source: trunk/server/doc/ @ 2156

Last change on this file since 2156 was 2113, checked in by ezyang, 13 years ago
Fix clobbered hunk.
File size: 19.7 KB
1# This document is a how-to for installing a Fedora server.
2# It is semi-vaguely in the form of a shell script, but is not really
3# runnable as it stands.
5# Notation
6# [PRODUCTION] Production server that will be put into the pool
7# [WIZARD]     Semi-production server that will only have
8#              daemon.scripts-security-upd bits, among other
9#              restricted permissions
10# [TESTSERVER] Completely untrusted server
12# 'branch' is the current svn branch you are on.  You want to
13# use trunk if your just installing a new server, and branches/fcXX-dev
14# if your preparing a server on a new Fedora release.
17# 'server' is the public hostname of your server, for SCP'ing files
18# to and from.
21# ----------------------------->8--------------------------------------
22#                       FIRST TIME INSTRUCTIONS
24# [PRODUCTION] If this is the first time you've installed this hostname,
25# you will need to update a bunch of files to add support for it. These
26# include:
27#   o Adding all aliases to /etc/httpd/conf.d/scripts-vhost-names.conf
28#     (usually this is hostname,, h-n,,
29#     scriptsN,, and the IP address.)
30#   o Adding routing rules for the static IP in
31#     /etc/sysconfig/network-scripts/route-eth1
32#   o Adding the IP address to the hosts file (same hosts as for
33#     scripts-vhost-names)
34#   o Update SSH config at
35#       - server/fedora/config/etc/ssh/shosts.equiv
36#       - server/fedora/config/etc/ssh/ssh_known_hosts
37#       - server/fedora/config/etc/ssh/sshd_config : DenyUsers
38#     (the last part is critical to ensure that rooting one server
39#     doesn't give you root to all the other servers)
40#   o Put the hostname information in LDAP so SVN and Git work
41#   o Set up Nagios monitoring on sipb-noc for the host
42#   o Set up the host as in the pool on r-b/r-b /etc/heartbeat/
43#   o Update locker/etc/known_hosts
44#   o Update website files:
45#       /mit/scripts/web_scripts/home/server.css.cgi
46#       /mit/scripts/web_scripts/heartbeat/heartbeat.php
48# You will also need to prepare the keytabs for credit-card.  In particular,
49# use ktutil to combine the host/ and
50# host/ keys with host/ in
51# the keytab.  Do not use 'k5srvutil change' on the combined keytab
52# or you'll break the other servers. (real servers only).  Be
53# careful about writing out the keytab: if you write it to an
54# existing file the keys will just get appended.  The correct
55# credential list should look like:
56#   ktutil:  l
57#   slot KVNO Principal
58#   ---- ---- ---------------------------------------------------------------------
59#      1    5 host/
60#      2    3 host/
61#      3    2 host/
62#      4    8 host/
64# The LDAP keytab should be by itself, so be sure to delete it and
65# put it in its own file.
67# ----------------------------->8--------------------------------------
68#                      INFINITE INSTALLATION
70# Start with a Scripts kickstarted install of Fedora (install-fedora)
72# IMPORTANT: If you are installing a server without the benefit of
73# Kickstart (for example, you are installing on XVM, it is VITALLY
74# IMPORTANT that you go through the kickstart and apply all of the
75# necessary changes--for example, disabling selinux or enabling
76# network.)
77#   XXX We should make Kickstart work for test servers too
79# Take updates, reboot if there's a kernel update.
80    yum update -y
82# Get rid of network manager (XXX figure out to make kickstarter do
83# this for us)
84    yum remove NetworkManager
86# Make sure sendmail isn't installed, replace it with postfix
87    yum shell <<EOF
88remove sendmail
89install postfix
94# Check out the scripts /etc configuration
95    cd /root
96    \cp -a etc /
97    chmod 0440 /etc/sudoers
99# [TEST] You'll need to fix some config now.  See bottom of document.
101# Make sure network is working.  Kickstart should have
102# configured eth0 and eth1 correctly; use service network restart
103# to add the new routes from etc in route-eth1.
