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/profile.h>
  16 #include <linux/sched.h>
  17 #include <linux/time.h>
  18 #include <linux/interrupt.h>
  19 #include <linux/efi.h>
  20
  21 #include <asm/delay.h>
  22 #include <asm/hw_irq.h>
  23 #include <asm/ptrace.h>
  24 #include <asm/sal.h>
  25 #include <asm/system.h>
  26
  27 extern unsigned long wall_jiffies;
  28 extern unsigned long last_nsec_offset;
  29
  30 u64 jiffies_64 = INITIAL_JIFFIES;
  31
  32 #ifdef CONFIG_IA64_DEBUG_IRQ
  33
  34 unsigned long last_cli_ip;
  35
  36 #endif
  37
  38 static void
  39 do_profile (unsigned long ip)
  40 {
  41         extern unsigned long prof_cpu_mask;
  42         extern char _stext;
  43
  44         if (!prof_buffer)
  45                 return;
  46
  47         if (!((1UL << smp_processor_id()) & prof_cpu_mask))
  48                 return;
  49
  50         ip -= (unsigned long) &_stext;
  51         ip >>= prof_shift;
  52         /*
  53          * Don't ignore out-of-bounds IP values silently, put them into the last
  54          * histogram slot, so if present, they will show up as a sharp peak.
  55          */
  56         if (ip > prof_len - 1)
  57                 ip = prof_len - 1;
  58
  59         atomic_inc((atomic_t *) &prof_buffer[ip]);
  60 }
  61
  62 /*
  63  * Return the number of nano-seconds that elapsed since the last update to jiffy.  The
  64  * xtime_lock must be at least read-locked when calling this routine.
  65  */
  66 static inline unsigned long
  67 gettimeoffset (void)
  68 {
  69         unsigned long elapsed_cycles, lost = jiffies - wall_jiffies;
  70         unsigned long now, last_tick;
  71 #       define time_keeper_id   0       /* smp_processor_id() of time-keeper */
  72
  73         last_tick = (cpu_data(time_keeper_id)->itm_next
  74                      - (lost + 1)*cpu_data(time_keeper_id)->itm_delta);
  75
  76         now = ia64_get_itc();
  77         if (unlikely((long) (now - last_tick) < 0)) {
  78                 printk(KERN_ERR "CPU %d: now < last_tick (now=0x%lx,last_tick=0x%lx)!\n",
  79                        smp_processor_id(), now, last_tick);
  80                 return last_nsec_offset;
  81         }
  82         elapsed_cycles = now - last_tick;
  83         return (elapsed_cycles*local_cpu_data->nsec_per_cyc) >> IA64_NSEC_PER_CYC_SHIFT;
  84 }
  85
  86 static inline void
  87 set_normalized_timespec (struct timespec *ts, time_t sec, long nsec)
  88 {
  89         while (nsec > NSEC_PER_SEC) {
  90                 nsec -= NSEC_PER_SEC;
  91                 ++sec;
  92         }
  93         while (nsec < 0) {
  94                 nsec += NSEC_PER_SEC;
  95                 --sec;
  96         }
  97         ts->tv_sec = sec;
  98         ts->tv_nsec = nsec;
  99 }
 100
 101 void
 102 do_settimeofday (struct timeval *tv)
 103 {
 104         time_t wtm_sec, sec = tv->tv_sec;
 105         long wtm_nsec, nsec = tv->tv_usec * 1000;
 106
 107         write_seqlock_irq(&xtime_lock);
 108         {
 109                 /*
 110                  * This is revolting. We need to set "xtime" correctly. However, the value
 111                  * in this location is the value at the most recent update of wall time.
 112                  * Discover what correction gettimeofday would have done, and then undo
 113                  * it!
 114                  */
 115                 nsec -= gettimeoffset();
 116
 117                 wtm_sec  = wall_to_monotonic.tv_sec + (xtime.tv_sec - sec);
 118                 wtm_nsec = wall_to_monotonic.tv_nsec + (xtime.tv_nsec - nsec);
 119
 120                 set_normalized_timespec(&xtime, sec, nsec);
 121                 set_normalized_timespec(&wall_to_monotonic, wtm_sec, wtm_nsec);
 122
 123                 time_adjust = 0;                /* stop active adjtime() */
 124                 time_status |= STA_UNSYNC;
 125                 time_maxerror = NTP_PHASE_LIMIT;
 126                 time_esterror = NTP_PHASE_LIMIT;
 127         }
 128         write_sequnlock_irq(&xtime_lock);
 129         clock_was_set();
 130 }
 131
 132 void
 133 do_gettimeofday (struct timeval *tv)
 134 {
 135         unsigned long seq, nsec, usec, sec, old, offset;
 136
 137         while (1) {
 138                 seq = read_seqbegin(&xtime_lock);
 139                 {
 140                         old = last_nsec_offset;
 141                         offset = gettimeoffset();
 142                         sec = xtime.tv_sec;
 143                         nsec = xtime.tv_nsec;
 144                 }
 145                 if (unlikely(read_seqretry(&xtime_lock, seq)))
 146                         continue;
 147                 /*
 148                  * Ensure that for any pair of causally ordered gettimeofday() calls, time
 149                  * never goes backwards (even when ITC on different CPUs are not perfectly
 150                  * synchronized).  (A pair of concurrent calls to gettimeofday() is by
 151                  * definition non-causal and hence it makes no sense to talk about
 152                  * time-continuity for such calls.)
