[0/2] prevent out of bounds read with checksum

  The functions for checksumming the packet payload don't perform bounds
checks, and are used by the TAP driver which does not do any bounds checks
on the incoming packet either. This means a packet received with an
incorrect IP header can read beyond the end of the mbuf.

In the worst case, where the length is specified as being smaller than the
IPv4 header, 32-bit wrap-around on subtraction occurs, meaning that approx
4GB of memory will be read.

To fix this, we can introduce a sanity check into the ipv4 function to
ensure that underflow does not occur. Since the checksum function does not
take the mbuf length as a parameter, we cannot check for overflow there,
so we instead perform the checks in the TAP driver directly.

Ideally, in a future release, all checksum functions should be modified to
take a max buffer length parameter to fix this issue globally.

NOTE: It appears that the dpaa driver also uses these functions, but from
what I can see there, they are only used on TX, which means that there
should be less need for parameter length checking, as the data does not
come from an untrusted source. Perhaps maintainers, Hemant and Shreyansh,
can confirm?

CC: Hemant Agrawal <hemant.agrawal@nxp.com>
CC: Shreyansh Jain <shreyansh.jain@nxp.com>

Bruce Richardson (2):
  net: fix underflow for checksum of invalid IPv4 packets
  net/tap: add buffer overflow checks before checksum

 drivers/net/tap/rte_eth_tap.c | 14 ++++++++++++++
 lib/librte_net/rte_ip.h       | 12 ++++++++----
 2 files changed, 22 insertions(+), 4 deletions(-)
  

HI Bruce,

On 17-Dec-18 9:20 PM, Bruce Richardson wrote:
> The functions for checksumming the packet payload don't perform bounds
> checks, and are used by the TAP driver which does not do any bounds checks
> on the incoming packet either. This means a packet received with an
> incorrect IP header can read beyond the end of the mbuf.
>
> In the worst case, where the length is specified as being smaller than the
> IPv4 header, 32-bit wrap-around on subtraction occurs, meaning that approx
> 4GB of memory will be read.
>
> To fix this, we can introduce a sanity check into the ipv4 function to
> ensure that underflow does not occur. Since the checksum function does not
> take the mbuf length as a parameter, we cannot check for overflow there,
> so we instead perform the checks in the TAP driver directly.
>
> Ideally, in a future release, all checksum functions should be modified to
> take a max buffer length parameter to fix this issue globally.
>
> NOTE: It appears that the dpaa driver also uses these functions, but from
> what I can see there, they are only used on TX, which means that there
> should be less need for parameter length checking, as the data does not
> come from an untrusted source. Perhaps maintainers, Hemant and Shreyansh,
> can confirm?

In DPAA, we are using software based checksum calculation for self 
generated packets largely.

They are mostly trust worthy unless someone is deliberately or 
mistakenly trying to send a corrupt packet.

We will check, if we can also add some checks in DPAA driver in these 
legs without making performance impact for self generated packets.


Regards,

Hemant

>
> CC: Hemant Agrawal <hemant.agrawal@nxp.com>
> CC: Shreyansh Jain <shreyansh.jain@nxp.com>
>
> Bruce Richardson (2):
>    net: fix underflow for checksum of invalid IPv4 packets
>    net/tap: add buffer overflow checks before checksum
>
>   drivers/net/tap/rte_eth_tap.c | 14 ++++++++++++++
>   lib/librte_net/rte_ip.h       | 12 ++++++++----
>   2 files changed, 22 insertions(+), 4 deletions(-)
>
  

> -----Original Message-----
> From: Hemant Agrawal [mailto:hemant.agrawal@nxp.com]
> Sent: Tuesday, December 18, 2018 12:50 PM
> To: Richardson, Bruce <bruce.richardson@intel.com>; Olivier Matz
> <olivier.matz@6wind.com>; Wiles, Keith <keith.wiles@intel.com>
> Cc: dev@dpdk.org; Shreyansh Jain <shreyansh.jain@nxp.com>
> Subject: Re: [dpdk-dev] [PATCH 0/2] prevent out of bounds read with
> checksum
> 
> HI Bruce,
> 
> On 17-Dec-18 9:20 PM, Bruce Richardson wrote:
> > The functions for checksumming the packet payload don't perform bounds
> > checks, and are used by the TAP driver which does not do any bounds
> > checks on the incoming packet either. This means a packet received
> > with an incorrect IP header can read beyond the end of the mbuf.
> >
> > In the worst case, where the length is specified as being smaller than
> > the
> > IPv4 header, 32-bit wrap-around on subtraction occurs, meaning that
> > approx 4GB of memory will be read.
> >
> > To fix this, we can introduce a sanity check into the ipv4 function to
> > ensure that underflow does not occur. Since the checksum function does
> > not take the mbuf length as a parameter, we cannot check for overflow
> > there, so we instead perform the checks in the TAP driver directly.
> >
> > Ideally, in a future release, all checksum functions should be
> > modified to take a max buffer length parameter to fix this issue
> globally.
> >
> > NOTE: It appears that the dpaa driver also uses these functions, but
> > from what I can see there, they are only used on TX, which means that
> > there should be less need for parameter length checking, as the data
> > does not come from an untrusted source. Perhaps maintainers, Hemant
> > and Shreyansh, can confirm?
> 
> In DPAA, we are using software based checksum calculation for self
> generated packets largely.
> 
> They are mostly trust worthy unless someone is deliberately or mistakenly
> trying to send a corrupt packet.
> 
> We will check, if we can also add some checks in DPAA driver in these legs
> without making performance impact for self generated packets.
> 

Right. Thanks for confirming it's not on RX path which would be the main risk.
I would assume that data coming from the app should be trusted, unless the
app is deliberately trying to crash itself. :-) (I didn't look to try and fix this
in DPAA because of that assumption, but glad you are looking into it.)
  

On 12/17/2018 3:50 PM, Bruce Richardson wrote:
> The functions for checksumming the packet payload don't perform bounds
> checks, and are used by the TAP driver which does not do any bounds checks
> on the incoming packet either. This means a packet received with an
> incorrect IP header can read beyond the end of the mbuf.
> 
> In the worst case, where the length is specified as being smaller than the
> IPv4 header, 32-bit wrap-around on subtraction occurs, meaning that approx
> 4GB of memory will be read.
> 
> To fix this, we can introduce a sanity check into the ipv4 function to
> ensure that underflow does not occur. Since the checksum function does not
> take the mbuf length as a parameter, we cannot check for overflow there,
> so we instead perform the checks in the TAP driver directly.
> 
> Ideally, in a future release, all checksum functions should be modified to
> take a max buffer length parameter to fix this issue globally.
> 
> NOTE: It appears that the dpaa driver also uses these functions, but from
> what I can see there, they are only used on TX, which means that there
> should be less need for parameter length checking, as the data does not
> come from an untrusted source. Perhaps maintainers, Hemant and Shreyansh,
> can confirm?
> 
> CC: Hemant Agrawal <hemant.agrawal@nxp.com>
> CC: Shreyansh Jain <shreyansh.jain@nxp.com>
> 
> Bruce Richardson (2):
>   net: fix underflow for checksum of invalid IPv4 packets
>   net/tap: add buffer overflow checks before checksum

Series applied to dpdk-next-net/master, thanks.