On Fri, Apr 08, 2016 at 05:50:03PM +0100, Andre Przywara wrote:
> Hi,
>
> On 08/04/16 12:05, Christoffer Dall wrote:
> > On Fri, Apr 08, 2016 at 11:43:50AM +0100, Andre Przywara wrote:
> >> Currently we call kvm_handle_mmio_return when we need to sync back
> >> register content into the guest after a trap.
> >> This function expects its arguments packaged into struct kvm_run,
> >> which we only naturally use when we emulate in userspace. With
> >> in-kernel emulation we have to copy the data into that strcut.
> >
> > s/strcut/struct/
> >
> >> This patch fixes this by explicitly passing the required variables,
> >> so we save the copying in two cases.
> >
> > this patch fixes what?
>
> The strcut ;-)
>
> It fixes the unnecessary copying. Indeed worded badly.
>
> >> Also since this function is actually a NOP for an MMIO write, we
> >> rename it to kvm_writeback_mmio_data and move the !is_write check to
> >> the callers.
> >> This fixes a bug where we were missing a writeback for one in-kernel
> >> emulation case.
> >>
> >> Signed-off-by: Andre Przywara <andre.przywara@xxxxxxx>
> >> ---
> >> Hi Christoffer,
> >>
> >> this is my take on fixing the missing MMIO writeback call.
> >> This rework kind of hides the actual bug-fix, that's why I dumped
> >> it in favour of the one-liner in my series.
> >>
> >
> > Hmmm, I thought I also sent a patch for this (off list), and I'm not
> > sure why your approach is better, but ok, I guess I can review your
> > patch...
>
> My approach is not better, just different. I just wanted some opinions
> because I wasn't satisfied either.
I just felt like I had implemented and tested a fix for a problem, and
then I also had to review an alternative approach, but oh well, no
worries.
> This run-> wrapping did annoy me already some months ago, so I just took
> the opportunity to kill it.
>
> >>
> >> arch/arm/include/asm/kvm_mmio.h | 3 ++-
> >> arch/arm/kvm/arm.c | 10 +++++---
> >> arch/arm/kvm/mmio.c | 54 ++++++++++++++++++---------------------
> >> arch/arm64/include/asm/kvm_mmio.h | 3 ++-
> >> virt/kvm/arm/vgic.c | 8 ++----
> >> 5 files changed, 38 insertions(+), 40 deletions(-)
> >>
> >> diff --git a/arch/arm/include/asm/kvm_mmio.h b/arch/arm/include/asm/kvm_mmio.h
> >> index d8e90c8..21f97ac 100644
> >> --- a/arch/arm/include/asm/kvm_mmio.h
> >> +++ b/arch/arm/include/asm/kvm_mmio.h
> >> @@ -28,7 +28,8 @@ struct kvm_decode {
> >> bool sign_extend;
> >> };
> >>
> >> -int kvm_handle_mmio_return(struct kvm_vcpu *vcpu, struct kvm_run *run);
> >> +int kvm_writeback_mmio_data(struct kvm_vcpu *vcpu, void *data, int len,
> >
> > I think this should be called kvm_write_back_mmio_data() instead.
> >
> >> + gpa_t addr);
> >> int io_mem_abort(struct kvm_vcpu *vcpu, struct kvm_run *run,
> >> phys_addr_t fault_ipa);
> >>
> >> diff --git a/arch/arm/kvm/arm.c b/arch/arm/kvm/arm.c
> >> index b538431..e9f593c 100644
> >> --- a/arch/arm/kvm/arm.c
> >> +++ b/arch/arm/kvm/arm.c
> >> @@ -555,9 +555,13 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu, struct kvm_run *run)
> >> return ret;
> >>
> >> if (run->exit_reason == KVM_EXIT_MMIO) {
> >> - ret = kvm_handle_mmio_return(vcpu, vcpu->run);
> >> - if (ret)
> >> - return ret;
> >> + if (!run->mmio.is_write) {
> >> + ret = kvm_writeback_mmio_data(vcpu, run->mmio.data,
> >> + run->mmio.len,
> >> + run->mmio.phys_addr);
> >> + if (ret)
> >> + return ret;
> >> + }
> >
> > I preferred all the logic be encapsulated in the function instead of
> > extracting values in the caller, but it's purely a cosmetic comment.
>
> You mean the "if (!is_write)" or wrapping the values into the kvm_run
> struct?
