AUTHOR: Matthias S. Benkmann
DATE: 2007-10-20 LICENSE: Creative Commons Attribution-Share Alike 3.0 (http://creativecommons.org/licenses/by-sa/3.0/) SYNOPSIS: More Control and Package Management using Package Users (v1.4) DESCRIPTION: -You want to know which packages your files belong to ? -You want to deinstall software that doesn't have make uninstall ? -You are bothered by programs installed setuid root behind your back ? -You don't like packages quietly overwriting files from other packages ? -You don't like package managers like RPM ? -YOU WANT TOTAL CONTROL USING ONLY UNIX BUILTINS ? ATTACHMENTS: http://www.linuxfromscratch.org/hints/downloads/attachments/more_control_and_pkg_man/more_control_helpers.tar.bz2 PREREQUISITES: For use with LFS book 6.2: Brain. For use with LFS book later than 6.2: Brain (awake, good working condition). HINT: ########################################################################### Contents ########################################################################### 1. Preface 2. Overview - PART 1: General Information - 3. Package Users 3.1 Introduction 3.2 User Name 3.3 Groups 3.4 Home Directory 4. Common Problems 4.1 Introduction 4.2 General Procedure 4.3 Permission Changes 4.4 Ownership Changes 4.5 Write to Non-Install Directory 4.6 Delete or Overwrite File 4.7 /sbin/ldconfig 4.8 What Commands to Run as a Package User 5. The more_control_helpers Archive 5.1 Overview 5.2 The Wrappers 5.3 add_package_user/install_package 5.4 forall_direntries_from 5.5 uninstall_package 5.6 list_suspicious_files/list_suspicious_files_from 5.7 list_package 5.8 grep_all_regular_files_for 5.9 The etc Directory 5.10 Temporary Files - PART 2: LFS Specifics - 6. Pre-Chroot Phase (Chapter 5) 7. Chroot Phase (Chapter 6) 7.1 Preparations 7.2 Walkthrough: Installing linux-libc-headers 7.3 Known Issues with LFS Packages 8. Sanity Checks 8.1 Suspicious Files 8.2 References to Temporary Files - APPENDICES - A. Security Issues A.1 NFS A.2 Daemons B. Package Categories C. Acknowledgements and Changelog ########################################################################### 1. Preface ########################################################################### Let's say I have written a program that you would like to use. To make it easier for you I come over to install it for you. Would you give me the root account and then leave the room ? No ? Then why do you give it to complete strangers who you have never seen in your life, to install software packages pulled from some Internet server, that come with no warranty and don't even list their contents in the README, although they will happily spread them all over your system ? It is a mystery why Unix admins who wouldn't even trust their employer with more than a normal user account carelessly execute complex and incomprehensible installation scripts with full root rights. Users and groups are the basic security principle in a Unix system. They have been used successfully for a long time to monitor who has created a file and to control who is allowed to delete or change it. But this control has only been imposed on the files of ordinary users. What a waste! I suggest to extend this control to all system files. The general idea is to create package users, i.e. user accounts with restricted rights, to build and install software packages, rather than doing these tasks as root. Not only does this give you more control over what build and install scripts may or may not do, it can also serve as a quite useful package management system. ############################################################################# 2. Overview ############################################################################# This hint is divided into 3 parts. The first part contains general information about the package user method. This part is the most important part of the hint. Read it thoroughly. The second part explains how to apply the package user method to the building of an LFS system. Finally, part 3 of this hint is the Appendix with information that would not fit anywhere else or that is not of general interest. It is inevitable that part 2 will become outdated with time as the LFS book changes and new versions of the software packages used with LFS are released. I make no attempt to track these changes. When someone reports an issue with a package I will incorporate it into the hint, but larger changes that might be required due to changes in the LFS build methodology could take a long time to get included. The reason for this (aside from lack of time) is that I consider part 2 as bonus material that helps people get started but is not essential. Part 1 describes the concepts, which are independent of package versions or the LFS book, and you will have to rely on this information whenever part 2 fails. Don't forget that part 2 only deals with the packages used by the LFS book. For all the other packages you install on your system after that even an up-to-date part 2 would offer no aid anyway. The previous paragraph might sound discouraging, and as you read more from the hint it is possible that you get the impression that the package user method is complicated, causes lots of difficult problems and is overall too much trouble for anyone but a real hardcore admin with programming experience. But you would be mistaken. First of all, many things experienced as installation problems when working with the package user system are in fact desirable features. If `make install' fails for some package, because it attempts to install a file with the same name as a pre-existing file from another package, you should not curse the fact that you have to spend additional time to resolve this issue. Instead you should be happy that you have been alerted of this collision that, had it gone unnoticed, could have messed up your system in more or less subtle ways. Secondly, the package user system is not an all-or-nothing approach. It works on a per-package basis. If a package gives you too much trouble, you can always decide to chicken out and finish the installation as root. Finally, the more_control_helpers archive provided with this hint contains several useful scripts that automate many aspects of software installation as a package user and, together with the tips given in this hint, add a lot of value to the package user system. So do not pass judgement until you have read at least the complete part 1, including the description of the more_control_helpers. ---------------------- PART 1: General Information -------------------------- ############################################################################# 3. Package Users ############################################################################# 3.1 Introduction ---------------- The basic idea of this scheme is easily explained. Every package belongs to a certain "package user". When you install a package, you build and install the package as this package user, causing all files that are installed to be owned by the package user. As a consequence all the usual package management tasks can be comfortably achieved through the use of standard command line utilities. A simple `ls -l ' will tell you, for instance, what package belongs to and a `find -user ...' command allows you to perform an operation on all the files belonging to a certain package, e.g. delete them to uninstall the package. But package management is not all that package users are good for. Because package users do not have root-rights, the installation of a package is limited in what it can do. One thing that a package user is not allowed to do, for example, is to overwrite files from a different package user. Clashes between different packages that want to install a binary, library or header file of the same name are more common than you might think. With package users you never run the risk of package B's installation destroying files from package A silently without you noticing. Every attempt of doing this during package B's installation will cause a "Permission denied" or "Operation not permitted" error so that you have the chance of taking appropriate steps. Another thing that package users are not allowed to do is install setuid root binaries. The decision to make a binary setuid root is also something that a prudent admin does not want to leave up to the creator of a software package. Usually package user accounts have no valid password so that only root can su to a package user, which ensures that package users do not open an additional way into the system and undermine security. But you *may* set passwords anyway to allow a co-admin who you want to be able to install and maintain certain software packages to do so without having access to the actual root account. This co-admin could for instance install, delete, change additional libraries which might be necessary for his workgroup. He would be unable, though, to remove or modify libraries which don't belong to him/her, such as libc. 3.2 User Name ------------- You don't need to drive yourself nuts trying to come up with 8 character names for the package users. I always use the name of the package without the version number, including dashes and possibly exceeding 8 characters in length, e.g. "util-linux", and in the several years that I've been using this