Re: [PATCH v5 mptcp-next 10/10] mptcp: leverage the backlog for RX packet processing

[Paolo Abeni <pabeni@redhat.com>]

On 10/21/25 1:32 AM, Mat Martineau wrote:
> On Mon, 6 Oct 2025, Paolo Abeni wrote:
>> diff --git a/net/mptcp/protocol.c b/net/mptcp/protocol.c
>> index 2d5d3da67d1ac..a97a92eccc502 100644
>> --- a/net/mptcp/protocol.c
>> +++ b/net/mptcp/protocol.c
> 
> ...
> 
>> @@ -3509,23 +3519,29 @@ void __mptcp_check_push(struct sock *sk, struct sock *ssk)
>>
>> #define MPTCP_FLAGS_PROCESS_CTX_NEED (BIT(MPTCP_PUSH_PENDING) | \
>> 				      BIT(MPTCP_RETRANSMIT) | \
>> -				      BIT(MPTCP_FLUSH_JOIN_LIST) | \
>> -				      BIT(MPTCP_DEQUEUE))
>> +				      BIT(MPTCP_FLUSH_JOIN_LIST))
>>
>> /* processes deferred events and flush wmem */
>> static void mptcp_release_cb(struct sock *sk)
>> 	__must_hold(&sk->sk_lock.slock)
>> {
>> 	struct mptcp_sock *msk = mptcp_sk(sk);
>> +	u32 delta = 0;
>>
>> 	for (;;) {
>> 		unsigned long flags = (msk->cb_flags & MPTCP_FLAGS_PROCESS_CTX_NEED);
>> -		struct list_head join_list;
>> +		LIST_HEAD(join_list);
>> +		LIST_HEAD(skbs);
>> +
>> +		sk_forward_alloc_add(sk, msk->borrowed_mem);
>> +		msk->borrowed_mem = 0;
>> +
>> +		if (sk_rmem_alloc_get(sk) < sk->sk_rcvbuf)
>> +			list_splice_init(&msk->backlog_list, &skbs);
>>
>> -		if (!flags)
>> +		if (!flags && list_empty(&skbs))
>> 			break;
>>
>> -		INIT_LIST_HEAD(&join_list);
>> 		list_splice_init(&msk->join_list, &join_list);
>>
>> 		/* the following actions acquire the subflow socket lock
>> @@ -3544,7 +3560,8 @@ static void mptcp_release_cb(struct sock *sk)
>> 			__mptcp_push_pending(sk, 0);
>> 		if (flags & BIT(MPTCP_RETRANSMIT))
>> 			__mptcp_retrans(sk);
>> -		if ((flags & BIT(MPTCP_DEQUEUE)) && __mptcp_move_skbs(sk)) {
>> +		if (!list_empty(&skbs) &&
>> +		    __mptcp_move_skbs(sk, &skbs, &delta)) {
>> 			/* notify ack seq update */
>> 			mptcp_cleanup_rbuf(msk, 0);
>> 			sk->sk_data_ready(sk);
>> @@ -3552,7 +3569,9 @@ static void mptcp_release_cb(struct sock *sk)
>>
>> 		cond_resched();
>> 		spin_lock_bh(&sk->sk_lock.slock);
>> +		list_splice(&skbs, &msk->backlog_list);
>> 	}
>> +	WRITE_ONCE(msk->backlog_len, msk->backlog_len - delta);
> 
> Hi Paolo -
> 
> Given the possible multiple calls to __mptcp_move_skbs() and that the 
> spinlock is released/reacquired (and the cond_resched) in the middle, 
> would it make sense to update msk->backlog_len for each iteration of the 
> loop so __mptcp_space() and mptcp_space() don't under-report available 
> space and mptcp_cleanup_rbuf() can make incremental progress?
> 
> I know we don't want to WRITE_ONCE() more than necessary, but it seems 
> like there won't typically be more than one loop iteration. In the cases 
> where it does repeat the loop that means data is arriving quickly and 
> reporting mptcp_space accurately will be important.

That WRITE_ONCE() is intentionally out of the loop, but not as an
optimization, but for a functional goal, similar to:

https://elixir.bootlin.com/linux/v6.17.4/source/net/core/sock.c#L3190

without it, in exceptional situation, the loop could run for an
unbounded amount of time.

Given this is MPTCP-level, and packets went already through the TCP
subflow possibly such scenario is even more unlikely, but I think it's
still possible and serious enough we want to avoid it.

WRT the receive buffer utilization, it should not change much, as either
on the backlog or in the receiver buffer, the skb should be accounted
for the whole truesize.

Cheers,

Paolo