seqlock: Get rid of SEQLOCK_UNLOCKED
[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(struct clocksource *cs);
37
38 struct fsyscall_gtod_data_t fsyscall_gtod_data = {
39         .lock = __SEQLOCK_UNLOCKED(fsyscall_gtod_data.lock),
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(struct clocksource *cs)
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         .flags          = CLOCK_SOURCE_IS_CONTINUOUS,
77 #ifdef CONFIG_PARAVIRT
78         .resume         = paravirt_clocksource_resume,
79 #endif
80 };
81 static struct clocksource *itc_clocksource;
82
83 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
84
85 #include <linux/kernel_stat.h>
86
87 extern cputime_t cycle_to_cputime(u64 cyc);
88
89 /*
90  * Called from the context switch with interrupts disabled, to charge all
91  * accumulated times to the current process, and to prepare accounting on
92  * the next process.
93  */
94 void ia64_account_on_switch(struct task_struct *prev, struct task_struct *next)
95 {
96         struct thread_info *pi = task_thread_info(prev);
97         struct thread_info *ni = task_thread_info(next);
98         cputime_t delta_stime, delta_utime;
99         __u64 now;
100
101         now = ia64_get_itc();
102
103         delta_stime = cycle_to_cputime(pi->ac_stime + (now - pi->ac_stamp));
104         if (idle_task(smp_processor_id()) != prev)
105                 account_system_time(prev, 0, delta_stime, delta_stime);
106         else
107                 account_idle_time(delta_stime);
108
109         if (pi->ac_utime) {
110                 delta_utime = cycle_to_cputime(pi->ac_utime);
111                 account_user_time(prev, delta_utime, delta_utime);
112         }
113
114         pi->ac_stamp = ni->ac_stamp = now;
115         ni->ac_stime = ni->ac_utime = 0;
116 }
117
118 /*
119  * Account time for a transition between system, hard irq or soft irq state.
120  * Note that this function is called with interrupts enabled.
121  */
122 void account_system_vtime(struct task_struct *tsk)
123 {
124         struct thread_info *ti = task_thread_info(tsk);
125         unsigned long flags;
126         cputime_t delta_stime;
127         __u64 now;
128
129         local_irq_save(flags);
130
131         now = ia64_get_itc();
132
133         delta_stime = cycle_to_cputime(ti->ac_stime + (now - ti->ac_stamp));
134         if (irq_count() || idle_task(smp_processor_id()) != tsk)
135                 account_system_time(tsk, 0, delta_stime, delta_stime);
136         else
137                 account_idle_time(delta_stime);
138         ti->ac_stime = 0;
139
140         ti->ac_stamp = now;
141
142         local_irq_restore(flags);
143 }
144 EXPORT_SYMBOL_GPL(account_system_vtime);
145
146 /*
147  * Called from the timer interrupt handler to charge accumulated user time
148  * to the current process.  Must be called with interrupts disabled.
149  */
150 void account_process_tick(struct task_struct *p, int user_tick)
151 {
152         struct thread_info *ti = task_thread_info(p);
153         cputime_t delta_utime;
154
155         if (ti->ac_utime) {
156                 delta_utime = cycle_to_cputime(ti->ac_utime);
157                 account_user_time(p, delta_utime, delta_utime);
158                 ti->ac_utime = 0;
159         }
160 }
161
162 #endif /* CONFIG_VIRT_CPU_ACCOUNTING */
163
164 static irqreturn_t
165 timer_interrupt (int irq, void *dev_id)
166 {
167         unsigned long new_itm;
168
169         if (cpu_is_offline(smp_processor_id())) {
170                 return IRQ_HANDLED;
171         }
172
173         platform_timer_interrupt(irq, dev_id);
174
175         new_itm = local_cpu_data->itm_next;
176
177         if (!time_after(ia64_get_itc(), new_itm))
178                 printk(KERN_ERR "Oops: timer tick before it's due (itc=%lx,itm=%lx)\n",
179                        ia64_get_itc(), new_itm);
180
181         profile_tick(CPU_PROFILING);
182
183         if (paravirt_do_steal_accounting(&new_itm))
184                 goto skip_process_time_accounting;
185
186         while (1) {
187                 update_process_times(user_mode(get_irq_regs()));
188
189                 new_itm += local_cpu_data->itm_delta;
190
191                 if (smp_processor_id() == time_keeper_id)
192                         xtime_update(1);
193
194                 local_cpu_data->itm_next = new_itm;
195
196                 if (time_after(new_itm, ia64_get_itc()))
197                         break;
198
199                 /*
200                  * Allow IPIs to interrupt the timer loop.
