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 #include <linux/module.h>
  12 #include <linux/kernel_stat.h>
  13 #include <linux/interrupt.h>
  14 #include <linux/init.h>
  15 #include <linux/mm.h>
  16 #include <linux/notifier.h>
  17 #include <linux/percpu.h>
  18 #include <linux/cpu.h>
  19 #include <linux/freezer.h>
  20 #include <linux/kthread.h>
  21 #include <linux/rcupdate.h>
  22 #include <linux/smp.h>
  23 #include <linux/tick.h>
  24
  25 #include <asm/irq.h>
  26 /*
  27    - No shared variables, all the data are CPU local.
  28    - If a softirq needs serialization, let it serialize itself
  29      by its own spinlocks.
  30    - Even if softirq is serialized, only local cpu is marked for
  31      execution. Hence, we get something sort of weak cpu binding.
  32      Though it is still not clear, will it result in better locality
  33      or will not.
  34
  35    Examples:
  36    - NET RX softirq. It is multithreaded and does not require
  37      any global serialization.
  38    - NET TX softirq. It kicks software netdevice queues, hence
  39      it is logically serialized per device, but this serialization
  40      is invisible to common code.
  41    - Tasklets: serialized wrt itself.
  42  */
  43
  44 #ifndef __ARCH_IRQ_STAT
  45 irq_cpustat_t irq_stat[NR_CPUS] ____cacheline_aligned;
  46 EXPORT_SYMBOL(irq_stat);
  47 #endif
  48
  49 static struct softirq_action softirq_vec[32] __cacheline_aligned_in_smp;
  50
  51 static DEFINE_PER_CPU(struct task_struct *, ksoftirqd);
  52
  53 /*
  54  * we cannot loop indefinitely here to avoid userspace starvation,
  55  * but we also don't want to introduce a worst case 1/HZ latency
  56  * to the pending events, so lets the scheduler to balance
  57  * the softirq load for us.
  58  */
  59 static inline void wakeup_softirqd(void)
  60 {
  61         /* Interrupts are disabled: no need to stop preemption */
  62         struct task_struct *tsk = __get_cpu_var(ksoftirqd);
  63
  64         if (tsk && tsk->state != TASK_RUNNING)
  65                 wake_up_process(tsk);
  66 }
  67
  68 /*
  69  * This one is for softirq.c-internal use,
  70  * where hardirqs are disabled legitimately:
  71  */
  72 #ifdef CONFIG_TRACE_IRQFLAGS
  73 static void __local_bh_disable(unsigned long ip)
  74 {
  75         unsigned long flags;
  76
  77         WARN_ON_ONCE(in_irq());
  78
  79         raw_local_irq_save(flags);
  80         add_preempt_count(SOFTIRQ_OFFSET);
  81         /*
  82          * Were softirqs turned off above:
  83          */
  84         if (softirq_count() == SOFTIRQ_OFFSET)
  85                 trace_softirqs_off(ip);
  86         raw_local_irq_restore(flags);
  87 }
  88 #else /* !CONFIG_TRACE_IRQFLAGS */
  89 static inline void __local_bh_disable(unsigned long ip)
  90 {
  91         add_preempt_count(SOFTIRQ_OFFSET);
  92         barrier();
  93 }
  94 #endif /* CONFIG_TRACE_IRQFLAGS */
  95
  96 void local_bh_disable(void)
  97 {
  98         __local_bh_disable((unsigned long)__builtin_return_address(0));
  99 }
 100
 101 EXPORT_SYMBOL(local_bh_disable);
 102
 103 void __local_bh_enable(void)
 104 {
 105         WARN_ON_ONCE(in_irq());
 106
