On 04/19/2017 03:31 PM, Marc-André Lureau wrote:
Hi
On Wed, Apr 19, 2017 at 10:38 AM Wang, Wei W <wei.w.w...@intel.com
<mailto:wei.w.w...@intel.com>> wrote:
Hi,
We made some design changes to the original vhost-pci design, and
want to open
a discussion about the latest design (labelled 2.0) and its
extension (2.1).
2.0 design: One VM shares the entire memory of another VM
2.1 design: One VM uses an intermediate memory shared with another
VM for
packet transmission.
For the convenience of discussion, I have some pictures presented
at this link:
_https://github.com/wei-w-wang/vhost-pci-discussion/blob/master/vhost-pci-rfc2.0.pdf_
Fig. 1 shows the common driver frame that we want use to build the
2.0 and 2.1
design. A TX/RX engine consists of a local ring and an exotic ring.
Isn't "external" (or "remote") more appropriate than "exotic" ?
OK, probably we can use "remote" here.
Local ring:
1) allocated by the driver itself;
2) registered with the device (i.e. virtio_add_queue())
Exotic ring:
1) ring memory comes from the outside (of the driver), and exposed
to the driver
via a BAR MMIO;
2) does not have a registration in the device, so no
ioeventfd/irqfd, configuration
registers allocated in the device
Fig. 2 shows how the driver frame is used to build the 2.0 design.
1) Asymmetric: vhost-pci-net <-> virtio-net
2) VM1 shares the entire memory of VM2, and the exotic rings are
the rings
from VM2.
3) Performance (in terms of copies between VMs):
TX: 0-copy (packets are put to VM2’s RX ring directly)
RX: 1-copy (the green arrow line in the VM1’s RX engine)
Why is the copy necessary?
Because the packet from the remote ring can't be delivered to the
network stack directly. To be more precise,
1) The buffer from the remote ring is not allocated by the guest
driver. If the
buffer is directly delivered to the network stack, the network
stack will free
the buffer that is not allocated by the guest;
2) When we think about the vring operation, after getting the buffer
from the
avail ring, we need to put back the used buffer to the used ring to
tell the
other end that the buffer has been used. The network stack won't do
this
operation.
So, based on the two points. I think we need to use a local ring, and
copy the
packet to the buffer from the local ring (i.e. buffer memory allocated
by the
guest driver), and the driver will do the "give back used buffer" operation
as explained in 2).
Fig. 3 shows how the driver frame is used to build the 2.1 design.
1) Symmetric: vhost-pci-net <-> vhost-pci-net
2) Share an intermediate memory, allocated by VM1’s vhost-pci device,
for data exchange, and the exotic rings are built on the shared memory
3) Performance:
TX: 1-copy
RX: 1-copy
Fig. 4 shows the inter-VM notification path for 2.0 (2.1 is similar).
The four eventfds are allocated by virtio-net, and shared with
vhost-pci-net:
Uses virtio-net’s TX/RX kickfd as the vhost-pci-net’s RX/TX callfd
Uses virtio-net’s TX/RX callfd as the vhost-pci-net’s RX/TX kickfd
Example of how it works:
After packets are put into vhost-pci-net’s TX, the driver kicks
TX, which
causes the an interrupt associated with fd3 to be injected to
virtio-net
The draft code of the 2.0 design is ready, and can be found here:
Qemu: _https://github.com/wei-w-wang/vhost-pci-device_
The repository contains a single big commit
(https://github.com/wei-w-wang/vhost-pci-device/commit/fa01ec5e41de176197dae505c05b659f5483187f).
Please try to provide a seperate patch or series of patch from an
upstream commit/release point.
It's the test-able version of the 2.0 design. I will separate it.
If possible, hope we can review the design first, especially the common
driver frame. Then I can make the related changes from the
discussion, and post out the patch series.
Best,
Wei