FILE
timer.c
File preview with syntax highlighting.
linux-6.19.14/kernel/time/timer.c
1// SPDX-License-Identifier: GPL-2.0 2/* 3 * Kernel internal timers 4 * 5 * Copyright (C) 1991, 1992 Linus Torvalds 6 * 7 * 1997-01-28 Modified by Finn Arne Gangstad to make timers scale better. 8 * 9 * 1997-09-10 Updated NTP code according to technical memorandum Jan '96 10 * "A Kernel Model for Precision Timekeeping" by Dave Mills 11 * 1998-12-24 Fixed a xtime SMP race (we need the xtime_lock rw spinlock to 12 * serialize accesses to xtime/lost_ticks). 13 * Copyright (C) 1998 Andrea Arcangeli 14 * 1999-03-10 Improved NTP compatibility by Ulrich Windl 15 * 2002-05-31 Move sys_sysinfo here and make its locking sane, Robert Love 16 * 2000-10-05 Implemented scalable SMP per-CPU timer handling. 17 * Copyright (C) 2000, 2001, 2002 Ingo Molnar 18 * Designed by David S. Miller, Alexey Kuznetsov and Ingo Molnar 19 */ 20 21#include <linux/kernel_stat.h> 22#include <linux/export.h> 23#include <linux/interrupt.h> 24#include <linux/percpu.h> 25#include <linux/init.h> 26#include <linux/mm.h> 27#include <linux/swap.h> 28#include <linux/pid_namespace.h> 29#include <linux/notifier.h> 30#include <linux/thread_info.h> 31#include <linux/time.h> 32#include <linux/jiffies.h> 33#include <linux/posix-timers.h> 34#include <linux/cpu.h> 35#include <linux/syscalls.h> 36#include <linux/delay.h> 37#include <linux/tick.h> 38#include <linux/kallsyms.h> 39#include <linux/irq_work.h> 40#include <linux/sched/sysctl.h> 41#include <linux/sched/nohz.h> 42#include <linux/sched/debug.h> 43#include <linux/slab.h> 44#include <linux/compat.h> 45#include <linux/random.h> 46#include <linux/sysctl.h> 47 48#include <linux/uaccess.h> 49#include <asm/unistd.h> 50#include <asm/div64.h> 51#include <asm/timex.h> 52#include <asm/io.h> 53 54#include "tick-internal.h" 55#include "timer_migration.h" 56 57#define CREATE_TRACE_POINTS 58#include <trace/events/timer.h> 59 60__visible u64 jiffies_64 __cacheline_aligned_in_smp = INITIAL_JIFFIES; 61 62EXPORT_SYMBOL(jiffies_64); 63 64/* 65 * The timer wheel has LVL_DEPTH array levels. Each level provides an array of 66 * LVL_SIZE buckets. Each level is driven by its own clock and therefore each 67 * level has a different granularity. 68 * 69 * The level granularity is: LVL_CLK_DIV ^ level 70 * The level clock frequency is: HZ / (LVL_CLK_DIV ^ level) 71 * 72 * The array level of a newly armed timer depends on the relative expiry 73 * time. The farther the expiry time is away the higher the array level and 74 * therefore the granularity becomes. 75 * 76 * Contrary to the original timer wheel implementation, which aims for 'exact' 77 * expiry of the timers, this implementation removes the need for recascading 78 * the timers into the lower array levels. The previous 'classic' timer wheel 79 * implementation of the kernel already violated the 'exact' expiry by adding 80 * slack to the expiry time to provide batched expiration. The granularity 81 * levels provide implicit batching. 82 * 83 * This is an optimization of the original timer wheel implementation for the 84 * majority of the timer wheel use cases: timeouts. The vast majority of 85 * timeout timers (networking, disk I/O ...) are canceled before expiry. If 86 * the timeout expires it indicates that normal operation is disturbed, so it 87 * does not matter much whether the timeout comes with a slight delay. 88 * 89 * The only exception to this are networking timers with a small expiry 90 * time. They rely on the granularity. Those fit into the first wheel level, 91 * which has HZ granularity. 92 * 93 * We don't have cascading anymore. timers with a expiry time above the 94 * capacity of the last wheel level are force expired at the maximum timeout 95 * value of the last wheel level. From data sampling we know that the maximum 96 * value observed is 5 days (network connection tracking), so this should not 97 * be an issue. 98 * 99 * The currently chosen array constants values are a good compromise between 100 * array size and granularity. 101 * 102 * This results in the following granularity and range levels: 103 * 104 * HZ 1000 steps 105 * Level Offset Granularity Range 106 * 0 0 1 ms 0 ms - 63 ms 107 * 1 64 8 ms 64 ms - 511 ms 108 * 2 128 64 ms 512 ms - 4095 ms (512ms - ~4s) 109 * 3 192 512 ms 4096 ms - 32767 ms (~4s - ~32s) 110 * 4 256 4096 ms (~4s) 32768 ms - 262143 ms (~32s - ~4m) 111 * 5 320 32768 ms (~32s) 262144 ms - 2097151 ms (~4m - ~34m) 112 * 6 384 262144 ms (~4m) 2097152 ms - 16777215 ms (~34m - ~4h) 113 * 7 448 2097152 ms (~34m) 16777216 ms - 134217727 ms (~4h - ~1d) 114 * 8 512 16777216 ms (~4h) 134217728 ms - 1073741822 ms (~1d - ~12d) 115 * 116 * HZ 300 117 * Level Offset Granularity Range 118 * 0 0 3 ms 0 ms - 210 ms 119 * 1 64 26 ms 213 ms - 1703 ms (213ms - ~1s) 120 * 2 128 213 ms 1706 ms - 13650 ms (~1s - ~13s) 121 * 3 192 1706 ms (~1s) 13653 ms - 109223 ms (~13s - ~1m) 122 * 4 256 13653 ms (~13s) 109226 ms - 873810 ms (~1m - ~14m) 123 * 5 320 109226 ms (~1m) 873813 ms - 6990503 ms (~14m - ~1h) 124 * 6 384 873813 ms (~14m) 6990506 ms - 55924050 ms (~1h - ~15h) 125 * 7 448 6990506 ms (~1h) 55924053 ms - 447392423 ms (~15h - ~5d) 126 * 8 512 55924053 ms (~15h) 447392426 ms - 3579139406 ms (~5d - ~41d) 127 * 128 * HZ 250 129 * Level Offset Granularity Range 130 * 0 0 4 ms 0 ms - 255 ms 131 * 1 64 32 ms 256 ms - 2047 ms (256ms - ~2s) 132 * 2 128 256 ms 2048 ms - 16383 ms (~2s - ~16s) 133 * 3 192 2048 ms (~2s) 16384 ms - 131071 ms (~16s - ~2m) 134 * 4 256 16384 ms (~16s) 131072 ms - 1048575 ms (~2m - ~17m) 135 * 5 320 131072 ms (~2m) 1048576 ms - 8388607 ms (~17m - ~2h) 136 * 6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h) 137 * 7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d) 138 * 8 512 67108864 ms (~18h) 536870912 ms - 4294967288 ms (~6d - ~49d) 139 * 140 * HZ 100 141 * Level Offset Granularity Range 142 * 0 0 10 ms 0 ms - 630 ms 143 * 1 64 80 ms 640 ms - 5110 ms (640ms - ~5s) 144 * 2 128 640 ms 5120 ms - 40950 ms (~5s - ~40s) 145 * 3 192 5120 ms (~5s) 40960 ms - 327670 ms (~40s - ~5m) 146 * 4 256 40960 ms (~40s) 327680 ms - 2621430 ms (~5m - ~43m) 147 * 5 320 327680 ms (~5m) 2621440 ms - 20971510 ms (~43m - ~5h) 148 * 6 384 2621440 ms (~43m) 20971520 ms - 167772150 ms (~5h - ~1d) 149 * 7 448 20971520 ms (~5h) 167772160 ms - 1342177270 ms (~1d - ~15d) 150 */ 151 152/* Clock divisor for the next level */ 153#define LVL_CLK_SHIFT 3 154#define LVL_CLK_DIV (1UL << LVL_CLK_SHIFT) 155#define LVL_CLK_MASK (LVL_CLK_DIV - 1) 156#define LVL_SHIFT(n) ((n) * LVL_CLK_SHIFT) 157#define LVL_GRAN(n) (1UL << LVL_SHIFT(n)) 158 159/* 160 * The time start value for each level to select the bucket at enqueue 161 * time. We start from the last possible delta of the previous level 162 * so that we can later add an extra LVL_GRAN(n) to n (see calc_index()). 163 */ 164#define LVL_START(n) ((LVL_SIZE - 1) << (((n) - 1) * LVL_CLK_SHIFT)) 165 166/* Size of each clock level */ 167#define LVL_BITS 6 168#define LVL_SIZE (1UL << LVL_BITS) 169#define LVL_MASK (LVL_SIZE - 1) 170#define LVL_OFFS(n) ((n) * LVL_SIZE) 171 172/* Level depth */ 173#if HZ > 100 174# define LVL_DEPTH 9 175# else 176# define LVL_DEPTH 8 177#endif 178 179/* The cutoff (max. capacity of the wheel) */ 180#define WHEEL_TIMEOUT_CUTOFF (LVL_START(LVL_DEPTH)) 181#define WHEEL_TIMEOUT_MAX (WHEEL_TIMEOUT_CUTOFF - LVL_GRAN(LVL_DEPTH - 1)) 182 183/* 184 * The resulting wheel size. If NOHZ is configured we allocate two 185 * wheels so we have a separate storage for the deferrable timers. 186 */ 187#define WHEEL_SIZE (LVL_SIZE * LVL_DEPTH) 188 189#ifdef CONFIG_NO_HZ_COMMON 190/* 191 * If multiple bases need to be locked, use the base ordering for lock 192 * nesting, i.e. lowest number first. 193 */ 194# define NR_BASES 3 195# define BASE_LOCAL 0 196# define BASE_GLOBAL 1 197# define BASE_DEF 2 198#else 199# define NR_BASES 1 200# define BASE_LOCAL 0 201# define BASE_GLOBAL 0 202# define BASE_DEF 0 203#endif 204 205/** 206 * struct timer_base - Per CPU timer base (number of base depends on config) 207 * @lock: Lock protecting the timer_base 208 * @running_timer: When expiring timers, the lock is dropped. To make 209 * sure not to race against deleting/modifying a 210 * currently running timer, the pointer is set to the 211 * timer, which expires at the moment. If no timer is 212 * running, the pointer is NULL. 213 * @expiry_lock: PREEMPT_RT only: Lock is taken in softirq around 214 * timer expiry callback execution and when trying to 215 * delete a running timer and it wasn't successful in 216 * the first glance. It prevents priority inversion 217 * when callback was preempted on a remote CPU and a 218 * caller tries to delete the running timer. It also 219 * prevents a life lock, when the task which tries to 220 * delete a timer preempted the softirq thread which 221 * is running the timer callback function. 222 * @timer_waiters: PREEMPT_RT only: Tells, if there is a waiter 223 * waiting for the end of the timer callback function 224 * execution. 225 * @clk: clock of the timer base; is updated before enqueue 226 * of a timer; during expiry, it is 1 offset ahead of 227 * jiffies to avoid endless requeuing to current 228 * jiffies 229 * @next_expiry: expiry value of the first timer; it is updated when 230 * finding the next timer and during enqueue; the 231 * value is not valid, when next_expiry_recalc is set 232 * @cpu: Number of CPU the timer base belongs to 233 * @next_expiry_recalc: States, whether a recalculation of next_expiry is 234 * required. Value is set true, when a timer was 235 * deleted. 236 * @is_idle: Is set, when timer_base is idle. It is triggered by NOHZ 237 * code. This state is only used in standard 238 * base. Deferrable timers, which are enqueued remotely 239 * never wake up an idle CPU. So no matter of supporting it 240 * for this base. 241 * @timers_pending: Is set, when a timer is pending in the base. It is only 242 * reliable when next_expiry_recalc is not set. 243 * @pending_map: bitmap of the timer wheel; each bit reflects a 244 * bucket of the wheel. When a bit is set, at least a 245 * single timer is enqueued in the related bucket. 246 * @vectors: Array of lists; Each array member reflects a bucket 247 * of the timer wheel. The list contains all timers 248 * which are enqueued into a specific bucket. 