104    systemctl restart network.service
105    # Check everything worked:
106    route
107    ifconfig
108    cat /etc/hosts
109    cat /etc/sysconfig/network-scripts/route-eth1
111# This is the point at which you should start updating scriptsified
112# packages for a new Fedora release.  Consult 'upgrade-tips' for more
113# information.
114    yum install -y scripts-base
115    # Some of these packages are naughty and clobber some of our files
116    cd /etc
117    svn revert resolv.conf hosts sysconfig/openafs nsswitch.conf
118    # Troubleshooting: if accountadm, tokensys and nscd fail to install
119    # you probably forgot to turn off selinux
121# Replace rsyslog with syslog-ng by doing:
122    yum shell <<EOF
123remove rsyslog
124install syslog-ng
128    systemctl enable syslog-ng.service
130# Install the full list of RPMs that users expect to be on the
131# servers.
132rpm -qa --queryformat "%{Name}.%{Arch}\n" | sort > packages.txt
133# arrange for packages.txt to be passed to the server, then run:
134    cd /tmp
135    yumdownloader --disablerepo=scripts ghc-cgi ghc-cgi-devel
136    yum localinstall ghc-cgi*.x86_64.rpm
137    yum install -y $(cat packages.txt)
138# The reason this works is that ghc-cgi is marked as installonlypkgs
139# in yum.conf, telling yum to install them side-by-side rather than
140# updating them. If it doesn't work, use --skip-broken on the yum
141# command line.
143# Check which packages are installed on your new server that are not
144# in the snapshot, and remove ones that aren't needed for some reason
145# on the new machine.  Otherwise, aside from bloat, you may end up
146# with undesirable things for security, like sendmail.
147    rpm -qa --queryformat "%{Name}.%{Arch}\n" | grep -v kernel | sort > newpackages.txt
148    diff -u packages.txt newpackages.txt | grep -v kernel | less
149    # here's a cute script that removes all extra packages
150    yum erase -y $(grep -Fxvf packages.txt newpackages.txt)
151    # 20101208 - Mysteriously we manage to get these extra packages
152    # from kickstart: mcelog mobile-broadband-provider-info
153    # ModemManager PackageKit
155# ----------------------------->8--------------------------------------
156#                      SPHEROID SHENANIGANS
158# Install the Python eggs and Ruby gems and PEAR/PECL doohickeys that are on
159# the other servers and do not have RPMs.
160# The general mode of operation will be to run the "list" command
161# on both servers, see what the differences are, check if those diffs
162# are packaged up as rpms, and install them (rpm if possible, native otherwise)
164# Note: Since ultimately we'd like to move away from using per-language
165# package manager and all of these be RPMs, it is of questionable
166# importance how much /good/ automation for these is necessary.
168# Warning: For a new release, we're supposed to check if Fedora has
169# packaged up the RPM.  Unfortunately we don't really have good incants
170# for this.
172# Warning: If you're installing a new server mid-lifecycle (or even if
173# this is the start of a cycle, but you've been staggering the
174# installation of servers), upstream may have moved on.  Because we
175# don't normally upgrade spheroid projects, that means executing these
176# instructions directly means that you will have mismatched versions
177# (the new servers will have newer versions.)  Please follow the
178# UPGRADE commentary attached to each of these.
180# Warning: The package lists that are generated are inconsistent on
181# the question of whether or not they contain all packages (locally
182# installed as well as distro packaged), or if they just contain locally
183# installed packages.  Check this carefully; many of the install incants
184# filter out already installed packages.
187# ---------
189# Install the full list of perl modules that users expect to be on the
190# servers.
