1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Kernel timekeeping code and accessor functions. Based on code from 4 * timer.c, moved in commit 8524070b7982. 5 */ 6 #include <linux/timekeeper_internal.h> 7 #include <linux/module.h> 8 #include <linux/interrupt.h> 9 #include <linux/percpu.h> 10 #include <linux/init.h> 11 #include <linux/mm.h> 12 #include <linux/nmi.h> 13 #include <linux/sched.h> 14 #include <linux/sched/loadavg.h> 15 #include <linux/sched/clock.h> 16 #include <linux/syscore_ops.h> 17 #include <linux/clocksource.h> 18 #include <linux/jiffies.h> 19 #include <linux/time.h> 20 #include <linux/tick.h> 21 #include <linux/stop_machine.h> 22 #include <linux/pvclock_gtod.h> 23 #include <linux/compiler.h> 24 #include <linux/audit.h> 25 26 #include "tick-internal.h" 27 #include "ntp_internal.h" 28 #include "timekeeping_internal.h" 29 30 #define TK_CLEAR_NTP (1 << 0) 31 #define TK_MIRROR (1 << 1) 32 #define TK_CLOCK_WAS_SET (1 << 2) 33 34 enum timekeeping_adv_mode { 35 /* Update timekeeper when a tick has passed */ 36 TK_ADV_TICK, 37 38 /* Update timekeeper on a direct frequency change */ 39 TK_ADV_FREQ 40 }; 41 42 DEFINE_RAW_SPINLOCK(timekeeper_lock); 43 44 /* 45 * The most important data for readout fits into a single 64 byte 46 * cache line. 47 */ 48 static struct { 49 seqcount_raw_spinlock_t seq; 50 struct timekeeper timekeeper; 51 } tk_core ____cacheline_aligned = { 52 .seq = SEQCNT_RAW_SPINLOCK_ZERO(tk_core.seq, &timekeeper_lock), 53 }; 54 55 static struct timekeeper shadow_timekeeper; 56 57 /* flag for if timekeeping is suspended */ 58 int __read_mostly timekeeping_suspended; 59 60 /** 61 * struct tk_fast - NMI safe timekeeper 62 * @seq: Sequence counter for protecting updates. The lowest bit 63 * is the index for the tk_read_base array 64 * @base: tk_read_base array. Access is indexed by the lowest bit of 65 * @seq. 66 * 67 * See @update_fast_timekeeper() below. 68 */ 69 struct tk_fast { 70 seqcount_latch_t seq; 71 struct tk_read_base base[2]; 72 }; 73 74 /* Suspend-time cycles value for halted fast timekeeper. */ 75 static u64 cycles_at_suspend; 76 77 static u64 dummy_clock_read(struct clocksource *cs) 78 { 79 if (timekeeping_suspended) 80 return cycles_at_suspend; 81 return local_clock(); 82 } 83 84 static struct clocksource dummy_clock = { 85 .read = dummy_clock_read, 86 }; 87 88 /* 89 * Boot time initialization which allows local_clock() to be utilized 90 * during early boot when clocksources are not available. local_clock() 91 * returns nanoseconds already so no conversion is required, hence mult=1 92 * and shift=0. When the first proper clocksource is installed then 93 * the fast time keepers are updated with the correct values. 94 */ 95 #define FAST_TK_INIT \ 96 { \ 97 .clock = &dummy_clock, \ 98 .mask = CLOCKSOURCE_MASK(64), \ 99 .mult = 1, \ 100 .shift = 0, \ 101 } 102 103 static struct tk_fast tk_fast_mono ____cacheline_aligned = { 104 .seq = SEQCNT_LATCH_ZERO(tk_fast_mono.seq), 105 .base[0] = FAST_TK_INIT, 106 .base[1] = FAST_TK_INIT, 107 }; 108 109 static struct tk_fast tk_fast_raw ____cacheline_aligned = { 110 .seq = SEQCNT_LATCH_ZERO(tk_fast_raw.seq), 111 .base[0] = FAST_TK_INIT, 112 .base[1] = FAST_TK_INIT, 113 }; 114 115 static inline void tk_normalize_xtime(struct timekeeper *tk) 116 { 117 while (tk->tkr_mono.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_mono.shift)) { 118 tk->tkr_mono.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_mono.shift; 119 tk->xtime_sec++; 120 } 121 while (tk->tkr_raw.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_raw.shift)) { 122 tk->tkr_raw.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_raw.shift; 123 tk->raw_sec++; 124 } 125 } 126 127 static inline struct timespec64 tk_xtime(const struct timekeeper *tk) 128 { 129 struct timespec64 ts; 130 131 ts.tv_sec = tk->xtime_sec; 132 ts.tv_nsec = (long)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift); 133 return ts; 134 } 135 136 static void tk_set_xtime(struct timekeeper *tk, const struct timespec64 *ts) 137 { 138 tk->xtime_sec = ts->tv_sec; 139 tk->tkr_mono.xtime_nsec = (u64)ts->tv_nsec << tk->tkr_mono.shift; 140 } 141 142 static void tk_xtime_add(struct timekeeper *tk, const struct timespec64 *ts) 143 { 144 tk->xtime_sec += ts->tv_sec; 145 tk->tkr_mono.xtime_nsec += (u64)ts->tv_nsec << tk->tkr_mono.shift; 146 tk_normalize_xtime(tk); 147 } 148 149 static void tk_set_wall_to_mono(struct timekeeper *tk, struct timespec64 wtm) 150 { 151 struct timespec64 tmp; 152 153 /* 154 * Verify consistency of: offset_real = -wall_to_monotonic 155 * before modifying anything 156 */ 157 set_normalized_timespec64(&tmp, -tk->wall_to_monotonic.tv_sec, 158 -tk->wall_to_monotonic.tv_nsec); 159 WARN_ON_ONCE(tk->offs_real != timespec64_to_ktime(tmp)); 160 tk->wall_to_monotonic = wtm; 161 set_normalized_timespec64(&tmp, -wtm.tv_sec, -wtm.tv_nsec); 162 tk->offs_real = timespec64_to_ktime(tmp); 163 tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tk->tai_offset, 0)); 164 } 165 166 static inline void tk_update_sleep_time(struct timekeeper *tk, ktime_t delta) 167 { 168 tk->offs_boot = ktime_add(tk->offs_boot, delta); 169 /* 170 * Timespec representation for VDSO update to avoid 64bit division 171 * on every update. 172 */ 173 tk->monotonic_to_boot = ktime_to_timespec64(tk->offs_boot); 174 } 175 176 /* 177 * tk_clock_read - atomic clocksource read() helper 178 * 179 * This helper is necessary to use in the read paths because, while the 180 * seqcount ensures we don't return a bad value while structures are updated, 181 * it doesn't protect from potential crashes. There is the possibility that 182 * the tkr's clocksource may change between the read reference, and the 183 * clock reference passed to the read function. This can cause crashes if 184 * the wrong clocksource is passed to the wrong read function. 185 * This isn't necessary to use when holding the timekeeper_lock or doing 186 * a read of the fast-timekeeper tkrs (which is protected by its own locking 187 * and update logic). 188 */ 189 static inline u64 tk_clock_read(const struct tk_read_base *tkr) 190 { 191 struct clocksource *clock = READ_ONCE(tkr->clock); 192 193 return clock->read(clock); 194 } 195 196 #ifdef CONFIG_DEBUG_TIMEKEEPING 197 #define WARNING_FREQ (HZ*300) /* 5 minute rate-limiting */ 198 199 static void timekeeping_check_update(struct timekeeper *tk, u64 offset) 200 { 201 202 u64 max_cycles = tk->tkr_mono.clock->max_cycles; 203 const char *name = tk->tkr_mono.clock->name; 204 205 if (offset > max_cycles) { 206 printk_deferred("WARNING: timekeeping: Cycle offset (%lld) is larger than allowed by the '%s' clock's max_cycles value (%lld): time overflow danger\n", 207 offset, name, max_cycles); 208 printk_deferred(" timekeeping: Your kernel is sick, but tries to cope by capping time updates\n"); 209 } else { 210 if (offset > (max_cycles >> 1)) { 211 printk_deferred("INFO: timekeeping: Cycle offset (%lld) is larger than the '%s' clock's 50%% safety margin (%lld)\n", 212 offset, name, max_cycles >> 1); 213 printk_deferred(" timekeeping: Your kernel is still fine, but is feeling a bit nervous\n"); 214 } 215 } 216 217 if (tk->underflow_seen) { 218 if (jiffies - tk->last_warning > WARNING_FREQ) { 219 printk_deferred("WARNING: Underflow in clocksource '%s' observed, time update ignored.\n", name); 220 printk_deferred(" Please report this, consider using a different clocksource, if possible.\n"); 221 printk_deferred(" Your kernel is probably still fine.\n"); 222 tk->last_warning = jiffies; 223 } 224 tk->underflow_seen = 0; 225 } 226 227 if (tk->overflow_seen) { 228 if (jiffies - tk->last_warning > WARNING_FREQ) { 229 printk_deferred("WARNING: Overflow in clocksource '%s' observed, time update capped.\n", name); 230 printk_deferred(" Please report this, consider using a different clocksource, if possible.\n"); 231 printk_deferred(" Your kernel is probably still fine.\n"); 232 tk->last_warning = jiffies; 233 } 234 tk->overflow_seen = 0; 235 } 236 } 237 238 static inline u64 timekeeping_get_delta(const struct tk_read_base *tkr) 239 { 240 struct timekeeper *tk = &tk_core.timekeeper; 241 u64 now, last, mask, max, delta; 242 unsigned int seq; 243 244 /* 245 * Since we're called holding a seqcount, the data may shift 246 * under us while we're doing the calculation. This can cause 247 * false positives, since we'd note a problem but throw the 248 * results away. So nest another seqcount here to atomically 249 * grab the points we are checking with. 250 */ 251 do { 252 seq = read_seqcount_begin(&tk_core.seq); 253 now = tk_clock_read(tkr); 254 last = tkr->cycle_last; 255 mask = tkr->mask; 256 max = tkr->clock->max_cycles; 257 } while (read_seqcount_retry(&tk_core.seq, seq)); 258 259 delta = clocksource_delta(now, last, mask); 260 261 /* 262 * Try to catch underflows by checking if we are seeing small 263 * mask-relative negative values. 264 */ 265 if (unlikely((~delta & mask) < (mask >> 3))) { 266 tk->underflow_seen = 1; 267 delta = 0; 268 } 269 270 /* Cap delta value to the max_cycles values to avoid mult overflows */ 271 if (unlikely(delta > max)) { 272 tk->overflow_seen = 1; 273 delta = tkr->clock->max_cycles; 274 } 275 276 return delta; 277 } 278 #else 279 static inline void timekeeping_check_update(struct timekeeper *tk, u64 offset) 280 { 281 } 282 static inline u64 timekeeping_get_delta(const struct tk_read_base *tkr) 283 { 284 u64 cycle_now, delta; 285 286 /* read clocksource */ 287 cycle_now = tk_clock_read(tkr); 288 289 /* calculate the delta since the last update_wall_time */ 290 delta = clocksource_delta(cycle_now, tkr->cycle_last, tkr->mask); 291 292 return delta; 293 } 294 #endif 295 296 /** 297 * tk_setup_internals - Set up internals to use clocksource clock. 298 * 299 * @tk: The target timekeeper to setup. 