Message-ID: <20150610133142.GB4062@localhost.localdomain>
Date:	Wed, 10 Jun 2015 10:31:42 -0300
From:	Marcelo Ricardo Leitner <mleitner@...hat.com>
To:	Neil Horman <nhorman@...driver.com>
Cc:	Hannes Frederic Sowa <hannes@...essinduktion.org>,
	netdev@...r.kernel.org, linux-sctp@...r.kernel.org,
	Daniel Borkmann <daniel@...earbox.net>,
	Vlad Yasevich <vyasevich@...il.com>,
	Michio Honda <micchie@....wide.ad.jp>
Subject: Re: [PATCH v3 1/2] sctp: rcu-ify addr_waitq

On Tue, Jun 09, 2015 at 04:32:59PM -0300, Marcelo Ricardo Leitner wrote:
> On Tue, Jun 09, 2015 at 07:36:38AM -0400, Neil Horman wrote:
> > On Mon, Jun 08, 2015 at 05:37:05PM +0200, Hannes Frederic Sowa wrote:
> > > On Mo, 2015-06-08 at 11:19 -0400, Neil Horman wrote:
> > > > On Mon, Jun 08, 2015 at 04:59:18PM +0200, Hannes Frederic Sowa wrote:
> > > > > On Mon, Jun 8, 2015, at 16:46, Hannes Frederic Sowa wrote:
> > > > > > Hi Marcelo,
> > > > > > 
> > > > > > a few hints on rcuification, sorry I reviewed the code so late:
> > > > > > 
> > > > > > On Fri, Jun 5, 2015, at 19:08, mleitner@...hat.com wrote:
> > > > > > > From: Marcelo Ricardo Leitner <marcelo.leitner@...il.com>
> > > > > > > 
> > > > > > > That's needed for the next patch, so we break the lock 
> > > > > > > inversion between
> > > > > > > netns_sctp->addr_wq_lock and socket lock on
> > > > > > > sctp_addr_wq_timeout_handler(). With this, we can traverse 
> > > > > > > addr_waitq
> > > > > > > without taking addr_wq_lock, taking it just for the write 
> > > > > > > operations.
> > > > > > > 
> > > > > > > Signed-off-by: Marcelo Ricardo Leitner <
> > > > > > > marcelo.leitner@...il.com>
> > > > > > > ---
> > > > > > > 
> > > > > > > Notes:
> > > > > > >     v2->v3:
> > > > > > >       placed break statement on sctp_free_addr_wq_entry()
> > > > > > >       removed unnecessary spin_lock noticed by Neil
> > > > > > > 
> > > > > > >  include/net/netns/sctp.h |  2 +-
> > > > > > >  net/sctp/protocol.c      | 80
> > > > > > >  +++++++++++++++++++++++++++++-------------------
> > > > > > >  2 files changed, 49 insertions(+), 33 deletions(-)
> > > > > > > 
> > > > > > > diff --git a/include/net/netns/sctp.h 
> > > > > > > b/include/net/netns/sctp.h
> > > > > > > index
> > > > > > > 3573a81815ad9e0efb6ceb721eb066d3726419f0..9e53412c4ed829e8e4577
> > > > > > > 7a6d95406d490dbaa75
> > > > > > > 100644
> > > > > > > --- a/include/net/netns/sctp.h
> > > > > > > +++ b/include/net/netns/sctp.h
> > > > > > > @@ -28,7 +28,7 @@ struct netns_sctp {
> > > > > > >          * It is a list of sctp_sockaddr_entry.
> > > > > > >          */
> > > > > > >         struct list_head local_addr_list;
> > > > > > > -       struct list_head addr_waitq;
> > > > > > > +       struct list_head __rcu addr_waitq;
> > > > > > >         struct timer_list addr_wq_timer;
> > > > > > >         struct list_head auto_asconf_splist;
> > > > > > >         spinlock_t addr_wq_lock;
> > > > > > > diff --git a/net/sctp/protocol.c b/net/sctp/protocol.c
> > > > > > > index
> > > > > > > 53b7acde9aa37bf3d4029c459421564d5270f4c0..9954fb8c9a9455d5ad7a6
> > > > > > > 27e2d7f9a1fef861fc2
> > > > > > > 100644
> > > > > > > --- a/net/sctp/protocol.c
> > > > > > > +++ b/net/sctp/protocol.c
> > > > > > > @@ -593,15 +593,47 @@ static void sctp_v4_ecn_capable(struct 
> > > > > > > sock *sk)
> > > > > > >         INET_ECN_xmit(sk);
> > > > > > >  }
> > > > > > >  
> > > > > > > +static void sctp_free_addr_wq(struct net *net)
> > > > > > > +{
> > > > > > > +       struct sctp_sockaddr_entry *addrw;
> > > > > > > +
