2 * Common time routines among all ppc machines.
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)
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).
15 * Speeded up do_gettimeofday by getting rid of references to
16 * xtime (which required locks for consistency). (mikejc@us.ibm.com)
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.
26 * 1997-09-10 Updated NTP code according to technical memorandum Jan '96
27 * "A Kernel Model for Precision Timekeeping" by Dave Mills
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.
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>
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>
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>
65 #include <asm/div64.h>
67 #include <asm/vdso_datapage.h>
69 #include <asm/firmware.h>
71 #ifdef CONFIG_PPC_ISERIES
72 #include <asm/iseries/it_lp_queue.h>
73 #include <asm/iseries/hv_call_xm.h>
76 /* keep track of when we need to update the rtc */
77 time_t last_rtc_update;
78 #ifdef CONFIG_PPC_ISERIES
79 static unsigned long __initdata iSeries_recal_titan;
80 static signed long __initdata iSeries_recal_tb;
83 /* The decrementer counts down by 128 every 128ns on a 601. */
84 #define DECREMENTER_COUNT_601 (1000000000 / HZ)
86 #define XSEC_PER_SEC (1024*1024)
89 #define SCALE_XSEC(xsec, max) (((xsec) * max) / XSEC_PER_SEC)
91 /* compute ((xsec << 12) * max) >> 32 */
92 #define SCALE_XSEC(xsec, max) mulhwu((xsec) << 12, max)
95 unsigned long tb_ticks_per_jiffy;
96 unsigned long tb_ticks_per_usec = 100; /* sane default */
97 EXPORT_SYMBOL(tb_ticks_per_usec);
98 unsigned long tb_ticks_per_sec;
99 EXPORT_SYMBOL(tb_ticks_per_sec); /* for cputime_t conversions */
103 #define TICKLEN_SCALE TICK_LENGTH_SHIFT
104 u64 last_tick_len; /* units are ns / 2^TICKLEN_SCALE */
105 u64 ticklen_to_xs; /* 0.64 fraction */
107 /* If last_tick_len corresponds to about 1/HZ seconds, then
108 last_tick_len << TICKLEN_SHIFT will be about 2^63. */
109 #define TICKLEN_SHIFT (63 - 30 - TICKLEN_SCALE + SHIFT_HZ)
111 DEFINE_SPINLOCK(rtc_lock);
112 EXPORT_SYMBOL_GPL(rtc_lock);
114 static u64 tb_to_ns_scale __read_mostly;
115 static unsigned tb_to_ns_shift __read_mostly;
116 static unsigned long boot_tb __read_mostly;
118 struct gettimeofday_struct do_gtod;
120 extern struct timezone sys_tz;
121 static long timezone_offset;
123 unsigned long ppc_proc_freq;
124 EXPORT_SYMBOL(ppc_proc_freq);
125 unsigned long ppc_tb_freq;
127 static u64 tb_last_jiffy __cacheline_aligned_in_smp;
128 static DEFINE_PER_CPU(u64, last_jiffy);
130 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
132 * Factors for converting from cputime_t (timebase ticks) to
133 * jiffies, milliseconds, seconds, and clock_t (1/USER_HZ seconds).
134 * These are all stored as 0.64 fixed-point binary fractions.
136 u64 __cputime_jiffies_factor;
137 EXPORT_SYMBOL(__cputime_jiffies_factor);
138 u64 __cputime_msec_factor;
139 EXPORT_SYMBOL(__cputime_msec_factor);
140 u64 __cputime_sec_factor;
141 EXPORT_SYMBOL(__cputime_sec_factor);
142 u64 __cputime_clockt_factor;
143 EXPORT_SYMBOL(__cputime_clockt_factor);
145 static void calc_cputime_factors(void)
147 struct div_result res;
149 div128_by_32(HZ, 0, tb_ticks_per_sec, &res);
150 __cputime_jiffies_factor = res.result_low;
151 div128_by_32(1000, 0, tb_ticks_per_sec, &res);
152 __cputime_msec_factor = res.result_low;
153 div128_by_32(1, 0, tb_ticks_per_sec, &res);
154 __cputime_sec_factor = res.result_low;
155 div128_by_32(USER_HZ, 0, tb_ticks_per_sec, &res);
156 __cputime_clockt_factor = res.result_low;
160 * Read the PURR on systems that have it, otherwise the timebase.
