Personal tools

Xmonad/Guided tour of the xmonad source

From HaskellWiki

< Xmonad(Difference between revisions)
Jump to: navigation, search
(additional content: the X monad)
(additional content: LayoutClass)
Line 25: Line 25:
   
 
darcs get http://code.haskell.org/xmonad
 
darcs get http://code.haskell.org/xmonad
  +
  +
I intend for this guided tour to keep abreast of the latest darcs changes; if you see something which is out of sync, report it on the xmonad mailing list, or -- even better -- fix it!
   
 
You may also want to refer to the [http://www.haskell.org/haddock/ Haddock]-generated documentation (it's all in the source code, of course, but may be nicer to read this way). You can build the documentation by going into the root of the xmonad source directory, and issuing the command:
 
You may also want to refer to the [http://www.haskell.org/haddock/ Haddock]-generated documentation (it's all in the source code, of course, but may be nicer to read this way). You can build the documentation by going into the root of the xmonad source directory, and issuing the command:
Line 296: Line 298:
 
=== <hask>ManageHook</hask> ===
 
=== <hask>ManageHook</hask> ===
   
=== Layout and Messages ===
+
This is a bit more advanced, and not really a central part of the system, so I'm skipping it for now, hopefully coming back to add more later.
  +
  +
=== <hask>LayoutClass</hask> ===
  +
  +
Next, let's take a look at the <hask>LayoutClass</hask>. This is one of the places that Haskell's type classes really shine. <hask>LayoutClass</hask> is a type class of which every layout must be an instance. It defines the basic functions which define what a layout is and how it behaves. The comments in the source code explain very clearly what each of these functions is supposed to do, but here are some highlights:
  +
  +
* Note that all the <hask>LayoutClass</hask> functions provide default implementations, so that <hask>LayoutClass</hask> instances do not have to provide implementations of those functions where the default behavior is desired. For example, by default, <hask>doLayout</hask> simply calls <hask>pureLayout</hask> by default, so a layout that does not require access to the X monad need only implement the <hask>pureLayout</hask> function. (For example, the Accordion, Square, and Grid layouts from the contrib library all use this approach.)
  +
  +
* Layouts can have their own private state, by storing this state in the <hask>LayoutClass</hask> instance and returning a modified structue (via <hask>doLayout</hask>) when the state changes.
  +
  +
* Both <hask>doLayout</hask> and <hask>handleMessage</hask> have corresponding "pure" versions, which do not give results in the X monad. These functions are never called directly by the xmonad core, which only calls <hask>doLayout</hask> and <hask>handleMessage</hask>, but a layout may choose to implement one (or both) of these "pure" functions, which will be called by the default implementation of the "impure" versions. Layouts which implement <hask>pureLayout</hask> or <hask>pureMessage</hask> are guaranteed to only make decisions about layout or messages (respectively) based on the internal layout state, and not on the state of the system or the window manager in general.
  +
  +
Now, every distinct layout will have a distinct type, although they all must be instances of <hask>LayoutClass</hask>. Given Haskell's strong typing, how can we store different layouts with different workspaces, or even change layouts on the fly? The solution is to wrap layouts using an <i>existential type</i> which hides the particular layout type and only exposes the fact that it is an instance of <hask>LayoutClass</hask>. Not only does this solve the problems caused by separate types for each layout, but it also guarantees that the xmonad core can only ever interact with a layout by calling functions from its <hask>LayoutClass</hask> instance. (Actually, this is a lie, since <hask>doLayout</hask> and <hask>handleMessage</hask> give results in the <hask>X</hask> monad, meaning they have access to the window manager state...)
  +
  +
<haskell>
  +
-- | An existential type that can hold any object that is in Read and LayoutClass.
  +
data Layout a = forall l. (LayoutClass l a, Read (l a)) => Layout (l a)
  +
</haskell>
  +
  +
The <hask>forall l.</hask> abstracts over all layout types (types which are an instance of <hask>LayoutClass</hask>); this is what makes it an existential type. Note that <hask>l</hask> does not appear on the left-hand side of the type declaration. A variable <hask>lay</hask> may have the highly specific type of some particular layout (for example, the types of layouts formed by stacking various layout combinators and transformers can get rather long and hairy!), but no matter what the type of <hask>lay</hask>, <hask>Layout lay</hask> will simply have the type <hask>Layout a</hask> for some window type <hask>a</hask> (usually <hask>Window</hask>).
  +
  +
Note that <hask>(l a)</hask> is also required to be an instance of the <hask>Read</hask> type class -- this is so the state of all the layouts can be serialized and then read back in during dynamic restarts.
  +
  +
=== Messages ===
  +
   
