Re: [PATCH RFC 2/2] genirq/matrix: WARN_ON_ONCE() when cm->allocated/m->total_allocated go negative

[Vitaly Kuznetsov <vkuznets@redhat.com>]

Thomas Gleixner <tglx@linutronix.de> writes:

> On Fri, Feb 19 2021 at 12:31, Vitaly Kuznetsov wrote:
>
>> When irq_matrix_assign()/irq_matrix_free() calls get unsynced, weird
>> effects are possible, e.g. when cm->allocated goes negative CPU hotplug
>> may get blocked. Add WARN_ON_ONCE() to simplify detecting such situations.
>>
>> Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
>> ---
>>  kernel/irq/matrix.c | 11 +++++++++--
>>  1 file changed, 9 insertions(+), 2 deletions(-)
>>
>> diff --git a/kernel/irq/matrix.c b/kernel/irq/matrix.c
>> index 651a4ad6d711..2438a4f9d726 100644
>> --- a/kernel/irq/matrix.c
>> +++ b/kernel/irq/matrix.c
>> @@ -189,7 +189,9 @@ void irq_matrix_assign_system(struct irq_matrix *m, unsigned int bit,
>>  	set_bit(bit, m->system_map);
>>  	if (replace) {
>>  		BUG_ON(!test_and_clear_bit(bit, cm->alloc_map));
>> +		WARN_ON_ONCE(!cm->allocated);
>>  		cm->allocated--;
>> +		WARN_ON_ONCE(!m->total_allocated);
>
> This hunk is not really useful. It already dies when the bit is not set
> in the alloc map.

This was to check for the hypothetical issue when then number of bits
set get out of sync with 'total_allocated' counter -- which is likely
impossible today but could maybe be useful as a future proof. In case
this seems to be too much I'm not against dropping it.

>
>>  		m->total_allocated--;
>>  	}
>>  	if (bit >= m->alloc_start && bit < m->alloc_end)
>> @@ -424,12 +426,17 @@ void irq_matrix_free(struct irq_matrix *m, unsigned int cpu,
>>  		return;
>>  
>>  	clear_bit(bit, cm->alloc_map);
>> +	WARN_ON_ONCE(!cm->allocated);
>>  	cm->allocated--;
>
> WARN and then decrement is not necessarily any better than just
> decrementing unconditionally. It's just more noisy.
>
> Why would you let the counter wrap into negative space if you already
> know it's 0?
>
> There is a way more useful way to handle this. In such a case the bit is
> NOT set in the alloc map. So:
>
>     if (!WARN_ON_ONCE(test_and_clear_bit(bit, cm->alloc_map)))
>     	   return;
>
> would have caught the problem at hand nicely and let the machine survive
> while just throwing warns and continuing is broken to begin with.

Thanks, I like the idea. I didn't do that probably because the problem
which triggered me to write these patches wasn't fatal, it was just
causing CPU0 offlining to fail.

>
> Thanks,
>
>         tglx
>

-- 
Vitaly