c9a7d2ebe0890fd9ea441bcc63f9f178a2565c10
[linux-3.10.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/timekeeper_internal.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
33 #include "fsyscall_gtod_data.h"
34
35 static cycle_t itc_get_cycles(struct clocksource *cs);
36
37 struct fsyscall_gtod_data_t fsyscall_gtod_data;
38
39 struct itc_jitter_data_t itc_jitter_data;
40
41 volatile int time_keeper_id = 0; /* smp_processor_id() of time-keeper */
42
43 #ifdef CONFIG_IA64_DEBUG_IRQ
44
45 unsigned long last_cli_ip;
46 EXPORT_SYMBOL(last_cli_ip);
47
48 #endif
49
50 #ifdef CONFIG_PARAVIRT
51 /* We need to define a real function for sched_clock, to override the
52    weak default version */
53 unsigned long long sched_clock(void)
54 {
55         return paravirt_sched_clock();
56 }
57 #endif
58
59 #ifdef CONFIG_PARAVIRT
60 static void
61 paravirt_clocksource_resume(struct clocksource *cs)
62 {
63         if (pv_time_ops.clocksource_resume)
64                 pv_time_ops.clocksource_resume();
65 }
66 #endif
67
68 static struct clocksource clocksource_itc = {
69         .name           = "itc",
70         .rating         = 350,
71         .read           = itc_get_cycles,
72         .mask           = CLOCKSOURCE_MASK(64),
73         .flags          = CLOCK_SOURCE_IS_CONTINUOUS,
74 #ifdef CONFIG_PARAVIRT
75         .resume         = paravirt_clocksource_resume,
76 #endif
77 };
78 static struct clocksource *itc_clocksource;
79
80 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
81
82 #include <linux/kernel_stat.h>
83
84 extern cputime_t cycle_to_cputime(u64 cyc);
85
86 void vtime_account_user(struct task_struct *tsk)
87 {
88         cputime_t delta_utime;
89         struct thread_info *ti = task_thread_info(tsk);
90
91         if (ti->ac_utime) {
92                 delta_utime = cycle_to_cputime(ti->ac_utime);
93                 account_user_time(tsk, delta_utime, delta_utime);
94                 ti->ac_utime = 0;
95         }
96 }
97
98 /*
99  * Called from the context switch with interrupts disabled, to charge all
100  * accumulated times to the current process, and to prepare accounting on
101  * the next process.
102  */
103 void arch_vtime_task_switch(struct task_struct *prev)
104 {
105         struct thread_info *pi = task_thread_info(prev);
106         struct thread_info *ni = task_thread_info(current);
107
108         pi->ac_stamp = ni->ac_stamp;
109         ni->ac_stime = ni->ac_utime = 0;
110 }
111
112 /*
113  * Account time for a transition between system, hard irq or soft irq state.
114  * Note that this function is called with interrupts enabled.
115  */
116 static cputime_t vtime_delta(struct task_struct *tsk)
117 {
118         struct thread_info *ti = task_thread_info(tsk);
119         cputime_t delta_stime;
120         __u64 now;
121
122         now = ia64_get_itc();
123
124         delta_stime = cycle_to_cputime(ti->ac_stime + (now - ti->ac_stamp));
125         ti->ac_stime = 0;
126         ti->ac_stamp = now;
127
128         return delta_stime;
129 }
130
131 void vtime_account_system(struct task_struct *tsk)
132 {
133         cputime_t delta = vtime_delta(tsk);
134
135         account_system_time(tsk, 0, delta, delta);
136 }
137
138 void vtime_account_idle(struct task_struct *tsk)
139 {
140         account_idle_time(vtime_delta(tsk));
141 }
142
143 #endif /* CONFIG_VIRT_CPU_ACCOUNTING */
144
145 static irqreturn_t
146 timer_interrupt (int irq, void *dev_id)
147 {
148         unsigned long new_itm;
149
150         if (cpu_is_offline(smp_processor_id())) {
151                 return IRQ_HANDLED;
152         }
153
154         platform_timer_interrupt(irq, dev_id);
155
156         new_itm = local_cpu_data->itm_next;
157
158         if (!time_after(ia64_get_itc(), new_itm))
159                 printk(KERN_ERR "Oops: timer tick before it's due (itc=%lx,itm=%lx)\n",
160                        ia64_get_itc(), new_itm);
161
162         profile_tick(CPU_PROFILING);
163
164         if (paravirt_do_steal_accounting(&new_itm))
165                 goto skip_process_time_accounting;
166
167         while (1) {
168                 update_process_times(user_mode(get_irq_regs()));
169
170                 new_itm += local_cpu_data->itm_delta;
171
172                 if (smp_processor_id() == time_keeper_id)
173                         xtime_update(1);
174
175                 local_cpu_data->itm_next = new_itm;
176
177                 if (time_after(new_itm, ia64_get_itc()))
178                         break;
179
180                 /*
181                  * Allow IPIs to interrupt the timer loop.
