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/timekeeper_internal.h> 12 #include <linux/module.h> 13 #include <linux/interrupt.h> 14 #include <linux/percpu.h> 15 #include <linux/init.h> 16 #include <linux/mm.h> 17 #include <linux/nmi.h> 18 #include <linux/sched.h> 19 #include <linux/sched/loadavg.h> 20 #include <linux/syscore_ops.h> 21 #include <linux/clocksource.h> 22 #include <linux/jiffies.h> 23 #include <linux/time.h> 24 #include <linux/tick.h> 25 #include <linux/stop_machine.h> 26 #include <linux/pvclock_gtod.h> 27 #include <linux/compiler.h> 28 29 #include "tick-internal.h" 30 #include "ntp_internal.h" 31 #include "timekeeping_internal.h" 32 33 #define TK_CLEAR_NTP (1 << 0) 34 #define TK_MIRROR (1 << 1) 35 #define TK_CLOCK_WAS_SET (1 << 2) 36 37 /* 38 * The most important data for readout fits into a single 64 byte 39 * cache line. 40 */ 41 static struct { 42 seqcount_t seq; 43 struct timekeeper timekeeper; 44 } tk_core ____cacheline_aligned; 45 46 static DEFINE_RAW_SPINLOCK(timekeeper_lock); 47 static struct timekeeper shadow_timekeeper; 48 49 /** 50 * struct tk_fast - NMI safe timekeeper 51 * @seq: Sequence counter for protecting updates. The lowest bit 52 * is the index for the tk_read_base array 53 * @base: tk_read_base array. Access is indexed by the lowest bit of 54 * @seq. 55 * 56 * See @update_fast_timekeeper() below. 57 */ 58 struct tk_fast { 59 seqcount_t seq; 60 struct tk_read_base base[2]; 61 }; 62 63 /* Suspend-time cycles value for halted fast timekeeper. */ 64 static u64 cycles_at_suspend; 65 66 static u64 dummy_clock_read(struct clocksource *cs) 67 { 68 return cycles_at_suspend; 69 } 70 71 static struct clocksource dummy_clock = { 72 .read = dummy_clock_read, 73 }; 74 75 static struct tk_fast tk_fast_mono ____cacheline_aligned = { 76 .base[0] = { .clock = &dummy_clock, }, 77 .base[1] = { .clock = &dummy_clock, }, 78 }; 79 80 static struct tk_fast tk_fast_raw ____cacheline_aligned = { 81 .base[0] = { .clock = &dummy_clock, }, 82 .base[1] = { .clock = &dummy_clock, }, 83 }; 84 85 /* flag for if timekeeping is suspended */ 86 int __read_mostly timekeeping_suspended; 87 88 static inline void tk_normalize_xtime(struct timekeeper *tk) 89 { 90 while (tk->tkr_mono.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_mono.shift)) { 91 tk->tkr_mono.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_mono.shift; 92 tk->xtime_sec++; 93 } 94 while (tk->tkr_raw.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_raw.shift)) { 95 tk->tkr_raw.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_raw.shift; 96 tk->raw_sec++; 97 } 98 } 99 100 static inline struct timespec64 tk_xtime(struct timekeeper *tk) 101 { 102 struct timespec64 ts; 103 104 ts.tv_sec = tk->xtime_sec; 105 ts.tv_nsec = (long)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift); 106 return ts; 107 } 108 109 static void tk_set_xtime(struct timekeeper *tk, const struct timespec64 *ts) 110 { 111 tk->xtime_sec = ts->tv_sec; 112 tk->tkr_mono.xtime_nsec = (u64)ts->tv_nsec << tk->tkr_mono.shift; 113 } 114 115 static void tk_xtime_add(struct timekeeper *tk, const struct timespec64 *ts) 116 { 117 tk->xtime_sec += ts->tv_sec; 118 tk->tkr_mono.xtime_nsec += (u64)ts->tv_nsec << tk->tkr_mono.shift; 119 tk_normalize_xtime(tk); 120 } 121 122 static void tk_set_wall_to_mono(struct timekeeper *tk, struct timespec64 wtm) 123 { 124 struct timespec64 tmp; 125 126 /* 127 * Verify consistency of: offset_real = -wall_to_monotonic 128 * before modifying anything 129 */ 130 set_normalized_timespec64(&tmp, -tk->wall_to_monotonic.tv_sec, 131 -tk->wall_to_monotonic.tv_nsec); 132 WARN_ON_ONCE(tk->offs_real != timespec64_to_ktime(tmp)); 133 tk->wall_to_monotonic = wtm; 134 set_normalized_timespec64(&tmp, -wtm.tv_sec, -wtm.tv_nsec); 135 tk->offs_real = timespec64_to_ktime(tmp); 136 tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tk->tai_offset, 0)); 137 } 138 139 static inline void tk_update_sleep_time(struct timekeeper *tk, ktime_t delta) 140 { 141 tk->offs_boot = ktime_add(tk->offs_boot, delta); 142 } 143 144 /* 145 * tk_clock_read - atomic clocksource read() helper 146 * 147 * This helper is necessary to use in the read paths because, while the 148 * seqlock ensures we don't return a bad value while structures are updated, 149 * it doesn't protect from potential crashes. There is the possibility that 150 * the tkr's clocksource may change between the read reference, and the 151 * clock reference passed to the read function. This can cause crashes if 152 * the wrong clocksource is passed to the wrong read function. 153 * This isn't necessary to use when holding the timekeeper_lock or doing 154 * a read of the fast-timekeeper tkrs (which is protected by its own locking 155 * and update logic). 156 */ 157 static inline u64 tk_clock_read(struct tk_read_base *tkr) 158 { 159 struct clocksource *clock = READ_ONCE(tkr->clock); 160 161 return clock->read(clock); 162 } 163 164 #ifdef CONFIG_DEBUG_TIMEKEEPING 165 #define WARNING_FREQ (HZ*300) /* 5 minute rate-limiting */ 166 167 static void timekeeping_check_update(struct timekeeper *tk, u64 offset) 168 { 169 170 u64 max_cycles = tk->tkr_mono.clock->max_cycles; 171 const char *name = tk->tkr_mono.clock->name; 172 173 if (offset > max_cycles) { 174 printk_deferred("WARNING: timekeeping: Cycle offset (%lld) is larger than allowed by the '%s' clock's max_cycles value (%lld): time overflow danger\n", 175 offset, name, max_cycles); 176 printk_deferred(" timekeeping: Your kernel is sick, but tries to cope by capping time updates\n"); 177 } else { 178 if (offset > (max_cycles >> 1)) { 179 printk_deferred("INFO: timekeeping: Cycle offset (%lld) is larger than the '%s' clock's 50%% safety margin (%lld)\n", 180 offset, name, max_cycles >> 1); 181 printk_deferred(" timekeeping: Your kernel is still fine, but is feeling a bit nervous\n"); 182 } 183 } 184 185 if (tk->underflow_seen) { 186 if (jiffies - tk->last_warning > WARNING_FREQ) { 187 printk_deferred("WARNING: Underflow in clocksource '%s' observed, time update ignored.\n", name); 188 printk_deferred(" Please report this, consider using a different clocksource, if possible.\n"); 189 printk_deferred(" Your kernel is probably still fine.\n"); 190 tk->last_warning = jiffies; 191 } 192 tk->underflow_seen = 0; 193 } 194 195 if (tk->overflow_seen) { 196 if (jiffies - tk->last_warning > WARNING_FREQ) { 197 printk_deferred("WARNING: Overflow in clocksource '%s' observed, time update capped.\n", name); 198 printk_deferred(" Please report this, consider using a different clocksource, if possible.\n"); 199 printk_deferred(" Your kernel is probably still fine.\n"); 200 tk->last_warning = jiffies; 201 } 202 tk->overflow_seen = 0; 203 } 204 } 205 206 static inline u64 timekeeping_get_delta(struct tk_read_base *tkr) 207 { 208 struct timekeeper *tk = &tk_core.timekeeper; 209 u64 now, last, mask, max, delta; 210 unsigned int seq; 211 212 /* 213 * Since we're called holding a seqlock, the data may shift 214 * under us while we're doing the calculation. This can cause 215 * false positives, since we'd note a problem but throw the 216 * results away. So nest another seqlock here to atomically 217 * grab the points we are checking with. 218 */ 219 do { 220 seq = read_seqcount_begin(&tk_core.seq); 221 now = tk_clock_read(tkr); 222 last = tkr->cycle_last; 223 mask = tkr->mask; 224 max = tkr->clock->max_cycles; 225 } while (read_seqcount_retry(&tk_core.seq, seq)); 226 227 delta = clocksource_delta(now, last, mask); 228 229 /* 230 * Try to catch underflows by checking if we are seeing small 231 * mask-relative negative values. 232 */ 233 if (unlikely((~delta & mask) < (mask >> 3))) { 234 tk->underflow_seen = 1; 235 delta = 0; 236 } 237 238 /* Cap delta value to the max_cycles values to avoid mult overflows */ 239 if (unlikely(delta > max)) { 240 tk->overflow_seen = 1; 241 delta = tkr->clock->max_cycles; 242 } 243 244 return delta; 245 } 246 #else 247 static inline void timekeeping_check_update(struct timekeeper *tk, u64 offset) 248 { 249 } 250 static inline u64 timekeeping_get_delta(struct tk_read_base *tkr) 251 { 252 u64 cycle_now, delta; 253 254 /* read clocksource */ 255 cycle_now = tk_clock_read(tkr); 256 257 /* calculate the delta since the last update_wall_time */ 258 delta = clocksource_delta(cycle_now, tkr->cycle_last, tkr->mask); 259 260 return delta; 261 } 262 #endif 263 264 /** 265 * tk_setup_internals - Set up internals to use clocksource clock. 266 * 267 * @tk: The target timekeeper to setup. 268 * @clock: Pointer to clocksource. 269 * 270 * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment 271 * pair and interval request. 272 * 273 * Unless you're the timekeeping code, you should not be using this! 274 */ 275 static void tk_setup_internals(struct timekeeper *tk, struct clocksource *clock) 276 { 277 u64 interval; 278 u64 tmp, ntpinterval; 279 struct clocksource *old_clock; 280 281 ++tk->cs_was_changed_seq; 282 old_clock = tk->tkr_mono.clock; 283 tk->tkr_mono.clock = clock; 284 tk->tkr_mono.mask = clock->mask; 285 tk->tkr_mono.cycle_last = tk_clock_read(&tk->tkr_mono); 286 287 tk->tkr_raw.clock = clock; 288 tk->tkr_raw.mask = clock->mask; 289 tk->tkr_raw.cycle_last = tk->tkr_mono.cycle_last; 290 291 /* Do the ns -> cycle conversion first, using original mult */ 292 tmp = NTP_INTERVAL_LENGTH; 293 tmp <<= clock->shift; 294 ntpinterval = tmp; 295 tmp += clock->mult/2; 296 do_div(tmp, clock->mult); 297 if (tmp == 0) 298 tmp = 1; 299 300 interval = (u64) tmp; 301 tk->cycle_interval = interval; 302 303 /* Go back from cycles -> shifted ns */ 304 tk->xtime_interval = interval * clock->mult; 305 tk->xtime_remainder = ntpinterval - tk->xtime_interval; 306 tk->raw_interval = interval * clock->mult; 307 308 /* if changing clocks, convert xtime_nsec shift units */ 309 if (old_clock) { 310 int shift_change = clock->shift - old_clock->shift; 311 if (shift_change < 0) { 312 tk->tkr_mono.xtime_nsec >>= -shift_change; 313 tk->tkr_raw.xtime_nsec >>= -shift_change; 314 } else { 315 tk->tkr_mono.xtime_nsec <<= shift_change; 316 tk->tkr_raw.xtime_nsec <<= shift_change; 317 } 318 } 319 320 tk->tkr_mono.shift = clock->shift; 321 tk->tkr_raw.shift = clock->shift; 322 323 tk->ntp_error = 0; 324 tk->ntp_error_shift = NTP_SCALE_SHIFT - clock->shift; 325 tk->ntp_tick = ntpinterval << tk->ntp_error_shift; 326 327 /* 328 * The timekeeper keeps its own mult values for the currently 329 * active clocksource. These value will be adjusted via NTP 330 * to counteract clock drifting. 