Re: [PATCH] virtio-net: Copy all for dhclient workaround

[Akihiko Odaki <akihiko.odaki@daynix.com>]

On 2025/04/11 22:20, Antoine Damhet wrote:
> On Fri, Apr 11, 2025 at 05:01:01PM +0900, Akihiko Odaki wrote:
>> On 2025/04/07 17:29, Antoine Damhet wrote:
>>> On Sat, Apr 05, 2025 at 05:04:28PM +0900, Akihiko Odaki wrote:
>>>> The goal of commit 7987d2be5a8b ("virtio-net: Copy received header to
>>>> buffer") was to remove the need to patch the (const) input buffer with a
>>>> recomputed UDP checksum by copying headers to a RW region and inject the
>>>> checksum there. The patch computed the checksum only from the header
>>>> fields (missing the rest of the payload) producing an invalid one
>>>> and making guests fail to acquire a DHCP lease.
>>>>
>>>> Fix the issue by copying the entire packet instead of only copying the
>>>> headers.
>>>>
>>>> Fixes: 7987d2be5a8b ("virtio-net: Copy received header to buffer")
>>>> Resolves: https://gitlab.com/qemu-project/qemu/-/issues/2727
>>>> Cc: qemu-stable@nongnu.org
>>>> Signed-off-by: Akihiko Odaki <akihiko.odaki@daynix.com>
>>>
>>> Tested-By: Antoine Damhet <adamhet@scaleway.com>
>>>
>>>> ---
>>>> This patch aims to resolves the issue the following one also does:
>>>> https://lore.kernel.org/qemu-devel/20250404151835.328368-1-adamhet@scaleway.com
>>>>
>>>> The difference from the mentioned patch is that this patch also
>>>> preserves that the original intent of regressing change, which is to
>>>> remove the need to patch the (const) input buffer with a recomputed UDP
>>>> checksum.
>>>>
>>>> To Antoine Damhet:
>>>> I confirmed that DHCP is currently not working and this patch fixes the
>>>> issue, but I would appreciate if you also confirm the fix as I already
>>>> have done testing badly for the regressing patch.
>>>
>>> Thanks for the swift response, ideally I'd like a non-regression test in
>>> the testsuite but a quick test showed me that I couldn't easily
>>> reproduce with user networking so unless someone has a great idea it
>>> would be a pain.
>>>
>>>> ---
>>>>    hw/net/virtio-net.c | 35 ++++++++++++++++-------------------
>>>>    1 file changed, 16 insertions(+), 19 deletions(-)
>>>>
>>>> diff --git a/hw/net/virtio-net.c b/hw/net/virtio-net.c
>>>> index de87cfadffe1..a920358a89c5 100644
>>>> --- a/hw/net/virtio-net.c
>>>> +++ b/hw/net/virtio-net.c
>>>> @@ -1687,6 +1687,11 @@ static void virtio_net_hdr_swap(VirtIODevice *vdev, struct virtio_net_hdr *hdr)
>>>>        virtio_tswap16s(vdev, &hdr->csum_offset);
>>>>    }
>>>> +typedef struct Header {
>>>> +    struct virtio_net_hdr_v1_hash virtio_net;
>>>> +    uint8_t payload[1500];
>>>> +} Header;
>>>> +
>>>>    /* dhclient uses AF_PACKET but doesn't pass auxdata to the kernel so
>>>>     * it never finds out that the packets don't have valid checksums.  This
>>>>     * causes dhclient to get upset.  Fedora's carried a patch for ages to
>>>> @@ -1701,7 +1706,7 @@ static void virtio_net_hdr_swap(VirtIODevice *vdev, struct virtio_net_hdr *hdr)
>>>>     * we should provide a mechanism to disable it to avoid polluting the host
>>>>     * cache.
>>>>     */
>>>> -static void work_around_broken_dhclient(struct virtio_net_hdr *hdr,
>>>> +static void work_around_broken_dhclient(struct Header *hdr,
>>>>                                            size_t *hdr_len, const uint8_t *buf,
>>>>                                            size_t buf_size, size_t *buf_offset)
>>>>    {
>>>> @@ -1711,20 +1716,20 @@ static void work_around_broken_dhclient(struct virtio_net_hdr *hdr,
