ext4 stable page writes question

ext4 stable page writes question

[Dmitry Monakhov]

According to stable write assumptions (1d1d1a767206fb)
grab_cache_page_write_begin() now calls relaxed method wait_for_stable_page()
which will wait for writeback to finish only if bdi demand that.
Commit message states that ext4 may not wait
But there are a lot of write-paths where we expect that:
BUG_ON(!PageLocked(page));
BUG_ON(PageWriteback(page));
And the only reason we avoid this bugon is because of commit 47564bfb95b
which use following trick to avoid lock inversion over journal_start:
page = grab_cache_page_write_begin()
unlock_page(page);
ext4_journal_start()
lock_page(page);
wait_on_page_writeback(page); <<<< unconditional wait

So as far as I understand this was done just by occasion because
ext4_page_mkwrite() use wait_for_stable_page().

So here is my question:  Do we have to wait for page's writeback to
finish for all write paths in ext4 code or we may use
wait_for_stable_page() and should cleanup all places where 
we may trigger BUG_ON(PageWriteback(page));
 

Re: ext4 stable page writes question

[Jan Kara]

On Tue 09-04-13 17:38:21, Dmitry Monakhov wrote:
> According to stable write assumptions (1d1d1a767206fb)
> grab_cache_page_write_begin() now calls relaxed method wait_for_stable_page()
> which will wait for writeback to finish only if bdi demand that.
  Yes.

> Commit message states that ext4 may not wait
> But there are a lot of write-paths where we expect that:
> BUG_ON(!PageLocked(page));
> BUG_ON(PageWriteback(page));
  Really? The only places I can find are in writeback path and there we
have wait_on_page_writeback() (either in write_cache_pages() or in
ext4_da_writepages()).

> And the only reason we avoid this bugon is because of commit 47564bfb95b
> which use following trick to avoid lock inversion over journal_start:
> page = grab_cache_page_write_begin()
> unlock_page(page);
> ext4_journal_start()
> lock_page(page);
> wait_on_page_writeback(page); <<<< unconditional wait
  No, I think this is really independent. ext4 should be fine when write &
writeback are running in parallel for a page.

> So as far as I understand this was done just by occasion because
> ext4_page_mkwrite() use wait_for_stable_page().
> 
> So here is my question:  Do we have to wait for page's writeback to
> finish for all write paths in ext4 code or we may use
> wait_for_stable_page() and should cleanup all places where 
> we may trigger BUG_ON(PageWriteback(page));
  If there's any such place, please tell me how we could trigger it...

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

Re: ext4 stable page writes question

[Dmitry Monakhov]

On Tue, 9 Apr 2013 16:28:05 +0200, Jan Kara <jack@suse.cz> wrote:
> On Tue 09-04-13 17:38:21, Dmitry Monakhov wrote:
> > According to stable write assumptions (1d1d1a767206fb)
> > grab_cache_page_write_begin() now calls relaxed method wait_for_stable_page()
> > which will wait for writeback to finish only if bdi demand that.
>   Yes.
> 
> > Commit message states that ext4 may not wait
> > But there are a lot of write-paths where we expect that:
> > BUG_ON(!PageLocked(page));
> > BUG_ON(PageWriteback(page));
>   Really? The only places I can find are in writeback path and there we
> have wait_on_page_writeback() (either in write_cache_pages() or in
> ext4_da_writepages()).
> 
> > And the only reason we avoid this bugon is because of commit 47564bfb95b
> > which use following trick to avoid lock inversion over journal_start:
> > page = grab_cache_page_write_begin()
> > unlock_page(page);
> > ext4_journal_start()
> > lock_page(page);
> > wait_on_page_writeback(page); <<<< unconditional wait
>   No, I think this is really independent. ext4 should be fine when write &
> writeback are running in parallel for a page.
> 
> > So as far as I understand this was done just by occasion because
> > ext4_page_mkwrite() use wait_for_stable_page().
> > 
> > So here is my question:  Do we have to wait for page's writeback to
> > finish for all write paths in ext4 code or we may use
> > wait_for_stable_page() and should cleanup all places where 
> > we may trigger BUG_ON(PageWriteback(page));
>   If there's any such place, please tell me how we could trigger it...
move_extent.c:mext_page_mkuptodate() expect that.
IMHO it is reasonable to force behavior here even if we may not do that
on other places because defragmentation code should be 101% reliable
and performance is not important here.
> 
> 								Honza
> -- 
> Jan Kara <jack@suse.cz>
> SUSE Labs, CR