[dpdk-dev] [PATCH v1] gso: fix marking TCP checksum flag in TCP segments

Ananyev, Konstantin konstantin.ananyev at intel.com
Tue Apr 24 14:31:27 CEST 2018


Hi Ophir,

> 
> Hi Konstantin,
> Please see inline
> 
> > -----Original Message-----
> > From: Ananyev, Konstantin [mailto:konstantin.ananyev at intel.com]
> > Sent: Tuesday, April 24, 2018 1:56 PM
> > To: Ophir Munk <ophirmu at mellanox.com>; Hu, Jiayu <jiayu.hu at intel.com>;
> > dev at dpdk.org
> > Cc: Thomas Monjalon <thomas at monjalon.net>; Olga Shern
> > <olgas at mellanox.com>; Pascal Mazon <pascal.mazon at 6wind.com>;
> > stable at dpdk.org
> > Subject: RE: [PATCH v1] gso: fix marking TCP checksum flag in TCP segments
> >
> >
> >
> > > -----Original Message-----
> > > From: Ophir Munk [mailto:ophirmu at mellanox.com]
> > > Sent: Tuesday, April 24, 2018 10:44 AM
> > > To: Hu, Jiayu <jiayu.hu at intel.com>; dev at dpdk.org; Ananyev, Konstantin
> > > <konstantin.ananyev at intel.com>
> > > Cc: Thomas Monjalon <thomas at monjalon.net>; Olga Shern
> > > <olgas at mellanox.com>; Pascal Mazon <pascal.mazon at 6wind.com>;
> > > stable at dpdk.org
> > > Subject: RE: [PATCH v1] gso: fix marking TCP checksum flag in TCP
> > > segments
> > >
> > > Hi Jiayu,
> > > Please find comments inline
> > >
> > > > -----Original Message-----
> > > > From: Hu, Jiayu [mailto:jiayu.hu at intel.com]
> > > > Sent: Monday, April 23, 2018 7:14 AM
> > > > To: Ophir Munk <ophirmu at mellanox.com>; dev at dpdk.org; Ananyev,
> > > > Konstantin <konstantin.ananyev at intel.com>
> > > > Cc: Thomas Monjalon <thomas at monjalon.net>; Olga Shern
> > > > <olgas at mellanox.com>; Pascal Mazon <pascal.mazon at 6wind.com>;
> > > > stable at dpdk.org
> > > > Subject: RE: [PATCH v1] gso: fix marking TCP checksum flag in TCP
> > > > segments
> > > >
> > > > Hi Ophir,
> > > >
> > > > In the GSO design, the GSO library doesn't care about checksums,
> > > > which means it doesn't check if input packets have correct
> > > > checksums, and it doesn't do any checksum related work for the
> > > > output GSO segments. It depends on the callers to use HW or SW
> > > > checksum calculation for output packets. This is why the GSO library
> > > > doesn't set PKT_TX_TCP_CKSUM. So I don't think it's a bug.
> > > >
> > >
> > > Can you please reconsider this design? I think the GSO library should
> > > imitate the HW behavior where TCP segments checksum is automatically
> > > calculated without explicitly requesting it. I am not saying that GSO library
> > itself should calculate the checksums - but at least it should mark each
> > segment as requiring this calculation.
> >
> > But gso has no idea how this packet will be processed after it.
> 
> GSO shouldn't know. It should only mark the fact that a new TCP segment was created without a TCP checksum.
> 

Ok, but new IP header was also created. And might be outer ip/udp (in case of tunnel).
If we go that way we'll have to set flags for each them.

> > Caller can choose to calculate L3/L4 cksum in SW or might be going to use
> > HW offloads.
> 
> Assuming TSO is configured then you suggest that TAP PMD will mark by itself the TCP_CKSUM flag for all packets returned from GSO
> library?

Yes.

