During the heartbeat loop, Update() is called on every SceneObjectGroup which in turn checks if any SceneObjectPart has changed. For large regions (> 100k prims) this work consumes 20-30% of a CPU even though there are only a few objects updating each frame.
There is only one other reason to check every object on every frame, and that is the case where a script has registered the object with an "at target" listener. We can easily track when an object is registered or unregistered with an AtTarget, so this is not a reason to check every object every heartbeat.
In the attached patch, I have added a dictionary to the scene which tracks the objects which have At Targets. Each heartbeat, the AtTarget() function will be called on every object registered with a listener for that event. Also, I added a dictionary to SceneGraph which stores references to objects which have been queued for updates during the heartbeat. At each heartbeat, Update() is called only on the objects which have generated updates during that beat.
* Added some missing implementations of IClientAPI.RemoteEndPoint
* Added a ClientManager.Remove(UUID) overload
* Removed a reference to a missing project from prebuild.xml
* Removed the confusing (and LL-specific) shutdowncircuit parameter from IClientAPI.Close()
* Updated the LLUDP code to only use ClientManager instead of trying to synchronize ClientManager and m_clients
* Remove clients asynchronously since it is a very slow operation (including a 2000ms sleep)
* Move ViewerEffect handling to Scene.PacketHandlers
* Removing the unused CloseAllAgents function
* Trimming ClientManager down. This class needs to be reworked to keep LLUDP circuit codes from intruding into the abstract OpenSim core code
* This causes time to be counted in ODECharacter and, when a collision occurs, the physics scene will report the collisions only if the the difference of last time it reported the collisions from now was more then the set ms.
* This is cool because the time accrues while collisions are not taking place and when they do take place again, you get an immediate update.
* Shrinks the largest in-memory object, the LLRAW.HeightmapLookupValue struct (only used for exporting to LLRAW terrain files), to the minimum possible size. This seems to have the odd side effect of cutting the size of the two double[256,256] terrain objects in half. Possibly an alignment optimization?
* Inspect incoming TextureEntry updates for bakes that do not exist on the simulator and request the missing textures
* Properly handle appearance updates that do not have a TextureEntry set
* Re-added openjpeg-dotnet files since they are used elsewhere in OpenSim * Updated prebuild.xml with a reference to CSJ2K
* Renamed IJ2KDecoder and J2KDecoder member names to follow standard naming conventions * Removed j2kDecodeCache cruft and replaced it with the OpenSim cache system * Rewrote the default layer boundary algorithm to use percentages instead of an exponent * Switched from an infinite in-memory cache to an expiring cache (10 minute timeout) * Slightly quieted logging errors for failed texture decodes
always behave like they are phantom, and llVolumeDetect
seems to operate in a random fashion.
Signed-off-by: dr scofield (aka dirk husemann) <drscofield@xyzzyxyzzy.net>
Commented out the Standalone teleport test because it's failing, and the scene setup is very confusing. I suspect it may be wrong -- the connectors-as-ISharedRegionModules are being instantiated several times when there are several scenes.
* Fixes 'Inconsistent Attachment State' when teleporting into another region besides the SW most region slot on a MegaRegion.
* Fixes a host of other unintended attachment border cross edge cases that lead to Inconsistent attachment state.
This patch allows the land owner to dynamically set the SIP address of a particular land parcel from script. This allows predetermined SIP addresses to be used, making it easier to allow non OpenSim users to join a regions voice channel.
Signed-off-by: dr scofield (aka dirk husemann) <drscofield@xyzzyxyzzy.net>
Eliminate dynamic capsule wobble. Instead introduce a small, fixed
tilt, and allow the tilt to rotate with the avatar while moving; the
tilt always faces away from the direction of avatar movement. The
rotation while moving should eliminate direction-dependent behavior
(e.g. only being able to climb on top of prims from certain directions).
Falling animation is still too frequently invoked.
Ideally the tilt should be completely eliminated, but doing so
currently causes the avatar to fall through the terrain.
llDie() not always completely working, I discovered I was
getting a lot (60+ over 6000 iterations of the test case)
null pointer exceptions in various physics related checks in
SceneObjectPart. It was apparent that the (frequent) checks for
PhysActor being non-null is an insufficient protection in a
highly asynchronous environment. The null reference exceptions
are one example of failure, but it could also happen that a
sequence started with one instance of a PhysicsActor might
finish with another?
