kernel/softirq.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *      linux/kernel/softirq.c
   4  *
   5  *      Copyright (C) 1992 Linus Torvalds
   6  *
   7  *      Rewritten. Old one was good in 2.2, but in 2.3 it was immoral. --ANK (990903)
   8  */
   9
  10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  11
  12 #include <linux/export.h>
  13 #include <linux/kernel_stat.h>
  14 #include <linux/interrupt.h>
  15 #include <linux/init.h>
  16 #include <linux/mm.h>
  17 #include <linux/notifier.h>
  18 #include <linux/percpu.h>
  19 #include <linux/cpu.h>
  20 #include <linux/freezer.h>
  21 #include <linux/kthread.h>
  22 #include <linux/rcupdate.h>
  23 #include <linux/ftrace.h>
  24 #include <linux/smp.h>
  25 #include <linux/smpboot.h>
  26 #include <linux/tick.h>
  27 #include <linux/irq.h>
  28
  29 #define CREATE_TRACE_POINTS
  30 #include <trace/events/irq.h>
  31
  32 /*
  33    - No shared variables, all the data are CPU local.
  34    - If a softirq needs serialization, let it serialize itself
  35      by its own spinlocks.
  36    - Even if softirq is serialized, only local cpu is marked for
  37      execution. Hence, we get something sort of weak cpu binding.
  38      Though it is still not clear, will it result in better locality
  39      or will not.
  40
  41    Examples:
  42    - NET RX softirq. It is multithreaded and does not require
  43      any global serialization.
  44    - NET TX softirq. It kicks software netdevice queues, hence
  45      it is logically serialized per device, but this serialization
  46      is invisible to common code.
  47    - Tasklets: serialized wrt itself.
  48  */
  49
  50 #ifndef __ARCH_IRQ_STAT
  51 DEFINE_PER_CPU_ALIGNED(irq_cpustat_t, irq_stat);
  52 EXPORT_PER_CPU_SYMBOL(irq_stat);
  53 #endif
  54
  55 static struct softirq_action softirq_vec[NR_SOFTIRQS] __cacheline_aligned_in_smp;
  56
  57 DEFINE_PER_CPU(struct task_struct *, ksoftirqd);
  58
  59 const char * const softirq_to_name[NR_SOFTIRQS] = {
  60         "HI", "TIMER", "NET_TX", "NET_RX", "BLOCK", "IRQ_POLL",
  61         "TASKLET", "SCHED", "HRTIMER", "RCU"
  62 };
  63
  64 /*
  65  * we cannot loop indefinitely here to avoid userspace starvation,
  66  * but we also don't want to introduce a worst case 1/HZ latency
  67  * to the pending events, so lets the scheduler to balance
  68  * the softirq load for us.
  69  */
  70 static void wakeup_softirqd(void)
  71 {
  72         /* Interrupts are disabled: no need to stop preemption */
  73         struct task_struct *tsk = __this_cpu_read(ksoftirqd);
  74
  75         if (tsk && tsk->state != TASK_RUNNING)
  76                 wake_up_process(tsk);
  77 }
  78
  79 /*
  80  * If ksoftirqd is scheduled, we do not want to process pending softirqs
  81  * right now. Let ksoftirqd handle this at its own rate, to get fairness,
  82  * unless we're doing some of the synchronous softirqs.
  83  */
  84 #define SOFTIRQ_NOW_MASK ((1 << HI_SOFTIRQ) | (1 << TASKLET_SOFTIRQ))
  85 static bool ksoftirqd_running(unsigned long pending)
  86 {
  87         struct task_struct *tsk = __this_cpu_read(ksoftirqd);
  88
  89         if (pending & SOFTIRQ_NOW_MASK)
  90                 return false;
  91         return tsk && (tsk->state == TASK_RUNNING) &&
  92                 !__kthread_should_park(tsk);
  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
 111 DEFINE_PER_CPU(int, hardirqs_enabled);
 112 DEFINE_PER_CPU(int, hardirq_context);
 113 EXPORT_PER_CPU_SYMBOL_GPL(hardirqs_enabled);
 114 EXPORT_PER_CPU_SYMBOL_GPL(hardirq_context);
 115
 116 void __local_bh_disable_ip(unsigned long ip, unsigned int cnt)
 117 {
 118         unsigned long flags;
 119
 120         WARN_ON_ONCE(in_irq());
 121
 122         raw_local_irq_save(flags);
 123         /*
 124          * The preempt tracer hooks into preempt_count_add and will break
 125          * lockdep because it calls back into lockdep after SOFTIRQ_OFFSET
 126          * is set and before current->softirq_enabled is cleared.
