// no absolute guarantees on the size of a log segment.
#define LOG_SEGMENT_SIZE (1 << 22)
-// Target amount of disk space to use for the journal and cache files, in
-// kilobytes.
-#define DISK_CACHE_SIZE_TARGET (64 * 1024)
-
#define HEADER_MAGIC 0x676f4c0a
#define FOOTER_MAGIC 0x2e435243
struct log_header {
uint32_t magic; // HEADER_MAGIC
- uint64_t offset; // Starting byte offset of the log header
- uint32_t size; // Size of the data item (bytes)
+ uint8_t type; // Object type + '0'
+ uint32_t offset; // Starting byte offset of the log header
+ uint32_t size1; // Size of the data item (bytes)
+ uint32_t size2; //
+ uint32_t size3; //
+ uint64_t inum; // Inode which owns this data, if any
BlueSkyCloudID id; // Object identifier
} __attribute__((packed));
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 = GUINT64_TO_LE(offset);
- header.size = GUINT32_TO_LE(item->data->len);
+ header.offset = GUINT32_TO_LE(offset);
+ 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);
footer.magic = GUINT32_TO_LE(FOOTER_MAGIC);
uint32_t crc = BLUESKY_CRC32C_SEED;
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));
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
* isn't yet on disk, it should at least be in the page cache and so
g_print("\n");
gboolean deletion_candidate = FALSE;
- if (g_atomic_int_get(&fs->log->disk_used) > DISK_CACHE_SIZE_TARGET
+ 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)
{
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
+ * 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)
+{
+ 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);
+ 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_string_unref(log_item->data);
+ log_item->data = NULL;
+ log_item->location_flags = CLOUDLOG_JOURNAL;
+
+ 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,
+ const char *buf, size_t len)
+{
+ 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);
+
+ g_mutex_lock(fs->lock);
+ BlueSkyCloudLog *log_item;
+ log_item = g_hash_table_lookup(fs->locations, &header->id);
+ if (log_item == NULL) {
+ log_item = bluesky_cloudlog_new(fs, &header->id);
+ g_hash_table_insert(fs->locations, &log_item->id, log_item);
+ g_mutex_lock(log_item->lock);
+ } else {
+ bluesky_cloudlog_ref(log_item);
+ g_mutex_lock(log_item->lock);
+ }
+ g_mutex_unlock(fs->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);
+ if (!bluesky_deserialize_inode(inode, log_item))
+ g_print("Error deserializing inode %"PRIu64"\n", inum);
+ fs->next_inum = MAX(fs->next_inum, inum + 1);
+ 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);
+ 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));
+ g_mapped_file_unref(map);
+ }
+ g_free(filename);
+
+ g_free(logfiles->data);
+ logfiles = g_list_delete_link(logfiles, logfiles);
+ }
+ 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;
+ int seq_num = 0;
+ while (TRUE) {
+ char *filename = g_strdup_printf("%s/journal-%08d",
+ log->log_directory, seq_num);
+ g_print("Replaying file %s\n", filename);
+ 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,
+ g_mapped_file_get_contents(map),
+ g_mapped_file_get_length(map));
+ g_mapped_file_unref(map);
+ seq_num++;
+ }
+
+ while (objects != NULL) {
+ bluesky_cloudlog_unref((BlueSkyCloudLog *)objects->data);
+ objects = g_list_delete_link(objects, objects);
+ }
+}
projects / bluesky.git / blobdiff