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

[Trond Myklebust <trond.myklebust@fys.uio.no>]

On Thu, 2010-08-19 at 15:16 -0400, Jeff Layton wrote:
> On Thu, 19 Aug 2010 10:58:25 -0400
> Trond Myklebust <trond.myklebust@fys.uio.no> wrote:
> 
> > On Thu, 2010-08-19 at 10:37 -0400, Christoph Hellwig wrote:
> > > On Thu, Aug 19, 2010 at 10:15:25AM -0400, Jeff Layton wrote:
> > > > I'm looking at backporting some upstream changes to earlier kernels,
> > > > and ran across something I don't quite understand...
> > > > 
> > > > In nfs_commit_unstable_pages, we set the flags to FLUSH_SYNC. We then
> > > > zero out the flags if wbc->nonblocking or wbc->for_background is set.
> > > > 
> > > > Shouldn't we also clear it out if wbc->sync_mode == WB_SYNC_NONE ?
> > > > WB_SYNC_NONE means "don't wait on anything", so shouldn't that include
> > > > not waiting on the COMMIT to complete?
> > > 
> > > I've been trying to figure out what the nonblocking flag is supposed
> > > to mean for a while now.
> > > 
> > > It basically disappeared in commit 0d99519efef15fd0cf84a849492c7b1deee1e4b7
> > > 
> > > 	"writeback: remove unused nonblocking and congestion checks"
> > > 
> > > from Wu.  What's left these days is a couple of places in local copies
> > > of write_cache_pages (afs, cifs), and a couple of checks in random
> > > writepages instances (afs, block_write_full_page, ceph, nfs, reiserfs, xfs)
> > > and the use in nfs_write_inode.  It's only actually set for memory
> > > migration and pageout, that is VM writeback.
> > > 
> > > To me it really doesn't make much sense, but maybe someone has a better
> > > idea what it is for.
> > > 
> > > > +	if (wbc->nonblocking || wbc->for_background ||
> > > > +	    wbc->sync_mode == WB_SYNC_NONE)
> > > 
> > > You could remove the nonblocking and for_background checks as
> > > these impliy WB_SYNC_NONE.
> > 
> > To me that sounds fine. I've also been trying to wrap my head around the
> > differences between 'nonblocking', 'for_background', 'for_reclaim' and
> > 'for_kupdate' and how the filesystem is supposed to treat them.
> > 
> > Aside from the above, I've used 'for_reclaim', 'for_kupdate' and
> > 'for_background' in order to adjust the RPC request's queuing priority
> > (high in the case of 'for_reclaim' and low for the other two).
> > 
> 
> Here's a lightly tested patch that turns the check for the two flags
> into a check for WB_SYNC_NONE. It seems to do the right thing, but I
> don't have a clear testcase for it. Does this look reasonable?

Looks fine to me. I'll queue it up for the post-2.6.36 merge window...

Cheers
  Trond