201                  */
202                 local_irq_enable();
203                 local_irq_disable();
204         }
205
206 skip_process_time_accounting:
207
208         do {
209                 /*
210                  * If we're too close to the next clock tick for
211                  * comfort, we increase the safety margin by
212                  * intentionally dropping the next tick(s).  We do NOT
213                  * update itm.next because that would force us to call
214                  * xtime_update() which in turn would let our clock run
215                  * too fast (with the potentially devastating effect
216                  * of losing monotony of time).
217                  */
218                 while (!time_after(new_itm, ia64_get_itc() + local_cpu_data->itm_delta/2))
219                         new_itm += local_cpu_data->itm_delta;
220                 ia64_set_itm(new_itm);
221                 /* double check, in case we got hit by a (slow) PMI: */
222         } while (time_after_eq(ia64_get_itc(), new_itm));
223         return IRQ_HANDLED;
224 }
225
226 /*
227  * Encapsulate access to the itm structure for SMP.
228  */
229 void
230 ia64_cpu_local_tick (void)
231 {
232         int cpu = smp_processor_id();
233         unsigned long shift = 0, delta;
234
235         /* arrange for the cycle counter to generate a timer interrupt: */
236         ia64_set_itv(IA64_TIMER_VECTOR);
237
238         delta = local_cpu_data->itm_delta;
239         /*
240          * Stagger the timer tick for each CPU so they don't occur all at (almost) the
241          * same time:
242          */
243         if (cpu) {
244                 unsigned long hi = 1UL << ia64_fls(cpu);
245                 shift = (2*(cpu - hi) + 1) * delta/hi/2;
246         }
247         local_cpu_data->itm_next = ia64_get_itc() + delta + shift;
248         ia64_set_itm(local_cpu_data->itm_next);
249 }
250
251 static int nojitter;
252
253 static int __init nojitter_setup(char *str)
254 {
255         nojitter = 1;
256         printk("Jitter checking for ITC timers disabled\n");
257         return 1;
258 }
259
260 __setup("nojitter", nojitter_setup);
261
262
263 void __devinit
264 ia64_init_itm (void)
265 {
266         unsigned long platform_base_freq, itc_freq;
267         struct pal_freq_ratio itc_ratio, proc_ratio;
268         long status, platform_base_drift, itc_drift;
269
270         /*
271          * According to SAL v2.6, we need to use a SAL call to determine the platform base
272          * frequency and then a PAL call to determine the frequency ratio between the ITC
273          * and the base frequency.
