1 /* Blue Sky: File Systems in the Cloud
3 * Copyright (C) 2009 The Regents of the University of California
4 * Written by Michael Vrable <mvrable@cs.ucsd.edu>
14 #include "bluesky-private.h"
16 #define WRITEBACK_DELAY (5 * 1000000)
18 /* Filesystem caching and cache coherency. */
20 static void writeback_complete(gpointer a, gpointer i)
22 BlueSkyInode *inode = (BlueSkyInode *)i;
24 g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
25 "Writeback for inode %"PRIu64" complete", inode->inum);
27 g_mutex_lock(inode->lock);
29 inode->change_commit = inode->change_pending;
30 if (inode->change_count == inode->change_commit) {
31 /* If inode is no longer dirty... */
32 inode->change_time = 0;
33 inode->change_pending = 0;
36 g_mutex_unlock(inode->lock);
39 static void flushd_inode(gpointer value, gpointer user_data)
41 BlueSkyFS *fs = (BlueSkyFS *)user_data;
43 BlueSkyInode *inode = (BlueSkyInode *)value;
45 g_mutex_lock(inode->lock);
47 if (inode->change_count == inode->change_commit) {
48 g_mutex_unlock(inode->lock);
49 bluesky_inode_unref(inode);
53 if (inode->change_pending) {
54 /* Waiting for an earlier writeback to finish, so don't start a new
56 g_mutex_unlock(inode->lock);
57 bluesky_inode_unref(inode);
61 uint64_t elapsed = bluesky_get_current_time() - inode->change_time;
62 if (elapsed < WRITEBACK_DELAY) {
63 /* Give a bit more time before starting writeback. */
64 g_mutex_unlock(inode->lock);
65 bluesky_inode_unref(inode);
69 inode->change_pending = inode->change_count;
71 g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
72 "Starting flush of inode %"PRIu64, inode->inum);
74 /* Create a store barrier. All operations part of the writeback will be
75 * added to this barrier, so when the barrier completes we know that the
76 * writeback is finished. */
77 BlueSkyStoreAsync *barrier = bluesky_store_async_new(fs->store);
78 barrier->op = STORE_OP_BARRIER;
80 bluesky_inode_start_sync(inode, barrier);
82 bluesky_store_async_add_notifier(barrier, writeback_complete, inode);
83 bluesky_store_async_submit(barrier);
84 bluesky_store_async_unref(barrier);
86 g_mutex_unlock(inode->lock);
87 bluesky_inode_unref(inode);
90 /* Scan through the cache for dirty data and start flushing it to stable
91 * storage. This does not guarantee that data is committed when it returns.
92 * Instead, this can be called occasionally to ensure that dirty data is
95 * We do not want to hold the filesystem lock while flushing individual inodes,
96 * a that could lead to deadlock. So first scan through the inode table to get
97 * a reference to all inodes, then process that queue of inodes after dropping
98 * the filesystem lock. */
99 static void gather_inodes(gpointer key, gpointer value, gpointer user_data)
101 GSList **list = (GSList **)user_data;
102 *list = g_slist_prepend(*list, value);
103 bluesky_inode_ref((BlueSkyInode *)value);
106 void bluesky_flushd_invoke(BlueSkyFS *fs)
110 g_mutex_lock(fs->lock);
111 g_hash_table_foreach(fs->inodes, gather_inodes, &list);
112 g_mutex_unlock(fs->lock);
114 list = g_slist_reverse(list);
115 g_slist_foreach(list, flushd_inode, fs);