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>
20 /* Various options to tweak for performance benchmarking purposes. */
22 /* Perform all get/put operations synchronously. */
23 int synchronous_stores;
25 /* Write data in cache immediately after file is modified. */
26 int writethrough_cache;
29 extern BlueSkyOptions bluesky_options;
31 /* BlueSky status and error codes. Various frontends should translate these to
32 * the appropriate error code for whatever protocol they implement. */
34 BSTATUS_OK = 0, /* No error */
35 BSTATUS_IOERR, /* I/O error of some form */
36 BSTATUS_NOENT, /* File does not exist */
39 void bluesky_init(void);
41 gchar *bluesky_lowercase(const gchar *s);
43 /* Reference-counted blocks of memory, used for passing data in and out of
44 * storage backends and in other places. */
51 BlueSkyRCStr *bluesky_string_new(gpointer data, gsize len);
52 BlueSkyRCStr *bluesky_string_new_from_gstring(GString *s);
53 void bluesky_string_ref(BlueSkyRCStr *string);
54 void bluesky_string_unref(BlueSkyRCStr *string);
55 BlueSkyRCStr *bluesky_string_dup(BlueSkyRCStr *string);
56 void bluesky_string_resize(BlueSkyRCStr *string, gsize len);
58 /* Cryptographic operations. */
59 #define CRYPTO_BLOCK_SIZE 16 /* 128-bit AES */
60 #define CRYPTO_KEY_SIZE 16
62 void bluesky_crypt_init();
63 void bluesky_crypt_hash_key(const char *keystr, uint8_t *out);
64 void bluesky_crypt_random_bytes(guchar *buf, gint len);
65 BlueSkyRCStr *bluesky_crypt_encrypt(BlueSkyRCStr *in, const uint8_t *key);
66 BlueSkyRCStr *bluesky_crypt_decrypt(BlueSkyRCStr *in, const uint8_t *key);
68 /* Storage interface. This presents a key-value store abstraction, and can
69 * have multiple implementations: in-memory, on-disk, in-cloud. */
71 typedef struct _BlueSkyStore BlueSkyStore;
73 void bluesky_store_init();
74 BlueSkyStore *bluesky_store_new(const gchar *type);
75 void bluesky_store_free(BlueSkyStore *store);
76 BlueSkyRCStr *bluesky_store_get(BlueSkyStore *store, const gchar *key);
77 void bluesky_store_put(BlueSkyStore *store,
78 const gchar *key, BlueSkyRCStr *val);
80 /* File types. The numeric values are chosen to match with those used in
84 BLUESKY_DIRECTORY = 2,
86 BLUESKY_CHARACTER = 4,
92 /* Filesystem state. Each filesystem which is exported is represented by a
93 * single bluesky_fs structure in memory. */
97 gchar *name; /* Descriptive name for the filesystem */
98 GHashTable *inodes; /* Cached inodes */
99 uint64_t next_inum; /* Next available inode for allocation */
104 /* Inode number of the root directory. */
105 #define BLUESKY_ROOT_INUM 1
107 /* Timestamp, measured in microseconds since the Unix epoch. */
108 typedef int64_t bluesky_time;
110 /* High-resolution timer, measured in nanoseconds. */
111 typedef int64_t bluesky_time_hires;
112 bluesky_time_hires bluesky_now_hires();
114 /* In-memory representation of an inode within a Blue Sky server. This
115 * corresponds roughly with information that is committed to persistent
116 * storage. Locking/refcounting rules:
117 * - To access or modify any data fields, the lock must be held. This
118 * includes file blocks.
119 * - One reference is held by the BlueSkyFS inode hash table. If that is the
120 * only reference (and the inode is unlocked), the inode is subject to
121 * dropping from the cache.
122 * - Any pending operations should hold extra references to the inode as
123 * appropriate to keep it available until the operation completes.
131 BlueSkyFileType type;
136 /* Rather than track an inode number and generation number, we will simply
137 * never re-use a fileid after a file is deleted. 64 bits should be enough
138 * that we don't exhaust the identifier space. */
141 /* change_count is increased with every operation which modifies the inode,
142 * and can be used to determine if cached data is still valid.
