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>
21 /* Various options to tweak for performance benchmarking purposes. */
23 /* OBSOLETE: Perform all get/put operations synchronously. */
24 int synchronous_stores;
26 /* OBSOLETE: Write data in cache immediately after file is modified. */
27 int writethrough_cache;
29 /* Can inodes be fetched asynchronously? (Inode object is initially
30 * created in a pending state, and not unlocked until the data is actually
32 int sync_inode_fetches;
34 /* Should frontends handle requests serially or allow operations to proceed
38 /* Target size of the disk cache at the proxy, in kilobytes. */
41 /* Full segment fetches (1) or use range requests (0) for fetching segments
43 int full_segment_fetches;
45 /* Disable aggregating of data into log segments. Each object will be
46 * stored in a separate segment. */
47 int disable_aggregation;
49 /* Disable cryptography. This is for benchmarking purposes. */
53 extern BlueSkyOptions bluesky_options;
55 /* Maximum number of threads to use in any particular thread pool, or -1 for no
57 extern int bluesky_max_threads;
59 /* A general-purpose counter for gathering run-time statistics. */
60 struct bluesky_stats {
65 struct bluesky_stats *bluesky_stats_new(const char *name);
66 void bluesky_stats_add(struct bluesky_stats *stats, int64_t value);
67 void bluesky_stats_dump_all();
68 void bluesky_stats_run_periodic_dump(FILE *f);
70 /* BlueSky status and error codes. Various frontends should translate these to
71 * the appropriate error code for whatever protocol they implement. */
73 BSTATUS_OK = 0, /* No error */
74 BSTATUS_IOERR, /* I/O error of some form */
75 BSTATUS_NOENT, /* File does not exist */
78 void bluesky_init(void);
80 gchar *bluesky_lowercase(const gchar *s);
82 struct BlueSkyCacheFile;
83 typedef struct BlueSkyCacheFile BlueSkyCacheFile;
87 BlueSkyCacheFile *mmap;
92 BlueSkyRCStr *bluesky_string_new(gpointer data, gsize len);
93 BlueSkyRCStr *bluesky_string_new_from_gstring(GString *s);
94 BlueSkyRCStr *bluesky_string_new_from_mmap(BlueSkyCacheFile *mmap,
95 int offset, gsize len);
96 void bluesky_string_ref(BlueSkyRCStr *string);
97 void bluesky_string_unref(BlueSkyRCStr *string);
98 BlueSkyRCStr *bluesky_string_dup(BlueSkyRCStr *string);
99 void bluesky_string_resize(BlueSkyRCStr *string, gsize len);
101 struct BlueSkyRangeset;
102 typedef struct BlueSkyRangeset BlueSkyRangeset;
104 uint64_t start, length;
106 } BlueSkyRangesetItem;
108 BlueSkyRangeset *bluesky_rangeset_new();
109 void bluesky_rangeset_free(BlueSkyRangeset *rangeset);
110 gboolean bluesky_rangeset_insert(BlueSkyRangeset *rangeset,
111 uint64_t start, uint64_t length,
113 const BlueSkyRangesetItem *bluesky_rangeset_lookup(BlueSkyRangeset *rangeset,
115 const BlueSkyRangesetItem *bluesky_rangeset_lookup_next(BlueSkyRangeset *rangeset, uint64_t offset);
116 void bluesky_rangeset_get_extents(BlueSkyRangeset *rangeset,
117 uint64_t *start, uint64_t *length);
119 /* Storage interface. This presents a key-value store abstraction, and can
120 * have multiple implementations: in-memory, on-disk, in-cloud. */
122 typedef struct BlueSkyStore BlueSkyStore;
125 typedef struct BlueSkyLog BlueSkyLog;
127 struct BlueSkyCloudLogState;
128 typedef struct BlueSkyCloudLogState BlueSkyCloudLogState;
130 struct BlueSkyCloudLog;
131 typedef struct BlueSkyCloudLog BlueSkyCloudLog;
133 void bluesky_store_init();
134 BlueSkyStore *bluesky_store_new(const gchar *type);
135 void bluesky_store_free(BlueSkyStore *store);
136 BlueSkyRCStr *bluesky_store_get(BlueSkyStore *store, const gchar *key);
137 void bluesky_store_put(BlueSkyStore *store,
138 const gchar *key, BlueSkyRCStr *val);
140 /* File types. The numeric values are chosen to match with those used in
144 BLUESKY_DIRECTORY = 2,
146 BLUESKY_CHARACTER = 4,
151 /* Special types used only internally. */
152 BLUESKY_PENDING = 0, /* Inode being loaded; type not yet determined */
153 BLUESKY_INVALID = -1, /* Inode is invalid (failed to load) */
156 /* Filesystem state. Each filesystem which is exported is represented by a
157 * single bluesky_fs structure in memory. */
158 struct BlueSkyCryptKeys;
163 gchar *name; /* Descriptive name for the filesystem */
164 GHashTable *inodes; /* Cached inodes */
165 uint64_t next_inum; /* Next available inode for allocation */
169 BlueSkyCloudLogState *log_state;
171 /* Filesystem crypto keys */
173 struct BlueSkyCryptKeys *keys;
175 /* Accounting for memory used for caches. Space is measured in blocks, not
176 * bytes. Updates to these variables must be made atomically. */
179 /* Like above, but tracking data stored in the cloudlog entries
181 * - cache_log_dirty: data uncommitted to journal and cloud
182 * - cache_log_writeback: data being written to journal
183 * - cache_log_journal: data committed to journal
