#define _BLUESKY_H
#include <stdint.h>
+#include <stdio.h>
#include <inttypes.h>
#include <glib.h>
/* Various options to tweak for performance benchmarking purposes. */
typedef struct {
- /* Perform all get/put operations synchronously. */
+ /* OBSOLETE: Perform all get/put operations synchronously. */
int synchronous_stores;
- /* Write data in cache immediately after file is modified. */
+ /* OBSOLETE: Write data in cache immediately after file is modified. */
int writethrough_cache;
+
+ /* Can inodes be fetched asynchronously? (Inode object is initially
+ * created in a pending state, and not unlocked until the data is actually
+ * available.) */
+ int sync_inode_fetches;
+
+ /* 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;
+
+ /* Full segment fetches (1) or use range requests (0) for fetching segments
+ * from cloud? */
+ int full_segment_fetches;
+
+ /* Disable aggregating of data into log segments. Each object will be
+ * stored in a separate segment. */
+ int disable_aggregation;
+
+ /* Disable cryptography. This is for benchmarking purposes. */
+ int disable_crypto;
+
+ /* Disable aggregation of read requests. Fetch items individually. */
+ int disable_read_aggregation;
} 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();
+void bluesky_stats_run_periodic_dump(FILE *f);
+
/* 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);
+struct BlueSkyRangeset;
+typedef struct BlueSkyRangeset BlueSkyRangeset;
+typedef struct {
+ uint64_t start, length;
+ gpointer data;
+} BlueSkyRangesetItem;
+
+BlueSkyRangeset *bluesky_rangeset_new();
+void bluesky_rangeset_free(BlueSkyRangeset *rangeset);
+gboolean bluesky_rangeset_insert(BlueSkyRangeset *rangeset,
+ uint64_t start, uint64_t length,
+ gpointer data);
+const BlueSkyRangesetItem *bluesky_rangeset_lookup(BlueSkyRangeset *rangeset,
+ uint64_t offset);
+const BlueSkyRangesetItem *bluesky_rangeset_lookup_next(BlueSkyRangeset *rangeset, uint64_t offset);
+void bluesky_rangeset_get_extents(BlueSkyRangeset *rangeset,
+ uint64_t *start, uint64_t *length);
/* 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 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);
BLUESKY_SYMLINK = 5,
BLUESKY_SOCKET = 6,
BLUESKY_FIFO = 7,
+
+ /* Special types used only internally. */
+ BLUESKY_PENDING = 0, /* Inode being loaded; type not yet determined */
+ BLUESKY_INVALID = -1, /* Inode is invalid (failed to load) */
} BlueSkyFileType;
/* 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. */
* dropping from the cache.
* - Any pending operations should hold extra references to the inode as
* appropriate to keep it available until the operation completes.
+ * - Locking dependency order is, when multiple locks are to be acquired, to
+ * acquire locks on parents in the filesystem tree before children.
+ * (TODO: What about rename when we acquire locks in unrelated parts of the
+ * filesystem?)
+ * - An inode should not be locked while the filesystem lock is already held,
+ * since some code may do an inode lookup (which acquires the filesystem
+ * lock) while a different inode is locked.
* */
typedef struct {
GMutex *lock;
/* 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;
/* Symlink-specific fields */
gchar *symlink_contents;
+
+ /* A field for short-term use internally while the lock is held. */
+ gpointer private_data;
} BlueSkyInode;
/* A directory entry. The name is UTF-8 and is a freshly-allocated string.
#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 bluesky_fs_alloc_inode(BlueSkyFS *fs);
+void bluesky_init_inode(BlueSkyInode *i, BlueSkyFileType type);
BlueSkyInode *bluesky_new_inode(uint64_t inum, BlueSkyFS *fs, BlueSkyFileType type);
BlueSkyInode *bluesky_get_inode(BlueSkyFS *fs, uint64_t inum);
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);
+void bluesky_inode_prefetch(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);
+/* Request response time tracking. */
+typedef struct BlueSkyProfile {
+ int magic;
+ GMutex *lock;
+ char *description;
+ GList *events;
+} BlueSkyProfile;
+
+BlueSkyProfile *bluesky_profile_new();
+void bluesky_profile_free(BlueSkyProfile *profile);
+void bluesky_profile_add_event(BlueSkyProfile *profile, char *message);
+void bluesky_profile_print(BlueSkyProfile *profile);
+BlueSkyProfile *bluesky_profile_get();
+void bluesky_profile_set(BlueSkyProfile *profile);
+void bluesky_profile_set_output(FILE *stream);
+
#ifdef __cplusplus
}
#endif
projects / bluesky.git / blobdiff