Kyle Swenson | 8d8f654 | 2021-03-15 11:02:55 -0600 | [diff] [blame] | 1 | /* |
| 2 | * linux/kernel/softirq.c |
| 3 | * |
| 4 | * Copyright (C) 1992 Linus Torvalds |
| 5 | * |
| 6 | * Distribute under GPLv2. |
| 7 | * |
| 8 | * Rewritten. Old one was good in 2.2, but in 2.3 it was immoral. --ANK (990903) |
| 9 | */ |
| 10 | |
| 11 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
| 12 | |
| 13 | #include <linux/export.h> |
| 14 | #include <linux/kernel_stat.h> |
| 15 | #include <linux/interrupt.h> |
| 16 | #include <linux/init.h> |
| 17 | #include <linux/mm.h> |
| 18 | #include <linux/notifier.h> |
| 19 | #include <linux/percpu.h> |
| 20 | #include <linux/cpu.h> |
| 21 | #include <linux/freezer.h> |
| 22 | #include <linux/kthread.h> |
| 23 | #include <linux/rcupdate.h> |
| 24 | #include <linux/ftrace.h> |
| 25 | #include <linux/smp.h> |
| 26 | #include <linux/smpboot.h> |
| 27 | #include <linux/tick.h> |
| 28 | #include <linux/irq.h> |
| 29 | |
| 30 | #define CREATE_TRACE_POINTS |
| 31 | #include <trace/events/irq.h> |
| 32 | |
| 33 | /* |
| 34 | - No shared variables, all the data are CPU local. |
| 35 | - If a softirq needs serialization, let it serialize itself |
| 36 | by its own spinlocks. |
| 37 | - Even if softirq is serialized, only local cpu is marked for |
| 38 | execution. Hence, we get something sort of weak cpu binding. |
| 39 | Though it is still not clear, will it result in better locality |
| 40 | or will not. |
| 41 | |
| 42 | Examples: |
| 43 | - NET RX softirq. It is multithreaded and does not require |
| 44 | any global serialization. |
| 45 | - NET TX softirq. It kicks software netdevice queues, hence |
| 46 | it is logically serialized per device, but this serialization |
| 47 | is invisible to common code. |
| 48 | - Tasklets: serialized wrt itself. |
| 49 | */ |
| 50 | |
| 51 | #ifndef __ARCH_IRQ_STAT |
| 52 | irq_cpustat_t irq_stat[NR_CPUS] ____cacheline_aligned; |
| 53 | EXPORT_SYMBOL(irq_stat); |
| 54 | #endif |
| 55 | |
| 56 | static struct softirq_action softirq_vec[NR_SOFTIRQS] __cacheline_aligned_in_smp; |
| 57 | |
| 58 | DEFINE_PER_CPU(struct task_struct *, ksoftirqd); |
| 59 | |
| 60 | const char * const softirq_to_name[NR_SOFTIRQS] = { |
| 61 | "HI", "TIMER", "NET_TX", "NET_RX", "BLOCK", "BLOCK_IOPOLL", |
| 62 | "TASKLET", "SCHED", "HRTIMER", "RCU" |
| 63 | }; |
| 64 | |
| 65 | /* |
| 66 | * we cannot loop indefinitely here to avoid userspace starvation, |
| 67 | * but we also don't want to introduce a worst case 1/HZ latency |
| 68 | * to the pending events, so lets the scheduler to balance |
| 69 | * the softirq load for us. |
| 70 | */ |
| 71 | static void wakeup_softirqd(void) |
| 72 | { |
| 73 | /* Interrupts are disabled: no need to stop preemption */ |
| 74 | struct task_struct *tsk = __this_cpu_read(ksoftirqd); |
| 75 | |
| 76 | if (tsk && tsk->state != TASK_RUNNING) |
| 77 | wake_up_process(tsk); |
| 78 | } |
| 79 | |
| 80 | /* |
Kyle Swenson | e01461f | 2021-03-15 11:14:57 -0600 | [diff] [blame] | 81 | * If ksoftirqd is scheduled, we do not want to process pending softirqs |
| 82 | * right now. Let ksoftirqd handle this at its own rate, to get fairness. |
| 83 | */ |
| 84 | static bool ksoftirqd_running(void) |
| 85 | { |
| 86 | struct task_struct *tsk = __this_cpu_read(ksoftirqd); |
| 87 | |
| 88 | return tsk && (tsk->state == TASK_RUNNING); |
| 89 | } |
| 90 | |
