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

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

[Arnd Bergmann]

From: Arnd Bergmann <arnd@arndb.de>

In general a 'struct device' is way too large to be put on the kernel
stack. Apparently something just caused it to grow a slightly larger,
which pushed the arm_lpae_do_selftests() function over the warning
limit in some configurations:

drivers/iommu/io-pgtable-arm.c:1423:19: error: stack frame size (1032) exceeds limit (1024) in 'arm_lpae_do_selftests' [-Werror,-Wframe-larger-than]
 1423 | static int __init arm_lpae_do_selftests(void)
      |                   ^

Change the function to use a dynamically allocated platform_device
instead of the on-stack device structure. The device is not actually
registered with the system, but is initialized enough to be used here.

Fixes: ca25ec247aad ("iommu/io-pgtable-arm: Remove iommu_dev==NULL special case")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
---
 drivers/iommu/io-pgtable-arm.c | 13 +++++++++----
 1 file changed, 9 insertions(+), 4 deletions(-)

diff --git a/drivers/iommu/io-pgtable-arm.c b/drivers/iommu/io-pgtable-arm.c
index 7632c80edea6..9f3bf0b5e8da 100644
--- a/drivers/iommu/io-pgtable-arm.c
+++ b/drivers/iommu/io-pgtable-arm.c
@@ -13,6 +13,7 @@
 #include <linux/bitops.h>
 #include <linux/io-pgtable.h>
 #include <linux/kernel.h>
+#include <linux/platform_device.h>
 #include <linux/sizes.h>
 #include <linux/slab.h>
 #include <linux/types.h>
@@ -1433,15 +1434,17 @@ static int __init arm_lpae_do_selftests(void)
 	};
 
 	int i, j, k, pass = 0, fail = 0;
-	struct device dev;
+	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);
+	if (!pdev)
+		return -ENOMEM;
+
+	cfg.iommu_dev = &pdev->dev;
 
 	for (i = 0; i < ARRAY_SIZE(pgsize); ++i) {
 		for (j = 0; j < ARRAY_SIZE(address_size); ++j) {
@@ -1461,6 +1464,8 @@ static int __init arm_lpae_do_selftests(void)
 	}
 
 	pr_info("selftest: completed with %d PASS %d FAIL\n", pass, fail);
+	platform_device_put(pdev);
+
 	return fail ? -EFAULT : 0;
 }
 subsys_initcall(arm_lpae_do_selftests);
-- 
2.39.5

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

[Robin Murphy]

On 11/04/2025 1:54 pm, Arnd Bergmann wrote:
> From: Arnd Bergmann <arnd@arndb.de>
> 
> In general a 'struct device' is way too large to be put on the kernel
> stack. Apparently something just caused it to grow a slightly larger,
> which pushed the arm_lpae_do_selftests() function over the warning
> limit in some configurations:
> 
> drivers/iommu/io-pgtable-arm.c:1423:19: error: stack frame size (1032) exceeds limit (1024) in 'arm_lpae_do_selftests' [-Werror,-Wframe-larger-than]
>   1423 | static int __init arm_lpae_do_selftests(void)
>        |                   ^
> 
> Change the function to use a dynamically allocated platform_device
> instead of the on-stack device structure. The device is not actually
> registered with the system, but is initialized enough to be used here.
> 
> Fixes: ca25ec247aad ("iommu/io-pgtable-arm: Remove iommu_dev==NULL special case")
> Signed-off-by: Arnd Bergmann <arnd@arndb.de>
> ---
>   drivers/iommu/io-pgtable-arm.c | 13 +++++++++----
>   1 file changed, 9 insertions(+), 4 deletions(-)
> 
> diff --git a/drivers/iommu/io-pgtable-arm.c b/drivers/iommu/io-pgtable-arm.c
> index 7632c80edea6..9f3bf0b5e8da 100644
> --- a/drivers/iommu/io-pgtable-arm.c
> +++ b/drivers/iommu/io-pgtable-arm.c
> @@ -13,6 +13,7 @@
>   #include <linux/bitops.h>
>   #include <linux/io-pgtable.h>
>   #include <linux/kernel.h>
> +#include <linux/platform_device.h>
>   #include <linux/sizes.h>
>   #include <linux/slab.h>
>   #include <linux/types.h>
> @@ -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...)

> +	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.

Thanks,
Robin.

> +	if (!pdev)
> +		return -ENOMEM;
> +
> +	cfg.iommu_dev = &pdev->dev;
>   
>   	for (i = 0; i < ARRAY_SIZE(pgsize); ++i) {
>   		for (j = 0; j < ARRAY_SIZE(address_size); ++j) {
> @@ -1461,6 +1464,8 @@ static int __init arm_lpae_do_selftests(void)
>   	}
>   
>   	pr_info("selftest: completed with %d PASS %d FAIL\n", pass, fail);
> +	platform_device_put(pdev);
> +
>   	return fail ? -EFAULT : 0;
>   }
>   subsys_initcall(arm_lpae_do_selftests);

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

[Arnd Bergmann]

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

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

[Robin Murphy]

On 11/04/2025 2:44 pm, Arnd Bergmann wrote:
> 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?

Sure, I'm happy to consciously err on the side of caution and 
robustness, just making sure :)

Thanks,
Robin.