300 * @clock: Pointer to clocksource. 301 * 302 * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment 303 * pair and interval request. 304 * 305 * Unless you're the timekeeping code, you should not be using this! 306 */ 307 static void tk_setup_internals(struct timekeeper *tk, struct clocksource *clock) 308 { 309 u64 interval; 310 u64 tmp, ntpinterval; 311 struct clocksource *old_clock; 312 313 ++tk->cs_was_changed_seq; 314 old_clock = tk->tkr_mono.clock; 315 tk->tkr_mono.clock = clock; 316 tk->tkr_mono.mask = clock->mask; 317 tk->tkr_mono.cycle_last = tk_clock_read(&tk->tkr_mono); 318 319 tk->tkr_raw.clock = clock; 320 tk->tkr_raw.mask = clock->mask; 321 tk->tkr_raw.cycle_last = tk->tkr_mono.cycle_last; 322 323 /* Do the ns -> cycle conversion first, using original mult */ 324 tmp = NTP_INTERVAL_LENGTH; 325 tmp <<= clock->shift; 326 ntpinterval = tmp; 327 tmp += clock->mult/2; 328 do_div(tmp, clock->mult); 329 if (tmp == 0) 330 tmp = 1; 331 332 interval = (u64) tmp; 333 tk->cycle_interval = interval; 334 335 /* Go back from cycles -> shifted ns */ 336 tk->xtime_interval = interval * clock->mult; 337 tk->xtime_remainder = ntpinterval - tk->xtime_interval; 338 tk->raw_interval = interval * clock->mult; 339 340 /* if changing clocks, convert xtime_nsec shift units */ 341 if (old_clock) { 342 int shift_change = clock->shift - old_clock->shift; 343 if (shift_change < 0) { 344 tk->tkr_mono.xtime_nsec >>= -shift_change; 345 tk->tkr_raw.xtime_nsec >>= -shift_change; 346 } else { 347 tk->tkr_mono.xtime_nsec <<= shift_change; 348 tk->tkr_raw.xtime_nsec <<= shift_change; 349 } 350 } 351 352 tk->tkr_mono.shift = clock->shift; 353 tk->tkr_raw.shift = clock->shift; 354 355 tk->ntp_error = 0; 356 tk->ntp_error_shift = NTP_SCALE_SHIFT - clock->shift; 357 tk->ntp_tick = ntpinterval << tk->ntp_error_shift; 358 359 /* 360 * The timekeeper keeps its own mult values for the currently 361 * active clocksource. These value will be adjusted via NTP 362 * to counteract clock drifting. 363 */ 364 tk->tkr_mono.mult = clock->mult; 365 tk->tkr_raw.mult = clock->mult; 366 tk->ntp_err_mult = 0; 367 tk->skip_second_overflow = 0; 368 } 369 370 /* Timekeeper helper functions. */ 371 372 static inline u64 timekeeping_delta_to_ns(const struct tk_read_base *tkr, u64 delta) 373 { 374 u64 nsec; 375 376 nsec = delta * tkr->mult + tkr->xtime_nsec; 377 nsec >>= tkr->shift; 378 379 return nsec; 380 } 381 382 static inline u64 timekeeping_get_ns(const struct tk_read_base *tkr) 383 { 384 u64 delta; 385 386 delta = timekeeping_get_delta(tkr); 387 return timekeeping_delta_to_ns(tkr, delta); 388 } 389 390 static inline u64 timekeeping_cycles_to_ns(const struct tk_read_base *tkr, u64 cycles) 391 { 392 u64 delta; 393 394 /* calculate the delta since the last update_wall_time */ 395 delta = clocksource_delta(cycles, tkr->cycle_last, tkr->mask); 396 return timekeeping_delta_to_ns(tkr, delta); 397 } 398 399 /** 400 * update_fast_timekeeper - Update the fast and NMI safe monotonic timekeeper. 401 * @tkr: Timekeeping readout base from which we take the update 402 * @tkf: Pointer to NMI safe timekeeper 403 * 404 * We want to use this from any context including NMI and tracing / 405 * instrumenting the timekeeping code itself. 406 * 407 * Employ the latch technique; see @raw_write_seqcount_latch. 408 * 409 * So if a NMI hits the update of base[0] then it will use base[1] 410 * which is still consistent. In the worst case this can result is a 411 * slightly wrong timestamp (a few nanoseconds). See 412 * @ktime_get_mono_fast_ns. 413 */ 414 static void update_fast_timekeeper(const struct tk_read_base *tkr, 415 struct tk_fast *tkf) 416 { 417 struct tk_read_base *base = tkf->base; 418 419 /* Force readers off to base[1] */ 420 raw_write_seqcount_latch(&tkf->seq); 421 422 /* Update base[0] */ 423 memcpy(base, tkr, sizeof(*base)); 424 425 /* Force readers back to base[0] */ 426 raw_write_seqcount_latch(&tkf->seq); 427 428 /* Update base[1] */ 429 memcpy(base + 1, base, sizeof(*base)); 430 } 431 432 static __always_inline u64 __ktime_get_fast_ns(struct tk_fast *tkf) 433 { 434 struct tk_read_base *tkr; 435 unsigned int seq; 436 u64 now; 437 438 do { 439 seq = raw_read_seqcount_latch(&tkf->seq); 440 tkr = tkf->base + (seq & 0x01); 441 now = ktime_to_ns(tkr->base); 442 443 now += timekeeping_delta_to_ns(tkr, 444 clocksource_delta( 445 tk_clock_read(tkr), 446 tkr->cycle_last, 447 tkr->mask)); 448 } while (read_seqcount_latch_retry(&tkf->seq, seq)); 449 450 return now; 451 } 452 453 /** 454 * ktime_get_mono_fast_ns - Fast NMI safe access to clock monotonic 455 * 456 * This timestamp is not guaranteed to be monotonic across an update. 457 * The timestamp is calculated by: 458 * 459 * now = base_mono + clock_delta * slope 460 * 461 * So if the update lowers the slope, readers who are forced to the 462 * not yet updated second array are still using the old steeper slope. 463 * 464 * tmono 465 * ^ 466 * | o n 467 * | o n 468 * | u 469 * | o 470 * |o 471 * |12345678---> reader order 472 * 473 * o = old slope 474 * u = update 475 * n = new slope 476 * 477 * So reader 6 will observe time going backwards versus reader 5. 478 * 479 * While other CPUs are likely to be able to observe that, the only way 480 * for a CPU local observation is when an NMI hits in the middle of 481 * the update. Timestamps taken from that NMI context might be ahead 482 * of the following timestamps. Callers need to be aware of that and 483 * deal with it. 484 */ 485 u64 ktime_get_mono_fast_ns(void) 486 { 487 return __ktime_get_fast_ns(&tk_fast_mono); 488 } 489 EXPORT_SYMBOL_GPL(ktime_get_mono_fast_ns); 490 491 /** 492 * ktime_get_raw_fast_ns - Fast NMI safe access to clock monotonic raw 493 * 494 * Contrary to ktime_get_mono_fast_ns() this is always correct because the 495 * conversion factor is not affected by NTP/PTP correction. 496 */ 497 u64 ktime_get_raw_fast_ns(void) 498 { 499 return __ktime_get_fast_ns(&tk_fast_raw); 500 } 501 EXPORT_SYMBOL_GPL(ktime_get_raw_fast_ns); 502 503 /** 504 * ktime_get_boot_fast_ns - NMI safe and fast access to boot clock. 505 * 506 * To keep it NMI safe since we're accessing from tracing, we're not using a 507 * separate timekeeper with updates to monotonic clock and boot offset 508 * protected with seqcounts. This has the following minor side effects: 509 * 510 * (1) Its possible that a timestamp be taken after the boot offset is updated 511 * but before the timekeeper is updated. If this happens, the new boot offset 512 * is added to the old timekeeping making the clock appear to update slightly 513 * earlier: 514 * CPU 0 CPU 1 515 * timekeeping_inject_sleeptime64() 516 * __timekeeping_inject_sleeptime(tk, delta); 517 * timestamp(); 518 * timekeeping_update(tk, TK_CLEAR_NTP...); 519 * 520 * (2) On 32-bit systems, the 64-bit boot offset (tk->offs_boot) may be 521 * partially updated. Since the tk->offs_boot update is a rare event, this 522 * should be a rare occurrence which postprocessing should be able to handle. 523 * 524 * The caveats vs. timestamp ordering as documented for ktime_get_fast_ns() 525 * apply as well. 526 */ 527 u64 notrace ktime_get_boot_fast_ns(void) 528 { 529 struct timekeeper *tk = &tk_core.timekeeper; 530 531 return (ktime_get_mono_fast_ns() + ktime_to_ns(tk->offs_boot)); 532 } 533 EXPORT_SYMBOL_GPL(ktime_get_boot_fast_ns); 534 535 static __always_inline u64 __ktime_get_real_fast(struct tk_fast *tkf, u64 *mono) 536 { 537 struct tk_read_base *tkr; 538 u64 basem, baser, delta; 539 unsigned int seq; 540 541 do { 542 seq = raw_read_seqcount_latch(&tkf->seq); 543 tkr = tkf->base + (seq & 0x01); 544 basem = ktime_to_ns(tkr->base); 545 baser = ktime_to_ns(tkr->base_real); 546 547 delta = timekeeping_delta_to_ns(tkr, 548 clocksource_delta(tk_clock_read(tkr), 549 tkr->cycle_last, tkr->mask)); 550 } while (read_seqcount_latch_retry(&tkf->seq, seq)); 551 552 if (mono) 553 *mono = basem + delta; 554 return baser + delta; 555 } 556 557 /** 558 * ktime_get_real_fast_ns: - NMI safe and fast access to clock realtime. 559 * 560 * See ktime_get_fast_ns() for documentation of the time stamp ordering. 561 */ 562 u64 ktime_get_real_fast_ns(void) 563 { 564 return __ktime_get_real_fast(&tk_fast_mono, NULL); 565 } 566 EXPORT_SYMBOL_GPL(ktime_get_real_fast_ns); 567 568 /** 569 * ktime_get_fast_timestamps: - NMI safe timestamps 570 * @snapshot: Pointer to timestamp storage 571 * 572 * Stores clock monotonic, boottime and realtime timestamps. 573 * 574 * Boot time is a racy access on 32bit systems if the sleep time injection 575 * happens late during resume and not in timekeeping_resume(). That could 576 * be avoided by expanding struct tk_read_base with boot offset for 32bit 577 * and adding more overhead to the update. As this is a hard to observe 578 * once per resume event which can be filtered with reasonable effort using 579 * the accurate mono/real timestamps, it's probably not worth the trouble. 580 * 581 * Aside of that it might be possible on 32 and 64 bit to observe the 582 * following when the sleep time injection happens late: 583 * 584 * CPU 0 CPU 1 585 * timekeeping_resume() 586 * ktime_get_fast_timestamps() 587 * mono, real = __ktime_get_real_fast() 588 * inject_sleep_time() 589 * update boot offset 590 * boot = mono + bootoffset; 591 * 592 * That means that boot time already has the sleep time adjustment, but 593 * real time does not. On the next readout both are in sync again. 594 * 595 * Preventing this for 64bit is not really feasible without destroying the 596 * careful cache layout of the timekeeper because the sequence count and 597 * struct tk_read_base would then need two cache lines instead of one. 598 * 599 * Access to the time keeper clock source is disabled accross the innermost 600 * steps of suspend/resume. The accessors still work, but the timestamps 601 * are frozen until time keeping is resumed which happens very early. 