Hi all,
I am looking for a way to have KVM_RUN exit back to userspace once the
vcpu encounters a WFI. It seems the kvm_run->request_interrupt_window
flag is currently ignored by arm64. So my solution thus far is to
patch kvm_handle_wfx in arch/arm64/kvm/handle_exit.c and return to
userspace with KVM_EXIT_IRQ_WINDOW_OPEN - working example attached.
Any chance to get this (or something similar) mainline?
--
Dr.-Ing. Jan Henrik Weinstock
Managing Director
MachineWare GmbH | http://www.machineware.de
Hühnermarkt 19, 52062 Aachen, Germany
Amtsgericht Aachen HRB25734
Geschäftsführung
Lukas Jünger
Dr.-Ing. Jan Henrik Weinstock
Hi Jan,
On Fri, 20 Oct 2023 19:45:05 +0100,
Jan Henrik Weinstock <[email protected]> wrote:
>
> Hi all,
>
> I am looking for a way to have KVM_RUN exit back to userspace once the
> vcpu encounters a WFI. It seems the kvm_run->request_interrupt_window
> flag is currently ignored by arm64.
Well, that's consistent with arm64 not being an x86 implementation. We
can inject interrupts any time, and there is no notion of "interrupt
window".
> So my solution thus far is to
> patch kvm_handle_wfx in arch/arm64/kvm/handle_exit.c and return to
> userspace with KVM_EXIT_IRQ_WINDOW_OPEN - working example attached.
> Any chance to get this (or something similar) mainline?
Certainly not as such. For start, this won't hit all WFIs, but only
those that actively trap. And we don't even *try* to trap WFx in a
number of cases (vcpu alone in its run queue and/or direct injection).
There isn't even any guarantee that WFx is anything other than a NOP
(it is architecturally only a hint), in which case no trap applies.
So your "working example" really isn't one, as the architecture
doesn't give you a way to do what you're asking for. If you want to
cause an exit, writing to 'immediate_exit' and delivering a signal is
the way.
Thanks,
M.
--
Without deviation from the norm, progress is not possible.
Hi Marc,
Thanks for your feedback. I understand that request_interrupt_window
is not to be used. I assume a setting a flag is a better way,
something similar to KVM_ARCH_FLAG_RETURN_NISV_IO_ABORT_TO_USER, e.g.
KVM_ARCH_FLAG_WFX_EXIT_TO_USER.
I will also check that WFx traps are always enabled while this mode is
active to make sure userspace does not get blocked/scheduled out.
The reason for this is that we cannot have the thread that executes
KVM_RUN to be blocked or scheduled out whenever it hits a WFI.
Nop-WFIs are not a problem, since the PE will just continue executing
instructions, which is fine. We are currently using a timeout signal
that kicks KVM_RUN back into userspace, but we are seeing a lot of
time wasted because our KVM thread hangs in WFI/WFEs. It would be
better if we could just return from KVM_RUN immediately if the thread
would otherwise be blocked.
Thanks
Jan
Am Fr., 20. Okt. 2023 um 21:56 Uhr schrieb Marc Zyngier <[email protected]>:
>
> Hi Jan,
>
> On Fri, 20 Oct 2023 19:45:05 +0100,
> Jan Henrik Weinstock <[email protected]> wrote:
> >
> > Hi all,
> >
> > I am looking for a way to have KVM_RUN exit back to userspace once the
> > vcpu encounters a WFI. It seems the kvm_run->request_interrupt_window
> > flag is currently ignored by arm64.
>
> Well, that's consistent with arm64 not being an x86 implementation. We
> can inject interrupts any time, and there is no notion of "interrupt
> window".
>
> > So my solution thus far is to
> > patch kvm_handle_wfx in arch/arm64/kvm/handle_exit.c and return to
> > userspace with KVM_EXIT_IRQ_WINDOW_OPEN - working example attached.
> > Any chance to get this (or something similar) mainline?
>
> Certainly not as such. For start, this won't hit all WFIs, but only
> those that actively trap. And we don't even *try* to trap WFx in a
> number of cases (vcpu alone in its run queue and/or direct injection).
> There isn't even any guarantee that WFx is anything other than a NOP
> (it is architecturally only a hint), in which case no trap applies.
