debug mode: only dump queue size stats when exceeding highwater mark

In order to make the queue size stats more useful, and not be
overwhelemd with too much output, this change makes it so that queue
sizes are only printed when they exceed the previous highwater mark.

Also, this adds a slight code cleanup by removing the `#if defined`
around the callsite for `dump_queue_size_stats()`, and the function is
defined as a no-op when `QPTPOOL_QUEUE_SIZE` is undefined.

src/QueuePerThreadPool.c

@@ -118,6 +118,10 @@ struct QPTPool {
     #if defined(DEBUG) && defined(PER_THREAD_STATS)
     struct OutputBuffers *debug_buffers;
     #endif
+
+    #if defined(DEBUG) && defined(QPTPOOL_QUEUE_SIZE)
+    size_t queue_size_highwater;  /* used when printing out queue size stats */
+    #endif
 };
 
 /* struct to pass into pthread_create */
@@ -268,29 +272,42 @@ static void claim_work(QPTPoolThreadData_t *tw) {
     pthread_mutex_unlock(&tw->claimed_mutex);
 }
 
-#if defined(DEBUG) && defined (QPTPOOL_QUEUE_SIZE)
 static void dump_queue_size_stats(QPTPool_t *ctx, QPTPoolThreadData_t *tw) {
+    #if defined(DEBUG) && defined (QPTPOOL_QUEUE_SIZE)
     pthread_mutex_lock(&ctx->mutex);
     pthread_mutex_lock(&print_mutex);
     tw->waiting.size = tw->claimed.size;
 
-    struct timespec now;
-    clock_gettime(CLOCK_MONOTONIC, &now);
-    fprintf(stderr, "qptpool_size %" PRIu64 " ", since_epoch(&now) - epoch);
+    char buf[4096];
+    char *p = buf;
 
     size_t sum = 0;
     for(size_t i = 0; i < ctx->nthreads; i++) {
-        fprintf(stderr, "%zu ", ctx->data[i].waiting.size);
+        p += snprintf(p, sizeof(buf), "%zu ", ctx->data[i].waiting.size);
         sum += ctx->data[i].waiting.size;
-        fprintf(stderr, "%zu ", ctx->data[i].deferred.size);
+        p += snprintf(p, sizeof(buf), "%zu ", ctx->data[i].deferred.size);
         sum += ctx->data[i].deferred.size;
     }
-    fprintf(stderr, "%zu\n", sum);
+
+    if (sum > ctx->queue_size_highwater) {
+        struct timespec now;
+        clock_gettime(CLOCK_MONOTONIC, &now);
+        fprintf(stderr, "qptpool_size %" PRIu64 " ", since_epoch(&now) - epoch);
+
+        fprintf(stderr, "%s", buf);
+
+        ctx->queue_size_highwater = sum;
+        fprintf(stderr, "total size of all queues: %zu\n", sum);
+    }
+
     tw->waiting.size = 0;
     pthread_mutex_unlock(&print_mutex);
     pthread_mutex_unlock(&ctx->mutex);
+    #else
+    (void) ctx;
+    (void) tw;
+    #endif
 }
-#endif
 
 /*
  * process_work() -
@@ -378,9 +395,7 @@ static void *worker_function(void *args) {
         claim_work(tw);
         timestamp_set_end(wf_move_queue);
 
-        #if defined(DEBUG) && defined (QPTPOOL_QUEUE_SIZE)
         dump_queue_size_stats(ctx, tw);
-        #endif
 
         pthread_mutex_unlock(&tw->mutex);
         /* tw->waiting is now empty and can be pushed to */