* rewriting log segments where needed and deleting old segments, is handled by
* the in-cloud cleaner component. */
+/* A specialized function for loading an object from the cloud, used just by
+ * the cleaner. When the cleaner runs, it will produce new objects that have
+ * the same object ID as before, differing only in the outgoing link locations
+ * as a result of cleaning. We can't use the normal functions for loading
+ * objects since those can only deal with one version of an object. Instead,
+ * we load the cleaned object with this function, and then merge the state into
+ * the official object as needed.
+ *
+ * The bluesky_cleaner_deserialize function is like
+ * bluesky_deserialize_cloudlog, but does a more limited deserialization and
+ * keeps the returned item entirely separate from any in-memory BlueSkyCloudLog
+ * objects. */
+
+typedef struct {
+ BlueSkyCloudLogType type;
+ BlueSkyCloudID id;
+ BlueSkyCloudPointer location;
+ uint64_t inum;
+ GArray *links;
+} BlueSkyCleanerItem;
+
+typedef struct {
+ BlueSkyCloudID id;
+ BlueSkyCloudPointer location;
+} BlueSkyCleanerLink;
+
+static BlueSkyCleanerItem *bluesky_cleaner_deserialize(BlueSkyRCStr *raw)
+{
+ const char *data = raw->data;
+ size_t len = raw->len;
+ const char *data1, *data2, *data3;
+ size_t len1, len2, len3;
+
+ g_assert(len > 4);
+ if (len < sizeof(struct cloudlog_header)
+ || memcmp(data, CLOUDLOG_MAGIC, 4) != 0)
+ {
+ g_warning("Deserializing garbage cloud log item from cleaner!");
+ return NULL;
+ };
+
+ struct cloudlog_header *header = (struct cloudlog_header *)data;
+ len1 = GUINT32_FROM_LE(header->size1);
+ len2 = GUINT32_FROM_LE(header->size2);
+ len3 = GUINT32_FROM_LE(header->size3);
+ data1 = data + sizeof(struct cloudlog_header);
+ data2 = data1 + len1;
+ data3 = data2 + len2;
+ g_assert(data3 + len3 - data <= len);
+
+ BlueSkyCleanerItem *item = g_new0(BlueSkyCleanerItem, 1);
+ item->type = header->type - '0';
+ item->inum = GUINT64_FROM_LE(header->inum);
+ memcpy(&item->id, &header->id, sizeof(BlueSkyCloudID));
+
+ int link_count = len2 / sizeof(BlueSkyCloudID);
+ g_print("Outgoing links: %d\n", link_count);
+ item->links = g_array_new(FALSE, TRUE, sizeof(BlueSkyCleanerLink));
+ for (int i = 0; i < link_count; i++) {
+ BlueSkyCleanerLink link;
+
+ g_assert(len2 >= sizeof(link.id));
+ memcpy(&link.id, data2, sizeof(link.id));
+ data2 += sizeof(link.id); len2 -= sizeof(link.id);
+
+ g_assert(len3 >= sizeof(link.location));
+ memcpy(&link.location, data3, sizeof(link.location));
+ data3 += sizeof(link.location); len3 -= sizeof(link.location);
+
+ g_array_append_val(item->links, link);
+ }
+
+ return item;
+}
+
+static void bluesky_cleaner_item_free(BlueSkyCleanerItem *item)
+{
+ if (item == NULL)
+ return;
+ g_array_unref(item->links);
+ g_free(item);
+}
+
/* Check the cleaner's logs to find the a more recent checkpoint record. This
* should be called occasionally to see if the cleaner has done any work since
* our last check. */
-void bluesky_cleaner_find_checkpoint(BlueSkyFS *fs)
+static BlueSkyCleanerItem *bluesky_cleaner_find_checkpoint(BlueSkyFS *fs)
{
char *prefix = g_strdup_printf("log-%08d", BLUESKY_CLOUD_DIR_CLEANER);
char *last_segment = bluesky_store_lookup_last(fs->store, prefix);
g_free(prefix);
if (last_segment == NULL)
- return;
+ return NULL;
- g_print("Last cloud log segment: %s\n", last_segment);
+ g_print("Last cloud log segment: %s (processed up to %d)\n",
+ last_segment, fs->log_state->latest_cleaner_seq_seen);
int seq = atoi(last_segment + 13);
g_free(last_segment);
if (seq <= fs->log_state->latest_cleaner_seq_seen)
- return;
