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  *      Remote softirq infrastructure is by Jens Axboe.
  11  */
  12
  13 #include <linux/module.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/tick.h>
  27
  28 #define CREATE_TRACE_POINTS
  29 #include <trace/events/irq.h>
  30
  31 #include <asm/irq.h>
  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 irq_cpustat_t irq_stat[NR_CPUS] ____cacheline_aligned;
  52 EXPORT_SYMBOL(irq_stat);
  53 #endif
  54
  55 static struct softirq_action softirq_vec[NR_SOFTIRQS] __cacheline_aligned_in_smp;
  56
  57 static DEFINE_PER_CPU(struct task_struct *, ksoftirqd);
  58
  59 char *softirq_to_name[NR_SOFTIRQS] = {
  60         "HI", "TIMER", "NET_TX", "NET_RX", "BLOCK", "BLOCK_IOPOLL",
  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 void wakeup_softirqd(void)
  71 {
  72         /* Interrupts are disabled: no need to stop preemption */
  73         struct task_struct *tsk = __get_cpu_var(ksoftirqd);
  74
  75         if (tsk && tsk->state != TASK_RUNNING)
  76                 wake_up_process(tsk);
  77 }
  78
  79 /*
  80  * preempt_count and SOFTIRQ_OFFSET usage:
  81  * - preempt_count is changed by SOFTIRQ_OFFSET on entering or leaving
  82  *   softirq processing.
  83  * - preempt_count is changed by SOFTIRQ_DISABLE_OFFSET (= 2 * SOFTIRQ_OFFSET)
  84  *   on local_bh_disable or local_bh_enable.
  85  * This lets us distinguish between whether we are currently processing
  86  * softirq and whether we just have bh disabled.
  87  */
  88
  89 /*
  90  * This one is for softirq.c-internal use,
  91  * where hardirqs are disabled legitimately:
  92  */
  93 #ifdef CONFIG_TRACE_IRQFLAGS
  94 static void __local_bh_disable(unsigned long ip, unsigned int cnt)
  95 {
  96         unsigned long flags;
  97
  98         WARN_ON_ONCE(in_irq());
  99
 100         raw_local_irq_save(flags);
 101         /*
 102          * The preempt tracer hooks into add_preempt_count and will break
 103          * lockdep because it calls back into lockdep after SOFTIRQ_OFFSET
 104          * is set and before current->softirq_enabled is cleared.
 105          * We must manually increment preempt_count here and manually
 106          * call the trace_preempt_off later.
 107          */
 108         preempt_count() += cnt;
 109         /*
 110          * Were softirqs turned off above:
 111          */
 112         if (softirq_count() == cnt)
 113                 trace_softirqs_off(ip);
 114         raw_local_irq_restore(flags);
 115
 116         if (preempt_count() == cnt)
 117                 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
 118 }
 119 #else /* !CONFIG_TRACE_IRQFLAGS */
 120 static inline void __local_bh_disable(unsigned long ip, unsigned int cnt)
 121 {
 122         add_preempt_count(cnt);
 123         barrier();
 124 }
 125 #endif /* CONFIG_TRACE_IRQFLAGS */
 126
 127 void local_bh_disable(void)
 128 {
 129         __local_bh_disable((unsigned long)__builtin_return_address(0),
 130                                 SOFTIRQ_DISABLE_OFFSET);
 131 }
 132
 133 EXPORT_SYMBOL(local_bh_disable);
 134
 135 static void __local_bh_enable(unsigned int cnt)
 136 {
 137         WARN_ON_ONCE(in_irq());
 138         WARN_ON_ONCE(!irqs_disabled());
 139
 140         if (softirq_count() == cnt)
 141                 trace_softirqs_on((unsigned long)__builtin_return_address(0));
 142         sub_preempt_count(cnt);
 143 }
 144
 145 /*
 146  * Special-case - softirqs can safely be enabled in
 147  * cond_resched_softirq(), or by __do_softirq(),
 148  * without processing still-pending softirqs:
 149  */
 150 void _local_bh_enable(void)
 151 {
 152         __local_bh_enable(SOFTIRQ_DISABLE_OFFSET);
 153 }
 154
 155 EXPORT_SYMBOL(_local_bh_enable);
 156
 157 static inline void _local_bh_enable_ip(unsigned long ip)
 158 {
 159         WARN_ON_ONCE(in_irq() || irqs_disabled());
 160 #ifdef CONFIG_TRACE_IRQFLAGS
 161         local_irq_disable();
 162 #endif
 163         /*
 164          * Are softirqs going to be turned on now:
