The Atomic implementation was only partially correct, and was very complex.
Use locks for sake of simplicity, following KISS principle.
Only remaining atomic operation use is time of day speed, because that
really is only read + written.
Also fixes a bug with m_time_conversion_skew only being decremented, never
incremented (Regresion from previous commit).
atomic.h changes:
* Add GenericAtomic<T> class for non-integral types like floats.
* Remove some last remainders from atomic.h of the volatile use.
It isn't possible to use atomic operations for floats, so don't use them there.
Having a lock is good out of other reasons too, because this way the float time
and the integer time both match, and can't get different values in a race,
e.g. when two setTimeofDay() get executed simultaneously.
Cleanup:
* Remove volatile keyword, it is of no use at all. [1]
* Remove the enable_if stuff. It had no use either.
The most likely explanation why the enable_if stuff was there is that it
was used as something like a STATIC_ASSERT to verify that sizeof(T) is not larger
than sizeof(void *). This check however is not just misplaced in a place where we
already use a lock, it isn't needed at all, as gcc will just generate a call to
to the runtime if it compiles for platforms that don't support atomic instructions.
The runtime will then most likely use locks.
Code style fixes:
* Prefix name of the mutex
* Line everything up nicely, where it makes things look nice
* Filling \ continuations with spaces is code style rule
Added operations on the atomic var:
* Compare and swap
* Swap
The second point of the cleanup also fixes the Android build of the next commit.
[1]: http://stackoverflow.com/q/2484980
On Windows Release x64 bit build this changes:
ProfilerGraph::put
1.68% -> 0.061%
ProfilerGraph::draw
12% -> 17.%
So yes, there is a tradeoff between saving profiling data
(executed always) and drawing the profiler graph (executed very rarely).
But usually you don't have the profiler graph open.
For several years now, the lua script lock has been completely broken.
This commit fixes the main issue (creation of a temporary rather than
scoped object), and fixes a subsequent deadlock issue caused by
nested script API calls by adding support for recursive mutexes.
When compiled with optimizations, the most recent versions of clang seem
to 'optimize' out a crucial "and %reg, 0x7FFFFFFF" instruction in noise2d(),
probably because it somehow assumed the variable n would never become greater
than that amount.
Indeed, signed integer underflow is undefined behavior in C and C++, so while
this optimization is "correct" in that sense, it breaks lots of existing code.
Solved by changing n to an unsigned type, making behavior well-defined.
Previously, the server called FATAL_ERROR when a Lua error occured.
This caused a (mostly useless) core dump.
The server now simply throws an exception, which is caught and printed before
exiting with a non-zero return value.
This also fixes a number of instances where errors were logged multiple times.
Increase default from 6 to 16 to help with mgv7 and mgfractal
Large-scale or alternative mapgens can result in a lowland spawn point not
being found, causing a spawn at (0, 0, 0) possibly buried underground
The max height is now settable to allow correct player spawn
in any mapgen or when using custom noise parameters
Use this macro to disallow copying of an object using the assignment
operator or copy constructor. This catches otherwise silent-but-deadly
mistakes such as "ServerMap map = env->getMap();" at compile time.
If so desired, it is still possible to copy a class, but it now requires
an explicit call to memcpy or std::copy.
-> Put access to time variables under the time lock.
-> Merge both time locks, there is no point to have two locks.
-> Fix the lock being released too early in Environment::setTimeOfDay
-> Add serverside getter so that you don't have to get
the environment if you only have the server
Previously, the invocation of Logger::addOutput(ILogOutput *out) led to
an out of bounds write of the m_outputs vector, resulting in the
m_silenced_levels array being modified.
Fortunately, the only caller of that method was android system logging,
and only since a few commits ago.
-> Get rid of Logger::logToSystem and use normal downstream output system for android instead
-> Give the downstream output system more information: enrich the log function of ILogOutput
with information and add ICombinedLogOutput for easier use.
-> Make Logger::getLevelLabel() static and public so that it can be used by downstream log output.
-> Add g_ and m_ prefixes where required
* Add key settings to setting table and ignore them later
This way they are added to the auto-generated minetest.conf.example
* Add flags type
* Add input validation for int, float and flags
* Break in-game graphic settings into multiple sections
* Parse settingtpes.txt in mods and games
* Improve description for a lot of settings
* Fix typos and wording in settingtypes.txt
* Convert language setting to an enum
Flushes the buffer of rollback actions that wait to get saved in two more situations:
1. Flushes in the destructor of the rollback. This makes the server not
forget the last < 500 rollback entries when it shuts down.
2. Flushes the rollback when /rollback_check is invoked. This is neccessary
as otherwise it leads to confusion if users want to test the rollback functionality
by placing a node and then executing the check on it, or if the actions were
very recent out of other reasons.
