2 * linux/arch/ia64/kernel/time.c
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>
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>
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>
30 #include <asm/sections.h>
31 #include <asm/system.h>
33 #include "fsyscall_gtod_data.h"
35 static cycle_t itc_get_cycles(void);
37 struct fsyscall_gtod_data_t fsyscall_gtod_data = {
38 .lock = SEQLOCK_UNLOCKED,
41 struct itc_jitter_data_t itc_jitter_data;
43 volatile int time_keeper_id = 0; /* smp_processor_id() of time-keeper */
45 #ifdef CONFIG_IA64_DEBUG_IRQ
47 unsigned long last_cli_ip;
48 EXPORT_SYMBOL(last_cli_ip);
52 #ifdef CONFIG_PARAVIRT
53 /* We need to define a real function for sched_clock, to override the
54 weak default version */
55 unsigned long long sched_clock(void)
57 return paravirt_sched_clock();
61 #ifdef CONFIG_PARAVIRT
63 paravirt_clocksource_resume(void)
65 if (pv_time_ops.clocksource_resume)
66 pv_time_ops.clocksource_resume();
70 static struct clocksource clocksource_itc = {
73 .read = itc_get_cycles,
74 .mask = CLOCKSOURCE_MASK(64),
75 .mult = 0, /*to be calculated*/
77 .flags = CLOCK_SOURCE_IS_CONTINUOUS,
78 #ifdef CONFIG_PARAVIRT
79 .resume = paravirt_clocksource_resume,
82 static struct clocksource *itc_clocksource;
84 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
86 #include <linux/kernel_stat.h>
88 extern cputime_t cycle_to_cputime(u64 cyc);
91 * Called from the context switch with interrupts disabled, to charge all
92 * accumulated times to the current process, and to prepare accounting on
95 void ia64_account_on_switch(struct task_struct *prev, struct task_struct *next)
97 struct thread_info *pi = task_thread_info(prev);
98 struct thread_info *ni = task_thread_info(next);
99 cputime_t delta_stime, delta_utime;
102 now = ia64_get_itc();
104 delta_stime = cycle_to_cputime(pi->ac_stime + (now - pi->ac_stamp));
105 if (idle_task(smp_processor_id()) != prev)
106 account_system_time(prev, 0, delta_stime, delta_stime);
108 account_idle_time(delta_stime);
111 delta_utime = cycle_to_cputime(pi->ac_utime);
112 account_user_time(prev, delta_utime, delta_utime);
115 pi->ac_stamp = ni->ac_stamp = now;
116 ni->ac_stime = ni->ac_utime = 0;
120 * Account time for a transition between system, hard irq or soft irq state.
121 * Note that this function is called with interrupts enabled.
123 void account_system_vtime(struct task_struct *tsk)
125 struct thread_info *ti = task_thread_info(tsk);
127 cputime_t delta_stime;
130 local_irq_save(flags);
132 now = ia64_get_itc();
134 delta_stime = cycle_to_cputime(ti->ac_stime + (now - ti->ac_stamp));
135 if (irq_count() || idle_task(smp_processor_id()) != tsk)
136 account_system_time(tsk, 0, delta_stime, delta_stime);
138 account_idle_time(delta_stime);
143 local_irq_restore(flags);
145 EXPORT_SYMBOL_GPL(account_system_vtime);
148 * Called from the timer interrupt handler to charge accumulated user time
149 * to the current process. Must be called with interrupts disabled.
151 void account_process_tick(struct task_struct *p, int user_tick)
153 struct thread_info *ti = task_thread_info(p);
154 cputime_t delta_utime;
157 delta_utime = cycle_to_cputime(ti->ac_utime);
158 account_user_time(p, delta_utime, delta_utime);
163 #endif /* CONFIG_VIRT_CPU_ACCOUNTING */
166 timer_interrupt (int irq, void *dev_id)
168 unsigned long new_itm;
170 if (unlikely(cpu_is_offline(smp_processor_id()))) {
174 platform_timer_interrupt(irq, dev_id);
176 new_itm = local_cpu_data->itm_next;
178 if (!time_after(ia64_get_itc(), new_itm))
179 printk(KERN_ERR "Oops: timer tick before it's due (itc=%lx,itm=%lx)\n",
180 ia64_get_itc(), new_itm);
182 profile_tick(CPU_PROFILING);
184 if (paravirt_do_steal_accounting(&new_itm))
185 goto skip_process_time_accounting;
188 update_process_times(user_mode(get_irq_regs()));
190 new_itm += local_cpu_data->itm_delta;
192 if (smp_processor_id() == time_keeper_id) {
194 * Here we are in the timer irq handler. We have irqs locally
195 * disabled, but we don't know if the timer_bh is running on
196 * another CPU. We need to avoid to SMP race by acquiring the
199 write_seqlock(&xtime_lock);
201 local_cpu_data->itm_next = new_itm;
202 write_sequnlock(&xtime_lock);
204 local_cpu_data->itm_next = new_itm;
206 if (time_after(new_itm, ia64_get_itc()))
210 * Allow IPIs to interrupt the timer loop.