182                  */
183                 local_irq_enable();
184                 local_irq_disable();
185         }
186
187 skip_process_time_accounting:
188
189         do {
190                 /*
191                  * If we're too close to the next clock tick for
192                  * comfort, we increase the safety margin by
193                  * intentionally dropping the next tick(s).  We do NOT
194                  * update itm.next because that would force us to call
195                  * xtime_update() which in turn would let our clock run
196                  * too fast (with the potentially devastating effect
197                  * of losing monotony of time).
198                  */
199                 while (!time_after(new_itm, ia64_get_itc() + local_cpu_data->itm_delta/2))
200                         new_itm += local_cpu_data->itm_delta;
201                 ia64_set_itm(new_itm);
202                 /* double check, in case we got hit by a (slow) PMI: */
203         } while (time_after_eq(ia64_get_itc(), new_itm));
204         return IRQ_HANDLED;
205 }
206
207 /*
208  * Encapsulate access to the itm structure for SMP.
209  */
210 void
211 ia64_cpu_local_tick (void)
212 {
213         int cpu = smp_processor_id();
214         unsigned long shift = 0, delta;
215
216         /* arrange for the cycle counter to generate a timer interrupt: */
217         ia64_set_itv(IA64_TIMER_VECTOR);
218
219         delta = local_cpu_data->itm_delta;
220         /*
221          * Stagger the timer tick for each CPU so they don't occur all at (almost) the
222          * same time:
223          */
224         if (cpu) {
225                 unsigned long hi = 1UL << ia64_fls(cpu);
226                 shift = (2*(cpu - hi) + 1) * delta/hi/2;
227         }
228         local_cpu_data->itm_next = ia64_get_itc() + delta + shift;
229         ia64_set_itm(local_cpu_data->itm_next);
230 }
231
232 static int nojitter;
233
234 static int __init nojitter_setup(char *str)
235 {
236         nojitter = 1;
237         printk("Jitter checking for ITC timers disabled\n");
238         return 1;
239 }
240
241 __setup("nojitter", nojitter_setup);
242
243
244 void __devinit
245 ia64_init_itm (void)
246 {
247         unsigned long platform_base_freq, itc_freq;
248         struct pal_freq_ratio itc_ratio, proc_ratio;
249         long status, platform_base_drift, itc_drift;
250
251         /*
252          * According to SAL v2.6, we need to use a SAL call to determine the platform base
253          * frequency and then a PAL call to determine the frequency ratio between the ITC
254          * and the base frequency.