602 * 603 * For regular suspend/resume there is no observable difference vs. sched 604 * clock, but it might affect some of the nasty low level debug printks. 605 * 606 * OTOH, access to sched clock is not guaranteed accross suspend/resume on 607 * all systems either so it depends on the hardware in use. 608 * 609 * If that turns out to be a real problem then this could be mitigated by 610 * using sched clock in a similar way as during early boot. But it's not as 611 * trivial as on early boot because it needs some careful protection 612 * against the clock monotonic timestamp jumping backwards on resume. 613 */ 614 void ktime_get_fast_timestamps(struct ktime_timestamps *snapshot) 615 { 616 struct timekeeper *tk = &tk_core.timekeeper; 617 618 snapshot->real = __ktime_get_real_fast(&tk_fast_mono, &snapshot->mono); 619 snapshot->boot = snapshot->mono + ktime_to_ns(data_race(tk->offs_boot)); 620 } 621 622 /** 623 * halt_fast_timekeeper - Prevent fast timekeeper from accessing clocksource. 624 * @tk: Timekeeper to snapshot. 625 * 626 * It generally is unsafe to access the clocksource after timekeeping has been 627 * suspended, so take a snapshot of the readout base of @tk and use it as the 628 * fast timekeeper's readout base while suspended. It will return the same 629 * number of cycles every time until timekeeping is resumed at which time the 630 * proper readout base for the fast timekeeper will be restored automatically. 631 */ 632 static void halt_fast_timekeeper(const struct timekeeper *tk) 633 { 634 static struct tk_read_base tkr_dummy; 635 const struct tk_read_base *tkr = &tk->tkr_mono; 636 637 memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy)); 638 cycles_at_suspend = tk_clock_read(tkr); 639 tkr_dummy.clock = &dummy_clock; 640 tkr_dummy.base_real = tkr->base + tk->offs_real; 641 update_fast_timekeeper(&tkr_dummy, &tk_fast_mono); 642 643 tkr = &tk->tkr_raw; 644 memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy)); 645 tkr_dummy.clock = &dummy_clock; 646 update_fast_timekeeper(&tkr_dummy, &tk_fast_raw); 647 } 648 649 static RAW_NOTIFIER_HEAD(pvclock_gtod_chain); 650 651 static void update_pvclock_gtod(struct timekeeper *tk, bool was_set) 652 { 653 raw_notifier_call_chain(&pvclock_gtod_chain, was_set, tk); 654 } 655 656 /** 657 * pvclock_gtod_register_notifier - register a pvclock timedata update listener 658 * @nb: Pointer to the notifier block to register 659 */ 660 int pvclock_gtod_register_notifier(struct notifier_block *nb) 661 { 662 struct timekeeper *tk = &tk_core.timekeeper; 663 unsigned long flags; 664 int ret; 665 666 raw_spin_lock_irqsave(&timekeeper_lock, flags); 667 ret = raw_notifier_chain_register(&pvclock_gtod_chain, nb); 668 update_pvclock_gtod(tk, true); 669 raw_spin_unlock_irqrestore(&timekeeper_lock, flags); 670 671 return ret; 672 } 673 EXPORT_SYMBOL_GPL(pvclock_gtod_register_notifier); 674 675 /** 676 * pvclock_gtod_unregister_notifier - unregister a pvclock 677 * timedata update listener 678 * @nb: Pointer to the notifier block to unregister 679 */ 680 int pvclock_gtod_unregister_notifier(struct notifier_block *nb) 681 { 682 unsigned long flags; 683 int ret; 684 685 raw_spin_lock_irqsave(&timekeeper_lock, flags); 686 ret = raw_notifier_chain_unregister(&pvclock_gtod_chain, nb); 687 raw_spin_unlock_irqrestore(&timekeeper_lock, flags); 688 689 return ret; 690 } 691 EXPORT_SYMBOL_GPL(pvclock_gtod_unregister_notifier); 692 693 /* 694 * tk_update_leap_state - helper to update the next_leap_ktime 695 */ 696 static inline void tk_update_leap_state(struct timekeeper *tk) 697 { 698 tk->next_leap_ktime = ntp_get_next_leap(); 699 if (tk->next_leap_ktime != KTIME_MAX) 700 /* Convert to monotonic time */ 701 tk->next_leap_ktime = ktime_sub(tk->next_leap_ktime, tk->offs_real); 702 } 703 704 /* 705 * Update the ktime_t based scalar nsec members of the timekeeper 706 */ 707 static inline void tk_update_ktime_data(struct timekeeper *tk) 708 { 709 u64 seconds; 710 u32 nsec; 711 712 /* 713 * The xtime based monotonic readout is: 714 * nsec = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec + now(); 715 * The ktime based monotonic readout is: 716 * nsec = base_mono + now(); 717 * ==> base_mono = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec 718 */ 719 seconds = (u64)(tk->xtime_sec + tk->wall_to_monotonic.tv_sec); 720 nsec = (u32) tk->wall_to_monotonic.tv_nsec; 721 tk->tkr_mono.base = ns_to_ktime(seconds * NSEC_PER_SEC + nsec); 722 723 /* 724 * The sum of the nanoseconds portions of xtime and 725 * wall_to_monotonic can be greater/equal one second. Take 726 * this into account before updating tk->ktime_sec. 727 */ 728 nsec += (u32)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift); 729 if (nsec >= NSEC_PER_SEC) 730 seconds++; 731 tk->ktime_sec = seconds; 732 733 /* Update the monotonic raw base */ 734 tk->tkr_raw.base = ns_to_ktime(tk->raw_sec * NSEC_PER_SEC); 735 } 736 737 /* must hold timekeeper_lock */ 738 static void timekeeping_update(struct timekeeper *tk, unsigned int action) 739 { 740 if (action & TK_CLEAR_NTP) { 741 tk->ntp_error = 0; 742 ntp_clear(); 743 } 744 745 tk_update_leap_state(tk); 746 tk_update_ktime_data(tk); 747 748 update_vsyscall(tk); 749 update_pvclock_gtod(tk, action & TK_CLOCK_WAS_SET); 750 751 tk->tkr_mono.base_real = tk->tkr_mono.base + tk->offs_real; 752 update_fast_timekeeper(&tk->tkr_mono, &tk_fast_mono); 753 update_fast_timekeeper(&tk->tkr_raw, &tk_fast_raw); 754 755 if (action & TK_CLOCK_WAS_SET) 756 tk->clock_was_set_seq++; 757 /* 758 * The mirroring of the data to the shadow-timekeeper needs 759 * to happen last here to ensure we don't over-write the 760 * timekeeper structure on the next update with stale data 761 */ 762 if (action & TK_MIRROR) 763 memcpy(&shadow_timekeeper, &tk_core.timekeeper, 764 sizeof(tk_core.timekeeper)); 765 } 766 767 /** 768 * timekeeping_forward_now - update clock to the current time 769 * @tk: Pointer to the timekeeper to update 770 * 771 * Forward the current clock to update its state since the last call to 772 * update_wall_time(). This is useful before significant clock changes, 773 * as it avoids having to deal with this time offset explicitly. 774 */ 775 static void timekeeping_forward_now(struct timekeeper *tk) 776 { 777 u64 cycle_now, delta; 778 779 cycle_now = tk_clock_read(&tk->tkr_mono); 780 delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last, tk->tkr_mono.mask); 781 tk->tkr_mono.cycle_last = cycle_now; 782 tk->tkr_raw.cycle_last = cycle_now; 783 784 tk->tkr_mono.xtime_nsec += delta * tk->tkr_mono.mult; 785 tk->tkr_raw.xtime_nsec += delta * tk->tkr_raw.mult; 786 787 tk_normalize_xtime(tk); 788 } 789 790 /** 791 * ktime_get_real_ts64 - Returns the time of day in a timespec64. 792 * @ts: pointer to the timespec to be set 793 * 794 * Returns the time of day in a timespec64 (WARN if suspended). 795 */ 796 void ktime_get_real_ts64(struct timespec64 *ts) 797 { 798 struct timekeeper *tk = &tk_core.timekeeper; 799 unsigned int seq; 800 u64 nsecs; 801 802 WARN_ON(timekeeping_suspended); 803 804 do { 805 seq = read_seqcount_begin(&tk_core.seq); 806 807 ts->tv_sec = tk->xtime_sec; 808 nsecs = timekeeping_get_ns(&tk->tkr_mono); 809 810 } while (read_seqcount_retry(&tk_core.seq, seq)); 811 812 ts->tv_nsec = 0; 813 timespec64_add_ns(ts, nsecs); 814 } 815 EXPORT_SYMBOL(ktime_get_real_ts64); 816 817 ktime_t ktime_get(void) 818 { 819 struct timekeeper *tk = &tk_core.timekeeper; 820 unsigned int seq; 821 ktime_t base; 822 u64 nsecs; 823 824 WARN_ON(timekeeping_suspended); 825 826 do { 827 seq = read_seqcount_begin(&tk_core.seq); 828 base = tk->tkr_mono.base; 829 nsecs = timekeeping_get_ns(&tk->tkr_mono); 830 831 } while (read_seqcount_retry(&tk_core.seq, seq)); 832 833 return ktime_add_ns(base, nsecs); 834 } 835 EXPORT_SYMBOL_GPL(ktime_get); 836 837 u32 ktime_get_resolution_ns(void) 838 { 839 struct timekeeper *tk = &tk_core.timekeeper; 840 unsigned int seq; 841 u32 nsecs; 842 843 WARN_ON(timekeeping_suspended); 844 845 do { 846 seq = read_seqcount_begin(&tk_core.seq); 847 nsecs = tk->tkr_mono.mult >> tk->tkr_mono.shift; 848 } while (read_seqcount_retry(&tk_core.seq, seq)); 849 850 return nsecs; 851 } 852 EXPORT_SYMBOL_GPL(ktime_get_resolution_ns); 853 854 static ktime_t *offsets[TK_OFFS_MAX] = { 855 [TK_OFFS_REAL] = &tk_core.timekeeper.offs_real, 856 [TK_OFFS_BOOT] = &tk_core.timekeeper.offs_boot, 857 [TK_OFFS_TAI] = &tk_core.timekeeper.offs_tai, 858 }; 859 860 ktime_t ktime_get_with_offset(enum tk_offsets offs) 861 { 862 struct timekeeper *tk = &tk_core.timekeeper; 863 unsigned int seq; 864 ktime_t base, *offset = offsets[offs]; 865 u64 nsecs; 866 867 WARN_ON(timekeeping_suspended); 868 869 do { 870 seq = read_seqcount_begin(&tk_core.seq); 871 base = ktime_add(tk->tkr_mono.base, *offset); 872 nsecs = timekeeping_get_ns(&tk->tkr_mono); 873 874 } while (read_seqcount_retry(&tk_core.seq, seq)); 875 876 return ktime_add_ns(base, nsecs); 877 878 } 879 EXPORT_SYMBOL_GPL(ktime_get_with_offset); 880 881 ktime_t ktime_get_coarse_with_offset(enum tk_offsets offs) 882 { 883 struct timekeeper *tk = &tk_core.timekeeper; 884 unsigned int seq; 885 ktime_t base, *offset = offsets[offs]; 886 u64 nsecs; 887 888 WARN_ON(timekeeping_suspended); 889 890 do { 891 seq = read_seqcount_begin(&tk_core.seq); 892 base = ktime_add(tk->tkr_mono.base, *offset); 893 nsecs = tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift; 894 895 } while (read_seqcount_retry(&tk_core.seq, seq)); 896 897 return ktime_add_ns(base, nsecs); 898 } 899 EXPORT_SYMBOL_GPL(ktime_get_coarse_with_offset); 900 901 /** 902 * ktime_mono_to_any() - convert mononotic time to any other time 903 * @tmono: time to convert. 