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 static void inode_fetch_task(gpointer a, gpointer b);
19 /* Core filesystem. Different proxies, such as the NFSv3 one, interface to
20 * this, but the core actually tracks the data which is stored. So far we just
21 * implement an in-memory filesystem, but eventually this will be state which
22 * is persisted to the cloud. */
24 /* Return the current time, in microseconds since the epoch. */
25 int64_t bluesky_get_current_time()
28 g_get_current_time(&t);
29 return (int64_t)t.tv_sec * 1000000 + t.tv_usec;
32 /* Update an inode to indicate that a modification was made. This increases
33 * the change counter, updates the ctime to the current time, and optionally
34 * updates the mtime. This also makes the inode contents subject to writeback
35 * to storage in the future. inode must already be locked. */
36 void bluesky_inode_update_ctime(BlueSkyInode *inode, gboolean update_mtime)
38 int64_t now = bluesky_get_current_time();
39 inode->change_count++;
44 if (inode->change_time == 0)
45 inode->change_time = now;
48 if (bluesky_options.writethrough_cache)
49 bluesky_file_flush(inode, NULL);
52 g_mutex_lock(inode->fs->lock);
53 bluesky_list_unlink(&inode->fs->unlogged_list, inode->unlogged_list);
54 inode->unlogged_list = bluesky_list_prepend(&inode->fs->unlogged_list, inode);
55 bluesky_list_unlink(&inode->fs->accessed_list, inode->accessed_list);
56 inode->accessed_list = bluesky_list_prepend(&inode->fs->accessed_list, inode);
57 g_mutex_unlock(inode->fs->lock);
60 /* Unfortunately a glib hash table is only guaranteed to be able to store
61 * 32-bit keys if we use the key directly. If we want 64-bit inode numbers,
62 * we'll have to allocate memory to store the 64-bit inumber, and use a pointer
63 * to it. Rather than allocate the memory for the key, we'll just include a
64 * pointer to the 64-bit inum stored in the inode itself, so that we don't need
65 * to do any more memory management. */
66 static guint bluesky_fs_key_hash_func(gconstpointer key)
68 uint64_t inum = *(const uint64_t *)key;
72 static gboolean bluesky_fs_key_equal_func(gconstpointer a, gconstpointer b)
74 uint64_t i1 = *(const uint64_t *)a;
75 uint64_t i2 = *(const uint64_t *)b;
79 /* Filesystem-level operations. A filesystem is like a directory tree that we
80 * are willing to export. */
81 BlueSkyFS *bluesky_new_fs(gchar *name)
83 BlueSkyFS *fs = g_new0(BlueSkyFS, 1);
84 fs->lock = g_mutex_new();
85 fs->name = g_strdup(name);
86 fs->inodes = g_hash_table_new(bluesky_fs_key_hash_func,
87 bluesky_fs_key_equal_func);
88 fs->next_inum = BLUESKY_ROOT_INUM + 1;
89 fs->store = bluesky_store_new("file");
90 fs->flushd_lock = g_mutex_new();
91 fs->flushd_cond = g_cond_new();
92 fs->locations = g_hash_table_new(bluesky_cloudlog_hash,
93 bluesky_cloudlog_equal);
94 fs->inode_map = g_sequence_new(NULL);
96 fs->log_state = g_new0(BlueSkyCloudLogState, 1);
97 fs->log_state->data = g_string_new("");
99 bluesky_cloudlog_threads_init(fs);
100 fs->inode_fetch_thread_pool = g_thread_pool_new(inode_fetch_task, NULL,
107 BlueSkyFS *bluesky_init_fs(gchar *name, BlueSkyStore *store,
108 const gchar *master_key)
110 BlueSkyFS *fs = bluesky_new_fs(name);
111 fs->master_key = g_strdup(master_key);
112 fs->keys = g_new(BlueSkyCryptKeys, 1);
113 bluesky_crypt_derive_keys(fs->keys, master_key);
115 fs->log = bluesky_log_new("journal");
118 if (bluesky_checkpoint_load(fs)) {
119 g_print("Filesystem checkpoint loaded, starting journal replay...\n");
121 g_print("Journal replay complete, filesystem ready.\n");
123 /* Initialize a fresh filesystem */
124 g_print("Initializing new filesystem...\n");
125 BlueSkyInode *root = bluesky_new_inode(BLUESKY_ROOT_INUM, fs,
129 bluesky_insert_inode(fs, root);
130 bluesky_inode_update_ctime(root, TRUE);
131 bluesky_inode_do_sync(root);
137 /* Inode reference counting. */
