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