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
23 extern int bluesky_verbose;
25 /* Target cache size levels. */
26 extern int bluesky_watermark_low_dirty;
27 extern int bluesky_watermark_medium_dirty;
28 extern int bluesky_watermark_med2_dirty;
29 extern int bluesky_watermark_high_dirty;
31 extern int bluesky_watermark_low_total;
32 extern int bluesky_watermark_medium_total;
33 extern int bluesky_watermark_med2_total;
34 extern int bluesky_watermark_high_total;
36 /* TODO: Make this go away entirely. */
37 BlueSkyFS *bluesky_new_fs(gchar *name);
39 void bluesky_inode_free_resources(BlueSkyInode *inode);
41 /* Linked list update functions for LRU lists. */
42 void bluesky_list_unlink(GList *head, GList *item);
43 GList *bluesky_list_prepend(GList *head, BlueSkyInode *inode);
44 GList *bluesky_list_append(GList *head, BlueSkyInode *inode);
45 BlueSkyInode *bluesky_list_head(GList *head);
46 BlueSkyInode *bluesky_list_tail(GList *head);
48 /* Serialization and deserialization of filesystem data for storing to
49 * persistent storage. */
50 void bluesky_serialize_superblock(GString *out, BlueSkyFS *fs);
51 BlueSkyFS *bluesky_deserialize_superblock(const gchar *buf);
52 BlueSkyCloudLog *bluesky_serialize_inode(BlueSkyInode *inode);
53 gboolean bluesky_deserialize_inode(BlueSkyInode *inode, BlueSkyCloudLog *item);
55 void bluesky_deserialize_cloudlog(BlueSkyCloudLog *item,
59 void bluesky_serialize_cloudlog(BlueSkyCloudLog *log,
61 GString *authenticated,
64 /* Cryptographic operations. */
65 #define CRYPTO_BLOCK_SIZE 16 /* 128-bit AES */
66 #define CRYPTO_KEY_SIZE 16
67 #define CRYPTO_HASH_SIZE 32 /* SHA-256 */
69 typedef struct BlueSkyCryptKeys {
70 uint8_t encryption_key[CRYPTO_KEY_SIZE];
71 uint8_t authentication_key[CRYPTO_HASH_SIZE];
74 void bluesky_crypt_init();
75 void bluesky_crypt_hash_key(const char *keystr, uint8_t *out);
76 void bluesky_crypt_random_bytes(guchar *buf, gint len);
77 void bluesky_crypt_derive_keys(BlueSkyCryptKeys *keys, const gchar *master);
78 BlueSkyRCStr *bluesky_crypt_encrypt(BlueSkyRCStr *in, const uint8_t *key);
79 BlueSkyRCStr *bluesky_crypt_decrypt(BlueSkyRCStr *in, const uint8_t *key);
81 void bluesky_crypt_block_encrypt(gchar *cloud_block, size_t len,
82 BlueSkyCryptKeys *keys);
83 gboolean bluesky_crypt_block_decrypt(gchar *cloud_block, size_t len,
84 BlueSkyCryptKeys *keys,
85 gboolean allow_unauth);
86 void bluesky_cloudlog_encrypt(GString *segment, BlueSkyCryptKeys *keys);
87 void bluesky_cloudlog_decrypt(char *segment, size_t len,
88 BlueSkyCryptKeys *keys,
89 BlueSkyRangeset *items,
90 gboolean allow_unauth);
92 /* Storage layer. Requests can be performed asynchronously, so these objects
93 * help keep track of operations in progress. */
99 STORE_OP_BARRIER, // Waits for other selected operations to complete
103 ASYNC_NEW, // Operation not yet submitted to storage layer
104 ASYNC_PENDING, // Submitted to storage layer
105 ASYNC_RUNNING, // Operation is in progress
106 ASYNC_COMPLETE, // Operation finished, results available
107 } BlueSkyAsyncStatus;
109 struct BlueSkyNotifierList;
110 typedef struct BlueSkyStoreAsync BlueSkyStoreAsync;
111 struct BlueSkyStoreAsync {
115 GCond *completion_cond; /* Used to wait for operation to complete. */
117 gint refcount; /* Reference count for destruction. */
119 BlueSkyAsyncStatus status;
122 gchar *key; /* Key to read/write */
123 BlueSkyRCStr *data; /* Data read/to write */
125 /* For range requests on reads: starting byte offset and length; len 0
126 * implies reading to the end of the object. At completion, the backend
127 * should set range_done if a range read was made; if not set the entire
128 * object was read and the storage layer will select out just the
129 * appropriate bytes. */
133 int result; /* Result code; 0 for success. */
134 struct BlueSkyNotifierList *notifiers;
137 /* The barrier waiting on this operation. Support for more than one
138 * barrier for a single async is not well-supported and should be avoided
140 BlueSkyStoreAsync *barrier;
142 bluesky_time_hires start_time; /* Time operation was submitted. */
143 bluesky_time_hires exec_time; /* Time processing started on operation. */
145 gpointer store_private; /* For use by the storage implementation */
