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