Merge branch 'linux-next' of git://git.kernel.org/pub/scm/linux/kernel/git/jbarnes...
[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/ptrace.h>
28 #include <asm/sal.h>
29 #include <asm/sections.h>
30 #include <asm/system.h>
31
32 #include "fsyscall_gtod_data.h"
33
34 static cycle_t itc_get_cycles(void);
35
36 struct fsyscall_gtod_data_t fsyscall_gtod_data = {
37         .lock = SEQLOCK_UNLOCKED,
38 };
39
40 struct itc_jitter_data_t itc_jitter_data;
41
42 volatile int time_keeper_id = 0; /* smp_processor_id() of time-keeper */
43
44 #ifdef CONFIG_IA64_DEBUG_IRQ
45
46 unsigned long last_cli_ip;
47 EXPORT_SYMBOL(last_cli_ip);
48
49 #endif
50
51 static struct clocksource clocksource_itc = {
52         .name           = "itc",
53         .rating         = 350,
54         .read           = itc_get_cycles,
55         .mask           = CLOCKSOURCE_MASK(64),
56         .mult           = 0, /*to be calculated*/
57         .shift          = 16,
58         .flags          = CLOCK_SOURCE_IS_CONTINUOUS,
59 };
60 static struct clocksource *itc_clocksource;
61
62 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
63
64 #include <linux/kernel_stat.h>
65
66 extern cputime_t cycle_to_cputime(u64 cyc);
67
68 /*
69  * Called from the context switch with interrupts disabled, to charge all
70  * accumulated times to the current process, and to prepare accounting on
71  * the next process.
72  */
73 void ia64_account_on_switch(struct task_struct *prev, struct task_struct *next)
74 {
75         struct thread_info *pi = task_thread_info(prev);
76         struct thread_info *ni = task_thread_info(next);
77         cputime_t delta_stime, delta_utime;
78         __u64 now;
79
80         now = ia64_get_itc();
81
82         delta_stime = cycle_to_cputime(pi->ac_stime + (now - pi->ac_stamp));
83         account_system_time(prev, 0, delta_stime);
84         account_system_time_scaled(prev, delta_stime);
85
86         if (pi->ac_utime) {
87                 delta_utime = cycle_to_cputime(pi->ac_utime);
88                 account_user_time(prev, delta_utime);
89                 account_user_time_scaled(prev, delta_utime);
90         }
91
92         pi->ac_stamp = ni->ac_stamp = now;
93         ni->ac_stime = ni->ac_utime = 0;
94 }
95
96 /*
97  * Account time for a transition between system, hard irq or soft irq state.
98  * Note that this function is called with interrupts enabled.
99  */
100 void account_system_vtime(struct task_struct *tsk)
101 {
102         struct thread_info *ti = task_thread_info(tsk);
103         unsigned long flags;
104         cputime_t delta_stime;
105         __u64 now;
106
107         local_irq_save(flags);
108
109         now = ia64_get_itc();
110
111         delta_stime = cycle_to_cputime(ti->ac_stime + (now - ti->ac_stamp));
112         account_system_time(tsk, 0, delta_stime);
113         account_system_time_scaled(tsk, delta_stime);
114         ti->ac_stime = 0;
115
116         ti->ac_stamp = now;
117
118         local_irq_restore(flags);
119 }
120 EXPORT_SYMBOL_GPL(account_system_vtime);
121
122 /*
123  * Called from the timer interrupt handler to charge accumulated user time
124  * to the current process.  Must be called with interrupts disabled.
125  */
126 void account_process_tick(struct task_struct *p, int user_tick)
127 {
128         struct thread_info *ti = task_thread_info(p);
129         cputime_t delta_utime;
130
131         if (ti->ac_utime) {
132                 delta_utime = cycle_to_cputime(ti->ac_utime);
133                 account_user_time(p, delta_utime);
134                 account_user_time_scaled(p, delta_utime);
135                 ti->ac_utime = 0;
136         }
137 }
138
139 #endif /* CONFIG_VIRT_CPU_ACCOUNTING */
140
141 static irqreturn_t
142 timer_interrupt (int irq, void *dev_id)
143 {
144         unsigned long new_itm;
145
146         if (unlikely(cpu_is_offline(smp_processor_id()))) {
147                 return IRQ_HANDLED;
148         }
149
150         platform_timer_interrupt(irq, dev_id);
151
152         new_itm = local_cpu_data->itm_next;
153
154         if (!time_after(ia64_get_itc(), new_itm))
155                 printk(KERN_ERR "Oops: timer tick before it's due (itc=%lx,itm=%lx)\n",
156                        ia64_get_itc(), new_itm);
157
158         profile_tick(CPU_PROFILING);
159
160         while (1) {
161                 update_process_times(user_mode(get_irq_regs()));
162
163                 new_itm += local_cpu_data->itm_delta;
164
165                 if (smp_processor_id() == time_keeper_id) {
166                         /*
167                          * Here we are in the timer irq handler. We have irqs locally
168                          * disabled, but we don't know if the timer_bh is running on
169                          * another CPU. We need to avoid to SMP race by acquiring the
170                          * xtime_lock.