 107         /*
 108          * softirqs should never be enabled by __local_bh_enable(),
 109          * it always nests inside local_bh_enable() sections:
 110          */
 111         WARN_ON_ONCE(softirq_count() == SOFTIRQ_OFFSET);
 112
 113         sub_preempt_count(SOFTIRQ_OFFSET);
 114 }
 115 EXPORT_SYMBOL_GPL(__local_bh_enable);
 116
 117 /*
 118  * Special-case - softirqs can safely be enabled in
 119  * cond_resched_softirq(), or by __do_softirq(),
 120  * without processing still-pending softirqs:
 121  */
 122 void _local_bh_enable(void)
 123 {
 124         WARN_ON_ONCE(in_irq());
 125         WARN_ON_ONCE(!irqs_disabled());
 126
 127         if (softirq_count() == SOFTIRQ_OFFSET)
 128                 trace_softirqs_on((unsigned long)__builtin_return_address(0));
 129         sub_preempt_count(SOFTIRQ_OFFSET);
 130 }
 131
 132 EXPORT_SYMBOL(_local_bh_enable);
 133
 134 void local_bh_enable(void)
 135 {
 136 #ifdef CONFIG_TRACE_IRQFLAGS
 137         unsigned long flags;
 138
 139         WARN_ON_ONCE(in_irq());
 140 #endif
 141         WARN_ON_ONCE(irqs_disabled());
 142
 143 #ifdef CONFIG_TRACE_IRQFLAGS
 144         local_irq_save(flags);
 145 #endif
 146         /*
 147          * Are softirqs going to be turned on now:
 148          */
 149         if (softirq_count() == SOFTIRQ_OFFSET)
 150                 trace_softirqs_on((unsigned long)__builtin_return_address(0));
 151         /*
 152          * Keep preemption disabled until we are done with
 153          * softirq processing:
 154          */
 155         sub_preempt_count(SOFTIRQ_OFFSET - 1);
 156
 157         if (unlikely(!in_interrupt() && local_softirq_pending()))
 158                 do_softirq();
 159
 160         dec_preempt_count();
 161 #ifdef CONFIG_TRACE_IRQFLAGS
 162         local_irq_restore(flags);
 163 #endif
 164         preempt_check_resched();
 165 }
 166 EXPORT_SYMBOL(local_bh_enable);
 167
 168 void local_bh_enable_ip(unsigned long ip)
 169 {
 170 #ifdef CONFIG_TRACE_IRQFLAGS
 171         unsigned long flags;
 172
 173         WARN_ON_ONCE(in_irq());
 174
 175         local_irq_save(flags);
 176 #endif
 177         /*
 178          * Are softirqs going to be turned on now:
 179          */
 180         if (softirq_count() == SOFTIRQ_OFFSET)
 181                 trace_softirqs_on(ip);
 182         /*
 183          * Keep preemption disabled until we are done with
 184          * softirq processing:
 185          */
 186         sub_preempt_count(SOFTIRQ_OFFSET - 1);
 187
 188         if (unlikely(!in_interrupt() && local_softirq_pending()))
 189                 do_softirq();
 190
 191         dec_preempt_count();
 192 #ifdef CONFIG_TRACE_IRQFLAGS
 193         local_irq_restore(flags);
 194 #endif
 195         preempt_check_resched();
 196 }
 197 EXPORT_SYMBOL(local_bh_enable_ip);
 198
 199 /*
 200  * We restart softirq processing MAX_SOFTIRQ_RESTART times,
 201  * and we fall back to softirqd after that.
 202  *
 203  * This number has been established via experimentation.
 204  * The two things to balance is latency against fairness -
 205  * we want to handle softirqs as soon as possible, but they
 206  * should not be able to lock up the box.
