lkml.org 
[lkml]   [2011]   [Apr]   [19]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 5/6] writeback: try more writeback as long as something was written
On Tue 19-04-11 19:16:01, Wu Fengguang wrote:
> On Tue, Apr 19, 2011 at 06:20:16PM +0800, Jan Kara wrote:
> > On Tue 19-04-11 11:00:08, Wu Fengguang wrote:
> > > writeback_inodes_wb()/__writeback_inodes_sb() are not aggressive in that
> > > they only populate possibly a subset of elegible inodes into b_io at
> > > entrance time. When the queued set of inodes are all synced, they just
> > > return, possibly with all queued inode pages written but still
> > > wbc.nr_to_write > 0.
> > >
> > > For kupdate and background writeback, there may be more eligible inodes
> > > sitting in b_dirty when the current set of b_io inodes are completed. So
> > > it is necessary to try another round of writeback as long as we made some
> > > progress in this round. When there are no more eligible inodes, no more
> > > inodes will be enqueued in queue_io(), hence nothing could/will be
> > > synced and we may safely bail.
> > Let me understand your concern here: You are afraid that if we do
> > for_background or for_kupdate writeback and we write less than
> > MAX_WRITEBACK_PAGES, we stop doing writeback although there could be more
> > inodes to write at the time we are stopping writeback - the two realistic
>
> Yes.
>
> > cases I can think of are:
> > a) when inodes just freshly expired during writeback
> > b) when bdi has less than MAX_WRITEBACK_PAGES of dirty data but we are over
> > background threshold due to data on some other bdi. And then while we are
> > doing writeback someone does dirtying at our bdi.
> > Or do you see some other case as well?
> >
> > The a) case does not seem like a big issue to me after your changes to
>
> Yeah (a) is not an issue with kupdate writeback.
>
> > move_expired_inodes(). The b) case maybe but do you think it will make any
> > difference?
>
> (b) seems also weird. What in my mind is this for_background case.
> Imagine 100 inodes
>
> i0, i1, i2, ..., i90, i91, i99
>
> At queue_io() time, i90-i99 happen to be expired and moved to s_io for
> IO. When finished successfully, if their total size is less than
> MAX_WRITEBACK_PAGES, nr_to_write will be > 0. Then wb_writeback() will
> quit the background work (w/o this patch) while it's still over
> background threshold.
>
> This will be a fairly normal/frequent case I guess.
Ah OK, I see. I missed this case your patch set has added. Also your
changes of
if (!wbc->for_kupdate || list_empty(&wb->b_io))
to
if (list_empty(&wb->b_io))
are going to cause more cases when we'd hit nr_to_write > 0 (e.g. when one
pass of b_io does not write all the inodes so some are left in b_io list
and then next call to writeback finds these inodes there but there's less
than MAX_WRITEBACK_PAGES in them). Frankly, it makes me like the above
change even less. I'd rather see writeback_inodes_wb /
__writeback_inodes_sb always work on a fresh set of inodes which is
initialized whenever we enter these functions. It just seems less
surprising to me...

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


\
 
 \ /
  Last update: 2011-04-19 23:13    [W:0.106 / U:24.652 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site