From mboxrd@z Thu Jan 1 00:00:00 1970 From: Andy Gospodarek Subject: Re: [Bugme-new] [Bug 9543] New: RTNL: assertion failed at net/ipv6/addrconf.c (2164)/RTNL: assertion failed at net/ipv4/devinet.c (1055) Date: Wed, 19 Dec 2007 09:42:08 -0500 Message-ID: <20071219144208.GB8728@gospo.usersys.redhat.com> References: <28503.1197481615@death> <20071214182638.GC25879@gospo.usersys.redhat.com> <20071214224722.GA8728@gospo.usersys.redhat.com> Mime-Version: 1.0 Content-Type: text/plain; charset=utf-8 Content-Transfer-Encoding: QUOTED-PRINTABLE Cc: Andy Gospodarek , Jay Vosburgh , Herbert Xu , Andrew Morton , bugme-daemon@bugzilla.kernel.org, shemminger@linux-foundation.org, davem@davemloft.net, netdev@vger.kernel.org To: Krzysztof Oledzki Return-path: Received: from mx1.redhat.com ([66.187.233.31]:48560 "EHLO mx1.redhat.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752713AbXLSOmo (ORCPT ); Wed, 19 Dec 2007 09:42:44 -0500 Content-Disposition: inline In-Reply-To: Sender: netdev-owner@vger.kernel.org List-ID: On Tue, Dec 18, 2007 at 08:53:39PM +0100, Krzysztof Oledzki wrote: >=20 >=20 > On Fri, 14 Dec 2007, Andy Gospodarek wrote: >=20 > >On Fri, Dec 14, 2007 at 07:57:42PM +0100, Krzysztof Oledzki wrote: > >> > >> > >>On Fri, 14 Dec 2007, Andy Gospodarek wrote: > >> > >>>On Fri, Dec 14, 2007 at 05:14:57PM +0100, Krzysztof Oledzki wrote: > >>>> > >>>> > >>>>On Wed, 12 Dec 2007, Jay Vosburgh wrote: > >>>> > >>>>>Herbert Xu wrote: > >>>>> > >>>>>>>diff -puN drivers/net/bonding/bond_sysfs.c~bonding-locking-fix > >>>>>>>drivers/net/bonding/bond_sysfs.c > >>>>>>>--- a/drivers/net/bonding/bond_sysfs.c~bonding-locking-fix > >>>>>>>+++ a/drivers/net/bonding/bond_sysfs.c > >>>>>>>@@ -1111,8 +1111,6 @@ static ssize_t bonding_store_primary(str > >>>>>>>out: > >>>>>>> write_unlock_bh(&bond->lock); > >>>>>>> > >>>>>>>- rtnl_unlock(); > >>>>>>>- > >>>>>> > >>>>>>Looking at the changeset that added this perhaps the intention > >>>>>>is to hold the lock? If so we should add an rtnl_lock to the st= art > >>>>>>of the function. > >>>>> > >>>>> Yes, this function needs to hold locks, and more than just > >>>>>what's there now. I believe the following should be correct; I = haven't > >>>>>tested it, though (I'm supposedly on vacation right now). > >>>>> > >>>>> The following change should be correct for the > >>>>>bonding_store_primary case discussed in this thread, and also co= rrects > >>>>>the bonding_store_active case which performs similar functions. > >>>>> > >>>>> The bond_change_active_slave and bond_select_active_slave > >>>>>functions both require rtnl, bond->lock for read and curr_slave_= lock=20 > >>>>>for > >>>>>write_bh, and no other locks. This is so that the lower level > >>>>>mode-specific functions can release locks down to just rtnl in o= rder to > >>>>>call, e.g., dev_set_mac_address with the locks it expects (rtnl = only). > >>>>> > >>>>>Signed-off-by: Jay Vosburgh > >>>>> > >>>>>diff --git a/drivers/net/bonding/bond_sysfs.c > >>>>>b/drivers/net/bonding/bond_sysfs.c > >>>>>index 11b76b3..28a2d80 100644 > >>>>>--- a/drivers/net/bonding/bond_sysfs.c > >>>>>+++ b/drivers/net/bonding/bond_sysfs.c > >>>>>@@ -1075,7 +1075,10 @@ static ssize_t bonding_store_primary(stru= ct=20 > >>>>>device > >>>>>*d, > >>>>> struct slave *slave; > >>>>> struct bonding *bond =3D to_bond(d); > >>>>> > >>>>>- write_lock_bh(&bond->lock); > >>>>>+ rtnl_lock(); > >>>>>+ read_lock(&bond->lock); > >>>>>+ write_lock_bh(&bond->curr_slave_lock); > >>>>>+ > >>>>> if (!USES_PRIMARY(bond->params.mode)) { > >>>>> printk(KERN_INFO DRV_NAME > >>>>> ": %s: Unable to set primary slave; %s is in mode > >>>>> %d\n", > >>>>>@@ -1109,8 +1112,8 @@ static ssize_t