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