[accept()] currently works, it's [PerlIO] that has problems.

I don't see how PerlIO can be a problem. It's a system to provide IO layers to handles. Individual functions in Perl's IO library might have bugs, but that doesn't mean PerlIO is a problem.

In fact, this replicates the problem outside of PerlIO (I think):

#! perl -lw
use strict;
use threads;
use IO::Socket;

my $lsn = new IO::Socket::INET(
    Listen => 5, LocalPort => '12345'
) or die "Failed to open listening port: $!\n";

async{
    print 'accept starting';
    my $c = $lsn->accept;
    print 'accept returned';
}->detach;

sleep 2;

print "-d starting";
-d $lsn;    # Freezes
print "-d returned";
[download]

The problem does not lie in the CRT. This does everything that my sample in the OP and your sample do, using the CRT directly, and there are no signs of any locking:

<Reveal this spoiler or all in this thread>

Output:

C:\test>828831.exe
828831.c(115):_fstat( lsn, &s ) == -1
        failed with error 9 (Bad file descriptor)
828831.c(116):_tell( lsn )
        failed with error 9 (Bad file descriptor)
828831.c(117):_dup( lsn ) == -1
        failed with error 9 (Bad file descriptor)
Connect from: 120
Thread started for client 120
Connect from: 128
Thread started for client 128
Thread completed 4895 echos for client 128
Thread completed 11575 echos for client 120
[download]

The locking is occuring within Perl itself, but damned if I can work out where. It must be in Win32 specific code given that other OSs aren't suffering the same problem. But given that OS conditional code is spread throughout the entire code base, it'll take someone clever than me to find it.

---

Examine what is said, not who speaks -- Silence betokens consent -- Love the truth but pardon error.

"Science is about questioning the status quo. Questioning authority".

In the absence of evidence, opinion is indistinguishable from prejudice.

"I'd rather go naked than blow up my ass"

Whenever I said accept, I meant Perl's accept. I didn't mean to imply there was a problem with the CRT. There's a problem with how Perl serializes access to the CRT. (At least I presume it's Perl that does the serializing.)

Granted, at first I thought it was a global lock, but it appears to be much finer grained than that. Perl's accept will not block all stat operations, just those on the socket passed to accept.

Is that necessary, or can something be done about it? Why isn't linux affected? I don't know.

But given that OS conditional code is spread throughout the entire code base, it'll take someone clever than me to find it.

The lock appears to be specific to a handle or descriptor. I would start by finding out which, then finding out where the lock is stored in the relevant structure, then find out what uses the lock.