255          */
256         status = ia64_sal_freq_base(SAL_FREQ_BASE_PLATFORM,
257                                     &platform_base_freq, &platform_base_drift);
258         if (status != 0) {
259                 printk(KERN_ERR "SAL_FREQ_BASE_PLATFORM failed: %s\n", ia64_sal_strerror(status));
260         } else {
261                 status = ia64_pal_freq_ratios(&proc_ratio, NULL, &itc_ratio);
262                 if (status != 0)
263                         printk(KERN_ERR "PAL_FREQ_RATIOS failed with status=%ld\n", status);
264         }
265         if (status != 0) {
266                 /* invent "random" values */
267                 printk(KERN_ERR
268                        "SAL/PAL failed to obtain frequency info---inventing reasonable values\n");
269                 platform_base_freq = 100000000;
270                 platform_base_drift = -1;       /* no drift info */
271                 itc_ratio.num = 3;
272                 itc_ratio.den = 1;
273         }
274         if (platform_base_freq < 40000000) {
275                 printk(KERN_ERR "Platform base frequency %lu bogus---resetting to 75MHz!\n",
276                        platform_base_freq);
277                 platform_base_freq = 75000000;
278                 platform_base_drift = -1;
279         }
280         if (!proc_ratio.den)
281                 proc_ratio.den = 1;     /* avoid division by zero */
282         if (!itc_ratio.den)
283                 itc_ratio.den = 1;      /* avoid division by zero */
284
285         itc_freq = (platform_base_freq*itc_ratio.num)/itc_ratio.den;
286
287         local_cpu_data->itm_delta = (itc_freq + HZ/2) / HZ;
288         printk(KERN_DEBUG "CPU %d: base freq=%lu.%03luMHz, ITC ratio=%u/%u, "
289                "ITC freq=%lu.%03luMHz", smp_processor_id(),
290                platform_base_freq / 1000000, (platform_base_freq / 1000) % 1000,
291                itc_ratio.num, itc_ratio.den, itc_freq / 1000000, (itc_freq / 1000) % 1000);
292
293         if (platform_base_drift != -1) {
294                 itc_drift = platform_base_drift*itc_ratio.num/itc_ratio.den;
295                 printk("+/-%ldppm\n", itc_drift);
296         } else {
297                 itc_drift = -1;
298                 printk("\n");
299         }
300
301         local_cpu_data->proc_freq = (platform_base_freq*proc_ratio.num)/proc_ratio.den;
302         local_cpu_data->itc_freq = itc_freq;
303         local_cpu_data->cyc_per_usec = (itc_freq + USEC_PER_SEC/2) / USEC_PER_SEC;
304         local_cpu_data->nsec_per_cyc = ((NSEC_PER_SEC<<IA64_NSEC_PER_CYC_SHIFT)
305                                         + itc_freq/2)/itc_freq;
306
307         if (!(sal_platform_features & IA64_SAL_PLATFORM_FEATURE_ITC_DRIFT)) {
308 #ifdef CONFIG_SMP
309                 /* On IA64 in an SMP configuration ITCs are never accurately synchronized.
310                  * Jitter compensation requires a cmpxchg which may limit
311                  * the scalability of the syscalls for retrieving time.
312                  * The ITC synchronization is usually successful to within a few
313                  * ITC ticks but this is not a sure thing. If you need to improve
314                  * timer performance in SMP situations then boot the kernel with the
315                  * "nojitter" option. However, doing so may result in time fluctuating (maybe
316                  * even going backward) if the ITC offsets between the individual CPUs
317                  * are too large.
318                  */
319                 if (!nojitter)
320                         itc_jitter_data.itc_jitter = 1;
321 #endif
322         } else
323                 /*
324                  * ITC is drifty and we have not synchronized the ITCs in smpboot.c.
325                  * ITC values may fluctuate significantly between processors.
326                  * Clock should not be used for hrtimers. Mark itc as only
327                  * useful for boot and testing.
328                  *
329                  * Note that jitter compensation is off! There is no point of
330                  * synchronizing ITCs since they may be large differentials
331                  * that change over time.
332                  *
333                  * The only way to fix this would be to repeatedly sync the
334                  * ITCs. Until that time we have to avoid ITC.
335                  */
336                 clocksource_itc.rating = 50;
337
338         paravirt_init_missing_ticks_accounting(smp_processor_id());
339
340         /* avoid softlock up message when cpu is unplug and plugged again. */
341         touch_softlockup_watchdog();
342
343         /* Setup the CPU local timer tick */
344         ia64_cpu_local_tick();
345
346         if (!itc_clocksource) {
347                 clocksource_register_hz(&clocksource_itc,
348                                                 local_cpu_data->itc_freq);
349                 itc_clocksource = &clocksource_itc;
350         }
351 }
352
353 static cycle_t itc_get_cycles(struct clocksource *cs)
354 {
355         unsigned long lcycle, now, ret;
356
357         if (!itc_jitter_data.itc_jitter)
358                 return get_cycles();
359
360         lcycle = itc_jitter_data.itc_lastcycle;
361         now = get_cycles();
362         if (lcycle && time_after(lcycle, now))
363                 return lcycle;
364
365         /*
366          * Keep track of the last timer value returned.