Anyway, I have implemented a safer mechanism that should
stop the errors. I re-ran my test case with the fix in place,
and completed nearly 1000 iterations without a single occurrence.
SceneObjectPart is seriously in need of rejigging, if not for
this reason, then for its ridiculous size.
Signed-off-by: dr scofield (aka dirk husemann) <drscofield@xyzzyxyzzy.net>
* Changed the Destination ID to 0 in the TeleportFinish Event (why did we have it as 3?)
* Added border based trigger teleports
* Fix MakeRootAgent border cross tests for ensuring that the position is inside the region to use the borders to figure out if it's outside the Region
* Uses a grid request multiple times, so the more regions are in the instance, the longer logging in takes. Consider this temporary until there's a RegionInfo request similar to the MapItem Request.
have been in this commit slipped in. If the last heartbeat is more than
2 seconds ago, kill that thread and start a new one. Untested.
his commit adds support to let the first heartbeat complete unconditionally,
since it is almost always longer.
* RegionCombinerModule replaces this delegate and distributes the CoarseLocationUpdates through the client connection in the region where the user would be if it was a separate region.
* Fixes Mini Map display on combined regions.
* Breaks up the Client event registrations in Scene into methods of similar event types. The generic event registration registers all of the event types.
*Created symmetrical Un-Registration methods.
* Made the Registration and Unregistration methods public
* Hooks The events required for creating new prim and rezzing new prim into the proper scene.
* This also contains some inactive preliminary code for disconnecting combined regions that will be used to make one root region a virtual region of a new root region.
* This time it might be the listening socket thread from HttpServer aborting with a non blocking thread abort exception. Hopefully calling Stop() on MainServer.Instance will solve that.
* LongVersion nIni may be causing the test thread death. Pausing OpenSimulator during startup causes a nIni error that makes debugging startup operations difficult for users. It might be because when it's in pause mode, something else reads from the nini config passed? If it is, it might not be fixable.. however, if it's concurrency that causes nini death it would make sense to give each section of the tests a new IConfigSource so that they don't read from the same configsource at the same time.
* Added a base inventory connector that does common processing of inventory among all reference connector implementations. E.g. AddItem requires additional processing before being forwarded to service.
* Added if (m_Enabled) upon RemoveRegion
|Date: Wed, 5 Aug 2009 09:51:52 -0700
|Subject: [PATCH] Closed two major memory leaks for scripted objects
|
|Two major memory leaks for the scripted objects were fixed
|- One leak had to do with remoting acrossing app domains. When a script and
| its controlling agent communicate across an application boundary, it calls
| functions on a stub proxy object that then invokes the remote method on
| the object in the other app domain. These stub objects (two for each script)
| were setup to have infinate lifetimes and were never being garbage collected.
|- The second leak was the result of adding a scene object part instance method
| to a scene event and never removing it. This cause the event's delegate list
| to maintain a link to that object which is then never freed as the scene event
| object is never destroyed.
Patch applied, please direct feedback to me. Possible issue: Longtime idle
scripts like vendors may fail.
[1] Added a new OnAttach event to Scene/EventManager
[2] Hooked up existing attach event handler in XEngine.
[3] Modified SceneGraph and Scene.Inventory to trigger
attach events at the appropriate places. I was forced
to distribut the changes across two files because of
the way attach processing is distributed across the
two files.
[4] In the case of RezSingleAttachmentFromInventory it is
necessary to handle event scheduling in a special way.
There is no synchronous path available, so the fact
that this object is attached, and who it is attached to,
is cached when the ScriptInstance is created. When
the script is started, the attached handler is driven
after on_rez (but before changed, this should be reviewed).
Signed-off-by: dr scofield (aka dirk husemann) <drscofield@xyzzyxyzzy.net>
* Test for prim obstructions between the avatar and camera. If there are obstructions, inform the client to move the camera closer. This makes it so that walls and objects don't obstruct your view while you're moving around. Try walking inside a hollowed tori. You'll see how much easier it is now because your camera automatically moves closer so you can still see.
* Created a way to know if the user's camera is alt + cammed or just following the avatar.
* Changes IClientAPI interface by adding SendCameraConstraint(Vector4 CameraConstraint)
This involved adding a new OnUpdatePrimSingleRotationPosition event to IClientAPI so that we can get the changed position from the client.
Btw adding new events to IClientAPI is really tedious where you have to copy the change across to at least 5 or 6 other files.
[Note this doesn't fix the bug where any rotation changes to the root prim (but not the whole linkset) cause rotation errors on the child prims.]
