kernel/softirq.c

   1 /*
   2  *      linux/kernel/softirq.c
   3  *
   4  *      Copyright (C) 1992 Linus Torvalds
   5  *
   6  *      Distribute under GPLv2.
   7  *
   8  *      Rewritten. Old one was good in 2.2, but in 2.3 it was immoral. --ANK (990903)
   9  *
  10  *      Remote softirq infrastructure is by Jens Axboe.
  11  */
  12
  13 #include <linux/module.h>
  14 #include <linux/kernel_stat.h>
  15 #include <linux/interrupt.h>
  16 #include <linux/init.h>
  17 #include <linux/mm.h>
  18 #include <linux/notifier.h>
  19 #include <linux/percpu.h>
  20 #include <linux/cpu.h>
  21 #include <linux/freezer.h>
  22 #include <linux/kthread.h>
  23 #include <linux/rcupdate.h>
  24 #include <linux/ftrace.h>
  25 #include <linux/smp.h>
  26 #include <linux/tick.h>
  27
  28 #define CREATE_TRACE_POINTS
  29 #include <trace/events/irq.h>
  30
  31 #include <asm/irq.h>
  32 /*
  33    - No shared variables, all the data are CPU local.
  34    - If a softirq needs serialization, let it serialize itself
  35      by its own spinlocks.
  36    - Even if softirq is serialized, only local cpu is marked for
  37      execution. Hence, we get something sort of weak cpu binding.
  38      Though it is still not clear, will it result in better locality
  39      or will not.
  40
  41    Examples:
  42    - NET RX softirq. It is multithreaded and does not require
  43      any global serialization.
  44    - NET TX softirq. It kicks software netdevice queues, hence
  45      it is logically serialized per device, but this serialization
  46      is invisible to common code.
  47    - Tasklets: serialized wrt itself.
  48  */
  49
  50 #ifndef __ARCH_IRQ_STAT
  51 irq_cpustat_t irq_stat[NR_CPUS] ____cacheline_aligned;
  52 EXPORT_SYMBOL(irq_stat);
  53 #endif
  54
  55 static struct softirq_action softirq_vec[NR_SOFTIRQS] __cacheline_aligned_in_smp;
  56
  57 static DEFINE_PER_CPU(struct task_struct *, ksoftirqd);
  58
  59 char *softirq_to_name[NR_SOFTIRQS] = {
  60         "HI", "TIMER", "NET_TX", "NET_RX", "BLOCK", "BLOCK_IOPOLL",
  61         "TASKLET", "SCHED", "HRTIMER",  "RCU"
  62 };
  63
  64 /*
  65  * we cannot loop indefinitely here to avoid userspace starvation,
  66  * but we also don't want to introduce a worst case 1/HZ latency
  67  * to the pending events, so lets the scheduler to balance
  68  * the softirq load for us.
  69  */
  70 void wakeup_softirqd(void)
  71 {
  72         /* Interrupts are disabled: no need to stop preemption */
  73         struct task_struct *tsk = __get_cpu_var(ksoftirqd);
  74
  75         if (tsk && tsk->state != TASK_RUNNING)
  76                 wake_up_process(tsk);
  77 }
  78
  79 /*
  80  * This one is for softirq.c-internal use,
  81  * where hardirqs are disabled legitimately:
  82  */
  83 #ifdef CONFIG_TRACE_IRQFLAGS
  84 static void __local_bh_disable(unsigned long ip)
  85 {
  86         unsigned long flags;
  87
  88         WARN_ON_ONCE(in_irq());
  89
  90         raw_local_irq_save(flags);
  91         /*
  92          * The preempt tracer hooks into add_preempt_count and will break
  93          * lockdep because it calls back into lockdep after SOFTIRQ_OFFSET
  94          * is set and before current->softirq_enabled is cleared.
  95          * We must manually increment preempt_count here and manually
  96          * call the trace_preempt_off later.
