- Correctly handle block state limit checks.
- Don't quash exceptions thrown in a region operation by having a "max-fails" check that can only be hit once....
Fixes#1127
- ImmutableBaseBlock is no longer needed as NBT was made immutable previously
- BlanketBaseBlock represents to masks that the block has NBT, but does not need to match a specific state
- Allow FuzzyBlockState to create BaseBlocks without NBT data (fixed tab completion issues)
- Required new argument parser, which should ultimately be made into a singe blockstate parser for best results, and proper tab-completion
- The layer brush itself is broken because of the changes to internal block retrieval from legacy
**Add a null-check to CharBlocks FULL section layer-retrieval.**
- It is possible to trim CharBlocks whilst it is attempting to read data due to the batching of chunks to help reduce memory
- This is done when the number of chunks sitting loaded in memory with having been "submitted" to the queue for writing to disk becomes high
- Seconday operations such as heightmap processing and lighting will quickly load chunks, meaning many chunks are submitted early
- This leads to much higher chances of the chunk being submitted and subsequently trimmed given heightmap and light processing is done layer-by-layer over many chunks, rather than chunk-by-chunk - thus leading to NPEs.
- By adding synchronisation to and around only the specific sections when loading/updating, and not blocking the whole chunk, many access can still be thread-safe without causing deadlocks
- This allows removal of lots of the needless and very-slowing synchronisation on get**Block** methods
**Remove much of the synchronisation from ChunkHolder**
- We shouldn't be synchronising with call() and safety should be added elsewhere. (plus it's making edits very very slow when queue target size is hit)
- Also remove much of synchronisation because we've added the null-check and section-specific synchronisation to CharBlocks
**Some QOL/thread-safe data access changes**
- Replaces the Array#clone seen in the get blocks classes with System#arraycopy as deep cloning is not required, and is also slower than arraycopy
- Add System#arraycopy when accessing chunk section data via history to ensure it is not altered whilst being written
- Renaming EMPTY to empty means it is not implied to be a static variable
Fixes https://github.com/IntellectualSites/FastAsyncWorldEdit/issues/1028
Fixes https://github.com/IntellectualSites/FastAsyncWorldEdit/issues/1025
Fixes https://github.com/IntellectualSites/FastAsyncWorldEdit/issues/1089
Fixes https://github.com/IntellectualSites/FastAsyncWorldEdit/issues/1091
Fixes https://github.com/IntellectualSites/FastAsyncWorldEdit/issues/1097
similar to ChunkHolder, it represents an internal chunk for which operations should not be accepted by multiple threads at once.
Co-authored-by: NotMyFault <[email protected]>
This is basically the main "chunk" class for internal FAWE. Chunk operations should (and are) almost always single-threaded operations, however, under certain circumstances it is possible for the chunk to be "called" (flushed: written to the world and sent to the player) from a separate thread. This would specifically occur from SingleThreadQueueExtent when there are a lot of chunks being loaded in memory by FAWE (where the chunk would then be submitted to a multi-threaded queue). It would therefore be possible for a thread accessing the chunk to attempt to access it in the middle of the call, which can lead to a number of issues, and it is my opinion that the most frequent of these is the NPE seen during lighting operations, where new chunks can be accessed/loaded very quickly, increasing the likelihood for the aforementioned synchronisation issue to occur.
Co-authored-by: Matt <[email protected]>
Since there's no new NMS revision (we're still on R3...?!) but there were changes, it would be more difficult to maintain both 1.16.4 and 1.16.5. Therefore, let's just stop supporting 1.16.4 as 1.16.5 is a minor update regardless.
Fixes#859
- Add a "loadPrivately" method to be used when GetChunks are called to avoid synchronocity issues with super classes being used on different threads
- Synchronise the call method so we're not attempting to call whilst also loading/updating
- Properly ensures we don't try to submit chunks that already have a monitor (prevent the FAWE-freezing)
- We can't detect if "no" threads hold the chunk's monitor, but equally that would also be kinda very bad practice.
- We don't want to flush if we're setting from the main thread, as that's going to be another plugin doing it
- It's better to have a concurrent error thrown than use a concurrent set, so concurrency issues can actually be fixed rather than handled unsafely
- Only send chunk packets up to once a tick (it really doesn't use any memory to cache IntPairs).
- With enough chunks waiting to write to history, it's possible to fill the secondary fork pool up with locked threads, and thus prevent any threads from actually writing their history.
- Have both sponge and fast r/w available but default to "fast"
- Correctly "offset" entities in the schematic, and add a legacy mode for loading old FAWE schematics with entities required.
- Lazily reading means it's read in order of appearance in the inputstream so we need to read schematic version first (skip past everything) and then reset the stream. Fixes#740
- Add an FAWEVersion to the metadata
- Correctly actually return a BlockArrayClipboard when required. Fixes#454
dordsor21
