TableOfContents

Overall Status

TODO List

Besides some missing features and software packages, these are the most pressing items:

=== Turn off irqbalance on SMP hosts ====

• error during shutdown - no problem, just ugly

NSCD Behaviour & Handling is a Total Mess (TM)

• causes trouble under SELinux (avc:denied messages), probably slowing things down
• (re)started by half a dozen features
• use it at all? only for some data (DNS)? depending on what NIS/LDAP is udes for?
• several VAMOS variables steering it?!

=> first think about this, then design, then code, then document !

Booting with init=/bin/sh does not work

• probably due to udev?
• will it help to add /dev/tty0 to /etc/udev?
• alternatively, devise a method for booting rescue systems over the net
  • PXE and/or GRUB entry

Terminal Settings

• settings for root are incompatible with those of normal users on SL3 etc.
• probably because root doesn't get the HEPiX environment any more
  • change back?

Application Software

As discussed in the [:UnixMinutes/Protokoll-2006-04-18:Unix meeting 2006-04-18]:

Java: 1.5 32-bit jpackage is default, 64bit build on amd64 in addition, nothing else without a case

• we'll provide the 32bit jpackage packages on both i386 and amd64
• amd64 systems will get the 64bit 1.5 build in /opt/products in addition
• 1.4.2 will only be provided (in /opt/products) if there is a physics case for it

Browsers: 32bit Firefox is the only supported one

• on both i386 and amd64, system packages as they come with SL4
• the version coming with SL4 (not necessarily the most recent)
• 32bit even on amd64 is the only way to provide the most important plugins:
  • flash: using the "official" RPM
  • java: using the plugin RPM from jpackage (see above)

Mail Client: pine recommended, Thunderbird provided as is

• we provide the pine package coming with SL
  • + DL_pine and DL_sslcerts

• thunderbird is provided as is without support, not recommended

Openoffice: 2.0.x system packages from openoffice.org

• 1.1.2 coming with SL4 is just too old
• no more hope for usable /opt/products package from HH (waiting for a year now...)
• only available as 32bit, who cares
• redhat desktop integration ok

• printing integration through printcap -> psprint.conf converter (student work)

  • done by Bernd, checks/rollout pending

Table of Missing Software

Table of Dropped Software

open questions

SELinux and preserved Partitions: Relabel /usr1?

• /usr1/tmp (the directory and the link) get the right label in %post, but no files in there are touched
• /usr1/scratch?

Disable Sound except on desktops?

Sound is automatically configured during installation. Maybe we should diable it on all classes of machines except desktops? And is it sufficient to comment out the *snd* lines in /etc/modprobe.conf?

Leave cpufreq enabled?

This is the default, and it works fine on some recent hardware.

• but is it completely stable?
• is there a performance penalty (servers, farmnodes)?

is DL_cron_sigpipe still needed?

Check whether this was fixed upstream.

roll out a development version of zsh with partial UTF8 support?

This was proposed in the [:UnixMinutes/Protokoll-2006-04-18:Unix meeting 2006-04-18] by at least two participants. At least two others think this is flirting with desaster.

replace NIS/LDAP with local files?

Why not finally rid ourselves of all NIS/LDAP/nsswicth(compat...) problems?

• create /etc/passwd & /etc/group from a distributed master file (taking into account who gets a shell on the system)

  • this is available now, but needs some more testing and refinement
  • this would also make it easy to change all users' shells to bash
  • impemented by Waltraut and Felix, being tested

This is now available, and works very well. Make it the default?

if yes: how to handle nscd?

• turn it off completely?
• or have different configurations (like: use for DNS only) ?
• need to adapt all features touching it
  • ldap
  • nagios
  • netgroup
  • nsswitch
• have a variable CF_nscd ?
  • values could be (off, on, dns_only, ...), feture would copy right config
  • which feature? nsswitch? or have a new one?

Any more s-bits to remove?

