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 #include <linux/config.h>
  12
  13 #include <linux/cpu.h>
  14 #include <linux/init.h>
  15 #include <linux/kernel.h>
  16 #include <linux/module.h>
  17 #include <linux/profile.h>
  18 #include <linux/sched.h>
  19 #include <linux/time.h>
  20 #include <linux/interrupt.h>
  21 #include <linux/efi.h>
  22 #include <linux/profile.h>
  23 #include <linux/timex.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 #include <asm/system.h>
  32
  33 extern unsigned long wall_jiffies;
  34
  35 u64 jiffies_64 = INITIAL_JIFFIES;
  36
  37 EXPORT_SYMBOL(jiffies_64);
  38
  39 #define TIME_KEEPER_ID  0       /* smp_processor_id() of time-keeper */
  40
  41 #ifdef CONFIG_IA64_DEBUG_IRQ
  42
  43 unsigned long last_cli_ip;
  44 EXPORT_SYMBOL(last_cli_ip);
  45
  46 #endif
  47
  48 unsigned long long
  49 sched_clock (void)
  50 {
  51         unsigned long offset = ia64_get_itc();
  52
  53         return (offset * local_cpu_data->nsec_per_cyc) >> IA64_NSEC_PER_CYC_SHIFT;
  54 }
  55
  56 static void
  57 itc_reset (void)
  58 {
  59 }
  60
  61 /*
  62  * Adjust for the fact that xtime has been advanced by delta_nsec (may be negative and/or
  63  * larger than NSEC_PER_SEC.
  64  */
  65 static void
  66 itc_update (long delta_nsec)
  67 {
  68 }
  69
  70 /*
  71  * Return the number of nano-seconds that elapsed since the last
  72  * update to jiffy.  It is quite possible that the timer interrupt
  73  * will interrupt this and result in a race for any of jiffies,
  74  * wall_jiffies or itm_next.  Thus, the xtime_lock must be at least
  75  * read synchronised when calling this routine (see do_gettimeofday()
  76  * below for an example).
  77  */
  78 unsigned long
  79 itc_get_offset (void)
  80 {
  81         unsigned long elapsed_cycles, lost = jiffies - wall_jiffies;
  82         unsigned long now = ia64_get_itc(), last_tick;
  83
  84         last_tick = (cpu_data(TIME_KEEPER_ID)->itm_next
  85                      - (lost + 1)*cpu_data(TIME_KEEPER_ID)->itm_delta);
  86
  87         elapsed_cycles = now - last_tick;
  88         return (elapsed_cycles*local_cpu_data->nsec_per_cyc) >> IA64_NSEC_PER_CYC_SHIFT;
  89 }
  90
  91 static struct time_interpolator itc_interpolator = {
  92         .get_offset =   itc_get_offset,
  93         .update =       itc_update,
  94         .reset =        itc_reset
  95 };
  96
  97 int
  98 do_settimeofday (struct timespec *tv)
  99 {
 100         time_t wtm_sec, sec = tv->tv_sec;
 101         long wtm_nsec, nsec = tv->tv_nsec;
 102
 103         if ((unsigned long)tv->tv_nsec >= NSEC_PER_SEC)
 104                 return -EINVAL;
 105
 106         write_seqlock_irq(&xtime_lock);
 107         {
 108                 /*
 109                  * This is revolting. We need to set "xtime" correctly. However, the value
 110                  * in this location is the value at the most recent update of wall time.
 111                  * Discover what correction gettimeofday would have done, and then undo
 112                  * it!
