seq_lseek, |
| 468 | .release = single_release, |
| 469 | }; |
| 470 | |
| 471 | # define schedstat_inc(rq, field) do { (rq)->field++; } while (0) |
| 472 | # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0) |
| 473 | #else /* !CONFIG_SCHEDSTATS */ |
| 474 | # define schedstat_inc(rq, field) do { } while (0) |
| 475 | # define schedstat_add(rq, field, amt) do { } while (0) |
| 476 | #endif |
| 477 | |
| 478 | /* |
| 479 | * rq_lock - lock a given runqueue and disable interrupts. |
| 480 | */ |
| 481 | static inline runqueue_t *this_rq_lock(void) |
| 482 | __acquires(rq->lock) |
| 483 | { |
| 484 | runqueue_t *rq; |
| 485 | |
| 486 | local_irq_disable(); |
| 487 | rq = this_rq(); |
| 488 | spin_lock(&rq->lock); |
| 489 | |
| 490 | return rq; |
| 491 | } |
| 492 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 493 | #ifdef CONFIG_SCHEDSTATS |
| 494 | /* |
| 495 | * Called when a process is dequeued from the active array and given |
| 496 | * the cpu. We should note that with the exception of interactive |
| 497 | * tasks, the expired queue will become the active queue after the active |
| 498 | * queue is empty, without explicitly dequeuing and requeuing tasks in the |
| 499 | * expired queue. (Interactive tasks may be requeued directly to the |
| 500 | * active queue, thus delaying tasks in the expired queue from running; |
| 501 | * see scheduler_tick()). |
| 502 | * |
| 503 | * This function is only called from sched_info_arrive(), rather than |
| 504 | * dequeue_task(). Even though a task may be queued and dequeued multiple |
| 505 | * times as it is shuffled about, we're really interested in knowing how |
| 506 | * long it was from the *first* time it was queued to the time that it |
| 507 | * finally hit a cpu. |
| 508 | */ |
| 509 | static inline void sched_info_dequeued(task_t *t) |
| 510 | { |
| 511 | t->sched_info.last_queued = 0; |
| 512 | } |
| 513 | |
| 514 | /* |
| 515 | * Called when a task finally hits the cpu. We can now calculate how |
| 516 | * long it was waiting to run. We also note when it began so that we |
| 517 | * can keep stats on how long its timeslice is. |
| 518 | */ |
Arjan van de Ven | 858119e | 2006-01-14 13:20:43 -0800 | [diff] [blame] | 519 | static void sched_info_arrive(task_t *t) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 520 | { |
| 521 | unsigned long now = jiffies, diff = 0; |
| 522 | struct runqueue *rq = task_rq(t); |
| 523 | |
| 524 | if (t->sched_info.last_queued) |
| 525 | diff = now - t->sched_info.last_queued; |
| 526 | sched_info_dequeued(t); |
| 527 | t->sched_info.run_delay += diff; |
| 528 | t->sched_info.last_arrival = now; |
| 529 | t->sched_info.pcnt++; |
| 530 | |
| 531 | if (!rq) |
| 532 | return; |
| 533 | |
| 534 | rq->rq_sched_info.run_delay += diff; |
| 535 | rq->rq_sched_info.pcnt++; |
| 536 | } |
| 537 | |
| 538 | /* |
| 539 | * Called when a process is queued into either the active or expired |
| 540 | * array. The time is noted and later used to determine how long we |
| 541 | * had to wait for us to reach the cpu. Since the expired queue will |
| 542 | * become the active queue after active queue is empty, without dequeuing |
| 543 | * and requeuing any tasks, we are interested in queuing to either. It |
| 544 | * is unusual but not impossible for tasks to be dequeued and immediately |
| 545 | * requeued in the same or another array: this can happen in sched_yield(), |
| 546 | * set_user_nice(), and even load_balance() as it moves tasks from runqueue |
| 547 | * to runqueue. |
| 548 | * |
| 549 | * This function is only called from enqueue_task(), but also only updates |
| 550 | * the timestamp if it is already not set. It's assumed that |
| 551 | * sched_info_dequeued() will clear that stamp when appropriate. |
| 552 | */ |
| 553 | static inline void sched_info_queued(task_t *t) |
| 554 | { |
| 555 | if (!t->sched_info.last_queued) |
| 556 | t->sched_info.last_queued = jiffies; |
| 557 | } |
| 558 | |
| 559 | /* |
| 560 | * Called when a process ceases being the active-running process, either |
| 561 | * voluntarily or involuntarily. Now we can calculate how long we ran. |
| 562 | */ |
| 563 | static inline void sched_info_depart(task_t *t) |
| 564 | { |
| 565 | struct runqueue *rq = task_rq(t); |
| 566 | unsigned long diff = jiffies - t->sched_info.last_arrival; |
| 567 | |
| 568 | t->sched_info.cpu_time += diff; |
| 569 | |
| 570 | if (rq) |
| 571 | rq->rq_sched_info.cpu_time += diff; |
| 572 | } |
| 573 | |
| 574 | /* |
| 575 | * Called when tasks are switched involuntarily due, typically, to expiring |
| 576 | * their time slice. (This may also be called when switching to or from |
| 577 | * the idle task.) We are only called when prev != next. |
| 578 | */ |
| 579 | static inline void sched_info_switch(task_t *prev, task_t *next) |
| 580 | { |
| 581 | struct runqueue *rq = task_rq(prev); |
| 582 | |
| 583 | /* |
| 584 | * prev now departs the cpu. It's not interesting to record |
| 585 | * stats about how efficient we were at scheduling the idle |
| 586 | * process, however. |
| 587 | */ |
| 588 | if (prev != rq->idle) |
| 589 | sched_info_depart(prev); |
| 590 | |
| 591 | if (next != rq->idle) |
| 592 | sched_info_arrive(next); |
| 593 | } |
| 594 | #else |
| 595 | #define sched_info_queued(t) do { } while (0) |
| 596 | #define sched_info_switch(t, next) do { } while (0) |
| 597 | #endif /* CONFIG_SCHEDSTATS */ |
| 598 | |
| 599 | /* |
| 600 | * Adding/removing a task to/from a priority array: |
| 601 | */ |
| 602 | static void dequeue_task(struct task_struct *p, prio_array_t *array) |
| 603 | { |
| 604 | array->nr_active--; |
| 605 | list_del(&p->run_list); |
| 606 | if (list_empty(array->queue + p->prio)) |
| 607 | __clear_bit(p->prio, array->bitmap); |
| 608 | } |
| 609 | |
| 610 | static void enqueue_task(struct task_struct *p, prio_array_t *array) |
| 611 | { |
| 612 | sched_info_queued(p); |
| 613 | list_add_tail(&p->run_list, array->queue + p->prio); |
| 614 | __set_bit(p->prio, array->bitmap); |
| 615 | array->nr_active++; |
| 616 | p->array = array; |
| 617 | } |
| 618 | |
| 619 | /* |
| 620 | * Put task to the end of the run list without the overhead of dequeue |
| 621 | * followed by enqueue. |
| 622 | */ |
| 623 | static void requeue_task(struct task_struct *p, prio_array_t *array) |
| 624 | { |
| 625 | list_move_tail(&p->run_list, array->queue + p->prio); |
| 626 | } |
| 627 | |
| 628 | static inline void enqueue_task_head(struct task_struct *p, prio_array_t *array) |
| 629 | { |
| 630 | list_add(&p->run_list, array->queue + p->prio); |
| 631 | __set_bit(p->prio, array->bitmap); |
| 632 | array->nr_active++; |
| 633 | p->array = array; |
| 634 | } |
| 635 | |
| 636 | /* |
| 637 | * effective_prio - return the priority that is based on the static |
| 638 | * priority but is modified by bonuses/penalties. |
| 639 | * |
| 640 | * We scale the actual sleep average [0 .... MAX_SLEEP_AVG] |
| 641 | * into the -5 ... 0 ... +5 bonus/penalty range. |
| 642 | * |
| 643 | * We use 25% of the full 0...39 priority range so that: |
| 644 | * |
| 645 | * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs. |
| 646 | * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks. |
| 647 | * |
| 648 | * Both properties are important to certain workloads. |
| 649 | */ |
| 650 | static int effective_prio(task_t *p) |
| 651 | { |
| 652 | int bonus, prio; |
| 653 | |
| 654 | if (rt_task(p)) |
| 655 | return p->prio; |
| 656 | |
| 657 | bonus = CURRENT_BONUS(p) - MAX_BONUS / 2; |
| 658 | |
| 659 | prio = p->static_prio - bonus; |
| 660 | if (prio < MAX_RT_PRIO) |
| 661 | prio = MAX_RT_PRIO; |
| 662 | if (prio > MAX_PRIO-1) |
| 663 | prio = MAX_PRIO-1; |
| 664 | return prio; |
| 665 | } |
| 666 | |
| 667 | /* |
Mike Galbraith | 5ce74ab | 2006-04-10 22:52:44 -0700 | [diff] [blame] | 668 | * We place interactive tasks back into the active array, if possible. |
| 669 | * |
| 670 | * To guarantee that this does not starve expired tasks we ignore the |
| 671 | * interactivity of a task if the first expired task had to wait more |
| 672 | * than a 'reasonable' amount of time. This deadline timeout is |
| 673 | * load-dependent, as the frequency of array switched decreases with |
| 674 | * increasing number of running tasks. We also ignore the interactivity |
| 675 | * if a better static_prio task has expired, and switch periodically |
| 676 | * regardless, to ensure that highly interactive tasks do not starve |
| 677 | * the less fortunate for unreasonably long periods. |
| 678 | */ |
| 679 | static inline int expired_starving(runqueue_t *rq) |
| 680 | { |
| 681 | int limit; |
| 682 | |
| 683 | /* |
| 684 | * Arrays were recently switched, all is well |
| 685 | */ |
| 686 | if (!rq->expired_timestamp) |
| 687 | return 0; |
| 688 | |
| 689 | limit = STARVATION_LIMIT * rq->nr_running; |
| 690 | |
| 691 | /* |
| 692 | * It's time to switch arrays |
| 693 | */ |
| 694 | if (jiffies - rq->expired_timestamp >= limit) |
| 695 | return 1; |
| 696 | |
| 697 | /* |
| 698 | * There's a better selection in the expired array |
| 699 | */ |
| 700 | if (rq->curr->static_prio > rq->best_expired_prio) |
| 701 | return 1; |
| 702 | |
| 703 | /* |
| 704 | * All is well |
| 705 | */ |
| 706 | return 0; |
| 707 | } |
| 708 | |
| 709 | /* |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 710 | * __activate_task - move a task to the runqueue. |
| 711 | */ |
Con Kolivas | d425b27 | 2006-03-31 02:31:29 -0800 | [diff] [blame] | 712 | static void __activate_task(task_t *p, runqueue_t *rq) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 713 | { |
Con Kolivas | d425b27 | 2006-03-31 02:31:29 -0800 | [diff] [blame] | 714 | prio_array_t *target = rq->active; |
| 715 | |
Mike Galbraith | 8a5bc07 | 2006-04-10 22:52:45 -0700 | [diff] [blame] | 716 | if (unlikely(batch_task(p) || (expired_starving(rq) && !rt_task(p)))) |
Con Kolivas | d425b27 | 2006-03-31 02:31:29 -0800 | [diff] [blame] | 717 | target = rq->expired; |
| 718 | enqueue_task(p, target); |
Nick Piggin | a200057 | 2006-02-10 01:51:02 -0800 | [diff] [blame] | 719 | rq->nr_running++; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 720 | } |
| 721 | |
| 722 | /* |
| 723 | * __activate_idle_task - move idle task to the _front_ of runqueue. |
| 724 | */ |
| 725 | static inline void __activate_idle_task(task_t *p, runqueue_t *rq) |
| 726 | { |
| 727 | enqueue_task_head(p, rq->active); |
Nick Piggin | a200057 | 2006-02-10 01:51:02 -0800 | [diff] [blame] | 728 | rq->nr_running++; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 729 | } |
| 730 | |
Chen Shang | a3464a1 | 2005-06-25 14:57:31 -0700 | [diff] [blame] | 731 | static int recalc_task_prio(task_t *p, unsigned long long now) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 732 | { |
| 733 | /* Caller must always ensure 'now >= p->timestamp' */ |
| 734 | unsigned long long __sleep_time = now - p->timestamp; |
| 735 | unsigned long sleep_time; |
| 736 | |
Con Kolivas | d425b27 | 2006-03-31 02:31:29 -0800 | [diff] [blame] | 737 | if (batch_task(p)) |
Ingo Molnar | b0a9499 | 2006-01-14 13:20:41 -0800 | [diff] [blame] | 738 | sleep_time = 0; |
| 739 | else { |
| 740 | if (__sleep_time > NS_MAX_SLEEP_AVG) |
| 741 | sleep_time = NS_MAX_SLEEP_AVG; |
| 742 | else |
| 743 | sleep_time = (unsigned long)__sleep_time; |
| 744 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 745 | |
| 746 | if (likely(sleep_time > 0)) { |
| 747 | /* |
| 748 | * User tasks that sleep a long time are categorised as |
Con Kolivas | e72ff0b | 2006-03-31 02:31:26 -0800 | [diff] [blame] | 749 | * idle. They will only have their sleep_avg increased to a |
| 750 | * level that makes them just interactive priority to stay |
| 751 | * active yet prevent them suddenly becoming cpu hogs and |
| 752 | * starving other processes. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 753 | */ |
Con Kolivas | 5138930 | 2006-03-31 02:31:27 -0800 | [diff] [blame] | 754 | if (p->mm && sleep_time > INTERACTIVE_SLEEP(p)) { |
Con Kolivas | e72ff0b | 2006-03-31 02:31:26 -0800 | [diff] [blame] | 755 | unsigned long ceiling; |
| 756 | |
| 757 | ceiling = JIFFIES_TO_NS(MAX_SLEEP_AVG - |
| 758 | DEF_TIMESLICE); |
| 759 | if (p->sleep_avg < ceiling) |
| 760 | p->sleep_avg = ceiling; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 761 | } else { |
| 762 | /* |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 763 | * Tasks waking from uninterruptible sleep are |
| 764 | * limited in their sleep_avg rise as they |
| 765 | * are likely to be waiting on I/O |
| 766 | */ |
Con Kolivas | 3dee386 | 2006-03-31 02:31:23 -0800 | [diff] [blame] | 767 | if (p->sleep_type == SLEEP_NONINTERACTIVE && p->mm) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 768 | if (p->sleep_avg >= INTERACTIVE_SLEEP(p)) |
| 769 | sleep_time = 0; |
| 770 | else if (p->sleep_avg + sleep_time >= |
| 771 | INTERACTIVE_SLEEP(p)) { |
| 772 | p->sleep_avg = INTERACTIVE_SLEEP(p); |
| 773 | sleep_time = 0; |
| 774 | } |
| 775 | } |
| 776 | |
| 777 | /* |
| 778 | * This code gives a bonus to interactive tasks. |
| 779 | * |
| 780 | * The boost works by updating the 'average sleep time' |
| 781 | * value here, based on ->timestamp. The more time a |
| 782 | * task spends sleeping, the higher the average gets - |
| 783 | * and the higher the priority boost gets as well. |
| 784 | */ |
| 785 | p->sleep_avg += sleep_time; |
| 786 | |
| 787 | if (p->sleep_avg > NS_MAX_SLEEP_AVG) |
| 788 | p->sleep_avg = NS_MAX_SLEEP_AVG; |
| 789 | } |
| 790 | } |
| 791 | |
Chen Shang | a3464a1 | 2005-06-25 14:57:31 -0700 | [diff] [blame] | 792 | return effective_prio(p); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 793 | } |
| 794 | |
| 795 | /* |
| 796 | * activate_task - move a task to the runqueue and do priority recalculation |
| 797 | * |
| 798 | * Update all the scheduling statistics stuff. (sleep average |
| 799 | * calculation, priority modifiers, etc.) |
| 800 | */ |
| 801 | static void activate_task(task_t *p, runqueue_t *rq, int local) |
| 802 | { |
| 803 | unsigned long long now; |
| 804 | |
| 805 | now = sched_clock(); |
| 806 | #ifdef CONFIG_SMP |
| 807 | if (!local) { |
| 808 | /* Compensate for drifting sched_clock */ |
| 809 | runqueue_t *this_rq = this_rq(); |
| 810 | now = (now - this_rq->timestamp_last_tick) |
| 811 | + rq->timestamp_last_tick; |
| 812 | } |
| 813 | #endif |
| 814 | |
Chen, Kenneth W | a47ab93 | 2005-11-09 15:45:29 -0800 | [diff] [blame] | 815 | if (!rt_task(p)) |
| 816 | p->prio = recalc_task_prio(p, now); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 817 | |
| 818 | /* |
| 819 | * This checks to make sure it's not an uninterruptible task |
| 820 | * that is now waking up. |
| 821 | */ |
Con Kolivas | 3dee386 | 2006-03-31 02:31:23 -0800 | [diff] [blame] | 822 | if (p->sleep_type == SLEEP_NORMAL) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 823 | /* |
| 824 | * Tasks which were woken up by interrupts (ie. hw events) |
| 825 | * are most likely of interactive nature. So we give them |
| 826 | * the credit of extending their sleep time to the period |
| 827 | * of time they spend on the runqueue, waiting for execution |
| 828 | * on a CPU, first time around: |
| 829 | */ |
| 830 | if (in_interrupt()) |
Con Kolivas | 3dee386 | 2006-03-31 02:31:23 -0800 | [diff] [blame] | 831 | p->sleep_type = SLEEP_INTERRUPTED; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 832 | else { |
| 833 | /* |
| 834 | * Normal first-time wakeups get a credit too for |
| 835 | * on-runqueue time, but it will be weighted down: |
| 836 | */ |
Con Kolivas | 3dee386 | 2006-03-31 02:31:23 -0800 | [diff] [blame] | 837 | p->sleep_type = SLEEP_INTERACTIVE; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 838 | } |
| 839 | } |
| 840 | p->timestamp = now; |
| 841 | |
| 842 | __activate_task(p, rq); |
| 843 | } |
| 844 | |
| 845 | /* |
| 846 | * deactivate_task - remove a task from the runqueue. |
| 847 | */ |
| 848 | static void deactivate_task(struct task_struct *p, runqueue_t *rq) |
| 849 | { |
Nick Piggin | a200057 | 2006-02-10 01:51:02 -0800 | [diff] [blame] | 850 | rq->nr_running--; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 851 | dequeue_task(p, p->array); |
| 852 | p->array = NULL; |
| 853 | } |
| 854 | |
| 855 | /* |
| 856 | * resched_task - mark a task 'to be rescheduled now'. |
| 857 | * |
| 858 | * On UP this means the setting of the need_resched flag, on SMP it |
| 859 | * might also involve a cross-CPU call to trigger the scheduler on |
| 860 | * the target CPU. |
| 861 | */ |
| 862 | #ifdef CONFIG_SMP |
| 863 | static void resched_task(task_t *p) |
| 864 | { |
Nick Piggin | 64c7c8f | 2005-11-08 21:39:04 -0800 | [diff] [blame] | 865 | int cpu; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 866 | |
| 867 | assert_spin_locked(&task_rq(p)->lock); |
| 868 | |
Nick Piggin | 64c7c8f | 2005-11-08 21:39:04 -0800 | [diff] [blame] | 869 | if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED))) |
| 870 | return; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 871 | |
Nick Piggin | 64c7c8f | 2005-11-08 21:39:04 -0800 | [diff] [blame] | 872 | set_tsk_thread_flag(p, TIF_NEED_RESCHED); |
| 873 | |
| 874 | cpu = task_cpu(p); |
| 875 | if (cpu == smp_processor_id()) |
| 876 | return; |
| 877 | |
| 878 | /* NEED_RESCHED must be visible before we test POLLING_NRFLAG */ |
| 879 | smp_mb(); |
| 880 | if (!test_tsk_thread_flag(p, TIF_POLLING_NRFLAG)) |
| 881 | smp_send_reschedule(cpu); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 882 | } |
| 883 | #else |
| 884 | static inline void resched_task(task_t *p) |
| 885 | { |
Nick Piggin | 64c7c8f | 2005-11-08 21:39:04 -0800 | [diff] [blame] | 886 | assert_spin_locked(&task_rq(p)->lock); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 887 | set_tsk_need_resched(p); |
| 888 | } |
| 889 | #endif |
| 890 | |
| 891 | /** |
| 892 | * task_curr - is this task currently executing on a CPU? |
| 893 | * @p: the task in question. |
| 894 | */ |
| 895 | inline int task_curr(const task_t *p) |
| 896 | { |
| 897 | return cpu_curr(task_cpu(p)) == p; |
| 898 | } |
| 899 | |
| 900 | #ifdef CONFIG_SMP |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 901 | typedef struct { |
| 902 | struct list_head list; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 903 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 904 | task_t *task; |
| 905 | int dest_cpu; |
| 906 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 907 | struct completion done; |
| 908 | } migration_req_t; |
| 909 | |
| 910 | /* |
| 911 | * The task's runqueue lock must be held. |
| 912 | * Returns true if you have to wait for migration thread. |
| 913 | */ |
| 914 | static int migrate_task(task_t *p, int dest_cpu, migration_req_t *req) |
| 915 | { |
| 916 | runqueue_t *rq = task_rq(p); |
| 917 | |
| 918 | /* |
| 919 | * If the task is not on a runqueue (and not running), then |
| 920 | * it is sufficient to simply update the task's cpu field. |
| 921 | */ |
| 922 | if (!p->array && !task_running(rq, p)) { |
| 923 | set_task_cpu(p, dest_cpu); |
| 924 | return 0; |
| 925 | } |
| 926 | |
| 927 | init_completion(&req->done); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 928 | req->task = p; |
| 929 | req->dest_cpu = dest_cpu; |
| 930 | list_add(&req->list, &rq->migration_queue); |
| 931 | return 1; |
| 932 | } |
| 933 | |
| 934 | /* |
| 935 | * wait_task_inactive - wait for a thread to unschedule. |
| 936 | * |
| 937 | * The caller must ensure that the task *will* unschedule sometime soon, |
| 938 | * else this function might spin for a *long* time. This function can't |
| 939 | * be called with interrupts off, or it may introduce deadlock with |
| 940 | * smp_call_function() if an IPI is sent by the same process we are |
| 941 | * waiting to become inactive. |
| 942 | */ |
Ingo Molnar | 95cdf3b | 2005-09-10 00:26:11 -0700 | [diff] [blame] | 943 | void wait_task_inactive(task_t *p) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 944 | { |
| 945 | unsigned long flags; |
| 946 | runqueue_t *rq; |
| 947 | int preempted; |
| 948 | |
| 949 | repeat: |
| 950 | rq = task_rq_lock(p, &flags); |
| 951 | /* Must be off runqueue entirely, not preempted. */ |
| 952 | if (unlikely(p->array || task_running(rq, p))) { |
| 953 | /* If it's preempted, we yield. It could be a while. */ |
| 954 | preempted = !task_running(rq, p); |
| 955 | task_rq_unlock(rq, &flags); |
| 956 | cpu_relax(); |
| 957 | if (preempted) |
| 958 | yield(); |
| 959 | goto repeat; |
| 960 | } |
| 961 | task_rq_unlock(rq, &flags); |
| 962 | } |
| 963 | |
| 964 | /*** |
| 965 | * kick_process - kick a running thread to enter/exit the kernel |
| 966 | * @p: the to-be-kicked thread |
| 967 | * |
| 968 | * Cause a process which is running on another CPU to enter |
| 969 | * kernel-mode, without any delay. (to get signals handled.) |
| 970 | * |
| 971 | * NOTE: this function doesnt have to take the runqueue lock, |
| 972 | * because all it wants to ensure is that the remote task enters |
| 973 | * the kernel. If the IPI races and the task has been migrated |
| 974 | * to another CPU then no harm is done and the purpose has been |
| 975 | * achieved as well. |
| 976 | */ |
| 977 | void kick_process(task_t *p) |
| 978 | { |
| 979 | int cpu; |
| 980 | |
| 981 | preempt_disable(); |
| 982 | cpu = task_cpu(p); |
| 983 | if ((cpu != smp_processor_id()) && task_curr(p)) |
| 984 | smp_send_reschedule(cpu); |
| 985 | preempt_enable(); |
| 986 | } |
| 987 | |
| 988 | /* |
| 989 | * Return a low guess at the load of a migration-source cpu. |
| 990 | * |
| 991 | * We want to under-estimate the load of migration sources, to |
| 992 | * balance conservatively. |
| 993 | */ |
Con Kolivas | b910472 | 2005-11-08 21:38:55 -0800 | [diff] [blame] | 994 | static inline unsigned long source_load(int cpu, int type) |
| 995 | { |
Nick Piggin | a200057 | 2006-02-10 01:51:02 -0800 | [diff] [blame] | 996 | runqueue_t *rq = cpu_rq(cpu); |
| 997 | unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE; |
| 998 | if (type == 0) |
| 999 | return load_now; |
| 1000 | |
| 1001 | return min(rq->cpu_load[type-1], load_now); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1002 | } |
| 1003 | |
| 1004 | /* |
| 1005 | * Return a high guess at the load of a migration-target cpu |
| 1006 | */ |
Con Kolivas | b910472 | 2005-11-08 21:38:55 -0800 | [diff] [blame] | 1007 | static inline unsigned long target_load(int cpu, int type) |
| 1008 | { |
Nick Piggin | a200057 | 2006-02-10 01:51:02 -0800 | [diff] [blame] | 1009 | runqueue_t *rq = cpu_rq(cpu); |
| 1010 | unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE; |
| 1011 | if (type == 0) |
| 1012 | return load_now; |
| 1013 | |
| 1014 | return max(rq->cpu_load[type-1], load_now); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1015 | } |
| 1016 | |
Nick Piggin | 147cbb4 | 2005-06-25 14:57:19 -0700 | [diff] [blame] | 1017 | /* |
| 1018 | * find_idlest_group finds and returns the least busy CPU group within the |
| 1019 | * domain. |
| 1020 | */ |
| 1021 | static struct sched_group * |
| 1022 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu) |
| 1023 | { |
| 1024 | struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups; |
| 1025 | unsigned long min_load = ULONG_MAX, this_load = 0; |
| 1026 | int load_idx = sd->forkexec_idx; |
| 1027 | int imbalance = 100 + (sd->imbalance_pct-100)/2; |
| 1028 | |
| 1029 | do { |
| 1030 | unsigned long load, avg_load; |
| 1031 | int local_group; |
| 1032 | int i; |
| 1033 | |
M.Baris Demiray | da5a552 | 2005-09-10 00:26:09 -0700 | [diff] [blame] | 1034 | /* Skip over this group if it has no CPUs allowed */ |
| 1035 | if (!cpus_intersects(group->cpumask, p->cpus_allowed)) |
| 1036 | goto nextgroup; |
| 1037 | |
Nick Piggin | 147cbb4 | 2005-06-25 14:57:19 -0700 | [diff] [blame] | 1038 | local_group = cpu_isset(this_cpu, group->cpumask); |
Nick Piggin | 147cbb4 | 2005-06-25 14:57:19 -0700 | [diff] [blame] | 1039 | |
| 1040 | /* Tally up the load of all CPUs in the group */ |
| 1041 | avg_load = 0; |
| 1042 | |
| 1043 | for_each_cpu_mask(i, group->cpumask) { |
| 1044 | /* Bias balancing toward cpus of our domain */ |
| 1045 | if (local_group) |
| 1046 | load = source_load(i, load_idx); |
| 1047 | else |
| 1048 | load = target_load(i, load_idx); |
| 1049 | |
| 1050 | avg_load += load; |
| 1051 | } |
| 1052 | |
| 1053 | /* Adjust by relative CPU power of the group */ |
| 1054 | avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power; |
| 1055 | |
| 1056 | if (local_group) { |
| 1057 | this_load = avg_load; |
| 1058 | this = group; |
| 1059 | } else if (avg_load < min_load) { |
| 1060 | min_load = avg_load; |
| 1061 | idlest = group; |
| 1062 | } |
M.Baris Demiray | da5a552 | 2005-09-10 00:26:09 -0700 | [diff] [blame] | 1063 | nextgroup: |
Nick Piggin | 147cbb4 | 2005-06-25 14:57:19 -0700 | [diff] [blame] | 1064 | group = group->next; |
| 1065 | } while (group != sd->groups); |
| 1066 | |
| 1067 | if (!idlest || 100*this_load < imbalance*min_load) |
| 1068 | return NULL; |
| 1069 | return idlest; |
| 1070 | } |
| 1071 | |
| 1072 | /* |
| 1073 | * find_idlest_queue - find the idlest runqueue among the cpus in group. |
| 1074 | */ |
Ingo Molnar | 95cdf3b | 2005-09-10 00:26:11 -0700 | [diff] [blame] | 1075 | static int |
| 1076 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) |
Nick Piggin | 147cbb4 | 2005-06-25 14:57:19 -0700 | [diff] [blame] | 1077 | { |
M.Baris Demiray | da5a552 | 2005-09-10 00:26:09 -0700 | [diff] [blame] | 1078 | cpumask_t tmp; |
Nick Piggin | 147cbb4 | 2005-06-25 14:57:19 -0700 | [diff] [blame] | 1079 | unsigned long load, min_load = ULONG_MAX; |
| 1080 | int idlest = -1; |
| 1081 | int i; |
| 1082 | |
M.Baris Demiray | da5a552 | 2005-09-10 00:26:09 -0700 | [diff] [blame] | 1083 | /* Traverse only the allowed CPUs */ |
| 1084 | cpus_and(tmp, group->cpumask, p->cpus_allowed); |
| 1085 | |
| 1086 | for_each_cpu_mask(i, tmp) { |
Nick Piggin | 147cbb4 | 2005-06-25 14:57:19 -0700 | [diff] [blame] | 1087 | load = source_load(i, 0); |
| 1088 | |
| 1089 | if (load < min_load || (load == min_load && i == this_cpu)) { |
| 1090 | min_load = load; |
| 1091 | idlest = i; |
| 1092 | } |
| 1093 | } |
| 1094 | |
| 1095 | return idlest; |
| 1096 | } |
| 1097 | |
Nick Piggin | 476d139 | 2005-06-25 14:57:29 -0700 | [diff] [blame] | 1098 | /* |
| 1099 | * sched_balance_self: balance the current task (running on cpu) in domains |
| 1100 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and |
| 1101 | * SD_BALANCE_EXEC. |
| 1102 | * |
| 1103 | * Balance, ie. select the least loaded group. |
| 1104 | * |
| 1105 | * Returns the target CPU number, or the same CPU if no balancing is needed. |
| 1106 | * |
| 1107 | * preempt must be disabled. |
| 1108 | */ |
| 1109 | static int sched_balance_self(int cpu, int flag) |
| 1110 | { |
| 1111 | struct task_struct *t = current; |
| 1112 | struct sched_domain *tmp, *sd = NULL; |
Nick Piggin | 147cbb4 | 2005-06-25 14:57:19 -0700 | [diff] [blame] | 1113 | |
Nick Piggin | 476d139 | 2005-06-25 14:57:29 -0700 | [diff] [blame] | 1114 | for_each_domain(cpu, tmp) |
| 1115 | if (tmp->flags & flag) |
| 1116 | sd = tmp; |
| 1117 | |
| 1118 | while (sd) { |
| 1119 | cpumask_t span; |
| 1120 | struct sched_group *group; |
| 1121 | int new_cpu; |
| 1122 | int weight; |
| 1123 | |
| 1124 | span = sd->span; |
| 1125 | group = find_idlest_group(sd, t, cpu); |
| 1126 | if (!group) |
| 1127 | goto nextlevel; |
| 1128 | |
M.Baris Demiray | da5a552 | 2005-09-10 00:26:09 -0700 | [diff] [blame] | 1129 | new_cpu = find_idlest_cpu(group, t, cpu); |
Nick Piggin | 476d139 | 2005-06-25 14:57:29 -0700 | [diff] [blame] | 1130 | if (new_cpu == -1 || new_cpu == cpu) |
| 1131 | goto nextlevel; |
| 1132 | |
| 1133 | /* Now try balancing at a lower domain level */ |
| 1134 | cpu = new_cpu; |
| 1135 | nextlevel: |
| 1136 | sd = NULL; |
| 1137 | weight = cpus_weight(span); |
| 1138 | for_each_domain(cpu, tmp) { |
| 1139 | if (weight <= cpus_weight(tmp->span)) |
| 1140 | break; |
| 1141 | if (tmp->flags & flag) |
| 1142 | sd = tmp; |
| 1143 | } |
| 1144 | /* while loop will break here if sd == NULL */ |
| 1145 | } |
| 1146 | |
| 1147 | return cpu; |
| 1148 | } |
| 1149 | |
| 1150 | #endif /* CONFIG_SMP */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1151 | |
| 1152 | /* |
| 1153 | * wake_idle() will wake a task on an idle cpu if task->cpu is |
| 1154 | * not idle and an idle cpu is available. The span of cpus to |
| 1155 | * search starts with cpus closest then further out as needed, |
| 1156 | * so we always favor a closer, idle cpu. |
| 1157 | * |
| 1158 | * Returns the CPU we should wake onto. |
| 1159 | */ |
| 1160 | #if defined(ARCH_HAS_SCHED_WAKE_IDLE) |
| 1161 | static int wake_idle(int cpu, task_t *p) |
| 1162 | { |
| 1163 | cpumask_t tmp; |
| 1164 | struct sched_domain *sd; |
| 1165 | int i; |
| 1166 | |
| 1167 | if (idle_cpu(cpu)) |
| 1168 | return cpu; |
| 1169 | |
| 1170 | for_each_domain(cpu, sd) { |
| 1171 | if (sd->flags & SD_WAKE_IDLE) { |
Nick Piggin | e0f364f | 2005-06-25 14:57:06 -0700 | [diff] [blame] | 1172 | cpus_and(tmp, sd->span, p->cpus_allowed); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1173 | for_each_cpu_mask(i, tmp) { |
| 1174 | if (idle_cpu(i)) |
| 1175 | return i; |
| 1176 | } |
| 1177 | } |
Nick Piggin | e0f364f | 2005-06-25 14:57:06 -0700 | [diff] [blame] | 1178 | else |
| 1179 | break; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1180 | } |
| 1181 | return cpu; |
| 1182 | } |
| 1183 | #else |
| 1184 | static inline int wake_idle(int cpu, task_t *p) |
| 1185 | { |
| 1186 | return cpu; |
| 1187 | } |
| 1188 | #endif |
| 1189 | |
| 1190 | /*** |
| 1191 | * try_to_wake_up - wake up a thread |
| 1192 | * @p: the to-be-woken-up thread |
| 1193 | * @state: the mask of task states that can be woken |
| 1194 | * @sync: do a synchronous wakeup? |
| 1195 | * |
| 1196 | * Put it on the run-queue if it's not already there. The "current" |
| 1197 | * thread is always on the run-queue (except when the actual |
| 1198 | * re-schedule is in progress), and as such you're allowed to do |