138 void bluesky_inode_ref(BlueSkyInode *inode)
140 g_atomic_int_inc(&inode->refcount);
143 /* Free most of the resources used by an inode structure, but do not free the
144 * inode itself. Can be used if the inode data will be reloaded from
145 * serialized form to clear out old information first. */
146 void bluesky_inode_free_resources(BlueSkyInode *inode)
148 switch (inode->type) {
149 case BLUESKY_REGULAR:
150 if (inode->blocks != NULL) {
151 for (int i = 0; i < inode->blocks->len; i++) {
152 BlueSkyBlock *b = &g_array_index(inode->blocks,
154 if (b->type == BLUESKY_BLOCK_DIRTY) {
155 g_error("Deleting an inode with dirty file data!");
157 bluesky_cloudlog_unref(b->ref);
158 bluesky_string_unref(b->dirty);
160 g_array_unref(inode->blocks);
161 inode->blocks = NULL;
165 case BLUESKY_DIRECTORY:
166 if (inode->dirhash != NULL)
167 g_hash_table_destroy(inode->dirhash);
168 inode->dirhash = NULL;
169 if (inode->dirhash_folded != NULL)
170 g_hash_table_destroy(inode->dirhash_folded);
171 inode->dirhash_folded = NULL;
172 if (inode->dirents != NULL)
173 g_sequence_free(inode->dirents);
174 inode->dirents = NULL;
177 case BLUESKY_SYMLINK:
178 g_free(inode->symlink_contents);
179 inode->symlink_contents = NULL;
187 void bluesky_inode_unref(BlueSkyInode *inode)
189 if (g_atomic_int_dec_and_test(&inode->refcount)) {
190 if (bluesky_verbose) {
191 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
192 "Reference count for inode %"PRIu64" dropped to zero.",
196 /* Sanity check: Is the inode clean? */
197 if (inode->change_commit < inode->change_count
198 || inode->accessed_list != NULL
199 || inode->unlogged_list != NULL
200 || inode->dirty_list != NULL) {
201 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);
204 /* These shouldn't be needed, but in case the above warning fires and
205 * we delete the inode anyway, we ought to be sure the inode is not on
207 g_mutex_lock(inode->fs->lock);
208 bluesky_list_unlink(&inode->fs->accessed_list, inode->accessed_list);
209 bluesky_list_unlink(&inode->fs->dirty_list, inode->dirty_list);
210 bluesky_list_unlink(&inode->fs->unlogged_list, inode->unlogged_list);
211 g_mutex_unlock(inode->fs->lock);
213 bluesky_inode_free_resources(inode);
215 g_mutex_free(inode->lock);
220 /* Allocate a fresh inode number which has not been used before within a
221 * filesystem. fs must already be locked. */
222 uint64_t bluesky_fs_alloc_inode(BlueSkyFS *fs)
226 inum = fs->next_inum;
232 /* Perform type-specification initialization of an inode. Normally performed
233 * in bluesky_new_inode, but can be separated if an inode is created first,
234 * then deserialized. */
235 void bluesky_init_inode(BlueSkyInode *i, BlueSkyFileType type)
240 case BLUESKY_REGULAR:
241 i->blocks = g_array_new(FALSE, TRUE, sizeof(BlueSkyBlock));
243 case BLUESKY_DIRECTORY:
244 i->dirents = g_sequence_new(bluesky_dirent_destroy);
245 i->dirhash = g_hash_table_new(g_str_hash, g_str_equal);
246 i->dirhash_folded = g_hash_table_new(g_str_hash, g_str_equal);
253 BlueSkyInode *bluesky_new_inode(uint64_t inum, BlueSkyFS *fs,
254 BlueSkyFileType type)
256 BlueSkyInode *i = g_new0(BlueSkyInode, 1);
258 i->lock = g_mutex_new();
263 bluesky_init_inode(i, type);
268 /* Retrieve an inode from the filesystem. Eventually this will be a cache and
269 * so we might need to go fetch the inode from elsewhere; for now all
270 * filesystem state is stored here. inode is returned with a reference held
272 BlueSkyInode *bluesky_get_inode(BlueSkyFS *fs, uint64_t inum)
274 BlueSkyInode *inode = NULL;
280 g_mutex_lock(fs->lock);
281 inode = (BlueSkyInode *)g_hash_table_lookup(fs->inodes, &inum);
284 bluesky_inode_fetch(fs, inum);
285 inode = (BlueSkyInode *)g_hash_table_lookup(fs->inodes, &inum);
289 bluesky_inode_ref(inode);
291 /* FIXME: We assume we can atomically update the in-memory access time
293 inode->access_time = bluesky_get_current_time();
