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/export.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 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 static void wakeup_softirqd(void)
  71 {
  72         /* Interrupts are disabled: no need to stop preemption */
  73         struct task_struct *tsk = __this_cpu_read(ksoftirqd);
  74
  75         if (tsk && tsk->state != TASK_RUNNING)
  76                 wake_up_process(tsk);
  77 }
  78
  79 /*
  80  * preempt_count and SOFTIRQ_OFFSET usage:
  81  * - preempt_count is changed by SOFTIRQ_OFFSET on entering or leaving
  82  *   softirq processing.
  83  * - preempt_count is changed by SOFTIRQ_DISABLE_OFFSET (= 2 * SOFTIRQ_OFFSET)
  84  *   on local_bh_disable or local_bh_enable.
  85  * This lets us distinguish between whether we are currently processing
  86  * softirq and whether we just have bh disabled.
  87  */
  88
  89 /*
  90  * This one is for softirq.c-internal use,
  91  * where hardirqs are disabled legitimately:
  92  */
  93 #ifdef CONFIG_TRACE_IRQFLAGS
  94 static void __local_bh_disable(unsigned long ip, unsigned int cnt)
  95 {
  96         unsigned long flags;
  97
  98         WARN_ON_ONCE(in_irq());
  99
 100         raw_local_irq_save(flags);
 101         /*
 102          * The preempt tracer hooks into add_preempt_count and will break
 103          * lockdep because it calls back into lockdep after SOFTIRQ_OFFSET
 104          * is set and before current->softirq_enabled is cleared.
 105          * We must manually increment preempt_count here and manually
 106          * call the trace_preempt_off later.
 107          */
 108         preempt_count() += cnt;
 109         /*
 110          * Were softirqs turned off above:
 111          */
 112         if (softirq_count() == cnt)
 113                 trace_softirqs_off(ip);
 114         raw_local_irq_restore(flags);
 115
 116         if (preempt_count() == cnt)
 117                 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
 118 }
 119 #else /* !CONFIG_TRACE_IRQFLAGS */
 120 static inline void __local_bh_disable(unsigned long ip, unsigned int cnt)
 121 {
 122         add_preempt_count(cnt);
 123         barrier();
 124 }
 125 #endif /* CONFIG_TRACE_IRQFLAGS */
 126
 127 void local_bh_disable(void)
 128 {
 129         __local_bh_disable((unsigned long)__builtin_return_address(0),
 130                                 SOFTIRQ_DISABLE_OFFSET);
 131 }
 132
 133 EXPORT_SYMBOL(local_bh_disable);
 134
 135 static void __local_bh_enable(unsigned int cnt)
 136 {
 137         WARN_ON_ONCE(in_irq());
 138         WARN_ON_ONCE(!irqs_disabled());
 139
 140         if (softirq_count() == cnt)
 141                 trace_softirqs_on((unsigned long)__builtin_return_address(0));
 142         sub_preempt_count(cnt);
 143 }
 144
 145 /*
 146  * Special-case - softirqs can safely be enabled in
 147  * cond_resched_softirq(), or by __do_softirq(),
 148  * without processing still-pending softirqs:
 149  */
 150 void _local_bh_enable(void)
 151 {
 152         __local_bh_enable(SOFTIRQ_DISABLE_OFFSET);
 153 }
 154
 155 EXPORT_SYMBOL(_local_bh_enable);
 156
 157 static inline void _local_bh_enable_ip(unsigned long ip)
 158 {
 159         WARN_ON_ONCE(in_irq() || irqs_disabled());
 160 #ifdef CONFIG_TRACE_IRQFLAGS
 161         local_irq_disable();
 162 #endif
 163         /*
 164          * Are softirqs going to be turned on now:
 165          */
 166         if (softirq_count() == SOFTIRQ_DISABLE_OFFSET)
 167                 trace_softirqs_on(ip);
 168         /*
 169          * Keep preemption disabled until we are done with
 170          * softirq processing:
 171          */
 172         sub_preempt_count(SOFTIRQ_DISABLE_OFFSET - 1);
 173
 174         if (unlikely(!in_interrupt() && local_softirq_pending()))
 175                 do_softirq();
 176
 177         dec_preempt_count();
 178 #ifdef CONFIG_TRACE_IRQFLAGS
 179         local_irq_enable();
 180 #endif
 181         preempt_check_resched();
 182 }
 183
 184 void local_bh_enable(void)
 185 {
 186         _local_bh_enable_ip((unsigned long)__builtin_return_address(0));
 187 }
 188 EXPORT_SYMBOL(local_bh_enable);
 189
 190 void local_bh_enable_ip(unsigned long ip)
 191 {
 192         _local_bh_enable_ip(ip);
 193 }
 194 EXPORT_SYMBOL(local_bh_enable_ip);
 195
 196 /*
 197  * We restart softirq processing MAX_SOFTIRQ_RESTART times,
 198  * and we fall back to softirqd after that.