184 * - cache_log_cloud: data written to cloud as well
185 * Log entries should progress from the top state to the bottom, and are
186 * only ever counted in one category at a time. */
187 gint cache_log_dirty, cache_log_writeback,
188 cache_log_journal, cache_log_cloud;
190 /* Linked list of inodes, sorted by access/modification times for cache
191 * management. Editing these lists is protected by the filesystem lock; to
192 * avoid deadlock do not attempt to take any other locks while the FS lock
193 * is held for list editing purposes. Items at the head of the list are
194 * most recently accessed/modified. */
195 GList unlogged_list; // Changes not yet synced to journal
196 GList dirty_list; // Not yet written to cloud storage
197 GList accessed_list; // All in-memory inodes
199 /* Mutex for the flush daemon, to prevent concurrent execution. */
202 /* Used to wait for the cache daemon to free up space */
205 /* Mapping of object identifiers (blocks, inodes) to physical location (in
206 * the local cache or in the logs in the cloud). */
207 GHashTable *locations;
209 /* The inode map, which maps inode numbers to the location of the most
211 GSequence *inode_map;
213 /* Queue for asynchronous cloudlog unrefs, where needed. */
214 GAsyncQueue *unref_queue;
216 /* Thread pool for asynchronous inode fetches */
217 GThreadPool *inode_fetch_thread_pool;
220 /* Inode number of the root directory. */
221 #define BLUESKY_ROOT_INUM 1
223 /* Timestamp, measured in microseconds since the Unix epoch. */
224 typedef int64_t bluesky_time;
226 /* High-resolution timer, measured in nanoseconds. */
227 typedef int64_t bluesky_time_hires;
228 bluesky_time_hires bluesky_now_hires();
230 /* In-memory representation of an inode within a Blue Sky server. This
231 * corresponds roughly with information that is committed to persistent
232 * storage. Locking/refcounting rules:
233 * - To access or modify any data fields, the lock must be held. This
234 * includes file blocks.
235 * - One reference is held by the BlueSkyFS inode hash table. If that is the
236 * only reference (and the inode is unlocked), the inode is subject to
237 * dropping from the cache.
238 * - Any pending operations should hold extra references to the inode as
239 * appropriate to keep it available until the operation completes.
240 * - Locking dependency order is, when multiple locks are to be acquired, to
241 * acquire locks on parents in the filesystem tree before children.
242 * (TODO: What about rename when we acquire locks in unrelated parts of the
244 * - An inode should not be locked while the filesystem lock is already held,
245 * since some code may do an inode lookup (which acquires the filesystem
246 * lock) while a different inode is locked.
254 BlueSkyFileType type;
259 /* Rather than track an inode number and generation number, we will simply
260 * never re-use a fileid after a file is deleted. 64 bits should be enough
261 * that we don't exhaust the identifier space. */
264 /* change_count is increased with every operation which modifies the inode,
265 * and can be used to determine if cached data is still valid.
266 * change_commit is the value of change_count when the inode was last
267 * committed to stable storage (the log).
268 * change_cloud tracks which version was last commited to cloud storage. */
269 uint64_t change_count, change_commit, change_cloud;
271 /* Timestamp for controlling when modified data is flushed to stable
272 * storage. When an inode is first modified from a clean state, this is
273 * set to the current time. If the inode is clean, it is set to zero. */
276 /* Last access time to this inode, for controlling cache evictions. */
279 /* Version of the object last serialized and committed to storage. */
280 BlueSkyCloudLog *committed_item;
282 /* Pointers to the linked-list elements for this inode in the accessed and
283 * dirty linked lists. We re-use the GList structure, using ->next to
284 * point to the head of the list and ->prev to point to the tail. The data
285 * element is unused. */
286 GList *unlogged_list, *accessed_list, *dirty_list;
288 int64_t atime; /* Microseconds since the Unix epoch */
291 int64_t ntime; /* "new time": time object was created */
293 /* File-specific fields */
297 /* Directory-specific fields */
298 GSequence *dirents; /* List of entries for READDIR */
299 GHashTable *dirhash; /* Hash table by name for LOOKUP */
300 GHashTable *dirhash_folded; /* As above, but case-folded */
301 uint64_t parent_inum; /* inode for ".."; 0 if the root directory */
303 /* Symlink-specific fields */
304 gchar *symlink_contents;
306 /* A field for short-term use internally while the lock is held. */
307 gpointer private_data;
310 /* A directory entry. The name is UTF-8 and is a freshly-allocated string.
