Commit | Line | Data |
---|---|---|
3e51f33f PZ |
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 | ||
35 | struct 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 | ||
50 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct sched_clock_data, sched_clock_data); | |
51 | ||
52 | static inline struct sched_clock_data *this_scd(void) | |
53 | { | |
54 | return &__get_cpu_var(sched_clock_data); | |
55 | } | |
56 | ||
57 | static inline struct sched_clock_data *cpu_sdc(int cpu) | |
58 | { | |
59 | return &per_cpu(sched_clock_data, cpu); | |
60 | } | |
61 | ||
a381759d PZ |
62 | static __read_mostly int sched_clock_running; |
63 | ||
3e51f33f PZ |
64 | void sched_clock_init(void) |
65 | { | |
66 | u64 ktime_now = ktime_to_ns(ktime_get()); | |
a381759d | 67 | unsigned long now_jiffies = jiffies; |
3e51f33f PZ |
68 | int cpu; |
69 | ||
70 | for_each_possible_cpu(cpu) { | |
71 | struct sched_clock_data *scd = cpu_sdc(cpu); | |
72 | ||
73 | scd->lock = (raw_spinlock_t)__RAW_SPIN_LOCK_UNLOCKED; | |
a381759d PZ |
74 | scd->prev_jiffies = now_jiffies; |
75 | scd->prev_raw = 0; | |
76 | scd->tick_raw = 0; | |
3e51f33f PZ |
77 | scd->tick_gtod = ktime_now; |
78 | scd->clock = ktime_now; | |
79 | } | |
a381759d PZ |
80 | |
81 | sched_clock_running = 1; | |
3e51f33f PZ |
82 | } |
83 | ||
84 | /* | |
85 | * update the percpu scd from the raw @now value | |
86 | * | |
87 | * - filter out backward motion | |
88 | * - use jiffies to generate a min,max window to clip the raw values | |
89 | */ | |
90 | static void __update_sched_clock(struct sched_clock_data *scd, u64 now) | |
91 | { | |
92 | unsigned long now_jiffies = jiffies; | |
93 | long delta_jiffies = now_jiffies - scd->prev_jiffies; | |
94 | u64 clock = scd->clock; | |
95 | u64 min_clock, max_clock; | |
96 | s64 delta = now - scd->prev_raw; | |
97 | ||
98 | WARN_ON_ONCE(!irqs_disabled()); | |
99 | min_clock = scd->tick_gtod + delta_jiffies * TICK_NSEC; | |
100 | ||
101 | if (unlikely(delta < 0)) { | |
102 | clock++; | |
103 | goto out; | |
104 | } | |
105 | ||
106 | max_clock = min_clock + TICK_NSEC; | |
107 | ||
108 | if (unlikely(clock + delta > max_clock)) { | |
109 | if (clock < max_clock) | |
110 | clock = max_clock; | |
111 | else | |
112 | clock++; | |
113 | } else { | |
114 | clock += delta; | |
115 | } | |
116 | ||
117 | out: | |
118 | if (unlikely(clock < min_clock)) | |
119 | clock = min_clock; | |
120 | ||
121 | scd->prev_raw = now; | |
122 | scd->prev_jiffies = now_jiffies; | |
123 | scd->clock = clock; | |
124 | } | |
125 | ||
126 | static void lock_double_clock(struct sched_clock_data *data1, | |
127 | struct sched_clock_data *data2) | |
128 | { | |
129 | if (data1 < data2) { | |
130 | __raw_spin_lock(&data1->lock); | |
131 | __raw_spin_lock(&data2->lock); | |
132 | } else { | |
133 | __raw_spin_lock(&data2->lock); | |
134 | __raw_spin_lock(&data1->lock); | |
135 | } | |
136 | } | |
137 | ||
138 | u64 sched_clock_cpu(int cpu) | |
139 | { | |
140 | struct sched_clock_data *scd = cpu_sdc(cpu); | |
141 | u64 now, clock; | |
142 | ||
a381759d PZ |
143 | if (unlikely(!sched_clock_running)) |
144 | return 0ull; | |
145 | ||
3e51f33f PZ |
146 | WARN_ON_ONCE(!irqs_disabled()); |
147 | now = sched_clock(); | |
148 | ||
149 | if (cpu != raw_smp_processor_id()) { | |
150 | /* | |
151 | * in order to update a remote cpu's clock based on our | |
