Message-ID: <CAKgT0Ufj2D2p9r4eX24BVyTmJAHZ9Bi4UcHOwAsphQNzim3LBA@mail.gmail.com>
Date:   Sat, 17 Feb 2018 09:41:00 -0800
From:   Alexander Duyck <alexander.duyck@...il.com>
To:     Jakub Kicinski <kubakici@...pl>
Cc:     Sridhar Samudrala <sridhar.samudrala@...el.com>,
        "Michael S. Tsirkin" <mst@...hat.com>,
        Stephen Hemminger <stephen@...workplumber.org>,
        David Miller <davem@...emloft.net>,
        Netdev <netdev@...r.kernel.org>,
        virtualization@...ts.linux-foundation.org,
        "Brandeburg, Jesse" <jesse.brandeburg@...el.com>,
        "Duyck, Alexander H" <alexander.h.duyck@...el.com>,
        Jason Wang <jasowang@...hat.com>,
        Siwei Liu <loseweigh@...il.com>,
        Or Gerlitz <gerlitz.or@...il.com>
Subject: Re: [RFC PATCH v3 2/3] virtio_net: Extend virtio to use VF datapath
 when available

On Fri, Feb 16, 2018 at 7:04 PM, Jakub Kicinski <kubakici@...pl> wrote:
> On Fri, 16 Feb 2018 10:11:21 -0800, Sridhar Samudrala wrote:
>> This patch enables virtio_net to switch over to a VF datapath when a VF
>> netdev is present with the same MAC address. It allows live migration
>> of a VM with a direct attached VF without the need to setup a bond/team
>> between a VF and virtio net device in the guest.
>>
>> The hypervisor needs to enable only one datapath at any time so that
>> packets don't get looped back to the VM over the other datapath. When a VF
>> is plugged, the virtio datapath link state can be marked as down. The
>> hypervisor needs to unplug the VF device from the guest on the source host
>> and reset the MAC filter of the VF to initiate failover of datapath to
>> virtio before starting the migration. After the migration is completed,
>> the destination hypervisor sets the MAC filter on the VF and plugs it back
>> to the guest to switch over to VF datapath.
>>
>> When BACKUP feature is enabled, an additional netdev(bypass netdev) is
>> created that acts as a master device and tracks the state of the 2 lower
>> netdevs. The original virtio_net netdev is marked as 'backup' netdev and a
>> passthru device with the same MAC is registered as 'active' netdev.
>>
>> This patch is based on the discussion initiated by Jesse on this thread.
>> https://marc.info/?l=linux-virtualization&m=151189725224231&w=2
>>
>> Signed-off-by: Sridhar Samudrala <sridhar.samudrala@...el.com>
>> Signed-off-by: Alexander Duyck <alexander.h.duyck@...el.com>
>
>> +static void
>> +virtnet_bypass_get_stats(struct net_device *dev,
>> +                      struct rtnl_link_stats64 *stats)
>> +{
>> +     struct virtnet_bypass_info *vbi = netdev_priv(dev);
>> +     const struct rtnl_link_stats64 *new;
>> +     struct rtnl_link_stats64 temp;
>> +     struct net_device *child_netdev;
>> +
>> +     spin_lock(&vbi->stats_lock);
>> +     memcpy(stats, &vbi->bypass_stats, sizeof(*stats));
>> +
>> +     rcu_read_lock();
>> +
>> +     child_netdev = rcu_dereference(vbi->active_netdev);
>> +     if (child_netdev) {
>> +             new = dev_get_stats(child_netdev, &temp);
>> +             virtnet_bypass_fold_stats(stats, new, &vbi->active_stats);
>> +             memcpy(&vbi->active_stats, new, sizeof(*new));
>> +     }
>> +
>> +     child_netdev = rcu_dereference(vbi->backup_netdev);
>> +     if (child_netdev) {
>> +             new = dev_get_stats(child_netdev, &temp);
>> +             virtnet_bypass_fold_stats(stats, new, &vbi->backup_stats);
>> +             memcpy(&vbi->backup_stats, new, sizeof(*new));
>> +     }
>> +
>> +     rcu_read_unlock();
>> +
>> +     memcpy(&vbi->bypass_stats, stats, sizeof(*stats));
>> +     spin_unlock(&vbi->stats_lock);
>> +}
>> +
>> +static int virtnet_bypass_change_mtu(struct net_device *dev, int new_mtu)
>> +{
>> +     struct virtnet_bypass_info *vbi = netdev_priv(dev);
>> +     struct net_device *child_netdev;
>> +     int ret = 0;
>> +
>> +     child_netdev = rcu_dereference(vbi->active_netdev);
>> +     if (child_netdev) {
>> +             ret = dev_set_mtu(child_netdev, new_mtu);
>> +             if (ret)
>> +                     return ret;
>> +     }
>> +
>> +     child_netdev = rcu_dereference(vbi->backup_netdev);
>> +     if (child_netdev) {
>> +             ret = dev_set_mtu(child_netdev, new_mtu);
>> +             if (ret)
>> +                     netdev_err(child_netdev,
>> +                                "Unexpected failure to set mtu to %d\n",
>> +                                new_mtu);
>
> You should probably unwind if set fails on one of the legs.

Actually if we know that the backup is always going to be a virtio I
don't know if we even need to worry about the call failing. Last I
knew virtio_net doesn't implement ndo_change_mtu so I don't think it
is an issue. Unless a notifier blows up about it somewhere I don't
think there is anything that should prevent us from updating the MTU.

