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 __always_inline u64
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
= 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 static __always_inline
unsigned long long __sched_clock(void)
85 struct clock_read_data
*rd
;
90 seq
= raw_read_seqcount_latch(&cd
.seq
);
91 rd
= cd
.read_data
+ (seq
& 1);
93 cyc
= (rd
->read_sched_clock() - rd
->epoch_cyc
) &
95 res
= rd
->epoch_ns
+ cyc_to_ns(cyc
, rd
->mult
, rd
->shift
);
96 } while (raw_read_seqcount_latch_retry(&cd
.seq
, seq
));
101 unsigned long long noinstr
sched_clock_noinstr(void)
103 return __sched_clock();
106 unsigned long long notrace
sched_clock(void)
108 unsigned long long ns
;
109 preempt_disable_notrace();
111 * All of __sched_clock() is a seqcount_latch reader critical section,
112 * but relies on the raw helpers which are uninstrumented. For KCSAN,
113 * mark all accesses in __sched_clock() as atomic.
115 kcsan_nestable_atomic_begin();
116 ns
= __sched_clock();
117 kcsan_nestable_atomic_end();
118 preempt_enable_notrace();
123 * Updating the data required to read the clock.
125 * sched_clock() will never observe mis-matched data even if called from
126 * an NMI. We do this by maintaining an odd/even copy of the data and
127 * steering sched_clock() to one or the other using a sequence counter.
128 * In order to preserve the data cache profile of sched_clock() as much
129 * as possible the system reverts back to the even copy when the update
130 * completes; the odd copy is used *only* during an update.
132 static void update_clock_read_data(struct clock_read_data
*rd
)
134 /* steer readers towards the odd copy */
135 write_seqcount_latch_begin(&cd
.seq
);
137 /* now its safe for us to update the normal (even) copy */
138 cd
.read_data
[0] = *rd
;
140 /* switch readers back to the even copy */
141 write_seqcount_latch(&cd
.seq
);
143 /* update the backup (odd) copy with the new data */
144 cd
.read_data
[1] = *rd
;
146 write_seqcount_latch_end(&cd
.seq
);
150 * Atomically update the sched_clock() epoch.
152 static void update_sched_clock(void)
156 struct clock_read_data rd
;
158 rd
= cd
.read_data
[0];
160 cyc
= cd
.actual_read_sched_clock();
161 ns
= rd
.epoch_ns
+ cyc_to_ns((cyc
- rd
.epoch_cyc
) & rd
.sched_clock_mask
, rd
.mult
, rd
.shift
);
166 update_clock_read_data(&rd
);
169 static enum hrtimer_restart
sched_clock_poll(struct hrtimer
*hrt
)
171 update_sched_clock();
172 hrtimer_forward_now(hrt
, cd
.wrap_kt
);
174 return HRTIMER_RESTART
;
178 sched_clock_register(u64 (*read
)(void), int bits
, unsigned long rate
)
180 u64 res
, wrap
, new_mask
, new_epoch
, cyc
, ns
;
181 u32 new_mult
, new_shift
;
182 unsigned long r
, flags
;
184 struct clock_read_data rd
;
189 /* Cannot register a sched_clock with interrupts on */
190 local_irq_save(flags
);
192 /* Calculate the mult/shift to convert counter ticks to ns. */
193 clocks_calc_mult_shift(&new_mult
, &new_shift
, rate
, NSEC_PER_SEC
, 3600);
195 new_mask
= CLOCKSOURCE_MASK(bits
);
198 /* Calculate how many nanosecs until we risk wrapping */
199 wrap
= clocks_calc_max_nsecs(new_mult
, new_shift
, 0, new_mask
, NULL
);
200 cd
.wrap_kt
= ns_to_ktime(wrap
