641c8b61c4f167200adf0e52646d5d823a509096
[linux-2.6.git] / arch / ia64 / kernel / time.c
1 /*
2  * linux/arch/ia64/kernel/time.c
3  *
4  * Copyright (C) 1998-2003 Hewlett-Packard Co
5  *      Stephane Eranian <eranian@hpl.hp.com>
6  *      David Mosberger <davidm@hpl.hp.com>
7  * Copyright (C) 1999 Don Dugger <don.dugger@intel.com>
8  * Copyright (C) 1999-2000 VA Linux Systems
9  * Copyright (C) 1999-2000 Walt Drummond <drummond@valinux.com>
10  */
11
12 #include <linux/cpu.h>
13 #include <linux/init.h>
14 #include <linux/kernel.h>
15 #include <linux/module.h>
16 #include <linux/profile.h>
17 #include <linux/sched.h>
18 #include <linux/time.h>
19 #include <linux/interrupt.h>
20 #include <linux/efi.h>
21 #include <linux/timex.h>
22 #include <linux/clocksource.h>
23 #include <linux/platform_device.h>
24
25 #include <asm/machvec.h>
26 #include <asm/delay.h>
27 #include <asm/hw_irq.h>
28 #include <asm/paravirt.h>
29 #include <asm/ptrace.h>
30 #include <asm/sal.h>
31 #include <asm/sections.h>
32 #include <asm/system.h>
33
34 #include "fsyscall_gtod_data.h"
35
36 static cycle_t itc_get_cycles(void);
37
38 struct fsyscall_gtod_data_t fsyscall_gtod_data = {
39         .lock = SEQLOCK_UNLOCKED,
40 };
41
42 struct itc_jitter_data_t itc_jitter_data;
43
44 volatile int time_keeper_id = 0; /* smp_processor_id() of time-keeper */
45
46 #ifdef CONFIG_IA64_DEBUG_IRQ
47
48 unsigned long last_cli_ip;
49 EXPORT_SYMBOL(last_cli_ip);
50
51 #endif
52
53 #ifdef CONFIG_PARAVIRT
54 /* We need to define a real function for sched_clock, to override the
55    weak default version */
56 unsigned long long sched_clock(void)
57 {
58         return paravirt_sched_clock();
59 }
60 #endif
61
62 #ifdef CONFIG_PARAVIRT
63 static void
64 paravirt_clocksource_resume(void)
65 {
66         if (pv_time_ops.clocksource_resume)
67                 pv_time_ops.clocksource_resume();
68 }
69 #endif
70
71 static struct clocksource clocksource_itc = {
72         .name           = "itc",
73         .rating         = 350,
74         .read           = itc_get_cycles,
75         .mask           = CLOCKSOURCE_MASK(64),
76         .mult           = 0, /*to be calculated*/
77         .shift          = 16,
78         .flags          = CLOCK_SOURCE_IS_CONTINUOUS,
79 #ifdef CONFIG_PARAVIRT
80         .resume         = paravirt_clocksource_resume,
81 #endif
82 };
83 static struct clocksource *itc_clocksource;
84
85 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
86
87 #include <linux/kernel_stat.h>
88
89 extern cputime_t cycle_to_cputime(u64 cyc);
90
91 /*
92  * Called from the context switch with interrupts disabled, to charge all
93  * accumulated times to the current process, and to prepare accounting on
94  * the next process.