 199  *
 200  * This number has been established via experimentation.
 201  * The two things to balance is latency against fairness -
 202  * we want to handle softirqs as soon as possible, but they
 203  * should not be able to lock up the box.
 204  */
 205 #define MAX_SOFTIRQ_RESTART 10
 206
 207 asmlinkage void __do_softirq(void)
 208 {
 209         struct softirq_action *h;
 210         __u32 pending;
 211         int max_restart = MAX_SOFTIRQ_RESTART;
 212         int cpu;
 213
 214         pending = local_softirq_pending();
 215         account_system_vtime(current);
 216
 217         __local_bh_disable((unsigned long)__builtin_return_address(0),
 218                                 SOFTIRQ_OFFSET);
 219         lockdep_softirq_enter();
 220
 221         cpu = smp_processor_id();
 222 restart:
 223         /* Reset the pending bitmask before enabling irqs */
 224         set_softirq_pending(0);
 225
 226         local_irq_enable();
 227
 228         h = softirq_vec;
 229
 230         do {
 231                 if (pending & 1) {
 232                         unsigned int vec_nr = h - softirq_vec;
 233                         int prev_count = preempt_count();
 234
 235                         kstat_incr_softirqs_this_cpu(vec_nr);
 236
 237                         trace_softirq_entry(vec_nr);
 238                         h->action(h);
 239                         trace_softirq_exit(vec_nr);
 240                         if (unlikely(prev_count != preempt_count())) {
 241                                 printk(KERN_ERR "huh, entered softirq %u %s %p"
 242                                        "with preempt_count %08x,"
 243                                        " exited with %08x?\n", vec_nr,
 244                                        softirq_to_name[vec_nr], h->action,
 245                                        prev_count, preempt_count());
 246                                 preempt_count() = prev_count;
 247                         }
 248
 249                         rcu_bh_qs(cpu);
 250                 }
 251                 h++;
 252                 pending >>= 1;
 253         } while (pending);
 254
 255         local_irq_disable();
 256
 257         pending = local_softirq_pending();
 258         if (pending && --max_restart)
 259                 goto restart;
 260
 261         if (pending)
 262                 wakeup_softirqd();
 263
 264         lockdep_softirq_exit();
 265
 266         account_system_vtime(current);
 267         __local_bh_enable(SOFTIRQ_OFFSET);
 268 }
 269
 270 #ifndef __ARCH_HAS_DO_SOFTIRQ
 271
 272 asmlinkage void do_softirq(void)
 273 {
 274         __u32 pending;
 275         unsigned long flags;
 276
 277         if (in_interrupt())
 278                 return;
 279
 280         local_irq_save(flags);
 281
 282         pending = local_softirq_pending();
 283
 284         if (pending)
 285                 __do_softirq();
 286
 287         local_irq_restore(flags);
 288 }
 289
 290 #endif
 291
 292 /*
 293  * Enter an interrupt context.
 294  */
 295 void irq_enter(void)
 296 {
 297         int cpu = smp_processor_id();
 298
 299         rcu_irq_enter();
 300         if (idle_cpu(cpu) && !in_interrupt()) {
 301                 /*
 302                  * Prevent raise_softirq from needlessly waking up ksoftirqd
 303                  * here, as softirq will be serviced on return from interrupt.
