Re: [RFC 0/2] guarantee natural alignment for kmalloc()

From: Vlastimil Babka
Date: Thu Mar 21 2019 - 03:42:15 EST


On 3/20/19 7:20 PM, Christopher Lameter wrote:
>>> Currently all kmalloc objects are aligned to KMALLOC_MIN_ALIGN. That will
>>> no longer be the case and alignments will become inconsistent.
>>
>> KMALLOC_MIN_ALIGN is still the minimum, but in practice it's larger
>> which is not a problem.
>
> "In practice" refers to the current way that slab allocators arrange
> objects within the page. They are free to do otherwise if new ideas come
> up for object arrangements etc.
>
> The slab allocators already may have to store data in addition to the user
> accessible part (f.e. for RCU or ctor). The "natural alighnment" of a
> power of 2 cache is no longer as you expect for these cases. Debugging is
> not the only case where we extend the object.

For plain kmalloc() caches, RCU and ctors don't apply, right.

>> Also let me stress again that nothing really changes except for SLOB,
>> and SLUB with debug options. The natural alignment for power-of-two
>> sizes already happens as SLAB and SLUB both allocate objects starting on
>> the page boundary. So people make assumptions based on that, and then
>> break with SLOB, or SLUB with debug. This patch just prevents that
>> breakage by guaranteeing those natural assumptions at all times.
>
> As explained before there is nothing "natural" here. Doing so restricts
> future features

Well, future features will have to deal with the existing named caches
created with specific alignment.

> and creates a mess within the allocator of exceptions for
> debuggin etc etc (see what happened to SLAB).

SLAB could be fixed, just nobody cares enough I guess. If I want to
debug wrong SL*B usage I'll use SLUB.

> "Natural" is just a
> simplistic thought of a user how he would arrange power of 2 objects.
> These assumption should not be made but specified explicitly.

Patch 1 does this explicitly for plain kmalloc(). It's unrealistic to
add 'align' parameter to plain kmalloc() as that would have to create
caches on-demand for 'new' values of align parameter.

>>> I think its valuable that alignment requirements need to be explicitly
>>> requested.
>>
>> That's still possible for named caches created by kmem_cache_create().
>
> So lets leave it as it is now then.

That however doesn't work well for the xfs/IO case where block sizes are
not known in advance:

https://lore.kernel.org/linux-fsdevel/20190225040904.5557-1-ming.lei@xxxxxxxxxx/T/#ec3a292c358d05a6b29cc4a9ce3ae6b2faf31a23f