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 struct _BlueSkyCacheFile;
67 typedef struct _BlueSkyCacheFile BlueSkyCacheFile;
71 BlueSkyCacheFile *mmap;
76 BlueSkyRCStr *bluesky_string_new(gpointer data, gsize len);
77 BlueSkyRCStr *bluesky_string_new_from_gstring(GString *s);
78 BlueSkyRCStr *bluesky_string_new_from_mmap(BlueSkyCacheFile *mmap,
79 int offset, gsize len);
80 void bluesky_string_ref(BlueSkyRCStr *string);
81 void bluesky_string_unref(BlueSkyRCStr *string);
82 BlueSkyRCStr *bluesky_string_dup(BlueSkyRCStr *string);
83 void bluesky_string_resize(BlueSkyRCStr *string, gsize len);
85 /* Cryptographic operations. */
86 #define CRYPTO_BLOCK_SIZE 16 /* 128-bit AES */
87 #define CRYPTO_KEY_SIZE 16
89 void bluesky_crypt_init();
90 void bluesky_crypt_hash_key(const char *keystr, uint8_t *out);
91 void bluesky_crypt_random_bytes(guchar *buf, gint len);
92 BlueSkyRCStr *bluesky_crypt_encrypt(BlueSkyRCStr *in, const uint8_t *key);
93 BlueSkyRCStr *bluesky_crypt_decrypt(BlueSkyRCStr *in, const uint8_t *key);
95 /* Storage interface. This presents a key-value store abstraction, and can
96 * have multiple implementations: in-memory, on-disk, in-cloud. */
98 typedef struct _BlueSkyStore BlueSkyStore;
101 typedef struct _BlueSkyLog BlueSkyLog;
103 struct _BlueSkyCloudLogState;
104 typedef struct _BlueSkyCloudLogState BlueSkyCloudLogState;
106 struct _BlueSkyCloudLog;
107 typedef struct _BlueSkyCloudLog BlueSkyCloudLog;
109 void bluesky_store_init();
110 BlueSkyStore *bluesky_store_new(const gchar *type);
111 void bluesky_store_free(BlueSkyStore *store);
112 BlueSkyRCStr *bluesky_store_get(BlueSkyStore *store, const gchar *key);
113 void bluesky_store_put(BlueSkyStore *store,
114 const gchar *key, BlueSkyRCStr *val);
116 /* File types. The numeric values are chosen to match with those used in
120 BLUESKY_DIRECTORY = 2,
122 BLUESKY_CHARACTER = 4,
127 /* Special types used only internally. */
128 BLUESKY_PENDING = 0, /* Inode being loaded; type not yet determined */
129 BLUESKY_INVALID = -1, /* Inode is invalid (failed to load) */
132 /* Filesystem state. Each filesystem which is exported is represented by a
133 * single bluesky_fs structure in memory. */
137 gchar *name; /* Descriptive name for the filesystem */
138 GHashTable *inodes; /* Cached inodes */
139 uint64_t next_inum; /* Next available inode for allocation */
143 BlueSkyCloudLogState *log_state;
145 /* Accounting for memory used for caches. Space is measured in blocks, not
146 * bytes. Updates to these variables must be made atomically. */
149 /* Like above, but tracking data stored in the cloudlog entries
151 * - cache_log_dirty: data uncommitted to journal and cloud
152 * - cache_log_writeback: data being written to journal
153 * - cache_log_journal: data committed to journal
154 * - cache_log_cloud: data written to cloud as well
155 * Log entries should progress from the top state to the bottom, and are
156 * only ever counted in one category at a time. */
157 gint cache_log_dirty, cache_log_writeback,
158 cache_log_journal, cache_log_cloud;
160 /* Linked list of inodes, sorted by access/modification times for cache
161 * management. Editing these lists is protected by the filesystem lock; to
162 * avoid deadlock do not attempt to take any other locks while the FS lock
163 * is held for list editing purposes. Items at the head of the list are
164 * most recently accessed/modified. */
165 GList unlogged_list; // Changes not yet synced to journal
166 GList dirty_list; // Not yet written to cloud storage
167 GList accessed_list; // All in-memory inodes
169 /* Mutex for the flush daemon, to prevent concurrent execution. */
172 /* Used to wait for the cache daemon to free up space */
175 /* Mapping of object identifiers (blocks, inodes) to physical location (in
176 * the local cache or in the logs in the cloud). */
177 GHashTable *locations;
180 /* Inode number of the root directory. */
181 #define BLUESKY_ROOT_INUM 1
183 /* Timestamp, measured in microseconds since the Unix epoch. */
184 typedef int64_t bluesky_time;
186 /* High-resolution timer, measured in nanoseconds. */
187 typedef int64_t bluesky_time_hires;
188 bluesky_time_hires bluesky_now_hires();
190 /* In-memory representation of an inode within a Blue Sky server. This
191 * corresponds roughly with information that is committed to persistent
192 * storage. Locking/refcounting rules:
193 * - To access or modify any data fields, the lock must be held. This
194 * includes file blocks.
195 * - One reference is held by the BlueSkyFS inode hash table. If that is the
196 * only reference (and the inode is unlocked), the inode is subject to
197 * dropping from the cache.
198 * - Any pending operations should hold extra references to the inode as
199 * appropriate to keep it available until the operation completes.
200 * - Locking dependency order is, when multiple locks are to be acquired, to
201 * acquire locks on parents in the filesystem tree before children.
202 * (TODO: What about rename when we acquire locks in unrelated parts of the
204 * - An inode should not be locked while the filesystem lock is already held,
205 * since some code may do an inode lookup (which acquires the filesystem
206 * lock) while a different inode is locked.
