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 struct BlueSkyRangeset;
89 typedef struct BlueSkyRangeset BlueSkyRangeset;
90 BlueSkyRangeset *bluesky_rangeset_new();
91 void bluesky_rangeset_free(BlueSkyRangeset *rangeset);
92 gboolean bluesky_rangeset_insert(BlueSkyRangeset *rangeset,
93 int start, int length, gpointer data);
94 gpointer bluesky_rangeset_lookup(BlueSkyRangeset *rangeset, int start);
96 /* Storage interface. This presents a key-value store abstraction, and can
97 * have multiple implementations: in-memory, on-disk, in-cloud. */
99 typedef struct BlueSkyStore BlueSkyStore;
102 typedef struct BlueSkyLog BlueSkyLog;
104 struct BlueSkyCloudLogState;
105 typedef struct BlueSkyCloudLogState BlueSkyCloudLogState;
107 struct BlueSkyCloudLog;
108 typedef struct BlueSkyCloudLog BlueSkyCloudLog;
110 void bluesky_store_init();
111 BlueSkyStore *bluesky_store_new(const gchar *type);
112 void bluesky_store_free(BlueSkyStore *store);
113 BlueSkyRCStr *bluesky_store_get(BlueSkyStore *store, const gchar *key);
114 void bluesky_store_put(BlueSkyStore *store,
115 const gchar *key, BlueSkyRCStr *val);
117 /* File types. The numeric values are chosen to match with those used in
121 BLUESKY_DIRECTORY = 2,
123 BLUESKY_CHARACTER = 4,
128 /* Special types used only internally. */
129 BLUESKY_PENDING = 0, /* Inode being loaded; type not yet determined */
130 BLUESKY_INVALID = -1, /* Inode is invalid (failed to load) */
133 /* Filesystem state. Each filesystem which is exported is represented by a
134 * single bluesky_fs structure in memory. */
135 struct BlueSkyCryptKeys;
140 gchar *name; /* Descriptive name for the filesystem */
141 GHashTable *inodes; /* Cached inodes */
142 uint64_t next_inum; /* Next available inode for allocation */
146 BlueSkyCloudLogState *log_state;
148 /* Filesystem crypto keys */
150 struct BlueSkyCryptKeys *keys;
152 /* Accounting for memory used for caches. Space is measured in blocks, not
153 * bytes. Updates to these variables must be made atomically. */
156 /* Like above, but tracking data stored in the cloudlog entries
158 * - cache_log_dirty: data uncommitted to journal and cloud
159 * - cache_log_writeback: data being written to journal
160 * - cache_log_journal: data committed to journal
161 * - cache_log_cloud: data written to cloud as well
162 * Log entries should progress from the top state to the bottom, and are
163 * only ever counted in one category at a time. */
164 gint cache_log_dirty, cache_log_writeback,
165 cache_log_journal, cache_log_cloud;
167 /* Linked list of inodes, sorted by access/modification times for cache
168 * management. Editing these lists is protected by the filesystem lock; to
169 * avoid deadlock do not attempt to take any other locks while the FS lock
170 * is held for list editing purposes. Items at the head of the list are
171 * most recently accessed/modified. */
172 GList unlogged_list; // Changes not yet synced to journal
173 GList dirty_list; // Not yet written to cloud storage
174 GList accessed_list; // All in-memory inodes
176 /* Mutex for the flush daemon, to prevent concurrent execution. */
179 /* Used to wait for the cache daemon to free up space */
182 /* Mapping of object identifiers (blocks, inodes) to physical location (in
183 * the local cache or in the logs in the cloud). */
184 GHashTable *locations;
186 /* The inode map, which maps inode numbers to the location of the most
188 GSequence *inode_map;
190 /* Queue for asynchronous cloudlog unrefs, where needed. */
191 GAsyncQueue *unref_queue;
193 /* Thread pool for asynchronous inode fetches */
194 GThreadPool *inode_fetch_thread_pool;
197 /* Inode number of the root directory. */
198 #define BLUESKY_ROOT_INUM 1
200 /* Timestamp, measured in microseconds since the Unix epoch. */
201 typedef int64_t bluesky_time;
203 /* High-resolution timer, measured in nanoseconds. */
204 typedef int64_t bluesky_time_hires;
205 bluesky_time_hires bluesky_now_hires();
207 /* In-memory representation of an inode within a Blue Sky server. This
208 * corresponds roughly with information that is committed to persistent
209 * storage. Locking/refcounting rules:
210 * - To access or modify any data fields, the lock must be held. This
211 * includes file blocks.
212 * - One reference is held by the BlueSkyFS inode hash table. If that is the
213 * only reference (and the inode is unlocked), the inode is subject to
214 * dropping from the cache.
215 * - Any pending operations should hold extra references to the inode as
216 * appropriate to keep it available until the operation completes.
217 * - Locking dependency order is, when multiple locks are to be acquired, to
218 * acquire locks on parents in the filesystem tree before children.
219 * (TODO: What about rename when we acquire locks in unrelated parts of the
221 * - An inode should not be locked while the filesystem lock is already held,
222 * since some code may do an inode lookup (which acquires the filesystem
223 * lock) while a different inode is locked.