bonding_store_primary(struc= t=20 > >>>>>device > >>>>>*d, > >>>>> } > >>>>> } > >>>>>out: > >>>>>- write_unlock_bh(&bond->lock); > >>>>>- > >>>>>+ write_unlock_bh(&bond->curr_slave_lock); > >>>>>+ read_unlock(&bond->lock); > >>>>> rtnl_unlock(); > >>>>> > >>>>> return count; > >>>>>@@ -1190,7 +1193,8 @@ static ssize_t bonding_store_active_slave(= struct > >>>>>device *d, > >>>>> struct bonding *bond =3D to_bond(d); > >>>>> > >>>>> rtnl_lock(); > >>>>>- write_lock_bh(&bond->lock); > >>>>>+ read_lock(&bond->lock); > >>>>>+ write_lock_bh(&bond->curr_slave_lock); > >>>>> > >>>>> if (!USES_PRIMARY(bond->params.mode)) { > >>>>> printk(KERN_INFO DRV_NAME > >>>>>@@ -1247,7 +1251,8 @@ static ssize_t bonding_store_active_slave(= struct > >>>>>device *d, > >>>>> } > >>>>> } > >>>>>out: > >>>>>- write_unlock_bh(&bond->lock); > >>>>>+ write_unlock_bh(&bond->curr_slave_lock); > >>>>>+ read_unlock(&bond->lock); > >>>>> rtnl_unlock(); > >>>>> > >>>>> return count; > >>>> > >>>>Vanilla 2.6.24-rc5 plus this patch: > >>>> > >>>>=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D= =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D= =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D > >>>>[ INFO: possible irq lock inversion dependency detected ] > >>>>2.6.24-rc5 #1 > >>>>--------------------------------------------------------- > >>>>events/0/9 just changed the state of lock: > >>>>(&mc->mca_lock){-+..}, at: [] mld_ifc_timer_expire+0x13= 0/0x1fb > >>>>but this lock took another, soft-read-irq-unsafe lock in the past= : > >>>>(&bond->lock){-.--} > >>>> > >>>>and interrupts could create inverse lock ordering between them. > >>>> > >>>> > >>> > >>>Grrr, I should have seen that -- sorry. Try your luck with this i= nstead: > >> > >> > >>No luck. > >> > > > > > >I'm guessing if we go back to using a write-lock for bond->lock this > >will go back to working again, but I'm not totally convinced since t= here > >are plenty of places where we used a read-lock with it. >=20 > Should I check this patch or rather, based on a future discussion, wa= it=20 > for another version? >=20 > > > >diff --git a/drivers/net/bonding/bond_sysfs.c=20 > >b/drivers/net/bonding/bond_sysfs.c > >index 11b76b3..635b857 100644 > >--- a/drivers/net/bonding/bond_sysfs.c > >+++ b/drivers/net/bonding/bond_sysfs.c > >@@ -1075,7 +1075,10 @@ static ssize_t bonding_store_primary(struct d= evice=20 > >*d, > > struct slave *slave; > > struct bonding *bond =3D to_bond(d); > > > >+ rtnl_lock(); > > write_lock_bh(&bond->lock); > >+ write_lock_bh(&bond->curr_slave_lock); > >+ > > if (!USES_PRIMARY(bond->params.mode)) { > > printk(KERN_INFO DRV_NAME > > ": %s: Unable to set primary slave; %s is in mode=20 > > %d\n", > >@@ -1109,8 +1112,8 @@ static ssize_t bonding_store_primary(struct de= vice=20 > >*d, > > } > > } > >out: > >+ write_unlock_bh(&bond->curr_slave_lock); > > write_unlock_bh(&bond->lock); > >- > > rtnl_unlock(); > > > > return count; > >@@ -1191,6 +1194,7 @@ static ssize_t bonding_store_active_slave(stru= ct=20 > >device *d, > > > > rtnl_lock(); > > write_lock_bh(&bond->lock); > >+ write_lock_bh(&bond->curr_slave_lock); > > > > if (!USES_PRIMARY(bond->params.mode)) { > > printk(KERN_INFO DRV_NAME > >@@ -1247,6 +1251,7 @@ static ssize_t bonding_store_active_slave(stru= ct=20 > >device *d, > > } > > } > >out: > >+ write_unlock_bh(&bond->curr_slave_lock); > > write_unlock_bh(&bond->lock); > > rtnl_unlock(); > > >=20 >=20 > Best regards, >=20 > Krzysztof Ol=C4=99dzki =46or now, I prefer Jay's original patch -- with the read_locks (rather than read/write_lock_bh) and the added rtnl_lock. There is still a lockdep issue that we need to sort-out, but this patch is needed first.