#include <glib.h>
#include <string.h>
-#include "bluesky.h"
+#include "bluesky-private.h"
+
+static void inode_fetch_task(gpointer a, gpointer b);
/* Core filesystem. Different proxies, such as the NFSv3 one, interface to
* this, but the core actually tracks the data which is stored. So far we just
/* Update an inode to indicate that a modification was made. This increases
* the change counter, updates the ctime to the current time, and optionally
- * updates the mtime. inode must already be locked. */
+ * updates the mtime. This also makes the inode contents subject to writeback
+ * to storage in the future. inode must already be locked. */
void bluesky_inode_update_ctime(BlueSkyInode *inode, gboolean update_mtime)
{
int64_t now = bluesky_get_current_time();
inode->ctime = now;
if (update_mtime)
inode->mtime = now;
+
+ if (inode->change_time == 0)
+ inode->change_time = now;
+
+#if 0
+ if (bluesky_options.writethrough_cache)
+ bluesky_file_flush(inode, NULL);
+#endif
+
+ g_mutex_lock(inode->fs->lock);
+ bluesky_list_unlink(&inode->fs->unlogged_list, inode->unlogged_list);
+ inode->unlogged_list = bluesky_list_prepend(&inode->fs->unlogged_list, inode);
+ bluesky_list_unlink(&inode->fs->accessed_list, inode->accessed_list);
+ inode->accessed_list = bluesky_list_prepend(&inode->fs->accessed_list, inode);
+ g_mutex_unlock(inode->fs->lock);
}
/* Unfortunately a glib hash table is only guaranteed to be able to store
bluesky_fs_key_equal_func);
fs->next_inum = BLUESKY_ROOT_INUM + 1;
fs->store = bluesky_store_new("file");
+ fs->flushd_lock = g_mutex_new();
+ fs->flushd_cond = g_cond_new();
+ fs->locations = g_hash_table_new(bluesky_cloudlog_hash,
+ bluesky_cloudlog_equal);
+ fs->inode_map = g_sequence_new(NULL);
+
+ fs->log_state = g_new0(BlueSkyCloudLogState, 1);
+ fs->log_state->data = g_string_new("");
+ fs->log_state->latest_cleaner_seq_seen = -1;
+ fs->log_state->uploads_pending_lock = g_mutex_new();
+ fs->log_state->uploads_pending_cond = g_cond_new();
+
+ bluesky_cloudlog_threads_init(fs);
+ fs->inode_fetch_thread_pool = g_thread_pool_new(inode_fetch_task, NULL,
+ bluesky_max_threads,
+ FALSE, NULL);
+
+ return fs;
+}
+
+BlueSkyFS *bluesky_init_fs(gchar *name, BlueSkyStore *store,
+ const gchar *master_key)
+{
+ BlueSkyFS *fs = bluesky_new_fs(name);
+ fs->master_key = g_strdup(master_key);
+ fs->keys = g_new(BlueSkyCryptKeys, 1);
+ bluesky_crypt_derive_keys(fs->keys, master_key);
+ fs->store = store;
+ fs->log = bluesky_log_new("journal");
+ fs->log->fs = fs;
+
+ if (bluesky_checkpoint_load(fs)) {
+ g_print("Filesystem checkpoint loaded, starting journal replay...\n");
+ bluesky_replay(fs);
+ g_print("Journal replay complete, filesystem ready.\n");
+ } else {
+ /* Initialize a fresh filesystem */
+ g_print("Initializing new filesystem...\n");
+ BlueSkyInode *root = bluesky_new_inode(BLUESKY_ROOT_INUM, fs,
+ BLUESKY_DIRECTORY);
+ root->nlink = 1;
+ root->mode = 0755;
+ bluesky_insert_inode(fs, root);
+ bluesky_inode_update_ctime(root, TRUE);
+ bluesky_inode_do_sync(root);
+ }
+
+ bluesky_cleaner_thread_launch(fs);
return fs;
}
g_atomic_int_inc(&inode->refcount);
}
+/* Free most of the resources used by an inode structure, but do not free the
