When there are many chunkholders and regions, the cost of collecting
and checking tick thread for each one for every region save
becomes the biggest cost for the save call. To avoid this from
happening, collect the chunk holders from the current region's
owned sections.
This showed significant speedup locally when running the
"walk test" found in RegionizedServer locally (>90% of time
was spent on the holder iteration/checking).
While for merging the synchronisation occured, it did not synchronise
for splitting. This resolves a possible CME that may occur while
splitting regions.
Player spawn position changing caused any player to log in while
on a boat to trip a thread check. To resolve this, simply reposition
any mounted entities if they are more than 5 blocks away from the player.
In general, this may happen to other entities that are loaded from
chunks as well. In these cases, we can delete the entity if it itself
is not saved in the correct chunk, and we can reposition the mounted
entities if they are not in the correct chunk. For tile entities,
we can simply remove them if they are not in the chunk.
These changes broadly should make loading player/chunk data more
resiliant to bad logic run by plugins.
Also, the Folia test revealed that the chunk generate rate limiter
also affected chunk loading, which was not intended and has been
resolved by exempting already FULL status chunks from the limit.
It becomes invalid switching dimensions or by moving far. If
it used after teleporting, then it may also trip thread checks.
Fixes https://github.com/PaperMC/Folia/issues/94
Starlight does not use the sky light sources, and there
appear to be issues with TE access on Folia when initializing
them. So, we can just delete it entirely.
Fix two regionizer issues:
In ThreadedRegionizer#addChunk, fix the incorrect handling
of merging two regions where one of the regions had
pending merges. If the first region had pending merges,
and the second was marked as "ready" then the merge would
cause a "ready" region to have pending merges. The fix is
to simply downgrade the "ready" region to "transient,"
as was previously done if the merge was delayed in the
case where the first region was "ticking."
Additionally, prevent the creation of empty regions
by checking if any new sections were created. This would
happen when a section existed, but had no marked chunks
in it AND all of the sections neighbours existed. In these
cases, no region needs to be created as no sections were
created.
The last save was based on region tick, but it was not adjusted
on player region change or region merge. To resolve this,
I have adjusted the last save to be based on time so that it does
not need adjustments on region change or region merge.
Additionally, fix the max per tick handling.
Mobs would use the evenness of server tick count plus id
to determine whether they eoilf tick only their running
goals or to tick the goal selector to find additional
goals. If the server had an even number of regions,
then every 50ms the server tick field would be incremented
by an even number and as a result would not change
the evenness of the mob goal check. This could put
some mobs in a state where they only ticked their
running goals, which would result in them
freezing.
Fixes https://github.com/PaperMC/Folia/issues/42
First, when a section update is stolen, the thread that acquires
the stolen update should remove the update from the update queue
before returning to mark it as completed and allow other threads
waiting on the update to continue. This fixes a deadlock issue
with section updates.
Fix incorrect decrease queue resize. Previously, it attempted
to resize the _increase_ queue, which is the wrong queue.
Use ALL_DIRECTIONS_BITSET for every decrease queue direction bitset
as decrease propagation cancellation due to neighbour values exceeding
the target decrease value cause some neighbour directions to not
be checked, which causes the final update grid to be incorrect.
We must attempt to synchronise when the returned queue is null
so that we can get a correct queue result or return false due to
the reference counter being released, or even to throw an exception
when the queue is null but the reference counter is not released.
I noticed during my stress testing that the total size of the
light list was far too large, which indicates many duplicates.
For me, this caused many GC problems which made stress testing
harder.
It turns out, it was possible for the light list recalculation
logic to occur _and_ the addition of the light list data from
the NBT data. Since there is no logic to de-duplicate this list,
every chunk load would re-add all light sources into the light
list and the light list would grow uncontrollably.
Since the recalculation logic would often run, I have
decided to solve this by discarding the data on disk and always
just calculating the list from the chunk data alone. Additionally,
I have applied an optimization from Vanilla 1.20 to avoid
searching sections without light sources by first checking the
palette for possible block sources.
Now my stress tests do not have issues with GC at all.
