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 fast_tk_get_delta_ns(struct tk_read_base *tkr) 433 { 434 u64 delta, cycles = tk_clock_read(tkr); 435 436 delta = clocksource_delta(cycles, tkr->cycle_last, tkr->mask); 437 return timekeeping_delta_to_ns(tkr, delta); 438 } 439 440 static __always_inline u64 __ktime_get_fast_ns(struct tk_fast *tkf) 441 { 442 struct tk_read_base *tkr; 443 unsigned int seq; 444 u64 now; 445 446 do { 447 seq = raw_read_seqcount_latch(&tkf->seq); 448 tkr = tkf->base + (seq & 0x01); 449 now = ktime_to_ns(tkr->base); 450 now += fast_tk_get_delta_ns(tkr); 451 } while (read_seqcount_latch_retry(&tkf->seq, seq)); 452 453 return now; 454 } 455 456 /** 457 * ktime_get_mono_fast_ns - Fast NMI safe access to clock monotonic 458 * 459 * This timestamp is not guaranteed to be monotonic across an update. 460 * The timestamp is calculated by: 461 * 462 * now = base_mono + clock_delta * slope 463 * 464 * So if the update lowers the slope, readers who are forced to the 465 * not yet updated second array are still using the old steeper slope. 466 * 467 * tmono 468 * ^ 469 * | o n 470 * | o n 471 * | u 472 * | o 473 * |o 474 * |12345678---> reader order 475 * 476 * o = old slope 477 * u = update 478 * n = new slope 479 * 480 * So reader 6 will observe time going backwards versus reader 5. 481 * 482 * While other CPUs are likely to be able to observe that, the only way 483 * for a CPU local observation is when an NMI hits in the middle of 484 * the update. Timestamps taken from that NMI context might be ahead 485 * of the following timestamps. Callers need to be aware of that and 486 * deal with it. 487 */ 488 u64 notrace ktime_get_mono_fast_ns(void) 489 { 490 return __ktime_get_fast_ns(&tk_fast_mono); 491 } 492 EXPORT_SYMBOL_GPL(ktime_get_mono_fast_ns); 493 494 /** 495 * ktime_get_raw_fast_ns - Fast NMI safe access to clock monotonic raw 496 * 497 * Contrary to ktime_get_mono_fast_ns() this is always correct because the 498 * conversion factor is not affected by NTP/PTP correction. 499 */ 500 u64 notrace ktime_get_raw_fast_ns(void) 501 { 502 return __ktime_get_fast_ns(&tk_fast_raw); 503 } 504 EXPORT_SYMBOL_GPL(ktime_get_raw_fast_ns); 505 506 /** 507 * ktime_get_boot_fast_ns - NMI safe and fast access to boot clock. 508 * 509 * To keep it NMI safe since we're accessing from tracing, we're not using a 510 * separate timekeeper with updates to monotonic clock and boot offset 511 * protected with seqcounts. This has the following minor side effects: 512 * 513 * (1) Its possible that a timestamp be taken after the boot offset is updated 514 * but before the timekeeper is updated. If this happens, the new boot offset 515 * is added to the old timekeeping making the clock appear to update slightly 516 * earlier: 517 * CPU 0 CPU 1 518 * timekeeping_inject_sleeptime64() 519 * __timekeeping_inject_sleeptime(tk, delta); 520 * timestamp(); 521 * timekeeping_update(tk, TK_CLEAR_NTP...); 522 * 523 * (2) On 32-bit systems, the 64-bit boot offset (tk->offs_boot) may be 524 * partially updated. Since the tk->offs_boot update is a rare event, this 525 * should be a rare occurrence which postprocessing should be able to handle. 526 * 527 * The caveats vs. timestamp ordering as documented for ktime_get_fast_ns() 528 * apply as well. 529 */ 530 u64 notrace ktime_get_boot_fast_ns(void) 531 { 532 struct timekeeper *tk = &tk_core.timekeeper; 533 534 return (ktime_get_mono_fast_ns() + ktime_to_ns(data_race(tk->offs_boot))); 535 } 536 EXPORT_SYMBOL_GPL(ktime_get_boot_fast_ns); 537 538 /** 539 * ktime_get_tai_fast_ns - NMI safe and fast access to tai clock. 540 * 541 * The same limitations as described for ktime_get_boot_fast_ns() apply. The 542 * mono time and the TAI offset are not read atomically which may yield wrong 543 * readouts. However, an update of the TAI offset is an rare event e.g., caused 544 * by settime or adjtimex with an offset. The user of this function has to deal 545 * with the possibility of wrong timestamps in post processing. 546 */ 547 u64 notrace ktime_get_tai_fast_ns(void) 548 { 549 struct timekeeper *tk = &tk_core.timekeeper; 550 551 return (ktime_get_mono_fast_ns() + ktime_to_ns(data_race(tk->offs_tai))); 552 } 553 EXPORT_SYMBOL_GPL(ktime_get_tai_fast_ns); 554 555 static __always_inline u64 __ktime_get_real_fast(struct tk_fast *tkf, u64 *mono) 556 { 557 struct tk_read_base *tkr; 558 u64 basem, baser, delta; 559 unsigned int seq; 560 561 do { 562 seq = raw_read_seqcount_latch(&tkf->seq); 563 tkr = tkf->base + (seq & 0x01); 564 basem = ktime_to_ns(tkr->base); 565 baser = ktime_to_ns(tkr->base_real); 566 delta = fast_tk_get_delta_ns(tkr); 567 } while (read_seqcount_latch_retry(&tkf->seq, seq)); 568 569 if (mono) 570 *mono = basem + delta; 571 return baser + delta; 572 } 573 574 /** 575 * ktime_get_real_fast_ns: - NMI safe and fast access to clock realtime. 576 * 577 * See ktime_get_fast_ns() for documentation of the time stamp ordering. 578 */ 579 u64 ktime_get_real_fast_ns(void) 580 { 581 return __ktime_get_real_fast(&tk_fast_mono, NULL); 582 } 583 EXPORT_SYMBOL_GPL(ktime_get_real_fast_ns); 584 585 /** 586 * ktime_get_fast_timestamps: - NMI safe timestamps 587 * @snapshot: Pointer to timestamp storage 588 * 589 * Stores clock monotonic, boottime and realtime timestamps. 590 * 591 * Boot time is a racy access on 32bit systems if the sleep time injection 592 * happens late during resume and not in timekeeping_resume(). That could 593 * be avoided by expanding struct tk_read_base with boot offset for 32bit 594 * and adding more overhead to the update. As this is a hard to observe 595 * once per resume event which can be filtered with reasonable effort using 596 * the accurate mono/real timestamps, it's probably not worth the trouble. 597 * 598 * Aside of that it might be possible on 32 and 64 bit to observe the 599 * following when the sleep time injection happens late: 600 * 601 * CPU 0 CPU 1 602 * timekeeping_resume() 603 * ktime_get_fast_timestamps() 604 * mono, real = __ktime_get_real_fast() 605 * inject_sleep_time() 606 * update boot offset 607 * boot = mono + bootoffset; 608 * 609 * That means that boot time already has the sleep time adjustment, but 610 * real time does not. On the next readout both are in sync again. 611 * 612 * Preventing this for 64bit is not really feasible without destroying the 613 * careful cache layout of the timekeeper because the sequence count and 614 * struct tk_read_base would then need two cache lines instead of one. 615 * 616 * Access to the time keeper clock source is disabled across the innermost 617 * steps of suspend/resume. The accessors still work, but the timestamps 618 * are frozen until time keeping is resumed which happens very early. 619 * 620 * For regular suspend/resume there is no observable difference vs. sched 621 * clock, but it might affect some of the nasty low level debug printks. 622 * 623 * OTOH, access to sched clock is not guaranteed across suspend/resume on 624 * all systems either so it depends on the hardware in use. 625 * 626 * If that turns out to be a real problem then this could be mitigated by 627 * using sched clock in a similar way as during early boot. But it's not as 628 * trivial as on early boot because it needs some careful protection 629 * against the clock monotonic timestamp jumping backwards on resume. 630 */ 631 void ktime_get_fast_timestamps(struct ktime_timestamps *snapshot) 632 { 633 struct timekeeper *tk = &tk_core.timekeeper; 634 635 snapshot->real = __ktime_get_real_fast(&tk_fast_mono, &snapshot->mono); 636 snapshot->boot = snapshot->mono + ktime_to_ns(data_race(tk->offs_boot)); 637 } 638 639 /** 640 * halt_fast_timekeeper - Prevent fast timekeeper from accessing clocksource. 641 * @tk: Timekeeper to snapshot. 642 * 643 * It generally is unsafe to access the clocksource after timekeeping has been 644 * suspended, so take a snapshot of the readout base of @tk and use it as the 645 * fast timekeeper's readout base while suspended. It will return the same 646 * number of cycles every time until timekeeping is resumed at which time the 647 * proper readout base for the fast timekeeper will be restored automatically. 