On 1/5/20 8:02 AM, Sasha Levin wrote:
On Fri, Jan 03, 2020 at 06:50:45AM -0800, Guenter Roeck wrote:
On 1/2/20 4:40 PM, Sasha Levin wrote:
On Thu, Jan 02, 2020 at 01:28:37PM -0800, Guenter Roeck wrote:
On Thu, Jan 02, 2020 at 10:01:19PM +0100, Greg Kroah-Hartman wrote:
On Wed, Jan 01, 2020 at 06:44:08PM -0800, Guenter Roeck wrote:
Hi,
I see a number of crashes in the latest v5.4.y-queue; please see below
for details. The problem bisects to commit 54a311c5d3988d ("clk: Fix memory
leak in clk_unregister()").
The context suggests recovery from a failed driver probe, and it appears
that the memory is released twice. Interestingly, I don't see the problem
in mainline.
I would suggest to drop that patch from the stable queue.
That does not look right, as you point out, so I will go drop it now.
The logic of the clk structure lifetimes seems crazy, messing with krefs
and just "knowing" the lifecycle of the other structures seems like a
problem just waiting to happen...
I agree. While the patch itself seems to be ok per Stephen's feedback,
we have to assume that there will be more secondary failures in addition
to the one I have discovered. Given that clocks are not normally
unregistered, I don't think fixing the memory leak is important enough
to risk the stability of stable releases.
With all that in mind, I'd rather have this in mainline for a prolonged
period of time before considering it for stable release (if at all).
I would very much like to circle back and add both this patch and it's
fix to the stable trees at some point in the future.
If the code is good enough for mainline it should be good enough for
stable as well. If it's broken - let's fix it now instead of deferring
this to when people try to upgrade their major kernel versions.
This is where we differ strongly, and where I think the Linux community will
have to make a decision sometime soon. If "good enough for mainline" is a
relevant criteria for inclusion of a patch into stable releases, we don't
need stable releases anymore (we are backporting all bugs into those anyway).
Just use mainline.
Really, stable releases should be limited to fixing severe bugs. This is not
a fix for a severe bug, and on top of that it has side effects. True, those
side effects are that it uncovers other bugs, but that just makes it worse.
If we assume that my marginal testing covers, optimistically, 1% of the kernel,
and it discovers one bug, we have the potential of many more bugs littered
throughout the kernel which are now exposed. I really don't want to export
that risk into stable releases.
The assumption here is that fixes introduce less bugs than newly
introduced features, so I'd like to think that we're not backporting
*all* bugs :)
It's hard to define "severe" given how widely the kernel is used and all
the weird usecases it has. Something that doesn't look severe might be
very critical in a specific usecase. I fear that if we have a strict
definition of "severe", our users will end up carrying more patches
out-of-tree to fix their "less severe" issue, causing fragmantation
which we really want to avoid.
I actually belive very much in the suggestion you've made in your first
paragraph: I'd love to see LTS and later on -stable kernels go away and
users just use mainline releases. Yes, it's unrealistic now, but I'd
like to think that we're working towards it, thus I want to keep picking
up more patches and develop our (as well as our user's) testing muscle
to be able to catch regressions.
The result will be that more users will abandon upstream stable releases.
I already have trouble explaining why that many patches are being backported;
there is quite some internal grumbling about it (along the line of "this
patch should not have been backported").
I think we are going into the absolutely wrong direction. Expecting that
everyone would use mainline is absolutely unrealistic, and backporting
more and more patches to stable branches can only result in destabilizing
stable branches, which in turn will drive people away from it (and _not_
to use mainline). The only reason it wasn't a disaster yet is that we have
better testing now. But we offset that better testing with backporting
more patches, which has the opposite effect. One stabilizes, the other
destabilizes. The end result is the same. Actually, it is worse - the
indiscriminate backporting not only causes unnecessary regressions,
it (again) gives people an argument against merging stable releases.
And, this time I have to admit they are right.
Guenter