162 static u64 read_purr(void)
164 if (cpu_has_feature(CPU_FTR_PURR))
165 return mfspr(SPRN_PURR);
170 * Account time for a transition between system, hard irq
173 void account_system_vtime(struct task_struct *tsk)
178 local_irq_save(flags);
180 delta = now - get_paca()->startpurr;
181 get_paca()->startpurr = now;
182 if (!in_interrupt()) {
183 delta += get_paca()->system_time;
184 get_paca()->system_time = 0;
186 account_system_time(tsk, 0, delta);
187 local_irq_restore(flags);
191 * Transfer the user and system times accumulated in the paca
192 * by the exception entry and exit code to the generic process
193 * user and system time records.
194 * Must be called with interrupts disabled.
196 void account_process_vtime(struct task_struct *tsk)
200 utime = get_paca()->user_time;
201 get_paca()->user_time = 0;
202 account_user_time(tsk, utime);
205 static void account_process_time(struct pt_regs *regs)
207 int cpu = smp_processor_id();
209 account_process_vtime(current);
211 if (rcu_pending(cpu))
212 rcu_check_callbacks(cpu, user_mode(regs));
214 run_posix_cpu_timers(current);
218 * Stuff for accounting stolen time.
220 struct cpu_purr_data {
221 int initialized; /* thread is running */
222 u64 tb; /* last TB value read */
223 u64 purr; /* last PURR value read */
227 * Each entry in the cpu_purr_data array is manipulated only by its
228 * "owner" cpu -- usually in the timer interrupt but also occasionally
229 * in process context for cpu online. As long as cpus do not touch
230 * each others' cpu_purr_data, disabling local interrupts is
231 * sufficient to serialize accesses.
233 static DEFINE_PER_CPU(struct cpu_purr_data, cpu_purr_data);
235 static void snapshot_tb_and_purr(void *data)
238 struct cpu_purr_data *p = &__get_cpu_var(cpu_purr_data);
240 local_irq_save(flags);
241 p->tb = get_tb_or_rtc();
242 p->purr = mfspr(SPRN_PURR);
245 local_irq_restore(flags);
249 * Called during boot when all cpus have come up.
251 void snapshot_timebases(void)
253 if (!cpu_has_feature(CPU_FTR_PURR))
255 on_each_cpu(snapshot_tb_and_purr, NULL, 0, 1);
259 * Must be called with interrupts disabled.
261 void calculate_steal_time(void)
265 struct cpu_purr_data *pme;
267 if (!cpu_has_feature(CPU_FTR_PURR))
269 pme = &per_cpu(cpu_purr_data, smp_processor_id());
270 if (!pme->initialized)
271 return; /* this can happen in early boot */
273 purr = mfspr(SPRN_PURR);
274 stolen = (tb - pme->tb) - (purr - pme->purr);
276 account_steal_time(current, stolen);
281 #ifdef CONFIG_PPC_SPLPAR
283 * Must be called before the cpu is added to the online map when
284 * a cpu is being brought up at runtime.
286 static void snapshot_purr(void)
288 struct cpu_purr_data *pme;
291 if (!cpu_has_feature(CPU_FTR_PURR))
293 local_irq_save(flags);
294 pme = &per_cpu(cpu_purr_data, smp_processor_id());
296 pme->purr = mfspr(SPRN_PURR);
297 pme->initialized = 1;
298 local_irq_restore(flags);
301 #endif /* CONFIG_PPC_SPLPAR */
303 #else /* ! CONFIG_VIRT_CPU_ACCOUNTING */
304 #define calc_cputime_factors()
305 #define account_process_time(regs) update_process_times(user_mode(regs))
306 #define calculate_steal_time() do { } while (0)
309 #if !(defined(CONFIG_VIRT_CPU_ACCOUNTING) && defined(CONFIG_PPC_SPLPAR))
310 #define snapshot_purr() do { } while (0)
314 * Called when a cpu comes up after the system has finished booting,
315 * i.e. as a result of a hotplug cpu action.
317 void snapshot_timebase(void)
319 __get_cpu_var(last_jiffy) = get_tb_or_rtc();
323 void __delay(unsigned long loops)
331 /* the RTCL register wraps at 1000000000 */
332 diff = get_rtcl() - start;
335 } while (diff < loops);
338 while (get_tbl() - start < loops)
343 EXPORT_SYMBOL(__delay);
345 void udelay(unsigned long usecs)
347 __delay(tb_ticks_per_usec * usecs);
349 EXPORT_SYMBOL(udelay);
351 static __inline__ void timer_check_rtc(void)
354 * update the rtc when needed, this should be performed on the
355 * right fraction of a second. Half or full second ?
356 * Full second works on mk48t59 clocks, others need testing.
357 * Note that this update is basically only used through
358 * the adjtimex system calls. Setting the HW clock in
359 * any other way is a /dev/rtc and userland business.
