Date: Fri, 23 Nov 2007 12:59:43 -0600
From: Matt Mackall <mpm@selenic.com>
To: Evgeniy Polyakov <johnpol@2ka.mipt.ru>
Cc: Andrew Morton <akpm@linux-foundation.org>,
       Simon Arlott <simon@fire.lp0.eu>,
       Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
       netdev@vger.kernel.org, Ingo Molnar <mingo@redhat.com>
Subject: Re: 2.6.23 WARNING: at kernel/softirq.c:139 local_bh_enable()
Message-ID: <20071123185943.GZ19691@waste.org>
References: <4745DCD7.8070805@simon.arlott.org.uk> <20071123002157.cb27f4a1.akpm@linux-foundation.org> <20071123105518.GA22062@2ka.mipt.ru> <20071123170756.GV19691@waste.org> <20071123175757.GA23991@2ka.mipt.ru>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20071123175757.GA23991@2ka.mipt.ru>
User-Agent: Mutt/1.5.13 (2006-08-11)
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 3620
Lines: 79

On Fri, Nov 23, 2007 at 08:57:57PM +0300, Evgeniy Polyakov wrote:
> On Fri, Nov 23, 2007 at 11:07:56AM -0600, Matt Mackall (mpm@selenic.com) wrote:
> > On Fri, Nov 23, 2007 at 01:55:19PM +0300, Evgeniy Polyakov wrote:
> > > On Fri, Nov 23, 2007 at 12:21:57AM -0800, Andrew Morton (akpm@linux-foundation.org) wrote:
> > > > > [2059664.615816] __iptables__: init4 IN=ppp0 OUT=ppp0 WARNING: at kernel/softirq.c:139 local_bh_enable()
> > > > > [2059664.620535]  [<80120364>] local_bh_enable+0x3c/0x97
> > > 
> > > > > [2059664.620657]  [<8011c205>] __call_console_drivers+0x61/0x6d
> > > > > [2059664.620669]  [<8011c3fc>] release_console_sem+0x164/0x1bf
> > > > > [2059664.620679]  [<8011c81f>] vprintk+0x27a/0x2ff
> > >  
> > > > If that trace is to be beieved we're doing nefilter stuff on packets which
> > > > were sent across netconsole.
> > > > 
> > > > This probably isn't anything the netfilter guys have thought about.  And
> > > > probably we don't want them to.  Is there some simple way in which we can
> > > > exempt netconsole from netfilter processing?
> > > 
> > > This is not about netfilter, but about freeing skb in interrupt context, 
> > > which is not allowed, and in interrupt skbs are queued to be freed in softirq,
> > > but netcnsole wants to flush softirq freeing queue. That is a question: why?
> > 
> > My memory here is hazy, but I think this exists to rescue netconsole
> > in low-memory situations. This bit originated with Ingo, so maybe he
> > can recall.
> > 
> > Netpoll can process an arbitrary number of skbs inside a single
> > interrupt. Think sysrq-t at one packet per line or kgdboe where the
> > entire trace session can happen inside one very long interrupt.
> > 
> > Perhaps we can refine this to mark netpoll's skbs (perhaps with
> > ->destructor?) and delete only skbs we own. As these are never passed
> > through any of the other route/xfrm/filter code, they should be safe
> > to delete even in irq context, yes?
> > 
> > > Removing zap_completion_queue() from find_skb() will fix the warning,
> > > but I'm not sure this is a correct fix. I've added Matt to the Cc list.
> > 
> > Care to try the sysrq-t or OOM message tests?
> 
> We basically can not free skbs there - if it is interrupt context and
> we are freeing some skb with destructor we will catch the warning anyway.

Perhaps I'm missing some context here. We don't free skbs with
destructors in irq context in zap_completion_queue. We reinsert them on the
completion list. We do this by calling dev_kfree_skb_any.

So I'd be surprised if that was a problem. But I can imagine having
problems for skbs without destructors which run into one of these in
__kfree_skb:

dst_release
secpath_put
nf_conntrack_put
nf_conntrack_put_reasm
nf_bridge_put

..some or all of which assume a softirq context.

> No matter if we are under memory pressure or whatever - it is not
> allowed - a lot of skbs are supposed to be freed in softirq context,
> that is why dev_kfree_skb_any() exists.

Some skbs we definitely -can- free in irq context. The only ones we
care about are the ones generated by netpoll. If there's a reason you
think netpoll's own skbs can't be freed, please describe it.

> I think we can drop skbs _without_ destructor from the queue though in
> that conditions given that we actually need only one.

Huh?

-- 
Mathematics is the supreme nostalgia of our time.
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/