17fda5293c6716392d17a62bc24329e759df8415
[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
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 static irqreturn_t
63 timer_interrupt (int irq, void *dev_id)
64 {
65         unsigned long new_itm;
66
67         if (unlikely(cpu_is_offline(smp_processor_id()))) {
68                 return IRQ_HANDLED;
69         }
70
71         platform_timer_interrupt(irq, dev_id);
72
73         new_itm = local_cpu_data->itm_next;
74
75         if (!time_after(ia64_get_itc(), new_itm))
76                 printk(KERN_ERR "Oops: timer tick before it's due (itc=%lx,itm=%lx)\n",
77                        ia64_get_itc(), new_itm);
78
79         profile_tick(CPU_PROFILING);
80
81         while (1) {
82                 update_process_times(user_mode(get_irq_regs()));
83
84                 new_itm += local_cpu_data->itm_delta;
85
86                 if (smp_processor_id() == time_keeper_id) {
87                         /*
88                          * Here we are in the timer irq handler. We have irqs locally
89                          * disabled, but we don't know if the timer_bh is running on
90                          * another CPU. We need to avoid to SMP race by acquiring the
91                          * xtime_lock.
92                          */
93                         write_seqlock(&xtime_lock);
94                         do_timer(1);
95                         local_cpu_data->itm_next = new_itm;
96                         write_sequnlock(&xtime_lock);
97                 } else
98                         local_cpu_data->itm_next = new_itm;
99
100                 if (time_after(new_itm, ia64_get_itc()))
101                         break;
102
103                 /*
104                  * Allow IPIs to interrupt the timer loop.
105                  */
106                 local_irq_enable();
107                 local_irq_disable();
108         }
109
110         do {
111                 /*
112                  * If we're too close to the next clock tick for
113                  * comfort, we increase the safety margin by
114                  * intentionally dropping the next tick(s).  We do NOT
115                  * update itm.next because that would force us to call
116                  * do_timer() which in turn would let our clock run
117                  * too fast (with the potentially devastating effect
118                  * of losing monotony of time).
119                  */
120                 while (!time_after(new_itm, ia64_get_itc() + local_cpu_data->itm_delta/2))
121                         new_itm += local_cpu_data->itm_delta;
122                 ia64_set_itm(new_itm);
123                 /* double check, in case we got hit by a (slow) PMI: */
124         } while (time_after_eq(ia64_get_itc(), new_itm));
125         return IRQ_HANDLED;
126 }
127
128 /*
129  * Encapsulate access to the itm structure for SMP.
130  */
131 void
132 ia64_cpu_local_tick (void)
133 {
134         int cpu = smp_processor_id();
135         unsigned long shift = 0, delta;
136
137         /* arrange for the cycle counter to generate a timer interrupt: */
138         ia64_set_itv(IA64_TIMER_VECTOR);
139
140         delta = local_cpu_data->itm_delta;
141         /*
142          * Stagger the timer tick for each CPU so they don't occur all at (almost) the
143          * same time:
144          */
145         if (cpu) {
146                 unsigned long hi = 1UL << ia64_fls(cpu);
147                 shift = (2*(cpu - hi) + 1) * delta/hi/2;
148         }
149         local_cpu_data->itm_next = ia64_get_itc() + delta + shift;
150         ia64_set_itm(local_cpu_data->itm_next);
151 }
152
153 static int nojitter;
154
155 static int __init nojitter_setup(char *str)
156 {
157         nojitter = 1;
158         printk("Jitter checking for ITC timers disabled\n");
159         return 1;
160 }
161
162 __setup("nojitter", nojitter_setup);
163
164
165 void __devinit
166 ia64_init_itm (void)
167 {
168         unsigned long platform_base_freq, itc_freq;
169         struct pal_freq_ratio itc_ratio, proc_ratio;
170         long status, platform_base_drift, itc_drift;
171
172         /*
173          * According to SAL v2.6, we need to use a SAL call to determine the platform base
174          * frequency and then a PAL call to determine the frequency ratio between the ITC
175          * and the base frequency.