| 91 | /* |
Kyle Swenson | 8d8f654 | 2021-03-15 11:02:55 -0600 | [diff] [blame] | 92 | * preempt_count and SOFTIRQ_OFFSET usage: |
| 93 | * - preempt_count is changed by SOFTIRQ_OFFSET on entering or leaving |
| 94 | * softirq processing. |
| 95 | * - preempt_count is changed by SOFTIRQ_DISABLE_OFFSET (= 2 * SOFTIRQ_OFFSET) |
| 96 | * on local_bh_disable or local_bh_enable. |
| 97 | * This lets us distinguish between whether we are currently processing |
| 98 | * softirq and whether we just have bh disabled. |
| 99 | */ |
| 100 | |
| 101 | /* |
| 102 | * This one is for softirq.c-internal use, |
| 103 | * where hardirqs are disabled legitimately: |
| 104 | */ |
| 105 | #ifdef CONFIG_TRACE_IRQFLAGS |
| 106 | void __local_bh_disable_ip(unsigned long ip, unsigned int cnt) |
| 107 | { |
| 108 | unsigned long flags; |
| 109 | |
| 110 | WARN_ON_ONCE(in_irq()); |
| 111 | |
| 112 | raw_local_irq_save(flags); |
| 113 | /* |
| 114 | * The preempt tracer hooks into preempt_count_add and will break |
| 115 | * lockdep because it calls back into lockdep after SOFTIRQ_OFFSET |
| 116 | * is set and before current->softirq_enabled is cleared. |
| 117 | * We must manually increment preempt_count here and manually |
| 118 | * call the trace_preempt_off later. |
| 119 | */ |
| 120 | __preempt_count_add(cnt); |
| 121 | /* |
| 122 | * Were softirqs turned off above: |
| 123 | */ |
| 124 | if (softirq_count() == (cnt & SOFTIRQ_MASK)) |
| 125 | trace_softirqs_off(ip); |
| 126 | raw_local_irq_restore(flags); |
| 127 | |
| 128 | if (preempt_count() == cnt) { |
| 129 | #ifdef CONFIG_DEBUG_PREEMPT |
| 130 | current->preempt_disable_ip = get_parent_ip(CALLER_ADDR1); |
| 131 | #endif |
| 132 | trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); |
| 133 | } |
| 134 | } |
| 135 | EXPORT_SYMBOL(__local_bh_disable_ip); |
| 136 | #endif /* CONFIG_TRACE_IRQFLAGS */ |
| 137 | |
| 138 | static void __local_bh_enable(unsigned int cnt) |
| 139 | { |
| 140 | WARN_ON_ONCE(!irqs_disabled()); |
| 141 | |
| 142 | if (softirq_count() == (cnt & SOFTIRQ_MASK)) |
| 143 | trace_softirqs_on(_RET_IP_); |
| 144 | preempt_count_sub(cnt); |
| 145 | } |
| 146 | |
| 147 | /* |
| 148 | * Special-case - softirqs can safely be enabled in |
| 149 | * cond_resched_softirq(), or by __do_softirq(), |
| 150 | * without processing still-pending softirqs: |
| 151 | */ |
| 152 | void _local_bh_enable(void) |
| 153 | { |
| 154 | WARN_ON_ONCE(in_irq()); |
| 155 | __local_bh_enable(SOFTIRQ_DISABLE_OFFSET); |
| 156 | } |
| 157 | EXPORT_SYMBOL(_local_bh_enable); |
| 158 | |
| 159 | void __local_bh_enable_ip(unsigned long ip, unsigned int cnt) |
| 160 | { |
| 161 | WARN_ON_ONCE(in_irq() || irqs_disabled()); |
| 162 | #ifdef CONFIG_TRACE_IRQFLAGS |
| 163 | local_irq_disable(); |
| 164 | #endif |
| 165 | /* |
| 166 | * Are softirqs going to be turned on now: |
| 167 | */ |
| 168 | if (softirq_count() == SOFTIRQ_DISABLE_OFFSET) |
| 169 | trace_softirqs_on(ip); |
| 170 | /* |
| 171 | * Keep preemption disabled until we are done with |
| 172 | * softirq processing: |
| 173 | */ |
| 174 | preempt_count_sub(cnt - 1); |
| 175 | |
| 176 | if (unlikely(!in_interrupt() && local_softirq_pending())) { |
| 177 | /* |
| 178 | * Run softirq if any pending. And do it in its own stack |
| 179 | * as we may be calling this deep in a task call stack already. |
| 180 | */ |
| 181 | do_softirq(); |
