2018-11-28 09:08:08

by Pan Bian

Subject: [PATCH V2] amd-xgbe: set skb to NULL after freeing it

The buffer skb is freed via dev_kfree_skb in a loop. After freeing skb,
the value of packet_count is updated via packet_count++. If packet_count
happens to equal the upper bound budget, the loop will be broken and skb
may be assigned to rdata->state.skb. Resulting that rdata->state.skb may
point to a freed memory chunk. To fix this, the patch sets skb to NULL
after dev_kfree_skb(skb).

Signed-off-by: Pan Bian <[email protected]>
---
V2: correct the commit log
---
drivers/net/ethernet/amd/xgbe/xgbe-drv.c | 2 ++
1 file changed, 2 insertions(+)

diff --git a/drivers/net/ethernet/amd/xgbe/xgbe-drv.c b/drivers/net/ethernet/amd/xgbe/xgbe-drv.c
index 0cc911f..ac6b82d 100644
--- a/drivers/net/ethernet/amd/xgbe/xgbe-drv.c
+++ b/drivers/net/ethernet/amd/xgbe/xgbe-drv.c
@@ -2754,6 +2754,7 @@ static int xgbe_rx_poll(struct xgbe_channel *channel, int budget)
netif_err(pdata, rx_err, netdev,
"error in received packet\n");
dev_kfree_skb(skb);
+ skb = NULL;
goto next_packet;
}

@@ -2806,6 +2807,7 @@ static int xgbe_rx_poll(struct xgbe_channel *channel, int budget)
netif_err(pdata, rx_err, netdev,
"packet length exceeds configured MTU\n");
dev_kfree_skb(skb);
+ skb = NULL;
goto next_packet;
}

--
2.7.4

2018-11-28 14:25:18

by Tom Lendacky

Subject: Re: [PATCH V2] amd-xgbe: set skb to NULL after freeing it

On 11/28/2018 03:06 AM, Pan Bian wrote:
> The buffer skb is freed via dev_kfree_skb in a loop. After freeing skb,
> the value of packet_count is updated via packet_count++. If packet_count
> happens to equal the upper bound budget, the loop will be broken and skb
> may be assigned to rdata->state.skb. Resulting that rdata->state.skb may

Except that rdata->state.skb is only assigned if we have received
something and it is (!last || context_next). If error is ever set,
then we read descriptors as long as (!last || context_next). Also,
packet->errors can only be set for the last packet which implies no
following context descriptors. So rdata->state.skb can never be set
with the "bad" skb value.

The only way I can see this occurring is if a context descriptor is
indicated (context_next) along with packet->errors. Are you actually
seeing an error/crash where this is occurring? If so, then the fix
should be to goto read_again if context_next is set because we want
to remove the context descriptor in this case and not save the state.

Thanks,
Tom

> point to a freed memory chunk. To fix this, the patch sets skb to NULL
> after dev_kfree_skb(skb).
>
> Signed-off-by: Pan Bian <[email protected]>
> ---
> V2: correct the commit log
> ---
> drivers/net/ethernet/amd/xgbe/xgbe-drv.c | 2 ++
> 1 file changed, 2 insertions(+)
>
> diff --git a/drivers/net/ethernet/amd/xgbe/xgbe-drv.c b/drivers/net/ethernet/amd/xgbe/xgbe-drv.c
> index 0cc911f..ac6b82d 100644
> --- a/drivers/net/ethernet/amd/xgbe/xgbe-drv.c
> +++ b/drivers/net/ethernet/amd/xgbe/xgbe-drv.c
> @@ -2754,6 +2754,7 @@ static int xgbe_rx_poll(struct xgbe_channel *channel, int budget)
> netif_err(pdata, rx_err, netdev,
> "error in received packet\n");
> dev_kfree_skb(skb);
> + skb = NULL;
> goto next_packet;
> }
>
> @@ -2806,6 +2807,7 @@ static int xgbe_rx_poll(struct xgbe_channel *channel, int budget)
> netif_err(pdata, rx_err, netdev,
> "packet length exceeds configured MTU\n");
> dev_kfree_skb(skb);
> + skb = NULL;
> goto next_packet;
> }
>
>