Re: PCI MSI issue with reinserting a driver

[John Garry <john.garry@huawei.com>]

On 03/02/2021 17:23, Marc Zyngier wrote:
> On 2021-02-02 15:46, John Garry wrote:
>> On 02/02/2021 14:48, Marc Zyngier wrote:
>>>>>
>>>>> Not sure. I also now notice an error for the SAS PCI driver on D06 
>>>>> when nr_cpus < 16, which means number of MSI vectors allocated < 
>>>>> 32, so looks the same problem. There we try to allocate 16 + max(nr 
>>>>> cpus, 16) MSI.
>>>>>
>>>>> Anyway, let me have a look today to see what is going wrong.
>>>>>
>>>> Could this be the problem:
>>>>
>>>> nr_cpus=11
>>>>
>>>> In alloc path, we have:
>>>>     its_alloc_device_irq(nvecs=27 = 16+11)
>>>>       bitmap_find_free_region(order = 5);
>>>> In free path, we have:
>>>>     its_irq_domain_free(nvecs = 1) and free each 27 vecs
>>>>       bitmap_release_region(order = 0)
>>>>
>>>> So we allocate 32 bits, but only free 27. And 2nd alloc for 32 fails.
>>
>> [ ... ]
>>
>>>>

Hi Marc,

Just a friendly reminder on this issue. Let me know if anything required 
some our side.

Cheers,
John

>>>>
>>>> But I'm not sure that we have any requirement for those map bits to be
>>>> consecutive.
>>>
>>> We can't really do that. All the events must be contiguous,
>>> and there is also a lot of assumptions in the ITS driver that
>>> LPI allocations is also contiguous.
>>>
>>> But there is also the fact that for Multi-MSI, we *must*
>>> allocate 32 vectors. Any driver could assume that if we have
>>> allocated 17 vectors, then there is another 15 available.
>>>
>>> My question still stand: how was this working with the previous
>>> behaviour?
>>
>> Because previously in this scenario we would allocate 32 bits and free
>> 32 bits in the map; but now we allocate 32 bits, yet only free 27 - so
>> leak 5 bits. And this comes from how irq_domain_free_irqs_hierarchy()
>> now frees per-interrupt, instead of all irqs per domain.
>>
>> Before:
>>  In free path, we have:
>>      its_irq_domain_free(nvecs = 27)
>>        bitmap_release_region(count order = 5 == 32bits)
>>
>> Current:
>>  In free path, we have:
>>      its_irq_domain_free(nvecs = 1) for free each 27 vecs
>>        bitmap_release_region(count order = 0 == 1bit)
> 
> Right. I was focusing on the patch and blindly ignored the explanation
> at the top of the email. Apologies for that.
> 
> I'm not overly keen on handling this in the ITS though, and I'd rather
> we try and do it in the generic code. How about this (compile tested
> only)?
> 
> Thanks,
> 
>          M.
> 
> diff --git a/kernel/irq/irqdomain.c b/kernel/irq/irqdomain.c
> index 6aacd342cd14..cfccad83c2df 100644
> --- a/kernel/irq/irqdomain.c
> +++ b/kernel/irq/irqdomain.c
> @@ -1399,8 +1399,19 @@ static void irq_domain_free_irqs_hierarchy(struct 
> irq_domain *domain,
>           return;
> 
>       for (i = 0; i < nr_irqs; i++) {
> -        if (irq_domain_get_irq_data(domain, irq_base + i))
> -            domain->ops->free(domain, irq_base + i, 1);
> +        int n ;
> +
> +        /* Find the largest possible span of IRQs to free in one go */
> +        for (n = 0;
> +             ((i + n) < nr_irqs &&
> +              irq_domain_get_irq_data(domain, irq_base + i + n));
> +             n++);
> +
> +        if (!n)
> +            continue;
> +
> +        domain->ops->free(domain, irq_base + i, n);
> +        i += n;
>       }
>   }
> 
>