Re: [PATCH net] net/smc: fix general protection fault in __smc_diag_dump

[Wang Liang]

在 2025/4/2 15:20, D. Wythe 写道:
> On Wed, Apr 02, 2025 at 10:37:24AM +0800, Wang Liang wrote:
> > 在 2025/4/1 19:01, Paolo Abeni 写道:
> > > On 3/31/25 10:10 AM, Wang Liang wrote:
> > > > diff --git a/net/smc/af_smc.c b/net/smc/af_smc.c
> > > > index 3e6cb35baf25..454801188514 100644
> > > > --- a/net/smc/af_smc.c
> > > > +++ b/net/smc/af_smc.c
> > > > @@ -371,6 +371,7 @@ void smc_sk_init(struct net *net, struct sock *sk, int protocol)
> > > >   	sk->sk_protocol = protocol;
> > > >   	WRITE_ONCE(sk->sk_sndbuf, 2 * READ_ONCE(net->smc.sysctl_wmem));
> > > >   	WRITE_ONCE(sk->sk_rcvbuf, 2 * READ_ONCE(net->smc.sysctl_rmem));
> > > > +	smc->clcsock = NULL;
> > > >   	INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
> > > >   	INIT_WORK(&smc->connect_work, smc_connect_work);
> > > >   	INIT_DELAYED_WORK(&smc->conn.tx_work, smc_tx_work);
> > > The syzkaller report has a few reproducers, have you tested this? AFAICS
> > > the smc socket is already zeroed on allocation by sk_alloc().
> >
> > Yes, I test it by the C repro:
> > https://syzkaller.appspot.com/text?tag=ReproC&x=13d2dc98580000
> >
> > The C repro is provided by the 2025/02/27 15:16 crash from
> >    https://syzkaller.appspot.com/bug?extid=271fed3ed6f24600c364
> >
> > After apply my patch, the crash no longer happens when running the C repro.
> >
> > "the smc socket is already zeroed on allocation by sk_alloc()", That
> > is right.
> > However, smc->clcsock may be modified indirectly in inet6_create().
> > The process like this:
> >
> >    __sys_socket
> >      __sys_socket_create
> >        sock_create
> >          __sock_create
> >            # pf->create
> >            inet6_create
> >              // init smc->clcsock = 0
> >              sk = sk_alloc()
> >
> >              // set smc->clcsock to invalid address
> >              inet = inet_sk(sk);
> >              inet_assign_bit(IS_ICSK, sk, INET_PROTOSW_ICSK & answer_flags);
> >              inet6_set_bit(MC6_LOOP, sk);
> >              inet6_set_bit(MC6_ALL, sk);
> >
> >              smc_inet_init_sock
> >                smc_sk_init
> >                  // add sk to smc_hash
> >                  smc_hash_sk
> >                    sk_add_node(sk, head);
> >                smc_create_clcsk
> >                  // set smc->clcsock
> >                  sock_create_kern(..., &smc->clcsock);)
> >
> > So initialize smc->clcsock to NULL explicitly in smc_sk_init() can fix
> > this crash scene. If the problem can be reproduced after this patch, I
> > guess it is not the same reason, and fix it by another patch is more
> > appropriate.
> This is actually because the current smc_sock is not an inet_sock,
> leading to two modules simultaneously modifying the same offset in
> memory but interpreting its structure differently. I previously proposed
> embedding an inet(6)_sock at the beginning of smc_sock, but the
> community had some objections...
>
> I'm not sure on the community's current stance on this matter, but if a
> fix is absolutely necessary, my recommendation would still be to embed
> an inet(6)_sock within the smc_sock structure
>
> D.

At present, I think initializing the smc in smc_sk_init() may be the 
most simple and effective method. :P

> > > /P