#define HEADER_MAGIC 0x676f4c0a
#define FOOTER_MAGIC 0x2e435243
-struct log_header {
- uint32_t magic; // HEADER_MAGIC
- uint8_t type; // Object type + '0'
- uint32_t offset; // Starting byte offset of the log header
- uint32_t size; // Size of the data item (bytes)
- uint64_t inum; // Inode which owns this data, if any
- BlueSkyCloudID id; // Object identifier
-} __attribute__((packed));
-
-struct log_footer {
- uint32_t magic; // FOOTER_MAGIC
- uint32_t crc; // Computed from log_header to log_footer.magic
-} __attribute__((packed));
-
static void writebuf(int fd, const char *buf, size_t len)
{
while (len > 0) {
g_mutex_lock(item->lock);
bluesky_cloudlog_stats_update(item, -1);
item->pending_write &= ~CLOUDLOG_JOURNAL;
- item->location_flags |= CLOUDLOG_JOURNAL;
+ item->location_flags
+ = (item->location_flags & ~CLOUDLOG_UNCOMMITTED) | CLOUDLOG_JOURNAL;
bluesky_cloudlog_stats_update(item, 1);
g_cond_signal(item->cond);
g_mutex_unlock(item->lock);
item->pending_write |= CLOUDLOG_JOURNAL;
bluesky_cloudlog_stats_update(item, 1);
+ GString *data1 = g_string_new("");
+ GString *data2 = g_string_new("");
+ GString *data3 = g_string_new("");
+ bluesky_serialize_cloudlog(item, data1, data2, data3);
+
struct log_header header;
struct log_footer footer;
- size_t size = sizeof(header) + sizeof(footer) + item->data->len;
+ size_t size = sizeof(header) + sizeof(footer);
+ size += data1->len + data2->len + data3->len;
off_t offset = 0;
if (log->fd >= 0)
offset = lseek(log->fd, 0, SEEK_CUR);
header.magic = GUINT32_TO_LE(HEADER_MAGIC);
header.offset = GUINT32_TO_LE(offset);
- header.size = GUINT32_TO_LE(item->data->len);
+ header.size1 = GUINT32_TO_LE(data1->len);
+ header.size2 = GUINT32_TO_LE(data2->len);
+ header.size3 = GUINT32_TO_LE(data3->len);
header.type = item->type + '0';
header.id = item->id;
header.inum = GUINT64_TO_LE(item->inum);
writebuf(log->fd, (const char *)&header, sizeof(header));
crc = crc32c(crc, (const char *)&header, sizeof(header));
- writebuf(log->fd, item->data->data, item->data->len);
- crc = crc32c(crc, item->data->data, item->data->len);
+ writebuf(log->fd, data1->str, data1->len);
+ crc = crc32c(crc, data1->str, data1->len);
+ writebuf(log->fd, data2->str, data2->len);
+ crc = crc32c(crc, data2->str, data2->len);
+ writebuf(log->fd, data3->str, data3->len);
+ crc = crc32c(crc, data3->str, data3->len);
crc = crc32c(crc, (const char *)&footer,
sizeof(footer) - sizeof(uint32_t));
writebuf(log->fd, (const char *)&footer, sizeof(footer));
item->log_seq = log->seq_num;
- item->log_offset = offset + sizeof(header);
- item->log_size = item->data->len;
+ item->log_offset = offset;
+ item->log_size = size;
+ item->data_size = item->data->len;
+
+ offset += size;
- offset += sizeof(header) + sizeof(footer) + item->data->len;
+ g_string_free(data1, TRUE);
+ g_string_free(data2, TRUE);
+ g_string_free(data3, TRUE);
/* Replace the log item's string data with a memory-mapped copy of the
* data, now that it has been written to the log file. (Even if it
item->data = NULL;
bluesky_cloudlog_fetch(item);
- log->committed = g_slist_prepend(log->committed, item);
+ log->committed = g_slist_prepend(log->committed, item);
g_atomic_int_add(&item->data_lock_count, -1);
g_mutex_unlock(item->lock);
log->mmap_lock = g_mutex_new();
log->mmap_cache = g_hash_table_new(g_str_hash, g_str_equal);
+ /* Determine the highest-numbered log file, so that we can start writing
+ * out new journal entries at the next sequence number. */
+ GDir *dir = g_dir_open(log_directory, 0, NULL);
+ if (dir != NULL) {
+ const gchar *file;
+ while ((file = g_dir_read_name(dir)) != NULL) {
+ if (strncmp(file, "journal-", 8) == 0) {
+ log->seq_num = MAX(log->seq_num, atoi(&file[8]) + 1);
