MIME-Version: 1.0
References: <04e27096-9ace-07eb-aa51-1663714a586d@nbd.name>
 <167475990764.1934330.11960904198087757911.stgit@localhost.localdomain>
 <cde24ed8-1852-ce93-69f3-ff378731f52c@huawei.com> <20230127212646.4cfeb475@kernel.org>
 <CANn89iKgZU4Q+THXupzZi4hETuKuCOvOB=iHpp5JzQTNv_Fg_A@mail.gmail.com> <ec534eacabf5c859930eb5ca7f417f7f01197d24.camel@redhat.com>
In-Reply-To: <ec534eacabf5c859930eb5ca7f417f7f01197d24.camel@redhat.com>
From:   Alexander Duyck <alexander.duyck@gmail.com>
Date:   Mon, 30 Jan 2023 08:17:58 -0800
Message-ID: <CAKgT0UcwrQ21F2mgaLK2ruWUsRbiuSd=T=V=e1P4GUpAfbxPgQ@mail.gmail.com>
Subject: Re: [net PATCH] skb: Do mix page pool and page referenced frags in GRO
To:     Paolo Abeni <pabeni@redhat.com>
Cc:     Eric Dumazet <edumazet@google.com>,
        Jakub Kicinski <kuba@kernel.org>,
        Yunsheng Lin <linyunsheng@huawei.com>, nbd@nbd.name,
        davem@davemloft.net, hawk@kernel.org, ilias.apalodimas@linaro.org,
        linux-kernel@vger.kernel.org, lorenzo@kernel.org,
        netdev@vger.kernel.org
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Mon, Jan 30, 2023 at 12:50 AM Paolo Abeni <pabeni@redhat.com> wrote:
>
> On Sat, 2023-01-28 at 08:08 +0100, Eric Dumazet wrote:
> > On Sat, Jan 28, 2023 at 6:26 AM Jakub Kicinski <kuba@kernel.org> wrote:
> > >
> > > On Sat, 28 Jan 2023 10:37:47 +0800 Yunsheng Lin wrote:
> > > > If we are not allowing gro for the above case, setting NAPI_GRO_CB(p)->flush
> > > > to 1 in gro_list_prepare() seems to be making more sense so that the above
> > > > case has the same handling as skb_has_frag_list() handling?
> > > > https://elixir.bootlin.com/linux/v6.2-rc4/source/net/core/gro.c#L503
> > > >
> > > > As it seems to avoid some unnecessary operation according to comment
> > > > in tcp4_gro_receive():
> > > > https://elixir.bootlin.com/linux/v6.2-rc4/source/net/ipv4/tcp_offload.c#L322
> > >
> > > The frag_list case can be determined with just the input skb.
> > > For pp_recycle we need to compare input skb's pp_recycle with
> > > the pp_recycle of the skb already held by GRO.
> > >
> > > I'll hold off with applying a bit longer tho, in case Eric
> > > wants to chime in with an ack or opinion.
> >
> > We can say that we are adding in the fast path an expensive check
> > about an unlikely condition.
> >
> > GRO is by far the most expensive component in our stack.
>
> Slightly related to the above: currently the GRO engine performs the
> skb metadata check for every packet. My understanding is that even with
> XDP enabled and ebpf running on the given packet, the skb should
> usually have meta_len == 0.
>
> What about setting 'skb->slow_gro' together with meta_len and moving
> the skb_metadata_differs() check under the slow_gro guard?

Makes sense to me, especially since we have to do a pointer chase to
get the metadata length out of the shared info.

Looking at the code one thing I was wondering about is if we should be
flagging frames where one is slow_gro and one is not as having a diff
and just skipping the checks since we know the slow_gro checks are
expensive and if they differ based on that flag odds are one will have
a field present that the other doesn't.