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
  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 static 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 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         cputime_t delta_stime;
 108         __u64 now;
 109
 110         now = ia64_get_itc();
 111
 112         delta_stime = cycle_to_cputime(pi->ac_stime + (now - pi->ac_stamp));
 113         if (idle_task(smp_processor_id()) != prev)
 114                 account_system_time(prev, 0, delta_stime, delta_stime);
 115         else
 116                 account_idle_time(delta_stime);
 117
 118         vtime_account_user(prev);
 119
 120         pi->ac_stamp = ni->ac_stamp = now;
 121         ni->ac_stime = ni->ac_utime = 0;
 122 }
 123
 124 /*
 125  * Account time for a transition between system, hard irq or soft irq state.
 126  * Note that this function is called with interrupts enabled.
 127  */
 128 static cputime_t vtime_delta(struct task_struct *tsk)
 129 {
 130         struct thread_info *ti = task_thread_info(tsk);
 131         cputime_t delta_stime;
 132         __u64 now;
 133
 134         now = ia64_get_itc();
 135
 136         delta_stime = cycle_to_cputime(ti->ac_stime + (now - ti->ac_stamp));
 137         ti->ac_stime = 0;
 138         ti->ac_stamp = now;
 139
 140         return delta_stime;
 141 }
 142
 143 void vtime_account_system(struct task_struct *tsk)
 144 {
 145         cputime_t delta = vtime_delta(tsk);
 146
 147         account_system_time(tsk, 0, delta, delta);
 148 }
 149
 150 void vtime_account_idle(struct task_struct *tsk)
 151 {
 152         account_idle_time(vtime_delta(tsk));
 153 }
 154
 155 /*
 156  * Called from the timer interrupt handler to charge accumulated user time
 157  * to the current process.  Must be called with interrupts disabled.
 158  */
 159 void account_process_tick(struct task_struct *p, int user_tick)
 160 {
 161         vtime_account_user(p);
 162 }
 163
 164 #endif /* CONFIG_VIRT_CPU_ACCOUNTING */
 165
 166 static irqreturn_t
 167 timer_interrupt (int irq, void *dev_id)
 168 {
 169         unsigned long new_itm;
 170
 171         if (cpu_is_offline(smp_processor_id())) {
 172                 return IRQ_HANDLED;
 173         }
 174
 175         platform_timer_interrupt(irq, dev_id);
 176
 177         new_itm = local_cpu_data->itm_next;
 178
 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);
 182
 183         profile_tick(CPU_PROFILING);
 184
 185         if (paravirt_do_steal_accounting(&new_itm))
 186                 goto skip_process_time_accounting;
 187
 188         while (1) {
 189                 update_process_times(user_mode(get_irq_regs()));
 190
 191                 new_itm += local_cpu_data->itm_delta;
 192
 193                 if (smp_processor_id() == time_keeper_id)
 194                         xtime_update(1);
 195
 196                 local_cpu_data->itm_next = new_itm;
 197
 198                 if (time_after(new_itm, ia64_get_itc()))
 199                         break;
 200
 201                 /*
 202                  * Allow IPIs to interrupt the timer loop.
 203                  */
 204                 local_irq_enable();
 205                 local_irq_disable();
 206         }
 207
 208 skip_process_time_accounting:
 209
 210         do {
 211                 /*
 212                  * If we're too close to the next clock tick for
 213                  * comfort, we increase the safety margin by
 214                  * intentionally dropping the next tick(s).  We do NOT
 215                  * update itm.next because that would force us to call
 216                  * xtime_update() which in turn would let our clock run
 217                  * too fast (with the potentially devastating effect
 218                  * of losing monotony of time).
 219                  */
 220                 while (!time_after(new_itm, ia64_get_itc() + local_cpu_data->itm_delta/2))
 221                         new_itm += local_cpu_data->itm_delta;
 222                 ia64_set_itm(new_itm);
 223                 /* double check, in case we got hit by a (slow) PMI: */
 224         } while (time_after_eq(ia64_get_itc(), new_itm));
 225         return IRQ_HANDLED;
 226 }
 227
 228 /*
 229  * Encapsulate access to the itm structure for SMP.
 230  */
 231 void
 232 ia64_cpu_local_tick (void)
 233 {
 234         int cpu = smp_processor_id();
 235         unsigned long shift = 0, delta;
 236
 237         /* arrange for the cycle counter to generate a timer interrupt: */
 238         ia64_set_itv(IA64_TIMER_VECTOR);
 239
 240         delta = local_cpu_data->itm_delta;
 241         /*
 242          * Stagger the timer tick for each CPU so they don't occur all at (almost) the
 243          * same time:
 244          */
 245         if (cpu) {
 246                 unsigned long hi = 1UL << ia64_fls(cpu);
 247                 shift = (2*(cpu - hi) + 1) * delta/hi/2;
 248         }
 249         local_cpu_data->itm_next = ia64_get_itc() + delta + shift;
 250         ia64_set_itm(local_cpu_data->itm_next);
 251 }
 252
 253 static int nojitter;
 254
 255 static int __init nojitter_setup(char *str)
 256 {
 257         nojitter = 1;
 258         printk("Jitter checking for ITC timers disabled\n");
 259         return 1;
 260 }
 261
 262 __setup("nojitter", nojitter_setup);
 263
 264
 265 void __devinit
 266 ia64_init_itm (void)
 267 {
 268         unsigned long platform_base_freq, itc_freq;
 269         struct pal_freq_ratio itc_ratio, proc_ratio;
 270         long status, platform_base_drift, itc_drift;
 271
 272         /*
 273          * According to SAL v2.6, we need to use a SAL call to determine the platform base
 274          * frequency and then a PAL call to determine the frequency ratio between the ITC
 275          * and the base frequency.