904 * @offs: which offset to use 905 */ 906 ktime_t ktime_mono_to_any(ktime_t tmono, enum tk_offsets offs) 907 { 908 ktime_t *offset = offsets[offs]; 909 unsigned int seq; 910 ktime_t tconv; 911 912 do { 913 seq = read_seqcount_begin(&tk_core.seq); 914 tconv = ktime_add(tmono, *offset); 915 } while (read_seqcount_retry(&tk_core.seq, seq)); 916 917 return tconv; 918 } 919 EXPORT_SYMBOL_GPL(ktime_mono_to_any); 920 921 /** 922 * ktime_get_raw - Returns the raw monotonic time in ktime_t format 923 */ 924 ktime_t ktime_get_raw(void) 925 { 926 struct timekeeper *tk = &tk_core.timekeeper; 927 unsigned int seq; 928 ktime_t base; 929 u64 nsecs; 930 931 do { 932 seq = read_seqcount_begin(&tk_core.seq); 933 base = tk->tkr_raw.base; 934 nsecs = timekeeping_get_ns(&tk->tkr_raw); 935 936 } while (read_seqcount_retry(&tk_core.seq, seq)); 937 938 return ktime_add_ns(base, nsecs); 939 } 940 EXPORT_SYMBOL_GPL(ktime_get_raw); 941 942 /** 943 * ktime_get_ts64 - get the monotonic clock in timespec64 format 944 * @ts: pointer to timespec variable 945 * 946 * The function calculates the monotonic clock from the realtime 947 * clock and the wall_to_monotonic offset and stores the result 948 * in normalized timespec64 format in the variable pointed to by @ts. 949 */ 950 void ktime_get_ts64(struct timespec64 *ts) 951 { 952 struct timekeeper *tk = &tk_core.timekeeper; 953 struct timespec64 tomono; 954 unsigned int seq; 955 u64 nsec; 956 957 WARN_ON(timekeeping_suspended); 958 959 do { 960 seq = read_seqcount_begin(&tk_core.seq); 961 ts->tv_sec = tk->xtime_sec; 962 nsec = timekeeping_get_ns(&tk->tkr_mono); 963 tomono = tk->wall_to_monotonic; 964 965 } while (read_seqcount_retry(&tk_core.seq, seq)); 966 967 ts->tv_sec += tomono.tv_sec; 968 ts->tv_nsec = 0; 969 timespec64_add_ns(ts, nsec + tomono.tv_nsec); 970 } 971 EXPORT_SYMBOL_GPL(ktime_get_ts64); 972 973 /** 974 * ktime_get_seconds - Get the seconds portion of CLOCK_MONOTONIC 975 * 976 * Returns the seconds portion of CLOCK_MONOTONIC with a single non 977 * serialized read. tk->ktime_sec is of type 'unsigned long' so this 978 * works on both 32 and 64 bit systems. On 32 bit systems the readout 979 * covers ~136 years of uptime which should be enough to prevent 980 * premature wrap arounds. 981 */ 982 time64_t ktime_get_seconds(void) 983 { 984 struct timekeeper *tk = &tk_core.timekeeper; 985 986 WARN_ON(timekeeping_suspended); 987 return tk->ktime_sec; 988 } 989 EXPORT_SYMBOL_GPL(ktime_get_seconds); 990 991 /** 992 * ktime_get_real_seconds - Get the seconds portion of CLOCK_REALTIME 993 * 994 * Returns the wall clock seconds since 1970. This replaces the 995 * get_seconds() interface which is not y2038 safe on 32bit systems. 996 * 997 * For 64bit systems the fast access to tk->xtime_sec is preserved. On 998 * 32bit systems the access must be protected with the sequence 999 * counter to provide "atomic" access to the 64bit tk->xtime_sec 1000 * value. 1001 */ 1002 time64_t ktime_get_real_seconds(void) 1003 { 1004 struct timekeeper *tk = &tk_core.timekeeper; 1005 time64_t seconds; 1006 unsigned int seq; 1007 1008 if (IS_ENABLED(CONFIG_64BIT)) 1009 return tk->xtime_sec; 1010 1011 do { 1012 seq = read_seqcount_begin(&tk_core.seq); 1013 seconds = tk->xtime_sec; 1014 1015 } while (read_seqcount_retry(&tk_core.seq, seq)); 1016 1017 return seconds; 1018 } 1019 EXPORT_SYMBOL_GPL(ktime_get_real_seconds); 1020 1021 /** 1022 * __ktime_get_real_seconds - The same as ktime_get_real_seconds 1023 * but without the sequence counter protect. This internal function 1024 * is called just when timekeeping lock is already held. 1025 */ 1026 noinstr time64_t __ktime_get_real_seconds(void) 1027 { 1028 struct timekeeper *tk = &tk_core.timekeeper; 1029 1030 return tk->xtime_sec; 1031 } 1032 1033 /** 1034 * ktime_get_snapshot - snapshots the realtime/monotonic raw clocks with counter 1035 * @systime_snapshot: pointer to struct receiving the system time snapshot 1036 */ 1037 void ktime_get_snapshot(struct system_time_snapshot *systime_snapshot) 1038 { 1039 struct timekeeper *tk = &tk_core.timekeeper; 1040 unsigned int seq; 1041 ktime_t base_raw; 1042 ktime_t base_real; 1043 u64 nsec_raw; 1044 u64 nsec_real; 1045 u64 now; 1046 1047 WARN_ON_ONCE(timekeeping_suspended); 1048 1049 do { 1050 seq = read_seqcount_begin(&tk_core.seq); 1051 now = tk_clock_read(&tk->tkr_mono); 1052 systime_snapshot->cs_was_changed_seq = tk->cs_was_changed_seq; 1053 systime_snapshot->clock_was_set_seq = tk->clock_was_set_seq; 1054 base_real = ktime_add(tk->tkr_mono.base, 1055 tk_core.timekeeper.offs_real); 1056 base_raw = tk->tkr_raw.base; 1057 nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono, now); 1058 nsec_raw = timekeeping_cycles_to_ns(&tk->tkr_raw, now); 1059 } while (read_seqcount_retry(&tk_core.seq, seq)); 1060 1061 systime_snapshot->cycles = now; 1062 systime_snapshot->real = ktime_add_ns(base_real, nsec_real); 1063 systime_snapshot->raw = ktime_add_ns(base_raw, nsec_raw); 1064 } 1065 EXPORT_SYMBOL_GPL(ktime_get_snapshot); 1066 1067 /* Scale base by mult/div checking for overflow */ 1068 static int scale64_check_overflow(u64 mult, u64 div, u64 *base) 1069 { 1070 u64 tmp, rem; 1071 1072 tmp = div64_u64_rem(*base, div, &rem); 1073 1074 if (((int)sizeof(u64)*8 - fls64(mult) < fls64(tmp)) || 1075 ((int)sizeof(u64)*8 - fls64(mult) < fls64(rem))) 1076 return -EOVERFLOW; 1077 tmp *= mult; 1078 1079 rem = div64_u64(rem * mult, div); 1080 *base = tmp + rem; 1081 return 0; 1082 } 1083 1084 /** 1085 * adjust_historical_crosststamp - adjust crosstimestamp previous to current interval 1086 * @history: Snapshot representing start of history 1087 * @partial_history_cycles: Cycle offset into history (fractional part) 1088 * @total_history_cycles: Total history length in cycles 1089 * @discontinuity: True indicates clock was set on history period 1090 * @ts: Cross timestamp that should be adjusted using 1091 * partial/total ratio 1092 * 1093 * Helper function used by get_device_system_crosststamp() to correct the 1094 * crosstimestamp corresponding to the start of the current interval to the 1095 * system counter value (timestamp point) provided by the driver. The 1096 * total_history_* quantities are the total history starting at the provided 1097 * reference point and ending at the start of the current interval. The cycle 1098 * count between the driver timestamp point and the start of the current 1099 * interval is partial_history_cycles. 1100 */ 1101 static int adjust_historical_crosststamp(struct system_time_snapshot *history, 1102 u64 partial_history_cycles, 1103 u64 total_history_cycles, 1104 bool discontinuity, 1105 struct system_device_crosststamp *ts) 1106 { 1107 struct timekeeper *tk = &tk_core.timekeeper; 1108 u64 corr_raw, corr_real; 1109 bool interp_forward; 1110 int ret; 1111 1112 if (total_history_cycles == 0 || partial_history_cycles == 0) 1113 return 0; 1114 1115 /* Interpolate shortest distance from beginning or end of history */ 1116 interp_forward = partial_history_cycles > total_history_cycles / 2; 1117 partial_history_cycles = interp_forward ? 1118 total_history_cycles - partial_history_cycles : 1119 partial_history_cycles; 1120 1121 /* 1122 * Scale the monotonic raw time delta by: 1123 * partial_history_cycles / total_history_cycles 1124 */ 1125 corr_raw = (u64)ktime_to_ns( 1126 ktime_sub(ts->sys_monoraw, history->raw)); 1127 ret = scale64_check_overflow(partial_history_cycles, 1128 total_history_cycles, &corr_raw); 1129 if (ret) 1130 return ret; 1131 1132 /* 1133 * If there is a discontinuity in the history, scale monotonic raw 1134 * correction by: 1135 * mult(real)/mult(raw) yielding the realtime correction 1136 * Otherwise, calculate the realtime correction similar to monotonic 1137 * raw calculation 1138 */ 1139 if (discontinuity) { 1140 corr_real = mul_u64_u32_div 1141 (corr_raw, tk->tkr_mono.mult, tk->tkr_raw.mult); 1142 } else { 1143 corr_real = (u64)ktime_to_ns( 1144 ktime_sub(ts->sys_realtime, history->real)); 1145 ret = scale64_check_overflow(partial_history_cycles, 1146 total_history_cycles, &corr_real); 1147 if (ret) 1148 return ret; 1149 } 1150 1151 /* Fixup monotonic raw and real time time values */ 1152 if (interp_forward) { 1153 ts->sys_monoraw = ktime_add_ns(history->raw, corr_raw); 1154 ts->sys_realtime = ktime_add_ns(history->real, corr_real); 1155 } else { 1156 ts->sys_monoraw = ktime_sub_ns(ts->sys_monoraw, corr_raw); 1157 ts->sys_realtime = ktime_sub_ns(ts->sys_realtime, corr_real); 1158 } 1159 1160 return 0; 1161 } 1162 1163 /* 1164 * cycle_between - true if test occurs chronologically between before and after 1165 */ 1166 static bool cycle_between(u64 before, u64 test, u64 after) 1167 { 1168 if (test > before && test < after) 1169 return true; 1170 if (test < before && before > after) 1171 return true; 1172 return false; 1173 } 1174 1175 /** 1176 * get_device_system_crosststamp - Synchronously capture system/device timestamp 1177 * @get_time_fn: Callback to get simultaneous device time and 1178 * system counter from the device driver 1179 * @ctx: Context passed to get_time_fn() 1180 * @history_begin: Historical reference point used to interpolate system 1181 * time when counter provided by the driver is before the current interval 1182 * @xtstamp: Receives simultaneously captured system and device time 1183 * 1184 * Reads a timestamp from a device and correlates it to system time 1185 */ 1186 int get_device_system_crosststamp(int (*get_time_fn) 1187 (ktime_t *device_time, 1188 struct system_counterval_t *sys_counterval, 1189 void *ctx), 1190 void *ctx, 1191 struct system_time_snapshot *history_begin, 1192 struct system_device_crosststamp *xtstamp) 1193 { 1194 struct system_counterval_t system_counterval; 1195 struct timekeeper *tk = &tk_core.timekeeper; 1196 u64 cycles, now, interval_start; 1197 unsigned int clock_was_set_seq = 0; 1198 ktime_t base_real, base_raw; 1199 u64 nsec_real, nsec_raw; 1200 u8 cs_was_changed_seq; 1201 unsigned int seq; 1202 bool do_interp; 1203 int ret; 1204 1205 do { 1206 seq = read_seqcount_begin(&tk_core.seq); 1207 /* 1208 * Try to synchronously capture device time and a system 1209 * counter value calling back into the device driver 1210 */ 1211 ret = get_time_fn(&xtstamp->device, &system_counterval, ctx); 1212 if (ret) 1213 return ret; 1214 1215 /* 1216 * Verify that the clocksource associated with the captured 1217 * system counter value is the same as the currently installed 1218 * timekeeper clocksource 1219 */ 1220 if (tk->tkr_mono.clock != system_counterval.cs) 1221 return -ENODEV; 1222 cycles = system_counterval.cycles; 1223 1224 /* 1225 * Check whether the system counter value provided by the 1226 * device driver is on the current timekeeping interval. 