Re: [PATCH mptcp-next v8 1/8] mptcp: use atomic64_t for msk->ack_seq

[sashiko-bot@kernel.org]

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:

Pre-existing issues:
- [High] Torn reads on locklessly accessed `u64` variables (`bytes_received`, `bytes_consumed`, `write_seq`, `snd_nxt`) on 32-bit architectures.
--

--- Patch [1]: [PATCH mptcp-next v8 1/8] mptcp: use atomic64_t for msk->ack_seq ---
commit 6ecb7c718880f7f40ce17752ce61e7a9dfad9bea
Author: Geliang Tang <geliang@kernel.org>

mptcp: use atomic64_t for msk->ack_seq

This commit converts msk->ack_seq from u64 to atomic64_t to guarantee
tear-free accesses on all architectures. It also caches the atomic64_read()
value in local variables in __mptcp_move_skb() and __mptcp_ofo_queue()
to avoid inconsistent repeated reads.

> diff --git a/net/mptcp/protocol.c b/net/mptcp/protocol.c
> index f22d64ab1c53b..88bdb65a9e87e 100644
> --- a/net/mptcp/protocol.c
> +++ b/net/mptcp/protocol.c
> @@ -464,7 +466,7 @@ static bool __mptcp_move_skb(struct sock *sk, struct sk_buff *skb)
>  		}
>  
>  		msk->bytes_received += copy_len;
> -		WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
> +		atomic64_add(copy_len, &msk->ack_seq);

[Severity: High]
This isn't a bug introduced by this patch, but since it explicitly sets out
to fix 32-bit torn reads for msk->ack_seq, should the other 64-bit socket
state variables also be converted?

For example, msk->bytes_received is updated here using a plain non-atomic +=
operation, and it is read locklessly in mptcp_data_avail():

net/mptcp/protocol.h:mptcp_data_avail() {
	...
	return READ_ONCE(msk->bytes_received) - READ_ONCE(msk->bytes_consumed);
}

Similarly, write_seq and snd_nxt are read locklessly:

net/mptcp/protocol.h:mptcp_stream_memory_free() {
	...
	notsent_bytes = READ_ONCE(msk->write_seq) - READ_ONCE(msk->snd_nxt);
}

A torn read on these 64-bit variables can produce a massively incorrect
value. When a counter crosses a 4GB boundary (e.g., transitioning from
0x00000000FFFFFFFF to 0x0000000100000000), a concurrent 32-bit read might see
0x00000001FFFFFFFF. In mptcp_poll(), this could lead to incorrectly
calculated data_avail or notsent_bytes, causing spurious EPOLLIN/EPOLLOUT
events or missing events that permanently stall the connection.

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/cover.1787644449.git.tanggeliang@kylinos.cn?part=1