On 05/10/2018 12:35 AM, Sebastian Andrzej Siewior wrote:
> On 2018-05-08 16:02:57 [-0700], Andrew Morton wrote:
>> On Mon, 7 May 2018 09:31:05 +0200 Vlastimil Babka <vbabka@xxxxxxx> wrote:
>>
>>> In any case I agree that the revert should be done immediately even
>>> before fixing the underlying bug. The preempt_disable/enable doesn't
>>> prevent the bug, it only prevents the debugging code from actually
>>> reporting it! Note that it's debugging code (CONFIG_DEBUG_PREEMPT) that
>>> production kernels most likely don't have enabled, so we are not even
>>> helping them not crash (while allowing possible data corruption).
>>
>> Grumble.
>>
>> I don't see much benefit in emitting warnings into end-users' logs for
>> bugs which we already know about.
>
> not end-users (not to mention that neither Debian Stretch nor F28 has
> preemption enabled in their kernels). And if so, they may provide
> additional information for someone to fix the bug in the end. I wasn't
Even if end-users have enabled preemption, they likely won't have
enabled CONFIG_DEBUG_PREEMPT anyway.
> able to reproduce the bug but I don't have access to anything MIPSish
> where I can boot my own kernels. At least two people were looking at the
> code after I posted the revert and nobody spotted the bug.
>
>> The only thing this buys us is that people will hassle us if we forget
>> to fix the bug, and how pathetic is that? I mean, we may as well put
>>
>> printk("don't forget to fix the vmstat_update() bug!\n");
>
> No that is different. That would be seen by everyone. The bug was only
> reported by Steven J. Hill which did not respond since. This message
> would also imply that we know how to fix the bug but didn't do it yet
> which is not the case. We seen that something was wrong but have no idea
> *how* it got there.
>
> The preempt_disable() was added by the end of v4.16. The
> smp_processor_id() in vmstat_update() was added in commit 7cc36bbddde5
> ("vmstat: on-demand vmstat workers V8") which was in v3.18-rc1. The
> hotplug rework took place in v4.10-rc1. And it took (counting from the
> hotplug rework) 6 kernel releases for someone to trigger that warning
> _if_ this was related to the hotplug rework.
>
> What we have *now* is way worse: We have a possible bug that triggered
> the warning. As we see in report the code in question was _already_
> invoked on the wrong CPU. The preempt_disable() just silences the
> warning, hiding the real issue so nobody will do a thing about it since
> it will be never reported again (in a kernel with preemption and debug
> enabled).
Fully agree with everything you said!
We could extend the warning to e.g. print affinity mask of the thread,
and e.g. state of cpus that are subject to ongoing hotplug/hotremove.
But maybe it's not so useful in general, as the common case is likely
indeed a missing preempt_disable, and this is an exception? In any case,
I would hope that Steven applies some patch locally and we get more
details about what's going on at that MIPS machine.