 113                  */
 114                 nsec -= time_interpolator_get_offset();
 115
 116                 wtm_sec  = wall_to_monotonic.tv_sec + (xtime.tv_sec - sec);
 117                 wtm_nsec = wall_to_monotonic.tv_nsec + (xtime.tv_nsec - nsec);
 118
 119                 set_normalized_timespec(&xtime, sec, nsec);
 120                 set_normalized_timespec(&wall_to_monotonic, wtm_sec, wtm_nsec);
 121
 122                 time_adjust = 0;                /* stop active adjtime() */
 123                 time_status |= STA_UNSYNC;
 124                 time_maxerror = NTP_PHASE_LIMIT;
 125                 time_esterror = NTP_PHASE_LIMIT;
 126                 time_interpolator_reset();
 127         }
 128         write_sequnlock_irq(&xtime_lock);
 129         clock_was_set();
 130         return 0;
 131 }
 132
 133 EXPORT_SYMBOL(do_settimeofday);
 134
 135 void
 136 do_gettimeofday (struct timeval *tv)
 137 {
 138         unsigned long seq, nsec, usec, sec, old, offset;
 139
 140         while (1) {
 141                 seq = read_seqbegin(&xtime_lock);
 142                 {
 143                         old = last_nsec_offset;
 144                         offset = time_interpolator_get_offset();
 145                         sec = xtime.tv_sec;
 146                         nsec = xtime.tv_nsec;
 147                 }
 148                 if (unlikely(read_seqretry(&xtime_lock, seq)))
 149                         continue;
 150                 /*
 151                  * Ensure that for any pair of causally ordered gettimeofday() calls, time
 152                  * never goes backwards (even when ITC on different CPUs are not perfectly
 153                  * synchronized).  (A pair of concurrent calls to gettimeofday() is by
 154                  * definition non-causal and hence it makes no sense to talk about
 155                  * time-continuity for such calls.)
 156                  *
 157                  * Doing this in a lock-free and race-free manner is tricky.  Here is why
 158                  * it works (most of the time): read_seqretry() just succeeded, which
 159                  * implies we calculated a consistent (valid) value for "offset".  If the
 160                  * cmpxchg() below succeeds, we further know that last_nsec_offset still
 161                  * has the same value as at the beginning of the loop, so there was
 162                  * presumably no timer-tick or other updates to last_nsec_offset in the
 163                  * meantime.  This isn't 100% true though: there _is_ a possibility of a
 164                  * timer-tick occurring right right after read_seqretry() and then getting
 165                  * zero or more other readers which will set last_nsec_offset to the same
 166                  * value as the one we read at the beginning of the loop.  If this
 167                  * happens, we'll end up returning a slightly newer time than we ought to
 168                  * (the jump forward is at most "offset" nano-seconds).  There is no
 169                  * danger of causing time to go backwards, though, so we are safe in that
 170                  * sense.  We could make the probability of this unlucky case occurring
 171                  * arbitrarily small by encoding a version number in last_nsec_offset, but
 172                  * even without versioning, the probability of this unlucky case should be
 173                  * so small that we won't worry about it.
 174                  */
 175                 if (offset <= old) {
 176                         offset = old;
 177                         break;
 178                 } else if (likely(cmpxchg(&last_nsec_offset, old, offset) == old))
 179                         break;
 180
 181                 /* someone else beat us to updating last_nsec_offset; try again */
 182         }
 183
 184         usec = (nsec + offset) / 1000;
 185
 186         while (unlikely(usec >= USEC_PER_SEC)) {
 187                 usec -= USEC_PER_SEC;
 188                 ++sec;
 189         }
 190
 191         tv->tv_sec = sec;
 192         tv->tv_usec = usec;
 193 }
 194
 195 EXPORT_SYMBOL(do_gettimeofday);
 196
 197 /*
 198  * The profiling function is SMP safe. (nothing can mess
 199  * around with "current", and the profiling counters are
 200  * updated with atomic operations). This is especially
 201  * useful with a profiling multiplier != 1
 202  */
 203 static inline void
 204 ia64_do_profile (struct pt_regs * regs)
 205 {
 206         unsigned long ip, slot;
 207         extern cpumask_t prof_cpu_mask;
 208
 209         profile_hook(regs);
 210
 211         if (user_mode(regs))
 212                 return;
 213
 214         if (!prof_buffer)
 215                 return;
 216
 217         ip = instruction_pointer(regs);
 218         /* Conserve space in histogram by encoding slot bits in address
 219          * bits 2 and 3 rather than bits 0 and 1.
 220          */
 221         slot = ip & 3;
 222         ip = (ip & ~3UL) + 4*slot;
 223
 224         /*
 225          * Only measure the CPUs specified by /proc/irq/prof_cpu_mask.
 226          * (default is all CPUs.)
 227          */
 228         if (!cpu_isset(smp_processor_id(), prof_cpu_mask))
 229                 return;
 230
 231         ip -= (unsigned long) &_stext;
 232         ip >>= prof_shift;
 233         /*
 234          * Don't ignore out-of-bounds IP values silently,
 235          * put them into the last histogram slot, so if
 236          * present, they will show up as a sharp peak.
