Re: knlist_empty locking fix

On Friday, January 27, 2012 2:14:55 pm Doug Ambrisko wrote:
> John Baldwin writes:
> | On Friday, January 27, 2012 12:52:18 pm Kostik Belousov wrote:
> | > On Fri, Jan 27, 2012 at 09:42:58AM -0800, Doug Ambrisko wrote:
> | > > Andrew Boyer writes:
> | > > | On Jan 27, 2012, at 12:15 PM, Doug Ambrisko wrote:
> | > > | 
> | > > | > John Baldwin writes:
> | > > | > | Agreed, I think the missing locking should just be added to the 
aio 
> | code.
> | > > | > 
> | > > | > Okay so then just:
> | > > | > 
> | > > | > Index: vfs_aio.c
> | > > | > 
===================================================================
> | > > | > RCS file: /usr/local/cvsroot/freebsd/src/sys/kern/vfs_aio.c,v
> | > > | > retrieving revision 1.243.2.3.4.1
> | > > | > diff -u -p -r1.243.2.3.4.1 vfs_aio.c
> | > > | > --- vfs_aio.c	21 Dec 2010 17:09:25 -0000	1.243.2.3.4.1
> | > > | > +++ vfs_aio.c	27 Jan 2012 17:07:11 -0000
> | > > | > _at__at_ -2509,9 +2509,12 _at__at_ static void
> | > > | > filt_aiodetach(struct knote *kn)
> | > > | > {
> | > > | > 	struct aiocblist *aiocbe = kn->kn_ptr.p_aio;
> | > > | > +	struct knlist *knl = &aiocbe->klist;
> | > > | > 
> | > > | > -	if (!knlist_empty(&aiocbe->klist))
> | > > | > -		knlist_remove(&aiocbe->klist, kn, 0);
> | > > | > +	knl->kl_lock(knl->kl_lockarg);
> | > > | > +	if (!knlist_empty(knl))
> | > > | > +		knlist_remove(knl, kn, 1);
> | > > | > +	knl->kl_unlock(knl->kl_lockarg);
> | > > | > }
> | > > | > 
> | > > | > /* kqueue filter function */
> | > > | > 
> | > > | > I was trying to be consistant with knlist_remove but this is a 
much
> | > > | > smaller change that can be merge to older branches.
> | > > |  
> | > > | Does filt_liodetach() need the same treatment?
> | > > 
> | > > Yes, I had that in the original.  I updated that and optimized
> | > > the knl to just get the structure needed.
> | > > 
> | > > Index: vfs_aio.c
> | > > ===================================================================
> | > > RCS file: /usr/local/cvsroot/freebsd/src/sys/kern/vfs_aio.c,v
> | > > retrieving revision 1.243.2.3.4.1
> | > > diff -u -p -r1.243.2.3.4.1 vfs_aio.c
> | > > --- vfs_aio.c	21 Dec 2010 17:09:25 -0000	1.243.2.3.4.1
> | > > +++ vfs_aio.c	27 Jan 2012 17:35:47 -0000
> | > > _at__at_ -2508,10 +2508,12 _at__at_ filt_aioattach(struct knote *kn)
> | > >  static void
> | > >  filt_aiodetach(struct knote *kn)
> | > >  {
> | > > -	struct aiocblist *aiocbe = kn->kn_ptr.p_aio;
> | > > +	struct knlist *knl = &kn->kn_ptr.p_aio->klist;
> | > >  
> | > > -	if (!knlist_empty(&aiocbe->klist))
> | > > -		knlist_remove(&aiocbe->klist, kn, 0);
> | > > +	knl->kl_lock(knl->kl_lockarg);
> | > > +	if (!knlist_empty(knl))
> | > > +		knlist_remove(knl, kn, 1);
> | > > +	knl->kl_unlock(knl->kl_lockarg);
> | > >  }
> | > >  
> | > >  /* kqueue filter function */
> | > > _at__at_ -2553,10 +2555,12 _at__at_ filt_lioattach(struct knote *kn)
> | > >  static void
> | > >  filt_liodetach(struct knote *kn)
> | > >  {
> | > > -	struct aioliojob * lj = kn->kn_ptr.p_lio;
> | > > +	struct knlist *knl = &kn->kn_ptr.p_lio->klist;
> | > It is easy to be style-compiant there and move initialization of knl
> | > after the blank line.
> | > 
> | > Do you need two different functions now ? I think you can leave just 
one.
> | 
> | Hmm, I think p_lio != p_aio, so two functions are required.
> 
> Correct, even thought they are together in a union, the structures are
> different so klist would have a different offset for each function.
>  
> | I think the patch looks fine.  The style fix to not assign 'knl' in its
> | declaration would be nice to fix as you suggested, but that's minor.
> 
> Here is the updated version that seems to work fine.

Looks good to me.

-- 
John Baldwin