Re: [PATCH net-next v6 04/19] tcp: add datapath logic for PSP with inline key exchange

On 8/14/25 9:18 AM, Paolo Abeni wrote:

On 8/12/25 2:29 AM, Daniel Zahka wrote:

quoted

@@ -2070,7 +2076,9 @@ bool tcp_add_backlog(struct sock *sk, struct sk_buff *skb,
  	     (TCPHDR_ECE | TCPHDR_CWR | TCPHDR_AE)) ||
  	    !tcp_skb_can_collapse_rx(tail, skb) ||
  	    thtail->doff != th->doff ||
-	    memcmp(thtail + 1, th + 1, hdrlen - sizeof(*th)))
+	    memcmp(thtail + 1, th + 1, hdrlen - sizeof(*th)) ||
+	    /* prior to PSP Rx policy check, retain exact PSP metadata */
+	    psp_skb_coalesce_diff(tail, skb))
  		goto no_coalesce;

The TCP stack will try to coalesce skbs in other places, too (i.e.
tcp_try_coalesce(), tcp_collapse()...) Why a similar check is not needed
there?

We handle coalescing of skb's in various places. On the tx path, we 
place a call to tcp_write_collapse_fence() in psp_sock_assoc_set_tx(), 
to prevent data written into the socket as cleartext from being merged 
with data written after the tx-assoc netlink operation has been 
performed. On the rx path, for dealing with coalescing before the skb is 
in the socket receive queue we call psp_skb_coalesce_diff() from 
tcp_add_backlog() and gro_list_prepare(). For skb's that are already on 
the socket receive queue, we rely on the fact that psp skb's will have 
skb->decrypted set, and all tcp functions that try to collapse skb's on 
the receive queue should call skb_cmp_decrypted() at some point. If we 
have missed a case, than that would be a bug.