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 /**
  21  * struct clock_read_data - data required to read from sched_clock()
  22  *
  23  * @epoch_ns:           sched_clock() value at last update
  24  * @epoch_cyc:          Clock cycle value at last update.
  25  * @sched_clock_mask:   Bitmask for two's complement subtraction of non 64bit
  26  *                      clocks.
  27  * @read_sched_clock:   Current clock source (or dummy source when suspended).
  28  * @mult:               Multipler for scaled math conversion.
  29  * @shift:              Shift value for scaled math conversion.
  30  *
  31  * Care must be taken when updating this structure; it is read by
  32  * some very hot code paths. It occupies <=40 bytes and, when combined
  33  * with the seqcount used to synchronize access, comfortably fits into
  34  * a 64 byte cache line.
  35  */
  36 struct clock_read_data {
  37         u64 epoch_ns;
  38         u64 epoch_cyc;
  39         u64 sched_clock_mask;
  40         u64 (*read_sched_clock)(void);
  41         u32 mult;
  42         u32 shift;
  43 };
  44
  45 /**
  46  * struct clock_data - all data needed for sched_clock() (including
  47  *                     registration of a new clock source)
  48  *
  49  * @seq:                Sequence counter for protecting updates. The lowest
  50  *                      bit is the index for @read_data.
  51  * @read_data:          Data required to read from sched_clock.
  52  * @wrap_kt:            Duration for which clock can run before wrapping.
  53  * @rate:               Tick rate of the registered clock.
  54  * @actual_read_sched_clock: Registered hardware level clock read function.
  55  *
  56  * The ordering of this structure has been chosen to optimize cache
  57  * performance. In particular 'seq' and 'read_data[0]' (combined) should fit
  58  * into a single 64-byte cache line.
  59  */
  60 struct clock_data {
  61         seqcount_t              seq;
  62         struct clock_read_data  read_data[2];
  63         ktime_t                 wrap_kt;
  64         unsigned long           rate;
  65
  66         u64 (*actual_read_sched_clock)(void);
  67 };
  68
  69 static struct hrtimer sched_clock_timer;
  70 static int irqtime = -1;
  71
  72 core_param(irqtime, irqtime, int, 0400);
  73
  74 static u64 notrace jiffy_sched_clock_read(void)
  75 {
  76         /*
  77          * We don't need to use get_jiffies_64 on 32-bit arches here
  78          * because we register with BITS_PER_LONG
  79          */
  80         return (u64)(jiffies - INITIAL_JIFFIES);
  81 }
  82
  83 static struct clock_data cd ____cacheline_aligned = {
  84         .read_data[0] = { .mult = NSEC_PER_SEC / HZ,
  85                           .read_sched_clock = jiffy_sched_clock_read, },
  86         .actual_read_sched_clock = jiffy_sched_clock_read,
  87 };
  88
  89 static inline u64 notrace cyc_to_ns(u64 cyc, u32 mult, u32 shift)
  90 {
  91         return (cyc * mult) >> shift;
  92 }
  93
  94 unsigned long long notrace sched_clock(void)
  95 {
  96         u64 cyc, res;
  97         unsigned int seq;
  98         struct clock_read_data *rd;
  99
 100         do {
 101                 seq = raw_read_seqcount(&cd.seq);
 102                 rd = cd.read_data + (seq & 1);
 103
 104                 cyc = (rd->read_sched_clock() - rd->epoch_cyc) &
 105                       rd->sched_clock_mask;
 106                 res = rd->epoch_ns + cyc_to_ns(cyc, rd->mult, rd->shift);
 107         } while (read_seqcount_retry(&cd.seq, seq));
 108
 109         return res;
 110 }
 111
 112 /*
 113  * Updating the data required to read the clock.
 114  *
 115  * sched_clock() will never observe mis-matched data even if called from
 116  * an NMI. We do this by maintaining an odd/even copy of the data and
 117  * steering sched_clock() to one or the other using a sequence counter.
 118  * In order to preserve the data cache profile of sched_clock() as much
 119  * as possible the system reverts back to the even copy when the update
 120  * completes; the odd copy is used *only* during an update.
 121  */
 122 static void update_clock_read_data(struct clock_read_data *rd)
 123 {
 124         /* update the backup (odd) copy with the new data */
 125         cd.read_data[1] = *rd;
 126
 127         /* steer readers towards the odd copy */
 128         raw_write_seqcount_latch(&cd.seq);
 129
 130         /* now its safe for us to update the normal (even) copy */
 131         cd.read_data[0] = *rd;
 132
 133         /* switch readers back to the even copy */
 134         raw_write_seqcount_latch(&cd.seq);
 135 }
 136
 137 /*
 138  * Atomically update the sched_clock() epoch.
