Back to Ceph

Submitting Patches to Ceph - Backports

SubmittingPatches-backports.rst

21.3.020.0 KB
Original Source

Submitting Patches to Ceph - Backports

Most likely you're reading this because you intend to submit a GitHub pull request ("PR") targeting one of the stable branches ("tentacle", etc.) at https://github.com/ceph/ceph.

Before you open that PR, please read this entire document or, at the very least, the following two sections: General principles_ and Cherry-picking rules_.

.. contents:: :depth: 3

General principles

.. note:: There are many automations that exist for backports. If you read nothing else in this document, please read ceph-backport_ section for automating backport creation.

To help the people who will review your backport, please state either in the backport PR, or in the backport tracker issue, or in the main branch tracker issue:

  1. what bug is fixed
  2. why this fix is the minimal way to do it
  3. why does this need to be fixed in <release>

The above should be followed especially in cases when the backport could be seen as introducing, into a stable branch, code that is not related to a particular bug or issue.

Rationale: every modification of a stable branch carries a certain risk of introducing a regression. To minimize this risk, backports should be as straightforward and narrowly-targeted as possible. As a stable release series ages, the importance of following these general principles rises.

Cherry-picking rules

Note: These rules are strictly enforced by the releng-audit GitHub Actions CI. Failure to adhere to them will block your PR.

The following rules, which have been codified from "best practices" developed over years of backporting, apply to the actual backport implementation:

  • all fixes should land in main first
  • commits to stable branches should be cherry-picked from main
  • if a commit is entirely specific to the backport and not cherry-picked from main, its commit message summary must begin with the target branch name (e.g., squid: fix compilation issue).
  • before starting to cherry-pick a set of commits from main, grep the main git history for the SHA1 of each main commit (using git log --grep) to check for follow-up fixes. Include any follow-up fixes found in the set of commits to be cherry-picked.
  • when backporting a main PR to a stable branch, double-check that the backport PR contains cherry-picks of all of the main PR's commits. If any commit needs to be omitted, declare and explain this in the PR.
  • cherry-picks must be done using git cherry-pick -x
  • if a cherry-pick from main is not feasible and a direct fix is being undertaken, this must be explained
  • the commit message generated by git cherry-pick -x must not be modified, except to add a "Conflicts" section below the "cherry picked from commit ..." line added by git
  • the "Conflicts" section must mention all files where changes had to be made manually (not just conflicts flagged by git)
  • the "Conflicts" section should also describe the manual changes that were made
  • if a change is to be backported to multiple stable branches, a tracker issue is needed, so the backports can be tracked (if a change is only to be backported to the most recent stable branch, a tracker issue is not strictly required)

For more information on tracker issues, see Tracker workflow_.

For more information on conflict resolution and writing the "Conflicts" section, see Conflict resolution_.

Adhering to these rules will make your backport easier for reviewers to understand. Not adhering to these rules creates additional work for reviewers and may cause your backport PR to be rejected.

Notes on the cherry-picking rules ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

What does "all fixes should land in main first" mean? What if I just need to fix the issue in <release>?

As the person fixing the issue, you are required to first check whether the issue exists in main. If it does, then the proper course of action is to create a main branch tracker (see Tracker workflow_) and fix the issue in main, first, and only then cherry-pick the fix to the stable branches that have the issue.

If the issue exists in the stable branch, but not in main, there are several possibilities:

  1. it's a regression introduced into the stable branch by a bad backport
  2. the issue was fixed in main by some massive refactoring that cannot be backported
  3. the issue was already fixed in main by a cherry-pickable commit

In cases 1 and 2, it's permissible to fix the issue directly in the most recent stable branch, subject to the rule "if a commit could not be cherry-picked from main, the commit message must explain why that was not possible". Once the fix has landed in the most recent stable branch, it can be cherry-picked to older stable branches from there.

In case 3, the issue should be handled like any other backport - read on.

Tracker workflow

Any change that is to be backported to multiple stable branches should have an associated tracker issue at https://tracker.ceph.com.

