Re: [PATCH v2 2/3] mm/swapfile: fix possible data races of inuse_pages

[Muchun Song]

On Mon, Jun 20, 2022 at 09:46:47AM -0400, Qian Cai wrote:
> On Mon, Jun 20, 2022 at 08:32:27PM +0800, Miaohe Lin wrote:
> > >>>>> --- a/mm/swapfile.c
> > >>>>> +++ b/mm/swapfile.c
> > >>>>> @@ -2646,7 +2646,7 @@ static int swap_show(struct seq_file *swap, void *v)
> > >>>>>  	}
> > >>>>>  
> > >>>>>  	bytes = si->pages << (PAGE_SHIFT - 10);
> > >>>>> -	inuse = si->inuse_pages << (PAGE_SHIFT - 10);
> > >>>>> +	inuse = READ_ONCE(si->inuse_pages) << (PAGE_SHIFT - 10);
> > >>>>>  
> > >>>>>  	file = si->swap_file;
> > >>>>>  	len = seq_file_path(swap, file, " \t\n\\");
> > >>>>> @@ -3265,7 +3265,7 @@ void si_swapinfo(struct sysinfo *val)
> > >>>>>  		struct swap_info_struct *si = swap_info[type];
> > >>>>>  
> > >>>>>  		if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
> > >>>>> -			nr_to_be_unused += si->inuse_pages;
> > >>>>> +			nr_to_be_unused += READ_ONCE(si->inuse_pages);
> > >>>>>  	}
> > >>>>>  	val->freeswap = atomic_long_read(&nr_swap_pages) + nr_to_be_unused;
> > >>>>>  	val->totalswap = total_swap_pages + nr_to_be_unused;
> > >>>>
> > >>>> READ_ONCE() should be paired with WRITE_ONCE().  So, change the writer
> > >>>> side too?
> > >>>
> > >>> READ_ONCE() is used to fix the complaint of concurrent accessing to si->inuse_pages from KCSAN here.
> > >>> The similar commit is 218209487c3d ("mm/swapfile: fix data races in try_to_unuse()"). IMHO, it's fine
> > >>
> > >> I think the fix 218209487c3d is incomplete. The write side in swap_range_free() should
> > >> also be fixed. Otherwise, IIUC, it cannot stop KCSAN complaining.
> > > 
> > > I tend to agree with you. READ_ONCE() should be paired with WRITE_ONCE() theoretically. But WRITTE_ONCE()
> > > is ignored while the commit is introduced. Add Qian Cai for helping verify it. It's very kind of @Qian Cai
> > > if he could tell us whether WRITTE_ONCE() is ignored deliberately.
> 
> The write side should be protected by the lock swap_info_struct::lock. Is
> that not the case here?
>

The lock does not protect the read sides. So the write side should be
fixed by WRITTE_ONCE().

Thanks.