Categories
Game Development Software Architecture Software Development

The Scourge of Global Mutable State

I’m not a perfect programmer, but I like to think that I’m getting better every day. I’ll admit that one of the things that I’ve had the most difficult time adapting to was what I perceived as other programmers being somewhat pedantic about avoiding what they called “global state” or sometimes more specifically “global mutable state.”

In the past, I held onto some conviction that while this is something to be avoided where possible, there are actually cases where it is appropriate. Recently I’ve decided to not only better conform to this well-known best-practice, but also to understand why it is that this is an issue, and what can be done to achieve what I, in the past, mistakenly perceived as the reason that global mutable state made sense, without violating this fundamental rule.

To learn something, you must learn the rules, but to become an expert, you must learn when to break these same rules. Programming is no different. However, over several years of programming I’ve come to find that the “when” of “when to break the rules” is an exception, not the rule. If you’re breaking the rule more often than not, then you’re probably misunderstanding something or being lazy. It pays to heed the words of the experts superior to you in experience even when you don’t fully understand why, because eventually, with more experience, you will understand why, and if you failed to heed their advice, you’ll find yourself chastising your past self for it.

Let me start by explaining why it is that I’ve so often resisted this best practice and gone ahead with global mutable state anyway. When writing larger code bases, which most projects inevitably aspire to be, you will find that some of your code will be, as a matter of necessity, used by a large proportion of the code base. Using dependency injection as your only distribution technique, as an example, might drastically increase the verbosity of your code. Who wants to pass an instance a “Log” class to every single instance of a “GameState” object? I’ll admit that it’s not really THAT much extra work, but it has a distinct feeling of not being “DRY,” since I’m essentially repeating myself on every object instantiation. There are some subsystems which, by nature, the entire system will need to access, and would be better off being organized.

A pet peeve not only of myself, but of many people, is that some people seem to think that using the singleton pattern is a solution to this problem. Singletons are global state, and moreover, it appears to me that by now, the singleton is obsolete. Any problem solvable with a singleton can be solved with a static class with mutable private fields, and frankly these static classes don’t lie about their true nature. But both are global state, and both ought to be avoided anyway. But if you’re going to ignore the advice anyway, you might as well just use the static class instead, so users can recognize the warning signs of nondeterministic behavior before it’s too late.

I’ve learned, however, that all of my arguments came from a lack of understanding of the actual problem, and the actual solutions.

  • Global mutable state generally excludes communication channels that allow decoupled non-global objects to communicate (such as Observers or Service Locators). This is a significant solution to this problem.
  • Global mutable state implies that the global items in question contain actual values that change between subsequent calls. Global classes are still perfectly valid for resolving the problem of subsystems that need to be accessed from across the system as whole, so long as they incorporate the solutions from the previous bullet.

I’ll document one solution that I solved with my TimberWolf game engine.

In the past, I’ve had a static class named Log, which was declared as follows:

#ifndef H_CLASS_LOG
#define H_CLASS_LOG

#include <iostream>
#include <string>
#include <iomanip>
#include <sstream>
#include <fstream>
#include <ctime>
#include <mutex>
#include <exception>
#include <cstdlib>
#include <atomic>
#include "../../enum/LogLevel/LogLevel.hpp"

class Log {

public:

    Log () = delete; // static only
    ~Log () = delete;

    Log (Log&&) = delete;
    Log& operator = (Log&&) = delete;

    Log (const Log&) = delete;
    Log& operator = (const Log&) = delete;

    static bool cliOutputEnabled ();
    static void enableCliOutput ();
    static void disableCliOutput ();

    static LogLevel getCliFilterLevel ();
    static void setCliFilterLevel (LogLevel);

    static bool fileOpen ();
    static bool openFile (const std::string&);
    static void closeFile ();

    static bool fileOutputEnabled ();
    static void enableFileOutput ();
    static void disableFileOutput ();

    static LogLevel getFileFilterLevel ();
    static void setFileFilterLevel (LogLevel);

    template<typename ...T>
    static void log (LogLevel messageType, T&&... message) {

        std::string out;

        if (
            (m_cliOutputEnabled || m_fileOutputEnabled) &&
            (
                messageType == LogLevel::UNDEFINED ||
                messageType >= m_cliFilterLevel ||
                messageType >= m_fileFilterLevel
            )
        ) {
            out = Log::formatMessage(messageType, Log::concatMessage(std::forward<T>(message)...));
        } else {
            return;
        }

        if (m_cliOutputEnabled) {
            if (messageType == LogLevel::UNDEFINED || messageType >= m_cliFilterLevel) {

                if (messageType == LogLevel::ERROR) {
                    std::unique_lock<std::mutex> lock_stderr(Log::mutex_stderr);
                    std::cerr << out << std::endl;
                } else {
                    std::unique_lock<std::mutex> lock_stdout(Log::mutex_stdout);
                    std::cout << out << std::endl;
                }

