Re: [PATCH 0/2] eventfd: new EFD_STATE flag

[Avi Kivity <avi@redhat.com>]

On 08/26/2009 01:29 PM, Michael S. Tsirkin wrote:
> How we wanted to solve it with EFD_STATE: Share a separate eventfd
> between each device and the hypervisor.  device sets state to either 0
> or 1.  hypervisor polls all eventfds, reads interrupt line on changes,
> calculates the interrupt level and updates guest.
>
> Alternative solution: shared memory where each device writes interrupt
> line value. This makes setup more complex (need to share around much more
> than just an fd), and makes access from interrupt impossible unless we
> lock the memory (and locking userspace memory introduces yet another set
> of issues).
>    

For completeness:

If the device is implemented in the same process as the hypervisor, an 
eventfd isn't really needed, as there is an ioctl which performs the 
same operation.

An important class of device implementations is real devices that are 
assigned to a guest.  We would like to forward the interrupt directly 
from the host interrupt handler to qemu.  Currently, we have a 
kvm-specific interrupt handler that forwards the interrupt using 
kvm-specific interfaces.  We would like to use a generic interrupt 
handler implemented by uio, so we want a generic interrupt transfer 
mechanism.

uio already supports edge-triggered interrupts using an eventfd-like 
mechanism.  So it makes sense to extend uio to support real eventfds, 
and to make it also support level-triggered interrupts.

We can work around the lack of state eventfd by having userspace wait on 
whatever mechanism uio uses to make the interrupt state visible, and 
then use the ioctl mentioned above to inform the hypervisor of this 
state.  But it's faster and nicer to give both components an eventfd and 
let them communicate directly.

-- 
error compiling committee.c: too many arguments to function