User talk:Jch

How to automate KVM creation

The goal is not only to have a working install but to have it at the after setup-alpine stage without human intervention... Tis is the first stages of a work in progress...

I want to pass a Block Device and a name as parameters. The block device could be an image file, a LV, a NBD, a hdd, a raid array, whatever.
Everything else should be fully automatic according to some config file (stating the http-proxy, the time server, the log server, ...).

The I will just run the script, watch my dhcp logs to discover the new IP assigned (that's why the name is a parameter), then log in with ssh without password to customize it further but at high level only (will be a robot and not me in fact).

I guess it would be something like emulate boot from usb key with specific overlay already on key...
then run setup-disk with proper parameters on the command line to avoid the interactive process (like setup-alpine does)...
Methink this could be done from a couple of scripts put in /etc/local.d/. The last.stop one deleting all of them to be clean at next reboot.
Let's start easy ;)

How to prepare a img file to emulate an USB key

first a working example done in console (accessed trough ssh).
Will build a script from it...

First, lets's prepare somme block device (here an image file but could be something else)

apk add qemu-img
qemu-img create -f raw usbkey.img 512M
apk del qemu-img
T="usbkey.img"

Next, let's install AL on this $T

apk add multipath-tools syslinux dosfstools
fdisk $T
kpartx -av $T
mkdosfs -F32 /dev/mapper/loop1p1
dd if=/usr/share/syslinux/mbr.bin of=/dev/mapper/loop1
syslinux /dev/mapper/loop1p1
mkdir key
mount -t vfat /dev/mapper/loop1p1 key
wget http://wiki.alpinelinux.org/cgi-bin/dl.cgi/v3.1/releases/x86_64/alpine-mini-3.1.1-x86_64.iso
mkdir cdrom
mount alpine-mini-3.1.1-x86_64.iso cdrom
cd cdrom
cp -a .alpine-release * ../key/
cd ..
umount key
umount cdrom
kpartx -d $T
apk del multipath-tools syslinux dosfstools
rm alpine-mini-3.1.1-x86_64.iso

This block device may now be use to boot some KVM for instance like:

screen -d -m -S KVM-builder \
qemu-system-x86_64 -name KVM-usb -enable-kvm -cpu qemu64 -curses \
-device nec-usb-xhci -drive if=none,id=usbstick,file=$T -device usb-storage,drive=usbstick

This is working fine. The problem is when adding a HDD to the lot, qemu try to boot from the hdd and does not even try to boot from the usb key. Enabling menu in boot let's one access the emulated bios which allows to select USB device to boot interactively but this break the goal of fully automated boot :( The stanza is for instance

screen -d -m -S KVM-builder \
qemu-system-x86_64 -name KVM-usb -enable-kvm -cpu qemu64 -curses \
-device nec-usb-xhci -drive if=none,id=usbstick,file=$T -device usb-storage,drive=usbstick \
-drive file=$T2 boot menu=on

qemu-doc states that very clearly:
> -boot [order=drives][,once=drives][,menu=on|off][,splash=sp_name][,splash-time=sp_time][,reboot-timeout=rb_timeout][,strict=on|off]
> Specify boot order drives as a string of drive letters. Valid drive letters depend on the target achitecture. The x86 PC uses: a, b (floppy 1 and 2), c (first hard disk), d (first CD-ROM), n-p (Etherboot from network adapter 1-4), hard disk boot is the default

Starting AL from network

As it does not seems possible to start qemu with a virtual USB key *and* a virtual HDD attached to the VM. Let's try something different: to start AL from the network and mount the HDD later on...

Usually this kind of setup needs

• a DHCP server to get an IP address and the location of the TFTP server
• a TFTP server to download the kernel and tje root file system to boot from
• a NFS server or a HTTP one to get the overlay used to configure the machine
• a NFS server to share files with others
• a NBD server to get his own block devices as storage
• a machine where to prepare initramfs

First, let's check what is vailable in AL and what is not...

• dhcpcd-6.6.7-r0
• tftp-hpa-5.2-r1
• nfs-utils-1.3.1-r2
• darkhttpd-1.10-r1
• nbd-3-10-r0

PXE_boot

We are trying to do something as in PXE_boot.

