Re: [PATCH] netdevsim: fix use-after-free in __nsim_dev_port_del

[Simon Horman]

This is an AI-generated review of your patch. The human sending this
email has considered the AI review valid, or at least plausible.
Full review at: https://netdev-ai.bots.linux.dev/sashiko/
---
Subject: netdevsim: fix use-after-free in __nsim_dev_port_del

The patch removes the per-port debugfs directory in nsim_destroy()
before free_netdev() so that debugfs files referencing fields inside
the netdevsim struct cannot be reached after the slab is released.

Link: https://syzkaller.appspot.com/bug?extid=6c25f4750230faf70be9

> diff --git a/drivers/net/netdevsim/netdev.c b/drivers/net/netdevsim/netdev.c
> index 27e5f109f933..08136e7990cb 100644
> --- a/drivers/net/netdevsim/netdev.c
> +++ b/drivers/net/netdevsim/netdev.c
> @@ -1214,6 +1214,13 @@ void nsim_destroy(struct netdevsim *ns)
>  		ns->page = NULL;
>  	}
>  
> +	/*
> +	 * Remove per-port debugfs files before free_netdev() releases the
> +	 * netdevsim struct to prevent use-after-free in concurrent readers.
> +	 */
> +	debugfs_remove_recursive(ns->nsim_dev_port->ddir);
> +	ns->nsim_dev_port->ddir = NULL;
> +
>  	free_netdev(dev);
>  }

[High]
Does the same use-after-free window remain on the nsim_create() error
path?

nsim_create() unconditionally calls nsim_ethtool_init(ns) before the
init step that can fail:

    SET_NETDEV_DEVLINK_PORT(dev, &nsim_dev_port->devlink_port);
    nsim_ethtool_init(ns);
    if (nsim_dev_port_is_pf(nsim_dev_port))
        err = nsim_init_netdevsim(ns);
    else
        err = nsim_init_netdevsim_vf(ns);
    if (err)
        goto err_free_netdev;
    ...
err_free_netdev:
    free_netdev(dev);
    return ERR_PTR(err);

nsim_ethtool_init() populates files under nsim_dev_port->ddir whose
data pointers reference fields inside ns, for example:

    ethtool = debugfs_create_dir("ethtool", ns->nsim_dev_port->ddir);
    debugfs_create_u32("get_err", 0600, ethtool, &ns->ethtool.get_err);
    debugfs_create_u32("set_err", 0600, ethtool, &ns->ethtool.set_err);
    ...
    dir = debugfs_create_dir("ring", ethtool);
    debugfs_create_u32("rx_max_pending", 0600, dir,
                      &ns->ethtool.ring.rx_max_pending);

If nsim_init_netdevsim() (or nsim_init_netdevsim_vf()) fails, the
err_free_netdev label runs free_netdev(dev) directly and returns
ERR_PTR(err). The caller then calls nsim_dev_port_debugfs_exit() to
tear down ddir, but in the meantime the netdevsim slab has already
been freed while these debugfs files are still live with data pointers
into it.

Should the same debugfs_remove_recursive(nsim_dev_port->ddir) be done
on the err_free_netdev path before free_netdev(dev), or should the
debugfs file creation be moved past the point of no return so the
fuzzer-reachable failure paths (alloc pressure on mock_phc_create,
register_netdevice, nsim_psp_init, etc.) cannot leave dangling
debugfs entries?

The Fixes: tag e05b2d141fef appears to cover this path as well.