From Fedora Project Wiki

Revision as of 17:59, 29 June 2014 by Dwa (talk | contribs) (→‎Work Items)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)

David Aquilina

Contact Info

E-mail: dwa (at) fedoraproject.org

IRC: 'dwa' on freenode

Everyone loves badges: Riddle Me This Speak Up! In Search of the Bull (Tester I) Junior Editor Senior Editor Associate Editor Involvement Let Me Introduce Myself White Rabbit Baby Badger Junior Badger (Badger I) Master Editor Embryo Froglet Tadpole Tadpole with Legs Secretary General If you build it... (Koji Success I) What goes up... (Koji Failure I) If you build it... (Koji Success II) What goes up... (Koji Failure II) If you build it... (Koji Success III) If you build it... (Koji Success IV) All your $arch are belong to us (Koji Success V) Helping Hand Hindenburg (Koji Failure V) Egg What goes up... (Koji Failure III) What goes up... (Koji Failure IV) Associate Badger (Badger 1.5) Origin Adult Frog Bona Fide Crypto Panda Taming the Beasts Paranoid Panda

Secondary Arch Releng

Builds

Principles

  • Never build an NVR before it's built on primary (koji.fedoraproject.org). The code used for a given NVR must match exactly between primary and secondary. koji-shadow uses source RPMs pulled from koji.fp.o, when building manually you should use a git url from pkgs.fedoraproject.org (e.g. git://pkgs.fedoraproject.org/kernel?#5394ba72a9d27667d10801685f71c003a7e205bc for kernel-3.9.6-301.fc19 )
  • Never build 'real' (i.e. not scratch) packages with code that hasn't been committed to pkgs.fp.o.
    • Milestone composes (TCs & RCs for alpha, beta and GA) should only ever be done with real packages. Including scratch builds in test composes should be kept to an absolute minimum.
  • Buildroots and dependencies should match primary as closely as possible. koji-shadow enforces this automatically, the only time you should build a package manually is when something is broken in a way which prevents koji-shadow from building it.
    • e.g. if eclipse-4.3.0-0.60.git7bf397.fc19 had a previous, broken-on-secondary version of eclipse in its buildroot, you should build eclipse-4.3.0-0.60.git7bf397.fc19 manually.
    • 'build a package manually' means submitting the build to ppc.koji.fp.o with e.g. ppc-koji build f19 git://blah
    • If not broken, build a single package with koji-shadow -c <config file> --build <nvr> <release>-build
      • This assumes <config file> specifies prefer_new and import_noarch, and you're using a koji-shadow from git which supports those options in conjunction with building a single NVR (previous versions of koji-shadow, using '--build <nvr>' caused it to match the buildroot exactly).
      • e.g.: /home/dwa/shadow/koji-shadow -c /home/dwa/shadow/f19.conf --build gnome-initial-setup-0.12-1.fc19 f19-build

Methodology

  1. koji-stalk.py monitors fedmsg for completed build notifications from koji.fp.o.
    • koji-stalk.py is run on the hub out of a screen session by dwa
    • Use /home/dwa/shadow/f19.conf as the koji-shadow configuration file.
    • Use /home/dwa/shadow/koji-shadow as the koji-shadow binary.
    • ./koji-stalk.py -c /home/dwa/shadow/f19.conf --shadow /home/dwa/shadow/koji-shadow
  2. When it receives a build notification, it kicks off koji-shadow for that NVR, and tags it into the -updates-candidate tag.
  3. Tags are automatically synced with primary
    • Tags are synced out of cron on the hub by ppc-sync-tagged-primary.py
    • dgilmore also runs sync-tagged-primary against all arches after every stable push.
  4. Signing the tags and mashing/pushing updates are all manually done.


  • koji-stalk.py should be the primary way that builds happen.
  • The koji-stalk logs should be monitored for failed builds to know where to investigate problems.
  • koji-shadow should be occasionally manually run against the stable tags (and possibly updates-testing?) to catch any builds that might have been missed by the script (e.g. it crashed, or was stopped for system maintenance or to update the script).
    • stable tags are the release tag pre-GA (e.g. 'f19'), and the updates tags (e.g. 'f18-updates') post-GA.
    • koji-stalk.py builds ahead of updates-candidate, so if unknown builds are missed the occasional runs against stable or updates-testing will eventually, but not immediately, catch them.
    • koji-stalk prints its queues (if any) every 10 minutes in the log, should you need to stop the script to update it.

