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();
58 /* BlueSky status and error codes. Various frontends should translate these to
59 * the appropriate error code for whatever protocol they implement. */
61 BSTATUS_OK = 0, /* No error */
62 BSTATUS_IOERR, /* I/O error of some form */
63 BSTATUS_NOENT, /* File does not exist */
66 void bluesky_init(void);
68 gchar *bluesky_lowercase(const gchar *s);
70 struct BlueSkyCacheFile;
71 typedef struct BlueSkyCacheFile BlueSkyCacheFile;
75 BlueSkyCacheFile *mmap;
80 BlueSkyRCStr *bluesky_string_new(gpointer data, gsize len);
81 BlueSkyRCStr *bluesky_string_new_from_gstring(GString *s);
82 BlueSkyRCStr *bluesky_string_new_from_mmap(BlueSkyCacheFile *mmap,
83 int offset, gsize len);
84 void bluesky_string_ref(BlueSkyRCStr *string);
85 void bluesky_string_unref(BlueSkyRCStr *string);
86 BlueSkyRCStr *bluesky_string_dup(BlueSkyRCStr *string);
87 void bluesky_string_resize(BlueSkyRCStr *string, gsize len);
89 struct BlueSkyRangeset;
90 typedef struct BlueSkyRangeset BlueSkyRangeset;
92 uint64_t start, length;
94 } BlueSkyRangesetItem;
96 BlueSkyRangeset *bluesky_rangeset_new();
97 void bluesky_rangeset_free(BlueSkyRangeset *rangeset);
98 gboolean bluesky_rangeset_insert(BlueSkyRangeset *rangeset,
99 uint64_t start, uint64_t length,
101 const BlueSkyRangesetItem *bluesky_rangeset_lookup(BlueSkyRangeset *rangeset,
103 const BlueSkyRangesetItem *bluesky_rangeset_lookup_next(BlueSkyRangeset *rangeset, uint64_t offset);
104 void bluesky_rangeset_get_extents(BlueSkyRangeset *rangeset,
105 uint64_t *start, uint64_t *length);
107 /* Storage interface. This presents a key-value store abstraction, and can
108 * have multiple implementations: in-memory, on-disk, in-cloud. */
110 typedef struct BlueSkyStore BlueSkyStore;
113 typedef struct BlueSkyLog BlueSkyLog;
115 struct BlueSkyCloudLogState;
116 typedef struct BlueSkyCloudLogState BlueSkyCloudLogState;
118 struct BlueSkyCloudLog;
119 typedef struct BlueSkyCloudLog BlueSkyCloudLog;
121 void bluesky_store_init();
122 BlueSkyStore *bluesky_store_new(const gchar *type);
123 void bluesky_store_free(BlueSkyStore *store);
124 BlueSkyRCStr *bluesky_store_get(BlueSkyStore *store, const gchar *key);
125 void bluesky_store_put(BlueSkyStore *store,
126 const gchar *key, BlueSkyRCStr *val);
128 /* File types. The numeric values are chosen to match with those used in
132 BLUESKY_DIRECTORY = 2,
134 BLUESKY_CHARACTER = 4,
139 /* Special types used only internally. */
140 BLUESKY_PENDING = 0, /* Inode being loaded; type not yet determined */
141 BLUESKY_INVALID = -1, /* Inode is invalid (failed to load) */
144 /* Filesystem state. Each filesystem which is exported is represented by a
145 * single bluesky_fs structure in memory. */
146 struct BlueSkyCryptKeys;
151 gchar *name; /* Descriptive name for the filesystem */
152 GHashTable *inodes; /* Cached inodes */
153 uint64_t next_inum; /* Next available inode for allocation */
157 BlueSkyCloudLogState *log_state;
159 /* Filesystem crypto keys */
161 struct BlueSkyCryptKeys *keys;
163 /* Accounting for memory used for caches. Space is measured in blocks, not
164 * bytes. Updates to these variables must be made atomically. */
167 /* Like above, but tracking data stored in the cloudlog entries
169 * - cache_log_dirty: data uncommitted to journal and cloud
170 * - cache_log_writeback: data being written to journal
171 * - cache_log_journal: data committed to journal
172 * - cache_log_cloud: data written to cloud as well
173 * Log entries should progress from the top state to the bottom, and are
174 * only ever counted in one category at a time. */
175 gint cache_log_dirty, cache_log_writeback,
176 cache_log_journal, cache_log_cloud;
178 /* Linked list of inodes, sorted by access/modification times for cache
179 * management. Editing these lists is protected by the filesystem lock; to
180 * avoid deadlock do not attempt to take any other locks while the FS lock
181 * is held for list editing purposes. Items at the head of the list are
182 * most recently accessed/modified. */
183 GList unlogged_list; // Changes not yet synced to journal