 165          */
 166         if (softirq_count() == SOFTIRQ_DISABLE_OFFSET)
 167                 trace_softirqs_on(ip);
 168         /*
 169          * Keep preemption disabled until we are done with
 170          * softirq processing:
 171          */
 172         sub_preempt_count(SOFTIRQ_DISABLE_OFFSET - 1);
 173
 174         if (unlikely(!in_interrupt() && local_softirq_pending()))
 175                 do_softirq();
 176
 177         dec_preempt_count();
 178 #ifdef CONFIG_TRACE_IRQFLAGS
 179         local_irq_enable();
 180 #endif
 181         preempt_check_resched();
 182 }
 183
 184 void local_bh_enable(void)
 185 {
 186         _local_bh_enable_ip((unsigned long)__builtin_return_address(0));
 187 }
 188 EXPORT_SYMBOL(local_bh_enable);
 189
 190 void local_bh_enable_ip(unsigned long ip)
 191 {
 192         _local_bh_enable_ip(ip);
 193 }
 194 EXPORT_SYMBOL(local_bh_enable_ip);
 195
 196 /*
 197  * We restart softirq processing MAX_SOFTIRQ_RESTART times,
 198  * and we fall back to softirqd after that.
 199  *
 200  * This number has been established via experimentation.
 201  * The two things to balance is latency against fairness -
 202  * we want to handle softirqs as soon as possible, but they
 203  * should not be able to lock up the box.
 204  */
 205 #define MAX_SOFTIRQ_RESTART 10
 206
 207 asmlinkage void __do_softirq(void)
 208 {
 209         struct softirq_action *h;
 210         __u32 pending;
 211         int max_restart = MAX_SOFTIRQ_RESTART;
 212         int cpu;
 213
 214         pending = local_softirq_pending();
 215         account_system_vtime(current);
 216
 217         __local_bh_disable((unsigned long)__builtin_return_address(0),
 218                                 SOFTIRQ_OFFSET);
 219         lockdep_softirq_enter();
 220
 221         cpu = smp_processor_id();
 222 restart:
 223         /* Reset the pending bitmask before enabling irqs */
 224         set_softirq_pending(0);
 225
 226         local_irq_enable();
 227
 228         h = softirq_vec;
 229
 230         do {
 231                 if (pending & 1) {
 232                         int prev_count = preempt_count();
 233                         kstat_incr_softirqs_this_cpu(h - softirq_vec);
 234
 235                         trace_softirq_entry(h, softirq_vec);
 236                         h->action(h);
 237                         trace_softirq_exit(h, softirq_vec);
 238                         if (unlikely(prev_count != preempt_count())) {
 239                                 printk(KERN_ERR "huh, entered softirq %td %s %p"
 240                                        "with preempt_count %08x,"
 241                                        " exited with %08x?\n", h - softirq_vec,
 242                                        softirq_to_name[h - softirq_vec],
 243                                        h->action, prev_count, preempt_count());
 244                                 preempt_count() = prev_count;
 245                         }
 246
 247                         rcu_bh_qs(cpu);
 248                 }
 249                 h++;
 250                 pending >>= 1;
 251         } while (pending);
 252
 253         local_irq_disable();
 254
 255         pending = local_softirq_pending();
 256         if (pending && --max_restart)
 257                 goto restart;
 258
 259         if (pending)
 260                 wakeup_softirqd();
 261
 262         lockdep_softirq_exit();
 263
 264         account_system_vtime(current);
 265         __local_bh_enable(SOFTIRQ_OFFSET);
 266 }
 267
 268 #ifndef __ARCH_HAS_DO_SOFTIRQ
 269
 270 asmlinkage void do_softirq(void)
 271 {
 272         __u32 pending;
 273         unsigned long flags;
 274
 275         if (in_interrupt())
 276                 return;
 277
 278         local_irq_save(flags);
 279
 280         pending = local_softirq_pending();
 281
 282         if (pending)
 283                 __do_softirq();
 284
 285         local_irq_restore(flags);
 286 }
 287
 288 #endif
 289
 290 /*
 291  * Enter an interrupt context.
