Re: [PATCH] remove throttle_vm_writeout()

[Fengguang Wu <wfg@mail.ustc.edu.cn>]

On Fri, Oct 05, 2007 at 08:32:19PM +0200, Peter Zijlstra wrote:
> 
> On Fri, 2007-10-05 at 13:50 -0400, Trond Myklebust wrote:
> > On Fri, 2007-10-05 at 12:57 +0200, Peter Zijlstra wrote:
> > > In this patch I totally ignored unstable, but I'm not sure that's the
> > > proper thing to do, I'd need to figure out what happens to an unstable
> > > page when passed into pageout() - or if its passed to pageout at all.
> > > 
> > > If unstable pages would be passed to pageout(), and it would properly
> > > convert them to writeback and clean them, then there is nothing wrong.
> > 
> > Why would we want to do that? That would be a hell of a lot of work
> > (locking pages, setting flags, unlocking pages, ...) for absolutely no
> > reason.
> > 
> > Unstable writes are writes which have been sent to the server, but which
> > haven't been written to disk on the server. A single RPC command is then
> > sent (COMMIT) which basically tells the server to call fsync(). After
> > that is successful, we can free up the pages, but we do that with no
> > extra manipulation of the pages themselves: no page locks, just removal
> > from the NFS private radix tree, and freeing up of the NFS private
> > structures.
> > 
> > We only need to touch the pages again in the unlikely case that the
> > COMMIT fails because the server has rebooted. In this case we have to
> > resend the writes, and so the pages are marked as dirty, so we can go
> > through the whole writepages() rigmarole again...
> > 
> > So, no. I don't see sending pages through pageout() as being at all
> > helpful.
> 
> Well, the thing is, we throttle pageout in throttle_vm_writeout(). As it
> stand we can deadlock there because it just waits for the numbers to
> drop, and unstable pages don't automagically dissapear. Only
> write_inodes() - normally called from balance_dirty_pages() will call
> COMMIT.

I wonder whether
        if (!bdi_nr_writeback)
                break;
or something like that could avoid the deadlock?

> So my thought was that calling pageout() on an unstable page would do
> the COMMIT - we're low on memory, otherwise we would not be paging, so
> getting rid of unstable pages seems to make sense to me.

I guess "many unstable pages" would be better if we are taking this way.