Re: [PATCH v2] x86/mce: Fix endless loop when run task works after #MC

[Ding Hui]

On 2021/7/7 11:39, Ding Hui wrote:
> On 2021/7/7 0:44, Luck, Tony wrote:
> > On Tue, Jul 06, 2021 at 08:16:06PM +0800, Ding Hui wrote:
> > > Recently we encounter multi #MC on the same task when it's
> > > task_work_run() has not been called, current->mce_kill_me was
> > > added to task_works list more than once, that make a circular
> > > linked task_works, so task_work_run() will do a endless loop.
> >
> > I saw the same and posted a similar fix a while back:
> >
> > https://www.spinics.net/lists/linux-mm/msg251006.html
> >
> > It didn't get merged because some validation tests began failing
> > around the same time.  I'm now pretty sure I understand what happened
> > with those other tests.
> >
> > I'll post my updated version (second patch in a three part series)
> > later today.
>
> Thanks for your fixes.
>
> After digging my original problem, maybe I find out why I met #MC flood.
>
> My test case:
> 1. run qemu-kvm guest VM, OS is memtest86+.iso
> 2. inject SRAR UE to VM memory and wait #MC
> When VM trigger #MC, I expect that qemu will receive SIGBUS signal ASAP, 
> and with the modifed qemu, I will kill VM.
>
> In this case, do_machine_check() maybe called by kvm_machine_check() in 
> vmx.c.
>
> Before [1], memory_failure() is called in do_machine_check(), so 
> TIF_SIGPENDING is set on due to SIGBUS signal, vcpu_run() checked the 
> pending singal, so return to qemu to handle SIGBUS.
>
> After [1], do_machine_check() only add task work but not send SIGBUS 
> directly, vcpu_run() will not break the for-loop because 
> vcpu_enter_guest() return 1 and not set TIF_SIGPENDING on, task works 
> never executed until sth else happen. So the kvm enter guest repeatedly 
> and the #MC is triggered repeatedly.

Sorry for my incorrect description.

I figure out that my test kernel is not the lastest, it's without [2] 
commit 72c3c0fe54a3 （"x86/kvm: Use generic xfer to guest work 
function"), so vcpu_run() only care about signal_pending but not 
TIF_NOTIFY_RESUME which set on in task_work_add().

After [2], #MC flood should not exist.

Also thank Thomas Gleixner.

> Can you consider to fix cases like this?
>
> And do you mind to give me some advice for my temporary workaround about 
> this #MC flood:
> I want to check the context of do_machine_check() is exception or kvm, 
> and fallback to call kill_me_xxx directly when in kvm context. (I 
> already tested simply and met my expection)

So ignore my ask, please.

> [1]: commit 5567d11c21a1 ("x86/mce: Send #MC singal from task work")


--
Thanks,
- Ding Hui