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
paravirt_clocksource_resume(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 */
ia64_account_on_switch(struct task_struct * prev,struct task_struct * next)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 if (idle_task(smp_processor_id()) != prev)
97 account_system_time(prev, 0, delta_stime, delta_stime);
98 else
99 account_idle_time(delta_stime);
100
101 if (pi->ac_utime) {
102 delta_utime = cycle_to_cputime(pi->ac_utime);
103 account_user_time(prev, delta_utime, delta_utime);
104 }
105
106 pi->ac_stamp = ni->ac_stamp = now;
107 ni->ac_stime = ni->ac_utime = 0;
108 }
109
110 /*
111 * Account time for a transition between system, hard irq or soft irq state.
112 * Note that this function is called with interrupts enabled.
113 */
account_system_vtime(struct task_struct * tsk)114 void account_system_vtime(struct task_struct *tsk)
115 {
116 struct thread_info *ti = task_thread_info(tsk);
117 unsigned long flags;
118 cputime_t delta_stime;
119 __u64 now;
120
121 local_irq_save(flags);
122
123 now = ia64_get_itc();
124
125 delta_stime = cycle_to_cputime(ti->ac_stime + (now - ti->ac_stamp));
126 if (irq_count() || idle_task(smp_processor_id()) != tsk)
127 account_system_time(tsk, 0, delta_stime, delta_stime);
128 else
129 account_idle_time(delta_stime);
130 ti->ac_stime = 0;
131
132 ti->ac_stamp = now;
133
134 local_irq_restore(flags);
135 }
136 EXPORT_SYMBOL_GPL(account_system_vtime);
137
138 /*
139 * Called from the timer interrupt handler to charge accumulated user time
140 * to the current process. Must be called with interrupts disabled.
141 */
account_process_tick(struct task_struct * p,int user_tick)142 void account_process_tick(struct task_struct *p, int user_tick)
143 {
144 struct thread_info *ti = task_thread_info(p);
145 cputime_t delta_utime;
146
147 if (ti->ac_utime) {
148 delta_utime = cycle_to_cputime(ti->ac_utime);
149 account_user_time(p, delta_utime, delta_utime);
150 ti->ac_utime = 0;
151 }
152 }
153
154 #endif /* CONFIG_VIRT_CPU_ACCOUNTING */
155
156 static irqreturn_t
timer_interrupt(int irq,void * dev_id)157 timer_interrupt (int irq, void *dev_id)
158 {
159 unsigned long new_itm;
160
161 if (unlikely(cpu_is_offline(smp_processor_id()))) {
162 return IRQ_HANDLED;
163 }
164
165 platform_timer_interrupt(irq, dev_id);
166
167 new_itm = local_cpu_data->itm_next;
168
169 if (!time_after(ia64_get_itc(), new_itm))
170 printk(KERN_ERR "Oops: timer tick before it's due (itc=%lx,itm=%lx)\n",
171 ia64_get_itc(), new_itm);
172
173 profile_tick(CPU_PROFILING);
174
175 if (paravirt_do_steal_accounting(&new_itm))
176 goto skip_process_time_accounting;
177
178 while (1) {
179 update_process_times(user_mode(get_irq_regs()));
180
181 new_itm += local_cpu_data->itm_delta;
182
183 if (smp_processor_id() == time_keeper_id) {
184 /*
185 * Here we are in the timer irq handler. We have irqs locally
186 * disabled, but we don't know if the timer_bh is running on
187 * another CPU. We need to avoid to SMP race by acquiring the
188 * xtime_lock.
189 */
190 write_seqlock(&xtime_lock);
191 do_timer(1);
192 local_cpu_data->itm_next = new_itm;
193 write_sequnlock(&xtime_lock);
194 } else
195 local_cpu_data->itm_next = new_itm;
196
197 if (time_after(new_itm, ia64_get_itc()))
198 break;
199
200 /*
201 * Allow IPIs to interrupt the timer loop.
