X-Git-Url: http://git.vrable.net/?a=blobdiff_plain;f=bluesky%2Fbluesky-private.h;h=a783629fb2e06071fed19cfdd6b5faa8123bb823;hb=7f01069131555a0e9ab332e578a9eb2815e1e12d;hp=e5b5b583919576614862bb79eb2cd2df06c86610;hpb=a82b60b3b683840a7074110831bcbaa16a40f0eb;p=bluesky.git diff --git a/bluesky/bluesky-private.h b/bluesky/bluesky-private.h index e5b5b58..a783629 100644 --- a/bluesky/bluesky-private.h +++ b/bluesky/bluesky-private.h @@ -14,6 +14,7 @@ #define _BLUESKY_PRIVATE_H #include "bluesky.h" +#include #ifdef __cplusplus extern "C" { @@ -24,15 +25,19 @@ extern int bluesky_verbose; /* Target cache size levels. */ extern int bluesky_watermark_low_dirty; extern int bluesky_watermark_medium_dirty; +extern int bluesky_watermark_med2_dirty; extern int bluesky_watermark_high_dirty; extern int bluesky_watermark_low_total; extern int bluesky_watermark_medium_total; +extern int bluesky_watermark_med2_total; extern int bluesky_watermark_high_total; /* TODO: Make this go away entirely. */ BlueSkyFS *bluesky_new_fs(gchar *name); +void bluesky_inode_free_resources(BlueSkyInode *inode); + /* Linked list update functions for LRU lists. */ void bluesky_list_unlink(GList *head, GList *item); GList *bluesky_list_prepend(GList *head, BlueSkyInode *inode); @@ -45,7 +50,44 @@ BlueSkyInode *bluesky_list_tail(GList *head); void bluesky_serialize_superblock(GString *out, BlueSkyFS *fs); BlueSkyFS *bluesky_deserialize_superblock(const gchar *buf); BlueSkyCloudLog *bluesky_serialize_inode(BlueSkyInode *inode); -gboolean bluesky_deserialize_inode(BlueSkyInode *inode, const gchar *buf); +gboolean bluesky_deserialize_inode(BlueSkyInode *inode, BlueSkyCloudLog *item); + +void bluesky_deserialize_cloudlog(BlueSkyCloudLog *item, + const char *data, + size_t len); + +void bluesky_serialize_cloudlog(BlueSkyCloudLog *log, + GString *encrypted, + GString *authenticated, + GString *writable); + +/* Cryptographic operations. */ +#define CRYPTO_BLOCK_SIZE 16 /* 128-bit AES */ +#define CRYPTO_KEY_SIZE 16 +#define CRYPTO_HASH_SIZE 32 /* SHA-256 */ + +typedef struct BlueSkyCryptKeys { + uint8_t encryption_key[CRYPTO_KEY_SIZE]; + uint8_t authentication_key[CRYPTO_HASH_SIZE]; +} BlueSkyCryptKeys; + +void bluesky_crypt_init(); +void bluesky_crypt_hash_key(const char *keystr, uint8_t *out); +void bluesky_crypt_random_bytes(guchar *buf, gint len); +void bluesky_crypt_derive_keys(BlueSkyCryptKeys *keys, const gchar *master); +BlueSkyRCStr *bluesky_crypt_encrypt(BlueSkyRCStr *in, const uint8_t *key); +BlueSkyRCStr *bluesky_crypt_decrypt(BlueSkyRCStr *in, const uint8_t *key); + +void bluesky_crypt_block_encrypt(gchar *cloud_block, size_t len, + BlueSkyCryptKeys *keys); +gboolean bluesky_crypt_block_decrypt(gchar *cloud_block, size_t len, + BlueSkyCryptKeys *keys, + gboolean allow_unauth); +void bluesky_cloudlog_encrypt(GString *segment, BlueSkyCryptKeys *keys); +void bluesky_cloudlog_decrypt(char *segment, size_t len, + BlueSkyCryptKeys *keys, + BlueSkyRangeset *items, + gboolean allow_unauth); /* Storage layer. Requests can be performed asynchronously, so these objects * help keep track of operations in progress. */ @@ -65,8 +107,8 @@ typedef enum { } BlueSkyAsyncStatus; struct BlueSkyNotifierList; -typedef struct _BlueSkyStoreAsync