Re: [PATCH] xhci: fix null pointer deref for xhci_urb_enqueue

[Greg KH <gregkh@linuxfoundation.org>]

On Fri, Nov 17, 2023 at 03:21:28PM +0800, Kuen-Han Tsai wrote:
> The null pointer dereference happens when xhci_free_dev() frees the
> xhci->devs[slot_id] virtual device while xhci_urb_enqueue() is
> processing a urb and checking the max packet size.
> 
> [106913.850735][ T2068] usb 2-1: USB disconnect, device number 2
> [106913.856999][ T4618] Unable to handle kernel NULL pointer dereference at virtual address 0000000000000010
> [106913.857488][ T4618] Call trace:
> [106913.857491][ T4618]  xhci_check_maxpacket+0x30/0x2dc
> [106913.857494][ T4618]  xhci_urb_enqueue+0x24c/0x47c
> [106913.857498][ T4618]  usb_hcd_submit_urb+0x1f4/0xf34
> [106913.857501][ T4618]  usb_submit_urb+0x4b8/0x4fc
> [106913.857503][ T4618]  usb_control_msg+0x144/0x238
> [106913.857507][ T4618]  do_proc_control+0x1f0/0x5bc
> [106913.857509][ T4618]  usbdev_ioctl+0xdd8/0x15a8
> 
> This patch adds a spinlock to the xhci_urb_enqueue function to make sure
> xhci_free_dev() and xhci_urb_enqueue() do not race and cause null
> pointer dereference.

I thought we had a lock for this already, what changed to cause this to
start triggering now, all these years later?

> 
> Signed-off-by: Kuen-Han Tsai <khtsai@google.com>

What commit id does this fix?


> ---
>  drivers/usb/host/xhci.c | 38 ++++++++++++++++++++++++--------------
>  1 file changed, 24 insertions(+), 14 deletions(-)
> 
> diff --git a/drivers/usb/host/xhci.c b/drivers/usb/host/xhci.c
> index 884b0898d9c9..e0766ebeff0e 100644
> --- a/drivers/usb/host/xhci.c
> +++ b/drivers/usb/host/xhci.c
> @@ -1522,23 +1522,32 @@ static int xhci_urb_enqueue(struct usb_hcd *hcd, struct urb *urb, gfp_t mem_flag
>  	struct urb_priv	*urb_priv;
>  	int num_tds;
>  
> -	if (!urb)
> -		return -EINVAL;
> -	ret = xhci_check_args(hcd, urb->dev, urb->ep,
> -					true, true, __func__);
> -	if (ret <= 0)
> -		return ret ? ret : -EINVAL;
> +	spin_lock_irqsave(&xhci->lock, flags);
> +
> +	if (!urb) {
> +		ret = -EINVAL;
> +		goto done;
> +	}

Why does this have to be inside the lock?  The urb can't change here,
can it?

> +
> +	ret = xhci_check_args(hcd, urb->dev, urb->ep, true, true, __func__);
> +	if (ret <= 0) {
> +		ret = ret ? ret : -EINVAL;
> +		goto done;
> +	}
>  
>  	slot_id = urb->dev->slot_id;
>  	ep_index = xhci_get_endpoint_index(&urb->ep->desc);
>  	ep_state = &xhci->devs[slot_id]->eps[ep_index].ep_state;
>  
> -	if (!HCD_HW_ACCESSIBLE(hcd))
> -		return -ESHUTDOWN;
> +	if (!HCD_HW_ACCESSIBLE(hcd)) {
> +		ret = -ESHUTDOWN;
> +		goto done;

Note, we now have completions, so all of this "goto done" doesn't need
to happen anymore.  Not a complaint, just a suggestion for future
changes or this one, your choice.

thanks,

greg k-h