shadow_scb() calls prepare_cpuflags() first, which can reject the SCB provided by guest 2 with a validity interception. This happens before the guest state (gpsw, gg14, gg15, gcr, cputm, ckc, todpr, pp) and the control fields (gbea, fpf) are copied from the original SCB into the shadow SCB. The error path still calls unshadow_scb(), which copies that state from the shadow SCB back into the original one. The shadow SCB does not hold guest 3 state at this point: - For a new or recycled vsie page, get_vsie_page() has cleared the shadow SCB, so the guest 3 state in the guest 2 SCB is overwritten with zeroes. - For a cached vsie page that is looked up again by SCB address, the shadow SCB still holds the state from the end of the previous run, so the current guest 3 state is replaced by the old one. No guest 3 instruction ran, yet guest 2 sees its guest 3 state (PSW, GRs, CRs, timers) modified. unshadow_scb() also clears FPF_BPBC if facility 82 is available, because fpf was cleared in the shadow SCB just before, and copies back a stale ihcpu. The SCB and the validity condition are controlled by guest 2, so this is a correctness problem for guest 2 only and does not cross a guest or host boundary. Only the interception data needs to be passed on in this case. Add unshadow_icpt() for that and use it from unshadow_scb() and from the prepare_cpuflags() failure path, which now returns directly. The other failure paths of shadow_scb() are reached after the guest state has been copied and keep using unshadow_scb(). On a cached vsie page, the interception data in the shadow SCB still holds the values of the previous run. set_validity_icpt() writes ipa, ipb and icptcode, but not icptstatus, so a stale icptstatus would be passed on together with the new validity interception. Also, it is not defined what SIE stores in icptstatus, ipa and ipb when it exits without an interception, so they must not be relied upon to be zero or unchanged. Clear them together with icptcode when the shadow SCB is reused, so that no stale interception data is passed on to guest 2. Fixes: a3508fbe9dc6 ("KVM: s390: vsie: initial support for nested virtualization") Reported-by: Christian Borntraeger Signed-off-by: Gautam Gala Reviewed-by: Christian Borntraeger --- arch/s390/kvm/s390/vsie.c | 24 +++++++++++++++++------- 1 file changed, 17 insertions(+), 7 deletions(-) diff --git a/arch/s390/kvm/s390/vsie.c b/arch/s390/kvm/s390/vsie.c index 5926689fa226..1f618bb274a8 100644 --- a/arch/s390/kvm/s390/vsie.c +++ b/arch/s390/kvm/s390/vsie.c @@ -424,6 +424,15 @@ static void prepare_ibc(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page) } } +static void unshadow_icpt(struct kvm_s390_sie_block *scb_o, + struct kvm_s390_sie_block *scb_s) +{ + scb_o->icptcode = scb_s->icptcode; + scb_o->icptstatus = scb_s->icptstatus; + scb_o->ipa = scb_s->ipa; + scb_o->ipb = scb_s->ipb; +} + /* unshadow the scb, copying parameters back to the real scb */ static void unshadow_scb(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page) { @@ -431,10 +440,7 @@ static void unshadow_scb(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page) struct kvm_s390_sie_block *scb_o = vsie_page->scb_o; /* interception */ - scb_o->icptcode = scb_s->icptcode; - scb_o->icptstatus = scb_s->icptstatus; - scb_o->ipa = scb_s->ipa; - scb_o->ipb = scb_s->ipb; + unshadow_icpt(scb_o, scb_s); scb_o->gbea = scb_s->gbea; /* timer */ @@ -490,6 +496,9 @@ static int shadow_scb(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page) /* make sure we don't have any leftovers when reusing the scb */ scb_s->icptcode = 0; + scb_s->icptstatus = 0; + scb_s->ipa = 0; + scb_s->ipb = 0; scb_s->eca = 0; scb_s->ecb = 0; scb_s->ecb2 = 0; @@ -499,8 +508,10 @@ static int shadow_scb(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page) scb_s->fpf = 0; rc = prepare_cpuflags(vcpu, vsie_page); - if (rc) - goto out; + if (rc) { + unshadow_icpt(scb_o, scb_s); + return rc; + } /* timer */ scb_s->cputm = scb_o->cputm; @@ -614,7 +625,6 @@ static int shadow_scb(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page) prepare_ibc(vcpu, vsie_page); rc = shadow_crycb(vcpu, vsie_page); -out: if (rc) unshadow_scb(vcpu, vsie_page); return rc; -- 2.56.0