Re: [RFC PATCH v3 1/5] tracing: Introduce faultable tracepoints (v3)

[Paul E. McKenney]

On Mon, Oct 02, 2023 at 09:19:36PM -0400, Steven Rostedt wrote:
> On Mon, 2 Oct 2023 17:14:39 -0700
> "Paul E. McKenney" <paulmck@xxxxxxxxxx> wrote:
> 
> > On Mon, Oct 02, 2023 at 07:10:23PM -0400, Steven Rostedt wrote:
> > > On Mon,  2 Oct 2023 16:25:27 -0400
> > > Mathieu Desnoyers <mathieu.desnoyers@xxxxxxxxxxxx> wrote:
> > >   
> > > > @@ -202,8 +198,12 @@ static inline struct tracepoint *tracepoint_ptr_deref(tracepoint_ptr_t *p)
> > > >  		if (WARN_ON_ONCE(RCUIDLE_COND(rcuidle)))		\
> > > >  			return;						\
> > > >  									\
> > > > -		/* keep srcu and sched-rcu usage consistent */		\
> > > > -		preempt_disable_notrace();				\
> > > > +		if (mayfault) {						\
> > > > +			rcu_read_lock_trace();				\  
> > > 
> > > I thought rcu_trace was for the case that a task can not voluntarily call
> > > schedule. If this tracepoint tries to read user space memory that isn't
> > > paged in, and faults, can't the faulting logic call schedule and break this
> > > requirement?  
> > 
> > Well, additional new uses of rcu_read_lock_trace() do bear close scrutiny,
> > but RCU Tasks Trace readers are permitted to block for page faults.
> > The BPF folks already use it for this purpose, so this should be OK.
> > (If for some unknown-to-me reason it isn't, I am sure that Alexei,
> > who is on CC, will not suffer in silence.)
> > 
> > One way of thinking of RCU Tasks Trace is as a form of SRCU with
> > lightweight readers.  Except that, unlike SRCU, there is only one global
> > RCU Tasks Trace.  This means that all RCU Tasks Trace users need to keep
> > each other informed, because one users' unruly readers will affect all
> > RCU Tasks Trace users.
> > 
> > But given that the BPF folks already have page faults in RCU Tasks Trace
> > readers, this one should be OK.
> 
> Then I think we should update the documentation.
> 
> From: Documentation/RCU/checklist.rst:
> 
>         If the updater uses call_rcu_tasks() or synchronize_rcu_tasks(),
>         then the readers must refrain from executing voluntary
>         context switches, that is, from blocking.  If the updater uses
>         call_rcu_tasks_trace() or synchronize_rcu_tasks_trace(), then
>         the corresponding readers must use rcu_read_lock_trace() and
>         rcu_read_unlock_trace().  If an updater uses call_rcu_tasks_rude()
>         or synchronize_rcu_tasks_rude(), then the corresponding readers
>         must use anything that disables preemption, for example,
>         preempt_disable() and preempt_enable().
> 
> Because it's all one paragraph it's a bit confusing to know what uses what.
> Perhaps it should be broken up a bit more?
> 
>         If the updater uses call_rcu_tasks() or synchronize_rcu_tasks(),
>         then the readers must refrain from executing voluntary
>         context switches, that is, from blocking.
> 
>         If the updater uses call_rcu_tasks_trace() or
>         synchronize_rcu_tasks_trace(), then the corresponding readers must
>         use rcu_read_lock_trace() and rcu_read_unlock_trace().
> 
>         If an updater uses call_rcu_tasks_rude() or synchronize_rcu_tasks_rude(),
>         then the corresponding readers must use anything that disables
>         preemption, for example, preempt_disable() and preempt_enable().
> 
> That way it is clear what uses what, as I read the original paragraph a
> couple of times and could have sworn that rcu_read_lock_trace() required
> tasks to not block.

That would work for me.  Would you like to send a patch, or would you
rather we made the adjustments?

							Thanx, Paul