184 GList dirty_list; // Not yet written to cloud storage
185 GList accessed_list; // All in-memory inodes
187 /* Mutex for the flush daemon, to prevent concurrent execution. */
190 /* Used to wait for the cache daemon to free up space */
193 /* Mapping of object identifiers (blocks, inodes) to physical location (in
194 * the local cache or in the logs in the cloud). */
195 GHashTable *locations;
197 /* The inode map, which maps inode numbers to the location of the most
199 GSequence *inode_map;
201 /* Queue for asynchronous cloudlog unrefs, where needed. */
202 GAsyncQueue *unref_queue;
204 /* Thread pool for asynchronous inode fetches */
205 GThreadPool *inode_fetch_thread_pool;
208 /* Inode number of the root directory. */
209 #define BLUESKY_ROOT_INUM 1
211 /* Timestamp, measured in microseconds since the Unix epoch. */
212 typedef int64_t bluesky_time;
214 /* High-resolution timer, measured in nanoseconds. */
215 typedef int64_t bluesky_time_hires;
216 bluesky_time_hires bluesky_now_hires();
218 /* In-memory representation of an inode within a Blue Sky server. This
219 * corresponds roughly with information that is committed to persistent
220 * storage. Locking/refcounting rules:
221 * - To access or modify any data fields, the lock must be held. This
222 * includes file blocks.
223 * - One reference is held by the BlueSkyFS inode hash table. If that is the
224 * only reference (and the inode is unlocked), the inode is subject to
225 * dropping from the cache.
226 * - Any pending operations should hold extra references to the inode as
227 * appropriate to keep it available until the operation completes.
228 * - Locking dependency order is, when multiple locks are to be acquired, to
229 * acquire locks on parents in the filesystem tree before children.
230 * (TODO: What about rename when we acquire locks in unrelated parts of the
232 * - An inode should not be locked while the filesystem lock is already held,
233 * since some code may do an inode lookup (which acquires the filesystem
234 * lock) while a different inode is locked.
242 BlueSkyFileType type;
247 /* Rather than track an inode number and generation number, we will simply
248 * never re-use a fileid after a file is deleted. 64 bits should be enough
249 * that we don't exhaust the identifier space. */
252 /* change_count is increased with every operation which modifies the inode,
253 * and can be used to determine if cached data is still valid.
254 * change_commit is the value of change_count when the inode was last
255 * committed to stable storage (the log).
256 * change_cloud tracks which version was last commited to cloud storage. */
257 uint64_t change_count, change_commit, change_cloud;
259 /* Timestamp for controlling when modified data is flushed to stable
260 * storage. When an inode is first modified from a clean state, this is
261 * set to the current time. If the inode is clean, it is set to zero. */
264 /* Last access time to this inode, for controlling cache evictions. */
267 /* Version of the object last serialized and committed to storage. */
268 BlueSkyCloudLog *committed_item;
270 /* Pointers to the linked-list elements for this inode in the accessed and
271 * dirty linked lists. We re-use the GList structure, using ->next to
272 * point to the head of the list and ->prev to point to the tail. The data
273 * element is unused. */
274 GList *unlogged_list, *accessed_list, *dirty_list;
276 int64_t atime; /* Microseconds since the Unix epoch */
279 int64_t ntime; /* "new time": time object was created */
281 /* File-specific fields */
285 /* Directory-specific fields */
286 GSequence *dirents; /* List of entries for READDIR */
287 GHashTable *dirhash; /* Hash table by name for LOOKUP */
288 GHashTable *dirhash_folded; /* As above, but case-folded */
289 uint64_t parent_inum; /* inode for ".."; 0 if the root directory */
291 /* Symlink-specific fields */
292 gchar *symlink_contents;
294 /* A field for short-term use internally while the lock is held. */
295 gpointer private_data;
298 /* A directory entry. The name is UTF-8 and is a freshly-allocated string.
