illumos gate

Bug is reproducible as early as https://github.com/illumos/illumos-gate/commit/c79df70e7f04e3742087189b820b431059193dfe (September 20) and not reproducible at https://github.com/illumos/illumos-gate/commit/582237e641696363bb029af5d31bf63b495e1b7d (August 15).
Perhaps, we can to bisect.

A bit of dtracing shows that zlogin is calling poll, and getting POLLRDNORM | POLLIN for the zone side stdout, but then getting 0 returned from a subsequent read() on the file descriptor which it handles as an EOF and closes the connection.

Changing zlogin.c:process_output() so that read() == 0 is not handled as an EOF makes zlogin stop exhibiting this problem.

What's strange is that this is happening on stdout and not stderr, yet the test program only triggers the problem when writing to stderr.

The behaviour of zlogin changes depending on whether it is built with gcc7 or gcc 4.4.4:

gcc 7¶

% pwd
/data/omnios-build/omniosorg/bloody/illumos/usr/src/cmd/zlogin
% make
...

% pfexec ./zlogin test
[Connected to zone 'test' pts/2]
Last login: Wed Dec 11 13:45:06 2019 on pts/2
OmniOS 5.11     omnios-master-f3e2549426        December 2019
root@test:~# ./zcons
.root@test:~#
[Connection to zone 'test' pts/2 closed]

gcc 4.4.4¶

% PRIMARY_CC=gcc4,/opt/gcc-4.4.4/bin/gcc,gnu make
...

% pfexec ./zlogin test
[Connected to zone 'test' pts/2]
Last login: Wed Dec 11 13:44:34 2019 on pts/2
OmniOS 5.11     omnios-master-f3e2549426        December 2019
root@test:~# ./zcons
.root@test:~#
root@test:~# logout

[Connection to zone 'test' pts/2 closed]

The problem here is not the compiler but rather that gcc7 is including values-xpg6.o in the link for C99 and above.

The problem was exposed by

commit d3b5f56344d8bfcdd6cfb82446af0e5e55ad9ebe
CommitDate: Fri Aug 16 01:57:38 2019 -0700

    11528 Makefile.noget can get gone
    11529 Use -Wno-maybe-initialized

because it changed zlogin to build with C99.

Building zlogin without the C99 flags produces a binary which does not exhibit this problem.

Summary:

In August, a change (11529) that was cleaning up the use of -Wno-initialized in
various places changed the zlogin Makefile to use -std=gnu99 instead of
-std=gnu98. This caused the illumos version of gcc7 to link values-xpg6.o into
the final object which has a number of side-effects.

Most important for zlogin is that when read() on a terminal returns 0, it no
longer means EOF, just that no data is available.