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/delay.h>
  25 #include <linux/ftrace.h>
  26 #include <linux/smp.h>
  27 #include <linux/smpboot.h>
  28 #include <linux/tick.h>
  29 #include <linux/locallock.h>
  30 #include <linux/irq.h>
  31
  32 #define CREATE_TRACE_POINTS
  33 #include <trace/events/irq.h>
  34
  35 /*
  36    - No shared variables, all the data are CPU local.
  37    - If a softirq needs serialization, let it serialize itself
  38      by its own spinlocks.
  39    - Even if softirq is serialized, only local cpu is marked for
  40      execution. Hence, we get something sort of weak cpu binding.
  41      Though it is still not clear, will it result in better locality
  42      or will not.
  43
  44    Examples:
  45    - NET RX softirq. It is multithreaded and does not require
  46      any global serialization.
  47    - NET TX softirq. It kicks software netdevice queues, hence
  48      it is logically serialized per device, but this serialization
  49      is invisible to common code.
  50    - Tasklets: serialized wrt itself.
  51  */
  52
  53 #ifndef __ARCH_IRQ_STAT
  54 irq_cpustat_t irq_stat[NR_CPUS] ____cacheline_aligned;
  55 EXPORT_SYMBOL(irq_stat);
  56 #endif
  57
  58 static struct softirq_action softirq_vec[NR_SOFTIRQS] __cacheline_aligned_in_smp;
  59
  60 DEFINE_PER_CPU(struct task_struct *, ksoftirqd);
  61
  62 const char * const softirq_to_name[NR_SOFTIRQS] = {
  63         "HI", "TIMER", "NET_TX", "NET_RX", "BLOCK", "BLOCK_IOPOLL",
  64         "TASKLET", "SCHED", "HRTIMER", "RCU"
  65 };
  66
  67 #ifdef CONFIG_NO_HZ_COMMON
  68 # ifdef CONFIG_PREEMPT_RT_FULL
  69
  70 struct softirq_runner {
  71         struct task_struct *runner[NR_SOFTIRQS];
  72 };
  73
  74 static DEFINE_PER_CPU(struct softirq_runner, softirq_runners);
  75
  76 static inline void softirq_set_runner(unsigned int sirq)
  77 {
  78         struct softirq_runner *sr = this_cpu_ptr(&softirq_runners);
  79
  80         sr->runner[sirq] = current;
  81 }
  82
  83 static inline void softirq_clr_runner(unsigned int sirq)
  84 {
  85         struct softirq_runner *sr = this_cpu_ptr(&softirq_runners);
  86
  87         sr->runner[sirq] = NULL;
  88 }
  89
  90 /*
  91  * On preempt-rt a softirq running context might be blocked on a
  92  * lock. There might be no other runnable task on this CPU because the
  93  * lock owner runs on some other CPU. So we have to go into idle with
  94  * the pending bit set. Therefor we need to check this otherwise we
  95  * warn about false positives which confuses users and defeats the
  96  * whole purpose of this test.
  97  *
  98  * This code is called with interrupts disabled.
  99  */
 100 void softirq_check_pending_idle(void)
 101 {
 102         static int rate_limit;
 103         struct softirq_runner *sr = this_cpu_ptr(&softirq_runners);
 104         u32 warnpending;
 105         int i;
 106
 107         if (rate_limit >= 10)
 108                 return;
 109
 110         warnpending = local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK;
 111         for (i = 0; i < NR_SOFTIRQS; i++) {
 112                 struct task_struct *tsk = sr->runner[i];
 113
 114                 /*
 115                  * The wakeup code in rtmutex.c wakes up the task
 116                  * _before_ it sets pi_blocked_on to NULL under
 117                  * tsk->pi_lock. So we need to check for both: state
 118                  * and pi_blocked_on.
 119                  */
 120                 if (tsk) {
 121                         raw_spin_lock(&tsk->pi_lock);
 122                         if (tsk->pi_blocked_on || tsk->state == TASK_RUNNING) {
 123                                 /* Clear all bits pending in that task */
 124                                 warnpending &= ~(tsk->softirqs_raised);
 125                                 warnpending &= ~(1 << i);
 126                         }
 127                         raw_spin_unlock(&tsk->pi_lock);
 128                 }
 129         }
 130
 131         if (warnpending) {
 132                 printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n",
 133                        warnpending);
 134                 rate_limit++;
 135         }
 136 }
 137 # else
 138 /*
 139  * On !PREEMPT_RT we just printk rate limited:
 140  */
 141 void softirq_check_pending_idle(void)
 142 {
 143         static int rate_limit;
 144
 145         if (rate_limit < 10 &&
 146                         (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) {
 147                 printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n",
 148                        local_softirq_pending());
 149                 rate_limit++;
 150         }
 151 }
 152 # endif
 153
 154 #else /* !CONFIG_NO_HZ_COMMON */
 155 static inline void softirq_set_runner(unsigned int sirq) { }
 156 static inline void softirq_clr_runner(unsigned int sirq) { }
 157 #endif
 158
 159 /*
 160  * we cannot loop indefinitely here to avoid userspace starvation,
 161  * but we also don't want to introduce a worst case 1/HZ latency
 162  * to the pending events, so lets the scheduler to balance
 163  * the softirq load for us.
