diff --git a/en_US.ISO8859-1/articles/portbuild/article.sgml b/en_US.ISO8859-1/articles/portbuild/article.sgml index 11f23d8a40..985ea217f8 100644 --- a/en_US.ISO8859-1/articles/portbuild/article.sgml +++ b/en_US.ISO8859-1/articles/portbuild/article.sgml @@ -1,1886 +1,1887 @@ %articles.ent; ]>
Package Building Procedures The &os; Ports Management Team $FreeBSD$ 2003 2004 2005 2006 2007 2008 2009 The &os; Ports Management Team &tm-attrib.freebsd; &tm-attrib.intel; &tm-attrib.sparc; &tm-attrib.general; Introduction and Conventions In order to provide pre-compiled binaries of third-party applications for &os;, the Ports Collection is regularly built on one of the Package Building Clusters. Currently, the main cluster in use is at . Most of the package building magic occurs under the /var/portbuild directory. Unless otherwise specified, all paths will be relative to this location. ${arch} will be used to specify one of the package architectures (amd64, &i386;, and &sparc64;), and ${branch} will be used to specify the build branch (6, 6-exp, 7, 7-exp, 8, 8-exp). Packages are no longer built for Release 4 or 5, nor for the alpha nor ia64 architectures. The scripts that control all of this live in /var/portbuild/scripts/. These are the checked-out copies from /usr/ports/Tools/portbuild/scripts/. Typically, incremental builds are done that use previous packages as dependendencies; this takes less time, and puts less load on the mirrors. Full builds are usually only done: right after release time, for the -STABLE branches periodically to test changes to -CURRENT for experimental builds Build Client Management The &i386; clients co-located with pointyhat netboot from it (connected nodes); all other clients (disconnected nodes) are either self-hosted or netboot from some other pxe host. In all cases they set themselves up at boot-time to prepare to build packages. The cluster master rsyncs the interesting data (ports and src trees, bindist tarballs, scripts, etc.) to disconnected nodes during the node-setup phase. Then, the disconnected portbuild directory is nullfs-mounted for chroot builds. The ports-${arch} user can &man.ssh.1; to the client nodes to monitor them. Use sudo and check the portbuild.hostname.conf for the user and access details. The scripts/allgohans script can be used to run a command on all of the ${arch} clients. The scripts/checkmachines script is used to monitor the load on all the nodes of the build cluster, and schedule which nodes build which ports. This script is not very robust, and has a tendency to die. It is best to start up this script on the build master (e.g. pointyhat) after boot time using a &man.while.1; loop. Chroot Build Environment Setup Package builds are performed in a chroot populated by the portbuild script using the ${arch}/${branch}/builds/${buildid}/bindist.tar file. The following command builds a world from the ${arch}/${branch}/src tree and installs it into ${worlddir}. The tree will be updated first unless -nocvs is specified. /var/portbuild&prompt.root; scripts/makeworld ${arch} ${branch} ${buildid} [-nocvs] The bindist.tar tarball is created from the previously installed world by the mkbindist script. It should be run as root with the following command: /var/portbuild&prompt.root; scripts/mkbindist ${arch} ${branch} ${buildid} The per-machine tarballs are located in ${arch}/clients. The bindist.tar file is extracted onto each client at client boot time, and at the start of each pass of the dopackages script. Starting the Build Several separate builds for each architecture - branch combination are supported. All data private to a build (ports tree, src tree, packages, distfiles, log files, bindist, Makefile, etc) are located under ${arch}/${branch}/builds/${buildid}. The last created build can be alternatively referenced under buildid latest, the one before is called previous. New builds are cloned from the latest, which is fast since it uses ZFS. <command>dopackages</command> scripts The scripts/dopackages* scripts are used to perform the builds. Most useful are: dopackages.6 - Perform a 6.X build dopackages.6-exp - Perform a 6.X build with experimental patches (6-exp branch) dopackages.7 - Perform a 7.X build dopackages.7-exp - Perform a 7.X build with experimental patches (7-exp branch) dopackages.8 - Perform a 8.X build dopackages.8-exp - Perform a 8.X build with experimental patches (8-exp branch) These are wrappers around dopackages, and are all symlinked to dopackages.wrapper. New branch wrapper scripts can be created by symlinking dopackages.