> As mentioned in the commit message, only actual userland emulation is
> using run->, so for in-kernel emulation we have to set it up, which is
> unnecessary copying IMHO.
> By un-wrapping the parameters we just would pass is_write to exit the
> function immediately if it was false, that's why the move to the callers.
> But I agree it is probably bike shedding.
>
> >> }
> >>
> >> if (vcpu->sigset_active)
> >> diff --git a/arch/arm/kvm/mmio.c b/arch/arm/kvm/mmio.c
> >> index 0f6600f..052aa02 100644
> >> --- a/arch/arm/kvm/mmio.c
> >> +++ b/arch/arm/kvm/mmio.c
> >> @@ -86,38 +86,33 @@ static unsigned long mmio_read_buf(char *buf, unsigned int len)
> >> }
> >>
> >> /**
> >> - * kvm_handle_mmio_return -- Handle MMIO loads after user space emulation
> >> + * kvm_writeback_mmio_data -- Write back emulation data into the guest's
> >> + * target register after return from userspace.
> >> * @vcpu: The VCPU pointer
> >> - * @run: The VCPU run struct containing the mmio data
> >> + * @data: A pointer to the data to be written back
> >> + * @len: The size of the read access
> >> + * @addr: The original MMIO address (for the tracepoint only)
> >> *
> >> * This should only be called after returning from userspace for MMIO load
> >> * emulation.
> >> */
> >> -int kvm_handle_mmio_return(struct kvm_vcpu *vcpu, struct kvm_run *run)
> >> +int kvm_writeback_mmio_data(struct kvm_vcpu *vcpu, void *data_ptr, int len,
> >> + gpa_t addr)
> >> {
> >> unsigned long data;
> >> - unsigned int len;
> >> int mask;
> >>
> >> - if (!run->mmio.is_write) {
> >> - len = run->mmio.len;
> >> - if (len > sizeof(unsigned long))
> >> - return -EINVAL;
> >> + data = mmio_read_buf(data_ptr, len);
> >>
> >> - data = mmio_read_buf(run->mmio.data, len);
> >> -
> >> - if (vcpu->arch.mmio_decode.sign_extend &&
> >> - len < sizeof(unsigned long)) {
> >> - mask = 1U << ((len * 8) - 1);
> >> - data = (data ^ mask) - mask;
> >> - }
> >> -
> >> - trace_kvm_mmio(KVM_TRACE_MMIO_READ, len, run->mmio.phys_addr,
> >> - data);
> >> - data = vcpu_data_host_to_guest(vcpu, data, len);
> >> - vcpu_set_reg(vcpu, vcpu->arch.mmio_decode.rt, data);
> >> + if (vcpu->arch.mmio_decode.sign_extend && len < sizeof(unsigned long)) {
> >> + mask = 1U << ((len * 8) - 1);
> >> + data = (data ^ mask) - mask;
> >> }
> >>
> >> + trace_kvm_mmio(KVM_TRACE_MMIO_READ, len, addr, data);
> >> + data = vcpu_data_host_to_guest(vcpu, data, len);
> >> + vcpu_set_reg(vcpu, vcpu->arch.mmio_decode.rt, data);
> >> +
> >> return 0;
> >> }
> >>
> >> @@ -202,22 +197,23 @@ int io_mem_abort(struct kvm_vcpu *vcpu, struct kvm_run *run,
> >> data_buf);
> >> }
> >>
> >> + if (!ret) {
> >> + /* We handled the access successfully in the kernel. */
> >> + vcpu->stat.mmio_exit_kernel++;
> >> + if (!is_write)
> >> + kvm_writeback_mmio_data(vcpu, data_buf, len, fault_ipa);
> >
> > are you not doing this writeback twice? Once in the vgic and once here?
> > That was the very thing I tried to address in my patch.
>
> Right, I missed that one (which was the actual reason the code worked
> before although we missed the write-back.
> Not sure why it was actually in in the first place.
>
> I guess I will just go with your patch, because it's simpler.
>
I do agree about the weirdness of using the mmio struct designed for
userspace in the kernel. On the other hand, that struct encapsulates
all that's needed, which why I think the code was done the way it is
currently.
I think we originally considered only the vgic as an in-kernel thing
that needed to do the write-back, hence the 'bug' or situation or
whatever we have today.
I don't care strongly which way we go about it. If you prefer building
the new vgic on top of your patch, then just address the double-copying
in the gic. It's really up to you, whatever makes the shitty job of
managing the giant new vgic implementation for you.
Thanks,
-Christoffer
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