[em64t] /root # find /bin /etc /lib /misc /opt /sbin /selinux /srv /tftpboot /tmp.orig /usr /var -perm +6000 | xargs ls -ldF | awk '{print $1 "  " $3 "  " $4 "    \t" $9}'
-rwsr-xr-x  root  root          /bin/mount*
-rwsr-xr-x  root  root          /bin/ping*
-rwsr-xr-x  root  root          /bin/ping6*
-rwsr-xr-x  root  root          /bin/su*
-rwsr-xr-x  root  root          /bin/umount*
-rwsr-xr-x  root  root          /opt/products/krb5/1.4/bin/ksu*
-rwsr-xr-x  root  root          /opt/products/krb5/1.4/bin/v4rcp*
-rwxr-sr-x  root  root          /sbin/netreport*
-r-s--x--x  root  root          /sbin/pam_timestamp_check*
-r-sr-xr-x  root  root          /sbin/pwdb_chkpwd*
-r-sr-xr-x  root  root          /sbin/unix_chkpwd*
-rwsr-xr-x  root  root          /usr/bin/chage*
-rwsr-xr-x  root  root          /usr/bin/gpasswd*
-rwsr-xr-x  root  root          /usr/bin/kgrantpty*
-rwsr-xr-x  root  root          /usr/bin/kpac_dhcp_helper*
-rwxr-sr-x  root  mail          /usr/bin/lockfile*
-rwsr-xr-x  root  root          /usr/bin/lppasswd*
-rws--x--x  root  root          /usr/bin/newgrp*
-r-s--x--x  root  root          /usr/bin/passwd*
-rwsr-xr-x  root  root          /usr/bin/rcp*
-rwsr-xr-x  root  root          /usr/bin/rlogin*
-rwsr-xr-x  root  root          /usr/bin/rsh*
-rwsr-xr-x  root  root          /usr/bin/sg*
-rwxr-sr-x  root  slocate       /usr/bin/slocate*
-rwxr-sr-x  root  nobody        /usr/bin/ssh-agent*
-r-xr-sr-x  root  tty           /usr/bin/wall*
-rwxr-sr-x  root  tty           /usr/bin/write*
-rwsr-xr-x  root  root          /usr/kerberos/bin/ksu*
-rws--x--x  root  root          /usr/libexec/openssh/ssh-keysign*
-rwsr-xr-x  root  root          /usr/libexec/pt_chown*
-rws--x--x  vcsa  root          /usr/lib/mc/cons.saver*
-rwx--s--x  root  utmp          /usr/lib/vte/gnome-pty-helper*
-rwsr-xr-x  root  root          /usr/sbin/kppp*
-rwxr-sr-x  root  lock          /usr/sbin/lockdev*
-rwxr-sr-x  root  mail          /usr/sbin/mlock*
-rwxr-sr-x  root  postdrop      /usr/sbin/postdrop*
-rwxr-sr-x  root  postdrop      /usr/sbin/postqueue*
-rws--x--x  root  root          /usr/sbin/userhelper*
-rwsr-xr-x  root  root          /usr/sbin/userisdnctl*
-rwsr-xr-x  root  root          /usr/sbin/usernetctl*
-rwxr-sr-x  root  utmp          /usr/sbin/utempter*

Answered Questions

support only bash as the login shell?

• zsh has UTF-8 issues
• can we stop supporting tcsh as a login shell?
  • much effort especially since we have to change the profiles
  • not recommended anyway
• mapping of users' shells in local /etc/passwd is implemented

Decision

It was ruled by the Unix Group that we shall support bash in addition to zsh and tcsh.

PAM & Krb5 & AFS & SSH & Screen Locks

The problem:

• The pam_krb5 coming with SL4 won't create AFS tokens from forwarded Kerberos5 tickets.
• It also cannot handle refreshing credentials when unlocking the screen.
• Our pam_krb5-1.3_rc7 from sourceforge hasn't seen any development in three years. And it segfaults on SL4.

Possible solutions:

1. Get our old pam_krb5 going, and build everything (including kerberos and ssh) ourselves, like on SL3. A lot of work, and not the direction we want to take generally.

2. Use Doug Engert's pam_afs2 module. This is currently implemented. It is a minimally invasive solution since it allows using the pam_krb5 and openssh coming with SL4. Alas, it doesn't solve the screen lock problem.
3. Try staying close to SL4 and the direction RHEL is taking:

   According to [https://lists.openafs.org/pipermail/openafs-info/2006-April/022127.html this] posting to the OpenAFS mailing list, there was some hope that Red Hats pam_krb5 module should do the job. This turned out to be true, alas only for more recent versions of the module than what comes with SL4. But the latest one from FC5 (2.2.6-2.2) works on SL4 and is easy to rebuild (nothing required except rpmbuild --rebuild ...src.rpm). It comes with new options, to be set in the appdefaults/pam part of krb5.conf:

   • external = sshdBR This enables getting a PAG and an AFS token upon gssapi authentication.

   If called correctly by the application, the module will refresh the credentials, which solves the screen lock problem. To make this work for AFS tokens as well, the following option is required:

   • tokens = sshd loginBR This will limit creation of a new PAG to the listed services. In addition, the tokens flag must not be used in /etc/pam.d/system_auth.