147 /* If storage operations should be charged to any particular profile, which
149 BlueSkyProfile *profile;
152 /* Support for notification lists. These are lists of one-shot functions which
153 * can be called when certain events--primarily, competed storage
154 * events--occur. Multiple notifiers can be added, but no particular order is
155 * guaranteed for the notification functions to be called. */
156 struct BlueSkyNotifierList {
157 struct BlueSkyNotifierList *next;
159 BlueSkyStoreAsync *async;
160 gpointer user_data; // Passed to the function when called
163 /* The abstraction layer for storage, allowing multiple implementations. */
165 /* Create a new store instance and return a handle to it. */
166 gpointer (*create)(const gchar *path);
168 /* Clean up any resources used by this store. */
169 void (*destroy)(gpointer store);
171 /* Submit an operation (get/put/delete) to the storage layer to be
172 * performed asynchronously. */
173 void (*submit)(gpointer store, BlueSkyStoreAsync *async);
175 /* Clean up any implementation-private data in a BlueSkyStoreAsync. */
176 void (*cleanup)(gpointer store, BlueSkyStoreAsync *async);
178 /* Find the lexicographically-largest file starting with the specified
180 char * (*lookup_last)(gpointer store, const gchar *prefix);
181 } BlueSkyStoreImplementation;
183 void bluesky_store_register(const BlueSkyStoreImplementation *impl,
186 char *bluesky_store_lookup_last(BlueSkyStore *store, const char *prefix);
187 BlueSkyStoreAsync *bluesky_store_async_new(BlueSkyStore *store);
188 gpointer bluesky_store_async_get_handle(BlueSkyStoreAsync *async);
189 void bluesky_store_async_ref(BlueSkyStoreAsync *async);
190 void bluesky_store_async_unref(BlueSkyStoreAsync *async);
191 void bluesky_store_async_wait(BlueSkyStoreAsync *async);
192 void bluesky_store_async_add_notifier(BlueSkyStoreAsync *async,
193 GFunc func, gpointer user_data);
194 void bluesky_store_async_mark_complete(BlueSkyStoreAsync *async);
195 void bluesky_store_async_submit(BlueSkyStoreAsync *async);
196 void bluesky_store_sync(BlueSkyStore *store);
198 void bluesky_store_add_barrier(BlueSkyStoreAsync *barrier,
199 BlueSkyStoreAsync *async);
201 void bluesky_inode_start_sync(BlueSkyInode *inode);
203 void bluesky_block_touch(BlueSkyInode *inode, uint64_t i, gboolean preserve);
204 void bluesky_block_fetch(BlueSkyInode *inode, BlueSkyBlock *block,
205 BlueSkyStoreAsync *barrier);
206 void bluesky_block_flush(BlueSkyInode *inode, BlueSkyBlock *block,
208 void bluesky_file_flush(BlueSkyInode *inode, GList **log_items);
209 void bluesky_file_drop_cached(BlueSkyInode *inode);
211 /* Writing of data to the cloud in log segments and tracking the location of
212 * various pieces of data (both where in the cloud and where cached locally).
215 /* Eventually we'll want to support multiple writers. But for now, hard-code
216 * separate namespaces in the cloud for the proxy and the cleaner to write to.
218 #define BLUESKY_CLOUD_DIR_PRIMARY 0
219 #define BLUESKY_CLOUD_DIR_CLEANER 1
230 } BlueSkyCloudPointer;
233 LOGTYPE_INVALID = -1,
237 LOGTYPE_INODE_MAP = 3,
238 LOGTYPE_CHECKPOINT = 4,
240 /* Used only as metadata in the local journal, not loaded as a
241 * BlueSkyCloudLogState nor stored in the cloud */
242 LOGTYPE_JOURNAL_MARKER = 16,
243 LOGTYPE_JOURNAL_CHECKPOINT = 17,
244 } BlueSkyCloudLogType;
246 /* Headers that go on items in local log segments and cloud log segments. */
248 uint32_t magic; // HEADER_MAGIC
249 uint8_t type; // Object type + '0'
250 uint32_t offset; // Starting byte offset of the log header
251 uint32_t size1; // Size of the data item (bytes)
254 uint64_t inum; // Inode which owns this data, if any
255 BlueSkyCloudID id; // Object identifier
256 } __attribute__((packed));
259 uint32_t magic; // FOOTER_MAGIC
260 uint32_t crc; // Computed from log_header to log_footer.magic
261 } __attribute__((packed));
263 struct cloudlog_header {
265 uint8_t crypt_auth[CRYPTO_HASH_SIZE];
266 uint8_t crypt_iv[CRYPTO_BLOCK_SIZE];
270 uint32_t size1, size2, size3;
271 } __attribute__((packed));
273 // Rough size limit for a log segment. This is not a firm limit and there are
274 // no absolute guarantees on the size of a log segment.
