Re: Library support for two-phase daemonization

[christos%astron.com@localhost (Christos Zoulas)]

In article <21223.61263.989443.746257%guava.gson.org@localhost>,
Andreas Gustafsson  <gson%NetBSD.org@localhost> wrote:
>Hi all,
>
>As discussed in PR 48282, some daemons background themselves before
>opening the sockets or registering the RPCs they receive requests on.
>This leads to race conditions: if you try to use the service provided
>by a daemon such as nfsd immediately after starting it, you get random
>failures because it may not yet be ready to provide service.
>
>In some cases, this can be fixed by moving the call to daemon(3)
>later, to a point where the socket/RPC setup is complete.  But this
>won't work in daemons that use kqueues or threads, because kqueues and
>threads are not inherited across a fork.
>
>I think the right way to fix the problem is to split the daemonization
>into two phases: the daemon should fork early, before creating threads
>or kqueues, but the parent process should only exit when the child is
>ready to accept requests for service.
>
>A proof of concept implementation of such a scheme is already in
>src/tests/lib/libc/net/h_dns_server.c, and I would now like to move
>this code into a library so that it can be used by other daemons, too.
>
>I propose to do this by adding the two functions daemon2_fork() and
>daemon2_detach() to libc.  A patch containing my proposed changes is
>at:
>
>   http://www.gson.org/netbsd/patches/daemon2-v9.patch
>
>The patch includes updates to the daemon(3) man page and test cases.
>Initially, two daemons would be switched to the new scheme: nfsd and
>rquotad (the ones involved in the test failures reported in PR 48282);
>others can follow later.
>
>Comments?  Objections?

- the assumption that pfd[0] + 1 == pfd[1] is not true. If you start
  with 0, 1, 2 open then close 1, and issue pipe(), you'll end up
  with pfd[0] = 1, pfd[1] = 3. So you should test that both pfd[0]
  and pfd[1] are > STDERR_FILENO.
- I think most fd's should be close on exec here. so you could use pipe2()
  to setup sigpipe and close-on-exec handlign.
- results of read/write are ssize_t
- I like to write fork as a switch instead of ifthenelif, r is pid_t.
- why isn't the detach daemon pipe static?

christos