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("");
98 fs->log_state->latest_cleaner_seq_seen = -1;
100 bluesky_cloudlog_threads_init(fs);
101 fs->inode_fetch_thread_pool = g_thread_pool_new(inode_fetch_task, NULL,
108 BlueSkyFS *bluesky_init_fs(gchar *name, BlueSkyStore *store,
109 const gchar *master_key)
111 BlueSkyFS *fs = bluesky_new_fs(name);
112 fs->master_key = g_strdup(master_key);
113 fs->keys = g_new(BlueSkyCryptKeys, 1);
114 bluesky_crypt_derive_keys(fs->keys, master_key);
116 fs->log = bluesky_log_new("journal");
119 if (bluesky_checkpoint_load(fs)) {
120 g_print("Filesystem checkpoint loaded, starting journal replay...\n");
122 g_print("Journal replay complete, filesystem ready.\n");
124 /* Initialize a fresh filesystem */
125 g_print("Initializing new filesystem...\n");
126 BlueSkyInode *root = bluesky_new_inode(BLUESKY_ROOT_INUM, fs,
130 bluesky_insert_inode(fs, root);
131 bluesky_inode_update_ctime(root, TRUE);
132 bluesky_inode_do_sync(root);
138 /* Inode reference counting. */
139 void bluesky_inode_ref(BlueSkyInode *inode)
141 g_atomic_int_inc(&inode->refcount);
144 /* Free most of the resources used by an inode structure, but do not free the
145 * inode itself. Can be used if the inode data will be reloaded from
146 * serialized form to clear out old information first. */
147 void bluesky_inode_free_resources(BlueSkyInode *inode)
149 switch (inode->type) {
150 case BLUESKY_REGULAR:
151 if (inode->blocks != NULL) {
152 for (int i = 0; i < inode->blocks->len; i++) {
153 BlueSkyBlock *b = &g_array_index(inode->blocks,
155 if (b->type == BLUESKY_BLOCK_DIRTY) {
156 g_error("Deleting an inode with dirty file data!");
158 bluesky_cloudlog_unref(b->ref);
159 bluesky_string_unref(b->dirty);
161 g_array_unref(inode->blocks);
162 inode->blocks = NULL;
166 case BLUESKY_DIRECTORY:
167 if (inode->dirhash != NULL)
168 g_hash_table_destroy(inode->dirhash);
169 inode->dirhash = NULL;
170 if (inode->dirhash_folded != NULL)
171 g_hash_table_destroy(inode->dirhash_folded);
172 inode->dirhash_folded = NULL;
173 if (inode->dirents != NULL)
174 g_sequence_free(inode->dirents);
175 inode->dirents = NULL;
178 case BLUESKY_SYMLINK:
179 g_free(inode->symlink_contents);
180 inode->symlink_contents = NULL;
188 void bluesky_inode_unref(BlueSkyInode *inode)
190 if (g_atomic_int_dec_and_test(&inode->refcount)) {
191 if (bluesky_verbose) {
192 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
193 "Reference count for inode %"PRIu64" dropped to zero.",
197 /* Sanity check: Is the inode clean? */
198 if (inode->change_commit < inode->change_count
199 || inode->accessed_list != NULL
200 || inode->unlogged_list != NULL
201 || inode->dirty_list != NULL) {
202 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);
205 /* These shouldn't be needed, but in case the above warning fires and
206 * we delete the inode anyway, we ought to be sure the inode is not on
208 g_mutex_lock(inode->fs->lock);
209 bluesky_list_unlink(&inode->fs->accessed_list, inode->accessed_list);
210 bluesky_list_unlink(&inode->fs->dirty_list, inode->dirty_list);
211 bluesky_list_unlink(&inode->fs->unlogged_list, inode->unlogged_list);
212 g_mutex_unlock(inode->fs->lock);
214 bluesky_inode_free_resources(inode);
216 g_mutex_free(inode->lock);
221 /* Allocate a fresh inode number which has not been used before within a
222 * filesystem. fs must already be locked. */
223 uint64_t bluesky_fs_alloc_inode(BlueSkyFS *fs)
227 inum = fs->next_inum;
233 /* Perform type-specification initialization of an inode. Normally performed
234 * in bluesky_new_inode, but can be separated if an inode is created first,
235 * then deserialized. */
236 void bluesky_init_inode(BlueSkyInode *i, BlueSkyFileType type)
241 case BLUESKY_REGULAR:
242 i->blocks = g_array_new(FALSE, TRUE, sizeof(BlueSkyBlock));
244 case BLUESKY_DIRECTORY:
245 i->dirents = g_sequence_new(bluesky_dirent_destroy);
246 i->dirhash = g_hash_table_new(g_str_hash, g_str_equal);
247 i->dirhash_folded = g_hash_table_new(g_str_hash, g_str_equal);
254 BlueSkyInode *bluesky_new_inode(uint64_t inum, BlueSkyFS *fs,
255 BlueSkyFileType type)
257 BlueSkyInode *i = g_new0(BlueSkyInode, 1);
259 i->lock = g_mutex_new();
264 bluesky_init_inode(i, type);
269 /* Retrieve an inode from the filesystem. Eventually this will be a cache and
270 * so we might need to go fetch the inode from elsewhere; for now all
271 * filesystem state is stored here. inode is returned with a reference held