For fixes already merged to main, this may have already been done - see the Fixes: line in the main PR. If the main PR has already been merged and there is no associated main branch tracker issue, you can create a main branch tracker issue and fill in the fields as described below.

This main branch tracker issue should be in the "Bug" or "Feature" trackers of the relevant subproject under the "Ceph" parent project (or in the "Ceph" project itself if none of the subprojects are a good fit). The stable branches to which the main changes are to be cherry-picked should be listed in the "Backport" field. For example::

Backport: mimic, nautilus

Once the PR targeting main is open, add the PR number assigned by GitHub to the tracker issue. For example, if the PR number is 99999::

Pull request ID: 99999

In general, you should stop here with manually managing the state of the tracker tickets. If the Pull request ID field of the main tracker is up-to-date, the redmine-upkeep Github CI will automatically detect a merge and move your tracker to the Pending backport state as long as the Backport field has release branches in it. Otherwise the ticket will be moved to Resolved.

However, if that CI fails to run for whatever reason, you may manually correct the ticket by setting the main branch tracker issue's Status field to "Pending Backport"::

Status: Pending Backport

If you do not have sufficient permissions to modify any field of the tracker issue, just add a comment describing what changes you would like to make. Someone with permissions will make the necessary modifications on your behalf.

.. _create backport tracker issues: .. _backport tracker issue:

Creating Backport tracker issues

To track backporting efforts, "backport tracker issues" can be created from a parent main branch tracker issue. The main branch tracker issue is described in the previous section, Tracker workflow_. This section focuses the backport tracker issue.

When a main branch tracker issue is in the Pending Backport state, another Github CI running backport-create-issue will trigger and create the backport tickets to be associated with the main tracker. This workflow will auto-populate the required metadata for the tracker tickets to track the backports.

The backport-create-issue script ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

.. warning:: It is discouraged to run this script yourself. However, sometimes CI checks fail to run, so please review for those circumstances.

The script used to create backport issues is located at src/script/backport-create-issue in the main branch. Though there might be an older version of this script in a stable branch, do not use it. Only use the most recent version from main.

Once you have the script somewhere in your PATH, you can proceed to install the dependencies.

The dependencies are:

  • python3
  • python-redmine

Python 3 should already be present on any recent Linux installation. The second dependency, python-redmine_, can be obtained from PyPi::

pip3 install --user python-redmine

.. _python-redmine: https://pypi.org/project/python-redmine/

Then, try to run the script::

python3 -c "$(git show main:src/script/backport-create-issue)" --help

This should produce a usage message.

Finally, run the script to actually create the Backport issues. For example, if the tracker issue number is 55555::

python3 -c "$(git show main:src/script/backport-create-issue)" --user <tracker_username> --password <tracker_password> 55555

The script needs to know your https://tracker.ceph.com credentials in order to authenticate to Redmine. In lieu of providing your literal username and password on the command line, you could also obtain a REST API key ("My account" -> "API access key"), put it in ~/.redmine_key and run the script like so::

python3 -c "$(git show main:src/script/backport-create-issue)" 55555

.. _stage backports: .. _stage the backport: .. _staging a backport:

Opening a backport PR

Once the Tracker workflow_ is completed and the backport tracker issue_ has been created, it's time to open a backport PR. One possibility is to do this manually, while taking care to follow the cherry-picking rules. However, this can result in a backport that is not properly staged. For example, the PR description might not contain a link to the backport tracker issue (a common oversight). You might even forget to update the backport tracker issue_.

In the past, much time was lost, and much frustration caused, by the necessity of staging backports manually. Now, fortunately, there is a script available which automates the process and takes away most of the guesswork.

.. _ceph-backport:

The ceph-backport.sh script ^^^^^^^^^^^^^^^^^^^^^^^^^^^

Similar to the case of creating backport tracker issues_, staging the actual backport PR and updating the Backport tracker issue is difficult - if not impossible - to get right if you're doing it manually, and quickly becomes tedious if you do it more than once in a long while.

