On Sat, Mar 11, 2023 at 10:54:36AM -0500, James Bottomley wrote:
On Sat, 2023-03-11 at 08:41 -0500, Sasha Levin wrote:
On Fri, Mar 10, 2023 at 03:07:04PM -0800, Eric Biggers wrote:
On Mon, Feb 27, 2023 at 07:41:31PM -0800, Eric Biggers wrote:
Well, probably more common is that prerequisites are in the same patchset, and the prerequisites are tagged for stable too. Whereas AUTOSEL often just picks patch X of N. Also, developers and maintainers who tag patches for stable are probably more likely to help with the stable process in general and make sure patches are backported correctly...
Anyway, the point is, AUTOSEL needs to be fixed to stop inappropriately cherry-picking patch X of N so often.
... and AUTOSEL strikes again, with the 6.1 and 6.2 kernels currently crashing whenever a block device is removed, due to patches 1 and 3 of a 3-patch series being AUTOSEL'ed (on the same day I started this discussion, no less):
https://lore.kernel.org/linux-block/CAOCAAm4reGhz400DSVrh0BetYD3Ljr2CZen7_3D...
Oh sorry, ignore this, it's just an anecdotal example.
Yes, clearly a problem with AUTOSEL and not with how sad the testing story is for stable releases.
Hey, that's a completely circular argument: if we had perfect testing then, of course, it would pick up every bad patch before anything got released; but we don't, and everyone knows it. Therefore, we have to be discriminating about what patches we put in. And we have to acknowledge that zero bugs in patches isn't feasible in spite of all the checking we do do. I also think we have to acknowledge that users play a role in the testing process because some bugs simply aren't picked up until they try out a release. So discouraging users from running mainline -rc's means we do get bugs in the released kernel that we might not have had if they did. Likewise if everyone only runs stable kernels, the bugs in the released kernel don't get found until stable. So this blame game really isn't helping.
I think the one thing everyone on this thread might agree on is that this bug wouldn't have happened if AUTOSEL could detect and backport series instead of individual patches. Sasha says that can't be done based on in information in Linus' tree[1] which is true but not a correct statement of the problem. The correct question is given all the information available, including lore, could we assist AUTOSEL in better detecting series and possibly making better decisions generally?
My argument was that this type of issue is no AUTOSEL specific, and we saw it happening multiple times with stable tagged patches as well.
It's something that needs to get solved, and I suspect that both Greg and myself will end up using it when it's there.
I think that's the challenge for anyone who actually wants to help rather than complain. At least the series detection bit sounds like it could be a reasonable summer of code project.
Right - I was trying to reply directly to Willy's question: this is something very useful, somewhat hard, and I don't think I could do in the near future - so help is welcome here.