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>
9 /* Declarations internal to the BlueSky library. This header file should not
10 * be included by any users of the library (such as any filesystem
11 * proxy)--external users should only include bluesky.h. */
13 #ifndef _BLUESKY_PRIVATE_H
14 #define _BLUESKY_PRIVATE_H
22 extern int bluesky_verbose;
24 /* Target cache size levels. */
25 extern int bluesky_watermark_low_dirty;
26 extern int bluesky_watermark_medium_dirty;
27 extern int bluesky_watermark_high_dirty;
29 extern int bluesky_watermark_low_total;
30 extern int bluesky_watermark_medium_total;
31 extern int bluesky_watermark_high_total;
33 /* TODO: Make this go away entirely. */
34 BlueSkyFS *bluesky_new_fs(gchar *name);
36 void bluesky_inode_free_resources(BlueSkyInode *inode);
38 /* Linked list update functions for LRU lists. */
39 void bluesky_list_unlink(GList *head, GList *item);
40 GList *bluesky_list_prepend(GList *head, BlueSkyInode *inode);
41 GList *bluesky_list_append(GList *head, BlueSkyInode *inode);
42 BlueSkyInode *bluesky_list_head(GList *head);
43 BlueSkyInode *bluesky_list_tail(GList *head);
45 /* Serialization and deserialization of filesystem data for storing to
46 * persistent storage. */
47 void bluesky_serialize_superblock(GString *out, BlueSkyFS *fs);
48 BlueSkyFS *bluesky_deserialize_superblock(const gchar *buf);
49 BlueSkyCloudLog *bluesky_serialize_inode(BlueSkyInode *inode);
50 gboolean bluesky_deserialize_inode(BlueSkyInode *inode, BlueSkyCloudLog *item);
52 void bluesky_deserialize_cloudlog(BlueSkyCloudLog *item,
56 void bluesky_serialize_cloudlog(BlueSkyCloudLog *log,
58 GString *authenticated,
61 /* Cryptographic operations. */
62 #define CRYPTO_BLOCK_SIZE 16 /* 128-bit AES */
63 #define CRYPTO_KEY_SIZE 16
64 #define CRYPTO_HASH_SIZE 32 /* SHA-256 */
66 typedef struct BlueSkyCryptKeys {
67 uint8_t encryption_key[CRYPTO_KEY_SIZE];
68 uint8_t authentication_key[CRYPTO_HASH_SIZE];
71 void bluesky_crypt_init();
72 void bluesky_crypt_hash_key(const char *keystr, uint8_t *out);
73 void bluesky_crypt_random_bytes(guchar *buf, gint len);
74 void bluesky_crypt_derive_keys(BlueSkyCryptKeys *keys, const gchar *master);
75 BlueSkyRCStr *bluesky_crypt_encrypt(BlueSkyRCStr *in, const uint8_t *key);
76 BlueSkyRCStr *bluesky_crypt_decrypt(BlueSkyRCStr *in, const uint8_t *key);
78 void bluesky_crypt_block_encrypt(gchar *cloud_block, size_t len,
79 BlueSkyCryptKeys *keys);
80 gboolean bluesky_crypt_block_decrypt(gchar *cloud_block, size_t len,
81 BlueSkyCryptKeys *keys);
82 void bluesky_cloudlog_encrypt(GString *segment, BlueSkyCryptKeys *keys);
83 void bluesky_cloudlog_decrypt(char *segment, size_t len,
84 BlueSkyCryptKeys *keys,
85 BlueSkyRangeset *items);
87 /* Storage layer. Requests can be performed asynchronously, so these objects
88 * help keep track of operations in progress. */
94 STORE_OP_BARRIER, // Waits for other selected operations to complete
98 ASYNC_NEW, // Operation not yet submitted to storage layer
99 ASYNC_PENDING, // Submitted to storage layer
100 ASYNC_RUNNING, // Operation is in progress
101 ASYNC_COMPLETE, // Operation finished, results available
102 } BlueSkyAsyncStatus;
104 struct BlueSkyNotifierList;
105 typedef struct BlueSkyStoreAsync BlueSkyStoreAsync;
106 struct BlueSkyStoreAsync {
110 GCond *completion_cond; /* Used to wait for operation to complete. */
112 gint refcount; /* Reference count for destruction. */
114 BlueSkyAsyncStatus status;
117 gchar *key; /* Key to read/write */