273 BlueSkyInode *bluesky_get_inode(BlueSkyFS *fs, uint64_t inum)
275 BlueSkyInode *inode = NULL;
281 g_mutex_lock(fs->lock);
282 inode = (BlueSkyInode *)g_hash_table_lookup(fs->inodes, &inum);
285 bluesky_inode_fetch(fs, inum);
286 inode = (BlueSkyInode *)g_hash_table_lookup(fs->inodes, &inum);
290 bluesky_inode_ref(inode);
292 /* FIXME: We assume we can atomically update the in-memory access time
294 inode->access_time = bluesky_get_current_time();
297 g_mutex_unlock(fs->lock);
302 /* Insert an inode into the filesystem inode cache. fs should be locked. */
303 void bluesky_insert_inode(BlueSkyFS *fs, BlueSkyInode *inode)
305 g_hash_table_insert(fs->inodes, &inode->inum, inode);
308 /* Start writeback of an inode and all associated data. */
309 void bluesky_inode_start_sync(BlueSkyInode *inode)
311 GList *log_items = NULL;
313 if (inode->type == BLUESKY_REGULAR)
314 bluesky_file_flush(inode, &log_items);
316 BlueSkyCloudLog *cloudlog = bluesky_serialize_inode(inode);
318 bluesky_cloudlog_unref(inode->committed_item);
319 inode->committed_item = cloudlog;
321 bluesky_cloudlog_sync(cloudlog);
322 bluesky_cloudlog_ref(cloudlog);
323 log_items = g_list_prepend(log_items, cloudlog);
325 /* Wait for all log items to be committed to disk. */
326 bluesky_log_finish_all(log_items);
328 /* Mark the inode as clean */
329 inode->change_commit = inode->change_count;
330 inode->change_time = 0;
331 g_mutex_lock(inode->fs->lock);
332 bluesky_list_unlink(&inode->fs->unlogged_list, inode->unlogged_list);
333 inode->unlogged_list = NULL;
335 /* Since a new version of the inode has been written to the log, also
336 * schedule a future flush of the new data to cloud storage. */
337 bluesky_list_unlink(&inode->fs->dirty_list, inode->dirty_list);
338 inode->dirty_list = bluesky_list_prepend(&inode->fs->dirty_list, inode);
339 inode->change_cloud = inode->change_count;
341 g_mutex_unlock(inode->fs->lock);
344 /* Write back an inode and all associated data and wait for completion. Inode
345 * should already be locked. */
346 void bluesky_inode_do_sync(BlueSkyInode *inode)
348 if (bluesky_verbose) {
349 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
350 "Synchronous writeback for inode %"PRIu64"...", inode->inum);
352 bluesky_inode_start_sync(inode);
353 if (bluesky_verbose) {
354 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
355 "Writeback for inode %"PRIu64" complete", inode->inum);
359 static void inode_fetch_task(gpointer a, gpointer b)
361 BlueSkyInode *inode = (BlueSkyInode *)a;
363 bluesky_profile_set((BlueSkyProfile *)inode->private_data);
365 BlueSkyCloudLog *item = inode->committed_item;
366 inode->committed_item = NULL;
367 g_print("Completing fetch of inode %"PRIu64"...\n", inode->inum);
369 g_mutex_lock(item->lock);
370 bluesky_cloudlog_fetch(item);
371 if (!bluesky_deserialize_inode(inode, item))
372 g_print("Error deserializing inode %"PRIu64"\n", inode->inum);
373 g_mutex_unlock(item->lock);
375 inode->access_time = bluesky_get_current_time();
376 g_mutex_lock(inode->fs->lock);
377 bluesky_list_unlink(&inode->fs->accessed_list, inode->accessed_list);
378 inode->accessed_list = bluesky_list_prepend(&inode->fs->accessed_list, inode);
379 g_mutex_unlock(inode->fs->lock);
381 g_mutex_unlock(inode->lock);
382 bluesky_cloudlog_unref(item);
383 bluesky_inode_unref(inode);
386 /* Fetch an inode from stable storage. The fetch can be performed
387 * asynchronously: the in-memory inode is allocated, but not filled with data
388 * immediately. It is kept locked until it has been filled in, so any users
389 * should try to acquire the lock on the inode before accessing any data. The
390 * fs lock must be held. */
391 void bluesky_inode_fetch(BlueSkyFS *fs, uint64_t inum)
393 InodeMapEntry *entry = bluesky_inode_map_lookup(fs->inode_map, inum, 0);
397 /* Non-portable behavior: We take the inode lock here, and release it in
398 * the fetching thread. This works with the default Linux pthreads
399 * implementation but is not guaranteed. */
401 BlueSkyInode *inode = bluesky_new_inode(inum, fs, BLUESKY_PENDING);
402 inode->change_count = 0;
403 bluesky_inode_ref(inode); // Extra ref held by fetching process
404 g_mutex_lock(inode->lock);
406 inode->committed_item = entry->item;
407 bluesky_cloudlog_ref(entry->item);
408 bluesky_insert_inode(fs, inode);
410 inode->private_data = bluesky_profile_get();
411 g_thread_pool_push(fs->inode_fetch_thread_pool, inode, NULL);