1 /* 2 * linux/kernel/time/timekeeping.c 3 * 4 * Kernel timekeeping code and accessor functions 5 * 6 * This code was moved from linux/kernel/timer.c. 7 * Please see that file for copyright and history logs. 8 * 9 */ 10 11 #include <linux/module.h> 12 #include <linux/interrupt.h> 13 #include <linux/percpu.h> 14 #include <linux/init.h> 15 #include <linux/mm.h> 16 #include <linux/sysdev.h> 17 #include <linux/clocksource.h> 18 #include <linux/jiffies.h> 19 #include <linux/time.h> 20 #include <linux/tick.h> 21 22 23 /* 24 * This read-write spinlock protects us from races in SMP while 25 * playing with xtime and avenrun. 26 */ 27 __cacheline_aligned_in_smp DEFINE_SEQLOCK(xtime_lock); 28 29 30 /* 31 * The current time 32 * wall_to_monotonic is what we need to add to xtime (or xtime corrected 33 * for sub jiffie times) to get to monotonic time. Monotonic is pegged 34 * at zero at system boot time, so wall_to_monotonic will be negative, 35 * however, we will ALWAYS keep the tv_nsec part positive so we can use 36 * the usual normalization. 37 * 38 * wall_to_monotonic is moved after resume from suspend for the monotonic 39 * time not to jump. We need to add total_sleep_time to wall_to_monotonic 40 * to get the real boot based time offset. 41 * 42 * - wall_to_monotonic is no longer the boot time, getboottime must be 43 * used instead. 44 */ 45 struct timespec xtime __attribute__ ((aligned (16))); 46 struct timespec wall_to_monotonic __attribute__ ((aligned (16))); 47 static unsigned long total_sleep_time; /* seconds */ 48 49 static struct timespec xtime_cache __attribute__ ((aligned (16))); 50 void update_xtime_cache(u64 nsec) 51 { 52 xtime_cache = xtime; 53 timespec_add_ns(&xtime_cache, nsec); 54 } 55 56 static struct clocksource *clock; /* pointer to current clocksource */ 57 58 59 #ifdef CONFIG_GENERIC_TIME 60 /** 61 * __get_nsec_offset - Returns nanoseconds since last call to periodic_hook 62 * 63 * private function, must hold xtime_lock lock when being 64 * called. Returns the number of nanoseconds since the 65 * last call to update_wall_time() (adjusted by NTP scaling) 66 */ 67 static inline s64 __get_nsec_offset(void) 68 { 69 cycle_t cycle_now, cycle_delta; 70 s64 ns_offset; 71 72 /* read clocksource: */ 73 cycle_now = clocksource_read(clock); 74 75 /* calculate the delta since the last update_wall_time: */ 76 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask; 77 78 /* convert to nanoseconds: */ 79 ns_offset = cyc2ns(clock, cycle_delta); 80 81 return ns_offset; 82 } 83 84 /** 85 * getnstimeofday - Returns the time of day in a timespec 86 * @ts: pointer to the timespec to be set 87 * 88 * Returns the time of day in a timespec. 89 */ 90 void getnstimeofday(struct timespec *ts) 91 { 92 unsigned long seq; 93 s64 nsecs; 94 95 do { 96 seq = read_seqbegin(&xtime_lock); 97 98 *ts = xtime; 99 nsecs = __get_nsec_offset(); 100 101 } while (read_seqretry(&xtime_lock, seq)); 102 103 timespec_add_ns(ts, nsecs); 104 } 105 106 EXPORT_SYMBOL(getnstimeofday); 107 108 /** 109 * do_gettimeofday - Returns the time of day in a timeval 110 * @tv: pointer to the timeval to be set 111 * 112 * NOTE: Users should be converted to using getnstimeofday() 113 */ 114 void do_gettimeofday(struct timeval *tv) 115 { 116 struct timespec now; 117 118 getnstimeofday(&now); 119 tv->tv_sec = now.tv_sec; 120 tv->tv_usec = now.tv_nsec/1000; 121 } 122 123 EXPORT_SYMBOL(do_gettimeofday); 124 /** 125 * do_settimeofday - Sets the time of day 126 * @tv: pointer to the timespec variable containing the new time 127 * 128 * Sets the time of day to the new time and update NTP and notify hrtimers 129 */ 130 int do_settimeofday(struct timespec *tv) 131 { 132 unsigned long flags; 133 time_t wtm_sec, sec = tv->tv_sec; 134 long wtm_nsec, nsec = tv->tv_nsec; 135 136 if ((unsigned