X-Git-Url: http://git.vrable.net/?a=blobdiff_plain;f=bluesky%2Fcache.c;h=00d0fec1d13ebd5a64eab26ca3dfd88cbb6d2341;hb=0e76979181d1e7b7bbaf24e7b196b58cea5d7879;hp=f953a6455a0736ee6cac142038a270404cbb25f7;hpb=0ef7fc934daee6ded318e3d52023521b758b295e;p=bluesky.git

diff --git a/bluesky/cache.c b/bluesky/cache.c
index f953a64..00d0fec 100644
--- a/bluesky/cache.c
+++ b/bluesky/cache.c
@@ -13,35 +13,304 @@
 
 #include "bluesky-private.h"
 
+#define WRITEBACK_DELAY (20 * 1000000)
+
 /* Filesystem caching and cache coherency. */
 
-static void flushd_inode(gpointer key, gpointer value, gpointer user_data)
+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;
 
-    if (inode->change_count == inode->change_commit)
+    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);
 
-    if (inode->type == BLUESKY_REGULAR)
-        bluesky_file_flush(inode);
-    bluesky_inode_flush(fs, inode);
+    /* 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
 
-/* 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. */
-void bluesky_flushd_invoke(BlueSkyFS *fs)
+static void flushd_dirty_inode(BlueSkyInode *inode)
 {
-    g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
-          "Writeout process invoked");
+    BlueSkyFS *fs = inode->fs;
 
     g_mutex_lock(fs->lock);
-    g_hash_table_foreach(fs->inodes, flushd_inode, fs);
+    bluesky_list_unlink(&fs->dirty_list, inode->dirty_list);
+    inode->dirty_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);
+}
+
+/* Try to flush dirty data to disk, either due to memory pressure or due to
+ * timeouts. */
+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->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);
+}
+
+/* Drop cached data for a given inode, if it is clean.  inode must be locked. */
+static void drop_caches(BlueSkyInode *inode)
+{
+    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);
+
+    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);
+    flushd_clean(fs);
+    g_mutex_unlock(fs->flushd_lock);
+
+    return NULL;
+}
+
+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;
+
+    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);
 }