clj-con
defines a set of concurrency operations modeled after their Clojure
counterparts. Sample operators include future
, promise
, deref
,
deliver
, and atom
. See the exported symbols in package.lisp for the full list.
Or, if you're familiar with the Clojure Cheatsheet, the project implements the following:
along with promise
and deliver
.
Mar-03-2024: Added to Ultralisp because quicklisp hasn't been updated in 5 months and there are updates I really wanted to get out.
If you didn't get this via quickload using a quicklisp/ultralisp repo, add it to
your ~/quicklisp/localprojects/
directory and update/wipe the
system-index.txt
file accordingly, and then you can quickload it.
;; See 'local-projects' note in preceding paragraph
(ql:quickload :clj-con) ; to use the code
or
;; To run the tests
(ql:quickload :clj-con-test)
(clj-con-test:run-tests)
This package will wishfully run on any lisp supporting bordeaux-threads
,
which is most of them.
Lisps supporting the atomics
package will use compare-and-swap
behavior
via atomics:cas
for the atom
implementation. At the time
of this writing (DEC 2023), that includes:
- Allegro
- CCL
- ECL
- LispWorks
- Mezzano
- SBCL
- CMUCL
- CLASP (*)
*CLASP isn't on the atomics README but does seem to be supported in the code.
If the atomics-provided :atomics-cas-svref
isn't in *features*
then the
implementation defaults to a locking behavior to emulate compare-and-swap
in the various atom
functions that require it.
Here are my experiences so far with Fedora 38 running on an Intel machine and tests run on some of the lisps. Note that I am primarily an SBCL user and am not particularly familiar with the other lisps, their heap configurations, or even how to debug them since I didn't bother to enable them for SLIME, I just ran them from the command line.
All tests were run with default memory configurations. I have no explanation for why some of them seem to be running out of memory, though some may be running with overly conservative heap sizes by default. The test suite allocates fewer than 50 threads, but the CLJ-CON package does expect threads to be (eventually) reclaimed when the lisp code running on them returns.
Given that the test suite does deliberately signal conditions in the bodies of many thread tests I suppose it's possible threads are hung and locks are not being released. There are caveats about unpredicable unwind behavior w.r.t. locks in some of the tools used.
All locking is done via bt2:with-lock-held
. If you want lisps that seem
to keep chugging along even with many allocations, look for the ones I've
labelled "GOOD".
-
SBCL GOOD
RUN-TESTS passes.
(dotimes (i 1000) (debug! 'test-suite))
passes. -
CCL GOOD
RUN-TESTS passes.
(dotimes (i 1000) (debug! 'test-suite))
passes.The test seemed to slow a bit toward the end of the 1000 iterations, as if perhaps a lot of gc activity were happening, but it did finish.
-
ABCL 1.9.0, OpenJDK 17.0.9 GOOD (No CAS)
RUN-TESTS passes.
(dotimes (i 1000) (debug! 'test-suite))
passes.This now works but required special timeout logic because BT2:CONDITION-WAIT always returns T on ABCL. Unfortunatey the logic which fixes ABCL breaks tests on CCL, and perhaps others.
-
LispWorks 8.0.1 Personal Edition. UNRELIABLE
RUN-TESTS passes.
(dotimes (i 1000) (debug! 'test-suite))
runs out of memory. Lots of messages on the console "Hanging Unknown thread 5612"On a personal note. No init file with personal edition. Really?
-
Allegro CL Express 11.0 (
alisp
executable) UNRELIABLERUN-TESTS passes.
(dotimes (i 1000) (debug! 'test-suite))
gets the following error after a number of iterations:Running test suite TEST-SUITE Running test PROMISE-DELIVERY .... Running test NO-TIMEOUT-WAITS Allegro CL(pid 1082089): System Error (gsgc) Object already pointing to target newspace half: 0x1000c8a1a68 The internal data structures in the running Lisp image have been corrupted and execution cannot continue. Check all foreign functions and any Lisp code that was compiled with high speed and/or low safety, as these are two common sources of this failure. If you cannot find anything incorrect in your code you should contact technical support for Allegro Common Lisp, and we will try to help determine whether this is a coding error or an internal bug.
The message suggests a gc bug, but maybe that's just a symptom of running out of memory.
-
ECL 21.2.1 UNRELIABLE
Works for the minimal (run once)
clj-con-test:run-tests
case, but runs out of memory if the test suite is run repeatedly.
compare-and-set!
now returns NIL and non-NIL, instead of strict NIL and T values.deliver
no longer returns the value delivered, it returns the input promise or nil according to clojure semantics, see the doc string fordeliver
.
