Re: [PATCH RFC -tip 0/4] v2 RCU cleanups and simplified preemptable RCU

[Ingo Molnar <mingo@elte.hu>]

* Paul E. McKenney <paulmck@linux.vnet.ibm.com> wrote:

> [Not yet for inclusion, but making good progress.]
> 
> This patchset contains some RCU cleanups as follows:
> 
> o	Move private definitions from include/linux/rcutree.h to
> 	kernel/rcutree.h
> 
> o	Rename variables and functions so that RCU-sched is an
> 	underlying definition, along with RCU-bh and (when so
> 	configured) RCU-preempt.  RCU then maps to either RCU-sched
> 	or RCU-preempt, depending on configuration.
> 
> o	Consolidate sparse and lockdep into include/linux/rcupdate.h
> 	so that all RCU implementations are set up properly.
> 
> With these in place, we add configurable preemptable-RCU functionality
> to kernel/rcutree.c.  This new implementation has much faster and
> simpler read-side primitives (roughly that of Classic RCU built with
> CONFIG_PREEMPT), and has update-side performance equal to Classic RCU (at
> least in absence of blocking/preemption of read-side critical sections).
> This new implementation should eventually replace the old preemptable RCU,
> which would remove 2099 lines of code from the kernel, for a net removal
> of more than 1000 lines of code.
> 
> This patchset is undoubtably buggy, and does not have RCU priority
> boosting, though it does have the necessary tracking of tasks blocked
> in RCU read-side critical sections.
> 
> Changes from v1 (http://lkml.org/lkml/2009/7/23/294):
> 
> o	Fixes some locking problems detected by lockdep.
> 
> o	Disable irqs in quiescent-state-detection code, and fix handling
> 	of scheduling-clock interrupt always occurring in RCU read-side
> 	critical section.
> 
> o	Fix sparse annotations.
> 
> o	Apply feedback from Mathieu Desnoyers.
> 
> o	Fix x86 kernel-build errors.
> 
> o	Now passes moderate (multi-hour) rcutorture tests.
> 
> Shortcomings:
> 
> o	Only moderately tested, probably still quite buggy.  CPU hotplug
> 	not yet tested heavily, for example.
> 
> o	Probably does not even compile for all of the relevant
> 	combinations of kernel configuration variables.
> 
> o	Lacks RCU priority boosting.
> 
>  b/Documentation/RCU/trace.txt |    7 
>  b/include/linux/init_task.h   |   15 +
>  b/include/linux/rcupdate.h    |   21 +-
>  b/include/linux/rcupreempt.h  |    4 
>  b/include/linux/rcutree.h     |  211 --------------------
>  b/include/linux/sched.h       |   37 +++
>  b/init/Kconfig                |   22 +-
>  b/kernel/Makefile             |    1 
>  b/kernel/exit.c               |    1 
>  b/kernel/fork.c               |    5 
>  b/kernel/rcupreempt.c         |    8 
>  b/kernel/rcutree.c            |    2 
>  b/kernel/rcutree.h            |  238 +++++++++++++++++++++++
>  b/kernel/rcutree_plugin.h     |  428 ++++++++++++++++++++++++++++++++++++++++++
>  b/kernel/rcutree_trace.c      |    2 
>  b/kernel/sched.c              |    2 
>  b/kernel/softirq.c            |    5 
>  include/linux/rcupdate.h      |   48 ++++
>  include/linux/rcupreempt.h    |    8 
>  include/linux/rcutree.h       |   43 ++--
>  kernel/rcutree.c              |  207 +++++++++++++-------
>  kernel/rcutree.h              |   13 +
>  kernel/rcutree_trace.c        |   40 ++-
>  23 files changed, 1018 insertions(+), 350 deletions(-)

The structure looks really nice. If feasible i'd suggest to remove 
preemptible-rcu in this same patch-set as well, to simplify the 
testing matrix and to get the net code removal effect as well.

An added benefit is that that way the -rt folks will test out the 
new preemptible-rcu code with enthusiasm as well ;-)

	Ingo