+ return NULL;
g_print("New log segment appeared in cleaner directory: %d\n", seq);
length = item->length;
}
- if (length > 0) {
- g_print("Last object: %"PRIu64" + %"PRIu64"\n", offset, length);
+ if (length == 0) {
+ bluesky_cachefile_unref(cachefile);
+ g_mutex_unlock(cachefile->lock);
+ return NULL;
}
+ g_print("Found a cleaner checkpoint record.\n");
+
+ BlueSkyRCStr *data = bluesky_cachefile_map_raw(cachefile, offset, length);
bluesky_cachefile_unref(cachefile);
g_mutex_unlock(cachefile->lock);
+
+ BlueSkyCleanerItem *checkpoint = bluesky_cleaner_deserialize(data);
+ checkpoint->location.directory = BLUESKY_CLOUD_DIR_CLEANER;
+ checkpoint->location.sequence = seq;
+ bluesky_string_unref(data);
+
+ return checkpoint;
+}
+
+static BlueSkyCleanerItem *cleaner_load_item(BlueSkyFS *fs,
+ BlueSkyCloudPointer location)
+{
+ g_print("Loading item %d/%d/%d...\n", location.directory, location.sequence, location.offset);
+
+ BlueSkyCacheFile *cachefile;
+ cachefile = bluesky_cachefile_lookup(fs, location.directory,
+ location.sequence, TRUE);
+ while (!cachefile->complete)
+ g_cond_wait(cachefile->cond, cachefile->lock);
+
+ /* TODO: Ought to check that we are loading an item which validated? */
+ BlueSkyRCStr *data = bluesky_cachefile_map_raw(cachefile, location.offset,
+ location.size);
+ bluesky_cachefile_unref(cachefile);
+ g_mutex_unlock(cachefile->lock);
+
+ BlueSkyCleanerItem *item = bluesky_cleaner_deserialize(data);
+ bluesky_string_unref(data);
+
+ return item;
+}
+
+/* Does the item at the given cloud location from the cleaner need merging? An
+ * item in the primary log does not need to be merged, as by definition we
+ * already know about it. Similarly, old items in the cleaner's log--those
+ * that we have already seen from a previous merge--do not need to be mergd
+ * again. */
+gboolean needs_merging(BlueSkyFS *fs, BlueSkyCloudPointer location)
+{
+ if (location.directory == BLUESKY_CLOUD_DIR_PRIMARY)
+ return FALSE;
+
+ if (location.directory == BLUESKY_CLOUD_DIR_CLEANER
+ && location.sequence <= fs->log_state->latest_cleaner_seq_seen)
+ return FALSE;
+
+ return TRUE;
+}
+
+/* For an inode which has been modified by the cleaner and must be flushed out
+ * to cloud storage, mark it as appropriately dirty. We will bypass writingt
+ * the inode to the journal if possible--but if there have ben other
+ * uncommitted changes besides what the cleaner did then we will force a
+ * journal write as well since the cloud shouldn't contain newer data than the
+ * journal. Inode must be locked. */
+static void cleaner_flush_inode(BlueSkyInode *inode)
+{
+ // if (inode->change_commit != inode->change_count) {
+ if (TRUE) {
+ /* bluesky_inode_start_sync schedules a flush to the cloud so we're all
+ * done. */
+ bluesky_inode_start_sync(inode);
+ return;
+ }
+
+ g_assert(inode->unlogged_list == NULL);
+
+ 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;
+}
+
+static void merge_inode(BlueSkyFS *fs, BlueSkyCleanerItem *cleaner_inode)
+{
+ /* There are two versions we are concerned with: cleaner_ is for the data
+ * stored in the cleaner's log, and proxy_ is for our most recent version,
+ * which the cleaner might or might not know about. */
+ uint64_t inum = cleaner_inode->inum;
+
+ g_print("Merging inode %"PRIu64" from cleaner\n", inum);
+
+ g_mutex_lock(fs->lock);
+ InodeMapEntry *entry = bluesky_inode_map_lookup(fs->inode_map, inum, 0);
+ if (entry == NULL) {
+ /* Inode doesn't exist: it was probably deleted so keep it that way. */
+ g_mutex_unlock(fs->lock);
+ return;
+ }
+
+ BlueSkyCloudLog *proxy_item = entry->item;
+ g_mutex_lock(proxy_item->lock);
