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>
77 /* keep track of when we need to update the rtc */
78 time_t last_rtc_update
;
79 #ifdef CONFIG_PPC_ISERIES
80 unsigned long iSeries_recal_titan
= 0;
81 unsigned long iSeries_recal_tb
= 0;
82 static unsigned long first_settimeofday
= 1;
85 /* The decrementer counts down by 128 every 128ns on a 601. */
86 #define DECREMENTER_COUNT_601 (1000000000 / HZ)
88 #define XSEC_PER_SEC (1024*1024)
91 #define SCALE_XSEC(xsec, max) (((xsec) * max) / XSEC_PER_SEC)
93 /* compute ((xsec << 12) * max) >> 32 */
94 #define SCALE_XSEC(xsec, max) mulhwu((xsec) << 12, max)
97 unsigned long tb_ticks_per_jiffy
;
98 unsigned long tb_ticks_per_usec
= 100; /* sane default */
99 EXPORT_SYMBOL(tb_ticks_per_usec
);
100 unsigned long tb_ticks_per_sec
;
101 EXPORT_SYMBOL(tb_ticks_per_sec
); /* for cputime_t conversions */
105 #define TICKLEN_SCALE TICK_LENGTH_SHIFT
106 u64 last_tick_len
; /* units are ns / 2^TICKLEN_SCALE */
107 u64 ticklen_to_xs
; /* 0.64 fraction */
109 /* If last_tick_len corresponds to about 1/HZ seconds, then
110 last_tick_len << TICKLEN_SHIFT will be about 2^63. */
111 #define TICKLEN_SHIFT (63 - 30 - TICKLEN_SCALE + SHIFT_HZ)
113 DEFINE_SPINLOCK(rtc_lock
);
114 EXPORT_SYMBOL_GPL(rtc_lock
);
117 unsigned tb_to_ns_shift
;
119 struct gettimeofday_struct do_gtod
;
121 extern struct timezone sys_tz
;
122 static long timezone_offset
;
124 unsigned long 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
);
217 #ifdef CONFIG_PPC_SPLPAR
219 * Stuff for accounting stolen time.
221 struct cpu_purr_data
{
222 int initialized
; /* thread is running */
223 u64 tb
; /* last TB value read */
224 u64 purr
; /* last PURR value read */
228 static DEFINE_PER_CPU(struct cpu_purr_data
, cpu_purr_data
);
230 static void snapshot_tb_and_purr(void *data
)
232 struct cpu_purr_data
*p
= &__get_cpu_var(cpu_purr_data
);
235 p
->purr
= mfspr(SPRN_PURR
);
241 * Called during boot when all cpus have come up.
243 void snapshot_timebases(void)
247 if (!cpu_has_feature(CPU_FTR_PURR
))
249 for_each_possible_cpu(cpu
)
250 spin_lock_init(&per_cpu(cpu_purr_data
, cpu
).lock
);
251 on_each_cpu(snapshot_tb_and_purr
, NULL
, 0, 1);
254 void calculate_steal_time(void)
258 struct cpu_purr_data
*pme
;
260 if (!cpu_has_feature(CPU_FTR_PURR
))
262 pme
= &per_cpu(cpu_purr_data
, smp_processor_id());
263 if (!pme
->initialized
)
264 return; /* this can happen in early boot */
265 spin_lock(&pme
->lock
);
267 purr
= mfspr(SPRN_PURR
);
268 stolen
= (tb
- pme
->tb
) - (purr
- pme
->purr
);
270 account_steal_time(current
, stolen
);
273 spin_unlock(&pme
->lock
);
277 * Must be called before the cpu is added to the online map when
278 * a cpu is being brought up at runtime.
280 static void snapshot_purr(void)
282 struct cpu_purr_data
*pme
;
285 if (!cpu_has_feature(CPU_FTR_PURR
))
287 pme
= &per_cpu(cpu_purr_data
, smp_processor_id());
288 spin_lock_irqsave(&pme
->lock
, flags
);
290 pme
->purr
= mfspr(SPRN_PURR
);
291 pme
->initialized
= 1;
292 spin_unlock_irqrestore(&pme
->lock
, flags
);
295 #endif /* CONFIG_PPC_SPLPAR */
297 #else /* ! CONFIG_VIRT_CPU_ACCOUNTING */
298 #define calc_cputime_factors()
299 #define account_process_time(regs) update_process_times(user_mode(regs))
300 #define calculate_steal_time() do { } while (0)
303 #if !(defined(CONFIG_VIRT_CPU_ACCOUNTING) && defined(CONFIG_PPC_SPLPAR))
304 #define snapshot_purr() do { } while (0)
308 * Called when a cpu comes up after the system has finished booting,
309 * i.e. as a result of a hotplug cpu action.
