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
  11 #include <linux/export.h>
  12 #include <linux/kernel_stat.h>
  13 #include <linux/interrupt.h>
  14 #include <linux/init.h>
  15 #include <linux/mm.h>
  16 #include <linux/notifier.h>
  17 #include <linux/percpu.h>
  18 #include <linux/cpu.h>
  19 #include <linux/freezer.h>
  20 #include <linux/kthread.h>
  21 #include <linux/rcupdate.h>
  22 #include <linux/ftrace.h>
  23 #include <linux/smp.h>
  24 #include <linux/smpboot.h>
  25 #include <linux/tick.h>
  26
  27 #define CREATE_TRACE_POINTS
  28 #include <trace/events/irq.h>
  29
  30 /*
  31    - No shared variables, all the data are CPU local.
  32    - If a softirq needs serialization, let it serialize itself
  33      by its own spinlocks.
  34    - Even if softirq is serialized, only local cpu is marked for
  35      execution. Hence, we get something sort of weak cpu binding.
  36      Though it is still not clear, will it result in better locality
  37      or will not.
  38
  39    Examples:
  40    - NET RX softirq. It is multithreaded and does not require
  41      any global serialization.
  42    - NET TX softirq. It kicks software netdevice queues, hence
  43      it is logically serialized per device, but this serialization
  44      is invisible to common code.
  45    - Tasklets: serialized wrt itself.
  46  */
  47
  48 #ifndef __ARCH_IRQ_STAT
  49 irq_cpustat_t irq_stat[NR_CPUS] ____cacheline_aligned;
  50 EXPORT_SYMBOL(irq_stat);
  51 #endif
  52
  53 static struct softirq_action softirq_vec[NR_SOFTIRQS] __cacheline_aligned_in_smp;
  54
  55 DEFINE_PER_CPU(struct task_struct *, ksoftirqd);
  56
  57 char *softirq_to_name[NR_SOFTIRQS] = {
  58         "HI", "TIMER", "NET_TX", "NET_RX", "BLOCK", "BLOCK_IOPOLL",
  59         "TASKLET", "SCHED", "HRTIMER", "RCU"
  60 };
  61
  62 /*
  63  * we cannot loop indefinitely here to avoid userspace starvation,
  64  * but we also don't want to introduce a worst case 1/HZ latency
  65  * to the pending events, so lets the scheduler to balance
  66  * the softirq load for us.
  67  */
  68 static void wakeup_softirqd(void)
  69 {
  70         /* Interrupts are disabled: no need to stop preemption */
  71         struct task_struct *tsk = __this_cpu_read(ksoftirqd);
  72
  73         if (tsk && tsk->state != TASK_RUNNING)
  74                 wake_up_process(tsk);
  75 }
  76
  77 /*
  78  * preempt_count and SOFTIRQ_OFFSET usage:
  79  * - preempt_count is changed by SOFTIRQ_OFFSET on entering or leaving
  80  *   softirq processing.
  81  * - preempt_count is changed by SOFTIRQ_DISABLE_OFFSET (= 2 * SOFTIRQ_OFFSET)
  82  *   on local_bh_disable or local_bh_enable.
  83  * This lets us distinguish between whether we are currently processing
  84  * softirq and whether we just have bh disabled.
  85  */
  86
  87 /*
  88  * This one is for softirq.c-internal use,
  89  * where hardirqs are disabled legitimately:
  90  */
  91 #ifdef CONFIG_TRACE_IRQFLAGS
  92 static void __local_bh_disable(unsigned long ip, unsigned int cnt)
  93 {
  94         unsigned long flags;
  95
  96         WARN_ON_ONCE(in_irq());
  97
  98         raw_local_irq_save(flags);
  99         /*
 100          * The preempt tracer hooks into preempt_count_add and will break
 101          * lockdep because it calls back into lockdep after SOFTIRQ_OFFSET
 102          * is set and before current->softirq_enabled is cleared.
 103          * We must manually increment preempt_count here and manually
 104          * call the trace_preempt_off later.