We did it on separate machine for each service. It forces us to deeply understand all interactions between processes.

dhcpd

192.168.1.1

with package dhcp from repo. Nothing special.

  filename "pxelinux.0";
  next-server 192.168.1.2;

and

# Disable RFC 2136 dynamic DNS updates.
ddns-update-style none;

# Define actions to take when leases are committed, released, or expired to
# accomplish dynamic DNS updates to djbdns. This does not use the RFC 2136
# update mechanism, because djbdns does not support it. However, it
# accomplishes the same thing.
# syntax "execute(cmd, arg, ...)"
### need to check if the two "on EVENT" must be nested or in sequence...
on commit {
  execute ("/usr/local/bin/dns-update-djb",
           "commit",
           lcase (option host-name),
           config-option domain-name,
           binary-to-ascii (10, 8, ".", leased-address));
  on release or expiry {
    execute ("/usr/local/bin/dns-update-djb",
             "release",
             binary-to-ascii (10, 8, ".", leased-address));
  }
}

with a custom /usr/local/bin/dns-update-djb script largely inspired from https://sites.google.com/site/dmoulding/dns-update-djb but adapted for a distant tinydns server and to the AL way.

tftp

192.168.1.2

tftp-hpa configured to serve some SYSLINUX files.

The config is in /etc/conf.d/in.tftpd
Then to issue:

rc-update add in.tftpd
rc-service in.tftpd start

We serve from /var/tftpboot.

We add to temporary install the syslinux apk to get pxelinix.0 and other libs needed.
We did prepare a "pxerd" initramfs file with virtio_net.ko, dhcp and nfs included; made sure loop and squashfs are included.

pxelinux.cfg/default looks like

PROMPT 0
TIMEOUT 3
default alpine
LABEL alpine
LINUX alpine/vmlinuz-grsec
INITRD alpine/pxerd
APPEND ip=dhcp alpine_dev=nfs:192.168.1.3:/srv/boot/alpine modloop=/boot/grsec.modloop.squashfs nomodeset quiet apkovl=http://192.168.1.4/localhost.apkovl.tar.gz
#APPEND modloop=http:/192.168.1.4/grsec.modloop.squashfs
#APPEND apkovl=http://192.168.1.4/localhost.apkovl.tar.gz # including the modloop hack
#APPEND alpine_repo=http://repo-url

Modules are loaded

/ # lsmod
Module                  Size  Used by    Not tainted
nfsv3                  22784  1
nfs                   144376  2 nfsv3
lockd                  71917  2 nfsv3,nfs
sunrpc                225574  6 nfsv3,nfs,lockd
af_packet              28735  0
sr_mod                 13487  0
cdrom                  40424  1 sr_mod
pata_acpi               3326  0
ata_piix               25601  0
ata_generic             3554  0
libata                181955  3 pata_acpi,ata_piix,ata_generic
virtio_net             19684  0
scsi_mod              113710  2 sr_mod,libata
virtio_pci              6485  0
virtio                  4933  2 virtio_net,virtio_pci
virtio_ring             9161  2 virtio_net,virtio_pci
squashfs               25893  1
loop                   18243  2

Network is up

/ # ifconfig
eth0      Link encap:Ethernet  HWaddr 52:54:33:B0:C2:D2
inet addr:192.168.1.108  Bcast:0.0.0.0  Mask:255.255.255.0  
UP BROADCAST RUNNING MULTICAST  MTU:1500  Metric:1  
RX packets:322 errors:0 dropped:0 overruns:0 frame:0  
TX packets:2 errors:0 dropped:0 overruns:0 carrier:0  
collisions:0 txqueuelen:1000
RX bytes:20514 (20.0 KiB)  TX bytes:684 (684.0 B)

but modloop does not load This patch fix this issue (hope to see it mainstream soon)

localhost:~# diff /etc/init.d/modloop modloop.new 
--- /etc/init.d/modloop 
+++ modloop.new
@@ -32,7 +32,7 @@
        local search_dev="$1" fstab="$2" 
        local dev mnt fs mntopts chk
        case "$search_dev" in
-               UUID=*|LABEL=*|/dev/*);;
+               UUID=*|LABEL=*|/dev/*|nfs);;
                *) search_dev=/dev/$search_dev;;
        esac
        local search_real_dev=$(resolve_dev $search_dev)
@@ -49,6 +49,10 @@
                                fi
                        done
                done
+               if [ "$fs" = "$search_dev" ]; then
+                       echo "$mnt" 
+                       return
+               fi
        done < $fstab 2>/dev/null
 } 

