BUG: "notmuch insert" fails with "Delivery of non-mail file"

[Alvaro Herrera]

Hi David, thanks for replying.

On 2019-Jan-19, David Bremner wrote:

> Alvaro Herrera <alvherre at alvh.no-ip.org> writes:
> 
> > In my read of the code ultimately comes from
> > g_mime_parser_construct_message rejecting the message.
> > I reported this to GMime, and they said that the problem is that notmuch
> > insert is using the mbox mode:
> > https://github.com/jstedfast/gmime/issues/58
> > (Sample email is attached there).
> 
> This issue (or a related one) has come up before
> 
>      https://nmbug.notmuchmail.org/nmweb/search/postfix+mbox
> 
> Generally it seems to be caused by tools that add mbox 'From ' headers,
> without actually mbox escaping the file. We haven't yet reached
> consensus on a good solution (generally people just want to fix their
> own mail, which is understandable). A workaround discussed in the
> messages I reference above is to strip the 'From ' header before passing
> to notmuch-insert. Perhaps some scholar of the RFCs can convince us that
> that is "always" the right thing for notmuch insert to do.

I'm not sure I follow.  As I understand, notmuch does not work with
mboxes, only with maildirs, so the behavior of splitting emails at "From
" is not strictly necessary, since one file always equals one message.

As for RFC scholarship, I spent some time looking at
https://tools.ietf.org/html/rfc5322 to see if it defined any sort of
message separator ... but as far as I can tell, it only defines what
does a valid message looks like.  It doesn't say where does one message
end.

On the other hand, in my world, it's been quite a while since 'From '
was considered a useful message separator.  This stopped being true in a
pretty extensive way when git-format-patches messages started being
posted as attachments.  But even before that, MUAs stopped adding the
">" at the start of a "From " line in human-written text.  Nowadays what
really governs the split is the Content-Length header, from the MIME
definitions.  Most tools do not escape lines starting with 'From '
anymore.  As far as I can tell, this is defined by RFC-2049,
https://tools.ietf.org/html/rfc2046#section-5.1.1 which states that the
implementation must look for the "boundary delimitir line".  Stopping at
a "From " line before finding the boundary delimiter line would be a
mistake, in my reading.

> > As far as I can tell, this is all coming from
> > _notmuch_message_file_parse() which sets the is_mbox flag when it sees
> > the "^From " line at the start of the file ... which kinda makes sense
> > in general terms, but for notmuch-insert I think that's the wrong thing
> > to do.  Maybe a solution is to pass a flag down from notmuch-insert.c's
> > add_file all the way down to _notmuch_message_file_parse telling it not
> > to treat the file as an mbox.
> 
> I'd be worried about letting notmuch-insert deliver messages that
> notmuch-new would not be able to parse. In particular we'd like to keep
> the property that a Maildir + the output of notmuch-dump should be
> enough to completely recover the notmuch database.

Hmm, that's a good point -- I assume that notmuch-new should be patched
similarly so that those messages are valid there too.

So maybe the solution (given that, as I said above, Notmuch does not
appear to handle mboxes at all) is to just set the mbox flag to false
completely ...

-- 
Álvaro Herrera                PostgreSQL Expert, https://www.2ndQuadrant.com/