linux/arch/ia64/kernel/time.c
<<
>>
Prefs
   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
  33#include "fsyscall_gtod_data.h"
  34
  35static cycle_t itc_get_cycles(struct clocksource *cs);
  36
  37struct fsyscall_gtod_data_t fsyscall_gtod_data;
  38
  39struct itc_jitter_data_t itc_jitter_data;
  40
  41volatile int time_keeper_id = 0; /* smp_processor_id() of time-keeper */
  42
  43#ifdef CONFIG_IA64_DEBUG_IRQ
  44
  45unsigned long last_cli_ip;
  46EXPORT_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 */
  53unsigned long long sched_clock(void)
  54{
  55        return paravirt_sched_clock();
  56}
  57#endif
  58
  59#ifdef CONFIG_PARAVIRT
  60static void
  61paravirt_clocksource_resume(struct clocksource *cs)
  62{
  63        if (pv_time_ops.clocksource_resume)
  64                pv_time_ops.clocksource_resume();
  65}
  66#endif
  67
  68static 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};
  78static struct clocksource *itc_clocksource;
  79
  80#ifdef CONFIG_VIRT_CPU_ACCOUNTING
  81
  82#include <linux/kernel_stat.h>
  83
  84extern cputime_t cycle_to_cputime(u64 cyc);
  85
  86/*
  87 * Called from the context switch with interrupts disabled, to charge all
  88 * accumulated times to the current process, and to prepare accounting on
  89 * the next process.
  90 */
  91void ia64_account_on_switch(struct task_struct *prev, struct task_struct *next)
  92{
  93        struct thread_info *pi = task_thread_info(prev);
  94        struct thread_info *ni = task_thread_info(next);
  95        cputime_t delta_stime, delta_utime;
  96        __u64 now;
  97
  98        now = ia64_get_itc();
  99
 100        delta_stime = cycle_to_cputime(pi->ac_stime + (now - pi->ac_stamp));
 101        if (idle_task(smp_processor_id()) != prev)
 102                account_system_time(prev, 0, delta_stime, delta_stime);
 103        else
 104                account_idle_time(delta_stime);
 105
 106        if (pi->ac_utime) {
 107                delta_utime = cycle_to_cputime(pi->ac_utime);
 108                account_user_time(prev, delta_utime, delta_utime);
 109        }
 110
 111        pi->ac_stamp = ni->ac_stamp = now;
 112        ni->ac_stime = ni->ac_utime = 0;
 113}
 114
 115/*
 116 * Account time for a transition between system, hard irq or soft irq state.
 117 * Note that this function is called with interrupts enabled.
 118 */
 119void account_system_vtime(struct task_struct *tsk)
 120{
 121        struct thread_info *ti = task_thread_info(tsk);
 122        unsigned long flags;
 123        cputime_t delta_stime;
 124        __u64 now;
 125
 126        local_irq_save(flags);
 127
 128        now = ia64_get_itc();
 129
 130        delta_stime = cycle_to_cputime(ti->ac_stime + (now - ti->ac_stamp));
 131        if (irq_count() || idle_task(smp_processor_id()) != tsk)
 132                account_system_time(tsk, 0, delta_stime, delta_stime);
 133        else
 134                account_idle_time(delta_stime);
 135        ti->ac_stime = 0;
 136
 137        ti->ac_stamp = now;
 138
 139        local_irq_restore(flags);
 140}
 141EXPORT_SYMBOL_GPL(account_system_vtime);
 142
 143/*
 144 * Called from the timer interrupt handler to charge accumulated user time
 145 * to the current process.  Must be called with interrupts disabled.