331 */ 332 tk->tkr_mono.mult = clock->mult; 333 tk->tkr_raw.mult = clock->mult; 334 tk->ntp_err_mult = 0; 335 } 336 337 /* Timekeeper helper functions. */ 338 339 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET 340 static u32 default_arch_gettimeoffset(void) { return 0; } 341 u32 (*arch_gettimeoffset)(void) = default_arch_gettimeoffset; 342 #else 343 static inline u32 arch_gettimeoffset(void) { return 0; } 344 #endif 345 346 static inline u64 timekeeping_delta_to_ns(struct tk_read_base *tkr, u64 delta) 347 { 348 u64 nsec; 349 350 nsec = delta * tkr->mult + tkr->xtime_nsec; 351 nsec >>= tkr->shift; 352 353 /* If arch requires, add in get_arch_timeoffset() */ 354 return nsec + arch_gettimeoffset(); 355 } 356 357 static inline u64 timekeeping_get_ns(struct tk_read_base *tkr) 358 { 359 u64 delta; 360 361 delta = timekeeping_get_delta(tkr); 362 return timekeeping_delta_to_ns(tkr, delta); 363 } 364 365 static inline u64 timekeeping_cycles_to_ns(struct tk_read_base *tkr, u64 cycles) 366 { 367 u64 delta; 368 369 /* calculate the delta since the last update_wall_time */ 370 delta = clocksource_delta(cycles, tkr->cycle_last, tkr->mask); 371 return timekeeping_delta_to_ns(tkr, delta); 372 } 373 374 /** 375 * update_fast_timekeeper - Update the fast and NMI safe monotonic timekeeper. 376 * @tkr: Timekeeping readout base from which we take the update 377 * 378 * We want to use this from any context including NMI and tracing / 379 * instrumenting the timekeeping code itself. 380 * 381 * Employ the latch technique; see @raw_write_seqcount_latch. 382 * 383 * So if a NMI hits the update of base[0] then it will use base[1] 384 * which is still consistent. In the worst case this can result is a 385 * slightly wrong timestamp (a few nanoseconds). See 386 * @ktime_get_mono_fast_ns. 387 */ 388 static void update_fast_timekeeper(struct tk_read_base *tkr, struct tk_fast *tkf) 389 { 390 struct tk_read_base *base = tkf->base; 391 392 /* Force readers off to base[1] */ 393 raw_write_seqcount_latch(&tkf->seq); 394 395 /* Update base[0] */ 396 memcpy(base, tkr, sizeof(*base)); 397 398 /* Force readers back to base[0] */ 399 raw_write_seqcount_latch(&tkf->seq); 400 401 /* Update base[1] */ 402 memcpy(base + 1, base, sizeof(*base)); 403 } 404 405 /** 406 * ktime_get_mono_fast_ns - Fast NMI safe access to clock monotonic 407 * 408 * This timestamp is not guaranteed to be monotonic across an update. 409 * The timestamp is calculated by: 410 * 411 * now = base_mono + clock_delta * slope 412 * 413 * So if the update lowers the slope, readers who are forced to the 414 * not yet updated second array are still using the old steeper slope. 415 * 416 * tmono 417 * ^ 418 * | o n 419 * | o n 420 * | u 421 * | o 422 * |o 423 * |12345678---> reader order 424 * 425 * o = old slope 426 * u = update 427 * n = new slope 428 * 429 * So reader 6 will observe time going backwards versus reader 5. 430 * 431 * While other CPUs are likely to be able observe that, the only way 432 * for a CPU local observation is when an NMI hits in the middle of 433 * the update. Timestamps taken from that NMI context might be ahead 434 * of the following timestamps. Callers need to be aware of that and 435 * deal with it. 436 */ 437 static __always_inline u64 __ktime_get_fast_ns(struct tk_fast *tkf) 438 { 439 struct tk_read_base *tkr; 440 unsigned int seq; 441 u64 now; 442 443 do { 444 seq = raw_read_seqcount_latch(&tkf->seq); 445 tkr = tkf->base + (seq & 0x01); 446 now = ktime_to_ns(tkr->base); 447 448 now += timekeeping_delta_to_ns(tkr, 449 clocksource_delta( 450 tk_clock_read(tkr), 451 tkr->cycle_last, 452 tkr->mask)); 453 } while (read_seqcount_retry(&tkf->seq, seq)); 454 455 return now; 456 } 457 458 u64 ktime_get_mono_fast_ns(void) 459 { 460 return __ktime_get_fast_ns(&tk_fast_mono); 461 } 462 EXPORT_SYMBOL_GPL(ktime_get_mono_fast_ns); 463 464 u64 ktime_get_raw_fast_ns(void) 465 { 466 return __ktime_get_fast_ns(&tk_fast_raw); 467 } 468 EXPORT_SYMBOL_GPL(ktime_get_raw_fast_ns); 469 470 /** 471 * ktime_get_boot_fast_ns - NMI safe and fast access to boot clock. 472 * 473 * To keep it NMI safe since we're accessing from tracing, we're not using a 474 * separate timekeeper with updates to monotonic clock and boot offset 475 * protected with seqlocks. This has the following minor side effects: 476 * 477 * (1) Its possible that a timestamp be taken after the boot offset is updated 478 * but before the timekeeper is updated. If this happens, the new boot offset 479 * is added to the old timekeeping making the clock appear to update slightly 480 * earlier: 481 * CPU 0 CPU 1 482 * timekeeping_inject_sleeptime64() 483 * __timekeeping_inject_sleeptime(tk, delta); 484 * timestamp(); 485 * timekeeping_update(tk, TK_CLEAR_NTP...); 486 * 487 * (2) On 32-bit systems, the 64-bit boot offset (tk->offs_boot) may be 488 * partially updated. Since the tk->offs_boot update is a rare event, this 489 * should be a rare occurrence which postprocessing should be able to handle. 490 */ 491 u64 notrace ktime_get_boot_fast_ns(void) 492 { 493 struct timekeeper *tk = &tk_core.timekeeper; 494 495 return (ktime_get_mono_fast_ns() + ktime_to_ns(tk->offs_boot)); 496 } 497 EXPORT_SYMBOL_GPL(ktime_get_boot_fast_ns); 498 499 500 /* 501 * See comment for __ktime_get_fast_ns() vs. timestamp ordering 502 */ 503 static __always_inline u64 __ktime_get_real_fast_ns(struct tk_fast *tkf) 504 { 505 struct tk_read_base *tkr; 506 unsigned int seq; 507 u64 now; 508 509 do { 510 seq = raw_read_seqcount_latch(&tkf->seq); 511 tkr = tkf->base + (seq & 0x01); 512 now = ktime_to_ns(tkr->base_real); 513 514 now += timekeeping_delta_to_ns(tkr, 515 clocksource_delta( 516 tk_clock_read(tkr), 517 tkr->cycle_last, 518 tkr->mask)); 519 } while (read_seqcount_retry(&tkf->seq, seq)); 520 521 return now; 522 } 523 524 /** 525 * ktime_get_real_fast_ns: - NMI safe and fast access to clock realtime. 526 */ 527 u64 ktime_get_real_fast_ns(void) 528 { 529 return __ktime_get_real_fast_ns(&tk_fast_mono); 530 } 531 EXPORT_SYMBOL_GPL(ktime_get_real_fast_ns); 532 533 /** 534 * halt_fast_timekeeper - Prevent fast timekeeper from accessing clocksource. 535 * @tk: Timekeeper to snapshot. 536 * 537 * It generally is unsafe to access the clocksource after timekeeping has been 538 * suspended, so take a snapshot of the readout base of @tk and use it as the 539 * fast timekeeper's readout base while suspended. It will return the same 540 * number of cycles every time until timekeeping is resumed at which time the 541 * proper readout base for the fast timekeeper will be restored automatically. 542 */ 543 static void halt_fast_timekeeper(struct timekeeper *tk) 544 { 545 static struct tk_read_base tkr_dummy; 546 struct tk_read_base *tkr = &tk->tkr_mono; 547 548 memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy)); 549 cycles_at_suspend = tk_clock_read(tkr); 550 tkr_dummy.clock = &dummy_clock; 551 tkr_dummy.base_real = tkr->base + tk->offs_real; 552 update_fast_timekeeper(&tkr_dummy, &tk_fast_mono); 553 554 tkr = &tk->tkr_raw; 555 memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy)); 556 tkr_dummy.clock = &dummy_clock; 557 update_fast_timekeeper(&tkr_dummy, &tk_fast_raw); 558 } 559 560 #ifdef CONFIG_GENERIC_TIME_VSYSCALL_OLD 561 #warning Please contact your maintainers, as GENERIC_TIME_VSYSCALL_OLD compatibity will disappear soon. 562 563 static inline void update_vsyscall(struct timekeeper *tk) 564 { 565 struct timespec xt, wm; 566 567 xt = timespec64_to_timespec(tk_xtime(tk)); 568 wm = timespec64_to_timespec(tk->wall_to_monotonic); 569 update_vsyscall_old(&xt, &wm, tk->tkr_mono.clock, tk->tkr_mono.mult, 570 tk->tkr_mono.cycle_last); 571 } 572 573 static inline void old_vsyscall_fixup(struct timekeeper *tk) 574 { 575 s64 remainder; 576 577 /* 578 * Store only full nanoseconds into xtime_nsec after rounding 579 * it up and add the remainder to the error difference. 580 * XXX - This is necessary to avoid small 1ns inconsistnecies caused 581 * by truncating the remainder in vsyscalls. However, it causes 582 * additional work to be done in timekeeping_adjust(). Once 583 * the vsyscall implementations are converted to use xtime_nsec 584 * (shifted nanoseconds), and CONFIG_GENERIC_TIME_VSYSCALL_OLD 585 * users are removed, this can be killed. 