Re: [PATCH 1/2] irq: sparseirq enabling v2

[Yinghai Lu <yinghai@kernel.org>]

Ingo Molnar wrote:
> * Yinghai Lu <yinghai@kernel.org> wrote:
> 
>> Ingo Molnar wrote:
>>> * Yinghai Lu <yinghai@kernel.org> wrote:
>>>
>>>> impact: new feature sparseirq
>>>> v2: use pointer array instead of hash
>>> ok, this looks pretty good!
>>>
>>> A few details:
>>>
>>>> +#ifdef CONFIG_SPARSE_IRQ
>>>> +#define set_ioapic_affinity_irq set_ioapic_affinity_irq_desc
>>>> +#else
>>>> +static void set_ioapic_affinity_irq(unsigned int irq, cpumask_t mask)
>>>> +{
>>>> +	struct irq_desc *desc;
>>>> +
>>>> +	desc = irq_to_desc(irq);
>>>> +
>>>> +	set_ioapic_affinity_irq_desc(desc, mask);
>>>> +}
>>>> +#endif
>>> i think this distinction can now go away?.
>> i may miss sth in your previous mail. you said we may put full hash 
>> back later, so need keep those change to avoid lookup costs later.
> 
> full hash as in no NR_IRQS limit?

32 bit irqs.

> 
> I dont think that's a good idea even in the far future: we can always 
> make NR_IRQS large enough, and turn the current array-of-pointers into 
> array-of-arrays-of-pointers scheme or an outright hash - but without 
> losing the fundamental abstraction that it's a 0..NR_IRQS array (just 
> accessed abstractly).

> 
>> also if we need move_irq_desc between node?
> 
> i'm not sure i understand. Do you mean we should deallocate and 
> reallocate the irq_desc to another node?

yes. the patch 2/2. [PATCH 2/2] irq: move irq_desc according to smp_affinity v2

> 
>>>> +#ifdef CONFIG_SPARSE_IRQ
>>>> +	irq_want = nr_irqs;
>>>> +#else
>>>> +	irq_want = NR_IRQS - 1;
>>>> +#endif
>>> ditto. I think we dont want 'nr_irqs' anymore - just remain with 
>>> NR_IRQS, right?
>>>
>> nr_irqs is the total GSI number when sparseirq is used. so MSI irq 
>> will start from that. ...
> 
> ah, okay. I think we should renumber MSIs to the scheme Thomas 
> suggested, independently of CONFIG_SPARSE_IRQ.

good, will add one local nr_irqs_gsi there..., and not touch nr_irqs.

> 
> CONFIG_SPARSE_IRQ should _only_ impact the irq_desc[] data structure: 
> it turns from a flat C array of irq_desc into an array of irq_desc 
> pointers.
> 
> No other semantic changes. Possibly all hidden away completely in the 
> irq-desc-array accessors.

YH