Re: [PATCH] rcu-tasks: Directly invoke rcuwait_wake_up() in call_rcu_tasks_generic()

[Paul E. McKenney]

On Thu, Feb 23, 2023 at 11:57:54AM -0500, Joel Fernandes wrote:
> 
> 
> > On Feb 23, 2023, at 11:10 AM, Joel Fernandes <joel@xxxxxxxxxxxxxxxxx> wrote:
> > 
> > On Thu, Feb 23, 2023 at 08:43:05AM +0000, Zhang, Qiang1 wrote:
> >>> From: Zqiang <qiang1.zhang@xxxxxxxxx>
> >>> Sent: Thursday, February 23, 2023 2:30 PM
> >>> To: paulmck@xxxxxxxxxx; frederic@xxxxxxxxxx; quic_neeraju@xxxxxxxxxxx;
> >>> joel@xxxxxxxxxxxxxxxxx
> >>> Cc: rcu@xxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx
> >>> Subject: [PATCH] rcu-tasks: Directly invoke rcuwait_wake_up() in
> >>> call_rcu_tasks_generic()
> >>> 
> >>> According to commit '3063b33a347c ("Avoid raw-spinlocked wakeups from
> >>> call_rcu_tasks_generic()")', the grace-period kthread is delayed to wakeup
> >>> using irq_work_queue() is because if the caller of
> >>> call_rcu_tasks_generic() holds a raw spinlock, when the kernel is built with
> >>> CONFIG_PROVE_RAW_LOCK_NESTING=y, due to a spinlock will be hold in
> >>> wake_up(), so the lockdep splats will happen. but now using
> >>> rcuwait_wake_up() to wakeup grace-period kthread instead of wake_up(), in
> >>> rcuwait_wake_up() no spinlock will be acquired, so this commit remove using
> >>> 
> >>> There are still spinlock-acquisition and spinlock-release invocations within the call path from rcuwait_wake_up().
> >>> 
> >>> rcuwait_wake_up() -> wake_up_process() -> try_to_wake_up(), then:
> >>> 
> >>>   raw_spin_lock_irqsave()
> >>>   ...
> >>>   raw_spin_unlock_irqrestore
> >> 
> >> Yes, but this is raw_spinlock acquisition and release(note: spinlock will convert to
> >> sleepable lock in Preempt-RT kernel, but raw spinlock is not change).
> >> 
> >> acquire raw_spinlock -> acquire spinlock  will trigger lockdep warning.
> > 
> > Is this really safe in the long run though? I seem to remember there are
> > weird locking dependencies if RCU is used from within the scheduler [1].
> > 
> > I prefer to keep it as irq_work_queue() unless you are seeing some benefit.
> > Generally, there has to be a 'win' or other justification for adding more
> > risk.
> 
> On second thought, you are deleting a decent number of lines.
> 
> What do others think?
> 
> I will take a closer look later, I am interested in researching the new lock dependency this adds.

One place to start is rcu_read_unlock_trace_special(), keeping firmly
in mind that rcu_read_unlock_trace() is intended to be invoked from a
great many places.

							Thanx, Paul

>  - Joel
> 
> > 
> > thanks,
> > 
> > - Joel
> > [1] http://www.joelfernandes.org/rcu/scheduler/locking/2019/09/02/rcu-schedlocks.html
> > 
> >>> irq_work_queue(), invoke rcuwait_wake_up() directly in
> >>> call_rcu_tasks_generic().
> >>> 
> >>> Signed-off-by: Zqiang <qiang1.zhang@xxxxxxxxx>
> >>> ---
> >>> kernel/rcu/tasks.h | 16 +---------------
> >>> 1 file changed, 1 insertion(+), 15 deletions(-)
> >>> 
> >>> diff --git a/kernel/rcu/tasks.h b/kernel/rcu/tasks.h index
> >>> baf7ec178155..757b8c6da1ad 100644
> >>> --- a/kernel/rcu/tasks.h
> >>> +++ b/kernel/rcu/tasks.h
> >>> @@ -39,7 +39,6 @@ struct rcu_tasks_percpu {
> >>>    unsigned long rtp_jiffies;
> >>>    unsigned long rtp_n_lock_retries;
> >>>    struct work_struct rtp_work;
> >>> -    struct irq_work rtp_irq_work;
> >>>    struct rcu_head barrier_q_head;
> >>>    struct list_head rtp_blkd_tasks;
> >>>    int cpu;
> >>> @@ -112,12 +111,9 @@ struct rcu_tasks {
> >>>    char *kname;
> >>> };
> >>> 
> >>> -static void call_rcu_tasks_iw_wakeup(struct irq_work *iwp);
> >>> -
> >>> #define DEFINE_RCU_TASKS(rt_name, gp, call, n)
> >>>            \
> >>> static DEFINE_PER_CPU(struct rcu_tasks_percpu, rt_name ## __percpu) = {
> >>>            \
> >>>    .lock = __RAW_SPIN_LOCK_UNLOCKED(rt_name ##
> >>> __percpu.cbs_pcpu_lock),        \
> >>> -    .rtp_irq_work = IRQ_WORK_INIT_HARD(call_rcu_tasks_iw_wakeup),
> >>>            \
> >>> };
> >>>        \
> >>> static struct rcu_tasks rt_name =
> >>>        \
> >>> {
> >>>        \
> >>> @@ -273,16 +269,6 @@ static void cblist_init_generic(struct rcu_tasks *rtp)
> >>>    pr_info("%s: Setting shift to %d and lim to %d.\n", __func__,
> >>> data_race(rtp->percpu_enqueue_shift), data_race(rtp-
> >>>> percpu_enqueue_lim));
> >>> }
> >>> 
> >>> -// IRQ-work handler that does deferred wakeup for call_rcu_tasks_generic().
> >>> -static void call_rcu_tasks_iw_wakeup(struct irq_work *iwp) -{
> >>> -    struct rcu_tasks *rtp;
> >>> -    struct rcu_tasks_percpu *rtpcp = container_of(iwp, struct
> >>> rcu_tasks_percpu, rtp_irq_work);
> >>> -
> >>> -    rtp = rtpcp->rtpp;
> >>> -    rcuwait_wake_up(&rtp->cbs_wait);
> >>> -}
> >>> -
> >>> // Enqueue a callback for the specified flavor of Tasks RCU.
> >>> static void call_rcu_tasks_generic(struct rcu_head *rhp, rcu_callback_t func,
> >>>                   struct rcu_tasks *rtp)
> >>> @@ -334,7 +320,7 @@ static void call_rcu_tasks_generic(struct rcu_head
> >>> *rhp, rcu_callback_t func,
> >>>    rcu_read_unlock();
> >>>    /* We can't create the thread unless interrupts are enabled. */
> >>>    if (needwake && READ_ONCE(rtp->kthread_ptr))
> >>> -        irq_work_queue(&rtpcp->rtp_irq_work);
> >>> +        rcuwait_wake_up(&rtp->cbs_wait);
> >>> }
> >>> 
> >>> // RCU callback function for rcu_barrier_tasks_generic().
> >>> --
> >>> 2.25.1
> >>