 292  */
 293 void irq_enter(void)
 294 {
 295         int cpu = smp_processor_id();
 296
 297         rcu_irq_enter();
 298         if (idle_cpu(cpu) && !in_interrupt()) {
 299                 __irq_enter();
 300                 tick_check_idle(cpu);
 301         } else
 302                 __irq_enter();
 303 }
 304
 305 #ifdef __ARCH_IRQ_EXIT_IRQS_DISABLED
 306 # define invoke_softirq()       __do_softirq()
 307 #else
 308 # define invoke_softirq()       do_softirq()
 309 #endif
 310
 311 /*
 312  * Exit an interrupt context. Process softirqs if needed and possible:
 313  */
 314 void irq_exit(void)
 315 {
 316         account_system_vtime(current);
 317         trace_hardirq_exit();
 318         sub_preempt_count(IRQ_EXIT_OFFSET);
 319         if (!in_interrupt() && local_softirq_pending())
 320                 invoke_softirq();
 321
 322         rcu_irq_exit();
 323 #ifdef CONFIG_NO_HZ
 324         /* Make sure that timer wheel updates are propagated */
 325         if (idle_cpu(smp_processor_id()) && !in_interrupt() && !need_resched())
 326                 tick_nohz_stop_sched_tick(0);
 327 #endif
 328         preempt_enable_no_resched();
 329 }
 330
 331 /*
 332  * This function must run with irqs disabled!
 333  */
 334 inline void raise_softirq_irqoff(unsigned int nr)
 335 {
 336         __raise_softirq_irqoff(nr);
 337
 338         /*
 339          * If we're in an interrupt or softirq, we're done
 340          * (this also catches softirq-disabled code). We will
 341          * actually run the softirq once we return from
 342          * the irq or softirq.
 343          *
 344          * Otherwise we wake up ksoftirqd to make sure we
 345          * schedule the softirq soon.
 346          */
 347         if (!in_interrupt())
 348                 wakeup_softirqd();
 349 }
 350
 351 void raise_softirq(unsigned int nr)
 352 {
 353         unsigned long flags;
 354
 355         local_irq_save(flags);
 356         raise_softirq_irqoff(nr);
 357         local_irq_restore(flags);
 358 }
 359
 360 void open_softirq(int nr, void (*action)(struct softirq_action *))
 361 {
 362         softirq_vec[nr].action = action;
 363 }
 364
 365 /*
 366  * Tasklets
 367  */
 368 struct tasklet_head
 369 {
 370         struct tasklet_struct *head;
 371         struct tasklet_struct **tail;
 372 };
 373
 374 static DEFINE_PER_CPU(struct tasklet_head, tasklet_vec);
 375 static DEFINE_PER_CPU(struct tasklet_head, tasklet_hi_vec);
 376
 377 void __tasklet_schedule(struct tasklet_struct *t)
 378 {
 379         unsigned long flags;
 380
 381         local_irq_save(flags);
 382         t->next = NULL;
 383         *__get_cpu_var(tasklet_vec).tail = t;
 384         __get_cpu_var(tasklet_vec).tail = &(t->next);
 385         raise_softirq_irqoff(TASKLET_SOFTIRQ);
 386         local_irq_restore(flags);
 387 }
 388
 389 EXPORT_SYMBOL(__tasklet_schedule);
 390
 391 void __tasklet_hi_schedule(struct tasklet_struct *t)
 392 {
 393         unsigned long flags;
 394
 395         local_irq_save(flags);
 396         t->next = NULL;
 397         *__get_cpu_var(tasklet_hi_vec).tail = t;
 398         __get_cpu_var(tasklet_hi_vec).tail = &(t->next);
 399         raise_softirq_irqoff(HI_SOFTIRQ);
 400         local_irq_restore(flags);
 401 }
 402
 403 EXPORT_SYMBOL(__tasklet_hi_schedule);
 404
 405 void __tasklet_hi_schedule_first(struct tasklet_struct *t)
 406 {
 407         BUG_ON(!irqs_disabled());
 408
 409         t->next = __get_cpu_var(tasklet_hi_vec).head;
 410         __get_cpu_var(tasklet_hi_vec).head = t;
 411         __raise_softirq_irqoff(HI_SOFTIRQ);
 412 }
 413
 414 EXPORT_SYMBOL(__tasklet_hi_schedule_first);
 415
 416 static void tasklet_action(struct softirq_action *a)
 417 {
 418         struct tasklet_struct *list;