249 */ 250struct timer_base { 251 raw_spinlock_t lock; 252 struct timer_list *running_timer; 253#ifdef CONFIG_PREEMPT_RT 254 spinlock_t expiry_lock; 255 atomic_t timer_waiters; 256#endif 257 unsigned long clk; 258 unsigned long next_expiry; 259 unsigned int cpu; 260 bool next_expiry_recalc; 261 bool is_idle; 262 bool timers_pending; 263 DECLARE_BITMAP(pending_map, WHEEL_SIZE); 264 struct hlist_head vectors[WHEEL_SIZE]; 265} ____cacheline_aligned; 266 267static DEFINE_PER_CPU(struct timer_base, timer_bases[NR_BASES]); 268 269#ifdef CONFIG_NO_HZ_COMMON 270 271static DEFINE_STATIC_KEY_FALSE(timers_nohz_active); 272static DEFINE_MUTEX(timer_keys_mutex); 273 274static void timer_update_keys(struct work_struct *work); 275static DECLARE_WORK(timer_update_work, timer_update_keys); 276 277#ifdef CONFIG_SMP 278static unsigned int sysctl_timer_migration = 1; 279 280DEFINE_STATIC_KEY_FALSE(timers_migration_enabled); 281 282static void timers_update_migration(void) 283{ 284 if (sysctl_timer_migration && tick_nohz_active) 285 static_branch_enable(&timers_migration_enabled); 286 else 287 static_branch_disable(&timers_migration_enabled); 288} 289 290#ifdef CONFIG_SYSCTL 291static int timer_migration_handler(const struct ctl_table *table, int write, 292 void *buffer, size_t *lenp, loff_t *ppos) 293{ 294 int ret; 295 296 mutex_lock(&timer_keys_mutex); 297 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); 298 if (!ret && write) 299 timers_update_migration(); 300 mutex_unlock(&timer_keys_mutex); 301 return ret; 302} 303 304static const struct ctl_table timer_sysctl[] = { 305 { 306 .procname = "timer_migration", 307 .data = &sysctl_timer_migration, 308 .maxlen = sizeof(unsigned int), 309 .mode = 0644, 310 .proc_handler = timer_migration_handler, 311 .extra1 = SYSCTL_ZERO, 312 .extra2 = SYSCTL_ONE, 313 }, 314}; 315 316static int __init timer_sysctl_init(void) 317{ 318 register_sysctl("kernel", timer_sysctl); 319 return 0; 320} 321device_initcall(timer_sysctl_init); 322#endif /* CONFIG_SYSCTL */ 323#else /* CONFIG_SMP */ 324static inline void timers_update_migration(void) { } 325#endif /* !CONFIG_SMP */ 326 327static void timer_update_keys(struct work_struct *work) 328{ 329 mutex_lock(&timer_keys_mutex); 330 timers_update_migration(); 331 static_branch_enable(&timers_nohz_active); 332 mutex_unlock(&timer_keys_mutex); 333} 334 335void timers_update_nohz(void) 336{ 337 schedule_work(&timer_update_work); 338} 339 340static inline bool is_timers_nohz_active(void) 341{ 342 return static_branch_unlikely(&timers_nohz_active); 343} 344#else 345static inline bool is_timers_nohz_active(void) { return false; } 346#endif /* NO_HZ_COMMON */ 347 348static unsigned long round_jiffies_common(unsigned long j, int cpu, 349 bool force_up) 350{ 351 int rem; 352 unsigned long original = j; 353 354 /* 355 * We don't want all cpus firing their timers at once hitting the 356 * same lock or cachelines, so we skew each extra cpu with an extra 357 * 3 jiffies. This 3 jiffies came originally from the mm/ code which 358 * already did this. 359 * The skew is done by adding 3*cpunr, then round, then subtract this 360 * extra offset again. 361 */ 362 j += cpu * 3; 363 364 rem = j % HZ; 365 366 /* 367 * If the target jiffy is just after a whole second (which can happen 368 * due to delays of the timer irq, long irq off times etc etc) then 369 * we should round down to the whole second, not up. Use 1/4th second 370 * as cutoff for this rounding as an extreme upper bound for this. 371 * But never round down if @force_up is set. 372 */ 373 if (rem < HZ/4 && !force_up) /* round down */ 374 j = j - rem; 375 else /* round up */ 376 j = j - rem + HZ; 377 378 /* now that we have rounded, subtract the extra skew again */ 379 j -= cpu * 3; 380 381 /* 382 * Make sure j is still in the future. Otherwise return the 383 * unmodified value. 384 */ 385 return time_is_after_jiffies(j) ? j : original; 386} 387 388/** 389 * __round_jiffies_relative - function to round jiffies to a full second 390 * @j: the time in (relative) jiffies that should be rounded 391 * @cpu: the processor number on which the timeout will happen 392 * 393 * __round_jiffies_relative() rounds a time delta in the future (in jiffies) 394 * up or down to (approximately) full seconds. This is useful for timers 395 * for which the exact time they fire does not matter too much, as long as 396 * they fire approximately every X seconds. 397 * 398 * By rounding these timers to whole seconds, all such timers will fire 399 * at the same time, rather than at various times spread out. The goal 400 * of this is to have the CPU wake up less, which saves power. 401 * 402 * The exact rounding is skewed for each processor to avoid all 403 * processors firing at the exact same time, which could lead 404 * to lock contention or spurious cache line bouncing. 405 * 406 * The return value is the rounded version of the @j parameter. 407 */ 408unsigned long __round_jiffies_relative(unsigned long j, int cpu) 409{ 410 unsigned long j0 = jiffies; 411 412 /* Use j0 because jiffies might change while we run */ 413 return round_jiffies_common(j + j0, cpu, false) - j0; 414} 415EXPORT_SYMBOL_GPL(__round_jiffies_relative); 416 417/** 418 * round_jiffies - function to round jiffies to a full second 419 * @j: the time in (absolute) jiffies that should be rounded 420 * 421 * round_jiffies() rounds an absolute time in the future (in jiffies) 422 * up or down to (approximately) full seconds. This is useful for timers 423 * for which the exact time they fire does not matter too much, as long as 424 * they fire approximately every X seconds. 425 * 426 * By rounding these timers to whole seconds, all such timers will fire 427 * at the same time, rather than at various times spread out. The goal 428 * of this is to have the CPU wake up less, which saves power. 429 * 430 * The return value is the rounded version of the @j parameter. 431 */ 432unsigned long round_jiffies(unsigned long j) 433{ 434 return round_jiffies_common(j, raw_smp_processor_id(), false); 435} 436EXPORT_SYMBOL_GPL(round_jiffies); 437 438/** 439 * round_jiffies_relative - function to round jiffies to a full second 440 * @j: the time in (relative) jiffies that should be rounded 441 * 442 * round_jiffies_relative() rounds a time delta in the future (in jiffies) 443 * up or down to (approximately) full seconds. This is useful for timers 444 * for which the exact time they fire does not matter too much, as long as 445 * they fire approximately every X seconds. 446 * 447 * By rounding these timers to whole seconds, all such timers will fire 448 * at the same time, rather than at various times spread out. The goal 449 * of this is to have the CPU wake up less, which saves power. 450 * 451 * The return value is the rounded version of the @j parameter. 452 */ 453unsigned long round_jiffies_relative(unsigned long j) 454{ 455 return __round_jiffies_relative(j, raw_smp_processor_id()); 456} 457EXPORT_SYMBOL_GPL(round_jiffies_relative); 458 459/** 460 * __round_jiffies_up_relative - function to round jiffies up to a full second 461 * @j: the time in (relative) jiffies that should be rounded 462 * @cpu: the processor number on which the timeout will happen 463 * 464 * This is the same as __round_jiffies_relative() except that it will never 465 * round down. This is useful for timeouts for which the exact time 466 * of firing does not matter too much, as long as they don't fire too 467 * early. 468 */ 469unsigned long __round_jiffies_up_relative(unsigned long j, int cpu) 470{ 471 unsigned long j0 = jiffies; 472 473 /* Use j0 because jiffies might change while we run */ 474 return round_jiffies_common(j + j0, cpu, true) - j0; 475} 476EXPORT_SYMBOL_GPL(__round_jiffies_up_relative); 477 478/** 479 * round_jiffies_up - function to round jiffies up to a full second 480 * @j: the time in (absolute) jiffies that should be rounded 481 * 482 * This is the same as round_jiffies() except that it will never 483 * round down. This is useful for timeouts for which the exact time 484 * of firing does not matter too much, as long as they don't fire too 485 * early. 486 */ 487unsigned long round_jiffies_up(unsigned long j) 488{ 489 return round_jiffies_common(j, raw_smp_processor_id(), true); 490} 491EXPORT_SYMBOL_GPL(round_jiffies_up); 492 493/** 494 * round_jiffies_up_relative - function to round jiffies up to a full second 495 * @j: the time in (relative) jiffies that should be rounded 496 * 497 * This is the same as round_jiffies_relative() except that it will never 498 * round down. This is useful for timeouts for which the exact time 499 * of firing does not matter too much, as long as they don't fire too 500 * early. 501 */ 502unsigned long round_jiffies_up_relative(unsigned long j) 503{ 504 return __round_jiffies_up_relative(j, raw_smp_processor_id()); 505} 506EXPORT_SYMBOL_GPL(round_jiffies_up_relative); 507 508 509static inline unsigned int timer_get_idx(struct timer_list *timer) 510{ 511 return (timer->flags & TIMER_ARRAYMASK) >> TIMER_ARRAYSHIFT; 512} 513 514static inline void timer_set_idx(struct timer_list *timer, unsigned int idx) 515{ 516 timer->flags = (timer->flags & ~TIMER_ARRAYMASK) | 517 idx << TIMER_ARRAYSHIFT; 518} 519 520/* 521 * Helper function to calculate the array index for a given expiry 522 * time. 523 */ 524static inline unsigned calc_index(unsigned long expires, unsigned lvl, 525 unsigned long *bucket_expiry) 526{ 527 528 /* 529 * The timer wheel has to guarantee that a timer does not fire 530 * early. Early expiry can happen due to: 531 * - Timer is armed at the edge of a tick 532 * - Truncation of the expiry time in the outer wheel levels 533 * 534 * Round up with level granularity to prevent this. 535 */ 536 expires = (expires >> LVL_SHIFT(lvl)) + 1; 537 *bucket_expiry = expires << LVL_SHIFT(lvl); 538 return LVL_OFFS(lvl) + (expires & LVL_MASK); 539} 540 541static int calc_wheel_index(unsigned long expires, unsigned long clk, 542 unsigned long *bucket_expiry) 543{ 544 unsigned long delta = expires - clk; 545 unsigned int idx; 546 547 if (delta < LVL_START(1)) { 548 idx = calc_index(expires, 0, bucket_expiry); 549 } else if (delta < LVL_START(2)) { 550 idx = calc_index(expires, 1, bucket_expiry); 551 } else if (delta < LVL_START(3)) { 552 idx = calc_index(expires, 2, bucket_expiry); 553 } else if (delta < LVL_START(4)) { 554 idx = calc_index(expires, 3, bucket_expiry); 555 } else if (delta < LVL_START(5)) { 556 idx = calc_index(expires, 4, bucket_expiry); 557 } else if (delta < LVL_START(6)) { 558 idx = calc_index(expires, 5, bucket_expiry); 559 } else if (delta < LVL_START(7)) { 560 idx = calc_index(expires, 6, bucket_expiry); 561 } else if (LVL_DEPTH > 8 && delta < LVL_START(8)) { 562 idx = calc_index(expires, 7, bucket_expiry); 563 } else if ((long) delta < 0) { 564 idx = clk & LVL_MASK; 565 *bucket_expiry = clk; 566 } else { 567 /* 568 * Force expire obscene large timeouts to expire at the 569 * capacity limit of the wheel. 