vtime: Warn if irqs aren't disabled on system time accounting APIs
[linux-3.10.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/timekeeper_internal.h>
23 #include <linux/platform_device.h>
24
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>
30 #include <asm/sal.h>
31 #include <asm/sections.h>
32
33 #include "fsyscall_gtod_data.h"
34
35 static cycle_t itc_get_cycles(struct clocksource *cs);
36
37 struct fsyscall_gtod_data_t fsyscall_gtod_data;
38
39 struct itc_jitter_data_t itc_jitter_data;
40
41 volatile int time_keeper_id = 0; /* smp_processor_id() of time-keeper */
42
43 #ifdef CONFIG_IA64_DEBUG_IRQ
44
45 unsigned long last_cli_ip;
46 EXPORT_SYMBOL(last_cli_ip);
47
48 #endif
49
50 #ifdef CONFIG_PARAVIRT
51 /* We need to define a real function for sched_clock, to override the
52    weak default version */
53 unsigned long long sched_clock(void)
54 {
55         return paravirt_sched_clock();
56 }
57 #endif
58
59 #ifdef CONFIG_PARAVIRT
60 static void
61 paravirt_clocksource_resume(struct clocksource *cs)
62 {
63         if (pv_time_ops.clocksource_resume)
64                 pv_time_ops.clocksource_resume();
65 }
66 #endif
67
68 static struct clocksource clocksource_itc = {
69         .name           = "itc",
70         .rating         = 350,
71         .read           = itc_get_cycles,
72         .mask           = CLOCKSOURCE_MASK(64),
73         .flags          = CLOCK_SOURCE_IS_CONTINUOUS,
74 #ifdef CONFIG_PARAVIRT
75         .resume         = paravirt_clocksource_resume,
76 #endif
77 };
78 static struct clocksource *itc_clocksource;
79
80 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
81
82 #include <linux/kernel_stat.h>
83
84 extern cputime_t cycle_to_cputime(u64 cyc);
85
86 void vtime_account_user(struct task_struct *tsk)
87 {
88         cputime_t delta_utime;
89         struct thread_info *ti = task_thread_info(tsk);
90
91         if (ti->ac_utime) {
92                 delta_utime = cycle_to_cputime(ti->ac_utime);
93                 account_user_time(tsk, delta_utime, delta_utime);
94                 ti->ac_utime = 0;
95         }
96 }
97
98 /*
99  * Called from the context switch with interrupts disabled, to charge all
100  * accumulated times to the current process, and to prepare accounting on
101  * the next process.
102  */
103 void arch_vtime_task_switch(struct task_struct *prev)
104 {
105         struct thread_info *pi = task_thread_info(prev);
106         struct thread_info *ni = task_thread_info(current);
107
108         pi->ac_stamp = ni->ac_stamp;
109         ni->ac_stime = ni->ac_utime = 0;
110 }
111
112 /*
113  * Account time for a transition between system, hard irq or soft irq state.
114  * Note that this function is called with interrupts enabled.
115  */
116 static cputime_t vtime_delta(struct task_struct *tsk)
117 {
118         struct thread_info *ti = task_thread_info(tsk);
119         cputime_t delta_stime;
120         __u64 now;
121
122         WARN_ON_ONCE(!irqs_disabled());
123
124         now = ia64_get_itc();
125
126         delta_stime = cycle_to_cputime(ti->ac_stime + (now - ti->ac_stamp));
127         ti->ac_stime = 0;
128         ti->ac_stamp = now;
129
130         return delta_stime;
131 }
132
133 void vtime_account_system(struct task_struct *tsk)
134 {
135         cputime_t delta = vtime_delta(tsk);
136
137         account_system_time(tsk, 0, delta, delta);
138 }
139
140 void vtime_account_idle(struct task_struct *tsk)
141 {
142         account_idle_time(vtime_delta(tsk));
143 }
144
145 #endif /* CONFIG_VIRT_CPU_ACCOUNTING */
146
147 static irqreturn_t
148 timer_interrupt (int irq, void *dev_id)
149 {
150         unsigned long new_itm;
151
152         if (cpu_is_offline(smp_processor_id())) {
153                 return IRQ_HANDLED;
154         }
155
156         platform_timer_interrupt(irq, dev_id);
157
158         new_itm = local_cpu_data->itm_next;
159
160         if (!time_after(ia64_get_itc(), new_itm))
161                 printk(KERN_ERR "Oops: timer tick before it's due (itc=%lx,itm=%lx)\n",
162                        ia64_get_itc(), new_itm);
163
164         profile_tick(CPU_PROFILING);
165
166         if (paravirt_do_steal_accounting(&new_itm))
167                 goto skip_process_time_accounting;
168
169         while (1) {
170                 update_process_times(user_mode(get_irq_regs()));
171
172                 new_itm += local_cpu_data->itm_delta;
173
174                 if (smp_processor_id() == time_keeper_id)
175                         xtime_update(1);
176
177                 local_cpu_data->itm_next = new_itm;
178
179                 if (time_after(new_itm, ia64_get_itc()))
180                         break;
181
182                 /*
183                  * Allow IPIs to interrupt the timer loop.
