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->dirty_list, inode->dirty_list);
52 inode->dirty_list = bluesky_list_prepend(&inode->fs->dirty_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();
93 BlueSkyFS *bluesky_init_fs(gchar *name, BlueSkyStore *store)
95 BlueSkyRCStr *data = bluesky_store_get(store, "superblock");
97 BlueSkyFS *fs = bluesky_deserialize_superblock(data->data);
100 fs->log = bluesky_log_new("journal");
101 g_print("Loaded filesystem superblock\n");
103 fs->name = g_strdup(name);
106 bluesky_string_unref(data);
109 g_print("Initializing fresh filesystem\n");
110 BlueSkyFS *fs = bluesky_new_fs(name);
112 fs->log = bluesky_log_new("journal");
114 BlueSkyInode *root = bluesky_new_inode(BLUESKY_ROOT_INUM, fs,
118 bluesky_insert_inode(fs, root);
119 bluesky_inode_update_ctime(root, TRUE);
121 bluesky_inode_flush(fs, root);
122 bluesky_superblock_flush(fs);
127 /* Inode reference counting. */
128 void bluesky_inode_ref(BlueSkyInode *inode)
130 g_atomic_int_inc(&inode->refcount);
133 void bluesky_inode_unref(BlueSkyInode *inode)
135 if (g_atomic_int_dec_and_test(&inode->refcount)) {
136 if (bluesky_verbose) {
137 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
138 "Reference count for inode %"PRIu64" dropped to zero.",
142 /* Sanity check: Is the inode clean? */
143 if (inode->change_commit < inode->change_count
144 || inode->accessed_list != NULL
145 || inode->dirty_list != NULL) {
146 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);
149 /* These shouldn't be needed, but in case the above warning fires and
150 * we delete the inode anyway, we ought to be sure the inode is not on
152 g_mutex_lock(inode->fs->lock);
153 bluesky_list_unlink(&inode->fs->accessed_list, inode->accessed_list);
154 bluesky_list_unlink(&inode->fs->dirty_list, inode->dirty_list);
155 g_mutex_unlock(inode->fs->lock);
157 /* Free file type specific data. It should be an error for there to be
158 * dirty data to commit when the reference count has reaches zero. */
159 switch (inode->type) {
160 case BLUESKY_REGULAR:
161 for (int i = 0; i < inode->blocks->len; i++) {
162 BlueSkyBlock *b = &g_array_index(inode->blocks,
164 if (b->type == BLUESKY_BLOCK_DIRTY) {
165 g_error("Deleting an inode with dirty file data!");
168 bluesky_string_unref(b->data);
170 g_array_unref(inode->blocks);
173 case BLUESKY_DIRECTORY:
174 g_hash_table_destroy(inode->dirhash);
175 g_hash_table_destroy(inode->dirhash_folded);
176 g_sequence_free(inode->dirents);
179 case BLUESKY_SYMLINK:
180 g_free(inode->symlink_contents);
187 g_mutex_free(inode->lock);
193 /* Allocate a fresh inode number which has not been used before within a
194 * filesystem. fs must already be locked. */
195 uint64_t bluesky_fs_alloc_inode(BlueSkyFS *fs)
199 inum = fs->next_inum;
202 bluesky_superblock_flush(fs);
207 /* Perform type-specification initialization of an inode. Normally performed
208 * in bluesky_new_inode, but can be separated if an inode is created first,
209 * then deserialized. */
210 void bluesky_init_inode(BlueSkyInode *i, BlueSkyFileType type)
215 case BLUESKY_REGULAR:
216 i->blocks = g_array_new(FALSE, TRUE, sizeof(BlueSkyBlock));
218 case BLUESKY_DIRECTORY:
219 i->dirents = g_sequence_new(bluesky_dirent_destroy);
220 i->dirhash = g_hash_table_new(g_str_hash, g_str_equal);
221 i->dirhash_folded = g_hash_table_new(g_str_hash, g_str_equal);
228 BlueSkyInode *bluesky_new_inode(uint64_t inum, BlueSkyFS *fs,
229 BlueSkyFileType type)
231 BlueSkyInode *i = g_new0(BlueSkyInode, 1);
233 i->lock = g_mutex_new();
238 bluesky_init_inode(i, type);
243 /* Retrieve an inode from the filesystem. Eventually this will be a cache and
244 * so we might need to go fetch the inode from elsewhere; for now all
245 * filesystem state is stored here. inode is returned with a reference held
247 BlueSkyInode *bluesky_get_inode(BlueSkyFS *fs, uint64_t inum)
249 BlueSkyInode *inode = NULL;
255 g_mutex_lock(fs->lock);
256 inode = (BlueSkyInode *)g_hash_table_lookup(fs->inodes, &inum);
259 bluesky_inode_fetch(fs, inum);
260 inode = (BlueSkyInode *)g_hash_table_lookup(fs->inodes, &inum);
264 bluesky_inode_ref(inode);
266 /* FIXME: We assume we can atomically update the in-memory access time
268 inode->access_time = bluesky_get_current_time();
271 g_mutex_unlock(fs->lock);
276 /* Insert an inode into the filesystem inode cache. fs should be locked. */
277 void bluesky_insert_inode(BlueSkyFS *fs, BlueSkyInode *inode)
279 g_hash_table_insert(fs->inodes, &inode->inum, inode);
282 /* Deprecated: Synchronize an inode to stable storage. */
283 void bluesky_inode_flush(BlueSkyFS *fs, BlueSkyInode *inode)
285 GString *buf = g_string_new("");
286 bluesky_serialize_inode(buf, inode);
287 BlueSkyRCStr *data = bluesky_string_new_from_gstring(buf);