95  */
96 void ia64_account_on_switch(struct task_struct *prev, struct task_struct *next)
97 {
98         struct thread_info *pi = task_thread_info(prev);
99         struct thread_info *ni = task_thread_info(next);
100         cputime_t delta_stime, delta_utime;
101         __u64 now;
102
103         now = ia64_get_itc();
104
105         delta_stime = cycle_to_cputime(pi->ac_stime + (now - pi->ac_stamp));
106         if (idle_task(smp_processor_id()) != prev)
107                 account_system_time(prev, 0, delta_stime, delta_stime);
108         else
109                 account_idle_time(delta_stime);
110
111         if (pi->ac_utime) {
112                 delta_utime = cycle_to_cputime(pi->ac_utime);
113                 account_user_time(prev, delta_utime, delta_utime);
114         }
115
116         pi->ac_stamp = ni->ac_stamp = now;
117         ni->ac_stime = ni->ac_utime = 0;
118 }
119
120 /*
121  * Account time for a transition between system, hard irq or soft irq state.
122  * Note that this function is called with interrupts enabled.
123  */
124 void account_system_vtime(struct task_struct *tsk)
125 {
126         struct thread_info *ti = task_thread_info(tsk);
127         unsigned long flags;
128         cputime_t delta_stime;
129         __u64 now;
130
131         local_irq_save(flags);
132
133         now = ia64_get_itc();
134
135         delta_stime = cycle_to_cputime(ti->ac_stime + (now - ti->ac_stamp));
136         if (irq_count() || idle_task(smp_processor_id()) != tsk)
137                 account_system_time(tsk, 0, delta_stime, delta_stime);
138         else
139                 account_idle_time(delta_stime);
140         ti->ac_stime = 0;
141
142         ti->ac_stamp = now;
143
144         local_irq_restore(flags);
145 }
146 EXPORT_SYMBOL_GPL(account_system_vtime);
147
148 /*
149  * Called from the timer interrupt handler to charge accumulated user time
150  * to the current process.  Must be called with interrupts disabled.
151  */
152 void account_process_tick(struct task_struct *p, int user_tick)
153 {
154         struct thread_info *ti = task_thread_info(p);
155         cputime_t delta_utime;
156
157         if (ti->ac_utime) {
158                 delta_utime = cycle_to_cputime(ti->ac_utime);
159                 account_user_time(p, delta_utime, delta_utime);
160                 ti->ac_utime = 0;
161         }
162 }
163
164 #endif /* CONFIG_VIRT_CPU_ACCOUNTING */
165
166 static irqreturn_t
167 timer_interrupt (int irq, void *dev_id)
168 {
169         unsigned long new_itm;
170
171         if (unlikely(cpu_is_offline(smp_processor_id()))) {
172                 return IRQ_HANDLED;
173         }
174
175         platform_timer_interrupt(irq, dev_id);
176
177         new_itm = local_cpu_data->itm_next;
178
179         if (!time_after(ia64_get_itc(), new_itm))
180                 printk(KERN_ERR "Oops: timer tick before it's due (itc=%lx,itm=%lx)\n",
181                        ia64_get_itc(), new_itm);
182
183         profile_tick(CPU_PROFILING);
184
185         if (paravirt_do_steal_accounting(&new_itm))
186                 goto skip_process_time_accounting;
187
188         while (1) {
189                 update_process_times(user_mode(get_irq_regs()));
190
191                 new_itm += local_cpu_data->itm_delta;
192
193                 if (smp_processor_id() == time_keeper_id) {
194                         /*
195                          * Here we are in the timer irq handler. We have irqs locally
196                          * disabled, but we don't know if the timer_bh is running on
197                          * another CPU. We need to avoid to SMP race by acquiring the
198                          * xtime_lock.
199                          */
200                         write_seqlock(&xtime_lock);
201                         do_timer(1);
202                         local_cpu_data->itm_next = new_itm;
203                         write_sequnlock(&xtime_lock);
204                 } else
205                         local_cpu_data->itm_next = new_itm;
206
207                 if (time_after(new_itm, ia64_get_itc()))
208                         break;
209
210                 /*
211                  * Allow IPIs to interrupt the timer loop.
