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
  24 #include <asm/machvec.h>
  25 #include <asm/delay.h>
  26 #include <asm/hw_irq.h>
  27 #include <asm/paravirt.h>
  28 #include <asm/ptrace.h>
  29 #include <asm/sal.h>
  30 #include <asm/sections.h>
  31 #include <asm/system.h>
  32
  33 #include "fsyscall_gtod_data.h"
  34
  35 static cycle_t itc_get_cycles(void);
  36
  37 struct fsyscall_gtod_data_t fsyscall_gtod_data = {
  38         .lock = SEQLOCK_UNLOCKED,
  39 };
  40
  41 struct itc_jitter_data_t itc_jitter_data;
  42
  43 volatile int time_keeper_id = 0; /* smp_processor_id() of time-keeper */
  44
  45 #ifdef CONFIG_IA64_DEBUG_IRQ
  46
  47 unsigned long last_cli_ip;
  48 EXPORT_SYMBOL(last_cli_ip);
  49
  50 #endif
  51
  52 #ifdef CONFIG_PARAVIRT
  53 /* We need to define a real function for sched_clock, to override the
  54    weak default version */
  55 unsigned long long sched_clock(void)
  56 {
  57         return paravirt_sched_clock();
  58 }
  59 #endif
  60
  61 #ifdef CONFIG_PARAVIRT
  62 static void
  63 paravirt_clocksource_resume(void)
  64 {
  65         if (pv_time_ops.clocksource_resume)
  66                 pv_time_ops.clocksource_resume();
  67 }
  68 #endif
  69
  70 static struct clocksource clocksource_itc = {
  71         .name           = "itc",
  72         .rating         = 350,
  73         .read           = itc_get_cycles,
  74         .mask           = CLOCKSOURCE_MASK(64),
  75         .mult           = 0, /*to be calculated*/
  76         .shift          = 16,
  77         .flags          = CLOCK_SOURCE_IS_CONTINUOUS,
  78 #ifdef CONFIG_PARAVIRT
  79         .resume         = paravirt_clocksource_resume,
  80 #endif
  81 };
  82 static struct clocksource *itc_clocksource;
  83
  84 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
  85
  86 #include <linux/kernel_stat.h>
  87
  88 extern cputime_t cycle_to_cputime(u64 cyc);
  89
  90 /*
  91  * Called from the context switch with interrupts disabled, to charge all
  92  * accumulated times to the current process, and to prepare accounting on
  93  * the next process.
  94  */
  95 void ia64_account_on_switch(struct task_struct *prev, struct task_struct *next)
  96 {
  97         struct thread_info *pi = task_thread_info(prev);
  98         struct thread_info *ni = task_thread_info(next);
  99         cputime_t delta_stime, delta_utime;
 100         __u64 now;
 101
 102         now = ia64_get_itc();
 103
 104         delta_stime = cycle_to_cputime(pi->ac_stime + (now - pi->ac_stamp));
 105         if (idle_task(smp_processor_id()) != prev)
 106                 account_system_time(prev, 0, delta_stime, delta_stime);
 107         else
 108                 account_idle_time(delta_stime);
 109
 110         if (pi->ac_utime) {
 111                 delta_utime = cycle_to_cputime(pi->ac_utime);
 112                 account_user_time(prev, delta_utime, delta_utime);
 113         }
 114
 115         pi->ac_stamp = ni->ac_stamp = now;
 116         ni->ac_stime = ni->ac_utime = 0;
 117 }
 118
 119 /*
 120  * Account time for a transition between system, hard irq or soft irq state.
 121  * Note that this function is called with interrupts enabled.
 122  */
 123 void account_system_vtime(struct task_struct *tsk)
 124 {
 125         struct thread_info *ti = task_thread_info(tsk);
 126         unsigned long flags;
 127         cputime_t delta_stime;
 128         __u64 now;
 129
 130         local_irq_save(flags);
 131
 132         now = ia64_get_itc();
 133
 134         delta_stime = cycle_to_cputime(ti->ac_stime + (now - ti->ac_stamp));
 135         if (irq_count() || idle_task(smp_processor_id()) != tsk)
 136                 account_system_time(tsk, 0, delta_stime, delta_stime);
 137         else
 138                 account_idle_time(delta_stime);
 139         ti->ac_stime = 0;
 140
 141         ti->ac_stamp = now;
 142
 143         local_irq_restore(flags);
 144 }
 145 EXPORT_SYMBOL_GPL(account_system_vtime);
 146
 147 /*
 148  * Called from the timer interrupt handler to charge accumulated user time
 149  * to the current process.  Must be called with interrupts disabled.
