Re: [PATCH net-next 2/2] net: stmmac: simplify GSO/TSO test in stmmac_xmit()

["Russell King (Oracle)" <linux@armlinux.org.uk>]

On Sat, Mar 28, 2026 at 05:25:58PM +0000, Russell King (Oracle) wrote:
> On Sat, Mar 28, 2026 at 09:31:46AM +0000, Russell King (Oracle) wrote:
> > On Sat, Mar 28, 2026 at 08:24:37AM +0000, Russell King (Oracle) wrote:
> > > On Fri, Mar 27, 2026 at 09:40:09AM +0000, Russell King (Oracle) wrote:
> > > > The test in stmmac_xmit() to see whether we should pass the skbuff to
> > > > stmmac_tso_xmit() is more complex than it needs to be. This test can
> > > > be simplified by storing the mask of GSO types that we will pass, and
> > > > setting it according to the enabled features.
> > > > 
> > > > Note that "tso" is a mis-nomer since commit b776620651a1 ("net:
> > > > stmmac: Implement UDP Segmentation Offload"). Also note that this
> > > > commit controls both via the TSO feature. We preserve this behaviour
> > > > in this commit.
> > > > 
> > > > Also, this commit unconditionally accessed skb_shinfo(skb)->gso_type
> > > > for all frames, even when skb_is_gso() was false. This access is
> > > > eliminated.
> > > > 
> > > > Signed-off-by: Russell King (Oracle) <rmk+kernel@armlinux.org.uk>
> > > 
> > > AI review of this patch regurgitates Jakub's point that was discussed.
> > > 
> > > > @@ -3700,7 +3700,7 @@ static int stmmac_hw_setup(struct net_device *dev)
> > > >  	stmmac_set_rings_length(priv);
> > > >  
> > > >  	/* Enable TSO */
> > > > -	if (priv->tso) {
> > > > +	if (priv->gso_enabled_types) {
> > > >  		for (chan = 0; chan < tx_cnt; chan++) {
> > > >  			struct stmmac_tx_queue *tx_q = &priv->dma_conf.tx_queue[chan];
> > > >  
> > > 
> > > ...
> > > 
> > > > @@ -7828,7 +7834,7 @@ static int __stmmac_dvr_probe(struct device *device,
> > > >  		ndev->hw_features |= NETIF_F_TSO | NETIF_F_TSO6;
> > > >  		if (priv->plat->core_type == DWMAC_CORE_GMAC4)
> > > >  			ndev->hw_features |= NETIF_F_GSO_UDP_L4;
> > > > -		priv->tso = true;
> > > > +		stmmac_set_gso_types(priv, true);
> > > 
> > > Clearly, the issue it is regurgitating has been there for a long time
> > > and isn't a new issue introduced by this patch.
> > > 
> > > AI needs to stop doing this, because it is encouraging multiple changes
> > > in a single patch, which is against the normal kernel process.
> > > 
> > > As already pointed out, there are multiple issues with stmmac TSO
> > > support, particularly with glue drivers that enable TSO on some
> > > queues/channels and not others, since netdev core TSO support is
> > > global across all channels.
> > > 
> > > So, won't the AI response in this patch - it's just another pre-
> > > existing issue that needs fixing in a separate patch.
> > 
> > Looking at the TSO vs TBS issue (which precludes the use of TSO on a
> > channel in stmmac) I can't find an obvious reason for this in the
> > available documentation. However, unfortunately, iMX8MP doesn't support
> > TSO, so the TSO bits are elided there, but does support TBS (needing
> > enhanced descriptors to be enabled). STM32MP151 on the other hand
> > supports TSO but not TBS, and thus fails to mention anything about
> > enhanced descriptors or TBS.
> > 
> > When stmmac_enable_tbs() enables TBS, it isn't actually enabling a
> > feature specific bit, but switching the channel to use enhanced
> > descriptor format. This format extends the basic descriptors by
> > placing four extra 32-bit words before the basic descriptor.
> > 
> > Looking at the enhanced normal descriptor format for TDES3, it
> > indicates that the format includes bit 18 in the control field, which
> > is the TSE bit (TCP segmentation enable for this packet.) So, it seems
> > it's not a limitation of the descriptor format.
> > 
> > So, either "TSO and TBS cannot co-exist" is incorrect, or there is a
> > hardware limitation that isn't documented between these two manuals.
> > 
> > One other interesting point is that stmmac_tso_xmit() seems to