586 */ 587 remainder = tk->tkr_mono.xtime_nsec & ((1ULL << tk->tkr_mono.shift) - 1); 588 if (remainder != 0) { 589 tk->tkr_mono.xtime_nsec -= remainder; 590 tk->tkr_mono.xtime_nsec += 1ULL << tk->tkr_mono.shift; 591 tk->ntp_error += remainder << tk->ntp_error_shift; 592 tk->ntp_error -= (1ULL << tk->tkr_mono.shift) << tk->ntp_error_shift; 593 } 594 } 595 #else 596 #define old_vsyscall_fixup(tk) 597 #endif 598 599 static RAW_NOTIFIER_HEAD(pvclock_gtod_chain); 600 601 static void update_pvclock_gtod(struct timekeeper *tk, bool was_set) 602 { 603 raw_notifier_call_chain(&pvclock_gtod_chain, was_set, tk); 604 } 605 606 /** 607 * pvclock_gtod_register_notifier - register a pvclock timedata update listener 608 */ 609 int pvclock_gtod_register_notifier(struct notifier_block *nb) 610 { 611 struct timekeeper *tk = &tk_core.timekeeper; 612 unsigned long flags; 613 int ret; 614 615 raw_spin_lock_irqsave(&timekeeper_lock, flags); 616 ret = raw_notifier_chain_register(&pvclock_gtod_chain, nb); 617 update_pvclock_gtod(tk, true); 618 raw_spin_unlock_irqrestore(&timekeeper_lock, flags); 619 620 return ret; 621 } 622 EXPORT_SYMBOL_GPL(pvclock_gtod_register_notifier); 623 624 /** 625 * pvclock_gtod_unregister_notifier - unregister a pvclock 626 * timedata update listener 627 */ 628 int pvclock_gtod_unregister_notifier(struct notifier_block *nb) 629 { 630 unsigned long flags; 631 int ret; 632 633 raw_spin_lock_irqsave(&timekeeper_lock, flags); 634 ret = raw_notifier_chain_unregister(&pvclock_gtod_chain, nb); 635 raw_spin_unlock_irqrestore(&timekeeper_lock, flags); 636 637 return ret; 638 } 639 EXPORT_SYMBOL_GPL(pvclock_gtod_unregister_notifier); 640 641 /* 642 * tk_update_leap_state - helper to update the next_leap_ktime 643 */ 644 static inline void tk_update_leap_state(struct timekeeper *tk) 645 { 646 tk->next_leap_ktime = ntp_get_next_leap(); 647 if (tk->next_leap_ktime != KTIME_MAX) 648 /* Convert to monotonic time */ 649 tk->next_leap_ktime = ktime_sub(tk->next_leap_ktime, tk->offs_real); 650 } 651 652 /* 653 * Update the ktime_t based scalar nsec members of the timekeeper 654 */ 655 static inline void tk_update_ktime_data(struct timekeeper *tk) 656 { 657 u64 seconds; 658 u32 nsec; 659 660 /* 661 * The xtime based monotonic readout is: 662 * nsec = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec + now(); 663 * The ktime based monotonic readout is: 664 * nsec = base_mono + now(); 665 * ==> base_mono = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec 666 */ 667 seconds = (u64)(tk->xtime_sec + tk->wall_to_monotonic.tv_sec); 668 nsec = (u32) tk->wall_to_monotonic.tv_nsec; 669 tk->tkr_mono.base = ns_to_ktime(seconds * NSEC_PER_SEC + nsec); 670 671 /* 672 * The sum of the nanoseconds portions of xtime and 673 * wall_to_monotonic can be greater/equal one second. Take 674 * this into account before updating tk->ktime_sec. 675 */ 676 nsec += (u32)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift); 677 if (nsec >= NSEC_PER_SEC) 678 seconds++; 679 tk->ktime_sec = seconds; 680 681 /* Update the monotonic raw base */ 682 tk->tkr_raw.base = ns_to_ktime(tk->raw_sec * NSEC_PER_SEC); 683 } 684 685 /* must hold timekeeper_lock */ 686 static void timekeeping_update(struct timekeeper *tk, unsigned int action) 687 { 688 if (action & TK_CLEAR_NTP) { 689 tk->ntp_error = 0; 690 ntp_clear(); 691 } 692 693 tk_update_leap_state(tk); 694 tk_update_ktime_data(tk); 695 696 update_vsyscall(tk); 697 update_pvclock_gtod(tk, action & TK_CLOCK_WAS_SET); 698 699 tk->tkr_mono.base_real = tk->tkr_mono.base + tk->offs_real; 700 update_fast_timekeeper(&tk->tkr_mono, &tk_fast_mono); 701 update_fast_timekeeper(&tk->tkr_raw, &tk_fast_raw); 702 703 if (action & TK_CLOCK_WAS_SET) 704 tk->clock_was_set_seq++; 705 /* 706 * The mirroring of the data to the shadow-timekeeper needs 707 * to happen last here to ensure we don't over-write the 708 * timekeeper structure on the next update with stale data 709 */ 710 if (action & TK_MIRROR) 711 memcpy(&shadow_timekeeper, &tk_core.timekeeper, 712 sizeof(tk_core.timekeeper)); 713 } 714 715 /** 716 * timekeeping_forward_now - update clock to the current time 717 * 718 * Forward the current clock to update its state since the last call to 719 * update_wall_time(). This is useful before significant clock changes, 720 * as it avoids having to deal with this time offset explicitly. 721 */ 722 static void timekeeping_forward_now(struct timekeeper *tk) 723 { 724 u64 cycle_now, delta; 725 726 cycle_now = tk_clock_read(&tk->tkr_mono); 727 delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last, tk->tkr_mono.mask); 728 tk->tkr_mono.cycle_last = cycle_now; 729 tk->tkr_raw.cycle_last = cycle_now; 730 731 tk->tkr_mono.xtime_nsec += delta * tk->tkr_mono.mult; 732 733 /* If arch requires, add in get_arch_timeoffset() */ 734 tk->tkr_mono.xtime_nsec += (u64)arch_gettimeoffset() << tk->tkr_mono.shift; 735 736 737 tk->tkr_raw.xtime_nsec += delta * tk->tkr_raw.mult; 738 739 /* If arch requires, add in get_arch_timeoffset() */ 740 tk->tkr_raw.xtime_nsec += (u64)arch_gettimeoffset() << tk->tkr_raw.shift; 741 742 tk_normalize_xtime(tk); 743 } 744 745 /** 746 * __getnstimeofday64 - Returns the time of day in a timespec64. 747 * @ts: pointer to the timespec to be set 748 * 749 * Updates the time of day in the timespec. 750 * Returns 0 on success, or -ve when suspended (timespec will be undefined). 751 */ 752 int __getnstimeofday64(struct timespec64 *ts) 753 { 754 struct timekeeper *tk = &tk_core.timekeeper; 755 unsigned long seq; 756 u64 nsecs; 757 758 do { 759 seq = read_seqcount_begin(&tk_core.seq); 760 761 ts->tv_sec = tk->xtime_sec; 762 nsecs = timekeeping_get_ns(&tk->tkr_mono); 763 764 } while (read_seqcount_retry(&tk_core.seq, seq)); 765 766 ts->tv_nsec = 0; 767 timespec64_add_ns(ts, nsecs); 768 769 /* 770 * Do not bail out early, in case there were callers still using 771 * the value, even in the face of the WARN_ON. 772 */ 773 if (unlikely(timekeeping_suspended)) 774 return -EAGAIN; 775 return 0; 776 } 777 EXPORT_SYMBOL(__getnstimeofday64); 778 779 /** 780 * getnstimeofday64 - Returns the time of day in a timespec64. 781 * @ts: pointer to the timespec64 to be set 782 * 783 * Returns the time of day in a timespec64 (WARN if suspended). 784 */ 785 void getnstimeofday64(struct timespec64 *ts) 786 { 787 WARN_ON(__getnstimeofday64(ts)); 788 } 789 EXPORT_SYMBOL(getnstimeofday64); 790 791 ktime_t ktime_get(void) 792 { 793 struct timekeeper *tk = &tk_core.timekeeper; 794 unsigned int seq; 795 ktime_t base; 796 u64 nsecs; 797 798 WARN_ON(timekeeping_suspended); 799 800 do { 801 seq = read_seqcount_begin(&tk_core.seq); 802 base = tk->tkr_mono.base; 803 nsecs = timekeeping_get_ns(&tk->tkr_mono); 804 805 } while (read_seqcount_retry(&tk_core.seq, seq)); 806 807 return ktime_add_ns(base, nsecs); 808 } 809 EXPORT_SYMBOL_GPL(ktime_get); 810 811 u32 ktime_get_resolution_ns(void) 812 { 813 struct timekeeper *tk = &tk_core.timekeeper; 814 unsigned int seq; 815 u32 nsecs; 816 817 WARN_ON(timekeeping_suspended); 818 819 do { 820 seq = read_seqcount_begin(&tk_core.seq); 821 nsecs = tk->tkr_mono.mult >> tk->tkr_mono.shift; 822 } while (read_seqcount_retry(&tk_core.seq, seq)); 823 824 return nsecs; 825 } 826 EXPORT_SYMBOL_GPL(ktime_get_resolution_ns); 827 828 static ktime_t *offsets[TK_OFFS_MAX] = { 829 [TK_OFFS_REAL] = &tk_core.timekeeper.offs_real, 830 [TK_OFFS_BOOT] = &tk_core.timekeeper.offs_boot, 831 [TK_OFFS_TAI] = &tk_core.timekeeper.offs_tai, 832 }; 833 834 ktime_t ktime_get_with_offset(enum tk_offsets offs) 835 { 836 struct timekeeper *tk = &tk_core.timekeeper; 837 unsigned int seq; 838 ktime_t base, *offset = offsets[offs]; 839 u64 nsecs; 840 841 WARN_ON(timekeeping_suspended); 842 843 do { 844 seq = read_seqcount_begin(&tk_core.seq); 845 base = ktime_add(tk->tkr_mono.base, *offset); 846 nsecs = timekeeping_get_ns(&tk->tkr_mono); 847 848 } while (read_seqcount_retry(&tk_core.seq, seq)); 849 850 return ktime_add_ns(base, nsecs); 851 852 } 853 EXPORT_SYMBOL_GPL(ktime_get_with_offset); 854 855 /** 856 * ktime_mono_to_any() - convert mononotic time to any other time 857 * @tmono: time to convert. 858 * @offs: which offset to use 859 */ 860 ktime_t ktime_mono_to_any(ktime_t tmono, enum tk_offsets offs) 861 { 862 ktime_t *offset = offsets[offs]; 863 unsigned long seq; 864 ktime_t tconv; 865 866 do { 867 seq = read_seqcount_begin(&tk_core.seq); 868 tconv = ktime_add(tmono, *offset); 869 } while (read_seqcount_retry(&tk_core.seq, seq)); 870 871 return tconv; 872 } 873 EXPORT_SYMBOL_GPL(ktime_mono_to_any); 874 875 /** 876 * ktime_get_raw - Returns the raw monotonic time in ktime_t format 877 */ 878 ktime_t ktime_get_raw(void) 879 { 880 struct timekeeper *tk = &tk_core.timekeeper; 881 unsigned int seq; 882 ktime_t base; 883 u64 nsecs; 884 885 do { 886 seq = read_seqcount_begin(&tk_core.seq); 887 base = tk->tkr_raw.base; 888 nsecs = timekeeping_get_ns(&tk->tkr_raw); 889 890 } while (read_seqcount_retry(&tk_core.seq, seq)); 891 892 return ktime_add_ns(base, nsecs); 893 } 894 EXPORT_SYMBOL_GPL(ktime_get_raw); 895 896 /** 897 * ktime_get_ts64 - get the monotonic clock in timespec64 format 898 * @ts: pointer to timespec variable 899 * 900 * The function calculates the monotonic clock from the realtime 901 * clock and the wall_to_monotonic offset and stores the result 902 * in normalized timespec64 format in the variable pointed to by @ts. 903 */ 904 void ktime_get_ts64(struct timespec64 *ts) 905 { 906 struct timekeeper *tk = &tk_core.timekeeper; 907 struct timespec64 tomono; 908 unsigned int seq; 909 u64 nsec; 910 911 WARN_ON(timekeeping_suspended); 912 913 do { 914 seq = read_seqcount_begin(&tk_core.seq); 915 ts->tv_sec = tk->xtime_sec; 916 nsec = timekeeping_get_ns(&tk->tkr_mono); 917 tomono = tk->wall_to_monotonic; 918 919 } while (read_seqcount_retry(&tk_core.seq, seq)); 920 921 ts->tv_sec += tomono.tv_sec; 922 ts->tv_nsec = 0; 923 timespec64_add_ns(ts, nsec + tomono.tv_nsec); 924 } 925 EXPORT_SYMBOL_GPL(ktime_get_ts64); 926 927 /** 928 * ktime_get_seconds - Get the seconds portion of CLOCK_MONOTONIC 929 * 930 * Returns the seconds portion of CLOCK_MONOTONIC with a single non 931 * serialized read. tk->ktime_sec is of type 'unsigned long' so this 932 * works on both 32 and 64 bit systems. On 32 bit systems the readout 933 * covers ~136 years of uptime which should be enough to prevent 934 * premature wrap arounds. 