| 1199 | * the simpler "current->state = TASK_RUNNING" to mark yourself |
| 1200 | * runnable without the overhead of this. |
| 1201 | * |
| 1202 | * returns failure only if the task is already active. |
| 1203 | */ |
Ingo Molnar | 95cdf3b | 2005-09-10 00:26:11 -0700 | [diff] [blame] | 1204 | static int try_to_wake_up(task_t *p, unsigned int state, int sync) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1205 | { |
| 1206 | int cpu, this_cpu, success = 0; |
| 1207 | unsigned long flags; |
| 1208 | long old_state; |
| 1209 | runqueue_t *rq; |
| 1210 | #ifdef CONFIG_SMP |
| 1211 | unsigned long load, this_load; |
Nick Piggin | 7897986 | 2005-06-25 14:57:13 -0700 | [diff] [blame] | 1212 | struct sched_domain *sd, *this_sd = NULL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1213 | int new_cpu; |
| 1214 | #endif |
| 1215 | |
| 1216 | rq = task_rq_lock(p, &flags); |
| 1217 | old_state = p->state; |
| 1218 | if (!(old_state & state)) |
| 1219 | goto out; |
| 1220 | |
| 1221 | if (p->array) |
| 1222 | goto out_running; |
| 1223 | |
| 1224 | cpu = task_cpu(p); |
| 1225 | this_cpu = smp_processor_id(); |
| 1226 | |
| 1227 | #ifdef CONFIG_SMP |
| 1228 | if (unlikely(task_running(rq, p))) |
| 1229 | goto out_activate; |
| 1230 | |
Nick Piggin | 7897986 | 2005-06-25 14:57:13 -0700 | [diff] [blame] | 1231 | new_cpu = cpu; |
| 1232 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1233 | schedstat_inc(rq, ttwu_cnt); |
| 1234 | if (cpu == this_cpu) { |
| 1235 | schedstat_inc(rq, ttwu_local); |
Nick Piggin | 7897986 | 2005-06-25 14:57:13 -0700 | [diff] [blame] | 1236 | goto out_set_cpu; |
| 1237 | } |
| 1238 | |
| 1239 | for_each_domain(this_cpu, sd) { |
| 1240 | if (cpu_isset(cpu, sd->span)) { |
| 1241 | schedstat_inc(sd, ttwu_wake_remote); |
| 1242 | this_sd = sd; |
| 1243 | break; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1244 | } |
| 1245 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1246 | |
Nick Piggin | 7897986 | 2005-06-25 14:57:13 -0700 | [diff] [blame] | 1247 | if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed))) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1248 | goto out_set_cpu; |
| 1249 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1250 | /* |
Nick Piggin | 7897986 | 2005-06-25 14:57:13 -0700 | [diff] [blame] | 1251 | * Check for affine wakeup and passive balancing possibilities. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1252 | */ |
Nick Piggin | 7897986 | 2005-06-25 14:57:13 -0700 | [diff] [blame] | 1253 | if (this_sd) { |
| 1254 | int idx = this_sd->wake_idx; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1255 | unsigned int imbalance; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1256 | |
Nick Piggin | a3f21bc | 2005-06-25 14:57:15 -0700 | [diff] [blame] | 1257 | imbalance = 100 + (this_sd->imbalance_pct - 100) / 2; |
| 1258 | |
Nick Piggin | 7897986 | 2005-06-25 14:57:13 -0700 | [diff] [blame] | 1259 | load = source_load(cpu, idx); |
| 1260 | this_load = target_load(this_cpu, idx); |
| 1261 | |
Nick Piggin | 7897986 | 2005-06-25 14:57:13 -0700 | [diff] [blame] | 1262 | new_cpu = this_cpu; /* Wake to this CPU if we can */ |
| 1263 | |
Nick Piggin | a3f21bc | 2005-06-25 14:57:15 -0700 | [diff] [blame] | 1264 | if (this_sd->flags & SD_WAKE_AFFINE) { |
| 1265 | unsigned long tl = this_load; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1266 | /* |
Nick Piggin | a3f21bc | 2005-06-25 14:57:15 -0700 | [diff] [blame] | 1267 | * If sync wakeup then subtract the (maximum possible) |
| 1268 | * effect of the currently running task from the load |
| 1269 | * of the current CPU: |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1270 | */ |
Nick Piggin | a3f21bc | 2005-06-25 14:57:15 -0700 | [diff] [blame] | 1271 | if (sync) |
| 1272 | tl -= SCHED_LOAD_SCALE; |
| 1273 | |
| 1274 | if ((tl <= load && |
| 1275 | tl + target_load(cpu, idx) <= SCHED_LOAD_SCALE) || |
| 1276 | 100*(tl + SCHED_LOAD_SCALE) <= imbalance*load) { |
| 1277 | /* |
| 1278 | * This domain has SD_WAKE_AFFINE and |
| 1279 | * p is cache cold in this domain, and |
| 1280 | * there is no bad imbalance. |
| 1281 | */ |
| 1282 | schedstat_inc(this_sd, ttwu_move_affine); |
| 1283 | goto out_set_cpu; |
| 1284 | } |
| 1285 | } |
| 1286 | |
| 1287 | /* |
| 1288 | * Start passive balancing when half the imbalance_pct |
| 1289 | * limit is reached. |
| 1290 | */ |
| 1291 | if (this_sd->flags & SD_WAKE_BALANCE) { |
| 1292 | if (imbalance*this_load <= 100*load) { |
| 1293 | schedstat_inc(this_sd, ttwu_move_balance); |
| 1294 | goto out_set_cpu; |
| 1295 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1296 | } |
| 1297 | } |
| 1298 | |
| 1299 | new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */ |
| 1300 | out_set_cpu: |
| 1301 | new_cpu = wake_idle(new_cpu, p); |
| 1302 | if (new_cpu != cpu) { |
| 1303 | set_task_cpu(p, new_cpu); |
| 1304 | task_rq_unlock(rq, &flags); |
| 1305 | /* might preempt at this point */ |
| 1306 | rq = task_rq_lock(p, &flags); |
| 1307 | old_state = p->state; |
| 1308 | if (!(old_state & state)) |
| 1309 | goto out; |
| 1310 | if (p->array) |
| 1311 | goto out_running; |
| 1312 | |
| 1313 | this_cpu = smp_processor_id(); |
| 1314 | cpu = task_cpu(p); |
| 1315 | } |
| 1316 | |
| 1317 | out_activate: |
| 1318 | #endif /* CONFIG_SMP */ |
| 1319 | if (old_state == TASK_UNINTERRUPTIBLE) { |
| 1320 | rq->nr_uninterruptible--; |
| 1321 | /* |
| 1322 | * Tasks on involuntary sleep don't earn |
| 1323 | * sleep_avg beyond just interactive state. |
| 1324 | */ |
Con Kolivas | 3dee386 | 2006-03-31 02:31:23 -0800 | [diff] [blame] | 1325 | p->sleep_type = SLEEP_NONINTERACTIVE; |
Con Kolivas | e7c38cb | 2006-03-31 02:31:25 -0800 | [diff] [blame] | 1326 | } else |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1327 | |
| 1328 | /* |
Ingo Molnar | d79fc0f | 2005-09-10 00:26:12 -0700 | [diff] [blame] | 1329 | * Tasks that have marked their sleep as noninteractive get |
Con Kolivas | e7c38cb | 2006-03-31 02:31:25 -0800 | [diff] [blame] | 1330 | * woken up with their sleep average not weighted in an |
| 1331 | * interactive way. |
Ingo Molnar | d79fc0f | 2005-09-10 00:26:12 -0700 | [diff] [blame] | 1332 | */ |
Con Kolivas | e7c38cb | 2006-03-31 02:31:25 -0800 | [diff] [blame] | 1333 | if (old_state & TASK_NONINTERACTIVE) |
| 1334 | p->sleep_type = SLEEP_NONINTERACTIVE; |
| 1335 | |
| 1336 | |
| 1337 | activate_task(p, rq, cpu == this_cpu); |
Ingo Molnar | d79fc0f | 2005-09-10 00:26:12 -0700 | [diff] [blame] | 1338 | /* |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1339 | * Sync wakeups (i.e. those types of wakeups where the waker |
| 1340 | * has indicated that it will leave the CPU in short order) |
| 1341 | * don't trigger a preemption, if the woken up task will run on |
| 1342 | * this cpu. (in this case the 'I will reschedule' promise of |
| 1343 | * the waker guarantees that the freshly woken up task is going |
| 1344 | * to be considered on this CPU.) |
| 1345 | */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1346 | if (!sync || cpu != this_cpu) { |
| 1347 | if (TASK_PREEMPTS_CURR(p, rq)) |
| 1348 | resched_task(rq->curr); |
| 1349 | } |
| 1350 | success = 1; |
| 1351 | |
| 1352 | out_running: |
| 1353 | p->state = TASK_RUNNING; |
| 1354 | out: |
| 1355 | task_rq_unlock(rq, &flags); |
| 1356 | |
| 1357 | return success; |
| 1358 | } |
| 1359 | |
Ingo Molnar | 95cdf3b | 2005-09-10 00:26:11 -0700 | [diff] [blame] | 1360 | int fastcall wake_up_process(task_t *p) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1361 | { |
| 1362 | return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED | |
| 1363 | TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0); |
| 1364 | } |
| 1365 | |
| 1366 | EXPORT_SYMBOL(wake_up_process); |
| 1367 | |
| 1368 | int fastcall wake_up_state(task_t *p, unsigned int state) |
| 1369 | { |
| 1370 | return try_to_wake_up(p, state, 0); |
| 1371 | } |
| 1372 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1373 | /* |
| 1374 | * Perform scheduler related setup for a newly forked process p. |
| 1375 | * p is forked by current. |
| 1376 | */ |
Nick Piggin | 476d139 | 2005-06-25 14:57:29 -0700 | [diff] [blame] | 1377 | void fastcall sched_fork(task_t *p, int clone_flags) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1378 | { |
Nick Piggin | 476d139 | 2005-06-25 14:57:29 -0700 | [diff] [blame] | 1379 | int cpu = get_cpu(); |
| 1380 | |
| 1381 | #ifdef CONFIG_SMP |
| 1382 | cpu = sched_balance_self(cpu, SD_BALANCE_FORK); |
| 1383 | #endif |
| 1384 | set_task_cpu(p, cpu); |
| 1385 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1386 | /* |
| 1387 | * We mark the process as running here, but have not actually |
| 1388 | * inserted it onto the runqueue yet. This guarantees that |
| 1389 | * nobody will actually run it, and a signal or other external |
| 1390 | * event cannot wake it up and insert it on the runqueue either. |
| 1391 | */ |
| 1392 | p->state = TASK_RUNNING; |
| 1393 | INIT_LIST_HEAD(&p->run_list); |
| 1394 | p->array = NULL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1395 | #ifdef CONFIG_SCHEDSTATS |
| 1396 | memset(&p->sched_info, 0, sizeof(p->sched_info)); |
| 1397 | #endif |
Chen, Kenneth W | d6077cb | 2006-02-14 13:53:10 -0800 | [diff] [blame] | 1398 | #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) |
Nick Piggin | 4866cde | 2005-06-25 14:57:23 -0700 | [diff] [blame] | 1399 | p->oncpu = 0; |
| 1400 | #endif |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1401 | #ifdef CONFIG_PREEMPT |
Nick Piggin | 4866cde | 2005-06-25 14:57:23 -0700 | [diff] [blame] | 1402 | /* Want to start with kernel preemption disabled. */ |
Al Viro | a1261f5 | 2005-11-13 16:06:55 -0800 | [diff] [blame] | 1403 | task_thread_info(p)->preempt_count = 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1404 | #endif |
| 1405 | /* |
| 1406 | * Share the timeslice between parent and child, thus the |
| 1407 | * total amount of pending timeslices in the system doesn't change, |
| 1408 | * resulting in more scheduling fairness. |
| 1409 | */ |
| 1410 | local_irq_disable(); |
| 1411 | p->time_slice = (current->time_slice + 1) >> 1; |
| 1412 | /* |
| 1413 | * The remainder of the first timeslice might be recovered by |
| 1414 | * the parent if the child exits early enough. |
| 1415 | */ |
| 1416 | p->first_time_slice = 1; |
| 1417 | current->time_slice >>= 1; |
| 1418 | p->timestamp = sched_clock(); |
| 1419 | if (unlikely(!current->time_slice)) { |
| 1420 | /* |
| 1421 | * This case is rare, it happens when the parent has only |
| 1422 | * a single jiffy left from its timeslice. Taking the |
| 1423 | * runqueue lock is not a problem. |
| 1424 | */ |
| 1425 | current->time_slice = 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1426 | scheduler_tick(); |
Nick Piggin | 476d139 | 2005-06-25 14:57:29 -0700 | [diff] [blame] | 1427 | } |
| 1428 | local_irq_enable(); |
| 1429 | put_cpu(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1430 | } |
| 1431 | |
| 1432 | /* |
| 1433 | * wake_up_new_task - wake up a newly created task for the first time. |
| 1434 | * |
| 1435 | * This function will do some initial scheduler statistics housekeeping |
| 1436 | * that must be done for every newly created context, then puts the task |
| 1437 | * on the runqueue and wakes it. |
| 1438 | */ |
Ingo Molnar | 95cdf3b | 2005-09-10 00:26:11 -0700 | [diff] [blame] | 1439 | void fastcall wake_up_new_task(task_t *p, unsigned long clone_flags) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1440 | { |
| 1441 | unsigned long flags; |
| 1442 | int this_cpu, cpu; |
| 1443 | runqueue_t *rq, *this_rq; |
| 1444 | |
| 1445 | rq = task_rq_lock(p, &flags); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1446 | BUG_ON(p->state != TASK_RUNNING); |
Nick Piggin | 147cbb4 | 2005-06-25 14:57:19 -0700 | [diff] [blame] | 1447 | this_cpu = smp_processor_id(); |
| 1448 | cpu = task_cpu(p); |
| 1449 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1450 | /* |
| 1451 | * We decrease the sleep average of forking parents |
| 1452 | * and children as well, to keep max-interactive tasks |
| 1453 | * from forking tasks that are max-interactive. The parent |
| 1454 | * (current) is done further down, under its lock. |
| 1455 | */ |
| 1456 | p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) * |
| 1457 | CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS); |
| 1458 | |
| 1459 | p->prio = effective_prio(p); |
| 1460 | |
| 1461 | if (likely(cpu == this_cpu)) { |
| 1462 | if (!(clone_flags & CLONE_VM)) { |
| 1463 | /* |
| 1464 | * The VM isn't cloned, so we're in a good position to |
| 1465 | * do child-runs-first in anticipation of an exec. This |
| 1466 | * usually avoids a lot of COW overhead. |
| 1467 | */ |
| 1468 | if (unlikely(!current->array)) |
| 1469 | __activate_task(p, rq); |
| 1470 | else { |
| 1471 | p->prio = current->prio; |
| 1472 | list_add_tail(&p->run_list, ¤t->run_list); |
| 1473 | p->array = current->array; |
| 1474 | p->array->nr_active++; |
Nick Piggin | a200057 | 2006-02-10 01:51:02 -0800 | [diff] [blame] | 1475 | rq->nr_running++; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1476 | } |
| 1477 | set_need_resched(); |
| 1478 | } else |
| 1479 | /* Run child last */ |
| 1480 | __activate_task(p, rq); |
| 1481 | /* |
| 1482 | * We skip the following code due to cpu == this_cpu |
| 1483 | * |
| 1484 | * task_rq_unlock(rq, &flags); |
| 1485 | * this_rq = task_rq_lock(current, &flags); |
| 1486 | */ |
| 1487 | this_rq = rq; |
| 1488 | } else { |
| 1489 | this_rq = cpu_rq(this_cpu); |
| 1490 | |
| 1491 | /* |
| 1492 | * Not the local CPU - must adjust timestamp. This should |
| 1493 | * get optimised away in the !CONFIG_SMP case. |
| 1494 | */ |
| 1495 | p->timestamp = (p->timestamp - this_rq->timestamp_last_tick) |
| 1496 | + rq->timestamp_last_tick; |
| 1497 | __activate_task(p, rq); |
| 1498 | if (TASK_PREEMPTS_CURR(p, rq)) |
| 1499 | resched_task(rq->curr); |
| 1500 | |
| 1501 | /* |
| 1502 | * Parent and child are on different CPUs, now get the |
| 1503 | * parent runqueue to update the parent's ->sleep_avg: |
| 1504 | */ |
| 1505 | task_rq_unlock(rq, &flags); |
| 1506 | this_rq = task_rq_lock(current, &flags); |
| 1507 | } |
| 1508 | current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) * |
| 1509 | PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS); |
| 1510 | task_rq_unlock(this_rq, &flags); |
| 1511 | } |
| 1512 | |
| 1513 | /* |
| 1514 | * Potentially available exiting-child timeslices are |
| 1515 | * retrieved here - this way the parent does not get |
| 1516 | * penalized for creating too many threads. |
| 1517 | * |
| 1518 | * (this cannot be used to 'generate' timeslices |
| 1519 | * artificially, because any timeslice recovered here |
| 1520 | * was given away by the parent in the first place.) |
| 1521 | */ |
Ingo Molnar | 95cdf3b | 2005-09-10 00:26:11 -0700 | [diff] [blame] | 1522 | void fastcall sched_exit(task_t *p) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1523 | { |
| 1524 | unsigned long flags; |
| 1525 | runqueue_t *rq; |
| 1526 | |
| 1527 | /* |
| 1528 | * If the child was a (relative-) CPU hog then decrease |
| 1529 | * the sleep_avg of the parent as well. |
| 1530 | */ |
| 1531 | rq = task_rq_lock(p->parent, &flags); |
Oleg Nesterov | 889dfaf | 2005-11-04 18:54:30 +0300 | [diff] [blame] | 1532 | if (p->first_time_slice && task_cpu(p) == task_cpu(p->parent)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1533 | p->parent->time_slice += p->time_slice; |
| 1534 | if (unlikely(p->parent->time_slice > task_timeslice(p))) |
| 1535 | p->parent->time_slice = task_timeslice(p); |
| 1536 | } |
| 1537 | if (p->sleep_avg < p->parent->sleep_avg) |
| 1538 | p->parent->sleep_avg = p->parent->sleep_avg / |
| 1539 | (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg / |
| 1540 | (EXIT_WEIGHT + 1); |
| 1541 | task_rq_unlock(rq, &flags); |
| 1542 | } |
| 1543 | |
| 1544 | /** |
Nick Piggin | 4866cde | 2005-06-25 14:57:23 -0700 | [diff] [blame] | 1545 | * prepare_task_switch - prepare to switch tasks |
| 1546 | * @rq: the runqueue preparing to switch |
| 1547 | * @next: the task we are going to switch to. |
| 1548 | * |
| 1549 | * This is called with the rq lock held and interrupts off. It must |
| 1550 | * be paired with a subsequent finish_task_switch after the context |
| 1551 | * switch. |
| 1552 | * |
| 1553 | * prepare_task_switch sets up locking and calls architecture specific |
| 1554 | * hooks. |
| 1555 | */ |
| 1556 | static inline void prepare_task_switch(runqueue_t *rq, task_t *next) |
| 1557 | { |
| 1558 | prepare_lock_switch(rq, next); |
| 1559 | prepare_arch_switch(next); |
| 1560 | } |
| 1561 | |
| 1562 | /** |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1563 | * finish_task_switch - clean up after a task-switch |
Jeff Garzik | 344baba | 2005-09-07 01:15:17 -0400 | [diff] [blame] | 1564 | * @rq: runqueue associated with task-switch |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1565 | * @prev: the thread we just switched away from. |
| 1566 | * |
Nick Piggin | 4866cde | 2005-06-25 14:57:23 -0700 | [diff] [blame] | 1567 | * finish_task_switch must be called after the context switch, paired |
| 1568 | * with a prepare_task_switch call before the context switch. |
| 1569 | * finish_task_switch will reconcile locking set up by prepare_task_switch, |
| 1570 | * and do any other architecture-specific cleanup actions. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1571 | * |
| 1572 | * Note that we may have delayed dropping an mm in context_switch(). If |
| 1573 | * so, we finish that here outside of the runqueue lock. (Doing it |
| 1574 | * with the lock held can cause deadlocks; see schedule() for |
| 1575 | * details.) |
| 1576 | */ |
Nick Piggin | 4866cde | 2005-06-25 14:57:23 -0700 | [diff] [blame] | 1577 | static inline void finish_task_switch(runqueue_t *rq, task_t *prev) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1578 | __releases(rq->lock) |
| 1579 | { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1580 | struct mm_struct *mm = rq->prev_mm; |
| 1581 | unsigned long prev_task_flags; |
| 1582 | |
| 1583 | rq->prev_mm = NULL; |
| 1584 | |
| 1585 | /* |
| 1586 | * A task struct has one reference for the use as "current". |
| 1587 | * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and |
| 1588 | * calls schedule one last time. The schedule call will never return, |
| 1589 | * and the scheduled task must drop that reference. |
| 1590 | * The test for EXIT_ZOMBIE must occur while the runqueue locks are |
| 1591 | * still held, otherwise prev could be scheduled on another cpu, die |
| 1592 | * there before we look at prev->state, and then the reference would |
| 1593 | * be dropped twice. |
| 1594 | * Manfred Spraul <manfred@colorfullife.com> |
| 1595 | */ |
| 1596 | prev_task_flags = prev->flags; |
Nick Piggin | 4866cde | 2005-06-25 14:57:23 -0700 | [diff] [blame] | 1597 | finish_arch_switch(prev); |
| 1598 | finish_lock_switch(rq, prev); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1599 | if (mm) |
| 1600 | mmdrop(mm); |
bibo mao | c6fd91f | 2006-03-26 01:38:20 -0800 | [diff] [blame] | 1601 | if (unlikely(prev_task_flags & PF_DEAD)) { |
| 1602 | /* |
| 1603 | * Remove function-return probe instances associated with this |
| 1604 | * task and put them back on the free list. |
| 1605 | */ |
| 1606 | kprobe_flush_task(prev); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1607 | put_task_struct(prev); |
bibo mao | c6fd91f | 2006-03-26 01:38:20 -0800 | [diff] [blame] | 1608 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1609 | } |
| 1610 | |
| 1611 | /** |
| 1612 | * schedule_tail - first thing a freshly forked thread must call. |
| 1613 | * @prev: the thread we just switched away from. |
| 1614 | */ |
| 1615 | asmlinkage void schedule_tail(task_t *prev) |
| 1616 | __releases(rq->lock) |
| 1617 | { |
Nick Piggin | 4866cde | 2005-06-25 14:57:23 -0700 | [diff] [blame] | 1618 | runqueue_t *rq = this_rq(); |
| 1619 | finish_task_switch(rq, prev); |
| 1620 | #ifdef __ARCH_WANT_UNLOCKED_CTXSW |
| 1621 | /* In this case, finish_task_switch does not reenable preemption */ |
| 1622 | preempt_enable(); |
| 1623 | #endif |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1624 | if (current->set_child_tid) |
| 1625 | put_user(current->pid, current->set_child_tid); |
| 1626 | } |
| 1627 | |
| 1628 | /* |
| 1629 | * context_switch - switch to the new MM and the new |
| 1630 | * thread's register state. |
| 1631 | */ |
| 1632 | static inline |
| 1633 | task_t * context_switch(runqueue_t *rq, task_t *prev, task_t *next) |
| 1634 | { |
| 1635 | struct mm_struct *mm = next->mm; |
| 1636 | struct mm_struct *oldmm = prev->active_mm; |
| 1637 | |
| 1638 | if (unlikely(!mm)) { |
| 1639 | next->active_mm = oldmm; |
| 1640 | atomic_inc(&oldmm->mm_count); |
| 1641 | enter_lazy_tlb(oldmm, next); |
| 1642 | } else |
| 1643 | switch_mm(oldmm, mm, next); |
| 1644 | |
| 1645 | if (unlikely(!prev->mm)) { |
| 1646 | prev->active_mm = NULL; |
| 1647 | WARN_ON(rq->prev_mm); |
| 1648 | rq->prev_mm = oldmm; |
| 1649 | } |
| 1650 | |
| 1651 | /* Here we just switch the register state and the stack. */ |
| 1652 | switch_to(prev, next, prev); |
| 1653 | |
| 1654 | return prev; |
| 1655 | } |
| 1656 | |
| 1657 | /* |
| 1658 | * nr_running, nr_uninterruptible and nr_context_switches: |
| 1659 | * |
| 1660 | * externally visible scheduler statistics: current number of runnable |
| 1661 | * threads, current number of uninterruptible-sleeping threads, total |
| 1662 | * number of context switches performed since bootup. |
| 1663 | */ |
| 1664 | unsigned long nr_running(void) |
| 1665 | { |
| 1666 | unsigned long i, sum = 0; |
| 1667 | |
| 1668 | for_each_online_cpu(i) |
| 1669 | sum += cpu_rq(i)->nr_running; |
| 1670 | |
| 1671 | return sum; |
| 1672 | } |
| 1673 | |
| 1674 | unsigned long nr_uninterruptible(void) |
| 1675 | { |
| 1676 | unsigned long i, sum = 0; |
| 1677 | |
KAMEZAWA Hiroyuki | 0a94502 | 2006-03-28 01:56:37 -0800 | [diff] [blame] | 1678 | for_each_possible_cpu(i) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1679 | sum += cpu_rq(i)->nr_uninterruptible; |
| 1680 | |
| 1681 | /* |
| 1682 | * Since we read the counters lockless, it might be slightly |
| 1683 | * inaccurate. Do not allow it to go below zero though: |
| 1684 | */ |
| 1685 | if (unlikely((long)sum < 0)) |
| 1686 | sum = 0; |
| 1687 | |
| 1688 | return sum; |
| 1689 | } |
| 1690 | |
| 1691 | unsigned long long nr_context_switches(void) |
| 1692 | { |
| 1693 | unsigned long long i, sum = 0; |
| 1694 | |
KAMEZAWA Hiroyuki | 0a94502 | 2006-03-28 01:56:37 -0800 | [diff] [blame] | 1695 | for_each_possible_cpu(i) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1696 | sum += cpu_rq(i)->nr_switches; |
| 1697 | |
| 1698 | return sum; |
| 1699 | } |
| 1700 | |
| 1701 | unsigned long nr_iowait(void) |
| 1702 | { |
| 1703 | unsigned long i, sum = 0; |
| 1704 | |
KAMEZAWA Hiroyuki | 0a94502 | 2006-03-28 01:56:37 -0800 | [diff] [blame] | 1705 | for_each_possible_cpu(i) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1706 | sum += atomic_read(&cpu_rq(i)->nr_iowait); |
| 1707 | |
| 1708 | return sum; |
| 1709 | } |
| 1710 | |
Jack Steiner | db1b1fe | 2006-03-31 02:31:21 -0800 | [diff] [blame] | 1711 | unsigned long nr_active(void) |
| 1712 | { |
| 1713 | unsigned long i, running = 0, uninterruptible = 0; |
| 1714 | |
| 1715 | for_each_online_cpu(i) { |
| 1716 | running += cpu_rq(i)->nr_running; |
| 1717 | uninterruptible += cpu_rq(i)->nr_uninterruptible; |
| 1718 | } |
| 1719 | |
| 1720 | if (unlikely((long)uninterruptible < 0)) |
| 1721 | uninterruptible = 0; |
| 1722 | |
| 1723 | return running + uninterruptible; |
| 1724 | } |
| 1725 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1726 | #ifdef CONFIG_SMP |
| 1727 | |
| 1728 | /* |
| 1729 | * double_rq_lock - safely lock two runqueues |
| 1730 | * |
Anton Blanchard | e9028b0 | 2006-03-23 02:59:20 -0800 | [diff] [blame] | 1731 | * We must take them in cpu order to match code in |
| 1732 | * dependent_sleeper and wake_dependent_sleeper. |
| 1733 | * |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1734 | * Note this does not disable interrupts like task_rq_lock, |
| 1735 | * you need to do so manually before calling. |
| 1736 | */ |
| 1737 | static void double_rq_lock(runqueue_t *rq1, runqueue_t *rq2) |
| 1738 | __acquires(rq1->lock) |
| 1739 | __acquires(rq2->lock) |
| 1740 | { |
| 1741 | if (rq1 == rq2) { |
| 1742 | spin_lock(&rq1->lock); |
| 1743 | __acquire(rq2->lock); /* Fake it out ;) */ |
| 1744 | } else { |
Anton Blanchard | e9028b0 | 2006-03-23 02:59:20 -0800 | [diff] [blame] | 1745 | if (rq1->cpu < rq2->cpu) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1746 | spin_lock(&rq1->lock); |
| 1747 | spin_lock(&rq2->lock); |
| 1748 | } else { |
| 1749 | spin_lock(&rq2->lock); |
| 1750 | spin_lock(&rq1->lock); |
| 1751 | } |
| 1752 | } |
| 1753 | } |
| 1754 | |
| 1755 | /* |
| 1756 | * double_rq_unlock - safely unlock two runqueues |
| 1757 | * |
| 1758 | * Note this does not restore interrupts like task_rq_unlock, |
| 1759 | * you need to do so manually after calling. |
| 1760 | */ |
| 1761 | static void double_rq_unlock(runqueue_t *rq1, runqueue_t *rq2) |
| 1762 | __releases(rq1->lock) |
| 1763 | __releases(rq2->lock) |
| 1764 | { |
| 1765 | spin_unlock(&rq1->lock); |
| 1766 | if (rq1 != rq2) |
| 1767 | spin_unlock(&rq2->lock); |
| 1768 | else |
| 1769 | __release(rq2->lock); |
| 1770 | } |
| 1771 | |
| 1772 | /* |
| 1773 | * double_lock_balance - lock the busiest runqueue, this_rq is locked already. |
| 1774 | */ |
| 1775 | static void double_lock_balance(runqueue_t *this_rq, runqueue_t *busiest) |
| 1776 | __releases(this_rq->lock) |
| 1777 | __acquires(busiest->lock) |
| 1778 | __acquires(this_rq->lock) |
| 1779 | { |
| 1780 | if (unlikely(!spin_trylock(&busiest->lock))) { |
Anton Blanchard | e9028b0 | 2006-03-23 02:59:20 -0800 | [diff] [blame] | 1781 | if (busiest->cpu < this_rq->cpu) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1782 | spin_unlock(&this_rq->lock); |
| 1783 | spin_lock(&busiest->lock); |
| 1784 | spin_lock(&this_rq->lock); |
| 1785 | } else |
| 1786 | spin_lock(&busiest->lock); |
| 1787 | } |
| 1788 | } |
| 1789 | |
| 1790 | /* |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1791 | * If dest_cpu is allowed for this process, migrate the task to it. |
| 1792 | * This is accomplished by forcing the cpu_allowed mask to only |
| 1793 | * allow dest_cpu, which will force the cpu onto dest_cpu. Then |
| 1794 | * the cpu_allowed mask is restored. |
| 1795 | */ |
| 1796 | static void sched_migrate_task(task_t *p, int dest_cpu) |
| 1797 | { |
| 1798 | migration_req_t req; |
| 1799 | runqueue_t *rq; |
| 1800 | unsigned long flags; |
| 1801 | |
| 1802 | rq = task_rq_lock(p, &flags); |
| 1803 | if (!cpu_isset(dest_cpu, p->cpus_allowed) |
| 1804 | || unlikely(cpu_is_offline(dest_cpu))) |
| 1805 | goto out; |
| 1806 | |
| 1807 | /* force the process onto the specified CPU */ |
| 1808 | if (migrate_task(p, dest_cpu, &req)) { |
| 1809 | /* Need to wait for migration thread (might exit: take ref). */ |
| 1810 | struct task_struct *mt = rq->migration_thread; |
| 1811 | get_task_struct(mt); |
| 1812 | task_rq_unlock(rq, &flags); |
| 1813 | wake_up_process(mt); |
| 1814 | put_task_struct(mt); |
| 1815 | wait_for_completion(&req.done); |
| 1816 | return; |
| 1817 | } |
| 1818 | out: |
| 1819 | task_rq_unlock(rq, &flags); |
| 1820 | } |
| 1821 | |
| 1822 | /* |
Nick Piggin | 476d139 | 2005-06-25 14:57:29 -0700 | [diff] [blame] | 1823 | * sched_exec - execve() is a valuable balancing opportunity, because at |
| 1824 | * this point the task has the smallest effective memory and cache footprint. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1825 | */ |
| 1826 | void sched_exec(void) |
| 1827 | { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1828 | int new_cpu, this_cpu = get_cpu(); |
Nick Piggin | 476d139 | 2005-06-25 14:57:29 -0700 | [diff] [blame] | 1829 | new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1830 | put_cpu(); |
Nick Piggin | 476d139 | 2005-06-25 14:57:29 -0700 | [diff] [blame] | 1831 | if (new_cpu != this_cpu) |
| 1832 | sched_migrate_task(current, new_cpu); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1833 | } |
| 1834 | |
| 1835 | /* |
| 1836 | * pull_task - move a task from a remote runqueue to the local runqueue. |
| 1837 | * Both runqueues must be locked. |
| 1838 | */ |
Arjan van de Ven | 858119e | 2006-01-14 13:20:43 -0800 | [diff] [blame] | 1839 | static |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1840 | void pull_task(runqueue_t *src_rq, prio_array_t *src_array, task_t *p, |
| 1841 | runqueue_t *this_rq, prio_array_t *this_array, int this_cpu) |
| 1842 | { |
| 1843 | dequeue_task(p, src_array); |
Nick Piggin | a200057 | 2006-02-10 01:51:02 -0800 | [diff] [blame] | 1844 | src_rq->nr_running--; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1845 | set_task_cpu(p, this_cpu); |
Nick Piggin | a200057 | 2006-02-10 01:51:02 -0800 | [diff] [blame] | 1846 | this_rq->nr_running++; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1847 | enqueue_task(p, this_array); |
| 1848 | p->timestamp = (p->timestamp - src_rq->timestamp_last_tick) |
| 1849 | + this_rq->timestamp_last_tick; |
| 1850 | /* |
| 1851 | * Note that idle threads have a prio of MAX_PRIO, for this test |
| 1852 | * to be always true for them. |
| 1853 | */ |
| 1854 | if (TASK_PREEMPTS_CURR(p, this_rq)) |
| 1855 | resched_task(this_rq->curr); |
| 1856 | } |
| 1857 | |
| 1858 | /* |
| 1859 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? |
| 1860 | */ |
Arjan van de Ven | 858119e | 2006-01-14 13:20:43 -0800 | [diff] [blame] | 1861 | static |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1862 | int can_migrate_task(task_t *p, runqueue_t *rq, int this_cpu, |
Ingo Molnar | 95cdf3b | 2005-09-10 00:26:11 -0700 | [diff] [blame] | 1863 | struct sched_domain *sd, enum idle_type idle, |