 304                  */
 305                 local_bh_disable();
 306                 tick_check_idle(cpu);
 307                 _local_bh_enable();
 308         }
 309
 310         __irq_enter();
 311 }
 312
 313 #ifdef __ARCH_IRQ_EXIT_IRQS_DISABLED
 314 static inline void invoke_softirq(void)
 315 {
 316         if (!force_irqthreads)
 317                 __do_softirq();
 318         else {
 319                 __local_bh_disable((unsigned long)__builtin_return_address(0),
 320                                 SOFTIRQ_OFFSET);
 321                 wakeup_softirqd();
 322                 __local_bh_enable(SOFTIRQ_OFFSET);
 323         }
 324 }
 325 #else
 326 static inline void invoke_softirq(void)
 327 {
 328         if (!force_irqthreads)
 329                 do_softirq();
 330         else {
 331                 __local_bh_disable((unsigned long)__builtin_return_address(0),
 332                                 SOFTIRQ_OFFSET);
 333                 wakeup_softirqd();
 334                 __local_bh_enable(SOFTIRQ_OFFSET);
 335         }
 336 }
 337 #endif
 338
 339 /*
 340  * Exit an interrupt context. Process softirqs if needed and possible:
 341  */
 342 void irq_exit(void)
 343 {
 344         account_system_vtime(current);
 345         trace_hardirq_exit();
 346         sub_preempt_count(IRQ_EXIT_OFFSET);
 347         if (!in_interrupt() && local_softirq_pending())
 348                 invoke_softirq();
 349
 350         rcu_irq_exit();
 351 #ifdef CONFIG_NO_HZ
 352         /* Make sure that timer wheel updates are propagated */
 353         if (idle_cpu(smp_processor_id()) && !in_interrupt() && !need_resched())
 354                 tick_nohz_stop_sched_tick(0);
 355 #endif
 356         preempt_enable_no_resched();
 357 }
 358
 359 /*
 360  * This function must run with irqs disabled!
 361  */
 362 inline void raise_softirq_irqoff(unsigned int nr)
 363 {
 364         __raise_softirq_irqoff(nr);
 365
 366         /*
 367          * If we're in an interrupt or softirq, we're done
 368          * (this also catches softirq-disabled code). We will
 369          * actually run the softirq once we return from
 370          * the irq or softirq.
 371          *
 372          * Otherwise we wake up ksoftirqd to make sure we
 373          * schedule the softirq soon.
 374          */
 375         if (!in_interrupt())
 376                 wakeup_softirqd();
 377 }
 378
 379 void raise_softirq(unsigned int nr)
 380 {
 381         unsigned long flags;
 382
 383         local_irq_save(flags);
 384         raise_softirq_irqoff(nr);
 385         local_irq_restore(flags);
 386 }
 387
 388 void open_softirq(int nr, void (*action)(struct softirq_action *))
 389 {
 390         softirq_vec[nr].action = action;
 391 }
 392
 393 /*
 394  * Tasklets
 395  */
 396 struct tasklet_head
 397 {
 398         struct tasklet_struct *head;
 399         struct tasklet_struct **tail;
 400 };
 401
 402 static DEFINE_PER_CPU(struct tasklet_head, tasklet_vec);
 403 static DEFINE_PER_CPU(struct tasklet_head, tasklet_hi_vec);
 404
 405 void __tasklet_schedule(struct tasklet_struct *t)
 406 {
 407         unsigned long flags;
 408
 409         local_irq_save(flags);
 410         t->next = NULL;
 411         *__this_cpu_read(tasklet_vec.tail) = t;
 412         __this_cpu_write(tasklet_vec.tail, &(t->next));
 413         raise_softirq_irqoff(TASKLET_SOFTIRQ);
 414         local_irq_restore(flags);
 415 }
 416
 417 EXPORT_SYMBOL(__tasklet_schedule);
 418
 419 void __tasklet_hi_schedule(struct tasklet_struct *t)
 420 {
 421         unsigned long flags;
 422
 423         local_irq_save(flags);
 424         t->next = NULL;
 425         *__this_cpu_read(tasklet_hi_vec.tail) = t;
 426         __this_cpu_write(tasklet_hi_vec.tail,  &(t->next));
 427         raise_softirq_irqoff(HI_SOFTIRQ);
 428         local_irq_restore(flags);
 429 }
 430
 431 EXPORT_SYMBOL(__tasklet_hi_schedule);
 432
 433 void __tasklet_hi_schedule_first(struct tasklet_struct *t)
 434 {
 435         BUG_ON(!irqs_disabled());
 436
 437         t->next = __this_cpu_read(tasklet_hi_vec.head);