long)tv->tv_nsec >= NSEC_PER_SEC) 137 return -EINVAL; 138 139 write_seqlock_irqsave(&xtime_lock, flags); 140 141 nsec -= __get_nsec_offset(); 142 143 wtm_sec = wall_to_monotonic.tv_sec + (xtime.tv_sec - sec); 144 wtm_nsec = wall_to_monotonic.tv_nsec + (xtime.tv_nsec - nsec); 145 146 set_normalized_timespec(&xtime, sec, nsec); 147 set_normalized_timespec(&wall_to_monotonic, wtm_sec, wtm_nsec); 148 update_xtime_cache(0); 149 150 clock->error = 0; 151 ntp_clear(); 152 153 update_vsyscall(&xtime, clock); 154 155 write_sequnlock_irqrestore(&xtime_lock, flags); 156 157 /* signal hrtimers about time change */ 158 clock_was_set(); 159 160 return 0; 161 } 162 163 EXPORT_SYMBOL(do_settimeofday); 164 165 /** 166 * change_clocksource - Swaps clocksources if a new one is available 167 * 168 * Accumulates current time interval and initializes new clocksource 169 */ 170 static void change_clocksource(void) 171 { 172 struct clocksource *new; 173 cycle_t now; 174 u64 nsec; 175 176 new = clocksource_get_next(); 177 178 if (clock == new) 179 return; 180 181 now = clocksource_read(new); 182 nsec = __get_nsec_offset(); 183 timespec_add_ns(&xtime, nsec); 184 185 clock = new; 186 clock->cycle_last = now; 187 188 clock->error = 0; 189 clock->xtime_nsec = 0; 190 clocksource_calculate_interval(clock, NTP_INTERVAL_LENGTH); 191 192 tick_clock_notify(); 193 194 /* 195 * We're holding xtime lock and waking up klogd would deadlock 196 * us on enqueue. So no printing! 197 printk(KERN_INFO "Time: %s clocksource has been installed.\n", 198 clock->name); 199 */ 200 } 201 #else 202 static inline void change_clocksource(void) { } 203 static inline s64 __get_nsec_offset(void) { return 0; } 204 #endif 205 206 /** 207 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres 208 */ 209 int timekeeping_valid_for_hres(void) 210 { 211 unsigned long seq; 212 int ret; 213 214 do { 215 seq = read_seqbegin(&xtime_lock); 216 217 ret = clock->flags & CLOCK_SOURCE_VALID_FOR_HRES; 218 219 } while (read_seqretry(&xtime_lock, seq)); 220 221 return ret; 222 } 223 224 /** 225 * read_persistent_clock - Return time in seconds from the persistent clock. 226 * 227 * Weak dummy function for arches that do not yet support it. 228 * Returns seconds from epoch using the battery backed persistent clock. 229 * Returns zero if unsupported. 230 * 231 * XXX - Do be sure to remove it once all arches implement it. 232 */ 233 unsigned long __attribute__((weak)) read_persistent_clock(void) 234 { 235 return 0; 236 } 237 238 /* 239 * timekeeping_init - Initializes the clocksource and common timekeeping values 240 */ 241 void __init timekeeping_init(void) 242 { 243 unsigned long flags; 244 unsigned long sec = read_persistent_clock(); 245 246 write_seqlock_irqsave(&xtime_lock, flags); 247 248 ntp_clear(); 249 250 clock = clocksource_get_next(); 251 clocksource_calculate_interval(clock, NTP_INTERVAL_LENGTH); 252 clock->cycle_last = clocksource_read(clock); 253 254 xtime.tv_sec = sec; 255 xtime.tv_nsec = 0; 256 set_normalized_timespec(&wall_to_monotonic, 257 -xtime.tv_sec, -xtime.tv_nsec); 258 update_xtime_cache(0); 259 total_sleep_time = 0; 260 write_sequnlock_irqrestore(&xtime_lock, flags); 261 } 262 263 /* flag for if timekeeping is suspended */ 264 static int timekeeping_suspended; 265 /* time in seconds when suspend began */ 266 static unsigned long timekeeping_suspend_time; 267 /* xtime offset when we went into suspend */ 268 static s64 timekeeping_suspend_nsecs; 269 270 /** 271 * timekeeping_resume - Resumes the generic timekeeping subsystem. 272 * @dev: unused 273 * 274 * This is for the generic clocksource timekeeping. 