Re: [PATCH 3/6] writeback: sync expired inodes first in backgroundwriteback

From: Dave Chinner
Date: Wed Apr 20 2011 - 23:02:06 EST


On Thu, Apr 21, 2011 at 10:06:17AM +0800, Wu Fengguang wrote:
> On Thu, Apr 21, 2011 at 08:45:47AM +0800, Dave Chinner wrote:
> > On Wed, Apr 20, 2011 at 10:53:21AM +0800, Wu Fengguang wrote:
> > > On Wed, Apr 20, 2011 at 09:21:20AM +0800, Dave Chinner wrote:
> > > > On Tue, Apr 19, 2011 at 08:56:16PM +0800, Wu Fengguang wrote:
> > > > > I actually started with wb_writeback() as a natural choice, and then
> > > > > found it much easier to do the expired-only=>all-inodes switching in
> > > > > move_expired_inodes() since it needs to know the @b_dirty and @tmp
> > > > > lists' emptiness to trigger the switch. It's not sane for
> > > > > wb_writeback() to look into such details. And once you do the switch
> > > > > part in move_expired_inodes(), the whole policy naturally follows.
> > > >
> > > > Well, not really. You didn't need to modify move_expired_inodes() at
> > > > all to implement these changes - all you needed to do was modify how
> > > > older_than_this is configured.
> > > >
> > > > writeback policy is defined by the struct writeback_control.
> > > > move_expired_inodes() is pure mechanism. What you've done is remove
> > > > policy from the struct wbc and moved it to move_expired_inodes(),
> > > > which now defines both policy and mechanism.
> > >
> > > > Furhter, this means that all the tracing that uses the struct wbc no
> > > > no longer shows the entire writeback policy that is being worked on,
> > > > so we lose visibility into policy decisions that writeback is
> > > > making.
> > >
> > > Good point! I'm convinced, visibility is a necessity for debugging the
> > > complex writeback behaviors.
> > >
> > > > This same change is as simple as updating wbc->older_than_this
> > > > appropriately after the wb_writeback() call for both background and
> > > > kupdate and leaving the lower layers untouched. It's just a policy
> > > > change. If you thinkthe mechanism is inefficient, copy
> > > > wbc->older_than_this to a local variable inside
> > > > move_expired_inodes()....
> > >
> > > Do you like something like this? (details will change a bit when
> > > rearranging the patchset)
> >
> > Yeah, this is close to what I had in mind.
> >
> > >
> > > --- linux-next.orig/fs/fs-writeback.c 2011-04-20 10:30:47.000000000 +0800
> > > +++ linux-next/fs/fs-writeback.c 2011-04-20 10:40:19.000000000 +0800
> > > @@ -660,11 +660,6 @@ static long wb_writeback(struct bdi_writ
> > > long write_chunk;
> > > struct inode *inode;
> > >
> > > - if (wbc.for_kupdate) {
> > > - wbc.older_than_this = &oldest_jif;
> > > - oldest_jif = jiffies -
> > > - msecs_to_jiffies(dirty_expire_interval * 10);
> > > - }
> >
> > Right here I'd do:
> >
> > if (work->for_kupdate || work->for_background)
> > wbc.older_than_this = &oldest_jif;
> >
> > so that the setting of wbc.older_than_this in the loop can trigger
> > on whether it is null or not.
>
> That's the tricky part that drove me to change move_expired_inodes()
> directly..
>
> One important thing to bear in mind is, the background work can run on
> for one hour, one day or whatever. During the time dirty inodes come
> and go, expired and cleaned. If we only reset wbc.older_than_this and
> never restore it _inside_ the loop, we'll quickly lose the ability to
> "start with expired inodes" shortly after f.g. 5 minutes.

However, there's not need to implicity switch back to expired inodes
on the next wb_writeback loop - it only needs to switch back when
b_io is emptied. And I suspect that it really only needs to switch
if there are inodes on b_more_io because if we didn't put any inodes
onto b_more_io, then then we most likely cleaned the entire list of
unexpired inodes in a single write chunk...

That is, something like this when updating the background state in
the loop tail:

if (work->for_background && list_empty(&wb->b_io)) {
if (wbc.older_than_this) {
if (list_empty(&wb->b_more_io)) {
wbc.older_than_this = NULL;
continue;
}
} else if (!list_empty(&wb->b_more_io)) {
wbc.older_than_this = &oldest_jif;
continue;
}
}

Still, given wb_writeback() is the only caller of both
__writeback_inodes_sb and writeback_inodes_wb(), I'm wondering if
moving the queue_io calls up into wb_writeback() would clean up this
logic somewhat. I think Jan mentioned doing something like this as
well elsewhere in the thread...

Cheers,

Dave.
--
Dave Chinner
david@xxxxxxxxxxxxx
--
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/