 164  */
 165 static void wakeup_softirqd(void)
 166 {
 167         /* Interrupts are disabled: no need to stop preemption */
 168         struct task_struct *tsk = __this_cpu_read(ksoftirqd);
 169
 170         if (tsk && tsk->state != TASK_RUNNING)
 171                 wake_up_process(tsk);
 172 }
 173
 174 static void handle_softirq(unsigned int vec_nr)
 175 {
 176         struct softirq_action *h = softirq_vec + vec_nr;
 177         int prev_count;
 178
 179         prev_count = preempt_count();
 180
 181         kstat_incr_softirqs_this_cpu(vec_nr);
 182
 183         trace_softirq_entry(vec_nr);
 184         h->action(h);
 185         trace_softirq_exit(vec_nr);
 186         if (unlikely(prev_count != preempt_count())) {
 187                 pr_err("huh, entered softirq %u %s %p with preempt_count %08x, exited with %08x?\n",
 188                        vec_nr, softirq_to_name[vec_nr], h->action,
 189                        prev_count, preempt_count());
 190                 preempt_count_set(prev_count);
 191         }
 192 }
 193
 194 #ifndef CONFIG_PREEMPT_RT_FULL
 195 static inline int ksoftirqd_softirq_pending(void)
 196 {
 197         return local_softirq_pending();
 198 }
 199
 200 static void handle_pending_softirqs(u32 pending, int need_rcu_bh_qs)
 201 {
 202         struct softirq_action *h = softirq_vec;
 203         int softirq_bit;
 204
 205         local_irq_enable();
 206
 207         h = softirq_vec;
 208
 209         while ((softirq_bit = ffs(pending))) {
 210                 unsigned int vec_nr;
 211
 212                 h += softirq_bit - 1;
 213                 vec_nr = h - softirq_vec;
 214                 handle_softirq(vec_nr);
 215
 216                 h++;
 217                 pending >>= softirq_bit;
 218         }
 219
 220         if (need_rcu_bh_qs)
 221                 rcu_bh_qs();
 222         local_irq_disable();
 223 }
 224
 225 static void run_ksoftirqd(unsigned int cpu)
 226 {
 227         local_irq_disable();
 228         if (ksoftirqd_softirq_pending()) {
 229                 __do_softirq();
 230                 local_irq_enable();
 231                 cond_resched_rcu_qs();
 232                 return;
 233         }
 234         local_irq_enable();
 235 }
 236
 237 /*
 238  * preempt_count and SOFTIRQ_OFFSET usage:
 239  * - preempt_count is changed by SOFTIRQ_OFFSET on entering or leaving
 240  *   softirq processing.
 241  * - preempt_count is changed by SOFTIRQ_DISABLE_OFFSET (= 2 * SOFTIRQ_OFFSET)
 242  *   on local_bh_disable or local_bh_enable.
 243  * This lets us distinguish between whether we are currently processing
 244  * softirq and whether we just have bh disabled.
 245  */
 246
 247 /*
 248  * This one is for softirq.c-internal use,
 249  * where hardirqs are disabled legitimately:
 250  */
 251 #ifdef CONFIG_TRACE_IRQFLAGS
 252 void __local_bh_disable_ip(unsigned long ip, unsigned int cnt)
 253 {
 254         unsigned long flags;
 255
 256         WARN_ON_ONCE(in_irq());
 257
 258         raw_local_irq_save(flags);
 259         /*
 260          * The preempt tracer hooks into preempt_count_add and will break
 261          * lockdep because it calls back into lockdep after SOFTIRQ_OFFSET
 262          * is set and before current->softirq_enabled is cleared.
 263          * We must manually increment preempt_count here and manually
 264          * call the trace_preempt_off later.
 265          */
 266         __preempt_count_add(cnt);
 267         /*
 268          * Were softirqs turned off above:
 269          */
 270         if (softirq_count() == (cnt & SOFTIRQ_MASK))
 271                 trace_softirqs_off(ip);
 272         raw_local_irq_restore(flags);
 273
 274         if (preempt_count() == cnt) {
 275 #ifdef CONFIG_DEBUG_PREEMPT
 276                 current->preempt_disable_ip = get_parent_ip(CALLER_ADDR1);
 277 #endif
 278                 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
 279         }
 280 }
 281 EXPORT_SYMBOL(__local_bh_disable_ip);
 282 #endif /* CONFIG_TRACE_IRQFLAGS */
 283
 284 static void __local_bh_enable(unsigned int cnt)
 285 {
 286         WARN_ON_ONCE(!irqs_disabled());
 287
 288         if (softirq_count() == (cnt & SOFTIRQ_MASK))
 289                 trace_softirqs_on(_RET_IP_);
 290         preempt_count_sub(cnt);
 291 }
 292
 293 /*
 294  * Special-case - softirqs can safely be enabled in
 295  * cond_resched_softirq(), or by __do_softirq(),
 296  * without processing still-pending softirqs:
 297  */
 298 void _local_bh_enable(void)
 299 {
 300         WARN_ON_ONCE(in_irq());
 301         __local_bh_enable(SOFTIRQ_DISABLE_OFFSET);
 302 }
 303 EXPORT_SYMBOL(_local_bh_enable);
 304
 305 void __local_bh_enable_ip(unsigned long ip, unsigned int cnt)
 306 {
 307         WARN_ON_ONCE(in_irq() || irqs_disabled());
 308 #ifdef CONFIG_TRACE_IRQFLAGS
 309         local_irq_disable();
 310 #endif
 311         /*
 312          * Are softirqs going to be turned on now:
 313          */
 314         if (softirq_count() == SOFTIRQ_DISABLE_OFFSET)
 315                 trace_softirqs_on(ip);
 316         /*
 317          * Keep preemption disabled until we are done with
 318          * softirq processing:
 319          */
 320         preempt_count_sub(cnt - 1);
 321
 322         if (unlikely(!in_interrupt() && local_softirq_pending())) {
 323                 /*
 324                  * Run softirq if any pending. And do it in its own stack
 325                  * as we may be calling this deep in a task call stack already.
