Re: [PATCH] hippi: fix possible buffer overflow caused by bad DMA value in rr_start_xmit()
From: Eric Dumazet <edumazet@google.com>
Date: 2024-06-25 08:13:39
Also in:
lkml
On Tue, Jun 25, 2024 at 9:52 AM Huai-Yuan Liu [off-list ref] wrote:
The value rrpriv->info->tx_ctrl is stored in DMA memory, and assigned to txctrl. However, txctrl->pi can be modified by malicious hardware at any time. Because txctrl->pi is assigned to index, buffer overflow may occur when the code "rrpriv->tx_skbuff[index]" is executed.
How txctrl->pi can be modified by malicious hardware ? This is host memory, not mapped to the device.
To address this issue, ensure index is checked.
Fixes: f33a7251c825 ("hippi: switch from 'pci_' to 'dma_' API")Is it just me or this Fixes: tag is not relevant ?
quoted hunk ↗ jump to hunk
Signed-off-by: Huai-Yuan Liu <redacted> --- V2: * In patch V2, we remove the first condition in if statement and use netdev_err() instead of printk(). Thanks Paolo Abeni for helpful advice. V3: * In patch V3, we stop the queue before return BUSY. Thanks to Jakub Kicinski for his advice. V4: * In patch V4, we revise the wording in the description. Thanks to Markus Elfring for pointing this out. --- drivers/net/hippi/rrunner.c | 6 ++++++ 1 file changed, 6 insertions(+)diff --git a/drivers/net/hippi/rrunner.c b/drivers/net/hippi/rrunner.c index aa8f828a0ae7..f4da342dd5cc 100644 --- a/drivers/net/hippi/rrunner.c +++ b/drivers/net/hippi/rrunner.c@@ -1440,6 +1440,12 @@ static netdev_tx_t rr_start_xmit(struct sk_buff *skb, txctrl = &rrpriv->info->tx_ctrl; index = txctrl->pi; + if (index >= TX_RING_ENTRIES) { + netdev_err(dev, "invalid index value %02x\n", index); + netif_stop_queue(dev); + spin_unlock_irqrestore(&rrpriv->lock, flags); + return NETDEV_TX_BUSY; + }
When seeing this patch, my initial reaction was : array_index_nospec() is missing. But then : If a malicious hardware can control txctl->pi, then it will cause multiple UAF or skb leaks if txctrl->pi is kept below TX_RING_ENTRIES. This driver should keep a trusted copy of txctrl->pi field, only seen by itself, and not read it from the untrusted device. Otherwise, this patch is only adding a lot of confusion.