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/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/profile.h>
  22 #include <linux/timex.h>
  23
  24 #include <asm/machvec.h>
  25 #include <asm/delay.h>
  26 #include <asm/hw_irq.h>
  27 #include <asm/ptrace.h>
  28 #include <asm/sal.h>
  29 #include <asm/sections.h>
  30 #include <asm/system.h>
  31
  32 extern unsigned long wall_jiffies;
  33
  34 u64 jiffies_64 = INITIAL_JIFFIES;
  35
  36 EXPORT_SYMBOL(jiffies_64);
  37
  38 #define TIME_KEEPER_ID  0       /* smp_processor_id() of time-keeper */
  39
  40 #ifdef CONFIG_IA64_DEBUG_IRQ
  41
  42 unsigned long last_cli_ip;
  43 EXPORT_SYMBOL(last_cli_ip);
  44
  45 #endif
  46
  47 unsigned long long
  48 sched_clock (void)
  49 {
  50         unsigned long offset = ia64_get_itc();
  51
  52         return (offset * local_cpu_data->nsec_per_cyc) >> IA64_NSEC_PER_CYC_SHIFT;
  53 }
  54
  55 static void
  56 itc_reset (void)
  57 {
  58 }
  59
  60 /*
  61  * Adjust for the fact that xtime has been advanced by delta_nsec (may be negative and/or
  62  * larger than NSEC_PER_SEC.
  63  */
  64 static void
  65 itc_update (long delta_nsec)
  66 {
  67 }
  68
  69 /*
  70  * Return the number of nano-seconds that elapsed since the last
  71  * update to jiffy.  It is quite possible that the timer interrupt
  72  * will interrupt this and result in a race for any of jiffies,
  73  * wall_jiffies or itm_next.  Thus, the xtime_lock must be at least
  74  * read synchronised when calling this routine (see do_gettimeofday()
  75  * below for an example).
  76  */
  77 unsigned long
  78 itc_get_offset (void)
  79 {
  80         unsigned long elapsed_cycles, lost = jiffies - wall_jiffies;
  81         unsigned long now = ia64_get_itc(), last_tick;
  82
  83         last_tick = (cpu_data(TIME_KEEPER_ID)->itm_next
  84                      - (lost + 1)*cpu_data(TIME_KEEPER_ID)->itm_delta);
  85
  86         elapsed_cycles = now - last_tick;
  87         return (elapsed_cycles*local_cpu_data->nsec_per_cyc) >> IA64_NSEC_PER_CYC_SHIFT;
  88 }
  89
  90 static struct time_interpolator itc_interpolator = {
  91         .get_offset =   itc_get_offset,
  92         .update =       itc_update,
  93         .reset =        itc_reset
  94 };
  95
  96 int
  97 do_settimeofday (struct timespec *tv)
  98 {
  99         time_t wtm_sec, sec = tv->tv_sec;
 100         long wtm_nsec, nsec = tv->tv_nsec;
 101
 102         if ((unsigned long)tv->tv_nsec >= NSEC_PER_SEC)
 103                 return -EINVAL;
 104
 105         write_seqlock_irq(&xtime_lock);
 106         {
 107                 /*
 108                  * This is revolting. We need to set "xtime" correctly. However, the value
 109                  * in this location is the value at the most recent update of wall time.
 110                  * Discover what correction gettimeofday would have done, and then undo
 111                  * it!
