@rittneje, do you want to send a CL to fix this?

(CC @ianlancetaylor @bradfitz @tklauser per https://dev.golang.org/owners)

@bcmills unix.Pselect is auto-generated code, so I'm not sure how to fix it.

unix.Pselect is generated from a directive in syscall_linux.go.

I expect the general form of the fix would be:

1. Change the existing directive to have a new, unexported name (pselect6) with the correct signature for the kernel syscall and re-run the generator, something like:

type pselect6Sigset_t {
	ss *Sigset_t
	ssLen uintptr  // Size (in bytes) of object pointed to by ss.
};
//sys pselect6(nfd int, r *FdSet, w *FdSet, e *FdSet, timeout *Timespec, sigmask *pselect6Sigset_t) (n int, err error)

2. Add a new, manually-written func Pselect that implements the POSIX signature in terms of the new (Linux-specific) function:

func Pselect(nfd int, r *FdSet, w *FdSet, e *FdSet, timeout *Timespec, sigmask *Sigset_t) (n int, err error) {
	// Per https://man7.org/linux/man-pages/man2/select.2.html#NOTES,
	// “The Linux pselect6() system call modifies its timeout argument.
	// [Not modifying the argument] is the behavior required by POSIX.1-2001.”
	var mutableTimeout *Timespec
	if timeout != nil {
		mutableTimeout = new(Timespec)
		*mutableTimeout = *timeout
	}

	// “The final argument of the pselect6() system call is not a
	// sigset_t * pointer, but is instead a structure …”
	var kernelMask *pselect6Sigset_t
	if sigmask != nil {
		kernelMask = &pselect6Sigset_t{
			ss: sigmask,
			ssLen: unsafe.Sizeof(*sigmask),
		}
	}

	return pselect6(nfd, r, w, e, mutableTimeout, kernelMask)
}

Plus a regression test, of course. 🙃

Not sure I will be able to submit the patch as it will be somewhat complicated.

Unfortunately, ss_len is not just sizeof(sigset_t) because that would be too easy. From traversing through glibc, these are the defines that are relevant:

#define __WORDSIZE	64 // or 32
#define ULONG_WIDTH __WORDSIZE
#define UCHAR_WIDTH 8

#define ALIGN_DOWN(base, size)	((base) & -((__typeof__ (base)) (size)))

#define ALIGN_UP(base, size)	ALIGN_DOWN ((base) + (size) - 1, (size))

#define __NSIG_WORDS (ALIGN_UP ((_NSIG - 1), ULONG_WIDTH) / ULONG_WIDTH)

#define __NSIG_BYTES (__NSIG_WORDS * (ULONG_WIDTH / UCHAR_WIDTH))

You then need to pass _NSIG_BYTES for ss_len. (On my machine it works out to be 8.)

Somehow x/sys/unix already knows _NSIG, so at least part of the work is done. Not sure exactly how that works though.

In terms of regression testing, this currently fails with an error "invalid argument":

	v := &unix.Sigset_t{}
	v.Val[0] = ^v.Val[0]
	n, err := unix.Pselect(0, nil, nil, nil, &unix.Timespec{Sec: 1}, v)

After correcting it to call pselect6 properly it succeeds. (I'm not sure how unix.Sigset_t is intended to be used since it doesn't appear sigaddset, etc. were mapped.)

I think this is how to compute ss_len: https://go.dev/play/p/p6sodpgiXoB

@rittneje In triage here, so I added the "help wanted" since I wasn't sure if you're pursuing the patch for this. Please feel free to let me know if you are, and I'll drop the label. Thanks!

Just pushed a CL: https://go-review.googlesource.com/c/sys/+/510195

This is fixed, it looks like! Thanks.