X-Git-Url: http://git.vrable.net/?a=blobdiff_plain;f=bluesky%2Fcache.c;h=d4d410e6811ddb6725e7c175df4346935f94eb89;hb=3380ff6d419c7d1370c95cf476520ed5398e5b6d;hp=6a91873c3750002e9956782a802d886ecf7e11f2;hpb=60b4792d65ba4b2a45733894f6a57e6581ddc487;p=bluesky.git

diff --git a/bluesky/cache.c b/bluesky/cache.c
index 6a91873..d4d410e 100644
--- a/bluesky/cache.c
+++ b/bluesky/cache.c
@@ -15,155 +15,32 @@
 
 #define WRITEBACK_DELAY (20 * 1000000)
 
-/* Filesystem caching and cache coherency. */
-
-static void writeback_complete(gpointer a, gpointer i)
-{
-    BlueSkyInode *inode = (BlueSkyInode *)i;
-
-    if (bluesky_verbose) {
-        g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
-              "Writeback for inode %"PRIu64" complete", inode->inum);
-    }
-
-    g_mutex_lock(inode->lock);
-
-    inode->change_commit = inode->change_pending;
-    inode->change_pending = 0;
-    if (inode->change_count == inode->change_commit) {
-        /* If inode is no longer dirty... */
-        inode->change_time = 0;
-        g_mutex_lock(inode->fs->lock);
-        bluesky_list_unlink(&inode->fs->dirty_list, inode->dirty_list);
-        inode->dirty_list = NULL;
-        g_mutex_unlock(inode->fs->lock);
-    }
-
-    g_mutex_unlock(inode->lock);
-}
-
-#if 0
-static void flushd_inode(gpointer value, gpointer user_data)
-{
-    BlueSkyFS *fs = (BlueSkyFS *)user_data;
-
-    BlueSkyInode *inode = (BlueSkyInode *)value;
-
-    g_mutex_lock(inode->lock);
-
-    if (inode->change_count == inode->change_commit) {
-        uint64_t delay = bluesky_get_current_time() - inode->access_time;
-        if (delay >= CACHE_CLEAN_DELAY) {
-            drop_caches(inode);
-
-            /* If the only references are the one we hold and the one in the
-             * filesystem inum->inode hash table...  First check the refcount
-             * without the lock for speed, but if the check looks good verify
-             * it after taking the filesystem lock. */
-            if (inode->refcount == 2) {
-                g_mutex_lock(fs->lock);
-                if (inode->refcount == 2) {
-                    g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
-                          "Trying to drop inode %"PRIu64" from cache",
-                          inode->inum);
-                    if (g_hash_table_remove(fs->inodes, &inode->inum))
-                        bluesky_inode_unref(inode);
-                }
-                bluesky_list_unlink(&inode->fs->accessed_list,
-                                    inode->accessed_list);
-                inode->accessed_list = NULL;
-                bluesky_list_unlink(&inode->fs->dirty_list,
-                                    inode->dirty_list);
-                inode->dirty_list = NULL;
-                g_mutex_unlock(fs->lock);
-            }
-        }
-
-        g_mutex_unlock(inode->lock);
-        bluesky_inode_unref(inode);
-        return;
-    }
-
-    if (inode->change_pending) {
-        /* Waiting for an earlier writeback to finish, so don't start a new
-         * writeback yet. */
-        g_mutex_unlock(inode->lock);
-        bluesky_inode_unref(inode);
-        return;
-    }
-
-    uint64_t elapsed = bluesky_get_current_time() - inode->change_time;
-    if (elapsed < WRITEBACK_DELAY) {
-        /* Give a bit more time before starting writeback. */
-        g_mutex_unlock(inode->lock);
-        bluesky_inode_unref(inode);
-        return;
-    }
-
-    inode->change_pending = inode->change_count;
-
-    g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
-          "Starting flush of inode %"PRIu64, inode->inum);
-
-    /* Create a store barrier.  All operations part of the writeback will be
-     * added to this barrier, so when the barrier completes we know that the
-     * writeback is finished. */
-    BlueSkyStoreAsync *barrier = bluesky_store_async_new(fs->store);
-    barrier->op = STORE_OP_BARRIER;
-
-    bluesky_inode_start_sync(inode, barrier);
-
-    bluesky_store_async_add_notifier(barrier, writeback_complete, inode);
-    bluesky_store_async_submit(barrier);
-    bluesky_store_async_unref(barrier);
-
-    g_mutex_unlock(inode->lock);
-    bluesky_inode_unref(inode);
-}
-#endif
+/* Filesystem caching and cache coherency.  There are actually a couple of
+ * different tasks that are performed here:
+ *   - Forcing data to the log if needed to reclaim memory or simply if the
+ *     data has been dirty in memory long enough.
+ *   - Writing batches of data to the cloud.
+ */
 
 static void flushd_dirty_inode(BlueSkyInode *inode)
 {
     BlueSkyFS *fs = inode->fs;
 
     g_mutex_lock(fs->lock);
-    bluesky_list_unlink(&fs->dirty_list, inode->dirty_list);
-    inode->dirty_list = NULL;
+    bluesky_list_unlink(&fs->unlogged_list, inode->unlogged_list);
+    inode->unlogged_list = NULL;
     g_mutex_unlock(fs->lock);
 
     /* Inode is clean; nothing to do. */
     if (inode->change_count == inode->change_commit)
         return;
 
