Re: [PATCH] nfsd: use a multithreaded workqueue for nfsd4_callbacks

[Jeff Layton <jlayton@poochiereds.net>]

On Mon, 12 Oct 2015 16:20:44 -0400
"J. Bruce Fields" <bfields@fieldses.org> wrote:

> On Mon, Oct 12, 2015 at 04:14:21PM -0400, Jeff Layton wrote:
> > On Mon, 12 Oct 2015 14:11:17 -0400
> > "J. Bruce Fields" <bfields@fieldses.org> wrote:
> > 
> > > On Fri, Oct 09, 2015 at 05:39:23PM -0400, Jeff Layton wrote:
> > > > On Fri, 9 Oct 2015 17:24:59 -0400
> > > > "J. Bruce Fields" <bfields@fieldses.org> wrote:
> > > > 
> > > > > On Sat, Oct 03, 2015 at 08:38:02AM -0400, Jeff Layton wrote:
> > > > > > I don't see any need to order callbacks with respect to one another.
> > > > > 
> > > > > I thought the code in nfsd4_process_cb_update really depended on this.
> > > > > The locking it has is against nfsd threads, it probably assumes it's
> > > > > only run from a cb thread and that it's the only one running at a time.
> > > > > 
> > > > > But I haven't reviewed it lately.
> > > > > 
> > > > > --b.
> > > > > 
> > > > 
> > > > Yikes -- ok. That's not at all obvious. That should prob be documented
> > > > if so.
> > > 
> > > Yes, my bad.
> > > 
> > > > Yeah, ok...I guess you could end up with multiple threads racing to
> > > > tear down the cb_client and xprt and create a new one, and it looks
> > > > like the cl_cb_client and cl_cred pointers could get clobbered by new
> > > > ones in that case.
> > > > 
> > > > Shouldn't be too hard to fix protecting those pointers with the
> > > > cl_lock. That said, I prob won't be able to spend time on it right now.
> > > > You can go ahead and drop that patch and I'll resend if/when I get
> > > > around to fixing that.
> > > 
> > > What's the problem that this fixes exactly?
> > > 
> > 
> > Unnecessary serialization of callbacks?
> > 
> > Not that that's necessarily a problem, but with the newer workqueue
> > implementation there is more overhead in running a single threaded
> > workqueue since it implies a rescuer thread and has to do extra work
> > to ensure that the jobs are run in sequential order.
> > 
> > If that serialization is not actually needed, then it's better to move
> > it to a multithreaded workqueue.
> 
> We'd still be serializing a lot of the same code, just using locks
> instead, wouldn't we?
> 
> --b.

Any serialization would be per-clnt instead of global like it is today.

Also, we'd likely only serialize changes to some of the pointers
instead of the entire operation.

-- 
Jeff Layton <jlayton@poochiereds.net>