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>
15 #include "bluesky-private.h"
17 /* Core filesystem. Different proxies, such as the NFSv3 one, interface to
18 * this, but the core actually tracks the data which is stored. So far we just
19 * implement an in-memory filesystem, but eventually this will be state which
20 * is persisted to the cloud. */
22 /* Return the current time, in microseconds since the epoch. */
23 int64_t bluesky_get_current_time()
26 g_get_current_time(&t);
27 return (int64_t)t.tv_sec * 1000000 + t.tv_usec;
30 /* Update an inode to indicate that a modification was made. This increases
31 * the change counter, updates the ctime to the current time, and optionally
32 * updates the mtime. This also makes the inode contents subject to writeback
33 * to storage in the future. inode must already be locked. */
34 void bluesky_inode_update_ctime(BlueSkyInode *inode, gboolean update_mtime)
36 int64_t now = bluesky_get_current_time();
37 inode->change_count++;
42 if (inode->change_time == 0)
43 inode->change_time = now;
46 if (bluesky_options.writethrough_cache)
47 bluesky_file_flush(inode, NULL);
50 g_mutex_lock(inode->fs->lock);
51 bluesky_list_unlink(&inode->fs->unlogged_list, inode->unlogged_list);
52 inode->unlogged_list = bluesky_list_prepend(&inode->fs->unlogged_list, inode);
53 bluesky_list_unlink(&inode->fs->accessed_list, inode->accessed_list);
54 inode->accessed_list = bluesky_list_prepend(&inode->fs->accessed_list, inode);
55 g_mutex_unlock(inode->fs->lock);
58 /* Unfortunately a glib hash table is only guaranteed to be able to store
59 * 32-bit keys if we use the key directly. If we want 64-bit inode numbers,
60 * we'll have to allocate memory to store the 64-bit inumber, and use a pointer
61 * to it. Rather than allocate the memory for the key, we'll just include a
62 * pointer to the 64-bit inum stored in the inode itself, so that we don't need
63 * to do any more memory management. */
64 static guint bluesky_fs_key_hash_func(gconstpointer key)
66 uint64_t inum = *(const uint64_t *)key;
70 static gboolean bluesky_fs_key_equal_func(gconstpointer a, gconstpointer b)
72 uint64_t i1 = *(const uint64_t *)a;
73 uint64_t i2 = *(const uint64_t *)b;
77 /* Filesystem-level operations. A filesystem is like a directory tree that we
78 * are willing to export. */
79 BlueSkyFS *bluesky_new_fs(gchar *name)
81 BlueSkyFS *fs = g_new0(BlueSkyFS, 1);
82 fs->lock = g_mutex_new();
83 fs->name = g_strdup(name);
84 fs->inodes = g_hash_table_new(bluesky_fs_key_hash_func,
85 bluesky_fs_key_equal_func);
86 fs->next_inum = BLUESKY_ROOT_INUM + 1;
87 fs->store = bluesky_store_new("file");
88 fs->flushd_lock = g_mutex_new();
89 fs->flushd_cond = g_cond_new();
90 fs->locations = g_hash_table_new(bluesky_cloudlog_hash,
91 bluesky_cloudlog_equal);
92 fs->inode_map = g_sequence_new(NULL);
94 fs->log_state = g_new0(BlueSkyCloudLogState, 1);
95 fs->log_state->data = g_string_new("");
100 BlueSkyFS *bluesky_init_fs(gchar *name, BlueSkyStore *store)
102 BlueSkyFS *fs = bluesky_new_fs(name);
104 fs->log = bluesky_log_new("journal");
107 if (bluesky_checkpoint_load(fs)) {
108 g_print("Filesystem checkpoint loaded, starting journal replay...\n");
110 g_print("Journal replay complete, filesystem ready.\n");
112 /* Initialize a fresh filesystem */
113 g_print("Initializing new filesystem...\n");
114 BlueSkyInode *root = bluesky_new_inode(BLUESKY_ROOT_INUM, fs,
118 bluesky_insert_inode(fs, root);
119 bluesky_inode_update_ctime(root, TRUE);
120 bluesky_inode_do_sync(root);
126 /* Inode reference counting. */
127 void bluesky_inode_ref(BlueSkyInode *inode)
129 g_atomic_int_inc(&inode->refcount);
132 /* Free most of the resources used by an inode structure, but do not free the
133 * inode itself. Can be used if the inode data will be reloaded from
