Re: [PATCH v3] xen/balloon: add late_initcall_sync() for initial ballooning done

[Juergen Gross]

On 30.10.21 01:44, Boris Ostrovsky wrote:
> On 10/29/21 6:18 PM, Marek Marczykowski-Górecki wrote:
> > On Fri, Oct 29, 2021 at 05:46:18PM -0400, Boris Ostrovsky wrote:
> > > On 10/29/21 10:20 AM, Juergen Gross wrote:
> > > > --- a/Documentation/ABI/stable/sysfs-devices-system-xen_memory
> > > > +++ b/Documentation/ABI/stable/sysfs-devices-system-xen_memory
> > > > @@ -84,3 +84,13 @@ Description:
> > > >            Control scrubbing pages before returning them to Xen for 
> > > > others domains
> > > >            use. Can be set with xen_scrub_pages cmdline
> > > >            parameter. Default value controlled with 
> > > > CONFIG_XEN_SCRUB_PAGES_DEFAULT.
> > > > +
> > > > +What:        /sys/devices/system/xen_memory/xen_memory0/boot_timeout
> > > > +Date:        November 2021
> > > > +KernelVersion:    5.16
> > > > +Contact:    xen-devel@xxxxxxxxxxxxxxxxxxxx
> > > > +Description:
> > > > +        The time (in seconds) to wait before giving up to boot in case
> > > > +        initial ballooning fails to free enough memory. Applies only
> > > > +        when running as HVM or PVH guest and started with less memory
> > > > +        configured than allowed at max.
> > >
> > > How is this going to be used? We only need this during boot.
> > >
> > >
> > > > -        state = update_schedule(state);
> > > > +        balloon_state = update_schedule(balloon_state);
> > >
> > > Now that balloon_state has whole file scope it can probably be 
> > > updated inside update_schedule().
> > >
> > >
> > > > +    while ((credit = current_credit()) < 0) {
> > > > +        if (credit != last_credit) {
> > > > +            last_changed = jiffies;
> > > > +            last_credit = credit;
> > > > +        }
> > > > +        if (balloon_state == BP_ECANCELED) {
> > >
> > > What about other states? We are really waiting for BP_DONE, aren't we?
> > BP_DONE is set also as an intermediate step:
> >
> >                         balloon_state = decrease_reservation(n_pages,
> >                                                              
> > GFP_BALLOON);
> >                         if (balloon_state == BP_DONE && n_pages != 
> > -credit &&
> >                              n_pages < totalreserve_pages)
> >                                 balloon_state = BP_EAGAIN;
> >
> > It would be bad to finish waiting in this case.
>
>
> RIght, but if we were to say 'if (balloon_state != BP_DONE)' the worst 
> that can happen is that we will continue on to the next iteration 
> without warning and/or panicing. Of course, there is a chance thaton the 
> next iteration the same thing will happen but I think chances of hitting 
> this race every time are infinitely low. We can also check for 
> current_credit() again.
>
>
> The question is whether we do want to continue waiting if we are in 
> BP_AGAIN. I don't think BP_WAIT is possible in this case although this 
> may change in the future and we will forget to update this code.

BP_EAGAIN should not stop waiting, as it might be intermediate in case
some caches or buffers are freed.


Juergen

Attachment: OpenPGP_0xB0DE9DD628BF132F.asc
Description: OpenPGP public key

Attachment: OpenPGP_signature
Description: OpenPGP digital signature