Re: [Xen-devel] [PATCH v3] xen: get rid of paravirt op adjust_exception_frame

[Andrew Cooper]

On 08/08/17 08:02, Juergen Gross wrote:
> On 07/08/17 22:56, Boris Ostrovsky wrote:
>>> diff --git a/arch/x86/xen/enlighten_pv.c b/arch/x86/xen/enlighten_pv.c
>>> index 811e4ddb3f37..a3dcd83187ce 100644
>>> --- a/arch/x86/xen/enlighten_pv.c
>>> +++ b/arch/x86/xen/enlighten_pv.c
>>> @@ -579,6 +579,71 @@ static void xen_write_ldt_entry(struct desc_struct *dt, int entrynum,
>>>  	preempt_enable();
>>>  }
>>>  
>>> +#ifdef CONFIG_X86_64
>>> +static struct {
>>> +	void (*orig)(void);
>>> +	void (*xen)(void);
>>> +	bool ist_okay;
>>> +	bool handle;
>>> +} trap_array[] = {
>>> +	{ debug, xen_xendebug, true, true },
>>> +	{ int3, xen_xenint3, true, true },
>>> +	{ double_fault, xen_double_fault, true, false },
>> Is it really worth adding 'handle' member to the structure because of a
>> single special case? We don't expect to ever have another such vector.
> Hmm, maybe you are right. We don't expect to ever see a double_fault in
> a pv domain, so we could just drop that special case by handling it like
> the other IST traps.

(This is steeped in a lot of history.)  There is no path where Xen will
raise #DF with a PV guest.

As Linux sets the DPL of the #DF handler to 0, the `int $8` emulation
will inject #GP (even for kernel uses!) rather than follow the #DF path.

I do however want to see about making Xen's behaviour rather more
architectural, e.g. raising #NP rather than crashing the guest.  If we
do get to a position of properly using contributory exceptions, #DF will
behave as #MC and #NMI currently do, by either switching from userspace
to the kernel stack, or pushed normally onto the current stack.

In some copious free time, it might also be good to invent a PV ABI to
behave like IST (even if only for the failsafe callback) so Linux can
actually recover sufficiently from a stack overflow to print some
diagnostics.

~Andrew