* TODO: Licensing
*/
+#include <stdio.h>
#include <stdint.h>
+#include <inttypes.h>
#include <glib.h>
#include <string.h>
-#include "bluesky.h"
+#include "bluesky-private.h"
/* 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. */
+ * 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;
-}
-
-/* Compute the HMAC keyed-hash function using the given hash algorithm, data,
- * and key. */
-void compute_hmac(GChecksumType algo,
- const guchar *data, gsize data_len,
- const guchar *key, gsize key_len,
- guint8 *buffer, gsize *digest_len)
-{
- int block_size;
-
- switch (algo) {
- case G_CHECKSUM_MD5:
- case G_CHECKSUM_SHA1:
- case G_CHECKSUM_SHA256:
- block_size = 64;
- break;
- default:
- g_error("Unknown hash algorithm for HMAC: %d\n", algo);
- }
- gsize digest_size = g_checksum_type_get_length(algo);
+ if (inode->change_time == 0)
+ inode->change_time = now;
- guchar keybuf[block_size];
- memset(keybuf, 0, block_size);
- memcpy(keybuf, key, MIN(block_size, key_len));
- for (int i = 0; i < block_size; i++)
- keybuf[i] ^= 0x36;
+#if 0
+ if (bluesky_options.writethrough_cache)
+ bluesky_file_flush(inode, NULL);
+#endif
- GChecksum *csum1 = g_checksum_new(algo);
- g_checksum_update(csum1, keybuf, block_size);
- g_checksum_update(csum1, data, data_len);
- guint8 digest[digest_size];
- g_checksum_get_digest(csum1, digest, &digest_size);
-
- memset(keybuf, 0, block_size);
- memcpy(keybuf, key, MIN(block_size, key_len));
- for (int i = 0; i < block_size; i++)
- keybuf[i] ^= 0x5c;
-
- GChecksum *csum2 = g_checksum_new(algo);
- g_checksum_update(csum2, keybuf, block_size);
- g_checksum_update(csum2, digest, digest_size);
-
- g_checksum_get_digest(csum2, buffer, digest_len);
-
- g_checksum_free(csum1);
- g_checksum_free(csum2);
+ 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
fs->inodes = g_hash_table_new(bluesky_fs_key_hash_func,
bluesky_fs_key_equal_func);
fs->next_inum = BLUESKY_ROOT_INUM + 1;
- fs->store = s3store_new();
+ 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("");
return fs;
}
+BlueSkyFS *bluesky_init_fs(gchar *name, BlueSkyStore *store)
+{
+ BlueSkyFS *fs = bluesky_new_fs(name);
+ 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);
+ }
+
+ return fs;
+}
+
+/* Inode reference counting. */
+void bluesky_inode_ref(BlueSkyInode *inode)
+{
+ 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)) {
+ 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);
+ }
+}
+
/* Allocate a fresh inode number which has not been used before within a
- * filesystem. */
+ * filesystem. fs must already be locked. */
uint64_t bluesky_fs_alloc_inode(BlueSkyFS *fs)
{
uint64_t inum;
- g_mutex_lock(fs->lock);
inum = fs->next_inum;
fs->next_inum++;
- g_mutex_unlock(fs->lock);
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->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;
}
/* 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. */
+ * 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_fetch(fs, inum);
+ 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);
return inode;
}
-/* Insert an inode into the filesystem inode cache. */
+/* Insert an inode into the filesystem inode cache. fs should be locked. */
void bluesky_insert_inode(BlueSkyFS *fs, BlueSkyInode *inode)
{
- g_mutex_lock(fs->lock);
g_hash_table_insert(fs->inodes, &inode->inum, inode);
- g_mutex_unlock(fs->lock);
-}
-
-/* Mark a given block dirty and make sure that data is faulted in so that it
- * can be written to. */
-void bluesky_block_touch(BlueSkyInode *inode, uint64_t i)
-{
- g_return_if_fail(i < inode->blocks->len);
- BlueSkyBlock *block = &g_array_index(inode->blocks, BlueSkyBlock, i);
-
- switch (block->type) {
- case BLUESKY_BLOCK_ZERO:
- block->data = bluesky_string_new(g_malloc0(BLUESKY_BLOCK_SIZE),
- BLUESKY_BLOCK_SIZE);
- break;
- case BLUESKY_BLOCK_REF:
- bluesky_block_fetch(inode->fs, block);
- g_assert(block->type == BLUESKY_BLOCK_CACHED);
- /* Fall through */
- case BLUESKY_BLOCK_CACHED:
- case BLUESKY_BLOCK_DIRTY:
- block->data = bluesky_string_dup(block->data);
- break;
- }
-
- block->type = BLUESKY_BLOCK_DIRTY;
}
-/* Set the size of a file. This will truncate or extend the file as needed.
