[PATCH] kill wall_jiffies
[linux-2.6/cjktty.git] / arch / powerpc / kernel / time.c
blob8b278d85ca4ed24b44f8b93f33257f79f6d3294c
1 /*
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>
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>
55 #include <asm/io.h>
56 #include <asm/processor.h>
57 #include <asm/nvram.h>
58 #include <asm/cache.h>
59 #include <asm/machdep.h>
60 #include <asm/uaccess.h>
61 #include <asm/time.h>
62 #include <asm/prom.h>
63 #include <asm/irq.h>
64 #include <asm/div64.h>
65 #include <asm/smp.h>
66 #include <asm/vdso_datapage.h>
67 #ifdef CONFIG_PPC64
68 #include <asm/firmware.h>
69 #endif
70 #ifdef CONFIG_PPC_ISERIES
71 #include <asm/iseries/it_lp_queue.h>
72 #include <asm/iseries/hv_call_xm.h>
73 #endif
74 #include <asm/smp.h>
76 /* keep track of when we need to update the rtc */
77 time_t last_rtc_update;
78 #ifdef CONFIG_PPC_ISERIES
79 unsigned long iSeries_recal_titan = 0;
80 unsigned long iSeries_recal_tb = 0;
81 static unsigned long first_settimeofday = 1;
82 #endif
84 /* The decrementer counts down by 128 every 128ns on a 601. */
85 #define DECREMENTER_COUNT_601 (1000000000 / HZ)
87 #define XSEC_PER_SEC (1024*1024)
89 #ifdef CONFIG_PPC64
90 #define SCALE_XSEC(xsec, max) (((xsec) * max) / XSEC_PER_SEC)
91 #else
92 /* compute ((xsec << 12) * max) >> 32 */
93 #define SCALE_XSEC(xsec, max) mulhwu((xsec) << 12, max)
94 #endif
96 unsigned long tb_ticks_per_jiffy;
97 unsigned long tb_ticks_per_usec = 100; /* sane default */
98 EXPORT_SYMBOL(tb_ticks_per_usec);
99 unsigned long tb_ticks_per_sec;
100 EXPORT_SYMBOL(tb_ticks_per_sec); /* for cputime_t conversions */
101 u64 tb_to_xs;
102 unsigned tb_to_us;
104 #define TICKLEN_SCALE TICK_LENGTH_SHIFT
105 u64 last_tick_len; /* units are ns / 2^TICKLEN_SCALE */
106 u64 ticklen_to_xs; /* 0.64 fraction */
108 /* If last_tick_len corresponds to about 1/HZ seconds, then
109 last_tick_len << TICKLEN_SHIFT will be about 2^63. */
110 #define TICKLEN_SHIFT (63 - 30 - TICKLEN_SCALE + SHIFT_HZ)
112 DEFINE_SPINLOCK(rtc_lock);
113 EXPORT_SYMBOL_GPL(rtc_lock);
115 u64 tb_to_ns_scale;
116 unsigned tb_to_ns_shift;
118 struct gettimeofday_struct do_gtod;
120 extern struct timezone sys_tz;
121 static long timezone_offset;
123 unsigned long ppc_proc_freq;
124 unsigned long ppc_tb_freq;
126 static u64 tb_last_jiffy __cacheline_aligned_in_smp;
127 static DEFINE_PER_CPU(u64, last_jiffy);
129 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
131 * Factors for converting from cputime_t (timebase ticks) to
132 * jiffies, milliseconds, seconds, and clock_t (1/USER_HZ seconds).
133 * These are all stored as 0.64 fixed-point binary fractions.
135 u64 __cputime_jiffies_factor;
136 EXPORT_SYMBOL(__cputime_jiffies_factor);
137 u64 __cputime_msec_factor;
138 EXPORT_SYMBOL(__cputime_msec_factor);
139 u64 __cputime_sec_factor;
140 EXPORT_SYMBOL(__cputime_sec_factor);
141 u64 __cputime_clockt_factor;
142 EXPORT_SYMBOL(__cputime_clockt_factor);
144 static void calc_cputime_factors(void)
146 struct div_result res;
148 div128_by_32(HZ, 0, tb_ticks_per_sec, &res);
149 __cputime_jiffies_factor = res.result_low;
150 div128_by_32(1000, 0, tb_ticks_per_sec, &res);
151 __cputime_msec_factor = res.result_low;
152 div128_by_32(1, 0, tb_ticks_per_sec, &res);
153 __cputime_sec_factor = res.result_low;
154 div128_by_32(USER_HZ, 0, tb_ticks_per_sec, &res);
155 __cputime_clockt_factor = res.result_low;
159 * Read the PURR on systems that have it, otherwise the timebase.
161 static u64 read_purr(void)
163 if (cpu_has_feature(CPU_FTR_PURR))
164 return mfspr(SPRN_PURR);
165 return mftb();
169 * Account time for a transition between system, hard irq
170 * or soft irq state.
172 void account_system_vtime(struct task_struct *tsk)
174 u64 now, delta;
175 unsigned long flags;
177 local_irq_save(flags);
178 now = read_purr();
179 delta = now - get_paca()->startpurr;
180 get_paca()->startpurr = now;
181 if (!in_interrupt()) {
182 delta += get_paca()->system_time;
183 get_paca()->system_time = 0;
185 account_system_time(tsk, 0, delta);
186 local_irq_restore(flags);
190 * Transfer the user and system times accumulated in the paca
191 * by the exception entry and exit code to the generic process
192 * user and system time records.
