1 /* 2 * Common time routines among all ppc machines. 3 * 4 * Written by Cort Dougan (cort@cs.nmt.edu) to merge 5 * Paul Mackerras' version and mine for PReP and Pmac. 6 * MPC8xx/MBX changes by Dan Malek (dmalek@jlc.net). 7 * Converted for 64-bit by Mike Corrigan (mikejc@us.ibm.com) 8 * 9 * First round of bugfixes by Gabriel Paubert (paubert@iram.es) 10 * to make clock more stable (2.4.0-test5). The only thing 11 * that this code assumes is that the timebases have been synchronized 12 * by firmware on SMP and are never stopped (never do sleep 13 * on SMP then, nap and doze are OK). 14 * 15 * Speeded up do_gettimeofday by getting rid of references to 16 * xtime (which required locks for consistency). (mikejc@us.ibm.com) 17 * 18 * TODO (not necessarily in this file): 19 * - improve precision and reproducibility of timebase frequency 20 * measurement at boot time. (for iSeries, we calibrate the timebase 21 * against the Titan chip's clock.) 22 * - for astronomical applications: add a new function to get 23 * non ambiguous timestamps even around leap seconds. This needs 24 * a new timestamp format and a good name. 25 * 26 * 1997-09-10 Updated NTP code according to technical memorandum Jan '96 27 * "A Kernel Model for Precision Timekeeping" by Dave Mills 28 * 29 * This program is free software; you can redistribute it and/or 30 * modify it under the terms of the GNU General Public License 31 * as published by the Free Software Foundation; either version 32 * 2 of the License, or (at your option) any later version. 33 */ 34 35 #include <linux/errno.h> 36 #include <linux/module.h> 37 #include <linux/sched.h> 38 #include <linux/kernel.h> 39 #include <linux/param.h> 40 #include <linux/string.h> 41 #include <linux/mm.h> 42 #include <linux/interrupt.h> 43 #include <linux/timex.h> 44 #include <linux/kernel_stat.h> 45 #include <linux/time.h> 46 #include <linux/init.h> 47 #include <linux/profile.h> 48 #include <linux/cpu.h> 49 #include <linux/security.h> 50 #include <linux/percpu.h> 51 #include <linux/rtc.h> 52 #include <linux/jiffies.h> 53 #include <linux/posix-timers.h> 54 #include <linux/irq.h> 55 56 #include <asm/io.h> 57 #include <asm/processor.h> 58 #include <asm/nvram.h> 59 #include <asm/cache.h> 60 #include <asm/machdep.h> 61 #include <asm/uaccess.h> 62 #include <asm/time.h> 63 #include <asm/prom.h> 64 #include <asm/irq.h> 65 #include <asm/div64.h> 66 #include <asm/smp.h> 67 #include <asm/vdso_datapage.h> 68 #include <asm/firmware.h> 69 #ifdef CONFIG_PPC_ISERIES 70 #include <asm/iseries/it_lp_queue.h> 71 #include <asm/iseries/hv_call_xm.h> 72 #endif 73 74 /* powerpc clocksource/clockevent code */ 75 76 #include <linux/clockchips.h> 77 #include <linux/clocksource.h> 78 79 static cycle_t rtc_read(void); 80 static struct clocksource clocksource_rtc = { 81 .name = "rtc", 82 .rating = 400, 83 .flags = CLOCK_SOURCE_IS_CONTINUOUS, 84 .mask = CLOCKSOURCE_MASK(64), 85 .shift = 22, 86 .mult = 0, /* To be filled in */ 87 .read = rtc_read, 88 }; 89 90 static cycle_t timebase_read(void); 91 static struct clocksource clocksource_timebase = { 92 .name = "timebase", 93 .rating = 400, 94 .flags = CLOCK_SOURCE_IS_CONTINUOUS, 95 .mask = CLOCKSOURCE_MASK(64), 96 .shift = 22, 97 .mult = 0, /* To be filled in */ 98 .read = timebase_read, 99 }; 100 101 #define DECREMENTER_MAX 0x7fffffff 102 103 static int decrementer_set_next_event(unsigned long evt, 104 struct clock_event_device *dev); 105 static void decrementer_set_mode(enum clock_event_mode mode, 106 struct clock_event_device *dev); 107 108 static struct clock_event_device decrementer_clockevent = { 109 .name = "decrementer", 110 .rating = 200, 111 .shift = 16, 112 .mult = 0, /* To be filled in */ 113 .irq = 0, 114 .set_next_event = decrementer_set_next_event, 115 .set_mode = decrementer_set_mode, 116 .features = CLOCK_EVT_FEAT_ONESHOT, 117 }; 118 119 static DEFINE_PER_CPU(struct clock_event_device, decrementers); 120 void init_decrementer_clockevent(void); 121 static DEFINE_PER_CPU(u64, decrementer_next_tb); 122 123 #ifdef CONFIG_PPC_ISERIES 124 static unsigned long __initdata iSeries_recal_titan; 125 static signed long __initdata iSeries_recal_tb; 126 127 /* Forward declaration is only needed for iSereis compiles */ 128 void __init