 419
 420         local_irq_disable();
 421         list = __get_cpu_var(tasklet_vec).head;
 422         __get_cpu_var(tasklet_vec).head = NULL;
 423         __get_cpu_var(tasklet_vec).tail = &__get_cpu_var(tasklet_vec).head;
 424         local_irq_enable();
 425
 426         while (list) {
 427                 struct tasklet_struct *t = list;
 428
 429                 list = list->next;
 430
 431                 if (tasklet_trylock(t)) {
 432                         if (!atomic_read(&t->count)) {
 433                                 if (!test_and_clear_bit(TASKLET_STATE_SCHED, &t->state))
 434                                         BUG();
 435                                 t->func(t->data);
 436                                 tasklet_unlock(t);
 437                                 continue;
 438                         }
 439                         tasklet_unlock(t);
 440                 }
 441
 442                 local_irq_disable();
 443                 t->next = NULL;
 444                 *__get_cpu_var(tasklet_vec).tail = t;
 445                 __get_cpu_var(tasklet_vec).tail = &(t->next);
 446                 __raise_softirq_irqoff(TASKLET_SOFTIRQ);
 447                 local_irq_enable();
 448         }
 449 }
 450
 451 static void tasklet_hi_action(struct softirq_action *a)
 452 {
 453         struct tasklet_struct *list;
 454
 455         local_irq_disable();
 456         list = __get_cpu_var(tasklet_hi_vec).head;
 457         __get_cpu_var(tasklet_hi_vec).head = NULL;
 458         __get_cpu_var(tasklet_hi_vec).tail = &__get_cpu_var(tasklet_hi_vec).head;
 459         local_irq_enable();
 460
 461         while (list) {
 462                 struct tasklet_struct *t = list;
 463
 464                 list = list->next;
 465
 466                 if (tasklet_trylock(t)) {
 467                         if (!atomic_read(&t->count)) {
 468                                 if (!test_and_clear_bit(TASKLET_STATE_SCHED, &t->state))
 469                                         BUG();
 470                                 t->func(t->data);
 471                                 tasklet_unlock(t);
 472                                 continue;
 473                         }
 474                         tasklet_unlock(t);
 475                 }
 476
 477                 local_irq_disable();
 478                 t->next = NULL;
 479                 *__get_cpu_var(tasklet_hi_vec).tail = t;
 480                 __get_cpu_var(tasklet_hi_vec).tail = &(t->next);
 481                 __raise_softirq_irqoff(HI_SOFTIRQ);
 482                 local_irq_enable();
 483         }
 484 }
 485
 486
 487 void tasklet_init(struct tasklet_struct *t,
 488                   void (*func)(unsigned long), unsigned long data)
 489 {
 490         t->next = NULL;
 491         t->state = 0;
 492         atomic_set(&t->count, 0);
 493         t->func = func;
 494         t->data = data;
 495 }
 496
 497 EXPORT_SYMBOL(tasklet_init);
 498
 499 void tasklet_kill(struct tasklet_struct *t)
 500 {
 501         if (in_interrupt())
 502                 printk("Attempt to kill tasklet from interrupt\n");
 503
 504         while (test_and_set_bit(TASKLET_STATE_SCHED, &t->state)) {
 505                 do {
 506                         yield();
 507                 } while (test_bit(TASKLET_STATE_SCHED, &t->state));
 508         }
 509         tasklet_unlock_wait(t);
 510         clear_bit(TASKLET_STATE_SCHED, &t->state);
 511 }
 512
 513 EXPORT_SYMBOL(tasklet_kill);
 514
 515 /*
 516  * tasklet_hrtimer
 517  */
 518
 519 /*
 520  * The trampoline is called when the hrtimer expires. It schedules a tasklet
 521  * to run __tasklet_hrtimer_trampoline() which in turn will call the intended
 522  * hrtimer callback, but from softirq context.