> > handle the case where TSO and TBS are enabled on the channel:
> > 
> >                 if (tx_q->tbs & STMMAC_TBS_AVAIL)
> >                         mss_desc = &tx_q->dma_entx[tx_q->cur_tx].basic;
> >                 else
> >                         mss_desc = &tx_q->dma_tx[tx_q->cur_tx];
> > 
> >                 stmmac_set_mss(priv, mss_desc, mss);
> > ...
> >         if (tx_q->tbs & STMMAC_TBS_AVAIL)
> >                 desc = &tx_q->dma_entx[first_entry].basic;
> >         else
> >                 desc = &tx_q->dma_tx[first_entry];
> >         first = desc;
> > 
> > etc.
> > 
> > Avoiding enabling TSO for a TBS channel was added by this commit:
> > 
> > commit 5e6038b88a5718910dd74b949946d9d9cee9a041
> > Author: Ong Boon Leong <boon.leong.ong@intel.com>
> > Date:   Wed Apr 21 17:11:49 2021 +0800
> > 
> >     net: stmmac: fix TSO and TBS feature enabling during driver open
> > 
> >     TSO and TBS cannot co-exist and current implementation requires two
> >     fixes:
> > 
> >      1) stmmac_open() does not need to call stmmac_enable_tbs() because
> >         the MAC is reset in stmmac_init_dma_engine() anyway.
> >      2) Inside stmmac_hw_setup(), we should call stmmac_enable_tso() for
> >         TX Q that is _not_ configured for TBS.
> > 
> >     Fixes: 579a25a854d4 ("net: stmmac: Initial support for TBS")
> >     Signed-off-by: Ong Boon Leong <boon.leong.ong@intel.com>
> >     Signed-off-by: David S. Miller <davem@davemloft.net>
> > 
> > which doesn't really explain the background, and leaves all the TBS
> > cruft in the TSO transmit path (nothing like properly updating the
> > driver, eh? No wonder stmmac is such a mess!)
> 
> The more I look at this, the more I'm convinced this commit is
> incorrect, even if it is the case that the hardware doesn't support
> TSO and TBS together.
> 
> When TSO is enabled (NETIF_F_TSO set in the netif's features) then
> the core net layer can submit skbuffs that need to be processed using
> TSO.
> 
> If such a skbuff hits a channel that has TSO disabled (because the
> above commit caused:
> 
> 	stmmac_enable_tso(priv, priv->ioaddr, 1, chan);
> 
> not to be called) then the TSE bit in the transmit control register
> will not be set, thereby disabling TSO on this particular channel.
> 
> However, stmmac_xmit() will still call through to stmmac_tso_xmit()
> which will dutifully populate the transmit ring with descriptors that
> assume TSE has been set in the transmit control register.
> 
> It seems to me _that_ is even more broken than "the hardware doesn't
> support TSO and TBS together" - I have no idea what the stmmac hardware
> does if it encounters descriptors with TSE set but TSE is disabled in
> the transmit control register. My guess would be it would ignore the
> TSE bit in the descriptor and assume that it's one very large packet
> to be sent - and either error out because it's longer than the
> maximum the hardware can support or it will just try to transmit it
> anyway.

Okay, I've found a statement in the stm32mp25xx documentation which
backs up Intel's commit, but that commit is still wrong because it
only half does the job.

However, I think I now have a solution - implementing
.ndo_features_check() which will mask out the NETIF_F_GSO_MASK
features if either the header length is greater than 1023 (the
hardware maximum) or the queue (as returned by
skb_get_queue_mapping(skb)) is for a queue which has TBS available,
and thus has TSO disabled.

I'm surprised we haven't had reports of brokenness in this area.

I think it's time to re-shuffle these patches (plus I think there's
a bit more scope to clean up some of the TSO code which would mean
placing stmmac_set_gso_types() in a different location to keep
all the TSO-related code together. I'll mark this series as
superseded once I have its replacement ready.

-- 
RMK's Patch system: https://www.armlinux.org.uk/developer/patches/
FTTP is here! 80Mbps down 10Mbps up. Decent connectivity at last!