aboutsummaryrefslogtreecommitdiffstats
path: root/kernel/sched_clock.c
diff options
context:
space:
mode:
authorPeter Zijlstra <a.p.zijlstra@chello.nl>2008-05-03 12:29:28 -0400
committerIngo Molnar <mingo@elte.hu>2008-05-05 17:56:18 -0400
commit3e51f33fcc7f55e6df25d15b55ed10c8b4da84cd (patch)
tree3752f9ea8e014ec40e95a1b197b0a3d18e1056a8 /kernel/sched_clock.c
parenta5574cf65b5f03ce9ade3918764fe22e5e2371e3 (diff)
sched: add optional support for CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
this replaces the rq->clock stuff (and possibly cpu_clock()). - architectures that have an 'imperfect' hardware clock can set CONFIG_HAVE_UNSTABLE_SCHED_CLOCK - the 'jiffie' window might be superfulous when we update tick_gtod before the __update_sched_clock() call in sched_clock_tick() - cpu_clock() might be implemented as: sched_clock_cpu(smp_processor_id()) if the accuracy proves good enough - how far can TSC drift in a single jiffie when considering the filtering and idle hooks? [ mingo@elte.hu: various fixes and cleanups ] Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
Diffstat (limited to 'kernel/sched_clock.c')
-rw-r--r--kernel/sched_clock.c236
1 files changed, 236 insertions, 0 deletions
diff --git a/kernel/sched_clock.c b/kernel/sched_clock.c
new file mode 100644
index 000000000000..9c597e37f7de
--- /dev/null
+++ b/kernel/sched_clock.c
@@ -0,0 +1,236 @@
1/*
2 * sched_clock for unstable cpu clocks
3 *
4 * Copyright (C) 2008 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
5 *
6 * Based on code by:
7 * Ingo Molnar <mingo@redhat.com>
8 * Guillaume Chazarain <guichaz@gmail.com>
9 *
10 * Create a semi stable clock from a mixture of other events, including:
11 * - gtod
12 * - jiffies
13 * - sched_clock()
14 * - explicit idle events
15 *
16 * We use gtod as base and the unstable clock deltas. The deltas are filtered,
17 * making it monotonic and keeping it within an expected window. This window
18 * is set up using jiffies.
19 *
20 * Furthermore, explicit sleep and wakeup hooks allow us to account for time
21 * that is otherwise invisible (TSC gets stopped).
22 *
23 * The clock: sched_clock_cpu() is monotonic per cpu, and should be somewhat
24 * consistent between cpus (never more than 1 jiffies difference).
25 */
26#include <linux/sched.h>
27#include <linux/percpu.h>
28#include <linux/spinlock.h>
29#include <linux/ktime.h>
30#include <linux/module.h>
31
32
33#ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
34
35struct sched_clock_data {
36 /*
37 * Raw spinlock - this is a special case: this might be called
38 * from within instrumentation code so we dont want to do any
39 * instrumentation ourselves.
40 */
41 raw_spinlock_t lock;
42
43 unsigned long prev_jiffies;
44 u64 prev_raw;
45 u64 tick_raw;
46 u64 tick_gtod;
47 u64 clock;
48};
49
50static DEFINE_PER_CPU_SHARED_ALIGNED(struct sched_clock_data, sched_clock_data);
51
52static inline struct sched_clock_data *this_scd(void)
53{
54 return &__get_cpu_var(sched_clock_data);
55}
56
57static inline struct sched_clock_data *cpu_sdc(int cpu)
58{
59 return &per_cpu(sched_clock_data, cpu);
60}
61
62void sched_clock_init(void)
63{
64 u64 ktime_now = ktime_to_ns(ktime_get());
65 u64 now = 0;
66 int cpu;
67
68 for_each_possible_cpu(cpu) {
69 struct sched_clock_data *scd = cpu_sdc(cpu);
70
71 scd->lock = (raw_spinlock_t)__RAW_SPIN_LOCK_UNLOCKED;
72 scd->prev_jiffies = jiffies;
73 scd->prev_raw = now;
74 scd->tick_raw = now;
75 scd->tick_gtod = ktime_now;
76 scd->clock = ktime_now;
77 }
78}
79
80/*
81 * update the percpu scd from the raw @now value
82 *
83 * - filter out backward motion
84 * - use jiffies to generate a min,max window to clip the raw values
85 */