 438         __this_cpu_write(tasklet_hi_vec.head, t);
 439         __raise_softirq_irqoff(HI_SOFTIRQ);
 440 }
 441
 442 EXPORT_SYMBOL(__tasklet_hi_schedule_first);
 443
 444 static void tasklet_action(struct softirq_action *a)
 445 {
 446         struct tasklet_struct *list;
 447
 448         local_irq_disable();
 449         list = __this_cpu_read(tasklet_vec.head);
 450         __this_cpu_write(tasklet_vec.head, NULL);
 451         __this_cpu_write(tasklet_vec.tail, &__get_cpu_var(tasklet_vec).head);
 452         local_irq_enable();
 453
 454         while (list) {
 455                 struct tasklet_struct *t = list;
 456
 457                 list = list->next;
 458
 459                 if (tasklet_trylock(t)) {
 460                         if (!atomic_read(&t->count)) {
 461                                 if (!test_and_clear_bit(TASKLET_STATE_SCHED, &t->state))
 462                                         BUG();
 463                                 t->func(t->data);
 464                                 tasklet_unlock(t);
 465                                 continue;
 466                         }
 467                         tasklet_unlock(t);
 468                 }
 469
 470                 local_irq_disable();
 471                 t->next = NULL;
 472                 *__this_cpu_read(tasklet_vec.tail) = t;
 473                 __this_cpu_write(tasklet_vec.tail, &(t->next));
 474                 __raise_softirq_irqoff(TASKLET_SOFTIRQ);
 475                 local_irq_enable();
 476         }
 477 }
 478
 479 static void tasklet_hi_action(struct softirq_action *a)
 480 {
 481         struct tasklet_struct *list;
 482
 483         local_irq_disable();
 484         list = __this_cpu_read(tasklet_hi_vec.head);
 485         __this_cpu_write(tasklet_hi_vec.head, NULL);
 486         __this_cpu_write(tasklet_hi_vec.tail, &__get_cpu_var(tasklet_hi_vec).head);
 487         local_irq_enable();
 488
 489         while (list) {
 490                 struct tasklet_struct *t = list;
 491
 492                 list = list->next;
 493
 494                 if (tasklet_trylock(t)) {
 495                         if (!atomic_read(&t->count)) {
 496                                 if (!test_and_clear_bit(TASKLET_STATE_SCHED, &t->state))
 497                                         BUG();
 498                                 t->func(t->data);
 499                                 tasklet_unlock(t);
 500                                 continue;
 501                         }
 502                         tasklet_unlock(t);
 503                 }
 504
 505                 local_irq_disable();
 506                 t->next = NULL;
 507                 *__this_cpu_read(tasklet_hi_vec.tail) = t;
 508                 __this_cpu_write(tasklet_hi_vec.tail, &(t->next));
 509                 __raise_softirq_irqoff(HI_SOFTIRQ);
 510                 local_irq_enable();
 511         }
 512 }
 513
 514
 515 void tasklet_init(struct tasklet_struct *t,
 516                   void (*func)(unsigned long), unsigned long data)
 517 {
 518         t->next = NULL;
 519         t->state = 0;
 520         atomic_set(&t->count, 0);
 521         t->func = func;
 522         t->data = data;
 523 }
 524
 525 EXPORT_SYMBOL(tasklet_init);
 526
 527 void tasklet_kill(struct tasklet_struct *t)
 528 {
 529         if (in_interrupt())
 530                 printk("Attempt to kill tasklet from interrupt\n");
 531
 532         while (test_and_set_bit(TASKLET_STATE_SCHED, &t->state)) {
 533                 do {
 534                         yield();
 535                 } while (test_bit(TASKLET_STATE_SCHED, &t->state));
 536         }
 537         tasklet_unlock_wait(t);
 538         clear_bit(TASKLET_STATE_SCHED, &t->state);
 539 }
 540
 541 EXPORT_SYMBOL(tasklet_kill);
 542
 543 /*
 544  * tasklet_hrtimer
 545  */
 546
 547 /*
 548  * The trampoline is called when the hrtimer expires. It schedules a tasklet
 549  * to run __tasklet_hrtimer_trampoline() which in turn will call the intended
 550  * hrtimer callback, but from softirq context.
