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> |
6409c4da | 28 | #include <linux/hardirq.h> |
3e51f33f | 29 | #include <linux/module.h> |
b342501c IM |
30 | #include <linux/percpu.h> |
31 | #include <linux/ktime.h> | |
32 | #include <linux/sched.h> | |
3e51f33f | 33 | |
2c3d103b HD |
34 | /* |
35 | * Scheduler clock - returns current time in nanosec units. | |
36 | * This is default implementation. | |
37 | * Architectures and sub-architectures can override this. | |
38 | */ | |
39 | unsigned long long __attribute__((weak)) sched_clock(void) | |
40 | { | |
41 | return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ); | |
42 | } | |
3e51f33f | 43 | |
c1955a3d PZ |
44 | static __read_mostly int sched_clock_running; |
45 | ||
3e51f33f | 46 | #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK |
b342501c | 47 | __read_mostly int sched_clock_stable; |
3e51f33f PZ |
48 | |
49 | struct sched_clock_data { | |
50 | /* | |
51 | * Raw spinlock - this is a special case: this might be called | |
52 | * from within instrumentation code so we dont want to do any | |
53 | * instrumentation ourselves. | |
54 | */ | |
55 | raw_spinlock_t lock; | |
56 | ||
3e51f33f PZ |
57 | u64 tick_raw; |
58 | u64 tick_gtod; | |
59 | u64 clock; | |
60 | }; | |
61 | ||
62 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct sched_clock_data, sched_clock_data); | |
63 | ||
64 | static inline struct sched_clock_data *this_scd(void) | |
65 | { | |
66 | return &__get_cpu_var(sched_clock_data); | |
67 | } | |
68 | ||
69 | static inline struct sched_clock_data *cpu_sdc(int cpu) | |
70 | { | |
71 | return &per_cpu(sched_clock_data, cpu); | |
72 | } | |
73 | ||
74 | void sched_clock_init(void) | |
75 | { | |
76 | u64 ktime_now = ktime_to_ns(ktime_get()); | |
3e51f33f PZ |
77 | int cpu; |
78 | ||
79 | for_each_possible_cpu(cpu) { | |
80 | struct sched_clock_data *scd = cpu_sdc(cpu); | |
81 | ||
82 | scd->lock = (raw_spinlock_t)__RAW_SPIN_LOCK_UNLOCKED; | |
a381759d | 83 | scd->tick_raw = 0; |
3e51f33f PZ |
84 | scd->tick_gtod = ktime_now; |
85 | scd->clock = ktime_now; | |
86 | } | |
a381759d PZ |
87 | |
88 | sched_clock_running = 1; | |
3e51f33f PZ |
89 | } |
90 | ||
354879bb | 91 | /* |
b342501c | 92 | * min, max except they take wrapping into account |
354879bb PZ |
93 | */ |
94 | ||
95 | static inline u64 wrap_min(u64 x, u64 y) | |
96 | { | |
97 | return (s64)(x - y) < 0 ? x : y; | |
98 | } | |
99 | ||
100 | static inline u64 wrap_max(u64 x, u64 y) | |
101 | { | |
102 | return (s64)(x - y) > 0 ? x : y; | |
103 | } | |
104 | ||
3e51f33f PZ |
105 | /* |
106 | * update the percpu scd from the raw @now value | |
107 | * | |
108 | * - filter out backward motion | |
354879bb | 109 | * - use the GTOD tick value to create a window to filter crazy TSC values |
3e51f33f | 110 | */ |
56b90612 | 111 | static u64 __update_sched_clock(struct sched_clock_data *scd, u64 now) |
3e51f33f | 112 | { |
18e4e36c | 113 | s64 delta = now - scd->tick_raw; |
354879bb | 114 | u64 clock, min_clock, max_clock; |
3e51f33f | 115 | |
354879bb PZ |
116 | if (unlikely(delta < 0)) |
117 | delta = 0; | |
3e51f33f | 118 | |
354879bb PZ |
119 | /* |
120 | * scd->clock = clamp(scd->tick_gtod + delta, | |
b342501c IM |
121 | * max(scd->tick_gtod, scd->clock), |
122 | * scd->tick_gtod + TICK_NSEC); | |
354879bb | 123 | */ |
3e51f33f | 124 | |
354879bb PZ |
125 | clock = scd->tick_gtod + delta; |
126 | min_clock = wrap_max(scd->tick_gtod, scd->clock); | |
1c5745aa | 127 | max_clock = wrap_max(scd->clock, scd->tick_gtod + TICK_NSEC); |
3e51f33f | 128 | |
354879bb PZ |
129 | clock = wrap_max(clock, min_clock); |
130 | clock = wrap_min(clock, max_clock); | |
3e51f33f | 131 | |
e4e4e534 | 132 | scd->clock = clock; |
56b90612 | 133 | |
354879bb | 134 | return scd->clock; |
3e51f33f PZ |
135 | } |
136 | ||
137 | static void lock_double_clock(struct sched_clock_data *data1, | |
138 | struct sched_clock_data *data2) | |
139 | { | |
140 | if (data1 < data2) { | |
141 | __raw_spin_lock(&data1->lock); | |
142 | __raw_spin_lock(&data2->lock); | |
143 | } else { | |
144 | __raw_spin_lock(&data2->lock); | |
145 | __raw_spin_lock(&data1->lock); | |
146 | } | |
147 | } | |
148 | ||
149 | u64 sched_clock_cpu(int cpu) | |
150 | { | |