  97          */
  98         preempt_count() += SOFTIRQ_OFFSET;
  99         /*
 100          * Were softirqs turned off above:
 101          */
 102         if (softirq_count() == SOFTIRQ_OFFSET)
 103                 trace_softirqs_off(ip);
 104         raw_local_irq_restore(flags);
 105
 106         if (preempt_count() == SOFTIRQ_OFFSET)
 107                 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
 108 }
 109 #else /* !CONFIG_TRACE_IRQFLAGS */
 110 static inline void __local_bh_disable(unsigned long ip)
 111 {
 112         add_preempt_count(SOFTIRQ_OFFSET);
 113         barrier();
 114 }
 115 #endif /* CONFIG_TRACE_IRQFLAGS */
 116
 117 void local_bh_disable(void)
 118 {
 119         __local_bh_disable((unsigned long)__builtin_return_address(0));
 120 }
 121
 122 EXPORT_SYMBOL(local_bh_disable);
 123
 124 /*
 125  * Special-case - softirqs can safely be enabled in
 126  * cond_resched_softirq(), or by __do_softirq(),
 127  * without processing still-pending softirqs:
 128  */
 129 void _local_bh_enable(void)
 130 {
 131         WARN_ON_ONCE(in_irq());
 132         WARN_ON_ONCE(!irqs_disabled());
 133
 134         if (softirq_count() == SOFTIRQ_OFFSET)
 135                 trace_softirqs_on((unsigned long)__builtin_return_address(0));
 136         sub_preempt_count(SOFTIRQ_OFFSET);
 137 }
 138
 139 EXPORT_SYMBOL(_local_bh_enable);
 140
 141 static inline void _local_bh_enable_ip(unsigned long ip)
 142 {
 143         WARN_ON_ONCE(in_irq() || irqs_disabled());
 144 #ifdef CONFIG_TRACE_IRQFLAGS
 145         local_irq_disable();
 146 #endif
 147         /*
 148          * Are softirqs going to be turned on now:
 149          */
 150         if (softirq_count() == SOFTIRQ_OFFSET)
 151                 trace_softirqs_on(ip);
 152         /*
 153          * Keep preemption disabled until we are done with
 154          * softirq processing:
 155          */
 156         sub_preempt_count(SOFTIRQ_OFFSET - 1);
 157
 158         if (unlikely(!in_interrupt() && local_softirq_pending()))
 159                 do_softirq();
 160
 161         dec_preempt_count();
 162 #ifdef CONFIG_TRACE_IRQFLAGS
 163         local_irq_enable();
 164 #endif
 165         preempt_check_resched();
 166 }
 167
 168 void local_bh_enable(void)
 169 {
 170         _local_bh_enable_ip((unsigned long)__builtin_return_address(0));
 171 }
 172 EXPORT_SYMBOL(local_bh_enable);
 173
 174 void local_bh_enable_ip(unsigned long ip)
 175 {
 176         _local_bh_enable_ip(ip);
 177 }
 178 EXPORT_SYMBOL(local_bh_enable_ip);
 179
 180 /*
 181  * We restart softirq processing MAX_SOFTIRQ_RESTART times,
 182  * and we fall back to softirqd after that.
 183  *
 184  * This number has been established via experimentation.
 185  * The two things to balance is latency against fairness -
 186  * we want to handle softirqs as soon as possible, but they
 187  * should not be able to lock up the box.