| 1864 | int *all_pinned) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1865 | { |
| 1866 | /* |
| 1867 | * We do not migrate tasks that are: |
| 1868 | * 1) running (obviously), or |
| 1869 | * 2) cannot be migrated to this CPU due to cpus_allowed, or |
| 1870 | * 3) are cache-hot on their current CPU. |
| 1871 | */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1872 | if (!cpu_isset(this_cpu, p->cpus_allowed)) |
| 1873 | return 0; |
Nick Piggin | 8102679 | 2005-06-25 14:57:07 -0700 | [diff] [blame] | 1874 | *all_pinned = 0; |
| 1875 | |
| 1876 | if (task_running(rq, p)) |
| 1877 | return 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1878 | |
| 1879 | /* |
| 1880 | * Aggressive migration if: |
Nick Piggin | cafb20c | 2005-06-25 14:57:17 -0700 | [diff] [blame] | 1881 | * 1) task is cache cold, or |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1882 | * 2) too many balance attempts have failed. |
| 1883 | */ |
| 1884 | |
Nick Piggin | cafb20c | 2005-06-25 14:57:17 -0700 | [diff] [blame] | 1885 | if (sd->nr_balance_failed > sd->cache_nice_tries) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1886 | return 1; |
| 1887 | |
| 1888 | if (task_hot(p, rq->timestamp_last_tick, sd)) |
Nick Piggin | 8102679 | 2005-06-25 14:57:07 -0700 | [diff] [blame] | 1889 | return 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1890 | return 1; |
| 1891 | } |
| 1892 | |
| 1893 | /* |
| 1894 | * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq, |
| 1895 | * as part of a balancing operation within "domain". Returns the number of |
| 1896 | * tasks moved. |
| 1897 | * |
| 1898 | * Called with both runqueues locked. |
| 1899 | */ |
| 1900 | static int move_tasks(runqueue_t *this_rq, int this_cpu, runqueue_t *busiest, |
| 1901 | unsigned long max_nr_move, struct sched_domain *sd, |
Nick Piggin | 8102679 | 2005-06-25 14:57:07 -0700 | [diff] [blame] | 1902 | enum idle_type idle, int *all_pinned) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1903 | { |
| 1904 | prio_array_t *array, *dst_array; |
| 1905 | struct list_head *head, *curr; |
Nick Piggin | 8102679 | 2005-06-25 14:57:07 -0700 | [diff] [blame] | 1906 | int idx, pulled = 0, pinned = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1907 | task_t *tmp; |
| 1908 | |
Nick Piggin | 8102679 | 2005-06-25 14:57:07 -0700 | [diff] [blame] | 1909 | if (max_nr_move == 0) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1910 | goto out; |
| 1911 | |
Nick Piggin | 8102679 | 2005-06-25 14:57:07 -0700 | [diff] [blame] | 1912 | pinned = 1; |
| 1913 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1914 | /* |
| 1915 | * We first consider expired tasks. Those will likely not be |
| 1916 | * executed in the near future, and they are most likely to |
| 1917 | * be cache-cold, thus switching CPUs has the least effect |
| 1918 | * on them. |
| 1919 | */ |
| 1920 | if (busiest->expired->nr_active) { |
| 1921 | array = busiest->expired; |
| 1922 | dst_array = this_rq->expired; |
| 1923 | } else { |
| 1924 | array = busiest->active; |
| 1925 | dst_array = this_rq->active; |
| 1926 | } |
| 1927 | |
| 1928 | new_array: |
| 1929 | /* Start searching at priority 0: */ |
| 1930 | idx = 0; |
| 1931 | skip_bitmap: |
| 1932 | if (!idx) |
| 1933 | idx = sched_find_first_bit(array->bitmap); |
| 1934 | else |
| 1935 | idx = find_next_bit(array->bitmap, MAX_PRIO, idx); |
| 1936 | if (idx >= MAX_PRIO) { |
| 1937 | if (array == busiest->expired && busiest->active->nr_active) { |
| 1938 | array = busiest->active; |
| 1939 | dst_array = this_rq->active; |
| 1940 | goto new_array; |
| 1941 | } |
| 1942 | goto out; |
| 1943 | } |
| 1944 | |
| 1945 | head = array->queue + idx; |
| 1946 | curr = head->prev; |
| 1947 | skip_queue: |
| 1948 | tmp = list_entry(curr, task_t, run_list); |
| 1949 | |
| 1950 | curr = curr->prev; |
| 1951 | |
Nick Piggin | 8102679 | 2005-06-25 14:57:07 -0700 | [diff] [blame] | 1952 | if (!can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1953 | if (curr != head) |
| 1954 | goto skip_queue; |
| 1955 | idx++; |
| 1956 | goto skip_bitmap; |
| 1957 | } |
| 1958 | |
| 1959 | #ifdef CONFIG_SCHEDSTATS |
| 1960 | if (task_hot(tmp, busiest->timestamp_last_tick, sd)) |
| 1961 | schedstat_inc(sd, lb_hot_gained[idle]); |
| 1962 | #endif |
| 1963 | |
| 1964 | pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu); |
| 1965 | pulled++; |
| 1966 | |
| 1967 | /* We only want to steal up to the prescribed number of tasks. */ |
| 1968 | if (pulled < max_nr_move) { |
| 1969 | if (curr != head) |
| 1970 | goto skip_queue; |
| 1971 | idx++; |
| 1972 | goto skip_bitmap; |
| 1973 | } |
| 1974 | out: |
| 1975 | /* |
| 1976 | * Right now, this is the only place pull_task() is called, |
| 1977 | * so we can safely collect pull_task() stats here rather than |
| 1978 | * inside pull_task(). |
| 1979 | */ |
| 1980 | schedstat_add(sd, lb_gained[idle], pulled); |
Nick Piggin | 8102679 | 2005-06-25 14:57:07 -0700 | [diff] [blame] | 1981 | |
| 1982 | if (all_pinned) |
| 1983 | *all_pinned = pinned; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1984 | return pulled; |
| 1985 | } |
| 1986 | |
| 1987 | /* |
| 1988 | * find_busiest_group finds and returns the busiest CPU group within the |
| 1989 | * domain. It calculates and returns the number of tasks which should be |
| 1990 | * moved to restore balance via the imbalance parameter. |
| 1991 | */ |
| 1992 | static struct sched_group * |
| 1993 | find_busiest_group(struct sched_domain *sd, int this_cpu, |
Nick Piggin | 5969fe0 | 2005-09-10 00:26:19 -0700 | [diff] [blame] | 1994 | unsigned long *imbalance, enum idle_type idle, int *sd_idle) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1995 | { |
| 1996 | struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups; |
| 1997 | unsigned long max_load, avg_load, total_load, this_load, total_pwr; |
Siddha, Suresh B | 0c117f1 | 2005-09-10 00:26:21 -0700 | [diff] [blame] | 1998 | unsigned long max_pull; |
Nick Piggin | 7897986 | 2005-06-25 14:57:13 -0700 | [diff] [blame] | 1999 | int load_idx; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2000 | |
| 2001 | max_load = this_load = total_load = total_pwr = 0; |
Nick Piggin | 7897986 | 2005-06-25 14:57:13 -0700 | [diff] [blame] | 2002 | if (idle == NOT_IDLE) |
| 2003 | load_idx = sd->busy_idx; |
| 2004 | else if (idle == NEWLY_IDLE) |
| 2005 | load_idx = sd->newidle_idx; |
| 2006 | else |
| 2007 | load_idx = sd->idle_idx; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2008 | |
| 2009 | do { |
| 2010 | unsigned long load; |
| 2011 | int local_group; |
| 2012 | int i; |
| 2013 | |
| 2014 | local_group = cpu_isset(this_cpu, group->cpumask); |
| 2015 | |
| 2016 | /* Tally up the load of all CPUs in the group */ |
| 2017 | avg_load = 0; |
| 2018 | |
| 2019 | for_each_cpu_mask(i, group->cpumask) { |
Nick Piggin | 5969fe0 | 2005-09-10 00:26:19 -0700 | [diff] [blame] | 2020 | if (*sd_idle && !idle_cpu(i)) |
| 2021 | *sd_idle = 0; |
| 2022 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2023 | /* Bias balancing toward cpus of our domain */ |
| 2024 | if (local_group) |
Nick Piggin | a200057 | 2006-02-10 01:51:02 -0800 | [diff] [blame] | 2025 | load = target_load(i, load_idx); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2026 | else |
Nick Piggin | a200057 | 2006-02-10 01:51:02 -0800 | [diff] [blame] | 2027 | load = source_load(i, load_idx); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2028 | |
| 2029 | avg_load += load; |
| 2030 | } |
| 2031 | |
| 2032 | total_load += avg_load; |
| 2033 | total_pwr += group->cpu_power; |
| 2034 | |
| 2035 | /* Adjust by relative CPU power of the group */ |
| 2036 | avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power; |
| 2037 | |
| 2038 | if (local_group) { |
| 2039 | this_load = avg_load; |
| 2040 | this = group; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2041 | } else if (avg_load > max_load) { |
| 2042 | max_load = avg_load; |
| 2043 | busiest = group; |
| 2044 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2045 | group = group->next; |
| 2046 | } while (group != sd->groups); |
| 2047 | |
Siddha, Suresh B | 0c117f1 | 2005-09-10 00:26:21 -0700 | [diff] [blame] | 2048 | if (!busiest || this_load >= max_load || max_load <= SCHED_LOAD_SCALE) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2049 | goto out_balanced; |
| 2050 | |
| 2051 | avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr; |
| 2052 | |
| 2053 | if (this_load >= avg_load || |
| 2054 | 100*max_load <= sd->imbalance_pct*this_load) |
| 2055 | goto out_balanced; |
| 2056 | |
| 2057 | /* |
| 2058 | * We're trying to get all the cpus to the average_load, so we don't |
| 2059 | * want to push ourselves above the average load, nor do we wish to |
| 2060 | * reduce the max loaded cpu below the average load, as either of these |
| 2061 | * actions would just result in more rebalancing later, and ping-pong |
| 2062 | * tasks around. Thus we look for the minimum possible imbalance. |
| 2063 | * Negative imbalances (*we* are more loaded than anyone else) will |
| 2064 | * be counted as no imbalance for these purposes -- we can't fix that |
| 2065 | * by pulling tasks to us. Be careful of negative numbers as they'll |
| 2066 | * appear as very large values with unsigned longs. |
| 2067 | */ |
Siddha, Suresh B | 0c117f1 | 2005-09-10 00:26:21 -0700 | [diff] [blame] | 2068 | |
| 2069 | /* Don't want to pull so many tasks that a group would go idle */ |
| 2070 | max_pull = min(max_load - avg_load, max_load - SCHED_LOAD_SCALE); |
| 2071 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2072 | /* How much load to actually move to equalise the imbalance */ |
Siddha, Suresh B | 0c117f1 | 2005-09-10 00:26:21 -0700 | [diff] [blame] | 2073 | *imbalance = min(max_pull * busiest->cpu_power, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2074 | (avg_load - this_load) * this->cpu_power) |
| 2075 | / SCHED_LOAD_SCALE; |
| 2076 | |
| 2077 | if (*imbalance < SCHED_LOAD_SCALE) { |
| 2078 | unsigned long pwr_now = 0, pwr_move = 0; |
| 2079 | unsigned long tmp; |
| 2080 | |
| 2081 | if (max_load - this_load >= SCHED_LOAD_SCALE*2) { |
| 2082 | *imbalance = 1; |
| 2083 | return busiest; |
| 2084 | } |
| 2085 | |
| 2086 | /* |
| 2087 | * OK, we don't have enough imbalance to justify moving tasks, |
| 2088 | * however we may be able to increase total CPU power used by |
| 2089 | * moving them. |
| 2090 | */ |
| 2091 | |
| 2092 | pwr_now += busiest->cpu_power*min(SCHED_LOAD_SCALE, max_load); |
| 2093 | pwr_now += this->cpu_power*min(SCHED_LOAD_SCALE, this_load); |
| 2094 | pwr_now /= SCHED_LOAD_SCALE; |
| 2095 | |
| 2096 | /* Amount of load we'd subtract */ |
| 2097 | tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/busiest->cpu_power; |
| 2098 | if (max_load > tmp) |
| 2099 | pwr_move += busiest->cpu_power*min(SCHED_LOAD_SCALE, |
| 2100 | max_load - tmp); |
| 2101 | |
| 2102 | /* Amount of load we'd add */ |
| 2103 | if (max_load*busiest->cpu_power < |
| 2104 | SCHED_LOAD_SCALE*SCHED_LOAD_SCALE) |
| 2105 | tmp = max_load*busiest->cpu_power/this->cpu_power; |
| 2106 | else |
| 2107 | tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/this->cpu_power; |
| 2108 | pwr_move += this->cpu_power*min(SCHED_LOAD_SCALE, this_load + tmp); |
| 2109 | pwr_move /= SCHED_LOAD_SCALE; |
| 2110 | |
| 2111 | /* Move if we gain throughput */ |
| 2112 | if (pwr_move <= pwr_now) |
| 2113 | goto out_balanced; |
| 2114 | |
| 2115 | *imbalance = 1; |
| 2116 | return busiest; |
| 2117 | } |
| 2118 | |
| 2119 | /* Get rid of the scaling factor, rounding down as we divide */ |
| 2120 | *imbalance = *imbalance / SCHED_LOAD_SCALE; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2121 | return busiest; |
| 2122 | |
| 2123 | out_balanced: |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2124 | |
| 2125 | *imbalance = 0; |
| 2126 | return NULL; |
| 2127 | } |
| 2128 | |
| 2129 | /* |
| 2130 | * find_busiest_queue - find the busiest runqueue among the cpus in group. |
| 2131 | */ |
Con Kolivas | b910472 | 2005-11-08 21:38:55 -0800 | [diff] [blame] | 2132 | static runqueue_t *find_busiest_queue(struct sched_group *group, |
| 2133 | enum idle_type idle) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2134 | { |
| 2135 | unsigned long load, max_load = 0; |
| 2136 | runqueue_t *busiest = NULL; |
| 2137 | int i; |
| 2138 | |
| 2139 | for_each_cpu_mask(i, group->cpumask) { |
Nick Piggin | a200057 | 2006-02-10 01:51:02 -0800 | [diff] [blame] | 2140 | load = source_load(i, 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2141 | |
| 2142 | if (load > max_load) { |
| 2143 | max_load = load; |
| 2144 | busiest = cpu_rq(i); |
| 2145 | } |
| 2146 | } |
| 2147 | |
| 2148 | return busiest; |
| 2149 | } |
| 2150 | |
| 2151 | /* |
Nick Piggin | 77391d7 | 2005-06-25 14:57:30 -0700 | [diff] [blame] | 2152 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but |
| 2153 | * so long as it is large enough. |
| 2154 | */ |
| 2155 | #define MAX_PINNED_INTERVAL 512 |
| 2156 | |
| 2157 | /* |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2158 | * Check this_cpu to ensure it is balanced within domain. Attempt to move |
| 2159 | * tasks if there is an imbalance. |
| 2160 | * |
| 2161 | * Called with this_rq unlocked. |
| 2162 | */ |
| 2163 | static int load_balance(int this_cpu, runqueue_t *this_rq, |
| 2164 | struct sched_domain *sd, enum idle_type idle) |
| 2165 | { |
| 2166 | struct sched_group *group; |
| 2167 | runqueue_t *busiest; |
| 2168 | unsigned long imbalance; |
Nick Piggin | 77391d7 | 2005-06-25 14:57:30 -0700 | [diff] [blame] | 2169 | int nr_moved, all_pinned = 0; |
Nick Piggin | 8102679 | 2005-06-25 14:57:07 -0700 | [diff] [blame] | 2170 | int active_balance = 0; |
Nick Piggin | 5969fe0 | 2005-09-10 00:26:19 -0700 | [diff] [blame] | 2171 | int sd_idle = 0; |
| 2172 | |
| 2173 | if (idle != NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER) |
| 2174 | sd_idle = 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2175 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2176 | schedstat_inc(sd, lb_cnt[idle]); |
| 2177 | |
Nick Piggin | 5969fe0 | 2005-09-10 00:26:19 -0700 | [diff] [blame] | 2178 | group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2179 | if (!group) { |
| 2180 | schedstat_inc(sd, lb_nobusyg[idle]); |
| 2181 | goto out_balanced; |
| 2182 | } |
| 2183 | |
Con Kolivas | b910472 | 2005-11-08 21:38:55 -0800 | [diff] [blame] | 2184 | busiest = find_busiest_queue(group, idle); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2185 | if (!busiest) { |
| 2186 | schedstat_inc(sd, lb_nobusyq[idle]); |
| 2187 | goto out_balanced; |
| 2188 | } |
| 2189 | |
Nick Piggin | db935db | 2005-06-25 14:57:11 -0700 | [diff] [blame] | 2190 | BUG_ON(busiest == this_rq); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2191 | |
| 2192 | schedstat_add(sd, lb_imbalance[idle], imbalance); |
| 2193 | |
| 2194 | nr_moved = 0; |
| 2195 | if (busiest->nr_running > 1) { |
| 2196 | /* |
| 2197 | * Attempt to move tasks. If find_busiest_group has found |
| 2198 | * an imbalance but busiest->nr_running <= 1, the group is |
| 2199 | * still unbalanced. nr_moved simply stays zero, so it is |
| 2200 | * correctly treated as an imbalance. |
| 2201 | */ |
Nick Piggin | e17224b | 2005-09-10 00:26:18 -0700 | [diff] [blame] | 2202 | double_rq_lock(this_rq, busiest); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2203 | nr_moved = move_tasks(this_rq, this_cpu, busiest, |
Nick Piggin | d6d5cfa | 2005-09-10 00:26:16 -0700 | [diff] [blame] | 2204 | imbalance, sd, idle, &all_pinned); |
Nick Piggin | e17224b | 2005-09-10 00:26:18 -0700 | [diff] [blame] | 2205 | double_rq_unlock(this_rq, busiest); |
Nick Piggin | 8102679 | 2005-06-25 14:57:07 -0700 | [diff] [blame] | 2206 | |
| 2207 | /* All tasks on this runqueue were pinned by CPU affinity */ |
| 2208 | if (unlikely(all_pinned)) |
| 2209 | goto out_balanced; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2210 | } |
Nick Piggin | 8102679 | 2005-06-25 14:57:07 -0700 | [diff] [blame] | 2211 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2212 | if (!nr_moved) { |
| 2213 | schedstat_inc(sd, lb_failed[idle]); |
| 2214 | sd->nr_balance_failed++; |
| 2215 | |
| 2216 | if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2217 | |
| 2218 | spin_lock(&busiest->lock); |
Siddha, Suresh B | fa3b6dd | 2005-09-10 00:26:21 -0700 | [diff] [blame] | 2219 | |
| 2220 | /* don't kick the migration_thread, if the curr |
| 2221 | * task on busiest cpu can't be moved to this_cpu |
| 2222 | */ |
| 2223 | if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) { |
| 2224 | spin_unlock(&busiest->lock); |
| 2225 | all_pinned = 1; |
| 2226 | goto out_one_pinned; |
| 2227 | } |
| 2228 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2229 | if (!busiest->active_balance) { |
| 2230 | busiest->active_balance = 1; |
| 2231 | busiest->push_cpu = this_cpu; |
Nick Piggin | 8102679 | 2005-06-25 14:57:07 -0700 | [diff] [blame] | 2232 | active_balance = 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2233 | } |
| 2234 | spin_unlock(&busiest->lock); |
Nick Piggin | 8102679 | 2005-06-25 14:57:07 -0700 | [diff] [blame] | 2235 | if (active_balance) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2236 | wake_up_process(busiest->migration_thread); |
| 2237 | |
| 2238 | /* |
| 2239 | * We've kicked active balancing, reset the failure |
| 2240 | * counter. |
| 2241 | */ |
Nick Piggin | 3950745 | 2005-06-25 14:57:09 -0700 | [diff] [blame] | 2242 | sd->nr_balance_failed = sd->cache_nice_tries+1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2243 | } |
Nick Piggin | 8102679 | 2005-06-25 14:57:07 -0700 | [diff] [blame] | 2244 | } else |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2245 | sd->nr_balance_failed = 0; |
| 2246 | |
Nick Piggin | 8102679 | 2005-06-25 14:57:07 -0700 | [diff] [blame] | 2247 | if (likely(!active_balance)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2248 | /* We were unbalanced, so reset the balancing interval */ |
| 2249 | sd->balance_interval = sd->min_interval; |
Nick Piggin | 8102679 | 2005-06-25 14:57:07 -0700 | [diff] [blame] | 2250 | } else { |
| 2251 | /* |
| 2252 | * If we've begun active balancing, start to back off. This |
| 2253 | * case may not be covered by the all_pinned logic if there |
| 2254 | * is only 1 task on the busy runqueue (because we don't call |
| 2255 | * move_tasks). |
| 2256 | */ |
| 2257 | if (sd->balance_interval < sd->max_interval) |
| 2258 | sd->balance_interval *= 2; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2259 | } |
| 2260 | |
Nick Piggin | 5969fe0 | 2005-09-10 00:26:19 -0700 | [diff] [blame] | 2261 | if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER) |
| 2262 | return -1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2263 | return nr_moved; |
| 2264 | |
| 2265 | out_balanced: |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2266 | schedstat_inc(sd, lb_balanced[idle]); |
| 2267 | |
Nick Piggin | 16cfb1c | 2005-06-25 14:57:08 -0700 | [diff] [blame] | 2268 | sd->nr_balance_failed = 0; |
Siddha, Suresh B | fa3b6dd | 2005-09-10 00:26:21 -0700 | [diff] [blame] | 2269 | |
| 2270 | out_one_pinned: |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2271 | /* tune up the balancing interval */ |
Nick Piggin | 77391d7 | 2005-06-25 14:57:30 -0700 | [diff] [blame] | 2272 | if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) || |
| 2273 | (sd->balance_interval < sd->max_interval)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2274 | sd->balance_interval *= 2; |
| 2275 | |
Nick Piggin | 5969fe0 | 2005-09-10 00:26:19 -0700 | [diff] [blame] | 2276 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER) |
| 2277 | return -1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2278 | return 0; |
| 2279 | } |
| 2280 | |
| 2281 | /* |
| 2282 | * Check this_cpu to ensure it is balanced within domain. Attempt to move |
| 2283 | * tasks if there is an imbalance. |
| 2284 | * |
| 2285 | * Called from schedule when this_rq is about to become idle (NEWLY_IDLE). |
| 2286 | * this_rq is locked. |
| 2287 | */ |
| 2288 | static int load_balance_newidle(int this_cpu, runqueue_t *this_rq, |
| 2289 | struct sched_domain *sd) |
| 2290 | { |
| 2291 | struct sched_group *group; |
| 2292 | runqueue_t *busiest = NULL; |
| 2293 | unsigned long imbalance; |
| 2294 | int nr_moved = 0; |
Nick Piggin | 5969fe0 | 2005-09-10 00:26:19 -0700 | [diff] [blame] | 2295 | int sd_idle = 0; |
| 2296 | |
| 2297 | if (sd->flags & SD_SHARE_CPUPOWER) |
| 2298 | sd_idle = 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2299 | |
| 2300 | schedstat_inc(sd, lb_cnt[NEWLY_IDLE]); |
Nick Piggin | 5969fe0 | 2005-09-10 00:26:19 -0700 | [diff] [blame] | 2301 | group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE, &sd_idle); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2302 | if (!group) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2303 | schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]); |
Nick Piggin | 16cfb1c | 2005-06-25 14:57:08 -0700 | [diff] [blame] | 2304 | goto out_balanced; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2305 | } |
| 2306 | |
Con Kolivas | b910472 | 2005-11-08 21:38:55 -0800 | [diff] [blame] | 2307 | busiest = find_busiest_queue(group, NEWLY_IDLE); |
Nick Piggin | db935db | 2005-06-25 14:57:11 -0700 | [diff] [blame] | 2308 | if (!busiest) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2309 | schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]); |
Nick Piggin | 16cfb1c | 2005-06-25 14:57:08 -0700 | [diff] [blame] | 2310 | goto out_balanced; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2311 | } |
| 2312 | |
Nick Piggin | db935db | 2005-06-25 14:57:11 -0700 | [diff] [blame] | 2313 | BUG_ON(busiest == this_rq); |
| 2314 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2315 | schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance); |
Nick Piggin | d6d5cfa | 2005-09-10 00:26:16 -0700 | [diff] [blame] | 2316 | |
| 2317 | nr_moved = 0; |
| 2318 | if (busiest->nr_running > 1) { |
| 2319 | /* Attempt to move tasks */ |
| 2320 | double_lock_balance(this_rq, busiest); |
| 2321 | nr_moved = move_tasks(this_rq, this_cpu, busiest, |
Nick Piggin | 8102679 | 2005-06-25 14:57:07 -0700 | [diff] [blame] | 2322 | imbalance, sd, NEWLY_IDLE, NULL); |
Nick Piggin | d6d5cfa | 2005-09-10 00:26:16 -0700 | [diff] [blame] | 2323 | spin_unlock(&busiest->lock); |
| 2324 | } |
| 2325 | |
Nick Piggin | 5969fe0 | 2005-09-10 00:26:19 -0700 | [diff] [blame] | 2326 | if (!nr_moved) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2327 | schedstat_inc(sd, lb_failed[NEWLY_IDLE]); |
Nick Piggin | 5969fe0 | 2005-09-10 00:26:19 -0700 | [diff] [blame] | 2328 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER) |
| 2329 | return -1; |
| 2330 | } else |
Nick Piggin | 16cfb1c | 2005-06-25 14:57:08 -0700 | [diff] [blame] | 2331 | sd->nr_balance_failed = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2332 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2333 | return nr_moved; |
Nick Piggin | 16cfb1c | 2005-06-25 14:57:08 -0700 | [diff] [blame] | 2334 | |
| 2335 | out_balanced: |
| 2336 | schedstat_inc(sd, lb_balanced[NEWLY_IDLE]); |
Nick Piggin | 5969fe0 | 2005-09-10 00:26:19 -0700 | [diff] [blame] | 2337 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER) |
| 2338 | return -1; |
Nick Piggin | 16cfb1c | 2005-06-25 14:57:08 -0700 | [diff] [blame] | 2339 | sd->nr_balance_failed = 0; |
| 2340 | return 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2341 | } |
| 2342 | |
| 2343 | /* |
| 2344 | * idle_balance is called by schedule() if this_cpu is about to become |
| 2345 | * idle. Attempts to pull tasks from other CPUs. |
| 2346 | */ |
Arjan van de Ven | 858119e | 2006-01-14 13:20:43 -0800 | [diff] [blame] | 2347 | static void idle_balance(int this_cpu, runqueue_t *this_rq) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2348 | { |
| 2349 | struct sched_domain *sd; |
| 2350 | |
| 2351 | for_each_domain(this_cpu, sd) { |
| 2352 | if (sd->flags & SD_BALANCE_NEWIDLE) { |
| 2353 | if (load_balance_newidle(this_cpu, this_rq, sd)) { |
| 2354 | /* We've pulled tasks over so stop searching */ |
| 2355 | break; |
| 2356 | } |
| 2357 | } |
| 2358 | } |
| 2359 | } |
| 2360 | |
| 2361 | /* |
| 2362 | * active_load_balance is run by migration threads. It pushes running tasks |
| 2363 | * off the busiest CPU onto idle CPUs. It requires at least 1 task to be |
| 2364 | * running on each physical CPU where possible, and avoids physical / |
| 2365 | * logical imbalances. |
| 2366 | * |
| 2367 | * Called with busiest_rq locked. |
| 2368 | */ |
| 2369 | static void active_load_balance(runqueue_t *busiest_rq, int busiest_cpu) |
| 2370 | { |
| 2371 | struct sched_domain *sd; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2372 | runqueue_t *target_rq; |
Nick Piggin | 3950745 | 2005-06-25 14:57:09 -0700 | [diff] [blame] | 2373 | int target_cpu = busiest_rq->push_cpu; |
| 2374 | |
| 2375 | if (busiest_rq->nr_running <= 1) |
| 2376 | /* no task to move */ |
| 2377 | return; |
| 2378 | |
| 2379 | target_rq = cpu_rq(target_cpu); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2380 | |
| 2381 | /* |
Nick Piggin | 3950745 | 2005-06-25 14:57:09 -0700 | [diff] [blame] | 2382 | * This condition is "impossible", if it occurs |
| 2383 | * we need to fix it. Originally reported by |
| 2384 | * Bjorn Helgaas on a 128-cpu setup. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2385 | */ |
Nick Piggin | 3950745 | 2005-06-25 14:57:09 -0700 | [diff] [blame] | 2386 | BUG_ON(busiest_rq == target_rq); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2387 | |
Nick Piggin | 3950745 | 2005-06-25 14:57:09 -0700 | [diff] [blame] | 2388 | /* move a task from busiest_rq to target_rq */ |
| 2389 | double_lock_balance(busiest_rq, target_rq); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2390 | |
Nick Piggin | 3950745 | 2005-06-25 14:57:09 -0700 | [diff] [blame] | 2391 | /* Search for an sd spanning us and the target CPU. */ |
| 2392 | for_each_domain(target_cpu, sd) |
| 2393 | if ((sd->flags & SD_LOAD_BALANCE) && |
| 2394 | cpu_isset(busiest_cpu, sd->span)) |
| 2395 | break; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2396 | |
Nick Piggin | 3950745 | 2005-06-25 14:57:09 -0700 | [diff] [blame] | 2397 | if (unlikely(sd == NULL)) |
| 2398 | goto out; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2399 | |
Nick Piggin | 3950745 | 2005-06-25 14:57:09 -0700 | [diff] [blame] | 2400 | schedstat_inc(sd, alb_cnt); |
| 2401 | |
| 2402 | if (move_tasks(target_rq, target_cpu, busiest_rq, 1, sd, SCHED_IDLE, NULL)) |
| 2403 | schedstat_inc(sd, alb_pushed); |
| 2404 | else |
| 2405 | schedstat_inc(sd, alb_failed); |
| 2406 | out: |
| 2407 | spin_unlock(&target_rq->lock); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2408 | } |
| 2409 | |
| 2410 | /* |
| 2411 | * rebalance_tick will get called every timer tick, on every CPU. |
| 2412 | * |
| 2413 | * It checks each scheduling domain to see if it is due to be balanced, |
| 2414 | * and initiates a balancing operation if so. |
| 2415 | * |
| 2416 | * Balancing parameters are set up in arch_init_sched_domains. |
| 2417 | */ |
| 2418 | |
| 2419 | /* Don't have all balancing operations going off at once */ |
| 2420 | #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS) |
| 2421 | |
| 2422 | static void rebalance_tick(int this_cpu, runqueue_t *this_rq, |
| 2423 | enum idle_type idle) |
| 2424 | { |
| 2425 | unsigned long old_load, this_load; |
| 2426 | unsigned long j = jiffies + CPU_OFFSET(this_cpu); |
| 2427 | struct sched_domain *sd; |
Nick Piggin | 7897986 | 2005-06-25 14:57:13 -0700 | [diff] [blame] | 2428 | int i; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2429 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2430 | this_load = this_rq->nr_running * SCHED_LOAD_SCALE; |
Nick Piggin | 7897986 | 2005-06-25 14:57:13 -0700 | [diff] [blame] | 2431 | /* Update our load */ |
| 2432 | for (i = 0; i < 3; i++) { |
| 2433 | unsigned long new_load = this_load; |
| 2434 | int scale = 1 << i; |
| 2435 | old_load = this_rq->cpu_load[i]; |
| 2436 | /* |
| 2437 | * Round up the averaging division if load is increasing. This |
| 2438 | * prevents us from getting stuck on 9 if the load is 10, for |
| 2439 | * example. |
| 2440 | */ |
| 2441 | if (new_load > old_load) |
| 2442 | new_load += scale-1; |
| 2443 | this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) / scale; |
| 2444 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2445 | |
| 2446 | for_each_domain(this_cpu, sd) { |
| 2447 | unsigned long interval; |
| 2448 | |
| 2449 | if (!(sd->flags & SD_LOAD_BALANCE)) |
| 2450 | continue; |
| 2451 | |
| 2452 | interval = sd->balance_interval; |
| 2453 | if (idle != SCHED_IDLE) |
| 2454 | interval *= sd->busy_factor; |
| 2455 | |
| 2456 | /* scale ms to jiffies */ |
| 2457 | interval = msecs_to_jiffies(interval); |
| 2458 | if (unlikely(!interval)) |
| 2459 | interval = 1; |
| 2460 | |
| 2461 | if (j - sd->last_balance >= interval) { |
| 2462 | if (load_balance(this_cpu, this_rq, sd, idle)) { |
Siddha, Suresh B | fa3b6dd | 2005-09-10 00:26:21 -0700 | [diff] [blame] | 2463 | /* |
| 2464 | * We've pulled tasks over so either we're no |
Nick Piggin | 5969fe0 | 2005-09-10 00:26:19 -0700 | [diff] [blame] | 2465 | * longer idle, or one of our SMT siblings is |
| 2466 | * not idle. |
| 2467 | */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2468 | idle = NOT_IDLE; |
| 2469 | } |
| 2470 | sd->last_balance += interval; |
| 2471 | } |
| 2472 | } |
| 2473 | } |
| 2474 | #else |
| 2475 | /* |
| 2476 | * on UP we do not need to balance between CPUs: |
| 2477 | */ |
| 2478 | static inline void rebalance_tick(int cpu, runqueue_t *rq, enum idle_type idle) |
| 2479 | { |
| 2480 | } |
| 2481 | static inline void idle_balance(int cpu, runqueue_t *rq) |
| 2482 | { |
| 2483 | } |
| 2484 | #endif |
| 2485 | |
| 2486 | static inline int wake_priority_sleeper(runqueue_t *rq) |
| 2487 | { |
| 2488 | int ret = 0; |
