Re: [PATCH] mm, page_alloc: capture page in task context only

[Hugh Dickins <hughd@google.com>]

On Fri, 12 Jun 2020, Vlastimil Babka wrote:
> On 6/10/20 10:48 PM, Hugh Dickins wrote:
> > While stressing compaction, one run oopsed on NULL capc->cc in
> > __free_one_page()'s task_capc(zone): compact_zone_order() had been
> > interrupted, and a page was being freed in the return from interrupt.
> > 
> > Though you would not expect it from the source, both gccs I was using
> > (a 4.8.1 and a 7.5.0) had chosen to compile compact_zone_order() with
> > the ".cc = &cc" implemented by mov %rbx,-0xb0(%rbp) immediately before
> > callq compact_zone - long after the "current->capture_control = &capc".
> > An interrupt in between those finds capc->cc NULL (zeroed by an earlier
> > rep stos).
> 
> Ugh, nasty. Same here with gcc 10.

Thanks for checking, nice to know that I'm in good company :)

> 
> > This could presumably be fixed by a barrier() before setting
> > current->capture_control in compact_zone_order(); but would also need
> > more care on return from compact_zone(), in order not to risk leaking
> > a page captured by interrupt just before capture_control is reset.
> 
> I was hoping a WRITE_ONCE(current->capture_control) would be enough,
> but apparently it's not (I tried).

Right, I don't think volatiles themselves actually constitute barriers;
but I'd better keep quiet, I notice the READ_ONCE/WRITE_ONCE/data_race
industry has been busy recently, and I'm likely out-of-date and mistaken.

> 
> > Maybe that is the preferable fix, but I felt safer for task_capc() to
> > exclude the rather surprising possibility of capture at interrupt time.
> 
> > Fixes: 5e1f0f098b46 ("mm, compaction: capture a page under direct compaction")
> > Cc: stable@vger.kernel.org # 5.1+
> > Signed-off-by: Hugh Dickins <hughd@google.com>
> 
> Acked-by: Vlastimil Babka <vbabka@suse.cz>

Thanks, and to Mel for his.

> 
> But perhaps I would also make sure that we don't expose the half initialized
> capture_control and run into this problem again later. It's not like this is a
> fast path where barriers hurt. Something like this then? (with added comments)

Would it be very rude if I leave that to you and to Mel? to add, or
to replace mine if you wish - go ahead.  I can easily see that more
sophistication at the compact_zone_order() end may be preferable to
another test and branch inside __free_one_page() (and would task_capc()
be better with an "unlikely" in it?).

But it seems unnecessary to have a fix at both ends, and I'm rather too
wound up in other things at the moment, to want to read up on the current
state of such barriers, and sign off on the Vlastipatch below myself (but
I do notice that READ_ONCE seems to have more in it today than I remember,
which probably accounts for why you did not put the barrier() I expected
to see on the way out).

Hugh

> 
> diff --git a/mm/compaction.c b/mm/compaction.c
> index fd988b7e5f2b..c89e26817278 100644
> --- a/mm/compaction.c
> +++ b/mm/compaction.c
> @@ -2316,15 +2316,17 @@ static enum compact_result compact_zone_order(struct zone *zone, int order,
>  		.page = NULL,
>  	};
>  
> -	current->capture_control = &capc;
> +	barrier();
> +
> +	WRITE_ONCE(current->capture_control, &capc);
>  
>  	ret = compact_zone(&cc, &capc);
>  
>  	VM_BUG_ON(!list_empty(&cc.freepages));
>  	VM_BUG_ON(!list_empty(&cc.migratepages));
>  
> -	*capture = capc.page;
> -	current->capture_control = NULL;
> +	WRITE_ONCE(current->capture_control, NULL);
> +	*capture = READ_ONCE(capc.page);
>  
>  	return ret;
>  }