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>
20 /* Various options to tweak for performance benchmarking purposes. */
22 /* Perform all get/put operations synchronously. */
23 int synchronous_stores;
25 /* Write data in cache immediately after file is modified. */
26 int writethrough_cache;
28 /* Can inodes be fetched asynchronously? (Inode object is initially
29 * created in a pending state, and not unlocked until the data is actually
31 int sync_inode_fetches;
33 /* Should frontends handle requests serially or allow operations to proceed
38 extern BlueSkyOptions bluesky_options;
40 /* Maximum number of threads to use in any particular thread pool, or -1 for no
42 extern int bluesky_max_threads;
44 /* A general-purpose counter for gathering run-time statistics. */
45 struct bluesky_stats {
50 struct bluesky_stats *bluesky_stats_new(const char *name);
51 void bluesky_stats_add(struct bluesky_stats *stats, int64_t value);
52 void bluesky_stats_dump_all();
54 /* BlueSky status and error codes. Various frontends should translate these to
55 * the appropriate error code for whatever protocol they implement. */
57 BSTATUS_OK = 0, /* No error */
58 BSTATUS_IOERR, /* I/O error of some form */
59 BSTATUS_NOENT, /* File does not exist */
62 void bluesky_init(void);
64 gchar *bluesky_lowercase(const gchar *s);
66 /* Reference-counted blocks of memory, used for passing data in and out of
67 * storage backends and in other places. */
74 BlueSkyRCStr *bluesky_string_new(gpointer data, gsize len);
75 BlueSkyRCStr *bluesky_string_new_from_gstring(GString *s);
76 void bluesky_string_ref(BlueSkyRCStr *string);
77 void bluesky_string_unref(BlueSkyRCStr *string);
78 BlueSkyRCStr *bluesky_string_dup(BlueSkyRCStr *string);
79 void bluesky_string_resize(BlueSkyRCStr *string, gsize len);
81 /* Cryptographic operations. */
82 #define CRYPTO_BLOCK_SIZE 16 /* 128-bit AES */
83 #define CRYPTO_KEY_SIZE 16
85 void bluesky_crypt_init();
86 void bluesky_crypt_hash_key(const char *keystr, uint8_t *out);
87 void bluesky_crypt_random_bytes(guchar *buf, gint len);
88 BlueSkyRCStr *bluesky_crypt_encrypt(BlueSkyRCStr *in, const uint8_t *key);
89 BlueSkyRCStr *bluesky_crypt_decrypt(BlueSkyRCStr *in, const uint8_t *key);
91 /* Storage interface. This presents a key-value store abstraction, and can
92 * have multiple implementations: in-memory, on-disk, in-cloud. */
94 typedef struct _BlueSkyStore BlueSkyStore;
97 typedef struct _BlueSkyLog BlueSkyLog;
99 void bluesky_store_init();
100 BlueSkyStore *bluesky_store_new(const gchar *type);
101 void bluesky_store_free(BlueSkyStore *store);
102 BlueSkyRCStr *bluesky_store_get(BlueSkyStore *store, const gchar *key);
103 void bluesky_store_put(BlueSkyStore *store,
104 const gchar *key, BlueSkyRCStr *val);
106 /* File types. The numeric values are chosen to match with those used in
110 BLUESKY_DIRECTORY = 2,
112 BLUESKY_CHARACTER = 4,
117 /* Special types used only internally. */
118 BLUESKY_PENDING = 0, /* Inode being loaded; type not yet determined */
119 BLUESKY_INVALID = -1, /* Inode is invalid (failed to load) */
122 /* Filesystem state. Each filesystem which is exported is represented by a
123 * single bluesky_fs structure in memory. */
127 gchar *name; /* Descriptive name for the filesystem */
128 GHashTable *inodes; /* Cached inodes */
129 uint64_t next_inum; /* Next available inode for allocation */
134 /* Accounting for memory used for caches. Space is measured in blocks, not
135 * bytes. We track both total data in the caches and dirty data (total
136 * data includes dirty data). Updates to these variables must be made
138 gint cache_total, cache_dirty;
140 /* Linked list of inodes, sorted by access/modification times for cache
141 * management. Editing these lists is protected by the filesystem lock; to
142 * avoid deadlock do not attempt to take any other locks while the FS lock
143 * is held for list editing purposes. Items at the head of the list are
144 * most recently accessed/modified. */
145 GList dirty_list, accessed_list;
147 /* Mutex for the flush daemon, to prevent concurrent execution. */
150 /* Mapping of object identifiers (blocks, inodes) to physical location (in
151 * the local cache or in the logs in the cloud). */
152 GHashTable *locations;
155 /* Inode number of the root directory. */
156 #define BLUESKY_ROOT_INUM 1
158 /* Timestamp, measured in microseconds since the Unix epoch. */
159 typedef int64_t bluesky_time;
161 /* High-resolution timer, measured in nanoseconds. */
162 typedef int64_t bluesky_time_hires;
163 bluesky_time_hires bluesky_now_hires();
165 /* In-memory representation of an inode within a Blue Sky server. This
166 * corresponds roughly with information that is committed to persistent
167 * storage. Locking/refcounting rules:
168 * - To access or modify any data fields, the lock must be held. This
169 * includes file blocks.
170 * - One reference is held by the BlueSkyFS inode hash table. If that is the
171 * only reference (and the inode is unlocked), the inode is subject to
172 * dropping from the cache.
173 * - Any pending operations should hold extra references to the inode as
174 * appropriate to keep it available until the operation completes.
175 * - Locking dependency order is, when multiple locks are to be acquired, to
176 * acquire locks on parents in the filesystem tree before children.
