When a player triggers a chunk load via walking around or teleporting there
is no need to stop everything and get this chunk on the main thread. The
client is used to having to wait some time for this chunk and the server
doesn't immediately do anything with it except send it to the player. At
the same time chunk loading is the last major source of file IO that still
runs on the main thread.
These two facts make it possible to offload chunks loaded for this reason
to another thread. However, not all parts of chunk loading can happen off
the main thread. For this we use the new AsynchronousExecutor system to
split chunk loading in to three pieces. The first is loading data from
disk, decompressing it, and parsing it in to an NBT structure. The second
piece is creating entities and tile entities in the chunk and adding them
to the world, this is still done on the main thread. The third piece is
informing everyone who requested a chunk load that the load is finished.
For this we register callbacks and then run them on the main thread once
the previous two stages are finished.
There are still cases where a chunk is needed immediately and these will
still trigger chunk loading entirely on the main thread. The most obvious
case is plugins using the API to request a chunk load. We also must load
the chunk immediately when something in the world tries to access it. In
these cases we ignore any possibly pending or in progress chunk loading
that is happening asynchronously as we will have the chunk loaded by the
time they are finished.
The hope is that overall this system will result in less CPU time and
pauses due to blocking file IO on the main thread thus giving more
consistent performance. Testing so far has shown that this also speeds up
chunk loading client side although some of this is likely to be because
we are sending less chunks at once for the client to process.
Thanks for @ammaraskar for help with the implementation of this feature.
Adds:
- Getting/Setting equipment
- getting/setting drop rates
- getting/setting ability to pick up items
-- As an added feature, players with this flag start off with a canceled PlayerPickupItemEvent
Currently when a plugin wants to get the location of something it calls
getLocation() which returns a new Location object. In some scenarios this
can cause enough object creation/destruction churn to be a significant
overhead. For this cases we add a method that updates a provided Location
object so there is no object creation done. This allows well written code
to work on several locations with only a single Location object getting
created.
Providing a more efficient way to set a location was also looked at but
the current solution is the fastest we can provide. You are not required
to create a new Location object every time you want to set something's
location so, with proper design, you can set locations with only a single
Location object being created.
As of 1.4 mobs have a flag to determine if they despawn when away from a
player or not. Unfortunately animals still use their own system to prevent
despawning instead of making use of this flag. This change modifies them
to use the new system (defaults to true) and to add API for plugins to adjust
this.
Stale player references will add a player back into the world when
teleporting them, causing a cascade of issues relating to ghost entities
and servers failing to stop.
This is a missed part of the original "[Bleeding] Use case from player data
for OfflinePlayer. Fixes BUKKIT-519" commit. It avoids doing (somewhat
expensive) lookups of player data to find the correct capitalization inside
getOfflinePlayers() as we're already loading their name from the player data
and thus have the correct capitalization.
If a plugin looks up a player that is offline they may not know the correct
capitalization for the name. In this case they're likely to get it wrong
and since we cache the result even after the player joins the server all
future request for an OfflinePlayer will return one with incorrect case.
When looking up a player who has played on the server before we can
get the correct case from the player data file saved by the server. If
the player has never played before this point we cannot do anything and
will still have the same issue but this is not a solvable problem.
Skulls need their tile entity in order to create an item correctly when
broken unlike every other block. Instead of sprinkling special cases all
over the code just override dropNaturally for skulls to read from their
tile entity and make sure everything that wants to drop them calls this
method before removing the block. There is only one case where this wasn't
already true so we end up with much less special casing.
The static assertions are not normally evaluated in the JVM, and failed
to fail when the enums went from size 25 to size 26. This meant missing
values would not be detected at runtime and instead return null,
compounding problems later. The switches should never evaluate to null
so will instead throw runtime assertion errors.
Additional unit tests were added to detect new paintings and assure they
have proper, unique mappings. The test checks both that a mapping
exists, is not null, and does not duplicate another mapping.
If a defensive copy is not used in the API, changes to the item are
reflected in memory, but never updated to the client. It also goes
against the general contract provided in Bukkit, where setItem should be
the only way to change the underlying item frame.
Skull blocks store their type in a tile entity and use their block data
as rotation. When breaking a block the block data is used for determining
what item to drop. Simply changing this to use the skull method for getting
their drop data is not enough because their tile entity is already gone.
Therefore we have to special case skulls to get the correct data _and_ get
that data before breaking the block.
CommandMap now contains the functionality for tab completion. This
commit replaces the vanilla implementation and simply delegates it to
the Bukkit API.
This change affects the old chat compatibility layer from an
implementation only standpoint. It does not queue the 'event' to fire,
but rather queues a runnable that allows the calling thread to wait for
execution to finish.
