Andrew Donnellan <ajd@xxxxxxxxxxxxx> writes: > On Thu, 2023-01-26 at 17:31 +0000, David Laight wrote: >> Changing the size to kzalloc() doesn't help. >> The alignment depends on the allocator and is only required to have >> a relatively small alignment (ARCH_MINALIGN?) regardless of the size. >> >> IIRC one of the allocators adds a small header to every item. >> It won't return 16 byte aligned items at all. > > I'm relying on the behaviour described in Documentation/core- > api/memory-allocation.rst: > > The address of a chunk allocated with kmalloc is aligned to at > least ARCH_KMALLOC_MINALIGN bytes. For sizes which are a power of > two, the alignment is also guaranteed to be at least the respective > size. > > Is this wrong? I believe it's correct. For SLAB and SLUB it boils down to: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/mm/slab_common.c?commit=830b3c68c1fb1e9176028d02ef86f3cf76aa2476#n640 That's where the kmalloc slabs are created (see create_kmalloc_cache()) just below. If you create your own slab (with kmem_cache_create()) then the alignment is up to you, so that's why there's no power-of-2 logic in calculate_alignment(). And SLOB (which we don't use) does something similar: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/mm/slob.c?commit=830b3c68c1fb1e9176028d02ef86f3cf76aa2476#n493 cheers