216 skip_process_time_accounting:
220 * If we're too close to the next clock tick for
221 * comfort, we increase the safety margin by
222 * intentionally dropping the next tick(s). We do NOT
223 * update itm.next because that would force us to call
224 * do_timer() which in turn would let our clock run
225 * too fast (with the potentially devastating effect
226 * of losing monotony of time).
228 while (!time_after(new_itm, ia64_get_itc() + local_cpu_data->itm_delta/2))
229 new_itm += local_cpu_data->itm_delta;
230 ia64_set_itm(new_itm);
231 /* double check, in case we got hit by a (slow) PMI: */
232 } while (time_after_eq(ia64_get_itc(), new_itm));
237 * Encapsulate access to the itm structure for SMP.
240 ia64_cpu_local_tick (void)
242 int cpu = smp_processor_id();
243 unsigned long shift = 0, delta;
245 /* arrange for the cycle counter to generate a timer interrupt: */
246 ia64_set_itv(IA64_TIMER_VECTOR);
248 delta = local_cpu_data->itm_delta;
250 * Stagger the timer tick for each CPU so they don't occur all at (almost) the
254 unsigned long hi = 1UL << ia64_fls(cpu);
255 shift = (2*(cpu - hi) + 1) * delta/hi/2;
257 local_cpu_data->itm_next = ia64_get_itc() + delta + shift;
258 ia64_set_itm(local_cpu_data->itm_next);
263 static int __init nojitter_setup(char *str)
266 printk("Jitter checking for ITC timers disabled\n");
270 __setup("nojitter", nojitter_setup);
276 unsigned long platform_base_freq, itc_freq;
277 struct pal_freq_ratio itc_ratio, proc_ratio;
278 long status, platform_base_drift, itc_drift;
281 * According to SAL v2.6, we need to use a SAL call to determine the platform base
282 * frequency and then a PAL call to determine the frequency ratio between the ITC
283 * and the base frequency.
285 status = ia64_sal_freq_base(SAL_FREQ_BASE_PLATFORM,
286 &platform_base_freq, &platform_base_drift);
288 printk(KERN_ERR "SAL_FREQ_BASE_PLATFORM failed: %s\n", ia64_sal_strerror(status));
290 status = ia64_pal_freq_ratios(&proc_ratio, NULL, &itc_ratio);
292 printk(KERN_ERR "PAL_FREQ_RATIOS failed with status=%ld\n", status);
295 /* invent "random" values */
297 "SAL/PAL failed to obtain frequency info---inventing reasonable values\n");
298 platform_base_freq = 100000000;
299 platform_base_drift = -1; /* no drift info */
303 if (platform_base_freq < 40000000) {
304 printk(KERN_ERR "Platform base frequency %lu bogus---resetting to 75MHz!\n",
306 platform_base_freq = 75000000;
307 platform_base_drift = -1;
310 proc_ratio.den = 1; /* avoid division by zero */
312 itc_ratio.den = 1; /* avoid division by zero */
314 itc_freq = (platform_base_freq*itc_ratio.num)/itc_ratio.den;
316 local_cpu_data->itm_delta = (itc_freq + HZ/2) / HZ;
317 printk(KERN_DEBUG "CPU %d: base freq=%lu.%03luMHz, ITC ratio=%u/%u, "
318 "ITC freq=%lu.%03luMHz", smp_processor_id(),
319 platform_base_freq / 1000000, (platform_base_freq / 1000) % 1000,
320 itc_ratio.num, itc_ratio.den, itc_freq / 1000000, (itc_freq / 1000) % 1000);
322 if (platform_base_drift != -1) {
323 itc_drift = platform_base_drift*itc_ratio.num/itc_ratio.den;
324 printk("+/-%ldppm\n", itc_drift);
330 local_cpu_data->proc_freq = (platform_base_freq*proc_ratio.num)/proc_ratio.den;
331 local_cpu_data->itc_freq = itc_freq;
332 local_cpu_data->cyc_per_usec = (itc_freq + USEC_PER_SEC/2) / USEC_PER_SEC;
333 local_cpu_data->nsec_per_cyc = ((NSEC_PER_SEC<<IA64_NSEC_PER_CYC_SHIFT)
334 + itc_freq/2)/itc_freq;
336 if (!(sal_platform_features & IA64_SAL_PLATFORM_FEATURE_ITC_DRIFT)) {
338 /* On IA64 in an SMP configuration ITCs are never accurately synchronized.
339 * Jitter compensation requires a cmpxchg which may limit
340 * the scalability of the syscalls for retrieving time.