);
202 rd
= cd
.read_data
[0];
204 /* Update epoch for new counter and update 'epoch_ns' from old counter*/
206 cyc
= cd
.actual_read_sched_clock();
207 ns
= rd
.epoch_ns
+ cyc_to_ns((cyc
- rd
.epoch_cyc
) & rd
.sched_clock_mask
, rd
.mult
, rd
.shift
);
208 cd
.actual_read_sched_clock
= read
;
210 rd
.read_sched_clock
= read
;
211 rd
.sched_clock_mask
= new_mask
;
213 rd
.shift
= new_shift
;
214 rd
.epoch_cyc
= new_epoch
;
217 update_clock_read_data(&rd
);
219 if (sched_clock_timer
.function
!= NULL
) {
220 /* update timeout for clock wrap */
221 hrtimer_start(&sched_clock_timer
, cd
.wrap_kt
,
222 HRTIMER_MODE_REL_HARD
);
227 r
= DIV_ROUND_CLOSEST(r
, 1000000);
229 } else if (r
>= 4000) {
230 r
= DIV_ROUND_CLOSEST(r
, 1000);
236 /* Calculate the ns resolution of this counter */
237 res
= cyc_to_ns(1ULL, new_mult
, new_shift
);
239 pr_info("sched_clock: %u bits at %lu%cHz, resolution %lluns, wraps every %lluns\n",
240 bits
, r
, r_unit
, res
, wrap
);
242 /* Enable IRQ time accounting if we have a fast enough sched_clock() */
243 if (irqtime
> 0 || (irqtime
== -1 && rate
>= 1000000))
244 enable_sched_clock_irqtime();
246 local_irq_restore(flags
);
248 pr_debug("Registered %pS as sched_clock source\n", read
);
251 void __init
generic_sched_clock_init(void)
254 * If no sched_clock() function has been provided at that point,
255 * make it the final one.
257 if (cd
.actual_read_sched_clock
== jiffy_sched_clock_read
)
258 sched_clock_register(jiffy_sched_clock_read
, BITS_PER_LONG
, HZ
);
260 update_sched_clock();
263 * Start the timer to keep sched_clock() properly updated and
264 * sets the initial epoch.
266 hrtimer_init(&sched_clock_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL_HARD
);
267 sched_clock_timer
.function
= sched_clock_poll
;
268 hrtimer_start(&sched_clock_timer
, cd
.wrap_kt
, HRTIMER_MODE_REL_HARD
);
272 * Clock read function for use when the clock is suspended.
274 * This function makes it appear to sched_clock() as if the clock
275 * stopped counting at its last update.
277 * This function must only be called from the critical
278 * section in sched_clock(). It relies on the read_seqcount_retry()
279 * at the end of the critical section to be sure we observe the
280 * correct copy of 'epoch_cyc'.
282 static u64 notrace
suspended_sched_clock_read(void)
284 unsigned int seq
= read_seqcount_latch(&cd
.seq
);
286 return cd
.read_data
[seq
& 1].epoch_cyc
;
289 int sched_clock_suspend(void)
291 struct clock_read_data
*rd
= &cd
.read_data
[0];
293 update_sched_clock();
294 hrtimer_cancel(&sched_clock_timer
);
295 rd
->read_sched_clock
= suspended_sched_clock_read
;
300 void sched_clock_resume(void)
302 struct clock_read_data
*rd
= &cd
.read_data
[0];
304 rd
->epoch_cyc
= cd
.actual_read_sched_clock();
305 hrtimer_start(&sched_clock_timer
, cd
.wrap_kt
, HRTIMER_MODE_REL_HARD
);
306 rd
->read_sched_clock
= cd
.actual_read_sched_clock
;
309 static struct syscore_ops sched_clock_ops
= {
310 .suspend
= sched_clock_suspend
,
311 .resume
= sched_clock_resume
,
314 static int __init
sched_clock_syscore_init(void)
316 register_syscore_ops(&sched_clock_ops
);
320 device_initcall(sched_clock_syscore_init
);