/* Should frontends handle requests serially or allow operations to proceed
* in parallel? */
int sync_frontends;
+
+ /* Target size of the disk cache at the proxy, in kilobytes. */
+ int cache_size;
} BlueSkyOptions;
extern BlueSkyOptions bluesky_options;
+/* Maximum number of threads to use in any particular thread pool, or -1 for no
+ * limit */
+extern int bluesky_max_threads;
+
+/* A general-purpose counter for gathering run-time statistics. */
+struct bluesky_stats {
+ const char *name;
+ int64_t count;
+ int64_t sum;
+};
+struct bluesky_stats *bluesky_stats_new(const char *name);
+void bluesky_stats_add(struct bluesky_stats *stats, int64_t value);
+void bluesky_stats_dump_all();
+
/* BlueSky status and error codes. Various frontends should translate these to
* the appropriate error code for whatever protocol they implement. */
typedef enum {
gchar *bluesky_lowercase(const gchar *s);
-/* Reference-counted blocks of memory, used for passing data in and out of
- * storage backends and in other places. */
+struct _BlueSkyCacheFile;
+typedef struct _BlueSkyCacheFile BlueSkyCacheFile;
+
typedef struct {
gint refcount;
+ BlueSkyCacheFile *mmap;
gchar *data;
gsize len;
} BlueSkyRCStr;
BlueSkyRCStr *bluesky_string_new(gpointer data, gsize len);
BlueSkyRCStr *bluesky_string_new_from_gstring(GString *s);
+BlueSkyRCStr *bluesky_string_new_from_mmap(BlueSkyCacheFile *mmap,
+ int offset, gsize len);
void bluesky_string_ref(BlueSkyRCStr *string);
void bluesky_string_unref(BlueSkyRCStr *string);
BlueSkyRCStr *bluesky_string_dup(BlueSkyRCStr *string);
void bluesky_string_resize(BlueSkyRCStr *string, gsize len);
-/* Cryptographic operations. */
-#define CRYPTO_BLOCK_SIZE 16 /* 128-bit AES */
-#define CRYPTO_KEY_SIZE 16
-
-void bluesky_crypt_init();
-void bluesky_crypt_hash_key(const char *keystr, uint8_t *out);
-void bluesky_crypt_random_bytes(guchar *buf, gint len);
-BlueSkyRCStr *bluesky_crypt_encrypt(BlueSkyRCStr *in, const uint8_t *key);
-BlueSkyRCStr *bluesky_crypt_decrypt(BlueSkyRCStr *in, const uint8_t *key);
-
/* Storage interface. This presents a key-value store abstraction, and can
* have multiple implementations: in-memory, on-disk, in-cloud. */
struct _BlueSkyStore;
typedef struct _BlueSkyStore BlueSkyStore;
+struct _BlueSkyLog;
+typedef struct _BlueSkyLog BlueSkyLog;
+
+struct _BlueSkyCloudLogState;
+typedef struct _BlueSkyCloudLogState BlueSkyCloudLogState;
+
+struct _BlueSkyCloudLog;
+typedef struct _BlueSkyCloudLog BlueSkyCloudLog;
+
void bluesky_store_init();
BlueSkyStore *bluesky_store_new(const gchar *type);
void bluesky_store_free(BlueSkyStore *store);
/* Filesystem state. Each filesystem which is exported is represented by a
* single bluesky_fs structure in memory. */
+struct BlueSkyCryptKeys;
+
typedef struct {
GMutex *lock;
uint64_t next_inum; /* Next available inode for allocation */
BlueSkyStore *store;
+ BlueSkyLog *log;
+ BlueSkyCloudLogState *log_state;
+
+ /* Filesystem crypto keys */
+ char *master_key;
+ struct BlueSkyCryptKeys *keys;
+
+ /* Accounting for memory used for caches. Space is measured in blocks, not
+ * bytes. Updates to these variables must be made atomically. */
+ gint cache_dirty;
+
+ /* Like above, but tracking data stored in the cloudlog entries
+ * specifically:
+ * - cache_log_dirty: data uncommitted to journal and cloud
+ * - cache_log_writeback: data being written to journal
+ * - cache_log_journal: data committed to journal
+ * - cache_log_cloud: data written to cloud as well
+ * Log entries should progress from the top state to the bottom, and are
+ * only ever counted in one category at a time. */
+ gint cache_log_dirty, cache_log_writeback,
+ cache_log_journal, cache_log_cloud;
+
+ /* Linked list of inodes, sorted by access/modification times for cache
+ * management. Editing these lists is protected by the filesystem lock; to
+ * avoid deadlock do not attempt to take any other locks while the FS lock
+ * is held for list editing purposes. Items at the head of the list are
+ * most recently accessed/modified. */
+ GList unlogged_list; // Changes not yet synced to journal
+ GList dirty_list; // Not yet written to cloud storage
+ GList accessed_list; // All in-memory inodes
+
+ /* Mutex for the flush daemon, to prevent concurrent execution. */
+ GMutex *flushd_lock;
+
+ /* Used to wait for the cache daemon to free up space */
+ GCond *flushd_cond;
+
+ /* Mapping of object identifiers (blocks, inodes) to physical location (in
+ * the local cache or in the logs in the cloud). */
+ GHashTable *locations;
+
+ /* The inode map, which maps inode numbers to the location of the most
+ * recent version. */
+ GSequence *inode_map;
+
+ /* Queue for asynchronous cloudlog unrefs, where needed. */
+ GAsyncQueue *unref_queue;
+
+ /* Thread pool for asynchronous inode fetches */
+ GThreadPool *inode_fetch_thread_pool;
} BlueSkyFS;
/* Inode number of the root directory. */
/* change_count is increased with every operation which modifies the inode,
* and can be used to determine if cached data is still valid.
