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 g_print("Initializing filesystem\n");
103 BlueSkyFS *fs = bluesky_new_fs(name);
105 fs->log = bluesky_log_new("journal");
108 bluesky_checkpoint_load(fs);
111 BlueSkyInode *root = bluesky_new_inode(BLUESKY_ROOT_INUM, fs,
115 bluesky_insert_inode(fs, root);
116 bluesky_inode_update_ctime(root, TRUE);
120 bluesky_inode_do_sync(root);
125 /* Inode reference counting. */
126 void bluesky_inode_ref(BlueSkyInode *inode)
128 g_atomic_int_inc(&inode->refcount);
131 /* Free most of the resources used by an inode structure, but do not free the
132 * inode itself. Can be used if the inode data will be reloaded from
133 * serialized form to clear out old information first. */
134 void bluesky_inode_free_resources(BlueSkyInode *inode)
136 switch (inode->type) {
137 case BLUESKY_REGULAR:
138 if (inode->blocks != NULL) {
139 for (int i = 0; i < inode->blocks->len; i++) {
140 BlueSkyBlock *b = &g_array_index(inode->blocks,
142 if (b->type == BLUESKY_BLOCK_DIRTY) {
143 g_error("Deleting an inode with dirty file data!");
145 bluesky_cloudlog_unref(b->ref);
146 bluesky_string_unref(b->dirty);
148 g_array_unref(inode->blocks);
149 inode->blocks = NULL;
153 case BLUESKY_DIRECTORY:
154 if (inode->dirhash != NULL)
155 g_hash_table_destroy(inode->dirhash);
156 inode->dirhash = NULL;
157 if (inode->dirhash_folded != NULL)
158 g_hash_table_destroy(inode->dirhash_folded);
159 inode->dirhash_folded = NULL;
160 if (inode->dirents != NULL)
161 g_sequence_free(inode->dirents);
162 inode->dirents = NULL;
165 case BLUESKY_SYMLINK:
166 g_free(inode->symlink_contents);
167 inode->symlink_contents = NULL;
175 void bluesky_inode_unref(BlueSkyInode *inode)
177 if (g_atomic_int_dec_and_test(&inode->refcount)) {
178 if (bluesky_verbose) {
179 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
180 "Reference count for inode %"PRIu64" dropped to zero.",
184 /* Sanity check: Is the inode clean? */
185 if (inode->change_commit < inode->change_count
186 || inode->accessed_list != NULL
187 || inode->unlogged_list != NULL
188 || inode->dirty_list != NULL) {
189 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);
192 /* These shouldn't be needed, but in case the above warning fires and
193 * we delete the inode anyway, we ought to be sure the inode is not on
195 g_mutex_lock(inode->fs->lock);
196 bluesky_list_unlink(&inode->fs->accessed_list, inode->accessed_list);
197 bluesky_list_unlink(&inode->fs->dirty_list, inode->dirty_list);
198 bluesky_list_unlink(&inode->fs->unlogged_list, inode->unlogged_list);
199 g_mutex_unlock(inode->fs->lock);
201 bluesky_inode_free_resources(inode);
203 g_mutex_free(inode->lock);
208 /* Allocate a fresh inode number which has not been used before within a
209 * filesystem. fs must already be locked. */
210 uint64_t bluesky_fs_alloc_inode(BlueSkyFS *fs)
214 inum = fs->next_inum;
220 /* Perform type-specification initialization of an inode. Normally performed
221 * in bluesky_new_inode, but can be separated if an inode is created first,
222 * then deserialized. */
223 void bluesky_init_inode(BlueSkyInode *i, BlueSkyFileType type)
228 case BLUESKY_REGULAR:
229 i->blocks = g_array_new(FALSE, TRUE, sizeof(BlueSkyBlock));
231 case BLUESKY_DIRECTORY:
232 i->dirents = g_sequence_new(bluesky_dirent_destroy);
233 i->dirhash = g_hash_table_new(g_str_hash, g_str_equal);
234 i->dirhash_folded = g_hash_table_new(g_str_hash, g_str_equal);
241 BlueSkyInode *bluesky_new_inode(uint64_t inum, BlueSkyFS *fs,
242 BlueSkyFileType type)
244 BlueSkyInode *i = g_new0(BlueSkyInode, 1);
246 i->lock = g_mutex_new();
251 bluesky_init_inode(i, type);
256 /* Retrieve an inode from the filesystem. Eventually this will be a cache and
257 * so we might need to go fetch the inode from elsewhere; for now all
258 * filesystem state is stored here. inode is returned with a reference held
260 BlueSkyInode *bluesky_get_inode(BlueSkyFS *fs, uint64_t inum)
262 BlueSkyInode *inode = NULL;
268 g_mutex_lock(fs->lock);
269 inode = (BlueSkyInode *)g_hash_table_lookup(fs->inodes, &inum);
272 bluesky_inode_fetch(fs, inum);
273 inode = (BlueSkyInode *)g_hash_table_lookup(fs->inodes, &inum);