311 void snapshot_timebase(void)
313 __get_cpu_var(last_jiffy
) = get_tb();
317 void __delay(unsigned long loops
)
325 /* the RTCL register wraps at 1000000000 */
326 diff
= get_rtcl() - start
;
329 } while (diff
< loops
);
332 while (get_tbl() - start
< loops
)
337 EXPORT_SYMBOL(__delay
);
339 void udelay(unsigned long usecs
)
341 __delay(tb_ticks_per_usec
* usecs
);
343 EXPORT_SYMBOL(udelay
);
345 static __inline__
void timer_check_rtc(void)
348 * update the rtc when needed, this should be performed on the
349 * right fraction of a second. Half or full second ?
350 * Full second works on mk48t59 clocks, others need testing.
351 * Note that this update is basically only used through
352 * the adjtimex system calls. Setting the HW clock in
353 * any other way is a /dev/rtc and userland business.
354 * This is still wrong by -0.5/+1.5 jiffies because of the
355 * timer interrupt resolution and possible delay, but here we
356 * hit a quantization limit which can only be solved by higher
357 * resolution timers and decoupling time management from timer
358 * interrupts. This is also wrong on the clocks
359 * which require being written at the half second boundary.
360 * We should have an rtc call that only sets the minutes and
361 * seconds like on Intel to avoid problems with non UTC clocks.
363 if (ppc_md
.set_rtc_time
&& ntp_synced() &&
364 xtime
.tv_sec
- last_rtc_update
>= 659 &&
365 abs((xtime
.tv_nsec
/1000) - (1000000-1000000/HZ
)) < 500000/HZ
) {
367 to_tm(xtime
.tv_sec
+ 1 + timezone_offset
, &tm
);
370 if (ppc_md
.set_rtc_time(&tm
) == 0)
371 last_rtc_update
= xtime
.tv_sec
+ 1;
373 /* Try again one minute later */
374 last_rtc_update
+= 60;
379 * This version of gettimeofday has microsecond resolution.
381 static inline void __do_gettimeofday(struct timeval
*tv
)
383 unsigned long sec
, usec
;
385 struct gettimeofday_vars
*temp_varp
;
386 u64 temp_tb_to_xs
, temp_stamp_xsec
;
389 * These calculations are faster (gets rid of divides)
390 * if done in units of 1/2^20 rather than microseconds.
391 * The conversion to microseconds at the end is done
392 * without a divide (and in fact, without a multiply)
394 temp_varp
= do_gtod
.varp
;
396 /* Sampling the time base must be done after loading
397 * do_gtod.varp in order to avoid racing with update_gtod.
399 data_barrier(temp_varp
);
400 tb_ticks
= get_tb() - temp_varp
->tb_orig_stamp
;
401 temp_tb_to_xs
= temp_varp
->tb_to_xs
;
402 temp_stamp_xsec
= temp_varp
->stamp_xsec
;
403 xsec
= temp_stamp_xsec
+ mulhdu(tb_ticks
, temp_tb_to_xs
);
404 sec
= xsec
/ XSEC_PER_SEC
;
405 usec
= (unsigned long)xsec
& (XSEC_PER_SEC
- 1);
406 usec
= SCALE_XSEC(usec
, 1000000);
412 void do_gettimeofday(struct timeval
*tv
)
415 /* do this the old way */
416 unsigned long flags
, seq
;
417 unsigned int sec
, nsec
, usec
;
420 seq
= read_seqbegin_irqsave(&xtime_lock
, flags
);
422 nsec
= xtime
.tv_nsec
+ tb_ticks_since(tb_last_jiffy
);
423 } while (read_seqretry_irqrestore(&xtime_lock
, seq
, flags
));
425 while (usec
>= 1000000) {
433 __do_gettimeofday(tv
);
436 EXPORT_SYMBOL(do_gettimeofday
);
439 * There are two copies of tb_to_xs and stamp_xsec so that no
440 * lock is needed to access and use these values in
441 * do_gettimeofday. We alternate the copies and as long as a
442 * reasonable time elapses between changes, there will never
443 * be inconsistent values. ntpd has a minimum of one minute
446 static inline void update_gtod(u64 new_tb_stamp
, u64 new_stamp_xsec
,
450 struct gettimeofday_vars
*temp_varp
;
452 temp_idx
= (do_gtod
.var_idx
== 0);
453 temp_varp
= &do_gtod
.vars
[temp_idx
];
455 temp_varp
->tb_to_xs
= new_tb_to_xs
;
456 temp_varp
->tb_orig_stamp
= new_tb_stamp
;
457 temp_varp
->stamp_xsec
= new_stamp_xsec
;
459 do_gtod
.varp
= temp_varp
;
460 do_gtod
.var_idx
= temp_idx
;
463 * tb_update_count is used to allow the userspace gettimeofday code
464 * to assure itself that it sees a consistent view of the tb_to_xs and
465 * stamp_xsec variables. It reads the tb_update_count, then reads
466 * tb_to_xs and stamp_xsec and then reads tb_update_count again. If
467 * the two values of tb_update_count match and are even then the
468 * tb_to_xs and stamp_xsec values are consistent. If not, then it
469 * loops back and reads them again until this criteria is met.