134 * serialized form to clear out old information first. */
135 void bluesky_inode_free_resources(BlueSkyInode *inode)
137 switch (inode->type) {
138 case BLUESKY_REGULAR:
139 if (inode->blocks != NULL) {
140 for (int i = 0; i < inode->blocks->len; i++) {
141 BlueSkyBlock *b = &g_array_index(inode->blocks,
143 if (b->type == BLUESKY_BLOCK_DIRTY) {
144 g_error("Deleting an inode with dirty file data!");
146 bluesky_cloudlog_unref(b->ref);
147 bluesky_string_unref(b->dirty);
149 g_array_unref(inode->blocks);
150 inode->blocks = NULL;
154 case BLUESKY_DIRECTORY:
155 if (inode->dirhash != NULL)
156 g_hash_table_destroy(inode->dirhash);
157 inode->dirhash = NULL;
158 if (inode->dirhash_folded != NULL)
159 g_hash_table_destroy(inode->dirhash_folded);
160 inode->dirhash_folded = NULL;
161 if (inode->dirents != NULL)
162 g_sequence_free(inode->dirents);
163 inode->dirents = NULL;
166 case BLUESKY_SYMLINK:
167 g_free(inode->symlink_contents);
168 inode->symlink_contents = NULL;
176 void bluesky_inode_unref(BlueSkyInode *inode)
178 if (g_atomic_int_dec_and_test(&inode->refcount)) {
179 if (bluesky_verbose) {
180 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
181 "Reference count for inode %"PRIu64" dropped to zero.",
185 /* Sanity check: Is the inode clean? */
186 if (inode->change_commit < inode->change_count
187 || inode->accessed_list != NULL
188 || inode->unlogged_list != NULL
189 || inode->dirty_list != NULL) {
190 g_warning("Dropping inode which is not clean (commit %"PRIi64" < change %"PRIi64"; accessed_list = %p; dirty_list = %p)\n", inode->change_commit, inode->change_count, inode->accessed_list, inode->dirty_list);
193 /* These shouldn't be needed, but in case the above warning fires and
194 * we delete the inode anyway, we ought to be sure the inode is not on
196 g_mutex_lock(inode->fs->lock);
197 bluesky_list_unlink(&inode->fs->accessed_list, inode->accessed_list);
198 bluesky_list_unlink(&inode->fs->dirty_list, inode->dirty_list);
199 bluesky_list_unlink(&inode->fs->unlogged_list, inode->unlogged_list);
200 g_mutex_unlock(inode->fs->lock);
202 bluesky_inode_free_resources(inode);
204 g_mutex_free(inode->lock);
209 /* Allocate a fresh inode number which has not been used before within a
210 * filesystem. fs must already be locked. */
211 uint64_t bluesky_fs_alloc_inode(BlueSkyFS *fs)
215 inum = fs->next_inum;
221 /* Perform type-specification initialization of an inode. Normally performed
222 * in bluesky_new_inode, but can be separated if an inode is created first,
223 * then deserialized. */
224 void bluesky_init_inode(BlueSkyInode *i, BlueSkyFileType type)
229 case BLUESKY_REGULAR:
230 i->blocks = g_array_new(FALSE, TRUE, sizeof(BlueSkyBlock));
232 case BLUESKY_DIRECTORY:
233 i->dirents = g_sequence_new(bluesky_dirent_destroy);
234 i->dirhash = g_hash_table_new(g_str_hash, g_str_equal);
235 i->dirhash_folded = g_hash_table_new(g_str_hash, g_str_equal);
242 BlueSkyInode *bluesky_new_inode(uint64_t inum, BlueSkyFS *fs,
243 BlueSkyFileType type)
245 BlueSkyInode *i = g_new0(BlueSkyInode, 1);
247 i->lock = g_mutex_new();
252 bluesky_init_inode(i, type);
257 /* Retrieve an inode from the filesystem. Eventually this will be a cache and
258 * so we might need to go fetch the inode from elsewhere; for now all
259 * filesystem state is stored here. inode is returned with a reference held
261 BlueSkyInode *bluesky_get_inode(BlueSkyFS *fs, uint64_t inum)
263 BlueSkyInode *inode = NULL;
269 g_mutex_lock(fs->lock);
270 inode = (BlueSkyInode *)g_hash_table_lookup(fs->inodes, &inum);
273 bluesky_inode_fetch(fs, inum);
274 inode = (BlueSkyInode *)g_hash_table_lookup(fs->inodes, &inum);
278 bluesky_inode_ref(inode);
280 /* FIXME: We assume we can atomically update the in-memory access time
