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 deltascaled = deltascaled * get_paca()->system_time / 245 (get_paca()->system_time + get_paca()->user_time); 246 delta += get_paca()->system_time; 247 get_paca()->system_time = 0; 248 } 249 account_system_time(tsk, 0, delta); 250 get_paca()->purrdelta = delta; 251 account_system_time_scaled(tsk, deltascaled); 252 get_paca()->spurrdelta = deltascaled; 253 local_irq_restore(flags); 254 } 255 256 /* 257 * Transfer the user and system times accumulated in the paca 258 * by the exception entry and exit code to the generic process 259 * user and system time records. 260 * Must be called with interrupts disabled. 261 */ 262 void account_process_tick(struct task_struct *tsk, int user_tick) 263 { 264 cputime_t utime, utimescaled; 265 266 utime = get_paca()->user_time; 267 get_paca()->user_time = 0; 268 account_user_time(tsk, utime); 269 270 /* Estimate the scaled utime by scaling the real utime based 271 * on the last spurr to purr ratio */ 272 utimescaled = utime * get_paca()->spurrdelta / get_paca()->purrdelta; 273 get_paca()->spurrdelta = get_paca()->purrdelta = 0; 274 account_user_time_scaled(tsk, utimescaled); 275 } 276 277 /* 278 * Stuff for accounting stolen time. 279 */ 280 struct cpu_purr_data { 281 int initialized; /* thread is running */ 282 u64 tb; /* last TB value read */ 283 u64 purr; /* last PURR value read */ 284 u64 spurr; /* last SPURR value read */ 285 }; 286 287 /* 288 * Each entry in the cpu_purr_data array is manipulated only by its 289 * "owner" cpu -- usually in the timer interrupt but also occasionally 290 * in process context for cpu online. As long as cpus do not touch 291 * each others' cpu_purr_data, disabling local interrupts is 292 * sufficient to serialize accesses. 293 */ 294 static DEFINE_PER_CPU(struct cpu_purr_data, cpu_purr_data); 295 296 static void snapshot_tb_and_purr(void *data) 297 { 298 unsigned long flags; 299 struct cpu_purr_data *p = &__get_cpu_var(cpu_purr_data); 300 301 local_irq_save(flags); 302 p->tb = get_tb_or_rtc(); 303 p->purr = mfspr(SPRN_PURR); 304 wmb(); 305 p->initialized = 1; 306 local_irq_restore(flags); 307 } 308 309 /* 310 * Called during boot when all cpus have come up. 311 */ 312 void snapshot_timebases(void) 313 { 314 if (!cpu_has_feature(CPU_FTR_PURR)) 315 return; 316 on_each_cpu(snapshot_tb_and_purr, NULL, 0, 1); 317 } 318 319 /* 320 * Must be called with interrupts disabled. 321 */ 322 void calculate_steal_time(void) 323 { 324 u64 tb, purr; 325 s64 stolen; 326 struct cpu_purr_data *pme; 327 328 if (!cpu_has_feature(CPU_FTR_PURR)) 329 return; 330 pme = &per_cpu(cpu_purr_data, smp_processor_id()); 331 if (!pme->initialized) 332 return; /* this can happen in early boot */ 333 tb = mftb(); 334 purr = mfspr(SPRN_PURR); 335 stolen = (tb - pme->tb) - (purr - pme->purr); 336 if (stolen > 0) 337 account_steal_time(current, stolen); 338 pme->tb = tb; 339 pme->purr = purr; 340 } 341 342 #ifdef CONFIG_PPC_SPLPAR 343 /* 344 * Must be called before the cpu is added to the online map when 345 * a cpu is being brought up at runtime. 