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. */
52 /* Disable aggregation of read requests. Fetch items individually. */
53 int disable_read_aggregation;
56 extern BlueSkyOptions bluesky_options;
58 /* Maximum number of threads to use in any particular thread pool, or -1 for no
60 extern int bluesky_max_threads;
62 /* A general-purpose counter for gathering run-time statistics. */
63 struct bluesky_stats {
68 struct bluesky_stats *bluesky_stats_new(const char *name);
69 void bluesky_stats_add(struct bluesky_stats *stats, int64_t value);
70 void bluesky_stats_dump_all();
71 void bluesky_stats_run_periodic_dump(FILE *f);
73 /* BlueSky status and error codes. Various frontends should translate these to
74 * the appropriate error code for whatever protocol they implement. */
76 BSTATUS_OK = 0, /* No error */
77 BSTATUS_IOERR, /* I/O error of some form */
78 BSTATUS_NOENT, /* File does not exist */
81 void bluesky_init(void);
83 gchar *bluesky_lowercase(const gchar *s);
85 struct BlueSkyCacheFile;
86 typedef struct BlueSkyCacheFile BlueSkyCacheFile;
90 BlueSkyCacheFile *mmap;
95 BlueSkyRCStr *bluesky_string_new(gpointer data, gsize len);
96 BlueSkyRCStr *bluesky_string_new_from_gstring(GString *s);
97 BlueSkyRCStr *bluesky_string_new_from_mmap(BlueSkyCacheFile *mmap,
98 int offset, gsize len);
99 void bluesky_string_ref(BlueSkyRCStr *string);
100 void bluesky_string_unref(BlueSkyRCStr *string);
101 BlueSkyRCStr *bluesky_string_dup(BlueSkyRCStr *string);
102 void bluesky_string_resize(BlueSkyRCStr *string, gsize len);
104 struct BlueSkyRangeset;
105 typedef struct BlueSkyRangeset BlueSkyRangeset;
107 uint64_t start, length;
109 } BlueSkyRangesetItem;
111 BlueSkyRangeset *bluesky_rangeset_new();
112 void bluesky_rangeset_free(BlueSkyRangeset *rangeset);
113 gboolean bluesky_rangeset_insert(BlueSkyRangeset *rangeset,
114 uint64_t start, uint64_t length,
116 const BlueSkyRangesetItem *bluesky_rangeset_lookup(BlueSkyRangeset *rangeset,
118 const BlueSkyRangesetItem *bluesky_rangeset_lookup_next(BlueSkyRangeset *rangeset, uint64_t offset);
119 void bluesky_rangeset_get_extents(BlueSkyRangeset *rangeset,
120 uint64_t *start, uint64_t *length);
122 /* Storage interface. This presents a key-value store abstraction, and can
123 * have multiple implementations: in-memory, on-disk, in-cloud. */
125 typedef struct BlueSkyStore BlueSkyStore;
128 typedef struct BlueSkyLog BlueSkyLog;
130 struct BlueSkyCloudLogState;
131 typedef struct BlueSkyCloudLogState BlueSkyCloudLogState;
133 struct BlueSkyCloudLog;
134 typedef struct BlueSkyCloudLog BlueSkyCloudLog;
136 void bluesky_store_init();
137 BlueSkyStore *bluesky_store_new(const gchar *type);
138 void bluesky_store_free(BlueSkyStore *store);
139 BlueSkyRCStr *bluesky_store_get(BlueSkyStore *store, const gchar *key);
140 void bluesky_store_put(BlueSkyStore *store,
141 const gchar *key, BlueSkyRCStr *val);
143 /* File types. The numeric values are chosen to match with those used in
147 BLUESKY_DIRECTORY = 2,
149 BLUESKY_CHARACTER = 4,
154 /* Special types used only internally. */
155 BLUESKY_PENDING = 0, /* Inode being loaded; type not yet determined */
156 BLUESKY_INVALID = -1, /* Inode is invalid (failed to load) */
159 /* Filesystem state. Each filesystem which is exported is represented by a
160 * single bluesky_fs structure in memory. */
161 struct BlueSkyCryptKeys;
166 gchar *name; /* Descriptive name for the filesystem */
167 GHashTable *inodes; /* Cached inodes */
168 uint64_t next_inum; /* Next available inode for allocation */
172 BlueSkyCloudLogState *log_state;
174 /* Filesystem crypto keys */
176 struct BlueSkyCryptKeys *keys;
178 /* Accounting for memory used for caches. Space is measured in blocks, not
179 * bytes. Updates to these variables must be made atomically. */
182 /* Like above, but tracking data stored in the cloudlog entries
184 * - cache_log_dirty: data uncommitted to journal and cloud
185 * - cache_log_writeback: data being written to journal
186 * - cache_log_journal: data committed to journal
187 * - cache_log_cloud: data written to cloud as well
188 * Log entries should progress from the top state to the bottom, and are
189 * only ever counted in one category at a time. */
190 gint cache_log_dirty, cache_log_writeback,
191 cache_log_journal, cache_log_cloud;
193 /* Linked list of inodes, sorted by access/modification times for cache
