[PATCH] driver: tg3: fix potential UAF in tigon3_dma_hwbug_workaround()

[PATCH] driver: tg3: fix potential UAF in tigon3_dma_hwbug_workaround()

[Gen Zhang]

In tigon3_dma_hwbug_workaround(), pskb is first stored in skb. And this
function is to store new_skb into pskb at the end. However, in the error
paths when new_skb is freed by dev_kfree_skb_any(), stroing new_skb to pskb
should be prevented.

And freeing skb with dev_consume_skb_any() should be executed after storing
new_skb to pskb, because freeing skb will free pskb (alias).

Signed-off-by: Gen Zhang <blackgod016574@gmail.com>
---
diff --git a/drivers/net/ethernet/broadcom/tg3.c b/drivers/net/ethernet/broadcom/tg3.c
index ca3aa12..dbfac26 100644
--- a/drivers/net/ethernet/broadcom/tg3.c
+++ b/drivers/net/ethernet/broadcom/tg3.c
@@ -7826,6 +7826,7 @@ static int tigon3_dma_hwbug_workaround(struct tg3_napi *tnapi,
 
 	if (!new_skb) {
 		ret = -1;
+		goto new_skb_err;
 	} else {
 		/* New SKB is guaranteed to be linear. */
 		new_addr = pci_map_single(tp->pdev, new_skb->data, new_skb->len,
@@ -7834,6 +7835,7 @@ static int tigon3_dma_hwbug_workaround(struct tg3_napi *tnapi,
 		if (pci_dma_mapping_error(tp->pdev, new_addr)) {
 			dev_kfree_skb_any(new_skb);
 			ret = -1;
+			goto new_skb_err;
 		} else {
 			u32 save_entry = *entry;
 
@@ -7849,12 +7851,14 @@ static int tigon3_dma_hwbug_workaround(struct tg3_napi *tnapi,
 				tg3_tx_skb_unmap(tnapi, save_entry, -1);
 				dev_kfree_skb_any(new_skb);
 				ret = -1;
+				goto new_skb_err;
 			}
 		}
 	}
 
-	dev_consume_skb_any(skb);
 	*pskb = new_skb;
+	dev_consume_skb_any(skb);
+new_skb_err:
 	return ret;
 }
 

Re: [PATCH] driver: tg3: fix potential UAF in tigon3_dma_hwbug_workaround()

[David Miller]

From: Gen Zhang <blackgod016574@gmail.com>
Date: Thu, 16 Jan 2020 11:30:44 +0800

> In tigon3_dma_hwbug_workaround(), pskb is first stored in skb. And this
> function is to store new_skb into pskb at the end. However, in the error
> paths when new_skb is freed by dev_kfree_skb_any(), stroing new_skb to pskb
> should be prevented.
> 
> And freeing skb with dev_consume_skb_any() should be executed after storing
> new_skb to pskb, because freeing skb will free pskb (alias).
> 
> Signed-off-by: Gen Zhang <blackgod016574@gmail.com>

There are no bugs here.

The caller never references "*pskb" when an error is returned.  So it is
safe to store any value whatsoever into that pointer.

'skb' never changes it's value even if we store something into *pskb
because we've loaded it into a local variable.  So it is always safe to
call dev_consume_skb_any() on 'skb' in any order with respect to that
assignment.

I'm not applying this until you can show a real bug resulting from
the current code, and if so you'll need to add that explanation to
your commit message.

Thanks.