570 */ 571 if (delta >= WHEEL_TIMEOUT_CUTOFF) 572 expires = clk + WHEEL_TIMEOUT_MAX; 573 574 idx = calc_index(expires, LVL_DEPTH - 1, bucket_expiry); 575 } 576 return idx; 577} 578 579static void 580trigger_dyntick_cpu(struct timer_base *base, struct timer_list *timer) 581{ 582 /* 583 * Deferrable timers do not prevent the CPU from entering dynticks and 584 * are not taken into account on the idle/nohz_full path. An IPI when a 585 * new deferrable timer is enqueued will wake up the remote CPU but 586 * nothing will be done with the deferrable timer base. Therefore skip 587 * the remote IPI for deferrable timers completely. 588 */ 589 if (!is_timers_nohz_active() || timer->flags & TIMER_DEFERRABLE) 590 return; 591 592 /* 593 * We might have to IPI the remote CPU if the base is idle and the 594 * timer is pinned. If it is a non pinned timer, it is only queued 595 * on the remote CPU, when timer was running during queueing. Then 596 * everything is handled by remote CPU anyway. If the other CPU is 597 * on the way to idle then it can't set base->is_idle as we hold 598 * the base lock: 599 */ 600 if (base->is_idle) { 601 WARN_ON_ONCE(!(timer->flags & TIMER_PINNED || 602 tick_nohz_full_cpu(base->cpu))); 603 wake_up_nohz_cpu(base->cpu); 604 } 605} 606 607/* 608 * Enqueue the timer into the hash bucket, mark it pending in 609 * the bitmap, store the index in the timer flags then wake up 610 * the target CPU if needed. 611 */ 612static void enqueue_timer(struct timer_base *base, struct timer_list *timer, 613 unsigned int idx, unsigned long bucket_expiry) 614{ 615 616 hlist_add_head(&timer->entry, base->vectors + idx); 617 __set_bit(idx, base->pending_map); 618 timer_set_idx(timer, idx); 619 620 trace_timer_start(timer, bucket_expiry); 621 622 /* 623 * Check whether this is the new first expiring timer. The 624 * effective expiry time of the timer is required here 625 * (bucket_expiry) instead of timer->expires. 626 */ 627 if (time_before(bucket_expiry, base->next_expiry)) { 628 /* 629 * Set the next expiry time and kick the CPU so it 630 * can reevaluate the wheel: 631 */ 632 WRITE_ONCE(base->next_expiry, bucket_expiry); 633 base->timers_pending = true; 634 base->next_expiry_recalc = false; 635 trigger_dyntick_cpu(base, timer); 636 } 637} 638 639static void internal_add_timer(struct timer_base *base, struct timer_list *timer) 640{ 641 unsigned long bucket_expiry; 642 unsigned int idx; 643 644 idx = calc_wheel_index(timer->expires, base->clk, &bucket_expiry); 645 enqueue_timer(base, timer, idx, bucket_expiry); 646} 647 648#ifdef CONFIG_DEBUG_OBJECTS_TIMERS 649 650static const struct debug_obj_descr timer_debug_descr; 651 652struct timer_hint { 653 void (*function)(struct timer_list *t); 654 long offset; 655}; 656 657#define TIMER_HINT(fn, container, timr, hintfn) \ 658 { \ 659 .function = fn, \ 660 .offset = offsetof(container, hintfn) - \ 661 offsetof(container, timr) \ 662 } 663 664static const struct timer_hint timer_hints[] = { 665 TIMER_HINT(delayed_work_timer_fn, 666 struct delayed_work, timer, work.func), 667 TIMER_HINT(kthread_delayed_work_timer_fn, 668 struct kthread_delayed_work, timer, work.func), 669}; 670 671static void *timer_debug_hint(void *addr) 672{ 673 struct timer_list *timer = addr; 674 int i; 675 676 for (i = 0; i < ARRAY_SIZE(timer_hints); i++) { 677 if (timer_hints[i].function == timer->function) { 678 void (**fn)(void) = addr + timer_hints[i].offset; 679 680 return *fn; 681 } 682 } 683 684 return timer->function; 685} 686 687static bool timer_is_static_object(void *addr) 688{ 689 struct timer_list *timer = addr; 690 691 return (timer->entry.pprev == NULL && 692 timer->entry.next == TIMER_ENTRY_STATIC); 693} 694 695/* 696 * timer_fixup_init is called when: 697 * - an active object is initialized 698 */ 699static bool timer_fixup_init(void *addr, enum debug_obj_state state) 700{ 701 struct timer_list *timer = addr; 702 703 switch (state) { 704 case ODEBUG_STATE_ACTIVE: 705 timer_delete_sync(timer); 706 debug_object_init(timer, &timer_debug_descr); 707 return true; 708 default: 709 return false; 710 } 711} 712 713/* Stub timer callback for improperly used timers. */ 714static void stub_timer(struct timer_list *unused) 715{ 716 WARN_ON(1); 717} 718 719/* 720 * timer_fixup_activate is called when: 721 * - an active object is activated 722 * - an unknown non-static object is activated 723 */ 724static bool timer_fixup_activate(void *addr, enum debug_obj_state state) 725{ 726 struct timer_list *timer = addr; 727 728 switch (state) { 729 case ODEBUG_STATE_NOTAVAILABLE: 730 timer_setup(timer, stub_timer, 0); 731 return true; 732 733 case ODEBUG_STATE_ACTIVE: 734 WARN_ON(1); 735 fallthrough; 736 default: 737 return false; 738 } 739} 740 741/* 742 * timer_fixup_free is called when: 743 * - an active object is freed 744 */ 745static bool timer_fixup_free(void *addr, enum debug_obj_state state) 746{ 747 struct timer_list *timer = addr; 748 749 switch (state) { 750 case ODEBUG_STATE_ACTIVE: 751 timer_delete_sync(timer); 752 debug_object_free(timer, &timer_debug_descr); 753 return true; 754 default: 755 return false; 756 } 757} 758 759/* 760 * timer_fixup_assert_init is called when: 761 * - an untracked/uninit-ed object is found 762 */ 763static bool timer_fixup_assert_init(void *addr, enum debug_obj_state state) 764{ 765 struct timer_list *timer = addr; 766 767 switch (state) { 768 case ODEBUG_STATE_NOTAVAILABLE: 769 timer_setup(timer, stub_timer, 0); 770 return true; 771 default: 772 return false; 773 } 774} 775 776static const struct debug_obj_descr timer_debug_descr = { 777 .name = "timer_list", 778 .debug_hint = timer_debug_hint, 779 .is_static_object = timer_is_static_object, 780 .fixup_init = timer_fixup_init, 781 .fixup_activate = timer_fixup_activate, 782 .fixup_free = timer_fixup_free, 783 .fixup_assert_init = timer_fixup_assert_init, 784}; 785 786static inline void debug_timer_init(struct timer_list *timer) 787{ 788 debug_object_init(timer, &timer_debug_descr); 789} 790 791static inline void debug_timer_activate(struct timer_list *timer) 792{ 793 debug_object_activate(timer, &timer_debug_descr); 794} 795 796static inline void debug_timer_deactivate(struct timer_list *timer) 797{ 798 debug_object_deactivate(timer, &timer_debug_descr); 799} 800 801static inline void debug_timer_assert_init(struct timer_list *timer) 802{ 803 debug_object_assert_init(timer, &timer_debug_descr); 804} 805 806static void do_init_timer(struct timer_list *timer, 807 void (*func)(struct timer_list *), 808 unsigned int flags, 809 const char *name, struct lock_class_key *key); 810 811void timer_init_key_on_stack(struct timer_list *timer, 812 void (*func)(struct timer_list *), 813 unsigned int flags, 814 const char *name, struct lock_class_key *key) 815{ 816 debug_object_init_on_stack(timer, &timer_debug_descr); 817 do_init_timer(timer, func, flags, name, key); 818} 819EXPORT_SYMBOL_GPL(timer_init_key_on_stack); 820 821void timer_destroy_on_stack(struct timer_list *timer) 822{ 823 debug_object_free(timer, &timer_debug_descr); 824} 825EXPORT_SYMBOL_GPL(timer_destroy_on_stack); 826 827#else 828static inline void debug_timer_init(struct timer_list *timer) { } 829static inline void debug_timer_activate(struct timer_list *timer) { } 830static inline void debug_timer_deactivate(struct timer_list *timer) { } 831static inline void debug_timer_assert_init(struct timer_list *timer) { } 832#endif 833 834static inline void debug_init(struct timer_list *timer) 835{ 836 debug_timer_init(timer); 837 trace_timer_init(timer); 838} 839 840static inline void debug_deactivate(struct timer_list *timer) 841{ 842 debug_timer_deactivate(timer); 843 trace_timer_cancel(timer); 844} 845 846static inline void debug_assert_init(struct timer_list *timer) 847{ 848 debug_timer_assert_init(timer); 849} 850 851static void do_init_timer(struct timer_list *timer, 852 void (*func)(struct timer_list *), 853 unsigned int flags, 854 const char *name, struct lock_class_key *key) 855{ 856 timer->entry.pprev = NULL; 857 timer->function = func; 858 if (WARN_ON_ONCE(flags & ~TIMER_INIT_FLAGS)) 859 flags &= TIMER_INIT_FLAGS; 860 timer->flags = flags | raw_smp_processor_id(); 861 lockdep_init_map(&timer->lockdep_map, name, key, 0); 862} 863 864/** 865 * timer_init_key - initialize a timer 866 * @timer: the timer to be initialized 867 * @func: timer callback function 868 * @flags: timer flags 869 * @name: name of the timer 870 * @key: lockdep class key of the fake lock used for tracking timer 871 * sync lock dependencies 872 * 873 * timer_init_key() must be done to a timer prior to calling *any* of the 874 * other timer functions. 875 */ 876void timer_init_key(struct timer_list *timer, 877 void (*func)(struct timer_list *), unsigned int flags, 878 const char *name, struct lock_class_key *key) 879{ 880 debug_init(timer); 881 do_init_timer(timer, func, flags, name, key); 882} 883EXPORT_SYMBOL(timer_init_key); 884 885static inline void detach_timer(struct timer_list *timer, bool clear_pending) 886{ 887 struct hlist_node *entry = &timer->entry; 888 889 debug_deactivate(timer); 890 891 __hlist_del(entry); 892 if (clear_pending) 893 entry->pprev = NULL; 894 entry->next = LIST_POISON2; 895} 896 897static int detach_if_pending(struct timer_list *timer, struct timer_base *base, 898 bool clear_pending) 899{ 900 unsigned idx = timer_get_idx(timer); 901 902 if (!timer_pending(timer)) 903 return 0; 904 905 if (hlist_is_singular_node(&timer->entry, base->vectors + idx)) { 906 __clear_bit(idx, base->pending_map); 907 base->next_expiry_recalc = true; 908 } 909 910 detach_timer(timer, clear_pending); 911 return 1; 912} 913 914static inline struct timer_base *get_timer_cpu_base(u32 tflags, u32 cpu) 915{ 916 int index = tflags & TIMER_PINNED ? BASE_LOCAL : BASE_GLOBAL; 917 918 /* 919 * If the timer is deferrable and NO_HZ_COMMON is set then we need 920 * to use the deferrable base. 921 */ 922 if (IS_ENABLED(CONFIG_NO_HZ_COMMON) && (tflags & TIMER_DEFERRABLE)) 923 index = BASE_DEF; 924 925 return per_cpu_ptr(&timer_bases[index], cpu); 926} 927 928static inline struct timer_base *get_timer_this_cpu_base(u32 tflags) 929{ 930 int index = tflags & TIMER_PINNED ? BASE_LOCAL : BASE_GLOBAL; 931 932 /* 933 * If the timer is deferrable and NO_HZ_COMMON is set then we need 934 * to use the deferrable base. 935 */ 936 if (IS_ENABLED(CONFIG_NO_HZ_COMMON) && (tflags & TIMER_DEFERRABLE)) 937 index = BASE_DEF; 938 939 return this_cpu_ptr(&timer_bases[index]); 940} 941 942static inline struct timer_base *get_timer_base(u32 tflags) 943{ 944 return get_timer_cpu_base(tflags, tflags & TIMER_CPUMASK); 945} 946 947static inline void __forward_timer_base(struct timer_base *base, 948 unsigned long basej) 949{ 950 /* 951 * Check whether we can forward the base. We can only do that when 952 * @basej is past base->clk otherwise we might rewind base->clk. 953 */ 954 if (time_before_eq(basej, base->clk)) 955 return; 956 957 /* 958 * If the next expiry value is > jiffies, then we fast forward to 959 * jiffies otherwise we forward to the next expiry value. 