 551  */
 552 static enum hrtimer_restart __hrtimer_tasklet_trampoline(struct hrtimer *timer)
 553 {
 554         struct tasklet_hrtimer *ttimer =
 555                 container_of(timer, struct tasklet_hrtimer, timer);
 556
 557         tasklet_hi_schedule(&ttimer->tasklet);
 558         return HRTIMER_NORESTART;
 559 }
 560
 561 /*
 562  * Helper function which calls the hrtimer callback from
 563  * tasklet/softirq context
 564  */
 565 static void __tasklet_hrtimer_trampoline(unsigned long data)
 566 {
 567         struct tasklet_hrtimer *ttimer = (void *)data;
 568         enum hrtimer_restart restart;
 569
 570         restart = ttimer->function(&ttimer->timer);
 571         if (restart != HRTIMER_NORESTART)
 572                 hrtimer_restart(&ttimer->timer);
 573 }
 574
 575 /**
 576  * tasklet_hrtimer_init - Init a tasklet/hrtimer combo for softirq callbacks
 577  * @ttimer:      tasklet_hrtimer which is initialized
 578  * @function:    hrtimer callback function which gets called from softirq context
 579  * @which_clock: clock id (CLOCK_MONOTONIC/CLOCK_REALTIME)
 580  * @mode:        hrtimer mode (HRTIMER_MODE_ABS/HRTIMER_MODE_REL)
 581  */
 582 void tasklet_hrtimer_init(struct tasklet_hrtimer *ttimer,
 583                           enum hrtimer_restart (*function)(struct hrtimer *),
 584                           clockid_t which_clock, enum hrtimer_mode mode)
 585 {
 586         hrtimer_init(&ttimer->timer, which_clock, mode);
 587         ttimer->timer.function = __hrtimer_tasklet_trampoline;
 588         tasklet_init(&ttimer->tasklet, __tasklet_hrtimer_trampoline,
 589                      (unsigned long)ttimer);
 590         ttimer->function = function;
 591 }
 592 EXPORT_SYMBOL_GPL(tasklet_hrtimer_init);
 593
 594 /*
 595  * Remote softirq bits
 596  */
 597
 598 DEFINE_PER_CPU(struct list_head [NR_SOFTIRQS], softirq_work_list);
 599 EXPORT_PER_CPU_SYMBOL(softirq_work_list);
 600
 601 static void __local_trigger(struct call_single_data *cp, int softirq)
 602 {
 603         struct list_head *head = &__get_cpu_var(softirq_work_list[softirq]);
 604
 605         list_add_tail(&cp->list, head);
 606
 607         /* Trigger the softirq only if the list was previously empty.  */
 608         if (head->next == &cp->list)
 609                 raise_softirq_irqoff(softirq);
 610 }
 611
 612 #ifdef CONFIG_USE_GENERIC_SMP_HELPERS
 613 static void remote_softirq_receive(void *data)
 614 {
 615         struct call_single_data *cp = data;
 616         unsigned long flags;
 617         int softirq;
 618
 619         softirq = cp->priv;
 620
 621         local_irq_save(flags);
 622         __local_trigger(cp, softirq);
 623         local_irq_restore(flags);
 624 }
 625
 626 static int __try_remote_softirq(struct call_single_data *cp, int cpu, int softirq)
 627 {
 628         if (cpu_online(cpu)) {
 629                 cp->func = remote_softirq_receive;
 630                 cp->info = cp;
 631                 cp->flags = 0;
 632                 cp->priv = softirq;
 633
 634                 __smp_call_function_single(cpu, cp, 0);
 635                 return 0;
 636         }
 637         return 1;
 638 }
 639 #else /* CONFIG_USE_GENERIC_SMP_HELPERS */
 640 static int __try_remote_softirq(struct call_single_data *cp, int cpu, int softirq)
 641 {
 642         return 1;
 643 }
 644 #endif
 645
 646 /**
 647  * __send_remote_softirq - try to schedule softirq work on a remote cpu
 648  * @cp: private SMP call function data area
 649  * @cpu: the remote cpu
 650  * @this_cpu: the currently executing cpu
 651  * @softirq: the softirq for the work
 652  *
 653  * Attempt to schedule softirq work on a remote cpu.  If this cannot be
 654  * done, the work is instead queued up on the local cpu.