clocksource_init(void); 129 #endif 130 131 #define XSEC_PER_SEC (1024*1024) 132 133 #ifdef CONFIG_PPC64 134 #define SCALE_XSEC(xsec, max) (((xsec) * max) / XSEC_PER_SEC) 135 #else 136 /* compute ((xsec << 12) * max) >> 32 */ 137 #define SCALE_XSEC(xsec, max) mulhwu((xsec) << 12, max) 138 #endif 139 140 unsigned long tb_ticks_per_jiffy; 141 unsigned long tb_ticks_per_usec = 100; /* sane default */ 142 EXPORT_SYMBOL(tb_ticks_per_usec); 143 unsigned long tb_ticks_per_sec; 144 EXPORT_SYMBOL(tb_ticks_per_sec); /* for cputime_t conversions */ 145 u64 tb_to_xs; 146 unsigned tb_to_us; 147 148 #define TICKLEN_SCALE TICK_LENGTH_SHIFT 149 u64 last_tick_len; /* units are ns / 2^TICKLEN_SCALE */ 150 u64 ticklen_to_xs; /* 0.64 fraction */ 151 152 /* If last_tick_len corresponds to about 1/HZ seconds, then 153 last_tick_len << TICKLEN_SHIFT will be about 2^63. */ 154 #define TICKLEN_SHIFT (63 - 30 - TICKLEN_SCALE + SHIFT_HZ) 155 156 DEFINE_SPINLOCK(rtc_lock); 157 EXPORT_SYMBOL_GPL(rtc_lock); 158 159 static u64 tb_to_ns_scale __read_mostly; 160 static unsigned tb_to_ns_shift __read_mostly; 161 static unsigned long boot_tb __read_mostly; 162 163 struct gettimeofday_struct do_gtod; 164 165 extern struct timezone sys_tz; 166 static long timezone_offset; 167 168 unsigned long ppc_proc_freq; 169 EXPORT_SYMBOL(ppc_proc_freq); 170 unsigned long ppc_tb_freq; 171 172 static u64 tb_last_jiffy __cacheline_aligned_in_smp; 173 static DEFINE_PER_CPU(u64, last_jiffy); 174 175 #ifdef CONFIG_VIRT_CPU_ACCOUNTING 176 /* 177 * Factors for converting from cputime_t (timebase ticks) to 178 * jiffies, milliseconds, seconds, and clock_t (1/USER_HZ seconds). 179 * These are all stored as 0.64 fixed-point binary fractions. 180 */ 181 u64 __cputime_jiffies_factor; 182 EXPORT_SYMBOL(__cputime_jiffies_factor); 183 u64 __cputime_msec_factor; 184 EXPORT_SYMBOL(__cputime_msec_factor); 185 u64 __cputime_sec_factor; 186 EXPORT_SYMBOL(__cputime_sec_factor); 187 u64 __cputime_clockt_factor; 188 EXPORT_SYMBOL(__cputime_clockt_factor); 189 190 static void calc_cputime_factors(void) 191 { 192 struct div_result res; 193 194 div128_by_32(HZ, 0, tb_ticks_per_sec, &res); 195 __cputime_jiffies_factor = res.result_low; 196 div128_by_32(1000, 0, tb_ticks_per_sec, &res); 197 __cputime_msec_factor = res.result_low; 198 div128_by_32(1, 0, tb_ticks_per_sec, &res); 199 __cputime_sec_factor = res.result_low; 200 div128_by_32(USER_HZ, 0, tb_ticks_per_sec, &res); 201 __cputime_clockt_factor = res.result_low; 202 } 203 204 /* 205 * Read the PURR on systems that have it, otherwise the timebase. 206 */ 207 static u64 read_purr(void) 208 { 209 if (cpu_has_feature(CPU_FTR_PURR)) 210 return mfspr(SPRN_PURR); 211 return mftb(); 212 } 213 214 /* 215 * Read the SPURR on systems that have it, otherwise the purr 216 */ 217 static u64 read_spurr(u64 purr) 218 { 219 if (cpu_has_feature(CPU_FTR_SPURR)) 220 return mfspr(SPRN_SPURR); 221 return purr; 222 } 223 224 /* 225 * Account time for a transition between system, hard irq 226 * or soft irq state. 227 */ 228 void account_system_vtime(struct task_struct *tsk) 229 { 230 u64 now, nowscaled, delta, deltascaled; 231 unsigned long flags; 232 233 local_irq_save(flags); 234 now = read_purr(); 235 delta = now - get_paca()->startpurr; 236 get_paca()->startpurr = now; 237 nowscaled = read_spurr(now); 238 deltascaled = nowscaled - get_paca()->startspurr; 239 get_paca()->startspurr = nowscaled; 240 if (!in_interrupt()) { 241 /* deltascaled includes both user and system time. 242 * Hence scale it based on the purr ratio to estimate 243 * the system time */ 244 if (get_paca()->user_time) 245 deltascaled = deltascaled * get_paca()->system_time / 246 (get_paca()->system_time + get_paca()->user_time); 247 delta += get_paca()->system_time; 248 get_paca()->system_time = 0; 249 } 250 account_system_time(tsk, 0, delta); 251 get_paca()->purrdelta = delta; 252 account_system_time_scaled(tsk, deltascaled); 253 get_paca()->spurrdelta = deltascaled; 254 local_irq_restore(flags); 255 } 256 257 /* 258 * Transfer the user and system times accumulated in the paca 259 * by the exception entry and exit code to the generic process 260 * user and system time records. 261 * Must be called with interrupts disabled. 