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 19:08:58 +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().