I remember that I once looke into this for implementing in Qsmtp. IIRC it was not that easy, so I would vote this for 1.09 or so, we have clearer things to do in 1.08.

For reference, here's a copy of my analysis from the qmail list:

I haven't noticed any problems with qmail-smtpd, but I think the patch adds a bug to qmail-remote.

To reproduce the bug:

1. Install a vanilla netqmail into /var/qmail-there

2. Apply the EAI patch, then install into /var/qmail-here

3. In /var/qmail-here/control/smtproutes, set qmail-there as qmail-here's smarthost

4. Run /var/qmail-there/bin/qmail-smtpd, wrapped in recordio

5. Run /var/qmail-here/bin/qmail-inject, sending to postmaster@qmail-there

6. Note that the message stays in qmail-here's queue

7. Inspect qmail-there's qmail-smtpd log

8. Observe that SMTP commands and message body lines are CRLF-terminated, but header lines are bare-LF-terminated

Temporarily wrapping qmail-there's qmail-smtpd with fixcrio then allows the message through.

The problem does not depend on whether the message is ASCII-only or UTF-8.

My analysis: qmail-remote.c:blast() normally munges every LF into CRLF. But when the EAI patch adds a "firstpart", those line endings don't get munged in the same way. The one-liner below munges firstpart's line endings a bit earlier. It appears to fix my bug, and maybe doesn't introduce new ones.

--- qmail-remote.c.orig 2017-07-30 18:24:37.000000000 +0000
+++ qmail-remote.c
@@ -253,6 +253,7 @@ void checkutf8message()
     if (r == 0) break;
     if (r == -1) temp_read();

+    if (ch == '\n' && !stralloc_append(&firstpart,"\r")) temp_nomem();
     if (!stralloc_append(&firstpart,&ch)) temp_nomem();

     if (ch == '\r')

The patch has one more tiny problem. The variable asciihost should be null terminated after stralloc_copys(&asciihost, ascii)

@@ -702,9 +793,17 @@
       relayhost[i] = 0;
     }
     if (!stralloc_copys(&host,relayhost)) temp_nomem();
+  } else {
+    char * ascii = 0;
+    host.s[host.len] = '\0';
+    switch (idn2_lookup_u8(host.s, (uint8_t**)&ascii, IDN2_NFC_INPUT)) {
+      case IDN2_OK: break;
+      case IDN2_MALLOC: temp_nomem();
+      default: perm_dns();
+    }
+    if (!stralloc_copys(&asciihost, ascii)) temp_nomem();
   }
 
-
   addrmangle(&sender,argv[2],&flagalias,0);
  
   if (!saa_readyplus(&reciplist,0)) temp_nomem();
@@ -723,7 +822,7 @@
 
  
   random = now() + (getpid() << 16);
-  switch (relayhost ? dns_ip(&ip,&host) : dns_mxip(&ip,&host,random)) {
+  switch (relayhost ? dns_ip(&ip,&host) : dns_mxip(&ip,&asciihost,random)) {

This is EH's version. https://www.fehcom.de/sqmail/doxygen/qmail-remote_8c_source.html

I ended up testing both versions.

1. Arnt's version has two bugs. One which you mentioned above and another one regarding null termination
2. EH's version has a smaller parser but fails if the Received header continuation line starts with tabs or starts with more than two white spaces. I have fixed EH's version by skipping extra white spaces.

int utf8read()
{
  int r;
  int i;
  int received = 0;
  char ch;

  /* we consider only our own last written header */

  for (;;) {
    r = substdio_get(subfdin,&ch,1);
    if (r == 0) break;
    if (r == -1) temp_read();

    if (ch == '\n') {
     if (!stralloc_append(&header,"\r")) temp_nomem(); /* received.c does not add '\r' */
     if (!stralloc_append(&header,"\n")) temp_nomem();  
     if (case_startb(receivedline.s,5,"Date:")) return 0;         /* header to quit asap */
     if (case_startb(receivedline.s,14,"Received: from")) received++; /* found Received header */
     if (received) {
      for (i = 0;i < receivedline.len; i++)
        if (*(receivedline.s+i) != ' ' && *(receivedline.s+i) != '\t')
         break;
      if (case_startb(receivedline.s + i,3,"by ")) {
        for (i += 3; i < receivedline.len; ++i)  
         if (*(receivedline.s+i) == ' ' || *(receivedline.s+i) == '\t')
          if (case_startb(receivedline.s+i+1,9,"with UTF8")) return 1;
        return 0;
      }
     }
     if (!stralloc_copys(&receivedline,"")) temp_nomem();
    } else {
     if (!stralloc_append(&header,&ch)) temp_nomem();
     if (!stralloc_catb(&receivedline,&ch,1)) temp_nomem();
    }
  }
  return 0;
}

Awesome! At a glance, that function looks like it might lend itself to being unit tested. Given the short but already buggy history of SMTPUTF8 in qmail, would you be willing to write some tests for common and edge cases?

That function is also a bit inefficient at a glance. For me it looks like if it is called multiple times, so I would say it should store it's result in a static variable and return early if that already has a value.

The function is called just once, before the blast() function, and exits the moment it sees the UTF8 in the received header or the Date: header, whichever comes first.

You are right, it's utf8received() that has this problem.