On Wed, Aug 14, 2019 at 12:41:33PM -0500, Wenwen Wang wrote:
> In usbnet_start_xmit(), 'urb->sg' is allocated through kmalloc_array() by
> invoking build_dma_sg(). Later on, if 'CONFIG_PM' is defined and the if
> branch is taken, the execution will go to the label 'deferred'. However,
> 'urb->sg' is not deallocated on this execution path, leading to a memory
> leak bug.
> 
> Signed-off-by: Wenwen Wang <wen...@cs.uga.edu>
> ---
>  drivers/net/usb/usbnet.c | 1 +
>  1 file changed, 1 insertion(+)
> 
> diff --git a/drivers/net/usb/usbnet.c b/drivers/net/usb/usbnet.c
> index 72514c4..f17fafa 100644
> --- a/drivers/net/usb/usbnet.c
> +++ b/drivers/net/usb/usbnet.c
> @@ -1433,6 +1433,7 @@ netdev_tx_t usbnet_start_xmit (struct sk_buff *skb,
>               usb_anchor_urb(urb, &dev->deferred);
>               /* no use to process more packets */
>               netif_stop_queue(net);
> +             kfree(urb->sg);
>               usb_put_urb(urb);

The URB itself is not getting freed here; it is merely added to the
anchor list and will be submitted later upon usbnet_resume(). Therefore
freeing the SG list is premature and incorrect, as it will get freed in
either the tx_complete/tx_done path or upon URB submission failure.

>               spin_unlock_irqrestore(&dev->txq.lock, flags);
>               netdev_dbg(dev->net, "Delaying transmission for resumption\n");

Jack
-- 
The Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
a Linux Foundation Collaborative Project

Reply via email to