From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mga04.intel.com (mga04.intel.com [192.55.52.120]) by dpdk.org (Postfix) with ESMTP id EBC862BEA; Tue, 24 Apr 2018 16:27:03 +0200 (CEST) X-Amp-Result: SKIPPED(no attachment in message) X-Amp-File-Uploaded: False Received: from orsmga007.jf.intel.com ([10.7.209.58]) by fmsmga104.fm.intel.com with ESMTP/TLS/DHE-RSA-AES256-GCM-SHA384; 24 Apr 2018 07:27:01 -0700 X-ExtLoop1: 1 X-IronPort-AV: E=Sophos;i="5.49,323,1520924400"; d="scan'208";a="35776909" Received: from irsmsx101.ger.corp.intel.com ([163.33.3.153]) by orsmga007.jf.intel.com with ESMTP; 24 Apr 2018 07:26:59 -0700 Received: from irsmsx102.ger.corp.intel.com ([169.254.2.164]) by IRSMSX101.ger.corp.intel.com ([169.254.1.118]) with mapi id 14.03.0319.002; Tue, 24 Apr 2018 15:26:58 +0100 From: "Ananyev, Konstantin" To: Ophir Munk , "Hu, Jiayu" , "dev@dpdk.org" CC: Thomas Monjalon , Olga Shern , Pascal Mazon , "stable@dpdk.org" Thread-Topic: [PATCH v1] gso: fix marking TCP checksum flag in TCP segments Thread-Index: AQHT2rl94CoPk27HYUCsaOLb2rnsLKQPm2oAgAAjyvD///4aAIAAGG2AgAAIEICAABVxoA== Date: Tue, 24 Apr 2018 14:26:58 +0000 Message-ID: <2601191342CEEE43887BDE71AB977258AEA53318@IRSMSX102.ger.corp.intel.com> References: <1524406859-29585-1-git-send-email-ophirmu@mellanox.com> <2601191342CEEE43887BDE71AB977258AEA5081C@IRSMSX102.ger.corp.intel.com> <2601191342CEEE43887BDE71AB977258AEA5221E@IRSMSX102.ger.corp.intel.com> In-Reply-To: Accept-Language: en-IE, en-US Content-Language: en-US X-MS-Has-Attach: X-MS-TNEF-Correlator: x-titus-metadata-40: eyJDYXRlZ29yeUxhYmVscyI6IiIsIk1ldGFkYXRhIjp7Im5zIjoiaHR0cDpcL1wvd3d3LnRpdHVzLmNvbVwvbnNcL0ludGVsMyIsImlkIjoiZjQzMTM5NmQtYjk2Mi00MzFiLTg4MTctNTA5MjA0MjBkYWQ5IiwicHJvcHMiOlt7Im4iOiJDVFBDbGFzc2lmaWNhdGlvbiIsInZhbHMiOlt7InZhbHVlIjoiQ1RQX05UIn1dfV19LCJTdWJqZWN0TGFiZWxzIjpbXSwiVE1DVmVyc2lvbiI6IjE2LjUuOS4zIiwiVHJ1c3RlZExhYmVsSGFzaCI6IjNVYnZ0ZVVtRTZKbTMzMW9uZURrR0laTEJCK0p6dFZZeEJmOEhlWEJ0ZHM9In0= x-ctpclassification: CTP_NT dlp-product: dlpe-windows dlp-version: 11.0.200.100 dlp-reaction: no-action x-originating-ip: [163.33.239.180] Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: quoted-printable MIME-Version: 1.0 Subject: Re: [dpdk-stable] [PATCH v1] gso: fix marking TCP checksum flag in TCP segments X-BeenThere: stable@dpdk.org X-Mailman-Version: 2.1.15 Precedence: list List-Id: patches for DPDK stable branches List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , X-List-Received-Date: Tue, 24 Apr 2018 14:27:07 -0000 > -----Original Message----- > From: Ophir Munk [mailto:ophirmu@mellanox.com] > Sent: Tuesday, April 24, 2018 2:42 PM > To: Ananyev, Konstantin ; Hu, Jiayu ; dev@dpdk.org > Cc: Thomas Monjalon ; Olga Shern ; Pascal Mazon ; > stable@dpdk.org > Subject: RE: [PATCH v1] gso: fix marking TCP checksum flag in TCP segment= s >=20 > Hi Konstantin, >=20 > > -----Original Message----- > > From: Ananyev, Konstantin [mailto:konstantin.ananyev@intel.com] > > Sent: Tuesday, April 24, 2018 3:31 PM > > To: Ophir Munk ; Hu, Jiayu ; > > dev@dpdk.org > > Cc: Thomas Monjalon ; Olga Shern > > ; Pascal Mazon ; > > stable@dpdk.org > > Subject: RE: [PATCH v1] gso: fix marking TCP checksum flag in TCP segme= nts > > > > > > Hi Ophir, > > > > > > > > Hi Konstantin, > > > Please