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 #include <linux/cputime.h>
  25
  26 #include <asm/machvec.h>
  27 #include <asm/delay.h>
  28 #include <asm/hw_irq.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 u64 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 static struct clocksource clocksource_itc = {
  51         .name           = "itc",
  52         .rating         = 350,
  53         .read           = itc_get_cycles,
  54         .mask           = CLOCKSOURCE_MASK(64),
  55         .flags          = CLOCK_SOURCE_IS_CONTINUOUS,
  56 };
  57 static struct clocksource *itc_clocksource;
  58
  59 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
  60
  61 #include <linux/kernel_stat.h>
  62
  63 extern u64 cycle_to_nsec(u64 cyc);
  64
  65 void vtime_flush(struct task_struct *tsk)
  66 {
  67         struct thread_info *ti = task_thread_info(tsk);
  68         u64 delta;
  69
  70         if (ti->utime)
  71                 account_user_time(tsk, cycle_to_nsec(ti->utime));
  72
  73         if (ti->gtime)
  74                 account_guest_time(tsk, cycle_to_nsec(ti->gtime));
  75
  76         if (ti->idle_time)
  77                 account_idle_time(cycle_to_nsec(ti->idle_time));
  78
  79         if (ti->stime) {
  80                 delta = cycle_to_nsec(ti->stime);
  81                 account_system_index_time(tsk, delta, CPUTIME_SYSTEM);
  82         }
  83
  84         if (ti->hardirq_time) {
  85                 delta = cycle_to_nsec(ti->hardirq_time);
  86                 account_system_index_time(tsk, delta, CPUTIME_IRQ);
  87         }
  88
  89         if (ti->softirq_time) {
  90                 delta = cycle_to_nsec(ti->softirq_time));
  91                 account_system_index_time(tsk, delta, CPUTIME_SOFTIRQ);
  92         }
  93
  94         ti->utime = 0;
  95         ti->gtime = 0;
  96         ti->idle_time = 0;
  97         ti->stime = 0;
  98         ti->hardirq_time = 0;
  99         ti->softirq_time = 0;
 100 }
 101
 102 /*
 103  * Called from the context switch with interrupts disabled, to charge all
 104  * accumulated times to the current process, and to prepare accounting on
 105  * the next process.
 106  */
 107 void arch_vtime_task_switch(struct task_struct *prev)
 108 {
 109         struct thread_info *pi = task_thread_info(prev);
 110         struct thread_info *ni = task_thread_info(current);
 111
 112         ni->ac_stamp = pi->ac_stamp;
 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 static __u64 vtime_delta(struct task_struct *tsk)
 121 {
 122         struct thread_info *ti = task_thread_info(tsk);
 123         __u64 now, delta_stime;
 124
 125         WARN_ON_ONCE(!irqs_disabled());
 126
 127         now = ia64_get_itc();
 128         delta_stime = now - ti->ac_stamp;
 129         ti->ac_stamp = now;
 130
 131         return delta_stime;
 132 }
 133
 134 void vtime_account_system(struct task_struct *tsk)
 135 {
 136         struct thread_info *ti = task_thread_info(tsk);
 137         __u64 stime = vtime_delta(tsk);
 138
 139         if ((tsk->flags & PF_VCPU) && !irq_count())
 140                 ti->gtime += stime;
 141         else if (hardirq_count())
 142                 ti->hardirq_time += stime;
 143         else if (in_serving_softirq())
 144                 ti->softirq_time += stime;
 145         else
 146                 ti->stime += stime;
 147 }
 148 EXPORT_SYMBOL_GPL(vtime_account_system);
 149
 150 void vtime_account_idle(struct task_struct *tsk)
 151 {
 152         struct thread_info *ti = task_thread_info(tsk);
 153
 154         ti->idle_time += vtime_delta(tsk);
 155 }
 156
 157 #endif /* CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
 158
 159 static irqreturn_t
 160 timer_interrupt (int irq, void *dev_id)
 161 {
 162         unsigned long new_itm;
 163
 164         if (cpu_is_offline(smp_processor_id())) {
 165                 return IRQ_HANDLED;
 166         }
 167
 168         platform_timer_interrupt(irq, dev_id);
 169
 170         new_itm = local_cpu_data->itm_next;
 171
 172         if (!time_after(ia64_get_itc(), new_itm))
 173                 printk(KERN_ERR "Oops: timer tick before it's due (itc=%lx,itm=%lx)\n",
 174                        ia64_get_itc(), new_itm);
 175
 176         profile_tick(CPU_PROFILING);
 177
 178         while (1) {
 179                 update_process_times(user_mode(get_irq_regs()));
 180
 181                 new_itm += local_cpu_data->itm_delta;
 182
 183                 if (smp_processor_id() == time_keeper_id)
 184                         xtime_update(1);
 185
 186                 local_cpu_data->itm_next = new_itm;
 187
 188                 if (time_after(new_itm, ia64_get_itc()))
 189                         break;
 190
 191                 /*
 192                  * Allow IPIs to interrupt the timer loop.
