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);
48 g_mutex_lock(inode->fs->lock);
49 bluesky_list_unlink(&inode->fs->dirty_list, inode->dirty_list);
50 inode->dirty_list = bluesky_list_prepend(&inode->fs->dirty_list, inode);
51 bluesky_list_unlink(&inode->fs->accessed_list, inode->accessed_list);
52 inode->accessed_list = bluesky_list_prepend(&inode->fs->accessed_list, inode);
53 g_mutex_unlock(inode->fs->lock);
56 /* Unfortunately a glib hash table is only guaranteed to be able to store
57 * 32-bit keys if we use the key directly. If we want 64-bit inode numbers,
58 * we'll have to allocate memory to store the 64-bit inumber, and use a pointer
59 * to it. Rather than allocate the memory for the key, we'll just include a
60 * pointer to the 64-bit inum stored in the inode itself, so that we don't need
61 * to do any more memory management. */
62 static guint bluesky_fs_key_hash_func(gconstpointer key)
64 uint64_t inum = *(const uint64_t *)key;
68 static gboolean bluesky_fs_key_equal_func(gconstpointer a, gconstpointer b)
70 uint64_t i1 = *(const uint64_t *)a;
71 uint64_t i2 = *(const uint64_t *)b;
75 /* Filesystem-level operations. A filesystem is like a directory tree that we
76 * are willing to export. */
77 BlueSkyFS *bluesky_new_fs(gchar *name)
79 BlueSkyFS *fs = g_new0(BlueSkyFS, 1);
80 fs->lock = g_mutex_new();
81 fs->name = g_strdup(name);
82 fs->inodes = g_hash_table_new(bluesky_fs_key_hash_func,
83 bluesky_fs_key_equal_func);
84 fs->next_inum = BLUESKY_ROOT_INUM + 1;
85 fs->store = bluesky_store_new("file");
90 BlueSkyFS *bluesky_init_fs(gchar *name, BlueSkyStore *store)
92 BlueSkyRCStr *data = bluesky_store_get(store, "superblock");
94 BlueSkyFS *fs = bluesky_deserialize_superblock(data->data);
97 g_print("Loaded filesystem superblock\n");
99 fs->name = g_strdup(name);
102 bluesky_string_unref(data);
105 g_print("Initializing fresh filesystem\n");
106 BlueSkyFS *fs = bluesky_new_fs(name);
109 BlueSkyInode *root = bluesky_new_inode(BLUESKY_ROOT_INUM, fs,
113 bluesky_insert_inode(fs, root);
114 bluesky_inode_update_ctime(root, TRUE);
116 bluesky_inode_flush(fs, root);
117 bluesky_superblock_flush(fs);
122 /* Inode reference counting. */
123 void bluesky_inode_ref(BlueSkyInode *inode)
125 g_atomic_int_inc(&inode->refcount);
128 void bluesky_inode_unref(BlueSkyInode *inode)
130 if (g_atomic_int_dec_and_test(&inode->refcount)) {
131 g_print("Reference count for inode %"PRIu64" dropped to zero.\n",
134 /* Sanity check: Is the inode clean? */
135 if (inode->change_commit < inode->change_count
136 || inode->dirty_list != NULL) {
137 g_warning("Dropping inode which is not clean (commit %"PRIi64" < change %"PRIi64"; dirty_list = %p)\n", inode->change_commit, inode->change_count, inode->dirty_list);
140 g_mutex_lock(inode->fs->lock);
141 bluesky_list_unlink(&inode->fs->accessed_list, inode->accessed_list);
142 bluesky_list_unlink(&inode->fs->dirty_list, inode->dirty_list);
143 g_mutex_unlock(inode->fs->lock);
145 /* Free file type specific data. It should be an error for there to be
146 * dirty data to commit when the reference count has reaches zero. */
147 switch (inode->type) {
148 case BLUESKY_REGULAR:
149 for (int i = 0; i < inode->blocks->len; i++) {
150 BlueSkyBlock *b = &g_array_index(inode->blocks,