 188  */
 189 #define MAX_SOFTIRQ_RESTART 10
 190
 191 asmlinkage void __do_softirq(void)
 192 {
 193         struct softirq_action *h;
 194         __u32 pending;
 195         int max_restart = MAX_SOFTIRQ_RESTART;
 196         int cpu;
 197
 198         pending = local_softirq_pending();
 199         account_system_vtime(current);
 200
 201         __local_bh_disable((unsigned long)__builtin_return_address(0));
 202         lockdep_softirq_enter();
 203
 204         cpu = smp_processor_id();
 205 restart:
 206         /* Reset the pending bitmask before enabling irqs */
 207         set_softirq_pending(0);
 208
 209         local_irq_enable();
 210
 211         h = softirq_vec;
 212
 213         do {
 214                 if (pending & 1) {
 215                         int prev_count = preempt_count();
 216                         kstat_incr_softirqs_this_cpu(h - softirq_vec);
 217
 218                         trace_softirq_entry(h, softirq_vec);
 219                         h->action(h);
 220                         trace_softirq_exit(h, softirq_vec);
 221                         if (unlikely(prev_count != preempt_count())) {
 222                                 printk(KERN_ERR "huh, entered softirq %td %s %p"
 223                                        "with preempt_count %08x,"
 224                                        " exited with %08x?\n", h - softirq_vec,
 225                                        softirq_to_name[h - softirq_vec],
 226                                        h->action, prev_count, preempt_count());
 227                                 preempt_count() = prev_count;
 228                         }
 229
 230                         rcu_bh_qs(cpu);
 231                 }
 232                 h++;
 233                 pending >>= 1;
 234         } while (pending);
 235
 236         local_irq_disable();
 237
 238         pending = local_softirq_pending();
 239         if (pending && --max_restart)
 240                 goto restart;
 241
 242         if (pending)
 243                 wakeup_softirqd();
 244
 245         lockdep_softirq_exit();
 246
 247         account_system_vtime(current);
 248         _local_bh_enable();
 249 }
 250
 251 #ifndef __ARCH_HAS_DO_SOFTIRQ
 252
 253 asmlinkage void do_softirq(void)
 254 {
 255         __u32 pending;
 256         unsigned long flags;
 257
 258         if (in_interrupt())
 259                 return;
 260
 261         local_irq_save(flags);
 262
 263         pending = local_softirq_pending();
 264
 265         if (pending)
 266                 __do_softirq();
 267
 268         local_irq_restore(flags);
 269 }
 270
 271 #endif
 272
 273 /*
 274  * Enter an interrupt context.
 275  */
 276 void irq_enter(void)
 277 {
 278         int cpu = smp_processor_id();
 279
 280         rcu_irq_enter();
 281         if (idle_cpu(cpu) && !in_interrupt()) {
 282                 __irq_enter();
 283                 tick_check_idle(cpu);
 284         } else
 285                 __irq_enter();
 286 }
 287
 288 #ifdef __ARCH_IRQ_EXIT_IRQS_DISABLED
 289 # define invoke_softirq()       __do_softirq()
 290 #else
 291 # define invoke_softirq()       do_softirq()
 292 #endif
 293
 294 /*
 295  * Exit an interrupt context. Process softirqs if needed and possible:
 296  */
 297 void irq_exit(void)
 298 {
 299         account_system_vtime(current);
 300         trace_hardirq_exit();
 301         sub_preempt_count(IRQ_EXIT_OFFSET);
 302         if (!in_interrupt() && local_softirq_pending())
 303                 invoke_softirq();
 304
 305         rcu_irq_exit();
 306 #ifdef CONFIG_NO_HZ
 307         /* Make sure that timer wheel updates are propagated */
 308         if (idle_cpu(smp_processor_id()) && !in_interrupt() && !need_resched())
 309                 tick_nohz_stop_sched_tick(0);
 310 #endif
 311         preempt_enable_no_resched();
 312 }
 313
 314 /*
 315  * This function must run with irqs disabled!
 316  */
 317 inline void raise_softirq_irqoff(unsigned int nr)
 318 {
 319         __raise_softirq_irqoff(nr);
 320
 321         /*
 322          * If we're in an interrupt or softirq, we're done
 323          * (this also catches softirq-disabled code). We will
 324          * actually run the softirq once we return from
 325          * the irq or softirq.
 326          *
 327          * Otherwise we wake up ksoftirqd to make sure we
 328          * schedule the softirq soon.