 523  */
 524 static enum hrtimer_restart __hrtimer_tasklet_trampoline(struct hrtimer *timer)
 525 {
 526         struct tasklet_hrtimer *ttimer =
 527                 container_of(timer, struct tasklet_hrtimer, timer);
 528
 529         tasklet_hi_schedule(&ttimer->tasklet);
 530         return HRTIMER_NORESTART;
 531 }
 532
 533 /*
 534  * Helper function which calls the hrtimer callback from
 535  * tasklet/softirq context
 536  */
 537 static void __tasklet_hrtimer_trampoline(unsigned long data)
 538 {
 539         struct tasklet_hrtimer *ttimer = (void *)data;
 540         enum hrtimer_restart restart;
 541
 542         restart = ttimer->function(&ttimer->timer);
 543         if (restart != HRTIMER_NORESTART)
 544                 hrtimer_restart(&ttimer->timer);
 545 }
 546
 547 /**
 548  * tasklet_hrtimer_init - Init a tasklet/hrtimer combo for softirq callbacks
 549  * @ttimer:      tasklet_hrtimer which is initialized
 550  * @function:    hrtimer callback funtion which gets called from softirq context
 551  * @which_clock: clock id (CLOCK_MONOTONIC/CLOCK_REALTIME)
 552  * @mode:        hrtimer mode (HRTIMER_MODE_ABS/HRTIMER_MODE_REL)
 553  */
 554 void tasklet_hrtimer_init(struct tasklet_hrtimer *ttimer,
 555                           enum hrtimer_restart (*function)(struct hrtimer *),
 556                           clockid_t which_clock, enum hrtimer_mode mode)
 557 {
 558         hrtimer_init(&ttimer->timer, which_clock, mode);
 559         ttimer->timer.function = __hrtimer_tasklet_trampoline;
 560         tasklet_init(&ttimer->tasklet, __tasklet_hrtimer_trampoline,
 561                      (unsigned long)ttimer);
 562         ttimer->function = function;
 563 }
 564 EXPORT_SYMBOL_GPL(tasklet_hrtimer_init);
 565
 566 /*
 567  * Remote softirq bits
 568  */
 569
 570 DEFINE_PER_CPU(struct list_head [NR_SOFTIRQS], softirq_work_list);
 571 EXPORT_PER_CPU_SYMBOL(softirq_work_list);
 572
 573 static void __local_trigger(struct call_single_data *cp, int softirq)
 574 {
 575         struct list_head *head = &__get_cpu_var(softirq_work_list[softirq]);
 576
 577         list_add_tail(&cp->list, head);
 578
 579         /* Trigger the softirq only if the list was previously empty.  */
 580         if (head->next == &cp->list)
 581                 raise_softirq_irqoff(softirq);
 582 }
 583
 584 #ifdef CONFIG_USE_GENERIC_SMP_HELPERS
 585 static void remote_softirq_receive(void *data)
 586 {
 587         struct call_single_data *cp = data;
 588         unsigned long flags;
 589         int softirq;
 590
 591         softirq = cp->priv;
 592
 593         local_irq_save(flags);
 594         __local_trigger(cp, softirq);
 595         local_irq_restore(flags);
 596 }
 597
 598 static int __try_remote_softirq(struct call_single_data *cp, int cpu, int softirq)
 599 {
 600         if (cpu_online(cpu)) {
 601                 cp->func = remote_softirq_receive;
 602                 cp->info = cp;
 603                 cp->flags = 0;
 604                 cp->priv = softirq;
 605
 606                 __smp_call_function_single(cpu, cp, 0);
 607                 return 0;
 608         }
 609         return 1;
 610 }
 611 #else /* CONFIG_USE_GENERIC_SMP_HELPERS */
 612 static int __try_remote_softirq(struct call_single_data *cp, int cpu, int softirq)
 613 {
 614         return 1;
 615 }
 616 #endif
 617
 618 /**
 619  * __send_remote_softirq - try to schedule softirq work on a remote cpu
 620  * @cp: private SMP call function data area
 621  * @cpu: the remote cpu
 622  * @this_cpu: the currently executing cpu
 623  * @softirq: the softirq for the work
 624  *
 625  * Attempt to schedule softirq work on a remote cpu.  If this cannot be
 626  * done, the work is instead queued up on the local cpu.
