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 #include <trace/hooks/epoch.h>
21
22 #include "timekeeping.h"
23
24 /**
25 * struct clock_data - all data needed for sched_clock() (including
26 * registration of a new clock source)
27 *
28 * @seq: Sequence counter for protecting updates. The lowest
29 * bit is the index for @read_data.
30 * @read_data: Data required to read from sched_clock.
31 * @wrap_kt: Duration for which clock can run before wrapping.
32 * @rate: Tick rate of the registered clock.
33 * @actual_read_sched_clock: Registered hardware level clock read function.
34 *
35 * The ordering of this structure has been chosen to optimize cache
36 * performance. In particular 'seq' and 'read_data[0]' (combined) should fit
37 * into a single 64-byte cache line.
38 */
39 struct clock_data {
40 seqcount_latch_t seq;
41 struct clock_read_data read_data[2];
42 ktime_t wrap_kt;
43 unsigned long rate;
44
45 u64 (*actual_read_sched_clock)(void);
46 };
47
48 static struct hrtimer sched_clock_timer;
49 static int irqtime = -1;
50
51 core_param(irqtime, irqtime, int, 0400);
52
jiffy_sched_clock_read(void)53 static u64 notrace jiffy_sched_clock_read(void)
54 {
55 /*
56 * We don't need to use get_jiffies_64 on 32-bit arches here
57 * because we register with BITS_PER_LONG
58 */
59 return (u64)(jiffies - INITIAL_JIFFIES);
60 }
61
62 static struct clock_data cd ____cacheline_aligned = {
63 .read_data[0] = { .mult = NSEC_PER_SEC / HZ,
64 .read_sched_clock = jiffy_sched_clock_read, },
65 .actual_read_sched_clock = jiffy_sched_clock_read,
66 };
67
cyc_to_ns(u64 cyc,u32 mult,u32 shift)68 static __always_inline u64 cyc_to_ns(u64 cyc, u32 mult, u32 shift)
69 {
70 return (cyc * mult) >> shift;
71 }
72
sched_clock_read_begin(unsigned int * seq)73 notrace struct clock_read_data *sched_clock_read_begin(unsigned int *seq)
74 {
75 *seq = raw_read_seqcount_latch(&cd.seq);
76 return cd.read_data + (*seq & 1);
77 }
78 EXPORT_SYMBOL_GPL(sched_clock_read_begin);
79
sched_clock_read_retry(unsigned int seq)80 notrace int sched_clock_read_retry(unsigned int seq)
81 {
82 return raw_read_seqcount_latch_retry(&cd.seq, seq);
83 }
84 EXPORT_SYMBOL_GPL(sched_clock_read_retry);
85
sched_clock_noinstr(void)86 unsigned long long noinstr sched_clock_noinstr(void)
87 {
88 struct clock_read_data *rd;
89 unsigned int seq;
90 u64 cyc, res;
91
92 do {
93 seq = raw_read_seqcount_latch(&cd.seq);
94 rd = cd.read_data + (seq & 1);
95
96 cyc = (rd->read_sched_clock() - rd->epoch_cyc) &
97 rd->sched_clock_mask;
98 res = rd->epoch_ns + cyc_to_ns(cyc, rd->mult, rd->shift);
99 } while (raw_read_seqcount_latch_retry(&cd.seq, seq));
100
101 return res;
102 }
103
sched_clock(void)104 unsigned long long notrace sched_clock(void)
105 {
106 unsigned long long ns;
107 preempt_disable_notrace();
108 ns = sched_clock_noinstr();
109 preempt_enable_notrace();
110 return ns;
111 }
112
113 /*
114 * Updating the data required to read the clock.
115 *
116 * sched_clock() will never observe mis-matched data even if called from
117 * an NMI. We do this by maintaining an odd/even copy of the data and
118 * steering sched_clock() to one or the other using a sequence counter.
119 * In order to preserve the data cache profile of sched_clock() as much
120 * as possible the system reverts back to the even copy when the update
121 * completes; the odd copy is used *only* during an update.
122 */
update_clock_read_data(struct clock_read_data * rd)123 static void update_clock_read_data(struct clock_read_data *rd)
124 {
125 /* update the backup (odd) copy with the new data */
126 cd.read_data[1] = *rd;
127
128 /* steer readers towards the odd copy */
129 raw_write_seqcount_latch(&cd.seq);
130
131 /* now its safe for us to update the normal (even) copy */
132 cd.read_data[0] = *rd;
133
134 /* switch readers back to the even copy */
135 raw_write_seqcount_latch(&cd.seq);
136 }
137
138 /*
139 * Atomically update the sched_clock() epoch.
