Re: [PATCH v4 net-next 02/12] ipv6: add IFLA_GSO_IPV6_MAX_SIZE

[Jakub Kicinski <kuba@kernel.org>]

On Fri, 6 May 2022 15:16:21 -0700 Alexander Duyck wrote:
> On Fri, May 6, 2022 at 2:50 PM Eric Dumazet <edumazet@google.com> wrote:
> > gso_max_size can not exceed GSO_MAX_SIZE.
> > This will break many drivers.
> > I do not want to change hundreds of them.  
> 
> Most drivers will not be impacted because they cannot exceed
> tso_max_size. The tso_max_size is the limit, not GSO_MAX_SIZE. Last I
> knew this patch set is overwriting that value to increase it beyond
> the legacy limits.
> 
> Right now the check is:
> if (max_size > GSO_MAX_SIZE || max_size > dev->tso_max_size)
> 
> What I am suggesting is that tso_max_size be used as the only limit,
> which is already defaulted to cap out at TSO_LEGACY_MAX_SIZE. So just
> remove the "max_size > GSO_MAX_SIZE ||" portion of the call. Then when
> you call netif_set_tso_max_size in the driver to enable jumbograms you
> are good to set gso_max_size to something larger than the standard
> 65536.

TBH that was my expectation as well.

Drivers should not pay any attention to dev->gso_* any longer.

> > Look, we chose this implementation so that chances of breaking things
> > are very small.
> > I understand this is frustrating, but I suggest you take the
> > responsibility of breaking things,
> > and not add this on us.  
> 
> What I have been trying to point out is your patch set will break things.
> 
> For all those cases out there where people are using gso_max_size to
> limit things you just poked a hole in that for IPv6 cases. What I am
> suggesting is that we don't do that as it will be likely to trigger a
> number of problems for people.
> 
> The primary reason gso_max_size was added was because there are cases
> out there where doing too big of a TSO was breaking things. For
> devices that are being used for LSOv2 I highly doubt they need to
> worry about cases less than 65536. As such they can just max out at
> 65536 for all non-IPv6 traffic and instead use gso_max_size as the
> limit for the IPv6/TSO case.

Good point. GSO limit is expected to be a cap, so we shouldn't go above
it. At the same time nothing wrong with IPv4 continuing to generate 64k
GSOs after the user raises the limit.