kernel/time/sched_clock.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Generic sched_clock() support, to extend low level hardware time
   4  * counters to full 64-bit ns values.
   5  */
   6 #include <linux/clocksource.h>
   7 #include <linux/init.h>
   8 #include <linux/jiffies.h>
   9 #include <linux/ktime.h>
  10 #include <linux/kernel.h>
  11 #include <linux/moduleparam.h>
  12 #include <linux/sched.h>
  13 #include <linux/sched/clock.h>
  14 #include <linux/syscore_ops.h>
  15 #include <linux/hrtimer.h>
  16 #include <linux/sched_clock.h>
  17 #include <linux/seqlock.h>
  18 #include <linux/bitops.h>
  19
  20 #include "timekeeping.h"
  21
  22 /**
  23  * struct clock_data - all data needed for sched_clock() (including
  24  *                     registration of a new clock source)
  25  *
  26  * @seq:                Sequence counter for protecting updates. The lowest
  27  *                      bit is the index for @read_data.
  28  * @read_data:          Data required to read from sched_clock.
  29  * @wrap_kt:            Duration for which clock can run before wrapping.
  30  * @rate:               Tick rate of the registered clock.
  31  * @actual_read_sched_clock: Registered hardware level clock read function.
  32  *
  33  * The ordering of this structure has been chosen to optimize cache
  34  * performance. In particular 'seq' and 'read_data[0]' (combined) should fit
  35  * into a single 64-byte cache line.
  36  */
  37 struct clock_data {
  38         seqcount_latch_t        seq;
  39         struct clock_read_data  read_data[2];
  40         ktime_t                 wrap_kt;
  41         unsigned long           rate;
  42
  43         u64 (*actual_read_sched_clock)(void);
  44 };
  45
  46 static struct hrtimer sched_clock_timer;
  47 static int irqtime = -1;
  48
  49 core_param(irqtime, irqtime, int, 0400);
  50
  51 static u64 notrace jiffy_sched_clock_read(void)
  52 {
  53         /*
  54          * We don't need to use get_jiffies_64 on 32-bit arches here
  55          * because we register with BITS_PER_LONG
  56          */
  57         return (u64)(jiffies - INITIAL_JIFFIES);
  58 }
  59
  60 static struct clock_data cd ____cacheline_aligned = {
  61         .read_data[0] = { .mult = NSEC_PER_SEC / HZ,
  62                           .read_sched_clock = jiffy_sched_clock_read, },
  63         .actual_read_sched_clock = jiffy_sched_clock_read,
  64 };
  65
  66 static inline u64 notrace cyc_to_ns(u64 cyc, u32 mult, u32 shift)
  67 {
  68         return (cyc * mult) >> shift;
  69 }
  70
  71 notrace struct clock_read_data *sched_clock_read_begin(unsigned int *seq)
  72 {
  73         *seq = raw_read_seqcount_latch(&cd.seq);
  74         return cd.read_data + (*seq & 1);
  75 }
  76
  77 notrace int sched_clock_read_retry(unsigned int seq)
  78 {
  79         return read_seqcount_latch_retry(&cd.seq, seq);
  80 }
  81
  82 unsigned long long notrace sched_clock(void)
  83 {
  84         u64 cyc, res;
  85         unsigned int seq;
  86         struct clock_read_data *rd;
  87
  88         do {
  89                 rd = sched_clock_read_begin(&seq);
  90
  91                 cyc = (rd->read_sched_clock() - rd->epoch_cyc) &
  92                       rd->sched_clock_mask;
  93                 res = rd->epoch_ns + cyc_to_ns(cyc, rd->mult, rd->shift);
  94         } while (sched_clock_read_retry(seq));
  95
  96         return res;
  97 }
  98
  99 /*
 100  * Updating the data required to read the clock.
 101  *
 102  * sched_clock() will never observe mis-matched data even if called from
 103  * an NMI. We do this by maintaining an odd/even copy of the data and
 104  * steering sched_clock() to one or the other using a sequence counter.
 105  * In order to preserve the data cache profile of sched_clock() as much
 106  * as possible the system reverts back to the even copy when the update
 107  * completes; the odd copy is used *only* during an update.
 108  */
 109 static void update_clock_read_data(struct clock_read_data *rd)
 110 {
 111         /* update the backup (odd) copy with the new data */
 112         cd.read_data[1] = *rd;
 113
 114         /* steer readers towards the odd copy */
 115         raw_write_seqcount_latch(&cd.seq);
 116
 117         /* now its safe for us to update the normal (even) copy */
 118         cd.read_data[0] = *rd;
 119
 120         /* switch readers back to the even copy */
 121         raw_write_seqcount_latch(&cd.seq);
 122 }
 123
 124 /*
 125  * Atomically update the sched_clock() epoch.
