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 gboolean allow_unauth);
83 void bluesky_cloudlog_encrypt(GString *segment, BlueSkyCryptKeys *keys);
84 void bluesky_cloudlog_decrypt(char *segment, size_t len,
85 BlueSkyCryptKeys *keys,
86 BlueSkyRangeset *items,
87 gboolean allow_unauth);
89 /* Storage layer. Requests can be performed asynchronously, so these objects
90 * help keep track of operations in progress. */
96 STORE_OP_BARRIER, // Waits for other selected operations to complete
100 ASYNC_NEW, // Operation not yet submitted to storage layer
101 ASYNC_PENDING, // Submitted to storage layer
102 ASYNC_RUNNING, // Operation is in progress
103 ASYNC_COMPLETE, // Operation finished, results available
104 } BlueSkyAsyncStatus;
106 struct BlueSkyNotifierList;
107 typedef struct BlueSkyStoreAsync BlueSkyStoreAsync;
108 struct BlueSkyStoreAsync {
112 GCond *completion_cond; /* Used to wait for operation to complete. */
114 gint refcount; /* Reference count for destruction. */
116 BlueSkyAsyncStatus status;
119 gchar *key; /* Key to read/write */
120 BlueSkyRCStr *data; /* Data read/to write */
122 /* For range requests on reads: starting byte offset and length; len 0
123 * implies reading to the end of the object. At completion, the backend
124 * should set range_done if a range read was made; if not set the entire
125 * object was read and the storage layer will select out just the
126 * appropriate bytes. */
130 int result; /* Result code; 0 for success. */
131 struct BlueSkyNotifierList *notifiers;
134 /* The barrier waiting on this operation. Support for more than one
135 * barrier for a single async is not well-supported and should be avoided
137 BlueSkyStoreAsync *barrier;
139 bluesky_time_hires start_time; /* Time operation was submitted. */
140 bluesky_time_hires exec_time; /* Time processing started on operation. */
142 gpointer store_private; /* For use by the storage implementation */
144 /* If storage operations should be charged to any particular profile, which
146 BlueSkyProfile *profile;
149 /* Support for notification lists. These are lists of one-shot functions which
150 * can be called when certain events--primarily, competed storage
151 * events--occur. Multiple notifiers can be added, but no particular order is
152 * guaranteed for the notification functions to be called. */
153 struct BlueSkyNotifierList {
154 struct BlueSkyNotifierList *next;
156 BlueSkyStoreAsync *async;
157 gpointer user_data; // Passed to the function when called
160 /* The abstraction layer for storage, allowing multiple implementations. */
162 /* Create a new store instance and return a handle to it. */
163 gpointer (*create)(const gchar *path);
165 /* Clean up any resources used by this store. */
166 void (*destroy)(gpointer store);
168 /* Submit an operation (get/put/delete) to the storage layer to be
169 * performed asynchronously. */
170 void (*submit)(gpointer store, BlueSkyStoreAsync *async);
172 /* Clean up any implementation-private data in a BlueSkyStoreAsync. */
173 void (*cleanup)(gpointer store, BlueSkyStoreAsync *async);
175 /* Find the lexicographically-largest file starting with the specified
177 char * (*lookup_last)(gpointer store, const gchar *prefix);
178 } BlueSkyStoreImplementation;
180 void bluesky_store_register(const BlueSkyStoreImplementation *impl,
183 char *bluesky_store_lookup_last(BlueSkyStore *store, const char *prefix);
184 BlueSkyStoreAsync *bluesky_store_async_new(BlueSkyStore *store);
185 gpointer bluesky_store_async_get_handle(BlueSkyStoreAsync *async);
186 void bluesky_store_async_ref(BlueSkyStoreAsync *async);
187 void bluesky_store_async_unref(BlueSkyStoreAsync *async);
188 void bluesky_store_async_wait(BlueSkyStoreAsync *async);
189 void bluesky_store_async_add_notifier(BlueSkyStoreAsync *async,
190 GFunc func, gpointer user_data);
191 void bluesky_store_async_mark_complete(BlueSkyStoreAsync *async);
192 void bluesky_store_async_submit(BlueSkyStoreAsync *async);
193 void bluesky_store_sync(BlueSkyStore *store);
195 void bluesky_store_add_barrier(BlueSkyStoreAsync *barrier,
196 BlueSkyStoreAsync *async);
198 void bluesky_inode_start_sync(BlueSkyInode *inode);
200 void bluesky_block_touch(BlueSkyInode *inode, uint64_t i);
201 void bluesky_block_fetch(BlueSkyInode *inode, BlueSkyBlock *block,
202 BlueSkyStoreAsync *barrier);
203 void bluesky_block_flush(BlueSkyInode *inode, BlueSkyBlock *block,
205 void bluesky_file_flush(BlueSkyInode *inode, GList **log_items);
206 void bluesky_file_drop_cached(BlueSkyInode *inode);
208 /* Writing of data to the cloud in log segments and tracking the location of
209 * various pieces of data (both where in the cloud and where cached locally).
