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;
45 if (bluesky_options.writethrough_cache)
46 bluesky_file_flush(inode, NULL);
49 /* Unfortunately a glib hash table is only guaranteed to be able to store
50 * 32-bit keys if we use the key directly. If we want 64-bit inode numbers,
51 * we'll have to allocate memory to store the 64-bit inumber, and use a pointer
52 * to it. Rather than allocate the memory for the key, we'll just include a
53 * pointer to the 64-bit inum stored in the inode itself, so that we don't need
54 * to do any more memory management. */
55 static guint bluesky_fs_key_hash_func(gconstpointer key)
57 uint64_t inum = *(const uint64_t *)key;
61 static gboolean bluesky_fs_key_equal_func(gconstpointer a, gconstpointer b)
63 uint64_t i1 = *(const uint64_t *)a;
64 uint64_t i2 = *(const uint64_t *)b;
68 /* Filesystem-level operations. A filesystem is like a directory tree that we
69 * are willing to export. */
70 BlueSkyFS *bluesky_new_fs(gchar *name)
72 BlueSkyFS *fs = g_new0(BlueSkyFS, 1);
73 fs->lock = g_mutex_new();
74 fs->name = g_strdup(name);
75 fs->inodes = g_hash_table_new(bluesky_fs_key_hash_func,
76 bluesky_fs_key_equal_func);
77 fs->next_inum = BLUESKY_ROOT_INUM + 1;
78 fs->store = bluesky_store_new("file");
83 BlueSkyFS *bluesky_init_fs(gchar *name, BlueSkyStore *store)
85 BlueSkyRCStr *data = bluesky_store_get(store, "superblock");
87 BlueSkyFS *fs = bluesky_deserialize_superblock(data->data);
90 g_print("Loaded filesystem superblock\n");
92 fs->name = g_strdup(name);
95 bluesky_string_unref(data);
98 g_print("Initializing fresh filesystem\n");
99 BlueSkyFS *fs = bluesky_new_fs(name);
102 BlueSkyInode *root = bluesky_new_inode(BLUESKY_ROOT_INUM, fs,
106 bluesky_insert_inode(fs, root);
108 bluesky_inode_flush(fs, root);
109 bluesky_superblock_flush(fs);
114 /* Inode reference counting. */
115 void bluesky_inode_ref(BlueSkyInode *inode)
117 g_atomic_int_inc(&inode->refcount);
120 void bluesky_inode_unref(BlueSkyInode *inode)
122 if (g_atomic_int_dec_and_test(&inode->refcount)) {
123 g_print("Reference count for inode %"PRIu64" dropped to zero.\n",
126 /* Sanity check: Is the inode clean? */
127 if (inode->change_commit < inode->change_count) {
128 g_warning("Dropping inode which is not clean (commit %"PRIi64" < change %"PRIi64")\n", inode->change_commit, inode->change_count);
131 /* Free file type specific data. It should be an error for there to be
132 * dirty data to commit when the reference count has reaches zero. */
133 switch (inode->type) {
134 case BLUESKY_REGULAR:
135 for (int i = 0; i < inode->blocks->len; i++) {
136 BlueSkyBlock *b = &g_array_index(inode->blocks,
138 if (b->type == BLUESKY_BLOCK_DIRTY) {
139 g_error("Deleting an inode with dirty file data!");
142 bluesky_string_unref(b->data);
144 g_array_unref(inode->blocks);
147 case BLUESKY_DIRECTORY:
148 g_hash_table_destroy(inode->dirhash);
149 g_hash_table_destroy(inode->dirhash_folded);
150 g_sequence_free(inode->dirents);
153 case BLUESKY_SYMLINK:
154 g_free(inode->symlink_contents);
161 g_mutex_free(inode->lock);
167 /* Allocate a fresh inode number which has not been used before within a
168 * filesystem. fs must already be locked. */
169 uint64_t bluesky_fs_alloc_inode(BlueSkyFS *fs)
173 inum = fs->next_inum;
176 bluesky_superblock_flush(fs);
181 /* Perform type-specification initialization of an inode. Normally performed
182 * in bluesky_new_inode, but can be separated if an inode is created first,
183 * then deserialized. */
184 void bluesky_init_inode(BlueSkyInode *i, BlueSkyFileType type)
189 case BLUESKY_REGULAR:
190 i->blocks = g_array_new(FALSE, TRUE, sizeof(BlueSkyBlock));
192 case BLUESKY_DIRECTORY:
193 i->dirents = g_sequence_new(bluesky_dirent_destroy);
194 i->dirhash = g_hash_table_new(g_str_hash, g_str_equal);
195 i->dirhash_folded = g_hash_table_new(g_str_hash, g_str_equal);
202 BlueSkyInode *bluesky_new_inode(uint64_t inum, BlueSkyFS *fs,
203 BlueSkyFileType type)
205 BlueSkyInode *i = g_new0(BlueSkyInode, 1);
207 i->lock = g_mutex_new();
212 bluesky_init_inode(i, type);
217 /* Retrieve an inode from the filesystem. Eventually this will be a cache and
218 * so we might need to go fetch the inode from elsewhere; for now all
219 * filesystem state is stored here. inode is returned with a reference held
221 BlueSkyInode *bluesky_get_inode(BlueSkyFS *fs, uint64_t inum)
223 BlueSkyInode *inode = NULL;
229 g_mutex_lock(fs->lock);
230 inode = (BlueSkyInode *)g_hash_table_lookup(fs->inodes, &inum);