 105          */
 106         __preempt_count_add(cnt);
 107         /*
 108          * Were softirqs turned off above:
 109          */
 110         if (softirq_count() == cnt)
 111                 trace_softirqs_off(ip);
 112         raw_local_irq_restore(flags);
 113
 114         if (preempt_count() == cnt)
 115                 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
 116 }
 117 #else /* !CONFIG_TRACE_IRQFLAGS */
 118 static inline void __local_bh_disable(unsigned long ip, unsigned int cnt)
 119 {
 120         preempt_count_add(cnt);
 121         barrier();
 122 }
 123 #endif /* CONFIG_TRACE_IRQFLAGS */
 124
 125 void local_bh_disable(void)
 126 {
 127         __local_bh_disable(_RET_IP_, SOFTIRQ_DISABLE_OFFSET);
 128 }
 129
 130 EXPORT_SYMBOL(local_bh_disable);
 131
 132 static void __local_bh_enable(unsigned int cnt)
 133 {
 134         WARN_ON_ONCE(!irqs_disabled());
 135
 136         if (softirq_count() == cnt)
 137                 trace_softirqs_on(_RET_IP_);
 138         preempt_count_sub(cnt);
 139 }
 140
 141 /*
 142  * Special-case - softirqs can safely be enabled in
 143  * cond_resched_softirq(), or by __do_softirq(),
 144  * without processing still-pending softirqs:
 145  */
 146 void _local_bh_enable(void)
 147 {
 148         WARN_ON_ONCE(in_irq());
 149         __local_bh_enable(SOFTIRQ_DISABLE_OFFSET);
 150 }
 151
 152 EXPORT_SYMBOL(_local_bh_enable);
 153
 154 static inline void _local_bh_enable_ip(unsigned long ip)
 155 {
 156         WARN_ON_ONCE(in_irq() || irqs_disabled());
 157 #ifdef CONFIG_TRACE_IRQFLAGS
 158         local_irq_disable();
 159 #endif
 160         /*
 161          * Are softirqs going to be turned on now:
 162          */
 163         if (softirq_count() == SOFTIRQ_DISABLE_OFFSET)
 164                 trace_softirqs_on(ip);
 165         /*
 166          * Keep preemption disabled until we are done with
 167          * softirq processing:
 168          */
 169         preempt_count_sub(SOFTIRQ_DISABLE_OFFSET - 1);
 170
 171         if (unlikely(!in_interrupt() && local_softirq_pending())) {
 172                 /*
 173                  * Run softirq if any pending. And do it in its own stack
 174                  * as we may be calling this deep in a task call stack already.
 175                  */
 176                 do_softirq();
 177         }
 178
 179         preempt_count_dec();
 180 #ifdef CONFIG_TRACE_IRQFLAGS
 181         local_irq_enable();
 182 #endif
 183         preempt_check_resched();
 184 }
 185
 186 void local_bh_enable(void)
 187 {
 188         _local_bh_enable_ip(_RET_IP_);
 189 }
 190 EXPORT_SYMBOL(local_bh_enable);
 191
 192 void local_bh_enable_ip(unsigned long ip)
 193 {
 194         _local_bh_enable_ip(ip);
 195 }
 196 EXPORT_SYMBOL(local_bh_enable_ip);
 197
 198 /*
 199  * We restart softirq processing for at most MAX_SOFTIRQ_RESTART times,
 200  * but break the loop if need_resched() is set or after 2 ms.
 201  * The MAX_SOFTIRQ_TIME provides a nice upper bound in most cases, but in
 202  * certain cases, such as stop_machine(), jiffies may cease to
 203  * increment and so we need the MAX_SOFTIRQ_RESTART limit as
 204  * well to make sure we eventually return from this method.
 205  *
 206  * These limits have been established via experimentation.
 207  * The two things to balance is latency against fairness -
 208  * we want to handle softirqs as soon as possible, but they
 209  * should not be able to lock up the box.
 210  */
 211 #define MAX_SOFTIRQ_TIME  msecs_to_jiffies(2)
 212 #define MAX_SOFTIRQ_RESTART 10
 213
 214 #ifdef CONFIG_TRACE_IRQFLAGS
 215 /*
 216  * When we run softirqs from irq_exit() and thus on the hardirq stack we need
 217  * to keep the lockdep irq context tracking as tight as possible in order to
 218  * not miss-qualify lock contexts and miss possible deadlocks.