152 if (b->type == BLUESKY_BLOCK_DIRTY) {
153 g_error("Deleting an inode with dirty file data!");
156 bluesky_string_unref(b->data);
158 g_array_unref(inode->blocks);
161 case BLUESKY_DIRECTORY:
162 g_hash_table_destroy(inode->dirhash);
163 g_hash_table_destroy(inode->dirhash_folded);
164 g_sequence_free(inode->dirents);
167 case BLUESKY_SYMLINK:
168 g_free(inode->symlink_contents);
175 g_mutex_free(inode->lock);
181 /* Allocate a fresh inode number which has not been used before within a
182 * filesystem. fs must already be locked. */
183 uint64_t bluesky_fs_alloc_inode(BlueSkyFS *fs)
187 inum = fs->next_inum;
190 bluesky_superblock_flush(fs);
195 /* Perform type-specification initialization of an inode. Normally performed
196 * in bluesky_new_inode, but can be separated if an inode is created first,
197 * then deserialized. */
198 void bluesky_init_inode(BlueSkyInode *i, BlueSkyFileType type)
203 case BLUESKY_REGULAR:
204 i->blocks = g_array_new(FALSE, TRUE, sizeof(BlueSkyBlock));
206 case BLUESKY_DIRECTORY:
207 i->dirents = g_sequence_new(bluesky_dirent_destroy);
208 i->dirhash = g_hash_table_new(g_str_hash, g_str_equal);
209 i->dirhash_folded = g_hash_table_new(g_str_hash, g_str_equal);
216 BlueSkyInode *bluesky_new_inode(uint64_t inum, BlueSkyFS *fs,
217 BlueSkyFileType type)
219 BlueSkyInode *i = g_new0(BlueSkyInode, 1);
221 i->lock = g_mutex_new();
226 bluesky_init_inode(i, type);
231 /* Retrieve an inode from the filesystem. Eventually this will be a cache and
232 * so we might need to go fetch the inode from elsewhere; for now all
233 * filesystem state is stored here. inode is returned with a reference held
235 BlueSkyInode *bluesky_get_inode(BlueSkyFS *fs, uint64_t inum)
237 BlueSkyInode *inode = NULL;
243 g_mutex_lock(fs->lock);
244 inode = (BlueSkyInode *)g_hash_table_lookup(fs->inodes, &inum);
247 bluesky_inode_fetch(fs, inum);
248 inode = (BlueSkyInode *)g_hash_table_lookup(fs->inodes, &inum);
252 bluesky_inode_ref(inode);
254 /* FIXME: We assume we can atomically update the in-memory access time
256 inode->access_time = bluesky_get_current_time();
259 g_mutex_unlock(fs->lock);
264 /* Insert an inode into the filesystem inode cache. fs should be locked. */
265 void bluesky_insert_inode(BlueSkyFS *fs, BlueSkyInode *inode)
267 g_hash_table_insert(fs->inodes, &inode->inum, inode);
270 /* Deprecated: Synchronize an inode to stable storage. */
271 void bluesky_inode_flush(BlueSkyFS *fs, BlueSkyInode *inode)
273 GString *buf = g_string_new("");
274 bluesky_serialize_inode(buf, inode);
275 BlueSkyRCStr *data = bluesky_string_new_from_gstring(buf);
278 sprintf(key, "inode-%016"PRIx64, inode->inum);
280 BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
281 async->op = STORE_OP_PUT;
282 async->key = g_strdup(key);
284 bluesky_store_async_submit(async);
285 bluesky_store_async_unref(async);
288 /* Start writeback of an inode and all associated data. */
289 void bluesky_inode_start_sync(BlueSkyInode *inode, BlueSkyStoreAsync *barrier)
291 BlueSkyFS *fs = inode->fs;
293 if (inode->type == BLUESKY_REGULAR)
294 bluesky_file_flush(inode, barrier);
296 GString *buf = g_string_new("");
297 bluesky_serialize_inode(buf, inode);
298 BlueSkyRCStr *data = bluesky_string_new_from_gstring(buf);
