From: Eric Dumazet <edumazet@xxxxxxxxxx>
Date: Tue, 27 Feb 2024 13:06:07 +0100
> On Tue, Feb 27, 2024 at 2:12 AM Kuniyuki Iwashima <kuniyu@xxxxxxxxxx> wrote:
> >
> > syzkaller reported a warning of netns tracker [0] followed by KASAN
> > splat [1] and another ref tracker warning [1].
> >
> > syzkaller could not find a repro, but in the log, the only suspicious
> > sequence was as follows:
> >
> > 18:26:22 executing program 1:
> > r0 = socket$inet6_mptcp(0xa, 0x1, 0x106)
> > ...
> > connect$inet6(r0, &(0x7f0000000080)={0xa, 0x4001, 0x0, @loopback}, 0x1c) (async)
> >
> > The notable thing here is 0x4001 in connect(), which is RDS_TCP_PORT.
> >
> > So, the scenario would be:
> >
> > 1. unshare(CLONE_NEWNET) creates a per netns tcp listener in
> > rds_tcp_listen_init().
> > 2. syz-executor connect()s to it and creates a reqsk.
> > 3. syz-executor exit()s immediately.
> > 4. netns is dismantled. [0]
> > 5. reqsk timer is fired, and UAF happens while freeing reqsk. [1]
> > 6. listener is freed after RCU grace period. [2]
> >
> > Basically, reqsk assumes that the listener guarantees netns safety
> > until all reqsk timers are expired by holding the listener's refcount.
> > However, this was not the case for kernel sockets.
> >
> > Commit 740ea3c4a0b2 ("tcp: Clean up kernel listener's reqsk in
> > inet_twsk_purge()") fixed this issue only for per-netns ehash, but
> > the issue still exists for the global ehash.
> >
> > We can apply the same fix, but this issue is specific to RDS.
> >
> > Instead of iterating ehash and purging reqsk during netns dismantle,
> > let's hold netns refcount for the kernel listener.
> >
> >
>
> > Reported-by: syzkaller <syzkaller@xxxxxxxxxxxxxxxx>
> > Suggested-by: Eric Dumazet <edumazet@xxxxxxxxxx>
> > Fixes: 467fa15356ac ("RDS-TCP: Support multiple RDS-TCP listen endpoints, one per netns.")
> > Signed-off-by: Kuniyuki Iwashima <kuniyu@xxxxxxxxxx>
> > ---
> > net/rds/tcp_listen.c | 4 ++--
> > 1 file changed, 2 insertions(+), 2 deletions(-)
> >
> > diff --git a/net/rds/tcp_listen.c b/net/rds/tcp_listen.c
> > index 05008ce5c421..2d40e523322c 100644
> > --- a/net/rds/tcp_listen.c
> > +++ b/net/rds/tcp_listen.c
> > @@ -274,8 +274,8 @@ struct socket *rds_tcp_listen_init(struct net *net, bool isv6)
> > int addr_len;
> > int ret;
> >
> > - ret = sock_create_kern(net, isv6 ? PF_INET6 : PF_INET, SOCK_STREAM,
> > - IPPROTO_TCP, &sock);
> > + ret = __sock_create(net, isv6 ? PF_INET6 : PF_INET, SOCK_STREAM,
> > + IPPROTO_TCP, &sock, SOCKET_KERN_NET_REF);
> > if (ret < 0) {
> > rdsdebug("could not create %s listener socket: %d\n",
> > isv6 ? "IPv6" : "IPv4", ret);
>
> If RDS module keeps a listener alive, not attached to a user process,
> netns dismantle will never occur.
>
> I think we have to cleanup SYN_RECV sockets in inet_twsk_purge()
Ah.. yes, __init_net ops hook must not take net ref..
I'll go that way in v3.
>
> Yes, it removes one optimization you did.
>
> Perhaps add a counter of all kernel sockets that were ever attached to
> a netns in order to decide to apply the optimization.
> (keeping a precise count of SYN_RECV would be too expensive)
I'll work on the follow-up for net-next after the right fix is merged.
Thanks!