176          */
177         status = ia64_sal_freq_base(SAL_FREQ_BASE_PLATFORM,
178                                     &platform_base_freq, &platform_base_drift);
179         if (status != 0) {
180                 printk(KERN_ERR "SAL_FREQ_BASE_PLATFORM failed: %s\n", ia64_sal_strerror(status));
181         } else {
182                 status = ia64_pal_freq_ratios(&proc_ratio, NULL, &itc_ratio);
183                 if (status != 0)
184                         printk(KERN_ERR "PAL_FREQ_RATIOS failed with status=%ld\n", status);
185         }
186         if (status != 0) {
187                 /* invent "random" values */
188                 printk(KERN_ERR
189                        "SAL/PAL failed to obtain frequency info---inventing reasonable values\n");
190                 platform_base_freq = 100000000;
191                 platform_base_drift = -1;       /* no drift info */
192                 itc_ratio.num = 3;
193                 itc_ratio.den = 1;
194         }
195         if (platform_base_freq < 40000000) {
196                 printk(KERN_ERR "Platform base frequency %lu bogus---resetting to 75MHz!\n",
197                        platform_base_freq);
198                 platform_base_freq = 75000000;
199                 platform_base_drift = -1;
200         }
201         if (!proc_ratio.den)
202                 proc_ratio.den = 1;     /* avoid division by zero */
203         if (!itc_ratio.den)
204                 itc_ratio.den = 1;      /* avoid division by zero */
205
206         itc_freq = (platform_base_freq*itc_ratio.num)/itc_ratio.den;
207
208         local_cpu_data->itm_delta = (itc_freq + HZ/2) / HZ;
209         printk(KERN_DEBUG "CPU %d: base freq=%lu.%03luMHz, ITC ratio=%u/%u, "
210                "ITC freq=%lu.%03luMHz", smp_processor_id(),
211                platform_base_freq / 1000000, (platform_base_freq / 1000) % 1000,
212                itc_ratio.num, itc_ratio.den, itc_freq / 1000000, (itc_freq / 1000) % 1000);
213
214         if (platform_base_drift != -1) {
215                 itc_drift = platform_base_drift*itc_ratio.num/itc_ratio.den;
216                 printk("+/-%ldppm\n", itc_drift);
217         } else {
218                 itc_drift = -1;
219                 printk("\n");
220         }
221
222         local_cpu_data->proc_freq = (platform_base_freq*proc_ratio.num)/proc_ratio.den;
223         local_cpu_data->itc_freq = itc_freq;
224         local_cpu_data->cyc_per_usec = (itc_freq + USEC_PER_SEC/2) / USEC_PER_SEC;
225         local_cpu_data->nsec_per_cyc = ((NSEC_PER_SEC<<IA64_NSEC_PER_CYC_SHIFT)
226                                         + itc_freq/2)/itc_freq;
227
228         if (!(sal_platform_features & IA64_SAL_PLATFORM_FEATURE_ITC_DRIFT)) {
229 #ifdef CONFIG_SMP
230                 /* On IA64 in an SMP configuration ITCs are never accurately synchronized.
231                  * Jitter compensation requires a cmpxchg which may limit
232                  * the scalability of the syscalls for retrieving time.
233                  * The ITC synchronization is usually successful to within a few
234                  * ITC ticks but this is not a sure thing. If you need to improve
235                  * timer performance in SMP situations then boot the kernel with the
236                  * "nojitter" option. However, doing so may result in time fluctuating (maybe
237                  * even going backward) if the ITC offsets between the individual CPUs
238                  * are too large.
239                  */
240                 if (!nojitter)
241                         itc_jitter_data.itc_jitter = 1;
242 #endif
243         } else
244                 /*
245                  * ITC is drifty and we have not synchronized the ITCs in smpboot.c.
246                  * ITC values may fluctuate significantly between processors.
247                  * Clock should not be used for hrtimers. Mark itc as only
248                  * useful for boot and testing.
249                  *
250                  * Note that jitter compensation is off! There is no point of
251                  * synchronizing ITCs since they may be large differentials
252                  * that change over time.
253                  *
254                  * The only way to fix this would be to repeatedly sync the
255                  * ITCs. Until that time we have to avoid ITC.
