On 8/26/19 11:33 AM, Greg KH wrote:
On Mon, Aug 26, 2019 at 11:23:58AM +0300, Nikolai Kondrashov wrote:
On 8/25/19 5:41 PM, Greg KH wrote:
On Sun, Aug 25, 2019 at 10:37:26AM -0400, CKI Project wrote:
Merge testing
-------------
We cloned this repository and checked out the following commit:
Repo: https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git
Commit: f7d5b3dc4792 - Linux 5.2.10
We grabbed the cc88f4442e50 commit of the stable queue repository.
We then merged the patchset with `git am`:
keys-trusted-allow-module-init-if-tpm-is-inactive-or-deactivated.patch
That file is not in the repo, I think your system is messed up :(
Sorry for the trouble, Greg, but I think it's a race between the changes to
the two repos.
The job which triggered this message was started right before the moment this
commit was made:
https://git.kernel.org/pub/scm/linux/kernel/git/stable/stable-queue.git/commit/?id=af2f46e26e770b3aa0bc304a13ecd24763f3b452
At that moment, the repo was still on this commit, about five hours old:
https://git.kernel.org/pub/scm/linux/kernel/git/stable/stable-queue.git/commit/?id=cc88f4442e505e9f1f21c8c119debe89cbf63ab2
which still had the file. And when the job finished, and the message reached
you, yes, the repo no longer contained it.
At the moment the job started, the latest commit to stable/linux.git
was about 22 minutes old:
https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/commit/?h=linux-5.2.y&id=f7d5b3dc4792a5fe0a4d6b8106a8f3eb20c3c24c
and the repo already contained the patches from the queue, including the one
the job tried to merge:
https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/commit/?h=linux-5.2.y&id=f820ecf609cc38676071ec6c6d3e96b26c73b747
How in the world are you seeing such a messed up tree?
The 5.2.10 commit moved things around, in one single atomic move.
IIRC, we agreed to not start testing both of the repos until the latest
commits are at least 5 minutes old. In this situation the latest commit was 22
minutes old, so the system started testing.
We could increase the window to, say, 30 minutes (or something else), to avoid
misfires like this, but then the response time would be increased accordingly.
It's your pick :)
Why is there any race at all?
Why do you not have a local mirror of the repo? When it updates, then
run the tests. Every commit in the tree is "stand alone" and things
should work at that point in time. Don't use a commit as a "time to go
mirror something at a later point in time", as you are ending up with
trees that are obviously not correct at all.
I think you need to rework your systems as no one else seems to have
this "stale random tree state" issue.
Git does commits in an atomic fashion, how you all are messing that up
shows you are doing _way_ more work than you probably need to :)
Sorry, I'm not the one who implemented and maintains the system, I'm just
generally aware of how it works and am looking at the code right now, so I
could be misunderstanding something. Please bear with me :)
However, I don't see how anything could be done, if we have two git repos,
which are inconsistent with each other, when CI comes to test them.
I'll try to draw the timeline of what was happening to explain what I think is
the problem. All times are in my timezone (UTC+03:00).
Time stable/linux.git stable/stable-queue.git Comments
branch linux-5.2.y branch master
subdir queue-5.2
--------------- ------------------- ----------------------- -----------------
Aug 5 19:44:27 aad39e30fb9e6e72, Repos are
"Linux 5.2.9", consistent
*doesn't have* the
patch that failed
Aug 25 11:53:25 cc88f4442e505e9f, Repos are
"Linux 4.4.190", consistent
*has* the patch
that failed
Aug 25 17:13:54 f7d5b3dc4792a5, Repos are
"Linux 5.2.10", inconsistent,
contains patches both contain
from the queue the same patches
above, including
the failed one
Aug 25 17:36:18 Our CI job starts
Aug 25 17:36:19 af2f46e26e770b3a Repos are
"Linux 5.2.10", consistent
"queue-5.2" dir is
removed, doesn't
have the failed
patch
Aug 25 17:37:23 Our CI sends
failure report
I.e. I think the problem was that both linux-5.2.y branch of stable/linux.git,
and the queue-5.2 subdir of master branch of stable/stable-queue.git contained
the same patches for about 22 minutes on Aug 25, when our CI started.
We sample the latest commits from both repos at the same time (well, as close
as Python and HTTP allow us), and we update our clones to those before
testing.
We also don't start testing if the commits in either are less than 5 minutes
old to avoid testing inconsistent repos, assuming that 5 minutes are enough to
update them both to keep them in consistency. We can increase that time to
what you think best fits your workflow, to avoid hitting these problems.
Nick