 326                  */
 327                 do_softirq();
 328         }
 329
 330         preempt_count_dec();
 331 #ifdef CONFIG_TRACE_IRQFLAGS
 332         local_irq_enable();
 333 #endif
 334         preempt_check_resched();
 335 }
 336 EXPORT_SYMBOL(__local_bh_enable_ip);
 337
 338 /*
 339  * We restart softirq processing for at most MAX_SOFTIRQ_RESTART times,
 340  * but break the loop if need_resched() is set or after 2 ms.
 341  * The MAX_SOFTIRQ_TIME provides a nice upper bound in most cases, but in
 342  * certain cases, such as stop_machine(), jiffies may cease to
 343  * increment and so we need the MAX_SOFTIRQ_RESTART limit as
 344  * well to make sure we eventually return from this method.
 345  *
 346  * These limits have been established via experimentation.
 347  * The two things to balance is latency against fairness -
 348  * we want to handle softirqs as soon as possible, but they
 349  * should not be able to lock up the box.
 350  */
 351 #define MAX_SOFTIRQ_TIME  msecs_to_jiffies(2)
 352 #define MAX_SOFTIRQ_RESTART 10
 353
 354 #ifdef CONFIG_TRACE_IRQFLAGS
 355 /*
 356  * When we run softirqs from irq_exit() and thus on the hardirq stack we need
 357  * to keep the lockdep irq context tracking as tight as possible in order to
 358  * not miss-qualify lock contexts and miss possible deadlocks.
 359  */
 360
 361 static inline bool lockdep_softirq_start(void)
 362 {
 363         bool in_hardirq = false;
 364
 365         if (trace_hardirq_context(current)) {
 366                 in_hardirq = true;
 367                 trace_hardirq_exit();
 368         }
 369
 370         lockdep_softirq_enter();
 371
 372         return in_hardirq;
 373 }
 374
 375 static inline void lockdep_softirq_end(bool in_hardirq)
 376 {
 377         lockdep_softirq_exit();
 378
 379         if (in_hardirq)
 380                 trace_hardirq_enter();
 381 }
 382 #else
 383 static inline bool lockdep_softirq_start(void) { return false; }
 384 static inline void lockdep_softirq_end(bool in_hardirq) { }
 385 #endif
 386
 387 asmlinkage __visible void __do_softirq(void)
 388 {
 389         unsigned long end = jiffies + MAX_SOFTIRQ_TIME;
 390         unsigned long old_flags = current->flags;
 391         int max_restart = MAX_SOFTIRQ_RESTART;
 392         bool in_hardirq;
 393         __u32 pending;
 394
 395         /*
 396          * Mask out PF_MEMALLOC s current task context is borrowed for the
 397          * softirq. A softirq handled such as network RX might set PF_MEMALLOC
 398          * again if the socket is related to swap
 399          */
 400         current->flags &= ~PF_MEMALLOC;
 401
 402         pending = local_softirq_pending();
 403         account_irq_enter_time(current);
 404
 405         __local_bh_disable_ip(_RET_IP_, SOFTIRQ_OFFSET);
 406         in_hardirq = lockdep_softirq_start();
 407
 408 restart:
 409         /* Reset the pending bitmask before enabling irqs */
 410         set_softirq_pending(0);
 411
 412         handle_pending_softirqs(pending, 1);
 413
 414         pending = local_softirq_pending();
 415         if (pending) {
 416                 if (time_before(jiffies, end) && !need_resched() &&
 417                     --max_restart)
 418                         goto restart;
 419
 420                 wakeup_softirqd();
 421         }
 422
 423         lockdep_softirq_end(in_hardirq);
 424         account_irq_exit_time(current);
 425         __local_bh_enable(SOFTIRQ_OFFSET);
 426         WARN_ON_ONCE(in_interrupt());
 427         tsk_restore_flags(current, old_flags, PF_MEMALLOC);
 428 }
 429
 430 asmlinkage __visible void do_softirq(void)
 431 {
 432         __u32 pending;
 433         unsigned long flags;
 434
 435         if (in_interrupt())
 436                 return;
 437
 438         local_irq_save(flags);
 439
 440         pending = local_softirq_pending();
 441
 442         if (pending)
 443                 do_softirq_own_stack();
 444
 445         local_irq_restore(flags);
 446 }
 447
 448 /*
 449  * This function must run with irqs disabled!
 450  */
 451 void raise_softirq_irqoff(unsigned int nr)
 452 {
 453         __raise_softirq_irqoff(nr);
 454
 455         /*
 456          * If we're in an interrupt or softirq, we're done
 457          * (this also catches softirq-disabled code). We will
 458          * actually run the softirq once we return from
 459          * the irq or softirq.
 460          *
 461          * Otherwise we wake up ksoftirqd to make sure we
 462          * schedule the softirq soon.
