1 // SPDX-License-Identifier: GPL-2.0 <<
2 /* 1 /*
3 * tracing clocks 2 * tracing clocks
4 * 3 *
5 * Copyright (C) 2009 Red Hat, Inc., Ingo Mol 4 * Copyright (C) 2009 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * 5 *
7 * Implements 3 trace clock variants, with dif 6 * Implements 3 trace clock variants, with differing scalability/precision
8 * tradeoffs: 7 * tradeoffs:
9 * 8 *
10 * - local: CPU-local trace clock 9 * - local: CPU-local trace clock
11 * - medium: scalable global clock with some 10 * - medium: scalable global clock with some jitter
12 * - global: globally monotonic, serialized 11 * - global: globally monotonic, serialized clock
13 * 12 *
14 * Tracer plugins will chose a default from th 13 * Tracer plugins will chose a default from these clocks.
15 */ 14 */
16 #include <linux/spinlock.h> 15 #include <linux/spinlock.h>
17 #include <linux/irqflags.h> 16 #include <linux/irqflags.h>
18 #include <linux/hardirq.h> 17 #include <linux/hardirq.h>
19 #include <linux/module.h> 18 #include <linux/module.h>
20 #include <linux/percpu.h> 19 #include <linux/percpu.h>
21 #include <linux/sched.h> 20 #include <linux/sched.h>
22 #include <linux/sched/clock.h> 21 #include <linux/sched/clock.h>
23 #include <linux/ktime.h> 22 #include <linux/ktime.h>
24 #include <linux/trace_clock.h> 23 #include <linux/trace_clock.h>
25 24
26 /* 25 /*
27 * trace_clock_local(): the simplest and least 26 * trace_clock_local(): the simplest and least coherent tracing clock.
28 * 27 *
29 * Useful for tracing that does not cross to o 28 * Useful for tracing that does not cross to other CPUs nor
30 * does it go through idle events. 29 * does it go through idle events.
31 */ 30 */
32 u64 notrace trace_clock_local(void) 31 u64 notrace trace_clock_local(void)
33 { 32 {
34 u64 clock; 33 u64 clock;
35 34
36 /* 35 /*
37 * sched_clock() is an architecture im 36 * sched_clock() is an architecture implemented, fast, scalable,
38 * lockless clock. It is not guarantee 37 * lockless clock. It is not guaranteed to be coherent across
39 * CPUs, nor across CPU idle events. 38 * CPUs, nor across CPU idle events.
40 */ 39 */
41 preempt_disable_notrace(); 40 preempt_disable_notrace();
42 clock = sched_clock(); 41 clock = sched_clock();
43 preempt_enable_notrace(); 42 preempt_enable_notrace();
44 43
45 return clock; 44 return clock;
46 } 45 }
47 EXPORT_SYMBOL_GPL(trace_clock_local); 46 EXPORT_SYMBOL_GPL(trace_clock_local);
48 47
49 /* 48 /*
50 * trace_clock(): 'between' trace clock. Not c 49 * trace_clock(): 'between' trace clock. Not completely serialized,
51 * but not completely incorrect when crossing 50 * but not completely incorrect when crossing CPUs either.
52 * 51 *
53 * This is based on cpu_clock(), which will al 52 * This is based on cpu_clock(), which will allow at most ~1 jiffy of
54 * jitter between CPUs. So it's a pretty scala 53 * jitter between CPUs. So it's a pretty scalable clock, but there
55 * can be offsets in the trace data. 54 * can be offsets in the trace data.
56 */ 55 */
57 u64 notrace trace_clock(void) 56 u64 notrace trace_clock(void)
58 { 57 {
59 return local_clock(); 58 return local_clock();
60 } 59 }
61 EXPORT_SYMBOL_GPL(trace_clock); 60 EXPORT_SYMBOL_GPL(trace_clock);
62 61
63 /* 62 /*
64 * trace_jiffy_clock(): Simply use jiffies as 63 * trace_jiffy_clock(): Simply use jiffies as a clock counter.
65 * Note that this use of jiffies_64 is not com 64 * Note that this use of jiffies_64 is not completely safe on
66 * 32-bit systems. But the window is tiny, and 65 * 32-bit systems. But the window is tiny, and the effect if
67 * we are affected is that we will have an obv 66 * we are affected is that we will have an obviously bogus
68 * timestamp on a trace event - i.e. not life 67 * timestamp on a trace event - i.e. not life threatening.