 112                  */
 113                 nsec -= time_interpolator_get_offset();
 114
 115                 wtm_sec  = wall_to_monotonic.tv_sec + (xtime.tv_sec - sec);
 116                 wtm_nsec = wall_to_monotonic.tv_nsec + (xtime.tv_nsec - nsec);
 117
 118                 set_normalized_timespec(&xtime, sec, nsec);
 119                 set_normalized_timespec(&wall_to_monotonic, wtm_sec, wtm_nsec);
 120
 121                 time_adjust = 0;                /* stop active adjtime() */
 122                 time_status |= STA_UNSYNC;
 123                 time_maxerror = NTP_PHASE_LIMIT;
 124                 time_esterror = NTP_PHASE_LIMIT;
 125                 time_interpolator_reset();
 126         }
 127         write_sequnlock_irq(&xtime_lock);
 128         clock_was_set();
 129         return 0;
 130 }
 131
 132 EXPORT_SYMBOL(do_settimeofday);
 133
 134 void
 135 do_gettimeofday (struct timeval *tv)
 136 {
 137         unsigned long seq, nsec, usec, sec, old, offset;
 138
 139         while (1) {
 140                 seq = read_seqbegin(&xtime_lock);
 141                 {
 142                         old = last_nsec_offset;
 143                         offset = time_interpolator_get_offset();
 144                         sec = xtime.tv_sec;
 145                         nsec = xtime.tv_nsec;
 146                 }
 147                 if (unlikely(read_seqretry(&xtime_lock, seq)))
 148                         continue;
 149                 /*
 150                  * Ensure that for any pair of causally ordered gettimeofday() calls, time
 151                  * never goes backwards (even when ITC on different CPUs are not perfectly
 152                  * synchronized).  (A pair of concurrent calls to gettimeofday() is by
 153                  * definition non-causal and hence it makes no sense to talk about
 154                  * time-continuity for such calls.)
 155                  *
 156                  * Doing this in a lock-free and race-free manner is tricky.  Here is why
 157                  * it works (most of the time): read_seqretry() just succeeded, which
 158                  * implies we calculated a consistent (valid) value for "offset".  If the
 159                  * cmpxchg() below succeeds, we further know that last_nsec_offset still
 160                  * has the same value as at the beginning of the loop, so there was
 161                  * presumably no timer-tick or other updates to last_nsec_offset in the
 162                  * meantime.  This isn't 100% true though: there _is_ a possibility of a
 163                  * timer-tick occurring right right after read_seqretry() and then getting
 164                  * zero or more other readers which will set last_nsec_offset to the same
 165                  * value as the one we read at the beginning of the loop.  If this
 166                  * happens, we'll end up returning a slightly newer time than we ought to
 167                  * (the jump forward is at most "offset" nano-seconds).  There is no
 168                  * danger of causing time to go backwards, though, so we are safe in that
 169                  * sense.  We could make the probability of this unlucky case occurring
 170                  * arbitrarily small by encoding a version number in last_nsec_offset, but
 171                  * even without versioning, the probability of this unlucky case should be
 172                  * so small that we won't worry about it.
 173                  */
 174                 if (offset <= old) {
 175                         offset = old;
 176                         break;
 177                 } else if (likely(cmpxchg(&last_nsec_offset, old, offset) == old))
 178                         break;
 179
 180                 /* someone else beat us to updating last_nsec_offset; try again */
 181         }
 182
 183         usec = (nsec + offset) / 1000;
 184
 185         while (unlikely(usec >= USEC_PER_SEC)) {
 186                 usec -= USEC_PER_SEC;
 187                 ++sec;
 188         }
 189
 190         tv->tv_sec = sec;
 191         tv->tv_usec = usec;
 192 }
 193
 194 EXPORT_SYMBOL(do_gettimeofday);
 195
 196 /*
 197  * The profiling function is SMP safe. (nothing can mess
 198  * around with "current", and the profiling counters are
 199  * updated with atomic operations). This is especially
 200  * useful with a profiling multiplier != 1
 201  */
 202 static inline void
 203 ia64_do_profile (struct pt_regs * regs)
 204 {
 205         unsigned long ip, slot;
 206         extern cpumask_t prof_cpu_mask;
 207
 208         profile_hook(regs);
 209
 210         if (user_mode(regs))
 211                 return;
 212
 213         if (!prof_buffer)
 214                 return;
 215
 216         ip = instruction_pointer(regs);
 217         /* Conserve space in histogram by encoding slot bits in address
 218          * bits 2 and 3 rather than bits 0 and 1.
 219          */
 220         slot = ip & 3;
 221         ip = (ip & ~3UL) + 4*slot;
 222
 223         /*
 224          * Only measure the CPUs specified by /proc/irq/prof_cpu_mask.
 225          * (default is all CPUs.)
 226          */
 227         if (!cpu_isset(smp_processor_id(), prof_cpu_mask))
 228                 return;
 229
 230         ip -= (unsigned long) &_stext;
 231         ip >>= prof_shift;
 232         /*
 233          * Don't ignore out-of-bounds IP values silently,
 234          * put them into the last histogram slot, so if
 235          * present, they will show up as a sharp peak.
