Re: tasks-trace RCU: question about grace period forward progress

[Mathieu Desnoyers]

----- On Feb 25, 2021, at 1:33 PM, paulmck paulmck@xxxxxxxxxx wrote:
[...]
> commit 581f79546b6be406a9c7280b2d3511b60821efe0
> Author: Paul E. McKenney <paulmck@xxxxxxxxxx>
> Date:   Thu Feb 25 10:26:00 2021 -0800
> 
>    rcu-tasks: Add block comment laying out RCU Tasks Trace design
>    
>    This commit adds a block comment that gives a high-level overview of
>    how RCU tasks trace grace periods progress.  It also adds a note about
>    how exiting tasks are handles, plus it gives an overview of the memory

handles -> handled

>    ordering.
>    
>    Reported-by: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
>    Reported-by: Mathieu Desnoyers <mathieu.desnoyers@xxxxxxxxxxxx>
>    Signed-off-by: Paul E. McKenney <paulmck@xxxxxxxxxx>
> 
> diff --git a/kernel/rcu/tasks.h b/kernel/rcu/tasks.h
> index 17c8ebe..f818357 100644
> --- a/kernel/rcu/tasks.h
> +++ b/kernel/rcu/tasks.h
> @@ -726,6 +726,42 @@ EXPORT_SYMBOL_GPL(show_rcu_tasks_rude_gp_kthread);
> // flavors, rcu_preempt and rcu_sched.  The fact that RCU Tasks Trace
> // readers can operate from idle, offline, and exception entry/exit in no
> // way allows rcu_preempt and rcu_sched readers to also do so.
> +//
> +// The implementation uses rcu_tasks_wait_gp(), which relies on function
> +// pointers in the rcu_tasks structure.  The rcu_spawn_tasks_trace_kthread()
> +// function sets these function pointers up so that rcu_tasks_wait_gp()
> +// invokes these functions in this order:
> +//
> +// rcu_tasks_trace_pregp_step():
> +//	Initialize the count of readers and block CPU-hotplug operations.
> +// rcu_tasks_trace_pertask(), invoked on every non-idle task:
> +//	Initialize per-task state and attempt to identify an immediate
> +//	quiescent state for that task, or, failing that, attempt to set
> +//	that task's .need_qs flag so that that task's next outermost
> +//	rcu_read_unlock_trace() will report the quiescent state (in which
> +//	case the count of readers is incremented).  If both attempts fail,
> +//	the task is added to a "holdout" list.
> +// rcu_tasks_trace_postscan():
> +//	Initialize state and attempt to identify an immediate quiescent
> +//	state as above (but only for idle tasks), unblock CPU-hotplug
> +//	operations, and wait for an RCU grace period to avoid races with
> +//	tasks that are in the process of exiting.
> +// check_all_holdout_tasks_trace(), repeatedly until holdout list is empty:
> +//	Scans the holdout list, attempting to identify a quiescent state
> +//	for each task on the list.  If there is a quiescent state, the
> +//	corresponding task is removed from the holdout list.
> +// rcu_tasks_trace_postgp():
> +//	Wait for the count of readers do drop to zero, reporting any stalls.
> +//	Also execute full memory barriers to maintain ordering with code
> +//	executing after the grace period.
> +//
> +// The exit_tasks_rcu_finish_trace() synchronizes with exiting tasks.
> +//
> +// Pre-grace-period update-side code is ordered before the grace
> +// period via the ->cbs_lock and barriers in rcu_tasks_kthread().
> +// Pre-grace-period read-side code is ordered before the grace period by
> +// atomic_dec_and_test() of the count of readers (for IPIed readers) and by
> +// scheduler context-switch ordering (for locked-down non-running readers).

The rest looks good, thanks!

Mathieu

> 
> // The lockdep state must be outside of #ifdef to be useful.
>  #ifdef CONFIG_DEBUG_LOCK_ALLOC

-- 
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com