118 BlueSkyRCStr *data; /* Data read/to write */
120 /* For range requests on reads: starting byte offset and length; len 0
121 * implies reading to the end of the object. At completion, the backend
122 * should set range_done if a range read was made; if not set the entire
123 * object was read and the storage layer will select out just the
124 * appropriate bytes. */
128 int result; /* Result code; 0 for success. */
129 struct BlueSkyNotifierList *notifiers;
132 /* The barrier waiting on this operation. Support for more than one
133 * barrier for a single async is not well-supported and should be avoided
135 BlueSkyStoreAsync *barrier;
137 bluesky_time_hires start_time; /* Time operation was submitted. */
138 bluesky_time_hires exec_time; /* Time processing started on operation. */
140 gpointer store_private; /* For use by the storage implementation */
143 /* Support for notification lists. These are lists of one-shot functions which
144 * can be called when certain events--primarily, competed storage
145 * events--occur. Multiple notifiers can be added, but no particular order is
146 * guaranteed for the notification functions to be called. */
147 struct BlueSkyNotifierList {
148 struct BlueSkyNotifierList *next;
150 BlueSkyStoreAsync *async;
151 gpointer user_data; // Passed to the function when called
154 /* The abstraction layer for storage, allowing multiple implementations. */
156 /* Create a new store instance and return a handle to it. */
157 gpointer (*create)(const gchar *path);
159 /* Clean up any resources used by this store. */
160 void (*destroy)(gpointer store);
162 /* Submit an operation (get/put/delete) to the storage layer to be
163 * performed asynchronously. */
164 void (*submit)(gpointer store, BlueSkyStoreAsync *async);
166 /* Clean up any implementation-private data in a BlueSkyStoreAsync. */
167 void (*cleanup)(gpointer store, BlueSkyStoreAsync *async);
169 /* Find the lexicographically-largest file starting with the specified
171 char * (*lookup_last)(gpointer store, const gchar *prefix);
172 } BlueSkyStoreImplementation;
174 void bluesky_store_register(const BlueSkyStoreImplementation *impl,
177 char *bluesky_store_lookup_last(BlueSkyStore *store, const char *prefix);
178 BlueSkyStoreAsync *bluesky_store_async_new(BlueSkyStore *store);
179 gpointer bluesky_store_async_get_handle(BlueSkyStoreAsync *async);
180 void bluesky_store_async_ref(BlueSkyStoreAsync *async);
181 void bluesky_store_async_unref(BlueSkyStoreAsync *async);
182 void bluesky_store_async_wait(BlueSkyStoreAsync *async);
183 void bluesky_store_async_add_notifier(BlueSkyStoreAsync *async,
184 GFunc func, gpointer user_data);
185 void bluesky_store_async_mark_complete(BlueSkyStoreAsync *async);
186 void bluesky_store_async_submit(BlueSkyStoreAsync *async);
187 void bluesky_store_sync(BlueSkyStore *store);
189 void bluesky_store_add_barrier(BlueSkyStoreAsync *barrier,
190 BlueSkyStoreAsync *async);
192 void bluesky_inode_start_sync(BlueSkyInode *inode);
194 void bluesky_block_touch(BlueSkyInode *inode, uint64_t i);
195 void bluesky_block_fetch(BlueSkyInode *inode, BlueSkyBlock *block,
196 BlueSkyStoreAsync *barrier);
197 void bluesky_block_flush(BlueSkyInode *inode, BlueSkyBlock *block,
199 void bluesky_file_flush(BlueSkyInode *inode, GList **log_items);
200 void bluesky_file_drop_cached(BlueSkyInode *inode);
202 /* Writing of data to the cloud in log segments and tracking the location of
203 * various pieces of data (both where in the cloud and where cached locally).
