On 07.08.2018 14:51, David Hildenbrand wrote:
> VCPU requests and VCPU blocking right now don't take care of the vSIE
> (as it was not necessary until now). But we want to have VCPU requests
> that will also be handled before running the vSIE again.
>
> So let's simulate a SIE entry when entering the vSIE loop and check
> for PROG_ flags. The existing infrastructure (e.g. exit_sie()) will then
> detect that the SIE (in form of the vSIE execution loop) is running and
> properly kick the vSIE CPU, resulting in it leaving the vSIE loop and
> therefore the vSIE interception handler, allowing it to handle VCPU
> requests.
>
> E.g. if we want to modify the crycb of the VCPU and make sure that any
> masks also get applied to the VSIE crycb shadow (which uses masks from the
> VCPU crycb), we will need a way to hinder the vSIE from running and make
> sure to process the updated crycb before reentering the vSIE again.
>
> Signed-off-by: David Hildenbrand <da...@redhat.com>
> ---
> arch/s390/kvm/kvm-s390.c | 9 ++++++++-
> arch/s390/kvm/kvm-s390.h | 1 +
> arch/s390/kvm/vsie.c | 20 ++++++++++++++++++--
> 3 files changed, 27 insertions(+), 3 deletions(-)
>
> diff --git a/arch/s390/kvm/kvm-s390.c b/arch/s390/kvm/kvm-s390.c
> index 91ad4a9425c0..c87734a31fdb 100644
> --- a/arch/s390/kvm/kvm-s390.c
> +++ b/arch/s390/kvm/kvm-s390.c
> @@ -2766,18 +2766,25 @@ static void kvm_s390_vcpu_request(struct kvm_vcpu
> *vcpu)
> exit_sie(vcpu);
> }
>
> +bool kvm_s390_vcpu_sie_inhibited(struct kvm_vcpu *vcpu)
> +{
> + return atomic_read(&vcpu->arch.sie_block->prog20) &
> + (PROG_BLOCK_SIE | PROG_REQUEST);
> +}
> +
> static void kvm_s390_vcpu_request_handled(struct kvm_vcpu *vcpu)
> {
> atomic_andnot(PROG_REQUEST, &vcpu->arch.sie_block->prog20);
> }
>
> /*
> - * Kick a guest cpu out of SIE and wait until SIE is not running.
> + * Kick a guest cpu out of (v)SIE and wait until (v)SIE is not running.
> * If the CPU is not running (e.g. waiting as idle) the function will
> * return immediately. */
> void exit_sie(struct kvm_vcpu *vcpu)
> {
> kvm_s390_set_cpuflags(vcpu, CPUSTAT_STOP_INT);
> + kvm_s390_vsie_kick(vcpu);
> while (vcpu->arch.sie_block->prog0c & PROG_IN_SIE)
> cpu_relax();
> }
> diff --git a/arch/s390/kvm/kvm-s390.h b/arch/s390/kvm/kvm-s390.h
> index 981e3ba97461..1f6e36cdce0d 100644
> --- a/arch/s390/kvm/kvm-s390.h
> +++ b/arch/s390/kvm/kvm-s390.h
> @@ -290,6 +290,7 @@ void kvm_s390_vcpu_start(struct kvm_vcpu *vcpu);
> void kvm_s390_vcpu_stop(struct kvm_vcpu *vcpu);
> void kvm_s390_vcpu_block(struct kvm_vcpu *vcpu);
> void kvm_s390_vcpu_unblock(struct kvm_vcpu *vcpu);
> +bool kvm_s390_vcpu_sie_inhibited(struct kvm_vcpu *vcpu);
> void exit_sie(struct kvm_vcpu *vcpu);
> void kvm_s390_sync_request(int req, struct kvm_vcpu *vcpu);
> int kvm_s390_vcpu_setup_cmma(struct kvm_vcpu *vcpu);
> diff --git a/arch/s390/kvm/vsie.c b/arch/s390/kvm/vsie.c
> index 63844b95c22c..faac06886f77 100644
> --- a/arch/s390/kvm/vsie.c
> +++ b/arch/s390/kvm/vsie.c
> @@ -989,6 +989,17 @@ static int vsie_run(struct kvm_vcpu *vcpu, struct
> vsie_page *vsie_page)
> struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
> int rc = 0;
>
> + /*
> + * Simulate a SIE entry of the VCPU (see sie64a), so VCPU blocking
> + * and VCPU requests can hinder the whole vSIE loop from running
> + * and lead to an immediate exit. We do it at this point (not
> + * earlier), so kvm_s390_vsie_kick() works correctly already.
> + */
> + vcpu->arch.sie_block->prog0c |= PROG_IN_SIE;
> + barrier();
> + if (kvm_s390_vcpu_sie_inhibited(vcpu))
> + return 0;
> +
> while (1) {
> rc = acquire_gmap_shadow(vcpu, vsie_page);
> if (!rc)
> @@ -1004,10 +1015,14 @@ static int vsie_run(struct kvm_vcpu *vcpu, struct
> vsie_page *vsie_page)
> if (rc == -EAGAIN)
> rc = 0;
> if (rc || scb_s->icptcode || signal_pending(current) ||
> - kvm_s390_vcpu_has_irq(vcpu, 0))
> + kvm_s390_vcpu_has_irq(vcpu, 0) ||
> + kvm_s390_vcpu_sie_inhibited(vcpu))
> break;
> }
>
> + barrier();
> + vcpu->arch.sie_block->prog0c &= ~PROG_IN_SIE;
> +
I am thinking about moving this down to the actual sie64 call. We
eventually take locks and even call into MM code (to resolve faults)
inside do_vsie_run(). I think this extra overhead can be avoided (where
any caller - e.g. on prefix unmaps has to wait).
--
Thanks,
David / dhildenb