1227 */ 1228 now = tk_clock_read(&tk->tkr_mono); 1229 interval_start = tk->tkr_mono.cycle_last; 1230 if (!cycle_between(interval_start, cycles, now)) { 1231 clock_was_set_seq = tk->clock_was_set_seq; 1232 cs_was_changed_seq = tk->cs_was_changed_seq; 1233 cycles = interval_start; 1234 do_interp = true; 1235 } else { 1236 do_interp = false; 1237 } 1238 1239 base_real = ktime_add(tk->tkr_mono.base, 1240 tk_core.timekeeper.offs_real); 1241 base_raw = tk->tkr_raw.base; 1242 1243 nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono, 1244 system_counterval.cycles); 1245 nsec_raw = timekeeping_cycles_to_ns(&tk->tkr_raw, 1246 system_counterval.cycles); 1247 } while (read_seqcount_retry(&tk_core.seq, seq)); 1248 1249 xtstamp->sys_realtime = ktime_add_ns(base_real, nsec_real); 1250 xtstamp->sys_monoraw = ktime_add_ns(base_raw, nsec_raw); 1251 1252 /* 1253 * Interpolate if necessary, adjusting back from the start of the 1254 * current interval 1255 */ 1256 if (do_interp) { 1257 u64 partial_history_cycles, total_history_cycles; 1258 bool discontinuity; 1259 1260 /* 1261 * Check that the counter value occurs after the provided 1262 * history reference and that the history doesn't cross a 1263 * clocksource change 1264 */ 1265 if (!history_begin || 1266 !cycle_between(history_begin->cycles, 1267 system_counterval.cycles, cycles) || 1268 history_begin->cs_was_changed_seq != cs_was_changed_seq) 1269 return -EINVAL; 1270 partial_history_cycles = cycles - system_counterval.cycles; 1271 total_history_cycles = cycles - history_begin->cycles; 1272 discontinuity = 1273 history_begin->clock_was_set_seq != clock_was_set_seq; 1274 1275 ret = adjust_historical_crosststamp(history_begin, 1276 partial_history_cycles, 1277 total_history_cycles, 1278 discontinuity, xtstamp); 1279 if (ret) 1280 return ret; 1281 } 1282 1283 return 0; 1284 } 1285 EXPORT_SYMBOL_GPL(get_device_system_crosststamp); 1286 1287 /** 1288 * do_settimeofday64 - Sets the time of day. 1289 * @ts: pointer to the timespec64 variable containing the new time 1290 * 1291 * Sets the time of day to the new time and update NTP and notify hrtimers 1292 */ 1293 int do_settimeofday64(const struct timespec64 *ts) 1294 { 1295 struct timekeeper *tk = &tk_core.timekeeper; 1296 struct timespec64 ts_delta, xt; 1297 unsigned long flags; 1298 int ret = 0; 1299 1300 if (!timespec64_valid_settod(ts)) 1301 return -EINVAL; 1302 1303 raw_spin_lock_irqsave(&timekeeper_lock, flags); 1304 write_seqcount_begin(&tk_core.seq); 1305 1306 timekeeping_forward_now(tk); 1307 1308 xt = tk_xtime(tk); 1309 ts_delta.tv_sec = ts->tv_sec - xt.tv_sec; 1310 ts_delta.tv_nsec = ts->tv_nsec - xt.tv_nsec; 1311 1312 if (timespec64_compare(&tk->wall_to_monotonic, &ts_delta) > 0) { 1313 ret = -EINVAL; 1314 goto out; 1315 } 1316 1317 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts_delta)); 1318 1319 tk_set_xtime(tk, ts); 1320 out: 1321 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET); 1322 1323 write_seqcount_end(&tk_core.seq); 1324 raw_spin_unlock_irqrestore(&timekeeper_lock, flags); 1325 1326 /* signal hrtimers about time change */ 1327 clock_was_set(); 1328 1329 if (!ret) 1330 audit_tk_injoffset(ts_delta); 1331 1332 return ret; 1333 } 1334 EXPORT_SYMBOL(do_settimeofday64); 1335 1336 /** 1337 * timekeeping_inject_offset - Adds or subtracts from the current time. 1338 * @ts: Pointer to the timespec variable containing the offset 1339 * 1340 * Adds or subtracts an offset value from the current time. 1341 */ 1342 static int timekeeping_inject_offset(const struct timespec64 *ts) 1343 { 1344 struct timekeeper *tk = &tk_core.timekeeper; 1345 unsigned long flags; 1346 struct timespec64 tmp; 1347 int ret = 0; 1348 1349 if (ts->tv_nsec < 0 || ts->tv_nsec >= NSEC_PER_SEC) 1350 return -EINVAL; 1351 1352 raw_spin_lock_irqsave(&timekeeper_lock, flags); 1353 write_seqcount_begin(&tk_core.seq); 1354 1355 timekeeping_forward_now(tk); 1356 1357 /* Make sure the proposed value is valid */ 1358 tmp = timespec64_add(tk_xtime(tk), *ts); 1359 if (timespec64_compare(&tk->wall_to_monotonic, ts) > 0 || 1360 !timespec64_valid_settod(&tmp)) { 1361 ret = -EINVAL; 1362 goto error; 1363 } 1364 1365 tk_xtime_add(tk, ts); 1366 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *ts)); 1367 1368 error: /* even if we error out, we forwarded the time, so call update */ 1369 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET); 1370 1371 write_seqcount_end(&tk_core.seq); 1372 raw_spin_unlock_irqrestore(&timekeeper_lock, flags); 1373 1374 /* signal hrtimers about time change */ 1375 clock_was_set(); 1376 1377 return ret; 1378 } 1379 1380 /* 1381 * Indicates if there is an offset between the system clock and the hardware 1382 * clock/persistent clock/rtc. 1383 */ 1384 int persistent_clock_is_local; 1385 1386 /* 1387 * Adjust the time obtained from the CMOS to be UTC time instead of 1388 * local time. 1389 * 1390 * This is ugly, but preferable to the alternatives. Otherwise we 1391 * would either need to write a program to do it in /etc/rc (and risk 1392 * confusion if the program gets run more than once; it would also be 1393 * hard to make the program warp the clock precisely n hours) or 1394 * compile in the timezone information into the kernel. Bad, bad.... 1395 * 1396 * - TYT, 1992-01-01 1397 * 1398 * The best thing to do is to keep the CMOS clock in universal time (UTC) 1399 * as real UNIX machines always do it. This avoids all headaches about 1400 * daylight saving times and warping kernel clocks. 1401 */ 1402 void timekeeping_warp_clock(void) 1403 { 1404 if (sys_tz.tz_minuteswest != 0) { 1405 struct timespec64 adjust; 1406 1407 persistent_clock_is_local = 1; 1408 adjust.tv_sec = sys_tz.tz_minuteswest * 60; 1409 adjust.tv_nsec = 0; 1410 timekeeping_inject_offset(&adjust); 1411 } 1412 } 1413 1414 /* 1415 * __timekeeping_set_tai_offset - Sets the TAI offset from UTC and monotonic 1416 */ 1417 static void __timekeeping_set_tai_offset(struct timekeeper *tk, s32 tai_offset) 1418 { 1419 tk->tai_offset = tai_offset; 1420 tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tai_offset, 0)); 1421 } 1422 1423 /* 1424 * change_clocksource - Swaps clocksources if a new one is available 1425 * 1426 * Accumulates current time interval and initializes new clocksource 1427 */ 1428 static int change_clocksource(void *data) 1429 { 1430 struct timekeeper *tk = &tk_core.timekeeper; 1431 struct clocksource *new, *old; 1432 unsigned long flags; 1433 1434 new = (struct clocksource *) data; 1435 1436 raw_spin_lock_irqsave(&timekeeper_lock, flags); 1437 write_seqcount_begin(&tk_core.seq); 1438 1439 timekeeping_forward_now(tk); 1440 /* 1441 * If the cs is in module, get a module reference. Succeeds 1442 * for built-in code (owner == NULL) as well. 1443 */ 1444 if (try_module_get(new->owner)) { 1445 if (!new->enable || new->enable(new) == 0) { 1446 old = tk->tkr_mono.clock; 1447 tk_setup_internals(tk, new); 1448 if (old->disable) 1449 old->disable(old); 1450 module_put(old->owner); 1451 } else { 1452 module_put(new->owner); 1453 } 1454 } 1455 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET); 1456 1457 write_seqcount_end(&tk_core.seq); 1458 raw_spin_unlock_irqrestore(&timekeeper_lock, flags); 1459 1460 return 0; 1461 } 1462 1463 /** 1464 * timekeeping_notify - Install a new clock source 1465 * @clock: pointer to the clock source 1466 * 1467 * This function is called from clocksource.c after a new, better clock 1468 * source has been registered. The caller holds the clocksource_mutex. 1469 */ 1470 int timekeeping_notify(struct clocksource *clock) 1471 { 1472 struct timekeeper *tk = &tk_core.timekeeper; 1473 1474 if (tk->tkr_mono.clock == clock) 1475 return 0; 1476 stop_machine(change_clocksource, clock, NULL); 1477 tick_clock_notify(); 1478 return tk->tkr_mono.clock == clock ? 0 : -1; 1479 } 1480 1481 /** 1482 * ktime_get_raw_ts64 - Returns the raw monotonic time in a timespec 1483 * @ts: pointer to the timespec64 to be set 1484 * 1485 * Returns the raw monotonic time (completely un-modified by ntp) 1486 */ 1487 void ktime_get_raw_ts64(struct timespec64 *ts) 1488 { 1489 struct timekeeper *tk = &tk_core.timekeeper; 1490 unsigned int seq; 1491 u64 nsecs; 1492 1493 do { 1494 seq = read_seqcount_begin(&tk_core.seq); 1495 ts->tv_sec = tk->raw_sec; 1496 nsecs = timekeeping_get_ns(&tk->tkr_raw); 1497 1498 } while (read_seqcount_retry(&tk_core.seq, seq)); 1499 1500 ts->tv_nsec = 0; 1501 timespec64_add_ns(ts, nsecs); 1502 } 1503 EXPORT_SYMBOL(ktime_get_raw_ts64); 1504 1505 1506 /** 1507 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres 1508 */ 1509 int timekeeping_valid_for_hres(void) 1510 { 1511 struct timekeeper *tk = &tk_core.timekeeper; 1512 unsigned int seq; 1513 int ret; 1514 1515 do { 1516 seq = read_seqcount_begin(&tk_core.seq); 1517 1518 ret = tk->tkr_mono.