648 */ 649 static void halt_fast_timekeeper(const struct timekeeper *tk) 650 { 651 static struct tk_read_base tkr_dummy; 652 const struct tk_read_base *tkr = &tk->tkr_mono; 653 654 memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy)); 655 cycles_at_suspend = tk_clock_read(tkr); 656 tkr_dummy.clock = &dummy_clock; 657 tkr_dummy.base_real = tkr->base + tk->offs_real; 658 update_fast_timekeeper(&tkr_dummy, &tk_fast_mono); 659 660 tkr = &tk->tkr_raw; 661 memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy)); 662 tkr_dummy.clock = &dummy_clock; 663 update_fast_timekeeper(&tkr_dummy, &tk_fast_raw); 664 } 665 666 static RAW_NOTIFIER_HEAD(pvclock_gtod_chain); 667 668 static void update_pvclock_gtod(struct timekeeper *tk, bool was_set) 669 { 670 raw_notifier_call_chain(&pvclock_gtod_chain, was_set, tk); 671 } 672 673 /** 674 * pvclock_gtod_register_notifier - register a pvclock timedata update listener 675 * @nb: Pointer to the notifier block to register 676 */ 677 int pvclock_gtod_register_notifier(struct notifier_block *nb) 678 { 679 struct timekeeper *tk = &tk_core.timekeeper; 680 unsigned long flags; 681 int ret; 682 683 raw_spin_lock_irqsave(&timekeeper_lock, flags); 684 ret = raw_notifier_chain_register(&pvclock_gtod_chain, nb); 685 update_pvclock_gtod(tk, true); 686 raw_spin_unlock_irqrestore(&timekeeper_lock, flags); 687 688 return ret; 689 } 690 EXPORT_SYMBOL_GPL(pvclock_gtod_register_notifier); 691 692 /** 693 * pvclock_gtod_unregister_notifier - unregister a pvclock 694 * timedata update listener 695 * @nb: Pointer to the notifier block to unregister 696 */ 697 int pvclock_gtod_unregister_notifier(struct notifier_block *nb) 698 { 699 unsigned long flags; 700 int ret; 701 702 raw_spin_lock_irqsave(&timekeeper_lock, flags); 703 ret = raw_notifier_chain_unregister(&pvclock_gtod_chain, nb); 704 raw_spin_unlock_irqrestore(&timekeeper_lock, flags); 705 706 return ret; 707 } 708 EXPORT_SYMBOL_GPL(pvclock_gtod_unregister_notifier); 709 710 /* 711 * tk_update_leap_state - helper to update the next_leap_ktime 712 */ 713 static inline void tk_update_leap_state(struct timekeeper *tk) 714 { 715 tk->next_leap_ktime = ntp_get_next_leap(); 716 if (tk->next_leap_ktime != KTIME_MAX) 717 /* Convert to monotonic time */ 718 tk->next_leap_ktime = ktime_sub(tk->next_leap_ktime, tk->offs_real); 719 } 720 721 /* 722 * Update the ktime_t based scalar nsec members of the timekeeper 723 */ 724 static inline void tk_update_ktime_data(struct timekeeper *tk) 725 { 726 u64 seconds; 727 u32 nsec; 728 729 /* 730 * The xtime based monotonic readout is: 731 * nsec = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec + now(); 732 * The ktime based monotonic readout is: 733 * nsec = base_mono + now(); 734 * ==> base_mono = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec 735 */ 736 seconds = (u64)(tk->xtime_sec + tk->wall_to_monotonic.tv_sec); 737 nsec = (u32) tk->wall_to_monotonic.tv_nsec; 738 tk->tkr_mono.base = ns_to_ktime(seconds * NSEC_PER_SEC + nsec); 739 740 /* 741 * The sum of the nanoseconds portions of xtime and 742 * wall_to_monotonic can be greater/equal one second. Take 743 * this into account before updating tk->ktime_sec. 744 */ 745 nsec += (u32)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift); 746 if (nsec >= NSEC_PER_SEC) 747 seconds++; 748 tk->ktime_sec = seconds; 749 750 /* Update the monotonic raw base */ 751 tk->tkr_raw.base = ns_to_ktime(tk->raw_sec * NSEC_PER_SEC); 752 } 753 754 /* must hold timekeeper_lock */ 755 static void timekeeping_update(struct timekeeper *tk, unsigned int action) 756 { 757 if (action & TK_CLEAR_NTP) { 758 tk->ntp_error = 0; 759 ntp_clear(); 760 } 761 762 tk_update_leap_state(tk); 763 tk_update_ktime_data(tk); 764 765 update_vsyscall(tk); 766 update_pvclock_gtod(tk, action & TK_CLOCK_WAS_SET); 767 768 tk->tkr_mono.base_real = tk->tkr_mono.base + tk->offs_real; 769 update_fast_timekeeper(&tk->tkr_mono, &tk_fast_mono); 770 update_fast_timekeeper(&tk->tkr_raw, &tk_fast_raw); 771 772 if (action & TK_CLOCK_WAS_SET) 773 tk->clock_was_set_seq++; 774 /* 775 * The mirroring of the data to the shadow-timekeeper needs 776 * to happen last here to ensure we don't over-write the 777 * timekeeper structure on the next update with stale data 778 */ 779 if (action & TK_MIRROR) 780 memcpy(&shadow_timekeeper, &tk_core.timekeeper, 781 sizeof(tk_core.timekeeper)); 782 } 783 784 /** 785 * timekeeping_forward_now - update clock to the current time 786 * @tk: Pointer to the timekeeper to update 787 * 788 * Forward the current clock to update its state since the last call to 789 * update_wall_time(). This is useful before significant clock changes, 790 * as it avoids having to deal with this time offset explicitly. 791 */ 792 static void timekeeping_forward_now(struct timekeeper *tk) 793 { 794 u64 cycle_now, delta; 795 796 cycle_now = tk_clock_read(&tk->tkr_mono); 797 delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last, tk->tkr_mono.mask); 798 tk->tkr_mono.cycle_last = cycle_now; 799 tk->tkr_raw.cycle_last = cycle_now; 800 801 tk->tkr_mono.xtime_nsec += delta * tk->tkr_mono.mult; 802 tk->tkr_raw.xtime_nsec += delta * tk->tkr_raw.mult; 803 804 tk_normalize_xtime(tk); 805 } 806 807 /** 808 * ktime_get_real_ts64 - Returns the time of day in a timespec64. 809 * @ts: pointer to the timespec to be set 810 * 811 * Returns the time of day in a timespec64 (WARN if suspended). 812 */ 813 void ktime_get_real_ts64(struct timespec64 *ts) 814 { 815 struct timekeeper *tk = &tk_core.timekeeper; 816 unsigned int seq; 817 u64 nsecs; 818 819 WARN_ON(timekeeping_suspended); 820 821 do { 822 seq = read_seqcount_begin(&tk_core.seq); 823 824 ts->tv_sec = tk->xtime_sec; 825 nsecs = timekeeping_get_ns(&tk->tkr_mono); 826 827 } while (read_seqcount_retry(&tk_core.seq, seq)); 828 829 ts->tv_nsec = 0; 830 timespec64_add_ns(ts, nsecs); 831 } 832 EXPORT_SYMBOL(ktime_get_real_ts64); 833 834 ktime_t ktime_get(void) 835 { 836 struct timekeeper *tk = &tk_core.timekeeper; 837 unsigned int seq; 838 ktime_t base; 839 u64 nsecs; 840 841 WARN_ON(timekeeping_suspended); 842 843 do { 844 seq = read_seqcount_begin(&tk_core.seq); 845 base = tk->tkr_mono.base; 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 EXPORT_SYMBOL_GPL(ktime_get); 853 854 u32 ktime_get_resolution_ns(void) 855 { 856 struct timekeeper *tk = &tk_core.timekeeper; 857 unsigned int seq; 858 u32 nsecs; 859 860 WARN_ON(timekeeping_suspended); 861 862 do { 863 seq = read_seqcount_begin(&tk_core.seq); 864 nsecs = tk->tkr_mono.mult >> tk->tkr_mono.shift; 865 } while (read_seqcount_retry(&tk_core.seq, seq)); 866 867 return nsecs; 868 } 869 EXPORT_SYMBOL_GPL(ktime_get_resolution_ns); 870 871 static ktime_t *offsets[TK_OFFS_MAX] = { 872 [TK_OFFS_REAL] = &tk_core.timekeeper.offs_real, 873 [TK_OFFS_BOOT] = &tk_core.timekeeper.offs_boot, 874 [TK_OFFS_TAI] = &tk_core.timekeeper.offs_tai, 875 }; 876 877 ktime_t ktime_get_with_offset(enum tk_offsets offs) 878 { 879 struct timekeeper *tk = &tk_core.timekeeper; 880 unsigned int seq; 881 ktime_t base, *offset = offsets[offs]; 882 u64 nsecs; 883 884 WARN_ON(timekeeping_suspended); 885 886 do { 887 seq = read_seqcount_begin(&tk_core.seq); 888 base = ktime_add(tk->tkr_mono.base, *offset); 889 nsecs = timekeeping_get_ns(&tk->tkr_mono); 890 891 } while (read_seqcount_retry(&tk_core.seq, seq)); 892 893 return ktime_add_ns(base, nsecs); 894 895 } 896 EXPORT_SYMBOL_GPL(ktime_get_with_offset); 897 898 ktime_t ktime_get_coarse_with_offset(enum tk_offsets offs) 899 { 900 struct timekeeper *tk = &tk_core.timekeeper; 901 unsigned int seq; 902 ktime_t base, *offset = offsets[offs]; 903 u64 nsecs; 904 905 WARN_ON(timekeeping_suspended); 906 907 do { 908 seq = read_seqcount_begin(&tk_core.seq); 909 base = ktime_add(tk->tkr_mono.base, *offset); 910 nsecs = tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift; 911 912 } while (read_seqcount_retry(&tk_core.seq, seq)); 913 914 return ktime_add_ns(base, nsecs); 915 } 916 EXPORT_SYMBOL_GPL(ktime_get_coarse_with_offset); 917 918 /** 919 * ktime_mono_to_any() - convert monotonic time to any other time 920 * @tmono: time to convert. 