On Thu, Sep 18, 2014 at 08:39:57PM +0800, Linhaifeng wrote: > > > On 2014/9/18 13:16, Michael S. Tsirkin wrote: > > On Thu, Sep 18, 2014 at 08:45:37AM +0800, Linhaifeng wrote: > >> > >> > >> On 2014/9/17 17:56, Michael S. Tsirkin wrote: > >>> On Wed, Sep 17, 2014 at 05:39:04PM +0800, Linhaifeng wrote: > >>>> I think maybe is not need for the backend to wait for response. > >>>> > >>>> There is another way.vhost-user send "VHOST_GET_MEM_TABLE" to qemu then > >>>> qemu send VHOST_SET_MEM_TABLE to update the regions of vhost-user.same > >>>> as other command. > >>>> If qemu could response the request of the vhost-user.the vhost-user > >>>> could update date at anytime. > >>> > >>> The updates are initiated by QEMU, as a result of IOMMU, > >>> memory hotplug or some other configuration change. > >> > >> How to deal with the vhost-user restart? > >> when vhost-user restart it will lost the infomation which QEMU send. > >> > >> In the kernel mode vhost will restart with QEMU but in the user mode vhost > >> will not. > > > > vhost-user must restart with qemu only. > > > > > Sometimes qemu not allowed to restart. e.g. The customer want to update the > vhost-user to a newer version but don't want to restart the VM.
I guess this means we'll have to implement the idea outlined below. Fixing known issues around vhost-user (such as lack of set_backend, and lack of synchronization around some commands) is probably higher priority. > > > The nature of virtio protocol is such that there's not enough in-memory > > state for host to gracefully recover from losing VQ state. > > > > We could add a new feature to allow recovery by reporting > > failure to guest, and disabling processing new requests. > > Guest could respond by recovering / discarding submitted > > requests, re-enabling the device (likely by executing a reset) > > and re-submitting requests. > > > > > > This would need a new feature bit though, and would have to > > be acknowledged by guest. As such this would have to be > > virtio 1.0 feature, virtio 0.x is frozen. > > > >>> > >>>> I think it's very useful for Commercialization. > >>>> > >>>> On 2014/9/17 16:38, Michael S. Tsirkin wrote: > >>>>> Reply-To: > >>>>> > >>>>> Thinking about the vhost-user protocol, VHOST_SET_MEM_TABLE > >>>>> is used to update the memory mappings. > >>>>> > >>>>> So shouldn't we want for response? > >>>>> Otherwise e.g. guest can start using the memory > >>>>> that vhost-user can't access. > >>>>> > >>>>> Similarly, with an IOMMU vhost-user might access memory it shouldn't. > >>>>> > >>>>> VHOST_SET_VRING_CALL is used for MSI-X masking. > >>>>> Again, after vector is masted by switching the call fd, > >>>>> backend shouldn't assert the old one. > >>>>> > >>>>> Thoughts? > >>>>> > >>>>> > >>> > >>> . > >>> > > > > . > >
