X-Git-Url: http://git.vrable.net/?a=blobdiff_plain;f=bluesky%2Fcache.c;h=979c9176e33b15f24cfd69a49975a12300f0fa33;hb=79fd07ce3738aceebc08fd363c14795ae9353057;hp=912c5fdff743b34638f3c369dc2e86a1b67555aa;hpb=a5b5871865301712bc22ffc3d86a425dc0b45991;p=bluesky.git

diff --git a/bluesky/cache.c b/bluesky/cache.c
index 912c5fd..979c917 100644
--- a/bluesky/cache.c
+++ b/bluesky/cache.c
@@ -36,24 +36,33 @@ static void writeback_complete(gpointer a, gpointer i)
     g_mutex_unlock(inode->lock);
 }
 
-static void flushd_inode(gpointer key, gpointer value, gpointer user_data)
+static void flushd_inode(gpointer value, gpointer user_data)
 {
     BlueSkyFS *fs = (BlueSkyFS *)user_data;
 
     BlueSkyInode *inode = (BlueSkyInode *)value;
 
-    if (inode->change_count == inode->change_commit)
+    g_mutex_lock(inode->lock);
+
+    if (inode->change_count == inode->change_commit) {
+        g_mutex_unlock(inode->lock);
+        bluesky_inode_unref(inode);
         return;
+    }
 
     if (inode->change_pending) {
         /* Waiting for an earlier writeback to finish, so don't start a new
          * writeback yet. */
+        g_mutex_unlock(inode->lock);
+        bluesky_inode_unref(inode);
         return;
     }
 
     uint64_t elapsed = bluesky_get_current_time() - inode->change_time;
     if (elapsed < WRITEBACK_DELAY) {
         /* Give a bit more time before starting writeback. */
+        g_mutex_unlock(inode->lock);
+        bluesky_inode_unref(inode);
         return;
     }
 
@@ -73,18 +82,37 @@ static void flushd_inode(gpointer key, gpointer value, gpointer user_data)
     bluesky_store_async_add_notifier(barrier, writeback_complete, inode);
     bluesky_store_async_submit(barrier);
     bluesky_store_async_unref(barrier);
+
+    g_mutex_unlock(inode->lock);
+    bluesky_inode_unref(inode);
 }
 
 /* Scan through the cache for dirty data and start flushing it to stable
  * storage.  This does not guarantee that data is committed when it returns.
  * Instead, this can be called occasionally to ensure that dirty data is
- * gradually flushed. */
+ * gradually flushed.
+ *
+ * We do not want to hold the filesystem lock while flushing individual inodes,
+ * a that could lead to deadlock.  So first scan through the inode table to get
+ * a reference to all inodes, then process that queue of inodes after dropping
+ * the filesystem lock. */
+static void gather_inodes(gpointer key, gpointer value, gpointer user_data)
+{
+    GSList **list = (GSList **)user_data;
+    *list = g_slist_prepend(*list, value);
+    bluesky_inode_ref((BlueSkyInode *)value);
+}
+
 void bluesky_flushd_invoke(BlueSkyFS *fs)
 {
-    g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
-          "Writeout process invoked");
+    GSList *list = NULL;
 
     g_mutex_lock(fs->lock);
-    g_hash_table_foreach(fs->inodes, flushd_inode, fs);
+    g_hash_table_foreach(fs->inodes, gather_inodes, &list);
     g_mutex_unlock(fs->lock);
+
+    list = g_slist_reverse(list);
+    g_slist_foreach(list, flushd_inode, fs);
+
+    g_slist_free(list);
 }