311 * Each directory entry is listed in two indices: dirents is indexed by cookie
312 * and dirhash by name. The cookie is a randomly-assigned 32-bit value, unique
313 * within the directory, that remains unchanged until the entry is deleted. It
314 * is used to provide a stable key for restarting a READDIR call. */
317 gchar *name_folded; /* Name, folded for case-insensitive lookup */
322 /* File data is divided into fixed-size blocks (except the last block which may
323 * be short?). These blocks are backed by storage in a key/value store, but
324 * may also be dirty if modifications have been made in-memory that have not
326 #define BLUESKY_BLOCK_SIZE 32768ULL
327 #define BLUESKY_MAX_FILE_SIZE (BLUESKY_BLOCK_SIZE << 24)
329 BLUESKY_BLOCK_ZERO = 0, /* Data is all zeroes, not explicitly stored */
330 BLUESKY_BLOCK_REF = 1, /* Reference to cloud log item, data clean */
331 BLUESKY_BLOCK_DIRTY = 2, /* Data needs to be committed to store */
335 BlueSkyBlockType type;
336 BlueSkyCloudLog *ref; /* if REF: cloud log entry with data */
337 BlueSkyRCStr *dirty; /* if DIRTY: raw data in memory */
340 BlueSkyFS *bluesky_init_fs(gchar *name, BlueSkyStore *store,
341 const gchar *master_key);
343 gboolean bluesky_inode_is_ready(BlueSkyInode *inode);
345 int64_t bluesky_get_current_time();
346 void bluesky_inode_update_ctime(BlueSkyInode *inode, gboolean update_mtime);
347 uint64_t bluesky_fs_alloc_inode(BlueSkyFS *fs);
348 void bluesky_init_inode(BlueSkyInode *i, BlueSkyFileType type);
349 BlueSkyInode *bluesky_new_inode(uint64_t inum, BlueSkyFS *fs, BlueSkyFileType type);
351 BlueSkyInode *bluesky_get_inode(BlueSkyFS *fs, uint64_t inum);
352 void bluesky_inode_ref(BlueSkyInode *inode);
353 void bluesky_inode_unref(BlueSkyInode *inode);
354 void bluesky_insert_inode(BlueSkyFS *fs, BlueSkyInode *inode);
356 void bluesky_dirent_destroy(gpointer dirent);
357 uint64_t bluesky_directory_lookup(BlueSkyInode *inode, gchar *name);
358 uint64_t bluesky_directory_ilookup(BlueSkyInode *inode, gchar *name);
359 BlueSkyDirent *bluesky_directory_read(BlueSkyInode *dir, uint32_t cookie);
360 gboolean bluesky_directory_insert(BlueSkyInode *dir, const gchar *name,
362 void bluesky_directory_dump(BlueSkyInode *dir);
364 void bluesky_file_truncate(BlueSkyInode *inode, uint64_t size);
365 void bluesky_file_write(BlueSkyInode *inode, uint64_t offset,
366 const char *data, gint len);
367 void bluesky_file_read(BlueSkyInode *inode, uint64_t offset,
368 char *buf, gint len);
370 void bluesky_inode_fetch(BlueSkyFS *fs, uint64_t inum);
371 void bluesky_inode_prefetch(BlueSkyFS *fs, uint64_t inum);
373 gint bluesky_dirent_compare(gconstpointer a, gconstpointer b,
376 void bluesky_flushd_invoke(BlueSkyFS *fs);
377 void bluesky_flushd_invoke_conditional(BlueSkyFS *fs);
378 void bluesky_inode_do_sync(BlueSkyInode *inode);
379 void bluesky_flushd_thread_launch(BlueSkyFS *fs);
381 void bluesky_debug_dump(BlueSkyFS *fs);
383 /* Request response time tracking. */
384 typedef struct BlueSkyProfile {
391 BlueSkyProfile *bluesky_profile_new();
392 void bluesky_profile_free(BlueSkyProfile *profile);
393 void bluesky_profile_add_event(BlueSkyProfile *profile, char *message);
394 void bluesky_profile_print(BlueSkyProfile *profile);
395 BlueSkyProfile *bluesky_profile_get();
396 void bluesky_profile_set(BlueSkyProfile *profile);
397 void bluesky_profile_set_output(FILE *stream);