360 * This is still wrong by -0.5/+1.5 jiffies because of the
361 * timer interrupt resolution and possible delay, but here we
362 * hit a quantization limit which can only be solved by higher
363 * resolution timers and decoupling time management from timer
364 * interrupts. This is also wrong on the clocks
365 * which require being written at the half second boundary.
366 * We should have an rtc call that only sets the minutes and
367 * seconds like on Intel to avoid problems with non UTC clocks.
369 if (ppc_md.set_rtc_time && ntp_synced() &&
370 xtime.tv_sec - last_rtc_update >= 659 &&
371 abs((xtime.tv_nsec/1000) - (1000000-1000000/HZ)) < 500000/HZ) {
373 to_tm(xtime.tv_sec + 1 + timezone_offset, &tm);
376 if (ppc_md.set_rtc_time(&tm) == 0)
377 last_rtc_update = xtime.tv_sec + 1;
379 /* Try again one minute later */
380 last_rtc_update += 60;
385 * This version of gettimeofday has microsecond resolution.
387 static inline void __do_gettimeofday(struct timeval *tv)
389 unsigned long sec, usec;
391 struct gettimeofday_vars *temp_varp;
392 u64 temp_tb_to_xs, temp_stamp_xsec;
395 * These calculations are faster (gets rid of divides)
396 * if done in units of 1/2^20 rather than microseconds.
397 * The conversion to microseconds at the end is done
398 * without a divide (and in fact, without a multiply)
400 temp_varp = do_gtod.varp;
402 /* Sampling the time base must be done after loading
403 * do_gtod.varp in order to avoid racing with update_gtod.
405 data_barrier(temp_varp);
406 tb_ticks = get_tb() - temp_varp->tb_orig_stamp;
407 temp_tb_to_xs = temp_varp->tb_to_xs;
408 temp_stamp_xsec = temp_varp->stamp_xsec;
409 xsec = temp_stamp_xsec + mulhdu(tb_ticks, temp_tb_to_xs);
410 sec = xsec / XSEC_PER_SEC;
411 usec = (unsigned long)xsec & (XSEC_PER_SEC - 1);
412 usec = SCALE_XSEC(usec, 1000000);
418 void do_gettimeofday(struct timeval *tv)
421 /* do this the old way */
422 unsigned long flags, seq;
423 unsigned int sec, nsec, usec;
426 seq = read_seqbegin_irqsave(&xtime_lock, flags);
428 nsec = xtime.tv_nsec + tb_ticks_since(tb_last_jiffy);
429 } while (read_seqretry_irqrestore(&xtime_lock, seq, flags));
431 while (usec >= 1000000) {
439 __do_gettimeofday(tv);
442 EXPORT_SYMBOL(do_gettimeofday);
445 * There are two copies of tb_to_xs and stamp_xsec so that no
446 * lock is needed to access and use these values in
447 * do_gettimeofday. We alternate the copies and as long as a
448 * reasonable time elapses between changes, there will never
449 * be inconsistent values. ntpd has a minimum of one minute
452 static inline void update_gtod(u64 new_tb_stamp, u64 new_stamp_xsec,
456 struct gettimeofday_vars *temp_varp;
458 temp_idx = (do_gtod.var_idx == 0);
459 temp_varp = &do_gtod.vars[temp_idx];
461 temp_varp->tb_to_xs = new_tb_to_xs;
462 temp_varp->tb_orig_stamp = new_tb_stamp;
463 temp_varp->stamp_xsec = new_stamp_xsec;
465 do_gtod.varp = temp_varp;
466 do_gtod.var_idx = temp_idx;
469 * tb_update_count is used to allow the userspace gettimeofday code
470 * to assure itself that it sees a consistent view of the tb_to_xs and
471 * stamp_xsec variables. It reads the tb_update_count, then reads
472 * tb_to_xs and stamp_xsec and then reads tb_update_count again. If
473 * the two values of tb_update_count match and are even then the
474 * tb_to_xs and stamp_xsec values are consistent. If not, then it
475 * loops back and reads them again until this criteria is met.
476 * We expect the caller to have done the first increment of
477 * vdso_data->tb_update_count already.