256                  */
257                 clocksource_itc.rating = 50;
258
259         /* Setup the CPU local timer tick */
260         ia64_cpu_local_tick();
261
262         if (!itc_clocksource) {
263                 /* Sort out mult/shift values: */
264                 clocksource_itc.mult =
265                         clocksource_hz2mult(local_cpu_data->itc_freq,
266                                                 clocksource_itc.shift);
267                 clocksource_register(&clocksource_itc);
268                 itc_clocksource = &clocksource_itc;
269         }
270 }
271
272 static cycle_t itc_get_cycles(void)
273 {
274         u64 lcycle, now, ret;
275
276         if (!itc_jitter_data.itc_jitter)
277                 return get_cycles();
278
279         lcycle = itc_jitter_data.itc_lastcycle;
280         now = get_cycles();
281         if (lcycle && time_after(lcycle, now))
282                 return lcycle;
283
284         /*
285          * Keep track of the last timer value returned.
286          * In an SMP environment, you could lose out in contention of
287          * cmpxchg. If so, your cmpxchg returns new value which the
288          * winner of contention updated to. Use the new value instead.
289          */
290         ret = cmpxchg(&itc_jitter_data.itc_lastcycle, lcycle, now);
291         if (unlikely(ret != lcycle))
292                 return ret;
293
294         return now;
295 }
296
297
298 static struct irqaction timer_irqaction = {
299         .handler =      timer_interrupt,
300         .flags =        IRQF_DISABLED | IRQF_IRQPOLL,
301         .name =         "timer"
302 };
303
304 void __devinit ia64_disable_timer(void)
305 {
306         ia64_set_itv(1 << 16);
307 }
308
309 void __init
310 time_init (void)
311 {
312         register_percpu_irq(IA64_TIMER_VECTOR, &timer_irqaction);
313         efi_gettimeofday(&xtime);
314         ia64_init_itm();
315
316         /*
317          * Initialize wall_to_monotonic such that adding it to xtime will yield zero, the
318          * tv_nsec field must be normalized (i.e., 0 <= nsec < NSEC_PER_SEC).
319          */
320         set_normalized_timespec(&wall_to_monotonic, -xtime.tv_sec, -xtime.tv_nsec);
321 }
322
323 /*
324  * Generic udelay assumes that if preemption is allowed and the thread
325  * migrates to another CPU, that the ITC values are synchronized across
326  * all CPUs.
327  */
328 static void
329 ia64_itc_udelay (unsigned long usecs)
330 {
331         unsigned long start = ia64_get_itc();
332         unsigned long end = start + usecs*local_cpu_data->cyc_per_usec;
333
334         while (time_before(ia64_get_itc(), end))
335                 cpu_relax();
336 }
337
338 void (*ia64_udelay)(unsigned long usecs) = &ia64_itc_udelay;
339
340 void
341 udelay (unsigned long usecs)
342 {
343         (*ia64_udelay)(usecs);
344 }
345 EXPORT_SYMBOL(udelay);
346
347 /* IA64 doesn't cache the timezone */
348 void update_vsyscall_tz(void)
349 {
350 }
351
352 void update_vsyscall(struct timespec *wall, struct clocksource *c)
353 {
354         unsigned long flags;
355
356         write_seqlock_irqsave(&fsyscall_gtod_data.lock, flags);
357
358         /* copy fsyscall clock data */
359         fsyscall_gtod_data.clk_mask = c->mask;
360         fsyscall_gtod_data.clk_mult = c->mult;
361         fsyscall_gtod_data.clk_shift = c->shift;
362         fsyscall_gtod_data.clk_fsys_mmio = c->fsys_mmio;
363         fsyscall_gtod_data.clk_cycle_last = c->cycle_last;
364
365         /* copy kernel time structures */
366         fsyscall_gtod_data.wall_time.tv_sec = wall->tv_sec;
367         fsyscall_gtod_data.wall_time.tv_nsec = wall->tv_nsec;
368         fsyscall_gtod_data.monotonic_time.tv_sec = wall_to_monotonic.tv_sec
369                                                         + wall->tv_sec;
370         fsyscall_gtod_data.monotonic_time.tv_nsec = wall_to_monotonic.tv_nsec
371                                                         + wall->tv_nsec;
372
373         /* normalize */
374         while (fsyscall_gtod_data.monotonic_time.tv_nsec >= NSEC_PER_SEC) {
375                 fsyscall_gtod_data.monotonic_time.tv_nsec -= NSEC_PER_SEC;
376                 fsyscall_gtod_data.monotonic_time.tv_sec++;
377         }
378
379         write_sequnlock_irqrestore(&fsyscall_gtod_data.lock, flags);
380 }
381