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>
23 #include <linux/platform_device.h>
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>
31 #include <asm/sections.h>
32 #include <asm/system.h>
34 #include "fsyscall_gtod_data.h"
36 static cycle_t
itc_get_cycles(struct clocksource
*cs
);
38 struct fsyscall_gtod_data_t fsyscall_gtod_data
= {
39 .lock
= SEQLOCK_UNLOCKED
,
42 struct itc_jitter_data_t itc_jitter_data
;
44 volatile int time_keeper_id
= 0; /* smp_processor_id() of time-keeper */
46 #ifdef CONFIG_IA64_DEBUG_IRQ
48 unsigned long last_cli_ip
;
49 EXPORT_SYMBOL(last_cli_ip
);
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)
58 return paravirt_sched_clock();
62 #ifdef CONFIG_PARAVIRT
64 paravirt_clocksource_resume(void)
66 if (pv_time_ops
.clocksource_resume
)
67 pv_time_ops
.clocksource_resume();
71 static struct clocksource clocksource_itc
= {
74 .read
= itc_get_cycles
,
75 .mask
= CLOCKSOURCE_MASK(64),
76 .mult
= 0, /*to be calculated*/
78 .flags
= CLOCK_SOURCE_IS_CONTINUOUS
,
79 #ifdef CONFIG_PARAVIRT
80 .resume
= paravirt_clocksource_resume
,
83 static struct clocksource
*itc_clocksource
;
85 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
87 #include <linux/kernel_stat.h>
89 extern cputime_t
cycle_to_cputime(u64 cyc
);
92 * Called from the context switch with interrupts disabled, to charge all
93 * accumulated times to the current process, and to prepare accounting on
96 void ia64_account_on_switch(struct task_struct
*prev
, struct task_struct
*next
)
98 struct thread_info
*pi
= task_thread_info(prev
);
99 struct thread_info
*ni
= task_thread_info(next
);
100 cputime_t delta_stime
, delta_utime
;
103 now
= ia64_get_itc();
105 delta_stime
= cycle_to_cputime(pi
->ac_stime
+ (now
- pi
->ac_stamp
));
106 if (idle_task(smp_processor_id()) != prev
)
107 account_system_time(prev
, 0, delta_stime
, delta_stime
);
109 account_idle_time(delta_stime
);
112 delta_utime
= cycle_to_cputime(pi
->ac_utime
);
113 account_user_time(prev
, delta_utime
, delta_utime
);
116 pi
->ac_stamp
= ni
->ac_stamp
= now
;
117 ni
->ac_stime
= ni
->ac_utime
= 0;
121 * Account time for a transition between system, hard irq or soft irq state.
122 * Note that this function is called with interrupts enabled.
124 void account_system_vtime(struct task_struct
*tsk
)
126 struct thread_info
*ti
= task_thread_info(tsk
);
128 cputime_t delta_stime
;
131 local_irq_save(flags
);
133 now
= ia64_get_itc();
135 delta_stime
= cycle_to_cputime(ti
->ac_stime
+ (now
- ti
->ac_stamp
));
136 if (irq_count() || idle_task(smp_processor_id()) != tsk
)
137 account_system_time(tsk
, 0, delta_stime
, delta_stime
);
139 account_idle_time(delta_stime
);
144 local_irq_restore(flags
);
146 EXPORT_SYMBOL_GPL(account_system_vtime
);
149 * Called from the timer interrupt handler to charge accumulated user time
150 * to the current process. Must be called with interrupts disabled.
