Re: [PATCH 0/5] kvm: fix latent guest entry/exit bugs

Re: [PATCH 0/5] kvm: fix latent guest entry/exit bugs

[Christian Borntraeger]

Am 18.01.22 um 14:12 schrieb Mark Rutland:
> On Tue, Jan 18, 2022 at 01:42:26PM +0100, Christian Borntraeger wrote:
>>
>>
>> Am 18.01.22 um 13:02 schrieb Mark Rutland:
>>> On Mon, Jan 17, 2022 at 06:45:36PM +0100, Paolo Bonzini wrote:
>>>> On 1/14/22 16:19, Mark Rutland wrote:
>>>>> I also think there is another issue here. When an IRQ is taken from SIE, will
>>>>> user_mode(regs) always be false, or could it be true if the guest userspace is
>>>>> running? If it can be true I think tha context tracking checks can complain,
>>>>> and it*might*  be possible to trigger a panic().
>>>>
>>>> I think that it would be false, because the guest PSW is in the SIE block
>>>> and switched on SIE entry and exit, but I might be incorrect.
>>>
>>> Ah; that's the crux of my confusion: I had thought the guest PSW would
>>> be placed in the regular lowcore *_old_psw slots. From looking at the
>>> entry asm it looks like the host PSW (around the invocation of SIE) is
>>> stored there, since that's what the OUTSIDE + SIEEXIT handling is
>>> checking for.
>>>
>>> Assuming that's correct, I agree this problem doesn't exist, and there's
>>> only the common RCU/tracing/lockdep management to fix.
>>
>> Will you provide an s390 patch in your next iteration or shall we then do
>> one as soon as there is a v2? We also need to look into vsie.c where we
>> also call sie64a
> 
> I'm having a go at that now; my plan is to try to have an s390 patch as
> part of v2 in the next day or so.
> 
> Now that I have a rough idea of how SIE and exception handling works on
> s390, I think the structural changes to kvm-s390.c:__vcpu_run() and
> vsie.c:do_vsie_run() are fairly simple.
> 
> The only open bit is exactly how/where to identify when the interrupt
> entry code needs to wake RCU. I can add a per-cpu variable or thread
> flag to indicate that we're inside that EQS, or or I could move the irq
> enable/disable into the sie64a asm and identify that as with the OUTSIDE
> macro in the entry asm.
What exactly would the low-level interrupt handler need to do?

CC Sven, Heiko for the entry.S changes.

Re: [PATCH 0/5] kvm: fix latent guest entry/exit bugs

[Mark Rutland]

On Tue, Jan 18, 2022 at 03:15:51PM +0100, Christian Borntraeger wrote:
> Am 18.01.22 um 14:12 schrieb Mark Rutland:
> > On Tue, Jan 18, 2022 at 01:42:26PM +0100, Christian Borntraeger wrote:
> > > 
> > > 
> > > Am 18.01.22 um 13:02 schrieb Mark Rutland:
> > > > On Mon, Jan 17, 2022 at 06:45:36PM +0100, Paolo Bonzini wrote:
> > > > > On 1/14/22 16:19, Mark Rutland wrote:
> > > > > > I also think there is another issue here. When an IRQ is taken from SIE, will
> > > > > > user_mode(regs) always be false, or could it be true if the guest userspace is
> > > > > > running? If it can be true I think tha context tracking checks can complain,
> > > > > > and it*might*  be possible to trigger a panic().
> > > > > 
> > > > > I think that it would be false, because the guest PSW is in the SIE block
> > > > > and switched on SIE entry and exit, but I might be incorrect.
> > > > 
> > > > Ah; that's the crux of my confusion: I had thought the guest PSW would
> > > > be placed in the regular lowcore *_old_psw slots. From looking at the
> > > > entry asm it looks like the host PSW (around the invocation of SIE) is
> > > > stored there, since that's what the OUTSIDE + SIEEXIT handling is
> > > > checking for.
> > > > 
> > > > Assuming that's correct, I agree this problem doesn't exist, and there's
> > > > only the common RCU/tracing/lockdep management to fix.
> > > 
> > > Will you provide an s390 patch in your next iteration or shall we then do
> > > one as soon as there is a v2? We also need to look into vsie.c where we
> > > also call sie64a
> > 
> > I'm having a go at that now; my plan is to try to have an s390 patch as
> > part of v2 in the next day or so.
> > 
> > Now that I have a rough idea of how SIE and exception handling works on
> > s390, I think the structural changes to kvm-s390.c:__vcpu_run() and
> > vsie.c:do_vsie_run() are fairly simple.
> > 
> > The only open bit is exactly how/where to identify when the interrupt
> > entry code needs to wake RCU. I can add a per-cpu variable or thread
> > flag to indicate that we're inside that EQS, or or I could move the irq
> > enable/disable into the sie64a asm and identify that as with the OUTSIDE
> > macro in the entry asm.
> What exactly would the low-level interrupt handler need to do?

Having looked around a bit, I think the best bet is to have
irqentry_enter() check PF_VCPU in addition to PF_IDLE (which it checks
via is_idle_task()), at which point nothing needs to change in the s390
entry code.

I'm currently implementing that, let me have a go, and then we can see
if that looks ok or whether we should do something else.

> CC Sven, Heiko for the entry.S changes.

I'll make sure you're all Cc'd when I send out vs with s390 patches.

Thanks,
Mark.