lkml.org 
[lkml]   [2019]   [Sep]   [2]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH block/for-next] writeback: add tracepoints for cgroup foreign writebacks
Hello Tejun,

On Fri 30-08-19 10:09:03, Tejun Heo wrote:
> On Fri, Aug 30, 2019 at 06:42:11PM +0200, Jan Kara wrote:
> > Well, but if you look at __set_page_dirty_nobuffers() it is careful. It
> > does:
> >
> > struct address_space *mapping = page_mapping(page);
> >
> > if (!mapping) {
> > bail
> > }
> > ... use mapping
> >
> > Exactly because page->mapping can become NULL under your hands if you don't
> > hold page lock. So I think you either need something similar in your
> > tracepoint or handle this in the caller.
>
> So, account_page_dirtied() is called from two places.
>
> __set_page_dirty() and __set_page_dirty_nobuffers(). The following is
> from the latter.
>
> lock_page_memcg(page);
> if (!TestSetPageDirty(page)) {
> struct address_space *mapping = page_mapping(page);
> ...
>
> if (!mapping) {
> unlock_page_memcg(page);
> return 1;
> }
>
> xa_lock_irqsave(&mapping->i_pages, flags);
> BUG_ON(page_mapping(page) != mapping);
> WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
> account_page_dirtied(page, mapping);
> ...
>
> If I'm reading it right, it's saying that at this point if mapping
> exists after setting page dirty, it must not change while locking
> i_pages.

Correct __set_page_dirty_nobuffers() is supposed to be called serialized
with truncation either through page lock or other means. At least the
comment says so and the code looks like that.

>
> __set_page_dirty_nobuffers() is more brief but seems to be making the
> same assumption.

I suppose you mean __set_page_dirty() here.

> xa_lock_irqsave(&mapping->i_pages, flags);
> if (page->mapping) { /* Race with truncate? */
> WARN_ON_ONCE(warn && !PageUptodate(page));
> account_page_dirtied(page, mapping);
> __xa_set_mark(&mapping->i_pages, page_index(page),
> PAGECACHE_TAG_DIRTY);
> }
> xa_unlock_irqrestore(&mapping->i_pages, flags);
>
> Both are clearly assuming that once i_pages is locked, mapping can't
> change. So, inside account_page_dirtied(), mapping clearly can't
> change. The TP in question - track_foreign_dirty - is invoked from
> mem_cgroup_track_foreign_dirty() which is only called from
> account_page_dirty(), so I'm failing to see how mapping would change
> there.

I'm not sure where we depend here on page->mapping not getting cleared. The
point is even if page->mapping is getting cleared while we work on the
page, we have 'mapping' stored locally so we just account everything
against the original mapping.

I've researched this a bit more and commit 2d6d7f982846 "mm: protect
set_page_dirty() from ongoing truncation" introduced the idea that
__set_page_dirty_nobuffers() should be only called synchronized with
truncation. Now I know for a fact that this is not always the case (e.g.
various RDMA drivers calling set_page_dirty() without a lock or any other
protection against truncate) but let's consider this a bug in the caller of
set_page_dirty(). So in the end I agree that you're fine with relying on
page_mapping() not changing under you.

Honza
--
Jan Kara <jack@suse.com>
SUSE Labs, CR

\
 
 \ /
  Last update: 2019-09-02 10:38    [W:0.044 / U:0.436 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site