152 void account_process_tick(struct task_struct
*p
, int user_tick
)
154 struct thread_info
*ti
= task_thread_info(p
);
155 cputime_t delta_utime
;
158 delta_utime
= cycle_to_cputime(ti
->ac_utime
);
159 account_user_time(p
, delta_utime
, delta_utime
);
164 #endif /* CONFIG_VIRT_CPU_ACCOUNTING */
167 timer_interrupt (int irq
, void *dev_id
)
169 unsigned long new_itm
;
171 if (unlikely(cpu_is_offline(smp_processor_id()))) {
175 platform_timer_interrupt(irq
, dev_id
);
177 new_itm
= local_cpu_data
->itm_next
;
179 if (!time_after(ia64_get_itc(), new_itm
))
180 printk(KERN_ERR
"Oops: timer tick before it's due (itc=%lx,itm=%lx)\n",
181 ia64_get_itc(), new_itm
);
183 profile_tick(CPU_PROFILING
);
185 if (paravirt_do_steal_accounting(&new_itm
))
186 goto skip_process_time_accounting
;
189 update_process_times(user_mode(get_irq_regs()));
191 new_itm
+= local_cpu_data
->itm_delta
;
193 if (smp_processor_id() == time_keeper_id
) {
195 * Here we are in the timer irq handler. We have irqs locally
196 * disabled, but we don't know if the timer_bh is running on
197 * another CPU. We need to avoid to SMP race by acquiring the
200 write_seqlock(&xtime_lock
);
202 local_cpu_data
->itm_next
= new_itm
;
203 write_sequnlock(&xtime_lock
);
205 local_cpu_data
->itm_next
= new_itm
;
207 if (time_after(new_itm
, ia64_get_itc()))
211 * Allow IPIs to interrupt the timer loop.
217 skip_process_time_accounting
:
221 * If we're too close to the next clock tick for
222 * comfort, we increase the safety margin by
223 * intentionally dropping the next tick(s). We do NOT
224 * update itm.next because that would force us to call
225 * do_timer() which in turn would let our clock run
226 * too fast (with the potentially devastating effect
227 * of losing monotony of time).
229 while (!time_after(new_itm
, ia64_get_itc() + local_cpu_data
->itm_delta
/2))
230 new_itm
+= local_cpu_data
->itm_delta
;
231 ia64_set_itm(new_itm
);
232 /* double check, in case we got hit by a (slow) PMI: */
233 } while (time_after_eq(ia64_get_itc(), new_itm
));
238 * Encapsulate access to the itm structure for SMP.
241 ia64_cpu_local_tick (void)
243 int cpu
= smp_processor_id();
244 unsigned long shift
= 0, delta
;
246 /* arrange for the cycle counter to generate a timer interrupt: */
247 ia64_set_itv(IA64_TIMER_VECTOR
);
249 delta
= local_cpu_data
->itm_delta
;
251 * Stagger the timer tick for each CPU so they don't occur all at (almost) the
255 unsigned long hi
= 1UL << ia64_fls(cpu
);
256 shift
= (2*(cpu
- hi
) + 1) * delta
/hi
/2;
258 local_cpu_data
->itm_next
= ia64_get_itc() + delta
+ shift
;
259 ia64_set_itm(local_cpu_data
->itm_next
);
264 static int __init
nojitter_setup(char *str
)
267 printk("Jitter checking for ITC timers disabled\n");
271 __setup("nojitter", nojitter_setup
);
277 unsigned long platform_base_freq
, itc_freq
;
278 struct pal_freq_ratio itc_ratio
, proc_ratio
;
279 long status
, platform_base_drift
, itc_drift
;
282 * According to SAL v2.6, we need to use a SAL call to determine the platform base
283 * frequency and then a PAL call to determine the frequency ratio between the ITC
284 * and the base frequency.
