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 s64 bm_cnt; 20 21 /* 22 * This gets called in a loop recording the time it took to write 23 * the tracepoint. What it writes is the time statistics of the last 24 * tracepoint write. As there is nothing to write the first time 25 * it simply writes "START". As the first write is cold cache and 26 * the rest is hot, we save off that time in bm_first and it is 27 * reported as "first", which is shown in the second write to the 28 * tracepoint. The "first" field is writen within the statics from 29 * then on but never changes. 30 */ 31 static void trace_do_benchmark(void) 32 { 33 u64 start; 34 u64 stop; 35 u64 delta; 36 u64 stddev; 37 u64 seed; 38 u64 last_seed; 39 unsigned int avg; 40 unsigned int std = 0; 41 42 /* Only run if the tracepoint is actually active */ 43 if (!trace_benchmark_event_enabled()) 44 return; 45 46 local_irq_disable(); 47 start = trace_clock_local(); 48 trace_benchmark_event(bm_str); 49 stop = trace_clock_local(); 50 local_irq_enable(); 51 52 bm_cnt++; 53 54 delta = stop - start; 55 56 /* 57 * The first read is cold cached, keep it separate from the 58 * other calculations. 59 */ 60 if (bm_cnt == 1) { 61 bm_first = delta; 62 scnprintf(bm_str, BENCHMARK_EVENT_STRLEN, 63 "first=%llu [COLD CACHED]", bm_first); 64 return; 65 } 66 67 bm_last = delta; 68 69 bm_total += delta; 70 bm_totalsq += delta * delta; 71 72 if (delta > bm_max) 73 bm_max = delta; 74 if (!bm_min || delta < bm_min) 75 bm_min = delta; 76 77 if (bm_cnt > 1) { 78 /* 79 * Apply Welford's method to calculate standard deviation: 80 * s^2 = 1 / (n * (n-1)) * (n * \Sum (x_i)^2 - (\Sum x_i)^2) 81 */ 82 stddev = (u64)bm_cnt * bm_totalsq - bm_total * bm_total; 83 do_div(stddev, bm_cnt); 84 do_div(stddev, bm_cnt - 1); 85 } else 86 stddev = 0; 87 88 delta = bm_total; 89 do_div(delta, bm_cnt); 90 avg = delta; 91 92 if (stddev > 0) { 93 int i = 0; 94 /* 95 * stddev is the square of standard deviation but 96 * we want the actualy number. Use the average 97 * as our seed to find the std. 98 * 99 * The next try is: 100 * x = (x + N/x) / 2 101 * 102 * Where N is the squared number to find the square 103 * root of. 104 */ 105 seed = avg; 106 do { 107 last_seed = seed; 108 seed = stddev; 109 if (!last_seed) 110 break; 111 do_div(seed, last_seed); 112 seed += last_seed; 113 do_div(seed, 2); 114 } while (i++ < 10 && last_seed != seed); 115 116 std = seed; 117 } 118 119 scnprintf(bm_str, BENCHMARK_EVENT_STRLEN, 120 "last=%llu first=%llu max=%llu min=%llu avg=%u std=%d std^2=%lld", 121 bm_last, bm_first, bm_max, bm_min, avg, std, stddev); 122 } 123 124 static int benchmark_event_kthread(void *arg) 125 { 126 /* sleep a bit to make sure the tracepoint gets activated */ 127 msleep(100); 128 129 while (!kthread_should_stop()) { 130 131 trace_do_benchmark(); 132 133 /* 134 * We don't go to sleep, but let others 135 * run as well. 136 */ 137 cond_resched(); 138 } 139 140 return 0; 141 } 142 143 /* 144 * When the benchmark tracepoint is enabled, it calls this 145 * function and the thread that calls the tracepoint is created. 146 */ 147 void trace_benchmark_reg(void) 148 { 149 bm_event_thread = kthread_run(benchmark_event_kthread, 150 NULL, "event_benchmark"); 151 WARN_ON(!bm_event_thread); 152 } 153 154 /* 155 * When the benchmark tracepoint is disabled, it calls this 156 * function and the thread that calls the tracepoint is deleted 157 * and all the numbers are reset. 158 */ 159 void trace_benchmark_unreg(void) 160 { 161 if (!bm_event_thread) 162 return; 163 164 kthread_stop(bm_event_thread); 165 166 strcpy(bm_str, "START"); 167 bm_total = 0; 168 bm_totalsq = 0; 169 bm_last = 0; 170 bm_max = 0; 171 bm_min = 0; 172 bm_cnt = 0; 173 /* bm_first doesn't need to be reset but reset it anyway */ 174 bm_first = 0; 175 } 176