Software - 2

Getting a serial console (for login purposes) on OpenSolaris seems to be a two-sided operation. Either it is insultingly simple, or a convoluted mess. The two options I've discovered so far, in order of difficulty.

h3. The easy way

By default Solaris starts a login console on /dev/console, whatever that happens to be at a given moment. Passing -B console=ttya on the kernel command line (via GRUB) will change the primary system console to the first serial port in the system (what Linux calls ttyS0), automagically creating a login prompt there when the system has finished booting.

h3. The hard way

Assuming that for whatever reason /dev/console is not supposed to be on ttya, things move into twilight territory.

Historically a getty was spawned from inittab, and that was it. Sometime in the past Solaris changed this behaviour to a more complicated scheme.

Login processes are handled by the Service Access Controller (sac(1M)), which handles port monitors. Port monitors bind services to ports, for example ttymon(1M), which handles baud rates and spawning the login process. I have to admin that I have not completely understood the architectural design behind all this. As far as I was able to deduce from the manual pages for pmadm(1M), ttyadm(1M) and sacadm(1M) the default setting ought to produce a login prompt on the first two serial ports. Alas, it does not.

It looks like ttymons are no longer handled by SAC, no matter what the documentation says, but are instead spawned directly by SMF, specifically by svc:/system/console-login. This SVC has several instances that handle the login prompts on the virtual consoles for VGA. Adding an instance for ttya makes a login prompt appear on the first serial port. The following creates and activates this instance:

# svccfg -f - <

Please note that due to the default settings in /etc/security/login root logins are not possible on this terminal.

Now I am waiting for one of two things to happen:


Someone telling me that I've got it all wrong and that there's a better way to do things
Someone telling me that this is the official way to do it and that it is documented somwhere.


I'm not holding my breath on either one.

            

For some reason or another finding instructions on how to actually configure iSCSI for a rather simple and common use case (one target on Solaris, one initiator on Fedora, CHAP authentication) seems to be pretty hard.

Either my google-fu is seriously broken today, or everyone except for me considers this to be so easy and obvious that it does not warrant documentation. Which, having done this for the last three hours, I seriously doubt. The Solaris side is actually documented quite well (I primarily used this blog entry by alasdair), the Linux side is lacking. It does not help matters that both sides use identically named tools that work in completely different ways.

So, the deal is as follows:

One OpenSolaris system, acting as iSCSI target (that is the system presenting the storage space).

One Fedora Linux system, acting as iSCSI initiator (that is the system that wants to use the storage space)

Create 100GB storage space on the target and let the initiator connect to this storage space and create a filesystem on in. The connection has to be authenticated one way (the initiator presents credentials to the target).

h3. The Solaris side

The system is running Nevada, the configuration here was done with build snv_110.

First, the iSCSI target software needs to be installed and enabled:

# gkp.pl -d /mnt/Solaris_11/Product SUNWiscsitgtu
# svcadm enable iscsitgt:default
#

The backing store for the iSCSI volumes shall be provided by ZFS:

# zfs create -o canmount=off tank/iscsi
# zfs create -V 100G -o shareiscsi=on tank/iscsi/vol001
# iscsitadm list target -v
Target: tank/iscsi/vol001
    iSCSI Name: iqn.1986-03.com.sun:02:218c35e0-0881-cb4f-d6bd-80e08bdc98d8
    Alias: tank/iscsi/vol001
[...]
#

The iSCSI Name will be different for each created volume. Per default this target will be available via all interfaces and IPs of the machine. Since this is not what I want in this special case the target is bound to a fixed IP by a construct called a Target Port Group Tag. The TPGT used here has the number 1.

# iscsitadm create tpgt 1
# iscsitadm modify tpgt -i 212.51.12.90 1
# iscsitadm list tpgt -v 1
TPGT: 1
    IP Address: 212.51.12.90
# iscsitadm modify target -p 1 tank/iscsi/vol001
# iscsitadm list target -v
Target: tank/iscsi/vol001
[...]
    TPGT list:
            TPGT: 1
[...]
#

In order to secure access to this volume some more the initiator is required to authenticate itself using CHAP before access is granted. To do this three pieces of information are needed:

• The iqn (basically a unique name) of the initiator
• A username
• A password

The initiator used here has an iqn of iqn.2005-03.com.max:01.cb5c4c (see below for the source of this information). The username is lain and the password is iscsipassword for the purpose of this demonstration. The password has to be between 12 and 16 charcters in length and should definitely be something stronger for production purposes.

