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