479 vdso_data->tb_orig_stamp = new_tb_stamp;
480 vdso_data->stamp_xsec = new_stamp_xsec;
481 vdso_data->tb_to_xs = new_tb_to_xs;
482 vdso_data->wtom_clock_sec = wall_to_monotonic.tv_sec;
483 vdso_data->wtom_clock_nsec = wall_to_monotonic.tv_nsec;
485 ++(vdso_data->tb_update_count);
489 * When the timebase - tb_orig_stamp gets too big, we do a manipulation
490 * between tb_orig_stamp and stamp_xsec. The goal here is to keep the
491 * difference tb - tb_orig_stamp small enough to always fit inside a
492 * 32 bits number. This is a requirement of our fast 32 bits userland
493 * implementation in the vdso. If we "miss" a call to this function
494 * (interrupt latency, CPU locked in a spinlock, ...) and we end up
495 * with a too big difference, then the vdso will fallback to calling
498 static __inline__ void timer_recalc_offset(u64 cur_tb)
500 unsigned long offset;
503 u64 tb, xsec_old, xsec_new;
504 struct gettimeofday_vars *varp;
508 tlen = current_tick_length();
509 offset = cur_tb - do_gtod.varp->tb_orig_stamp;
510 if (tlen == last_tick_len && offset < 0x80000000u)
512 if (tlen != last_tick_len) {
513 t2x = mulhdu(tlen << TICKLEN_SHIFT, ticklen_to_xs);
514 last_tick_len = tlen;
516 t2x = do_gtod.varp->tb_to_xs;
517 new_stamp_xsec = (u64) xtime.tv_nsec * XSEC_PER_SEC;
518 do_div(new_stamp_xsec, 1000000000);
519 new_stamp_xsec += (u64) xtime.tv_sec * XSEC_PER_SEC;
521 ++vdso_data->tb_update_count;
525 * Make sure time doesn't go backwards for userspace gettimeofday.
529 xsec_old = mulhdu(tb - varp->tb_orig_stamp, varp->tb_to_xs)
531 xsec_new = mulhdu(tb - cur_tb, t2x) + new_stamp_xsec;
532 if (xsec_new < xsec_old)
533 new_stamp_xsec += xsec_old - xsec_new;
535 update_gtod(cur_tb, new_stamp_xsec, t2x);
539 unsigned long profile_pc(struct pt_regs *regs)
541 unsigned long pc = instruction_pointer(regs);
543 if (in_lock_functions(pc))
548 EXPORT_SYMBOL(profile_pc);
551 #ifdef CONFIG_PPC_ISERIES
554 * This function recalibrates the timebase based on the 49-bit time-of-day
555 * value in the Titan chip. The Titan is much more accurate than the value
556 * returned by the service processor for the timebase frequency.
559 static int __init iSeries_tb_recal(void)
561 struct div_result divres;
562 unsigned long titan, tb;
564 /* Make sure we only run on iSeries */
565 if (!firmware_has_feature(FW_FEATURE_ISERIES))
569 titan = HvCallXm_loadTod();
570 if ( iSeries_recal_titan ) {
571 unsigned long tb_ticks = tb - iSeries_recal_tb;
572 unsigned long titan_usec = (titan - iSeries_recal_titan) >> 12;
573 unsigned long new_tb_ticks_per_sec = (tb_ticks * USEC_PER_SEC)/titan_usec;
574 unsigned long new_tb_ticks_per_jiffy = (new_tb_ticks_per_sec+(HZ/2))/HZ;
575 long tick_diff = new_tb_ticks_per_jiffy - tb_ticks_per_jiffy;
577 /* make sure tb_ticks_per_sec and tb_ticks_per_jiffy are consistent */
578 new_tb_ticks_per_sec = new_tb_ticks_per_jiffy * HZ;
580 if ( tick_diff < 0 ) {
581 tick_diff = -tick_diff;
585 if ( tick_diff < tb_ticks_per_jiffy/25 ) {
586 printk( "Titan recalibrate: new tb_ticks_per_jiffy = %lu (%c%ld)\n",
587 new_tb_ticks_per_jiffy, sign, tick_diff );
588 tb_ticks_per_jiffy = new_tb_ticks_per_jiffy;
589 tb_ticks_per_sec = new_tb_ticks_per_sec;
590 calc_cputime_factors();
591 div128_by_32( XSEC_PER_SEC, 0, tb_ticks_per_sec, &divres );
592 do_gtod.tb_ticks_per_sec = tb_ticks_per_sec;
593 tb_to_xs = divres.result_low;
594 do_gtod.varp->tb_to_xs = tb_to_xs;
595 vdso_data->tb_ticks_per_sec = tb_ticks_per_sec;
596 vdso_data->tb_to_xs = tb_to_xs;
599 printk( "Titan recalibrate: FAILED (difference > 4 percent)\n"
600 " new tb_ticks_per_jiffy = %lu\n"
601 " old tb_ticks_per_jiffy = %lu\n",
602 new_tb_ticks_per_jiffy, tb_ticks_per_jiffy );
606 iSeries_recal_titan = titan;
607 iSeries_recal_tb = tb;
611 late_initcall(iSeries_tb_recal);
613 /* Called from platform early init */
614 void __init iSeries_time_init_early(void)
616 iSeries_recal_tb = get_tb();
617 iSeries_recal_titan = HvCallXm_loadTod();
619 #endif /* CONFIG_PPC_ISERIES */
622 * For iSeries shared processors, we have to let the hypervisor
623 * set the hardware decrementer. We set a virtual decrementer
624 * in the lppaca and call the hypervisor if the virtual
625 * decrementer is less than the current value in the hardware
626 * decrementer. (almost always the new decrementer value will
627 * be greater than the current hardware decementer so the hypervisor
628 * call will not be needed)
632 * timer_interrupt - gets called when the decrementer overflows,
633 * with interrupts disabled.