275 * xtime/wall_to_monotonic/jiffies/etc are 276 * still managed by arch specific suspend/resume code. 277 */ 278 static int timekeeping_resume(struct sys_device *dev) 279 { 280 unsigned long flags; 281 unsigned long now = read_persistent_clock(); 282 283 clocksource_resume(); 284 285 write_seqlock_irqsave(&xtime_lock, flags); 286 287 if (now && (now > timekeeping_suspend_time)) { 288 unsigned long sleep_length = now - timekeeping_suspend_time; 289 290 xtime.tv_sec += sleep_length; 291 wall_to_monotonic.tv_sec -= sleep_length; 292 total_sleep_time += sleep_length; 293 } 294 /* Make sure that we have the correct xtime reference */ 295 timespec_add_ns(&xtime, timekeeping_suspend_nsecs); 296 update_xtime_cache(0); 297 /* re-base the last cycle value */ 298 clock->cycle_last = clocksource_read(clock); 299 clock->error = 0; 300 timekeeping_suspended = 0; 301 write_sequnlock_irqrestore(&xtime_lock, flags); 302 303 touch_softlockup_watchdog(); 304 305 clockevents_notify(CLOCK_EVT_NOTIFY_RESUME, NULL); 306 307 /* Resume hrtimers */ 308 hres_timers_resume(); 309 310 return 0; 311 } 312 313 static int timekeeping_suspend(struct sys_device *dev, pm_message_t state) 314 { 315 unsigned long flags; 316 317 timekeeping_suspend_time = read_persistent_clock(); 318 319 write_seqlock_irqsave(&xtime_lock, flags); 320 /* Get the current xtime offset */ 321 timekeeping_suspend_nsecs = __get_nsec_offset(); 322 timekeeping_suspended = 1; 323 write_sequnlock_irqrestore(&xtime_lock, flags); 324 325 clockevents_notify(CLOCK_EVT_NOTIFY_SUSPEND, NULL); 326 327 return 0; 328 } 329 330 /* sysfs resume/suspend bits for timekeeping */ 331 static struct sysdev_class timekeeping_sysclass = { 332 .name = "timekeeping", 333 .resume = timekeeping_resume, 334 .suspend = timekeeping_suspend, 335 }; 336 337 static struct sys_device device_timer = { 338 .id = 0, 339 .cls = &timekeeping_sysclass, 340 }; 341 342 static int __init timekeeping_init_device(void) 343 { 344 int error = sysdev_class_register(&timekeeping_sysclass); 345 if (!error) 346 error = sysdev_register(&device_timer); 347 return error; 348 } 349 350 device_initcall(timekeeping_init_device); 351 352 /* 353 * If the error is already larger, we look ahead even further 354 * to compensate for late or lost adjustments. 355 */ 356 static __always_inline int clocksource_bigadjust(s64 error, s64 *interval, 357 s64 *offset) 358 { 359 s64 tick_error, i; 360 u32 look_ahead, adj; 361 s32 error2, mult; 362 363 /* 364 * Use the current error value to determine how much to look ahead. 365 * The larger the error the slower we adjust for it to avoid problems 366 * with losing too many ticks, otherwise we would overadjust and 367 * produce an even larger error. The smaller the adjustment the 368 * faster we try to adjust for it, as lost ticks can do less harm 369 * here. This is tuned so that an error of about 1 msec is adjusted 370 * within about 1 sec (or 2^20 nsec in 2^SHIFT_HZ ticks). 371 */ 372 error2 = clock->error >> (TICK_LENGTH_SHIFT + 22 - 2 * SHIFT_HZ); 373 error2 = abs(error2); 374 for (look_ahead = 0; error2 > 0; look_ahead++) 375 error2 >>= 2; 376 377 /* 378 * Now calculate the error in (1 << look_ahead) ticks, but first 379 * remove the single look ahead already included in the error. 380 */ 381 tick_error = current_tick_length() >> 382 (TICK_LENGTH_SHIFT - clock->shift + 1); 383 tick_error -= clock->xtime_interval >> 1; 384 error = ((error - tick_error) >> look_ahead) + tick_error; 385 386 /* Finally calculate the adjustment shift value. */ 387 i = *interval; 388 mult = 1; 389 if (error < 0) { 390 error = -error; 391 *interval = -*interval; 392 *offset = -*offset; 393 mult = -1; 394 } 395 for (adj = 0; error > i; adj++) 396 error >>= 1; 397 398 *interval <<= adj; 399 *offset <<= adj; 400 return mult << adj; 401 } 402 403 /* 404 * Adjust the multiplier to reduce the error value, 405 * this is optimized for the most common adjustments of -1,0,1, 406 * for other values we can do a bit