Everywhere where wstrgettext was used, its output was converted back
to utf8. As wstrgettext internally converts the return value
from utf8 to wstring, it has been a waste. Remove the function, and
use strgettext instead.
- setOpenedTrees(): this internal function was calling setSelected()
to update m_selected. Since setSelected() calls autoScroll(),
this caused the scrollbar to scroll back to the selected row
in some cases when that shouldn't be done.
For example, clicking the "+" to open a tree caused autoscroll.
Fix this by making setOpenedTrees() modify m_selected directly.
- setDynamicData(): set scrollbar position after calling
setSelected(), not before. This avoids setSelected()'s autoscroll
messing up the scrollbar position again.
- setSelected(): If an invisible row is selected, open all parents
of the selected row in order to make the selected row visible.
This fixes the issue where all the trees are closed again whenever
you return from the setting edit dialog to the settings tab.
- Fix some incompatibilities with obscure platforms (AIX and WinCE)
- Clean up Thread class interface
- Add m_ prefix to private member variables
- Simplify platform-dependent logic, reducing preprocessor
conditional clauses and improving readibility
- Add Thread class documentation
The invsize formspec element is outdated. Even though,
it is still supported, only a deprecation warning is shown,
introduced by commit [1]. The lua context passed to the
log_deprecated method added by commit [1] is NULL for the
invsize deprecation warning, as its run on the client and not
the server.
Commit [1] has removed checks for NULL inside the log_deprecated
method, resulting in a crash when a formspec with an invsize
element is parsed. This commit puts the check back.
Fixes#3260.
Referenced commits:
[1]: b5acec0a3c5701c53854ff7afdf4008863e6e8df "Add proper lua api deprecated handling"
[2]: 7b8d372947aae232ddf598155e972bb4dda157a "Use warningstream for deprecated field messages and refactor log_deprecated"
Commit
d2ca662569427d36642660314668e416bf68f3c8 "Enforce hiding nametag"
didn't fix the issue for "client" instances, where the nametag update
was received before the object was added to the scene. This resulted
in the grey shadow on the nametag that commit tried to fix.
Thanks to @neoascetic for pointing out that there still is a shadow.
- Add warning log level
- Change debug_log_level setting to enumeration string
- Map Irrlicht log events to MT log events
- Encapsulate log_* functions and global variables into a class, Logger
- Unify dstream with standard logging mechanism
- Unify core.debug() with standard core.log() script API
- SRP: print the address only once, not twice
- Legacy: previously the address was not printed at all
- Make both messages structurally the same, to facilitate log analyzers
Player now spawns on julia set due to julia offset
Add commented-out '#include profiler.h' for timetaker use
Use v3fs to reduce number of parameters
Tune tunnel width to match mgv7
Fixes a bug where the old list ring remained when a new formspec
was displayed over the old one. This created the list-ring of the new formspec
to be partly ignored.
Thanks to @VanessaE to report the bug, and @DonBatman to produce the code that
exposed it.
This might break some mods, but it is important for all uses of the param2 to
be documented.
This doesn't need a serialisation version or network protocol version change,
as old clients will still work on new servers, and it is bearable to have
new clients getting non rotated plants on old servers.
It returns the index used in mg->biomemap for a given biome name.
The biomemap is useless without this unless you re-register all existing biomes,
which could cause problems for anyone else trying to use biomemap.
With this, you can quickly create a lookup table of ids and names.
Get facedir by using lowest 5 bits of param2 and limiting to 23
More robust, frees up higher param2 bits for other uses
Change lookup table and table index to u8
1. Check for entity addition success in spawn_item implementation
2. Check for success in item_drop callback, so that the player
doesn't lose the item if they are outside bounds and try to drop it.
3. When existing player joins game, check that their position is inside
map bounds. If not, set their position to the return value of findSpawnPos().
4. Make findSpawnPos() respect the border
2 fixes a lua crash if a player drops an item outside map bounds.
3 fixes an assertion crash if a player leaves when being outside map bounds,
and then rejoins.
Because the count is serialized as u16, this would cause overflow.
If minetest later deserialized a mapblock with an incorrect
static object count, it would be unable to find the NameIdMapping
(which comes after the StaticObjectList) and abort with an error
such as "Invalid block data in database: unsupported NameIdMapping
version" (issue #2610).
Sidelen larger than 16 is essential for low density decorations
With sidelen > 16 chunksize may not be divisable by sidelen if
chunksize is changed, in this situation setting sidelen = chunksize
is desirable and should not create error messages.
Change a remaining assert(0) call to FATAL_ERROR(msg).
