Matthew Wilcox <willy@infradead.org> 于2024年11月4日周一 21:27写道：

> On Mon, Nov 04, 2024 at 02:16:17PM +0800, Ke Sun wrote:
> > Use the KMEM_CACHE() macro instead of kmem_cache_create() to simplify
> > the creation of SLAB caches.
>
> Did you even test this?  It is REQUIRED that maple_node be aligned to
>

I have tests on an ARM64 QEMU VM and enabled some Maple Tree debugging
configs:
CONFIG_DEBUG_VM_MAPLE_TREE=y
CONFIG_DEBUG_MAPLE_TREE=y
CONFIG_TEST_MAPLE_TREE=y

The results show:
[    1.434017][    T1] TEST STARTING
[    1.434017][    T1]
[   78.257756][    T1] maple_tree: 172860522 of 172860522 tests passed
[   78.271770][    T1] atomic64_test: passed


> its size (eg 256 bytes) as the bottom bits of pointers to nodes are used
> for other purposes.  KMEM_CACHE() does not give us this guarantee.
>

The code comments of kmem_cache say: 'The alignment of the struct
determines object alignment.'
 #define KMEM_CACHE(__struct, __flags)                                   \
__kmem_cache_create_args(#__struct, sizeof(struct __struct),    \
&(struct kmem_cache_args) { \
.align = __alignof__(struct __struct), \
}, (__flags))


>
> NACK.
>
>