212                  */
213                 local_irq_enable();
214                 local_irq_disable();
215         }
216
217 skip_process_time_accounting:
218
219         do {
220                 /*
221                  * If we're too close to the next clock tick for
222                  * comfort, we increase the safety margin by
223                  * intentionally dropping the next tick(s).  We do NOT
224                  * update itm.next because that would force us to call
225                  * do_timer() which in turn would let our clock run
226                  * too fast (with the potentially devastating effect
227                  * of losing monotony of time).
228                  */
229                 while (!time_after(new_itm, ia64_get_itc() + local_cpu_data->itm_delta/2))
230                         new_itm += local_cpu_data->itm_delta;
231                 ia64_set_itm(new_itm);
232                 /* double check, in case we got hit by a (slow) PMI: */
233         } while (time_after_eq(ia64_get_itc(), new_itm));
234         return IRQ_HANDLED;
235 }
236
237 /*
238  * Encapsulate access to the itm structure for SMP.
239  */
240 void
241 ia64_cpu_local_tick (void)
242 {
243         int cpu = smp_processor_id();
244         unsigned long shift = 0, delta;
245
246         /* arrange for the cycle counter to generate a timer interrupt: */
247         ia64_set_itv(IA64_TIMER_VECTOR);
248
249         delta = local_cpu_data->itm_delta;
250         /*
251          * Stagger the timer tick for each CPU so they don't occur all at (almost) the
252          * same time:
253          */
254         if (cpu) {
255                 unsigned long hi = 1UL << ia64_fls(cpu);
256                 shift = (2*(cpu - hi) + 1) * delta/hi/2;
257         }
258         local_cpu_data->itm_next = ia64_get_itc() + delta + shift;
259         ia64_set_itm(local_cpu_data->itm_next);
260 }
261
262 static int nojitter;
263
264 static int __init nojitter_setup(char *str)
265 {
266         nojitter = 1;
267         printk("Jitter checking for ITC timers disabled\n");
268         return 1;
269 }
270
271 __setup("nojitter", nojitter_setup);
272
273
274 void __devinit
275 ia64_init_itm (void)
276 {
277         unsigned long platform_base_freq, itc_freq;
278         struct pal_freq_ratio itc_ratio, proc_ratio;
279         long status, platform_base_drift, itc_drift;
280
281         /*
282          * According to SAL v2.6, we need to use a SAL call to determine the platform base
283          * frequency and then a PAL call to determine the frequency ratio between the ITC
284          * and the base frequency.
285          */
286         status = ia64_sal_freq_base(SAL_FREQ_BASE_PLATFORM,
287                                     &platform_base_freq, &platform_base_drift);
288         if (status != 0) {
289                 printk(KERN_ERR "SAL_FREQ_BASE_PLATFORM failed: %s\n", ia64_sal_strerror(status));
290         } else {
291                 status = ia64_pal_freq_ratios(&proc_ratio, NULL, &itc_ratio);
292                 if (status != 0)
293                         printk(KERN_ERR "PAL_FREQ_RATIOS failed with status=%ld\n", status);
294         }
295         if (status != 0) {
296                 /* invent "random" values */
297                 printk(KERN_ERR
298                        "SAL/PAL failed to obtain frequency info---inventing reasonable values\n");
299                 platform_base_freq = 100000000;
300                 platform_base_drift = -1;       /* no drift info */
301                 itc_ratio.num = 3;
302                 itc_ratio.den = 1;
303         }
304         if (platform_base_freq < 40000000) {
305                 printk(KERN_ERR "Platform base frequency %lu bogus---resetting to 75MHz!\n",
306                        platform_base_freq);
307                 platform_base_freq = 75000000;
308                 platform_base_drift = -1;
309         }
310         if (!proc_ratio.den)
311                 proc_ratio.den = 1;     /* avoid division by zero */
312         if (!itc_ratio.den)
313                 itc_ratio.den = 1;      /* avoid division by zero */
314
315         itc_freq = (platform_base_freq*itc_ratio.num)/itc_ratio.den;
316