 329          */
 330         if (!in_interrupt())
 331                 wakeup_softirqd();
 332 }
 333
 334 void raise_softirq(unsigned int nr)
 335 {
 336         unsigned long flags;
 337
 338         local_irq_save(flags);
 339         raise_softirq_irqoff(nr);
 340         local_irq_restore(flags);
 341 }
 342
 343 void open_softirq(int nr, void (*action)(struct softirq_action *))
 344 {
 345         softirq_vec[nr].action = action;
 346 }
 347
 348 /*
 349  * Tasklets
 350  */
 351 struct tasklet_head
 352 {
 353         struct tasklet_struct *head;
 354         struct tasklet_struct **tail;
 355 };
 356
 357 static DEFINE_PER_CPU(struct tasklet_head, tasklet_vec);
 358 static DEFINE_PER_CPU(struct tasklet_head, tasklet_hi_vec);
 359
 360 void __tasklet_schedule(struct tasklet_struct *t)
 361 {
 362         unsigned long flags;
 363
 364         local_irq_save(flags);
 365         t->next = NULL;
 366         *__get_cpu_var(tasklet_vec).tail = t;
 367         __get_cpu_var(tasklet_vec).tail = &(t->next);
 368         raise_softirq_irqoff(TASKLET_SOFTIRQ);
 369         local_irq_restore(flags);
 370 }
 371
 372 EXPORT_SYMBOL(__tasklet_schedule);
 373
 374 void __tasklet_hi_schedule(struct tasklet_struct *t)
 375 {
 376         unsigned long flags;
 377
 378         local_irq_save(flags);
 379         t->next = NULL;
 380         *__get_cpu_var(tasklet_hi_vec).tail = t;
 381         __get_cpu_var(tasklet_hi_vec).tail = &(t->next);
 382         raise_softirq_irqoff(HI_SOFTIRQ);
 383         local_irq_restore(flags);
 384 }
 385
 386 EXPORT_SYMBOL(__tasklet_hi_schedule);
 387
 388 void __tasklet_hi_schedule_first(struct tasklet_struct *t)
 389 {
 390         BUG_ON(!irqs_disabled());
 391
 392         t->next = __get_cpu_var(tasklet_hi_vec).head;
 393         __get_cpu_var(tasklet_hi_vec).head = t;
 394         __raise_softirq_irqoff(HI_SOFTIRQ);
 395 }
 396
 397 EXPORT_SYMBOL(__tasklet_hi_schedule_first);
 398
 399 static void tasklet_action(struct softirq_action *a)
 400 {
 401         struct tasklet_struct *list;
 402
 403         local_irq_disable();
 404         list = __get_cpu_var(tasklet_vec).head;
 405         __get_cpu_var(tasklet_vec).head = NULL;
 406         __get_cpu_var(tasklet_vec).tail = &__get_cpu_var(tasklet_vec).head;
 407         local_irq_enable();
 408
 409         while (list) {
 410                 struct tasklet_struct *t = list;
 411
 412                 list = list->next;
 413
 414                 if (tasklet_trylock(t)) {
 415                         if (!atomic_read(&t->count)) {
 416                                 if (!test_and_clear_bit(TASKLET_STATE_SCHED, &t->state))
 417                                         BUG();
 418                                 t->func(t->data);
 419                                 tasklet_unlock(t);
 420                                 continue;
 421                         }
 422                         tasklet_unlock(t);
 423                 }
 424
 425                 local_irq_disable();
 426                 t->next = NULL;
 427                 *__get_cpu_var(tasklet_vec).tail = t;
 428                 __get_cpu_var(tasklet_vec).tail = &(t->next);
 429                 __raise_softirq_irqoff(TASKLET_SOFTIRQ);
 430                 local_irq_enable();
 431         }
 432 }
 433
 434 static void tasklet_hi_action(struct softirq_action *a)
 435 {
 436         struct tasklet_struct *list;
 437
 438         local_irq_disable();
 439         list = __get_cpu_var(tasklet_hi_vec).head;
 440         __get_cpu_var(tasklet_hi_vec).head = NULL;
 441         __get_cpu_var(tasklet_hi_vec).tail = &__get_cpu_var(tasklet_hi_vec).head;
 442         local_irq_enable();
 443
 444         while (list) {
 445                 struct tasklet_struct *t = list;
 446
 447                 list = list->next;
 448
 449                 if (tasklet_trylock(t)) {
 450                         if (!atomic_read(&t->count)) {
 451                                 if (!test_and_clear_bit(TASKLET_STATE_SCHED, &t->state))
 452                                         BUG();
 453                                 t->func(t->data);
 454                                 tasklet_unlock(t);
 455                                 continue;
 456                         }
 457                         tasklet_unlock(t);
 458                 }
 459
 460                 local_irq_disable();
 461                 t->next = NULL;
 462                 *__get_cpu_var(tasklet_hi_vec).tail = t;
 463                 __get_cpu_var(tasklet_hi_vec).tail = &(t->next);
 464                 __raise_softirq_irqoff(HI_SOFTIRQ);
 465                 local_irq_enable();
 466         }
 467 }
 468
 469
 470 void tasklet_init(struct tasklet_struct *t,
 471                   void (*func)(unsigned long), unsigned long data)
 472 {
 473         t->next = NULL;
 474         t->state = 0;
 475         atomic_set(&t->count, 0);
 476         t->func = func;
 477         t->data = data;
 478 }
 479
 480 EXPORT_SYMBOL(tasklet_init);
 481
 482 void tasklet_kill(struct tasklet_struct *t)
 483 {
 484         if (in_interrupt())
 485                 printk("Attempt to kill tasklet from interrupt\n");
 486
 487         while (test_and_set_bit(TASKLET_STATE_SCHED, &t->state)) {
 488                 do {
 489                         yield();
 490                 } while (test_bit(TASKLET_STATE_SCHED, &t->state));
 491         }
 492         tasklet_unlock_wait(t);
 493         clear_bit(TASKLET_STATE_SCHED, &t->state);
 494 }
 495
 496 EXPORT_SYMBOL(tasklet_kill);
 497
 498 /*
 499  * tasklet_hrtimer
 500  */
 501
 502 /*
 503  * The trampoline is called when the hrtimer expires. It schedules a tasklet
 504  * to run __tasklet_hrtimer_trampoline() which in turn will call the intended
 505  * hrtimer callback, but from softirq context.
