Re: [syzbot] [netfilter?] possible deadlock in nf_tables_dumpreset_obj

[Pablo Neira Ayuso <pablo@netfilter.org>]

Sorry, I pressed sent too fast... see below.

On Mon, Dec 22, 2025 at 10:22:02AM +0100, Pablo Neira Ayuso wrote:
> On Mon, Dec 22, 2025 at 12:16:53AM +0100, Florian Westphal wrote:
> > syzbot <syzbot+ff16b505ec9152e5f448@syzkaller.appspotmail.com> wrote:
> > > syz.3.970/9330 is trying to acquire lock:
> > > ffff888012d4ccd8 (&nft_net->commit_mutex){+.+.}-{4:4}, at: nf_tables_dumpreset_obj+0x6f/0xa0 net/netfilter/nf_tables_api.c:8491
> > > 
> > > but task is already holding lock:
> > > ffff88802bce36f0 (nlk_cb_mutex-NETFILTER){+.+.}-{4:4}, at: __netlink_dump_start+0x150/0x990 net/netlink/af_netlink.c:2404
> > > 
> > > which lock already depends on the new lock.
> > 
> > I think this is a real bug:
> 
> Yes, I think so too, it was a bad idea to use the commit_mutex for this.
> 
> > CPU0: 'nft reset'.
> > CPU1: 'ipset list' (anything in ipset doing a netlink dump op)
> > CPU2: 'iptables-nft -A ... -m set ...'
> > 
> > ... can result in:
> > 
> > CPU0                    CPU1                            CPU2
> > ----                    ----                            ----
> > lock(nlk_cb_mutex-NETFILTER);
> >                         lock(nfnl_subsys_ipset);
> >                                                        lock(&nft_net->commit_mutex);
> >                         lock(nlk_cb_mutex-NETFILTER);
> >                                                        lock(nfnl_subsys_ipset);
> > lock(&nft_net->commit_mutex);

Would it work to use a separated mutex for reset itself?

> > CPU0 is waiting for CPU2 to release transaction mutex.
> > CPU1 is waiting for CPU0 to release the netlink dump mutex
> > CPU2 is waiting for CPU1 to release the ipset subsys mutex
> > 
> > This bug was added when 'nft reset' started to grab the transaction
> > mutex from the dump callback path in nf_tables.
> > 
> > Not yet sure how to avoid it.
> > Maybe we could get rid of 'lock(nfnl_subsys_ipset);'
> > from the xt_set module call paths.
> > 
> > Or add a new lock (spinlock?) to protect the 'reset' object info
> > instead of using the transaction mutex.
> > 
> > I haven't given it much thought yet and will likely not
> > investigate further for the next two weeks.
> 
>