346 */ 347 static void snapshot_purr(void) 348 { 349 struct cpu_purr_data *pme; 350 unsigned long flags; 351 352 if (!cpu_has_feature(CPU_FTR_PURR)) 353 return; 354 local_irq_save(flags); 355 pme = &per_cpu(cpu_purr_data, smp_processor_id()); 356 pme->tb = mftb(); 357 pme->purr = mfspr(SPRN_PURR); 358 pme->initialized = 1; 359 local_irq_restore(flags); 360 } 361 362 #endif /* CONFIG_PPC_SPLPAR */ 363 364 #else /* ! CONFIG_VIRT_CPU_ACCOUNTING */ 365 #define calc_cputime_factors() 366 #define calculate_steal_time() do { } while (0) 367 #endif 368 369 #if !(defined(CONFIG_VIRT_CPU_ACCOUNTING) && defined(CONFIG_PPC_SPLPAR)) 370 #define snapshot_purr() do { } while (0) 371 #endif 372 373 /* 374 * Called when a cpu comes up after the system has finished booting, 375 * i.e. as a result of a hotplug cpu action. 376 */ 377 void snapshot_timebase(void) 378 { 379 __get_cpu_var(last_jiffy) = get_tb_or_rtc(); 380 snapshot_purr(); 381 } 382 383 void __delay(unsigned long loops) 384 { 385 unsigned long start; 386 int diff; 387 388 if (__USE_RTC()) { 389 start = get_rtcl(); 390 do { 391 /* the RTCL register wraps at 1000000000 */ 392 diff = get_rtcl() - start; 393 if (diff < 0) 394 diff += 1000000000; 395 } while (diff < loops); 396 } else { 397 start = get_tbl(); 398 while (get_tbl() - start < loops) 399 HMT_low(); 400 HMT_medium(); 401 } 402 } 403 EXPORT_SYMBOL(__delay); 404 405 void udelay(unsigned long usecs) 406 { 407 __delay(tb_ticks_per_usec * usecs); 408 } 409 EXPORT_SYMBOL(udelay); 410 411 412 /* 413 * There are two copies of tb_to_xs and stamp_xsec so that no 414 * lock is needed to access and use these values in 415 * do_gettimeofday. We alternate the copies and as long as a 416 * reasonable time elapses between changes, there will never 417 * be inconsistent values. ntpd has a minimum of one minute 418 * between updates. 419 */ 420 static inline void update_gtod(u64 new_tb_stamp, u64 new_stamp_xsec, 421 u64 new_tb_to_xs) 422 { 423 unsigned temp_idx; 424 struct gettimeofday_vars *temp_varp; 425 426 temp_idx = (do_gtod.var_idx == 0); 427 temp_varp = &do_gtod.vars[temp_idx]; 428 429 temp_varp->tb_to_xs = new_tb_to_xs; 430 temp_varp->tb_orig_stamp = new_tb_stamp; 431 temp_varp->stamp_xsec = new_stamp_xsec; 432 smp_mb(); 433 do_gtod.varp = temp_varp; 434 do_gtod.var_idx = temp_idx; 435 436 /* 437 * tb_update_count is used to allow the userspace gettimeofday code 438 * to assure itself that it sees a consistent view of the tb_to_xs and 439 * stamp_xsec variables. It reads the tb_update_count, then reads 440 * tb_to_xs and stamp_xsec and then reads tb_update_count again. If 441 * the two values of tb_update_count match and are even then the 442 * tb_to_xs and stamp_xsec values are consistent. If not, then it 443 * loops back and reads them again until this criteria is met. 444 * We expect the caller to have done the first increment of 445 * vdso_data->tb_update_count already. 446 */ 447 vdso_data->tb_orig_stamp = new_tb_stamp; 448 vdso_data->stamp_xsec = new_stamp_xsec; 449 vdso_data->tb_to_xs = new_tb_to_xs; 450 vdso_data->wtom_clock_sec = wall_to_monotonic.tv_sec; 451 vdso_data->wtom_clock_nsec = wall_to_monotonic.tv_nsec; 452 smp_wmb(); 453 ++(vdso_data->tb_update_count); 454 } 455 456 #ifdef CONFIG_SMP 457 unsigned long profile_pc(struct pt_regs *regs) 458 { 459 unsigned long pc = instruction_pointer(regs); 460 461 if (in_lock_functions(pc)) 462 return regs->link; 463 464 return pc; 465 } 466 EXPORT_SYMBOL(profile_pc); 467 #endif 468 469 #ifdef CONFIG_PPC_ISERIES 470 471 /* 472 * This function recalibrates the timebase based on the 49-bit time-of-day 473 * value in the Titan chip. The Titan is much more accurate than the value 474 * returned by the service processor for the timebase frequency. 