The Architecture of Open Source Applications (Volume 1)
Battle for Wesnoth

Richard Shimooka and David White

Programming tends to be considered a straightforward problem solving activity; a developer has a requirement and codes a solution. Beauty is often judged on the technical implementation's elegance or effectiveness; this book is replete with excellent examples. Yet beyond its immediate computing functions, code can have a profound effect on people's lives. It can inspire people to participate and create new content. Unfortunately, serious barriers exist that prevent individuals from participating in a project.

Most programming languages require significant technical expertise to utilize, which is out of reach for many. In addition, enhancing the accessibility of code is technically difficult and is not necessary for many programs. It rarely translates into neat coding scripts or clever programming solutions. Achieving accessibility requires considerable forethought in project and program design, which often runs counter-intuitive to normal programming standards. Moreover most projects rely upon an established staff of skilled professionals that are expected to operate at a reasonably high level. They do not require additional programming resources. Thus, code accessibility becomes an afterthought, if considered at all.

Our project, the Battle for Wesnoth, attempted to address this issue from its origins. The program is a turn-based fantasy strategy game, produced in an open source model based on a GPL2 license. It has been a moderate success, with over four million downloads at the time of this writing. While this is an impressive metric, we believe the real beauty of our project is the development model that allowed a band of volunteers from widely different skill levels to interact in a productive way.

Enhancing accessibility was not a vague objective set by developers, it was viewed as essential for the project's survival. Wesnoth's open source approach meant that the project could not immediately expect large numbers of highly skilled developers. Making the project accessible to a wide a number of contributors, with varying skill levels, would ensure its long-term viability.

Our developers attempted to lay the foundations for broadening accessibility right from its earliest iteration. This would have undeniable consequences for all aspect of the programming architecture. Major decisions were made largely with this objective in mind. This chapter will provide an in-depth examination of our program with a focus on the efforts to increase accessibility.

The first part of this chapter offers a general overview of the project's programming, covering its language, dependencies and architecture. The second part will focus on Wesnoth's unique data storage language, known as Wesnoth Markup Language (WML). It will explain the specific functions of WML, with a particular emphasis on its effects on in-game units. The next section covers multiplayer implementation and external programs. The chapter will end with some concluding observations on our structure and the challenges of broadening participation.

25.1. Project Overview

Wesnoth's core engine is written in C++, totalling around 200,000 lines at the time of this publication. This represents the core game engine, approximately half of the code base without any content. The program also allows in game content to be defined in a unique data language known as Wesnoth Markup Language (WML). The game ships with another 250,000 lines of WML code. The proportion has shifted over the project's existence. As the program matured, game content that was hardcoded in C++ has increasingly been rewritten so that WML can be used to define its operation. Figure 25.1 gives a rough picture of the program's architecture; green areas are maintained by Wesnoth developers, while white areas are external dependencies.

[Program Architecture]

Figure 25.1: Program Architecture

Overall, the project attempts to minimize dependencies in most cases so as to maximize the portability of the application. This has the added benefit of reducing the program's complexity, and decreases the need for developers to learn the nuances of a large number of third party APIs. At the same time, the prudent use of some dependencies can actually achieve the same effect. For example, Wesnoth uses the Simple Directmedia Layer (SDL) for video, I/O and event handling. It was chosen because it is easy to use and provides a common I/O interface across many platforms. This allows it to be portable to a wide array of platforms, rather than the alternative of coding to specific APIs on different platforms. This comes at a price however; it is harder to take advantage of some platform specific features. SDL also has an accompanying family of libraries that are used by Wesnoth for various purposes:

Additionally, Wesnoth uses several other libraries:

Throughout Wesnoth's engine, the use of WML objects—that is, string dictionaries with child nodes—is fairly ubiquitous. Many objects can be constructed from a WML node, and also serialize themselves to a WML node. Some parts of the engine keep data in this WML dictionary based format, interpreting it directly rather than parsing it into a C++ data structure.

Wesnoth utilizes several important subsystems, most of which are as self-contained as possible. This segmented structure has advantages for accessibility. An interested party can easily work a code in a specific area and introduce changes without damaging the rest of the program. The major subdivisions include:

There are also different modules for controlling different parts of the game flow:

The "play game" module and the main display module are the largest within Wesnoth. Their purpose is the least well defined, as their function is ever-changing and thus difficult to have a clear specification for. Consequently, the modules has often been in danger of suffering from the Blob anti-pattern over the program's history—i.e., becoming huge dominant segments without well-defined behaviors. The code in the display and play game modules are regularly reviewed to see if any of it can be separated into a module of its own.