193 * Must be called with interrupts disabled.
195 void account_process_vtime(struct task_struct *tsk)
197 cputime_t utime;
199 utime = get_paca()->user_time;
200 get_paca()->user_time = 0;
201 account_user_time(tsk, utime);
204 static void account_process_time(struct pt_regs *regs)
206 int cpu = smp_processor_id();
208 account_process_vtime(current);
209 run_local_timers();
210 if (rcu_pending(cpu))
211 rcu_check_callbacks(cpu, user_mode(regs));
212 scheduler_tick();
213 run_posix_cpu_timers(current);
216 #ifdef CONFIG_PPC_SPLPAR
218 * Stuff for accounting stolen time.
220 struct cpu_purr_data {
221 int initialized; /* thread is running */
222 u64 tb0; /* timebase at origin time */
223 u64 purr0; /* PURR at origin time */
224 u64 tb; /* last TB value read */
225 u64 purr; /* last PURR value read */
226 u64 stolen; /* stolen time so far */
227 spinlock_t lock;
230 static DEFINE_PER_CPU(struct cpu_purr_data, cpu_purr_data);
232 static void snapshot_tb_and_purr(void *data)
234 struct cpu_purr_data *p = &__get_cpu_var(cpu_purr_data);
236 p->tb0 = mftb();
237 p->purr0 = mfspr(SPRN_PURR);
238 p->tb = p->tb0;
239 p->purr = 0;
240 wmb();
241 p->initialized = 1;
245 * Called during boot when all cpus have come up.
247 void snapshot_timebases(void)
249 int cpu;
251 if (!cpu_has_feature(CPU_FTR_PURR))
252 return;
253 for_each_possible_cpu(cpu)
254 spin_lock_init(&per_cpu(cpu_purr_data, cpu).lock);
255 on_each_cpu(snapshot_tb_and_purr, NULL, 0, 1);
258 void calculate_steal_time(void)
260 u64 tb, purr, t0;
261 s64 stolen;
262 struct cpu_purr_data *p0, *pme, *phim;
263 int cpu;
265 if (!cpu_has_feature(CPU_FTR_PURR))
266 return;
267 cpu = smp_processor_id();
268 pme = &per_cpu(cpu_purr_data, cpu);
269 if (!pme->initialized)
270 return; /* this can happen in early boot */
271 p0 = &per_cpu(cpu_purr_data, cpu & ~1);
272 phim = &per_cpu(cpu_purr_data, cpu ^ 1);
273 spin_lock(&p0->lock);
274 tb = mftb();
275 purr = mfspr(SPRN_PURR) - pme->purr0;
276 if (!phim->initialized || !cpu_online(cpu ^ 1)) {
277 stolen = (tb - pme->tb) - (purr - pme->purr);
278 } else {
279 t0 = pme->tb0;
280 if (phim->tb0 < t0)
281 t0 = phim->tb0;
282 stolen = phim->tb - t0 - phim->purr - purr - p0->stolen;
284 if (stolen > 0) {
285 account_steal_time(current, stolen);
286 p0->stolen += stolen;
288 pme->tb = tb;
289 pme->purr = purr;
290 spin_unlock(&p0->lock);
294 * Must be called before the cpu is added to the online map when
295 * a cpu is being brought up at runtime.
297 static void snapshot_purr(void)
299 int cpu;
300 u64 purr;
301 struct cpu_purr_data *p0, *pme, *phim;
302 unsigned long flags;
304 if (!cpu_has_feature(CPU_FTR_PURR))
305 return;
306 cpu = smp_processor_id();
307 pme = &per_cpu(cpu_purr_data, cpu);
308 p0 = &per_cpu(cpu_purr_data, cpu & ~1);
309 phim = &per_cpu(cpu_purr_data, cpu ^ 1);
310 spin_lock_irqsave(&p0->lock, flags);
311 pme->tb = pme->tb0 = mftb();
312 purr = mfspr(SPRN_PURR);
313 if (!phim->initialized) {
314 pme->purr = 0;
315 pme->purr0 = purr;
316 } else {
317 /* set p->purr and p->purr0 for no change in p0->stolen */
318 pme->purr = phim->tb - phim->tb0 - phim->purr - p0->stolen;
319 pme->purr0 = purr - pme->purr;
321 pme->initialized = 1;
322 spin_unlock_irqrestore(&p0->lock, flags);
325 #endif /* CONFIG_PPC_SPLPAR */
327 #else /* ! CONFIG_VIRT_CPU_ACCOUNTING */
328 #define calc_cputime_factors()
329 #define account_process_time(regs) update_process_times(user_mode(regs))
330 #define calculate_steal_time() do { } while (0)
331 #endif
333 #if !(defined(CONFIG_VIRT_CPU_ACCOUNTING) && defined(CONFIG_PPC_SPLPAR))
334 #define snapshot_purr() do { } while (0)
335 #endif
338 * Called when a cpu comes up after the system has finished booting,
339 * i.e. as a result of a hotplug cpu action.
