+ if (bluesky_verbose || async->result != 0)
+ g_print("Fetch of %s from cloud complete, status = %d\n",
+ async->key, async->result);
+
+ g_mutex_lock(cachefile->lock);
+ if (async->result >= 0) {
+ if (async->len == 0) {
+ if (bluesky_verbose)
+ g_print("Complete object was fetched.\n");
+ cachefile->complete = TRUE;
+ }
+
+ /* Descrypt items fetched and write valid items out to the local log,
+ * but only if they do not overlap existing objects. This will protect
+ * against an attack by the cloud provider where one valid object is
+ * moved to another offset and used to overwrite data that we already
+ * have fetched. */
+ BlueSkyRangeset *items = bluesky_rangeset_new();
+ int fd = openat(cachefile->log->dirfd, cachefile->filename, O_WRONLY);
+ if (fd >= 0) {
+ gboolean allow_unauth;
+ async->data = bluesky_string_dup(async->data);
+ allow_unauth = cachefile->log_dir == BLUESKY_CLOUD_DIR_CLEANER;
+ bluesky_cloudlog_decrypt(async->data->data, async->data->len,
+ cachefile->fs->keys, items, allow_unauth);
+ uint64_t item_offset = 0;
+ while (TRUE) {
+ const BlueSkyRangesetItem *item;
+ item = bluesky_rangeset_lookup_next(items, item_offset);
+ if (item == NULL)
+ break;
+ if (bluesky_verbose) {
+ g_print(" item offset from range request: %d\n",
+ (int)(item->start + async->start));
+ }
+ if (bluesky_rangeset_insert(cachefile->items,
+ async->start + item->start,
+ item->length, item->data))
+ {
+ robust_pwrite(fd, async->data->data + item->start,
+ item->length, async->start + item->start);
+ } else {
+ g_print(" item overlaps existing data!\n");
+ }
+ item_offset = item->start + 1;
+ }
+ /* TODO: Iterate over items and merge into cached file. */
+ close(fd);
+ } else {
+ g_warning("Unable to open and write to cache file %s: %m",
+ cachefile->filename);
+ }
+
+ bluesky_rangeset_free(items);
+ } else {
+ g_print("Error fetching %s from cloud, retrying...\n", async->key);
+ cloudlog_partial_fetch_start(cachefile, async->start, async->len);
+ }
+
+ /* Update disk-space usage statistics, since the writes above may have
+ * consumed more space. */
+ g_atomic_int_add(&cachefile->log->disk_used, -cachefile->disk_used);
+ struct stat statbuf;
+ if (fstatat(cachefile->log->dirfd, cachefile->filename, &statbuf, 0) >= 0) {
+ /* Convert from 512-byte blocks to 1-kB units */
+ cachefile->disk_used = (statbuf.st_blocks + 1) / 2;
+ }
+ g_atomic_int_add(&cachefile->log->disk_used, cachefile->disk_used);
+
+ bluesky_cachefile_unref(cachefile);
+ g_cond_broadcast(cachefile->cond);
+ g_mutex_unlock(cachefile->lock);
+}
+
+static void cloudlog_fetch_start(BlueSkyCacheFile *cachefile)
+{
+ g_atomic_int_inc(&cachefile->refcount);
+ cachefile->fetching = TRUE;
+ if (bluesky_verbose)
+ g_print("Starting fetch of %s from cloud\n", cachefile->filename);
+ BlueSkyStoreAsync *async = bluesky_store_async_new(cachefile->fs->store);
+ async->op = STORE_OP_GET;
+ async->key = g_strdup(cachefile->filename);
+ async->profile = bluesky_profile_get();
+ bluesky_store_async_add_notifier(async,
+ (GFunc)cloudlog_partial_fetch_complete,
+ cachefile);
+ bluesky_store_async_submit(async);
+ bluesky_store_async_unref(async);
+}
+
+/* Map and return a read-only version of a byte range from a cached file. The
+ * CacheFile object must be locked. */
+BlueSkyRCStr *bluesky_cachefile_map_raw(BlueSkyCacheFile *cachefile,
+ off_t offset, size_t size)
+{
+ cachefile->atime = bluesky_get_current_time();
+
+ /* Easy case: the needed data is already in memory */
+ if (cachefile->addr != NULL && offset + size <= cachefile->len)
+ return bluesky_string_new_from_mmap(cachefile, offset, size);
+
+ int fd = openat(cachefile->log->dirfd, cachefile->filename, O_RDONLY);
+ if (fd < 0) {
+ fprintf(stderr, "Error opening logfile %s: %m\n",
+ cachefile->filename);
+ return NULL;
+ }
+
+ off_t length = lseek(fd, 0, SEEK_END);
+ if (offset + size > length) {
+ close(fd);
+ return NULL;
+ }
+
+ /* File is not mapped in memory. Map the entire file in, then return a
+ * pointer to just the required data. */
+ if (cachefile->addr == NULL) {
+ cachefile->addr = (const char *)mmap(NULL, length, PROT_READ,
+ MAP_SHARED, fd, 0);
+ cachefile->len = length;
+ g_atomic_int_inc(&cachefile->refcount);
+
+ close(fd);
+ return bluesky_string_new_from_mmap(cachefile, offset, size);
+ }
+
+ /* Otherwise, the file was mapped in but doesn't cover the data we need.