470 * We expect the caller to have done the first increment of
471 * vdso_data->tb_update_count already.
473 vdso_data
->tb_orig_stamp
= new_tb_stamp
;
474 vdso_data
->stamp_xsec
= new_stamp_xsec
;
475 vdso_data
->tb_to_xs
= new_tb_to_xs
;
476 vdso_data
->wtom_clock_sec
= wall_to_monotonic
.tv_sec
;
477 vdso_data
->wtom_clock_nsec
= wall_to_monotonic
.tv_nsec
;
479 ++(vdso_data
->tb_update_count
);
483 * When the timebase - tb_orig_stamp gets too big, we do a manipulation
484 * between tb_orig_stamp and stamp_xsec. The goal here is to keep the
485 * difference tb - tb_orig_stamp small enough to always fit inside a
486 * 32 bits number. This is a requirement of our fast 32 bits userland
487 * implementation in the vdso. If we "miss" a call to this function
488 * (interrupt latency, CPU locked in a spinlock, ...) and we end up
489 * with a too big difference, then the vdso will fallback to calling
492 static __inline__
void timer_recalc_offset(u64 cur_tb
)
494 unsigned long offset
;
497 u64 tb
, xsec_old
, xsec_new
;
498 struct gettimeofday_vars
*varp
;
502 tlen
= current_tick_length();
503 offset
= cur_tb
- do_gtod
.varp
->tb_orig_stamp
;
504 if (tlen
== last_tick_len
&& offset
< 0x80000000u
)
506 if (tlen
!= last_tick_len
) {
507 t2x
= mulhdu(tlen
<< TICKLEN_SHIFT
, ticklen_to_xs
);
508 last_tick_len
= tlen
;
510 t2x
= do_gtod
.varp
->tb_to_xs
;
511 new_stamp_xsec
= (u64
) xtime
.tv_nsec
* XSEC_PER_SEC
;
512 do_div(new_stamp_xsec
, 1000000000);
513 new_stamp_xsec
+= (u64
) xtime
.tv_sec
* XSEC_PER_SEC
;
515 ++vdso_data
->tb_update_count
;
519 * Make sure time doesn't go backwards for userspace gettimeofday.
523 xsec_old
= mulhdu(tb
- varp
->tb_orig_stamp
, varp
->tb_to_xs
)
525 xsec_new
= mulhdu(tb
- cur_tb
, t2x
) + new_stamp_xsec
;
526 if (xsec_new
< xsec_old
)
527 new_stamp_xsec
+= xsec_old
- xsec_new
;
529 update_gtod(cur_tb
, new_stamp_xsec
, t2x
);
533 unsigned long profile_pc(struct pt_regs
*regs
)
535 unsigned long pc
= instruction_pointer(regs
);
537 if (in_lock_functions(pc
))
542 EXPORT_SYMBOL(profile_pc
);
545 #ifdef CONFIG_PPC_ISERIES
548 * This function recalibrates the timebase based on the 49-bit time-of-day
549 * value in the Titan chip. The Titan is much more accurate than the value
550 * returned by the service processor for the timebase frequency.
