>From: Hu, Jiayu
>Sent: Thursday, September 14, 2017 7:07 AM
>To: Ananyev, Konstantin <konstantin.anan...@intel.com>
>Cc: dev@dpdk.org; Kavanagh, Mark B <mark.b.kavan...@intel.com>; Tan, Jianfeng
><jianfeng....@intel.com>
>Subject: Re: [PATCH v3 2/5] gso: add TCP/IPv4 GSO support
>
>Hi Konstantin,
>
>On Thu, Sep 14, 2017 at 06:10:37AM +0800, Ananyev, Konstantin wrote:
>>
>> Hi Jiayu,
>>
>> > >
>> > >
>> > > > -----Original Message-----
>> > > > From: Ananyev, Konstantin
>> > > > Sent: Tuesday, September 12, 2017 12:18 PM
>> > > > To: Hu, Jiayu <jiayu...@intel.com>; dev@dpdk.org
>> > > > Cc: Kavanagh, Mark B <mark.b.kavan...@intel.com>; Tan, Jianfeng
><jianfeng....@intel.com>
>> > > > Subject: RE: [PATCH v3 2/5] gso: add TCP/IPv4 GSO support
>> > > >
>> > > > > result, when all of its GSOed segments are freed, the packet is
>freed
>> > > > > automatically.
>> > > > > diff --git a/lib/librte_gso/rte_gso.c b/lib/librte_gso/rte_gso.c
>> > > > > index dda50ee..95f6ea6 100644
>> > > > > --- a/lib/librte_gso/rte_gso.c
>> > > > > +++ b/lib/librte_gso/rte_gso.c
>> > > > > @@ -33,18 +33,53 @@
>> > > > >
>> > > > > #include <errno.h>
>> > > > >
>> > > > > +#include <rte_log.h>
>> > > > > +
>> > > > > #include "rte_gso.h"
>> > > > > +#include "gso_common.h"
>> > > > > +#include "gso_tcp4.h"
>> > > > >
>> > > > > int
>> > > > > rte_gso_segment(struct rte_mbuf *pkt,
>> > > > > - struct rte_gso_ctx gso_ctx __rte_unused,
>> > > > > + struct rte_gso_ctx gso_ctx,
>> > > > > struct rte_mbuf **pkts_out,
>> > > > > uint16_t nb_pkts_out)
>> > > > > {
>> > > > > + struct rte_mempool *direct_pool, *indirect_pool;
>> > > > > + struct rte_mbuf *pkt_seg;
>> > > > > + uint16_t gso_size;
>> > > > > + uint8_t ipid_delta;
>> > > > > + int ret = 1;
>> > > > > +
>> > > > > if (pkt == NULL || pkts_out == NULL || nb_pkts_out < 1)
>> > > > > return -EINVAL;
>> > > > >
>> > > > > - pkts_out[0] = pkt;
>> > > > > + if (gso_ctx.gso_size >= pkt->pkt_len ||
>> > > > > + (pkt->packet_type & gso_ctx.gso_types) !=
>> > > > > + pkt->packet_type) {
>> > > > > + pkts_out[0] = pkt;
>> > > > > + return ret;
>> > > > > + }
>> > > > > +
>> > > > > + direct_pool = gso_ctx.direct_pool;
>> > > > > + indirect_pool = gso_ctx.indirect_pool;
>> > > > > + gso_size = gso_ctx.gso_size;
>> > > > > + ipid_delta = gso_ctx.ipid_flag == RTE_GSO_IPID_INCREASE;
>> > > > > +
>> > > > > + if (is_ipv4_tcp(pkt->packet_type)) {
>> > > >
>> > > > Probably we need here:
>> > > > If (is_ipv4_tcp(pkt->packet_type) && (gso_ctx->gso_types &
>DEV_TX_OFFLOAD_TCP_TSO) != 0) {...
>> > >
>> > > Sorry, actually it probably should be:
>> > > If (pkt->ol_flags & (PKT_TX_TCP_SEG | PKT_TX_IPV4) == PKT_TX_IPV4 &&
>> > > (gso_ctx->gso_types & DEV_TX_OFFLOAD_TCP_TSO) != 0) {...
>> >
>> > I don't quite understand why the GSO library should be aware if the TSO
>> > flag is set or not. Applications can query device TSO capability before
>> > they call the GSO library. Do I misundertsand anything?
>> >
>> > Additionally, we don't need to check if the packet is a TCP/IPv4 packet
>here?
>>
>> Well, right now PMD we doesn't rely on ptype to figure out what type of
>packet and
>> what TX offload have to be performed.
>> Instead it looks at TX part of ol_flags, and
>> My thought was that as what we doing is actually TSO in SW, it would be good
>> to use the same API here too.
>> Also with that approach, by setting ol_flags properly user can use the same
>gso_ctx and still
>> specify what segmentation to perform on a per-packet basis.
>>
>> Alternative way is to rely on ptype to distinguish should segmentation be
>performed on that package or not.
>> The only advantage I see here is that if someone would like to add GSO for
>some new protocol,
>> he wouldn't need to introduce new TX flag value for mbuf.ol_flags.
>> Though he still would need to update TX_OFFLOAD_* capabilities and probably
>packet_type definitions.
>>
>> So from my perspective first variant (use HW TSO API) is more plausible.
>> Wonder what is your and Mark opinions here?
>
>In the first choice, you mean:
>the GSO library uses gso_ctx->gso_types and mbuf->ol_flags to call a specific
>GSO
>segmentation function (e.g. gso_tcp4_segment(), gso_tunnel_xxx()) for each
>input packet.
>Applications should parse the packet type, and set an exactly correct
>DEV_TX_OFFLOAD_*_TSO
>flag to gso_types and ol_flags according to the packet type. That is, the
>value of gso_types
>is on a per-packet basis. Using gso_ctx->gso_types and mbuf->ol_flags at the
>same time
>is because that DEV_TX_OFFLOAD_*_TSO only tells tunnelling type and the inner
>L4 type, and
>we need to know L3 type by ol_flags. With this design, HW segmentation and SW
>segmentation
>are indeed consistent.
>
>If I understand it correctly, applications need to set 'ol_flags =
>PKT_TX_IPV4' and
>'gso_types = DEV_TX_OFFLOAD_VXLAN_TNL_TSO' for a
>"ether+ipv4+udp+vxlan+ether+ipv4+
>tcp+payload" packet. But PKT_TX_IPV4 just present the inner L3 type for
>tunneled packet.
>How about the outer L3 type? Always assume the inner and the outer L3 type are
>the same?
Hi Jiayu,
If I'm not mistaken, I think what Konstantin is suggesting is as follows:
- The DEV_TX_OFFLOAD_*_TSO flags are currently used to describe a NIC's TSO
capabilities; the GSO capabilities may also be described using the same macros,
to provide a consistent view of segmentation capabilities across the HW and SW
implementations.
- As part of segmentation, it's still a case of checking the packet type, but
then setting the appropriate ol_flags in the mbuf, which the GSO library can
use to segment the packet.
Thanks,
Mark
>
>Jiayu
>> Konstantin
