bionic: benchmarks: Add BM_semaphore_sem_post

This test tries its best to report the producer side underlying futex
wake syscall overhead wthin sem_post. It does not measure the time it
takes for the wakeup to propagate to the consumer. It suffers from
clock_gettime syscall overhead, so subtract that. Lock the CPU speed
for consistent results as we may not reach >50% cpu utilization.

Change-Id: I02fa9dab2e6ac27202f0290115150bd3c8de00f2

diff --git a/benchmarks/semaphore_benchmark.cpp b/benchmarks/semaphore_benchmark.cpp
index a11fcc1..f383da4 100644 (file)
--- a/benchmarks/semaphore_benchmark.cpp
+++ b/benchmarks/semaphore_benchmark.cpp
@@ -16,7 +16,10 @@
 
 #include "benchmark.h"
 
+#include <pthread.h>
 #include <semaphore.h>
+#include <stdatomic.h>
+#include <stdio.h>
 
 static void BM_semaphore_sem_getvalue(int iters) {
   StopBenchmarkTiming();
@@ -47,3 +50,94 @@ static void BM_semaphore_sem_wait_sem_post(int iters) {
   StopBenchmarkTiming();
 }
 BENCHMARK(BM_semaphore_sem_wait_sem_post);
+
+/*
+ *    This test reports the overhead of the underlying futex wake syscall on
+ * the producer. It does not report the overhead from issuing the wake to the
+ * point where the posted consumer thread wakes up. It suffers from
+ * clock_gettime syscall overhead. Lock the CPU speed for consistent results
+ * as we may not reach >50% cpu utilization.
+ *
+ *    We will run a background thread that catches the sem_post wakeup and
+ * loops immediately returning back to sleep in sem_wait for the next one. This
+ * thread is run with policy SCHED_OTHER (normal policy), a middle policy.
+ *
+ *    The primary thread will run at SCHED_IDLE (lowest priority policy) when
+ * monitoring the background thread to detect when it hits sem_wait sleep. It
+ * will do so with no clock running. Once we are ready, we will switch to
+ * SCHED_FIFO (highest priority policy) to time the act of running sem_post
+ * with the benchmark clock running. This ensures nothing else in the system
+ * can preempt our timed activity, including the background thread. We are
+ * also protected with the scheduling policy of letting a process hit a
+ * resource limit rather than get hit with a context switch.
+ *
+ *    The background thread will start executing either on another CPU, or
+ * after we back down from SCHED_FIFO, but certainly not in the context of
+ * the timing of the sem_post.
+ */
+static atomic_int BM_semaphore_sem_post_running;
+
+static void *BM_semaphore_sem_post_start_thread(void *obj) {
+    sem_t *semaphore = reinterpret_cast<sem_t *>(obj);
+
+    while ((BM_semaphore_sem_post_running > 0) && !sem_wait(semaphore)) {
+        ;
+    }
+    BM_semaphore_sem_post_running = -1;
+    return NULL;
+}
+
+static void BM_semaphore_sem_post(int iters) {
+  StopBenchmarkTiming();
+
+  sem_t semaphore;
+  sem_init(&semaphore, 0, 0);
+
+  pthread_attr_t attr;
+  pthread_attr_init(&attr);
+  BM_semaphore_sem_post_running = 1;
+  struct sched_param param = { 0, };
+  pthread_attr_setschedparam(&attr, &param);
+  pthread_attr_setschedpolicy(&attr, SCHED_OTHER);
+  pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED);
+  pthread_t pthread;
+  pthread_create(&pthread, &attr, BM_semaphore_sem_post_start_thread, &semaphore);
+  pthread_attr_destroy(&attr);
+
+  sched_setscheduler((pid_t)0, SCHED_IDLE, &param);
+  for (int i = 0; i < iters; ++i) {
+    int trys = 3, dummy = 0;
+    do {
+      if (BM_semaphore_sem_post_running < 0) {
+        sched_setscheduler((pid_t)0, SCHED_OTHER, &param);
+        fprintf(stderr, "BM_semaphore_sem_post: start_thread died unexpectedly\n");
+        return;
+      }
+      sched_yield();
+      sem_getvalue(&semaphore, &dummy);
+      if (dummy < 0) {  // POSIX.1-2001 possibility 1
+        break;
+      }
+      if (dummy == 0) { // POSIX.1-2001 possibility 2
+        --trys;
+      }
+    } while (trys);
+    param.sched_priority = 1;
+    sched_setscheduler((pid_t)0, SCHED_FIFO, &param);
+    StartBenchmarkTiming();
+    sem_post(&semaphore);
+    StopBenchmarkTiming(); // Remember to subtract clock syscall overhead
+    param.sched_priority = 0;
+    sched_setscheduler((pid_t)0, SCHED_IDLE, &param);
+  }
+  sched_setscheduler((pid_t)0, SCHED_OTHER, &param);
+
+  if (BM_semaphore_sem_post_running > 0) {
+    BM_semaphore_sem_post_running = 0;
+  }
+  do {
+    sem_post(&semaphore);
+    sched_yield();
+  } while (!BM_semaphore_sem_post_running);
+}
+BENCHMARK(BM_semaphore_sem_post);