 219  */
 220
 221 static inline bool lockdep_softirq_start(void)
 222 {
 223         bool in_hardirq = false;
 224
 225         if (trace_hardirq_context(current)) {
 226                 in_hardirq = true;
 227                 trace_hardirq_exit();
 228         }
 229
 230         lockdep_softirq_enter();
 231
 232         return in_hardirq;
 233 }
 234
 235 static inline void lockdep_softirq_end(bool in_hardirq)
 236 {
 237         lockdep_softirq_exit();
 238
 239         if (in_hardirq)
 240                 trace_hardirq_enter();
 241 }
 242 #else
 243 static inline bool lockdep_softirq_start(void) { return false; }
 244 static inline void lockdep_softirq_end(bool in_hardirq) { }
 245 #endif
 246
 247 asmlinkage void __do_softirq(void)
 248 {
 249         unsigned long end = jiffies + MAX_SOFTIRQ_TIME;
 250         unsigned long old_flags = current->flags;
 251         int max_restart = MAX_SOFTIRQ_RESTART;
 252         struct softirq_action *h;
 253         bool in_hardirq;
 254         __u32 pending;
 255         int cpu;
 256
 257         /*
 258          * Mask out PF_MEMALLOC s current task context is borrowed for the
 259          * softirq. A softirq handled such as network RX might set PF_MEMALLOC
 260          * again if the socket is related to swap
 261          */
 262         current->flags &= ~PF_MEMALLOC;
 263
 264         pending = local_softirq_pending();
 265         account_irq_enter_time(current);
 266
 267         __local_bh_disable(_RET_IP_, SOFTIRQ_OFFSET);
 268         in_hardirq = lockdep_softirq_start();
 269
 270         cpu = smp_processor_id();
 271 restart:
 272         /* Reset the pending bitmask before enabling irqs */
 273         set_softirq_pending(0);
 274
 275         local_irq_enable();
 276
 277         h = softirq_vec;
 278
 279         do {
 280                 if (pending & 1) {
 281                         unsigned int vec_nr = h - softirq_vec;
 282                         int prev_count = preempt_count();
 283
 284                         kstat_incr_softirqs_this_cpu(vec_nr);
 285
 286                         trace_softirq_entry(vec_nr);
 287                         h->action(h);
 288                         trace_softirq_exit(vec_nr);
 289                         if (unlikely(prev_count != preempt_count())) {
 290                                 printk(KERN_ERR "huh, entered softirq %u %s %p"
 291                                        "with preempt_count %08x,"
 292                                        " exited with %08x?\n", vec_nr,
 293                                        softirq_to_name[vec_nr], h->action,
 294                                        prev_count, preempt_count());
 295                                 preempt_count_set(prev_count);
 296                         }
 297
 298                         rcu_bh_qs(cpu);
 299                 }
 300                 h++;
 301                 pending >>= 1;
 302         } while (pending);
 303
 304         local_irq_disable();
 305
 306         pending = local_softirq_pending();
 307         if (pending) {
 308                 if (time_before(jiffies, end) && !need_resched() &&
 309                     --max_restart)
 310                         goto restart;
 311
 312                 wakeup_softirqd();
 313         }
 314
 315         lockdep_softirq_end(in_hardirq);
 316         account_irq_exit_time(current);
 317         __local_bh_enable(SOFTIRQ_OFFSET);
 318         WARN_ON_ONCE(in_interrupt());
 319         tsk_restore_flags(current, old_flags, PF_MEMALLOC);
 320 }
 321
 322 asmlinkage void do_softirq(void)
 323 {
 324         __u32 pending;
 325         unsigned long flags;
 326
 327         if (in_interrupt())
 328                 return;
 329
 330         local_irq_save(flags);
 331
 332         pending = local_softirq_pending();
 333
 334         if (pending)
 335                 do_softirq_own_stack();
 336
 337         local_irq_restore(flags);
 338 }
 339
 340 /*
 341  * Enter an interrupt context.
 342  */
 343 void irq_enter(void)
 344 {
 345         int cpu = smp_processor_id();
 346
 347         rcu_irq_enter();
 348         if (is_idle_task(current) && !in_interrupt()) {
 349                 /*
 350                  * Prevent raise_softirq from needlessly waking up ksoftirqd
 351                  * here, as softirq will be serviced on return from interrupt.