scheme I have not encountered any problems, nor has anyone else reported trouble. The 8-character limit on user names seems to be a thing of the past. TIP: You can use bash's programmable completion feature to save yourself some typing when entering commands that take a user name as an argument, such as su, finger or pinky. The command complete -o default -o nospace -A user su finger pinky tells bash to tab-complete words as user names for the commands su, finger and pinky. With this in place you can simply type `su linux-li' and bash will complete this to `su linux-libc-headers' (assuming that you have a package user named "linux-libc-headers"). "-o default" tells bash that if a suitable user name does not exist, the default completion shall be attempted. "-o nospace" prevents the addition of a space after the completed word. This is a very useful command to put into root's .bashrc and .bash_profile. BTW, at http://freshmeat.net/projects/bashcompletion/ you can find a project that offers sophisticated completions for many other commands. Or switch to zsh (http://freshmeat.net/projects/zsh/). It's more powerful and less buggy than bash. 3.3 Groups ---------- Every package user belongs to at least 2 groups. One of these groups is the "install" group, which all package users (and only package users) belong to. All directories that packages are allowed to install stuff in belong to the install group. This includes directories such as /bin and /usr/bin but excludes directories like /root or /. The directories owned by the install group are always group-writable. This would be enough for the package management aspects, but without further preparation this would not give added security or control because every package could replace the files from a different package (the change would be visible in the output from `ls -l', though). For this reason all install directories get the sticky attribute. This allows users to create new files and delete or modify their own files in the directory, but files from other users can not be modified or removed. In the rest of this hint, whenever the term "install directory" is used, it refers to a directory that belongs to group install, is group-writable and sticky. IOW, to turn into an install directory you would do chgrp install && chmod g+w,o+t Although the install group is crucial for the package user system, it is implemented as a supplementary group, rather than as the primary group for package users. This has at least 2 advantages. One advantage is that this makes it easy to get a list of all packages installed on the system with the command grep install /etc/group A more important point, however, is that the primary group is the one that files created by the package user will belong to. So it will be printed in the output of `ls -l' and is subject to find's "-group" test. This makes it very useful for organizational purposes. Following are some suggestions for how to use the primary group. 1. group name = user name Under this scheme the package user for the bash package would be bash:bash. `ls -l /bin/bash' would show something like this -rwxr-xr-x 1 bash bash 1731859 Feb 30 2005 /bin/bash An important advantage of this scheme is that the user information is not lost when you make a file setuid root, which requires changing the file's owner. Because of this advantage, this scheme is the one recommended by this hint. However, the hint's instructions will work fine if you choose a different scheme. 2. group name = package category Under this scheme, you would have certain package categories, such as games, system, net,... and bash, being a system program, would possibly belong to the system group, so that `ls -l /bin/bash' would show something like this -rwxr-xr-x 1 bash system 1731859 Jul 4 1776 /bin/bash This system is nice, but probably not as useful as #1 above, unless you have a real use for this categorization. For a possible categorization see Appendix B at the end of this hint. 3. group name = identifier for a real group of people Under this scheme, the group would correspond to a real group of people in meatspace, e.g. the group of admins responsible for the package. If you need something like this you'll know best what it looks like and how to implement it, so no further discussion of this method will be given here. 3.4 Home Directory ------------------ Although it is well possible not to have a valid home directory for package users or to have just one home directory shared by all package users, that would be a wasted opportunity. Having individual home directories for the package users offers a nice way to organize tarballs, patches, build scripts, notes and all the other per-package information that you accumulate with time. I suggest to use the home directory /usr/src/ for a package user called with the contents detailed below. The more_control_helpers archive contains scripts and skeleton files that implement this suggestion. .bash_profile: You will usually want to have the same environment for all package users, so it is a good idea to make .bash_profile a symbolic link to a file in a central location. The more_control_helpers example uses /etc/pkgusr/bash_profile for this purpose. .bashrc: As for .bash_profile a symlink is a good idea for .bashrc. The more_control_helpers example uses /etc/pkgusr/bashrc as link target. Under normal circumstances package users are not (and even can not be) used for logging into the system, so there is little reason to distinguish between login and non-login shells for package users. Therefore, the example bashrc from more_control_helpers simply sources .bash_profile. This makes sure that the same environment will be used, regardless of whether `su ' or `su - ' is used to become the package user. .project: The contents of this file are printed by the commands `finger -l ' and 'pinky -l ' so .project is a good place for putting information about a package. You should keep the contents of the .project files for your package users up-to-date. source code: The package user's home directory is the perfect place for storing a package's source code. This includes tarballs for different versions, CVS checkouts, unpacked source trees for building,... build script(s): Package user installations require more careful examination of build and install messages than installations done as root, because of the package user-specific problems that can occur. Therefore it is unwise to simply copy'n'paste installation instructions from the LFS book. Build scripts allow you to use sophisticated output redirection for logging purposes that is impractical for direct entry on the command line. The build script skeleton included in the more_control_helpers archive demonstrates this. ############################################################################ 4. Common Problems ############################################################################ 4.1 Introduction ---------------- Software installation is the crux of the package user system. Because installation scripts are often written under the assumption that they will be executed as root, they sometimes fail when executed by a package user. Once this hurdle is passed and a package has been installed, there's usually no difference to a root-installation. A few programs insist that certain security-sensitive files be owned by root and will not execute otherwise, but this is the rare exception. This chapter presents some more or less common problems that you will encounter when using package user accounts to install software, together with guidelines on how to deal with these issues. Although I've said it before I will say it again: Many of the problems you encounter during a package user installation are desirable features of the package user system. You want installation to fail rather than have potentially dangerous actions performed behind your back with root rights. 4.2 General Procedure --------------------- When an installation fails it is almost always due to a "Permission denied" or "Operation not permitted" error while executing a command during `make install'. The first thing you have to do is identify the command that is causing the problem. Usually you will find this in the make output right before the error message. Once you have identified the culprit, you have to decide whether the action that is attempted is illegitimate, partially legitimate or completely legitimate. Illegitimate commands can simply be removed from the Makefile. The other 2 possibilities are more difficult to deal with. You either have to change the condition that makes the command fail or you have to change or sometimes remove the command and make a note if your change suppresses a legitimate action. After you've made changes to solve a certain problem, you reattempt the installation and solve any remaining problems until the installation succeeds. Once you've reached that point it is time to perform any remaining legitimate actions that you've had to disable, such as make certain binaries setuid root. Note that often Makefiles are generated during the configure step, sometimes even later in the build process. If you want to apply changes before the configure step you will usually have to edit files called "Makefile.in". 