Re: [PATCH] bdi_sync_writeback should WB_SYNC_NONE first

[Andrew Morton <akpm@linux-foundation.org>]

On Sun, 27 Sep 2009 18:55:14 +0200 Jens Axboe <jens.axboe@oracle.com> wrote:

> > I wasn't referring to this patch actually.  The code as it stands in
> > Linus's tree right now attempts to write back up to 2^63 pages...  
> 
> I agree, it could make the fs sync take a looong time. This is not a new
> issue, though.

It _should_ be a new issue.  The old code would estimate the number of
dirty pages up-front and would then add a +50% fudge factor, so if we
started the sync with 1GB dirty memory, we write back a max of 1.5GB.

However that might have got broken.

void sync_inodes_sb(struct super_block *sb, int wait)
{
	struct writeback_control wbc = {
		.sync_mode	= wait ? WB_SYNC_ALL : WB_SYNC_NONE,
		.range_start	= 0,
		.range_end	= LLONG_MAX,
	};

	if (!wait) {
		unsigned long nr_dirty = global_page_state(NR_FILE_DIRTY);
		unsigned long nr_unstable = global_page_state(NR_UNSTABLE_NFS);

		wbc.nr_to_write = nr_dirty + nr_unstable +
			(inodes_stat.nr_inodes - inodes_stat.nr_unused);
	} else
		wbc.nr_to_write = LONG_MAX; /* doesn't actually matter */

	sync_sb_inodes(sb, &wbc);
}

a) the +50% isn't there in 2.6.31

b) the wait=true case appears to be vulnerable to livelock in 2.6.31.

whodidthat

38f21977663126fef53f5585e7f1653d8ebe55c4 did that back in January.