X-Git-Url: http://git.vrable.net/?a=blobdiff_plain;f=bluesky%2Fcloudlog.c;h=ec168e82380aaaa4f94234c8d65bded9c3c157ab;hb=7daddca3715fee8a3bfff417e769ff08dd7f4cbf;hp=bf7d22eb384a3526ec306d56f6949db721e1046e;hpb=6732e9143e2c773a8c94947713bc8fc22613f702;p=bluesky.git

diff --git a/bluesky/cloudlog.c b/bluesky/cloudlog.c
index bf7d22e..ec168e8 100644
--- a/bluesky/cloudlog.c
+++ b/bluesky/cloudlog.c
@@ -147,7 +147,10 @@ void bluesky_cloudlog_unref(BlueSkyCloudLog *log)
             return;
         }
 
-        g_hash_table_remove(fs->locations, &log->id);
+        if (!g_hash_table_remove(fs->locations, &log->id)) {
+            if (bluesky_verbose)
+                g_warning("Could not find and remove cloud log item from hash table!");
+        }
         g_mutex_unlock(fs->lock);
 
         bluesky_cloudlog_stats_update(log, -1);
@@ -165,6 +168,61 @@ void bluesky_cloudlog_unref(BlueSkyCloudLog *log)
     }
 }
 
+/* For locking reasons cloudlog unrefs may sometimes need to be performed in
+ * the future.  We launch a thread for handling these delayed unreference
+ * requests. */
+static gpointer cloudlog_unref_thread(gpointer q)
+{
+    GAsyncQueue *queue = (GAsyncQueue *)q;
+
+    while (TRUE) {
+        BlueSkyCloudLog *item = (BlueSkyCloudLog *)g_async_queue_pop(queue);
+        bluesky_cloudlog_unref(item);
+    }
+
+    return NULL;
+}
+
+void bluesky_cloudlog_unref_delayed(BlueSkyCloudLog *log)
+{
+    if (log != NULL)
+        g_async_queue_push(log->fs->unref_queue, log);
+}
+
+void bluesky_cloudlog_threads_init(BlueSkyFS *fs)
+{
+    fs->unref_queue = g_async_queue_new();
+    g_thread_create(cloudlog_unref_thread, fs->unref_queue, FALSE, NULL);
+}
+
+/* Erase the information contained within the in-memory cloud log
+ * representation.  This does not free up the item itself, but frees the data
+ * and references to other log items and resets the type back to unknown.  If
+ * the object was written out to persistent storage, all state about it can be
+ * recovered by loading the object back in.  The object must be locked before
+ * calling this function. */
+void bluesky_cloudlog_erase(BlueSkyCloudLog *log)
+{
+    g_assert(log->data_lock_count == 0);
+
+    if (log->type == LOGTYPE_UNKNOWN)
+        return;
+
+    log->type = LOGTYPE_UNKNOWN;
+    log->data_size = 0;
+    bluesky_string_unref(log->data);
+    log->data = NULL;
+    log->data_lock_count = 0;
+
+    for (int i = 0; i < log->links->len; i++) {
+        BlueSkyCloudLog *c = g_array_index(log->links,
+                                           BlueSkyCloudLog *, i);
+        bluesky_cloudlog_unref(c);
+    }
+    g_array_unref(log->links);
+    log->links = g_array_new(FALSE, TRUE, sizeof(BlueSkyCloudLog *));
+}
+
 /* Start a write of the object to the local log. */
 void bluesky_cloudlog_sync(BlueSkyCloudLog *log)
 {
@@ -173,29 +231,41 @@ void bluesky_cloudlog_sync(BlueSkyCloudLog *log)
 
 /* Add the given entry to the global hash table containing cloud log entries.
  * Takes ownership of the caller's reference. */
+void bluesky_cloudlog_insert_locked(BlueSkyCloudLog *log)
+{
+    g_hash_table_insert(log->fs->locations, &log->id, log);
+}
+
 void bluesky_cloudlog_insert(BlueSkyCloudLog *log)
 {
     g_mutex_lock(log->fs->lock);
-    g_hash_table_insert(log->fs->locations, &log->id, log);
+    bluesky_cloudlog_insert_locked(log);
     g_mutex_unlock(log->fs->lock);
 }
 
-struct log_header {
-    char magic[4];
-    uint32_t size;
-    BlueSkyCloudID id;
-    uint32_t pointer_count;
-} __attribute__((packed));
+/* Look up the cloud log entry for the given ID.  If create is TRUE and the
+ * item does not exist, create a special pending entry that can later be filled
+ * in when the real item is loaded.  The returned item has a reference held.
+ * As a special case, if a null ID is provided then NULL is returned. */
+BlueSkyCloudLog *bluesky_cloudlog_get(BlueSkyFS *fs, BlueSkyCloudID id)
+{
+    static BlueSkyCloudID id0 = {{0}};
 
-struct logref {
-    BlueSkyCloudID id;
-    BlueSkyCloudPointer location;
-} __attribute__((packed));
+    if (memcmp(&id, &id0, sizeof(BlueSkyCloudID)) == 0)
+        return NULL;
 
-struct log_footer {
-    char refmagic[4];
-    struct logref refs[0];
-};
+    g_mutex_lock(fs->lock);
+    BlueSkyCloudLog *item;
+    item = g_hash_table_lookup(fs->locations, &id);
+    if (item == NULL) {
+        item = bluesky_cloudlog_new(fs, &id);
+        bluesky_cloudlog_insert_locked(item);
+    } else {
+        bluesky_cloudlog_ref(item);
+    }
+    g_mutex_unlock(fs->lock);
+    return item;
+}
 