+ //BlueSkyCloudPointer proxy_location = entry->item->location;
+ //BlueSkyCloudID proxy_id = entry->item->id;
+ g_mutex_unlock(proxy_item->lock);
+ g_mutex_unlock(fs->lock);
+
+ /* If the cleaner and the proxy have the same ID, and if the proxy's
+ * in-memory copy is unmodified, then we can simply use the cleaner's
+ * version of the inode. */
+ /* TODO */
+
+ /* Merge file data together for a regular file. Iterate over the file
+ * blocks in the proxy's copy of the inode. If the block ID is unchanged
+ * in the cleaner but the location was updated, then update the location in
+ * the cleaner because the block was relocated. Otherwise ignore the
+ * cleaner's version for that block because the proxy's information is more
+ * recent. */
+ BlueSkyInode *proxy_inode = bluesky_get_inode(fs, inum);
+ g_mutex_lock(proxy_inode->lock);
+ if (proxy_inode->type == BLUESKY_REGULAR) {
+ for (int i = 0; i < proxy_inode->blocks->len; i++) {
+ BlueSkyBlock *b = &g_array_index(proxy_inode->blocks,
+ BlueSkyBlock, i);
+ if (b->type != BLUESKY_BLOCK_REF)
+ continue;
+ if (i >= cleaner_inode->links->len)
+ continue;
+ BlueSkyCleanerLink *cb = &g_array_index(cleaner_inode->links,
+ BlueSkyCleanerLink, i);
+ if (memcmp(&b->ref->id, &cb->id, sizeof(BlueSkyCloudPointer)) != 0)
+ continue;
+
+ g_print(" Updating block %d pointer\n", i);
+ b->ref->location = cb->location;
+ }
+ }
+ cleaner_flush_inode(proxy_inode);
+ g_mutex_unlock(proxy_inode->lock);
+
+ /* Mark the inode as modified so it will get written back to the cloud. We
+ * don't actually need to force a synchronous write to our local journal
+ * since there have been no logical modifications. */
+#if 0
+ 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);
+ g_mutex_unlock(inode->fs->lock);
+#endif
+}
+
+void bluesky_cleaner_merge(BlueSkyFS *fs)
+{
+ BlueSkyCleanerItem *checkpoint = bluesky_cleaner_find_checkpoint(fs);
+
+ if (checkpoint == NULL) {
+ g_warning("Unable to load cleaner checkpoint record!");
+ return;
+ }
+
+ if (checkpoint->type != LOGTYPE_CHECKPOINT) {
+ g_warning("Last cleaner object not a checkpoint; cleaning probably in progress.");
+ bluesky_cleaner_item_free(checkpoint);
+ return;
+ }
+
+ /* Iterate over each of the inode map sections in the checkpoint */
+ for (int i = 0; i < checkpoint->links->len; i++) {
+ BlueSkyCleanerLink *link = &g_array_index(checkpoint->links,
+ BlueSkyCleanerLink, i);
+ if (!needs_merging(fs, link->location))
+ continue;
+
+ BlueSkyCleanerItem *imap = cleaner_load_item(fs, link->location);
+ if (imap == NULL) {
+ g_warning("Unable to load cleaner inode map");
+ continue;
+ }
+
+ /* Iterate over all inodes found in the inode map section */
+ for (int j = 0; j < imap->links->len; j++) {
+ BlueSkyCleanerLink *link = &g_array_index(imap->links,
+ BlueSkyCleanerLink, j);
+ if (!needs_merging(fs, link->location))
+ continue;
+ BlueSkyCleanerItem *inode = cleaner_load_item(fs, link->location);
+ if (inode != NULL) {
+ merge_inode(fs, inode);
+ }
+ bluesky_cleaner_item_free(inode);
+ }
+
+ bluesky_cleaner_item_free(imap);
+ }
+
+ fs->log_state->latest_cleaner_seq_seen = checkpoint->location.sequence;
+ bluesky_cleaner_item_free(checkpoint);
+}
+
+/* Run the cleaner as a background task. */
+static gpointer cleaner_thread(BlueSkyFS *fs)
+{
+ while (TRUE) {
+ struct timespec delay;
+ delay.tv_sec = 30;
+ delay.tv_nsec = 0;
+ nanosleep(&delay, NULL);
+ bluesky_cleaner_merge(fs);
+ }
+
+ return NULL;
+}
+
+void bluesky_cleaner_thread_launch(BlueSkyFS *fs)
+{
+ g_thread_create((GThreadFunc)cleaner_thread, fs, FALSE, NULL);
}
projects / bluesky.git / blobdiff