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 /* Perform all get/put operations synchronously. */
24 int synchronous_stores;
26 /* 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. */
42 extern BlueSkyOptions bluesky_options;
44 /* Maximum number of threads to use in any particular thread pool, or -1 for no
46 extern int bluesky_max_threads;
48 /* A general-purpose counter for gathering run-time statistics. */
49 struct bluesky_stats {
54 struct bluesky_stats *bluesky_stats_new(const char *name);
55 void bluesky_stats_add(struct bluesky_stats *stats, int64_t value);
56 void bluesky_stats_dump_all();
57 void bluesky_stats_run_periodic_dump(FILE *f);
59 /* BlueSky status and error codes. Various frontends should translate these to
60 * the appropriate error code for whatever protocol they implement. */
62 BSTATUS_OK = 0, /* No error */
63 BSTATUS_IOERR, /* I/O error of some form */
64 BSTATUS_NOENT, /* File does not exist */
67 void bluesky_init(void);
69 gchar *bluesky_lowercase(const gchar *s);
71 struct BlueSkyCacheFile;
72 typedef struct BlueSkyCacheFile BlueSkyCacheFile;
76 BlueSkyCacheFile *mmap;
81 BlueSkyRCStr *bluesky_string_new(gpointer data, gsize len);
82 BlueSkyRCStr *bluesky_string_new_from_gstring(GString *s);
83 BlueSkyRCStr *bluesky_string_new_from_mmap(BlueSkyCacheFile *mmap,
84 int offset, gsize len);
85 void bluesky_string_ref(BlueSkyRCStr *string);
86 void bluesky_string_unref(BlueSkyRCStr *string);
87 BlueSkyRCStr *bluesky_string_dup(BlueSkyRCStr *string);
88 void bluesky_string_resize(BlueSkyRCStr *string, gsize len);
90 struct BlueSkyRangeset;
91 typedef struct BlueSkyRangeset BlueSkyRangeset;
93 uint64_t start, length;
95 } BlueSkyRangesetItem;
97 BlueSkyRangeset *bluesky_rangeset_new();
98 void bluesky_rangeset_free(BlueSkyRangeset *rangeset);
99 gboolean bluesky_rangeset_insert(BlueSkyRangeset *rangeset,
100 uint64_t start, uint64_t length,
102 const BlueSkyRangesetItem *bluesky_rangeset_lookup(BlueSkyRangeset *rangeset,
104 const BlueSkyRangesetItem *bluesky_rangeset_lookup_next(BlueSkyRangeset *rangeset, uint64_t offset);
105 void bluesky_rangeset_get_extents(BlueSkyRangeset *rangeset,
106 uint64_t *start, uint64_t *length);
108 /* Storage interface. This presents a key-value store abstraction, and can
109 * have multiple implementations: in-memory, on-disk, in-cloud. */
111 typedef struct BlueSkyStore BlueSkyStore;
114 typedef struct BlueSkyLog BlueSkyLog;
116 struct BlueSkyCloudLogState;
117 typedef struct BlueSkyCloudLogState BlueSkyCloudLogState;
119 struct BlueSkyCloudLog;
120 typedef struct BlueSkyCloudLog BlueSkyCloudLog;
122 void bluesky_store_init();
123 BlueSkyStore *bluesky_store_new(const gchar *type);
124 void bluesky_store_free(BlueSkyStore *store);
125 BlueSkyRCStr *bluesky_store_get(BlueSkyStore *store, const gchar *key);
126 void bluesky_store_put(BlueSkyStore *store,
127 const gchar *key, BlueSkyRCStr *val);
129 /* File types. The numeric values are chosen to match with those used in
133 BLUESKY_DIRECTORY = 2,
135 BLUESKY_CHARACTER = 4,
140 /* Special types used only internally. */
141 BLUESKY_PENDING = 0, /* Inode being loaded; type not yet determined */
142 BLUESKY_INVALID = -1, /* Inode is invalid (failed to load) */
145 /* Filesystem state. Each filesystem which is exported is represented by a
146 * single bluesky_fs structure in memory. */
147 struct BlueSkyCryptKeys;
152 gchar *name; /* Descriptive name for the filesystem */
153 GHashTable *inodes; /* Cached inodes */
154 uint64_t next_inum; /* Next available inode for allocation */
158 BlueSkyCloudLogState *log_state;
160 /* Filesystem crypto keys */
162 struct BlueSkyCryptKeys *keys;
164 /* Accounting for memory used for caches. Space is measured in blocks, not
165 * bytes. Updates to these variables must be made atomically. */
168 /* Like above, but tracking data stored in the cloudlog entries
170 * - cache_log_dirty: data uncommitted to journal and cloud
171 * - cache_log_writeback: data being written to journal
172 * - cache_log_journal: data committed to journal
173 * - cache_log_cloud: data written to cloud as well
174 * Log entries should progress from the top state to the bottom, and are
175 * only ever counted in one category at a time. */
176 gint cache_log_dirty, cache_log_writeback,
177 cache_log_journal, cache_log_cloud;
179 /* Linked list of inodes, sorted by access/modification times for cache