 139  */
 140 static void update_sched_clock(void)
 141 {
 142         u64 cyc;
 143         u64 ns;
 144         struct clock_read_data rd;
 145
 146         rd = cd.read_data[0];
 147
 148         cyc = cd.actual_read_sched_clock();
 149         ns = rd.epoch_ns + cyc_to_ns((cyc - rd.epoch_cyc) & rd.sched_clock_mask, rd.mult, rd.shift);
 150
 151         rd.epoch_ns = ns;
 152         rd.epoch_cyc = cyc;
 153
 154         update_clock_read_data(&rd);
 155 }
 156
 157 static enum hrtimer_restart sched_clock_poll(struct hrtimer *hrt)
 158 {
 159         update_sched_clock();
 160         hrtimer_forward_now(hrt, cd.wrap_kt);
 161
 162         return HRTIMER_RESTART;
 163 }
 164
 165 void __init
 166 sched_clock_register(u64 (*read)(void), int bits, unsigned long rate)
 167 {
 168         u64 res, wrap, new_mask, new_epoch, cyc, ns;
 169         u32 new_mult, new_shift;
 170         unsigned long r;
 171         char r_unit;
 172         struct clock_read_data rd;
 173
 174         if (cd.rate > rate)
 175                 return;
 176
 177         WARN_ON(!irqs_disabled());
 178
 179         /* Calculate the mult/shift to convert counter ticks to ns. */
 180         clocks_calc_mult_shift(&new_mult, &new_shift, rate, NSEC_PER_SEC, 3600);
 181
 182         new_mask = CLOCKSOURCE_MASK(bits);
 183         cd.rate = rate;
 184
 185         /* Calculate how many nanosecs until we risk wrapping */
 186         wrap = clocks_calc_max_nsecs(new_mult, new_shift, 0, new_mask, NULL);
 187         cd.wrap_kt = ns_to_ktime(wrap);
 188
 189         rd = cd.read_data[0];
 190
 191         /* Update epoch for new counter and update 'epoch_ns' from old counter*/
 192         new_epoch = read();
 193         cyc = cd.actual_read_sched_clock();
 194         ns = rd.epoch_ns + cyc_to_ns((cyc - rd.epoch_cyc) & rd.sched_clock_mask, rd.mult, rd.shift);
 195         cd.actual_read_sched_clock = read;
 196
 197         rd.read_sched_clock     = read;
 198         rd.sched_clock_mask     = new_mask;
 199         rd.mult                 = new_mult;
 200         rd.shift                = new_shift;
 201         rd.epoch_cyc            = new_epoch;
 202         rd.epoch_ns             = ns;
 203
 204         update_clock_read_data(&rd);
 205
 206         if (sched_clock_timer.function != NULL) {
 207                 /* update timeout for clock wrap */
 208                 hrtimer_start(&sched_clock_timer, cd.wrap_kt, HRTIMER_MODE_REL);
 209         }
 210
 211         r = rate;
 212         if (r >= 4000000) {
 213                 r /= 1000000;
 214                 r_unit = 'M';
 215         } else {
 216                 if (r >= 1000) {
 217                         r /= 1000;
 218                         r_unit = 'k';
 219                 } else {
 220                         r_unit = ' ';
 221                 }
 222         }
 223
 224         /* Calculate the ns resolution of this counter */
 225         res = cyc_to_ns(1ULL, new_mult, new_shift);
 226
 227         pr_info("sched_clock: %u bits at %lu%cHz, resolution %lluns, wraps every %lluns\n",
 228                 bits, r, r_unit, res, wrap);
 229
 230         /* Enable IRQ time accounting if we have a fast enough sched_clock() */
 231         if (irqtime > 0 || (irqtime == -1 && rate >= 1000000))
 232                 enable_sched_clock_irqtime();
 233
 234         pr_debug("Registered %pS as sched_clock source\n", read);
 235 }
 236
 237 void __init generic_sched_clock_init(void)
 238 {
 239         /*
 240          * If no sched_clock() function has been provided at that point,
 241          * make it the final one one.
 242          */
 243         if (cd.actual_read_sched_clock == jiffy_sched_clock_read)
 244                 sched_clock_register(jiffy_sched_clock_read, BITS_PER_LONG, HZ);
 245
 246         update_sched_clock();
 247
 248         /*
 249          * Start the timer to keep sched_clock() properly updated and
 250          * sets the initial epoch.
 251          */
 252         hrtimer_init(&sched_clock_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 253         sched_clock_timer.function = sched_clock_poll;
 254         hrtimer_start(&sched_clock_timer, cd.wrap_kt, HRTIMER_MODE_REL);
 255 }
 256
 257 /*
 258  * Clock read function for use when the clock is suspended.
 259  *
 260  * This function makes it appear to sched_clock() as if the clock
 261  * stopped counting at its last update.
 262  *
 263  * This function must only be called from the critical
 264  * section in sched_clock(). It relies on the read_seqcount_retry()
 265  * at the end of the critical section to be sure we observe the
 266  * correct copy of 'epoch_cyc'.
 267  */
 268 static u64 notrace suspended_sched_clock_read(void)
 269 {
 270         unsigned int seq = raw_read_seqcount(&cd.seq);
 271
 272         return cd.read_data[seq & 1].epoch_cyc;
 273 }
 274
 275 int sched_clock_suspend(void)
 276 {
 277         struct clock_read_data *rd = &cd.read_data[0];
 278
 279         update_sched_clock();
 280         hrtimer_cancel(&sched_clock_timer);
 281         rd->read_sched_clock = suspended_sched_clock_read;
 282
 283         return 0;
 284 }
 285
 286 void sched_clock_resume(void)
 287 {
 288         struct clock_read_data *rd = &cd.read_data[0];
 289
 290         rd->epoch_cyc = cd.actual_read_sched_clock();
 291         hrtimer_start(&sched_clock_timer, cd.wrap_kt, HRTIMER_MODE_REL);
 292         rd->read_sched_clock = cd.actual_read_sched_clock;
 293 }
 294
 295 static struct syscore_ops sched_clock_ops = {
 296         .suspend        = sched_clock_suspend,
 297         .resume         = sched_clock_resume,
 298 };
 299
 300 static int __init sched_clock_syscore_init(void)
 301 {
 302         register_syscore_ops(&sched_clock_ops);
 303
 304         return 0;
 305 }
 306 device_initcall(sched_clock_syscore_init);