475 */ 476 477 static int __init iSeries_tb_recal(void) 478 { 479 struct div_result divres; 480 unsigned long titan, tb; 481 482 /* Make sure we only run on iSeries */ 483 if (!firmware_has_feature(FW_FEATURE_ISERIES)) 484 return -ENODEV; 485 486 tb = get_tb(); 487 titan = HvCallXm_loadTod(); 488 if ( iSeries_recal_titan ) { 489 unsigned long tb_ticks = tb - iSeries_recal_tb; 490 unsigned long titan_usec = (titan - iSeries_recal_titan) >> 12; 491 unsigned long new_tb_ticks_per_sec = (tb_ticks * USEC_PER_SEC)/titan_usec; 492 unsigned long new_tb_ticks_per_jiffy = (new_tb_ticks_per_sec+(HZ/2))/HZ; 493 long tick_diff = new_tb_ticks_per_jiffy - tb_ticks_per_jiffy; 494 char sign = '+'; 495 /* make sure tb_ticks_per_sec and tb_ticks_per_jiffy are consistent */ 496 new_tb_ticks_per_sec = new_tb_ticks_per_jiffy * HZ; 497 498 if ( tick_diff < 0 ) { 499 tick_diff = -tick_diff; 500 sign = '-'; 501 } 502 if ( tick_diff ) { 503 if ( tick_diff < tb_ticks_per_jiffy/25 ) { 504 printk( "Titan recalibrate: new tb_ticks_per_jiffy = %lu (%c%ld)\n", 505 new_tb_ticks_per_jiffy, sign, tick_diff ); 506 tb_ticks_per_jiffy = new_tb_ticks_per_jiffy; 507 tb_ticks_per_sec = new_tb_ticks_per_sec; 508 calc_cputime_factors(); 509 div128_by_32( XSEC_PER_SEC, 0, tb_ticks_per_sec, &divres ); 510 do_gtod.tb_ticks_per_sec = tb_ticks_per_sec; 511 tb_to_xs = divres.result_low; 512 do_gtod.varp->tb_to_xs = tb_to_xs; 513 vdso_data->tb_ticks_per_sec = tb_ticks_per_sec; 514 vdso_data->tb_to_xs = tb_to_xs; 515 } 516 else { 517 printk( "Titan recalibrate: FAILED (difference > 4 percent)\n" 518 " new tb_ticks_per_jiffy = %lu\n" 519 " old tb_ticks_per_jiffy = %lu\n", 520 new_tb_ticks_per_jiffy, tb_ticks_per_jiffy ); 521 } 522 } 523 } 524 iSeries_recal_titan = titan; 525 iSeries_recal_tb = tb; 526 527 /* Called here as now we know accurate values for the timebase */ 528 clocksource_init(); 529 return 0; 530 } 531 late_initcall(iSeries_tb_recal); 532 533 /* Called from platform early init */ 534 void __init iSeries_time_init_early(void) 535 { 536 iSeries_recal_tb = get_tb(); 537 iSeries_recal_titan = HvCallXm_loadTod(); 538 } 539 #endif /* CONFIG_PPC_ISERIES */ 540 541 /* 542 * For iSeries shared processors, we have to let the hypervisor 543 * set the hardware decrementer. We set a virtual decrementer 544 * in the lppaca and call the hypervisor if the virtual 545 * decrementer is less than the current value in the hardware 546 * decrementer. (almost always the new decrementer value will 547 * be greater than the current hardware decementer so the hypervisor 548 * call will not be needed) 549 */ 550 551 /* 552 * timer_interrupt - gets called when the decrementer overflows, 553 * with interrupts disabled. 