935 */ 936 time64_t ktime_get_seconds(void) 937 { 938 struct timekeeper *tk = &tk_core.timekeeper; 939 940 WARN_ON(timekeeping_suspended); 941 return tk->ktime_sec; 942 } 943 EXPORT_SYMBOL_GPL(ktime_get_seconds); 944 945 /** 946 * ktime_get_real_seconds - Get the seconds portion of CLOCK_REALTIME 947 * 948 * Returns the wall clock seconds since 1970. This replaces the 949 * get_seconds() interface which is not y2038 safe on 32bit systems. 950 * 951 * For 64bit systems the fast access to tk->xtime_sec is preserved. On 952 * 32bit systems the access must be protected with the sequence 953 * counter to provide "atomic" access to the 64bit tk->xtime_sec 954 * value. 955 */ 956 time64_t ktime_get_real_seconds(void) 957 { 958 struct timekeeper *tk = &tk_core.timekeeper; 959 time64_t seconds; 960 unsigned int seq; 961 962 if (IS_ENABLED(CONFIG_64BIT)) 963 return tk->xtime_sec; 964 965 do { 966 seq = read_seqcount_begin(&tk_core.seq); 967 seconds = tk->xtime_sec; 968 969 } while (read_seqcount_retry(&tk_core.seq, seq)); 970 971 return seconds; 972 } 973 EXPORT_SYMBOL_GPL(ktime_get_real_seconds); 974 975 /** 976 * __ktime_get_real_seconds - The same as ktime_get_real_seconds 977 * but without the sequence counter protect. This internal function 978 * is called just when timekeeping lock is already held. 979 */ 980 time64_t __ktime_get_real_seconds(void) 981 { 982 struct timekeeper *tk = &tk_core.timekeeper; 983 984 return tk->xtime_sec; 985 } 986 987 /** 988 * ktime_get_snapshot - snapshots the realtime/monotonic raw clocks with counter 989 * @systime_snapshot: pointer to struct receiving the system time snapshot 990 */ 991 void ktime_get_snapshot(struct system_time_snapshot *systime_snapshot) 992 { 993 struct timekeeper *tk = &tk_core.timekeeper; 994 unsigned long seq; 995 ktime_t base_raw; 996 ktime_t base_real; 997 u64 nsec_raw; 998 u64 nsec_real; 999 u64 now; 1000 1001 WARN_ON_ONCE(timekeeping_suspended); 1002 1003 do { 1004 seq = read_seqcount_begin(&tk_core.seq); 1005 now = tk_clock_read(&tk->tkr_mono); 1006 systime_snapshot->cs_was_changed_seq = tk->cs_was_changed_seq; 1007 systime_snapshot->clock_was_set_seq = tk->clock_was_set_seq; 1008 base_real = ktime_add(tk->tkr_mono.base, 1009 tk_core.timekeeper.offs_real); 1010 base_raw = tk->tkr_raw.base; 1011 nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono, now); 1012 nsec_raw = timekeeping_cycles_to_ns(&tk->tkr_raw, now); 1013 } while (read_seqcount_retry(&tk_core.seq, seq)); 1014 1015 systime_snapshot->cycles = now; 1016 systime_snapshot->real = ktime_add_ns(base_real, nsec_real); 1017 systime_snapshot->raw = ktime_add_ns(base_raw, nsec_raw); 1018 } 1019 EXPORT_SYMBOL_GPL(ktime_get_snapshot); 1020 1021 /* Scale base by mult/div checking for overflow */ 1022 static int scale64_check_overflow(u64 mult, u64 div, u64 *base) 1023 { 1024 u64 tmp, rem; 1025 1026 tmp = div64_u64_rem(*base, div, &rem); 1027 1028 if (((int)sizeof(u64)*8 - fls64(mult) < fls64(tmp)) || 1029 ((int)sizeof(u64)*8 - fls64(mult) < fls64(rem))) 1030 return -EOVERFLOW; 1031 tmp *= mult; 1032 rem *= mult; 1033 1034 do_div(rem, div); 1035 *base = tmp + rem; 1036 return 0; 1037 } 1038 1039 /** 1040 * adjust_historical_crosststamp - adjust crosstimestamp previous to current interval 1041 * @history: Snapshot representing start of history 1042 * @partial_history_cycles: Cycle offset into history (fractional part) 1043 * @total_history_cycles: Total history length in cycles 1044 * @discontinuity: True indicates clock was set on history period 1045 * @ts: Cross timestamp that should be adjusted using 1046 * partial/total ratio 1047 * 1048 * Helper function used by get_device_system_crosststamp() to correct the 1049 * crosstimestamp corresponding to the start of the current interval to the 1050 * system counter value (timestamp point) provided by the driver. The 1051 * total_history_* quantities are the total history starting at the provided 1052 * reference point and ending at the start of the current interval. The cycle 1053 * count between the driver timestamp point and the start of the current 1054 * interval is partial_history_cycles. 1055 */ 1056 static int adjust_historical_crosststamp(struct system_time_snapshot *history, 1057 u64 partial_history_cycles, 1058 u64 total_history_cycles, 1059 bool discontinuity, 1060 struct system_device_crosststamp *ts) 1061 { 1062 struct timekeeper *tk = &tk_core.timekeeper; 1063 u64 corr_raw, corr_real; 1064 bool interp_forward; 1065 int ret; 1066 1067 if (total_history_cycles == 0 || partial_history_cycles == 0) 1068 return 0; 1069 1070 /* Interpolate shortest distance from beginning or end of history */ 1071 interp_forward = partial_history_cycles > total_history_cycles / 2; 1072 partial_history_cycles = interp_forward ? 1073 total_history_cycles - partial_history_cycles : 1074 partial_history_cycles; 1075 1076 /* 1077 * Scale the monotonic raw time delta by: 1078 * partial_history_cycles / total_history_cycles 1079 */ 1080 corr_raw = (u64)ktime_to_ns( 1081 ktime_sub(ts->sys_monoraw, history->raw)); 1082 ret = scale64_check_overflow(partial_history_cycles, 1083 total_history_cycles, &corr_raw); 1084 if (ret) 1085 return ret; 1086 1087 /* 1088 * If there is a discontinuity in the history, scale monotonic raw 1089 * correction by: 1090 * mult(real)/mult(raw) yielding the realtime correction 1091 * Otherwise, calculate the realtime correction similar to monotonic 1092 * raw calculation 1093 */ 1094 if (discontinuity) { 1095 corr_real = mul_u64_u32_div 1096 (corr_raw, tk->tkr_mono.mult, tk->tkr_raw.mult); 1097 } else { 1098 corr_real = (u64)ktime_to_ns( 1099 ktime_sub(ts->sys_realtime, history->real)); 1100 ret = scale64_check_overflow(partial_history_cycles, 1101 total_history_cycles, &corr_real); 1102 if (ret) 1103 return ret; 1104 } 1105 1106 /* Fixup monotonic raw and real time time values */ 1107 if (interp_forward) { 1108 ts->sys_monoraw = ktime_add_ns(history->raw, corr_raw); 1109 ts->sys_realtime = ktime_add_ns(history->real, corr_real); 1110 } else { 1111 ts->sys_monoraw = ktime_sub_ns(ts->sys_monoraw, corr_raw); 1112 ts->sys_realtime = ktime_sub_ns(ts->sys_realtime, corr_real); 1113 } 1114 1115 return 0; 1116 } 1117 1118 /* 1119 * cycle_between - true if test occurs chronologically between before and after 1120 */ 1121 static bool cycle_between(u64 before, u64 test, u64 after) 1122 { 1123 if (test > before && test < after) 1124 return true; 1125 if (test < before && before > after) 1126 return true; 1127 return false; 1128 } 1129 1130 /** 1131 * get_device_system_crosststamp - Synchronously capture system/device timestamp 1132 * @get_time_fn: Callback to get simultaneous device time and 1133 * system counter from the device driver 1134 * @ctx: Context passed to get_time_fn() 1135 * @history_begin: Historical reference point used to interpolate system 1136 * time when counter provided by the driver is before the current interval 1137 * @xtstamp: Receives simultaneously captured system and device time 1138 * 1139 * Reads a timestamp from a device and correlates it to system time 1140 */ 1141 int get_device_system_crosststamp(int (*get_time_fn) 1142 (ktime_t *device_time, 1143 struct system_counterval_t *sys_counterval, 1144 void *ctx), 1145 void *ctx, 1146 struct system_time_snapshot *history_begin, 1147 struct system_device_crosststamp *xtstamp) 1148 { 1149 struct system_counterval_t system_counterval; 1150 struct timekeeper *tk = &tk_core.timekeeper; 1151 u64 cycles, now, interval_start; 1152 unsigned int clock_was_set_seq = 0; 1153 ktime_t base_real, base_raw; 1154 u64 nsec_real, nsec_raw; 1155 u8 cs_was_changed_seq; 1156 unsigned long seq; 1157 bool do_interp; 1158 int ret; 1159 1160 do { 1161 seq = read_seqcount_begin(&tk_core.seq); 1162 /* 1163 * Try to synchronously capture device time and a system 1164 * counter value calling back into the device driver 1165 */ 1166 ret = get_time_fn(&xtstamp->device, &system_counterval, ctx); 1167 if (ret) 1168 return ret; 1169 1170 /* 1171 * Verify that the clocksource associated with the captured 1172 * system counter value is the same as the currently installed 1173 * timekeeper clocksource 1174 */ 1175 if (tk->tkr_mono.clock != system_counterval.cs) 1176 return -ENODEV; 1177 cycles = system_counterval.cycles; 1178 1179 /* 1180 * Check whether the system counter value provided by the 1181 * device driver is on the current timekeeping interval. 