See "Tested Lisps" above.
Added conditional use of the atomics
package for a real compare-and-swap
behavior in the atom
implementation.
Recent testing with ECL found that ECL doesn't like condition broadcasts with recursive locks. The recursive locks were changed to non-recursive locks, hopefully without loss of functionality or introduction of bugs.
The motiviation was to use CONDITION-BROADCAST
which is not in APIV1
and was forcing CLJ-CON code to loop on CONDITION-NOTIFY
.
Only tested with SBCL and ABCL, known to be broken on ECL.
reset-vals!
and swap-vals!
return vectors in Clojure but return
multiple values here. I couldn't see the point of returning vectors when CL
has no destructuring bind that works on vectors. At least you can use
multiple-value-bind if you want, though it doesn't destructure either.
Tip: metabang-bind (available in quicklisp) provides a nice destructuring tool that also handles multiple values.
YMMV if you use the atomics-enabled compare-and-swap
behavior on Common
Lisp characters and numbers, because it uses EQ
semantics, not EQL
, and
EQ is not necessarily true for numbers and characters.
In SBCL, fixnums are usually EQ, and (eq #\a #\a)
will likely return
true, but have a care.
If you're going to (use :clj-con)
note that atom
requires a
(:shadowing-import-from #:clj-con #:atom)
.
The Java Virtual Machine's threading tools are really a marvelous thing. If
you've been in that ecosystem a long time, going back to pthreads with some of
its limitations (or lisp oddities built on them), will feel fragile, and
reading the various SBCL source comments on interrupt-thread
doesn't do much
to prevent that feeling.
The test suite does test future-cancel
and other ways of unwinding the
thread stack, and seems to work on all tested platforms. But it may still be a
source of bugs, such as the memory problems noted on some lisps.
Have a care if you are repeatedly interrupting threads or using complicated mission critical handlers in the threads unless you have taken to heart the use of SBCL's WITHOUT-INTERRUPTS and other appropriate implementation dependent tools. I didn't hit any problems with my simple tests but that isn't saying much.
There is no attempt here to bring clojure syntax or persistent data structures to Common Lisp. Fortunately neither of those things is particularly prevalent in Clojure's concurrency operator model, at least not in the clojure.core namespace.
Some enterprising person might want to make a readtable that maps @
to
deref
, assuming it doesn't conflict with ,@
, but that hasn't been done
here so you'll just have to call deref
.
If you're missing clojure.core.async and want some blocking queues for producer/consumer
situations, take a look at the lparallel.queue
package (ql:quickload :lparallel)
. Unlike clojure.core.async it has a peek
operator which I find useful
when I need to speculatively try something on a queue element without losing FIFO ordering.
The Atomics maintainers were considering adding some queue capabilities in 2023, so you may wish to check there as well. It isn't in the quicklisp distribution as of June 2023 though.
I recommend reading documentation on the bordeaux-threads make-thread function for cautions about interactions between threads and dynamic variables.
You also need to mentally prepare yourself for how values you've closed
over in the body of your future
can mutate. Consider this example:
(dotimes (i 20)
(future ... (print i) ...))
You may be expecting the first value of i
printed by the first future created
would be zero because dotimes starts at zero. However depending on your lisp
implementation it may actually print one, or some other value, depending on
time of evaluation and whether the reference to the location/register
holding 'i' has been incremented or not by the time the future body is
executed on the new thread.
The clj-con-test
package has a test case where this exact issue was encountered on
SBCL, and the workaround was to use something like this:
(dotimes (i 20)
(let ((i2 i)) ;of course I could have rebound 'i' as well
(future ... (print i2) ...)))
This way the future is referencing a binding that won't change.
Binding semantics such as the above may vary by lisp implementation and has nothing to do with parallelism. E.g., you might get this:
(let ((funs nil))
(dotimes (i 3) (push (lambda () i) funs))
(dotimes (j 3) (print (funcall (elt funs j)))))
3
3
3
The behavior is related to closing over bindings for mutable data. The CL spec for dotimes says this:
"It is implementation-dependent whether dotimes establishes a new binding
of var on each iteration or whether it establishes a binding for var once
at the beginning and then assigns it on any subsequent iterations."
When in doubt, add a binding that won't change for use in your closed over
future
(or other) bodies.
(reverse "moc.liamg@ynnet.evad")
This is a secondary address that isn't monitored every day. Feel free to submit Github issues if appropriate.