]>
Commit | Line | Data |
---|---|---|
bcea3f96 | 1 | // SPDX-License-Identifier: GPL-2.0 |
14131f2f IM |
2 | /* |
3 | * tracing clocks | |
4 | * | |
5 | * Copyright (C) 2009 Red Hat, Inc., Ingo Molnar <[email protected]> | |
6 | * | |
7 | * Implements 3 trace clock variants, with differing scalability/precision | |
8 | * tradeoffs: | |
9 | * | |
10 | * - local: CPU-local trace clock | |
11 | * - medium: scalable global clock with some jitter | |
12 | * - global: globally monotonic, serialized clock | |
13 | * | |
14 | * Tracer plugins will chose a default from these clocks. | |
15 | */ | |
16 | #include <linux/spinlock.h> | |
ae1f3038 | 17 | #include <linux/irqflags.h> |
14131f2f IM |
18 | #include <linux/hardirq.h> |
19 | #include <linux/module.h> | |
20 | #include <linux/percpu.h> | |
21 | #include <linux/sched.h> | |
e6017571 | 22 | #include <linux/sched/clock.h> |
14131f2f | 23 | #include <linux/ktime.h> |
b8b94265 | 24 | #include <linux/trace_clock.h> |
14131f2f IM |
25 | |
26 | /* | |
27 | * trace_clock_local(): the simplest and least coherent tracing clock. | |
28 | * | |
29 | * Useful for tracing that does not cross to other CPUs nor | |
30 | * does it go through idle events. | |
31 | */ | |
32 | u64 notrace trace_clock_local(void) | |
33 | { | |
6cc3c6e1 PZ |
34 | u64 clock; |
35 | ||
14131f2f IM |
36 | /* |
37 | * sched_clock() is an architecture implemented, fast, scalable, | |
38 | * lockless clock. It is not guaranteed to be coherent across | |
39 | * CPUs, nor across CPU idle events. | |
40 | */ | |
5168ae50 | 41 | preempt_disable_notrace(); |
6cc3c6e1 | 42 | clock = sched_clock(); |
5168ae50 | 43 | preempt_enable_notrace(); |
6cc3c6e1 PZ |
44 | |
45 | return clock; | |
14131f2f | 46 | } |
dc975e94 | 47 | EXPORT_SYMBOL_GPL(trace_clock_local); |
14131f2f IM |
48 | |
49 | /* | |
25985edc | 50 | * trace_clock(): 'between' trace clock. Not completely serialized, |
14131f2f IM |
51 | * but not completely incorrect when crossing CPUs either. |
52 | * | |
53 | * This is based on cpu_clock(), which will allow at most ~1 jiffy of | |
54 | * jitter between CPUs. So it's a pretty scalable clock, but there | |
55 | * can be offsets in the trace data. | |
56 | */ | |
57 | u64 notrace trace_clock(void) | |
58 | { | |
c676329a | 59 | return local_clock(); |
14131f2f | 60 | } |
7e255d34 | 61 | EXPORT_SYMBOL_GPL(trace_clock); |
14131f2f | 62 | |
8aacf017 SRRH |
63 | /* |
64 | * trace_jiffy_clock(): Simply use jiffies as a clock counter. | |
58d4e21e TL |
65 | * Note that this use of jiffies_64 is not completely safe on |
66 | * 32-bit systems. But the window is tiny, and the effect if | |
67 | * we are affected is that we will have an obviously bogus | |
68 | * timestamp on a trace event - i.e. not life threatening. | |
8aacf017 SRRH |
69 | */ |
70 | u64 notrace trace_clock_jiffies(void) | |
71 | { | |
58d4e21e | 72 | return jiffies_64_to_clock_t(jiffies_64 - INITIAL_JIFFIES); |
8aacf017 | 73 | } |
7e255d34 | 74 | EXPORT_SYMBOL_GPL(trace_clock_jiffies); |
14131f2f IM |
75 | |
76 | /* | |
77 | * trace_clock_global(): special globally coherent trace clock | |
78 | * | |
79 | * It has higher overhead than the other trace clocks but is still | |
80 | * an order of magnitude faster than GTOD derived hardware clocks. | |
81 | * | |
82 | * Used by plugins that need globally coherent timestamps. | |
83 | */ | |
84 | ||
6ca6cca3 SR |
85 | /* keep prev_time and lock in the same cacheline. */ |
86 | static struct { | |
87 | u64 prev_time; | |
445c8951 | 88 | arch_spinlock_t lock; |
6ca6cca3 SR |
89 | } trace_clock_struct ____cacheline_aligned_in_smp = |
90 | { | |
edc35bd7 | 91 | .lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED, |
6ca6cca3 | 92 | }; |
14131f2f IM |
93 | |
94 | u64 notrace trace_clock_global(void) | |
95 | { | |
96 | unsigned long flags; | |
97 | int this_cpu; | |
98 | u64 now; | |
99 | ||
f7a1570d | 100 | raw_local_irq_save(flags); |
14131f2f IM |
101 | |
102 | this_cpu = raw_smp_processor_id(); | |
5e67b51e | 103 | now = sched_clock_cpu(this_cpu); |
14131f2f IM |
104 | /* |
105 | * If in an NMI context then dont risk lockups and return the | |
106 | * cpu_clock() time: | |
107 | */ | |
108 | if (unlikely(in_nmi())) | |
109 | goto out; | |
110 | ||
0199c4e6 | 111 | arch_spin_lock(&trace_clock_struct.lock); |
14131f2f IM |
112 | |
113 | /* | |
114 | * TODO: if this happens often then maybe we should reset | |
6ca6cca3 | 115 | * my_scd->clock to prev_time+1, to make sure |
14131f2f IM |
116 | * we start ticking with the local clock from now on? |
117 | */ | |
6ca6cca3 SR |
118 | if ((s64)(now - trace_clock_struct.prev_time) < 0) |
119 | now = trace_clock_struct.prev_time + 1; | |
14131f2f | 120 | |
6ca6cca3 | 121 | trace_clock_struct.prev_time = now; |
14131f2f | 122 | |
0199c4e6 | 123 | arch_spin_unlock(&trace_clock_struct.lock); |
14131f2f IM |
124 | |
125 | out: | |
f7a1570d | 126 | raw_local_irq_restore(flags); |
14131f2f IM |
127 | |
128 | return now; | |
129 | } | |
7e255d34 | 130 | EXPORT_SYMBOL_GPL(trace_clock_global); |
6249687f SR |
131 | |
132 | static atomic64_t trace_counter; | |
133 | ||
134 | /* | |
135 | * trace_clock_counter(): simply an atomic counter. | |
136 | * Use the trace_counter "counter" for cases where you do not care | |
137 | * about timings, but are interested in strict ordering. | |
138 | */ | |
139 | u64 notrace trace_clock_counter(void) | |
140 | { | |
141 | return atomic64_add_return(1, &trace_counter); | |
142 | } |