-    /* Inode writeback is in progress; put back on the dirty list. */
-    if (inode->change_pending) {
-        /* Waiting for an earlier writeback to finish, so don't start a new
-         * writeback yet. */
-        g_mutex_lock(fs->lock);
-        inode->change_time = bluesky_get_current_time();
-        bluesky_list_unlink(&fs->dirty_list, inode->dirty_list);
-        inode->dirty_list = bluesky_list_prepend(&fs->dirty_list, inode);
-        g_mutex_unlock(fs->lock);
-        return;
-    }
-
     if (bluesky_verbose) {
         g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
             "Starting flush of inode %"PRIu64, inode->inum);
     }
-    inode->change_pending = inode->change_count;
-
-    /* Create a store barrier.  All operations part of the writeback will be
-     * added to this barrier, so when the barrier completes we know that the
-     * writeback is finished. */
-    BlueSkyStoreAsync *barrier = bluesky_store_async_new(fs->store);
-    barrier->op = STORE_OP_BARRIER;
-
-    bluesky_inode_start_sync(inode, barrier);
 
-    bluesky_store_async_add_notifier(barrier, writeback_complete, inode);
-    bluesky_store_async_submit(barrier);
-    bluesky_store_async_unref(barrier);
+    bluesky_inode_start_sync(inode);
 }
 
 /* Try to flush dirty data to disk, either due to memory pressure or due to
@@ -173,6 +50,48 @@ static void flushd_dirty(BlueSkyFS *fs)
     int64_t start_time = bluesky_get_current_time();
     g_mutex_lock(fs->lock);
 
+    while (1) {
+        BlueSkyInode *inode;
+        if (fs->unlogged_list.prev == NULL)
+            break;
+        inode = fs->unlogged_list.prev->data;
+
+        if (bluesky_verbose) {
+            g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
+                  "Considering flushing inode %"PRIu64, inode->inum);
+        }
+
+        /* Stop processing dirty inodes if we both have enough memory available
+         * and the oldest inode is sufficiently new that it need not be flushed
+         * out. */
+        uint64_t elapsed = bluesky_get_current_time() - inode->change_time;
+        if (g_atomic_int_get(&fs->cache_dirty) < bluesky_watermark_low_dirty
+                && elapsed < WRITEBACK_DELAY)
+            break;
+        if (inode->change_time > start_time)
+            break;
+
+        bluesky_inode_ref(inode);
+
+        g_mutex_unlock(fs->lock);
+
+        g_mutex_lock(inode->lock);
+        flushd_dirty_inode(inode);
+        g_mutex_unlock(inode->lock);
+        bluesky_inode_unref(inode);
+
+        g_mutex_lock(fs->lock);
+    }
+
+    g_mutex_unlock(fs->lock);
+}
+
+/* Try to flush dirty data to the cloud. */
+static void flushd_cloud(BlueSkyFS *fs)
+{
+    int64_t start_time = bluesky_get_current_time();
+    g_mutex_lock(fs->lock);
+
     while (1) {
         BlueSkyInode *inode;
         if (fs->dirty_list.prev == NULL)
@@ -181,7 +100,7 @@ static void flushd_dirty(BlueSkyFS *fs)
 
         if (bluesky_verbose) {
             g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
-                  "Considering flushing inode %"PRIu64, inode->inum);
+                  "Flushing inode %"PRIu64" to cloud", inode->inum);
         }
 
         /* Stop processing dirty inodes if we both have enough memory available
@@ -200,13 +119,25 @@ static void flushd_dirty(BlueSkyFS *fs)
 
         g_mutex_lock(inode->lock);
         flushd_dirty_inode(inode);
+        g_mutex_lock(fs->lock);
+        bluesky_list_unlink(&fs->dirty_list, inode->dirty_list);
+        inode->dirty_list = NULL;
+        g_mutex_unlock(fs->lock);
+
+        BlueSkyCloudLog *log = inode->committed_item;
+        bluesky_cloudlog_ref(log);
         g_mutex_unlock(inode->lock);
+
+        if (log != NULL)
+            bluesky_cloudlog_serialize(log, fs);
         bluesky_inode_unref(inode);
+        bluesky_cloudlog_ref(log);
 
         g_mutex_lock(fs->lock);
     }
 
     g_mutex_unlock(fs->lock);
+    bluesky_cloudlog_flush(fs);
 }
 
 /* Drop cached data for a given inode, if it is clean.  inode must be locked. */
@@ -216,7 +147,7 @@ static void drop_caches(BlueSkyInode *inode)
         bluesky_file_drop_cached(inode);
 }
 
-/* Drop clean data fromt the cache if needed due to memory pressure. */
+/* Drop clean data from the cache if needed due to memory pressure. */
 static void flushd_clean(BlueSkyFS *fs)
 {
     g_mutex_lock(fs->lock);
@@ -226,8 +157,10 @@ static void flushd_clean(BlueSkyFS *fs)
         inode_count = 1;
 
     while (inode_count-- > 0) {
+#if 0
         if (g_atomic_int_get(&fs->cache_total) < bluesky_watermark_medium_total)
             break;
+#endif
 
         BlueSkyInode *inode;
         if (fs->accessed_list.prev == NULL)
@@ -269,8 +202,8 @@ static gpointer flushd_task(BlueSkyFS *fs)
     if (!g_mutex_trylock(fs->flushd_lock))
         return NULL;
     flushd_dirty(fs);
+    flushd_cloud(fs);
     flushd_clean(fs);
-    bluesky_cloudlog_write_log(fs);
     g_mutex_unlock(fs->flushd_lock);
 
     return NULL;