There are also ancillary features that are part of the overall project, but are separate from the main program. This includes a multiplayer server that facilitates multiplayer network games, as well as a content server that allows users to upload their content to a common server and share it with others. Both are written in C++.

25.2. Wesnoth Markup Language

As an extensible game engine, Wesnoth uses a simple data language to store and load all game data. Although XML was considered initially, we decided that we wanted something a little more friendly to non-technical users, and a little more relaxed with regard to use of visual data. We therefore developed our own data language, called Wesnoth Markup Language (WML). It was designed with the least technical of users in mind: the hope was that even users who find Python or HTML intimidating would be able to make sense of a WML file. All Wesnoth game data is stored in WML, including unit definitions, campaigns, scenarios, GUI definitions, and other game logic configuration.

WML shares the same basic attributes as XML: elements and attributes, though it doesn't support text within elements. WML attributes are represented simply as a dictionary mapping strings to strings, with the program logic responsible for interpretation of attributes. A simple example of WML is a trimmed definition for the Elvish Fighter unit within the game:

[unit_type]
    id=Elvish Fighter
    name= _ "Elvish Fighter"
    race=elf
    image="units/elves-wood/fighter.png"
    profile="portraits/elves/fighter.png"
    hitpoints=33
    movement_type=woodland
    movement=5
    experience=40
    level=1
    alignment=neutral
    advances_to=Elvish Captain,Elvish Hero
    cost=14
    usage=fighter
    {LESS_NIMBLE_ELF}
    [attack]
        name=sword
        description=_"sword"
        icon=attacks/sword-elven.png
        type=blade
        range=melee
        damage=5
        number=4
    [/attack]
[/unit_type]

Since internationalization is important in Wesnoth, WML does have direct support for it: attribute values which have an underscore prefix are translatable. Any translatable string is converted using GNU gettext to the translated version of the string when the WML is parsed.

Rather than have many different WML documents, Wesnoth opts for the approach of all main game data being presented to the game engine in just a single document. This allows for a single global variable to hold the document, and when the game is loaded all unit definitions, for instance, are loaded by looking for elements with the name unit_type within a units element.

Though all data is stored in a single conceptual WML document, it would be unwieldy to have it all in a single file. Wesnoth therefore supports a preprocessor that is run over all WML before parsing. This preprocessor allows one file to include the contents of another file, or an entire directory. For instance:

{gui/default/window/}

will include all the .cfg files within gui/default/window/.

Since WML can become very verbose, the preprocessor also allows macros to be defined to condense things. For instance, the {LESS_NIMBLE_ELF} invocation in the definition of the Elvish Fighter is a call to a macro that makes certain elf units less nimble under certain conditions, such as when they are stationed in a forest:

#define LESS_NIMBLE_ELF
    [defense]
        forest=40
    [/defense]
#enddef

This design has the advantage of making the engine agnostic to how the WML document is broken up into files. It is the responsibility of WML authors to decide how to structure and divide all game data into different files and directories.

When the game engine loads the WML document, it also defines some preprocessor symbols according to various game settings. For instance, a Wesnoth campaign can define different difficulty settings, with each difficulty setting resulting in a different preprocessor symbol being defined. As an example, a common way to vary difficulty is by varying the amount of resources given to an opponent (represented by gold). To facilitate this, there is a WML macro defined like this:

#define GOLD EASY_AMOUNT NORMAL_AMOUNT HARD_AMOUNT
  #ifdef EASY
    gold={EASY_AMOUNT}
  #endif
  #ifdef NORMAL
    gold={NORMAL_AMOUNT}
  #endif
  #ifdef HARD
    gold={HARD_AMOUNT}
  #endif
#enddef

This macro can be invoked using, for instance, {GOLD 50 100 200} within the definition of an opponent to define how much gold the opponent has based on the difficulty level.

Since the WML is processed conditionally, if any of the symbols provided to the WML document change during execution of the Wesnoth engine, the entire WML document must be re-loaded and processed. For instance, when the user starts the game, the WML document is loaded and available campaigns among other things are loaded. But then, if the user chooses to start a campaign and chooses a certain difficulty level—easy for instance—then the entire document will have to be re-loaded with EASY defined.

This design is convenient in that a single document contains all game data, and that symbols can allow easy configuration of the WML document. However, as a successful project, more and more content is available for Wesnoth, including much downloadable content—all of which ends up inserted into the core document tree—which means the WML document is many megabytes in size. This has become a performance issue for Wesnoth: Loading the document may take up to a minute on some computers, causing delays in-game any time the document needs to be reloaded. Additionally, it uses a substantial amount of memory. Some measures are used to counter this: when a campaign is loaded, it has a symbol unique to that campaign defined in the preprocessor. This means that any content specific to that campaign can be #ifdefed to only be used when that campaign is needed.