BlueSkyStoreAsync; -struct _BlueSkyStoreAsync { +typedef struct BlueSkyStoreAsync BlueSkyStoreAsync; +struct BlueSkyStoreAsync { BlueSkyStore *store; GMutex *lock; @@ -80,6 +122,14 @@ struct _BlueSkyStoreAsync { gchar *key; /* Key to read/write */ BlueSkyRCStr *data; /* Data read/to write */ + /* For range requests on reads: starting byte offset and length; len 0 + * implies reading to the end of the object. At completion, the backend + * should set range_done if a range read was made; if not set the entire + * object was read and the storage layer will select out just the + * appropriate bytes. */ + size_t start, len; + gboolean range_done; + int result; /* Result code; 0 for success. */ struct BlueSkyNotifierList *notifiers; gint notifier_count; @@ -93,6 +143,10 @@ struct _BlueSkyStoreAsync { bluesky_time_hires exec_time; /* Time processing started on operation. */ gpointer store_private; /* For use by the storage implementation */ + + /* If storage operations should be charged to any particular profile, which + * one? */ + BlueSkyProfile *profile; }; /* Support for notification lists. These are lists of one-shot functions which @@ -120,11 +174,16 @@ typedef struct { /* Clean up any implementation-private data in a BlueSkyStoreAsync. */ void (*cleanup)(gpointer store, BlueSkyStoreAsync *async); + + /* Find the lexicographically-largest file starting with the specified + * prefix. */ + char * (*lookup_last)(gpointer store, const gchar *prefix); } BlueSkyStoreImplementation; void bluesky_store_register(const BlueSkyStoreImplementation *impl, const gchar *name); +char *bluesky_store_lookup_last(BlueSkyStore *store, const char *prefix); BlueSkyStoreAsync *bluesky_store_async_new(BlueSkyStore *store); gpointer bluesky_store_async_get_handle(BlueSkyStoreAsync *async); void bluesky_store_async_ref(BlueSkyStoreAsync *async); @@ -141,7 +200,7 @@ void bluesky_store_add_barrier(BlueSkyStoreAsync *barrier, void bluesky_inode_start_sync(BlueSkyInode *inode); -void bluesky_block_touch(BlueSkyInode *inode, uint64_t i); +void bluesky_block_touch(BlueSkyInode *inode, uint64_t i, gboolean preserve); void bluesky_block_fetch(BlueSkyInode *inode, BlueSkyBlock *block, BlueSkyStoreAsync *barrier); void bluesky_block_flush(BlueSkyInode *inode, BlueSkyBlock *block, @@ -153,6 +212,12 @@ void bluesky_file_drop_cached(BlueSkyInode *inode); * various pieces of data (both where in the cloud and where cached locally). * */ +/* Eventually we'll want to support multiple writers. But for now, hard-code + * separate namespaces in the cloud for the proxy and the cleaner to write to. + * */ +#define BLUESKY_CLOUD_DIR_PRIMARY 0 +#define BLUESKY_CLOUD_DIR_CLEANER 1 + typedef struct { char bytes[16]; } BlueSkyCloudID; @@ -171,15 +236,55 @@ typedef enum { LOGTYPE_INODE = 2, LOGTYPE_INODE_MAP = 3, LOGTYPE_CHECKPOINT = 4, - LOGTYPE_CHECKPOINT_PTR = 5, + + /* Used only as metadata in the local journal, not loaded as a + * BlueSkyCloudLogState nor stored in the cloud */ + LOGTYPE_JOURNAL_MARKER = 16, + LOGTYPE_JOURNAL_CHECKPOINT = 17, } BlueSkyCloudLogType; +/* Headers that go on items in local log segments and cloud log segments. */ +struct log_header { + uint32_t magic; // HEADER_MAGIC + uint8_t type; // Object