X-Git-Url: http://git.vrable.net/?a=blobdiff_plain;f=bluesky%2Fcache.c;h=2c1bc92c32a39f78ad639977c916923830caf613;hb=83fd6b61a6e092a22d4d5e59ed95f05f5e287f11;hp=979c9176e33b15f24cfd69a49975a12300f0fa33;hpb=79fd07ce3738aceebc08fd363c14795ae9353057;p=bluesky.git

diff --git a/bluesky/cache.c b/bluesky/cache.c
index 979c917..2c1bc92 100644
--- a/bluesky/cache.c
+++ b/bluesky/cache.c
@@ -13,106 +13,271 @@
 
 #include "bluesky-private.h"
 
-#define WRITEBACK_DELAY (5 * 1000000)
+#define WRITEBACK_DELAY (20 * 1000000)
+#define CACHE_DROP_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)
+static void flushd_dirty_inode(BlueSkyInode *inode)
 {
-    BlueSkyInode *inode = (BlueSkyInode *)i;
+    BlueSkyFS *fs = inode->fs;
 
-    g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
-          "Writeback for inode %"PRIu64" complete", inode->inum);
+    g_mutex_lock(fs->lock);
+    bluesky_list_unlink(&fs->unlogged_list, inode->unlogged_list);
+    inode->unlogged_list = NULL;
+    g_mutex_unlock(fs->lock);
 
-    g_mutex_lock(inode->lock);
+    /* Inode is clean; nothing to do. */
+    if (inode->change_count == inode->change_commit)
+        return;
 
-    inode->change_commit = inode->change_pending;
-    if (inode->change_count == inode->change_commit) {
-        /* If inode is no longer dirty... */
-        inode->change_time = 0;
-        inode->change_pending = 0;
+    if (bluesky_verbose) {
+        g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
+            "Starting flush of inode %"PRIu64, inode->inum);
     }
 
-    g_mutex_unlock(inode->lock);
+    bluesky_inode_start_sync(inode);
 }
 
-static void flushd_inode(gpointer value, gpointer user_data)
+/* Check whether memory usage may have dropped below critical thresholds for
+ * waking up waiting threads. */
+void flushd_check_wakeup(BlueSkyFS *fs)
 {
-    BlueSkyFS *fs = (BlueSkyFS *)user_data;
+    int dirty = g_atomic_int_get(&fs->cache_dirty);
+    dirty += g_atomic_int_get(&fs->cache_log_dirty);
 
-    BlueSkyInode *inode = (BlueSkyInode *)value;
+    if (dirty <= bluesky_watermark_high_dirty)
+        g_cond_broadcast(fs->flushd_cond);
+}
 
-    g_mutex_lock(inode->lock);
+/* 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);
 
-    if (inode->change_count == inode->change_commit) {
-        g_mutex_unlock(inode->lock);
-        bluesky_inode_unref(inode);
-        return;
-    }
+    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);
 
-    if (inode->change_pending) {
-        /* Waiting for an earlier writeback to finish, so don't start a new
-         * writeback yet. */
+        g_mutex_unlock(fs->lock);
+
+        g_mutex_lock(inode->lock);
+        flushd_dirty_inode(inode);
         g_mutex_unlock(inode->lock);
         bluesky_inode_unref(inode);
-        return;
+
+        g_mutex_lock(fs->lock);
+        flushd_check_wakeup(fs);
     }
 
-    uint64_t elapsed = bluesky_get_current_time() - inode->change_time;
-    if (elapsed < WRITEBACK_DELAY) {
-        /* Give a bit more time before starting writeback. */
+    g_cond_broadcast(fs->flushd_cond);
+
+    g_mutex_unlock(fs->lock);
+}
+
+/* Try to flush dirty data to the cloud.
+ * TODO: Rewrite this to work on cloud log items rather than inodes, so we can
+ * better track which logs are fully synchronized to the cloud and can be
+ * garbage collected if needed? */
+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)
+            break;
+        inode = fs->dirty_list.prev->data;
+
+        if (bluesky_verbose) {
+            g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
+                  "Flushing inode %"PRIu64" to cloud", 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_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);
-        return;
+        bluesky_cloudlog_unref(log);
+
+        g_mutex_lock(fs->lock);
     }
 