184                  */
185                 local_irq_enable();
186                 local_irq_disable();
187         }
188
189 skip_process_time_accounting:
190
191         do {
192                 /*
193                  * If we're too close to the next clock tick for
194                  * comfort, we increase the safety margin by
195                  * intentionally dropping the next tick(s).  We do NOT
196                  * update itm.next because that would force us to call
197                  * xtime_update() which in turn would let our clock run
198                  * too fast (with the potentially devastating effect
199                  * of losing monotony of time).
200                  */
201                 while (!time_after(new_itm, ia64_get_itc() + local_cpu_data->itm_delta/2))
202                         new_itm += local_cpu_data->itm_delta;
203                 ia64_set_itm(new_itm);
204                 /* double check, in case we got hit by a (slow) PMI: */
205         } while (time_after_eq(ia64_get_itc(), new_itm));
206         return IRQ_HANDLED;
207 }
208
209 /*
210  * Encapsulate access to the itm structure for SMP.
211  */
212 void
213 ia64_cpu_local_tick (void)
214 {
215         int cpu = smp_processor_id();
216         unsigned long shift = 0, delta;
217
218         /* arrange for the cycle counter to generate a timer interrupt: */
219         ia64_set_itv(IA64_TIMER_VECTOR);
220
221         delta = local_cpu_data->itm_delta;
222         /*
223          * Stagger the timer tick for each CPU so they don't occur all at (almost) the
224          * same time:
225          */
226         if (cpu) {
227                 unsigned long hi = 1UL << ia64_fls(cpu);
228                 shift = (2*(cpu - hi) + 1) * delta/hi/2;
229         }
230         local_cpu_data->itm_next = ia64_get_itc() + delta + shift;
231         ia64_set_itm(local_cpu_data->itm_next);
232 }
233
234 static int nojitter;
235
236 static int __init nojitter_setup(char *str)
237 {
238         nojitter = 1;
239         printk("Jitter checking for ITC timers disabled\n");
240         return 1;
241 }
242
243 __setup("nojitter", nojitter_setup);
244
245
246 void __devinit
247 ia64_init_itm (void)
248 {
249         unsigned long platform_base_freq, itc_freq;
250         struct pal_freq_ratio itc_ratio, proc_ratio;
251         long status, platform_base_drift, itc_drift;
252
253         /*
254          * According to SAL v2.6, we need to use a SAL call to determine the platform base
255          * frequency and then a PAL call to determine the frequency ratio between the ITC
256          * and the base frequency.
