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