 
=== Utility functions ===
 
=== Utility functions ===

Revision as of 18:30, 30 January 2008

Contents

1 Introduction

Do you know a little Haskell and want to see how it can profitably be applied in a real-world situation? Would you like to quickly get up to speed on the xmonad source code so you can contribute modules and patches? Do you aspire to be as cool of a hacker as the xmonad authors? If so, this might be for you. Specifically, this document aims to:

  • Provide a readable overview of the xmonad source code for Haskell non-experts interested in contributing extensions or modifications to xmonad, or who are just curious.
  • Highlight some of the uniquenesses of xmonad and the things that make functional languages in general, and Haskell in particular, so ideally suited to this domain.

This is not a Haskell tutorial. I assume that you already know some basic Haskell: defining functions and data; the type system; standard functions, types, and type classes from the Standard Prelude; and at least a basic familiarity with monads. With that said, however, I do take frequent detours to highlight and explain more advanced topics and features of Haskell as they arise.

2 First things first

You'll want to have your own version of the xmonad source code to refer to as you read through the guided tour. In particular, you'll want the latest darcs version, which you can easily download by issuing the command:

darcs get http://code.haskell.org/xmonad

I intend for this guided tour to keep abreast of the latest darcs changes; if you see something which is out of sync, report it on the xmonad mailing list, or -- even better -- fix it!

You may also want to refer to the Haddock-generated documentation (it's all in the source code, of course, but may be nicer to read this way). You can build the documentation by going into the root of the xmonad source directory, and issuing the command:

runhaskell Setup haddock

which will generate HTML documentation in dist/doc/html/xmonad/.

Without further ado, let's begin!

3 StackSet.hs

StackSet.hs is the pure, functional heart of xmonad. Far removed from corrupting pollutants such as the IO monad and the X server, it is a beatiful, limpid pool of pure code which defines most of the basic data structures used to store the state of xmonad. It is heavily validated by QuickCheck tests; the combination of good use of types and QuickCheck validation means that we can be very confident of the correctness of the code in StackSet.hs.

3.1
StackSet

The
StackSet
data type is the mother-type which stores (almost) all

of xmonad's state. Let's take a look at the definition of the

StackSet
data type itself:
data StackSet i l a sid sd =
    StackSet { current  :: !(Screen i l a sid sd)    -- ^ currently focused workspace
             , visible  :: [Screen i l a sid sd]     -- ^ non-focused workspaces, visible in xinerama
             , hidden   :: [Workspace i l a]         -- ^ workspaces not visible anywhere
             , floating :: M.Map a RationalRect      -- ^ floating windows
             } deriving (Show, Read, Eq)
First of all, what's up with
i l a sid sd
? These are type parameters to
StackSet
---five types which must be provided to form a concrete instance of
StackSet
. It's not obvious just from this definition what they represent, so let's talk about them first, so we have a better idea of what's going on when they keep coming up later.
  • The first type parameter, here represented by
    i
    , is the type of workspace tags. Each workspace has a tag which uniquely identifies it (and which is shown in your status bar if you use the DynamicLog extension). At the moment, these tags are simply
    String
    s---but, as you can see, the definition of
    StackSet
    doesn't depend on knowing exactly what they are. If, in the future, the xmonad developers decided that
    Complex Double
    s would make better workspace tags, no changes would be required to any of the code in StackSet.hs!
  • The second type parameter
    l
    is somewhat mysterious---there isn't much code in StackSet.hs that does much of anything with it. For now, it's enough to know that the type
    l
    has something to do with layouts;
    StackSet
    is completely independent of particular window layouts, so there's not much to see here.
  • The third type parameter,
    a
    , is the type of a single window.
  • sid
    is a screen id, which identifies a physical screen; as we'll see later, it is (essentially)
    Int
    .
  • sd
    , the last type parameter to
    StackSet
    , represents details about a physical screen.