286 status
= ia64_sal_freq_base(SAL_FREQ_BASE_PLATFORM
,
287 &platform_base_freq
, &platform_base_drift
);
289 printk(KERN_ERR
"SAL_FREQ_BASE_PLATFORM failed: %s\n", ia64_sal_strerror(status
));
291 status
= ia64_pal_freq_ratios(&proc_ratio
, NULL
, &itc_ratio
);
293 printk(KERN_ERR
"PAL_FREQ_RATIOS failed with status=%ld\n", status
);
296 /* invent "random" values */
298 "SAL/PAL failed to obtain frequency info---inventing reasonable values\n");
299 platform_base_freq
= 100000000;
300 platform_base_drift
= -1; /* no drift info */
304 if (platform_base_freq
< 40000000) {
305 printk(KERN_ERR
"Platform base frequency %lu bogus---resetting to 75MHz!\n",
307 platform_base_freq
= 75000000;
308 platform_base_drift
= -1;
311 proc_ratio
.den
= 1; /* avoid division by zero */
313 itc_ratio
.den
= 1; /* avoid division by zero */
315 itc_freq
= (platform_base_freq
*itc_ratio
.num
)/itc_ratio
.den
;
317 local_cpu_data
->itm_delta
= (itc_freq
+ HZ
/2) / HZ
;
318 printk(KERN_DEBUG
"CPU %d: base freq=%lu.%03luMHz, ITC ratio=%u/%u, "
319 "ITC freq=%lu.%03luMHz", smp_processor_id(),
320 platform_base_freq
/ 1000000, (platform_base_freq
/ 1000) % 1000,
321 itc_ratio
.num
, itc_ratio
.den
, itc_freq
/ 1000000, (itc_freq
/ 1000) % 1000);
323 if (platform_base_drift
!= -1) {
324 itc_drift
= platform_base_drift
*itc_ratio
.num
/itc_ratio
.den
;
325 printk("+/-%ldppm\n", itc_drift
);
331 local_cpu_data
->proc_freq
= (platform_base_freq
*proc_ratio
.num
)/proc_ratio
.den
;
332 local_cpu_data
->itc_freq
= itc_freq
;
333 local_cpu_data
->cyc_per_usec
= (itc_freq
+ USEC_PER_SEC
/2) / USEC_PER_SEC
;
334 local_cpu_data
->nsec_per_cyc
= ((NSEC_PER_SEC
<<IA64_NSEC_PER_CYC_SHIFT
)
335 + itc_freq
/2)/itc_freq
;
337 if (!(sal_platform_features
& IA64_SAL_PLATFORM_FEATURE_ITC_DRIFT
)) {
339 /* On IA64 in an SMP configuration ITCs are never accurately synchronized.
340 * Jitter compensation requires a cmpxchg which may limit
341 * the scalability of the syscalls for retrieving time.
342 * The ITC synchronization is usually successful to within a few
343 * ITC ticks but this is not a sure thing. If you need to improve
344 * timer performance in SMP situations then boot the kernel with the
345 * "nojitter" option. However, doing so may result in time fluctuating (maybe
346 * even going backward) if the ITC offsets between the individual CPUs
350 itc_jitter_data
.itc_jitter
= 1;
354 * ITC is drifty and we have not synchronized the ITCs in smpboot.c.
355 * ITC values may fluctuate significantly between processors.
356 * Clock should not be used for hrtimers. Mark itc as only
357 * useful for boot and testing.
359 * Note that jitter compensation is off! There is no point of
360 * synchronizing ITCs since they may be large differentials
361 * that change over time.
363 * The only way to fix this would be to repeatedly sync the
364 * ITCs. Until that time we have to avoid ITC.
366 clocksource_itc
.rating
= 50;
368 paravirt_init_missing_ticks_accounting(smp_processor_id());
370 /* avoid softlock up message when cpu is unplug and plugged again. */
371 touch_softlockup_watchdog();
373 /* Setup the CPU local timer tick */
374 ia64_cpu_local_tick();
376 if (!itc_clocksource
) {
377 /* Sort out mult/shift values: */
378 clocksource_itc
.mult
=
379 clocksource_hz2mult(local_cpu_data
->itc_freq
,
380 clocksource_itc
.shift
);
381 clocksource_register(&clocksource_itc
);
382 itc_clocksource
= &clocksource_itc
;
386 static cycle_t
itc_get_cycles(struct clocksource
*cs
)
388 unsigned long lcycle
, now
, ret
;
390 if (!itc_jitter_data
.itc_jitter
)
393 lcycle
= itc_jitter_data
.itc_lastcycle
;
395 if (lcycle
&& time_after(lcycle
, now
))
399 * Keep track of the last timer value returned.
