kernel/time/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/ktime.h>
  12 #include <linux/kernel.h>
  13 #include <linux/moduleparam.h>
  14 #include <linux/sched.h>
  15 #include <linux/syscore_ops.h>
  16 #include <linux/hrtimer.h>
  17 #include <linux/sched_clock.h>
  18 #include <linux/seqlock.h>
  19 #include <linux/bitops.h>
  20
  21 struct clock_data {
  22         ktime_t wrap_kt;
  23         u64 epoch_ns;
  24         u64 epoch_cyc;
  25         seqcount_t seq;
  26         unsigned long rate;
  27         u32 mult;
  28         u32 shift;
  29         bool suspended;
  30 };
  31
  32 static struct hrtimer sched_clock_timer;
  33 static int irqtime = -1;
  34
  35 core_param(irqtime, irqtime, int, 0400);
  36
  37 static struct clock_data cd = {
  38         .mult   = NSEC_PER_SEC / HZ,
  39 };
  40
  41 static u64 __read_mostly sched_clock_mask;
  42
  43 static u64 notrace jiffy_sched_clock_read(void)
  44 {
  45         /*
  46          * We don't need to use get_jiffies_64 on 32-bit arches here
  47          * because we register with BITS_PER_LONG
  48          */
  49         return (u64)(jiffies - INITIAL_JIFFIES);
  50 }
  51
  52 static u32 __read_mostly (*read_sched_clock_32)(void);
  53
  54 static u64 notrace read_sched_clock_32_wrapper(void)
  55 {
  56         return read_sched_clock_32();
  57 }
  58
  59 static u64 __read_mostly (*read_sched_clock)(void) = jiffy_sched_clock_read;
  60
  61 static inline u64 notrace cyc_to_ns(u64 cyc, u32 mult, u32 shift)
  62 {
  63         return (cyc * mult) >> shift;
  64 }
  65
  66 unsigned long long notrace sched_clock(void)
  67 {
  68         u64 epoch_ns;
  69         u64 epoch_cyc;
  70         u64 cyc;
  71         unsigned long seq;
  72
  73         if (cd.suspended)
  74                 return cd.epoch_ns;
  75
  76         do {
  77                 seq = raw_read_seqcount_begin(&cd.seq);
  78                 epoch_cyc = cd.epoch_cyc;
  79                 epoch_ns = cd.epoch_ns;
  80         } while (read_seqcount_retry(&cd.seq, seq));
  81
  82         cyc = read_sched_clock();
  83         cyc = (cyc - epoch_cyc) & sched_clock_mask;
  84         return epoch_ns + cyc_to_ns(cyc, cd.mult, cd.shift);
  85 }
  86
  87 /*
  88  * Atomically update the sched_clock epoch.
  89  */
  90 static void notrace update_sched_clock(void)
  91 {
  92         unsigned long flags;
  93         u64 cyc;
  94         u64 ns;
  95
  96         cyc = read_sched_clock();
  97         ns = cd.epoch_ns +
  98                 cyc_to_ns((cyc - cd.epoch_cyc) & sched_clock_mask,
  99                           cd.mult, cd.shift);
 100
 101         raw_local_irq_save(flags);
 102         raw_write_seqcount_begin(&cd.seq);
 103         cd.epoch_ns = ns;
 104         cd.epoch_cyc = cyc;
 105         raw_write_seqcount_end(&cd.seq);
 106         raw_local_irq_restore(flags);
 107 }
 108
 109 static enum hrtimer_restart sched_clock_poll(struct hrtimer *hrt)
 110 {
 111         update_sched_clock();
 112         hrtimer_forward_now(hrt, cd.wrap_kt);
 113         return HRTIMER_RESTART;
 114 }
 115
 116 void __init sched_clock_register(u64 (*read)(void), int bits,
 117                                  unsigned long rate)
 118 {
 119         u64 res, wrap, new_mask, new_epoch, cyc, ns;
 120         u32 new_mult, new_shift;
 121         ktime_t new_wrap_kt;
 122         unsigned long r;
 123         char r_unit;
 124
 125         if (cd.rate > rate)
 126                 return;
 127
 128         WARN_ON(!irqs_disabled());
 129
 130         /* calculate the mult/shift to convert counter ticks to ns. */
