arch/arm/kernel/sched_clock.c

   1 /*
   2  * sched_clock.c: support for extending counters to full 64-bit ns counter
   3  *
   4  * This program is free software; you can redistribute it and/or modify
   5  * it under the terms of the GNU General Public License version 2 as
   6  * published by the Free Software Foundation.
   7  */
   8 #include <linux/clocksource.h>
   9 #include <linux/init.h>
  10 #include <linux/jiffies.h>
  11 #include <linux/kernel.h>
  12 #include <linux/sched.h>
  13 #include <linux/timer.h>
  14
  15 #include <asm/sched_clock.h>
  16
  17 struct clock_data {
  18         u64 epoch_ns;
  19         u32 epoch_cyc;
  20         u32 epoch_cyc_copy;
  21         u32 mult;
  22         u32 shift;
  23 };
  24
  25 static void sched_clock_poll(unsigned long wrap_ticks);
  26 static DEFINE_TIMER(sched_clock_timer, sched_clock_poll, 0, 0);
  27
  28 static struct clock_data cd = {
  29         .mult   = NSEC_PER_SEC / HZ,
  30 };
  31
  32 static u32 __read_mostly sched_clock_mask = 0xffffffff;
  33
  34 static u32 notrace jiffy_sched_clock_read(void)
  35 {
  36         return (u32)(jiffies - INITIAL_JIFFIES);
  37 }
  38
  39 static u32 __read_mostly (*read_sched_clock)(void) = jiffy_sched_clock_read;
  40
  41 static inline u64 cyc_to_ns(u64 cyc, u32 mult, u32 shift)
  42 {
  43         return (cyc * mult) >> shift;
  44 }
  45
  46 static unsigned long long cyc_to_sched_clock(u32 cyc, u32 mask)
  47 {
  48         u64 epoch_ns;
  49         u32 epoch_cyc;
  50
  51         /*
  52          * Load the epoch_cyc and epoch_ns atomically.  We do this by
  53          * ensuring that we always write epoch_cyc, epoch_ns and
  54          * epoch_cyc_copy in strict order, and read them in strict order.
  55          * If epoch_cyc and epoch_cyc_copy are not equal, then we're in
  56          * the middle of an update, and we should repeat the load.
  57          */
  58         do {
  59                 epoch_cyc = cd.epoch_cyc;
  60                 smp_rmb();
  61                 epoch_ns = cd.epoch_ns;
  62                 smp_rmb();
  63         } while (epoch_cyc != cd.epoch_cyc_copy);
  64
  65         return epoch_ns + cyc_to_ns((cyc - epoch_cyc) & mask, cd.mult, cd.shift);
  66 }
  67
  68 /*
  69  * Atomically update the sched_clock epoch.
  70  */
  71 static void notrace update_sched_clock(void)
  72 {
  73         unsigned long flags;
  74         u32 cyc;
  75         u64 ns;
  76
  77         cyc = read_sched_clock();
  78         ns = cd.epoch_ns +
  79                 cyc_to_ns((cyc - cd.epoch_cyc) & sched_clock_mask,
  80                           cd.mult, cd.shift);
  81         /*
  82          * Write epoch_cyc and epoch_ns in a way that the update is
  83          * detectable in cyc_to_fixed_sched_clock().
  84          */
  85         raw_local_irq_save(flags);
  86         cd.epoch_cyc = cyc;
  87         smp_wmb();
  88         cd.epoch_ns = ns;
  89         smp_wmb();
  90         cd.epoch_cyc_copy = cyc;
  91         raw_local_irq_restore(flags);
  92 }
  93
  94 static void sched_clock_poll(unsigned long wrap_ticks)
  95 {
  96         mod_timer(&sched_clock_timer, round_jiffies(jiffies + wrap_ticks));
  97         update_sched_clock();
  98 }
  99
 100 void __init setup_sched_clock(u32 (*read)(void), int bits, unsigned long rate)
 101 {
 102         unsigned long r, w;
 103         u64 res, wrap;
 104         char r_unit;
 105
 106         BUG_ON(bits > 32);
 107         WARN_ON(!irqs_disabled());
 108         WARN_ON(read_sched_clock != jiffy_sched_clock_read);
 109         read_sched_clock = read;
 110         sched_clock_mask = (1 << bits) - 1;
 111
 112         /* calculate the mult/shift to convert counter ticks to ns. */
 113         clocks_calc_mult_shift(&cd.mult, &cd.shift, rate, NSEC_PER_SEC, 0);
 114
 115         r = rate;
 116         if (r >= 4000000) {
 117                 r /= 1000000;
 118                 r_unit = 'M';
 119         } else if (r >= 1000) {
 120                 r /= 1000;
 121                 r_unit = 'k';
 122         } else
 123                 r_unit = ' ';
 124
 125         /* calculate how many ns until we wrap */
 126         wrap = cyc_to_ns((1ULL << bits) - 1, cd.mult, cd.shift);
 127         do_div(wrap, NSEC_PER_MSEC);
 128         w = wrap;
 129
 130         /* calculate the ns resolution of this counter */
 131         res = cyc_to_ns(1ULL, cd.mult, cd.shift);
 132         pr_info("sched_clock: %u bits at %lu%cHz, resolution %lluns, wraps every %lums\n",
 133                 bits, r, r_unit, res, w);
 134
 135         /*
 136          * Start the timer to keep sched_clock() properly updated and
 137          * sets the initial epoch.
 138          */
 139         sched_clock_timer.data = msecs_to_jiffies(w - (w / 10));
 140         update_sched_clock();
 141
 142         /*
 143          * Ensure that sched_clock() starts off at 0ns
 144          */
 145         cd.epoch_ns = 0;
 146
 147         pr_debug("Registered %pF as sched_clock source\n", read);
 148 }
 149
 150 unsigned long long notrace sched_clock(void)
 151 {
 152         u32 cyc = read_sched_clock();
 153         return cyc_to_sched_clock(cyc, sched_clock_mask);
 154 }
 155
 156 void __init sched_clock_postinit(void)
 157 {
 158         /*
 159          * If no sched_clock function has been provided at that point,
 160          * make it the final one one.
 161          */
 162         if (read_sched_clock == jiffy_sched_clock_read)
 163                 setup_sched_clock(jiffy_sched_clock_read, 32, HZ);
 164
 165         sched_clock_poll(sched_clock_timer.data);
 166 }