 193                  */
 194                 local_irq_enable();
 195                 local_irq_disable();
 196         }
 197
 198         do {
 199                 /*
 200                  * If we're too close to the next clock tick for
 201                  * comfort, we increase the safety margin by
 202                  * intentionally dropping the next tick(s).  We do NOT
 203                  * update itm.next because that would force us to call
 204                  * xtime_update() which in turn would let our clock run
 205                  * too fast (with the potentially devastating effect
 206                  * of losing monotony of time).
 207                  */
 208                 while (!time_after(new_itm, ia64_get_itc() + local_cpu_data->itm_delta/2))
 209                         new_itm += local_cpu_data->itm_delta;
 210                 ia64_set_itm(new_itm);
 211                 /* double check, in case we got hit by a (slow) PMI: */
 212         } while (time_after_eq(ia64_get_itc(), new_itm));
 213         return IRQ_HANDLED;
 214 }
 215
 216 /*
 217  * Encapsulate access to the itm structure for SMP.
 218  */
 219 void
 220 ia64_cpu_local_tick (void)
 221 {
 222         int cpu = smp_processor_id();
 223         unsigned long shift = 0, delta;
 224
 225         /* arrange for the cycle counter to generate a timer interrupt: */
 226         ia64_set_itv(IA64_TIMER_VECTOR);
 227
 228         delta = local_cpu_data->itm_delta;
 229         /*
 230          * Stagger the timer tick for each CPU so they don't occur all at (almost) the
 231          * same time:
 232          */
 233         if (cpu) {
 234                 unsigned long hi = 1UL << ia64_fls(cpu);
 235                 shift = (2*(cpu - hi) + 1) * delta/hi/2;
 236         }
 237         local_cpu_data->itm_next = ia64_get_itc() + delta + shift;
 238         ia64_set_itm(local_cpu_data->itm_next);
 239 }
 240
 241 static int nojitter;
 242
 243 static int __init nojitter_setup(char *str)
 244 {
 245         nojitter = 1;
 246         printk("Jitter checking for ITC timers disabled\n");
 247         return 1;
 248 }
 249
 250 __setup("nojitter", nojitter_setup);
 251
 252
 253 void ia64_init_itm(void)
 254 {
 255         unsigned long platform_base_freq, itc_freq;
 256         struct pal_freq_ratio itc_ratio, proc_ratio;
 257         long status, platform_base_drift, itc_drift;
 258
 259         /*
 260          * According to SAL v2.6, we need to use a SAL call to determine the platform base
 261          * frequency and then a PAL call to determine the frequency ratio between the ITC
 262          * and the base frequency.