341 void snapshot_timebase(void)
343 __get_cpu_var(last_jiffy) = get_tb();
344 snapshot_purr();
347 void __delay(unsigned long loops)
349 unsigned long start;
350 int diff;
352 if (__USE_RTC()) {
353 start = get_rtcl();
354 do {
355 /* the RTCL register wraps at 1000000000 */
356 diff = get_rtcl() - start;
357 if (diff < 0)
358 diff += 1000000000;
359 } while (diff < loops);
360 } else {
361 start = get_tbl();
362 while (get_tbl() - start < loops)
363 HMT_low();
364 HMT_medium();
367 EXPORT_SYMBOL(__delay);
369 void udelay(unsigned long usecs)
371 __delay(tb_ticks_per_usec * usecs);
373 EXPORT_SYMBOL(udelay);
375 static __inline__ void timer_check_rtc(void)
378 * update the rtc when needed, this should be performed on the
379 * right fraction of a second. Half or full second ?
380 * Full second works on mk48t59 clocks, others need testing.
381 * Note that this update is basically only used through
382 * the adjtimex system calls. Setting the HW clock in
383 * any other way is a /dev/rtc and userland business.
384 * This is still wrong by -0.5/+1.5 jiffies because of the
385 * timer interrupt resolution and possible delay, but here we
386 * hit a quantization limit which can only be solved by higher
387 * resolution timers and decoupling time management from timer
388 * interrupts. This is also wrong on the clocks
389 * which require being written at the half second boundary.
390 * We should have an rtc call that only sets the minutes and
391 * seconds like on Intel to avoid problems with non UTC clocks.
393 if (ppc_md.set_rtc_time && ntp_synced() &&
394 xtime.tv_sec - last_rtc_update >= 659 &&
395 abs((xtime.tv_nsec/1000) - (1000000-1000000/HZ)) < 500000/HZ) {
396 struct rtc_time tm;
397 to_tm(xtime.tv_sec + 1 + timezone_offset, &tm);
398 tm.tm_year -= 1900;
399 tm.tm_mon -= 1;
400 if (ppc_md.set_rtc_time(&tm) == 0)
401 last_rtc_update = xtime.tv_sec + 1;
402 else
403 /* Try again one minute later */
404 last_rtc_update += 60;
409 * This version of gettimeofday has microsecond resolution.
411 static inline void __do_gettimeofday(struct timeval *tv)
413 unsigned long sec, usec;
414 u64 tb_ticks, xsec;
415 struct gettimeofday_vars *temp_varp;
416 u64 temp_tb_to_xs, temp_stamp_xsec;
419 * These calculations are faster (gets rid of divides)
420 * if done in units of 1/2^20 rather than microseconds.
421 * The conversion to microseconds at the end is done
422 * without a divide (and in fact, without a multiply)
424 temp_varp = do_gtod.varp;
426 /* Sampling the time base must be done after loading
427 * do_gtod.varp in order to avoid racing with update_gtod.
429 data_barrier(temp_varp);
430 tb_ticks = get_tb() - temp_varp->tb_orig_stamp;
431 temp_tb_to_xs = temp_varp->tb_to_xs;
432 temp_stamp_xsec = temp_varp->stamp_xsec;
433 xsec = temp_stamp_xsec + mulhdu(tb_ticks, temp_tb_to_xs);
434 sec = xsec / XSEC_PER_SEC;
435 usec = (unsigned long)xsec & (XSEC_PER_SEC - 1);
436 usec = SCALE_XSEC(usec, 1000000);
438 tv->tv_sec = sec;
439 tv->tv_usec = usec;
442 void do_gettimeofday(struct timeval *tv)
444 if (__USE_RTC()) {
445 /* do this the old way */
446 unsigned long flags, seq;
447 unsigned int sec, nsec, usec;
449 do {
450 seq = read_seqbegin_irqsave(&xtime_lock, flags);
451 sec = xtime.tv_sec;
452 nsec = xtime.tv_nsec + tb_ticks_since(tb_last_jiffy);
453 } while (read_seqretry_irqrestore(&xtime_lock, seq, flags));
454 usec = nsec / 1000;
455 while (usec >= 1000000) {
456 usec -= 1000000;
457 ++sec;
459 tv->tv_sec = sec;
460 tv->tv_usec = usec;
461 return;
463 __do_gettimeofday(tv);
466 EXPORT_SYMBOL(do_gettimeofday);
469 * There are two copies of tb_to_xs and stamp_xsec so that no
470 * lock is needed to access and use these values in
471 * do_gettimeofday. We alternate the copies and as long as a
472 * reasonable time elapses between changes, there will never
473 * be inconsistent values. ntpd has a minimum of one minute
474 * between updates.
476 static inline void update_gtod(u64 new_tb_stamp, u64 new_stamp_xsec,
477 u64 new_tb_to_xs)
479 unsigned temp_idx;
480 struct gettimeofday_vars *temp_varp;
482 temp_idx = (do_gtod.var_idx == 0);
483 temp_varp = &do_gtod.vars[temp_idx];
485 temp_varp->tb_to_xs = new_tb_to_xs;
486 temp_varp->tb_orig_stamp = new_tb_stamp;
487 temp_varp->stamp_xsec = new_stamp_xsec;
488 smp_mb();
489 do_gtod.varp = temp_varp;
490 do_gtod.var_idx = temp_idx;
493 * tb_update_count is used to allow the userspace gettimeofday code
494 * to assure itself that it sees a consistent view of the tb_to_xs and
495 * stamp_xsec variables. It reads the tb_update_count, then reads
496 * tb_to_xs and stamp_xsec and then reads tb_update_count again. If
497 * the two values of tb_update_count match and are even then the
498 * tb_to_xs and stamp_xsec values are consistent. If not, then it
499 * loops back and reads them again until this criteria is met.
