improved perf_test to measure steady state behavior

[nbds] / runtime / rcu.c

diff --git a/runtime/rcu.c b/runtime/rcu.c
index 50205e8670173c2d2ab55089ae1cac4b4bb80901..12c37a3d4e37df2d5a340010ed039cde7efc1b3c 100644 (file)
--- a/runtime/rcu.c
+++ b/runtime/rcu.c
@@ -1,4 +1,4 @@
-/* 
+/*
  * Written by Josh Dybnis and released to the public domain, as explained at
  * http://creativecommons.org/licenses/publicdomain
  *
@@ -31,7 +31,7 @@ static fifo_t *pending_[MAX_NUM_THREADS] = {};
 static int num_threads_ = 0;
 
 static fifo_t *fifo_alloc(int scale) {
-    fifo_t *q = (fifo_t *)nbd_malloc(sizeof(fifo_t) + (1 << scale) * sizeof(void *)); 
+    fifo_t *q = (fifo_t *)nbd_malloc(sizeof(fifo_t) + (1ULL << scale) * sizeof(void *));
     memset(q, 0, sizeof(fifo_t));
     q->scale = scale;
     q->head = 0;
@@ -43,7 +43,7 @@ void rcu_thread_init (int id) {
     assert(id < MAX_NUM_THREADS);
     if (pending_[id] == NULL) {
         pending_[id] = fifo_alloc(RCU_QUEUE_SCALE);
-        SYNC_ADD(&num_threads_, 1);
+        (void)SYNC_ADD(&num_threads_, 1);
     }
 }
 
@@ -51,6 +51,7 @@ void rcu_update (void) {
     LOCALIZE_THREAD_LOCAL(tid_, int);
     assert(tid_ < num_threads_);
     int next_thread_id = (tid_ + 1) % num_threads_;
+    TRACE("r1", "rcu_update: updating thread %llu", next_thread_id, 0);
     int i;
     for (i = 0; i < num_threads_; ++i) {
         if (i == tid_)
@@ -62,13 +63,16 @@ void rcu_update (void) {
 
         uint64_t x = rcu_[tid_][i];
         rcu_[next_thread_id][i] = rcu_last_posted_[tid_][i] = x;
+        TRACE("r2", "rcu_update: posted updated value (%llu) for thread %llu", x, i);
     }
 
     // free
-    while (pending_[tid_]->tail != rcu_[tid_][tid_]) {
-        fifo_t *q = pending_[tid_];
-        uint32_t i = MOD_SCALE(q->tail++, q->scale);
+    fifo_t *q = pending_[tid_];
+    while (q->tail != rcu_[tid_][tid_]) {
+        uint32_t i = MOD_SCALE(q->tail, q->scale);
+        TRACE("r0", "rcu_update: freeing %p from queue at position %llu", q->x[i], q->tail);
         nbd_free(q->x[i]);
+        q->tail++;
     }
 }
 
@@ -77,13 +81,14 @@ void rcu_defer_free (void *x) {
     LOCALIZE_THREAD_LOCAL(tid_, int);
     fifo_t *q = pending_[tid_];
     assert(MOD_SCALE(q->head + 1, q->scale) != MOD_SCALE(q->tail, q->scale));
-    uint32_t i = MOD_SCALE(q->head++, q->scale);
+    uint32_t i = MOD_SCALE(q->head, q->scale);
     q->x[i] = x;
-    TRACE("r0", "rcu_defer_free: put %p on queue at position %llu", x, pending_[tid_]->head);
+    TRACE("r0", "rcu_defer_free: put %p on queue at position %llu", x, q->head);
+    q->head++;
 
-    if (pending_[tid_]->head - rcu_last_posted_[tid_][tid_] < RCU_POST_THRESHOLD)
-        return;
-    TRACE("r0", "rcu_defer_free: posting %llu", pending_[tid_]->head, 0);
-    int next_thread_id = (tid_ + 1) % num_threads_;
-    rcu_[next_thread_id][tid_] = rcu_last_posted_[tid_][tid_] = pending_[tid_]->head;
+    if (pending_[tid_]->head - rcu_last_posted_[tid_][tid_] >= RCU_POST_THRESHOLD) {
+        TRACE("r0", "rcu_defer_free: posting %llu", pending_[tid_]->head, 0);
+        int next_thread_id = (tid_ + 1) % num_threads_;
+        rcu_[next_thread_id][tid_] = rcu_last_posted_[tid_][tid_] = pending_[tid_]->head;
+    }
 }