921 * @offs: which offset to use 922 */ 923 ktime_t ktime_mono_to_any(ktime_t tmono, enum tk_offsets offs) 924 { 925 ktime_t *offset = offsets[offs]; 926 unsigned int seq; 927 ktime_t tconv; 928 929 do { 930 seq = read_seqcount_begin(&tk_core.seq); 931 tconv = ktime_add(tmono, *offset); 932 } while (read_seqcount_retry(&tk_core.seq, seq)); 933 934 return tconv; 935 } 936 EXPORT_SYMBOL_GPL(ktime_mono_to_any); 937 938 /** 939 * ktime_get_raw - Returns the raw monotonic time in ktime_t format 940 */ 941 ktime_t ktime_get_raw(void) 942 { 943 struct timekeeper *tk = &tk_core.timekeeper; 944 unsigned int seq; 945 ktime_t base; 946 u64 nsecs; 947 948 do { 949 seq = read_seqcount_begin(&tk_core.seq); 950 base = tk->tkr_raw.base; 951 nsecs = timekeeping_get_ns(&tk->tkr_raw); 952 953 } while (read_seqcount_retry(&tk_core.seq, seq)); 954 955 return ktime_add_ns(base, nsecs); 956 } 957 EXPORT_SYMBOL_GPL(ktime_get_raw); 958 959 /** 960 * ktime_get_ts64 - get the monotonic clock in timespec64 format 961 * @ts: pointer to timespec variable 962 * 963 * The function calculates the monotonic clock from the realtime 964 * clock and the wall_to_monotonic offset and stores the result 965 * in normalized timespec64 format in the variable pointed to by @ts. 966 */ 967 void ktime_get_ts64(struct timespec64 *ts) 968 { 969 struct timekeeper *tk = &tk_core.timekeeper; 970 struct timespec64 tomono; 971 unsigned int seq; 972 u64 nsec; 973 974 WARN_ON(timekeeping_suspended); 975 976 do { 977 seq = read_seqcount_begin(&tk_core.seq); 978 ts->tv_sec = tk->xtime_sec; 979 nsec = timekeeping_get_ns(&tk->tkr_mono); 980 tomono = tk->wall_to_monotonic; 981 982 } while (read_seqcount_retry(&tk_core.seq, seq)); 983 984 ts->tv_sec += tomono.tv_sec; 985 ts->tv_nsec = 0; 986 timespec64_add_ns(ts, nsec + tomono.tv_nsec); 987 } 988 EXPORT_SYMBOL_GPL(ktime_get_ts64); 989 990 /** 991 * ktime_get_seconds - Get the seconds portion of CLOCK_MONOTONIC 992 * 993 * Returns the seconds portion of CLOCK_MONOTONIC with a single non 994 * serialized read. tk->ktime_sec is of type 'unsigned long' so this 995 * works on both 32 and 64 bit systems. On 32 bit systems the readout 996 * covers ~136 years of uptime which should be enough to prevent 997 * premature wrap arounds. 998 */ 999 time64_t ktime_get_seconds(void) 1000 { 1001 struct timekeeper *tk = &tk_core.timekeeper; 1002 1003 WARN_ON(timekeeping_suspended); 1004 return tk->ktime_sec; 1005 } 1006 EXPORT_SYMBOL_GPL(ktime_get_seconds); 1007 1008 /** 1009 * ktime_get_real_seconds - Get the seconds portion of CLOCK_REALTIME 1010 * 1011 * Returns the wall clock seconds since 1970. 1012 * 1013 * For 64bit systems the fast access to tk->xtime_sec is preserved. On 1014 * 32bit systems the access must be protected with the sequence 1015 * counter to provide "atomic" access to the 64bit tk->xtime_sec 1016 * value. 1017 */ 1018 time64_t ktime_get_real_seconds(void) 1019 { 1020 struct timekeeper *tk = &tk_core.timekeeper; 1021 time64_t seconds; 1022 unsigned int seq; 1023 1024 if (IS_ENABLED(CONFIG_64BIT)) 1025 return tk->xtime_sec; 1026 1027 do { 1028 seq = read_seqcount_begin(&tk_core.seq); 1029 seconds = tk->xtime_sec; 1030 1031 } while (read_seqcount_retry(&tk_core.seq, seq)); 1032 1033 return seconds; 1034 } 1035 EXPORT_SYMBOL_GPL(ktime_get_real_seconds); 1036 1037 /** 1038 * __ktime_get_real_seconds - The same as ktime_get_real_seconds 1039 * but without the sequence counter protect. This internal function 1040 * is called just when timekeeping lock is already held. 1041 */ 1042 noinstr time64_t __ktime_get_real_seconds(void) 1043 { 1044 struct timekeeper *tk = &tk_core.timekeeper; 1045 1046 return tk->xtime_sec; 1047 } 1048 1049 /** 1050 * ktime_get_snapshot - snapshots the realtime/monotonic raw clocks with counter 1051 * @systime_snapshot: pointer to struct receiving the system time snapshot 1052 */ 1053 void ktime_get_snapshot(struct system_time_snapshot *systime_snapshot) 1054 { 1055 struct timekeeper *tk = &tk_core.timekeeper; 1056 unsigned int seq; 1057 ktime_t base_raw; 1058 ktime_t base_real; 1059 u64 nsec_raw; 1060 u64 nsec_real; 1061 u64 now; 1062 1063 WARN_ON_ONCE(timekeeping_suspended); 1064 1065 do { 1066 seq = read_seqcount_begin(&tk_core.seq); 1067 now = tk_clock_read(&tk->tkr_mono); 1068 systime_snapshot->cs_id = tk->tkr_mono.clock->id; 1069 systime_snapshot->cs_was_changed_seq = tk->cs_was_changed_seq; 1070 systime_snapshot->clock_was_set_seq = tk->clock_was_set_seq; 1071 base_real = ktime_add(tk->tkr_mono.base, 1072 tk_core.timekeeper.offs_real); 1073 base_raw = tk->tkr_raw.base; 1074 nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono, now); 1075 nsec_raw = timekeeping_cycles_to_ns(&tk->tkr_raw, now); 1076 } while (read_seqcount_retry(&tk_core.seq, seq)); 1077 1078 systime_snapshot->cycles = now; 1079 systime_snapshot->real = ktime_add_ns(base_real, nsec_real); 1080 systime_snapshot->raw = ktime_add_ns(base_raw, nsec_raw); 1081 } 1082 EXPORT_SYMBOL_GPL(ktime_get_snapshot); 1083 1084 /* Scale base by mult/div checking for overflow */ 1085 static int scale64_check_overflow(u64 mult, u64 div, u64 *base) 1086 { 1087 u64 tmp, rem; 1088 1089 tmp = div64_u64_rem(*base, div, &rem); 1090 1091 if (((int)sizeof(u64)*8 - fls64(mult) < fls64(tmp)) || 1092 ((int)sizeof(u64)*8 - fls64(mult) < fls64(rem))) 1093 return -EOVERFLOW; 1094 tmp *= mult; 1095 1096 rem = div64_u64(rem * mult, div); 1097 *base = tmp + rem; 1098 return 0; 1099 } 1100 1101 /** 1102 * adjust_historical_crosststamp - adjust crosstimestamp previous to current interval 1103 * @history: Snapshot representing start of history 1104 * @partial_history_cycles: Cycle offset into history (fractional part) 1105 * @total_history_cycles: Total history length in cycles 1106 * @discontinuity: True indicates clock was set on history period 1107 * @ts: Cross timestamp that should be adjusted using 1108 * partial/total ratio 1109 * 1110 * Helper function used by get_device_system_crosststamp() to correct the 1111 * crosstimestamp corresponding to the start of the current interval to the 1112 * system counter value (timestamp point) provided by the driver. The 1113 * total_history_* quantities are the total history starting at the provided 1114 * reference point and ending at the start of the current interval. The cycle 1115 * count between the driver timestamp point and the start of the current 1116 * interval is partial_history_cycles. 1117 */ 1118 static int adjust_historical_crosststamp(struct system_time_snapshot *history, 1119 u64 partial_history_cycles, 1120 u64 total_history_cycles, 1121 bool discontinuity, 1122 struct system_device_crosststamp *ts) 1123 { 1124 struct timekeeper *tk = &tk_core.timekeeper; 1125 u64 corr_raw, corr_real; 1126 bool interp_forward; 1127 int ret; 1128 1129 if (total_history_cycles == 0 || partial_history_cycles == 0) 1130 return 0; 1131 1132 /* Interpolate shortest distance from beginning or end of history */ 1133 interp_forward = partial_history_cycles > total_history_cycles / 2; 1134 partial_history_cycles = interp_forward ? 1135 total_history_cycles - partial_history_cycles : 1136 partial_history_cycles; 1137 1138 /* 1139 * Scale the monotonic raw time delta by: 1140 * partial_history_cycles / total_history_cycles 1141 */ 1142 corr_raw = (u64)ktime_to_ns( 1143 ktime_sub(ts->sys_monoraw, history->raw)); 1144 ret = scale64_check_overflow(partial_history_cycles, 1145 total_history_cycles, &corr_raw); 1146 if (ret) 1147 return ret; 1148 1149 /* 1150 * If there is a discontinuity in the history, scale monotonic raw 1151 * correction by: 1152 * mult(real)/mult(raw) yielding the realtime correction 1153 * Otherwise, calculate the realtime correction similar to monotonic 1154 * raw calculation 1155 */ 1156 if (discontinuity) { 1157 corr_real = mul_u64_u32_div 1158 (corr_raw, tk->tkr_mono.mult, tk->tkr_raw.mult); 1159 } else { 1160 corr_real = (u64)ktime_to_ns( 1161 ktime_sub(ts->sys_realtime, history->real)); 1162 ret = scale64_check_overflow(partial_history_cycles, 1163 total_history_cycles, &corr_real); 1164 if (ret) 1165 return ret; 1166 } 1167 1168 /* Fixup monotonic raw and real time time values */ 1169 if (interp_forward) { 1170 ts->sys_monoraw = ktime_add_ns(history->raw, corr_raw); 1171 ts->sys_realtime = ktime_add_ns(history->real, corr_real); 1172 } else { 1173 ts->sys_monoraw = ktime_sub_ns(ts->sys_monoraw, corr_raw); 1174 ts->sys_realtime = ktime_sub_ns(ts->sys_realtime, corr_real); 1175 } 1176 1177 return 0; 1178 } 1179 1180 /* 1181 * cycle_between - true if test occurs