1182 */ 1183 now = tk_clock_read(&tk->tkr_mono); 1184 interval_start = tk->tkr_mono.cycle_last; 1185 if (!cycle_between(interval_start, cycles, now)) { 1186 clock_was_set_seq = tk->clock_was_set_seq; 1187 cs_was_changed_seq = tk->cs_was_changed_seq; 1188 cycles = interval_start; 1189 do_interp = true; 1190 } else { 1191 do_interp = false; 1192 } 1193 1194 base_real = ktime_add(tk->tkr_mono.base, 1195 tk_core.timekeeper.offs_real); 1196 base_raw = tk->tkr_raw.base; 1197 1198 nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono, 1199 system_counterval.cycles); 1200 nsec_raw = timekeeping_cycles_to_ns(&tk->tkr_raw, 1201 system_counterval.cycles); 1202 } while (read_seqcount_retry(&tk_core.seq, seq)); 1203 1204 xtstamp->sys_realtime = ktime_add_ns(base_real, nsec_real); 1205 xtstamp->sys_monoraw = ktime_add_ns(base_raw, nsec_raw); 1206 1207 /* 1208 * Interpolate if necessary, adjusting back from the start of the 1209 * current interval 1210 */ 1211 if (do_interp) { 1212 u64 partial_history_cycles, total_history_cycles; 1213 bool discontinuity; 1214 1215 /* 1216 * Check that the counter value occurs after the provided 1217 * history reference and that the history doesn't cross a 1218 * clocksource change 1219 */ 1220 if (!history_begin || 1221 !cycle_between(history_begin->cycles, 1222 system_counterval.cycles, cycles) || 1223 history_begin->cs_was_changed_seq != cs_was_changed_seq) 1224 return -EINVAL; 1225 partial_history_cycles = cycles - system_counterval.cycles; 1226 total_history_cycles = cycles - history_begin->cycles; 1227 discontinuity = 1228 history_begin->clock_was_set_seq != clock_was_set_seq; 1229 1230 ret = adjust_historical_crosststamp(history_begin, 1231 partial_history_cycles, 1232 total_history_cycles, 1233 discontinuity, xtstamp); 1234 if (ret) 1235 return ret; 1236 } 1237 1238 return 0; 1239 } 1240 EXPORT_SYMBOL_GPL(get_device_system_crosststamp); 1241 1242 /** 1243 * do_gettimeofday - Returns the time of day in a timeval 1244 * @tv: pointer to the timeval to be set 1245 * 1246 * NOTE: Users should be converted to using getnstimeofday() 1247 */ 1248 void do_gettimeofday(struct timeval *tv) 1249 { 1250 struct timespec64 now; 1251 1252 getnstimeofday64(&now); 1253 tv->tv_sec = now.tv_sec; 1254 tv->tv_usec = now.tv_nsec/1000; 1255 } 1256 EXPORT_SYMBOL(do_gettimeofday); 1257 1258 /** 1259 * do_settimeofday64 - Sets the time of day. 1260 * @ts: pointer to the timespec64 variable containing the new time 1261 * 1262 * Sets the time of day to the new time and update NTP and notify hrtimers 1263 */ 1264 int do_settimeofday64(const struct timespec64 *ts) 1265 { 1266 struct timekeeper *tk = &tk_core.timekeeper; 1267 struct timespec64 ts_delta, xt; 1268 unsigned long flags; 1269 int ret = 0; 1270 1271 if (!timespec64_valid_strict(ts)) 1272 return -EINVAL; 1273 1274 raw_spin_lock_irqsave(&timekeeper_lock, flags); 1275 write_seqcount_begin(&tk_core.seq); 1276 1277 timekeeping_forward_now(tk); 1278 1279 xt = tk_xtime(tk); 1280 ts_delta.tv_sec = ts->tv_sec - xt.tv_sec; 1281 ts_delta.tv_nsec = ts->tv_nsec - xt.tv_nsec; 1282 1283 if (timespec64_compare(&tk->wall_to_monotonic, &ts_delta) > 0) { 1284 ret = -EINVAL; 1285 goto out; 1286 } 1287 1288 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts_delta)); 1289 1290 tk_set_xtime(tk, ts); 1291 out: 1292 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET); 1293 1294 write_seqcount_end(&tk_core.seq); 1295 raw_spin_unlock_irqrestore(&timekeeper_lock, flags); 1296 1297 /* signal hrtimers about time change */ 1298 clock_was_set(); 1299 1300 return ret; 1301 } 1302 EXPORT_SYMBOL(do_settimeofday64); 1303 1304 /** 1305 * timekeeping_inject_offset - Adds or subtracts from the current time. 1306 * @tv: pointer to the timespec variable containing the offset 1307 * 1308 * Adds or subtracts an offset value from the current time. 1309 */ 1310 static int timekeeping_inject_offset(struct timespec64 *ts) 1311 { 1312 struct timekeeper *tk = &tk_core.timekeeper; 1313 unsigned long flags; 1314 struct timespec64 tmp; 1315 int ret = 0; 1316 1317 if (ts->tv_nsec < 0 || ts->tv_nsec >= NSEC_PER_SEC) 1318 return -EINVAL; 1319 1320 raw_spin_lock_irqsave(&timekeeper_lock, flags); 1321 write_seqcount_begin(&tk_core.seq); 1322 1323 timekeeping_forward_now(tk); 1324 1325 /* Make sure the proposed value is valid */ 1326 tmp = timespec64_add(tk_xtime(tk), *ts); 1327 if (timespec64_compare(&tk->wall_to_monotonic, ts) > 0 || 1328 !timespec64_valid_strict(&tmp)) { 1329 ret = -EINVAL; 1330 goto error; 1331 } 1332 1333 tk_xtime_add(tk, ts); 1334 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *ts)); 1335 1336 error: /* even if we error out, we forwarded the time, so call update */ 1337 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET); 1338 1339 write_seqcount_end(&tk_core.seq); 1340 raw_spin_unlock_irqrestore(&timekeeper_lock, flags); 1341 1342 /* signal hrtimers about time change */ 1343 clock_was_set(); 1344 1345 return ret; 1346 } 1347 1348 /* 1349 * Indicates if there is an offset between the system clock and the hardware 1350 * clock/persistent clock/rtc. 1351 */ 1352 int persistent_clock_is_local; 1353 1354 /* 1355 * Adjust the time obtained from the CMOS to be UTC time instead of 1356 * local time. 1357 * 1358 * This is ugly, but preferable to the alternatives. Otherwise we 1359 * would either need to write a program to do it in /etc/rc (and risk 1360 * confusion if the program gets run more than once; it would also be 1361 * hard to make the program warp the clock precisely n hours) or 1362 * compile in the timezone information into the kernel. Bad, bad.... 1363 * 1364 * - TYT, 1992-01-01 1365 * 1366 * The best thing to do is to keep the CMOS clock in universal time (UTC) 1367 * as real UNIX machines always do it. This avoids all headaches about 1368 * daylight saving times and warping kernel clocks. 1369 */ 1370 void timekeeping_warp_clock(void) 1371 { 1372 if (sys_tz.tz_minuteswest != 0) { 1373 struct timespec64 adjust; 1374 1375 persistent_clock_is_local = 1; 1376 adjust.tv_sec = sys_tz.tz_minuteswest * 60; 1377 adjust.tv_nsec = 0; 1378 timekeeping_inject_offset(&adjust); 1379 } 1380 } 1381 1382 /** 1383 * __timekeeping_set_tai_offset - Sets the TAI offset from UTC and monotonic 1384 * 1385 */ 1386 static void __timekeeping_set_tai_offset(struct timekeeper *tk, s32 tai_offset) 1387 { 1388 tk->tai_offset = tai_offset; 1389 tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tai_offset, 0)); 1390 } 1391 1392 /** 1393 * change_clocksource - Swaps clocksources if a new one is available 1394 * 1395 * Accumulates current time interval and initializes new clocksource 1396 */ 1397 static int change_clocksource(void *data) 1398 { 1399 struct timekeeper *tk = &tk_core.timekeeper; 1400 struct clocksource *new, *old; 1401 unsigned long flags; 1402 1403 new = (struct clocksource *) data; 1404 1405 raw_spin_lock_irqsave(&timekeeper_lock, flags); 1406 write_seqcount_begin(&tk_core.seq); 1407 1408 timekeeping_forward_now(tk); 1409 /* 1410 * If the cs is in module, get a module reference. Succeeds 1411 * for built-in code (owner == NULL) as well. 1412 */ 1413 if (try_module_get(new->owner)) { 1414 if (!new->enable || new->enable(new) == 0) { 1415 old = tk->tkr_mono.clock; 1416 tk_setup_internals(tk, new); 1417 if (old->disable) 1418 old->disable(old); 1419 module_put(old->owner); 1420 } else { 1421 module_put(new->owner); 1422 } 1423 } 1424 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET); 1425 1426 write_seqcount_end(&tk_core.seq); 1427 raw_spin_unlock_irqrestore(&timekeeper_lock, flags); 1428 1429 return 0; 1430 } 1431 1432 /** 1433 * timekeeping_notify - Install a new clock source 1434 * @clock: pointer to the clock source 1435 * 1436 * This function is called from clocksource.c after a new, better clock 1437 * source has been registered. The caller holds the clocksource_mutex. 1438 */ 1439 int timekeeping_notify(struct clocksource *clock) 1440 { 1441 struct timekeeper *tk = &tk_core.timekeeper; 1442 1443 if (tk->tkr_mono.clock == clock) 1444 return 0; 1445 stop_machine(change_clocksource, clock, NULL); 1446 tick_clock_notify(); 1447 return tk->tkr_mono.clock == clock ? 0 : -1; 1448 } 1449 1450 /** 1451 * getrawmonotonic64 - Returns the raw monotonic time in a timespec 1452 * @ts: pointer to the timespec64 to be set 1453 * 1454 * Returns the raw monotonic time (completely un-modified by ntp) 1455 */ 1456 void getrawmonotonic64(struct timespec64 *ts) 1457 { 1458 struct timekeeper *tk = &tk_core.timekeeper; 1459 unsigned long seq; 1460 u64 nsecs; 1461 1462 do { 1463 seq = read_seqcount_begin(&tk_core.seq); 1464 ts->tv_sec = tk->raw_sec; 1465 nsecs = timekeeping_get_ns(&tk->tkr_raw); 1466 1467 } while (read_seqcount_retry(&tk_core.seq, seq)); 1468 1469 ts->tv_nsec = 0; 1470 timespec64_add_ns(ts, nsecs); 1471 } 1472 EXPORT_SYMBOL(getrawmonotonic64); 1473 1474 1475 /** 1476 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres 1477 */ 1478 int timekeeping_valid_for_hres(void) 1479 { 1480 struct timekeeper *tk = &tk_core.timekeeper; 1481 unsigned long seq; 1482 int ret; 1483 1484 do { 1485 seq = read_seqcount_begin(&tk_core.seq); 1486 1487 ret = tk->tkr_mono.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES; 1488 1489 } while (read_seqcount_retry(&tk_core.seq, seq)); 1490 1491 return ret; 1492 } 1493 1494 /** 1495 * timekeeping_max_deferment - Returns max time the clocksource can be deferred 1496 */ 1497 u64 timekeeping_max_deferment(void) 1498 { 1499 struct timekeeper *tk = &tk_core.timekeeper; 1500 unsigned long seq; 1501 u64 ret; 1502 1503 do { 1504 seq = read_seqcount_begin(&tk_core.seq); 1505 1506 ret = tk->tkr_mono.clock->max_idle_ns; 1507 1508 } while (read_seqcount_retry(&tk_core.seq, seq)); 1509 1510 return ret; 1511 } 1512 1513 /** 1514 * read_persistent_clock - Return time from the persistent clock. 1515 * 1516 * Weak dummy function for arches that do not yet support it. 1517 * Reads the time from the battery backed persistent clock. 1518 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported. 1519 * 1520 * XXX - Do be sure to remove it once all arches implement it. 1521 */ 1522 void __weak read_persistent_clock(struct timespec *ts) 1523 { 1524 ts->tv_sec = 0; 1525 ts->tv_nsec = 0; 1526 } 1527 1528 void __weak read_persistent_clock64(struct timespec64 *ts64) 1529 { 1530 struct timespec ts; 1531 1532 read_persistent_clock(&ts); 1533 *ts64 = timespec_to_timespec64(ts); 1534 } 1535 1536 /** 1537 * read_boot_clock64 - Return time of the system start. 1538 * 1539 * Weak dummy function for arches that do not yet support it. 1540 * Function to read the exact time the system has been started. 1541 * Returns a timespec64 with tv_sec=0 and tv_nsec=0 if unsupported. 1542 * 1543 * XXX - Do be sure to remove it once all arches implement it. 1544 */ 1545 void __weak read_boot_clock64(struct timespec64 *ts) 1546 { 1547 ts->tv_sec = 0; 1548 ts->tv_nsec = 0; 1549 } 1550 1551 /* Flag for if timekeeping_resume() has injected sleeptime */ 1552 static bool sleeptime_injected; 1553 1554 /* Flag for if there is a persistent clock on this platform */ 1555 static bool persistent_clock_exists; 1556 1557 /* 1558 * timekeeping_init - Initializes the clocksource and common timekeeping values 1559 */ 1560 void __init timekeeping_init(void) 1561 { 1562 struct timekeeper *tk = &tk_core.timekeeper; 1563 struct clocksource *clock; 1564 unsigned long flags; 1565 struct timespec64 now, boot, tmp; 1566 1567 read_persistent_clock64(&now); 1568 if (!timespec64_valid_strict(&now)) { 1569 pr_warn("WARNING: Persistent clock returned invalid value!