Add partial journal replay to filesystem recovery.

After loading a cloud checkpoint, replay just the last portion of the
journal that may not have been committed to the cloud.

diff --git a/bluesky/bluesky-private.h b/bluesky/bluesky-private.h
index 5bb18f9..85eb99c 100644 (file)
--- a/bluesky/bluesky-private.h
+++ b/bluesky/bluesky-private.h
@@ -187,6 +187,11 @@ typedef enum {
     LOGTYPE_INODE = 2,
     LOGTYPE_INODE_MAP = 3,
     LOGTYPE_CHECKPOINT = 4,
+
+    /* Used only as metadata in the local journal, not loaded as a
+     * BlueSkyCloudLogState nor stored in the cloud */
+    LOGTYPE_JOURNAL_MARKER = 16,
+    LOGTYPE_JOURNAL_CHECKPOINT = 17,
 } BlueSkyCloudLogType;
 
 /* Headers that go on items in local log segments and cloud log segments. */
@@ -344,6 +349,9 @@ struct _BlueSkyCacheFile {
 BlueSkyLog *bluesky_log_new(const char *log_directory);
 void bluesky_log_item_submit(BlueSkyCloudLog *item, BlueSkyLog *log);
 void bluesky_log_finish_all(GList *log_items);
+BlueSkyCloudLog *bluesky_log_get_commit_point(BlueSkyFS *fs);
+void bluesky_log_write_commit_point(BlueSkyFS *fs, BlueSkyCloudLog *marker);
+
 BlueSkyRCStr *bluesky_log_map_object(BlueSkyFS *fs, int log_dir, int log_seq,
                                      int log_offset, int log_size);
 void bluesky_mmap_unref(BlueSkyCacheFile *mmap);

diff --git a/bluesky/cache.c b/bluesky/cache.c
index 454e06f..c27d39d 100644 (file)
--- a/bluesky/cache.c
+++ b/bluesky/cache.c
@@ -114,6 +114,7 @@ static void flushd_cloud(BlueSkyFS *fs)
 
     /* TODO: Locking?  Since we're reading a single variable this is probably
      * atomic but a lock could be safer. */
+    BlueSkyCloudLog *marker = bluesky_log_get_commit_point(fs);
     int journal_seq_start = fs->log->seq_num;
 
     while (1) {
@@ -177,6 +178,8 @@ static void flushd_cloud(BlueSkyFS *fs)
                                  fs->log_state->pending_segments);
     }
 
+    bluesky_log_write_commit_point(fs, marker);
+
     g_print("All segments have been flushed, journal < %d is clean\n",
             journal_seq_start);
 

diff --git a/bluesky/inode.c b/bluesky/inode.c
index 41227fb..e11bfeb 100644 (file)
--- a/bluesky/inode.c
+++ b/bluesky/inode.c
@@ -106,7 +106,7 @@ BlueSkyFS *bluesky_init_fs(gchar *name, BlueSkyStore *store)
 
     if (bluesky_checkpoint_load(fs)) {
         g_print("Filesystem checkpoint loaded, starting journal replay...\n");
-        //bluesky_replay(fs);
+        bluesky_replay(fs);
         g_print("Journal replay complete, filesystem ready.\n");
     } else {
         /* Initialize a fresh filesystem */

diff --git a/bluesky/log.c b/bluesky/log.c
index 82db346..97b4f30 100644 (file)
--- a/bluesky/log.c
+++ b/bluesky/log.c
@@ -285,6 +285,46 @@ void bluesky_log_finish_all(GList *log_items)
     }
 }
 
+/* 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_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_sync(commit);
+
+    g_mutex_lock(commit->lock);
+    while ((commit->pending_write & CLOUDLOG_JOURNAL))
+        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;
@@ -632,7 +672,8 @@ static gboolean validate_journal_item(const char *buf, size_t len, off_t offset)
 
 /* 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;
@@ -642,6 +683,13 @@ static void bluesky_replay_scan_journal(const char *buf, size_t len)
         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);
     }
 }
@@ -690,11 +738,11 @@ static void reload_item(BlueSkyCloudLog *log_item,
 }
 
 static void bluesky_replay_scan_journal2(BlueSkyFS *fs, GList **objects,
-                                         int log_seq,
+                                         int log_seq, int start_offset,
                                          const char *buf, size_t len)
 {
     const struct log_header *header;
-    off_t offset = 0;
+    off_t offset = start_offset;
 
     while (validate_journal_item(buf, len, offset)) {
         header = (const struct log_header *)(buf + offset);
@@ -765,6 +813,7 @@ void bluesky_replay(BlueSkyFS *fs)
     /* 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);
@@ -774,13 +823,16 @@ void bluesky_replay(BlueSkyFS *fs)
             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_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);
@@ -791,11 +843,10 @@ void bluesky_replay(BlueSkyFS *fs)
      * 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);
+        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) {
@@ -803,11 +854,12 @@ void bluesky_replay(BlueSkyFS *fs)
             break;
         }
 
-        bluesky_replay_scan_journal2(fs, &objects, seq_num,
+        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) {