> > It should be once, from the host memory address space to the guest > > memory address space (and vice-versa in the other direction). > > I can not understand 'be once' > > In vhost, VM share *down* the memory to host, why host 'copy the pkts > from the host memory address space to the guest memory address space'?
Yes, this is what happens. When a frame arrives in the NIC, it is DMA'ed into the host memory. At that point, the vswitch decides where it should be sent. Then the vswitch will copy the frame into the destination guest's address space so that the guest can access it. > > thks for reply > > ********************* > > RTFSC > > ********************* > > > > > ------------------------------------------------------------------ > > 发件人：Gray, Mark D <mark.d.g...@intel.com> > > 发送时间：2015年6月4日(星期四) 06:10 > > 收件人：Ethan Jackson <et...@nicira.com>， > dev@openvswitch.org <dev@openvswitch.org> > > 主 题：Re: [ovs-dev] dpdk vhost-user patch questions. > > > > > Hi Ethan > > > > > > > > > I have some general questions about the vhost user feature which > aren't > > > totally clear to me. I would greatly appreciate some clarification, > or > pointers > > > to documentation. > > > > > Have you watched the presentation that Kevin and Maryam did at > the conference > > last year? It describes vhost-cuse pretty well. The only major > difference between > > cuse and user is the use of the socket for signaling. The datapath is > basically the same. > > > > > > > > > 1) Since vhost-user is using standard Qemu interfaces, will it > support all of the > > > features of the standard vhost implementation? I.E. can you do > things like > > > live migration on a VM which is using this? My understanding is you > can't > > > with ivshmem. > > > > > Yes, although I think live migration is currently broken in DPDK. There > is a > > patch that needs to be upstreamed. > > > > > Currently, you can't live migrate with ivshmem. That’s not to say it > couldn’t be done > > but there are challenges. With Ivshmem, you create all the data > structures (rings > > and buffer pools) for all the vms in the host memory space in one or > more shared pages and > > then you share it *up* to the guest. With vhost, it's kind of the > opposite. You > > create all the data structures for an individual vm in that vm's > memory space. Then > > you share the memory *down* to the host. So if you need to live > migrate > > with vhost, it's easier as you just need to migrate the guest memory, > which > > you have to do anyway as it contains the state of the guest operating > system. > > With ivshmem, it’s a little trickier as all the data structures for all > the > vms are > > in the same page (or pages) in the host. Therefore, you would need > to figure > > out which were the correct ones to migrate. Also, migrating pages > that are > > really part of the host memory space (but shared up) is not a normal > migration > > operation. > > > > > > > > > 2) What are the security implications of vhost-user? Can a miss- > behaving > > > guest crash the vswitch? Can guests crash each other? > > > Is there a separate shared memory region per vif, or is it shared > globally? > > > > > This is an advantage of vhost. The rings and buffers are present in the > > guest address space and shared down which keeps them isolated > from other > > virtual machines preventing security issues like this. > > > > > ivshmem is different as the host memory is getting shared > everywhere. Again, > > this could be resolved. We did some work in dpdk that allowed you > to share > > individual memory objects to a guest. For example, share a single ring > descriptor. > > You could then, in theory, partition the memory into zones. If two > guests > > were in the same zone, they could share the same ring and buffer > pool and then > > communicate in a zero-copy manner. If they were not, you would > then do > > a copy in the host which would give you isolation but it wouldn’t be > zero-copy. > > > > > > 3) What actually happens to packets transmitted over vhost-user? > Are they > > > copied once? Twice? Zero-copy? It looks like a unix domain socket > is used for > > > signalling. Is that done per packet or just at setup? > > > > > It should be once, from the host memory address space to the guest > > memory address space (and vice-versa in the other direction). > > > > > The socket is only used for signaling at setup. I think it is also used > when > > the guest uses a virtio-net driver (interrupt driven) but not for > virtio- > pmd driver > > (poll mode). > > > > > > Please excuse me if this stuff is answered somewhere else already, > I can't > > > seem to find it. Thanks in advance for the response, this will help > me > > > message the feature to users on our end. > > > > > There is some documentation in the dpdk repo. > > > > > http://www.dpdk.org/doc/guides/prog_guide/vhost_lib.html > > http://www.dpdk.org/doc/api/rte__virtio__net_8h.html > > > > > > > > > Ethan > > > _______________________________________________ > > > dev mailing list > > > dev@openvswitch.org > > > http://openvswitch.org/mailman/listinfo/dev > > _______________________________________________ > > dev mailing list > > dev@openvswitch.org > > http://openvswitch.org/mailman/listinfo/dev _______________________________________________ dev mailing list dev@openvswitch.org http://openvswitch.org/mailman/listinfo/dev