960 */ 961 if (time_after(base->next_expiry, basej)) { 962 base->clk = basej; 963 } else { 964 if (WARN_ON_ONCE(time_before(base->next_expiry, base->clk))) 965 return; 966 base->clk = base->next_expiry; 967 } 968 969} 970 971static inline void forward_timer_base(struct timer_base *base) 972{ 973 __forward_timer_base(base, READ_ONCE(jiffies)); 974} 975 976/* 977 * We are using hashed locking: Holding per_cpu(timer_bases[x]).lock means 978 * that all timers which are tied to this base are locked, and the base itself 979 * is locked too. 980 * 981 * So __run_timers/migrate_timers can safely modify all timers which could 982 * be found in the base->vectors array. 983 * 984 * When a timer is migrating then the TIMER_MIGRATING flag is set and we need 985 * to wait until the migration is done. 986 */ 987static struct timer_base *lock_timer_base(struct timer_list *timer, 988 unsigned long *flags) 989 __acquires(timer->base->lock) 990{ 991 for (;;) { 992 struct timer_base *base; 993 u32 tf; 994 995 /* 996 * We need to use READ_ONCE() here, otherwise the compiler 997 * might re-read @tf between the check for TIMER_MIGRATING 998 * and spin_lock(). 999 */ 1000 tf = READ_ONCE(timer->flags); 1001 1002 if (!(tf & TIMER_MIGRATING)) { 1003 base = get_timer_base(tf); 1004 raw_spin_lock_irqsave(&base->lock, *flags); 1005 if (timer->flags == tf) 1006 return base; 1007 raw_spin_unlock_irqrestore(&base->lock, *flags); 1008 } 1009 cpu_relax(); 1010 } 1011} 1012 1013#define MOD_TIMER_PENDING_ONLY 0x01 1014#define MOD_TIMER_REDUCE 0x02 1015#define MOD_TIMER_NOTPENDING 0x04 1016 1017static inline int 1018__mod_timer(struct timer_list *timer, unsigned long expires, unsigned int options) 1019{ 1020 unsigned long clk = 0, flags, bucket_expiry; 1021 struct timer_base *base, *new_base; 1022 unsigned int idx = UINT_MAX; 1023 int ret = 0; 1024 1025 debug_assert_init(timer); 1026 1027 /* 1028 * This is a common optimization triggered by the networking code - if 1029 * the timer is re-modified to have the same timeout or ends up in the 1030 * same array bucket then just return: 1031 */ 1032 if (!(options & MOD_TIMER_NOTPENDING) && timer_pending(timer)) { 1033 /* 1034 * The downside of this optimization is that it can result in 1035 * larger granularity than you would get from adding a new 1036 * timer with this expiry. 1037 */ 1038 long diff = timer->expires - expires; 1039 1040 if (!diff) 1041 return 1; 1042 if (options & MOD_TIMER_REDUCE && diff <= 0) 1043 return 1; 1044 1045 /* 1046 * We lock timer base and calculate the bucket index right 1047 * here. If the timer ends up in the same bucket, then we 1048 * just update the expiry time and avoid the whole 1049 * dequeue/enqueue dance. 1050 */ 1051 base = lock_timer_base(timer, &flags); 1052 /* 1053 * Has @timer been shutdown? This needs to be evaluated 1054 * while holding base lock to prevent a race against the 1055 * shutdown code. 1056 */ 1057 if (!timer->function) 1058 goto out_unlock; 1059 1060 forward_timer_base(base); 1061 1062 if (timer_pending(timer) && (options & MOD_TIMER_REDUCE) && 1063 time_before_eq(timer->expires, expires)) { 1064 ret = 1; 1065 goto out_unlock; 1066 } 1067 1068 clk = base->clk; 1069 idx = calc_wheel_index(expires, clk, &bucket_expiry); 1070 1071 /* 1072 * Retrieve and compare the array index of the pending 1073 * timer. If it matches set the expiry to the new value so a 1074 * subsequent call will exit in the expires check above. 1075 */ 1076 if (idx == timer_get_idx(timer)) { 1077 if (!(options & MOD_TIMER_REDUCE)) 1078 timer->expires = expires; 1079 else if (time_after(timer->expires, expires)) 1080 timer->expires = expires; 1081 ret = 1; 1082 goto out_unlock; 1083 } 1084 } else { 1085 base = lock_timer_base(timer, &flags); 1086 /* 1087 * Has @timer been shutdown? This needs to be evaluated 1088 * while holding base lock to prevent a race against the 1089 * shutdown code. 1090 */ 1091 if (!timer->function) 1092 goto out_unlock; 1093 1094 forward_timer_base(base); 1095 } 1096 1097 ret = detach_if_pending(timer, base, false); 1098 if (!ret && (options & MOD_TIMER_PENDING_ONLY)) 1099 goto out_unlock; 1100 1101 new_base = get_timer_this_cpu_base(timer->flags); 1102 1103 if (base != new_base) { 1104 /* 1105 * We are trying to schedule the timer on the new base. 1106 * However we can't change timer's base while it is running, 1107 * otherwise timer_delete_sync() can't detect that the timer's 1108 * handler yet has not finished. This also guarantees that the 1109 * timer is serialized wrt itself. 1110 */ 1111 if (likely(base->running_timer != timer)) { 1112 /* See the comment in lock_timer_base() */ 1113 timer->flags |= TIMER_MIGRATING; 1114 1115 raw_spin_unlock(&base->lock); 1116 base = new_base; 1117 raw_spin_lock(&base->lock); 1118 WRITE_ONCE(timer->flags, 1119 (timer->flags & ~TIMER_BASEMASK) | base->cpu); 1120 forward_timer_base(base); 1121 } 1122 } 1123 1124 debug_timer_activate(timer); 1125 1126 timer->expires = expires; 1127 /* 1128 * If 'idx' was calculated above and the base time did not advance 1129 * between calculating 'idx' and possibly switching the base, only 1130 * enqueue_timer() is required. Otherwise we need to (re)calculate 1131 * the wheel index via internal_add_timer(). 1132 */ 1133 if (idx != UINT_MAX && clk == base->clk) 1134 enqueue_timer(base, timer, idx, bucket_expiry); 1135 else 1136 internal_add_timer(base, timer); 1137 1138out_unlock: 1139 raw_spin_unlock_irqrestore(&base->lock, flags); 1140 1141 return ret; 1142} 1143 1144/** 1145 * mod_timer_pending - Modify a pending timer's timeout 1146 * @timer: The pending timer to be modified 1147 * @expires: New absolute timeout in jiffies 1148 * 1149 * mod_timer_pending() is the same for pending timers as mod_timer(), but 1150 * will not activate inactive timers. 1151 * 1152 * If @timer->function == NULL then the start operation is silently 1153 * discarded. 1154 * 1155 * Return: 1156 * * %0 - The timer was inactive and not modified or was in 1157 * shutdown state and the operation was discarded 1158 * * %1 - The timer was active and requeued to expire at @expires 1159 */ 1160int mod_timer_pending(struct timer_list *timer, unsigned long expires) 1161{ 1162 return __mod_timer(timer, expires, MOD_TIMER_PENDING_ONLY); 1163} 1164EXPORT_SYMBOL(mod_timer_pending); 1165 1166/** 1167 * mod_timer - Modify a timer's timeout 1168 * @timer: The timer to be modified 1169 * @expires: New absolute timeout in jiffies 1170 * 1171 * mod_timer(timer, expires) is equivalent to: 1172 * 1173 * timer_delete(timer); timer->expires = expires; add_timer(timer); 1174 * 1175 * mod_timer() is more efficient than the above open coded sequence. In 1176 * case that the timer is inactive, the timer_delete() part is a NOP. The 1177 * timer is in any case activated with the new expiry time @expires. 1178 * 1179 * Note that if there are multiple unserialized concurrent users of the 1180 * same timer, then mod_timer() is the only safe way to modify the timeout, 1181 * since add_timer() cannot modify an already running timer. 1182 * 1183 * If @timer->function == NULL then the start operation is silently 1184 * discarded. In this case the return value is 0 and meaningless. 1185 * 1186 * Return: 1187 * * %0 - The timer was inactive and started or was in shutdown 1188 * state and the operation was discarded 1189 * * %1 - The timer was active and requeued to expire at @expires or 1190 * the timer was active and not modified because @expires did 1191 * not change the effective expiry time 1192 */ 1193int mod_timer(struct timer_list *timer, unsigned long expires) 1194{ 1195 return __mod_timer(timer, expires, 0); 1196} 1197EXPORT_SYMBOL(mod_timer); 1198 1199/** 1200 * timer_reduce - Modify a timer's timeout if it would reduce the timeout 1201 * @timer: The timer to be modified 1202 * @expires: New absolute timeout in jiffies 1203 * 1204 * timer_reduce() is very similar to mod_timer(), except that it will only 1205 * modify an enqueued timer if that would reduce the expiration time. If 1206 * @timer is not enqueued it starts the timer. 1207 * 1208 * If @timer->function == NULL then the start operation is silently 1209 * discarded. 1210 * 1211 * Return: 1212 * * %0 - The timer was inactive and started or was in shutdown 1213 * state and the operation was discarded 1214 * * %1 - The timer was active and requeued to expire at @expires or 1215 * the timer was active and not modified because @expires 1216 * did not change the effective expiry time such that the 1217 * timer would expire earlier than already scheduled 1218 */ 1219int timer_reduce(struct timer_list *timer, unsigned long expires) 1220{ 1221 return __mod_timer(timer, expires, MOD_TIMER_REDUCE); 1222} 1223EXPORT_SYMBOL(timer_reduce); 1224 1225/** 1226 * add_timer - Start a timer 1227 * @timer: The timer to be started 1228 * 1229 * Start @timer to expire at @timer->expires in the future. @timer->expires 1230 * is the absolute expiry time measured in 'jiffies'. When the timer expires 1231 * timer->function(timer) will be invoked from soft interrupt context. 1232 * 1233 * The @timer->expires and @timer->function fields must be set prior 1234 * to calling this function. 1235 * 1236 * If @timer->function == NULL then the start operation is silently 1237 * discarded. 1238 * 1239 * If @timer->expires is already in the past @timer will be queued to 1240 * expire at the next timer tick. 1241 * 1242 * This can only operate on an inactive timer. Attempts to invoke this on 1243 * an active timer are rejected with a warning. 1244 */ 1245void add_timer(struct timer_list *timer) 1246{ 1247 if (WARN_ON_ONCE(timer_pending(timer))) 1248 return; 1249 __mod_timer(timer, timer->expires, MOD_TIMER_NOTPENDING); 1250} 1251EXPORT_SYMBOL(add_timer); 1252 1253/** 1254 * add_timer_local() - Start a timer on the local CPU 1255 * @timer: The timer to be started 1256 * 1257 * Same as add_timer() except that the timer flag TIMER_PINNED is set. 1258 * 1259 * See add_timer() for further details. 1260 */ 1261void add_timer_local(struct timer_list *timer) 1262{ 1263 if (WARN_ON_ONCE(timer_pending(timer))) 1264 return; 1265 timer->flags |= TIMER_PINNED; 1266 __mod_timer(timer, timer->expires, MOD_TIMER_NOTPENDING); 1267} 1268EXPORT_SYMBOL(add_timer_local); 1269 1270/** 1271 * add_timer_global() - Start a timer without TIMER_PINNED flag set 1272 * @timer: The timer to be started 1273 * 1274 * Same as add_timer() except that the timer flag TIMER_PINNED is unset. 1275 * 1276 * See add_timer() for further details. 