Re: [PATCH] atomic: Specify natural alignment for atomic_t

[Lance Yang <lance.yang@linux.dev>]

Hi Geert,

On 2025/9/1 16:45, Geert Uytterhoeven wrote:
> Hi Lance,
> 
> On Thu, 28 Aug 2025 at 04:05, Lance Yang <lance.yang@linux.dev> wrote:
>> On 2025/8/28 07:43, Finn Thain wrote:
>>> On Mon, 25 Aug 2025, Lance Yang wrote:
>>>> Same here, using a global static variable instead of a local one. The
>>>> result is consistently misaligned.
>>>>
>>>> ```
>>>> #include <linux/module.h>
>>>> #include <linux/init.h>
>>>>
>>>> static struct __attribute__((packed)) test_container {
>>>>       char padding[49];
>>>>       struct mutex io_lock;
>>>> } cont;
>>>>
>>>> static int __init alignment_init(void)
>>>> {
>>>>       pr_info("Container base address      : %px\n", &cont);
>>>>       pr_info("io_lock member address      : %px\n", &cont.io_lock);
>>>>       pr_info("io_lock address offset mod 4: %lu\n", (unsigned long)&cont.io_lock % 4);
>>>>       return 0;
>>>> }
>>>>
>>>> static void __exit alignment_exit(void)
>>>> {
>>>>       pr_info("Module unloaded\n");
>>>> }
>>>>
>>>> module_init(alignment_init);
>>>> module_exit(alignment_exit);
>>>> MODULE_LICENSE("GPL");
>>>> MODULE_AUTHOR("x");
>>>> MODULE_DESCRIPTION("x");
>>>> ```
>>>>
>>>> Result from dmesg:
>>>>
>>>> ```
>>>> [Mon Aug 25 19:33:28 2025] Container base address      : ffffffffc28f0940
>>>> [Mon Aug 25 19:33:28 2025] io_lock member address      : ffffffffc28f0971
>>>> [Mon Aug 25 19:33:28 2025] io_lock address offset mod 4: 1
>>>> ```
>>>>
>>>
>>> FTR, I was able to reproduce that result (i.e. static storage):
>>>
>>> [    0.320000] Container base address      : 0055d9d0
>>> [    0.320000] io_lock member address      : 0055da01
>>> [    0.320000] io_lock address offset mod 4: 1
>>>
>>> I think the experiments you sent previously would have demonstrated the
>>> same result, except for the unpredictable base address that you sensibly
>>> logged in this version.
>>
>> Thanks for taking the time to reproduce it!
>>
>> This proves the problem can happen in practice (e.g., with packed structs),
>> so we need to ignore the unaligned pointers on the architectures that don't
>> trap for now.
> 
> Putting locks inside a packed struct is definitely a Very Bad Idea
> and a No Go.  Packed structs are meant to describe memory data and

Right. That's definitely not how packed structs should be used ;)

> MMIO register layouts, and must not contain control data for critical
> sections.

Unfortunately, this patten was found in an in-tree driver, as reported[1]
by kernel test robot, and there might be other undiscovered instances ...

[1] 
https://lore.kernel.org/oe-kbuild-all/202508240539.ARmC1Umu-lkp@intel.com

Cheers,
Lance