From: Birger Koblitz <mail@birger-koblitz.de>
To: netdev-bot+sashiko@kernel.org
Cc: andrew+netdev@lunn.ch, davem@davemloft.net, edumazet@google.com,
kuba@kernel.org, pabeni@redhat.com, linux-usb@vger.kernel.org,
netdev@vger.kernel.org, linux-kernel@vger.kernel.org
Subject: Re: [PATCH net-next] aqc111: Correct endianness handling in rx/tx_fixup
Date: Sun, 22 Nov 2026 11:55:57 +0100 [thread overview]
Message-ID: <a6cac285-5df6-4440-bd58-99e29c71b0df@birger-koblitz.de> (raw)
In-Reply-To: <179515614261.3136.9007298128427052621@kernel.org>
On 20/11/2026 7:29 am, netdev-bot+sashiko@kernel.org wrote:
> Thank you for your contribution! Sashiko AI review found 3 potential
> issue(s) to consider.
>
> Critical: 0 · High: 0 (1) · Medium: 1 · Low: 2
>
> If you believe none of the issues are legitimate, please reply to each
> issue of Medium severity or higher and explain why the AI is wrong.
> If any of them are legitimate and you plan to respin, please let patchwork
> know by including "pw-bot: cr" as a separate line at the end of your reply
> (one such reply per series is enough).
>
> - [Medium] The new `skb_cloned(skb) ||` term in aqc111_tx_fixup()
> (drivers/net/usb/aqc111.c:1223) sends every cloned TX skb to…
> - [Low] The commit message misdescribes the change.
> - [Low] The patch fixes real runtime bugs but has no Fixes: tag and no
> stable Cc, and its message reads like a sparse cleanup.
>
> Pre-existing issues:
> - [High] This problem wasn't introduced by this patch, but
> aqc111_rx_fixup() takes pkt_len from the device-controlled per-packet…
>
> --
>
> aqc111: Correct endianness handling in rx/tx_fixup
>
> This commit moves the RX and TX descriptor accesses in aqc111_rx_fixup()
> and aqc111_tx_fixup() to __le64 pointers with get_unaligned_le64() and
> put_unaligned_le64(), to address sparse warnings. It also makes
> aqc111_tx_fixup() always copy cloned skbs via skb_copy_expand().
>
>> aqc111: Correct endianness handling in rx/tx_fixup
>
> [Severity: Low]
> Does "endianness handling" accurately describe what changes here?
>
> The old code already got the byte order right. le64_to_cpup() is
> le64_to_cpu(*p), and the TX path stored cpu_to_le64(tx_desc). Sparse was
> only complaining about the missing __le64 annotations.
>
> The runtime changes in the patch are different, and the message doesn't
> describe them:
>
> - Plain 64-bit loads and stores become get_unaligned_le64() and
> put_unaligned_le64(), which is an alignment fix. aqc111 never sets
> EVENT_NO_IP_ALIGN, so RX skbs carry the NET_IP_ALIGN offset, and
> desc_offset is a multiple of 8. On NET_IP_ALIGN=2 architectures, the
> trailing descriptor header and the descriptor array both sit at 2 mod 8.
> On TX, skb->data - 8 is normally 2 mod 4.
>
> - aqc111_tx_fixup() now always copies cloned skbs, and the message covers
> that with only this sentence:
>
>> Fix these problems and additionally also make sure that cloned
>> skbs are always copied for expansion.
>
> Could the subject and body describe the alignment change and the
> cloned-skb change, so the patch isn't mistaken for a cosmetic sparse
> cleanup?
I will adapt the commit message and subject and make sure this is not
misunderstood as cosmetics.
>
> [Severity: Low]
> Should this carry Fixes: tags, and maybe be split into an alignment fix and
> a cloned-skb fix?
>
> Two problems are fixed here. The 64-bit descriptor accesses in
> aqc111_rx_fixup() and aqc111_tx_fixup() are misaligned. aqc111_tx_fixup()
> also writes the TX descriptor and padding in place into the headroom and
> tailroom of cloned skbs. Both go back to the original aqc111 RX/TX data
> path commits.
>
> Without Fixes: tags or a stable Cc, these fixes may not reach stable
> kernels, and the patch may go to net-next instead of net.
>
>> diff --git a/drivers/net/usb/aqc111.c b/drivers/net/usb/aqc111.c
>> index dd53f413c38f6..39d8755ca8e7b 100644
>> --- a/drivers/net/usb/aqc111.c
>> +++ b/drivers/net/usb/aqc111.c
>
> [ ... ]
>
>> @@ -1116,10 +1116,10 @@ static int aqc111_rx_fixup(struct usbnet *dev, struct sk_buff *skb)
>> goto err;
>>
>> /* Get the first RX packet descriptor */
>> - pkt_desc_ptr = (u64 *)(skb->data + desc_offset);
>> + pkt_desc_ptr = (__le64 *)(skb->data + desc_offset);
>>
>> while (pkt_count--) {
>> - u64 pkt_desc = le64_to_cpup(pkt_desc_ptr);
>> + u64 pkt_desc = get_unaligned_le64(pkt_desc_ptr);
>> u32 pkt_len_with_padd = 0;
>> u32 pkt_len = 0;
>
I will keep this together but describe the fix better. The fix comprises
very few spot changes and should also be understandable as is.
> [Severity: High]
> This is a pre-existing issue, not one introduced by this patch, but is
> aqc111_rx_fixup() missing a lower bound on pkt_len?
