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);
97 g_print("Initializing fresh filesystem\n");
98 BlueSkyFS *fs = bluesky_new_fs(name);
101 BlueSkyInode *root = bluesky_new_inode(BLUESKY_ROOT_INUM, fs,
105 bluesky_insert_inode(fs, root);
107 bluesky_inode_flush(fs, root);
108 bluesky_superblock_flush(fs);
113 /* Inode reference counting. */
114 void bluesky_inode_ref(BlueSkyInode *inode)
116 g_atomic_int_inc(&inode->refcount);
119 void bluesky_inode_unref(BlueSkyInode *inode)
121 if (g_atomic_int_dec_and_test(&inode->refcount)) {
122 g_error("Reference count for inode %"PRIu64" dropped to zero!\n",
127 /* Allocate a fresh inode number which has not been used before within a
128 * filesystem. fs must already be locked. */
129 uint64_t bluesky_fs_alloc_inode(BlueSkyFS *fs)
133 inum = fs->next_inum;
136 bluesky_superblock_flush(fs);
141 /* Perform type-specification initialization of an inode. Normally performed
142 * in bluesky_new_inode, but can be separated if an inode is created first,
143 * then deserialized. */
144 void bluesky_init_inode(BlueSkyInode *i, BlueSkyFileType type)
149 case BLUESKY_REGULAR:
150 i->blocks = g_array_new(FALSE, TRUE, sizeof(BlueSkyBlock));
152 case BLUESKY_DIRECTORY:
153 i->dirents = g_sequence_new(bluesky_dirent_destroy);
154 i->dirhash = g_hash_table_new(g_str_hash, g_str_equal);
155 i->dirhash_folded = g_hash_table_new(g_str_hash, g_str_equal);
162 BlueSkyInode *bluesky_new_inode(uint64_t inum, BlueSkyFS *fs,
163 BlueSkyFileType type)
165 BlueSkyInode *i = g_new0(BlueSkyInode, 1);
167 i->lock = g_mutex_new();
172 bluesky_init_inode(i, type);
177 /* Retrieve an inode from the filesystem. Eventually this will be a cache and
178 * so we might need to go fetch the inode from elsewhere; for now all
179 * filesystem state is stored here. inode is returned with a reference held
181 BlueSkyInode *bluesky_get_inode(BlueSkyFS *fs, uint64_t inum)
183 BlueSkyInode *inode = NULL;
189 g_mutex_lock(fs->lock);
190 inode = (BlueSkyInode *)g_hash_table_lookup(fs->inodes, &inum);
193 bluesky_inode_fetch(fs, inum);
194 inode = (BlueSkyInode *)g_hash_table_lookup(fs->inodes, &inum);
198 bluesky_inode_ref(inode);
201 g_mutex_unlock(fs->lock);
206 /* Insert an inode into the filesystem inode cache. fs should be locked. */
207 void bluesky_insert_inode(BlueSkyFS *fs, BlueSkyInode *inode)
209 g_hash_table_insert(fs->inodes, &inode->inum, inode);
212 /* Deprecated: Synchronize an inode to stable storage. */
213 void bluesky_inode_flush(BlueSkyFS *fs, BlueSkyInode *inode)
215 GString *buf = g_string_new("");
216 bluesky_serialize_inode(buf, inode);
217 BlueSkyRCStr *data = bluesky_string_new_from_gstring(buf);
220 sprintf(key, "inode-%016"PRIx64, inode->inum);
222 BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
223 async->op = STORE_OP_PUT;
224 async->key = g_strdup(key);
226 bluesky_store_async_submit(async);
227 bluesky_store_async_unref(async);
230 /* Start writeback of an inode and all associated data. */
231 void bluesky_inode_start_sync(BlueSkyInode *inode, BlueSkyStoreAsync *barrier)
233 BlueSkyFS *fs = inode->fs;
235 GString *buf = g_string_new("");
236 bluesky_serialize_inode(buf, inode);
237 BlueSkyRCStr *data = bluesky_string_new_from_gstring(buf);
239 if (inode->type == BLUESKY_REGULAR)
240 bluesky_file_flush(inode, barrier);
243 sprintf(key, "inode-%016"PRIx64, inode->inum);
245 BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
246 async->op = STORE_OP_PUT;
247 async->key = g_strdup(key);
249 bluesky_store_async_submit(async);
251 bluesky_store_add_barrier(barrier, async);
252 bluesky_store_async_unref(async);
255 /* Fetch an inode from stable storage. */
256 void bluesky_inode_fetch(BlueSkyFS *fs, uint64_t inum)
259 sprintf(key, "inode-%016"PRIx64, inum);
260 BlueSkyRCStr *data = bluesky_store_get(fs->store, key);
264 BlueSkyInode *inode = bluesky_new_inode(inum, fs, BLUESKY_PENDING);
265 bluesky_inode_ref(inode);
266 g_mutex_lock(inode->lock);
267 bluesky_insert_inode(fs, inode);
269 if (!bluesky_deserialize_inode(inode, data->data)) {
270 g_hash_table_remove(fs->inodes, &inode->inum);
271 bluesky_inode_unref(inode);
274 g_mutex_unlock(inode->lock);
275 bluesky_inode_unref(inode);
278 /* Synchronize filesystem superblock to stable storage. */
279 void bluesky_superblock_flush(BlueSkyFS *fs)
281 GString *buf = g_string_new("");
282 bluesky_serialize_superblock(buf, fs);
283 BlueSkyRCStr *data = bluesky_string_new_from_gstring(buf);
285 BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
286 async->op = STORE_OP_PUT;
287 async->key = g_strdup("superblock");
289 bluesky_store_async_submit(async);
290 bluesky_store_async_unref(async);
292 bluesky_store_sync(fs->store);