On Tue, Feb 22, 2022 at 04:21:26PM +0100, Nicolas Saenz Julienne wrote:
> On Tue, 2022-02-22 at 11:47 -0300, Marcelo Tosatti wrote:
> > For the per-CPU LRU page vectors, augment the local lock protected
> > code sections with rcu_read_lock.
> >
> > This makes it possible to replace the queueing of work items on all
> > CPUs by synchronize_rcu (which is necessary to run FIFO:1 applications
> > uninterrupted on isolated CPUs).
>
> I don't think this is needed. In RT local_locks use a spinlock. See
> kernel/locking/spinlock_rt.c:
>
> "The RT [spinlock] substitutions explicitly disable migration and take
> rcu_read_lock() across the lock held section."
Nice! Then the migrate_disable from __local_lock and friends seems unnecessary as
well
#define __local_lock(__lock) \
do { \
migrate_disable(); \
spin_lock(this_cpu_ptr((__lock))); \
} while (0)
Since:
static __always_inline void __rt_spin_lock(spinlock_t *lock)
{
rtlock_might_resched();
rtlock_lock(&lock->lock);
rcu_read_lock();
migrate_disable();
}
Will resend -v2.
On Tue, 2022-02-22 at 12:51 -0300, Marcelo Tosatti wrote:
> On Tue, Feb 22, 2022 at 04:21:26PM +0100, Nicolas Saenz Julienne wrote:
> > On Tue, 2022-02-22 at 11:47 -0300, Marcelo Tosatti wrote:
> > > For the per-CPU LRU page vectors, augment the local lock protected
> > > code sections with rcu_read_lock.
> > >
> > > This makes it possible to replace the queueing of work items on all
> > > CPUs by synchronize_rcu (which is necessary to run FIFO:1 applications
> > > uninterrupted on isolated CPUs).
> >
> > I don't think this is needed. In RT local_locks use a spinlock. See
> > kernel/locking/spinlock_rt.c:
> >
> > "The RT [spinlock] substitutions explicitly disable migration and take
> > rcu_read_lock() across the lock held section."
>
> Nice! Then the migrate_disable from __local_lock and friends seems unnecessary as
> well
>
> #define __local_lock(__lock) \
> do { \
> migrate_disable(); \
> spin_lock(this_cpu_ptr((__lock))); \
> } while (0)
>
It's needed as you might migrate between:
cpu1_lock = this_cpu_ptr(__lock);
// migrate here to cpu2
spin_lock(cpu1_lock);
// unprotected write into cpu2 lists
Regards,
--
Nicolás Sáenz