References

http://www.syslinux.org/wiki/index.php/PXELINUX

nfs

192.168.1.3

see http://wiki.alpinelinux.org/wiki/User_talk:Jch#NFS_bug_study
It is now working with http://dev.alpinelinux.org/~clandmeter/rpcbind-0.2.3_rc2-r0.apk

We serve the content of an usb key (iso) in ro as

/srv/boot/alpine	*(ro,no_root_squash,no_subtree_check)

http

192.168.1.4

With package Darkhttpd from repo serving from /var/tftpboot/ to serve files needed to boot (kernel, rootfs, apkovl.tar.gz)

nbd

192.168.1.5

I really would like to have xnbd-server in AL. nbd-3.1.0 was just added to edge/testing repo; need to try it in real situation...
For now, we have a qcow2 debian image added to the apkovl with lbu add; lbu ci.
This image is used to launch a first KVM with /dev/mdX as second drive.
In turn, inside the KVM, vdb is used to define a lvm2 volume.
The LV are published with xnbd-server.

Later on, the same KVM will be able to connect to RBD device and re-publish it as NBD.

xnbd-server allows live migration of Block Devices while live. And has a powerfull proxy mode.

All other KVM are running from FS accessed trough NBD from such SAN. Even other SAN.
As soon as those KVM-NBD are up, they may be used to launch others or to provide datastores.

We put that image on every USB key we use along with mdadm and OpenVSwitch (and collectd).

dns

192.168.1.6

tinydns from repo with split-dns config.

Building a complete infrastucture with AL

I'm doing it. It's for real! That's my daily job at present ^^

I'm building a full private cloud bootstraped with only an AlpineLinux USB key for each physical machine. But next ones will be able to boot from network; not even USB keys will be needed. As a matter of fact, we used more than only one physical USB key because we didn't started from scratch but had a live migration from Debian to Alpine for most of the services and machines...

If there is some feed-back, I may develop config files and so on ;)

As I started from scratch and OpenVSwitch was not available in Alpine at that time yet, It took me a while to build everything. But to reproduce it, it would be piece of cake!

We use qemu-kvm for KVM. But I guess one may use whatever Virtual Machine technology one likes.

This is the presentation of a use case. Not a HOW TO. And it's still a work in progess...

Network

Firewall

We put a dedicated physical machine on each link between our LAN and other networks. It just run iptables and some paquets accounting metrology.

Router

Physical machine connected to our LAN and other networks (trough a firewall). A static routing table do the trick.

Switches

All physical machines run OpenVSwitch reproducing virtually all physical switches we have plus some virtuals only.

VPN

All physical machines run openVPN as client to as many switch defined less the physical interfaces of the machine. There is an openVPN server somewhere running in a KVM connected to needed switches.

Storage

SAN

On each physical machine, a couple of HDD are mounted in raid1 witch mdadm. This raid array is passed as parameter to a KVM who in turn mount it as physical volume for LVM. The created LV are published as NBD with xnbd-server. For the time being, this KVM is running debian 7.8 as xnbd is not in Alpine (yet?)..

The SAN also connects to the CEPH cluster as client and publish reached RBD as NBD with xnbd-server. For the time being, this KVM is running debian 7.8 as no xnbd nor RBD are in Alpine (yet?)..

NAS

Some KVM is mounting some NBD as local drives and publishing some directories as NFS shares.
We now have nfs and nbd in AL.

CEPH

KVM with physical HDD as parameters are used for building OSD and MON needed to operate a CEPH cluster. One KVM is the "console" to drive it from a single point of presence (usefull but not "needed").For the time being, those KVM are running debian 7.8 as CEPH and RBD are not in Alpine (yet?)..