257          */
258         status = ia64_sal_freq_base(SAL_FREQ_BASE_PLATFORM,
259                                     &platform_base_freq, &platform_base_drift);
260         if (status != 0) {
261                 printk(KERN_ERR "SAL_FREQ_BASE_PLATFORM failed: %s\n", ia64_sal_strerror(status));
262         } else {
263                 status = ia64_pal_freq_ratios(&proc_ratio, NULL, &itc_ratio);
264                 if (status != 0)
265                         printk(KERN_ERR "PAL_FREQ_RATIOS failed with status=%ld\n", status);
266         }
267         if (status != 0) {
268                 /* invent "random" values */
269                 printk(KERN_ERR
270                        "SAL/PAL failed to obtain frequency info---inventing reasonable values\n");
271                 platform_base_freq = 100000000;
272                 platform_base_drift = -1;       /* no drift info */
273                 itc_ratio.num = 3;
274                 itc_ratio.den = 1;
275         }
276         if (platform_base_freq < 40000000) {
277                 printk(KERN_ERR "Platform base frequency %lu bogus---resetting to 75MHz!\n",
278                        platform_base_freq);
279                 platform_base_freq = 75000000;
280                 platform_base_drift = -1;
281         }
282         if (!proc_ratio.den)
283                 proc_ratio.den = 1;     /* avoid division by zero */
284         if (!itc_ratio.den)
285                 itc_ratio.den = 1;      /* avoid division by zero */
286
287         itc_freq = (platform_base_freq*itc_ratio.num)/itc_ratio.den;
288
289         local_cpu_data->itm_delta = (itc_freq + HZ/2) / HZ;
290         printk(KERN_DEBUG "CPU %d: base freq=%lu.%03luMHz, ITC ratio=%u/%u, "
291                "ITC freq=%lu.%03luMHz", smp_processor_id(),
292                platform_base_freq / 1000000, (platform_base_freq / 1000) % 1000,
293                itc_ratio.num, itc_ratio.den, itc_freq / 1000000, (itc_freq / 1000) % 1000);
294
295         if (platform_base_drift != -1) {
296                 itc_drift = platform_base_drift*itc_ratio.num/itc_ratio.den;
297                 printk("+/-%ldppm\n", itc_drift);
298         } else {
299                 itc_drift = -1;
300                 printk("\n");
301         }
302
303         local_cpu_data->proc_freq = (platform_base_freq*proc_ratio.num)/proc_ratio.den;
304         local_cpu_data->itc_freq = itc_freq;
305         local_cpu_data->cyc_per_usec = (itc_freq + USEC_PER_SEC/2) / USEC_PER_SEC;
306         local_cpu_data->nsec_per_cyc = ((NSEC_PER_SEC<<IA64_NSEC_PER_CYC_SHIFT)
307                                         + itc_freq/2)/itc_freq;
308
309         if (!(sal_platform_features & IA64_SAL_PLATFORM_FEATURE_ITC_DRIFT)) {
310 #ifdef CONFIG_SMP
311                 /* On IA64 in an SMP configuration ITCs are never accurately synchronized.
312                  * Jitter compensation requires a cmpxchg which may limit
313                  * the scalability of the syscalls for retrieving time.
314                  * The ITC synchronization is usually successful to within a few
315                  * ITC ticks but this is not a sure thing. If you need to improve
316                  * timer performance in SMP situations then boot the kernel with the
317                  * "nojitter" option. However, doing so may result in time fluctuating (maybe
318                  * even going backward) if the ITC offsets between the individual CPUs
319                  * are too large.
320                  */
321                 if (!nojitter)
322                         itc_jitter_data.itc_jitter = 1;
323 #endif
324         } else
325                 /*
326                  * ITC is drifty and we have not synchronized the ITCs in smpboot.c.
327                  * ITC values may fluctuate significantly between processors.
328                  * Clock should not be used for hrtimers. Mark itc as only
329                  * useful for boot and testing.
330                  *
331                  * Note that jitter compensation is off! There is no point of
332                  * synchronizing ITCs since they may be large differentials
333                  * that change over time.
334                  *
335                  * The only way to fix this would be to repeatedly sync the
336                  * ITCs. Until that time we have to avoid ITC.
337                  */
338                 clocksource_itc.rating = 50;
339
340         paravirt_init_missing_ticks_accounting(smp_processor_id());
341
342         /* avoid softlock up message when cpu is unplug and plugged again. */
343         touch_softlockup_watchdog();
344
345         /* Setup the CPU local timer tick */
346         ia64_cpu_local_tick();
347
348         if (!itc_clocksource) {
349                 clocksource_register_hz(&clocksource_itc,
350                                                 local_cpu_data->itc_freq);
351                 itc_clocksource = &clocksource_itc;
352         }
353 }
354
355 static cycle_t itc_get_cycles(struct clocksource *cs)
356 {
357         unsigned long lcycle, now, ret;
358
359         if (!itc_jitter_data.itc_jitter)
360                 return get_cycles();
361
362         lcycle = itc_jitter_data.itc_lastcycle;
363         now = get_cycles();
364         if (lcycle && time_after(lcycle, now))
365                 return lcycle;
366
367         /*
368          * Keep track of the last timer value returned.
