Two buildfarm runs[1][2] from the last 90 days have failed in
src/test/authentication, like this:
t/001_password.pl ..
Failed 3/8 subtests
t/002_saslprep.pl .. ok 1815 ms ( 0.00 usr 0.00 sys + 0.89 cusr 0.26 csys = 1.15 CPU)
Test Summary Report
-------------------
t/001_password.pl (Wstat: 13 Tests: 5 Failed: 0)
Non-zero wait status: 13
Parse errors: Bad plan. You planned 8 tests but ran 5.
Files=2, Tests=17, 3 wallclock secs ( 0.04 usr 0.00 sys + 1.67 cusr 0.50 csys = 2.21 CPU)
Result: FAIL
Compared to a good run, the other logs just end suddenly after the expected
"FATAL: password authentication failed". "Wstat: 13" means the Perl process
died to signal 13 (SIGPIPE). This test invokes psql in a way that fails
authentication, and it writes "SELECT 1" to psql's stdin. The SIGPIPE happens
if the psql process exits before that write. I can reproduce this
consistently by patching a delay into IPC::Run:
--- IPC/Run.pm~ 2017-05-12 06:46:43.000000000 -0700
+++ IPC/Run.pm 2017-12-09 00:24:06.876132628 -0800
@@ -1449,6 +1449,8 @@
sub run {
local $in_run = 1; ## Allow run()-only optimizations.
my IPC::Run $self = start(@_);
+ use Time::HiRes 'usleep';
+ usleep 100_000;
The two src/test/authentication tests then fail, but nothing else fails.
Let's ignore SIGPIPE in all TAP tests, which leaves some evidence in
regress_log_001_password:
ack Broken pipe: write( 13, 'SELECT 1' ) at /home/nm/sw/cpan/lib/perl5/IPC/Run/IO.pm line 549.
To fix the actual failures, we can cease sending "SELECT 1"; it's enough to
disconnect immediately. Patch attached.
[1] https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=crake&dt=2017-11-23%2008%3A37%3A17
[2] https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=mandrill&dt=2017-10-30%2000%3A14%3A09