 463          */
 464         if (!in_interrupt())
 465                 wakeup_softirqd();
 466 }
 467
 468 void __raise_softirq_irqoff(unsigned int nr)
 469 {
 470         trace_softirq_raise(nr);
 471         or_softirq_pending(1UL << nr);
 472 }
 473
 474 static inline void local_bh_disable_nort(void) { local_bh_disable(); }
 475 static inline void _local_bh_enable_nort(void) { _local_bh_enable(); }
 476 static void ksoftirqd_set_sched_params(unsigned int cpu) { }
 477 static void ksoftirqd_clr_sched_params(unsigned int cpu, bool online) { }
 478
 479 #else /* !PREEMPT_RT_FULL */
 480
 481 /*
 482  * On RT we serialize softirq execution with a cpu local lock per softirq
 483  */
 484 static DEFINE_PER_CPU(struct local_irq_lock [NR_SOFTIRQS], local_softirq_locks);
 485
 486 void __init softirq_early_init(void)
 487 {
 488         int i;
 489
 490         for (i = 0; i < NR_SOFTIRQS; i++)
 491                 local_irq_lock_init(local_softirq_locks[i]);
 492 }
 493
 494 static void lock_softirq(int which)
 495 {
 496         local_lock(local_softirq_locks[which]);
 497 }
 498
 499 static void unlock_softirq(int which)
 500 {
 501         local_unlock(local_softirq_locks[which]);
 502 }
 503
 504 static void do_single_softirq(int which, int need_rcu_bh_qs)
 505 {
 506         unsigned long old_flags = current->flags;
 507
 508         current->flags &= ~PF_MEMALLOC;
 509         vtime_account_irq_enter(current);
 510         current->flags |= PF_IN_SOFTIRQ;
 511         lockdep_softirq_enter();
 512         local_irq_enable();
 513         handle_softirq(which);
 514         local_irq_disable();
 515         lockdep_softirq_exit();
 516         current->flags &= ~PF_IN_SOFTIRQ;
 517         vtime_account_irq_enter(current);
 518         tsk_restore_flags(current, old_flags, PF_MEMALLOC);
 519 }
 520
 521 /*
 522  * Called with interrupts disabled. Process softirqs which were raised
 523  * in current context (or on behalf of ksoftirqd).
 524  */
 525 static void do_current_softirqs(int need_rcu_bh_qs)
 526 {
 527         while (current->softirqs_raised) {
 528                 int i = __ffs(current->softirqs_raised);
 529                 unsigned int pending, mask = (1U << i);
 530
 531                 current->softirqs_raised &= ~mask;
 532                 local_irq_enable();
 533
 534                 /*
 535                  * If the lock is contended, we boost the owner to
 536                  * process the softirq or leave the critical section
 537                  * now.
 538                  */
 539                 lock_softirq(i);
 540                 local_irq_disable();
 541                 softirq_set_runner(i);
 542                 /*
 543                  * Check with the local_softirq_pending() bits,
 544                  * whether we need to process this still or if someone
 545                  * else took care of it.
 546                  */
 547                 pending = local_softirq_pending();
 548                 if (pending & mask) {
 549                         set_softirq_pending(pending & ~mask);
 550                         do_single_softirq(i, need_rcu_bh_qs);
 551                 }
 552                 softirq_clr_runner(i);
 553                 unlock_softirq(i);
 554                 WARN_ON(current->softirq_nestcnt != 1);
 555         }
 556 }
 557
 558 static void __local_bh_disable(void)
 559 {
 560         if (++current->softirq_nestcnt == 1)
 561                 migrate_disable();
 562 }
 563
 564 void local_bh_disable(void)
 565 {
 566         __local_bh_disable();
 567 }
 568 EXPORT_SYMBOL(local_bh_disable);
 569
 570 void __local_bh_disable_ip(unsigned long ip, unsigned int cnt)
 571 {
 572         __local_bh_disable();
 573         if (cnt & PREEMPT_CHECK_OFFSET)
 574                 preempt_disable();
 575 }
 576
 577 static void __local_bh_enable(void)
 578 {
 579         if (WARN_ON(current->softirq_nestcnt == 0))
 580                 return;
 581
 582         local_irq_disable();
 583         if (current->softirq_nestcnt == 1 && current->softirqs_raised)
 584                 do_current_softirqs(1);
 585         local_irq_enable();
 586
 587         if (--current->softirq_nestcnt == 0)
 588                 migrate_enable();
 589 }
 590
 591 void local_bh_enable(void)
 592 {
 593         __local_bh_enable();
 594 }
 595 EXPORT_SYMBOL(local_bh_enable);
 596
 597 extern void __local_bh_enable_ip(unsigned long ip, unsigned int cnt)
 598 {
 599         __local_bh_enable();
 600         if (cnt & PREEMPT_CHECK_OFFSET)
 601                 preempt_enable();
 602 }
 603
 604 void local_bh_enable_ip(unsigned long ip)
 605 {
 606         local_bh_enable();
 607 }
 608 EXPORT_SYMBOL(local_bh_enable_ip);
 609
 610 void _local_bh_enable(void)
 611 {
 612         if (WARN_ON(current->softirq_nestcnt == 0))
 613                 return;
 614         if (--current->softirq_nestcnt == 0)
 615                 migrate_enable();
 616 }
 617 EXPORT_SYMBOL(_local_bh_enable);
 618
 619 int in_serving_softirq(void)
 620 {
 621         return current->flags & PF_IN_SOFTIRQ;
 622 }
 623 EXPORT_SYMBOL(in_serving_softirq);
 624
 625 /* Called with preemption disabled */
 626 static void run_ksoftirqd(unsigned int cpu)
 627 {
 628         local_irq_disable();
 629         current->softirq_nestcnt++;
 630
 631         do_current_softirqs(1);
 632         current->softirq_nestcnt--;
 633         rcu_note_context_switch();
 634         local_irq_enable();
 635 }
 636
 637 /*
 638  * Called from netif_rx_ni(). Preemption enabled, but migration
 639  * disabled. So the cpu can't go away under us.
 640  */
 641 void thread_do_softirq(void)
 642 {
 643         if (!in_serving_softirq() && current->softirqs_raised) {
 644                 current->softirq_nestcnt++;
 645                 do_current_softirqs(0);
 646                 current->softirq_nestcnt--;
 647         }
 648 }
 649
 650 static void do_raise_softirq_irqoff(unsigned int nr)
 651 {
 652         trace_softirq_raise(nr);
 653         or_softirq_pending(1UL << nr);
 654
 655         /*
 656          * If we are not in a hard interrupt and inside a bh disabled
 657          * region, we simply raise the flag on current. local_bh_enable()
 658          * will make sure that the softirq is executed. Otherwise we
 659          * delegate it to ksoftirqd.
