#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
* implement an in-memory filesystem, but eventually this will be state which
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->dirty_list, inode->dirty_list);
- inode->dirty_list = bluesky_list_prepend(&inode->fs->dirty_list, inode);
+ 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);
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("");
+
+ 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)
+BlueSkyFS *bluesky_init_fs(gchar *name, BlueSkyStore *store,
+ const gchar *master_key)
{
- BlueSkyRCStr *data = bluesky_store_get(store, "superblock");
- if (data != NULL) {
- BlueSkyFS *fs = bluesky_deserialize_superblock(data->data);
- if (fs != NULL) {
- fs->store = store;
- g_print("Loaded filesystem superblock\n");
- g_free(fs->name);
- fs->name = g_strdup(name);
- return fs;
- }
- bluesky_string_unref(data);
- }
-
- g_print("Initializing fresh filesystem\n");
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;
-
- 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_flush(fs, root);
- bluesky_superblock_flush(fs);
+ 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);
+ }
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_print("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);
}
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);
- /* Free file type specific data. It should be an error for there to be
- * dirty data to commit when the reference count has reaches zero. */
- switch (inode->type) {
- case BLUESKY_REGULAR:
- 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!");
- }
- g_free(b->ref);
- bluesky_string_unref(b->data);
- }
- g_array_unref(inode->blocks);
- break;
-
- case BLUESKY_DIRECTORY:
- g_hash_table_destroy(inode->dirhash);
- g_hash_table_destroy(inode->dirhash_folded);
- g_sequence_free(inode->dirents);
- break;
-
- case BLUESKY_SYMLINK:
- g_free(inode->symlink_contents);
- break;
-
- default:
- break;
- }
+ bluesky_inode_free_resources(inode);
g_mutex_free(inode->lock);
-
g_free(inode);
}
}
inum = fs->next_inum;
fs->next_inum++;
- bluesky_superblock_flush(fs);
-
return inum;
}
g_hash_table_insert(fs->inodes, &inode->inum, inode);
}
-/* Deprecated: Synchronize an inode to stable storage. */
-void bluesky_inode_flush(BlueSkyFS *fs, BlueSkyInode *inode)
-{
- GString *buf = g_string_new("");
- bluesky_serialize_inode(buf, inode);
- BlueSkyRCStr *data = bluesky_string_new_from_gstring(buf);
-
- char key[64];
- sprintf(key, "inode-%016"PRIx64, inode->inum);
-
- BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
- async->op = STORE_OP_PUT;
- async->key = g_strdup(key);
- async->data = data;
- bluesky_store_async_submit(async);
- bluesky_store_async_unref(async);
-}
-
/* Start writeback of an inode and all associated data. */
-void bluesky_inode_start_sync(BlueSkyInode *inode, BlueSkyStoreAsync *barrier)
+void bluesky_inode_start_sync(BlueSkyInode *inode)
{
- BlueSkyFS *fs = inode->fs;
+ GList *log_items = NULL;
if (inode->type == BLUESKY_REGULAR)
- bluesky_file_flush(inode, barrier);
-
- GString *buf = g_string_new("");
- bluesky_serialize_inode(buf, inode);
- BlueSkyRCStr *data = bluesky_string_new_from_gstring(buf);
-
- char key[64];
- sprintf(key, "inode-%016"PRIx64, inode->inum);
-
- BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
- async->op = STORE_OP_PUT;
- async->key = g_strdup(key);
- async->data = data;
- bluesky_store_async_submit(async);
- if (barrier != NULL)
- bluesky_store_add_barrier(barrier, async);
- bluesky_store_async_unref(async);
+ bluesky_file_flush(inode, &log_items);
+
+ BlueSkyCloudLog *cloudlog = bluesky_serialize_inode(inode);
+
+ 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)
{
- BlueSkyStoreAsync *barrier = bluesky_store_async_new(inode->fs->store);
- barrier->op = STORE_OP_BARRIER;
-
- g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
- "Synchronous writeback for inode %"PRIu64"...", inode->inum);
- bluesky_inode_start_sync(inode, barrier);
- bluesky_store_async_submit(barrier);
- bluesky_store_async_wait(barrier);
- bluesky_store_async_unref(barrier);
- g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
- "Writeback for inode %"PRIu64" complete", inode->inum);
+ 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);
+ }
}
-static void complete_inode_fetch(BlueSkyStoreAsync *async, BlueSkyInode *inode)
+static void inode_fetch_task(gpointer a, gpointer b)
{
+ BlueSkyInode *inode = (BlueSkyInode *)a;
+
+ BlueSkyCloudLog *item = inode->committed_item;
+ inode->committed_item = NULL;
g_print("Completing fetch of inode %"PRIu64"...\n", inode->inum);
- if (async->result != 0
- || !bluesky_deserialize_inode(inode, async->data->data))
- {
- g_print(" failed to load inode, cleaning up\n");
- g_mutex_lock(inode->fs->lock);
- g_hash_table_remove(inode->fs->inodes, &inode->inum);
- g_mutex_unlock(inode->fs->lock);
- bluesky_inode_unref(inode);
- }
+ 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);
}
* fs lock must be held. */
void bluesky_inode_fetch(BlueSkyFS *fs, uint64_t inum)
{
- char key[64];
- sprintf(key, "inode-%016"PRIx64, inum);
+ InodeMapEntry *entry = bluesky_inode_map_lookup(fs->inode_map, inum, 0);
+ if (entry == NULL)
+ return;
+
+ /* 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);
- bluesky_insert_inode(fs, inode);
-
- BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
- async->op = STORE_OP_GET;
- async->key = g_strdup(key);
-
- bluesky_store_async_add_notifier(async, (GFunc)complete_inode_fetch, inode);
- bluesky_store_async_submit(async);
- if (bluesky_options.sync_inode_fetches) {
- bluesky_store_async_wait(async);
- }
-
- bluesky_store_async_unref(async);
-}
+ inode->committed_item = entry->item;
+ bluesky_cloudlog_ref(entry->item);
+ bluesky_insert_inode(fs, inode);
-/* Synchronize filesystem superblock to stable storage. */
-void bluesky_superblock_flush(BlueSkyFS *fs)
-{
- GString *buf = g_string_new("");
- bluesky_serialize_superblock(buf, fs);
- BlueSkyRCStr *data = bluesky_string_new_from_gstring(buf);
-
- BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
- async->op = STORE_OP_PUT;
- async->key = g_strdup("superblock");
- async->data = data;
- bluesky_store_async_submit(async);
- bluesky_store_async_unref(async);
-
- bluesky_store_sync(fs->store);
+ g_thread_pool_push(fs->inode_fetch_thread_pool, inode, NULL);
}