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 | * | |
c300ba25 SR |
6 | * Updates and enhancements: |
7 | * Copyright (C) 2008 Red Hat, Inc. Steven Rostedt <srostedt@redhat.com> | |
8 | * | |
3e51f33f PZ |
9 | * Based on code by: |
10 | * Ingo Molnar <mingo@redhat.com> | |
11 | * Guillaume Chazarain <guichaz@gmail.com> | |
12 | * | |
13 | * Create a semi stable clock from a mixture of other events, including: | |
14 | * - gtod | |
3e51f33f PZ |
15 | * - sched_clock() |
16 | * - explicit idle events | |
17 | * | |
18 | * We use gtod as base and the unstable clock deltas. The deltas are filtered, | |
354879bb | 19 | * making it monotonic and keeping it within an expected window. |
3e51f33f PZ |
20 | * |
21 | * Furthermore, explicit sleep and wakeup hooks allow us to account for time | |
22 | * that is otherwise invisible (TSC gets stopped). | |
23 | * | |
24 | * The clock: sched_clock_cpu() is monotonic per cpu, and should be somewhat | |
354879bb | 25 | * consistent between cpus (never more than 2 jiffies difference). |
3e51f33f | 26 | */ |
3e51f33f | 27 | #include <linux/spinlock.h> |
3e51f33f | 28 | #include <linux/module.h> |
b342501c IM |
29 | #include <linux/percpu.h> |
30 | #include <linux/ktime.h> | |
31 | #include <linux/sched.h> | |
3e51f33f | 32 | |
2c3d103b HD |
33 | /* |
34 | * Scheduler clock - returns current time in nanosec units. | |
35 | * This is default implementation. | |
36 | * Architectures and sub-architectures can override this. | |
37 | */ | |
38 | unsigned long long __attribute__((weak)) sched_clock(void) | |
39 | { | |
40 | return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ); | |
41 | } | |
3e51f33f | 42 | |
c1955a3d PZ |
43 | static __read_mostly int sched_clock_running; |
44 | ||
3e51f33f | 45 | #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK |
b342501c IM |
46 | __read_mostly int sched_clock_stable; |
47 | #else | |
48 | static const int sched_clock_stable = 1; | |
49 | #endif | |
3e51f33f PZ |
50 | |
51 | struct sched_clock_data { | |
52 | /* | |
53 | * Raw spinlock - this is a special case: this might be called | |
54 | * from within instrumentation code so we dont want to do any | |
55 | * instrumentation ourselves. | |
56 | */ | |
57 | raw_spinlock_t lock; | |
58 | ||
3e51f33f PZ |
59 | u64 tick_raw; |
60 | u64 tick_gtod; | |
61 | u64 clock; | |
62 | }; | |
63 | ||
64 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct sched_clock_data, sched_clock_data); | |
65 | ||
66 | static inline struct sched_clock_data *this_scd(void) | |
67 | { | |
68 | return &__get_cpu_var(sched_clock_data); | |
69 | } | |
70 | ||
71 | static inline struct sched_clock_data *cpu_sdc(int cpu) | |
72 | { | |
73 | return &per_cpu(sched_clock_data, cpu); | |
74 | } | |
75 | ||
76 | void sched_clock_init(void) | |
77 | { | |
78 | u64 ktime_now = ktime_to_ns(ktime_get()); | |
3e51f33f PZ |
79 | int cpu; |
80 | ||
81 | for_each_possible_cpu(cpu) { | |
82 | struct sched_clock_data *scd = cpu_sdc(cpu); | |
83 | ||
84 | scd->lock = (raw_spinlock_t)__RAW_SPIN_LOCK_UNLOCKED; | |
a381759d | 85 | scd->tick_raw = 0; |
3e51f33f PZ |
86 | scd->tick_gtod = ktime_now; |
87 | scd->clock = ktime_now; | |
88 | } | |
a381759d PZ |
89 | |
90 | sched_clock_running = 1; | |
3e51f33f PZ |
91 | } |
92 | ||
354879bb | 93 | /* |
b342501c | 94 | * min, max except they take wrapping into account |
354879bb PZ |
95 | */ |
96 | ||
97 | static inline u64 wrap_min(u64 x, u64 y) | |
98 | { | |
99 | return (s64)(x - y) < 0 ? x : y; | |
100 | } | |
101 | ||
102 | static inline u64 wrap_max(u64 x, u64 y) | |
103 | { | |
104 | return (s64)(x - y) > 0 ? x : y; | |
105 | } | |
106 | ||
3e51f33f PZ |
107 | /* |
108 | * update the percpu scd from the raw @now value | |
109 | * | |
110 | * - filter out backward motion | |
354879bb | 111 | * - use the GTOD tick value to create a window to filter crazy TSC values |
3e51f33f | 112 | */ |
56b90612 | 113 | static u64 __update_sched_clock(struct sched_clock_data *scd, u64 now) |
3e51f33f | 114 | { |
18e4e36c | 115 | s64 delta = now - scd->tick_raw; |
354879bb | 116 | u64 clock, min_clock, max_clock; |
3e51f33f PZ |
117 | |
118 | WARN_ON_ONCE(!irqs_disabled()); | |
3e51f33f | 119 | |
354879bb PZ |
120 | if (unlikely(delta < 0)) |
121 | delta = 0; | |
3e51f33f | 122 | |
b342501c IM |
123 | if (unlikely(!sched_clock_running)) |
124 | return 0ull; | |
125 | ||
354879bb PZ |
126 | /* |
127 | * scd->clock = clamp(scd->tick_gtod + delta, | |
b342501c IM |
128 | * max(scd->tick_gtod, scd->clock), |
