Re: [PATCH] mm/page_alloc.c: avoid inheritting current's flags when invoked in interrupt

[Xu, Yanfei]

On 9/16/20 9:17 AM, Andrew Morton wrote:
> On Tue, 15 Sep 2020 15:56:35 +0800 <yanfei.xu@xxxxxxxxxxxxx> wrote:
>
> > From: Yanfei Xu <yanfei.xu@xxxxxxxxxxxxx>
> >
> > alloc_mask shouldn't inherit the current task's flags when
> > __alloc_pages_nodemask is invoked in interrupt.
> >
> > ...
> >
> > --- a/mm/page_alloc.c
> > +++ b/mm/page_alloc.c
> > @@ -4889,7 +4889,8 @@ __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
> >   	 * from a particular context which has been marked by
> >   	 * memalloc_no{fs,io}_{save,restore}.
> >   	 */
> > -	alloc_mask = current_gfp_context(gfp_mask);
> > +	if (!in_interrupt())
> > +		alloc_mask = current_gfp_context(gfp_mask);
> >   	ac.spread_dirty_pages = false;
> >   
> >   	/*
>
> hm, yes, and perhaps other callsites in page_alloc.c.
>
> I assume this doesn't actually make any runtime difference?  Because
> gfp_mask in interrupt contexts isn't going to have __GFP_IO or __GFP_FS
> anyway.
Thanks for your reply!

Yes, It doesn't make any runtime difference. Theoretically, GPF_ATOMIC 
or GFP_NOWAIT should be used in interrupt context for allocate pages, so
that gfp_mask isn't going to have __GFP_IO or __GFP_FS.

But if somebody use wrong gfp_masks, __GFP_IO or __GFP_FS will be 
introduced, with the process interrupted has PF_MEMALLOC_NOIO or 
PF_MEMALLOC_NOFS, current_gfp_context may help to hide these wrong 
usages. I don't think it is the original purpose of that piece of
codes.

And how about add BUG_ON or WARN_ON to figure out the situation which
introduce __GFP_IO or __GFP_FS in interrupt context?

Regards,
Yanfei