 237          */
 238         if (ip > prof_len-1)
 239                 ip = prof_len-1;
 240         atomic_inc((atomic_t *)&prof_buffer[ip]);
 241 }
 242
 243 static irqreturn_t
 244 timer_interrupt (int irq, void *dev_id, struct pt_regs *regs)
 245 {
 246         unsigned long new_itm;
 247
 248         if (unlikely(cpu_is_offline(smp_processor_id()))) {
 249                 return IRQ_HANDLED;
 250         }
 251
 252         platform_timer_interrupt(irq, dev_id, regs);
 253
 254         new_itm = local_cpu_data->itm_next;
 255
 256         if (!time_after(ia64_get_itc(), new_itm))
 257                 printk(KERN_ERR "Oops: timer tick before it's due (itc=%lx,itm=%lx)\n",
 258                        ia64_get_itc(), new_itm);
 259
 260         ia64_do_profile(regs);
 261
 262         while (1) {
 263 #ifdef CONFIG_SMP
 264                 /*
 265                  * For UP, this is done in do_timer().  Weird, but
 266                  * fixing that would require updates to all
 267                  * platforms.
 268                  */
 269                 update_process_times(user_mode(regs));
 270 #endif
 271                 new_itm += local_cpu_data->itm_delta;
 272
 273                 if (smp_processor_id() == TIME_KEEPER_ID) {
 274                         /*
 275                          * Here we are in the timer irq handler. We have irqs locally
 276                          * disabled, but we don't know if the timer_bh is running on
 277                          * another CPU. We need to avoid to SMP race by acquiring the
 278                          * xtime_lock.
 279                          */
 280                         write_seqlock(&xtime_lock);
 281                         do_timer(regs);
 282                         local_cpu_data->itm_next = new_itm;
 283                         write_sequnlock(&xtime_lock);
 284                 } else
 285                         local_cpu_data->itm_next = new_itm;
 286
 287                 if (time_after(new_itm, ia64_get_itc()))
 288                         break;
 289         }
 290
 291         do {
 292                 /*
 293                  * If we're too close to the next clock tick for
 294                  * comfort, we increase the safety margin by
 295                  * intentionally dropping the next tick(s).  We do NOT
 296                  * update itm.next because that would force us to call
 297                  * do_timer() which in turn would let our clock run
 298                  * too fast (with the potentially devastating effect
 299                  * of losing monotony of time).
 300                  */
 301                 while (!time_after(new_itm, ia64_get_itc() + local_cpu_data->itm_delta/2))
 302                         new_itm += local_cpu_data->itm_delta;
 303                 ia64_set_itm(new_itm);
 304                 /* double check, in case we got hit by a (slow) PMI: */
 305         } while (time_after_eq(ia64_get_itc(), new_itm));
 306         return IRQ_HANDLED;
 307 }
 308
 309 /*
 310  * Encapsulate access to the itm structure for SMP.
 311  */
 312 void
 313 ia64_cpu_local_tick (void)
 314 {
 315         int cpu = smp_processor_id();
 316         unsigned long shift = 0, delta;
 317
 318         /* arrange for the cycle counter to generate a timer interrupt: */
 319         ia64_set_itv(IA64_TIMER_VECTOR);
 320
 321         delta = local_cpu_data->itm_delta;
 322         /*
 323          * Stagger the timer tick for each CPU so they don't occur all at (almost) the
 324          * same time:
 325          */
 326         if (cpu) {
 327                 unsigned long hi = 1UL << ia64_fls(cpu);
 328                 shift = (2*(cpu - hi) + 1) * delta/hi/2;
 329         }
 330         local_cpu_data->itm_next = ia64_get_itc() + delta + shift;
 331         ia64_set_itm(local_cpu_data->itm_next);
 332 }
 333
 334 void __devinit
 335 ia64_init_itm (void)
 336 {
 337         unsigned long platform_base_freq, itc_freq;
 338         struct pal_freq_ratio itc_ratio, proc_ratio;
 339         long status, platform_base_drift, itc_drift;
 340
 341         /*
 342          * According to SAL v2.6, we need to use a SAL call to determine the platform base
 343          * frequency and then a PAL call to determine the frequency ratio between the ITC
 344          * and the base frequency.