>
> So your "working example" really isn't one, as the architecture
> doesn't give you a way to do what you're asking for. If you want to
> cause an exit, writing to 'immediate_exit' and delivering a signal is
> the way.
>
> Thanks,
>
> M.
>
> --
> Without deviation from the norm, progress is not possible.
--
Dr.-Ing. Jan Henrik Weinstock
Managing Director
MachineWare GmbH | http://www.machineware.de
Hühnermarkt 19, 52062 Aachen, Germany
Amtsgericht Aachen HRB25734
Geschäftsführung
Lukas Jünger
Dr.-Ing. Jan Henrik Weinstock
On Wed, 25 Oct 2023 13:12:14 +0100,
Jan Henrik Weinstock <[email protected]> wrote:
>
> Hi Marc,
>
> Thanks for your feedback. I understand that request_interrupt_window
> is not to be used. I assume a setting a flag is a better way,
> something similar to KVM_ARCH_FLAG_RETURN_NISV_IO_ABORT_TO_USER, e.g.
> KVM_ARCH_FLAG_WFX_EXIT_TO_USER.
>
> I will also check that WFx traps are always enabled while this mode is
> active to make sure userspace does not get blocked/scheduled out.
Why would that be an acceptable behaviour?
> The reason for this is that we cannot have the thread that executes
> KVM_RUN to be blocked or scheduled out whenever it hits a WFI.
Why? If that's not acceptable, how do you even cope with the basic
preemption?
> Nop-WFIs are not a problem, since the PE will just continue executing
> instructions, which is fine. We are currently using a timeout signal
> that kicks KVM_RUN back into userspace, but we are seeing a lot of
> time wasted because our KVM thread hangs in WFI/WFEs. It would be
> better if we could just return from KVM_RUN immediately if the thread
> would otherwise be blocked.
On the face of it, this makes little sense:
- While in userspace, no interrupt source that normally delivered
without any userpsace intervention will be blocked (timers,
VLPIs...). I cannot how this can be a good idea.
- Trapping WFE is an important scheduling hint, and returning to
userspace defeats it. Contended spinlocks, for example, will be even
slower to acquire.
I'm sure you have a particular use case for such a degraded behaviour,
but since you are not describing it, I'm not at all inclined to
actively break KVM's performance and scalability.
Thanks,
M.
--
Without deviation from the norm, progress is not possible.
Hi Marc,
the basic idea behind this is to have a (single-threaded) execution loop,
something like this:
vcpu-thread: vcpu-run | process-io-devices | vcpu-run | process-io...
^
WFX or timeout
We switch to simulating IO devices whenever the vcpu is idle (wfi) or exceeds
a certain budget of instructions (counted via pmu). Our fallback currently is
to kick the vcpu out of its execution using a signal (via a timeout/alarm). But
of course, if the cpu is stuck at a wfi, we are wasting a lot of time.
I understand that the proposed behavior is not desirable for most use cases,
which is why I suggest locking it behind a flag, e.g.
KVM_ARCH_FLAG_WFX_EXIT_TO_USER.
Am Mi., 25. Okt. 2023 um 14:42 Uhr schrieb Marc Zyngier <[email protected]>:
>
> On Wed, 25 Oct 2023 13:12:14 +0100,
> Jan Henrik Weinstock <[email protected]> wrote:
> >
> > Hi Marc,
> >
> > Thanks for your feedback. I understand that request_interrupt_window
> > is not to be used. I assume a setting a flag is a better way,
> > something similar to KVM_ARCH_FLAG_RETURN_NISV_IO_ABORT_TO_USER, e.g.
> > KVM_ARCH_FLAG_WFX_EXIT_TO_USER.
> >
> > I will also check that WFx traps are always enabled while this mode is
> > active to make sure userspace does not get blocked/scheduled out.
>
> Why would that be an acceptable behaviour?
>
> > The reason for this is that we cannot have the thread that executes
> > KVM_RUN to be blocked or scheduled out whenever it hits a WFI.
>
> Why? If that's not acceptable, how do you even cope with the basic
> preemption?
>
> > Nop-WFIs are not a problem, since the PE will just continue executing
> > instructions, which is fine. We are currently using a timeout signal
> > that kicks KVM_RUN back into userspace, but we are seeing a lot of
> > time wasted because our KVM thread hangs in WFI/WFEs. It would be
> > better if we could just return from KVM_RUN immediately if the thread
> > would otherwise be blocked.