191    cd /root
192    export PERL_MM_USE_DEFAULT=1
193    cpan # this is interactive, enter the next two lines
194        o conf prerequisites_policy follow
195        o conf commit
196# on a reference server
197perldoc -u perllocal | grep head2 | cut -f 3 -d '<' | cut -f 1 -d '|' | sort -u | perl -ne 'chomp; print "notest install $_\n" if system("rpm -q --whatprovides \"perl($_)\" >/dev/null 2>/dev/null")' > perl-packages.txt
198# arrange for perl-packages.txt to be transferred to server
199    # Package list only contains new packages
200    cat perl-packages.txt | perl -MCPAN -e shell
201# These are in /usr/local
203# UPGRADE: Installing old versions of CPAN modules requires you to
204# specify the full path of a module, e.g.
205# M/MS/MSCHWERN/Test-Simple-0.62.tar.gz.  It is not currently clear how
206# to get this information programatically.  Furthermore, we have a lot
207# of CPAN managed modules.  Since CPAN is the only thing
208# placed in /usr/local at this point, it may be easier to simple tar and
209# cp the Perl modules from one server to another, to keep them
210# consistent.  But doing this is fiddly XXX
213# -----------
215# - Look at /usr/lib/python2.7/site-packages and
216#           /usr/lib64/python2.7/site-packages for Python eggs and modules.
217#   There will be a lot of gunk that was installed from packages;
218#   easy-install.pth in /usr/lib/ will tell you what was easy_installed.
219#   First use 'yum search' to see if the relevant package is now available
220#   as an RPM, and install that if it is.  If not, then use easy_install.
221#   Pass -Z to easy_install to install them unzipped, as some zipped eggs
222#   want to be able to write to ~/.python-eggs.  (Also makes sourcediving
223#   easier.)
224# 'easy_install AuthKit jsonlib2 pygit'
225cat /usr/lib/python2.7/site-packages/easy-install.pth | grep "^./" | cut -c3- | cut -f1 -d- > egg.txt
226    # Package list only contains new packages
227    cat egg.txt | xargs easy_install -Z
228# These are in /usr
230# UPGRADE: Use 'easy_install -n' to see what new versions are installed, and if there
231# are updates validate them and upgrade them on the old servers.  Since
232# we have a really small package list (around 4) checking these manually
233# should be fine.  Note that dry run is slightly buggy and may fail
234# midway processing files on account of a missing build directory.
237# ---------
239# - Look at `gem list` for Ruby gems.
240#   Again, use 'yum search' and prefer RPMs, but failing that, 'gem install'.
241#       ezyang: rspec-rails depends on rspec, and will override the Yum
242#       package, so... don't use that RPM yet
243# XXX This doesn't do the right thing for old version gems
244gem list --no-version > gem.txt
245    # Package list contains distro gems too
246    gem install $(gem list --no-version | grep -Fxvf - gem.txt)
247    # Also, we need to install the old rails version
248    gem install -v=2.3.5 rails
249# These are in /usr
251# UPGRADE:  You can either upgrade out-of-date gems, or leave them at
252# the old version.  We recommend the latter (see below for the
253# rationale), but note that the install script described here doesn't
254# pin against version, so you'll need to supply the -v parameters
255# manually (the gems we install manually don't move too quickly, so this
256# is fairly tractable if you check 'gem outdated'.)
258# If you want to upgrade, do NOT use wildcard 'gem update'; use 'gem
259# outdated' to find out all gems that are out of date, and verify this
260# against our locally installed gems (there will be a lot of out of date
261# gems, but this is simply because Fedora packaging lags behind the
262# canonical versions (this is a good thing).  Manually upgrade just
263# those gems.  Note that this doesn't save you from having to install
264# old gems on the servers that are being installed out-of-cycle,
265# because Ruby supports pinning against old versions, and if those gems
266# then mysteriously disappear, things will be sad (note that this isn't
267# a *huge* problem, because usually when you pin gems it's in
268# conjunction with rvm, so they have their local copy of the gem.)
271# --------
273# - Look at `pear list` for Pear fruits (or whatever they're called).
274#   Yet again, 'yum search' for RPMs before resorting to 'pear install'.  Note
275#   that for things in the beta repo, you'll need 'pear install package-beta'.
