login
Search
What it is: vhost net is a character device that can be used to reduce the number of system calls involved in virtio networking. Existing virtio net code is used in the guest without modification. There's similarity with vringfd, with some differences and reduced scope - uses eventfd for signalling - structures can be moved around in memory at any time (good for migration) - support memory table and not just an offset (needed for kvm) common virtio related code has been put in a separate file vhost.c and can be made into a separate module if/when more backend appear. I used Rusty's lguest.c as the source for developing this part : this supplied me with witty comments I wouldn't be able to write myself. What it is not: vhost net is not a bus, and not a generic new system call. No assumptions are made on how guest performs hypercalls. Userspace hypervisors are supported as well as kvm. How it works: Basically, we connect virtio frontend (configured by userspace) to a backend. The backend could be a network device, or a tun-like device. In this version I only support raw socket as a backend, which can be bound to e.g. SR IOV, or to macvlan device. Backend is also configured by userspace, including vlan/mac etc. Status: This works for me, and I haven't see any crashes. I have not run any benchmarks yet, compared to userspace, I expect to see improved latency (as I save up to 4 system calls per packet) but not yet bandwidth/CPU (as TSO and interrupt mitigation are not yet supported). Features that I plan to look at in the future: - TSO - interrupt mitigation - zero copy Signed-off-by: Michael S. Tsirkin <mst@redhat.com> --- MAINTAINERS | 10 + arch/x86/kvm/Kconfig | 1 + drivers/Makefile | 1 + drivers/block/virtio_blk.c | 3 + drivers/vhost/Kconfig | 11 + drivers/vhost/Makefile | 2 + drivers/vhost/net.c | 462 ++++++++++++++++++++++++++++++ drivers/vhost/vhost.c | 663 ...
Very nice, I loved reading it. It's getting rather late in my time
zone, so this comments only on the network driver. I'll go through
[style] I think the code gets more readable if you reorder it
what is the difference between vhost_net_reset_owner(n)
This is missing a compat_ioctl pointer. It should simply be
static long vhost_net_compat_ioctl(struct file *f,
unsigned int ioctl, unsigned long arg)
{
return f, ioctl, (unsigned long)compat_ptr(arg);
Why do you need a kernel thread here? If the data transfer functions
all get called from a guest intercept, shouldn't you already be
I guess that this is where one could plug into macvlan directly, using
sock_alloc_send_skb/memcpy_fromiovec/dev_queue_xmit directly,
instead of filling a msghdr for each, if we want to combine this
with the work I did on that.
Arnd <><
--
set socket to -1 will only stop the device. reset owner will let another process take over the device. It also needs to reset all parameters to make it safe for that other process, so in particular the device is stopped. I tried explaining this in the header vhost.h - does the comment I had the impression that if there's no compat_ioctl, several reasons :) - I get called under lock, so can't block - eventfd can be passed to another process, and I won't be in guest context at all - this also gets called outside guest context from socket poll - vcpu is blocked while it's doing i/o. it is better to free it up --
It will issue a kernel warning but not call unlocked_ioctl, so you need either a compat_ioctl method or list the numbers Right, that works as well, but may get more complicated once we try to add zero-copy or other optimizations. Arnd <>< --
We discussed this before, and I still think this could be directly derived from struct virtqueue, in the same way that vring_virtqueue is derived from struct virtqueue. That would make it possible for simple device drivers to use the same driver in both host and guest, similar to how Ira Snyder used virtqueues to make virtio_net run between two hosts running the same code [1]. Ideally, I guess you should be able to even make virtio_net work in the host if you do that, but that could bring other complexities. Arnd <>< [1] http://lkml.org/lkml/2009/2/23/353 --
I have no comments about the vhost code itself, I haven't reviewed it. It might be interesting to try using a virtio-net in the host kernel to communicate with the virtio-net running in the guest kernel. The lack of a management interface is the biggest problem you will face (setting MAC addresses, negotiating features, etc. doesn't work intuitively). Getting the network interfaces talking is relatively easy. Ira --
That was one of the reasons I decided to move most of code out to
userspace. My kernel driver only handles datapath,
Tried this, but
- guest memory isn't pinned, so copy_to_user
to access it, errors need to be handled in a sane way
- used/available roles are reversed
- kick/interrupt roles are reversed
So most of the code then looks like
if (host) {
} else {
}
return
The only common part is walking the descriptor list,
but that's like 10 lines of code.
At which point it's better to keep host/guest code separate, IMO.
--
MST
--
Ok, that makes sense. Let me see if I understand the concept of the
driver. Here's a picture of what makes sense to me:
guest system
---------------------------------
| userspace applications |
---------------------------------
| kernel network stack |
---------------------------------
| virtio-net |
---------------------------------
| transport (virtio-ring, etc.) |
---------------------------------
|
|
---------------------------------
| transport (virtio-ring, etc.) |
---------------------------------
| some driver (maybe vhost?) | <-- [1]
---------------------------------
| kernel network stack |
---------------------------------
host system
From the host's network stack, packets can be forwarded out to the
physical network, or be consumed by a normal userspace application on
the host. Just as if this were any other network interface.