262 */ 263 void account_process_tick(struct task_struct *tsk, int user_tick) 264 { 265 cputime_t utime, utimescaled; 266 267 utime = get_paca()->user_time; 268 get_paca()->user_time = 0; 269 account_user_time(tsk, utime); 270 271 /* Estimate the scaled utime by scaling the real utime based 272 * on the last spurr to purr ratio */ 273 utimescaled = utime * get_paca()->spurrdelta / get_paca()->purrdelta; 274 get_paca()->spurrdelta = get_paca()->purrdelta = 0; 275 account_user_time_scaled(tsk, utimescaled); 276 } 277 278 /* 279 * Stuff for accounting stolen time. 280 */ 281 struct cpu_purr_data { 282 int initialized; /* thread is running */ 283 u64 tb; /* last TB value read */ 284 u64 purr; /* last PURR value read */ 285 u64 spurr; /* last SPURR value read */ 286 }; 287 288 /* 289 * Each entry in the cpu_purr_data array is manipulated only by its 290 * "owner" cpu -- usually in the timer interrupt but also occasionally 291 * in process context for cpu online. As long as cpus do not touch 292 * each others' cpu_purr_data, disabling local interrupts is 293 * sufficient to serialize accesses. 294 */ 295 static DEFINE_PER_CPU(struct cpu_purr_data, cpu_purr_data); 296 297 static void snapshot_tb_and_purr(void *data) 298 { 299 unsigned long flags; 300 struct cpu_purr_data *p = &__get_cpu_var(cpu_purr_data); 301 302 local_irq_save(flags); 303 p->tb = get_tb_or_rtc(); 304 p->purr = mfspr(SPRN_PURR); 305 wmb(); 306 p->initialized = 1; 307 local_irq_restore(flags); 308 } 309 310 /* 311 * Called during boot when all cpus have come up. 312 */ 313 void snapshot_timebases(void) 314 { 315 if (!cpu_has_feature(CPU_FTR_PURR)) 316 return; 317 on_each_cpu(snapshot_tb_and_purr, NULL, 0, 1); 318 } 319 320 /* 321 * Must be called with interrupts disabled. 322 */ 323 void calculate_steal_time(void) 324 { 325 u64 tb, purr; 326 s64 stolen; 327 struct cpu_purr_data *pme; 328 329 if (!cpu_has_feature(CPU_FTR_PURR)) 330 return; 331 pme = &per_cpu(cpu_purr_data, smp_processor_id()); 332 if (!pme->initialized) 333 return; /* this can happen in early boot */ 334 tb = mftb(); 335 purr = mfspr(SPRN_PURR); 336 stolen = (tb - pme->tb) - (purr - pme->purr); 337 if (stolen > 0) 338 account_steal_time(current, stolen); 339 pme->tb = tb; 340 pme->purr = purr; 341 } 342 343 #ifdef CONFIG_PPC_SPLPAR 344 /* 345 * Must be called before the cpu is added to the online map when 346 * a cpu is being brought up at runtime. 347 */ 348 static void snapshot_purr(void) 349 { 350 struct cpu_purr_data *pme; 351 unsigned long flags; 352 353 if (!cpu_has_feature(CPU_FTR_PURR)) 354 return; 355 local_irq_save(flags); 356 pme = &per_cpu(cpu_purr_data, smp_processor_id()); 357 pme->tb = mftb(); 358 pme->purr = mfspr(SPRN_PURR); 359 pme->initialized = 1; 360 local_irq_restore(flags); 361 } 362 363 #endif /* CONFIG_PPC_SPLPAR */ 364 365 #else /* ! CONFIG_VIRT_CPU_ACCOUNTING */ 366 #define calc_cputime_factors() 367 #define calculate_steal_time() do { } while (0) 368 #endif 369 370 #if !(defined(CONFIG_VIRT_CPU_ACCOUNTING) && defined(CONFIG_PPC_SPLPAR)) 371 #define snapshot_purr() do { } while (0) 372 #endif 373 374 /* 375 * Called when a cpu comes up after the system has finished booting, 376 * i.e. as a result of a hotplug cpu action. 377 */ 378 void snapshot_timebase(void) 379 { 380 __get_cpu_var(last_jiffy) = get_tb_or_rtc(); 381 snapshot_purr(); 382 } 383 384 void __delay(unsigned long loops) 385 { 386 unsigned long start; 387 int diff; 388 389 if (__USE_RTC()) { 390 start = get_rtcl(); 391 do { 392 /* the RTCL register wraps at 1000000000 */ 393 diff = get_rtcl() - start; 394 if (diff < 0) 395 diff += 1000000000; 396 } while (diff < loops); 397 } else { 398 start = get_tbl(); 399 while (get_tbl() - start < loops) 400 HMT_low(); 401 HMT_medium(); 402 } 403 } 404 EXPORT_SYMBOL(__delay); 405 406 void udelay(unsigned long usecs) 407 { 408 __delay(tb_ticks_per_usec * usecs); 409 } 410 EXPORT_SYMBOL(udelay); 411 412 413 /* 414 * There are two copies of tb_to_xs and stamp_xsec so that no 415 * lock is needed to access and use these values in 416 * do_gettimeofday. We alternate the copies and as long as a 417 * reasonable time elapses between changes, there will never 418 * be inconsistent values. ntpd has a minimum of one minute 419 * between updates. 