+
+/* 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);
+ bluesky_cloudlog_unref(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. */
+static void drop_caches(BlueSkyInode *inode)
+{
+ if (inode->type == BLUESKY_REGULAR)
+ bluesky_file_drop_cached(inode);
+
+ BlueSkyCloudLog *log = inode->committed_item;
+ if (log != NULL) {
+ g_mutex_lock(log->lock);
+ if (log->data != NULL
+ && g_atomic_int_get(&log->data_lock_count) == 0
+ && (log->location_flags != 0))
+ {
+ bluesky_cloudlog_stats_update(log, -1);
+ bluesky_string_unref(log->data);
+ log->data = NULL;
+ bluesky_cloudlog_stats_update(log, 1);
+ }
+ g_mutex_unlock(log->lock);
+ }
+}
+
+/* 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;
+
+ 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_cloud(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_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));
+ }
+
+ bluesky_flushd_invoke(fs);
+
+ /* 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);
+}