1277 */ 1278void add_timer_global(struct timer_list *timer) 1279{ 1280 if (WARN_ON_ONCE(timer_pending(timer))) 1281 return; 1282 timer->flags &= ~TIMER_PINNED; 1283 __mod_timer(timer, timer->expires, MOD_TIMER_NOTPENDING); 1284} 1285EXPORT_SYMBOL(add_timer_global); 1286 1287/** 1288 * add_timer_on - Start a timer on a particular CPU 1289 * @timer: The timer to be started 1290 * @cpu: The CPU to start it on 1291 * 1292 * Same as add_timer() except that it starts the timer on the given CPU and 1293 * the TIMER_PINNED flag is set. When timer shouldn't be a pinned timer in 1294 * the next round, add_timer_global() should be used instead as it unsets 1295 * the TIMER_PINNED flag. 1296 * 1297 * See add_timer() for further details. 1298 */ 1299void add_timer_on(struct timer_list *timer, int cpu) 1300{ 1301 struct timer_base *new_base, *base; 1302 unsigned long flags; 1303 1304 debug_assert_init(timer); 1305 1306 if (WARN_ON_ONCE(timer_pending(timer))) 1307 return; 1308 1309 /* Make sure timer flags have TIMER_PINNED flag set */ 1310 timer->flags |= TIMER_PINNED; 1311 1312 new_base = get_timer_cpu_base(timer->flags, cpu); 1313 1314 /* 1315 * If @timer was on a different CPU, it should be migrated with the 1316 * old base locked to prevent other operations proceeding with the 1317 * wrong base locked. See lock_timer_base(). 1318 */ 1319 base = lock_timer_base(timer, &flags); 1320 /* 1321 * Has @timer been shutdown? This needs to be evaluated while 1322 * holding base lock to prevent a race against the shutdown code. 1323 */ 1324 if (!timer->function) 1325 goto out_unlock; 1326 1327 if (base != new_base) { 1328 timer->flags |= TIMER_MIGRATING; 1329 1330 raw_spin_unlock(&base->lock); 1331 base = new_base; 1332 raw_spin_lock(&base->lock); 1333 WRITE_ONCE(timer->flags, 1334 (timer->flags & ~TIMER_BASEMASK) | cpu); 1335 } 1336 forward_timer_base(base); 1337 1338 debug_timer_activate(timer); 1339 internal_add_timer(base, timer); 1340out_unlock: 1341 raw_spin_unlock_irqrestore(&base->lock, flags); 1342} 1343EXPORT_SYMBOL_GPL(add_timer_on); 1344 1345/** 1346 * __timer_delete - Internal function: Deactivate a timer 1347 * @timer: The timer to be deactivated 1348 * @shutdown: If true, this indicates that the timer is about to be 1349 * shutdown permanently. 1350 * 1351 * If @shutdown is true then @timer->function is set to NULL under the 1352 * timer base lock which prevents further rearming of the time. In that 1353 * case any attempt to rearm @timer after this function returns will be 1354 * silently ignored. 1355 * 1356 * Return: 1357 * * %0 - The timer was not pending 1358 * * %1 - The timer was pending and deactivated 1359 */ 1360static int __timer_delete(struct timer_list *timer, bool shutdown) 1361{ 1362 struct timer_base *base; 1363 unsigned long flags; 1364 int ret = 0; 1365 1366 debug_assert_init(timer); 1367 1368 /* 1369 * If @shutdown is set then the lock has to be taken whether the 1370 * timer is pending or not to protect against a concurrent rearm 1371 * which might hit between the lockless pending check and the lock 1372 * acquisition. By taking the lock it is ensured that such a newly 1373 * enqueued timer is dequeued and cannot end up with 1374 * timer->function == NULL in the expiry code. 1375 * 1376 * If timer->function is currently executed, then this makes sure 1377 * that the callback cannot requeue the timer. 1378 */ 1379 if (timer_pending(timer) || shutdown) { 1380 base = lock_timer_base(timer, &flags); 1381 ret = detach_if_pending(timer, base, true); 1382 if (shutdown) 1383 timer->function = NULL; 1384 raw_spin_unlock_irqrestore(&base->lock, flags); 1385 } 1386 1387 return ret; 1388} 1389 1390/** 1391 * timer_delete - Deactivate a timer 1392 * @timer: The timer to be deactivated 1393 * 1394 * The function only deactivates a pending timer, but contrary to 1395 * timer_delete_sync() it does not take into account whether the timer's 1396 * callback function is concurrently executed on a different CPU or not. 1397 * It neither prevents rearming of the timer. If @timer can be rearmed 1398 * concurrently then the return value of this function is meaningless. 1399 * 1400 * Return: 1401 * * %0 - The timer was not pending 1402 * * %1 - The timer was pending and deactivated 1403 */ 1404int timer_delete(struct timer_list *timer) 1405{ 1406 return __timer_delete(timer, false); 1407} 1408EXPORT_SYMBOL(timer_delete); 1409 1410/** 1411 * timer_shutdown - Deactivate a timer and prevent rearming 1412 * @timer: The timer to be deactivated 1413 * 1414 * The function does not wait for an eventually running timer callback on a 1415 * different CPU but it prevents rearming of the timer. Any attempt to arm 1416 * @timer after this function returns will be silently ignored. 1417 * 1418 * This function is useful for teardown code and should only be used when 1419 * timer_shutdown_sync() cannot be invoked due to locking or context constraints. 1420 * 1421 * Return: 1422 * * %0 - The timer was not pending 1423 * * %1 - The timer was pending 1424 */ 1425int timer_shutdown(struct timer_list *timer) 1426{ 1427 return __timer_delete(timer, true); 1428} 1429EXPORT_SYMBOL_GPL(timer_shutdown); 1430 1431/** 1432 * __try_to_del_timer_sync - Internal function: Try to deactivate a timer 1433 * @timer: Timer to deactivate 1434 * @shutdown: If true, this indicates that the timer is about to be 1435 * shutdown permanently. 1436 * 1437 * If @shutdown is true then @timer->function is set to NULL under the 1438 * timer base lock which prevents further rearming of the timer. Any 1439 * attempt to rearm @timer after this function returns will be silently 1440 * ignored. 1441 * 1442 * This function cannot guarantee that the timer cannot be rearmed 1443 * right after dropping the base lock if @shutdown is false. That 1444 * needs to be prevented by the calling code if necessary. 1445 * 1446 * Return: 1447 * * %0 - The timer was not pending 1448 * * %1 - The timer was pending and deactivated 1449 * * %-1 - The timer callback function is running on a different CPU 1450 */ 1451static int __try_to_del_timer_sync(struct timer_list *timer, bool shutdown) 1452{ 1453 struct timer_base *base; 1454 unsigned long flags; 1455 int ret = -1; 1456 1457 debug_assert_init(timer); 1458 1459 base = lock_timer_base(timer, &flags); 1460 1461 if (base->running_timer != timer) { 1462 ret = detach_if_pending(timer, base, true); 1463 if (shutdown) 1464 timer->function = NULL; 1465 } 1466 1467 raw_spin_unlock_irqrestore(&base->lock, flags); 1468 1469 return ret; 1470} 1471 1472/** 1473 * timer_delete_sync_try - Try to deactivate a timer 1474 * @timer: Timer to deactivate 1475 * 1476 * This function tries to deactivate a timer. On success the timer is not 1477 * queued and the timer callback function is not running on any CPU. 1478 * 1479 * This function does not guarantee that the timer cannot be rearmed right 1480 * after dropping the base lock. That needs to be prevented by the calling 1481 * code if necessary. 1482 * 1483 * Return: 1484 * * %0 - The timer was not pending 1485 * * %1 - The timer was pending and deactivated 1486 * * %-1 - The timer callback function is running on a different CPU 1487 */ 1488int timer_delete_sync_try(struct timer_list *timer) 1489{ 1490 return __try_to_del_timer_sync(timer, false); 1491} 1492EXPORT_SYMBOL(timer_delete_sync_try); 1493 1494#ifdef CONFIG_PREEMPT_RT 1495static __init void timer_base_init_expiry_lock(struct timer_base *base) 1496{ 1497 spin_lock_init(&base->expiry_lock); 1498} 1499 1500static inline void timer_base_lock_expiry(struct timer_base *base) 1501{ 1502 spin_lock(&base->expiry_lock); 1503} 1504 1505static inline void timer_base_unlock_expiry(struct timer_base *base) 1506{ 1507 spin_unlock(&base->expiry_lock); 1508} 1509 1510/* 1511 * The counterpart to del_timer_wait_running(). 1512 * 1513 * If there is a waiter for base->expiry_lock, then it was waiting for the 1514 * timer callback to finish. Drop expiry_lock and reacquire it. That allows 1515 * the waiter to acquire the lock and make progress. 1516 */ 1517static void timer_sync_wait_running(struct timer_base *base) 1518 __releases(&base->lock) __releases(&base->expiry_lock) 1519 __acquires(&base->expiry_lock) __acquires(&base->lock) 1520{ 1521 if (atomic_read(&base->timer_waiters)) { 1522 raw_spin_unlock_irq(&base->lock); 1523 spin_unlock(&base->expiry_lock); 1524 spin_lock(&base->expiry_lock); 1525 raw_spin_lock_irq(&base->lock); 1526 } 1527} 1528 1529/* 1530 * This function is called on PREEMPT_RT kernels when the fast path 1531 * deletion of a timer failed because the timer callback function was 1532 * running. 1533 * 1534 * This prevents priority inversion, if the softirq thread on a remote CPU 1535 * got preempted, and it prevents a life lock when the task which tries to 1536 * delete a timer preempted the softirq thread running the timer callback 1537 * function. 1538 */ 1539static void del_timer_wait_running(struct timer_list *timer) 1540{ 1541 u32 tf; 1542 1543 tf = READ_ONCE(timer->flags); 1544 if (!(tf & (TIMER_MIGRATING | TIMER_IRQSAFE))) { 1545 struct timer_base *base = get_timer_base(tf); 1546 1547 /* 1548 * Mark the base as contended and grab the expiry lock, 1549 * which is held by the softirq across the timer 1550 * callback. Drop the lock immediately so the softirq can 1551 * expire the next timer. In theory the timer could already 1552 * be running again, but that's more than unlikely and just 1553 * causes another wait loop. 1554 */ 1555 atomic_inc(&base->timer_waiters); 1556 spin_lock_bh(&base->expiry_lock); 1557 atomic_dec(&base->timer_waiters); 1558 spin_unlock_bh(&base->expiry_lock); 1559 } 1560} 1561#else 1562static inline void timer_base_init_expiry_lock(struct timer_base *base) { } 1563static inline void timer_base_lock_expiry(struct timer_base *base) { } 1564static inline void timer_base_unlock_expiry(struct timer_base *base) { } 1565static inline void timer_sync_wait_running(struct timer_base *base) { } 1566static inline void del_timer_wait_running(struct timer_list *timer) { } 1567#endif 1568 1569/** 1570 * __timer_delete_sync - Internal function: Deactivate a timer and wait 1571 * for the handler to finish. 1572 * @timer: The timer to be deactivated 1573 * @shutdown: If true, @timer->function will be set to NULL under the 1574 * timer base lock which prevents rearming of @timer 1575 * 1576 * If @shutdown is not set the timer can be rearmed later. If the timer can 1577 * be rearmed concurrently, i.e. after dropping the base lock then the 1578 * return value is meaningless. 1579 * 1580 * If @shutdown is set then @timer->function is set to NULL under timer 1581 * base lock which prevents rearming of the timer. Any attempt to rearm 1582 * a shutdown timer is silently ignored. 1583 * 1584 * If the timer should be reused after shutdown it has to be initialized 1585 * again. 