            }
        }

        if (m_fileOutputEnabled && fileOpen()) {
            if (messageType == LogLevel::UNDEFINED || messageType >= m_fileFilterLevel) {

                std::unique_lock<std::mutex> lock_file(Log::mutex_file);
                m_file << out << std::endl;

            }
        }

    }

    template<typename ...T>
    static void verbose (T&&... message) {
        Log::log(LogLevel::VERBOSE, std::forward<T>(message)...);
    }

    template<typename ...T>
    static void notice (T&&... message) {
        Log::log(LogLevel::NOTICE, std::forward<T>(message)...);
    }

    template<typename ...T>
    static void warning (T&&... message) {
        Log::log(LogLevel::WARNING, std::forward<T>(message)...);
    }

    template<typename ...T>
    static void error (T&&... message) {
        Log::log(LogLevel::ERROR, std::forward<T>(message)...);
    }

    static void bindUnhandledException ();
    static void unhandledException ();

private:

    template<typename ...T>
    static std::string concatMessage (T&&... message) {

        std::ostringstream oss;
        (oss << ... << std::forward<T>(message));
        return oss.str();

    }

    static std::string formatMessage (LogLevel, const std::string&);

    static std::atomic<bool> m_cliOutputEnabled;
    static std::atomic<bool> m_fileOutputEnabled;

    static std::atomic<LogLevel> m_cliFilterLevel;
    static std::atomic<LogLevel> m_fileFilterLevel;

    static std::ofstream m_file;

    static std::mutex mutex_stdout;
    static std::mutex mutex_stderr;
    static std::mutex mutex_file;

};

#endif

As you can see, everything is static, and the class cannot be instantiated. This makes sense, because the Log class is something that needs to be accessed from all over the code base, and we generally want all log messages to funnel through a single location so none of them accidentally go astray before they are sent or stored. This was all fine and dandy when it was originally written, because it only had one purpose then, and that was to write error messages to std::cout or std::cerr. It interacted with a global I/O, but it didn’t store any kind of state.

After a while, though, I decided that a few features would be nice to have in the logging subsystem. First of all, I wanted to be able to filter log messages by their LogLevel enum. This was simple enough. Add a static field to store the minimum log level, and then compare the log level of each message before writing it to I/O. But now the log class contains the global state for the current log level. Anybody could change that, and it might make it hard to debug stuff because its behavior becomes nondeterministic. The behavior of the Log class depends on the order of modifications. If some code somewhere deep in the code base calls Log::setFilterLevel(LogLevel::ERROR) because the developer was getting inundated with verbose calls and wanted to temporarily silence them, and then forgot to remove it, all LogLevel::WARNING messages sent after that piece of code gets invoked will fail to display. The developer might mistakenly think that no errors are occurring. And there’s no way to receive those messages elsewhere unless the filter level is dropped again.

This hasn’t actually been a problem for me yet, but I was slightly aware of the possibility, but I hadn’t yet done anything to resolve it. I added additional features, such as allowing Log to write to a log file, and allowing a separate filter level for the log file. I noticed some additional potential problems after adding these features, however. First of all, I can only write to a single log file. What if I decide later on that I want to write only warnings and errors to one file, and all messages to another file. Or what if I want a console window in my game somewhere to display the error messages, without requiring the end-user to look at their terminal or open up a log file?

The Log subsystem is a problem that can be solved by the observer pattern. I can leave the API for sending log messages alone, because it implements everything the observables need (but I am going to add a context string that can be used for additional filtering, in case we want the log messages pertaining only to the sound system to go somewhere, for example). There will need to be a separate API for RECEIVING the log messages, however. One of these will be used to receive messages and route them to the command line. Another one will be used to put messages into a file. But more can be created and removed as needed.