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES; 1519 1520 } while (read_seqcount_retry(&tk_core.seq, seq)); 1521 1522 return ret; 1523 } 1524 1525 /** 1526 * timekeeping_max_deferment - Returns max time the clocksource can be deferred 1527 */ 1528 u64 timekeeping_max_deferment(void) 1529 { 1530 struct timekeeper *tk = &tk_core.timekeeper; 1531 unsigned int seq; 1532 u64 ret; 1533 1534 do { 1535 seq = read_seqcount_begin(&tk_core.seq); 1536 1537 ret = tk->tkr_mono.clock->max_idle_ns; 1538 1539 } while (read_seqcount_retry(&tk_core.seq, seq)); 1540 1541 return ret; 1542 } 1543 1544 /** 1545 * read_persistent_clock64 - Return time from the persistent clock. 1546 * @ts: Pointer to the storage for the readout value 1547 * 1548 * Weak dummy function for arches that do not yet support it. 1549 * Reads the time from the battery backed persistent clock. 1550 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported. 1551 * 1552 * XXX - Do be sure to remove it once all arches implement it. 1553 */ 1554 void __weak read_persistent_clock64(struct timespec64 *ts) 1555 { 1556 ts->tv_sec = 0; 1557 ts->tv_nsec = 0; 1558 } 1559 1560 /** 1561 * read_persistent_wall_and_boot_offset - Read persistent clock, and also offset 1562 * from the boot. 1563 * 1564 * Weak dummy function for arches that do not yet support it. 1565 * @wall_time: - current time as returned by persistent clock 1566 * @boot_offset: - offset that is defined as wall_time - boot_time 1567 * 1568 * The default function calculates offset based on the current value of 1569 * local_clock(). This way architectures that support sched_clock() but don't 1570 * support dedicated boot time clock will provide the best estimate of the 1571 * boot time. 1572 */ 1573 void __weak __init 1574 read_persistent_wall_and_boot_offset(struct timespec64 *wall_time, 1575 struct timespec64 *boot_offset) 1576 { 1577 read_persistent_clock64(wall_time); 1578 *boot_offset = ns_to_timespec64(local_clock()); 1579 } 1580 1581 /* 1582 * Flag reflecting whether timekeeping_resume() has injected sleeptime. 1583 * 1584 * The flag starts of false and is only set when a suspend reaches 1585 * timekeeping_suspend(), timekeeping_resume() sets it to false when the 1586 * timekeeper clocksource is not stopping across suspend and has been 1587 * used to update sleep time. If the timekeeper clocksource has stopped 1588 * then the flag stays true and is used by the RTC resume code to decide 1589 * whether sleeptime must be injected and if so the flag gets false then. 1590 * 1591 * If a suspend fails before reaching timekeeping_resume() then the flag 1592 * stays false and prevents erroneous sleeptime injection. 1593 */ 1594 static bool suspend_timing_needed; 1595 1596 /* Flag for if there is a persistent clock on this platform */ 1597 static bool persistent_clock_exists; 1598 1599 /* 1600 * timekeeping_init - Initializes the clocksource and common timekeeping values 1601 */ 1602 void __init timekeeping_init(void) 1603 { 1604 struct timespec64 wall_time, boot_offset, wall_to_mono; 1605 struct timekeeper *tk = &tk_core.timekeeper; 1606 struct clocksource *clock; 1607 unsigned long flags; 1608 1609 read_persistent_wall_and_boot_offset(&wall_time, &boot_offset); 1610 if (timespec64_valid_settod(&wall_time) && 1611 timespec64_to_ns(&wall_time) > 0) { 1612 persistent_clock_exists = true; 1613 } else if (timespec64_to_ns(&wall_time) != 0) { 1614 pr_warn("Persistent clock returned invalid value"); 1615 wall_time = (struct timespec64){0}; 1616 } 1617 1618 if (timespec64_compare(&wall_time, &boot_offset) < 0) 1619 boot_offset = (struct timespec64){0}; 1620 1621 /* 1622 * We want set wall_to_mono, so the following is true: 1623 * wall time + wall_to_mono = boot time 1624 */ 1625 wall_to_mono = timespec64_sub(boot_offset, wall_time); 1626 1627 raw_spin_lock_irqsave(&timekeeper_lock, flags); 1628 write_seqcount_begin(&tk_core.seq); 1629 ntp_init(); 1630 1631 clock = clocksource_default_clock(); 1632 if (clock->enable) 1633 clock->enable(clock); 1634 tk_setup_internals(tk, clock); 1635 1636 tk_set_xtime(tk, &wall_time); 1637 tk->raw_sec = 0; 1638 1639 tk_set_wall_to_mono(tk, wall_to_mono); 1640 1641 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET); 1642 1643 write_seqcount_end(&tk_core.seq); 1644 raw_spin_unlock_irqrestore(&timekeeper_lock, flags); 1645 } 1646 1647 /* time in seconds when suspend began for persistent clock */ 1648 static struct timespec64 timekeeping_suspend_time; 1649 1650 /** 1651 * __timekeeping_inject_sleeptime - Internal function to add sleep interval 1652 * @tk: Pointer to the timekeeper to be updated 1653 * @delta: Pointer to the delta value in timespec64 format 1654 * 1655 * Takes a timespec offset measuring a suspend interval and properly 1656 * adds the sleep offset to the timekeeping variables. 1657 */ 1658 static void __timekeeping_inject_sleeptime(struct timekeeper *tk, 1659 const struct timespec64 *delta) 1660 { 1661 if (!timespec64_valid_strict(delta)) { 1662 printk_deferred(KERN_WARNING 1663 "__timekeeping_inject_sleeptime: Invalid " 1664 "sleep delta value!\n"); 1665 return; 1666 } 1667 tk_xtime_add(tk, delta); 1668 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *delta)); 1669 tk_update_sleep_time(tk, timespec64_to_ktime(*delta)); 1670 tk_debug_account_sleep_time(delta); 1671 } 1672 1673 #if defined(CONFIG_PM_SLEEP) && defined(CONFIG_RTC_HCTOSYS_DEVICE) 1674 /** 1675 * We have three kinds of time sources to use for sleep time 1676 * injection, the preference order is: 1677 * 1) non-stop clocksource 1678 * 2) persistent clock (ie: RTC accessible when irqs are off) 1679 * 3) RTC 1680 * 1681 * 1) and 2) are used by timekeeping, 3) by RTC subsystem. 1682 * If system has neither 1) nor 2), 3) will be used finally. 1683 * 1684 * 1685 * If timekeeping has injected sleeptime via either 1) or 2), 1686 * 3) becomes needless, so in this case we don't need to call 1687 * rtc_resume(), and this is what timekeeping_rtc_skipresume() 1688 * means. 1689 */ 1690 bool timekeeping_rtc_skipresume(void) 1691 { 1692 return !suspend_timing_needed; 1693 } 1694 1695 /** 1696 * 1) can be determined whether to use or not only when doing 1697 * timekeeping_resume() which is invoked after rtc_suspend(), 1698 * so we can't skip rtc_suspend() surely if system has 1). 1699 * 1700 * But if system has 2), 2) will definitely be used, so in this 1701 * case we don't need to call rtc_suspend(), and this is what 1702 * timekeeping_rtc_skipsuspend() means. 1703 */ 1704 bool timekeeping_rtc_skipsuspend(void) 1705 { 1706 return persistent_clock_exists; 1707 } 1708 1709 /** 1710 * timekeeping_inject_sleeptime64 - Adds suspend interval to timeekeeping values 1711 * @delta: pointer to a timespec64 delta value 1712 * 1713 * This hook is for architectures that cannot support read_persistent_clock64 1714 * because their RTC/persistent clock is only accessible when irqs are enabled. 1715 * and also don't have an effective nonstop clocksource. 1716 * 1717 * This function should only be called by rtc_resume(), and allows 1718 * a suspend offset to be injected into the timekeeping values. 1719 */ 1720 void timekeeping_inject_sleeptime64(const struct timespec64 *delta) 1721 { 1722 struct timekeeper *tk = &tk_core.timekeeper; 1723 unsigned long flags; 1724 1725 raw_spin_lock_irqsave(&timekeeper_lock, flags); 1726 write_seqcount_begin(&tk_core.seq); 1727 1728 suspend_timing_needed = false; 1729 1730 timekeeping_forward_now(tk); 1731 1732 __timekeeping_inject_sleeptime(tk, delta); 1733 1734 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET); 1735 1736 write_seqcount_end(&tk_core.seq); 1737 raw_spin_unlock_irqrestore(&timekeeper_lock, flags); 1738 1739 /* signal hrtimers about time change */ 1740 clock_was_set(); 1741 } 1742 #endif 1743 1744 /** 1745 * timekeeping_resume - Resumes the generic timekeeping subsystem. 1746 */ 1747 void timekeeping_resume(void) 1748 { 1749 struct timekeeper *tk = &tk_core.timekeeper; 1750 struct clocksource *clock = tk->tkr_mono.clock; 1751 unsigned long flags; 1752 struct timespec64 ts_new, ts_delta; 1753 u64 cycle_now, nsec; 1754 bool inject_sleeptime = false; 1755 1756 read_persistent_clock64(&ts_new); 1757 1758 clockevents_resume(); 1759 clocksource_resume(); 1760 1761 raw_spin_lock_irqsave(&timekeeper_lock, flags); 1762 write_seqcount_begin(&tk_core.seq); 1763 1764 /* 1765 * After system resumes, we need to calculate the suspended time and 1766 * compensate it for the OS time. There are 3 sources that could be 1767 * used: Nonstop clocksource during suspend, persistent clock and rtc 1768 * device. 1769 * 1770 * One specific platform may have 1 or 2 or all of them, and the 1771 * preference will be: 1772 * suspend-nonstop clocksource -> persistent clock -> rtc 1773 * The less preferred source will only be tried if there is no better 1774 * usable source. The rtc part is handled separately in rtc core code. 1775 */ 1776 cycle_now = tk_clock_read(&tk->tkr_mono); 1777 nsec = clocksource_stop_suspend_timing(clock, cycle_now); 1778 if (nsec > 0) { 1779 ts_delta = ns_to_timespec64(nsec); 1780 inject_sleeptime = true; 1781 } else if (timespec64_compare(&ts_new, &timekeeping_suspend_time) > 0) { 1782 ts_delta = timespec64_sub(ts_new, timekeeping_suspend_time); 1783 inject_sleeptime = true; 1784 } 1785 1786 if (inject_sleeptime) { 1787 suspend_timing_needed = false; 1788 __timekeeping_inject_sleeptime(tk, &ts_delta); 1789 } 1790 1791 /* Re-base the last cycle value */ 1792 tk->tkr_mono.cycle_last = cycle_now; 1793 tk->tkr_raw.cycle_last = cycle_now; 1794 1795 tk->ntp_error = 0; 1796 timekeeping_suspended = 0; 1797 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET); 1798 write_seqcount_end(&tk_core.seq); 1799 raw_spin_unlock_irqrestore(&timekeeper_lock, flags); 1800 1801 touch_softlockup_watchdog(); 1802 1803 tick_resume(); 1804 hrtimers_resume(); 1805 } 1806 1807 int timekeeping_suspend(void) 1808 { 1809 struct timekeeper *tk = &tk_core.timekeeper; 1810 unsigned long flags; 1811 struct timespec64 delta, delta_delta; 1812 static struct timespec64 old_delta; 1813 struct clocksource *curr_clock; 1814 u64 cycle_now; 1815 1816 read_persistent_clock64(&timekeeping_suspend_time); 1817 1818 /* 1819 * On some systems the persistent_clock can not be detected at 1820 * timekeeping_init by its return value, so if we see a valid 1821 * value returned, update the persistent_clock_exists flag. 1822 */ 1823 if (timekeeping_suspend_time.tv_sec || timekeeping_suspend_time.tv_nsec) 1824 persistent_clock_exists = true; 1825 1826 suspend_timing_needed = true; 1827 1828 raw_spin_lock_irqsave(&timekeeper_lock, flags); 1829 write_seqcount_begin(&tk_core.seq); 1830 timekeeping_forward_now(tk); 1831 timekeeping_suspended = 1; 1832 1833 /* 1834 * Since we've called forward_now, cycle_last stores the value 1835 * just read from the current clocksource. Save this to potentially 1836 * use in suspend timing. 