sched_clock.c 4.6 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192
  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/moduleparam.h>
  13. #include <linux/sched.h>
  14. #include <linux/syscore_ops.h>
  15. #include <linux/timer.h>
  16. #include <asm/sched_clock.h>
  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. static void sched_clock_poll(unsigned long wrap_ticks);
  25. static DEFINE_TIMER(sched_clock_timer, sched_clock_poll, 0, 0);
  26. static int irqtime = -1;
  27. core_param(irqtime, irqtime, int, 0400);
  28. static struct clock_data cd = {
  29. .mult = NSEC_PER_SEC / HZ,
  30. };
  31. static u32 __read_mostly sched_clock_mask = 0xffffffff;
  32. static u32 notrace jiffy_sched_clock_read(void)
  33. {
  34. return (u32)(jiffies - INITIAL_JIFFIES);
  35. }
  36. static u32 __read_mostly (*read_sched_clock)(void) = jiffy_sched_clock_read;
  37. static inline u64 cyc_to_ns(u64 cyc, u32 mult, u32 shift)
  38. {
  39. return (cyc * mult) >> shift;
  40. }
  41. static unsigned long long cyc_to_sched_clock(u32 cyc, u32 mask)
  42. {
  43. u64 epoch_ns;
  44. u32 epoch_cyc;
  45. /*
  46. * Load the epoch_cyc and epoch_ns atomically. We do this by
  47. * ensuring that we always write epoch_cyc, epoch_ns and
  48. * epoch_cyc_copy in strict order, and read them in strict order.
  49. * If epoch_cyc and epoch_cyc_copy are not equal, then we're in
  50. * the middle of an update, and we should repeat the load.
  51. */
  52. do {
  53. epoch_cyc = cd.epoch_cyc;
  54. smp_rmb();
  55. epoch_ns = cd.epoch_ns;
  56. smp_rmb();
  57. } while (epoch_cyc != cd.epoch_cyc_copy);
  58. return epoch_ns + cyc_to_ns((cyc - epoch_cyc) & mask, cd.mult, cd.shift);
  59. }
  60. /*
  61. * Atomically update the sched_clock epoch.
  62. */
  63. static void notrace update_sched_clock(void)
  64. {
  65. unsigned long flags;
  66. u32 cyc;
  67. u64 ns;
  68. cyc = read_sched_clock();
  69. ns = cd.epoch_ns +
  70. cyc_to_ns((cyc - cd.epoch_cyc) & sched_clock_mask,
  71. cd.mult, cd.shift);
  72. /*
  73. * Write epoch_cyc and epoch_ns in a way that the update is
  74. * detectable in cyc_to_fixed_sched_clock().
  75. */
  76. raw_local_irq_save(flags);
  77. cd.epoch_cyc = cyc;
  78. smp_wmb();
  79. cd.epoch_ns = ns;
  80. smp_wmb();
  81. cd.epoch_cyc_copy = cyc;
  82. raw_local_irq_restore(flags);
  83. }
  84. static void sched_clock_poll(unsigned long wrap_ticks)
  85. {
  86. mod_timer(&sched_clock_timer, round_jiffies(jiffies + wrap_ticks));
  87. update_sched_clock();
  88. }
  89. void __init setup_sched_clock(u32 (*read)(void), int bits, unsigned long rate)
  90. {
  91. unsigned long r, w;
  92. u64 res, wrap;
  93. char r_unit;
  94. BUG_ON(bits > 32);
  95. WARN_ON(!irqs_disabled());
  96. WARN_ON(read_sched_clock != jiffy_sched_clock_read);
  97. read_sched_clock = read;
  98. sched_clock_mask = (1 << bits) - 1;
  99. /* calculate the mult/shift to convert counter ticks to ns. */
  100. clocks_calc_mult_shift(&cd.mult, &cd.shift, rate, NSEC_PER_SEC, 0);
  101. r = rate;
  102. if (r >= 4000000) {
  103. r /= 1000000;
  104. r_unit = 'M';
  105. } else if (r >= 1000) {
  106. r /= 1000;
  107. r_unit = 'k';
  108. } else
  109. r_unit = ' ';
  110. /* calculate how many ns until we wrap */
  111. wrap = cyc_to_ns((1ULL << bits) - 1, cd.mult, cd.shift);
  112. do_div(wrap, NSEC_PER_MSEC);
  113. w = wrap;
  114. /* calculate the ns resolution of this counter */
  115. res = cyc_to_ns(1ULL, cd.mult, cd.shift);
  116. pr_info("sched_clock: %u bits at %lu%cHz, resolution %lluns, wraps every %lums\n",
  117. bits, r, r_unit, res, w);
  118. /*
  119. * Start the timer to keep sched_clock() properly updated and
  120. * sets the initial epoch.
  121. */
  122. sched_clock_timer.data = msecs_to_jiffies(w - (w / 10));
  123. update_sched_clock();
  124. /*
  125. * Ensure that sched_clock() starts off at 0ns
  126. */
  127. cd.epoch_ns = 0;
  128. /* Enable IRQ time accounting if we have a fast enough sched_clock */
  129. if (irqtime > 0 || (irqtime == -1 && rate >= 1000000))
  130. enable_sched_clock_irqtime();
  131. pr_debug("Registered %pF as sched_clock source\n", read);
  132. }
  133. unsigned long long notrace sched_clock(void)
  134. {
  135. u32 cyc = read_sched_clock();
  136. return cyc_to_sched_clock(cyc, sched_clock_mask);
  137. }
  138. void __init sched_clock_postinit(void)
  139. {
  140. /*
  141. * If no sched_clock function has been provided at that point,
  142. * make it the final one one.
  143. */
  144. if (read_sched_clock == jiffy_sched_clock_read)
  145. setup_sched_clock(jiffy_sched_clock_read, 32, HZ);
  146. sched_clock_poll(sched_clock_timer.data);
  147. }
  148. static int sched_clock_suspend(void)
  149. {
  150. sched_clock_poll(sched_clock_timer.data);
  151. return 0;
  152. }
  153. static struct syscore_ops sched_clock_ops = {
  154. .suspend = sched_clock_suspend,
  155. };
  156. static int __init sched_clock_syscore_init(void)
  157. {
  158. register_syscore_ops(&sched_clock_ops);
  159. return 0;
  160. }
  161. device_initcall(sched_clock_syscore_init);