The ceph-backport.sh script automates the entire process of cherry-picking the commits from the main PR, opening the GitHub backport PR, and cross-linking the GitHub backport PR with the correct Backport tracker issue. The script can also be used to good effect if you have already manually prepared the backport branch with the cherry-picks in it.

The script is located at src/script/ceph-backport.sh in the main branch. Though there might be an older version of this script in a stable branch, do not use it. Only use the most recent version from the main branch. To do this from anywhere and from any branch use the following alias that will use the most recent script in upstream/main of your local ceph clone on every call::

alias ceph-backport.sh="bash <(git --git-dir=$pathToCephClone/.git --no-pager show upstream/main:src/script/ceph-backport.sh)"

ceph-backport.sh is just a bash script, so the only dependency is bash itself, but it does need to be run in the top level of a local clone of ceph/ceph.git. A small up-front time investment is required to get the script working in your environment. This is because the script needs to authenticate itself (i.e., as you) in order to use the GitHub and Redmine REST API services.

The script is self-documenting. Just run the script and proceed from there.

Once the script has been set up properly, you can validate the setup like so::

ceph-backport.sh --setup

Once you have this saying "Overall setup is OK", you have two options for staging the backport: either leave everything to the script, or prepare the backport branch yourself and use the script only for creating the PR and updating the Backport tracker issue.

If you prefer to leave everything to the script, just provide the Backport tracker issue number to the script::

ceph-backport.sh 55555

The script will start by creating the backport branch in your local git clone. The script always uses the following format for naming the branch::

wip-<backport_issue_number>-<name_of_stable_branch>

For example, if the Backport tracker issue number is 55555 and it's targeting the stable branch "nautilus", the backport branch would be named::

wip-55555-nautilus

If you prefer to create the backport branch yourself, just do that. Be sure to name the backport branch as described above. (It's a good idea to use this branch naming convention for all your backporting work.) Then, run the script::

ceph-backport.sh 55555

The script will see that the backport branch already exists, and use it.

Once the script hits the first cherry-pick conflict, it will no longer provide any cherry-picking assistance, so in that case it's up to you to resolve the conflict(s) (as described in Conflict resolution_) and finish cherry-picking all of the remaining commits. Once you are satisfied that the backport is complete in your local branch, ceph-backport.sh can finish the job of creating the pull request and updating the backport tracker issue. To make that happen, just re-run the script exactly as you did before::

ceph-backport.sh $BACKPORT_TRACKER_ID

The script will detect that it is running from a branch with the same name as the one it would normally create on the first run and continues after the cherry-picking phase.

For a quick reference on CLI, that contains above information, you can run::

ceph-backport.sh --usage

Milestones and Labels """""""""""""""""""""

The ceph-backport.sh script automates the process of setting the Milestone tag to the stable release the backport PR is targeting.

PR labels (such as component labels) are added automatically by automations; the backport author does not need to do anything manually.

Conflict resolution ^^^^^^^^^^^^^^^^^^^

Automated Conflict Simulation: The releng-audit CI will perform a dry-run of the cherry-pick and compare the resulting tree to your PR. If you make any manual changes (even to fix a bug introduced by the backport), you must document it in the Conflicts: block. If a legitimate manual resolution is flagged as an unapproved deviation, any user with "maintain" or "admin" rights on the repository, or a member of the @ceph/ceph-release-manager team, must comment /audit override or manually apply the releng-audit-override label on the PR to bypass the simulation check.

If git reports conflicts, the script will abort to allow you to resolve the conflicts manually. Once the conflicts are resolved, complete the cherry-pick ::

git cherry-pick --continue

Git will present a draft commit message with a "Conflicts" section.

Unfortunately, in recent versions of git, the Conflicts section is commented out. Since the Conflicts section is mandatory for Ceph backports that do not apply cleanly, you will need to uncomment the entire "Conflicts" section of the commit message before committing the cherry-pick. You can also include commentary on what the conflicts were and how you resolved them. For example::

Conflicts:
        src/foo/bar.cc
- mimic does not have blatz; use batlo instead

When editing the cherry-pick commit message, leave everything before the "cherry picked from" line unchanged. Any edits you make should be in the part following that line. Here is an example::

osd: check batlo before setting blatz

Setting blatz requires special precautions. Check batlo first.

