Re: [PATCH v2 6/7] mm, slab: add static key for should_failslab()
On 6/25/24 7:12 PM, Alexei Starovoitov wrote: > On Tue, Jun 25, 2024 at 7:24 AM Vlastimil Babka wrote: >> >> On 6/20/24 12:49 AM, Vlastimil Babka wrote: >> > --- a/mm/slub.c >> > +++ b/mm/slub.c >> > @@ -3874,13 +3874,37 @@ static __always_inline void >> > maybe_wipe_obj_freeptr(struct kmem_cache *s, >> > 0, sizeof(void *)); >> > } >> > >> > -noinline int should_failslab(struct kmem_cache *s, gfp_t gfpflags) >> > +#if defined(CONFIG_FUNCTION_ERROR_INJECTION) || defined(CONFIG_FAILSLAB) >> > +DEFINE_STATIC_KEY_FALSE(should_failslab_active); >> > + >> > +#ifdef CONFIG_FUNCTION_ERROR_INJECTION >> > +noinline >> > +#else >> > +static inline >> > +#endif >> > +int should_failslab(struct kmem_cache *s, gfp_t gfpflags) >> >> Note that it has been found that (regardless of this series) gcc may clone >> this to a should_failslab.constprop.0 in case the function is empty because >> __should_failslab is compiled out (CONFIG_FAILSLAB=n). The "noinline" >> doesn't help - the original function stays but only the clone is actually >> being called, thus overriding the original function achieves nothing, see: >> https://github.com/bpftrace/bpftrace/issues/3258 >> >> So we could use __noclone to prevent that, and I was thinking by adding >> something this to error-injection.h: >> >> #ifdef CONFIG_FUNCTION_ERROR_INJECTION >> #define __error_injectable(alternative) noinline __noclone > > To prevent such compiler transformations we typically use > __used noinline > > We didn't have a need for __noclone yet. If __used is enough I'd stick to > that. __used made no difference here (gcc 13.3), __noclone did
Re: [PATCH v2 6/7] mm, slab: add static key for should_failslab()
On Tue, Jun 25, 2024 at 7:24 AM Vlastimil Babka wrote: > > On 6/20/24 12:49 AM, Vlastimil Babka wrote: > > --- a/mm/slub.c > > +++ b/mm/slub.c > > @@ -3874,13 +3874,37 @@ static __always_inline void > > maybe_wipe_obj_freeptr(struct kmem_cache *s, > > 0, sizeof(void *)); > > } > > > > -noinline int should_failslab(struct kmem_cache *s, gfp_t gfpflags) > > +#if defined(CONFIG_FUNCTION_ERROR_INJECTION) || defined(CONFIG_FAILSLAB) > > +DEFINE_STATIC_KEY_FALSE(should_failslab_active); > > + > > +#ifdef CONFIG_FUNCTION_ERROR_INJECTION > > +noinline > > +#else > > +static inline > > +#endif > > +int should_failslab(struct kmem_cache *s, gfp_t gfpflags) > > Note that it has been found that (regardless of this series) gcc may clone > this to a should_failslab.constprop.0 in case the function is empty because > __should_failslab is compiled out (CONFIG_FAILSLAB=n). The "noinline" > doesn't help - the original function stays but only the clone is actually > being called, thus overriding the original function achieves nothing, see: > https://github.com/bpftrace/bpftrace/issues/3258 > > So we could use __noclone to prevent that, and I was thinking by adding > something this to error-injection.h: > > #ifdef CONFIG_FUNCTION_ERROR_INJECTION > #define __error_injectable(alternative) noinline __noclone To prevent such compiler transformations we typically use __used noinline We didn't have a need for __noclone yet. If __used is enough I'd stick to that.
Re: [PATCH v2 6/7] mm, slab: add static key for should_failslab()
On 6/20/24 12:49 AM, Vlastimil Babka wrote:
> --- a/mm/slub.c
> +++ b/mm/slub.c
> @@ -3874,13 +3874,37 @@ static __always_inline void
> maybe_wipe_obj_freeptr(struct kmem_cache *s,
> 0, sizeof(void *));
> }
>
> -noinline int should_failslab(struct kmem_cache *s, gfp_t gfpflags)
> +#if defined(CONFIG_FUNCTION_ERROR_INJECTION) || defined(CONFIG_FAILSLAB)
> +DEFINE_STATIC_KEY_FALSE(should_failslab_active);
> +
> +#ifdef CONFIG_FUNCTION_ERROR_INJECTION
> +noinline
> +#else
> +static inline
> +#endif
> +int should_failslab(struct kmem_cache *s, gfp_t gfpflags)
Note that it has been found that (regardless of this series) gcc may clone
this to a should_failslab.constprop.0 in case the function is empty because
__should_failslab is compiled out (CONFIG_FAILSLAB=n). The "noinline"
doesn't help - the original function stays but only the clone is actually
being called, thus overriding the original function achieves nothing, see:
https://github.com/bpftrace/bpftrace/issues/3258
So we could use __noclone to prevent that, and I was thinking by adding
something this to error-injection.h:
#ifdef CONFIG_FUNCTION_ERROR_INJECTION
#define __error_injectable(alternative) noinline __noclone
#else
#define __error_injectable(alternative) alternative
#endif
and the usage here would be:
__error_injectable(static inline) int should_failslab(...)
Does that look acceptable, or is it too confusing that "static inline" is
specified there as the storage class to use when error injection is actually
disabled?
> {
> if (__should_failslab(s, gfpflags))
> return -ENOMEM;
> return 0;
> }
> -ALLOW_ERROR_INJECTION(should_failslab, ERRNO);
> +ALLOW_ERROR_INJECTION_KEY(should_failslab, ERRNO, &should_failslab_active);
> +
> +static __always_inline int should_failslab_wrapped(struct kmem_cache *s,
> +gfp_t gfp)
> +{
> + if (static_branch_unlikely(&should_failslab_active))
> + return should_failslab(s, gfp);
> + else
> + return 0;
> +}
> +#else
> +static __always_inline int should_failslab_wrapped(struct kmem_cache *s,
> +gfp_t gfp)
> +{
> + return false;
> +}
> +#endif
>
> static __fastpath_inline
> struct kmem_cache *slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
> @@ -3889,7 +3913,7 @@ struct kmem_cache *slab_pre_alloc_hook(struct
> kmem_cache *s, gfp_t flags)
>
> might_alloc(flags);
>
> - if (unlikely(should_failslab(s, flags)))
> + if (should_failslab_wrapped(s, flags))
> return NULL;
>
> return s;
>