Although it's helpful to know what these types represent, it's

important to understand that as far as
StackSet
is concerned, the particular types don't matter. A
StackSet
simply organizes data

with these types in particular ways, so it has no need to know the actual types.

The
StackSet
data type has four members:
current
stores the currently focused workspace;
visible
stores a list of those

workspaces which are not focused but are still visible on other

physical screens;
hidden
stores those workspaces which are, well, hidden; and
floating
stores any windows which are in the floating

layer.

A few comments are in order:

  • visible
    is only needed to support multiple physical screens with Xinerama; in a non-Xinerama setup,
    visible
    will always be the empty list.
  • Notice that
    current
    and
    visible
    store
    Screen
    s, whereas
    hidden
    stores
    Workspace
    s. This might seem confusing until you realize that a
    Screen
    is really just a glorified
    Workspace
    , with a little extra information to keep track of which physical screen it is currently being displayed on:
data Screen i l a sid sd = Screen { workspace :: !(Workspace i l a)
                                  , screen :: !sid
                                  , screenDetail :: !sd }
    deriving (Show, Read, Eq)
  • A note about those exclamation points, as in
    workspace :: !(Workspace i l a)
    : they are strictness annotations which specify that the fields in question should never contain thunks (unevaluated expressions). This helps ensure that we don't get huge memory blowups with fields whose values aren't needed for a while and lazily accumulate large unevaluated expressions. Such fields could also potentially cause sudden slowdowns, freezing, etc. when their values are finally needed, so the strictness annotations also help ensure that xmonad runs smoothly by spreading out the work.
  • The
    floating
    field stores a
    Map
    from windows (type
    a
    ,remember?) to
    RationalRect
    s, which simply store x position, y position, width, and height. Note that floating windows are still stored in a
    Workspace
    in addition to being a key of
    floating
    , which means that floating/sinking a window is a simple matter of inserting/deleting it from
    floating
    , without having to mess with any
    Workspace
    data.

3.2
StackSet
functions

StackSet.hs also provides a few functions for dealing directly with

StackSet
values:
new
,
view
, and
greedyView
. For example, here's
new
:
new :: (Integral s) => l -> [i] -> [sd] -> StackSet i l a s sd
new l wids m | not (null wids) && length m <= length wids = StackSet cur visi unseen M.empty
  where (seen,unseen) = L.splitAt (length m) $ map (\i -> Workspace i l Nothing) wids
        (cur:visi)    = [ Screen i s sd |  (i, s, sd) <- zip3 seen [0..] m ]
                -- now zip up visibles with their screen id
new _ _ _ = abort "non-positive argument to StackSet.new"
If you're
new
(haha) to Haskell, this might seem dauntingly complex,

but it isn't actually all that bad. In general, if you just take things slowly and break them down piece by piece, you'll probably be surprised how much you understand after all.

new
takes a layout thingy (
l
), a list of workspace tags (
[i]
), and a list of screen descriptors (
[sd]
), and produces a new
StackSet
. First, there's a guard, which requires
wids
to be nonempty (there must be at least one workspace), and
length m
to be at most
length wids
(there can't be more screens than workspaces). If those conditions are met, it constructs a
StackSet
by creating a list of empty
Workspace
s, splitting them into
seen
and
unseen

workspaces (depending on the number of physical screens), combining

the
seen
workspaces with screen information, and finally picking the

first screen to be current. If the conditions on the guard are not met, it aborts with an error. Since this function will only ever be

called internally, the call to
abort
isn't a problem: it's there

just so we can test to make sure it's never called! If this were a function which might be called by users from their xmonad.hs configuration file, aborting would be a huge no-no: by design, xmonad should never crash for any reason (even user stupidity!).

Now take a look at
view
and
greedyView
.
view
takes a workspace tag and a
StackSet
, and returns a new
StackSet
in which the given workspace has been made current.
greedyView
only differs in the way it treats Xinerama screens:
greedyView
will always swap the

requested workspace so it is now on the current screen even if it was

already visible, whereas calling
view
on a visible workspace will

just switch the focus to whatever screen it happens to be on. For single-head setups, of course, there isn't any difference in behavior

between
view
and
greedyView
. Note that
view
/
greedyView
do not modify a
StackSet
, but simply

