xref: /linux/kernel/trace/trace_clock.c (revision 27eb2c4b3d3e13f376a359e293c212a2e9407af5)
1 /*
2  * tracing clocks
3  *
4  *  Copyright (C) 2009 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
5  *
6  * Implements 3 trace clock variants, with differing scalability/precision
7  * tradeoffs:
8  *
9  *  -   local: CPU-local trace clock
10  *  -  medium: scalable global clock with some jitter
11  *  -  global: globally monotonic, serialized clock
12  *
13  * Tracer plugins will chose a default from these clocks.
14  */
15 #include <linux/spinlock.h>
16 #include <linux/irqflags.h>
17 #include <linux/hardirq.h>
18 #include <linux/module.h>
19 #include <linux/percpu.h>
20 #include <linux/sched.h>
21 #include <linux/ktime.h>
22 #include <linux/trace_clock.h>
23 
24 /*
25  * trace_clock_local(): the simplest and least coherent tracing clock.
26  *
27  * Useful for tracing that does not cross to other CPUs nor
28  * does it go through idle events.
29  */
30 u64 notrace trace_clock_local(void)
31 {
32 	u64 clock;
33 
34 	/*
35 	 * sched_clock() is an architecture implemented, fast, scalable,
36 	 * lockless clock. It is not guaranteed to be coherent across
37 	 * CPUs, nor across CPU idle events.
38 	 */
39 	preempt_disable_notrace();
40 	clock = sched_clock();
41 	preempt_enable_notrace();
42 
43 	return clock;
44 }
45 EXPORT_SYMBOL_GPL(trace_clock_local);
46 
47 /*
48  * trace_clock(): 'between' trace clock. Not completely serialized,
49  * but not completely incorrect when crossing CPUs either.
50  *
51  * This is based on cpu_clock(), which will allow at most ~1 jiffy of
52  * jitter between CPUs. So it's a pretty scalable clock, but there
53  * can be offsets in the trace data.
54  */
55 u64 notrace trace_clock(void)
56 {
57 	return local_clock();
58 }
59 
60 /*
61  * trace_jiffy_clock(): Simply use jiffies as a clock counter.
62  */
63 u64 notrace trace_clock_jiffies(void)
64 {
65 	u64 jiffy = jiffies - INITIAL_JIFFIES;
66 
67 	/* Return nsecs */
68 	return (u64)jiffies_to_usecs(jiffy) * 1000ULL;
69 }
70 
71 /*
72  * trace_clock_global(): special globally coherent trace clock
73  *
74  * It has higher overhead than the other trace clocks but is still
75  * an order of magnitude faster than GTOD derived hardware clocks.
76  *
77  * Used by plugins that need globally coherent timestamps.
78  */
79 
80 /* keep prev_time and lock in the same cacheline. */
81 static struct {
82 	u64 prev_time;
83 	arch_spinlock_t lock;
84 } trace_clock_struct ____cacheline_aligned_in_smp =
85 	{
86 		.lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED,
87 	};
88 
89 u64 notrace trace_clock_global(void)
90 {
91 	unsigned long flags;
92 	int this_cpu;
93 	u64 now;
94 
95 	local_irq_save(flags);
96 
97 	this_cpu = raw_smp_processor_id();
98 	now = sched_clock_cpu(this_cpu);
99 	/*
100 	 * If in an NMI context then dont risk lockups and return the
101 	 * cpu_clock() time:
102 	 */
103 	if (unlikely(in_nmi()))
104 		goto out;
105 
106 	arch_spin_lock(&trace_clock_struct.lock);
107 
108 	/*
109 	 * TODO: if this happens often then maybe we should reset
110 	 * my_scd->clock to prev_time+1, to make sure
111 	 * we start ticking with the local clock from now on?
112 	 */
113 	if ((s64)(now - trace_clock_struct.prev_time) < 0)
114 		now = trace_clock_struct.prev_time + 1;
115 
116 	trace_clock_struct.prev_time = now;
117 
118 	arch_spin_unlock(&trace_clock_struct.lock);
119 
120  out:
121 	local_irq_restore(flags);
122 
123 	return now;
124 }
125 
126 static atomic64_t trace_counter;
127 
128 /*
129  * trace_clock_counter(): simply an atomic counter.
130  * Use the trace_counter "counter" for cases where you do not care
131  * about timings, but are interested in strict ordering.
132  */
133 u64 notrace trace_clock_counter(void)
134 {
135 	return atomic64_add_return(1, &trace_counter);
136 }
137