Loading...
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/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>
20
21#include "timekeeping.h"
22
23/**
24 * struct clock_data - all data needed for sched_clock() (including
25 * registration of a new clock source)
26 *
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.
33 *
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.
37 */
38struct clock_data {
39 seqcount_latch_t seq;
40 struct clock_read_data read_data[2];
41 ktime_t wrap_kt;
42 unsigned long rate;
43
44 u64 (*actual_read_sched_clock)(void);
45};
46
47static struct hrtimer sched_clock_timer;
48static int irqtime = -1;
49
50core_param(irqtime, irqtime, int, 0400);
51
52static u64 notrace jiffy_sched_clock_read(void)
53{
54 /*
55 * We don't need to use get_jiffies_64 on 32-bit arches here
56 * because we register with BITS_PER_LONG
57 */
58 return (u64)(jiffies - INITIAL_JIFFIES);
59}
60
61static 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,
65};
66
67static inline u64 notrace cyc_to_ns(u64 cyc, u32 mult, u32 shift)
68{
69 return (cyc * mult) >> shift;
70}
71
72notrace struct clock_read_data *sched_clock_read_begin(unsigned int *seq)
73{
74 *seq = raw_read_seqcount_latch(&cd.seq);
75 return cd.read_data + (*seq & 1);
76}
77
78notrace int sched_clock_read_retry(unsigned int seq)
79{
80 return read_seqcount_latch_retry(&cd.seq, seq);
81}
82
83unsigned long long notrace sched_clock(void)
84{
85 u64 cyc, res;
86 unsigned int seq;
87 struct clock_read_data *rd;
88
89 do {
90 rd = sched_clock_read_begin(&seq);
91
92 cyc = (rd->read_sched_clock() - rd->epoch_cyc) &
93 rd->sched_clock_mask;
94 res = rd->epoch_ns + cyc_to_ns(cyc, rd->mult, rd->shift);
95 } while (sched_clock_read_retry(seq));
96
97 return res;
98}
99
100/*
101 * Updating the data required to read the clock.
102 *
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.
109 */
110static void update_clock_read_data(struct clock_read_data *rd)
111{
112 /* update the backup (odd) copy with the new data */
113 cd.read_data[1] = *rd;
114
115 /* steer readers towards the odd copy */
116 raw_write_seqcount_latch(&cd.seq);
117
118 /* now its safe for us to update the normal (even) copy */
119 cd.read_data[0] = *rd;
120
121 /* switch readers back to the even copy */
122 raw_write_seqcount_latch(&cd.seq);
123}
124
125/*
126 * Atomically update the sched_clock() epoch.
127 */
128static void update_sched_clock(void)
129{
130 u64 cyc;
131 u64 ns;
132 struct clock_read_data rd;
133
134 rd = cd.read_data[0];
135
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);
138
139 rd.epoch_ns = ns;
140 rd.epoch_cyc = cyc;
141
142 update_clock_read_data(&rd);
143}
144
145static enum hrtimer_restart sched_clock_poll(struct hrtimer *hrt)
146{
147 update_sched_clock();
148 hrtimer_forward_now(hrt, cd.wrap_kt);
149
150 return HRTIMER_RESTART;
151}
152
153void __init
154sched_clock_register(u64 (*read)(void), int bits, unsigned long rate)
155{
156 u64 res, wrap, new_mask, new_epoch, cyc, ns;
157 u32 new_mult, new_shift;
158 unsigned long r, flags;
159 char r_unit;
160 struct clock_read_data rd;
161
162 if (cd.rate > rate)
163 return;
164
165 /* Cannot register a sched_clock with interrupts on */
166 local_irq_save(flags);
167
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);
170
171 new_mask = CLOCKSOURCE_MASK(bits);
172 cd.rate = rate;
173
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);
177
178 rd = cd.read_data[0];
179
180 /* Update epoch for new counter and update 'epoch_ns' from old counter*/
181 new_epoch = read();
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;
185
186 rd.read_sched_clock = read;
187 rd.sched_clock_mask = new_mask;
188 rd.mult = new_mult;
189 rd.shift = new_shift;
190 rd.epoch_cyc = new_epoch;
191 rd.epoch_ns = ns;
192
193 update_clock_read_data(&rd);
194
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);
199 }
200
201 r = rate;
202 if (r >= 4000000) {
203 r = DIV_ROUND_CLOSEST(r, 1000000);
204 r_unit = 'M';
205 } else if (r >= 4000) {
206 r = DIV_ROUND_CLOSEST(r, 1000);
207 r_unit = 'k';
208 } else {
209 r_unit = ' ';
210 }
211
212 /* Calculate the ns resolution of this counter */
213 res = cyc_to_ns(1ULL, new_mult, new_shift);
214
215 pr_info("sched_clock: %u bits at %lu%cHz, resolution %lluns, wraps every %lluns\n",
216 bits, r, r_unit, res, wrap);
217
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();
221
222 local_irq_restore(flags);
223
224 pr_debug("Registered %pS as sched_clock source\n", read);
225}
226
227void __init generic_sched_clock_init(void)
228{
229 /*
230 * If no sched_clock() function has been provided at that point,
231 * make it the final one.
