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();
89 fs->locations = g_hash_table_new(bluesky_cloudlog_hash,
90 bluesky_cloudlog_equal);
95 BlueSkyFS *bluesky_init_fs(gchar *name, BlueSkyStore *store)
97 BlueSkyRCStr *data = bluesky_store_get(store, "superblock");
99 BlueSkyFS *fs = bluesky_deserialize_superblock(data->data);
102 fs->log = bluesky_log_new("journal");
103 g_print("Loaded filesystem superblock\n");
105 fs->name = g_strdup(name);
108 bluesky_string_unref(data);
111 g_print("Initializing fresh filesystem\n");
112 BlueSkyFS *fs = bluesky_new_fs(name);
114 fs->log = bluesky_log_new("journal");
116 BlueSkyInode *root = bluesky_new_inode(BLUESKY_ROOT_INUM, fs,
120 bluesky_insert_inode(fs, root);
121 bluesky_inode_update_ctime(root, TRUE);
123 bluesky_inode_flush(fs, root);
124 bluesky_superblock_flush(fs);
129 /* Inode reference counting. */
130 void bluesky_inode_ref(BlueSkyInode *inode)
132 g_atomic_int_inc(&inode->refcount);
135 void bluesky_inode_unref(BlueSkyInode *inode)
137 if (g_atomic_int_dec_and_test(&inode->refcount)) {
138 if (bluesky_verbose) {
139 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
140 "Reference count for inode %"PRIu64" dropped to zero.",
144 /* Sanity check: Is the inode clean? */
145 if (inode->change_commit < inode->change_count
146 || inode->accessed_list != NULL
147 || inode->dirty_list != NULL) {
148 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);
151 /* These shouldn't be needed, but in case the above warning fires and
152 * we delete the inode anyway, we ought to be sure the inode is not on
154 g_mutex_lock(inode->fs->lock);
155 bluesky_list_unlink(&inode->fs->accessed_list, inode->accessed_list);
156 bluesky_list_unlink(&inode->fs->dirty_list, inode->dirty_list);
157 g_mutex_unlock(inode->fs->lock);
159 /* Free file type specific data. It should be an error for there to be
160 * dirty data to commit when the reference count has reaches zero. */
161 switch (inode->type) {
162 case BLUESKY_REGULAR:
163 for (int i = 0; i < inode->blocks->len; i++) {
164 BlueSkyBlock *b = &g_array_index(inode->blocks,
166 if (b->type == BLUESKY_BLOCK_DIRTY) {
167 g_error("Deleting an inode with dirty file data!");
170 bluesky_string_unref(b->data);
172 g_array_unref(inode->blocks);
175 case BLUESKY_DIRECTORY:
176 g_hash_table_destroy(inode->dirhash);
177 g_hash_table_destroy(inode->dirhash_folded);
178 g_sequence_free(inode->dirents);
181 case BLUESKY_SYMLINK:
182 g_free(inode->symlink_contents);
189 g_mutex_free(inode->lock);
195 /* Allocate a fresh inode number which has not been used before within a
196 * filesystem. fs must already be locked. */
197 uint64_t bluesky_fs_alloc_inode(BlueSkyFS *fs)
201 inum = fs->next_inum;
204 bluesky_superblock_flush(fs);
209 /* Perform type-specification initialization of an inode. Normally performed
210 * in bluesky_new_inode, but can be separated if an inode is created first,
211 * then deserialized. */
212 void bluesky_init_inode(BlueSkyInode *i, BlueSkyFileType type)
217 case BLUESKY_REGULAR:
218 i->blocks = g_array_new(FALSE, TRUE, sizeof(BlueSkyBlock));
220 case BLUESKY_DIRECTORY:
221 i->dirents = g_sequence_new(bluesky_dirent_destroy);
222 i->dirhash = g_hash_table_new(g_str_hash, g_str_equal);
223 i->dirhash_folded = g_hash_table_new(g_str_hash, g_str_equal);
230 BlueSkyInode *bluesky_new_inode(uint64_t inum, BlueSkyFS *fs,
231 BlueSkyFileType type)
233 BlueSkyInode *i = g_new0(BlueSkyInode, 1);
235 i->lock = g_mutex_new();
240 bluesky_init_inode(i, type);
245 /* Retrieve an inode from the filesystem. Eventually this will be a cache and
246 * so we might need to go fetch the inode from elsewhere; for now all
247 * filesystem state is stored here. inode is returned with a reference held
249 BlueSkyInode *bluesky_get_inode(BlueSkyFS *fs, uint64_t inum)
251 BlueSkyInode *inode = NULL;
257 g_mutex_lock(fs->lock);
258 inode = (BlueSkyInode *)g_hash_table_lookup(fs->inodes, &inum);
261 bluesky_inode_fetch(fs, inum);
262 inode = (BlueSkyInode *)g_hash_table_lookup(fs->inodes, &inum);
266 bluesky_inode_ref(inode);
268 /* FIXME: We assume we can atomically update the in-memory access time
270 inode->access_time = bluesky_get_current_time();
273 g_mutex_unlock(fs->lock);
278 /* Insert an inode into the filesystem inode cache. fs should be locked. */
279 void bluesky_insert_inode(BlueSkyFS *fs, BlueSkyInode *inode)
281 g_hash_table_insert(fs->inodes, &inode->inum, inode);
284 /* Deprecated: Synchronize an inode to stable storage. */
285 void bluesky_inode_flush(BlueSkyFS *fs, BlueSkyInode *inode)
287 GString *buf = g_string_new("");
288 bluesky_serialize_inode(buf, inode);
289 BlueSkyRCStr *data = bluesky_string_new_from_gstring(buf);
292 sprintf(key, "inode-%016"PRIx64, inode->inum);
294 BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