553 static void iSeries_tb_recal(void)
555 struct div_result divres
;
556 unsigned long titan
, tb
;
558 titan
= HvCallXm_loadTod();
559 if ( iSeries_recal_titan
) {
560 unsigned long tb_ticks
= tb
- iSeries_recal_tb
;
561 unsigned long titan_usec
= (titan
- iSeries_recal_titan
) >> 12;
562 unsigned long new_tb_ticks_per_sec
= (tb_ticks
* USEC_PER_SEC
)/titan_usec
;
563 unsigned long new_tb_ticks_per_jiffy
= (new_tb_ticks_per_sec
+(HZ
/2))/HZ
;
564 long tick_diff
= new_tb_ticks_per_jiffy
- tb_ticks_per_jiffy
;
566 /* make sure tb_ticks_per_sec and tb_ticks_per_jiffy are consistent */
567 new_tb_ticks_per_sec
= new_tb_ticks_per_jiffy
* HZ
;
569 if ( tick_diff
< 0 ) {
570 tick_diff
= -tick_diff
;
574 if ( tick_diff
< tb_ticks_per_jiffy
/25 ) {
575 printk( "Titan recalibrate: new tb_ticks_per_jiffy = %lu (%c%ld)\n",
576 new_tb_ticks_per_jiffy
, sign
, tick_diff
);
577 tb_ticks_per_jiffy
= new_tb_ticks_per_jiffy
;
578 tb_ticks_per_sec
= new_tb_ticks_per_sec
;
579 calc_cputime_factors();
580 div128_by_32( XSEC_PER_SEC
, 0, tb_ticks_per_sec
, &divres
);
581 do_gtod
.tb_ticks_per_sec
= tb_ticks_per_sec
;
582 tb_to_xs
= divres
.result_low
;
583 do_gtod
.varp
->tb_to_xs
= tb_to_xs
;
584 vdso_data
->tb_ticks_per_sec
= tb_ticks_per_sec
;
585 vdso_data
->tb_to_xs
= tb_to_xs
;
588 printk( "Titan recalibrate: FAILED (difference > 4 percent)\n"
589 " new tb_ticks_per_jiffy = %lu\n"
590 " old tb_ticks_per_jiffy = %lu\n",
591 new_tb_ticks_per_jiffy
, tb_ticks_per_jiffy
);
595 iSeries_recal_titan
= titan
;
596 iSeries_recal_tb
= tb
;
601 * For iSeries shared processors, we have to let the hypervisor
602 * set the hardware decrementer. We set a virtual decrementer
603 * in the lppaca and call the hypervisor if the virtual
604 * decrementer is less than the current value in the hardware
605 * decrementer. (almost always the new decrementer value will
606 * be greater than the current hardware decementer so the hypervisor
607 * call will not be needed)
611 * timer_interrupt - gets called when the decrementer overflows,
612 * with interrupts disabled.
614 void timer_interrupt(struct pt_regs
* regs
)
616 struct pt_regs
*old_regs
;
618 int cpu
= smp_processor_id();
623 if (atomic_read(&ppc_n_lost_interrupts
) != 0)
627 old_regs
= set_irq_regs(regs
);
630 profile_tick(CPU_PROFILING
);
631 calculate_steal_time();
633 #ifdef CONFIG_PPC_ISERIES
634 if (firmware_has_feature(FW_FEATURE_ISERIES
))
635 get_lppaca()->int_dword
.fields
.decr_int
= 0;
638 while ((ticks
= tb_ticks_since(per_cpu(last_jiffy
, cpu
)))
639 >= tb_ticks_per_jiffy
) {
640 /* Update last_jiffy */
641 per_cpu(last_jiffy
, cpu
) += tb_ticks_per_jiffy
;
642 /* Handle RTCL overflow on 601 */
643 if (__USE_RTC() && per_cpu(last_jiffy
, cpu
) >= 1000000000)
644 per_cpu(last_jiffy
, cpu
) -= 1000000000;
647 * We cannot disable the decrementer, so in the period
648 * between this cpu's being marked offline in cpu_online_map
649 * and calling stop-self, it is taking timer interrupts.
650 * Avoid calling into the scheduler rebalancing code if this
653 if (!cpu_is_offline(cpu
))
654 account_process_time(regs
);
657 * No need to check whether cpu is offline here; boot_cpuid
658 * should have been fixed up by now.
