X-Git-Url: http://git.vrable.net/?a=blobdiff_plain;f=bluesky%2Fcloudlog.c;h=134f3d798600f839e8f52e01ba8b958c385800c9;hb=ddaec40a37a5e65e53546b14632b1b0b35613264;hp=82f0f65fd80be70f57fccbbb9fe508363530b7ea;hpb=e9af23b41ae0c11245fbc6886a794e1dfb04a6bd;p=bluesky.git

diff --git a/bluesky/cloudlog.c b/bluesky/cloudlog.c
index 82f0f65..134f3d7 100644
--- a/bluesky/cloudlog.c
+++ b/bluesky/cloudlog.c
@@ -13,13 +13,9 @@
 
 #include "bluesky-private.h"
 
-/* The locations hash table in the file system is used to map objects to their locations.  Objects are named using 128- */
-
-typedef struct {
-    BlueSkyCloudID id;
-
-    BlueSkyCloudPointer *cloud_loc;
-} BlueSkyLocationEntry;
+// 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 CLOUDLOG_SEGMENT_SIZE (4 << 20)
 
 BlueSkyCloudID bluesky_cloudlog_new_id()
 {
@@ -91,21 +87,58 @@ BlueSkyCloudLog *bluesky_cloudlog_new(BlueSkyFS *fs)
     log->fs = fs;
     log->type = LOGTYPE_UNKNOWN;
     log->id = bluesky_cloudlog_new_id();
-    log->pointers = g_array_new(FALSE, TRUE, sizeof(BlueSkyCloudID));
+    log->links = g_array_new(FALSE, TRUE, sizeof(BlueSkyCloudLog *));
     g_atomic_int_set(&log->refcount, 1);
 
     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);
+
+        log->type = LOGTYPE_INVALID;
+        g_mutex_free(log->lock);
+        g_cond_free(log->cond);
+        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);
+        bluesky_string_unref(log->data);
+        g_free(log);
     }
 }
 
@@ -141,29 +174,40 @@ struct log_footer {
     struct logref refs[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_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);
+}
+
 BlueSkyCloudPointer bluesky_cloudlog_serialize(BlueSkyCloudLog *log,
-                                               BlueSkyCloudLogState *state)
+                                               BlueSkyFS *fs)
 {
+    BlueSkyCloudLogState *state = fs->log_state;
+
     if (log->location_flags & CLOUDLOG_CLOUD) {
         return log->location;
     }
 
-    g_print("Flushing object %s to cloud...\n",
-            bluesky_cloudlog_id_to_string(log->id));
-
-    for (int i = 0; i < log->pointers->len; i++) {
-        BlueSkyCloudID id = g_array_index(log->pointers, BlueSkyCloudID, i);
-        g_print("  ...checking reference %s...\n",
-                bluesky_cloudlog_id_to_string(id));
-        g_mutex_lock(log->fs->lock);
-        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_serialize(log2, state);
+    for (int i = 0; i < log->links->len; i++) {
+        BlueSkyCloudLog *ref = g_array_index(log->links,
+                                             BlueSkyCloudLog *, i);
+        if (ref != NULL)
+            bluesky_cloudlog_serialize(ref, fs);
     }
 
+    g_mutex_lock(log->lock);
+    bluesky_cloudlog_fetch(log);
     g_assert(log->data != NULL);
 
     log->location = state->location;
@@ -182,6 +226,10 @@ 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);
+
+    if (state->data->len > CLOUDLOG_SEGMENT_SIZE)
+        bluesky_cloudlog_flush(fs);
 
     return log->location;
 }
@@ -198,25 +246,16 @@ static void find_inodes(gpointer key, gpointer value, gpointer user_data)
     state->inode_list = g_list_prepend(state->inode_list, item);
 }
 
-void bluesky_cloudlog_write_log(BlueSkyFS *fs)
+/* Finish up a partially-written cloud log segment and flush it to storage. */
+void bluesky_cloudlog_flush(BlueSkyFS *fs)
 {
-    g_print("Starting cloudlog write...\n");
-
     BlueSkyCloudLogState *state = fs->log_state;
-    state->data = g_string_new("");
-
-    g_mutex_lock(fs->lock);
-    g_hash_table_foreach(fs->locations, find_inodes, state);
-    g_mutex_unlock(fs->lock);
+    if (state->data == NULL || state->data->len == 0)
+        return;
 
-    while (state->inode_list != NULL) {
-        BlueSkyCloudLog *log = (BlueSkyCloudLog *)state->inode_list->data;
-        bluesky_cloudlog_serialize(log, state);
-        state->inode_list = g_list_delete_link(state->inode_list,
-                                               state->inode_list);
-    }
+    /* TODO: Append some type of commit record to the log segment? */
 
-    g_print("Serialized %zd bytes of data\n", state->data->len);
+    g_print("Serializing %zd bytes of data to cloud\n", state->data->len);
 
     BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
     async->op = STORE_OP_PUT;
@@ -228,7 +267,31 @@ void bluesky_cloudlog_write_log(BlueSkyFS *fs)
     bluesky_store_async_wait(async);
     bluesky_store_async_unref(async);
 
-    state->data = NULL;
     state->location.sequence++;
     state->location.offset = 0;
+    state->data = g_string_new("");
+}
+
+void bluesky_cloudlog_write_log(BlueSkyFS *fs)
+{
+    BlueSkyCloudLogState *state = fs->log_state;
+    if (state->data == NULL)
+        state->data = g_string_new("");
+
+    g_mutex_lock(fs->lock);
+    g_hash_table_foreach(fs->locations, find_inodes, state);
+    g_mutex_unlock(fs->lock);
+
+    while (state->inode_list != NULL) {
+        BlueSkyCloudLog *log = (BlueSkyCloudLog *)state->inode_list->data;
+        bluesky_cloudlog_serialize(log, fs);
+        bluesky_cloudlog_unref(log);
+        state->inode_list = g_list_delete_link(state->inode_list,
+                                               state->inode_list);
+
+        if (state->data->len > CLOUDLOG_SEGMENT_SIZE)
+            bluesky_cloudlog_flush(fs);
+    }
+
+    bluesky_cloudlog_flush(fs);
 }