500 * We expect the caller to have done the first increment of
501 * vdso_data->tb_update_count already.
503 vdso_data->tb_orig_stamp = new_tb_stamp;
504 vdso_data->stamp_xsec = new_stamp_xsec;
505 vdso_data->tb_to_xs = new_tb_to_xs;
506 vdso_data->wtom_clock_sec = wall_to_monotonic.tv_sec;
507 vdso_data->wtom_clock_nsec = wall_to_monotonic.tv_nsec;
508 smp_wmb();
509 ++(vdso_data->tb_update_count);
513 * When the timebase - tb_orig_stamp gets too big, we do a manipulation
514 * between tb_orig_stamp and stamp_xsec. The goal here is to keep the
515 * difference tb - tb_orig_stamp small enough to always fit inside a
516 * 32 bits number. This is a requirement of our fast 32 bits userland
517 * implementation in the vdso. If we "miss" a call to this function
518 * (interrupt latency, CPU locked in a spinlock, ...) and we end up
519 * with a too big difference, then the vdso will fallback to calling
520 * the syscall
522 static __inline__ void timer_recalc_offset(u64 cur_tb)
524 unsigned long offset;
525 u64 new_stamp_xsec;
526 u64 tlen, t2x;
527 u64 tb, xsec_old, xsec_new;
528 struct gettimeofday_vars *varp;
530 if (__USE_RTC())
531 return;
532 tlen = current_tick_length();
533 offset = cur_tb - do_gtod.varp->tb_orig_stamp;
534 if (tlen == last_tick_len && offset < 0x80000000u)
535 return;
536 if (tlen != last_tick_len) {
537 t2x = mulhdu(tlen << TICKLEN_SHIFT, ticklen_to_xs);
538 last_tick_len = tlen;
539 } else
540 t2x = do_gtod.varp->tb_to_xs;
541 new_stamp_xsec = (u64) xtime.tv_nsec * XSEC_PER_SEC;
542 do_div(new_stamp_xsec, 1000000000);
543 new_stamp_xsec += (u64) xtime.tv_sec * XSEC_PER_SEC;
545 ++vdso_data->tb_update_count;
546 smp_mb();
549 * Make sure time doesn't go backwards for userspace gettimeofday.
551 tb = get_tb();
552 varp = do_gtod.varp;
553 xsec_old = mulhdu(tb - varp->tb_orig_stamp, varp->tb_to_xs)
554 + varp->stamp_xsec;
555 xsec_new = mulhdu(tb - cur_tb, t2x) + new_stamp_xsec;
556 if (xsec_new < xsec_old)
557 new_stamp_xsec += xsec_old - xsec_new;
559 update_gtod(cur_tb, new_stamp_xsec, t2x);
562 #ifdef CONFIG_SMP
563 unsigned long profile_pc(struct pt_regs *regs)
565 unsigned long pc = instruction_pointer(regs);
567 if (in_lock_functions(pc))
568 return regs->link;
570 return pc;
572 EXPORT_SYMBOL(profile_pc);
573 #endif
575 #ifdef CONFIG_PPC_ISERIES
578 * This function recalibrates the timebase based on the 49-bit time-of-day
579 * value in the Titan chip. The Titan is much more accurate than the value
580 * returned by the service processor for the timebase frequency.
583 static void iSeries_tb_recal(void)
585 struct div_result divres;
586 unsigned long titan, tb;
587 tb = get_tb();
588 titan = HvCallXm_loadTod();
589 if ( iSeries_recal_titan ) {
590 unsigned long tb_ticks = tb - iSeries_recal_tb;
591 unsigned long titan_usec = (titan - iSeries_recal_titan) >> 12;
592 unsigned long new_tb_ticks_per_sec = (tb_ticks * USEC_PER_SEC)/titan_usec;
593 unsigned long new_tb_ticks_per_jiffy = (new_tb_ticks_per_sec+(HZ/2))/HZ;
594 long tick_diff = new_tb_ticks_per_jiffy - tb_ticks_per_jiffy;
595 char sign = '+';
596 /* make sure tb_ticks_per_sec and tb_ticks_per_jiffy are consistent */
597 new_tb_ticks_per_sec = new_tb_ticks_per_jiffy * HZ;
599 if ( tick_diff < 0 ) {
600 tick_diff = -tick_diff;
601 sign = '-';
603 if ( tick_diff ) {
604 if ( tick_diff < tb_ticks_per_jiffy/25 ) {
605 printk( "Titan recalibrate: new tb_ticks_per_jiffy = %lu (%c%ld)\n",
606 new_tb_ticks_per_jiffy, sign, tick_diff );
607 tb_ticks_per_jiffy = new_tb_ticks_per_jiffy;
608 tb_ticks_per_sec = new_tb_ticks_per_sec;
609 calc_cputime_factors();
610 div128_by_32( XSEC_PER_SEC, 0, tb_ticks_per_sec, &divres );
611 do_gtod.tb_ticks_per_sec = tb_ticks_per_sec;
612 tb_to_xs = divres.result_low;
613 do_gtod.varp->tb_to_xs = tb_to_xs;
614 vdso_data->tb_ticks_per_sec = tb_ticks_per_sec;
615 vdso_data->tb_to_xs = tb_to_xs;
617 else {
618 printk( "Titan recalibrate: FAILED (difference > 4 percent)\n"
619 " new tb_ticks_per_jiffy = %lu\n"
620 " old tb_ticks_per_jiffy = %lu\n",
621 new_tb_ticks_per_jiffy, tb_ticks_per_jiffy );
625 iSeries_recal_titan = titan;
626 iSeries_recal_tb = tb;
628 #endif
631 * For iSeries shared processors, we have to let the hypervisor
632 * set the hardware decrementer. We set a virtual decrementer
633 * in the lppaca and call the hypervisor if the virtual
634 * decrementer is less than the current value in the hardware
635 * decrementer. (almost always the new decrementer value will
636 * be greater than the current hardware decementer so the hypervisor
637 * call will not be needed)
641 * timer_interrupt - gets called when the decrementer overflows,
642 * with interrupts disabled.