${branch} to dopackages.wrapper. These scripts take a number of arguments. For example: dopackages.6 ${arch} [-options] [-options] may be zero or more of the following: -keep - Do not delete this build in the future, when it would be normally deleted as part of the latest - previous cycle. Don't forget to clean it up manually when you no longer need it. -nofinish - Do not perform post-processing once the build is complete. Useful if you expect that the build will need to be restarted once it finishes. If you use this option, don't forget to cleanup the clients when you don't need the build anymore. -finish - Perform post-processing only. -nocleanup - By default, when the -finish stage of the build is complete, the build data will be deleted from the clients. This option will prevent that. -restart - Restart an interrupted (or non-finished) build from the beginning. Ports that failed on the previous build will be rebuilt. -continue - Restart an interrupted (or non-finished) build. Will not rebuild ports that failed on the previous build. -incremental - Compare the interesting fields of the new INDEX with the previous one, remove packages and log files for the old ports that have changed, and rebuild the rest. This cuts down on build times substantially since unchanged ports do not get rebuilt every time. -cdrom - This package build is intended to end up on a CD-ROM, so NO_CDROM packages and distfiles should be deleted in post-processing. -nobuild - Perform all the preprocessing steps, but do not actually do the package build. -noindex - Do not rebuild INDEX during preprocessing. -noduds - Do not rebuild the duds file (ports that are never built, e.g. those marked IGNORE, NO_PACKAGE, etc.) during preprocessing. -trybroken - Try to build BROKEN ports (off by default because the amd64/&i386; clusters are fast enough now that when doing incremental builds, more time was spent rebuilding things that were going to fail anyway. Conversely, the other clusters are slow enough that it would be a waste of time to try and build BROKEN ports). -nosrc - Do not update the src tree from the ZFS snapshot, keep the tree from previous build instead. -srccvs - Do not update the src tree from the ZFS snapshot, update it with cvs update instead. -noports - Do not update the ports tree from the ZFS snapshot, keep the tree from previous build instead. -portscvs - Do not update the ports tree from the ZFS snapshot, update it with cvs update instead. -norestr - Do not attempt to build RESTRICTED ports. -plistcheck - Make it fatal for ports to leave behind files after deinstallation. -nodistfiles - Do not collect distfiles that pass make checksum for later uploading to ftp-master. -fetch-original - Fetch the distfile from the original MASTER_SITES rather than ftp-master. If the last build finished cleanly you do not need to delete anything. If it was interrupted, or you selected -nocleanup, you need to clean up clients by running build cleanup ${arch} ${branch} ${buildid} -full errors/, logs/, packages/, and so forth, are cleaned by the scripts. If you are short of space, you can also clean out ports/distfiles/. Leave the latest/ directory alone; it is a symlink for the webserver. dosetupnodes is supposed to be run from the dopackages script in the -restart case, but it can be a good idea to run it by hand and then verify that the clients all have the expected job load. Sometimes, dosetupnode cannot clean up a build and you need to do it by hand. (This is a bug.) Make sure the ${arch} build is run as the ports-${arch} user or it will complain loudly. The actual package build itself occurs in two identical phases. The reason for this is that sometimes transient problems (e.g. NFS failures, FTP sites being unreachable, etc.) may halt a build. Doing things in two phases is a workaround for these types of problems. Be careful that ports/Makefile does not specify any empty subdirectories. This is especially important if you are doing an -exp build. If the build process encounters an empty subdirectory, both package build phases will stop short, and an error similar to the following will be written to ${arch}/${branch}/make.[0|1]: don't know how to make dns-all(continuing) To correct this problem, simply comment out or remove the SUBDIR entries that point to empty subdirectories. After doing this, you can restart the build by running the proper dopackages command with the -restart option. This problem also appears if you create a new category Makefile with no SUBDIRs in it. This is probably a bug. Update the i386-6 tree and do a complete build dopackages.6 i386 -nosrc -norestr -nofinish Restart an interrupted amd64-8 build without updating dopackages.8 amd64 -nosrc -noports -norestr -continue -noindex -noduds -nofinish Post-process a completed sparc64-7 tree dopackages.7 sparc64 -finish Hint: it us usually best to run the dopackages command inside of screen(1). <command>build</command> command You may need to manipulate the build data before starting it, especially for experimental builds. This is done with build command. build list arch branch - Shows the current set of build ids. build clone arch branch oldid [newid] - Clones oldid to newid (or a datestamp if not specified). build srcupdate arch branch buildid - Replaces the src tree with a new ZFS snapshot. Don't forget to use -nosrc flag to dopackages later! build portsupdate arch branch buildid - Replaces the ports tree with a new ZFS snapshot. Don't forget to use -noports flag to dopackages later! Building a single package Sometimes there is a need to rebuild a single package from the package set. This can be accomplished with the following invocation: /var/portbuild/evil/qmanager/packagebuild amd64 7-exp 20080904212103 aclock-0.2.3_2 Anatomy of a Build A full build without any -no options performs the following operations in the specified order: An update of the current ports tree from the ZFS snapshot [*] An update of the running branch's src tree from the ZFS snapshot [*] Checks which ports do not have a SUBDIR entry in their respective category's Makefile [*] Creates the duds file, which is a list of ports not to build [*] [+] Generates a fresh INDEX file [*] [+] Sets up the nodes that will be used in the build [*] [+] Builds a list of restricted ports [*] [+] Builds packages (phase 1) [++] Performs another node setup [+] Builds packages (phase 2) [++] [*] Status of these steps can be found in ${arch}/${branch}/build.log as well as on stderr of the tty running the dopackages command. [+] If any of these steps fail, the build will stop cold in its tracks. [++] Status of these steps can be found in ${arch}/${branch}/make.[0|1], where make.0 is the log file used by phase 1 of the package build and make.1 is the log file used by phase 2. Individual ports will write their build logs to ${arch}/${branch}/logs and their error logs to ${arch}/${branch}/errors. Formerly the docs tree was also checked out, however, it has been found to be unnecessary. Interrupting a Build Interrupting a build is a bit messy. First you need to identify the tty in which it's running (either record the output of &man.tty.1; when you start the build, or use ps x to identify it. You need to make sure that nothing else important is running in this tty, e.g. ps -t p1 or whatever. If there is not, you can just kill off the whole term easily with pkill -t pts/1; otherwise issue a kill -HUP in there by, for example, ps -t pts/1 -o pid= | xargs kill -HUP. Replace p1 by whatever the tty is, of course. The package builds dispatched by make to the client machines will clean themselves up after a few minutes (check with ps x until they all go away). If you do not kill &man.make.1;, then it will spawn more jobs. If you do not kill dopackages, then it will restart the entire build. If you do not kill the pdispatch processes, they'll keep going (or respawn) until they've built their package. To free up resources, you will need to clean up client machines by running build cleanup command. For example: &prompt.user; /var/portbuild/scripts/build cleanup i386 6-exp 20080714120411 -full If you forget to do this, then the old build chroots will not be cleaned up for 24 hours, and no new jobs will be dispatched in their place since pointyhat thinks the job slot is still occupied. To check, cat ~/loads/* to display the status of client machines; the first column is the number of jobs it thinks is running, and this should be roughly concordant with the load average. loads is refreshed every 2 minutes. If you do ps x | grep pdispatch and it's less than the number of jobs that loads thinks are in use, you're in trouble. You may have problem with the umount commands hanging. If so, you are going to have to use the allgohans script to run an &man.ssh.1; command across all clients for that buildenv. For example: ssh -l root gohan24 df will get you a df, and allgohans "umount -f pointyhat.freebsd.org:/var/portbuild/i386/6-exp/ports" allgohans "umount -f pointyhat.freebsd.org:/var/portbuild/i386/6-exp/src" are supposed to get rid of the hanging mounts. You will have to keep doing them since there can be multiple mounts. Ignore the following: umount: pointyhat.freebsd.org:/var/portbuild/i386/6-exp/ports: statfs: No such file or directory umount: pointyhat.freebsd.org:/var/portbuild/i386/6-exp/ports: unknown file system umount: Cleanup of /x/tmp/6-exp/chroot/53837/compat/linux/proc failed! /x/tmp/6-exp/chroot/53837/compat/linux/proc: not a file system root directory The former 2 mean that that client did not have those mounted; the latter 2 are a bug. You may also see messages about procfs. After you have done all the above, remove the ${arch}/lock file before trying to restart the build. If you do not, dopackages will simply exit. If you have to do a ports tree update before restarting, you may have to rebuild either duds, INDEX, or both. If you are doing the latter manually, you will also have to rebuild packages/All/Makefile via the makeparallel script. Monitoring the Build You can use qclient command to monitor the status of build nodes, and to list the currently scheduled jobs: python /var/portbuild/evil/qmanager/qclient jobs python /var/portbuild/evil/qmanager/qclient status The scripts/stats ${branch} command shows the number of packages already built. Running cat /var/portbuild/*/loads/* shows the client loads and number of concurrent builds in progress. The files that have been recently updated are the clients that are online; the others are the offline