Additionally, Wesnoth uses a caching system to cache the fully preprocessed version of the WML document for a given set of key definitions. Naturally this caching system must inspect the timestamp of all WML files so that if any have changed, the cached document is regenerated.

25.3. Units in Wesnoth

The protagonists of Wesnoth are its units. An Elvish Fighter and an Elvish Shaman might battle against a Troll Warrior and an Orcish Grunt. All units share the same basic behavior, but many have special abilities that alter the normal flow of gameplay. For example, a troll regenerates some of its health every turn, an Elvish shaman slows its opponents with an entangling root, and a Wose is invisible in a forest.

What is the best way to represent this in an engine? It is tempting to make a base unit class in C++, with different types of units derived from it. For instance, a wose_unit class could derive from unit, and unit could have a virtual function, bool is_invisible() const, which returns false, which the wose_unit overrides, returning true if the unit happens to be in forest.

Such an approach would work reasonably well for a game with a limited set of rules. Unfortunately Wesnoth is quite a large game and such an approach is not easily extendable. If a person wanted to add a new type of unit under this approach, it would require the addition of a new C++ class to the game. Additionally, it does not allow different characteristics to be combined well: what if you had a unit that regenerated, could slow enemies with a net, and was invisible in a forest? You would have to write an entirely new class that duplicates code in the other classes.

Wesnoth's unit system doesn't use inheritance at all to accomplish this task. Instead, it uses a unit class to represent instances of units, and a unit_type class, which represents the immutable characteristics that all units of a certain type share. The unit class has a reference to the type of object that it is. All the possible unit_type objects are stored in a globally held dictionary that is loaded when the main WML document is loaded.

A unit type has a list of all the abilities that that unit has. For instance, a Troll has the "regeneration" ability that makes it heal life every turn. A Saurian Skirmisher has the "skirmisher" ability that allows it to move through enemy lines. Recognition of these abilities is built into the engine—for instance, the pathfinding algorithms will check if a unit has the "skirmisher" flag set to see if it can move freely past enemy lines. This approach allows an individual to add new units, which have any combination of abilities made by the engine, by only editing WML. Of course, it doesn't allow adding completely new abilities and unit behavior without modifying the engine.

Additionally, each unit in Wesnoth may have any number of ways to attack. For instance, an Elvish Archer has a long-range bow attack and also a short-range sword attack. Each deals different damage amounts and characteristics. To represent an attack, there is an attack_type class, with every unit_type instance having a list of possible attack_types.

To give each unit more character, Wesnoth has a feature known as traits. Upon recruitment, most units are assigned two traits at random from a predefined list. For instance, a strong unit does more damage with its melee attacks, while an intelligent unit needs less experience before it "levels up." Also, it is possible for units to acquire equipment during the game that make them more powerful. For instance, there might be a sword a unit can pick up that makes their attacks do more damage. To implement traits and equipment Wesnoth allows modifications on units, which are WML-defined alterations to a unit's statistics. The modification can even be applied to certain types of attacks. For instance, the strong trait gives strong units more damage when attacking in melee, but not when using a ranged strike.

Allowing completely configurable unit behavior with WML would be an admirable goal, so it is instructional to consider why Wesnoth has never achieved such a goal. WML would need to be much more flexible than it is if it were to allow arbitrary unit behavior. Rather than being a data-oriented language, WML would have to be extended into a full-fledged programming language and that would be intimidating for many aspiring contributors.

Additionally, the Wesnoth AI, which is developed in C++, recognizes the abilities present in the game. It takes into account regeneration, invisibility, and so forth, and attempts to maneuver its units to take best advantage of these different abilities. Even if a unit ability could be created using WML, it would be difficult to make the AI sophisticated enough to recognize this ability to take advantage of it. Implementing an ability but not having it accounted for by the AI would not be a very satisfying implementation. Similarly, implementing an ability in WML and then having to modify the AI in C++ to account for the ability would be awkward. Thus, having units definable in WML, but having abilities hard-wired into the engine is considered a reasonable compromise that works best for Wesnoth's specific requirements.

