Re: [PATCH] iommu/io-pgtable-arm: dynamically allocate selftest device struct

["Arnd Bergmann" <arnd@arndb.de>]

On Fri, Apr 11, 2025, at 15:19, Robin Murphy wrote:
> On 11/04/2025 1:54 pm, Arnd Bergmann wrote:
>> @@ -1433,15 +1434,17 @@ static int __init arm_lpae_do_selftests(void)
>>   	};
>>   
>>   	int i, j, k, pass = 0, fail = 0;
>> -	struct device dev;
>
> Could we not simply make this static? Per the comment it's only here to 
> serve a NUMA node lookup buried deep in the pagetable allocator (TBH my 
> first thought was to just put an int on the stack and contrive a pointer 
> as the inverse of dev_to_node(), but I decided that would probably be 
> too contentious...)

A static device would work here, but that has other (small)
downsides:

 - static devices are discouraged for any real purpose because
   of the problematic lifetime rules. I think Greg would still
   want to eliminate these entirely.

 - there is slightly more memory usage: the __init function
   gets eliminated after boot, while a static allocation says
   around. It could perhaps be made __initdata.

 - If we ever need anything beyond the NUMA node from it, the
   dynamic allocation is probably close enough to make that
   work.

>> +	struct platform_device *pdev;
>>   	struct io_pgtable_cfg cfg = {
>>   		.tlb = &dummy_tlb_ops,
>>   		.coherent_walk = true,
>> -		.iommu_dev = &dev,
>>   	};
>>   
>> -	/* __arm_lpae_alloc_pages() merely needs dev_to_node() to work */
>> -	set_dev_node(&dev, NUMA_NO_NODE);
>> +	pdev = platform_device_alloc("io-pgtable-test", 0);
>
> Otherwise, this would seem to be another perfect case for the new 
> faux_device.

Good point, that is clearly better than platform_device in this
case. Shall I send a new version with that?

     Arnd