Date:   Wed, 8 Feb 2023 17:27:57 +0100
From:   Maciej Fijalkowski <maciej.fijalkowski@...el.com>
To:     Alexander H Duyck <alexander.duyck@...il.com>
CC:     Jason Xing <kerneljasonxing@...il.com>,
        <jesse.brandeburg@...el.com>, <anthony.l.nguyen@...el.com>,
        <davem@...emloft.net>, <edumazet@...gle.com>, <kuba@...nel.org>,
        <pabeni@...hat.com>, <richardcochran@...il.com>, <ast@...nel.org>,
        <daniel@...earbox.net>, <hawk@...nel.org>,
        <john.fastabend@...il.com>, <alexandr.lobakin@...el.com>,
        <intel-wired-lan@...ts.osuosl.org>, <netdev@...r.kernel.org>,
        <linux-kernel@...r.kernel.org>, <bpf@...r.kernel.org>,
        Jason Xing <kernelxing@...cent.com>
Subject: Re: [PATCH net v4 1/3] ixgbe: allow to increase MTU to 3K with XDP
 enabled

On Wed, Feb 08, 2023 at 07:37:57AM -0800, Alexander H Duyck wrote:
> On Wed, 2023-02-08 at 10:43 +0800, Jason Xing wrote:
> > From: Jason Xing <kernelxing@...cent.com>
> > 
> > Recently I encountered one case where I cannot increase the MTU size
> > directly from 1500 to a much bigger value with XDP enabled if the
> > server is equipped with IXGBE card, which happened on thousands of
> > servers in production environment. After appling the current patch,
> > we can set the maximum MTU size to 3K.
> > 
> > This patch follows the behavior of changing MTU as i40e/ice does.
> > 
> > Referrences:
> > [1] commit 23b44513c3e6 ("ice: allow 3k MTU for XDP")
> > [2] commit 0c8493d90b6b ("i40e: add XDP support for pass and drop actions")
> > 
> > Fixes: fabf1bce103a ("ixgbe: Prevent unsupported configurations with XDP")
> > Signed-off-by: Jason Xing <kernelxing@...cent.com>
> 
> This is based on the broken premise that w/ XDP we are using a 4K page.
> The ixgbe driver isn't using page pool and is therefore running on
> different limitations. The ixgbe driver is only using 2K slices of the
> 4K page. In addition that is reduced to 1.5K to allow for headroom and
> the shared info in the buffer.
> 
> Currently the only way a 3K buffer would work is if FCoE is enabled and
> in that case the driver is using order 1 pages and still using the
> split buffer approach.

Hey Alex, interesting, we based this on the following logic from
ixgbe_set_rx_buffer_len() I guess:

#if (PAGE_SIZE < 8192)
		if (adapter->flags2 & IXGBE_FLAG2_RSC_ENABLED)
			set_bit(__IXGBE_RX_3K_BUFFER, &rx_ring->state);

		if (IXGBE_2K_TOO_SMALL_WITH_PADDING ||
		    (max_frame > (ETH_FRAME_LEN + ETH_FCS_LEN)))
			set_bit(__IXGBE_RX_3K_BUFFER, &rx_ring->state);
#endif

so we assumed that ixgbe is no different than i40e/ice in these terms, but
we ignored whole overhead of LRO/RSC that ixgbe carries.

I am not actively working with ixgbe but I know that you were the main dev
of it, so without premature dive into the datasheet and codebase, are you
really sure that 3k mtu for XDP is a no go?

> 
> Changing the MTU to more than 1.5K will allow multi-buffer frames which
> would break things when you try to use XDP_REDIRECT or XDP_TX on frames
> over 1.5K in size. For things like XDP_PASS, XDP_DROP, and XDP_ABORT it
> should still work as long as you don't attempt to reach beyond the 1.5K
> boundary.
> 
> Until this driver supports XDP multi-buffer I don't think you can
> increase the MTU past 1.5K. If you are wanting a larger MTU you should
> look at enabling XDP multi-buffer and then just drop the XDP
> limitations entirely.
> 
> > ---
> > v4:
> > 1) use ':' instead of '-' for kdoc
> > 
> > v3:
> > 1) modify the titile and body message.
> > 
> > v2:
> > 1) change the commit message.
> > 2) modify the logic when changing MTU size suggested by Maciej and Alexander.
> > ---
> >  drivers/net/ethernet/intel/ixgbe/ixgbe_main.c | 25 ++++++++++++-------
> >  1 file changed, 16 insertions(+), 9 deletions(-)
> > 
> > diff --git a/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c b/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c
> > index ab8370c413f3..25ca329f7d3c 100644
> > --- a/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c
> > +++ b/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c
> > @@ -6777,6 +6777,18 @@ static void ixgbe_free_all_rx_resources(struct ixgbe_adapter *adapter)
> >  			ixgbe_free_rx_resources(adapter->rx_ring[i]);
> >  }
> >  
> > +/**
> > + * ixgbe_max_xdp_frame_size - returns the maximum allowed frame size for XDP
> > + * @adapter: device handle, pointer to adapter
> > + */
> > +static int ixgbe_max_xdp_frame_size(struct ixgbe_adapter *adapter)
> > +{
> > +	if (PAGE_SIZE >= 8192 || adapter->flags2 & IXGBE_FLAG2_RX_LEGACY)
> > +		return IXGBE_RXBUFFER_2K;
> > +	else
> > +		return IXGBE_RXBUFFER_3K;
> > +}
> > +
> 
> There is no difference in the buffer allocation approach for LEGACY vs
> non-legacy. The difference is if we are building the frame around the
> buffer using build_skb or we are adding it as a frag and then copying
> out the header.
> 
> >  /**
> >   * ixgbe_change_mtu - Change the Maximum Transfer Unit
> >   * @netdev: network interface device structure
> > @@ -6788,18 +6800,13 @@ static int ixgbe_change_mtu(struct net_device *netdev, int new_mtu)
> >  {
> >  	struct ixgbe_adapter *adapter = netdev_priv(netdev);
> >  
> > -	if (adapter->xdp_prog) {
> > +	if (ixgbe_enabled_xdp_adapter(adapter)) {
> >  		int new_frame_size = new_mtu + ETH_HLEN + ETH_FCS_LEN +
> >  				     VLAN_HLEN;
> > -		int i;
> > -
> > -		for (i = 0; i < adapter->num_rx_queues; i++) {
> > -			struct ixgbe_ring *ring = adapter->rx_ring[i];
> >  
> > -			if (new_frame_size > ixgbe_rx_bufsz(ring)) {
> > -				e_warn(probe, "Requested MTU size is not supported with XDP\n");
> > -				return -EINVAL;
> > -			}
> > +		if (new_frame_size > ixgbe_max_xdp_frame_size(adapter)) {
> > +			e_warn(probe, "Requested MTU size is not supported with XDP\n");
> > +			return -EINVAL;
> >  		}
> >  	}
> >  
> 

Powered by blists - more mailing lists