 263          */
 264         status = ia64_sal_freq_base(SAL_FREQ_BASE_PLATFORM,
 265                                     &platform_base_freq, &platform_base_drift);
 266         if (status != 0) {
 267                 printk(KERN_ERR "SAL_FREQ_BASE_PLATFORM failed: %s\n", ia64_sal_strerror(status));
 268         } else {
 269                 status = ia64_pal_freq_ratios(&proc_ratio, NULL, &itc_ratio);
 270                 if (status != 0)
 271                         printk(KERN_ERR "PAL_FREQ_RATIOS failed with status=%ld\n", status);
 272         }
 273         if (status != 0) {
 274                 /* invent "random" values */
 275                 printk(KERN_ERR
 276                        "SAL/PAL failed to obtain frequency info---inventing reasonable values\n");
 277                 platform_base_freq = 100000000;
 278                 platform_base_drift = -1;       /* no drift info */
 279                 itc_ratio.num = 3;
 280                 itc_ratio.den = 1;
 281         }
 282         if (platform_base_freq < 40000000) {
 283                 printk(KERN_ERR "Platform base frequency %lu bogus---resetting to 75MHz!\n",
 284                        platform_base_freq);
 285                 platform_base_freq = 75000000;
 286                 platform_base_drift = -1;
 287         }
 288         if (!proc_ratio.den)
 289                 proc_ratio.den = 1;     /* avoid division by zero */
 290         if (!itc_ratio.den)
 291                 itc_ratio.den = 1;      /* avoid division by zero */
 292
 293         itc_freq = (platform_base_freq*itc_ratio.num)/itc_ratio.den;
 294
 295         local_cpu_data->itm_delta = (itc_freq + HZ/2) / HZ;
 296         printk(KERN_DEBUG "CPU %d: base freq=%lu.%03luMHz, ITC ratio=%u/%u, "
 297                "ITC freq=%lu.%03luMHz", smp_processor_id(),
 298                platform_base_freq / 1000000, (platform_base_freq / 1000) % 1000,
 299                itc_ratio.num, itc_ratio.den, itc_freq / 1000000, (itc_freq / 1000) % 1000);
 300
 301         if (platform_base_drift != -1) {
 302                 itc_drift = platform_base_drift*itc_ratio.num/itc_ratio.den;
 303                 printk("+/-%ldppm\n", itc_drift);
 304         } else {
 305                 itc_drift = -1;
 306                 printk("\n");
 307         }
 308
 309         local_cpu_data->proc_freq = (platform_base_freq*proc_ratio.num)/proc_ratio.den;
 310         local_cpu_data->itc_freq = itc_freq;
 311         local_cpu_data->cyc_per_usec = (itc_freq + USEC_PER_SEC/2) / USEC_PER_SEC;
 312         local_cpu_data->nsec_per_cyc = ((NSEC_PER_SEC<<IA64_NSEC_PER_CYC_SHIFT)
 313                                         + itc_freq/2)/itc_freq;
 314
 315         if (!(sal_platform_features & IA64_SAL_PLATFORM_FEATURE_ITC_DRIFT)) {
 316 #ifdef CONFIG_SMP
 317                 /* On IA64 in an SMP configuration ITCs are never accurately synchronized.
 318                  * Jitter compensation requires a cmpxchg which may limit
 319                  * the scalability of the syscalls for retrieving time.
 320                  * The ITC synchronization is usually successful to within a few
 321                  * ITC ticks but this is not a sure thing. If you need to improve
 322                  * timer performance in SMP situations then boot the kernel with the
 323                  * "nojitter" option. However, doing so may result in time fluctuating (maybe
 324                  * even going backward) if the ITC offsets between the individual CPUs
 325                  * are too large.
 326                  */
 327                 if (!nojitter)
 328                         itc_jitter_data.itc_jitter = 1;
 329 #endif
 330         } else
 331                 /*
 332                  * ITC is drifty and we have not synchronized the ITCs in smpboot.c.
 333                  * ITC values may fluctuate significantly between processors.
 334                  * Clock should not be used for hrtimers. Mark itc as only
 335                  * useful for boot and testing.
 336                  *
 337                  * Note that jitter compensation is off! There is no point of
 338                  * synchronizing ITCs since they may be large differentials
 339                  * that change over time.
 340                  *
 341                  * The only way to fix this would be to repeatedly sync the
 342                  * ITCs. Until that time we have to avoid ITC.