202 */
203 local_irq_enable();
204 local_irq_disable();
205 }
206
207 skip_process_time_accounting:
208
209 do {
210 /*
211 * If we're too close to the next clock tick for
212 * comfort, we increase the safety margin by
213 * intentionally dropping the next tick(s). We do NOT
214 * update itm.next because that would force us to call
215 * do_timer() which in turn would let our clock run
216 * too fast (with the potentially devastating effect
217 * of losing monotony of time).
218 */
219 while (!time_after(new_itm, ia64_get_itc() + local_cpu_data->itm_delta/2))
220 new_itm += local_cpu_data->itm_delta;
221 ia64_set_itm(new_itm);
222 /* double check, in case we got hit by a (slow) PMI: */
223 } while (time_after_eq(ia64_get_itc(), new_itm));
224 return IRQ_HANDLED;
225 }
226
227 /*
228 * Encapsulate access to the itm structure for SMP.
229 */
230 void
ia64_cpu_local_tick(void)231 ia64_cpu_local_tick (void)
232 {
233 int cpu = smp_processor_id();
234 unsigned long shift = 0, delta;
235
236 /* arrange for the cycle counter to generate a timer interrupt: */
237 ia64_set_itv(IA64_TIMER_VECTOR);
238
239 delta = local_cpu_data->itm_delta;
240 /*
241 * Stagger the timer tick for each CPU so they don't occur all at (almost) the
242 * same time:
243 */
244 if (cpu) {
245 unsigned long hi = 1UL << ia64_fls(cpu);
246 shift = (2*(cpu - hi) + 1) * delta/hi/2;
247 }
248 local_cpu_data->itm_next = ia64_get_itc() + delta + shift;
249 ia64_set_itm(local_cpu_data->itm_next);
250 }
251
252 static int nojitter;
253
nojitter_setup(char * str)254 static int __init nojitter_setup(char *str)
255 {
256 nojitter = 1;
257 printk("Jitter checking for ITC timers disabled\n");
258 return 1;
259 }
260
261 __setup("nojitter", nojitter_setup);
262
263
264 void __devinit
ia64_init_itm(void)265 ia64_init_itm (void)
266 {
267 unsigned long platform_base_freq, itc_freq;
268 struct pal_freq_ratio itc_ratio, proc_ratio;
269 long status, platform_base_drift, itc_drift;
270
271 /*
272 * According to SAL v2.6, we need to use a SAL call to determine the platform base
273 * frequency and then a PAL call to determine the frequency ratio between the ITC
274 * and the base frequency.
275 */
276 status = ia64_sal_freq_base(SAL_FREQ_BASE_PLATFORM,
277 &platform_base_freq, &platform_base_drift);
278 if (status != 0) {
279 printk(KERN_ERR "SAL_FREQ_BASE_PLATFORM failed: %s\n", ia64_sal_strerror(status));
280 } else {
281 status = ia64_pal_freq_ratios(&proc_ratio, NULL, &itc_ratio);
282 if (status != 0)
283 printk(KERN_ERR "PAL_FREQ_RATIOS failed with status=%ld\n", status);
284 }
285 if (status != 0) {
286 /* invent "random" values */
287 printk(KERN_ERR
288 "SAL/PAL failed to obtain frequency info---inventing reasonable values\n");
289 platform_base_freq = 100000000;
290 platform_base_drift = -1; /* no drift info */
291 itc_ratio.num = 3;
292 itc_ratio.den = 1;
293 }
294 if (platform_base_freq < 40000000) {
295 printk(KERN_ERR "Platform base frequency %lu bogus---resetting to 75MHz!\n",
296 platform_base_freq);
297 platform_base_freq = 75000000;
298 platform_base_drift = -1;
299 }
300 if (!proc_ratio.den)
301 proc_ratio.den = 1; /* avoid division by zero */
302 if (!itc_ratio.den)
