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");
86 fs->flushd_lock = g_mutex_new();
91 BlueSkyFS *bluesky_init_fs(gchar *name, BlueSkyStore *store)
93 BlueSkyRCStr *data = bluesky_store_get(store, "superblock");
95 BlueSkyFS *fs = bluesky_deserialize_superblock(data->data);
98 g_print("Loaded filesystem superblock\n");
100 fs->name = g_strdup(name);
103 bluesky_string_unref(data);
106 g_print("Initializing fresh filesystem\n");
107 BlueSkyFS *fs = bluesky_new_fs(name);
109 fs->log = bluesky_log_new("journal");
111 BlueSkyInode *root = bluesky_new_inode(BLUESKY_ROOT_INUM, fs,
115 bluesky_insert_inode(fs, root);
116 bluesky_inode_update_ctime(root, TRUE);
118 bluesky_inode_flush(fs, root);
119 bluesky_superblock_flush(fs);
124 /* Inode reference counting. */
125 void bluesky_inode_ref(BlueSkyInode *inode)
127 g_atomic_int_inc(&inode->refcount);
130 void bluesky_inode_unref(BlueSkyInode *inode)
132 if (g_atomic_int_dec_and_test(&inode->refcount)) {
133 if (bluesky_verbose) {
134 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
135 "Reference count for inode %"PRIu64" dropped to zero.",
139 /* Sanity check: Is the inode clean? */
140 if (inode->change_commit < inode->change_count
141 || inode->accessed_list != NULL
142 || inode->dirty_list != NULL) {
143 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);
146 /* These shouldn't be needed, but in case the above warning fires and
147 * we delete the inode anyway, we ought to be sure the inode is not on
149 g_mutex_lock(inode->fs->lock);
150 bluesky_list_unlink(&inode->fs->accessed_list, inode->accessed_list);
151 bluesky_list_unlink(&inode->fs->dirty_list, inode->dirty_list);
152 g_mutex_unlock(inode->fs->lock);
154 /* Free file type specific data. It should be an error for there to be
155 * dirty data to commit when the reference count has reaches zero. */
156 switch (inode->type) {
157 case BLUESKY_REGULAR:
158 for (int i = 0; i < inode->blocks->len; i++) {
159 BlueSkyBlock *b = &g_array_index(inode->blocks,
161 if (b->type == BLUESKY_BLOCK_DIRTY) {
162 g_error("Deleting an inode with dirty file data!");
165 bluesky_string_unref(b->data);
167 g_array_unref(inode->blocks);
170 case BLUESKY_DIRECTORY:
171 g_hash_table_destroy(inode->dirhash);
172 g_hash_table_destroy(inode->dirhash_folded);
173 g_sequence_free(inode->dirents);
176 case BLUESKY_SYMLINK:
177 g_free(inode->symlink_contents);
184 g_mutex_free(inode->lock);
190 /* Allocate a fresh inode number which has not been used before within a
191 * filesystem. fs must already be locked. */
192 uint64_t bluesky_fs_alloc_inode(BlueSkyFS *fs)
196 inum = fs->next_inum;
199 bluesky_superblock_flush(fs);
204 /* Perform type-specification initialization of an inode. Normally performed
205 * in bluesky_new_inode, but can be separated if an inode is created first,
206 * then deserialized. */
207 void bluesky_init_inode(BlueSkyInode *i, BlueSkyFileType type)
212 case BLUESKY_REGULAR:
213 i->blocks = g_array_new(FALSE, TRUE, sizeof(BlueSkyBlock));
215 case BLUESKY_DIRECTORY:
216 i->dirents = g_sequence_new(bluesky_dirent_destroy);
217 i->dirhash = g_hash_table_new(g_str_hash, g_str_equal);
218 i->dirhash_folded = g_hash_table_new(g_str_hash, g_str_equal);
225 BlueSkyInode *bluesky_new_inode(uint64_t inum, BlueSkyFS *fs,
226 BlueSkyFileType type)
228 BlueSkyInode *i = g_new0(BlueSkyInode, 1);
230 i->lock = g_mutex_new();
235 bluesky_init_inode(i, type);
240 /* Retrieve an inode from the filesystem. Eventually this will be a cache and
241 * so we might need to go fetch the inode from elsewhere; for now all
242 * filesystem state is stored here. inode is returned with a reference held
244 BlueSkyInode *bluesky_get_inode(BlueSkyFS *fs, uint64_t inum)
246 BlueSkyInode *inode = NULL;
252 g_mutex_lock(fs->lock);
253 inode = (BlueSkyInode *)g_hash_table_lookup(fs->inodes, &inum);
256 bluesky_inode_fetch(fs, inum);
257 inode = (BlueSkyInode *)g_hash_table_lookup(fs->inodes, &inum);
261 bluesky_inode_ref(inode);
263 /* FIXME: We assume we can atomically update the in-memory access time
265 inode->access_time = bluesky_get_current_time();
268 g_mutex_unlock(fs->lock);
273 /* Insert an inode into the filesystem inode cache. fs should be locked. */
274 void bluesky_insert_inode(BlueSkyFS *fs, BlueSkyInode *inode)
276 g_hash_table_insert(fs->inodes, &inode->inum, inode);
279 /* Deprecated: Synchronize an inode to stable storage. */
280 void bluesky_inode_flush(BlueSkyFS *fs, BlueSkyInode *inode)
282 GString *buf = g_string_new("");
283 bluesky_serialize_inode(buf, inode);