 352                  */
 353                 local_bh_disable();
 354                 tick_check_idle(cpu);
 355                 _local_bh_enable();
 356         }
 357
 358         __irq_enter();
 359 }
 360
 361 static inline void invoke_softirq(void)
 362 {
 363         if (!force_irqthreads) {
 364 #ifdef CONFIG_HAVE_IRQ_EXIT_ON_IRQ_STACK
 365                 /*
 366                  * We can safely execute softirq on the current stack if
 367                  * it is the irq stack, because it should be near empty
 368                  * at this stage.
 369                  */
 370                 __do_softirq();
 371 #else
 372                 /*
 373                  * Otherwise, irq_exit() is called on the task stack that can
 374                  * be potentially deep already. So call softirq in its own stack
 375                  * to prevent from any overrun.
 376                  */
 377                 do_softirq_own_stack();
 378 #endif
 379         } else {
 380                 wakeup_softirqd();
 381         }
 382 }
 383
 384 static inline void tick_irq_exit(void)
 385 {
 386 #ifdef CONFIG_NO_HZ_COMMON
 387         int cpu = smp_processor_id();
 388
 389         /* Make sure that timer wheel updates are propagated */
 390         if ((idle_cpu(cpu) && !need_resched()) || tick_nohz_full_cpu(cpu)) {
 391                 if (!in_interrupt())
 392                         tick_nohz_irq_exit();
 393         }
 394 #endif
 395 }
 396
 397 /*
 398  * Exit an interrupt context. Process softirqs if needed and possible:
 399  */
 400 void irq_exit(void)
 401 {
 402 #ifndef __ARCH_IRQ_EXIT_IRQS_DISABLED
 403         local_irq_disable();
 404 #else
 405         WARN_ON_ONCE(!irqs_disabled());
 406 #endif
 407
 408         account_irq_exit_time(current);
 409         preempt_count_sub(HARDIRQ_OFFSET);
 410         if (!in_interrupt() && local_softirq_pending())
 411                 invoke_softirq();
 412
 413         tick_irq_exit();
 414         rcu_irq_exit();
 415         trace_hardirq_exit(); /* must be last! */
 416 }
 417
 418 /*
 419  * This function must run with irqs disabled!
 420  */
 421 inline void raise_softirq_irqoff(unsigned int nr)
 422 {
 423         __raise_softirq_irqoff(nr);
 424
 425         /*
 426          * If we're in an interrupt or softirq, we're done
 427          * (this also catches softirq-disabled code). We will
 428          * actually run the softirq once we return from
 429          * the irq or softirq.
 430          *
 431          * Otherwise we wake up ksoftirqd to make sure we
 432          * schedule the softirq soon.
 433          */
 434         if (!in_interrupt())
 435                 wakeup_softirqd();
 436 }
 437
 438 void raise_softirq(unsigned int nr)
 439 {
 440         unsigned long flags;
 441
 442         local_irq_save(flags);
 443         raise_softirq_irqoff(nr);
 444         local_irq_restore(flags);
 445 }
 446
 447 void __raise_softirq_irqoff(unsigned int nr)
 448 {
 449         trace_softirq_raise(nr);
 450         or_softirq_pending(1UL << nr);
 451 }
 452
 453 void open_softirq(int nr, void (*action)(struct softirq_action *))
 454 {
 455         softirq_vec[nr].action = action;
 456 }
 457
 458 /*
 459  * Tasklets
 460  */
 461 struct tasklet_head
 462 {
 463         struct tasklet_struct *head;
 464         struct tasklet_struct **tail;
 465 };
 466
 467 static DEFINE_PER_CPU(struct tasklet_head, tasklet_vec);
 468 static DEFINE_PER_CPU(struct tasklet_head, tasklet_hi_vec);
 469
 470 void __tasklet_schedule(struct tasklet_struct *t)
 471 {
 472         unsigned long flags;
 473
 474         local_irq_save(flags);
 475         t->next = NULL;
 476         *__this_cpu_read(tasklet_vec.tail) = t;
 477         __this_cpu_write(tasklet_vec.tail, &(t->next));
 478         raise_softirq_irqoff(TASKLET_SOFTIRQ);
 479         local_irq_restore(flags);
 480 }
 481
 482 EXPORT_SYMBOL(__tasklet_schedule);
 483
 484 void __tasklet_hi_schedule(struct tasklet_struct *t)
 485 {
 486         unsigned long flags;
 487
 488         local_irq_save(flags);
 489         t->next = NULL;
 490         *__this_cpu_read(tasklet_hi_vec.tail) = t;
 491         __this_cpu_write(tasklet_hi_vec.tail,  &(t->next));
 492         raise_softirq_irqoff(HI_SOFTIRQ);
 493         local_irq_restore(flags);
