In-progress work on better cache flushing.

[bluesky.git] / bluesky / cache.c

diff --git a/bluesky/cache.c b/bluesky/cache.c
index 4fe4671..211c2b9 100644 (file)
--- a/bluesky/cache.c
+++ b/bluesky/cache.c
@@ -14,6 +14,7 @@
 #include "bluesky-private.h"
 
 #define WRITEBACK_DELAY (5 * 1000000)
+#define CACHE_CLEAN_DELAY (30 * 1000000)
 
 /* Filesystem caching and cache coherency. */
 
@@ -27,33 +28,80 @@ static void writeback_complete(gpointer a, gpointer i)
     g_mutex_lock(inode->lock);
 
     inode->change_commit = inode->change_pending;
+    inode->change_pending = 0;
     if (inode->change_count == inode->change_commit) {
         /* If inode is no longer dirty... */
         inode->change_time = 0;
-        inode->change_pending = 0;
+        g_mutex_lock(inode->fs->lock);
+        bluesky_list_unlink(&inode->fs->dirty_list, inode->dirty_list);
+        inode->dirty_list = NULL;
+        g_mutex_unlock(inode->fs->lock);
     }
 
     g_mutex_unlock(inode->lock);
 }
 
-static void flushd_inode(gpointer key, gpointer value, gpointer user_data)
+/* Drop cached data for a given inode, if it is clean.  inode must be locked. */
+static void drop_caches(BlueSkyInode *inode)
+{
+    if (inode->type == BLUESKY_REGULAR)
+        bluesky_file_drop_cached(inode);
+}
+
+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) {
+        uint64_t delay = bluesky_get_current_time() - inode->access_time;
+        if (delay >= CACHE_CLEAN_DELAY) {
+            drop_caches(inode);
+
+            /* If the only references are the one we hold and the one in the
+             * filesystem inum->inode hash table...  First check the refcount
+             * without the lock for speed, but if the check looks good verify
+             * it after taking the filesystem lock. */
+            if (inode->refcount == 2) {
+                g_mutex_lock(fs->lock);
+                if (inode->refcount == 2) {
+                    g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
+                          "Trying to drop inode %"PRIu64" from cache",
+                          inode->inum);
+                    if (g_hash_table_remove(fs->inodes, &inode->inum))
+                        bluesky_inode_unref(inode);
+                }
+                bluesky_list_unlink(&inode->fs->accessed_list,
+                                    inode->accessed_list);
+                inode->accessed_list = NULL;
+                bluesky_list_unlink(&inode->fs->dirty_list,
+                                    inode->dirty_list);
+                inode->dirty_list = NULL;
+                g_mutex_unlock(fs->lock);
+            }
+        }
+
+        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;
     }
 
@@ -73,15 +121,37 @@ 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)
 {
+    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);
 }