296 g_mutex_unlock(fs->lock);
301 /* Insert an inode into the filesystem inode cache. fs should be locked. */
302 void bluesky_insert_inode(BlueSkyFS *fs, BlueSkyInode *inode)
304 g_hash_table_insert(fs->inodes, &inode->inum, inode);
307 /* Start writeback of an inode and all associated data. */
308 void bluesky_inode_start_sync(BlueSkyInode *inode)
310 GList *log_items = NULL;
312 if (inode->type == BLUESKY_REGULAR)
313 bluesky_file_flush(inode, &log_items);
315 BlueSkyCloudLog *cloudlog = bluesky_serialize_inode(inode);
317 bluesky_cloudlog_unref(inode->committed_item);
318 inode->committed_item = cloudlog;
320 bluesky_cloudlog_sync(cloudlog);
321 bluesky_cloudlog_ref(cloudlog);
322 log_items = g_list_prepend(log_items, cloudlog);
324 /* Wait for all log items to be committed to disk. */
325 bluesky_log_finish_all(log_items);
327 /* Mark the inode as clean */
328 inode->change_commit = inode->change_count;
329 inode->change_time = 0;
330 g_mutex_lock(inode->fs->lock);
331 bluesky_list_unlink(&inode->fs->unlogged_list, inode->unlogged_list);
332 inode->unlogged_list = NULL;
334 /* Since a new version of the inode has been written to the log, also
335 * schedule a future flush of the new data to cloud storage. */
336 bluesky_list_unlink(&inode->fs->dirty_list, inode->dirty_list);
337 inode->dirty_list = bluesky_list_prepend(&inode->fs->dirty_list, inode);
338 inode->change_cloud = inode->change_count;
340 g_mutex_unlock(inode->fs->lock);
343 /* Write back an inode and all associated data and wait for completion. Inode
344 * should already be locked. */
345 void bluesky_inode_do_sync(BlueSkyInode *inode)
347 if (bluesky_verbose) {
348 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
349 "Synchronous writeback for inode %"PRIu64"...", inode->inum);
351 bluesky_inode_start_sync(inode);
352 if (bluesky_verbose) {
353 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
354 "Writeback for inode %"PRIu64" complete", inode->inum);
358 static void inode_fetch_task(gpointer a, gpointer b)
360 BlueSkyInode *inode = (BlueSkyInode *)a;
362 BlueSkyCloudLog *item = inode->committed_item;
363 inode->committed_item = NULL;
364 g_print("Completing fetch of inode %"PRIu64"...\n", inode->inum);
366 g_mutex_lock(item->lock);
367 bluesky_cloudlog_fetch(item);
368 if (!bluesky_deserialize_inode(inode, item))
369 g_print("Error deserializing inode %"PRIu64"\n", inode->inum);
370 g_mutex_unlock(item->lock);
372 inode->access_time = bluesky_get_current_time();
373 g_mutex_lock(inode->fs->lock);
374 bluesky_list_unlink(&inode->fs->accessed_list, inode->accessed_list);
375 inode->accessed_list = bluesky_list_prepend(&inode->fs->accessed_list, inode);
376 g_mutex_unlock(inode->fs->lock);
378 g_mutex_unlock(inode->lock);
379 bluesky_cloudlog_unref(item);
380 bluesky_inode_unref(inode);
383 /* Fetch an inode from stable storage. The fetch can be performed
384 * asynchronously: the in-memory inode is allocated, but not filled with data
385 * immediately. It is kept locked until it has been filled in, so any users
386 * should try to acquire the lock on the inode before accessing any data. The
387 * fs lock must be held. */
388 void bluesky_inode_fetch(BlueSkyFS *fs, uint64_t inum)
390 InodeMapEntry *entry = bluesky_inode_map_lookup(fs->inode_map, inum, 0);
394 /* Non-portable behavior: We take the inode lock here, and release it in
395 * the fetching thread. This works with the default Linux pthreads
396 * implementation but is not guaranteed. */
398 BlueSkyInode *inode = bluesky_new_inode(inum, fs, BLUESKY_PENDING);
399 inode->change_count = 0;
400 bluesky_inode_ref(inode); // Extra ref held by fetching process
401 g_mutex_lock(inode->lock);
403 inode->committed_item = entry->item;
404 bluesky_cloudlog_ref(entry->item);
405 bluesky_insert_inode(fs, inode);
407 g_thread_pool_push(fs->inode_fetch_thread_pool, inode, NULL);