 150  */
 151 void account_process_tick(struct task_struct *p, int user_tick)
 152 {
 153         struct thread_info *ti = task_thread_info(p);
 154         cputime_t delta_utime;
 155
 156         if (ti->ac_utime) {
 157                 delta_utime = cycle_to_cputime(ti->ac_utime);
 158                 account_user_time(p, delta_utime, delta_utime);
 159                 ti->ac_utime = 0;
 160         }
 161 }
 162
 163 #endif /* CONFIG_VIRT_CPU_ACCOUNTING */
 164
 165 static irqreturn_t
 166 timer_interrupt (int irq, void *dev_id)
 167 {
 168         unsigned long new_itm;
 169
 170         if (unlikely(cpu_is_offline(smp_processor_id()))) {
 171                 return IRQ_HANDLED;
 172         }
 173
 174         platform_timer_interrupt(irq, dev_id);
 175
 176         new_itm = local_cpu_data->itm_next;
 177
 178         if (!time_after(ia64_get_itc(), new_itm))
 179                 printk(KERN_ERR "Oops: timer tick before it's due (itc=%lx,itm=%lx)\n",
 180                        ia64_get_itc(), new_itm);
 181
 182         profile_tick(CPU_PROFILING);
 183
 184         if (paravirt_do_steal_accounting(&new_itm))
 185                 goto skip_process_time_accounting;
 186
 187         while (1) {
 188                 update_process_times(user_mode(get_irq_regs()));
 189
 190                 new_itm += local_cpu_data->itm_delta;
 191
 192                 if (smp_processor_id() == time_keeper_id) {
 193                         /*
 194                          * Here we are in the timer irq handler. We have irqs locally
 195                          * disabled, but we don't know if the timer_bh is running on
 196                          * another CPU. We need to avoid to SMP race by acquiring the
 197                          * xtime_lock.
 198                          */
 199                         write_seqlock(&xtime_lock);
 200                         do_timer(1);
 201                         local_cpu_data->itm_next = new_itm;
 202                         write_sequnlock(&xtime_lock);
 203                 } else
 204                         local_cpu_data->itm_next = new_itm;
 205
 206                 if (time_after(new_itm, ia64_get_itc()))
 207                         break;
 208
 209                 /*
 210                  * Allow IPIs to interrupt the timer loop.
 211                  */
 212                 local_irq_enable();
 213                 local_irq_disable();
 214         }
 215
 216 skip_process_time_accounting:
 217
 218         do {
 219                 /*
 220                  * If we're too close to the next clock tick for
 221                  * comfort, we increase the safety margin by
 222                  * intentionally dropping the next tick(s).  We do NOT
 223                  * update itm.next because that would force us to call
 224                  * do_timer() which in turn would let our clock run
 225                  * too fast (with the potentially devastating effect
 226                  * of losing monotony of time).
 227                  */
 228                 while (!time_after(new_itm, ia64_get_itc() + local_cpu_data->itm_delta/2))
 229                         new_itm += local_cpu_data->itm_delta;
 230                 ia64_set_itm(new_itm);
 231                 /* double check, in case we got hit by a (slow) PMI: */
 232         } while (time_after_eq(ia64_get_itc(), new_itm));
 233         return IRQ_HANDLED;
 234 }
 235
 236 /*
 237  * Encapsulate access to the itm structure for SMP.
 238  */
 239 void
 240 ia64_cpu_local_tick (void)
 241 {
 242         int cpu = smp_processor_id();
 243         unsigned long shift = 0, delta;
 244
 245         /* arrange for the cycle counter to generate a timer interrupt: */
 246         ia64_set_itv(IA64_TIMER_VECTOR);
 247
 248         delta = local_cpu_data->itm_delta;
 249         /*
 250          * Stagger the timer tick for each CPU so they don't occur all at (almost) the
 251          * same time:
 252          */
 253         if (cpu) {
 254                 unsigned long hi = 1UL << ia64_fls(cpu);
 255                 shift = (2*(cpu - hi) + 1) * delta/hi/2;
 256         }
 257         local_cpu_data->itm_next = ia64_get_itc() + delta + shift;
 258         ia64_set_itm(local_cpu_data->itm_next);
 259 }
 260
 261 static int nojitter;
 262
 263 static int __init nojitter_setup(char *str)
 264 {
 265         nojitter = 1;
 266         printk("Jitter checking for ITC timers disabled\n");
 267         return 1;
 268 }
 269
 270 __setup("nojitter", nojitter_setup);
 271
 272
 273 void __devinit
 274 ia64_init_itm (void)
 275 {
 276         unsigned long platform_base_freq, itc_freq;
 277         struct pal_freq_ratio itc_ratio, proc_ratio;
 278         long status, platform_base_drift, itc_drift;
 279
 280         /*
 281          * According to SAL v2.6, we need to use a SAL call to determine the platform base
 282          * frequency and then a PAL call to determine the frequency ratio between the ITC
 283          * and the base frequency.
