X-Git-Url: http://git.vrable.net/?a=blobdiff_plain;f=bluesky%2Fcache.c;h=441e6c4ef50af1b6c643fd9e52b8cdec505217a7;hb=db4915b4f7fc8f74958c4d1891dc69b76cbbe383;hp=f58a949e497058bfd0a5a568d9d37ebeffce15cb;hpb=e26a903ddf80011e3b72a780d7392a8333c996af;p=bluesky.git

diff --git a/bluesky/cache.c b/bluesky/cache.c
index f58a949..441e6c4 100644
--- a/bluesky/cache.c
+++ b/bluesky/cache.c
@@ -13,17 +13,23 @@
 
 #include "bluesky-private.h"
 
-#define WRITEBACK_DELAY (5 * 1000000)
-#define CACHE_CLEAN_DELAY (30 * 1000000)
+#define WRITEBACK_DELAY (20 * 1000000)
 
-/* Filesystem caching and cache coherency. */
+/* 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 writeback_complete(gpointer a, gpointer i)
 {
     BlueSkyInode *inode = (BlueSkyInode *)i;
 
-    g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
-          "Writeback for inode %"PRIu64" complete", inode->inum);
+    if (bluesky_verbose) {
+        g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
+              "Writeback for inode %"PRIu64" complete", inode->inum);
+    }
 
     g_mutex_lock(inode->lock);
 
@@ -41,69 +47,37 @@ static void writeback_complete(gpointer a, gpointer i)
     g_mutex_unlock(inode->lock);
 }
 
-/* Drop cached data for a given inode, if it is clean.  inode must be locked. */
-static void drop_caches(BlueSkyInode *inode)
+static void flushd_dirty_inode(BlueSkyInode *inode)
 {
-    if (inode->type == BLUESKY_REGULAR)
-        bluesky_file_drop_cached(inode);
-}
-
-static void flushd_inode(gpointer value, gpointer user_data)
-{
-    BlueSkyFS *fs = (BlueSkyFS *)user_data;
-
-    BlueSkyInode *inode = (BlueSkyInode *)value;
-
-    g_mutex_lock(inode->lock);
+    BlueSkyFS *fs = inode->fs;
 
-    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);
-                }
-                g_mutex_unlock(fs->lock);
-            }
-        }
+    g_mutex_lock(fs->lock);
+    bluesky_list_unlink(&fs->dirty_list, inode->dirty_list);
+    inode->dirty_list = NULL;
+    g_mutex_unlock(fs->lock);
 
-        g_mutex_unlock(inode->lock);
-        bluesky_inode_unref(inode);
+    /* 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_unlock(inode->lock);
-        bluesky_inode_unref(inode);
+        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;
     }
 
-    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;
+    if (bluesky_verbose) {
+        g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
+            "Starting flush of inode %"PRIu64, inode->inum);
     }
-
     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. */
@@ -115,37 +89,154 @@ static void flushd_inode(gpointer value, gpointer user_data)
     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);
 }
 
-/* Scan through the cache for dirty data and start flushing it to stable
- * storage.  This does not guarantee that data is committed when it returns.
- * Instead, this can be called occasionally to ensure that dirty data is
- * gradually flushed.
- *
- * We do not want to hold the filesystem lock while flushing individual inodes,
- * a that could lead to deadlock.  So first scan through the inode table to get
- * a reference to all inodes, then process that queue of inodes after dropping
- * the filesystem lock. */
-static void gather_inodes(gpointer key, gpointer value, gpointer user_data)
+/* Try to flush dirty data to disk, either due to memory pressure or due to
+ * timeouts. */
+static void flushd_dirty(BlueSkyFS *fs)
 {
-    GSList **list = (GSList **)user_data;
-    *list = g_slist_prepend(*list, value);
-    bluesky_inode_ref((BlueSkyInode *)value);
+    int64_t start_time = bluesky_get_current_time();
+    g_mutex_lock(fs->lock);
+
+    while (1) {
+        BlueSkyInode *inode;
+        if (fs->dirty_list.prev == NULL)
+            break;
+        inode = fs->dirty_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);
 }
 
