[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20200512093231.7ce29f30@kicinski-fedora-pc1c0hjn.dhcp.thefacebook.com>
Date: Tue, 12 May 2020 09:32:31 -0700
From: Jakub Kicinski <kuba@...nel.org>
To: Madhuparna Bhowmik <madhuparnabhowmik10@...il.com>
Cc: sfr@...b.auug.org.au, Amol Grover <frextrite@...il.com>,
"David S . Miller" <davem@...emloft.net>,
Alexey Kuznetsov <kuznet@....inr.ac.ru>,
Hideaki YOSHIFUJI <yoshfuji@...ux-ipv6.org>,
netdev@...r.kernel.org, linux-kernel@...r.kernel.org,
linux-kernel-mentees@...ts.linuxfoundation.org,
Joel Fernandes <joel@...lfernandes.org>,
"Paul E . McKenney" <paulmck@...nel.org>
Subject: Re: [PATCH net 2/2 RESEND] ipmr: Add lockdep expression to
ipmr_for_each_table macro
On Tue, 12 May 2020 10:47:05 +0530 Madhuparna Bhowmik wrote:
> > > #ifdef CONFIG_IP_MROUTE_MULTIPLE_TABLES
> > > -#define ipmr_for_each_table(mrt, net) \
> > > - list_for_each_entry_rcu(mrt, &net->ipv4.mr_tables, list, \
> > > - lockdep_rtnl_is_held())
> > > +#define ipmr_for_each_table(mrt, net) \
> > > + list_for_each_entry_rcu(mrt, &net->ipv4.mr_tables, list, \
> > > + lockdep_rtnl_is_held() || \
> > > + lockdep_is_held(&pernet_ops_rwsem))
> >
> > This is a strange condition, IMHO. How can we be fine with either
> > lock.. This is supposed to be the writer side lock, one can't have
> > two writer side locks..
> >
> > I think what is happening is this:
> >
> > ipmr_net_init() -> ipmr_rules_init() -> ipmr_new_table()
> >
> > ipmr_new_table() returns an existing table if there is one, but
> > obviously none can exist at init. So a better fix would be:
> >
> > #define ipmr_for_each_table(mrt, net) \
> > list_for_each_entry_rcu(mrt, &net->ipv4.mr_tables, list, \
> > lockdep_rtnl_is_held() || \
> > list_empty(&net->ipv4.mr_tables))
> >
> (adding Stephen)
>
> Hi Jakub,
>
> Thank you for your suggestion about this patch.
> Here is a stack trace for ipmr.c:
>
> [...]
Thanks!
> > Thoughts?
>
> Do you think a similar fix (the one you suggested) is also applicable
> in the ip6mr case.
Yes, looking at the code it seems ip6mr has the exact same flow for
netns init.
Powered by blists - more mailing lists