Re: [PATCH mptcp-next v11 3/4] mptcp: support MSG_ERRQUEUE on the parent socket

On 5/31/26 4:59 PM, David Carlier wrote:
> Splice pending err skbs from each subflow's error queue onto the parent
> msk's error queue at error-report time, so poll() and recvmsg(MSG_ERRQUEUE)
> on the parent socket observe TX timestamps and MSG_ZEROCOPY completion
> notifications through the standard inet ABI.
> 
> The splice filters by SO_EE_ORIGIN: TIMESTAMPING / ZEROCOPY / LOCAL
> events forward to the parent because they are tied to user-handed data,
> not to a specific path; subflow-level ICMP errors are dropped because
> the legacy RECVERR ABI cannot meaningfully convey their per-subflow peer
> identity to single-path-aware userspace. Such events will be carried by
> a future MPTCP_RECERR channel.
> 
> Forwarded events all go through sock_queue_err_skb(), which re-homes
> skb->sk onto the parent and charges sk_rmem_alloc, so the parent's error
> queue stays bounded by sk_rcvbuf and is dropped under rmem pressure
> (sk_rmem_alloc + truesize >= sk_rcvbuf), matching tcp's sk_rcvbuf-gated
> tx-timestamp path and ip_icmp_error() / ipv6_icmp_error(). MPTCP itself
> never originates MSG_ZEROCOPY or OPT_ID tx-timestamp completions -- its
> data path copies into msk-owned pages and bypasses tcp_sendmsg_locked()
> -- so no subflow-relative ee_data sequence is ever forwarded to the
> parent. The MSG_ERRQUEUE branch of mptcp_recvmsg() forwards to
> inet_recv_error() directly, and poll() advertises EPOLLERR purely on the
> parent's sk_err / sk_error_queue, matching tcp_poll().
> 
> Suggested-by: Paolo Abeni <[email protected]>

Did I? I honestly can't recall anymore?!? (same for the previous patches).

This patch should possibly be applied before 1 && 2, to avoid sashiko
noise there. More relevantly, generally speaking its' better to
implement first the core infrastructure and only later expose it to the
user-space.

> Signed-off-by: David Carlier <[email protected]>
> ---
>  net/mptcp/protocol.c | 63 +++++++++++++++++++++++++++++++++++++-------
>  1 file changed, 54 insertions(+), 9 deletions(-)
> 
> diff --git a/net/mptcp/protocol.c b/net/mptcp/protocol.c
> index f1d74d4b28cf..42a355311c81 100644
> --- a/net/mptcp/protocol.c
> +++ b/net/mptcp/protocol.c
> @@ -11,6 +11,7 @@
>  #include <linux/netdevice.h>
>  #include <linux/sched/signal.h>
>  #include <linux/atomic.h>
> +#include <linux/errqueue.h>
>  #include <net/aligned_data.h>
>  #include <net/rps.h>
>  #include <net/sock.h>
> @@ -894,21 +895,61 @@ static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
>  	return moved;
>  }
>  
> +static bool mptcp_errqueue_skb_forwardable(const struct sk_buff *skb)
> +{
> +	u8 origin = SKB_EXT_ERR(skb)->ee.ee_origin;
> +
> +	return origin == SO_EE_ORIGIN_LOCAL ||
> +		origin == SO_EE_ORIGIN_ZEROCOPY ||
> +		origin == SO_EE_ORIGIN_LOCAL;

Since SO_EE_ORIGIN_LOCAL and SO_EE_ORIGIN_ZEROCOPY can't yet be
generated, why including them here? If there are plans to support such
origins later, this nice helper could be updated at due time.

I *guess* that reducing its' scope...

> +}
> +
> +static bool __mptcp_subflow_splice_errqueue(struct sock *sk, struct sock *ssk)
> +{
> +	struct sk_buff *skb;
> +	bool moved = false;
> +
> +	while ((skb = skb_dequeue(&ssk->sk_error_queue))) {
> +		if (!mptcp_errqueue_skb_forwardable(skb)) {
> +			kfree_skb(skb);  /* path-specific (ICMP) — belongs in MPTCP_RECERR */
> +			continue;
> +		}
> +		/* sock_queue_err_skb() re-homes skb->sk onto the parent and
> +		 * charges its sk_rmem_alloc, so the error queue stays bounded by
> +		 * sk_rcvbuf; drop on overflow, matching tcp's tx-timestamp path.
> +		 * MPTCP never originates MSG_ZEROCOPY or OPT_ID tx-timestamp
> +		 * completions (the data path copies and bypasses
> +		 * tcp_sendmsg_locked()), so no subflow-relative ee_data sequence
> +		 * is ever forwarded.

... will allow dropping this rater verbose comment, which looks like
it's targeting mostly sashiko.

/P