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_serialize_cloudlog(BlueSkyCloudLog *log,
54 GString *authenticated,
57 /* Storage layer. Requests can be performed asynchronously, so these objects
58 * help keep track of operations in progress. */
64 STORE_OP_BARRIER, // Waits for other selected operations to complete
68 ASYNC_NEW, // Operation not yet submitted to storage layer
69 ASYNC_PENDING, // Submitted to storage layer
70 ASYNC_RUNNING, // Operation is in progress
71 ASYNC_COMPLETE, // Operation finished, results available
74 struct BlueSkyNotifierList;
75 typedef struct _BlueSkyStoreAsync BlueSkyStoreAsync;
76 struct _BlueSkyStoreAsync {
80 GCond *completion_cond; /* Used to wait for operation to complete. */
82 gint refcount; /* Reference count for destruction. */
84 BlueSkyAsyncStatus status;
87 gchar *key; /* Key to read/write */
88 BlueSkyRCStr *data; /* Data read/to write */
90 int result; /* Result code; 0 for success. */
91 struct BlueSkyNotifierList *notifiers;
94 /* The barrier waiting on this operation. Support for more than one
95 * barrier for a single async is not well-supported and should be avoided
97 BlueSkyStoreAsync *barrier;
99 bluesky_time_hires start_time; /* Time operation was submitted. */
100 bluesky_time_hires exec_time; /* Time processing started on operation. */
102 gpointer store_private; /* For use by the storage implementation */
105 /* Support for notification lists. These are lists of one-shot functions which
106 * can be called when certain events--primarily, competed storage
107 * events--occur. Multiple notifiers can be added, but no particular order is
108 * guaranteed for the notification functions to be called. */
109 struct BlueSkyNotifierList {
110 struct BlueSkyNotifierList *next;
112 BlueSkyStoreAsync *async;
113 gpointer user_data; // Passed to the function when called
116 /* The abstraction layer for storage, allowing multiple implementations. */
118 /* Create a new store instance and return a handle to it. */
119 gpointer (*create)(const gchar *path);
121 /* Clean up any resources used by this store. */
122 void (*destroy)(gpointer store);
124 /* Submit an operation (get/put/delete) to the storage layer to be
125 * performed asynchronously. */
126 void (*submit)(gpointer store, BlueSkyStoreAsync *async);
128 /* Clean up any implementation-private data in a BlueSkyStoreAsync. */
129 void (*cleanup)(gpointer store, BlueSkyStoreAsync *async);
130 } BlueSkyStoreImplementation;
132 void bluesky_store_register(const BlueSkyStoreImplementation *impl,
135 BlueSkyStoreAsync *bluesky_store_async_new(BlueSkyStore *store);
136 gpointer bluesky_store_async_get_handle(BlueSkyStoreAsync *async);
137 void bluesky_store_async_ref(BlueSkyStoreAsync *async);
138 void bluesky_store_async_unref(BlueSkyStoreAsync *async);
139 void bluesky_store_async_wait(BlueSkyStoreAsync *async);
140 void bluesky_store_async_add_notifier(BlueSkyStoreAsync *async,
141 GFunc func, gpointer user_data);
142 void bluesky_store_async_mark_complete(BlueSkyStoreAsync *async);
143 void bluesky_store_async_submit(BlueSkyStoreAsync *async);
144 void bluesky_store_sync(BlueSkyStore *store);
146 void bluesky_store_add_barrier(BlueSkyStoreAsync *barrier,
147 BlueSkyStoreAsync *async);
149 void bluesky_inode_start_sync(BlueSkyInode *inode);
151 void bluesky_block_touch(BlueSkyInode *inode, uint64_t i);
152 void bluesky_block_fetch(BlueSkyInode *inode, BlueSkyBlock *block,
153 BlueSkyStoreAsync *barrier);
154 void bluesky_block_flush(BlueSkyInode *inode, BlueSkyBlock *block,
156 void bluesky_file_flush(BlueSkyInode *inode, GList **log_items);
157 void bluesky_file_drop_cached(BlueSkyInode *inode);
159 /* Writing of data to the cloud in log segments and tracking the location of
160 * various pieces of data (both where in the cloud and where cached locally).
172 } BlueSkyCloudPointer;
175 LOGTYPE_INVALID = -1,
179 LOGTYPE_INODE_MAP = 3,
180 LOGTYPE_CHECKPOINT = 4,
181 LOGTYPE_CHECKPOINT_PTR = 5,
182 } BlueSkyCloudLogType;
184 /* Headers that go on items in local log segments and cloud log segments. */
186 uint32_t magic; // HEADER_MAGIC
187 uint8_t type; // Object type + '0'
188 uint32_t offset; // Starting byte offset of the log header
189 uint32_t size1; // Size of the data item (bytes)
192 uint64_t inum; // Inode which owns this data, if any
193 BlueSkyCloudID id; // Object identifier
194 } __attribute__((packed));
197 uint32_t magic; // FOOTER_MAGIC
198 uint32_t crc; // Computed from log_header to log_footer.magic
199 } __attribute__((packed));
201 struct cloudlog_header {
205 uint32_t size1, size2, size3;
206 } __attribute__((packed));
208 #define JOURNAL_MAGIC "\nLog"
209 #define CLOUDLOG_MAGIC "AgI-"
211 /* A record which tracks an object which has been written to a local log,
212 * cached, locally, and/or written to the cloud. */
213 #define CLOUDLOG_JOURNAL 0x01
214 #define CLOUDLOG_CLOUD 0x02
215 #define CLOUDLOG_CACHE 0x04
216 struct _BlueSkyCloudLog {
223 BlueSkyCloudLogType type;
225 // Bitmask of CLOUDLOG_* flags indicating where the object exists.
