tcp: prevent collapsing skbs across boundary in rtx queue

In the Linux kernel, the following vulnerability has been resolved:

tcp: prevent collapsing skbs across boundary in rtx queue

tcp_write_collapse_fence() sets TCP_SKB_CB(skb)->eor = 1 on
tcp_write_queue_tail(sk) to prevent skbs queued after a switch to
device encryption from being collapsed into earlier skbs.

The fence is a no-op if all earlier data has already been transmitted
when the switch happens: sk->sk_write_queue is empty. The not yet
acknowledged earlier skbs wait in sk->tcp_rtx_queue with eor 0.

On a subsequent retransmit or SACK shift, tcp_retrans_try_collapse() or
tcp_shift_skb_data() can then merge an skb queued after the switch into
one queued before it.

Both users of the fence are affected:

- psp: devices only encrypt skbs with skb->decrypted set. The merged skb
keeps decrypted = 0 from the earlier skb, so merged data sent after
psp_sock_assoc_set_tx() is retransmitted in cleartext.

- tls device offload: the merged skb straddles the start marker set in
tls_set_device_offload(). The software fallback (fill_sg_in() returns
-EINVAL) and the mlx5, nfp and funeth drivers cannot handle such an
skb and drop it. Every retransmit rebuilds the same skb, so the
connection stalls.

Fix this in two places, for defense in depth:

1. Fall back to tcp_rtx_queue_tail(sk) in tcp_write_collapse_fence()
when tcp_write_queue_tail(sk) is NULL.

2. Check !skb_cmp_decrypted(to, from) in tcp_skb_can_collapse(), as
tcp_skb_can_collapse_rx() does on receive. skb_shift(), which both
collapse paths call, already has a DEBUG_NET_WARN_ON_ONCE() for this
condition.