Low-level services

No service at all is running in the AL on bare metal. All are running is some KVM connected to needed switches by the means of the OpenVSwitches. The apkovl on the USB keys contains only the scripts to launch KVM and one image file to launch the first SAN. Other KVM are launched from LV in the SAN.

dhcp

Exactly two KVM stored in different SAN, primary and secondary in failover mode, are running dhcpd from repo.
We just have to configure it properly.

We have to test if dhcpd may run in a LXC instead of a KVM?

DNS

tinydns from repo with split-dns config.

Resolver

With dnscache from repo.

Those KVM have manually assigned IP address in the LAN and does know a gateway to the Internet.
They use themselves as resolver...
They know the direct manually assigned IP address in the LAN of the main DNS server of selected domains (for split dns configuration).

PXEboot

kernel and initrd files in tftp server.
copy of usb content in nfs server.
apkovl files in darkhttpd server.

Time server

The router (who has access to internet) usr ntpd (or similar) from repo, to act as client to the WAN and server to the LAN.

syslog

With syslog-ng from repo, we receive the logs from all machines be it physical or virtual.
It's the only place who needs logrotate from repo.

HTTP proxy/cache

The web proxy/cache squid, from repo, uses a NBD as cache. It has a link to the internet to forward requests and one to the LAN.

Because of him, no machine, as they are all connected to the LAN, be it physical or virtual, needs a published default gateway. And all machines are able to install/upgrade packages or to see the WWW as client.

We point all AL boxes to this KVM with setup-proxy.

Monitoring

shinken from sources in some LXC with barely only the python package installed

Metrology

Collectd (one LXC as server, all other machines, be it physical or virtual, as client) with collectd-network from repo.
A couple of lines in CGP config file is enough for now.

Backups

with common tools: rsync, tar, nc, bzip2, openssh, cron

LDAP

openldap with openldap-back-hdb, both from repo.

http://www.openldap.org/doc/admin24/backends.html states
> The hdb backend to slapd(8) is the recommended primary backend for a normal slapd database.
And
> Note: The hdb backend has superseded the bdb backend, and both will soon be deprecated in favor of the new mdb backend.

> The mdb backend to slapd(8) is the upcoming primary backend for a normal slapd database. It uses OpenLDAP's own Lightning Memory-Mapped Database (LMDB) library to store data and is intended to replace the Berkeley DB backends.

Unfortunately there is no openldap-back-mdb package in AL yet.

High-level services

in LXC AL whenever possible.
in LXC Debian as second choice
in KVM otherwise.

x2goserver

I did package nx-libs and x2goserver. I'm waiting for the packages to be included in edge/testing. They are already being used for single app access. Next step is full desktop but we are not sure if AL is the right choice for full desktop usage for our customers...

unfortunately, x2goclient pops up "kex error : did not find one of algos diffie-hellman-group1-sha1 in list curve25519-sha256@libssh.org,ecdh-sha2-nistp256,ecdh-sha2-nistp384,ecdh-sha2-nistp521,diffie-hellman-group-exchange-sha256,diffie-hellman-group14-sha1 for kex algos" need to specify diffie-hellman-group1-sha1 in sshd_config. Luckyly a fix exists and my business partner is looking after a way to enhance it's security upstream.

ejabberd

with ejabberd from edge/testing repo. I migrate the mnesia DB from an old debian squeeze just copying the files and changing ownership in a LXC-AL. I just had to disable mod_pubsub to have it run properly. Authentification is done with openLDAP. I now plan to migrate a very very old jabberd (11 years I guess) running on a debian etch to it if I find a way to keep users's password and rosters... I also would like to use it as a gateway to IRC to follow #alpine, #alpine-devel and #x2go channels ;) Some other ejabberd features are interesting to my organisation and we will experiment more in depth, namely mod_sip, mod_stun, mod_proxy65...

