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 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 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         WARN_ON_ONCE(!irqs_disabled());
 145
 146         if (softirq_count() == (cnt & SOFTIRQ_MASK))
 147                 trace_softirqs_on(_RET_IP_);
 148         preempt_count_sub(cnt);
 149 }
 150
 151 /*
 152  * Special-case - softirqs can safely be enabled in
 153  * cond_resched_softirq(), or by __do_softirq(),
 154  * without processing still-pending softirqs:
 155  */
 156 void _local_bh_enable(void)
 157 {
 158         WARN_ON_ONCE(in_irq());
 159         __local_bh_enable(SOFTIRQ_DISABLE_OFFSET);
 160 }
 161 EXPORT_SYMBOL(_local_bh_enable);
 162
 163 void __local_bh_enable_ip(unsigned long ip, unsigned int cnt)
 164 {
 165         WARN_ON_ONCE(in_irq() || irqs_disabled());
 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         tsk_restore_flags(current, 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(pending))
 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(local_softirq_pending()))
 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         WARN_ON_ONCE(!irqs_disabled());
 404 #endif
 405
 406         account_irq_exit_time(current);
 407         preempt_count_sub(HARDIRQ_OFFSET);
 408         if (!in_interrupt() && local_softirq_pending())
 409                 invoke_softirq();
 410
 411         tick_irq_exit();
 412         rcu_irq_exit();
 413         trace_hardirq_exit(); /* must be last! */
 414 }
 415
 416 /*
 417  * This function must run with irqs disabled!
 418  */
 419 inline void raise_softirq_irqoff(unsigned int nr)
 420 {
 421         __raise_softirq_irqoff(nr);
 422
 423         /*
 424          * If we're in an interrupt or softirq, we're done
 425          * (this also catches softirq-disabled code). We will
 426          * actually run the softirq once we return from
 427          * the irq or softirq.
 428          *
 429          * Otherwise we wake up ksoftirqd to make sure we
 430          * schedule the softirq soon.
 431          */
 432         if (!in_interrupt())
 433                 wakeup_softirqd();
 434 }
 435
 436 void raise_softirq(unsigned int nr)
 437 {
 438         unsigned long flags;
 439
 440         local_irq_save(flags);
 441         raise_softirq_irqoff(nr);
 442         local_irq_restore(flags);
 443 }
 444
 445 void __raise_softirq_irqoff(unsigned int nr)
 446 {
 447         trace_softirq_raise(nr);
 448         or_softirq_pending(1UL << nr);
 449 }
 450
 451 void open_softirq(int nr, void (*action)(struct softirq_action *))
 452 {
 453         softirq_vec[nr].action = action;
 454 }
 455
 456 /*
 457  * Tasklets
 458  */
 459 struct tasklet_head {
 460         struct tasklet_struct *head;
 461         struct tasklet_struct **tail;
 462 };
 463
 464 static DEFINE_PER_CPU(struct tasklet_head, tasklet_vec);
 465 static DEFINE_PER_CPU(struct tasklet_head, tasklet_hi_vec);
 466
 467 void __tasklet_schedule(struct tasklet_struct *t)
 468 {
 469         unsigned long flags;
 470
 471         local_irq_save(flags);
 472         t->next = NULL;
 473         *__this_cpu_read(tasklet_vec.tail) = t;
 474         __this_cpu_write(tasklet_vec.tail, &(t->next));
 475         raise_softirq_irqoff(TASKLET_SOFTIRQ);
 476         local_irq_restore(flags);
 477 }
 478 EXPORT_SYMBOL(__tasklet_schedule);
 479
 480 void __tasklet_hi_schedule(struct tasklet_struct *t)
 481 {
 482         unsigned long flags;
 483
 484         local_irq_save(flags);
 485         t->next = NULL;
 486         *__this_cpu_read(tasklet_hi_vec.tail) = t;
 487         __this_cpu_write(tasklet_hi_vec.tail,  &(t->next));
 488         raise_softirq_irqoff(HI_SOFTIRQ);
 489         local_irq_restore(flags);
 490 }
 491 EXPORT_SYMBOL(__tasklet_hi_schedule);
 492
 493 void __tasklet_hi_schedule_first(struct tasklet_struct *t)
 494 {
 495         BUG_ON(!irqs_disabled());
 496
 497         t->next = __this_cpu_read(tasklet_hi_vec.head);
 498         __this_cpu_write(tasklet_hi_vec.head, t);
 499         __raise_softirq_irqoff(HI_SOFTIRQ);