 655  *
 656  * Interrupts must be disabled.
 657  */
 658 void __send_remote_softirq(struct call_single_data *cp, int cpu, int this_cpu, int softirq)
 659 {
 660         if (cpu == this_cpu || __try_remote_softirq(cp, cpu, softirq))
 661                 __local_trigger(cp, softirq);
 662 }
 663 EXPORT_SYMBOL(__send_remote_softirq);
 664
 665 /**
 666  * send_remote_softirq - try to schedule softirq work on a remote cpu
 667  * @cp: private SMP call function data area
 668  * @cpu: the remote cpu
 669  * @softirq: the softirq for the work
 670  *
 671  * Like __send_remote_softirq except that disabling interrupts and
 672  * computing the current cpu is done for the caller.
 673  */
 674 void send_remote_softirq(struct call_single_data *cp, int cpu, int softirq)
 675 {
 676         unsigned long flags;
 677         int this_cpu;
 678
 679         local_irq_save(flags);
 680         this_cpu = smp_processor_id();
 681         __send_remote_softirq(cp, cpu, this_cpu, softirq);
 682         local_irq_restore(flags);
 683 }
 684 EXPORT_SYMBOL(send_remote_softirq);
 685
 686 static int __cpuinit remote_softirq_cpu_notify(struct notifier_block *self,
 687                                                unsigned long action, void *hcpu)
 688 {
 689         /*
 690          * If a CPU goes away, splice its entries to the current CPU
 691          * and trigger a run of the softirq
 692          */
 693         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
 694                 int cpu = (unsigned long) hcpu;
 695                 int i;
 696
 697                 local_irq_disable();
 698                 for (i = 0; i < NR_SOFTIRQS; i++) {
 699                         struct list_head *head = &per_cpu(softirq_work_list[i], cpu);
 700                         struct list_head *local_head;
 701
 702                         if (list_empty(head))
 703                                 continue;
 704
 705                         local_head = &__get_cpu_var(softirq_work_list[i]);
 706                         list_splice_init(head, local_head);
 707                         raise_softirq_irqoff(i);
 708                 }
 709                 local_irq_enable();
 710         }
 711
 712         return NOTIFY_OK;
 713 }
 714
 715 static struct notifier_block __cpuinitdata remote_softirq_cpu_notifier = {
 716         .notifier_call  = remote_softirq_cpu_notify,
 717 };
 718
 719 void __init softirq_init(void)
 720 {
 721         int cpu;
 722
 723         for_each_possible_cpu(cpu) {
 724                 int i;
 725
 726                 per_cpu(tasklet_vec, cpu).tail =
 727                         &per_cpu(tasklet_vec, cpu).head;
 728                 per_cpu(tasklet_hi_vec, cpu).tail =
 729                         &per_cpu(tasklet_hi_vec, cpu).head;
 730                 for (i = 0; i < NR_SOFTIRQS; i++)
 731                         INIT_LIST_HEAD(&per_cpu(softirq_work_list[i], cpu));
 732         }
 733
 734         register_hotcpu_notifier(&remote_softirq_cpu_notifier);
 735
 736         open_softirq(TASKLET_SOFTIRQ, tasklet_action);
 737         open_softirq(HI_SOFTIRQ, tasklet_hi_action);
 738 }
 739
 740 static int run_ksoftirqd(void * __bind_cpu)
 741 {
 742         set_current_state(TASK_INTERRUPTIBLE);
 743
 744         while (!kthread_should_stop()) {
 745                 preempt_disable();
 746                 if (!local_softirq_pending()) {
 747                         preempt_enable_no_resched();
 748                         schedule();
 749                         preempt_disable();
 750                 }
 751
 752                 __set_current_state(TASK_RUNNING);
 753
 754                 while (local_softirq_pending()) {
 755                         /* Preempt disable stops cpu going offline.
