Re: [PATCH v2 net] net: Fix sk->sk_stamp race in sock_recv_cmsgs().

[Kuniyuki Iwashima <kuniyu@amazon.com>]

From: Eric Dumazet <edumazet@google.com>
Date: Mon, 8 May 2023 19:31:12 +0200
> On Mon, May 8, 2023 at 7:20 PM Kuniyuki Iwashima <kuniyu@amazon.com> wrote:
> >
> > From: Eric Dumazet <edumazet@google.com>
> > Date: Mon, 8 May 2023 19:08:58 +0200
> > > On Mon, May 8, 2023 at 6:58 PM Kuniyuki Iwashima <kuniyu@amazon.com> wrote:
> > > >
> > > > KCSAN found a data race in sock_recv_cmsgs() [0] where the read access
> > > > to sk->sk_stamp needs READ_ONCE().
> > > >
> > > > Also, there is another race like below.  If the torn load of the high
> > > > 32-bits precedes WRITE_ONCE(sk, skb->tstamp) and later the written
> > > > lower 32-bits happens to match with SK_DEFAULT_STAMP, the final result
> > > > of sk->sk_stamp could be 0.
> > > >
> > > >   sock_recv_cmsgs()  ioctl(SIOCGSTAMP)      sock_recv_cmsgs()
> > > >   |                  |                      |
> > > >   |- if (sock_flag(sk, SOCK_TIMESTAMP))     |
> > > >   |                  |                      |
> > > >   |                  `- sock_set_flag(sk, SOCK_TIMESTAMP)
> > > >   |                                         |
> > > >   |                                          `- if (sock_flag(sk, SOCK_TIMESTAMP))
> > > >   `- if (sk->sk_stamp == SK_DEFAULT_STAMP)      `- sock_write_timestamp(sk, skb->tstamp)
> > > >       `- sock_write_timestamp(sk, 0)
> > > >
> > > > Even with READ_ONCE(), we could get the same result if READ_ONCE() precedes
> > > > WRITE_ONCE() because the SK_DEFAULT_STAMP check and WRITE_ONCE(sk_stamp, 0)
> > > > are not atomic.
> > > >
> > > > Let's avoid the race by cmpxchg() on 64-bits architecture or seqlock on
> > > > 32-bits machines.
> > > >
> > >
> > > I disagree. Please use WRITE_ONCE(), even if we know it is racy on 32bit.
> > >
> > > sock_read_timestamp() and sock_write_timestamp() already are racy, and
> > > we do not care.
> >
> > I think it's not racy since commit 3a0ed3e96197 ("sock: Make sock->sk_stamp
> > thread-safe"), which introduced seqlock in sock_read_timestamp() and
> > sock_write_timestamp().
> 
> Please note I do not care of 32bit.
> 
> It is definitely racy on 64bit.
> 
> Please look at
> 
> commit f75359f3ac855940c5718af10ba089b8977bf339
> Author: Eric Dumazet <edumazet@google.com>
> Date:   Mon Nov 4 21:38:43 2019 -0800
> 
>     net: prevent load/store tearing on sk->sk_stamp
> 
> 
> We can not use cmpxchg() only in one place and not the others.

Ah, I understand.  I'll post v3 with this diff to silence KCSAN.

---8<---
diff --git a/include/net/sock.h b/include/net/sock.h
index 8b7ed7167243..656ea89f60ff 100644
--- a/include/net/sock.h
+++ b/include/net/sock.h
@@ -2718,7 +2718,7 @@ static inline void sock_recv_cmsgs(struct msghdr *msg, struct sock *sk,
 		__sock_recv_cmsgs(msg, sk, skb);
 	else if (unlikely(sock_flag(sk, SOCK_TIMESTAMP)))
 		sock_write_timestamp(sk, skb->tstamp);
-	else if (unlikely(sk->sk_stamp == SK_DEFAULT_STAMP))
+	else if (unlikely(sock_read_timestamp(sk) == SK_DEFAULT_STAMP))
 		sock_write_timestamp(sk, 0);
 }
 
---8<---

Thank you!


> 
> cmpxchg() is expensive, we do not want it here on our fast path.
>
> Thanks.