One interesting thing we may want to take a look at would be to tweak
the ordering of things based on if we are increasing or decreasing the
MTU. In the case of a increase we need to work from the bottom up, but
in the case of a decrease I wonder if we shouldn't be working from the
top down in order to guarantee we don't create an MTU mismatch
somewhere in the path.

>> +     }
>> +
>> +     dev->mtu = new_mtu;
>> +     return 0;
>> +}
>
> nit: stats, mtu, all those mundane things are implemented in team
>      already.  If we had this as kernel-internal team mode we wouldn't
>      have to reimplement them...  You probably did investigate that
>      option, for my edification, would you mind saying what the
>      challenges/downsides were?

So I tried working with the bonding driver to get down to what we
needed. The issue is there are so many controls and such exposed that
trying to pull them out and generate a simple bond became very
difficult to get done. It was just much easier to start over versus
trying to take an existing interface and pare it down to just what we
needed. What may make more sense is to in the future create either a
lib or some sort of file in net/core/ that we could consolidate the
core functionality of these type of devices into and leave the
user-space interfaces, debugfs, ioctls, and such out of and leave
those driver specific.

>> +static struct net_device *
>> +get_virtnet_bypass_bymac(struct net *net, const u8 *mac)
>> +{
>> +     struct net_device *dev;
>> +
>> +     ASSERT_RTNL();
>> +
>> +     for_each_netdev(net, dev) {
>> +             if (dev->netdev_ops != &virtnet_bypass_netdev_ops)
>> +                     continue;       /* not a virtnet_bypass device */
>
> Is there anything inherently wrong with enslaving another virtio dev
> now?  I was expecting something like a hash map to map MAC addr ->
> master and then one can check if dev is already enslaved to that master.
> Just a random thought, I'm probably missing something...

This isn't about enslaving, this is just finding the master device.
Basically the virtnet_bypass uses a separate set of netdev ops so we
just are using that instead of maintaining a global or per-net hash.
You could have two virtio devices enslaved by the same virtnet_bypass
and it shouldn't be an issue.

>> +             if (ether_addr_equal(mac, dev->perm_addr))
>> +                     return dev;
>> +     }
>> +
>> +     return NULL;
>> +}
>> +
>> +static struct net_device *
>> +get_virtnet_bypass_byref(struct net_device *child_netdev)
>> +{
>> +     struct net *net = dev_net(child_netdev);
>> +     struct net_device *dev;
>> +
>> +     ASSERT_RTNL();
>> +
>> +     for_each_netdev(net, dev) {
>> +             struct virtnet_bypass_info *vbi;
>> +
>> +             if (dev->netdev_ops != &virtnet_bypass_netdev_ops)
>> +                     continue;       /* not a virtnet_bypass device */
>> +
>> +             vbi = netdev_priv(dev);
>> +
>> +             if ((rtnl_dereference(vbi->active_netdev) == child_netdev) ||
>> +                 (rtnl_dereference(vbi->backup_netdev) == child_netdev))
>
> nit: parens not needed

Yeah, it is a habit of mine since I do that for readability more than
anything else. Some people can't track the order of operations.

If they need to go they can.

>> +                     return dev;     /* a match */
>> +     }
>> +
>> +     return NULL;
>> +}
>
>> +static int virtnet_bypass_create(struct virtnet_info *vi)
>> +{
>> +     struct net_device *backup_netdev = vi->dev;
>> +     struct device *dev = &vi->vdev->dev;
>> +     struct net_device *bypass_netdev;
>> +     int res;
>> +
>> +     /* Alloc at least 2 queues, for now we are going with 16 assuming
>> +      * that most devices being bonded won't have too many queues.
>> +      */
>> +     bypass_netdev = alloc_etherdev_mq(sizeof(struct virtnet_bypass_info),
>> +                                       16);
>> +     if (!bypass_netdev) {
>> +             dev_err(dev, "Unable to allocate bypass_netdev!\n");
>> +             return -ENOMEM;
>> +     }
>
> Maybe it's just me but referring to master as bypass seems slightly
> confusing.  I know you don't like team and bond, but perhaps we can
> come up with a better name?  For me bypass device is the other leg,
> i.e. the VF, not the master.  Perhaps others disagree.

The choice of naming is based on some basic plumbing ideas. You could
almost think of the bypass netdev as more of a bypass valve. Basically
what we are doing is rerouting the traffic at the bypass interface and
sending it to the VF instead of routing it to the virtio_net. That was
why I thought it would be a fitting term. It gets us out of the
"bond", "team", and other such concepts because it isn't really any of
those. This is supposed to be a very simple device that will shunt the
traffic off of the virtio_net and re-route it through the VF. In
addition this isn't a complete solution by itself either since there
will still be some configuration required on the host to take care of
applying filters on the PF and/or tap interface to make certain we
don't end up with any loops in the traffic and that receives are
directed to the VF.

>> +     dev_net_set(bypass_netdev, dev_net(backup_netdev));
>> +     SET_NETDEV_DEV(bypass_netdev, dev);
>> +
>> +     bypass_netdev->netdev_ops = &virtnet_bypass_netdev_ops;
>> +     bypass_netdev->ethtool_ops = &virtnet_bypass_ethtool_ops;
>
> Thanks!

Powered by blists - more mailing lists