If those changes are implemented, The Log class, which is the global portion of the subsystem, will no longer contain any global state at all (the vectors of observers are not technically “state” in this context, as they are not going to have any effect on one another, the log subsystem, or the game as a whole). Any code in the code base can send messages to it, but the output, which is the real tangible effect of the log subsystem, will only be controlled by non-global objects. Code far across the code base will not be able to create nondeterministic behavior in the logs of an unrelated system.

This is, for all intents and purposes, the best of both worlds. I can expose a global API without needing to define behavior globally. I can create a log subsystem that doesn’t require dependency injection, but also doesn’t expose any state to the global scope, while still supporting stateful behavior at one end of the pipeline.

With that explanation out of the way now, allow me to show you the finished solution. Note that between the old version and this current one, the namespace “tw” was added to everything.

Here’s the new Log class declaration:

#ifndef H_CLASS_LOG
#define H_CLASS_LOG

#include <iostream>
#include <string>
#include <vector>
#include <memory>
#include <sstream>
#include <mutex>
#include <functional>
#include "../LogObserver/LogObserver.hpp"
#include "../../enum/LogLevel/LogLevel.hpp"

namespace tw {
class Log {

public:

    Log () = delete; // static only
    ~Log () = delete;

    Log (Log&&) = delete;
    Log& operator = (Log&&) = delete;

    Log (const Log&) = delete;
    Log& operator = (const Log&) = delete;

    template <typename ...T>
    static void log (LogLevel messageType, const std::string& context, T&&... message) {

        std::unique_lock<std::mutex> lock(m_mutex);

        for (unsigned int i = 0; i < m_observers.size(); ++i) {

            m_observers[i]->notify(messageType, context, concatMessage(std::forward<T>(message)...));

        }

    }

    template <typename ...T>
    static void verbose (const std::string& context, T&&... message) {
        Log::log(LogLevel::VERBOSE, context, std::forward<T>(message)...);
    }

    template <typename ...T>
    static void notice (const std::string& context, T&&... message) {
        Log::log(LogLevel::NOTICE, context, std::forward<T>(message)...);
    }

    template <typename ...T>
    static void warning (const std::string& context, T&&... message) {
        Log::log(LogLevel::WARNING, context, std::forward<T>(message)...);
    }

    template <typename ...T>
    static void error (const std::string& context, T&&... message) {
        Log::log(LogLevel::ERROR, context, std::forward<T>(message)...);
    }

    static void bindUnhandledException ();

    template <typename T, typename ...Targ>
    static LogObserver* makeObserver (Targ&&... args) {

        auto observer = std::make_unique<T>(std::forward<Targ>(args)...);
        auto observerPtr = observer.get();

        m_observers.push_back(std::move(observer));

        return observerPtr;

    }
    static void registerObserver (std::unique_ptr<LogObserver>&&);
    static void registerObserver (LogObserver*);

private:

    template <typename ...T>
    static std::string concatMessage (T&&... message) {

        std::ostringstream oss;
        (oss << ... << std::forward<T>(message));
        return oss.str();

    }

    static std::mutex m_mutex;
    static std::vector<std::unique_ptr<LogObserver>> m_observers;

};
}

#endif

As you can see, the Log class is now nothing more than a shell, a notification system for a collection of observers (and also a system for creating and managing said observers). It presents a new class known as LogObserver, which is a virtual class.

#ifndef H_CLASS_LOGOBSERVER
#define H_CLASS_LOGOBSERVER

#include <string>
#include <set>
#include "../../enum/LogLevel/LogLevel.hpp"
namespace tw{ class Log; }

namespace tw {
class LogObserver {

// WARNING: Never call Log::* functions from inside subclasses of this.
// It will cause infinite recursion and you will be sad.

public:

    static inline constexpr unsigned int ALLOW_UNDEFINED = 0b00001;
    static inline constexpr unsigned int ALLOW_VERBOSE =   0b00010;
    static inline constexpr unsigned int ALLOW_NOTICE =    0b00100;
    static inline constexpr unsigned int ALLOW_WARNING =   0b01000;
    static inline constexpr unsigned int ALLOW_ERROR =     0b10000;