| 2489 | #ifdef CONFIG_SCHED_SMT |
| 2490 | spin_lock(&rq->lock); |
| 2491 | /* |
| 2492 | * If an SMT sibling task has been put to sleep for priority |
| 2493 | * reasons reschedule the idle task to see if it can now run. |
| 2494 | */ |
| 2495 | if (rq->nr_running) { |
| 2496 | resched_task(rq->idle); |
| 2497 | ret = 1; |
| 2498 | } |
| 2499 | spin_unlock(&rq->lock); |
| 2500 | #endif |
| 2501 | return ret; |
| 2502 | } |
| 2503 | |
| 2504 | DEFINE_PER_CPU(struct kernel_stat, kstat); |
| 2505 | |
| 2506 | EXPORT_PER_CPU_SYMBOL(kstat); |
| 2507 | |
| 2508 | /* |
| 2509 | * This is called on clock ticks and on context switches. |
| 2510 | * Bank in p->sched_time the ns elapsed since the last tick or switch. |
| 2511 | */ |
| 2512 | static inline void update_cpu_clock(task_t *p, runqueue_t *rq, |
| 2513 | unsigned long long now) |
| 2514 | { |
| 2515 | unsigned long long last = max(p->timestamp, rq->timestamp_last_tick); |
| 2516 | p->sched_time += now - last; |
| 2517 | } |
| 2518 | |
| 2519 | /* |
| 2520 | * Return current->sched_time plus any more ns on the sched_clock |
| 2521 | * that have not yet been banked. |
| 2522 | */ |
| 2523 | unsigned long long current_sched_time(const task_t *tsk) |
| 2524 | { |
| 2525 | unsigned long long ns; |
| 2526 | unsigned long flags; |
| 2527 | local_irq_save(flags); |
| 2528 | ns = max(tsk->timestamp, task_rq(tsk)->timestamp_last_tick); |
| 2529 | ns = tsk->sched_time + (sched_clock() - ns); |
| 2530 | local_irq_restore(flags); |
| 2531 | return ns; |
| 2532 | } |
| 2533 | |
| 2534 | /* |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2535 | * Account user cpu time to a process. |
| 2536 | * @p: the process that the cpu time gets accounted to |
| 2537 | * @hardirq_offset: the offset to subtract from hardirq_count() |
| 2538 | * @cputime: the cpu time spent in user space since the last update |
| 2539 | */ |
| 2540 | void account_user_time(struct task_struct *p, cputime_t cputime) |
| 2541 | { |
| 2542 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; |
| 2543 | cputime64_t tmp; |
| 2544 | |
| 2545 | p->utime = cputime_add(p->utime, cputime); |
| 2546 | |
| 2547 | /* Add user time to cpustat. */ |
| 2548 | tmp = cputime_to_cputime64(cputime); |
| 2549 | if (TASK_NICE(p) > 0) |
| 2550 | cpustat->nice = cputime64_add(cpustat->nice, tmp); |
| 2551 | else |
| 2552 | cpustat->user = cputime64_add(cpustat->user, tmp); |
| 2553 | } |
| 2554 | |
| 2555 | /* |
| 2556 | * Account system cpu time to a process. |
| 2557 | * @p: the process that the cpu time gets accounted to |
| 2558 | * @hardirq_offset: the offset to subtract from hardirq_count() |
| 2559 | * @cputime: the cpu time spent in kernel space since the last update |
| 2560 | */ |
| 2561 | void account_system_time(struct task_struct *p, int hardirq_offset, |
| 2562 | cputime_t cputime) |
| 2563 | { |
| 2564 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; |
| 2565 | runqueue_t *rq = this_rq(); |
| 2566 | cputime64_t tmp; |
| 2567 | |
| 2568 | p->stime = cputime_add(p->stime, cputime); |
| 2569 | |
| 2570 | /* Add system time to cpustat. */ |
| 2571 | tmp = cputime_to_cputime64(cputime); |
| 2572 | if (hardirq_count() - hardirq_offset) |
| 2573 | cpustat->irq = cputime64_add(cpustat->irq, tmp); |
| 2574 | else if (softirq_count()) |
| 2575 | cpustat->softirq = cputime64_add(cpustat->softirq, tmp); |
| 2576 | else if (p != rq->idle) |
| 2577 | cpustat->system = cputime64_add(cpustat->system, tmp); |
| 2578 | else if (atomic_read(&rq->nr_iowait) > 0) |
| 2579 | cpustat->iowait = cputime64_add(cpustat->iowait, tmp); |
| 2580 | else |
| 2581 | cpustat->idle = cputime64_add(cpustat->idle, tmp); |
| 2582 | /* Account for system time used */ |
| 2583 | acct_update_integrals(p); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2584 | } |
| 2585 | |
| 2586 | /* |
| 2587 | * Account for involuntary wait time. |
| 2588 | * @p: the process from which the cpu time has been stolen |
| 2589 | * @steal: the cpu time spent in involuntary wait |
| 2590 | */ |
| 2591 | void account_steal_time(struct task_struct *p, cputime_t steal) |
| 2592 | { |
| 2593 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; |
| 2594 | cputime64_t tmp = cputime_to_cputime64(steal); |
| 2595 | runqueue_t *rq = this_rq(); |
| 2596 | |
| 2597 | if (p == rq->idle) { |
| 2598 | p->stime = cputime_add(p->stime, steal); |
| 2599 | if (atomic_read(&rq->nr_iowait) > 0) |
| 2600 | cpustat->iowait = cputime64_add(cpustat->iowait, tmp); |
| 2601 | else |
| 2602 | cpustat->idle = cputime64_add(cpustat->idle, tmp); |
| 2603 | } else |
| 2604 | cpustat->steal = cputime64_add(cpustat->steal, tmp); |
| 2605 | } |
| 2606 | |
| 2607 | /* |
| 2608 | * This function gets called by the timer code, with HZ frequency. |
| 2609 | * We call it with interrupts disabled. |
| 2610 | * |
| 2611 | * It also gets called by the fork code, when changing the parent's |
| 2612 | * timeslices. |
| 2613 | */ |
| 2614 | void scheduler_tick(void) |
| 2615 | { |
| 2616 | int cpu = smp_processor_id(); |
| 2617 | runqueue_t *rq = this_rq(); |
| 2618 | task_t *p = current; |
| 2619 | unsigned long long now = sched_clock(); |
| 2620 | |
| 2621 | update_cpu_clock(p, rq, now); |
| 2622 | |
| 2623 | rq->timestamp_last_tick = now; |
| 2624 | |
| 2625 | if (p == rq->idle) { |
| 2626 | if (wake_priority_sleeper(rq)) |
| 2627 | goto out; |
| 2628 | rebalance_tick(cpu, rq, SCHED_IDLE); |
| 2629 | return; |
| 2630 | } |
| 2631 | |
| 2632 | /* Task might have expired already, but not scheduled off yet */ |
| 2633 | if (p->array != rq->active) { |
| 2634 | set_tsk_need_resched(p); |
| 2635 | goto out; |
| 2636 | } |
| 2637 | spin_lock(&rq->lock); |
| 2638 | /* |
| 2639 | * The task was running during this tick - update the |
| 2640 | * time slice counter. Note: we do not update a thread's |
| 2641 | * priority until it either goes to sleep or uses up its |
| 2642 | * timeslice. This makes it possible for interactive tasks |
| 2643 | * to use up their timeslices at their highest priority levels. |
| 2644 | */ |
| 2645 | if (rt_task(p)) { |
| 2646 | /* |
| 2647 | * RR tasks need a special form of timeslice management. |
| 2648 | * FIFO tasks have no timeslices. |
| 2649 | */ |
| 2650 | if ((p->policy == SCHED_RR) && !--p->time_slice) { |
| 2651 | p->time_slice = task_timeslice(p); |
| 2652 | p->first_time_slice = 0; |
| 2653 | set_tsk_need_resched(p); |
| 2654 | |
| 2655 | /* put it at the end of the queue: */ |
| 2656 | requeue_task(p, rq->active); |
| 2657 | } |
| 2658 | goto out_unlock; |
| 2659 | } |
| 2660 | if (!--p->time_slice) { |
| 2661 | dequeue_task(p, rq->active); |
| 2662 | set_tsk_need_resched(p); |
| 2663 | p->prio = effective_prio(p); |
| 2664 | p->time_slice = task_timeslice(p); |
| 2665 | p->first_time_slice = 0; |
| 2666 | |
| 2667 | if (!rq->expired_timestamp) |
| 2668 | rq->expired_timestamp = jiffies; |
Mike Galbraith | 5ce74ab | 2006-04-10 22:52:44 -0700 | [diff] [blame] | 2669 | if (!TASK_INTERACTIVE(p) || expired_starving(rq)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2670 | enqueue_task(p, rq->expired); |
| 2671 | if (p->static_prio < rq->best_expired_prio) |
| 2672 | rq->best_expired_prio = p->static_prio; |
| 2673 | } else |
| 2674 | enqueue_task(p, rq->active); |
| 2675 | } else { |
| 2676 | /* |
| 2677 | * Prevent a too long timeslice allowing a task to monopolize |
| 2678 | * the CPU. We do this by splitting up the timeslice into |
| 2679 | * smaller pieces. |
| 2680 | * |
| 2681 | * Note: this does not mean the task's timeslices expire or |
| 2682 | * get lost in any way, they just might be preempted by |
| 2683 | * another task of equal priority. (one with higher |
| 2684 | * priority would have preempted this task already.) We |
| 2685 | * requeue this task to the end of the list on this priority |
| 2686 | * level, which is in essence a round-robin of tasks with |
| 2687 | * equal priority. |
| 2688 | * |
| 2689 | * This only applies to tasks in the interactive |
| 2690 | * delta range with at least TIMESLICE_GRANULARITY to requeue. |
| 2691 | */ |
| 2692 | if (TASK_INTERACTIVE(p) && !((task_timeslice(p) - |
| 2693 | p->time_slice) % TIMESLICE_GRANULARITY(p)) && |
| 2694 | (p->time_slice >= TIMESLICE_GRANULARITY(p)) && |
| 2695 | (p->array == rq->active)) { |
| 2696 | |
| 2697 | requeue_task(p, rq->active); |
| 2698 | set_tsk_need_resched(p); |
| 2699 | } |
| 2700 | } |
| 2701 | out_unlock: |
| 2702 | spin_unlock(&rq->lock); |
| 2703 | out: |
| 2704 | rebalance_tick(cpu, rq, NOT_IDLE); |
| 2705 | } |
| 2706 | |
| 2707 | #ifdef CONFIG_SCHED_SMT |
Con Kolivas | fc38ed7 | 2005-09-10 00:26:08 -0700 | [diff] [blame] | 2708 | static inline void wakeup_busy_runqueue(runqueue_t *rq) |
| 2709 | { |
| 2710 | /* If an SMT runqueue is sleeping due to priority reasons wake it up */ |
| 2711 | if (rq->curr == rq->idle && rq->nr_running) |
| 2712 | resched_task(rq->idle); |
| 2713 | } |
| 2714 | |
Arjan van de Ven | 858119e | 2006-01-14 13:20:43 -0800 | [diff] [blame] | 2715 | static void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2716 | { |
Nick Piggin | 41c7ce9 | 2005-06-25 14:57:24 -0700 | [diff] [blame] | 2717 | struct sched_domain *tmp, *sd = NULL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2718 | cpumask_t sibling_map; |
| 2719 | int i; |
| 2720 | |
Nick Piggin | 41c7ce9 | 2005-06-25 14:57:24 -0700 | [diff] [blame] | 2721 | for_each_domain(this_cpu, tmp) |
| 2722 | if (tmp->flags & SD_SHARE_CPUPOWER) |
| 2723 | sd = tmp; |
| 2724 | |
| 2725 | if (!sd) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2726 | return; |
| 2727 | |
| 2728 | /* |
| 2729 | * Unlock the current runqueue because we have to lock in |
| 2730 | * CPU order to avoid deadlocks. Caller knows that we might |
| 2731 | * unlock. We keep IRQs disabled. |
| 2732 | */ |
| 2733 | spin_unlock(&this_rq->lock); |
| 2734 | |
| 2735 | sibling_map = sd->span; |
| 2736 | |
| 2737 | for_each_cpu_mask(i, sibling_map) |
| 2738 | spin_lock(&cpu_rq(i)->lock); |
| 2739 | /* |
| 2740 | * We clear this CPU from the mask. This both simplifies the |
| 2741 | * inner loop and keps this_rq locked when we exit: |
| 2742 | */ |
| 2743 | cpu_clear(this_cpu, sibling_map); |
| 2744 | |
| 2745 | for_each_cpu_mask(i, sibling_map) { |
| 2746 | runqueue_t *smt_rq = cpu_rq(i); |
| 2747 | |
Con Kolivas | fc38ed7 | 2005-09-10 00:26:08 -0700 | [diff] [blame] | 2748 | wakeup_busy_runqueue(smt_rq); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2749 | } |
| 2750 | |
| 2751 | for_each_cpu_mask(i, sibling_map) |
| 2752 | spin_unlock(&cpu_rq(i)->lock); |
| 2753 | /* |
| 2754 | * We exit with this_cpu's rq still held and IRQs |
| 2755 | * still disabled: |
| 2756 | */ |
| 2757 | } |
| 2758 | |
Ingo Molnar | 67f9a61 | 2005-09-10 00:26:16 -0700 | [diff] [blame] | 2759 | /* |
| 2760 | * number of 'lost' timeslices this task wont be able to fully |
| 2761 | * utilize, if another task runs on a sibling. This models the |
| 2762 | * slowdown effect of other tasks running on siblings: |
| 2763 | */ |
| 2764 | static inline unsigned long smt_slice(task_t *p, struct sched_domain *sd) |
| 2765 | { |
| 2766 | return p->time_slice * (100 - sd->per_cpu_gain) / 100; |
| 2767 | } |
| 2768 | |
Arjan van de Ven | 858119e | 2006-01-14 13:20:43 -0800 | [diff] [blame] | 2769 | static int dependent_sleeper(int this_cpu, runqueue_t *this_rq) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2770 | { |
Nick Piggin | 41c7ce9 | 2005-06-25 14:57:24 -0700 | [diff] [blame] | 2771 | struct sched_domain *tmp, *sd = NULL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2772 | cpumask_t sibling_map; |
| 2773 | prio_array_t *array; |
| 2774 | int ret = 0, i; |
| 2775 | task_t *p; |
| 2776 | |
Nick Piggin | 41c7ce9 | 2005-06-25 14:57:24 -0700 | [diff] [blame] | 2777 | for_each_domain(this_cpu, tmp) |
| 2778 | if (tmp->flags & SD_SHARE_CPUPOWER) |
| 2779 | sd = tmp; |
| 2780 | |
| 2781 | if (!sd) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2782 | return 0; |
| 2783 | |
| 2784 | /* |
| 2785 | * The same locking rules and details apply as for |
| 2786 | * wake_sleeping_dependent(): |
| 2787 | */ |
| 2788 | spin_unlock(&this_rq->lock); |
| 2789 | sibling_map = sd->span; |
| 2790 | for_each_cpu_mask(i, sibling_map) |
| 2791 | spin_lock(&cpu_rq(i)->lock); |
| 2792 | cpu_clear(this_cpu, sibling_map); |
| 2793 | |
| 2794 | /* |
| 2795 | * Establish next task to be run - it might have gone away because |
| 2796 | * we released the runqueue lock above: |
| 2797 | */ |
| 2798 | if (!this_rq->nr_running) |
| 2799 | goto out_unlock; |
| 2800 | array = this_rq->active; |
| 2801 | if (!array->nr_active) |
| 2802 | array = this_rq->expired; |
| 2803 | BUG_ON(!array->nr_active); |
| 2804 | |
| 2805 | p = list_entry(array->queue[sched_find_first_bit(array->bitmap)].next, |
| 2806 | task_t, run_list); |
| 2807 | |
| 2808 | for_each_cpu_mask(i, sibling_map) { |
| 2809 | runqueue_t *smt_rq = cpu_rq(i); |
| 2810 | task_t *smt_curr = smt_rq->curr; |
| 2811 | |
Con Kolivas | fc38ed7 | 2005-09-10 00:26:08 -0700 | [diff] [blame] | 2812 | /* Kernel threads do not participate in dependent sleeping */ |
| 2813 | if (!p->mm || !smt_curr->mm || rt_task(p)) |
| 2814 | goto check_smt_task; |
| 2815 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2816 | /* |
| 2817 | * If a user task with lower static priority than the |
| 2818 | * running task on the SMT sibling is trying to schedule, |
| 2819 | * delay it till there is proportionately less timeslice |
| 2820 | * left of the sibling task to prevent a lower priority |
| 2821 | * task from using an unfair proportion of the |
| 2822 | * physical cpu's resources. -ck |
| 2823 | */ |
Con Kolivas | fc38ed7 | 2005-09-10 00:26:08 -0700 | [diff] [blame] | 2824 | if (rt_task(smt_curr)) { |
| 2825 | /* |
| 2826 | * With real time tasks we run non-rt tasks only |
| 2827 | * per_cpu_gain% of the time. |
| 2828 | */ |
| 2829 | if ((jiffies % DEF_TIMESLICE) > |
| 2830 | (sd->per_cpu_gain * DEF_TIMESLICE / 100)) |
| 2831 | ret = 1; |
| 2832 | } else |
Ingo Molnar | 67f9a61 | 2005-09-10 00:26:16 -0700 | [diff] [blame] | 2833 | if (smt_curr->static_prio < p->static_prio && |
| 2834 | !TASK_PREEMPTS_CURR(p, smt_rq) && |
| 2835 | smt_slice(smt_curr, sd) > task_timeslice(p)) |
Con Kolivas | fc38ed7 | 2005-09-10 00:26:08 -0700 | [diff] [blame] | 2836 | ret = 1; |
| 2837 | |
| 2838 | check_smt_task: |
| 2839 | if ((!smt_curr->mm && smt_curr != smt_rq->idle) || |
| 2840 | rt_task(smt_curr)) |
| 2841 | continue; |
| 2842 | if (!p->mm) { |
| 2843 | wakeup_busy_runqueue(smt_rq); |
| 2844 | continue; |
| 2845 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2846 | |
| 2847 | /* |
Con Kolivas | fc38ed7 | 2005-09-10 00:26:08 -0700 | [diff] [blame] | 2848 | * Reschedule a lower priority task on the SMT sibling for |
| 2849 | * it to be put to sleep, or wake it up if it has been put to |
| 2850 | * sleep for priority reasons to see if it should run now. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2851 | */ |
Con Kolivas | fc38ed7 | 2005-09-10 00:26:08 -0700 | [diff] [blame] | 2852 | if (rt_task(p)) { |
| 2853 | if ((jiffies % DEF_TIMESLICE) > |
| 2854 | (sd->per_cpu_gain * DEF_TIMESLICE / 100)) |
| 2855 | resched_task(smt_curr); |
| 2856 | } else { |
Ingo Molnar | 67f9a61 | 2005-09-10 00:26:16 -0700 | [diff] [blame] | 2857 | if (TASK_PREEMPTS_CURR(p, smt_rq) && |
| 2858 | smt_slice(p, sd) > task_timeslice(smt_curr)) |
Con Kolivas | fc38ed7 | 2005-09-10 00:26:08 -0700 | [diff] [blame] | 2859 | resched_task(smt_curr); |
| 2860 | else |
| 2861 | wakeup_busy_runqueue(smt_rq); |
| 2862 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2863 | } |
| 2864 | out_unlock: |
| 2865 | for_each_cpu_mask(i, sibling_map) |
| 2866 | spin_unlock(&cpu_rq(i)->lock); |
| 2867 | return ret; |
| 2868 | } |
| 2869 | #else |
| 2870 | static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq) |
| 2871 | { |
| 2872 | } |
| 2873 | |
| 2874 | static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq) |
| 2875 | { |
| 2876 | return 0; |
| 2877 | } |
| 2878 | #endif |
| 2879 | |
| 2880 | #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT) |
| 2881 | |
| 2882 | void fastcall add_preempt_count(int val) |
| 2883 | { |
| 2884 | /* |
| 2885 | * Underflow? |
| 2886 | */ |
Jesper Juhl | be5b4fb | 2005-06-23 00:09:09 -0700 | [diff] [blame] | 2887 | BUG_ON((preempt_count() < 0)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2888 | preempt_count() += val; |
| 2889 | /* |
| 2890 | * Spinlock count overflowing soon? |
| 2891 | */ |
| 2892 | BUG_ON((preempt_count() & PREEMPT_MASK) >= PREEMPT_MASK-10); |
| 2893 | } |
| 2894 | EXPORT_SYMBOL(add_preempt_count); |
| 2895 | |
| 2896 | void fastcall sub_preempt_count(int val) |
| 2897 | { |
| 2898 | /* |
| 2899 | * Underflow? |
| 2900 | */ |
| 2901 | BUG_ON(val > preempt_count()); |
| 2902 | /* |
| 2903 | * Is the spinlock portion underflowing? |
| 2904 | */ |
| 2905 | BUG_ON((val < PREEMPT_MASK) && !(preempt_count() & PREEMPT_MASK)); |
| 2906 | preempt_count() -= val; |
| 2907 | } |
| 2908 | EXPORT_SYMBOL(sub_preempt_count); |
| 2909 | |
| 2910 | #endif |
| 2911 | |
Con Kolivas | 3dee386 | 2006-03-31 02:31:23 -0800 | [diff] [blame] | 2912 | static inline int interactive_sleep(enum sleep_type sleep_type) |
| 2913 | { |
| 2914 | return (sleep_type == SLEEP_INTERACTIVE || |
| 2915 | sleep_type == SLEEP_INTERRUPTED); |
| 2916 | } |
| 2917 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2918 | /* |
| 2919 | * schedule() is the main scheduler function. |
| 2920 | */ |
| 2921 | asmlinkage void __sched schedule(void) |
| 2922 | { |
| 2923 | long *switch_count; |
| 2924 | task_t *prev, *next; |
| 2925 | runqueue_t *rq; |
| 2926 | prio_array_t *array; |
| 2927 | struct list_head *queue; |
| 2928 | unsigned long long now; |
| 2929 | unsigned long run_time; |
Chen Shang | a3464a1 | 2005-06-25 14:57:31 -0700 | [diff] [blame] | 2930 | int cpu, idx, new_prio; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2931 | |
| 2932 | /* |
| 2933 | * Test if we are atomic. Since do_exit() needs to call into |
| 2934 | * schedule() atomically, we ignore that path for now. |
| 2935 | * Otherwise, whine if we are scheduling when we should not be. |
| 2936 | */ |
Andreas Mohr | 77e4bfb | 2006-03-27 01:15:20 -0800 | [diff] [blame] | 2937 | if (unlikely(in_atomic() && !current->exit_state)) { |
| 2938 | printk(KERN_ERR "BUG: scheduling while atomic: " |
| 2939 | "%s/0x%08x/%d\n", |
| 2940 | current->comm, preempt_count(), current->pid); |
| 2941 | dump_stack(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2942 | } |
| 2943 | profile_hit(SCHED_PROFILING, __builtin_return_address(0)); |
| 2944 | |
| 2945 | need_resched: |
| 2946 | preempt_disable(); |
| 2947 | prev = current; |
| 2948 | release_kernel_lock(prev); |
| 2949 | need_resched_nonpreemptible: |
| 2950 | rq = this_rq(); |
| 2951 | |
| 2952 | /* |
| 2953 | * The idle thread is not allowed to schedule! |
| 2954 | * Remove this check after it has been exercised a bit. |
| 2955 | */ |
| 2956 | if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) { |
| 2957 | printk(KERN_ERR "bad: scheduling from the idle thread!\n"); |
| 2958 | dump_stack(); |
| 2959 | } |
| 2960 | |
| 2961 | schedstat_inc(rq, sched_cnt); |
| 2962 | now = sched_clock(); |
Ingo Molnar | 238628e | 2005-04-18 10:58:36 -0700 | [diff] [blame] | 2963 | if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2964 | run_time = now - prev->timestamp; |
Ingo Molnar | 238628e | 2005-04-18 10:58:36 -0700 | [diff] [blame] | 2965 | if (unlikely((long long)(now - prev->timestamp) < 0)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2966 | run_time = 0; |
| 2967 | } else |
| 2968 | run_time = NS_MAX_SLEEP_AVG; |
| 2969 | |
| 2970 | /* |
| 2971 | * Tasks charged proportionately less run_time at high sleep_avg to |
| 2972 | * delay them losing their interactive status |
| 2973 | */ |
| 2974 | run_time /= (CURRENT_BONUS(prev) ? : 1); |
| 2975 | |
| 2976 | spin_lock_irq(&rq->lock); |
| 2977 | |
| 2978 | if (unlikely(prev->flags & PF_DEAD)) |
| 2979 | prev->state = EXIT_DEAD; |
| 2980 | |
| 2981 | switch_count = &prev->nivcsw; |
| 2982 | if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { |
| 2983 | switch_count = &prev->nvcsw; |
| 2984 | if (unlikely((prev->state & TASK_INTERRUPTIBLE) && |
| 2985 | unlikely(signal_pending(prev)))) |
| 2986 | prev->state = TASK_RUNNING; |
| 2987 | else { |
| 2988 | if (prev->state == TASK_UNINTERRUPTIBLE) |
| 2989 | rq->nr_uninterruptible++; |
| 2990 | deactivate_task(prev, rq); |
| 2991 | } |
| 2992 | } |
| 2993 | |
| 2994 | cpu = smp_processor_id(); |
| 2995 | if (unlikely(!rq->nr_running)) { |
| 2996 | go_idle: |
| 2997 | idle_balance(cpu, rq); |
| 2998 | if (!rq->nr_running) { |
| 2999 | next = rq->idle; |
| 3000 | rq->expired_timestamp = 0; |
| 3001 | wake_sleeping_dependent(cpu, rq); |
| 3002 | /* |
| 3003 | * wake_sleeping_dependent() might have released |
| 3004 | * the runqueue, so break out if we got new |
| 3005 | * tasks meanwhile: |
| 3006 | */ |
| 3007 | if (!rq->nr_running) |
| 3008 | goto switch_tasks; |
| 3009 | } |
| 3010 | } else { |
| 3011 | if (dependent_sleeper(cpu, rq)) { |
| 3012 | next = rq->idle; |
| 3013 | goto switch_tasks; |
| 3014 | } |
| 3015 | /* |
| 3016 | * dependent_sleeper() releases and reacquires the runqueue |
| 3017 | * lock, hence go into the idle loop if the rq went |
| 3018 | * empty meanwhile: |
| 3019 | */ |
| 3020 | if (unlikely(!rq->nr_running)) |
| 3021 | goto go_idle; |
| 3022 | } |
| 3023 | |
| 3024 | array = rq->active; |
| 3025 | if (unlikely(!array->nr_active)) { |
| 3026 | /* |
| 3027 | * Switch the active and expired arrays. |
| 3028 | */ |
| 3029 | schedstat_inc(rq, sched_switch); |
| 3030 | rq->active = rq->expired; |
| 3031 | rq->expired = array; |
| 3032 | array = rq->active; |
| 3033 | rq->expired_timestamp = 0; |
| 3034 | rq->best_expired_prio = MAX_PRIO; |
| 3035 | } |
| 3036 | |
| 3037 | idx = sched_find_first_bit(array->bitmap); |
| 3038 | queue = array->queue + idx; |
| 3039 | next = list_entry(queue->next, task_t, run_list); |
| 3040 | |
Con Kolivas | 3dee386 | 2006-03-31 02:31:23 -0800 | [diff] [blame] | 3041 | if (!rt_task(next) && interactive_sleep(next->sleep_type)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3042 | unsigned long long delta = now - next->timestamp; |
Ingo Molnar | 238628e | 2005-04-18 10:58:36 -0700 | [diff] [blame] | 3043 | if (unlikely((long long)(now - next->timestamp) < 0)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3044 | delta = 0; |
| 3045 | |
Con Kolivas | 3dee386 | 2006-03-31 02:31:23 -0800 | [diff] [blame] | 3046 | if (next->sleep_type == SLEEP_INTERACTIVE) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3047 | delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128; |
| 3048 | |
| 3049 | array = next->array; |
Chen Shang | a3464a1 | 2005-06-25 14:57:31 -0700 | [diff] [blame] | 3050 | new_prio = recalc_task_prio(next, next->timestamp + delta); |
| 3051 | |
| 3052 | if (unlikely(next->prio != new_prio)) { |
| 3053 | dequeue_task(next, array); |
| 3054 | next->prio = new_prio; |
| 3055 | enqueue_task(next, array); |
Con Kolivas | 7c4bb1f | 2006-03-31 02:31:29 -0800 | [diff] [blame] | 3056 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3057 | } |
Con Kolivas | 3dee386 | 2006-03-31 02:31:23 -0800 | [diff] [blame] | 3058 | next->sleep_type = SLEEP_NORMAL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3059 | switch_tasks: |
| 3060 | if (next == rq->idle) |
| 3061 | schedstat_inc(rq, sched_goidle); |
| 3062 | prefetch(next); |
Chen, Kenneth W | 383f283 | 2005-09-09 13:02:02 -0700 | [diff] [blame] | 3063 | prefetch_stack(next); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3064 | clear_tsk_need_resched(prev); |
| 3065 | rcu_qsctr_inc(task_cpu(prev)); |
| 3066 | |
| 3067 | update_cpu_clock(prev, rq, now); |
| 3068 | |
| 3069 | prev->sleep_avg -= run_time; |
| 3070 | if ((long)prev->sleep_avg <= 0) |
| 3071 | prev->sleep_avg = 0; |
| 3072 | prev->timestamp = prev->last_ran = now; |
| 3073 | |
| 3074 | sched_info_switch(prev, next); |
| 3075 | if (likely(prev != next)) { |
| 3076 | next->timestamp = now; |
| 3077 | rq->nr_switches++; |
| 3078 | rq->curr = next; |
| 3079 | ++*switch_count; |
| 3080 | |
Nick Piggin | 4866cde | 2005-06-25 14:57:23 -0700 | [diff] [blame] | 3081 | prepare_task_switch(rq, next); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3082 | prev = context_switch(rq, prev, next); |
| 3083 | barrier(); |
Nick Piggin | 4866cde | 2005-06-25 14:57:23 -0700 | [diff] [blame] | 3084 | /* |
| 3085 | * this_rq must be evaluated again because prev may have moved |
| 3086 | * CPUs since it called schedule(), thus the 'rq' on its stack |
| 3087 | * frame will be invalid. |
| 3088 | */ |
| 3089 | finish_task_switch(this_rq(), prev); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3090 | } else |
| 3091 | spin_unlock_irq(&rq->lock); |
| 3092 | |
| 3093 | prev = current; |
| 3094 | if (unlikely(reacquire_kernel_lock(prev) < 0)) |
| 3095 | goto need_resched_nonpreemptible; |
| 3096 | preempt_enable_no_resched(); |
| 3097 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) |
| 3098 | goto need_resched; |
| 3099 | } |
| 3100 | |
| 3101 | EXPORT_SYMBOL(schedule); |
| 3102 | |
| 3103 | #ifdef CONFIG_PREEMPT |
| 3104 | /* |
| 3105 | * this is is the entry point to schedule() from in-kernel preemption |
| 3106 | * off of preempt_enable. Kernel preemptions off return from interrupt |
| 3107 | * occur there and call schedule directly. |
| 3108 | */ |
| 3109 | asmlinkage void __sched preempt_schedule(void) |
| 3110 | { |
| 3111 | struct thread_info *ti = current_thread_info(); |
| 3112 | #ifdef CONFIG_PREEMPT_BKL |
| 3113 | struct task_struct *task = current; |
| 3114 | int saved_lock_depth; |
| 3115 | #endif |
| 3116 | /* |
| 3117 | * If there is a non-zero preempt_count or interrupts are disabled, |
| 3118 | * we do not want to preempt the current task. Just return.. |
| 3119 | */ |
| 3120 | if (unlikely(ti->preempt_count || irqs_disabled())) |
| 3121 | return; |
| 3122 | |
| 3123 | need_resched: |
| 3124 | add_preempt_count(PREEMPT_ACTIVE); |
| 3125 | /* |
| 3126 | * We keep the big kernel semaphore locked, but we |
| 3127 | * clear ->lock_depth so that schedule() doesnt |
| 3128 | * auto-release the semaphore: |
| 3129 | */ |
| 3130 | #ifdef CONFIG_PREEMPT_BKL |
| 3131 | saved_lock_depth = task->lock_depth; |
| 3132 | task->lock_depth = -1; |
| 3133 | #endif |
| 3134 | schedule(); |
| 3135 | #ifdef CONFIG_PREEMPT_BKL |
| 3136 | task->lock_depth = saved_lock_depth; |
| 3137 | #endif |
| 3138 | sub_preempt_count(PREEMPT_ACTIVE); |
| 3139 | |
| 3140 | /* we could miss a preemption opportunity between schedule and now */ |
| 3141 | barrier(); |
| 3142 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) |
| 3143 | goto need_resched; |
| 3144 | } |
| 3145 | |
| 3146 | EXPORT_SYMBOL(preempt_schedule); |
| 3147 | |
| 3148 | /* |
| 3149 | * this is is the entry point to schedule() from kernel preemption |
| 3150 | * off of irq context. |
| 3151 | * Note, that this is called and return with irqs disabled. This will |
| 3152 | * protect us against recursive calling from irq. |
| 3153 | */ |
| 3154 | asmlinkage void __sched preempt_schedule_irq(void) |
| 3155 | { |
| 3156 | struct thread_info *ti = current_thread_info(); |
| 3157 | #ifdef CONFIG_PREEMPT_BKL |
| 3158 | struct task_struct *task = current; |
| 3159 | int saved_lock_depth; |
| 3160 | #endif |
| 3161 | /* Catch callers which need to be fixed*/ |
| 3162 | BUG_ON(ti->preempt_count || !irqs_disabled()); |
| 3163 | |
| 3164 | need_resched: |
| 3165 | add_preempt_count(PREEMPT_ACTIVE); |
| 3166 | /* |
| 3167 | * We keep the big kernel semaphore locked, but we |
| 3168 | * clear ->lock_depth so that schedule() doesnt |
| 3169 | * auto-release the semaphore: |
| 3170 | */ |
| 3171 | #ifdef CONFIG_PREEMPT_BKL |
| 3172 | saved_lock_depth = task->lock_depth; |
| 3173 | task->lock_depth = -1; |
| 3174 | #endif |
| 3175 | local_irq_enable(); |
| 3176 | schedule(); |
| 3177 | local_irq_disable(); |
| 3178 | #ifdef CONFIG_PREEMPT_BKL |
| 3179 | task->lock_depth = saved_lock_depth; |
| 3180 | #endif |
| 3181 | sub_preempt_count(PREEMPT_ACTIVE); |
| 3182 | |
| 3183 | /* we could miss a preemption opportunity between schedule and now */ |
| 3184 | barrier(); |
| 3185 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) |
| 3186 | goto need_resched; |
| 3187 | } |
| 3188 | |
| 3189 | #endif /* CONFIG_PREEMPT */ |
| 3190 | |
Ingo Molnar | 95cdf3b | 2005-09-10 00:26:11 -0700 | [diff] [blame] | 3191 | int default_wake_function(wait_queue_t *curr, unsigned mode, int sync, |
| 3192 | void *key) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3193 | { |
Benjamin LaHaise | c43dc2f | 2005-06-23 00:10:27 -0700 | [diff] [blame] | 3194 | task_t *p = curr->private; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3195 | return try_to_wake_up(p, mode, sync); |
| 3196 | } |
| 3197 | |
| 3198 | EXPORT_SYMBOL(default_wake_function); |
| 3199 | |
| 3200 | /* |
| 3201 | * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just |
| 3202 | * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve |
| 3203 | * number) then we wake all the non-exclusive tasks and one exclusive task. |
| 3204 | * |
| 3205 | * There are circumstances in which we can try to wake a task which has already |
| 3206 | * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns |
| 3207 | * zero in this (rare) case, and we handle it by continuing to scan the queue. |
| 3208 | */ |
| 3209 | static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, |
| 3210 | int nr_exclusive, int sync, void *key) |
| 3211 | { |
| 3212 | struct list_head *tmp, *next; |
| 3213 | |
| 3214 | list_for_each_safe(tmp, next, &q->task_list) { |
| 3215 | wait_queue_t *curr; |
| 3216 | unsigned flags; |
| 3217 | curr = list_entry(tmp, wait_queue_t, task_list); |
| 3218 | flags = curr->flags; |
| 3219 | if (curr->func(curr, mode, sync, key) && |
| 3220 | (flags & WQ_FLAG_EXCLUSIVE) && |
| 3221 | !--nr_exclusive) |