86static void __update_sched_clock(struct sched_clock_data *scd, u64 now)
87{
88 unsigned long now_jiffies = jiffies;
89 long delta_jiffies = now_jiffies - scd->prev_jiffies;
90 u64 clock = scd->clock;
91 u64 min_clock, max_clock;
92 s64 delta = now - scd->prev_raw;
93
94 WARN_ON_ONCE(!irqs_disabled());
95 min_clock = scd->tick_gtod + delta_jiffies * TICK_NSEC;
96
97 if (unlikely(delta < 0)) {
98 clock++;
99 goto out;
100 }
101
102 max_clock = min_clock + TICK_NSEC;
103
104 if (unlikely(clock + delta > max_clock)) {
105 if (clock < max_clock)
106 clock = max_clock;
107 else
108 clock++;
109 } else {
110 clock += delta;
111 }
112
113 out:
114 if (unlikely(clock < min_clock))
115 clock = min_clock;
116
117 scd->prev_raw = now;
118 scd->prev_jiffies = now_jiffies;
119 scd->clock = clock;
120}
121
122static void lock_double_clock(struct sched_clock_data *data1,
123 struct sched_clock_data *data2)
124{
125 if (data1 < data2) {
126 __raw_spin_lock(&data1->lock);
127 __raw_spin_lock(&data2->lock);
128 } else {
129 __raw_spin_lock(&data2->lock);
130 __raw_spin_lock(&data1->lock);
131 }
132}
133
134u64 sched_clock_cpu(int cpu)
135{
136 struct sched_clock_data *scd = cpu_sdc(cpu);
137 u64 now, clock;
138
139 WARN_ON_ONCE(!irqs_disabled());
140 now = sched_clock();
141
142 if (cpu != raw_smp_processor_id()) {
143 /*
144 * in order to update a remote cpu's clock based on our
145 * unstable raw time rebase it against:
146 * tick_raw (offset between raw counters)
147 * tick_gotd (tick offset between cpus)
148 */
149 struct sched_clock_data *my_scd = this_scd();
150
151 lock_double_clock(scd, my_scd);
152
153 now -= my_scd->tick_raw;
154 now += scd->tick_raw;
155
156 now -= my_scd->tick_gtod;
157 now += scd->tick_gtod;
158
159 __raw_spin_unlock(&my_scd->lock);
160 } else {
161 __raw_spin_lock(&scd->lock);
162 }
163
164 __update_sched_clock(scd, now);
165 clock = scd->clock;
166
167 __raw_spin_unlock(&scd->lock);
168
169 return clock;
170}
171
172void sched_clock_tick(void)
173{
174 struct sched_clock_data *scd = this_scd();
175 u64 now, now_gtod;
176
177 WARN_ON_ONCE(!irqs_disabled());
178
179 now = sched_clock();
180 now_gtod = ktime_to_ns(ktime_get());
181
182 __raw_spin_lock(&scd->lock);
183 __update_sched_clock(scd, now);
184 /*
185 * update tick_gtod after __update_sched_clock() because that will
186 * already observe 1 new jiffy; adding a new tick_gtod to that would
187 * increase the clock 2 jiffies.
188 */
189 scd->tick_raw = now;
190 scd->tick_gtod = now_gtod;
191 __raw_spin_unlock(&scd->lock);
192}
193
194/*
195 * We are going deep-idle (irqs are disabled):
196 */
197void sched_clock_idle_sleep_event(void)
198{
199 sched_clock_cpu(smp_processor_id());
200}
201EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
202
203/*
204 * We just idled delta nanoseconds (called with irqs disabled):
205 */
206void sched_clock_idle_wakeup_event(u64 delta_ns)
207{
208 struct sched_clock_data *scd = this_scd();
209 u64 now = sched_clock();
210
211 /*
212 * Override the previous timestamp and ignore all
213 * sched_clock() deltas that occured while we idled,
214 * and use the PM-provided delta_ns to advance the
215 * rq clock:
216 */
217 __raw_spin_lock(&scd->lock);
218 scd->prev_raw = now;
219 scd->clock += delta_ns;
220 __raw_spin_unlock(&scd->lock);
221
222 touch_softlockup_watchdog();
223}
224EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
225
226#endif
227
228/*
229 * Scheduler clock - returns current time in nanosec units.
230 * This is default implementation.
231 * Architectures and sub-architectures can override this.
232 */
233unsigned long long __attribute__((weak)) sched_clock(void)
234{
235 return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ);
236}