194 * management. Editing these lists is protected by the filesystem lock; to
195 * avoid deadlock do not attempt to take any other locks while the FS lock
196 * is held for list editing purposes. Items at the head of the list are
197 * most recently accessed/modified. */
198 GList unlogged_list; // Changes not yet synced to journal
199 GList dirty_list; // Not yet written to cloud storage
200 GList accessed_list; // All in-memory inodes
202 /* Mutex for the flush daemon, to prevent concurrent execution. */
205 /* Used to wait for the cache daemon to free up space */
208 /* Mapping of object identifiers (blocks, inodes) to physical location (in
209 * the local cache or in the logs in the cloud). */
210 GHashTable *locations;
212 /* The inode map, which maps inode numbers to the location of the most
214 GSequence *inode_map;
216 /* Queue for asynchronous cloudlog unrefs, where needed. */
217 GAsyncQueue *unref_queue;
219 /* Thread pool for asynchronous inode fetches */
220 GThreadPool *inode_fetch_thread_pool;
223 /* Inode number of the root directory. */
224 #define BLUESKY_ROOT_INUM 1
226 /* Timestamp, measured in microseconds since the Unix epoch. */
227 typedef int64_t bluesky_time;
229 /* High-resolution timer, measured in nanoseconds. */
230 typedef int64_t bluesky_time_hires;
231 bluesky_time_hires bluesky_now_hires();
233 /* In-memory representation of an inode within a Blue Sky server. This
234 * corresponds roughly with information that is committed to persistent
235 * storage. Locking/refcounting rules:
236 * - To access or modify any data fields, the lock must be held. This
237 * includes file blocks.
238 * - One reference is held by the BlueSkyFS inode hash table. If that is the
239 * only reference (and the inode is unlocked), the inode is subject to
240 * dropping from the cache.
241 * - Any pending operations should hold extra references to the inode as
242 * appropriate to keep it available until the operation completes.
243 * - Locking dependency order is, when multiple locks are to be acquired, to
244 * acquire locks on parents in the filesystem tree before children.
245 * (TODO: What about rename when we acquire locks in unrelated parts of the
247 * - An inode should not be locked while the filesystem lock is already held,
248 * since some code may do an inode lookup (which acquires the filesystem
249 * lock) while a different inode is locked.
257 BlueSkyFileType type;
262 /* Rather than track an inode number and generation number, we will simply
263 * never re-use a fileid after a file is deleted. 64 bits should be enough
264 * that we don't exhaust the identifier space. */
267 /* change_count is increased with every operation which modifies the inode,
268 * and can be used to determine if cached data is still valid.
269 * change_commit is the value of change_count when the inode was last
270 * committed to stable storage (the log).
271 * change_cloud tracks which version was last commited to cloud storage. */
272 uint64_t change_count, change_commit, change_cloud;
274 /* Timestamp for controlling when modified data is flushed to stable
275 * storage. When an inode is first modified from a clean state, this is
276 * set to the current time. If the inode is clean, it is set to zero. */
279 /* Last access time to this inode, for controlling cache evictions. */
282 /* Version of the object last serialized and committed to storage. */
283 BlueSkyCloudLog *committed_item;
285 /* Pointers to the linked-list elements for this inode in the accessed and
286 * dirty linked lists. We re-use the GList structure, using ->next to
287 * point to the head of the list and ->prev to point to the tail. The data
288 * element is unused. */
289 GList *unlogged_list, *accessed_list, *dirty_list;
291 int64_t atime; /* Microseconds since the Unix epoch */
294 int64_t ntime; /* "new time": time object was created */
296 /* File-specific fields */
300 /* Directory-specific fields */
301 GSequence *dirents; /* List of entries for READDIR */
302 GHashTable *dirhash; /* Hash table by name for LOOKUP */
303 GHashTable *dirhash_folded; /* As above, but case-folded */
304 uint64_t parent_inum; /* inode for ".."; 0 if the root directory */
306 /* Symlink-specific fields */
307 gchar *symlink_contents;
309 /* A field for short-term use internally while the lock is held. */
310 gpointer private_data;
313 /* A directory entry. The name is UTF-8 and is a freshly-allocated string.
