1// SPDX-License-Identifier: GPL-2.0
2#include <linux/delay.h>
3#include <linux/module.h>
4#include <linux/kthread.h>
5#include <linux/trace_clock.h>
6
7#define CREATE_TRACE_POINTS
8#include "trace_benchmark.h"
9
10static struct task_struct *bm_event_thread;
11
12static char bm_str[BENCHMARK_EVENT_STRLEN] = "START";
13
14static u64 bm_total;
15static u64 bm_totalsq;
16static u64 bm_last;
17static u64 bm_max;
18static u64 bm_min;
19static u64 bm_first;
20static u64 bm_cnt;
21static u64 bm_stddev;
22static unsigned int bm_avg;
23static unsigned int bm_std;
24
25static bool ok_to_run;
26
27/*
28 * This gets called in a loop recording the time it took to write
29 * the tracepoint. What it writes is the time statistics of the last
30 * tracepoint write. As there is nothing to write the first time
31 * it simply writes "START". As the first write is cold cache and
32 * the rest is hot, we save off that time in bm_first and it is
33 * reported as "first", which is shown in the second write to the
34 * tracepoint. The "first" field is written within the statics from
35 * then on but never changes.
36 */
37static void trace_do_benchmark(void)
38{
39 u64 start;
40 u64 stop;
41 u64 delta;
42 u64 stddev;
43 u64 seed;
44 u64 last_seed;
45 unsigned int avg;
46 unsigned int std = 0;
47
48 /* Only run if the tracepoint is actually active */
49 if (!trace_benchmark_event_enabled() || !tracing_is_on())
50 return;
51
52 local_irq_disable();
53 start = trace_clock_local();
54 trace_benchmark_event(str: bm_str, delta: bm_last);
55 stop = trace_clock_local();
56 local_irq_enable();
57
58 bm_cnt++;
59
60 delta = stop - start;
61
62 /*
63 * The first read is cold cached, keep it separate from the
64 * other calculations.
65 */
66 if (bm_cnt == 1) {
67 bm_first = delta;
68 scnprintf(buf: bm_str, BENCHMARK_EVENT_STRLEN,
69 fmt: "first=%llu [COLD CACHED]", bm_first);
70 return;
71 }
72
73 bm_last = delta;
74
75 if (delta > bm_max)
76 bm_max = delta;
77 if (!bm_min || delta < bm_min)
78 bm_min = delta;
79
80 /*
81 * When bm_cnt is greater than UINT_MAX, it breaks the statistics
82 * accounting. Freeze the statistics when that happens.
83 * We should have enough data for the avg and stddev anyway.
84 */
85 if (bm_cnt > UINT_MAX) {
86 scnprintf(buf: bm_str, BENCHMARK_EVENT_STRLEN,
87 fmt: "last=%llu first=%llu max=%llu min=%llu ** avg=%u std=%d std^2=%lld",
88 bm_last, bm_first, bm_max, bm_min, bm_avg, bm_std, bm_stddev);
89 return;
90 }
91
92 bm_total += delta;
93 bm_totalsq += delta * delta;
94
95
96 if (bm_cnt > 1) {
97 /*
98 * Apply Welford's method to calculate standard deviation:
99 * s^2 = 1 / (n * (n-1)) * (n * \Sum (x_i)^2 - (\Sum x_i)^2)
100 */
101 stddev = (u64)bm_cnt * bm_totalsq - bm_total * bm_total;
102 do_div(stddev, (u32)bm_cnt);
103 do_div(stddev, (u32)bm_cnt - 1);
104 } else
105 stddev = 0;
106
107 delta = bm_total;
108 do_div(delta, bm_cnt);
109 avg = delta;
110
111 if (stddev > 0) {
112 int i = 0;
113 /*
114 * stddev is the square of standard deviation but
115 * we want the actually number. Use the average
116 * as our seed to find the std.
117 *
118 * The next try is:
119 * x = (x + N/x) / 2
120 *
121 * Where N is the squared number to find the square
122 * root of.
123 */
124 seed = avg;
125 do {
126 last_seed = seed;
127 seed = stddev;
128 if (!last_seed)
129 break;
130 do_div(seed, last_seed);
131 seed += last_seed;
132 do_div(seed, 2);
133 } while (i++ < 10 && last_seed != seed);
134
135 std = seed;
136 }
137
138 scnprintf(buf: bm_str, BENCHMARK_EVENT_STRLEN,
139 fmt: "last=%llu first=%llu max=%llu min=%llu avg=%u std=%d std^2=%lld",
140 bm_last, bm_first, bm_max, bm_min, avg, std, stddev);
141
142 bm_std = std;
143 bm_avg = avg;
144 bm_stddev = stddev;
145}
146
147static int benchmark_event_kthread(void *arg)
148{
149 /* sleep a bit to make sure the tracepoint gets activated */
150 msleep(msecs: 100);
151
152 while (!kthread_should_stop()) {
153
154 trace_do_benchmark();
155
156 /*
157 * We don't go to sleep, but let others run as well.
158 * This is basically a "yield()" to let any task that
159 * wants to run, schedule in, but if the CPU is idle,
160 * we'll keep burning cycles.
161 *
162 * Note the tasks_rcu_qs() version of cond_resched() will
163 * notify synchronize_rcu_tasks() that this thread has
164 * passed a quiescent state for rcu_tasks. Otherwise
165 * this thread will never voluntarily schedule which would
166 * block synchronize_rcu_tasks() indefinitely.
167 */
168 cond_resched_tasks_rcu_qs();
169 }
170
171 return 0;
172}
173
174/*
175 * When the benchmark tracepoint is enabled, it calls this
176 * function and the thread that calls the tracepoint is created.
177 */
178int trace_benchmark_reg(void)
179{
180 if (!ok_to_run) {
181 pr_warn("trace benchmark cannot be started via kernel command line\n");
182 return -EBUSY;
183 }
184
185 bm_event_thread = kthread_run(benchmark_event_kthread,
186 NULL, "event_benchmark");
187 if (IS_ERR(ptr: bm_event_thread)) {
188 pr_warn("trace benchmark failed to create kernel thread\n");
189 return PTR_ERR(ptr: bm_event_thread);
190 }
191
192 return 0;
193}
194
195/*
196 * When the benchmark tracepoint is disabled, it calls this
197 * function and the thread that calls the tracepoint is deleted
198 * and all the numbers are reset.
199 */
200void trace_benchmark_unreg(void)
201{
202 if (!bm_event_thread)
203 return;
204
205 kthread_stop(k: bm_event_thread);
206 bm_event_thread = NULL;
207
208 strcpy(p: bm_str, q: "START");
209 bm_total = 0;
210 bm_totalsq = 0;
211 bm_last = 0;
212 bm_max = 0;
213 bm_min = 0;
214 bm_cnt = 0;
215 /* These don't need to be reset but reset them anyway */
216 bm_first = 0;
217 bm_std = 0;
218 bm_avg = 0;
219 bm_stddev = 0;
220}
221
222static __init int ok_to_run_trace_benchmark(void)
223{
224 ok_to_run = true;
225
226 return 0;
227}
228
229early_initcall(ok_to_run_trace_benchmark);
230

source code of linux/kernel/trace/trace_benchmark.c