\n" 1570 " Check your CMOS/BIOS settings.\n"); 1571 now.tv_sec = 0; 1572 now.tv_nsec = 0; 1573 } else if (now.tv_sec || now.tv_nsec) 1574 persistent_clock_exists = true; 1575 1576 read_boot_clock64(&boot); 1577 if (!timespec64_valid_strict(&boot)) { 1578 pr_warn("WARNING: Boot clock returned invalid value!\n" 1579 " Check your CMOS/BIOS settings.\n"); 1580 boot.tv_sec = 0; 1581 boot.tv_nsec = 0; 1582 } 1583 1584 raw_spin_lock_irqsave(&timekeeper_lock, flags); 1585 write_seqcount_begin(&tk_core.seq); 1586 ntp_init(); 1587 1588 clock = clocksource_default_clock(); 1589 if (clock->enable) 1590 clock->enable(clock); 1591 tk_setup_internals(tk, clock); 1592 1593 tk_set_xtime(tk, &now); 1594 tk->raw_sec = 0; 1595 if (boot.tv_sec == 0 && boot.tv_nsec == 0) 1596 boot = tk_xtime(tk); 1597 1598 set_normalized_timespec64(&tmp, -boot.tv_sec, -boot.tv_nsec); 1599 tk_set_wall_to_mono(tk, tmp); 1600 1601 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET); 1602 1603 write_seqcount_end(&tk_core.seq); 1604 raw_spin_unlock_irqrestore(&timekeeper_lock, flags); 1605 } 1606 1607 /* time in seconds when suspend began for persistent clock */ 1608 static struct timespec64 timekeeping_suspend_time; 1609 1610 /** 1611 * __timekeeping_inject_sleeptime - Internal function to add sleep interval 1612 * @delta: pointer to a timespec delta value 1613 * 1614 * Takes a timespec offset measuring a suspend interval and properly 1615 * adds the sleep offset to the timekeeping variables. 1616 */ 1617 static void __timekeeping_inject_sleeptime(struct timekeeper *tk, 1618 struct timespec64 *delta) 1619 { 1620 if (!timespec64_valid_strict(delta)) { 1621 printk_deferred(KERN_WARNING 1622 "__timekeeping_inject_sleeptime: Invalid " 1623 "sleep delta value!\n"); 1624 return; 1625 } 1626 tk_xtime_add(tk, delta); 1627 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *delta)); 1628 tk_update_sleep_time(tk, timespec64_to_ktime(*delta)); 1629 tk_debug_account_sleep_time(delta); 1630 } 1631 1632 #if defined(CONFIG_PM_SLEEP) && defined(CONFIG_RTC_HCTOSYS_DEVICE) 1633 /** 1634 * We have three kinds of time sources to use for sleep time 1635 * injection, the preference order is: 1636 * 1) non-stop clocksource 1637 * 2) persistent clock (ie: RTC accessible when irqs are off) 1638 * 3) RTC 1639 * 1640 * 1) and 2) are used by timekeeping, 3) by RTC subsystem. 1641 * If system has neither 1) nor 2), 3) will be used finally. 1642 * 1643 * 1644 * If timekeeping has injected sleeptime via either 1) or 2), 1645 * 3) becomes needless, so in this case we don't need to call 1646 * rtc_resume(), and this is what timekeeping_rtc_skipresume() 1647 * means. 1648 */ 1649 bool timekeeping_rtc_skipresume(void) 1650 { 1651 return sleeptime_injected; 1652 } 1653 1654 /** 1655 * 1) can be determined whether to use or not only when doing 1656 * timekeeping_resume() which is invoked after rtc_suspend(), 1657 * so we can't skip rtc_suspend() surely if system has 1). 1658 * 1659 * But if system has 2), 2) will definitely be used, so in this 1660 * case we don't need to call rtc_suspend(), and this is what 1661 * timekeeping_rtc_skipsuspend() means. 1662 */ 1663 bool timekeeping_rtc_skipsuspend(void) 1664 { 1665 return persistent_clock_exists; 1666 } 1667 1668 /** 1669 * timekeeping_inject_sleeptime64 - Adds suspend interval to timeekeeping values 1670 * @delta: pointer to a timespec64 delta value 1671 * 1672 * This hook is for architectures that cannot support read_persistent_clock64 1673 * because their RTC/persistent clock is only accessible when irqs are enabled. 1674 * and also don't have an effective nonstop clocksource. 1675 * 1676 * This function should only be called by rtc_resume(), and allows 1677 * a suspend offset to be injected into the timekeeping values. 1678 */ 1679 void timekeeping_inject_sleeptime64(struct timespec64 *delta) 1680 { 1681 struct timekeeper *tk = &tk_core.timekeeper; 1682 unsigned long flags; 1683 1684 raw_spin_lock_irqsave(&timekeeper_lock, flags); 1685 write_seqcount_begin(&tk_core.seq); 1686 1687 timekeeping_forward_now(tk); 1688 1689 __timekeeping_inject_sleeptime(tk, delta); 1690 1691 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET); 1692 1693 write_seqcount_end(&tk_core.seq); 1694 raw_spin_unlock_irqrestore(&timekeeper_lock, flags); 1695 1696 /* signal hrtimers about time change */ 1697 clock_was_set(); 1698 } 1699 #endif 1700 1701 /** 1702 * timekeeping_resume - Resumes the generic timekeeping subsystem. 1703 */ 1704 void timekeeping_resume(void) 1705 { 1706 struct timekeeper *tk = &tk_core.timekeeper; 1707 struct clocksource *clock = tk->tkr_mono.clock; 1708 unsigned long flags; 1709 struct timespec64 ts_new, ts_delta; 1710 u64 cycle_now; 1711 1712 sleeptime_injected = false; 1713 read_persistent_clock64(&ts_new); 1714 1715 clockevents_resume(); 1716 clocksource_resume(); 1717 1718 raw_spin_lock_irqsave(&timekeeper_lock, flags); 1719 write_seqcount_begin(&tk_core.seq); 1720 1721 /* 1722 * After system resumes, we need to calculate the suspended time and 1723 * compensate it for the OS time. There are 3 sources that could be 1724 * used: Nonstop clocksource during suspend, persistent clock and rtc 1725 * device. 1726 * 1727 * One specific platform may have 1 or 2 or all of them, and the 1728 * preference will be: 1729 * suspend-nonstop clocksource -> persistent clock -> rtc 1730 * The less preferred source will only be tried if there is no better 1731 * usable source. The rtc part is handled separately in rtc core code. 1732 */ 1733 cycle_now = tk_clock_read(&tk->tkr_mono); 1734 if ((clock->flags & CLOCK_SOURCE_SUSPEND_NONSTOP) && 1735 cycle_now > tk->tkr_mono.cycle_last) { 1736 u64 nsec, cyc_delta; 1737 1738 cyc_delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last, 1739 tk->tkr_mono.mask); 1740 nsec = mul_u64_u32_shr(cyc_delta, clock->mult, clock->shift); 1741 ts_delta = ns_to_timespec64(nsec); 1742 sleeptime_injected = true; 1743 } else if (timespec64_compare(&ts_new, &timekeeping_suspend_time) > 0) { 1744 ts_delta = timespec64_sub(ts_new, timekeeping_suspend_time); 1745 sleeptime_injected = true; 1746 } 1747 1748 if (sleeptime_injected) 1749 __timekeeping_inject_sleeptime(tk, &ts_delta); 1750 1751 /* Re-base the last cycle value */ 1752 tk->tkr_mono.cycle_last = cycle_now; 1753 tk->tkr_raw.cycle_last = cycle_now; 1754 1755 tk->ntp_error = 0; 1756 timekeeping_suspended = 0; 1757 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET); 1758 write_seqcount_end(&tk_core.seq); 1759 raw_spin_unlock_irqrestore(&timekeeper_lock, flags); 1760 1761 touch_softlockup_watchdog(); 1762 1763 tick_resume(); 1764 hrtimers_resume(); 1765 } 1766 1767 int timekeeping_suspend(void) 1768 { 1769 struct timekeeper *tk = &tk_core.timekeeper; 1770 unsigned long flags; 1771 struct timespec64 delta, delta_delta; 1772 static struct timespec64 old_delta; 1773 1774 read_persistent_clock64(&timekeeping_suspend_time); 1775 1776 /* 1777 * On some systems the persistent_clock can not be detected at 1778 * timekeeping_init by its return value, so if we see a valid 1779 * value returned, update the persistent_clock_exists flag. 1780 */ 1781 if (timekeeping_suspend_time.tv_sec || timekeeping_suspend_time.tv_nsec) 1782 persistent_clock_exists = true; 1783 1784 raw_spin_lock_irqsave(&timekeeper_lock, flags); 1785 write_seqcount_begin(&tk_core.seq); 1786 timekeeping_forward_now(tk); 1787 timekeeping_suspended = 1; 1788 1789 if (persistent_clock_exists) { 1790 /* 1791 * To avoid drift caused by repeated suspend/resumes, 1792 * which each can add ~1 second drift error, 1793 * try to compensate so the difference in system time 1794 * and persistent_clock time stays close to constant. 1795 */ 1796 delta = timespec64_sub(tk_xtime(tk), timekeeping_suspend_time); 1797 delta_delta = timespec64_sub(delta, old_delta); 1798 if (abs(delta_delta.tv_sec) >= 2) { 1799 /* 1800 * if delta_delta is too large, assume time correction 1801 * has occurred and set old_delta to the current delta. 1802 */ 1803 old_delta = delta; 1804 } else { 1805 /* Otherwise try to adjust old_system to compensate */ 1806 timekeeping_suspend_time = 1807 timespec64_add(timekeeping_suspend_time, delta_delta); 1808 } 1809 } 1810 1811 timekeeping_update(tk, TK_MIRROR); 1812 halt_fast_timekeeper(tk); 1813 write_seqcount_end(&tk_core.seq); 1814 raw_spin_unlock_irqrestore(&timekeeper_lock, flags); 1815 1816 tick_suspend(); 1817 clocksource_suspend(); 1818 clockevents_suspend(); 1819 1820 return 0; 1821 } 1822 1823 /* sysfs resume/suspend bits for timekeeping */ 1824 static struct syscore_ops timekeeping_syscore_ops = { 1825 .resume = timekeeping_resume, 1826 .suspend = timekeeping_suspend, 1827 }; 1828 1829 static int __init timekeeping_init_ops(void) 1830 { 1831 register_syscore_ops(&timekeeping_syscore_ops); 1832 return 0; 1833 } 1834 device_initcall(timekeeping_init_ops); 1835 1836 /* 1837 * Apply a multiplier adjustment to the timekeeper 1838 */ 1839 static __always_inline void timekeeping_apply_adjustment(struct timekeeper *tk, 1840 s64 offset, 1841 bool negative, 1842 int adj_scale) 1843 { 1844 s64 interval = tk->cycle_interval; 1845 s32 mult_adj = 1; 1846 1847 if (negative) { 1848 mult_adj = -mult_adj; 1849 interval = -interval; 1850 offset = -offset; 1851 } 1852 mult_adj <<= adj_scale; 1853 interval <<= adj_scale; 1854 offset <<= adj_scale; 1855 1856 /* 1857 * So the following can be confusing. 