1586 * 1587 * Return: 1588 * * %0 - The timer was not pending 1589 * * %1 - The timer was pending and deactivated 1590 */ 1591static int __timer_delete_sync(struct timer_list *timer, bool shutdown) 1592{ 1593 int ret; 1594 1595#ifdef CONFIG_LOCKDEP 1596 unsigned long flags; 1597 1598 /* 1599 * If lockdep gives a backtrace here, please reference 1600 * the synchronization rules above. 1601 */ 1602 local_irq_save(flags); 1603 lock_map_acquire(&timer->lockdep_map); 1604 lock_map_release(&timer->lockdep_map); 1605 local_irq_restore(flags); 1606#endif 1607 /* 1608 * don't use it in hardirq context, because it 1609 * could lead to deadlock. 1610 */ 1611 WARN_ON(in_hardirq() && !(timer->flags & TIMER_IRQSAFE)); 1612 1613 /* 1614 * Must be able to sleep on PREEMPT_RT because of the slowpath in 1615 * del_timer_wait_running(). 1616 */ 1617 if (IS_ENABLED(CONFIG_PREEMPT_RT) && !(timer->flags & TIMER_IRQSAFE)) 1618 lockdep_assert_preemption_enabled(); 1619 1620 do { 1621 ret = __try_to_del_timer_sync(timer, shutdown); 1622 1623 if (unlikely(ret < 0)) { 1624 del_timer_wait_running(timer); 1625 cpu_relax(); 1626 } 1627 } while (ret < 0); 1628 1629 return ret; 1630} 1631 1632/** 1633 * timer_delete_sync - Deactivate a timer and wait for the handler to finish. 1634 * @timer: The timer to be deactivated 1635 * 1636 * Synchronization rules: Callers must prevent restarting of the timer, 1637 * otherwise this function is meaningless. It must not be called from 1638 * interrupt contexts unless the timer is an irqsafe one. The caller must 1639 * not hold locks which would prevent completion of the timer's callback 1640 * function. The timer's handler must not call add_timer_on(). Upon exit 1641 * the timer is not queued and the handler is not running on any CPU. 1642 * 1643 * For !irqsafe timers, the caller must not hold locks that are held in 1644 * interrupt context. Even if the lock has nothing to do with the timer in 1645 * question. Here's why:: 1646 * 1647 * CPU0 CPU1 1648 * ---- ---- 1649 * <SOFTIRQ> 1650 * call_timer_fn(); 1651 * base->running_timer = mytimer; 1652 * spin_lock_irq(somelock); 1653 * <IRQ> 1654 * spin_lock(somelock); 1655 * timer_delete_sync(mytimer); 1656 * while (base->running_timer == mytimer); 1657 * 1658 * Now timer_delete_sync() will never return and never release somelock. 1659 * The interrupt on the other CPU is waiting to grab somelock but it has 1660 * interrupted the softirq that CPU0 is waiting to finish. 1661 * 1662 * This function cannot guarantee that the timer is not rearmed again by 1663 * some concurrent or preempting code, right after it dropped the base 1664 * lock. If there is the possibility of a concurrent rearm then the return 1665 * value of the function is meaningless. 1666 * 1667 * If such a guarantee is needed, e.g. for teardown situations then use 1668 * timer_shutdown_sync() instead. 1669 * 1670 * Return: 1671 * * %0 - The timer was not pending 1672 * * %1 - The timer was pending and deactivated 1673 */ 1674int timer_delete_sync(struct timer_list *timer) 1675{ 1676 return __timer_delete_sync(timer, false); 1677} 1678EXPORT_SYMBOL(timer_delete_sync); 1679 1680/** 1681 * timer_shutdown_sync - Shutdown a timer and prevent rearming 1682 * @timer: The timer to be shutdown 1683 * 1684 * When the function returns it is guaranteed that: 1685 * - @timer is not queued 1686 * - The callback function of @timer is not running 1687 * - @timer cannot be enqueued again. Any attempt to rearm 1688 * @timer is silently ignored. 1689 * 1690 * See timer_delete_sync() for synchronization rules. 1691 * 1692 * This function is useful for final teardown of an infrastructure where 1693 * the timer is subject to a circular dependency problem. 1694 * 1695 * A common pattern for this is a timer and a workqueue where the timer can 1696 * schedule work and work can arm the timer. On shutdown the workqueue must 1697 * be destroyed and the timer must be prevented from rearming. Unless the 1698 * code has conditionals like 'if (mything->in_shutdown)' to prevent that 1699 * there is no way to get this correct with timer_delete_sync(). 1700 * 1701 * timer_shutdown_sync() is solving the problem. The correct ordering of 1702 * calls in this case is: 1703 * 1704 * timer_shutdown_sync(&mything->timer); 1705 * workqueue_destroy(&mything->workqueue); 1706 * 1707 * After this 'mything' can be safely freed. 1708 * 1709 * This obviously implies that the timer is not required to be functional 1710 * for the rest of the shutdown operation. 1711 * 1712 * Return: 1713 * * %0 - The timer was not pending 1714 * * %1 - The timer was pending 1715 */ 1716int timer_shutdown_sync(struct timer_list *timer) 1717{ 1718 return __timer_delete_sync(timer, true); 1719} 1720EXPORT_SYMBOL_GPL(timer_shutdown_sync); 1721 1722static void call_timer_fn(struct timer_list *timer, 1723 void (*fn)(struct timer_list *), 1724 unsigned long baseclk) 1725{ 1726 int count = preempt_count(); 1727 1728#ifdef CONFIG_LOCKDEP 1729 /* 1730 * It is permissible to free the timer from inside the 1731 * function that is called from it, this we need to take into 1732 * account for lockdep too. To avoid bogus "held lock freed" 1733 * warnings as well as problems when looking into 1734 * timer->lockdep_map, make a copy and use that here. 1735 */ 1736 struct lockdep_map lockdep_map; 1737 1738 lockdep_copy_map(&lockdep_map, &timer->lockdep_map); 1739#endif 1740 /* 1741 * Couple the lock chain with the lock chain at 1742 * timer_delete_sync() by acquiring the lock_map around the fn() 1743 * call here and in timer_delete_sync(). 1744 */ 1745 lock_map_acquire(&lockdep_map); 1746 1747 trace_timer_expire_entry(timer, baseclk); 1748 fn(timer); 1749 trace_timer_expire_exit(timer); 1750 1751 lock_map_release(&lockdep_map); 1752 1753 if (count != preempt_count()) { 1754 WARN_ONCE(1, "timer: %pS preempt leak: %08x -> %08x\n", 1755 fn, count, preempt_count()); 1756 /* 1757 * Restore the preempt count. That gives us a decent 1758 * chance to survive and extract information. If the 1759 * callback kept a lock held, bad luck, but not worse 1760 * than the BUG() we had. 1761 */ 1762 preempt_count_set(count); 1763 } 1764} 1765 1766static void expire_timers(struct timer_base *base, struct hlist_head *head) 1767{ 1768 /* 1769 * This value is required only for tracing. base->clk was 1770 * incremented directly before expire_timers was called. But expiry 1771 * is related to the old base->clk value. 1772 */ 1773 unsigned long baseclk = base->clk - 1; 1774 1775 while (!hlist_empty(head)) { 1776 struct timer_list *timer; 1777 void (*fn)(struct timer_list *); 1778 1779 timer = hlist_entry(head->first, struct timer_list, entry); 1780 1781 base->running_timer = timer; 1782 detach_timer(timer, true); 1783 1784 fn = timer->function; 1785 1786 if (WARN_ON_ONCE(!fn)) { 1787 /* Should never happen. Emphasis on should! */ 1788 base->running_timer = NULL; 1789 continue; 1790 } 1791 1792 if (timer->flags & TIMER_IRQSAFE) { 1793 raw_spin_unlock(&base->lock); 1794 call_timer_fn(timer, fn, baseclk); 1795 raw_spin_lock(&base->lock); 1796 base->running_timer = NULL; 1797 } else { 1798 raw_spin_unlock_irq(&base->lock); 1799 call_timer_fn(timer, fn, baseclk); 1800 raw_spin_lock_irq(&base->lock); 1801 base->running_timer = NULL; 1802 timer_sync_wait_running(base); 1803 } 1804 } 1805} 1806 1807static int collect_expired_timers(struct timer_base *base, 1808 struct hlist_head *heads) 1809{ 1810 unsigned long clk = base->clk = base->next_expiry; 1811 struct hlist_head *vec; 1812 int i, levels = 0; 1813 unsigned int idx; 1814 1815 for (i = 0; i < LVL_DEPTH; i++) { 1816 idx = (clk & LVL_MASK) + i * LVL_SIZE; 1817 1818 if (__test_and_clear_bit(idx, base->pending_map)) { 1819 vec = base->vectors + idx; 1820 hlist_move_list(vec, heads++); 1821 levels++; 1822 } 1823 /* Is it time to look at the next level? */ 1824 if (clk & LVL_CLK_MASK) 1825 break; 1826 /* Shift clock for the next level granularity */ 1827 clk >>= LVL_CLK_SHIFT; 1828 } 1829 return levels; 1830} 1831 1832/* 1833 * Find the next pending bucket of a level. Search from level start (@offset) 1834 * + @clk upwards and if nothing there, search from start of the level 1835 * (@offset) up to @offset + clk. 1836 */ 1837static int next_pending_bucket(struct timer_base *base, unsigned offset, 1838 unsigned clk) 1839{ 1840 unsigned pos, start = offset + clk; 1841 unsigned end = offset + LVL_SIZE; 1842 1843 pos = find_next_bit(base->pending_map, end, start); 1844 if (pos < end) 1845 return pos - start; 1846 1847 pos = find_next_bit(base->pending_map, start, offset); 1848 return pos < start ? pos + LVL_SIZE - start : -1; 1849} 1850 1851/* 1852 * Search the first expiring timer in the various clock levels. Caller must 1853 * hold base->lock. 1854 * 1855 * Store next expiry time in base->next_expiry. 1856 */ 1857static void timer_recalc_next_expiry(struct timer_base *base) 1858{ 1859 unsigned long clk, next, adj; 1860 unsigned lvl, offset = 0; 1861 1862 next = base->clk + TIMER_NEXT_MAX_DELTA; 1863 clk = base->clk; 1864 for (lvl = 0; lvl < LVL_DEPTH; lvl++, offset += LVL_SIZE) { 1865 int pos = next_pending_bucket(base, offset, clk & LVL_MASK); 1866 unsigned long lvl_clk = clk & LVL_CLK_MASK; 1867 1868 if (pos >= 0) { 1869 unsigned long tmp = clk + (unsigned long) pos; 1870 1871 tmp <<= LVL_SHIFT(lvl); 1872 if (time_before(tmp, next)) 1873 next = tmp; 1874 1875 /* 1876 * If the next expiration happens before we reach 1877 * the next level, no need to check further. 1878 */ 1879 if (pos <= ((LVL_CLK_DIV - lvl_clk) & LVL_CLK_MASK)) 1880 break; 1881 } 1882 /* 1883 * Clock for the next level. If the current level clock lower 1884 * bits are zero, we look at the next level as is. If not we 1885 * need to advance it by one because that's going to be the 1886 * next expiring bucket in that level. base->clk is the next 1887 * expiring jiffy. So in case of: 1888 * 1889 * LVL5 LVL4 LVL3 LVL2 LVL1 LVL0 1890 * 0 0 0 0 0 0 1891 * 1892 * we have to look at all levels @index 0. With 1893 * 1894 * LVL5 LVL4 LVL3 LVL2 LVL1 LVL0 1895 * 0 0 0 0 0 2 1896 * 1897 * LVL0 has the next expiring bucket @index 2. The upper 1898 * levels have the next expiring bucket @index 1. 1899 * 1900 * In case that the propagation wraps the next level the same 1901 * rules apply: 1902 * 1903 * LVL5 LVL4 LVL3 LVL2 LVL1 LVL0 1904 * 0 0 0 0 F 2 1905 * 1906 * So after looking at LVL0 we get: 1907 * 1908 * LVL5 LVL4 LVL3 LVL2 LVL1 1909 * 0 0 0 1 0 1910 * 1911 * So no propagation from LVL1 to LVL2 because that happened 1912 * with the add already, but then we need to propagate further 1913 * from LVL2 to LVL3. 1914 * 1915 * So the simple check whether the lower bits of the current 1916 * level are 0 or not is sufficient for all cases. 