303 itc_ratio.den = 1; /* avoid division by zero */
304
305 itc_freq = (platform_base_freq*itc_ratio.num)/itc_ratio.den;
306
307 local_cpu_data->itm_delta = (itc_freq + HZ/2) / HZ;
308 printk(KERN_DEBUG "CPU %d: base freq=%lu.%03luMHz, ITC ratio=%u/%u, "
309 "ITC freq=%lu.%03luMHz", smp_processor_id(),
310 platform_base_freq / 1000000, (platform_base_freq / 1000) % 1000,
311 itc_ratio.num, itc_ratio.den, itc_freq / 1000000, (itc_freq / 1000) % 1000);
312
313 if (platform_base_drift != -1) {
314 itc_drift = platform_base_drift*itc_ratio.num/itc_ratio.den;
315 printk("+/-%ldppm\n", itc_drift);
316 } else {
317 itc_drift = -1;
318 printk("\n");
319 }
320
321 local_cpu_data->proc_freq = (platform_base_freq*proc_ratio.num)/proc_ratio.den;
322 local_cpu_data->itc_freq = itc_freq;
323 local_cpu_data->cyc_per_usec = (itc_freq + USEC_PER_SEC/2) / USEC_PER_SEC;
324 local_cpu_data->nsec_per_cyc = ((NSEC_PER_SEC<<IA64_NSEC_PER_CYC_SHIFT)
325 + itc_freq/2)/itc_freq;
326
327 if (!(sal_platform_features & IA64_SAL_PLATFORM_FEATURE_ITC_DRIFT)) {
328 #ifdef CONFIG_SMP
329 /* On IA64 in an SMP configuration ITCs are never accurately synchronized.
330 * Jitter compensation requires a cmpxchg which may limit
331 * the scalability of the syscalls for retrieving time.
332 * The ITC synchronization is usually successful to within a few
333 * ITC ticks but this is not a sure thing. If you need to improve
334 * timer performance in SMP situations then boot the kernel with the
335 * "nojitter" option. However, doing so may result in time fluctuating (maybe
336 * even going backward) if the ITC offsets between the individual CPUs
337 * are too large.
338 */
339 if (!nojitter)
340 itc_jitter_data.itc_jitter = 1;
341 #endif
342 } else
343 /*
344 * ITC is drifty and we have not synchronized the ITCs in smpboot.c.
345 * ITC values may fluctuate significantly between processors.
346 * Clock should not be used for hrtimers. Mark itc as only
347 * useful for boot and testing.
348 *
349 * Note that jitter compensation is off! There is no point of
350 * synchronizing ITCs since they may be large differentials
351 * that change over time.
352 *
353 * The only way to fix this would be to repeatedly sync the
354 * ITCs. Until that time we have to avoid ITC.
355 */
356 clocksource_itc.rating = 50;
357
358 paravirt_init_missing_ticks_accounting(smp_processor_id());
359
360 /* avoid softlock up message when cpu is unplug and plugged again. */
361 touch_softlockup_watchdog();
362
363 /* Setup the CPU local timer tick */
364 ia64_cpu_local_tick();
365
366 if (!itc_clocksource) {
367 /* Sort out mult/shift values: */
368 clocksource_itc.mult =
369 clocksource_hz2mult(local_cpu_data->itc_freq,
370 clocksource_itc.shift);
371 clocksource_register(&clocksource_itc);
372 itc_clocksource = &clocksource_itc;
373 }
374 }
375
itc_get_cycles(void)376 static cycle_t itc_get_cycles(void)
377 {
378 u64 lcycle, now, ret;
379
380 if (!itc_jitter_data.itc_jitter)
381 return get_cycles();
382
383 lcycle = itc_jitter_data.itc_lastcycle;
384 now = get_cycles();
385 if (lcycle && time_after(lcycle, now))
386 return lcycle;
387
388 /*
389 * Keep track of the last timer value returned.
