X-Git-Url: http://git.vrable.net/?a=blobdiff_plain;f=bluesky%2Fbluesky-private.h;h=e5b5b583919576614862bb79eb2cd2df06c86610;hb=a82b60b3b683840a7074110831bcbaa16a40f0eb;hp=ba9df02bbdb7b8e929468ed1013869f76dffa966;hpb=641411b0a7ce9afa309bea28c01320eda913b2d2;p=bluesky.git diff --git a/bluesky/bluesky-private.h b/bluesky/bluesky-private.h index ba9df02..e5b5b58 100644 --- a/bluesky/bluesky-private.h +++ b/bluesky/bluesky-private.h @@ -19,6 +19,8 @@ extern "C" { #endif +extern int bluesky_verbose; + /* Target cache size levels. */ extern int bluesky_watermark_low_dirty; extern int bluesky_watermark_medium_dirty; @@ -42,7 +44,7 @@ BlueSkyInode *bluesky_list_tail(GList *head); * persistent storage. */ void bluesky_serialize_superblock(GString *out, BlueSkyFS *fs); BlueSkyFS *bluesky_deserialize_superblock(const gchar *buf); -void bluesky_serialize_inode(GString *out, BlueSkyInode *inode); +BlueSkyCloudLog *bluesky_serialize_inode(BlueSkyInode *inode); gboolean bluesky_deserialize_inode(BlueSkyInode *inode, const gchar *buf); /* Storage layer. Requests can be performed asynchronously, so these objects @@ -137,16 +139,217 @@ void bluesky_store_sync(BlueSkyStore *store); void bluesky_store_add_barrier(BlueSkyStoreAsync *barrier, BlueSkyStoreAsync *async); -void bluesky_inode_start_sync(BlueSkyInode *inode, BlueSkyStoreAsync *barrier); +void bluesky_inode_start_sync(BlueSkyInode *inode); void bluesky_block_touch(BlueSkyInode *inode, uint64_t i); -void bluesky_block_fetch(BlueSkyFS *fs, BlueSkyBlock *block, - BlueSkyStoreAsync *barrier); -void bluesky_block_flush(BlueSkyFS *fs, BlueSkyBlock *block, +void bluesky_block_fetch(BlueSkyInode *inode, BlueSkyBlock *block, BlueSkyStoreAsync *barrier); -void bluesky_file_flush(BlueSkyInode *inode, BlueSkyStoreAsync *barrier); +void bluesky_block_flush(BlueSkyInode *inode, BlueSkyBlock *block, + GList **log_items); +void bluesky_file_flush(BlueSkyInode *inode, GList **log_items); void bluesky_file_drop_cached(BlueSkyInode *inode); +/* Writing of data to the cloud in log segments and tracking the location of + * various pieces of data (both where in the cloud and where cached locally). + * */ + +typedef struct { + char bytes[16]; +} BlueSkyCloudID; + +typedef struct { + uint32_t directory; + uint32_t sequence; + uint32_t offset; + uint32_t size; +} BlueSkyCloudPointer; + +typedef enum { + LOGTYPE_INVALID = -1, + LOGTYPE_UNKNOWN = 0, + LOGTYPE_DATA = 1, + LOGTYPE_INODE = 2, + LOGTYPE_INODE_MAP = 3, + LOGTYPE_CHECKPOINT = 4, + LOGTYPE_CHECKPOINT_PTR = 5, +} BlueSkyCloudLogType; + +/* A record which tracks an object which has been written to a local log, + * cached, locally, and/or written to the cloud. */ +#define CLOUDLOG_JOURNAL 0x01 +#define CLOUDLOG_CLOUD 0x02 +#define CLOUDLOG_CACHE 0x04 +struct _BlueSkyCloudLog { + gint refcount; + GMutex *lock; + GCond *cond; + + BlueSkyFS *fs; + + BlueSkyCloudLogType type; + + // Bitmask of CLOUDLOG_* flags indicating where the object exists. + int location_flags; + int pending_read, pending_write; + + // A stable identifier for the object (only changes when authenticated data + // is written out, but stays the same when the in-cloud cleaner relocates + // the object). + BlueSkyCloudID id; + + // The inode which owns this data, if any, and an offset. + uint64_t inum; + int32_t inum_offset; + + // The location of the object in the cloud, if available. + BlueSkyCloudPointer location; + + // TODO: Location in journal/cache + int log_seq, log_offset, log_size; + + // Pointers to other objects. Each link counts towards the reference count + // of the pointed-to object. To avoid memory leaks there should be no + // cycles in the reference graph. + GArray *links; + + // Serialized data, if available in memory (otherwise NULL), and a lock + // count which tracks if there are users that require the data to be kept + // around. + BlueSkyRCStr *data; + int data_lock_count; +}; + +/* Serialize objects into a log segment to be written to the cloud. */ +struct _BlueSkyCloudLogState { + GString *data; + BlueSkyCloudPointer location; + GList *inode_list; + GSList *writeback_list; // Items which are being serialized right now + GList *pending_segments; // Segments which are being