Re: [PATCH v2] net/e1000e_core: adjust count if RDH exceeds RDT in e1000e_ring_advance()

[Jason Wang <jasowang@redhat.com>]

On 2020/11/27 下午10:49, Mauro Matteo Cascella wrote:
> On Fri, Nov 27, 2020 at 6:21 AM Jason Wang <jasowang@redhat.com> wrote:
>>
>> On 2020/11/24 上午5:30, Mauro Matteo Cascella wrote:
>>> On Thu, Nov 19, 2020 at 6:57 AM Jason Wang <jasowang@redhat.com> wrote:
>>>> On 2020/11/18 下午4:53, Mauro Matteo Cascella wrote:
>>>>> On Wed, Nov 18, 2020 at 4:56 AM Jason Wang <jasowang@redhat.com> wrote:
>>>>>> On 2020/11/13 下午6:31, Mauro Matteo Cascella wrote:
>>>>>>> The e1000e_write_packet_to_guest() function iterates over a set of
>>>>>>> receive descriptors by advancing rx descriptor head register (RDH) from
>>>>>>> its initial value to rx descriptor tail register (RDT). The check in
>>>>>>> e1000e_ring_empty() is responsible for detecting whether RDH has reached
>>>>>>> RDT, terminating the loop if that's the case. Additional checks have
>>>>>>> been added in the past to deal with bogus values submitted by the guest
>>>>>>> to prevent possible infinite loop. This is done by "wrapping around" RDH
>>>>>>> at some point and detecting whether it assumes the original value during
>>>>>>> the loop.
>>>>>>>
>>>>>>> However, when e1000e is configured to use the packet split feature, RDH is
>>>>>>> incremented by two instead of one, as the packet split descriptors are
>>>>>>> 32 bytes while regular descriptors are 16 bytes. A malicious or buggy
>>>>>>> guest may set RDT to an odd value and transmit only null RX descriptors.
>>>>>>> This corner case would prevent RDH from ever matching RDT, leading to an
>>>>>>> infinite loop. This patch adds a check in e1000e_ring_advance() to make sure
>>>>>>> RDH does not exceed RDT in a single incremental step, adjusting the count
>>>>>>> value accordingly.
>>>>>> Can this patch solve this issue in another way?
>>>>>>
>>>>>> https://patchew.org/QEMU/20201111130636.2208620-1-ppandit@redhat.com/
>>>>>>
>>>>>> Thanks
>>>>>>
>>>>> Yes, it does work nicely. Still, I think this patch is useful to avoid
>>>>> possible inconsistent state in e1000e_ring_advance() when count > 1.
>>>> So if RDT is odd, it looks to me the following codes in
>>>> e1000e_write_packet_to_guest() needs to be fixed as well.
>>>>
>>>>
>>>>            base = e1000e_ring_head_descr(core, rxi);
>>>>
>>>>            pci_dma_read(d, base, &desc, core->rx_desc_len);
>>>>
>>>> Otherwise e1000e may try to read out of descriptor ring.
>>> Sorry, I'm not sure I understand what you mean. Isn't the base address
>>> computed from RDH? How can e1000e read out of the descriptor ring if
>>> RDT is odd?
>>>
>>>> Thanks
>>> On Thu, Nov 19, 2020 at 6:57 AM Jason Wang <jasowang@redhat.com> wrote:
>>>> On 2020/11/18 下午4:53, Mauro Matteo Cascella wrote:
>>>>> On Wed, Nov 18, 2020 at 4:56 AM Jason Wang <jasowang@redhat.com> wrote:
>>>>>> On 2020/11/13 下午6:31, Mauro Matteo Cascella wrote:
>>>>>>> The e1000e_write_packet_to_guest() function iterates over a set of
>>>>>>> receive descriptors by advancing rx descriptor head register (RDH) from
>>>>>>> its initial value to rx descriptor tail register (RDT). The check in
>>>>>>> e1000e_ring_empty() is responsible for detecting whether RDH has reached
>>>>>>> RDT, terminating the loop if that's the case. Additional checks have
>>>>>>> been added in the past to deal with bogus values submitted by the guest
>>>>>>> to prevent possible infinite loop. This is done by "wrapping around" RDH
>>>>>>> at some point and detecting whether it assumes the original value during
>>>>>>> the loop.
>>>>>>>
>>>>>>> However, when e1000e is configured to use the packet split feature, RDH is
>>>>>>> incremented by two instead of one, as the packet split descriptors are
>>>>>>> 32 bytes while regular descriptors are 16 bytes. A malicious or buggy
>>>>>>> guest may set RDT to an odd value and transmit only null RX descriptors.
>>>>>>> This corner case would prevent RDH from ever matching RDT, leading to an
>>>>>>> infinite loop. This patch adds a check in e1000e_ring_advance() to make sure
>>>>>>> RDH does not exceed RDT in a single incremental step, adjusting the count
>>>>>>> value accordingly.
>>>>>> Can this patch solve this issue in another way?
>>>>>>
>>>>>> https://patchew.org/QEMU/20201111130636.2208620-1-ppandit@redhat.com/
>>>>>>
>>>>>> Thanks
>>>>>>
>>>>> Yes, it does work nicely. Still, I think this patch is useful to avoid
>>>>> possible inconsistent state in e1000e_ring_advance() when count > 1.
>>>> So if RDT is odd, it looks to me the following codes in
>>>> e1000e_write_packet_to_guest() needs to be fixed as well.
>>>>
>>>>
>>>>            base = e1000e_ring_head_descr(core, rxi);
>>>>
>>>>            pci_dma_read(d, base, &desc, core->rx_desc_len);
>>>>
>>>> Otherwise e1000e may try to read out of descriptor ring.
>>>>
>>>> Thanks
>>
>> Sorry, I meant RDH actually, when packet split descriptor is used, it
>> doesn't check whether DH exceeds DLEN?
>>
> When the packet split feature is used (i.e., count > 1) this patch
> basically sets RDH=RDT in case the increment would exceed RDT.


Can software set RDH to an odd value? If not, I think we are probably fine.

Thanks


> The
> next iteration should detect that RDH equals RDT in
> e1000e_ring_empty(), and exit the loop right before pci_dma_read(). On
> the other hand RDH is set to zero if it exceeds DLEN in
> e1000e_ring_advance() so we should be fine in either case, unless I'm
> missing something?
>
>
> Thank you for your time,
> --
> Mauro Matteo Cascella
> Red Hat Product Security
> PGP-Key ID: BB3410B0
>