On Thu, Dec 03, 2015 at 10:53:16AM +0100, Didier Pallard wrote: > Hi, > > I recently did some stress tests of a vhost-user interface using an UDP > traffic generator. Traffic generator was connected to 2 physical ports > that are in turn connected to 2 virtio ports through a linux bridge, VM > (running linux) doing routing to forward packets between the 2 virtio ports. > When traffic reaches high pps rates of small packets, I faced the 2 following > problems: > > - at some time, my qemu socket becomes full, causing qemu to send incomplete > SET_VRING_CALL messages to vhost-user backend (without proper fd set in > ancillary data). > - after some time, some interrupts are lost, causing the VM to stop > transmitting packets. > > Both problems come from the fact that interrupt masking/unmasking of the VM > is deferred to vhost-user backend through the linux socket. > First problem comes from the fact that socket buffer gets full; it is > corrected > in the first patch of the serie. > Second problem is a bit more complex. From what i understand of the code, > when VM wants to mask/unmask interrupts, qemu traps the command and sends a > SET_VRING_CALL to the vhost-user to swap interrupt notifications either to > a dummy descriptor or to an fd that was given to kvm module to route > interruption to the VM. After sending SET_VRING_CALL message through > the socket, VM code continues to run, assuming that the interrupts are now > masked; but due to linux socket, this message may be buffered and not > currently > treated by the vhost-user backend, ie interrupts are not really > masked/unmasked > as they ought to be. > I think it can be solved in two different ways: > - by waiting for an acknowledgement of vhost-user backend before exiting > interrupt masking function, this ensures that interrupt masking is correctly > taken into account before giving back hand to the VM, but it has a very high > cost in term of cycles. Moreover, unless specifying a new option, it will > break current API, since existing vhost-user implementations do not > expect a return message on a SET_VRING_CALL call. > - second way could be, in case vhost-user is involved, to restore the > initial behaviour of interrupt masking (ie masking/unmasking interrupts > by taking/giving vring_call fd from/to kernel kvm module). > Patches 2 and 3 of the serie propose an implementation of this second option. > > Didier >
Tested-by: Victor Kaplansky <vict...@redhat.com> > Didier Pallard (3): > char: fix vhost-user socket full > virtio-pci: add an option to bypass guest_notifier_mask > vhost-net: force guest_notifier_mask bypass in vhost-user case > > hw/net/vhost_net.c | 19 ++++++++++++++++++- > hw/virtio/vhost.c | 13 +++++++++++++ > hw/virtio/virtio-pci.c | 29 +++++++++++++++++++++++------ > hw/virtio/virtio-pci.h | 6 ++++++ > qemu-char.c | 10 ++++++++++ > 5 files changed, 70 insertions(+), 7 deletions(-) > > -- > 2.1.4 > >
