[dpdk-dev] [PATCH v3 0/6] vhost: add vhost-user client mode and reconnect ability

[Yuanhan Liu]

v3: - make the "reconnect" feature be default for client mode, as it's
      good to handle guest OS restart with less effort.
    - fix var not-initilized error pointed out by Rich


NOTE: I created a branch at dpdk.org [0] for more convenient testing:

    [0]: git://dpdk.org/next/dpdk-next-virtio for-testing


When the DPDK vhost-user application (such as OVS) restarts (due to
crash, or update), the vhost-user connection between DPDK and QEMU
won't be established automatically again. In another word, the virtio
net is broken.

The reason it doesn't work is that DPDK just acts as server only.
A restart of the server needs a reconnection from the client (QEMU).
However, reconnect from QEMU is not supported from QEMU.

Adding the support of client mode and let DPDK be the client somehow
would resolve above issue a bit easier: a restart of DPDK would naturally
try to connect to the server (QEMU) automatically.

Therefore, this patchset implements the DPDK vhost-user client mode, by
introducing a new arg (flags) for API rte_vhost_driver_register(). And the
client mode is enabled when RTE_VHOST_USER_CLIENT is given. Note that this
implies an API breakage. However, since this release deals with ABI/API
refactoring, it should not be an issue.

Another interesting thing to make it work is that you not only have
to consider that case the DPDK vhost-user app might restart, but also
have to think that QEMU might restart as well: guest OS sometimes
just reboots. In such case, when the server is down, the client has
to keep reconnecting with the server until the server is back and the
connection is established again. And that's what "reconnect" patch for.

Note that current QEMU doesn't not support a second time connection
from client, thus a restart of DPDK vhost-user will not work. This is
because current QEMU won't be able to detect the disconnect from
restart, thus it will not listen for later connections. Patches [1] have
been sent, it's just not merged yet. But unlike the vhost-user mulitple
queue case, that we have critical depends on QEMU implementation, here
we have no such dependency, therefore, I think it's okay to make DPDK
be ready for the "reconnect" stuff first. (note that I also mentioned
this fact in the release doc).

    [1]: http://lists.nongnu.org/archive/html/qemu-devel/2016-05/msg01507.html

v2: - added release doc
    - do not remove socket file for the client mode
    - create one thread ony to handle all reconnects


Thanks.
	--yliu

---
Yuanhan Liu (6):
  vhost: rename structs for enabling client mode
  vhost: add vhost-user client mode
  vhost: add reconnect ability
  vhost: workaround stale vring base
  examples/vhost: add client option
  vhost: add pmd client option

 doc/guides/rel_notes/release_16_07.rst       |  21 ++
 drivers/net/vhost/rte_eth_vhost.c            |  38 ++-
 examples/vhost/main.c                        |  12 +-
 lib/librte_vhost/rte_virtio_net.h            |  12 +-
 lib/librte_vhost/vhost_cuse/vhost-net-cdev.c |   8 +-
 lib/librte_vhost/vhost_user/vhost-net-user.c | 403 ++++++++++++++++++---------
 lib/librte_vhost/vhost_user/vhost-net-user.h |   6 -
 lib/librte_vhost/virtio-net.c                |   9 +
 8 files changed, 361 insertions(+), 148 deletions(-)

-- 
1.9.0