// Rough size limit for a log segment. This is not a firm limit and there are
// no absolute guarantees on the size of a log segment.
-#define LOG_SEGMENT_SIZE (1 << 24)
+#define LOG_SEGMENT_SIZE (1 << 22)
#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
+ 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));
log->fd = -1;
}
+ if (log->current_log != NULL) {
+ bluesky_cachefile_unref(log->current_log);
+ log->current_log = NULL;
+ }
+
while (log->fd < 0) {
g_snprintf(logname, sizeof(logname), "journal-%08d", log->seq_num);
log->fd = openat(log->dirfd, logname, O_CREAT|O_WRONLY|O_EXCL, 0600);
}
}
+ log->current_log = bluesky_cachefile_lookup(log->fs, -1, log->seq_num);
+ g_assert(log->current_log != NULL);
+ g_mutex_unlock(log->current_log->lock);
+
if (ftruncate(log->fd, LOG_SEGMENT_SIZE) < 0) {
fprintf(stderr, "Unable to truncate logfile %s: %m\n", logname);
}
}
header.magic = GUINT32_TO_LE(HEADER_MAGIC);
- header.offset = GUINT64_TO_LE(offset);
+ header.offset = GUINT32_TO_LE(offset);
header.size = GUINT32_TO_LE(item->data->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;
g_atomic_int_add(&cachefile->refcount, -1);
}
+static void cloudlog_fetch_complete(BlueSkyStoreAsync *async,
+ BlueSkyCacheFile *cachefile);
+
+static void cloudlog_fetch_start(BlueSkyCacheFile *cachefile)
+{
+ g_atomic_int_inc(&cachefile->refcount);
+ cachefile->fetching = TRUE;
+ 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);
+ bluesky_store_async_add_notifier(async,
+ (GFunc)cloudlog_fetch_complete,
+ cachefile);
+ bluesky_store_async_submit(async);
+ bluesky_store_async_unref(async);
+}
+
static void cloudlog_fetch_complete(BlueSkyStoreAsync *async,
BlueSkyCacheFile *cachefile)
{
g_print("Fetch of %s from cloud complete, status = %d\n",
async->key, async->result);
- if (async->result < 0)
- return;
-
g_mutex_lock(cachefile->lock);
- char *pathname = g_strdup_printf("%s/%s", cachefile->log->log_directory,
- cachefile->filename);
- if (!g_file_set_contents(pathname, async->data->data, async->data->len,
- NULL))
- {
- g_print("Error writing out fetched file to cache!\n");
+ if (async->result >= 0) {
+ char *pathname = g_strdup_printf("%s/%s",
+ cachefile->log->log_directory,
+ cachefile->filename);
+ if (!g_file_set_contents(pathname, async->data->data, async->data->len,
+ NULL))
+ g_print("Error writing out fetched file to cache!\n");
+ g_free(pathname);
+
+ cachefile->fetching = FALSE;
+ cachefile->ready = TRUE;
+ } else {
+ g_print("Error fetching from cloud, retrying...\n");
+ cloudlog_fetch_start(cachefile);
}
- g_free(pathname);
- cachefile->fetching = FALSE;
- cachefile->ready = TRUE;
+
bluesky_cachefile_unref(cachefile);
g_cond_broadcast(cachefile->cond);
g_mutex_unlock(cachefile->lock);
BlueSkyLog *log = fs->log;
+ struct stat statbuf;
char logname[64];
int type;
g_mutex_lock(log->mmap_lock);
map = g_hash_table_lookup(log->mmap_cache, logname);
- if (map == NULL) {
- /* TODO: stat() call */
+ if (map == NULL
+ && type == CLOUDLOG_JOURNAL
+ && fstatat(log->dirfd, logname, &statbuf, 0) < 0) {
+ /* A stale reference to a journal file which doesn't exist any longer
+ * because it was reclaimed. Return NULL. */
+ } else if (map == NULL) {
+ g_print("Adding cache file %s\n", logname);
+
map = g_new0(BlueSkyCacheFile, 1);
- map->type = CLOUDLOG_JOURNAL;
+ map->fs = fs;
+ map->type = type;
map->lock = g_mutex_new();
map->type = type;
g_mutex_lock(map->lock);
g_hash_table_insert(log->mmap_cache, map->filename, map);
// If the log file is stored in the cloud, we may need to fetch it
- if (clouddir >= 0) {
- g_atomic_int_inc(&map->refcount);
- map->fetching = TRUE;
- g_print("Starting fetch of %s from cloud\n", logname);
- BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
- async->op = STORE_OP_GET;
- async->key = g_strdup(logname);
- bluesky_store_async_add_notifier(async,
- (GFunc)cloudlog_fetch_complete,
- map);
- bluesky_store_async_submit(async);
- bluesky_store_async_unref(async);
- }
+ if (clouddir >= 0)
+ cloudlog_fetch_start(map);
} else {
g_mutex_lock(map->lock);
}
g_mutex_unlock(log->mmap_lock);
- g_atomic_int_inc(&map->refcount);
+ if (map != NULL)
+ g_atomic_int_inc(&map->refcount);
return map;
}
g_print(" (fetching)");
g_print("\n");
- if (g_atomic_int_get(&cachefile->refcount) == 0
- && g_atomic_int_get(&cachefile->mapcount) == 0
- && cachefile->type == CLOUDLOG_CLOUD /* FIXME: journals too */)
+ 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_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->size);
+
+ 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->size);
+ offset += sizeof(struct log_header) + size + sizeof(struct log_footer);
+ }
+}
+
+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->size);
+
+ 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);
+
+ bluesky_string_unref(log_item->data);
+ log_item->location_flags = CLOUDLOG_JOURNAL;
+ log_item->data = NULL;
+ log_item->log_seq = log_seq;
+ log_item->log_offset = offset + sizeof(struct log_header);
+ log_item->log_size = header->size;
+ 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