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 #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 #include <asm/system.h>
  33
  34 #include "fsyscall_gtod_data.h"
  35
  36 static cycle_t itc_get_cycles(struct clocksource *cs);
  37
  38 struct fsyscall_gtod_data_t fsyscall_gtod_data;
  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 #ifdef CONFIG_PARAVIRT
  52 /* We need to define a real function for sched_clock, to override the
  53    weak default version */
  54 unsigned long long sched_clock(void)
  55 {
  56         return paravirt_sched_clock();
  57 }
  58 #endif
  59
  60 #ifdef CONFIG_PARAVIRT
  61 static void
  62 paravirt_clocksource_resume(struct clocksource *cs)
  63 {
  64         if (pv_time_ops.clocksource_resume)
  65                 pv_time_ops.clocksource_resume();
  66 }
  67 #endif
  68
  69 static struct clocksource clocksource_itc = {
  70         .name           = "itc",
  71         .rating         = 350,
  72         .read           = itc_get_cycles,
  73         .mask           = CLOCKSOURCE_MASK(64),
  74         .flags          = CLOCK_SOURCE_IS_CONTINUOUS,
  75 #ifdef CONFIG_PARAVIRT
  76         .resume         = paravirt_clocksource_resume,
  77 #endif
  78 };
  79 static struct clocksource *itc_clocksource;
  80
  81 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
  82
  83 #include <linux/kernel_stat.h>
  84
  85 extern cputime_t cycle_to_cputime(u64 cyc);
  86
  87 /*
  88  * Called from the context switch with interrupts disabled, to charge all
  89  * accumulated times to the current process, and to prepare accounting on
  90  * the next process.
  91  */
  92 void ia64_account_on_switch(struct task_struct *prev, struct task_struct *next)
  93 {
  94         struct thread_info *pi = task_thread_info(prev);
  95         struct thread_info *ni = task_thread_info(next);
  96         cputime_t delta_stime, delta_utime;
  97         __u64 now;
  98
  99         now = ia64_get_itc();
 100
 101         delta_stime = cycle_to_cputime(pi->ac_stime + (now - pi->ac_stamp));
 102         if (idle_task(smp_processor_id()) != prev)
 103                 account_system_time(prev, 0, delta_stime, delta_stime);
 104         else
 105                 account_idle_time(delta_stime);
 106
 107         if (pi->ac_utime) {
 108                 delta_utime = cycle_to_cputime(pi->ac_utime);
 109                 account_user_time(prev, delta_utime, delta_utime);
 110         }
 111
 112         pi->ac_stamp = ni->ac_stamp = now;
 113         ni->ac_stime = ni->ac_utime = 0;
 114 }
 115
 116 /*
 117  * Account time for a transition between system, hard irq or soft irq state.
 118  * Note that this function is called with interrupts enabled.
 119  */
 120 void account_system_vtime(struct task_struct *tsk)
 121 {
 122         struct thread_info *ti = task_thread_info(tsk);
 123         unsigned long flags;
 124         cputime_t delta_stime;
 125         __u64 now;
 126
 127         local_irq_save(flags);
 128
 129         now = ia64_get_itc();
 130
 131         delta_stime = cycle_to_cputime(ti->ac_stime + (now - ti->ac_stamp));
 132         if (irq_count() || idle_task(smp_processor_id()) != tsk)
 133                 account_system_time(tsk, 0, delta_stime, delta_stime);
 134         else
 135                 account_idle_time(delta_stime);
 136         ti->ac_stime = 0;
 137
 138         ti->ac_stamp = now;
 139
 140         local_irq_restore(flags);
 141 }
 142 EXPORT_SYMBOL_GPL(account_system_vtime);
 143
 144 /*
 145  * Called from the timer interrupt handler to charge accumulated user time
 146  * to the current process.  Must be called with interrupts disabled.
