kernel: userspace: aligned memory allocation for dynamic objects

[dcpleung]

This allows allocating dynamic kernel objects with memory alignment
requirements. The first candidate is for thread objects for x86
and x86_64 where saving/restoring SSE/FPU registers require strict
alignment.

Fixes #29589
Fixes #29629

[cfriedt]

Looks like there is some potential conflict with #30363 which has been open for a while in preparation for dynamic pthtread stacks (#29029, which has been there for quite a long time too).

@dcpleung - I wonder if this ticket could focus more on dynamic kobject creation while #30363 could focus more on allocation.

Please let me know if you can think of a better alternative.

[dcpleung]

Looks like there is some potential conflict with #30363 which has been open for a while in preparation for dynamic pthtread stacks (#29029, which has been there for quite a long time too).

@dcpleung - I wonder if this ticket could focus more on dynamic kobject creation while #30363 could focus more on allocation.

Please let me know if you can think of a better alternative.

Once your PR is merged, I can rebase this on top of your changes.

[cfriedt]

Looks good - just a slight change request to have consistent naming. Are any additional tests required?

[npitre]

On Tue, 12 Jan 2021, Christopher Friedt wrote: + return z_dynamic_object_aligned_create(1, size); I think elsewhere that `sizeof(void*)` is used in place of an unspecified alignment because it's a reasonable safe default alignment.

Personally, I'd prefer using 0 when no particular alignment is necessary. The code will already take care to return memory with the minimum alignment for all basic types for a given platform in that case.

[dcpleung]

On Tue, 12 Jan 2021, Christopher Friedt wrote: + return z_dynamic_object_aligned_create(1, size); I think elsewhere that sizeof(void*) is used in place of an unspecified alignment because it's a reasonable safe default alignment.
Personally, I'd prefer using 0 when no particular alignment is necessary. The code will already take care to return memory with the minimum alignment for all basic types for a given platform in that case.

Hm... sys_heap_aligned_alloc() with alignment 0 simply calls sys_heap_alloc() which has no minimal alignment (AFAIK). So probably better to pass a value over.

[dcpleung]

Looks good - just a slight change request to have consistent naming. Are any additional tests required?

Updated the naming and alignment value.

[npitre]

On Tue, 12 Jan 2021, Daniel Leung wrote: Hm... `sys_heap_aligned_alloc()` with alignment `0` simply calls `sys_heap_alloc()` which has no minimal alignment (AFAIK). So probably better to pass a value over.

It sure does. sys_heap_alloc() must have minimal alignment for the platform no matter what. And I know it does.

[dcpleung]

On Tue, 12 Jan 2021, Daniel Leung wrote: Hm... sys_heap_aligned_alloc() with alignment 0 simply calls sys_heap_alloc() which has no minimal alignment (AFAIK). So probably better to pass a value over.
It sure does. sys_heap_alloc() must have minimal alignment for the platform no matter what. And I know it does.

Changed the value to 0. There is another issue where they reported the minimal alignment was 4, but I recalled it as 1, hence the confusion.

[npitre]

On Tue, 12 Jan 2021, Daniel Leung wrote: Changed the value to `0`. There is another issue where they reported the minimal alignment was 4, but I recalled it as 1, hence the confusion.

Thanks. There is a 1 in `z_thread_malloc()` that could be turned into a 0 as well for consistency.

[dcpleung]

On Tue, 12 Jan 2021, Daniel Leung wrote: Changed the value to 0. There is another issue where they reported the minimal alignment was 4, but I recalled it as 1, hence the confusion.
Thanks. There is a 1 in z_thread_malloc() that could be turned into a 0 as well for consistency.

Changed that too.

[cfriedt]

Works for me