On Tue, Dec 18, 2018 at 9:55 PM Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> wrote: > > On Tue, 18 Dec 2018 14:30:33 +0100 Andrey Konovalov <andreyknvl@xxxxxxxxxx> wrote: > > > Instead of changing cache->align to be aligned to KASAN_SHADOW_SCALE_SIZE > > in kasan_cache_create() we can reuse the ARCH_SLAB_MINALIGN macro. > > > > ... > > > > --- a/arch/arm64/include/asm/kasan.h > > +++ b/arch/arm64/include/asm/kasan.h > > @@ -36,6 +36,10 @@ > > #define KASAN_SHADOW_OFFSET (KASAN_SHADOW_END - (1ULL << \ > > (64 - KASAN_SHADOW_SCALE_SHIFT))) > > > > +#ifdef CONFIG_KASAN_SW_TAGS > > +#define ARCH_SLAB_MINALIGN (1ULL << KASAN_SHADOW_SCALE_SHIFT) > > +#endif > > + > > void kasan_init(void); > > void kasan_copy_shadow(pgd_t *pgdir); > > asmlinkage void kasan_early_init(void); > > This looks unreliable. include/linux/slab.h has > > /* > * Setting ARCH_SLAB_MINALIGN in arch headers allows a different alignment. > * Intended for arches that get misalignment faults even for 64 bit integer > * aligned buffers. > */ > #ifndef ARCH_SLAB_MINALIGN > #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long) > #endif > > so if a .c file includes arch/arm64/include/asm/kasan.h after > include/linux/slab.h, it can get a macro-redefined warning. If the .c > file includes those headers in the other order, ARCH_SLAB_MINALIGN will > get a different value compared to other .c files. > > Or something like that. > > Different architectures define ARCH_SLAB_MINALIGN in different place: > > ./arch/microblaze/include/asm/page.h:#define ARCH_SLAB_MINALIGN L1_CACHE_BYTES > ./arch/arm/include/asm/cache.h:#define ARCH_SLAB_MINALIGN 8 > ./arch/sh/include/asm/page.h:#define ARCH_SLAB_MINALIGN 8 > ./arch/c6x/include/asm/cache.h:#define ARCH_SLAB_MINALIGN L1_CACHE_BYTES > ./arch/sparc/include/asm/cache.h:#define ARCH_SLAB_MINALIGN __alignof__(unsigned long long) > ./arch/xtensa/include/asm/processor.h:#define ARCH_SLAB_MINALIGN STACK_ALIGN > > which is rather bad of us. > > But still. I think your definition should occur in an arch header file > which is reliably included from slab.h. And kasan code should get its > definition of ARCH_SLAB_MINALIGN by including slab.h. > KASAN code doesn't use this macro directly, so I don't think it needs to get it's definition. What do you think about adding #include <linux/kasan.h> into linux/slab.h? Perhaps with a comment that this is needed to get definition of ARCH_SLAB_MINALIGN?