+ * inode itself. Can be used if the inode data will be reloaded from
+ * serialized form to clear out old information first. */
+void bluesky_inode_free_resources(BlueSkyInode *inode)
+{
+ switch (inode->type) {
+ case BLUESKY_REGULAR:
+ if (inode->blocks != NULL) {
+ for (int i = 0; i < inode->blocks->len; i++) {
+ BlueSkyBlock *b = &g_array_index(inode->blocks,
+ BlueSkyBlock, i);
+ if (b->type == BLUESKY_BLOCK_DIRTY) {
+ g_error("Deleting an inode with dirty file data!");
+ }
+ bluesky_cloudlog_unref(b->ref);
+ bluesky_string_unref(b->dirty);
+ }
+ g_array_unref(inode->blocks);
+ inode->blocks = NULL;
+ }
+ break;
+
+ case BLUESKY_DIRECTORY:
+ if (inode->dirhash != NULL)
+ g_hash_table_destroy(inode->dirhash);
+ inode->dirhash = NULL;
+ if (inode->dirhash_folded != NULL)
+ g_hash_table_destroy(inode->dirhash_folded);
+ inode->dirhash_folded = NULL;
+ if (inode->dirents != NULL)
+ g_sequence_free(inode->dirents);
+ inode->dirents = NULL;
+ break;
+
+ case BLUESKY_SYMLINK:
+ g_free(inode->symlink_contents);
+ inode->symlink_contents = NULL;
+ break;
+
+ default:
+ break;
+ }
+}
+
void bluesky_inode_unref(BlueSkyInode *inode)
{
if (g_atomic_int_dec_and_test(&inode->refcount)) {
- g_error("Reference count for inode %"PRIu64" dropped to zero!\n",
- inode->inum);
+ if (bluesky_verbose) {
+ g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
+ "Reference count for inode %"PRIu64" dropped to zero.",
+ inode->inum);
+ }
+
+ /* Sanity check: Is the inode clean? */
+ if (inode->change_commit < inode->change_count
+ || inode->accessed_list != NULL
+ || inode->unlogged_list != NULL
+ || inode->dirty_list != NULL) {
+ g_warning("Dropping inode which is not clean (commit %"PRIi64" < change %"PRIi64"; accessed_list = %p; dirty_list = %p)\n", inode->change_commit, inode->change_count, inode->accessed_list, inode->dirty_list);
+ }
+
+ /* These shouldn't be needed, but in case the above warning fires and
+ * we delete the inode anyway, we ought to be sure the inode is not on
+ * any LRU list. */
+ g_mutex_lock(inode->fs->lock);
+ bluesky_list_unlink(&inode->fs->accessed_list, inode->accessed_list);
+ bluesky_list_unlink(&inode->fs->dirty_list, inode->dirty_list);
+ bluesky_list_unlink(&inode->fs->unlogged_list, inode->unlogged_list);
+ g_mutex_unlock(inode->fs->lock);
+
+ bluesky_inode_free_resources(inode);
+
+ g_mutex_free(inode->lock);
+ g_free(inode);
}
}
return inum;
}
-BlueSkyInode *bluesky_new_inode(uint64_t inum, BlueSkyFS *fs,
- BlueSkyFileType type)
+/* Perform type-specification initialization of an inode. Normally performed
+ * in bluesky_new_inode, but can be separated if an inode is created first,
+ * then deserialized. */
+void bluesky_init_inode(BlueSkyInode *i, BlueSkyFileType type)
{
- BlueSkyInode *i = g_new0(BlueSkyInode, 1);
-
- i->lock = g_mutex_new();
- i->refcount = 1;
i->type = type;
- i->fs = fs;
- i->inum = inum;
switch (type) {
case BLUESKY_REGULAR:
case BLUESKY_DIRECTORY:
i->dirents = g_sequence_new(bluesky_dirent_destroy);
i->dirhash = g_hash_table_new(g_str_hash, g_str_equal);
+ i->dirhash_folded = g_hash_table_new(g_str_hash, g_str_equal);
break;