232 */
233 if (cd.actual_read_sched_clock == jiffy_sched_clock_read)
234 sched_clock_register(jiffy_sched_clock_read, BITS_PER_LONG, HZ);
235
236 update_sched_clock();
237
238 /*
239 * Start the timer to keep sched_clock() properly updated and
240 * sets the initial epoch.
241 */
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);
245}
246
247/*
248 * Clock read function for use when the clock is suspended.
249 *
250 * This function makes it appear to sched_clock() as if the clock
251 * stopped counting at its last update.
252 *
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'.
257 */
258static u64 notrace suspended_sched_clock_read(void)
259{
260 unsigned int seq = raw_read_seqcount_latch(&cd.seq);
261
262 return cd.read_data[seq & 1].epoch_cyc;
263}
264
265int sched_clock_suspend(void)
266{
267 struct clock_read_data *rd = &cd.read_data[0];
268
269 update_sched_clock();
270 hrtimer_cancel(&sched_clock_timer);
271 rd->read_sched_clock = suspended_sched_clock_read;
272
273 return 0;
274}
275
276void sched_clock_resume(void)
277{
278 struct clock_read_data *rd = &cd.read_data[0];
279
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;
283}
284
285static struct syscore_ops sched_clock_ops = {
286 .suspend = sched_clock_suspend,
287 .resume = sched_clock_resume,
288};
289
290static int __init sched_clock_syscore_init(void)
291{
292 register_syscore_ops(&sched_clock_ops);
293
294 return 0;
295}
296device_initcall(sched_clock_syscore_init);
1/*
2 * sched_clock.c: Generic sched_clock() support, to extend low level
3 * hardware time counters to full 64-bit ns values.
4 *
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License version 2 as
7 * published by the Free Software Foundation.
8 */
9#include <linux/clocksource.h>
10#include <linux/init.h>
11#include <linux/jiffies.h>
12#include <linux/ktime.h>
13#include <linux/kernel.h>
14#include <linux/moduleparam.h>
15#include <linux/sched.h>
16#include <linux/sched/clock.h>
17#include <linux/syscore_ops.h>
18#include <linux/hrtimer.h>
19#include <linux/sched_clock.h>
20#include <linux/seqlock.h>
21#include <linux/bitops.h>
22
23/**
24 * struct clock_read_data - data required to read from sched_clock()
25 *
26 * @epoch_ns: sched_clock() value at last update
27 * @epoch_cyc: Clock cycle value at last update.
28 * @sched_clock_mask: Bitmask for two's complement subtraction of non 64bit
29 * clocks.
30 * @read_sched_clock: Current clock source (or dummy source when suspended).
31 * @mult: Multipler for scaled math conversion.
32 * @shift: Shift value for scaled math conversion.
33 *
34 * Care must be taken when updating this structure; it is read by
35 * some very hot code paths. It occupies <=40 bytes and, when combined
36 * with the seqcount used to synchronize access, comfortably fits into
37 * a 64 byte cache line.
38 */
39struct clock_read_data {
40 u64 epoch_ns;
41 u64 epoch_cyc;
42 u64 sched_clock_mask;
43 u64 (*read_sched_clock)(void);
44 u32 mult;
45 u32 shift;
46};
47
48/**
49 * struct clock_data - all data needed for sched_clock() (including
50 * registration of a new clock source)
51 *
52 * @seq: Sequence counter for protecting updates. The lowest
53 * bit is the index for @read_data.
54 * @read_data: Data required to read from sched_clock.
55 * @wrap_kt: Duration for which clock can run before wrapping.
56 * @rate: Tick rate of the registered clock.