* Remove the draconic locking around adding an avatar to the Scene
* Handle an extreme error case when border crossing fails and user uses map to teleport to a different region on the same instance causing control commands to go to a child agent.
* Make the Set Appearance method use the proper 'remove from physics scene' method.
* It *may* help border crossings.
* It *may* help the 'on avatar rez' lag, that people have been seeing the past week.
* It may also cause physics to crash more often on failed teleports (though.. I think I got the cases covered).
The patch included updates the root and child prims' AttchedAvatar
with the right UUID. It also cleans the AttachedAvatar properties
for the root and child prims on Drop and Detach
Attached is a patch that changes the oar file saving of creation date/time to an integer
instead of a string. I did this after justincc emailed me saying there is a problem
with internationalisation doing it the old way and I said I'd fix it. Its been
tested with MySQL and I've made the changes for MSSQL but that hasn't been well tested.
By extracting this code, it should make it easier to experiment with different ways of managing the update process. [Next step to make this module based, could be to create a SceneViewerFactoryModule]
I also fixed the issue where the code also loops through the m_forcesList and copies each force to the ScenePresence's movementVector. Which resulted in only the last force in the list actually be acted on. As each copy overrode the last one. So now it only copies the last force in the list.
out which client connection has closed. So, in multi-region sims, things
can get messy fast.
This introduces a second parameters, which is a Scene object ref. Minor
adjustments to custom modules may be required due to this change.
- Set OwnerID = GroupID for deeded objects.
- Close a security loophole that would have allowed a user with deed rights in a group to deed ANY object to that group, even if it's not owned by them and/or not set to that group
- Set LastOwnerID correctly. Handle objects vs. prims correctly.
Several improvements in the connectors themselves.
Several improvements in configurations.
Needed to add a hack in IUserService and UserManagerBase, to be removed when that service is refactored.
This change addresses two issues:
[1] It adds a flag field to the blendface call which allows the
caller to indicate whether or not the generated asset is
temporary, and whether or not the asset being replaced should
be explicitly retired fromt the memory cache. The decimal
values correspond to:
0 - Permanent asset, do not expire old asset
1 - Permanent asset, expire old asset
2 - Temporary asset, do not expire old asset
3 - Temporary asset, expire old asset
'3' corresponds to the default behavior seen today, and is
the continued behavior of the non-blendface calls.
[2] The dynamic texture routines are highly-asynchronous and can
be scheduled simultaneously on a multi-core machine. The nature
of the texture management interfaece is such that updates may
be lost, and the nature of asynchornous operation means that
they may be processed out of order. A lock has been added to
ensure that updates are at least atomic. No attempt has been
made to enforce ordering. The lock applies to the SceneObjectPart
being updated and is held for the lifetime of the TextureEntry
used to carry texture updates (the one instance carries all
faces supported by the prim).
Users of these services should remember that the dynamic texture
call is asynchronous and control will be returned *before* the
texture update has actually occurred. As a result, a isubsequent
GetTexture call may not return the expected asset id. A script
must wait for the corresponding TEXTURE_CHANGED event before
retrieving any texture information.
* For some reason, if a null was recieved (indicating a missing asset), the code had stopped passing that on to the waiting lock, resulting in a perpetual freeze
* This change passes the null on correctly
* Many thanks to thomax for being insistent in presenting his analysis of the problem :)
The attached patch implements llPassTouches. It has been added
to the export/import XML along with the flag for AllowedInventoryDrop.
The MySQL backend has been updated as well, though I haven't
done one of those before so could do with a check. I added
the migration mysql file as well.
The other data backends need updating as well.
The attached patch implements llPassTouches. It has been added
to the export/import XML along with the flag for AllowedInventoryDrop.
The MySQL backend has been updated as well, though I haven't
done one of those before so could do with a check. I added
the migration mysql file as well.
The other data backends need updating as well.
Changes to support client-side image pre-caching in the region. This
commit adds an additional calling sequence to the DynamicTexture data
and URL calls. The new interface allows a dynamic image to be loaded
into a specific object face (rather than the mandatory ALL_SIDES
supported today.
This is in part fulfilment of ticket #458.
This may break a lot of things, but it needs to go in. It was tested in standalone and the UCI grid, but it needs a lot more testing.