>
> On the face of it, this makes little sense:
>
> - While in userspace, no interrupt source that normally delivered
> without any userpsace intervention will be blocked (timers,
> VLPIs...). I cannot how this can be a good idea.
>
> - Trapping WFE is an important scheduling hint, and returning to
> userspace defeats it. Contended spinlocks, for example, will be even
> slower to acquire.
>
> I'm sure you have a particular use case for such a degraded behaviour,
> but since you are not describing it, I'm not at all inclined to
> actively break KVM's performance and scalability.
>
> Thanks,
>
> M.
>
> --
> Without deviation from the norm, progress is not possible.
--
Dr.-Ing. Jan Henrik Weinstock
Managing Director
MachineWare GmbH | http://www.machineware.de
Hühnermarkt 19, 52062 Aachen, Germany
Amtsgericht Aachen HRB25734
Geschäftsführung
Lukas Jünger
Dr.-Ing. Jan Henrik Weinstock
[please make an effort not to top-post]
On Fri, 27 Oct 2023 18:41:44 +0100,
Jan Henrik Weinstock <[email protected]> wrote:
>
> Hi Marc,
>
> the basic idea behind this is to have a (single-threaded) execution loop,
> something like this:
>
> vcpu-thread: vcpu-run | process-io-devices | vcpu-run | process-io...
> ^
> WFX or timeout
>
> We switch to simulating IO devices whenever the vcpu is idle (wfi) or exceeds
> a certain budget of instructions (counted via pmu). Our fallback currently is
> to kick the vcpu out of its execution using a signal (via a timeout/alarm). But
> of course, if the cpu is stuck at a wfi, we are wasting a lot of time.
>
> I understand that the proposed behavior is not desirable for most use cases,
> which is why I suggest locking it behind a flag, e.g.
> KVM_ARCH_FLAG_WFX_EXIT_TO_USER.
But how do you reconcile the fact that exposing this to userspace
breaks fundamental expectations that the guest has, such as getting
its timer interrupts and directly injected LPIs? Implementing WFI in
userspace breaks it. What about the case where we don't trap WFx and
let the *guest* wait for an interrupt?
Honestly, what you are describing seems to be a use model that doesn't
fit KVM, which is a general purpose hypervisor, but more a simulation
environment. Yes, the primitives are the same, but the plumbing is
wildly different.
*If* that's the stuff you're looking at, then I'm afraid you'll have
to do it in different way, because what you are suggesting is
fundamentally incompatible with the guarantees that KVM gives to guest
and userspace. Because your KVM_ARCH_FLAG_WFX_EXIT_TO_USER is really a
lie. It should really be named something more along the lines of
KVM_ARCH_FLAG_WFX_EXIT_TO_USER_SOMETIME_AND_I_DONT_EVEN_KNOW_WHEN
(probably with additional clauses related to breaking things).
Overall, you are still asking for something that is not guaranteed at
the architecture level, even less in KVM, and I'm not going to add
support for something that can only work "sometime".
M.
--
Without deviation from the norm, progress is not possible.
Am Mo., 30. Okt. 2023 um 13:36 Uhr schrieb Marc Zyngier <[email protected]>:
>
> [please make an effort not to top-post]
>
> On Fri, 27 Oct 2023 18:41:44 +0100,
> Jan Henrik Weinstock <[email protected]> wrote:
> >
> > Hi Marc,
> >
> > the basic idea behind this is to have a (single-threaded) execution loop,
> > something like this:
> >
> > vcpu-thread: vcpu-run | process-io-devices | vcpu-run | process-io...
> > ^
> > WFX or timeout
> >
> > We switch to simulating IO devices whenever the vcpu is idle (wfi) or exceeds
> > a certain budget of instructions (counted via pmu). Our fallback currently is
> > to kick the vcpu out of its execution using a signal (via a timeout/alarm). But
> > of course, if the cpu is stuck at a wfi, we are wasting a lot of time.
> >
> > I understand that the proposed behavior is not desirable for most use cases,
> > which is why I suggest locking it behind a flag, e.g.
> > KVM_ARCH_FLAG_WFX_EXIT_TO_USER.