290 sprintf(key, "inode-%016"PRIx64, inode->inum);
292 BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
293 async->op = STORE_OP_PUT;
294 async->key = g_strdup(key);
296 bluesky_store_async_submit(async);
297 bluesky_store_async_unref(async);
300 /* Start writeback of an inode and all associated data. */
301 void bluesky_inode_start_sync(BlueSkyInode *inode, BlueSkyStoreAsync *barrier)
303 GList *log_items = NULL;
304 BlueSkyFS *fs = inode->fs;
306 if (inode->type == BLUESKY_REGULAR)
307 bluesky_file_flush(inode, barrier, &log_items);
309 GString *buf = g_string_new("");
310 bluesky_serialize_inode(buf, inode);
311 BlueSkyRCStr *data = bluesky_string_new_from_gstring(buf);
314 sprintf(key, "inode-%016"PRIx64, inode->inum);
316 BlueSkyLogItem *log_item = bluesky_log_item_new();
317 log_item->key = g_strdup(key);
318 log_item->data = data;
319 bluesky_string_ref(data);
320 bluesky_log_item_submit(log_item, fs->log);
321 log_items = g_list_prepend(log_items, log_item);
323 /* Wait for all log items to be committed to disk. */
324 while (log_items != NULL) {
325 log_item = (BlueSkyLogItem *)log_items->data;
326 bluesky_log_item_finish(log_item);
327 log_items = g_list_delete_link(log_items, log_items);
330 BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
331 async->op = STORE_OP_PUT;
332 async->key = g_strdup(key);
334 bluesky_store_async_submit(async);
336 bluesky_store_add_barrier(barrier, async);
337 bluesky_store_async_unref(async);
340 /* Write back an inode and all associated data and wait for completion. Inode
341 * should already be locked. */
342 void bluesky_inode_do_sync(BlueSkyInode *inode)
344 BlueSkyStoreAsync *barrier = bluesky_store_async_new(inode->fs->store);
345 barrier->op = STORE_OP_BARRIER;
347 if (bluesky_verbose) {
348 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
349 "Synchronous writeback for inode %"PRIu64"...", inode->inum);
351 bluesky_inode_start_sync(inode, barrier);
352 bluesky_store_async_submit(barrier);
353 bluesky_store_async_wait(barrier);
354 bluesky_store_async_unref(barrier);
355 if (bluesky_verbose) {
356 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
357 "Writeback for inode %"PRIu64" complete", inode->inum);
361 static void complete_inode_fetch(BlueSkyStoreAsync *async, BlueSkyInode *inode)
363 if (bluesky_verbose) {
364 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
365 "Completing fetch of inode %"PRIu64"...", inode->inum);
368 if (async->result != 0
369 || !bluesky_deserialize_inode(inode, async->data->data))
371 if (bluesky_verbose) {
372 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
373 " failed to load inode, cleaning up");
375 g_mutex_lock(inode->fs->lock);
376 g_hash_table_remove(inode->fs->inodes, &inode->inum);
377 bluesky_list_unlink(&inode->fs->accessed_list, inode->accessed_list);
378 inode->accessed_list = NULL;
379 g_mutex_unlock(inode->fs->lock);
380 bluesky_inode_unref(inode);
383 inode->access_time = bluesky_get_current_time();
384 g_mutex_lock(inode->fs->lock);
385 bluesky_list_unlink(&inode->fs->accessed_list, inode->accessed_list);
386 inode->accessed_list = bluesky_list_prepend(&inode->fs->accessed_list, inode);
387 g_mutex_unlock(inode->fs->lock);
389 g_mutex_unlock(inode->lock);
390 bluesky_inode_unref(inode);
393 /* Fetch an inode from stable storage. The fetch can be performed
394 * asynchronously: the in-memory inode is allocated, but not filled with data
395 * immediately. It is kept locked until it has been filled in, so any users
396 * should try to acquire the lock on the inode before accessing any data. The
397 * fs lock must be held. */
398 void bluesky_inode_fetch(BlueSkyFS *fs, uint64_t inum)
401 sprintf(key, "inode-%016"PRIx64, inum);
403 BlueSkyInode *inode = bluesky_new_inode(inum, fs, BLUESKY_PENDING);
404 inode->change_count = 0;
405 bluesky_inode_ref(inode); // Extra ref held by fetching process
406 g_mutex_lock(inode->lock);
407 bluesky_insert_inode(fs, inode);
409 BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
410 async->op = STORE_OP_GET;
411 async->key = g_strdup(key);
413 bluesky_store_async_add_notifier(async, (GFunc)complete_inode_fetch, inode);
414 bluesky_store_async_submit(async);
416 if (bluesky_options.sync_inode_fetches) {
417 bluesky_store_async_wait(async);
420 bluesky_store_async_unref(async);
423 /* Synchronize filesystem superblock to stable storage. */
424 void bluesky_superblock_flush(BlueSkyFS *fs)
426 GString *buf = g_string_new("");
427 bluesky_serialize_superblock(buf, fs);
428 BlueSkyRCStr *data = bluesky_string_new_from_gstring(buf);
430 BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
431 async->op = STORE_OP_PUT;
432 async->key = g_strdup("superblock");
434 bluesky_store_async_submit(async);
435 bluesky_store_async_unref(async);
437 //bluesky_store_sync(fs->store);