This also makes the detection in SP.FindNextAvailableSitTarget() and SendSitResponse() identical.
Previously they varied slightly (SendSitResponse didn't check for an older type of invalid quaternion) but the practical effect is most probably zero.
This is to fix a regression starting from 5301648 where attachments had to start being deleted before persistence in order to avoid race conditions with hud update threads.
This prevents a stack overflow where a get position on the avatar will refer to the attachment which will in turn refer back to the avatar.
This required recording of all sitting avatars on a prim which is done separately from recording the sit target avatar.
Recording HashSet is null if there are no sitting avatars in order to save memory.
This is always done later on in SceneGraph.AddSceneObject() if the call hasn't failed due to sanity checks.
There's no other purpose for this method to exist and it's dangerous/pointless to call in other conditions.
This works like osForceAttachToAvatar() but allows an object to be directly specified from the script object's inventory rather than forcing it to be rezzed in the scene first.
Still only attaches objects to the owner of the script.
This allows one to bypass the complicated co-ordination of first rezzing objects in the scene before attaching them.
Threat level high.
If delete doesn't occur first then the update thread can outrace the IsAttachment = false necessary to save attachments and send hud artifacts to other viewers.
Serialization of attachments requires IsAttachment = false so that correct positions are serialized instead of avatar position.
However, doing this when a hud is still attached allows race conditions with update threads, resulting in hud artifacts on other viewers.
This change sets SOG.IsDeleted before serialization changes take place (IsDeleted itself is not a serialized property).
LLClientView then screens out any deleted SOGs before sending updates to viewers.
Change SOP.SendFullUpdateToClient() and SoundModule.PlayAttachedSound() to use this rather than different magic number formulations.
This also corrects a bug in PlayAttachedSound() where the code assumed that all attachment points over 30 were HUDs.
It appears this is no longer true with Neck and Root (Avatar Center)
Stop checking IsLoggingOut on these listeners, if called with a root agent then we always want to perform these actions.
This covers cases where the client is closed due to manual kick, simulator shutdown, etc.
This is always done as part of Scene.RemoveClient()
Also refactors try/catching in Scene.RemoveClient() to log NREs instead of silently discarding, since these are useful symptoms of problems.
This is done by making the kick user command call IClientAPI.Close() rather than routing through Scene.IncomingCloseAgent(), which also called IClientAPI.Close()
DisableSimulator for child agents is moved from IncomingCloseAgent() to RemoveClient(), this is not a functional change since IncomingCloseAgent() always ends up calling RemoveClient()
This makes frame time stats properly tally with fps, which saves confusion and makes it easier to interpret numbers.
In some ways this is not so artifical - physics FPS runs at the higher rate.
This is to make these statistics actually match their names (and also be more accurate as number of frames can vary under heavy load)
Currently using scene frames (11.23 every second) instead of physics frames (56.18 per second)
If active, the physics module can return arbitrary stat counters that can be seen via the MonitoringModule
(http://opensimulator.org/wiki/Monitoring_Module)
This is only active in OdeScene if collect_stats = true in [ODEPhysicsSettings].
This patch allows OdeScene to collect elapsed time information for calls to the ODE native collision methods to assess what proportion of time this takes compared to total physics processing.
This data is returned as ODENativeCollisionFrameMS in the monitoring module, updated every 3 seconds.
The performance effect of collecting stats is probably extremely minor, dwarfed by the rest of the physics code.
If this is done before then on ODE agent update calls still incoming can fail as they try to use a raycastmanager that has been disposed.
Bullet plugin does nothing on Dispose()
However, I wouldn't be at all surprised if individual region restarting was buggy in lots of other areas.
If this is allowed, then the client usually gets forcibly logged out and data structures might be put into bad states.
To prevent this, the binary state machine of EMT.m_agentsInTransit is replaced with a 4 state machine (Preparing, Transferring, ReceivedAtDestination, CleaningUp).