214 BlueSkyFileType type;
219 /* Rather than track an inode number and generation number, we will simply
220 * never re-use a fileid after a file is deleted. 64 bits should be enough
221 * that we don't exhaust the identifier space. */
224 /* change_count is increased with every operation which modifies the inode,
225 * and can be used to determine if cached data is still valid.
226 * change_commit is the value of change_count when the inode was last
227 * committed to stable storage (the log).
228 * change_cloud tracks which version was last commited to cloud storage. */
229 uint64_t change_count, change_commit, change_cloud;
231 /* Timestamp for controlling when modified data is flushed to stable
232 * storage. When an inode is first modified from a clean state, this is
233 * set to the current time. If the inode is clean, it is set to zero. */
236 /* Last access time to this inode, for controlling cache evictions. */
239 /* Version of the object last serialized and committed to storage. */
240 BlueSkyCloudLog *committed_item;
242 /* Pointers to the linked-list elements for this inode in the accessed and
243 * dirty linked lists. We re-use the GList structure, using ->next to
244 * point to the head of the list and ->prev to point to the tail. The data
245 * element is unused. */
246 GList *unlogged_list, *accessed_list, *dirty_list;
248 int64_t atime; /* Microseconds since the Unix epoch */
251 int64_t ntime; /* "new time": time object was created */
253 /* File-specific fields */
257 /* Directory-specific fields */
258 GSequence *dirents; /* List of entries for READDIR */
259 GHashTable *dirhash; /* Hash table by name for LOOKUP */
260 GHashTable *dirhash_folded; /* As above, but case-folded */
261 uint64_t parent_inum; /* inode for ".."; 0 if the root directory */
263 /* Symlink-specific fields */
264 gchar *symlink_contents;
267 /* A directory entry. The name is UTF-8 and is a freshly-allocated string.
268 * Each directory entry is listed in two indices: dirents is indexed by cookie
269 * and dirhash by name. The cookie is a randomly-assigned 32-bit value, unique
270 * within the directory, that remains unchanged until the entry is deleted. It
271 * is used to provide a stable key for restarting a READDIR call. */
274 gchar *name_folded; /* Name, folded for case-insensitive lookup */
279 /* File data is divided into fixed-size blocks (except the last block which may
280 * be short?). These blocks are backed by storage in a key/value store, but
281 * may also be dirty if modifications have been made in-memory that have not
283 #define BLUESKY_BLOCK_SIZE 32768ULL
284 #define BLUESKY_MAX_FILE_SIZE (BLUESKY_BLOCK_SIZE << 24)
286 BLUESKY_BLOCK_ZERO = 0, /* Data is all zeroes, not explicitly stored */
287 BLUESKY_BLOCK_REF = 1, /* Reference to cloud log item, data clean */
288 BLUESKY_BLOCK_DIRTY = 2, /* Data needs to be committed to store */
292 BlueSkyBlockType type;
293 BlueSkyCloudLog *ref; /* if REF: cloud log entry with data */
294 BlueSkyRCStr *dirty; /* if DIRTY: raw data in memory */
297 BlueSkyFS *bluesky_init_fs(gchar *name, BlueSkyStore *store);
298 void bluesky_superblock_flush(BlueSkyFS *fs);
300 gboolean bluesky_inode_is_ready(BlueSkyInode *inode);
302 int64_t bluesky_get_current_time();
303 void bluesky_inode_update_ctime(BlueSkyInode *inode, gboolean update_mtime);
304 uint64_t bluesky_fs_alloc_inode(BlueSkyFS *fs);
305 void bluesky_init_inode(BlueSkyInode *i, BlueSkyFileType type);
306 BlueSkyInode *bluesky_new_inode(uint64_t inum, BlueSkyFS *fs, BlueSkyFileType type);
308 BlueSkyInode *bluesky_get_inode(BlueSkyFS *fs, uint64_t inum);
309 void bluesky_inode_ref(BlueSkyInode *inode);
310 void bluesky_inode_unref(BlueSkyInode *inode);
311 void bluesky_insert_inode(BlueSkyFS *fs, BlueSkyInode *inode);
313 void bluesky_dirent_destroy(gpointer dirent);
314 uint64_t bluesky_directory_lookup(BlueSkyInode *inode, gchar *name);
315 uint64_t bluesky_directory_ilookup(BlueSkyInode *inode, gchar *name);
316 BlueSkyDirent *bluesky_directory_read(BlueSkyInode *dir, uint32_t cookie);
317 gboolean bluesky_directory_insert(BlueSkyInode *dir, const gchar *name,
319 void bluesky_directory_dump(BlueSkyInode *dir);
321 void bluesky_file_truncate(BlueSkyInode *inode, uint64_t size);
322 void bluesky_file_write(BlueSkyInode *inode, uint64_t offset,
323 const char *data, gint len);
324 void bluesky_file_read(BlueSkyInode *inode, uint64_t offset,
325 char *buf, gint len);
327 void bluesky_inode_fetch(BlueSkyFS *fs, uint64_t inum);
329 gint bluesky_dirent_compare(gconstpointer a, gconstpointer b,
332 void bluesky_flushd_invoke(BlueSkyFS *fs);
333 void bluesky_flushd_invoke_conditional(BlueSkyFS *fs);
334 void bluesky_inode_do_sync(BlueSkyInode *inode);
335 void bluesky_flushd_thread_launch(BlueSkyFS *fs);
337 void bluesky_debug_dump(BlueSkyFS *fs);