 506  */
 507 static enum hrtimer_restart __hrtimer_tasklet_trampoline(struct hrtimer *timer)
 508 {
 509         struct tasklet_hrtimer *ttimer =
 510                 container_of(timer, struct tasklet_hrtimer, timer);
 511
 512         tasklet_hi_schedule(&ttimer->tasklet);
 513         return HRTIMER_NORESTART;
 514 }
 515
 516 /*
 517  * Helper function which calls the hrtimer callback from
 518  * tasklet/softirq context
 519  */
 520 static void __tasklet_hrtimer_trampoline(unsigned long data)
 521 {
 522         struct tasklet_hrtimer *ttimer = (void *)data;
 523         enum hrtimer_restart restart;
 524
 525         restart = ttimer->function(&ttimer->timer);
 526         if (restart != HRTIMER_NORESTART)
 527                 hrtimer_restart(&ttimer->timer);
 528 }
 529
 530 /**
 531  * tasklet_hrtimer_init - Init a tasklet/hrtimer combo for softirq callbacks
 532  * @ttimer:      tasklet_hrtimer which is initialized
 533  * @function:    hrtimer callback funtion which gets called from softirq context
 534  * @which_clock: clock id (CLOCK_MONOTONIC/CLOCK_REALTIME)
 535  * @mode:        hrtimer mode (HRTIMER_MODE_ABS/HRTIMER_MODE_REL)
 536  */
 537 void tasklet_hrtimer_init(struct tasklet_hrtimer *ttimer,
 538                           enum hrtimer_restart (*function)(struct hrtimer *),
 539                           clockid_t which_clock, enum hrtimer_mode mode)
 540 {
 541         hrtimer_init(&ttimer->timer, which_clock, mode);
 542         ttimer->timer.function = __hrtimer_tasklet_trampoline;
 543         tasklet_init(&ttimer->tasklet, __tasklet_hrtimer_trampoline,
 544                      (unsigned long)ttimer);
 545         ttimer->function = function;
 546 }
 547 EXPORT_SYMBOL_GPL(tasklet_hrtimer_init);
 548
 549 /*
 550  * Remote softirq bits
 551  */
 552
 553 DEFINE_PER_CPU(struct list_head [NR_SOFTIRQS], softirq_work_list);
 554 EXPORT_PER_CPU_SYMBOL(softirq_work_list);
 555
 556 static void __local_trigger(struct call_single_data *cp, int softirq)
 557 {
 558         struct list_head *head = &__get_cpu_var(softirq_work_list[softirq]);
 559
 560         list_add_tail(&cp->list, head);
 561
 562         /* Trigger the softirq only if the list was previously empty.  */
 563         if (head->next == &cp->list)
 564                 raise_softirq_irqoff(softirq);
 565 }
 566
 567 #ifdef CONFIG_USE_GENERIC_SMP_HELPERS
 568 static void remote_softirq_receive(void *data)
 569 {
 570         struct call_single_data *cp = data;
 571         unsigned long flags;
 572         int softirq;
 573
 574         softirq = cp->priv;
 575
 576         local_irq_save(flags);
 577         __local_trigger(cp, softirq);
 578         local_irq_restore(flags);
 579 }
 580
 581 static int __try_remote_softirq(struct call_single_data *cp, int cpu, int softirq)
 582 {
 583         if (cpu_online(cpu)) {
 584                 cp->func = remote_softirq_receive;
 585                 cp->info = cp;
 586                 cp->flags = 0;
 587                 cp->priv = softirq;
 588
 589                 __smp_call_function_single(cpu, cp, 0);
 590                 return 0;
 591         }
 592         return 1;
 593 }
 594 #else /* CONFIG_USE_GENERIC_SMP_HELPERS */
 595 static int __try_remote_softirq(struct call_single_data *cp, int cpu, int softirq)
 596 {
 597         return 1;
 598 }
 599 #endif
 600
 601 /**
 602  * __send_remote_softirq - try to schedule softirq work on a remote cpu
 603  * @cp: private SMP call function data area
 604  * @cpu: the remote cpu
 605  * @this_cpu: the currently executing cpu
 606  * @softirq: the softirq for the work
 607  *
 608  * Attempt to schedule softirq work on a remote cpu.  If this cannot be
 609  * done, the work is instead queued up on the local cpu.