The other effect of this change is that rcon connects now have their
commands queued to be run on next server tick using the same
implementation.
The internal implementation is in org.bukkit.craftbukkit.util.Waitable.
It is very similar to a Future<T> task, but only contains minimal
implementation with object.wait() and object.notify() calls
under the hood of waitable.get() and waitable.run().
PlayerPreLoginEvent now properly implements thread-safe event execution
by queuing the events similar to chat and rcon. This is still a poor way
albeit proper way to implement thread-safety; PlayerPreLoginEvent will
stay deprecated.
The implementation for the new methods mimics the old methods. The final
call for the old methods now maps to the new methods with an additional
call to get id.
If two players (or a player and any other entity) are teleported to the
same location in the same tick they will both get added to the other's
destroy queue then have a new entity spawn packet sent. Next tick the
destroy queue will be processed and they will then be invisible to each
other. To prevent this situation we remove the entity from the destroy
queue when sending out a spawn packet for them.
If a plugin calls player.hidePlayer(other); then player.showPlayer(other);
in the same tick the other player will be added to the entity destroy queue
then a spawn packet will be sent. On the next tick the queue will be
processed and a destroy packet will be sent that renders the other player
invisible. To correct this we ensure the destroy queue is in sync with use
of the vanish API.
An internal method for making the debug output for CraftScheduler's
async tasks was erroneously using the 'this' reference when the loop
should be referencing the current task.
This change was done to remove the internal sound names from the API.
Along with moving the internal names into CraftBukkit, a unit test was
added for any new sounds added in the API to assure they have a non-null
mapping.
After further testing it appears that while the original LongHashtable
has issues with object creation churn and is severly slower than even
java.util.HashMap in general case benchmarks it is in fact very efficient
for our use case.
With this in mind I wrote a replacement LongObjectHashMap modeled after
LongHashtable. Unlike the original implementation this one does not use
Entry objects for storage so does not have the same object creation churn.
It also uses a 2D array instead of a 3D one and does not use a cache as
benchmarking shows this is more efficient. The "bucket size" was chosen
based on benchmarking performance of the HashMap with contents that would
be plausible for a 200+ player server. This means it uses a little extra
memory for smaller servers but almost always uses less than the normal
java.util.HashMap.
To make up for the original LongHashtable being a poor choice for generic
datasets I added a mixer to the new implementation based on code from
MurmurHash. While this has no noticable effect positive or negative with
our normal use of chunk coordinates it makes the HashMap perform just as
well with nearly any kind of dataset.
After these changes ChunkProviderServer.isChunkLoaded() goes from using
20% CPU time while sampling to not even showing up after 45 minutes of
sampling due to the CPU usage being too low to be noticed.
The new setting is located at "ticks-per.autosave". By changing this
value, it affects how often a full save is automatically executed,
measured in ticks.
This value is defaulting to 0 (off) because we believe that the vast
majority of servers already have a third-party solution to automatically
saving the server at set intervals. Having the built in auto-save disabled
by default ensures that we are not saving things twice; doing so leads to
absolutely no benefits, but results in detrimental and noticeable
unnecessary performance decrease.
For servers that do not use an automated external script to perform saves,
this setting can be turned on by setting the value higher than 0, with 900
being the value used in vanilla.
Refactoring dependencies 'changes' the string literal in the code. This
commit changes the literal to instead use a char[] to initialize a new
String. On a bytecode level, there will not exist a String literal for these
two values; the shade plugin will no longer refactor them.
Refactoring jline also changes the other String literals we use for
notifying jline of the current state. To insure that our local code reflects
the inner logic in jline, the key value was changed to the static final
variable located in TerminalFactory. Likewise, UnsupportedTerminal uses the
explicit class name (as reflection is used later with the value that has
been set).
Async tasks are notorious for causing CMEs and corrupted data when
accessing the API. This change makes a linked list to track recent tasks
that may no longer be running. It is accessed via the toString method on
the scheduler. This behavior is not guaranteed, but it is accessible as
such currently.
Although toString is located in the scheduler, its contract does not
guarantee an accurate or up to date call when accessed from a second
thread.
When 1.3.1 was released, a try-catch block was removed from the tick
loop that called the method in NMS to handle commands. This restores a
try-catch to prevent the console from crashing the server.
Previously, the timeout would erroneously get converted to milliseconds
twice. The second conversion was removed.
Spurious wakeups were not handled properly, and would instead throw a
TimeoutException even if the waited time was not reached..