 153                  *
 154                  * Doing this in a lock-free and race-free manner is tricky.  Here is why
 155                  * it works (most of the time): read_seqretry() just succeeded, which
 156                  * implies we calculated a consistent (valid) value for "offset".  If the
 157                  * cmpxchg() below succeeds, we further know that last_nsec_offset still
 158                  * has the same value as at the beginning of the loop, so there was
 159                  * presumably no timer-tick or other updates to last_nsec_offset in the
 160                  * meantime.  This isn't 100% true though: there _is_ a possibility of a
 161                  * timer-tick occurring right right after read_seqretry() and then getting
 162                  * zero or more other readers which will set last_nsec_offset to the same
 163                  * value as the one we read at the beginning of the loop.  If this
 164                  * happens, we'll end up returning a slightly newer time than we ought to
 165                  * (the jump forward is at most "offset" nano-seconds).  There is no
 166                  * danger of causing time to go backwards, though, so we are safe in that
 167                  * sense.  We could make the probability of this unlucky case occurring
 168                  * arbitrarily small by encoding a version number in last_nsec_offset, but
 169                  * even without versioning, the probability of this unlucky case should be
 170                  * so small that we won't worry about it.
 171                  */
 172                 if (offset <= old) {
 173                         offset = old;
 174                         break;
 175                 } else if (likely(cmpxchg(&last_nsec_offset, old, offset) == old))
 176                         break;
 177
 178                 /* someone else beat us to updating last_nsec_offset; try again */
 179         }
 180
 181         usec = (nsec + offset) / 1000;
 182
 183         while (unlikely(usec >= USEC_PER_SEC)) {
 184                 usec -= USEC_PER_SEC;
 185                 ++sec;
 186         }
 187
 188         tv->tv_sec = sec;
 189         tv->tv_usec = usec;
 190 }
 191
 192 static irqreturn_t
 193 timer_interrupt (int irq, void *dev_id, struct pt_regs *regs)
 194 {
 195         unsigned long new_itm;
 196
 197         new_itm = local_cpu_data->itm_next;
 198
 199         if (!time_after(ia64_get_itc(), new_itm))
 200                 printk(KERN_ERR "Oops: timer tick before it's due (itc=%lx,itm=%lx)\n",
 201                        ia64_get_itc(), new_itm);
 202
 203         while (1) {
 204                 /*
 205                  * Do kernel PC profiling here.  We multiply the instruction number by
 206                  * four so that we can use a prof_shift of 2 to get instruction-level
 207                  * instead of just bundle-level accuracy.
 208                  */
 209                 if (!user_mode(regs))
 210                         do_profile(regs->cr_iip + 4*ia64_psr(regs)->ri);
 211
 212 #ifdef CONFIG_SMP
 213                 smp_do_timer(regs);
 214 #endif
 215                 new_itm += local_cpu_data->itm_delta;
 216
 217                 if (smp_processor_id() == 0) {
 218                         /*
 219                          * Here we are in the timer irq handler. We have irqs locally
 220                          * disabled, but we don't know if the timer_bh is running on
 221                          * another CPU. We need to avoid to SMP race by acquiring the
 222                          * xtime_lock.
 223                          */
 224                         write_seqlock(&xtime_lock);
 225                         do_timer(regs);
 226                         local_cpu_data->itm_next = new_itm;
 227                         write_sequnlock(&xtime_lock);
 228                 } else
 229                         local_cpu_data->itm_next = new_itm;
 230
 231                 if (time_after(new_itm, ia64_get_itc()))
 232                         break;
 233         }
 234
 235         do {
 236             /*
 237              * If we're too close to the next clock tick for comfort, we increase the
 238              * safety margin by intentionally dropping the next tick(s).  We do NOT update
 239              * itm.next because that would force us to call do_timer() which in turn would
 240              * let our clock run too fast (with the potentially devastating effect of
 241              * losing monotony of time).
 242              */
 243             while (!time_after(new_itm, ia64_get_itc() + local_cpu_data->itm_delta/2))
 244               new_itm += local_cpu_data->itm_delta;
 245             ia64_set_itm(new_itm);
 246             /* double check, in case we got hit by a (slow) PMI: */
 247         } while (time_after_eq(ia64_get_itc(), new_itm));
 248         return IRQ_HANDLED;
 249 }
 250
 251 /*
 252  * Encapsulate access to the itm structure for SMP.