 276          */
 277         status = ia64_sal_freq_base(SAL_FREQ_BASE_PLATFORM,
 278                                     &platform_base_freq, &platform_base_drift);
 279         if (status != 0) {
 280                 printk(KERN_ERR "SAL_FREQ_BASE_PLATFORM failed: %s\n", ia64_sal_strerror(status));
 281         } else {
 282                 status = ia64_pal_freq_ratios(&proc_ratio, NULL, &itc_ratio);
 283                 if (status != 0)
 284                         printk(KERN_ERR "PAL_FREQ_RATIOS failed with status=%ld\n", status);
 285         }
 286         if (status != 0) {
 287                 /* invent "random" values */
 288                 printk(KERN_ERR
 289                        "SAL/PAL failed to obtain frequency info---inventing reasonable values\n");
 290                 platform_base_freq = 100000000;
 291                 platform_base_drift = -1;       /* no drift info */
 292                 itc_ratio.num = 3;
 293                 itc_ratio.den = 1;
 294         }
 295         if (platform_base_freq < 40000000) {
 296                 printk(KERN_ERR "Platform base frequency %lu bogus---resetting to 75MHz!\n",
 297                        platform_base_freq);
 298                 platform_base_freq = 75000000;
 299                 platform_base_drift = -1;
 300         }
 301         if (!proc_ratio.den)
 302                 proc_ratio.den = 1;     /* avoid division by zero */
 303         if (!itc_ratio.den)
 304                 itc_ratio.den = 1;      /* avoid division by zero */
 305
 306         itc_freq = (platform_base_freq*itc_ratio.num)/itc_ratio.den;
 307
 308         local_cpu_data->itm_delta = (itc_freq + HZ/2) / HZ;
 309         printk(KERN_DEBUG "CPU %d: base freq=%lu.%03luMHz, ITC ratio=%u/%u, "
 310                "ITC freq=%lu.%03luMHz", smp_processor_id(),
 311                platform_base_freq / 1000000, (platform_base_freq / 1000) % 1000,
 312                itc_ratio.num, itc_ratio.den, itc_freq / 1000000, (itc_freq / 1000) % 1000);
 313
 314         if (platform_base_drift != -1) {
 315                 itc_drift = platform_base_drift*itc_ratio.num/itc_ratio.den;
 316                 printk("+/-%ldppm\n", itc_drift);
 317         } else {
 318                 itc_drift = -1;
 319                 printk("\n");
 320         }
 321
 322         local_cpu_data->proc_freq = (platform_base_freq*proc_ratio.num)/proc_ratio.den;
 323         local_cpu_data->itc_freq = itc_freq;
 324         local_cpu_data->cyc_per_usec = (itc_freq + USEC_PER_SEC/2) / USEC_PER_SEC;
 325         local_cpu_data->nsec_per_cyc = ((NSEC_PER_SEC<<IA64_NSEC_PER_CYC_SHIFT)
 326                                         + itc_freq/2)/itc_freq;
 327
 328         if (!(sal_platform_features & IA64_SAL_PLATFORM_FEATURE_ITC_DRIFT)) {
 329 #ifdef CONFIG_SMP
 330                 /* On IA64 in an SMP configuration ITCs are never accurately synchronized.
 331                  * Jitter compensation requires a cmpxchg which may limit
 332                  * the scalability of the syscalls for retrieving time.
 333                  * The ITC synchronization is usually successful to within a few
 334                  * ITC ticks but this is not a sure thing. If you need to improve
 335                  * timer performance in SMP situations then boot the kernel with the
 336                  * "nojitter" option. However, doing so may result in time fluctuating (maybe
 337                  * even going backward) if the ITC offsets between the individual CPUs
 338                  * are too large.
 339                  */
 340                 if (!nojitter)
 341                         itc_jitter_data.itc_jitter = 1;
 342 #endif
 343         } else
 344                 /*
 345                  * ITC is drifty and we have not synchronized the ITCs in smpboot.c.
 346                  * ITC values may fluctuate significantly between processors.
 347                  * Clock should not be used for hrtimers. Mark itc as only
 348                  * useful for boot and testing.
 349                  *
 350                  * Note that jitter compensation is off! There is no point of
 351                  * synchronizing ITCs since they may be large differentials
 352                  * that change over time.
 353                  *
 354                  * The only way to fix this would be to repeatedly sync the
 355                  * ITCs. Until that time we have to avoid ITC.