| 182 | } |
| 183 | |
| 184 | preempt_count_dec(); |
| 185 | #ifdef CONFIG_TRACE_IRQFLAGS |
| 186 | local_irq_enable(); |
| 187 | #endif |
| 188 | preempt_check_resched(); |
| 189 | } |
| 190 | EXPORT_SYMBOL(__local_bh_enable_ip); |
| 191 | |
| 192 | /* |
| 193 | * We restart softirq processing for at most MAX_SOFTIRQ_RESTART times, |
| 194 | * but break the loop if need_resched() is set or after 2 ms. |
| 195 | * The MAX_SOFTIRQ_TIME provides a nice upper bound in most cases, but in |
| 196 | * certain cases, such as stop_machine(), jiffies may cease to |
| 197 | * increment and so we need the MAX_SOFTIRQ_RESTART limit as |
| 198 | * well to make sure we eventually return from this method. |
| 199 | * |
| 200 | * These limits have 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_TIME msecs_to_jiffies(2) |
| 206 | #define MAX_SOFTIRQ_RESTART 10 |
| 207 | |
| 208 | #ifdef CONFIG_TRACE_IRQFLAGS |
| 209 | /* |
| 210 | * When we run softirqs from irq_exit() and thus on the hardirq stack we need |
| 211 | * to keep the lockdep irq context tracking as tight as possible in order to |
| 212 | * not miss-qualify lock contexts and miss possible deadlocks. |
| 213 | */ |
| 214 | |
| 215 | static inline bool lockdep_softirq_start(void) |
| 216 | { |
| 217 | bool in_hardirq = false; |
| 218 | |
| 219 | if (trace_hardirq_context(current)) { |
| 220 | in_hardirq = true; |
| 221 | trace_hardirq_exit(); |
| 222 | } |
| 223 | |
| 224 | lockdep_softirq_enter(); |
| 225 | |
| 226 | return in_hardirq; |
| 227 | } |
| 228 | |
| 229 | static inline void lockdep_softirq_end(bool in_hardirq) |
| 230 | { |
| 231 | lockdep_softirq_exit(); |
| 232 | |
| 233 | if (in_hardirq) |
| 234 | trace_hardirq_enter(); |
| 235 | } |
| 236 | #else |
| 237 | static inline bool lockdep_softirq_start(void) { return false; } |
| 238 | static inline void lockdep_softirq_end(bool in_hardirq) { } |
| 239 | #endif |
| 240 | |
| 241 | asmlinkage __visible void __do_softirq(void) |
| 242 | { |
| 243 | unsigned long end = jiffies + MAX_SOFTIRQ_TIME; |
| 244 | unsigned long old_flags = current->flags; |
| 245 | int max_restart = MAX_SOFTIRQ_RESTART; |
| 246 | struct softirq_action *h; |
| 247 | bool in_hardirq; |
| 248 | __u32 pending; |
| 249 | int softirq_bit; |
| 250 | |
| 251 | /* |
| 252 | * Mask out PF_MEMALLOC s current task context is borrowed for the |
| 253 | * softirq. A softirq handled such as network RX might set PF_MEMALLOC |
| 254 | * again if the socket is related to swap |
| 255 | */ |
| 256 | current->flags &= ~PF_MEMALLOC; |
| 257 | |
| 258 | pending = local_softirq_pending(); |
| 259 | account_irq_enter_time(current); |
| 260 | |
| 261 | __local_bh_disable_ip(_RET_IP_, SOFTIRQ_OFFSET); |
| 262 | in_hardirq = lockdep_softirq_start(); |
| 263 | |
| 264 | restart: |
| 265 | /* Reset the pending bitmask before enabling irqs */ |
| 266 | set_softirq_pending(0); |
| 267 | |
| 268 | local_irq_enable(); |
| 269 | |
| 270 | h = softirq_vec; |
| 271 | |
| 272 | while ((softirq_bit = ffs(pending))) { |
| 273 | unsigned int vec_nr; |
| 274 | int prev_count; |
| 275 | |
| 276 | h += softirq_bit - 1; |
| 277 | |
| 278 | vec_nr = h - softirq_vec; |
| 279 | prev_count = preempt_count(); |
| 280 | |
| 281 | kstat_incr_softirqs_this_cpu(vec_nr); |
| 282 | |
| 283 | trace_softirq_entry(vec_nr); |
| 284 | h->action(h); |
| 285 | trace_softirq_exit(vec_nr); |