| 3222 | break; |
| 3223 | } |
| 3224 | } |
| 3225 | |
| 3226 | /** |
| 3227 | * __wake_up - wake up threads blocked on a waitqueue. |
| 3228 | * @q: the waitqueue |
| 3229 | * @mode: which threads |
| 3230 | * @nr_exclusive: how many wake-one or wake-many threads to wake up |
Martin Waitz | 67be2dd | 2005-05-01 08:59:26 -0700 | [diff] [blame] | 3231 | * @key: is directly passed to the wakeup function |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3232 | */ |
| 3233 | void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode, |
Ingo Molnar | 95cdf3b | 2005-09-10 00:26:11 -0700 | [diff] [blame] | 3234 | int nr_exclusive, void *key) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3235 | { |
| 3236 | unsigned long flags; |
| 3237 | |
| 3238 | spin_lock_irqsave(&q->lock, flags); |
| 3239 | __wake_up_common(q, mode, nr_exclusive, 0, key); |
| 3240 | spin_unlock_irqrestore(&q->lock, flags); |
| 3241 | } |
| 3242 | |
| 3243 | EXPORT_SYMBOL(__wake_up); |
| 3244 | |
| 3245 | /* |
| 3246 | * Same as __wake_up but called with the spinlock in wait_queue_head_t held. |
| 3247 | */ |
| 3248 | void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode) |
| 3249 | { |
| 3250 | __wake_up_common(q, mode, 1, 0, NULL); |
| 3251 | } |
| 3252 | |
| 3253 | /** |
Martin Waitz | 67be2dd | 2005-05-01 08:59:26 -0700 | [diff] [blame] | 3254 | * __wake_up_sync - wake up threads blocked on a waitqueue. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3255 | * @q: the waitqueue |
| 3256 | * @mode: which threads |
| 3257 | * @nr_exclusive: how many wake-one or wake-many threads to wake up |
| 3258 | * |
| 3259 | * The sync wakeup differs that the waker knows that it will schedule |
| 3260 | * away soon, so while the target thread will be woken up, it will not |
| 3261 | * be migrated to another CPU - ie. the two threads are 'synchronized' |
| 3262 | * with each other. This can prevent needless bouncing between CPUs. |
| 3263 | * |
| 3264 | * On UP it can prevent extra preemption. |
| 3265 | */ |
Ingo Molnar | 95cdf3b | 2005-09-10 00:26:11 -0700 | [diff] [blame] | 3266 | void fastcall |
| 3267 | __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3268 | { |
| 3269 | unsigned long flags; |
| 3270 | int sync = 1; |
| 3271 | |
| 3272 | if (unlikely(!q)) |
| 3273 | return; |
| 3274 | |
| 3275 | if (unlikely(!nr_exclusive)) |
| 3276 | sync = 0; |
| 3277 | |
| 3278 | spin_lock_irqsave(&q->lock, flags); |
| 3279 | __wake_up_common(q, mode, nr_exclusive, sync, NULL); |
| 3280 | spin_unlock_irqrestore(&q->lock, flags); |
| 3281 | } |
| 3282 | EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ |
| 3283 | |
| 3284 | void fastcall complete(struct completion *x) |
| 3285 | { |
| 3286 | unsigned long flags; |
| 3287 | |
| 3288 | spin_lock_irqsave(&x->wait.lock, flags); |
| 3289 | x->done++; |
| 3290 | __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE, |
| 3291 | 1, 0, NULL); |
| 3292 | spin_unlock_irqrestore(&x->wait.lock, flags); |
| 3293 | } |
| 3294 | EXPORT_SYMBOL(complete); |
| 3295 | |
| 3296 | void fastcall complete_all(struct completion *x) |
| 3297 | { |
| 3298 | unsigned long flags; |
| 3299 | |
| 3300 | spin_lock_irqsave(&x->wait.lock, flags); |
| 3301 | x->done += UINT_MAX/2; |
| 3302 | __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE, |
| 3303 | 0, 0, NULL); |
| 3304 | spin_unlock_irqrestore(&x->wait.lock, flags); |
| 3305 | } |
| 3306 | EXPORT_SYMBOL(complete_all); |
| 3307 | |
| 3308 | void fastcall __sched wait_for_completion(struct completion *x) |
| 3309 | { |
| 3310 | might_sleep(); |
| 3311 | spin_lock_irq(&x->wait.lock); |
| 3312 | if (!x->done) { |
| 3313 | DECLARE_WAITQUEUE(wait, current); |
| 3314 | |
| 3315 | wait.flags |= WQ_FLAG_EXCLUSIVE; |
| 3316 | __add_wait_queue_tail(&x->wait, &wait); |
| 3317 | do { |
| 3318 | __set_current_state(TASK_UNINTERRUPTIBLE); |
| 3319 | spin_unlock_irq(&x->wait.lock); |
| 3320 | schedule(); |
| 3321 | spin_lock_irq(&x->wait.lock); |
| 3322 | } while (!x->done); |
| 3323 | __remove_wait_queue(&x->wait, &wait); |
| 3324 | } |
| 3325 | x->done--; |
| 3326 | spin_unlock_irq(&x->wait.lock); |
| 3327 | } |
| 3328 | EXPORT_SYMBOL(wait_for_completion); |
| 3329 | |
| 3330 | unsigned long fastcall __sched |
| 3331 | wait_for_completion_timeout(struct completion *x, unsigned long timeout) |
| 3332 | { |
| 3333 | might_sleep(); |
| 3334 | |
| 3335 | spin_lock_irq(&x->wait.lock); |
| 3336 | if (!x->done) { |
| 3337 | DECLARE_WAITQUEUE(wait, current); |
| 3338 | |
| 3339 | wait.flags |= WQ_FLAG_EXCLUSIVE; |
| 3340 | __add_wait_queue_tail(&x->wait, &wait); |
| 3341 | do { |
| 3342 | __set_current_state(TASK_UNINTERRUPTIBLE); |
| 3343 | spin_unlock_irq(&x->wait.lock); |
| 3344 | timeout = schedule_timeout(timeout); |
| 3345 | spin_lock_irq(&x->wait.lock); |
| 3346 | if (!timeout) { |
| 3347 | __remove_wait_queue(&x->wait, &wait); |
| 3348 | goto out; |
| 3349 | } |
| 3350 | } while (!x->done); |
| 3351 | __remove_wait_queue(&x->wait, &wait); |
| 3352 | } |
| 3353 | x->done--; |
| 3354 | out: |
| 3355 | spin_unlock_irq(&x->wait.lock); |
| 3356 | return timeout; |
| 3357 | } |
| 3358 | EXPORT_SYMBOL(wait_for_completion_timeout); |
| 3359 | |
| 3360 | int fastcall __sched wait_for_completion_interruptible(struct completion *x) |
| 3361 | { |
| 3362 | int ret = 0; |
| 3363 | |
| 3364 | might_sleep(); |
| 3365 | |
| 3366 | spin_lock_irq(&x->wait.lock); |
| 3367 | if (!x->done) { |
| 3368 | DECLARE_WAITQUEUE(wait, current); |
| 3369 | |
| 3370 | wait.flags |= WQ_FLAG_EXCLUSIVE; |
| 3371 | __add_wait_queue_tail(&x->wait, &wait); |
| 3372 | do { |
| 3373 | if (signal_pending(current)) { |
| 3374 | ret = -ERESTARTSYS; |
| 3375 | __remove_wait_queue(&x->wait, &wait); |
| 3376 | goto out; |
| 3377 | } |
| 3378 | __set_current_state(TASK_INTERRUPTIBLE); |
| 3379 | spin_unlock_irq(&x->wait.lock); |
| 3380 | schedule(); |
| 3381 | spin_lock_irq(&x->wait.lock); |
| 3382 | } while (!x->done); |
| 3383 | __remove_wait_queue(&x->wait, &wait); |
| 3384 | } |
| 3385 | x->done--; |
| 3386 | out: |
| 3387 | spin_unlock_irq(&x->wait.lock); |
| 3388 | |
| 3389 | return ret; |
| 3390 | } |
| 3391 | EXPORT_SYMBOL(wait_for_completion_interruptible); |
| 3392 | |
| 3393 | unsigned long fastcall __sched |
| 3394 | wait_for_completion_interruptible_timeout(struct completion *x, |
| 3395 | unsigned long timeout) |
| 3396 | { |
| 3397 | might_sleep(); |
| 3398 | |
| 3399 | spin_lock_irq(&x->wait.lock); |
| 3400 | if (!x->done) { |
| 3401 | DECLARE_WAITQUEUE(wait, current); |
| 3402 | |
| 3403 | wait.flags |= WQ_FLAG_EXCLUSIVE; |
| 3404 | __add_wait_queue_tail(&x->wait, &wait); |
| 3405 | do { |
| 3406 | if (signal_pending(current)) { |
| 3407 | timeout = -ERESTARTSYS; |
| 3408 | __remove_wait_queue(&x->wait, &wait); |
| 3409 | goto out; |
| 3410 | } |
| 3411 | __set_current_state(TASK_INTERRUPTIBLE); |
| 3412 | spin_unlock_irq(&x->wait.lock); |
| 3413 | timeout = schedule_timeout(timeout); |
| 3414 | spin_lock_irq(&x->wait.lock); |
| 3415 | if (!timeout) { |
| 3416 | __remove_wait_queue(&x->wait, &wait); |
| 3417 | goto out; |
| 3418 | } |
| 3419 | } while (!x->done); |
| 3420 | __remove_wait_queue(&x->wait, &wait); |
| 3421 | } |
| 3422 | x->done--; |
| 3423 | out: |
| 3424 | spin_unlock_irq(&x->wait.lock); |
| 3425 | return timeout; |
| 3426 | } |
| 3427 | EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); |
| 3428 | |
| 3429 | |
| 3430 | #define SLEEP_ON_VAR \ |
| 3431 | unsigned long flags; \ |
| 3432 | wait_queue_t wait; \ |
| 3433 | init_waitqueue_entry(&wait, current); |
| 3434 | |
| 3435 | #define SLEEP_ON_HEAD \ |
| 3436 | spin_lock_irqsave(&q->lock,flags); \ |
| 3437 | __add_wait_queue(q, &wait); \ |
| 3438 | spin_unlock(&q->lock); |
| 3439 | |
| 3440 | #define SLEEP_ON_TAIL \ |
| 3441 | spin_lock_irq(&q->lock); \ |
| 3442 | __remove_wait_queue(q, &wait); \ |
| 3443 | spin_unlock_irqrestore(&q->lock, flags); |
| 3444 | |
| 3445 | void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q) |
| 3446 | { |
| 3447 | SLEEP_ON_VAR |
| 3448 | |
| 3449 | current->state = TASK_INTERRUPTIBLE; |
| 3450 | |
| 3451 | SLEEP_ON_HEAD |
| 3452 | schedule(); |
| 3453 | SLEEP_ON_TAIL |
| 3454 | } |
| 3455 | |
| 3456 | EXPORT_SYMBOL(interruptible_sleep_on); |
| 3457 | |
Ingo Molnar | 95cdf3b | 2005-09-10 00:26:11 -0700 | [diff] [blame] | 3458 | long fastcall __sched |
| 3459 | interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3460 | { |
| 3461 | SLEEP_ON_VAR |
| 3462 | |
| 3463 | current->state = TASK_INTERRUPTIBLE; |
| 3464 | |
| 3465 | SLEEP_ON_HEAD |
| 3466 | timeout = schedule_timeout(timeout); |
| 3467 | SLEEP_ON_TAIL |
| 3468 | |
| 3469 | return timeout; |
| 3470 | } |
| 3471 | |
| 3472 | EXPORT_SYMBOL(interruptible_sleep_on_timeout); |
| 3473 | |
| 3474 | void fastcall __sched sleep_on(wait_queue_head_t *q) |
| 3475 | { |
| 3476 | SLEEP_ON_VAR |
| 3477 | |
| 3478 | current->state = TASK_UNINTERRUPTIBLE; |
| 3479 | |
| 3480 | SLEEP_ON_HEAD |
| 3481 | schedule(); |
| 3482 | SLEEP_ON_TAIL |
| 3483 | } |
| 3484 | |
| 3485 | EXPORT_SYMBOL(sleep_on); |
| 3486 | |
| 3487 | long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) |
| 3488 | { |
| 3489 | SLEEP_ON_VAR |
| 3490 | |
| 3491 | current->state = TASK_UNINTERRUPTIBLE; |
| 3492 | |
| 3493 | SLEEP_ON_HEAD |
| 3494 | timeout = schedule_timeout(timeout); |
| 3495 | SLEEP_ON_TAIL |
| 3496 | |
| 3497 | return timeout; |
| 3498 | } |
| 3499 | |
| 3500 | EXPORT_SYMBOL(sleep_on_timeout); |
| 3501 | |
| 3502 | void set_user_nice(task_t *p, long nice) |
| 3503 | { |
| 3504 | unsigned long flags; |
| 3505 | prio_array_t *array; |
| 3506 | runqueue_t *rq; |
| 3507 | int old_prio, new_prio, delta; |
| 3508 | |
| 3509 | if (TASK_NICE(p) == nice || nice < -20 || nice > 19) |
| 3510 | return; |
| 3511 | /* |
| 3512 | * We have to be careful, if called from sys_setpriority(), |
| 3513 | * the task might be in the middle of scheduling on another CPU. |
| 3514 | */ |
| 3515 | rq = task_rq_lock(p, &flags); |
| 3516 | /* |
| 3517 | * The RT priorities are set via sched_setscheduler(), but we still |
| 3518 | * allow the 'normal' nice value to be set - but as expected |
| 3519 | * it wont have any effect on scheduling until the task is |
Ingo Molnar | b0a9499 | 2006-01-14 13:20:41 -0800 | [diff] [blame] | 3520 | * not SCHED_NORMAL/SCHED_BATCH: |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3521 | */ |
| 3522 | if (rt_task(p)) { |
| 3523 | p->static_prio = NICE_TO_PRIO(nice); |
| 3524 | goto out_unlock; |
| 3525 | } |
| 3526 | array = p->array; |
Nick Piggin | a200057 | 2006-02-10 01:51:02 -0800 | [diff] [blame] | 3527 | if (array) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3528 | dequeue_task(p, array); |
| 3529 | |
| 3530 | old_prio = p->prio; |
| 3531 | new_prio = NICE_TO_PRIO(nice); |
| 3532 | delta = new_prio - old_prio; |
| 3533 | p->static_prio = NICE_TO_PRIO(nice); |
| 3534 | p->prio += delta; |
| 3535 | |
| 3536 | if (array) { |
| 3537 | enqueue_task(p, array); |
| 3538 | /* |
| 3539 | * If the task increased its priority or is running and |
| 3540 | * lowered its priority, then reschedule its CPU: |
| 3541 | */ |
| 3542 | if (delta < 0 || (delta > 0 && task_running(rq, p))) |
| 3543 | resched_task(rq->curr); |
| 3544 | } |
| 3545 | out_unlock: |
| 3546 | task_rq_unlock(rq, &flags); |
| 3547 | } |
| 3548 | |
| 3549 | EXPORT_SYMBOL(set_user_nice); |
| 3550 | |
Matt Mackall | e43379f | 2005-05-01 08:59:00 -0700 | [diff] [blame] | 3551 | /* |
| 3552 | * can_nice - check if a task can reduce its nice value |
| 3553 | * @p: task |
| 3554 | * @nice: nice value |
| 3555 | */ |
| 3556 | int can_nice(const task_t *p, const int nice) |
| 3557 | { |
Matt Mackall | 024f474 | 2005-08-18 11:24:19 -0700 | [diff] [blame] | 3558 | /* convert nice value [19,-20] to rlimit style value [1,40] */ |
| 3559 | int nice_rlim = 20 - nice; |
Matt Mackall | e43379f | 2005-05-01 08:59:00 -0700 | [diff] [blame] | 3560 | return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur || |
| 3561 | capable(CAP_SYS_NICE)); |
| 3562 | } |
| 3563 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3564 | #ifdef __ARCH_WANT_SYS_NICE |
| 3565 | |
| 3566 | /* |
| 3567 | * sys_nice - change the priority of the current process. |
| 3568 | * @increment: priority increment |
| 3569 | * |
| 3570 | * sys_setpriority is a more generic, but much slower function that |
| 3571 | * does similar things. |
| 3572 | */ |
| 3573 | asmlinkage long sys_nice(int increment) |
| 3574 | { |
| 3575 | int retval; |
| 3576 | long nice; |
| 3577 | |
| 3578 | /* |
| 3579 | * Setpriority might change our priority at the same moment. |
| 3580 | * We don't have to worry. Conceptually one call occurs first |
| 3581 | * and we have a single winner. |
| 3582 | */ |
Matt Mackall | e43379f | 2005-05-01 08:59:00 -0700 | [diff] [blame] | 3583 | if (increment < -40) |
| 3584 | increment = -40; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3585 | if (increment > 40) |
| 3586 | increment = 40; |
| 3587 | |
| 3588 | nice = PRIO_TO_NICE(current->static_prio) + increment; |
| 3589 | if (nice < -20) |
| 3590 | nice = -20; |
| 3591 | if (nice > 19) |
| 3592 | nice = 19; |
| 3593 | |
Matt Mackall | e43379f | 2005-05-01 08:59:00 -0700 | [diff] [blame] | 3594 | if (increment < 0 && !can_nice(current, nice)) |
| 3595 | return -EPERM; |
| 3596 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3597 | retval = security_task_setnice(current, nice); |
| 3598 | if (retval) |
| 3599 | return retval; |
| 3600 | |
| 3601 | set_user_nice(current, nice); |
| 3602 | return 0; |
| 3603 | } |
| 3604 | |
| 3605 | #endif |
| 3606 | |
| 3607 | /** |
| 3608 | * task_prio - return the priority value of a given task. |
| 3609 | * @p: the task in question. |
| 3610 | * |
| 3611 | * This is the priority value as seen by users in /proc. |
| 3612 | * RT tasks are offset by -200. Normal tasks are centered |
| 3613 | * around 0, value goes from -16 to +15. |
| 3614 | */ |
| 3615 | int task_prio(const task_t *p) |
| 3616 | { |
| 3617 | return p->prio - MAX_RT_PRIO; |
| 3618 | } |
| 3619 | |
| 3620 | /** |
| 3621 | * task_nice - return the nice value of a given task. |
| 3622 | * @p: the task in question. |
| 3623 | */ |
| 3624 | int task_nice(const task_t *p) |
| 3625 | { |
| 3626 | return TASK_NICE(p); |
| 3627 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3628 | EXPORT_SYMBOL_GPL(task_nice); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3629 | |
| 3630 | /** |
| 3631 | * idle_cpu - is a given cpu idle currently? |
| 3632 | * @cpu: the processor in question. |
| 3633 | */ |
| 3634 | int idle_cpu(int cpu) |
| 3635 | { |
| 3636 | return cpu_curr(cpu) == cpu_rq(cpu)->idle; |
| 3637 | } |
| 3638 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3639 | /** |
| 3640 | * idle_task - return the idle task for a given cpu. |
| 3641 | * @cpu: the processor in question. |
| 3642 | */ |
| 3643 | task_t *idle_task(int cpu) |
| 3644 | { |
| 3645 | return cpu_rq(cpu)->idle; |
| 3646 | } |
| 3647 | |
| 3648 | /** |
| 3649 | * find_process_by_pid - find a process with a matching PID value. |
| 3650 | * @pid: the pid in question. |
| 3651 | */ |
| 3652 | static inline task_t *find_process_by_pid(pid_t pid) |
| 3653 | { |
| 3654 | return pid ? find_task_by_pid(pid) : current; |
| 3655 | } |
| 3656 | |
| 3657 | /* Actually do priority change: must hold rq lock. */ |
| 3658 | static void __setscheduler(struct task_struct *p, int policy, int prio) |
| 3659 | { |
| 3660 | BUG_ON(p->array); |
| 3661 | p->policy = policy; |
| 3662 | p->rt_priority = prio; |
Ingo Molnar | b0a9499 | 2006-01-14 13:20:41 -0800 | [diff] [blame] | 3663 | if (policy != SCHED_NORMAL && policy != SCHED_BATCH) { |
Steven Rostedt | d46523e | 2005-07-25 16:28:39 -0400 | [diff] [blame] | 3664 | p->prio = MAX_RT_PRIO-1 - p->rt_priority; |
Ingo Molnar | b0a9499 | 2006-01-14 13:20:41 -0800 | [diff] [blame] | 3665 | } else { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3666 | p->prio = p->static_prio; |
Ingo Molnar | b0a9499 | 2006-01-14 13:20:41 -0800 | [diff] [blame] | 3667 | /* |
| 3668 | * SCHED_BATCH tasks are treated as perpetual CPU hogs: |
| 3669 | */ |
| 3670 | if (policy == SCHED_BATCH) |
| 3671 | p->sleep_avg = 0; |
| 3672 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3673 | } |
| 3674 | |
| 3675 | /** |
| 3676 | * sched_setscheduler - change the scheduling policy and/or RT priority of |
| 3677 | * a thread. |
| 3678 | * @p: the task in question. |
| 3679 | * @policy: new policy. |
| 3680 | * @param: structure containing the new RT priority. |
| 3681 | */ |
Ingo Molnar | 95cdf3b | 2005-09-10 00:26:11 -0700 | [diff] [blame] | 3682 | int sched_setscheduler(struct task_struct *p, int policy, |
| 3683 | struct sched_param *param) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3684 | { |
| 3685 | int retval; |
| 3686 | int oldprio, oldpolicy = -1; |
| 3687 | prio_array_t *array; |
| 3688 | unsigned long flags; |
| 3689 | runqueue_t *rq; |
| 3690 | |
| 3691 | recheck: |
| 3692 | /* double check policy once rq lock held */ |
| 3693 | if (policy < 0) |
| 3694 | policy = oldpolicy = p->policy; |
| 3695 | else if (policy != SCHED_FIFO && policy != SCHED_RR && |
Ingo Molnar | b0a9499 | 2006-01-14 13:20:41 -0800 | [diff] [blame] | 3696 | policy != SCHED_NORMAL && policy != SCHED_BATCH) |
| 3697 | return -EINVAL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3698 | /* |
| 3699 | * Valid priorities for SCHED_FIFO and SCHED_RR are |
Ingo Molnar | b0a9499 | 2006-01-14 13:20:41 -0800 | [diff] [blame] | 3700 | * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and |
| 3701 | * SCHED_BATCH is 0. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3702 | */ |
| 3703 | if (param->sched_priority < 0 || |
Ingo Molnar | 95cdf3b | 2005-09-10 00:26:11 -0700 | [diff] [blame] | 3704 | (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || |
Steven Rostedt | d46523e | 2005-07-25 16:28:39 -0400 | [diff] [blame] | 3705 | (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3706 | return -EINVAL; |
Ingo Molnar | b0a9499 | 2006-01-14 13:20:41 -0800 | [diff] [blame] | 3707 | if ((policy == SCHED_NORMAL || policy == SCHED_BATCH) |
| 3708 | != (param->sched_priority == 0)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3709 | return -EINVAL; |
| 3710 | |
Olivier Croquette | 37e4ab3 | 2005-06-25 14:57:32 -0700 | [diff] [blame] | 3711 | /* |
| 3712 | * Allow unprivileged RT tasks to decrease priority: |
| 3713 | */ |
| 3714 | if (!capable(CAP_SYS_NICE)) { |
Ingo Molnar | b0a9499 | 2006-01-14 13:20:41 -0800 | [diff] [blame] | 3715 | /* |
| 3716 | * can't change policy, except between SCHED_NORMAL |
| 3717 | * and SCHED_BATCH: |
| 3718 | */ |
| 3719 | if (((policy != SCHED_NORMAL && p->policy != SCHED_BATCH) && |
| 3720 | (policy != SCHED_BATCH && p->policy != SCHED_NORMAL)) && |
| 3721 | !p->signal->rlim[RLIMIT_RTPRIO].rlim_cur) |
Olivier Croquette | 37e4ab3 | 2005-06-25 14:57:32 -0700 | [diff] [blame] | 3722 | return -EPERM; |
| 3723 | /* can't increase priority */ |
Ingo Molnar | b0a9499 | 2006-01-14 13:20:41 -0800 | [diff] [blame] | 3724 | if ((policy != SCHED_NORMAL && policy != SCHED_BATCH) && |
Olivier Croquette | 37e4ab3 | 2005-06-25 14:57:32 -0700 | [diff] [blame] | 3725 | param->sched_priority > p->rt_priority && |
| 3726 | param->sched_priority > |
| 3727 | p->signal->rlim[RLIMIT_RTPRIO].rlim_cur) |
| 3728 | return -EPERM; |
| 3729 | /* can't change other user's priorities */ |
| 3730 | if ((current->euid != p->euid) && |
| 3731 | (current->euid != p->uid)) |
| 3732 | return -EPERM; |
| 3733 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3734 | |
| 3735 | retval = security_task_setscheduler(p, policy, param); |
| 3736 | if (retval) |
| 3737 | return retval; |
| 3738 | /* |
| 3739 | * To be able to change p->policy safely, the apropriate |
| 3740 | * runqueue lock must be held. |
| 3741 | */ |
| 3742 | rq = task_rq_lock(p, &flags); |
| 3743 | /* recheck policy now with rq lock held */ |
| 3744 | if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { |
| 3745 | policy = oldpolicy = -1; |
| 3746 | task_rq_unlock(rq, &flags); |
| 3747 | goto recheck; |
| 3748 | } |
| 3749 | array = p->array; |
| 3750 | if (array) |
| 3751 | deactivate_task(p, rq); |
| 3752 | oldprio = p->prio; |
| 3753 | __setscheduler(p, policy, param->sched_priority); |
| 3754 | if (array) { |
| 3755 | __activate_task(p, rq); |
| 3756 | /* |
| 3757 | * Reschedule if we are currently running on this runqueue and |
| 3758 | * our priority decreased, or if we are not currently running on |
| 3759 | * this runqueue and our priority is higher than the current's |
| 3760 | */ |
| 3761 | if (task_running(rq, p)) { |
| 3762 | if (p->prio > oldprio) |
| 3763 | resched_task(rq->curr); |
| 3764 | } else if (TASK_PREEMPTS_CURR(p, rq)) |
| 3765 | resched_task(rq->curr); |
| 3766 | } |
| 3767 | task_rq_unlock(rq, &flags); |
| 3768 | return 0; |
| 3769 | } |
| 3770 | EXPORT_SYMBOL_GPL(sched_setscheduler); |
| 3771 | |
Ingo Molnar | 95cdf3b | 2005-09-10 00:26:11 -0700 | [diff] [blame] | 3772 | static int |
| 3773 | do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3774 | { |
| 3775 | int retval; |
| 3776 | struct sched_param lparam; |
| 3777 | struct task_struct *p; |
| 3778 | |
| 3779 | if (!param || pid < 0) |
| 3780 | return -EINVAL; |
| 3781 | if (copy_from_user(&lparam, param, sizeof(struct sched_param))) |
| 3782 | return -EFAULT; |
| 3783 | read_lock_irq(&tasklist_lock); |
| 3784 | p = find_process_by_pid(pid); |
| 3785 | if (!p) { |
| 3786 | read_unlock_irq(&tasklist_lock); |
| 3787 | return -ESRCH; |
| 3788 | } |
| 3789 | retval = sched_setscheduler(p, policy, &lparam); |
| 3790 | read_unlock_irq(&tasklist_lock); |
| 3791 | return retval; |
| 3792 | } |
| 3793 | |
| 3794 | /** |
| 3795 | * sys_sched_setscheduler - set/change the scheduler policy and RT priority |
| 3796 | * @pid: the pid in question. |
| 3797 | * @policy: new policy. |
| 3798 | * @param: structure containing the new RT priority. |
| 3799 | */ |
| 3800 | asmlinkage long sys_sched_setscheduler(pid_t pid, int policy, |
| 3801 | struct sched_param __user *param) |
| 3802 | { |
Jason Baron | c21761f | 2006-01-18 17:43:03 -0800 | [diff] [blame] | 3803 | /* negative values for policy are not valid */ |
| 3804 | if (policy < 0) |
| 3805 | return -EINVAL; |
| 3806 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3807 | return do_sched_setscheduler(pid, policy, param); |
| 3808 | } |
| 3809 | |
| 3810 | /** |
| 3811 | * sys_sched_setparam - set/change the RT priority of a thread |
| 3812 | * @pid: the pid in question. |
| 3813 | * @param: structure containing the new RT priority. |
| 3814 | */ |
| 3815 | asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param) |
| 3816 | { |
| 3817 | return do_sched_setscheduler(pid, -1, param); |
| 3818 | } |
| 3819 | |
| 3820 | /** |
| 3821 | * sys_sched_getscheduler - get the policy (scheduling class) of a thread |
| 3822 | * @pid: the pid in question. |
| 3823 | */ |
| 3824 | asmlinkage long sys_sched_getscheduler(pid_t pid) |
| 3825 | { |
| 3826 | int retval = -EINVAL; |
| 3827 | task_t *p; |
| 3828 | |
| 3829 | if (pid < 0) |
| 3830 | goto out_nounlock; |
| 3831 | |
| 3832 | retval = -ESRCH; |
| 3833 | read_lock(&tasklist_lock); |
| 3834 | p = find_process_by_pid(pid); |
| 3835 | if (p) { |
| 3836 | retval = security_task_getscheduler(p); |
| 3837 | if (!retval) |
| 3838 | retval = p->policy; |
| 3839 | } |
| 3840 | read_unlock(&tasklist_lock); |
| 3841 | |
| 3842 | out_nounlock: |
| 3843 | return retval; |
| 3844 | } |
| 3845 | |
| 3846 | /** |
| 3847 | * sys_sched_getscheduler - get the RT priority of a thread |
| 3848 | * @pid: the pid in question. |
| 3849 | * @param: structure containing the RT priority. |
| 3850 | */ |
| 3851 | asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param) |
| 3852 | { |
| 3853 | struct sched_param lp; |
| 3854 | int retval = -EINVAL; |
| 3855 | task_t *p; |
| 3856 | |
| 3857 | if (!param || pid < 0) |
| 3858 | goto out_nounlock; |
| 3859 | |
| 3860 | read_lock(&tasklist_lock); |
| 3861 | p = find_process_by_pid(pid); |
| 3862 | retval = -ESRCH; |
| 3863 | if (!p) |
| 3864 | goto out_unlock; |
| 3865 | |
| 3866 | retval = security_task_getscheduler(p); |
| 3867 | if (retval) |
| 3868 | goto out_unlock; |
| 3869 | |
| 3870 | lp.sched_priority = p->rt_priority; |
| 3871 | read_unlock(&tasklist_lock); |
| 3872 | |
| 3873 | /* |
| 3874 | * This one might sleep, we cannot do it with a spinlock held ... |
| 3875 | */ |
| 3876 | retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; |
| 3877 | |
| 3878 | out_nounlock: |
| 3879 | return retval; |
| 3880 | |
| 3881 | out_unlock: |
| 3882 | read_unlock(&tasklist_lock); |
| 3883 | return retval; |
| 3884 | } |
| 3885 | |
| 3886 | long sched_setaffinity(pid_t pid, cpumask_t new_mask) |
| 3887 | { |
| 3888 | task_t *p; |
| 3889 | int retval; |
| 3890 | cpumask_t cpus_allowed; |
| 3891 | |
| 3892 | lock_cpu_hotplug(); |
| 3893 | read_lock(&tasklist_lock); |
| 3894 | |
| 3895 | p = find_process_by_pid(pid); |
| 3896 | if (!p) { |
| 3897 | read_unlock(&tasklist_lock); |
| 3898 | unlock_cpu_hotplug(); |
| 3899 | return -ESRCH; |
| 3900 | } |
| 3901 | |
| 3902 | /* |
| 3903 | * It is not safe to call set_cpus_allowed with the |
| 3904 | * tasklist_lock held. We will bump the task_struct's |
| 3905 | * usage count and then drop tasklist_lock. |
| 3906 | */ |
| 3907 | get_task_struct(p); |
| 3908 | read_unlock(&tasklist_lock); |
| 3909 | |
| 3910 | retval = -EPERM; |
| 3911 | if ((current->euid != p->euid) && (current->euid != p->uid) && |
| 3912 | !capable(CAP_SYS_NICE)) |
| 3913 | goto out_unlock; |
| 3914 | |
| 3915 | cpus_allowed = cpuset_cpus_allowed(p); |
| 3916 | cpus_and(new_mask, new_mask, cpus_allowed); |
| 3917 | retval = set_cpus_allowed(p, new_mask); |
| 3918 | |
| 3919 | out_unlock: |
| 3920 | put_task_struct(p); |
| 3921 | unlock_cpu_hotplug(); |
| 3922 | return retval; |
| 3923 | } |
| 3924 | |
| 3925 | static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, |
| 3926 | cpumask_t *new_mask) |
| 3927 | { |
| 3928 | if (len < sizeof(cpumask_t)) { |
| 3929 | memset(new_mask, 0, sizeof(cpumask_t)); |
| 3930 | } else if (len > sizeof(cpumask_t)) { |
| 3931 | len = sizeof(cpumask_t); |
| 3932 | } |
| 3933 | return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; |
| 3934 | } |
| 3935 | |
| 3936 | /** |
| 3937 | * sys_sched_setaffinity - set the cpu affinity of a process |
| 3938 | * @pid: pid of the process |
| 3939 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr |
| 3940 | * @user_mask_ptr: user-space pointer to the new cpu mask |
| 3941 | */ |
| 3942 | asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len, |
| 3943 | unsigned long __user *user_mask_ptr) |
| 3944 | { |
| 3945 | cpumask_t new_mask; |
| 3946 | int retval; |
| 3947 | |
| 3948 | retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask); |
| 3949 | if (retval) |
| 3950 | return retval; |
| 3951 | |
| 3952 | return sched_setaffinity(pid, new_mask); |
| 3953 | } |
| 3954 | |
| 3955 | /* |
| 3956 | * Represents all cpu's present in the system |
| 3957 | * In systems capable of hotplug, this map could dynamically grow |
| 3958 | * as new cpu's are detected in the system via any platform specific |
| 3959 | * method, such as ACPI for e.g. |
| 3960 | */ |
| 3961 | |
Andi Kleen | 4cef0c6 | 2006-01-11 22:44:57 +0100 | [diff] [blame] | 3962 | cpumask_t cpu_present_map __read_mostly; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3963 | EXPORT_SYMBOL(cpu_present_map); |
| 3964 | |
| 3965 | #ifndef CONFIG_SMP |
Andi Kleen | 4cef0c6 | 2006-01-11 22:44:57 +0100 | [diff] [blame] | 3966 | cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL; |
| 3967 | cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3968 | #endif |
| 3969 | |
| 3970 | long sched_getaffinity(pid_t pid, cpumask_t *mask) |
| 3971 | { |
| 3972 | int retval; |
| 3973 | task_t *p; |
| 3974 | |
| 3975 | lock_cpu_hotplug(); |
| 3976 | read_lock(&tasklist_lock); |
| 3977 | |
| 3978 | retval = -ESRCH; |
| 3979 | p = find_process_by_pid(pid); |
| 3980 | if (!p) |
| 3981 | goto out_unlock; |
| 3982 | |
| 3983 | retval = 0; |
Jack Steiner | 2f7016d | 2006-02-01 03:05:18 -0800 | [diff] [blame] | 3984 | cpus_and(*mask, p->cpus_allowed, cpu_online_map); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3985 | |
| 3986 | out_unlock: |
| 3987 | read_unlock(&tasklist_lock); |
| 3988 | unlock_cpu_hotplug(); |
| 3989 | if (retval) |
| 3990 | return retval; |
| 3991 | |
| 3992 | return 0; |
| 3993 | } |
| 3994 | |
| 3995 | /** |
| 3996 | * sys_sched_getaffinity - get the cpu affinity of a process |
| 3997 | * @pid: pid of the process |
| 3998 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr |
| 3999 | * @user_mask_ptr: user-space pointer to hold the current cpu mask |
| 4000 | */ |
| 4001 | asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len, |
| 4002 | unsigned long __user *user_mask_ptr) |
| 4003 | { |
| 4004 | int ret; |
| 4005 | cpumask_t mask; |
| 4006 | |
| 4007 | if (len < sizeof(cpumask_t)) |
| 4008 | return -EINVAL; |
| 4009 | |
| 4010 | ret = sched_getaffinity(pid, &mask); |
| 4011 | if (ret < 0) |
| 4012 | return ret; |
| 4013 | |
| 4014 | if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t))) |
| 4015 | return -EFAULT; |
| 4016 | |
| 4017 | return sizeof(cpumask_t); |
| 4018 | } |
| 4019 | |
| 4020 | /** |
| 4021 | * sys_sched_yield - yield the current processor to other threads. |
| 4022 | * |
| 4023 | * this function yields the current CPU by moving the calling thread |
| 4024 | * to the expired array. If there are no other threads running on this |
| 4025 | * CPU then this function will return. |
| 4026 | */ |
| 4027 | asmlinkage long sys_sched_yield(void) |
| 4028 | { |
| 4029 | runqueue_t *rq = this_rq_lock(); |
| 4030 | prio_array_t *array = current->array; |
| 4031 | prio_array_t *target = rq->expired; |
| 4032 | |
| 4033 | schedstat_inc(rq, yld_cnt); |
| 4034 | /* |
| 4035 | * We implement yielding by moving the task into the expired |
| 4036 | * queue. |
| 4037 | * |