redmine

in a brand new LXC with edge/main and edge/testing repos
mostly following Redmine page
I use a mariaDB server on another host where I did create the user and pushed the sqldump from a running redmine 3.0.0 instance

apk update
apk upgrade
reboot
setup-timezone
apk add redmine
apk add ruby-unicorn
cp /etc/unicorn/redmine.conf.rb.sample /etc/unicorn/redmine.conf.rb
vi /etc/conf.d/unicorn
vi /etc/redmine/database.yml 
apk add sudo
apk add ruby-mysql2
apk add ruby-yard
apk add tzdata
cd /usr/share/webapps/redmine
sudo -u redmine rake generate_secret_token
sudo -u redmine RAILS_ENV=production rake db:migrate

Master key

We want to be able to bootstrap the full infrastructure from only one usb key and one machine with physical access (to insert the usb key obviously).

This key will run AL stable. With only very few packages installed. But some images on the sorage.

Initial packages

dhcpd tftp syslinux nfs darkhttp openssh vim openvswitch mdadm qemu screen collectd collectd-network gptdisk irqbalance ssmtp mailx

Bootstrap

First, we setup the network. Remember, this is a bootstrap. We assume nothing exect to have (or not) an internet connexion on some ethernet cable plugged into some NIC. It means we may take any decision we see fit.

Our primary machine is the only fixed point for now. Let's give it the number 1.
All machines are connected to the LAN. We know nothing yet about other NICs.
First we must decide about the LAN IP range. For instance be it 192.168.1.0/24.
We will use complicated network setup, let's start by installing openvswitch on bare metal

apk add openvswitch
rc-update add ovs-modules
rc-update add ovsdb-server
rc-update add ovs-vswitch
rc-service ovs-modules start
rc-service ovsdb-server start
rc-service ovs-vswitch start
ovs-vsctl add-br lan
ovs-vsctl add-port lan eth0
vi /etc/network/interfaces #iface eth0 inet manual
                           #iface lan inet dhcp

No machine will offer any service from bare metal.

apk add qemu-system-x86_64 screen libusb
modprobe kvm
modprobe kvm-intel
modprobe kvm-amd
modprobe tun
screen -m -d -S KVM-infra qemu-system-x86_64 -kvm -kernel /kernel -initrd /initrd -append alpine_dev=...,apkovl=... -net -net -drive /dev/usb

La suite immédiate se fait dans cette VM

screen -r KVM-infra

We need the storage space from the usb key to handle boot images and apkovl files.
In KVM-infra

setup-alpine (fixed IP)
apk add openvswitch
rc-update add ovs-modules
rc-update add ovsdb-server
rc-update add ovs-vswitch
rc-service ovs-modules start
rc-service ovsdb-server start
rc-service ovs-vswitch start
ovs-vsctl add-br lan
ovs-vsctl add-port lan eth0
vi /etc/network/interfaces #iface eth0 inet manual
                           #iface lan inet static
mkdir -p /srv
mount /vda2 /srv
mkdir -p /srv/nfs/alpine
mount /vda1 /srv/nfs/alpine

Next we will have other networked devices. We need dhcpd (in KVM-infra)

apk add dhcpd 
rc-update dhcpd
vi /etc/dhcp/dhcpd.conf  #filename "pxelinux.0";
                         #next-server 192.168.1.1;
rc-service dhcpd start

To boot others, we need tftp, nfs and http (tftp and http in LXC, nfs in KVM-infra)

apk add tftp-hla nfs-utils darkhttp
rc-update add nfs
rc-update add tftp
rc-update add darkhttpd
mkdir -p /srv/nfs
mkdir -p /srv/tftp
mkdir -p /srv/http
vi /etc/exports
vi /etc/tftp 
vi /etc/darkhttp
rc-service tftp start
rc-service darkhttpd start
rc-service nfs start

Let's populate those servers.

