On Fri, Aug 09, 2019 at 03:04:49PM +0200, Thierry Reding wrote:
On Fri, Aug 09, 2019 at 10:52:53AM +0200, Greg Kroah-Hartman wrote:
On Mon, Aug 05, 2019 at 01:48:21PM +0200, Thierry Reding wrote:
Hi Greg,
Sorry for the delay, this got pushed down my queue...
I stumbled across something as I was attempting to automate more parts of our process to generate these reports. The original test results were from a different version of the tree: 5.2.6-rc1-gdbc7f5c7df28. I suspect that's the same thing that you were discussing with Pavel regarding the IP tunnel patch that was added subsequent to the announcement.
Just for my understanding, does this mean that the patch still makes it into the 5.2.6 release, or was it supposed to go into 5.2.7?
One problem that I ran into was that when I tried to correlate the test results with your announcement email, there's no indication other than the branch name and the release candidate name (5.2.6-rc1 in this case), so there's no way to uniquely identify which test run belongs to the announcement. Given that there are no tags for the release candidates means that that's also not an option to uniquely associate with the builds and tests.
While the differences between the two builds are very minor here, I wonder if there could perhaps in the future be a problem where I report successful results for a test, but the same tests would be broken by a patch added to the stable-rc branch subsequent to the announcement. The test report would be misleading in that case.
I noticed that you do add a couple of X-KernelTest-* headers to your announcement emails, so I'm wondering if perhaps it was possible for you to add another one that contains the exact SHA1 that corresponds to the snapshot that's the release candidate. That would allow everyone to uniquely associate test results with a specific release candidate.
That said, perhaps I've just got this all wrong and there's already a way to connect all the dots that I'm not aware of. Or maybe I'm being too pedantic here?
You aren't being pedantic, I think you are missing exactly what the linux-stable-rc tree is for and how it is created.
Granted, it's not really documented anywhere so it's not your fault :)
The linux-stable-rc tree is there ONLY for people who want to test the -rc kernels and can not, or do not want to, use the quilt tree of patches in the stable-queue.git tree on kernel.org. I generate the branches there from a script that throws away the current -rc branch and rebuilds it "from scratch" by applying the patches that are in the stable-quilt tree and then adds the -rc patch as well. This tree is generated multiple times when I am working on the queues and then when I want to do a "real" -rc release.
The branches are constantly rebased, so you can not rely on 'git pull' doing the right thing (unless you add --rebase to it), and are there for testing only.
Yes, you will see different results of a "-rc1" release in the tree depending on the time of day/week when I created the tree, and sometimes I generate them multiple times a day just to have some of the auto-builders give me results quickly (Linaro does pull from it and sends me results within the hour usually).
So does that help? Ideally everyone would just use the quilt trees from stable-queue.git, but no everyone likes that, so the -rc git tree is there to make automated testing "easier". If that works with your workflow, wonderful, feel free to use it. If not, then go with the stable-quilt.git tree, or the tarballs on kernel.org.
I'll have to look into that, to see if that'd work. I have to admit that having a git tree to point scripts at is rather convenient for automation.
And as for the SHA1 being in the emails, that doesn't make all that much sense as that SHA1 doesn't live very long. When I create the trees locally, I instantly delete them after pushing them to kernel.org so I can't regenerate them or do anything with them.
Fair enough. I suppose the worst thing that could happen is that we may fail to test a couple of commits occasionally. In the rare case where this actually matters we'll likely end up reporting the failure for the stable release, in which case it can be fixed in the next one.
DOes that help explain things better?
Yes, makes a lot more sense now. Thanks for taking the time to explain it. Do you want me to snapshot this and submit it as documentation somewhere for later reference?
It probably should go somewhere, but as the number of people that do "test stable kernels in an automated way" is very low, so far I've been doing ok with a one-by-one explaination. I guess if it's more obvious, more people would test, so sure, this should be cleaned up...
thanks,
greg k-h