 500 }
 501 EXPORT_SYMBOL(__tasklet_hi_schedule_first);
 502
 503 static __latent_entropy void tasklet_action(struct softirq_action *a)
 504 {
 505         struct tasklet_struct *list;
 506
 507         local_irq_disable();
 508         list = __this_cpu_read(tasklet_vec.head);
 509         __this_cpu_write(tasklet_vec.head, NULL);
 510         __this_cpu_write(tasklet_vec.tail, this_cpu_ptr(&tasklet_vec.head));
 511         local_irq_enable();
 512
 513         while (list) {
 514                 struct tasklet_struct *t = list;
 515
 516                 list = list->next;
 517
 518                 if (tasklet_trylock(t)) {
 519                         if (!atomic_read(&t->count)) {
 520                                 if (!test_and_clear_bit(TASKLET_STATE_SCHED,
 521                                                         &t->state))
 522                                         BUG();
 523                                 t->func(t->data);
 524                                 tasklet_unlock(t);
 525                                 continue;
 526                         }
 527                         tasklet_unlock(t);
 528                 }
 529
 530                 local_irq_disable();
 531                 t->next = NULL;
 532                 *__this_cpu_read(tasklet_vec.tail) = t;
 533                 __this_cpu_write(tasklet_vec.tail, &(t->next));
 534                 __raise_softirq_irqoff(TASKLET_SOFTIRQ);
 535                 local_irq_enable();
 536         }
 537 }
 538
 539 static __latent_entropy void tasklet_hi_action(struct softirq_action *a)
 540 {
 541         struct tasklet_struct *list;
 542
 543         local_irq_disable();
 544         list = __this_cpu_read(tasklet_hi_vec.head);
 545         __this_cpu_write(tasklet_hi_vec.head, NULL);
 546         __this_cpu_write(tasklet_hi_vec.tail, this_cpu_ptr(&tasklet_hi_vec.head));
 547         local_irq_enable();
 548
 549         while (list) {
 550                 struct tasklet_struct *t = list;
 551
 552                 list = list->next;
 553
 554                 if (tasklet_trylock(t)) {
 555                         if (!atomic_read(&t->count)) {
 556                                 if (!test_and_clear_bit(TASKLET_STATE_SCHED,
 557                                                         &t->state))
 558                                         BUG();
 559                                 t->func(t->data);
 560                                 tasklet_unlock(t);
 561                                 continue;
 562                         }
 563                         tasklet_unlock(t);
 564                 }
 565
 566                 local_irq_disable();
 567                 t->next = NULL;
 568                 *__this_cpu_read(tasklet_hi_vec.tail) = t;
 569                 __this_cpu_write(tasklet_hi_vec.tail, &(t->next));
 570                 __raise_softirq_irqoff(HI_SOFTIRQ);
 571                 local_irq_enable();
 572         }
 573 }
 574
 575 void tasklet_init(struct tasklet_struct *t,
 576                   void (*func)(unsigned long), unsigned long data)
 577 {
 578         t->next = NULL;
 579         t->state = 0;
 580         atomic_set(&t->count, 0);
 581         t->func = func;
 582         t->data = data;
 583 }
 584 EXPORT_SYMBOL(tasklet_init);
 585
 586 void tasklet_kill(struct tasklet_struct *t)
 587 {
 588         if (in_interrupt())
 589                 pr_notice("Attempt to kill tasklet from interrupt\n");
 590
 591         while (test_and_set_bit(TASKLET_STATE_SCHED, &t->state)) {
 592                 do {
 593                         yield();
 594                 } while (test_bit(TASKLET_STATE_SCHED, &t->state));
 595         }
 596         tasklet_unlock_wait(t);
 597         clear_bit(TASKLET_STATE_SCHED, &t->state);
 598 }
 599 EXPORT_SYMBOL(tasklet_kill);
 600
 601 /*
 602  * tasklet_hrtimer
 603  */
 604
 605 /*
 606  * The trampoline is called when the hrtimer expires. It schedules a tasklet
 607  * to run __tasklet_hrtimer_trampoline() which in turn will call the intended
 608  * hrtimer callback, but from softirq context.