| 286 | if (unlikely(prev_count != preempt_count())) { |
| 287 | pr_err("huh, entered softirq %u %s %p with preempt_count %08x, exited with %08x?\n", |
| 288 | vec_nr, softirq_to_name[vec_nr], h->action, |
| 289 | prev_count, preempt_count()); |
| 290 | preempt_count_set(prev_count); |
| 291 | } |
| 292 | h++; |
| 293 | pending >>= softirq_bit; |
| 294 | } |
| 295 | |
| 296 | rcu_bh_qs(); |
| 297 | local_irq_disable(); |
| 298 | |
| 299 | pending = local_softirq_pending(); |
| 300 | if (pending) { |
| 301 | if (time_before(jiffies, end) && !need_resched() && |
| 302 | --max_restart) |
| 303 | goto restart; |
| 304 | |
| 305 | wakeup_softirqd(); |
| 306 | } |
| 307 | |
| 308 | lockdep_softirq_end(in_hardirq); |
| 309 | account_irq_exit_time(current); |
| 310 | __local_bh_enable(SOFTIRQ_OFFSET); |
| 311 | WARN_ON_ONCE(in_interrupt()); |
| 312 | tsk_restore_flags(current, old_flags, PF_MEMALLOC); |
| 313 | } |
| 314 | |
| 315 | asmlinkage __visible void do_softirq(void) |
| 316 | { |
| 317 | __u32 pending; |
| 318 | unsigned long flags; |
| 319 | |
| 320 | if (in_interrupt()) |
| 321 | return; |
| 322 | |
| 323 | local_irq_save(flags); |
| 324 | |
| 325 | pending = local_softirq_pending(); |
| 326 | |
Kyle Swenson | e01461f | 2021-03-15 11:14:57 -0600 | [diff] [blame] | 327 | if (pending && !ksoftirqd_running()) |
Kyle Swenson | 8d8f654 | 2021-03-15 11:02:55 -0600 | [diff] [blame] | 328 | do_softirq_own_stack(); |
| 329 | |
| 330 | local_irq_restore(flags); |
| 331 | } |
| 332 | |
| 333 | /* |
| 334 | * Enter an interrupt context. |
| 335 | */ |
| 336 | void irq_enter(void) |
| 337 | { |
| 338 | rcu_irq_enter(); |
| 339 | if (is_idle_task(current) && !in_interrupt()) { |
| 340 | /* |
| 341 | * Prevent raise_softirq from needlessly waking up ksoftirqd |
| 342 | * here, as softirq will be serviced on return from interrupt. |
| 343 | */ |
| 344 | local_bh_disable(); |
| 345 | tick_irq_enter(); |
| 346 | _local_bh_enable(); |
| 347 | } |
| 348 | |
| 349 | __irq_enter(); |
| 350 | } |
| 351 | |
| 352 | static inline void invoke_softirq(void) |
| 353 | { |
Kyle Swenson | e01461f | 2021-03-15 11:14:57 -0600 | [diff] [blame] | 354 | if (ksoftirqd_running()) |
| 355 | return; |
| 356 | |
Kyle Swenson | 8d8f654 | 2021-03-15 11:02:55 -0600 | [diff] [blame] | 357 | if (!force_irqthreads) { |
| 358 | #ifdef CONFIG_HAVE_IRQ_EXIT_ON_IRQ_STACK |
| 359 | /* |
| 360 | * We can safely execute softirq on the current stack if |
| 361 | * it is the irq stack, because it should be near empty |
| 362 | * at this stage. |
| 363 | */ |
| 364 | __do_softirq(); |
| 365 | #else |
| 366 | /* |
| 367 | * Otherwise, irq_exit() is called on the task stack that can |
| 368 | * be potentially deep already. So call softirq in its own stack |
| 369 | * to prevent from any overrun. |
| 370 | */ |
| 371 | do_softirq_own_stack(); |
| 372 | #endif |
| 373 | } else { |
| 374 | wakeup_softirqd(); |
| 375 | } |
| 376 | } |
| 377 | |
| 378 | static inline void tick_irq_exit(void) |
| 379 | { |
| 380 | #ifdef CONFIG_NO_HZ_COMMON |
| 381 | int cpu = smp_processor_id(); |
| 382 | |
| 383 | /* Make sure that timer wheel updates are propagated */ |
| 384 | if ((idle_cpu(cpu) && !need_resched()) || tick_nohz_full_cpu(cpu)) { |
| 385 | if (!in_interrupt()) |
| 386 | tick_nohz_irq_exit(); |
| 387 | } |
| 388 | #endif |
| 389 | } |
| 390 | |
| 391 | /* |