 236          */
 237         if (ip > prof_len-1)
 238                 ip = prof_len-1;
 239         atomic_inc((atomic_t *)&prof_buffer[ip]);
 240 }
 241
 242 static irqreturn_t
 243 timer_interrupt (int irq, void *dev_id, struct pt_regs *regs)
 244 {
 245         unsigned long new_itm;
 246
 247         platform_timer_interrupt(irq, dev_id, regs);
 248
 249         new_itm = local_cpu_data->itm_next;
 250
 251         if (!time_after(ia64_get_itc(), new_itm))
 252                 printk(KERN_ERR "Oops: timer tick before it's due (itc=%lx,itm=%lx)\n",
 253                        ia64_get_itc(), new_itm);
 254
 255         ia64_do_profile(regs);
 256
 257         while (1) {
 258 #ifdef CONFIG_SMP
 259                 /*
 260                  * For UP, this is done in do_timer().  Weird, but
 261                  * fixing that would require updates to all
 262                  * platforms.
 263                  */
 264                 update_process_times(user_mode(regs));
 265 #endif
 266                 new_itm += local_cpu_data->itm_delta;
 267
 268                 if (smp_processor_id() == TIME_KEEPER_ID) {
 269                         /*
 270                          * Here we are in the timer irq handler. We have irqs locally
 271                          * disabled, but we don't know if the timer_bh is running on
 272                          * another CPU. We need to avoid to SMP race by acquiring the
 273                          * xtime_lock.
 274                          */
 275                         write_seqlock(&xtime_lock);
 276                         do_timer(regs);
 277                         local_cpu_data->itm_next = new_itm;
 278                         write_sequnlock(&xtime_lock);
 279                 } else
 280                         local_cpu_data->itm_next = new_itm;
 281
 282                 if (time_after(new_itm, ia64_get_itc()))
 283                         break;
 284         }
 285
 286         do {
 287                 /*
 288                  * If we're too close to the next clock tick for
 289                  * comfort, we increase the safety margin by
 290                  * intentionally dropping the next tick(s).  We do NOT
 291                  * update itm.next because that would force us to call
 292                  * do_timer() which in turn would let our clock run
 293                  * too fast (with the potentially devastating effect
 294                  * of losing monotony of time).
 295                  */
 296                 while (!time_after(new_itm, ia64_get_itc() + local_cpu_data->itm_delta/2))
 297                         new_itm += local_cpu_data->itm_delta;
 298                 ia64_set_itm(new_itm);
 299                 /* double check, in case we got hit by a (slow) PMI: */
 300         } while (time_after_eq(ia64_get_itc(), new_itm));
 301         return IRQ_HANDLED;
 302 }
 303
 304 /*
 305  * Encapsulate access to the itm structure for SMP.
 306  */
 307 void
 308 ia64_cpu_local_tick (void)
 309 {
 310         int cpu = smp_processor_id();
 311         unsigned long shift = 0, delta;
 312
 313         /* arrange for the cycle counter to generate a timer interrupt: */
 314         ia64_set_itv(IA64_TIMER_VECTOR);
 315
 316         delta = local_cpu_data->itm_delta;
 317         /*
 318          * Stagger the timer tick for each CPU so they don't occur all at (almost) the
 319          * same time:
 320          */
 321         if (cpu) {
 322                 unsigned long hi = 1UL << ia64_fls(cpu);
 323                 shift = (2*(cpu - hi) + 1) * delta/hi/2;
 324         }
 325         local_cpu_data->itm_next = ia64_get_itc() + delta + shift;
 326         ia64_set_itm(local_cpu_data->itm_next);
 327 }
 328
 329 void __init
 330 ia64_init_itm (void)
 331 {
 332         unsigned long platform_base_freq, itc_freq;
 333         struct pal_freq_ratio itc_ratio, proc_ratio;
 334         long status, platform_base_drift, itc_drift;
 335
 336         /*
 337          * According to SAL v2.6, we need to use a SAL call to determine the platform base
 338          * frequency and then a PAL call to determine the frequency ratio between the ITC
 339          * and the base frequency.