 356                  */
 357                 clocksource_itc.rating = 50;
 358
 359         paravirt_init_missing_ticks_accounting(smp_processor_id());
 360
 361         /* avoid softlock up message when cpu is unplug and plugged again. */
 362         touch_softlockup_watchdog();
 363
 364         /* Setup the CPU local timer tick */
 365         ia64_cpu_local_tick();
 366
 367         if (!itc_clocksource) {
 368                 clocksource_register_hz(&clocksource_itc,
 369                                                 local_cpu_data->itc_freq);
 370                 itc_clocksource = &clocksource_itc;
 371         }
 372 }
 373
 374 static cycle_t itc_get_cycles(struct clocksource *cs)
 375 {
 376         unsigned long lcycle, now, ret;
 377
 378         if (!itc_jitter_data.itc_jitter)
 379                 return get_cycles();
 380
 381         lcycle = itc_jitter_data.itc_lastcycle;
 382         now = get_cycles();
 383         if (lcycle && time_after(lcycle, now))
 384                 return lcycle;
 385
 386         /*
 387          * Keep track of the last timer value returned.
 388          * In an SMP environment, you could lose out in contention of
 389          * cmpxchg. If so, your cmpxchg returns new value which the
 390          * winner of contention updated to. Use the new value instead.
 391          */
 392         ret = cmpxchg(&itc_jitter_data.itc_lastcycle, lcycle, now);
 393         if (unlikely(ret != lcycle))
 394                 return ret;
 395
 396         return now;
 397 }
 398
 399
 400 static struct irqaction timer_irqaction = {
 401         .handler =      timer_interrupt,
 402         .flags =        IRQF_DISABLED | IRQF_IRQPOLL,
 403         .name =         "timer"
 404 };
 405
 406 static struct platform_device rtc_efi_dev = {
 407         .name = "rtc-efi",
 408         .id = -1,
 409 };
 410
 411 static int __init rtc_init(void)
 412 {
 413         if (platform_device_register(&rtc_efi_dev) < 0)
 414                 printk(KERN_ERR "unable to register rtc device...\n");
 415
 416         /* not necessarily an error */
 417         return 0;
 418 }
 419 module_init(rtc_init);
 420
 421 void read_persistent_clock(struct timespec *ts)
 422 {
 423         efi_gettimeofday(ts);
 424 }
 425
 426 void __init
 427 time_init (void)
 428 {
 429         register_percpu_irq(IA64_TIMER_VECTOR, &timer_irqaction);
 430         ia64_init_itm();
 431 }
 432
 433 /*
 434  * Generic udelay assumes that if preemption is allowed and the thread
 435  * migrates to another CPU, that the ITC values are synchronized across
 436  * all CPUs.
 437  */
 438 static void
 439 ia64_itc_udelay (unsigned long usecs)
 440 {
 441         unsigned long start = ia64_get_itc();
 442         unsigned long end = start + usecs*local_cpu_data->cyc_per_usec;
 443
 444         while (time_before(ia64_get_itc(), end))
 445                 cpu_relax();
 446 }
 447
 448 void (*ia64_udelay)(unsigned long usecs) = &ia64_itc_udelay;
 449
 450 void
 451 udelay (unsigned long usecs)
 452 {
 453         (*ia64_udelay)(usecs);
 454 }
 455 EXPORT_SYMBOL(udelay);
 456
 457 /* IA64 doesn't cache the timezone */
 458 void update_vsyscall_tz(void)
 459 {
 460 }
 461
 462 void update_vsyscall(struct timespec *wall, struct timespec *wtm,
 463                         struct clocksource *c, u32 mult)
 464 {
 465         write_seqcount_begin(&fsyscall_gtod_data.seq);
 466
 467         /* copy fsyscall clock data */
 468         fsyscall_gtod_data.clk_mask = c->mask;
 469         fsyscall_gtod_data.clk_mult = mult;
 470         fsyscall_gtod_data.clk_shift = c->shift;
 471         fsyscall_gtod_data.clk_fsys_mmio = c->archdata.fsys_mmio;
 472         fsyscall_gtod_data.clk_cycle_last = c->cycle_last;
 473
 474         /* copy kernel time structures */
 475         fsyscall_gtod_data.wall_time.tv_sec = wall->tv_sec;
 476         fsyscall_gtod_data.wall_time.tv_nsec = wall->tv_nsec;
 477         fsyscall_gtod_data.monotonic_time.tv_sec = wtm->tv_sec
 478                                                         + wall->tv_sec;
 479         fsyscall_gtod_data.monotonic_time.tv_nsec = wtm->tv_nsec
 480                                                         + wall->tv_nsec;
 481
 482         /* normalize */
 483         while (fsyscall_gtod_data.monotonic_time.tv_nsec >= NSEC_PER_SEC) {
 484                 fsyscall_gtod_data.monotonic_time.tv_nsec -= NSEC_PER_SEC;
 485                 fsyscall_gtod_data.monotonic_time.tv_sec++;
 486         }
 487
 488         write_seqcount_end(&fsyscall_gtod_data.seq);
 489 }
 490