 127          * We must manually increment preempt_count here and manually
 128          * call the trace_preempt_off later.
 129          */
 130         __preempt_count_add(cnt);
 131         /*
 132          * Were softirqs turned off above:
 133          */
 134         if (softirq_count() == (cnt & SOFTIRQ_MASK))
 135                 lockdep_softirqs_off(ip);
 136         raw_local_irq_restore(flags);
 137
 138         if (preempt_count() == cnt) {
 139 #ifdef CONFIG_DEBUG_PREEMPT
 140                 current->preempt_disable_ip = get_lock_parent_ip();
 141 #endif
 142                 trace_preempt_off(CALLER_ADDR0, get_lock_parent_ip());
 143         }
 144 }
 145 EXPORT_SYMBOL(__local_bh_disable_ip);
 146 #endif /* CONFIG_TRACE_IRQFLAGS */
 147
 148 static void __local_bh_enable(unsigned int cnt)
 149 {
 150         lockdep_assert_irqs_disabled();
 151
 152         if (preempt_count() == cnt)
 153                 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
 154
 155         if (softirq_count() == (cnt & SOFTIRQ_MASK))
 156                 lockdep_softirqs_on(_RET_IP_);
 157
 158         __preempt_count_sub(cnt);
 159 }
 160
 161 /*
 162  * Special-case - softirqs can safely be enabled by __do_softirq(),
 163  * without processing still-pending softirqs:
 164  */
 165 void _local_bh_enable(void)
 166 {
 167         WARN_ON_ONCE(in_irq());
 168         __local_bh_enable(SOFTIRQ_DISABLE_OFFSET);
 169 }
 170 EXPORT_SYMBOL(_local_bh_enable);
 171
 172 void __local_bh_enable_ip(unsigned long ip, unsigned int cnt)
 173 {
 174         WARN_ON_ONCE(in_irq());
 175         lockdep_assert_irqs_enabled();
 176 #ifdef CONFIG_TRACE_IRQFLAGS
 177         local_irq_disable();
 178 #endif
 179         /*
 180          * Are softirqs going to be turned on now:
 181          */
 182         if (softirq_count() == SOFTIRQ_DISABLE_OFFSET)
 183                 lockdep_softirqs_on(ip);
 184         /*
 185          * Keep preemption disabled until we are done with
 186          * softirq processing:
 187          */
 188         preempt_count_sub(cnt - 1);
 189
 190         if (unlikely(!in_interrupt() && local_softirq_pending())) {
 191                 /*
 192                  * Run softirq if any pending. And do it in its own stack
 193                  * as we may be calling this deep in a task call stack already.
 194                  */
 195                 do_softirq();
 196         }
 197
 198         preempt_count_dec();
 199 #ifdef CONFIG_TRACE_IRQFLAGS
 200         local_irq_enable();
 201 #endif
 202         preempt_check_resched();
 203 }
 204 EXPORT_SYMBOL(__local_bh_enable_ip);
 205
 206 /*
 207  * We restart softirq processing for at most MAX_SOFTIRQ_RESTART times,
 208  * but break the loop if need_resched() is set or after 2 ms.
 209  * The MAX_SOFTIRQ_TIME provides a nice upper bound in most cases, but in
 210  * certain cases, such as stop_machine(), jiffies may cease to
 211  * increment and so we need the MAX_SOFTIRQ_RESTART limit as
 212  * well to make sure we eventually return from this method.
 213  *
 214  * These limits have been established via experimentation.
 215  * The two things to balance is latency against fairness -
 216  * we want to handle softirqs as soon as possible, but they
 217  * should not be able to lock up the box.
 218  */
 219 #define MAX_SOFTIRQ_TIME  msecs_to_jiffies(2)
 220 #define MAX_SOFTIRQ_RESTART 10
 221
 222 #ifdef CONFIG_TRACE_IRQFLAGS
 223 /*
 224  * When we run softirqs from irq_exit() and thus on the hardirq stack we need
 225  * to keep the lockdep irq context tracking as tight as possible in order to
 226  * not miss-qualify lock contexts and miss possible deadlocks.
