Re: [PATCH tip/core/rcu 0/55] Preview of RCU changes for 3.2

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Sun, Oct 16, 2011 at 07:06:13PM -0700, Josh Triplett wrote:
> On Tue, Sep 06, 2011 at 11:00:15AM -0700, Paul E. McKenney wrote:
> > Hello!
> > 
> > This patchset adds RCU event tracing, improved diagnostics and
> > documentation, and fixes a number of bugs, including several from an
> > ongoing top-to-bottom inspection of RCU.  The patches are as follows:
> > 
> > 1.	Place per-CPU kthreads' stack and task struct on the corresponding
> > 	node on NUMA systems (courtesy of Eric Dumazet).
> > 2.	Avoid unnecessary self-wakeups for per-CPU kthreads
> > 	(courtesy of Shaohua Li).
> > 3,6,10,12,25,28,33.
> > 	Documentations updates (some courtesy Wanlong Gao).
> > 4.	Add replacement checks for blocking within an RCU read-side
> > 	critical section.
> > 5.	Header-file untangling part 1 of N: move rcu_head to types.h.
> > 7.	Fix mismatched variable declaration (courtesy of Andi Kleen).
> > 8.	Abstract out common grace-period-primitive code.
> > 9.	Update rcutorture to test newish RCU API members.
> > 11.	Drive RCU algorithm selection directly from SMP and PREEMPT.
> > 13.	Make rcu_torture_boost() wait for callbacks before telling
> > 	debug-objects that they are done.
> > 14-17,20,22.
> > 	Add event tracing for RCU.
> > 18.	Update comments to reflect kthreads being used only when
> > 	RCU priority boosting is enabled.
> > 19.	Move RCU_BOOSt data declarations to alow compiler to detect
> > 	mismatches.
> > 20.	Make TINY_RCU use softirqs for RCU_BOOST=n.
> > 23.	Simplify quiescent-state accounting.
> > 24.	Stop passing rcu_read_lock_held() to rcu_dereference_protected()
> > 	(courtesy of Michal Hocko).
> > 26.	Remove unused and redundant RCU API members.
> > 27.	Allow rcutorture's stat_interval parameter to be changed at runtime
> > 	to make it easier to test RCU in guest OSes.
> > 28.	Removed unused nohz_cpu_mask (courtesy of Alex Shi).
> > 30.	Eliminate in_irq() checks in rcu_enter_nohz().
> > 31.	Fix rcu_implicit_dynticks_qs() local-variable size mismatches.
> > 32.	Make rcu_assign_pointer() unconditionally emit memory barrier
> > 	to silence new gcc warnings (courtesy of Eric Dumazet).
> > 34.	Move __rcu_read_lock()'s barrier within if-statement.
> > 35.	Dump local stack for CPU stall warnings if cannot dump all stacks.
> > 36.	Prevent early-boot set_need_resched() from __rcu_pending().
> > 37.	Simplify unboosting checks.
> > 38.	Prohibit RCU grace periods during early boot.
> > 39.	Suppress NMI backtraces when CPU stall ends before dump.
> > 40.	Avoid just-online CPU needlessly rescheding itself.
> > 41.	Permit rt_mutex_unlock() with irqs disabled.
> > 42-43.	Prevent end-of-test rcutorture hangs.
> > 44.	Wire up RCU_BOOST_PRIO, use conventional kthread naming scheme
> > 	(courtesy of Mike Galbraith).
> > 45.	Check for entering dyntick-idle in RCU read-side critical section.
> > 46.	Adjust RCU_FAST_NO_HZ to avoid false quiescent states.
> > 47.	Avoid concurrent end of old GP with start of new GP.
> > 48.	Strengthen powerpc value-returning atomic memory ordering.
> > 49-51.	Detect illegal RCU use from dyntick-idle mode (courtesy of
> > 	Frederic Weisbecker).
> > 52.	Remove an unnecessary layer of abstraction from PROVE_RCU checking.
> > 53.	Detect illegal SRCU use from dyntick-idle mode.
> > 54.	Make SRCU use common lockdep-splat code.
> > 55.	Placeholder patch that disables illegal tracing from dyntick-idle
> > 	mode (illegal because tracing uses RCU).
> 
> I responded to a few of the patches with comments and potential issues.
> I also don't consider myself qualified to review patch 48/55, "powerpc:
> strengthen value-returning-atomics memory barriers".  For the rest:
> 
> Reviewed-by: Josh Triplett <josh@joshtriplett.org>

Thank you very much for your review and comments!!!

							Thanx, Paul