type + '0' + uint32_t offset; // Starting byte offset of the log header + uint32_t size1; // Size of the data item (bytes) + uint32_t size2; // + uint32_t size3; // + uint64_t inum; // Inode which owns this data, if any + BlueSkyCloudID id; // Object identifier +} __attribute__((packed)); + +struct log_footer { + uint32_t magic; // FOOTER_MAGIC + uint32_t crc; // Computed from log_header to log_footer.magic +} __attribute__((packed)); + +struct cloudlog_header { + char magic[4]; + uint8_t crypt_auth[CRYPTO_HASH_SIZE]; + uint8_t crypt_iv[CRYPTO_BLOCK_SIZE]; + uint8_t type; + BlueSkyCloudID id; + uint64_t inum; + uint32_t size1, size2, size3; +} __attribute__((packed)); + +// Rough size limit for a log segment. This is not a firm limit and there are +// no absolute guarantees on the size of a log segment. +#define LOG_SEGMENT_SIZE (1 << 22) + +#define JOURNAL_MAGIC "\nLog" +#define CLOUDLOG_MAGIC "AgI-" +#define CLOUDLOG_MAGIC_ENCRYPTED "AgI=" // CLOUDLOG_MAGIC[3] ^= 0x10 + /* 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 { +#define CLOUDLOG_UNCOMMITTED 0x10 +struct BlueSkyCloudLog { gint refcount; GMutex *lock; GCond *cond; @@ -201,6 +306,9 @@ struct _BlueSkyCloudLog { uint64_t inum; int32_t inum_offset; + // The size of encrypted object data, not including any headers + int data_size; + // The location of the object in the cloud, if available. BlueSkyCloudPointer location; @@ -220,25 +328,42 @@ struct _BlueSkyCloudLog { }; /* Serialize objects into a log segment to be written to the cloud. */ -struct _BlueSkyCloudLogState { +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 + + int uploads_pending; // Count of uploads in progress, not completed + GMutex *uploads_pending_lock; + GCond *uploads_pending_cond; + + /* What is the most recent sequence number written by the cleaner which we + * have processed and incorporated into our own log? This gets + * incorporated into the version vector written out with our checkpoint + * records. */ + int latest_cleaner_seq_seen; }; gboolean bluesky_cloudlog_equal(gconstpointer a, gconstpointer b); guint bluesky_cloudlog_hash(gconstpointer a); -BlueSkyCloudLog *bluesky_cloudlog_new(BlueSkyFS *fs); +BlueSkyCloudLog *bluesky_cloudlog_new(BlueSkyFS *fs, const BlueSkyCloudID *id); gchar *bluesky_cloudlog_id_to_string(BlueSkyCloudID id); BlueSkyCloudID bluesky_cloudlog_id_from_string(const gchar *idstr); +void bluesky_cloudlog_threads_init(BlueSkyFS *fs); void bluesky_cloudlog_ref(BlueSkyCloudLog *log); void bluesky_cloudlog_unref(BlueSkyCloudLog *log); +void bluesky_cloudlog_unref_delayed(BlueSkyCloudLog *log); +void bluesky_cloudlog_erase(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_insert_locked(BlueSkyCloudLog *log); +BlueSkyCloudLog *bluesky_cloudlog_get(BlueSkyFS *fs, BlueSkyCloudID id); +void bluesky_cloudlog_prefetch(BlueSkyCloudLog *log); void bluesky_cloudlog_fetch(BlueSkyCloudLog *log); +void bluesky_cloudlog_background_fetch(BlueSkyCloudLog *item); BlueSkyCloudPointer bluesky_cloudlog_serialize(BlueSkyCloudLog *log, BlueSkyFS *fs); void bluesky_cloudlog_flush(BlueSkyFS *fs); @@ -246,10 +371,11 @@ void bluesky_cloudlog_flush(BlueSkyFS *fs); /* Logging infrastructure for ensuring operations are persistently