On Fri, Jun 21, 2024 at 11:41 AM Daniel Borkmann <daniel@xxxxxxxxxxxxx> wrote:
>
> On 6/21/24 6:00 PM, Yan Zhai wrote:
> > On Fri, Jun 21, 2024 at 8:13 AM Daniel Borkmann <daniel@xxxxxxxxxxxxx> wrote:
> >> On 6/21/24 2:15 PM, Willem de Bruijn wrote:
> >>> Yan Zhai wrote:
> >>>> Software GRO is currently controlled by a single switch, i.e.
> >>>>
> >>>> ethtool -K dev gro on|off
> >>>>
> >>>> However, this is not always desired. When GRO is enabled, even if the
> >>>> kernel cannot GRO certain traffic, it has to run through the GRO receive
> >>>> handlers with no benefit.
> >>>>
> >>>> There are also scenarios that turning off GRO is a requirement. For
> >>>> example, our production environment has a scenario that a TC egress hook
> >>>> may add multiple encapsulation headers to forwarded skbs for load
> >>>> balancing and isolation purpose. The encapsulation is implemented via
> >>>> BPF. But the problem arises then: there is no way to properly offload a
> >>>> double-encapsulated packet, since skb only has network_header and
> >>>> inner_network_header to track one layer of encapsulation, but not two.
> >>>> On the other hand, not all the traffic through this device needs double
> >>>> encapsulation. But we have to turn off GRO completely for any ingress
> >>>> device as a result.
> >>>>
> >>>> Introduce a bit on skb so that GRO engine can be notified to skip GRO on
> >>>> this skb, rather than having to be 0-or-1 for all traffic.
> >>>>
> >>>> Signed-off-by: Yan Zhai <yan@xxxxxxxxxxxxxx>
> >>>> ---
> >>>> include/linux/netdevice.h | 9 +++++++--
> >>>> include/linux/skbuff.h | 10 ++++++++++
> >>>> net/Kconfig | 10 ++++++++++
> >>>> net/core/gro.c | 2 +-
> >>>> net/core/gro_cells.c | 2 +-
> >>>> net/core/skbuff.c | 4 ++++
> >>>> 6 files changed, 33 insertions(+), 4 deletions(-)
> >>>>
> >>>> diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h
> >>>> index c83b390191d4..2ca0870b1221 100644
> >>>> --- a/include/linux/netdevice.h
> >>>> +++ b/include/linux/netdevice.h
> >>>> @@ -2415,11 +2415,16 @@ struct net_device {
> >>>> ((dev)->devlink_port = (port)); \
> >>>> })
> >>>>
> >>>> -static inline bool netif_elide_gro(const struct net_device *dev)
> >>>> +static inline bool netif_elide_gro(const struct sk_buff *skb)
> >>>> {
> >>>> - if (!(dev->features & NETIF_F_GRO) || dev->xdp_prog)
> >>>> + if (!(skb->dev->features & NETIF_F_GRO) || skb->dev->xdp_prog)
> >>>> return true;
> >>>> +
> >>>> +#ifdef CONFIG_SKB_GRO_CONTROL
> >>>> + return skb->gro_disabled;
> >>>> +#else
> >>>> return false;
> >>>> +#endif
> >>>
> >>> Yet more branches in the hot path.
> >>>
> >>> Compile time configurability does not help, as that will be
> >>> enabled by distros.
> >>>
> >>> For a fairly niche use case. Where functionality of GRO already
> >>> works. So just a performance for a very rare case at the cost of a
> >>> regression in the common case. A small regression perhaps, but death
> >>> by a thousand cuts.
> >>
> >> Mentioning it here b/c it perhaps fits in this context, longer time ago
> >> there was the idea mentioned to have BPF operating as GRO engine which
> >> might also help to reduce attack surface by only having to handle packets
> >> of interest for the concrete production use case. Perhaps here meta data
> >> buffer could be used to pass a notification from XDP to exit early w/o
> >> aggregation.
> >
> > Metadata is in fact one of our interests as well. We discussed using
> > metadata instead of a skb bit to carry this information internally.
> > Since metadata is opaque atm so it seems the only option is to have a
> > GRO control hook before napi_gro_receive, and let BPF decide
> > netif_receive_skb or napi_gro_receive (echo what Paolo said). With BPF
> > it could indeed be more flexible, but the cons is that it could be
> > even more slower than taking a bit on skb. I am actually open to
> > either approach, as long as it gives us more control on when to enable
> > GRO :)
>
> Oh wait, one thing that just came to mind.. have you tried u64 per-CPU
> counter map in XDP? For packets which should not be GRO-aggregated you
> add count++ into the meta data area, and this forces GRO to not aggregate
> since meta data that needs to be transported to tc BPF layer mismatches
> (and therefore the contract/intent is that tc BPF needs to see the different
> meta data passed to it).
>
Very very sorry to resendx2 :( Not sure why my laptop also decided to
switch on html... I removed CCs from the message hopefully it reduces
some noises...
We did this before accidentally (we put a timestamp for debugging
purposes in metadata) and this actually caused about 20% of OoO for
TCP in production: all PSH packets are reordered. GRO does not fire
the packet to the upper layer when a diff in metadata is found for a
non-PSH packet, instead it is queued as a “new flow” on the GRO list
and waits for flushing. When a PSH packet arrives, its semantic is to
flush this packet immediately and thus precedes earlier packets of the
same flow.
The artifact of this behavior can also cause latency increase and hash
degradation since the mismatch does not result in flushing, it results
in extra queuing on the same hash list, until the list is flushed.
It’s another reason we want to disable GRO when we know metadata can
be set differently for tracing purposes (I didn’t mention this though
because it seems distracting).
Thanks
Yan
> Thanks,
> Daniel
