Fellow haskellers,
I have a proposal I would like to enter into the eternal top-level IO
debate. The proposal
involves a minor language extension and some runtime support for
thread local reference cells.
I believe it has the potential to meet many of the needs of those
requesting top level IO.
My apologies for this rather lengthy message, but given the
volatility of discussion on this matter
in the past, it seemed best to lay out my thoughts as precisely as
possible at the beginning.
So without further ado, the proposal.
The language extension:
-- Add a new keyword 'threadlocal' and a top level declaration with
the syntax
threadlocal <identifier> <type expression> <initializer
expression>
-- The declaration consists of three parts 1) the name of the
reference cell 2) the type of
data stored in the cell and 3) an initializer action.
-- The name of the cell declared with 'threadlocal' shares the
function namespace and introduces
a CAF with the type 'TLRef a', where 'a' stands for the type
given in the declaration.
-- The initializer action must be an expression with type 'TL a',
where 'a' stands for the
the type given in the 'threadlocal' declaration, and 'TL' is the
thread-local initialization
monad (a little like the ACIO monad, more on this below).
The semantics:
-- Each thread in a program (the main thread and threads sparked from
forkIO and forkOS) has a "bank"
of thread local variables. Writes to and reads from a thread-
local cell are only written to/read
from the bank of the thread performing the write/read.
-- For any given bank, a thread-local cell may be "empty" (which
means it holds no value) or "full"
with a value of its declared type.
-- There is a phantom bank of thread-local values belonging to no
thread in which the value of all
thread-local cells is "empty". This represents the state of
thread local variables before program
start.
-- Whenever a thread is sparked (including the main thread) and
before it begins executing, its
thread-local variables are initialized. For each declared
thread-local variable (in the transitive
closure of imported modules), the declared initilzation action
is run and the generated value initializes
the thread-local cell for that thread. The initializer actions
are run in an unspecified order.
-- The primitives of the TL are strictly limited and include only
actions which have no observable
side effects (a proposed list of primitives is listed below). A
TL action may read from (but NOT write
to) thread-local cells in the bank of the sparking thread (the
bank of the thread calling forkIO, or the
special phantom bank for the main thread).
-- Any exceptions generated during a thread-local initilization
action are propigated to the thread
which called forkIO/forkOS or, in the case of the main thread,
directly to the runtime system just
as though an uncaught exception bubbled off the main thread.
-- New IO primitives are added to read from, write to and clear (set
to empty) thread-local variables.
Advantages:
This proposal seems to hit most of the use cases that I recall having
seen (including the very important
allocate-a-top-level-concurrency-variable use case) and seems to
provide a nice way to reinterpret some
of the "magic" currently in the standard libraries. In addition,
this proposal does not suffer from the
module loading order problem that some previous proposals have;
because thread local initializer actions
depend only on the "previous" bank of values, the order in which they
are run makes very little difference
(only for the primitives that read clock time or some such). The
value of a thread-local cell is always well-defined,
even before the main thread starts. Values in a thread-local have a
well defined lifetime that is tied to the owning
thread. I think that efficient implementation is possible (maybe we
can play some copy-on-write games?).
I especially like that variables are only as "global" as desired for
any given task; if a library writer
uses thread-locals for some manner of shared "global" state, later
users are always able to write programs
that use more than one instance of the "global" state without needing
to alter the library.
Disadvantages:
Requires a language extension (but I don't know of a serious
alternate proposal that doesn't). Requires
non-trivial runtime system support. Not sure what effect this has on
garbage collection. Adds overhead to
thread creation (this could perhaps be mitigated by introducing new
primitives that distinguish heavyweight
threads with their own thread-local banks from lightweight threads,
which do not have separate thread-local banks).
Its a bit complicated. You can shoot yourself in the foot (true of
most of the other proposals).
Some representative use cases:
-- Implicit parameter style use case:
You want to provide a default value that you expect will be
rarely changed. Threading the
parameter deeply through the code obsfucates meaning and the
code is in the IO monad.
