lkml.org 
[lkml]   [2021]   [Oct]   [4]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v2 1/2] mm, thp: check page mapping when truncating page cache
On Mon, Oct 04, 2021 at 07:58:10PM -0700, Hugh Dickins wrote:
> On Tue, 5 Oct 2021, Rongwei Wang wrote:
>
> > Hi,
> > I have run our cases these two days to stress test new Patch #1. The new Patch
> > #1 mainly add filemap_invalidate_{un}lock before and after
> > truncate_pagecache(), basing on original Patch #1. And the crash has not
> > happened.
> >
> > Now, I keep the original Patch #1, then adding the code below which suggested
> > by liu song (I'm not sure which one I should add in the next version,
> > Suggested-by or Signed-off-by? If you know, please remind me).
> >
> > - if (filemap_nr_thps(inode->i_mapping))
> > + if (filemap_nr_thps(inode->i_mapping)) {
> > + filemap_invalidate_lock(inode->i_mapping);
> > truncate_pagecache(inode, 0);
> > + filemap_invalidate_unlock(inode->i_mapping);
> > + }
>
> I won't NAK that patch; but I still believe it's unnecessary, and don't
> see how it protects against all the races (collapse_file() does not use
> that lock, whereas collapse_file() does use page lock). And if you're
> hoping to fix 5.10, then you will have to backport those invalidate_lock
> patches there too (they're really intended to protect hole-punching).

I believe all we really need to do is protect against calling
truncate_pagecache() simultaneously to avoid one of the calls seeing
a tail page. i_mutex would work for this purpose just as well as
filemap_invalidate_lock(). See, for example, ext4_zero_range() which
first takes inode_lock(), then filemap_invalidate_lock() before calling
truncate_pagecache_range().

> > And the reason for keeping the original Patch #1 is mainly to fix the race
> > between collapse_file and truncate_pagecache. It seems necessary. Despite the
> > two-day test, I did not reproduce this race any more.
> >
> > In addition, I also test the below method:
> >
> > diff --git a/mm/truncate.c b/mm/truncate.c
> > index 3f47190f98a8..33604e4ce60a 100644
> > --- a/mm/truncate.c
> > +++ b/mm/truncate.c
> > @@ -210,8 +210,6 @@ invalidate_complete_page(struct address_space *mapping,
> > struct page *page)
> >
> > int truncate_inode_page(struct address_space *mapping, struct page *page)
> > {
> > - VM_BUG_ON_PAGE(PageTail(page), page);
> > -
> > if (page->mapping != mapping)
> > return -EIO;
> >
> > I am not very sure this VM_BUG_ON_PAGE(PageTail) is what Hugh means. And
> > the test results show that only removing this VM_BUG_ON_PAGE(PageTail) has no
> > effect. So, I still keep the original Patch #1 to fix one race.
>
> Yes, that's exactly what I meant, and thank you for intending to try it.
>
> But if that patch had "no effect", then I think you were not running the
> kernel with that patch applied: because it deletes the BUG on line 213
> of mm/truncate.c, which is what you reported in the first mail!
>
> Or, is line 213 of mm/truncate.c in your 5.10.46-hugetext+ kernel
> something else? I've been looking at 5.15-rc.
>
> But I wasn't proposing to delete it merely to hide the BUG: as I hope
> I explained, we could move it below the page->mapping check, but it
> wouldn't really be of any value there since tails have NULL page->mapping
> anyway (well, I didn't check first and second tails, maybe mapping gets
> reused for some compound page field in those). I was proposing to delete
> it because the page->mapping check then weeds out the racy case once
> we're holding page lock, without the need for adding anything special.

I think if we remove the race with the above mutex lock then we'll never
see a tail page in this routine.

\
 
 \ /
  Last update: 2021-10-05 05:09    [W:0.488 / U:0.256 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site