test(wasix): cover POSIX signal semantics (SA_SIGINFO, SA_RESETHAND, signal masks)

[Arshia001]

What

Seven fixtures under lib/wasix/tests/wasm_tests/signal/ covering POSIX signal semantics:

Fixture	Covers
sigaction-siginfo	Both dispatch forms: sa_handler (one argument) and sa_sigaction + SA_SIGINFO (three arguments)
sigaction-resethand	SA_RESETHAND resets the disposition to SIG_DFL on entry
sigaction-mask-defers	A blocked signal is deferred, not nested
sigaction-nodefer-reenters	SA_NODEFER allows re-entry
sigaction-mask-blocks-other	sa_mask blocks other named signals during a handler
sigprocmask-blocks-and-pends	sigprocmask blocks, sigpending reports, unblocking delivers
pthread-sigmask-inherit	A thread inherits its creator's mask, and the mask is per-thread

Why

On wasm the argument count is part of the function type, so a three-argument handler dispatched through the one-argument signature traps with indirect call type mismatch rather than harmlessly ignoring the surplus arguments as native ABIs do. The runtime maps that to WasiError::Exit(Errno::Intr) — exit code 27 — killing the instance. wasix-libc's __wasm_signal did exactly that.

Fixing it exposed two more: SA_RESETHAND was never implemented (a handler re-raising its own signal re-entered itself until the host stack was exhausted), and there was no signal mask at all — sa_mask was stored but never applied, pthread_sigmask was a stub, so every handler ran as though SA_NODEFER were set.

All three are fixed in wasix-org/wasix-libc#128; these are the regression tests.

The mask pair

sigaction-mask-defers and sigaction-nodefer-reenters are the same program differing only in SA_NODEFER, which pins the semantics from both sides. The handler re-raises its own signal and records how often it ran and how deep it nested:

Fixture	Expected
sigaction-mask-defers	calls=5 max_depth=1 — deferred, handler runs sequentially
sigaction-nodefer-reenters	calls=5 max_depth=5 — re-enters, nesting observable

The re-raise is gated on a counter in guest code, so recursion is bounded to 5 frames however libc behaves — no host stack exhaustion even with masking entirely broken. Against today's libc both report max_depth=5, since every handler currently behaves as SA_NODEFER, so mask-defers is the fixture that proves the fix.

Two details worth knowing when reading them:

• nodefer-reenters must use an empty sa_mask: sa_mask is applied on top of SA_NODEFER, so a sigfillset() mask would block the signal anyway and defeat the opt-out being tested.
• mask-defers re-raises from inside each invocation rather than raising 5 times up front, because standard signals do not queue — several raises while blocked collapse into one pending delivery, which would legitimately give calls=2.

sigaction-resethand asserts the disposition reset directly via sigaction(sig, NULL, &old) rather than by re-raising, since unbounded re-entry would take the test process down instead of failing one test.

Test status

Each fixture is red against a sysroot lacking the corresponding fix. For example sigaction-siginfo:

signal/sigaction-siginfo expected exit code 0, got 27
stdout: plain handler ok
error: WASI error: WASI exited with code: ExitCode::27

With sysroots built from wasix-org/wasix-libc#128 all pass, and the full suite is 1100 passed, 0 failed, 71 ignored.

Before merge

All seven carry //#MinimalLibc: v2026-07-30.1 so pinned older sysroots (v2026-05-12.1) skip rather than fail. That version is a placeholder — update it to the libc release that ships the fixes, and land this after that release is available.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Adds a WASIX regression test ensuring sigaction dispatch respects SA_SIGINFO so a 3-argument sa_sigaction handler is not invoked via the 1-argument sa_handler ABI (which would trap on wasm due to function type mismatch).

Changes:

• Add signal/sigaction-siginfo wasm_test that exercises both sa_handler (1-arg) and sa_sigaction+SA_SIGINFO (3-arg) handler paths.
• Pin the test behind a //#MinimalLibc: gate so older sysroots skip the test until the libc fix is available.

[marxin]

LGTM - depends on not yet released wasix-libc, right?