341 * The ITC synchronization is usually successful to within a few
342 * ITC ticks but this is not a sure thing. If you need to improve
343 * timer performance in SMP situations then boot the kernel with the
344 * "nojitter" option. However, doing so may result in time fluctuating (maybe
345 * even going backward) if the ITC offsets between the individual CPUs
349 itc_jitter_data.itc_jitter = 1;
353 * ITC is drifty and we have not synchronized the ITCs in smpboot.c.
354 * ITC values may fluctuate significantly between processors.
355 * Clock should not be used for hrtimers. Mark itc as only
356 * useful for boot and testing.
358 * Note that jitter compensation is off! There is no point of
359 * synchronizing ITCs since they may be large differentials
360 * that change over time.
362 * The only way to fix this would be to repeatedly sync the
363 * ITCs. Until that time we have to avoid ITC.
365 clocksource_itc.rating = 50;
367 paravirt_init_missing_ticks_accounting(smp_processor_id());
369 /* avoid softlock up message when cpu is unplug and plugged again. */
370 touch_softlockup_watchdog();
372 /* Setup the CPU local timer tick */
373 ia64_cpu_local_tick();
375 if (!itc_clocksource) {
376 /* Sort out mult/shift values: */
377 clocksource_itc.mult =
378 clocksource_hz2mult(local_cpu_data->itc_freq,
379 clocksource_itc.shift);
380 clocksource_register(&clocksource_itc);
381 itc_clocksource = &clocksource_itc;
385 static cycle_t itc_get_cycles(void)
387 u64 lcycle, now, ret;
389 if (!itc_jitter_data.itc_jitter)
392 lcycle = itc_jitter_data.itc_lastcycle;
394 if (lcycle && time_after(lcycle, now))
398 * Keep track of the last timer value returned.
399 * In an SMP environment, you could lose out in contention of
400 * cmpxchg. If so, your cmpxchg returns new value which the
401 * winner of contention updated to. Use the new value instead.
403 ret = cmpxchg(&itc_jitter_data.itc_lastcycle, lcycle, now);
404 if (unlikely(ret != lcycle))
411 static struct irqaction timer_irqaction = {
412 .handler = timer_interrupt,
413 .flags = IRQF_DISABLED | IRQF_IRQPOLL,
420 register_percpu_irq(IA64_TIMER_VECTOR, &timer_irqaction);
421 efi_gettimeofday(&xtime);
425 * Initialize wall_to_monotonic such that adding it to xtime will yield zero, the
426 * tv_nsec field must be normalized (i.e., 0 <= nsec < NSEC_PER_SEC).
428 set_normalized_timespec(&wall_to_monotonic, -xtime.tv_sec, -xtime.tv_nsec);
432 * Generic udelay assumes that if preemption is allowed and the thread
433 * migrates to another CPU, that the ITC values are synchronized across
437 ia64_itc_udelay (unsigned long usecs)
439 unsigned long start = ia64_get_itc();
440 unsigned long end = start + usecs*local_cpu_data->cyc_per_usec;
442 while (time_before(ia64_get_itc(), end))
446 void (*ia64_udelay)(unsigned long usecs) = &ia64_itc_udelay;
449 udelay (unsigned long usecs)
451 (*ia64_udelay)(usecs);
453 EXPORT_SYMBOL(udelay);
455 /* IA64 doesn't cache the timezone */
456 void update_vsyscall_tz(void)
460 void update_vsyscall(struct timespec *wall, struct clocksource *c)
464 write_seqlock_irqsave(&fsyscall_gtod_data.lock, flags);
466 /* copy fsyscall clock data */
467 fsyscall_gtod_data.clk_mask = c->mask;
468 fsyscall_gtod_data.clk_mult = c->mult;
469 fsyscall_gtod_data.clk_shift = c->shift;
470 fsyscall_gtod_data.clk_fsys_mmio = c->fsys_mmio;
471 fsyscall_gtod_data.clk_cycle_last = c->cycle_last;
473 /* copy kernel time structures */
474 fsyscall_gtod_data.wall_time.tv_sec = wall->tv_sec;
475 fsyscall_gtod_data.wall_time.tv_nsec = wall->tv_nsec;
476 fsyscall_gtod_data.monotonic_time.tv_sec = wall_to_monotonic.tv_sec
478 fsyscall_gtod_data.monotonic_time.tv_nsec = wall_to_monotonic.tv_nsec
482 while (fsyscall_gtod_data.monotonic_time.tv_nsec >= NSEC_PER_SEC) {
483 fsyscall_gtod_data.monotonic_time.tv_nsec -= NSEC_PER_SEC;
484 fsyscall_gtod_data.monotonic_time.tv_sec++;
487 write_sequnlock_irqrestore(&fsyscall_gtod_data.lock, flags);