Re: splice: move balance_dirty_pages_ratelimited() outside of splice actor

[Jens Axboe]

On Mon, Jun 11 2007, Andrew Morton wrote:
> On Mon, 11 Jun 2007 21:59:15 GMT
> Linux Kernel Mailing List <linux-kernel@xxxxxxxxxxxxxxx> wrote:
> 
> > Gitweb:     http://git.kernel.org/git/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commit;h=20d698db67059a63d217030dfd02872cb5f88dfb
> > Commit:     20d698db67059a63d217030dfd02872cb5f88dfb
> > Parent:     17374ff1aa9ce2a0597416a16729474b538af443
> > Author:     Jens Axboe <jens.axboe@xxxxxxxxxx>
> > AuthorDate: Tue Jun 5 11:05:11 2007 +0200
> > Committer:  Jens Axboe <jens.axboe@xxxxxxxxxx>
> > CommitDate: Fri Jun 8 08:33:59 2007 +0200
> > 
> >     splice: move balance_dirty_pages_ratelimited() outside of splice actor
> >     
> >     I've seen inode related deadlocks, so move this call outside of the
> >     actor itself, which may hold the inode lock.
> >     
> 
> eh?  If the pipe_to_file() caller holds inode_lock, our problems are large.
> 
> I doubt if that's true, so what problem is this patch really fixing??

I can repeatedly lock up balance_dirty_pages() if it's called inside the
double lock of splice_from_pipe(). I'll fire up the test box and
reproduce, then post the backtraces.

> > ---
> >  fs/splice.c |    3 ++-
> >  1 files changed, 2 insertions(+), 1 deletions(-)
> > 
> > diff --git a/fs/splice.c b/fs/splice.c
> > index b78a7f0..6349d31 100644
> > --- a/fs/splice.c
> > +++ b/fs/splice.c
> > @@ -652,7 +652,6 @@ find_page:
> >  	 * accessed, we are now done!
> >  	 */
> >  	mark_page_accessed(page);
> > -	balance_dirty_pages_ratelimited(mapping);
> >  out:
> >  	page_cache_release(page);
> >  	unlock_page(page);
> > @@ -823,6 +822,7 @@ generic_file_splice_write_nolock(struct pipe_inode_info *pipe, struct file *out,
> >  			if (err)
> >  				ret = err;
> >  		}
> > +		balance_dirty_pages_ratelimited(mapping);
> >  	}
> >  
> >  	return ret;
> > @@ -876,6 +876,7 @@ generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
> >  			if (err)
> >  				ret = err;
> >  		}
> > +		balance_dirty_pages_ratelimited(mapping);
> >  	}
> >  
> 
> balance_dirty_pages_ratelimited() is supposed to be called
> once-per-dirtied-page.  This caller can dirty an arbitrarily large amount
> of memory and hence should use balance_dirty_pages_ratelimited_nr().
> 
> As things stand, a large splice() could potentially cause the dirty limits
> to be exceeded.

It's mostly (at most) 16 pages, not an arbitrarily large amount. So I
doubt it makes a lot of difference.

> btw, can we please arrange to get patches reviewed prior to them being
> merged?

Sure, I should have posted the series here. Mostly simple stuff though,
and others have seen them.

-- 
Jens Axboe

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/