 660          */
 661         if (!in_irq() && current->softirq_nestcnt)
 662                 current->softirqs_raised |= (1U << nr);
 663         else if (__this_cpu_read(ksoftirqd))
 664                 __this_cpu_read(ksoftirqd)->softirqs_raised |= (1U << nr);
 665 }
 666
 667 void __raise_softirq_irqoff(unsigned int nr)
 668 {
 669         do_raise_softirq_irqoff(nr);
 670         if (!in_irq() && !current->softirq_nestcnt)
 671                 wakeup_softirqd();
 672 }
 673
 674 /*
 675  * This function must run with irqs disabled!
 676  */
 677 void raise_softirq_irqoff(unsigned int nr)
 678 {
 679         do_raise_softirq_irqoff(nr);
 680
 681         /*
 682          * If we're in an hard interrupt we let irq return code deal
 683          * with the wakeup of ksoftirqd.
 684          */
 685         if (in_irq())
 686                 return;
 687         /*
 688          * If we are in thread context but outside of a bh disabled
 689          * region, we need to wake ksoftirqd as well.
 690          *
 691          * CHECKME: Some of the places which do that could be wrapped
 692          * into local_bh_disable/enable pairs. Though it's unclear
 693          * whether this is worth the effort. To find those places just
 694          * raise a WARN() if the condition is met.
 695          */
 696         if (!current->softirq_nestcnt)
 697                 wakeup_softirqd();
 698 }
 699
 700 static inline int ksoftirqd_softirq_pending(void)
 701 {
 702         return current->softirqs_raised;
 703 }
 704
 705 static inline void local_bh_disable_nort(void) { }
 706 static inline void _local_bh_enable_nort(void) { }
 707
 708 static inline void ksoftirqd_set_sched_params(unsigned int cpu)
 709 {
 710         struct sched_param param = { .sched_priority = 1 };
 711
 712         sched_setscheduler(current, SCHED_FIFO, &param);
 713         /* Take over all pending softirqs when starting */
 714         local_irq_disable();
 715         current->softirqs_raised = local_softirq_pending();
 716         local_irq_enable();
 717 }
 718
 719 static inline void ksoftirqd_clr_sched_params(unsigned int cpu, bool online)
 720 {
 721         struct sched_param param = { .sched_priority = 0 };
 722
 723         sched_setscheduler(current, SCHED_NORMAL, &param);
 724 }
 725
 726 #endif /* PREEMPT_RT_FULL */
 727 /*
 728  * Enter an interrupt context.
 729  */
 730 void irq_enter(void)
 731 {
 732         rcu_irq_enter();
 733         if (is_idle_task(current) && !in_interrupt()) {
 734                 /*
 735                  * Prevent raise_softirq from needlessly waking up ksoftirqd
 736                  * here, as softirq will be serviced on return from interrupt.
 737                  */
 738                 local_bh_disable_nort();
 739                 tick_irq_enter();
 740                 _local_bh_enable_nort();
 741         }
 742
 743         __irq_enter();
 744 }
 745
 746 static inline void invoke_softirq(void)
 747 {
 748 #ifndef CONFIG_PREEMPT_RT_FULL
 749         if (!force_irqthreads) {
 750 #ifdef CONFIG_HAVE_IRQ_EXIT_ON_IRQ_STACK
 751                 /*
 752                  * We can safely execute softirq on the current stack if
 753                  * it is the irq stack, because it should be near empty
 754                  * at this stage.
 755                  */
 756                 __do_softirq();
 757 #else
 758                 /*
 759                  * Otherwise, irq_exit() is called on the task stack that can
 760                  * be potentially deep already. So call softirq in its own stack
 761                  * to prevent from any overrun.
 762                  */
 763                 do_softirq_own_stack();
 764 #endif
 765         } else {
 766                 wakeup_softirqd();
 767         }
 768 #else /* PREEMPT_RT_FULL */
 769         unsigned long flags;
 770
 771         local_irq_save(flags);
 772         if (__this_cpu_read(ksoftirqd) &&
 773                         __this_cpu_read(ksoftirqd)->softirqs_raised)
 774                 wakeup_softirqd();
 775         local_irq_restore(flags);
 776 #endif
 777 }
 778
 779 static inline void tick_irq_exit(void)
 780 {
 781 #ifdef CONFIG_NO_HZ_COMMON
 782         int cpu = smp_processor_id();
 783
 784         /* Make sure that timer wheel updates are propagated */
 785         if ((idle_cpu(cpu) && !need_resched()) || tick_nohz_full_cpu(cpu)) {
 786                 if (!in_interrupt())
 787                         tick_nohz_irq_exit();
 788         }
 789 #endif
 790 }
 791
 792 /*
 793  * Exit an interrupt context. Process softirqs if needed and possible:
 794  */
 795 void irq_exit(void)
 796 {
 797 #ifndef __ARCH_IRQ_EXIT_IRQS_DISABLED
 798         local_irq_disable();
 799 #else
 800         WARN_ON_ONCE(!irqs_disabled());
 801 #endif
 802
 803         account_irq_exit_time(current);
 804         preempt_count_sub(HARDIRQ_OFFSET);
 805         if (!in_interrupt() && local_softirq_pending())
 806                 invoke_softirq();
 807
 808         tick_irq_exit();
 809         rcu_irq_exit();
 810         trace_hardirq_exit(); /* must be last! */
 811 }
 812
 813 void raise_softirq(unsigned int nr)
 814 {
 815         unsigned long flags;
 816
 817         local_irq_save(flags);
 818         raise_softirq_irqoff(nr);
 819         local_irq_restore(flags);
 820 }
 821
 822 void open_softirq(int nr, void (*action)(struct softirq_action *))
 823 {
 824         softirq_vec[nr].action = action;
 825 }
 826
 827 /*
 828  * Tasklets
 829  */
 830 struct tasklet_head {
 831         struct tasklet_struct *head;
 832         struct tasklet_struct **tail;
 833 };
 834
 835 static DEFINE_PER_CPU(struct tasklet_head, tasklet_vec);
 836 static DEFINE_PER_CPU(struct tasklet_head, tasklet_hi_vec);
 837
 838 static void inline
 839 __tasklet_common_schedule(struct tasklet_struct *t, struct tasklet_head *head, unsigned int nr)
 840 {
 841         if (tasklet_trylock(t)) {
 842 again:
 843                 /* We may have been preempted before tasklet_trylock
 844                  * and __tasklet_action may have already run.