635 void timer_interrupt(struct pt_regs * regs)
637 struct pt_regs *old_regs;
639 int cpu = smp_processor_id();
644 if (atomic_read(&ppc_n_lost_interrupts) != 0)
648 old_regs = set_irq_regs(regs);
651 profile_tick(CPU_PROFILING);
652 calculate_steal_time();
654 #ifdef CONFIG_PPC_ISERIES
655 if (firmware_has_feature(FW_FEATURE_ISERIES))
656 get_lppaca()->int_dword.fields.decr_int = 0;
659 while ((ticks = tb_ticks_since(per_cpu(last_jiffy, cpu)))
660 >= tb_ticks_per_jiffy) {
661 /* Update last_jiffy */
662 per_cpu(last_jiffy, cpu) += tb_ticks_per_jiffy;
663 /* Handle RTCL overflow on 601 */
664 if (__USE_RTC() && per_cpu(last_jiffy, cpu) >= 1000000000)
665 per_cpu(last_jiffy, cpu) -= 1000000000;
668 * We cannot disable the decrementer, so in the period
669 * between this cpu's being marked offline in cpu_online_map
670 * and calling stop-self, it is taking timer interrupts.
671 * Avoid calling into the scheduler rebalancing code if this
674 if (!cpu_is_offline(cpu))
675 account_process_time(regs);
678 * No need to check whether cpu is offline here; boot_cpuid
679 * should have been fixed up by now.
681 if (cpu != boot_cpuid)
684 write_seqlock(&xtime_lock);
685 tb_next_jiffy = tb_last_jiffy + tb_ticks_per_jiffy;
686 if (__USE_RTC() && tb_next_jiffy >= 1000000000)
687 tb_next_jiffy -= 1000000000;
688 if (per_cpu(last_jiffy, cpu) >= tb_next_jiffy) {
689 tb_last_jiffy = tb_next_jiffy;
691 timer_recalc_offset(tb_last_jiffy);
694 write_sequnlock(&xtime_lock);
697 next_dec = tb_ticks_per_jiffy - ticks;
700 #ifdef CONFIG_PPC_ISERIES
701 if (firmware_has_feature(FW_FEATURE_ISERIES) && hvlpevent_is_pending())
702 process_hvlpevents();
706 /* collect purr register values often, for accurate calculations */
707 if (firmware_has_feature(FW_FEATURE_SPLPAR)) {
708 struct cpu_usage *cu = &__get_cpu_var(cpu_usage_array);
709 cu->current_tb = mfspr(SPRN_PURR);
714 set_irq_regs(old_regs);
717 void wakeup_decrementer(void)
722 * The timebase gets saved on sleep and restored on wakeup,
723 * so all we need to do is to reset the decrementer.
725 ticks = tb_ticks_since(__get_cpu_var(last_jiffy));
726 if (ticks < tb_ticks_per_jiffy)
727 ticks = tb_ticks_per_jiffy - ticks;
734 void __init smp_space_timers(unsigned int max_cpus)
737 u64 previous_tb = per_cpu(last_jiffy, boot_cpuid);
739 /* make sure tb > per_cpu(last_jiffy, cpu) for all cpus always */
740 previous_tb -= tb_ticks_per_jiffy;
742 for_each_possible_cpu(i) {
745 per_cpu(last_jiffy, i) = previous_tb;
751 * Scheduler clock - returns current time in nanosec units.
753 * Note: mulhdu(a, b) (multiply high double unsigned) returns
754 * the high 64 bits of a * b, i.e. (a * b) >> 64, where a and b
755 * are 64-bit unsigned numbers.
