1 // SPDX-License-Identifier: GPL-2.0
3 * Generic sched_clock() support, to extend low level hardware time
4 * counters to full 64-bit ns values.
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/math.h>
12 #include <linux/moduleparam.h>
13 #include <linux/sched.h>
14 #include <linux/sched/clock.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>
21 #include "timekeeping.h"
24 * struct clock_data - all data needed for sched_clock() (including
25 * registration of a new clock source)
27 * @seq: Sequence counter for protecting updates. The lowest
28 * bit is the index for @read_data.
29 * @read_data: Data required to read from sched_clock.
30 * @wrap_kt: Duration for which clock can run before wrapping.
31 * @rate: Tick rate of the registered clock.
32 * @actual_read_sched_clock: Registered hardware level clock read function.
34 * The ordering of this structure has been chosen to optimize cache
35 * performance. In particular 'seq' and 'read_data[0]' (combined) should fit
36 * into a single 64-byte cache line.
40 struct clock_read_data read_data[2];
44 u64 (*actual_read_sched_clock)(void);
47 static struct hrtimer sched_clock_timer;
48 static int irqtime = -1;
50 core_param(irqtime, irqtime, int, 0400);
52 static u64 notrace jiffy_sched_clock_read(void)
55 * We don't need to use get_jiffies_64 on 32-bit arches here
56 * because we register with BITS_PER_LONG
58 return (u64)(jiffies - INITIAL_JIFFIES);
61 static struct clock_data cd ____cacheline_aligned = {
62 .read_data[0] = { .mult = NSEC_PER_SEC / HZ,
63 .read_sched_clock = jiffy_sched_clock_read, },
64 .actual_read_sched_clock = jiffy_sched_clock_read,
67 static inline u64 notrace cyc_to_ns(u64 cyc, u32 mult, u32 shift)
69 return (cyc * mult) >> shift;
72 notrace struct clock_read_data *sched_clock_read_begin(unsigned int *seq)
74 *seq = raw_read_seqcount_latch(&cd.seq);
75 return cd.read_data + (*seq & 1);
78 notrace int sched_clock_read_retry(unsigned int seq)
80 return read_seqcount_latch_retry(&cd.seq, seq);
83 unsigned long long notrace sched_clock(void)
87 struct clock_read_data *rd;
90 rd = sched_clock_read_begin(&seq);
92 cyc = (rd->read_sched_clock() - rd->epoch_cyc) &
94 res = rd->epoch_ns + cyc_to_ns(cyc, rd->mult, rd->shift);
95 } while (sched_clock_read_retry(seq));
101 * Updating the data required to read the clock.
103 * sched_clock() will never observe mis-matched data even if called from
104 * an NMI. We do this by maintaining an odd/even copy of the data and
105 * steering sched_clock() to one or the other using a sequence counter.
106 * In order to preserve the data cache profile of sched_clock() as much
107 * as possible the system reverts back to the even copy when the update
108 * completes; the odd copy is used *only* during an update.
110 static void update_clock_read_data(struct clock_read_data *rd)
112 /* update the backup (odd) copy with the new data */
113 cd.read_data[1] = *rd;
115 /* steer readers towards the odd copy */
116 raw_write_seqcount_latch(&cd.seq);
118 /* now its safe for us to update the normal (even) copy */
119 cd.read_data[0] = *rd;
121 /* switch readers back to the even copy */
122 raw_write_seqcount_latch(&cd.seq);
126 * Atomically update the sched_clock() epoch.
