Skip to contentSkip to conversation

btrfs: don't add delayed refs to an aborted transaction

PATCH In review Adarsh Das 2 participants 2 messages Mar 1 Btrfs file system

lore
Share & tools
lore

Adarsh Das

author
#0 lore raw

PATCH 1/1 btrfs: don't add delayed refs to an aborted transaction

When a transaction aborts, cleanup_transaction() calls btrfs_cleanup_one_transaction() which drains all pending delayed refs via btrfs_destroy_delayed_refs().

But, btrfs_cleanup_one_transaction() then wakes up tasks waiting
on transaction_blocked_wait and sets the transaction state to TRANS_STATE_UNBLOCKED. These woken tasks can then call btrfs_add_delayed_tree_ref(), btrfs_add_delayed_data_ref(), or btrfs_add_delayed_extent_op() on the already-aborted transaction, inserting new entries into the head_refs xarray after it was just drained.

When btrfs_put_transaction() subsequently drops the refcount to zero, it hits:

  WARN_ON(!xa_empty(&transaction->delayed_refs.head_refs));

This patch fixes this by checking TRANS_ABORTED() at the start of add_delayed_ref() and btrfs_add_delayed_extent_op() before inserting into the xarray. btrfs_abort_transaction() is called at the start of cleanup_transaction(), before btrfs_destroy_delayed_refs(), so the aborted flag should always be set before any wakeups occur.

---
 fs/btrfs/delayed-ref.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)
fs/btrfs/delayed-ref.c +9 -1
index 3766ff29fbbb..b994f9702c32 100644
@@ -327,7 +327,7 @@ static int cmp_refs_node(const struct rb_node *new, const struct rb_node *exist)
 	return comp_refs(new_node, exist_node, true);
 }
 
-static struct btrfs_delayed_ref_node* tree_insert(struct rb_root_cached *root,
+static struct btrfs_delayed_ref_node *tree_insert(struct rb_root_cached *root,
 		struct btrfs_delayed_ref_node *ins)
 {
 	struct rb_node *node = &ins->ref_node;
@@ -1025,6 +1025,10 @@ static int add_delayed_ref(struct btrfs_trans_handle *trans,
 	}
 
 	delayed_refs = &trans->transaction->delayed_refs;
+	if (TRANS_ABORTED(trans->transaction)) {
+		ret = -EIO;
+		goto free_head_ref;
+	}
 
 	if (btrfs_qgroup_full_accounting(fs_info) && !generic_ref->skip_qgroup) {
 		record = kzalloc_obj(*record, GFP_NOFS);
@@ -1153,6 +1157,10 @@ int btrfs_add_delayed_extent_op(struct btrfs_trans_handle *trans,
 	head_ref->extent_op = extent_op;
 
 	delayed_refs = &trans->transaction->delayed_refs;
+	if (TRANS_ABORTED(trans->transaction)) {
+		kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
+		return -EIO;
+	}
 
 	ret = xa_reserve(&delayed_refs->head_refs, index, GFP_NOFS);
 	if (ret) {
Signature · 1 line
2.53.0
#1 lore raw
Quoted 8 lines · Adarsh Das

When a transaction aborts, cleanup_transaction() calls btrfs_cleanup_one_transaction() which drains all pending delayed refs via btrfs_destroy_delayed_refs().

But, btrfs_cleanup_one_transaction() then wakes up tasks waiting
on transaction_blocked_wait and sets the transaction state to TRANS_STATE_UNBLOCKED. These woken tasks can then call btrfs_add_delayed_tree_ref(),

No, that's not correct.

You need to take into account transaction states.

Once a transaction is in the state TRANS_STATE_COMMIT_START, it means no one else is holding a transaction handle for that transaction (except the task calling btrfs_commit_transaction()).
Also no one else is able to start a new transaction until the current transaction state is >= TRANS_STATE_UNBLOCKED.
Take a look at the array btrfs_blocked_trans_types, join_transaction() and start_transaction().

We call cleanup_transaction() only in btrfs_commit_transaction(), if some error happens, after the state is set to TRANS_STATE_COMMIT_START.
Once we call cleanup_transaction() we call btrfs_abort_transaction() and mark the filesystem in error state, setting fs_info->fs_error among other things.
So after btrfs_cleanup_one_transaction() goes through the remaining transaction states and does the wakeups, no one can start a new transaction - join_transaction() sees that fs_info->fs_error is not 0 (through BTRFS_FS_ERROR()) and fails.

Quoted 3 lines

btrfs_add_delayed_data_ref(), or btrfs_add_delayed_extent_op() on the already-aborted transaction, inserting new entries into the head_refs xarray after it was just drained.

Nop... if the transaction was aborted by cleanup_transaction(), it means that, except for the task calling btrfs_commit_transaction(), no one is holding a handle for that transaction.

Think about how chaotic it would be if every piece of code holding a transaction handle had to check if the transaction was aborted before doing anything...
We would have not only to avoid adding delayed refs but pretty much anything.

Thanks.

Quoted 57 lines

When btrfs_put_transaction() subsequently drops the refcount to zero, it hits:

  WARN_ON(!xa_empty(&transaction->delayed_refs.head_refs));

This patch fixes this by checking TRANS_ABORTED() at the start of add_delayed_ref() and btrfs_add_delayed_extent_op() before inserting into the xarray. btrfs_abort_transaction() is called at the start of cleanup_transaction(), before btrfs_destroy_delayed_refs(), so the aborted flag should always be set before any wakeups occur.

---
 fs/btrfs/delayed-ref.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)
fs/btrfs/delayed-ref.c +9 -2
index 3766ff29fbbb..b994f9702c32 100644
@@ -327,7 +327,7 @@ static int cmp_refs_node(const struct rb_node *new, const struct rb_node *exist)
        return comp_refs(new_node, exist_node, true);
 }
 
-static struct btrfs_delayed_ref_node* tree_insert(struct rb_root_cached *root,
+static struct btrfs_delayed_ref_node *tree_insert(struct rb_root_cached *root,
                struct btrfs_delayed_ref_node *ins)
 {
        struct rb_node *node = &ins->ref_node;
@@ -1025,6 +1025,10 @@ static int add_delayed_ref(struct btrfs_trans_handle *trans,
        }
 
        delayed_refs = &trans->transaction->delayed_refs;
+       if (TRANS_ABORTED(trans->transaction)) {
+               ret = -EIO;
+               goto free_head_ref;
+       }
 
        if (btrfs_qgroup_full_accounting(fs_info) && !generic_ref->skip_qgroup) {
                record = kzalloc_obj(*record, GFP_NOFS);
@@ -1153,6 +1157,10 @@ int btrfs_add_delayed_extent_op(struct btrfs_trans_handle *trans,
        head_ref->extent_op = extent_op;
 
        delayed_refs = &trans->transaction->delayed_refs;
+       if (TRANS_ABORTED(trans->transaction)) {
+               kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
+               return -EIO;
+       }
 
        ret = xa_reserve(&delayed_refs->head_refs, index, GFP_NOFS);
        if (ret) {
--

2.53.0