369          * In an SMP environment, you could lose out in contention of
370          * cmpxchg. If so, your cmpxchg returns new value which the
371          * winner of contention updated to. Use the new value instead.
372          */
373         ret = cmpxchg(&itc_jitter_data.itc_lastcycle, lcycle, now);
374         if (unlikely(ret != lcycle))
375                 return ret;
376
377         return now;
378 }
379
380
381 static struct irqaction timer_irqaction = {
382         .handler =      timer_interrupt,
383         .flags =        IRQF_DISABLED | IRQF_IRQPOLL,
384         .name =         "timer"
385 };
386
387 static struct platform_device rtc_efi_dev = {
388         .name = "rtc-efi",
389         .id = -1,
390 };
391
392 static int __init rtc_init(void)
393 {
394         if (platform_device_register(&rtc_efi_dev) < 0)
395                 printk(KERN_ERR "unable to register rtc device...\n");
396
397         /* not necessarily an error */
398         return 0;
399 }
400 module_init(rtc_init);
401
402 void read_persistent_clock(struct timespec *ts)
403 {
404         efi_gettimeofday(ts);
405 }
406
407 void __init
408 time_init (void)
409 {
410         register_percpu_irq(IA64_TIMER_VECTOR, &timer_irqaction);
411         ia64_init_itm();
412 }
413
414 /*
415  * Generic udelay assumes that if preemption is allowed and the thread
416  * migrates to another CPU, that the ITC values are synchronized across
417  * all CPUs.
418  */
419 static void
420 ia64_itc_udelay (unsigned long usecs)
421 {
422         unsigned long start = ia64_get_itc();
423         unsigned long end = start + usecs*local_cpu_data->cyc_per_usec;
424
425         while (time_before(ia64_get_itc(), end))
426                 cpu_relax();
427 }
428
429 void (*ia64_udelay)(unsigned long usecs) = &ia64_itc_udelay;
430
431 void
432 udelay (unsigned long usecs)
433 {
434         (*ia64_udelay)(usecs);
435 }
436 EXPORT_SYMBOL(udelay);
437
438 /* IA64 doesn't cache the timezone */
439 void update_vsyscall_tz(void)
440 {
441 }
442
443 void update_vsyscall_old(struct timespec *wall, struct timespec *wtm,
444                         struct clocksource *c, u32 mult)
445 {
446         write_seqcount_begin(&fsyscall_gtod_data.seq);
447
448         /* copy fsyscall clock data */
449         fsyscall_gtod_data.clk_mask = c->mask;
450         fsyscall_gtod_data.clk_mult = mult;
451         fsyscall_gtod_data.clk_shift = c->shift;
452         fsyscall_gtod_data.clk_fsys_mmio = c->archdata.fsys_mmio;
453         fsyscall_gtod_data.clk_cycle_last = c->cycle_last;
454
455         /* copy kernel time structures */
456         fsyscall_gtod_data.wall_time.tv_sec = wall->tv_sec;
457         fsyscall_gtod_data.wall_time.tv_nsec = wall->tv_nsec;
458         fsyscall_gtod_data.monotonic_time.tv_sec = wtm->tv_sec
459                                                         + wall->tv_sec;
460         fsyscall_gtod_data.monotonic_time.tv_nsec = wtm->tv_nsec
461                                                         + wall->tv_nsec;
462
463         /* normalize */
464         while (fsyscall_gtod_data.monotonic_time.tv_nsec >= NSEC_PER_SEC) {
465                 fsyscall_gtod_data.monotonic_time.tv_nsec -= NSEC_PER_SEC;
466                 fsyscall_gtod_data.monotonic_time.tv_sec++;
467         }
468
469         write_seqcount_end(&fsyscall_gtod_data.seq);
470 }
471