 146 */
 147void account_process_tick(struct task_struct *p, int user_tick)
 148{
 149        struct thread_info *ti = task_thread_info(p);
 150        cputime_t delta_utime;
 151
 152        if (ti->ac_utime) {
 153                delta_utime = cycle_to_cputime(ti->ac_utime);
 154                account_user_time(p, delta_utime, delta_utime);
 155                ti->ac_utime = 0;
 156        }
 157}
 158
 159#endif /* CONFIG_VIRT_CPU_ACCOUNTING */
 160
 161static irqreturn_t
 162timer_interrupt (int irq, void *dev_id)
 163{
 164        unsigned long new_itm;
 165
 166        if (cpu_is_offline(smp_processor_id())) {
 167                return IRQ_HANDLED;
 168        }
 169
 170        platform_timer_interrupt(irq, dev_id);
 171
 172        new_itm = local_cpu_data->itm_next;
 173
 174        if (!time_after(ia64_get_itc(), new_itm))
 175                printk(KERN_ERR "Oops: timer tick before it's due (itc=%lx,itm=%lx)\n",
 176                       ia64_get_itc(), new_itm);
 177
 178        profile_tick(CPU_PROFILING);
 179
 180        if (paravirt_do_steal_accounting(&new_itm))
 181                goto skip_process_time_accounting;
 182
 183        while (1) {
 184                update_process_times(user_mode(get_irq_regs()));
 185
 186                new_itm += local_cpu_data->itm_delta;
 187
 188                if (smp_processor_id() == time_keeper_id)
 189                        xtime_update(1);
 190
 191                local_cpu_data->itm_next = new_itm;
 192
 193                if (time_after(new_itm, ia64_get_itc()))
 194                        break;
 195
 196                /*
 197                 * Allow IPIs to interrupt the timer loop.
 198                 */
 199                local_irq_enable();
 200                local_irq_disable();
 201        }
 202
 203skip_process_time_accounting:
 204
 205        do {
 206                /*
 207                 * If we're too close to the next clock tick for
 208                 * comfort, we increase the safety margin by
 209                 * intentionally dropping the next tick(s).  We do NOT
 210                 * update itm.next because that would force us to call
 211                 * xtime_update() which in turn would let our clock run
 212                 * too fast (with the potentially devastating effect
 213                 * of losing monotony of time).
 214                 */
 215                while (!time_after(new_itm, ia64_get_itc() + local_cpu_data->itm_delta/2))
 216                        new_itm += local_cpu_data->itm_delta;
 217                ia64_set_itm(new_itm);
 218                /* double check, in case we got hit by a (slow) PMI: */
 219        } while (time_after_eq(ia64_get_itc(), new_itm));
 220        return IRQ_HANDLED;
 221}
 222
 223/*
 224 * Encapsulate access to the itm structure for SMP.
 225 */
 226void
 227ia64_cpu_local_tick (void)
 228{
 229        int cpu = smp_processor_id();
 230        unsigned long shift = 0, delta;
 231
 232        /* arrange for the cycle counter to generate a timer interrupt: */
 233        ia64_set_itv(IA64_TIMER_VECTOR);
 234
 235        delta = local_cpu_data->itm_delta;
 236        /*
 237         * Stagger the timer tick for each CPU so they don't occur all at (almost) the
 238         * same time:
 239         */
 240        if (cpu) {
 241                unsigned long hi = 1UL << ia64_fls(cpu);
 242                shift = (2*(cpu - hi) + 1) * delta/hi/2;
 243        }
 244        local_cpu_data->itm_next = ia64_get_itc() + delta + shift;
 245        ia64_set_itm(local_cpu_data->itm_next);
 246}
 247
 248static int nojitter;
 249
 250static int __init nojitter_setup(char *str)
 251{
 252        nojitter = 1;
 253        printk("Jitter checking for ITC timers disabled\n");
 254        return 1;
 255}
 256
 257__setup("nojitter", nojitter_setup);
 258
 259
 260void __devinit
 261ia64_init_itm (void)
 262{
 263        unsigned long platform_base_freq, itc_freq;
 264        struct pal_freq_ratio itc_ratio, proc_ratio;
 265        long status, platform_base_drift, itc_drift;
 266
 267        /*
 268         * According to SAL v2.6, we need to use a SAL call to determine the platform base
 269         * frequency and then a PAL call to determine the frequency ratio between the ITC
 270         * and the base frequency.
