On Mon 21-10-19 14:58:49, Oscar Salvador wrote:
> On Fri, Oct 18, 2019 at 02:06:15PM +0200, Michal Hocko wrote:
> > On Thu 17-10-19 16:21:17, Oscar Salvador wrote:
> > [...]
> > > +bool take_page_off_buddy(struct page *page)
> > > + {
> > > + struct zone *zone = page_zone(page);
> > > + unsigned long pfn = page_to_pfn(page);
> > > + unsigned long flags;
> > > + unsigned int order;
> > > + bool ret = false;
> > > +
> > > + spin_lock_irqsave(&zone->lock, flags);
> >
> > What prevents the page to be allocated in the meantime? Also what about
> > free pages on the pcp lists? Also the page could be gone by the time you
> > have reached here.
>
> Nothing prevents the page to be allocated in the meantime.
> We would just bail out and return -EBUSY to userspace.
> Since we do not do __anything__ to the page until we are sure we took it off,
> and it is completely isolated from the memory, there is no danger.
Wouldn't it be better to simply check the PageBuddy state after the lock
has been taken?
> Since soft-offline is kinda "best effort" mode, it is something like:
> "Sorry, could not poison the page, try again".
Well, I would disagree here. While madvise is indeed a best effort
operation please keep in mind that the sole purpose of this interface is
to allow real MCE behavior. And that operation should better try
_really_ hard to make sure we try to recover as gracefully as possible.
> Now, thinking about this a bit more, I guess we could be more clever here
> and call the routine that handles in-use pages if we see that the page
> was allocated by the time we reach take_page_off_buddy.
>
> About pcp pages, you are right.
> I thought that we were already handling that case, and we do, but looking closer the
> call to shake_page() (that among other things spills pcppages into buddy)
> is performed at a later stage.
> I think we need to adjust __get_any_page to recognize pcp pages as well.
Yeah, pcp pages are PITA. We cannot really recognize them now. Dropping
all pcp pages is certainly a way to go but we need to mark the page
before that happens.
--
Michal Hocko
SUSE Labs
