Re: Hang with v4.15-rc trying to swap back in

[Hugh Dickins <hughd@google.com>]

[-- Attachment #1: Type: TEXT/PLAIN, Size: 4163 bytes --]

On Wed, 17 Jan 2018, Andrew Morton wrote:
> On Wed, 17 Jan 2018 14:33:21 -0800 (PST) Hugh Dickins <hughd@google.com> wrote:
> > On Thu, 28 Dec 2017, James Bottomley wrote:
> > > On Thu, 2017-12-28 at 09:41 -0800, James Bottomley wrote:
> > > > I'd guess that since they're both in io_schedule, the problem is that
> > > > the io_scheduler is taking far too long servicing the requests due to
> > > > some priority issue you've introduced.
> > > 
> > > OK, so after some analysis, that turned out to be incorrect.  The
> > > problem seems to be that we're exiting do_swap_page() with locked pages
> > > that have been read in from swap.
> > > 
> > > Your changelogs are entirely unclear on why you changed the swapcache
> > > setting logic in this patch:
> > > 
> > > commit 0bcac06f27d7528591c27ac2b093ccd71c5d0168
> > > Author: Minchan Kim <minchan@kernel.org>
> > > Date:   Wed Nov 15 17:33:07 2017 -0800
> > > 
> > >     mm, swap: skip swapcache for swapin of synchronous device
> > > 
> > > But I think you're using swapcache == NULL as a signal the page came
> > > from a synchronous device.  In which case the bug is that you've
> > > forgotten we may already have picked up a page in
> > > swap_readahead_detect() which you're wrongly keeping swapcache == NULL
> > > for and the fix is this (it works on my system, although I'm still
> > > getting an unaccountable shutdown delay).
> > > 
> > > I still think we should revert this series, because this may not be the
> > > only bug lurking in the code, so it should go through a lot more
> > > rigorous testing than it has.
> > 
> > Andrew, neither the fix below (works for me, though I have seen other
> > swap funniness, most probably unrelated), nor the reversion preferred
> > by James and Minchan (later in this linux-mm thread), was in 4.15-rc8:
> > the sands of time are running out...
> 
> Yup.  I'm actually planning on sending in this one.  OK by you?

Thanks, yes, that looks equivalent to what I've been running with.

> 
> 
> From: Minchan Kim <minchan@kernel.org>
> Subject: mm/memory.c: release locked page in do_swap_page()
> 
> James reported a bug in swap paging-in from his testing.  It is that
> do_swap_page doesn't release locked page so system hang-up happens due to
> a deadlock on PG_locked.
> 
> It was introduced by 0bcac06f27d7 ("mm, swap: skip swapcache for swapin of
> synchronous device") because I missed swap cache hit places to update
> swapcache variable to work well with other logics against swapcache in
> do_swap_page.
> 
> This patch fixes it.
> 
> Debugged by James Bottomley.
> 
> Link: http://lkml.kernel.org/r/<1514407817.4169.4.camel@HansenPartnership.com>
> Link: http://lkml.kernel.org/r/20180102235606.GA19438@bbox
> Signed-off-by: Minchan Kim <minchan@kernel.org>
> Reported-by: James Bottomley <James.Bottomley@hansenpartnership.com>
> Cc: Hugh Dickins <hughd@google.com>

Acked-by: Hugh Dickins <hughd@google.com>

> Cc: Sergey Senozhatsky <sergey.senozhatsky@gmail.com>
> Cc: Huang Ying <ying.huang@intel.com>
> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
> ---
> 
>  mm/memory.c |   10 ++++++++--
>  1 file changed, 8 insertions(+), 2 deletions(-)
> 
> diff -puN mm/memory.c~mm-release-locked-page-in-do_swap_page mm/memory.c
> --- a/mm/memory.c~mm-release-locked-page-in-do_swap_page
> +++ a/mm/memory.c
> @@ -2857,8 +2857,11 @@ int do_swap_page(struct vm_fault *vmf)
>  	int ret = 0;
>  	bool vma_readahead = swap_use_vma_readahead();
>  
> -	if (vma_readahead)
> +	if (vma_readahead) {
>  		page = swap_readahead_detect(vmf, &swap_ra);
> +		swapcache = page;
> +	}
> +
>  	if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte)) {
>  		if (page)
>  			put_page(page);
> @@ -2889,9 +2892,12 @@ int do_swap_page(struct vm_fault *vmf)
>  
>  
>  	delayacct_set_flag(DELAYACCT_PF_SWAPIN);
> -	if (!page)
> +	if (!page) {
>  		page = lookup_swap_cache(entry, vma_readahead ? vma : NULL,
>  					 vmf->address);
> +		swapcache = page;
> +	}
> +
>  	if (!page) {
>  		struct swap_info_struct *si = swp_swap_info(entry);
>  
> _