Re: [PATCH] net: tls: fix possible race condition between do_tls_getsockopt_conf() and do_tls_setsockopt_conf()

[Hangyu Hua]

On 25/2/2023 06:17, Jakub Kicinski wrote:
> On Fri, 24 Feb 2023 22:48:57 +0100 Sabrina Dubroca wrote:
> > 2023-02-24, 13:06:25 -0800, Jakub Kicinski wrote:
> > > On Fri, 24 Feb 2023 21:22:43 +0100 Sabrina Dubroca wrote:
> >   [...]
> > > > I suggested a change of locking in do_tls_getsockopt_conf this
> > > > morning [1]. The issue reported last seemed valid, but this patch is not
> > > > at all what I had in mind.
> > > > [1] https://lore.kernel.org/all/Y/ht6gQL+u6fj3dG@hog/
> > >
> > > Ack, I read the messages out of order, sorry.
> > >    
> > > > do_tls_setsockopt_conf fills crypto_info immediately from what
> > > > userspace gives us (and clears it on exit in case of failure), which
> > > > getsockopt could see since it's not locking the socket when it checks
> > > > TLS_CRYPTO_INFO_READY. So getsockopt would progress up to the point it
> > > > finally locks the socket, but if setsockopt failed, we could have
> > > > cleared TLS_CRYPTO_INFO_READY and freed iv/rec_seq.
> > >
> > > Makes sense. We should just take the socket lock around all of
> > > do_tls_getsockopt(), then?
> >
> > That would make things simple and consistent. My idea was just taking
> > the existing lock_sock in do_tls_getsockopt_conf out of the switch and
> > put it just above TLS_CRYPTO_INFO_READY.

I know what you mean. I just think lock crypto_info can fix this simply.

The original situation is:

thread1				thread2(do_tls_getsockopt_conf)

lock_sock(sk)
do_tls_setsockopt_conf(crypto_info->cipher_type set)

				crypto_info = xxx
				cctx = &ctx->tx
				if(!TLS_CRYPTO_INFO_READY(crypto_info))
				
tls_set_device_offload(kmalloc cctx->iv)
tls_set_sw_offload(fail and cctx->iv may not set to NULL)
do_tls_setsockopt_conf(set crypto_info->cipher_type to NULL)
release_sock(sk)

				lock_sock(sk)
				memcpy(xxx, cctx->iv, xxx)
				release_sock(sk)

If we lock crypto_info:

thread1				thread2(do_tls_getsockopt_conf)

lock_sock(sk)
do_tls_setsockopt_conf(crypto_info->cipher_type set)			
tls_set_device_offload(kmalloc cctx->iv)
tls_set_sw_offload(fail and cctx->iv may not set to NULL)
do_tls_setsockopt_conf(set crypto_info->cipher_type to NULL)
release_sock(sk)

				lock_sock(sk)
				crypto_info = xxx
				cctx = &ctx->tx
				release_sock(sk)
				if(!TLS_CRYPTO_INFO_READY(crypto_info))
				lock_sock(sk)
				memcpy(xxx, cctx->iv, xxx)
				release_sock(sk)

> > While we're at it, should we move the
> >
> >      ctx->prot_info.version != TLS_1_3_VERSION
> >
> > check in do_tls_setsockopt_no_pad under lock_sock?
>
> Yes, or READ_ONCE(), same for do_tls_getsockopt_tx_zc() and its access
> on ctx->zerocopy_sendfile.
>
> >   I don't think that
> > can do anything wrong (we'd have to get past this check just before a
> > failing setsockopt clears crypto_info, and even then we're just
> > reading a bit from the context), it just looks a bit strange. Or just
> > lock the socket around all of do_tls_setsockopt_no_pad, like the other
> > options we have.
>
> The delayed locking feels like a premature optimization, we'll keep
> having such issues with new options. Hence my vote to lock all of
> do_tls_getsockopt().

In order to reduce ambiguity, I think it may be a good idea only to
lock do_tls_getsockopt_conf() like we did in do_tls_setsockopt()

It will look like:

static int do_tls_getsockopt(struct sock *sk, int optname,
			     char __user *optval, int __user *optlen)
{
	int rc = 0;

	switch (optname) {
	case TLS_TX:
	case TLS_RX:
+		lock_sock(sk);
		rc = do_tls_getsockopt_conf(sk, optval, optlen,
					    optname == TLS_TX);
+		release_sock(sk);
		break;
	case TLS_TX_ZEROCOPY_RO:
		rc = do_tls_getsockopt_tx_zc(sk, optval, optlen);
		break;
	case TLS_RX_EXPECT_NO_PAD:
		rc = do_tls_getsockopt_no_pad(sk, optval, optlen);
		break;
	default:
		rc = -ENOPROTOOPT;
		break;
	}
	return rc;
}

Of cause, I will clean the lock in do_tls_getsockopt_conf(). What do you
guys think?