4.3 Permission Changes ---------------------- Some unsophisticated build systems that don't use the mkinstalldirs script to create installation target directories are very poorly written. Instead of testing whether a target directory exists, they simply attempt to create it with default permissions. This problem usually manifests as a line such as "install -d $(prefix)/bin" in the Makefile. In the common case where prefix=/usr this would attempt to create the /usr/bin directory. If the target directory already exists, as in this case, install will attempt to change its permissions to the default permissions (or those passed on the command line). Of course a package user is not allowed to change the permissions of /usr/bin and so the command fails with a message like "install: cannot change permissions of `/usr/bin': Operation not permitted" This is an example of a completely illegitimate command. Just remove it from the Makefile and everything's fine. 4.4 Ownership Changes --------------------- The most common situation when a package wants to change the ownership of files during installation is when it wants to install setuid root binaries. A common command to do this would be something like "install -c -m 4755 -o root name /usr/bin/name" and the error message would look like this: "install: cannot change ownership of `name': Operation not permitted" The change of ownership is hidden in the "-o root" switch to install, which tells it to make the target file owned by root. The command is at least partially legitimate, because you probably want the binary to be installed. Whether you actually want it to be setuid root is a different matter. The fact that a binary is commonly installed as setuid root doesn't mean that you should make it so. You'll have to ask yourself if normal users absolutely need to execute that binary. If you think they can live without it you're better off not making it setuid root, because every setuid root binary is a potential security hole. In any case you will have to edit the Makefile and remove the offending switch, "-o root" in this case, so that the installation can succeed. Note that this will cause the binary to be installed setuid , which of course makes no sense at all. If you don't intend to make the binary setuid root after the installation, it is best to change the "-m 4755" to "-m 755", so that it won't be installed setuid at all. TIP: When you make a binary setuid root after the installation, use `chown root /usr/bin/name' and not `chown root:root /usr/bin/name'. This way you can keep original group of the file (i.e. the group of the package user) intact. With the user name = group name scheme recommended for package users this makes sure that you can identify the source package of the binary even after making it setuid root. Note that as a security measure chown resets the setuid bit, so you will have to do `chmod u+s /usr/bin/name' after the chown. 4.5 Write to Non-Install Directory ---------------------------------- Sometimes packages want to create files or directories in non-install directories. 3 situations have to be distinguished in this case. The 1st possibility is that the target directory should be an install directory. An example of this is /usr/share/aclocal. This directory is not among the standard system directories created when building an LFS system. It will be created by the first package that has files to install there and will be owned by the corresponding package user. The next package that wants to write in it will fail to install. The remedy is simple. Just make the directory an install directory. You don't even need to be root to do it. The package user that owns the directory has the rights to make that change. The 2nd possible reason for a package wanting to write to a non-install directory is that the failing command is only partially legitimate, i.e. you do want to have installed whatever it is meant to install, but you want it in a different location. For example some packages install binaries that are not meant to be called directly. The default location for these binaries is sometimes called libexec and with prefix=/usr the package will attempt to create /usr/libexec. In cases such as this you often don't have to change any Makefiles. There is either a configure switch to change the directory in question or it is just a matter of overriding a Makefile variable as in `make libexecdir=/usr/lib install'. The 3rd possible reason for an attempt to write to a non-install directory is that the command in question is illegitimate, i.e. you don't want to have installed whatever the package wants to install. As usual with illegitimate commands you can edit the Makefile and just remove them. In the case of a whole directory whose installation you want to suppress it could be too much effort to remove all of the offending commands that want to install files there. In this case an approach similar to that from the previous paragraph can be more effective. Either through configure switches or overriding of variables you change the directory in question to something like /foobar, where is the directory in which build commands are run (i.e. usually the top of the unpackaged source tree). This will cause the package to create the unwanted directory inside the build tree, which doesn't cause any permission problems and has the nice side effect that it'll be deleted together with the build directory when you clean up after the build. 4.6 Delete or Overwrite File ---------------------------- In a perfect world one package should not mess with another package's files, but in the real world conflicts do happen occasionally. While a normal sysadmin installing as root won't notice this until it's too late, an admin employing the package user system will have to deal with conflicts right away. When a package tries to overwrite or delete a file or directory that is owned by another package the attempt will fail. It will fail even inside install directories because of the sticky bit. Although sometimes difficult to implement, the solution to such a conflict is easy to describe. You need to either remove (or rename) the old file or directory before installing, or suppress the installation of the new file or directory. The installation of individual binaries is sometimes easy to prevent. If you find a line such as "PROGRAMS=foo bar fubar barfu" in the Makefile and "foo" is the name of the conflicting binary, just try removing it from that list. That may be sufficient to prevent it from being installed. 4.7 /sbin/ldconfig ------------------ Packages that install libraries sometimes run /sbin/ldconfig as part of their installation so that the dynamic libraries are properly registered on the system. Because a package user is not allowed to overwrite /etc/ld.so.cache ldconfig fails. This failure is commonly ignored in Makefiles, but you should take note of it anyway, because you need to run ldconfig as root after the installation. Alternatively, the more_control_helpers contain a wrapper program that calls /sbin/ldconfig and can be made setuid root. 4.8 What Commands to Run as a Package User ------------------------------------------ A common problem that new users of this hint have is to decide which commands to run as a package user and which commands to run as root. The general rule is that the only commands to run as a package user are those for building, installing, removing and modifying the files that belong to *that* package user's package. Everything else should be run as root as usual. Some things you CAN/SHOULD NOT DO as a package user include - starting daemons - running udevstart - stripping /lib/* ############################################################################ 5. The more_control_helpers Archive ############################################################################ 5.1 Overview ------------ The more_control_helpers archive contains files to help you with building and maintaining a system that uses the package user method. One thing that the more_control_helpers archive contains are some LFS-specific temporary files that are only needed for the building of your LFS system and will not remain installed in a permanent location. Then there are the previously mentioned example files that demonstrate the suggested use of the package user home directories discussed earlier. Another group of files contained in the archive is a set of scripts that help with package management aspects, such as creating new package users and checking which files a particular package has installed. Finally the more_control_helpers archive contains wrapper scripts for some commands that handle many of the common problems discussed in the previous chapter and make package user installations a lot easier. 