chronologically between before and after 1182 */ 1183 static bool cycle_between(u64 before, u64 test, u64 after) 1184 { 1185 if (test > before && test < after) 1186 return true; 1187 if (test < before && before > after) 1188 return true; 1189 return false; 1190 } 1191 1192 /** 1193 * get_device_system_crosststamp - Synchronously capture system/device timestamp 1194 * @get_time_fn: Callback to get simultaneous device time and 1195 * system counter from the device driver 1196 * @ctx: Context passed to get_time_fn() 1197 * @history_begin: Historical reference point used to interpolate system 1198 * time when counter provided by the driver is before the current interval 1199 * @xtstamp: Receives simultaneously captured system and device time 1200 * 1201 * Reads a timestamp from a device and correlates it to system time 1202 */ 1203 int get_device_system_crosststamp(int (*get_time_fn) 1204 (ktime_t *device_time, 1205 struct system_counterval_t *sys_counterval, 1206 void *ctx), 1207 void *ctx, 1208 struct system_time_snapshot *history_begin, 1209 struct system_device_crosststamp *xtstamp) 1210 { 1211 struct system_counterval_t system_counterval; 1212 struct timekeeper *tk = &tk_core.timekeeper; 1213 u64 cycles, now, interval_start; 1214 unsigned int clock_was_set_seq = 0; 1215 ktime_t base_real, base_raw; 1216 u64 nsec_real, nsec_raw; 1217 u8 cs_was_changed_seq; 1218 unsigned int seq; 1219 bool do_interp; 1220 int ret; 1221 1222 do { 1223 seq = read_seqcount_begin(&tk_core.seq); 1224 /* 1225 * Try to synchronously capture device time and a system 1226 * counter value calling back into the device driver 1227 */ 1228 ret = get_time_fn(&xtstamp->device, &system_counterval, ctx); 1229 if (ret) 1230 return ret; 1231 1232 /* 1233 * Verify that the clocksource associated with the captured 1234 * system counter value is the same as the currently installed 1235 * timekeeper clocksource 1236 */ 1237 if (tk->tkr_mono.clock != system_counterval.cs) 1238 return -ENODEV; 1239 cycles = system_counterval.cycles; 1240 1241 /* 1242 * Check whether the system counter value provided by the 1243 * device driver is on the current timekeeping interval. 1244 */ 1245 now = tk_clock_read(&tk->tkr_mono); 1246 interval_start = tk->tkr_mono.cycle_last; 1247 if (!cycle_between(interval_start, cycles, now)) { 1248 clock_was_set_seq = tk->clock_was_set_seq; 1249 cs_was_changed_seq = tk->cs_was_changed_seq; 1250 cycles = interval_start; 1251 do_interp = true; 1252 } else { 1253 do_interp = false; 1254 } 1255 1256 base_real = ktime_add(tk->tkr_mono.base, 1257 tk_core.timekeeper.offs_real); 1258 base_raw = tk->tkr_raw.base; 1259 1260 nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono, 1261 system_counterval.cycles); 1262 nsec_raw = timekeeping_cycles_to_ns(&tk->tkr_raw, 1263 system_counterval.cycles); 1264 } while (read_seqcount_retry(&tk_core.seq, seq)); 1265 1266 xtstamp->sys_realtime = ktime_add_ns(base_real, nsec_real); 1267 xtstamp->sys_monoraw = ktime_add_ns(base_raw, nsec_raw); 1268 1269 /* 1270 * Interpolate if necessary, adjusting back from the start of the 1271 * current interval 1272 */ 1273 if (do_interp) { 1274 u64 partial_history_cycles, total_history_cycles; 1275 bool discontinuity; 1276 1277 /* 1278 * Check that the counter value occurs after the provided 1279 * history reference and that the history doesn't cross a 1280 * clocksource change 1281 */ 1282 if (!history_begin || 1283 !cycle_between(history_begin->cycles, 1284 system_counterval.cycles, cycles) || 1285 history_begin->cs_was_changed_seq != cs_was_changed_seq) 1286 return -EINVAL; 1287 partial_history_cycles = cycles - system_counterval.cycles; 1288 total_history_cycles = cycles - history_begin->cycles; 1289 discontinuity = 1290 history_begin->clock_was_set_seq != clock_was_set_seq; 1291 1292 ret = adjust_historical_crosststamp(history_begin, 1293 partial_history_cycles, 1294 total_history_cycles, 1295 discontinuity, xtstamp); 1296 if (ret) 1297 return ret; 1298 } 1299 1300 return 0; 1301 } 1302 EXPORT_SYMBOL_GPL(get_device_system_crosststamp); 1303 1304 /** 1305 * do_settimeofday64 - Sets the time of day. 1306 * @ts: pointer to the timespec64 variable containing the new time 1307 * 1308 * Sets the time of day to the new time and update NTP and notify hrtimers 1309 */ 1310 int do_settimeofday64(const struct timespec64 *ts) 1311 { 1312 struct timekeeper *tk = &tk_core.timekeeper; 1313 struct timespec64 ts_delta, xt; 1314 unsigned long flags; 1315 int ret = 0; 1316 1317 if (!timespec64_valid_settod(ts)) 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 xt = tk_xtime(tk); 1326 ts_delta = timespec64_sub(*ts, xt); 1327 1328 if (timespec64_compare(&tk->wall_to_monotonic, &ts_delta) > 0) { 1329 ret = -EINVAL; 1330 goto out; 1331 } 1332 1333 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts_delta)); 1334 1335 tk_set_xtime(tk, ts); 1336 out: 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(CLOCK_SET_WALL); 1344 1345 if (!ret) 1346 audit_tk_injoffset(ts_delta); 1347 1348 return ret; 1349 } 1350 EXPORT_SYMBOL(do_settimeofday64); 1351 1352 /** 1353 * timekeeping_inject_offset - Adds or subtracts from the current time. 1354 * @ts: Pointer to the timespec variable containing the offset 1355 * 1356 * Adds or subtracts an offset value from the current time. 1357 */ 1358 static int timekeeping_inject_offset(const struct timespec64 *ts) 1359 { 1360 struct timekeeper *tk = &tk_core.timekeeper; 1361 unsigned long flags; 1362 struct timespec64 tmp; 1363 int ret = 0; 1364 1365 if (ts->tv_nsec < 0 || ts->tv_nsec >= NSEC_PER_SEC) 1366 return -EINVAL; 1367 1368 raw_spin_lock_irqsave(&timekeeper_lock, flags); 1369 write_seqcount_begin(&tk_core.seq); 1370 1371 timekeeping_forward_now(tk); 1372 1373 /* Make sure the proposed value is valid */ 1374 tmp = timespec64_add(tk_xtime(tk), *ts); 1375 if (timespec64_compare(&tk->wall_to_monotonic, ts) > 0 || 1376 !timespec64_valid_settod(&tmp)) { 1377 ret = -EINVAL; 1378 goto error; 1379 } 1380 1381 tk_xtime_add(tk, ts); 1382 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *ts)); 1383 1384 error: /* even if we error out, we forwarded the time, so call update */ 1385 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET); 1386 1387 write_seqcount_end(&tk_core.seq); 1388 raw_spin_unlock_irqrestore(&timekeeper_lock, flags); 1389 1390 /* Signal hrtimers about time change */ 1391 clock_was_set(CLOCK_SET_WALL); 1392 1393 return ret; 1394 } 1395 1396 /* 1397 * Indicates if there is an offset between the system clock and the hardware 1398 * clock/persistent clock/rtc. 1399 */ 1400 int persistent_clock_is_local; 1401 1402 /* 1403 * Adjust the time obtained from the CMOS to be UTC time instead of 1404 * local time. 1405 * 1406 * This is ugly, but preferable to the alternatives. Otherwise we 1407 * would either need to write a program to do it in /etc/rc (and risk 1408 * confusion if the program gets run more than once; it would also be 1409 * hard to make the program warp the clock precisely n hours) or 1410 * compile in the timezone information into the kernel. Bad, bad.... 1411 * 1412 * - TYT, 1992-01-01 1413 * 1414 * The best thing to do is to keep the CMOS clock in universal time (UTC) 1415 * as real UNIX machines always do it. This avoids all headaches about 1416 * daylight saving times and warping kernel clocks. 