 227  */
 228
 229 static inline bool lockdep_softirq_start(void)
 230 {
 231         bool in_hardirq = false;
 232
 233         if (lockdep_hardirq_context()) {
 234                 in_hardirq = true;
 235                 lockdep_hardirq_exit();
 236         }
 237
 238         lockdep_softirq_enter();
 239
 240         return in_hardirq;
 241 }
 242
 243 static inline void lockdep_softirq_end(bool in_hardirq)
 244 {
 245         lockdep_softirq_exit();
 246
 247         if (in_hardirq)
 248                 lockdep_hardirq_enter();
 249 }
 250 #else
 251 static inline bool lockdep_softirq_start(void) { return false; }
 252 static inline void lockdep_softirq_end(bool in_hardirq) { }
 253 #endif
 254
 255 asmlinkage __visible void __softirq_entry __do_softirq(void)
 256 {
 257         unsigned long end = jiffies + MAX_SOFTIRQ_TIME;
 258         unsigned long old_flags = current->flags;
 259         int max_restart = MAX_SOFTIRQ_RESTART;
 260         struct softirq_action *h;
 261         bool in_hardirq;
 262         __u32 pending;
 263         int softirq_bit;
 264
 265         /*
 266          * Mask out PF_MEMALLOC as the current task context is borrowed for the
 267          * softirq. A softirq handled, such as network RX, might set PF_MEMALLOC
 268          * again if the socket is related to swapping.
 269          */
 270         current->flags &= ~PF_MEMALLOC;
 271
 272         pending = local_softirq_pending();
 273         account_irq_enter_time(current);
 274
 275         __local_bh_disable_ip(_RET_IP_, SOFTIRQ_OFFSET);
 276         in_hardirq = lockdep_softirq_start();
 277
 278 restart:
 279         /* Reset the pending bitmask before enabling irqs */
 280         set_softirq_pending(0);
 281
 282         local_irq_enable();
 283
 284         h = softirq_vec;
 285
 286         while ((softirq_bit = ffs(pending))) {
 287                 unsigned int vec_nr;
 288                 int prev_count;
 289
 290                 h += softirq_bit - 1;
 291
 292                 vec_nr = h - softirq_vec;
 293                 prev_count = preempt_count();
 294
 295                 kstat_incr_softirqs_this_cpu(vec_nr);
 296
 297                 trace_softirq_entry(vec_nr);
 298                 h->action(h);
 299                 trace_softirq_exit(vec_nr);
 300                 if (unlikely(prev_count != preempt_count())) {
 301                         pr_err("huh, entered softirq %u %s %p with preempt_count %08x, exited with %08x?\n",
 302                                vec_nr, softirq_to_name[vec_nr], h->action,
 303                                prev_count, preempt_count());
 304                         preempt_count_set(prev_count);
 305                 }
 306                 h++;
 307                 pending >>= softirq_bit;
 308         }
 309
 310         if (__this_cpu_read(ksoftirqd) == current)
 311                 rcu_softirq_qs();
 312         local_irq_disable();
 313
 314         pending = local_softirq_pending();
 315         if (pending) {
 316                 if (time_before(jiffies, end) && !need_resched() &&
 317                     --max_restart)
 318                         goto restart;
 319
 320                 wakeup_softirqd();
 321         }
 322
 323         lockdep_softirq_end(in_hardirq);
 324         account_irq_exit_time(current);
 325         __local_bh_enable(SOFTIRQ_OFFSET);
 326         WARN_ON_ONCE(in_interrupt());
 327         current_restore_flags(old_flags, PF_MEMALLOC);
 328 }
 329
 330 asmlinkage __visible void do_softirq(void)
 331 {
 332         __u32 pending;
 333         unsigned long flags;
 334
 335         if (in_interrupt())
 336                 return;
 337
 338         local_irq_save(flags);
 339
 340         pending = local_softirq_pending();
 341
 342         if (pending && !ksoftirqd_running(pending))
 343                 do_softirq_own_stack();
 344
 345         local_irq_restore(flags);
 346 }
 347
 348 /**
 349  * irq_enter_rcu - Enter an interrupt context with RCU watching
 350  */
 351 void irq_enter_rcu(void)
 352 {
 353         if (is_idle_task(current) && !in_interrupt()) {
 354                 /*
 355                  * Prevent raise_softirq from needlessly waking up ksoftirqd
 356                  * here, as softirq will be serviced on return from interrupt.
