I'm trying to formalize the process for merging code into the drm/i915 driver. Here's a first draft, please send along your comments.
-keith
Right now, I'm merging patches destined for the 3.0 release in a kernel.org tree:
git://git.kernel.org/pub/scm/linux/kernel/git/keithp/linux-2.6.git
(yes, I need to fix the name, I'll send a note when I do).
There are two important branches here (and I'll try to keep clutter down by pruning dead branches):
1) drm-intel-next
This contains work destined for the 'next' release, it may include new functionality and performance enhancements. It may also cause regressions on some hardware. The tip of this tree will be sent to Dave Airlie for inclusion in the kernel during the next merge window. I've already sent all of what is on this branch to him for 3.0
This tree should be tested during the development process to try and catch bugs and regressions before they are merged. The most critical time for testing this is just *before* the release of the current RC kernel version as that is when it should have most of the code planned for the *next* release.
2) drm-intel-fixes
This contains bug fixes after the merge window has closed. I will fast-forwarded to each RC when possible so that we test the fully integrated kernel for regressions.
This tree should be tested during the stabilization process after RC1 comes out as patches are applied.
If we need another tree or two to test integrated patch sequences for major rework or new features, we can create other ones, but I hope to keep it to just these two for ongoing work.
So, here's a typical kernel release cycle with notes on where we should be focusing our efforts:
2.6.39 The merge window is open and patches are flowing from drm-intel-next into Dave Airlie's drm-next branch.
Testing should focus on drm-intel-next to ensure that we let as few bugs as possible leak into the main tree. The drm-intel-fixes branch should be essentially idle at this time, although if there are serious bugs in 2.6.39, fixes for those should be tested on drm-intel-fixes and then sent along to stable@ for integration in 2.6.39.1. I don't anticipate keeping a drm-intel-fixes branch open for each stable release.
3.0-RC1 The merge window is closed. New features and major non-regression bug fixes will continue to land on drm-intel-next. Minor bug fixes and any regression fixes will land on drm-intel-fixes. The drm-intel-fixes tree will be sent directly to Linus for integration into the next RC or final release.
Regressions on drm-intel-fixes will be dealt with swiftly -- if new feature or performance work that came from drm-intel-next cause regressions on any hardware, they must either be fixed or the new work will be reverted from the kernel. A regression which malingers through more than one RC is likely to be subjected to this harsh treatment.
During early RC versions, we should focus our efforts on the drm-intel-fixes tree to track reported bugs, especially regressions, to make sure they are fixed as quickly as possible.
However, we cannot afford to ignore the drm-intel-next branch. If major new code lands on that branch during this time, we need to schedule testing of that as soon as possible; otherwise, it will not be ready for the next kernel version.
... 3.0-RCx
As we get close to the next release, there will be an accumulation of new code in the drm-intel-next branch and we must ensure that it is ready before the merge window opens. Plus, as the release gets closer, we must also ensure that any patches we do make on drm-intel-fixes cause no further problems. At this point, drm-intel-fixes and drm-intel-next will be significantly different.
During later RC versions, we must split our attention between drm-intel-fixes to track issues in the next kernel version and drm-intel-next to ensure that code merged for the subsequent kernel are ready. By reducing the volume of patches to drm-intel-fixes, I hope to make this tractable.
3.0 And the merge window opens, drm-intel-next becomes our primary focus as that is merged to drm-next and then to master.
I've been working with Dave Airlie to figure out when I can move the base of the intel driver branches forward along master. Here's what I've got so far:
drm-intel-fixes
After the merge window closes and no further patches from drm-intel-next are anticipated, this branch will get pulled forward to -RC1. There will be nothing on this branch which is not in -RC1, so it will be a simple fast-forward (or something is messed up and needs fixing).
Subsequently, when a PULL request is made and has been merged to master, drm-intel-fixes will get fast-forwarded to that merge point. If no further patches are pending between the merge point and the subsequent -RC release, drm-intel-fixes will get fast forwarded to that -RC.
This includes pulling drm-intel-fixes forward to the final release in case we end up needing to prepare emergency patches for the .1 minor release.
The part I'm less sure about is how to deal with patches that affect both drm/i915 and drm itself. In that case, we may have patches on both drm-intel-fixes and drm-fixes which work together to resolve an issue. I think the right thing will be to have drm-intel-fixes get merged into drm and from there be merged into master. That's what I've done with a set of fixes posted after 3.0-RC1.
I'm interested in hearing comments about whether this will cause too many additional issues with merges from other parts of the kernel, or whether we'll end up far behind other trees somehow.
drm-intel-next
While the merge window is closed, this tree will be accumulating patches in preparation for the subsequent merge window. However, to reduce the divergence from drm-intel-fixes, I'm proposing that drm-intel-fixes get merged to drm-intel-next whenever it is merged to master. I fully expect this to cause merge conflicts, but resolving those within our own tree before they are sent upstream should reduce the conflicts at that level. As drm-intel-fixes is periodically fast-forwarded to points along master, those merges will get pulled across to drm-intel-next.
At the opening of the merge window, drm-intel-next should contain all of drm-intel-fixes from the release, and most of the rest of the release as well.