Compose & Build scripts

  • Scripts for mashing & composing milestones, bleed repos, and nightly images are on my people page.
    • mash-milestone.sh mashes a repo for a particular release while enforcing strict signing, unlike buildbranched. Edit the DATE variable before running it.
    • compose.sh composes test releases. It uses the fedora-install-fedora.ks which points to the branched & bleed repos, and does not create source or debuginfo trees. Edit the DATE variable before running it.
    • compose-milestone.sh composes milestone releases. Ensure there's a $milestone-fedora-install-fedora.ks which points to a repo which was mashed with strict signing (such as what mash-milestone.sh creates). Edit the DATE and MILESTONE variables before running it.
    • make-bleed.sh creates a bleed repo of packages from the PKGS variable. Edit the PKGS variable before running it.
    • All of the above should be run from ppc-composer.
  • koji-stalk.py is a script for monitoring fedmsg and kicking off builds as soon as they finish on primary.
  • All build scripts (koji-stalk.py and koji-shadow) are currently run from the hub.
  • All compose and mash scripts, pushing updates, and signing are done from the composer.
  • koji-stalk.py and sigulsign_unsigned.py are available from the releng git repo.

TC & RC Composes

  • TCs are Test Composes. They will never be formal releases so the only thing that makes them different from daily ISOs is the package list for the bleed repo is taken from primary's releng ticket for that release/milestone, e.g. https://fedorahosted.org/rel-eng/ticket/5623
    1. Grab the list of NVRs from the ticket, then edit make-bleed.sh and put that (space separated) list in the PKGS variable.
    2. To make life easier for you down the road, it's best to sign those builds now. Run sigulsign_unsigned.py --arch ppc -v fedora-19-secondary <space separated list of NVRs>, and take note of any builds which don't exist yet (then build them, and sign).
    3. Run make-bleed.sh -s ba094068 to build the bleed repos.
      • ba094068 is the key ID for fedora-19-secondary. As an argument, it is case-sensitive.
    4. Edit compose.sh to reflect the correct DATE label (e.g. f19-20130618-GA-TC5) and run it.
  • RCs are Release Candidates. They may become formal releases, so we need to ensure that all packages in an RC are signed. RCs also have full source and debuginfo trees generated as well, so they take a little longer to make.
    1. Sign the entire tag for the release.
      • sigulsign_unsigned.py --arch ppc -v --write --tag f19 fedora-19-secondary
    2. Ensure the bleed repo only contains signed packages.
      • The following command should not return any output: rpm -q --qf="%{RSAHEADER}\n" -p /mnt/koji/mash/bleed-repo/*/*.rpm | grep -v ba094068
    3. Edit mash-milestone.sh to have the appropriate date label, then run it. Have a cup of coffee.
      • If mash-milestone.sh complains of unsigned packages, you may need to re-run sigulsign_unsigned.py with the --write-all option then try again.
      • If mash *still* complains of unsigned packages, ensure the correct key is specified in the mash config.
    4. Edit compose-milestone.sh for the appropriate milestone & date labels. the OLD variable will affect which tree is used in repodiff.
    5. Ensure that /mnt/data/kickstarts/f19-kickstarts/$MILESTONE-fedora-install-fedora.ks exists and has an appropriate package list & exclude list.
    6. Run compose-milestone.sh. The install tree will automatically be placed in /mnt/koji/stage. Be sure to check the repoclosure output before asking people to test.
    7. If you see errors such as 'Cannot find source RPM for foo-1.2-3.fc20.src.rpm', check to make sure if you're excluding a package in the repo line of your kickstart file you're also excluding all subpackages. E.g. the blender source RPM creates both blender and fonts-blender packages.

Staging Releases

  • Sign the iso/CHECKSUM file:
cat Fedora-18-ppc64-CHECKSUM > /tmp/sum
sigul sign-text -o /tmp/signed fedora-18-secondary /tmp/sum
sudo mv /tmp/signed Fedora-18-ppc64-CHECKSUM
  • Also sign the ppc64/os/.treeinfo file and name it .treeinfo.signed.
    • There should be both signed and unsigned copies present in ppc64/os
  • Be sure to fix the permissions afterwards, sigul will write files as 0600.
  • Run build_composeinfo from the releng repo on the Fedora/ppc64 directory.
  • Copy the Fedora directory to:
    • /mnt/koji/tree/releases/(version)/Fedora/ for GA releases
    • /mnt/koji/tree/releases/test/(version)-(milestone)/Fedora for alpha & beta releases
  • For GA, copy the latest branched mash to /mnt/koji/tree/releases/(version)/Everything
    • Ensure delta RPMs, .treeinfo, .discinfo, and any other install bits are removed from it.

