bugfix: udptcp_checksum should tread tcp and udp differently
diff mbox series

Message ID 20200527144127.5792-1-guohongzhi1@huawei.com
State Superseded
Delegated to: Thomas Monjalon
Headers show
Series
  • bugfix: udptcp_checksum should tread tcp and udp differently
Related show

Checks

Context Check Description
ci/Intel-compilation success Compilation OK
ci/travis-robot success Travis build: passed
ci/iol-testing warning Testing issues
ci/iol-mellanox-Performance success Performance Testing PASS
ci/iol-nxp-Performance success Performance Testing PASS
ci/iol-intel-Performance success Performance Testing PASS
ci/checkpatch success coding style OK

Commit Message

Hongzhi Guo May 27, 2020, 2:41 p.m. UTC
From: Hongzhi Guo <guohongzhi1@huawei.com>

RFC 768 for UDP specifies:
If the computed  checksum  is zero,  it is transmitted  as all ones.
An all zero transmitted checksum  value means that the transmitter
generated  no checksum.

RFC 793 for TCP has no such special treatment for the checksum of zero.

Signed-off-by: Hongzhi Guo <guohongzhi1@huawei.com>
---
 lib/librte_net/rte_ip.h | 13 +++++++++----
 1 file changed, 9 insertions(+), 4 deletions(-)

Comments

Stephen Hemminger May 27, 2020, 3:02 p.m. UTC | #1
On Wed, 27 May 2020 22:41:27 +0800
guohongzhi <guohongzhi1@huawei.com> wrote:

> +	/* For Udp, if the computed checksum is zero,
> +	 * it is transmitted as all ones.RFC768
> +	 */
> +	if (cksum == 0 && ipv4_hdr->next_proto_id == IPPROTO_UDP)
>  		cksum = 0xffff;
>  


The comment should be reformatted to be clearer.

Maybe:
	/*
	 * Per RFC768:
	 * If the computed  checksum  is zero,it is transmitted  as all ones
	 * (the equivalent  in one's complement  arithmetic).
	 */

There is no special case required here, it is true for TCP as well.
In TCP it appears 0 is allowed but not generally used. Most implementations
use common checksum for both TCP and UDP
Morten Brørup May 27, 2020, 3:36 p.m. UTC | #2
> From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Stephen Hemminger
> Sent: Wednesday, May 27, 2020 5:03 PM
> 
> On Wed, 27 May 2020 22:41:27 +0800
> guohongzhi <guohongzhi1@huawei.com> wrote:
> 
> > +	/* For Udp, if the computed checksum is zero,
> > +	 * it is transmitted as all ones.RFC768
> > +	 */
> > +	if (cksum == 0 && ipv4_hdr->next_proto_id == IPPROTO_UDP)
> >  		cksum = 0xffff;
> >
> 
> 
> The comment should be reformatted to be clearer.
> 
> Maybe:
> 	/*
> 	 * Per RFC768:
> 	 * If the computed  checksum  is zero,it is transmitted  as all
> ones
> 	 * (the equivalent  in one's complement  arithmetic).
> 	 */
> 

But without the double spaces. :-)

> There is no special case required here, it is true for TCP as well.

I disagree. I researched this topic when Hongzhi Guo initially submitted the patches, and have only seen the special exception mentioned in RFC 768 (describing UDP), where a transmitted value of 0x0000 means that the checksum has not been generated. None of the RFCs regarding Internet Checksum or TCP checksum mentions this special case.

> In TCP it appears 0 is allowed but not generally used. Most
> implementations
> use common checksum for both TCP and UDP

Then most implementations are wrong.

Jon Postel famously wrote: "Be liberal in what you accept, and conservative in what you send."

This function is in the transmission path, so we should calculate it correctly.

In the receive path, when checking the checksum as described in RFC 1071, any of the two values (0x0000 or 0xFFFF) in the Checksum field will yield the correct result. Which is why most implementations can get away with doing it wrong.
Olivier Matz July 6, 2020, 7:58 a.m. UTC | #3
Hi,

Here is a suggestion for the title:

  net: fix uneeded replacement of 0 by ffff for TCP checksum

The commit log looks good to me, but you can add:

  Fixes: 6006818cfb26 ("net: new checksum functions")
  Cc: stable@dpdk.org

On Wed, May 27, 2020 at 05:36:59PM +0200, Morten Brørup wrote:
> > From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Stephen Hemminger
> > Sent: Wednesday, May 27, 2020 5:03 PM
> > 
> > On Wed, 27 May 2020 22:41:27 +0800
> > guohongzhi <guohongzhi1@huawei.com> wrote:
> >
> > > --- a/lib/librte_net/rte_ip.h
> > > +++ b/lib/librte_net/rte_ip.h
> > > @@ -324,8 +324,7 @@ rte_ipv4_phdr_cksum(const struct rte_ipv4_hdr *ipv4_hdr, uint64_t ol_flags)
> > >   * @param l4_hdr
> > >   *   The pointer to the beginning of the L4 header.
> > >   * @return
> > > - *   The complemented checksum to set in the IP packet
> > > - *   or 0 on error
> > > + *   The complemented checksum to set in the IP packet.
> > >   */
> > >  static inline uint16_t
> > >  rte_ipv4_udptcp_cksum(const struct rte_ipv4_hdr *ipv4_hdr, const void *l4_hdr)

It still returns 0 when ipv4_hdr->total_length < sizeof(struct rte_ipv4_hdr).
I suggest to keep this case in the API comment, maybe like this:

  The complemented checksum to set in the IP packet
  or 0 if the IP length is invalid in the header.


> > > +	/* For Udp, if the computed checksum is zero,
> > > +	 * it is transmitted as all ones.RFC768
> > > +	 */
> > > +	if (cksum == 0 && ipv4_hdr->next_proto_id == IPPROTO_UDP)
> > >  		cksum = 0xffff;
> > >
> > 
> > 
> > The comment should be reformatted to be clearer.
> > 
> > Maybe:
> > 	/*
> > 	 * Per RFC768:
> > 	 * If the computed  checksum  is zero,it is transmitted  as all
> > ones
> > 	 * (the equivalent  in one's complement  arithmetic).
> > 	 */
> > 
> 
> But without the double spaces. :-)

I agree, Stephen's wording looks better, can you please use it?

> > There is no special case required here, it is true for TCP as well.
> 
> I disagree. I researched this topic when Hongzhi Guo initially submitted the patches, and have only seen the special exception mentioned in RFC 768 (describing UDP), where a transmitted value of 0x0000 means that the checksum has not been generated. None of the RFCs regarding Internet Checksum or TCP checksum mentions this special case.
> 
> > In TCP it appears 0 is allowed but not generally used. Most
> > implementations
> > use common checksum for both TCP and UDP
> 
> Then most implementations are wrong.
> 
> Jon Postel famously wrote: "Be liberal in what you accept, and conservative in what you send."
> 
> This function is in the transmission path, so we should calculate it correctly.
> 
> In the receive path, when checking the checksum as described in RFC 1071, any of the two values (0x0000 or 0xFFFF) in the Checksum field will yield the correct result. Which is why most implementations can get away with doing it wrong.

I agree with Morten, it looks more strict like this.

Patch
diff mbox series

diff --git a/lib/librte_net/rte_ip.h b/lib/librte_net/rte_ip.h
index 1ceb7b7..8bef313 100644
--- a/lib/librte_net/rte_ip.h
+++ b/lib/librte_net/rte_ip.h
@@ -324,8 +324,7 @@  rte_ipv4_phdr_cksum(const struct rte_ipv4_hdr *ipv4_hdr, uint64_t ol_flags)
  * @param l4_hdr
  *   The pointer to the beginning of the L4 header.
  * @return
- *   The complemented checksum to set in the IP packet
- *   or 0 on error
+ *   The complemented checksum to set in the IP packet.
  */
 static inline uint16_t
 rte_ipv4_udptcp_cksum(const struct rte_ipv4_hdr *ipv4_hdr, const void *l4_hdr)
@@ -344,7 +343,10 @@  rte_ipv4_udptcp_cksum(const struct rte_ipv4_hdr *ipv4_hdr, const void *l4_hdr)
 
 	cksum = ((cksum & 0xffff0000) >> 16) + (cksum & 0xffff);
 	cksum = (~cksum) & 0xffff;
-	if (cksum == 0)
+	/* For Udp, if the computed checksum is zero,
+	 * it is transmitted as all ones.RFC768
+	 */
+	if (cksum == 0 && ipv4_hdr->next_proto_id == IPPROTO_UDP)
 		cksum = 0xffff;
 
 	return (uint16_t)cksum;
@@ -436,7 +438,10 @@  rte_ipv6_udptcp_cksum(const struct rte_ipv6_hdr *ipv6_hdr, const void *l4_hdr)
 
 	cksum = ((cksum & 0xffff0000) >> 16) + (cksum & 0xffff);
 	cksum = (~cksum) & 0xffff;
-	if (cksum == 0)
+	/* For Udp, if the computed checksum is zero,
+	 * it is transmitted as all ones.RFC768
+	 */
+	if (cksum == 0 && ipv6_hdr->proto == IPPROTO_UDP)
 		cksum = 0xffff;
 
 	return (uint16_t)cksum;