275 #define LOG_SEGMENT_SIZE (1 << 22)
277 #define JOURNAL_MAGIC "\nLog"
278 #define CLOUDLOG_MAGIC "AgI-"
279 #define CLOUDLOG_MAGIC_ENCRYPTED "AgI=" // CLOUDLOG_MAGIC[3] ^= 0x10
281 /* A record which tracks an object which has been written to a local log,
282 * cached, locally, and/or written to the cloud. */
283 #define CLOUDLOG_JOURNAL 0x01
284 #define CLOUDLOG_CLOUD 0x02
285 #define CLOUDLOG_CACHE 0x04
286 #define CLOUDLOG_UNCOMMITTED 0x10
287 struct BlueSkyCloudLog {
294 BlueSkyCloudLogType type;
296 // Bitmask of CLOUDLOG_* flags indicating where the object exists.
298 int pending_read, pending_write;
300 // A stable identifier for the object (only changes when authenticated data
301 // is written out, but stays the same when the in-cloud cleaner relocates
305 // The inode which owns this data, if any, and an offset.
309 // The size of encrypted object data, not including any headers
312 // The location of the object in the cloud, if available.
313 BlueSkyCloudPointer location;
315 // TODO: Location in journal/cache
316 int log_seq, log_offset, log_size;
318 // Pointers to other objects. Each link counts towards the reference count
319 // of the pointed-to object. To avoid memory leaks there should be no
320 // cycles in the reference graph.
323 // Serialized data, if available in memory (otherwise NULL), and a lock
324 // count which tracks if there are users that require the data to be kept
330 /* Serialize objects into a log segment to be written to the cloud. */
331 struct BlueSkyCloudLogState {
333 BlueSkyCloudPointer location;
335 GSList *writeback_list; // Items which are being serialized right now
336 GList *pending_segments; // Segments which are being uploaded now
338 int uploads_pending; // Count of uploads in progress, not completed
339 GMutex *uploads_pending_lock;
340 GCond *uploads_pending_cond;
342 /* What is the most recent sequence number written by the cleaner which we
343 * have processed and incorporated into our own log? This gets
344 * incorporated into the version vector written out with our checkpoint
346 int latest_cleaner_seq_seen;
349 gboolean bluesky_cloudlog_equal(gconstpointer a, gconstpointer b);
350 guint bluesky_cloudlog_hash(gconstpointer a);
351 BlueSkyCloudLog *bluesky_cloudlog_new(BlueSkyFS *fs, const BlueSkyCloudID *id);
352 gchar *bluesky_cloudlog_id_to_string(BlueSkyCloudID id);
353 BlueSkyCloudID bluesky_cloudlog_id_from_string(const gchar *idstr);
354 void bluesky_cloudlog_threads_init(BlueSkyFS *fs);
355 void bluesky_cloudlog_ref(BlueSkyCloudLog *log);
356 void bluesky_cloudlog_unref(BlueSkyCloudLog *log);
357 void bluesky_cloudlog_unref_delayed(BlueSkyCloudLog *log);
358 void bluesky_cloudlog_erase(BlueSkyCloudLog *log);
359 void bluesky_cloudlog_stats_update(BlueSkyCloudLog *log, int type);
360 void bluesky_cloudlog_sync(BlueSkyCloudLog *log);
361 void bluesky_cloudlog_insert(BlueSkyCloudLog *log);
362 void bluesky_cloudlog_insert_locked(BlueSkyCloudLog *log);
363 BlueSkyCloudLog *bluesky_cloudlog_get(BlueSkyFS *fs, BlueSkyCloudID id);
364 void bluesky_cloudlog_prefetch(BlueSkyCloudLog *log);
365 void bluesky_cloudlog_fetch(BlueSkyCloudLog *log);
366 void bluesky_cloudlog_background_fetch(BlueSkyCloudLog *item);
367 BlueSkyCloudPointer bluesky_cloudlog_serialize(BlueSkyCloudLog *log,
369 void bluesky_cloudlog_flush(BlueSkyFS *fs);
371 /* Logging infrastructure for ensuring operations are persistently recorded to
373 #define BLUESKY_CRC32C_SEED (~(uint32_t)0)
374 #define BLUESKY_CRC32C_VALIDATOR ((uint32_t)0xb798b438UL)
375 uint32_t crc32c(uint32_t crc, const char *buf, unsigned int length);
