Implement new scheme for retaining needed journal segments.

[bluesky.git] / bluesky / cache.c

diff --git a/bluesky/cache.c b/bluesky/cache.c
index 50c0f36..a9fda5b 100644 (file)
--- a/bluesky/cache.c
+++ b/bluesky/cache.c
@@ -101,15 +101,21 @@ static void flushd_dirty(BlueSkyFS *fs)
     g_mutex_unlock(fs->lock);
 }
 
-/* Try to flush dirty data to the cloud.
- * TODO: Rewrite this to work on cloud log items rather than inodes, so we can
- * better track which logs are fully synchronized to the cloud and can be
- * garbage collected if needed? */
+/* Try to flush dirty data to the cloud.  This will take a snapshot of the
+ * entire filesystem (though only point-in-time consistent for isolated inodes
+ * and not the filesystem as a whole) and ensure all data is written to the
+ * cloud.  When the write completes, we will allow old journal segments (those
+ * that were fully written _before_ the snapshot process started) to be garbage
+ * collected.  Newer journal segments can't be collected yet since they may
+ * still contain data which has not been written persistently to the cloud. */
 static void flushd_cloud(BlueSkyFS *fs)
 {
-    int64_t start_time = bluesky_get_current_time();
     g_mutex_lock(fs->lock);
 
+    /* TODO: Locking?  Since we're reading a single variable this is probably
+     * atomic but a lock could be safer. */
+    int journal_seq_start = fs->log->seq_num;
+
     while (1) {
         BlueSkyInode *inode;
         if (fs->dirty_list.prev == NULL)
@@ -121,16 +127,6 @@ static void flushd_cloud(BlueSkyFS *fs)
                   "Flushing inode %"PRIu64" to cloud", inode->inum);
         }
 
-        /* Stop processing dirty inodes if we both have enough memory available
-         * and the oldest inode is sufficiently new that it need not be flushed
-         * out. */
-        uint64_t elapsed = bluesky_get_current_time() - inode->change_time;
-        if (g_atomic_int_get(&fs->cache_dirty) < bluesky_watermark_low_dirty
-                && elapsed < WRITEBACK_DELAY)
-            break;
-        if (inode->change_time > start_time)
-            break;
-
         bluesky_inode_ref(inode);
 
         g_mutex_unlock(fs->lock);
@@ -156,6 +152,30 @@ static void flushd_cloud(BlueSkyFS *fs)
 
     g_mutex_unlock(fs->lock);
     bluesky_cloudlog_flush(fs);
+
+    /* Wait until all segments have been written to the cloud, so that it
+     * becomes safe to free up journal segments. */
+    while (fs->log_state->pending_segments != NULL) {
+        SerializedRecord *segment
+            = (SerializedRecord *)fs->log_state->pending_segments->data;
+        g_mutex_lock(segment->lock);
+        while (!segment->complete)
+            g_cond_wait(segment->cond, segment->lock);
+        g_mutex_unlock(segment->lock);
+
+        g_mutex_free(segment->lock);
+        g_cond_free(segment->cond);
+        g_free(segment);
+
+        fs->log_state->pending_segments
+            = g_list_delete_link(fs->log_state->pending_segments,
+                                 fs->log_state->pending_segments);
+    }
+
+    g_print("All segments have been flushed, journal < %d is clean\n",
+            journal_seq_start);
+
+    fs->log->journal_watermark = journal_seq_start;
 }
 
 /* Drop cached data for a given inode, if it is clean.  inode must be locked. */