171                          */
172                         write_seqlock(&xtime_lock);
173                         do_timer(1);
174                         local_cpu_data->itm_next = new_itm;
175                         write_sequnlock(&xtime_lock);
176                 } else
177                         local_cpu_data->itm_next = new_itm;
178
179                 if (time_after(new_itm, ia64_get_itc()))
180                         break;
181
182                 /*
183                  * Allow IPIs to interrupt the timer loop.
184                  */
185                 local_irq_enable();
186                 local_irq_disable();
187         }
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                  * do_timer() 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         /* Setup the CPU local timer tick */
339         ia64_cpu_local_tick();
340
341         if (!itc_clocksource) {
342                 /* Sort out mult/shift values: */
343                 clocksource_itc.mult =
344                         clocksource_hz2mult(local_cpu_data->itc_freq,
345                                                 clocksource_itc.shift);
346                 clocksource_register(&clocksource_itc);
347                 itc_clocksource = &clocksource_itc;
348         }
349 }
350
351 static cycle_t itc_get_cycles(void)
352 {
353         u64 lcycle, now, ret;
354
355         if (!itc_jitter_data.itc_jitter)
356                 return get_cycles();
357
358         lcycle = itc_jitter_data.itc_lastcycle;
359         now = get_cycles();
360         if (lcycle && time_after(lcycle, now))
361                 return lcycle;
362
363         /*
364          * Keep track of the last timer value returned.
365          * In an SMP environment, you could lose out in contention of
366          * cmpxchg. If so, your cmpxchg returns new value which the
367          * winner of contention updated to. Use the new value instead.
368          */
369         ret = cmpxchg(&itc_jitter_data.itc_lastcycle, lcycle, now);
370         if (unlikely(ret != lcycle))
371                 return ret;
372
373         return now;
374 }
375
376
377 static struct irqaction timer_irqaction = {
378         .handler =      timer_interrupt,
379         .flags =        IRQF_DISABLED | IRQF_IRQPOLL,
380         .name =         "timer"
381 };
382
383 void __init
384 time_init (void)
385 {
386         register_percpu_irq(IA64_TIMER_VECTOR, &timer_irqaction);
387         efi_gettimeofday(&xtime);
388         ia64_init_itm();
389
390         /*
391          * Initialize wall_to_monotonic such that adding it to xtime will yield zero, the
392          * tv_nsec field must be normalized (i.e., 0 <= nsec < NSEC_PER_SEC).
393          */
394         set_normalized_timespec(&wall_to_monotonic, -xtime.tv_sec, -xtime.tv_nsec);
395 }
396
397 /*
398  * Generic udelay assumes that if preemption is allowed and the thread
399  * migrates to another CPU, that the ITC values are synchronized across
400  * all CPUs.
401  */
402 static void
403 ia64_itc_udelay (unsigned long usecs)
404 {
405         unsigned long start = ia64_get_itc();
406         unsigned long end = start + usecs*local_cpu_data->cyc_per_usec;
407
408         while (time_before(ia64_get_itc(), end))
409                 cpu_relax();
410 }
411
412 void (*ia64_udelay)(unsigned long usecs) = &ia64_itc_udelay;
413
414 void
415 udelay (unsigned long usecs)
416 {
417         (*ia64_udelay)(usecs);
418 }
419 EXPORT_SYMBOL(udelay);
420
421 /* IA64 doesn't cache the timezone */
422 void update_vsyscall_tz(void)
423 {
424 }
425
426 void update_vsyscall(struct timespec *wall, struct clocksource *c)
427 {
428         unsigned long flags;
429
430         write_seqlock_irqsave(&fsyscall_gtod_data.lock, flags);
431
432         /* copy fsyscall clock data */
433         fsyscall_gtod_data.clk_mask = c->mask;
434         fsyscall_gtod_data.clk_mult = c->mult;
435         fsyscall_gtod_data.clk_shift = c->shift;
436         fsyscall_gtod_data.clk_fsys_mmio = c->fsys_mmio;
437         fsyscall_gtod_data.clk_cycle_last = c->cycle_last;
438
439         /* copy kernel time structures */
440         fsyscall_gtod_data.wall_time.tv_sec = wall->tv_sec;
441         fsyscall_gtod_data.wall_time.tv_nsec = wall->tv_nsec;
442         fsyscall_gtod_data.monotonic_time.tv_sec = wall_to_monotonic.tv_sec
443                                                         + wall->tv_sec;
444         fsyscall_gtod_data.monotonic_time.tv_nsec = wall_to_monotonic.tv_nsec
445                                                         + wall->tv_nsec;
446
447         /* normalize */
448         while (fsyscall_gtod_data.monotonic_time.tv_nsec >= NSEC_PER_SEC) {
449                 fsyscall_gtod_data.monotonic_time.tv_nsec -= NSEC_PER_SEC;
450                 fsyscall_gtod_data.monotonic_time.tv_sec++;
451         }
452
453         write_sequnlock_irqrestore(&fsyscall_gtod_data.lock, flags);
454 }
455