On 09/08/2022 19:50, Elliot Berman wrote:
On 8/9/2022 4:29 AM, Marc Zyngier wrote:
On Mon, 08 Aug 2022 23:22:48 +0100,
Elliot Berman <quic_eberman@xxxxxxxxxxx> wrote:
In a future series, I'll add the support to load other virtual
machines. When running other virtual machines, additional gunyah
devices are needed for doorbells (e.g. to emulate interrupts for
paravirtualized devices) and to represent the vCPUs of that other
VM. Other gunyah devices are also possible, but those are the
immediate devices coming over the horizon.
Can you elaborate on this "doorbell" aspect? If you signal interrupts
to guests, they should be signalled as actual interrupts, not as some
hypervisor-specific events, as we rely on the interrupt semantics for
most things.
Or are you talking about injecting an interrupt from a guest into
another, the doorbell representing an interrupt source?
Doorbells can operate either of these modes:
1. As simple interrupt sources. The doorbell sender makes a hypercall
and an interrupt is raised on the receiver. The hypervisor can be
configured to raise a specific SPI on the receiver VM and simply
acknowledging the SPI is enough to clear the interrupt assert. No
hypervisor-specific code is needed on the receiver to handle these
interrupts. This is the mode one would expect to use for
paravirtualized devices.
This sounds good.
2. As hypervisor-specific events which must be acknowledged using
hypercalls. We aren't currently using this advanced use-case and no
plans currently to post these. However, I can try to briefly
explain: These doorbells can operate on a bitfield and the sender
can assert flags on the bitmask; the receiver can decide which bits
should trigger the interrupt and which SPI the doorbell "runs" on.
The "user story" for this doorbell is to support multiple sender
using the same doorbell object. Each sender has a few designated
bits they should set. The receiver can choose which events it wants
an interrupt to be raised for and then can process all the pending
events. To re-iterate, we don't have an interesting use-case for
this yet, so don't plan on post patches for this second mode of
doorbell.
Well. For me this sounds like 'we have such capability, no real usecase,
but we want to support it anyway' kind of story. As history has shown
multiple times, the order should be the opposite one. First you have the
use case, then you create the API for it. Otherwise it is very easy to
end up with the abstraction that looks good on the API side, but is very
hard to fit into the actual user code.
I would suggest to drop the second bullet for now and focus on getting
the simple doorbells done and accepted into mainline.
--
With best wishes
Dmitry