+ }
+ }
+ g_dir_close(dir);
+ g_print("Starting journal at sequence number %d\n", log->seq_num);
+ }
+
log->dirfd = open(log->log_directory, O_DIRECTORY);
if (log->dirfd < 0) {
fprintf(stderr, "Unable to open logging directory: %m\n");
void bluesky_log_item_submit(BlueSkyCloudLog *item, BlueSkyLog *log)
{
- bluesky_cloudlog_ref(item);
- g_atomic_int_add(&item->data_lock_count, 1);
- g_async_queue_push(log->queue, item);
+ if (!(item->location_flags & CLOUDLOG_JOURNAL)) {
+ bluesky_cloudlog_ref(item);
+ item->location_flags |= CLOUDLOG_UNCOMMITTED;
+ g_atomic_int_add(&item->data_lock_count, 1);
+ g_async_queue_push(log->queue, item);
+ }
}
void bluesky_log_finish_all(GList *log_items)
BlueSkyCloudLog *item = (BlueSkyCloudLog *)log_items->data;
g_mutex_lock(item->lock);
- while ((item->pending_write & CLOUDLOG_JOURNAL))
+ while ((item->location_flags & CLOUDLOG_UNCOMMITTED))
g_cond_wait(item->cond, item->lock);
g_mutex_unlock(item->lock);
bluesky_cloudlog_unref(item);
}
}
+/* Return a committed cloud log record that can be used as a watermark for how
+ * much of the journal has been written. */
+BlueSkyCloudLog *bluesky_log_get_commit_point(BlueSkyFS *fs)
+{
+ BlueSkyCloudLog *marker = bluesky_cloudlog_new(fs, NULL);
+ marker->type = LOGTYPE_JOURNAL_MARKER;
+ marker->data = bluesky_string_new(g_strdup(""), 0);
+ bluesky_cloudlog_stats_update(marker, 1);
+ bluesky_cloudlog_sync(marker);
+
+ g_mutex_lock(marker->lock);
+ while ((marker->pending_write & CLOUDLOG_JOURNAL))
+ g_cond_wait(marker->cond, marker->lock);
+ g_mutex_unlock(marker->lock);
+
+ return marker;
+}
+
+void bluesky_log_write_commit_point(BlueSkyFS *fs, BlueSkyCloudLog *marker)
+{
+ BlueSkyCloudLog *commit = bluesky_cloudlog_new(fs, NULL);
+ commit->type = LOGTYPE_JOURNAL_CHECKPOINT;
+
+ uint32_t seq, offset;
+ seq = GUINT32_TO_LE(marker->log_seq);
+ offset = GUINT32_TO_LE(marker->log_offset);
+ GString *loc = g_string_new("");
+ g_string_append_len(loc, (const gchar *)&seq, sizeof(seq));
+ g_string_append_len(loc, (const gchar *)&offset, sizeof(offset));
+ commit->data = bluesky_string_new_from_gstring(loc);
+ bluesky_cloudlog_stats_update(commit, 1);
+ bluesky_cloudlog_sync(commit);
+
+ g_mutex_lock(commit->lock);
+ while ((commit->location_flags & CLOUDLOG_UNCOMMITTED))
+ g_cond_wait(commit->cond, commit->lock);
+ g_mutex_unlock(commit->lock);
+
+ bluesky_cloudlog_unref(marker);
+ bluesky_cloudlog_unref(commit);
+}
+
/* Memory-map the given log object into memory (read-only) and return a pointer
* to it. */
static int page_size = 0;
char *pathname = g_strdup_printf("%s/%s",
cachefile->log->log_directory,
cachefile->filename);
+ async->data = bluesky_string_dup(async->data);
+ bluesky_cloudlog_decrypt(async->data->data, async->data->len,
+ cachefile->fs->keys);
if (!g_file_set_contents(pathname, async->data->data, async->data->len,
NULL))
g_print("Error writing out fetched file to cache!\n");
close(fd);
}
- g_mutex_unlock(log->mmap_lock);
-
BlueSkyRCStr *str;
map->atime = bluesky_get_current_time();
str = bluesky_string_new_from_mmap(map, log_offset, log_size);
}
}
-/* Scan through all currently-stored files in the journal/cache and garbage
- * collect old unused ones, if needed. */
-static void gather_cachefiles(gpointer key, gpointer value, gpointer user_data)
-{
- GList **files = (GList **)user_data;
- *files = g_list_prepend(*files, value);
-}
-
-static gint compare_cachefiles(gconstpointer a, gconstpointer b)
-{
- int64_t ta, tb;
-
- ta = ((BlueSkyCacheFile *)a)->atime;
- tb = ((BlueSkyCacheFile *)b)->atime;
- if (ta < tb)
- return -1;
- else if (ta > tb)
- return 1;
- else
- return 0;
-}
-
-void bluesky_cachefile_gc(BlueSkyFS *fs)
-{
- GList *files = NULL;
-
- g_mutex_lock(fs->log->mmap_lock);