>
> pkt_len comes from the device-supplied descriptor, and only an upper bound
> is checked:
>
> drivers/net/usb/aqc111.c:aqc111_rx_fixup() {
> ...
> if (pkt_desc & AQ_RX_PD_DROP ||
> !(pkt_desc & AQ_RX_PD_RX_OK) ||
> pkt_len > (dev->hard_mtu + AQ_RX_HW_PAD)) {
> ...
> skb_put(new_skb, pkt_len);
> memcpy(new_skb->data, skb->data, pkt_len);
> skb_pull(new_skb, AQ_RX_HW_PAD);
> ...
> usbnet_skb_return(dev, new_skb);
> ...
> }
>
> A malicious or broken USB device can report a pkt_len below
> ETH_HLEN + AQ_RX_HW_PAD and still keep pkt_total_offset equal to
> desc_offset. When pkt_len < 2, the skb_pull() above silently fails.
>
> The runt skb then goes through usbnet_skb_return()->eth_type_trans(), which
> doesn't check the length. The 14 byte ethhdr is read from bytes that were
> never written, so skb->protocol and pkt_type come from uninitialized
> memory. The skb is then passed to netif_rx() without ETH_HLEN having been
> pulled.
>
> If the device is bridged, br_dev_queue_push_xmit() calls
> skb_push(skb, ETH_HLEN) into headroom that was never written. That could
> transmit about 12 bytes of stale kernel memory on another port. An
> AF_PACKET reader could also get uninitialized bytes in sll_addr.
>
> Would rejecting pkt_len < ETH_HLEN + AQ_RX_HW_PAD in the drop condition
> above prevent this?
>
I will add this fix.
> [ ... ]
>
>> @@ -1220,7 +1220,8 @@ static struct sk_buff *aqc111_tx_fixup(struct usbnet *dev, struct sk_buff *skb,
>> headroom = skb_headroom(skb);
>> tailroom = skb_tailroom(skb);
>>
>> - if (!(headroom >= sizeof(tx_desc) && tailroom >= padding_size)) {
>> + if (skb_cloned(skb) ||
>> + !(headroom >= sizeof(tx_desc) && tailroom >= padding_size)) {
>> new_skb = skb_copy_expand(skb, sizeof(tx_desc),
>> padding_size, flags);
>> dev_kfree_skb_any(skb);
>
> [Severity: Medium]
> Does this new skb_cloned() check force a full payload copy for every TCP
> data skb?
>
> TCP transmits clones from __tcp_transmit_skb(), so skb_cloned() is true
> for these skbs by the time they reach ndo_start_xmit. TCP also marks them
> nohdr, so skb_header_cloned() is false. Writing the 8 byte descriptor into
> their headroom was therefore already allowed.
>
> skb_copy_expand() allocates a new linear buffer of
> 8 + skb->len + padding_size using the caller's GFP flags. It then copies
> the whole payload, including the frags.
>
> This bites hardest when usb_device_no_sg_constraint() sets can_dma_sg in
> aqc111_bind() and AQ_SUPPORT_FEATURE enables NETIF_F_SG and NETIF_F_TSO.
> In that case the skb_linearize() call earlier in aqc111_tx_fixup() is
> skipped.
>
> Before this patch, a nonlinear TSO skb with padding_size == 0 went straight
> to skb_push(). Now it gets a linearizing copy of up to about 64KB. That
> needs a high-order atomic allocation, and the packet is dropped if the
> allocation fails.
>
> Linear cloned skbs are copied too, for example every packet while an
> AF_PACKET tap is active. Nonlinear skbs that need padding were already
> copied before this patch, because skb_tailroom() returns 0 for them.
>
> The sibling ax88179_tx_fixup() only unshares the head:
>
> drivers/net/usb/ax88179_178a.c:ax88179_tx_fixup() {
> ...
> if ((skb_header_cloned(skb) || headroom < 0) &&
> pskb_expand_head(skb, headroom < 0 ? 8 : 0, 0, GFP_ATOMIC)) {
> ...
> }
>
> Would skb_header_cloned() plus pskb_expand_head(), sized for the
> descriptor and the padding, be enough here?
The change fixes a definite error in case the skb was cloned. The fix
may be relatively resource-intensive in case it is triggered, however it
is preferable for the code to be correct. I cannot further optimize this
code as I do not have actual hardware to test.
>
> The commit message also doesn't say which problem the cloned check fixes.
I will add this.
prev parent reply other threads:[~2026-11-22 10:56 UTC|newest]
Thread overview: 7+ messages / expand[flat|nested] mbox.gz Atom feed top
2026-11-17 18:28 Birger Koblitz
2026-11-17 18:33 ` netdev-bot+sinfo
2026-11-19 19:41 ` Andrew Lunn
2026-11-22 7:38 ` Birger Koblitz
2026-11-22 12:47 ` Daniel Golle
2026-11-20 6:29 ` netdev-bot+sashiko
2026-11-22 10:55 ` Birger Koblitz [this message]
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=a6cac285-5df6-4440-bd58-99e29c71b0df@birger-koblitz.de \
--to=mail@birger-koblitz.de \
--cc=andrew+netdev@lunn.ch \
--cc=davem@davemloft.net \
--cc=edumazet@google.com \
--cc=kuba@kernel.org \
--cc=linux-kernel@vger.kernel.org \
--cc=linux-usb@vger.kernel.org \
--cc=netdev-bot+sashiko@kernel.org \
--cc=netdev@vger.kernel.org \
--cc=pabeni@redhat.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox
all inboxes | Powered by JetHome®