420 */ 421 static inline void update_gtod(u64 new_tb_stamp, u64 new_stamp_xsec, 422 u64 new_tb_to_xs) 423 { 424 unsigned temp_idx; 425 struct gettimeofday_vars *temp_varp; 426 427 temp_idx = (do_gtod.var_idx == 0); 428 temp_varp = &do_gtod.vars[temp_idx]; 429 430 temp_varp->tb_to_xs = new_tb_to_xs; 431 temp_varp->tb_orig_stamp = new_tb_stamp; 432 temp_varp->stamp_xsec = new_stamp_xsec; 433 smp_mb(); 434 do_gtod.varp = temp_varp; 435 do_gtod.var_idx = temp_idx; 436 437 /* 438 * tb_update_count is used to allow the userspace gettimeofday code 439 * to assure itself that it sees a consistent view of the tb_to_xs and 440 * stamp_xsec variables. It reads the tb_update_count, then reads 441 * tb_to_xs and stamp_xsec and then reads tb_update_count again. If 442 * the two values of tb_update_count match and are even then the 443 * tb_to_xs and stamp_xsec values are consistent. If not, then it 444 * loops back and reads them again until this criteria is met. 445 * We expect the caller to have done the first increment of 446 * vdso_data->tb_update_count already. 447 */ 448 vdso_data->tb_orig_stamp = new_tb_stamp; 449 vdso_data->stamp_xsec = new_stamp_xsec; 450 vdso_data->tb_to_xs = new_tb_to_xs; 451 vdso_data->wtom_clock_sec = wall_to_monotonic.tv_sec; 452 vdso_data->wtom_clock_nsec = wall_to_monotonic.tv_nsec; 453 smp_wmb(); 454 ++(vdso_data->tb_update_count); 455 } 456 457 #ifdef CONFIG_SMP 458 unsigned long profile_pc(struct pt_regs *regs) 459 { 460 unsigned long pc = instruction_pointer(regs); 461 462 if (in_lock_functions(pc)) 463 return regs->link; 464 465 return pc; 466 } 467 EXPORT_SYMBOL(profile_pc); 468 #endif 469 470 #ifdef CONFIG_PPC_ISERIES 471 472 /* 473 * This function recalibrates the timebase based on the 49-bit time-of-day 474 * value in the Titan chip. The Titan is much more accurate than the value 475 * returned by the service processor for the timebase frequency. 476 */ 477 478 static int __init iSeries_tb_recal(void) 479 { 480 struct div_result divres; 481 unsigned long titan, tb; 482 483 /* Make sure we only run on iSeries */ 484 if (!firmware_has_feature(FW_FEATURE_ISERIES)) 485 return -ENODEV; 486 487 tb = get_tb(); 488 titan = HvCallXm_loadTod(); 489 if ( iSeries_recal_titan ) { 490 unsigned long tb_ticks = tb - iSeries_recal_tb; 491 unsigned long titan_usec = (titan - iSeries_recal_titan) >> 12; 492 unsigned long new_tb_ticks_per_sec = (tb_ticks * USEC_PER_SEC)/titan_usec; 493 unsigned long new_tb_ticks_per_jiffy = (new_tb_ticks_per_sec+(HZ/2))/HZ; 494 long tick_diff = new_tb_ticks_per_jiffy - tb_ticks_per_jiffy; 495 char sign = '+'; 496 /* make sure tb_ticks_per_sec and tb_ticks_per_jiffy are consistent */ 497 new_tb_ticks_per_sec = new_tb_ticks_per_jiffy * HZ; 498 499 if ( tick_diff < 0 ) { 500 tick_diff = -tick_diff; 501 sign = '-'; 502 } 503 if ( tick_diff ) { 504 if ( tick_diff < tb_ticks_per_jiffy/25 ) { 505 printk( "Titan recalibrate: new tb_ticks_per_jiffy = %lu (%c%ld)\n", 506 new_tb_ticks_per_jiffy, sign, tick_diff ); 507 tb_ticks_per_jiffy = new_tb_ticks_per_jiffy; 508 tb_ticks_per_sec = new_tb_ticks_per_sec; 509 calc_cputime_factors(); 510 div128_by_32( XSEC_PER_SEC, 0, tb_ticks_per_sec, &divres ); 511 do_gtod.tb_ticks_per_sec = tb_ticks_per_sec; 512 tb_to_xs = divres.result_low; 513 do_gtod.varp->tb_to_xs = tb_to_xs; 514 vdso_data->tb_ticks_per_sec = tb_ticks_per_sec; 515 vdso_data->tb_to_xs = tb_to_xs; 516 } 517 else { 518 printk( "Titan recalibrate: FAILED (difference > 4 percent)\n" 519 " new tb_ticks_per_jiffy = %lu\n" 520 " old tb_ticks_per_jiffy = %lu\n", 521 new_tb_ticks_per_jiffy, tb_ticks_per_jiffy ); 522 } 523 } 524 } 525 iSeries_recal_titan = titan; 526 iSeries_recal_tb = tb; 527 528 /* Called here as now we know accurate values for the timebase */ 529 clocksource_init(); 530 return 0; 531 } 532 late_initcall(iSeries_tb_recal); 533 534 /* Called from platform early init */ 535 void __init iSeries_time_init_early(void) 536 { 537 iSeries_recal_tb = get_tb(); 538 iSeries_recal_titan = HvCallXm_loadTod(); 539 } 540 #endif /* CONFIG_PPC_ISERIES */ 541 542 /* 543 * For iSeries shared processors, we have to let the hypervisor 544 * set the hardware decrementer. We set a virtual decrementer 545 * in the lppaca and call the hypervisor if the virtual 546 * decrementer is less than the current value in the hardware 547 * decrementer. (almost always the new decrementer value will 548 * be greater than the current hardware decementer so the hypervisor 549 * call will not be needed) 550 */ 551 552 /* 553 * timer_interrupt - gets called when the decrementer overflows, 554 * with interrupts disabled. 