5.2 The Wrappers ---------------- The previous chapter discussed some common problems encountered during package user builds and how to solve them. The solution to an installation failure usually requires editing of one or more Makefiles. Making such changes manually is annoying, even if it happens only occasionally, and whenever you reinstall a package you have to make the changes again. Sed scripts and patches can help with the latter problem, but they still have to be custom fitted to every package that needs them. There is a better solution, though. While there exist countless ways to install files, only very few are commonly used by packages. The 5 commands mkdir, chgrp, chown, chmod and install are responsible for most of the problems that arise during an LFS installation. This prompted me to write wrapper scripts for these 5 commands that recognize certain problematic patterns and deal with them automatically. The instructions given in this hint in the LFS-specific part will instruct you to install these wrappers in /usr/lib/pkgusr. If you do that and make sure that this directory is the first entry in the PATH of every package user, then they will save you a lot of time and effort in dealing with recurring issues. Note that if you want to choose a directory other than /usr/lib/pkgusr for the wrappers, you need to be careful. Some configure scripts ignore certain locations. A subdirectory of /etc would not work, for instance, because /etc is one of these locations. It is important that you understand the limitations of the wrapper scripts. They can fix some problems without user intervention, such as turning newly created directories in /usr/share/locale into install directories. But other problems by their very nature require manual intervention. When a program tries to install a setuid root binary, for instance, the wrapper scripts will suppress the attempt to change ownership of an installed file to root. While that allows `make install' to complete without error, it is only a partial solution. The wrapper scripts can not (and should not) take away your responsibility for deciding whether the program in question should be setuid root and to make it so, if necessary. To account for this, the wrapper scripts will output warning lines to standard error that start with "***" whenever they encounter a situation that needs to be reviewed. Following the "***" in the message will be the original command that the installation attempted to perform. You *must* check these "***" lines, examine the affected files or directories and take appropriate action. Because of this it is imperative that you log the messages output during a package installation and check these logs religiously. The `build' script contained in the more_control_helpers archive demonstrates some useful output redirection tricks to be used for this purpose. The following 3 examples shall illustrate what kind of things you will have to look for: Example 1: "*** install -c rsh -o root -m 4775 /usr/bin/rsh" This message is output by the install wrapper during the installation of inetutils. The package wants to install the rsh binary setuid root. The install wrapper removes the "-o root" and changes the "-m 4775" to "-m 755" before passing the command on to the real install program. The important thing here is the "-m 4xxx", because this wants to set the setuid bit. Some install scripts throw in a "-o root" for good measure when installing an otherwise normal binary. In that case it's enough that the install wrapper strips out the "-o root" and you don't have to take further action. But when, as in the case of inetutils, the permissions indicate an attempt to make a binary setuid or setgid, then you will have to investigate. You need to decide if you want rsh to be setuid root and if you decide you do, you need to become root and issue commands like this: chown root /usr/bin/rsh chmod u+s /usr/bin/rsh TIP: Be conservative with making binaries setuid. If you're unsure whether you will ever use a program (as non-root), you probably don't want it to be setuid root. Keep in mind that you can always make the change later when you need it. When you apply this reasoning to rsh, for instance, you'll probably end up not making it setuid root. Example 2: "*** chgrp tty /usr/bin/write" This is output by the chgrp wrapper during the util-linux installation. The util-linux package wants to install the write program as setgid tty, so that it is allowed to access other users' terminals. The chgrp wrapper prevents the changing of the group and the chmod wrapper prevents the setting of the setgid bit. You need to decide if you want the program to be setgid and if you decide in favor of this, do as root chgrp tty /usr/bin/write chmod g+s /usr/bin/write Example 3: "*** install -d -m 755 /sbin" This is also from the util-linux installation. Util-linux, for no good reason, tries to recreate the /sbin directory. The install wrapper prevents this and you don't have to take any further action. 5.3 add_package_user/install_package ------------------------------------ Whenever you install a new package on your system, you first have to create a new user account, possibly create a new group and if you follow the advice from this hint about making productive use of a package user's home directory, you will have to set up that one, too. If you were to do all of this manually, it would be a lot of work. The add_package_user and install_package scripts in the more_control_helpers archive were written to automate this. The install_package script is the one you will normally use to prepare for installing a new package. It takes 3 parameters: the description of the package, the name of the package user account to create and the name of the package user's primary group. So if you use the user=group scheme recommended by this hint and are as creative with your package descriptions as I am, then the command you'll use to prepare for installing package "foo" will be install_package foo foo foo This command does 2 things. First it calls the add_package_user script with the provided name, group and description plus sensible default values for add_package_user's other parameters. Then, after add_package_user has created the package user account, install_package automatically uses the su-command to switch to the newly created account. If the default .bashrc and .bash_profile scripts you use for package users contain the command "cd" as do the examples in the more_control_helpers archive, you will be put right into your package user's home directory so that you can start installing right away. TIP: The install_package script can be called with a single argument that will be used as user name, group name and description. So instead of the above command a simple `install_package foo' would have sufficed. The add_package_user script is responsible for the actual work of creating a new package user account. Given a name, a group name and a description, it will create a new user account with the provided primary group and the install group as supplementary group. The groups will be created if necessary. add_package_user takes several arguments that determine the numeric ranges from which it will pick the new user's UID and the GIDs for groups it needs to create. add_package_user does not only create the package user account. It will set up a home directory for it, too. You can either specify the directory or go with the default, which is /usr/src/, where is the name provided for the new account. If the home directory already exists, its ownership and that of any existing contents will be changed to the new user. If it doesn't exist, it will be created. The contents of /etc/pkgusr/skel-package will be copied into the new package user's home directory (without overwriting pre-existing files). The more_control_helpers archive contains an example of a useful skel-package directory. Note that symlinks are copied as symlinks, so skel-package is the perfect place to put .bashrc and .bash_profile symlinks to a central location that will ensure that all package user accounts have the same environment. This is especially useful to make sure that all package users have the wrappers directory in their PATH. 5.4 forall_direntries_from -------------------------- The forall_direntries_from script is a very useful tool for common package management tasks. It can roughly be described as a shortcut for "find / -user -or -group ", where is the first parameter to forall_direntries_from and are the remaining parameters. However, forall_direntries_from takes care of making sure that only relevant filesystems are scanned and shields you from certain unpleasant surprises such as "Oops, I forgot that -depth negates -prune and have accidentally wiped out my home directory." or "Oops, I forgot to -prune /proc and now I'm getting parity errors on my SCSI bus.". IMPORTANT NOTE: By default the forall_direntries_from script will only scan the / filesystem and will not traverse other filesystems. If you have relevant directories that need to be scanned on other filesystems, you will need to edit the script and add the respective mount point(s) to the fs_to_scan list at the beginning of the script. The most likely candidate for addition is "/usr". Application examples: Example 1: Create a tar.gz archive of all files that belong to , e.g. for installing on another machine without having to recompile it there. forall_direntries_from -fprint0 /tmp/files.lst tar --null -P -czf /tmp/archive.tar.gz --files-from=/tmp/files.lst Example 2: Print out all setuid root binaries installed by . (This only works if you use the user=group scheme.) forall_direntries_from -perm +u+s -print Example 3: List all binaries in /bin and /usr/bin belonging to "me" (i.e. the package user executing the command) in alphabetical order. forall_direntries_from $(whoami) -path "*/bin/*" -printf "%f\n" | sort Example 4: Uninstall . See following section about the uninstall_package script. 