Fix a deadlock and a few memory leaks.

[bluesky.git] / bluesky / cache.c

diff --git a/bluesky/cache.c b/bluesky/cache.c
index f953a64..979c917 100644 (file)
--- a/bluesky/cache.c
+++ b/bluesky/cache.c
@@ -13,35 +13,106 @@
 
 #include "bluesky-private.h"
 
+#define WRITEBACK_DELAY (5 * 1000000)
+
 /* Filesystem caching and cache coherency. */
 
-static void flushd_inode(gpointer key, gpointer value, gpointer user_data)
+static void writeback_complete(gpointer a, gpointer i)
+{
+    BlueSkyInode *inode = (BlueSkyInode *)i;
+
+    g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
+          "Writeback for inode %"PRIu64" complete", inode->inum);
+
+    g_mutex_lock(inode->lock);
+
+    inode->change_commit = inode->change_pending;
+    if (inode->change_count == inode->change_commit) {
+        /* If inode is no longer dirty... */
+        inode->change_time = 0;
+        inode->change_pending = 0;
+    }
+
+    g_mutex_unlock(inode->lock);
+}
+
+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;
+    }
+
+    inode->change_pending = inode->change_count;
 
     g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
           "Starting flush of inode %"PRIu64, inode->inum);
 
-    if (inode->type == BLUESKY_REGULAR)
-        bluesky_file_flush(inode);
-    bluesky_inode_flush(fs, inode);
+    /* Create a store barrier.  All operations part of the writeback will be
+     * added to this barrier, so when the barrier completes we know that the
+     * writeback is finished. */
+    BlueSkyStoreAsync *barrier = bluesky_store_async_new(fs->store);
+    barrier->op = STORE_OP_BARRIER;
+
+    bluesky_inode_start_sync(inode, barrier);
+
+    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);
 }