Partial work on parallel data fetches from S3 for large reads.

[bluesky.git] / bluesky / bluesky.h

diff --git a/bluesky/bluesky.h b/bluesky/bluesky.h
index 201cc17..e30f5b0 100644 (file)
--- a/bluesky/bluesky.h
+++ b/bluesky/bluesky.h
@@ -24,6 +24,15 @@ typedef struct {
 
     /* 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;
@@ -87,6 +96,10 @@ 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
@@ -121,6 +134,13 @@ bluesky_time_hires bluesky_now_hires();
  *     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;
@@ -149,6 +169,9 @@ typedef struct {
      * 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 */
 
@@ -208,6 +231,7 @@ 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);
@@ -223,8 +247,6 @@ 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);
@@ -238,6 +260,7 @@ gint bluesky_dirent_compare(gconstpointer a, gconstpointer b,
                             gpointer unused);
 
 void bluesky_flushd_invoke(BlueSkyFS *fs);
+void bluesky_inode_do_sync(BlueSkyInode *inode);
 
 void bluesky_debug_dump(BlueSkyFS *fs);