In 1.18, every chunk section is initialised to a non-null value
and recalcBlockCounts() is invoked for each section.
However, in a standard world, most sections are empty. In such cases,
recalcBlockCounts() would iterate over ever position - even though
the block data would all be air. To avoid this, we skip
searching the section unless the palette indicates there _could_ be
a non-air block state or non-empty fluid state.
Chunk loading initially showed that recalcBlockCounts() over
sections with a ZeroBitStorage data to to take ~20% of the process,
now it takes <1%.
Since the chunk load task was not scheduled, the entity/poi load
task fields will not be set, but the task complete counter
will not be adjusted. Thus, the chunk load task will not complete.
To resolve this, detect when the entity/poi tasks were not scheduled
and decrement the task complete counter in such cases.
It must be marked as completed during that lock hold since the
waiters field is set to null. Thus, any other thread attempting
a cancellation will fail to remove from waiters. Also, any
other thread attempting to cancel may set the completed field
to true which would cause accept() to fail as well.
Completion was always designed to happen while holding the
scheduling lock to prevent these race conditions. The code
was originally set up to complete while not holding the
scheduling lock to avoid invoking callbacks while holding the
lock, however the access to the completion field was not
considered.
Resolve this by marking the callback as completed during the
lock, but invoking the accept() function after releasing
the lock. This will prevent any cancellation attempts to be
blocked, and allow the current thread to complete the callback
without any issues.
Since multiple regions can exist, there are concurrent accesses
in this class. To prevent deadlock, the monitor is not held
when recalculating permissions, as Permissable holds its own
lock.
This fixes CMEs originating from this class.
The concurrent access occurs on the Netty IO threads when
serializing packets. Thus, it seems it was an oversight of
the implementator of this function as there are typically
more than one Netty IO thread.
Fixes https://github.com/PaperMC/Folia/issues/11
A significant overhead in Folia comes from the chunk system's
locks, the ticket lock and the scheduling lock. The public
test server, which had ~330 players, had signficant performance
problems with these locks: ~80% of the time spent ticking
was _waiting_ for the locks to free. Given that it used
around 15 cores total at peak, this is a complete and utter loss
of potential.
To address this issue, I have replaced the ticket lock and scheduling
lock with two ReentrantAreaLocks. The ReentrantAreaLock takes a
shift, which is used internally to group positions into sections.
This grouping is neccessary, as the possible radius of area that
needs to be acquired for any given lock usage is up to 64. As such,
the shift is critical to reduce the number of areas required to lock
for any lock operation. Currently, it is set to a shift of 6, which
is identical to the ticket level propagation shift (and, it must be
at least the ticket level propagation shift AND the region shift).
The chunk system locking changes required a complete rewrite of the
chunk system tick, chunk system unload, and chunk system ticket level
propagation - as all of the previous logic only works with a single
global lock.
This does introduce two other section shifts: the lock shift, and the
ticket shift. The lock shift is simply what shift the area locks use,
and the ticket shift represents the size of the ticket sections.
Currently, these values are just set to the region shift for simplicity.
However, they are not arbitrary: the lock shift must be at least the size
of the ticket shift and must be at least the size of the region shift.
The ticket shift must also be >= the ceil(log2(max ticket level source)).
The chunk system's ticket propagator is now global state, instead of
region state. This cleans up the logic for ticket levels significantly,
and removes usage of the region lock in this area, but it also means
that the addition of a ticket no longer creates a region. To alleviate
the side effects of this change, the global tick thread now processes
ticket level updates for each world every tick to guarantee eventual
ticket level processing. The chunk system also provides a hook to
process ticket level changes in a given _section_, so that the
region queue can guarantee that after adding its reference counter
that the region section is created/exists/wont be destroyed.
The ticket propagator operates by updating the sources in a single ticket
section, and propagating the updates to its 1 radius neighbours. This
allows the ticket updates to occur in parallel or selectively (see above).
Currently, the process ticket level update function operates by
polling from a concurrent queue of sections to update and simply
invoking the single section update logic. This allows the function
to operate completely in parallel, provided the queue is ordered right.