This is necessary because the source region needs to know when the destination region has received the user but a teleport back cannot happen until the source region has cleaned up.
Tested on standalone, grid and with v1 and v3 clients.
This is to help relieve a race condition when an agent teleports then immediately attempts to teleport back before the source region has properly cleaned up/demoted the old ScenePresence.
This is rare in viewers but much more possible via scripting or region module.
However, more needs to be done since virtually all clean up happens after the transit flag is cleared .
Possibly need to add a 'cleaning up' state to in transit.
This change required making the EntityTransferModule and HGEntityTransferModule per-region rather than shared, in order to allow separate transit lists.
Changes were also required in LocalSimulationConnector.
Tested in standalone, grid and with local and remote region crossings with attachments.
This includes prim count, script count, avatar count, etc.
Information is currently the same as "show stats", though show stats can only show one scene at a time because it listens for the latest outgoing stats packet (a bad approach that needs to change).
Might be better to tie this module into the other stats module to display arbitrary stats rather than fetching directly from scene.SimStatsReporter.
Console command is "show scene" because "show region" already exists for the grid service, which is unfortunate.
Might need to make a distinction between "scene" relating to a live scene and "region" relating to more static region data (url, coords, etc.)
This is because the attachments module code was setting the 'object slam' bit by using PermissionMask.All
Solution here is to route the attachment item creation call through the existing inventory code in BasicInventoryAccessModule rather than copy/pasted code in AttachmentsModule itself.
From sl docs such as http://community.secondlife.com/t5/English-Knowledge-Base/Managing-Private-Regions/ta-p/700115
agent should apply to avatars only.
This makes sense from a user perspective, and also from a code perspective since child agents with no physics or actions take up a fraction of root agent resources.
As such, the check is now only performed in Scene.QueryAccess() - cross and teleport check this before allowing an agent to translocate.
This also removes an off-by-one error that could occur in certain circumstances on teleport when a new child agent was double counted when a pre-teleport agent update was performed.
This does not affect an existing bug where limits or other QueryAccess() checks are not applied to avatars logging directly into a region.
This was because the parts in scene objects were sometimes not serialized in link order.
This is perfectly fine since the parts still have the right link numbers, but an extra fix to adjust for this
had not been done in the SerialiserModule methods that OAR loading used.
Add regression test for same.
Addresses http://opensimulator.org/mantis/view.php?id=5948, http://opensimulator.org/mantis/view.php?id=5749
Will use one of three selected methods to route avatar landing
points when using Telehubs. The setting is in [Startup] using
SpawnPointRouting = closest/random/sequence
closest: The default setting. Routes avatar to the nearest SpawnPoint
to the location.
random: Picks random SpawnPoints to land the avatar.
sequence: Follows a sequence to place the avatar on the next available
SpawnPoint location
Conflicts:
OpenSim/Region/Framework/Scenes/Scene.cs
DuplicateObject() relies on source object having correct link numbers for the duration of the dupe.
Both link and delink can change link numbers such that they are not consistent for short periods of time.
Viewers 1 and 3 are fine with doing this immediately. However, Imprudence has a small delay (<200ms, >500ms) after receiving the AgentCompleteMovement reply packet on the destination region before regarding that region as the currnet region.
If Imprudence receives a DisableSimulator in this period, it quits.
We are not restoring the full 5000ms delay since this brings back a bug where teleports permanently fail if an avatar tries to teleport back too quickly.
This commit also sends the AgentCompleteMovement packet to the client before telling the source region to release its old agent, in order to further cut down any possibility of the DisableSimulator being recieved before the AgentMovementComplete.
This adds a non-advertised wait_for_callback option in [EntityTransfer]. Default is always true.
Teleport tests disable the wait for callback from the destination region in order to run within a single thread.