1417 */ 1418 void timekeeping_warp_clock(void) 1419 { 1420 if (sys_tz.tz_minuteswest != 0) { 1421 struct timespec64 adjust; 1422 1423 persistent_clock_is_local = 1; 1424 adjust.tv_sec = sys_tz.tz_minuteswest * 60; 1425 adjust.tv_nsec = 0; 1426 timekeeping_inject_offset(&adjust); 1427 } 1428 } 1429 1430 /* 1431 * __timekeeping_set_tai_offset - Sets the TAI offset from UTC and monotonic 1432 */ 1433 static void __timekeeping_set_tai_offset(struct timekeeper *tk, s32 tai_offset) 1434 { 1435 tk->tai_offset = tai_offset; 1436 tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tai_offset, 0)); 1437 } 1438 1439 /* 1440 * change_clocksource - Swaps clocksources if a new one is available 1441 * 1442 * Accumulates current time interval and initializes new clocksource 1443 */ 1444 static int change_clocksource(void *data) 1445 { 1446 struct timekeeper *tk = &tk_core.timekeeper; 1447 struct clocksource *new, *old = NULL; 1448 unsigned long flags; 1449 bool change = false; 1450 1451 new = (struct clocksource *) data; 1452 1453 /* 1454 * If the cs is in module, get a module reference. Succeeds 1455 * for built-in code (owner == NULL) as well. 1456 */ 1457 if (try_module_get(new->owner)) { 1458 if (!new->enable || new->enable(new) == 0) 1459 change = true; 1460 else 1461 module_put(new->owner); 1462 } 1463 1464 raw_spin_lock_irqsave(&timekeeper_lock, flags); 1465 write_seqcount_begin(&tk_core.seq); 1466 1467 timekeeping_forward_now(tk); 1468 1469 if (change) { 1470 old = tk->tkr_mono.clock; 1471 tk_setup_internals(tk, new); 1472 } 1473 1474 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET); 1475 1476 write_seqcount_end(&tk_core.seq); 1477 raw_spin_unlock_irqrestore(&timekeeper_lock, flags); 1478 1479 if (old) { 1480 if (old->disable) 1481 old->disable(old); 1482 1483 module_put(old->owner); 1484 } 1485 1486 return 0; 1487 } 1488 1489 /** 1490 * timekeeping_notify - Install a new clock source 1491 * @clock: pointer to the clock source 1492 * 1493 * This function is called from clocksource.c after a new, better clock 1494 * source has been registered. The caller holds the clocksource_mutex. 1495 */ 1496 int timekeeping_notify(struct clocksource *clock) 1497 { 1498 struct timekeeper *tk = &tk_core.timekeeper; 1499 1500 if (tk->tkr_mono.clock == clock) 1501 return 0; 1502 stop_machine(change_clocksource, clock, NULL); 1503 tick_clock_notify(); 1504 return tk->tkr_mono.clock == clock ? 0 : -1; 1505 } 1506 1507 /** 1508 * ktime_get_raw_ts64 - Returns the raw monotonic time in a timespec 1509 * @ts: pointer to the timespec64 to be set 1510 * 1511 * Returns the raw monotonic time (completely un-modified by ntp) 1512 */ 1513 void ktime_get_raw_ts64(struct timespec64 *ts) 1514 { 1515 struct timekeeper *tk = &tk_core.timekeeper; 1516 unsigned int seq; 1517 u64 nsecs; 1518 1519 do { 1520 seq = read_seqcount_begin(&tk_core.seq); 1521 ts->tv_sec = tk->raw_sec; 1522 nsecs = timekeeping_get_ns(&tk->tkr_raw); 1523 1524 } while (read_seqcount_retry(&tk_core.seq, seq)); 1525 1526 ts->tv_nsec = 0; 1527 timespec64_add_ns(ts, nsecs); 1528 } 1529 EXPORT_SYMBOL(ktime_get_raw_ts64); 1530 1531 1532 /** 1533 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres 1534 */ 1535 int timekeeping_valid_for_hres(void) 1536 { 1537 struct timekeeper *tk = &tk_core.timekeeper; 1538 unsigned int seq; 1539 int ret; 1540 1541 do { 1542 seq = read_seqcount_begin(&tk_core.seq); 1543 1544 ret = tk->tkr_mono.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES; 1545 1546 } while (read_seqcount_retry(&tk_core.seq, seq)); 1547 1548 return ret; 1549 } 1550 1551 /** 1552 * timekeeping_max_deferment - Returns max time the clocksource can be deferred 1553 */ 1554 u64 timekeeping_max_deferment(void) 1555 { 1556 struct timekeeper *tk = &tk_core.timekeeper; 1557 unsigned int seq; 1558 u64 ret; 1559 1560 do { 1561 seq = read_seqcount_begin(&tk_core.seq); 1562 1563 ret = tk->tkr_mono.clock->max_idle_ns; 1564 1565 } while (read_seqcount_retry(&tk_core.seq, seq)); 1566 1567 return ret; 1568 } 1569 1570 /** 1571 * read_persistent_clock64 - Return time from the persistent clock. 1572 * @ts: Pointer to the storage for the readout value 1573 * 1574 * Weak dummy function for arches that do not yet support it. 1575 * Reads the time from the battery backed persistent clock. 1576 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported. 1577 * 1578 * XXX - Do be sure to remove it once all arches implement it. 1579 */ 1580 void __weak read_persistent_clock64(struct timespec64 *ts) 1581 { 1582 ts->tv_sec = 0; 1583 ts->tv_nsec = 0; 1584 } 1585 1586 /** 1587 * read_persistent_wall_and_boot_offset - Read persistent clock, and also offset 1588 * from the boot. 1589 * 1590 * Weak dummy function for arches that do not yet support it. 1591 * @wall_time: - current time as returned by persistent clock 1592 * @boot_offset: - offset that is defined as wall_time - boot_time 1593 * 1594 * The default function calculates offset based on the current value of 1595 * local_clock(). This way architectures that support sched_clock() but don't 1596 * support dedicated boot time clock will provide the best estimate of the 1597 * boot time. 1598 */ 1599 void __weak __init 1600 read_persistent_wall_and_boot_offset(struct timespec64 *wall_time, 1601 struct timespec64 *boot_offset) 1602 { 1603 read_persistent_clock64(wall_time); 1604 *boot_offset = ns_to_timespec64(local_clock()); 1605 } 1606 1607 /* 1608 * Flag reflecting whether timekeeping_resume() has injected sleeptime. 1609 * 1610 * The flag starts of false and is only set when a suspend reaches 1611 * timekeeping_suspend(), timekeeping_resume() sets it to false when the 1612 * timekeeper clocksource is not stopping across suspend and has been 1613 * used to update sleep time. If the timekeeper clocksource has stopped 1614 * then the flag stays true and is used by the RTC resume code to decide 1615 * whether sleeptime must be injected and if so the flag gets false then. 1616 * 1617 * If a suspend fails before reaching timekeeping_resume() then the flag 1618 * stays false and prevents erroneous sleeptime injection. 1619 */ 1620 static bool suspend_timing_needed; 1621 1622 /* Flag for if there is a persistent clock on this platform */ 1623 static bool persistent_clock_exists; 1624 1625 /* 1626 * timekeeping_init - Initializes the clocksource and common timekeeping values 1627 */ 1628 void __init timekeeping_init(void) 1629 { 1630 struct timespec64 wall_time, boot_offset, wall_to_mono; 1631 struct timekeeper *tk = &tk_core.timekeeper; 1632 struct clocksource *clock; 1633 unsigned long flags; 1634 1635 read_persistent_wall_and_boot_offset(&wall_time, &boot_offset); 1636 if (timespec64_valid_settod(&wall_time) && 1637 timespec64_to_ns(&wall_time) > 0) { 1638 persistent_clock_exists = true; 1639 } else if (timespec64_to_ns(&wall_time) != 0) { 1640 pr_warn("Persistent clock returned invalid value"); 1641 wall_time = (struct timespec64){0}; 1642 } 1643 1644 if (timespec64_compare(&wall_time, &boot_offset) < 0) 1645 boot_offset = (struct timespec64){0}; 1646 1647 /* 1648 * We want set wall_to_mono, so the following is true: 1649 * wall time + wall_to_mono = boot time 1650 */ 1651 wall_to_mono = timespec64_sub(boot_offset, wall_time); 1652 1653 raw_spin_lock_irqsave(&timekeeper_lock, flags); 1654 write_seqcount_begin(&tk_core.seq); 1655 ntp_init(); 1656 1657 clock = clocksource_default_clock(); 1658 if (clock->enable) 1659 clock->enable(clock); 1660 tk_setup_internals(tk, clock); 1661 1662 tk_set_xtime(tk, &wall_time); 1663 tk->raw_sec = 0; 1664 1665 tk_set_wall_to_mono(tk, wall_to_mono); 1666 1667 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET); 1668 1669 write_seqcount_end(&tk_core.seq); 1670 raw_spin_unlock_irqrestore(&timekeeper_lock, flags); 1671 } 1672 1673 /* time in seconds when suspend began for persistent clock */ 1674 static struct timespec64 timekeeping_suspend_time; 1675 1676 /** 1677 * __timekeeping_inject_sleeptime - Internal function to add sleep interval 1678 * @tk: Pointer to the timekeeper to be updated 1679 * @delta: Pointer to the delta value in timespec64 format 1680 * 1681 * Takes a timespec offset measuring a suspend interval and properly 1682 * adds the sleep offset to the timekeeping variables. 1683 */ 1684 static void __timekeeping_inject_sleeptime(struct timekeeper *tk, 1685 const struct timespec64 *delta) 1686 { 1687 if (!timespec64_valid_strict(delta)) { 1688 printk_deferred(KERN_WARNING 1689 "__timekeeping_inject_sleeptime: Invalid " 1690 "sleep delta value!\n"); 1691 return; 1692 } 1693 tk_xtime_add(tk, delta); 1694 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *delta)); 1695 tk_update_sleep_time(tk, timespec64_to_ktime(*delta)); 1696 tk_debug_account_sleep_time(delta); 1697 } 1698 1699 #if defined(CONFIG_PM_SLEEP) && defined(CONFIG_RTC_HCTOSYS_DEVICE) 1700 /** 1701 * We have three kinds of time sources to use for sleep time 1702 * injection, the preference order is: 1703 * 1) non-stop clocksource 1704 * 2) persistent clock (ie: RTC accessible when irqs are off) 1705 * 3) RTC 1706 * 1707 * 1) and 2) are used by timekeeping, 3) by RTC subsystem. 1708 * If system has neither 1) nor 2), 3) will be used finally. 