There was a regression since commit
ced6d20295a8263757d57c02a07ffcb66688a163 "For usages of assert() that are meant to persist in Release builds (when NDEBUG is defined), replace those usages with persistent alternatives"
where when an "uncatched" exception is thrown inside a "side thread",
the program doesn't abort anymore.
This led to the problem @netinetwalker experienced where the emergethread
got an unhandled exception for loading a mapblock while redis was loading,
(see #3196) and then jmped outside its loop to work down its queue.
This resulted in the server not doing any emerges anymore.
Thanks to @netinetwalker for spotting the error, proposing a fix, and testing it.
Error due to @est31's merging changes to PR #3202 to add more error reporting for invalid reply types, commit:
524a7656e3e5cd671b05c13e2ad69cb84bad0423 "redis: throw error if block request failed"
Now we branch out on the valid reply type "not found".
On world load the collision code can not see node boxes, since the
nodes have not been loaded. Thus it collided only at the next full
node. However, standing on a slab on world load leaded to sinking into
it until the world finished loading. Then one maybe fell further, if
the node below was not walkable.
Now, with this commit, when no node around the player has been loaded
it simply does not move the player.
Fixes#3196. Before, we didn't throw an error, and the engine thought the
block isn't occupied. But in fact it might be that redis is still loading,
and the block does exist in the database. The result was a cheesy map.
Why doing things simple ? Use pointer instead of strings to save players and remove them.
Saving players by name does a lookup to find pointer we already have ! Idem with removePlayer
Also remove unused removePlayer(peer_id), it's never called
-> Don't pass pointer to whole IGameDef to NodeMetadata constructors
and deserializers, but only to IItemDefManager, which is needed
-> Remove the unused content_mapnode_get_new_name() method
-> Fix style for MapBlock::deSerialize and MapBlock::deSerialize_pre22,
improving accuracy of error messages a bit
-> Fix style at other serialisation methods too
-> Improve accuracy of some comments
Use std::queue instead of std::set, we don't need such a heavy container.
Don't convert position to int to convert it back to float in the next function.
The legacy code added in commit
d879a539cd19ddd1ee34afec2512fb2238de2822 - "Add minetest.swap_node"
for sending the whole mapblock to older clients on the case of a node
modification with swap_node, had the problem that the block chosen to be
sent to the client was referenced with node coordinates and not with
block coordinates, resulting in getting the wrong block sent to the client.
Entity positions are serialized as F1000. Disallow placing
entities outside safe borders with the minetest.add_entity
call.
Note that this patch only enforces those boundaries for
placing entities, moving entities that move outside boundaries
aren't affected.
Thanks to @nanepiwo for pointing this out.
1. Do two renames:
* SER_FMT_CLIENT_VER_LOWEST -> SER_FMT_VER_LOWEST_WRITE
* SER_FMT_VER_LOWEST -> SER_FMT_VER_LOWEST_READ
Now the two define values are consistently named with the _WRITE defines
SER_FMT_VER_{HIGHEST,LOWEST}_WRITE, and to better point out what the two
serialisation versions actually are for.
2. wrap some lines in doc/worldformat.txt, and point out that the node
timers are serialized at a later point, as this can cause confusion about
what now happens (if one doesn't strictly read the if block's conditions).
3. some whitespace fixes in NodeTimerList::serialize, and one new comment.
Modders are now able to select the range of ore column height,
and the midpoint at which they 'grow' starting from.
This commit adds three new parameters for the 'sheet' ore type:
column_height_min, column_height_max, and column_midpoint_factor.
clust_size is now deprecated for this ore type.
Its obvious that "u16 command" is inside every packet, therefore this
commit removes all mentions of the command, if non-array like notation
is used. We already didn't add "u16 command" to new packets or removed
it at packet changes, so now we remove it from existing packets.
Use in-place locale directory if that exists, and
static one (RUN_IN_PLACE or CUSTOM_LOCALEDIR) doesn't exist.
Report to errorstream if neither static nor in-place locale
dirs exist, and report successfully found paths to infostreem.
Fixes two bugs:
-> Regression of commit [1] where if we use RUN_IN_PLACE=false,
but don't make install, locales aren't found. One might
think this is no regression, as its no bug, but all other
paths (mainmenu, etc.) are detected properly.
-> Regression of commit [1] where locales don't work on windows.
References:
[1]: Commit 645e2086734e3d2d1ec95f50faa39f0f24304761 "Use CUSTOM_LOCALEDIR if specified" by @ShadowNinja
Commit
e4bff8be94c0db4f94e63ad448d0eeb869ccdbbd - Clean up threading
by @ShadowNinja has broken the OSX build.
Including things inside a namespace isn't good.
Also fixes#3124.