233 bluesky_inode_fetch(fs, inum);
234 inode = (BlueSkyInode *)g_hash_table_lookup(fs->inodes, &inum);
238 bluesky_inode_ref(inode);
240 /* FIXME: We assume we can atomically update the in-memory access time
242 inode->access_time = bluesky_get_current_time();
245 g_mutex_unlock(fs->lock);
250 /* Insert an inode into the filesystem inode cache. fs should be locked. */
251 void bluesky_insert_inode(BlueSkyFS *fs, BlueSkyInode *inode)
253 g_hash_table_insert(fs->inodes, &inode->inum, inode);
256 /* Deprecated: Synchronize an inode to stable storage. */
257 void bluesky_inode_flush(BlueSkyFS *fs, BlueSkyInode *inode)
259 GString *buf = g_string_new("");
260 bluesky_serialize_inode(buf, inode);
261 BlueSkyRCStr *data = bluesky_string_new_from_gstring(buf);
264 sprintf(key, "inode-%016"PRIx64, inode->inum);
266 BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
267 async->op = STORE_OP_PUT;
268 async->key = g_strdup(key);
270 bluesky_store_async_submit(async);
271 bluesky_store_async_unref(async);
274 /* Start writeback of an inode and all associated data. */
275 void bluesky_inode_start_sync(BlueSkyInode *inode, BlueSkyStoreAsync *barrier)
277 BlueSkyFS *fs = inode->fs;
279 if (inode->type == BLUESKY_REGULAR)
280 bluesky_file_flush(inode, barrier);
282 GString *buf = g_string_new("");
283 bluesky_serialize_inode(buf, inode);
284 BlueSkyRCStr *data = bluesky_string_new_from_gstring(buf);
287 sprintf(key, "inode-%016"PRIx64, inode->inum);
289 BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
290 async->op = STORE_OP_PUT;
291 async->key = g_strdup(key);
293 bluesky_store_async_submit(async);
295 bluesky_store_add_barrier(barrier, async);
296 bluesky_store_async_unref(async);
299 /* Write back an inode and all associated data and wait for completion. Inode
300 * should already be locked. */
301 void bluesky_inode_do_sync(BlueSkyInode *inode)
303 BlueSkyStoreAsync *barrier = bluesky_store_async_new(inode->fs->store);
304 barrier->op = STORE_OP_BARRIER;
306 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
307 "Synchronous writeback for inode %"PRIu64"...", inode->inum);
308 bluesky_inode_start_sync(inode, barrier);
309 bluesky_store_async_submit(barrier);
310 bluesky_store_async_wait(barrier);
311 bluesky_store_async_unref(barrier);
312 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
313 "Writeback for inode %"PRIu64" complete", inode->inum);
316 static void complete_inode_fetch(BlueSkyStoreAsync *async, BlueSkyInode *inode)
318 g_print("Completing fetch of inode %"PRIu64"...\n", inode->inum);
320 if (async->result != 0
321 || !bluesky_deserialize_inode(inode, async->data->data))
323 g_print(" failed to load inode, cleaning up\n");
324 g_mutex_lock(inode->fs->lock);
325 g_hash_table_remove(inode->fs->inodes, &inode->inum);
326 g_mutex_unlock(inode->fs->lock);
327 bluesky_inode_unref(inode);
330 inode->access_time = bluesky_get_current_time();
332 g_mutex_unlock(inode->lock);
333 bluesky_inode_unref(inode);
336 /* Fetch an inode from stable storage. The fetch can be performed
337 * asynchronously: the in-memory inode is allocated, but not filled with data
338 * immediately. It is kept locked until it has been filled in, so any users
339 * should try to acquire the lock on the inode before accessing any data. The
340 * fs lock must be held. */
341 void bluesky_inode_fetch(BlueSkyFS *fs, uint64_t inum)
344 sprintf(key, "inode-%016"PRIx64, inum);
346 BlueSkyInode *inode = bluesky_new_inode(inum, fs, BLUESKY_PENDING);
347 bluesky_inode_ref(inode); // Extra ref held by fetching process
348 g_mutex_lock(inode->lock);
349 bluesky_insert_inode(fs, inode);
351 BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
352 async->op = STORE_OP_GET;
353 async->key = g_strdup(key);
355 bluesky_store_async_add_notifier(async, (GFunc)complete_inode_fetch, inode);
356 bluesky_store_async_submit(async);
358 if (bluesky_options.sync_inode_fetches) {
359 bluesky_store_async_wait(async);
362 bluesky_store_async_unref(async);
365 /* Synchronize filesystem superblock to stable storage. */
366 void bluesky_superblock_flush(BlueSkyFS *fs)
368 GString *buf = g_string_new("");
369 bluesky_serialize_superblock(buf, fs);
370 BlueSkyRCStr *data = bluesky_string_new_from_gstring(buf);
372 BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
373 async->op = STORE_OP_PUT;
374 async->key = g_strdup("superblock");
376 bluesky_store_async_submit(async);
377 bluesky_store_async_unref(async);
379 bluesky_store_sync(fs->store);