 343                  */
 344                 clocksource_itc.rating = 50;
 345
 346         /* avoid softlock up message when cpu is unplug and plugged again. */
 347         touch_softlockup_watchdog();
 348
 349         /* Setup the CPU local timer tick */
 350         ia64_cpu_local_tick();
 351
 352         if (!itc_clocksource) {
 353                 clocksource_register_hz(&clocksource_itc,
 354                                                 local_cpu_data->itc_freq);
 355                 itc_clocksource = &clocksource_itc;
 356         }
 357 }
 358
 359 static u64 itc_get_cycles(struct clocksource *cs)
 360 {
 361         unsigned long lcycle, now, ret;
 362
 363         if (!itc_jitter_data.itc_jitter)
 364                 return get_cycles();
 365
 366         lcycle = itc_jitter_data.itc_lastcycle;
 367         now = get_cycles();
 368         if (lcycle && time_after(lcycle, now))
 369                 return lcycle;
 370
 371         /*
 372          * Keep track of the last timer value returned.
 373          * In an SMP environment, you could lose out in contention of
 374          * cmpxchg. If so, your cmpxchg returns new value which the
 375          * winner of contention updated to. Use the new value instead.
 376          */
 377         ret = cmpxchg(&itc_jitter_data.itc_lastcycle, lcycle, now);
 378         if (unlikely(ret != lcycle))
 379                 return ret;
 380
 381         return now;
 382 }
 383
 384
 385 static struct irqaction timer_irqaction = {
 386         .handler =      timer_interrupt,
 387         .flags =        IRQF_IRQPOLL,
 388         .name =         "timer"
 389 };
 390
 391 void read_persistent_clock64(struct timespec64 *ts)
 392 {
 393         efi_gettimeofday(ts);
 394 }
 395
 396 void __init
 397 time_init (void)
 398 {
 399         register_percpu_irq(IA64_TIMER_VECTOR, &timer_irqaction);
 400         ia64_init_itm();
 401 }
 402
 403 /*
 404  * Generic udelay assumes that if preemption is allowed and the thread
 405  * migrates to another CPU, that the ITC values are synchronized across
 406  * all CPUs.
 407  */
 408 static void
 409 ia64_itc_udelay (unsigned long usecs)
 410 {
 411         unsigned long start = ia64_get_itc();
 412         unsigned long end = start + usecs*local_cpu_data->cyc_per_usec;
 413
 414         while (time_before(ia64_get_itc(), end))
 415                 cpu_relax();
 416 }
 417
 418 void (*ia64_udelay)(unsigned long usecs) = &ia64_itc_udelay;
 419
 420 void
 421 udelay (unsigned long usecs)
 422 {
 423         (*ia64_udelay)(usecs);
 424 }
 425 EXPORT_SYMBOL(udelay);
 426
 427 /* IA64 doesn't cache the timezone */
 428 void update_vsyscall_tz(void)
 429 {
 430 }
 431
 432 void update_vsyscall_old(struct timespec *wall, struct timespec *wtm,
 433                          struct clocksource *c, u32 mult, u64 cycle_last)
 434 {
 435         write_seqcount_begin(&fsyscall_gtod_data.seq);
 436
 437         /* copy fsyscall clock data */
 438         fsyscall_gtod_data.clk_mask = c->mask;
 439         fsyscall_gtod_data.clk_mult = mult;
 440         fsyscall_gtod_data.clk_shift = c->shift;
 441         fsyscall_gtod_data.clk_fsys_mmio = c->archdata.fsys_mmio;
 442         fsyscall_gtod_data.clk_cycle_last = cycle_last;
 443
 444         /* copy kernel time structures */
 445         fsyscall_gtod_data.wall_time.tv_sec = wall->tv_sec;
 446         fsyscall_gtod_data.wall_time.tv_nsec = wall->tv_nsec;
 447         fsyscall_gtod_data.monotonic_time.tv_sec = wtm->tv_sec
 448                                                         + wall->tv_sec;
 449         fsyscall_gtod_data.monotonic_time.tv_nsec = wtm->tv_nsec
 450                                                         + wall->tv_nsec;
 451
 452         /* normalize */
 453         while (fsyscall_gtod_data.monotonic_time.tv_nsec >= NSEC_PER_SEC) {
 454                 fsyscall_gtod_data.monotonic_time.tv_nsec -= NSEC_PER_SEC;
 455                 fsyscall_gtod_data.monotonic_time.tv_sec++;
 456         }
 457
 458         write_seqcount_end(&fsyscall_gtod_data.seq);
 459 }
 460