> > > > > > > +       spin_lock_bh(&net->sctp.addr_wq_lock);
> > > > > > 
> > > > > > Instead of holding spin_lock_bh you need to hold 
> > > > > > rcu_read_lock_bh, so
> > > > > > kfree_rcu does not call free function at once (in theory ;) ).
> > > > > > 
> > > > > > > +       del_timer(&net->sctp.addr_wq_timer);
> > > > > > > +       list_for_each_entry_rcu(addrw, &net->sctp.addr_waitq, 
> > > > > > > list) {
> > > > > > > +               list_del_rcu(&addrw->list);
> > > > > > > +               kfree_rcu(addrw, rcu);
> > > > > > > +       }
> > > > > > > +       spin_unlock_bh(&net->sctp.addr_wq_lock);
> > > > > > > +}
> > > > > > > +
> > > > > > > +/* As there is no refcnt on sctp_sockaddr_entry, we must check 
> > > > > > > inside
> > > > > > > + * the lock if it wasn't removed from addr_waitq already, 
> > > > > > > otherwise we
> > > > > > > + * could double-free it.
> > > > > > > + */
> > > > > > > +static void sctp_free_addr_wq_entry(struct net *net,
> > > > > > > +                                   struct sctp_sockaddr_entry 
> > > > > > > *addrw)
> > > > > > > +{
> > > > > > > +       struct sctp_sockaddr_entry *temp;
> > > > > > > +
> > > > > > > +       spin_lock_bh(&net->sctp.addr_wq_lock);
> > > > > > 
> > > > > > I don't think this spin_lock operation is needed. The del_timer
> > > > > > functions do synchronize themselves.
> > > > > > 
> > > > > 
> > > > > Sorry, those above two locks are needed, they are not implied by 
> > > > > other
> > > > > locks.
> > > > > 
> > > > What makes you say that? Multiple contexts can issue mod_timer calls 
> > > > on the
> > > > same timer safely no, because of the internal locking?
> > > 
> > > That's true for timer handling but not to protect net->sctp.addr_waitq
> > > list (Marcelo just explained it to me off-list). Looking at the patch
> > > only in patchworks lost quite a lot of context you were already
> > > discussing. ;)
> > > 
> > I can imagine :)
> > 
> > > We are currently checking if the double iteration can be avoided by
> > > splicing addr_waitq on the local stack while holding the spin_lock and
> > > later on notifying the sockets.
> > > 
> > As we discussed, this I think would make a good alternate approach.
> 
> I was experimenting on this but this would introduce another complex
> logic instead, as not all elements are pruned from net->sctp.addr_waitq
> at sctp_addr_wq_timeout_handler(), mainly ipv6 addresses in DAD state
> (which I believe that break statement is misplaced and should be a
> continue instead, I'll check on this later)
> 
> That means we would have to do the splice, process the loop, merge the
> remaining elements with the new net->sctp.addr_waitq that was possibly
> was built meanwhile and then squash oppositve events (logic currently in
> sctp_addr_wq_mgmt() ), otherwise we could be issuing spurious events.
> 
> But it will probably do more harm than good as the double search will
> usually hit the first list element on this 2nd search, unless the
> element we are trying to remove was already removed from it (which is
> rare, it's when user add and remove addresses too fast) or some other
> address was skipped (DAD addresses).

Better thinking.. actually it may be the way to go. If we rcu-cify
addr_waitq like that and if the user manage to add an address and remove
it while the timeout handler is running, the system may emit just the
address add and not the remove, while if we splice the list, this won't
happen. 

  Marcelo

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

Powered by blists - more mailing lists