(ITS#6610) Client receives SIGPIPE when connected via ldapi with TLS

[jzeleny@redhat.com]

Full_Name: Jan Zeleny
Version: 2.4.23
OS: Linux
URL: http://jzeleny.fedorapeople.org/debug/openldap/sigpipe-traces.tar.bz2
Submission from: (NULL) (209.132.186.34)


When running slapd listening on local socket (ldapi:///), clients connecting to
it will sometimes SIGPIPE when using TLS. This happens in about 70% times.

How to reproduce:
generate a pem certificate
slapd -h ldapi:///
ldapsearch -H ldapi:/// -ZZ -x -d -1

I'm attaching straces from both slapd and ldapsearch. What seems to be happening
is that slapd receives EAGAIN during the read from socket, marks it for another
read, but then terminates a reading thread and closes the connection, while
client still wants to write some data. When doing ldapsearch, it does this after
result was returned, that's why it can be seen probably only in debugging mode.

The issue was originally reported on 2.3.43, but I successfully reproduced it on
newer versions, including 2.4.23. The only exception was Fedora rawhide version
(currently 2.4.22), which is built with NSS instead of OpenSSL. NSS (and NSPR)
doesn't seem to support local sockets at all, so it is not possible to use ldapi
with -ZZ any more.

I'm attaching straces from both successful and unsuccessful run. For complete
information here is URL of relevant redhat bugzilla:
https://bugzilla.redhat.com/show_bug.cgi?id=564108