This happened because the scripts were notified about control changes (e.g., the user stopped pressing the Forward key) when the animation was still WALK, so the script didn't stop the walking animation. Fixing this required: a) Update the movement animation *before* notifying the script; b) Add locking to prevent clashes with the Heartbeat thread (which also updates the animations); c) Handle the case of a user who stops walking just as the avatar is in the air: the avatar should STAND in that case, not WALK.
This reverts commit feef1dd732.
This involves a large amount of change in test scene setup code to allow test scenes to share shared modules
SetupScene is now an instance method that requires an instantiation of SceneHelpers, though other SceneHelpers methods are still static
May split these out into separate classes in the future.
folder when deleting objects from a scene. The use of the trash
folder causes assets to be created and stored everytime you delete
an object from the scene (slows down the delete and adds mostly useless
assets to your database).
Default is on (use the trash folder) which is the standard behavior.
These are identical apart from setting Velocity = zero, which has no practical effect anyway since this is zeroed when the avatar is added back to the physics scene.
This was an oversight when removing some race conditions from PhysicsActor setting recently.
Regression tests extended to probe this code path.
Extending regression tests required implementation of a BasicPhysicsPrim (there was none before). However, BasicPhysics plugin is still of no current practical use other than to fill in as a component for other parts of regression testing.
On region restart, the whole object would become physical as expected.
Observed behaviour from elsewhere is that all prims in a new linkset should take on the status of the root prim.
Add regression test for this behaviour.
This makes it consistent with other parts of OpenSimulator that are treating ESTATE_MANAGER and ESTATE_OWNER as different entities.
As per opensim-dev mailing list.
Another thread could come and turn off physics for a part (null PhysicsActor) at any point.
Had to turn off localCopy on warp3D CoreModules section in prebuild.xml since on current nant this copies all DLLs in bin/ which can be a very large number with compiled DLLs
No obvious reason for doing that copy - nothing else does it.
This adds ScenePresence to IClientAPI.SceneAgent earlier on in the add client process so that its information is available to EventManager.OnNewClient() and OnClientLogin()
Also add a code comment as to why we're caching friend information for child agents.
We need to cache child agents so that friends object edit/delete permissions will work across boarders on regions hosted by different simulators.
This reverts commit d9f7b8549b.
This allows us to avoid unnecessary multiple calls to the friends service.
All friends functions originate from the root agent and only go to other root agents in existing code.
This also allows us to eliminate complex ref counting.
On osgrid and other places, I have observed that manually sending appearance updates from the console often relieves grey avatar syndrome.
Despite hunting high and low, I haven't been able to find where this packet is sometimes being lost - it might be a persistent viewer bug for all I know.
Therefore, this experimental setting resends appearance data for everybody in the scene every 60 seconds. These packets are small and the viewer only fetches texture
data if it doesn't already have it.
Default is false.
The idea is to make the critical main scene loop as skinny as possible - it doesn't need to run things that aren't time critical and don't depend on update ordering.
This will be done gradually over time to try and uncover any issues. Many non-criticial scene loop activities are being launched on separate threadpool threads anyway.
This may also allow modules to register their own maintenance jobs without having to maintain their own timers and threads.
Currently the maintenance loop runs once a second, as opposed to the 89ms scene loop.
This lets us remove the dependency of OpenSim.Framework.dll on data/avataranimations.xml, which is not necessary for ROBUST.
This commit also takes care of the odd situation where animations are stored and used internally with uppercase names (e.g. "STAND")
but scripts refer to them with lowercase names (e.g. "sit").
Update() now accepts a frames parameter which can control the number of frames updated.
-1 will update until shutdown.
The watchdog updating moves above the maintc recalculation for any required sleep since it should be accounted for within the frame.
The previous lines-per-second measurement used for top scripts report was inaccurate, since lines executed does not reflect time taken to execute.
Also, every fetch of the report would reset all the numbers limiting its usefulness and we weren't even guaranteed to see the top 100.