129 | * scd->tick_gtod + TICK_NSEC); | |
354879bb | 130 | */ |
3e51f33f | 131 | |
354879bb PZ |
132 | clock = scd->tick_gtod + delta; |
133 | min_clock = wrap_max(scd->tick_gtod, scd->clock); | |
1c5745aa | 134 | max_clock = wrap_max(scd->clock, scd->tick_gtod + TICK_NSEC); |
3e51f33f | 135 | |
354879bb PZ |
136 | clock = wrap_max(clock, min_clock); |
137 | clock = wrap_min(clock, max_clock); | |
3e51f33f | 138 | |
e4e4e534 | 139 | scd->clock = clock; |
56b90612 | 140 | |
354879bb | 141 | return scd->clock; |
3e51f33f PZ |
142 | } |
143 | ||
144 | static void lock_double_clock(struct sched_clock_data *data1, | |
145 | struct sched_clock_data *data2) | |
146 | { | |
147 | if (data1 < data2) { | |
148 | __raw_spin_lock(&data1->lock); | |
149 | __raw_spin_lock(&data2->lock); | |
150 | } else { | |
151 | __raw_spin_lock(&data2->lock); | |
152 | __raw_spin_lock(&data1->lock); | |
153 | } | |
154 | } | |
155 | ||
156 | u64 sched_clock_cpu(int cpu) | |
157 | { | |
4a273f20 | 158 | u64 now, clock, this_clock, remote_clock; |
b342501c | 159 | struct sched_clock_data *scd; |
3e51f33f | 160 | |
b342501c IM |
161 | if (sched_clock_stable) |
162 | return sched_clock(); | |
a381759d | 163 | |
b342501c | 164 | scd = cpu_sdc(cpu); |
3e51f33f PZ |
165 | WARN_ON_ONCE(!irqs_disabled()); |
166 | now = sched_clock(); | |
167 | ||
168 | if (cpu != raw_smp_processor_id()) { | |
3e51f33f PZ |
169 | struct sched_clock_data *my_scd = this_scd(); |
170 | ||
171 | lock_double_clock(scd, my_scd); | |
172 | ||
4a273f20 IM |
173 | this_clock = __update_sched_clock(my_scd, now); |
174 | remote_clock = scd->clock; | |
175 | ||
176 | /* | |
177 | * Use the opportunity that we have both locks | |
178 | * taken to couple the two clocks: we take the | |
179 | * larger time as the latest time for both | |
180 | * runqueues. (this creates monotonic movement) | |
181 | */ | |
354879bb | 182 | if (likely((s64)(remote_clock - this_clock) < 0)) { |
4a273f20 IM |
183 | clock = this_clock; |
184 | scd->clock = clock; | |
185 | } else { | |
186 | /* | |
187 | * Should be rare, but possible: | |
188 | */ | |
189 | clock = remote_clock; | |
190 | my_scd->clock = remote_clock; | |
191 | } | |
3e51f33f PZ |
192 | |
193 | __raw_spin_unlock(&my_scd->lock); | |
194 | } else { | |
195 | __raw_spin_lock(&scd->lock); | |
4a273f20 | 196 | clock = __update_sched_clock(scd, now); |
3e51f33f PZ |
197 | } |
198 | ||
e4e4e534 IM |
199 | __raw_spin_unlock(&scd->lock); |
200 | ||
3e51f33f PZ |
201 | return clock; |
202 | } | |
203 | ||
b342501c IM |
204 | #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK |
205 | ||
3e51f33f PZ |
206 | void sched_clock_tick(void) |
207 | { | |
208 | struct sched_clock_data *scd = this_scd(); | |
209 | u64 now, now_gtod; | |
210 | ||
a381759d PZ |
211 | if (unlikely(!sched_clock_running)) |
212 | return; | |
213 | ||
3e51f33f PZ |
214 | WARN_ON_ONCE(!irqs_disabled()); |
215 | ||
3e51f33f | 216 | now_gtod = ktime_to_ns(ktime_get()); |
a83bc47c | 217 | now = sched_clock(); |
3e51f33f PZ |
218 | |
219 | __raw_spin_lock(&scd->lock); | |
3e51f33f PZ |
220 | scd->tick_raw = now; |
221 | scd->tick_gtod = now_gtod; | |
354879bb | 222 | __update_sched_clock(scd, now); |
3e51f33f PZ |
223 | __raw_spin_unlock(&scd->lock); |
224 | } | |
225 | ||
226 | /* | |
227 | * We are going deep-idle (irqs are disabled): | |
228 | */ | |
229 | void sched_clock_idle_sleep_event(void) | |
230 | { | |
231 | sched_clock_cpu(smp_processor_id()); | |
232 | } | |
233 | EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event); | |
234 | ||
235 | /* | |
236 | * We just idled delta nanoseconds (called with irqs disabled): | |
237 | */ | |
238 | void sched_clock_idle_wakeup_event(u64 delta_ns) | |
239 | { | |
1c5745aa TG |
240 | if (timekeeping_suspended) |
241 | return; | |
242 | ||
354879bb | 243 | sched_clock_tick(); |
3e51f33f PZ |
244 | touch_softlockup_watchdog(); |
245 | } | |
246 | EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event); | |
247 | ||
b342501c | 248 | #endif /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */ |
3e51f33f | 249 | |
76a2a6ee PZ |
250 | unsigned long long cpu_clock(int cpu) |
251 | { | |
252 | unsigned long long clock; | |
253 | unsigned long flags; | |
254 | ||
2d452c9b | 255 | local_irq_save(flags); |
76a2a6ee | 256 | clock = sched_clock_cpu(cpu); |
2d452c9b | 257 | local_irq_restore(flags); |
76a2a6ee PZ |
258 | |
259 | return clock; | |
260 | } | |
4c9fe8ad | 261 | EXPORT_SYMBOL_GPL(cpu_clock); |