314 * Each directory entry is listed in two indices: dirents is indexed by cookie
315 * and dirhash by name. The cookie is a randomly-assigned 32-bit value, unique
316 * within the directory, that remains unchanged until the entry is deleted. It
317 * is used to provide a stable key for restarting a READDIR call. */
320 gchar *name_folded; /* Name, folded for case-insensitive lookup */
325 /* File data is divided into fixed-size blocks (except the last block which may
326 * be short?). These blocks are backed by storage in a key/value store, but
327 * may also be dirty if modifications have been made in-memory that have not
329 #define BLUESKY_BLOCK_SIZE 32768ULL
330 #define BLUESKY_MAX_FILE_SIZE (BLUESKY_BLOCK_SIZE << 24)
332 BLUESKY_BLOCK_ZERO = 0, /* Data is all zeroes, not explicitly stored */
333 BLUESKY_BLOCK_REF = 1, /* Reference to cloud log item, data clean */
334 BLUESKY_BLOCK_DIRTY = 2, /* Data needs to be committed to store */
338 BlueSkyBlockType type;
339 BlueSkyCloudLog *ref; /* if REF: cloud log entry with data */
340 BlueSkyRCStr *dirty; /* if DIRTY: raw data in memory */
343 BlueSkyFS *bluesky_init_fs(gchar *name, BlueSkyStore *store,
344 const gchar *master_key);
346 gboolean bluesky_inode_is_ready(BlueSkyInode *inode);
348 int64_t bluesky_get_current_time();
349 void bluesky_inode_update_ctime(BlueSkyInode *inode, gboolean update_mtime);
350 uint64_t bluesky_fs_alloc_inode(BlueSkyFS *fs);
351 void bluesky_init_inode(BlueSkyInode *i, BlueSkyFileType type);
352 BlueSkyInode *bluesky_new_inode(uint64_t inum, BlueSkyFS *fs, BlueSkyFileType type);
354 BlueSkyInode *bluesky_get_inode(BlueSkyFS *fs, uint64_t inum);
355 void bluesky_inode_ref(BlueSkyInode *inode);
356 void bluesky_inode_unref(BlueSkyInode *inode);
357 void bluesky_insert_inode(BlueSkyFS *fs, BlueSkyInode *inode);
359 void bluesky_dirent_destroy(gpointer dirent);
360 uint64_t bluesky_directory_lookup(BlueSkyInode *inode, gchar *name);
361 uint64_t bluesky_directory_ilookup(BlueSkyInode *inode, gchar *name);
362 BlueSkyDirent *bluesky_directory_read(BlueSkyInode *dir, uint32_t cookie);
363 gboolean bluesky_directory_insert(BlueSkyInode *dir, const gchar *name,
365 void bluesky_directory_dump(BlueSkyInode *dir);
367 void bluesky_file_truncate(BlueSkyInode *inode, uint64_t size);
368 void bluesky_file_write(BlueSkyInode *inode, uint64_t offset,
369 const char *data, gint len);
370 void bluesky_file_read(BlueSkyInode *inode, uint64_t offset,
371 char *buf, gint len);
373 void bluesky_inode_fetch(BlueSkyFS *fs, uint64_t inum);
374 void bluesky_inode_prefetch(BlueSkyFS *fs, uint64_t inum);
376 gint bluesky_dirent_compare(gconstpointer a, gconstpointer b,
379 void bluesky_flushd_invoke(BlueSkyFS *fs);
380 void bluesky_flushd_invoke_conditional(BlueSkyFS *fs);
381 void bluesky_inode_do_sync(BlueSkyInode *inode);
382 void bluesky_flushd_thread_launch(BlueSkyFS *fs);
384 void bluesky_debug_dump(BlueSkyFS *fs);
386 /* Request response time tracking. */
387 typedef struct BlueSkyProfile {
394 BlueSkyProfile *bluesky_profile_new();
395 void bluesky_profile_free(BlueSkyProfile *profile);
396 void bluesky_profile_add_event(BlueSkyProfile *profile, char *message);
397 void bluesky_profile_print(BlueSkyProfile *profile);
398 BlueSkyProfile *bluesky_profile_get();
399 void bluesky_profile_set(BlueSkyProfile *profile);
400 void bluesky_profile_set_output(FILE *stream);