-void bluesky_flushd_invoke(BlueSkyFS *fs)
+/* Drop cached data for a given inode, if it is clean.  inode must be locked. */
+static void drop_caches(BlueSkyInode *inode)
 {
-    GSList *list = NULL;
+    if (inode->type == BLUESKY_REGULAR)
+        bluesky_file_drop_cached(inode);
+}
 
+/* Drop clean data fromt the cache if needed due to memory pressure. */
+static void flushd_clean(BlueSkyFS *fs)
+{
     g_mutex_lock(fs->lock);
-    g_hash_table_foreach(fs->inodes, gather_inodes, &list);
+
+    size_t inode_count = g_hash_table_size(fs->inodes);
+    if (!inode_count)
+        inode_count = 1;
+
+    while (inode_count-- > 0) {
+        if (g_atomic_int_get(&fs->cache_total) < bluesky_watermark_medium_total)
+            break;
+
+        BlueSkyInode *inode;
+        if (fs->accessed_list.prev == NULL)
+            break;
+        inode = fs->accessed_list.prev->data;
+
+        if (bluesky_verbose) {
+            g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
+                  "Considering dropping cached data for inode %"PRIu64,
+                  inode->inum);
+        }
+
+        bluesky_inode_ref(inode);
+
+        g_mutex_unlock(fs->lock);
+
+        g_mutex_lock(inode->lock);
+
+        g_mutex_lock(fs->lock);
+        bluesky_list_unlink(&fs->accessed_list, inode->accessed_list);
+        inode->accessed_list = bluesky_list_prepend(&fs->accessed_list, inode);
+        g_mutex_unlock(fs->lock);
+
+        drop_caches(inode);
+
+        g_mutex_unlock(inode->lock);
+        bluesky_inode_unref(inode);
+
+        g_mutex_lock(fs->lock);
+    }
+
     g_mutex_unlock(fs->lock);
+}
+
+/* Run the flush daemon for a single iteration, though if it is already
+ * executing returns immediately. */
+static gpointer flushd_task(BlueSkyFS *fs)
+{
+    if (!g_mutex_trylock(fs->flushd_lock))
+        return NULL;
+    flushd_dirty(fs);
+    bluesky_cloudlog_write_log(fs);
+    flushd_clean(fs);
+    g_mutex_unlock(fs->flushd_lock);
+
+    return NULL;
+}
 
-    list = g_slist_reverse(list);
-    g_slist_foreach(list, flushd_inode, fs);
+void bluesky_flushd_invoke(BlueSkyFS *fs)
+{
+    g_thread_create((GThreadFunc)flushd_task, fs, FALSE, NULL);
+}
+
+void bluesky_flushd_invoke_conditional(BlueSkyFS *fs)
+{
+    if (g_atomic_int_get(&fs->cache_dirty) < bluesky_watermark_high_dirty
+        && g_atomic_int_get(&fs->cache_total) < bluesky_watermark_high_total)
+        return;
 
-    g_slist_free(list);
+    if (bluesky_verbose) {
+        g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
+              "Too much data; invoking flushd: dirty=%d total=%d",
+              g_atomic_int_get(&fs->cache_dirty),
+              g_atomic_int_get(&fs->cache_total));
+    }
+
+    bluesky_flushd_invoke(fs);
+}
+
+/* Start a perpetually-running thread that flushes the cache occasionally. */
+static gpointer flushd_thread(BlueSkyFS *fs)
+{
+    while (TRUE) {
+        bluesky_flushd_invoke(fs);
+        struct timespec delay;
+        delay.tv_sec = 2;
+        delay.tv_nsec = 0;
+        nanosleep(&delay, NULL);
+    }
+
+    return NULL;
+}
+
+void bluesky_flushd_thread_launch(BlueSkyFS *fs)
+{
+    g_thread_create((GThreadFunc)flushd_thread, fs, FALSE, NULL);
 }