 /* Ensure that a cloud log item is loaded in memory, and if not read it in.
  * TODO: Make asynchronous, and make this also fetch from the cloud.  Right now
@@ -205,20 +275,52 @@ void bluesky_cloudlog_fetch(BlueSkyCloudLog *log)
     if (log->data != NULL)
         return;
 
+    /* There are actually two cases: a full deserialization if we have not ever
+     * read the object before, and a partial deserialization where the metadata
+     * is already in memory and we just need to remap the data.  If the object
+     * type has not yet been set, we'll need to read and parse the metadata.
+     * Once that is done, we can fall through the case of remapping the data
+     * itself. */
+    if (log->type == LOGTYPE_UNKNOWN) {
+        BlueSkyRCStr *raw = NULL;
+        if ((log->location_flags | log->pending_write) & CLOUDLOG_JOURNAL) {
+            raw = bluesky_log_map_object(log->fs, -1, log->log_seq,
+                                         log->log_offset, log->log_size);
+        }
+
+        if (raw == NULL && (log->location_flags & CLOUDLOG_CLOUD)) {
+            log->location_flags &= ~CLOUDLOG_JOURNAL;
+            raw = bluesky_log_map_object(log->fs,
+                                         log->location.directory,
+                                         log->location.sequence,
+                                         log->location.offset,
+                                         log->location.size);
+        }
+
+        g_assert(raw != NULL);
+        bluesky_deserialize_cloudlog(log, raw->data, raw->len);
+        bluesky_string_unref(raw);
+    }
+
+    /* At this point all metadata should be available and we need only remap
+     * the object data. */
+
+    int offset;
     if ((log->location_flags | log->pending_write) & CLOUDLOG_JOURNAL) {
         bluesky_cloudlog_stats_update(log, -1);
+        offset = log->log_offset + sizeof(struct log_header);
         log->data = bluesky_log_map_object(log->fs, -1, log->log_seq,
-                                           log->log_offset, log->log_size);
+                                           offset, log->data_size);
         bluesky_cloudlog_stats_update(log, 1);
     }
 
     if (log->data == NULL && (log->location_flags & CLOUDLOG_CLOUD)) {
         log->location_flags &= ~CLOUDLOG_JOURNAL;
         bluesky_cloudlog_stats_update(log, -1);
+        offset = log->location.offset + sizeof(struct cloudlog_header);
         log->data = bluesky_log_map_object(log->fs, log->location.directory,
                                            log->location.sequence,
-                                           log->location.offset,
-                                           log->location.size);
+                                           offset, log->data_size);
         bluesky_cloudlog_stats_update(log, 1);
     }
 
@@ -251,30 +353,44 @@ BlueSkyCloudPointer bluesky_cloudlog_serialize(BlueSkyCloudLog *log,
 
     bluesky_cloudlog_stats_update(log, -1);
 
-    /* TODO: Right now offset/size are set to the raw data, but we should add
-     * header parsing to the code which loads objects back in. */
+    GString *data1 = g_string_new("");
+    GString *data2 = g_string_new("");
+    GString *data3 = g_string_new("");
+    bluesky_serialize_cloudlog(log, data1, data2, data3);
+
     log->location = state->location;
-    log->location.offset = state->data->len + sizeof(struct log_header);
-    log->location.size = log->data->len;
-        /* = sizeof(struct log_header) + sizeof(BlueSkyCloudID) * 0
-           + log->data->len; */
-
-    struct log_header header;
-    memcpy(header.magic, "AgI ", 4);
-    header.size = GUINT32_TO_LE(log->location.size);
+    log->location.offset = state->data->len;
+    log->data_size = data1->len;
+
+    struct cloudlog_header header;
+    memcpy(header.magic, CLOUDLOG_MAGIC, 4);
+    memset(header.crypt_auth, sizeof(header.crypt_auth), 0);
+    memset(header.crypt_iv, sizeof(header.crypt_iv), 0);
+    header.type = log->type + '0';
+    header.size1 = GUINT32_TO_LE(data1->len);
+    header.size2 = GUINT32_TO_LE(data2->len);
+    header.size3 = GUINT32_TO_LE(data3->len);
     header.id = log->id;
-    header.pointer_count = GUINT32_TO_LE(0);
+    header.inum = GUINT64_TO_LE(log->inum);
 
     g_string_append_len(state->data, (const char *)&header, sizeof(header));
-    g_string_append_len(state->data, log->data->data, log->data->len);
+    g_string_append_len(state->data, data1->str, data1->len);
+    g_string_append_len(state->data, data2->str, data2->len);
+    g_string_append_len(state->data, data3->str, data3->len);
+
+    log->location.size = state->data->len - log->location.offset;
+
+    g_string_free(data1, TRUE);
+    g_string_free(data2, TRUE);
+    g_string_free(data3, TRUE);
 
     /* If the object we flushed was an inode, update the inode map. */
     if (log->type == LOGTYPE_INODE) {
         g_mutex_lock(fs->lock);
         InodeMapEntry *entry = bluesky_inode_map_lookup(fs->inode_map,
                                                         log->inum, 1);
-        entry->id = log->id;
-        entry->location = log->location;
+        entry->item = log;
+        bluesky_cloudlog_ref(entry->item);
         g_mutex_unlock(fs->lock);
     }