component/userhosts: use github.com/figiel/hosts for compatibility with recent libc

In ubuntu 21.04 (libc6 2.33-0ubuntu5) or current debian testing
(libc6 2.32-5) no longer use open to open /etc/hosts, but what appears
as "openat" in strace output - but can not be replaced by defining an
openat function.

This uses https://github.com/figiel/hosts which uses another approach
of replacing getaddrinfo, gethostbyname* and inet_aton.

Users have been updated a bit, because there are some small differences:
 - the /etc/hosts replacement file is defined by HOSTS_FILE environment
   variable, not HOSTS
 - the library name is libuserhosts.so, not userhosts.so

Other notable differences, for which we did not need code change are:
 - the new library also try to load a file when HOSTS_FILE is not set
 - the new library still use original /etc/hosts file
 - the new library supports aliases to hostnames, not only ip addresses

component/userhosts/buildout.cfg

 [userhosts]
 recipe = slapos.recipe.cmmi
 shared = true
-url = https://lab.nexedi.com/nexedi/userhosts/repository/${:revision}/archive.tar.gz
-revision = a05fe5a3a5cb7005351ef4ec41460089f3ce4d0a
-md5sum = 5a80b4d962d975f290a60cf790c3334d
+url = https://github.com/figiel/hosts/archive/${:revision}.tar.gz
+revision = fdb45fe219593d63453f4be55cfc2a1199c18f59
+md5sum = 9ff5407132a2a67200ef1036a02f0ee2
 configure-command = true
 make-options = PREFIX=@@LOCATION@@
-make-targets = check install
+make-targets = test install

software/slapos-testing/buildout.hash.cfg

@@ -15,4 +15,4 @@
 
 [template]
 filename = instance.cfg
-md5sum = 597c29546519aabe7259e416d0b92095
+md5sum = ba6c572678687081ce111d3805be3cee

software/slapos-testing/instance.cfg

@@ -35,10 +35,10 @@ repository = ${caucase-repository:location}
 [caucase-test-runner]
 recipe = slapos.recipe.template:jinja2
 template = inline:#!/bin/sh
-  export HOSTS="$(mktemp)"
-  trap 'rm "$HOSTS"' EXIT
-  printf '%s testhost\n%s testhost\n' "$SLAPOS_TEST_IPV4" "$SLAPOS_TEST_IPV6" > "$HOSTS"
-  export CAUCASE_NETLOC=testhost:8000 LD_PRELOAD=${userhosts:location}/lib/userhosts.so:$LD_PRELOAD
+  export HOSTS_FILE="$(mktemp)"
+  trap 'rm "$HOSTS_FILE"' EXIT
+  printf '%s testhost\n%s testhost\n' "$SLAPOS_TEST_IPV4" "$SLAPOS_TEST_IPV6" > "$HOSTS_FILE"
+  export CAUCASE_NETLOC=testhost:8000 LD_PRELOAD=${userhosts:location}/lib/libuserhosts.so:$LD_PRELOAD
   exec python -m unittest discover -v
 rendered = $${caucase:location}/host_setting.sh
 

stack/erp5/buildout.hash.cfg

@@ -70,7 +70,7 @@ md5sum = 7a14019abf48ca100eb94d9add20f5ae
 
 [template]
 filename = instance.cfg.in
-md5sum = bbef65b4edeb342f08309604ca3717d5
+md5sum = 0444d646607f8643e987bb62d1cbbc1d
 
 [template-erp5]
 filename = instance-erp5.cfg.in
@@ -86,7 +86,7 @@ md5sum = bc821f9f9696953b10a03ad7b59a1936
 
 [template-zope]
 filename = instance-zope.cfg.in
-md5sum = f3121380ab4d31ba5f4984aec74d0a2f
+md5sum = 09895deebca3206425fada3caf1d92cf
 
 [template-balancer]
 filename = instance-balancer.cfg.in

stack/erp5/instance-zope.cfg.in

@@ -185,7 +185,7 @@ context = key host_dict hosts-parameter:host-dict
 recipe = slapos.recipe.template:jinja2
 mode = 755
 template = inline:#!{{ parameter_dict['dash'] }}/bin/dash
-  export HOSTS=${hosts:rendered}
+  export HOSTS_FILE=${hosts:rendered}
   export LD_PRELOAD={{ parameter_dict['userhosts'] }}:$LD_PRELOAD
   exec "$@"
 rendered = ${directory:bin}/userhosts

stack/erp5/instance.cfg.in

@@ -110,7 +110,7 @@ link-binary = {{ dumps(zope_link_binary) }}
 fonts = {{ dumps(zope_fonts) }}
 fontconfig-includes = {{ dumps(zope_fontconfig_includes) }}
 template-fonts-conf = {{ dumps(template_fonts_conf) }}
-userhosts = {{ userhosts_location }}/lib/userhosts.so
+userhosts = {{ userhosts_location }}/lib/libuserhosts.so
 site-zcml = {{ site_zcml }}
 extra-path-list = {{ dumps(extra_path_list) }}
 matplotlibrc = {{ matplotlibrc_location }}

[Jérome Perrin]

I just noticed now, but unlike the previous userhost, this does not support the scenario on which caucase test relies on, which causes a test to be skipped (see test result):

testHttpNetlocIPv6 (caucase.test.CaucaseTest) ... skipped "Cannot resolve 'testhost' as AF_INET6"

This caucase test uses a host file like this

127.0.0.1 testhost
::1 testhost

but the approach in https://github.com/figiel/hosts/blob/fdb45fe219593d63453f4be55cfc2a1199c18f59/userhosts.c is to hook system calls like

       int getaddrinfo(const char *restrict node,
                       const char *restrict service,
                       const struct addrinfo *restrict hints,
                       struct addrinfo **restrict res);

and, lookup the node from the first matching line in the hosts file:

127.0.0.1 testhost     <--
::1 testhost

and call the original method with the alias.

With this example file, when called with

getaddrinfo("testhost", service, hints, res)

the "real" getaddrinfo will be called with the resolved value:

getaddrinfo("127.0.0.1", service, hints, res)

I don't see how this could easily support names with both IPv4 and IPv6 addresses like this with this approach.

In this caucase test, it's used like this https://lab.nexedi.com/nexedi/caucase/blob/0d663926441f7be6413745f89b1d8756824bbb05/caucase/test.py#L3528-3532 so hints argument will contain AF_INET6 , we could add support for this exact case, but that seems nonsense because we won't be able to fully support a host files like that: with AF_UNSPEC it should return both 127.0.0.1 and ::1, which seems impossible with this approach.

For now, we can say that such a hosts file is not supported by https://github.com/figiel/hosts/ and I don't know how we could make this test run again.

This use case does not seem critically important, we use this only in caucase test and in ERP5. In ERP5 we are only using IPv4 so the problem is only in caucase test, maybe there's a way to change the caucase test to use another approach.

[Vincent Pelletier]

I don't see how this could easily support names with both IPv4 and IPv6 addresses like this with this approach.

+1

This is a bug in userhosts.c, and not immediately trivial to fix.

This use case does not seem critically important, we use this only in caucase test and in ERP5. In ERP5 we are only using IPv4 so the problem is only in caucase test, maybe there's a way to change the caucase test to use another approach.

The plain caucased software release (which I rely on) makes caucased listen on IPv6, so having this test skipped is not ideal.