 284          */
 285         status = ia64_sal_freq_base(SAL_FREQ_BASE_PLATFORM,
 286                                     &platform_base_freq, &platform_base_drift);
 287         if (status != 0) {
 288                 printk(KERN_ERR "SAL_FREQ_BASE_PLATFORM failed: %s\n", ia64_sal_strerror(status));
 289         } else {
 290                 status = ia64_pal_freq_ratios(&proc_ratio, NULL, &itc_ratio);
 291                 if (status != 0)
 292                         printk(KERN_ERR "PAL_FREQ_RATIOS failed with status=%ld\n", status);
 293         }
 294         if (status != 0) {
 295                 /* invent "random" values */
 296                 printk(KERN_ERR
 297                        "SAL/PAL failed to obtain frequency info---inventing reasonable values\n");
 298                 platform_base_freq = 100000000;
 299                 platform_base_drift = -1;       /* no drift info */
 300                 itc_ratio.num = 3;
 301                 itc_ratio.den = 1;
 302         }
 303         if (platform_base_freq < 40000000) {
 304                 printk(KERN_ERR "Platform base frequency %lu bogus---resetting to 75MHz!\n",
 305                        platform_base_freq);
 306                 platform_base_freq = 75000000;
 307                 platform_base_drift = -1;
 308         }
 309         if (!proc_ratio.den)
 310                 proc_ratio.den = 1;     /* avoid division by zero */
 311         if (!itc_ratio.den)
 312                 itc_ratio.den = 1;      /* avoid division by zero */
 313
 314         itc_freq = (platform_base_freq*itc_ratio.num)/itc_ratio.den;
 315
 316         local_cpu_data->itm_delta = (itc_freq + HZ/2) / HZ;
 317         printk(KERN_DEBUG "CPU %d: base freq=%lu.%03luMHz, ITC ratio=%u/%u, "
 318                "ITC freq=%lu.%03luMHz", smp_processor_id(),
 319                platform_base_freq / 1000000, (platform_base_freq / 1000) % 1000,
 320                itc_ratio.num, itc_ratio.den, itc_freq / 1000000, (itc_freq / 1000) % 1000);
 321
 322         if (platform_base_drift != -1) {
 323                 itc_drift = platform_base_drift*itc_ratio.num/itc_ratio.den;
 324                 printk("+/-%ldppm\n", itc_drift);
 325         } else {
 326                 itc_drift = -1;
 327                 printk("\n");
 328         }
 329
 330         local_cpu_data->proc_freq = (platform_base_freq*proc_ratio.num)/proc_ratio.den;
 331         local_cpu_data->itc_freq = itc_freq;
 332         local_cpu_data->cyc_per_usec = (itc_freq + USEC_PER_SEC/2) / USEC_PER_SEC;
 333         local_cpu_data->nsec_per_cyc = ((NSEC_PER_SEC<<IA64_NSEC_PER_CYC_SHIFT)
 334                                         + itc_freq/2)/itc_freq;
 335
 336         if (!(sal_platform_features & IA64_SAL_PLATFORM_FEATURE_ITC_DRIFT)) {
 337 #ifdef CONFIG_SMP
 338                 /* On IA64 in an SMP configuration ITCs are never accurately synchronized.
 339                  * Jitter compensation requires a cmpxchg which may limit
 340                  * the scalability of the syscalls for retrieving time.
 341                  * The ITC synchronization is usually successful to within a few
 342                  * ITC ticks but this is not a sure thing. If you need to improve
 343                  * timer performance in SMP situations then boot the kernel with the
 344                  * "nojitter" option. However, doing so may result in time fluctuating (maybe
 345                  * even going backward) if the ITC offsets between the individual CPUs
 346                  * are too large.
 347                  */
 348                 if (!nojitter)
 349                         itc_jitter_data.itc_jitter = 1;
 350 #endif
 351         } else
 352                 /*
 353                  * ITC is drifty and we have not synchronized the ITCs in smpboot.c.
 354                  * ITC values may fluctuate significantly between processors.
 355                  * Clock should not be used for hrtimers. Mark itc as only
 356                  * useful for boot and testing.
 357                  *
 358                  * Note that jitter compensation is off! There is no point of
 359                  * synchronizing ITCs since they may be large differentials
 360                  * that change over time.
 361                  *
 362                  * The only way to fix this would be to repeatedly sync the
 363                  * ITCs. Until that time we have to avoid ITC.