+ * This shouldn't happen much, if at all, but if it does just read the data
+ * we need directly from the file. We lose memory-management benefits of
+ * using mmapped data, but otherwise this works. */
+ char *buf = g_malloc(size);
+ size_t actual_size = readbuf(fd, buf, size);
+ close(fd);
+ if (actual_size != size) {
+ g_free(buf);
+ return NULL;
+ } else {
+ return bluesky_string_new(buf, size);
+ }
+}
+
+/* The arguments are mostly straightforward. log_dir is -1 for access from the
+ * journal, and non-negative for access to a cloud log segment. map_data
+ * should be TRUE for the case that are mapping just the data of an item where
+ * we have already parsed the item headers; this surpresses the error when the
+ * access is not to the first bytes of the item. */
+BlueSkyRCStr *bluesky_log_map_object(BlueSkyCloudLog *item, gboolean map_data)
+{
+ BlueSkyFS *fs = item->fs;
+ BlueSkyCacheFile *map = NULL;
+ BlueSkyRCStr *str = NULL;
+ int location = 0;
+ size_t file_offset = 0, file_size = 0;
+ gboolean range_request = bluesky_options.full_segment_fetches
+ ? FALSE : TRUE;
+
+ if (page_size == 0) {
+ page_size = getpagesize();
+ }
+
+ bluesky_cloudlog_stats_update(item, -1);
+
+ /* First, check to see if the journal still contains a copy of the item and
+ * if so use that. */
+ if ((item->location_flags | item->pending_write) & CLOUDLOG_JOURNAL) {
+ map = bluesky_cachefile_lookup(fs, -1, item->log_seq, TRUE);
+ if (map != NULL) {
+ location = CLOUDLOG_JOURNAL;
+ file_offset = item->log_offset;
+ file_size = item->log_size;
+ }
+ }
+
+ if (location == 0 && (item->location_flags & CLOUDLOG_CLOUD)) {
+ item->location_flags &= ~CLOUDLOG_JOURNAL;
+ map = bluesky_cachefile_lookup(fs,
+ item->location.directory,
+ item->location.sequence,
+ !range_request);
+ if (map == NULL) {
+ g_warning("Unable to remap cloud log segment!");
+ goto exit1;
+ }
+ location = CLOUDLOG_CLOUD;
+ file_offset = item->location.offset;
+ file_size = item->location.size;
+ }
+
+ /* Log segments fetched from the cloud might only be partially-fetched.