555 */ 556 void timer_interrupt(struct pt_regs * regs) 557 { 558 struct pt_regs *old_regs; 559 int cpu = smp_processor_id(); 560 struct clock_event_device *evt = &per_cpu(decrementers, cpu); 561 u64 now; 562 563 /* Ensure a positive value is written to the decrementer, or else 564 * some CPUs will continuue to take decrementer exceptions */ 565 set_dec(DECREMENTER_MAX); 566 567 #ifdef CONFIG_PPC32 568 if (atomic_read(&ppc_n_lost_interrupts) != 0) 569 do_IRQ(regs); 570 #endif 571 572 now = get_tb_or_rtc(); 573 if (now < per_cpu(decrementer_next_tb, cpu)) { 574 /* not time for this event yet */ 575 now = per_cpu(decrementer_next_tb, cpu) - now; 576 if (now <= DECREMENTER_MAX) 577 set_dec((int)now); 578 return; 579 } 580 old_regs = set_irq_regs(regs); 581 irq_enter(); 582 583 calculate_steal_time(); 584 585 #ifdef CONFIG_PPC_ISERIES 586 if (firmware_has_feature(FW_FEATURE_ISERIES)) 587 get_lppaca()->int_dword.fields.decr_int = 0; 588 #endif 589 590 if (evt->event_handler) 591 evt->event_handler(evt); 592 593 #ifdef CONFIG_PPC_ISERIES 594 if (firmware_has_feature(FW_FEATURE_ISERIES) && hvlpevent_is_pending()) 595 process_hvlpevents(); 596 #endif 597 598 #ifdef CONFIG_PPC64 599 /* collect purr register values often, for accurate calculations */ 600 if (firmware_has_feature(FW_FEATURE_SPLPAR)) { 601 struct cpu_usage *cu = &__get_cpu_var(cpu_usage_array); 602 cu->current_tb = mfspr(SPRN_PURR); 603 } 604 #endif 605 606 irq_exit(); 607 set_irq_regs(old_regs); 608 } 609 610 void wakeup_decrementer(void) 611 { 612 unsigned long ticks; 613 614 /* 615 * The timebase gets saved on sleep and restored on wakeup, 616 * so all we need to do is to reset the decrementer. 617 */ 618 ticks = tb_ticks_since(__get_cpu_var(last_jiffy)); 619 if (ticks < tb_ticks_per_jiffy) 620 ticks = tb_ticks_per_jiffy - ticks; 621 else 622 ticks = 1; 623 set_dec(ticks); 624 } 625 626 #ifdef CONFIG_SMP 627 void __init smp_space_timers(unsigned int max_cpus) 628 { 629 int i; 630 u64 previous_tb = per_cpu(last_jiffy, boot_cpuid); 631 632 /* make sure tb > per_cpu(last_jiffy, cpu) for all cpus always */ 633 previous_tb -= tb_ticks_per_jiffy; 634 635 for_each_possible_cpu(i) { 636 if (i == boot_cpuid) 637 continue; 638 per_cpu(last_jiffy, i) = previous_tb; 639 } 640 } 641 #endif 642 643 /* 644 * Scheduler clock - returns current time in nanosec units. 645 * 646 * Note: mulhdu(a, b) (multiply high double unsigned) returns 647 * the high 64 bits of a * b, i.e. (a * b) >> 64, where a and b 648 * are 64-bit unsigned numbers. 649 */ 650 unsigned long long sched_clock(void) 651 { 652 if (__USE_RTC()) 653 return get_rtc(); 654 return mulhdu(get_tb() - boot_tb, tb_to_ns_scale) << tb_to_ns_shift; 655 } 656 657 static int __init get_freq(char *name, int cells, unsigned long *val) 658 { 659 struct device_node *cpu; 660 const unsigned int *fp; 661 int found = 0; 662 663 /* The cpu node should have timebase and clock frequency properties */ 664 cpu = of_find_node_by_type(NULL, "cpu"); 665 666 if (cpu) { 667 fp = of_get_property(cpu, name, NULL); 668 if (fp) { 669 found = 1; 670 *val = of_read_ulong(fp, cells); 671 } 672 673 of_node_put(cpu); 674 } 675 676 return found; 677 } 678 679 void __init generic_calibrate_decr(void) 680 { 681 ppc_tb_freq = DEFAULT_TB_FREQ; /* hardcoded default */ 682 683 if (!get_freq("ibm,extended-timebase-frequency", 2, &ppc_tb_freq) && 684 !get_freq("timebase-frequency", 1, &ppc_tb_freq)) { 685 686 printk(KERN_ERR "WARNING: Estimating decrementer frequency " 687 "(not found)\n"); 688 } 689 690 ppc_proc_freq = DEFAULT_PROC_FREQ; /* hardcoded default */ 691 692 if (!get_freq("ibm,extended-clock-frequency", 2, &ppc_proc_freq) && 693 !get_freq("clock-frequency", 1, &ppc_proc_freq)) { 694 695 printk(KERN_ERR "WARNING: Estimating processor frequency " 696 "(not found)\n"); 697 } 698 699 #if defined(CONFIG_BOOKE) || defined(CONFIG_40x) 700 /* Set the time base to zero */ 701 mtspr(SPRN_TBWL, 0); 702 mtspr(SPRN_TBWU, 0); 703 704 /* Clear any pending timer interrupts */ 705 mtspr(SPRN_TSR, TSR_ENW | TSR_WIS | TSR_DIS | TSR_FIS); 706 707 /* Enable decrementer interrupt */ 708 mtspr(SPRN_TCR, TCR_DIE); 709 #endif 710 } 711 712 int update_persistent_clock(struct timespec now) 713 { 714 struct rtc_time tm; 715 716 if (!ppc_md.set_rtc_time) 717 return 0; 718 719 to_tm(now.tv_sec + 1 + timezone_offset, &tm); 720 tm.tm_year -= 1900; 721 tm.tm_mon -= 1; 722 723 return ppc_md.set_rtc_time(&tm); 724 } 725 726 unsigned long read_persistent_clock(void) 727 { 728 struct rtc_time tm; 729 static int first = 1; 730 731 /* XXX this is a litle fragile but will work okay in the short term */ 732 if (first) { 733 first = 0; 734 if (ppc_md.time_init) 735 timezone_offset = ppc_md.time_init(); 736 737 /* get_boot_time() isn't guaranteed to be safe to call late */ 738 if (ppc_md.get_boot_time) 739 return ppc_md.get_boot_time() -timezone_offset; 740 } 741 if (!ppc_md.get_rtc_time) 742 return 0; 743 ppc_md.get_rtc_time(&tm); 744 return mktime(tm.tm_year+1900, tm.tm_mon+1, tm.tm_mday, 745 tm.tm_hour, tm.tm_min, tm.tm_sec); 746 } 747 748 /* clocksource code */ 749 static cycle_t rtc_read(void) 750 { 751 return (cycle_t)get_rtc(); 752 } 753 754 static cycle_t timebase_read(void) 755 { 756 return (cycle_t)get_tb(); 757 } 758 759 void update_vsyscall(struct timespec *wall_time, struct clocksource *clock) 760 { 761 u64 t2x, stamp_xsec; 762 763 if (clock != &clocksource_timebase) 764 return; 765 766 /* Make userspace gettimeofday spin until we're done. */ 767 ++vdso_data->tb_update_count; 768 smp_mb(); 769 770 /* XXX this assumes clock->shift == 22 */ 771 /* 4611686018 ~= 2^(20+64-22) / 1e9 */ 772 t2x = (u64) clock->mult * 4611686018ULL; 773 stamp_xsec = (u64) xtime.tv_nsec * XSEC_PER_SEC; 774 do_div(stamp_xsec, 1000000000); 775 stamp_xsec += (u64) xtime.tv_sec * XSEC_PER_SEC; 776 update_gtod(clock->cycle_last, stamp_xsec, t2x); 777 } 778 779 void update_vsyscall_tz(void) 780 { 781 /* Make userspace gettimeofday spin until we're done. */ 782 ++vdso_data->tb_update_count; 783 smp_mb(); 784 vdso_data->tz_minuteswest = sys_tz.tz_minuteswest; 785 vdso_data->tz_dsttime = sys_tz.tz_dsttime; 786 smp_mb(); 787 ++vdso_data->tb_update_count; 788 } 789 790 void __init clocksource_init(void) 791 { 792 struct clocksource *clock; 793 794 if (__USE_RTC()) 795 clock = &clocksource_rtc; 796 else 797 clock = &clocksource_timebase; 798 799 clock->mult = clocksource_hz2mult(tb_ticks_per_sec, clock->shift); 800 801 if (clocksource_register(clock)) { 802 printk(KERN_ERR "clocksource: %s is already registered\n", 803 clock->name); 804 return; 805 } 806 807 printk(KERN_INFO "clocksource: %s mult[%x] shift[%d] registered\n", 808 clock->name, clock->mult, clock->shift); 809 } 810 811 static int decrementer_set_next_event(unsigned long evt, 812 struct clock_event_device *dev) 813 { 814 __get_cpu_var(decrementer_next_tb) = get_tb_or_rtc() + evt; 815 set_dec(evt); 816 return 0; 817 } 818 819 static void decrementer_set_mode(enum clock_event_mode mode, 820 struct clock_event_device *dev) 821 { 822 if (mode != CLOCK_EVT_MODE_ONESHOT) 823 decrementer_set_next_event(DECREMENTER_MAX, dev); 824 } 825 826 static void register_decrementer_clockevent(int cpu) 827 { 828 struct clock_event_device *dec = &per_cpu(decrementers, cpu); 829 830 *dec = decrementer_clockevent; 831 dec->cpumask = cpumask_of_cpu(cpu); 832 833 printk(KERN_DEBUG "clockevent: %s mult[%lx] shift[%d] cpu[%d]\n", 834 dec->name, dec->mult, dec->shift, cpu); 835 836 clockevents_register_device(dec); 837 } 838 839 void init_decrementer_clockevent(void) 840 { 841 int cpu = smp_processor_id(); 842 843 decrementer_clockevent.mult = div_sc(ppc_tb_freq, NSEC_PER_SEC, 844 decrementer_clockevent.shift); 845 decrementer_clockevent.max_delta_ns = 846 clockevent_delta2ns(DECREMENTER_MAX, &decrementer_clockevent); 847 decrementer_clockevent.min_delta_ns = 848 clockevent_delta2ns(2, &decrementer_clockevent); 849 850 register_decrementer_clockevent(cpu); 851 } 852 853 void secondary_cpu_time_init(void) 854 { 855 /* FIME: Should make unrelatred change to move snapshot_timebase 856 * call here ! */ 857 register_decrementer_clockevent(smp_processor_id()); 858 } 859 860 /* This function is only called on the boot processor */ 861 void __init time_init(void) 862 { 863 unsigned long flags; 864 struct div_result res; 865 u64 scale, x; 866 unsigned shift; 867 868 if (__USE_RTC()) { 869 /* 601 processor: dec counts down by 128 every 128ns */ 870 ppc_tb_freq = 1000000000; 871 tb_last_jiffy = get_rtcl(); 872 } else { 873 /* Normal PowerPC with timebase register */ 874 ppc_md.calibrate_decr(); 875 printk(KERN_DEBUG "time_init: decrementer frequency = %lu.