> 
> > In later case nothing stops the caller to update mbuf->ol_flags in a way he
> > likes (TCP_CKSUM, IP_CKSUM, etc.).
> > Konstantin
> >
> 
> Please note that TCP_SEG flag is cleared by GSO library in 2 different cases:
> 1. Packet length equals to or is bigger than GSO size. In this case new TCP segments are created with no TCP checksum.
> 2. Packet length is smaller than GSO size. In this case no TCP segmentation is required. The original packet is returned and its existing TCP
> checksum is OK.
> 
> In the latter case TAP PMD will always calculate TCP checksum in SW (performance concerns) where this could have been saved.
> I am thinking of a practical scenario where TSO is configured but all packets are smaller than GSO size, however TAP PMD unnecessarily
> recalculates their checksums.
> 
> How do you suggest to avoid this scenario?

Probably something like that:

rc = rte_gso_segment(pkt_in, gso_ctx, pkts_out, nb_pkts_out);
if (rc == 1 && pkt_in == pkts_out[0] == pkt_in) {
     /* no gso was performed */
} else {
   /* new packets, update ol_flags if needed */
}

?

Another possibility - might be make chages in librte_gso to allow user to
specify what flags to set for the output packets (probably via  rte_gso_ctx.flag) 

Konstantin

> >
> > > > In my opinion, it's not a good idea to enable HW TCP checksum
> > > > calculation silently, and without the aware of the caller. In fact,
> > > > the caller always know it does SW TSO (i.e. GSO), instead of real HW TSO.
> > >
> > > This is not correct. Consider net_failsafe with 2 sub-devices: one is
> > > a HW PCI device, the other one is a SW TAP device. Failsafe must work
> > transparently with these two sub-devices and the caller cannot tell if TSO is
> > done in SW or HW.
> > >
> > > > If the caller wants HW
> > > > checksum calculation, it can add PKT_TX_TCP_CKSUM to ol_flags before
> > > > or after calling the GSO library.
> > > >
> > >
> > > FYI - TAP TSO patches were submitted to dpdk.org mailing list. These
> > patches use the GSO library.
> > >
> > https://emea01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fdpd
> > >
> > k.org%2Fdev%2Fpatchwork%2Fpatch%2F38666%2F&data=02%7C01%7Coph
> > irmu%40me
> > >
> > llanox.com%7C7455c8e31c7a4364bc7108d5a9d20008%7Ca652971c7d2e4d
> > 9ba6a4d1
> > >
> > 49256f461b%7C0%7C0%7C636601641779974217&sdata=CF7EvhXG%2BrH%
> > 2BPiQEbvM0
> > > mC%2FSpqobneKaoV03j5VrSDw%3D&reserved=0
> > >
> > https://emea01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fdpd
> > >
> > k.org%2Fdev%2Fpatchwork%2Fpatch%2F38667%2F&data=02%7C01%7Coph
> > irmu%40me
> > >
> > llanox.com%7C7455c8e31c7a4364bc7108d5a9d20008%7Ca652971c7d2e4d
> > 9ba6a4d1
> > >
> > 49256f461b%7C0%7C0%7C636601641779974217&sdata=j9WVIj%2FKq6EN
> > WXu3mr5By1
> > > toSowU8bqJRGZ19SxiGoI%3D&reserved=0
> > >
> > > Running testpmd with TAP TSO is currently broken without the suggested
> > librte_gso patch.
> > > Please note testpmd implementation (app/test-pmd/csumonly.c
> > > b/app/test-pmd/csumonly.c) in case *both* TSO and TCP CKSUM are
> > > configured:
> > >
> > >   if (tso_segsz)
> > >       ol_flags |= PKT_TX_TCP_SEG;    // *** if TSO is applicable - the packet
> > flags are only marked with PKT_TX_TCP_SEG and no
> > > PKT_TX_TCP_CKSUM ***
> > >    else if (tx_offloads & DEV_TX_OFFLOAD_TCP_CKSUM)