| 392 | * Exit an interrupt context. Process softirqs if needed and possible: |
| 393 | */ |
| 394 | void irq_exit(void) |
| 395 | { |
| 396 | #ifndef __ARCH_IRQ_EXIT_IRQS_DISABLED |
| 397 | local_irq_disable(); |
| 398 | #else |
| 399 | WARN_ON_ONCE(!irqs_disabled()); |
| 400 | #endif |
| 401 | |
| 402 | account_irq_exit_time(current); |
| 403 | preempt_count_sub(HARDIRQ_OFFSET); |
| 404 | if (!in_interrupt() && local_softirq_pending()) |
| 405 | invoke_softirq(); |
| 406 | |
| 407 | tick_irq_exit(); |
| 408 | rcu_irq_exit(); |
| 409 | trace_hardirq_exit(); /* must be last! */ |
| 410 | } |
| 411 | |
| 412 | /* |
| 413 | * This function must run with irqs disabled! |
| 414 | */ |
| 415 | inline void raise_softirq_irqoff(unsigned int nr) |
| 416 | { |
| 417 | __raise_softirq_irqoff(nr); |
| 418 | |
| 419 | /* |
| 420 | * If we're in an interrupt or softirq, we're done |
| 421 | * (this also catches softirq-disabled code). We will |
| 422 | * actually run the softirq once we return from |
| 423 | * the irq or softirq. |
| 424 | * |
| 425 | * Otherwise we wake up ksoftirqd to make sure we |
| 426 | * schedule the softirq soon. |
| 427 | */ |
| 428 | if (!in_interrupt()) |
| 429 | wakeup_softirqd(); |
| 430 | } |
| 431 | |
| 432 | void raise_softirq(unsigned int nr) |
| 433 | { |
| 434 | unsigned long flags; |
| 435 | |
| 436 | local_irq_save(flags); |
| 437 | raise_softirq_irqoff(nr); |
| 438 | local_irq_restore(flags); |
| 439 | } |
| 440 | |
| 441 | void __raise_softirq_irqoff(unsigned int nr) |
| 442 | { |
| 443 | trace_softirq_raise(nr); |
| 444 | or_softirq_pending(1UL << nr); |
| 445 | } |
| 446 | |
| 447 | void open_softirq(int nr, void (*action)(struct softirq_action *)) |
| 448 | { |
| 449 | softirq_vec[nr].action = action; |
| 450 | } |
| 451 | |
| 452 | /* |
| 453 | * Tasklets |
| 454 | */ |
| 455 | struct tasklet_head { |
| 456 | struct tasklet_struct *head; |
| 457 | struct tasklet_struct **tail; |
| 458 | }; |
| 459 | |
| 460 | static DEFINE_PER_CPU(struct tasklet_head, tasklet_vec); |
| 461 | static DEFINE_PER_CPU(struct tasklet_head, tasklet_hi_vec); |
| 462 | |
| 463 | void __tasklet_schedule(struct tasklet_struct *t) |
| 464 | { |
| 465 | unsigned long flags; |
| 466 | |
| 467 | local_irq_save(flags); |
| 468 | t->next = NULL; |
| 469 | *__this_cpu_read(tasklet_vec.tail) = t; |
| 470 | __this_cpu_write(tasklet_vec.tail, &(t->next)); |
| 471 | raise_softirq_irqoff(TASKLET_SOFTIRQ); |
| 472 | local_irq_restore(flags); |
| 473 | } |
| 474 | EXPORT_SYMBOL(__tasklet_schedule); |
| 475 | |
| 476 | void __tasklet_hi_schedule(struct tasklet_struct *t) |
| 477 | { |
| 478 | unsigned long flags; |
| 479 | |
| 480 | local_irq_save(flags); |
| 481 | t->next = NULL; |
| 482 | *__this_cpu_read(tasklet_hi_vec.tail) = t; |
| 483 | __this_cpu_write(tasklet_hi_vec.tail, &(t->next)); |
| 484 | raise_softirq_irqoff(HI_SOFTIRQ); |
| 485 | local_irq_restore(flags); |
| 486 | } |
| 487 | EXPORT_SYMBOL(__tasklet_hi_schedule); |
| 488 | |
| 489 | void __tasklet_hi_schedule_first(struct tasklet_struct *t) |
| 490 | { |
| 491 | BUG_ON(!irqs_disabled()); |
| 492 | |
| 493 | t->next = __this_cpu_read(tasklet_hi_vec.head); |
| 494 | __this_cpu_write(tasklet_hi_vec.head, t); |
| 495 | __raise_softirq_irqoff(HI_SOFTIRQ); |
| 496 | } |
| 497 | EXPORT_SYMBOL(__tasklet_hi_schedule_first); |
| 498 | |
