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