return a new one computed from the old one. This is a common purely functional paradigm: functions which would modify a data structure in an imperative/non-pure paradigm are recast as functions which take an old version of a data structure as input and produce a new version. This might seem horribly inefficient to someone used to a non-pure paradigm, but it actually isn't, for (at least) two reasons. First, a lot of work has gone into memory allocation and garbage collection, so that in a modern functional language such as Haskell, these processes are quite efficient. Second, and more importantly, the fact that Haskell is pure (modifying values is not allowed) means that when a new structure is constructed out of an old one with only a small change, usually the new structure can actually share most of the old one, with new memory being allocated only for the part that changed. In an impure language, this kind of sharing would be a big no-no, since modifying the old value later would suddenly cause the new value to change as well.

3.3
Workspace

The
Workspace
type is quite simple. It stores a tag, a layout, and possibly a
Stack
:
data Workspace i l a = Workspace  { tag :: !i, layout :: l, stack :: Maybe (Stack a) }
    deriving (Show, Read, Eq)
If there are no windows in a given workspace,
stack
will be
Nothing
; if there are windows, it will be
Just s
, where
s
is a non-empty
Stack
of windows. There's not much else to say about it, which makes this a perfect chance to talk about record syntax. The basic way to define the
Workspace
type would be:
data Workspace i l a = Workspace i l (Maybe (Stack a))
This simply specifies a single constructor for the
Workspace
type (perhaps somewhat confusingly, also called
Workspace
, although these are two different things) which has three components, of types
i
,
l
, and
Maybe (Stack a)
, respectively. The record syntax in the actual code wraps the components in curly braces, and adds a name associated with each component. These names automatically turn into accessor functions which allow us to extract the corresponding component from a value of type
Workspace i l a
. For example,
tag
becomes a function of type
tag :: Workspace i l a -> i

Hence, we have two ways to get at the internals of any value whose type is defined using record syntax: pattern-matching, or accessor functions.

3.4
Stack

The
Stack
type stores a list of the actual windows on a given workspace, along with a notion of the "current" window. Now, the "obvious" way to do this in an imperative language would be to store an array of windows along with an index into the array. However, this approach has several disadvantages:
  • Creating a new window or deleting the current one would be O(n) operations, as all the windows to the right of the current location would have to be shifted by one in the array.
  • In Haskell, indexing into a list is O(n) anyway, and using an array library would be unwieldy here.
  • Much work must go into maintaining guarantees such as always having the current index be a valid index into the array, maintaining the ordering of the windows when shifting them around in the array, and so on.
Instead, a
Stack
uses an ingenious structure known as a list zipper:
data Stack a = Stack { focus  :: !a        -- focused thing in this set
                     , up     :: [a]       -- clowns to the left
                     , down   :: [a] }     -- jokers to the right
    deriving (Show, Read, Eq)
Instead of using a single list with some sort of index, the list is broken into three pieces: a current window (
focus
), the windows before that, in reverse order (
up
), and the windows after it (
down
). This has several nice properties:
  • A
    Stack a
    cannot be empty, since it must always contain a current element. Remember, the possibility of an empty workspace is handled by the type of
    Workspace
    's
    stack
    field,
    Maybe (Stack a)
    .
  • Shifting focus, adding a new window next to the current one, and reversing the window list are all simple O(1) operations.
  • There is not even the possibility of any sort of index-out-of-bounds errors while keeping track of the current window.

For more information on zippers, the Zipper page on the Haskell wiki and the chapter on zippers in the Haskell wikibook are good starting places.

3.5 Other functions

At this point you should spend some time studying the rest of the functions in StackSet.hs, which provide various operations on
Stack
s and
StackSet
s. There are quite a few, but they are, for the most part, quite straightforward. Some general notes and commentary:
  • The functions
    with
    ,
    modify
    , and
    modify'
    are great examples of higher-order functions, functions which take other functions as input. Haskell (and most functional languages) make such a thing easy and natural. For example,
    with
    applies a function to the current workspace's stack;
    modify
    essentially transforms a function on
    Stack
    s to a function on
    StackSet
    s, with some
    Maybe
    types thrown in to handle empty cases.
  • Note how the implementation of functions such as
    focusUp
    /
    Down
    ,
    swapUp
    /
    Down
    , and
    reverseStack
    are quite simple, thanks to higher-order functions and the zipper structure of
    Stack
    s.
  • I sort of lied when I said that moving focus is O(1) with the zipper structure: in the one case that focus wraps around the end of the list, it is O(n). But it is still takes O(1) amortized time.