 271         */
 272        status = ia64_sal_freq_base(SAL_FREQ_BASE_PLATFORM,
 273                                    &platform_base_freq, &platform_base_drift);
 274        if (status != 0) {
 275                printk(KERN_ERR "SAL_FREQ_BASE_PLATFORM failed: %s\n", ia64_sal_strerror(status));
 276        } else {
 277                status = ia64_pal_freq_ratios(&proc_ratio, NULL, &itc_ratio);
 278                if (status != 0)
 279                        printk(KERN_ERR "PAL_FREQ_RATIOS failed with status=%ld\n", status);
 280        }
 281        if (status != 0) {
 282                /* invent "random" values */
 283                printk(KERN_ERR
 284                       "SAL/PAL failed to obtain frequency info---inventing reasonable values\n");
 285                platform_base_freq = 100000000;
 286                platform_base_drift = -1;       /* no drift info */
 287                itc_ratio.num = 3;
 288                itc_ratio.den = 1;
 289        }
 290        if (platform_base_freq < 40000000) {
 291                printk(KERN_ERR "Platform base frequency %lu bogus---resetting to 75MHz!\n",
 292                       platform_base_freq);
 293                platform_base_freq = 75000000;
 294                platform_base_drift = -1;
 295        }
 296        if (!proc_ratio.den)
 297                proc_ratio.den = 1;     /* avoid division by zero */
 298        if (!itc_ratio.den)
 299                itc_ratio.den = 1;      /* avoid division by zero */
 300
 301        itc_freq = (platform_base_freq*itc_ratio.num)/itc_ratio.den;
 302
 303        local_cpu_data->itm_delta = (itc_freq + HZ/2) / HZ;
 304        printk(KERN_DEBUG "CPU %d: base freq=%lu.%03luMHz, ITC ratio=%u/%u, "
 305               "ITC freq=%lu.%03luMHz", smp_processor_id(),
 306               platform_base_freq / 1000000, (platform_base_freq / 1000) % 1000,
 307               itc_ratio.num, itc_ratio.den, itc_freq / 1000000, (itc_freq / 1000) % 1000);
 308
 309        if (platform_base_drift != -1) {
 310                itc_drift = platform_base_drift*itc_ratio.num/itc_ratio.den;
 311                printk("+/-%ldppm\n", itc_drift);
 312        } else {
 313                itc_drift = -1;
 314                printk("\n");
 315        }
 316
 317        local_cpu_data->proc_freq = (platform_base_freq*proc_ratio.num)/proc_ratio.den;
 318        local_cpu_data->itc_freq = itc_freq;
 319        local_cpu_data->cyc_per_usec = (itc_freq + USEC_PER_SEC/2) / USEC_PER_SEC;
 320        local_cpu_data->nsec_per_cyc = ((NSEC_PER_SEC<<IA64_NSEC_PER_CYC_SHIFT)
 321                                        + itc_freq/2)/itc_freq;
 322
 323        if (!(sal_platform_features & IA64_SAL_PLATFORM_FEATURE_ITC_DRIFT)) {
 324#ifdef CONFIG_SMP
 325                /* On IA64 in an SMP configuration ITCs are never accurately synchronized.
 326                 * Jitter compensation requires a cmpxchg which may limit
 327                 * the scalability of the syscalls for retrieving time.
 328                 * The ITC synchronization is usually successful to within a few
 329                 * ITC ticks but this is not a sure thing. If you need to improve
 330                 * timer performance in SMP situations then boot the kernel with the
 331                 * "nojitter" option. However, doing so may result in time fluctuating (maybe
 332                 * even going backward) if the ITC offsets between the individual CPUs
 333                 * are too large.
 334                 */
 335                if (!nojitter)
 336                        itc_jitter_data.itc_jitter = 1;
 337#endif
 338        } else
 339                /*
 340                 * ITC is drifty and we have not synchronized the ITCs in smpboot.c.
 341                 * ITC values may fluctuate significantly between processors.
 342                 * Clock should not be used for hrtimers. Mark itc as only
 343                 * useful for boot and testing.
 344                 *
 345                 * Note that jitter compensation is off! There is no point of
 346                 * synchronizing ITCs since they may be large differentials
 347                 * that change over time.
 348                 *
 349                 * The only way to fix this would be to repeatedly sync the
 350                 * ITCs. Until that time we have to avoid ITC.
 351                 */
 352                clocksource_itc.rating = 50;
 353
 354        paravirt_init_missing_ticks_accounting(smp_processor_id());
 355
 356        /* avoid softlock up message when cpu is unplug and plugged again. */
 357        touch_softlockup_watchdog();
 358
 359        /* Setup the CPU local timer tick */
 360        ia64_cpu_local_tick();
 361
 362        if (!itc_clocksource) {
 363                clocksource_register_hz(&clocksource_itc,
 364                                                local_cpu_data->itc_freq);
 365                itc_clocksource = &clocksource_itc;
 366        }
 367}
 368
 369static cycle_t itc_get_cycles(struct clocksource *cs)
 370{
 371        unsigned long lcycle, now, ret;
 372
 373        if (!itc_jitter_data.itc_jitter)
 374                return get_cycles();
 375
 376        lcycle = itc_jitter_data.itc_lastcycle;
 377        now = get_cycles();
 378        if (lcycle && time_after(lcycle, now))
 379                return lcycle;
 380
 381        /*
 382         * Keep track of the last timer value returned.
