X-Git-Url: http://git.vrable.net/?a=blobdiff_plain;f=bluesky%2Fbluesky.h;h=820fa123fc5c0c4f295439a6be6cc0804339e3c4;hb=83fd6b61a6e092a22d4d5e59ed95f05f5e287f11;hp=6ddf605cdb6ea8e946f28489383355af28bed3fd;hpb=d514caf49faff9295d0e497d3b6b8856fe83f8d0;p=bluesky.git diff --git a/bluesky/bluesky.h b/bluesky/bluesky.h index 6ddf605..820fa12 100644 --- a/bluesky/bluesky.h +++ b/bluesky/bluesky.h @@ -28,11 +28,29 @@ typedef struct { /* Can inodes be fetched asynchronously? (Inode object is initially * created in a pending state, and not unlocked until the data is actually * available.) */ - int async_inode_fetches; + int sync_inode_fetches; + + /* Should frontends handle requests serially or allow operations to proceed + * in parallel? */ + int sync_frontends; } 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 { @@ -45,16 +63,20 @@ void bluesky_init(void); 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 _BlueSkyMmap; +typedef struct _BlueSkyMmap BlueSkyMmap; + typedef struct { gint refcount; + BlueSkyMmap *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(BlueSkyMmap *mmap, + int offset, gsize len); void bluesky_string_ref(BlueSkyRCStr *string); void bluesky_string_unref(BlueSkyRCStr *string); BlueSkyRCStr *bluesky_string_dup(BlueSkyRCStr *string); @@ -75,6 +97,15 @@ BlueSkyRCStr *bluesky_crypt_decrypt(BlueSkyRCStr *in, const uint8_t *key); 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); @@ -108,6 +139,42 @@ typedef struct { uint64_t next_inum; /* Next available inode for allocation */ BlueSkyStore *store; + BlueSkyLog *log; + BlueSkyCloudLogState *log_state; + + /* 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; } BlueSkyFS; /* Inode number of the root directory. */ @@ -134,6 +201,9 @@ bluesky_time_hires bluesky_now_hires(); * 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; @@ -154,16 +224,26 @@ typedef struct { /* 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; @@ -204,20 +284,21 @@ typedef struct { #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); +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); @@ -237,21 +318,21 @@ gboolean bluesky_directory_insert(BlueSkyInode *dir, const gchar *name, 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);