[RFC PATCH v0 5/5] pseries: Asynchronous page fault support

Bharata B Rao bharata at linux.ibm.com
Fri Aug 13 14:54:39 AEST 2021


On Fri, Aug 13, 2021 at 02:06:40PM +1000, Nicholas Piggin wrote:
> Excerpts from Bharata B Rao's message of August 5, 2021 5:24 pm:
> > Add asynchronous page fault support for pseries guests.
> > 
> > 1. Setup the guest to handle async-pf
> >    - Issue H_REG_SNS hcall to register the SNS region.
> >    - Setup the subvention interrupt irq.
> >    - Enable async-pf by updating the byte_b9 of VPA for each
> >      CPU.
> > 2. Check if the page fault is an expropriation notification
> >    (SRR1_PROGTRAP set in SRR1) and if so put the task on
> >    wait queue based on the expropriation correlation number
> >    read from the VPA.
> > 3. Handle subvention interrupt to wake any waiting tasks.
> >    The wait and wakeup mechanism from x86 async-pf implementation
> >    is being reused here.
> 
> I don't know too much about the background of this.
> 
> How much benefit does this give? What situations?

I haven't yet gotten into measuring the benefit of this. Once
the patches are bit more stable than what they are currently,
we need to measure and evaluate the benefits.

> Does PowerVM implement it?

I suppose so, need to check though.

> Do other architectures KVM have something similar?

Yes, x86 and s390 KVM have had this feature for a while now
and generic KVM interfaces exist to support it.

> 
> The SRR1 setting for the DSI is in PAPR? In that case it should be okay,
> it might be good to add a small comment in exceptions-64s.S.

Yes, SRR1 setting is part of PAPR.

> 
> [...]
> 
> > @@ -395,6 +395,11 @@ static int ___do_page_fault(struct pt_regs *regs, unsigned long address,
> >  	vm_fault_t fault, major = 0;
> >  	bool kprobe_fault = kprobe_page_fault(regs, 11);
> >  
> > +#ifdef CONFIG_PPC_PSERIES
> > +	if (handle_async_page_fault(regs, address))
> > +		return 0;
> > +#endif
> > +
> >  	if (unlikely(debugger_fault_handler(regs) || kprobe_fault))
> >  		return 0;
> 
> [...]
> 
> > +int handle_async_page_fault(struct pt_regs *regs, unsigned long addr)
> > +{
> > +	struct async_pf_sleep_node n;
> > +	DECLARE_SWAITQUEUE(wait);
> > +	unsigned long exp_corr_nr;
> > +
> > +	/* Is this Expropriation notification? */
> > +	if (!(mfspr(SPRN_SRR1) & SRR1_PROGTRAP))
> > +		return 0;
> 
> Yep this should be an inline that is guarded by a static key, and then 
> probably have an inline check for SRR1_PROGTRAP. You shouldn't need to
> mfspr here, but just use regs->msr.

Right.

> 
> > +
> > +	if (unlikely(!user_mode(regs)))
> > +		panic("Host injected async PF in kernel mode\n");
> 
> Hmm. Is there anything in the PAPR interface that specifies that the
> OS can only deal with problem state access faults here? Or is that
> inherent in the expropriation feature?

Didn't see anything specific to that effect in PAPR. However since
this puts the faulting guest process to sleep until the page
becomes ready in the host, I have limited it to guest user space
faults.

Regards,
Bharata.


More information about the Linuxppc-dev mailing list