Fixes: https://tracker.ceph.com/issues/99999
Signed-off-by: Random J Developer <[email protected]>
(cherry picked from commit 01d73020da12f40ccd95ea1e49cfcf663f1a3a75)

Conflicts:
	src/osd/batlo.cc
- add_batlo_check has an extra arg in newer code

Naturally, the Fixes line points to the main issue. You might be tempted to modify it so it points to the backport issue, but - please - don't do that. First, the main issue points to all the backport issues, and second, any editing of the original commit message calls the entire backport into doubt, simply because there is no good reason for such editing.

The part below the (cherry picked from commit ...) line is fair game for editing. If you need to add additional information to the cherry-pick commit message, append that information below this line. Once again: do not modify the original commit message.

If you use ceph-backport.sh for your backport creation (which is recommended), read up at the end of The ceph-backport.sh script_ on how to continue from here.

Automated Audit Workflow

Once your backport PR is open, it will be automatically audited by the releng-audit GitHub Actions CI. This workflow enforces backport rules and ensures consistency between GitHub and Redmine.

The audit performs the following checks:

  • Merge Conflict: Verifies that the PR can merge cleanly into the target base branch without git conflicts.
  • Commit Parity: Ensures all commits from the original main PR(s) are present in the backport. It flags missing commits, unmerged cherry-picks, or invalid commit message formats.
  • Conflict Simulation: Dry-runs the cherry-pick sequence to dynamically verify conflict resolutions. It will fail if there are undocumented or unapproved deviations from a clean cherry-pick.
  • Redmine Linkage: Checks that the backport PR is properly linked to its Redmine backport tracker, and that the original main PR is correctly linked to the parent tracker ticket.

Addressing Failures (PR Author Workflow):

If the audit fails, the CI bot will post a review detailing the issues and apply the releng-audit-fail label.

  • Fix the issues: Address the failures as guided by the bot's review (e.g., fix Redmine tracker links, rebase your branch, fix commit messages, or adjust the code to match upstream patches).
  • Retest: When you are ready for a new audit, remove the releng-audit-fail label or comment /audit retest on the PR.
  • Note: You cannot manually apply the releng-audit-pass or releng-audit-fail labels; they are strictly managed by the CI bot and your manual changes to them will be rejected.

Bypassing the Audit (Component Lead / Release Manager Workflow):

In some cases, the audit may flag a legitimate manual conflict resolution or an intentional deviation in the code.

  • If a deviation is intentional, documented in the commit message, and approved, authorized users (repository admins/maintainers or members of the @ceph/ceph-release-manager team) can bypass the check.
  • To apply the bypass, comment /audit override on the PR or manually apply the releng-audit-override label.

Audit Labels Explained:

  • releng-audit-pass: Applied automatically by the bot when all checks pass.
  • releng-audit-fail: Applied automatically by the bot when checks fail. Removing this label triggers a new audit.
  • releng-audit-override: Applied manually by authorized leads/managers to override an audit failure.

The bot will leave a failing CI check if a backport PR does not have either releng-audit-pass or releng-audit-override.

.. _backport PR reviewing: .. _backport PR testing: .. _backport PR merging:

Reviewing, testing, and merging of backport PRs

Once your backport PR is open, it will be reviewed and tested. When the PR has been reviewed and tested, it will be merged.

  • Prerequisite: The PR must have the releng-audit-pass or releng-audit-override label applied before it is eligible for merging.

If you would like to facilitate this process, you can solicit reviews and run integration tests on the PR. In this case, add comments to the PR describing the tests you ran and their results.

Once the PR has been reviewed and deemed to have undergone sufficient testing, it will be merged. Even if you have sufficient GitHub permissions to merge the PR, please do not merge it yourself. (Uncontrolled merging to stable branches unnecessarily complicates the release preparation process, which is done by volunteers.)