Re: [PATCH 2/2] mm: slince possible data races about pgdat->kswapd

[Kefeng Wang]

On 2022/8/25 16:22, David Hildenbrand wrote:
> On 25.08.22 04:34, Kefeng Wang wrote:
> > On 2022/8/24 16:24, David Hildenbrand wrote:
> > > On 24.08.22 09:19, Kefeng Wang wrote:
> > > > The pgdat->kswapd could be accessed concurrently by kswapd_run() and
> > > > kcompactd(), it don't be protected by any lock, which could leads to
> > > > data races, adding READ/WRITE_ONCE() to slince it.
> > > Okay, I think this patch here makes it clearer that we really just want
> > > proper synchronization instead of hacking around it.
> > >
> > > What speaks against protecting pgdat->kswapd this using some proper
> > > locking primitive?
> > as comments about kswapd in struct pglist_data,  pgdat->kswapd should be
> >
> > protected by mem_hotplug_begin/done(), how about this way?
> >
> > diff --git a/mm/compaction.c b/mm/compaction.c
> > index 640fa76228dd..62018f35242a 100644
> > --- a/mm/compaction.c
> > +++ b/mm/compaction.c
> > @@ -1983,7 +1983,13 @@ static inline bool is_via_compact_memory(int order)
> >
> >    static bool kswapd_is_running(pg_data_t *pgdat)
> >    {
> > -       return pgdat->kswapd && task_is_running(pgdat->kswapd);
> > +       bool running;
> > +
> > +       mem_hotplug_begin();
> > +       running = pgdat->kswapd && task_is_running(pgdat->kswapd);
> > +       mem_hotplug_end();
> > +
> > +       return running;
> >    }
> I'd much rather just use a dedicated lock that does not involve memory
> hotplug.

The issue only occurred due memory hotplug, without mem-hotplug,

the kswapd won't stop or re-run, there is no above issue too, add a new

lock would be duplicated, but the scope of protection is smaller, I could

repost with new lock if no more comment.

>