internal/fakecgo: pthread_sigmask uses Linux constants on Darwin/FreeBSD/NetBSD (SIG_SETMASK unblocks all signals) #552

PureGo Version

v0.10.2 (via github.com/shirou/gopsutil/v4 v4.26.7 )

Operating System

• Windows
• macOS
• Linux
• FreeBSD
• NetBSD
• Android
• iOS

Go Version (go version)

1.27.2

What steps will reproduce the problem?

run go.chromium.org/build/siso with profiler enabled.

What is the expected result?

not crash

What happens instead?

signal 27 received on thread with no signal stack
golang/go#77588

Anything else you feel useful to add?

Problem Description

In internal/fakecgo/libcgo.go (with build tag //go:build !cgo && (darwin || freebsd || linux || netbsd)), the constants for pthread_sigmask's how argument are hardcoded to Linux values:

// for pthread_sigmask:
type sighow int32
const (
	SIG_BLOCK   sighow = 0
	SIG_UNBLOCK sighow = 1
	SIG_SETMASK sighow = 2
)

However, on Darwin, FreeBSD, and NetBSD (and other BSD-derived systems), SIG_BLOCK, SIG_UNBLOCK, and SIG_SETMASK are 1-based rather than 0-based:

• SIG_BLOCK = 1
• SIG_UNBLOCK = 2
• SIG_SETMASK = 3
  (For reference, see <sys/signal.h> on Darwin/FreeBSD/NetBSD, or the Go runtime definitions in src/runtime/os_darwin.go, src/runtime/os_freebsd.go, and src/runtime/os_netbsd.go).

Consequences

In _cgo_sys_thread_start (internal/fakecgo/go_darwin.go:20-31, and identically in go_freebsd.go and go_netbsd.go):

sigfillset(&ign)
pthread_sigmask(SIG_SETMASK, &ign, &oset)
...
err = _cgo_try_pthread_create(&p, &attr, unsafe.Pointer(threadentry_trampolineABI0), ts)
pthread_sigmask(SIG_SETMASK, &oset, nil)

Because SIG_SETMASK evaluates to 2, calling pthread_sigmask(2, ...) on Darwin and BSDs executes SIG_UNBLOCK (how = 2) instead of SIG_SETMASK:

1. Instead of temporarily blocking all signals in ign across pthread_create, all signals are unblocked on the calling thread, and the newly spawned thread inherits an all-unblocked signal mask.
2. After pthread_create, pthread_sigmask(SIG_SETMASK /* 2 = SIG_UNBLOCK */, &oset, nil) unblocks oset rather than restoring the previous signal mask.

Impact / Crash Reproduction

When a Go program enables CPU profiling (runtime/pprof.StartCPUProfile) or receives signals while running with purego/internal/fakecgo active (e.g. built with CGO_ENABLED=0 on Darwin):

1. A newly spawned OS thread begins executing threadentry with sigaltstack disabled (SS_DISABLE, the OS default for a new pthread before Go sets up its signal stack).
2. threadentry calls setg_trampoline(...), installing the thread's g0 pointer into TLS. At this moment, ts.g.stackhi is set to the stack byte size (e.g., 0x80000) and ts.g.stacklo is 0, because runtime.mstart0 has not yet had a chance to set g0.stack to the actual OS stack bounds.
3. If SIGPROF (signal 27) arrives on the new thread in the window between setg_trampoline and mstart0, it is delivered on the thread's normal stack (since signals were never masked, and sigaltstack is disabled).
4. Darwin's signal trampoline runtime.sigtramp runs, loads the non-nil g0 from TLS, and calls runtime.sigtrampgo.
5. runtime.sigtrampgo observes that sigaltstack is SS_DISABLE and checks if the current stack pointer sp is within g0.stack ([0, stackhi]). Since sp is in the thread's real virtual memory space (e.g. 0x16... or 0x7f...), this check fails.
6. The Go runtime assumes an external non-Go thread disabled sigaltstack and immediately crashes:

   signal 27 received on thread with no signal stack
   fatal error: non-Go code disabled sigaltstack
   

   (This is the root cause behind runtime: signal 27 received on thread with no signal stack golang/go#77588).

Suggested Fix

Move the definition of SIG_BLOCK, SIG_UNBLOCK, and SIG_SETMASK into OS-specific files (e.g., in libcgo_linux.go and libcgo_darwin.go/libcgo_freebsd.go/libcgo_netbsd.go, or dedicated build-tagged files):
Linux:

const (
	SIG_BLOCK   sighow = 0
	SIG_UNBLOCK sighow = 1
	SIG_SETMASK sighow = 2
)

Darwin, FreeBSD, NetBSD:

const (
	SIG_BLOCK   sighow = 1
	SIG_UNBLOCK sighow = 2
	SIG_SETMASK sighow = 3
)