284 BlueSkyRCStr *data = bluesky_string_new_from_gstring(buf);
287 sprintf(key, "inode-%016"PRIx64, inode->inum);
289 BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
290 async->op = STORE_OP_PUT;
291 async->key = g_strdup(key);
293 bluesky_store_async_submit(async);
294 bluesky_store_async_unref(async);
297 /* Start writeback of an inode and all associated data. */
298 void bluesky_inode_start_sync(BlueSkyInode *inode, BlueSkyStoreAsync *barrier)
300 BlueSkyFS *fs = inode->fs;
302 if (inode->type == BLUESKY_REGULAR)
303 bluesky_file_flush(inode, barrier);
305 GString *buf = g_string_new("");
306 bluesky_serialize_inode(buf, inode);
307 BlueSkyRCStr *data = bluesky_string_new_from_gstring(buf);
310 sprintf(key, "inode-%016"PRIx64, inode->inum);
312 BlueSkyLogItem *log_item = bluesky_log_item_new();
313 log_item->key = g_strdup(key);
314 log_item->data = data;
315 bluesky_string_ref(data);
316 bluesky_log_item_submit(log_item, fs->log);
317 bluesky_log_item_finish(log_item);
319 BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
320 async->op = STORE_OP_PUT;
321 async->key = g_strdup(key);
323 bluesky_store_async_submit(async);
325 bluesky_store_add_barrier(barrier, async);
326 bluesky_store_async_unref(async);
329 /* Write back an inode and all associated data and wait for completion. Inode
330 * should already be locked. */
331 void bluesky_inode_do_sync(BlueSkyInode *inode)
333 BlueSkyStoreAsync *barrier = bluesky_store_async_new(inode->fs->store);
334 barrier->op = STORE_OP_BARRIER;
336 if (bluesky_verbose) {
337 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
338 "Synchronous writeback for inode %"PRIu64"...", inode->inum);
340 bluesky_inode_start_sync(inode, barrier);
341 bluesky_store_async_submit(barrier);
342 bluesky_store_async_wait(barrier);
343 bluesky_store_async_unref(barrier);
344 if (bluesky_verbose) {
345 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
346 "Writeback for inode %"PRIu64" complete", inode->inum);
350 static void complete_inode_fetch(BlueSkyStoreAsync *async, BlueSkyInode *inode)
352 if (bluesky_verbose) {
353 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
354 "Completing fetch of inode %"PRIu64"...", inode->inum);
357 if (async->result != 0
358 || !bluesky_deserialize_inode(inode, async->data->data))
360 if (bluesky_verbose) {
361 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
362 " failed to load inode, cleaning up");
364 g_mutex_lock(inode->fs->lock);
365 g_hash_table_remove(inode->fs->inodes, &inode->inum);
366 bluesky_list_unlink(&inode->fs->accessed_list, inode->accessed_list);
367 inode->accessed_list = NULL;
368 g_mutex_unlock(inode->fs->lock);
369 bluesky_inode_unref(inode);
372 inode->access_time = bluesky_get_current_time();
373 g_mutex_lock(inode->fs->lock);
374 bluesky_list_unlink(&inode->fs->accessed_list, inode->accessed_list);
375 inode->accessed_list = bluesky_list_prepend(&inode->fs->accessed_list, inode);
376 g_mutex_unlock(inode->fs->lock);
378 g_mutex_unlock(inode->lock);
379 bluesky_inode_unref(inode);
382 /* Fetch an inode from stable storage. The fetch can be performed
383 * asynchronously: the in-memory inode is allocated, but not filled with data
384 * immediately. It is kept locked until it has been filled in, so any users
385 * should try to acquire the lock on the inode before accessing any data. The
386 * fs lock must be held. */
387 void bluesky_inode_fetch(BlueSkyFS *fs, uint64_t inum)
390 sprintf(key, "inode-%016"PRIx64, inum);
392 BlueSkyInode *inode = bluesky_new_inode(inum, fs, BLUESKY_PENDING);
393 inode->change_count = 0;
394 bluesky_inode_ref(inode); // Extra ref held by fetching process
395 g_mutex_lock(inode->lock);
396 bluesky_insert_inode(fs, inode);
398 BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
399 async->op = STORE_OP_GET;
400 async->key = g_strdup(key);
402 bluesky_store_async_add_notifier(async, (GFunc)complete_inode_fetch, inode);
403 bluesky_store_async_submit(async);
405 if (bluesky_options.sync_inode_fetches) {
406 bluesky_store_async_wait(async);
409 bluesky_store_async_unref(async);
412 /* Synchronize filesystem superblock to stable storage. */
413 void bluesky_superblock_flush(BlueSkyFS *fs)
415 GString *buf = g_string_new("");
416 bluesky_serialize_superblock(buf, fs);
417 BlueSkyRCStr *data = bluesky_string_new_from_gstring(buf);
419 BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
420 async->op = STORE_OP_PUT;
421 async->key = g_strdup("superblock");
423 bluesky_store_async_submit(async);
424 bluesky_store_async_unref(async);
426 //bluesky_store_sync(fs->store);