> > >        ol_flags |= PKT_TX_TCP_CKSUM;     // *** PKT_TX_TCP_CKSUM is
> > marked only if TSO is not applicable ***
> > >   else {
> > >       tcp_hdr->cksum =
> > >          get_udptcp_checksum(l3_hdr, tcp_hdr,
> > >
> > > In other words - testpmd does not set TCP_CKSUM along with TCP_SEG
> > > therefore using testpmd with TAP/TSO will result in TCP segments with 0
> > (incorrect) TCP checksums.
> > >
> > > In addition - please note the comments in lib/librte_mbuf/rte_mbuf.h
> > > which specify that PKT_TX_TCP_SEG flag implies the PKT_TX_TCP_CKSUM
> > > (hence it is not required to be explicitly set by the caller)
> > >
> > > /**
> > > * TCP segmentation offload. To enable this offload feature for a
> > > * packet to be transmitted on hardware supporting TSO:
> > > *  - set the PKT_TX_TCP_SEG flag in mbuf->ol_flags (this flag implies
> > > *    PKT_TX_TCP_CKSUM)
> > > ...
> > >
> > > > Add Konstantin for more suggestions.
> > > >
> > > > Thanks,
> > > > Jiayu
> > > >
> > > > > -----Original Message-----
> > > > > From: Ophir Munk [mailto:ophirmu at mellanox.com]
> > > > > Sent: Sunday, April 22, 2018 10:21 PM
> > > > > To: dev at dpdk.org; Hu, Jiayu <jiayu.hu at intel.com>
> > > > > Cc: Thomas Monjalon <thomas at monjalon.net>; Olga Shern
> > > > > <olgas at mellanox.com>; Pascal Mazon <pascal.mazon at 6wind.com>;
> > > > Ophir
> > > > > Munk <ophirmu at mellanox.com>; stable at dpdk.org
> > > > > Subject: [PATCH v1] gso: fix marking TCP checksum flag in TCP
> > > > > segments
> > > > >
> > > > > Large TCP packets which are marked with PKT_TX_TCP_SEG flag are
> > > > > segmented and the flag is cleared in the resulting segments,
> > > > > however, the segments checksum is not updated. It is therefore
> > > > > required to set the PKT_TX_TCP_CKSUM flag in each TCP segment in
> > > > > order to mark for the sending driver the need to update the TCP
> > > > > checksum before transmitting the segment.
> > > > >
> > > > > Fixes: 119583797b6a ("gso: support TCP/IPv4 GSO")
> > > > > Cc: stable at dpdk.org
> > > > >
> > > > > Signed-off-by: Ophir Munk <ophirmu at mellanox.com>
> > > > > ---
> > > > >  lib/librte_gso/rte_gso.c | 2 ++
> > > > >  1 file changed, 2 insertions(+)
> > > > >
> > > > > diff --git a/lib/librte_gso/rte_gso.c b/lib/librte_gso/rte_gso.c
> > > > > index a44e3d4..e9ce9ce 100644
> > > > > --- a/lib/librte_gso/rte_gso.c
> > > > > +++ b/lib/librte_gso/rte_gso.c
> > > > > @@ -50,12 +50,14 @@ rte_gso_segment(struct rte_mbuf *pkt,
> > > > >  			((IS_IPV4_GRE_TCP4(pkt->ol_flags) &&
> > > > >  			 (gso_ctx->gso_types &
> > > > > DEV_TX_OFFLOAD_GRE_TNL_TSO)))) {
> > > > >  		pkt->ol_flags &= (~PKT_TX_TCP_SEG);
> > > > > +		pkt->ol_flags |= PKT_TX_TCP_CKSUM;
> > > > >  		ret = gso_tunnel_tcp4_segment(pkt, gso_size, ipid_delta,
> > > > >  				direct_pool, indirect_pool,
> > > > >  				pkts_out, nb_pkts_out);
> > > > >  	} else if (IS_IPV4_TCP(pkt->ol_flags) &&
> > > > >  			(gso_ctx->gso_types &
> > > > > DEV_TX_OFFLOAD_TCP_TSO)) {
> > > > >  		pkt->ol_flags &= (~PKT_TX_TCP_SEG);
> > > > > +		pkt->ol_flags |= PKT_TX_TCP_CKSUM;
> > > > >  		ret = gso_tcp4_segment(pkt, gso_size, ipid_delta,
> > > > >  				direct_pool, indirect_pool,
> > > > >  				pkts_out, nb_pkts_out);
> > > > > --
> > > > > 2.7.4



More information about the dev mailing list