Before, calling get_areas_in_area for an areastore with both
include_borders and include_data would result in a lua error,
if there was at least one area as result:
attempt to index a number value in function 'get_areas_in_area'
The commit
8f9af57314f71aae1cc77e13f9996e13015d776d "Add core.get_dir_list" by @ShadowNinja
has removed the implementation of the l_get_dirlist function and all its usages
from the l_mainmenu.cpp file, but hasn't removed it from the header file.
The reason why this hasn't been detected earlier is that C++ has this interesting
feature to still make it possible to create instances of classes whose never used
private methods are declared but not defined.
Preserves the rare unbroken protruding dungeons
Fix random range for first room roomplace
Fix checked volume for first room 'fits' bool
and check for 'untouchable' flag instead of 'inside'
Remove 'enable floating dungeons' setting
* Rename everything.
* Strip J prefix.
* Change UpperCamelCase functions to lowerCamelCase.
* Remove global (!) semaphore count mutex on OSX.
* Remove semaphore count getter (unused, unsafe, depended on internal
API functions on Windows, and used a hack on OSX).
* Add `Atomic<type>`.
* Make `Thread` handle thread names.
* Add support for C++11 multi-threading.
* Combine pthread and win32 sources.
* Remove `ThreadStarted` (unused, unneeded).
* Move some includes from the headers to the sources.
* Move all of `Event` into its header (allows inlining with no new includes).
* Make `Event` use `Semaphore` (except on Windows).
* Move some porting functions into `Thread`.
* Integrate logging with `Thread`.
* Add threading test.
This bypass had to be re-enabled as some users reported issues,
even after the iconv build fix.
While utf8_to_wide works well, wide_to_utf8 is quite broken
on android, for some reason, and some devices (unrelated from build
configuration).
Before, this lua code led to a crash:
local pcg = PcgRandom(42)
local value = pcg:next()
This was because if you called s32 PcgRandom::range(min, max) with the
minimum and maximum possible values for s32 integers (which the lua
binding code did), u32 PcgRandom::range(bound) got called with 0 as the
bound. The bound however is one above the maximum value, so 0 is a "special"
value to pass to this function. This commit fixes the lua crash by
assigning the RNG's full range to the bound 0, which is also fits to the
"maximum is bound - 1" principle, as (u32)-1 is the maximum value in the
u32 range.
When tunnels entirely above ground were avoided, the
missing pseudorandom calls changed the allowed caves.
Now, above ground tunnels are not placed while
still running all previous pseudorandom calls.
Multiplying by a factor of 1/1000.f (rather than dividing by 1000.f) directly
introduces an error of 1 ULP. With this patch, an exact comparison of a
floating point literal with the deserialized F1000 form representing it is now
guaranteed to be successful.
In addition, the maxmium and minimum safely representible floating point
numbers are now well-defined as constants.
Before, android_version.h got changed at every new commit. Now, we
only change it with new minetest releases. Analogous to how cmake
does it, we add an android_version_githash.h file that communicates
the git hash to C++ code.
Also, unify VERS_MAJOR, VERS_MINOR and VERS_PATCH variable
calculation inside the whole makefile.
-> Remove the SendableMediaAnnouncement struct
-> Forge the packet directly in the m_media loop, spare one loop and the construction of a vector
-> Use preincrement to spare iterator copies
Before, our libiconv build was a joke. We first called configure for our own build host system,
then called make, before we executed a Android.mk script we provided as patch. The first "native make"
always failed, and the LIBICONV_LIB file setting in our Makefile didn't match the built one,
resulting in an always-rebuild of iconv.
This commit cleans up this total mess, removes the double-build, and the Android.mk, and properly calls
./configure with the according target platform, and uses a built toolchain.
As we have to deal with the android bug "NDK: Support for prebuild libs with full sonames"
https://code.google.com/p/android/issues/detail?id=55868
as the 2013 patch
https://lists.gnu.org/archive/html/libtool-patches/2013-06/msg00002.html
by Google's David Turner wasn't inside the 2011 libtool, we pass -avoid-version to
libtool.
Thanks to the proper build, wide_to_utf8 works for android now, removing us of the need to disable it.
This was needed due to the TileDef serialisation changes. Originally it has been planned
to also add utf-8 based chat to 26, but chat changes aren't final yet, so they are done
in one change, after the release, and not two small ones, causing us having to be compliant
to three versions of the packet.
-> Fix server crash with protocol >=25 if a default password is set.
-> Remove some useless and possibly confusion causing code for the TOCLIENT_FIRST_SRP packet handler
-> Fix bug where MoveSomewhere from an infinite source would fill the destination inventory with copies of itself.
-> Fix bug where MoveSomewhere would needlessly call callbacks.
-> Remove trailing whitespaces