 756                            If already offline, we'll be on wrong CPU:
 757                            don't process */
 758                         if (cpu_is_offline((long)__bind_cpu))
 759                                 goto wait_to_die;
 760                         local_irq_disable();
 761                         if (local_softirq_pending())
 762                                 __do_softirq();
 763                         local_irq_enable();
 764                         preempt_enable_no_resched();
 765                         cond_resched();
 766                         preempt_disable();
 767                         rcu_note_context_switch((long)__bind_cpu);
 768                 }
 769                 preempt_enable();
 770                 set_current_state(TASK_INTERRUPTIBLE);
 771         }
 772         __set_current_state(TASK_RUNNING);
 773         return 0;
 774
 775 wait_to_die:
 776         preempt_enable();
 777         /* Wait for kthread_stop */
 778         set_current_state(TASK_INTERRUPTIBLE);
 779         while (!kthread_should_stop()) {
 780                 schedule();
 781                 set_current_state(TASK_INTERRUPTIBLE);
 782         }
 783         __set_current_state(TASK_RUNNING);
 784         return 0;
 785 }
 786
 787 #ifdef CONFIG_HOTPLUG_CPU
 788 /*
 789  * tasklet_kill_immediate is called to remove a tasklet which can already be
 790  * scheduled for execution on @cpu.
 791  *
 792  * Unlike tasklet_kill, this function removes the tasklet
 793  * _immediately_, even if the tasklet is in TASKLET_STATE_SCHED state.
 794  *
 795  * When this function is called, @cpu must be in the CPU_DEAD state.
 796  */
 797 void tasklet_kill_immediate(struct tasklet_struct *t, unsigned int cpu)
 798 {
 799         struct tasklet_struct **i;
 800
 801         BUG_ON(cpu_online(cpu));
 802         BUG_ON(test_bit(TASKLET_STATE_RUN, &t->state));
 803
 804         if (!test_bit(TASKLET_STATE_SCHED, &t->state))
 805                 return;
 806
 807         /* CPU is dead, so no lock needed. */
 808         for (i = &per_cpu(tasklet_vec, cpu).head; *i; i = &(*i)->next) {
 809                 if (*i == t) {
 810                         *i = t->next;
 811                         /* If this was the tail element, move the tail ptr */
 812                         if (*i == NULL)
 813                                 per_cpu(tasklet_vec, cpu).tail = i;
 814                         return;
 815                 }
 816         }
 817         BUG();
 818 }
 819
 820 static void takeover_tasklets(unsigned int cpu)
 821 {
 822         /* CPU is dead, so no lock needed. */
 823         local_irq_disable();
 824
 825         /* Find end, append list for that CPU. */
 826         if (&per_cpu(tasklet_vec, cpu).head != per_cpu(tasklet_vec, cpu).tail) {
 827                 *__this_cpu_read(tasklet_vec.tail) = per_cpu(tasklet_vec, cpu).head;
 828                 this_cpu_write(tasklet_vec.tail, per_cpu(tasklet_vec, cpu).tail);
 829                 per_cpu(tasklet_vec, cpu).head = NULL;
 830                 per_cpu(tasklet_vec, cpu).tail = &per_cpu(tasklet_vec, cpu).head;
 831         }
 832         raise_softirq_irqoff(TASKLET_SOFTIRQ);
 833
 834         if (&per_cpu(tasklet_hi_vec, cpu).head != per_cpu(tasklet_hi_vec, cpu).tail) {
 835                 *__this_cpu_read(tasklet_hi_vec.tail) = per_cpu(tasklet_hi_vec, cpu).head;
 836                 __this_cpu_write(tasklet_hi_vec.tail, per_cpu(tasklet_hi_vec, cpu).tail);