    LogObserver () = default;
    explicit LogObserver (unsigned int);
    explicit LogObserver (const std::set<std::string>&);
    explicit LogObserver (const std::string&...);
    LogObserver (unsigned int, const std::set<std::string>&);
    LogObserver (unsigned int, const std::string&...);
    virtual ~LogObserver () = default;

    LogObserver (LogObserver&&) = default;
    LogObserver& operator = (LogObserver&&) = default;

    LogObserver (const LogObserver&) = default;
    LogObserver& operator = (const LogObserver&) = default;

    void notify (LogLevel, const std::string&, const std::string&);

    bool undefinedAllowed () const;
    void allowUndefined ();
    void blockUndefined ();
    bool verboseAllowed () const;
    void allowVerbose ();
    void blockVerbose ();
    bool noticeAllowed () const;
    void allowNotice ();
    void blockNotice ();
    bool warningAllowed () const;
    void allowWarning ();
    void blockWarning ();
    bool errorAllowed () const;
    void allowError ();
    void blockError ();

    bool contextAllowed (const std::string&...) const;
    void allowContext (const std::string&...);
    void blockContext (const std::string&...);
    bool allContextAllowed () const;
    void allowAllContext ();
    void restrictContext ();

protected:

    virtual void notifyCallback (LogLevel, const std::string&, const std::string&) = 0;

    unsigned int m_allowedLevelFlags {
        ALLOW_UNDEFINED |
        ALLOW_VERBOSE |
        ALLOW_NOTICE |
        ALLOW_WARNING |
        ALLOW_ERROR
    };

    std::set<std::string> m_allowedContexts;
    bool m_allContextsAllowed {true};

};
}

#endif

LogObserver classes contain non-overrideable fields and methods for storing and manipulating data about the log levels and contexts that it’s interested in. The notify method is markedly non-virtual because it implements the code for determining if the LogObserver even cares about the message that’s being sent to notify. As such, notify is the method that is initially invoked whenever a log message is being processed. If all of the checks pass, then it is finally forwarded on to the protected virtual method notifyCallback, which is to be overridden to the tastes and needs of the implementor.

Let’s look at some examples of implementations, the few that I’ve already made, perhaps all that I’ll ever need:

#ifndef H_CLASS_LOGOBSERVER
#define H_CLASS_LOGOBSERVER

#include <string>
#include <set>
#include "../../enum/LogLevel/LogLevel.hpp"
namespace tw{ class Log; }

namespace tw {
class LogObserver {

// WARNING: Never call Log::* functions from inside subclasses of this.
// It will cause infinite recursion and you will be sad.

public:

    static inline constexpr unsigned int ALLOW_UNDEFINED = 0b00001;
    static inline constexpr unsigned int ALLOW_VERBOSE =   0b00010;
    static inline constexpr unsigned int ALLOW_NOTICE =    0b00100;
    static inline constexpr unsigned int ALLOW_WARNING =   0b01000;
    static inline constexpr unsigned int ALLOW_ERROR =     0b10000;

    LogObserver () = default;
    explicit LogObserver (unsigned int);
    explicit LogObserver (const std::set<std::string>&);
    explicit LogObserver (const std::string&...);
    LogObserver (unsigned int, const std::set<std::string>&);
    LogObserver (unsigned int, const std::string&...);
    virtual ~LogObserver () = default;

    LogObserver (LogObserver&&) = default;
    LogObserver& operator = (LogObserver&&) = default;

    LogObserver (const LogObserver&) = default;
    LogObserver& operator = (const LogObserver&) = default;

    void notify (LogLevel, const std::string&, const std::string&);

    bool undefinedAllowed () const;
    void allowUndefined ();
    void blockUndefined ();
    bool verboseAllowed () const;
    void allowVerbose ();
    void blockVerbose ();
    bool noticeAllowed () const;
    void allowNotice ();
    void blockNotice ();
    bool warningAllowed () const;
    void allowWarning ();
    void blockWarning ();
    bool errorAllowed () const;
    void allowError ();
    void blockError ();

    bool contextAllowed (const std::string&...) const;
    void allowContext (const std::string&...);
    void blockContext (const std::string&...);
    bool allContextAllowed () const;
    void allowAllContext ();
    void restrictContext ();

protected:

    virtual void notifyCallback (LogLevel, const std::string&, const std::string&) = 0;

    unsigned int m_allowedLevelFlags {
        ALLOW_UNDEFINED |
        ALLOW_VERBOSE |
        ALLOW_NOTICE |
        ALLOW_WARNING |
        ALLOW_ERROR
    };

    std::set<std::string> m_allowedContexts;
    bool m_allContextsAllowed {true};

};
}

#endif

ConsoleLogObserver is a specialization that sends the messages to the console.