 845                  * So double check the sched bit while the takslet
 846                  * is locked before adding it to the list.
 847                  */
 848                 if (test_bit(TASKLET_STATE_SCHED, &t->state)) {
 849                         t->next = NULL;
 850                         *head->tail = t;
 851                         head->tail = &(t->next);
 852                         raise_softirq_irqoff(nr);
 853                         tasklet_unlock(t);
 854                 } else {
 855                         /* This is subtle. If we hit the corner case above
 856                          * It is possible that we get preempted right here,
 857                          * and another task has successfully called
 858                          * tasklet_schedule(), then this function, and
 859                          * failed on the trylock. Thus we must be sure
 860                          * before releasing the tasklet lock, that the
 861                          * SCHED_BIT is clear. Otherwise the tasklet
 862                          * may get its SCHED_BIT set, but not added to the
 863                          * list
 864                          */
 865                         if (!tasklet_tryunlock(t))
 866                                 goto again;
 867                 }
 868         }
 869 }
 870
 871 void __tasklet_schedule(struct tasklet_struct *t)
 872 {
 873         unsigned long flags;
 874
 875         local_irq_save(flags);
 876         __tasklet_common_schedule(t, this_cpu_ptr(&tasklet_vec), TASKLET_SOFTIRQ);
 877         local_irq_restore(flags);
 878 }
 879 EXPORT_SYMBOL(__tasklet_schedule);
 880
 881 void __tasklet_hi_schedule(struct tasklet_struct *t)
 882 {
 883         unsigned long flags;
 884
 885         local_irq_save(flags);
 886         __tasklet_common_schedule(t, this_cpu_ptr(&tasklet_hi_vec), HI_SOFTIRQ);
 887         local_irq_restore(flags);
 888 }
 889 EXPORT_SYMBOL(__tasklet_hi_schedule);
 890
 891 void __tasklet_hi_schedule_first(struct tasklet_struct *t)
 892 {
 893         BUG_ON(!irqs_disabled());
 894
 895         __tasklet_hi_schedule(t);
 896 }
 897 EXPORT_SYMBOL(__tasklet_hi_schedule_first);
 898
 899 void  tasklet_enable(struct tasklet_struct *t)
 900 {
 901         if (!atomic_dec_and_test(&t->count))
 902                 return;
 903         if (test_and_clear_bit(TASKLET_STATE_PENDING, &t->state))
 904                 tasklet_schedule(t);
 905 }
 906 EXPORT_SYMBOL(tasklet_enable);
 907
 908 static void __tasklet_action(struct softirq_action *a,
 909                              struct tasklet_struct *list)
 910 {
 911         int loops = 1000000;
 912
 913         while (list) {
 914                 struct tasklet_struct *t = list;
 915
 916                 list = list->next;
 917
 918                 /*
 919                  * Should always succeed - after a tasklist got on the
 920                  * list (after getting the SCHED bit set from 0 to 1),
 921                  * nothing but the tasklet softirq it got queued to can
 922                  * lock it:
 923                  */
 924                 if (!tasklet_trylock(t)) {
 925                         WARN_ON(1);
 926                         continue;
 927                 }
 928
 929                 t->next = NULL;
 930
 931                 /*
 932                  * If we cannot handle the tasklet because it's disabled,
 933                  * mark it as pending. tasklet_enable() will later
 934                  * re-schedule the tasklet.
 935                  */
 936                 if (unlikely(atomic_read(&t->count))) {
 937 out_disabled:
 938                         /* implicit unlock: */
 939                         wmb();
 940                         t->state = TASKLET_STATEF_PENDING;
 941                         continue;
 942                 }
 943
 944                 /*
 945                  * After this point on the tasklet might be rescheduled
 946                  * on another CPU, but it can only be added to another
 947                  * CPU's tasklet list if we unlock the tasklet (which we
 948                  * dont do yet).
 949                  */
 950                 if (!test_and_clear_bit(TASKLET_STATE_SCHED, &t->state))
 951                         WARN_ON(1);
 952
 953 again:
 954                 t->func(t->data);
 955
 956                 /*
 957                  * Try to unlock the tasklet. We must use cmpxchg, because
 958                  * another CPU might have scheduled or disabled the tasklet.
 959                  * We only allow the STATE_RUN -> 0 transition here.
