RemoteServer and PluginManager side would need complete redesign to be
data race free and concurrent. But as that would be unlikely to be
required from DFHack I decided simpler solution that is fixing data
ownership to a thread and all ServerConnection share a single lock which
allows access to PluginManager and Core.
The old CoreSuspender requires processing from Core::Update to allow
commands execute. But that causes issues if Core::Shutdown wants
quarentee cleanup order with std:🧵:join. Fixing shutdown ordering
adds too many branches to already fairly complex code.
I decided to try to refactor CoreSuspender to use simpler locking
locking using a std::recusive_muted as primary synchronization
primitive.
To help control when Core::Update unlocks the primary mutex there is
std::contition_variable_any and std::atomic<size_t> queue lenght
counter.
The last state variable is std::atomic<std:🧵:id> that is used to
keep track of owner thread for Core::IsSuspended query.
This should be merged only just after a release to make sure that it
gets maximum testing in develop branch before next release.
Fixes#1066
There is a minor chance that console or init thread would access already
freed memory when core is shutting down and cleaning up state. To avoid
any danger of having random bugs caused by the potential data race I
decided to make sure the shutdown code waits for the thread to exit
first.
Windows change is completely untested. It is purely based on msdn
documentation.
I noticed that tthread is missing some c++11 features that make thread
handling code a bit easier. To be able to use those features I decided
to convert Core.cpp to use equivalent standard classes.
This patch has no functional changes.
command-prompt viewscreen may affect command execution if they are
looking for UI state. To make commands execute simillary to hotkeys or
console commands the viewscreen needs to removed from the top position.
Fixes#1194
gcc supports type checks for printf parameters which can catch some hard
to reproduce bugs. Possible bugs happen when the parameter value is
intepreted differently to the variable value.
Example warnings follow
../library/LuaWrapper.cpp:1011:86: warning: format ‘%llu’ expects argument
of type ‘long long unsigned int’, but argument 3 has type ‘uint64_t
{aka long unsigned int}’ [-Wformat=]
../plugins/follow.cpp:159:35: warning: format not a string literal and no
format arguments [-Wformat-security]
* Several fatal errors that occurred during core initialization didn't
stop initialization or set 'errorstate' properly, which caused
update hooks and other code to crash later. This has been fixed and
should address crashes like the one mentioned in #470.
* Errors when loading dfhack.lua now cause Lua::Open() to fail, which
triggers a fatal error in Core::Init()
* Failure to initialize the console no longer results in a call to
fatal() (since it didn't actually stop initialization previously)
- This context requires core suspend lock and asserts it in a few places.
- Special 'event' objects are introduced. They can be invoked as
functions, in which case they iterate all their fields and call
them as functions. Errors are printed and consumed.
- When a plugin is opened by the core context, events registered in
a special array are linked to it. The system is organized so as to
avoid even trying to pass the event to lua if the module isn't loaded.
* Paint, filter, and brush state is now saved between calls.
* Added 'all' paint option to set material, shape, special, and variant at
the same time.
* Added tiletypes-here (like liquids here, except is uses the saved brush
settings)
* Added tiletypes-here-point (like liquids here, always only the tile under
the cursor)
* Added tiletypes-command: runs tiletypes commands seperated by ';' tokens
(affects saved state)
* Make the internal workings match liquids a bit more
* Give brush objects a descriptor string
* Make Core::cheap_tokenise available
- To ensure reload safety functions have to be wrapped. Every call
checks the loaded state and locks a mutex in Plugin. If the plugin
is unloaded, calling its functions throws a lua error. Therefore,
plugins may not create closures or export yieldable functions.
- The set of function argument and return types supported by
LuaWrapper is severely limited when compared to being compiled
inside the main library.
Currently supported types: numbers, bool, std::string, df::foo,
df::foo*, std::vector<bool>, std::vector<df::foo*>.
- To facilitate postponing initialization until after all plugins
have been loaded, the core sends a SC_CORE_INITIALIZED event.
- As an example, the burrows plugin now exports its functions.
The idea is that if you have to execute many RPC calls, it is
faster to suspend once. The service class takes care to auto-resume
in the destructor in case the client just disappears.
This is an incompatible change to the plugin ABI.
The Console is not thread-safe unless used indirectly
via color_ostream_proxy, so everything should use their
per-thread stream.