# iscsitadm create initiator -n iqn.2005-03.com.max:01.cb5c4c lain
# iscsitadm modify initiator --chap-name lain lain
# iscsitadm modify initiator --chap-secret lain
[...]
# iscsitadm list initiator -v
Initiator: lain
    iSCSI Name: iqn.2005-03.com.max:01.cb5c4c
    CHAP Name: lain
    CHAP Secret: Set
# iscsitadm modify target --acl lain tank/iscsi/vol001
# iscsitadm list target -v
Target: tank/iscsi/vol001
[...]
    ACL list:
            Initiator: lain
[...]
#

This concludes the Solaris side of things.

h3. The Fedora side

The system is running Fedora Rawhide, close to the Fedora 11 Beta release at the time of this writing.

On the Linux side iSCSI is handled by the open-iscsi toolchain, packaged as iscsi-initiator-utils in Fedora. To install it:

# yum install iscsi-initiator-utils
[...]
#

Since the system in question is a notebook, and the iSCSI target may not be available at all times, the iSCSI service must be instructed not to connect to configured devices automatically:

# perl -pi -e 's/node.startup.*/node.startup = manual/' /etc/iscsi/iscsid.conf
#

The initiator name mentioned in the Solaris section above can be configured freely on the system. A random value is created during package installation and saved in /etc/iscsi/initiatorname.iscsi:

# cat /etc/iscsi/initiatorname.iscsi
InitiatorName=iqn.2005-03.com.max:01.cb5c4c
#

Be sure the name configured there matches the name defined in the initiator on the Solaris side. Even though a username/password pair was defined on the target credentials are not needed for target discovery (the process by which an initiator asks a target which iSCSI volumes are available):

# iscsiadm -m discovery -t st -p 212.51.12.90
212.51.12.90:3260,1 iqn.1986-03.com.sun:02:218c35e0-0881-cb4f-d6bd-80e08bdc98d8
# iscsiadm -m node
212.51.12.90:3260,1 iqn.1986-03.com.sun:02:218c35e0-0881-cb4f-d6bd-80e08bdc98d8
#

The discovery process has found a single volume exported by the target and added it to the local node list. The iqn matches the value shown above in the Solaris section. Now the CHAP credentials have to be added to the node so the initiator can actually connect to the volume:

# iscsiadm -m node --target 'iqn.1986-03.com.sun:02:218c35e0-0881-cb4f-d6bd-80e08bdc98d8' \
--name 'node.session.auth.authmethod' -v 'CHAP'
# iscsiadm -m node --target 'iqn.1986-03.com.sun:02:218c35e0-0881-cb4f-d6bd-80e08bdc98d8' \
--name 'node.session.auth.username' -v 'lain'
# iscsiadm -m node --target 'iqn.1986-03.com.sun:02:218c35e0-0881-cb4f-d6bd-80e08bdc98d8' \
--name 'node.session.auth.password' -v 'iscsipassword'
# iscsiadm -m node --target 'iqn.1986-03.com.sun:02:218c35e0-0881-cb4f-d6bd-80e08bdc98d8'
node.name = iqn.1986-03.com.sun:02:218c35e0-0881-cb4f-d6bd-80e08bdc98d8
node.tpgt = 1
node.startup = manual
[...]
node.session.auth.authmethod = CHAP
node.session.auth.username = lain
node.session.auth.password = ********
[...]
#

Now the volume can finally be accessed:

# iscsiadm -m node --target 'iqn.1986-03.com.sun:02:218c35e0-0881-cb4f-d6bd-80e08bdc98d8' --login
Logging in to [iface: default, target: iqn.1986-03.com.sun:02:218c35e0-0881-cb4f-d6bd-80e08bdc98d8,
portal: 212.51.12.90,3260]
Login to [iface: default, target: iqn.1986-03.com.sun:02:218c35e0-0881-cb4f-d6bd-80e08bdc98d8,
portal: 212.51.12.90,3260]: successful
#

After some seconds a device node for this volume should appear in /dev/disk/by-path:

# ls /dev/disk/by-path/
ip-212.51.12.90:3260-iscsi-iqn.1986-03.com.sun:02:218c35e0-0881-cb4f-d6bd-80e08bdc98d8-lun-0
[...]
#

The disk is ready to be used.

h3. Initial install

pkgsrc can be obtained in two main ways. The first is via anonymous CVS checkout, the second is via quarterly tarball releases. Since the whole tree is several hundred megabytes (even without sources) and the NetBSD CVS servers are permanently overloaded a tarball is the way to go for the initial install. The tarballs can be found in the NetBSD mirror structure, the main mirror being here. At the time of this writing pkgsrc-2008Q4 is current.

The tarball already contains a pkgsrc directory at the top level, so it has to be unpacked under /usr.

# wget ftp://ftp.netbsd.org/pub/pkgsrc/current/pkgsrc.tar.gz
# gtar -xzf pkgsrc.tar.gz -C /usr
# chown -R builder: /usr/pkgsrc

This will unpack the tarball into the filesystem mounted at /usr/pkgsrc and change the ownership to the build user created earlier.

h3. Patching gcc

Unfortunately the gcc suite as delivered with Solaris has a small flaw, which will cause some packages to be built incorrectly (the most famous example is openssl, which will build but not work afterwards). Fortunately the bug has been tracked down and a bandaid is available here.

