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