Re: [PATCH] fs: fput() can be called from interrupt context

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

On Fri, 2009-03-13 at 18:40 -0700, Davide Libenzi wrote:
> On Fri, 13 Mar 2009, Trond Myklebust wrote:
> 
> > On Thu, 2009-03-12 at 06:39 -0700, Davide Libenzi wrote:
> > > On Wed, 11 Mar 2009, Andrew Morton wrote:
> > > 
> > > > > Take the time to check how fs/aio.c handle the fput(req->ki_filp) case
> > > > > (or read my 2nd patch, it should spot the thing)
> > > > 
> > > > Well yes, a kludge like that seems a bit safer.
> > > > 
> > > > It's somewhat encouraging that we're apparently already doing fput()
> > > > from within keventd (although how frequently?).  There might be
> > > > problems with file locking, security code, etc from doing fput() from
> > > > an unexpected thread.  And then there are all the usual weird problem
> > > > with using the keventd queues which take a long time to get discovered.
> > > 
> > > Would it be a huge problem, performance-wise, to stop making ->f_count 
> > > tricks in __aio_put_req, and always offload to fput_work the task of 
> > > releasing the requests?
> > > If that's a huge problem, IMO the lower impact fix would be to use 
> > > aio_fput_routine to loop into a second list, releasing the eventual 
> > > eventfd file*. There's no need, IMO, to turn the whole fput() into 
> > > IRQ-callable just for this case, when we can contain it into the 
> > > particular KAIO+eventfd usage.
> > > 
> > 
> > Do you really want to see eventd doing umounts and remote flock() calls?
> > This really needs to be run in a thread that can cope with __long__
> > waits, unavailable servers, etc...
> 
> Did I miss something? This wouldn't be (eventually) on keventd shoulders, 
> but on aio work queue (aio_wq).
> I guess we could do the same optimization we're already doing for ki_filp, 
> for ki_eventfd ...

Last I checked, a call to schedule_work(&fddef->wq) would still run
fd_defer_queue() under keventd.

That said, even if you were to run it under aio_wq, the argument remains
the same: do you really want to add potentially long lasting sleeps onto
a work queue that was designed to service fast i/o requests?

Trond