5.5 uninstall_package --------------------- The uninstall_package script is basically a forall_direntries_from application example in script form. The command `uninstall_package foo' prints out the forall_direntries_from call that you have to use to delete all the files of package "foo" (except for those in directories that forall_direntries_from is instructed not to scan) together with some explanations. So in order to delete the files from package foo, you would execute `uninstall_package foo' and then copy'n'paste the command it prints to the command line. As a safeguard the forall_direntries_from call has an "echo" in front of the "rm" and "rmdir" calls, so if you execute it, the files will not actually be deleted unless you remove both instances of "echo". It is recommended that you execute the command once with the echos and check the output to make sure that only the files you intend to be deleted are in the list. After you've confirmed that, you can use the shell's history to recall the command, edit out the instances of "echo" and remove the files for real. 5.6 list_suspicious_files/list_suspicious_files_from ---------------------------------------------------- list_suspicious_files looks for filesystem entries that are out of the ordinary in some way and prints a categorized list of them. Things that qualify as suspicious include setuid and setgid binaries, world-writable files, symlinks that are possibly broken, hard links, install directories with unusual permissions and other stuff. You should run this script after you've finished your new LFS system and in regular intervals after that. Investigate the listing closely. TIP: When you check the list of setuid and setgid files, don't forget to look at the actual user or group ownership of the file. It's easy to forget that, especially in the setuid case, because we often equate setuid with setuid root since setuid is seldom used with other user accounts. list_suspicious_files_from takes a user or group name or a UID/GID as an argument and reports suspicious entries only when they are owned by the given user or group. Usually you would not call this script directly but instead use list_package (described in the next section), whose output includes that from list_suspicious_files_from. IMPORTANT NOTE: By default the list_suspicious_files script will only scan the / filesystem and will not traverse other filesystems. If you have relevant directories that need to be scanned on other filesystems, you will need to edit the script and add the respective mount point(s) to the fs_to_scan list at the beginning of the script. The most likely candidate for addition is "/usr". 5.7 list_package ---------------- list_package tells you everything about a package's installed files. In general you will want to execute something like list_package $(whoami) >pkg.lst right after installing a package and you can forget about the chronically inaccurate content listings in the (B)LFS book. The following (shortened) output for util-linux speaks for itself: PS1> list_package util-linux EXECUTABLES (in */bin or */sbin) agetty, arch, blockdev, cal, cfdisk, [...] vidmode(->rdev), whereis, write EXECUTABLES WITH NO MANPAGE (in */bin or */sbin) fsck.cramfs, mkfs.cramfs MANPAGE SUMMARIES OF EXECUTABLES (in */bin or */sbin) agetty: alternative Linux getty arch: print machine architecture blockdev: call block device ioctls from the command line cal: displays a calendar cfdisk: Curses based disk partition table manipulator for Linux chkdupexe: find duplicate executables col: filter reverse line feeds from input [...] swapon: enable/disable devices and files for paging and swapping tailf: follow the growth of a log file tunelp: set various parameters for the lp device ul: do underlining umount: unmount file systems vidmode: query/set image root device, RAM disk size, or video mode whereis: locate the binary, source, and manual page files for a command write: send a message to another user EXTRA MANPAGES /usr/share/man/man5/fstab.5 /usr/share/man/man5/nfs.5 /usr/share/man/man8/sln.8 EXTRA EXECUTABLES (not in */bin or */sbin) /usr/share/misc/getopt/getopt-parse.bash /usr/share/misc/getopt/getopt-parse.tcsh /usr/share/misc/getopt/getopt-test.bash /usr/share/misc/getopt/getopt-test.tcsh ALL FILES /etc/fdprm /sbin/agetty /sbin/blockdev /sbin/cfdisk /sbin/ctrlaltdel /sbin/elvtune /sbin/fdisk /sbin/fsck.cramfs /sbin/fsck.minix /sbin/hwclock /sbin/losetup /sbin/mkfs /sbin/mkfs.bfs [...] /usr/share/man/man8/rootflags.8 /usr/share/man/man8/setfdprm.8 /usr/share/man/man8/setsid.8 /usr/share/man/man8/sfdisk.8 /usr/share/man/man8/sln.8 /usr/share/man/man8/swapoff.8 /usr/share/man/man8/swapon.8 /usr/share/man/man8/tunelp.8 /usr/share/man/man8/umount.8 /usr/share/man/man8/vidmode.8 /usr/share/misc/getopt /usr/share/misc/getopt/getopt-parse.bash /usr/share/misc/getopt/getopt-parse.tcsh /usr/share/misc/getopt/getopt-test.bash /usr/share/misc/getopt/getopt-test.tcsh SETUID FILES -rwsr-xr-x "/usr/bin/mount" root:util-linux -rwsr-xr-x "/usr/bin/umount" root:util-linux SETGID FILES -rwxr-sr-x "/usr/bin/write" util-linux:tty FILES WITH UNUSUAL PERMISSIONS -rwsr-xr-x "/usr/bin/mount" root:util-linux -rwsr-xr-x "/usr/bin/umount" root:util-linux -rwxr-sr-x "/usr/bin/write" util-linux:tty Note: list_package works regardless of the prefix you've installed the package with, so you can for instance configure with --prefix=/opt/package and list_package will work just fine (provided that /opt is on a filesystem configured to be scanned by forall_direntries_from and list_suspicious_files). Note: list_package only considers manpages actually owned by the package to list. It will not consider manpages installed by another package. This means that you may see executables identified as not having a manpage although they do have one courtesy of another package (usually man-pages). 5.8 grep_all_regular_files_for ------------------------------ This script is not really related to the package user system, but because of its similarity to the other scripts I've included it anyway. The sole purpose of this script is to identify files that store references to the build environment, specifically the /tools directory. Such references may point out problems, since the /tools directory is supposed to be transient. Don't forget that results for unstripped binaries and libraries are not reliable, because debugging information often includes references to the build environment. These do not cause trouble (unless you're trying to debug the objects in question after deleting /tools). IMPORTANT NOTE: By default the grep_all_regular_files_for script will only scan the / filesystem and will not traverse other filesystems. If you have relevant directories that need to be scanned on other filesystems, you will need to edit the script and add the respective mount point(s) to the fs_to_scan list at the beginning of the script. The most likely candidate for addition is "/usr". 5.9 The etc Directory --------------------- If you follow the instructions provided in the LFS-specific part of this hint, the contents of the etc directory will be installed in /etc/pkgusr. The directory contains a bashrc and bash_profile for package users that takes care of package user specific details such as putting the wrappers directory at the beginning of the PATH and calling cd, so that `su ' will put you right into the package user's home directory. Also contained in the etc directory is a skel-package directory as used by install_package/add_package_user to populate the home directories of newly created package users. 5.10 ldconfig.c -------------------- A lot of packages contain libraries. Having to manually call /sbin/ldconfig as root after installing these packages can become annoying. It would be much easier if one could grant package users permission to use /sbin/ldconfig. Making ldconfig setuid root would be a simple and effective solution, but there are some pitfalls. First of all it is imperative that ordinary users be prohibited from executing ldconfig with elevated privileges. Otherwise an ordinary user can overwrite and possibly read arbitrary files on the system. This can be prevented by making ldconfig owned by group install and removing the o+x bit from the file mode. While this setup is no less secure than running `make install' as root, one reason why we're using package users is because we don't feel safe doing that. To protect against the (admittedly very theoretical) danger of a malicious package user, the more_control_helpers provide ldconfig.c. The only thing this program does is to call `/sbin/ldconfig -v' with an empty environment. Because it doesn't evaluate any user input and doesn't pass any user-provided data to ldconfig, it can safely be made setuid root. 5.11 Temporary Files -------------------- 3 files in the more_control_helpers archive are only used during the installation of the base LFS system and are not installed permanently. The first of them is the installdirs.lst file that contains a list of directories that should be install directories. The second file is sbin/useradd, which is a very primitive shell script that adds a new entry to /etc/passwd. It