cljw builds for wasm32-wasi (ADR-0193). That build is a WASI command, so a
native cljw can run it with wasm/run: untrusted Clojure evaluated inside the
zwasm sandbox, under fuel, memory, output and wall-clock budgets, with only
the authority the host passes in.
(wasm/run "zig-out/bin/cljw.wasm"
{:args ["cljw" "-"]
:stdin "(println (reduce + (range 100)))"
:dir "/tmp/cljw-guest-scratch"})
;=> {:out "4950\n", :err "", :exit 0}zig build -Dtarget=wasm32-wasi -Doptimize=ReleaseSmall # zig-out/bin/cljw.wasm
cljw docs/examples/wasm/clojure-guest/guest.cljThe host must be wasm-enabled (cljw --version shows wasm; the released
binaries are). The .wasm is a build output and is not committed. Set
CLJW_GUEST to run a guest built elsewhere.
The guest needs one preopened directory. At startup it reads deps.edn from
its working directory, and under WASI a relative path needs a preopen to
resolve against; with none, path_open gets EBADF and the guest exits with
error: Unexpected before it reads any input. The same happens under Node's
WASI, so it is the guest, not the engine. An empty scratch directory passed as
:dir is enough, and it is then the guest's whole filesystem.
2026-11-11. Guest: this tree (v1.15.7), Zig 0.16.0, ReleaseSmall. Host: cljw v1.15.3 (ReleaseSafe, wasm), x86_64 Linux, with the machine under load (load average about 34 on 22 cores), so treat the times as upper bounds.
cljw.wasm size |
3,692,268 B |
| imports | wasi_snapshot_preview1 only (32 functions) |
| exports | memory, _start |
| memory floor | starts at 448 pages (28 MiB); 384 pages fails in bootstrap |
| first run (compile, instantiate, eval) | 600 to 700 ms |
| later runs (module cached, instantiate and eval) | 140 to 230 ms |
Every program gave the JVM Clojure answer: (reduce + (range 100)) 4950, a
recursive fib 6765 and 75025, a map/filter/reduce pipeline 1330,
clojure.string split, reverse, join and upper-case "C-B-A", a caught
ex-info with its message and data. An uncaught exception is {:exit 1} with
the message on :err.
Both zwasm engines run it. wasm/run has no :engine option today (it always
uses :auto, JIT first), so the per-engine numbers come from the zwasm v2.7.0
CLI, zwasm run --engine <e> --dir <empty>:. cljw.wasm -:
| interpreter | JIT | |
|---|---|---|
nil (startup) |
2.8 s | 0.66 s |
(fib 25) |
19.1 s | 1.1 s |
(loop [] (recur)), --fuel 100000000 |
out_of_fuel trap, 2.5 s | out_of_fuel trap, 4.8 s |
cons forever, --max-memory 64 MiB |
OutOfMemory, 22 s |
OutOfMemory, 3.1 s |
What each limit does, through wasm/run with only :args, :stdin and an
empty :dir:
- Fuel.
(loop [] (recur))with:fuel 100000000trapsout_of_fuelin about 4 s and returns{:out "" :err "" :exit 1}. With no:fuelthe default budget traps it in about 44 s.:timeout-ms 2000with:fuel 0interrupts it at 2.0 s. The host is never hung and gets data, not an exception. The trap text (zwasm: trap kind=out_of_fuel) goes to the host process's stderr, not into:err, so a caller cannot yet tell a fuel trap from a timeout from any other exit 1. - Memory. Consing forever under
:max-memory-pages 1024(64 MiB) ends in about 2 s with{:err "<stdin>: error: OutOfMemory\n" :exit 1}: the guest's allocator seesmemory.growrefused and fails cleanly. Under the default page budget the same happens in about 18 s. - Files.
(slurp "/etc/passwd")throws inside the guest, catchable there, asslurp '/etc/passwd' failed: AccessDenied. So does../../../../etc/passwd. The guest can write in the scratch directory (spitthere succeeds), so pass a fresh directory per run if runs must not see each other. - Environment.
(System/getenv "HOME")isniland(System/getenv)is{}.
Single-threaded: future, send to an agent, pmap and Thread.start each
end the guest with error: ThreadsUnsupported and exit 1, which a try inside
the guest does not catch. No network (sockets and HTTP are compiled out). No FFI.
It is a stdin-to-stdout evaluator: each wasm/run is a fresh instance with no
state carried between runs.