Re: why are WB_SYNC_NONE COMMITs being done with FLUSH_SYNC set ?

[Jeff Layton <jlayton@redhat.com>]

On Fri, 20 Aug 2010 05:19:04 -0400
Christoph Hellwig <hch@infradead.org> wrote:

> On Fri, Aug 20, 2010 at 07:55:53AM +0800, Wu Fengguang wrote:
> > Since migration and pageout still set nonblocking for ->writepage, we
> > may keep them in the near future, until VM does not start IO on itself.
> 
> Why does pageout() and memory migration need to be even more
> non-blocking than the already non-blockig WB_SYNC_NONE writeout?
> 

Just an idle thought on this...

I think a lot of the confusion here comes from the fact that we have
sync_mode and a bunch of flags, and it's not at all clear how
filesystems are supposed to treat the union of them. There are also
possible unions of flags/sync_modes that never happen in practice. It's
not always obvious though and as filesystem implementors we have to
consider the possibility that they might occur (consider WB_SYNC_ALL +
for_background).

Perhaps a lot of this confusion could be lifted by getting rid of the
extra flags and adding new sync_mode's. Maybe something like:

WB_SYNC_ALL /* wait on everything to complete */
WB_SYNC_NONE /* don't wait on anything */
WB_SYNC_FOR_RECLAIM /* sync for reclaim */
WB_SYNC_FOR_KUPDATED /* sync by kupdate */
...etc...

That does mean that all of the filesystem specific code may need to be
touched when new modes are added and removed. I think it would be
clearer though about what you're supposed to do in ->writepages.

-- 
Jeff Layton <jlayton@redhat.com>