177 * (TODO: What about rename when we acquire locks in unrelated parts of the
179 * - An inode should not be locked while the filesystem lock is already held,
180 * since some code may do an inode lookup (which acquires the filesystem
181 * lock) while a different inode is locked.
189 BlueSkyFileType type;
194 /* Rather than track an inode number and generation number, we will simply
195 * never re-use a fileid after a file is deleted. 64 bits should be enough
196 * that we don't exhaust the identifier space. */
199 /* change_count is increased with every operation which modifies the inode,
200 * and can be used to determine if cached data is still valid.
201 * change_commit is the value of change_count when the inode was last
202 * committed to stable storage (the log).
203 * change_cloud tracks which version was last commited to cloud storage. */
204 uint64_t change_count, change_commit, change_cloud;
206 /* Timestamp for controlling when modified data is flushed to stable
207 * storage. When an inode is first modified from a clean state, this is
208 * set to the current time. If the inode is clean, it is set to zero. */
211 /* Last access time to this inode, for controlling cache evictions. */
214 /* Additional state for tracking cache writeback status. */
215 uint64_t change_pending; /* change_count version currently being committed to storage */
217 /* Pointers to the linked-list elements for this inode in the accessed and
218 * dirty linked lists. We re-use the GList structure, using ->next to
219 * point to the head of the list and ->prev to point to the tail. The data
220 * element is unused. */
221 GList *accessed_list, *dirty_list;
223 int64_t atime; /* Microseconds since the Unix epoch */
226 int64_t ntime; /* "new time": time object was created */
228 /* File-specific fields */
232 /* Directory-specific fields */
233 GSequence *dirents; /* List of entries for READDIR */
234 GHashTable *dirhash; /* Hash table by name for LOOKUP */
235 GHashTable *dirhash_folded; /* As above, but case-folded */
236 uint64_t parent_inum; /* inode for ".."; 0 if the root directory */
238 /* Symlink-specific fields */
239 gchar *symlink_contents;
242 /* A directory entry. The name is UTF-8 and is a freshly-allocated string.
243 * Each directory entry is listed in two indices: dirents is indexed by cookie
244 * and dirhash by name. The cookie is a randomly-assigned 32-bit value, unique
245 * within the directory, that remains unchanged until the entry is deleted. It
246 * is used to provide a stable key for restarting a READDIR call. */
249 gchar *name_folded; /* Name, folded for case-insensitive lookup */
254 /* File data is divided into fixed-size blocks (except the last block which may
255 * be short?). These blocks are backed by storage in a key/value store, but
256 * may also be dirty if modifications have been made in-memory that have not
258 #define BLUESKY_BLOCK_SIZE 32768ULL
259 #define BLUESKY_MAX_FILE_SIZE (BLUESKY_BLOCK_SIZE << 24)
261 BLUESKY_BLOCK_ZERO = 0, /* Data is all zeroes, not explicitly stored */
262 BLUESKY_BLOCK_REF = 1, /* Reference to key/value store, not cached */
263 BLUESKY_BLOCK_CACHED = 2, /* Data is cached in memory, clean */
264 BLUESKY_BLOCK_DIRTY = 3, /* Data needs to be committed to store */
268 BlueSkyBlockType type;
269 gchar *ref; /* Name of data block in the backing store */
270 BlueSkyRCStr *data; /* Pointer to data in memory if cached */
273 BlueSkyFS *bluesky_init_fs(gchar *name, BlueSkyStore *store);
274 void bluesky_superblock_flush(BlueSkyFS *fs);
276 gboolean bluesky_inode_is_ready(BlueSkyInode *inode);
278 int64_t bluesky_get_current_time();
279 void bluesky_inode_update_ctime(BlueSkyInode *inode, gboolean update_mtime);
280 uint64_t bluesky_fs_alloc_inode(BlueSkyFS *fs);
281 void bluesky_init_inode(BlueSkyInode *i, BlueSkyFileType type);
282 BlueSkyInode *bluesky_new_inode(uint64_t inum, BlueSkyFS *fs, BlueSkyFileType type);
284 BlueSkyInode *bluesky_get_inode(BlueSkyFS *fs, uint64_t inum);
285 void bluesky_inode_ref(BlueSkyInode *inode);
286 void bluesky_inode_unref(BlueSkyInode *inode);
287 void bluesky_insert_inode(BlueSkyFS *fs, BlueSkyInode *inode);
289 void bluesky_dirent_destroy(gpointer dirent);
290 uint64_t bluesky_directory_lookup(BlueSkyInode *inode, gchar *name);
291 uint64_t bluesky_directory_ilookup(BlueSkyInode *inode, gchar *name);
292 BlueSkyDirent *bluesky_directory_read(BlueSkyInode *dir, uint32_t cookie);
293 gboolean bluesky_directory_insert(BlueSkyInode *dir, const gchar *name,
295 void bluesky_directory_dump(BlueSkyInode *dir);
297 void bluesky_file_truncate(BlueSkyInode *inode, uint64_t size);
298 void bluesky_file_write(BlueSkyInode *inode, uint64_t offset,
299 const char *data, gint len);
300 void bluesky_file_read(BlueSkyInode *inode, uint64_t offset,
301 char *buf, gint len);
303 void bluesky_inode_flush(BlueSkyFS *fs, BlueSkyInode *inode);
304 void bluesky_inode_fetch(BlueSkyFS *fs, uint64_t inum);
306 gint bluesky_dirent_compare(gconstpointer a, gconstpointer b,
309 void bluesky_flushd_invoke(BlueSkyFS *fs);
310 void bluesky_flushd_invoke_conditional(BlueSkyFS *fs);
311 void bluesky_inode_do_sync(BlueSkyInode *inode);
312 void bluesky_flushd_thread_launch(BlueSkyFS *fs);
314 void bluesky_debug_dump(BlueSkyFS *fs);