X-Git-Url: http://git.vrable.net/?a=blobdiff_plain;f=bluesky%2Fcloudlog.c;h=0d91e0dcb43bdaf456ec56367d0a9413fb752567;hb=ac5113f4988176ca8b4f8f575df82ba7d05495cd;hp=82f0f65fd80be70f57fccbbb9fe508363530b7ea;hpb=e9af23b41ae0c11245fbc6886a794e1dfb04a6bd;p=bluesky.git

diff --git a/bluesky/cloudlog.c b/bluesky/cloudlog.c
index 82f0f65..0d91e0d 100644
--- a/bluesky/cloudlog.c
+++ b/bluesky/cloudlog.c
@@ -97,15 +97,46 @@ BlueSkyCloudLog *bluesky_cloudlog_new(BlueSkyFS *fs)
     return log;
 }
 
+/* The reference held by the hash table does not count towards the reference
+ * count.  When a new object is created, it initially has a reference count of
+ * 1 for the creator, and similarly fetching an item from the hash table will
+ * also create a reference.  If the reference count drops to zero,
+ * bluesky_cloudlog_unref attempts to remove the object from the hash
+ * table--but there is a potential race since another thread might read the
+ * object from the hash table at the same time.  So an object with a reference
+ * count of zero may still be resurrected, in which case we need to abort the
+ * destruction.  Once the object is gone from the hash table, and if the
+ * reference count is still zero, it can actually be deleted. */
 void bluesky_cloudlog_ref(BlueSkyCloudLog *log)
 {
+    if (log == NULL)
+        return;
+
     g_atomic_int_inc(&log->refcount);
 }
 
 void bluesky_cloudlog_unref(BlueSkyCloudLog *log)
 {
+    if (log == NULL)
+        return;
+
     if (g_atomic_int_dec_and_test(&log->refcount)) {
-        g_print("Cloud log refcount dropped to zero.\n");
+        BlueSkyFS *fs = log->fs;
+
+        g_mutex_lock(fs->lock);
+        if (g_atomic_int_get(&log->refcount) > 0) {
+            g_mutex_unlock(fs->lock);
+            return;
+        }
+
+        g_hash_table_remove(fs->locations, &log->id);
+        g_mutex_unlock(fs->lock);
+
+        g_mutex_free(log->lock);
+        g_cond_free(log->cond);
+        g_array_unref(log->pointers);
+        bluesky_string_unref(log->data);
+        g_free(log);
     }
 }
 
@@ -159,11 +190,15 @@ BlueSkyCloudPointer bluesky_cloudlog_serialize(BlueSkyCloudLog *log,
         BlueSkyCloudLog *log2
             = (BlueSkyCloudLog *)g_hash_table_lookup(log->fs->locations, &id);
         // TODO: refcount
-        g_mutex_unlock(log->fs->lock);
         g_assert(log2 != NULL);
+        bluesky_cloudlog_ref(log2);
+        g_mutex_unlock(log->fs->lock);
         bluesky_cloudlog_serialize(log2, state);
+        bluesky_cloudlog_unref(log2);
     }
 
+    g_mutex_lock(log->lock);
+    bluesky_cloudlog_fetch(log);
     g_assert(log->data != NULL);
 
     log->location = state->location;
@@ -182,6 +217,7 @@ BlueSkyCloudPointer bluesky_cloudlog_serialize(BlueSkyCloudLog *log,
     g_string_append_len(state->data, log->data->data, log->data->len);
 
     log->location_flags |= CLOUDLOG_CLOUD;
+    g_mutex_unlock(log->lock);
 
     return log->location;
 }
@@ -200,10 +236,9 @@ static void find_inodes(gpointer key, gpointer value, gpointer user_data)
 
 void bluesky_cloudlog_write_log(BlueSkyFS *fs)
 {
-    g_print("Starting cloudlog write...\n");
-
     BlueSkyCloudLogState *state = fs->log_state;
-    state->data = g_string_new("");
+    if (state->data == NULL)
+        state->data = g_string_new("");
 
     g_mutex_lock(fs->lock);
     g_hash_table_foreach(fs->locations, find_inodes, state);
@@ -216,19 +251,41 @@ void bluesky_cloudlog_write_log(BlueSkyFS *fs)
                                                state->inode_list);
     }
 
-    g_print("Serialized %zd bytes of data\n", state->data->len);
-
-    BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
-    async->op = STORE_OP_PUT;
-    async->key = g_strdup_printf("log-%08d-%08d",
-                                 state->location.directory,
-                                 state->location.sequence);
-    async->data = bluesky_string_new_from_gstring(state->data);
-    bluesky_store_async_submit(async);
-    bluesky_store_async_wait(async);
-    bluesky_store_async_unref(async);
+    if (state->data->len > 0) {
+        g_print("Serialized %zd bytes of data to cloud\n", state->data->len);
+
+        BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
+        async->op = STORE_OP_PUT;
+        async->key = g_strdup_printf("log-%08d-%08d",
+                                     state->location.directory,
+                                     state->location.sequence);
+        async->data = bluesky_string_new_from_gstring(state->data);
+        bluesky_store_async_submit(async);
+        bluesky_store_async_wait(async);
+        bluesky_store_async_unref(async);
+
+        state->location.sequence++;
+        state->location.offset = 0;
+    }
 
     state->data = NULL;
-    state->location.sequence++;
-    state->location.offset = 0;
+}
+
+/* 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
+ * we only read from the log.  Log item must be locked. */
+void bluesky_cloudlog_fetch(BlueSkyCloudLog *log)
+{
+    if (log->data != NULL)
+        return;
+
+    g_print("Re-mapping log entry %d/%d/%d...\n",
+            log->log_seq, log->log_offset, log->log_size);
+
+    g_assert(log->location_flags & CLOUDLOG_JOURNAL);
+
+    log->data = bluesky_log_map_object(log->fs->log, log->log_seq,
+                                       log->log_offset, log->log_size);
+
+    g_cond_broadcast(log->cond);
 }