Re: [RFC PATCH] slab: don't assume alignment on allocators that may return ZERO_SIZE_PTR

[Vlastimil Babka (SUSE)]

On 7/13/26 07:18, Harry Yoo wrote:
> 
> [+Cc Catalin for ARCH_KMALLOC_MINALIGN bits]
> 
> On 7/12/26 9:07 PM, Karl Mehltretter wrote:
>> __kmalloc_noprof(), __kmalloc_node_noprof(), and __kmalloc_flags_noprof()
>> are annotated with __assume_kmalloc_alignment, which expands to
>> __assume_aligned(ARCH_KMALLOC_MINALIGN). All three can return
>> ZERO_SIZE_PTR, currently (void *)16, for zero-sized requests. When
>> ARCH_KMALLOC_MINALIGN exceeds 16, this contradicts the annotation.
> 
> Ouch.
> 
>> Compilers may use this false assumption to reduce ZERO_OR_NULL_PTR() to a
>> NULL check, losing recognition of ZERO_SIZE_PTR. Current GCC and Clang
>> retain the existing range check through kmalloc's inline wrapper, so no
>> functional miscompile was reproduced with the current source form. However,
>> both eliminate an exact ZERO_SIZE_PTR comparison on the same return value.
>> With Clang, UBSAN also reports the invalid alignment assumption at boot.
>> 
>> Changing ZERO_SIZE_PTR to satisfy the annotation would make its value and
>> the range accepted by ZERO_OR_NULL_PTR() architecture-dependent. Avoid that
>> semantic change by dropping the annotation from the general kmalloc entry
>> points. Retain it on the cache helpers, which cannot return the sentinel.
>> Allocation behavior is unchanged.
>> 
>> Kernels before v7.2 do not have __kmalloc_flags_noprof().
>> Backports to those kernels only need the include/linux/slab.h change.
>> 
>> Fixes: 94a58c360a45 ("slab.h: sprinkle __assume_aligned attributes")
>> Fixes: f6d50ab29afd ("mm/slab: introduce kmalloc_flags()")
>> Cc: <stable@xxxxxxxxxxxxxxx> # needs adjustment before v7.2
>> Assisted-by: Claude:claude-opus-4-8
>> Signed-off-by: Karl Mehltretter <kmehltretter@xxxxxxxxx>
>> ---
>> 
>> Notes:
>>     This is RFC because the alignment contract can be corrected in three ways:
> 
> I think it's still worth having at least a weaker alignment guarantee
> and would like to vote for option 3 (or option 2, if 3 turns out to be
> infeasible) unless there's something unexpected that prevents us from
> doing that.

Agreed.

>>       1. Remove the annotation from entry points that can return ZERO_SIZE_PTR
>>          (this patch). On armv5 this increases .text by 1120 bytes (0.02%);
>>          no measurable change was seen on arm64.
>>     
>>       2. Cap the assumed alignment at 16. This preserves some alignment
>>          information but couples the annotation to the current sentinel value.
>>
>>       3. Change ZERO_SIZE_PTR to satisfy ARCH_KMALLOC_MINALIGN. This makes the
>>          long-standing sentinel architecture-dependent and either expands the
>>          range accepted by ZERO_OR_NULL_PTR() or requires changing that macro.
> 
> Some code (e.g., kmem_dump_obj()) performs (addr < PAGE_SIZE) check
> either NULL or ZERO_SIZE_PTR) to see if the address is valid.
> 
> Bumping ZERO_SIZE_PTR to something smaller than PAGE_SIZE should still
> work: (addr < PAGE_SIZE) check still works, and accessing it still
> causes a fault. No arch should have ARCH_KMALLOC_MINALIGN >= PAGE_SIZE?

Yeah, I don't see why anything should mind if the value changes to a larger
one, if it's still well below PAGE_SIZE.