Re: [PATCH v3 1/2] binder: Refactor binder_node print synchronization

[Tiffany Yang <ynaffit@google.com>]

Alice Ryhl <aliceryhl@google.com> writes:

> On Wed, May 07, 2025 at 09:10:05PM +0000, Tiffany Y. Yang wrote:
>> +/**
>> + * print_next_binder_node_ilocked() - Print binder_node from a locked list
>> + * @m:          struct seq_file for output via seq_printf()
>> + * @node:       struct binder_node to print fields of
>> + * @prev_node:	struct binder_node we hold a temporary reference to (if any)
>> + *
>> + * Helper function to handle synchronization around printing a struct
>> + * binder_node while iterating through @node->proc->nodes or the dead nodes
>> + * list. Caller must hold either @node->proc->inner_lock (for live nodes) or
>> + * binder_dead_nodes_lock. This lock will be released during the body of this
>> + * function, but it will be reacquired before returning to the caller.
>> + *
>> + * Return:	pointer to the struct binder_node we hold a tmpref on
>> + */
>> +static struct binder_node *
>> +print_next_binder_node_ilocked(struct seq_file *m, struct binder_node *node,
>> +			       struct binder_node *prev_node)
>> +{
>> +	/*
>> +	 * Take a temporary reference on the node so that isn't removed from
>> +	 * its proc's tree or the dead nodes list while we print it.
>> +	 */
>> +	binder_inc_node_tmpref_ilocked(node);
>> +	/*
>> +	 * Live nodes need to drop the inner proc lock and dead nodes need to
>> +	 * drop the binder_dead_nodes_lock before trying to take the node lock.
>> +	 */
>> +	if (node->proc)
>> +		binder_inner_proc_unlock(node->proc);
>> +	else
>> +		spin_unlock(&binder_dead_nodes_lock);
>> +	if (prev_node)
>> +		binder_put_node(prev_node);
>
> I don't buy this logic. Imagine the following scenario:
>
> 1. print_binder_proc is called, and we loop over proc->nodes.
> 2. We call binder_inner_proc_unlock(node->proc).
> 3. On another thread, binder_deferred_release() is called.
> 4. The node is removed from proc->nodes and node->proc is set to NULL.
> 5. Back in print_next_binder_node_ilocked(), we now call
>    spin_lock(&binder_dead_nodes_lock) and return.
> 6. In print_binder_proc(), we think that we hold the proc lock, but
>    actually we hold the dead nodes lock instead. BOOM.
>
> What happens with the current code is that print_binder_proc() takes the
> proc lock again after the node was removed from proc->nodes, and then it
> exits the loop because rb_next(n) returns NULL when called on a node not
> in any rb-tree.
>
> Alice


Thanks for catching this!! I think this race could be solved by passing
"proc" in as a parameter (NULL if iterating over the dead_nodes_list),
and locking/unlocking based on that instead of node->proc. WDYT?

-- 
Tiffany Y. Yang