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 struct _BlueSkyCloudLogState;
100 typedef struct _BlueSkyCloudLogState BlueSkyCloudLogState;
102 void bluesky_store_init();
103 BlueSkyStore *bluesky_store_new(const gchar *type);
104 void bluesky_store_free(BlueSkyStore *store);
105 BlueSkyRCStr *bluesky_store_get(BlueSkyStore *store, const gchar *key);
106 void bluesky_store_put(BlueSkyStore *store,
107 const gchar *key, BlueSkyRCStr *val);
109 /* File types. The numeric values are chosen to match with those used in
113 BLUESKY_DIRECTORY = 2,
115 BLUESKY_CHARACTER = 4,
120 /* Special types used only internally. */
121 BLUESKY_PENDING = 0, /* Inode being loaded; type not yet determined */
122 BLUESKY_INVALID = -1, /* Inode is invalid (failed to load) */
125 /* Filesystem state. Each filesystem which is exported is represented by a
126 * single bluesky_fs structure in memory. */
130 gchar *name; /* Descriptive name for the filesystem */
131 GHashTable *inodes; /* Cached inodes */
132 uint64_t next_inum; /* Next available inode for allocation */
136 BlueSkyCloudLogState *log_state;
138 /* Accounting for memory used for caches. Space is measured in blocks, not
139 * bytes. We track both total data in the caches and dirty data (total
140 * data includes dirty data). Updates to these variables must be made
142 gint cache_total, cache_dirty;
144 /* Linked list of inodes, sorted by access/modification times for cache
145 * management. Editing these lists is protected by the filesystem lock; to
146 * avoid deadlock do not attempt to take any other locks while the FS lock
147 * is held for list editing purposes. Items at the head of the list are
148 * most recently accessed/modified. */
149 GList unlogged_list; // Changes not yet synced to journal
150 GList dirty_list; // Not yet written to cloud storage
151 GList accessed_list; // All in-memory inodes
153 /* Mutex for the flush daemon, to prevent concurrent execution. */
156 /* Mapping of object identifiers (blocks, inodes) to physical location (in
157 * the local cache or in the logs in the cloud). */
158 GHashTable *locations;
161 /* Inode number of the root directory. */
162 #define BLUESKY_ROOT_INUM 1
164 /* Timestamp, measured in microseconds since the Unix epoch. */
165 typedef int64_t bluesky_time;
167 /* High-resolution timer, measured in nanoseconds. */
168 typedef int64_t bluesky_time_hires;
169 bluesky_time_hires bluesky_now_hires();
171 /* In-memory representation of an inode within a Blue Sky server. This
172 * corresponds roughly with information that is committed to persistent
173 * storage. Locking/refcounting rules:
174 * - To access or modify any data fields, the lock must be held. This
175 * includes file blocks.
176 * - One reference is held by the BlueSkyFS inode hash table. If that is the
177 * only reference (and the inode is unlocked), the inode is subject to
178 * dropping from the cache.
179 * - Any pending operations should hold extra references to the inode as
180 * appropriate to keep it available until the operation completes.
181 * - Locking dependency order is, when multiple locks are to be acquired, to
182 * acquire locks on parents in the filesystem tree before children.
183 * (TODO: What about rename when we acquire locks in unrelated parts of the
185 * - An inode should not be locked while the filesystem lock is already held,
186 * since some code may do an inode lookup (which acquires the filesystem
187 * lock) while a different inode is locked.
195 BlueSkyFileType type;
200 /* Rather than track an inode number and generation number, we will simply
201 * never re-use a fileid after a file is deleted. 64 bits should be enough
202 * that we don't exhaust the identifier space. */
205 /* change_count is increased with every operation which modifies the inode,
206 * and can be used to determine if cached data is still valid.
207 * change_commit is the value of change_count when the inode was last
208 * committed to stable storage (the log).
209 * change_cloud tracks which version was last commited to cloud storage. */
210 uint64_t change_count, change_commit, change_cloud;
212 /* Timestamp for controlling when modified data is flushed to stable
213 * storage. When an inode is first modified from a clean state, this is
214 * set to the current time. If the inode is clean, it is set to zero. */
217 /* Last access time to this inode, for controlling cache evictions. */
220 /* Additional state for tracking cache writeback status. */
221 uint64_t change_pending; /* change_count version currently being committed to storage */
223 /* Pointers to the linked-list elements for this inode in the accessed and
224 * dirty linked lists. We re-use the GList structure, using ->next to
225 * point to the head of the list and ->prev to point to the tail. The data
226 * element is unused. */
227 GList *unlogged_list, *accessed_list, *dirty_list;
229 int64_t atime; /* Microseconds since the Unix epoch */
232 int64_t ntime; /* "new time": time object was created */
234 /* File-specific fields */
238 /* Directory-specific fields */
239 GSequence *dirents; /* List of entries for READDIR */
240 GHashTable *dirhash; /* Hash table by name for LOOKUP */
241 GHashTable *dirhash_folded; /* As above, but case-folded */
242 uint64_t parent_inum; /* inode for ".."; 0 if the root directory */
244 /* Symlink-specific fields */
245 gchar *symlink_contents;
248 /* A directory entry. The name is UTF-8 and is a freshly-allocated string.