 494 }
 495
 496 EXPORT_SYMBOL(__tasklet_hi_schedule);
 497
 498 void __tasklet_hi_schedule_first(struct tasklet_struct *t)
 499 {
 500         BUG_ON(!irqs_disabled());
 501
 502         t->next = __this_cpu_read(tasklet_hi_vec.head);
 503         __this_cpu_write(tasklet_hi_vec.head, t);
 504         __raise_softirq_irqoff(HI_SOFTIRQ);
 505 }
 506
 507 EXPORT_SYMBOL(__tasklet_hi_schedule_first);
 508
 509 static void tasklet_action(struct softirq_action *a)
 510 {
 511         struct tasklet_struct *list;
 512
 513         local_irq_disable();
 514         list = __this_cpu_read(tasklet_vec.head);
 515         __this_cpu_write(tasklet_vec.head, NULL);
 516         __this_cpu_write(tasklet_vec.tail, &__get_cpu_var(tasklet_vec).head);
 517         local_irq_enable();
 518
 519         while (list) {
 520                 struct tasklet_struct *t = list;
 521
 522                 list = list->next;
 523
 524                 if (tasklet_trylock(t)) {
 525                         if (!atomic_read(&t->count)) {
 526                                 if (!test_and_clear_bit(TASKLET_STATE_SCHED, &t->state))
 527                                         BUG();
 528                                 t->func(t->data);
 529                                 tasklet_unlock(t);
 530                                 continue;
 531                         }
 532                         tasklet_unlock(t);
 533                 }
 534
 535                 local_irq_disable();
 536                 t->next = NULL;
 537                 *__this_cpu_read(tasklet_vec.tail) = t;
 538                 __this_cpu_write(tasklet_vec.tail, &(t->next));
 539                 __raise_softirq_irqoff(TASKLET_SOFTIRQ);
 540                 local_irq_enable();
 541         }
 542 }
 543
 544 static void tasklet_hi_action(struct softirq_action *a)
 545 {
 546         struct tasklet_struct *list;
 547
 548         local_irq_disable();
 549         list = __this_cpu_read(tasklet_hi_vec.head);
 550         __this_cpu_write(tasklet_hi_vec.head, NULL);
 551         __this_cpu_write(tasklet_hi_vec.tail, &__get_cpu_var(tasklet_hi_vec).head);
 552         local_irq_enable();
 553
 554         while (list) {
 555                 struct tasklet_struct *t = list;
 556
 557                 list = list->next;
 558
 559                 if (tasklet_trylock(t)) {
 560                         if (!atomic_read(&t->count)) {
 561                                 if (!test_and_clear_bit(TASKLET_STATE_SCHED, &t->state))
 562                                         BUG();
 563                                 t->func(t->data);
 564                                 tasklet_unlock(t);
 565                                 continue;
 566                         }
 567                         tasklet_unlock(t);
 568                 }
 569
 570                 local_irq_disable();
 571                 t->next = NULL;
 572                 *__this_cpu_read(tasklet_hi_vec.tail) = t;
 573                 __this_cpu_write(tasklet_hi_vec.tail, &(t->next));
 574                 __raise_softirq_irqoff(HI_SOFTIRQ);
 575                 local_irq_enable();
 576         }
 577 }
 578
 579
 580 void tasklet_init(struct tasklet_struct *t,
 581                   void (*func)(unsigned long), unsigned long data)
 582 {
 583         t->next = NULL;
 584         t->state = 0;
 585         atomic_set(&t->count, 0);
 586         t->func = func;
 587         t->data = data;
 588 }
 589
 590 EXPORT_SYMBOL(tasklet_init);
 591
 592 void tasklet_kill(struct tasklet_struct *t)
 593 {
 594         if (in_interrupt())
 595                 printk("Attempt to kill tasklet from interrupt\n");
 596
 597         while (test_and_set_bit(TASKLET_STATE_SCHED, &t->state)) {
 598                 do {
 599                         yield();
 600                 } while (test_bit(TASKLET_STATE_SCHED, &t->state));
 601         }
 602         tasklet_unlock_wait(t);
 603         clear_bit(TASKLET_STATE_SCHED, &t->state);
 604 }
 605
 606 EXPORT_SYMBOL(tasklet_kill);
 607
 608 /*
 609  * tasklet_hrtimer
 610  */
 611
 612 /*
 613  * The trampoline is called when the hrtimer expires. It schedules a tasklet
 614  * to run __tasklet_hrtimer_trampoline() which in turn will call the intended
 615  * hrtimer callback, but from softirq context.
