On 23.08.2018 17:43, Christian Borntraeger wrote: > > > On 08/22/2018 11:53 AM, David Hildenbrand wrote: >> When DATA exceptions and vector-processing exceptions (program interrupts) >> are injected, the DXC/VXC is also to be stored in the fpc, if AFP is >> enabled in CR0. >> >> This can happen inside KVM when reinjecting an interrupt during program >> interrupt intercepts. These are triggered for example when debugging the >> guest (concurrent PER events result in an intercept instead of an >> injection of such interrupts). >> >> Signed-off-by: David Hildenbrand <david@xxxxxxxxxx> >> --- >> >> Only compile-tested. > > I checked the Linux code (arch/s390/kernel/traps.c) and Linux uses the FPC (and > not the lowcore field) to decide about the signal (SIGFPE) and si_code. So we want > to have the correct DXC/VXC value. > > Now, I wrote a short test program that does > feenableexcept(FE_DIVBYZERO); > and a division by zero. > and attached gdb to that guest together with a breakpoint on the divide (and the instruction > after). > I get the pint exit for the instruction after (as it is suppressing) and at this point in > time the guest fpc already contains the correct DXC value. So you patch will certainly not > hurt, but it seems not necessary. Thanks for trying. Wonder if that is documented behavior or just works by pure luck. E.g. it would be interesting to see what other instructions do that usually don't touch the DXC, except when injecting an exception. E.g. CRT. But if you believe this is not needed, we can also drop it. (if ever somebody would want to inject from QEMU, he could also just set the fpc directly) > > Still trying to look further if I missed something. > >> >> arch/s390/include/asm/ctl_reg.h | 1 + >> arch/s390/kvm/interrupt.c | 8 ++++++++ >> 2 files changed, 9 insertions(+) >> >> diff --git a/arch/s390/include/asm/ctl_reg.h b/arch/s390/include/asm/ctl_reg.h >> index 4600453536c2..88f3f14baee9 100644 >> --- a/arch/s390/include/asm/ctl_reg.h >> +++ b/arch/s390/include/asm/ctl_reg.h >> @@ -11,6 +11,7 @@ >> #include <linux/const.h> >> >> #define CR0_CLOCK_COMPARATOR_SIGN _BITUL(63 - 10) >> +#define CR0_AFP_REGISTER_CONTROL _BITUL(63 - 45) >> #define CR0_EMERGENCY_SIGNAL_SUBMASK _BITUL(63 - 49) >> #define CR0_EXTERNAL_CALL_SUBMASK _BITUL(63 - 50) >> #define CR0_CLOCK_COMPARATOR_SUBMASK _BITUL(63 - 52) >> diff --git a/arch/s390/kvm/interrupt.c b/arch/s390/kvm/interrupt.c >> index fcb55b02990e..5b5754d8f460 100644 >> --- a/arch/s390/kvm/interrupt.c >> +++ b/arch/s390/kvm/interrupt.c >> @@ -765,6 +765,14 @@ static int __must_check __deliver_prog(struct kvm_vcpu *vcpu) >> break; >> case PGM_VECTOR_PROCESSING: >> case PGM_DATA: >> + if (vcpu->arch.sie_block->gcr[0] & CR0_AFP_REGISTER_CONTROL) { >> + /* make sure the new fpc will be lazily loaded */ >> + save_fpu_regs(); >> + /* the DXC/VXC cannot make the fpc invalid */ >> + current->thread.fpu.fpc &= ~0xff00u; >> + current->thread.fpu.fpc |= (pgm_info.data_exc_code << 8) >> + & 0xff00u; > > maybe reuse FPC_DXC_MASK instead of 0xff00 ? > Sure, didn't know about that. -- Thanks, David / dhildenb