69 */ 68 */
70 u64 notrace trace_clock_jiffies(void) 69 u64 notrace trace_clock_jiffies(void)
71 { 70 {
72 return jiffies_64_to_clock_t(jiffies_6 71 return jiffies_64_to_clock_t(jiffies_64 - INITIAL_JIFFIES);
73 } 72 }
74 EXPORT_SYMBOL_GPL(trace_clock_jiffies); 73 EXPORT_SYMBOL_GPL(trace_clock_jiffies);
75 74
76 /* 75 /*
77 * trace_clock_global(): special globally cohe 76 * trace_clock_global(): special globally coherent trace clock
78 * 77 *
79 * It has higher overhead than the other trace 78 * It has higher overhead than the other trace clocks but is still
80 * an order of magnitude faster than GTOD deri 79 * an order of magnitude faster than GTOD derived hardware clocks.
81 * 80 *
82 * Used by plugins that need globally coherent 81 * Used by plugins that need globally coherent timestamps.
83 */ 82 */
84 83
85 /* keep prev_time and lock in the same cacheli 84 /* keep prev_time and lock in the same cacheline. */
86 static struct { 85 static struct {
87 u64 prev_time; 86 u64 prev_time;
88 arch_spinlock_t lock; 87 arch_spinlock_t lock;
89 } trace_clock_struct ____cacheline_aligned_in_ 88 } trace_clock_struct ____cacheline_aligned_in_smp =
90 { 89 {
91 .lock = (arch_spinlock_t)__ARC 90 .lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED,
92 }; 91 };
93 92
94 u64 notrace trace_clock_global(void) 93 u64 notrace trace_clock_global(void)
95 { 94 {
96 unsigned long flags; 95 unsigned long flags;
97 int this_cpu; 96 int this_cpu;
98 u64 now, prev_time; !! 97 u64 now;
99 98
100 raw_local_irq_save(flags); 99 raw_local_irq_save(flags);
101 100
102 this_cpu = raw_smp_processor_id(); 101 this_cpu = raw_smp_processor_id();
103 !! 102 now = sched_clock_cpu(this_cpu);
104 /* 103 /*
105 * The global clock "guarantees" that !! 104 * If in an NMI context then dont risk lockups and return the
106 * between CPUs. But if two events on !! 105 * cpu_clock() time:
107 * trace_clock_global at roughly the s <<
108 * not matter which one gets the earli <<
109 * that the same CPU will always show <<
110 * <<
111 * Use a read memory barrier to get th <<
112 * time that was recorded. <<
113 */ 106 */
114 smp_rmb(); !! 107 if (unlikely(in_nmi()))
115 prev_time = READ_ONCE(trace_clock_stru !! 108 goto out;
116 now = sched_clock_cpu(this_cpu); <<
117 109
118 /* Make sure that now is always greate !! 110 arch_spin_lock(&trace_clock_struct.lock);
119 if ((s64)(now - prev_time) < 0) <<
120 now = prev_time; <<
121 111
122 /* 112 /*
123 * If in an NMI context then dont risk !! 113 * TODO: if this happens often then maybe we should reset
124 * the current time. !! 114 * my_scd->clock to prev_time+1, to make sure
>> 115 * we start ticking with the local clock from now on?
125 */ 116 */
126 if (unlikely(in_nmi())) !! 117 if ((s64)(now - trace_clock_struct.prev_time) < 0)
127 goto out; !! 118 now = trace_clock_struct.prev_time + 1;
>> 119
>> 120 trace_clock_struct.prev_time = now;
>> 121
>> 122 arch_spin_unlock(&trace_clock_struct.lock);
128 123
129 /* Tracing can cause strange recursion <<
130 if (arch_spin_trylock(&trace_clock_str <<
131 /* Reread prev_time in case it <<
132 prev_time = READ_ONCE(trace_cl <<
133 if ((s64)(now - prev_time) < 0 <<
134 now = prev_time; <<
135 <<
136 trace_clock_struct.prev_time = <<
137 <<
138 /* The unlock acts as the wmb <<
139 arch_spin_unlock(&trace_clock_ <<
140 } <<
141 out: 124 out:
142 raw_local_irq_restore(flags); 125 raw_local_irq_restore(flags);
143 126
144 return now; 127 return now;
145 } 128 }
146 EXPORT_SYMBOL_GPL(trace_clock_global); 129 EXPORT_SYMBOL_GPL(trace_clock_global);
147 130
148 static atomic64_t trace_counter; 131 static atomic64_t trace_counter;
149 132
150 /* 133 /*
151 * trace_clock_counter(): simply an atomic cou 134 * trace_clock_counter(): simply an atomic counter.
152 * Use the trace_counter "counter" for cases w 135 * Use the trace_counter "counter" for cases where you do not care
153 * about timings, but are interested in strict 136 * about timings, but are interested in strict ordering.
154 */ 137 */
155 u64 notrace trace_clock_counter(void) 138 u64 notrace trace_clock_counter(void)
156 { 139 {
157 return atomic64_add_return(1, &trace_c 140 return atomic64_add_return(1, &trace_counter);
158 } 141 }
159 142
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.