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)
17 #define CACHE_CLEAN_DELAY (30 * 1000000)
19 /* Filesystem caching and cache coherency. */
21 static void writeback_complete(gpointer a, gpointer i)
23 BlueSkyInode *inode = (BlueSkyInode *)i;
25 g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
26 "Writeback for inode %"PRIu64" complete", inode->inum);
28 g_mutex_lock(inode->lock);
30 inode->change_commit = inode->change_pending;
31 inode->change_pending = 0;
32 if (inode->change_count == inode->change_commit) {
33 /* If inode is no longer dirty... */
34 inode->change_time = 0;
35 g_mutex_lock(inode->fs->lock);
36 bluesky_list_unlink(&inode->fs->dirty_list, inode->dirty_list);
37 inode->dirty_list = NULL;
38 g_mutex_unlock(inode->fs->lock);
41 g_mutex_unlock(inode->lock);
44 /* Drop cached data for a given inode, if it is clean. inode must be locked. */
45 static void drop_caches(BlueSkyInode *inode)
47 if (inode->type == BLUESKY_REGULAR)
48 bluesky_file_drop_cached(inode);
51 static void flushd_inode(gpointer value, gpointer user_data)
53 BlueSkyFS *fs = (BlueSkyFS *)user_data;
55 BlueSkyInode *inode = (BlueSkyInode *)value;
57 g_mutex_lock(inode->lock);
59 if (inode->change_count == inode->change_commit) {
60 uint64_t delay = bluesky_get_current_time() - inode->access_time;
61 if (delay >= CACHE_CLEAN_DELAY) {
64 /* If the only references are the one we hold and the one in the
65 * filesystem inum->inode hash table... First check the refcount
66 * without the lock for speed, but if the check looks good verify
67 * it after taking the filesystem lock. */
68 if (inode->refcount == 2) {
69 g_mutex_lock(fs->lock);
70 if (inode->refcount == 2) {
71 g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
72 "Trying to drop inode %"PRIu64" from cache",
74 if (g_hash_table_remove(fs->inodes, &inode->inum))
75 bluesky_inode_unref(inode);
77 g_mutex_unlock(fs->lock);
81 g_mutex_unlock(inode->lock);
82 bluesky_inode_unref(inode);
86 if (inode->change_pending) {
87 /* Waiting for an earlier writeback to finish, so don't start a new
89 g_mutex_unlock(inode->lock);
90 bluesky_inode_unref(inode);
94 uint64_t elapsed = bluesky_get_current_time() - inode->change_time;
95 if (elapsed < WRITEBACK_DELAY) {
96 /* Give a bit more time before starting writeback. */
97 g_mutex_unlock(inode->lock);
98 bluesky_inode_unref(inode);
102 inode->change_pending = inode->change_count;
104 g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
105 "Starting flush of inode %"PRIu64, inode->inum);
107 /* Create a store barrier. All operations part of the writeback will be
108 * added to this barrier, so when the barrier completes we know that the
109 * writeback is finished. */
110 BlueSkyStoreAsync *barrier = bluesky_store_async_new(fs->store);
111 barrier->op = STORE_OP_BARRIER;
113 bluesky_inode_start_sync(inode, barrier);
115 bluesky_store_async_add_notifier(barrier, writeback_complete, inode);
116 bluesky_store_async_submit(barrier);
117 bluesky_store_async_unref(barrier);
119 g_mutex_unlock(inode->lock);
120 bluesky_inode_unref(inode);
123 /* Scan through the cache for dirty data and start flushing it to stable
124 * storage. This does not guarantee that data is committed when it returns.
125 * Instead, this can be called occasionally to ensure that dirty data is
128 * We do not want to hold the filesystem lock while flushing individual inodes,
129 * a that could lead to deadlock. So first scan through the inode table to get
130 * a reference to all inodes, then process that queue of inodes after dropping
131 * the filesystem lock. */
132 static void gather_inodes(gpointer key, gpointer value, gpointer user_data)
134 GSList **list = (GSList **)user_data;
135 *list = g_slist_prepend(*list, value);
136 bluesky_inode_ref((BlueSkyInode *)value);
139 void bluesky_flushd_invoke(BlueSkyFS *fs)
143 g_mutex_lock(fs->lock);
144 g_hash_table_foreach(fs->inodes, gather_inodes, &list);
145 g_mutex_unlock(fs->lock);
147 list = g_slist_reverse(list);
148 g_slist_foreach(list, flushd_inode, fs);