diff mbox

[1/7] nfsd4: fix recall_lock use in unhash_delegation

Message ID 20131030143552.GE3227@fieldses.org (mailing list archive)
State New, archived
Headers show

Commit Message

J. Bruce Fields Oct. 30, 2013, 2:35 p.m. UTC
On Wed, Oct 30, 2013 at 04:02:47PM +0200, Benny Halevy wrote:
> On 2013-10-29 16:46, J. Bruce Fields wrote:
> > On Tue, Oct 29, 2013 at 11:39:04AM +0200, Benny Halevy wrote:
> >> Access to dp->dl_perclnt must be synchronized by the recall_lock
> > 
> > Are you sure?  recall_lock is for stuff that's needed in the delegation
> > break callback (nfsd_break_deleg_cb() and any of the subsequent callback
> > handling).  I don't think that includes dl_perclnt.
> 
> I was mislead by the fact that destroy_client and nfs4_state_shutdown_net
> care to acquire the recall_lock for traversing the clp->cl_delegations
> and nn->del_recall_lru lists, respectively.
> 
> Now nfs4_state_shutdown_net, although its comment says it should be
> called with the state lock held (SHOULD or should, or should it be MUST? :)
> It doesn't seem like nfsd_shutdown_net acquires the state lock.

Good point.

> Therefore, we either need to grab the state lock in nfs4_state_shutdown_net
> which adds yet another problem to fix,

That sounds simplest as the short-term fix....  What problem do you see?

--b.

commit 2c9407ae694dad8a1ad6394d9bf6077024b9fae7
Author: J. Bruce Fields <bfields@redhat.com>
Date:   Wed Oct 30 10:33:09 2013 -0400

    nfsd4: nfsd_shutdown_net needs state lock
    
    A comment claims the caller should take it, but that's not being done.
    Note we don't want it around the cancel_delayed_work_sync since that may
    wait on work which holds the client lock.
    
    Signed-off-by: J. Bruce Fields <bfields@redhat.com>

--
To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Comments

Benny Halevy Oct. 30, 2013, 2:43 p.m. UTC | #1
On 2013-10-30 16:35, J. Bruce Fields wrote:
> On Wed, Oct 30, 2013 at 04:02:47PM +0200, Benny Halevy wrote:
>> On 2013-10-29 16:46, J. Bruce Fields wrote:
>>> On Tue, Oct 29, 2013 at 11:39:04AM +0200, Benny Halevy wrote:
>>>> Access to dp->dl_perclnt must be synchronized by the recall_lock
>>>
>>> Are you sure?  recall_lock is for stuff that's needed in the delegation
>>> break callback (nfsd_break_deleg_cb() and any of the subsequent callback
>>> handling).  I don't think that includes dl_perclnt.
>>
>> I was mislead by the fact that destroy_client and nfs4_state_shutdown_net
>> care to acquire the recall_lock for traversing the clp->cl_delegations
>> and nn->del_recall_lru lists, respectively.
>>
>> Now nfs4_state_shutdown_net, although its comment says it should be
>> called with the state lock held (SHOULD or should, or should it be MUST? :)
>> It doesn't seem like nfsd_shutdown_net acquires the state lock.
> 
> Good point.
> 
>> Therefore, we either need to grab the state lock in nfs4_state_shutdown_net
>> which adds yet another problem to fix,
> 
> That sounds simplest as the short-term fix....  What problem do you see?

Nothing specific, just the general use of the global state lock...

> 
> --b.
> 
> commit 2c9407ae694dad8a1ad6394d9bf6077024b9fae7
> Author: J. Bruce Fields <bfields@redhat.com>
> Date:   Wed Oct 30 10:33:09 2013 -0400
> 
>     nfsd4: nfsd_shutdown_net needs state lock
>     
>     A comment claims the caller should take it, but that's not being done.
>     Note we don't want it around the cancel_delayed_work_sync since that may
>     wait on work which holds the client lock.
>     
>     Signed-off-by: J. Bruce Fields <bfields@redhat.com>