Known problems:
* HG asset transfers are borked for now
* missing texture is missing
* 3 unit tests commented out for now
further ACL stuff:
- adding StrictAccessControl variable: DON'T set this to false if you
want to enforce ACL, it will disable ACLs right now. Default is
true.
once we've got code added to allow child agents but prevent them
from becoming root agents when the ACL denies access to the avatar,
setting this to false will then allow avatars to see into a
neighboring region but not enter it (currently ACL prevent both,
seeing and entering).
- enhancing log statements
- AssetType isn't InventoryType. Those enums contain different numbers. Use AssetType for the asset type, InventoryType for the inventory type.
- The ToString method (or ToLower) of AssetType/InventoryType doesn't necessarily return the correct LLSD string.
- Replaced several magic numbers by their corresponding enum.
- Fixed the invType for gestures and animations in the library.
This should fix Mantis #3610 and the non-terminating inventory loading
Scene.NewUserConnection()
- adding reason reporting
this enforces estate bans very early on and prevents us from
circulating client objects that we'd then have to retract once we
realize that the client is not allowed into the region
Move json stats to non-published resource name
Remove well-known resource name for json stats,
creating dynamic uris with private keys and add
a user configurable resource name for region owner usage.
Don't allow packets to be resent before they have actually been sent for the
first time. Switch from serializing a packet to get it's length to the LibOMV
provided Length property. Fix resend timing. Fix the use of dangling references
to Acked packets. Fix the packet handler to play nice with the packet pool.
Fix the packet pool. Add data block recycling to the packet pool. Packet pool
is now ENABLED by default. Add config option to disable packet and data block
reuse. Add ObjectUpdate and ImprovedTerseObjectUpdate to the packets being
recycled.
arbitrary combinations of root and child prims from one or multiple
link sets. Please test throughly and consider things UNSTABLE until
this is proven out.
If an avatar is sitting when the client disconnects, the avatar
is not disassociated from the SOG on which (s)he was sat. This
produces any, and varied, effects.
I have updated RemoveCLient in Scene, to check, and stand the
client up immediately prior to disconnect. This seems like the
most robust way to handle the situation. Though in this case
it might be worth factoring out the animations from other
standup processing. It does no harm, but in this case it is
entirely redundant.
Added support for access control lists.
Scene: Added test to AddNewClient for an entry in the access
list when connecting to a region with limited access.
EstateSettings: Added an HasAccess(UUID) property to test for
an entry in the estate's access list.
RemoteAdmin: Add RPC calls for admin_acl_list, clear, add,
and remove.
llSetPrimitiveParams in a large linkset can disrupt the
entire region. However, when the script is in a large
linkset, it appears to totally lag out the scene and
stops updates from being sent.
Changes to enable script state persistence across non-restart
serialization situations (inventory/OAR/attachments)
Also fixing test cases for OAR and IAR so they don't barf with the new code.
An attachment with the physical checkbox checked will not allow the
phantom checkbox to be cleared. This interfers with scripting
functions such as llMoveToTarget(), which won't work while an object
is phantom. If the prim containing the script is rezzed to the ground,
it will then allow the phantom checlbox to be cleared, and the script
works as expected.
This patch adds few properties to ScenePresence and
thus allows region module or MRM script:
1. Force flying for avatar or,
2. Disable flying from avatar
This hooks up the LandManagementModule to handle the DeedParcelToGroup
packet. Now people can start testing land assigned to and owned by groups.
Also fixes a viewer crash issue when searching for and then joining a group
with an agent that is not already being tracked by groups server.
* This may eliminate the occasional archive test freezes, since they appeared to occur when somehow the asset server didn't pick up on the presence of a request in the asset
quque
* WebStatsModule doesn't crash on restart. GodsModule doesn't crash when there is no Dialog Module. LLUDPServer doesn't crash when the Operation was Aborted.
* ODEPlugin does 'Almost NaN' sanity checks.
* ODEPlugin sacrifices NaN avatars to the NaN black hole to appease it and keep it from sucking the rest of the world in.
This is sent to all root prims of all attachments of an avatar when the
animation state changes. llGetAnimation() can thenbe used to find the
new movement animation. This eliminates the need for fast timers in AOs
Addresses Mantis #3381
The current implementation works as expected if the object has no rotation or
only rotation around the Z axis; you can spin the object left or right (around
the world Z axis).
It works a little unexpectedly if the object has a non-Z-axis rotation; in this
case the body is spun about its local Z axis, not the world Z-axis. (But SL
also behaves oddly with a spin on an arbitrarily rotated object.)
Dan Lake