I'd mentioned this when rt29 was announced, but there was yet one
more piece of the puzzle I hadn't figured out until rt30 (#4 below).
The executive summary is that we inserted an invalid dependency, but
then via another incorrect patch, reverted that dependency, and hence
it is still possible to select RT_FULL without PREEMPT_RCU. We can
dump two patches from the queue by cleaning this up.
This commit message (from #2 below):
PREEMPT_RT relies on PREEMPT_RCU - only allow RCU to be configured
interactively in the !PREEMPT_RT case.
captures what we want(ed) to happen but wasn't what was actually
happening (note that PREEMPT_RT implicitly means RT_FULL here;
the log was written before RT_FULL existed).
A stacking effect of errors that is typical of large patch queues
that get carried forward over time took place. It goes back at
least as far as rt-3.2
Here are the gory details:
1) We have "sched-disable-rt-group-sched-on-rt.patch" which adds a
"depends !PREEMPT_RT_FULL" to RT_GROUP_SCHED.
2) The patch "rcu-force-preempt-rcu-for-rt.patch" got mangled during
a carry forward. It has the commit log quoted above, but instead
of doing that, it just reverts the sched patch above in #1.
3) The bug that #1 fixed got "rediscovered" and so
"rt-disable-rt-group-sched.patch" got added as an accidental
duplicate of the earlier sched patch in #1. (instead of fixing #2)
4) Fixing patch #2, so it adds a "depends !PREEMPT_RT_FULL" to the
proper block (i.e. PREEMPT_RCU instead of RT_GROUP_SCHED) doesn't
have the desired effect of making PREEMPT_RCU enabled and non
interactive, instead it simply makes it go away entirely when
RT_FULL is enabled (and then you get compile failures, etc.)
So, to fix all of that, we throw away the duplicate from #3,
delete #2 entirely, since it is an accidental revert and can't
be salvaged (as per #4). And finally, to achieve the goal of
having PREEMPT_RCU be always set non-interactively for RT_FULL=y
we simply add a select for it in the RT_FULL Kconfig control block.
That is added into the patch which introduces RT_FULL.
Overall, this is primarily a queue maintenance thing and not really
that interesting to the end users, since diffing the two trees after
running the updated queue results in only this difference:
-------------------------------------
$ git diff v3.6.11-rt30 v3.6.11-rt30-fixes
diff --git a/kernel/Kconfig.preempt b/kernel/Kconfig.preempt
index 3050029..38cecfe 100644
--- a/kernel/Kconfig.preempt
+++ b/kernel/Kconfig.preempt
@@ -77,6 +77,7 @@ config PREEMPT_RT_FULL
bool "Fully Preemptible Kernel (RT)"
depends on IRQ_FORCED_THREADING
select PREEMPT_RT_BASE
+ select PREEMPT_RCU
help
All and everything
$
-------------------------------------
I'm posting this since even if there isn't a v3.6.11-rt31, ideally we'd
like to see this not get carried forward to yet another release.
Thanks,
Paul.
---
The following changes since commit bb102e108f5d60068b3a5564f0f1f314d4c72332:
patches-3.6.11-rt30.tar.xz (2013-02-19 18:33:41 -0500)
are available in the git repository at:
git://git.kernel.org/pub/scm/linux/kernel/git/paulg/3.6-rt-patches.git v3.6.11-rt30-fixes
for you to fetch changes up to 55c97947531d88af6cf9f56ccccb41e26c7f8cd5:
preempt-rcu: untangle kconfig RT_FULL/PREEMPT_RCU relationship (2013-02-20 19:05:38 -0500)
----------------------------------------------------------------
Paul Gortmaker (3):
rt: drop duplicated rt-disable-rt-group-sched.patch
rt: fix mispatch of Kconfig by rcu-force-preempt-rcu-for-rt.patch
preempt-rcu: untangle kconfig RT_FULL/PREEMPT_RCU relationship
kconfig-preempt-rt-full.patch | 3 ++-
rcu-force-preempt-rcu-for-rt.patch | 28 ----------------------------
rt-disable-rt-group-sched.patch | 29 -----------------------------
series | 4 ++--
4 files changed, 4 insertions(+), 60 deletions(-)
delete mode 100644 rcu-force-preempt-rcu-for-rt.patch
delete mode 100644 rt-disable-rt-group-sched.patch
--
To unsubscribe from this list: send the line "unsubscribe linux-rt-users" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