277 bluesky_inode_ref(inode);
279 /* FIXME: We assume we can atomically update the in-memory access time
281 inode->access_time = bluesky_get_current_time();
284 g_mutex_unlock(fs->lock);
289 /* Insert an inode into the filesystem inode cache. fs should be locked. */
290 void bluesky_insert_inode(BlueSkyFS *fs, BlueSkyInode *inode)
292 g_hash_table_insert(fs->inodes, &inode->inum, inode);
295 /* Start writeback of an inode and all associated data. */
296 void bluesky_inode_start_sync(BlueSkyInode *inode)
298 GList *log_items = NULL;
300 if (inode->type == BLUESKY_REGULAR)
301 bluesky_file_flush(inode, &log_items);
303 BlueSkyCloudLog *cloudlog = bluesky_serialize_inode(inode);
305 bluesky_cloudlog_unref(inode->committed_item);
306 inode->committed_item = cloudlog;
308 bluesky_cloudlog_sync(cloudlog);
309 bluesky_cloudlog_ref(cloudlog);
310 log_items = g_list_prepend(log_items, cloudlog);
312 /* Wait for all log items to be committed to disk. */
313 bluesky_log_finish_all(log_items);
315 /* Mark the inode as clean */
316 inode->change_commit = inode->change_count;
317 inode->change_time = 0;
318 g_mutex_lock(inode->fs->lock);
319 bluesky_list_unlink(&inode->fs->unlogged_list, inode->unlogged_list);
320 inode->unlogged_list = NULL;
322 /* Since a new version of the inode has been written to the log, also
323 * schedule a future flush of the new data to cloud storage. */
324 bluesky_list_unlink(&inode->fs->dirty_list, inode->dirty_list);
325 inode->dirty_list = bluesky_list_prepend(&inode->fs->dirty_list, inode);
326 inode->change_cloud = inode->change_count;
328 g_mutex_unlock(inode->fs->lock);
331 /* Write back an inode and all associated data and wait for completion. Inode
332 * should already be locked. */
333 void bluesky_inode_do_sync(BlueSkyInode *inode)
335 if (bluesky_verbose) {
336 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
337 "Synchronous writeback for inode %"PRIu64"...", inode->inum);
339 bluesky_inode_start_sync(inode);
340 if (bluesky_verbose) {
341 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
342 "Writeback for inode %"PRIu64" complete", inode->inum);
346 static void complete_inode_fetch(BlueSkyStoreAsync *async, BlueSkyInode *inode)
348 if (bluesky_verbose) {
349 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
350 "Completing fetch of inode %"PRIu64"...", inode->inum);
353 if (async->result != 0 || FALSE)
355 if (bluesky_verbose) {
356 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
357 " failed to load inode, cleaning up");
359 g_mutex_lock(inode->fs->lock);
360 g_hash_table_remove(inode->fs->inodes, &inode->inum);
361 bluesky_list_unlink(&inode->fs->accessed_list, inode->accessed_list);
362 inode->accessed_list = NULL;
363 g_mutex_unlock(inode->fs->lock);
364 bluesky_inode_unref(inode);
367 inode->access_time = bluesky_get_current_time();
368 g_mutex_lock(inode->fs->lock);
369 bluesky_list_unlink(&inode->fs->accessed_list, inode->accessed_list);
370 inode->accessed_list = bluesky_list_prepend(&inode->fs->accessed_list, inode);
371 g_mutex_unlock(inode->fs->lock);
373 g_mutex_unlock(inode->lock);
374 bluesky_inode_unref(inode);
377 /* Fetch an inode from stable storage. The fetch can be performed
378 * asynchronously: the in-memory inode is allocated, but not filled with data
379 * immediately. It is kept locked until it has been filled in, so any users
380 * should try to acquire the lock on the inode before accessing any data. The
381 * fs lock must be held. */
382 void bluesky_inode_fetch(BlueSkyFS *fs, uint64_t inum)
385 sprintf(key, "inode-%016"PRIx64, inum);
387 BlueSkyInode *inode = bluesky_new_inode(inum, fs, BLUESKY_PENDING);
388 inode->change_count = 0;
389 bluesky_inode_ref(inode); // Extra ref held by fetching process
390 g_mutex_lock(inode->lock);
391 bluesky_insert_inode(fs, inode);
393 BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
394 async->op = STORE_OP_GET;
395 async->key = g_strdup(key);
397 bluesky_store_async_add_notifier(async, (GFunc)complete_inode_fetch, inode);
398 bluesky_store_async_submit(async);
400 if (bluesky_options.sync_inode_fetches) {
401 bluesky_store_async_wait(async);
404 bluesky_store_async_unref(async);