Linux Virtualization - May 2023 - [RFC PATCH net 2/3] virtio-net: allow usage of vrings smaller than MAX_SKB_FRAGS + 2

On Sun, Apr 30, 2023 at 06:54:08PM +0000, Alvaro Karsz
wrote:
> > > At the moment, if a network device uses vrings with less than
> > > MAX_SKB_FRAGS + 2 entries, the device won't be functional.
> > >
> > > The following condition vq->num_free >= 2 + MAX_SKB_FRAGS
will always
> > > evaluate to false, leading to TX timeouts.
> > >
> > > This patch introduces a new variable, single_pkt_max_descs, that
holds
> > > the max number of descriptors we may need to handle a single
packet.
> > >
> > > This patch also detects the small vring during probe, blocks some
> > > features that can't be used with small vrings, and fails
probe,
> > > leading to a reset and features re-negotiation.
> > >
> > > Features that can't be used with small vrings:
> > > GRO features (VIRTIO_NET_F_GUEST_*):
> > > When we use small vrings, we may not have enough entries in the
ring to
> > > chain page size buffers and form a 64K buffer.
> > > So we may need to allocate 64k of continuous memory, which may be
too
> > > much when the system is stressed.
> > >
> > > This patch also fixes the MTU size in small vring cases to be up
to the
> > > default one, 1500B.
> > 
> > and then it should clear VIRTIO_NET_F_MTU?
> > 
> 
> Following [1], I was thinking to accept the feature and a let the device
figure out that it can't transmit a big packet, since the RX buffers are not
big enough (without VIRTIO_NET_F_MRG_RXBUF).
> But, I think that we may need to block the MTU feature after all.
> Quoting the spec:
> 
> A driver SHOULD negotiate VIRTIO_NET_F_MTU if the device offers it.
> If the driver negotiates VIRTIO_NET_F_MTU, it MUST supply enough receive
buffers to receive at least one receive packet of size mtu (plus low level
ethernet header length) with gso_type NONE or ECN.
> 
> So, if VIRTIO_NET_F_MTU is negotiated, we MUST supply enough receive
buffers.
> So I think that blocking VIRTIO_NET_F_MTU  should be the way to go, If mtu
> 1500.
> 
> [1] https://lore.kernel.org/lkml/20230417031052-mutt-send-email-mst at
kernel.org/

First up to 4k should not be a problem. Even jumbo frames e.g. 9k
is highly likely to succeed. And a probe time which is often boot
even 64k isn't a problem ...

Hmm. We could allocate large buffers at probe time. Reuse them and
copy data over.

IOW reusing  GOOD_COPY_LEN flow for this case.  Not yet sure how I feel
about this. OTOH it removes the need for the whole feature blocking
approach, does it not?
WDYT?

> > > +     /* How many ring descriptors we may need to transmit a
single packet */
> > > +     u16 single_pkt_max_descs;
> > > +
> > > +     /* Do we have virtqueues with small vrings? */
> > > +     bool svring;
> > > +
> > >       /* CPU hotplug instances for online & dead */
> > >       struct hlist_node node;
> > >       struct hlist_node node_dead;
> > 
> > worth checking that all these layout changes don't push useful
things to
> > a different cache line. can you add that analysis?
> > 
> 
> Good point.
> I think that we can just move these to the bottom of the struct.
> 
> > 
> > I see confusiong here wrt whether some rings are "small"?
all of them?
> > some rx rings? some tx rings? names should make it clear.
> 
> The small vring is a device attribute, not a vq attribute. It blocks
features, which affects the entire device.
> Maybe we can call it "small vring mode".
> 
> > also do we really need bool svring? can't we just check
single_pkt_max_descs
> > all the time?
> > 
> 
> We can work without the bool, we could always check if single_pkt_max_descs
!= MAX_SKB_FRAGS + 2.
> It doesn't really matter to me, I was thinking it may be more readable
this way.
> 
> > > +static bool virtnet_uses_svring(struct virtnet_info *vi)
> > > +{
> > > +     u32 i;
> > > +
> > > +     /* If a transmit/receive virtqueue is small,
> > > +      * we cannot handle fragmented packets.
> > > +      */
> > > +     for (i = 0; i < vi->max_queue_pairs; i++) {
> > > +             if
(IS_SMALL_VRING(virtqueue_get_vring_size(vi->sq[i].vq)) ||
> > > +                
IS_SMALL_VRING(virtqueue_get_vring_size(vi->rq[i].vq)))
> > > +                     return true;
> > > +     }
> > > +
> > > +     return false;
> > > +}
> > 
> > I see even if only some rings are too small we force everything to use
> > small ones. Wouldn't it be better to just disable small ones in
this
> > case? That would not need a reset.
> > 
> 
> I'm not sure. It may complicate things.
> 
> What if all TX vqs are small?
> What if all RX vqs are small?
> What if we end up with an unbalanced number of TX vqs and RX vqs? is this
allowed by the spec?
> What if we end up disabling the RX default vq (receiveq1)?
> 
> I guess we could do it, after checking some conditions.
> Maybe we can do it in a follow up patch?
> Do you think it's important for it to be included since day 1?
> 
> I think that the question is: what's more important, to use all the vqs
while blocking some features, or to use part of the vqs without blocking
features?
> 
> > > +
> > > +/* Function returns the number of features it blocked */
> > 
> > We don't need the # though. Make it bool?
> > 
> 
> Sure.
>

> First up to 4k should not be a problem. Even jumbo frames e.g. 9k
> is highly likely to succeed. And a probe time which is often boot
> even 64k isn't a problem ...
> 
> Hmm. We could allocate large buffers at probe time. Reuse them and
> copy data over.
> 
> IOW reusing  GOOD_COPY_LEN flow for this case.  Not yet sure how I feel
> about this. OTOH it removes the need for the whole feature blocking
> approach, does it not?
> WDYT?
> 
It could work..

In order to remove completely the feature blocking approach, we'll need to
let the control commands fail (as you mentioned in the other patch).
I'm not sure I like it, it means many warnings from virtnet..
And it means accepting features that we know for sure that are not going to
work.