644 void timer_interrupt(struct pt_regs * regs)
646 int next_dec;
647 int cpu = smp_processor_id();
648 unsigned long ticks;
649 u64 tb_next_jiffy;
651 #ifdef CONFIG_PPC32
652 if (atomic_read(&ppc_n_lost_interrupts) != 0)
653 do_IRQ(regs);
654 #endif
656 irq_enter();
658 profile_tick(CPU_PROFILING, regs);
659 calculate_steal_time();
661 #ifdef CONFIG_PPC_ISERIES
662 get_lppaca()->int_dword.fields.decr_int = 0;
663 #endif
665 while ((ticks = tb_ticks_since(per_cpu(last_jiffy, cpu)))
666 >= tb_ticks_per_jiffy) {
667 /* Update last_jiffy */
668 per_cpu(last_jiffy, cpu) += tb_ticks_per_jiffy;
669 /* Handle RTCL overflow on 601 */
670 if (__USE_RTC() && per_cpu(last_jiffy, cpu) >= 1000000000)
671 per_cpu(last_jiffy, cpu) -= 1000000000;
674 * We cannot disable the decrementer, so in the period
675 * between this cpu's being marked offline in cpu_online_map
676 * and calling stop-self, it is taking timer interrupts.
677 * Avoid calling into the scheduler rebalancing code if this
678 * is the case.
680 if (!cpu_is_offline(cpu))
681 account_process_time(regs);
684 * No need to check whether cpu is offline here; boot_cpuid
685 * should have been fixed up by now.
687 if (cpu != boot_cpuid)
688 continue;
690 write_seqlock(&xtime_lock);
691 tb_next_jiffy = tb_last_jiffy + tb_ticks_per_jiffy;
692 if (per_cpu(last_jiffy, cpu) >= tb_next_jiffy) {
693 tb_last_jiffy = tb_next_jiffy;
694 do_timer(1);
695 timer_recalc_offset(tb_last_jiffy);
696 timer_check_rtc();
698 write_sequnlock(&xtime_lock);
701 next_dec = tb_ticks_per_jiffy - ticks;
702 set_dec(next_dec);
704 #ifdef CONFIG_PPC_ISERIES
705 if (hvlpevent_is_pending())
706 process_hvlpevents(regs);
707 #endif
709 #ifdef CONFIG_PPC64
710 /* collect purr register values often, for accurate calculations */
711 if (firmware_has_feature(FW_FEATURE_SPLPAR)) {
712 struct cpu_usage *cu = &__get_cpu_var(cpu_usage_array);
713 cu->current_tb = mfspr(SPRN_PURR);
715 #endif
717 irq_exit();
720 void wakeup_decrementer(void)
722 unsigned long ticks;
725 * The timebase gets saved on sleep and restored on wakeup,
726 * so all we need to do is to reset the decrementer.
728 ticks = tb_ticks_since(__get_cpu_var(last_jiffy));
729 if (ticks < tb_ticks_per_jiffy)
730 ticks = tb_ticks_per_jiffy - ticks;
731 else
732 ticks = 1;
733 set_dec(ticks);
736 #ifdef CONFIG_SMP
737 void __init smp_space_timers(unsigned int max_cpus)
739 int i;
740 unsigned long half = tb_ticks_per_jiffy / 2;
741 unsigned long offset = tb_ticks_per_jiffy / max_cpus;
742 u64 previous_tb = per_cpu(last_jiffy, boot_cpuid);
744 /* make sure tb > per_cpu(last_jiffy, cpu) for all cpus always */
745 previous_tb -= tb_ticks_per_jiffy;
747 * The stolen time calculation for POWER5 shared-processor LPAR
748 * systems works better if the two threads' timebase interrupts
749 * are staggered by half a jiffy with respect to each other.
751 for_each_possible_cpu(i) {
752 if (i == boot_cpuid)
753 continue;
754 if (i == (boot_cpuid ^ 1))
755 per_cpu(last_jiffy, i) =
756 per_cpu(last_jiffy, boot_cpuid) - half;
757 else if (i & 1)
758 per_cpu(last_jiffy, i) =
759 per_cpu(last_jiffy, i ^ 1) + half;
760 else {
761 previous_tb += offset;
762 per_cpu(last_jiffy, i) = previous_tb;
766 #endif
769 * Scheduler clock - returns current time in nanosec units.
771 * Note: mulhdu(a, b) (multiply high double unsigned) returns
772 * the high 64 bits of a * b, i.e. (a * b) >> 64, where a and b
773 * are 64-bit unsigned numbers.