allows us to add package users before we have installed shadow, which provides a real useradd. Finally there is groupadd, which is like useradd, only for /etc/group. Both scripts, useradd as well as groupadd, do very little error checking and only support the syntax needed by install_package/add_package_user. So don't try anything funky with them. ------------------------ PART 2: LFS Specifics ------------------------------ ############################################################################# 6. Pre-Chroot Phase (Chapter 6) ############################################################################# Build Chapter 5 explained by the LFS book with the following changes: more_control_helpers: When you have reached the end of Chapter 5, before you begin with Chapter 6 you will need to install the helper scripts in the /tools directory so that they are available once you've entered the chroot environment. Use the following commands to install the more_control_helpers in /tools: cd $LFS/tools && tar xjf /path/to/more_control_helpers.tar.bz2 && cd more_control_helpers && cp ./sbin/* $LFS/sbin ############################################################################# 7. Chroot Phase (Chapter 7) ############################################################################# 7.1 Preparations ---------------- Skip the chapter in which the ownership of installed files is changed. Currently this is "7.2. Changing Ownership". Enter the chroot environment and follow the instructions from the book up to but *not* including the installation of the first package, unless is is libstdc++ Install libstdc++ and then compile the shadow package (currently chapter 8.25) and copy the su binary into its proper location: cp src/su /usr/bin Now install the more_control_helpers files in their proper locations on the new LFS system: cp -a /tools/more_control_helpers/etc /etc/pkgusr && chown -R 0:0 /etc/pkgusr && cp -a /tools/more_control_helpers/lib /usr/lib/pkgusr && chown -R 0:0 /usr/lib/pkgusr && cp /tools/more_control_helpers/bin/* /usr/bin && cp /tools/more_control_helpers/sbin/* /usr/sbin && rm /usr/sbin/{useradd,groupadd} Note that the useradd and groupadd scripts are not installed on the new LFS system. These scripts are just temporary workarounds we will use as long as the real useradd and groupadd are not available. Therefore they should only be in /tools/bin. ATTENTION! If you decide to use a different directory than /usr/lib/pkgusr for the wrappers, you have to be careful, because at least the glibc configure script ignores certain directories when looking for programs. The list of ignored directories for glibc includes, among others, everything that starts with "/etc", "/usr/etc" and "/sbin". Wrappers put into a directory that matches any of these patterns would be ineffective. Now it's time to create the install group: groupadd -g 9999 install The GID 9999 has been chosen because the default range used by add_package_user for package user GIDs starts at 10000. Choose whatever number you like. Once the install group has been created you have to create a temporary user that owns all currently installed tools: useradd useradd -s /bin/bash -g lfs -m -k /dev/null -u 9999 lfs The UID 9999 has been chosen because the default range used by add_package_user for package user UIDs starts at 10000. Choose whatever number you like. This user will be deleted after the install process is finished. Now change the ownership of all currently installed tools and make them group-writeable: chown -R lfs:install /{usr,lib,var,etc,bin,sbin,tools} chmod -R ug=rwx /{usr,lib,var,etc,bin,sbin,tools} case $(uname -m) in x86_64) chown -R lfs:install $LFS/lib64 chmod -R ug=rwx $LFS/lib64 ;; esac After this, turn all the directories that packages will install files in into install directories. To make this easier I have compiled a list of install directories that can be found in the file /tools/more_control_helpers/installdirs.lst. The following command uses this list to assign the necessary directories to the install group. Note that you will get several error messages because of non-existent directories. This is because the list contains some directories not created by LFS. chown 0:9999 $(cat /tools/more_control_helpers/installdirs.lst) To be usable by package users, the directories will have to be group-writable and should be sticky as has been explained in the beginning of this hint. However, for now the sticky bit must not be set, as this would prevent the temporary from being overwritten by the package users. The following command sets the permissions appropriately. You will get the same error messages as for the previous command. chmod ug=rwx,o=rx $(cat /tools/more_control_helpers/installdirs.lst) 7.2 Walkthrough: Installing linux-libc-headers ---------------------------------------------- At this point everything has been set up for creating the first package user. At the time of this writing the first package installed in the LFS book is Linux-Libc-Headers, so this package will serve as an example for how things are done. The command install_package 'Linux Headers' linux-libc-headers linux-libc-headers will create a package user with user and group name linux-libc-headers. If you don't want to use the user=group scheme, change the last argument to the desired group name. The description is arbitrary but needs to meet the requirements for the description field of an /etc/passwd entry. TIP: Remember that you can call install_package with just one argument, if you want user name, group name and description to be the same. The directory /usr/src/linux-libc-headers will be set up as the home directory for the package user, automatically populated with the contents of /etc/pkgusr/skel-package. The install_package command also issues the command `su - linux-libc-headers' to assume the identity of the newly created package user. If you're using the bashrc and bash_profile scripts from the more_control_helpers archive, you will be put straight into the directory /usr/src/linux-libc-headers and your prompt will look like this package linux-libc-headers:/usr/src/linux-libc-headers> to show you that you're working as package user linux-libc-headers and that your current working directory is /usr/src/linux-libc-headers. Use the command echo $PATH to verify that your PATH starts with "/usr/lib/pkgusr", the directory that contains the wrappers, and ends with "/tools/bin". Now everything is prepared for installing the package according to the instructions in the LFS book. Note that at the time of this writing the LFS book tells you to execute a chown command to make sure that the headers are owned by root. This is just because the packager has made a very common mistake when creating the tarball for the headers: He has archived the files with a non-root user/group assignment. When unpacking such a tarball as root, the files end up being owned by some weird user/group combination, which may open a security hole. When you're working as a package user this can not happen and you don't want to chown the headers to root:root, because that would defeat the whole point of installing the headers with a package user. This is one of the small points on which you will have to deviate from the standard LFS instructions when using package users. More package user related issues with the current LFS book can be found in the next section. After you've installed the headers, simply issue the command exit to become root again. Now would be a good time to think about useful customizations for /etc/pkgusr/{bash_profile,bashrc} and/or /etc/pkgusr/skel-package, if you've not already customized them. Once you're satisfied with your setup, install the rest of the packages. The following section will help you with some problems that you will run into. 7.3 Cleaning up --------------- When you reach the end of the chroot phase, you will need to do some manual cleanup of file ownership. List all remaining files owned by the temporary lfs user by using the following command: find / -user lfs Many of the listed files belong to the linux api headers and glibc packages. The following commands correct their ownership: chown -R linux: /usr/include/{asm,asm-generic,drm,linux,misc,mtd,rdma,scsi,\ sound,video,xen} chown -R glibc: /usr/include/{arpa,bits,gnu,net,netash,netatalk,netax25,\ neteconet,netinet,netipx,netiucv,nepacket,netrom,netrose,nfs,protocols,\ rpc,sys,audit,gconv} Use the apropriate package names for your install. TIP: You will need to use chown -h to reassign symlinks. Finally, add the sticky bit to all install directories to complete the chroot stage: chmod o+t $(cat /tools/more_control_helpers/installdirs.lst) 7.4 Known Issues with LFS Packages ---------------------------------- This section has details on the package user related problems you will face when building your LFS system. You should copy the information from this section to the INSTALL NOTES of the relevant .project files for the packages concerned, together with any of your own notes. NOTE: If you're building by an LFS book later than 6.2 it is recommended that you read this complete chapter before you start building any packages. If your LFS version is 6.2 then it's fine to read this section package by package as you progress with your build. linux-libc-headers: At the time of this writing the LFS book tells you to execute a chown command to make sure that the headers are owned by root. This is just because the packager has made a very common mistake when creating the tarball for the headers: He has archived the files with a non-root user/group assignment. When unpacking such a tarball as root, the files end up being owned by some weird user/group combination, which may open a security hole. When you're working as a package user this can not happen and you don't want to chown the headers to root:root, because that would defeat the whole point of installing the headers with a package user. There used to be another packaging error in the linux-libc-headers. Version 2.6.12.0 (current as of this writing) doesn't have it anymore, but older versions used to contain files with permissions set incorrectly. All headers are supposed to be world-readable, but they weren't. More about this later in the glibc notes. man-pages: If the name you use for the man-pages package user is not exactly "man-pages", then you will have to change the variable "manpagesowner" right at the beginning of the wrapper script `install'. Recent versions of man-pages contain POSIX manpages that the package tries to install in /usr/share/man/man{0,1,3}p. There's also a manpage that man-pages wants to install to /usr/share/man/man9. As /usr/share/man is not an install directory and the LFS book does not have instructions to create these directories at the time of this writing, the installation will fail and the respective man-pages will not be installed. Possible remedies: 1. Make /usr/share/man an install directory. Consequence: All Packages will be able to create new subdirectories in /usr/share/man. I find this undesirable because there are packages that create directories for manpages in foreign languages that I don't want. YMMV. 2. Ignore the problem and live without the POSIX manpages. Unless you are a developer (including script writer) who is interested in writing portable programs/scripts this is a good solution. 3. Create the directories /usr/share/man/man{0,1,3}p and man9 as root prior to installing man-pages. You'll have to either chown them to the man-pages package user or make them install directories. This is my preferred solution. glibc: It is kind of unfortunate that the packaging error of libc-linux-headers concerning the permissions doesn't exist in the current version. It provided for a great learning experience. I've kept the following section in the hint for this reason even though it doesn't apply anymore. Please take the time to read it. --------------------- old stuff start ---------------------------------------- Because of the error, the headers in /tools/include are not world-readable. Unfortunately the LFS book (as of this writing) does not correct this in Chapter 5 like it does in Chapter 6. For a standard LFS build this is no problem, because glibc is built as root and root can access everything regardless of permissions. The glibc package user, however, does not have permission to access these headers. This will cause several configure tests to fail, because the respective test programs can not be compiled. The end result is the error message "/lib/cpp fails sanity check", which is completely nonsensical as we don't have a /lib/cpp. This is the perfect opportunity to introduce rule #1 of error diagnostics: NEVER TRUST DIAGNOSTIC MESSAGES ! There are 2 kinds of diagnostic messages: 1. Those that are unnecessary, because once you see which component has failed, the source of the problem is obvious. 2. Those that grossly misdiagnose the source of the problem and lead you to draw the wrong conclusions. No, there is no other kind. Trust me ;-) In this case, /lib/cpp has nothing to do with the problem. It doesn't exist and that's fine. The message just wants to trick you into doing something stupid such as create a symlink /lib/cpp -> /tools/bin/cpp. But that would be totally wrong. Before you jump to any premature conclusions you should always try to get as much *low-level* information as you can. Diagnostic messages are *high-level* information. They represent a filtered view of the problem, which is usually of little help. Fortunately the message (the complete one, not the part quoted above) also points at the source for the necessary low-level information. In this case that is the file config.log (not to be confused with configure.log, the file created by the build script included in the more_control_helpers archive). config.log is created by all autoconf-created configures (not just that of glibc) and it contains, among other things, the test programs used by configure and messages output while building and running them. Whenever a configure script fails or gives weird results, check config.log. And always remember rule #2 of error diagnostics ALWAYS START AT THE FIRST ERROR This seems pretty obvious, but nevertheless people commonly do the exact opposite. It's just too tempting to start at the point of the final failure and try to work backwards. In this case many people would open config.log and scroll to the point of the failed /lib/cpp sanity check. After all, that's what caused configure to abort and so that's what needs to be fixed, right? WRONG! Someone who takes this approach just sees the error message "/lib/cpp: No such file or directory" and is even more convinced that a missing /lib/cpp symlink (or program) is the problem. The correct way to approach such a problem is to start at the beginning of config.log, to scroll down to first error message and to check if it is an issue that needs to be fixed (error messages in config.log are not always signs for a problem). If the issue needs to be fixed, then it needs to be fixed first, because all later errors could be rooted in this issue (even if, no, *especially* if you don't believe this is the case). If we apply this advice to the problem at hand, we quickly get to the first serious error in config.log: "/tools/include/linux/limits.h: Permission denied" A quick check with ls reveals that indeed the directory with the linux headers is not world-readable, which is obviously wrong. The fix is easy. Just make (as root) the header directories /tools/include/{linux,asm} world-readable with commands similar to those the LFS book presents in Chapter 6 for the installation of linux-libc-headers. Once this change has been made, glibc's configure succeeds. --------------------- old stuff end ----------------------------------------- TIP: Even when configure completes successfully, you should still check the output carefully to see if there is anything odd. E.g. if you're using the wrappers, you should check that configure outputs the line checking for a BSD-compatible install... /usr/lib/pkgusr/install -c If configure detects a different install, such as /tools/bin/install, something is wrong. Maybe there's a typo in the PATH for the package user, or you've put the wrappers into a directory that is ignored by configure. With the wrappers the glibc build and install should work smoothly. The wrapper script for install makes sure that the /usr/share/locale/* directories become install directories so that other programs can install their localized messages. One thing that the wrappers do not take care of, however, is the file /usr/share/info/dir. Because in the current LFS build order glibc is the first package that installs info files, dir is owned by and only writable by glibc. In order to allow other packages to install info pages, execute the following commands as root: chown root:install /usr/share/info/dir && chmod ug=rw,o=r /usr/share/info/dir NOTE: glibc wants to install the program pt_chown as setuid root. If you install as a package user, the program will get installed but not given root privileges (because of the install wrapper). The following info is from the glibc docs: One auxiliary program, `/usr/libexec/pt_chown', is installed setuid `root'. This program is invoked by the `grantpt' function; it sets the permissions on a pseudoterminal so it can be used by the calling process. This means programs like `xterm' and `screen' do not have to be setuid to get a pty. (There may be other reasons why they need privileges.) If you are using a 2.1 or newer Linux kernel with the `devptsfs' or `devfs' filesystems providing pty slaves, you don't need this program; otherwise you do. The source for `pt_chown' is in `login/programs/pt_chown.c'. So unless you're building a system that does not use devpts (which would be quite unusual), this does not need to concern you. TIP: In case you were wondering if you should create /etc/nsswitch.conf and /etc/ld.so.conf as root or glibc, I recommend to assign all files that you manually create or manually edit to the root account. That way you can distinguish between those files that can be regenerated automatically and those that can not. Assigning even automatically generated files to root once you make the first manual edit, ensures that a later reinstallation of a package won't silently do away with your manual tweaks. ldconfig: Now that glibc has installed /sbin/ldconfig you can activate the ldconfig wrapper if you want to. Perform the following steps as root AFTER re-adjusting the toolchain, just before starting with binutils: cd /usr/lib/pkgusr gcc -O2 -W -Wall -o ldconfig ldconfig.c chown root:install ldconfig chmod u=rwxs,g=rxs,o= ldconfig These instructions make the ldconfig wrapper setuid root and setgid install and prevent non-root users not in the install group from executing it. The setuid root is required so that it can replace /etc/ld.so.cache. The setgid install is not strictly required, but without it /etc/ld.so.cache will end up with the group of the last package user that touched it. If you use the user name=group name scheme this will cause the more_control_helpers scripts to believe that /etc/ld.so.cache belongs to the package in question which can be confusing. binutils: Have you make /usr/share/info/dir group-writable as explained above in the glibc notes? If you've missed that part, go back and do it now. The installation of binutils should complete without problems. It does however cause minor conflicts with autoconf (see later). NOTE: At the time of this writing the version of bash used in the LFS book has a bug that causes the list_package script to spit out errors and to list all manpages of binutils (and other packages) as Broken. This bug is already fixed by the bash patch used by the book but the patch is not applied in chapter 5. Since the manpage summary functionality of list_package requires man which is not installed until after bash is rebuilt, this doesn't really matter, because while patching the chapter 5 bash would get rid of the error messages, it wouldn't result in usable manpage summaries. gcc: Because the /usr/lib/libgcc_s.so* symlinks created at the beginning of Chapter 6 is owned by root, gcc's installation cannot remove it. So you will have to remove it as root before `make install'. Alternatively use chown -h gcc: /usr/lib/libgcc* to change ownership of the files in question after creating the gcc package user. Note the -h option which has to be used to change ownership of the symlinks themselves rather than their target files. db: It should be obvious that you don't want to change the ownership of the installed files. coreutils: Because the /bin/cat, /bin/pwd and /bin/stty symlinks are owned by root, coreutils' installation cannot remove them. So you will have to remove them manually before `make install'. Alternatively use the command chown -h coreutils: /bin/{cat,pwd,stty} after creating the coreutils package user. Note the -h switch that makes chown change the ownership of the symlinks themselves rather than their target files. The chapter 6 instructions move the coreutils binaries to /bin, including the mv binary itself. You need to make sure that hashing is turned off for this to work. The LFS book does this by putting `set +h' into the LFS user's .bashrc. If you're following this hint, you're likely using build scripts, so you need to put this command into the build script before the mv commands. NOTE: The man-pages package has already installed manpages for the binaries from coreutils. The install wrapper will prevent coreutils from overwriting those. This is done because the manpages from the man-pages package are of superior quality (although not necessarily uptodate). It also prevents errors during `make install' that would otherwise occur because the coreutils package user cannot overwrite manpages owned by another user. If you don't like the above behaviour and would rather have the original package manpages (because they are uptodate), you can set the variable manpagesowner at the beginning of the install wrapper to a string that doesn't correspond to a package user name (it must not be empty, though!). If you do this, you will have to resolve manpage conflicts in another way. The easiest way to handle this is probably to not install the man-pages package at the beginning of Chapter 6 but at the end, after all the other packages have already installed their manpages. Then you need only deal with the conflicts once, when installing man-pages. ncurses: The installation of ncurses (like that of other packages that include libraries) wants to run /sbin/ldconfig to update /etc/ld.so.cache. This fails because the package user doesn't have permission to replace /etc/ld.so.cache. Making /etc/ld.so.cache group-writable by the install group doesn't help, because the permissions would be reset on the next call to /sbin/ldconfig. This error will usually not abort the installation and you can just run /sbin/ldconfig manually as root afterwards. Alternatively you can use the ldconfig wrapper as described earlier. aclocal directory: At the time of this writing the directory /usr/share/aclocal is created during the bison installation. This directory contains macros for autoconf. Other packages want to install files into this directory, so you should make it writable by the install group and sticky. perl: Before you do `make install', you will have to `chown -h perl: /usr/bin/perl' so that the perl package user is allowed to remove the /usr/bin/perl symlink. If you will install add-on packages for perl as their own package users into /usr/lib/perl5/site_perl, then you will need to turn /usr/lib/perl5/site_perl/ and its subdirectories into install directories. You don't need to do this now as you'll notice it anyway when installing a perl add-on fails. autoconf: The autoconf package wants to install its own copy of standards.info, which fails because binutils has already installed this file. You can either ignore the error or remove the binutils version of standards.info before `make install'. bash: configure: The bash configure script tests for the presence of the special devices /dev/std* and /dev/fd/*. Unfortunately at the time of this writing the test for /dev/fd/* is buggy (the test for /dev/stdin used to be broke, too in bash-2.x, but has been fixed since). It ends up testing read access to standard input, which is the (pseudo)terminal you're building your system in. Unfortunately su doesn't change ownership of the terminal device, so when you're su'd to a package user account, the terminal still belongs to the login user. As the package user doesn't have read access to the device, the tests fail. There is a simple way to get around this. Simply run ./configure like this ./configure .... ' !) to refer to /dev/null. Unlike the terminal device, /dev/null is world-readable and world-writable, so the tests succeed as they should. If you don't like this trick, you can also chown the terminal device in question (see `ls -la /dev/fd/0') to the package user before building bash. make check: When running the test suite as a package user, the test "run-test" will fail with the output like this: 33d32 < *** chmod g+s /tmp/test.setgid 35c34 < 1 --- > 0 64d62 < *** chmod u+s /tmp/test.setuid 66c64 < 1 --- > 0 154c152 < 1 --- > 0 160c158 < 1 --- > 0 The first 2 failures are caused by the chmod wrapper which prevents the test from setting the setuid and setgid bits and outputs the *** warning. The failures are harmless. You can get rid of them by removing the wrappers directory from the PATH before running the tests. The last 2 failures are not specific to package users but will occur whenever you run the tests su'd to another user. The reasons are the same as for the configure problem described earlier. The same solutions apply. Either use chown (if you chowned before configure you're already done, of course) or run make check like this make check now (i.e. start a login shell). -build script now handles unpacking of tarballs and allows calling the different stages individually. -useradd uses the -s provided shell and no longer hardwires bash. -chapter 6 bash notes now properly address the configure and make check issues. 2007-03-21 Matthias Benkmann -changed forall_direntries_from to avoid warning message from find when -depth is used. -added 4.8 What Commands to Run as a Package User 2005-12-22 Matthias Benkmann -added advice on how to cope with the moving mv problem to coreutils note. 2005-11-13 Matthias Benkmann -fixed list_suspicious_files and list_package to work with recent more POSIX-conforming versions of GNU find -released version 1.2 2005-01-01 Matthias Benkmann -fixed bug in skel-package/build script that caused it to report all steps as successful, even if they failed -released version 1.1 2004-11-01 Matthias Benkmann -capitalized title -released version 1.0 2004-10-14 Matthias Benkmann -started developing the more_control_helpers utilities 2004-08-14 Matthias Benkmann -started major rewrite (update for new LFS version, new hint format, textual improvements,...) 2002-04-20 Matthias Benkmann -changed LFS VERSION header to be more conservative -added
tags to the synopsis for the sake of the hints index -added group mmedia to the list of suggested groups -submitted v0.8 2002-03-16 Matthias Benkmann -added note, that on Linux make doesn't need to be setgid kmem 2002-02-18 Matthias Benkmann -added section "Security issues with NFS" -submitted v0.7 2002-01-30 Matthias Benkmann -added Changelog -moved "chown 0.10000 `cat /tmp/installdirs`" command up (before glibc package user is created) -add_package_user: create home directory with "mkdir -p" use $grpfile everywhere instead of /etc/group -improved mammoth sentence in Introduction -added note about possibility to have user name==group name -source bashrc_basic in bashrc_package -minor textual changes