367          * In an SMP environment, you could lose out in contention of
368          * cmpxchg. If so, your cmpxchg returns new value which the
369          * winner of contention updated to. Use the new value instead.
370          */
371         ret = cmpxchg(&itc_jitter_data.itc_lastcycle, lcycle, now);
372         if (unlikely(ret != lcycle))
373                 return ret;
374
375         return now;
376 }
377
378
379 static struct irqaction timer_irqaction = {
380         .handler =      timer_interrupt,
381         .flags =        IRQF_DISABLED | IRQF_IRQPOLL,
382         .name =         "timer"
383 };
384
385 static struct platform_device rtc_efi_dev = {
386         .name = "rtc-efi",
387         .id = -1,
388 };
389
390 static int __init rtc_init(void)
391 {
392         if (platform_device_register(&rtc_efi_dev) < 0)
393                 printk(KERN_ERR "unable to register rtc device...\n");
394
395         /* not necessarily an error */
396         return 0;
397 }
398 module_init(rtc_init);
399
400 void read_persistent_clock(struct timespec *ts)
401 {
402         efi_gettimeofday(ts);
403 }
404
405 void __init
406 time_init (void)
407 {
408         register_percpu_irq(IA64_TIMER_VECTOR, &timer_irqaction);
409         ia64_init_itm();
410 }
411
412 /*
413  * Generic udelay assumes that if preemption is allowed and the thread
414  * migrates to another CPU, that the ITC values are synchronized across
415  * all CPUs.
416  */
417 static void
418 ia64_itc_udelay (unsigned long usecs)
419 {
420         unsigned long start = ia64_get_itc();
421         unsigned long end = start + usecs*local_cpu_data->cyc_per_usec;
422
423         while (time_before(ia64_get_itc(), end))
424                 cpu_relax();
425 }
426
427 void (*ia64_udelay)(unsigned long usecs) = &ia64_itc_udelay;
428
429 void
430 udelay (unsigned long usecs)
431 {
432         (*ia64_udelay)(usecs);
433 }
434 EXPORT_SYMBOL(udelay);
435
436 /* IA64 doesn't cache the timezone */
437 void update_vsyscall_tz(void)
438 {
439 }
440
441 void update_vsyscall_old(struct timespec *wall, struct timespec *wtm,
442                         struct clocksource *c, u32 mult)
443 {
444         write_seqcount_begin(&fsyscall_gtod_data.seq);
445
446         /* copy fsyscall clock data */
447         fsyscall_gtod_data.clk_mask = c->mask;
448         fsyscall_gtod_data.clk_mult = mult;
449         fsyscall_gtod_data.clk_shift = c->shift;
450         fsyscall_gtod_data.clk_fsys_mmio = c->archdata.fsys_mmio;
451         fsyscall_gtod_data.clk_cycle_last = c->cycle_last;
452
453         /* copy kernel time structures */
454         fsyscall_gtod_data.wall_time.tv_sec = wall->tv_sec;
455         fsyscall_gtod_data.wall_time.tv_nsec = wall->tv_nsec;
456         fsyscall_gtod_data.monotonic_time.tv_sec = wtm->tv_sec
457                                                         + wall->tv_sec;
458         fsyscall_gtod_data.monotonic_time.tv_nsec = wtm->tv_nsec
459                                                         + wall->tv_nsec;
460
461         /* normalize */
462         while (fsyscall_gtod_data.monotonic_time.tv_nsec >= NSEC_PER_SEC) {
463                 fsyscall_gtod_data.monotonic_time.tv_nsec -= NSEC_PER_SEC;
464                 fsyscall_gtod_data.monotonic_time.tv_sec++;
465         }
466
467         write_seqcount_end(&fsyscall_gtod_data.seq);
468 }
469