   The pam configuration for xscreensaver can now simply stack system_auth! The only caveat is that the xscreensaver coming with SL4 will use krb5 and not pam. Hence this requires

   rebuilding the package, with a trivial change: replace configure --with-pam with configure --with-pam --without-kerberos in the spec. And in order not to pick up a dependency on /opt/products/bin/perl5 add export PATH=/usr/bin:/usr/X11R6/bin:/bin to the beginning of the %build and %install sections.

   Now unlocking the screen works, but only with xscreensaver. The KDE desktop lock does not call pam the right way. Since it also works different than in the past (when we could patch the string identifying the pam servicename in the executable), this probably will not be supported any longer. We recommend using GNOME anyway. The last remaining problem is then that sshd cannot be used in privsep mode. While the new pam module supports an option

   • use_shmem = sshd

   it only works with recent openssh releases. Rebuilding 4.3p2-4 from FC5 on SL4 requires a few minor changes only. From the spec's changelog:

   -  replaced openssh-selinux.patch by openssh-selinux-sl4 .patch 
      (the one from 3.9p1-8.RHEL4.12) to match lower libselinux version on SL4
   -  downgraded libselinux requirements to >= 1.17.9
   -  disabled AUDIT (requires the new selinux patch - although the differences
      are trivial, saved the effort since we don;t need audit support)
   -  changed build requirement from libX11-devel to xorg-x11-devel 
   -  disabled patch26 (pam-no-stack.patch)

   SELinux support in ssh is probably not needed, hence this could even be done much simpler.

Solution 3 is nearly perfect, moderately expensive to support in the long run, and what things will be like on SL5 anyway. It is also more or less what CERN is doing.

Decision

In the [:UnixMinutes/Protokoll-2006-04-18:Unix meeting 2006-04-18], it was ruled by the majority of participants that solution 1 is the way to go. Andreas will carry out the work since he was the only volunteer. Providing a public preview will have to wait until this is done.

Update

The dignified pam_krb5afs module (build statically against heimdal) works. /etc/pam.d/system-auth is using it. xscreensaver and kdesktop_lock work with adapted pam configuration. The SL_fix_kdesktop_lock rpm was reanimated to make kdesktop_lock use a pam configuration file different to /etc/pam.d/kde (named kde-kss). The pam_afs2 pam module is no longer needed.

roadmap

1. other priority1 features (sue) (?profiles?)
2. Milestone: Regular test systems
3. priority2 features
4. Milestone: Ready for special purposes
5. priority3 features
6. Milestone: public preview

little todos

• set CF_no_root_passwd to true for appropriate defaults
• microcode_ctl fails, as the device is missing. From Mandriva's /etc/udev/rules.d/50-mdk.rules
  • 2006-04-13 sw: fixed in 4.3

VAMOS features

Traps & Pitfalls

Files in /etc/cron.d

These must have mode 644. Crond does not accept 755.

SELinux and cfengine

the problem

Any files read by daemons covered by the targeted policy must have the correct type. The file /etc/passwd, for example, is read by nscd:

[root@satyr2 ~]# ls -Z /etc/passwd
-rw-r--r--  root     root     system_u:object_r:etc_t          /etc/passwd

Unfortuantely, cfengine is not aware of security contexts. Hence files it modifies may end up with the wrong context, and daemons may no longer be able to access them:

[root@satyr2 ~]# ls -Z /etc/passwd
-rw-r--r--  root     root     system_u:object_r:etc_runtime_t  /etc/passwd

These actions are known to have this problem:

• editfiles
• copy
• link

Even more unfortunately, many files (but not all) will still have the right context if a feature is run interactively, or even by cron, since in this case they run in the unconfined_t domain. This may change in the future, but so far the problem only affects features run from /etc/init.d/sue_boot The reason is that then processes run in domain initrc_t, with the result that newly created files may get a different context.

Hence to test the effect of a feature, run sue.update like this:

[root@satyr2 ~]# runcon -t initrc_t /products/sue/etc/sue.update ...

the solution(s)

modify the sue_boot init script

One step is obviously to run sue_boot in a different context, by adding this to /etc/init.d/sue_boot

RUNCON=""
if [ -x /usr/sbin/selinuxenabled ]; then
    /usr/sbin/selinuxenabled && RUNCON="/usr/bin/runcon -t unconfined_t --"
fi

and then changing the sue.update command to this:

$RUNCON /products/sue/etc/sue.update -v 2>&1 | cat

This seems to work the way it does normally (which is less than perfect since all output ends up in the sue update logs even if nothing was done), and it indeed solves most problems. Hence this should probably be implemented.

