[RFC PATCH 1/4] ARM: KVM: on unhandled IO mem abort, route the call to the KVM MMIO bus

Nikolay Nikolaev n.nikolaev at virtualopensystems.com
Thu Nov 13 02:45:42 PST 2014


On Mon, Nov 10, 2014 at 6:27 PM, Christoffer Dall
<christoffer.dall at linaro.org> wrote:
> On Mon, Nov 10, 2014 at 05:09:07PM +0200, Nikolay Nikolaev wrote:
>> Hello,
>>
>> On Fri, Mar 28, 2014 at 9:09 PM, Christoffer Dall
>> <christoffer.dall at linaro.org> wrote:
>> >
>> > On Thu, Mar 13, 2014 at 04:57:26PM +0100, Antonios Motakis wrote:
>> > > On an unhandled IO memory abort, use the kvm_io_bus_* API in order to
>> > > handle the MMIO access through any registered read/write callbacks. This
>> > > is a dependency for eventfd support (ioeventfd and irqfd).
>> > >
>> > > However, accesses to the VGIC are still left implemented independently,
>> > > since the kvm_io_bus_* API doesn't pass the VCPU pointer doing the access.
>> > >
>> > > Signed-off-by: Antonios Motakis <a.motakis at virtualopensystems.com>
>> > > Signed-off-by: Nikolay Nikolaev <n.nikolaev at virtualopensystems.com>
>> > > ---
>> > >  arch/arm/kvm/mmio.c | 32 ++++++++++++++++++++++++++++++++
>> > >  virt/kvm/arm/vgic.c |  5 ++++-
>> > >  2 files changed, 36 insertions(+), 1 deletion(-)
>> > >
>> > > diff --git a/arch/arm/kvm/mmio.c b/arch/arm/kvm/mmio.c
>> > > index 4cb5a93..1d17831 100644
>> > > --- a/arch/arm/kvm/mmio.c
>> > > +++ b/arch/arm/kvm/mmio.c
>> > > @@ -162,6 +162,35 @@ static int decode_hsr(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
>> > >       return 0;
>> > >  }
>> > >
>> > > +/**
>> > > + * kvm_handle_mmio - handle an in-kernel MMIO access
>> > > + * @vcpu:    pointer to the vcpu performing the access
>> > > + * @run:     pointer to the kvm_run structure
>> > > + * @mmio:    pointer to the data describing the access
>> > > + *
>> > > + * returns true if the MMIO access has been performed in kernel space,
>> > > + * and false if it needs to be emulated in user space.
>> > > + */
>> > > +static bool handle_kernel_mmio(struct kvm_vcpu *vcpu, struct kvm_run *run,
>> > > +             struct kvm_exit_mmio *mmio)
>> > > +{
>> > > +     int ret;
>> > > +     if (mmio->is_write) {
>> > > +             ret = kvm_io_bus_write(vcpu->kvm, KVM_MMIO_BUS, mmio->phys_addr,
>> > > +                             mmio->len, &mmio->data);
>> > > +
>> > > +     } else {
>> > > +             ret = kvm_io_bus_read(vcpu->kvm, KVM_MMIO_BUS, mmio->phys_addr,
>> > > +                             mmio->len, &mmio->data);
>> > > +     }
>> > > +     if (!ret) {
>> > > +             kvm_prepare_mmio(run, mmio);
>> > > +             kvm_handle_mmio_return(vcpu, run);
>> > > +     }
>> > > +
>> > > +     return !ret;
>> > > +}
>> > > +
>> > >  int io_mem_abort(struct kvm_vcpu *vcpu, struct kvm_run *run,
>> > >                phys_addr_t fault_ipa)
>> > >  {
>> > > @@ -200,6 +229,9 @@ int io_mem_abort(struct kvm_vcpu *vcpu, struct kvm_run *run,
>> > >       if (vgic_handle_mmio(vcpu, run, &mmio))
>> > >               return 1;
>> > >
>> > > +     if (handle_kernel_mmio(vcpu, run, &mmio))
>> > > +             return 1;
>> > > +
>>
>>
>> We're reconsidering ioeventfds patchseries and we tried to evaluate
>> what you suggested here.
>>
>> >
>> > this special-casing of the vgic is now really terrible.  Is there
>> > anything holding you back from doing the necessary restructure of the
>> > kvm_bus_io_*() API instead?
>>
>> Restructuring the kvm_io_bus_ API is not a big thing (we actually did
>> it), but is not directly related to the these patches.
>> Of course it can be justified if we do it in the context of removing
>> vgic_handle_mmio and leaving only handle_kernel_mmio.
>>
>> >
>> > That would allow us to get rid of the ugly
>> > Fix it! in the vgic driver as well.
>>
>> Going through the vgic_handle_mmio we see that it will require large
>> refactoring:
>>  - there are 15 MMIO ranges for the vgic now - each should be
>> registered as a separate device
>>  - the handler of each range should be split into read and write
>>  - all handlers take 'struct kvm_exit_mmio', and pass it to
>> 'vgic_reg_access', 'mmio_data_read' and 'mmio_data_read'
>>
>> To sum up - if we do this refactoring of vgic's MMIO handling +
>> kvm_io_bus_ API getting 'vcpu" argument we'll get a 'much' cleaner
>> vgic code and as a bonus we'll get 'ioeventfd' capabilities.
>>
>> We have 3 questions:
>>  - is the kvm_io_bus_ getting 'vcpu' argument acceptable for the other
>> architectures too?
>>  - is this huge vgic MMIO handling redesign acceptable/desired (it
>> touches a lot of code)?
>>  - is there a way that ioeventfd is accepted leaving vgic.c in it's
>> current state?
>>
> Not sure how the latter question is relevant to this, but check with
> Andre who recently looked at this as well and decided that for GICv3 the
> only sane thing was to remove that comment for the gic.
@Andre - what's your experience with the GICv3 and MMIO handling,
anything specific?
>
> I don't recall the details of what you were trying to accomplish here
> (it's been 8 months or so) but the surely the vgic handling code should
> *somehow* be integrated into the handle_kernel_mmio (like Paolo
> suggested), unless you come back and tell me that that would involve a
> complete rewrite of the vgic code.
I'm experimenting now - it's not exactly rewrite of whole vgic code,
but it will touch a lot of it  - all MMIO access handlers and the
supporting functions.
We're ready to spend the effort. My question is  - is this acceptable?

regards,
Nikolay Nikolaev
Virtual Open Systems
>
> -Christoffer



More information about the linux-arm-kernel mailing list