180 * management. Editing these lists is protected by the filesystem lock; to
181 * avoid deadlock do not attempt to take any other locks while the FS lock
182 * is held for list editing purposes. Items at the head of the list are
183 * most recently accessed/modified. */
184 GList unlogged_list; // Changes not yet synced to journal
185 GList dirty_list; // Not yet written to cloud storage
186 GList accessed_list; // All in-memory inodes
188 /* Mutex for the flush daemon, to prevent concurrent execution. */
191 /* Used to wait for the cache daemon to free up space */
194 /* Mapping of object identifiers (blocks, inodes) to physical location (in
195 * the local cache or in the logs in the cloud). */
196 GHashTable *locations;
198 /* The inode map, which maps inode numbers to the location of the most
200 GSequence *inode_map;
202 /* Queue for asynchronous cloudlog unrefs, where needed. */
203 GAsyncQueue *unref_queue;
205 /* Thread pool for asynchronous inode fetches */
206 GThreadPool *inode_fetch_thread_pool;
209 /* Inode number of the root directory. */
210 #define BLUESKY_ROOT_INUM 1
212 /* Timestamp, measured in microseconds since the Unix epoch. */
213 typedef int64_t bluesky_time;
215 /* High-resolution timer, measured in nanoseconds. */
216 typedef int64_t bluesky_time_hires;
217 bluesky_time_hires bluesky_now_hires();
219 /* In-memory representation of an inode within a Blue Sky server. This
220 * corresponds roughly with information that is committed to persistent
221 * storage. Locking/refcounting rules:
222 * - To access or modify any data fields, the lock must be held. This
223 * includes file blocks.
224 * - One reference is held by the BlueSkyFS inode hash table. If that is the
225 * only reference (and the inode is unlocked), the inode is subject to
226 * dropping from the cache.
227 * - Any pending operations should hold extra references to the inode as
228 * appropriate to keep it available until the operation completes.
229 * - Locking dependency order is, when multiple locks are to be acquired, to
230 * acquire locks on parents in the filesystem tree before children.
231 * (TODO: What about rename when we acquire locks in unrelated parts of the
233 * - An inode should not be locked while the filesystem lock is already held,
234 * since some code may do an inode lookup (which acquires the filesystem
235 * lock) while a different inode is locked.
243 BlueSkyFileType type;
248 /* Rather than track an inode number and generation number, we will simply
249 * never re-use a fileid after a file is deleted. 64 bits should be enough
250 * that we don't exhaust the identifier space. */
253 /* change_count is increased with every operation which modifies the inode,
254 * and can be used to determine if cached data is still valid.
255 * change_commit is the value of change_count when the inode was last
256 * committed to stable storage (the log).
257 * change_cloud tracks which version was last commited to cloud storage. */
258 uint64_t change_count, change_commit, change_cloud;
260 /* Timestamp for controlling when modified data is flushed to stable
261 * storage. When an inode is first modified from a clean state, this is
262 * set to the current time. If the inode is clean, it is set to zero. */
265 /* Last access time to this inode, for controlling cache evictions. */
268 /* Version of the object last serialized and committed to storage. */
269 BlueSkyCloudLog *committed_item;
271 /* Pointers to the linked-list elements for this inode in the accessed and
272 * dirty linked lists. We re-use the GList structure, using ->next to
273 * point to the head of the list and ->prev to point to the tail. The data
274 * element is unused. */
275 GList *unlogged_list, *accessed_list, *dirty_list;
277 int64_t atime; /* Microseconds since the Unix epoch */
280 int64_t ntime; /* "new time": time object was created */
282 /* File-specific fields */
286 /* Directory-specific fields */
287 GSequence *dirents; /* List of entries for READDIR */
288 GHashTable *dirhash; /* Hash table by name for LOOKUP */
289 GHashTable *dirhash_folded; /* As above, but case-folded */
290 uint64_t parent_inum; /* inode for ".."; 0 if the root directory */
292 /* Symlink-specific fields */
293 gchar *symlink_contents;
295 /* A field for short-term use internally while the lock is held. */
296 gpointer private_data;
299 /* A directory entry. The name is UTF-8 and is a freshly-allocated string.