128 static void update_sched_clock(void)
132 struct clock_read_data rd;
134 rd = cd.read_data[0];
136 cyc = cd.actual_read_sched_clock();
137 ns = rd.epoch_ns + cyc_to_ns((cyc - rd.epoch_cyc) & rd.sched_clock_mask, rd.mult, rd.shift);
142 update_clock_read_data(&rd);
145 static enum hrtimer_restart sched_clock_poll(struct hrtimer *hrt)
147 update_sched_clock();
148 hrtimer_forward_now(hrt, cd.wrap_kt);
150 return HRTIMER_RESTART;
154 sched_clock_register(u64 (*read)(void), int bits, unsigned long rate)
156 u64 res, wrap, new_mask, new_epoch, cyc, ns;
157 u32 new_mult, new_shift;
158 unsigned long r, flags;
160 struct clock_read_data rd;
165 /* Cannot register a sched_clock with interrupts on */
166 local_irq_save(flags);
168 /* Calculate the mult/shift to convert counter ticks to ns. */
169 clocks_calc_mult_shift(&new_mult, &new_shift, rate, NSEC_PER_SEC, 3600);
171 new_mask = CLOCKSOURCE_MASK(bits);
174 /* Calculate how many nanosecs until we risk wrapping */
175 wrap = clocks_calc_max_nsecs(new_mult, new_shift, 0, new_mask, NULL);
176 cd.wrap_kt = ns_to_ktime(wrap);
178 rd = cd.read_data[0];
180 /* Update epoch for new counter and update 'epoch_ns' from old counter*/
182 cyc = cd.actual_read_sched_clock();
183 ns = rd.epoch_ns + cyc_to_ns((cyc - rd.epoch_cyc) & rd.sched_clock_mask, rd.mult, rd.shift);
184 cd.actual_read_sched_clock = read;
186 rd.read_sched_clock = read;
187 rd.sched_clock_mask = new_mask;
189 rd.shift = new_shift;
190 rd.epoch_cyc = new_epoch;
193 update_clock_read_data(&rd);
195 if (sched_clock_timer.function != NULL) {
196 /* update timeout for clock wrap */
197 hrtimer_start(&sched_clock_timer, cd.wrap_kt,
198 HRTIMER_MODE_REL_HARD);
203 r = DIV_ROUND_CLOSEST(r, 1000000);
205 } else if (r >= 4000) {
206 r = DIV_ROUND_CLOSEST(r, 1000);
212 /* Calculate the ns resolution of this counter */
213 res = cyc_to_ns(1ULL, new_mult, new_shift);
215 pr_info("sched_clock: %u bits at %lu%cHz, resolution %lluns, wraps every %lluns\n",
216 bits, r, r_unit, res, wrap);
218 /* Enable IRQ time accounting if we have a fast enough sched_clock() */
219 if (irqtime > 0 || (irqtime == -1 && rate >= 1000000))
220 enable_sched_clock_irqtime();
222 local_irq_restore(flags);
224 pr_debug("Registered %pS as sched_clock source\n", read);
227 void __init generic_sched_clock_init(void)
230 * If no sched_clock() function has been provided at that point,
231 * make it the final one.
233 if (cd.actual_read_sched_clock == jiffy_sched_clock_read)
234 sched_clock_register(jiffy_sched_clock_read, BITS_PER_LONG, HZ);
236 update_sched_clock();
239 * Start the timer to keep sched_clock() properly updated and
240 * sets the initial epoch.
242 hrtimer_init(&sched_clock_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
243 sched_clock_timer.function = sched_clock_poll;
244 hrtimer_start(&sched_clock_timer, cd.wrap_kt, HRTIMER_MODE_REL_HARD);
248 * Clock read function for use when the clock is suspended.
250 * This function makes it appear to sched_clock() as if the clock
251 * stopped counting at its last update.
253 * This function must only be called from the critical
254 * section in sched_clock(). It relies on the read_seqcount_retry()
255 * at the end of the critical section to be sure we observe the
256 * correct copy of 'epoch_cyc'.
258 static u64 notrace suspended_sched_clock_read(void)
260 unsigned int seq = raw_read_seqcount_latch(&cd.seq);
262 return cd.read_data[seq & 1].epoch_cyc;
265 int sched_clock_suspend(void)
267 struct clock_read_data *rd = &cd.read_data[0];
269 update_sched_clock();
270 hrtimer_cancel(&sched_clock_timer);
271 rd->read_sched_clock = suspended_sched_clock_read;
276 void sched_clock_resume(void)
278 struct clock_read_data *rd = &cd.read_data[0];
280 rd->epoch_cyc = cd.actual_read_sched_clock();
281 hrtimer_start(&sched_clock_timer, cd.wrap_kt, HRTIMER_MODE_REL_HARD);
282 rd->read_sched_clock = cd.actual_read_sched_clock;
285 static struct syscore_ops sched_clock_ops = {
286 .suspend = sched_clock_suspend,
287 .resume = sched_clock_resume,
290 static int __init sched_clock_syscore_init(void)
292 register_syscore_ops(&sched_clock_ops);
296 device_initcall(sched_clock_syscore_init);