249 * Each directory entry is listed in two indices: dirents is indexed by cookie
250 * and dirhash by name. The cookie is a randomly-assigned 32-bit value, unique
251 * within the directory, that remains unchanged until the entry is deleted. It
252 * is used to provide a stable key for restarting a READDIR call. */
255 gchar *name_folded; /* Name, folded for case-insensitive lookup */
260 /* File data is divided into fixed-size blocks (except the last block which may
261 * be short?). These blocks are backed by storage in a key/value store, but
262 * may also be dirty if modifications have been made in-memory that have not
264 #define BLUESKY_BLOCK_SIZE 32768ULL
265 #define BLUESKY_MAX_FILE_SIZE (BLUESKY_BLOCK_SIZE << 24)
267 BLUESKY_BLOCK_ZERO = 0, /* Data is all zeroes, not explicitly stored */
268 BLUESKY_BLOCK_REF = 1, /* Reference to key/value store, not cached */
269 BLUESKY_BLOCK_CACHED = 2, /* Data is cached in memory, clean */
270 BLUESKY_BLOCK_DIRTY = 3, /* Data needs to be committed to store */
274 BlueSkyBlockType type;
275 gchar *ref; /* Name of data block in the backing store */
276 BlueSkyRCStr *data; /* Pointer to data in memory if cached */
279 BlueSkyFS *bluesky_init_fs(gchar *name, BlueSkyStore *store);
280 void bluesky_superblock_flush(BlueSkyFS *fs);
282 gboolean bluesky_inode_is_ready(BlueSkyInode *inode);
284 int64_t bluesky_get_current_time();
285 void bluesky_inode_update_ctime(BlueSkyInode *inode, gboolean update_mtime);
286 uint64_t bluesky_fs_alloc_inode(BlueSkyFS *fs);
287 void bluesky_init_inode(BlueSkyInode *i, BlueSkyFileType type);
288 BlueSkyInode *bluesky_new_inode(uint64_t inum, BlueSkyFS *fs, BlueSkyFileType type);
290 BlueSkyInode *bluesky_get_inode(BlueSkyFS *fs, uint64_t inum);
291 void bluesky_inode_ref(BlueSkyInode *inode);
292 void bluesky_inode_unref(BlueSkyInode *inode);
293 void bluesky_insert_inode(BlueSkyFS *fs, BlueSkyInode *inode);
295 void bluesky_dirent_destroy(gpointer dirent);
296 uint64_t bluesky_directory_lookup(BlueSkyInode *inode, gchar *name);
297 uint64_t bluesky_directory_ilookup(BlueSkyInode *inode, gchar *name);
298 BlueSkyDirent *bluesky_directory_read(BlueSkyInode *dir, uint32_t cookie);
299 gboolean bluesky_directory_insert(BlueSkyInode *dir, const gchar *name,
301 void bluesky_directory_dump(BlueSkyInode *dir);
303 void bluesky_file_truncate(BlueSkyInode *inode, uint64_t size);
304 void bluesky_file_write(BlueSkyInode *inode, uint64_t offset,
305 const char *data, gint len);
306 void bluesky_file_read(BlueSkyInode *inode, uint64_t offset,
307 char *buf, gint len);
309 void bluesky_inode_flush(BlueSkyFS *fs, BlueSkyInode *inode);
310 void bluesky_inode_fetch(BlueSkyFS *fs, uint64_t inum);
312 gint bluesky_dirent_compare(gconstpointer a, gconstpointer b,
315 void bluesky_flushd_invoke(BlueSkyFS *fs);
316 void bluesky_flushd_invoke_conditional(BlueSkyFS *fs);
317 void bluesky_inode_do_sync(BlueSkyInode *inode);
318 void bluesky_flushd_thread_launch(BlueSkyFS *fs);
320 void bluesky_debug_dump(BlueSkyFS *fs);