25.4. Wesnoth's Multiplayer Implementation

The Wesnoth multiplayer implementation uses a simple-as-possible approach to implementing multiplayer in Wesnoth. It attempts to mitigate the possibility of malicious attacks on the server, but doesn't make a serious attempt to prevent cheating. Any movement that is made in a Wesnoth game—moving of a unit, attacking an enemy, recruiting a unit, and so forth—can be saved as a WML node. For instance, a command to move a unit might be saved into WML like this:

[move]
    x="11,11,10,9,8,7"
    y="6,7,7,8,8,9"
[/move]

This shows the path that a unit follows as a result of a player's commands. The game then has a facility to execute any such WML command given to it. This is very useful because it means that a complete replay can be saved, by storing the initial state of the game and then all subsequent commands. Being able to replay games is useful both for players to observe each other playing, as well as to help in certain kinds of bug reports.

We decided that the community would try to focus on friendly, casual games for the network multiplayer implementation of Wesnoth. Rather than fight a technical battle against anti-social crackers trying to compromise cheat prevention systems, the project would simply not try hard to prevent cheating. An analysis of other multiplayer games indicated that competitive ranking systems were a key source of anti-social behavior. Deliberately preventing such functions on the server greatly reduced the motivation for individuals to cheat. Moreover the moderators try to encourage a positive gaming community where individuals develop personal rapport with other players and play with them. This placed a greater emphasis on relationships rather than competition. The outcome of these efforts has been deemed successful, as thus far efforts to maliciously hack the game have been largely isolated.

Wesnoth's multiplayer implementation consists of a typical client-server infrastructure. A server, known as wesnothd, accepts connections from the Wesnoth client, and sends the client a summary of available games. Wesnoth will display a 'lobby' to the player who can choose to join a game or create a new game for others to join. Once players are in a game and the game starts, each instance of Wesnoth will generate WML commands describing the actions the player makes. These commands are sent to the server, and then the server relays them on to all the other clients in the game. The server will thus act as a very thin, simple relay. The replay system is used on the other clients to execute the WML commands. Since Wesnoth is a turn-based game, TCP/IP is used for all network communication.

This system also allows observers to easily watch a game. An observer can join a game in-progress, in which case the server will send the WML representing the initial state of the game, followed by a history of all commands that have been carried out since the start of the game. This allows new observers to get up to speed on the state of the game. They can see a history of the game, although it does take time for the observer to get to the game's current position—the history of commands can be fast forwarded but it still takes time. The alternative would be to have one of the clients generate a snapshot of the game's current state as WML and send it to the new observer; however this approach would burden clients with overhead based on observers, and could facilitate denial-of-service attacks by having many observers join a game.

Of course, since Wesnoth clients do not share any kind of game state with each other, only sending commands, it is important that they agree on the rules of the game. The server is segmented by version, with only players using the same version of the game able to interact. Players are immediately alerted if their client's game becomes out of sync with others. This also is a useful system to prevent cheating. Although it is rather easy for a player to cheat by modifying their client, any difference between versions will immediately be identified to players where it can be dealt with.

25.5. Conclusion

We believe that the beauty of the Battle for Wesnoth as a program is how it made coding accessible to a wide variety of individuals. To achieve this aim, the project often made compromises that do not look elegant whatsoever in the code. It should be noted that many of the project's more talented programmers frown upon WML for its inefficient syntax. Yet this compromise enabled one of the project's greatest successes. Today Wesnoth can boast of hundreds of user-made campaigns and eras, created mostly by users with little or no programming experience. Furthermore it has inspired a number of people to take up programming as a profession, using the project as a learning tool. Those are tangible accomplishments that few programs can equal.

One of the key lessons a reader should take away from Wesnoth's efforts is to consider the challenges faced by lesser skilled programmers. It requires an awareness of what blocks contributors from actually performing coding and developing their skills. For example an individual might want to contribute to the program but does not possess any programming skills. Dedicated technological editors like emacs or vim possess a significant learning curve that might prove daunting for such an individual. Thus WML was designed to allow a simple text editor to open up its files, giving anybody the tools to contribute.

However, increasing a code base's accessibility is not a simple objective to achieve. There are no hard and fast rules for increasing code's accessibility. Rather it requires a balance between different considerations, which can have negative consequences that the community must be aware of. This is apparent in how the program dealt with dependencies. In some cases, dependencies can actually increase barriers to participation, while in others they can allow people to contribute more easily. Every issue must be considered on a case-by-case basis.

We should also be careful not to overstate some of Wesnoth's successes. The project enjoyed some advantages that are not easily replicated by other programs. Making code accessible to a wider public is partly a result of the program's setting. As an open source program, Wesnoth had several advantages in this regard. Legally the GNU license allows someone to open up an existing file, understand how it works and makes changes. Individuals are encouraged to experiment, learn and share within this culture, which might not be appropriate for other programs. Nevertheless we hope that there are certain elements that might prove useful for all developers and help them in their effort to find beauty in coding.