%.6lu MHz\n", 876 ppc_tb_freq / 1000000, ppc_tb_freq % 1000000); 877 printk(KERN_DEBUG "time_init: processor frequency = %lu.%.6lu MHz\n", 878 ppc_proc_freq / 1000000, ppc_proc_freq % 1000000); 879 tb_last_jiffy = get_tb(); 880 } 881 882 tb_ticks_per_jiffy = ppc_tb_freq / HZ; 883 tb_ticks_per_sec = ppc_tb_freq; 884 tb_ticks_per_usec = ppc_tb_freq / 1000000; 885 tb_to_us = mulhwu_scale_factor(ppc_tb_freq, 1000000); 886 calc_cputime_factors(); 887 888 /* 889 * Calculate the length of each tick in ns. It will not be 890 * exactly 1e9/HZ unless ppc_tb_freq is divisible by HZ. 891 * We compute 1e9 * tb_ticks_per_jiffy / ppc_tb_freq, 892 * rounded up. 893 */ 894 x = (u64) NSEC_PER_SEC * tb_ticks_per_jiffy + ppc_tb_freq - 1; 895 do_div(x, ppc_tb_freq); 896 tick_nsec = x; 897 last_tick_len = x << TICKLEN_SCALE; 898 899 /* 900 * Compute ticklen_to_xs, which is a factor which gets multiplied 901 * by (last_tick_len << TICKLEN_SHIFT) to get a tb_to_xs value. 902 * It is computed as: 903 * ticklen_to_xs = 2^N / (tb_ticks_per_jiffy * 1e9) 904 * where N = 64 + 20 - TICKLEN_SCALE - TICKLEN_SHIFT 905 * which turns out to be N = 51 - SHIFT_HZ. 906 * This gives the result as a 0.64 fixed-point fraction. 907 * That value is reduced by an offset amounting to 1 xsec per 908 * 2^31 timebase ticks to avoid problems with time going backwards 909 * by 1 xsec when we do timer_recalc_offset due to losing the 910 * fractional xsec. That offset is equal to ppc_tb_freq/2^51 911 * since there are 2^20 xsec in a second. 912 */ 913 div128_by_32((1ULL << 51) - ppc_tb_freq, 0, 914 tb_ticks_per_jiffy << SHIFT_HZ, &res); 915 div128_by_32(res.result_high, res.result_low, NSEC_PER_SEC, &res); 916 ticklen_to_xs = res.result_low; 917 918 /* Compute tb_to_xs from tick_nsec */ 919 tb_to_xs = mulhdu(last_tick_len << TICKLEN_SHIFT, ticklen_to_xs); 920 921 /* 922 * Compute scale factor for sched_clock. 923 * The calibrate_decr() function has set tb_ticks_per_sec, 924 * which is the timebase frequency. 925 * We compute 1e9 * 2^64 / tb_ticks_per_sec and interpret 926 * the 128-bit result as a 64.64 fixed-point number. 927 * We then shift that number right until it is less than 1.0, 928 * giving us the scale factor and shift count to use in 929 * sched_clock(). 930 */ 931 div128_by_32(1000000000, 0, tb_ticks_per_sec, &res); 932 scale = res.result_low; 933 for (shift = 0; res.result_high != 0; ++shift) { 934 scale = (scale >> 1) | (res.result_high << 63); 935 res.result_high >>= 1; 936 } 937 tb_to_ns_scale = scale; 938 tb_to_ns_shift = shift; 939 /* Save the current timebase to pretty up CONFIG_PRINTK_TIME */ 940 boot_tb = get_tb_or_rtc(); 941 942 write_seqlock_irqsave(&xtime_lock, flags); 943 944 /* If platform provided a timezone (pmac), we correct the time */ 945 if (timezone_offset) { 946 sys_tz.tz_minuteswest = -timezone_offset / 60; 947 sys_tz.tz_dsttime = 0; 948 } 949 950 do_gtod.varp = &do_gtod.vars[0]; 951 do_gtod.var_idx = 0; 952 do_gtod.varp->tb_orig_stamp = tb_last_jiffy; 953 __get_cpu_var(last_jiffy) = tb_last_jiffy; 954 do_gtod.varp->stamp_xsec = (u64) xtime.tv_sec * XSEC_PER_SEC; 955 do_gtod.tb_ticks_per_sec = tb_ticks_per_sec; 956 do_gtod.varp->tb_to_xs = tb_to_xs; 957 do_gtod.tb_to_us = tb_to_us; 958 959 vdso_data->tb_orig_stamp = tb_last_jiffy; 960 vdso_data->tb_update_count = 0; 961 vdso_data->tb_ticks_per_sec = tb_ticks_per_sec; 962 vdso_data->stamp_xsec = (u64) xtime.tv_sec * XSEC_PER_SEC; 963 vdso_data->tb_to_xs = tb_to_xs; 964 965 time_freq = 0; 966 967 write_sequnlock_irqrestore(&xtime_lock, flags); 968 969 /* Register the clocksource, if we're not running on iSeries */ 970 if (!firmware_has_feature(FW_FEATURE_ISERIES)) 971 clocksource_init(); 972 973 init_decrementer_clockevent(); 974 } 975 976 977 #define FEBRUARY 2 978 #define STARTOFTIME 1970 979 #define SECDAY 86400L 980 #define SECYR (SECDAY * 365) 981 #define leapyear(year) ((year) % 4 == 0 && \ 982 ((year) % 100 != 0 || (year) % 400 == 0)) 983 #define days_in_year(a) (leapyear(a) ? 