clients. The pdispatch command does the dispatching of work onto the client, and post-processing. ptimeout.host is a watchdog that kills a build after timeouts. So, having 50 pdispatch processes but only 4 &man.ssh.1; processes means 46 pdispatches are idle, waiting to get an idle node. Running tail -f ${arch}/${branch}/build.log shows the overall build progress. If a port build is failing, and it is not immediately obvious from the log as to why, you can preserve the WRKDIR for further analysis. To do this, touch a file called .keep in the port's directory. The next time the cluster tries to build this port, it will tar, compress, and copy the WRKDIR to ${arch}/${branch}/wrkdirs. If you find that the system is looping trying to build the same package over and over again, you may be able to fix the problem by rebuilding the offending package by hand. If all the builds start failing with complaints that they cannot load the dependent packages, check to see that httpd is still running, and restart it if not. Keep an eye on &man.df.1; output. If the /var/portbuild file system becomes full then Bad Things happen. The status of all current builds is generated twice an hour and posted to . For each buildenv, the following is displayed: cvs date is the contents of cvsdone. This is why we recommend that you update cvsdone for -exp runs (see below). date of latest log number of lines in INDEX the number of current build logs the number of completed packages the number of errors the number of duds (shown as skipped) missing shows the difference between INDEX and the other columns. If you have restarted a run after a cvs update, there will likely be duplicates in the packages and error columns, and this column will be meaningless. (The script is naive). running and completed are guesses based on a &man.grep.1; of build.log. Dealing With Build Errors The easiest way to track build failures is to receive the emailed logs and sort them to a folder, so you can maintain a running list of current failures and detect new ones easily. To do this, add an email address to ${branch}/portbuild.conf. You can easily bounce the new ones to maintainers. After a port appears broken on every build combination multiple times, it is time to mark it BROKEN. Two weeks' notification for the maintainers seems fair. To avoid build errors with ports that need to be manually fetched, put the distfiles into ~ftp/pub/FreeBSD/distfiles. Release Builds When building packages for a release, it may be necessary to manually update the ports and src trees to the release tag and use -nocvs and -noportscvs. To build package sets intended for use on a CD-ROM, use the -cdrom option to dopackages. If the disk space is not available on the cluster, use -nodistfiles to avoid collecting distfiles. After the initial build completes, restart the build with -restart -fetch-original to collect updated distfiles as well. Then, once the build is post-processed, take an inventory of the list of files fetched: &prompt.user; cd ${arch}/${branch} &prompt.user; find distfiles > distfiles-${release} This inventory file typically lives in i386/${branch} on the cluster master. This is useful to aid in periodically cleaning out the distfiles from ftp-master. When space gets tight, distfiles from recent releases can be kept while others can be thrown away. Once the distfiles have been uploaded (see below), the final release package set must be created. Just to be on the safe side, run the ${arch}/${branch}/cdrom.sh script by hand to make sure all the CD-ROM restricted packages and distfiles have been pruned. Then, copy the ${arch}/${branch}/packages directory to ${arch}/${branch}/packages-${release}. Once the packages are safely moved off, contact the &a.re; and inform them of the release package location. Remember to coordinate with the &a.re; about the timing and status of the release builds. Uploading Packages Once a build has completed, packages and/or distfiles can be transferred to ftp-master for propagation to the FTP mirror network. If the build was run with -nofinish, then make sure to follow up with dopackages -finish to post-process the packages (removes RESTRICTED and NO_CDROM packages where appropriate, prunes packages not listed in INDEX, removes from INDEX references to packages not built, and generates a CHECKSUM.MD5 summary); and distfiles (moves them from the temporary distfiles/.pbtmp directory into distfiles/ and removes RESTRICTED and NO_CDROM distfiles). It is usually a good idea to run the restricted.sh and/or cdrom.sh scripts by hand after dopackages finishes just to be safe. Run the restricted.sh script before uploading to ftp-master, then run cdrom.sh before preparing the final package set for a release. The package subdirectories are named by whether they are for release, stable, or current. Examples: packages-6.4-release packages-6-stable packages-7.2-release packages-7-stable packages-8-current Some of the directories on ftp-master are, in fact, symlinks. Examples: packages-stable