diff options
author | Sean Dague | 2008-11-10 12:42:22 +0000 |
---|---|---|
committer | Sean Dague | 2008-11-10 12:42:22 +0000 |
commit | bfc2d8c231234c35ae8134d885c26f0fe14935e2 (patch) | |
tree | 8673d0e1cb8947ee96bb2a7a92df7436e8dd3c49 /OpenSim/Region/Environment/Scenes | |
parent | The region proxy for the load balancer module works again. The incoming proxy... (diff) | |
download | opensim-SC_OLD-bfc2d8c231234c35ae8134d885c26f0fe14935e2.zip opensim-SC_OLD-bfc2d8c231234c35ae8134d885c26f0fe14935e2.tar.gz opensim-SC_OLD-bfc2d8c231234c35ae8134d885c26f0fe14935e2.tar.bz2 opensim-SC_OLD-bfc2d8c231234c35ae8134d885c26f0fe14935e2.tar.xz |
From: Christopher Yeoh <cyeoh@au1.ibm.com>
This patch addresses mantis bug 2576.
http://opensimulator.org/mantis/view.php?id=2576
Briefly, if you call llDie from many scripts at the same time (say a
build is cleaning up excess objects) then OpenSim deadlocks. Avatars
are unable to move, and whilst the console is active you can't do much
without it also locking up. This only occurs with the XEngine script
engine enabled.
I have attached a patch which works, but I'm not sure its the right way
to address the problem. The fundamental problem is that a lock on a
SceneObjectGroup's m_parts is taken when the object is deleted, a
callback to the script engine occurs and a fair way down the callchain,
potentially there are locks taken on several other SceneObjectGroup's
m_parts. Deadlock then occurs if you get unlucky enough
to get in the situation where with several llDie's are called and
SceneObjectGroups
have taken a lock on their own m_parts, and end up waiting on each
other's
locks to become available.
The patch adds a lock at a high level so that that the removal of script
instances
from an object only occurs once per scene at a time. This avoids the
potential
of deadlock. Theoretically there could be some performance hit but
AFAICT
the path taken is not a common occurrence.
Would welcome any suggestions for a better solution, otherwise feel free
to apply :-)
Note this patch was built against the 0.6.0 freeze as trunk was
rather broken for me this morning (creating a script killed the client
connection).
Diffstat (limited to '')
-rw-r--r-- | OpenSim/Region/Environment/Scenes/Scene.cs | 8 |
1 files changed, 7 insertions, 1 deletions
diff --git a/OpenSim/Region/Environment/Scenes/Scene.cs b/OpenSim/Region/Environment/Scenes/Scene.cs index 6634028..f1c4c6c 100644 --- a/OpenSim/Region/Environment/Scenes/Scene.cs +++ b/OpenSim/Region/Environment/Scenes/Scene.cs | |||
@@ -164,6 +164,8 @@ namespace OpenSim.Region.Environment.Scenes | |||
164 | private Thread HeartbeatThread; | 164 | private Thread HeartbeatThread; |
165 | private volatile bool shuttingdown = false; | 165 | private volatile bool shuttingdown = false; |
166 | 166 | ||
167 | private object m_deleting_scene_object = new object(); | ||
168 | |||
167 | #endregion | 169 | #endregion |
168 | 170 | ||
169 | #region Properties | 171 | #region Properties |
@@ -1824,7 +1826,11 @@ namespace OpenSim.Region.Environment.Scenes | |||
1824 | { | 1826 | { |
1825 | //SceneObjectPart rootPart = group.GetChildPart(group.UUID); | 1827 | //SceneObjectPart rootPart = group.GetChildPart(group.UUID); |
1826 | 1828 | ||
1827 | group.RemoveScriptInstances(); | 1829 | // Serialise calls to RemoveScriptInstances to avoid |
1830 | // deadlocking on m_parts inside SceneObjectGroup | ||
1831 | lock (m_deleting_scene_object) { | ||
1832 | group.RemoveScriptInstances(); | ||
1833 | } | ||
1828 | 1834 | ||
1829 | foreach (SceneObjectPart part in group.Children.Values) | 1835 | foreach (SceneObjectPart part in group.Children.Values) |
1830 | { | 1836 | { |