212 /* Eventually we'll want to support multiple writers. But for now, hard-code
213 * separate namespaces in the cloud for the proxy and the cleaner to write to.
215 #define BLUESKY_CLOUD_DIR_PRIMARY 0
216 #define BLUESKY_CLOUD_DIR_CLEANER 1
227 } BlueSkyCloudPointer;
230 LOGTYPE_INVALID = -1,
234 LOGTYPE_INODE_MAP = 3,
235 LOGTYPE_CHECKPOINT = 4,
237 /* Used only as metadata in the local journal, not loaded as a
238 * BlueSkyCloudLogState nor stored in the cloud */
239 LOGTYPE_JOURNAL_MARKER = 16,
240 LOGTYPE_JOURNAL_CHECKPOINT = 17,
241 } BlueSkyCloudLogType;
243 /* Headers that go on items in local log segments and cloud log segments. */
245 uint32_t magic; // HEADER_MAGIC
246 uint8_t type; // Object type + '0'
247 uint32_t offset; // Starting byte offset of the log header
248 uint32_t size1; // Size of the data item (bytes)
251 uint64_t inum; // Inode which owns this data, if any
252 BlueSkyCloudID id; // Object identifier
253 } __attribute__((packed));
256 uint32_t magic; // FOOTER_MAGIC
257 uint32_t crc; // Computed from log_header to log_footer.magic
258 } __attribute__((packed));
260 struct cloudlog_header {
262 uint8_t crypt_auth[CRYPTO_HASH_SIZE];
263 uint8_t crypt_iv[CRYPTO_BLOCK_SIZE];
267 uint32_t size1, size2, size3;
268 } __attribute__((packed));
270 #define JOURNAL_MAGIC "\nLog"
271 #define CLOUDLOG_MAGIC "AgI-"
272 #define CLOUDLOG_MAGIC_ENCRYPTED "AgI=" // CLOUDLOG_MAGIC[3] ^= 0x10
274 /* A record which tracks an object which has been written to a local log,
275 * cached, locally, and/or written to the cloud. */
276 #define CLOUDLOG_JOURNAL 0x01
277 #define CLOUDLOG_CLOUD 0x02
278 #define CLOUDLOG_CACHE 0x04
279 #define CLOUDLOG_UNCOMMITTED 0x10
280 struct BlueSkyCloudLog {
287 BlueSkyCloudLogType type;
289 // Bitmask of CLOUDLOG_* flags indicating where the object exists.
291 int pending_read, pending_write;
293 // A stable identifier for the object (only changes when authenticated data
294 // is written out, but stays the same when the in-cloud cleaner relocates
298 // The inode which owns this data, if any, and an offset.
302 // The size of encrypted object data, not including any headers
305 // The location of the object in the cloud, if available.
306 BlueSkyCloudPointer location;
308 // TODO: Location in journal/cache
309 int log_seq, log_offset, log_size;
311 // Pointers to other objects. Each link counts towards the reference count
312 // of the pointed-to object. To avoid memory leaks there should be no
313 // cycles in the reference graph.
316 // Serialized data, if available in memory (otherwise NULL), and a lock
317 // count which tracks if there are users that require the data to be kept
323 /* Serialize objects into a log segment to be written to the cloud. */
324 struct BlueSkyCloudLogState {
326 BlueSkyCloudPointer location;
328 GSList *writeback_list; // Items which are being serialized right now
329 GList *pending_segments; // Segments which are being uploaded now
331 /* What is the most recent sequence number written by the cleaner which we
332 * have processed and incorporated into our own log? This gets
333 * incorporated into the version vector written out with our checkpoint
335 int latest_cleaner_seq_seen;
338 gboolean bluesky_cloudlog_equal(gconstpointer a, gconstpointer b);
339 guint bluesky_cloudlog_hash(gconstpointer a);
340 BlueSkyCloudLog *bluesky_cloudlog_new(BlueSkyFS *fs, const BlueSkyCloudID *id);
341 gchar *bluesky_cloudlog_id_to_string(BlueSkyCloudID id);
342 BlueSkyCloudID bluesky_cloudlog_id_from_string(const gchar *idstr);
343 void bluesky_cloudlog_threads_init(BlueSkyFS *fs);
344 void bluesky_cloudlog_ref(BlueSkyCloudLog *log);
345 void bluesky_cloudlog_unref(BlueSkyCloudLog *log);
346 void bluesky_cloudlog_unref_delayed(BlueSkyCloudLog *log);
347 void bluesky_cloudlog_erase(BlueSkyCloudLog *log);
348 void bluesky_cloudlog_stats_update(BlueSkyCloudLog *log, int type);
349 void bluesky_cloudlog_sync(BlueSkyCloudLog *log);