660 if (cpu
!= boot_cpuid
)
663 write_seqlock(&xtime_lock
);
664 tb_next_jiffy
= tb_last_jiffy
+ tb_ticks_per_jiffy
;
665 if (per_cpu(last_jiffy
, cpu
) >= tb_next_jiffy
) {
666 tb_last_jiffy
= tb_next_jiffy
;
668 timer_recalc_offset(tb_last_jiffy
);
671 write_sequnlock(&xtime_lock
);
674 next_dec
= tb_ticks_per_jiffy
- ticks
;
677 #ifdef CONFIG_PPC_ISERIES
678 if (firmware_has_feature(FW_FEATURE_ISERIES
) && hvlpevent_is_pending())
679 process_hvlpevents();
683 /* collect purr register values often, for accurate calculations */
684 if (firmware_has_feature(FW_FEATURE_SPLPAR
)) {
685 struct cpu_usage
*cu
= &__get_cpu_var(cpu_usage_array
);
686 cu
->current_tb
= mfspr(SPRN_PURR
);
691 set_irq_regs(old_regs
);
694 void wakeup_decrementer(void)
699 * The timebase gets saved on sleep and restored on wakeup,
700 * so all we need to do is to reset the decrementer.
702 ticks
= tb_ticks_since(__get_cpu_var(last_jiffy
));
703 if (ticks
< tb_ticks_per_jiffy
)
704 ticks
= tb_ticks_per_jiffy
- ticks
;
711 void __init
smp_space_timers(unsigned int max_cpus
)
714 unsigned long half
= tb_ticks_per_jiffy
/ 2;
715 unsigned long offset
= tb_ticks_per_jiffy
/ max_cpus
;
716 u64 previous_tb
= per_cpu(last_jiffy
, boot_cpuid
);
718 /* make sure tb > per_cpu(last_jiffy, cpu) for all cpus always */
719 previous_tb
-= tb_ticks_per_jiffy
;
721 * The stolen time calculation for POWER5 shared-processor LPAR
722 * systems works better if the two threads' timebase interrupts
723 * are staggered by half a jiffy with respect to each other.
725 for_each_possible_cpu(i
) {
728 if (i
== (boot_cpuid
^ 1))
729 per_cpu(last_jiffy
, i
) =
730 per_cpu(last_jiffy
, boot_cpuid
) - half
;
732 per_cpu(last_jiffy
, i
) =
733 per_cpu(last_jiffy
, i
^ 1) + half
;
735 previous_tb
+= offset
;
736 per_cpu(last_jiffy
, i
) = previous_tb
;
743 * Scheduler clock - returns current time in nanosec units.
745 * Note: mulhdu(a, b) (multiply high double unsigned) returns
746 * the high 64 bits of a * b, i.e. (a * b) >> 64, where a and b
747 * are 64-bit unsigned numbers.
749 unsigned long long sched_clock(void)
753 return mulhdu(get_tb(), tb_to_ns_scale
) << tb_to_ns_shift
;
756 int do_settimeofday(struct timespec
*tv
)
758 time_t wtm_sec
, new_sec
= tv
->tv_sec
;
759 long wtm_nsec
, new_nsec
= tv
->tv_nsec
;
762 unsigned long tb_delta
;
764 if ((unsigned long)tv
->tv_nsec
>= NSEC_PER_SEC
)
767 write_seqlock_irqsave(&xtime_lock
, flags
);
770 * Updating the RTC is not the job of this code. If the time is
771 * stepped under NTP, the RTC will be updated after STA_UNSYNC
772 * is cleared. Tools like clock/hwclock either copy the RTC
773 * to the system time, in which case there is no point in writing
774 * to the RTC again, or write to the RTC but then they don't call
775 * settimeofday to perform this operation.
777 #ifdef CONFIG_PPC_ISERIES
778 if (firmware_has_feature(FW_FEATURE_ISERIES
) && first_settimeofday
) {
780 first_settimeofday
= 0;
784 /* Make userspace gettimeofday spin until we're done. */
785 ++vdso_data
->tb_update_count
;
789 * Subtract off the number of nanoseconds since the
790 * beginning of the last tick.
792 tb_delta
= tb_ticks_since(tb_last_jiffy
);
793 tb_delta
= mulhdu(tb_delta
, do_gtod
.varp
->tb_to_xs
); /* in xsec */
794 new_nsec
-= SCALE_XSEC(tb_delta
, 1000000000);
796 wtm_sec
= wall_to_monotonic
.tv_sec
+ (xtime
.tv_sec
- new_sec
);
797 wtm_nsec
= wall_to_monotonic
.tv_nsec
+ (xtime
.tv_nsec
- new_nsec
);
799 set_normalized_timespec(&xtime
, new_sec
, new_nsec
);
800 set_normalized_timespec(&wall_to_monotonic
, wtm_sec
, wtm_nsec
);
802 /* In case of a large backwards jump in time with NTP, we want the
803 * clock to be updated as soon as the PLL is again in lock.
