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/syscore_ops.h>
17#include <linux/hrtimer.h>
18#include <linux/sched_clock.h>
19#include <linux/seqlock.h>
20#include <linux/bitops.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 long 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	r = rate;
209	if (r >= 4000000) {
210		r /= 1000000;
211		r_unit = 'M';
212	} else {
213		if (r >= 1000) {
214			r /= 1000;
215			r_unit = 'k';
216		} else {
217			r_unit = ' ';
218		}
219	}
220
221	/* Calculate the ns resolution of this counter */
222	res = cyc_to_ns(1ULL, new_mult, new_shift);
223
224	pr_info("sched_clock: %u bits at %lu%cHz, resolution %lluns, wraps every %lluns\n",
225		bits, r, r_unit, res, wrap);
226
227	/* Enable IRQ time accounting if we have a fast enough sched_clock() */
228	if (irqtime > 0 || (irqtime == -1 && rate >= 1000000))
229		enable_sched_clock_irqtime();
230
231	pr_debug("Registered %pF as sched_clock source\n", read);
232}
233
234void __init sched_clock_postinit(void)
235{
236	/*
237	 * If no sched_clock() function has been provided at that point,
238	 * make it the final one one.
239	 */
240	if (cd.actual_read_sched_clock == jiffy_sched_clock_read)
241		sched_clock_register(jiffy_sched_clock_read, BITS_PER_LONG, HZ);
242
243	update_sched_clock();
244
245	/*
246	 * Start the timer to keep sched_clock() properly updated and
247	 * sets the initial epoch.
248	 */
249	hrtimer_init(&sched_clock_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
250	sched_clock_timer.function = sched_clock_poll;
251	hrtimer_start(&sched_clock_timer, cd.wrap_kt, HRTIMER_MODE_REL);
252}
253
254/*
255 * Clock read function for use when the clock is suspended.
256 *
257 * This function makes it appear to sched_clock() as if the clock
258 * stopped counting at its last update.
259 *
260 * This function must only be called from the critical
261 * section in sched_clock(). It relies on the read_seqcount_retry()
262 * at the end of the critical section to be sure we observe the
263 * correct copy of 'epoch_cyc'.
264 */
265static u64 notrace suspended_sched_clock_read(void)
266{
267	unsigned long seq = raw_read_seqcount(&cd.seq);
268
269	return cd.read_data[seq & 1].epoch_cyc;
270}
271
272static int sched_clock_suspend(void)
273{
274	struct clock_read_data *rd = &cd.read_data[0];
275
276	update_sched_clock();
277	hrtimer_cancel(&sched_clock_timer);
278	rd->read_sched_clock = suspended_sched_clock_read;
279
280	return 0;
281}
282
283static void sched_clock_resume(void)
284{
285	struct clock_read_data *rd = &cd.read_data[0];
286
287	rd->epoch_cyc = cd.actual_read_sched_clock();
288	hrtimer_start(&sched_clock_timer, cd.wrap_kt, HRTIMER_MODE_REL);
289	rd->read_sched_clock = cd.actual_read_sched_clock;
290}
291
292static struct syscore_ops sched_clock_ops = {
293	.suspend	= sched_clock_suspend,
294	.resume		= sched_clock_resume,
295};
296
297static int __init sched_clock_syscore_init(void)
298{
299	register_syscore_ops(&sched_clock_ops);
300
301	return 0;
302}
303device_initcall(sched_clock_syscore_init);
304