- g_hash_table_foreach(fs->log->mmap_cache, gather_cachefiles, &files);
-
- /* Sort based on atime. The atime should be stable since it shouln't be
- * updated except by threads which can grab the mmap_lock, which we already
- * hold. */
- files = g_list_sort(files, compare_cachefiles);
-
- /* Walk the list of files, starting with the oldest, deleting files if
- * possible until enough space has been reclaimed. */
- g_print("\nScanning cache: (total size = %d kB)\n", fs->log->disk_used);
- while (files != NULL) {
- BlueSkyCacheFile *cachefile = (BlueSkyCacheFile *)files->data;
- /* Try to lock the structure, but if the lock is held by another thread
- * then we'll just skip the file on this pass. */
- if (g_mutex_trylock(cachefile->lock)) {
- int64_t age = bluesky_get_current_time() - cachefile->atime;
- g_print("%s addr=%p mapcount=%d refcount=%d atime_age=%f",
- cachefile->filename, cachefile->addr, cachefile->mapcount,
- cachefile->refcount, age / 1e6);
- if (cachefile->fetching)
- g_print(" (fetching)");
- g_print("\n");
-
- gboolean deletion_candidate = FALSE;
- if (g_atomic_int_get(&fs->log->disk_used)
- > bluesky_options.cache_size
- && g_atomic_int_get(&cachefile->refcount) == 0
- && g_atomic_int_get(&cachefile->mapcount) == 0)
- {
- deletion_candidate = TRUE;
- }
-
- /* Don't allow journal files to be reclaimed until all data is
- * known to be durably stored in the cloud. */
- if (cachefile->type == CLOUDLOG_JOURNAL
- && cachefile->log_seq >= fs->log->journal_watermark)
- {
- deletion_candidate = FALSE;
- }
-
- if (deletion_candidate) {
- g_print(" ...deleting\n");
- if (unlinkat(fs->log->dirfd, cachefile->filename, 0) < 0) {
- fprintf(stderr, "Unable to unlink journal %s: %m\n",
- cachefile->filename);
- }
-
- g_atomic_int_add(&fs->log->disk_used, -(cachefile->len / 1024));
- g_hash_table_remove(fs->log->mmap_cache, cachefile->filename);
- g_mutex_unlock(cachefile->lock);
- g_mutex_free(cachefile->lock);
- g_cond_free(cachefile->cond);
- g_free(cachefile->filename);
- g_free(cachefile);
- } else {
- g_mutex_unlock(cachefile->lock);
- }
- }
- files = g_list_delete_link(files, files);
- }
- g_list_free(files);
-
- g_mutex_unlock(fs->log->mmap_lock);
-}
-
/******************************* 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
* 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)
{
return FALSE;
if (GUINT32_FROM_LE(header->offset) != offset)
return FALSE;
- size_t size = GUINT32_FROM_LE(header->size);
+ 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)
uint32_t crc = crc32c(BLUESKY_CRC32C_SEED, buf + offset,
sizeof(struct log_header) + sizeof(struct log_footer)
+ size);
- if (crc != 0)
+ 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)
+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);
- 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);
+
+ 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]);
+ }
- size_t size = GUINT32_FROM_LE(header->size);
+ 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;
- char logname[64];
- int seq_num = 0;
- int fd;
-
- g_snprintf(logname, sizeof(logname), "journal-%08d", seq_num);
- fd = openat(log->dirfd, logname, O_RDONLY);
- if (fd < 0)
- return;
+ 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);
- off_t length = lseek(fd, 0, SEEK_END);
- void *addr = mmap(NULL, length, PROT_READ, MAP_SHARED, fd, 0);
- if (addr == NULL)
- return;
+ 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_journal((const char *)addr, length);
+ 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;
+ }
- munmap(addr, length);
+ while (objects != NULL) {
+ bluesky_cloudlog_unref((BlueSkyCloudLog *)objects->data);
+ objects = g_list_delete_link(objects, objects);
+ }
}