Re: [PATCH] KVM: s390: vsie: fix crycb virtual vs physical usage

From: Pierre Morel
Date: Mon Sep 12 2022 - 04:03:24 EST



Hi David,

sorry for the delay, just came back from vacation.

On 9/5/22 18:32, David Hildenbrand wrote:
On 05.09.22 10:41, Pierre Morel wrote:
Prepare VSIE for architectural changes where lowmem kernel real and
kernel virtual address are different.

Bear with me, it used to be

    crycb = (struct kvm_s390_crypto_cb *) (unsigned long)crycb_o;
    apcb_o = (unsigned long) &crycb->apcb0;

and now it's

    apcb_o = crycb_o + offsetof(struct kvm_s390_crypto_cb, apcb0);


So the real issue seems to be

    crycb = (struct kvm_s390_crypto_cb *) (unsigned long)crycb_o;

because crycb_o actually is a guest address and not a host address.

Yes, right this is the real issue, however...



But now I'm confused, because I would have thought that the result produced by both code would be identical (I completely agree that the new variant is better).

How does this interact with "lowmem kernel real and kernel virtual address are different." -- I would have thought that &crycb->apcb0 doesn't actually access any memory and only performs arithmetical operations?

...you are right and the result is identical.




When we get the original crycb from the guest crycb we can use the
phys_to_virt transformation, which will use the host transformations,
but we must use an offset to calculate the guest real address apcb
and give it to read_guest_real().

Can you elaborate where phys_to_virt() comes into play?

No, it does not have to do with phys_to_virt(), I first started to work on the line:
- crycb = (struct kvm_s390_crypto_cb *) (unsigned long)crycb_o;

and did not notice that after the simplifications the result was identical.
So The comment is wrong and this patch is only making the code clearer.

Thanks for your clarifications,

Regards,
Pierre



If this is an actual fix (as indicated in the patch subject), should this carry a

    Fixes: 56019f9aca22 ("KVM: s390: vsie: Allow CRYCB FORMAT-2")


--
Pierre Morel
IBM Lab Boeblingen