The actual measurement value should be script execution time per frame but XEngine does not work this way.
Therefore, we use actual script execution time scaled by the measurement period and an idealised frame time.
This is still not ideal but gives reasonable results and allows scripts to be compared.
This commit moves script execution time calculations from SceneGraph into IScriptModule implementations.
The first llDie() could lock Scene.m_deleting_scene_object.
The second llDie() would then wait at this lock.
The first llDie() would go on to remove the second script but always abort it since the second script's WorkItem would not go away.
Easiest solution here is to remove the m_deleting_scene_object since it's no longer justified - we no longer lock m_parts but take a copy instead.
This also requires an adjustment in XEngine.OnRemoveScript not to use instance.ObjectID instead when firing the OnObjectRemoved event.
However, it looks like we should retain SP.ParentID since it's much easier to use that in places where another thread could change ParentPart to null.
Otherwise one has to clumsily put ParentPart in a reference, etc. to avoid a race.
This is to deal with the hundred lines of command splurge when one previously typed "help"
Modelled somewhat on the mysql console
One can still type help <command> to get per command help at any point.
Categories capitalized to avoid conflict with the all-lowercase commands (except for commander system, as of yet).
Does not affect command parsing or any other aspects of the console apart from the help system.
Backwards compatible with existing modules.
We can now do this since the entire scene and all objects within it are now successfully gc'd at the end of these tests.
This greatly improves the time taken to run each test (by reducing teardown time, not the time to actually do the test work that we're interested in).
Slightly simplifies config read in Scene constructor to help facilitate this.
OnNewScript fires when a script is added to a scene
OnRezScript fires when the script actually runs (i.e. after permission checks, state retrieval, etc.)
This allows scripts to run in child prims that are outside region boundaries.
This is an interim patch applied from http://opensimulator.org/mantis/view.php?id=5899 though it does not resolve that bug
Thanks tglion!
On the first frame, all startup scene objects are added to the physics scene.
This can cause a considerable delay, so we don't start raising the alarm on scene loop timeouts until the second frame.
This commit also slightly changes the behaviour of timeout reporting.
Previously, a report was made for the very first timed out thread, ignoring all others until the next watchdog check.
Instead, we now report every timed out thread, though we still only do this once no matter how long the timeout.
This involves
1) On forcible teleport, call m_scene.RequestTeleportLocation() rather than ScenePresence.Teleport() - only EntityTransferModule now should call SP.Teleport()
2) When avatar is being forcibly moved due to banlines, use a 'stop movement' tolerance of 0.2 to requested position rather than 1
This prevents the avatar sometimes being stuck to banlines until they teleport somewhere else.
This aims to fix some problems in http://opensimulator.org/mantis/view.php?id=5822
potential bad update that places an object at the opposite side of the
origin sim for a moment before actually crossing it. Especially important in
grids like OSG where lag between sims is high.
This fixes the problem by fixing the permissions module to look at root part permissions rather than having to do this for every caller.
Resolves http://opensimulator.org/mantis/view.php?id=5569
At least on mono 2.6.4, running GC.Collect() is not guaranteed to force gc of all objects when run in the same method where those objects had references.
Therefore, GC.Collect() is now being done in the per-script teardown of ObjectTortureTests.
In addition, scene loop update is being run after garbage collection in order to clean out the viewer update list of scene objects in the SceneGraph.
These measures mean that scene objects/parts are now garbage collected after a test run if deleted from the scene, resulting in a much better memory usage report (though probably still not very accurate).
However, deletion takes a very long time - what's really needed is to find out now why the entire scene isn't being GC'd by this measure.
This change hasn't yet been applied to the other stress tests.
Logging level was DEBUG before 312e145 (Fri Feb 3 2012).
312e145 also accidentally removed the 'general error' log message if any shape deserialization failed.
This commit restores it, though this has no functional impact.
Justin Clark-Casey (justincc)