On 08/15/2016 05:13 PM, Stefan Hajnoczi wrote:
On Mon, Aug 15, 2016 at 02:15:38AM +0300, Michael S. Tsirkin wrote:
On Sat, Aug 13, 2016 at 12:21:51PM +0200, ggar...@abra.uab.cat wrote:
From: Gerard Garcia <ggar...@deic.uab.cat>
This patch applies over the mst vhost git repository:
http://git.kernel.org/cgit/linux/kernel/git/mst/vhost.git
So I do like where this is going, but it gives me pause
that there's a global list of taps, where all sockets
seem to multicast to them all.
In particular, this won't play well with things
like containers.
vsock currently has no network namespace support. I agree that the tap
instances should be per-namespace when we add namespace support.
As each socket is bound to a physical device, how about binding
the monitor there as well?
Sockets aren't bound to physical devices, they are bound globally in the
af_vsock.ko module. The module currently doesn't allow multiple
instances (you cannot have multiple VMCI or virtio transports).
Only sockets from this device
would do the forwarding, and only one monitor per
device would be supported.
In a sense this will make it more like macvtap than tap.
Restricting the number of monitors could make userspace cumbersome.
Imagine two scripts that want to capture packets. The two scripts have
no knowledge of each other and create their own vsockmon interfaces. If
we restrict vsockmon to just 1 interface then users need to agree on
sharing just 1 vsockmon interface. I don't think this is beneficial.
So I think this global list is acceptable until we introduce network
namespace support. At that point it will become per-namespace.
Sorry, I was out last week.
I don't have much to add to what Stefan said. I agree that when vsock
introduces namespace support it will be necessary to have monitors
divided per-namespace. Right now, if only one af_vsock instance is
allowed, I think it makes sense to have a global list of taps.
Gerard
