Hi,
There's an odd thing about the nr_active field in arrays of runqueue_t:
it is actually never initialized to 0!... This doesn't yet trigger a
bug probably because the way runqueues are allocated make it so that
it is already initialized to 0, but that's not a safe way. Here is a
patch:
Initialize to zero the nr_active field of arrays of runqueues.
(fixes future potential dynamic allocation of runqueues).
Signed-off-by: Samuel Thibault <[email protected]>
--- linux-2.6.17-orig/kernel/sched.c 2006-06-18 19:22:40.000000000 +0200
+++ linux/kernel/sched.c 2006-08-02 14:23:02.000000000 +0200
@@ -6132,6 +6132,7 @@
for (j = 0; j < 2; j++) {
array = rq->arrays + j;
+ array->nr_active = 0;
for (k = 0; k < MAX_PRIO; k++) {
INIT_LIST_HEAD(array->queue + k);
__clear_bit(k, array->bitmap);
* Samuel Thibault <[email protected]> wrote:
> Hi,
>
> There's an odd thing about the nr_active field in arrays of
> runqueue_t: it is actually never initialized to 0!... This doesn't
> yet trigger a bug probably because the way runqueues are allocated
> make it so that it is already initialized to 0, but that's not a safe
> way. Here is a patch:
we do rely on zero initialization of bss (and percpu) data in a number
of places.
Ingo
Ingo Molnar, le Wed 02 Aug 2006 17:24:19 +0200, a ?crit :
>
> * Samuel Thibault <[email protected]> wrote:
>
> > Hi,
> >
> > There's an odd thing about the nr_active field in arrays of
> > runqueue_t: it is actually never initialized to 0!... This doesn't
> > yet trigger a bug probably because the way runqueues are allocated
> > make it so that it is already initialized to 0, but that's not a safe
> > way. Here is a patch:
>
> we do rely on zero initialization of bss (and percpu) data in a number
> of places.
The rest of runqueue initialization doesn't rely on that, and as
a result people might think that it is safe to allocate runqueues
dynamically.
Samuel
On Wed, 2006-08-02 at 18:57 +0200, Samuel Thibault wrote:
> Ingo Molnar, le Wed 02 Aug 2006 17:24:19 +0200, a ?crit :
> >
> > * Samuel Thibault <[email protected]> wrote:
> >
> > > Hi,
> > >
> > > There's an odd thing about the nr_active field in arrays of
> > > runqueue_t: it is actually never initialized to 0!... This doesn't
> > > yet trigger a bug probably because the way runqueues are allocated
> > > make it so that it is already initialized to 0, but that's not a safe
> > > way. Here is a patch:
> >
> > we do rely on zero initialization of bss (and percpu) data in a number
> > of places.
>
> The rest of runqueue initialization doesn't rely on that, and as
> a result people might think that it is safe to allocate runqueues
> dynamically.
I don't buy the "safe to allocate runqueues dynamically" bit since they
are local to sched.c and if you do do that (I did for a customer once)
you better know what you're doing.
That said, ...
Hmm, Ingo I guess he's right on the first part:
<sched_init snipit>
rq->nr_running = 0;
[...]
#ifdef CONFIG_SMP
rq->sd = NULL;
for (j = 1; j < 3; j++)
rq->cpu_load[j] = 0;
rq->active_balance = 0;
rq->push_cpu = 0;
rq->migration_thread = NULL;
</sched_init snipit>
So I guess we should add his zero initializer, or we should remove all
the other zero initializers. Either way, we should be consistent.
-- Steve
Steven Rostedt, le Thu 03 Aug 2006 11:07:37 -0400, a ?crit :
> On Wed, 2006-08-02 at 18:57 +0200, Samuel Thibault wrote:
> > Ingo Molnar, le Wed 02 Aug 2006 17:24:19 +0200, a ?crit :
> > >
> > > * Samuel Thibault <[email protected]> wrote:
> > >
> > > > Hi,
> > > >
> > > > There's an odd thing about the nr_active field in arrays of
> > > > runqueue_t: it is actually never initialized to 0!... This doesn't
> > > > yet trigger a bug probably because the way runqueues are allocated
> > > > make it so that it is already initialized to 0, but that's not a safe
> > > > way. Here is a patch:
> > >
> > > we do rely on zero initialization of bss (and percpu) data in a number
> > > of places.
> >
> > The rest of runqueue initialization doesn't rely on that, and as
> > a result people might think that it is safe to allocate runqueues
> > dynamically.
>
> I don't buy the "safe to allocate runqueues dynamically" bit since they
> are local to sched.c and if you do do that (I did for a customer once)
> you better know what you're doing.
Yes, but as you agreed, initializing some members to 0 and not others
doesn't help to know what you're doing :)
Samuel