 207  */
 208 #define MAX_SOFTIRQ_RESTART 10
 209
 210 asmlinkage void __do_softirq(void)
 211 {
 212         struct softirq_action *h;
 213         __u32 pending;
 214         int max_restart = MAX_SOFTIRQ_RESTART;
 215         int cpu;
 216
 217         pending = local_softirq_pending();
 218         account_system_vtime(current);
 219
 220         __local_bh_disable((unsigned long)__builtin_return_address(0));
 221         trace_softirq_enter();
 222
 223         cpu = smp_processor_id();
 224 restart:
 225         /* Reset the pending bitmask before enabling irqs */
 226         set_softirq_pending(0);
 227
 228         local_irq_enable();
 229
 230         h = softirq_vec;
 231
 232         do {
 233                 if (pending & 1) {
 234                         h->action(h);
 235                         rcu_bh_qsctr_inc(cpu);
 236                 }
 237                 h++;
 238                 pending >>= 1;
 239         } while (pending);
 240
 241         local_irq_disable();
 242
 243         pending = local_softirq_pending();
 244         if (pending && --max_restart)
 245                 goto restart;
 246
 247         if (pending)
 248                 wakeup_softirqd();
 249
 250         trace_softirq_exit();
 251
 252         account_system_vtime(current);
 253         _local_bh_enable();
 254 }
 255
 256 #ifndef __ARCH_HAS_DO_SOFTIRQ
 257
 258 asmlinkage void do_softirq(void)
 259 {
 260         __u32 pending;
 261         unsigned long flags;
 262
 263         if (in_interrupt())
 264                 return;
 265
 266         local_irq_save(flags);
 267
 268         pending = local_softirq_pending();
 269
 270         if (pending)
 271                 __do_softirq();
 272
 273         local_irq_restore(flags);
 274 }
 275
 276 #endif
 277
 278 /*
 279  * Enter an interrupt context.
 280  */
 281 void irq_enter(void)
 282 {
 283 #ifdef CONFIG_NO_HZ
 284         int cpu = smp_processor_id();
 285         if (idle_cpu(cpu) && !in_interrupt())
 286                 tick_nohz_stop_idle(cpu);
 287 #endif
 288         __irq_enter();
 289 #ifdef CONFIG_NO_HZ
 290         if (idle_cpu(cpu))
 291                 tick_nohz_update_jiffies();
 292 #endif
 293 }
 294
 295 #ifdef __ARCH_IRQ_EXIT_IRQS_DISABLED
 296 # define invoke_softirq()       __do_softirq()
 297 #else
 298 # define invoke_softirq()       do_softirq()
 299 #endif
 300
 301 /*
 302  * Exit an interrupt context. Process softirqs if needed and possible:
 303  */
 304 void irq_exit(void)
 305 {
 306         account_system_vtime(current);
 307         trace_hardirq_exit();
 308         sub_preempt_count(IRQ_EXIT_OFFSET);
 309         if (!in_interrupt() && local_softirq_pending())
 310                 invoke_softirq();
 311
 312 #ifdef CONFIG_NO_HZ
 313         /* Make sure that timer wheel updates are propagated */
 314         if (!in_interrupt() && idle_cpu(smp_processor_id()) && !need_resched())
 315                 tick_nohz_stop_sched_tick();
 316 <<<<<<< HEAD:kernel/softirq.c
 317 =======
 318         rcu_irq_exit();
 319 >>>>>>> 264e3e889d86e552b4191d69bb60f4f3b383135a:kernel/softirq.c
 320 #endif
 321         preempt_enable_no_resched();
 322 }
 323
 324 /*
 325  * This function must run with irqs disabled!
 326  */
 327 inline void raise_softirq_irqoff(unsigned int nr)
 328 {
 329         __raise_softirq_irqoff(nr);
 330
 331         /*
 332          * If we're in an interrupt or softirq, we're done
 333          * (this also catches softirq-disabled code). We will
 334          * actually run the softirq once we return from
 335          * the irq or softirq.
 336          *
 337          * Otherwise we wake up ksoftirqd to make sure we
 338          * schedule the softirq soon.
