[nginx] Core: when changing binary, newly forked master process should not remove unix domain socket unless old master has quit.

[Maxim Dounin]

Hello!

On Fri, Dec 02, 2016 at 01:23:36AM -0800, Datong Sun via nginx-devel wrote:

> # HG changeset patch
> # User Datong Sun <dndx at idndx.com>
> # Date 1480669905 21600
> #      Fri Dec 02 03:11:45 2016 -0600
> # Node ID 92a7853a439708584bbc45b416df464c64ec446e
> # Parent  1b7a6785d1f3c76f436a9dea5cb0fc86e1db5f2b
> Core: when changing binary, newly forked master process should
> not remove unix domain socket unless old master has quit.

Style, please keep summary line under 67 symbols.  See 
http://nginx.org/en/docs/contributing_changes.html for more 
details.

> 
> This commit fixes a bug that during a binary change, when both
> old and new master processes are running, sending TERM to the
> new master process will accidently remove unix domain socket
> from disk while old master is still listening on it.
> 
> diff -r 1b7a6785d1f3 -r 92a7853a4397 src/core/ngx_connection.c
> --- a/src/core/ngx_connection.c Fri Dec 02 03:06:48 2016 -0600
> +++ b/src/core/ngx_connection.c Fri Dec 02 03:11:45 2016 -0600
> @@ -1007,7 +1007,8 @@
> 
>          if (ls[i].sockaddr->sa_family == AF_UNIX
>              && ngx_process <= NGX_PROCESS_MASTER
> -            && ngx_new_binary == 0)
> +            && ngx_new_binary == 0
> +            && getppid() == 1)
>          {
>              u_char *name = ls[i].addr_text.data + sizeof("unix:") - 1;

The getppid() call doesn't look right in platform-independent 
code.  Especially keeping in mind that the test in question known 
to produce incorrect results in some cases - at least in Solaris 
zones, and AFAIR somehow with systemd.

Also note that with your previous patch this problem starts to 
affect not only TERM, but also QUIT.

-- 
Maxim Dounin
http://nginx.org/