This file is a nice little hack in itself as it is a shell script and a C source file at the same time. If executed by a shell this file will be put though gcc three times to produce three object files, which are placed into a directory where the compiler can find them and use them instead of files that were delivered with the compiler.

# bash ./values.c /usr/sfw/bin/gcc
[....]
# ls -l $(dirname $(/usr/sfw/bin/gcc -print-libgcc-file-name))
[...]
-rw-r--r-- 1 root root    763 2009-02-21 17:34 values-Xa.o
-rw-r--r-- 1 root root    763 2009-02-21 17:34 values-Xc.o
-rw-r--r-- 1 root root    763 2009-02-21 17:34 values-Xt.o
#

h3. Bootstrapping

With this particular bug out of the way pkgsrc needs to be bootstrapped. This means that the initial set of programs needed to build further packages (the package management itself and some helper programs) need to be built and installed on the host system. This has to be done as root. In order to find some programs it is necessary to expand the $PATH a bit.

# cd /usr/pkgsrc/bootstrap
# PATH="$PATH:/usr/sfw/bin:/usr/xpg4/bin" ./bootstrap
[....]
#

Bootstrapping will take a while, but it should run though cleanly. Afterwards there should be some files in /usr/pkg and running /usr/pkg/sbin/pkg_info should list a handful of installed packages.

h3. Vulnerabilities and updates

Packages need to be kept up-to-date, for new features as much as for possible vulnerabilites. In the latter department the bootstrap installed two programs to help with identifying such programs.

/usr/pkg/sbin/download-vulnerability-list will fetch a list of vulnerable program versions. Builds of such versions will fail, unless explicitly requested otherwise. For already installed programs there is /usr/sbin/pkg/audit-packages which will identify and list installed vulnerable versions.

Both programs should be used on a regular basis.

Updates of packages (whether due to a vulnerability or not) are best done via CVS. Updates can be done on the whole package tree or on individual subtrees, as needed. Assuming an older installed version the following command will update the whole tree to 2008Q4:

# su - builder
$ cd /usr/pkgsrc
$ cvs up -dPR -rpkgsrc-2008Q4
[...]
$

It's important to check the output of the CVS command for eventual problems, especially if local modifications to packages have been done.

h3. Host tools

There are a number of tools that are used by a large number of packages during the build process. pkgsrc will build all of these from source if necessary, but most of them can be supplied by the host system instead.

Among these tools are

• GNU m4
• GNU make
• GNU gettext
• GNU tar
• Perl
• GNU patch
• unzip
• BISON & FLEX

GNU make and GNU tar are already installed, the rest of the packages can be found on the Solaris install media. The rest can easily be installed:

# gkp.pl -d /mnt/Solaris_11/Product SUNWgm4 SUNWgnu-gettext SUNWgpch SUNWunzip \
SUNWbison SUNWflexlex
[...]
#

Now pkgsrc has to be made aware of these tools in order to use them. The main config file is /usr/pkg/etc/mk.conf. Adding the following statements somewhere above the final .endif will make pkgsrc use the host version of the above mentioned tools instead of building it's own:

# SUNWgm4
TOOLS_PLATFORM.m4=              /usr/bin/gm4
TOOLS_PLATFORM.gm4=             /usr/bin/gm4
# SUNWgmake
TOOLS_PLATFORM.gmake=           /usr/bin/gmake
# SUNWgnu-gettext
TOOLS_PLATFORM.msgfmt=          /usr/bin/gmsgfmt
# SUNWgtar
TOOLS_PLATFORM.tar=             /usr/bin/gtar
TOOLS_PLATFORM.bsdtar=          /usr/bin/gtar
# SUNWgpch
TOOLS_PLATFORM.patch=           /usr/bin/gpatch
TOOLS_PLATFORM.gpatch=          /usr/bin/gpatch
#
TOOLS_PLATFORM.perl=            /usr/bin/perl
# SUNWunzip
TOOLS_PLATFORM.unzip=           /usr/bin/unzip
# SUNWbison
TOOLS_PLATFORM.bison=           /usr/bin/bison
TOOLS_PLATFORM.bison-yacc=      /usr/bin/bison -y
# SUNWflexlex
TOOLS_PLATFORM.lex=             /usr/bin/flex

Preparing the system for pkgsrc involves several steps. A compiler has to be installed to bootstrap pkgsrc, and a directory structure has to be prepared in which the files downloaded and generated will be stored. Moreover, it is a good idea to create a special user which will be used to build software.

h3. Compiler environment

The Solaris core installation does not come with a compiler, but a version of GCC 3 can be found on the DVD (3.4.3 in snv_109). pkgsrc will also work with the proprietary Sun compiler suite, but that is beyond the scope here.