- case BLUESKY_BLOCK:
- case BLUESKY_CHARACTER:
- case BLUESKY_SYMLINK:
- case BLUESKY_SOCKET:
- case BLUESKY_FIFO:
+ default:
break;
}
+}
+
+BlueSkyInode *bluesky_new_inode(uint64_t inum, BlueSkyFS *fs,
+ BlueSkyFileType type)
+{
+ BlueSkyInode *i = g_new0(BlueSkyInode, 1);
+
+ i->lock = g_mutex_new();
+ i->refcount = 1;
+ i->fs = fs;
+ i->inum = inum;
+ i->change_count = 1;
+ bluesky_init_inode(i, type);
return i;
}
+/* Issue a prefetch hint for an inode. This signals that the inode may be
+ * needed soon. Does not return any useful data. */
+void bluesky_inode_prefetch(BlueSkyFS *fs, uint64_t inum)
+{
+ BlueSkyInode *inode = NULL;
+
+ g_mutex_lock(fs->lock);
+ inode = (BlueSkyInode *)g_hash_table_lookup(fs->inodes, &inum);
+
+ if (inode != NULL) {
+ /* Inode is already available, no need for any prefetching... */
+ g_mutex_unlock(fs->lock);
+ return;
+ }
+
+ InodeMapEntry *entry = bluesky_inode_map_lookup(fs->inode_map, inum, 0);
+ if (entry != NULL) {
+ bluesky_cloudlog_prefetch(entry->item);
+ }
+
+ g_mutex_unlock(fs->lock);
+ return;
+}
+
/* Retrieve an inode from the filesystem. Eventually this will be a cache and
* so we might need to go fetch the inode from elsewhere; for now all
- * filesystem state is stored here. inode is returned locked with a reference
- * held. */
+ * filesystem state is stored here. inode is returned with a reference held
+ * but not locked. */
BlueSkyInode *bluesky_get_inode(BlueSkyFS *fs, uint64_t inum)
{
BlueSkyInode *inode = NULL;
+ if (inum == 0) {
+ return NULL;
+ }
+
g_mutex_lock(fs->lock);
inode = (BlueSkyInode *)g_hash_table_lookup(fs->inodes, &inum);
if (inode != NULL) {
bluesky_inode_ref(inode);
+
+ /* FIXME: We assume we can atomically update the in-memory access time
+ * without a lock. */
+ inode->access_time = bluesky_get_current_time();
}
g_mutex_unlock(fs->lock);
g_hash_table_insert(fs->inodes, &inode->inum, inode);
}
-/* Synchronize an inode to stable storage. */
-void bluesky_inode_flush(BlueSkyFS *fs, BlueSkyInode *inode)
+/* Start writeback of an inode and all associated data. */
+void bluesky_inode_start_sync(BlueSkyInode *inode)
{
- GString *buf = g_string_new("");
- bluesky_serialize_inode(buf, inode);
+ GList *log_items = NULL;
- gsize len = buf->len;
- BlueSkyRCStr *data = bluesky_string_new(g_string_free(buf, FALSE), len);
+ if (inode->type == BLUESKY_REGULAR)
+ bluesky_file_flush(inode, &log_items);
- char key[64];
- sprintf(key, "inode-%016"PRIx64, inode->inum);
+ BlueSkyCloudLog *cloudlog = bluesky_serialize_inode(inode);
- bluesky_store_put(fs->store, key, data);
+ bluesky_cloudlog_unref(inode->committed_item);
+ inode->committed_item = cloudlog;
+
+ bluesky_cloudlog_sync(cloudlog);
+ bluesky_cloudlog_ref(cloudlog);
+ log_items = g_list_prepend(log_items, cloudlog);
+
+ /* Wait for all log items to be committed to disk. */
+ bluesky_log_finish_all(log_items);
+
+ /* Mark the inode as clean */
+ inode->change_commit = inode->change_count;
+ inode->change_time = 0;
+ g_mutex_lock(inode->fs->lock);
+ bluesky_list_unlink(&inode->fs->unlogged_list, inode->unlogged_list);
+ inode->unlogged_list = NULL;
+
+ /* Since a new version of the inode has been written to the log, also