| 499 | static void tasklet_action(struct softirq_action *a) |
| 500 | { |
| 501 | struct tasklet_struct *list; |
| 502 | |
| 503 | local_irq_disable(); |
| 504 | list = __this_cpu_read(tasklet_vec.head); |
| 505 | __this_cpu_write(tasklet_vec.head, NULL); |
| 506 | __this_cpu_write(tasklet_vec.tail, this_cpu_ptr(&tasklet_vec.head)); |
| 507 | local_irq_enable(); |
| 508 | |
| 509 | while (list) { |
| 510 | struct tasklet_struct *t = list; |
| 511 | |
| 512 | list = list->next; |
| 513 | |
| 514 | if (tasklet_trylock(t)) { |
| 515 | if (!atomic_read(&t->count)) { |
| 516 | if (!test_and_clear_bit(TASKLET_STATE_SCHED, |
| 517 | &t->state)) |
| 518 | BUG(); |
| 519 | t->func(t->data); |
| 520 | tasklet_unlock(t); |
| 521 | continue; |
| 522 | } |
| 523 | tasklet_unlock(t); |
| 524 | } |
| 525 | |
| 526 | local_irq_disable(); |
| 527 | t->next = NULL; |
| 528 | *__this_cpu_read(tasklet_vec.tail) = t; |
| 529 | __this_cpu_write(tasklet_vec.tail, &(t->next)); |
| 530 | __raise_softirq_irqoff(TASKLET_SOFTIRQ); |
| 531 | local_irq_enable(); |
| 532 | } |
| 533 | } |
| 534 | |
| 535 | static void tasklet_hi_action(struct softirq_action *a) |
| 536 | { |
| 537 | struct tasklet_struct *list; |
| 538 | |
| 539 | local_irq_disable(); |
| 540 | list = __this_cpu_read(tasklet_hi_vec.head); |
| 541 | __this_cpu_write(tasklet_hi_vec.head, NULL); |
| 542 | __this_cpu_write(tasklet_hi_vec.tail, this_cpu_ptr(&tasklet_hi_vec.head)); |
| 543 | local_irq_enable(); |
| 544 | |
| 545 | while (list) { |
| 546 | struct tasklet_struct *t = list; |
| 547 | |
| 548 | list = list->next; |
| 549 | |
| 550 | if (tasklet_trylock(t)) { |
| 551 | if (!atomic_read(&t->count)) { |
| 552 | if (!test_and_clear_bit(TASKLET_STATE_SCHED, |
| 553 | &t->state)) |
| 554 | BUG(); |
| 555 | t->func(t->data); |
| 556 | tasklet_unlock(t); |
| 557 | continue; |
| 558 | } |
| 559 | tasklet_unlock(t); |
| 560 | } |
| 561 | |
| 562 | local_irq_disable(); |
| 563 | t->next = NULL; |
| 564 | *__this_cpu_read(tasklet_hi_vec.tail) = t; |
| 565 | __this_cpu_write(tasklet_hi_vec.tail, &(t->next)); |
| 566 | __raise_softirq_irqoff(HI_SOFTIRQ); |
| 567 | local_irq_enable(); |
| 568 | } |
| 569 | } |
| 570 | |
| 571 | void tasklet_init(struct tasklet_struct *t, |
| 572 | void (*func)(unsigned long), unsigned long data) |
| 573 | { |
| 574 | t->next = NULL; |
| 575 | t->state = 0; |
| 576 | atomic_set(&t->count, 0); |
| 577 | t->func = func; |
| 578 | t->data = data; |
| 579 | } |
| 580 | EXPORT_SYMBOL(tasklet_init); |
| 581 | |
| 582 | void tasklet_kill(struct tasklet_struct *t) |
| 583 | { |
| 584 | if (in_interrupt()) |
| 585 | pr_notice("Attempt to kill tasklet from interrupt\n"); |
| 586 | |
| 587 | while (test_and_set_bit(TASKLET_STATE_SCHED, &t->state)) { |
| 588 | do { |
| 589 | yield(); |
| 590 | } while (test_bit(TASKLET_STATE_SCHED, &t->state)); |
| 591 | } |
| 592 | tasklet_unlock_wait(t); |
| 593 | clear_bit(TASKLET_STATE_SCHED, &t->state); |
| 594 | } |
| 595 | EXPORT_SYMBOL(tasklet_kill); |
| 596 | |
| 597 | /* |
| 598 | * tasklet_hrtimer |
| 599 | */ |
| 600 | |
| 601 | /* |
| 602 | * The trampoline is called when the hrtimer expires. It schedules a tasklet |
| 603 | * to run __tasklet_hrtimer_trampoline() which in turn will call the intended |
| 604 | * hrtimer callback, but from softirq context. |
| 605 | */ |
| 606 | static enum hrtimer_restart __hrtimer_tasklet_trampoline(struct hrtimer *timer) |