1837 */ 1838 curr_clock = tk->tkr_mono.clock; 1839 cycle_now = tk->tkr_mono.cycle_last; 1840 clocksource_start_suspend_timing(curr_clock, cycle_now); 1841 1842 if (persistent_clock_exists) { 1843 /* 1844 * To avoid drift caused by repeated suspend/resumes, 1845 * which each can add ~1 second drift error, 1846 * try to compensate so the difference in system time 1847 * and persistent_clock time stays close to constant. 1848 */ 1849 delta = timespec64_sub(tk_xtime(tk), timekeeping_suspend_time); 1850 delta_delta = timespec64_sub(delta, old_delta); 1851 if (abs(delta_delta.tv_sec) >= 2) { 1852 /* 1853 * if delta_delta is too large, assume time correction 1854 * has occurred and set old_delta to the current delta. 1855 */ 1856 old_delta = delta; 1857 } else { 1858 /* Otherwise try to adjust old_system to compensate */ 1859 timekeeping_suspend_time = 1860 timespec64_add(timekeeping_suspend_time, delta_delta); 1861 } 1862 } 1863 1864 timekeeping_update(tk, TK_MIRROR); 1865 halt_fast_timekeeper(tk); 1866 write_seqcount_end(&tk_core.seq); 1867 raw_spin_unlock_irqrestore(&timekeeper_lock, flags); 1868 1869 tick_suspend(); 1870 clocksource_suspend(); 1871 clockevents_suspend(); 1872 1873 return 0; 1874 } 1875 1876 /* sysfs resume/suspend bits for timekeeping */ 1877 static struct syscore_ops timekeeping_syscore_ops = { 1878 .resume = timekeeping_resume, 1879 .suspend = timekeeping_suspend, 1880 }; 1881 1882 static int __init timekeeping_init_ops(void) 1883 { 1884 register_syscore_ops(&timekeeping_syscore_ops); 1885 return 0; 1886 } 1887 device_initcall(timekeeping_init_ops); 1888 1889 /* 1890 * Apply a multiplier adjustment to the timekeeper 1891 */ 1892 static __always_inline void timekeeping_apply_adjustment(struct timekeeper *tk, 1893 s64 offset, 1894 s32 mult_adj) 1895 { 1896 s64 interval = tk->cycle_interval; 1897 1898 if (mult_adj == 0) { 1899 return; 1900 } else if (mult_adj == -1) { 1901 interval = -interval; 1902 offset = -offset; 1903 } else if (mult_adj != 1) { 1904 interval *= mult_adj; 1905 offset *= mult_adj; 1906 } 1907 1908 /* 1909 * So the following can be confusing. 1910 * 1911 * To keep things simple, lets assume mult_adj == 1 for now. 1912 * 1913 * When mult_adj != 1, remember that the interval and offset values 1914 * have been appropriately scaled so the math is the same. 1915 * 1916 * The basic idea here is that we're increasing the multiplier 1917 * by one, this causes the xtime_interval to be incremented by 1918 * one cycle_interval. This is because: 1919 * xtime_interval = cycle_interval * mult 1920 * So if mult is being incremented by one: 1921 * xtime_interval = cycle_interval * (mult + 1) 1922 * Its the same as: 1923 * xtime_interval = (cycle_interval * mult) + cycle_interval 1924 * Which can be shortened to: 1925 * xtime_interval += cycle_interval 1926 * 1927 * So offset stores the non-accumulated cycles. Thus the current 1928 * time (in shifted nanoseconds) is: 1929 * now = (offset * adj) + xtime_nsec 1930 * Now, even though we're adjusting the clock frequency, we have 1931 * to keep time consistent. In other words, we can't jump back 1932 * in time, and we also want to avoid jumping forward in time. 1933 * 1934 * So given the same offset value, we need the time to be the same 1935 * both before and after the freq adjustment. 1936 * now = (offset * adj_1) + xtime_nsec_1 1937 * now = (offset * adj_2) + xtime_nsec_2 1938 * So: 1939 * (offset * adj_1) + xtime_nsec_1 = 1940 * (offset * adj_2) + xtime_nsec_2 1941 * And we know: 1942 * adj_2 = adj_1 + 1 1943 * So: 1944 * (offset * adj_1) + xtime_nsec_1 = 1945 * (offset * (adj_1+1)) + xtime_nsec_2 1946 * (offset * adj_1) + xtime_nsec_1 = 1947 * (offset * adj_1) + offset + xtime_nsec_2 1948 * Canceling the sides: 1949 * xtime_nsec_1 = offset + xtime_nsec_2 1950 * Which gives us: 1951 * xtime_nsec_2 = xtime_nsec_1 - offset 1952 * Which simplfies to: 1953 * xtime_nsec -= offset 1954 */ 1955 if ((mult_adj > 0) && (tk->tkr_mono.mult + mult_adj < mult_adj)) { 1956 /* NTP adjustment caused clocksource mult overflow */ 1957 WARN_ON_ONCE(1); 1958 return; 1959 } 1960 1961 tk->tkr_mono.mult += mult_adj; 1962 tk->xtime_interval += interval; 1963 tk->tkr_mono.xtime_nsec -= offset; 1964 } 1965 1966 /* 1967 * Adjust the timekeeper's multiplier to the correct frequency 1968 * and also to reduce the accumulated error value. 1969 */ 1970 static void timekeeping_adjust(struct timekeeper *tk, s64 offset) 1971 { 1972 u32 mult; 1973 1974 /* 1975 * Determine the multiplier from the current NTP tick length. 1976 * Avoid expensive division when the tick length doesn't change. 1977 */ 1978 if (likely(tk->ntp_tick == ntp_tick_length())) { 1979 mult = tk->tkr_mono.mult - tk->ntp_err_mult; 1980 } else { 1981 tk->ntp_tick = ntp_tick_length(); 1982 mult = div64_u64((tk->ntp_tick >> tk->ntp_error_shift) - 1983 tk->xtime_remainder, tk->cycle_interval); 1984 } 1985 1986 /* 1987 * If the clock is behind the NTP time, increase the multiplier by 1 1988 * to catch up with it. If it's ahead and there was a remainder in the 1989 * tick division, the clock will slow down. Otherwise it will stay 1990 * ahead until the tick length changes to a non-divisible value. 1991 */ 1992 tk->ntp_err_mult = tk->ntp_error > 0 ? 1 : 0; 1993 mult += tk->ntp_err_mult; 1994 1995 timekeeping_apply_adjustment(tk, offset, mult - tk->tkr_mono.mult); 1996 1997 if (unlikely(tk->tkr_mono.clock->maxadj && 1998 (abs(tk->tkr_mono.mult - tk->tkr_mono.clock->mult) 1999 > tk->tkr_mono.clock->maxadj))) { 2000 printk_once(KERN_WARNING 2001 "Adjusting %s more than 11%% (%ld vs %ld)\n", 2002 tk->tkr_mono.clock->name, (long)tk->tkr_mono.mult, 2003 (long)tk->tkr_mono.clock->mult + tk->tkr_mono.clock->maxadj); 2004 } 2005 2006 /* 2007 * It may be possible that when we entered this function, xtime_nsec 2008 * was very small. Further, if we're slightly speeding the clocksource 2009 * in the code above, its possible the required corrective factor to 2010 * xtime_nsec could cause it to underflow. 2011 * 2012 * Now, since we have already accumulated the second and the NTP 2013 * subsystem has been notified via second_overflow(), we need to skip 2014 * the next update. 2015 */ 2016 if (unlikely((s64)tk->tkr_mono.xtime_nsec < 0)) { 2017 tk->tkr_mono.xtime_nsec += (u64)NSEC_PER_SEC << 2018 tk->tkr_mono.shift; 2019 tk->xtime_sec--; 2020 tk->skip_second_overflow = 1; 2021 } 2022 } 2023 2024 /* 2025 * accumulate_nsecs_to_secs - Accumulates nsecs into secs 2026 * 2027 * Helper function that accumulates the nsecs greater than a second 2028 * from the xtime_nsec field to the xtime_secs field. 2029 * It also calls into the NTP code to handle leapsecond processing. 2030 */ 2031 static inline unsigned int accumulate_nsecs_to_secs(struct timekeeper *tk) 2032 { 2033 u64 nsecps = (u64)NSEC_PER_SEC << tk->tkr_mono.shift; 2034 unsigned int clock_set = 0; 2035 2036 while (tk->tkr_mono.xtime_nsec >= nsecps) { 2037 int leap; 2038 2039 tk->tkr_mono.xtime_nsec -= nsecps; 2040 tk->xtime_sec++; 2041 2042 /* 2043 * Skip NTP update if this second was accumulated before, 2044 * i.e. xtime_nsec underflowed in timekeeping_adjust() 2045 */ 2046 if (unlikely(tk->skip_second_overflow)) { 2047 tk->skip_second_overflow = 0; 2048 continue; 2049 } 2050 2051 /* Figure out if its a leap sec and apply if needed */ 2052 leap = second_overflow(tk->xtime_sec); 2053 if (unlikely(leap)) { 2054 struct timespec64 ts; 2055 2056 tk->xtime_sec += leap; 2057 2058 ts.tv_sec = leap; 2059 ts.tv_nsec = 0; 2060 tk_set_wall_to_mono(tk, 2061 timespec64_sub(tk->wall_to_monotonic, ts)); 2062 2063 __timekeeping_set_tai_offset(tk, tk->tai_offset - leap); 2064 2065 clock_set = TK_CLOCK_WAS_SET; 2066 } 2067 } 2068 return clock_set; 2069 } 2070 2071 /* 2072 * logarithmic_accumulation - shifted accumulation of cycles 2073 * 2074 * This functions accumulates a shifted interval of cycles into 2075 * a shifted interval nanoseconds. Allows for O(log) accumulation 2076 * loop. 2077 * 2078 * Returns the unconsumed cycles. 