recorded to * disk. */ #define BLUESKY_CRC32C_SEED (~(uint32_t)0) +#define BLUESKY_CRC32C_VALIDATOR ((uint32_t)0xb798b438UL) uint32_t crc32c(uint32_t crc, const char *buf, unsigned int length); uint32_t crc32c_finalize(uint32_t crc); -struct _BlueSkyLog { +struct BlueSkyLog { BlueSkyFS *fs; char *log_directory; GAsyncQueue *queue; @@ -276,7 +402,7 @@ struct _BlueSkyLog { /* 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 { +struct BlueSkyCacheFile { GMutex *lock; GCond *cond; gint refcount; @@ -287,27 +413,41 @@ struct _BlueSkyCacheFile { gint mapcount; // References to the mmaped data const char *addr; // May be null if data is not mapped in memory size_t len; + int disk_used; BlueSkyFS *fs; BlueSkyLog *log; - gboolean fetching, ready; // Cloud data: downloading or ready for use + gboolean fetching; // Cloud data: downloading or ready for use + gboolean complete; // Complete file has been fetched from cloud int64_t atime; // Access time, for cache management + BlueSkyRangeset *items; // Locations of valid items + BlueSkyRangeset *prefetches;// Locations we have been requested to prefetch }; 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); +BlueSkyCloudLog *bluesky_log_get_commit_point(BlueSkyFS *fs); +void bluesky_log_write_commit_point(BlueSkyFS *fs, BlueSkyCloudLog *marker); + +BlueSkyRCStr *bluesky_cachefile_map_raw(BlueSkyCacheFile *cachefile, + off_t offset, size_t size); +BlueSkyRCStr *bluesky_log_map_object(BlueSkyCloudLog *item, gboolean map_data); void bluesky_mmap_unref(BlueSkyCacheFile *mmap); void bluesky_cachefile_unref(BlueSkyCacheFile *cachefile); BlueSkyCacheFile *bluesky_cachefile_lookup(BlueSkyFS *fs, - int clouddir, int log_seq); + int clouddir, int log_seq, + gboolean start_fetch); void bluesky_cachefile_gc(BlueSkyFS *fs); +void bluesky_replay(BlueSkyFS *fs); + /* Used to track log segments that are being written to the cloud. */ typedef struct { - BlueSkyRCStr *data; + BlueSkyFS *fs; + char *key; /* File name for log segment in backend */ + GString *raw_data; /* Data before encryption */ + BlueSkyRCStr *data; /* Data after encryption */ GSList *items; GMutex *lock; GCond *cond; @@ -321,14 +461,9 @@ typedef struct { 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. */ + /* A pointer to the cloud log entry for this inode. This may or may not + * actually have data loaded (it might just contain pointers to the data + * location, and in fact this will likely often be the case). */ BlueSkyCloudLog *item; } InodeMapEntry; @@ -340,8 +475,8 @@ typedef struct { /* 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; + /* The serialized version of the inode map data. */ + BlueSkyCloudLog *serialized; /* Have there been changes that require writing this section out again? */ gboolean dirty; @@ -349,6 +484,14 @@ typedef struct { InodeMapEntry *bluesky_inode_map_lookup(GSequence *inode_map, uint64_t inum, int action); +BlueSkyCloudLog *bluesky_inode_map_serialize(BlueSkyFS *fs); +void bluesky_inode_map_minimize(BlueSkyFS *fs); + +gboolean bluesky_checkpoint_load(BlueSkyFS *fs); + +/* Merging of log state with the work of the cleaner. */ +void bluesky_cleaner_merge(BlueSkyFS *fs); +void bluesky_cleaner_thread_launch(BlueSkyFS *fs); #ifdef __cplusplus }