The new scheduler uses a non-blocking methodology. Combining volatile
references to make a linked reference chain, with the atomic reference
handling the tail, tasks are queued without waiting for locks. The main
thread will no longer limit the length of time spend for scheduled tasks,
but no task will run twice in the same tick. Scheduling a new task inside of
a synchronous task will always run the new task during the same tick,
assuming there is no supplied delay > 0.
Asynchronous tasks are now run using a thread pool. Any thread-local
implemenation should now account for threads being reused between
executions.
Race conditions were carefully examined and the order of logic is now very
important. Each task is placed in a secondary collection before removal from
primary collections. Thus, by reading tasks from the collections in the same
order they travel, it retains state-safety. This does make modifications
less responsive in some situations, as the task may be transitioning before
the modifier accesses it. This cost outweighs the requirement to synchronize
on the scheduler; previously any conflict would be first-come-first-serve,
with the main thread backing out arbitrarily.
Many codepaths only end up with one iterator being used at a time and
most of the rest only get up to two being used so using a static pool of
three is wasteful. This also allows us to efficiently handle cases that
exceed 3 iterators in use. Overall this dramatically increases the hit rate
and results in less iterators being created.
This ArrayList duplicates part of the functionality of the much more
efficient chunk map so can be removed as the map can be used in the few
places this was needed.
Replace uses of LongHashtable and LongHashset with new implementations.
Remove EntryBase, LongBaseHashtable, LongHashset, and LongHashtable as they
are no longer used.
LongObjectHashMap does not use Entry or EntryBase classes internally for
storage so has much lower object churn and greater performance. LongHashSet
is not as much of performance win for our use case but for general use is
up to seventeen times faster than the old implementation and is in fact
faster than alternatives from "high performance" java libraries. This is
being added so that if someone tries to use it in the future in a place
unrelated to its current use they don't accidentally end up with something
slower than the Java collections HashSet implementation.
They can spawn any valid entities now. What is a "valid" entity? A "valid" entity is an EntityType with a non-null getName(). (for example: PRIMED_TNT, FALLING_BLOCK)
Added two utility collections for use with PlayerChatEvents allowing lazier
initialization of events and less need to synchronize against the player
list.
Provided a hidden queue system for similar logic to pre-1.3 chat. When a
plugin is listening for the deprecated PlayerChatEvent, all chat will be
delayed to be mirror executed from the main thread. All developers are
encouraged to immediately update to the developmental Bukkit chat API as a
minimum transition for server stability.
Additionally, changes were required to bring thread-safety to the flow
logic. CopyOnWriteArrayList is the only viable means to produce thread
safety with minimal diff; using a sane pre-implemented collection would
require reworking of sections of NMS logic.
As a minor change, implemented expected functionality for
PlayerCommandPreProcessEvent. Setting the player should now change the
player executing the command.
This allows previous causes to be available during the event, as well as making the damage cause a valid one. If EntityDamageEvent is canceled, then it's not the last DamageCause.
Also prevents setting DamageCause involuntarily through construction.
TextWrapper used to try to ensure a message would wrap correctly on the
client by counting the width of the characters in pixels and wrapping
before hitting that limit. This was needed because the client would lose
color information when wrapping and could not handle long lines of text.
Now that both of these problems are solved in the client we can replace
TextWrapper with simple code to split the message into multiple packets on
newlines and ensure chat colors carry across to the new packet.
Make EntityLiving call AI logic every tick again.
Rework PathfinderGoalSelector logic.
Adds UnsafeList for use in places where we use ArrayList and know we won't
get index out of range errors. Added usage to World's tickEntities, Chunk's
entitySlices to speed up searching for entities, and to PathfinderGoalSelector
to speed up dealing with AI goals.
Reworked logic in PathfinderGoalSelector with help from fullwall. This code
no longer uses an extra ArrayList for setting up goals and only updates which
goals should be run every other time it is called.
Removed only calling PathfinderGoalSelector every other tick from EntityLiving
as we now only setup new goals every other tick. This ensures existing goals
run every tick to properly update mob movement.
- Allows drops in creative mode by adding items to the getDrops() list
- Contents of containers are not reported
- Contents of storage minecarts are not reported
Details:
- The attributes of custom inventory views are no longer ignored
- Enchanting or crafting inventories no longer ignore the passed inventory and open a new one
- Inventories associated with tile entities no longer raise a class cast exception if there was no associated tile entity
- InventoryOpenEvent and InventoryCloseEvent (if they already had some other inventory open) now fire in all cases
- If for any reason the inventory failed to open, the method now returns null instead of returned the previous inventory they had open (or the default inventory, if none)
Also adds CraftEventFactory.callEvent(Event), which returns the event called. Currently only used for n.m.s.BlockStationary's lava
BlockIgniteEvent calls.