376 uint32_t crc32c_finalize(uint32_t crc);
386 /* The currently-open log file. */
387 BlueSkyCacheFile *current_log;
389 /* Cache of log segments which have been memory-mapped. */
391 GHashTable *mmap_cache;
393 /* A count of the disk space consumed (in 1024-byte units) by all files
394 * tracked by mmap_cache (whether mapped or not, actually). */
397 /* The smallest journal sequence number which may still contain data that
398 * must be preserved (since it it not yet in the cloud). */
399 int journal_watermark;
402 /* An object for tracking log files which are stored locally--either the
403 * journal for filesystem consistency or log segments which have been fetched
404 * back from cloud storage. */
405 struct BlueSkyCacheFile {
409 int type; // Only one of CLOUDLOG_{JOURNAL,CLOUD}
412 char *filename; // Local filename, relateive to log directory
413 gint mapcount; // References to the mmaped data
414 const char *addr; // May be null if data is not mapped in memory
419 gboolean fetching; // Cloud data: downloading or ready for use
420 gboolean complete; // Complete file has been fetched from cloud
421 int64_t atime; // Access time, for cache management
422 BlueSkyRangeset *items; // Locations of valid items
423 BlueSkyRangeset *prefetches;// Locations we have been requested to prefetch
426 BlueSkyLog *bluesky_log_new(const char *log_directory);
427 void bluesky_log_item_submit(BlueSkyCloudLog *item, BlueSkyLog *log);
428 void bluesky_log_finish_all(GList *log_items);
429 BlueSkyCloudLog *bluesky_log_get_commit_point(BlueSkyFS *fs);
430 void bluesky_log_write_commit_point(BlueSkyFS *fs, BlueSkyCloudLog *marker);
432 BlueSkyRCStr *bluesky_cachefile_map_raw(BlueSkyCacheFile *cachefile,
433 off_t offset, size_t size);
434 BlueSkyRCStr *bluesky_log_map_object(BlueSkyCloudLog *item, gboolean map_data);
435 void bluesky_mmap_unref(BlueSkyCacheFile *mmap);
436 void bluesky_cachefile_unref(BlueSkyCacheFile *cachefile);
438 BlueSkyCacheFile *bluesky_cachefile_lookup(BlueSkyFS *fs,
439 int clouddir, int log_seq,
440 gboolean start_fetch);
441 void bluesky_cachefile_gc(BlueSkyFS *fs);
443 void bluesky_replay(BlueSkyFS *fs);
445 /* Used to track log segments that are being written to the cloud. */
448 char *key; /* File name for log segment in backend */
449 GString *raw_data; /* Data before encryption */
450 BlueSkyRCStr *data; /* Data after encryption */
457 /***** Inode map management *****/
459 /* Mapping information for a single inode number. These are grouped together
460 * into InodeMapRange objects. */
464 /* A pointer to the cloud log entry for this inode. This may or may not
465 * actually have data loaded (it might just contain pointers to the data
466 * location, and in fact this will likely often be the case). */
467 BlueSkyCloudLog *item;
471 /* Starting and ending inode number values that fall in this section.
472 * Endpoint values are inclusive. */
475 /* A sorted list (by inode number) of InodeMapEntry objects. */
476 GSequence *map_entries;
478 /* The serialized version of the inode map data. */
479 BlueSkyCloudLog *serialized;
481 /* Have there been changes that require writing this section out again? */
485 InodeMapEntry *bluesky_inode_map_lookup(GSequence *inode_map, uint64_t inum,
487 BlueSkyCloudLog *bluesky_inode_map_serialize(BlueSkyFS *fs);
488 void bluesky_inode_map_minimize(BlueSkyFS *fs);
490 gboolean bluesky_checkpoint_load(BlueSkyFS *fs);
492 /* Merging of log state with the work of the cleaner. */
493 void bluesky_cleaner_merge(BlueSkyFS *fs);
494 void bluesky_cleaner_thread_launch(BlueSkyFS *fs);