 147  */
 148 void account_process_tick(struct task_struct *p, int user_tick)
 149 {
 150         struct thread_info *ti = task_thread_info(p);
 151         cputime_t delta_utime;
 152
 153         if (ti->ac_utime) {
 154                 delta_utime = cycle_to_cputime(ti->ac_utime);
 155                 account_user_time(p, delta_utime, delta_utime);
 156                 ti->ac_utime = 0;
 157         }
 158 }
 159
 160 #endif /* CONFIG_VIRT_CPU_ACCOUNTING */
 161
 162 static irqreturn_t
 163 timer_interrupt (int irq, void *dev_id)
 164 {
 165         unsigned long new_itm;
 166
 167         if (cpu_is_offline(smp_processor_id())) {
 168                 return IRQ_HANDLED;
 169         }
 170
 171         platform_timer_interrupt(irq, dev_id);
 172
 173         new_itm = local_cpu_data->itm_next;
 174
 175         if (!time_after(ia64_get_itc(), new_itm))
 176                 printk(KERN_ERR "Oops: timer tick before it's due (itc=%lx,itm=%lx)\n",
 177                        ia64_get_itc(), new_itm);
 178
 179         profile_tick(CPU_PROFILING);
 180
 181         if (paravirt_do_steal_accounting(&new_itm))
 182                 goto skip_process_time_accounting;
 183
 184         while (1) {
 185                 update_process_times(user_mode(get_irq_regs()));
 186
 187                 new_itm += local_cpu_data->itm_delta;
 188
 189                 if (smp_processor_id() == time_keeper_id)
 190                         xtime_update(1);
 191
 192                 local_cpu_data->itm_next = new_itm;
 193
 194                 if (time_after(new_itm, ia64_get_itc()))
 195                         break;
 196
 197                 /*
 198                  * Allow IPIs to interrupt the timer loop.
 199                  */
 200                 local_irq_enable();
 201                 local_irq_disable();
 202         }
 203
 204 skip_process_time_accounting:
 205
 206         do {
 207                 /*
 208                  * If we're too close to the next clock tick for
 209                  * comfort, we increase the safety margin by
 210                  * intentionally dropping the next tick(s).  We do NOT
 211                  * update itm.next because that would force us to call
 212                  * xtime_update() which in turn would let our clock run
 213                  * too fast (with the potentially devastating effect
 214                  * of losing monotony of time).
 215                  */
 216                 while (!time_after(new_itm, ia64_get_itc() + local_cpu_data->itm_delta/2))
 217                         new_itm += local_cpu_data->itm_delta;
 218                 ia64_set_itm(new_itm);
 219                 /* double check, in case we got hit by a (slow) PMI: */
 220         } while (time_after_eq(ia64_get_itc(), new_itm));
 221         return IRQ_HANDLED;
 222 }
 223
 224 /*
 225  * Encapsulate access to the itm structure for SMP.
 226  */
 227 void
 228 ia64_cpu_local_tick (void)
 229 {
 230         int cpu = smp_processor_id();
 231         unsigned long shift = 0, delta;
 232
 233         /* arrange for the cycle counter to generate a timer interrupt: */
 234         ia64_set_itv(IA64_TIMER_VECTOR);
 235
 236         delta = local_cpu_data->itm_delta;
 237         /*
 238          * Stagger the timer tick for each CPU so they don't occur all at (almost) the
 239          * same time:
 240          */
 241         if (cpu) {
 242                 unsigned long hi = 1UL << ia64_fls(cpu);
 243                 shift = (2*(cpu - hi) + 1) * delta/hi/2;
 244         }
 245         local_cpu_data->itm_next = ia64_get_itc() + delta + shift;
 246         ia64_set_itm(local_cpu_data->itm_next);
 247 }
 248
 249 static int nojitter;
 250
 251 static int __init nojitter_setup(char *str)
 252 {
 253         nojitter = 1;
 254         printk("Jitter checking for ITC timers disabled\n");
 255         return 1;
 256 }
 257
 258 __setup("nojitter", nojitter_setup);
 259
 260
 261 void __devinit
 262 ia64_init_itm (void)
 263 {
 264         unsigned long platform_base_freq, itc_freq;
 265         struct pal_freq_ratio itc_ratio, proc_ratio;
 266         long status, platform_base_drift, itc_drift;
 267
 268         /*
 269          * According to SAL v2.6, we need to use a SAL call to determine the platform base
 270          * frequency and then a PAL call to determine the frequency ratio between the ITC
 271          * and the base frequency.
