Message-Id: <1525731046-10989-1-git-send-email-sridhar.samudrala@intel.com>
Date:   Mon,  7 May 2018 15:10:42 -0700
From:   Sridhar Samudrala <sridhar.samudrala@...el.com>
To:     mst@...hat.com, stephen@...workplumber.org, davem@...emloft.net,
        netdev@...r.kernel.org, virtualization@...ts.linux-foundation.org,
        virtio-dev@...ts.oasis-open.org, jesse.brandeburg@...el.com,
        alexander.h.duyck@...el.com, kubakici@...pl,
        sridhar.samudrala@...el.com, jasowang@...hat.com,
        loseweigh@...il.com, jiri@...nulli.us, aaron.f.brown@...el.com
Subject: [PATCH net-next v10 0/4] Enable virtio_net to act as a standby for a passthru device

The main motivation for this patch is to enable cloud service providers
to provide an accelerated datapath to virtio-net enabled VMs in a 
transparent manner with no/minimal guest userspace changes. This also
enables hypervisor controlled live migration to be supported with VMs that
have direct attached SR-IOV VF devices.

Patch 1 introduces a new feature bit VIRTIO_NET_F_STANDBY that can be
used by hypervisor to indicate that virtio_net interface should act as
a standby for another device with the same MAC address.

Patch 2 introduces a failover module that provides a generic interface for 
paravirtual drivers to listen for netdev register/unregister/link change
events from pci ethernet devices with the same MAC and takeover their
datapath. The notifier and event handling code is based on the existing
netvsc implementation. It provides 2 sets of interfaces to paravirtual 
drivers to support 2-netdev(netvsc) and 3-netdev(virtio_net) models.

Patch 3 extends virtio_net to use alternate datapath when available and
registered. When STANDBY feature is enabled, virtio_net driver creates
an additional 'failover' netdev that acts as a master device and controls
2 slave devices.  The original virtio_net netdev is registered as
'standby' netdev and a passthru/vf device with the same MAC gets
registered as 'primary' netdev. Both 'standby' and 'failover' netdevs are
associated with the same 'pci' device.  The user accesses the network
interface via 'failover' netdev. The 'failover' netdev chooses 'primary'
netdev as default for transmits when it is available with link up and
running.

Patch 4 refactors netvsc to use the registration/notification framework
supported by failover module.

As this patch series is initially focusing on usecases where hypervisor 
fully controls the VM networking and the guest is not expected to directly 
configure any hardware settings, it doesn't expose all the ndo/ethtool ops
that are supported by virtio_net at this time. To support additional usecases,
it should be possible to enable additional ops later by caching the state
in virtio netdev and replaying when the 'primary' netdev gets registered. 
 
At the time of live migration, 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.

This patch is based on the discussion initiated by Jesse on this thread.
https://marc.info/?l=linux-virtualization&m=151189725224231&w=2

v10:
- Tested live migration with virtio-net/AVF(i40evf) configured in failover
  mode while running iperf in background. Tried static ip and dhcp
  configurations using 'network' scripts and Network Manager.
  - To avoid dhcp requests to be sent over the lower 'standby' and 'primary'
    netdevs, ifcfg-xxx scripts for the lower devices can be created.
- Build tested netvsc module.
Updates:
- fix net_failover_open() to update failover CARRIER correctly based on
  standby and primary states. 
- fix net_failover_handle_frame() to handle frames received on standby
  when primary is present.
- replace netdev_upper_dev_link with netdev_master_upper_dev_link and
  handle lower dev state changes.
- fix net_failver_create() and net_failover_register() interfaces to
  use ERR_PTR and avoid arg **
- disable setting mac address when virtio-net in STANDBY mode
- document exported symbols
- added entry to MAINTAINERS file

v9:
Select NET_FAILOVER automatically when VIRTIO_NET/HYPERV_NET 
are enabled. (stephen)

v8:
- Made the failover managment routines more robust by updating the feature 
  bits/other fields in the failover netdev when slave netdevs are 
  registered/unregistered. (mst)
- added support for handling vlans.
- Limited the changes in netvsc to only use the notifier/event/lookups
  from the failover module. The slave register/unregister/link-change 
  handlers are only updated to use the getbymac routine to get the 
  upper netdev. There is no change in their functionality. (stephen)
- renamed structs/function/file names to use net_failover prefix. (mst)

v7
- Rename 'bypass/active/backup' terminology with 'failover/primary/standy'
  (jiri, mst)
- re-arranged dev_open() and dev_set_mtu() calls in the register routines
  so that they don't get called for 2-netdev model. (stephen)
- fixed select_queue() routine to do queue selection based on VF if it is
  registered as primary. (stephen)
-  minor bugfixes

v6 RFC:
  Simplified virtio_net changes by moving all the ndo_ops of the 
  bypass_netdev and create/destroy of bypass_netdev to 'bypass' module.
  avoided 2 phase registration(driver + instances).
  introduced IFF_BYPASS/IFF_BYPASS_SLAVE dev->priv_flags 
  replaced mutex with a spinlock

v5 RFC:
  Based on Jiri's comments, moved the common functionality to a 'bypass'
  module so that the same notifier and event handlers to handle child
  register/unregister/link change events can be shared between virtio_net
  and netvsc.
  Improved error handling based on Siwei's comments.
v4:
- Based on the review comments on the v3 version of the RFC patch and
  Jakub's suggestion for the naming issue with 3 netdev solution,
  proposed 3 netdev in-driver bonding solution for virtio-net.
v3 RFC:
- Introduced 3 netdev model and pointed out a couple of issues with
  that model and proposed 2 netdev model to avoid these issues.
- Removed broadcast/multicast optimization and only use virtio as
  backup path when VF is unplugged.
v2 RFC:
- Changed VIRTIO_NET_F_MASTER to VIRTIO_NET_F_BACKUP (mst)
- made a small change to the virtio-net xmit path to only use VF datapath
  for unicasts. Broadcasts/multicasts use virtio datapath. This avoids
  east-west broadcasts to go over the PCI link.
- added suppport for the feature bit in qemu

Sridhar Samudrala (4):
  virtio_net: Introduce VIRTIO_NET_F_STANDBY feature bit
  net: Introduce generic failover module
  virtio_net: Extend virtio to use VF datapath when available
  netvsc: refactor notifier/event handling code to use the failover
    framework

 MAINTAINERS                     |    7 +
 drivers/net/Kconfig             |    1 +
 drivers/net/hyperv/Kconfig      |    1 +
 drivers/net/hyperv/hyperv_net.h |    2 +
 drivers/net/hyperv/netvsc_drv.c |  134 +----
 drivers/net/virtio_net.c        |   40 +-
 include/linux/netdevice.h       |   16 +
 include/net/net_failover.h      |   52 ++
 include/uapi/linux/virtio_net.h |    3 +
 net/Kconfig                     |   10 +
 net/core/Makefile               |    1 +
 net/core/net_failover.c         | 1044 +++++++++++++++++++++++++++++++++++++++
 12 files changed, 1198 insertions(+), 113 deletions(-)
 create mode 100644 include/net/net_failover.h
 create mode 100644 net/core/net_failover.c

-- 
2.14.3

Powered by blists - more mailing lists