Re: [PATCH v2 0/7] kvm: fix latent guest entry/exit bugs

[Christian Borntraeger]

Am 20.01.22 um 16:14 schrieb Christian Borntraeger:
> Am 20.01.22 um 13:03 schrieb Mark Rutland:
> > On Thu, Jan 20, 2022 at 12:28:09PM +0100, Paolo Bonzini wrote:
> > > On 1/19/22 20:22, Mark Rutland wrote:
> > > > I wonder, is the s390 guest entry/exit*preemptible*  ?
> > > >
> > > > If a timer IRQ can preempt in the middle of the EQS, we wouldn't balance
> > > > things before a ctx-switch to the idle thread, which would then be able
> > > > to hit this.
> > > >
> > > > I'll need to go audit the other architectures for similar.
> > >
> > > They don't enable interrupts in the entry/exit path so they should be okay.
> >
> > True.
> >
> > So it sounds like for s390 adding an explicit preempt_{disable,enable}() is the
> > right thing to do. I'll add that and explanatory commentary.
>
> That would not be trivial I guess. We do allow (and need) page faults on sie for guest
> demand paging and
>
> this piece of arch/s390/mm/fault.c
>
>         case GMAP_FAULT:
>                  if (faulthandler_disabled() || !mm)
>                          goto out;
>                  break;
>          }
>
> would no longer work since faulthandler_disabled checks for the preempt count.


Something like this


diff --git a/arch/s390/mm/fault.c b/arch/s390/mm/fault.c
index d30f5986fa85..1c7d45346e12 100644
--- a/arch/s390/mm/fault.c
+++ b/arch/s390/mm/fault.c
@@ -385,10 +385,18 @@ static inline vm_fault_t do_exception(struct pt_regs *regs, int access)
                        return 0;
                goto out;
        case USER_FAULT:
-       case GMAP_FAULT:
                if (faulthandler_disabled() || !mm)
                        goto out;
                break;
+               /*
+                * We know that we interrupted SIE and we are not in a IRQ.
+                * preemption might be disabled thus checking for in_atomic
+                * would result in failures
+                */
+       case GMAP_FAULT:
+               if (pagefault_disabled() || !mm)
+                       goto out;
+               break;
        }
 
        perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);

seems to work with preemption disabled around sie. Not sure yet if this is correct.