On 2021-06-14 13:33:43 [+0200], Vlastimil Babka wrote:

On 6/14/21 1:16 PM, Sebastian Andrzej Siewior wrote:

quoted

I haven't looked at the series and I have just this tiny question: why
did migrate_disable() crash for Mel on !RT and why do you expect that it
does not happen on PREEMPT_RT?

Right, so it's because __slab_alloc() has this optimization to avoid re-reading
'c' in case there is no preemption enabled at all (or it's just voluntary).

#ifdef CONFIG_PREEMPTION
        /*
         * We may have been preempted and rescheduled on a different
         * cpu before disabling preemption. Need to reload cpu area
         * pointer.
         */
        c = slub_get_cpu_ptr(s->cpu_slab);
#endif

Mel's config has CONFIG_PREEMPT_VOLUNTARY, which means CONFIG_PREEMPTION is not
enabled.

But then later in ___slab_alloc() we have

        slub_put_cpu_ptr(s->cpu_slab);
        page = new_slab(s, gfpflags, node);
        c = slub_get_cpu_ptr(s->cpu_slab);

And this is not hidden under CONFIG_PREEMPTION, so with the #ifdef bug the
slub_put_cpu_ptr did a migrate_enable() with Mel's config, without prior
migrate_disable().

Ach, right. The update to this field is done with cmpxchg-double (if I
remember correctly) but I don't remember if this is also re-entry safe. 

If there wasn't the #ifdef PREEMPT_RT bug:
- this slub_put_cpu_ptr() would translate to put_cpu_ptr() thus
preempt_enable(), which on this config is just a barrier(), so it doesn't matter
that there was no matching preempt_disable() before.
- with PREEMPT_RT the CONFIG_PREEMPTION would be enabled, so the
slub_get_cpu_ptr() would do a migrate_disable() and there's no imbalance.

But now that I dig into this in detail, I can see there might be another
instance of this imbalance bug, if CONFIG_PREEMPTION is disabled, but
CONFIG_PREEMPT_COUNT is enabled, which seems to be possible in some debug
scenarios. Because then preempt_disable()/preempt_enable() still manipulate the
preempt counter and compiling them out in __slab_alloc() will cause imbalance.

So I think the guards in __slab_alloc() should be using CONFIG_PREEMPT_COUNT
instead of CONFIG_PREEMPT to be correct on all configs. I dare not remove them
completely :)

:)

Sebastian