Received-SPF: pass (google.com: domain of linux-kernel+bounces-97123-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:4601:e00::3 as permitted sender) client-ip=2604:1380:4601:e00::3;
Precedence: bulk
MIME-Version: 1.0
References: <20240308115722.14671-1-gakula@marvell.com>
In-Reply-To: <20240308115722.14671-1-gakula@marvell.com>
From: Eric Dumazet <edumazet@google.com>
Date: Fri, 8 Mar 2024 14:58:31 +0100
Message-ID: <CANn89iJ2euG8SgmTpifRK2DV1N+PSPAgiSoZP-W+7YWE3Ygv6w@mail.gmail.com>
Subject: Re: [net PATCH] octeontx2-pf: Do not use HW TSO when gso_size < 16bytes
To: Geetha sowjanya <gakula@marvell.com>
Cc: netdev@vger.kernel.org, linux-kernel@vger.kernel.org, kuba@kernel.org, 
	davem@davemloft.net, pabeni@redhat.com, sgoutham@marvell.com, 
	sbhatta@marvell.com, hkelam@marvell.com
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

On Fri, Mar 8, 2024 at 12:57=E2=80=AFPM Geetha sowjanya <gakula@marvell.com=
> wrote:
>
> Hardware doesn't support packet segmentation when segment size
> is < 16 bytes. Hence add an additional check and use SW
> segmentation in such case.
>
> Fixes: 86d7476078b8 ("octeontx2-pf: TCP segmentation offload support").
> Signed-off-by: Geetha sowjanya <gakula@marvell.com>
> ---
>  drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c | 7 +++++++
>  1 file changed, 7 insertions(+)
>
> diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c b/dri=
vers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c
> index f828d32737af..2ac56abb3a0e 100644
> --- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c
> +++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c
> @@ -967,6 +967,13 @@ static bool is_hw_tso_supported(struct otx2_nic *pfv=
f,
>  {
>         int payload_len, last_seg_size;
>
> +       /* Due to hw issue segment size less than 16 bytes
> +        * are not supported. Hence do not offload such TSO
> +        * segments.
> +        */
> +       if (skb_shinfo(skb)->gso_size < 16)
> +               return false;
> +
>         if (test_bit(HW_TSO, &pfvf->hw.cap_flag))
>                 return true;

How is this driver doing SW segmentation at this stage ?

You might perform this check in ndo_features_check() instead ?

otx2_sq_append_skb() is also forcing a linearization if
skb_shinfo(skb)->nr_frags >=3D OTX2_MAX_FRAGS_IN_SQE

This is quite bad, this one definitely should use ndo_features_check()
to mask NETIF_F_GSO_MASK for GSO packets.

Look at typhoon_features_check() for an example.