-    inode->change_pending = inode->change_count;
+    g_mutex_unlock(fs->lock);
+    bluesky_cloudlog_flush(fs);
+}
+
+/* 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 from the cache if needed.  Clean data should generally be
+ * memory-mapped from log file or similar, so the kernel can drop this clean
+ * data from memory for us and hence memory management isn't too important.
+ * Mainly, we'll want to drop references to data that hasn't been accessed in a
+ * while so that it is possible to reclaim log segments on disk. */
+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) {
+        BlueSkyInode *inode;
+        if (fs->accessed_list.prev == NULL)
+            break;
+        inode = fs->accessed_list.prev->data;
+
+        uint64_t elapsed = bluesky_get_current_time() - inode->access_time;
+        if (elapsed < CACHE_DROP_DELAY)
+            break;
 
-    g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
-          "Starting flush of inode %"PRIu64, inode->inum);
+        if (bluesky_verbose) {
+            g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
+                  "Considering dropping cached data for 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_ref(inode);
 
-    bluesky_inode_start_sync(inode, barrier);
+        g_mutex_unlock(fs->lock);
 
-    bluesky_store_async_add_notifier(barrier, writeback_complete, inode);
-    bluesky_store_async_submit(barrier);
-    bluesky_store_async_unref(barrier);
+        g_mutex_lock(inode->lock);
 
-    g_mutex_unlock(inode->lock);
-    bluesky_inode_unref(inode);
+        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);
 }
 
-/* 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)
+/* Run the flush daemon for a single iteration, though if it is already
+ * executing returns immediately. */
+static gpointer flushd_task(BlueSkyFS *fs)
 {
-    GSList **list = (GSList **)user_data;
-    *list = g_slist_prepend(*list, value);
-    bluesky_inode_ref((BlueSkyInode *)value);
+    if (!g_mutex_trylock(fs->flushd_lock))
+        return NULL;
+    flushd_dirty(fs);
+    flushd_cloud(fs);
+    flushd_clean(fs);
+    g_mutex_unlock(fs->flushd_lock);
+
+    return NULL;
 }
 
 void bluesky_flushd_invoke(BlueSkyFS *fs)
 {
-    GSList *list = NULL;
+    g_thread_create((GThreadFunc)flushd_task, fs, FALSE, NULL);
+}
 
-    g_mutex_lock(fs->lock);
-    g_hash_table_foreach(fs->inodes, gather_inodes, &list);
-    g_mutex_unlock(fs->lock);
+void bluesky_flushd_invoke_conditional(BlueSkyFS *fs)
+{
+    if (g_atomic_int_get(&fs->cache_dirty) < bluesky_watermark_medium_dirty)
+        return;
+
+    if (bluesky_verbose) {
+        g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
+              "Too much data; invoking flushd: dirty=%d",
+              g_atomic_int_get(&fs->cache_dirty));
+    }
 
-    list = g_slist_reverse(list);
-    g_slist_foreach(list, flushd_inode, fs);
+    bluesky_flushd_invoke(fs);
 
-    g_slist_free(list);
+    /* If the system is under heavy memory pressure, actually delay execution
+     * so the flush daemon can catch up. */
+    while (g_atomic_int_get(&fs->cache_dirty)
+                + g_atomic_int_get(&fs->cache_log_dirty)
+           > bluesky_watermark_high_dirty) {
+        g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
+              "Waiting due to memory pressure, dirty=%d + %d",
+              g_atomic_int_get(&fs->cache_dirty),
+              g_atomic_int_get(&fs->cache_log_dirty));
+        g_mutex_lock(fs->lock);
+        g_cond_wait(fs->flushd_cond, fs->lock);
+        g_mutex_unlock(fs->lock);
+    }
+}
+
+/* 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);
 }