Solution:
* Define a thread-local variable for the value and set the
initializer to set the variable to
some default value if it was empty, or to copy the parent
thread's value otherwise.
* If desired, change the variable's value early in main
(before any other threads are sparked);
the new value will be propagated to all new threads and be
available in main.
* If desired, different threads can set different values of
the parameter which will then
be propagated to their sub-threads.
-- Top level synchronization variable use case:
You need an MVar to manage some "global" resource.
Solution:
* Define a thread-local variable to hold the MVar. Define the
initializer to create a new MVar
if the TLRef was empty, and to copy the parent thread's
value otherwise.
* Read the MVar from the thread-local var, and use it as usual.
* Also, allows you to partition the program into sandboxes
which use distinct MVars to manage
distinct pools of the "global" resource without needing to
change or complicate any underlying
libraries.
-- Running time statistics use case:
You want to easily keep track of how long each thread in a
program has been running in wall-clock time.
Solution:
* Define a thread-local variable with an initializer that
reads the current wall-clock time.
* Calculate the thread running time by taking the difference
between the current wall-clock time
and the time in the thread-local var.
-- Give better semantics to standard handles use case:
You want to make the handling of stdin, stdout and stderr in
System.IO less "magic" and baked-in.
Solution:
* Define a thread-local variable for each of stdin, stdout,
and stderr. The initializer action
creates appropriate handles for each one in the case that
the thread-local was empty, and
copies the parent's value otherwise.
* Nice feature: allows you to override the values of stdin,
stdout and stderr for sub-threads,
shell style.
-- Give better semantics to getArgs use case:
You want getArgs/withArgs and getProgName/withProgName to have
better semantics.
Solution:
* Define a thread-local variable for the list of arguments and
for the program name. Define an
initializer which reads these values from C land when the
thread-local value is empty and copies
the parent thread's value otherwise.
* Allows you to spark multiple threads in a single program
with different "command line" arguments.
The proposed list of new primitives:
-- thread-locals maintained by the standard libs
currentWorkingDirectory :: TLRef FilePath
stdin :: TLRef Handle
stdout :: TLRef Handle
stderr :: TLRef Handle
-- in the TL monad
readTL :: TLRef a -> TL a -- ^ Reads a thread-local variable in
the bank of the parent thread.
-- Returns bottom if the cell is
empty.
tryReadTL :: TLRef a -> TL (Maybe a)
-- ^ Reads a thread-local variable in
the bank of the parent thread.
-- Returns Nothing if the cell is
empty.
getClocktimeTL :: TL ClockTime
getCPUTimeTL :: TL Integer
getDefaultStdinHandle :: TL Handle
getDefaultStdoutHandle :: TL Handle
getDefaultStderrHandle :: TL Handle
getDefaultArgs :: TL [String]
getDefaultProgramName :: TL String
newIORefTL :: a -> TL (IORef a)
newMVarTL :: a -> TL (MVar a)
newEmptyMVarTL :: TL (MVar a)
newSTRefTL :: a -> TL (STRef a)
-- in the IO monad
readThreadLocal :: TLRef a -> IO a -- ^ bottom on empty
tryReadThreadLocal :: TLRef a -> IO (Maybe a) -- ^ Nothing on
empty
writeThreadLocal :: a -> TLRef a -> IO ()
clearThreadLocal :: TLRef a -> IO () -- ^ Reset the
thread-local to empty
clearBank :: IO () -- ^ Clear all
thread-local cells in the current thread
------------------------------------------------------------------------
----------------
In order to get discussion flowing and experiment with the semantics
I have put together
a demonstration module which implements thread-local variables using
the standard
"unsafePerformIO" hacks. The module and an example usage are attached.
If there is any interest in these ideas, I will post this proposal to
the wiki.
Please respond with thoughts and comments,
Rob Dockins
Speak softly and drive a Sherman tank.
Laugh hard; it's a long way to the bank.
-- TMBG