see inline > > > > > > > -----Original Message----- > > > > From: Ananyev, Konstantin [mailto:konstantin.ananyev@intel.com] > > > > Sent: Tuesday, April 24, 2018 1:56 PM > > > > To: Ophir Munk ; Hu, Jiayu > > > > ; dev@dpdk.org > > > > Cc: Thomas Monjalon ; Olga Shern > > > > ; Pascal Mazon ; > > > > stable@dpdk.org > > > > Subject: RE: [PATCH v1] gso: fix marking TCP checksum flag in TCP > > > > segments > > > > > > > > > > > > > > > > > -----Original Message----- > > > > > From: Ophir Munk [mailto:ophirmu@mellanox.com] > > > > > Sent: Tuesday, April 24, 2018 10:44 AM > > > > > To: Hu, Jiayu ; dev@dpdk.org; Ananyev, > > > > > Konstantin > > > > > Cc: Thomas Monjalon ; Olga Shern > > > > > ; Pascal Mazon ; > > > > > stable@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@intel.com] > > > > > > Sent: Monday, April 23, 2018 7:14 AM > > > > > > To: Ophir Munk ; dev@dpdk.org; Ananyev, > > > > > > Konstantin > > > > > > Cc: Thomas Monjalon ; Olga Shern > > > > > > ; Pascal Mazon ; > > > > > > stable@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 b= ug. > > > > > > > > > > > > > > > > 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 segme= nt > > was created without a TCP checksum. > > > > > > > Ok, but new IP header was also created. And might be outer ip/udp (in c= ase > > of tunnel). > > If we go that way we'll have to set flags for each them. > > >=20 > No. The application should set the PKT_TX_IP_CKSUM or other flags which w= ill be forwarded by TAP PMD to GSO library which will clone > them in the new TCP segments. The only concern is about PKT_TX_TCP_SEG ve= rsus PTK_TX_TCP_CKSUM flags. > By dpdk design the PKT_TX_TCP_SEGS and PKT_TX_TCP_CKSUM flags are mutual = exclusive. They do not co-exist, but they are replaceable. > You can verify it in testpmd implementation, other PMDs and code document= ation (please note previous code snippets examples in this > thread). I still don't see how it differs. Let say user has a packet with already calculated ip nd tcp cksum. He passes it to rte_gso_segment(). If the segmentation did happen - he would need to recalculate both ip and t= cp cksum for each segment. Same for tunneling. So if we'll decide follow your logic - gso() would have to set all related = flags. >>From other side - if application doesn't have cksums calculated, and applic= ation plans to use HW offload for TCP cksum - nothing prevents it from setting PKT_TX_TCP_= CKSUM before calling rte_gso_segment(). >=20 > > > > 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 differen= t cases: > > > 1. Packet length equals to or is bigger than GSO size. In this case n= ew 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 exi= sting > > 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. Actually the question is why it happens that TAP PMD receives a packets with TCP_SEG flag on and valid TCP checksum already calculated? As you mentioned above TCP_SEG implies TCP_CKSUM for each segment? > > > > > > How do you suggest to avoid this scenario? > > > > Probably something like that: > > > > rc =3D rte_gso_segment(pkt_in, gso_ctx, pkts_out, nb_pkts_out); if (rc = =3D=3D 1 && > > pkt_in =3D=3D pkts_out[0] =3D=3D pkt_in) { > > /* no gso was performed */ > > } else { > > /* new packets, update ol_flags