231 BlueSkyFileType type;
236 /* Rather than track an inode number and generation number, we will simply
237 * never re-use a fileid after a file is deleted. 64 bits should be enough
238 * that we don't exhaust the identifier space. */
241 /* change_count is increased with every operation which modifies the inode,
242 * and can be used to determine if cached data is still valid.
243 * change_commit is the value of change_count when the inode was last
244 * committed to stable storage (the log).
245 * change_cloud tracks which version was last commited to cloud storage. */
246 uint64_t change_count, change_commit, change_cloud;
248 /* Timestamp for controlling when modified data is flushed to stable
249 * storage. When an inode is first modified from a clean state, this is
250 * set to the current time. If the inode is clean, it is set to zero. */
253 /* Last access time to this inode, for controlling cache evictions. */
256 /* Version of the object last serialized and committed to storage. */
257 BlueSkyCloudLog *committed_item;
259 /* Pointers to the linked-list elements for this inode in the accessed and
260 * dirty linked lists. We re-use the GList structure, using ->next to
261 * point to the head of the list and ->prev to point to the tail. The data
262 * element is unused. */
263 GList *unlogged_list, *accessed_list, *dirty_list;
265 int64_t atime; /* Microseconds since the Unix epoch */
268 int64_t ntime; /* "new time": time object was created */
270 /* File-specific fields */
274 /* Directory-specific fields */
275 GSequence *dirents; /* List of entries for READDIR */
276 GHashTable *dirhash; /* Hash table by name for LOOKUP */
277 GHashTable *dirhash_folded; /* As above, but case-folded */
278 uint64_t parent_inum; /* inode for ".."; 0 if the root directory */
280 /* Symlink-specific fields */
281 gchar *symlink_contents;
284 /* A directory entry. The name is UTF-8 and is a freshly-allocated string.
285 * Each directory entry is listed in two indices: dirents is indexed by cookie
286 * and dirhash by name. The cookie is a randomly-assigned 32-bit value, unique
287 * within the directory, that remains unchanged until the entry is deleted. It
288 * is used to provide a stable key for restarting a READDIR call. */
291 gchar *name_folded; /* Name, folded for case-insensitive lookup */
296 /* File data is divided into fixed-size blocks (except the last block which may
297 * be short?). These blocks are backed by storage in a key/value store, but
298 * may also be dirty if modifications have been made in-memory that have not
300 #define BLUESKY_BLOCK_SIZE 32768ULL
301 #define BLUESKY_MAX_FILE_SIZE (BLUESKY_BLOCK_SIZE << 24)
303 BLUESKY_BLOCK_ZERO = 0, /* Data is all zeroes, not explicitly stored */
304 BLUESKY_BLOCK_REF = 1, /* Reference to cloud log item, data clean */
305 BLUESKY_BLOCK_DIRTY = 2, /* Data needs to be committed to store */
309 BlueSkyBlockType type;
310 BlueSkyCloudLog *ref; /* if REF: cloud log entry with data */
311 BlueSkyRCStr *dirty; /* if DIRTY: raw data in memory */
314 BlueSkyFS *bluesky_init_fs(gchar *name, BlueSkyStore *store,
315 const gchar *master_key);
317 gboolean bluesky_inode_is_ready(BlueSkyInode *inode);
319 int64_t bluesky_get_current_time();
320 void bluesky_inode_update_ctime(BlueSkyInode *inode, gboolean update_mtime);
321 uint64_t bluesky_fs_alloc_inode(BlueSkyFS *fs);
322 void bluesky_init_inode(BlueSkyInode *i, BlueSkyFileType type);
323 BlueSkyInode *bluesky_new_inode(uint64_t inum, BlueSkyFS *fs, BlueSkyFileType type);
325 BlueSkyInode *bluesky_get_inode(BlueSkyFS *fs, uint64_t inum);
326 void bluesky_inode_ref(BlueSkyInode *inode);
327 void bluesky_inode_unref(BlueSkyInode *inode);
328 void bluesky_insert_inode(BlueSkyFS *fs, BlueSkyInode *inode);
330 void bluesky_dirent_destroy(gpointer dirent);
331 uint64_t bluesky_directory_lookup(BlueSkyInode *inode, gchar *name);
332 uint64_t bluesky_directory_ilookup(BlueSkyInode *inode, gchar *name);
333 BlueSkyDirent *bluesky_directory_read(BlueSkyInode *dir, uint32_t cookie);
334 gboolean bluesky_directory_insert(BlueSkyInode *dir, const gchar *name,
336 void bluesky_directory_dump(BlueSkyInode *dir);
338 void bluesky_file_truncate(BlueSkyInode *inode, uint64_t size);
339 void bluesky_file_write(BlueSkyInode *inode, uint64_t offset,
340 const char *data, gint len);
341 void bluesky_file_read(BlueSkyInode *inode, uint64_t offset,
342 char *buf, gint len);
344 void bluesky_inode_fetch(BlueSkyFS *fs, uint64_t inum);
346 gint bluesky_dirent_compare(gconstpointer a, gconstpointer b,
349 void bluesky_flushd_invoke(BlueSkyFS *fs);
350 void bluesky_flushd_invoke_conditional(BlueSkyFS *fs);
351 void bluesky_inode_do_sync(BlueSkyInode *inode);
352 void bluesky_flushd_thread_launch(BlueSkyFS *fs);
354 void bluesky_debug_dump(BlueSkyFS *fs);