 272          */
 273         status = ia64_sal_freq_base(SAL_FREQ_BASE_PLATFORM,
 274                                     &platform_base_freq, &platform_base_drift);
 275         if (status != 0) {
 276                 printk(KERN_ERR "SAL_FREQ_BASE_PLATFORM failed: %s\n", ia64_sal_strerror(status));
 277         } else {
 278                 status = ia64_pal_freq_ratios(&proc_ratio, NULL, &itc_ratio);
 279                 if (status != 0)
 280                         printk(KERN_ERR "PAL_FREQ_RATIOS failed with status=%ld\n", status);
 281         }
 282         if (status != 0) {
 283                 /* invent "random" values */
 284                 printk(KERN_ERR
 285                        "SAL/PAL failed to obtain frequency info---inventing reasonable values\n");
 286                 platform_base_freq = 100000000;
 287                 platform_base_drift = -1;       /* no drift info */
 288                 itc_ratio.num = 3;
 289                 itc_ratio.den = 1;
 290         }
 291         if (platform_base_freq < 40000000) {
 292                 printk(KERN_ERR "Platform base frequency %lu bogus---resetting to 75MHz!\n",
 293                        platform_base_freq);
 294                 platform_base_freq = 75000000;
 295                 platform_base_drift = -1;
 296         }
 297         if (!proc_ratio.den)
 298                 proc_ratio.den = 1;     /* avoid division by zero */
 299         if (!itc_ratio.den)
 300                 itc_ratio.den = 1;      /* avoid division by zero */
 301
 302         itc_freq = (platform_base_freq*itc_ratio.num)/itc_ratio.den;
 303
 304         local_cpu_data->itm_delta = (itc_freq + HZ/2) / HZ;
 305         printk(KERN_DEBUG "CPU %d: base freq=%lu.%03luMHz, ITC ratio=%u/%u, "
 306                "ITC freq=%lu.%03luMHz", smp_processor_id(),
 307                platform_base_freq / 1000000, (platform_base_freq / 1000) % 1000,
 308                itc_ratio.num, itc_ratio.den, itc_freq / 1000000, (itc_freq / 1000) % 1000);
 309
 310         if (platform_base_drift != -1) {
 311                 itc_drift = platform_base_drift*itc_ratio.num/itc_ratio.den;
 312                 printk("+/-%ldppm\n", itc_drift);
 313         } else {
 314                 itc_drift = -1;
 315                 printk("\n");
 316         }
 317
 318         local_cpu_data->proc_freq = (platform_base_freq*proc_ratio.num)/proc_ratio.den;
 319         local_cpu_data->itc_freq = itc_freq;
 320         local_cpu_data->cyc_per_usec = (itc_freq + USEC_PER_SEC/2) / USEC_PER_SEC;
 321         local_cpu_data->nsec_per_cyc = ((NSEC_PER_SEC<<IA64_NSEC_PER_CYC_SHIFT)
 322                                         + itc_freq/2)/itc_freq;
 323
 324         if (!(sal_platform_features & IA64_SAL_PLATFORM_FEATURE_ITC_DRIFT)) {
 325 #ifdef CONFIG_SMP
 326                 /* On IA64 in an SMP configuration ITCs are never accurately synchronized.
 327                  * Jitter compensation requires a cmpxchg which may limit
 328                  * the scalability of the syscalls for retrieving time.
 329                  * The ITC synchronization is usually successful to within a few
 330                  * ITC ticks but this is not a sure thing. If you need to improve
 331                  * timer performance in SMP situations then boot the kernel with the
 332                  * "nojitter" option. However, doing so may result in time fluctuating (maybe
 333                  * even going backward) if the ITC offsets between the individual CPUs
 334                  * are too large.
 335                  */
 336                 if (!nojitter)
 337                         itc_jitter_data.itc_jitter = 1;
 338 #endif
 339         } else
 340                 /*
 341                  * ITC is drifty and we have not synchronized the ITCs in smpboot.c.
 342                  * ITC values may fluctuate significantly between processors.
 343                  * Clock should not be used for hrtimers. Mark itc as only
 344                  * useful for boot and testing.
 345                  *
 346                  * Note that jitter compensation is off! There is no point of
 347                  * synchronizing ITCs since they may be large differentials
 348                  * that change over time.
 349                  *
 350                  * The only way to fix this would be to repeatedly sync the
 351                  * ITCs. Until that time we have to avoid ITC.
 352                  */
 353                 clocksource_itc.rating = 50;
 354
 355         paravirt_init_missing_ticks_accounting(smp_processor_id());
 356
 357         /* avoid softlock up message when cpu is unplug and plugged again. */
 358         touch_softlockup_watchdog();
 359
 360         /* Setup the CPU local timer tick */
 361         ia64_cpu_local_tick();
 362
 363         if (!itc_clocksource) {
 364                 clocksource_register_hz(&clocksource_itc,
 365                                                 local_cpu_data->itc_freq);
 366                 itc_clocksource = &clocksource_itc;
 367         }
 368 }
 369
 370 static cycle_t itc_get_cycles(struct clocksource *cs)
 371 {
 372         unsigned long lcycle, now, ret;
 373
 374         if (!itc_jitter_data.itc_jitter)
 375                 return get_cycles();
 376
 377         lcycle = itc_jitter_data.itc_lastcycle;
 378         now = get_cycles();
 379         if (lcycle && time_after(lcycle, now))
 380                 return lcycle;
 381
 382         /*
 383          * Keep track of the last timer value returned.