 383         * In an SMP environment, you could lose out in contention of
 384         * cmpxchg. If so, your cmpxchg returns new value which the
 385         * winner of contention updated to. Use the new value instead.
 386         */
 387        ret = cmpxchg(&itc_jitter_data.itc_lastcycle, lcycle, now);
 388        if (unlikely(ret != lcycle))
 389                return ret;
 390
 391        return now;
 392}
 393
 394
 395static struct irqaction timer_irqaction = {
 396        .handler =      timer_interrupt,
 397        .flags =        IRQF_DISABLED | IRQF_IRQPOLL,
 398        .name =         "timer"
 399};
 400
 401static struct platform_device rtc_efi_dev = {
 402        .name = "rtc-efi",
 403        .id = -1,
 404};
 405
 406static int __init rtc_init(void)
 407{
 408        if (platform_device_register(&rtc_efi_dev) < 0)
 409                printk(KERN_ERR "unable to register rtc device...\n");
 410
 411        /* not necessarily an error */
 412        return 0;
 413}
 414module_init(rtc_init);
 415
 416void read_persistent_clock(struct timespec *ts)
 417{
 418        efi_gettimeofday(ts);
 419}
 420
 421void __init
 422time_init (void)
 423{
 424        register_percpu_irq(IA64_TIMER_VECTOR, &timer_irqaction);
 425        ia64_init_itm();
 426}
 427
 428/*
 429 * Generic udelay assumes that if preemption is allowed and the thread
 430 * migrates to another CPU, that the ITC values are synchronized across
 431 * all CPUs.
 432 */
 433static void
 434ia64_itc_udelay (unsigned long usecs)
 435{
 436        unsigned long start = ia64_get_itc();
 437        unsigned long end = start + usecs*local_cpu_data->cyc_per_usec;
 438
 439        while (time_before(ia64_get_itc(), end))
 440                cpu_relax();
 441}
 442
 443void (*ia64_udelay)(unsigned long usecs) = &ia64_itc_udelay;
 444
 445void
 446udelay (unsigned long usecs)
 447{
 448        (*ia64_udelay)(usecs);
 449}
 450EXPORT_SYMBOL(udelay);
 451
 452/* IA64 doesn't cache the timezone */
 453void update_vsyscall_tz(void)
 454{
 455}
 456
 457void update_vsyscall(struct timespec *wall, struct timespec *wtm,
 458                        struct clocksource *c, u32 mult)
 459{
 460        write_seqcount_begin(&fsyscall_gtod_data.seq);
 461
 462        /* copy fsyscall clock data */
 463        fsyscall_gtod_data.clk_mask = c->mask;
 464        fsyscall_gtod_data.clk_mult = mult;
 465        fsyscall_gtod_data.clk_shift = c->shift;
 466        fsyscall_gtod_data.clk_fsys_mmio = c->archdata.fsys_mmio;
 467        fsyscall_gtod_data.clk_cycle_last = c->cycle_last;
 468
 469        /* copy kernel time structures */
 470        fsyscall_gtod_data.wall_time.tv_sec = wall->tv_sec;
 471        fsyscall_gtod_data.wall_time.tv_nsec = wall->tv_nsec;
 472        fsyscall_gtod_data.monotonic_time.tv_sec = wtm->tv_sec
 473                                                        + wall->tv_sec;
 474        fsyscall_gtod_data.monotonic_time.tv_nsec = wtm->tv_nsec
 475                                                        + wall->tv_nsec;
 476
 477        /* normalize */
 478        while (fsyscall_gtod_data.monotonic_time.tv_nsec >= NSEC_PER_SEC) {
 479                fsyscall_gtod_data.monotonic_time.tv_nsec -= NSEC_PER_SEC;
 480                fsyscall_gtod_data.monotonic_time.tv_sec++;
 481        }
 482
 483        write_seqcount_end(&fsyscall_gtod_data.seq);
 484}
 485
 486