Hi,
I've found a race condition in libpq. It is about the initialization of
the my_bio_methods static variable in fe-secure-openssl.c, which is not
protected by any lock. The race condition may make the initialization of
the connection fail, and as an additional weird consequence, it might
cause openssl call close(0), so stdin of the client application gets
closed.
I've prepared a patch to protect the initialization of my_bio_methods
from the race. This is my first patch submission to the postgresql
project, so I hope I didn't miss anything. Any comments and suggestions
are of course very welcome.
I also prepared a testcase. In the testcase tarball, there is a patch
that adds sleeps at the right positions to make the close(0) problem
occur practically always. It also includes comments to explain how the
race can end up calling close(0).
Concerning the patch, it is only tested on Linux. I'm unsure about
whether the simple initialization of the mutex would work nowadays also
on Windows or whether the more complicated initialization also to be
found for the ssl_config_mutex in the same source file needs to be used.
Let me know whether I should adapt that.
We discovered the problem with release 11.5, but the patch and the
testcase are against the master branch.
Regards,
Willi
--
___________________________________________________
Dr. Willi Mann | Principal Software Engineer, Tech Lead PQL
Celonis SE | Theresienstrasse 6 | 80333 Munich, Germany
F: +4989416159679
w.mann@celonis.com | www.celonis.com | LinkedIn | Twitter | Xing
AG Munich HRB 225439 | Management: Martin Klenk, Bastian Nominacher,
Alexander Rinke