1858 * 1859 * To keep things simple, lets assume mult_adj == 1 for now. 1860 * 1861 * When mult_adj != 1, remember that the interval and offset values 1862 * have been appropriately scaled so the math is the same. 1863 * 1864 * The basic idea here is that we're increasing the multiplier 1865 * by one, this causes the xtime_interval to be incremented by 1866 * one cycle_interval. This is because: 1867 * xtime_interval = cycle_interval * mult 1868 * So if mult is being incremented by one: 1869 * xtime_interval = cycle_interval * (mult + 1) 1870 * Its the same as: 1871 * xtime_interval = (cycle_interval * mult) + cycle_interval 1872 * Which can be shortened to: 1873 * xtime_interval += cycle_interval 1874 * 1875 * So offset stores the non-accumulated cycles. Thus the current 1876 * time (in shifted nanoseconds) is: 1877 * now = (offset * adj) + xtime_nsec 1878 * Now, even though we're adjusting the clock frequency, we have 1879 * to keep time consistent. In other words, we can't jump back 1880 * in time, and we also want to avoid jumping forward in time. 1881 * 1882 * So given the same offset value, we need the time to be the same 1883 * both before and after the freq adjustment. 1884 * now = (offset * adj_1) + xtime_nsec_1 1885 * now = (offset * adj_2) + xtime_nsec_2 1886 * So: 1887 * (offset * adj_1) + xtime_nsec_1 = 1888 * (offset * adj_2) + xtime_nsec_2 1889 * And we know: 1890 * adj_2 = adj_1 + 1 1891 * So: 1892 * (offset * adj_1) + xtime_nsec_1 = 1893 * (offset * (adj_1+1)) + xtime_nsec_2 1894 * (offset * adj_1) + xtime_nsec_1 = 1895 * (offset * adj_1) + offset + xtime_nsec_2 1896 * Canceling the sides: 1897 * xtime_nsec_1 = offset + xtime_nsec_2 1898 * Which gives us: 1899 * xtime_nsec_2 = xtime_nsec_1 - offset 1900 * Which simplfies to: 1901 * xtime_nsec -= offset 1902 * 1903 * XXX - TODO: Doc ntp_error calculation. 1904 */ 1905 if ((mult_adj > 0) && (tk->tkr_mono.mult + mult_adj < mult_adj)) { 1906 /* NTP adjustment caused clocksource mult overflow */ 1907 WARN_ON_ONCE(1); 1908 return; 1909 } 1910 1911 tk->tkr_mono.mult += mult_adj; 1912 tk->xtime_interval += interval; 1913 tk->tkr_mono.xtime_nsec -= offset; 1914 tk->ntp_error -= (interval - offset) << tk->ntp_error_shift; 1915 } 1916 1917 /* 1918 * Calculate the multiplier adjustment needed to match the frequency 1919 * specified by NTP 1920 */ 1921 static __always_inline void timekeeping_freqadjust(struct timekeeper *tk, 1922 s64 offset) 1923 { 1924 s64 interval = tk->cycle_interval; 1925 s64 xinterval = tk->xtime_interval; 1926 u32 base = tk->tkr_mono.clock->mult; 1927 u32 max = tk->tkr_mono.clock->maxadj; 1928 u32 cur_adj = tk->tkr_mono.mult; 1929 s64 tick_error; 1930 bool negative; 1931 u32 adj_scale; 1932 1933 /* Remove any current error adj from freq calculation */ 1934 if (tk->ntp_err_mult) 1935 xinterval -= tk->cycle_interval; 1936 1937 tk->ntp_tick = ntp_tick_length(); 1938 1939 /* Calculate current error per tick */ 1940 tick_error = ntp_tick_length() >> tk->ntp_error_shift; 1941 tick_error -= (xinterval + tk->xtime_remainder); 1942 1943 /* Don't worry about correcting it if its small */ 1944 if (likely((tick_error >= 0) && (tick_error <= interval))) 1945 return; 1946 1947 /* preserve the direction of correction */ 1948 negative = (tick_error < 0); 1949 1950 /* If any adjustment would pass the max, just return */ 1951 if (negative && (cur_adj - 1) <= (base - max)) 1952 return; 1953 if (!negative && (cur_adj + 1) >= (base + max)) 1954 return; 1955 /* 1956 * Sort out the magnitude of the correction, but 1957 * avoid making so large a correction that we go 1958 * over the max adjustment. 1959 */ 1960 adj_scale = 0; 1961 tick_error = abs(tick_error); 1962 while (tick_error > interval) { 1963 u32 adj = 1 << (adj_scale + 1); 1964 1965 /* Check if adjustment gets us within 1 unit from the max */ 1966 if (negative && (cur_adj - adj) <= (base - max)) 1967 break; 1968 if (!negative && (cur_adj + adj) >= (base + max)) 1969 break; 1970 1971 adj_scale++; 1972 tick_error >>= 1; 1973 } 1974 1975 /* scale the corrections */ 1976 timekeeping_apply_adjustment(tk, offset, negative, adj_scale); 1977 } 1978 1979 /* 1980 * Adjust the timekeeper's multiplier to the correct frequency 1981 * and also to reduce the accumulated error value. 1982 */ 1983 static void timekeeping_adjust(struct timekeeper *tk, s64 offset) 1984 { 1985 /* Correct for the current frequency error */ 1986 timekeeping_freqadjust(tk, offset); 1987 1988 /* Next make a small adjustment to fix any cumulative error */ 1989 if (!tk->ntp_err_mult && (tk->ntp_error > 0)) { 1990 tk->ntp_err_mult = 1; 1991 timekeeping_apply_adjustment(tk, offset, 0, 0); 1992 } else if (tk->ntp_err_mult && (tk->ntp_error <= 0)) { 1993 /* Undo any existing error adjustment */ 1994 timekeeping_apply_adjustment(tk, offset, 1, 0); 1995 tk->ntp_err_mult = 0; 1996 } 1997 1998 if (unlikely(tk->tkr_mono.clock->maxadj && 1999 (abs(tk->tkr_mono.mult - tk->tkr_mono.clock->mult) 2000 > tk->tkr_mono.clock->maxadj))) { 2001 printk_once(KERN_WARNING 2002 "Adjusting %s more than 11%% (%ld vs %ld)\n", 2003 tk->tkr_mono.clock->name, (long)tk->tkr_mono.mult, 2004 (long)tk->tkr_mono.clock->mult + tk->tkr_mono.clock->maxadj); 2005 } 2006 2007 /* 2008 * It may be possible that when we entered this function, xtime_nsec 2009 * was very small. Further, if we're slightly speeding the clocksource 2010 * in the code above, its possible the required corrective factor to 2011 * xtime_nsec could cause it to underflow. 2012 * 2013 * Now, since we already accumulated the second, cannot simply roll 2014 * the accumulated second back, since the NTP subsystem has been 2015 * notified via second_overflow. So instead we push xtime_nsec forward 2016 * by the amount we underflowed, and add that amount into the error. 2017 * 2018 * We'll correct this error next time through this function, when 2019 * xtime_nsec is not as small. 2020 */ 2021 if (unlikely((s64)tk->tkr_mono.xtime_nsec < 0)) { 2022 s64 neg = -(s64)tk->tkr_mono.xtime_nsec; 2023 tk->tkr_mono.xtime_nsec = 0; 2024 tk->ntp_error += neg << tk->ntp_error_shift; 2025 } 2026 } 2027 2028 /** 2029 * accumulate_nsecs_to_secs - Accumulates nsecs into secs 2030 * 2031 * Helper function that accumulates the nsecs greater than a second 2032 * from the xtime_nsec field to the xtime_secs field. 2033 * It also calls into the NTP code to handle leapsecond processing. 2034 * 2035 */ 2036 static inline unsigned int accumulate_nsecs_to_secs(struct timekeeper *tk) 2037 { 2038 u64 nsecps = (u64)NSEC_PER_SEC << tk->tkr_mono.shift; 2039 unsigned int clock_set = 0; 2040 2041 while (tk->tkr_mono.xtime_nsec >= nsecps) { 2042 int leap; 2043 2044 tk->tkr_mono.xtime_nsec -= nsecps; 2045 tk->xtime_sec++; 2046 2047 /* Figure out if its a leap sec and apply if needed */ 2048 leap = second_overflow(tk->xtime_sec); 2049 if (unlikely(leap)) { 2050 struct timespec64 ts; 2051 2052 tk->xtime_sec += leap; 2053 2054 ts.tv_sec = leap; 2055 ts.tv_nsec = 0; 2056 tk_set_wall_to_mono(tk, 2057 timespec64_sub(tk->wall_to_monotonic, ts)); 2058 2059 __timekeeping_set_tai_offset(tk, tk->tai_offset - leap); 2060 2061 clock_set = TK_CLOCK_WAS_SET; 2062 } 2063 } 2064 return clock_set; 2065 } 2066 2067 /** 2068 * logarithmic_accumulation - shifted accumulation of cycles 2069 * 2070 * This functions accumulates a shifted interval of cycles into 2071 * into a shifted interval nanoseconds. Allows for O(log) accumulation 2072 * loop. 2073 * 2074 * Returns the unconsumed cycles. 2075 */ 2076 static u64 logarithmic_accumulation(struct timekeeper *tk, u64 offset, 2077 u32 shift, unsigned int *clock_set) 2078 { 2079 u64 interval = tk->cycle_interval << shift; 2080 u64 snsec_per_sec; 2081 2082 /* If the offset is smaller than a shifted interval, do nothing */ 2083 if (offset < interval) 2084 return offset; 2085 2086 /* Accumulate one shifted interval */ 2087 offset -= interval; 2088 tk->tkr_mono.cycle_last += interval; 2089 tk->tkr_raw.cycle_last += interval; 2090 2091 tk->tkr_mono.xtime_nsec += tk->xtime_interval << shift; 2092 *clock_set |= accumulate_nsecs_to_secs(tk); 2093 2094 /* Accumulate raw time */ 2095 tk->tkr_raw.xtime_nsec += tk->raw_interval << shift; 2096 snsec_per_sec = (u64)NSEC_PER_SEC << tk->tkr_raw.shift; 2097 while (tk->tkr_raw.xtime_nsec >= snsec_per_sec) { 2098 tk->tkr_raw.xtime_nsec -= snsec_per_sec; 2099 tk->raw_sec++; 2100 } 2101 2102 /* Accumulate error between NTP and clock interval */ 2103 tk->ntp_error += tk->ntp_tick << shift; 2104 tk->ntp_error -= (tk->xtime_interval + tk->xtime_remainder) << 2105 (tk->ntp_error_shift + shift); 2106 2107 return offset; 2108 } 2109 2110 /** 2111 * update_wall_time - Uses the current clocksource to increment the wall time 2112 * 2113 */ 2114 void update_wall_time(void) 2115 { 2116 struct timekeeper *real_tk = &tk_core.timekeeper; 2117 struct timekeeper *tk = &shadow_timekeeper; 2118 u64 offset; 2119 int shift = 0, maxshift; 2120 unsigned int clock_set = 0; 2121 unsigned long flags; 2122 2123 raw_spin_lock_irqsave(&timekeeper_lock, flags); 2124 2125 /* Make sure we're fully resumed: */ 2126 if (unlikely(timekeeping_suspended)) 2127 goto out; 2128 2129 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET 2130 offset = real_tk->cycle_interval; 2131 #else 2132 offset = clocksource_delta(tk_clock_read(&tk->tkr_mono), 2133 tk->tkr_mono.cycle_last, tk->tkr_mono.mask); 2134 #endif 2135 2136 /* Check if there's really nothing to do */ 2137 if (offset < real_tk->cycle_interval) 2138 goto out; 2139 2140 /* Do some additional sanity checking */ 2141 timekeeping_check_update(tk, offset); 2142 2143 /* 2144 * With NO_HZ we may have to accumulate many cycle_intervals 2145 * (think "ticks") worth of time at once. To do this efficiently, 2146 * we calculate the largest doubling multiple of cycle_intervals 2147 * that is smaller than the offset. We then accumulate that 2148 * chunk in one go, and then try to consume the next smaller 2149 * doubled multiple. 