Re: [Qemu-devel] [PATCH 0/3] vhost-user reconnect

[Marc-André Lureau]

Hi

On Mon, Aug 20, 2018 at 2:51 PM, Yury Kotov <address@hidden> wrote:
> 16.08.2018, 18:36, "Marc-André Lureau" <address@hidden>:
>> On Thu, Aug 16, 2018 at 5:32 PM, Yury Kotov <address@hidden> wrote:
>>>  We are using QEMU (2.12.0) with SPDK (18.04.1) over vhost-user to emulate 
>>> block
>>>  devices. One of our cases it to restart SPDK without restarting VM (in case
>>>  of some updates or smth like it). We tried to use the 'reconnect' option 
>>> for
>>>  the '-chardev' device:
>>>    -object 
>>> memory-backend-file,id=mem0,size=1G,mem-path=/dev/hugepages,share=on \
>>>    -numa node,memdev=mem0 \
>>>    -chardev socket,id=spdk_vhost_blk1,path=/var/tmp/vhost.1,reconnect=10 \
>>>    -device vhost-user-blk-pci,chardev=spdk_vhost_blk1,num-queues=4
>>>
>>>  After this, vhost-user-blk initialization fails with an error below:
>>>    qemu-system-x86_64: -device ...: Failed to set msg fds.
>>>    qemu-system-x86_64: -device ...: vhost-user-blk: vhost initialization 
>>> failed:
>>>                                     Operation not permitted
>>>
>>>  We got the same error with the latest QEMU (c542a9f9794ec8e0bc3f).
>>>
>>>  We made some investigations and found out that there are several issues:
>>>
>>>  1. Reconnect option postpones the first connection till machine init done 
>>> event.
>>>     But we need this connection during vhost blk device initialization which
>>>     happens before the machine init done handling.
>>>
>>>  2. If the connection is forced, then the reconnection will be successful
>>>     after SPDK restart. The problem is that virtual queue will not start.
>>>     The reason for it is that virtual queue initialization commands
>>>     should be resent:
>>>     * VHOST_USER_SET_FEATURES
>>>     * VHOST_USER_SET_MEM_TABLE
>>>     * VHOST_USER_SET_VRING_NUM
>>>     * VHOST_USER_SET_VRING_BASE
>>>     * VHOST_USER_SET_VRING_ADDR
>>>     * VHOST_USER_SET_VRING_KICK
>>>     * VHOST_USER_SET_VRING_CALL
>>>
>>>  The patch set resolves both of these issues.
>>>
>>>  Test case:
>>>
>>>  1. Start fio process (inside VM):
>>>       fio --name test --ioengine=libaio --iodepth=64 --bs=4096 \
>>>           --rw=randrw --direct=1 --sync=1 --verify=md5 \
>>>           --size=64M --filename=/dev/vda --loops=100
>>>
>>>  2. Restart SPDK many times.
>>>     We are expecting that during SPDK restart fio will pause and fio should
>>>     continue to work after restart completion.
>>>
>>>  3. fio process completed successfully without any error.
>>
>> Can you write a test case in vhost-user-test.c ? (perhaps under
>> QTEST_VHOST_USER_FIXME scope...)
>>
>
> This is a great idea and we were definitely going to do that during coming 
> couple of weeks. We thought that we could make a follow up commit with 
> necessary tests added a bit later though, since currently we need to figure 
> out the state of vhost-user tests in general, before we can try to add any 
> new stuff, and that will take some time. So far we have stress-tested these 
> fixes manually.

Yes, some vhost-user tests are disabled by default (sadly for travis
CI reason - not a really bug), and it's easy to introduce regressions.

I sent a related series "[PATCH 0/4] Fix socket chardev regression" to
make it work again.

> Do you suggest we wait with this series as well until we have all tests 
> ready? Or do we proceed now and make a follow up series with vhost user tests 
> later like we suggested?

I would rather have the tests with the series.

>
>>>  Yury Kotov (3):
>>>    chardev: prevent extra connection attempt in tcp_chr_machine_done_hook
>>>    vhost: refactor vhost_dev_start and vhost_virtqueue_start
>>>    vhost-user: add reconnect support for vhost-user
>>>
>>>   chardev/char-socket.c | 5 +-
>>>   hw/virtio/vhost-user.c | 65 ++++++++++++--
>>>   hw/virtio/vhost.c | 223 +++++++++++++++++++++++++++++++---------------
>>>   include/hw/virtio/vhost.h | 2 +
>>>   4 files changed, 215 insertions(+), 80 deletions(-)
>>>
>>>  --
>>>  2.7.4