packages-current Be sure you move the new packages directory over the real destination directory, and not one of the symlinks that points to it. If you are doing a completely new package set (e.g. for a new release), copy packages to the staging area on ftp-master with something like the following: &prompt.root; cd /var/portbuild/${arch}/${branch} &prompt.root; tar cfv - packages/ | ssh portmgr@ftp-master tar xfC - w/ports/${arch}/tmp/${subdir} Then log into ftp-master, verify that the package set was transferred successfully, remove the package set that the new package set is to replace (in ~/w/ports/${arch}), and move the new set into place. (w/ is merely a shortcut.) For incremental builds, packages should be uploaded using rsync so we do not put too much strain on the mirrors. ALWAYS use -n first with rsync and check the output to make sure it is sane. If it looks good, re-run the rsync without the -n option. Example rsync command for incremental package upload: &prompt.root; rsync -n -r -v -l -t -p --delete packages/ portmgr@ftp-master:w/ports/${arch}/${subdir}/ | tee log Distfiles can be transferred with the cpdistfiles script: &prompt.root; /var/portbuild/scripts/cpdistfiles ${arch} ${branch} Or you can do it by hand using rsync command: &prompt.root; cd /var/portbuild/${arch}/${branch} &prompt.root; rsync -n -r -v -l -p -c distfiles/ portmgr@ftp-master:w/ports/distfiles/ | tee log Again, run the command without the -n option after you have checked it. Experimental Patches Builds Experimental patches builds are run from time to time to new features or bugfixes to the ports infrastructure (i.e. bsd.port.mk), or to test large sweeping upgrades. The current experimental patches branch is 7-exp on the amd64 architecture. In general, an experimental patches build is run the same way as any other build, except that you should first update the ports tree to the latest version and then apply your patches. To do the former, you can use the following: &prompt.user; cvs -R update -dP > update.out &prompt.user; date > cvsdone This will most closely simulate what the dopackages script does. (While cvsdone is merely informative, it can be a help.) You will need to edit update.out to look for lines beginning with ^M, ^C, or ^? and then deal with them. It is always a good idea to save original copies of all changed files, as well as a list of what you are changing. You can then look back on this list when doing the final commit, to make sure you are committing exactly what you tested. Since the machine is shared, someone else may delete your changes by mistake, so keep a copy of them in e.g. your home directory on freefall. Do not use tmp/; since pointyhat itself runs some version of -CURRENT, you can expect reboots (if nothing else, for updates). In order to have a good control case with which to compare failures, you should first do a package build of the branch on which the experimental patches branch is based for the &i386; architecture (currently this is 7). Then, when preparing for the experimental patches build, checkout a ports tree and a src tree with the same date as was used for the control build. This will ensure an apples-to-apples comparison later. Once the build finishes, compare the control build failures to those of the experimental patches build. Use the following commands to facilitate this (this assumes the 7 branch is the control branch, and the 7-exp branch is the experimental patches branch): &prompt.user; cd /var/portbuild/i386/7-exp/errors &prompt.user; find . -name \*.log\* | sort > /tmp/7-exp-errs &prompt.user; cd /var/portbuild/i386/7/errors &prompt.user; find . -name \*.log\* | sort > /tmp/7-errs If it has been a long time since one of the builds finished, the logs may have been automatically compressed with bzip2. In that case, you must use sort | sed 's,\.bz2,,g' instead. &prompt.user; comm -3 /tmp/7-errs /tmp/7-exp-errs | less This last command will produce a two-column report. The first column is ports that failed on the control build but not in the experimental patches build; the second column is vice versa. Reasons that the port might be in the first column include: Port was fixed since the control build was run, or was upgraded to a newer version that is also broken (thus the newer version should appear in the second column) Port is fixed by the patches in the experimental patches build Port did not build under the experimental patches build due to a dependency failure Reasons for a port appearing in the second column include: Port was broken by the experimental patches [1] Port was upgraded since the control build and has become broken [2] Port was broken due to a transient error (e.g. FTP site down, package client error, etc.) Both columns should be investigated and the reason for the errors understood before committing the experimental patches set. To differentiate between [1] and [2] above, you can do a rebuild of the affected packages under the control branch: &prompt.user; cd /var/portbuild/i386/7/ports