 339          */
 340         if (!in_interrupt())
 341                 wakeup_softirqd();
 342 }
 343
 344 void raise_softirq(unsigned int nr)
 345 {
 346         unsigned long flags;
 347
 348         local_irq_save(flags);
 349         raise_softirq_irqoff(nr);
 350         local_irq_restore(flags);
 351 }
 352
 353 void open_softirq(int nr, void (*action)(struct softirq_action*), void *data)
 354 {
 355         softirq_vec[nr].data = data;
 356         softirq_vec[nr].action = action;
 357 }
 358
 359 /* Tasklets */
 360 struct tasklet_head
 361 {
 362         struct tasklet_struct *list;
 363 };
 364
 365 /* Some compilers disobey section attribute on statics when not
 366    initialized -- RR */
 367 static DEFINE_PER_CPU(struct tasklet_head, tasklet_vec) = { NULL };
 368 static DEFINE_PER_CPU(struct tasklet_head, tasklet_hi_vec) = { NULL };
 369
 370 void __tasklet_schedule(struct tasklet_struct *t)
 371 {
 372         unsigned long flags;
 373
 374         local_irq_save(flags);
 375         t->next = __get_cpu_var(tasklet_vec).list;
 376         __get_cpu_var(tasklet_vec).list = t;
 377         raise_softirq_irqoff(TASKLET_SOFTIRQ);
 378         local_irq_restore(flags);
 379 }
 380
 381 EXPORT_SYMBOL(__tasklet_schedule);
 382
 383 void __tasklet_hi_schedule(struct tasklet_struct *t)
 384 {
 385         unsigned long flags;
 386
 387         local_irq_save(flags);
 388         t->next = __get_cpu_var(tasklet_hi_vec).list;
 389         __get_cpu_var(tasklet_hi_vec).list = t;
 390         raise_softirq_irqoff(HI_SOFTIRQ);
 391         local_irq_restore(flags);
 392 }
 393
 394 EXPORT_SYMBOL(__tasklet_hi_schedule);
 395
 396 static void tasklet_action(struct softirq_action *a)
 397 {
 398         struct tasklet_struct *list;
 399
 400         local_irq_disable();
 401         list = __get_cpu_var(tasklet_vec).list;
 402         __get_cpu_var(tasklet_vec).list = NULL;
 403         local_irq_enable();
 404
 405         while (list) {
 406                 struct tasklet_struct *t = list;
 407
 408                 list = list->next;
 409
 410                 if (tasklet_trylock(t)) {
 411                         if (!atomic_read(&t->count)) {
 412                                 if (!test_and_clear_bit(TASKLET_STATE_SCHED, &t->state))
 413                                         BUG();
 414                                 t->func(t->data);
 415                                 tasklet_unlock(t);
 416                                 continue;
 417                         }
 418                         tasklet_unlock(t);
 419                 }
 420
 421                 local_irq_disable();
 422                 t->next = __get_cpu_var(tasklet_vec).list;
 423                 __get_cpu_var(tasklet_vec).list = t;
 424                 __raise_softirq_irqoff(TASKLET_SOFTIRQ);
 425                 local_irq_enable();
 426         }
 427 }
 428
 429 static void tasklet_hi_action(struct softirq_action *a)
 430 {
 431         struct tasklet_struct *list;
 432
 433         local_irq_disable();
 434         list = __get_cpu_var(tasklet_hi_vec).list;
 435         __get_cpu_var(tasklet_hi_vec).list = NULL;
 436         local_irq_enable();
 437
 438         while (list) {
 439                 struct tasklet_struct *t = list;
 440
 441                 list = list->next;
 442
 443                 if (tasklet_trylock(t)) {
 444                         if (!atomic_read(&t->count)) {
 445                                 if (!test_and_clear_bit(TASKLET_STATE_SCHED, &t->state))
 446                                         BUG();
 447                                 t->func(t->data);