Mashing & Pushing Updates

  • The following tags are mashed and pushed regularly (1-3x/week):
    • f20-updates
    • f20-updates-testing
    • f19-updates
    • f19-updates-testing
  • All steps are performed on ppc-composer in a mock chroot which matches the release being mashed.

Step by step

Using f19-updates-testing as an example:

  1. Sign the tag
    • sigulsign_unsigned.py --arch ppc -v --tag f19-updates-testing fedora-19-secondary
    • Use fedora-20-secondary for f20 tags.
  2. Switch to the 'masher' user.
    • The 'masher' user has all of the below aliases already in place and has permissions to write to /mnt/koji/tree.
  3. Mash the tag.
    • Run mash-19-updates-testing on ppc-composer.
    • If mashing complains of unsigned packages, run sigulsign_unsigned.py --arch ppc -v --tag f19-updates-testing --write-all fedora-19-secondary
  4. Copy the mash to /mnt/koji/tree/updates/testing/19.
    • Switch to the directory the mash completed in (e.g. /mnt/koji/mash/updates-testing-f19-20130612/19-updates-testing) and run push-updates /mnt/koji/tree/updates/testing/19
    • For stable tags (i.e. f20-updates and f19-updates) they are pushed to /mnt/koji/tree/updates/(release).


alias mash-17-updates="mash -o /mnt/koji/mash/updates-f17-date +%Y%m%d 17-updates -f ~/comps/comps-f17.xml -c /etc/mash/mash.ppc.conf -p /mnt/koji/tree/updates/17"
alias mash-18-updates='mock -r fedora-18-ppc64 --chroot "mash -o /mnt/koji/mash/updates-f18-date +%Y%m%d 18-updates -f /tmp/comps-f18.xml -c /etc/mash/mash.ppc.conf -p /mnt/koji/tree/updates/18/"'
alias mash-18-updates-testing='mock -r fedora-18-ppc64 --chroot "mash -o /mnt/koji/mash/updates-testing-f18-date +%Y%m%d 18-updates-testing -f /tmp/comps-f18.xml -c /etc/mash/mash.ppc.conf -p /mnt/koji/tree/updates/testing/18/"'
alias mash-19-updates-testing='mock -r fedora-19-ppc64 --uniqueext=mash --chroot "mash -o /mnt/koji/mash/updates-testing-f19-date +%Y%m%d 19-updates-testing -f /tmp/comps-f19.xml -c /etc/mash/mash.ppc.conf -p /mnt/koji/tree/updates/testing/19/"'
alias mash-19-updates='mock -r fedora-19-ppc64 --uniqueext=mash --chroot "mash -o /mnt/koji/mash/updates-f19-date +%Y%m%d 19-updates -f /tmp/comps-f19.xml -c /etc/mash/mash.ppc.conf -p /mnt/koji/tree/updates/19/"'
push-updates() { rsync -rlptDHhv --delay-updates --delete-after ppc ppc64 source "$@" ;}

Creating New Mash Configs

  • Your mash host (or chroot) needs to be at least Fedora 18 to properly mash ppc64p7 RPMs.
  • Using 19-updates.ppc.mash as an example, replace the configuration name (i.e. '19-updates'), the tag name ('f19-updates'), the key ID ('ba094068'), the repoviewurl, and repoviewtitle with the new information and save it in a new file in /etc/mash.
  • You will also need to have mash.ppc.conf present and reference it in your mash command line (-c /etc/mash/mash.ppc.conf) to fetch packages from ppc.koji.

Branching

Draft Information
This is draft documentation and may not be correct. It also does not account for Fedora.Next changes (e.g. there are different kickstarts for each product. You are likely to be eaten by a grue.

The examples in this section are for f20 branching from rawhide.