if needed */ } > > > > ? > > >=20 > Regarding the check: "pkts_out[0] =3D=3D pkt_in " - I would prefer an "A= PI approach" rather than relying on internal code specifics. >=20 > > 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) > > >=20 > I will gladly review any new enhancement. However, for 18.05, can we plea= se accept this patch as a practical solution for now? I am not sure that it is ok to do things that way - we'll introduce behavio= r change here. If there is no better way - it seems safer for 18.05 would be to use the wo= rkaround suggested above. >=20 > > 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: on= e > > > > > 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=3Dhttps%3A%2F%2Fd > > > > pd > > > > > > > > > > > k.org%2Fdev%2Fpatchwork%2Fpatch%2F38666%2F&data=3D02%7C01%7Coph > > > > irmu%40me > > > > > > > > > > > llanox.com%7C7455c8e31c7a4364bc7108d5a9d20008%7Ca652971c7d2e4d > > > > 9ba6a4d1 > > > > > > > > > > > 49256f461b%7C0%7C0%7C636601641779974217&sdata=3DCF7EvhXG%2BrH% > > > > 2BPiQEbvM0 > > > > > mC%2FSpqobneKaoV03j5VrSDw%3D&reserved=3D0 > > > > > > > > > > > https://emea01.safelinks.protection.outlook.com/?url=3Dhttps%3A%2F%2Fd > > > > pd > > > > > > > > > > > k.org%2Fdev%2Fpatchwork%2Fpatch%2F38667%2F&data=3D02%7C01%7Coph > > > > irmu%40me > > > > > > > > > > > llanox.com%7C7455c8e31c7a4364bc7108d5a9d20008%7Ca652971c7d2e4d > > > > 9ba6a4d1 > > > > > > > > > > > 49256f461b%7C0%7C0%7C636601641779974217&sdata=3Dj9WVIj%2FKq6EN > > > > WXu3mr5By1 > > > > > toSowU8bqJRGZ19SxiGoI%3D&reserved=3D0 > > > > > > > > > > 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 |=3D PKT_TX_TCP_SEG; // *** if TSO is applicabl= e - 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 |=3D PKT_TX_TCP_CKSUM; // *** PKT_TX_TCP_CKSU= M is > > > > marked only if TSO is not applicable *** > > > > > else { > > > > > tcp_hdr->cksum =3D > > > > > get_udptcp_checksum(l3_hdr, tcp_hdr, > > > > > > > > > > In other words - testpmd does not set TCP_CKSUM along with TCP_SE= G > > > > > 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 imp= lies > > > > > * PKT_TX_TCP_CKSUM) > > > > > ... > > > > > > > > > > > Add Konstantin for more suggestions. > > > > > > > > > > > > Thanks, > > > > > > Jiayu > > > > > > > > > > > > > -----Original Message----- > > > > > > > From: Ophir Munk [mailto:ophirmu@mellanox.com] > > > > > > > Sent: Sunday, April 22, 2018 10:21 PM > > > > > > > To: dev@dpdk.org; Hu, Jiayu > > > > > > > Cc: Thomas Monjalon ; Olga Shern > > > > > > > ; Pascal Mazon > > ; > > > > > > Ophir > > > > > > > Munk ; stable@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 i= s > > > > > > > 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@dpdk.org > > > > > > > > > > > > > > Signed-off-by: Ophir Munk > > > > > > > --- > > > > > > > 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 &=3D (~PKT_TX_TCP_SEG); > > > > > > > + pkt->ol_flags |=3D PKT_TX_TCP_CKSUM; > > > > > > > ret =3D 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 &=3D (~PKT_TX_TCP_SEG); > > > > > > > + pkt->ol_flags |=3D PKT_TX_TCP_CKSUM; > > > > > > > ret =3D gso_tcp4_segment(pkt, gso_size, ipid_delta, > > > > > > > direct_pool, indirect_pool, > > > > > > > pkts_out, nb_pkts_out); > > > > > > > -- > > > > > > > 2.7.4