Re: [PATCH] mm: fix long time stall from mm_populate

[Andrew Morton <akpm@linux-foundation.org>]

On Wed, 12 Feb 2020 15:12:10 -0800 Minchan Kim <minchan@kernel.org> wrote:

> On Wed, Feb 12, 2020 at 02:24:35PM -0800, Andrew Morton wrote:
> > On Wed, 12 Feb 2020 11:53:22 -0800 Minchan Kim <minchan@kernel.org> wrote:
> > 
> > > > That's definitely wrong.  It'll clear PageReclaim and then pretend it did
> > > > nothing wrong.
> > > > 
> > > > 	return !PageWriteback(page) ||
> > > > 		test_and_clear_bit(PG_reclaim, &page->flags);
> > > > 
> > > 
> > > Much better, Thanks for the review, Matthew!
> > > If there is no objection, I will send two patches to Andrew.
> > > One is PageReadahead strict, the other is limit retry from mm_populate.
> > 
> > With much more detailed changelogs, please!
> > 
> > This all seems rather screwy.  if a page is under writeback then it is
> > uptodate and we should be able to fault it in immediately.
> 
> Hi Andrew,
> 
> This description in cover-letter will work? If so, I will add each part
> below in each patch.
> 
> Subject: [PATCH 0/3] fixing mm_populate long stall
> 
> I got several reports major page fault takes several seconds sometime.
> When I review drop mmap_sem in page fault hanlder, I found several bugs.
> 
>    CPU 1							CPU 2
> mm_populate
>  for ()
>    ..
>    ret = populate_vma_page_range
>      __get_user_pages
>        faultin_page
>          handle_mm_fault
> 	   filemap_fault
> 	     do_async_mmap_readahead
> 	     						shrink_page_list
> 							  pageout
> 							    SetPageReclaim(=SetPageReadahead)
> 							      writepage
> 							        SetPageWriteback
> 	       if (PageReadahead(page))
> 	         maybe_unlock_mmap_for_io
> 		   up_read(mmap_sem)
> 		 page_cache_async_readahead()
> 		   if (PageWriteback(page))
> 		     return;
> 
>     here, since ret from populate_vma_page_range is zero,
>     the loop continue to run with same address with previous
>     iteration. It will repeat the loop until the page's
>     writeout is done(ie, PG_writeback or PG_reclaim clear).

The populate_vma_page_range() kerneldoc is wrong.  "return 0 on
success, negative error code on error".  Care to fix that please?

> We could fix the above specific case via adding PageWriteback. IOW,
> 
>    ret = populate_vma_page_range
>    	   ...
> 	   ...
> 	   filemap_fault
> 	     do_async_mmap_readahead
> 	       if (!PageWriteback(page) && PageReadahead(page))
> 	         maybe_unlock_mmap_for_io
> 		   up_read(mmap_sem)
> 		 page_cache_async_readahead()
> 		   if (PageWriteback(page))
> 		     return;

Well yes, but the testing of PageWriteback() is a hack added in
fe3cba17c49471 to permit the sharing of PG_reclaim and PG_readahead. 
If we didn't need that hack then we could avoid adding new hacks to
hack around the old hack :(.  Have you considered anything along those
lines?  Rework how we handle PG_reclaim/PG_readahead?

> That's a thing [3/3] is fixing here. Even though it could fix the
> problem effectively, it has still livelock problem theoretically
> because the page of faulty address could be reclaimed and then
> allocated/become readahead marker on other CPUs during faulty
> process is retrying in mm_populate's loop.

Really?  filemap_fault()'s

	if (!lock_page_maybe_drop_mmap(vmf, page, &fpin))
		goto out_retry;

	/* Did it get truncated? */
	if (unlikely(compound_head(page)->mapping != mapping)) {
		unlock_page(page);
		put_page(page);
		goto retry_find;
	}

should handle such cases?

> [2/3] is fixing the
> such livelock via limiting retry count.

I wouldn't call that "fixing" :(

> There is another hole for the livelock or hang of the process as well
> as ageWriteback - ra_pages.
> 
> mm_populate
>  for ()
>    ..
>    ret = populate_vma_page_range
>      __get_user_pages
>        faultin_page
>          handle_mm_fault
> 	   filemap_fault
> 	     do_async_mmap_readahead
> 	       if (PageReadahead(page))
> 	         maybe_unlock_mmap_for_io
> 		   up_read(mmap_sem)
> 		 page_cache_async_readahead()
> 		   if (!ra->ra_pages)
> 		     return;
> 
> It will repeat the loop until ra->ra_pages become non-zero.
> [1/3] is fixing the problem.
>