317         local_cpu_data->itm_delta = (itc_freq + HZ/2) / HZ;
318         printk(KERN_DEBUG "CPU %d: base freq=%lu.%03luMHz, ITC ratio=%u/%u, "
319                "ITC freq=%lu.%03luMHz", smp_processor_id(),
320                platform_base_freq / 1000000, (platform_base_freq / 1000) % 1000,
321                itc_ratio.num, itc_ratio.den, itc_freq / 1000000, (itc_freq / 1000) % 1000);
322
323         if (platform_base_drift != -1) {
324                 itc_drift = platform_base_drift*itc_ratio.num/itc_ratio.den;
325                 printk("+/-%ldppm\n", itc_drift);
326         } else {
327                 itc_drift = -1;
328                 printk("\n");
329         }
330
331         local_cpu_data->proc_freq = (platform_base_freq*proc_ratio.num)/proc_ratio.den;
332         local_cpu_data->itc_freq = itc_freq;
333         local_cpu_data->cyc_per_usec = (itc_freq + USEC_PER_SEC/2) / USEC_PER_SEC;
334         local_cpu_data->nsec_per_cyc = ((NSEC_PER_SEC<<IA64_NSEC_PER_CYC_SHIFT)
335                                         + itc_freq/2)/itc_freq;
336
337         if (!(sal_platform_features & IA64_SAL_PLATFORM_FEATURE_ITC_DRIFT)) {
338 #ifdef CONFIG_SMP
339                 /* On IA64 in an SMP configuration ITCs are never accurately synchronized.
340                  * Jitter compensation requires a cmpxchg which may limit
341                  * the scalability of the syscalls for retrieving time.
342                  * The ITC synchronization is usually successful to within a few
343                  * ITC ticks but this is not a sure thing. If you need to improve
344                  * timer performance in SMP situations then boot the kernel with the
345                  * "nojitter" option. However, doing so may result in time fluctuating (maybe
346                  * even going backward) if the ITC offsets between the individual CPUs
347                  * are too large.
348                  */
349                 if (!nojitter)
350                         itc_jitter_data.itc_jitter = 1;
351 #endif
352         } else
353                 /*
354                  * ITC is drifty and we have not synchronized the ITCs in smpboot.c.
355                  * ITC values may fluctuate significantly between processors.
356                  * Clock should not be used for hrtimers. Mark itc as only
357                  * useful for boot and testing.
358                  *
359                  * Note that jitter compensation is off! There is no point of
360                  * synchronizing ITCs since they may be large differentials
361                  * that change over time.
362                  *
363                  * The only way to fix this would be to repeatedly sync the
364                  * ITCs. Until that time we have to avoid ITC.
365                  */
366                 clocksource_itc.rating = 50;
367
368         paravirt_init_missing_ticks_accounting(smp_processor_id());
369
370         /* avoid softlock up message when cpu is unplug and plugged again. */
371         touch_softlockup_watchdog();
372
373         /* Setup the CPU local timer tick */
374         ia64_cpu_local_tick();
375
376         if (!itc_clocksource) {
377                 /* Sort out mult/shift values: */
378                 clocksource_itc.mult =
379                         clocksource_hz2mult(local_cpu_data->itc_freq,
380                                                 clocksource_itc.shift);
381                 clocksource_register(&clocksource_itc);
382                 itc_clocksource = &clocksource_itc;
383         }
384 }
385
386 static cycle_t itc_get_cycles(void)
387 {
388         u64 lcycle, now, ret;
389
390         if (!itc_jitter_data.itc_jitter)
391                 return get_cycles();
392
393         lcycle = itc_jitter_data.itc_lastcycle;
394         now = get_cycles();
395         if (lcycle && time_after(lcycle, now))
396                 return lcycle;
397
398         /*
399          * Keep track of the last timer value returned.
400          * In an SMP environment, you could lose out in contention of
401          * cmpxchg. If so, your cmpxchg returns new value which the
402          * winner of contention updated to. Use the new value instead.