>
> But how do you reconcile the fact that exposing this to userspace
> breaks fundamental expectations that the guest has, such as getting
> its timer interrupts and directly injected LPIs? Implementing WFI in
> userspace breaks it. What about the case where we don't trap WFx and
> let the *guest* wait for an interrupt?
Timer interrupts etc. will be injected into the vcpu during the
io-phases. When there are no interrupts present and the guest performs
a WFI, we can just skip forward to the next timer event.
> Honestly, what you are describing seems to be a use model that doesn't
> fit KVM, which is a general purpose hypervisor, but more a simulation
> environment. Yes, the primitives are the same, but the plumbing is
> wildly different.
Agreed.
> *If* that's the stuff you're looking at, then I'm afraid you'll have
> to do it in different way, because what you are suggesting is
> fundamentally incompatible with the guarantees that KVM gives to guest
> and userspace. Because your KVM_ARCH_FLAG_WFX_EXIT_TO_USER is really a
> lie. It should really be named something more along the lines of
> KVM_ARCH_FLAG_WFX_EXIT_TO_USER_SOMETIME_AND_I_DONT_EVEN_KNOW_WHEN
> (probably with additional clauses related to breaking things).
I have attached a reworked version of the patch as a reference (based
on my 5.15 kernel). It puts the modified behavior behind a new
capability so as to not interfere with the current expectations
towards handling WFI/WFE.
I think it should now trap all blocking calls to WFx on the vcpu and
reliably return to the userspace. If I have missed something that
would cause the vcpu to not trap on a WFI kindly let me know.
> Overall, you are still asking for something that is not guaranteed at
> the architecture level, even less in KVM, and I'm not going to add
> support for something that can only work "sometime".
I am not quite sure what you mean with "sometime". Are you referring
to WFIs as NOPs? Or WFIs that do not yield because of pending
interrupts?
The point of my patch is not to accurately count every single WFI. The
point is to prevent the host cpu from sleeping just because my vcpu
executed a WFI somewhere in the guest software. If a WFI is executed
by the guest and that does not result in my vcpu thread to block (in
other words: the vcpu continues executing instructions beyond the WFI)
then it also should not exit to userspace. So instead of
"KVM_ARCH_FLAG_WFX_EXIT_TO_USER_SOMETIME_AND_I_DONT_EVEN_KNOW_WHEN" it
is really "KVM_ARCH_FLAG_WFX_EXIT_TO_USER_WHENEVER_YOU_WOULD_OTHERWISE_YIELD_AND_I_CANNOT_GET_MY_THREAD_BACK".
> M.
>
> --
> Without deviation from the norm, progress is not possible.
--
Dr.-Ing. Jan Henrik Weinstock
Managing Director
MachineWare GmbH | http://www.machineware.de
Hühnermarkt 19, 52062 Aachen, Germany
Amtsgericht Aachen HRB25734
Geschäftsführung
Lukas Jünger
Dr.-Ing. Jan Henrik Weinstock
On Tue, 31 Oct 2023 19:21:16 +0000,
Jan Henrik Weinstock <[email protected]> wrote:
>
> Am Mo., 30. Okt. 2023 um 13:36 Uhr schrieb Marc Zyngier <[email protected]>:
> >
> > [please make an effort not to top-post]
> >
> > On Fri, 27 Oct 2023 18:41:44 +0100,
> > Jan Henrik Weinstock <[email protected]> wrote:
> > >
> > > Hi Marc,
> > >
> > > the basic idea behind this is to have a (single-threaded) execution loop,
> > > something like this:
> > >
> > > vcpu-thread: vcpu-run | process-io-devices | vcpu-run | process-io...
> > > ^
> > > WFX or timeout
> > >
> > > We switch to simulating IO devices whenever the vcpu is idle (wfi) or exceeds
> > > a certain budget of instructions (counted via pmu). Our fallback currently is
> > > to kick the vcpu out of its execution using a signal (via a timeout/alarm). But
> > > of course, if the cpu is stuck at a wfi, we are wasting a lot of time.
> > >
> > > I understand that the proposed behavior is not desirable for most use cases,
> > > which is why I suggest locking it behind a flag, e.g.
> > > KVM_ARCH_FLAG_WFX_EXIT_TO_USER.