 357                  */
 358                 local_bh_disable();
 359                 tick_irq_enter();
 360                 _local_bh_enable();
 361         }
 362         __irq_enter();
 363 }
 364
 365 /**
 366  * irq_enter - Enter an interrupt context including RCU update
 367  */
 368 void irq_enter(void)
 369 {
 370         rcu_irq_enter();
 371         irq_enter_rcu();
 372 }
 373
 374 static inline void invoke_softirq(void)
 375 {
 376         if (ksoftirqd_running(local_softirq_pending()))
 377                 return;
 378
 379         if (!force_irqthreads) {
 380 #ifdef CONFIG_HAVE_IRQ_EXIT_ON_IRQ_STACK
 381                 /*
 382                  * We can safely execute softirq on the current stack if
 383                  * it is the irq stack, because it should be near empty
 384                  * at this stage.
 385                  */
 386                 __do_softirq();
 387 #else
 388                 /*
 389                  * Otherwise, irq_exit() is called on the task stack that can
 390                  * be potentially deep already. So call softirq in its own stack
 391                  * to prevent from any overrun.
 392                  */
 393                 do_softirq_own_stack();
 394 #endif
 395         } else {
 396                 wakeup_softirqd();
 397         }
 398 }
 399
 400 static inline void tick_irq_exit(void)
 401 {
 402 #ifdef CONFIG_NO_HZ_COMMON
 403         int cpu = smp_processor_id();
 404
 405         /* Make sure that timer wheel updates are propagated */
 406         if ((idle_cpu(cpu) && !need_resched()) || tick_nohz_full_cpu(cpu)) {
 407                 if (!in_irq())
 408                         tick_nohz_irq_exit();
 409         }
 410 #endif
 411 }
 412
 413 static inline void __irq_exit_rcu(void)
 414 {
 415 #ifndef __ARCH_IRQ_EXIT_IRQS_DISABLED
 416         local_irq_disable();
 417 #else
 418         lockdep_assert_irqs_disabled();
 419 #endif
 420         account_irq_exit_time(current);
 421         preempt_count_sub(HARDIRQ_OFFSET);
 422         if (!in_interrupt() && local_softirq_pending())
 423                 invoke_softirq();
 424
 425         tick_irq_exit();
 426 }
 427
 428 /**
 429  * irq_exit_rcu() - Exit an interrupt context without updating RCU
 430  *
 431  * Also processes softirqs if needed and possible.
 432  */
 433 void irq_exit_rcu(void)
 434 {
 435         __irq_exit_rcu();
 436          /* must be last! */
 437         lockdep_hardirq_exit();
 438 }
 439
 440 /**
 441  * irq_exit - Exit an interrupt context, update RCU and lockdep
 442  *
 443  * Also processes softirqs if needed and possible.
 444  */
 445 void irq_exit(void)
 446 {
 447         __irq_exit_rcu();
 448         rcu_irq_exit();
 449          /* must be last! */
 450         lockdep_hardirq_exit();
 451 }
 452
 453 /*
 454  * This function must run with irqs disabled!
 455  */
 456 inline void raise_softirq_irqoff(unsigned int nr)
 457 {
 458         __raise_softirq_irqoff(nr);
 459
 460         /*
 461          * If we're in an interrupt or softirq, we're done
 462          * (this also catches softirq-disabled code). We will
 463          * actually run the softirq once we return from
 464          * the irq or softirq.
 465          *
 466          * Otherwise we wake up ksoftirqd to make sure we
 467          * schedule the softirq soon.