 837                 per_cpu(tasklet_hi_vec, cpu).head = NULL;
 838                 per_cpu(tasklet_hi_vec, cpu).tail = &per_cpu(tasklet_hi_vec, cpu).head;
 839         }
 840         raise_softirq_irqoff(HI_SOFTIRQ);
 841
 842         local_irq_enable();
 843 }
 844 #endif /* CONFIG_HOTPLUG_CPU */
 845
 846 static int __cpuinit cpu_callback(struct notifier_block *nfb,
 847                                   unsigned long action,
 848                                   void *hcpu)
 849 {
 850         int hotcpu = (unsigned long)hcpu;
 851         struct task_struct *p;
 852
 853         switch (action) {
 854         case CPU_UP_PREPARE:
 855         case CPU_UP_PREPARE_FROZEN:
 856                 p = kthread_create_on_node(run_ksoftirqd,
 857                                            hcpu,
 858                                            cpu_to_node(hotcpu),
 859                                            "ksoftirqd/%d", hotcpu);
 860                 if (IS_ERR(p)) {
 861                         printk("ksoftirqd for %i failed\n", hotcpu);
 862                         return notifier_from_errno(PTR_ERR(p));
 863                 }
 864                 kthread_bind(p, hotcpu);
 865                 per_cpu(ksoftirqd, hotcpu) = p;
 866                 break;
 867         case CPU_ONLINE:
 868         case CPU_ONLINE_FROZEN:
 869                 wake_up_process(per_cpu(ksoftirqd, hotcpu));
 870                 break;
 871 #ifdef CONFIG_HOTPLUG_CPU
 872         case CPU_UP_CANCELED:
 873         case CPU_UP_CANCELED_FROZEN:
 874                 if (!per_cpu(ksoftirqd, hotcpu))
 875                         break;
 876                 /* Unbind so it can run.  Fall thru. */
 877                 kthread_bind(per_cpu(ksoftirqd, hotcpu),
 878                              cpumask_any(cpu_online_mask));
 879         case CPU_DEAD:
 880         case CPU_DEAD_FROZEN: {
 881                 static const struct sched_param param = {
 882                         .sched_priority = MAX_RT_PRIO-1
 883                 };
 884
 885                 p = per_cpu(ksoftirqd, hotcpu);
 886                 per_cpu(ksoftirqd, hotcpu) = NULL;
 887                 sched_setscheduler_nocheck(p, SCHED_FIFO, &param);
 888                 kthread_stop(p);
 889                 takeover_tasklets(hotcpu);
 890                 break;
 891         }
 892 #endif /* CONFIG_HOTPLUG_CPU */
 893         }
 894         return NOTIFY_OK;
 895 }
 896
 897 static struct notifier_block __cpuinitdata cpu_nfb = {
 898         .notifier_call = cpu_callback
 899 };
 900
 901 static __init int spawn_ksoftirqd(void)
 902 {
 903         void *cpu = (void *)(long)smp_processor_id();
 904         int err = cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);
 905
 906         BUG_ON(err != NOTIFY_OK);
 907         cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);
 908         register_cpu_notifier(&cpu_nfb);
 909         return 0;
 910 }
 911 early_initcall(spawn_ksoftirqd);
 912
 913 /*
 914  * [ These __weak aliases are kept in a separate compilation unit, so that
 915  *   GCC does not inline them incorrectly. ]
 916  */
 917
 918 int __init __weak early_irq_init(void)
 919 {
 920         return 0;
 921 }
 922
 923 #ifdef CONFIG_GENERIC_HARDIRQS
 924 int __init __weak arch_probe_nr_irqs(void)
 925 {
 926         return NR_IRQS_LEGACY;
 927 }
 928
 929 int __init __weak arch_early_irq_init(void)
 930 {
 931         return 0;
 932 }
 933 #endif