757 unsigned long long sched_clock(void)
761 return mulhdu(get_tb() - boot_tb, tb_to_ns_scale) << tb_to_ns_shift;
764 int do_settimeofday(struct timespec *tv)
766 time_t wtm_sec, new_sec = tv->tv_sec;
767 long wtm_nsec, new_nsec = tv->tv_nsec;
770 unsigned long tb_delta;
772 if ((unsigned long)tv->tv_nsec >= NSEC_PER_SEC)
775 write_seqlock_irqsave(&xtime_lock, flags);
778 * Updating the RTC is not the job of this code. If the time is
779 * stepped under NTP, the RTC will be updated after STA_UNSYNC
780 * is cleared. Tools like clock/hwclock either copy the RTC
781 * to the system time, in which case there is no point in writing
782 * to the RTC again, or write to the RTC but then they don't call
783 * settimeofday to perform this operation.
786 /* Make userspace gettimeofday spin until we're done. */
787 ++vdso_data->tb_update_count;
791 * Subtract off the number of nanoseconds since the
792 * beginning of the last tick.
794 tb_delta = tb_ticks_since(tb_last_jiffy);
795 tb_delta = mulhdu(tb_delta, do_gtod.varp->tb_to_xs); /* in xsec */
796 new_nsec -= SCALE_XSEC(tb_delta, 1000000000);
798 wtm_sec = wall_to_monotonic.tv_sec + (xtime.tv_sec - new_sec);
799 wtm_nsec = wall_to_monotonic.tv_nsec + (xtime.tv_nsec - new_nsec);
801 set_normalized_timespec(&xtime, new_sec, new_nsec);
802 set_normalized_timespec(&wall_to_monotonic, wtm_sec, wtm_nsec);
804 /* In case of a large backwards jump in time with NTP, we want the
805 * clock to be updated as soon as the PLL is again in lock.
807 last_rtc_update = new_sec - 658;
811 new_xsec = xtime.tv_nsec;
813 new_xsec *= XSEC_PER_SEC;
814 do_div(new_xsec, NSEC_PER_SEC);
816 new_xsec += (u64)xtime.tv_sec * XSEC_PER_SEC;
817 update_gtod(tb_last_jiffy, new_xsec, do_gtod.varp->tb_to_xs);
819 vdso_data->tz_minuteswest = sys_tz.tz_minuteswest;
820 vdso_data->tz_dsttime = sys_tz.tz_dsttime;
822 write_sequnlock_irqrestore(&xtime_lock, flags);
827 EXPORT_SYMBOL(do_settimeofday);
829 static int __init get_freq(char *name, int cells, unsigned long *val)
831 struct device_node *cpu;
832 const unsigned int *fp;
835 /* The cpu node should have timebase and clock frequency properties */
836 cpu = of_find_node_by_type(NULL, "cpu");
839 fp = of_get_property(cpu, name, NULL);
842 *val = of_read_ulong(fp, cells);
851 void __init generic_calibrate_decr(void)
853 ppc_tb_freq = DEFAULT_TB_FREQ; /* hardcoded default */
855 if (!get_freq("ibm,extended-timebase-frequency", 2, &ppc_tb_freq) &&
856 !get_freq("timebase-frequency", 1, &ppc_tb_freq)) {
858 printk(KERN_ERR "WARNING: Estimating decrementer frequency "
862 ppc_proc_freq = DEFAULT_PROC_FREQ; /* hardcoded default */
864 if (!get_freq("ibm,extended-clock-frequency", 2, &ppc_proc_freq) &&
865 !get_freq("clock-frequency", 1, &ppc_proc_freq)) {
867 printk(KERN_ERR "WARNING: Estimating processor frequency "
871 #if defined(CONFIG_BOOKE) || defined(CONFIG_40x)
872 /* Set the time base to zero */
876 /* Clear any pending timer interrupts */
877 mtspr(SPRN_TSR, TSR_ENW | TSR_WIS | TSR_DIS | TSR_FIS);
879 /* Enable decrementer interrupt */
880 mtspr(SPRN_TCR, TCR_DIE);
884 unsigned long get_boot_time(void)
888 if (ppc_md.get_boot_time)
889 return ppc_md.get_boot_time();
890 if (!ppc_md.get_rtc_time)
892 ppc_md.get_rtc_time(&tm);
893 return mktime(tm.tm_year+1900, tm.tm_mon+1, tm.tm_mday,
894 tm.tm_hour, tm.tm_min, tm.tm_sec);
897 /* This function is only called on the boot processor */
898 void __init time_init(void)
901 unsigned long tm = 0;
902 struct div_result res;
906 if (ppc_md.time_init != NULL)
907 timezone_offset = ppc_md.time_init();
910 /* 601 processor: dec counts down by 128 every 128ns */
911 ppc_tb_freq = 1000000000;
912 tb_last_jiffy = get_rtcl();
914 /* Normal PowerPC with timebase register */
915 ppc_md.calibrate_decr();
916 printk(KERN_DEBUG "time_init: decrementer frequency = %lu.%.6lu MHz\n",
917 ppc_tb_freq / 1000000, ppc_tb_freq % 1000000);
918 printk(KERN_DEBUG "time_init: processor frequency = %lu.%.6lu MHz\n",
919 ppc_proc_freq / 1000000, ppc_proc_freq % 1000000);
920 tb_last_jiffy = get_tb();
923 tb_ticks_per_jiffy = ppc_tb_freq / HZ;
924 tb_ticks_per_sec = ppc_tb_freq;
925 tb_ticks_per_usec = ppc_tb_freq / 1000000;
926 tb_to_us = mulhwu_scale_factor(ppc_tb_freq, 1000000);
927 calc_cputime_factors();
930 * Calculate the length of each tick in ns. It will not be
931 * exactly 1e9/HZ unless ppc_tb_freq is divisible by HZ.