 610  *
 611  * Interrupts must be disabled.
 612  */
 613 void __send_remote_softirq(struct call_single_data *cp, int cpu, int this_cpu, int softirq)
 614 {
 615         if (cpu == this_cpu || __try_remote_softirq(cp, cpu, softirq))
 616                 __local_trigger(cp, softirq);
 617 }
 618 EXPORT_SYMBOL(__send_remote_softirq);
 619
 620 /**
 621  * send_remote_softirq - try to schedule softirq work on a remote cpu
 622  * @cp: private SMP call function data area
 623  * @cpu: the remote cpu
 624  * @softirq: the softirq for the work
 625  *
 626  * Like __send_remote_softirq except that disabling interrupts and
 627  * computing the current cpu is done for the caller.
 628  */
 629 void send_remote_softirq(struct call_single_data *cp, int cpu, int softirq)
 630 {
 631         unsigned long flags;
 632         int this_cpu;
 633
 634         local_irq_save(flags);
 635         this_cpu = smp_processor_id();
 636         __send_remote_softirq(cp, cpu, this_cpu, softirq);
 637         local_irq_restore(flags);
 638 }
 639 EXPORT_SYMBOL(send_remote_softirq);
 640
 641 static int __cpuinit remote_softirq_cpu_notify(struct notifier_block *self,
 642                                                unsigned long action, void *hcpu)
 643 {
 644         /*
 645          * If a CPU goes away, splice its entries to the current CPU
 646          * and trigger a run of the softirq
 647          */
 648         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
 649                 int cpu = (unsigned long) hcpu;
 650                 int i;
 651
 652                 local_irq_disable();
 653                 for (i = 0; i < NR_SOFTIRQS; i++) {
 654                         struct list_head *head = &per_cpu(softirq_work_list[i], cpu);
 655                         struct list_head *local_head;
 656
 657                         if (list_empty(head))
 658                                 continue;
 659
 660                         local_head = &__get_cpu_var(softirq_work_list[i]);
 661                         list_splice_init(head, local_head);
 662                         raise_softirq_irqoff(i);
 663                 }
 664                 local_irq_enable();
 665         }
 666
 667         return NOTIFY_OK;
 668 }
 669
 670 static struct notifier_block __cpuinitdata remote_softirq_cpu_notifier = {
 671         .notifier_call  = remote_softirq_cpu_notify,
 672 };
 673
 674 void __init softirq_init(void)
 675 {
 676         int cpu;
 677
 678         for_each_possible_cpu(cpu) {
 679                 int i;
 680
 681                 per_cpu(tasklet_vec, cpu).tail =
 682                         &per_cpu(tasklet_vec, cpu).head;
 683                 per_cpu(tasklet_hi_vec, cpu).tail =
 684                         &per_cpu(tasklet_hi_vec, cpu).head;
 685                 for (i = 0; i < NR_SOFTIRQS; i++)
 686                         INIT_LIST_HEAD(&per_cpu(softirq_work_list[i], cpu));
 687         }
 688
 689         register_hotcpu_notifier(&remote_softirq_cpu_notifier);
 690
 691         open_softirq(TASKLET_SOFTIRQ, tasklet_action);
 692         open_softirq(HI_SOFTIRQ, tasklet_hi_action);
 693 }
 694
 695 static int run_ksoftirqd(void * __bind_cpu)
 696 {
 697         set_current_state(TASK_INTERRUPTIBLE);
 698
 699         while (!kthread_should_stop()) {
 700                 preempt_disable();
 701                 if (!local_softirq_pending()) {
 702                         preempt_enable_no_resched();
 703                         schedule();
 704                         preempt_disable();
 705                 }
 706
 707                 __set_current_state(TASK_RUNNING);
 708
 709                 while (local_softirq_pending()) {
 710                         /* Preempt disable stops cpu going offline.
