6

If our server (running on a device) starts before a DHCP lease had been acquired then it can never connect using a hostname.

If that happens it can find hosts by IP address but not by DNS.

I initially thought that the Curl DNS cache was at fault as the curl connections failed. But I used CURLOPT_DNS_CACHE_TIMEOUT to prevent curl from caching address but connections still failed.

mat_geek
  • 2,481
  • 3
  • 24
  • 29
  • May I dredge up this ancient question by asking if there's any other solution? The trouble is that `res_init()` is much less portable than the `libcurl` API. On some Curl-supporting platforms it doesn't exist; on many others it's not thread-safe. – NickJH Nov 09 '17 at 11:06
  • This issue was finally resolved in glibc 2.26 back in late 2017. It was incorporated into RHE and CentOS in version 7.5: https://sourceware.org/bugzilla/show_bug.cgi?id=984 – jschultz410 May 26 '21 at 14:45

1 Answers1

11

It turns out that glibc gethostbyname_r won't automatically reload it's configuration if that configuration changes. You have to manually call res_init. See bug report below.

Note: Neither the man page for gethostbyname_r nor for rer_init mentioned this limitation.

My solution is very specific. It works for our long running server but it is not my ideal solution.

I have a function that checks the mtime of the /etc/resolv.conf against the last known mtime (0 for DNE). If the two mtime differ then I call res_init. This function is called on program startup and then periodically to optionally reload the configuration.

The glibc bug report

libc caches resolv.conf forever

...

That's what res_init() is for, call it.

Community
  • 1
  • 1
mat_geek
  • 2,481
  • 3
  • 24
  • 29
  • This issue was finally resolved in glibc 2.26 back in late 2017. It was incorporated into RHE and CentOS in version 7.5: sourceware.org/bugzilla/show_bug.cgi?id=984 – jschultz410 May 26 '21 at 14:49