2150 */ 2151 shift = ilog2(offset) - ilog2(tk->cycle_interval); 2152 shift = max(0, shift); 2153 /* Bound shift to one less than what overflows tick_length */ 2154 maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1; 2155 shift = min(shift, maxshift); 2156 while (offset >= tk->cycle_interval) { 2157 offset = logarithmic_accumulation(tk, offset, shift, 2158 &clock_set); 2159 if (offset < tk->cycle_interval<<shift) 2160 shift--; 2161 } 2162 2163 /* correct the clock when NTP error is too big */ 2164 timekeeping_adjust(tk, offset); 2165 2166 /* 2167 * XXX This can be killed once everyone converts 2168 * to the new update_vsyscall. 2169 */ 2170 old_vsyscall_fixup(tk); 2171 2172 /* 2173 * Finally, make sure that after the rounding 2174 * xtime_nsec isn't larger than NSEC_PER_SEC 2175 */ 2176 clock_set |= accumulate_nsecs_to_secs(tk); 2177 2178 write_seqcount_begin(&tk_core.seq); 2179 /* 2180 * Update the real timekeeper. 2181 * 2182 * We could avoid this memcpy by switching pointers, but that 2183 * requires changes to all other timekeeper usage sites as 2184 * well, i.e. move the timekeeper pointer getter into the 2185 * spinlocked/seqcount protected sections. And we trade this 2186 * memcpy under the tk_core.seq against one before we start 2187 * updating. 2188 */ 2189 timekeeping_update(tk, clock_set); 2190 memcpy(real_tk, tk, sizeof(*tk)); 2191 /* The memcpy must come last. Do not put anything here! */ 2192 write_seqcount_end(&tk_core.seq); 2193 out: 2194 raw_spin_unlock_irqrestore(&timekeeper_lock, flags); 2195 if (clock_set) 2196 /* Have to call _delayed version, since in irq context*/ 2197 clock_was_set_delayed(); 2198 } 2199 2200 /** 2201 * getboottime64 - Return the real time of system boot. 2202 * @ts: pointer to the timespec64 to be set 2203 * 2204 * Returns the wall-time of boot in a timespec64. 2205 * 2206 * This is based on the wall_to_monotonic offset and the total suspend 2207 * time. Calls to settimeofday will affect the value returned (which 2208 * basically means that however wrong your real time clock is at boot time, 2209 * you get the right time here). 2210 */ 2211 void getboottime64(struct timespec64 *ts) 2212 { 2213 struct timekeeper *tk = &tk_core.timekeeper; 2214 ktime_t t = ktime_sub(tk->offs_real, tk->offs_boot); 2215 2216 *ts = ktime_to_timespec64(t); 2217 } 2218 EXPORT_SYMBOL_GPL(getboottime64); 2219 2220 unsigned long get_seconds(void) 2221 { 2222 struct timekeeper *tk = &tk_core.timekeeper; 2223 2224 return tk->xtime_sec; 2225 } 2226 EXPORT_SYMBOL(get_seconds); 2227 2228 struct timespec __current_kernel_time(void) 2229 { 2230 struct timekeeper *tk = &tk_core.timekeeper; 2231 2232 return timespec64_to_timespec(tk_xtime(tk)); 2233 } 2234 2235 struct timespec64 current_kernel_time64(void) 2236 { 2237 struct timekeeper *tk = &tk_core.timekeeper; 2238 struct timespec64 now; 2239 unsigned long seq; 2240 2241 do { 2242 seq = read_seqcount_begin(&tk_core.seq); 2243 2244 now = tk_xtime(tk); 2245 } while (read_seqcount_retry(&tk_core.seq, seq)); 2246 2247 return now; 2248 } 2249 EXPORT_SYMBOL(current_kernel_time64); 2250 2251 struct timespec64 get_monotonic_coarse64(void) 2252 { 2253 struct timekeeper *tk = &tk_core.timekeeper; 2254 struct timespec64 now, mono; 2255 unsigned long seq; 2256 2257 do { 2258 seq = read_seqcount_begin(&tk_core.seq); 2259 2260 now = tk_xtime(tk); 2261 mono = tk->wall_to_monotonic; 2262 } while (read_seqcount_retry(&tk_core.seq, seq)); 2263 2264 set_normalized_timespec64(&now, now.tv_sec + mono.tv_sec, 2265 now.tv_nsec + mono.tv_nsec); 2266 2267 return now; 2268 } 2269 EXPORT_SYMBOL(get_monotonic_coarse64); 2270 2271 /* 2272 * Must hold jiffies_lock 2273 */ 2274 void do_timer(unsigned long ticks) 2275 { 2276 jiffies_64 += ticks; 2277 calc_global_load(ticks); 2278 } 2279 2280 /** 2281 * ktime_get_update_offsets_now - hrtimer helper 2282 * @cwsseq: pointer to check and store the clock was set sequence number 2283 * @offs_real: pointer to storage for monotonic -> realtime offset 2284 * @offs_boot: pointer to storage for monotonic -> boottime offset 2285 * @offs_tai: pointer to storage for monotonic -> clock tai offset 2286 * 2287 * Returns current monotonic time and updates the offsets if the 2288 * sequence number in @cwsseq and timekeeper.clock_was_set_seq are 2289 * different. 2290 * 2291 * Called from hrtimer_interrupt() or retrigger_next_event() 2292 */ 2293 ktime_t ktime_get_update_offsets_now(unsigned int *cwsseq, ktime_t *offs_real, 2294 ktime_t *offs_boot, ktime_t *offs_tai) 2295 { 2296 struct timekeeper *tk = &tk_core.timekeeper; 2297 unsigned int seq; 2298 ktime_t base; 2299 u64 nsecs; 2300 2301 do { 2302 seq = read_seqcount_begin(&tk_core.seq); 2303 2304 base = tk->tkr_mono.base; 2305 nsecs = timekeeping_get_ns(&tk->tkr_mono); 2306 base = ktime_add_ns(base, nsecs); 2307 2308 if (*cwsseq != tk->clock_was_set_seq) { 2309 *cwsseq = tk->clock_was_set_seq; 2310 *offs_real = tk->offs_real; 2311 *offs_boot = tk->offs_boot; 2312 *offs_tai = tk->offs_tai; 2313 } 2314 2315 /* Handle leapsecond insertion adjustments */ 2316 if (unlikely(base >= tk->next_leap_ktime)) 2317 *offs_real = ktime_sub(tk->offs_real, ktime_set(1, 0)); 2318 2319 } while (read_seqcount_retry(&tk_core.seq, seq)); 2320 2321 return base; 2322 } 2323 2324 /** 2325 * timekeeping_validate_timex - Ensures the timex is ok for use in do_adjtimex 2326 */ 2327 static int timekeeping_validate_timex(struct timex *txc) 2328 { 2329 if (txc->modes & ADJ_ADJTIME) { 2330 /* singleshot must not be used with any other mode bits */ 2331 if (!(txc->modes & ADJ_OFFSET_SINGLESHOT)) 2332 return -EINVAL; 2333 if (!(txc->modes & ADJ_OFFSET_READONLY) && 2334 !capable(CAP_SYS_TIME)) 2335 return -EPERM; 2336 } else { 2337 /* In order to modify anything, you gotta be super-user! */ 2338 if (txc->modes && !capable(CAP_SYS_TIME)) 2339 return -EPERM; 2340 /* 2341 * if the quartz is off by more than 10% then 2342 * something is VERY wrong! 2343 */ 2344 if (txc->modes & ADJ_TICK && 2345 (txc->tick < 900000/USER_HZ || 2346 txc->tick > 1100000/USER_HZ)) 2347 return -EINVAL; 2348 } 2349 2350 if (txc->modes & ADJ_SETOFFSET) { 2351 /* In order to inject time, you gotta be super-user! */ 2352 if (!capable(CAP_SYS_TIME)) 2353 return -EPERM; 2354 2355 /* 2356 * Validate if a timespec/timeval used to inject a time 2357 * offset is valid. Offsets can be postive or negative, so 2358 * we don't check tv_sec. The value of the timeval/timespec 2359 * is the sum of its fields,but *NOTE*: 2360 * The field tv_usec/tv_nsec must always be non-negative and 2361 * we can't have more nanoseconds/microseconds than a second. 2362 */ 2363 if (txc->time.tv_usec < 0) 2364 return -EINVAL; 2365 2366 if (txc->modes & ADJ_NANO) { 2367 if (txc->time.tv_usec >= NSEC_PER_SEC) 2368 return -EINVAL; 2369 } else { 2370 if (txc->time.tv_usec >= USEC_PER_SEC) 2371 return -EINVAL; 2372 } 2373 } 2374 2375 /* 2376 * Check for potential multiplication overflows that can 2377 * only happen on 64-bit systems: 2378 */ 2379 if ((txc->modes & ADJ_FREQUENCY) && (BITS_PER_LONG == 64)) { 2380 if (LLONG_MIN / PPM_SCALE > txc->freq) 2381 return -EINVAL; 2382 if (LLONG_MAX / PPM_SCALE < txc->freq) 2383 return -EINVAL; 2384 } 2385 2386 return 0; 2387 } 2388 2389 2390 /** 2391 * do_adjtimex() - Accessor function to NTP __do_adjtimex function 2392 */ 2393 int do_adjtimex(struct timex *txc) 2394 { 2395 struct timekeeper *tk = &tk_core.timekeeper; 2396 unsigned long flags; 2397 struct timespec64 ts; 2398 s32 orig_tai, tai; 2399 int ret; 2400 2401 /* Validate the data before disabling interrupts */ 2402 ret = timekeeping_validate_timex(txc); 2403 if (ret) 2404 return ret; 2405 2406 if (txc->modes & ADJ_SETOFFSET) { 2407 struct timespec64 delta; 2408 delta.tv_sec = txc->time.tv_sec; 2409 delta.tv_nsec = txc->time.tv_usec; 2410 if (!(txc->modes & ADJ_NANO)) 2411 delta.tv_nsec *= 1000; 2412 ret = timekeeping_inject_offset(&delta); 2413 if (ret) 2414 return ret; 2415 } 2416 2417 getnstimeofday64(&ts); 2418 2419 raw_spin_lock_irqsave(&timekeeper_lock, flags); 2420 write_seqcount_begin(&tk_core.seq); 2421 2422 orig_tai = tai = tk->tai_offset; 2423 ret = __do_adjtimex(txc, &ts, &tai); 2424 2425 if (tai != orig_tai) { 2426 __timekeeping_set_tai_offset(tk, tai); 2427 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET); 2428 } 2429 tk_update_leap_state(tk); 2430 2431 write_seqcount_end(&tk_core.seq); 2432 raw_spin_unlock_irqrestore(&timekeeper_lock, flags); 2433 2434 if (tai != orig_tai) 2435 clock_was_set(); 2436 2437 ntp_notify_cmos_timer(); 2438 2439 return ret; 2440 } 2441 2442 #ifdef CONFIG_NTP_PPS 2443 /** 2444 * hardpps() - Accessor function to NTP __hardpps function 2445 */ 2446 void hardpps(const struct timespec64 *phase_ts, const struct timespec64 *raw_ts) 2447 { 2448 unsigned long flags; 2449 2450 raw_spin_lock_irqsave(&timekeeper_lock, flags); 2451 write_seqcount_begin(&tk_core.seq); 2452 2453 __hardpps(phase_ts, raw_ts); 2454 2455 write_seqcount_end(&tk_core.seq); 2456 raw_spin_unlock_irqrestore(&timekeeper_lock, flags); 2457 } 2458 EXPORT_SYMBOL(hardpps); 2459 #endif /* CONFIG_NTP_PPS */ 2460 2461 /** 2462 * xtime_update() - advances the timekeeping infrastructure 2463 * @ticks: number of ticks, that have elapsed since the last call. 2464 * 2465 * Must be called with interrupts disabled. 2466 */ 2467 void xtime_update(unsigned long ticks) 2468 { 2469 write_seqlock(&jiffies_lock); 2470 do_timer(ticks); 2471 write_sequnlock(&jiffies_lock); 2472 update_wall_time(); 2473 } 2474