152 | * unstable raw time rebase it against: | |
153 | * tick_raw (offset between raw counters) | |
154 | * tick_gotd (tick offset between cpus) | |
155 | */ | |
156 | struct sched_clock_data *my_scd = this_scd(); | |
157 | ||
158 | lock_double_clock(scd, my_scd); | |
159 | ||
160 | now -= my_scd->tick_raw; | |
161 | now += scd->tick_raw; | |
162 | ||
163 | now -= my_scd->tick_gtod; | |
164 | now += scd->tick_gtod; | |
165 | ||
166 | __raw_spin_unlock(&my_scd->lock); | |
167 | } else { | |
168 | __raw_spin_lock(&scd->lock); | |
169 | } | |
170 | ||
171 | __update_sched_clock(scd, now); | |
172 | clock = scd->clock; | |
173 | ||
174 | __raw_spin_unlock(&scd->lock); | |
175 | ||
176 | return clock; | |
177 | } | |
178 | ||
179 | void sched_clock_tick(void) | |
180 | { | |
181 | struct sched_clock_data *scd = this_scd(); | |
182 | u64 now, now_gtod; | |
183 | ||
a381759d PZ |
184 | if (unlikely(!sched_clock_running)) |
185 | return; | |
186 | ||
3e51f33f PZ |
187 | WARN_ON_ONCE(!irqs_disabled()); |
188 | ||
189 | now = sched_clock(); | |
190 | now_gtod = ktime_to_ns(ktime_get()); | |
191 | ||
192 | __raw_spin_lock(&scd->lock); | |
193 | __update_sched_clock(scd, now); | |
194 | /* | |
195 | * update tick_gtod after __update_sched_clock() because that will | |
196 | * already observe 1 new jiffy; adding a new tick_gtod to that would | |
197 | * increase the clock 2 jiffies. | |
198 | */ | |
199 | scd->tick_raw = now; | |
200 | scd->tick_gtod = now_gtod; | |
201 | __raw_spin_unlock(&scd->lock); | |
202 | } | |
203 | ||
204 | /* | |
205 | * We are going deep-idle (irqs are disabled): | |
206 | */ | |
207 | void sched_clock_idle_sleep_event(void) | |
208 | { | |
209 | sched_clock_cpu(smp_processor_id()); | |
210 | } | |
211 | EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event); | |
212 | ||
213 | /* | |
214 | * We just idled delta nanoseconds (called with irqs disabled): | |
215 | */ | |
216 | void sched_clock_idle_wakeup_event(u64 delta_ns) | |
217 | { | |
218 | struct sched_clock_data *scd = this_scd(); | |
219 | u64 now = sched_clock(); | |
220 | ||
221 | /* | |
222 | * Override the previous timestamp and ignore all | |
223 | * sched_clock() deltas that occured while we idled, | |
224 | * and use the PM-provided delta_ns to advance the | |
225 | * rq clock: | |
226 | */ | |
227 | __raw_spin_lock(&scd->lock); | |
228 | scd->prev_raw = now; | |
229 | scd->clock += delta_ns; | |
230 | __raw_spin_unlock(&scd->lock); | |
231 | ||
232 | touch_softlockup_watchdog(); | |
233 | } | |
234 | EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event); | |
235 | ||
236 | #endif | |
237 | ||
238 | /* | |
239 | * Scheduler clock - returns current time in nanosec units. | |
240 | * This is default implementation. | |
241 | * Architectures and sub-architectures can override this. | |
242 | */ | |
243 | unsigned long long __attribute__((weak)) sched_clock(void) | |
244 | { | |
245 | return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ); | |
246 | } | |
76a2a6ee PZ |
247 | |
248 | unsigned long long cpu_clock(int cpu) | |
249 | { | |
250 | unsigned long long clock; | |
251 | unsigned long flags; | |
252 | ||
253 | raw_local_irq_save(flags); | |
254 | clock = sched_clock_cpu(cpu); | |
255 | raw_local_irq_restore(flags); | |
256 | ||
257 | return clock; | |
258 | } |