| 607 | { |
| 608 | struct tasklet_hrtimer *ttimer = |
| 609 | container_of(timer, struct tasklet_hrtimer, timer); |
| 610 | |
| 611 | tasklet_hi_schedule(&ttimer->tasklet); |
| 612 | return HRTIMER_NORESTART; |
| 613 | } |
| 614 | |
| 615 | /* |
| 616 | * Helper function which calls the hrtimer callback from |
| 617 | * tasklet/softirq context |
| 618 | */ |
| 619 | static void __tasklet_hrtimer_trampoline(unsigned long data) |
| 620 | { |
| 621 | struct tasklet_hrtimer *ttimer = (void *)data; |
| 622 | enum hrtimer_restart restart; |
| 623 | |
| 624 | restart = ttimer->function(&ttimer->timer); |
| 625 | if (restart != HRTIMER_NORESTART) |
| 626 | hrtimer_restart(&ttimer->timer); |
| 627 | } |
| 628 | |
| 629 | /** |
| 630 | * tasklet_hrtimer_init - Init a tasklet/hrtimer combo for softirq callbacks |
| 631 | * @ttimer: tasklet_hrtimer which is initialized |
| 632 | * @function: hrtimer callback function which gets called from softirq context |
| 633 | * @which_clock: clock id (CLOCK_MONOTONIC/CLOCK_REALTIME) |
| 634 | * @mode: hrtimer mode (HRTIMER_MODE_ABS/HRTIMER_MODE_REL) |
| 635 | */ |
| 636 | void tasklet_hrtimer_init(struct tasklet_hrtimer *ttimer, |
| 637 | enum hrtimer_restart (*function)(struct hrtimer *), |
| 638 | clockid_t which_clock, enum hrtimer_mode mode) |
| 639 | { |
| 640 | hrtimer_init(&ttimer->timer, which_clock, mode); |
| 641 | ttimer->timer.function = __hrtimer_tasklet_trampoline; |
| 642 | tasklet_init(&ttimer->tasklet, __tasklet_hrtimer_trampoline, |
| 643 | (unsigned long)ttimer); |
| 644 | ttimer->function = function; |
| 645 | } |
| 646 | EXPORT_SYMBOL_GPL(tasklet_hrtimer_init); |
| 647 | |
| 648 | void __init softirq_init(void) |
| 649 | { |
| 650 | int cpu; |
| 651 | |
| 652 | for_each_possible_cpu(cpu) { |
| 653 | per_cpu(tasklet_vec, cpu).tail = |
| 654 | &per_cpu(tasklet_vec, cpu).head; |
| 655 | per_cpu(tasklet_hi_vec, cpu).tail = |
| 656 | &per_cpu(tasklet_hi_vec, cpu).head; |
| 657 | } |
| 658 | |
| 659 | open_softirq(TASKLET_SOFTIRQ, tasklet_action); |
| 660 | open_softirq(HI_SOFTIRQ, tasklet_hi_action); |
| 661 | } |
| 662 | |
| 663 | static int ksoftirqd_should_run(unsigned int cpu) |
| 664 | { |
| 665 | return local_softirq_pending(); |
| 666 | } |
| 667 | |
| 668 | static void run_ksoftirqd(unsigned int cpu) |
| 669 | { |
| 670 | local_irq_disable(); |
| 671 | if (local_softirq_pending()) { |
| 672 | /* |
| 673 | * We can safely run softirq on inline stack, as we are not deep |
| 674 | * in the task stack here. |
| 675 | */ |
| 676 | __do_softirq(); |
| 677 | local_irq_enable(); |
| 678 | cond_resched_rcu_qs(); |
| 679 | return; |
| 680 | } |
| 681 | local_irq_enable(); |
| 682 | } |
| 683 | |
| 684 | #ifdef CONFIG_HOTPLUG_CPU |
| 685 | /* |
| 686 | * tasklet_kill_immediate is called to remove a tasklet which can already be |
| 687 | * scheduled for execution on @cpu. |
| 688 | * |
| 689 | * Unlike tasklet_kill, this function removes the tasklet |
| 690 | * _immediately_, even if the tasklet is in TASKLET_STATE_SCHED state. |
| 691 | * |
| 692 | * When this function is called, @cpu must be in the CPU_DEAD state. |
| 693 | */ |
| 694 | void tasklet_kill_immediate(struct tasklet_struct *t, unsigned int cpu) |
| 695 | { |
| 696 | struct tasklet_struct **i; |
| 697 | |
| 698 | BUG_ON(cpu_online(cpu)); |
| 699 | BUG_ON(test_bit(TASKLET_STATE_RUN, &t->state)); |
