Btrfs: change how we wait for pending ordered extents

We have a mechanism to make sure we don't lose updates for ordered extents that
were logged in the transaction that is currently running. We add the ordered
extent to a transaction list and then the transaction waits on all the ordered
extents in that list. However are substantially large file systems this list
can be extremely large, and can give us soft lockups, since the ordered extents
don't remove themselves from the list when they do complete.

To fix this we simply add a counter to the transaction that is incremented any
time we have a logged extent that needs to be completed in the current
transaction. Then when the ordered extent finally completes it decrements the
per transaction counter and wakes up the transaction if we are the last ones.
This will eliminate the softlockup. Thanks,

Signed-off-by: Josef Bacik <[e-mail]>
Signed-off-by: Chris Mason <[e-mail]>

fs/btrfs/disk-io.c

···
 	return 0;
 }
 
-static void btrfs_free_pending_ordered(struct btrfs_transaction *cur_trans,
-				       struct btrfs_fs_info *fs_info)
-{
-	struct btrfs_ordered_extent *ordered;
-
-	spin_lock(&fs_info->trans_lock);
-	while (!list_empty(&cur_trans->pending_ordered)) {
-		ordered = list_first_entry(&cur_trans->pending_ordered,
-					   struct btrfs_ordered_extent,
-					   trans_list);
-		list_del_init(&ordered->trans_list);
-		spin_unlock(&fs_info->trans_lock);
-
-		btrfs_put_ordered_extent(ordered);
-		spin_lock(&fs_info->trans_lock);
-	}
-	spin_unlock(&fs_info->trans_lock);
-}
-
 void btrfs_cleanup_one_transaction(struct btrfs_transaction *cur_trans,
 				   struct btrfs_root *root)
 {
···
 	cur_trans->state = TRANS_STATE_UNBLOCKED;
 	wake_up(&root->fs_info->transaction_wait);
 
-	btrfs_free_pending_ordered(cur_trans, root->fs_info);
 	btrfs_destroy_delayed_inodes(root);
 	btrfs_assert_delayed_root_empty(root);
 

fs/btrfs/ordered-data.c

···
 
 	spin_lock_irq(&log->log_extents_lock[index]);
 	while (!list_empty(&log->logged_list[index])) {
+		struct inode *inode;
 		ordered = list_first_entry(&log->logged_list[index],
 					   struct btrfs_ordered_extent,
 					   log_list);
 		list_del_init(&ordered->log_list);
+		inode = ordered->inode;
 		spin_unlock_irq(&log->log_extents_lock[index]);
 
 		if (!test_bit(BTRFS_ORDERED_IO_DONE, &ordered->flags) &&
 		    !test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags)) {
-			struct inode *inode = ordered->inode;
 			u64 start = ordered->file_offset;
 			u64 end = ordered->file_offset + ordered->len - 1;
 
···
 						   &ordered->flags));
 
 		/*
-		 * If our ordered extent completed it means it updated the
-		 * fs/subvol and csum trees already, so no need to make the
-		 * current transaction's commit wait for it, as we end up
-		 * holding memory unnecessarily and delaying the inode's iput
-		 * until the transaction commit (we schedule an iput for the
-		 * inode when the ordered extent's refcount drops to 0), which
-		 * prevents it from being evictable until the transaction
-		 * commits.
+		 * In order to keep us from losing our ordered extent
+		 * information when committing the transaction we have to make
+		 * sure that any logged extents are completed when we go to
+		 * commit the transaction.  To do this we simply increase the
+		 * current transactions pending_ordered counter and decrement it
+		 * when the ordered extent completes.
 		 */
-		if (test_bit(BTRFS_ORDERED_COMPLETE, &ordered->flags))
-			btrfs_put_ordered_extent(ordered);
-		else
-			list_add_tail(&ordered->trans_list, &trans->ordered);
+		if (!test_bit(BTRFS_ORDERED_COMPLETE, &ordered->flags)) {
+			struct btrfs_ordered_inode_tree *tree;
 
+			tree = &BTRFS_I(inode)->ordered_tree;
+			spin_lock_irq(&tree->lock);
+			if (!test_bit(BTRFS_ORDERED_COMPLETE, &ordered->flags)) {
+				set_bit(BTRFS_ORDERED_PENDING, &ordered->flags);
+				atomic_inc(&trans->transaction->pending_ordered);
+			}
+			spin_unlock_irq(&tree->lock);
+		}
+		btrfs_put_ordered_extent(ordered);
 		spin_lock_irq(&log->log_extents_lock[index]);
 	}
 	spin_unlock_irq(&log->log_extents_lock[index]);
···
 	struct btrfs_ordered_inode_tree *tree;
 	struct btrfs_root *root = BTRFS_I(inode)->root;
 	struct rb_node *node;
+	bool dec_pending_ordered = false;
 
 	tree = &BTRFS_I(inode)->ordered_tree;
 	spin_lock_irq(&tree->lock);
···
 	if (tree->last == node)
 		tree->last = NULL;
 	set_bit(BTRFS_ORDERED_COMPLETE, &entry->flags);
+	if (test_and_clear_bit(BTRFS_ORDERED_PENDING, &entry->flags))
+		dec_pending_ordered = true;
 	spin_unlock_irq(&tree->lock);
+
+	/*
+	 * The current running transaction is waiting on us, we need to let it
+	 * know that we're complete and wake it up.
+	 */
+	if (dec_pending_ordered) {
+		struct btrfs_transaction *trans;
+
+		/*
+		 * The checks for trans are just a formality, it should be set,
+		 * but if it isn't we don't want to deref/assert under the spin
+		 * lock, so be nice and check if trans is set, but ASSERT() so
+		 * if it isn't set a developer will notice.
+		 */
+		spin_lock(&root->fs_info->trans_lock);
+		trans = root->fs_info->running_transaction;
+		if (trans)
+			atomic_inc(&trans->use_count);
+		spin_unlock(&root->fs_info->trans_lock);
+
+		ASSERT(trans);
+		if (trans) {
+			if (atomic_dec_and_test(&trans->pending_ordered))
+				wake_up(&trans->pending_wait);
+			btrfs_put_transaction(trans);
+		}
+	}
 