932 * We compute 1e9 * tb_ticks_per_jiffy / ppc_tb_freq,
935 x = (u64) NSEC_PER_SEC * tb_ticks_per_jiffy + ppc_tb_freq - 1;
936 do_div(x, ppc_tb_freq);
938 last_tick_len = x << TICKLEN_SCALE;
941 * Compute ticklen_to_xs, which is a factor which gets multiplied
942 * by (last_tick_len << TICKLEN_SHIFT) to get a tb_to_xs value.
944 * ticklen_to_xs = 2^N / (tb_ticks_per_jiffy * 1e9)
945 * where N = 64 + 20 - TICKLEN_SCALE - TICKLEN_SHIFT
946 * which turns out to be N = 51 - SHIFT_HZ.
947 * This gives the result as a 0.64 fixed-point fraction.
948 * That value is reduced by an offset amounting to 1 xsec per
949 * 2^31 timebase ticks to avoid problems with time going backwards
950 * by 1 xsec when we do timer_recalc_offset due to losing the
951 * fractional xsec. That offset is equal to ppc_tb_freq/2^51
952 * since there are 2^20 xsec in a second.
954 div128_by_32((1ULL << 51) - ppc_tb_freq, 0,
955 tb_ticks_per_jiffy << SHIFT_HZ, &res);
956 div128_by_32(res.result_high, res.result_low, NSEC_PER_SEC, &res);
957 ticklen_to_xs = res.result_low;
959 /* Compute tb_to_xs from tick_nsec */
960 tb_to_xs = mulhdu(last_tick_len << TICKLEN_SHIFT, ticklen_to_xs);
963 * Compute scale factor for sched_clock.
964 * The calibrate_decr() function has set tb_ticks_per_sec,
965 * which is the timebase frequency.
966 * We compute 1e9 * 2^64 / tb_ticks_per_sec and interpret
967 * the 128-bit result as a 64.64 fixed-point number.
968 * We then shift that number right until it is less than 1.0,
969 * giving us the scale factor and shift count to use in
972 div128_by_32(1000000000, 0, tb_ticks_per_sec, &res);
973 scale = res.result_low;
974 for (shift = 0; res.result_high != 0; ++shift) {
975 scale = (scale >> 1) | (res.result_high << 63);
976 res.result_high >>= 1;
978 tb_to_ns_scale = scale;
979 tb_to_ns_shift = shift;
980 /* Save the current timebase to pretty up CONFIG_PRINTK_TIME */
981 boot_tb = get_tb_or_rtc();
983 tm = get_boot_time();
985 write_seqlock_irqsave(&xtime_lock, flags);
987 /* If platform provided a timezone (pmac), we correct the time */
988 if (timezone_offset) {
989 sys_tz.tz_minuteswest = -timezone_offset / 60;
990 sys_tz.tz_dsttime = 0;
991 tm -= timezone_offset;
996 do_gtod.varp = &do_gtod.vars[0];
998 do_gtod.varp->tb_orig_stamp = tb_last_jiffy;
999 __get_cpu_var(last_jiffy) = tb_last_jiffy;
1000 do_gtod.varp->stamp_xsec = (u64) xtime.tv_sec * XSEC_PER_SEC;
1001 do_gtod.tb_ticks_per_sec = tb_ticks_per_sec;
1002 do_gtod.varp->tb_to_xs = tb_to_xs;
1003 do_gtod.tb_to_us = tb_to_us;
1005 vdso_data->tb_orig_stamp = tb_last_jiffy;
1006 vdso_data->tb_update_count = 0;
1007 vdso_data->tb_ticks_per_sec = tb_ticks_per_sec;
1008 vdso_data->stamp_xsec = (u64) xtime.tv_sec * XSEC_PER_SEC;
1009 vdso_data->tb_to_xs = tb_to_xs;
1013 last_rtc_update = xtime.tv_sec;
1014 set_normalized_timespec(&wall_to_monotonic,
1015 -xtime.tv_sec, -xtime.tv_nsec);
1016 write_sequnlock_irqrestore(&xtime_lock, flags);
1018 /* Not exact, but the timer interrupt takes care of this */
1019 set_dec(tb_ticks_per_jiffy);
1024 #define STARTOFTIME 1970
1025 #define SECDAY 86400L
1026 #define SECYR (SECDAY * 365)
1027 #define leapyear(year) ((year) % 4 == 0 && \
1028 ((year) % 100 != 0 || (year) % 400 == 0))
1029 #define days_in_year(a) (leapyear(a) ? 366 : 365)