 340          */
 341         status = ia64_sal_freq_base(SAL_FREQ_BASE_PLATFORM,
 342                                     &platform_base_freq, &platform_base_drift);
 343         if (status != 0) {
 344                 printk(KERN_ERR "SAL_FREQ_BASE_PLATFORM failed: %s\n", ia64_sal_strerror(status));
 345         } else {
 346                 status = ia64_pal_freq_ratios(&proc_ratio, 0, &itc_ratio);
 347                 if (status != 0)
 348                         printk(KERN_ERR "PAL_FREQ_RATIOS failed with status=%ld\n", status);
 349         }
 350         if (status != 0) {
 351                 /* invent "random" values */
 352                 printk(KERN_ERR
 353                        "SAL/PAL failed to obtain frequency info---inventing reasonable values\n");
 354                 platform_base_freq = 100000000;
 355                 platform_base_drift = -1;       /* no drift info */
 356                 itc_ratio.num = 3;
 357                 itc_ratio.den = 1;
 358         }
 359         if (platform_base_freq < 40000000) {
 360                 printk(KERN_ERR "Platform base frequency %lu bogus---resetting to 75MHz!\n",
 361                        platform_base_freq);
 362                 platform_base_freq = 75000000;
 363                 platform_base_drift = -1;
 364         }
 365         if (!proc_ratio.den)
 366                 proc_ratio.den = 1;     /* avoid division by zero */
 367         if (!itc_ratio.den)
 368                 itc_ratio.den = 1;      /* avoid division by zero */
 369
 370         itc_freq = (platform_base_freq*itc_ratio.num)/itc_ratio.den;
 371         if (platform_base_drift != -1)
 372                 itc_drift = platform_base_drift*itc_ratio.num/itc_ratio.den;
 373         else
 374                 itc_drift = -1;
 375
 376         local_cpu_data->itm_delta = (itc_freq + HZ/2) / HZ;
 377         printk(KERN_INFO "CPU %d: base freq=%lu.%03luMHz, ITC ratio=%lu/%lu, "
 378                "ITC freq=%lu.%03luMHz+/-%ldppm\n", smp_processor_id(),
 379                platform_base_freq / 1000000, (platform_base_freq / 1000) % 1000,
 380                itc_ratio.num, itc_ratio.den, itc_freq / 1000000, (itc_freq / 1000) % 1000,
 381                itc_drift);
 382
 383         local_cpu_data->proc_freq = (platform_base_freq*proc_ratio.num)/proc_ratio.den;
 384         local_cpu_data->itc_freq = itc_freq;
 385         local_cpu_data->cyc_per_usec = (itc_freq + USEC_PER_SEC/2) / USEC_PER_SEC;
 386         local_cpu_data->nsec_per_cyc = ((NSEC_PER_SEC<<IA64_NSEC_PER_CYC_SHIFT)
 387                                         + itc_freq/2)/itc_freq;
 388
 389         if (!(sal_platform_features & IA64_SAL_PLATFORM_FEATURE_ITC_DRIFT)) {
 390                 itc_interpolator.frequency = local_cpu_data->itc_freq;
 391                 itc_interpolator.drift = itc_drift;
 392                 register_time_interpolator(&itc_interpolator);
 393         }
 394
 395         /* Setup the CPU local timer tick */
 396         ia64_cpu_local_tick();
 397 }
 398
 399 static struct irqaction timer_irqaction = {
 400         .handler =      timer_interrupt,
 401         .flags =        SA_INTERRUPT,
 402         .name =         "timer"
 403 };
 404
 405 void __init
 406 time_init (void)
 407 {
 408         register_percpu_irq(IA64_TIMER_VECTOR, &timer_irqaction);
 409         efi_gettimeofday(&xtime);
 410         ia64_init_itm();
 411
 412         /*
 413          * Initialize wall_to_monotonic such that adding it to xtime will yield zero, the
 414          * tv_nsec field must be normalized (i.e., 0 <= nsec < NSEC_PER_SEC).
 415          */
 416         set_normalized_timespec(&wall_to_monotonic, -xtime.tv_sec, -xtime.tv_nsec);
 417 }