mkdir -p /srv/tftp/alpine
cp /media/usb/boot/vmlinuz* /srv/tftp/alpine/
cp /media/usb/boot/modloop* /srv/tftp/alpine/
apk add mkinitfs
cd /etc/mkinitfs
vi features.d/network.modules
vi features.d/dhcp.files
vi features.d/dhcp.modules
vi features.d/nfs.modules
vi mkinitfs.conf # add network, dhcp, nfs and squashfs
mkinitfs -o /srv/tftp/alpine/pxerd
apk del mkinitfs
apk add syslinux
cp /usr/share/syslinux/pxelinux.0 /srv/tftp/
cp /usr/share/syslinux/ldlinux.c32 /srv/tftp/
apk del syslinux
mkdir -p /srv/tftp/pxelinux.cfg
vi /srv/tftp/pxelinux.cfg/default

All we need now to boot another AL machines (be it physical or virtual) are some {MAC}.apkovl.tar.gz files served by darkhttpd. We badly need name resolution at this stage. DNS and resolver are needed. DNS to be updated dynamically by dhcp server with split-dns. Resolver knowing the fixed IP address on the DNS and the default route if known at this stage. Both may run in LXC inside this KVM-infra (like the other previous services). DNS will be djbdns and resolver will be dnscache (both from repo).

Then (for now) we need image of a debian install with xnbd-server and lvm2 to build SAN.
Also, on bare-metal we need mdadm to assemble raid1 arrays.
A new SAN is therefore, a MAC address (for debian boot as san), some BD as vda (raid1 from mdadm).
A new server is therefore, a MAC address (for AL boot), a apkovl file (MAC named), some data NBD from some SAN.
The apkovl will be downloaded at boot time with PXE provided address; before launching openvswitch! The IP address will then change because of the apparent MAC change when OVS becomes active.
We may use symlinks to MAC named config files to have a more human friendly view.

It is to be noted that after bootstrap KVM may move to other physical machines. While some KVM-infra is somehow connected to the LAN, everything stay alive! This precise image will be reproduced in every SAN build.

Deploy

After bootstraping, we dispose of a way to boot any AL KVM or bare-bone in about 10 sec.

First we deploy KVM-SAN on bare-metal.

Next we deploy KVM-AL grouping (or not) some LXC (AL or debian).

Second we deploy low-level services: syslog-ng, fail2ban, openVPN, la_console, http-reverse-proxy (primary and secondary), http-proxy, smtp relay, secondary resolver, secondary dns, ldap (primary and secondary), NAS, mariaDB, backups, collectd, shinken, local AL repo, git

Third intermediary services: smtp in, smtp out, antivirus, antispam, smtp store, imap, pop3, http, php, sip, jabber

High level services: x2goserver, lamp, mail toaster, webdav, redmine, etc

For each of those services, we provide a template in the form of a {kvm-template}.apkovl.tar.gz.
After customisation, "lbu package" followed by sending the a.tgz to the central repository is all needed.
We follow a naming convention for MAC:

For bare metal, the 3 first bytes of the MAC is the manufacturer ID.
We symlink that to the baremetal.apkovl.tar.gz.

For KVM, we fix the MAC ourself.
The first 2 bytes (AA:BB) are fixed.
The third one (CC) is the level type of the KVM.
The fourth one (DD) is the specific type of the template.
The last 2 ones are incremental unique ID.
So we are able to define pxelinux.cfg/AA:BB:CC:DD symlinks to config files defining use of {kvm-template}.apkovl.tar.gz.

As {kvm-template}.apkovl.tar.gz tend to be small, we can store a lot of those on the initial USB stick.
Depending on available space on the USB stick, we could offer {lxc-template}s that way from the USB stick to be downloaded from darkhttpd with wget to the right KVM. Or later on from any wanted NAS.

We add a couple of other OVS (WAN, STORAGE) in every machines. Some are connected to NIC. Some are connected to VPN. Netflow will be used in the future to manage the network (naas: network as a service). One of those OVS (WAN) allows connected machines to access the internet trough a default route passing through a physical firewall. STORAGE is used for data replication between SAN and NAS.

We have the list of bare-metal machines.
Those may launch KVM in one command.

We have the list of SAN KVM.
Those may create and publish NBD in two commands.
Even on diskless' machines those are present to offer nbd-proxy in one command.

All those command are grouped as one-liner scripts in some redundant NAS available from la_console.

Waiting for CEPH, we need a strategy for duplicating NBD accros SAN.