Vhost kick & call principle

1. vhost vring Set up

VHOST_SET_VRING_KICK：

         establish virtio Front end to vhost Back end notification mechanism ;

VHOST_SET_VRING_CALL：

         establish vhost Back to virtio Front end notification mechanism ;

2. VHOST_SET_VRING_KICK principle

vhost The back-end driver uses , Such as vsock Registered packages kick function ,

vsock->vqs[VSOCK_VQ_TX].handle_kick = vhost_vsock_handle_tx_kick; // be used for host Receive package

vsock->vqs[VSOCK_VQ_RX].handle_kick = vhost_vsock_handle_rx_kick; // be used for host Contract awarding

The following steps explain evenfd And kick How to establish a connection between .

1)  handle_kick Where did you register ?

VSOCK_VQ_RX = 0, /* for host to guest data */

VSOCK_VQ_TX = 1, /* for guest to host data */ 

vhost_poll_init call init_waitqueue_func_entry Set up (struct vhost_poll *) poll->wait.func = vhost_poll_wakeup; At the same time vhost_poll_func And poll table Make connections ：(&poll->table) ->pt->_qproc = vhost_poll_func, Process below VHOST_SET_VRING_KICK when , call vhost_poll_start->vfs_poll->eventfd_poll->poll_wait when , Will execute pt->_qproc namely vhost_poll_func, This function will poll->work Added below as KICK Registered eventfd Waiting queue （ctx->wqh） in ,

vhost_poll_init->vhost_work_init, so handle_tick Registered to work->fn.

2)  work->fn Where is it called ？

vhost_poll_wakeup -> "work->fn"( Namely vq Of handle_kick), but “work->fn” Only in vhost_dev No settings use_worker Will be called directly ,

         

If set use_worker（ here vsock Just set up use_worker）, Just call vhost_poll_queue Wake up worker, And then in worker Of handler In the implementation of work->fn(handle_kick),

vhost_poll_queue
    vhost_work_queue
        wake_up_process(dev->worker) // The corresponding vhost_dev Of work, Such as vsock->dev->worker
            vhost_worker //worker handler
                work->fn(work);

Above worker Need to be in LKVM Call in ioctl VHOST_SET_OWNER Will be the corresponding vhost Driver creation worker.kernel vhost worker optimization commit: 01fcb1cbc

3)  vhost_poll_wakeup When called ？

HOST received GUEST Of notify after （GUEST Write the correspondence virtio The equipment MMIO Spatial VIRTIO_MMIO_QUEUE_NOTIFY Field ）,

from kvm_io_bus_write The function starts to determine whether the system is the virtio The equipment MMIO Spatial VIRTIO_MMIO_QUEUE_NOTIFY The address is registered eventfd, If you register, you can pass NOTIFY Address found registered evenfd, To get evenfd Corresponding struct eventfd_ctx *ctx structure （ Every eventfd The kernel corresponds to a eventfd_ctx structure ）, And then get its waiting queue ctx->wqh, Get out of the queue work（ In the above analysis vhost_poll_func Function has poll->work Has joined wqh It's in ） To perform its fn that will do （__wake_up_common To realize ）,fn That is vhost_poll_wakeup, The complete call stack is as follows ,

vhost_work_queue
vhost_poll_wakeup
__wake_up_common
__wake_up_locked_key
eventfd_signal
ioeventfd_write
__kvm_io_bus_write
kvm_io_bus_write
kvm_{arch}_vcpu_exit
kvm_arch_vcpu_ioctl_run
kvm_vcpu_ioctl
sys_ioctl
ret_from_syscall

In the above steps ,vhost_poll_init call init_waitqueue_func_entry Set up poll->wait.func = vhost_poll_wakeup, So here vhost_poll_wakeup It is in the following logic that the call is triggered ,

4) LKVM How to register eventfd Of ？

• LKVM Side treatment

start-up GUEST Stage ,LKVM received guest Write MMIO Trigger settings QUEUE_PFN when , For the sake of QUEUE Corresponding to the device MMIO Spatial NITOFY Field settings eventfd,

• KERNEL Side treatment

vhost_vring_ioctl(VHOST_SET_VRING_KICK),

You know vq->kick As the corresponding eventfd Corresponding file, It can be seen from the above code , The eventfd file By ioctl adopt VHOST_SET_VRING_KICK Handed over .vhost_poll_start take eventfd Corresponding file Add to poll table In the middle ,

2. VHOST_SET_VRING_CALL principle

• LKVM Side treatment

gsi It's for irqfd Bound one guest Side interrupt number ,

• KERNEL Side treatment

1)  Kernel processing KVM_IRQFD

kvm_irqfd->kvm_irqfd_assign Established inject And shutdown Two work, And set the processing function , Be responsible for reporting to guest Trigger injection interrupt and shutdown interrupt ,

At the same time kvm_irqfd_assign Function irqfd Waiting queue handler irqfd_wakeup;init_poll_funcptr Function will irqfd_ptable_queue_proc And irqfd Of poll table（irqfd->pt） binding ,

And then call vfs_poll（evenfd file Of poll The function is eventfd_poll）,eventfd_poll Will call poll_wait(file, &ctx->wqh, wait) -> "p->_qproc(filp, &ctx->wqh, wait);", and _qproc It is registered above irqfd_ptable_queue_proc function , This function will call add_wait_queue(&ctx->wqh, &irqfd->wait) take irqfd->wait( namely wait_queue_entry_t) Add to evenfd Waiting queue wqh in , This is a crucial step . And the wait entry Registered func It is irqfd_wakeup, such irqfd and eventfd A two-way relationship is established .

At the same time, if you find that there are signals available now , Immediately call schedule_work, Scheduling execution irqfd->inject work, Execute interrupt injection ,

         

2)  Kernel processing VHOST_SET_VRING_CALL

The kernel will be the same eventfd Set it to vq->call_ctx.ctx,

3） host towards guest Inject interrupt process

stay host Need to be informed guest Will call vhost_signal->eventfd_signal, The first parameter is the one set above evenfd Of ctx,

As with the introduction KICK The call stack is the same ,eventfd_signal Will finally execute evenfd ctx->wqh Queue work->fn, namely irqfd_wakeup, Execute... In this function schedule_work(&irqfd->inject) Used to inject interrupts .

Reference resources ：

https://xzpeter.org/htmls/2017_12_07_kvm_irqfd/kvm_irqfd_implementation.html

The appended drawings ：

KVM VHOST in irqfd Use - jack.chen - Blog Garden

Dive into irqfd(KVM side) mechanism - L