57 * @actual_read_sched_clock: Registered hardware level clock read function.
58 *
59 * The ordering of this structure has been chosen to optimize cache
60 * performance. In particular 'seq' and 'read_data[0]' (combined) should fit
61 * into a single 64-byte cache line.
62 */
63struct clock_data {
64 seqcount_t seq;
65 struct clock_read_data read_data[2];
66 ktime_t wrap_kt;
67 unsigned long rate;
68
69 u64 (*actual_read_sched_clock)(void);
70};
71
72static struct hrtimer sched_clock_timer;
73static int irqtime = -1;
74
75core_param(irqtime, irqtime, int, 0400);
76
77static u64 notrace jiffy_sched_clock_read(void)
78{
79 /*
80 * We don't need to use get_jiffies_64 on 32-bit arches here
81 * because we register with BITS_PER_LONG
82 */
83 return (u64)(jiffies - INITIAL_JIFFIES);
84}
85
86static struct clock_data cd ____cacheline_aligned = {
87 .read_data[0] = { .mult = NSEC_PER_SEC / HZ,
88 .read_sched_clock = jiffy_sched_clock_read, },
89 .actual_read_sched_clock = jiffy_sched_clock_read,
90};
91
92static inline u64 notrace cyc_to_ns(u64 cyc, u32 mult, u32 shift)
93{
94 return (cyc * mult) >> shift;
95}
96
97unsigned long long notrace sched_clock(void)
98{
99 u64 cyc, res;
100 unsigned long seq;
101 struct clock_read_data *rd;
102
103 do {
104 seq = raw_read_seqcount(&cd.seq);
105 rd = cd.read_data + (seq & 1);
106
107 cyc = (rd->read_sched_clock() - rd->epoch_cyc) &
108 rd->sched_clock_mask;
109 res = rd->epoch_ns + cyc_to_ns(cyc, rd->mult, rd->shift);
110 } while (read_seqcount_retry(&cd.seq, seq));
111
112 return res;
113}
114
115/*
116 * Updating the data required to read the clock.
117 *
118 * sched_clock() will never observe mis-matched data even if called from
119 * an NMI. We do this by maintaining an odd/even copy of the data and
120 * steering sched_clock() to one or the other using a sequence counter.
121 * In order to preserve the data cache profile of sched_clock() as much
122 * as possible the system reverts back to the even copy when the update
123 * completes; the odd copy is used *only* during an update.
124 */
125static void update_clock_read_data(struct clock_read_data *rd)
126{
127 /* update the backup (odd) copy with the new data */
128 cd.read_data[1] = *rd;
129
130 /* steer readers towards the odd copy */
131 raw_write_seqcount_latch(&cd.seq);
132
133 /* now its safe for us to update the normal (even) copy */
134 cd.read_data[0] = *rd;
135
136 /* switch readers back to the even copy */
137 raw_write_seqcount_latch(&cd.seq);
138}
139
140/*
141 * Atomically update the sched_clock() epoch.
142 */
143static void update_sched_clock(void)
144{
145 u64 cyc;
146 u64 ns;
147 struct clock_read_data rd;
148
149 rd = cd.read_data[0];
150
151 cyc = cd.actual_read_sched_clock();
152 ns = rd.epoch_ns + cyc_to_ns((cyc - rd.epoch_cyc) & rd.sched_clock_mask, rd.mult, rd.shift);
153
154 rd.epoch_ns = ns;
155 rd.epoch_cyc = cyc;
156
157 update_clock_read_data(&rd);
158}
159
160static enum hrtimer_restart sched_clock_poll(struct hrtimer *hrt)
161{
162 update_sched_clock();
163 hrtimer_forward_now(hrt, cd.wrap_kt);
164
165 return HRTIMER_RESTART;
166}
167
168void __init
169sched_clock_register(u64 (*read)(void), int bits, unsigned long rate)
170{
171 u64 res, wrap, new_mask, new_epoch, cyc, ns;
172 u32 new_mult, new_shift;
173 unsigned long r;
174 char r_unit;
175 struct clock_read_data rd;
176
177 if (cd.rate > rate)
178 return;
179
180 WARN_ON(!irqs_disabled());
181
182 /* Calculate the mult/shift to convert counter ticks to ns. */
183 clocks_calc_mult_shift(&new_mult, &new_shift, rate, NSEC_PER_SEC, 3600);
184
185 new_mask = CLOCKSOURCE_MASK(bits);