140 */
update_sched_clock(void)141 static void update_sched_clock(void)
142 {
143 u64 cyc;
144 u64 ns;
145 struct clock_read_data rd;
146
147 rd = cd.read_data[0];
148
149 cyc = cd.actual_read_sched_clock();
150 ns = rd.epoch_ns + cyc_to_ns((cyc - rd.epoch_cyc) & rd.sched_clock_mask, rd.mult, rd.shift);
151
152 rd.epoch_ns = ns;
153 rd.epoch_cyc = cyc;
154
155 update_clock_read_data(&rd);
156 }
157
sched_clock_poll(struct hrtimer * hrt)158 static enum hrtimer_restart sched_clock_poll(struct hrtimer *hrt)
159 {
160 update_sched_clock();
161 hrtimer_forward_now(hrt, cd.wrap_kt);
162
163 return HRTIMER_RESTART;
164 }
165
166 void __init
sched_clock_register(u64 (* read)(void),int bits,unsigned long rate)167 sched_clock_register(u64 (*read)(void), int bits, unsigned long rate)
168 {
169 u64 res, wrap, new_mask, new_epoch, cyc, ns;
170 u32 new_mult, new_shift;
171 unsigned long r, flags;
172 char r_unit;
173 struct clock_read_data rd;
174
175 if (cd.rate > rate)
176 return;
177
178 /* Cannot register a sched_clock with interrupts on */
179 local_irq_save(flags);
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,
211 HRTIMER_MODE_REL_HARD);
212 }
213
214 r = rate;
215 if (r >= 4000000) {
216 r = DIV_ROUND_CLOSEST(r, 1000000);
217 r_unit = 'M';
218 } else if (r >= 4000) {
219 r = DIV_ROUND_CLOSEST(r, 1000);
220 r_unit = 'k';
221 } else {
222 r_unit = ' ';
223 }
224
225 /* Calculate the ns resolution of this counter */
226 res = cyc_to_ns(1ULL, new_mult, new_shift);
227
228 pr_info("sched_clock: %u bits at %lu%cHz, resolution %lluns, wraps every %lluns\n",
229 bits, r, r_unit, res, wrap);
230
231 /* Enable IRQ time accounting if we have a fast enough sched_clock() */
232 if (irqtime > 0 || (irqtime == -1 && rate >= 1000000))
233 enable_sched_clock_irqtime();
234
235 local_irq_restore(flags);
236
237 pr_debug("Registered %pS as sched_clock source\n", read);
238 }
239
generic_sched_clock_init(void)240 void __init generic_sched_clock_init(void)
241 {
242 /*
243 * If no sched_clock() function has been provided at that point,
244 * make it the final 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_HARD);
256 sched_clock_timer.function = sched_clock_poll;
257 hrtimer_start(&sched_clock_timer, cd.wrap_kt, HRTIMER_MODE_REL_HARD);
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 */
suspended_sched_clock_read(void)271 static u64 notrace suspended_sched_clock_read(void)
272 {
273 unsigned int seq = raw_read_seqcount_latch(&cd.seq);
274
275 return cd.read_data[seq & 1].epoch_cyc;
276 }
277
sched_clock_suspend(void)278 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 trace_android_vh_show_suspend_epoch_val(rd->epoch_ns, rd->epoch_cyc);
286
287 return 0;
288 }
289
sched_clock_resume(void)290 void sched_clock_resume(void)
291 {
292 struct clock_read_data *rd = &cd.read_data[0];
293
294 rd->epoch_cyc = cd.actual_read_sched_clock();
295 hrtimer_start(&sched_clock_timer, cd.wrap_kt, HRTIMER_MODE_REL_HARD);
296 rd->read_sched_clock = cd.actual_read_sched_clock;
297 trace_android_vh_show_resume_epoch_val(rd->epoch_cyc);
298 }
299
300 static struct syscore_ops sched_clock_ops = {
301 .suspend = sched_clock_suspend,
302 .resume = sched_clock_resume,
303 };
304
sched_clock_syscore_init(void)305 static int __init sched_clock_syscore_init(void)
306 {
307 register_syscore_ops(&sched_clock_ops);
308
309 return 0;
310 }
311 device_initcall(sched_clock_syscore_init);
312