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_error("Reference count for inode %"PRIu64" dropped to zero!\n",
128 /* Allocate a fresh inode number which has not been used before within a
129 * filesystem. fs must already be locked. */
130 uint64_t bluesky_fs_alloc_inode(BlueSkyFS *fs)
134 inum = fs->next_inum;
137 bluesky_superblock_flush(fs);
142 /* Perform type-specification initialization of an inode. Normally performed
143 * in bluesky_new_inode, but can be separated if an inode is created first,
144 * then deserialized. */
145 void bluesky_init_inode(BlueSkyInode *i, BlueSkyFileType type)
150 case BLUESKY_REGULAR:
151 i->blocks = g_array_new(FALSE, TRUE, sizeof(BlueSkyBlock));
153 case BLUESKY_DIRECTORY:
154 i->dirents = g_sequence_new(bluesky_dirent_destroy);
155 i->dirhash = g_hash_table_new(g_str_hash, g_str_equal);
156 i->dirhash_folded = g_hash_table_new(g_str_hash, g_str_equal);
163 BlueSkyInode *bluesky_new_inode(uint64_t inum, BlueSkyFS *fs,
164 BlueSkyFileType type)
166 BlueSkyInode *i = g_new0(BlueSkyInode, 1);
168 i->lock = g_mutex_new();
173 bluesky_init_inode(i, type);
178 /* Retrieve an inode from the filesystem. Eventually this will be a cache and
179 * so we might need to go fetch the inode from elsewhere; for now all
180 * filesystem state is stored here. inode is returned with a reference held
182 BlueSkyInode *bluesky_get_inode(BlueSkyFS *fs, uint64_t inum)
184 BlueSkyInode *inode = NULL;
190 g_mutex_lock(fs->lock);
191 inode = (BlueSkyInode *)g_hash_table_lookup(fs->inodes, &inum);
194 bluesky_inode_fetch(fs, inum);
195 inode = (BlueSkyInode *)g_hash_table_lookup(fs->inodes, &inum);
199 bluesky_inode_ref(inode);
202 g_mutex_unlock(fs->lock);
207 /* Insert an inode into the filesystem inode cache. fs should be locked. */
208 void bluesky_insert_inode(BlueSkyFS *fs, BlueSkyInode *inode)
210 g_hash_table_insert(fs->inodes, &inode->inum, inode);
213 /* Deprecated: Synchronize an inode to stable storage. */
214 void bluesky_inode_flush(BlueSkyFS *fs, BlueSkyInode *inode)
216 GString *buf = g_string_new("");
217 bluesky_serialize_inode(buf, inode);
218 BlueSkyRCStr *data = bluesky_string_new_from_gstring(buf);
221 sprintf(key, "inode-%016"PRIx64, inode->inum);
223 BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
224 async->op = STORE_OP_PUT;
225 async->key = g_strdup(key);
227 bluesky_store_async_submit(async);
228 bluesky_store_async_unref(async);
231 /* Start writeback of an inode and all associated data. */
232 void bluesky_inode_start_sync(BlueSkyInode *inode, BlueSkyStoreAsync *barrier)
234 BlueSkyFS *fs = inode->fs;
236 if (inode->type == BLUESKY_REGULAR)
237 bluesky_file_flush(inode, barrier);
239 GString *buf = g_string_new("");
240 bluesky_serialize_inode(buf, inode);
241 BlueSkyRCStr *data = bluesky_string_new_from_gstring(buf);
244 sprintf(key, "inode-%016"PRIx64, inode->inum);
246 BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
247 async->op = STORE_OP_PUT;
248 async->key = g_strdup(key);
250 bluesky_store_async_submit(async);
252 bluesky_store_add_barrier(barrier, async);
253 bluesky_store_async_unref(async);
256 /* Write back an inode and all associated data and wait for completion. Inode
257 * should already be locked. */
258 void bluesky_inode_do_sync(BlueSkyInode *inode)
260 BlueSkyStoreAsync *barrier = bluesky_store_async_new(inode->fs->store);
261 barrier->op = STORE_OP_BARRIER;
263 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
264 "Synchronous writeback for inode %"PRIu64"...", inode->inum);
265 bluesky_inode_start_sync(inode, barrier);
266 bluesky_store_async_submit(barrier);
267 bluesky_store_async_wait(barrier);
268 bluesky_store_async_unref(barrier);
269 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
270 "Writeback for inode %"PRIu64" complete", inode->inum);
273 static void complete_inode_fetch(BlueSkyStoreAsync *async, BlueSkyInode *inode)
275 g_print("Completing fetch of inode %"PRIu64"...\n", inode->inum);
277 if (async->result != 0
278 || !bluesky_deserialize_inode(inode, async->data->data))
280 g_print(" failed to load inode, cleaning up\n");
281 g_mutex_lock(inode->fs->lock);
282 g_hash_table_remove(inode->fs->inodes, &inode->inum);
283 g_mutex_unlock(inode->fs->lock);
284 bluesky_inode_unref(inode);
287 g_mutex_unlock(inode->lock);
288 bluesky_inode_unref(inode);
291 /* Fetch an inode from stable storage. The fetch can be performed
292 * asynchronously: the in-memory inode is allocated, but not filled with data
293 * immediately. It is kept locked until it has been filled in, so any users
294 * should try to acquire the lock on the inode before accessing any data. The
295 * fs lock must be held. */
296 void bluesky_inode_fetch(BlueSkyFS *fs, uint64_t inum)
299 sprintf(key, "inode-%016"PRIx64, inum);
301 BlueSkyInode *inode = bluesky_new_inode(inum, fs, BLUESKY_PENDING);
302 bluesky_inode_ref(inode); // Extra ref held by fetching process
303 g_mutex_lock(inode->lock);
304 bluesky_insert_inode(fs, inode);
306 BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
307 async->op = STORE_OP_GET;
308 async->key = g_strdup(key);
310 bluesky_store_async_add_notifier(async, (GFunc)complete_inode_fetch, inode);
311 bluesky_store_async_submit(async);
313 if (bluesky_options.sync_inode_fetches) {
314 bluesky_store_async_wait(async);
317 bluesky_store_async_unref(async);
320 /* Synchronize filesystem superblock to stable storage. */
321 void bluesky_superblock_flush(BlueSkyFS *fs)
323 GString *buf = g_string_new("");
324 bluesky_serialize_superblock(buf, fs);
325 BlueSkyRCStr *data = bluesky_string_new_from_gstring(buf);
327 BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
328 async->op = STORE_OP_PUT;
329 async->key = g_strdup("superblock");
331 bluesky_store_async_submit(async);
332 bluesky_store_async_unref(async);
334 bluesky_store_sync(fs->store);