 960                  */
 961                 while (!tasklet_tryunlock(t)) {
 962                         /*
 963                          * If it got disabled meanwhile, bail out:
 964                          */
 965                         if (atomic_read(&t->count))
 966                                 goto out_disabled;
 967                         /*
 968                          * If it got scheduled meanwhile, re-execute
 969                          * the tasklet function:
 970                          */
 971                         if (test_and_clear_bit(TASKLET_STATE_SCHED, &t->state))
 972                                 goto again;
 973                         if (!--loops) {
 974                                 printk("hm, tasklet state: %08lx\n", t->state);
 975                                 WARN_ON(1);
 976                                 tasklet_unlock(t);
 977                                 break;
 978                         }
 979                 }
 980         }
 981 }
 982
 983 static void tasklet_action(struct softirq_action *a)
 984 {
 985         struct tasklet_struct *list;
 986
 987         local_irq_disable();
 988
 989         list = __this_cpu_read(tasklet_vec.head);
 990         __this_cpu_write(tasklet_vec.head, NULL);
 991         __this_cpu_write(tasklet_vec.tail, this_cpu_ptr(&tasklet_vec.head));
 992
 993         local_irq_enable();
 994
 995         __tasklet_action(a, list);
 996 }
 997
 998 static void tasklet_hi_action(struct softirq_action *a)
 999 {
1000         struct tasklet_struct *list;
1001
1002         local_irq_disable();
1003
1004         list = __this_cpu_read(tasklet_hi_vec.head);
1005         __this_cpu_write(tasklet_hi_vec.head, NULL);
1006         __this_cpu_write(tasklet_hi_vec.tail, this_cpu_ptr(&tasklet_hi_vec.head));
1007
1008         local_irq_enable();
1009
1010         __tasklet_action(a, list);
1011 }
1012
1013 void tasklet_init(struct tasklet_struct *t,
1014                   void (*func)(unsigned long), unsigned long data)
1015 {
1016         t->next = NULL;
1017         t->state = 0;
1018         atomic_set(&t->count, 0);
1019         t->func = func;
1020         t->data = data;
1021 }
1022 EXPORT_SYMBOL(tasklet_init);
1023
1024 void tasklet_kill(struct tasklet_struct *t)
1025 {
1026         if (in_interrupt())
1027                 pr_notice("Attempt to kill tasklet from interrupt\n");
1028
1029         while (test_and_set_bit(TASKLET_STATE_SCHED, &t->state)) {
1030                 do {
1031                         msleep(1);
1032                 } while (test_bit(TASKLET_STATE_SCHED, &t->state));
1033         }
1034         tasklet_unlock_wait(t);
1035         clear_bit(TASKLET_STATE_SCHED, &t->state);
1036 }
1037 EXPORT_SYMBOL(tasklet_kill);
1038
1039 /*
1040  * tasklet_hrtimer
1041  */
1042
1043 /*
1044  * The trampoline is called when the hrtimer expires. It schedules a tasklet
1045  * to run __tasklet_hrtimer_trampoline() which in turn will call the intended
1046  * hrtimer callback, but from softirq context.
1047  */
1048 static enum hrtimer_restart __hrtimer_tasklet_trampoline(struct hrtimer *timer)
1049 {
1050         struct tasklet_hrtimer *ttimer =
1051                 container_of(timer, struct tasklet_hrtimer, timer);
1052
1053         tasklet_hi_schedule(&ttimer->tasklet);
1054         return HRTIMER_NORESTART;
1055 }
1056
1057 /*
1058  * Helper function which calls the hrtimer callback from
1059  * tasklet/softirq context
1060  */
1061 static void __tasklet_hrtimer_trampoline(unsigned long data)
1062 {
1063         struct tasklet_hrtimer *ttimer = (void *)data;
1064         enum hrtimer_restart restart;
1065
1066         restart = ttimer->function(&ttimer->timer);
1067         if (restart != HRTIMER_NORESTART)
1068                 hrtimer_restart(&ttimer->timer);
1069 }
1070
1071 /**
1072  * tasklet_hrtimer_init - Init a tasklet/hrtimer combo for softirq callbacks
1073  * @ttimer:      tasklet_hrtimer which is initialized
1074  * @function:    hrtimer callback function which gets called from softirq context
1075  * @which_clock: clock id (CLOCK_MONOTONIC/CLOCK_REALTIME)
1076  * @mode:        hrtimer mode (HRTIMER_MODE_ABS/HRTIMER_MODE_REL)
1077  */
1078 void tasklet_hrtimer_init(struct tasklet_hrtimer *ttimer,
1079                           enum hrtimer_restart (*function)(struct hrtimer *),
1080                           clockid_t which_clock, enum hrtimer_mode mode)
1081 {
1082         hrtimer_init(&ttimer->timer, which_clock, mode);
1083         ttimer->timer.function = __hrtimer_tasklet_trampoline;
1084         tasklet_init(&ttimer->tasklet, __tasklet_hrtimer_trampoline,
1085                      (unsigned long)ttimer);
1086         ttimer->function = function;
1087 }
1088 EXPORT_SYMBOL_GPL(tasklet_hrtimer_init);
1089
1090 void __init softirq_init(void)
1091 {
1092         int cpu;
1093
1094         for_each_possible_cpu(cpu) {
1095                 per_cpu(tasklet_vec, cpu).tail =
1096                         &per_cpu(tasklet_vec, cpu).head;
1097                 per_cpu(tasklet_hi_vec, cpu).tail =
1098                         &per_cpu(tasklet_hi_vec, cpu).head;
1099         }
1100
1101         open_softirq(TASKLET_SOFTIRQ, tasklet_action);
1102         open_softirq(HI_SOFTIRQ, tasklet_hi_action);
1103 }
1104
1105 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT_FULL)
1106 void tasklet_unlock_wait(struct tasklet_struct *t)
1107 {
1108         while (test_bit(TASKLET_STATE_RUN, &(t)->state)) {
1109                 /*
1110                  * Hack for now to avoid this busy-loop:
1111                  */
1112 #ifdef CONFIG_PREEMPT_RT_FULL
1113                 msleep(1);
1114 #else
1115                 barrier();
1116 #endif
1117         }
1118 }
1119 EXPORT_SYMBOL(tasklet_unlock_wait);
1120 #endif
1121
1122 static int ksoftirqd_should_run(unsigned int cpu)
1123 {
1124         return ksoftirqd_softirq_pending();
1125 }
1126
1127 #ifdef CONFIG_HOTPLUG_CPU
1128 /*
1129  * tasklet_kill_immediate is called to remove a tasklet which can already be
1130  * scheduled for execution on @cpu.