554 */ 555 void timer_interrupt(struct pt_regs * regs) 556 { 557 struct pt_regs *old_regs; 558 int cpu = smp_processor_id(); 559 struct clock_event_device *evt = &per_cpu(decrementers, cpu); 560 u64 now; 561 562 /* Ensure a positive value is written to the decrementer, or else 563 * some CPUs will continuue to take decrementer exceptions */ 564 set_dec(DECREMENTER_MAX); 565 566 #ifdef CONFIG_PPC32 567 if (atomic_read(&ppc_n_lost_interrupts) != 0) 568 do_IRQ(regs); 569 #endif 570 571 now = get_tb_or_rtc(); 572 if (now < per_cpu(decrementer_next_tb, cpu)) { 573 /* not time for this event yet */ 574 now = per_cpu(decrementer_next_tb, cpu) - now; 575 if (now <= DECREMENTER_MAX) 576 set_dec((unsigned int)now - 1); 577 return; 578 } 579 old_regs = set_irq_regs(regs); 580 irq_enter(); 581 582 calculate_steal_time(); 583 584 #ifdef CONFIG_PPC_ISERIES 585 if (firmware_has_feature(FW_FEATURE_ISERIES)) 586 get_lppaca()->int_dword.fields.decr_int = 0; 587 #endif 588 589 if (evt->event_handler) 590 evt->event_handler(evt); 591 else 592 evt->set_next_event(DECREMENTER_MAX, evt); 593 594 #ifdef CONFIG_PPC_ISERIES 595 if (firmware_has_feature(FW_FEATURE_ISERIES) && hvlpevent_is_pending()) 596 process_hvlpevents(); 597 #endif 598 599 #ifdef CONFIG_PPC64 600 /* collect purr register values often, for accurate calculations */ 601 if (firmware_has_feature(FW_FEATURE_SPLPAR)) { 602 struct cpu_usage *cu = &__get_cpu_var(cpu_usage_array); 603 cu->current_tb = mfspr(SPRN_PURR); 604 } 605 #endif 606 607 irq_exit(); 608 set_irq_regs(old_regs); 609 } 610 611 void wakeup_decrementer(void) 612 { 613 unsigned long ticks; 614 615 /* 616 * The timebase gets saved on sleep and restored on wakeup, 617 * so all we need to do is to reset the decrementer. 618 */ 619 ticks = tb_ticks_since(__get_cpu_var(last_jiffy)); 620 if (ticks < tb_ticks_per_jiffy) 621 ticks = tb_ticks_per_jiffy - ticks; 622 else 623 ticks = 1; 624 set_dec(ticks); 625 } 626 627 #ifdef CONFIG_SMP 628 void __init smp_space_timers(unsigned int max_cpus) 629 { 630 int i; 631 u64 previous_tb = per_cpu(last_jiffy, boot_cpuid); 632 633 /* make sure tb > per_cpu(last_jiffy, cpu) for all cpus always */ 634 previous_tb -= tb_ticks_per_jiffy; 635 636 for_each_possible_cpu(i) { 637 if (i == boot_cpuid) 638 continue; 639 per_cpu(last_jiffy, i) = previous_tb; 640 } 641 } 642 #endif 643 644 /* 645 * Scheduler clock - returns current time in nanosec units. 646 * 647 * Note: mulhdu(a, b) (multiply high double unsigned) returns 648 * the high 64 bits of a * b, i.e. (a * b) >> 64, where a and b 649 * are 64-bit unsigned numbers. 650 */ 651 unsigned long long sched_clock(void) 652 { 653 if (__USE_RTC()) 654 return get_rtc(); 655 return mulhdu(get_tb() - boot_tb, tb_to_ns_scale) << tb_to_ns_shift; 656 } 657 658 static int __init get_freq(char *name, int cells, unsigned long *val) 659 { 660 struct device_node *cpu; 661 const unsigned int *fp; 662 int found = 0; 663 664 /* The cpu node should have timebase and clock frequency properties */ 665 cpu = of_find_node_by_type(NULL, "cpu"); 666 667 if (cpu) { 668 fp = of_get_property(cpu, name, NULL); 669 if (fp) { 670 found = 1; 671 *val = of_read_ulong(fp, cells); 672 } 673 674 of_node_put(cpu); 675 } 676 677 return found; 678 } 679 680 void __init generic_calibrate_decr(void) 681 { 682 ppc_tb_freq = DEFAULT_TB_FREQ; /* hardcoded default */ 683 684 if (!get_freq("ibm,extended-timebase-frequency", 2, &ppc_tb_freq) && 685 !get_freq("timebase-frequency", 1, &ppc_tb_freq)) { 686 687 printk(KERN_ERR "WARNING: Estimating decrementer frequency " 688 "(not found)\n"); 