805 last_rtc_update
= new_sec
- 658;
809 new_xsec
= xtime
.tv_nsec
;
811 new_xsec
*= XSEC_PER_SEC
;
812 do_div(new_xsec
, NSEC_PER_SEC
);
814 new_xsec
+= (u64
)xtime
.tv_sec
* XSEC_PER_SEC
;
815 update_gtod(tb_last_jiffy
, new_xsec
, do_gtod
.varp
->tb_to_xs
);
817 vdso_data
->tz_minuteswest
= sys_tz
.tz_minuteswest
;
818 vdso_data
->tz_dsttime
= sys_tz
.tz_dsttime
;
820 write_sequnlock_irqrestore(&xtime_lock
, flags
);
825 EXPORT_SYMBOL(do_settimeofday
);
827 static int __init
get_freq(char *name
, int cells
, unsigned long *val
)
829 struct device_node
*cpu
;
830 const unsigned int *fp
;
833 /* The cpu node should have timebase and clock frequency properties */
834 cpu
= of_find_node_by_type(NULL
, "cpu");
837 fp
= of_get_property(cpu
, name
, NULL
);
840 *val
= of_read_ulong(fp
, cells
);
849 void __init
generic_calibrate_decr(void)
851 ppc_tb_freq
= DEFAULT_TB_FREQ
; /* hardcoded default */
853 if (!get_freq("ibm,extended-timebase-frequency", 2, &ppc_tb_freq
) &&
854 !get_freq("timebase-frequency", 1, &ppc_tb_freq
)) {
856 printk(KERN_ERR
"WARNING: Estimating decrementer frequency "
860 ppc_proc_freq
= DEFAULT_PROC_FREQ
; /* hardcoded default */
862 if (!get_freq("ibm,extended-clock-frequency", 2, &ppc_proc_freq
) &&
863 !get_freq("clock-frequency", 1, &ppc_proc_freq
)) {
865 printk(KERN_ERR
"WARNING: Estimating processor frequency "
870 /* Set the time base to zero */
874 /* Clear any pending timer interrupts */
875 mtspr(SPRN_TSR
, TSR_ENW
| TSR_WIS
| TSR_DIS
| TSR_FIS
);
877 /* Enable decrementer interrupt */
878 mtspr(SPRN_TCR
, TCR_DIE
);
882 unsigned long get_boot_time(void)
886 if (ppc_md
.get_boot_time
)
887 return ppc_md
.get_boot_time();
888 if (!ppc_md
.get_rtc_time
)
890 ppc_md
.get_rtc_time(&tm
);
891 return mktime(tm
.tm_year
+1900, tm
.tm_mon
+1, tm
.tm_mday
,
892 tm
.tm_hour
, tm
.tm_min
, tm
.tm_sec
);
895 /* This function is only called on the boot processor */
896 void __init
time_init(void)
899 unsigned long tm
= 0;
900 struct div_result res
;
904 if (ppc_md
.time_init
!= NULL
)
905 timezone_offset
= ppc_md
.time_init();
908 /* 601 processor: dec counts down by 128 every 128ns */
909 ppc_tb_freq
= 1000000000;
910 tb_last_jiffy
= get_rtcl();
912 /* Normal PowerPC with timebase register */
913 ppc_md
.calibrate_decr();
914 printk(KERN_DEBUG
"time_init: decrementer frequency = %lu.%.6lu MHz\n",
915 ppc_tb_freq
/ 1000000, ppc_tb_freq
% 1000000);
916 printk(KERN_DEBUG
"time_init: processor frequency = %lu.%.6lu MHz\n",
917 ppc_proc_freq
/ 1000000, ppc_proc_freq
% 1000000);
918 tb_last_jiffy
= get_tb();
921 tb_ticks_per_jiffy
= ppc_tb_freq
/ HZ
;
922 tb_ticks_per_sec
= ppc_tb_freq
;
923 tb_ticks_per_usec
= ppc_tb_freq
/ 1000000;
924 tb_to_us
= mulhwu_scale_factor(ppc_tb_freq
, 1000000);
925 calc_cputime_factors();
928 * Calculate the length of each tick in ns. It will not be
929 * exactly 1e9/HZ unless ppc_tb_freq is divisible by HZ.