1917 */ 1918 adj = lvl_clk ? 1 : 0; 1919 clk >>= LVL_CLK_SHIFT; 1920 clk += adj; 1921 } 1922 1923 WRITE_ONCE(base->next_expiry, next); 1924 base->next_expiry_recalc = false; 1925 base->timers_pending = !(next == base->clk + TIMER_NEXT_MAX_DELTA); 1926} 1927 1928#ifdef CONFIG_NO_HZ_COMMON 1929/* 1930 * Check, if the next hrtimer event is before the next timer wheel 1931 * event: 1932 */ 1933static u64 cmp_next_hrtimer_event(u64 basem, u64 expires) 1934{ 1935 u64 nextevt = hrtimer_get_next_event(); 1936 1937 /* 1938 * If high resolution timers are enabled 1939 * hrtimer_get_next_event() returns KTIME_MAX. 1940 */ 1941 if (expires <= nextevt) 1942 return expires; 1943 1944 /* 1945 * If the next timer is already expired, return the tick base 1946 * time so the tick is fired immediately. 1947 */ 1948 if (nextevt <= basem) 1949 return basem; 1950 1951 /* 1952 * Round up to the next jiffy. High resolution timers are 1953 * off, so the hrtimers are expired in the tick and we need to 1954 * make sure that this tick really expires the timer to avoid 1955 * a ping pong of the nohz stop code. 1956 * 1957 * Use DIV_ROUND_UP_ULL to prevent gcc calling __divdi3 1958 */ 1959 return DIV_ROUND_UP_ULL(nextevt, TICK_NSEC) * TICK_NSEC; 1960} 1961 1962static unsigned long next_timer_interrupt(struct timer_base *base, 1963 unsigned long basej) 1964{ 1965 if (base->next_expiry_recalc) 1966 timer_recalc_next_expiry(base); 1967 1968 /* 1969 * Move next_expiry for the empty base into the future to prevent an 1970 * unnecessary raise of the timer softirq when the next_expiry value 1971 * will be reached even if there is no timer pending. 1972 * 1973 * This update is also required to make timer_base::next_expiry values 1974 * easy comparable to find out which base holds the first pending timer. 1975 */ 1976 if (!base->timers_pending) 1977 WRITE_ONCE(base->next_expiry, basej + TIMER_NEXT_MAX_DELTA); 1978 1979 return base->next_expiry; 1980} 1981 1982static unsigned long fetch_next_timer_interrupt(unsigned long basej, u64 basem, 1983 struct timer_base *base_local, 1984 struct timer_base *base_global, 1985 struct timer_events *tevt) 1986{ 1987 unsigned long nextevt, nextevt_local, nextevt_global; 1988 bool local_first; 1989 1990 nextevt_local = next_timer_interrupt(base_local, basej); 1991 nextevt_global = next_timer_interrupt(base_global, basej); 1992 1993 local_first = time_before_eq(nextevt_local, nextevt_global); 1994 1995 nextevt = local_first ? nextevt_local : nextevt_global; 1996 1997 /* 1998 * If the @nextevt is at max. one tick away, use @nextevt and store 1999 * it in the local expiry value. The next global event is irrelevant in 2000 * this case and can be left as KTIME_MAX. 2001 */ 2002 if (time_before_eq(nextevt, basej + 1)) { 2003 /* If we missed a tick already, force 0 delta */ 2004 if (time_before(nextevt, basej)) 2005 nextevt = basej; 2006 tevt->local = basem + (u64)(nextevt - basej) * TICK_NSEC; 2007 2008 /* 2009 * This is required for the remote check only but it doesn't 2010 * hurt, when it is done for both call sites: 2011 * 2012 * * The remote callers will only take care of the global timers 2013 * as local timers will be handled by CPU itself. When not 2014 * updating tevt->global with the already missed first global 2015 * timer, it is possible that it will be missed completely. 2016 * 2017 * * The local callers will ignore the tevt->global anyway, when 2018 * nextevt is max. one tick away. 2019 */ 2020 if (!local_first) 2021 tevt->global = tevt->local; 2022 return nextevt; 2023 } 2024 2025 /* 2026 * Update tevt.* values: 2027 * 2028 * If the local queue expires first, then the global event can be 2029 * ignored. If the global queue is empty, nothing to do either. 2030 */ 2031 if (!local_first && base_global->timers_pending) 2032 tevt->global = basem + (u64)(nextevt_global - basej) * TICK_NSEC; 2033 2034 if (base_local->timers_pending) 2035 tevt->local = basem + (u64)(nextevt_local - basej) * TICK_NSEC; 2036 2037 return nextevt; 2038} 2039 2040# ifdef CONFIG_SMP 2041/** 2042 * fetch_next_timer_interrupt_remote() - Store next timers into @tevt 2043 * @basej: base time jiffies 2044 * @basem: base time clock monotonic 2045 * @tevt: Pointer to the storage for the expiry values 2046 * @cpu: Remote CPU 2047 * 2048 * Stores the next pending local and global timer expiry values in the 2049 * struct pointed to by @tevt. If a queue is empty the corresponding 2050 * field is set to KTIME_MAX. If local event expires before global 2051 * event, global event is set to KTIME_MAX as well. 2052 * 2053 * Caller needs to make sure timer base locks are held (use 2054 * timer_lock_remote_bases() for this purpose). 2055 */ 2056void fetch_next_timer_interrupt_remote(unsigned long basej, u64 basem, 2057 struct timer_events *tevt, 2058 unsigned int cpu) 2059{ 2060 struct timer_base *base_local, *base_global; 2061 2062 /* Preset local / global events */ 2063 tevt->local = tevt->global = KTIME_MAX; 2064 2065 base_local = per_cpu_ptr(&timer_bases[BASE_LOCAL], cpu); 2066 base_global = per_cpu_ptr(&timer_bases[BASE_GLOBAL], cpu); 2067 2068 lockdep_assert_held(&base_local->lock); 2069 lockdep_assert_held(&base_global->lock); 2070 2071 fetch_next_timer_interrupt(basej, basem, base_local, base_global, tevt); 2072} 2073 2074/** 2075 * timer_unlock_remote_bases - unlock timer bases of cpu 2076 * @cpu: Remote CPU 2077 * 2078 * Unlocks the remote timer bases. 2079 */ 2080void timer_unlock_remote_bases(unsigned int cpu) 2081 __releases(timer_bases[BASE_LOCAL]->lock) 2082 __releases(timer_bases[BASE_GLOBAL]->lock) 2083{ 2084 struct timer_base *base_local, *base_global; 2085 2086 base_local = per_cpu_ptr(&timer_bases[BASE_LOCAL], cpu); 2087 base_global = per_cpu_ptr(&timer_bases[BASE_GLOBAL], cpu); 2088 2089 raw_spin_unlock(&base_global->lock); 2090 raw_spin_unlock(&base_local->lock); 2091} 2092 2093/** 2094 * timer_lock_remote_bases - lock timer bases of cpu 2095 * @cpu: Remote CPU 2096 * 2097 * Locks the remote timer bases. 2098 */ 2099void timer_lock_remote_bases(unsigned int cpu) 2100 __acquires(timer_bases[BASE_LOCAL]->lock) 2101 __acquires(timer_bases[BASE_GLOBAL]->lock) 2102{ 2103 struct timer_base *base_local, *base_global; 2104 2105 base_local = per_cpu_ptr(&timer_bases[BASE_LOCAL], cpu); 2106 base_global = per_cpu_ptr(&timer_bases[BASE_GLOBAL], cpu); 2107 2108 lockdep_assert_irqs_disabled(); 2109 2110 raw_spin_lock(&base_local->lock); 2111 raw_spin_lock_nested(&base_global->lock, SINGLE_DEPTH_NESTING); 2112} 2113 2114/** 2115 * timer_base_is_idle() - Return whether timer base is set idle 2116 * 2117 * Returns value of local timer base is_idle value. 2118 */ 2119bool timer_base_is_idle(void) 2120{ 2121 return __this_cpu_read(timer_bases[BASE_LOCAL].is_idle); 2122} 2123 2124static void __run_timer_base(struct timer_base *base); 2125 2126/** 2127 * timer_expire_remote() - expire global timers of cpu 2128 * @cpu: Remote CPU 2129 * 2130 * Expire timers of global base of remote CPU. 2131 */ 2132void timer_expire_remote(unsigned int cpu) 2133{ 2134 struct timer_base *base = per_cpu_ptr(&timer_bases[BASE_GLOBAL], cpu); 2135 2136 __run_timer_base(base); 2137} 2138 2139static void timer_use_tmigr(unsigned long basej, u64 basem, 2140 unsigned long *nextevt, bool *tick_stop_path, 2141 bool timer_base_idle, struct timer_events *tevt) 2142{ 2143 u64 next_tmigr; 2144 2145 if (timer_base_idle) 2146 next_tmigr = tmigr_cpu_new_timer(tevt->global); 2147 else if (tick_stop_path) 2148 next_tmigr = tmigr_cpu_deactivate(tevt->global); 2149 else 2150 next_tmigr = tmigr_quick_check(tevt->global); 2151 2152 /* 2153 * If the CPU is the last going idle in timer migration hierarchy, make 2154 * sure the CPU will wake up in time to handle remote timers. 2155 * next_tmigr == KTIME_MAX if other CPUs are still active. 2156 */ 2157 if (next_tmigr < tevt->local) { 2158 u64 tmp; 2159 2160 /* If we missed a tick already, force 0 delta */ 2161 if (next_tmigr < basem) 2162 next_tmigr = basem; 2163 2164 tmp = div_u64(next_tmigr - basem, TICK_NSEC); 2165 2166 *nextevt = basej + (unsigned long)tmp; 2167 tevt->local = next_tmigr; 2168 } 2169} 2170# else 2171static void timer_use_tmigr(unsigned long basej, u64 basem, 2172 unsigned long *nextevt, bool *tick_stop_path, 2173 bool timer_base_idle, struct timer_events *tevt) 2174{ 2175 /* 2176 * Make sure first event is written into tevt->local to not miss a 2177 * timer on !SMP systems. 2178 */ 2179 tevt->local = min_t(u64, tevt->local, tevt->global); 2180} 2181# endif /* CONFIG_SMP */ 2182 2183static inline u64 __get_next_timer_interrupt(unsigned long basej, u64 basem, 2184 bool *idle) 2185{ 2186 struct timer_events tevt = { .local = KTIME_MAX, .global = KTIME_MAX }; 2187 struct timer_base *base_local, *base_global; 2188 unsigned long nextevt; 2189 bool idle_is_possible; 2190 2191 /* 2192 * When the CPU is offline, the tick is cancelled and nothing is supposed 2193 * to try to stop it. 2194 */ 2195 if (WARN_ON_ONCE(cpu_is_offline(smp_processor_id()))) { 2196 if (idle) 2197 *idle = true; 2198 return tevt.local; 2199 } 2200 2201 base_local = this_cpu_ptr(&timer_bases[BASE_LOCAL]); 2202 base_global = this_cpu_ptr(&timer_bases[BASE_GLOBAL]); 2203 2204 raw_spin_lock(&base_local->lock); 2205 raw_spin_lock_nested(&base_global->lock, SINGLE_DEPTH_NESTING); 2206 2207 nextevt = fetch_next_timer_interrupt(basej, basem, base_local, 2208 base_global, &tevt); 2209 2210 /* 2211 * If the next event is only one jiffy ahead there is no need to call 2212 * timer migration hierarchy related functions. The value for the next 2213 * global timer in @tevt struct equals then KTIME_MAX. This is also 2214 * true, when the timer base is idle. 2215 * 2216 * The proper timer migration hierarchy function depends on the callsite 2217 * and whether timer base is idle or not. @nextevt will be updated when 2218 * this CPU needs to handle the first timer migration hierarchy 2219 * event. See timer_use_tmigr() for detailed information. 2220 */ 2221 idle_is_possible = time_after(nextevt, basej + 1); 2222 if (idle_is_possible) 2223 timer_use_tmigr(basej, basem, &nextevt, idle, 2224 base_local->is_idle, &tevt); 2225 2226 /* 2227 * We have a fresh next event. Check whether we can forward the 2228 * base. 2229 */ 2230 __forward_timer_base(base_local, basej); 2231 __forward_timer_base(base_global, basej); 2232 2233 /* 2234 * Set base->is_idle only when caller is timer_base_try_to_set_idle() 2235 */ 2236 if (idle) { 2237 /* 2238 * Bases are idle if the next event is more than a tick 2239 * away. Caution: @nextevt could have changed by enqueueing a 2240 * global timer into timer migration hierarchy. Therefore a new 2241 * check is required here. 2242 * 2243 * If the base is marked idle then any timer add operation must 2244 * forward the base clk itself to keep granularity small. This 2245 * idle logic is only maintained for the BASE_LOCAL and 2246 * BASE_GLOBAL base, deferrable timers may still see large 2247 * granularity skew (by design). 