282 inode->access_time = bluesky_get_current_time();
285 g_mutex_unlock(fs->lock);
290 /* Insert an inode into the filesystem inode cache. fs should be locked. */
291 void bluesky_insert_inode(BlueSkyFS *fs, BlueSkyInode *inode)
293 g_hash_table_insert(fs->inodes, &inode->inum, inode);
296 /* Start writeback of an inode and all associated data. */
297 void bluesky_inode_start_sync(BlueSkyInode *inode)
299 GList *log_items = NULL;
301 if (inode->type == BLUESKY_REGULAR)
302 bluesky_file_flush(inode, &log_items);
304 BlueSkyCloudLog *cloudlog = bluesky_serialize_inode(inode);
306 bluesky_cloudlog_unref(inode->committed_item);
307 inode->committed_item = cloudlog;
309 bluesky_cloudlog_sync(cloudlog);
310 bluesky_cloudlog_ref(cloudlog);
311 log_items = g_list_prepend(log_items, cloudlog);
313 /* Wait for all log items to be committed to disk. */
314 bluesky_log_finish_all(log_items);
316 /* Mark the inode as clean */
317 inode->change_commit = inode->change_count;
318 inode->change_time = 0;
319 g_mutex_lock(inode->fs->lock);
320 bluesky_list_unlink(&inode->fs->unlogged_list, inode->unlogged_list);
321 inode->unlogged_list = NULL;
323 /* Since a new version of the inode has been written to the log, also
324 * schedule a future flush of the new data to cloud storage. */
325 bluesky_list_unlink(&inode->fs->dirty_list, inode->dirty_list);
326 inode->dirty_list = bluesky_list_prepend(&inode->fs->dirty_list, inode);
327 inode->change_cloud = inode->change_count;
329 g_mutex_unlock(inode->fs->lock);
332 /* Write back an inode and all associated data and wait for completion. Inode
333 * should already be locked. */
334 void bluesky_inode_do_sync(BlueSkyInode *inode)
336 if (bluesky_verbose) {
337 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
338 "Synchronous writeback for inode %"PRIu64"...", inode->inum);
340 bluesky_inode_start_sync(inode);
341 if (bluesky_verbose) {
342 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
343 "Writeback for inode %"PRIu64" complete", inode->inum);
347 static void complete_inode_fetch(BlueSkyInode *inode)
349 BlueSkyCloudLog *item = inode->committed_item;
350 inode->committed_item = NULL;
351 g_print("Completing fetch of inode %"PRIu64"...", inode->inum);
353 g_mutex_lock(item->lock);
354 bluesky_cloudlog_fetch(item);
355 if (!bluesky_deserialize_inode(inode, item))
356 g_print("Error deserializing inode %"PRIu64"\n", inode->inum);
357 g_mutex_unlock(item->lock);
359 inode->access_time = bluesky_get_current_time();
360 g_mutex_lock(inode->fs->lock);
361 bluesky_list_unlink(&inode->fs->accessed_list, inode->accessed_list);
362 inode->accessed_list = bluesky_list_prepend(&inode->fs->accessed_list, inode);
363 g_mutex_unlock(inode->fs->lock);
365 g_mutex_unlock(inode->lock);
366 bluesky_cloudlog_unref(item);
367 bluesky_inode_unref(inode);
370 /* Fetch an inode from stable storage. The fetch can be performed
371 * asynchronously: the in-memory inode is allocated, but not filled with data
372 * immediately. It is kept locked until it has been filled in, so any users
373 * should try to acquire the lock on the inode before accessing any data. The
374 * fs lock must be held. */
375 void bluesky_inode_fetch(BlueSkyFS *fs, uint64_t inum)
377 InodeMapEntry *entry = bluesky_inode_map_lookup(fs->inode_map, inum, 0);
381 BlueSkyInode *inode = bluesky_new_inode(inum, fs, BLUESKY_PENDING);
382 inode->change_count = 0;
383 bluesky_inode_ref(inode); // Extra ref held by fetching process
384 g_mutex_lock(inode->lock);
385 bluesky_insert_inode(fs, inode);
386 inode->committed_item = entry->item;
387 bluesky_cloudlog_ref(entry->item);
389 /* TODO: Thread pool or other better async method. */
390 g_thread_create((GThreadFunc)complete_inode_fetch,
391 (gpointer)inode, FALSE, NULL);