390 * In an SMP environment, you could lose out in contention of
391 * cmpxchg. If so, your cmpxchg returns new value which the
392 * winner of contention updated to. Use the new value instead.
393 */
394 ret = cmpxchg(&itc_jitter_data.itc_lastcycle, lcycle, now);
395 if (unlikely(ret != lcycle))
396 return ret;
397
398 return now;
399 }
400
401
402 static struct irqaction timer_irqaction = {
403 .handler = timer_interrupt,
404 .flags = IRQF_DISABLED | IRQF_IRQPOLL,
405 .name = "timer"
406 };
407
408 void __init
time_init(void)409 time_init (void)
410 {
411 register_percpu_irq(IA64_TIMER_VECTOR, &timer_irqaction);
412 efi_gettimeofday(&xtime);
413 ia64_init_itm();
414
415 /*
416 * Initialize wall_to_monotonic such that adding it to xtime will yield zero, the
417 * tv_nsec field must be normalized (i.e., 0 <= nsec < NSEC_PER_SEC).
418 */
419 set_normalized_timespec(&wall_to_monotonic, -xtime.tv_sec, -xtime.tv_nsec);
420 }
421
422 /*
423 * Generic udelay assumes that if preemption is allowed and the thread
424 * migrates to another CPU, that the ITC values are synchronized across
425 * all CPUs.
426 */
427 static void
ia64_itc_udelay(unsigned long usecs)428 ia64_itc_udelay (unsigned long usecs)
429 {
430 unsigned long start = ia64_get_itc();
431 unsigned long end = start + usecs*local_cpu_data->cyc_per_usec;
432
433 while (time_before(ia64_get_itc(), end))
434 cpu_relax();
435 }
436
437 void (*ia64_udelay)(unsigned long usecs) = &ia64_itc_udelay;
438
439 void
udelay(unsigned long usecs)440 udelay (unsigned long usecs)
441 {
442 (*ia64_udelay)(usecs);
443 }
444 EXPORT_SYMBOL(udelay);
445
446 /* IA64 doesn't cache the timezone */
update_vsyscall_tz(void)447 void update_vsyscall_tz(void)
448 {
449 }
450
update_vsyscall(struct timespec * wall,struct clocksource * c)451 void update_vsyscall(struct timespec *wall, struct clocksource *c)
452 {
453 unsigned long flags;
454
455 write_seqlock_irqsave(&fsyscall_gtod_data.lock, flags);
456
457 /* copy fsyscall clock data */
458 fsyscall_gtod_data.clk_mask = c->mask;
459 fsyscall_gtod_data.clk_mult = c->mult;
460 fsyscall_gtod_data.clk_shift = c->shift;
461 fsyscall_gtod_data.clk_fsys_mmio = c->fsys_mmio;
462 fsyscall_gtod_data.clk_cycle_last = c->cycle_last;
463
464 /* copy kernel time structures */
465 fsyscall_gtod_data.wall_time.tv_sec = wall->tv_sec;
466 fsyscall_gtod_data.wall_time.tv_nsec = wall->tv_nsec;
467 fsyscall_gtod_data.monotonic_time.tv_sec = wall_to_monotonic.tv_sec
468 + wall->tv_sec;
469 fsyscall_gtod_data.monotonic_time.tv_nsec = wall_to_monotonic.tv_nsec
470 + wall->tv_nsec;
471
472 /* normalize */
473 while (fsyscall_gtod_data.monotonic_time.tv_nsec >= NSEC_PER_SEC) {
474 fsyscall_gtod_data.monotonic_time.tv_nsec -= NSEC_PER_SEC;
475 fsyscall_gtod_data.monotonic_time.tv_sec++;
476 }
477
478 write_sequnlock_irqrestore(&fsyscall_gtod_data.lock, flags);
479 }
480
481