Before this patch, destroying an obsidian Nether portal frame destroyed
Nether portal nodes recursively. In Minetest binaries compiled with Lua
5.1 (i.e. without LuaJIT) this would cause a stack overflow when a huge
portal (23×23) was destroyed, crashing the server.
This patch implements Nether portal destruction using node timers. When
a portal node's timer triggers, it starts the timers of adjacent portal
nodes with the same orientation and no active timer and deletes itself.
Attempts to solve this problem using minetest.after() seemed promising,
but rubenwardy pointed out that anything relying on minetest.after() is
bound to fail if a server shuts down while portal nodes are destroyed.
Before this patch, when placing a fire above a node that would turn it
into eternal fire (e.g. Netherrack or Magma) the spawn_fire() function
would call itself infinitely via the on_construct() handler of eternal
fire – because the latter called spawn_fire() itself.
On an x86 machine, this caused a memory leak, hanging Minetest. On an
x86_64 machine though, Minetest crashed immediately, showing an error
message about a stack overflow.
Even the mitigated timers seem to have lead to slow
memory leaks. Once Minetest has used up all the RAM,
it will free some, then quickly use memory up again,
then repeat it ad nauseum, requiring 100% CPU. On a
PC with 2GB of RAM this could be reliably triggered
by having a fire burn a forest for 20 to 30 minutes.
This patch removes fire node timers completely and
instead extinguishes fire using an ABM.
This patch initializes the random number generator used in mcl_fire with
the current Unix timestamp. It also corrects two biases in fire spread
that were caused by nodes being iterated over in a predictable way.
This patch adds a new test structure to the “/spawnstruct” command which
can be spawned with this command: /spawnstruct test_structure_fireproof
The structure can be used to verify that eternal fire can be spawned by
the structure placement code. It can also be used to debug fire spread,
as according to Minecraft fire spread rules, it should be fireproof.
See <https://minecraft.fandom.com/wiki/Fire#Spread> for those rules.
In Minetest 5.4.1, calling get_player_control() on a mob returned the
empty string. Minetest commit 5eb45e1ea03c6104f007efec6dd9c351f310193d
changed this, so now calling get_player_control() on a mob returns nil.
As mcl_boats defines boats that can have a player or a mob as a driver,
code like the following crashes with a changed get_player_control() API:
local ctrl = driver:get_player_control()
if ctrl.sneak then
detach_object(driver, true)
end
Furthermore, once a world has crashed, joining it near a mob that is the
driver of a boat with such control code immediately crashes again.
When I reported this bug to Minetest, several Minetest core developers
stated that they disliked the old API and proposed other return values
for calling a mob's get_player_control() function – all different from
the empty string. Since I have some doubts that this bug will be fixed
in Minetest 5.5.0, boat code must take into account a nil return value.
Minetest issue: https://github.com/minetest/minetest/issues/11989
Before this patch, wearing a pumpkin would make body armor render as a
corrupted item held in the player's left hand instead of on their body.
Some parts of their armor would also get rendered on the player's feet.
(cherry picked from commit 12192d1a8d)
Before this patch it was possible for any user to to crash Minetest in
creative mode. This was possible because queries in the search field
were interpreted as search patterns for string.find().
A search for a single square bracket would reliably crash the server.
Also, a search for 6000 times the string “a?” would hang the server.
The solution to both bugs is to not interpret the query as a pattern.
Before this patch, growing a gourd (e.g. melon, pumpkin) would always
convert a node west of the node below the stem to dirt if belonged to
the group “dirtifies_below_solid”. This happened because of a loop in
which the variables floorpos and floor were re-used without setting a
new value … therefore, both floorpos and floor were always containing
the last values set in a previous loop instead of the correct values.
This patch fixes the problem by setting both variables in both loops.
The comparison and setting logic in the previous patch that set player
bone positions and properties conditionally incorrectly did not update
some values (like player eye level position) when they changed. This
patch fixes it and adds asserts to ensure the code works as intended.
Before this patch, Mineclonia set (and therefore sent) player bone
positions and player properties in every globalstep. This results in
about 30 TOCLIENT_ACTIVE_OBJECT_MESSAGE per second per player. This
patch adds conditional functions to set bone positions and properties.
The functions set values only when they have changed, reducing traffic.
Commit 5252952555 used string.gsub() to
strip newlines from tools in death messages. The second return value of
string.gsub() (the number of substitutions) was erroneously returned too
by get_tool_name(). This bug caused a crash whenever a skeleton killed a
player using its bow.
Players naming their tools with newlines can mess up chat with death
messages. This commit strips out newlines (\r and \n ) in the tool
name for death messages.