689 } 690 691 ppc_proc_freq = DEFAULT_PROC_FREQ; /* hardcoded default */ 692 693 if (!get_freq("ibm,extended-clock-frequency", 2, &ppc_proc_freq) && 694 !get_freq("clock-frequency", 1, &ppc_proc_freq)) { 695 696 printk(KERN_ERR "WARNING: Estimating processor frequency " 697 "(not found)\n"); 698 } 699 700 #if defined(CONFIG_BOOKE) || defined(CONFIG_40x) 701 /* Set the time base to zero */ 702 mtspr(SPRN_TBWL, 0); 703 mtspr(SPRN_TBWU, 0); 704 705 /* Clear any pending timer interrupts */ 706 mtspr(SPRN_TSR, TSR_ENW | TSR_WIS | TSR_DIS | TSR_FIS); 707 708 /* Enable decrementer interrupt */ 709 mtspr(SPRN_TCR, TCR_DIE); 710 #endif 711 } 712 713 int update_persistent_clock(struct timespec now) 714 { 715 struct rtc_time tm; 716 717 if (!ppc_md.set_rtc_time) 718 return 0; 719 720 to_tm(now.tv_sec + 1 + timezone_offset, &tm); 721 tm.tm_year -= 1900; 722 tm.tm_mon -= 1; 723 724 return ppc_md.set_rtc_time(&tm); 725 } 726 727 unsigned long read_persistent_clock(void) 728 { 729 struct rtc_time tm; 730 static int first = 1; 731 732 /* XXX this is a litle fragile but will work okay in the short term */ 733 if (first) { 734 first = 0; 735 if (ppc_md.time_init) 736 timezone_offset = ppc_md.time_init(); 737 738 /* get_boot_time() isn't guaranteed to be safe to call late */ 739 if (ppc_md.get_boot_time) 740 return ppc_md.get_boot_time() -timezone_offset; 741 } 742 if (!ppc_md.get_rtc_time) 743 return 0; 744 ppc_md.get_rtc_time(&tm); 745 return mktime(tm.tm_year+1900, tm.tm_mon+1, tm.tm_mday, 746 tm.tm_hour, tm.tm_min, tm.tm_sec); 747 } 748 749 /* clocksource code */ 750 static cycle_t rtc_read(void) 751 { 752 return (cycle_t)get_rtc(); 753 } 754 755 static cycle_t timebase_read(void) 756 { 757 return (cycle_t)get_tb(); 758 } 759 760 void update_vsyscall(struct timespec *wall_time, struct clocksource *clock) 761 { 762 u64 t2x, stamp_xsec; 763 764 if (clock != &clocksource_timebase) 765 return; 766 767 /* Make userspace gettimeofday spin until we're done. */ 768 ++vdso_data->tb_update_count; 769 smp_mb(); 770 771 /* XXX this assumes clock->shift == 22 */ 772 /* 4611686018 ~= 2^(20+64-22) / 1e9 */ 773 t2x = (u64) clock->mult * 4611686018ULL; 774 stamp_xsec = (u64) xtime.tv_nsec * XSEC_PER_SEC; 775 do_div(stamp_xsec, 1000000000); 776 stamp_xsec += (u64) xtime.tv_sec * XSEC_PER_SEC; 777 update_gtod(clock->cycle_last, stamp_xsec, t2x); 778 } 779 780 void update_vsyscall_tz(void) 781 { 782 /* Make userspace gettimeofday spin until we're done. */ 783 ++vdso_data->tb_update_count; 784 smp_mb(); 785 vdso_data->tz_minuteswest = sys_tz.tz_minuteswest; 786 vdso_data->tz_dsttime = sys_tz.tz_dsttime; 787 smp_mb(); 788 ++vdso_data->tb_update_count; 789 } 790 791 void __init clocksource_init(void) 792 { 793 struct clocksource *clock; 794 795 if (__USE_RTC()) 796 clock = &clocksource_rtc; 797 else 798 clock = &clocksource_timebase; 799 800 clock->mult = clocksource_hz2mult(tb_ticks_per_sec, clock->shift); 801 802 if (clocksource_register(clock)) { 803 printk(KERN_ERR "clocksource: %s is already registered\n", 804 clock->name); 805 return; 806 } 807 808 printk(KERN_INFO "clocksource: %s mult[%x] shift[%d] registered\n", 809 clock->name, clock->mult, clock->shift); 810 } 811 812 static int decrementer_set_next_event(unsigned long evt, 813 struct clock_event_device *dev) 814 { 815 __get_cpu_var(decrementer_next_tb) = get_tb_or_rtc() + evt; 816 /* The decrementer interrupts on the 0 -> -1 transition */ 817 if (evt) 818 --evt; 819 set_dec(evt); 820 return 0; 821 } 822 823 static void decrementer_set_mode(enum clock_event_mode mode, 824 struct clock_event_device *dev) 825 { 826 if (mode != CLOCK_EVT_MODE_ONESHOT) 827 decrementer_set_next_event(DECREMENTER_MAX, dev); 828 } 829 830 static void register_decrementer_clockevent(int cpu) 831 { 832 struct clock_event_device *dec = &per_cpu(decrementers, cpu); 833 834 *dec = decrementer_clockevent; 835 dec->cpumask = cpumask_of_cpu(cpu); 836 837 printk(KERN_INFO "clockevent: %s mult[%lx] shift[%d] cpu[%d]\n", 838 dec->name, dec->mult, dec->shift, cpu); 839 840 clockevents_register_device(dec); 841 } 842 843 void init_decrementer_clockevent(void) 844 { 845 int cpu = smp_processor_id(); 846 847 decrementer_clockevent.mult = div_sc(ppc_tb_freq, NSEC_PER_SEC, 848 decrementer_clockevent.shift); 849 decrementer_clockevent.max_delta_ns = 850 clockevent_delta2ns(DECREMENTER_MAX, &decrementer_clockevent); 851 decrementer_clockevent.min_delta_ns = 1000; 852 853 register_decrementer_clockevent(cpu); 854 } 855 856 void secondary_cpu_time_init(void) 857 { 858 /* FIME: Should make unrelatred change to move snapshot_timebase 859 * call here ! */ 860 register_decrementer_clockevent(smp_processor_id()); 861 } 862 863 /* This function is only called on the boot processor */ 864 void __init time_init(void) 865 { 866 unsigned long flags; 867 struct div_result res; 868 u64 scale, x; 869 unsigned shift; 870 871 if (__USE_RTC()) { 872 /* 601 processor: dec counts down by 128 every 128ns */ 873 ppc_tb_freq = 1000000000; 874 tb_last_jiffy = get_rtcl(); 875 } else { 876 /* Normal PowerPC with timebase register */ 877 ppc_md.calibrate_decr(); 878 printk(KERN_DEBUG "time_init: decrementer frequency = %lu.%.6lu MHz\n", 879 ppc_tb_freq / 1000000, ppc_tb_freq % 1000000); 880 printk(KERN_DEBUG "time_init: processor frequency = %lu.%.6lu MHz\n", 881 ppc_proc_freq / 1000000, ppc_proc_freq % 1000000); 882 tb_last_jiffy = get_tb(); 883 } 884 885 tb_ticks_per_jiffy = ppc_tb_freq / HZ; 886 tb_ticks_per_sec = ppc_tb_freq; 887 tb_ticks_per_usec = ppc_tb_freq / 1000000; 888 tb_to_us = mulhwu_scale_factor(ppc_tb_freq, 1000000); 889 calc_cputime_factors(); 890 891 /* 892 * Calculate the length of each tick in ns. It will not be 893 * exactly 1e9/HZ unless ppc_tb_freq is divisible by HZ. 894 * We compute 1e9 * tb_ticks_per_jiffy / ppc_tb_freq, 895 * rounded up. 896 */ 897 x = (u64) NSEC_PER_SEC * tb_ticks_per_jiffy + ppc_tb_freq - 1; 898 do_div(x, ppc_tb_freq); 899 tick_nsec = x; 900 last_tick_len = x << TICKLEN_SCALE; 901 902 /* 903 * Compute ticklen_to_xs, which is a factor which gets multiplied 904 * by (last_tick_len << TICKLEN_SHIFT) to get a tb_to_xs value. 905 * It is computed as: 906 * ticklen_to_xs = 2^N / (tb_ticks_per_jiffy * 1e9) 907 * where N = 64 + 20 - TICKLEN_SCALE - TICKLEN_SHIFT 908 * which turns out to be N = 51 - SHIFT_HZ. 909 * This gives the result as a 0.64 fixed-point fraction. 910 * That value is reduced by an offset amounting to 1 xsec per 911 * 2^31 timebase ticks to avoid problems with time going backwards 912 * by 1 xsec when we do timer_recalc_offset due to losing the 913 * fractional xsec. That offset is equal to ppc_tb_freq/2^51 914 * since there are 2^20 xsec in a second. 