276#   (you might get complaints about the php_scripts module; ignore them)
277pear list | tail -n +4 | cut -f 1 -d " " > pear.txt
278    # Package list contains distro packages
279    pear config-set preferred_state beta
280    pear channel-update
281    pear install $(pear list | tail -n +4 | cut -f 1 -d " " | grep -Fxvf - pear.txt)
282# These are in /usr
285# --------
287# - Look at `pecl list` for PECL things.  'yum search', and if you must,
288#   'pecl install' needed items. If it doesn't work, try 'pear install
289#   pecl/foo' or 'pecl install foo-beta' or those two combined.
290pecl list | tail -n +4 | cut -f 1 -d " " > pecl.txt
291    # Package list contains distro packages
292    pecl install --nodeps $(pecl list | tail -n +4 | cut -f 1 -d " " | grep -Fxvf - pecl.txt)
293# These are in /usr
295# ----------------------------->8--------------------------------------
296#                       INFINITE CONFIGURATION
298# [PROD] Create fedora-ds user (needed for credit-card)
299useradd -u 103 -r -d /var/lib/dirsrv fedora-ds
301# Run credit-card to clone in credentials and make things runabble
302# NOTE: You may be tempted to run credit-card earlier in the install
303# process in order, for example, to be able to SSH in to the servers
304# with Kerberos.  However, it is better to install the credentials
305# *after* we have run a boatload untrusted code as part of the
306# spheroids objects process.  So don't move this step earlier!
307python push $server
309# This is superseded by credit-card, but only for [PRODUCTION]
310# Don't use credit-card on [WIZARD]: it will put in the wrong creds!
312#   # All types of servers will have an /etc/daemon.keytab file, however,
313#   # different types of server will have different credentials in this
314#   # keytab.
315#   #   [PRODUCTION] daemon.scripts
316#   #   [WIZARD]     daemon.scripts-security-upd
317#   #   [TESTSERVER] daemon.scripts-test
319# Test that zephyr is working
320    systemctl enable zhm.service
321    systemctl start zhm.service
322    echo 'Test!' | zwrite -d -c scripts -i test
324# Check out the scripts /usr/vice/etc configuration
325    cd /root/vice
326    \cp -a etc /usr/vice
327# [TESTSERVER] If you're installing a test server, this needs to be
328# much smaller; the max filesize on XVM is 10GB.  Pick something like
329# 500000. Also, some of the AFS parameters are kind of retarded (and if
330# you're low on disk space, will actually exhaust our inodes).  Edit
331# these parameters in /etc/sysconfig/openafs (I just chopped a zero
332# off of all of our parameters)
333    echo "/afs:/usr/vice/cache:500000" > /usr/vice/etc/cacheinfo
334    vim /etc/sysconfig/openafs
336# [PRODUCTION] Set up replication (see ./install-ldap).
337# You'll need the LDAP keytab for this server: be sure to chown it
338# fedora-ds after you create the fedora-ds user
339    ls -l /etc/dirsrv/keytab
340    cat install-ldap
342# Enable lots of services
343    systemctl enable openafs-client.service
344    systemctl enable dirsrv.service
345    systemctl enable nslcd.service
346    systemctl enable nscd.service
347    systemctl enable postfix.service
348    systemctl enable nrpe.service
349    systemctl enable httpd.service # not for [WIZARD]
351    systemctl start openafs-client.service
352    systemctl start dirsrv.service
353    systemctl start nslcd.service
354    systemctl start nscd.service
355    systemctl start postfix.service
356    systemctl start nrpe.service
357    systemctl start httpd.service # not for [WIZARD]
359# Note about OpenAFS: Check that fs sysname is correct.  You should see,
360# among others, 'amd64_fedoraX_scripts' (vary X) and 'scripts'. If it's
361# not, you probably did a distro upgrade and should update
362# /etc/sysconfig/openafs (XXX this is wrong: figuring out new
363# systemd world order).
364    fs sysname
366# Postfix doesn't actually deliver mail; fix this
367    cd /etc/postfix
368    postmap virtual
370# Munin might not be monitoring packages that were installed after it
371    munin-node-configure --suggest --shell | sh
373# Run fmtutil-sys --all, which does something that makes TeX work.
