Re: [PATCH v1 1/3] virtio-pci: submit msi route changes in batch

[Longpeng (Mike, Cloud Infrastructure Service Product Dept.)]

在 2023/2/28 19:20, Longpeng (Mike, Cloud Infrastructure Service Product 
Dept.) 写道:
> 在 2023/2/28 18:17, Michael S. Tsirkin 写道:
> > On Tue, Feb 28, 2023 at 05:39:35PM +0800, Longpeng(Mike) wrote:
> > > From: Longpeng <longpeng2@huawei.com>
> > >
> > > The kvm_irqchip_commit_routes() is a time-intensive operation, it needs
> > > scan and update all irqfds that are already assigned during each 
> > > invocation,
> > > so more vectors means need more time to process them.
> >
> > I think the real reason is it's the write side of RCU.
>
> Yes, so we can reduce the invocation of it in this way.
>
> I'll send other optimizations in the next step, including irqbypass, 
> kvm_irqfd, etc.

Iterates the irqfds list is also time-consuming, it would iterate all 
existing irqfds when we commit, so the time complexity is O(n^2) without 
this patch.

> > > For virtio-pci, we
> > > can just submit once when enabling vectors of a virtio-pci device.
> > >
> > > This can reduce the downtime when migrating a VM with vhost-vdpa 
> > > devices.
> > >
> > > Signed-off-by: Longpeng <longpeng2@huawei.com>
> > > ---
> > >   hw/virtio/virtio-pci.c | 24 +++++++++++++++++++++---
> > >   1 file changed, 21 insertions(+), 3 deletions(-)
> > >
> > > diff --git a/hw/virtio/virtio-pci.c b/hw/virtio/virtio-pci.c
> > > index 247325c193..22e76e3902 100644
> > > --- a/hw/virtio/virtio-pci.c
> > > +++ b/hw/virtio/virtio-pci.c
> > > @@ -49,6 +49,19 @@
> > >    * configuration space */
> > >   #define VIRTIO_PCI_CONFIG_SIZE(dev)     
> > > VIRTIO_PCI_CONFIG_OFF(msix_enabled(dev))
> > > +/* Protected by the BQL */
> > > +static KVMRouteChange virtio_pci_route_change;
> > > +
> > > +static inline void virtio_pci_begin_route_changes(void)
> > > +{
> > > +    virtio_pci_route_change = 
> > > kvm_irqchip_begin_route_changes(kvm_state);
> > > +}
> > > +
> > > +static inline void virtio_pci_commit_route_changes(void)
> > > +{
> > > +    kvm_irqchip_commit_route_changes(&virtio_pci_route_change);
> > > +}
> > > +
> > >   static void virtio_pci_bus_new(VirtioBusState *bus, size_t bus_size,
> > >                                  VirtIOPCIProxy *dev);
> > >   static void virtio_pci_reset(DeviceState *qdev);
> > > @@ -790,12 +803,11 @@ static int 
> > > kvm_virtio_pci_vq_vector_use(VirtIOPCIProxy *proxy,
> > >       int ret;
> > >       if (irqfd->users == 0) {
> > > -        KVMRouteChange c = kvm_irqchip_begin_route_changes(kvm_state);
> > > -        ret = kvm_irqchip_add_msi_route(&c, vector, &proxy->pci_dev);
> > > +        ret = kvm_irqchip_add_msi_route(&virtio_pci_route_change, 
> > > vector,
> > > +                                        &proxy->pci_dev);
> > >           if (ret < 0) {
> > >               return ret;
> > >           }
> > > -        kvm_irqchip_commit_route_changes(&c);
> > >           irqfd->virq = ret;
> > >       }
> > >       irqfd->users++;
> > > @@ -903,12 +915,18 @@ static int 
> > > kvm_virtio_pci_vector_vq_use(VirtIOPCIProxy *proxy, int nvqs)
> > >       int ret = 0;
> > >       VirtIODevice *vdev = virtio_bus_get_device(&proxy->bus);
> > > +    virtio_pci_begin_route_changes();
> > > +
> > >       for (queue_no = 0; queue_no < nvqs; queue_no++) {
> > >           if (!virtio_queue_get_num(vdev, queue_no)) {
> > > +            virtio_pci_commit_route_changes();
> > >               return -1;
> > >           }
> > >           ret = kvm_virtio_pci_vector_use_one(proxy, queue_no);
> > >       }
> > > +
> > > +    virtio_pci_commit_route_changes();
> > > +
> > >       return ret;
> > >   }
> > > --
> > > 2.23.0
> >
> > .