- * Newly-allocated bytes are zeroed. */
-void bluesky_file_truncate(BlueSkyInode *inode, uint64_t size)
+/* Start writeback of an inode and all associated data. */
+void bluesky_inode_start_sync(BlueSkyInode *inode)
{
- g_return_if_fail(size <= BLUESKY_MAX_FILE_SIZE);
+ GList *log_items = NULL;
- if (size == inode->size)
- return;
+ if (inode->type == BLUESKY_REGULAR)
+ bluesky_file_flush(inode, &log_items);
- uint64_t blocks = (size + BLUESKY_BLOCK_SIZE - 1) / BLUESKY_BLOCK_SIZE;
-
- if (blocks > inode->blocks->len) {
- /* Need to add new blocks to the end of a file. New block structures
- * are automatically zeroed, which initializes them to be pointers to
- * zero blocks so we don't need to do any more work. */
- g_array_set_size(inode->blocks, blocks);
- } else if (blocks < inode->blocks->len) {
- /* Delete blocks from a file. Must reclaim memory. */
- for (guint i = inode->blocks->len; i < blocks; i++) {
- BlueSkyBlock *b = &g_array_index(inode->blocks, BlueSkyBlock, i);
- g_free(b->ref);
- bluesky_string_unref(b->data);
- }
- g_array_set_size(inode->blocks, blocks);
- }
+ BlueSkyCloudLog *cloudlog = bluesky_serialize_inode(inode);
- /* If the file size is being decreased, ensure that any trailing data in
- * the last block is zeroed. */
- if (size < inode->size) {
- BlueSkyBlock *b = &g_array_index(inode->blocks, BlueSkyBlock,
- blocks - 1);
- if (b->type != BLUESKY_BLOCK_ZERO) {
- bluesky_block_touch(inode, blocks - 1);
- int end_offset = size % BLUESKY_BLOCK_SIZE;
- if (end_offset > 0) {
- memset(&b->data->data[end_offset], 0,
- BLUESKY_BLOCK_SIZE - end_offset);
- }
- }
- }
+ bluesky_cloudlog_unref(inode->committed_item);
+ inode->committed_item = cloudlog;
- inode->size = size;
- bluesky_inode_update_ctime(inode, 1);
-}
+ bluesky_cloudlog_sync(cloudlog);
+ bluesky_cloudlog_ref(cloudlog);
+ log_items = g_list_prepend(log_items, cloudlog);
-void bluesky_file_write(BlueSkyInode *inode, uint64_t offset,
- const char *data, gint len)
-{
- g_return_if_fail(inode->type == BLUESKY_REGULAR);
- g_return_if_fail(offset < inode->size);
- g_return_if_fail(len <= inode->size - offset);
+ /* Wait for all log items to be committed to disk. */
+ bluesky_log_finish_all(log_items);
- if (len == 0)
- return;
-
- while (len > 0) {
- uint64_t block_num = offset / BLUESKY_BLOCK_SIZE;
- gint block_offset = offset % BLUESKY_BLOCK_SIZE;
- gint bytes = MIN(BLUESKY_BLOCK_SIZE - block_offset, len);
+ /* 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;
- bluesky_block_touch(inode, block_num);
- BlueSkyBlock *b = &g_array_index(inode->blocks, BlueSkyBlock,
- block_num);
- memcpy(&b->data->data[block_offset], data, bytes);
- bluesky_block_flush(inode->fs, b);
-
- offset += bytes;
- data += bytes;
- len -= bytes;
- }
+ /* 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;
- bluesky_inode_update_ctime(inode, 1);