930 * We compute 1e9 * tb_ticks_per_jiffy / ppc_tb_freq,
933 x
= (u64
) NSEC_PER_SEC
* tb_ticks_per_jiffy
+ ppc_tb_freq
- 1;
934 do_div(x
, ppc_tb_freq
);
936 last_tick_len
= x
<< TICKLEN_SCALE
;
939 * Compute ticklen_to_xs, which is a factor which gets multiplied
940 * by (last_tick_len << TICKLEN_SHIFT) to get a tb_to_xs value.
942 * ticklen_to_xs = 2^N / (tb_ticks_per_jiffy * 1e9)
943 * where N = 64 + 20 - TICKLEN_SCALE - TICKLEN_SHIFT
944 * which turns out to be N = 51 - SHIFT_HZ.
945 * This gives the result as a 0.64 fixed-point fraction.
946 * That value is reduced by an offset amounting to 1 xsec per
947 * 2^31 timebase ticks to avoid problems with time going backwards
948 * by 1 xsec when we do timer_recalc_offset due to losing the
949 * fractional xsec. That offset is equal to ppc_tb_freq/2^51
950 * since there are 2^20 xsec in a second.
952 div128_by_32((1ULL << 51) - ppc_tb_freq
, 0,
953 tb_ticks_per_jiffy
<< SHIFT_HZ
, &res
);
954 div128_by_32(res
.result_high
, res
.result_low
, NSEC_PER_SEC
, &res
);
955 ticklen_to_xs
= res
.result_low
;
957 /* Compute tb_to_xs from tick_nsec */
958 tb_to_xs
= mulhdu(last_tick_len
<< TICKLEN_SHIFT
, ticklen_to_xs
);
961 * Compute scale factor for sched_clock.
962 * The calibrate_decr() function has set tb_ticks_per_sec,
963 * which is the timebase frequency.
964 * We compute 1e9 * 2^64 / tb_ticks_per_sec and interpret
965 * the 128-bit result as a 64.64 fixed-point number.
966 * We then shift that number right until it is less than 1.0,
967 * giving us the scale factor and shift count to use in
970 div128_by_32(1000000000, 0, tb_ticks_per_sec
, &res
);
971 scale
= res
.result_low
;
972 for (shift
= 0; res
.result_high
!= 0; ++shift
) {
973 scale
= (scale
>> 1) | (res
.result_high
<< 63);
974 res
.result_high
>>= 1;
976 tb_to_ns_scale
= scale
;
977 tb_to_ns_shift
= shift
;
979 tm
= get_boot_time();
981 write_seqlock_irqsave(&xtime_lock
, flags
);
983 /* If platform provided a timezone (pmac), we correct the time */
984 if (timezone_offset
) {
985 sys_tz
.tz_minuteswest
= -timezone_offset
/ 60;
986 sys_tz
.tz_dsttime
= 0;
987 tm
-= timezone_offset
;
992 do_gtod
.varp
= &do_gtod
.vars
[0];
994 do_gtod
.varp
->tb_orig_stamp
= tb_last_jiffy
;
995 __get_cpu_var(last_jiffy
) = tb_last_jiffy
;
996 do_gtod
.varp
->stamp_xsec
= (u64
) xtime
.tv_sec
* XSEC_PER_SEC
;
997 do_gtod
.tb_ticks_per_sec
= tb_ticks_per_sec
;
998 do_gtod
.varp
->tb_to_xs
= tb_to_xs
;
999 do_gtod
.tb_to_us
= tb_to_us
;
1001 vdso_data
->tb_orig_stamp
= tb_last_jiffy
;
1002 vdso_data
->tb_update_count
= 0;
1003 vdso_data
->tb_ticks_per_sec
= tb_ticks_per_sec
;
1004 vdso_data
->stamp_xsec
= (u64
) xtime
.tv_sec
* XSEC_PER_SEC
;
1005 vdso_data
->tb_to_xs
= tb_to_xs
;
1009 last_rtc_update
= xtime
.tv_sec
;
1010 set_normalized_timespec(&wall_to_monotonic
,
1011 -xtime
.tv_sec
, -xtime
.tv_nsec
);
1012 write_sequnlock_irqrestore(&xtime_lock
, flags
);
1014 /* Not exact, but the timer interrupt takes care of this */
1015 set_dec(tb_ticks_per_jiffy
);
1020 #define STARTOFTIME 1970
1021 #define SECDAY 86400L
1022 #define SECYR (SECDAY * 365)
1023 #define leapyear(year) ((year) % 4 == 0 && \
1024 ((year) % 100 != 0 || (year) % 400 == 0))
1025 #define days_in_year(a) (leapyear(a) ? 366 : 365)
1026 #define days_in_month(a) (month_days[(a) - 1])