 131         clocks_calc_mult_shift(&new_mult, &new_shift, rate, NSEC_PER_SEC, 3600);
 132
 133         new_mask = CLOCKSOURCE_MASK(bits);
 134
 135         /* calculate how many ns until we wrap */
 136         wrap = clocks_calc_max_nsecs(new_mult, new_shift, 0, new_mask);
 137         new_wrap_kt = ns_to_ktime(wrap - (wrap >> 3));
 138
 139         /* update epoch for new counter and update epoch_ns from old counter*/
 140         new_epoch = read();
 141         cyc = read_sched_clock();
 142         ns = cd.epoch_ns + cyc_to_ns((cyc - cd.epoch_cyc) & sched_clock_mask,
 143                           cd.mult, cd.shift);
 144
 145         raw_write_seqcount_begin(&cd.seq);
 146         read_sched_clock = read;
 147         sched_clock_mask = new_mask;
 148         cd.rate = rate;
 149         cd.wrap_kt = new_wrap_kt;
 150         cd.mult = new_mult;
 151         cd.shift = new_shift;
 152         cd.epoch_cyc = new_epoch;
 153         cd.epoch_ns = ns;
 154         raw_write_seqcount_end(&cd.seq);
 155
 156         r = rate;
 157         if (r >= 4000000) {
 158                 r /= 1000000;
 159                 r_unit = 'M';
 160         } else if (r >= 1000) {
 161                 r /= 1000;
 162                 r_unit = 'k';
 163         } else
 164                 r_unit = ' ';
 165
 166         /* calculate the ns resolution of this counter */
 167         res = cyc_to_ns(1ULL, new_mult, new_shift);
 168
 169         pr_info("sched_clock: %u bits at %lu%cHz, resolution %lluns, wraps every %lluns\n",
 170                 bits, r, r_unit, res, wrap);
 171
 172         /* Enable IRQ time accounting if we have a fast enough sched_clock */
 173         if (irqtime > 0 || (irqtime == -1 && rate >= 1000000))
 174                 enable_sched_clock_irqtime();
 175
 176         pr_debug("Registered %pF as sched_clock source\n", read);
 177 }
 178
 179 void __init setup_sched_clock(u32 (*read)(void), int bits, unsigned long rate)
 180 {
 181         read_sched_clock_32 = read;
 182         sched_clock_register(read_sched_clock_32_wrapper, bits, rate);
 183 }
 184
 185 void __init sched_clock_postinit(void)
 186 {
 187         /*
 188          * If no sched_clock function has been provided at that point,
 189          * make it the final one one.
 190          */
 191         if (read_sched_clock == jiffy_sched_clock_read)
 192                 sched_clock_register(jiffy_sched_clock_read, BITS_PER_LONG, HZ);
 193
 194         update_sched_clock();
 195
 196         /*
 197          * Start the timer to keep sched_clock() properly updated and
 198          * sets the initial epoch.
 199          */
 200         hrtimer_init(&sched_clock_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 201         sched_clock_timer.function = sched_clock_poll;
 202         hrtimer_start(&sched_clock_timer, cd.wrap_kt, HRTIMER_MODE_REL);
 203 }
 204
 205 static int sched_clock_suspend(void)
 206 {
 207         sched_clock_poll(&sched_clock_timer);
 208         cd.suspended = true;
 209         return 0;
 210 }
 211
 212 static void sched_clock_resume(void)
 213 {
 214         cd.epoch_cyc = read_sched_clock();
 215         cd.suspended = false;
 216 }
 217
 218 static struct syscore_ops sched_clock_ops = {
 219         .suspend = sched_clock_suspend,
 220         .resume = sched_clock_resume,
 221 };
 222
 223 static int __init sched_clock_syscore_init(void)
 224 {
 225         register_syscore_ops(&sched_clock_ops);
 226         return 0;
 227 }
 228 device_initcall(sched_clock_syscore_init);