1709 * 1710 * 1711 * If timekeeping has injected sleeptime via either 1) or 2), 1712 * 3) becomes needless, so in this case we don't need to call 1713 * rtc_resume(), and this is what timekeeping_rtc_skipresume() 1714 * means. 1715 */ 1716 bool timekeeping_rtc_skipresume(void) 1717 { 1718 return !suspend_timing_needed; 1719 } 1720 1721 /** 1722 * 1) can be determined whether to use or not only when doing 1723 * timekeeping_resume() which is invoked after rtc_suspend(), 1724 * so we can't skip rtc_suspend() surely if system has 1). 1725 * 1726 * But if system has 2), 2) will definitely be used, so in this 1727 * case we don't need to call rtc_suspend(), and this is what 1728 * timekeeping_rtc_skipsuspend() means. 1729 */ 1730 bool timekeeping_rtc_skipsuspend(void) 1731 { 1732 return persistent_clock_exists; 1733 } 1734 1735 /** 1736 * timekeeping_inject_sleeptime64 - Adds suspend interval to timeekeeping values 1737 * @delta: pointer to a timespec64 delta value 1738 * 1739 * This hook is for architectures that cannot support read_persistent_clock64 1740 * because their RTC/persistent clock is only accessible when irqs are enabled. 1741 * and also don't have an effective nonstop clocksource. 1742 * 1743 * This function should only be called by rtc_resume(), and allows 1744 * a suspend offset to be injected into the timekeeping values. 1745 */ 1746 void timekeeping_inject_sleeptime64(const struct timespec64 *delta) 1747 { 1748 struct timekeeper *tk = &tk_core.timekeeper; 1749 unsigned long flags; 1750 1751 raw_spin_lock_irqsave(&timekeeper_lock, flags); 1752 write_seqcount_begin(&tk_core.seq); 1753 1754 suspend_timing_needed = false; 1755 1756 timekeeping_forward_now(tk); 1757 1758 __timekeeping_inject_sleeptime(tk, delta); 1759 1760 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET); 1761 1762 write_seqcount_end(&tk_core.seq); 1763 raw_spin_unlock_irqrestore(&timekeeper_lock, flags); 1764 1765 /* Signal hrtimers about time change */ 1766 clock_was_set(CLOCK_SET_WALL | CLOCK_SET_BOOT); 1767 } 1768 #endif 1769 1770 /** 1771 * timekeeping_resume - Resumes the generic timekeeping subsystem. 1772 */ 1773 void timekeeping_resume(void) 1774 { 1775 struct timekeeper *tk = &tk_core.timekeeper; 1776 struct clocksource *clock = tk->tkr_mono.clock; 1777 unsigned long flags; 1778 struct timespec64 ts_new, ts_delta; 1779 u64 cycle_now, nsec; 1780 bool inject_sleeptime = false; 1781 1782 read_persistent_clock64(&ts_new); 1783 1784 clockevents_resume(); 1785 clocksource_resume(); 1786 1787 raw_spin_lock_irqsave(&timekeeper_lock, flags); 1788 write_seqcount_begin(&tk_core.seq); 1789 1790 /* 1791 * After system resumes, we need to calculate the suspended time and 1792 * compensate it for the OS time. There are 3 sources that could be 1793 * used: Nonstop clocksource during suspend, persistent clock and rtc 1794 * device. 1795 * 1796 * One specific platform may have 1 or 2 or all of them, and the 1797 * preference will be: 1798 * suspend-nonstop clocksource -> persistent clock -> rtc 1799 * The less preferred source will only be tried if there is no better 1800 * usable source. The rtc part is handled separately in rtc core code. 1801 */ 1802 cycle_now = tk_clock_read(&tk->tkr_mono); 1803 nsec = clocksource_stop_suspend_timing(clock, cycle_now); 1804 if (nsec > 0) { 1805 ts_delta = ns_to_timespec64(nsec); 1806 inject_sleeptime = true; 1807 } else if (timespec64_compare(&ts_new, &timekeeping_suspend_time) > 0) { 1808 ts_delta = timespec64_sub(ts_new, timekeeping_suspend_time); 1809 inject_sleeptime = true; 1810 } 1811 1812 if (inject_sleeptime) { 1813 suspend_timing_needed = false; 1814 __timekeeping_inject_sleeptime(tk, &ts_delta); 1815 } 1816 1817 /* Re-base the last cycle value */ 1818 tk->tkr_mono.cycle_last = cycle_now; 1819 tk->tkr_raw.cycle_last = cycle_now; 1820 1821 tk->ntp_error = 0; 1822 timekeeping_suspended = 0; 1823 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET); 1824 write_seqcount_end(&tk_core.seq); 1825 raw_spin_unlock_irqrestore(&timekeeper_lock, flags); 1826 1827 touch_softlockup_watchdog(); 1828 1829 /* Resume the clockevent device(s) and hrtimers */ 1830 tick_resume(); 1831 /* Notify timerfd as resume is equivalent to clock_was_set() */ 1832 timerfd_resume(); 1833 } 1834 1835 int timekeeping_suspend(void) 1836 { 1837 struct timekeeper *tk = &tk_core.timekeeper; 1838 unsigned long flags; 1839 struct timespec64 delta, delta_delta; 1840 static struct timespec64 old_delta; 1841 struct clocksource *curr_clock; 1842 u64 cycle_now; 1843 1844 read_persistent_clock64(&timekeeping_suspend_time); 1845 1846 /* 1847 * On some systems the persistent_clock can not be detected at 1848 * timekeeping_init by its return value, so if we see a valid 1849 * value returned, update the persistent_clock_exists flag. 1850 */ 1851 if (timekeeping_suspend_time.tv_sec || timekeeping_suspend_time.tv_nsec) 1852 persistent_clock_exists = true; 1853 1854 suspend_timing_needed = true; 1855 1856 raw_spin_lock_irqsave(&timekeeper_lock, flags); 1857 write_seqcount_begin(&tk_core.seq); 1858 timekeeping_forward_now(tk); 1859 timekeeping_suspended = 1; 1860 1861 /* 1862 * Since we've called forward_now, cycle_last stores the value 1863 * just read from the current clocksource. Save this to potentially 1864 * use in suspend timing. 1865 */ 1866 curr_clock = tk->tkr_mono.clock; 1867 cycle_now = tk->tkr_mono.cycle_last; 1868 clocksource_start_suspend_timing(curr_clock, cycle_now); 1869 1870 if (persistent_clock_exists) { 1871 /* 1872 * To avoid drift caused by repeated suspend/resumes, 1873 * which each can add ~1 second drift error, 1874 * try to compensate so the difference in system time 1875 * and persistent_clock time stays close to constant. 1876 */ 1877 delta = timespec64_sub(tk_xtime(tk), timekeeping_suspend_time); 1878 delta_delta = timespec64_sub(delta, old_delta); 1879 if (abs(delta_delta.tv_sec) >= 2) { 1880 /* 1881 * if delta_delta is too large, assume time correction 1882 * has occurred and set old_delta to the current delta. 1883 */ 1884 old_delta = delta; 1885 } else { 1886 /* Otherwise try to adjust old_system to compensate */ 1887 timekeeping_suspend_time = 1888 timespec64_add(timekeeping_suspend_time, delta_delta); 1889 } 1890 } 1891 1892 timekeeping_update(tk, TK_MIRROR); 1893 halt_fast_timekeeper(tk); 1894 write_seqcount_end(&tk_core.seq); 1895 raw_spin_unlock_irqrestore(&timekeeper_lock, flags); 1896 1897 tick_suspend(); 1898 clocksource_suspend(); 1899 clockevents_suspend(); 1900 1901 return 0; 1902 } 1903 1904 /* sysfs resume/suspend bits for timekeeping */ 1905 static struct syscore_ops timekeeping_syscore_ops = { 1906 .resume = timekeeping_resume, 1907 .suspend = timekeeping_suspend, 1908 }; 1909 1910 static int __init timekeeping_init_ops(void) 1911 { 1912 register_syscore_ops(&timekeeping_syscore_ops); 1913 return 0; 1914 } 1915 device_initcall(timekeeping_init_ops); 1916 1917 /* 1918 * Apply a multiplier adjustment to the timekeeper 1919 */ 1920 static __always_inline void timekeeping_apply_adjustment(struct timekeeper *tk, 1921 s64 offset, 1922 s32 mult_adj) 1923 { 1924 s64 interval = tk->cycle_interval; 1925 1926 if (mult_adj == 0) { 1927 return; 1928 } else if (mult_adj == -1) { 1929 interval = -interval; 1930 offset = -offset; 1931 } else if (mult_adj != 1) { 1932 interval *= mult_adj; 1933 offset *= mult_adj; 1934 } 1935 1936 /* 1937 * So the following can be confusing. 1938 * 1939 * To keep things simple, lets assume mult_adj == 1 for now. 1940 * 1941 * When mult_adj != 1, remember that the interval and offset values 1942 * have been appropriately scaled so the math is the same. 1943 * 1944 * The basic idea here is that we're increasing the multiplier 1945 * by one, this causes the xtime_interval to be incremented by 1946 * one cycle_interval. This is because: 1947 * xtime_interval = cycle_interval * mult 1948 * So if mult is being incremented by one: 1949 * xtime_interval = cycle_interval * (mult + 1) 1950 * Its the same as: 1951 * xtime_interval = (cycle_interval * mult) + cycle_interval 1952 * Which can be shortened to: 1953 * xtime_interval += cycle_interval 1954 * 1955 * So offset stores the non-accumulated cycles. Thus the current 1956 * time (in shifted nanoseconds) is: 1957 * now = (offset * adj) + xtime_nsec 1958 * Now, even though we're adjusting the clock frequency, we have 1959 * to keep time consistent. In other words, we can't jump back 1960 * in time, and we also want to avoid jumping forward in time. 1961 * 1962 * So given the same offset value, we need the time to be the same 1963 * both before and after the freq adjustment. 