In addition to the compiler itself several other programs are needed to get pkgsrc off the ground.

• Binutils (SUNWbinutils)
• GNU Make (SUNWgmake)
• System header files (SUNWhea)
• System libraries (SUNWlibmr, SUNWlibm)
• XPG4 compatible system utilities (SUNWxcu4)

The complete command line for installing all this is

# gkp.pl -d /mnt/Solaris_11/Product SUNWarc SUNWgccruntime SUNWbinutils SUNWgcc \
SUNWgmake SUNWhea SUNWlibmr SUNWlibm SUNWxcu4 SUNWsprot

The compiler is installed into /usr/sfw/bin by default, which is not in the system search path.

h3. Directories and users

By default pkgsrc uses three directories on the system, in which all files are installed.

• /usr/pkg for installed packages
• /usr/pkgsrc for the source tree
• /var/db/pkg for the package database

There is no good reason to change these defaults here. /usr/pkg and /usr/pkgsrc will get their own ZFS filesystems, while /var/db/pkg will just be a directory.

# zfs create -o mountpoint=/usr/pkg rpool/pkg
# zfs create -o mountpoint=/usr/pkgsrc rpool/pkgsrc

These two commands create two new filesystems on the root pool and set the mount points to the correct directories. The filesystems are automatically mounted and are immediately ready to be used.

Having a special user for building packages instead of using root is a good idea in general. It protects the system from eventual errors in the build system, which might cause files to be written outside the build root (even if those files are completely harmless and not malicious they are a nuisance nonetheless). The build user should also have no special privileges on the system.

# useradd -d /export/home/builder -m -s /usr/bin/bash -c "pkgsrc build user" builder

Even though Solaris comes with a large selection of software these days (and some of it in recent versions) there still will be software that is not on the install media. In my case the greatest itch was rtorrent.

I was not too keen on the idea of building everything from scratch the classic way (tar; configure; make; make install;), as this makes tracking software versions and removing software somewhat more complicated than I prefer it to be. If there had to be compiling it would have to come with some kind of system that kept track of the various programs it had compiled.

All this sounded a lot like the FreeBSD ports system to me which I found to be unavailable on Solaris. But the general idea was sound, and in the immediate neighbourhood I discovered the NetBSD pkgsrc system.

pkgsrc is what the ports system is to FreeBSD, a directory tree containing files describing the method to build a certain piece of software (dependencies, source files, patches and so on). It is primarily developed for NetBSD, of course, but over the years support for other UNIX like operating systems has been added, allowing the system to build software on non-NetBSD systems. Solaris has been on the list of supported systems for several years now.

pkgsrc is able to build a self-contained software repository on the guest system, including it's own compiler, binutils and all the other utilities needed to build software. Of course it will initially need a compiler supplied by the guest system, but it's not necessary to keep using that compiler. In other words, pkgsrc can pull a complete Münchhausen, if needed.

On the other hand it can be instructed to use the compiler and tools supplied by the guest operating system, if those are available and recent enough, so that just software that the guest system does not supply is built by pkgsrc.

pkgsrc will by default use it's own dedicated directory tree to keep the software built by it and write only a minimum amount of files into the guest systems file system space. Getting rid of pkgsrc can be as simple as removing three directories (the source tree, the binary tree and the package database tree).pkgsrc also builds it's own package management tools that allow clean deinstallation of it's packages.

Since I tend to forget these things, here's a short list showing various package management tasks.

|_. Task |_. RPM |_. pkg (Solaris) | | List all installed packages | rpm -qa | pkginfo | | Show all files belonging to an installed package | rpm -ql [package] | pkgchk -l [package] | grep Pathname | | Find package owning a file | rpm -qf [filename] | pkgchk -l -p [filename] |

Some releases ago Solaris introduces a Role Bases Access System (details and introduction here). The upshot of this is that Profiles can be attached to a user account, giving this account elevated privileges for some tasks.

The main administrator account should probably be able to execute all command with root privileges. Linux/BSD systems use sudo for this, Solaris has pfexec. From a high level standpoint those two commands serve much the same purpose.

Almost all tutorials about pfexec will mention the "Primary Administrator" profile, which will give an account the permissions to execute every command as root. The rub is that a (reduced) Solaris install does not contain this profile. Instead, it is hidden in the SUNWwbcor package. It has to be installed separately.

# gkp.pl -d /mnt/Solaris_11/Product SUNWwbcor
[...]
# usermod -P'Primary Administrator' admin
# su - admin
$ id
uid=500(admin) gid=100(users)
$ pfexec id
uid=0(root) gid=0(root)

The admin user can now execute every command as root without needing the root password (or any password, for that matter) simply by prefixing it with pfexec

After the newly installed system has booted for the first time it is time to make Solaris a bit more homely. This involves installing a number of packages, which brings me neatly to the first thing that annoys me (and probably a lot of other people used to current Linux distributions): the package management is an absolute mess.