775 unsigned long long sched_clock(void)
777 if (__USE_RTC())
778 return get_rtc();
779 return mulhdu(get_tb(), tb_to_ns_scale) << tb_to_ns_shift;
782 int do_settimeofday(struct timespec *tv)
784 time_t wtm_sec, new_sec = tv->tv_sec;
785 long wtm_nsec, new_nsec = tv->tv_nsec;
786 unsigned long flags;
787 u64 new_xsec;
788 unsigned long tb_delta;
790 if ((unsigned long)tv->tv_nsec >= NSEC_PER_SEC)
791 return -EINVAL;
793 write_seqlock_irqsave(&xtime_lock, flags);
796 * Updating the RTC is not the job of this code. If the time is
797 * stepped under NTP, the RTC will be updated after STA_UNSYNC
798 * is cleared. Tools like clock/hwclock either copy the RTC
799 * to the system time, in which case there is no point in writing
800 * to the RTC again, or write to the RTC but then they don't call
801 * settimeofday to perform this operation.
803 #ifdef CONFIG_PPC_ISERIES
804 if (first_settimeofday) {
805 iSeries_tb_recal();
806 first_settimeofday = 0;
808 #endif
810 /* Make userspace gettimeofday spin until we're done. */
811 ++vdso_data->tb_update_count;
812 smp_mb();
815 * Subtract off the number of nanoseconds since the
816 * beginning of the last tick.
818 tb_delta = tb_ticks_since(tb_last_jiffy);
819 tb_delta = mulhdu(tb_delta, do_gtod.varp->tb_to_xs); /* in xsec */
820 new_nsec -= SCALE_XSEC(tb_delta, 1000000000);
822 wtm_sec = wall_to_monotonic.tv_sec + (xtime.tv_sec - new_sec);
823 wtm_nsec = wall_to_monotonic.tv_nsec + (xtime.tv_nsec - new_nsec);
825 set_normalized_timespec(&xtime, new_sec, new_nsec);
826 set_normalized_timespec(&wall_to_monotonic, wtm_sec, wtm_nsec);
828 /* In case of a large backwards jump in time with NTP, we want the
829 * clock to be updated as soon as the PLL is again in lock.
831 last_rtc_update = new_sec - 658;
833 ntp_clear();
835 new_xsec = xtime.tv_nsec;
836 if (new_xsec != 0) {
837 new_xsec *= XSEC_PER_SEC;
838 do_div(new_xsec, NSEC_PER_SEC);
840 new_xsec += (u64)xtime.tv_sec * XSEC_PER_SEC;
841 update_gtod(tb_last_jiffy, new_xsec, do_gtod.varp->tb_to_xs);
843 vdso_data->tz_minuteswest = sys_tz.tz_minuteswest;
844 vdso_data->tz_dsttime = sys_tz.tz_dsttime;
846 write_sequnlock_irqrestore(&xtime_lock, flags);
847 clock_was_set();
848 return 0;
851 EXPORT_SYMBOL(do_settimeofday);
853 static int __init get_freq(char *name, int cells, unsigned long *val)
855 struct device_node *cpu;
856 const unsigned int *fp;
857 int found = 0;
859 /* The cpu node should have timebase and clock frequency properties */
860 cpu = of_find_node_by_type(NULL, "cpu");
862 if (cpu) {
863 fp = get_property(cpu, name, NULL);
864 if (fp) {
865 found = 1;
866 *val = of_read_ulong(fp, cells);
869 of_node_put(cpu);
872 return found;
875 void __init generic_calibrate_decr(void)
877 ppc_tb_freq = DEFAULT_TB_FREQ; /* hardcoded default */
879 if (!get_freq("ibm,extended-timebase-frequency", 2, &ppc_tb_freq) &&
880 !get_freq("timebase-frequency", 1, &ppc_tb_freq)) {
882 printk(KERN_ERR "WARNING: Estimating decrementer frequency "
883 "(not found)\n");
886 ppc_proc_freq = DEFAULT_PROC_FREQ; /* hardcoded default */
888 if (!get_freq("ibm,extended-clock-frequency", 2, &ppc_proc_freq) &&
889 !get_freq("clock-frequency", 1, &ppc_proc_freq)) {
891 printk(KERN_ERR "WARNING: Estimating processor frequency "
892 "(not found)\n");
895 #ifdef CONFIG_BOOKE
896 /* Set the time base to zero */
897 mtspr(SPRN_TBWL, 0);
898 mtspr(SPRN_TBWU, 0);
900 /* Clear any pending timer interrupts */
901 mtspr(SPRN_TSR, TSR_ENW | TSR_WIS | TSR_DIS | TSR_FIS);
903 /* Enable decrementer interrupt */
904 mtspr(SPRN_TCR, TCR_DIE);
905 #endif
908 unsigned long get_boot_time(void)
910 struct rtc_time tm;
912 if (ppc_md.get_boot_time)
913 return ppc_md.get_boot_time();
914 if (!ppc_md.get_rtc_time)
915 return 0;
916 ppc_md.get_rtc_time(&tm);
917 return mktime(tm.tm_year+1900, tm.tm_mon+1, tm.tm_mday,
918 tm.tm_hour, tm.tm_min, tm.tm_sec);
921 /* This function is only called on the boot processor */
922 void __init time_init(void)
924 unsigned long flags;
925 unsigned long tm = 0;
926 struct div_result res;
927 u64 scale, x;
928 unsigned shift;
930 if (ppc_md.time_init != NULL)
931 timezone_offset = ppc_md.time_init();
933 if (__USE_RTC()) {
934 /* 601 processor: dec counts down by 128 every 128ns */
935 ppc_tb_freq = 1000000000;
936 tb_last_jiffy = get_rtcl();
937 } else {
938 /* Normal PowerPC with timebase register */
939 ppc_md.calibrate_decr();
940 printk(KERN_DEBUG "time_init: decrementer frequency = %lu.%.6lu MHz\n",
941 ppc_tb_freq / 1000000, ppc_tb_freq % 1000000);
942 printk(KERN_DEBUG "time_init: processor frequency = %lu.%.6lu MHz\n",
943 ppc_proc_freq / 1000000, ppc_proc_freq % 1000000);
944 tb_last_jiffy = get_tb();
947 tb_ticks_per_jiffy = ppc_tb_freq / HZ;
948 tb_ticks_per_sec = ppc_tb_freq;
949 tb_ticks_per_usec = ppc_tb_freq / 1000000;
950 tb_to_us = mulhwu_scale_factor(ppc_tb_freq, 1000000);
951 calc_cputime_factors();
954 * Calculate the length of each tick in ns. It will not be
955 * exactly 1e9/HZ unless ppc_tb_freq is divisible by HZ.
