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