366 : 365) 984 #define days_in_month(a) (month_days[(a) - 1]) 985 986 static int month_days[12] = { 987 31, 28, 31, 30, 31, 30, 31, 31, 30, 31, 30, 31 988 }; 989 990 /* 991 * This only works for the Gregorian calendar - i.e. after 1752 (in the UK) 992 */ 993 void GregorianDay(struct rtc_time * tm) 994 { 995 int leapsToDate; 996 int lastYear; 997 int day; 998 int MonthOffset[] = { 0, 31, 59, 90, 120, 151, 181, 212, 243, 273, 304, 334 }; 999 1000 lastYear = tm->tm_year - 1; 1001 1002 /* 1003 * Number of leap corrections to apply up to end of last year 1004 */ 1005 leapsToDate = lastYear / 4 - lastYear / 100 + lastYear / 400; 1006 1007 /* 1008 * This year is a leap year if it is divisible by 4 except when it is 1009 * divisible by 100 unless it is divisible by 400 1010 * 1011 * e.g. 1904 was a leap year, 1900 was not, 1996 is, and 2000 was 1012 */ 1013 day = tm->tm_mon > 2 && leapyear(tm->tm_year); 1014 1015 day += lastYear*365 + leapsToDate + MonthOffset[tm->tm_mon-1] + 1016 tm->tm_mday; 1017 1018 tm->tm_wday = day % 7; 1019 } 1020 1021 void to_tm(int tim, struct rtc_time * tm) 1022 { 1023 register int i; 1024 register long hms, day; 1025 1026 day = tim / SECDAY; 1027 hms = tim % SECDAY; 1028 1029 /* Hours, minutes, seconds are easy */ 1030 tm->tm_hour = hms / 3600; 1031 tm->tm_min = (hms % 3600) / 60; 1032 tm->tm_sec = (hms % 3600) % 60; 1033 1034 /* Number of years in days */ 1035 for (i = STARTOFTIME; day >= days_in_year(i); i++) 1036 day -= days_in_year(i); 1037 tm->tm_year = i; 1038 1039 /* Number of months in days left */ 1040 if (leapyear(tm->tm_year)) 1041 days_in_month(FEBRUARY) = 29; 1042 for (i = 1; day >= days_in_month(i); i++) 1043 day -= days_in_month(i); 1044 days_in_month(FEBRUARY) = 28; 1045 tm->tm_mon = i; 1046 1047 /* Days are what is left over (+1) from all that. */ 1048 tm->tm_mday = day + 1; 1049 1050 /* 1051 * Determine the day of week 1052 */ 1053 GregorianDay(tm); 1054 } 1055 1056 /* Auxiliary function to compute scaling factors */ 1057 /* Actually the choice of a timebase running at 1/4 the of the bus 1058 * frequency giving resolution of a few tens of nanoseconds is quite nice. 1059 * It makes this computation very precise (27-28 bits typically) which 1060 * is optimistic considering the stability of most processor clock 1061 * oscillators and the precision with which the timebase frequency 1062 * is measured but does not harm. 1063 */ 1064 unsigned mulhwu_scale_factor(unsigned inscale, unsigned outscale) 1065 { 1066 unsigned mlt=0, tmp, err; 1067 /* No concern for performance, it's done once: use a stupid 1068 * but safe and compact method to find the multiplier. 1069 */ 1070 1071 for (tmp = 1U<<31; tmp != 0; tmp >>= 1) { 1072 if (mulhwu(inscale, mlt|tmp) < outscale) 1073 mlt |= tmp; 1074 } 1075 1076 /* We might still be off by 1 for the best approximation. 1077 * A side effect of this is that if outscale is too large 1078 * the returned value will be zero. 1079 * Many corner cases have been checked and seem to work, 1080 * some might have been forgotten in the test however. 1081 */ 1082 1083 err = inscale * (mlt+1); 1084 if (err <= inscale/2) 1085 mlt++; 1086 return mlt; 1087 } 1088 1089 /* 1090 * Divide a 128-bit dividend by a 32-bit divisor, leaving a 128 bit 1091 * result. 1092 */ 1093 void div128_by_32(u64 dividend_high, u64 dividend_low, 1094 unsigned divisor, struct div_result *dr) 1095 { 1096 unsigned long a, b, c, d; 1097 unsigned long w, x, y, z; 1098 u64 ra, rb, rc; 1099 1100 a = dividend_high >> 32; 1101 b = dividend_high & 0xffffffff; 1102 c = dividend_low >> 32; 1103 d = dividend_low & 0xffffffff; 1104 1105 w = a / divisor; 1106 ra = ((u64)(a - (w * divisor)) << 32) + b; 1107 1108 rb = ((u64) do_div(ra, divisor) << 32) + c; 1109 x = ra; 1110 1111 rc = ((u64) do_div(rb, divisor) << 32) + d; 1112 y = rb; 1113 1114 do_div(rc, divisor); 1115 z = rc; 1116 1117 dr->result_high = ((u64)w << 32) + x; 1118 dr->result_low = ((u64)y << 32) + z; 1119 1120 } 1121