 711                            If already offline, we'll be on wrong CPU:
 712                            don't process */
 713                         if (cpu_is_offline((long)__bind_cpu))
 714                                 goto wait_to_die;
 715                         do_softirq();
 716                         preempt_enable_no_resched();
 717                         cond_resched();
 718                         preempt_disable();
 719                         rcu_note_context_switch((long)__bind_cpu);
 720                 }
 721                 preempt_enable();
 722                 set_current_state(TASK_INTERRUPTIBLE);
 723         }
 724         __set_current_state(TASK_RUNNING);
 725         return 0;
 726
 727 wait_to_die:
 728         preempt_enable();
 729         /* Wait for kthread_stop */
 730         set_current_state(TASK_INTERRUPTIBLE);
 731         while (!kthread_should_stop()) {
 732                 schedule();
 733                 set_current_state(TASK_INTERRUPTIBLE);
 734         }
 735         __set_current_state(TASK_RUNNING);
 736         return 0;
 737 }
 738
 739 #ifdef CONFIG_HOTPLUG_CPU
 740 /*
 741  * tasklet_kill_immediate is called to remove a tasklet which can already be
 742  * scheduled for execution on @cpu.
 743  *
 744  * Unlike tasklet_kill, this function removes the tasklet
 745  * _immediately_, even if the tasklet is in TASKLET_STATE_SCHED state.
 746  *
 747  * When this function is called, @cpu must be in the CPU_DEAD state.
 748  */
 749 void tasklet_kill_immediate(struct tasklet_struct *t, unsigned int cpu)
 750 {
 751         struct tasklet_struct **i;
 752
 753         BUG_ON(cpu_online(cpu));
 754         BUG_ON(test_bit(TASKLET_STATE_RUN, &t->state));
 755
 756         if (!test_bit(TASKLET_STATE_SCHED, &t->state))
 757                 return;
 758
 759         /* CPU is dead, so no lock needed. */
 760         for (i = &per_cpu(tasklet_vec, cpu).head; *i; i = &(*i)->next) {
 761                 if (*i == t) {
 762                         *i = t->next;
 763                         /* If this was the tail element, move the tail ptr */
 764                         if (*i == NULL)
 765                                 per_cpu(tasklet_vec, cpu).tail = i;
 766                         return;
 767                 }
 768         }
 769         BUG();
 770 }
 771
 772 static void takeover_tasklets(unsigned int cpu)
 773 {
 774         /* CPU is dead, so no lock needed. */
 775         local_irq_disable();
 776
 777         /* Find end, append list for that CPU. */
 778         if (&per_cpu(tasklet_vec, cpu).head != per_cpu(tasklet_vec, cpu).tail) {
 779                 *(__get_cpu_var(tasklet_vec).tail) = per_cpu(tasklet_vec, cpu).head;
 780                 __get_cpu_var(tasklet_vec).tail = per_cpu(tasklet_vec, cpu).tail;
 781                 per_cpu(tasklet_vec, cpu).head = NULL;
 782                 per_cpu(tasklet_vec, cpu).tail = &per_cpu(tasklet_vec, cpu).head;
 783         }
 784         raise_softirq_irqoff(TASKLET_SOFTIRQ);
 785
 786         if (&per_cpu(tasklet_hi_vec, cpu).head != per_cpu(tasklet_hi_vec, cpu).tail) {
 787                 *__get_cpu_var(tasklet_hi_vec).tail = per_cpu(tasklet_hi_vec, cpu).head;
 788                 __get_cpu_var(tasklet_hi_vec).tail = per_cpu(tasklet_hi_vec, cpu).tail;
 789                 per_cpu(tasklet_hi_vec, cpu).head = NULL;
 790                 per_cpu(tasklet_hi_vec, cpu).tail = &per_cpu(tasklet_hi_vec, cpu).head;
 791         }
 792         raise_softirq_irqoff(HI_SOFTIRQ);