300 * Each directory entry is listed in two indices: dirents is indexed by cookie
301 * and dirhash by name. The cookie is a randomly-assigned 32-bit value, unique
302 * within the directory, that remains unchanged until the entry is deleted. It
303 * is used to provide a stable key for restarting a READDIR call. */
306 gchar *name_folded; /* Name, folded for case-insensitive lookup */
311 /* File data is divided into fixed-size blocks (except the last block which may
312 * be short?). These blocks are backed by storage in a key/value store, but
313 * may also be dirty if modifications have been made in-memory that have not
315 #define BLUESKY_BLOCK_SIZE 32768ULL
316 #define BLUESKY_MAX_FILE_SIZE (BLUESKY_BLOCK_SIZE << 24)
318 BLUESKY_BLOCK_ZERO = 0, /* Data is all zeroes, not explicitly stored */
319 BLUESKY_BLOCK_REF = 1, /* Reference to cloud log item, data clean */
320 BLUESKY_BLOCK_DIRTY = 2, /* Data needs to be committed to store */
324 BlueSkyBlockType type;
325 BlueSkyCloudLog *ref; /* if REF: cloud log entry with data */
326 BlueSkyRCStr *dirty; /* if DIRTY: raw data in memory */
329 BlueSkyFS *bluesky_init_fs(gchar *name, BlueSkyStore *store,
330 const gchar *master_key);
332 gboolean bluesky_inode_is_ready(BlueSkyInode *inode);
334 int64_t bluesky_get_current_time();
335 void bluesky_inode_update_ctime(BlueSkyInode *inode, gboolean update_mtime);
336 uint64_t bluesky_fs_alloc_inode(BlueSkyFS *fs);
337 void bluesky_init_inode(BlueSkyInode *i, BlueSkyFileType type);
338 BlueSkyInode *bluesky_new_inode(uint64_t inum, BlueSkyFS *fs, BlueSkyFileType type);
340 BlueSkyInode *bluesky_get_inode(BlueSkyFS *fs, uint64_t inum);
341 void bluesky_inode_ref(BlueSkyInode *inode);
342 void bluesky_inode_unref(BlueSkyInode *inode);
343 void bluesky_insert_inode(BlueSkyFS *fs, BlueSkyInode *inode);
345 void bluesky_dirent_destroy(gpointer dirent);
346 uint64_t bluesky_directory_lookup(BlueSkyInode *inode, gchar *name);
347 uint64_t bluesky_directory_ilookup(BlueSkyInode *inode, gchar *name);
348 BlueSkyDirent *bluesky_directory_read(BlueSkyInode *dir, uint32_t cookie);
349 gboolean bluesky_directory_insert(BlueSkyInode *dir, const gchar *name,
351 void bluesky_directory_dump(BlueSkyInode *dir);
353 void bluesky_file_truncate(BlueSkyInode *inode, uint64_t size);
354 void bluesky_file_write(BlueSkyInode *inode, uint64_t offset,
355 const char *data, gint len);
356 void bluesky_file_read(BlueSkyInode *inode, uint64_t offset,
357 char *buf, gint len);
359 void bluesky_inode_fetch(BlueSkyFS *fs, uint64_t inum);
360 void bluesky_inode_prefetch(BlueSkyFS *fs, uint64_t inum);
362 gint bluesky_dirent_compare(gconstpointer a, gconstpointer b,
365 void bluesky_flushd_invoke(BlueSkyFS *fs);
366 void bluesky_flushd_invoke_conditional(BlueSkyFS *fs);
367 void bluesky_inode_do_sync(BlueSkyInode *inode);
368 void bluesky_flushd_thread_launch(BlueSkyFS *fs);
370 void bluesky_debug_dump(BlueSkyFS *fs);
372 /* Request response time tracking. */
373 typedef struct BlueSkyProfile {
380 BlueSkyProfile *bluesky_profile_new();
381 void bluesky_profile_free(BlueSkyProfile *profile);
382 void bluesky_profile_add_event(BlueSkyProfile *profile, char *message);
383 void bluesky_profile_print(BlueSkyProfile *profile);
384 BlueSkyProfile *bluesky_profile_get();
385 void bluesky_profile_set(BlueSkyProfile *profile);
386 void bluesky_profile_set_output(FILE *stream);