1028 static int month_days
[12] = {
1029 31, 28, 31, 30, 31, 30, 31, 31, 30, 31, 30, 31
1033 * This only works for the Gregorian calendar - i.e. after 1752 (in the UK)
1035 void GregorianDay(struct rtc_time
* tm
)
1040 int MonthOffset
[] = { 0, 31, 59, 90, 120, 151, 181, 212, 243, 273, 304, 334 };
1042 lastYear
= tm
->tm_year
- 1;
1045 * Number of leap corrections to apply up to end of last year
1047 leapsToDate
= lastYear
/ 4 - lastYear
/ 100 + lastYear
/ 400;
1050 * This year is a leap year if it is divisible by 4 except when it is
1051 * divisible by 100 unless it is divisible by 400
1053 * e.g. 1904 was a leap year, 1900 was not, 1996 is, and 2000 was
1055 day
= tm
->tm_mon
> 2 && leapyear(tm
->tm_year
);
1057 day
+= lastYear
*365 + leapsToDate
+ MonthOffset
[tm
->tm_mon
-1] +
1060 tm
->tm_wday
= day
% 7;
1063 void to_tm(int tim
, struct rtc_time
* tm
)
1066 register long hms
, day
;
1071 /* Hours, minutes, seconds are easy */
1072 tm
->tm_hour
= hms
/ 3600;
1073 tm
->tm_min
= (hms
% 3600) / 60;
1074 tm
->tm_sec
= (hms
% 3600) % 60;
1076 /* Number of years in days */
1077 for (i
= STARTOFTIME
; day
>= days_in_year(i
); i
++)
1078 day
-= days_in_year(i
);
1081 /* Number of months in days left */
1082 if (leapyear(tm
->tm_year
))
1083 days_in_month(FEBRUARY
) = 29;
1084 for (i
= 1; day
>= days_in_month(i
); i
++)
1085 day
-= days_in_month(i
);
1086 days_in_month(FEBRUARY
) = 28;
1089 /* Days are what is left over (+1) from all that. */
1090 tm
->tm_mday
= day
+ 1;
1093 * Determine the day of week
1098 /* Auxiliary function to compute scaling factors */
1099 /* Actually the choice of a timebase running at 1/4 the of the bus
1100 * frequency giving resolution of a few tens of nanoseconds is quite nice.
1101 * It makes this computation very precise (27-28 bits typically) which
1102 * is optimistic considering the stability of most processor clock
1103 * oscillators and the precision with which the timebase frequency
1104 * is measured but does not harm.
1106 unsigned mulhwu_scale_factor(unsigned inscale
, unsigned outscale
)
1108 unsigned mlt
=0, tmp
, err
;
1109 /* No concern for performance, it's done once: use a stupid
1110 * but safe and compact method to find the multiplier.
1113 for (tmp
= 1U<<31; tmp
!= 0; tmp
>>= 1) {
1114 if (mulhwu(inscale
, mlt
|tmp
) < outscale
)
1118 /* We might still be off by 1 for the best approximation.
1119 * A side effect of this is that if outscale is too large
1120 * the returned value will be zero.
1121 * Many corner cases have been checked and seem to work,
1122 * some might have been forgotten in the test however.
1125 err
= inscale
* (mlt
+1);
1126 if (err
<= inscale
/2)
1132 * Divide a 128-bit dividend by a 32-bit divisor, leaving a 128 bit
1135 void div128_by_32(u64 dividend_high
, u64 dividend_low
,
1136 unsigned divisor
, struct div_result
*dr
)
1138 unsigned long a
, b
, c
, d
;
1139 unsigned long w
, x
, y
, z
;
1142 a
= dividend_high
>> 32;
1143 b
= dividend_high
& 0xffffffff;
1144 c
= dividend_low
>> 32;
1145 d
= dividend_low
& 0xffffffff;
1148 ra
= ((u64
)(a
- (w
* divisor
)) << 32) + b
;
1150 rb
= ((u64
) do_div(ra
, divisor
) << 32) + c
;
1153 rc
= ((u64
) do_div(rb
, divisor
) << 32) + d
;
1156 do_div(rc
, divisor
);
1159 dr
->result_high
= ((u64
)w
<< 32) + x
;
1160 dr
->result_low
= ((u64
)y
<< 32) + z
;