Try out a fix for timeouts with limited bandwidth to cloud.

[bluesky.git] / bluesky / cache.c

diff --git a/bluesky/cache.c b/bluesky/cache.c
index 16e3d81..cdd5c65 100644 (file)
--- a/bluesky/cache.c
+++ b/bluesky/cache.c
@@ -421,9 +421,68 @@ void bluesky_flushd_invoke(BlueSkyFS *fs)
     g_thread_create((GThreadFunc)flushd_task, fs, FALSE, NULL);
 }
 
+/* How urgent is flushing out data?  Returns one of several values:
+ *   0 - memory state is fine
+ *   1 - should launch flushd if not already running
+ *   2 - should delay writers while memory frees up
+ *   3 - should block writers until memory frees up
+ */
+static int compute_pressure(BlueSkyFS *fs)
+{
+    /* LEVEL 3 */
+    /* Too much dirty data in memory? */
+    if (g_atomic_int_get(&fs->cache_dirty)
+                + g_atomic_int_get(&fs->cache_log_dirty)
+           > bluesky_watermark_high_dirty)
+    {
+        g_print("pressure: too much dirty data (3)\n");
+        return 3;
+    }
+
+    /* Too much uncommitted data in the journal on disk, not yet flushed to the
+     * cloud? */
+    /*printf("Dirty journals: %d to %d\n",
+           fs->log->journal_watermark, fs->log->seq_num);*/
+    int dirty_limit;
+    dirty_limit = bluesky_options.cache_size / (LOG_SEGMENT_SIZE / 1024) / 2;
+    int dirty_journals = fs->log->seq_num - fs->log->journal_watermark + 1;
+    if (dirty_journals > 1 && dirty_journals >= dirty_limit) {
+        printf("pressure: too many dirty journals (%d >= %d) (3)\n",
+               dirty_journals, dirty_limit);
+        return 3;
+    }
+
+    /* LEVEL 2 */
+    if (g_atomic_int_get(&fs->cache_dirty) > bluesky_watermark_med2_dirty) {
+        g_print("pressure: too much dirty data (2)\n");
+        return 2;
+    }
+
+    if (dirty_journals > 1 && dirty_journals > dirty_limit * 3 / 4) {
+        printf("pressure: many dirty journals (%d), should start writeback (2)\n",
+               dirty_journals);
+        return 2;
+    }
+
+    /* LEVEL 1 */
+    if (g_atomic_int_get(&fs->cache_dirty) > bluesky_watermark_medium_dirty) {
+        g_print("pressure: too much dirty data (1)\n");
+        return 1;
+    }
+
+    if (dirty_journals > 1 && dirty_journals > dirty_limit / 2) {
+        printf("pressure: many dirty journals (%d), should start writeback (1)\n",
+               dirty_journals);
+        return 1;
+    }
+
+    return 0;
+}
+
 void bluesky_flushd_invoke_conditional(BlueSkyFS *fs)
 {
-    if (g_atomic_int_get(&fs->cache_dirty) < bluesky_watermark_medium_dirty)
+    int pressure = compute_pressure(fs);
+    if (pressure == 0)
         return;
 
     if (bluesky_verbose) {
@@ -436,16 +495,26 @@ void bluesky_flushd_invoke_conditional(BlueSkyFS *fs)
 
     /* If the system is under heavy memory pressure, actually delay execution
      * so the flush daemon can catch up. */
-    while (g_atomic_int_get(&fs->cache_dirty)
-                + g_atomic_int_get(&fs->cache_log_dirty)
-           > bluesky_watermark_high_dirty) {
+    while (pressure > 1) {
         g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
               "Waiting due to memory pressure, dirty=%d + %d",
               g_atomic_int_get(&fs->cache_dirty),
               g_atomic_int_get(&fs->cache_log_dirty));
         g_mutex_lock(fs->lock);
-        g_cond_wait(fs->flushd_cond, fs->lock);
+        pressure = compute_pressure(fs);
+        if (pressure > 2) {
+            g_cond_wait(fs->flushd_cond, fs->lock);
+        } else if (pressure > 1) {
+            /* Wait for up to 10 seconds. */
+            GTimeVal timeout;
+            g_get_current_time(&timeout);
+            g_time_val_add(&timeout, 10 * 1000 * 1000);
+            g_cond_timed_wait(fs->flushd_cond, fs->lock, &timeout);
+        }
         g_mutex_unlock(fs->lock);
+        pressure = compute_pressure(fs);
+        if (pressure == 1)
+            break;  /* Do not loop indefinitely for a pressure of 1 */
     }
 }