Re: [PATCH 0/2] incorrect cpumask behavior with CPUMASK_OFFSTACK

[Oleg Drokin]

Hello!

On Feb 27, 2015, at 6:46 AM, Rusty Russell wrote:

> green@xxxxxxxxxxxxxx writes:
>> From: Oleg Drokin <green@xxxxxxxxxxxxxx>
>> 
>> I just got a report today from Tyson Whitehead <twhitehead@xxxxxxxxx>
>> that Lustre crashes when CPUMASK_OFFSTACK is enabled.
>> 
>> A little investigation revealed that this code:
>>        cpumask_t                       mask;
>> ...
>>        cpumask_copy(&mask, topology_thread_cpumask(0));
>>        weight = cpus_weight(mask);
> Yes.  cpumask_weight should have been used here.  The old cpus_* are
> deprecated.

Oh! I guess we missed the announcement.
I'll convert it over.

Should I also do a patch converting all other users and removing the deprecated
functions while I am at it?

>> The second patch that I am not sure if we wnat, but it seems to be useful
>> until struct cpumask is fully dynamic is to convert what looks like
>> whole-set operations e.g. copies, namely:
>> cpumask_setall, cpumask_clear, cpumask_copy to always operate on NR_CPUS
>> bits to ensure there's no stale garbage left in the mask should the
>> cpu count increases later.
> You can't do this, because dynamically allocated cpumasks don't have
> NR_CPUS bits.

Well, right now they certainly have. As in, it's a static define and we allocate
a bitmap to fit the (in my case) up to 8192 bits into such off-stack masks.

The concern is since number of cpus is not really a fixed variable, when you
do cpumask initialization, and then number of cpus grows, the end of the mask
could be garbage? Am I overthinking this and it's not really a problem?

> Let's just kill all the cpus_ functions.  This wasn't done originally
> because archs which didn't care about offline cpumasks didn't want the
> churn.  In particular, they must not copy struct cpumask by assignment,
> and fixing those is a fair bit of churn.

Ah, copy masks by assignment, I see.
Well, I guess we can eliminate the users outside of the affected arch trees
(I assume in x86 there would be no objections?) and perhaps add a warning to
checkpatch.pl?

Thanks!

Bye,
    Oleg--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/