274          */
275         status = ia64_sal_freq_base(SAL_FREQ_BASE_PLATFORM,
276                                     &platform_base_freq, &platform_base_drift);
277         if (status != 0) {
278                 printk(KERN_ERR "SAL_FREQ_BASE_PLATFORM failed: %s\n", ia64_sal_strerror(status));
279         } else {
280                 status = ia64_pal_freq_ratios(&proc_ratio, NULL, &itc_ratio);
281                 if (status != 0)
282                         printk(KERN_ERR "PAL_FREQ_RATIOS failed with status=%ld\n", status);
283         }
284         if (status != 0) {
285                 /* invent "random" values */
286                 printk(KERN_ERR
287                        "SAL/PAL failed to obtain frequency info---inventing reasonable values\n");
288                 platform_base_freq = 100000000;
289                 platform_base_drift = -1;       /* no drift info */
290                 itc_ratio.num = 3;
291                 itc_ratio.den = 1;
292         }
293         if (platform_base_freq < 40000000) {
294                 printk(KERN_ERR "Platform base frequency %lu bogus---resetting to 75MHz!\n",
295                        platform_base_freq);
296                 platform_base_freq = 75000000;
297                 platform_base_drift = -1;
298         }
299         if (!proc_ratio.den)
300                 proc_ratio.den = 1;     /* avoid division by zero */
301         if (!itc_ratio.den)
302                 itc_ratio.den = 1;      /* avoid division by zero */
303
304         itc_freq = (platform_base_freq*itc_ratio.num)/itc_ratio.den;
305
306         local_cpu_data->itm_delta = (itc_freq + HZ/2) / HZ;
307         printk(KERN_DEBUG "CPU %d: base freq=%lu.%03luMHz, ITC ratio=%u/%u, "
308                "ITC freq=%lu.%03luMHz", smp_processor_id(),
309                platform_base_freq / 1000000, (platform_base_freq / 1000) % 1000,
310                itc_ratio.num, itc_ratio.den, itc_freq / 1000000, (itc_freq / 1000) % 1000);
311
312         if (platform_base_drift != -1) {
313                 itc_drift = platform_base_drift*itc_ratio.num/itc_ratio.den;
314                 printk("+/-%ldppm\n", itc_drift);
315         } else {
316                 itc_drift = -1;
317                 printk("\n");
318         }
319
320         local_cpu_data->proc_freq = (platform_base_freq*proc_ratio.num)/proc_ratio.den;
321         local_cpu_data->itc_freq = itc_freq;
322         local_cpu_data->cyc_per_usec = (itc_freq + USEC_PER_SEC/2) / USEC_PER_SEC;
323         local_cpu_data->nsec_per_cyc = ((NSEC_PER_SEC<<IA64_NSEC_PER_CYC_SHIFT)
324                                         + itc_freq/2)/itc_freq;
325
326         if (!(sal_platform_features & IA64_SAL_PLATFORM_FEATURE_ITC_DRIFT)) {
327 #ifdef CONFIG_SMP
328                 /* On IA64 in an SMP configuration ITCs are never accurately synchronized.
329                  * Jitter compensation requires a cmpxchg which may limit
330                  * the scalability of the syscalls for retrieving time.
331                  * The ITC synchronization is usually successful to within a few
332                  * ITC ticks but this is not a sure thing. If you need to improve
333                  * timer performance in SMP situations then boot the kernel with the
334                  * "nojitter" option. However, doing so may result in time fluctuating (maybe
335                  * even going backward) if the ITC offsets between the individual CPUs
336                  * are too large.
337                  */
338                 if (!nojitter)
339                         itc_jitter_data.itc_jitter = 1;
340 #endif
341         } else
342                 /*
343                  * ITC is drifty and we have not synchronized the ITCs in smpboot.c.
344                  * ITC values may fluctuate significantly between processors.
345                  * Clock should not be used for hrtimers. Mark itc as only
346                  * useful for boot and testing.
347                  *
348                  * Note that jitter compensation is off! There is no point of
349                  * synchronizing ITCs since they may be large differentials
350                  * that change over time.
351                  *
352                  * The only way to fix this would be to repeatedly sync the
353                  * ITCs. Until that time we have to avoid ITC.
354                  */
355                 clocksource_itc.rating = 50;
356
357         paravirt_init_missing_ticks_accounting(smp_processor_id());
358
359         /* avoid softlock up message when cpu is unplug and plugged again. */
360         touch_softlockup_watchdog();
361
362         /* Setup the CPU local timer tick */
363         ia64_cpu_local_tick();
364
365         if (!itc_clocksource) {
366                 clocksource_register_hz(&clocksource_itc,
367                                                 local_cpu_data->itc_freq);
368                 itc_clocksource = &clocksource_itc;
369         }
370 }
371
372 static cycle_t itc_get_cycles(struct clocksource *cs)
373 {
374         unsigned long lcycle, now, ret;
375
376         if (!itc_jitter_data.itc_jitter)
377                 return get_cycles();
378
379         lcycle = itc_jitter_data.itc_lastcycle;
380         now = get_cycles();
381         if (lcycle && time_after(lcycle, now))
382                 return lcycle;
383
384         /*
385          * Keep track of the last timer value returned.