uploaded now +}; + +gboolean bluesky_cloudlog_equal(gconstpointer a, gconstpointer b); +guint bluesky_cloudlog_hash(gconstpointer a); +BlueSkyCloudLog *bluesky_cloudlog_new(BlueSkyFS *fs); +gchar *bluesky_cloudlog_id_to_string(BlueSkyCloudID id); +BlueSkyCloudID bluesky_cloudlog_id_from_string(const gchar *idstr); +void bluesky_cloudlog_ref(BlueSkyCloudLog *log); +void bluesky_cloudlog_unref(BlueSkyCloudLog *log); +void bluesky_cloudlog_stats_update(BlueSkyCloudLog *log, int type); +void bluesky_cloudlog_sync(BlueSkyCloudLog *log); +void bluesky_cloudlog_insert(BlueSkyCloudLog *log); +void bluesky_cloudlog_fetch(BlueSkyCloudLog *log); +BlueSkyCloudPointer bluesky_cloudlog_serialize(BlueSkyCloudLog *log, + BlueSkyFS *fs); +void bluesky_cloudlog_flush(BlueSkyFS *fs); + +/* Logging infrastructure for ensuring operations are persistently recorded to + * disk. */ +#define BLUESKY_CRC32C_SEED (~(uint32_t)0) +uint32_t crc32c(uint32_t crc, const char *buf, unsigned int length); +uint32_t crc32c_finalize(uint32_t crc); + +struct _BlueSkyLog { + BlueSkyFS *fs; + char *log_directory; + GAsyncQueue *queue; + int fd, dirfd; + int seq_num; + GSList *committed; + + /* The currently-open log file. */ + BlueSkyCacheFile *current_log; + + /* Cache of log segments which have been memory-mapped. */ + GMutex *mmap_lock; + GHashTable *mmap_cache; + + /* A count of the disk space consumed (in 1024-byte units) by all files + * tracked by mmap_cache (whether mapped or not, actually). */ + gint disk_used; + + /* The smallest journal sequence number which may still contain data that + * must be preserved (since it it not yet in the cloud). */ + int journal_watermark; +}; + +/* An object for tracking log files which are stored locally--either the + * journal for filesystem consistency or log segments which have been fetched + * back from cloud storage. */ +struct _BlueSkyCacheFile { + GMutex *lock; + GCond *cond; + gint refcount; + int type; // Only one of CLOUDLOG_{JOURNAL,CLOUD} + int log_dir; + int log_seq; + char *filename; // Local filename, relateive to log directory + gint mapcount; // References to the mmaped data + const char *addr; // May be null if data is not mapped in memory + size_t len; + BlueSkyFS *fs; + BlueSkyLog *log; + gboolean fetching, ready; // Cloud data: downloading or ready for use + int64_t atime; // Access time, for cache management +}; + +BlueSkyLog *bluesky_log_new(const char *log_directory); +void bluesky_log_item_submit(BlueSkyCloudLog *item, BlueSkyLog *log); +void bluesky_log_finish_all(GList *log_items); +BlueSkyRCStr *bluesky_log_map_object(BlueSkyFS *fs, int log_dir, int log_seq, + int log_offset, int log_size); +void bluesky_mmap_unref(BlueSkyCacheFile *mmap); +void bluesky_cachefile_unref(BlueSkyCacheFile *cachefile); + +BlueSkyCacheFile *bluesky_cachefile_lookup(BlueSkyFS *fs, + int clouddir, int log_seq); +void bluesky_cachefile_gc(BlueSkyFS *fs); + +/* Used to track log segments that are being written to the cloud. */ +typedef struct { + BlueSkyRCStr *data; + GSList *items; + GMutex *lock; + GCond *cond; + gboolean complete; +} SerializedRecord; + +/***** Inode map management *****/ + +/* Mapping information for a single inode number. These are grouped together + * into InodeMapRange objects. */ +typedef struct { + uint64_t inum; + + /* The ID of the most recent version of the inode. */ + BlueSkyCloudID id; + + /* The location where that version is written in the cloud. */ + BlueSkyCloudPointer location; + + /* If the cloud log entry exists in memory, then a pointer to it, otherwise + * NULL. */ + BlueSkyCloudLog *item; +} InodeMapEntry; + +typedef struct { + /* Starting and ending inode number values that fall in this section. + * Endpoint values are inclusive. */ + uint64_t start, end; + + /* A sorted list (by inode number) of InodeMapEntry objects. */ + GSequence *map_entries; + + /* The location where this inode map section is stored in the cloud. */ + BlueSkyCloudPointer location; + + /* Have there been changes that require writing this section out again? */ + gboolean dirty; +} InodeMapRange; + +InodeMapEntry *bluesky_inode_map_lookup(GSequence *inode_map, uint64_t inum, + int action); + #ifdef __cplusplus } #endif