more work. 407 */ 408 static void clocksource_adjust(s64 offset) 409 { 410 s64 error, interval = clock->cycle_interval; 411 int adj; 412 413 error = clock->error >> (TICK_LENGTH_SHIFT - clock->shift - 1); 414 if (error > interval) { 415 error >>= 2; 416 if (likely(error <= interval)) 417 adj = 1; 418 else 419 adj = clocksource_bigadjust(error, &interval, &offset); 420 } else if (error < -interval) { 421 error >>= 2; 422 if (likely(error >= -interval)) { 423 adj = -1; 424 interval = -interval; 425 offset = -offset; 426 } else 427 adj = clocksource_bigadjust(error, &interval, &offset); 428 } else 429 return; 430 431 clock->mult += adj; 432 clock->xtime_interval += interval; 433 clock->xtime_nsec -= offset; 434 clock->error -= (interval - offset) << 435 (TICK_LENGTH_SHIFT - clock->shift); 436 } 437 438 /** 439 * update_wall_time - Uses the current clocksource to increment the wall time 440 * 441 * Called from the timer interrupt, must hold a write on xtime_lock. 442 */ 443 void update_wall_time(void) 444 { 445 cycle_t offset; 446 447 /* Make sure we're fully resumed: */ 448 if (unlikely(timekeeping_suspended)) 449 return; 450 451 #ifdef CONFIG_GENERIC_TIME 452 offset = (clocksource_read(clock) - clock->cycle_last) & clock->mask; 453 #else 454 offset = clock->cycle_interval; 455 #endif 456 clock->xtime_nsec += (s64)xtime.tv_nsec << clock->shift; 457 458 /* normally this loop will run just once, however in the 459 * case of lost or late ticks, it will accumulate correctly. 460 */ 461 while (offset >= clock->cycle_interval) { 462 /* accumulate one interval */ 463 clock->xtime_nsec += clock->xtime_interval; 464 clock->cycle_last += clock->cycle_interval; 465 offset -= clock->cycle_interval; 466 467 if (clock->xtime_nsec >= (u64)NSEC_PER_SEC << clock->shift) { 468 clock->xtime_nsec -= (u64)NSEC_PER_SEC << clock->shift; 469 xtime.tv_sec++; 470 second_overflow(); 471 } 472 473 /* accumulate error between NTP and clock interval */ 474 clock->error += current_tick_length(); 475 clock->error -= clock->xtime_interval << (TICK_LENGTH_SHIFT - clock->shift); 476 } 477 478 /* correct the clock when NTP error is too big */ 479 clocksource_adjust(offset); 480 481 /* store full nanoseconds into xtime */ 482 xtime.tv_nsec = (s64)clock->xtime_nsec >> clock->shift; 483 clock->xtime_nsec -= (s64)xtime.tv_nsec << clock->shift; 484 485 update_xtime_cache(cyc2ns(clock, offset)); 486 487 /* check to see if there is a new clocksource to use */ 488 change_clocksource(); 489 update_vsyscall(&xtime, clock); 490 } 491 492 /** 493 * getboottime - Return the real time of system boot. 494 * @ts: pointer to the timespec to be set 495 * 496 * Returns the time of day in a timespec. 497 * 498 * This is based on the wall_to_monotonic offset and the total suspend 499 * time. Calls to settimeofday will affect the value returned (which 500 * basically means that however wrong your real time clock is at boot time, 501 * you get the right time here). 502 */ 503 void getboottime(struct timespec *ts) 504 { 505 set_normalized_timespec(ts, 506 - (wall_to_monotonic.tv_sec + total_sleep_time), 507 - wall_to_monotonic.tv_nsec); 508 } 509 510 /** 511 * monotonic_to_bootbased - Convert the monotonic time to boot based. 512 * @ts: pointer to the timespec to be converted 513 */ 514 void monotonic_to_bootbased(struct timespec *ts) 515 { 516 ts->tv_sec += total_sleep_time; 517 } 518 519 unsigned long get_seconds(void) 520 { 521 return xtime_cache.tv_sec; 522 } 523 EXPORT_SYMBOL(get_seconds); 524 525 526 struct timespec current_kernel_time(void) 527 { 528 struct timespec now; 529 unsigned long seq; 530 531 do { 532 seq = read_seqbegin(&xtime_lock); 533 534 now = xtime_cache; 535 } while (read_seqretry(&xtime_lock, seq)); 536 537 return now; 538 } 539 EXPORT_SYMBOL(current_kernel_time); 540