915 */ 916 div128_by_32((1ULL << 51) - ppc_tb_freq, 0, 917 tb_ticks_per_jiffy << SHIFT_HZ, &res); 918 div128_by_32(res.result_high, res.result_low, NSEC_PER_SEC, &res); 919 ticklen_to_xs = res.result_low; 920 921 /* Compute tb_to_xs from tick_nsec */ 922 tb_to_xs = mulhdu(last_tick_len << TICKLEN_SHIFT, ticklen_to_xs); 923 924 /* 925 * Compute scale factor for sched_clock. 926 * The calibrate_decr() function has set tb_ticks_per_sec, 927 * which is the timebase frequency. 928 * We compute 1e9 * 2^64 / tb_ticks_per_sec and interpret 929 * the 128-bit result as a 64.64 fixed-point number. 930 * We then shift that number right until it is less than 1.0, 931 * giving us the scale factor and shift count to use in 932 * sched_clock(). 933 */ 934 div128_by_32(1000000000, 0, tb_ticks_per_sec, &res); 935 scale = res.result_low; 936 for (shift = 0; res.result_high != 0; ++shift) { 937 scale = (scale >> 1) | (res.result_high << 63); 938 res.result_high >>= 1; 939 } 940 tb_to_ns_scale = scale; 941 tb_to_ns_shift = shift; 942 /* Save the current timebase to pretty up CONFIG_PRINTK_TIME */ 943 boot_tb = get_tb_or_rtc(); 944 945 write_seqlock_irqsave(&xtime_lock, flags); 946 947 /* If platform provided a timezone (pmac), we correct the time */ 948 if (timezone_offset) { 949 sys_tz.tz_minuteswest = -timezone_offset / 60; 950 sys_tz.tz_dsttime = 0; 951 } 952 953 do_gtod.varp = &do_gtod.vars[0]; 954 do_gtod.var_idx = 0; 955 do_gtod.varp->tb_orig_stamp = tb_last_jiffy; 956 __get_cpu_var(last_jiffy) = tb_last_jiffy; 957 do_gtod.varp->stamp_xsec = (u64) xtime.tv_sec * XSEC_PER_SEC; 958 do_gtod.tb_ticks_per_sec = tb_ticks_per_sec; 959 do_gtod.varp->tb_to_xs = tb_to_xs; 960 do_gtod.tb_to_us = tb_to_us; 961 962 vdso_data->tb_orig_stamp = tb_last_jiffy; 963 vdso_data->tb_update_count = 0; 964 vdso_data->tb_ticks_per_sec = tb_ticks_per_sec; 965 vdso_data->stamp_xsec = (u64) xtime.tv_sec * XSEC_PER_SEC; 966 vdso_data->tb_to_xs = tb_to_xs; 967 968 time_freq = 0; 969 970 write_sequnlock_irqrestore(&xtime_lock, flags); 971 972 /* Register the clocksource, if we're not running on iSeries */ 973 if (!firmware_has_feature(FW_FEATURE_ISERIES)) 974 clocksource_init(); 975 976 init_decrementer_clockevent(); 977 } 978 979 980 #define FEBRUARY 2 981 #define STARTOFTIME 1970 982 #define SECDAY 86400L 983 #define SECYR (SECDAY * 365) 984 #define leapyear(year) ((year) % 4 == 0 && \ 985 ((year) % 100 != 0 || (year) % 400 == 0)) 986 #define days_in_year(a) (leapyear(a) ? 366 : 365) 987 #define days_in_month(a) (month_days[(a) - 1]) 988 989 static int month_days[12] = { 990 31, 28, 31, 30, 31, 30, 31, 31, 30, 31, 30, 31 991 }; 992 993 /* 994 * This only works for the Gregorian calendar - i.e. after 1752 (in the UK) 995 */ 996 void GregorianDay(struct rtc_time * tm) 997 { 998 int leapsToDate; 999 int lastYear; 1000 int day; 1001 int MonthOffset[] = { 0, 31, 59, 90, 120, 151, 181, 212, 243, 273, 304, 334 }; 1002 1003 lastYear = tm->tm_year - 1; 1004 1005 /* 1006 * Number of leap corrections to apply up to end of last year 1007 */ 1008 leapsToDate = lastYear / 4 - lastYear / 100 + lastYear / 400; 1009 1010 /* 1011 * This year is a leap year if