Applied to dpdk-next-virtio.
--yliu
On Tue, Jun 07, 2016 at 12:05:02PM +0800, Yuanhan Liu wrote:
> 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