215 } BlueSkyCloudPointer;
218 LOGTYPE_INVALID = -1,
222 LOGTYPE_INODE_MAP = 3,
223 LOGTYPE_CHECKPOINT = 4,
225 /* Used only as metadata in the local journal, not loaded as a
226 * BlueSkyCloudLogState nor stored in the cloud */
227 LOGTYPE_JOURNAL_MARKER = 16,
228 LOGTYPE_JOURNAL_CHECKPOINT = 17,
229 } BlueSkyCloudLogType;
231 /* Headers that go on items in local log segments and cloud log segments. */
233 uint32_t magic; // HEADER_MAGIC
234 uint8_t type; // Object type + '0'
235 uint32_t offset; // Starting byte offset of the log header
236 uint32_t size1; // Size of the data item (bytes)
239 uint64_t inum; // Inode which owns this data, if any
240 BlueSkyCloudID id; // Object identifier
241 } __attribute__((packed));
244 uint32_t magic; // FOOTER_MAGIC
245 uint32_t crc; // Computed from log_header to log_footer.magic
246 } __attribute__((packed));
248 struct cloudlog_header {
250 uint8_t crypt_auth[CRYPTO_HASH_SIZE];
251 uint8_t crypt_iv[CRYPTO_BLOCK_SIZE];
255 uint32_t size1, size2, size3;
256 } __attribute__((packed));
258 #define JOURNAL_MAGIC "\nLog"
259 #define CLOUDLOG_MAGIC "AgI-"
260 #define CLOUDLOG_MAGIC_ENCRYPTED "AgI=" // CLOUDLOG_MAGIC[3] ^= 0x10
262 /* A record which tracks an object which has been written to a local log,
263 * cached, locally, and/or written to the cloud. */
264 #define CLOUDLOG_JOURNAL 0x01
265 #define CLOUDLOG_CLOUD 0x02
266 #define CLOUDLOG_CACHE 0x04
267 #define CLOUDLOG_UNCOMMITTED 0x10
268 struct BlueSkyCloudLog {
275 BlueSkyCloudLogType type;
277 // Bitmask of CLOUDLOG_* flags indicating where the object exists.
279 int pending_read, pending_write;
281 // A stable identifier for the object (only changes when authenticated data
282 // is written out, but stays the same when the in-cloud cleaner relocates
286 // The inode which owns this data, if any, and an offset.
290 // The size of encrypted object data, not including any headers
293 // The location of the object in the cloud, if available.
294 BlueSkyCloudPointer location;
296 // TODO: Location in journal/cache
297 int log_seq, log_offset, log_size;
299 // Pointers to other objects. Each link counts towards the reference count
300 // of the pointed-to object. To avoid memory leaks there should be no
301 // cycles in the reference graph.
304 // Serialized data, if available in memory (otherwise NULL), and a lock
305 // count which tracks if there are users that require the data to be kept
311 /* Serialize objects into a log segment to be written to the cloud. */
312 struct BlueSkyCloudLogState {
314 BlueSkyCloudPointer location;
316 GSList *writeback_list; // Items which are being serialized right now
317 GList *pending_segments; // Segments which are being uploaded now
320 gboolean bluesky_cloudlog_equal(gconstpointer a, gconstpointer b);
321 guint bluesky_cloudlog_hash(gconstpointer a);
322 BlueSkyCloudLog *bluesky_cloudlog_new(BlueSkyFS *fs, const BlueSkyCloudID *id);
323 gchar *bluesky_cloudlog_id_to_string(BlueSkyCloudID id);
324 BlueSkyCloudID bluesky_cloudlog_id_from_string(const gchar *idstr);
325 void bluesky_cloudlog_threads_init(BlueSkyFS *fs);
326 void bluesky_cloudlog_ref(BlueSkyCloudLog *log);
327 void bluesky_cloudlog_unref(BlueSkyCloudLog *log);
328 void bluesky_cloudlog_unref_delayed(BlueSkyCloudLog *log);
329 void bluesky_cloudlog_erase(BlueSkyCloudLog *log);
330 void bluesky_cloudlog_stats_update(BlueSkyCloudLog *log, int type);
331 void bluesky_cloudlog_sync(BlueSkyCloudLog *log);
332 void bluesky_cloudlog_insert(BlueSkyCloudLog *log);
333 void bluesky_cloudlog_insert_locked(BlueSkyCloudLog *log);