#ifndef H_CLASS_FILELOGOBSERVER
#define H_CLASS_FILELOGOBSERVER

#include <string>
#include <set>
#include <ctime>
#include <iomanip>
#include "../LogObserver/LogObserver.hpp"
#include "../File/File.hpp"

namespace tw {
class FileLogObserver : public LogObserver {

public:

    FileLogObserver () = default;
    explicit FileLogObserver (const std::string&);
    FileLogObserver (const std::string&, unsigned int);
    FileLogObserver (const std::string&, const std::set<std::string>&);
    FileLogObserver (const std::string&, const std::string&...);
    FileLogObserver (const std::string&, unsigned int, const std::set<std::string>&);
    FileLogObserver (const std::string&, unsigned int, const std::string&...);
    ~FileLogObserver () override = default;

    FileLogObserver (FileLogObserver&&) = default;
    FileLogObserver& operator = (FileLogObserver&&) = default;

    FileLogObserver (const FileLogObserver&) = default;
    FileLogObserver& operator = (const FileLogObserver&) = default;

    std::string getFilePath () const;
    bool setFilePath (const std::string&);

protected:

    virtual void notifyCallback (LogLevel, const std::string&, const std::string&) override;

    virtual std::string formatMessage (LogLevel, const std::string&, const std::string&);

    File m_file {"./log.txt", File::ENABLE_WRITE | File::ENABLE_APPEND};

};
}

#endif

FileLogObserver is a specialization that stores a file handle and writes log messages to that file.

And finally, FunctionLogObserver is a specialization that stores and manages a collection of function objects that receive the message (so in a way, it can do anything).

#ifndef H_CLASS_FUNCTIONLOGOBSERVER
#define H_CLASS_FUNCTIONLOGOBSERVER

#include <string>
#include <set>
#include <vector>
#include <functional>
#include <mutex>
#include "../LogObserver/LogObserver.hpp"

namespace tw {
class FunctionLogObserver : public LogObserver {

public:

    typedef std::function<void(LogLevel, const std::string&, const std::string&)> Callback;

    FunctionLogObserver () = default;
    explicit FunctionLogObserver (const std::vector<Callback>&);
    explicit FunctionLogObserver (const Callback&...);
    FunctionLogObserver (const std::vector<Callback>&, unsigned int);
    FunctionLogObserver (const Callback&, unsigned int);
    FunctionLogObserver (const std::vector<Callback>&, const std::set<std::string>&);
    FunctionLogObserver (const Callback&, const std::set<std::string>&);
    FunctionLogObserver (const std::vector<Callback>&, const std::string&...);
    FunctionLogObserver (const Callback&, const std::string&...);
    FunctionLogObserver (const std::vector<Callback>&, unsigned int, const std::set<std::string>&);
    FunctionLogObserver (const Callback&, unsigned int, const std::set<std::string>&);
    FunctionLogObserver (const std::vector<Callback>&, unsigned int, const std::string&...);
    FunctionLogObserver (const Callback&, unsigned int, const std::string&...);
    ~FunctionLogObserver () = default;

    FunctionLogObserver (FunctionLogObserver&&) = default;
    FunctionLogObserver& operator = (FunctionLogObserver&&) = default;

    FunctionLogObserver (const FunctionLogObserver&) = default;
    FunctionLogObserver& operator = (const FunctionLogObserver&) = default;

    const std::vector<Callback>& getCallbacks () const;
    void setCallbacks (const std::vector<Callback>&);
    void addCallback (const Callback&...);
    void clearCallbacks ();

protected:

    virtual void notifyCallback (LogLevel, const std::string&, const std::string&);

    std::mutex m_mutex;
    std::vector<Callback> m_callbacks;

};
}

#endif

If you’d like to read the implementation .cpp files, those are available in the git repo. They didn’t seem to be necessary to illustrate the design pattern. Also, this example code isn’t perfect, and I’m sure in short order I’ll find reason to modify it. At the time of writing, TimberWolf doesn’t have a stable API, so please don’t use this as documentation for it.