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