 468          */
 469         if (!in_interrupt())
 470                 wakeup_softirqd();
 471 }
 472
 473 void raise_softirq(unsigned int nr)
 474 {
 475         unsigned long flags;
 476
 477         local_irq_save(flags);
 478         raise_softirq_irqoff(nr);
 479         local_irq_restore(flags);
 480 }
 481
 482 void __raise_softirq_irqoff(unsigned int nr)
 483 {
 484         lockdep_assert_irqs_disabled();
 485         trace_softirq_raise(nr);
 486         or_softirq_pending(1UL << nr);
 487 }
 488
 489 void open_softirq(int nr, void (*action)(struct softirq_action *))
 490 {
 491         softirq_vec[nr].action = action;
 492 }
 493
 494 /*
 495  * Tasklets
 496  */
 497 struct tasklet_head {
 498         struct tasklet_struct *head;
 499         struct tasklet_struct **tail;
 500 };
 501
 502 static DEFINE_PER_CPU(struct tasklet_head, tasklet_vec);
 503 static DEFINE_PER_CPU(struct tasklet_head, tasklet_hi_vec);
 504
 505 static void __tasklet_schedule_common(struct tasklet_struct *t,
 506                                       struct tasklet_head __percpu *headp,
 507                                       unsigned int softirq_nr)
 508 {
 509         struct tasklet_head *head;
 510         unsigned long flags;
 511
 512         local_irq_save(flags);
 513         head = this_cpu_ptr(headp);
 514         t->next = NULL;
 515         *head->tail = t;
 516         head->tail = &(t->next);
 517         raise_softirq_irqoff(softirq_nr);
 518         local_irq_restore(flags);
 519 }
 520
 521 void __tasklet_schedule(struct tasklet_struct *t)
 522 {
 523         __tasklet_schedule_common(t, &tasklet_vec,
 524                                   TASKLET_SOFTIRQ);
 525 }
 526 EXPORT_SYMBOL(__tasklet_schedule);
 527
 528 void __tasklet_hi_schedule(struct tasklet_struct *t)
 529 {
 530         __tasklet_schedule_common(t, &tasklet_hi_vec,
 531                                   HI_SOFTIRQ);
 532 }
 533 EXPORT_SYMBOL(__tasklet_hi_schedule);
 534
 535 static void tasklet_action_common(struct softirq_action *a,
 536                                   struct tasklet_head *tl_head,
 537                                   unsigned int softirq_nr)
 538 {
 539         struct tasklet_struct *list;
 540
 541         local_irq_disable();
 542         list = tl_head->head;
 543         tl_head->head = NULL;
 544         tl_head->tail = &tl_head->head;
 545         local_irq_enable();
 546
 547         while (list) {
 548                 struct tasklet_struct *t = list;
 549
 550                 list = list->next;
 551
 552                 if (tasklet_trylock(t)) {
 553                         if (!atomic_read(&t->count)) {
 554                                 if (!test_and_clear_bit(TASKLET_STATE_SCHED,
 555                                                         &t->state))
 556                                         BUG();
 557                                 if (t->use_callback)
 558                                         t->callback(t);
 559                                 else
 560                                         t->func(t->data);
 561                                 tasklet_unlock(t);
 562                                 continue;
 563                         }
 564                         tasklet_unlock(t);
 565                 }
 566
 567                 local_irq_disable();
 568                 t->next = NULL;
 569                 *tl_head->tail = t;
 570                 tl_head->tail = &t->next;
 571                 __raise_softirq_irqoff(softirq_nr);
 572                 local_irq_enable();
 573         }
 574 }
 575
 576 static __latent_entropy void tasklet_action(struct softirq_action *a)
 577 {
 578         tasklet_action_common(a, this_cpu_ptr(&tasklet_vec), TASKLET_SOFTIRQ);
 579 }
 580
 581 static __latent_entropy void tasklet_hi_action(struct softirq_action *a)
 582 {
 583         tasklet_action_common(a, this_cpu_ptr(&tasklet_hi_vec), HI_SOFTIRQ);
 584 }
 585
 586 void tasklet_setup(struct tasklet_struct *t,
 587                    void (*callback)(struct tasklet_struct *))
 588 {
 589         t->next = NULL;
 590         t->state = 0;
 591         atomic_set(&t->count, 0);
 592         t->callback = callback;
 593         t->use_callback = true;
 594         t->data = 0;
 595 }
 596 EXPORT_SYMBOL(tasklet_setup);
 597
 598 void tasklet_init(struct tasklet_struct *t,
 599                   void (*func)(unsigned long), unsigned long data)
 600 {
 601         t->next = NULL;
 602         t->state = 0;
 603         atomic_set(&t->count, 0);
 604         t->func = func;
 605         t->use_callback = false;
 606         t->data = data;
 607 }
 608 EXPORT_SYMBOL(tasklet_init);
 609
 610 void tasklet_kill(struct tasklet_struct *t)
 611 {
 612         if (in_interrupt())
 613                 pr_notice("Attempt to kill tasklet from interrupt\n");
 614
 615         while (test_and_set_bit(TASKLET_STATE_SCHED, &t->state)) {
 616                 do {
 617                         yield();
 618                 } while (test_bit(TASKLET_STATE_SCHED, &t->state));
 619         }
 620         tasklet_unlock_wait(t);
 621         clear_bit(TASKLET_STATE_SCHED, &t->state);
 622 }
 623 EXPORT_SYMBOL(tasklet_kill);
 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 }