+ * Check whether the object we are interested in is available. */
+ if (location == CLOUDLOG_CLOUD) {
+ while (TRUE) {
+ const BlueSkyRangesetItem *rangeitem;
+ rangeitem = bluesky_rangeset_lookup(map->items, file_offset);
+ if (rangeitem != NULL && (rangeitem->start != file_offset
+ || rangeitem->length != file_size)) {
+ g_warning("log-%d: Item offset %zd seems to be invalid!",
+ (int)item->location.sequence, file_offset);
+ goto exit2;
+ }
+ if (rangeitem == NULL) {
+ if (bluesky_verbose) {
+ g_print("Item at offset 0x%zx not available, need to fetch.\n",
+ file_offset);
+ }
+ if (range_request) {
+ uint64_t start = file_offset, length = file_size, end;
+ if (map->prefetches != NULL)
+ bluesky_rangeset_get_extents(map->prefetches,
+ &start, &length);
+ start = MIN(start, file_offset);
+ end = MAX(start + length, file_offset + file_size);
+ length = end - start;
+ cloudlog_partial_fetch_start(map, start, length);
+ if (map->prefetches != NULL) {
+ bluesky_rangeset_free(map->prefetches);
+ map->prefetches = NULL;
+ }
+ }
+ g_cond_wait(map->cond, map->lock);
+ } else if (rangeitem->start == file_offset
+ && rangeitem->length == file_size) {
+ if (bluesky_verbose)
+ g_print("Item %zd now available.\n", file_offset);
+ break;
+ }
+ }
+ }
+
+ if (map_data) {
+ if (location == CLOUDLOG_JOURNAL)
+ file_offset += sizeof(struct log_header);
+ else
+ file_offset += sizeof(struct cloudlog_header);
+
+ file_size = item->data_size;
+ }
+ str = bluesky_cachefile_map_raw(map, file_offset, file_size);
+
+exit2:
+ bluesky_cachefile_unref(map);
+ g_mutex_unlock(map->lock);
+exit1:
+ bluesky_cloudlog_stats_update(item, 1);
+ return str;
+}
+
+void bluesky_mmap_unref(BlueSkyCacheFile *mmap)
+{
+ if (mmap == NULL)
+ return;
+
+ if (g_atomic_int_dec_and_test(&mmap->mapcount)) {
+ g_mutex_lock(mmap->lock);
+ if (mmap->addr != NULL && g_atomic_int_get(&mmap->mapcount) == 0) {
+ if (bluesky_verbose)
+ g_print("Unmapped log segment %d...\n", mmap->log_seq);
+ munmap((void *)mmap->addr, mmap->len);
+ mmap->addr = NULL;
+ g_atomic_int_add(&mmap->refcount, -1);
+ }
+ g_mutex_unlock(mmap->lock);
+ }
+ g_assert(g_atomic_int_get(&mmap->mapcount) >= 0);
+}
+
+/******************************* JOURNAL REPLAY *******************************
+ * The journal replay code is used to recover filesystem state after a
+ * filesystem restart. We first look for the most recent commit record in the
+ * journal, which indicates the point before which all data in the journal has
+ * also been committed to the cloud. Then, we read in all data in the log past
+ * that point.
+ */
+static GList *directory_contents(const char *dirname)
+{
+ GList *contents = NULL;
+ GDir *dir = g_dir_open(dirname, 0, NULL);
+ if (dir == NULL) {
+ g_warning("Unable to open journal directory: %s", dirname);
+ return NULL;
+ }
+
+ const gchar *file;
+ while ((file = g_dir_read_name(dir)) != NULL) {
+ if (strncmp(file, "journal-", 8) == 0)
+ contents = g_list_prepend(contents, g_strdup(file));
+ }
+ g_dir_close(dir);
+
+ contents = g_list_sort(contents, (GCompareFunc)strcmp);
+
+ return contents;
+}
+
+static gboolean validate_journal_item(const char *buf, size_t len, off_t offset)
+{
+ const struct log_header *header;
+ const struct log_footer *footer;
+
+ if (offset + sizeof(struct log_header) + sizeof(struct log_footer) > len)
+ return FALSE;
+
+ header = (const struct log_header *)(buf + offset);
+ if (GUINT32_FROM_LE(header->magic) != HEADER_MAGIC)
+ return FALSE;
+ if (GUINT32_FROM_LE(header->offset) != offset)
+ return FALSE;
+ size_t size = GUINT32_FROM_LE(header->size1)
+ + GUINT32_FROM_LE(header->size2)
+ + GUINT32_FROM_LE(header->size3);
+
+ off_t footer_offset = offset + sizeof(struct log_header) + size;
