On 11/29/22 13:56, Marco Elver wrote:
> On Tue, 29 Nov 2022 at 13:53, Feng Tang <feng.tang@xxxxxxxxx> wrote:
>>
>> On Tue, Nov 29, 2022 at 08:02:51PM +0800, Vlastimil Babka wrote:
>> > On 11/29/22 12:48, Marco Elver wrote:
>> > > On Tue, 29 Nov 2022 at 12:01, Vlastimil Babka <vbabka@xxxxxxx> wrote:
>> > >>
>> > >> On 11/29/22 10:31, Marco Elver wrote:
>> > >> > On Tue, 29 Nov 2022 at 07:37, Feng Tang <feng.tang@xxxxxxxxx> wrote:
>> > >
>> > >> For SLAB_SKIP_KFENCE, we could also add the flag after creation to avoid
>> > >> this trouble? After all there is a sysfs file to control it at runtime
>> > >> anyway (via skip_kfence_store()).
>> > >> In that case patch 1 would have to wrap kmem_cache_create() and the flag
>> > >> addition with a new function to avoid repeating. That function could also be
>> > >> adding SLAB_NO_USER_FLAGS to kmem_cache_create(), instead of the #define
>> > >> DEFAULT_FLAGS.
>> > >
>> > > I wouldn't overcomplicate it, all we need is a way to say "this flag
>> > > should not be used directly" - and only have it available via an
>> > > indirect step. Availability via sysfs is one such step.
>> > >
>> > > And for tests, there are 2 options:
>> > >
>> > > 1. we could provide a function "kmem_cache_set_test_flags(cache,
>> > > gfp_flags)" and define SLAB_TEST_FLAGS (which would include
>> > > SLAB_SKIP_KFENCE). This still allows to set it generally, but should
>> > > make abuse less likely due to the "test" in the name of that function.
>> > >
>> > > 2. just set it directly, s->flags |= SLAB_SKIP_KFENCE.
>> > >
>> > > If you're fine with #2, that seems simplest and would be my preference.
>> >
>> > Yeah, that's what I meant. But slub_kunit.c could still have own internal
>> > cache creation function so the "|SLAB_NO_USER_FLAGS" and "s->flags |=
>> > SLAB_SKIP_KFENCE" is not repeated X times.
>>
>> I just quickly tried adding a new wrapper, like
>>
>> struct kmem_cache *debug_kmem_cache_create(const char *name, unsigned int size,
>> unsigned int align, slab_flags_t flags,
>> void (*ctor)(void *), slab_flags_t debug_flags);
>>
>> and found that, IIUC, both SLAB_KMALLOC and SLAB_NO_USER are creation
>> time flag, while SLAB_SKIP_KFENCE is an allocation runtime flag which
>> could be set after creation.
>>
>> So how about use the initial suggestion from Vlastimil to set the
>> SKIP_KFENCE flag through an internal wrapper in slub_kunit.c?
>>
>> /* Only for debug and test use, to skip kfence allocation */
>> static inline void kmem_cache_skip_kfence(struct kmem_cache *s)
>> {
>> s->flags |= SLAB_SKIP_KFENCE;
>> }
>
> Yes, that's fine - as long as it's local to slub_kunit.c, this seems
> very reasonable.
Wrapping just |= SLAB_SKIP_KFENCE won't help that much as you'd need to add
a call to kmem_cache_skip_kfence() after each kmem_cache_create() in
slub_kunit.c. That's why I propose a wrapper, *also internally defined in
slub_kunit.c*, that calls kmem_cache_create() with flags
|SLAB_NO_USER_FLAGS, then does s->flags |= SLAB_SKIP_KFENCE; then returns s.
At this point said wrapper wouldn't even need align and ctor parameters and
could pass 0 and NULL to kmem_cache_create() by itself, as no test uses
different values.
> Thanks,
> -- Marco