Be sure to cvs update this tree to the same date as the experimental patches tree. The following command will set up the control branch for the partial build: &prompt.user; /var/portbuild/scripts/dopackages.7 -noportscvs -nobuild -nocvs -nofinish The builds must be performed from the packages/All directory. This directory should initially be empty except for the Makefile symlink. If this symlink does not exist, it must be created: &prompt.user; cd /var/portbuild/i386/7/packages/All &prompt.user; ln -sf ../../Makefile . &prompt.user; make -k -j<#> <list of packages to build> <#> is the concurrency of the build to attempt. It is usually the sum of the weights listed in /var/portbuild/i386/mlist unless you have a reason to run a heavier or lighter build. The list of packages to build should be a list of package names (including versions) as they appear in INDEX. The PKGSUFFIX (i.e. .tgz or .tbz) is optional. This will build only those packages listed as well as all of their dependencies. You can check the progress of this partial build the same way you would a regular build. Once all the errors have been resolved, you can commit the package set. After committing, it is customary to send a HEADS UP email to ports@FreeBSD.org and copy ports-developers@FreeBSD.org informing people of the changes. A summary of all changes should also be committed to /usr/ports/CHANGES. How to configure a new package building node Before following these steps, please coordinate with portmgr. Node requirements portmgr is still working on characterizing what a node needs to be generally useful. CPU capacity: TBA. However, we have several dual-CPU P-III &i386; 1.0GHz machines available, so anything with less horsepower than that is not as likely to be useful. (However, many of our &sparc64;s are single-CPU, 500MHz machines, so our requirements are lower.) We are able to adjust the number of jobs dispatched to each machine, and we generally tune the number to use 100% of CPU. RAM: TBA. Again, we have been tuning to one job per 512M of RAM. (Anything less than 512M is very unlikely to be useful.) disk: at least 20G is needed for filesystem; 32G is needed for swap. Best performance will be if multiple disks are used, and configured as geom stripes. Performance numbers are also TBA. Package building will test disk drives to destruction. Be aware of what you are signing up for! network bandwidth: TBA. However, an 8-job machine has been shown to saturate a cable modem line. Preparation Pick a unique hostname. It does not have to be a publicly resolvable hostname (it can be a name on your internal network). By default, package building requires the following TCP ports to be accessible: 22 (ssh), 414 (infoseek), and 8649 (ganglia). If these are not accessible, pick others and ensure that an ssh tunnel is set up (see below). (Note: if you have more than one machine at your site, you will need an individual TCP port for each service on each machine, and thus ssh tunnels will be necessary. As such, you will probably need to configure port forwarding on your firewall.) Decide if you will be booting natively or via pxeboot. You will find that it is easier to keep up with changes to -current with the latter, especially if you have multiple machines at your site. Pick a directory to hold ports configuration and chroot subdirectories. It may be best to put it this on its own partition. (Example: /usr2/.) Configuring <literal>src</literal> Create a directory to contain the latest -current source tree and check it out. (Since your machine will likely be asked to build packages for -current, the kernel it runs should be reasonably up-to-date with the bindist that will be exported by our scripts.) If you are using pxeboot: create a directory to contain the install bits. You will probably want to use a subdirectory of /pxeroot, e.g., /pxeroot/${arch}-${branch}. Export that as DESTDIR. If you are cross-building, export TARGET_ARCH=${arch}. The procedure for cross-building ports is not yet defined. Generate a kernel config file. Include GENERIC (or, if you are using more than 3.5G on &i386;, PAE). Suggested options: options GEOM_CONCAT options GEOM_STRIPE options NULLFS options TMPFS options SHMMAXPGS=65536 options SEMMNI=40 options SEMMNS=240 options SEMUME=40 options SEMMNU=120 options ALT_BREAK_TO_DEBUGGER options PRINTF_BUFR_SIZE=128 For PAE, it is not currently possible to load modules. Therefore, you will need to add: options COMPAT_LINUX options LINPROCFS As root, do the usual build steps, e.g.: make -j4 buildworld make buildkernel KERNCONF=${kernconf} make installkernel KERNCONF=${kernconf} make installworld The install steps use DESTDIR. Customize files in etc/. Whether you do this on the client itself, or another machine, will depend on whether you are using pxeboot. If you are using pxeboot: create a subdirectory of ${DESTDIR} called conf/. Create one subdirectory default/etc/, and (if your site will host multiple nodes), subdirectories ${ip-address}/etc/ to contain override