> >
> > But how do you reconcile the fact that exposing this to userspace
> > breaks fundamental expectations that the guest has, such as getting
> > its timer interrupts and directly injected LPIs? Implementing WFI in
> > userspace breaks it. What about the case where we don't trap WFx and
> > let the *guest* wait for an interrupt?
>
> Timer interrupts etc. will be injected into the vcpu during the
> io-phases. When there are no interrupts present and the guest performs
> a WFI, we can just skip forward to the next timer event.
Skip forward? What does that mean? Compress time and move along?
>
> > Honestly, what you are describing seems to be a use model that doesn't
> > fit KVM, which is a general purpose hypervisor, but more a simulation
> > environment. Yes, the primitives are the same, but the plumbing is
> > wildly different.
>
> Agreed.
>
> > *If* that's the stuff you're looking at, then I'm afraid you'll have
> > to do it in different way, because what you are suggesting is
> > fundamentally incompatible with the guarantees that KVM gives to guest
> > and userspace. Because your KVM_ARCH_FLAG_WFX_EXIT_TO_USER is really a
> > lie. It should really be named something more along the lines of
> > KVM_ARCH_FLAG_WFX_EXIT_TO_USER_SOMETIME_AND_I_DONT_EVEN_KNOW_WHEN
> > (probably with additional clauses related to breaking things).
>
> I have attached a reworked version of the patch as a reference (based
> on my 5.15 kernel). It puts the modified behavior behind a new
> capability so as to not interfere with the current expectations
> towards handling WFI/WFE.
> I think it should now trap all blocking calls to WFx on the vcpu and
> reliably return to the userspace. If I have missed something that
> would cause the vcpu to not trap on a WFI kindly let me know.
Oh FFS. Please read my previous emails, the architecture spec, and
understand that WFx is a *hint*. Given your line of work, I would hope
you understand the implications of this.
>
> > Overall, you are still asking for something that is not guaranteed at
> > the architecture level, even less in KVM, and I'm not going to add
> > support for something that can only work "sometime".
>
> I am not quite sure what you mean with "sometime". Are you referring
> to WFIs as NOPs? Or WFIs that do not yield because of pending
> interrupts?
NOP is a valid implementation of WFx. WFx doesn't have to trap. Its
only requirements are not to lose state. Nothing else. Trapping is a
'quality of implementation' feature, and doesn't affect correctness.
And yes, there are machines out there that will absolutely ignore any
request for trapping.
From the architecture spec (ARM DDI 0487J.a, D19.2.48, TWI):
<quote>
Since a WFI can complete at any time, even without a Wakeup event, the
traps on WFI are not guaranteed to be taken, even if the WFI is
executed when there is no Wakeup event. The only guarantee is that if
the instruction does not complete in finite time in the absence of a
Wakeup event, the trap will be taken.
</quote>
Similar verbiage exists for WFE. Do you now see why your proposal
makes little sense?
>
> The point of my patch is not to accurately count every single WFI. The
> point is to prevent the host cpu from sleeping just because my vcpu
> executed a WFI somewhere in the guest software. If a WFI is executed
> by the guest and that does not result in my vcpu thread to block (in
> other words: the vcpu continues executing instructions beyond the WFI)
> then it also should not exit to userspace. So instead of
> "KVM_ARCH_FLAG_WFX_EXIT_TO_USER_SOMETIME_AND_I_DONT_EVEN_KNOW_WHEN" it
> is really "KVM_ARCH_FLAG_WFX_EXIT_TO_USER_WHENEVER_YOU_WOULD_OTHERWISE_YIELD_AND_I_CANNOT_GET_MY_THREAD_BACK".
You already must be able to handle a guest spinning in a loop without
a WFI. So why would WFI be of interest more than anything else? You
can always make an interrupt pending at any point, without having to
wait for WFI to occur. Just make the interrupt pending (which, if you
emulate everything in userspace, is just giving the vcpu thread a
signal).
My hunch is that your SW is trying to do the interrupt injection from
the vcpu thread, which is a pretty broken model (it would badly model
the concept of an interrupt being an asynchronous event).
Honestly, if there was one thing I would add to the kernel, it would
be an option to *prevent* any trap of WFx, because that at least is
something we can universally enforce and guarantee to userspace.
Anything else is only wishful thinking.
M.
--
Without deviation from the norm, progress is not possible.