4a273f20 | 151 | u64 now, clock, this_clock, remote_clock; |
b342501c | 152 | struct sched_clock_data *scd; |
3e51f33f | 153 | |
b342501c IM |
154 | if (sched_clock_stable) |
155 | return sched_clock(); | |
a381759d | 156 | |
b342501c | 157 | scd = cpu_sdc(cpu); |
3e51f33f | 158 | |
6409c4da IM |
159 | /* |
160 | * Normally this is not called in NMI context - but if it is, | |
161 | * trying to do any locking here is totally lethal. | |
162 | */ | |
163 | if (unlikely(in_nmi())) | |
164 | return scd->clock; | |
165 | ||
a381759d PZ |
166 | if (unlikely(!sched_clock_running)) |
167 | return 0ull; | |
168 | ||
3e51f33f PZ |
169 | WARN_ON_ONCE(!irqs_disabled()); |
170 | now = sched_clock(); | |
171 | ||
172 | if (cpu != raw_smp_processor_id()) { | |
3e51f33f PZ |
173 | struct sched_clock_data *my_scd = this_scd(); |
174 | ||
175 | lock_double_clock(scd, my_scd); | |
176 | ||
4a273f20 IM |
177 | this_clock = __update_sched_clock(my_scd, now); |
178 | remote_clock = scd->clock; | |
179 | ||
180 | /* | |
181 | * Use the opportunity that we have both locks | |
182 | * taken to couple the two clocks: we take the | |
183 | * larger time as the latest time for both | |
184 | * runqueues. (this creates monotonic movement) | |
185 | */ | |
354879bb | 186 | if (likely((s64)(remote_clock - this_clock) < 0)) { |
4a273f20 IM |
187 | clock = this_clock; |
188 | scd->clock = clock; | |
189 | } else { | |
190 | /* | |
191 | * Should be rare, but possible: | |
192 | */ | |
193 | clock = remote_clock; | |
194 | my_scd->clock = remote_clock; | |
195 | } | |
3e51f33f PZ |
196 | |
197 | __raw_spin_unlock(&my_scd->lock); | |
198 | } else { | |
199 | __raw_spin_lock(&scd->lock); | |
4a273f20 | 200 | clock = __update_sched_clock(scd, now); |
3e51f33f PZ |
201 | } |
202 | ||
e4e4e534 IM |
203 | __raw_spin_unlock(&scd->lock); |
204 | ||
3e51f33f PZ |
205 | return clock; |
206 | } | |
207 | ||
208 | void sched_clock_tick(void) | |
209 | { | |
8325d9c0 | 210 | struct sched_clock_data *scd; |
3e51f33f PZ |
211 | u64 now, now_gtod; |
212 | ||
8325d9c0 PZ |
213 | if (sched_clock_stable) |
214 | return; | |
215 | ||
a381759d PZ |
216 | if (unlikely(!sched_clock_running)) |
217 | return; | |
218 | ||
3e51f33f PZ |
219 | WARN_ON_ONCE(!irqs_disabled()); |
220 | ||
8325d9c0 | 221 | scd = this_scd(); |
3e51f33f | 222 | now_gtod = ktime_to_ns(ktime_get()); |
a83bc47c | 223 | now = sched_clock(); |
3e51f33f PZ |
224 | |
225 | __raw_spin_lock(&scd->lock); | |
3e51f33f PZ |
226 | scd->tick_raw = now; |
227 | scd->tick_gtod = now_gtod; | |
354879bb | 228 | __update_sched_clock(scd, now); |
3e51f33f PZ |
229 | __raw_spin_unlock(&scd->lock); |
230 | } | |
231 | ||
232 | /* | |
233 | * We are going deep-idle (irqs are disabled): | |
234 | */ | |
235 | void sched_clock_idle_sleep_event(void) | |
236 | { | |
237 | sched_clock_cpu(smp_processor_id()); | |
238 | } | |
239 | EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event); | |
240 | ||
241 | /* | |
242 | * We just idled delta nanoseconds (called with irqs disabled): | |
243 | */ | |
244 | void sched_clock_idle_wakeup_event(u64 delta_ns) | |
245 | { | |
1c5745aa TG |
246 | if (timekeeping_suspended) |
247 | return; | |
248 | ||
354879bb | 249 | sched_clock_tick(); |
3e51f33f PZ |
250 | touch_softlockup_watchdog(); |
251 | } | |
252 | EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event); | |
253 | ||
8325d9c0 PZ |
254 | #else /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */ |
255 | ||
256 | void sched_clock_init(void) | |
257 | { | |
258 | sched_clock_running = 1; | |
259 | } | |
260 | ||
261 | u64 sched_clock_cpu(int cpu) | |
262 | { | |
263 | if (unlikely(!sched_clock_running)) | |
264 | return 0; | |
265 | ||
266 | return sched_clock(); | |
267 | } | |
268 | ||
b342501c | 269 | #endif /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */ |
3e51f33f | 270 | |
76a2a6ee PZ |
271 | unsigned long long cpu_clock(int cpu) |
272 | { | |
273 | unsigned long long clock; | |
274 | unsigned long flags; | |
275 | ||
2d452c9b | 276 | local_irq_save(flags); |
76a2a6ee | 277 | clock = sched_clock_cpu(cpu); |
2d452c9b | 278 | local_irq_restore(flags); |
76a2a6ee PZ |
279 | |
280 | return clock; | |
281 | } | |
4c9fe8ad | 282 | EXPORT_SYMBOL_GPL(cpu_clock); |