| 700 | |
| 701 | if (!test_bit(TASKLET_STATE_SCHED, &t->state)) |
| 702 | return; |
| 703 | |
| 704 | /* CPU is dead, so no lock needed. */ |
| 705 | for (i = &per_cpu(tasklet_vec, cpu).head; *i; i = &(*i)->next) { |
| 706 | if (*i == t) { |
| 707 | *i = t->next; |
| 708 | /* If this was the tail element, move the tail ptr */ |
| 709 | if (*i == NULL) |
| 710 | per_cpu(tasklet_vec, cpu).tail = i; |
| 711 | return; |
| 712 | } |
| 713 | } |
| 714 | BUG(); |
| 715 | } |
| 716 | |
| 717 | static void takeover_tasklets(unsigned int cpu) |
| 718 | { |
| 719 | /* CPU is dead, so no lock needed. */ |
| 720 | local_irq_disable(); |
| 721 | |
| 722 | /* Find end, append list for that CPU. */ |
| 723 | if (&per_cpu(tasklet_vec, cpu).head != per_cpu(tasklet_vec, cpu).tail) { |
| 724 | *__this_cpu_read(tasklet_vec.tail) = per_cpu(tasklet_vec, cpu).head; |
| 725 | this_cpu_write(tasklet_vec.tail, per_cpu(tasklet_vec, cpu).tail); |
| 726 | per_cpu(tasklet_vec, cpu).head = NULL; |
| 727 | per_cpu(tasklet_vec, cpu).tail = &per_cpu(tasklet_vec, cpu).head; |
| 728 | } |
| 729 | raise_softirq_irqoff(TASKLET_SOFTIRQ); |
| 730 | |
| 731 | if (&per_cpu(tasklet_hi_vec, cpu).head != per_cpu(tasklet_hi_vec, cpu).tail) { |
| 732 | *__this_cpu_read(tasklet_hi_vec.tail) = per_cpu(tasklet_hi_vec, cpu).head; |
| 733 | __this_cpu_write(tasklet_hi_vec.tail, per_cpu(tasklet_hi_vec, cpu).tail); |
| 734 | per_cpu(tasklet_hi_vec, cpu).head = NULL; |
| 735 | per_cpu(tasklet_hi_vec, cpu).tail = &per_cpu(tasklet_hi_vec, cpu).head; |
| 736 | } |
| 737 | raise_softirq_irqoff(HI_SOFTIRQ); |
| 738 | |
| 739 | local_irq_enable(); |
| 740 | } |
| 741 | #endif /* CONFIG_HOTPLUG_CPU */ |
| 742 | |
| 743 | static int cpu_callback(struct notifier_block *nfb, unsigned long action, |
| 744 | void *hcpu) |
| 745 | { |
| 746 | switch (action) { |
| 747 | #ifdef CONFIG_HOTPLUG_CPU |
| 748 | case CPU_DEAD: |
| 749 | case CPU_DEAD_FROZEN: |
| 750 | takeover_tasklets((unsigned long)hcpu); |
| 751 | break; |
| 752 | #endif /* CONFIG_HOTPLUG_CPU */ |
| 753 | } |
| 754 | return NOTIFY_OK; |
| 755 | } |
| 756 | |
| 757 | static struct notifier_block cpu_nfb = { |
| 758 | .notifier_call = cpu_callback |
| 759 | }; |
| 760 | |
| 761 | static struct smp_hotplug_thread softirq_threads = { |
| 762 | .store = &ksoftirqd, |
| 763 | .thread_should_run = ksoftirqd_should_run, |
| 764 | .thread_fn = run_ksoftirqd, |
| 765 | .thread_comm = "ksoftirqd/%u", |
| 766 | }; |
| 767 | |
| 768 | static __init int spawn_ksoftirqd(void) |
| 769 | { |
| 770 | register_cpu_notifier(&cpu_nfb); |
| 771 | |
| 772 | BUG_ON(smpboot_register_percpu_thread(&softirq_threads)); |
| 773 | |
| 774 | return 0; |
| 775 | } |
| 776 | early_initcall(spawn_ksoftirqd); |
| 777 | |
| 778 | /* |
| 779 | * [ These __weak aliases are kept in a separate compilation unit, so that |
| 780 | * GCC does not inline them incorrectly. ] |
| 781 | */ |
| 782 | |
| 783 | int __init __weak early_irq_init(void) |
| 784 | { |
| 785 | return 0; |
| 786 | } |
| 787 | |
| 788 | int __init __weak arch_probe_nr_irqs(void) |
| 789 | { |
| 790 | return NR_IRQS_LEGACY; |
| 791 | } |
| 792 | |
| 793 | int __init __weak arch_early_irq_init(void) |
| 794 | { |
| 795 | return 0; |
| 796 | } |
| 797 | |
| 798 | unsigned int __weak arch_dynirq_lower_bound(unsigned int from) |
| 799 | { |
| 800 | return from; |
| 801 | } |