 609  */
 610 static enum hrtimer_restart __hrtimer_tasklet_trampoline(struct hrtimer *timer)
 611 {
 612         struct tasklet_hrtimer *ttimer =
 613                 container_of(timer, struct tasklet_hrtimer, timer);
 614
 615         tasklet_hi_schedule(&ttimer->tasklet);
 616         return HRTIMER_NORESTART;
 617 }
 618
 619 /*
 620  * Helper function which calls the hrtimer callback from
 621  * tasklet/softirq context
 622  */
 623 static void __tasklet_hrtimer_trampoline(unsigned long data)
 624 {
 625         struct tasklet_hrtimer *ttimer = (void *)data;
 626         enum hrtimer_restart restart;
 627
 628         restart = ttimer->function(&ttimer->timer);
 629         if (restart != HRTIMER_NORESTART)
 630                 hrtimer_restart(&ttimer->timer);
 631 }
 632
 633 /**
 634  * tasklet_hrtimer_init - Init a tasklet/hrtimer combo for softirq callbacks
 635  * @ttimer:      tasklet_hrtimer which is initialized
 636  * @function:    hrtimer callback function which gets called from softirq context
 637  * @which_clock: clock id (CLOCK_MONOTONIC/CLOCK_REALTIME)
 638  * @mode:        hrtimer mode (HRTIMER_MODE_ABS/HRTIMER_MODE_REL)
 639  */
 640 void tasklet_hrtimer_init(struct tasklet_hrtimer *ttimer,
 641                           enum hrtimer_restart (*function)(struct hrtimer *),
 642                           clockid_t which_clock, enum hrtimer_mode mode)
 643 {
 644         hrtimer_init(&ttimer->timer, which_clock, mode);
 645         ttimer->timer.function = __hrtimer_tasklet_trampoline;
 646         tasklet_init(&ttimer->tasklet, __tasklet_hrtimer_trampoline,
 647                      (unsigned long)ttimer);
 648         ttimer->function = function;
 649 }
 650 EXPORT_SYMBOL_GPL(tasklet_hrtimer_init);
 651
 652 void __init softirq_init(void)
 653 {
 654         int cpu;
 655
 656         for_each_possible_cpu(cpu) {
 657                 per_cpu(tasklet_vec, cpu).tail =
 658                         &per_cpu(tasklet_vec, cpu).head;
 659                 per_cpu(tasklet_hi_vec, cpu).tail =
 660                         &per_cpu(tasklet_hi_vec, cpu).head;
 661         }
 662
 663         open_softirq(TASKLET_SOFTIRQ, tasklet_action);
 664         open_softirq(HI_SOFTIRQ, tasklet_hi_action);
 665 }
 666
 667 static int ksoftirqd_should_run(unsigned int cpu)
 668 {
 669         return local_softirq_pending();
 670 }
 671
 672 static void run_ksoftirqd(unsigned int cpu)
 673 {
 674         local_irq_disable();
 675         if (local_softirq_pending()) {
 676                 /*
 677                  * We can safely run softirq on inline stack, as we are not deep
 678                  * in the task stack here.
 679                  */
 680                 __do_softirq();
 681                 local_irq_enable();
 682                 cond_resched_rcu_qs();
 683                 return;
 684         }
 685         local_irq_enable();
 686 }
 687
 688 #ifdef CONFIG_HOTPLUG_CPU
 689 /*
 690  * tasklet_kill_immediate is called to remove a tasklet which can already be
 691  * scheduled for execution on @cpu.
 692  *
 693  * Unlike tasklet_kill, this function removes the tasklet
 694  * _immediately_, even if the tasklet is in TASKLET_STATE_SCHED state.