 448                                 tasklet_unlock(t);
 449                                 continue;
 450                         }
 451                         tasklet_unlock(t);
 452                 }
 453
 454                 local_irq_disable();
 455                 t->next = __get_cpu_var(tasklet_hi_vec).list;
 456                 __get_cpu_var(tasklet_hi_vec).list = t;
 457                 __raise_softirq_irqoff(HI_SOFTIRQ);
 458                 local_irq_enable();
 459         }
 460 }
 461
 462
 463 void tasklet_init(struct tasklet_struct *t,
 464                   void (*func)(unsigned long), unsigned long data)
 465 {
 466         t->next = NULL;
 467         t->state = 0;
 468         atomic_set(&t->count, 0);
 469         t->func = func;
 470         t->data = data;
 471 }
 472
 473 EXPORT_SYMBOL(tasklet_init);
 474
 475 void tasklet_kill(struct tasklet_struct *t)
 476 {
 477         if (in_interrupt())
 478                 printk("Attempt to kill tasklet from interrupt\n");
 479
 480         while (test_and_set_bit(TASKLET_STATE_SCHED, &t->state)) {
 481                 do
 482                         yield();
 483                 while (test_bit(TASKLET_STATE_SCHED, &t->state));
 484         }
 485         tasklet_unlock_wait(t);
 486         clear_bit(TASKLET_STATE_SCHED, &t->state);
 487 }
 488
 489 EXPORT_SYMBOL(tasklet_kill);
 490
 491 void __init softirq_init(void)
 492 {
 493         open_softirq(TASKLET_SOFTIRQ, tasklet_action, NULL);
 494         open_softirq(HI_SOFTIRQ, tasklet_hi_action, NULL);
 495 }
 496
 497 static int ksoftirqd(void * __bind_cpu)
 498 {
 499         set_current_state(TASK_INTERRUPTIBLE);
 500
 501         while (!kthread_should_stop()) {
 502                 preempt_disable();
 503                 if (!local_softirq_pending()) {
 504                         preempt_enable_no_resched();
 505                         schedule();
 506                         preempt_disable();
 507                 }
 508
 509                 __set_current_state(TASK_RUNNING);
 510
 511                 while (local_softirq_pending()) {
 512                         /* Preempt disable stops cpu going offline.
 513                            If already offline, we'll be on wrong CPU:
 514                            don't process */
 515                         if (cpu_is_offline((long)__bind_cpu))
 516                                 goto wait_to_die;
 517                         do_softirq();
 518                         preempt_enable_no_resched();
 519                         cond_resched();
 520                         preempt_disable();
 521                 }
 522                 preempt_enable();
 523                 set_current_state(TASK_INTERRUPTIBLE);
 524         }
 525         __set_current_state(TASK_RUNNING);
 526         return 0;
 527
 528 wait_to_die:
 529         preempt_enable();
 530         /* Wait for kthread_stop */
 531         set_current_state(TASK_INTERRUPTIBLE);
 532         while (!kthread_should_stop()) {
 533                 schedule();
 534                 set_current_state(TASK_INTERRUPTIBLE);
 535         }
 536         __set_current_state(TASK_RUNNING);
 537         return 0;
 538 }
 539
 540 #ifdef CONFIG_HOTPLUG_CPU
 541 /*
 542  * tasklet_kill_immediate is called to remove a tasklet which can already be
 543  * scheduled for execution on @cpu.
 544  *
 545  * Unlike tasklet_kill, this function removes the tasklet
 546  * _immediately_, even if the tasklet is in TASKLET_STATE_SCHED state.
 547  *
 548  * When this function is called, @cpu must be in the CPU_DEAD state.