Most Linux distributions take a two-tiered approach to package management: there are low level tools that allow installation of one or more packages, reading from a local filesystem, removal of packages, listing information about packages and so on. RPM is an example of this, as is DPKG.

Then there are high level tools that know about local and remote collections of packages, can resolve dependencies for package installation and removal. YUM, APT and PackageKit are examples of these tools.

Unless something is severely broken the high level tools are used for package management in the usual cases. Noone actually wants to do all the manual dependency solving and package downloading themselves, that's what computers are for, after all.

Unfortunately all Solaris provides are tools of the first category. While I understand that changing the low level tools is impossible for binary compatibility reasons I do not see a reason why high level tools for handling all the grunt work are not being provided.

The logical response to this dilemma is to write a rudimentary high level tool which takes away some of the pain. My version is written in Perl (which, astoundingly, is part of the core install, and not even insultingly old). Currently all it can do is install packages. It relies on the dependency information in the Solaris packages, which is spotty at best.

So, lets make this system a bit more habitable.

mount -F nfs 10.200.200.1:/jumpstart /mnt
gkp.pl -d /mnt/Solaris/Product SUNWgcmn SUNWwgetr SUNWwgetu SUNWntpr SUNWntpu SUNWgtar \
SUNWsshcu SUNWsshr SUNWsshu SUNWsshdr SUNWsshdu SUNWbash SUNWless SUNWdoc SUNWman \
SUNWgnu-coreutils SUNWtoo

Among other things this will install

• Bash
• wget
• GNU tar
• SSH server and client
• man pages

h3. Setting the root shell

I probably won't make many friends here, but I like bash. And I also like a root shell I can actually work with. So, root gets a bash. Contrary to what other people might say this is an absolutely harmless change on modern Solaris systems. /bin/sh still points to ksh (and that should not be changed), but the shell of the root account is not important to the work of the system itself. And should the root shell be unavailable or broken Solaris will fall back to /bin/sh.

usermod -s /usr/bin/bash root

h3. Configuring SSH

Working on the console itself is tedious, so getting SSH up and running is important. Even though the packages are installed they are not configured or enabled yet.

First, though, I usually define a special group to which accounts which are allowed to ssh into the system are added. Then the keys are generated, and finally ssh is enabled.

groupadd sshusers
echo "AllowGroups sshusers" >> /etc/ssh/sshd_config
/lib/svc/method/sshd -c
svcadm enable ssh

h3. Security settings

By default Solaris creates crypt encrypted passwords. While this is compatible with almost every UNIX system since the later jurassic, it is certainly no longer up to date. The following will change the password algorithm to the BSD (and Linux) compatible version of MD5.

perl -pi -e 's/^CRYPT_DEFAULT=.*/CRYPT_DEFAULT=1/' /etc/security/policy.conf

This will not change existing accounts. Only newly created password hashes are affected. This means that the root password should be set again, to create a new hash.

h3. Adding normal users

Working as root is a bad idea, so adding normal users is the order of the day to do normal work on the system. I prefer to have a special group containing all normal users, so this has to be added as well.

In addition, giving every user a separate ZFS filesystem as home directory is a good idea. The dataset rpool/export/home does already exist, the installer created it.

Since this is going to be a storage system without a lot of users there is no big problem with /export/home living on rpool.

groupadd users
zfs create rpool/export/home/admin
useradd -g users -G sshusers -d /export/home/admin -s /usr/bin/bash \
 -c "Admin User" -m admin
chown -R admin:users ~admin

The user becomes a member of the sshusers group and can thus login via SSH (after a password is set).

Installing Solaris can be a strange experience for someone who is only used to modern time Linux installers. Yes, there is a graphical installer, but it consists of little more than an X windows which basically asks the same questions as the text mode installer. Unless you already know how to install Solaris, and what the installer expects of you some of the questions and dialogs seem a little strange.

Due to Solaris' focus on binary compatibility some of the defaults don't make that much sense anymore, either, but changing them to more sensible defaults would cause confusion, or so it seems.

For the install on the storate system, though, most of the defaults are sensible, and since Solaris does not have to share any disks withother operating systems the partitioning process is not that painful, either.

The first question the installer asks (always in text mode) is about the general installaton mode the user wishes to perform (roughly graphical/text based or rescue shell). Interactive/text mode (option 4) is usually fine.

If the system has booted from the network the installer will not ask about IP configuration for the network cards but assume DHCP for IPv4.

The question about the name resolution service is one of the odd quirks in the installer. The naming service defaults to NIS, which is probably wrong for almost any new installation on this planet. Usually DNS is the right choice here. The installer will then ask for the DNS server IPs and default domains. If the installer can not resolve the current machine IP via these nameservers it will explicitly ask for confirmation that the data is really right.