186 cd.rate = rate;
187
188 /* Calculate how many nanosecs until we risk wrapping */
189 wrap = clocks_calc_max_nsecs(new_mult, new_shift, 0, new_mask, NULL);
190 cd.wrap_kt = ns_to_ktime(wrap);
191
192 rd = cd.read_data[0];
193
194 /* Update epoch for new counter and update 'epoch_ns' from old counter*/
195 new_epoch = read();
196 cyc = cd.actual_read_sched_clock();
197 ns = rd.epoch_ns + cyc_to_ns((cyc - rd.epoch_cyc) & rd.sched_clock_mask, rd.mult, rd.shift);
198 cd.actual_read_sched_clock = read;
199
200 rd.read_sched_clock = read;
201 rd.sched_clock_mask = new_mask;
202 rd.mult = new_mult;
203 rd.shift = new_shift;
204 rd.epoch_cyc = new_epoch;
205 rd.epoch_ns = ns;
206
207 update_clock_read_data(&rd);
208
209 if (sched_clock_timer.function != NULL) {
210 /* update timeout for clock wrap */
211 hrtimer_start(&sched_clock_timer, cd.wrap_kt, HRTIMER_MODE_REL);
212 }
213
214 r = rate;
215 if (r >= 4000000) {
216 r /= 1000000;
217 r_unit = 'M';
218 } else {
219 if (r >= 1000) {
220 r /= 1000;
221 r_unit = 'k';
222 } else {
223 r_unit = ' ';
224 }
225 }
226
227 /* Calculate the ns resolution of this counter */
228 res = cyc_to_ns(1ULL, new_mult, new_shift);
229
230 pr_info("sched_clock: %u bits at %lu%cHz, resolution %lluns, wraps every %lluns\n",
231 bits, r, r_unit, res, wrap);
232
233 /* Enable IRQ time accounting if we have a fast enough sched_clock() */
234 if (irqtime > 0 || (irqtime == -1 && rate >= 1000000))
235 enable_sched_clock_irqtime();
236
237 pr_debug("Registered %pF as sched_clock source\n", read);
238}
239
240void __init sched_clock_postinit(void)
241{
242 /*
243 * If no sched_clock() function has been provided at that point,
244 * make it the final one one.
245 */
246 if (cd.actual_read_sched_clock == jiffy_sched_clock_read)
247 sched_clock_register(jiffy_sched_clock_read, BITS_PER_LONG, HZ);
248
249 update_sched_clock();
250
251 /*
252 * Start the timer to keep sched_clock() properly updated and
253 * sets the initial epoch.
254 */
255 hrtimer_init(&sched_clock_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
256 sched_clock_timer.function = sched_clock_poll;
257 hrtimer_start(&sched_clock_timer, cd.wrap_kt, HRTIMER_MODE_REL);
258}
259
260/*
261 * Clock read function for use when the clock is suspended.
262 *
263 * This function makes it appear to sched_clock() as if the clock
264 * stopped counting at its last update.
265 *
266 * This function must only be called from the critical
267 * section in sched_clock(). It relies on the read_seqcount_retry()
268 * at the end of the critical section to be sure we observe the
269 * correct copy of 'epoch_cyc'.
270 */
271static u64 notrace suspended_sched_clock_read(void)
272{
273 unsigned long seq = raw_read_seqcount(&cd.seq);
274
275 return cd.read_data[seq & 1].epoch_cyc;
276}
277
278static int sched_clock_suspend(void)
279{
280 struct clock_read_data *rd = &cd.read_data[0];
281
282 update_sched_clock();
283 hrtimer_cancel(&sched_clock_timer);
284 rd->read_sched_clock = suspended_sched_clock_read;
285
286 return 0;
287}
288
289static void sched_clock_resume(void)
290{
291 struct clock_read_data *rd = &cd.read_data[0];
292
293 rd->epoch_cyc = cd.actual_read_sched_clock();
294 hrtimer_start(&sched_clock_timer, cd.wrap_kt, HRTIMER_MODE_REL);
295 rd->read_sched_clock = cd.actual_read_sched_clock;
296}
297
298static struct syscore_ops sched_clock_ops = {
299 .suspend = sched_clock_suspend,
300 .resume = sched_clock_resume,
301};
302
303static int __init sched_clock_syscore_init(void)
304{
305 register_syscore_ops(&sched_clock_ops);
306
307 return 0;
308}
309device_initcall(sched_clock_syscore_init);