| 4038 | * (special rule: RT tasks will just roundrobin in the active |
| 4039 | * array.) |
| 4040 | */ |
| 4041 | if (rt_task(current)) |
| 4042 | target = rq->active; |
| 4043 | |
Renaud Lienhart | 5927ad7 | 2005-09-10 00:26:20 -0700 | [diff] [blame] | 4044 | if (array->nr_active == 1) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4045 | schedstat_inc(rq, yld_act_empty); |
| 4046 | if (!rq->expired->nr_active) |
| 4047 | schedstat_inc(rq, yld_both_empty); |
| 4048 | } else if (!rq->expired->nr_active) |
| 4049 | schedstat_inc(rq, yld_exp_empty); |
| 4050 | |
| 4051 | if (array != target) { |
| 4052 | dequeue_task(current, array); |
| 4053 | enqueue_task(current, target); |
| 4054 | } else |
| 4055 | /* |
| 4056 | * requeue_task is cheaper so perform that if possible. |
| 4057 | */ |
| 4058 | requeue_task(current, array); |
| 4059 | |
| 4060 | /* |
| 4061 | * Since we are going to call schedule() anyway, there's |
| 4062 | * no need to preempt or enable interrupts: |
| 4063 | */ |
| 4064 | __release(rq->lock); |
| 4065 | _raw_spin_unlock(&rq->lock); |
| 4066 | preempt_enable_no_resched(); |
| 4067 | |
| 4068 | schedule(); |
| 4069 | |
| 4070 | return 0; |
| 4071 | } |
| 4072 | |
| 4073 | static inline void __cond_resched(void) |
| 4074 | { |
Ingo Molnar | 5bbcfd9 | 2005-07-07 17:57:04 -0700 | [diff] [blame] | 4075 | /* |
| 4076 | * The BKS might be reacquired before we have dropped |
| 4077 | * PREEMPT_ACTIVE, which could trigger a second |
| 4078 | * cond_resched() call. |
| 4079 | */ |
| 4080 | if (unlikely(preempt_count())) |
| 4081 | return; |
Linus Torvalds | 8ba7b0a | 2006-03-06 17:38:49 -0800 | [diff] [blame] | 4082 | if (unlikely(system_state != SYSTEM_RUNNING)) |
| 4083 | return; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4084 | do { |
| 4085 | add_preempt_count(PREEMPT_ACTIVE); |
| 4086 | schedule(); |
| 4087 | sub_preempt_count(PREEMPT_ACTIVE); |
| 4088 | } while (need_resched()); |
| 4089 | } |
| 4090 | |
| 4091 | int __sched cond_resched(void) |
| 4092 | { |
| 4093 | if (need_resched()) { |
| 4094 | __cond_resched(); |
| 4095 | return 1; |
| 4096 | } |
| 4097 | return 0; |
| 4098 | } |
| 4099 | |
| 4100 | EXPORT_SYMBOL(cond_resched); |
| 4101 | |
| 4102 | /* |
| 4103 | * cond_resched_lock() - if a reschedule is pending, drop the given lock, |
| 4104 | * call schedule, and on return reacquire the lock. |
| 4105 | * |
| 4106 | * This works OK both with and without CONFIG_PREEMPT. We do strange low-level |
| 4107 | * operations here to prevent schedule() from being called twice (once via |
| 4108 | * spin_unlock(), once by hand). |
| 4109 | */ |
Ingo Molnar | 95cdf3b | 2005-09-10 00:26:11 -0700 | [diff] [blame] | 4110 | int cond_resched_lock(spinlock_t *lock) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4111 | { |
Jan Kara | 6df3cec | 2005-06-13 15:52:32 -0700 | [diff] [blame] | 4112 | int ret = 0; |
| 4113 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4114 | if (need_lockbreak(lock)) { |
| 4115 | spin_unlock(lock); |
| 4116 | cpu_relax(); |
Jan Kara | 6df3cec | 2005-06-13 15:52:32 -0700 | [diff] [blame] | 4117 | ret = 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4118 | spin_lock(lock); |
| 4119 | } |
| 4120 | if (need_resched()) { |
| 4121 | _raw_spin_unlock(lock); |
| 4122 | preempt_enable_no_resched(); |
| 4123 | __cond_resched(); |
Jan Kara | 6df3cec | 2005-06-13 15:52:32 -0700 | [diff] [blame] | 4124 | ret = 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4125 | spin_lock(lock); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4126 | } |
Jan Kara | 6df3cec | 2005-06-13 15:52:32 -0700 | [diff] [blame] | 4127 | return ret; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4128 | } |
| 4129 | |
| 4130 | EXPORT_SYMBOL(cond_resched_lock); |
| 4131 | |
| 4132 | int __sched cond_resched_softirq(void) |
| 4133 | { |
| 4134 | BUG_ON(!in_softirq()); |
| 4135 | |
| 4136 | if (need_resched()) { |
| 4137 | __local_bh_enable(); |
| 4138 | __cond_resched(); |
| 4139 | local_bh_disable(); |
| 4140 | return 1; |
| 4141 | } |
| 4142 | return 0; |
| 4143 | } |
| 4144 | |
| 4145 | EXPORT_SYMBOL(cond_resched_softirq); |
| 4146 | |
| 4147 | |
| 4148 | /** |
| 4149 | * yield - yield the current processor to other threads. |
| 4150 | * |
| 4151 | * this is a shortcut for kernel-space yielding - it marks the |
| 4152 | * thread runnable and calls sys_sched_yield(). |
| 4153 | */ |
| 4154 | void __sched yield(void) |
| 4155 | { |
| 4156 | set_current_state(TASK_RUNNING); |
| 4157 | sys_sched_yield(); |
| 4158 | } |
| 4159 | |
| 4160 | EXPORT_SYMBOL(yield); |
| 4161 | |
| 4162 | /* |
| 4163 | * This task is about to go to sleep on IO. Increment rq->nr_iowait so |
| 4164 | * that process accounting knows that this is a task in IO wait state. |
| 4165 | * |
| 4166 | * But don't do that if it is a deliberate, throttling IO wait (this task |
| 4167 | * has set its backing_dev_info: the queue against which it should throttle) |
| 4168 | */ |
| 4169 | void __sched io_schedule(void) |
| 4170 | { |
Ingo Molnar | 39c715b | 2005-06-21 17:14:34 -0700 | [diff] [blame] | 4171 | struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id()); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4172 | |
| 4173 | atomic_inc(&rq->nr_iowait); |
| 4174 | schedule(); |
| 4175 | atomic_dec(&rq->nr_iowait); |
| 4176 | } |
| 4177 | |
| 4178 | EXPORT_SYMBOL(io_schedule); |
| 4179 | |
| 4180 | long __sched io_schedule_timeout(long timeout) |
| 4181 | { |
Ingo Molnar | 39c715b | 2005-06-21 17:14:34 -0700 | [diff] [blame] | 4182 | struct runqueue *rq = &per_cpu(runqueues, raw_smp_processor_id()); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4183 | long ret; |
| 4184 | |
| 4185 | atomic_inc(&rq->nr_iowait); |
| 4186 | ret = schedule_timeout(timeout); |
| 4187 | atomic_dec(&rq->nr_iowait); |
| 4188 | return ret; |
| 4189 | } |
| 4190 | |
| 4191 | /** |
| 4192 | * sys_sched_get_priority_max - return maximum RT priority. |
| 4193 | * @policy: scheduling class. |
| 4194 | * |
| 4195 | * this syscall returns the maximum rt_priority that can be used |
| 4196 | * by a given scheduling class. |
| 4197 | */ |
| 4198 | asmlinkage long sys_sched_get_priority_max(int policy) |
| 4199 | { |
| 4200 | int ret = -EINVAL; |
| 4201 | |
| 4202 | switch (policy) { |
| 4203 | case SCHED_FIFO: |
| 4204 | case SCHED_RR: |
| 4205 | ret = MAX_USER_RT_PRIO-1; |
| 4206 | break; |
| 4207 | case SCHED_NORMAL: |
Ingo Molnar | b0a9499 | 2006-01-14 13:20:41 -0800 | [diff] [blame] | 4208 | case SCHED_BATCH: |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4209 | ret = 0; |
| 4210 | break; |
| 4211 | } |
| 4212 | return ret; |
| 4213 | } |
| 4214 | |
| 4215 | /** |
| 4216 | * sys_sched_get_priority_min - return minimum RT priority. |
| 4217 | * @policy: scheduling class. |
| 4218 | * |
| 4219 | * this syscall returns the minimum rt_priority that can be used |
| 4220 | * by a given scheduling class. |
| 4221 | */ |
| 4222 | asmlinkage long sys_sched_get_priority_min(int policy) |
| 4223 | { |
| 4224 | int ret = -EINVAL; |
| 4225 | |
| 4226 | switch (policy) { |
| 4227 | case SCHED_FIFO: |
| 4228 | case SCHED_RR: |
| 4229 | ret = 1; |
| 4230 | break; |
| 4231 | case SCHED_NORMAL: |
Ingo Molnar | b0a9499 | 2006-01-14 13:20:41 -0800 | [diff] [blame] | 4232 | case SCHED_BATCH: |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4233 | ret = 0; |
| 4234 | } |
| 4235 | return ret; |
| 4236 | } |
| 4237 | |
| 4238 | /** |
| 4239 | * sys_sched_rr_get_interval - return the default timeslice of a process. |
| 4240 | * @pid: pid of the process. |
| 4241 | * @interval: userspace pointer to the timeslice value. |
| 4242 | * |
| 4243 | * this syscall writes the default timeslice value of a given process |
| 4244 | * into the user-space timespec buffer. A value of '0' means infinity. |
| 4245 | */ |
| 4246 | asmlinkage |
| 4247 | long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval) |
| 4248 | { |
| 4249 | int retval = -EINVAL; |
| 4250 | struct timespec t; |
| 4251 | task_t *p; |
| 4252 | |
| 4253 | if (pid < 0) |
| 4254 | goto out_nounlock; |
| 4255 | |
| 4256 | retval = -ESRCH; |
| 4257 | read_lock(&tasklist_lock); |
| 4258 | p = find_process_by_pid(pid); |
| 4259 | if (!p) |
| 4260 | goto out_unlock; |
| 4261 | |
| 4262 | retval = security_task_getscheduler(p); |
| 4263 | if (retval) |
| 4264 | goto out_unlock; |
| 4265 | |
| 4266 | jiffies_to_timespec(p->policy & SCHED_FIFO ? |
| 4267 | 0 : task_timeslice(p), &t); |
| 4268 | read_unlock(&tasklist_lock); |
| 4269 | retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; |
| 4270 | out_nounlock: |
| 4271 | return retval; |
| 4272 | out_unlock: |
| 4273 | read_unlock(&tasklist_lock); |
| 4274 | return retval; |
| 4275 | } |
| 4276 | |
| 4277 | static inline struct task_struct *eldest_child(struct task_struct *p) |
| 4278 | { |
| 4279 | if (list_empty(&p->children)) return NULL; |
| 4280 | return list_entry(p->children.next,struct task_struct,sibling); |
| 4281 | } |
| 4282 | |
| 4283 | static inline struct task_struct *older_sibling(struct task_struct *p) |
| 4284 | { |
| 4285 | if (p->sibling.prev==&p->parent->children) return NULL; |
| 4286 | return list_entry(p->sibling.prev,struct task_struct,sibling); |
| 4287 | } |
| 4288 | |
| 4289 | static inline struct task_struct *younger_sibling(struct task_struct *p) |
| 4290 | { |
| 4291 | if (p->sibling.next==&p->parent->children) return NULL; |
| 4292 | return list_entry(p->sibling.next,struct task_struct,sibling); |
| 4293 | } |
| 4294 | |
Ingo Molnar | 95cdf3b | 2005-09-10 00:26:11 -0700 | [diff] [blame] | 4295 | static void show_task(task_t *p) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4296 | { |
| 4297 | task_t *relative; |
| 4298 | unsigned state; |
| 4299 | unsigned long free = 0; |
| 4300 | static const char *stat_nam[] = { "R", "S", "D", "T", "t", "Z", "X" }; |
| 4301 | |
| 4302 | printk("%-13.13s ", p->comm); |
| 4303 | state = p->state ? __ffs(p->state) + 1 : 0; |
| 4304 | if (state < ARRAY_SIZE(stat_nam)) |
| 4305 | printk(stat_nam[state]); |
| 4306 | else |
| 4307 | printk("?"); |
| 4308 | #if (BITS_PER_LONG == 32) |
| 4309 | if (state == TASK_RUNNING) |
| 4310 | printk(" running "); |
| 4311 | else |
| 4312 | printk(" %08lX ", thread_saved_pc(p)); |
| 4313 | #else |
| 4314 | if (state == TASK_RUNNING) |
| 4315 | printk(" running task "); |
| 4316 | else |
| 4317 | printk(" %016lx ", thread_saved_pc(p)); |
| 4318 | #endif |
| 4319 | #ifdef CONFIG_DEBUG_STACK_USAGE |
| 4320 | { |
Al Viro | 10ebffd | 2005-11-13 16:06:56 -0800 | [diff] [blame] | 4321 | unsigned long *n = end_of_stack(p); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4322 | while (!*n) |
| 4323 | n++; |
Al Viro | 10ebffd | 2005-11-13 16:06:56 -0800 | [diff] [blame] | 4324 | free = (unsigned long)n - (unsigned long)end_of_stack(p); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4325 | } |
| 4326 | #endif |
| 4327 | printk("%5lu %5d %6d ", free, p->pid, p->parent->pid); |
| 4328 | if ((relative = eldest_child(p))) |
| 4329 | printk("%5d ", relative->pid); |
| 4330 | else |
| 4331 | printk(" "); |
| 4332 | if ((relative = younger_sibling(p))) |
| 4333 | printk("%7d", relative->pid); |
| 4334 | else |
| 4335 | printk(" "); |
| 4336 | if ((relative = older_sibling(p))) |
| 4337 | printk(" %5d", relative->pid); |
| 4338 | else |
| 4339 | printk(" "); |
| 4340 | if (!p->mm) |
| 4341 | printk(" (L-TLB)\n"); |
| 4342 | else |
| 4343 | printk(" (NOTLB)\n"); |
| 4344 | |
| 4345 | if (state != TASK_RUNNING) |
| 4346 | show_stack(p, NULL); |
| 4347 | } |
| 4348 | |
| 4349 | void show_state(void) |
| 4350 | { |
| 4351 | task_t *g, *p; |
| 4352 | |
| 4353 | #if (BITS_PER_LONG == 32) |
| 4354 | printk("\n" |
| 4355 | " sibling\n"); |
| 4356 | printk(" task PC pid father child younger older\n"); |
| 4357 | #else |
| 4358 | printk("\n" |
| 4359 | " sibling\n"); |
| 4360 | printk(" task PC pid father child younger older\n"); |
| 4361 | #endif |
| 4362 | read_lock(&tasklist_lock); |
| 4363 | do_each_thread(g, p) { |
| 4364 | /* |
| 4365 | * reset the NMI-timeout, listing all files on a slow |
| 4366 | * console might take alot of time: |
| 4367 | */ |
| 4368 | touch_nmi_watchdog(); |
| 4369 | show_task(p); |
| 4370 | } while_each_thread(g, p); |
| 4371 | |
| 4372 | read_unlock(&tasklist_lock); |
Ingo Molnar | de5097c | 2006-01-09 15:59:21 -0800 | [diff] [blame] | 4373 | mutex_debug_show_all_locks(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4374 | } |
| 4375 | |
Ingo Molnar | f340c0d | 2005-06-28 16:40:42 +0200 | [diff] [blame] | 4376 | /** |
| 4377 | * init_idle - set up an idle thread for a given CPU |
| 4378 | * @idle: task in question |
| 4379 | * @cpu: cpu the idle task belongs to |
| 4380 | * |
| 4381 | * NOTE: this function does not set the idle thread's NEED_RESCHED |
| 4382 | * flag, to make booting more robust. |
| 4383 | */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4384 | void __devinit init_idle(task_t *idle, int cpu) |
| 4385 | { |
| 4386 | runqueue_t *rq = cpu_rq(cpu); |
| 4387 | unsigned long flags; |
| 4388 | |
Ingo Molnar | 81c29a8 | 2006-03-07 21:55:27 -0800 | [diff] [blame] | 4389 | idle->timestamp = sched_clock(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4390 | idle->sleep_avg = 0; |
| 4391 | idle->array = NULL; |
| 4392 | idle->prio = MAX_PRIO; |
| 4393 | idle->state = TASK_RUNNING; |
| 4394 | idle->cpus_allowed = cpumask_of_cpu(cpu); |
| 4395 | set_task_cpu(idle, cpu); |
| 4396 | |
| 4397 | spin_lock_irqsave(&rq->lock, flags); |
| 4398 | rq->curr = rq->idle = idle; |
Nick Piggin | 4866cde | 2005-06-25 14:57:23 -0700 | [diff] [blame] | 4399 | #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) |
| 4400 | idle->oncpu = 1; |
| 4401 | #endif |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4402 | spin_unlock_irqrestore(&rq->lock, flags); |
| 4403 | |
| 4404 | /* Set the preempt count _outside_ the spinlocks! */ |
| 4405 | #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL) |
Al Viro | a1261f5 | 2005-11-13 16:06:55 -0800 | [diff] [blame] | 4406 | task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4407 | #else |
Al Viro | a1261f5 | 2005-11-13 16:06:55 -0800 | [diff] [blame] | 4408 | task_thread_info(idle)->preempt_count = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4409 | #endif |
| 4410 | } |
| 4411 | |
| 4412 | /* |
| 4413 | * In a system that switches off the HZ timer nohz_cpu_mask |
| 4414 | * indicates which cpus entered this state. This is used |
| 4415 | * in the rcu update to wait only for active cpus. For system |
| 4416 | * which do not switch off the HZ timer nohz_cpu_mask should |
| 4417 | * always be CPU_MASK_NONE. |
| 4418 | */ |
| 4419 | cpumask_t nohz_cpu_mask = CPU_MASK_NONE; |
| 4420 | |
| 4421 | #ifdef CONFIG_SMP |
| 4422 | /* |
| 4423 | * This is how migration works: |
| 4424 | * |
| 4425 | * 1) we queue a migration_req_t structure in the source CPU's |
| 4426 | * runqueue and wake up that CPU's migration thread. |
| 4427 | * 2) we down() the locked semaphore => thread blocks. |
| 4428 | * 3) migration thread wakes up (implicitly it forces the migrated |
| 4429 | * thread off the CPU) |
| 4430 | * 4) it gets the migration request and checks whether the migrated |
| 4431 | * task is still in the wrong runqueue. |
| 4432 | * 5) if it's in the wrong runqueue then the migration thread removes |
| 4433 | * it and puts it into the right queue. |
| 4434 | * 6) migration thread up()s the semaphore. |
| 4435 | * 7) we wake up and the migration is done. |
| 4436 | */ |
| 4437 | |
| 4438 | /* |
| 4439 | * Change a given task's CPU affinity. Migrate the thread to a |
| 4440 | * proper CPU and schedule it away if the CPU it's executing on |
| 4441 | * is removed from the allowed bitmask. |
| 4442 | * |
| 4443 | * NOTE: the caller must have a valid reference to the task, the |
| 4444 | * task must not exit() & deallocate itself prematurely. The |
| 4445 | * call is not atomic; no spinlocks may be held. |
| 4446 | */ |
| 4447 | int set_cpus_allowed(task_t *p, cpumask_t new_mask) |
| 4448 | { |
| 4449 | unsigned long flags; |
| 4450 | int ret = 0; |
| 4451 | migration_req_t req; |
| 4452 | runqueue_t *rq; |
| 4453 | |
| 4454 | rq = task_rq_lock(p, &flags); |
| 4455 | if (!cpus_intersects(new_mask, cpu_online_map)) { |
| 4456 | ret = -EINVAL; |
| 4457 | goto out; |
| 4458 | } |
| 4459 | |
| 4460 | p->cpus_allowed = new_mask; |
| 4461 | /* Can the task run on the task's current CPU? If so, we're done */ |
| 4462 | if (cpu_isset(task_cpu(p), new_mask)) |
| 4463 | goto out; |
| 4464 | |
| 4465 | if (migrate_task(p, any_online_cpu(new_mask), &req)) { |
| 4466 | /* Need help from migration thread: drop lock and wait. */ |
| 4467 | task_rq_unlock(rq, &flags); |
| 4468 | wake_up_process(rq->migration_thread); |
| 4469 | wait_for_completion(&req.done); |
| 4470 | tlb_migrate_finish(p->mm); |
| 4471 | return 0; |
| 4472 | } |
| 4473 | out: |
| 4474 | task_rq_unlock(rq, &flags); |
| 4475 | return ret; |
| 4476 | } |
| 4477 | |
| 4478 | EXPORT_SYMBOL_GPL(set_cpus_allowed); |
| 4479 | |
| 4480 | /* |
| 4481 | * Move (not current) task off this cpu, onto dest cpu. We're doing |
| 4482 | * this because either it can't run here any more (set_cpus_allowed() |
| 4483 | * away from this CPU, or CPU going down), or because we're |
| 4484 | * attempting to rebalance this task on exec (sched_exec). |
| 4485 | * |
| 4486 | * So we race with normal scheduler movements, but that's OK, as long |
| 4487 | * as the task is no longer on this CPU. |
| 4488 | */ |
| 4489 | static void __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) |
| 4490 | { |
| 4491 | runqueue_t *rq_dest, *rq_src; |
| 4492 | |
| 4493 | if (unlikely(cpu_is_offline(dest_cpu))) |
| 4494 | return; |
| 4495 | |
| 4496 | rq_src = cpu_rq(src_cpu); |
| 4497 | rq_dest = cpu_rq(dest_cpu); |
| 4498 | |
| 4499 | double_rq_lock(rq_src, rq_dest); |
| 4500 | /* Already moved. */ |
| 4501 | if (task_cpu(p) != src_cpu) |
| 4502 | goto out; |
| 4503 | /* Affinity changed (again). */ |
| 4504 | if (!cpu_isset(dest_cpu, p->cpus_allowed)) |
| 4505 | goto out; |
| 4506 | |
| 4507 | set_task_cpu(p, dest_cpu); |
| 4508 | if (p->array) { |
| 4509 | /* |
| 4510 | * Sync timestamp with rq_dest's before activating. |
| 4511 | * The same thing could be achieved by doing this step |
| 4512 | * afterwards, and pretending it was a local activate. |
| 4513 | * This way is cleaner and logically correct. |
| 4514 | */ |
| 4515 | p->timestamp = p->timestamp - rq_src->timestamp_last_tick |
| 4516 | + rq_dest->timestamp_last_tick; |
| 4517 | deactivate_task(p, rq_src); |
| 4518 | activate_task(p, rq_dest, 0); |
| 4519 | if (TASK_PREEMPTS_CURR(p, rq_dest)) |
| 4520 | resched_task(rq_dest->curr); |
| 4521 | } |
| 4522 | |
| 4523 | out: |
| 4524 | double_rq_unlock(rq_src, rq_dest); |
| 4525 | } |
| 4526 | |
| 4527 | /* |
| 4528 | * migration_thread - this is a highprio system thread that performs |
| 4529 | * thread migration by bumping thread off CPU then 'pushing' onto |
| 4530 | * another runqueue. |
| 4531 | */ |
Ingo Molnar | 95cdf3b | 2005-09-10 00:26:11 -0700 | [diff] [blame] | 4532 | static int migration_thread(void *data) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4533 | { |
| 4534 | runqueue_t *rq; |
| 4535 | int cpu = (long)data; |
| 4536 | |
| 4537 | rq = cpu_rq(cpu); |
| 4538 | BUG_ON(rq->migration_thread != current); |
| 4539 | |
| 4540 | set_current_state(TASK_INTERRUPTIBLE); |
| 4541 | while (!kthread_should_stop()) { |
| 4542 | struct list_head *head; |
| 4543 | migration_req_t *req; |
| 4544 | |
Christoph Lameter | 3e1d1d2 | 2005-06-24 23:13:50 -0700 | [diff] [blame] | 4545 | try_to_freeze(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4546 | |
| 4547 | spin_lock_irq(&rq->lock); |
| 4548 | |
| 4549 | if (cpu_is_offline(cpu)) { |
| 4550 | spin_unlock_irq(&rq->lock); |
| 4551 | goto wait_to_die; |
| 4552 | } |
| 4553 | |
| 4554 | if (rq->active_balance) { |
| 4555 | active_load_balance(rq, cpu); |
| 4556 | rq->active_balance = 0; |
| 4557 | } |
| 4558 | |
| 4559 | head = &rq->migration_queue; |
| 4560 | |
| 4561 | if (list_empty(head)) { |
| 4562 | spin_unlock_irq(&rq->lock); |
| 4563 | schedule(); |
| 4564 | set_current_state(TASK_INTERRUPTIBLE); |
| 4565 | continue; |
| 4566 | } |
| 4567 | req = list_entry(head->next, migration_req_t, list); |
| 4568 | list_del_init(head->next); |
| 4569 | |
Nick Piggin | 674311d | 2005-06-25 14:57:27 -0700 | [diff] [blame] | 4570 | spin_unlock(&rq->lock); |
| 4571 | __migrate_task(req->task, cpu, req->dest_cpu); |
| 4572 | local_irq_enable(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4573 | |
| 4574 | complete(&req->done); |
| 4575 | } |
| 4576 | __set_current_state(TASK_RUNNING); |
| 4577 | return 0; |
| 4578 | |
| 4579 | wait_to_die: |
| 4580 | /* Wait for kthread_stop */ |
| 4581 | set_current_state(TASK_INTERRUPTIBLE); |
| 4582 | while (!kthread_should_stop()) { |
| 4583 | schedule(); |
| 4584 | set_current_state(TASK_INTERRUPTIBLE); |
| 4585 | } |
| 4586 | __set_current_state(TASK_RUNNING); |
| 4587 | return 0; |
| 4588 | } |
| 4589 | |
| 4590 | #ifdef CONFIG_HOTPLUG_CPU |
| 4591 | /* Figure out where task on dead CPU should go, use force if neccessary. */ |
| 4592 | static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *tsk) |
| 4593 | { |
| 4594 | int dest_cpu; |
| 4595 | cpumask_t mask; |
| 4596 | |
| 4597 | /* On same node? */ |
| 4598 | mask = node_to_cpumask(cpu_to_node(dead_cpu)); |
| 4599 | cpus_and(mask, mask, tsk->cpus_allowed); |
| 4600 | dest_cpu = any_online_cpu(mask); |
| 4601 | |
| 4602 | /* On any allowed CPU? */ |
| 4603 | if (dest_cpu == NR_CPUS) |
| 4604 | dest_cpu = any_online_cpu(tsk->cpus_allowed); |
| 4605 | |
| 4606 | /* No more Mr. Nice Guy. */ |
| 4607 | if (dest_cpu == NR_CPUS) { |
Paul Jackson | b39c4fa | 2005-05-20 13:59:15 -0700 | [diff] [blame] | 4608 | cpus_setall(tsk->cpus_allowed); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4609 | dest_cpu = any_online_cpu(tsk->cpus_allowed); |
| 4610 | |
| 4611 | /* |
| 4612 | * Don't tell them about moving exiting tasks or |
| 4613 | * kernel threads (both mm NULL), since they never |
| 4614 | * leave kernel. |
| 4615 | */ |
| 4616 | if (tsk->mm && printk_ratelimit()) |
| 4617 | printk(KERN_INFO "process %d (%s) no " |
| 4618 | "longer affine to cpu%d\n", |
| 4619 | tsk->pid, tsk->comm, dead_cpu); |
| 4620 | } |
| 4621 | __migrate_task(tsk, dead_cpu, dest_cpu); |
| 4622 | } |
| 4623 | |
| 4624 | /* |
| 4625 | * While a dead CPU has no uninterruptible tasks queued at this point, |
| 4626 | * it might still have a nonzero ->nr_uninterruptible counter, because |
| 4627 | * for performance reasons the counter is not stricly tracking tasks to |
| 4628 | * their home CPUs. So we just add the counter to another CPU's counter, |
| 4629 | * to keep the global sum constant after CPU-down: |
| 4630 | */ |
| 4631 | static void migrate_nr_uninterruptible(runqueue_t *rq_src) |
| 4632 | { |
| 4633 | runqueue_t *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL)); |
| 4634 | unsigned long flags; |
| 4635 | |
| 4636 | local_irq_save(flags); |
| 4637 | double_rq_lock(rq_src, rq_dest); |
| 4638 | rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible; |
| 4639 | rq_src->nr_uninterruptible = 0; |
| 4640 | double_rq_unlock(rq_src, rq_dest); |
| 4641 | local_irq_restore(flags); |
| 4642 | } |
| 4643 | |
| 4644 | /* Run through task list and migrate tasks from the dead cpu. */ |
| 4645 | static void migrate_live_tasks(int src_cpu) |
| 4646 | { |
| 4647 | struct task_struct *tsk, *t; |
| 4648 | |
| 4649 | write_lock_irq(&tasklist_lock); |
| 4650 | |
| 4651 | do_each_thread(t, tsk) { |
| 4652 | if (tsk == current) |
| 4653 | continue; |
| 4654 | |
| 4655 | if (task_cpu(tsk) == src_cpu) |
| 4656 | move_task_off_dead_cpu(src_cpu, tsk); |
| 4657 | } while_each_thread(t, tsk); |
| 4658 | |
| 4659 | write_unlock_irq(&tasklist_lock); |
| 4660 | } |
| 4661 | |
| 4662 | /* Schedules idle task to be the next runnable task on current CPU. |
| 4663 | * It does so by boosting its priority to highest possible and adding it to |
| 4664 | * the _front_ of runqueue. Used by CPU offline code. |
| 4665 | */ |
| 4666 | void sched_idle_next(void) |
| 4667 | { |
| 4668 | int cpu = smp_processor_id(); |
| 4669 | runqueue_t *rq = this_rq(); |
| 4670 | struct task_struct *p = rq->idle; |
| 4671 | unsigned long flags; |
| 4672 | |
| 4673 | /* cpu has to be offline */ |
| 4674 | BUG_ON(cpu_online(cpu)); |
| 4675 | |
| 4676 | /* Strictly not necessary since rest of the CPUs are stopped by now |
| 4677 | * and interrupts disabled on current cpu. |
| 4678 | */ |
| 4679 | spin_lock_irqsave(&rq->lock, flags); |
| 4680 | |
| 4681 | __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1); |
| 4682 | /* Add idle task to _front_ of it's priority queue */ |
| 4683 | __activate_idle_task(p, rq); |
| 4684 | |
| 4685 | spin_unlock_irqrestore(&rq->lock, flags); |
| 4686 | } |
| 4687 | |
| 4688 | /* Ensures that the idle task is using init_mm right before its cpu goes |
| 4689 | * offline. |
| 4690 | */ |
| 4691 | void idle_task_exit(void) |
| 4692 | { |
| 4693 | struct mm_struct *mm = current->active_mm; |
| 4694 | |
| 4695 | BUG_ON(cpu_online(smp_processor_id())); |
| 4696 | |
| 4697 | if (mm != &init_mm) |
| 4698 | switch_mm(mm, &init_mm, current); |
| 4699 | mmdrop(mm); |
| 4700 | } |
| 4701 | |
| 4702 | static void migrate_dead(unsigned int dead_cpu, task_t *tsk) |
| 4703 | { |
| 4704 | struct runqueue *rq = cpu_rq(dead_cpu); |
| 4705 | |
| 4706 | /* Must be exiting, otherwise would be on tasklist. */ |
| 4707 | BUG_ON(tsk->exit_state != EXIT_ZOMBIE && tsk->exit_state != EXIT_DEAD); |
| 4708 | |
| 4709 | /* Cannot have done final schedule yet: would have vanished. */ |
| 4710 | BUG_ON(tsk->flags & PF_DEAD); |
| 4711 | |
| 4712 | get_task_struct(tsk); |
| 4713 | |
| 4714 | /* |
| 4715 | * Drop lock around migration; if someone else moves it, |
| 4716 | * that's OK. No task can be added to this CPU, so iteration is |
| 4717 | * fine. |
| 4718 | */ |
| 4719 | spin_unlock_irq(&rq->lock); |
| 4720 | move_task_off_dead_cpu(dead_cpu, tsk); |
| 4721 | spin_lock_irq(&rq->lock); |
| 4722 | |
| 4723 | put_task_struct(tsk); |
| 4724 | } |
| 4725 | |
| 4726 | /* release_task() removes task from tasklist, so we won't find dead tasks. */ |
| 4727 | static void migrate_dead_tasks(unsigned int dead_cpu) |
| 4728 | { |
| 4729 | unsigned arr, i; |
| 4730 | struct runqueue *rq = cpu_rq(dead_cpu); |
| 4731 | |
| 4732 | for (arr = 0; arr < 2; arr++) { |
| 4733 | for (i = 0; i < MAX_PRIO; i++) { |
| 4734 | struct list_head *list = &rq->arrays[arr].queue[i]; |
| 4735 | while (!list_empty(list)) |
| 4736 | migrate_dead(dead_cpu, |
| 4737 | list_entry(list->next, task_t, |
| 4738 | run_list)); |
| 4739 | } |
| 4740 | } |
| 4741 | } |
| 4742 | #endif /* CONFIG_HOTPLUG_CPU */ |
| 4743 | |
| 4744 | /* |
| 4745 | * migration_call - callback that gets triggered when a CPU is added. |
| 4746 | * Here we can start up the necessary migration thread for the new CPU. |
| 4747 | */ |
| 4748 | static int migration_call(struct notifier_block *nfb, unsigned long action, |
| 4749 | void *hcpu) |
| 4750 | { |
| 4751 | int cpu = (long)hcpu; |
| 4752 | struct task_struct *p; |
| 4753 | struct runqueue *rq; |
| 4754 | unsigned long flags; |
| 4755 | |
| 4756 | switch (action) { |
| 4757 | case CPU_UP_PREPARE: |
| 4758 | p = kthread_create(migration_thread, hcpu, "migration/%d",cpu); |
| 4759 | if (IS_ERR(p)) |
| 4760 | return NOTIFY_BAD; |
| 4761 | p->flags |= PF_NOFREEZE; |
| 4762 | kthread_bind(p, cpu); |
| 4763 | /* Must be high prio: stop_machine expects to yield to it. */ |
| 4764 | rq = task_rq_lock(p, &flags); |
| 4765 | __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1); |
| 4766 | task_rq_unlock(rq, &flags); |
| 4767 | cpu_rq(cpu)->migration_thread = p; |
| 4768 | break; |
| 4769 | case CPU_ONLINE: |
| 4770 | /* Strictly unneccessary, as first user will wake it. */ |
| 4771 | wake_up_process(cpu_rq(cpu)->migration_thread); |
| 4772 | break; |
| 4773 | #ifdef CONFIG_HOTPLUG_CPU |
| 4774 | case CPU_UP_CANCELED: |
| 4775 | /* Unbind it from offline cpu so it can run. Fall thru. */ |
Heiko Carstens | a4c4af7 | 2005-11-07 00:58:38 -0800 | [diff] [blame] | 4776 | kthread_bind(cpu_rq(cpu)->migration_thread, |
| 4777 | any_online_cpu(cpu_online_map)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4778 | kthread_stop(cpu_rq(cpu)->migration_thread); |
| 4779 | cpu_rq(cpu)->migration_thread = NULL; |
| 4780 | break; |
| 4781 | case CPU_DEAD: |
| 4782 | migrate_live_tasks(cpu); |
| 4783 | rq = cpu_rq(cpu); |
| 4784 | kthread_stop(rq->migration_thread); |
| 4785 | rq->migration_thread = NULL; |
| 4786 | /* Idle task back to normal (off runqueue, low prio) */ |
| 4787 | rq = task_rq_lock(rq->idle, &flags); |
| 4788 | deactivate_task(rq->idle, rq); |
| 4789 | rq->idle->static_prio = MAX_PRIO; |
| 4790 | __setscheduler(rq->idle, SCHED_NORMAL, 0); |
| 4791 | migrate_dead_tasks(cpu); |
| 4792 | task_rq_unlock(rq, &flags); |
| 4793 | migrate_nr_uninterruptible(rq); |
| 4794 | BUG_ON(rq->nr_running != 0); |
| 4795 | |
| 4796 | /* No need to migrate the tasks: it was best-effort if |
| 4797 | * they didn't do lock_cpu_hotplug(). Just wake up |
| 4798 | * the requestors. */ |
| 4799 | spin_lock_irq(&rq->lock); |
| 4800 | while (!list_empty(&rq->migration_queue)) { |
| 4801 | migration_req_t *req; |
| 4802 | req = list_entry(rq->migration_queue.next, |
| 4803 | migration_req_t, list); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4804 | list_del_init(&req->list); |
| 4805 | complete(&req->done); |
| 4806 | } |
| 4807 | spin_unlock_irq(&rq->lock); |
| 4808 | break; |
| 4809 | #endif |
| 4810 | } |
| 4811 | return NOTIFY_OK; |
| 4812 | } |
| 4813 | |
| 4814 | /* Register at highest priority so that task migration (migrate_all_tasks) |
| 4815 | * happens before everything else. |
| 4816 | */ |
| 4817 | static struct notifier_block __devinitdata migration_notifier = { |
| 4818 | .notifier_call = migration_call, |
| 4819 | .priority = 10 |
| 4820 | }; |
