On Sun, Nov 6, 2022 at 11:55 PM Michal Hocko <mhocko@xxxxxxxx> wrote:
>
> On Fri 04-11-22 13:52:52, Yang Shi wrote:
> > On Fri, Nov 4, 2022 at 12:51 PM Michal Hocko <mhocko@xxxxxxxx> wrote:
> > >
> > > On Fri 04-11-22 10:42:45, Yang Shi wrote:
> > > > On Fri, Nov 4, 2022 at 2:56 AM Michal Hocko <mhocko@xxxxxxxx> wrote:
> > > > >
> > > > > On Fri 04-11-22 10:35:21, Michal Hocko wrote:
> > > > > [...]
> > > > > > diff --git a/include/linux/gfp.h b/include/linux/gfp.h
> > > > > > index ef4aea3b356e..308daafc4871 100644
> > > > > > --- a/include/linux/gfp.h
> > > > > > +++ b/include/linux/gfp.h
> > > > > > @@ -227,7 +227,10 @@ static inline
> > > > > > struct folio *__folio_alloc_node(gfp_t gfp, unsigned int order, int nid)
> > > > > > {
> > > > > > VM_BUG_ON(nid < 0 || nid >= MAX_NUMNODES);
> > > > > > - VM_WARN_ON((gfp & __GFP_THISNODE) && !node_online(nid));
> > > > > > + if((gfp & __GFP_THISNODE) && !node_online(nid)) {
> > > > >
> > > > > or maybe even better
> > > > > if ((gfp & (__GFP_THISNODE|__GFP_NOWARN) == __GFP_THISNODE|__GFP_NOWARN) && !node_online(nid))
> > > > >
> > > > > because it doesn't really make much sense to dump this information if
> > > > > the allocation failure is going to provide sufficient (and even more
> > > > > comprehensive) context for the failure. It looks more hairy but this can
> > > > > be hidden in a nice little helper shared between the two callers.
> > > >
> > > > Thanks a lot for the suggestion, printing warning if the gfp flag
> > > > allows sounds like a good idea to me. Will adopt it. But the check
> > > > should look like:
> > > >
> > > > if ((gfp & __GFP_THISNODE) && !(gfp & __GFP_NOWARN) && !node_online(nid))
> > >
> > > The idea was to warn if __GFP_NOWARN _was_ specified. Otherwise we will
> > > get an allocation failure splat from the page allocator and there it
> > > will be clear that the node doesn't have any memory associated. It is
> > > exactly __GFP_NOWARN case that would be a silent failure and potentially
> > > a buggy code (like this THP collapse path). See my point?
> >
> > Aha, yeah, see your point now. I didn't see the splat from the
> > allocator from the bug report, then I realized it had not called into
> > allocator yet before the warning was triggered.
>
> And it would trigger even if it did because GFP_TRANSHUGE has
> __GFP_NOWARN
Yeah, the syzbot has panic on warn set, so kernel just panicked before
entering the allocator.
> --
> Michal Hocko
> SUSE Labs