it is divisible by 4 except when it is 1012 * divisible by 100 unless it is divisible by 400 1013 * 1014 * e.g. 1904 was a leap year, 1900 was not, 1996 is, and 2000 was 1015 */ 1016 day = tm->tm_mon > 2 && leapyear(tm->tm_year); 1017 1018 day += lastYear*365 + leapsToDate + MonthOffset[tm->tm_mon-1] + 1019 tm->tm_mday; 1020 1021 tm->tm_wday = day % 7; 1022 } 1023 1024 void to_tm(int tim, struct rtc_time * tm) 1025 { 1026 register int i; 1027 register long hms, day; 1028 1029 day = tim / SECDAY; 1030 hms = tim % SECDAY; 1031 1032 /* Hours, minutes, seconds are easy */ 1033 tm->tm_hour = hms / 3600; 1034 tm->tm_min = (hms % 3600) / 60; 1035 tm->tm_sec = (hms % 3600) % 60; 1036 1037 /* Number of years in days */ 1038 for (i = STARTOFTIME; day >= days_in_year(i); i++) 1039 day -= days_in_year(i); 1040 tm->tm_year = i; 1041 1042 /* Number of months in days left */ 1043 if (leapyear(tm->tm_year)) 1044 days_in_month(FEBRUARY) = 29; 1045 for (i = 1; day >= days_in_month(i); i++) 1046 day -= days_in_month(i); 1047 days_in_month(FEBRUARY) = 28; 1048 tm->tm_mon = i; 1049 1050 /* Days are what is left over (+1) from all that. */ 1051 tm->tm_mday = day + 1; 1052 1053 /* 1054 * Determine the day of week 1055 */ 1056 GregorianDay(tm); 1057 } 1058 1059 /* Auxiliary function to compute scaling factors */ 1060 /* Actually the choice of a timebase running at 1/4 the of the bus 1061 * frequency giving resolution of a few tens of nanoseconds is quite nice. 1062 * It makes this computation very precise (27-28 bits typically) which 1063 * is optimistic considering the stability of most processor clock 1064 * oscillators and the precision with which the timebase frequency 1065 * is measured but does not harm. 1066 */ 1067 unsigned mulhwu_scale_factor(unsigned inscale, unsigned outscale) 1068 { 1069 unsigned mlt=0, tmp, err; 1070 /* No concern for performance, it's done once: use a stupid 1071 * but safe and compact method to find the multiplier. 1072 */ 1073 1074 for (tmp = 1U<<31; tmp != 0; tmp >>= 1) { 1075 if (mulhwu(inscale, mlt|tmp) < outscale) 1076 mlt |= tmp; 1077 } 1078 1079 /* We might still be off by 1 for the best approximation. 1080 * A side effect of this is that if outscale is too large 1081 * the returned value will be zero. 1082 * Many corner cases have been checked and seem to work, 1083 * some might have been forgotten in the test however. 1084 */ 1085 1086 err = inscale * (mlt+1); 1087 if (err <= inscale/2) 1088 mlt++; 1089 return mlt; 1090 } 1091 1092 /* 1093 * Divide a 128-bit dividend by a 32-bit divisor, leaving a 128 bit 1094 * result. 1095 */ 1096 void div128_by_32(u64 dividend_high, u64 dividend_low, 1097 unsigned divisor, struct div_result *dr) 1098 { 1099 unsigned long a, b, c, d; 1100 unsigned long w, x, y, z; 1101 u64 ra, rb, rc; 1102 1103 a = dividend_high >> 32; 1104 b = dividend_high & 0xffffffff; 1105 c = dividend_low >> 32; 1106 d = dividend_low & 0xffffffff; 1107 1108 w = a / divisor; 1109 ra = ((u64)(a - (w * divisor)) << 32) + b; 1110 1111 rb = ((u64) do_div(ra, divisor) << 32) + c; 1112 x = ra; 1113 1114 rc = ((u64) do_div(rb, divisor) << 32) + d; 1115 y = rb; 1116 1117 do_div(rc, divisor); 1118 z = rc; 1119 1120 dr->result_high = ((u64)w << 32) + x; 1121 dr->result_low = ((u64)y << 32) + z; 1122 1123 } 1124