386          * In an SMP environment, you could lose out in contention of
387          * cmpxchg. If so, your cmpxchg returns new value which the
388          * winner of contention updated to. Use the new value instead.
389          */
390         ret = cmpxchg(&itc_jitter_data.itc_lastcycle, lcycle, now);
391         if (unlikely(ret != lcycle))
392                 return ret;
393
394         return now;
395 }
396
397
398 static struct irqaction timer_irqaction = {
399         .handler =      timer_interrupt,
400         .flags =        IRQF_DISABLED | IRQF_IRQPOLL,
401         .name =         "timer"
402 };
403
404 static struct platform_device rtc_efi_dev = {
405         .name = "rtc-efi",
406         .id = -1,
407 };
408
409 static int __init rtc_init(void)
410 {
411         if (platform_device_register(&rtc_efi_dev) < 0)
412                 printk(KERN_ERR "unable to register rtc device...\n");
413
414         /* not necessarily an error */
415         return 0;
416 }
417 module_init(rtc_init);
418
419 void read_persistent_clock(struct timespec *ts)
420 {
421         efi_gettimeofday(ts);
422 }
423
424 void __init
425 time_init (void)
426 {
427         register_percpu_irq(IA64_TIMER_VECTOR, &timer_irqaction);
428         ia64_init_itm();
429 }
430
431 /*
432  * Generic udelay assumes that if preemption is allowed and the thread
433  * migrates to another CPU, that the ITC values are synchronized across
434  * all CPUs.
435  */
436 static void
437 ia64_itc_udelay (unsigned long usecs)
438 {
439         unsigned long start = ia64_get_itc();
440         unsigned long end = start + usecs*local_cpu_data->cyc_per_usec;
441
442         while (time_before(ia64_get_itc(), end))
443                 cpu_relax();
444 }
445
446 void (*ia64_udelay)(unsigned long usecs) = &ia64_itc_udelay;
447
448 void
449 udelay (unsigned long usecs)
450 {
451         (*ia64_udelay)(usecs);
452 }
453 EXPORT_SYMBOL(udelay);
454
455 /* IA64 doesn't cache the timezone */
456 void update_vsyscall_tz(void)
457 {
458 }
459
460 void update_vsyscall(struct timespec *wall, struct timespec *wtm,
461                         struct clocksource *c, u32 mult)
462 {
463         unsigned long flags;
464
465         write_seqlock_irqsave(&fsyscall_gtod_data.lock, flags);
466
467         /* copy fsyscall clock data */
468         fsyscall_gtod_data.clk_mask = c->mask;
469         fsyscall_gtod_data.clk_mult = mult;
470         fsyscall_gtod_data.clk_shift = c->shift;
471         fsyscall_gtod_data.clk_fsys_mmio = c->fsys_mmio;
472         fsyscall_gtod_data.clk_cycle_last = c->cycle_last;
473
474         /* copy kernel time structures */
475         fsyscall_gtod_data.wall_time.tv_sec = wall->tv_sec;
476         fsyscall_gtod_data.wall_time.tv_nsec = wall->tv_nsec;
477         fsyscall_gtod_data.monotonic_time.tv_sec = wtm->tv_sec
478                                                         + wall->tv_sec;
479         fsyscall_gtod_data.monotonic_time.tv_nsec = wtm->tv_nsec
480                                                         + wall->tv_nsec;
481
482         /* normalize */
483         while (fsyscall_gtod_data.monotonic_time.tv_nsec >= NSEC_PER_SEC) {
484                 fsyscall_gtod_data.monotonic_time.tv_nsec -= NSEC_PER_SEC;
485                 fsyscall_gtod_data.monotonic_time.tv_sec++;
486         }
487
488         write_sequnlock_irqrestore(&fsyscall_gtod_data.lock, flags);
489 }
490