 345          */
 346         status = ia64_sal_freq_base(SAL_FREQ_BASE_PLATFORM,
 347                                     &platform_base_freq, &platform_base_drift);
 348         if (status != 0) {
 349                 printk(KERN_ERR "SAL_FREQ_BASE_PLATFORM failed: %s\n", ia64_sal_strerror(status));
 350         } else {
 351                 status = ia64_pal_freq_ratios(&proc_ratio, 0, &itc_ratio);
 352                 if (status != 0)
 353                         printk(KERN_ERR "PAL_FREQ_RATIOS failed with status=%ld\n", status);
 354         }
 355         if (status != 0) {
 356                 /* invent "random" values */
 357                 printk(KERN_ERR
 358                        "SAL/PAL failed to obtain frequency info---inventing reasonable values\n");
 359                 platform_base_freq = 100000000;
 360                 platform_base_drift = -1;       /* no drift info */
 361                 itc_ratio.num = 3;
 362                 itc_ratio.den = 1;
 363         }
 364         if (platform_base_freq < 40000000) {
 365                 printk(KERN_ERR "Platform base frequency %lu bogus---resetting to 75MHz!\n",
 366                        platform_base_freq);
 367                 platform_base_freq = 75000000;
 368                 platform_base_drift = -1;
 369         }
 370         if (!proc_ratio.den)
 371                 proc_ratio.den = 1;     /* avoid division by zero */
 372         if (!itc_ratio.den)
 373                 itc_ratio.den = 1;      /* avoid division by zero */
 374
 375         itc_freq = (platform_base_freq*itc_ratio.num)/itc_ratio.den;
 376         if (platform_base_drift != -1)
 377                 itc_drift = platform_base_drift*itc_ratio.num/itc_ratio.den;
 378         else
 379                 itc_drift = -1;
 380
 381         local_cpu_data->itm_delta = (itc_freq + HZ/2) / HZ;
 382         printk(KERN_INFO "CPU %d: base freq=%lu.%03luMHz, ITC ratio=%lu/%lu, "
 383                "ITC freq=%lu.%03luMHz+/-%ldppm\n", smp_processor_id(),
 384                platform_base_freq / 1000000, (platform_base_freq / 1000) % 1000,
 385                itc_ratio.num, itc_ratio.den, itc_freq / 1000000, (itc_freq / 1000) % 1000,
 386                itc_drift);
 387
 388         local_cpu_data->proc_freq = (platform_base_freq*proc_ratio.num)/proc_ratio.den;
 389         local_cpu_data->itc_freq = itc_freq;
 390         local_cpu_data->cyc_per_usec = (itc_freq + USEC_PER_SEC/2) / USEC_PER_SEC;
 391         local_cpu_data->nsec_per_cyc = ((NSEC_PER_SEC<<IA64_NSEC_PER_CYC_SHIFT)
 392                                         + itc_freq/2)/itc_freq;
 393
 394         if (!(sal_platform_features & IA64_SAL_PLATFORM_FEATURE_ITC_DRIFT)) {
 395                 itc_interpolator.frequency = local_cpu_data->itc_freq;
 396                 itc_interpolator.drift = itc_drift;
 397                 register_time_interpolator(&itc_interpolator);
 398         }
 399
 400         /* Setup the CPU local timer tick */
 401         ia64_cpu_local_tick();
 402 }
 403
 404 static struct irqaction timer_irqaction = {
 405         .handler =      timer_interrupt,
 406         .flags =        SA_INTERRUPT,
 407         .name =         "timer"
 408 };
 409
 410 void __init
 411 time_init (void)
 412 {
 413         register_percpu_irq(IA64_TIMER_VECTOR, &timer_irqaction);
 414         efi_gettimeofday(&xtime);
 415         ia64_init_itm();
 416
 417         /*
 418          * Initialize wall_to_monotonic such that adding it to xtime will yield zero, the
 419          * tv_nsec field must be normalized (i.e., 0 <= nsec < NSEC_PER_SEC).
 420          */
 421         set_normalized_timespec(&wall_to_monotonic, -xtime.tv_sec, -xtime.tv_nsec);
 422 }