The default answers for the next questions (Kerberos/NFS4) are sensible in the usual cases.

When asked for the file system to use for the root filesystem the default is UFS. Change it to ZFS. I prefer to use separate datasets for / and /var, but that is a matter of personal taste.

The (almost) final question is for the amount of packages to be installed. The installer offers five predefined groups, ranging from several hundred to almost three gigabytes of installed data. Selecting the smallest set will do fine here, the system will boot, have network and NFS client support, which is enough to get at the rest of the packages to install later.

That's it, basically. The installer will now copy the files to the boot disk, prepare the bootloader and restart the system.

After solving the cache issue on the MSI board it was time to install the OS for real. I had planned to do the install via the network, for two reasons.

• To avoid having to burn a DVD for this
• To find out how to do it

Installing Solaris via the network has been supported for a very long time, and Solaris being what it is the process has not changed very much. That means that there are some quirks in it.

Because this was the first Solaris installation in my network some non-Solaris machine had to take over the job of providing the various services needed for an installation. The job was delegated to my notebook running Linux.

The following services are needed to install Solaris over the network: * A DHCP server * A TFTP server * An NFS server

In addition to that some software: * A Solaris medium (obviously) * A bootloader (I'll use SYSLINUX, at least version 3.73 required)

First and foremost, though, the network card in the system that is to be the target of the installation needs to support PXE. PXE is a method that defines a way for a network card BIOS to obtain an IP address via DHCP and load a piece of software from the network, which is then executed by the system. In addition PXE provides a handful of library functions that give the just loaded software a way to talk to the network itself (to load even more software, for example).

Most modern network cards and BIOSes support PXE and booting from the network. If this is not the case the nice people over at etherboot.org have a large library of network card specific code that can be booted via a floppy disk or a bootable CDROM which will provide the network card with the appropriate capabilities.

A dedicated network will be used for the installation, namely 10.200.200.0/24. The install server has the IP 10.200.200.1.

h3. Preparing the tftpboot directory

The TFTP server will serve it's files from the /tftpboot directory. The following structure is expected there:

/tftpboot/
|-- mboot.c32
|-- pxelinux.0
|-- pxelinux.cfg
|   `-- default
`-- solaris
    |-- platform
    |   `-- i86pc
    |       `-- kernel
    |           `-- unix
    `-- x86.miniroot

The mboot.c32 and pxelinux.0 files come from the SYSLINUX bootloader package. pxelinux.cfg/default is the configuration file for the PXE bootloader. It is loaded and read when pxelinux.0 runs. It has the following content:

DEFAULT jumpstart
LABEL jumpstart
  KERNEL mboot.c32
  APPEND -solaris solaris/platform/i86pc/kernel/unix -v -m verbose -B install_media=10.200.200.1:/jumpstart  --- solaris/x86.miniroot

The parameters describe the path of the kernel image under the TFTP server root (solaris/platform/i86pc/kernel/unix), request a verbose startup (-v -m verbose), select the install method, server and path (-B install_media=10.200.200.1:/jumpstart) and the describe the path to the miniroot which contains the installer (solaris/x86.miniroot).

The two files in the solaris directory are found on the Solaris install media under the /boot directory in a similar directory structure. It is important that this structure is retained (even if it seems a bit pointless).

h3. Preparing the DHCP server

There is not much to this, really, all that is required (besides the obvious IP address and netmask) is the IP address of the TFTP server and the filename of the SYSLINUX bootloader in the TFTP directory structure. The complete config file (for the ISC DHCP server) looks like this:

subnet 10.200.200.0 netmask 255.255.255.0 {
    range 10.200.200.128 10.200.200.200;
    option routers 10.200.200.1;
    option subnet-mask 255.255.255.0;
    next-server 10.200.200.1;
    filename "/pxelinux.0";
}

h3. Preparing the NFS server

There are several ways to present the contents of the install media via NFS, but for this install the method that worked best for me was to simply mount the ISO into a directory and share that via NFS.

# mkdir /jumpstart
# mount -o ro,loop /tmp/nv105.iso /jumpstart

The entry in /etc/exports looks like this:

/jumpstart      *(ro,no_subtree_check,sec=sys)

h3. Putting it all together

That should be it, basically. If the machine is turned on and set to boot from the network the following chain of events will take place, provided all goes well:

• The PXE BIOS sends DHCP requests
• The DHCP server answers the requests, assigning an IP address and the TFTP server information
• The PXE BIOS loads /pxelinux.0 from the TFTP server and executes it
• /pxelinux.0 loads /pxelinux.cfg/default from the TFTP server and interpets it as a config file
• /pxelinux.0 loads the Solaris kernel and miniroot from the TFTP server and runs the kernel
• The Solaris kernel boots, mounts the miniroot and starts the installer and re-requests the DHCP address
• The installer mounts the NFS directory

After the hardware was assembled I made a quick attempt to boot Nevara 105 from a DVD in order to see how things went (the installation proper will be made via the network).