 793
 794         local_irq_enable();
 795 }
 796 #endif /* CONFIG_HOTPLUG_CPU */
 797
 798 static int __cpuinit cpu_callback(struct notifier_block *nfb,
 799                                   unsigned long action,
 800                                   void *hcpu)
 801 {
 802         int hotcpu = (unsigned long)hcpu;
 803         struct task_struct *p;
 804
 805         switch (action) {
 806         case CPU_UP_PREPARE:
 807         case CPU_UP_PREPARE_FROZEN:
 808                 p = kthread_create(run_ksoftirqd, hcpu, "ksoftirqd/%d", hotcpu);
 809                 if (IS_ERR(p)) {
 810                         printk("ksoftirqd for %i failed\n", hotcpu);
 811                         return notifier_from_errno(PTR_ERR(p));
 812                 }
 813                 kthread_bind(p, hotcpu);
 814                 per_cpu(ksoftirqd, hotcpu) = p;
 815                 break;
 816         case CPU_ONLINE:
 817         case CPU_ONLINE_FROZEN:
 818                 wake_up_process(per_cpu(ksoftirqd, hotcpu));
 819                 break;
 820 #ifdef CONFIG_HOTPLUG_CPU
 821         case CPU_UP_CANCELED:
 822         case CPU_UP_CANCELED_FROZEN:
 823                 if (!per_cpu(ksoftirqd, hotcpu))
 824                         break;
 825                 /* Unbind so it can run.  Fall thru. */
 826                 kthread_bind(per_cpu(ksoftirqd, hotcpu),
 827                              cpumask_any(cpu_online_mask));
 828         case CPU_DEAD:
 829         case CPU_DEAD_FROZEN: {
 830                 struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
 831
 832                 p = per_cpu(ksoftirqd, hotcpu);
 833                 per_cpu(ksoftirqd, hotcpu) = NULL;
 834                 sched_setscheduler_nocheck(p, SCHED_FIFO, &param);
 835                 kthread_stop(p);
 836                 takeover_tasklets(hotcpu);
 837                 break;
 838         }
 839 #endif /* CONFIG_HOTPLUG_CPU */
 840         }
 841         return NOTIFY_OK;
 842 }
 843
 844 static struct notifier_block __cpuinitdata cpu_nfb = {
 845         .notifier_call = cpu_callback
 846 };
 847
 848 static __init int spawn_ksoftirqd(void)
 849 {
 850         void *cpu = (void *)(long)smp_processor_id();
 851         int err = cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);
 852
 853         BUG_ON(err == NOTIFY_BAD);
 854         cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);
 855         register_cpu_notifier(&cpu_nfb);
 856         return 0;
 857 }
 858 early_initcall(spawn_ksoftirqd);
 859
 860 #ifdef CONFIG_SMP
 861 /*
 862  * Call a function on all processors
 863  */
 864 int on_each_cpu(void (*func) (void *info), void *info, int wait)
 865 {
 866         int ret = 0;
 867
 868         preempt_disable();
 869         ret = smp_call_function(func, info, wait);
 870         local_irq_disable();
 871         func(info);
 872         local_irq_enable();
 873         preempt_enable();
 874         return ret;
 875 }
 876 EXPORT_SYMBOL(on_each_cpu);
 877 #endif
 878
 879 /*
 880  * [ These __weak aliases are kept in a separate compilation unit, so that
 881  *   GCC does not inline them incorrectly. ]
 882  */
 883
 884 int __init __weak early_irq_init(void)
 885 {
 886         return 0;
 887 }
 888
 889 int __init __weak arch_probe_nr_irqs(void)
 890 {
 891         return 0;
 892 }
 893
 894 int __init __weak arch_early_irq_init(void)
 895 {
 896         return 0;
 897 }
 898
 899 int __weak arch_init_chip_data(struct irq_desc *desc, int node)
 900 {
 901         return 0;
 902 }