3.6 QuickCheck properties

Now take a look at tests/Properties.hs, which contains a large collection of QuickCheck properties for testing StackSet.hs. Here's the way it works: a QuickCheck property is essentially a function of type
a -> Bool
which should always evaluate to
True
for any value of type
a
. (Actually, it's a little more complex than this: using a special return type
Property
instead of
Bool
, a property can also specify only a subset of values for which the property should hold.) The type
a
must be an instance of the type class
Arbitrary
, which defines a way to generate random values of that type. Running QuickCheck on a property generates a large amount of random test data of the appropriate type, and ensures that the property always evaluates to true; if not, it displays the input for which the property failed. Of course, this is not proof that the property is always true, but a set of well-defined properties is surprisingly good at catching bugs and forgotten corner cases.

These tests are run automatically with every recorded change to the xmonad source repository -- indeed, they must all pass in order for a change to be recorded.

4 Core.hs

The next source file to examine is Core.hs. It defines several core data types and some of the core functionality of xmonad. If StackSet.hs is the heart of xmonad, Core.hs is its guts.

4.1
XState
,
XConf
, and
XConfig

These three record types make up the core of xmonad's state and configuration:

  • A value of type
    XState
    stores xmonad's mutable runtime state, consisting of the list of workspaces, a set of mapped windows, something to do with keeping track of pending UnmapEvents, and something to do with dragging.
data XState = XState
    { windowset    :: !WindowSet           -- ^ workspace list
    , mapped       :: !(S.Set Window)      -- ^ the Set of mapped windows
    , waitingUnmap :: !(M.Map Window Int)  -- ^ the number of expected UnmapEvents
    , dragging     :: !(Maybe (Position -> Position -> X (), X ())) }
Note that the
WindowSet
type is just a
StackSet
with various concrete types substituted for its type parameters:
type WindowSet   = StackSet  WorkspaceId (Layout Window) Window ScreenId ScreenDetail
type WindowSpace = Workspace WorkspaceId (Layout Window) Window
 
-- | Virtual workspace indicies
type WorkspaceId = String
 
-- | Physical screen indicies
newtype ScreenId    = S Int deriving (Eq,Ord,Show,Read,Enum,Num,Integral,Real)
 
-- | The 'Rectangle' with screen dimensions and the list of gaps
data ScreenDetail   = SD { screenRect :: !Rectangle
                         , statusGap  :: !(Int,Int,Int,Int) -- ^ width of status bar on the screen
                         } deriving (Eq,Show, Read)
The type of workspace tags,
WorkspaceId
, is just
String
; a
ScreenDetail
stores information about the screen dimension as well as any gaps which should be left at the edges of the screen for status bars and other such things. Now, you may wonder why we have
newtype ScreenId = S Int deriving (...)
rather than just
type ScreenId = Int
? The reason is that if
type ScreenId = Int
, it would be possible to accidentally do arithmetic with
ScreenId
s mixed with other
Int
values, which clearly doesn't make sense. Making
ScreenId
a separate type means that the type system will enforce non-mixing of
ScreenId
s with other types, while using
newtype
with automatic instance deriving means none of the convenience is lost -- we can write code just as if
ScreenId
s are normal
Int
values, but be sure that we can't accidentally get mixed up and do something silly like add a
ScreenId
to the width of a window. This is the power of a rich static type system like Haskell's -- we can encode certain invariants and constraints in the type system, and have them automatically checked at compile time.
  • An
    XConf
    record stores xmonad's (immutable) configuration data, such as window border colors, the keymap, information about the X11 display and root window, and other user-specified configuration information. The reason this record is separated from
    XState
    is that, as we'll see later, xmonad's code provides a static guarantee that the data stored in this record is truly read-only, and cannot be changed while xmonad is running.
  • XConfig
    provides a way for the user to customize xmonad's configuration, by defining an
    XConfig
    record in their
    xmonad.hs
    file. You're probably already familiar with this record type.