files for individual hosts. (You may find it handy to symlink each of those directories to a hostname.) Copy the entire contents of ${DESTDIR}/etc/ to default/etc/; that is where you will edit your files. The by-ip-address etc/ directories will probably only need customized rc.conf files. In either case, apply the following steps: Create a ports-${arch} user and group. Add it to the wheel group. It can have the '*' password. Create /home/ports-${arch}/.ssh/ and populate authorized_keys. Also add the following users: squid:*:100:100::0:0:User &:/usr/local/squid:/bin/sh ganglia:*:102:102::0:0:User &:/usr/local/ganglia:/bin/sh Add them to etc/group as well. Create the appropriate files in etc/.ssh/. In etc/crontab: add * * * * * root /var/portbuild/scripts/client-metrics Create the appropriate etc/fstab. (If you have multiple, different, machines, you will need to put those in the override directories.) In etc/inetd.conf: add infoseek stream tcp nowait nobody /var/portbuild/scripts/reportload reportload ${arch} We run the cluster on UTC: cp /usr/share/zoneinfo/Etc/UTC etc/localtime Create the appropriate - etc/rc.conf. (If you have multiple, + etc/rc.conf. (If you are using + pxeboot, and have multiple, different, machines, you will need to put those in the override directories.) Recommended entries: hostname="${hostname} inetd_enable="YES" linux_enable="YES" nfs_client_enable="YES" ntpd_enable="YES" ntpdate_enable="YES" ntpdate_flags="north-america.pool.ntp.org" sendmail_enable="NONE" sshd_enable="YES" sshd_program="/usr/local/sbin/sshd" gmond_enable="YES" squid_enable="YES" Create etc/resolv.conf, if necessary. Modify etc/sysctl.conf: 9a10,30 > kern.corefile=/tmp/%N.core > kern.sugid_coredump=1 > #debug.witness_ddb=0 > #debug.witness_watch=0 > > # squid needs a lot of fds (leak?) > kern.maxfiles=40000 > kern.maxfilesperproc=30000 > > # Since the NFS root is static we don't need to check frequently for file changes > # This saves >75% of NFS traffic > vfs.nfs.access_cache_timeout=300 > debug.debugger_on_panic=1 > > # For jailing > security.jail.sysvipc_allowed=1 > security.jail.allow_raw_sockets=1 > security.jail.chflags_allowed=1 > security.jail.enforce_statfs=1 > > vfs.lookup_shared=1 If desired, modify etc/syslog.conf to change the logging destinations to @pointyhat.freebsd.org. Configuring <literal>ports</literal> Install the following ports: net/rsync security/openssh-portable (with HPN on) security/sudo sysutils/ganglia-monitor-core (with GMETAD off) www/squid (with SQUID_AUFS on) There is a WIP to create a meta-port, but it is not yet complete. Customize files in usr/local/etc/. Whether you do this on the client itself, or another machine, will depend on whether you are using pxeboot. The trick of using conf override subdirectories is less effective here, because you would need to copy over all subdirectories of usr/. This is an implementation detail of how the pxeboot works. Apply the following steps: Modify usr/local/etc/gmond.conf: 21,22c21,22 < name = "unspecified" < owner = "unspecified" --- > name = "${arch} package build cluster" > owner = "portmgr@FreeBSD.org" 24c24 < url = "unspecified" --- > url = "http://pointyhat.freebsd.org" If there are machines from more than one cluster in the same multicast domain (basically = LAN) then change the multicast groups to different values (.71, .72, etc). Create usr/local/etc/rc.d/portbuild.sh: #!/bin/sh # # Configure the system post-boot ln -sf /usr2/portbuild /var/portbuild # Identify builds ready for use cd /var/portbuild/${arch} for i in */builds/*; do if [ -f ${i}/.ready ]; then mkdir /tmp/.setup-${i##*/} fi done # Flag that we are ready to accept jobs touch /tmp/.boot_finished If there are machines from more than one cluster in the same multicast domain (basically = LAN) then change the multicast groups to different values (.71, .72, etc). Modify usr/local/etc/squid/squid.conf: 288,290c288,290 < #auth_param basic children 5 < #auth_param basic realm Squid proxy-caching web server < #auth_param basic credentialsttl 2 hours --- > auth_param basic children 5 > auth_param basic realm Squid proxy-caching web server > auth_param basic credentialsttl 2 hours 611a612 > acl localnet src 127.0.0.0/255.0.0.0 655a657 > http_access allow localnet 2007a2011 > maximum_object_size 400 MB 2828a2838 > negative_ttl 0 minutes Also, change usr/local to usr2 in the definitions of cache_dir, access_log, cache_log, cache_store_log, pid_filename, netdb_filename, coredump_dir. Finally, change the cache_dir storage scheme from ufs to aufs (offers better performance). Configure ssh: copy /etc/ssh to /usr/local/etc/ssh and add NoneEnabled yes to sshd_config. Modify usr/local/etc/sudoers: 38a39,42 > > # local changes for package building > %wheel ALL=(ALL) ALL > ports-${arch} ALL=(ALL) NOPASSWD: ALL Configuration on the client itself Change into the port/package directory you picked above, e.g., cd /usr2. As root: mkdir portbuild chown ports-${arch}:ports-${arch} portbuild mkdir pkgbuild chown ports-${arch}:ports-${arch} pkgbuild