301 sprintf(key, "inode-%016"PRIx64, inode->inum);
303 BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
304 async->op = STORE_OP_PUT;
305 async->key = g_strdup(key);
307 bluesky_store_async_submit(async);
309 bluesky_store_add_barrier(barrier, async);
310 bluesky_store_async_unref(async);
313 /* Write back an inode and all associated data and wait for completion. Inode
314 * should already be locked. */
315 void bluesky_inode_do_sync(BlueSkyInode *inode)
317 BlueSkyStoreAsync *barrier = bluesky_store_async_new(inode->fs->store);
318 barrier->op = STORE_OP_BARRIER;
320 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
321 "Synchronous writeback for inode %"PRIu64"...", inode->inum);
322 bluesky_inode_start_sync(inode, barrier);
323 bluesky_store_async_submit(barrier);
324 bluesky_store_async_wait(barrier);
325 bluesky_store_async_unref(barrier);
326 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
327 "Writeback for inode %"PRIu64" complete", inode->inum);
330 static void complete_inode_fetch(BlueSkyStoreAsync *async, BlueSkyInode *inode)
332 g_print("Completing fetch of inode %"PRIu64"...\n", inode->inum);
334 if (async->result != 0
335 || !bluesky_deserialize_inode(inode, async->data->data))
337 g_print(" failed to load inode, cleaning up\n");
338 g_mutex_lock(inode->fs->lock);
339 g_hash_table_remove(inode->fs->inodes, &inode->inum);
340 g_mutex_unlock(inode->fs->lock);
341 bluesky_inode_unref(inode);
344 inode->access_time = bluesky_get_current_time();
345 g_mutex_lock(inode->fs->lock);
346 inode->accessed_list = bluesky_list_prepend(&inode->fs->accessed_list, inode);
347 g_mutex_unlock(inode->fs->lock);
349 g_mutex_unlock(inode->lock);
350 bluesky_inode_unref(inode);
353 /* Fetch an inode from stable storage. The fetch can be performed
354 * asynchronously: the in-memory inode is allocated, but not filled with data
355 * immediately. It is kept locked until it has been filled in, so any users
356 * should try to acquire the lock on the inode before accessing any data. The
357 * fs lock must be held. */
358 void bluesky_inode_fetch(BlueSkyFS *fs, uint64_t inum)
361 sprintf(key, "inode-%016"PRIx64, inum);
363 BlueSkyInode *inode = bluesky_new_inode(inum, fs, BLUESKY_PENDING);
364 bluesky_inode_ref(inode); // Extra ref held by fetching process
365 g_mutex_lock(inode->lock);
366 bluesky_insert_inode(fs, inode);
368 BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
369 async->op = STORE_OP_GET;
370 async->key = g_strdup(key);
372 bluesky_store_async_add_notifier(async, (GFunc)complete_inode_fetch, inode);
373 bluesky_store_async_submit(async);
375 if (bluesky_options.sync_inode_fetches) {
376 bluesky_store_async_wait(async);
379 bluesky_store_async_unref(async);
382 /* Synchronize filesystem superblock to stable storage. */
383 void bluesky_superblock_flush(BlueSkyFS *fs)
385 GString *buf = g_string_new("");
386 bluesky_serialize_superblock(buf, fs);
387 BlueSkyRCStr *data = bluesky_string_new_from_gstring(buf);
389 BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
390 async->op = STORE_OP_PUT;
391 async->key = g_strdup("superblock");
393 bluesky_store_async_submit(async);
394 bluesky_store_async_unref(async);
396 bluesky_store_sync(fs->store);