295 async->op = STORE_OP_PUT;
296 async->key = g_strdup(key);
298 bluesky_store_async_submit(async);
299 bluesky_store_async_unref(async);
302 /* Start writeback of an inode and all associated data. */
303 void bluesky_inode_start_sync(BlueSkyInode *inode, BlueSkyStoreAsync *barrier)
305 GList *log_items = NULL;
306 BlueSkyFS *fs = inode->fs;
308 if (inode->type == BLUESKY_REGULAR)
309 bluesky_file_flush(inode, barrier, &log_items);
311 GString *buf = g_string_new("");
312 bluesky_serialize_inode(buf, inode);
313 BlueSkyRCStr *data = bluesky_string_new_from_gstring(buf);
316 sprintf(key, "inode-%016"PRIx64, inode->inum);
318 BlueSkyCloudLog *cloudlog = bluesky_cloudlog_new(fs);
319 cloudlog->type = LOGTYPE_DATA;
320 cloudlog->inum = inode->inum;
321 cloudlog->data = data;
322 bluesky_string_ref(data);
324 log_items = g_list_prepend(log_items, bluesky_cloudlog_sync(cloudlog));
326 bluesky_cloudlog_insert(cloudlog);
328 /* Wait for all log items to be committed to disk. */
329 while (log_items != NULL) {
330 BlueSkyLogItem *log_item = (BlueSkyLogItem *)log_items->data;
331 bluesky_log_item_finish(log_item);
332 log_items = g_list_delete_link(log_items, log_items);
335 BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
336 async->op = STORE_OP_PUT;
337 async->key = g_strdup(key);
339 bluesky_store_async_submit(async);
341 bluesky_store_add_barrier(barrier, async);
342 bluesky_store_async_unref(async);
345 /* Write back an inode and all associated data and wait for completion. Inode
346 * should already be locked. */
347 void bluesky_inode_do_sync(BlueSkyInode *inode)
349 BlueSkyStoreAsync *barrier = bluesky_store_async_new(inode->fs->store);
350 barrier->op = STORE_OP_BARRIER;
352 if (bluesky_verbose) {
353 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
354 "Synchronous writeback for inode %"PRIu64"...", inode->inum);
356 bluesky_inode_start_sync(inode, barrier);
357 bluesky_store_async_submit(barrier);
358 bluesky_store_async_wait(barrier);
359 bluesky_store_async_unref(barrier);
360 if (bluesky_verbose) {
361 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
362 "Writeback for inode %"PRIu64" complete", inode->inum);
366 static void complete_inode_fetch(BlueSkyStoreAsync *async, BlueSkyInode *inode)
368 if (bluesky_verbose) {
369 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
370 "Completing fetch of inode %"PRIu64"...", inode->inum);
373 if (async->result != 0
374 || !bluesky_deserialize_inode(inode, async->data->data))
376 if (bluesky_verbose) {
377 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
378 " failed to load inode, cleaning up");
380 g_mutex_lock(inode->fs->lock);
381 g_hash_table_remove(inode->fs->inodes, &inode->inum);
382 bluesky_list_unlink(&inode->fs->accessed_list, inode->accessed_list);
383 inode->accessed_list = NULL;
384 g_mutex_unlock(inode->fs->lock);
385 bluesky_inode_unref(inode);
388 inode->access_time = bluesky_get_current_time();
389 g_mutex_lock(inode->fs->lock);
390 bluesky_list_unlink(&inode->fs->accessed_list, inode->accessed_list);
391 inode->accessed_list = bluesky_list_prepend(&inode->fs->accessed_list, inode);
392 g_mutex_unlock(inode->fs->lock);
394 g_mutex_unlock(inode->lock);
395 bluesky_inode_unref(inode);
398 /* Fetch an inode from stable storage. The fetch can be performed
399 * asynchronously: the in-memory inode is allocated, but not filled with data
400 * immediately. It is kept locked until it has been filled in, so any users
401 * should try to acquire the lock on the inode before accessing any data. The
402 * fs lock must be held. */
403 void bluesky_inode_fetch(BlueSkyFS *fs, uint64_t inum)
406 sprintf(key, "inode-%016"PRIx64, inum);
408 BlueSkyInode *inode = bluesky_new_inode(inum, fs, BLUESKY_PENDING);
409 inode->change_count = 0;
410 bluesky_inode_ref(inode); // Extra ref held by fetching process
411 g_mutex_lock(inode->lock);
412 bluesky_insert_inode(fs, inode);
414 BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
415 async->op = STORE_OP_GET;
416 async->key = g_strdup(key);
418 bluesky_store_async_add_notifier(async, (GFunc)complete_inode_fetch, inode);
419 bluesky_store_async_submit(async);
421 if (bluesky_options.sync_inode_fetches) {
422 bluesky_store_async_wait(async);
425 bluesky_store_async_unref(async);
428 /* Synchronize filesystem superblock to stable storage. */
429 void bluesky_superblock_flush(BlueSkyFS *fs)
431 GString *buf = g_string_new("");
432 bluesky_serialize_superblock(buf, fs);
433 BlueSkyRCStr *data = bluesky_string_new_from_gstring(buf);
435 BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
436 async->op = STORE_OP_PUT;
437 async->key = g_strdup("superblock");
439 bluesky_store_async_submit(async);
440 bluesky_store_async_unref(async);
442 //bluesky_store_sync(fs->store);