143 * change_commit is the value of change_count when the inode was last
144 * committed to stable storage. */
145 uint64_t change_count, change_commit;
147 /* Timestamp for controlling when modified data is flushed to stable
148 * storage. When an inode is first modified from a clean state, this is
149 * set to the current time. If the inode is clean, it is set to zero. */
152 /* Additional state for tracking cache writeback status. */
153 uint64_t change_pending; /* change_count version currently being committed to storage */
155 int64_t atime; /* Microseconds since the Unix epoch */
158 int64_t ntime; /* "new time": time object was created */
160 /* File-specific fields */
164 /* Directory-specific fields */
165 GSequence *dirents; /* List of entries for READDIR */
166 GHashTable *dirhash; /* Hash table by name for LOOKUP */
167 GHashTable *dirhash_folded; /* As above, but case-folded */
168 uint64_t parent_inum; /* inode for ".."; 0 if the root directory */
170 /* Symlink-specific fields */
171 gchar *symlink_contents;
174 /* A directory entry. The name is UTF-8 and is a freshly-allocated string.
175 * Each directory entry is listed in two indices: dirents is indexed by cookie
176 * and dirhash by name. The cookie is a randomly-assigned 32-bit value, unique
177 * within the directory, that remains unchanged until the entry is deleted. It
178 * is used to provide a stable key for restarting a READDIR call. */
181 gchar *name_folded; /* Name, folded for case-insensitive lookup */
186 /* File data is divided into fixed-size blocks (except the last block which may
187 * be short?). These blocks are backed by storage in a key/value store, but
188 * may also be dirty if modifications have been made in-memory that have not
190 #define BLUESKY_BLOCK_SIZE 32768ULL
191 #define BLUESKY_MAX_FILE_SIZE (BLUESKY_BLOCK_SIZE << 24)
193 BLUESKY_BLOCK_ZERO = 0, /* Data is all zeroes, not explicitly stored */
194 BLUESKY_BLOCK_REF = 1, /* Reference to key/value store, not cached */
195 BLUESKY_BLOCK_CACHED = 2, /* Data is cached in memory, clean */
196 BLUESKY_BLOCK_DIRTY = 3, /* Data needs to be committed to store */
200 BlueSkyBlockType type;
201 gchar *ref; /* Name of data block in the backing store */
202 BlueSkyRCStr *data; /* Pointer to data in memory if cached */
205 BlueSkyFS *bluesky_init_fs(gchar *name, BlueSkyStore *store);
206 void bluesky_superblock_flush(BlueSkyFS *fs);
208 int64_t bluesky_get_current_time();
209 void bluesky_inode_update_ctime(BlueSkyInode *inode, gboolean update_mtime);
210 uint64_t bluesky_fs_alloc_inode(BlueSkyFS *fs);
211 BlueSkyInode *bluesky_new_inode(uint64_t inum, BlueSkyFS *fs, BlueSkyFileType type);
213 BlueSkyInode *bluesky_get_inode(BlueSkyFS *fs, uint64_t inum);
214 void bluesky_insert_inode(BlueSkyFS *fs, BlueSkyInode *inode);
216 void bluesky_dirent_destroy(gpointer dirent);
217 uint64_t bluesky_directory_lookup(BlueSkyInode *inode, gchar *name);
218 uint64_t bluesky_directory_ilookup(BlueSkyInode *inode, gchar *name);
219 BlueSkyDirent *bluesky_directory_read(BlueSkyInode *dir, uint32_t cookie);
220 gboolean bluesky_directory_insert(BlueSkyInode *dir, const gchar *name,
222 void bluesky_directory_dump(BlueSkyInode *dir);
224 void bluesky_block_touch(BlueSkyInode *inode, uint64_t i);
225 void bluesky_block_fetch(BlueSkyFS *fs, BlueSkyBlock *block);
226 void bluesky_file_truncate(BlueSkyInode *inode, uint64_t size);
227 void bluesky_file_write(BlueSkyInode *inode, uint64_t offset,
228 const char *data, gint len);
229 void bluesky_file_read(BlueSkyInode *inode, uint64_t offset,
230 char *buf, gint len);
232 void bluesky_inode_flush(BlueSkyFS *fs, BlueSkyInode *inode);
233 void bluesky_inode_fetch(BlueSkyFS *fs, uint64_t inum);
235 gint bluesky_dirent_compare(gconstpointer a, gconstpointer b,
238 void bluesky_flushd_invoke(BlueSkyFS *fs);