299 * Each directory entry is listed in two indices: dirents is indexed by cookie
300 * and dirhash by name. The cookie is a randomly-assigned 32-bit value, unique
301 * within the directory, that remains unchanged until the entry is deleted. It
302 * is used to provide a stable key for restarting a READDIR call. */
305 gchar *name_folded; /* Name, folded for case-insensitive lookup */
310 /* File data is divided into fixed-size blocks (except the last block which may
311 * be short?). These blocks are backed by storage in a key/value store, but
312 * may also be dirty if modifications have been made in-memory that have not
314 #define BLUESKY_BLOCK_SIZE 32768ULL
315 #define BLUESKY_MAX_FILE_SIZE (BLUESKY_BLOCK_SIZE << 24)
317 BLUESKY_BLOCK_ZERO = 0, /* Data is all zeroes, not explicitly stored */
318 BLUESKY_BLOCK_REF = 1, /* Reference to cloud log item, data clean */
319 BLUESKY_BLOCK_DIRTY = 2, /* Data needs to be committed to store */
323 BlueSkyBlockType type;
324 BlueSkyCloudLog *ref; /* if REF: cloud log entry with data */
325 BlueSkyRCStr *dirty; /* if DIRTY: raw data in memory */
328 BlueSkyFS *bluesky_init_fs(gchar *name, BlueSkyStore *store,
329 const gchar *master_key);
331 gboolean bluesky_inode_is_ready(BlueSkyInode *inode);
333 int64_t bluesky_get_current_time();
334 void bluesky_inode_update_ctime(BlueSkyInode *inode, gboolean update_mtime);
335 uint64_t bluesky_fs_alloc_inode(BlueSkyFS *fs);
336 void bluesky_init_inode(BlueSkyInode *i, BlueSkyFileType type);
337 BlueSkyInode *bluesky_new_inode(uint64_t inum, BlueSkyFS *fs, BlueSkyFileType type);
339 BlueSkyInode *bluesky_get_inode(BlueSkyFS *fs, uint64_t inum);
340 void bluesky_inode_ref(BlueSkyInode *inode);
341 void bluesky_inode_unref(BlueSkyInode *inode);
342 void bluesky_insert_inode(BlueSkyFS *fs, BlueSkyInode *inode);
344 void bluesky_dirent_destroy(gpointer dirent);
345 uint64_t bluesky_directory_lookup(BlueSkyInode *inode, gchar *name);
346 uint64_t bluesky_directory_ilookup(BlueSkyInode *inode, gchar *name);
347 BlueSkyDirent *bluesky_directory_read(BlueSkyInode *dir, uint32_t cookie);
348 gboolean bluesky_directory_insert(BlueSkyInode *dir, const gchar *name,
350 void bluesky_directory_dump(BlueSkyInode *dir);
352 void bluesky_file_truncate(BlueSkyInode *inode, uint64_t size);
353 void bluesky_file_write(BlueSkyInode *inode, uint64_t offset,
354 const char *data, gint len);
355 void bluesky_file_read(BlueSkyInode *inode, uint64_t offset,
356 char *buf, gint len);
358 void bluesky_inode_fetch(BlueSkyFS *fs, uint64_t inum);
360 gint bluesky_dirent_compare(gconstpointer a, gconstpointer b,
363 void bluesky_flushd_invoke(BlueSkyFS *fs);
364 void bluesky_flushd_invoke_conditional(BlueSkyFS *fs);
365 void bluesky_inode_do_sync(BlueSkyInode *inode);
366 void bluesky_flushd_thread_launch(BlueSkyFS *fs);
368 void bluesky_debug_dump(BlueSkyFS *fs);
370 /* Request response time tracking. */
371 typedef struct BlueSkyProfile {
378 BlueSkyProfile *bluesky_profile_new();
379 void bluesky_profile_free(BlueSkyProfile *profile);
380 void bluesky_profile_add_event(BlueSkyProfile *profile, char *message);
381 void bluesky_profile_print(BlueSkyProfile *profile);
382 BlueSkyProfile *bluesky_profile_get();
383 void bluesky_profile_set(BlueSkyProfile *profile);
384 void bluesky_profile_set_output(FILE *stream);