 364                  */
 365                 clocksource_itc.rating = 50;
 366
 367         paravirt_init_missing_ticks_accounting(smp_processor_id());
 368
 369         /* avoid softlock up message when cpu is unplug and plugged again. */
 370         touch_softlockup_watchdog();
 371
 372         /* Setup the CPU local timer tick */
 373         ia64_cpu_local_tick();
 374
 375         if (!itc_clocksource) {
 376                 /* Sort out mult/shift values: */
 377                 clocksource_itc.mult =
 378                         clocksource_hz2mult(local_cpu_data->itc_freq,
 379                                                 clocksource_itc.shift);
 380                 clocksource_register(&clocksource_itc);
 381                 itc_clocksource = &clocksource_itc;
 382         }
 383 }
 384
 385 static cycle_t itc_get_cycles(void)
 386 {
 387         u64 lcycle, now, ret;
 388
 389         if (!itc_jitter_data.itc_jitter)
 390                 return get_cycles();
 391
 392         lcycle = itc_jitter_data.itc_lastcycle;
 393         now = get_cycles();
 394         if (lcycle && time_after(lcycle, now))
 395                 return lcycle;
 396
 397         /*
 398          * Keep track of the last timer value returned.
 399          * In an SMP environment, you could lose out in contention of
 400          * cmpxchg. If so, your cmpxchg returns new value which the
 401          * winner of contention updated to. Use the new value instead.
 402          */
 403         ret = cmpxchg(&itc_jitter_data.itc_lastcycle, lcycle, now);
 404         if (unlikely(ret != lcycle))
 405                 return ret;
 406
 407         return now;
 408 }
 409
 410
 411 static struct irqaction timer_irqaction = {
 412         .handler =      timer_interrupt,
 413         .flags =        IRQF_DISABLED | IRQF_IRQPOLL,
 414         .name =         "timer"
 415 };
 416
 417 void __init
 418 time_init (void)
 419 {
 420         register_percpu_irq(IA64_TIMER_VECTOR, &timer_irqaction);
 421         efi_gettimeofday(&xtime);
 422         ia64_init_itm();
 423
 424         /*
 425          * Initialize wall_to_monotonic such that adding it to xtime will yield zero, the
 426          * tv_nsec field must be normalized (i.e., 0 <= nsec < NSEC_PER_SEC).
 427          */
 428         set_normalized_timespec(&wall_to_monotonic, -xtime.tv_sec, -xtime.tv_nsec);
 429 }
 430
 431 /*
 432  * Generic udelay assumes that if preemption is allowed and the thread
 433  * migrates to another CPU, that the ITC values are synchronized across
 434  * all CPUs.
 435  */
 436 static void
 437 ia64_itc_udelay (unsigned long usecs)
 438 {
 439         unsigned long start = ia64_get_itc();
 440         unsigned long end = start + usecs*local_cpu_data->cyc_per_usec;
 441
 442         while (time_before(ia64_get_itc(), end))
 443                 cpu_relax();
 444 }
 445
 446 void (*ia64_udelay)(unsigned long usecs) = &ia64_itc_udelay;
 447
 448 void
 449 udelay (unsigned long usecs)
 450 {
 451         (*ia64_udelay)(usecs);
 452 }
 453 EXPORT_SYMBOL(udelay);
 454
 455 /* IA64 doesn't cache the timezone */
 456 void update_vsyscall_tz(void)
 457 {
 458 }
 459
 460 void update_vsyscall(struct timespec *wall, struct clocksource *c)
 461 {
 462         unsigned long flags;
 463
 464         write_seqlock_irqsave(&fsyscall_gtod_data.lock, flags);
 465
 466         /* copy fsyscall clock data */
 467         fsyscall_gtod_data.clk_mask = c->mask;
 468         fsyscall_gtod_data.clk_mult = c->mult;
 469         fsyscall_gtod_data.clk_shift = c->shift;
 470         fsyscall_gtod_data.clk_fsys_mmio = c->fsys_mmio;
 471         fsyscall_gtod_data.clk_cycle_last = c->cycle_last;
 472
 473         /* copy kernel time structures */
 474         fsyscall_gtod_data.wall_time.tv_sec = wall->tv_sec;
 475         fsyscall_gtod_data.wall_time.tv_nsec = wall->tv_nsec;
 476         fsyscall_gtod_data.monotonic_time.tv_sec = wall_to_monotonic.tv_sec
 477                                                         + wall->tv_sec;
 478         fsyscall_gtod_data.monotonic_time.tv_nsec = wall_to_monotonic.tv_nsec
 479                                                         + wall->tv_nsec;
 480
 481         /* normalize */
 482         while (fsyscall_gtod_data.monotonic_time.tv_nsec >= NSEC_PER_SEC) {
 483                 fsyscall_gtod_data.monotonic_time.tv_nsec -= NSEC_PER_SEC;
 484                 fsyscall_gtod_data.monotonic_time.tv_sec++;
 485         }
 486
 487         write_sequnlock_irqrestore(&fsyscall_gtod_data.lock, flags);
 488 }
 489