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