In the paper Tackling the awkward squad, Simon Peyton Jones has provided a "possible implementation" of a Channel
.
type Channel a = (MVar (Stream a) , -- Read end
MVar (Stream a) ) -- Write end (the hole)
type Stream a = MVar (Item a)
data Item a = MkItem a (Stream a)
Now, he implements a function putChan :: Channel a -> a -> IO ()
like this
putChan (read, write) val
= do { new_hole <- newEmptyVar ;
old_hole <- takeMVar write ;
putMVar write new_hole ;
putMVar old_hole (MkItem val new_hole) }
The function above takes a MVar out of write, then puts an empty MVar into it.
Then it writes to the old_hole it extracted from write.
The question is, why does it write to old_hole? It has been taken out from write and its scope is limited to the current block only, then what difference does it make?
Not quite.
old_hole
is "in scope" on the read side. You have to look atnewChan
for the full picture:So right after calling
newChan
the "old_hole
" fromputChan
is the sameMVar
ashole
innewChan
. And as channel operations progress, thatold_hole
is always somewhere nested in theMVar
s ofread
.I found the design of linked list-style channels truly difficult to wrap my head around at first. The illustration from that paper does a decent job of showing the structure, but the basic idea is readers "peel off" a layer of MVar to reveal a value, while writers are inserting values "at the bottom" of the pile of MVars, mainting a pointer to the bottom-most one.
By the way this is the design used in Control.Concurrent.Chan