2015-06-05 13:46:36

by Jonas Jensen

Subject: [PATCH] net: moxa: fix TX overrun memory leak

moxart_mac_start_xmit() doesn't care where tx_tail is, tx_head can
catch and pass tx_tail, which is bad because moxart_tx_finished()
isn't guaranteed to catch up on freeing resources from tx_tail.

Add a check in moxart_mac_start_xmit(), exit to make sure tx_head
never passes tx_tail.

Move tx_head and place tx_tail within spinlock, avoiding read-write
race that could happen when interrupts hit at the wrong moment.

Addresses https://bugzilla.kernel.org/show_bug.cgi?id=99451

Signed-off-by: Jonas Jensen <[email protected]>
---

Notes:
Applies to next-20150604

drivers/net/ethernet/moxa/moxart_ether.c | 14 ++++++++++++--
1 file changed, 12 insertions(+), 2 deletions(-)

diff --git a/drivers/net/ethernet/moxa/moxart_ether.c b/drivers/net/ethernet/moxa/moxart_ether.c
index becbb5f..e0af86c 100644
--- a/drivers/net/ethernet/moxa/moxart_ether.c
+++ b/drivers/net/ethernet/moxa/moxart_ether.c
@@ -25,6 +25,7 @@
#include <linux/of_irq.h>
#include <linux/crc32.h>
#include <linux/crc32c.h>
+#include <linux/circ_buf.h>

#include "moxart_ether.h"

@@ -312,13 +313,22 @@ static int moxart_mac_start_xmit(struct sk_buff *skb, struct net_device *ndev)
struct moxart_mac_priv_t *priv = netdev_priv(ndev);
void __iomem *desc;
unsigned int len;
- unsigned int tx_head = priv->tx_head;
+ unsigned int tx_head, tx_tail;
u32 txdes1;
int ret = NETDEV_TX_BUSY;

+ spin_lock_irq(&priv->txlock);
+
+ tx_head = priv->tx_head;
+ tx_tail = priv->tx_tail;
+
desc = priv->tx_desc_base + (TX_REG_DESC_SIZE * tx_head);

- spin_lock_irq(&priv->txlock);
+ if (CIRC_SPACE(tx_head, tx_tail, TX_DESC_NUM) == 0) {
+ priv->stats.tx_dropped++;
+ goto out_unlock;
+ }
+
if (readl(desc + TX_REG_OFFSET_DESC0) & TX_DESC0_DMA_OWN) {
net_dbg_ratelimited("no TX space for packet\n");
priv->stats.tx_dropped++;
--
1.8.2.1

2015-06-08 02:22:10

by David Miller

Subject: Re: [PATCH] net: moxa: fix TX overrun memory leak

From: Jonas Jensen <[email protected]>
Date: Fri, 5 Jun 2015 15:46:18 +0200

> desc = priv->tx_desc_base + (TX_REG_DESC_SIZE * tx_head);
>
> - spin_lock_irq(&priv->txlock);
> + if (CIRC_SPACE(tx_head, tx_tail, TX_DESC_NUM) == 0) {
> + priv->stats.tx_dropped++;
> + goto out_unlock;
> + }

You can't do this.

Whatever code creates this condition _must_ stop the TX queue
so that the core netdevice transmit layer does not call your
transmit method.