2079 */ 2080 static u64 logarithmic_accumulation(struct timekeeper *tk, u64 offset, 2081 u32 shift, unsigned int *clock_set) 2082 { 2083 u64 interval = tk->cycle_interval << shift; 2084 u64 snsec_per_sec; 2085 2086 /* If the offset is smaller than a shifted interval, do nothing */ 2087 if (offset < interval) 2088 return offset; 2089 2090 /* Accumulate one shifted interval */ 2091 offset -= interval; 2092 tk->tkr_mono.cycle_last += interval; 2093 tk->tkr_raw.cycle_last += interval; 2094 2095 tk->tkr_mono.xtime_nsec += tk->xtime_interval << shift; 2096 *clock_set |= accumulate_nsecs_to_secs(tk); 2097 2098 /* Accumulate raw time */ 2099 tk->tkr_raw.xtime_nsec += tk->raw_interval << shift; 2100 snsec_per_sec = (u64)NSEC_PER_SEC << tk->tkr_raw.shift; 2101 while (tk->tkr_raw.xtime_nsec >= snsec_per_sec) { 2102 tk->tkr_raw.xtime_nsec -= snsec_per_sec; 2103 tk->raw_sec++; 2104 } 2105 2106 /* Accumulate error between NTP and clock interval */ 2107 tk->ntp_error += tk->ntp_tick << shift; 2108 tk->ntp_error -= (tk->xtime_interval + tk->xtime_remainder) << 2109 (tk->ntp_error_shift + shift); 2110 2111 return offset; 2112 } 2113 2114 /* 2115 * timekeeping_advance - Updates the timekeeper to the current time and 2116 * current NTP tick length 2117 */ 2118 static void timekeeping_advance(enum timekeeping_adv_mode mode) 2119 { 2120 struct timekeeper *real_tk = &tk_core.timekeeper; 2121 struct timekeeper *tk = &shadow_timekeeper; 2122 u64 offset; 2123 int shift = 0, maxshift; 2124 unsigned int clock_set = 0; 2125 unsigned long flags; 2126 2127 raw_spin_lock_irqsave(&timekeeper_lock, flags); 2128 2129 /* Make sure we're fully resumed: */ 2130 if (unlikely(timekeeping_suspended)) 2131 goto out; 2132 2133 offset = clocksource_delta(tk_clock_read(&tk->tkr_mono), 2134 tk->tkr_mono.cycle_last, tk->tkr_mono.mask); 2135 2136 /* Check if there's really nothing to do */ 2137 if (offset < real_tk->cycle_interval && mode == TK_ADV_TICK) 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 /* Adjust the multiplier to correct NTP error */ 2164 timekeeping_adjust(tk, offset); 2165 2166 /* 2167 * Finally, make sure that after the rounding 2168 * xtime_nsec isn't larger than NSEC_PER_SEC 2169 */ 2170 clock_set |= accumulate_nsecs_to_secs(tk); 2171 2172 write_seqcount_begin(&tk_core.seq); 2173 /* 2174 * Update the real timekeeper. 2175 * 2176 * We could avoid this memcpy by switching pointers, but that 2177 * requires changes to all other timekeeper usage sites as 2178 * well, i.e. move the timekeeper pointer getter into the 2179 * spinlocked/seqcount protected sections. And we trade this 2180 * memcpy under the tk_core.seq against one before we start 2181 * updating. 2182 */ 2183 timekeeping_update(tk, clock_set); 2184 memcpy(real_tk, tk, sizeof(*tk)); 2185 /* The memcpy must come last. Do not put anything here! */ 2186 write_seqcount_end(&tk_core.seq); 2187 out: 2188 raw_spin_unlock_irqrestore(&timekeeper_lock, flags); 2189 if (clock_set) 2190 /* Have to call _delayed version, since in irq context*/ 2191 clock_was_set_delayed(); 2192 } 2193 2194 /** 2195 * update_wall_time - Uses the current clocksource to increment the wall time 2196 * 2197 */ 2198 void update_wall_time(void) 2199 { 2200 timekeeping_advance(TK_ADV_TICK); 2201 } 2202 2203 /** 2204 * getboottime64 - Return the real time of system boot. 2205 * @ts: pointer to the timespec64 to be set 2206 * 2207 * Returns the wall-time of boot in a timespec64. 2208 * 2209 * This is based on the wall_to_monotonic offset and the total suspend 2210 * time. Calls to settimeofday will affect the value returned (which 2211 * basically means that however wrong your real time clock is at boot time, 2212 * you get the right time here). 2213 */ 2214 void getboottime64(struct timespec64 *ts) 2215 { 2216 struct timekeeper *tk = &tk_core.timekeeper; 2217 ktime_t t = ktime_sub(tk->offs_real, tk->offs_boot); 2218 2219 *ts = ktime_to_timespec64(t); 2220 } 2221 EXPORT_SYMBOL_GPL(getboottime64); 2222 2223 void ktime_get_coarse_real_ts64(struct timespec64 *ts) 2224 { 2225 struct timekeeper *tk = &tk_core.timekeeper; 2226 unsigned int seq; 2227 2228 do { 2229 seq = read_seqcount_begin(&tk_core.seq); 2230 2231 *ts = tk_xtime(tk); 2232 } while (read_seqcount_retry(&tk_core.seq, seq)); 2233 } 2234 EXPORT_SYMBOL(ktime_get_coarse_real_ts64); 2235 2236 void ktime_get_coarse_ts64(struct timespec64 *ts) 2237 { 2238 struct timekeeper *tk = &tk_core.timekeeper; 2239 struct timespec64 now, mono; 2240 unsigned int seq; 2241 2242 do { 2243 seq = read_seqcount_begin(&tk_core.seq); 2244 2245 now = tk_xtime(tk); 2246 mono = tk->wall_to_monotonic; 2247 } while (read_seqcount_retry(&tk_core.seq, seq)); 2248 2249 set_normalized_timespec64(ts, now.tv_sec + mono.tv_sec, 2250 now.tv_nsec + mono.tv_nsec); 2251 } 2252 EXPORT_SYMBOL(ktime_get_coarse_ts64); 2253 2254 /* 2255 * Must hold jiffies_lock 2256 */ 2257 void do_timer(unsigned long ticks) 2258 { 2259 jiffies_64 += ticks; 2260 calc_global_load(); 2261 } 2262 2263 /** 2264 * ktime_get_update_offsets_now - hrtimer helper 2265 * @cwsseq: pointer to check and store the clock was set sequence number 2266 * @offs_real: pointer to storage for monotonic -> realtime offset 2267 * @offs_boot: pointer to storage for monotonic -> boottime offset 2268 * @offs_tai: pointer to storage for monotonic -> clock tai offset 2269 * 2270 * Returns current monotonic time and updates the offsets if the 2271 * sequence number in @cwsseq and timekeeper.clock_was_set_seq are 2272 * different. 2273 * 2274 * Called from hrtimer_interrupt() or retrigger_next_event() 2275 */ 2276 ktime_t ktime_get_update_offsets_now(unsigned int *cwsseq, ktime_t *offs_real, 2277 ktime_t *offs_boot, ktime_t *offs_tai) 2278 { 2279 struct timekeeper *tk = &tk_core.timekeeper; 2280 unsigned int seq; 2281 ktime_t base; 2282 u64 nsecs; 2283 2284 do { 2285 seq = read_seqcount_begin(&tk_core.seq); 2286 2287 base = tk->tkr_mono.base; 2288 nsecs = timekeeping_get_ns(&tk->tkr_mono); 2289 base = ktime_add_ns(base, nsecs); 2290 2291 if (*cwsseq != tk->clock_was_set_seq) { 2292 *cwsseq = tk->clock_was_set_seq; 2293 *offs_real = tk->offs_real; 2294 *offs_boot = tk->offs_boot; 2295 *offs_tai = tk->offs_tai; 2296 } 2297 2298 /* Handle leapsecond insertion adjustments */ 2299 if (unlikely(base >= tk->next_leap_ktime)) 2300 *offs_real = ktime_sub(tk->offs_real, ktime_set(1, 0)); 2301 2302 } while (read_seqcount_retry(&tk_core.seq, seq)); 2303 2304 return base; 2305 } 2306 2307 /* 2308 * timekeeping_validate_timex - Ensures the timex is ok for use in do_adjtimex 2309 */ 2310 static int timekeeping_validate_timex(const struct __kernel_timex *txc) 2311 { 2312 if (txc->modes & ADJ_ADJTIME) { 2313 /* singleshot must not be used with any other mode bits */ 2314 if (!(txc->modes & ADJ_OFFSET_SINGLESHOT)) 2315 return -EINVAL; 2316 if (!(txc->modes & ADJ_OFFSET_READONLY) && 2317 !capable(CAP_SYS_TIME)) 2318 return -EPERM; 2319 } else { 2320 /* In order to modify anything, you gotta be super-user! */ 2321 if (txc->modes && !capable(CAP_SYS_TIME)) 2322 return -EPERM; 2323 /* 2324 * if the quartz is off by more than 10% then 2325 * something is VERY wrong! 2326 */ 2327 if (txc->modes & ADJ_TICK && 2328 (txc->tick < 900000/USER_HZ || 2329 txc->tick > 1100000/USER_HZ)) 2330 return -EINVAL; 2331 } 2332 2333 if (txc->modes & ADJ_SETOFFSET) { 2334 /* In order to inject time, you gotta be super-user! */ 2335 if (!capable(CAP_SYS_TIME)) 2336 return -EPERM; 2337 2338 /* 2339 * Validate if a timespec/timeval used to inject a time 2340 * offset is valid. Offsets can be postive or negative, so 2341 * we don't check tv_sec. The value of the timeval/timespec 2342 * is the sum of its fields,but *NOTE*: 2343 * The field tv_usec/tv_nsec must always be non-negative and 2344 * we can't have more nanoseconds/microseconds than a second. 2345 */ 2346 if (txc->time.tv_usec < 0) 2347 return -EINVAL; 2348 2349 if (txc->modes & ADJ_NANO) { 2350 if (txc->time.tv_usec >= NSEC_PER_SEC) 2351 return -EINVAL; 2352 } else { 2353 if (txc->time.tv_usec >= USEC_PER_SEC) 2354 return -EINVAL; 2355 } 2356 } 2357 2358 /* 2359 * Check for potential multiplication overflows that can 2360 * only happen on 64-bit systems: 2361 */ 2362 if ((txc->modes & ADJ_FREQUENCY) && (BITS_PER_LONG == 64)) { 2363 if (LLONG_MIN / PPM_SCALE > txc->freq) 2364 return -EINVAL; 2365 if (LLONG_MAX / PPM_SCALE < txc->freq) 2366 return -EINVAL; 2367 } 2368 2369 return 0; 2370 } 2371 2372 2373 /** 2374 * do_adjtimex() - Accessor function to NTP __do_adjtimex function 2375 */ 2376 int do_adjtimex(struct __kernel_timex *txc) 2377 { 2378 struct timekeeper *tk = &tk_core.timekeeper; 2379 struct audit_ntp_data ad; 2380 unsigned long flags; 2381 struct timespec64 ts; 2382 s32 orig_tai, tai; 2383 int ret; 2384 2385 /* Validate the data before disabling interrupts */ 2386 ret = timekeeping_validate_timex(txc); 2387 if (ret) 2388 return ret; 2389 2390 if (txc->modes & ADJ_SETOFFSET) { 2391 struct timespec64 delta; 2392 delta.tv_sec = txc->time.tv_sec; 2393 delta.tv_nsec = txc->time.tv_usec; 2394 if (!(txc->modes & ADJ_NANO)) 2395 delta.tv_nsec *= 1000; 2396 ret = timekeeping_inject_offset(&delta); 2397 if (ret) 2398 return ret; 2399 2400 audit_tk_injoffset(delta); 2401 } 2402 2403 audit_ntp_init(&ad); 2404 2405 ktime_get_real_ts64(&ts); 2406 2407 raw_spin_lock_irqsave(&timekeeper_lock, flags); 2408 write_seqcount_begin(&tk_core.seq); 2409 2410 orig_tai = tai = tk->tai_offset; 2411 ret = __do_adjtimex(txc, &ts, &tai, &ad); 2412 2413 if (tai != orig_tai) { 2414 __timekeeping_set_tai_offset(tk, tai); 2415 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET); 2416 } 2417 tk_update_leap_state(tk); 2418 2419 write_seqcount_end(&tk_core.seq); 2420 raw_spin_unlock_irqrestore(&timekeeper_lock, flags); 2421 2422 audit_ntp_log(&ad); 2423 2424 /* Update the multiplier immediately if frequency was set directly */ 2425 if (txc->modes & (ADJ_FREQUENCY | ADJ_TICK)) 2426 timekeeping_advance(TK_ADV_FREQ); 2427 2428 if (tai != orig_tai) 2429 clock_was_set(); 2430 2431 ntp_notify_cmos_timer(); 2432 2433 return ret; 2434 } 2435 2436 #ifdef CONFIG_NTP_PPS 2437 /** 2438 * hardpps() - Accessor function to NTP __hardpps function 2439 */ 2440 void hardpps(const struct timespec64 *phase_ts, const struct timespec64 *raw_ts) 2441 { 2442 unsigned long flags; 2443 2444 raw_spin_lock_irqsave(&timekeeper_lock, flags); 2445 write_seqcount_begin(&tk_core.seq); 2446 2447 __hardpps(phase_ts, raw_ts); 2448 2449 write_seqcount_end(&tk_core.seq); 2450 raw_spin_unlock_irqrestore(&timekeeper_lock, flags); 2451 } 2452 EXPORT_SYMBOL(hardpps); 2453 #endif /* CONFIG_NTP_PPS */ 2454