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
37 /* Target size of the disk cache at the proxy, in kilobytes. */
41 extern BlueSkyOptions bluesky_options;
43 /* Maximum number of threads to use in any particular thread pool, or -1 for no
45 extern int bluesky_max_threads;
47 /* A general-purpose counter for gathering run-time statistics. */
48 struct bluesky_stats {
53 struct bluesky_stats *bluesky_stats_new(const char *name);
54 void bluesky_stats_add(struct bluesky_stats *stats, int64_t value);
55 void bluesky_stats_dump_all();
57 /* BlueSky status and error codes. Various frontends should translate these to
58 * the appropriate error code for whatever protocol they implement. */
60 BSTATUS_OK = 0, /* No error */
61 BSTATUS_IOERR, /* I/O error of some form */
62 BSTATUS_NOENT, /* File does not exist */
65 void bluesky_init(void);
67 gchar *bluesky_lowercase(const gchar *s);
69 struct BlueSkyCacheFile;
70 typedef struct BlueSkyCacheFile BlueSkyCacheFile;
74 BlueSkyCacheFile *mmap;
79 BlueSkyRCStr *bluesky_string_new(gpointer data, gsize len);
80 BlueSkyRCStr *bluesky_string_new_from_gstring(GString *s);
81 BlueSkyRCStr *bluesky_string_new_from_mmap(BlueSkyCacheFile *mmap,
82 int offset, gsize len);
83 void bluesky_string_ref(BlueSkyRCStr *string);
84 void bluesky_string_unref(BlueSkyRCStr *string);
85 BlueSkyRCStr *bluesky_string_dup(BlueSkyRCStr *string);
86 void bluesky_string_resize(BlueSkyRCStr *string, gsize len);
88 /* Storage interface. This presents a key-value store abstraction, and can
89 * have multiple implementations: in-memory, on-disk, in-cloud. */
91 typedef struct BlueSkyStore BlueSkyStore;
94 typedef struct BlueSkyLog BlueSkyLog;
96 struct BlueSkyCloudLogState;
97 typedef struct BlueSkyCloudLogState BlueSkyCloudLogState;
99 struct BlueSkyCloudLog;
100 typedef struct BlueSkyCloudLog BlueSkyCloudLog;
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. */
127 struct BlueSkyCryptKeys;
132 gchar *name; /* Descriptive name for the filesystem */
133 GHashTable *inodes; /* Cached inodes */
134 uint64_t next_inum; /* Next available inode for allocation */
138 BlueSkyCloudLogState *log_state;
140 /* Filesystem crypto keys */
142 struct BlueSkyCryptKeys *keys;
144 /* Accounting for memory used for caches. Space is measured in blocks, not
145 * bytes. Updates to these variables must be made atomically. */
148 /* Like above, but tracking data stored in the cloudlog entries
150 * - cache_log_dirty: data uncommitted to journal and cloud
151 * - cache_log_writeback: data being written to journal
152 * - cache_log_journal: data committed to journal
153 * - cache_log_cloud: data written to cloud as well
154 * Log entries should progress from the top state to the bottom, and are
155 * only ever counted in one category at a time. */
156 gint cache_log_dirty, cache_log_writeback,
157 cache_log_journal, cache_log_cloud;
159 /* Linked list of inodes, sorted by access/modification times for cache
160 * management. Editing these lists is protected by the filesystem lock; to
161 * avoid deadlock do not attempt to take any other locks while the FS lock
162 * is held for list editing purposes. Items at the head of the list are
163 * most recently accessed/modified. */
164 GList unlogged_list; // Changes not yet synced to journal
165 GList dirty_list; // Not yet written to cloud storage
166 GList accessed_list; // All in-memory inodes
168 /* Mutex for the flush daemon, to prevent concurrent execution. */
171 /* Used to wait for the cache daemon to free up space */
174 /* Mapping of object identifiers (blocks, inodes) to physical location (in
175 * the local cache or in the logs in the cloud). */
176 GHashTable *locations;
178 /* The inode map, which maps inode numbers to the location of the most
180 GSequence *inode_map;
182 /* Queue for asynchronous cloudlog unrefs, where needed. */
183 GAsyncQueue *unref_queue;
185 /* Thread pool for asynchronous inode fetches */
186 GThreadPool *inode_fetch_thread_pool;
189 /* Inode number of the root directory. */
190 #define BLUESKY_ROOT_INUM 1
192 /* Timestamp, measured in microseconds since the Unix epoch. */
193 typedef int64_t bluesky_time;
195 /* High-resolution timer, measured in nanoseconds. */
196 typedef int64_t bluesky_time_hires;
197 bluesky_time_hires bluesky_now_hires();
199 /* In-memory representation of an inode within a Blue Sky server. This
200 * corresponds roughly with information that is committed to persistent
201 * storage. Locking/refcounting rules:
202 * - To access or modify any data fields, the lock must be held. This
203 * includes file blocks.
204 * - One reference is held by the BlueSkyFS inode hash table. If that is the
205 * only reference (and the inode is unlocked), the inode is subject to
206 * dropping from the cache.
207 * - Any pending operations should hold extra references to the inode as
208 * appropriate to keep it available until the operation completes.
209 * - Locking dependency order is, when multiple locks are to be acquired, to
210 * acquire locks on parents in the filesystem tree before children.
211 * (TODO: What about rename when we acquire locks in unrelated parts of the
213 * - An inode should not be locked while the filesystem lock is already held,
214 * since some code may do an inode lookup (which acquires the filesystem
215 * lock) while a different inode is locked.