>>>>        buf += *buf_offset;
>>>>        buf_size -= *buf_offset;
>>>> -    if ((hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) && /* missing csum */
>>>> -        (buf_size >= csum_size && buf_size < 1500) && /* normal sized MTU */
>>>> +    if ((hdr->virtio_net.hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) && /* missing csum */
>>>> +        (buf_size >= csum_size && buf_size < sizeof(hdr->payload)) && /* normal sized MTU */
>>>>            (buf[12] == 0x08 && buf[13] == 0x00) && /* ethertype == IPv4 */
>>>>            (buf[23] == 17) && /* ip.protocol == UDP */
>>>>            (buf[34] == 0 && buf[35] == 67)) { /* udp.srcport == bootps */
>>>> -        memcpy((uint8_t *)hdr + *hdr_len, buf, csum_size);
>>>> -        net_checksum_calculate((uint8_t *)hdr + *hdr_len, csum_size, CSUM_UDP);
>>>> -        hdr->flags &= ~VIRTIO_NET_HDR_F_NEEDS_CSUM;
>>>> -        *hdr_len += csum_size;
>>>> -        *buf_offset += csum_size;
>>>> +        memcpy((uint8_t *)hdr + *hdr_len, buf, buf_size);
>>>> +        net_checksum_calculate((uint8_t *)hdr + *hdr_len, buf_size, CSUM_UDP);
>>>> +        hdr->virtio_net.hdr.flags &= ~VIRTIO_NET_HDR_F_NEEDS_CSUM;
>>>> +        *hdr_len += buf_size;
>>>> +        *buf_offset += buf_size;
>>>>        }
>>>>    }
>>>> -static size_t receive_header(VirtIONet *n, struct virtio_net_hdr *hdr,
>>>> +static size_t receive_header(VirtIONet *n, Header *hdr,
>>>>                                 const void *buf, size_t buf_size,
>>>>                                 size_t *buf_offset)
>>>
>>> `receive_header` can now "receive" the whole packet that's kinda
>>> misleading. I though another approach would be to only do the
>>> detection/flag patching from receive_header and recompute the checksum
>>> directly in the final `iov`, this would also eliminate the extra payload
>>> copy.
>>
>> It is possible to avoid copying but I chose not to do that because this is
>> not a hot path and the code complexity required for that does not look
>> worthwhile for me.
> 
> Understood and OK.
> 
>>
>> But I agree that the names of receive_header() and Header structure are
>> misleading. The reasoning I used to convince myself is that the "Header" is
>> at the head of the packet at least. I'd like to hear if you have an idea of
>> better naming; otherwise I would rather leave it as is.
> 
> Maybe we can sidestep this entirely, do we need to do the workaround
> _inside_ `receive_header` ? WDYT of the following pseudocode:
> 
> ```
> guest_offset = receive_header(&header);
> iov_from_buf(&header);
> work_around_broken_dhclient(&header, &payload);
> iov_from_buf(&payload);
> ```

net_checksum_calculate() currently needs a contiguous buffer so it needs 
to be changed and it also requires one additional iov_from_buf() call. 
It's a bit too complicated to workaround the naming problem I think.

> 
> If not maybe something along the line of "PacketPrefix" or
> "PacketStart".

Now I'm inclined for "PacketPrefix". In a normal context, "prefix" and 
"start" are no different from "header", but in the networking context, 
"header" is frequently used to describe the metadata and implies it 
doesn't contain data. Usually I don't like to choose wordings according 
to such an implied nuance, but avoiding the word "header" here has a 
practical value.

I'll probably choose "prefix" instead of "start" since it sounds more 
specific than "start".

Regards,
Akihiko Odaki

> 
> Regards,
>