Ack

> 
> diff --git a/fs/nfsd/nfs4state.c b/fs/nfsd/nfs4state.c
> index 21eb678..e03e8ef 100644
> --- a/fs/nfsd/nfs4state.c
> +++ b/fs/nfsd/nfs4state.c
> @@ -5124,7 +5124,6 @@ out_recovery:
>  	return ret;
>  }
>  
> -/* should be called with the state lock held */
>  void
>  nfs4_state_shutdown_net(struct net *net)
>  {
> @@ -5135,6 +5134,7 @@ nfs4_state_shutdown_net(struct net *net)
>  	cancel_delayed_work_sync(&nn->laundromat_work);
>  	locks_end_grace(&nn->nfsd4_manager);
>  
> +	nfs4_lock_state();
>  	INIT_LIST_HEAD(&reaplist);
>  	spin_lock(&recall_lock);
>  	list_for_each_safe(pos, next, &nn->del_recall_lru) {
> @@ -5149,6 +5149,7 @@ nfs4_state_shutdown_net(struct net *net)
>  
>  	nfsd4_client_tracking_exit(net);
>  	nfs4_state_destroy_net(net);
> +	nfs4_unlock_state();
>  }
>  
>  void
> --
> To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> 
--
To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
J. Bruce Fields Oct. 30, 2013, 10:19 p.m. UTC | #2
On Wed, Oct 30, 2013 at 04:43:03PM +0200, Benny Halevy wrote:
> On 2013-10-30 16:35, J. Bruce Fields wrote:
> > On Wed, Oct 30, 2013 at 04:02:47PM +0200, Benny Halevy wrote:
> >> On 2013-10-29 16:46, J. Bruce Fields wrote:
> >>> On Tue, Oct 29, 2013 at 11:39:04AM +0200, Benny Halevy wrote:
> >>>> Access to dp->dl_perclnt must be synchronized by the recall_lock
> >>>
> >>> Are you sure?  recall_lock is for stuff that's needed in the delegation
> >>> break callback (nfsd_break_deleg_cb() and any of the subsequent callback
> >>> handling).  I don't think that includes dl_perclnt.
> >>
> >> I was mislead by the fact that destroy_client and nfs4_state_shutdown_net
> >> care to acquire the recall_lock for traversing the clp->cl_delegations
> >> and nn->del_recall_lru lists, respectively.
> >>
> >> Now nfs4_state_shutdown_net, although its comment says it should be
> >> called with the state lock held (SHOULD or should, or should it be MUST? :)
> >> It doesn't seem like nfsd_shutdown_net acquires the state lock.
> > 
> > Good point.
> > 
> >> Therefore, we either need to grab the state lock in nfs4_state_shutdown_net
> >> which adds yet another problem to fix,
> > 
> > That sounds simplest as the short-term fix....  What problem do you see?
> 
> Nothing specific, just the general use of the global state lock...
> 
> > 
> > --b.
> > 
> > commit 2c9407ae694dad8a1ad6394d9bf6077024b9fae7
> > Author: J. Bruce Fields <bfields@redhat.com>
> > Date:   Wed Oct 30 10:33:09 2013 -0400
> > 
> >     nfsd4: nfsd_shutdown_net needs state lock
> >     
> >     A comment claims the caller should take it, but that's not being done.
> >     Note we don't want it around the cancel_delayed_work_sync since that may
> >     wait on work which holds the client lock.
> >     
> >     Signed-off-by: J. Bruce Fields <bfields@redhat.com>
> 
> Ack

Thanks, done, and pushed out to

	git://linux-nfs.org/~bfields/linux.git for-3.13

--b.

> 
> > 
> > diff --git a/fs/nfsd/nfs4state.c b/fs/nfsd/nfs4state.c
> > index 21eb678..e03e8ef 100644
> > --- a/fs/nfsd/nfs4state.c
> > +++ b/fs/nfsd/nfs4state.c
> > @@ -5124,7 +5124,6 @@ out_recovery:
> >  	return ret;
> >  }
> >  
> > -/* should be called with the state lock held */
> >  void
> >  nfs4_state_shutdown_net(struct net *net)
> >  {
> > @@ -5135,6 +5134,7 @@ nfs4_state_shutdown_net(struct net *net)
> >  	cancel_delayed_work_sync(&nn->laundromat_work);
> >  	locks_end_grace(&nn->nfsd4_manager);
> >  
> > +	nfs4_lock_state();
> >  	INIT_LIST_HEAD(&reaplist);
> >  	spin_lock(&recall_lock);
> >  	list_for_each_safe(pos, next, &nn->del_recall_lru) {
> > @@ -5149,6 +5149,7 @@ nfs4_state_shutdown_net(struct net *net)
> >  
> >  	nfsd4_client_tracking_exit(net);
> >  	nfs4_state_destroy_net(net);
> > +	nfs4_unlock_state();
> >  }
> >  
> >  void
> > --
> > To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
> > the body of a message to majordomo@vger.kernel.org
> > More majordomo info at  http://vger.kernel.org/majordomo-info.html
> > 
--
To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
diff mbox

Patch

diff --git a/fs/nfsd/nfs4state.c b/fs/nfsd/nfs4state.c
index 21eb678..e03e8ef 100644
--- a/fs/nfsd/nfs4state.c
+++ b/fs/nfsd/nfs4state.c
@@ -5124,7 +5124,6 @@  out_recovery:
 	return ret;
 }
 
-/* should be called with the state lock held */
 void
 nfs4_state_shutdown_net(struct net *net)
 {
@@ -5135,6 +5134,7 @@  nfs4_state_shutdown_net(struct net *net)
 	cancel_delayed_work_sync(&nn->laundromat_work);
 	locks_end_grace(&nn->nfsd4_manager);
 
+	nfs4_lock_state();
 	INIT_LIST_HEAD(&reaplist);
 	spin_lock(&recall_lock);
 	list_for_each_safe(pos, next, &nn->del_recall_lru) {
@@ -5149,6 +5149,7 @@  nfs4_state_shutdown_net(struct net *net)
 
 	nfsd4_client_tracking_exit(net);
 	nfs4_state_destroy_net(net);
+	nfs4_unlock_state();
 }
 
 void