+ * schedule a future flush of the new data to cloud storage. */
+ bluesky_list_unlink(&inode->fs->dirty_list, inode->dirty_list);
+ inode->dirty_list = bluesky_list_prepend(&inode->fs->dirty_list, inode);
+ inode->change_cloud = inode->change_count;
+
+ g_mutex_unlock(inode->fs->lock);
+}
+
+/* Write back an inode and all associated data and wait for completion. Inode
+ * should already be locked. */
+void bluesky_inode_do_sync(BlueSkyInode *inode)
+{
+ if (bluesky_verbose) {
+ g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
+ "Synchronous writeback for inode %"PRIu64"...", inode->inum);
+ }
+ bluesky_inode_start_sync(inode);
+ if (bluesky_verbose) {
+ g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
+ "Writeback for inode %"PRIu64" complete", inode->inum);
+ }
}
-/* Fetch an inode from stable storage. */
+static void inode_fetch_task(gpointer a, gpointer b)
+{
+ BlueSkyInode *inode = (BlueSkyInode *)a;
+
+ bluesky_profile_set((BlueSkyProfile *)inode->private_data);
+
+ BlueSkyCloudLog *item = inode->committed_item;
+ inode->committed_item = NULL;
+ g_print("Completing fetch of inode %"PRIu64"...\n", inode->inum);
+
+ g_mutex_lock(item->lock);
+ bluesky_cloudlog_fetch(item);
+ if (!bluesky_deserialize_inode(inode, item))
+ g_print("Error deserializing inode %"PRIu64"\n", inode->inum);
+ g_mutex_unlock(item->lock);
+
+ inode->access_time = bluesky_get_current_time();
+ g_mutex_lock(inode->fs->lock);
+ bluesky_list_unlink(&inode->fs->accessed_list, inode->accessed_list);
+ inode->accessed_list = bluesky_list_prepend(&inode->fs->accessed_list, inode);
+ g_mutex_unlock(inode->fs->lock);
+
+ g_mutex_unlock(inode->lock);
+ bluesky_cloudlog_unref(item);
+ bluesky_inode_unref(inode);
+}
+
+/* Fetch an inode from stable storage. The fetch can be performed
+ * asynchronously: the in-memory inode is allocated, but not filled with data
+ * immediately. It is kept locked until it has been filled in, so any users
+ * should try to acquire the lock on the inode before accessing any data. The
+ * fs lock must be held. */
void bluesky_inode_fetch(BlueSkyFS *fs, uint64_t inum)
{
- char key[64];
- sprintf(key, "inode-%016"PRIx64, inum);
- BlueSkyRCStr *data = bluesky_store_get(fs->store, key);
- if (data == NULL)
+ InodeMapEntry *entry = bluesky_inode_map_lookup(fs->inode_map, inum, 0);
+ if (entry == NULL)
return;
- BlueSkyInode *inode = bluesky_deserialize_inode(fs, data->data);
- if (inode != NULL) {
- bluesky_insert_inode(fs, inode);
- g_print("Loaded inode %"PRIu64"\n", inum);
- }
+ /* Non-portable behavior: We take the inode lock here, and release it in
+ * the fetching thread. This works with the default Linux pthreads
+ * implementation but is not guaranteed. */
+
+ BlueSkyInode *inode = bluesky_new_inode(inum, fs, BLUESKY_PENDING);
+ inode->change_count = 0;
+ bluesky_inode_ref(inode); // Extra ref held by fetching process
+ g_mutex_lock(inode->lock);
+
+ inode->committed_item = entry->item;
+ bluesky_cloudlog_ref(entry->item);
+ bluesky_insert_inode(fs, inode);
+
+ inode->private_data = bluesky_profile_get();
+ g_thread_pool_push(fs->inode_fetch_thread_pool, inode, NULL);
}