 	spin_lock(&root->ordered_extent_lock);
 	list_del_init(&entry->root_extent_list);

fs/btrfs/ordered-data.h

···
 
 #define BTRFS_ORDERED_LOGGED 10 /* Set when we've waited on this ordered extent
 				 * in the logging code. */
+#define BTRFS_ORDERED_PENDING 11 /* We are waiting for this ordered extent to
+				  * complete in the current transaction. */
 struct btrfs_ordered_extent {
 	/* logical offset in the file */
 	u64 file_offset;

fs/btrfs/transaction.c

···
 	extwriter_counter_init(cur_trans, type);
 	init_waitqueue_head(&cur_trans->writer_wait);
 	init_waitqueue_head(&cur_trans->commit_wait);
+	init_waitqueue_head(&cur_trans->pending_wait);
 	cur_trans->state = TRANS_STATE_RUNNING;
 	/*
 	 * One for this trans handle, one so it will live on until we
···
 	 */
 	atomic_set(&cur_trans->use_count, 2);
 	cur_trans->have_free_bgs = 0;
+	atomic_set(&cur_trans->pending_ordered, 0);
 	cur_trans->start_time = get_seconds();
 	cur_trans->dirty_bg_run = 0;
 
···
 	INIT_LIST_HEAD(&cur_trans->pending_snapshots);
 	INIT_LIST_HEAD(&cur_trans->pending_chunks);
 	INIT_LIST_HEAD(&cur_trans->switch_commits);
-	INIT_LIST_HEAD(&cur_trans->pending_ordered);
 	INIT_LIST_HEAD(&cur_trans->dirty_bgs);
 	INIT_LIST_HEAD(&cur_trans->io_bgs);
 	INIT_LIST_HEAD(&cur_trans->dropped_roots);
···
 	h->can_flush_pending_bgs = true;
 	INIT_LIST_HEAD(&h->qgroup_ref_list);
 	INIT_LIST_HEAD(&h->new_bgs);
-	INIT_LIST_HEAD(&h->ordered);
 
 	smp_mb();
 	if (cur_trans->state >= TRANS_STATE_BLOCKED &&
···
 
 	if (!list_empty(&trans->new_bgs))
 		btrfs_create_pending_block_groups(trans, root);
-
-	if (!list_empty(&trans->ordered)) {
-		spin_lock(&info->trans_lock);
-		list_splice_init(&trans->ordered, &cur_trans->pending_ordered);
-		spin_unlock(&info->trans_lock);
-	}
 
 	trans->delayed_ref_updates = 0;
 	if (!trans->sync) {
···
 }
 
 static inline void
-btrfs_wait_pending_ordered(struct btrfs_transaction *cur_trans,
-			   struct btrfs_fs_info *fs_info)
+btrfs_wait_pending_ordered(struct btrfs_transaction *cur_trans)
 {
-	struct btrfs_ordered_extent *ordered;
-
-	spin_lock(&fs_info->trans_lock);
-	while (!list_empty(&cur_trans->pending_ordered)) {
-		ordered = list_first_entry(&cur_trans->pending_ordered,
-					   struct btrfs_ordered_extent,
-					   trans_list);
-		list_del_init(&ordered->trans_list);
-		spin_unlock(&fs_info->trans_lock);
-
-		wait_event(ordered->wait, test_bit(BTRFS_ORDERED_COMPLETE,
-						   &ordered->flags));
-		btrfs_put_ordered_extent(ordered);
-		spin_lock(&fs_info->trans_lock);
-	}
-	spin_unlock(&fs_info->trans_lock);
+	wait_event(cur_trans->pending_wait,
+		   atomic_read(&cur_trans->pending_ordered) == 0);
 }
 
 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
···
 	}
 
 	spin_lock(&root->fs_info->trans_lock);
-	list_splice_init(&trans->ordered, &cur_trans->pending_ordered);
 	if (cur_trans->state >= TRANS_STATE_COMMIT_START) {
 		spin_unlock(&root->fs_info->trans_lock);
 		atomic_inc(&cur_trans->use_count);
···
 
 	btrfs_wait_delalloc_flush(root->fs_info);
 
-	btrfs_wait_pending_ordered(cur_trans, root->fs_info);
+	btrfs_wait_pending_ordered(cur_trans);
 
 	btrfs_scrub_pause(root);
 	/*

fs/btrfs/transaction.h

···
 	 */
 	atomic_t num_writers;
 	atomic_t use_count;
+	atomic_t pending_ordered;
 
 	/*
 	 * true if there is free bgs operations in this transaction
···
 	unsigned long start_time;
 	wait_queue_head_t writer_wait;
 	wait_queue_head_t commit_wait;
+	wait_queue_head_t pending_wait;
 	struct list_head pending_snapshots;
 	struct list_head pending_chunks;
-	struct list_head pending_ordered;
 	struct list_head switch_commits;
 	struct list_head dirty_bgs;
 	struct list_head io_bgs;
···
 	 */
 	struct btrfs_root *root;
 	struct seq_list delayed_ref_elem;
-	struct list_head ordered;
 	struct list_head qgroup_ref_list;
 	struct list_head new_bgs;
 };