 627  *
 628  * Interrupts must be disabled.
 629  */
 630 void __send_remote_softirq(struct call_single_data *cp, int cpu, int this_cpu, int softirq)
 631 {
 632         if (cpu == this_cpu || __try_remote_softirq(cp, cpu, softirq))
 633                 __local_trigger(cp, softirq);
 634 }
 635 EXPORT_SYMBOL(__send_remote_softirq);
 636
 637 /**
 638  * send_remote_softirq - try to schedule softirq work on a remote cpu
 639  * @cp: private SMP call function data area
 640  * @cpu: the remote cpu
 641  * @softirq: the softirq for the work
 642  *
 643  * Like __send_remote_softirq except that disabling interrupts and
 644  * computing the current cpu is done for the caller.
 645  */
 646 void send_remote_softirq(struct call_single_data *cp, int cpu, int softirq)
 647 {
 648         unsigned long flags;
 649         int this_cpu;
 650
 651         local_irq_save(flags);
 652         this_cpu = smp_processor_id();
 653         __send_remote_softirq(cp, cpu, this_cpu, softirq);
 654         local_irq_restore(flags);
 655 }
 656 EXPORT_SYMBOL(send_remote_softirq);
 657
 658 static int __cpuinit remote_softirq_cpu_notify(struct notifier_block *self,
 659                                                unsigned long action, void *hcpu)
 660 {
 661         /*
 662          * If a CPU goes away, splice its entries to the current CPU
 663          * and trigger a run of the softirq
 664          */
 665         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
 666                 int cpu = (unsigned long) hcpu;
 667                 int i;
 668
 669                 local_irq_disable();
 670                 for (i = 0; i < NR_SOFTIRQS; i++) {
 671                         struct list_head *head = &per_cpu(softirq_work_list[i], cpu);
 672                         struct list_head *local_head;
 673
 674                         if (list_empty(head))
 675                                 continue;
 676
 677                         local_head = &__get_cpu_var(softirq_work_list[i]);
 678                         list_splice_init(head, local_head);
 679                         raise_softirq_irqoff(i);
 680                 }
 681                 local_irq_enable();
 682         }
 683
 684         return NOTIFY_OK;
 685 }
 686
 687 static struct notifier_block __cpuinitdata remote_softirq_cpu_notifier = {
 688         .notifier_call  = remote_softirq_cpu_notify,
 689 };
 690
 691 void __init softirq_init(void)
 692 {
 693         int cpu;
 694
 695         for_each_possible_cpu(cpu) {
 696                 int i;
 697
 698                 per_cpu(tasklet_vec, cpu).tail =
 699                         &per_cpu(tasklet_vec, cpu).head;
 700                 per_cpu(tasklet_hi_vec, cpu).tail =
 701                         &per_cpu(tasklet_hi_vec, cpu).head;
 702                 for (i = 0; i < NR_SOFTIRQS; i++)
 703                         INIT_LIST_HEAD(&per_cpu(softirq_work_list[i], cpu));
 704         }
 705
 706         register_hotcpu_notifier(&remote_softirq_cpu_notifier);
 707
 708         open_softirq(TASKLET_SOFTIRQ, tasklet_action);
 709         open_softirq(HI_SOFTIRQ, tasklet_hi_action);
 710 }
 711
 712 static int run_ksoftirqd(void * __bind_cpu)
 713 {
 714         set_current_state(TASK_INTERRUPTIBLE);
 715
 716         current->flags |= PF_KSOFTIRQD;
 717         while (!kthread_should_stop()) {
 718                 preempt_disable();
 719                 if (!local_softirq_pending()) {
 720                         preempt_enable_no_resched();
 721                         schedule();
 722                         preempt_disable();
 723                 }
 724
 725                 __set_current_state(TASK_RUNNING);
 726
 727                 while (local_softirq_pending()) {
 728                         /* Preempt disable stops cpu going offline.