956 * We compute 1e9 * tb_ticks_per_jiffy / ppc_tb_freq,
957 * rounded up.
959 x = (u64) NSEC_PER_SEC * tb_ticks_per_jiffy + ppc_tb_freq - 1;
960 do_div(x, ppc_tb_freq);
961 tick_nsec = x;
962 last_tick_len = x << TICKLEN_SCALE;
965 * Compute ticklen_to_xs, which is a factor which gets multiplied
966 * by (last_tick_len << TICKLEN_SHIFT) to get a tb_to_xs value.
967 * It is computed as:
968 * ticklen_to_xs = 2^N / (tb_ticks_per_jiffy * 1e9)
969 * where N = 64 + 20 - TICKLEN_SCALE - TICKLEN_SHIFT
970 * which turns out to be N = 51 - SHIFT_HZ.
971 * This gives the result as a 0.64 fixed-point fraction.
972 * That value is reduced by an offset amounting to 1 xsec per
973 * 2^31 timebase ticks to avoid problems with time going backwards
974 * by 1 xsec when we do timer_recalc_offset due to losing the
975 * fractional xsec. That offset is equal to ppc_tb_freq/2^51
976 * since there are 2^20 xsec in a second.
978 div128_by_32((1ULL << 51) - ppc_tb_freq, 0,
979 tb_ticks_per_jiffy << SHIFT_HZ, &res);
980 div128_by_32(res.result_high, res.result_low, NSEC_PER_SEC, &res);
981 ticklen_to_xs = res.result_low;
983 /* Compute tb_to_xs from tick_nsec */
984 tb_to_xs = mulhdu(last_tick_len << TICKLEN_SHIFT, ticklen_to_xs);
987 * Compute scale factor for sched_clock.
988 * The calibrate_decr() function has set tb_ticks_per_sec,
989 * which is the timebase frequency.
990 * We compute 1e9 * 2^64 / tb_ticks_per_sec and interpret
991 * the 128-bit result as a 64.64 fixed-point number.
992 * We then shift that number right until it is less than 1.0,
993 * giving us the scale factor and shift count to use in
994 * sched_clock().
996 div128_by_32(1000000000, 0, tb_ticks_per_sec, &res);
997 scale = res.result_low;
998 for (shift = 0; res.result_high != 0; ++shift) {
999 scale = (scale >> 1) | (res.result_high << 63);
1000 res.result_high >>= 1;
1002 tb_to_ns_scale = scale;
1003 tb_to_ns_shift = shift;
1005 tm = get_boot_time();
1007 write_seqlock_irqsave(&xtime_lock, flags);
1009 /* If platform provided a timezone (pmac), we correct the time */
1010 if (timezone_offset) {
1011 sys_tz.tz_minuteswest = -timezone_offset / 60;
1012 sys_tz.tz_dsttime = 0;
1013 tm -= timezone_offset;
1016 xtime.tv_sec = tm;
1017 xtime.tv_nsec = 0;
1018 do_gtod.varp = &do_gtod.vars[0];
1019 do_gtod.var_idx = 0;
1020 do_gtod.varp->tb_orig_stamp = tb_last_jiffy;
1021 __get_cpu_var(last_jiffy) = tb_last_jiffy;
1022 do_gtod.varp->stamp_xsec = (u64) xtime.tv_sec * XSEC_PER_SEC;
1023 do_gtod.tb_ticks_per_sec = tb_ticks_per_sec;
1024 do_gtod.varp->tb_to_xs = tb_to_xs;
1025 do_gtod.tb_to_us = tb_to_us;
1027 vdso_data->tb_orig_stamp = tb_last_jiffy;
1028 vdso_data->tb_update_count = 0;
1029 vdso_data->tb_ticks_per_sec = tb_ticks_per_sec;
1030 vdso_data->stamp_xsec = (u64) xtime.tv_sec * XSEC_PER_SEC;
1031 vdso_data->tb_to_xs = tb_to_xs;
1033 time_freq = 0;
1035 last_rtc_update = xtime.tv_sec;
1036 set_normalized_timespec(&wall_to_monotonic,
1037 -xtime.tv_sec, -xtime.tv_nsec);
1038 write_sequnlock_irqrestore(&xtime_lock, flags);
1040 /* Not exact, but the timer interrupt takes care of this */
1041 set_dec(tb_ticks_per_jiffy);
1045 #define FEBRUARY 2
1046 #define STARTOFTIME 1970
1047 #define SECDAY 86400L
1048 #define SECYR (SECDAY * 365)
1049 #define leapyear(year) ((year) % 4 == 0 && \
1050 ((year) % 100 != 0 || (year) % 400 == 0))
1051 #define days_in_year(a) (leapyear(a) ? 366 : 365)
1052 #define days_in_month(a) (month_days[(a) - 1])
1054 static int month_days[12] = {
1055 31, 28, 31, 30, 31, 30, 31, 31, 30, 31, 30, 31