The system booted, but was very slow. It took over 10 minutes to get to the first prompt (which asks about the kind of installation you want to perform, and which is usually reached in a few seconds). Older releases and OpenSolaris 08/11 behaved the same. A Linux system booted from an USB stick behaved normally, though.

I wrestled with this for two days, but then noticed something while running memtest86+ on the system. Since the system has 4GB of RAM, and quite a lot of the physical address space between 3 and 4 GB is used by PCI devices, quite a lot of physical memory is remapped to physical addresses above the 4GB mark. memtest shows the start and end address of the block it is currently testing (this is why start and end addresses above 4GB can show up there, even if the system has less than 4GB of RAM). While testing the relocated memory block memtest slowed to a crawl, while the memory below 4GB was tested at normal speed. It looks like memory accesses above 4GB are not covered by the processor cache. Linux seems to put it's kernel below this magic mark, and thus runs normally, while Solaris lands above it, and is less than usable.

h3. Memory layout

In order to explain what is going on here (and why it is bad) a small detour is in order.

Physical address space is a shared resource on most architectures, the Intel x86 platform (in 32 and 64 bit) included. It is shared between real, physical memory and IO memory. Physical memory is the kind that comes (usually and these days) in DIMM memory sticks that go into the appropriate slots on the main board. IO memory, on the other hand, is a way of talking to extension cards, for example network adapters and graphics cards.

These adapters register one or more areas of memory with the BIOS during system startup. Accessing these adresses results in reading or writing to these extension cards instead of real memory.

The use of this is that programs can treat extension cards just like normal memory.

The end result of this is that a given physical memory address can have one of three "backgrounds": * A physical memory cell * A device * Nothing

The memory ranges that devices register usually live between the 3GB and 4GB physical address space. This was all well and good as long as practically no system had that much real memory, so there was no contention for address space.

However, two things happened in the last years: memory got incredibly cheap, and devices got more hungry for memory space. Modern graphics cards, for example, map a large chunk of their on board graphics RAM into the memory space, sometimes all of it, easily taking up half a gigabyte or more of address space.

There are two ways to handle address conflicts in this situation.

The easy way is to simply ignore the physical memory in the address ranges claimed by devices. The physical memory cell becomes inaccessible, and the storage it provides is lost. This is obviously not a popular solution.

The other way is to relocate the physical memory from the contended address spaces into non-contended space. This usually means "above the 4GB border". So although you may only have 4GB of physical RAM in the system some of it must be accessed at addresses above 4GB.

The exact layout of the memory is passed from the BIOS to the system in the so-called E820 memory maps.

On the MSI IM-GM45 board with 4GB of memory installed it looks like this:

BIOS-e820: 0000000000000000 - 0000000000099000 (usable)
BIOS-e820: 0000000000099000 - 00000000000a0000 (reserved)
BIOS-e820: 00000000000e0000 - 0000000000100000 (reserved)
BIOS-e820: 0000000000100000 - 00000000bdc80000 (usable)
BIOS-e820: 00000000bdc80000 - 00000000bdc8e000 (ACPI data)
BIOS-e820: 00000000bdc8e000 - 00000000bdcd0000 (ACPI NVS)
BIOS-e820: 00000000bdcd0000 - 00000000bdce0000 (reserved)
BIOS-e820: 00000000bdcec000 - 00000000bde00000 (reserved)
BIOS-e820: 00000000fee00000 - 00000000fee01000 (reserved)
BIOS-e820: 00000000ffb00000 - 0000000100000000 (reserved)
BIOS-e820: 0000000100000000 - 000000013c000000 (usable)

The lines marked with (usable) represent physical memory, the others are memory ranges reserved for other purposes.

The last line shows the physical memory that has been relocated above the 4GB memory barrier. Almost one gigabyte of physical memory has been relocated.

h3. Processor caches

These days there are multiple layers of caches between the CPU core and the main memory. Data read from memory is held in the caches, for it might be needed again soon, and data written to memory is held for the same reasons.

Data written to memory addresses beloning to devices, however, may or may not be eligible to caching. While memory is expected to keep it's content stable (unless explicitly written to) devices may change the content of their memory maps as they see fit, and caching the values read or written would screw with the CPUs world view.

CPUs therefore contain a list of memory ranges and the cache policies associated with those ranges. These lists are called memory type range registers (MTRR) on Intel CPUs. Below is the list from a different system, also with 4GB of physical memory:

reg00: base=0x100000000 (4096MB), size= 512MB: write-back, count=1
reg01: base=0x120000000 (4608MB), size= 256MB: write-back, count=1
reg02: base=0x00000000 (   0MB), size=2048MB: write-back, count=1
reg03: base=0x80000000 (2048MB), size=1024MB: write-back, count=1
reg04: base=0xc0000000 (3072MB), size= 256MB: write-back, count=1
reg05: base=0xcff00000 (3327MB), size=   1MB: uncachable, count=1

Three quarters of the memory range between 3 and 4GB do not appear on this list, and are thus considered uncacheable by the CPU. The memory above 4GB, however, is marked as cacheable (reg00 and reg01).

