Description of problem: NetworkManager can reconfigure /etc/resolv.conf and it can even notify anyone who registers. But the glibc resolver never finds out /etc/resolv.conf has changed beneath. Version-Release number of selected component (if applicable): rawhide-20080409.i386 (exact glibc version upon request) How reproducible: Seen once, verified /etc/resolv.conf is never read on its change. Steps to Reproduce: 1. Put invalid nameserver into `/etc/resolv.conf'. 2. Run the testcase (./netdb). 3. Put valid nameserver into `/etc/resolv.conf'. Actual results: Testcase still prints (nil) Expected results: Testcase should print 0xhexaddr for immediate valid name resolving: killall -USR1 netdb Additional info: NetworkManagerDispatch provides the directory `/etc/NetworkManager/dispatcher.d' where glibc can place its hook to notify all the running applications to reload `/etc/resolv.conf'. But I am not aware how to notify glibc without colliding with any application using glibc vs. the application's signals/mainloop. Not filed upstream as NetworkManager is more Fedora specific. nscd was not installed in this case.
Created attachment 302209 [details] Testcase not seeing changed `/etc/resolv.conf'.
I think the preferred upstream solution proposed by the glibc maintainers is to use nscd. NM already does: /etc/init.d/nscd condrestart /usr/sbin/nscd -i hosts whenever it updates resolv.conf.
In such case NetworkManager should rpm-Require the package nscd and it should start the nscd daemon along. In Bug 442171 was shown the practical case of a minimal Fedora installation where nscd is missing and in such case "it just does not work". It would be best if NetworkManager would be able to call res_init() in all the running applications using libresolv without nscd but neither Jakub nor me knows a safe+cheap method of such notification (shmat(2) of a flag-page?).
NM isn't going to require nscd because it's not a hard dep and not everyone wants to run it. People that care need to start nscd if they want it. Having NM call res_init() in each process seems like a gross hack though; if that's the lengths NM has to go to, then the call should be in glibc itself...
Dan, glibc should call res_init() itself but it needs to know when to do so. Any kernel call (such as stat() or time()) is not acceptable for the maintainers. Going to try to provide a `rdtsc'-protected stat() call if it will be feasible.
Changing version to '9' as part of upcoming Fedora 9 GA. More information and reason for this action is here: http://fedoraproject.org/wiki/BugZappers/HouseKeeping
Could you please reply to the previous question? If you won't reply in one month, I will have to close this bug as INSUFFICIENT_DATA. Thank you.
The current rawhide nscd version automatically will discover when /etc/resolv.conf is changed and it invalidates the database. It is necessary to have check-files passwd yes in /etc/nscd.conf.
I'ts of course check-files hosts yes
Ulrich, this Bug is whole about a configuration without nscd installed. If nscd is required for the proper functionality of glibc nscd should be Required by the rpm packaging. It is currently not the case. (As I still did not provide a patch the Bug should remain CLOSED, though.)
(In reply to comment #10) > this Bug is whole about a configuration without nscd installed. If nscd is > required for the proper functionality of glibc nscd should be Required by the > rpm packaging. nscd is required to make changes work. Period. I won't add code into the stub resolver to check the file every time. This is far too slow. We have a perfectly fine solution, so use it. I've been arguing to change the default to use nscd by default for a long time. In fact, we should completely remove the option to not use nscd.