#define _BLUESKY_H
#include <stdint.h>
+#include <inttypes.h>
#include <glib.h>
#ifdef __cplusplus
extern "C" {
#endif
-struct S3Store;
+/* Various options to tweak for performance benchmarking purposes. */
+typedef struct {
+ /* Perform all get/put operations synchronously. */
+ int synchronous_stores;
+
+ /* 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;
+} 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 {
+ BSTATUS_OK = 0, /* No error */
+ BSTATUS_IOERR, /* I/O error of some form */
+ BSTATUS_NOENT, /* File does not exist */
+} BlueSkyStatus;
+
+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. */
} BlueSkyRCStr;
BlueSkyRCStr *bluesky_string_new(gpointer data, gsize len);
+BlueSkyRCStr *bluesky_string_new_from_gstring(GString *s);
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;
+
+void bluesky_store_init();
+BlueSkyStore *bluesky_store_new(const gchar *type);
+void bluesky_store_free(BlueSkyStore *store);
+BlueSkyRCStr *bluesky_store_get(BlueSkyStore *store, const gchar *key);
+void bluesky_store_put(BlueSkyStore *store,
+ const gchar *key, BlueSkyRCStr *val);
+
/* File types. The numeric values are chosen to match with those used in
* NFSv3. */
typedef enum {
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
GHashTable *inodes; /* Cached inodes */
uint64_t next_inum; /* Next available inode for allocation */
- struct S3Store *store;
+ BlueSkyStore *store;
+
+ /* Accounting for memory used for caches. Space is measured in blocks, not
+ * bytes. We track both total data in the caches and dirty data (total
+ * data includes dirty data). Updates to these variables must be made
+ * atomically. */
+ gint cache_total, cache_dirty;
- uint8_t *encryption_key;
+ /* 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 dirty_list, accessed_list;
+
+ /* Mutex for the flush daemon, to prevent concurrent execution. */
+ GMutex *flushd_lock;
} BlueSkyFS;
/* Inode number of the root directory. */
/* Timestamp, measured in microseconds since the Unix epoch. */
typedef int64_t bluesky_time;
+/* High-resolution timer, measured in nanoseconds. */
+typedef int64_t bluesky_time_hires;
+bluesky_time_hires bluesky_now_hires();
+
/* In-memory representation of an inode within a Blue Sky server. This
* corresponds roughly with information that is committed to persistent
* storage. Locking/refcounting rules:
* 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;
* that we don't exhaust the identifier space. */
uint64_t inum;
- uint64_t change_count; /* Incremented each with each change made */
+ /* 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;
+
+ /* 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;
+
+ /* Last access time to this inode, for controlling cache evictions. */
+ int64_t access_time;
+
+ /* Additional state for tracking cache writeback status. */
+ uint64_t change_pending; /* change_count version currently being committed to storage */
+
+ /* 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 *accessed_list, *dirty_list;
+
int64_t atime; /* Microseconds since the Unix epoch */
int64_t ctime;
int64_t mtime;
/* Directory-specific fields */
GSequence *dirents; /* List of entries for READDIR */
GHashTable *dirhash; /* Hash table by name for LOOKUP */
+ GHashTable *dirhash_folded; /* As above, but case-folded */
uint64_t parent_inum; /* inode for ".."; 0 if the root directory */
+
+ /* Symlink-specific fields */
+ gchar *symlink_contents;
} BlueSkyInode;
/* A directory entry. The name is UTF-8 and is a freshly-allocated string.
* is used to provide a stable key for restarting a READDIR call. */
typedef struct {
gchar *name;
+ gchar *name_folded; /* Name, folded for case-insensitive lookup */
uint32_t cookie;
uint64_t inum;
} BlueSkyDirent;
BlueSkyRCStr *data; /* Pointer to data in memory if cached */
} BlueSkyBlock;
-BlueSkyFS *bluesky_new_fs(gchar *name);
+BlueSkyFS *bluesky_init_fs(gchar *name, BlueSkyStore *store);
+void bluesky_superblock_flush(BlueSkyFS *fs);
+
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);
+void bluesky_inode_ref(BlueSkyInode *inode);
+void bluesky_inode_unref(BlueSkyInode *inode);
void bluesky_insert_inode(BlueSkyFS *fs, BlueSkyInode *inode);
void bluesky_dirent_destroy(gpointer dirent);
-uint64_t bluesky_directory_hash(gchar *name);
uint64_t bluesky_directory_lookup(BlueSkyInode *inode, gchar *name);
-gboolean bluesky_directory_insert(BlueSkyInode *dir, gchar *name,
+uint64_t bluesky_directory_ilookup(BlueSkyInode *inode, gchar *name);
+BlueSkyDirent *bluesky_directory_read(BlueSkyInode *dir, uint32_t cookie);
+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_block_flush(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);
-struct S3Store *s3store_new();
-BlueSkyRCStr *s3store_get(struct S3Store *store, const gchar *key);
-void s3store_put(struct S3Store *store, const gchar *key, BlueSkyRCStr *val);
+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_debug_dump(BlueSkyFS *fs);
#ifdef __cplusplus
}
projects / bluesky.git / blobdiff