[GIT RFC PULL rcu/urgent] Revert to fix RCU-related deadlock/softlockup

[GIT RFC PULL rcu/urgent] Revert to fix RCU-related deadlock/softlockup

[Paul E. McKenney]

Hello, Ingo,

This series has a single revert from the ill-starred attempt to inline
__rcu_read_lock() for preemptible RCU.  Without this revert, on mainline
kernels using CONFIG_RCU_BOOST there is a low-probability deadlock on the
runqueue locks, but one that actually appeared in Sasha Levin's testing.
With the revert, and with an diagnostic patch that increased probability
of the deadlock to a MTBF of roughly 10 seconds, Sasha's tests ran for
two days with no failure.

The sequence of events leading to the deadlock is as follows:

1.	A task enters an RCU read-side critical section, and is both
	preempted and subjected to RCU priority boosting.
2.	The task starts to exit its RCU read-side critical section,
	but is preempted in __rcu_read_unlock() just after the assignment
	setting t->rcu_read_lock_nesting to INT_MIN.  (The diagnostic
	patch mentioned above expands this window by ten microseconds,
	and is available in -rcu as a debug option queued for 3.7.)
3.	The task enters the scheduler, where it acquires the corresponding
	runqueue lock, then invokes rcu_switch_from() which in turn
	invokes rcu_preempt_note_context_switch(), which in turn invokes
	rcu_read_unlock_special(), which attempts to deboost the task.
4.	The attempt to deboost the task recursively enters the scheduler
	with a runqueue lock held, which can result in deadlock.

The revert moves the point at which rcu_preempt_note_context_switch() is
called to a point in the scheduler code before the runqueue lock is
acquired, avoiding the deadlock.

This pull is marked "RFC" because CONFIG_RCU_BOOST=y is not used much
outside of the real-time community.  I will be sending another pull
request later today (Pacific Time) for 3.6 RCU commits, which will
include this commit as well.  Your choice.  ;-)

This change is available in the git repository at:
  git://git.kernel.org/pub/scm/linux/kernel/git/paulmck/linux-rcu.git rcu/urgent

							Thanx, Paul

------------------>
Paul E. McKenney (1):
      Revert "rcu: Move PREEMPT_RCU preemption to switch_to() invocation"

 arch/um/drivers/mconsole_kern.c |    1 -
 include/linux/rcupdate.h        |    1 -
 include/linux/rcutiny.h         |    6 ++++++
 include/linux/sched.h           |   10 ----------
 kernel/rcutree.c                |    1 +
 kernel/rcutree.h                |    1 +
 kernel/rcutree_plugin.h         |   14 +++++++++++---
 kernel/sched/core.c             |    1 -
 8 files changed, 19 insertions(+), 16 deletions(-)

Re: [GIT RFC PULL rcu/urgent] Revert to fix RCU-related deadlock/softlockup

[Ingo Molnar]

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

> Hello, Ingo,
> 
> This series has a single revert from the ill-starred attempt to inline
> __rcu_read_lock() for preemptible RCU.  Without this revert, on mainline
> kernels using CONFIG_RCU_BOOST there is a low-probability deadlock on the
> runqueue locks, but one that actually appeared in Sasha Levin's testing.
> With the revert, and with an diagnostic patch that increased probability
> of the deadlock to a MTBF of roughly 10 seconds, Sasha's tests ran for
> two days with no failure.
> 
> The sequence of events leading to the deadlock is as follows:
> 
> 1.	A task enters an RCU read-side critical section, and is both
> 	preempted and subjected to RCU priority boosting.
> 2.	The task starts to exit its RCU read-side critical section,
> 	but is preempted in __rcu_read_unlock() just after the assignment
> 	setting t->rcu_read_lock_nesting to INT_MIN.  (The diagnostic
> 	patch mentioned above expands this window by ten microseconds,
> 	and is available in -rcu as a debug option queued for 3.7.)
> 3.	The task enters the scheduler, where it acquires the corresponding
> 	runqueue lock, then invokes rcu_switch_from() which in turn
> 	invokes rcu_preempt_note_context_switch(), which in turn invokes
> 	rcu_read_unlock_special(), which attempts to deboost the task.
> 4.	The attempt to deboost the task recursively enters the scheduler
> 	with a runqueue lock held, which can result in deadlock.
> 
> The revert moves the point at which rcu_preempt_note_context_switch() is
> called to a point in the scheduler code before the runqueue lock is
> acquired, avoiding the deadlock.
> 
> This pull is marked "RFC" because CONFIG_RCU_BOOST=y is not used much
> outside of the real-time community.  I will be sending another pull
> request later today (Pacific Time) for 3.6 RCU commits, which will
> include this commit as well.  Your choice.  ;-)

It got introduced in this cycle so I agree with you that the fix 
for the regression is justified.

> This change is available in the git repository at:
>   git://git.kernel.org/pub/scm/linux/kernel/git/paulmck/linux-rcu.git rcu/urgent
> 
> 							Thanx, Paul
> 
> ------------------>
> Paul E. McKenney (1):
>       Revert "rcu: Move PREEMPT_RCU preemption to switch_to() invocation"
> 
>  arch/um/drivers/mconsole_kern.c |    1 -
>  include/linux/rcupdate.h        |    1 -
>  include/linux/rcutiny.h         |    6 ++++++
>  include/linux/sched.h           |   10 ----------
>  kernel/rcutree.c                |    1 +
>  kernel/rcutree.h                |    1 +
>  kernel/rcutree_plugin.h         |   14 +++++++++++---
>  kernel/sched/core.c             |    1 -
>  8 files changed, 19 insertions(+), 16 deletions(-)

Pulled, thanks Paul!

	Ingo