* change_commit is the value of change_count when the inode was last
- * committed to stable storage. */
- uint64_t change_count, change_commit;
+ * committed to stable storage (the log).
+ * change_cloud tracks which version was last commited to cloud storage. */
+ uint64_t change_count, change_commit, change_cloud;
/* Timestamp for controlling when modified data is flushed to stable
* storage. When an inode is first modified from a clean state, this is
* set to the current time. If the inode is clean, it is set to zero. */
int64_t change_time;
- /* Additional state for tracking cache writeback status. */
- uint64_t change_pending; /* change_count version currently being committed to storage */
+ /* Last access time to this inode, for controlling cache evictions. */
+ int64_t access_time;
+
+ /* Version of the object last serialized and committed to storage. */
+ BlueSkyCloudLog *committed_item;
+
+ /* Pointers to the linked-list elements for this inode in the accessed and
+ * dirty linked lists. We re-use the GList structure, using ->next to
+ * point to the head of the list and ->prev to point to the tail. The data
+ * element is unused. */
+ GList *unlogged_list, *accessed_list, *dirty_list;
int64_t atime; /* Microseconds since the Unix epoch */
int64_t ctime;
#define BLUESKY_MAX_FILE_SIZE (BLUESKY_BLOCK_SIZE << 24)
typedef enum {
BLUESKY_BLOCK_ZERO = 0, /* Data is all zeroes, not explicitly stored */
- BLUESKY_BLOCK_REF = 1, /* Reference to key/value store, not cached */
- BLUESKY_BLOCK_CACHED = 2, /* Data is cached in memory, clean */
- BLUESKY_BLOCK_DIRTY = 3, /* Data needs to be committed to store */
+ BLUESKY_BLOCK_REF = 1, /* Reference to cloud log item, data clean */
+ BLUESKY_BLOCK_DIRTY = 2, /* Data needs to be committed to store */
} BlueSkyBlockType;
typedef struct {
BlueSkyBlockType type;
- gchar *ref; /* Name of data block in the backing store */
- BlueSkyRCStr *data; /* Pointer to data in memory if cached */
+ BlueSkyCloudLog *ref; /* if REF: cloud log entry with data */
+ BlueSkyRCStr *dirty; /* if DIRTY: raw data in memory */
} BlueSkyBlock;
-BlueSkyFS *bluesky_init_fs(gchar *name, BlueSkyStore *store);
-void bluesky_superblock_flush(BlueSkyFS *fs);
+BlueSkyFS *bluesky_init_fs(gchar *name, BlueSkyStore *store,
+ const gchar *master_key);
+
+gboolean bluesky_inode_is_ready(BlueSkyInode *inode);
int64_t bluesky_get_current_time();
void bluesky_inode_update_ctime(BlueSkyInode *inode, gboolean update_mtime);
uint64_t inum);
void bluesky_directory_dump(BlueSkyInode *dir);
-void bluesky_block_touch(BlueSkyInode *inode, uint64_t i);
-void bluesky_block_fetch(BlueSkyFS *fs, BlueSkyBlock *block);
void bluesky_file_truncate(BlueSkyInode *inode, uint64_t size);
void bluesky_file_write(BlueSkyInode *inode, uint64_t offset,
const char *data, gint len);
void bluesky_file_read(BlueSkyInode *inode, uint64_t offset,
char *buf, gint len);
-void bluesky_inode_flush(BlueSkyFS *fs, BlueSkyInode *inode);
void bluesky_inode_fetch(BlueSkyFS *fs, uint64_t inum);
gint bluesky_dirent_compare(gconstpointer a, gconstpointer b,
gpointer unused);
void bluesky_flushd_invoke(BlueSkyFS *fs);
+void bluesky_flushd_invoke_conditional(BlueSkyFS *fs);
void bluesky_inode_do_sync(BlueSkyInode *inode);
+void bluesky_flushd_thread_launch(BlueSkyFS *fs);
void bluesky_debug_dump(BlueSkyFS *fs);