 384          * In an SMP environment, you could lose out in contention of
 385          * cmpxchg. If so, your cmpxchg returns new value which the
 386          * winner of contention updated to. Use the new value instead.
 387          */
 388         ret = cmpxchg(&itc_jitter_data.itc_lastcycle, lcycle, now);
 389         if (unlikely(ret != lcycle))
 390                 return ret;
 391
 392         return now;
 393 }
 394
 395
 396 static struct irqaction timer_irqaction = {
 397         .handler =      timer_interrupt,
 398         .flags =        IRQF_DISABLED | IRQF_IRQPOLL,
 399         .name =         "timer"
 400 };
 401
 402 static struct platform_device rtc_efi_dev = {
 403         .name = "rtc-efi",
 404         .id = -1,
 405 };
 406
 407 static int __init rtc_init(void)
 408 {
 409         if (platform_device_register(&rtc_efi_dev) < 0)
 410                 printk(KERN_ERR "unable to register rtc device...\n");
 411
 412         /* not necessarily an error */
 413         return 0;
 414 }
 415 module_init(rtc_init);
 416
 417 void read_persistent_clock(struct timespec *ts)
 418 {
 419         efi_gettimeofday(ts);
 420 }
 421
 422 void __init
 423 time_init (void)
 424 {
 425         register_percpu_irq(IA64_TIMER_VECTOR, &timer_irqaction);
 426         ia64_init_itm();
 427 }
 428
 429 /*
 430  * Generic udelay assumes that if preemption is allowed and the thread
 431  * migrates to another CPU, that the ITC values are synchronized across
 432  * all CPUs.
 433  */
 434 static void
 435 ia64_itc_udelay (unsigned long usecs)
 436 {
 437         unsigned long start = ia64_get_itc();
 438         unsigned long end = start + usecs*local_cpu_data->cyc_per_usec;
 439
 440         while (time_before(ia64_get_itc(), end))
 441                 cpu_relax();
 442 }
 443
 444 void (*ia64_udelay)(unsigned long usecs) = &ia64_itc_udelay;
 445
 446 void
 447 udelay (unsigned long usecs)
 448 {
 449         (*ia64_udelay)(usecs);
 450 }
 451 EXPORT_SYMBOL(udelay);
 452
 453 /* IA64 doesn't cache the timezone */
 454 void update_vsyscall_tz(void)
 455 {
 456 }
 457
 458 void update_vsyscall(struct timespec *wall, struct timespec *wtm,
 459                         struct clocksource *c, u32 mult)
 460 {
 461         write_seqcount_begin(&fsyscall_gtod_data.seq);
 462
 463         /* copy fsyscall clock data */
 464         fsyscall_gtod_data.clk_mask = c->mask;
 465         fsyscall_gtod_data.clk_mult = mult;
 466         fsyscall_gtod_data.clk_shift = c->shift;
 467         fsyscall_gtod_data.clk_fsys_mmio = c->archdata.fsys_mmio;
 468         fsyscall_gtod_data.clk_cycle_last = c->cycle_last;
 469
 470         /* copy kernel time structures */
 471         fsyscall_gtod_data.wall_time.tv_sec = wall->tv_sec;
 472         fsyscall_gtod_data.wall_time.tv_nsec = wall->tv_nsec;
 473         fsyscall_gtod_data.monotonic_time.tv_sec = wtm->tv_sec
 474                                                         + wall->tv_sec;
 475         fsyscall_gtod_data.monotonic_time.tv_nsec = wtm->tv_nsec
 476                                                         + wall->tv_nsec;
 477
 478         /* normalize */
 479         while (fsyscall_gtod_data.monotonic_time.tv_nsec >= NSEC_PER_SEC) {
 480                 fsyscall_gtod_data.monotonic_time.tv_nsec -= NSEC_PER_SEC;
 481                 fsyscall_gtod_data.monotonic_time.tv_sec++;
 482         }
 483
 484         write_seqcount_end(&fsyscall_gtod_data.seq);
 485 }
 486