Re: [PATCH 3/3] writeback: avoid extra sync work at enqueue time

[Wu Fengguang <fengguang.wu@intel.com>]

On Thu, May 05, 2011 at 10:10:39PM +0800, Wu Fengguang wrote:
> On Thu, May 05, 2011 at 10:01:34PM +0800, Jan Kara wrote:
> > On Thu 05-05-11 20:27:32, Wu Fengguang wrote:
> > > On Thu, May 05, 2011 at 05:24:27AM +0800, Jan Kara wrote:
> > > > On Mon 02-05-11 11:17:53, Wu Fengguang wrote:
> > > > > This removes writeback_control.wb_start and does more straightforward
> > > > > sync livelock prevention by setting .older_than_this to prevent extra
> > > > > inodes from being enqueued in the first place.
> > > > > 
> > > > > --- linux-next.orig/fs/fs-writeback.c	2011-05-02 11:17:24.000000000 +0800
> > > > > +++ linux-next/fs/fs-writeback.c	2011-05-02 11:17:27.000000000 +0800
> > > > > @@ -683,10 +672,12 @@ static long wb_writeback(struct bdi_writ
> > > > >  	 *                   (quickly) tag currently dirty pages
> > > > >  	 *                   (maybe slowly) sync all tagged pages
> > > > >  	 */
> > > > > -	if (wbc.sync_mode == WB_SYNC_ALL || wbc.tagged_sync)
> > > > > +	if (wbc.sync_mode == WB_SYNC_ALL || wbc.tagged_sync) {
> > > > >  		write_chunk = LONG_MAX;
> > > > > +		oldest_jif = jiffies;
> > > > > +		wbc.older_than_this = &oldest_jif;
> > > > > +	}
> > > >   What are the implications of not doing dirty-time livelock avoidance for
> > > > other types of writeback? Is that a mistake? I'd prefer to have in
> > > > wb_writeback():
> > > > if (wbc.for_kupdate)
> > > > 	oldest_jif = jiffies - msecs_to_jiffies(dirty_expire_interval * 10);
> > > > else
> > > > 	oldest_jif = jiffies;
> > > > wbc.older_than_this = &oldest_jif;
> > > > 
> > > > And when you have this, you can make wbc.older_than_this just a plain
> > > > number and remove all those checks for wbc.older_than_this == NULL.
> > > 
> > > Good point. Here is the fixed patch. Will you send the patch to change
> > > the type when the current patches are settled down?
> >   OK, I will do that.
> 
> Thank you.
> 
> > > @@ -686,7 +674,9 @@ static long wb_writeback(struct bdi_writ
> > >  	if (wbc.sync_mode == WB_SYNC_ALL || wbc.tagged_sync)
> > >  		write_chunk = LONG_MAX;
> > >  
> > > -	wbc.wb_start = jiffies; /* livelock avoidance */
> > > +	oldest_jif = jiffies;
> > > +	wbc.older_than_this = &oldest_jif;
> > > +
> >   I might be already confused with all the code moving around but won't
> > this overwrite the value set for the for_kupdate case?
> 
> It's the opposite -- it will be overwritten inside the loop by
> for_kupdate, which may run for long time and hence need to update
> oldest_jif from time to time.

The code is now:

        oldest_jif = jiffies;
        work->older_than_this = &oldest_jif;

        for (;;) {
                // ...
               
                if (work->for_kupdate || work->for_background) {
                        oldest_jif = jiffies -
                                msecs_to_jiffies(dirty_expire_interval * 10);
                        work->older_than_this = &oldest_jif;
                }

retry:         
                // ...
               
                /*
                 * background writeback will start with expired inodes, and
                 * if none is found, fallback to all inodes. This order helps
                 * reduce the number of dirty pages reaching the end of LRU
                 * lists and cause trouble to the page reclaim.
                 */
                if (work->for_background &&
                    work->older_than_this &&
                    list_empty(&wb->b_io) &&
                    list_empty(&wb->b_more_io)) {
                        work->older_than_this = NULL;
                        goto retry;
                }

                // ...
        }

Thanks,
Fengguang