 729                            If already offline, we'll be on wrong CPU:
 730                            don't process */
 731                         if (cpu_is_offline((long)__bind_cpu))
 732                                 goto wait_to_die;
 733                         do_softirq();
 734                         preempt_enable_no_resched();
 735                         cond_resched();
 736                         preempt_disable();
 737                         rcu_note_context_switch((long)__bind_cpu);
 738                 }
 739                 preempt_enable();
 740                 set_current_state(TASK_INTERRUPTIBLE);
 741         }
 742         __set_current_state(TASK_RUNNING);
 743         return 0;
 744
 745 wait_to_die:
 746         preempt_enable();
 747         /* Wait for kthread_stop */
 748         set_current_state(TASK_INTERRUPTIBLE);
 749         while (!kthread_should_stop()) {
 750                 schedule();
 751                 set_current_state(TASK_INTERRUPTIBLE);
 752         }
 753         __set_current_state(TASK_RUNNING);
 754         return 0;
 755 }
 756
 757 #ifdef CONFIG_HOTPLUG_CPU
 758 /*
 759  * tasklet_kill_immediate is called to remove a tasklet which can already be
 760  * scheduled for execution on @cpu.
 761  *
 762  * Unlike tasklet_kill, this function removes the tasklet
 763  * _immediately_, even if the tasklet is in TASKLET_STATE_SCHED state.
 764  *
 765  * When this function is called, @cpu must be in the CPU_DEAD state.
 766  */
 767 void tasklet_kill_immediate(struct tasklet_struct *t, unsigned int cpu)
 768 {
 769         struct tasklet_struct **i;
 770
 771         BUG_ON(cpu_online(cpu));
 772         BUG_ON(test_bit(TASKLET_STATE_RUN, &t->state));
 773
 774         if (!test_bit(TASKLET_STATE_SCHED, &t->state))
 775                 return;
 776
 777         /* CPU is dead, so no lock needed. */
 778         for (i = &per_cpu(tasklet_vec, cpu).head; *i; i = &(*i)->next) {
 779                 if (*i == t) {
 780                         *i = t->next;
 781                         /* If this was the tail element, move the tail ptr */
 782                         if (*i == NULL)
 783                                 per_cpu(tasklet_vec, cpu).tail = i;
 784                         return;
 785                 }
 786         }
 787         BUG();
 788 }
 789
 790 static void takeover_tasklets(unsigned int cpu)
 791 {
 792         /* CPU is dead, so no lock needed. */
 793         local_irq_disable();
 794
 795         /* Find end, append list for that CPU. */
 796         if (&per_cpu(tasklet_vec, cpu).head != per_cpu(tasklet_vec, cpu).tail) {
 797                 *(__get_cpu_var(tasklet_vec).tail) = per_cpu(tasklet_vec, cpu).head;
 798                 __get_cpu_var(tasklet_vec).tail = per_cpu(tasklet_vec, cpu).tail;
 799                 per_cpu(tasklet_vec, cpu).head = NULL;
 800                 per_cpu(tasklet_vec, cpu).tail = &per_cpu(tasklet_vec, cpu).head;
 801         }
 802         raise_softirq_irqoff(TASKLET_SOFTIRQ);
 803
 804         if (&per_cpu(tasklet_hi_vec, cpu).head != per_cpu(tasklet_hi_vec, cpu).tail) {
 805                 *__get_cpu_var(tasklet_hi_vec).tail = per_cpu(tasklet_hi_vec, cpu).head;
 806                 __get_cpu_var(tasklet_hi_vec).tail = per_cpu(tasklet_hi_vec, cpu).tail;
 807                 per_cpu(tasklet_hi_vec, cpu).head = NULL;
 808                 per_cpu(tasklet_hi_vec, cpu).tail = &per_cpu(tasklet_hi_vec, cpu).head;
 809         }
 810         raise_softirq_irqoff(HI_SOFTIRQ);