1059 * This only works for the Gregorian calendar - i.e. after 1752 (in the UK)
1061 void GregorianDay(struct rtc_time * tm)
1063 int leapsToDate;
1064 int lastYear;
1065 int day;
1066 int MonthOffset[] = { 0, 31, 59, 90, 120, 151, 181, 212, 243, 273, 304, 334 };
1068 lastYear = tm->tm_year - 1;
1071 * Number of leap corrections to apply up to end of last year
1073 leapsToDate = lastYear / 4 - lastYear / 100 + lastYear / 400;
1076 * This year is a leap year if it is divisible by 4 except when it is
1077 * divisible by 100 unless it is divisible by 400
1079 * e.g. 1904 was a leap year, 1900 was not, 1996 is, and 2000 was
1081 day = tm->tm_mon > 2 && leapyear(tm->tm_year);
1083 day += lastYear*365 + leapsToDate + MonthOffset[tm->tm_mon-1] +
1084 tm->tm_mday;
1086 tm->tm_wday = day % 7;
1089 void to_tm(int tim, struct rtc_time * tm)
1091 register int i;
1092 register long hms, day;
1094 day = tim / SECDAY;
1095 hms = tim % SECDAY;
1097 /* Hours, minutes, seconds are easy */
1098 tm->tm_hour = hms / 3600;
1099 tm->tm_min = (hms % 3600) / 60;
1100 tm->tm_sec = (hms % 3600) % 60;
1102 /* Number of years in days */
1103 for (i = STARTOFTIME; day >= days_in_year(i); i++)
1104 day -= days_in_year(i);
1105 tm->tm_year = i;
1107 /* Number of months in days left */
1108 if (leapyear(tm->tm_year))
1109 days_in_month(FEBRUARY) = 29;
1110 for (i = 1; day >= days_in_month(i); i++)
1111 day -= days_in_month(i);
1112 days_in_month(FEBRUARY) = 28;
1113 tm->tm_mon = i;
1115 /* Days are what is left over (+1) from all that. */
1116 tm->tm_mday = day + 1;
1119 * Determine the day of week
1121 GregorianDay(tm);
1124 /* Auxiliary function to compute scaling factors */
1125 /* Actually the choice of a timebase running at 1/4 the of the bus
1126 * frequency giving resolution of a few tens of nanoseconds is quite nice.
1127 * It makes this computation very precise (27-28 bits typically) which
1128 * is optimistic considering the stability of most processor clock
1129 * oscillators and the precision with which the timebase frequency
1130 * is measured but does not harm.
1132 unsigned mulhwu_scale_factor(unsigned inscale, unsigned outscale)
1134 unsigned mlt=0, tmp, err;
1135 /* No concern for performance, it's done once: use a stupid
1136 * but safe and compact method to find the multiplier.
1139 for (tmp = 1U<<31; tmp != 0; tmp >>= 1) {
1140 if (mulhwu(inscale, mlt|tmp) < outscale)
1141 mlt |= tmp;
1144 /* We might still be off by 1 for the best approximation.
1145 * A side effect of this is that if outscale is too large
1146 * the returned value will be zero.
1147 * Many corner cases have been checked and seem to work,
1148 * some might have been forgotten in the test however.
1151 err = inscale * (mlt+1);
1152 if (err <= inscale/2)
1153 mlt++;
1154 return mlt;
1158 * Divide a 128-bit dividend by a 32-bit divisor, leaving a 128 bit
1159 * result.
1161 void div128_by_32(u64 dividend_high, u64 dividend_low,
1162 unsigned divisor, struct div_result *dr)
1164 unsigned long a, b, c, d;
1165 unsigned long w, x, y, z;
1166 u64 ra, rb, rc;
1168 a = dividend_high >> 32;
1169 b = dividend_high & 0xffffffff;
1170 c = dividend_low >> 32;
1171 d = dividend_low & 0xffffffff;
1173 w = a / divisor;
1174 ra = ((u64)(a - (w * divisor)) << 32) + b;
1176 rb = ((u64) do_div(ra, divisor) << 32) + c;
1177 x = ra;
1179 rc = ((u64) do_div(rb, divisor) << 32) + d;
1180 y = rb;
1182 do_div(rc, divisor);
1183 z = rc;
1185 dr->result_high = ((u64)w << 32) + x;
1186 dr->result_low = ((u64)y << 32) + z;