 549  */
 550 void tasklet_kill_immediate(struct tasklet_struct *t, unsigned int cpu)
 551 {
 552         struct tasklet_struct **i;
 553
 554         BUG_ON(cpu_online(cpu));
 555         BUG_ON(test_bit(TASKLET_STATE_RUN, &t->state));
 556
 557         if (!test_bit(TASKLET_STATE_SCHED, &t->state))
 558                 return;
 559
 560         /* CPU is dead, so no lock needed. */
 561         for (i = &per_cpu(tasklet_vec, cpu).list; *i; i = &(*i)->next) {
 562                 if (*i == t) {
 563                         *i = t->next;
 564                         return;
 565                 }
 566         }
 567         BUG();
 568 }
 569
 570 static void takeover_tasklets(unsigned int cpu)
 571 {
 572         struct tasklet_struct **i;
 573
 574         /* CPU is dead, so no lock needed. */
 575         local_irq_disable();
 576
 577         /* Find end, append list for that CPU. */
 578         for (i = &__get_cpu_var(tasklet_vec).list; *i; i = &(*i)->next);
 579         *i = per_cpu(tasklet_vec, cpu).list;
 580         per_cpu(tasklet_vec, cpu).list = NULL;
 581         raise_softirq_irqoff(TASKLET_SOFTIRQ);
 582
 583         for (i = &__get_cpu_var(tasklet_hi_vec).list; *i; i = &(*i)->next);
 584         *i = per_cpu(tasklet_hi_vec, cpu).list;
 585         per_cpu(tasklet_hi_vec, cpu).list = NULL;
 586         raise_softirq_irqoff(HI_SOFTIRQ);
 587
 588         local_irq_enable();
 589 }
 590 #endif /* CONFIG_HOTPLUG_CPU */
 591
 592 static int __cpuinit cpu_callback(struct notifier_block *nfb,
 593                                   unsigned long action,
 594                                   void *hcpu)
 595 {
 596         int hotcpu = (unsigned long)hcpu;
 597         struct task_struct *p;
 598
 599         switch (action) {
 600         case CPU_UP_PREPARE:
 601         case CPU_UP_PREPARE_FROZEN:
 602                 p = kthread_create(ksoftirqd, hcpu, "ksoftirqd/%d", hotcpu);
 603                 if (IS_ERR(p)) {
 604                         printk("ksoftirqd for %i failed\n", hotcpu);
 605                         return NOTIFY_BAD;
 606                 }
 607                 kthread_bind(p, hotcpu);
 608                 per_cpu(ksoftirqd, hotcpu) = p;
 609                 break;
 610         case CPU_ONLINE:
 611         case CPU_ONLINE_FROZEN:
 612                 wake_up_process(per_cpu(ksoftirqd, hotcpu));
 613                 break;
 614 #ifdef CONFIG_HOTPLUG_CPU
 615         case CPU_UP_CANCELED:
 616         case CPU_UP_CANCELED_FROZEN:
 617                 if (!per_cpu(ksoftirqd, hotcpu))
 618                         break;
 619                 /* Unbind so it can run.  Fall thru. */
 620                 kthread_bind(per_cpu(ksoftirqd, hotcpu),
 621                              any_online_cpu(cpu_online_map));
 622         case CPU_DEAD:
 623         case CPU_DEAD_FROZEN: {
 624                 struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
 625
 626                 p = per_cpu(ksoftirqd, hotcpu);
 627                 per_cpu(ksoftirqd, hotcpu) = NULL;
 628                 sched_setscheduler(p, SCHED_FIFO, &param);
 629                 kthread_stop(p);
 630                 takeover_tasklets(hotcpu);
 631                 break;
 632         }
 633 #endif /* CONFIG_HOTPLUG_CPU */
 634         }
 635         return NOTIFY_OK;
 636 }
 637
 638 static struct notifier_block __cpuinitdata cpu_nfb = {
 639         .notifier_call = cpu_callback
 640 };
 641
 642 __init int spawn_ksoftirqd(void)
 643 {
 644         void *cpu = (void *)(long)smp_processor_id();
 645         int err = cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);
 646
 647         BUG_ON(err == NOTIFY_BAD);
 648         cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);
 649         register_cpu_notifier(&cpu_nfb);
 650         return 0;
 651 }
 652
 653 #ifdef CONFIG_SMP
 654 /*
 655  * Call a function on all processors
 656  */
 657 int on_each_cpu(void (*func) (void *info), void *info, int retry, int wait)
 658 {
 659         int ret = 0;
 660
 661         preempt_disable();
 662         ret = smp_call_function(func, info, retry, wait);
 663         local_irq_disable();
 664         func(info);
 665         local_irq_enable();
 666         preempt_enable();
 667         return ret;
 668 }
 669 EXPORT_SYMBOL(on_each_cpu);
 670 #endif