In my patch, [1] was the virtio-net driver, completely unmodified.
So, does this patch accomplish the above diagram? If so, why the
copy_to_user(), etc? Maybe I'm confusing this with my system, where the
"guest" is another physical system, separated by the PCI bus.
Ira
--
Not exactly. vhost passes packets to a physical device, Guest memory is not pinned. Memory access needs to go through Yes, that's different. --
I prefer keeping it simple. Much of abstraction in virtio is due to the fact that it needs to work on top of different hardware emulations: lguest,kvm, possibly others in the future. vhost is always working on I don't think so. For example, there's a callback field that gets invoked in guest when buffers are consumed. It could be overloaded to mean "buffers are available" in host but you never handle both As I pointed out earlier, most code in virtio net is asymmetrical: guest provides buffers, host consumes them. Possibly, one could use virtio rings in a symmetrical way, but support of existing guest virtio net means there's almost no shared code. -- MST --
Well, that was my point: virtio can already work on a number of abstractions, The trick is to swap the virtqueues instead. virtio-net is actually mostly symmetric in just the same way that the physical wires on a twisted pair ethernet are symmetric (I like how that analogy fits). virtio_net kicks the transmit virtqueue when it has data and it kicks the receive queue when it has empty buffers to fill, and it has callbacks when the two are done. You can do the same in both the guest and the host, but then the guests input virtqueue is the hosts output virtqueue and vice versa. Once a virtqueue got kicked from both sides, the vhost_virtqueue implementation between the two only needs to do a copy_from_user or copy_to_user (possibly from a thread if it is in atomic context) and then call the two callback functions. This is basically the same thing you do already, except that you use slightly different names for the components. Arnd <>< --
It's already been done between two guests. See http://article.gmane.org/gmane.linux.kernel.virtualization/5423 Regards, Anthony Liguori --
Yes, this works by copying data (see PATCH 5/5). Another possibility is page flipping. Either will kill performance. -- MST --
You need to really squint hard for it to look symmetric. For example, for RX, virtio allocates an skb, puts a descriptor on a ring and waits for host to fill it in. Host system can not do the same: guest does not have access to host memory. You can do a copy in transport to hide this fact, but it will kill performance. -- MST --
Yes, that is what I was suggesting all along. The actual copy operation has to be done by the host transport, which is obviously different from the guest transport that just calls the host using vring_kick(). Right now, the number of copy operations in your code is the same. You are doing the copy a little bit later in skb_copy_datagram_iovec(), which is indeed a very nice hack. Changing to a virtqueue based method would imply that the host needs to add each skb_frag_t to its outbound virtqueue, which then gets copied into the guests inbound virtqueue. Unfortunately, this also implies that you could no longer simply use the packet socket interface as you do currently, as I realized only now. This obviously has a significant impact on your user space interface. Arnd <>< --
Also, if we do the copy in the transport, it definitely means that we can't get to zero-copy RX/TX from guest space any more. The current vhost_net driver doesn't do that yet, but could be extended in the same way that I'm hoping to do it for macvtap. Arnd <>< --
The best way to do this IMO would be to add zero copy support to raw sockets, vhost will then get it basically for free. -- MST --
Yes, that would be nice. I wonder if that could lead to security problems on TX though. I guess It will only bring significant performance improvements if we leave the data writable in the user space or guest during the operation. If the user finds the right timing, it could modify the frame headers after they have been checked using netfilter, or while the frames are being consumed in the kernel (e.g. an NFS server running in a guest). Ardn <>< --
IIRC when the kernel consumes data it linearizes the skb. We just need to make sure all the zerocopy data is in the nonlinear part, and the kernel will copy if/when it needs to access packet data. -- error compiling committee.c: too many arguments to function --
For this reason, we always linearize parts of packets we're filtering. ie. copy. Cheers, Rusty. --
And, it will remove our ability to implement zero copy down the road (when raw sockets support it). -- MST --
Well, I don't see this part as much of a problem, because the code already exists in virtio_net. If we really wanted to go down that road, just using virtio_net would solve the problem of frame handling entirely, but create new problems elsewhere, as we have mentioned. Arnd <>< --
Actually, vhost may not always be limited to real hardware. We may on day use vhost as the basis of a driver domain. There's quite a lot of interest in this for networking. At any rate, I'd like to see performance results before we consider trying to reuse virtio code. Regards, Anthony Liguori --
Yes, any ethernet device will do. What I mean is that vhost does not --
Yes, I agree. I'd also like to do more work on the macvlan extensions to see if it works out without involving a socket. Passing the socket into the vhost_net device is a nice feature of the current implementation that we'd have to give up for something else (e.g. making the vhost a real network interface that you can hook up to a bridge) if it were to use virtio. Unless I can come up with a solution that is clearly superior, I'm taking back my objections on that part for now. Arnd <>< --
Much better -- a couple of documentation nits below. How about something like "Therefore the beginning of workqueue execution acts as rcu_read_lock() and the end of workqueue execution acts as rcu_read_lock()"? It would also be good to add comments to the workqueue functions themselves saying that they act as read-side critical sections for your kind of RCU. --
