bug#14884: TLS connection not terminated properly

[Ludovic Courtès]

As reported by Mark Weaver and others, fetching from
https://archive.apache.org leads an error:

--8<---------------cut here---------------start------------->8---
$ guix build -S subversion --no-substitutes
The following derivation will be built:
   /nix/store/0qm0bggyhrdhrk1ks8hs2pya5n0ikx57-subversion-1.7.8.tar.bz2.drv
@ build-started 
/nix/store/0qm0bggyhrdhrk1ks8hs2pya5n0ikx57-subversion-1.7.8.tar.bz2.drv - 
x86_64-linux 
/nix/var/log/nix/drvs/0q//m0bggyhrdhrk1ks8hs2pya5n0ikx57-subversion-1.7.8.tar.bz2.drv.bz2
starting download of 
`/nix/store/i35q1vm2sl27sjhs7mx8n2m05056ya9x-subversion-1.7.8.tar.bz2' from 
`https://archive.apache.org/dist/subversion/subversion-1.7.8.tar.bz2'...
https://archive.apache.org/.../subversion-1.7.8.tar.bz2  99.0% of 5882.7 
KiBERROR: Throw to key `gnutls-error' with args `(#<gnutls-error-enum The TLS 
connection was non-properly terminated.> fill_session_record_port_input)'.
failed to download 
"/nix/store/i35q1vm2sl27sjhs7mx8n2m05056ya9x-subversion-1.7.8.tar.bz2" from 
"https://archive.apache.org/dist/subversion/subversion-1.7.8.tar.bz2";
--8<---------------cut here---------------end--------------->8---

We discussed it on IRC some time ago:

<mark_weaver> I just tried, and the wget from guix also works.
<civodul> ok
<mark_weaver> maybe wget is ignoring that particular TLS error, dunno.
* civodul tries  [23:22]
<civodul> i can reproduce it
<mark_weaver> I see something about it on this page:
              http://download.opensuse.org/distribution/12.1/repo/oss/ChangeLog
                                                                        [23:29]
<mark_weaver> For glib-networking update to version 2.29.92, it says "Fixed a
              problem when linking against GNUTLS 3.0, where connections would
              sometimes return the error "The TLS connection was non-properly
              terminated". (bgo#659233)"  [23:30]
<mark_weaver> I'm not sure what bug tracking system that bug number is in.
<civodul> the rationale is discussed at
          http://comments.gmane.org/gmane.comp.encryption.gpg.gnutls.devel/4842
                                                                        [23:32]
<mark_weaver> https://bugzilla.gnome.org/show_bug.cgi?id=659233  [23:33]
<mark_weaver> well, I suppose we could just use plain http for that URL.
                                                                        [23:35]
<civodul> sure :-)  [23:36]
<civodul> though the problem is worth fixing
<mark_weaver> is it a problem on our end, or on the apache archive server?
                                                                        [23:37]
<mark_weaver> given that we will check the SHAsum on the downloaded file, I
              suppose there's no harm in ignoring that error for downloads, in
              any case.  [23:38]
<civodul> yes, that's what i was thinking  [23:39]
<civodul> but it's actually tricky to ignore
<civodul> because we pass a TLS port to the download code
<mark_weaver> here's what glib-networking did, fwiw:
              
https://bug659233.bugzilla-attachments.gnome.org/attachment.cgi?id=196741
                                                                        [23:40]

The problem is that the exception is raised by the TLS session record
port’s fill_input method, so there’s no nice call site to wrap into
‘catch’.

We could catch around the ‘dump-port’ call in (guix build download), but
we’d lose info about how much data has actually been transferred.

So for now, I will just:

  1. use http://archive.apache.org instead of https;
  2. ignore this problem altogether, unless this behavior is found to be
     widespread.

Comments welcome.

Ludo’.