Now the same list on the MSI system:

reg00: base=0x00000000 (   0MB), size=4096MB: write-back, count=1

This is wrong in several ways, but the most important one is that the physical memory above 4GB is no longer cached by the CPU. The result of that is access to that memory becomes painfully slow.

This also explains why the problem goes away when only 2GB of memory are present: no physisical memory is relocated beyond the 4GB barrier, because not enough address contentions exist, so all physical memory is being cached again.

I have filed a support request with MSI and hope someone there understands the problem.

h3. [Update 1]

After a short discussion about supported operating systems, MSI has sent me a newer BIOS file. I'll try that this evening.

h3. [Update 2]

The new BIOS does indeed fix this issue. The new MTRRs look like this:

reg00: base=0x13c000000 (5056MB), size=  64MB: uncachable, count=1
reg01: base=0x00000000 (   0MB), size=4096MB: write-back, count=1
reg02: base=0x100000000 (4096MB), size=1024MB: write-back, count=1
reg03: base=0xc0000000 (3072MB), size=1024MB: uncachable, count=1
reg04: base=0xbdd00000 (3037MB), size=   1MB: uncachable, count=1
reg05: base=0xbde00000 (3038MB), size=   2MB: uncachable, count=1
reg06: base=0xbe000000 (3040MB), size=  32MB: uncachable, count=1

There are still some weird edges about this, but by and large it does what it should: all memory (minus that for the onborad graphics) is cached now.

My thanks to the MSI support for the fast (and working) response.

If you ever tried to take a look around an OSX install DVD unter Linux you may have been surprised by the marvelous compression algorithms that Apple uses, because there seems to be preciously little data on those disks:

$ du -sh /media/OSX86DVD/
132K    /media/OSX86DVD/

There is no magic here, as seen by the fact that the same disk inserted into a running OSX system shows a completely different file structure with more than 4GB of data.

The explanation for this is that there are two filesystems on the DVD:

First, a normal ISO9660 CD file system, which contains some bootloader files, and not much else. This filesystem is the one that Linux (and Windows) see by default.

Behind the ISO9660 filesystem there is another one, which spans the rest of the disk. This is a complete hard disk image with it's own partition table, and which contains the real installer data. As Linux does not expect this it does not try to access this filesystem, and the files remain invisible.

In order to get at the files in the second part of the disk some command line magic is necessary. You will need root privileges for the following operations. I assume that the DVD device is /dev/sr0, adjust as neccesary.

The Linux kernel has a nice block device mapping layer which allows us to take a slice out of an existing block device and present this slice as a new device. Even more, there is a helper tool that inspects a block device, looks for partitions and automatically creates such a mapping for each partition. This tool is called kpartx.

So we use kpartx to inspect the DVD device. The command for this is as follows:

# kpartx -a /dev/sr0
device-mapper: reload ioctl failed: Invalid argument
create/reload failed on sr0p1
device-mapper: reload ioctl failed: Invalid argument
create/reload failed on sr0p2
#

That did not work too well. For some reason the device mapper does not like CDROM/DVD devices. So we'll have to get a bit inventive. We create a loop device which is backed by the DVD device, and use that.

# /sbin/losetup /dev/loop0 /dev/sr0
# /sbin/kpartx -av /dev/loop0 
add map loop0p1 (253:5): 0 63 linear /dev/loop0 1
add map loop0p2 (253:6): 0 9178688 linear /dev/loop0 448
#

The command found and created two partitions. The first partition is the ISO9660 file system at the beginning of the disk, the second one is the partition we are interested in. Let's look what's in it.

# file -s /dev/mapper/loop0p2
/dev/mapper/loop0p2: Macintosh HFS Extended version 4 data last mounted by: '10.0', created: 
Tue Oct 30 00:32:01 2007, last modified: Wed Dec 19 16:45:14 2007, last checked: Tue Oct 30 00:32:01
2007, block size: 4096, number of blocks: 1147336, free blocks: 192685
#

That looks good. We could now mount the filesystem and look around.

# mount /dev/mapper/loop0p2 /mnt
# ls /mnt
Applications  Desktop DB  dev  Install Mac OS X.app  mach_kernel  sbin    tmp  vanilla  Volumes
bin           Desktop DF  etc  Library               private      System  usr  var
#

To get rid of all this again we have to unmount the file system, destroy the mappings and release the loop device:

# umount /mnt
# kpartx -d /dev/loop0
# losetup -d /dev/loop0