350 void bluesky_cloudlog_insert(BlueSkyCloudLog *log);
351 void bluesky_cloudlog_insert_locked(BlueSkyCloudLog *log);
352 BlueSkyCloudLog *bluesky_cloudlog_get(BlueSkyFS *fs, BlueSkyCloudID id);
353 void bluesky_cloudlog_prefetch(BlueSkyCloudLog *log);
354 void bluesky_cloudlog_fetch(BlueSkyCloudLog *log);
355 BlueSkyCloudPointer bluesky_cloudlog_serialize(BlueSkyCloudLog *log,
357 void bluesky_cloudlog_flush(BlueSkyFS *fs);
359 /* Logging infrastructure for ensuring operations are persistently recorded to
361 #define BLUESKY_CRC32C_SEED (~(uint32_t)0)
362 #define BLUESKY_CRC32C_VALIDATOR ((uint32_t)0xb798b438UL)
363 uint32_t crc32c(uint32_t crc, const char *buf, unsigned int length);
364 uint32_t crc32c_finalize(uint32_t crc);
374 /* The currently-open log file. */
375 BlueSkyCacheFile *current_log;
377 /* Cache of log segments which have been memory-mapped. */
379 GHashTable *mmap_cache;
381 /* A count of the disk space consumed (in 1024-byte units) by all files
382 * tracked by mmap_cache (whether mapped or not, actually). */
385 /* The smallest journal sequence number which may still contain data that
386 * must be preserved (since it it not yet in the cloud). */
387 int journal_watermark;
390 /* An object for tracking log files which are stored locally--either the
391 * journal for filesystem consistency or log segments which have been fetched
392 * back from cloud storage. */
393 struct BlueSkyCacheFile {
397 int type; // Only one of CLOUDLOG_{JOURNAL,CLOUD}
400 char *filename; // Local filename, relateive to log directory
401 gint mapcount; // References to the mmaped data
402 const char *addr; // May be null if data is not mapped in memory
407 gboolean fetching; // Cloud data: downloading or ready for use
408 gboolean complete; // Complete file has been fetched from cloud
409 int64_t atime; // Access time, for cache management
410 BlueSkyRangeset *items; // Locations of valid items
411 BlueSkyRangeset *prefetches;// Locations we have been requested to prefetch
414 BlueSkyLog *bluesky_log_new(const char *log_directory);
415 void bluesky_log_item_submit(BlueSkyCloudLog *item, BlueSkyLog *log);
416 void bluesky_log_finish_all(GList *log_items);
417 BlueSkyCloudLog *bluesky_log_get_commit_point(BlueSkyFS *fs);
418 void bluesky_log_write_commit_point(BlueSkyFS *fs, BlueSkyCloudLog *marker);
420 BlueSkyRCStr *bluesky_log_map_object(BlueSkyCloudLog *item, gboolean map_data);
421 void bluesky_mmap_unref(BlueSkyCacheFile *mmap);
422 void bluesky_cachefile_unref(BlueSkyCacheFile *cachefile);
424 BlueSkyCacheFile *bluesky_cachefile_lookup(BlueSkyFS *fs,
425 int clouddir, int log_seq,
426 gboolean start_fetch);
427 void bluesky_cachefile_gc(BlueSkyFS *fs);
429 void bluesky_replay(BlueSkyFS *fs);
431 /* Used to track log segments that are being written to the cloud. */
440 /***** Inode map management *****/
442 /* Mapping information for a single inode number. These are grouped together
443 * into InodeMapRange objects. */
447 /* A pointer to the cloud log entry for this inode. This may or may not
448 * actually have data loaded (it might just contain pointers to the data
449 * location, and in fact this will likely often be the case). */
450 BlueSkyCloudLog *item;
454 /* Starting and ending inode number values that fall in this section.
455 * Endpoint values are inclusive. */
458 /* A sorted list (by inode number) of InodeMapEntry objects. */
459 GSequence *map_entries;
461 /* The serialized version of the inode map data. */
462 BlueSkyCloudLog *serialized;
464 /* Have there been changes that require writing this section out again? */
468 InodeMapEntry *bluesky_inode_map_lookup(GSequence *inode_map, uint64_t inum,
470 BlueSkyCloudLog *bluesky_inode_map_serialize(BlueSkyFS *fs);
471 void bluesky_inode_map_minimize(BlueSkyFS *fs);
473 gboolean bluesky_checkpoint_load(BlueSkyFS *fs);
475 /* Merging of log state with the work of the cleaner. */
476 void bluesky_cleaner_find_checkpoint(BlueSkyFS *fs);