  • Create empty updates & updates-testing repositories (in /mnt/koji/tree/updates{/testing}/20/{ppc,ppc64,source/SRPMS})
  • Set the /mnt/koji/mash/branched link to point to the last pre-branch rawhide
  • Update the mock configs on composer
    • for branched mocks, change 'fc19' to 'fc20', any repos that point to /mnt/koji/repos/f19-build to /mnt/koji/repos/f20-build
    • for rawhide mocks, change 'fc20' to 'fc21', etc.
  • Update serverstatus to reflect F20 status
    • edit ppc-hub:/usr/local/bin/get-no-of-unsigned.py to reflect new key
    • edit ppc-hub:/usr/local/bin/get-no-of-unsigned.sh to include f20 & f20-updates-testing
    • edit ppc-hub:/usr/share/koji-web/scripts/serverstatus.chtml to include new tags/releases
    • edit ppc-hub:/usr/share/koji-web/scripts/index.py to include new tags/releases
      • backup index.py in case someone accidentally updates koji and overwrites it [fixme - can we break out the serverstatus stuff into a separate file?]
  • Update /home/dwa/bin/ppc-sync-tagged-primary.py to remove EOL releases and add new ones.
  • Update /home/dwa/bin/sync-override-tags.py to remove EOL releases and add new ones.
  • Update nightly-compose.sh, compose-milestone.sh, compose.sh, mash-milestone.sh for new release number.
  • Grab the latest fedora-install-fedora.ks (from the spin-kickstarts git repo) and make any necessary changes to it for our local copy in ppc-composer:/mnt/data/kickstarts/f20-kickstarts

Miscellaneous

Corrupt noarch RPMs

When parsing a noarch RPM, koji-shadow will import it without rebuilding it. Rarely, the RPMs will become corrupt on ppc.koji in the process. Use the koji-reimport.py script from the releng git repo to fix it. Note you must edit the script before running it to specify the NVRs and the tag which they belong to.

mis-tagged shadowbuild packages

  • Find packages that shouldn't be there:
    ppc-koji latest-pkg SHADOWBUILD-f16-build --all --quiet | grep fc18 | awk {'print $1'} > /tmp/mistagged-pkgs 
  • Find the latest versions that should be.:
    for i in `awk -F- '{if (NF > 4) {print $1"-"$2"-"$3} else if (NF == 4) {print $1"-"$2} else {print $1}}' /tmp/mistagged-pkgs`; do ppc-koji latest-pkg f16-updates $i --quiet | awk {'print $1'}; done | grep -v Warning > /tmp/shouldbetagged-pkgs
    • 'grep -v Warning' is needed depending on your koji configuration, to supress the pkgurl deprecated option warning.
  • Untag the offenders:
    ppc-koji untag-pkg SHADOWBUILD-f16-build `cat /tmp/mistagged-pkgs`
  • Tag the right ones:
    ppc-koji tag-pkg SHADOWBUILD-f16-build `cat /tmp/shouldbetagged-pkgs`
  • Regen the repo:
    ppc-koji regen-repo SHADOWBUILD-f16-build

sigul client setup

Sigul for Primary

See Sigul_Client_Setup_SOP

Sigul for Secondary

in ~/.sigul/client.conf:

[client]
bridge-hostname: secondary-signer
server-hostname: secondary-signer-server

in /etc/hosts (internal to phx):

10.5.124.145  sigul-bridge.ausil.us sigul-bridge secondary-signer

General Sigul Notes

If your client is RHEL 6.4, in ~/.bashrc:

export NSS_HASH_ALG_SUPPORT=+MD5

Whenever your Fedora certificate expires, after putting the new one into place, run the following commands:

certutil -d ~/.sigul -D -n sigul-client-cert
sigul_setup_client

Infrastructure Maintenance

/mnt/koji

If /mnt/koji needs to be remounted, e.g. to migrate to a new filer, some downtime is unavoidable. The following should be done:

  1. Disable all builders (ppc-koji disable-host ppc-builder{1..7}) and wait until their existing tasks have finished.
    • Any tasks that haven't yet been picked up by a builder will be queued and preserved when you restart everything.
  2. Stop koji-stalk (since it logs into /mnt/koji/reports)
  3. Stop the kojira and httpd services on the hub
  4. Unmount /mnt/koji/repos from the createrepo hosts (from the hub, pssh -h ~dwa/hosts/createrepo umount /mnt/koji/repos)
  5. Unmount /mnt/koji from all hosts (pssh -h ~dwa/hosts/all umount /mnt/koji)
    • If unmount fails for any reason, use lsof /mnt/koji as root to figure out what still has /mnt/koji open
  6. Do the needful things & stuff for the specific maintenance / migration you're doing.
  7. Remount /mnt/koji
  8. Remount /mnt/koji/repos on createrepo hosts
  9. Start httpd on the hub
  10. Start kojira on the hub
  11. Restart koji-stalk
  12. Re-enable the builders
  13. profit!