On 2019/9/4 18:19, Eric Dumazet wrote: > > > On 9/4/19 11:42 AM, Mao Wenan wrote: >> When dma_map_single is failed to map buffer, skb can't be freed >> before sonic driver return to stack with NETDEV_TX_BUSY, because >> this skb may be requeued to qdisc, it might trigger use-after-free. >> >> Fixes: d9fb9f384292 ("*sonic/natsemi/ns83829: Move the National Semi-conductor drivers") >> Signed-off-by: Mao Wenan <maowenan@xxxxxxxxxx> >> --- >> drivers/net/ethernet/natsemi/sonic.c | 1 - >> 1 file changed, 1 deletion(-) >> >> diff --git a/drivers/net/ethernet/natsemi/sonic.c b/drivers/net/ethernet/natsemi/sonic.c >> index d0a01e8f000a..248a8f22a33b 100644 >> --- a/drivers/net/ethernet/natsemi/sonic.c >> +++ b/drivers/net/ethernet/natsemi/sonic.c >> @@ -233,7 +233,6 @@ static int sonic_send_packet(struct sk_buff *skb, struct net_device *dev) >> laddr = dma_map_single(lp->device, skb->data, length, DMA_TO_DEVICE); >> if (!laddr) { >> printk(KERN_ERR "%s: failed to map tx DMA buffer.\n", dev->name); >> - dev_kfree_skb(skb); >> return NETDEV_TX_BUSY; >> } >> >> > > That is the wrong way to fix this bug. > > What guarantee do we have that the mapping operation will succeed next time we attempt > the transmit (and the dma_map_single() operation) ? > > NETDEV_TX_BUSY is very dangerous, this might trigger an infinite loop. > > I would rather leave the dev_kfree_skb(skb), and return NETDEV_TX_OK yes, right, it will go to infinite loop if dma_map_single failed always. v2 will be sent later. > > Also the printk(KERN_ERR ...) should be replaced by pr_err_ratelimited(...) > > NETDEV_TX_BUSY really should only be used by drivers that call netif_tx_stop_queue() > at the wrong moment. It will call netif_tx_stop_queue() then return NETDEV_TX_BUSY, and give a chance to netif_tx_wake_queue(), then stack can be resent packet to driver? > > . >