During the merge window, patches that have been reviewed and tested on drm-intel-next will be included in a PULL request to Dave Airlie for merging into the drm tree.
For patches which involve code in the core drm, we'll have those changes included in the drm-intel-next tree and they will get merged to drm-next from there.
Other feature branches
I do not plan to maintain any additional branches for specific feature work in the drm/i915 tree. Those should be done in personal repositories, where they can be rebased as necessary before being merged to the drm/i915 tree.
Of course, there will probably exceptions to this plan. Those should be clearly marked and suitable comments placed in commit messages.
Patches and commit messages
As I will be reviewing a significant number of patches heading to the kernel, I will be requiring substantive, complete and correct commit messages from developers. If the commit message does not include a detailed description of the problem and the fix, expect to get a request for more information. Please add any clarification to the commit message the next time the patch is submitted.
For inclusion in drm-intel-fixes, any patch should have an associated bug, and the commit message should include a link to the related bugzilla entry. Fixes for regressions should be marked clearly as such. During the later RC sequences, patches must be short and address the bug directly, without any fancy bits.
All patches and pull requests should be sent to the intel-gfx, dri-devel and linux-kernel mailing lists. Long sequences of patches are best done with a pull request instead; as review comments come in, the branch should be rebased so that it contains all of the updates, including Tested-by, Acked-by and Reviewed-by lines.
Pull requests should be based from the related tree, and branch from a point close to the current tip. Patches and pulls which don't merge cleanly should be fixed by the author and re-submitted.
On Sat, 2011-06-04 at 23:05 -0700, Keith Packard wrote:
I'm trying to formalize the process for merging code into the drm/i915 driver. Here's a first draft, please send along your comments.
Okay so you made the same mistake a lot of people make, the merge window for -next from me to Linus is from when he releases a .x release. However the merge window for you to me, should happen before this, multiple -next merges are allowed before Linus releases a kernel, though I'd really like to close the door around -rc6 in general. Generally I've found though with graphics we are crunch time for late found regression around rc6 time so I rarely push back too hard as I'd rather the kernel release with no regressions than worry overly about -next.
So when Linus is releasing rc4/5 you should really be cutting down stuff going into your -next, and getting the tree ready for me to take. We can probably add your tree direct to the -next git list as well so that we can get the benefits of -next testing earlier.
Some misc stuff inline:
<snip>
Yes if you have any interdepends then you need to coordinate with me, hopefully outside of -next we don't have that sort of work happening, as generally individual drm patches should be in my tree anyways.
Linus rule is to avoid misc merges, however if you are only getting fast forwards then it seems like it should be fine as we are getting people testing from a better base.
Just make sure any merge commits are documented in the commit, this generally means having to git commit --amend the merge commit since most of the time it doesn't allow you to write anything unless you get a conflict, esp if its a strange non i915 commit (like say some ACPI or bluetooth commit that fixes a test laptop).
Otherwise sounds good.
Dave.
On Sun, 05 Jun 2011 16:23:43 +1000, Dave Airlie airlied@redhat.com wrote:
Right, even earlier than I had thought. That should be pretty easy to follow.
Thanks for your comments; I'll try to have -next ready by rc4/rc5 time this release.
On Sat, 04 Jun 2011 23:05:23 -0700 "Keith Packard" keithp@keithp.com wrote:
Can you keep drm-intel-next fairly up to date with respect to the fixes branch? I.e. keep it a superset of drm-intel-fixes for the most part?
In PCI-land, this means re-basing my -next tree periodically before the merge window opens (though not right before the merge window unless something unexpected happens, like a patch needing to be dropped; in that case I'll delay the merge window push a bit to allow for more testing).
Downstream PCI developers requested this since it makes feature development much easier (you get all the bug fixes destined for Linus while working on a new feature for the next window), and as a downstream gfx developer I'd like to see this on the Intel side as well, unless other developers have big objections...
Thanks,
On Mon, 6 Jun 2011 13:36:18 -0700, Jesse Barnes jbarnes@virtuousgeek.org wrote:
Can you keep drm-intel-next fairly up to date with respect to the fixes branch? I.e. keep it a superset of drm-intel-fixes for the most part?
Yes, I wanted to do that now, but -fixes is not a fast-forward from -next and I thought I shouldn't be doing rebases.
If a rebase around -RC1 is reasonable, then I'll do that now and move changes over to that.
On Mon, 06 Jun 2011 16:24:46 -0700 Keith Packard keithp@keithp.com wrote:
You shouldn't if your downstream is using git trees and you're pulling from them, but it depends on your downstream. In my particular case, I'm ok with rebases if it means I get fixes. :)
On Mon, 6 Jun 2011 16:30:25 -0700 Jesse Barnes jbarnes@virtuousgeek.org wrote:
Oh and the other big reason is testing. rebase generally voids previous testing since you have new bits, so Linus really hates to see a rebase just before a pull request, and I think Dave does too.
But rebasing for good reason (e.g. to make your -next branch a superset of your -fixes branch) on occasion is fine.
dri-devel@lists.freedesktop.org
-
Dave Airlie -
Jesse Barnes -
Keith Packard