On Wed, 2011-01-05 at 10:43 -0500, Jason Baron wrote:
> Hi,
>
> The first patch uses the storage space of the jump label key address
> as a pointer into the update table. In this way, we can find all
> the addresses that need to be updated without hashing.
>
> The second patch introduces:
>
> static __always_inline bool static_branch(struct jump_label_key *key);
>
> instead of the old JUMP_LABEL(key, label) macro.
>
> In this way, jump labels become really easy to use:
>
> Define:
>
> struct jump_label_key jump_key;
>
> Can be used as:
>
> if (static_branch(&jump_key))
> do unlikely code
>
> enable/disale via:
>
> jump_label_enable(&jump_key);
> jump_label_disable(&jump_key);
>
> that's it!
>
> For perf, which also uses jump labels, I've left the reference counting
> out of the jump label layer, thus removing the 'jump_label_inc()' and
> 'jump_label_dec()' interface. Hopefully, this is a more palatable solution.
Right, lets go with this. Maybe we'll manage to come up with something
saner than _else_atomic_read(), but for now its an improvement over what
we have.
* Peter Zijlstra ([email protected]) wrote:
> On Wed, 2011-01-05 at 10:43 -0500, Jason Baron wrote:
> > Hi,
> >
> > The first patch uses the storage space of the jump label key address
> > as a pointer into the update table. In this way, we can find all
> > the addresses that need to be updated without hashing.
> >
> > The second patch introduces:
> >
> > static __always_inline bool static_branch(struct jump_label_key *key);
> >
> > instead of the old JUMP_LABEL(key, label) macro.
> >
> > In this way, jump labels become really easy to use:
> >
> > Define:
> >
> > struct jump_label_key jump_key;
> >
> > Can be used as:
> >
> > if (static_branch(&jump_key))
> > do unlikely code
> >
> > enable/disale via:
> >
> > jump_label_enable(&jump_key);
> > jump_label_disable(&jump_key);
> >
> > that's it!
> >
> > For perf, which also uses jump labels, I've left the reference counting
> > out of the jump label layer, thus removing the 'jump_label_inc()' and
> > 'jump_label_dec()' interface. Hopefully, this is a more palatable solution.
>
> Right, lets go with this. Maybe we'll manage to come up with something
> saner than _else_atomic_read(), but for now its an improvement over what
> we have.
I agree that keeping jump_label.h with the minimal clean API is a good
goal, and this patchset is almost there (maybe except for the
_else_atomic_read() part).
Hrm, given that the atomic inc/dec return and test for 1/0 is moved into
the Perf code, I wonder if it would make sense to move the
"_else_atomic_read()" oddness into the perf code too ? Perf could
declare, in its own header, a wrapper over __static_branch, e.g. put in
perf_event.h:
#ifdef HAVE_JUMP_LABEL
static __always_inline
bool perf_sw_event_static_branch_refcount(struct jump_label_key *key,
atomic_t *ref)
{
return __static_branch(key);
}
#else
static __always_inline
bool perf_sw_event_static_branch_refcount(struct jump_label_key *key,
atomic_t *ref)
{
if (unlikely(atomic_read(ref)))
return true;
return false;
}
#endif
Otherwise, jump_label_ref.h looks like an half-baked interface that only
provides the "test" API but not the ref/unref. If we have only a single
user interested in refcounting, it might make more sense to put the code
in perf_event.h. If we have many users using an atomic refcount like
this, then we should extend jump_label_ref.h to also provide the
ref/unref API too. I don't care much about where it ends up, as long as
it's a consistent choice.
Thoughts ?
Mathieu
--
Mathieu Desnoyers
Operating System Efficiency R&D Consultant
EfficiOS Inc.
http://www.efficios.com