 695  *
 696  * When this function is called, @cpu must be in the CPU_DEAD state.
 697  */
 698 void tasklet_kill_immediate(struct tasklet_struct *t, unsigned int cpu)
 699 {
 700         struct tasklet_struct **i;
 701
 702         BUG_ON(cpu_online(cpu));
 703         BUG_ON(test_bit(TASKLET_STATE_RUN, &t->state));
 704
 705         if (!test_bit(TASKLET_STATE_SCHED, &t->state))
 706                 return;
 707
 708         /* CPU is dead, so no lock needed. */
 709         for (i = &per_cpu(tasklet_vec, cpu).head; *i; i = &(*i)->next) {
 710                 if (*i == t) {
 711                         *i = t->next;
 712                         /* If this was the tail element, move the tail ptr */
 713                         if (*i == NULL)
 714                                 per_cpu(tasklet_vec, cpu).tail = i;
 715                         return;
 716                 }
 717         }
 718         BUG();
 719 }
 720
 721 static int takeover_tasklets(unsigned int cpu)
 722 {
 723         /* CPU is dead, so no lock needed. */
 724         local_irq_disable();
 725
 726         /* Find end, append list for that CPU. */
 727         if (&per_cpu(tasklet_vec, cpu).head != per_cpu(tasklet_vec, cpu).tail) {
 728                 *__this_cpu_read(tasklet_vec.tail) = per_cpu(tasklet_vec, cpu).head;
 729                 this_cpu_write(tasklet_vec.tail, per_cpu(tasklet_vec, cpu).tail);
 730                 per_cpu(tasklet_vec, cpu).head = NULL;
 731                 per_cpu(tasklet_vec, cpu).tail = &per_cpu(tasklet_vec, cpu).head;
 732         }
 733         raise_softirq_irqoff(TASKLET_SOFTIRQ);
 734
 735         if (&per_cpu(tasklet_hi_vec, cpu).head != per_cpu(tasklet_hi_vec, cpu).tail) {
 736                 *__this_cpu_read(tasklet_hi_vec.tail) = per_cpu(tasklet_hi_vec, cpu).head;
 737                 __this_cpu_write(tasklet_hi_vec.tail, per_cpu(tasklet_hi_vec, cpu).tail);
 738                 per_cpu(tasklet_hi_vec, cpu).head = NULL;
 739                 per_cpu(tasklet_hi_vec, cpu).tail = &per_cpu(tasklet_hi_vec, cpu).head;
 740         }
 741         raise_softirq_irqoff(HI_SOFTIRQ);
 742
 743         local_irq_enable();
 744         return 0;
 745 }
 746 #else
 747 #define takeover_tasklets       NULL
 748 #endif /* CONFIG_HOTPLUG_CPU */
 749
 750 static struct smp_hotplug_thread softirq_threads = {
 751         .store                  = &ksoftirqd,
 752         .thread_should_run      = ksoftirqd_should_run,
 753         .thread_fn              = run_ksoftirqd,
 754         .thread_comm            = "ksoftirqd/%u",
 755 };
 756
 757 static __init int spawn_ksoftirqd(void)
 758 {
 759         cpuhp_setup_state_nocalls(CPUHP_SOFTIRQ_DEAD, "softirq:dead", NULL,
 760                                   takeover_tasklets);
 761         BUG_ON(smpboot_register_percpu_thread(&softirq_threads));
 762
 763         return 0;
 764 }
 765 early_initcall(spawn_ksoftirqd);
 766
 767 /*
 768  * [ These __weak aliases are kept in a separate compilation unit, so that
 769  *   GCC does not inline them incorrectly. ]
 770  */
 771
 772 int __init __weak early_irq_init(void)
 773 {
 774         return 0;
 775 }
 776
 777 int __init __weak arch_probe_nr_irqs(void)
 778 {
 779         return NR_IRQS_LEGACY;
 780 }
 781
 782 int __init __weak arch_early_irq_init(void)
 783 {
 784         return 0;
 785 }
 786
 787 unsigned int __weak arch_dynirq_lower_bound(unsigned int from)
 788 {
 789         return from;
 790 }