Skip to contentSkip to conversation

Markers : fix reentrancy

PATCH In review Mathieu Desnoyers 2 participants 3 messages Sep 30, 2008

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 Markers : fix reentrancy

Quoted 4 lines · Lai Jiangshan

marker_synchronize_unregister must be called _also_ between unregistration and destruction the data that unregistration-ed probes need to make sure there is no caller executing a probe when it's data is destroyed.

Ah, you're right. I looked again through your patch description and I think using a :

if (entry->rcu_pending)

                rcu_barrier_sched();

After each time the markers_mutex is taken would keep the fast batch registration/unregistration and fix the reentrancy problem. The following patch should address the problem.

Thanks,

Mathieu

Lai Jiangshan discovered a reentrancy issue with markers. This patch implements a version of the fix which won't slow down marker batch registration/unregistration.

CC Ingo Molnar <[email protected]>
CC Lai Jiangshan <[email protected]>
CC "Frank Ch. Eigler" <[email protected]>
---
 kernel/marker.c |    6 ++++++
 1 file changed, 6 insertions(+)
kernel/marker.c +6 -0
Index: linux-2.6-lttng/kernel/marker.c
===================================================================
@@ -674,6 +674,8 @@ int marker_probe_register(const char *na
 	mutex_lock(&markers_mutex);
 	entry = get_marker(name);
 	WARN_ON(!entry);
+	if (entry->rcu_pending)
+		rcu_barrier_sched();
 	entry->oldptr = old;
 	entry->rcu_pending = 1;
 	/* write rcu_pending before calling the RCU callback */
@@ -717,6 +719,8 @@ int marker_probe_unregister(const char *
 	entry = get_marker(name);
 	if (!entry)
 		goto end;
+	if (entry->rcu_pending)
+		rcu_barrier_sched();
 	entry->oldptr = old;
 	entry->rcu_pending = 1;
 	/* write rcu_pending before calling the RCU callback */
@@ -795,6 +799,8 @@ int marker_probe_unregister_private_data
 	mutex_lock(&markers_mutex);
 	entry = get_marker_from_private_data(probe, probe_private);
 	WARN_ON(!entry);
+	if (entry->rcu_pending)
+		rcu_barrier_sched();
 	entry->oldptr = old;
 	entry->rcu_pending = 1;
 	/* write rcu_pending before calling the RCU callback */
Signature · 2 lines
Mathieu Desnoyers
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68
#1 lore raw
Quoted 10 lines · Mathieu Desnoyers
Quoted 4 lines · Lai Jiangshan

marker_synchronize_unregister must be called _also_ between unregistration and destruction the data that unregistration-ed probes need to make sure there is no caller executing a probe when it's data is destroyed.

Ah, you're right. I looked again through your patch description and I think using a :

Quoted 6 lines

if (entry->rcu_pending)

                rcu_barrier_sched();

After each time the markers_mutex is taken would keep the fast batch registration/unregistration and fix the reentrancy problem. The following patch should address the problem.

could you please send a delta patch against tip/master? Lai's patch is already applied and is tested. Thanks,

Ingo

#2 lore raw
Quoted 16 lines · * Ingo Molnar ([email protected]) wrote:
Quoted 13 lines · Mathieu Desnoyers
Quoted 4 lines · Lai Jiangshan

marker_synchronize_unregister must be called _also_ between unregistration and destruction the data that unregistration-ed probes need to make sure there is no caller executing a probe when it's data is destroyed.

Ah, you're right. I looked again through your patch description and I think using a :

if (entry->rcu_pending)

                rcu_barrier_sched();
Quoted 8 lines
Quoted 3 lines

After each time the markers_mutex is taken would keep the fast batch registration/unregistration and fix the reentrancy problem. The following patch should address the problem.

could you please send a delta patch against tip/master? Lai's patch is already applied and is tested. Thanks,

Ingo

Hrm,

Most of Lai "simplifications" to use synchronize_sched() instead of call_rcu_sched() will make batch registration/unregistration much slower (very noticeable on ~50 markers on a loaded machine). It also contains two different fixes in one, one of which has been nacked.

The correct list of changes against tip now becomes :

Revert commit d587284c1d26b796bf52a6e3d3f783db4e462119 Apply patch "Markers : marker_synchronize_unregister()" Apply patch "Markers : probe example fix teardown" Apply patch "Markers : documentation fix teardown" Apply patch "sputrace : use marker_synchronize_unregister()" Apply patch "Markers : fix reentrancy"

Sorry, but rebasing those marker fixes will just make all those changes really messy. I think the cleanest way is to just revert. :-/

Mathieu

Signature · 2 lines
Mathieu Desnoyers
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68