 811
 812         local_irq_enable();
 813 }
 814 #endif /* CONFIG_HOTPLUG_CPU */
 815
 816 static int __cpuinit cpu_callback(struct notifier_block *nfb,
 817                                   unsigned long action,
 818                                   void *hcpu)
 819 {
 820         int hotcpu = (unsigned long)hcpu;
 821         struct task_struct *p;
 822
 823         switch (action) {
 824         case CPU_UP_PREPARE:
 825         case CPU_UP_PREPARE_FROZEN:
 826                 p = kthread_create(run_ksoftirqd, hcpu, "ksoftirqd/%d", hotcpu);
 827                 if (IS_ERR(p)) {
 828                         printk("ksoftirqd for %i failed\n", hotcpu);
 829                         return notifier_from_errno(PTR_ERR(p));
 830                 }
 831                 kthread_bind(p, hotcpu);
 832                 per_cpu(ksoftirqd, hotcpu) = p;
 833                 break;
 834         case CPU_ONLINE:
 835         case CPU_ONLINE_FROZEN:
 836                 wake_up_process(per_cpu(ksoftirqd, hotcpu));
 837                 break;
 838 #ifdef CONFIG_HOTPLUG_CPU
 839         case CPU_UP_CANCELED:
 840         case CPU_UP_CANCELED_FROZEN:
 841                 if (!per_cpu(ksoftirqd, hotcpu))
 842                         break;
 843                 /* Unbind so it can run.  Fall thru. */
 844                 kthread_bind(per_cpu(ksoftirqd, hotcpu),
 845                              cpumask_any(cpu_online_mask));
 846         case CPU_DEAD:
 847         case CPU_DEAD_FROZEN: {
 848                 struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
 849
 850                 p = per_cpu(ksoftirqd, hotcpu);
 851                 per_cpu(ksoftirqd, hotcpu) = NULL;
 852                 sched_setscheduler_nocheck(p, SCHED_FIFO, &param);
 853                 kthread_stop(p);
 854                 takeover_tasklets(hotcpu);
 855                 break;
 856         }
 857 #endif /* CONFIG_HOTPLUG_CPU */
 858         }
 859         return NOTIFY_OK;
 860 }
 861
 862 static struct notifier_block __cpuinitdata cpu_nfb = {
 863         .notifier_call = cpu_callback
 864 };
 865
 866 static __init int spawn_ksoftirqd(void)
 867 {
 868         void *cpu = (void *)(long)smp_processor_id();
 869         int err = cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);
 870
 871         BUG_ON(err != NOTIFY_OK);
 872         cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);
 873         register_cpu_notifier(&cpu_nfb);
 874         return 0;
 875 }
 876 early_initcall(spawn_ksoftirqd);
 877
 878 #ifdef CONFIG_SMP
 879 /*
 880  * Call a function on all processors
 881  */
 882 int on_each_cpu(void (*func) (void *info), void *info, int wait)
 883 {
 884         int ret = 0;
 885
 886         preempt_disable();
 887         ret = smp_call_function(func, info, wait);
 888         local_irq_disable();
 889         func(info);
 890         local_irq_enable();
 891         preempt_enable();
 892         return ret;
 893 }
 894 EXPORT_SYMBOL(on_each_cpu);
 895 #endif
 896
 897 /*
 898  * [ These __weak aliases are kept in a separate compilation unit, so that
 899  *   GCC does not inline them incorrectly. ]
 900  */
 901
 902 int __init __weak early_irq_init(void)
 903 {
 904         return 0;
 905 }
 906
 907 int __init __weak arch_probe_nr_irqs(void)
 908 {
 909         return 0;
 910 }
 911
 912 int __init __weak arch_early_irq_init(void)
 913 {
 914         return 0;
 915 }
 916
 917 int __weak arch_init_chip_data(struct irq_desc *desc, int node)
 918 {
 919         return 0;
 920 }