| 4821 | |
| 4822 | int __init migration_init(void) |
| 4823 | { |
| 4824 | void *cpu = (void *)(long)smp_processor_id(); |
| 4825 | /* Start one for boot CPU. */ |
| 4826 | migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); |
| 4827 | migration_call(&migration_notifier, CPU_ONLINE, cpu); |
| 4828 | register_cpu_notifier(&migration_notifier); |
| 4829 | return 0; |
| 4830 | } |
| 4831 | #endif |
| 4832 | |
| 4833 | #ifdef CONFIG_SMP |
Dinakar Guniguntala | 1a20ff2 | 2005-06-25 14:57:33 -0700 | [diff] [blame] | 4834 | #undef SCHED_DOMAIN_DEBUG |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4835 | #ifdef SCHED_DOMAIN_DEBUG |
| 4836 | static void sched_domain_debug(struct sched_domain *sd, int cpu) |
| 4837 | { |
| 4838 | int level = 0; |
| 4839 | |
Nick Piggin | 41c7ce9 | 2005-06-25 14:57:24 -0700 | [diff] [blame] | 4840 | if (!sd) { |
| 4841 | printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); |
| 4842 | return; |
| 4843 | } |
| 4844 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4845 | printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); |
| 4846 | |
| 4847 | do { |
| 4848 | int i; |
| 4849 | char str[NR_CPUS]; |
| 4850 | struct sched_group *group = sd->groups; |
| 4851 | cpumask_t groupmask; |
| 4852 | |
| 4853 | cpumask_scnprintf(str, NR_CPUS, sd->span); |
| 4854 | cpus_clear(groupmask); |
| 4855 | |
| 4856 | printk(KERN_DEBUG); |
| 4857 | for (i = 0; i < level + 1; i++) |
| 4858 | printk(" "); |
| 4859 | printk("domain %d: ", level); |
| 4860 | |
| 4861 | if (!(sd->flags & SD_LOAD_BALANCE)) { |
| 4862 | printk("does not load-balance\n"); |
| 4863 | if (sd->parent) |
| 4864 | printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent"); |
| 4865 | break; |
| 4866 | } |
| 4867 | |
| 4868 | printk("span %s\n", str); |
| 4869 | |
| 4870 | if (!cpu_isset(cpu, sd->span)) |
| 4871 | printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu); |
| 4872 | if (!cpu_isset(cpu, group->cpumask)) |
| 4873 | printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu); |
| 4874 | |
| 4875 | printk(KERN_DEBUG); |
| 4876 | for (i = 0; i < level + 2; i++) |
| 4877 | printk(" "); |
| 4878 | printk("groups:"); |
| 4879 | do { |
| 4880 | if (!group) { |
| 4881 | printk("\n"); |
| 4882 | printk(KERN_ERR "ERROR: group is NULL\n"); |
| 4883 | break; |
| 4884 | } |
| 4885 | |
| 4886 | if (!group->cpu_power) { |
| 4887 | printk("\n"); |
| 4888 | printk(KERN_ERR "ERROR: domain->cpu_power not set\n"); |
| 4889 | } |
| 4890 | |
| 4891 | if (!cpus_weight(group->cpumask)) { |
| 4892 | printk("\n"); |
| 4893 | printk(KERN_ERR "ERROR: empty group\n"); |
| 4894 | } |
| 4895 | |
| 4896 | if (cpus_intersects(groupmask, group->cpumask)) { |
| 4897 | printk("\n"); |
| 4898 | printk(KERN_ERR "ERROR: repeated CPUs\n"); |
| 4899 | } |
| 4900 | |
| 4901 | cpus_or(groupmask, groupmask, group->cpumask); |
| 4902 | |
| 4903 | cpumask_scnprintf(str, NR_CPUS, group->cpumask); |
| 4904 | printk(" %s", str); |
| 4905 | |
| 4906 | group = group->next; |
| 4907 | } while (group != sd->groups); |
| 4908 | printk("\n"); |
| 4909 | |
| 4910 | if (!cpus_equal(sd->span, groupmask)) |
| 4911 | printk(KERN_ERR "ERROR: groups don't span domain->span\n"); |
| 4912 | |
| 4913 | level++; |
| 4914 | sd = sd->parent; |
| 4915 | |
| 4916 | if (sd) { |
| 4917 | if (!cpus_subset(groupmask, sd->span)) |
| 4918 | printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n"); |
| 4919 | } |
| 4920 | |
| 4921 | } while (sd); |
| 4922 | } |
| 4923 | #else |
| 4924 | #define sched_domain_debug(sd, cpu) {} |
| 4925 | #endif |
| 4926 | |
Dinakar Guniguntala | 1a20ff2 | 2005-06-25 14:57:33 -0700 | [diff] [blame] | 4927 | static int sd_degenerate(struct sched_domain *sd) |
Suresh Siddha | 245af2c | 2005-06-25 14:57:25 -0700 | [diff] [blame] | 4928 | { |
| 4929 | if (cpus_weight(sd->span) == 1) |
| 4930 | return 1; |
| 4931 | |
| 4932 | /* Following flags need at least 2 groups */ |
| 4933 | if (sd->flags & (SD_LOAD_BALANCE | |
| 4934 | SD_BALANCE_NEWIDLE | |
| 4935 | SD_BALANCE_FORK | |
| 4936 | SD_BALANCE_EXEC)) { |
| 4937 | if (sd->groups != sd->groups->next) |
| 4938 | return 0; |
| 4939 | } |
| 4940 | |
| 4941 | /* Following flags don't use groups */ |
| 4942 | if (sd->flags & (SD_WAKE_IDLE | |
| 4943 | SD_WAKE_AFFINE | |
| 4944 | SD_WAKE_BALANCE)) |
| 4945 | return 0; |
| 4946 | |
| 4947 | return 1; |
| 4948 | } |
| 4949 | |
Dinakar Guniguntala | 1a20ff2 | 2005-06-25 14:57:33 -0700 | [diff] [blame] | 4950 | static int sd_parent_degenerate(struct sched_domain *sd, |
Suresh Siddha | 245af2c | 2005-06-25 14:57:25 -0700 | [diff] [blame] | 4951 | struct sched_domain *parent) |
| 4952 | { |
| 4953 | unsigned long cflags = sd->flags, pflags = parent->flags; |
| 4954 | |
| 4955 | if (sd_degenerate(parent)) |
| 4956 | return 1; |
| 4957 | |
| 4958 | if (!cpus_equal(sd->span, parent->span)) |
| 4959 | return 0; |
| 4960 | |
| 4961 | /* Does parent contain flags not in child? */ |
| 4962 | /* WAKE_BALANCE is a subset of WAKE_AFFINE */ |
| 4963 | if (cflags & SD_WAKE_AFFINE) |
| 4964 | pflags &= ~SD_WAKE_BALANCE; |
| 4965 | /* Flags needing groups don't count if only 1 group in parent */ |
| 4966 | if (parent->groups == parent->groups->next) { |
| 4967 | pflags &= ~(SD_LOAD_BALANCE | |
| 4968 | SD_BALANCE_NEWIDLE | |
| 4969 | SD_BALANCE_FORK | |
| 4970 | SD_BALANCE_EXEC); |
| 4971 | } |
| 4972 | if (~cflags & pflags) |
| 4973 | return 0; |
| 4974 | |
| 4975 | return 1; |
| 4976 | } |
| 4977 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4978 | /* |
| 4979 | * Attach the domain 'sd' to 'cpu' as its base domain. Callers must |
| 4980 | * hold the hotplug lock. |
| 4981 | */ |
John Hawkes | 9c1cfda | 2005-09-06 15:18:14 -0700 | [diff] [blame] | 4982 | static void cpu_attach_domain(struct sched_domain *sd, int cpu) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4983 | { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4984 | runqueue_t *rq = cpu_rq(cpu); |
Suresh Siddha | 245af2c | 2005-06-25 14:57:25 -0700 | [diff] [blame] | 4985 | struct sched_domain *tmp; |
| 4986 | |
| 4987 | /* Remove the sched domains which do not contribute to scheduling. */ |
| 4988 | for (tmp = sd; tmp; tmp = tmp->parent) { |
| 4989 | struct sched_domain *parent = tmp->parent; |
| 4990 | if (!parent) |
| 4991 | break; |
| 4992 | if (sd_parent_degenerate(tmp, parent)) |
| 4993 | tmp->parent = parent->parent; |
| 4994 | } |
| 4995 | |
| 4996 | if (sd && sd_degenerate(sd)) |
| 4997 | sd = sd->parent; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4998 | |
| 4999 | sched_domain_debug(sd, cpu); |
| 5000 | |
Nick Piggin | 674311d | 2005-06-25 14:57:27 -0700 | [diff] [blame] | 5001 | rcu_assign_pointer(rq->sd, sd); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5002 | } |
| 5003 | |
| 5004 | /* cpus with isolated domains */ |
John Hawkes | 9c1cfda | 2005-09-06 15:18:14 -0700 | [diff] [blame] | 5005 | static cpumask_t __devinitdata cpu_isolated_map = CPU_MASK_NONE; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5006 | |
| 5007 | /* Setup the mask of cpus configured for isolated domains */ |
| 5008 | static int __init isolated_cpu_setup(char *str) |
| 5009 | { |
| 5010 | int ints[NR_CPUS], i; |
| 5011 | |
| 5012 | str = get_options(str, ARRAY_SIZE(ints), ints); |
| 5013 | cpus_clear(cpu_isolated_map); |
| 5014 | for (i = 1; i <= ints[0]; i++) |
| 5015 | if (ints[i] < NR_CPUS) |
| 5016 | cpu_set(ints[i], cpu_isolated_map); |
| 5017 | return 1; |
| 5018 | } |
| 5019 | |
| 5020 | __setup ("isolcpus=", isolated_cpu_setup); |
| 5021 | |
| 5022 | /* |
| 5023 | * init_sched_build_groups takes an array of groups, the cpumask we wish |
| 5024 | * to span, and a pointer to a function which identifies what group a CPU |
| 5025 | * belongs to. The return value of group_fn must be a valid index into the |
| 5026 | * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we |
| 5027 | * keep track of groups covered with a cpumask_t). |
| 5028 | * |
| 5029 | * init_sched_build_groups will build a circular linked list of the groups |
| 5030 | * covered by the given span, and will set each group's ->cpumask correctly, |
| 5031 | * and ->cpu_power to 0. |
| 5032 | */ |
John Hawkes | 9c1cfda | 2005-09-06 15:18:14 -0700 | [diff] [blame] | 5033 | static void init_sched_build_groups(struct sched_group groups[], cpumask_t span, |
| 5034 | int (*group_fn)(int cpu)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5035 | { |
| 5036 | struct sched_group *first = NULL, *last = NULL; |
| 5037 | cpumask_t covered = CPU_MASK_NONE; |
| 5038 | int i; |
| 5039 | |
| 5040 | for_each_cpu_mask(i, span) { |
| 5041 | int group = group_fn(i); |
| 5042 | struct sched_group *sg = &groups[group]; |
| 5043 | int j; |
| 5044 | |
| 5045 | if (cpu_isset(i, covered)) |
| 5046 | continue; |
| 5047 | |
| 5048 | sg->cpumask = CPU_MASK_NONE; |
| 5049 | sg->cpu_power = 0; |
| 5050 | |
| 5051 | for_each_cpu_mask(j, span) { |
| 5052 | if (group_fn(j) != group) |
| 5053 | continue; |
| 5054 | |
| 5055 | cpu_set(j, covered); |
| 5056 | cpu_set(j, sg->cpumask); |
| 5057 | } |
| 5058 | if (!first) |
| 5059 | first = sg; |
| 5060 | if (last) |
| 5061 | last->next = sg; |
| 5062 | last = sg; |
| 5063 | } |
| 5064 | last->next = first; |
| 5065 | } |
| 5066 | |
John Hawkes | 9c1cfda | 2005-09-06 15:18:14 -0700 | [diff] [blame] | 5067 | #define SD_NODES_PER_DOMAIN 16 |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5068 | |
akpm@osdl.org | 198e2f1 | 2006-01-12 01:05:30 -0800 | [diff] [blame] | 5069 | /* |
| 5070 | * Self-tuning task migration cost measurement between source and target CPUs. |
| 5071 | * |
| 5072 | * This is done by measuring the cost of manipulating buffers of varying |
| 5073 | * sizes. For a given buffer-size here are the steps that are taken: |
| 5074 | * |
| 5075 | * 1) the source CPU reads+dirties a shared buffer |
| 5076 | * 2) the target CPU reads+dirties the same shared buffer |
| 5077 | * |
| 5078 | * We measure how long they take, in the following 4 scenarios: |
| 5079 | * |
| 5080 | * - source: CPU1, target: CPU2 | cost1 |
| 5081 | * - source: CPU2, target: CPU1 | cost2 |
| 5082 | * - source: CPU1, target: CPU1 | cost3 |
| 5083 | * - source: CPU2, target: CPU2 | cost4 |
| 5084 | * |
| 5085 | * We then calculate the cost3+cost4-cost1-cost2 difference - this is |
| 5086 | * the cost of migration. |
| 5087 | * |
| 5088 | * We then start off from a small buffer-size and iterate up to larger |
| 5089 | * buffer sizes, in 5% steps - measuring each buffer-size separately, and |
| 5090 | * doing a maximum search for the cost. (The maximum cost for a migration |
| 5091 | * normally occurs when the working set size is around the effective cache |
| 5092 | * size.) |
| 5093 | */ |
| 5094 | #define SEARCH_SCOPE 2 |
| 5095 | #define MIN_CACHE_SIZE (64*1024U) |
| 5096 | #define DEFAULT_CACHE_SIZE (5*1024*1024U) |
Ingo Molnar | 70b4d63 | 2006-01-30 20:24:38 +0100 | [diff] [blame] | 5097 | #define ITERATIONS 1 |
akpm@osdl.org | 198e2f1 | 2006-01-12 01:05:30 -0800 | [diff] [blame] | 5098 | #define SIZE_THRESH 130 |
| 5099 | #define COST_THRESH 130 |
| 5100 | |
| 5101 | /* |
| 5102 | * The migration cost is a function of 'domain distance'. Domain |
| 5103 | * distance is the number of steps a CPU has to iterate down its |
| 5104 | * domain tree to share a domain with the other CPU. The farther |
| 5105 | * two CPUs are from each other, the larger the distance gets. |
| 5106 | * |
| 5107 | * Note that we use the distance only to cache measurement results, |
| 5108 | * the distance value is not used numerically otherwise. When two |
| 5109 | * CPUs have the same distance it is assumed that the migration |
| 5110 | * cost is the same. (this is a simplification but quite practical) |
| 5111 | */ |
| 5112 | #define MAX_DOMAIN_DISTANCE 32 |
| 5113 | |
| 5114 | static unsigned long long migration_cost[MAX_DOMAIN_DISTANCE] = |
Ingo Molnar | 4bbf39c | 2006-02-17 13:52:44 -0800 | [diff] [blame] | 5115 | { [ 0 ... MAX_DOMAIN_DISTANCE-1 ] = |
| 5116 | /* |
| 5117 | * Architectures may override the migration cost and thus avoid |
| 5118 | * boot-time calibration. Unit is nanoseconds. Mostly useful for |
| 5119 | * virtualized hardware: |
| 5120 | */ |
| 5121 | #ifdef CONFIG_DEFAULT_MIGRATION_COST |
| 5122 | CONFIG_DEFAULT_MIGRATION_COST |
| 5123 | #else |
| 5124 | -1LL |
| 5125 | #endif |
| 5126 | }; |
akpm@osdl.org | 198e2f1 | 2006-01-12 01:05:30 -0800 | [diff] [blame] | 5127 | |
| 5128 | /* |
| 5129 | * Allow override of migration cost - in units of microseconds. |
| 5130 | * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost |
| 5131 | * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs: |
| 5132 | */ |
| 5133 | static int __init migration_cost_setup(char *str) |
| 5134 | { |
| 5135 | int ints[MAX_DOMAIN_DISTANCE+1], i; |
| 5136 | |
| 5137 | str = get_options(str, ARRAY_SIZE(ints), ints); |
| 5138 | |
| 5139 | printk("#ints: %d\n", ints[0]); |
| 5140 | for (i = 1; i <= ints[0]; i++) { |
| 5141 | migration_cost[i-1] = (unsigned long long)ints[i]*1000; |
| 5142 | printk("migration_cost[%d]: %Ld\n", i-1, migration_cost[i-1]); |
| 5143 | } |
| 5144 | return 1; |
| 5145 | } |
| 5146 | |
| 5147 | __setup ("migration_cost=", migration_cost_setup); |
| 5148 | |
| 5149 | /* |
| 5150 | * Global multiplier (divisor) for migration-cutoff values, |
| 5151 | * in percentiles. E.g. use a value of 150 to get 1.5 times |
| 5152 | * longer cache-hot cutoff times. |
| 5153 | * |
| 5154 | * (We scale it from 100 to 128 to long long handling easier.) |
| 5155 | */ |
| 5156 | |
| 5157 | #define MIGRATION_FACTOR_SCALE 128 |
| 5158 | |
| 5159 | static unsigned int migration_factor = MIGRATION_FACTOR_SCALE; |
| 5160 | |
| 5161 | static int __init setup_migration_factor(char *str) |
| 5162 | { |
| 5163 | get_option(&str, &migration_factor); |
| 5164 | migration_factor = migration_factor * MIGRATION_FACTOR_SCALE / 100; |
| 5165 | return 1; |
| 5166 | } |
| 5167 | |
| 5168 | __setup("migration_factor=", setup_migration_factor); |
| 5169 | |
| 5170 | /* |
| 5171 | * Estimated distance of two CPUs, measured via the number of domains |
| 5172 | * we have to pass for the two CPUs to be in the same span: |
| 5173 | */ |
| 5174 | static unsigned long domain_distance(int cpu1, int cpu2) |
| 5175 | { |
| 5176 | unsigned long distance = 0; |
| 5177 | struct sched_domain *sd; |
| 5178 | |
| 5179 | for_each_domain(cpu1, sd) { |
| 5180 | WARN_ON(!cpu_isset(cpu1, sd->span)); |
| 5181 | if (cpu_isset(cpu2, sd->span)) |
| 5182 | return distance; |
| 5183 | distance++; |
| 5184 | } |
| 5185 | if (distance >= MAX_DOMAIN_DISTANCE) { |
| 5186 | WARN_ON(1); |
| 5187 | distance = MAX_DOMAIN_DISTANCE-1; |
| 5188 | } |
| 5189 | |
| 5190 | return distance; |
| 5191 | } |
| 5192 | |
| 5193 | static unsigned int migration_debug; |
| 5194 | |
| 5195 | static int __init setup_migration_debug(char *str) |
| 5196 | { |
| 5197 | get_option(&str, &migration_debug); |
| 5198 | return 1; |
| 5199 | } |
| 5200 | |
| 5201 | __setup("migration_debug=", setup_migration_debug); |
| 5202 | |
| 5203 | /* |
| 5204 | * Maximum cache-size that the scheduler should try to measure. |
| 5205 | * Architectures with larger caches should tune this up during |
| 5206 | * bootup. Gets used in the domain-setup code (i.e. during SMP |
| 5207 | * bootup). |
| 5208 | */ |
| 5209 | unsigned int max_cache_size; |
| 5210 | |
| 5211 | static int __init setup_max_cache_size(char *str) |
| 5212 | { |
| 5213 | get_option(&str, &max_cache_size); |
| 5214 | return 1; |
| 5215 | } |
| 5216 | |
| 5217 | __setup("max_cache_size=", setup_max_cache_size); |
| 5218 | |
| 5219 | /* |
| 5220 | * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This |
| 5221 | * is the operation that is timed, so we try to generate unpredictable |
| 5222 | * cachemisses that still end up filling the L2 cache: |
| 5223 | */ |
| 5224 | static void touch_cache(void *__cache, unsigned long __size) |
| 5225 | { |
| 5226 | unsigned long size = __size/sizeof(long), chunk1 = size/3, |
| 5227 | chunk2 = 2*size/3; |
| 5228 | unsigned long *cache = __cache; |
| 5229 | int i; |
| 5230 | |
| 5231 | for (i = 0; i < size/6; i += 8) { |
| 5232 | switch (i % 6) { |
| 5233 | case 0: cache[i]++; |
| 5234 | case 1: cache[size-1-i]++; |
| 5235 | case 2: cache[chunk1-i]++; |
| 5236 | case 3: cache[chunk1+i]++; |
| 5237 | case 4: cache[chunk2-i]++; |
| 5238 | case 5: cache[chunk2+i]++; |
| 5239 | } |
| 5240 | } |
| 5241 | } |
| 5242 | |
| 5243 | /* |
| 5244 | * Measure the cache-cost of one task migration. Returns in units of nsec. |
| 5245 | */ |
| 5246 | static unsigned long long measure_one(void *cache, unsigned long size, |
| 5247 | int source, int target) |
| 5248 | { |
| 5249 | cpumask_t mask, saved_mask; |
| 5250 | unsigned long long t0, t1, t2, t3, cost; |
| 5251 | |
| 5252 | saved_mask = current->cpus_allowed; |
| 5253 | |
| 5254 | /* |
| 5255 | * Flush source caches to RAM and invalidate them: |
| 5256 | */ |
| 5257 | sched_cacheflush(); |
| 5258 | |
| 5259 | /* |
| 5260 | * Migrate to the source CPU: |
| 5261 | */ |
| 5262 | mask = cpumask_of_cpu(source); |
| 5263 | set_cpus_allowed(current, mask); |
| 5264 | WARN_ON(smp_processor_id() != source); |
| 5265 | |
| 5266 | /* |
| 5267 | * Dirty the working set: |
| 5268 | */ |
| 5269 | t0 = sched_clock(); |
| 5270 | touch_cache(cache, size); |
| 5271 | t1 = sched_clock(); |
| 5272 | |
| 5273 | /* |
| 5274 | * Migrate to the target CPU, dirty the L2 cache and access |
| 5275 | * the shared buffer. (which represents the working set |
| 5276 | * of a migrated task.) |
| 5277 | */ |
| 5278 | mask = cpumask_of_cpu(target); |
| 5279 | set_cpus_allowed(current, mask); |
| 5280 | WARN_ON(smp_processor_id() != target); |
| 5281 | |
| 5282 | t2 = sched_clock(); |
| 5283 | touch_cache(cache, size); |
| 5284 | t3 = sched_clock(); |
| 5285 | |
| 5286 | cost = t1-t0 + t3-t2; |
| 5287 | |
| 5288 | if (migration_debug >= 2) |
| 5289 | printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n", |
| 5290 | source, target, t1-t0, t1-t0, t3-t2, cost); |
| 5291 | /* |
| 5292 | * Flush target caches to RAM and invalidate them: |
| 5293 | */ |
| 5294 | sched_cacheflush(); |
| 5295 | |
| 5296 | set_cpus_allowed(current, saved_mask); |
| 5297 | |
| 5298 | return cost; |
| 5299 | } |
| 5300 | |
| 5301 | /* |
| 5302 | * Measure a series of task migrations and return the average |
| 5303 | * result. Since this code runs early during bootup the system |
| 5304 | * is 'undisturbed' and the average latency makes sense. |
| 5305 | * |
| 5306 | * The algorithm in essence auto-detects the relevant cache-size, |
| 5307 | * so it will properly detect different cachesizes for different |
| 5308 | * cache-hierarchies, depending on how the CPUs are connected. |
| 5309 | * |
| 5310 | * Architectures can prime the upper limit of the search range via |
| 5311 | * max_cache_size, otherwise the search range defaults to 20MB...64K. |
| 5312 | */ |
| 5313 | static unsigned long long |
| 5314 | measure_cost(int cpu1, int cpu2, void *cache, unsigned int size) |
| 5315 | { |
| 5316 | unsigned long long cost1, cost2; |
| 5317 | int i; |
| 5318 | |
| 5319 | /* |
| 5320 | * Measure the migration cost of 'size' bytes, over an |
| 5321 | * average of 10 runs: |
| 5322 | * |
| 5323 | * (We perturb the cache size by a small (0..4k) |
| 5324 | * value to compensate size/alignment related artifacts. |
| 5325 | * We also subtract the cost of the operation done on |
| 5326 | * the same CPU.) |
| 5327 | */ |
| 5328 | cost1 = 0; |
| 5329 | |
| 5330 | /* |
| 5331 | * dry run, to make sure we start off cache-cold on cpu1, |
| 5332 | * and to get any vmalloc pagefaults in advance: |
| 5333 | */ |
| 5334 | measure_one(cache, size, cpu1, cpu2); |
| 5335 | for (i = 0; i < ITERATIONS; i++) |
| 5336 | cost1 += measure_one(cache, size - i*1024, cpu1, cpu2); |
| 5337 | |
| 5338 | measure_one(cache, size, cpu2, cpu1); |
| 5339 | for (i = 0; i < ITERATIONS; i++) |
| 5340 | cost1 += measure_one(cache, size - i*1024, cpu2, cpu1); |
| 5341 | |
| 5342 | /* |
| 5343 | * (We measure the non-migrating [cached] cost on both |
| 5344 | * cpu1 and cpu2, to handle CPUs with different speeds) |
| 5345 | */ |
| 5346 | cost2 = 0; |
| 5347 | |
| 5348 | measure_one(cache, size, cpu1, cpu1); |
| 5349 | for (i = 0; i < ITERATIONS; i++) |
| 5350 | cost2 += measure_one(cache, size - i*1024, cpu1, cpu1); |
| 5351 | |
| 5352 | measure_one(cache, size, cpu2, cpu2); |
| 5353 | for (i = 0; i < ITERATIONS; i++) |
| 5354 | cost2 += measure_one(cache, size - i*1024, cpu2, cpu2); |
| 5355 | |
| 5356 | /* |
| 5357 | * Get the per-iteration migration cost: |
| 5358 | */ |
| 5359 | do_div(cost1, 2*ITERATIONS); |
| 5360 | do_div(cost2, 2*ITERATIONS); |
| 5361 | |
| 5362 | return cost1 - cost2; |
| 5363 | } |
| 5364 | |
| 5365 | static unsigned long long measure_migration_cost(int cpu1, int cpu2) |
| 5366 | { |
| 5367 | unsigned long long max_cost = 0, fluct = 0, avg_fluct = 0; |
| 5368 | unsigned int max_size, size, size_found = 0; |
| 5369 | long long cost = 0, prev_cost; |
| 5370 | void *cache; |
| 5371 | |
| 5372 | /* |
| 5373 | * Search from max_cache_size*5 down to 64K - the real relevant |
| 5374 | * cachesize has to lie somewhere inbetween. |
| 5375 | */ |
| 5376 | if (max_cache_size) { |
| 5377 | max_size = max(max_cache_size * SEARCH_SCOPE, MIN_CACHE_SIZE); |
| 5378 | size = max(max_cache_size / SEARCH_SCOPE, MIN_CACHE_SIZE); |
| 5379 | } else { |
| 5380 | /* |
| 5381 | * Since we have no estimation about the relevant |
| 5382 | * search range |
| 5383 | */ |
| 5384 | max_size = DEFAULT_CACHE_SIZE * SEARCH_SCOPE; |
| 5385 | size = MIN_CACHE_SIZE; |
| 5386 | } |
| 5387 | |
| 5388 | if (!cpu_online(cpu1) || !cpu_online(cpu2)) { |
| 5389 | printk("cpu %d and %d not both online!\n", cpu1, cpu2); |
| 5390 | return 0; |
| 5391 | } |
| 5392 | |
| 5393 | /* |
| 5394 | * Allocate the working set: |
| 5395 | */ |
| 5396 | cache = vmalloc(max_size); |
| 5397 | if (!cache) { |
| 5398 | printk("could not vmalloc %d bytes for cache!\n", 2*max_size); |
| 5399 | return 1000000; // return 1 msec on very small boxen |
| 5400 | } |
| 5401 | |
| 5402 | while (size <= max_size) { |
| 5403 | prev_cost = cost; |
| 5404 | cost = measure_cost(cpu1, cpu2, cache, size); |
| 5405 | |
| 5406 | /* |
| 5407 | * Update the max: |
| 5408 | */ |
| 5409 | if (cost > 0) { |
| 5410 | if (max_cost < cost) { |
| 5411 | max_cost = cost; |
| 5412 | size_found = size; |
| 5413 | } |
| 5414 | } |
| 5415 | /* |
| 5416 | * Calculate average fluctuation, we use this to prevent |
| 5417 | * noise from triggering an early break out of the loop: |
| 5418 | */ |
| 5419 | fluct = abs(cost - prev_cost); |
| 5420 | avg_fluct = (avg_fluct + fluct)/2; |
| 5421 | |
| 5422 | if (migration_debug) |
| 5423 | printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): (%8Ld %8Ld)\n", |
| 5424 | cpu1, cpu2, size, |
| 5425 | (long)cost / 1000000, |
| 5426 | ((long)cost / 100000) % 10, |
| 5427 | (long)max_cost / 1000000, |
| 5428 | ((long)max_cost / 100000) % 10, |
| 5429 | domain_distance(cpu1, cpu2), |
| 5430 | cost, avg_fluct); |
| 5431 | |
| 5432 | /* |
| 5433 | * If we iterated at least 20% past the previous maximum, |
| 5434 | * and the cost has dropped by more than 20% already, |
| 5435 | * (taking fluctuations into account) then we assume to |
| 5436 | * have found the maximum and break out of the loop early: |
| 5437 | */ |
| 5438 | if (size_found && (size*100 > size_found*SIZE_THRESH)) |
| 5439 | if (cost+avg_fluct <= 0 || |
| 5440 | max_cost*100 > (cost+avg_fluct)*COST_THRESH) { |
| 5441 | |
| 5442 | if (migration_debug) |
| 5443 | printk("-> found max.\n"); |
| 5444 | break; |
| 5445 | } |
| 5446 | /* |
Ingo Molnar | 70b4d63 | 2006-01-30 20:24:38 +0100 | [diff] [blame] | 5447 | * Increase the cachesize in 10% steps: |
akpm@osdl.org | 198e2f1 | 2006-01-12 01:05:30 -0800 | [diff] [blame] | 5448 | */ |
Ingo Molnar | 70b4d63 | 2006-01-30 20:24:38 +0100 | [diff] [blame] | 5449 | size = size * 10 / 9; |
akpm@osdl.org | 198e2f1 | 2006-01-12 01:05:30 -0800 | [diff] [blame] | 5450 | } |
| 5451 | |
| 5452 | if (migration_debug) |
| 5453 | printk("[%d][%d] working set size found: %d, cost: %Ld\n", |
| 5454 | cpu1, cpu2, size_found, max_cost); |
| 5455 | |
| 5456 | vfree(cache); |
| 5457 | |
| 5458 | /* |
| 5459 | * A task is considered 'cache cold' if at least 2 times |
| 5460 | * the worst-case cost of migration has passed. |
| 5461 | * |
| 5462 | * (this limit is only listened to if the load-balancing |
| 5463 | * situation is 'nice' - if there is a large imbalance we |
| 5464 | * ignore it for the sake of CPU utilization and |
| 5465 | * processing fairness.) |
| 5466 | */ |
| 5467 | return 2 * max_cost * migration_factor / MIGRATION_FACTOR_SCALE; |
| 5468 | } |
| 5469 | |
| 5470 | static void calibrate_migration_costs(const cpumask_t *cpu_map) |
| 5471 | { |
| 5472 | int cpu1 = -1, cpu2 = -1, cpu, orig_cpu = raw_smp_processor_id(); |
| 5473 | unsigned long j0, j1, distance, max_distance = 0; |
| 5474 | struct sched_domain *sd; |
| 5475 | |
| 5476 | j0 = jiffies; |
| 5477 | |
| 5478 | /* |
| 5479 | * First pass - calculate the cacheflush times: |
| 5480 | */ |
| 5481 | for_each_cpu_mask(cpu1, *cpu_map) { |
| 5482 | for_each_cpu_mask(cpu2, *cpu_map) { |
| 5483 | if (cpu1 == cpu2) |
| 5484 | continue; |
| 5485 | distance = domain_distance(cpu1, cpu2); |
| 5486 | max_distance = max(max_distance, distance); |
| 5487 | /* |
| 5488 | * No result cached yet? |
| 5489 | */ |
| 5490 | if (migration_cost[distance] == -1LL) |
| 5491 | migration_cost[distance] = |
| 5492 | measure_migration_cost(cpu1, cpu2); |
| 5493 | } |
| 5494 | } |
| 5495 | /* |
| 5496 | * Second pass - update the sched domain hierarchy with |
| 5497 | * the new cache-hot-time estimations: |
| 5498 | */ |
| 5499 | for_each_cpu_mask(cpu, *cpu_map) { |
| 5500 | distance = 0; |
| 5501 | for_each_domain(cpu, sd) { |
| 5502 | sd->cache_hot_time = migration_cost[distance]; |
| 5503 | distance++; |
| 5504 | } |
| 5505 | } |
| 5506 | /* |
| 5507 | * Print the matrix: |
| 5508 | */ |
| 5509 | if (migration_debug) |
| 5510 | printk("migration: max_cache_size: %d, cpu: %d MHz:\n", |
| 5511 | max_cache_size, |
| 5512 | #ifdef CONFIG_X86 |
| 5513 | cpu_khz/1000 |
| 5514 | #else |
| 5515 | -1 |
| 5516 | #endif |
| 5517 | ); |
Chuck Ebbert | bd576c9 | 2006-02-04 23:27:42 -0800 | [diff] [blame] | 5518 | if (system_state == SYSTEM_BOOTING) { |
| 5519 | printk("migration_cost="); |
| 5520 | for (distance = 0; distance <= max_distance; distance++) { |
| 5521 | if (distance) |
| 5522 | printk(","); |
| 5523 | printk("%ld", (long)migration_cost[distance] / 1000); |
| 5524 | } |
| 5525 | printk("\n"); |
akpm@osdl.org | 198e2f1 | 2006-01-12 01:05:30 -0800 | [diff] [blame] | 5526 | } |
akpm@osdl.org | 198e2f1 | 2006-01-12 01:05:30 -0800 | [diff] [blame] | 5527 | j1 = jiffies; |
| 5528 | if (migration_debug) |
| 5529 | printk("migration: %ld seconds\n", (j1-j0)/HZ); |
| 5530 | |
| 5531 | /* |
| 5532 | * Move back to the original CPU. NUMA-Q gets confused |
| 5533 | * if we migrate to another quad during bootup. |
| 5534 | */ |
| 5535 | if (raw_smp_processor_id() != orig_cpu) { |
| 5536 | cpumask_t mask = cpumask_of_cpu(orig_cpu), |
| 5537 | saved_mask = current->cpus_allowed; |
| 5538 | |
| 5539 | set_cpus_allowed(current, mask); |
| 5540 | set_cpus_allowed(current, saved_mask); |
| 5541 | } |
| 5542 | } |
| 5543 | |
John Hawkes | 9c1cfda | 2005-09-06 15:18:14 -0700 | [diff] [blame] | 5544 | #ifdef CONFIG_NUMA |
akpm@osdl.org | 198e2f1 | 2006-01-12 01:05:30 -0800 | [diff] [blame] | 5545 | |
John Hawkes | 9c1cfda | 2005-09-06 15:18:14 -0700 | [diff] [blame] | 5546 | /** |
| 5547 | * find_next_best_node - find the next node to include in a sched_domain |
| 5548 | * @node: node whose sched_domain we're building |
| 5549 | * @used_nodes: nodes already in the sched_domain |
| 5550 | * |
| 5551 | * Find the next node to include in a given scheduling domain. Simply |
| 5552 | * finds the closest node not already in the @used_nodes map. |
| 5553 | * |
| 5554 | * Should use nodemask_t. |
| 5555 | */ |
| 5556 | static int find_next_best_node(int node, unsigned long *used_nodes) |
| 5557 | { |
| 5558 | int i, n, val, min_val, best_node = 0; |
| 5559 | |
| 5560 | min_val = INT_MAX; |
| 5561 | |
| 5562 | for (i = 0; i < MAX_NUMNODES; i++) { |
| 5563 | /* Start at @node */ |
| 5564 | n = (node + i) % MAX_NUMNODES; |
| 5565 | |
| 5566 | if (!nr_cpus_node(n)) |
| 5567 | continue; |
| 5568 | |
| 5569 | /* Skip already used nodes */ |
| 5570 | if (test_bit(n, used_nodes)) |
| 5571 | continue; |
| 5572 | |
| 5573 | /* Simple min distance search */ |
| 5574 | val = node_distance(node, n); |
| 5575 | |
| 5576 | if (val < min_val) { |
| 5577 | min_val = val; |
| 5578 | best_node = n; |
| 5579 | } |
| 5580 | } |
| 5581 | |
| 5582 | set_bit(best_node, used_nodes); |
| 5583 | return best_node; |
| 5584 | } |
| 5585 | |
| 5586 | /** |
| 5587 | * sched_domain_node_span - get a cpumask for a node's sched_domain |
| 5588 | * @node: node whose cpumask we're constructing |
| 5589 | * @size: number of nodes to include in this span |
| 5590 | * |
| 5591 | * Given a node, construct a good cpumask for its sched_domain to span. It |
| 5592 | * should be one that prevents unnecessary balancing, but also spreads tasks |
| 5593 | * out optimally. |
| 5594 | */ |
| 5595 | static cpumask_t sched_domain_node_span(int node) |
| 5596 | { |
| 5597 | int i; |
| 5598 | cpumask_t span, nodemask; |
| 5599 | DECLARE_BITMAP(used_nodes, MAX_NUMNODES); |
| 5600 | |
| 5601 | cpus_clear(span); |
| 5602 | bitmap_zero(used_nodes, MAX_NUMNODES); |
| 5603 | |
| 5604 | nodemask = node_to_cpumask(node); |
| 5605 | cpus_or(span, span, nodemask); |
| 5606 | set_bit(node, used_nodes); |
| 5607 | |
| 5608 | for (i = 1; i < SD_NODES_PER_DOMAIN; i++) { |
| 5609 | int next_node = find_next_best_node(node, used_nodes); |
| 5610 | nodemask = node_to_cpumask(next_node); |
| 5611 | cpus_or(span, span, nodemask); |
| 5612 | } |
| 5613 | |
| 5614 | return span; |
| 5615 | } |
| 5616 | #endif |
| 5617 | |