227 int pending_read, pending_write;
229 // A stable identifier for the object (only changes when authenticated data
230 // is written out, but stays the same when the in-cloud cleaner relocates
234 // The inode which owns this data, if any, and an offset.
238 // The size of encrypted object data, not including any headers
241 // The location of the object in the cloud, if available.
242 BlueSkyCloudPointer location;
244 // TODO: Location in journal/cache
245 int log_seq, log_offset, log_size;
247 // Pointers to other objects. Each link counts towards the reference count
248 // of the pointed-to object. To avoid memory leaks there should be no
249 // cycles in the reference graph.
252 // Serialized data, if available in memory (otherwise NULL), and a lock
253 // count which tracks if there are users that require the data to be kept
259 /* Serialize objects into a log segment to be written to the cloud. */
260 struct _BlueSkyCloudLogState {
262 BlueSkyCloudPointer location;
264 GSList *writeback_list; // Items which are being serialized right now
265 GList *pending_segments; // Segments which are being uploaded now
268 gboolean bluesky_cloudlog_equal(gconstpointer a, gconstpointer b);
269 guint bluesky_cloudlog_hash(gconstpointer a);
270 BlueSkyCloudLog *bluesky_cloudlog_new(BlueSkyFS *fs, const BlueSkyCloudID *id);
271 gchar *bluesky_cloudlog_id_to_string(BlueSkyCloudID id);
272 BlueSkyCloudID bluesky_cloudlog_id_from_string(const gchar *idstr);
273 void bluesky_cloudlog_ref(BlueSkyCloudLog *log);
274 void bluesky_cloudlog_unref(BlueSkyCloudLog *log);
275 void bluesky_cloudlog_stats_update(BlueSkyCloudLog *log, int type);
276 void bluesky_cloudlog_sync(BlueSkyCloudLog *log);
277 void bluesky_cloudlog_insert(BlueSkyCloudLog *log);
278 BlueSkyCloudLog *bluesky_cloudlog_get(BlueSkyFS *fs, BlueSkyCloudID id);
279 void bluesky_cloudlog_fetch(BlueSkyCloudLog *log);
280 BlueSkyCloudPointer bluesky_cloudlog_serialize(BlueSkyCloudLog *log,
282 void bluesky_cloudlog_flush(BlueSkyFS *fs);
284 /* Logging infrastructure for ensuring operations are persistently recorded to
286 #define BLUESKY_CRC32C_SEED (~(uint32_t)0)
287 #define BLUESKY_CRC32C_VALIDATOR ((uint32_t)0xb798b438UL)
288 uint32_t crc32c(uint32_t crc, const char *buf, unsigned int length);
289 uint32_t crc32c_finalize(uint32_t crc);
299 /* The currently-open log file. */
300 BlueSkyCacheFile *current_log;
302 /* Cache of log segments which have been memory-mapped. */
304 GHashTable *mmap_cache;
306 /* A count of the disk space consumed (in 1024-byte units) by all files
307 * tracked by mmap_cache (whether mapped or not, actually). */
310 /* The smallest journal sequence number which may still contain data that
311 * must be preserved (since it it not yet in the cloud). */
312 int journal_watermark;
315 /* An object for tracking log files which are stored locally--either the
316 * journal for filesystem consistency or log segments which have been fetched
317 * back from cloud storage. */
318 struct _BlueSkyCacheFile {
322 int type; // Only one of CLOUDLOG_{JOURNAL,CLOUD}
325 char *filename; // Local filename, relateive to log directory
326 gint mapcount; // References to the mmaped data
327 const char *addr; // May be null if data is not mapped in memory
331 gboolean fetching, ready; // Cloud data: downloading or ready for use
332 int64_t atime; // Access time, for cache management
335 BlueSkyLog *bluesky_log_new(const char *log_directory);
336 void bluesky_log_item_submit(BlueSkyCloudLog *item, BlueSkyLog *log);
337 void bluesky_log_finish_all(GList *log_items);
338 BlueSkyRCStr *bluesky_log_map_object(BlueSkyFS *fs, int log_dir, int log_seq,
339 int log_offset, int log_size);
340 void bluesky_mmap_unref(BlueSkyCacheFile *mmap);
341 void bluesky_cachefile_unref(BlueSkyCacheFile *cachefile);
343 BlueSkyCacheFile *bluesky_cachefile_lookup(BlueSkyFS *fs,
344 int clouddir, int log_seq);
345 void bluesky_cachefile_gc(BlueSkyFS *fs);
347 void bluesky_replay(BlueSkyFS *fs);
349 /* Used to track log segments that are being written to the cloud. */
358 /***** Inode map management *****/
360 /* Mapping information for a single inode number. These are grouped together
361 * into InodeMapRange objects. */
365 /* The ID of the most recent version of the inode. */
368 /* The location where that version is written in the cloud. */
369 BlueSkyCloudPointer location;
371 /* If the cloud log entry exists in memory, then a pointer to it, otherwise
373 BlueSkyCloudLog *item;
377 /* Starting and ending inode number values that fall in this section.
378 * Endpoint values are inclusive. */
381 /* A sorted list (by inode number) of InodeMapEntry objects. */
382 GSequence *map_entries;
384 /* The serialized version of the inode map data. */
385 BlueSkyCloudLog *serialized;
387 /* Have there been changes that require writing this section out again? */
391 InodeMapEntry *bluesky_inode_map_lookup(GSequence *inode_map, uint64_t inum,
393 BlueSkyCloudLog *bluesky_inode_map_serialize(BlueSkyFS *fs);