1131  *
1132  * Unlike tasklet_kill, this function removes the tasklet
1133  * _immediately_, even if the tasklet is in TASKLET_STATE_SCHED state.
1134  *
1135  * When this function is called, @cpu must be in the CPU_DEAD state.
1136  */
1137 void tasklet_kill_immediate(struct tasklet_struct *t, unsigned int cpu)
1138 {
1139         struct tasklet_struct **i;
1140
1141         BUG_ON(cpu_online(cpu));
1142         BUG_ON(test_bit(TASKLET_STATE_RUN, &t->state));
1143
1144         if (!test_bit(TASKLET_STATE_SCHED, &t->state))
1145                 return;
1146
1147         /* CPU is dead, so no lock needed. */
1148         for (i = &per_cpu(tasklet_vec, cpu).head; *i; i = &(*i)->next) {
1149                 if (*i == t) {
1150                         *i = t->next;
1151                         /* If this was the tail element, move the tail ptr */
1152                         if (*i == NULL)
1153                                 per_cpu(tasklet_vec, cpu).tail = i;
1154                         return;
1155                 }
1156         }
1157         BUG();
1158 }
1159
1160 static void takeover_tasklets(unsigned int cpu)
1161 {
1162         /* CPU is dead, so no lock needed. */
1163         local_irq_disable();
1164
1165         /* Find end, append list for that CPU. */
1166         if (&per_cpu(tasklet_vec, cpu).head != per_cpu(tasklet_vec, cpu).tail) {
1167                 *__this_cpu_read(tasklet_vec.tail) = per_cpu(tasklet_vec, cpu).head;
1168                 this_cpu_write(tasklet_vec.tail, per_cpu(tasklet_vec, cpu).tail);
1169                 per_cpu(tasklet_vec, cpu).head = NULL;
1170                 per_cpu(tasklet_vec, cpu).tail = &per_cpu(tasklet_vec, cpu).head;
1171         }
1172         raise_softirq_irqoff(TASKLET_SOFTIRQ);
1173
1174         if (&per_cpu(tasklet_hi_vec, cpu).head != per_cpu(tasklet_hi_vec, cpu).tail) {
1175                 *__this_cpu_read(tasklet_hi_vec.tail) = per_cpu(tasklet_hi_vec, cpu).head;
1176                 __this_cpu_write(tasklet_hi_vec.tail, per_cpu(tasklet_hi_vec, cpu).tail);
1177                 per_cpu(tasklet_hi_vec, cpu).head = NULL;
1178                 per_cpu(tasklet_hi_vec, cpu).tail = &per_cpu(tasklet_hi_vec, cpu).head;
1179         }
1180         raise_softirq_irqoff(HI_SOFTIRQ);
1181
1182         local_irq_enable();
1183 }
1184 #endif /* CONFIG_HOTPLUG_CPU */
1185
1186 static int cpu_callback(struct notifier_block *nfb, unsigned long action,
1187                         void *hcpu)
1188 {
1189         switch (action) {
1190 #ifdef CONFIG_HOTPLUG_CPU
1191         case CPU_DEAD:
1192         case CPU_DEAD_FROZEN:
1193                 takeover_tasklets((unsigned long)hcpu);
1194                 break;
1195 #endif /* CONFIG_HOTPLUG_CPU */
1196         }
1197         return NOTIFY_OK;
1198 }
1199
1200 static struct notifier_block cpu_nfb = {
1201         .notifier_call = cpu_callback
1202 };
1203
1204 static struct smp_hotplug_thread softirq_threads = {
1205         .store                  = &ksoftirqd,
1206         .setup                  = ksoftirqd_set_sched_params,
1207         .cleanup                = ksoftirqd_clr_sched_params,
1208         .thread_should_run      = ksoftirqd_should_run,
1209         .thread_fn              = run_ksoftirqd,
1210         .thread_comm            = "ksoftirqd/%u",
1211 };
1212
1213 static __init int spawn_ksoftirqd(void)
1214 {
1215         register_cpu_notifier(&cpu_nfb);
1216
1217         BUG_ON(smpboot_register_percpu_thread(&softirq_threads));
1218
1219         return 0;
1220 }
1221 early_initcall(spawn_ksoftirqd);
1222
1223 /*
1224  * [ These __weak aliases are kept in a separate compilation unit, so that
1225  *   GCC does not inline them incorrectly. ]
1226  */
1227
1228 int __init __weak early_irq_init(void)
1229 {
1230         return 0;
1231 }
1232
1233 int __init __weak arch_probe_nr_irqs(void)
1234 {
1235         return NR_IRQS_LEGACY;
1236 }
1237
1238 int __init __weak arch_early_irq_init(void)
1239 {
1240         return 0;
1241 }
1242
1243 unsigned int __weak arch_dynirq_lower_bound(unsigned int from)
1244 {
1245         return from;
1246 }