223 BlueSkyFileType type;
228 /* Rather than track an inode number and generation number, we will simply
229 * never re-use a fileid after a file is deleted. 64 bits should be enough
230 * that we don't exhaust the identifier space. */
233 /* change_count is increased with every operation which modifies the inode,
234 * and can be used to determine if cached data is still valid.
235 * change_commit is the value of change_count when the inode was last
236 * committed to stable storage (the log).
237 * change_cloud tracks which version was last commited to cloud storage. */
238 uint64_t change_count, change_commit, change_cloud;
240 /* Timestamp for controlling when modified data is flushed to stable
241 * storage. When an inode is first modified from a clean state, this is
242 * set to the current time. If the inode is clean, it is set to zero. */
245 /* Last access time to this inode, for controlling cache evictions. */
248 /* Version of the object last serialized and committed to storage. */
249 BlueSkyCloudLog *committed_item;
251 /* Pointers to the linked-list elements for this inode in the accessed and
252 * dirty linked lists. We re-use the GList structure, using ->next to
253 * point to the head of the list and ->prev to point to the tail. The data
254 * element is unused. */
255 GList *unlogged_list, *accessed_list, *dirty_list;
257 int64_t atime; /* Microseconds since the Unix epoch */
260 int64_t ntime; /* "new time": time object was created */
262 /* File-specific fields */
266 /* Directory-specific fields */
267 GSequence *dirents; /* List of entries for READDIR */
268 GHashTable *dirhash; /* Hash table by name for LOOKUP */
269 GHashTable *dirhash_folded; /* As above, but case-folded */
270 uint64_t parent_inum; /* inode for ".."; 0 if the root directory */
272 /* Symlink-specific fields */
273 gchar *symlink_contents;
276 /* A directory entry. The name is UTF-8 and is a freshly-allocated string.
277 * Each directory entry is listed in two indices: dirents is indexed by cookie
278 * and dirhash by name. The cookie is a randomly-assigned 32-bit value, unique
279 * within the directory, that remains unchanged until the entry is deleted. It
280 * is used to provide a stable key for restarting a READDIR call. */
283 gchar *name_folded; /* Name, folded for case-insensitive lookup */
288 /* File data is divided into fixed-size blocks (except the last block which may
289 * be short?). These blocks are backed by storage in a key/value store, but
290 * may also be dirty if modifications have been made in-memory that have not
292 #define BLUESKY_BLOCK_SIZE 32768ULL
293 #define BLUESKY_MAX_FILE_SIZE (BLUESKY_BLOCK_SIZE << 24)
295 BLUESKY_BLOCK_ZERO = 0, /* Data is all zeroes, not explicitly stored */
296 BLUESKY_BLOCK_REF = 1, /* Reference to cloud log item, data clean */
297 BLUESKY_BLOCK_DIRTY = 2, /* Data needs to be committed to store */
301 BlueSkyBlockType type;
302 BlueSkyCloudLog *ref; /* if REF: cloud log entry with data */
303 BlueSkyRCStr *dirty; /* if DIRTY: raw data in memory */
306 BlueSkyFS *bluesky_init_fs(gchar *name, BlueSkyStore *store,
307 const gchar *master_key);
309 gboolean bluesky_inode_is_ready(BlueSkyInode *inode);
311 int64_t bluesky_get_current_time();
312 void bluesky_inode_update_ctime(BlueSkyInode *inode, gboolean update_mtime);
313 uint64_t bluesky_fs_alloc_inode(BlueSkyFS *fs);
314 void bluesky_init_inode(BlueSkyInode *i, BlueSkyFileType type);
315 BlueSkyInode *bluesky_new_inode(uint64_t inum, BlueSkyFS *fs, BlueSkyFileType type);
317 BlueSkyInode *bluesky_get_inode(BlueSkyFS *fs, uint64_t inum);
318 void bluesky_inode_ref(BlueSkyInode *inode);
319 void bluesky_inode_unref(BlueSkyInode *inode);
320 void bluesky_insert_inode(BlueSkyFS *fs, BlueSkyInode *inode);
322 void bluesky_dirent_destroy(gpointer dirent);
323 uint64_t bluesky_directory_lookup(BlueSkyInode *inode, gchar *name);
324 uint64_t bluesky_directory_ilookup(BlueSkyInode *inode, gchar *name);
325 BlueSkyDirent *bluesky_directory_read(BlueSkyInode *dir, uint32_t cookie);
326 gboolean bluesky_directory_insert(BlueSkyInode *dir, const gchar *name,
328 void bluesky_directory_dump(BlueSkyInode *dir);
330 void bluesky_file_truncate(BlueSkyInode *inode, uint64_t size);
331 void bluesky_file_write(BlueSkyInode *inode, uint64_t offset,
332 const char *data, gint len);
333 void bluesky_file_read(BlueSkyInode *inode, uint64_t offset,
334 char *buf, gint len);
336 void bluesky_inode_fetch(BlueSkyFS *fs, uint64_t inum);
338 gint bluesky_dirent_compare(gconstpointer a, gconstpointer b,
341 void bluesky_flushd_invoke(BlueSkyFS *fs);
342 void bluesky_flushd_invoke_conditional(BlueSkyFS *fs);
343 void bluesky_inode_do_sync(BlueSkyInode *inode);
344 void bluesky_flushd_thread_launch(BlueSkyFS *fs);
346 void bluesky_debug_dump(BlueSkyFS *fs);