Re: Random panic in load_balance() with 3.16-rc

[Linus Torvalds]

On Wed, Jul 23, 2014 at 10:12 AM, Linus Torvalds
<torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
>
> A cpumask_var is TOTALLY DIFFERENT. It's *either* a cpumask _or_ just
> a pointer to an externally allocated cpumask.
>
> sched_init() definitely does _not_ allocate a cpumask_var.

I take that back. It does end up allocating it properly, it just
avoids all the correct abstractions.

In general, the rule of thumb should be:

 - stack allocations should use "cpumask_var_t cpus" and they
absolutely *have* to be paired with an "alloc_cpumask_var(&cpus,
GFP_KERNEL)". Having a "struct cpumask" on stack is very wrong.

 - random single nonstack allocations should probably just use a plain
"struct cpumask" (or cpumask_t, but we really shouldn't use typedef's
unless they actively abstract some per-config *changing* type).

 - dynamic allocations that are size-conscious (because there's a lot
of them) should allocate a "struct cpumask *" by using
"cpumask_size()". They have a pointer anyway, they allocate things
dynamically anyway, extra indirection through a cpumask_var_t would
just be unnecessary.

 - *static* per-cpu allocations might want to use "cpumask_var_t" (to
avoid having a full "struct cpumask_t") along with doing a
"zalloc_cpumask_var_node(..)" for each cpu.

sched_init() follows that last pattern, except it open-codes that
zalloc_cpumask_var_node() in an odd way that confused me.

So I take my patch back. It's wrong, because it only allocates that
cpumask_size() if CONFIG_CPUMASK_OFFSTACK is true.

Ugh, that code really is unreadable.

               Linus
--
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/