1030 #define days_in_month(a) (month_days[(a) - 1])
1032 static int month_days[12] = {
1033 31, 28, 31, 30, 31, 30, 31, 31, 30, 31, 30, 31
1037 * This only works for the Gregorian calendar - i.e. after 1752 (in the UK)
1039 void GregorianDay(struct rtc_time * tm)
1044 int MonthOffset[] = { 0, 31, 59, 90, 120, 151, 181, 212, 243, 273, 304, 334 };
1046 lastYear = tm->tm_year - 1;
1049 * Number of leap corrections to apply up to end of last year
1051 leapsToDate = lastYear / 4 - lastYear / 100 + lastYear / 400;
1054 * This year is a leap year if it is divisible by 4 except when it is
1055 * divisible by 100 unless it is divisible by 400
1057 * e.g. 1904 was a leap year, 1900 was not, 1996 is, and 2000 was
1059 day = tm->tm_mon > 2 && leapyear(tm->tm_year);
1061 day += lastYear*365 + leapsToDate + MonthOffset[tm->tm_mon-1] +
1064 tm->tm_wday = day % 7;
1067 void to_tm(int tim, struct rtc_time * tm)
1070 register long hms, day;
1075 /* Hours, minutes, seconds are easy */
1076 tm->tm_hour = hms / 3600;
1077 tm->tm_min = (hms % 3600) / 60;
1078 tm->tm_sec = (hms % 3600) % 60;
1080 /* Number of years in days */
1081 for (i = STARTOFTIME; day >= days_in_year(i); i++)
1082 day -= days_in_year(i);
1085 /* Number of months in days left */
1086 if (leapyear(tm->tm_year))
1087 days_in_month(FEBRUARY) = 29;
1088 for (i = 1; day >= days_in_month(i); i++)
1089 day -= days_in_month(i);
1090 days_in_month(FEBRUARY) = 28;
1093 /* Days are what is left over (+1) from all that. */
1094 tm->tm_mday = day + 1;
1097 * Determine the day of week
1102 /* Auxiliary function to compute scaling factors */
1103 /* Actually the choice of a timebase running at 1/4 the of the bus
1104 * frequency giving resolution of a few tens of nanoseconds is quite nice.
1105 * It makes this computation very precise (27-28 bits typically) which
1106 * is optimistic considering the stability of most processor clock
1107 * oscillators and the precision with which the timebase frequency
1108 * is measured but does not harm.
1110 unsigned mulhwu_scale_factor(unsigned inscale, unsigned outscale)
1112 unsigned mlt=0, tmp, err;
1113 /* No concern for performance, it's done once: use a stupid
1114 * but safe and compact method to find the multiplier.
1117 for (tmp = 1U<<31; tmp != 0; tmp >>= 1) {
1118 if (mulhwu(inscale, mlt|tmp) < outscale)
1122 /* We might still be off by 1 for the best approximation.
1123 * A side effect of this is that if outscale is too large
1124 * the returned value will be zero.
1125 * Many corner cases have been checked and seem to work,
1126 * some might have been forgotten in the test however.
1129 err = inscale * (mlt+1);
1130 if (err <= inscale/2)
1136 * Divide a 128-bit dividend by a 32-bit divisor, leaving a 128 bit
1139 void div128_by_32(u64 dividend_high, u64 dividend_low,
1140 unsigned divisor, struct div_result *dr)
1142 unsigned long a, b, c, d;
1143 unsigned long w, x, y, z;
1146 a = dividend_high >> 32;
1147 b = dividend_high & 0xffffffff;
1148 c = dividend_low >> 32;
1149 d = dividend_low & 0xffffffff;
1152 ra = ((u64)(a - (w * divisor)) << 32) + b;
1154 rb = ((u64) do_div(ra, divisor) << 32) + c;
1157 rc = ((u64) do_div(rb, divisor) << 32) + d;
1160 do_div(rc, divisor);
1163 dr->result_high = ((u64)w << 32) + x;
1164 dr->result_low = ((u64)y << 32) + z;