1964 * now = (offset * adj_1) + xtime_nsec_1 1965 * now = (offset * adj_2) + xtime_nsec_2 1966 * So: 1967 * (offset * adj_1) + xtime_nsec_1 = 1968 * (offset * adj_2) + xtime_nsec_2 1969 * And we know: 1970 * adj_2 = adj_1 + 1 1971 * So: 1972 * (offset * adj_1) + xtime_nsec_1 = 1973 * (offset * (adj_1+1)) + xtime_nsec_2 1974 * (offset * adj_1) + xtime_nsec_1 = 1975 * (offset * adj_1) + offset + xtime_nsec_2 1976 * Canceling the sides: 1977 * xtime_nsec_1 = offset + xtime_nsec_2 1978 * Which gives us: 1979 * xtime_nsec_2 = xtime_nsec_1 - offset 1980 * Which simplifies to: 1981 * xtime_nsec -= offset 1982 */ 1983 if ((mult_adj > 0) && (tk->tkr_mono.mult + mult_adj < mult_adj)) { 1984 /* NTP adjustment caused clocksource mult overflow */ 1985 WARN_ON_ONCE(1); 1986 return; 1987 } 1988 1989 tk->tkr_mono.mult += mult_adj; 1990 tk->xtime_interval += interval; 1991 tk->tkr_mono.xtime_nsec -= offset; 1992 } 1993 1994 /* 1995 * Adjust the timekeeper's multiplier to the correct frequency 1996 * and also to reduce the accumulated error value. 1997 */ 1998 static void timekeeping_adjust(struct timekeeper *tk, s64 offset) 1999 { 2000 u32 mult; 2001 2002 /* 2003 * Determine the multiplier from the current NTP tick length. 2004 * Avoid expensive division when the tick length doesn't change. 2005 */ 2006 if (likely(tk->ntp_tick == ntp_tick_length())) { 2007 mult = tk->tkr_mono.mult - tk->ntp_err_mult; 2008 } else { 2009 tk->ntp_tick = ntp_tick_length(); 2010 mult = div64_u64((tk->ntp_tick >> tk->ntp_error_shift) - 2011 tk->xtime_remainder, tk->cycle_interval); 2012 } 2013 2014 /* 2015 * If the clock is behind the NTP time, increase the multiplier by 1 2016 * to catch up with it. If it's ahead and there was a remainder in the 2017 * tick division, the clock will slow down. Otherwise it will stay 2018 * ahead until the tick length changes to a non-divisible value. 2019 */ 2020 tk->ntp_err_mult = tk->ntp_error > 0 ? 1 : 0; 2021 mult += tk->ntp_err_mult; 2022 2023 timekeeping_apply_adjustment(tk, offset, mult - tk->tkr_mono.mult); 2024 2025 if (unlikely(tk->tkr_mono.clock->maxadj && 2026 (abs(tk->tkr_mono.mult - tk->tkr_mono.clock->mult) 2027 > tk->tkr_mono.clock->maxadj))) { 2028 printk_once(KERN_WARNING 2029 "Adjusting %s more than 11%% (%ld vs %ld)\n", 2030 tk->tkr_mono.clock->name, (long)tk->tkr_mono.mult, 2031 (long)tk->tkr_mono.clock->mult + tk->tkr_mono.clock->maxadj); 2032 } 2033 2034 /* 2035 * It may be possible that when we entered this function, xtime_nsec 2036 * was very small. Further, if we're slightly speeding the clocksource 2037 * in the code above, its possible the required corrective factor to 2038 * xtime_nsec could cause it to underflow. 2039 * 2040 * Now, since we have already accumulated the second and the NTP 2041 * subsystem has been notified via second_overflow(), we need to skip 2042 * the next update. 2043 */ 2044 if (unlikely((s64)tk->tkr_mono.xtime_nsec < 0)) { 2045 tk->tkr_mono.xtime_nsec += (u64)NSEC_PER_SEC << 2046 tk->tkr_mono.shift; 2047 tk->xtime_sec--; 2048 tk->skip_second_overflow = 1; 2049 } 2050 } 2051 2052 /* 2053 * accumulate_nsecs_to_secs - Accumulates nsecs into secs 2054 * 2055 * Helper function that accumulates the nsecs greater than a second 2056 * from the xtime_nsec field to the xtime_secs field. 2057 * It also calls into the NTP code to handle leapsecond processing. 2058 */ 2059 static inline unsigned int accumulate_nsecs_to_secs(struct timekeeper *tk) 2060 { 2061 u64 nsecps = (u64)NSEC_PER_SEC << tk->tkr_mono.shift; 2062 unsigned int clock_set = 0; 2063 2064 while (tk->tkr_mono.xtime_nsec >= nsecps) { 2065 int leap; 2066 2067 tk->tkr_mono.xtime_nsec -= nsecps; 2068 tk->xtime_sec++; 2069 2070 /* 2071 * Skip NTP update if this second was accumulated before, 2072 * i.e. xtime_nsec underflowed in timekeeping_adjust() 2073 */ 2074 if (unlikely(tk->skip_second_overflow)) { 2075 tk->skip_second_overflow = 0; 2076 continue; 2077 } 2078 2079 /* Figure out if its a leap sec and apply if needed */ 2080 leap = second_overflow(tk->xtime_sec); 2081 if (unlikely(leap)) { 2082 struct timespec64 ts; 2083 2084 tk->xtime_sec += leap; 2085 2086 ts.tv_sec = leap; 2087 ts.tv_nsec = 0; 2088 tk_set_wall_to_mono(tk, 2089 timespec64_sub(tk->wall_to_monotonic, ts)); 2090 2091 __timekeeping_set_tai_offset(tk, tk->tai_offset - leap); 2092 2093 clock_set = TK_CLOCK_WAS_SET; 2094 } 2095 } 2096 return clock_set; 2097 } 2098 2099 /* 2100 * logarithmic_accumulation - shifted accumulation of cycles 2101 * 2102 * This functions accumulates a shifted interval of cycles into 2103 * a shifted interval nanoseconds. Allows for O(log) accumulation 2104 * loop. 2105 * 2106 * Returns the unconsumed cycles. 2107 */ 2108 static u64 logarithmic_accumulation(struct timekeeper *tk, u64 offset, 2109 u32 shift, unsigned int *clock_set) 2110 { 2111 u64 interval = tk->cycle_interval << shift; 2112 u64 snsec_per_sec; 2113 2114 /* If the offset is smaller than a shifted interval, do nothing */ 2115 if (offset < interval) 2116 return offset; 2117 2118 /* Accumulate one shifted interval */ 2119 offset -= interval; 2120 tk->tkr_mono.cycle_last += interval; 2121 tk->tkr_raw.cycle_last += interval; 2122 2123 tk->tkr_mono.xtime_nsec += tk->xtime_interval << shift; 2124 *clock_set |= accumulate_nsecs_to_secs(tk); 2125 2126 /* Accumulate raw time */ 2127 tk->tkr_raw.xtime_nsec += tk->raw_interval << shift; 2128 snsec_per_sec = (u64)NSEC_PER_SEC << tk->tkr_raw.shift; 2129 while (tk->tkr_raw.xtime_nsec >= snsec_per_sec) { 2130 tk->tkr_raw.xtime_nsec -= snsec_per_sec; 2131 tk->raw_sec++; 2132 } 2133 2134 /* Accumulate error between NTP and clock interval */ 2135 tk->ntp_error += tk->ntp_tick << shift; 2136 tk->ntp_error -= (tk->xtime_interval + tk->xtime_remainder) << 2137 (tk->ntp_error_shift + shift); 2138 2139 return offset; 2140 } 2141 2142 /* 2143 * timekeeping_advance - Updates the timekeeper to the current time and 2144 * current NTP tick length 2145 */ 2146 static bool timekeeping_advance(enum timekeeping_adv_mode mode) 2147 { 2148 struct timekeeper *real_tk = &tk_core.timekeeper; 2149 struct timekeeper *tk = &shadow_timekeeper; 2150 u64 offset; 2151 int shift = 0, maxshift; 2152 unsigned int clock_set = 0; 2153 unsigned long flags; 2154 2155 raw_spin_lock_irqsave(&timekeeper_lock, flags); 2156 2157 /* Make sure we're fully resumed: */ 2158 if (unlikely(timekeeping_suspended)) 2159 goto out; 2160 2161 offset = clocksource_delta(tk_clock_read(&tk->tkr_mono), 2162 tk->tkr_mono.cycle_last, tk->tkr_mono.mask); 2163 2164 /* Check if there's really nothing to do */ 2165 if (offset < real_tk->cycle_interval && mode == TK_ADV_TICK) 2166 goto out; 2167 2168 /* Do some additional sanity checking */ 2169 timekeeping_check_update(tk, offset); 2170 2171 /* 2172 * With NO_HZ we may have to accumulate many cycle_intervals 2173 * (think "ticks") worth of time at once. To do this efficiently, 2174 * we calculate the largest doubling multiple of cycle_intervals 2175 * that is smaller than the offset. We then accumulate that 2176 * chunk in one go, and then try to consume the next smaller 2177 * doubled multiple. 2178 */ 2179 shift = ilog2(offset) - ilog2(tk->cycle_interval); 2180 shift = max(0, shift); 2181 /* Bound shift to one less than what overflows tick_length */ 2182 maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1; 2183 shift = min(shift, maxshift); 2184 while (offset >= tk->cycle_interval) { 2185 offset = logarithmic_accumulation(tk, offset, shift, 2186 &clock_set); 2187 if (offset < tk->cycle_interval<<shift) 2188 shift--; 2189 } 2190 2191 /* Adjust the multiplier to correct NTP error */ 2192 timekeeping_adjust(tk, offset); 2193 2194 /* 2195 * Finally, make sure that after the rounding 2196 * xtime_nsec isn't larger than NSEC_PER_SEC 2197 */ 2198 clock_set |= accumulate_nsecs_to_secs(tk); 2199 2200 write_seqcount_begin(&tk_core.seq); 2201 /* 2202 * Update the real timekeeper. 