400 * In an SMP environment, you could lose out in contention of
401 * cmpxchg. If so, your cmpxchg returns new value which the
402 * winner of contention updated to. Use the new value instead.
404 ret
= cmpxchg(&itc_jitter_data
.itc_lastcycle
, lcycle
, now
);
405 if (unlikely(ret
!= lcycle
))
412 static struct irqaction timer_irqaction
= {
413 .handler
= timer_interrupt
,
414 .flags
= IRQF_DISABLED
| IRQF_IRQPOLL
,
418 static struct platform_device rtc_efi_dev
= {
423 static int __init
rtc_init(void)
425 if (platform_device_register(&rtc_efi_dev
) < 0)
426 printk(KERN_ERR
"unable to register rtc device...\n");
428 /* not necessarily an error */
431 module_init(rtc_init
);
436 register_percpu_irq(IA64_TIMER_VECTOR
, &timer_irqaction
);
437 efi_gettimeofday(&xtime
);
441 * Initialize wall_to_monotonic such that adding it to xtime will yield zero, the
442 * tv_nsec field must be normalized (i.e., 0 <= nsec < NSEC_PER_SEC).
444 set_normalized_timespec(&wall_to_monotonic
, -xtime
.tv_sec
, -xtime
.tv_nsec
);
448 * Generic udelay assumes that if preemption is allowed and the thread
449 * migrates to another CPU, that the ITC values are synchronized across
453 ia64_itc_udelay (unsigned long usecs
)
455 unsigned long start
= ia64_get_itc();
456 unsigned long end
= start
+ usecs
*local_cpu_data
->cyc_per_usec
;
458 while (time_before(ia64_get_itc(), end
))
462 void (*ia64_udelay
)(unsigned long usecs
) = &ia64_itc_udelay
;
465 udelay (unsigned long usecs
)
467 (*ia64_udelay
)(usecs
);
469 EXPORT_SYMBOL(udelay
);
471 /* IA64 doesn't cache the timezone */
472 void update_vsyscall_tz(void)
476 void update_vsyscall(struct timespec
*wall
, struct clocksource
*c
, u32 mult
)
480 write_seqlock_irqsave(&fsyscall_gtod_data
.lock
, flags
);
482 /* copy fsyscall clock data */
483 fsyscall_gtod_data
.clk_mask
= c
->mask
;
484 fsyscall_gtod_data
.clk_mult
= mult
;
485 fsyscall_gtod_data
.clk_shift
= c
->shift
;
486 fsyscall_gtod_data
.clk_fsys_mmio
= c
->fsys_mmio
;
487 fsyscall_gtod_data
.clk_cycle_last
= c
->cycle_last
;
489 /* copy kernel time structures */
490 fsyscall_gtod_data
.wall_time
.tv_sec
= wall
->tv_sec
;
491 fsyscall_gtod_data
.wall_time
.tv_nsec
= wall
->tv_nsec
;
492 fsyscall_gtod_data
.monotonic_time
.tv_sec
= wall_to_monotonic
.tv_sec
494 fsyscall_gtod_data
.monotonic_time
.tv_nsec
= wall_to_monotonic
.tv_nsec
498 while (fsyscall_gtod_data
.monotonic_time
.tv_nsec
>= NSEC_PER_SEC
) {
499 fsyscall_gtod_data
.monotonic_time
.tv_nsec
-= NSEC_PER_SEC
;
500 fsyscall_gtod_data
.monotonic_time
.tv_sec
++;
503 write_sequnlock_irqrestore(&fsyscall_gtod_data
.lock
, flags
);