Re: [PATCH 2/4] Convert epoll to a bitlock

[Andrew Morton <akpm@linux-foundation.org>]

On Tue, 3 Feb 2009 16:19:31 -0700
Jonathan Corbet <corbet@lwn.net> wrote:

> On Tue, 3 Feb 2009 14:53:46 -0800
> Andrew Morton <akpm@linux-foundation.org> wrote:
> 
> > Well.  We _could_ whack part of this nut with my usual hammer: protect
> > f_flags with file->f_dentry->d_inode->i_lock.  IIRC there was some
> > objection to that - performance?
> 
> Andi has objected to the addition of locks, but i_lock is maybe
> sufficiently dispersed to pass muster there.

Hope so.

I'd wrap it in a lock_file_flags(file*) thing so we can change it later
on (add a lock to struct file, take a global, lock, etc).

>  I had an instinctive
> reaction to using a lock which is three pointers away, but I can get
> over that.  I'll admit a bit of ignorance, though: if a given struct
> file exists, do we know for sure that file->f_dentry->d_inode exists?

It should.  A NULL ->d_inode especially signifies a negative dentry.

> > One problem here seems to be that we're trying to change multiple
> > things at the same time.  We can blame the BKL for that.
> > 
> > Can we break the problem into manageable chunks?  Your patchset did
> > that, I guess.  What were those chunks again? ;)
> 
> I'm not really sure how to break it down any further.  If we take the
> i_lock approach, the chunks would be something like:
> 
>  1) Use i_lock to protect accesses to f_flags.  This would enable some 
>     BKL usage to be removed, but would not fix fasync.
> 
>  2) Move responsibility for the FASYNC bit into ->fasync(), with
>     fasync_helper() doing it in almost all situations.  The remaining
>     BKL usage would then go away.
> 
>  3) The same optional fasync() return values cleanup.
> 
> Make sense?

yup.

If the ->i_lock think is no good then we can trivially switch over to a
global lock.  Heck, we could even go back to lock_kernel() ;)