2203 * 2204 * We could avoid this memcpy by switching pointers, but that 2205 * requires changes to all other timekeeper usage sites as 2206 * well, i.e. move the timekeeper pointer getter into the 2207 * spinlocked/seqcount protected sections. And we trade this 2208 * memcpy under the tk_core.seq against one before we start 2209 * updating. 2210 */ 2211 timekeeping_update(tk, clock_set); 2212 memcpy(real_tk, tk, sizeof(*tk)); 2213 /* The memcpy must come last. Do not put anything here! */ 2214 write_seqcount_end(&tk_core.seq); 2215 out: 2216 raw_spin_unlock_irqrestore(&timekeeper_lock, flags); 2217 2218 return !!clock_set; 2219 } 2220 2221 /** 2222 * update_wall_time - Uses the current clocksource to increment the wall time 2223 * 2224 */ 2225 void update_wall_time(void) 2226 { 2227 if (timekeeping_advance(TK_ADV_TICK)) 2228 clock_was_set_delayed(); 2229 } 2230 2231 /** 2232 * getboottime64 - Return the real time of system boot. 2233 * @ts: pointer to the timespec64 to be set 2234 * 2235 * Returns the wall-time of boot in a timespec64. 2236 * 2237 * This is based on the wall_to_monotonic offset and the total suspend 2238 * time. Calls to settimeofday will affect the value returned (which 2239 * basically means that however wrong your real time clock is at boot time, 2240 * you get the right time here). 2241 */ 2242 void getboottime64(struct timespec64 *ts) 2243 { 2244 struct timekeeper *tk = &tk_core.timekeeper; 2245 ktime_t t = ktime_sub(tk->offs_real, tk->offs_boot); 2246 2247 *ts = ktime_to_timespec64(t); 2248 } 2249 EXPORT_SYMBOL_GPL(getboottime64); 2250 2251 void ktime_get_coarse_real_ts64(struct timespec64 *ts) 2252 { 2253 struct timekeeper *tk = &tk_core.timekeeper; 2254 unsigned int seq; 2255 2256 do { 2257 seq = read_seqcount_begin(&tk_core.seq); 2258 2259 *ts = tk_xtime(tk); 2260 } while (read_seqcount_retry(&tk_core.seq, seq)); 2261 } 2262 EXPORT_SYMBOL(ktime_get_coarse_real_ts64); 2263 2264 void ktime_get_coarse_ts64(struct timespec64 *ts) 2265 { 2266 struct timekeeper *tk = &tk_core.timekeeper; 2267 struct timespec64 now, mono; 2268 unsigned int seq; 2269 2270 do { 2271 seq = read_seqcount_begin(&tk_core.seq); 2272 2273 now = tk_xtime(tk); 2274 mono = tk->wall_to_monotonic; 2275 } while (read_seqcount_retry(&tk_core.seq, seq)); 2276 2277 set_normalized_timespec64(ts, now.tv_sec + mono.tv_sec, 2278 now.tv_nsec + mono.tv_nsec); 2279 } 2280 EXPORT_SYMBOL(ktime_get_coarse_ts64); 2281 2282 /* 2283 * Must hold jiffies_lock 2284 */ 2285 void do_timer(unsigned long ticks) 2286 { 2287 jiffies_64 += ticks; 2288 calc_global_load(); 2289 } 2290 2291 /** 2292 * ktime_get_update_offsets_now - hrtimer helper 2293 * @cwsseq: pointer to check and store the clock was set sequence number 2294 * @offs_real: pointer to storage for monotonic -> realtime offset 2295 * @offs_boot: pointer to storage for monotonic -> boottime offset 2296 * @offs_tai: pointer to storage for monotonic -> clock tai offset 2297 * 2298 * Returns current monotonic time and updates the offsets if the 2299 * sequence number in @cwsseq and timekeeper.clock_was_set_seq are 2300 * different. 2301 * 2302 * Called from hrtimer_interrupt() or retrigger_next_event() 2303 */ 2304 ktime_t ktime_get_update_offsets_now(unsigned int *cwsseq, ktime_t *offs_real, 2305 ktime_t *offs_boot, ktime_t *offs_tai) 2306 { 2307 struct timekeeper *tk = &tk_core.timekeeper; 2308 unsigned int seq; 2309 ktime_t base; 2310 u64 nsecs; 2311 2312 do { 2313 seq = read_seqcount_begin(&tk_core.seq); 2314 2315 base = tk->tkr_mono.base; 2316 nsecs = timekeeping_get_ns(&tk->tkr_mono); 2317 base = ktime_add_ns(base, nsecs); 2318 2319 if (*cwsseq != tk->clock_was_set_seq) { 2320 *cwsseq = tk->clock_was_set_seq; 2321 *offs_real = tk->offs_real; 2322 *offs_boot = tk->offs_boot; 2323 *offs_tai = tk->offs_tai; 2324 } 2325 2326 /* Handle leapsecond insertion adjustments */ 2327 if (unlikely(base >= tk->next_leap_ktime)) 2328 *offs_real = ktime_sub(tk->offs_real, ktime_set(1, 0)); 2329 2330 } while (read_seqcount_retry(&tk_core.seq, seq)); 2331 2332 return base; 2333 } 2334 2335 /* 2336 * timekeeping_validate_timex - Ensures the timex is ok for use in do_adjtimex 2337 */ 2338 static int timekeeping_validate_timex(const struct __kernel_timex *txc) 2339 { 2340 if (txc->modes & ADJ_ADJTIME) { 2341 /* singleshot must not be used with any other mode bits */ 2342 if (!(txc->modes & ADJ_OFFSET_SINGLESHOT)) 2343 return -EINVAL; 2344 if (!(txc->modes & ADJ_OFFSET_READONLY) && 2345 !capable(CAP_SYS_TIME)) 2346 return -EPERM; 2347 } else { 2348 /* In order to modify anything, you gotta be super-user! */ 2349 if (txc->modes && !capable(CAP_SYS_TIME)) 2350 return -EPERM; 2351 /* 2352 * if the quartz is off by more than 10% then 2353 * something is VERY wrong! 2354 */ 2355 if (txc->modes & ADJ_TICK && 2356 (txc->tick < 900000/USER_HZ || 2357 txc->tick > 1100000/USER_HZ)) 2358 return -EINVAL; 2359 } 2360 2361 if (txc->modes & ADJ_SETOFFSET) { 2362 /* In order to inject time, you gotta be super-user! */ 2363 if (!capable(CAP_SYS_TIME)) 2364 return -EPERM; 2365 2366 /* 2367 * Validate if a timespec/timeval used to inject a time 2368 * offset is valid. Offsets can be positive or negative, so 2369 * we don't check tv_sec. The value of the timeval/timespec 2370 * is the sum of its fields,but *NOTE*: 2371 * The field tv_usec/tv_nsec must always be non-negative and 2372 * we can't have more nanoseconds/microseconds than a second. 2373 */ 2374 if (txc->time.tv_usec < 0) 2375 return -EINVAL; 2376 2377 if (txc->modes & ADJ_NANO) { 2378 if (txc->time.tv_usec >= NSEC_PER_SEC) 2379 return -EINVAL; 2380 } else { 2381 if (txc->time.tv_usec >= USEC_PER_SEC) 2382 return -EINVAL; 2383 } 2384 } 2385 2386 /* 2387 * Check for potential multiplication overflows that can 2388 * only happen on 64-bit systems: 2389 */ 2390 if ((txc->modes & ADJ_FREQUENCY) && (BITS_PER_LONG == 64)) { 2391 if (LLONG_MIN / PPM_SCALE > txc->freq) 2392 return -EINVAL; 2393 if (LLONG_MAX / PPM_SCALE < txc->freq) 2394 return -EINVAL; 2395 } 2396 2397 return 0; 2398 } 2399 2400 2401 /** 2402 * do_adjtimex() - Accessor function to NTP __do_adjtimex function 2403 */ 2404 int do_adjtimex(struct __kernel_timex *txc) 2405 { 2406 struct timekeeper *tk = &tk_core.timekeeper; 2407 struct audit_ntp_data ad; 2408 bool clock_set = false; 2409 struct timespec64 ts; 2410 unsigned long flags; 2411 s32 orig_tai, tai; 2412 int ret; 2413 2414 /* Validate the data before disabling interrupts */ 2415 ret = timekeeping_validate_timex(txc); 2416 if (ret) 2417 return ret; 2418 2419 if (txc->modes & ADJ_SETOFFSET) { 2420 struct timespec64 delta; 2421 delta.tv_sec = txc->time.tv_sec; 2422 delta.tv_nsec = txc->time.tv_usec; 2423 if (!(txc->modes & ADJ_NANO)) 2424 delta.tv_nsec *= 1000; 2425 ret = timekeeping_inject_offset(&delta); 2426 if (ret) 2427 return ret; 2428 2429 audit_tk_injoffset(delta); 2430 } 2431 2432 audit_ntp_init(&ad); 2433 2434 ktime_get_real_ts64(&ts); 2435 2436 raw_spin_lock_irqsave(&timekeeper_lock, flags); 2437 write_seqcount_begin(&tk_core.seq); 2438 2439 orig_tai = tai = tk->tai_offset; 2440 ret = __do_adjtimex(txc, &ts, &tai, &ad); 2441 2442 if (tai != orig_tai) { 2443 __timekeeping_set_tai_offset(tk, tai); 2444 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET); 2445 clock_set = true; 2446 } 2447 tk_update_leap_state(tk); 2448 2449 write_seqcount_end(&tk_core.seq); 2450 raw_spin_unlock_irqrestore(&timekeeper_lock, flags); 2451 2452 audit_ntp_log(&ad); 2453 2454 /* Update the multiplier immediately if frequency was set directly */ 2455 if (txc->modes & (ADJ_FREQUENCY | ADJ_TICK)) 2456 clock_set |= timekeeping_advance(TK_ADV_FREQ); 2457 2458 if (clock_set) 2459 clock_was_set(CLOCK_REALTIME); 2460 2461 ntp_notify_cmos_timer(); 2462 2463 return ret; 2464 } 2465 2466 #ifdef CONFIG_NTP_PPS 2467 /** 2468 * hardpps() - Accessor function to NTP __hardpps function 2469 */ 2470 void hardpps(const struct timespec64 *phase_ts, const struct timespec64 *raw_ts) 2471 { 2472 unsigned long flags; 2473 2474 raw_spin_lock_irqsave(&timekeeper_lock, flags); 2475 write_seqcount_begin(&tk_core.seq); 2476 2477 __hardpps(phase_ts, raw_ts); 2478 2479 write_seqcount_end(&tk_core.seq); 2480 raw_spin_unlock_irqrestore(&timekeeper_lock, flags); 2481 } 2482 EXPORT_SYMBOL(hardpps); 2483 #endif /* CONFIG_NTP_PPS */ 2484