[Qemu-devel] Re: [PATCH v2] disable sigcld handling before calling pclose()

[Paolo Bonzini <pbonzini@redhat.com>]

On 12/21/2010 05:05 AM, Wen Congyang wrote:
> When I use the command 'virsh save' to save the domain state,
> I receive the following error message:
> operation failed: Migration unexpectedly failed.
>
> I debug the qemu by adding some printf(), and find the function
> pclose() returns -1.
>
> I use strace to trace qemu, the log is as the following:
> ======
> close(17)                               = 0
> --- SIGCHLD (Child exited) @ 0 (0) ---
> wait4(-1, NULL, WNOHANG, NULL)          = 22016
> rt_sigreturn(0)                         = 0
> wait4(22016, 0x7fff7f1034fc, 0, NULL)   = -1 ECHILD (No child processes)
> ======
>
> We wait the child twice: one is in signal SIGCHLD handling and the other
> one is in pclose().
>
> We should disable sigcld handling before calling pclose().

I wondered whether we need SIGCHLD handling at all.  fork is called only 
from:

- xen_domain_watcher in hw/xen_domainbuild.c

- launch_script in net/tap.c

- SLIRP's fork_exec ("mini inetd")

For the first, the child will always "outlive" the parent.  For the 
second, we do waitpid in the function.  So SLIRP is the only real user 
of the SIGCHLD handler and in fact this:

http://www.google.com/codesearch/p?hl=en#tGk9u3ZS0cw/pub/Linux/system/network/serial/slirp-1.0.9.tar.gz|s3yKHVXI6eg/slirp-1.0.9/src/main.c

suggests that the handler came from there (search for do_wait).  Would 
anybody object to removing the support for Samba under SLIRP and all the 
resulting cruft?

Paolo