 616  */
 617 static enum hrtimer_restart __hrtimer_tasklet_trampoline(struct hrtimer *timer)
 618 {
 619         struct tasklet_hrtimer *ttimer =
 620                 container_of(timer, struct tasklet_hrtimer, timer);
 621
 622         tasklet_hi_schedule(&ttimer->tasklet);
 623         return HRTIMER_NORESTART;
 624 }
 625
 626 /*
 627  * Helper function which calls the hrtimer callback from
 628  * tasklet/softirq context
 629  */
 630 static void __tasklet_hrtimer_trampoline(unsigned long data)
 631 {
 632         struct tasklet_hrtimer *ttimer = (void *)data;
 633         enum hrtimer_restart restart;
 634
 635         restart = ttimer->function(&ttimer->timer);
 636         if (restart != HRTIMER_NORESTART)
 637                 hrtimer_restart(&ttimer->timer);
 638 }
 639
 640 /**
 641  * tasklet_hrtimer_init - Init a tasklet/hrtimer combo for softirq callbacks
 642  * @ttimer:      tasklet_hrtimer which is initialized
 643  * @function:    hrtimer callback function which gets called from softirq context
 644  * @which_clock: clock id (CLOCK_MONOTONIC/CLOCK_REALTIME)
 645  * @mode:        hrtimer mode (HRTIMER_MODE_ABS/HRTIMER_MODE_REL)
 646  */
 647 void tasklet_hrtimer_init(struct tasklet_hrtimer *ttimer,
 648                           enum hrtimer_restart (*function)(struct hrtimer *),
 649                           clockid_t which_clock, enum hrtimer_mode mode)
 650 {
 651         hrtimer_init(&ttimer->timer, which_clock, mode);
 652         ttimer->timer.function = __hrtimer_tasklet_trampoline;
 653         tasklet_init(&ttimer->tasklet, __tasklet_hrtimer_trampoline,
 654                      (unsigned long)ttimer);
 655         ttimer->function = function;
 656 }
 657 EXPORT_SYMBOL_GPL(tasklet_hrtimer_init);
 658
 659 void __init softirq_init(void)
 660 {
 661         int cpu;
 662
 663         for_each_possible_cpu(cpu) {
 664                 per_cpu(tasklet_vec, cpu).tail =
 665                         &per_cpu(tasklet_vec, cpu).head;
 666                 per_cpu(tasklet_hi_vec, cpu).tail =
 667                         &per_cpu(tasklet_hi_vec, cpu).head;
 668         }
 669
 670         open_softirq(TASKLET_SOFTIRQ, tasklet_action);
 671         open_softirq(HI_SOFTIRQ, tasklet_hi_action);
 672 }
 673
 674 static int ksoftirqd_should_run(unsigned int cpu)
 675 {
 676         return local_softirq_pending();
 677 }
 678
 679 static void run_ksoftirqd(unsigned int cpu)
 680 {
 681         local_irq_disable();
 682         if (local_softirq_pending()) {
 683                 /*
 684                  * We can safely run softirq on inline stack, as we are not deep
 685                  * in the task stack here.
 686                  */
 687                 __do_softirq();
 688                 rcu_note_context_switch(cpu);
 689                 local_irq_enable();
 690                 cond_resched();
 691                 return;
 692         }
 693         local_irq_enable();
 694 }
 695
 696 #ifdef CONFIG_HOTPLUG_CPU
 697 /*
 698  * tasklet_kill_immediate is called to remove a tasklet which can already be
 699  * scheduled for execution on @cpu.
 700  *
 701  * Unlike tasklet_kill, this function removes the tasklet
 702  * _immediately_, even if the tasklet is in TASKLET_STATE_SCHED state.
 703  *
 704  * When this function is called, @cpu must be in the CPU_DEAD state.
