Question about synchronize_sched_expedited()
Hello, Tejun,
I was taking another look at synchronize_sched_expedited(), and was concerned about the scenario listed out in the following commit. Is this scenario a real problem, or am I missing the synchronization that makes it safe?
(If my concerns are valid, I should also be able to change this to non-atomically increment synchronize_sched_expedited_count, but one step at a time...)
Thanx, Paul
------------------------------------------------------------------------
commit 1c2f788a742b87f8fae692b0b3014732124ee3c6
Author: Paul E. McKenney <[email protected]> Date: Mon Oct 25 07:39:22 2010 -0700
rcu: fix race condition in synchronize_sched_expedited()
The new (early 2010) implementation of synchronize_sched_expedited() uses try_stop_cpu() to force a context switch on every CPU. It also permits concurrent calls to synchronize_sched_expedited() to share a single call to try_stop_cpu() through use of an atomically incremented synchronize_sched_expedited_count variable. Unfortunately, this is subject to failure as follows:
o Task A invokes synchronize_sched_expedited(), try_stop_cpus()
succeeds, but Task A is preempted before getting to the atomic increment of synchronize_sched_expedited_count.
o Task B also invokes synchronize_sched_expedited(), with exactly
the same outcome as Task A.
o Task C also invokes synchronize_sched_expedited(), again with
exactly the same outcome as Tasks A and B.
o Task D also invokes synchronize_sched_expedited(), but only
gets as far as acquiring the mutex within try_stop_cpus() before being preempted, interrupted, or otherwise delayed.
o Task E also invokes synchronize_sched_expedited(), but only
gets to the snapshotting of synchronize_sched_expedited_count.
o Tasks A, B, and C all increment synchronize_sched_expedited_count.
o Task E fails to get the mutex, so checks the new value
of synchronize_sched_expedited_count. It finds that the value has increased, so (wrongly) assumes that its work has been done, returning despite there having been no expedited grace period since it began.
The solution is to have the lowest-numbered CPU atomically increment the synchronize_sched_expedited_count variable within the synchronize_sched_expedited_cpu_stop() function, which is under the protection of the mutex acquired by try_stop_cpus().
kernel/rcutree_plugin.h +2 -1
@@ -1041,6 +1041,8 @@ static int synchronize_sched_expedited_cpu_stop(void *data)* robustness against future implementation changes. */ smp_mb ();/* See above comment block. */ +if (cpumask_first (cpu_online_mask) ==smp_processor_id ()) +atomic_inc (&synchronize_sched_expedited_count);return 0 ; }
@@ -1077,7 +1079,6 @@ void synchronize_sched_expedited(void) }get_online_cpus (); } -atomic_inc (&synchronize_sched_expedited_count);smp_mb__after_atomic_inc ();/* ensure post-GP actions seen after GP. */ put_online_cpus (); }
Hello, Paul.
Quoted 10 lines · Paul E. McKenney
Hello, Tejun,
I was taking another look at
synchronize_sched_expedited(), and was concerned about the scenario listed out in the following commit. Is this scenario a real problem, or am I missing the synchronization that makes it safe?(If my concerns are valid, I should also be able to change this to non-atomically increment
synchronize_sched_expedited_count, but one step at a time...)
I think your concern is valid and this can happen w/o preemption given enough cpus and perfect timing. Was the original code free from this problem?
IMHO the counter based mechanism is a bit too difficult to ponder and verify. Can we do more conventional double queueing (ie. flipping pending and executing queues so that multiple sync calls can get coalesced while another one is in progress)? That's what the code is trying to achieve anyway, right?
Thanks.
Signature · 1 line
tejun
Quoted 10 lines · Tejun Heo
Hello, Paul.
Quoted 7 lines · Paul E. McKenney
Hello, Tejun,
I was taking another look at
synchronize_sched_expedited(), and was concerned about the scenario listed out in the following commit. Is this scenario a real problem, or am I missing the synchronization that makes it safe?
Quoted 7 lines
Quoted 3 lines
(If my concerns are valid, I should also be able to change this to non-atomically increment
synchronize_sched_expedited_count, but one step at a time...)I think your concern is valid and this can happen w/o preemption given enough cpus and perfect timing. Was the original code free from this problem?
I believe so -- there was a mutex guarding the whole operation, including the increment.
Quoted 5 lines
IMHO the counter based mechanism is a bit too difficult to ponder and verify. Can we do more conventional double queueing (ie. flipping pending and executing queues so that multiple sync calls can get coalesced while another one is in progress)? That's what the code is trying to achieve anyway, right?
Hmmm... But it would be necessary to flip the queues somewhere, and wouldn't determining where that somewhere was involve the same analysis and complexity as determining where to increment the counter?
Thanx, Paul
Hello, Paul.
Quoted 6 lines · Paul E. McKenney
Quoted 3 lines
I think your concern is valid and this can happen w/o preemption given enough cpus and perfect timing. Was the original code free from this problem?
I believe so -- there was a mutex guarding the whole operation, including the increment.
I see.
Quoted 9 lines
Quoted 5 lines
IMHO the counter based mechanism is a bit too difficult to ponder and verify. Can we do more conventional double queueing (ie. flipping pending and executing queues so that multiple sync calls can get coalesced while another one is in progress)? That's what the code is trying to achieve anyway, right?
Hmmm... But it would be necessary to flip the queues somewhere, and wouldn't determining where that somewhere was involve the same analysis and complexity as determining where to increment the counter?
I was thinking something like the following.
lock; if (list_empty(running)) add myself to running unlock; else remember list_empty(pending) append myself to pending queue; unlock and sleep; if (pending wasn't empty) return; do it; lock; wake up all on running and clear it; list_splice_init(pending, running); wake up the first of running; unlock;
Thanks.
Signature · 1 line
tejun