However, if the default type created in a directory is not the desired one, action must be taken in the features. Example:

• /etc/dhcpd.conf should be dhcp_etc_t, not etc_t

  • even if in this case dhcpd is still allowed to read the file, this may change eventually

  • and the file is not protected from other daemons allowed to read files of type etc_t but have no business with dhcp

use restorecon in features where appropriate

The restorecon command looks up the right file context(s) in /etc/selinux/targeted/contexts/files/file_contexts and sets the context accordingly for all files passed as arguments. It only works if SELinux is available in the kernel and enabled. If it is available but disabled, the command will complete successfully but the contexts will not be changed, hence no point in running it.

These are the ingredients for (cfengine2 only) feature to correct the filetypes:

•   control:
  
    actionsequence =
      (
        shellcommands.PASS1
        copy
        editfiles
        shellcommands.PASS2
      )
  
      AddInstallable = ( RESTORECON )

  It will often be necessary to introduce several shellcomands passes, since restorecon should obviously run after the files are modified, but before any daemons that need them are (re)started.

    groups:
  
      HAS_SEL = ( FileExists(/usr/sbin/selinuxenabled) )
      HAS_SEL::
        SELINUX = ( ReturnsZero(/usr/sbin/selinuxenabled) )

  This should work silently and with minimal overhead on any system. If the SELINUX class is set, restorecon should work.

    editfiles:
      { /some/file.1
        ...
        DefineClasses "RESTORECON"
      }
  
    copy:
      /repository/some/file.2
      dest=/some/file.2 type=checksum define=RESTORECON
  
  
    links:
      CLASS1::
        /some/file ->! ./some/file.1  type=relative define=RESTORECON
      CLASS2::
        /some/file ->! ./some/file.2  type=relative define=RESTORECON

  Here we make sure the RESTORECON class gets set if there is something to do for restorecon.

  At least in some cases, this only works if RESTORECON is declared AddInstallable, as shown above.

    shellcommands:
      PASS1::
        "/some/other command ..."
  
      PASS2::
        "/sbin/restorecon /some/file /some/file.1 /some/file.2"

  One could do this more fine grained, but it's probably not worth it.

SELinux and httpd

• local files

  Must be of type httpd_sys_content_t. The sync-a.pl script now does this with chcon. To work on a whole tree:

   chcon -R -h -t httpd_sys_content_t /some/tree

  The -h switch is important if symlinks should be usable through http. Notice the root directory of new filesystems is unlabelled, and chcon will refuse to apply a partial context to it. Hence, after mounting a new filesystem, run something like this:

   chcon system_u:object_r:file_t /new/mounted/filesystem

• AFS By default, httpd is not allowed to read files in AFS. It takes two steps to make this work (and, of course, appropriate httpd configuration in /etc/httpd/conf.d/):

  1. Enable the SELinux booleans httpd_enable_homedirs and use_nfs_home_dirs. This allows httpd to read files of type nfs_t. The latest policy (requires selinux-policy-targeted-1.17.30-2.120 or later) applies a genfs context of nfs_t to anything below /afs. It is easiest to do this with system-config-securitylevel.

  2. Modify the policy.

     After step 1 and with an unmodified policy, it almost works. But access is really slow, and the logs fill with messages

     avc:  denied  { write } for  pid=3980 comm="httpd" lport=7001 scontext=root:system_r:httpd_t tcontext=root:system_r:initrc_t tclass=udp_socket

     Using the audit2allow utility, the following rule was generated:

     # allow AFS callbacks triggered by httpd:
     # This is much too generous, but I haven't found out yet how to limit this to
     # port 7001. Anyway, it's better than running httpd in unconfined_t.
     allow httpd_t initrc_t:udp_socket write;
     # needed if afs is ever restarted:
     allow httpd_t unconfined_t:udp_socket write;

     To get this into a binary policy and load it:

     • install the selinux-policy-targeted-sources rpm

     • save the rule above in the file /etc/selinux/targeted/src/policy/domains/local.te

     • make -C /etc/selinux/targeted/src/policy load

       Afterwards, updating the policy rpm will no longer load the new policy.BRBR To get the new policy with our change, install the new selinux-policy-targeted-sources and rerun the make load step. This happens automatically during an update of selinux-policy-targeted-sources.BRBRDo not play with the policy on production systems. Never install the policy sources on a production system unless you have to modify the binary policy on it and know what you're doing.

Bugs in the distribution we need to fix

• SL 4.2: service dhcpd status buggy, always returns 0

  • bug is present in many other init scripts as well
  • should be fixed with ~TUTILS/chkservice changes as of 2006-05-25