 705  */
 706 void tasklet_kill_immediate(struct tasklet_struct *t, unsigned int cpu)
 707 {
 708         struct tasklet_struct **i;
 709
 710         BUG_ON(cpu_online(cpu));
 711         BUG_ON(test_bit(TASKLET_STATE_RUN, &t->state));
 712
 713         if (!test_bit(TASKLET_STATE_SCHED, &t->state))
 714                 return;
 715
 716         /* CPU is dead, so no lock needed. */
 717         for (i = &per_cpu(tasklet_vec, cpu).head; *i; i = &(*i)->next) {
 718                 if (*i == t) {
 719                         *i = t->next;
 720                         /* If this was the tail element, move the tail ptr */
 721                         if (*i == NULL)
 722                                 per_cpu(tasklet_vec, cpu).tail = i;
 723                         return;
 724                 }
 725         }
 726         BUG();
 727 }
 728
 729 static void takeover_tasklets(unsigned int cpu)
 730 {
 731         /* CPU is dead, so no lock needed. */
 732         local_irq_disable();
 733
 734         /* Find end, append list for that CPU. */
 735         if (&per_cpu(tasklet_vec, cpu).head != per_cpu(tasklet_vec, cpu).tail) {
 736                 *__this_cpu_read(tasklet_vec.tail) = per_cpu(tasklet_vec, cpu).head;
 737                 this_cpu_write(tasklet_vec.tail, per_cpu(tasklet_vec, cpu).tail);
 738                 per_cpu(tasklet_vec, cpu).head = NULL;
 739                 per_cpu(tasklet_vec, cpu).tail = &per_cpu(tasklet_vec, cpu).head;
 740         }
 741         raise_softirq_irqoff(TASKLET_SOFTIRQ);
 742
 743         if (&per_cpu(tasklet_hi_vec, cpu).head != per_cpu(tasklet_hi_vec, cpu).tail) {
 744                 *__this_cpu_read(tasklet_hi_vec.tail) = per_cpu(tasklet_hi_vec, cpu).head;
 745                 __this_cpu_write(tasklet_hi_vec.tail, per_cpu(tasklet_hi_vec, cpu).tail);
 746                 per_cpu(tasklet_hi_vec, cpu).head = NULL;
 747                 per_cpu(tasklet_hi_vec, cpu).tail = &per_cpu(tasklet_hi_vec, cpu).head;
 748         }
 749         raise_softirq_irqoff(HI_SOFTIRQ);
 750
 751         local_irq_enable();
 752 }
 753 #endif /* CONFIG_HOTPLUG_CPU */
 754
 755 static int cpu_callback(struct notifier_block *nfb,
 756                                   unsigned long action,
 757                                   void *hcpu)
 758 {
 759         switch (action) {
 760 #ifdef CONFIG_HOTPLUG_CPU
 761         case CPU_DEAD:
 762         case CPU_DEAD_FROZEN:
 763                 takeover_tasklets((unsigned long)hcpu);
 764                 break;
 765 #endif /* CONFIG_HOTPLUG_CPU */
 766         }
 767         return NOTIFY_OK;
 768 }
 769
 770 static struct notifier_block cpu_nfb = {
 771         .notifier_call = cpu_callback
 772 };
 773
 774 static struct smp_hotplug_thread softirq_threads = {
 775         .store                  = &ksoftirqd,
 776         .thread_should_run      = ksoftirqd_should_run,
 777         .thread_fn              = run_ksoftirqd,
 778         .thread_comm            = "ksoftirqd/%u",
 779 };
 780
 781 static __init int spawn_ksoftirqd(void)
 782 {
 783         register_cpu_notifier(&cpu_nfb);
 784
 785         BUG_ON(smpboot_register_percpu_thread(&softirq_threads));
 786
 787         return 0;
 788 }
 789 early_initcall(spawn_ksoftirqd);
 790
 791 /*
 792  * [ These __weak aliases are kept in a separate compilation unit, so that
 793  *   GCC does not inline them incorrectly. ]
 794  */
 795
 796 int __init __weak early_irq_init(void)
 797 {
 798         return 0;
 799 }
 800
 801 int __init __weak arch_probe_nr_irqs(void)
 802 {
 803         return NR_IRQS_LEGACY;
 804 }
 805
 806 int __init __weak arch_early_irq_init(void)
 807 {
 808         return 0;
 809 }