Additionally, this limits the area used in the ticket/scheduling lock
when processing updates, which should massively increase parallelism compared
to before.
The chunk system ticket addition for expirable ticket types has been modified
to no longer track exact tick deadlines, as this relies on what region the
ticket is in. Instead, the chunk system tracks a map of
lock section -> (chunk coordinate -> expire ticket count) and every ticket
has been changed to have a removeDelay count that is decremented each tick.
Each region searches its own sections to find tickets to try to expire.
Chunk system unloading has been modified to track unloads by lock section.
The ordering is determined by which section a chunk resides in.
The unload process now removes from unload sections and processes
the full unload stages (1, 2, 3) before moving to the next section, if possible.
This allows the unload logic to only hold one lock section at a time for
each lock, which is a massive parallelism increase.
In stress testing, these changes lowered the locking overhead to only 5%
from ~70%, which completely fix the original problem as described.
Instead, we can just check the loaded chunk's block position for
the lodestone block, as that is at least safe enough for the light
engine compared to the POI access. This should make it safe for
off-region access.
Fixes https://github.com/PaperMC/Folia/issues/60
In general, worldstate read/write is unacceptable during
data deserialization and is racey even in Vanilla. But in Folia,
some accesses may throw and as such we need to fix this directly.
Fixes https://github.com/PaperMC/Folia/issues/57
The returned TE may be in the world, in which case it is unsafe
for the current thread to modify or access its contents.
Fixes https://github.com/PaperMC/Folia/issues/52
This is to prevent block physics from tripping thread checks by
far exceeding the bounds of the current region. While this does
add explicit block update suppression techniques, it's better
than the server crashing.
Perform thread checks on the chunk send and warn when the
world is mismatched. I suspect that the world mismatches for
an unknown reason, but need to confirm it to chase it down.
Change overview:
- Rework limiting
- Remove mid tick updates
- Introduce consistency checks
The old limiting logic used an intervalled counter, but
did not account for possible slight changes in mid tick
invoke rate as it relied heavily on mid-tick logic. Due to
the removal of mid tick updates, it is now important that
the logic functions correctly no matter what rate it is invoked
at. The new logic directly tracks the last update time and
allocates an amount based proportional on the rate targetted,
which makes the logic call rate independent.
The removal of mid tick updates is done to eliminate recursive
call risk, and to additionally reduce the lock pressure on the
chunk system by grouping chunk loads onto one part of the tick
rather than spreading it out. The limiting rework should ensure
that this does not negatively affect rates, but it will decrease
the perceived smoothness of chunk generation/loading at low rates.
Introduce more consistency checks such as correct tick thread
and ticking-after-removal checks. Also, perform checks during the player
chunk loader tick to avoid updating potentially removed
players during the tick.
The checks are primarily made to try to hunt down a bug that
is causing the player chunk loader to double send a chunk
to a player.
Due to TPS catchup being removed, a lost tick will always
affect block breaking.
Additionally, constant low TPS would affect block breaking
in an intrusive manner to the user when there is no need
for that to occur.
The expected behavior is that the entity is only dismounted
_if_ the teleport takes place, not regardless of whether
the teleport takes place.
To adhere to the expected behavior, we need to create a new teleport
flag so that the NMS teleportAsync can perform the dismount.
Required to add some basic hacks to expose a regionstats object
in the TickRegions data class, but this ensures that the retrieval
is thread-safe and could possibly support other data exposure
for async reads.
Additionally, since DecimalFormat is not thread-safe we need
to use ThreadLocals to instantiate them. Change the format
as well to use commas to separate groups of digits when
formatting large numbers so they are easier to read.
For example, 1000 becomes 1,000.
Vanilla inserted the alive check so that dead players
could see the chunks around them, but in Folia we do not
remove dead players so chunks will still load for them.
This prevents the client from having chunks in memory it
should not.
The inventory was being cleared by removing the entity
from the world, but Folia changed the remove to be
before the entity copy which caused the new entity
to have a cleared inventory. Fix this by creating
a post-dimension copy callback that will clear
the inventory of the old entity.
Fixes https://github.com/PaperMC/Folia/issues/29