 253  */
 254 void
 255 ia64_cpu_local_tick (void)
 256 {
 257         int cpu = smp_processor_id();
 258         unsigned long shift = 0, delta;
 259
 260         /* arrange for the cycle counter to generate a timer interrupt: */
 261         ia64_set_itv(IA64_TIMER_VECTOR);
 262
 263         delta = local_cpu_data->itm_delta;
 264         /*
 265          * Stagger the timer tick for each CPU so they don't occur all at (almost) the
 266          * same time:
 267          */
 268         if (cpu) {
 269                 unsigned long hi = 1UL << ia64_fls(cpu);
 270                 shift = (2*(cpu - hi) + 1) * delta/hi/2;
 271         }
 272         local_cpu_data->itm_next = ia64_get_itc() + delta + shift;
 273         ia64_set_itm(local_cpu_data->itm_next);
 274 }
 275
 276 void __init
 277 ia64_init_itm (void)
 278 {
 279         unsigned long platform_base_freq, itc_freq, drift;
 280         struct pal_freq_ratio itc_ratio, proc_ratio;
 281         long status;
 282
 283         /*
 284          * According to SAL v2.6, we need to use a SAL call to determine the platform base
 285          * frequency and then a PAL call to determine the frequency ratio between the ITC
 286          * and the base frequency.
 287          */
 288         status = ia64_sal_freq_base(SAL_FREQ_BASE_PLATFORM, &platform_base_freq, &drift);
 289         if (status != 0) {
 290                 printk(KERN_ERR "SAL_FREQ_BASE_PLATFORM failed: %s\n", ia64_sal_strerror(status));
 291         } else {
 292                 status = ia64_pal_freq_ratios(&proc_ratio, 0, &itc_ratio);
 293                 if (status != 0)
 294                         printk(KERN_ERR "PAL_FREQ_RATIOS failed with status=%ld\n", status);
 295         }
 296         if (status != 0) {
 297                 /* invent "random" values */
 298                 printk(KERN_ERR
 299                        "SAL/PAL failed to obtain frequency info---inventing reasonable values\n");
 300                 platform_base_freq = 100000000;
 301                 itc_ratio.num = 3;
 302                 itc_ratio.den = 1;
 303         }
 304         if (platform_base_freq < 40000000) {
 305                 printk(KERN_ERR "Platform base frequency %lu bogus---resetting to 75MHz!\n",
 306                        platform_base_freq);
 307                 platform_base_freq = 75000000;
 308         }
 309         if (!proc_ratio.den)
 310                 proc_ratio.den = 1;     /* avoid division by zero */
 311         if (!itc_ratio.den)
 312                 itc_ratio.den = 1;      /* avoid division by zero */
 313
 314         itc_freq = (platform_base_freq*itc_ratio.num)/itc_ratio.den;
 315         local_cpu_data->itm_delta = (itc_freq + HZ/2) / HZ;
 316         printk(KERN_INFO "CPU %d: base freq=%lu.%03luMHz, ITC ratio=%lu/%lu, "
 317                "ITC freq=%lu.%03luMHz\n", smp_processor_id(),
 318                platform_base_freq / 1000000, (platform_base_freq / 1000) % 1000,
 319                itc_ratio.num, itc_ratio.den, itc_freq / 1000000, (itc_freq / 1000) % 1000);
 320
 321         local_cpu_data->proc_freq = (platform_base_freq*proc_ratio.num)/proc_ratio.den;
 322         local_cpu_data->itc_freq = itc_freq;
 323         local_cpu_data->cyc_per_usec = (itc_freq + USEC_PER_SEC/2) / USEC_PER_SEC;
 324         local_cpu_data->nsec_per_cyc = ((NSEC_PER_SEC<<IA64_NSEC_PER_CYC_SHIFT)
 325                                         + itc_freq/2)/itc_freq;
 326
 327         /* Setup the CPU local timer tick */
 328         ia64_cpu_local_tick();
 329 }
 330
 331 static struct irqaction timer_irqaction = {
 332         .handler =      timer_interrupt,
 333         .flags =        SA_INTERRUPT,
 334         .name =         "timer"
 335 };
 336
 337 void __init
 338 time_init (void)
 339 {
 340         register_percpu_irq(IA64_TIMER_VECTOR, &timer_irqaction);
 341         efi_gettimeofday(&xtime);
 342         ia64_init_itm();
 343
 344         /*
 345          * Initialize wall_to_monotonic such that adding it to xtime will yield zero, the
 346          * tv_nsec field must be normalized (i.e., 0 <= nsec < NSEC_PER_SEC).
 347          */
 348         set_normalized_timespec(&wall_to_monotonic, -xtime.tv_sec, -xtime.tv_nsec);
 349 }