| 5618 | /* |
| 5619 | * At the moment, CONFIG_SCHED_SMT is never defined, but leave it in so we |
| 5620 | * can switch it on easily if needed. |
| 5621 | */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5622 | #ifdef CONFIG_SCHED_SMT |
| 5623 | static DEFINE_PER_CPU(struct sched_domain, cpu_domains); |
| 5624 | static struct sched_group sched_group_cpus[NR_CPUS]; |
Dinakar Guniguntala | 1a20ff2 | 2005-06-25 14:57:33 -0700 | [diff] [blame] | 5625 | static int cpu_to_cpu_group(int cpu) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5626 | { |
| 5627 | return cpu; |
| 5628 | } |
| 5629 | #endif |
| 5630 | |
Siddha, Suresh B | 1e9f28f | 2006-03-27 01:15:22 -0800 | [diff] [blame] | 5631 | #ifdef CONFIG_SCHED_MC |
| 5632 | static DEFINE_PER_CPU(struct sched_domain, core_domains); |
| 5633 | static struct sched_group sched_group_core[NR_CPUS]; |
| 5634 | #endif |
| 5635 | |
| 5636 | #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT) |
| 5637 | static int cpu_to_core_group(int cpu) |
| 5638 | { |
| 5639 | return first_cpu(cpu_sibling_map[cpu]); |
| 5640 | } |
| 5641 | #elif defined(CONFIG_SCHED_MC) |
| 5642 | static int cpu_to_core_group(int cpu) |
| 5643 | { |
| 5644 | return cpu; |
| 5645 | } |
| 5646 | #endif |
| 5647 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5648 | static DEFINE_PER_CPU(struct sched_domain, phys_domains); |
| 5649 | static struct sched_group sched_group_phys[NR_CPUS]; |
Dinakar Guniguntala | 1a20ff2 | 2005-06-25 14:57:33 -0700 | [diff] [blame] | 5650 | static int cpu_to_phys_group(int cpu) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5651 | { |
Siddha, Suresh B | 1e9f28f | 2006-03-27 01:15:22 -0800 | [diff] [blame] | 5652 | #if defined(CONFIG_SCHED_MC) |
| 5653 | cpumask_t mask = cpu_coregroup_map(cpu); |
| 5654 | return first_cpu(mask); |
| 5655 | #elif defined(CONFIG_SCHED_SMT) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5656 | return first_cpu(cpu_sibling_map[cpu]); |
| 5657 | #else |
| 5658 | return cpu; |
| 5659 | #endif |
| 5660 | } |
| 5661 | |
| 5662 | #ifdef CONFIG_NUMA |
John Hawkes | 9c1cfda | 2005-09-06 15:18:14 -0700 | [diff] [blame] | 5663 | /* |
| 5664 | * The init_sched_build_groups can't handle what we want to do with node |
| 5665 | * groups, so roll our own. Now each node has its own list of groups which |
| 5666 | * gets dynamically allocated. |
| 5667 | */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5668 | static DEFINE_PER_CPU(struct sched_domain, node_domains); |
John Hawkes | d1b5513 | 2005-09-06 15:18:14 -0700 | [diff] [blame] | 5669 | static struct sched_group **sched_group_nodes_bycpu[NR_CPUS]; |
John Hawkes | 9c1cfda | 2005-09-06 15:18:14 -0700 | [diff] [blame] | 5670 | |
| 5671 | static DEFINE_PER_CPU(struct sched_domain, allnodes_domains); |
John Hawkes | d1b5513 | 2005-09-06 15:18:14 -0700 | [diff] [blame] | 5672 | static struct sched_group *sched_group_allnodes_bycpu[NR_CPUS]; |
John Hawkes | 9c1cfda | 2005-09-06 15:18:14 -0700 | [diff] [blame] | 5673 | |
| 5674 | static int cpu_to_allnodes_group(int cpu) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5675 | { |
| 5676 | return cpu_to_node(cpu); |
| 5677 | } |
Siddha, Suresh B | 0806903 | 2006-03-27 01:15:23 -0800 | [diff] [blame] | 5678 | static void init_numa_sched_groups_power(struct sched_group *group_head) |
| 5679 | { |
| 5680 | struct sched_group *sg = group_head; |
| 5681 | int j; |
| 5682 | |
| 5683 | if (!sg) |
| 5684 | return; |
| 5685 | next_sg: |
| 5686 | for_each_cpu_mask(j, sg->cpumask) { |
| 5687 | struct sched_domain *sd; |
| 5688 | |
| 5689 | sd = &per_cpu(phys_domains, j); |
| 5690 | if (j != first_cpu(sd->groups->cpumask)) { |
| 5691 | /* |
| 5692 | * Only add "power" once for each |
| 5693 | * physical package. |
| 5694 | */ |
| 5695 | continue; |
| 5696 | } |
| 5697 | |
| 5698 | sg->cpu_power += sd->groups->cpu_power; |
| 5699 | } |
| 5700 | sg = sg->next; |
| 5701 | if (sg != group_head) |
| 5702 | goto next_sg; |
| 5703 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5704 | #endif |
| 5705 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5706 | /* |
Dinakar Guniguntala | 1a20ff2 | 2005-06-25 14:57:33 -0700 | [diff] [blame] | 5707 | * Build sched domains for a given set of cpus and attach the sched domains |
| 5708 | * to the individual cpus |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5709 | */ |
John Hawkes | 9c1cfda | 2005-09-06 15:18:14 -0700 | [diff] [blame] | 5710 | void build_sched_domains(const cpumask_t *cpu_map) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5711 | { |
| 5712 | int i; |
John Hawkes | d1b5513 | 2005-09-06 15:18:14 -0700 | [diff] [blame] | 5713 | #ifdef CONFIG_NUMA |
| 5714 | struct sched_group **sched_group_nodes = NULL; |
| 5715 | struct sched_group *sched_group_allnodes = NULL; |
| 5716 | |
| 5717 | /* |
| 5718 | * Allocate the per-node list of sched groups |
| 5719 | */ |
| 5720 | sched_group_nodes = kmalloc(sizeof(struct sched_group*)*MAX_NUMNODES, |
| 5721 | GFP_ATOMIC); |
| 5722 | if (!sched_group_nodes) { |
| 5723 | printk(KERN_WARNING "Can not alloc sched group node list\n"); |
| 5724 | return; |
| 5725 | } |
| 5726 | sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes; |
| 5727 | #endif |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5728 | |
| 5729 | /* |
Dinakar Guniguntala | 1a20ff2 | 2005-06-25 14:57:33 -0700 | [diff] [blame] | 5730 | * Set up domains for cpus specified by the cpu_map. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5731 | */ |
Dinakar Guniguntala | 1a20ff2 | 2005-06-25 14:57:33 -0700 | [diff] [blame] | 5732 | for_each_cpu_mask(i, *cpu_map) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5733 | int group; |
| 5734 | struct sched_domain *sd = NULL, *p; |
| 5735 | cpumask_t nodemask = node_to_cpumask(cpu_to_node(i)); |
| 5736 | |
Dinakar Guniguntala | 1a20ff2 | 2005-06-25 14:57:33 -0700 | [diff] [blame] | 5737 | cpus_and(nodemask, nodemask, *cpu_map); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5738 | |
| 5739 | #ifdef CONFIG_NUMA |
John Hawkes | d1b5513 | 2005-09-06 15:18:14 -0700 | [diff] [blame] | 5740 | if (cpus_weight(*cpu_map) |
John Hawkes | 9c1cfda | 2005-09-06 15:18:14 -0700 | [diff] [blame] | 5741 | > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) { |
John Hawkes | d1b5513 | 2005-09-06 15:18:14 -0700 | [diff] [blame] | 5742 | if (!sched_group_allnodes) { |
| 5743 | sched_group_allnodes |
| 5744 | = kmalloc(sizeof(struct sched_group) |
| 5745 | * MAX_NUMNODES, |
| 5746 | GFP_KERNEL); |
| 5747 | if (!sched_group_allnodes) { |
| 5748 | printk(KERN_WARNING |
| 5749 | "Can not alloc allnodes sched group\n"); |
| 5750 | break; |
| 5751 | } |
| 5752 | sched_group_allnodes_bycpu[i] |
| 5753 | = sched_group_allnodes; |
| 5754 | } |
John Hawkes | 9c1cfda | 2005-09-06 15:18:14 -0700 | [diff] [blame] | 5755 | sd = &per_cpu(allnodes_domains, i); |
| 5756 | *sd = SD_ALLNODES_INIT; |
| 5757 | sd->span = *cpu_map; |
| 5758 | group = cpu_to_allnodes_group(i); |
| 5759 | sd->groups = &sched_group_allnodes[group]; |
| 5760 | p = sd; |
| 5761 | } else |
| 5762 | p = NULL; |
| 5763 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5764 | sd = &per_cpu(node_domains, i); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5765 | *sd = SD_NODE_INIT; |
John Hawkes | 9c1cfda | 2005-09-06 15:18:14 -0700 | [diff] [blame] | 5766 | sd->span = sched_domain_node_span(cpu_to_node(i)); |
| 5767 | sd->parent = p; |
| 5768 | cpus_and(sd->span, sd->span, *cpu_map); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5769 | #endif |
| 5770 | |
| 5771 | p = sd; |
| 5772 | sd = &per_cpu(phys_domains, i); |
| 5773 | group = cpu_to_phys_group(i); |
| 5774 | *sd = SD_CPU_INIT; |
| 5775 | sd->span = nodemask; |
| 5776 | sd->parent = p; |
| 5777 | sd->groups = &sched_group_phys[group]; |
| 5778 | |
Siddha, Suresh B | 1e9f28f | 2006-03-27 01:15:22 -0800 | [diff] [blame] | 5779 | #ifdef CONFIG_SCHED_MC |
| 5780 | p = sd; |
| 5781 | sd = &per_cpu(core_domains, i); |
| 5782 | group = cpu_to_core_group(i); |
| 5783 | *sd = SD_MC_INIT; |
| 5784 | sd->span = cpu_coregroup_map(i); |
| 5785 | cpus_and(sd->span, sd->span, *cpu_map); |
| 5786 | sd->parent = p; |
| 5787 | sd->groups = &sched_group_core[group]; |
| 5788 | #endif |
| 5789 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5790 | #ifdef CONFIG_SCHED_SMT |
| 5791 | p = sd; |
| 5792 | sd = &per_cpu(cpu_domains, i); |
| 5793 | group = cpu_to_cpu_group(i); |
| 5794 | *sd = SD_SIBLING_INIT; |
| 5795 | sd->span = cpu_sibling_map[i]; |
Dinakar Guniguntala | 1a20ff2 | 2005-06-25 14:57:33 -0700 | [diff] [blame] | 5796 | cpus_and(sd->span, sd->span, *cpu_map); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5797 | sd->parent = p; |
| 5798 | sd->groups = &sched_group_cpus[group]; |
| 5799 | #endif |
| 5800 | } |
| 5801 | |
| 5802 | #ifdef CONFIG_SCHED_SMT |
| 5803 | /* Set up CPU (sibling) groups */ |
John Hawkes | 9c1cfda | 2005-09-06 15:18:14 -0700 | [diff] [blame] | 5804 | for_each_cpu_mask(i, *cpu_map) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5805 | cpumask_t this_sibling_map = cpu_sibling_map[i]; |
Dinakar Guniguntala | 1a20ff2 | 2005-06-25 14:57:33 -0700 | [diff] [blame] | 5806 | cpus_and(this_sibling_map, this_sibling_map, *cpu_map); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5807 | if (i != first_cpu(this_sibling_map)) |
| 5808 | continue; |
| 5809 | |
| 5810 | init_sched_build_groups(sched_group_cpus, this_sibling_map, |
| 5811 | &cpu_to_cpu_group); |
| 5812 | } |
| 5813 | #endif |
| 5814 | |
Siddha, Suresh B | 1e9f28f | 2006-03-27 01:15:22 -0800 | [diff] [blame] | 5815 | #ifdef CONFIG_SCHED_MC |
| 5816 | /* Set up multi-core groups */ |
| 5817 | for_each_cpu_mask(i, *cpu_map) { |
| 5818 | cpumask_t this_core_map = cpu_coregroup_map(i); |
| 5819 | cpus_and(this_core_map, this_core_map, *cpu_map); |
| 5820 | if (i != first_cpu(this_core_map)) |
| 5821 | continue; |
| 5822 | init_sched_build_groups(sched_group_core, this_core_map, |
| 5823 | &cpu_to_core_group); |
| 5824 | } |
| 5825 | #endif |
| 5826 | |
| 5827 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5828 | /* Set up physical groups */ |
| 5829 | for (i = 0; i < MAX_NUMNODES; i++) { |
| 5830 | cpumask_t nodemask = node_to_cpumask(i); |
| 5831 | |
Dinakar Guniguntala | 1a20ff2 | 2005-06-25 14:57:33 -0700 | [diff] [blame] | 5832 | cpus_and(nodemask, nodemask, *cpu_map); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5833 | if (cpus_empty(nodemask)) |
| 5834 | continue; |
| 5835 | |
| 5836 | init_sched_build_groups(sched_group_phys, nodemask, |
| 5837 | &cpu_to_phys_group); |
| 5838 | } |
| 5839 | |
| 5840 | #ifdef CONFIG_NUMA |
| 5841 | /* Set up node groups */ |
John Hawkes | d1b5513 | 2005-09-06 15:18:14 -0700 | [diff] [blame] | 5842 | if (sched_group_allnodes) |
| 5843 | init_sched_build_groups(sched_group_allnodes, *cpu_map, |
| 5844 | &cpu_to_allnodes_group); |
John Hawkes | 9c1cfda | 2005-09-06 15:18:14 -0700 | [diff] [blame] | 5845 | |
| 5846 | for (i = 0; i < MAX_NUMNODES; i++) { |
| 5847 | /* Set up node groups */ |
| 5848 | struct sched_group *sg, *prev; |
| 5849 | cpumask_t nodemask = node_to_cpumask(i); |
| 5850 | cpumask_t domainspan; |
| 5851 | cpumask_t covered = CPU_MASK_NONE; |
| 5852 | int j; |
| 5853 | |
| 5854 | cpus_and(nodemask, nodemask, *cpu_map); |
John Hawkes | d1b5513 | 2005-09-06 15:18:14 -0700 | [diff] [blame] | 5855 | if (cpus_empty(nodemask)) { |
| 5856 | sched_group_nodes[i] = NULL; |
John Hawkes | 9c1cfda | 2005-09-06 15:18:14 -0700 | [diff] [blame] | 5857 | continue; |
John Hawkes | d1b5513 | 2005-09-06 15:18:14 -0700 | [diff] [blame] | 5858 | } |
John Hawkes | 9c1cfda | 2005-09-06 15:18:14 -0700 | [diff] [blame] | 5859 | |
| 5860 | domainspan = sched_domain_node_span(i); |
| 5861 | cpus_and(domainspan, domainspan, *cpu_map); |
| 5862 | |
| 5863 | sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL); |
| 5864 | sched_group_nodes[i] = sg; |
| 5865 | for_each_cpu_mask(j, nodemask) { |
| 5866 | struct sched_domain *sd; |
| 5867 | sd = &per_cpu(node_domains, j); |
| 5868 | sd->groups = sg; |
| 5869 | if (sd->groups == NULL) { |
| 5870 | /* Turn off balancing if we have no groups */ |
| 5871 | sd->flags = 0; |
| 5872 | } |
| 5873 | } |
| 5874 | if (!sg) { |
| 5875 | printk(KERN_WARNING |
| 5876 | "Can not alloc domain group for node %d\n", i); |
| 5877 | continue; |
| 5878 | } |
| 5879 | sg->cpu_power = 0; |
| 5880 | sg->cpumask = nodemask; |
| 5881 | cpus_or(covered, covered, nodemask); |
| 5882 | prev = sg; |
| 5883 | |
| 5884 | for (j = 0; j < MAX_NUMNODES; j++) { |
| 5885 | cpumask_t tmp, notcovered; |
| 5886 | int n = (i + j) % MAX_NUMNODES; |
| 5887 | |
| 5888 | cpus_complement(notcovered, covered); |
| 5889 | cpus_and(tmp, notcovered, *cpu_map); |
| 5890 | cpus_and(tmp, tmp, domainspan); |
| 5891 | if (cpus_empty(tmp)) |
| 5892 | break; |
| 5893 | |
| 5894 | nodemask = node_to_cpumask(n); |
| 5895 | cpus_and(tmp, tmp, nodemask); |
| 5896 | if (cpus_empty(tmp)) |
| 5897 | continue; |
| 5898 | |
| 5899 | sg = kmalloc(sizeof(struct sched_group), GFP_KERNEL); |
| 5900 | if (!sg) { |
| 5901 | printk(KERN_WARNING |
| 5902 | "Can not alloc domain group for node %d\n", j); |
| 5903 | break; |
| 5904 | } |
| 5905 | sg->cpu_power = 0; |
| 5906 | sg->cpumask = tmp; |
| 5907 | cpus_or(covered, covered, tmp); |
| 5908 | prev->next = sg; |
| 5909 | prev = sg; |
| 5910 | } |
| 5911 | prev->next = sched_group_nodes[i]; |
| 5912 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5913 | #endif |
| 5914 | |
| 5915 | /* Calculate CPU power for physical packages and nodes */ |
Dinakar Guniguntala | 1a20ff2 | 2005-06-25 14:57:33 -0700 | [diff] [blame] | 5916 | for_each_cpu_mask(i, *cpu_map) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5917 | int power; |
| 5918 | struct sched_domain *sd; |
| 5919 | #ifdef CONFIG_SCHED_SMT |
| 5920 | sd = &per_cpu(cpu_domains, i); |
| 5921 | power = SCHED_LOAD_SCALE; |
| 5922 | sd->groups->cpu_power = power; |
| 5923 | #endif |
Siddha, Suresh B | 1e9f28f | 2006-03-27 01:15:22 -0800 | [diff] [blame] | 5924 | #ifdef CONFIG_SCHED_MC |
| 5925 | sd = &per_cpu(core_domains, i); |
| 5926 | power = SCHED_LOAD_SCALE + (cpus_weight(sd->groups->cpumask)-1) |
| 5927 | * SCHED_LOAD_SCALE / 10; |
| 5928 | sd->groups->cpu_power = power; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5929 | |
| 5930 | sd = &per_cpu(phys_domains, i); |
Siddha, Suresh B | 1e9f28f | 2006-03-27 01:15:22 -0800 | [diff] [blame] | 5931 | |
| 5932 | /* |
| 5933 | * This has to be < 2 * SCHED_LOAD_SCALE |
| 5934 | * Lets keep it SCHED_LOAD_SCALE, so that |
| 5935 | * while calculating NUMA group's cpu_power |
| 5936 | * we can simply do |
| 5937 | * numa_group->cpu_power += phys_group->cpu_power; |
| 5938 | * |
| 5939 | * See "only add power once for each physical pkg" |
| 5940 | * comment below |
| 5941 | */ |
| 5942 | sd->groups->cpu_power = SCHED_LOAD_SCALE; |
| 5943 | #else |
| 5944 | sd = &per_cpu(phys_domains, i); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5945 | power = SCHED_LOAD_SCALE + SCHED_LOAD_SCALE * |
| 5946 | (cpus_weight(sd->groups->cpumask)-1) / 10; |
| 5947 | sd->groups->cpu_power = power; |
Siddha, Suresh B | 1e9f28f | 2006-03-27 01:15:22 -0800 | [diff] [blame] | 5948 | #endif |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5949 | } |
| 5950 | |
John Hawkes | 9c1cfda | 2005-09-06 15:18:14 -0700 | [diff] [blame] | 5951 | #ifdef CONFIG_NUMA |
Siddha, Suresh B | 0806903 | 2006-03-27 01:15:23 -0800 | [diff] [blame] | 5952 | for (i = 0; i < MAX_NUMNODES; i++) |
| 5953 | init_numa_sched_groups_power(sched_group_nodes[i]); |
John Hawkes | 9c1cfda | 2005-09-06 15:18:14 -0700 | [diff] [blame] | 5954 | |
Siddha, Suresh B | 0806903 | 2006-03-27 01:15:23 -0800 | [diff] [blame] | 5955 | init_numa_sched_groups_power(sched_group_allnodes); |
John Hawkes | 9c1cfda | 2005-09-06 15:18:14 -0700 | [diff] [blame] | 5956 | #endif |
| 5957 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5958 | /* Attach the domains */ |
Dinakar Guniguntala | 1a20ff2 | 2005-06-25 14:57:33 -0700 | [diff] [blame] | 5959 | for_each_cpu_mask(i, *cpu_map) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5960 | struct sched_domain *sd; |
| 5961 | #ifdef CONFIG_SCHED_SMT |
| 5962 | sd = &per_cpu(cpu_domains, i); |
Siddha, Suresh B | 1e9f28f | 2006-03-27 01:15:22 -0800 | [diff] [blame] | 5963 | #elif defined(CONFIG_SCHED_MC) |
| 5964 | sd = &per_cpu(core_domains, i); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5965 | #else |
| 5966 | sd = &per_cpu(phys_domains, i); |
| 5967 | #endif |
| 5968 | cpu_attach_domain(sd, i); |
| 5969 | } |
akpm@osdl.org | 198e2f1 | 2006-01-12 01:05:30 -0800 | [diff] [blame] | 5970 | /* |
| 5971 | * Tune cache-hot values: |
| 5972 | */ |
| 5973 | calibrate_migration_costs(cpu_map); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5974 | } |
Dinakar Guniguntala | 1a20ff2 | 2005-06-25 14:57:33 -0700 | [diff] [blame] | 5975 | /* |
| 5976 | * Set up scheduler domains and groups. Callers must hold the hotplug lock. |
| 5977 | */ |
John Hawkes | 9c1cfda | 2005-09-06 15:18:14 -0700 | [diff] [blame] | 5978 | static void arch_init_sched_domains(const cpumask_t *cpu_map) |
Dinakar Guniguntala | 1a20ff2 | 2005-06-25 14:57:33 -0700 | [diff] [blame] | 5979 | { |
| 5980 | cpumask_t cpu_default_map; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5981 | |
Dinakar Guniguntala | 1a20ff2 | 2005-06-25 14:57:33 -0700 | [diff] [blame] | 5982 | /* |
| 5983 | * Setup mask for cpus without special case scheduling requirements. |
| 5984 | * For now this just excludes isolated cpus, but could be used to |
| 5985 | * exclude other special cases in the future. |
| 5986 | */ |
| 5987 | cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map); |
| 5988 | |
| 5989 | build_sched_domains(&cpu_default_map); |
| 5990 | } |
| 5991 | |
| 5992 | static void arch_destroy_sched_domains(const cpumask_t *cpu_map) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5993 | { |
John Hawkes | 9c1cfda | 2005-09-06 15:18:14 -0700 | [diff] [blame] | 5994 | #ifdef CONFIG_NUMA |
| 5995 | int i; |
John Hawkes | d1b5513 | 2005-09-06 15:18:14 -0700 | [diff] [blame] | 5996 | int cpu; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5997 | |
John Hawkes | d1b5513 | 2005-09-06 15:18:14 -0700 | [diff] [blame] | 5998 | for_each_cpu_mask(cpu, *cpu_map) { |
| 5999 | struct sched_group *sched_group_allnodes |
| 6000 | = sched_group_allnodes_bycpu[cpu]; |
| 6001 | struct sched_group **sched_group_nodes |
| 6002 | = sched_group_nodes_bycpu[cpu]; |
| 6003 | |
| 6004 | if (sched_group_allnodes) { |
| 6005 | kfree(sched_group_allnodes); |
| 6006 | sched_group_allnodes_bycpu[cpu] = NULL; |
| 6007 | } |
| 6008 | |
| 6009 | if (!sched_group_nodes) |
John Hawkes | 9c1cfda | 2005-09-06 15:18:14 -0700 | [diff] [blame] | 6010 | continue; |
| 6011 | |
John Hawkes | d1b5513 | 2005-09-06 15:18:14 -0700 | [diff] [blame] | 6012 | for (i = 0; i < MAX_NUMNODES; i++) { |
| 6013 | cpumask_t nodemask = node_to_cpumask(i); |
| 6014 | struct sched_group *oldsg, *sg = sched_group_nodes[i]; |
| 6015 | |
| 6016 | cpus_and(nodemask, nodemask, *cpu_map); |
| 6017 | if (cpus_empty(nodemask)) |
| 6018 | continue; |
| 6019 | |
| 6020 | if (sg == NULL) |
| 6021 | continue; |
| 6022 | sg = sg->next; |
John Hawkes | 9c1cfda | 2005-09-06 15:18:14 -0700 | [diff] [blame] | 6023 | next_sg: |
John Hawkes | d1b5513 | 2005-09-06 15:18:14 -0700 | [diff] [blame] | 6024 | oldsg = sg; |
| 6025 | sg = sg->next; |
| 6026 | kfree(oldsg); |
| 6027 | if (oldsg != sched_group_nodes[i]) |
| 6028 | goto next_sg; |
| 6029 | } |
| 6030 | kfree(sched_group_nodes); |
| 6031 | sched_group_nodes_bycpu[cpu] = NULL; |
John Hawkes | 9c1cfda | 2005-09-06 15:18:14 -0700 | [diff] [blame] | 6032 | } |
| 6033 | #endif |
| 6034 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 6035 | |
Dinakar Guniguntala | 1a20ff2 | 2005-06-25 14:57:33 -0700 | [diff] [blame] | 6036 | /* |
| 6037 | * Detach sched domains from a group of cpus specified in cpu_map |
| 6038 | * These cpus will now be attached to the NULL domain |
| 6039 | */ |
Arjan van de Ven | 858119e | 2006-01-14 13:20:43 -0800 | [diff] [blame] | 6040 | static void detach_destroy_domains(const cpumask_t *cpu_map) |
Dinakar Guniguntala | 1a20ff2 | 2005-06-25 14:57:33 -0700 | [diff] [blame] | 6041 | { |
| 6042 | int i; |
| 6043 | |
| 6044 | for_each_cpu_mask(i, *cpu_map) |
| 6045 | cpu_attach_domain(NULL, i); |
| 6046 | synchronize_sched(); |
| 6047 | arch_destroy_sched_domains(cpu_map); |
| 6048 | } |
| 6049 | |
| 6050 | /* |
| 6051 | * Partition sched domains as specified by the cpumasks below. |
| 6052 | * This attaches all cpus from the cpumasks to the NULL domain, |
| 6053 | * waits for a RCU quiescent period, recalculates sched |
| 6054 | * domain information and then attaches them back to the |
| 6055 | * correct sched domains |
| 6056 | * Call with hotplug lock held |
| 6057 | */ |
| 6058 | void partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2) |
| 6059 | { |
| 6060 | cpumask_t change_map; |
| 6061 | |
| 6062 | cpus_and(*partition1, *partition1, cpu_online_map); |
| 6063 | cpus_and(*partition2, *partition2, cpu_online_map); |
| 6064 | cpus_or(change_map, *partition1, *partition2); |
| 6065 | |
| 6066 | /* Detach sched domains from all of the affected cpus */ |
| 6067 | detach_destroy_domains(&change_map); |
| 6068 | if (!cpus_empty(*partition1)) |
| 6069 | build_sched_domains(partition1); |
| 6070 | if (!cpus_empty(*partition2)) |
| 6071 | build_sched_domains(partition2); |
| 6072 | } |
| 6073 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 6074 | #ifdef CONFIG_HOTPLUG_CPU |
| 6075 | /* |
| 6076 | * Force a reinitialization of the sched domains hierarchy. The domains |
| 6077 | * and groups cannot be updated in place without racing with the balancing |
Nick Piggin | 41c7ce9 | 2005-06-25 14:57:24 -0700 | [diff] [blame] | 6078 | * code, so we temporarily attach all running cpus to the NULL domain |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 6079 | * which will prevent rebalancing while the sched domains are recalculated. |
| 6080 | */ |
| 6081 | static int update_sched_domains(struct notifier_block *nfb, |
| 6082 | unsigned long action, void *hcpu) |
| 6083 | { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 6084 | switch (action) { |
| 6085 | case CPU_UP_PREPARE: |
| 6086 | case CPU_DOWN_PREPARE: |
Dinakar Guniguntala | 1a20ff2 | 2005-06-25 14:57:33 -0700 | [diff] [blame] | 6087 | detach_destroy_domains(&cpu_online_map); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 6088 | return NOTIFY_OK; |
| 6089 | |
| 6090 | case CPU_UP_CANCELED: |
| 6091 | case CPU_DOWN_FAILED: |
| 6092 | case CPU_ONLINE: |
| 6093 | case CPU_DEAD: |
| 6094 | /* |
| 6095 | * Fall through and re-initialise the domains. |
| 6096 | */ |
| 6097 | break; |
| 6098 | default: |
| 6099 | return NOTIFY_DONE; |
| 6100 | } |
| 6101 | |
| 6102 | /* The hotplug lock is already held by cpu_up/cpu_down */ |
Dinakar Guniguntala | 1a20ff2 | 2005-06-25 14:57:33 -0700 | [diff] [blame] | 6103 | arch_init_sched_domains(&cpu_online_map); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 6104 | |
| 6105 | return NOTIFY_OK; |
| 6106 | } |
| 6107 | #endif |
| 6108 | |
| 6109 | void __init sched_init_smp(void) |
| 6110 | { |
| 6111 | lock_cpu_hotplug(); |
Dinakar Guniguntala | 1a20ff2 | 2005-06-25 14:57:33 -0700 | [diff] [blame] | 6112 | arch_init_sched_domains(&cpu_online_map); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 6113 | unlock_cpu_hotplug(); |
| 6114 | /* XXX: Theoretical race here - CPU may be hotplugged now */ |
| 6115 | hotcpu_notifier(update_sched_domains, 0); |
| 6116 | } |
| 6117 | #else |
| 6118 | void __init sched_init_smp(void) |
| 6119 | { |
| 6120 | } |
| 6121 | #endif /* CONFIG_SMP */ |
| 6122 | |
| 6123 | int in_sched_functions(unsigned long addr) |
| 6124 | { |
| 6125 | /* Linker adds these: start and end of __sched functions */ |
| 6126 | extern char __sched_text_start[], __sched_text_end[]; |
| 6127 | return in_lock_functions(addr) || |
| 6128 | (addr >= (unsigned long)__sched_text_start |
| 6129 | && addr < (unsigned long)__sched_text_end); |
| 6130 | } |
| 6131 | |
| 6132 | void __init sched_init(void) |
| 6133 | { |
| 6134 | runqueue_t *rq; |
| 6135 | int i, j, k; |
| 6136 | |
KAMEZAWA Hiroyuki | 0a94502 | 2006-03-28 01:56:37 -0800 | [diff] [blame] | 6137 | for_each_possible_cpu(i) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 6138 | prio_array_t *array; |
| 6139 | |
| 6140 | rq = cpu_rq(i); |
| 6141 | spin_lock_init(&rq->lock); |
Nick Piggin | 7897986 | 2005-06-25 14:57:13 -0700 | [diff] [blame] | 6142 | rq->nr_running = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 6143 | rq->active = rq->arrays; |
| 6144 | rq->expired = rq->arrays + 1; |
| 6145 | rq->best_expired_prio = MAX_PRIO; |
| 6146 | |
| 6147 | #ifdef CONFIG_SMP |
Nick Piggin | 41c7ce9 | 2005-06-25 14:57:24 -0700 | [diff] [blame] | 6148 | rq->sd = NULL; |
Nick Piggin | 7897986 | 2005-06-25 14:57:13 -0700 | [diff] [blame] | 6149 | for (j = 1; j < 3; j++) |
| 6150 | rq->cpu_load[j] = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 6151 | rq->active_balance = 0; |
| 6152 | rq->push_cpu = 0; |
| 6153 | rq->migration_thread = NULL; |
| 6154 | INIT_LIST_HEAD(&rq->migration_queue); |
Anton Blanchard | e9028b0 | 2006-03-23 02:59:20 -0800 | [diff] [blame] | 6155 | rq->cpu = i; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 6156 | #endif |
| 6157 | atomic_set(&rq->nr_iowait, 0); |
| 6158 | |
| 6159 | for (j = 0; j < 2; j++) { |
| 6160 | array = rq->arrays + j; |
| 6161 | for (k = 0; k < MAX_PRIO; k++) { |
| 6162 | INIT_LIST_HEAD(array->queue + k); |
| 6163 | __clear_bit(k, array->bitmap); |
| 6164 | } |
| 6165 | // delimiter for bitsearch |
| 6166 | __set_bit(MAX_PRIO, array->bitmap); |
| 6167 | } |
| 6168 | } |
| 6169 | |
| 6170 | /* |
| 6171 | * The boot idle thread does lazy MMU switching as well: |
| 6172 | */ |
| 6173 | atomic_inc(&init_mm.mm_count); |
| 6174 | enter_lazy_tlb(&init_mm, current); |
| 6175 | |
| 6176 | /* |
| 6177 | * Make us the idle thread. Technically, schedule() should not be |
| 6178 | * called from this thread, however somewhere below it might be, |
| 6179 | * but because we are the idle thread, we just pick up running again |
| 6180 | * when this runqueue becomes "idle". |
| 6181 | */ |
| 6182 | init_idle(current, smp_processor_id()); |
| 6183 | } |
| 6184 | |
| 6185 | #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP |
| 6186 | void __might_sleep(char *file, int line) |
| 6187 | { |
| 6188 | #if defined(in_atomic) |
| 6189 | static unsigned long prev_jiffy; /* ratelimiting */ |
| 6190 | |
| 6191 | if ((in_atomic() || irqs_disabled()) && |
| 6192 | system_state == SYSTEM_RUNNING && !oops_in_progress) { |
| 6193 | if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) |
| 6194 | return; |
| 6195 | prev_jiffy = jiffies; |
Ingo Molnar | 91368d7 | 2006-03-23 03:00:54 -0800 | [diff] [blame] | 6196 | printk(KERN_ERR "BUG: sleeping function called from invalid" |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 6197 | " context at %s:%d\n", file, line); |
| 6198 | printk("in_atomic():%d, irqs_disabled():%d\n", |
| 6199 | in_atomic(), irqs_disabled()); |
| 6200 | dump_stack(); |
| 6201 | } |
| 6202 | #endif |
| 6203 | } |
| 6204 | EXPORT_SYMBOL(__might_sleep); |
| 6205 | #endif |
| 6206 | |
| 6207 | #ifdef CONFIG_MAGIC_SYSRQ |
| 6208 | void normalize_rt_tasks(void) |
| 6209 | { |
| 6210 | struct task_struct *p; |
| 6211 | prio_array_t *array; |
| 6212 | unsigned long flags; |
| 6213 | runqueue_t *rq; |
| 6214 | |
| 6215 | read_lock_irq(&tasklist_lock); |
| 6216 | for_each_process (p) { |
| 6217 | if (!rt_task(p)) |
| 6218 | continue; |
| 6219 | |
| 6220 | rq = task_rq_lock(p, &flags); |
| 6221 | |
| 6222 | array = p->array; |
| 6223 | if (array) |
| 6224 | deactivate_task(p, task_rq(p)); |
| 6225 | __setscheduler(p, SCHED_NORMAL, 0); |
| 6226 | if (array) { |
| 6227 | __activate_task(p, task_rq(p)); |
| 6228 | resched_task(rq->curr); |
| 6229 | } |
| 6230 | |
| 6231 | task_rq_unlock(rq, &flags); |
| 6232 | } |
| 6233 | read_unlock_irq(&tasklist_lock); |
| 6234 | } |
| 6235 | |
| 6236 | #endif /* CONFIG_MAGIC_SYSRQ */ |
Linus Torvalds | 1df5c10 | 2005-09-12 07:59:21 -0700 | [diff] [blame] | 6237 | |
| 6238 | #ifdef CONFIG_IA64 |
| 6239 | /* |
| 6240 | * These functions are only useful for the IA64 MCA handling. |
| 6241 | * |
| 6242 | * They can only be called when the whole system has been |
| 6243 | * stopped - every CPU needs to be quiescent, and no scheduling |
| 6244 | * activity can take place. Using them for anything else would |
| 6245 | * be a serious bug, and as a result, they aren't even visible |
| 6246 | * under any other configuration. |
| 6247 | */ |
| 6248 | |
| 6249 | /** |
| 6250 | * curr_task - return the current task for a given cpu. |
| 6251 | * @cpu: the processor in question. |
| 6252 | * |
| 6253 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! |
| 6254 | */ |
| 6255 | task_t *curr_task(int cpu) |
| 6256 | { |
| 6257 | return cpu_curr(cpu); |
| 6258 | } |
| 6259 | |
| 6260 | /** |
| 6261 | * set_curr_task - set the current task for a given cpu. |
| 6262 | * @cpu: the processor in question. |
| 6263 | * @p: the task pointer to set. |
| 6264 | * |
| 6265 | * Description: This function must only be used when non-maskable interrupts |
| 6266 | * are serviced on a separate stack. It allows the architecture to switch the |
| 6267 | * notion of the current task on a cpu in a non-blocking manner. This function |
| 6268 | * must be called with all CPU's synchronized, and interrupts disabled, the |
| 6269 | * and caller must save the original value of the current task (see |
| 6270 | * curr_task() above) and restore that value before reenabling interrupts and |
| 6271 | * re-starting the system. |
| 6272 | * |
| 6273 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! |
| 6274 | */ |
| 6275 | void set_curr_task(int cpu, task_t *p) |
| 6276 | { |
| 6277 | cpu_curr(cpu) = p; |
| 6278 | } |
| 6279 | |
| 6280 | #endif |