+ if (footer_offset + sizeof(struct log_footer) > len)
+ return FALSE;
+ footer = (const struct log_footer *)(buf + footer_offset);
+
+ if (GUINT32_FROM_LE(footer->magic) != FOOTER_MAGIC)
+ return FALSE;
+
+ uint32_t crc = crc32c(BLUESKY_CRC32C_SEED, buf + offset,
+ sizeof(struct log_header) + sizeof(struct log_footer)
+ + size);
+ if (crc != BLUESKY_CRC32C_VALIDATOR) {
+ g_warning("Journal entry failed to validate: CRC %08x != %08x",
+ crc, BLUESKY_CRC32C_VALIDATOR);
+ return FALSE;
+ }
+
+ return TRUE;
+}
+
+/* Scan through a journal segment to extract correctly-written items (those
+ * that pass sanity checks and have a valid checksum). */
+static void bluesky_replay_scan_journal(const char *buf, size_t len,
+ uint32_t *seq, uint32_t *start_offset)
+{
+ const struct log_header *header;
+ off_t offset = 0;
+
+ while (validate_journal_item(buf, len, offset)) {
+ header = (const struct log_header *)(buf + offset);
+ size_t size = GUINT32_FROM_LE(header->size1)
+ + GUINT32_FROM_LE(header->size2)
+ + GUINT32_FROM_LE(header->size3);
+
+ if (header->type - '0' == LOGTYPE_JOURNAL_CHECKPOINT) {
+ const uint32_t *data = (const uint32_t *)((const char *)header + sizeof(struct log_header));
+ *seq = GUINT32_FROM_LE(data[0]);
+ *start_offset = GUINT32_FROM_LE(data[1]);
+ }
+
+ offset += sizeof(struct log_header) + size + sizeof(struct log_footer);
+ }
+}
+
+static void reload_item(BlueSkyCloudLog *log_item,
+ const char *data,
+ size_t len1, size_t len2, size_t len3)
+{
+ BlueSkyFS *fs = log_item->fs;
+ /*const char *data1 = data;*/
+ const BlueSkyCloudID *data2
+ = (const BlueSkyCloudID *)(data + len1);
+ /*const BlueSkyCloudPointer *data3
+ = (const BlueSkyCloudPointer *)(data + len1 + len2);*/
+
+ bluesky_cloudlog_stats_update(log_item, -1);
+ bluesky_string_unref(log_item->data);
+ log_item->data = NULL;
+ log_item->location_flags = CLOUDLOG_JOURNAL;
+ bluesky_cloudlog_stats_update(log_item, 1);
+
+ BlueSkyCloudID id0;
+ memset(&id0, 0, sizeof(id0));
+
+ int link_count = len2 / sizeof(BlueSkyCloudID);
+ GArray *new_links = g_array_new(FALSE, TRUE, sizeof(BlueSkyCloudLog *));
+ for (int i = 0; i < link_count; i++) {
+ BlueSkyCloudID id = data2[i];
+ BlueSkyCloudLog *ref = NULL;
+ if (memcmp(&id, &id0, sizeof(BlueSkyCloudID)) != 0) {
+ g_mutex_lock(fs->lock);
+ ref = g_hash_table_lookup(fs->locations, &id);
+ if (ref != NULL) {
+ bluesky_cloudlog_ref(ref);
+ }
+ g_mutex_unlock(fs->lock);
+ }
+ g_array_append_val(new_links, ref);
+ }
+
+ for (int i = 0; i < log_item->links->len; i++) {
+ BlueSkyCloudLog *c = g_array_index(log_item->links,
+ BlueSkyCloudLog *, i);
+ bluesky_cloudlog_unref(c);
+ }
+ g_array_unref(log_item->links);
+ log_item->links = new_links;
+}
+
+static void bluesky_replay_scan_journal2(BlueSkyFS *fs, GList **objects,
+ int log_seq, int start_offset,
+ const char *buf, size_t len)
+{
+ const struct log_header *header;
+ off_t offset = start_offset;
+
+ while (validate_journal_item(buf, len, offset)) {
+ header = (const struct log_header *)(buf + offset);
+ g_print("In replay found valid item at offset %zd\n", offset);
+ size_t size = GUINT32_FROM_LE(header->size1)
+ + GUINT32_FROM_LE(header->size2)
+ + GUINT32_FROM_LE(header->size3);
+
+ BlueSkyCloudLog *log_item = bluesky_cloudlog_get(fs, header->id);
+ g_mutex_lock(log_item->lock);
+ *objects = g_list_prepend(*objects, log_item);
+
+ log_item->inum = GUINT64_FROM_LE(header->inum);
+ reload_item(log_item, buf + offset + sizeof(struct log_header),
+ GUINT32_FROM_LE(header->size1),
+ GUINT32_FROM_LE(header->size2),
+ GUINT32_FROM_LE(header->size3));
+ log_item->log_seq = log_seq;
+ log_item->log_offset = offset + sizeof(struct log_header);
+ log_item->log_size = header->size1;
+