334 BlueSkyCloudLog *bluesky_cloudlog_get(BlueSkyFS *fs, BlueSkyCloudID id);
335 void bluesky_cloudlog_prefetch(BlueSkyCloudLog *log);
336 void bluesky_cloudlog_fetch(BlueSkyCloudLog *log);
337 BlueSkyCloudPointer bluesky_cloudlog_serialize(BlueSkyCloudLog *log,
339 void bluesky_cloudlog_flush(BlueSkyFS *fs);
341 /* Logging infrastructure for ensuring operations are persistently recorded to
343 #define BLUESKY_CRC32C_SEED (~(uint32_t)0)
344 #define BLUESKY_CRC32C_VALIDATOR ((uint32_t)0xb798b438UL)
345 uint32_t crc32c(uint32_t crc, const char *buf, unsigned int length);
346 uint32_t crc32c_finalize(uint32_t crc);
356 /* The currently-open log file. */
357 BlueSkyCacheFile *current_log;
359 /* Cache of log segments which have been memory-mapped. */
361 GHashTable *mmap_cache;
363 /* A count of the disk space consumed (in 1024-byte units) by all files
364 * tracked by mmap_cache (whether mapped or not, actually). */
367 /* The smallest journal sequence number which may still contain data that
368 * must be preserved (since it it not yet in the cloud). */
369 int journal_watermark;
372 /* An object for tracking log files which are stored locally--either the
373 * journal for filesystem consistency or log segments which have been fetched
374 * back from cloud storage. */
375 struct BlueSkyCacheFile {
379 int type; // Only one of CLOUDLOG_{JOURNAL,CLOUD}
382 char *filename; // Local filename, relateive to log directory
383 gint mapcount; // References to the mmaped data
384 const char *addr; // May be null if data is not mapped in memory
389 gboolean fetching, ready; // Cloud data: downloading or ready for use
390 int64_t atime; // Access time, for cache management
391 BlueSkyRangeset *items; // Locations of valid items
394 BlueSkyLog *bluesky_log_new(const char *log_directory);
395 void bluesky_log_item_submit(BlueSkyCloudLog *item, BlueSkyLog *log);
396 void bluesky_log_finish_all(GList *log_items);
397 BlueSkyCloudLog *bluesky_log_get_commit_point(BlueSkyFS *fs);
398 void bluesky_log_write_commit_point(BlueSkyFS *fs, BlueSkyCloudLog *marker);
400 BlueSkyRCStr *bluesky_log_map_object(BlueSkyCloudLog *item, gboolean map_data);
401 void bluesky_mmap_unref(BlueSkyCacheFile *mmap);
402 void bluesky_cachefile_unref(BlueSkyCacheFile *cachefile);
404 BlueSkyCacheFile *bluesky_cachefile_lookup(BlueSkyFS *fs,
405 int clouddir, int log_seq,
406 gboolean start_fetch);
407 void bluesky_cachefile_gc(BlueSkyFS *fs);
409 void bluesky_replay(BlueSkyFS *fs);
411 /* Used to track log segments that are being written to the cloud. */
420 /***** Inode map management *****/
422 /* Mapping information for a single inode number. These are grouped together
423 * into InodeMapRange objects. */
427 /* A pointer to the cloud log entry for this inode. This may or may not
428 * actually have data loaded (it might just contain pointers to the data
429 * location, and in fact this will likely often be the case). */
430 BlueSkyCloudLog *item;
434 /* Starting and ending inode number values that fall in this section.
435 * Endpoint values are inclusive. */
438 /* A sorted list (by inode number) of InodeMapEntry objects. */
439 GSequence *map_entries;
441 /* The serialized version of the inode map data. */
442 BlueSkyCloudLog *serialized;
444 /* Have there been changes that require writing this section out again? */
448 InodeMapEntry *bluesky_inode_map_lookup(GSequence *inode_map, uint64_t inum,
450 BlueSkyCloudLog *bluesky_inode_map_serialize(BlueSkyFS *fs);
451 void bluesky_inode_map_minimize(BlueSkyFS *fs);
453 gboolean bluesky_checkpoint_load(BlueSkyFS *fs);