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