+ bluesky_string_unref(log_item->data);
+ log_item->data = bluesky_string_new(g_memdup(buf + offset + sizeof(struct log_header), GUINT32_FROM_LE(header->size1)), GUINT32_FROM_LE(header->size1));
+
+ /* For any inodes which were read from the journal, deserialize the
+ * inode information, overwriting any old inode data. */
+ if (header->type - '0' == LOGTYPE_INODE) {
+ uint64_t inum = GUINT64_FROM_LE(header->inum);
+ BlueSkyInode *inode;
+ g_mutex_lock(fs->lock);
+ inode = (BlueSkyInode *)g_hash_table_lookup(fs->inodes, &inum);
+ if (inode == NULL) {
+ inode = bluesky_new_inode(inum, fs, BLUESKY_PENDING);
+ inode->change_count = 0;
+ bluesky_insert_inode(fs, inode);
+ }
+ g_mutex_lock(inode->lock);
+ bluesky_inode_free_resources(inode);
+ if (!bluesky_deserialize_inode(inode, log_item))
+ g_print("Error deserializing inode %"PRIu64"\n", inum);
+ fs->next_inum = MAX(fs->next_inum, inum + 1);
+ bluesky_list_unlink(&fs->accessed_list, inode->accessed_list);
+ inode->accessed_list = bluesky_list_prepend(&fs->accessed_list, inode);
+ bluesky_list_unlink(&fs->dirty_list, inode->dirty_list);
+ inode->dirty_list = bluesky_list_prepend(&fs->dirty_list, inode);
+ bluesky_list_unlink(&fs->unlogged_list, inode->unlogged_list);
+ inode->unlogged_list = NULL;
+ inode->change_cloud = inode->change_commit;
+ bluesky_cloudlog_ref(log_item);
+ bluesky_cloudlog_unref(inode->committed_item);
+ inode->committed_item = log_item;
+ g_mutex_unlock(inode->lock);
+ g_mutex_unlock(fs->lock);
+ }
+ bluesky_string_unref(log_item->data);
+ log_item->data = NULL;
+ g_mutex_unlock(log_item->lock);
+
+ offset += sizeof(struct log_header) + size + sizeof(struct log_footer);
+ }
+}
+
+void bluesky_replay(BlueSkyFS *fs)
+{
+ BlueSkyLog *log = fs->log;
+ GList *logfiles = directory_contents(log->log_directory);
+
+ /* Scan through log files in reverse order to find the most recent commit
+ * record. */
+ logfiles = g_list_reverse(logfiles);
+ uint32_t seq_num = 0, start_offset = 0;
+ while (logfiles != NULL) {
+ char *filename = g_strdup_printf("%s/%s", log->log_directory,
+ (char *)logfiles->data);
+ g_print("Scanning file %s\n", filename);
+ GMappedFile *map = g_mapped_file_new(filename, FALSE, NULL);
+ if (map == NULL) {
+ g_warning("Mapping logfile %s failed!\n", filename);
+ } else {
+ bluesky_replay_scan_journal(g_mapped_file_get_contents(map),
+ g_mapped_file_get_length(map),
+ &seq_num, &start_offset);
+ g_mapped_file_unref(map);
+ }
+ g_free(filename);
+
+ g_free(logfiles->data);
+ logfiles = g_list_delete_link(logfiles, logfiles);
+ if (seq_num != 0 || start_offset != 0)
+ break;
+ }
+ g_list_foreach(logfiles, (GFunc)g_free, NULL);
+ g_list_free(logfiles);
+
+ /* Now, scan forward starting from the given point in the log to
+ * reconstruct all filesystem state. As we reload objects we hold a
+ * reference to each loaded object. At the end we free all these
+ * references, so that any objects which were not linked into persistent
+ * filesystem data structures are freed. */
+ GList *objects = NULL;
+ while (TRUE) {
+ char *filename = g_strdup_printf("%s/journal-%08d",
+ log->log_directory, seq_num);
+ g_print("Replaying file %s from offset %d\n", filename, start_offset);
+ GMappedFile *map = g_mapped_file_new(filename, FALSE, NULL);
+ g_free(filename);
+ if (map == NULL) {
+ g_warning("Mapping logfile failed, assuming end of journal\n");
+ break;
+ }
+
+ bluesky_replay_scan_journal2(fs, &objects, seq_num, start_offset,
+ g_mapped_file_get_contents(map),
+ g_mapped_file_get_length(map));
+ g_mapped_file_unref(map);
+ seq_num++;
+ start_offset = 0;
+ }
+
+ while (objects != NULL) {
+ bluesky_cloudlog_unref((BlueSkyCloudLog *)objects->data);
+ objects = g_list_delete_link(objects, objects);
+ }