2248 */ 2249 if (!base_local->is_idle && time_after(nextevt, basej + 1)) { 2250 base_local->is_idle = true; 2251 /* 2252 * Global timers queued locally while running in a task 2253 * in nohz_full mode need a self-IPI to kick reprogramming 2254 * in IRQ tail. 2255 */ 2256 if (tick_nohz_full_cpu(base_local->cpu)) 2257 base_global->is_idle = true; 2258 trace_timer_base_idle(true, base_local->cpu); 2259 } 2260 *idle = base_local->is_idle; 2261 2262 /* 2263 * When timer base is not set idle, undo the effect of 2264 * tmigr_cpu_deactivate() to prevent inconsistent states - active 2265 * timer base but inactive timer migration hierarchy. 2266 * 2267 * When timer base was already marked idle, nothing will be 2268 * changed here. 2269 */ 2270 if (!base_local->is_idle && idle_is_possible) 2271 tmigr_cpu_activate(); 2272 } 2273 2274 raw_spin_unlock(&base_global->lock); 2275 raw_spin_unlock(&base_local->lock); 2276 2277 return cmp_next_hrtimer_event(basem, tevt.local); 2278} 2279 2280/** 2281 * get_next_timer_interrupt() - return the time (clock mono) of the next timer 2282 * @basej: base time jiffies 2283 * @basem: base time clock monotonic 2284 * 2285 * Returns the tick aligned clock monotonic time of the next pending timer or 2286 * KTIME_MAX if no timer is pending. If timer of global base was queued into 2287 * timer migration hierarchy, first global timer is not taken into account. If 2288 * it was the last CPU of timer migration hierarchy going idle, first global 2289 * event is taken into account. 2290 */ 2291u64 get_next_timer_interrupt(unsigned long basej, u64 basem) 2292{ 2293 return __get_next_timer_interrupt(basej, basem, NULL); 2294} 2295 2296/** 2297 * timer_base_try_to_set_idle() - Try to set the idle state of the timer bases 2298 * @basej: base time jiffies 2299 * @basem: base time clock monotonic 2300 * @idle: pointer to store the value of timer_base->is_idle on return; 2301 * *idle contains the information whether tick was already stopped 2302 * 2303 * Returns the tick aligned clock monotonic time of the next pending timer or 2304 * KTIME_MAX if no timer is pending. When tick was already stopped KTIME_MAX is 2305 * returned as well. 2306 */ 2307u64 timer_base_try_to_set_idle(unsigned long basej, u64 basem, bool *idle) 2308{ 2309 if (*idle) 2310 return KTIME_MAX; 2311 2312 return __get_next_timer_interrupt(basej, basem, idle); 2313} 2314 2315/** 2316 * timer_clear_idle - Clear the idle state of the timer base 2317 * 2318 * Called with interrupts disabled 2319 */ 2320void timer_clear_idle(void) 2321{ 2322 /* 2323 * We do this unlocked. The worst outcome is a remote pinned timer 2324 * enqueue sending a pointless IPI, but taking the lock would just 2325 * make the window for sending the IPI a few instructions smaller 2326 * for the cost of taking the lock in the exit from idle 2327 * path. Required for BASE_LOCAL only. 2328 */ 2329 __this_cpu_write(timer_bases[BASE_LOCAL].is_idle, false); 2330 if (tick_nohz_full_cpu(smp_processor_id())) 2331 __this_cpu_write(timer_bases[BASE_GLOBAL].is_idle, false); 2332 trace_timer_base_idle(false, smp_processor_id()); 2333 2334 /* Activate without holding the timer_base->lock */ 2335 tmigr_cpu_activate(); 2336} 2337#endif 2338 2339/** 2340 * __run_timers - run all expired timers (if any) on this CPU. 2341 * @base: the timer vector to be processed. 2342 */ 2343static inline void __run_timers(struct timer_base *base) 2344{ 2345 struct hlist_head heads[LVL_DEPTH]; 2346 int levels; 2347 2348 lockdep_assert_held(&base->lock); 2349 2350 if (base->running_timer) 2351 return; 2352 2353 while (time_after_eq(jiffies, base->clk) && 2354 time_after_eq(jiffies, base->next_expiry)) { 2355 levels = collect_expired_timers(base, heads); 2356 /* 2357 * The two possible reasons for not finding any expired 2358 * timer at this clk are that all matching timers have been 2359 * dequeued or no timer has been queued since 2360 * base::next_expiry was set to base::clk + 2361 * TIMER_NEXT_MAX_DELTA. 2362 */ 2363 WARN_ON_ONCE(!levels && !base->next_expiry_recalc 2364 && base->timers_pending); 2365 /* 2366 * While executing timers, base->clk is set 1 offset ahead of 2367 * jiffies to avoid endless requeuing to current jiffies. 2368 */ 2369 base->clk++; 2370 timer_recalc_next_expiry(base); 2371 2372 while (levels--) 2373 expire_timers(base, heads + levels); 2374 } 2375} 2376 2377static void __run_timer_base(struct timer_base *base) 2378{ 2379 /* Can race against a remote CPU updating next_expiry under the lock */ 2380 if (time_before(jiffies, READ_ONCE(base->next_expiry))) 2381 return; 2382 2383 timer_base_lock_expiry(base); 2384 raw_spin_lock_irq(&base->lock); 2385 __run_timers(base); 2386 raw_spin_unlock_irq(&base->lock); 2387 timer_base_unlock_expiry(base); 2388} 2389 2390static void run_timer_base(int index) 2391{ 2392 struct timer_base *base = this_cpu_ptr(&timer_bases[index]); 2393 2394 __run_timer_base(base); 2395} 2396 2397/* 2398 * This function runs timers and the timer-tq in bottom half context. 2399 */ 2400static __latent_entropy void run_timer_softirq(void) 2401{ 2402 run_timer_base(BASE_LOCAL); 2403 if (IS_ENABLED(CONFIG_NO_HZ_COMMON)) { 2404 run_timer_base(BASE_GLOBAL); 2405 run_timer_base(BASE_DEF); 2406 2407 if (is_timers_nohz_active()) 2408 tmigr_handle_remote(); 2409 } 2410} 2411 2412/* 2413 * Called by the local, per-CPU timer interrupt on SMP. 2414 */ 2415static void run_local_timers(void) 2416{ 2417 struct timer_base *base = this_cpu_ptr(&timer_bases[BASE_LOCAL]); 2418 2419 hrtimer_run_queues(); 2420 2421 for (int i = 0; i < NR_BASES; i++, base++) { 2422 /* 2423 * Raise the softirq only if required. 2424 * 2425 * timer_base::next_expiry can be written by a remote CPU while 2426 * holding the lock. If this write happens at the same time than 2427 * the lockless local read, sanity checker could complain about 2428 * data corruption. 2429 * 2430 * There are two possible situations where 2431 * timer_base::next_expiry is written by a remote CPU: 2432 * 2433 * 1. Remote CPU expires global timers of this CPU and updates 2434 * timer_base::next_expiry of BASE_GLOBAL afterwards in 2435 * next_timer_interrupt() or timer_recalc_next_expiry(). The 2436 * worst outcome is a superfluous raise of the timer softirq 2437 * when the not yet updated value is read. 2438 * 2439 * 2. A new first pinned timer is enqueued by a remote CPU 2440 * and therefore timer_base::next_expiry of BASE_LOCAL is 2441 * updated. When this update is missed, this isn't a 2442 * problem, as an IPI is executed nevertheless when the CPU 2443 * was idle before. When the CPU wasn't idle but the update 2444 * is missed, then the timer would expire one jiffy late - 2445 * bad luck. 2446 * 2447 * Those unlikely corner cases where the worst outcome is only a 2448 * one jiffy delay or a superfluous raise of the softirq are 2449 * not that expensive as doing the check always while holding 2450 * the lock. 2451 * 2452 * Possible remote writers are using WRITE_ONCE(). Local reader 2453 * uses therefore READ_ONCE(). 2454 */ 2455 if (time_after_eq(jiffies, READ_ONCE(base->next_expiry)) || 2456 (i == BASE_DEF && tmigr_requires_handle_remote())) { 2457 raise_timer_softirq(TIMER_SOFTIRQ); 2458 return; 2459 } 2460 } 2461} 2462 2463/* 2464 * Called from the timer interrupt handler to charge one tick to the current 2465 * process. user_tick is 1 if the tick is user time, 0 for system. 2466 */ 2467void update_process_times(int user_tick) 2468{ 2469 struct task_struct *p = current; 2470 2471 /* Note: this timer irq context must be accounted for as well. */ 2472 account_process_tick(p, user_tick); 2473 run_local_timers(); 2474 rcu_sched_clock_irq(user_tick); 2475#ifdef CONFIG_IRQ_WORK 2476 if (in_hardirq()) 2477 irq_work_tick(); 2478#endif 2479 sched_tick(); 2480 if (IS_ENABLED(CONFIG_POSIX_TIMERS)) 2481 run_posix_cpu_timers(); 2482} 2483 2484#ifdef CONFIG_HOTPLUG_CPU 2485static void migrate_timer_list(struct timer_base *new_base, struct hlist_head *head) 2486{ 2487 struct timer_list *timer; 2488 int cpu = new_base->cpu; 2489 2490 while (!hlist_empty(head)) { 2491 timer = hlist_entry(head->first, struct timer_list, entry); 2492 detach_timer(timer, false); 2493 timer->flags = (timer->flags & ~TIMER_BASEMASK) | cpu; 2494 internal_add_timer(new_base, timer); 2495 } 2496} 2497 2498int timers_prepare_cpu(unsigned int cpu) 2499{ 2500 struct timer_base *base; 2501 int b; 2502 2503 for (b = 0; b < NR_BASES; b++) { 2504 base = per_cpu_ptr(&timer_bases[b], cpu); 2505 base->clk = jiffies; 2506 base->next_expiry = base->clk + TIMER_NEXT_MAX_DELTA; 2507 base->next_expiry_recalc = false; 2508 base->timers_pending = false; 2509 base->is_idle = false; 2510 } 2511 return 0; 2512} 2513 2514int timers_dead_cpu(unsigned int cpu) 2515{ 2516 struct timer_base *old_base; 2517 struct timer_base *new_base; 2518 int b, i; 2519 2520 for (b = 0; b < NR_BASES; b++) { 2521 old_base = per_cpu_ptr(&timer_bases[b], cpu); 2522 new_base = get_cpu_ptr(&timer_bases[b]); 2523 /* 2524 * The caller is globally serialized and nobody else 2525 * takes two locks at once, deadlock is not possible. 2526 */ 2527 raw_spin_lock_irq(&new_base->lock); 2528 raw_spin_lock_nested(&old_base->lock, SINGLE_DEPTH_NESTING); 2529 2530 /* 2531 * The current CPUs base clock might be stale. Update it 2532 * before moving the timers over. 2533 */ 2534 forward_timer_base(new_base); 2535 2536 WARN_ON_ONCE(old_base->running_timer); 2537 old_base->running_timer = NULL; 2538 2539 for (i = 0; i < WHEEL_SIZE; i++) 2540 migrate_timer_list(new_base, old_base->vectors + i); 2541 2542 raw_spin_unlock(&old_base->lock); 2543 raw_spin_unlock_irq(&new_base->lock); 2544 put_cpu_ptr(&timer_bases); 2545 } 2546 return 0; 2547} 2548 2549#endif /* CONFIG_HOTPLUG_CPU */ 2550 2551static void __init init_timer_cpu(int cpu) 2552{ 2553 struct timer_base *base; 2554 int i; 2555 2556 for (i = 0; i < NR_BASES; i++) { 2557 base = per_cpu_ptr(&timer_bases[i], cpu); 2558 base->cpu = cpu; 2559 raw_spin_lock_init(&base->lock); 2560 base->clk = jiffies; 2561 base->next_expiry = base->clk + TIMER_NEXT_MAX_DELTA; 2562 timer_base_init_expiry_lock(base); 2563 } 2564} 2565 2566static void __init init_timer_cpus(void) 2567{ 2568 int cpu; 2569 2570 for_each_possible_cpu(cpu) 2571 init_timer_cpu(cpu); 2572} 2573 2574void __init timers_init(void) 2575{ 2576 init_timer_cpus(); 2577 posix_cputimers_init_work(); 2578 open_softirq(TIMER_SOFTIRQ, run_timer_softirq); 2579}
~/
Read-only file preview.