Date: Mon, 25 Mar 2024 13:26:19 +0800
From: Z qiang <qiang.zhang1211@...il.com>
To: paulmck@...nel.org
Cc: frederic@...nel.org, neeraj.upadhyay@...nel.org, joel@...lfernandes.org, 
	rcu@...r.kernel.org, linux-kernel@...r.kernel.org
Subject: Re: [PATCH] rcutorture: Make rcutorture support srcu double call test

>
> On Sun, Mar 24, 2024 at 08:42:24PM +0800, Zqiang wrote:
> > This commit also allows rcutorture to support srcu double call test
> > with CONFIG_DEBUG_OBJECTS_RCU_HEAD option enabled. since the spinlock
>
>                                                    ^ Comma ","?
>
> > will be called in call_srcu(), in RT-kernel, the spinlock is sleepable,
>
> You lost me on "the spinlock will be called in call_srcu()".

Hi, Paul

I mean that
call_srcu()
->srcu_gp_start_if_needed
    ->spin_lock_irqsave_sdp_contention
         -> spin_trylock_irqsave_rcu_node     (may be return false)
          ->spin_lock_irqsave_rcu_node(ssp->srcu_sup, *flags);   <---spinlock

>
> > therefore remove disable-irq and disable-preempt protection.
> >
> > Signed-off-by: Zqiang <qiang.zhang1211@...il.com>
>
> Nice!  A question below.
>
> > ---
> >  kernel/rcu/rcutorture.c | 36 +++++++++++++++++++++---------------
> >  1 file changed, 21 insertions(+), 15 deletions(-)
> >
> > diff --git a/kernel/rcu/rcutorture.c b/kernel/rcu/rcutorture.c
> > index 3f9c3766f52b..6571a69142f8 100644
> > --- a/kernel/rcu/rcutorture.c
> > +++ b/kernel/rcu/rcutorture.c
> > @@ -388,6 +388,7 @@ struct rcu_torture_ops {
> >       int extendables;
> >       int slow_gps;
> >       int no_pi_lock;
> > +     int debug_objects;
> >       const char *name;
> >  };
> >
> > @@ -573,6 +574,7 @@ static struct rcu_torture_ops rcu_ops = {
> >       .irq_capable            = 1,
> >       .can_boost              = IS_ENABLED(CONFIG_RCU_BOOST),
> >       .extendables            = RCUTORTURE_MAX_EXTEND,
> > +     .debug_objects          = 1,
> >       .name                   = "rcu"
> >  };
> >
> > @@ -743,6 +745,7 @@ static struct rcu_torture_ops srcu_ops = {
> >       .cbflood_max    = 50000,
> >       .irq_capable    = 1,
> >       .no_pi_lock     = IS_ENABLED(CONFIG_TINY_SRCU),
> > +     .debug_objects  = 1,
> >       .name           = "srcu"
> >  };
> >
> > @@ -782,6 +785,7 @@ static struct rcu_torture_ops srcud_ops = {
> >       .cbflood_max    = 50000,
> >       .irq_capable    = 1,
> >       .no_pi_lock     = IS_ENABLED(CONFIG_TINY_SRCU),
> > +     .debug_objects  = 1,
> >       .name           = "srcud"
> >  };
> >
> > @@ -3481,35 +3485,37 @@ static void rcu_test_debug_objects(void)
> >  #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD
> >       struct rcu_head rh1;
> >       struct rcu_head rh2;
> > +     int idx;
> > +
> > +     if (!cur_ops->debug_objects || !cur_ops->call ||
> > +                     !cur_ops->cb_barrier)
>
> If this is built-in, could we please WARN if there is a conflict?

WARN_ON_ONCE(!cur_ops->debug_objects) ?

> Otherwise, it looks like the test succeeded.
>
> > +             return;
> > +
> >       struct rcu_head *rhp = kmalloc(sizeof(*rhp), GFP_KERNEL);
> >
> >       init_rcu_head_on_stack(&rh1);
> >       init_rcu_head_on_stack(&rh2);
> > -     pr_alert("%s: WARN: Duplicate call_rcu() test starting.\n", KBUILD_MODNAME);
> > +     pr_alert("%s: WARN: Duplicate call_%s() test starting.\n", KBUILD_MODNAME, cur_ops->name);
> >
> >       /* Try to queue the rh2 pair of callbacks for the same grace period. */
> > -     preempt_disable(); /* Prevent preemption from interrupting test. */
>
> What makes us not need this preempt_disable() in the RCU case?

the cur_ops->readlock/unlock() can guarantee that the callback will
not be called
when in the readlock/unlock() critical section.
Besides, for srcu, if invoke preempt_disable(), and the call_srcu()
internally calls
spinlock, which will trigger a lockdep warning in RT-kernels.


>
> > -     rcu_read_lock(); /* Make it impossible to finish a grace period. */
> > -     call_rcu_hurry(&rh1, rcu_torture_leak_cb); /* Start grace period. */
> > -     local_irq_disable(); /* Make it harder to start a new grace period. */
>
> Same question for the local_irq_disable()?
>
> > -     call_rcu_hurry(&rh2, rcu_torture_leak_cb);
> > -     call_rcu_hurry(&rh2, rcu_torture_err_cb); /* Duplicate callback. */
> > +     idx = cur_ops->readlock(); /* Make it impossible to finish a grace period. */
> > +     cur_ops->call(&rh1, rcu_torture_leak_cb); /* Start grace period. */
> > +     cur_ops->call(&rh2, rcu_torture_leak_cb);
> > +     cur_ops->call(&rh2, rcu_torture_err_cb); /* Duplicate callback. */
> >       if (rhp) {
> > -             call_rcu_hurry(rhp, rcu_torture_leak_cb);
> > -             call_rcu_hurry(rhp, rcu_torture_err_cb); /* Another duplicate callback. */
> > +             cur_ops->call(rhp, rcu_torture_leak_cb);
> > +             cur_ops->call(rhp, rcu_torture_err_cb); /* Another duplicate callback. */
> >       }
> > -     local_irq_enable();
> > -     rcu_read_unlock();
> > -     preempt_enable();
> > +     cur_ops->readunlock(idx);
> >
> >       /* Wait for them all to get done so we can safely return. */
> > -     rcu_barrier();
> > -     pr_alert("%s: WARN: Duplicate call_rcu() test complete.\n", KBUILD_MODNAME);
> > +     cur_ops->cb_barrier();
> > +     pr_alert("%s: WARN: Duplicate call_%s() test complete.\n", KBUILD_MODNAME, cur_ops->name);
> >       destroy_rcu_head_on_stack(&rh1);
> >       destroy_rcu_head_on_stack(&rh2);
> >       kfree(rhp);
> >  #else /* #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD */
> > -     pr_alert("%s: !CONFIG_DEBUG_OBJECTS_RCU_HEAD, not testing duplicate call_rcu()\n", KBUILD_MODNAME);
> > +     pr_alert("%s: !CONFIG_DEBUG_OBJECTS_RCU_HEAD, not testing duplicate call_%s()\n", KBUILD_MODNAME, cur_ops->name);
> >  #endif /* #else #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD */
>
> It might be possible to simplify the code by turning this #ifdef into
> IS_ENABLED().

mean that IS_ENABLED(CONFIG_DEBUG_OBJECTS_RCU_HEAD)?

Thanks
Zqiang

>
>                                                         Thanx, Paul
>
> >  }
> >
> > --
> > 2.17.1
> >

Powered by blists - more mailing lists