+ g_mutex_unlock(inode->fs->lock);
}
-void bluesky_file_read(BlueSkyInode *inode, uint64_t offset,
- char *buf, gint len)
+/* Write back an inode and all associated data and wait for completion. Inode
+ * should already be locked. */
+void bluesky_inode_do_sync(BlueSkyInode *inode)
{
- g_return_if_fail(inode->type == BLUESKY_REGULAR);
- g_return_if_fail(offset < inode->size);
- g_return_if_fail(len <= inode->size - offset);
-
- while (len > 0) {
- uint64_t block_num = offset / BLUESKY_BLOCK_SIZE;
- gint block_offset = offset % BLUESKY_BLOCK_SIZE;
- gint bytes = MIN(BLUESKY_BLOCK_SIZE - block_offset, len);
-
- BlueSkyBlock *b = &g_array_index(inode->blocks, BlueSkyBlock,
- block_num);
- switch (b->type) {
- case BLUESKY_BLOCK_ZERO:
- memset(buf, 0, bytes);
- break;
- case BLUESKY_BLOCK_REF:
- bluesky_block_fetch(inode->fs, b);
- /* Fall through */
- case BLUESKY_BLOCK_CACHED:
- case BLUESKY_BLOCK_DIRTY:
- memcpy(buf, &b->data->data[block_offset], bytes);
- break;
- }
-
- offset += bytes;
- buf += bytes;
- len -= bytes;
+ 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);
}
}
-/* Read the given block from cloud-backed storage if the data is not already
- * cached. */
-void bluesky_block_fetch(BlueSkyFS *fs, BlueSkyBlock *block)
+static void complete_inode_fetch(BlueSkyInode *inode)
{
- if (block->type != BLUESKY_BLOCK_REF)
- return;
-
- g_print("Fetching block from %s\n", block->ref);
- BlueSkyRCStr *string = s3store_get(fs->store, block->ref);
-
- bluesky_string_unref(block->data);
- block->data = string;
- block->type = BLUESKY_BLOCK_CACHED;
+ BlueSkyCloudLog *item = inode->committed_item;
+ inode->committed_item = NULL;
+ g_print("Completing fetch of inode %"PRIu64"...", 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);
}
-/* Write the given block to cloud-backed storage and mark it clean. */
-void bluesky_block_flush(BlueSkyFS *fs, BlueSkyBlock *block)
+/* 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)
{
- if (block->type != BLUESKY_BLOCK_DIRTY)
+ InodeMapEntry *entry = bluesky_inode_map_lookup(fs->inode_map, inum, 0);
+ if (entry == NULL)
return;
- GChecksum *csum = g_checksum_new(G_CHECKSUM_SHA256);
- g_checksum_update(csum, block->data->data, block->data->len);
- gchar *name = g_strdup(g_checksum_get_string(csum));
-
- g_print("Flushing block as %s\n", name);
- s3store_put(fs->store, name, block->data);
- g_free(block->ref);
- block->ref = name;
-
- /* block->type = BLUESKY_BLOCK_CACHED; */
- bluesky_string_unref(block->data);
- block->data = NULL;
- block->type = BLUESKY_BLOCK_REF;
-
- g_checksum_free(csum);
+ 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);
+ inode->committed_item = entry->item;
+ bluesky_cloudlog_ref(entry->item);
+
+ /* TODO: Thread pool or other better async method. */
+ g_thread_create((GThreadFunc)complete_inode_fetch,
+ (gpointer)inode, FALSE, NULL);
}