mkdir squid mkdir squid/cache mkdir squid/logs chown -R squid:squid squid If clients preserve /var/portbuild between boots then they must either preserve their /tmp, or revalidate their available builds at boot time (see the script on the amd64 machines). They must also clean up stale chroots from previous builds before creating /tmp/.boot_finished. Boot the client. As root, initialize the squid directories: squid -z Configuration on <literal>pointyhat</literal> These steps need to be taken by a portmgr acting as root on pointyhat. If any of the default TCP ports is not available (see above), you will need to create an ssh tunnel for it and include it in the appropriate crontab. Add an entry to /home/ports-${arch}/.ssh/config to specify the public IP address, TCP port for ssh, username, and any other necessary information. Add the public IP address to /etc/hosts.allow. (Remember, multiple machines can be on the same IP address.) Create /var/portbuild/${arch}/clients/bindist-${hostname}.tar. Copy one of the existing ones as a template and unpack it in a temporary directory. Customize etc/resolv.conf and etc/make.conf for the local site. tar it up and move it to the right location. Hint: you will need one of these for each machine; however, if you have multiple machines at one site, you may be able to create a site-specific one and symlink to it. Create /var/portbuild/${arch}/portbuild-${hostname} using one of the existing ones as a guide. This file contains overrides to /var/portbuild/${arch}/portbuild.conf. Suggested values: disconnected=1 http_proxy="http://localhost:3128/" squid_dir=/usr2/squid scratchdir=/usr2/pkgbuild client_user=ports-${arch} sudo_cmd="sudo -H" rsync_gzip=-z infoseek_host=localhost infoseek_port=${tunelled-tcp-port} Possible other values: use_md_swap=1 md_size=9g use_zfs=1 scp_cmd="/usr/local/bin/scp" ssh_cmd="/usr/local/bin/ssh" Add an appropriate data_source entry to /usr/local/etc/gmetad.conf: data_source "arch/location Package Build Cluster" 30 hostname Enabling the node These steps need to be taken by a portmgr acting as ports-arch on pointyhat. Ensure that ssh is working by executing ssh hostname. Populate /var/portbuild/scripts/ by something like /var/portbuild/dosetupnode arch major latest hostname. Verify that you now have files in that directory. Test the other TCP ports by executing telnet hostname portnumber. 414 (or its tunnel) should give you a few lines of status information including arch and osversion; 8649 should give you an XML response from ganglia. This step needs to be taken by a portmgr acting as root on pointyhat. Tell qmanager about the node. Example: python /var/portbuild/evil/qmanager/qclient add name=uniquename arch=arch osversion=osversion numcpus=number haszfs=0 online=1 domain=domain primarypool=package pools="package all" maxjobs=1 acl="ports-arch,deny_all" Procedures for dealing with disk failures When a machine has a disk failure (e.g. panics due to read errors, etc), then we should do the following steps: Note the time and failure mode (e.g. paste in the relevant console output) in /var/portbuild/${arch}/reboots For i386 gohan clients, scrub the disk by touching /SCRUB in the nfsroot (e.g. /a/nfs/8.dir1/SCRUB) and rebooting. This will dd if=/dev/zero of=/dev/ad0 and force the drive to remap any bad sectors it finds, if it has enough spares left. This is a temporary measure to extend the lifetime of a drive that is on the way out. For the i386 blade systems another signal of a failing disk seems to be that the blade will completely hang and be unresponsive to either console break, or even NMI. For other build systems that don't newfs their disk at boot (e.g. amd64 systems) this step has to be skipped. If the problem recurs, then the disk is probably toast. Take the machine out of mlist and (for ata disks) run smartctl on the drive: smartctl -t long /dev/ad0 It will take about 1/2 hour: gohan51# smartctl -t long /dev/ad0 smartctl version 5.38 [i386-portbld-freebsd8.0] Copyright (C) 2002-8 Bruce Allen Home page is http://smartmontools.sourceforge.net/ === START OF OFFLINE IMMEDIATE AND SELF-TEST SECTION === Sending command: "Execute SMART Extended self-test routine immediately in off-line mode". Drive command "Execute SMART Extended self-test routine immediately in off-line mode" successful. Testing has begun. Please wait 31 minutes for test to complete. Test will complete after Fri Jul 4 03:59:56 2008 Use smartctl -X to abort test. Then smartctl -a /dev/ad0 shows the status after it finishes: # SMART Self-test log structure revision number 1 # Num Test_Description Status Remaining LifeTime(hours) LBA_of_first_error # 1 Extended offline Completed: read failure 80% 15252 319286 It will also display other data including a log of previous drive errors. It is possible for the drive to show previous DMA errors without failing the self-test though (because of sector remapping). When a disk has failed, please inform the cluster administrators so we can try to get it replaced.