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@xxxxxxxxxx>
> 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...
Hugh
>
> James
>
> ---
>
> diff --git a/mm/memory.c b/mm/memory.c
> index ca5674cbaff2..31f9845c340e 100644
> --- a/mm/memory.c
> +++ b/mm/memory.c
> @@ -2847,7 +2847,7 @@ EXPORT_SYMBOL(unmap_mapping_range);
> int do_swap_page(struct vm_fault *vmf)
> {
> struct vm_area_struct *vma = vmf->vma;
> - struct page *page = NULL, *swapcache = NULL;
> + struct page *page = NULL, *swapcache;
> struct mem_cgroup *memcg;
> struct vma_swap_readahead swap_ra;
> swp_entry_t entry;
> @@ -2892,6 +2892,7 @@ int do_swap_page(struct vm_fault *vmf)
> 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);
>