403          */
404         ret = cmpxchg(&itc_jitter_data.itc_lastcycle, lcycle, now);
405         if (unlikely(ret != lcycle))
406                 return ret;
407
408         return now;
409 }
410
411
412 static struct irqaction timer_irqaction = {
413         .handler =      timer_interrupt,
414         .flags =        IRQF_DISABLED | IRQF_IRQPOLL,
415         .name =         "timer"
416 };
417
418 static struct platform_device rtc_efi_dev = {
419         .name = "rtc-efi",
420         .id = -1,
421 };
422
423 static int __init rtc_init(void)
424 {
425         if (platform_device_register(&rtc_efi_dev) < 0)
426                 printk(KERN_ERR "unable to register rtc device...\n");
427
428         /* not necessarily an error */
429         return 0;
430 }
431 module_init(rtc_init);
432
433 void __init
434 time_init (void)
435 {
436         register_percpu_irq(IA64_TIMER_VECTOR, &timer_irqaction);
437         efi_gettimeofday(&xtime);
438         ia64_init_itm();
439
440         /*
441          * Initialize wall_to_monotonic such that adding it to xtime will yield zero, the
442          * tv_nsec field must be normalized (i.e., 0 <= nsec < NSEC_PER_SEC).
443          */
444         set_normalized_timespec(&wall_to_monotonic, -xtime.tv_sec, -xtime.tv_nsec);
445 }
446
447 /*
448  * Generic udelay assumes that if preemption is allowed and the thread
449  * migrates to another CPU, that the ITC values are synchronized across
450  * all CPUs.
451  */
452 static void
453 ia64_itc_udelay (unsigned long usecs)
454 {
455         unsigned long start = ia64_get_itc();
456         unsigned long end = start + usecs*local_cpu_data->cyc_per_usec;
457
458         while (time_before(ia64_get_itc(), end))
459                 cpu_relax();
460 }
461
462 void (*ia64_udelay)(unsigned long usecs) = &ia64_itc_udelay;
463
464 void
465 udelay (unsigned long usecs)
466 {
467         (*ia64_udelay)(usecs);
468 }
469 EXPORT_SYMBOL(udelay);
470
471 /* IA64 doesn't cache the timezone */
472 void update_vsyscall_tz(void)
473 {
474 }
475
476 void update_vsyscall(struct timespec *wall, struct clocksource *c)
477 {
478         unsigned long flags;
479
480         write_seqlock_irqsave(&fsyscall_gtod_data.lock, flags);
481
482         /* copy fsyscall clock data */
483         fsyscall_gtod_data.clk_mask = c->mask;
484         fsyscall_gtod_data.clk_mult = c->mult;
485         fsyscall_gtod_data.clk_shift = c->shift;
486         fsyscall_gtod_data.clk_fsys_mmio = c->fsys_mmio;
487         fsyscall_gtod_data.clk_cycle_last = c->cycle_last;
488
489         /* copy kernel time structures */
490         fsyscall_gtod_data.wall_time.tv_sec = wall->tv_sec;
491         fsyscall_gtod_data.wall_time.tv_nsec = wall->tv_nsec;
492         fsyscall_gtod_data.monotonic_time.tv_sec = wall_to_monotonic.tv_sec
493                                                         + wall->tv_sec;
494         fsyscall_gtod_data.monotonic_time.tv_nsec = wall_to_monotonic.tv_nsec
495                                                         + wall->tv_nsec;
496
497         /* normalize */
498         while (fsyscall_gtod_data.monotonic_time.tv_nsec >= NSEC_PER_SEC) {
499                 fsyscall_gtod_data.monotonic_time.tv_nsec -= NSEC_PER_SEC;
500                 fsyscall_gtod_data.monotonic_time.tv_sec++;
501         }
502
503         write_sequnlock_irqrestore(&fsyscall_gtod_data.lock, flags);
504 }
505