374# (Note: this errors on XeTeX which is ok.)
375    fmtutil-sys --all
377# Fix etc by making sure none of our config files got overwritten
378    cd /etc
379    svn status -q
380    # Some usual candidates for clobbering include nsswitch.conf,
381    # resolv.conf and sysconfig/openafs
382    # [WIZARD/TEST] Remember that changes you made should not get
383    # reverted!
385# Reboot the machine to restore a consistent state, in case you
386# changed anything. (Note: Starting kdump fails (this is ok))
388# ------------------------------->8-------------------------------
391# [OPTIONAL] Your machine's hostname is baked in at install time;
392# in the rare case you need to change it: it appears to be in:
393#   o /etc/sysconfig/network
394#   o your lvm thingies; probably don't need to edit
396# [TESTSERVER] Enable password log in
397        vim /etc/ssh/sshd_config
398        service sshd reload
399        vim /etc/pam.d/sshd
400# Replace the first auth block with:
401#           # If they're not root, but their user exists (success),
402#           auth    [success=ignore ignore=ignore default=1] uid > 0
403#           # print the "You don't have tickets" error:
404#           auth    [success=die ignore=reset default=die] file=/etc/
405#           # If !(they are root),
406#           auth    [success=1 ignore=ignore default=ignore] uid eq 0
407#           # print the "your account doesn't exist" error:
408#           auth    [success=die ignore=reset default=die] file=/etc/
411# [WIZARD/TESTSERVER] If you are setting up a non-production server,
412# there are some services that it won't provide, and you will need to
413# make it talk to a real server instead.  In particular:
414#   - We don't serve the web, so don't bind
415#   - We don't serve LDAP, so use another server
416# This involves editing the following files:
417        \rm /etc/sysconfig/network-scripts/ifcfg-lo:{0,1,2,3}
418        \rm /etc/sysconfig/network-scripts/route-eth1 # [TESTSERVER] only
419#   o /etc/nslcd.conf
420#       replace: uri ldapi://%2fvar%2frun%2fdirsrv%2fslapd-scripts.socket/
421#       with: uri ldap://
422#           (what happened to nss-ldapd?)
423#   o /etc/openldap/ldap.conf
424#       add: URI ldap://
425#            BASE dc=scripts,dc=mit,dc=edu
426#   o /etc/httpd/conf.d/vhost_ldap.conf
427#       replace: VhostLDAPUrl "ldap://,dc=scripts,dc=mit,dc=edu"
428#       with: VhostLDAPUrl "ldap://,dc=scripts,dc=mit,dc=edu"
429#   o /etc/postfix/virtual-alias-{domains,maps}
430#       replace: server_host ldapi://%2fvar%2frun%2fdirsrv%2fslapd-scripts.socket/
431#       with: server_host = ldap://
432# to use instead of localhost.
433# XXX: someone should write sed scripts to do this
435# [WIZARD/TESTSERVER] If you are setting up a non-production server,
436# afsagent's cronjob will attempt to be renewing with the wrong
437# credentials (daemon.scripts). Change this:
438    vim /home/afsagent/renew # replace all mentions of
441#   - You need a self-signed SSL cert or Apache will refuse to start
442#     or do SSL.  Generate with:
443    openssl req -new -x509 -keyout /etc/pki/tls/private/scripts.key -out /etc/pki/tls/certs/scripts.cert -nodes
444    ln -s /etc/pki/tls/private/scripts.key /etc/pki/tls/private/scripts-1024.key
445#     Also make /etc/pki/tls/certs/ca.pem match up
446    openssl rsa -in /etc/pki/tls/private/scripts.key -pubout > /etc/pki/tls/certs/ca.pem
448# [TESTSERVER] More stuff for test servers
449#   - Make (/etc/aliases) root mail go to /dev/null, so we don't spam people
450#   - Edit /etc/httpd/conf.d/scripts-vhost-names.conf to have
451#     be an accepted vhost name
452#   - Look at the old test server and see what config changes are floating around
Note: See TracBrowser for help on using the repository browser.