4.2 The
X
monad

And now, what you've all been waiting for: the X monad!

newtype X a = X (ReaderT XConf (StateT XState IO) a)
    deriving (Functor, Monad, MonadIO, MonadState XState, MonadReader XConf)
The X monad represents a common pattern for building custom monad instances: using monad transformers, one can simply 'layer' the capabilities and effects of several monads into one, and then use GHC's newtype deriving capabilities to automatically derive instances of the relevant type classes. In this case, the base monad out of which the X monad is built is IO; this is necessary since communicating with the X server involves IO operations. On top of that is
StateT XState
, which automatically threads a mutable
XState
record through computations in the X monad; finally there is a
ReaderT XConf
which also threads a read-only
XConf
record through. As noted in the comments in the source, the
XState
record can be accessed with any functions in the <div class="inline-code">
MonadState
type class, such as
get
,
put
,
gets
, and
modify
; the
XConf
record can be accessed with
MonadReader
</div>
functions, such as
ask
.

For more information on monad transformers in general, I recommend reading Martin Grabmüller's excellent tutorial paper, Monad Transformers Step-by-Step; for more information on this particular style of composing monad transformers and using automatic newtype deriving, read Cale Gibbard's tutorial, How To Use Monad Transformers.

Along with the X monad, several utility functions are provided, such as
runX
, which turns an action in the X monad into an IO action, and
catchX
, which provides error handling for X actions. There are also several higher-order functions provided for convenience, including
withDisplay
(apply a function producing an X action to the current display) and
withWindowSet
(apply a function to the current window set).

4.3
ManageHook

This is a bit more advanced, and not really a central part of the system, so I'm skipping it for now, hopefully coming back to add more later.

4.4
LayoutClass

Next, let's take a look at the
LayoutClass
. This is one of the places that Haskell's type classes really shine.
LayoutClass
is a type class of which every layout must be an instance. It defines the basic functions which define what a layout is and how it behaves. The comments in the source code explain very clearly what each of these functions is supposed to do, but here are some highlights:
  • Note that all the
    LayoutClass
    functions provide default implementations, so that
    LayoutClass
    instances do not have to provide implementations of those functions where the default behavior is desired. For example, by default,
    doLayout
    simply calls
    pureLayout
    by default, so a layout that does not require access to the X monad need only implement the
    pureLayout
    function. (For example, the Accordion, Square, and Grid layouts from the contrib library all use this approach.)
  • Layouts can have their own private state, by storing this state in the
    LayoutClass
    instance and returning a modified structue (via
    doLayout
    ) when the state changes.
  • Both
    doLayout
    and
    handleMessage
    have corresponding "pure" versions, which do not give results in the X monad. These functions are never called directly by the xmonad core, which only calls
    doLayout
    and
    handleMessage
    , but a layout may choose to implement one (or both) of these "pure" functions, which will be called by the default implementation of the "impure" versions. Layouts which implement
    pureLayout
    or
    pureMessage
    are guaranteed to only make decisions about layout or messages (respectively) based on the internal layout state, and not on the state of the system or the window manager in general.
Now, every distinct layout will have a distinct type, although they all must be instances of
LayoutClass
. Given Haskell's strong typing, how can we store different layouts with different workspaces, or even change layouts on the fly? The solution is to wrap layouts using an existential type which hides the particular layout type and only exposes the fact that it is an instance of
LayoutClass
. Not only does this solve the problems caused by separate types for each layout, but it also guarantees that the xmonad core can only ever interact with a layout by calling functions from its
LayoutClass
instance. (Actually, this is a lie, since
doLayout
and
handleMessage
give results in the
X
monad, meaning they have access to the window manager state...)
-- | An existential type that can hold any object that is in Read and LayoutClass.
data Layout a = forall l. (LayoutClass l a, Read (l a)) => Layout (l a)
The
forall l.
abstracts over all layout types (types which are an instance of
LayoutClass
); this is what makes it an existential type. Note that
l
does not appear on the left-hand side of the type declaration. A variable
lay
may have the highly specific type of some particular layout (for example, the types of layouts formed by stacking various layout combinators and transformers can get rather long and hairy!), but no matter what the type of
lay
,
Layout lay
will simply have the type
Layout a
for some window type
a
(usually
Window
). Note that
(l a)
is also required to be an instance of the
Read
type class -- this is so the state of all the layouts can be serialized and then read back in during dynamic restarts.

4.5 Messages

4.6 Utility functions

4.7 On-the-fly recompilation