Messages in this thread | | | Date | Tue, 30 Jun 2015 15:25:30 -0700 | From | "Paul E. McKenney" <> | Subject | [PATCH RFC tip/core/rcu 0/14] Rework expedited grace periods |
| |
Hello!
This RFC series reworks the handling of expedited grace periods, eliminating the previous use of try_stop_cpus(), improving scalability, and adding an RCU CPU stall-warning capability. This series is as follows:
1. Replace try_stop_cpus() with stop_one_cpu(), courtesy of Peter Zijlstra.
2. Improve counter handling to more efficiently detect that some other task has already done the required work.
3. Replace synchronize_sched_expedited()'s polling loop with funnel locking.
4. Replace stop_one_cpu() with stop_one_cpu_nowait() to avoid the latency hit from #1 above, also courtesy of Peter Zijlstra.
5. Abstract the counter handling.
6. Apply the newly abstracted counter handling to synchronize_rcu_expedited().
7. Abstract the funnel locking.
8. Fix a stupid type error in synchronize_sched_expedited().
9. Replace synchronize_rcu_expedited()'s polling with the newly abstracted funnel locking.
10. Apply the abstracted counter handling to _rcu_barrier().
11. Consolidate the last expedited open-coded expedited memory barrier into the counter handling.
12. Extend the funnel locking to the rcu_data structure in order to further increase scalability.
13. Add stall warnings to synchronize_sched_expedited().
14. Document the newly added stall warnings.
Thanx, Paul
------------------------------------------------------------------------
b/Documentation/RCU/stallwarn.txt | 17 + b/include/trace/events/rcu.h | 1 b/kernel/rcu/tree.c | 581 +++++++++++++++++++------------------- b/kernel/rcu/tree.h | 31 +- b/kernel/rcu/tree_plugin.h | 53 +-- b/kernel/rcu/tree_trace.c | 25 - 6 files changed, 369 insertions(+), 339 deletions(-)
| |