 126  */
 127 static void update_sched_clock(void)
 128 {
 129         u64 cyc;
 130         u64 ns;
 131         struct clock_read_data rd;
 132
 133         rd = cd.read_data[0];
 134
 135         cyc = cd.actual_read_sched_clock();
 136         ns = rd.epoch_ns + cyc_to_ns((cyc - rd.epoch_cyc) & rd.sched_clock_mask, rd.mult, rd.shift);
 137
 138         rd.epoch_ns = ns;
 139         rd.epoch_cyc = cyc;
 140
 141         update_clock_read_data(&rd);
 142 }
 143
 144 static enum hrtimer_restart sched_clock_poll(struct hrtimer *hrt)
 145 {
 146         update_sched_clock();
 147         hrtimer_forward_now(hrt, cd.wrap_kt);
 148
 149         return HRTIMER_RESTART;
 150 }
 151
 152 void __init
 153 sched_clock_register(u64 (*read)(void), int bits, unsigned long rate)
 154 {
 155         u64 res, wrap, new_mask, new_epoch, cyc, ns;
 156         u32 new_mult, new_shift;
 157         unsigned long r, flags;
 158         char r_unit;
 159         struct clock_read_data rd;
 160
 161         if (cd.rate > rate)
 162                 return;
 163
 164         /* Cannot register a sched_clock with interrupts on */
 165         local_irq_save(flags);
 166
 167         /* Calculate the mult/shift to convert counter ticks to ns. */
 168         clocks_calc_mult_shift(&new_mult, &new_shift, rate, NSEC_PER_SEC, 3600);
 169
 170         new_mask = CLOCKSOURCE_MASK(bits);
 171         cd.rate = rate;
 172
 173         /* Calculate how many nanosecs until we risk wrapping */
 174         wrap = clocks_calc_max_nsecs(new_mult, new_shift, 0, new_mask, NULL);
 175         cd.wrap_kt = ns_to_ktime(wrap);
 176
 177         rd = cd.read_data[0];
 178
 179         /* Update epoch for new counter and update 'epoch_ns' from old counter*/
 180         new_epoch = read();
 181         cyc = cd.actual_read_sched_clock();
 182         ns = rd.epoch_ns + cyc_to_ns((cyc - rd.epoch_cyc) & rd.sched_clock_mask, rd.mult, rd.shift);
 183         cd.actual_read_sched_clock = read;
 184
 185         rd.read_sched_clock     = read;
 186         rd.sched_clock_mask     = new_mask;
 187         rd.mult                 = new_mult;
 188         rd.shift                = new_shift;
 189         rd.epoch_cyc            = new_epoch;
 190         rd.epoch_ns             = ns;
 191
 192         update_clock_read_data(&rd);
 193
 194         if (sched_clock_timer.function != NULL) {
 195                 /* update timeout for clock wrap */
 196                 hrtimer_start(&sched_clock_timer, cd.wrap_kt,
 197                               HRTIMER_MODE_REL_HARD);
 198         }
 199
 200         r = rate;
 201         if (r >= 4000000) {
 202                 r /= 1000000;
 203                 r_unit = 'M';
 204         } else {
 205                 if (r >= 1000) {
 206                         r /= 1000;
 207                         r_unit = 'k';
 208                 } else {
 209                         r_unit = ' ';
 210                 }
 211         }
 212
 213         /* Calculate the ns resolution of this counter */
 214         res = cyc_to_ns(1ULL, new_mult, new_shift);
 215
 216         pr_info("sched_clock: %u bits at %lu%cHz, resolution %lluns, wraps every %lluns\n",
 217                 bits, r, r_unit, res, wrap);
 218
 219         /* Enable IRQ time accounting if we have a fast enough sched_clock() */
 220         if (irqtime > 0 || (irqtime == -1 && rate >= 1000000))
 221                 enable_sched_clock_irqtime();
 222
 223         local_irq_restore(flags);
 224
 225         pr_debug("Registered %pS as sched_clock source\n", read);
 226 }
 227
 228 void __init generic_sched_clock_init(void)
 229 {
 230         /*
 231          * If no sched_clock() function has been provided at that point,
 232          * make it the final one.
 233          */
 234         if (cd.actual_read_sched_clock == jiffy_sched_clock_read)
 235                 sched_clock_register(jiffy_sched_clock_read, BITS_PER_LONG, HZ);
 236
 237         update_sched_clock();
 238
 239         /*
 240          * Start the timer to keep sched_clock() properly updated and
 241          * sets the initial epoch.
 242          */
 243         hrtimer_init(&sched_clock_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
 244         sched_clock_timer.function = sched_clock_poll;
 245         hrtimer_start(&sched_clock_timer, cd.wrap_kt, HRTIMER_MODE_REL_HARD);
 246 }
 247
 248 /*
 249  * Clock read function for use when the clock is suspended.
 250  *
 251  * This function makes it appear to sched_clock() as if the clock
 252  * stopped counting at its last update.
 253  *
 254  * This function must only be called from the critical
 255  * section in sched_clock(). It relies on the read_seqcount_retry()
 256  * at the end of the critical section to be sure we observe the
 257  * correct copy of 'epoch_cyc'.
 258  */
 259 static u64 notrace suspended_sched_clock_read(void)
 260 {
 261         unsigned int seq = raw_read_seqcount_latch(&cd.seq);
 262
 263         return cd.read_data[seq & 1].epoch_cyc;
 264 }
 265
 266 int sched_clock_suspend(void)
 267 {
 268         struct clock_read_data *rd = &cd.read_data[0];
 269
 270         update_sched_clock();
 271         hrtimer_cancel(&sched_clock_timer);
 272         rd->read_sched_clock = suspended_sched_clock_read;
 273
 274         return 0;
 275 }
 276
 277 void sched_clock_resume(void)
 278 {
 279         struct clock_read_data *rd = &cd.read_data[0];
 280
 281         rd->epoch_cyc = cd.actual_read_sched_clock();
 282         hrtimer_start(&sched_clock_timer, cd.wrap_kt, HRTIMER_MODE_REL_HARD);
 283         rd->read_sched_clock = cd.actual_read_sched_clock;
 284 }
 285
 286 static struct syscore_ops sched_clock_ops = {
 287         .suspend        = sched_clock_suspend,
 288         .resume         = sched_clock_resume,
 289 };
 290
 291 static int __init sched_clock_syscore_init(void)
 292 {
 293         register_syscore_ops(&sched_clock_ops);
 294
 295         return 0;
 296 }
 297 device_initcall(sched_clock_syscore_init);