Re: [PATCH v3 net-next 6/9] af_unix: Move SOCK_PASS{CRED,PIDFD,SEC} to struct sock.

From: Willem de Bruijn <willemdebruijn.kernel@gmail.com>
Date: Thu, 15 May 2025 18:07:59 -0400

Kuniyuki Iwashima wrote:

quoted

From: Willem de Bruijn <willemdebruijn.kernel@gmail.com>
Date: Thu, 15 May 2025 14:44:14 -0400

quoted

Kuniyuki Iwashima wrote:

quoted

As explained in the next patch, SO_PASSRIGHTS would have a problem
if we assigned a corresponding bit to socket->flags, so it must be
managed in struct sock.

Mixing socket->flags and sk->sk_flags for similar options will look
confusing, and sk->sk_flags does not have enough space on 32bit system.

Also, as mentioned in commit 16e572626961 ("af_unix: dont send
SCM_CREDENTIALS by default"), SOCK_PASSCRED and SOCK_PASSPID handling
is known to be slow, and managing the flags in struct socket cannot
avoid that for embryo sockets.

Let's move SOCK_PASS{CRED,PIDFD,SEC} to struct sock.

While at it, other SOCK_XXX flags in net.h are grouped as enum.

Note that assign_bit() was atomic, so the writer side is moved down
after lock_sock() in setsockopt(), but the bit is only read once
in sendmsg() and recvmsg(), so lock_sock() is not needed there.

Because the socket lock is already held there?

No, for example, scm_recv_unix() is called without lock_sock(),
but it's okay because reading a single bit is always a matter
of timing, when to snapshot the flag, (unless there is another
dependency or the bit is read more than once).

With this, write happens before/after the if block:

                               <-- write could happen here
  lock_sock()
  if (sk->sk_scm_credentials) {
    do something
  }
  lock_unlock()
                               <-- or here (not related to logic)

but this is same without lock_sock() if the bit is read only
once:

                               <-- write could happen here
  if (sk->sk_scm_credentials) {
    do something               <-- or here (not related to logic)
  }
                               <-- or here (not related to logic)

So for SOCK_PASSXXX bits, lock_sock() prevents data-race
between writers as you pointed out, but it does nothing
for readers.

Essentially you're saying that a single bit read is a natural
word read, so atomic anyway? I.e., yes this is a data race, safe.

Yes, even there is load-tearing on the word read, bit is
never torn and is alwyas atomic.

Will KCSAN report on the race regardless?

Not sure, but given we can't use READ_ONCE() for a bit field,
I guess/hope no ?

All other single bit cases in sk_getsockopt use sk_flags
and sock_flag, so are not a good existing example. But the single
bit reads in do_tcp_getsockopt do the same. So I guess it's fine.
Indeed constant_test_bit does nothing special either.

Sounds good, thanks.

Thanks for checking!