[syslog-ng] Question about syntax for source

[Balazs Scheidler]

On Fri, 2011-04-08 at 11:14 +0200, Sandor Geller wrote:
> Hello Bazsi,
> 
> On Fri, Apr 8, 2011 at 12:36 AM, Balazs Scheidler <bazsi at balabit.hu> wrote:
> > On Mon, 2011-04-04 at 12:18 +0200, Sandor Geller wrote:
> >> Hello,
> >>
> >> On Sun, Apr 3, 2011 at 9:02 PM, Gergely Nagy <algernon at balabit.hu> wrote:
> >> > Phusion <phusion2k at gmail.com> writes:
> >> >
> >> >> When trying to convert this for syslog-ng 3.x under OpenBSD. I have an
> >> >> error. When I use syslog-ng -s -f ../syslog-ng.conf it works, but when
> >> >> I run syslog-ng -p /var/run/syslog-ng.pid, I get the following error.
> >> >>
> >> >> # /usr/local/sbin/syslog-ng -p /var/run/syslog-ng.pid
> >> >> WARNING: you are using the pipe driver, underlying file is not a FIFO,
> >> >> it should be used by file(); filename='/dev/klog'
> >> >
> >> > The warning pretty much tells you what to do, and why: change pipe() to
> >> > file(), because /dev/klog appears to be something else than your common
> >> > pipe.
> >>
> >> This warning is misleading. As reported a while ago using file() for
> >> anything else than non-regular files could lead to severe problems.
> >>
> >> > In syslog-ng 2.x, things worked a bit differently (the details escape
> >> > me, I'm afraid), which changed in 3.x, and thus, the warning is printed.
> >> >
> >> >> Here is the current source line using syslog-ng 3.0 under OpenBSD 4.8.
> >> >>
> >> >> source local { internal(); pipe("/dev/klog" program_override("kernel:
> >> >> ")); unix-dgram("/dev/log"); };
> >> >
> >> > Change it to something like this:
> >> >
> >> > source local {
> >> >       internal();
> >> >       file("/dev/klog" program_override("kernel: "));
> >> >       unix-dgram("/dev/log");
> >> > };
> >>
> >> Don't do this. Commit 61940d18c205d36cb7dd0b30dba741cc8459e2ac fixed
> >> the underlying problem in the 3.2 branch. When a new version will get
> >> released then the warning would remain but at least syslog-ng would
> >> actually check that the source is a regular file and assume
> >> readability only in this case, otherwise it will poll() the source
> >> which is the wanted behaviour for character devices and pipes.
> >
> > But If I remember correctly, that only affected 3.2, right?
> 
> In 3.0.8 using pipe() on a non-fifo source or using file() on a fifo
> source was a fatal error. In 3.1.3 the errors were downgraded to
> warnings, 3.2 behaves more or less the same. I haven't checked 3.3
> yet.
> 
> In 3.2 git there is an additional check so LW_ALWAYS_WRITABLE is set
> only for regular file destinations which fixes the blocking write
> problem reported on the list a while ago but there is still a warning
> which could get suppressed. I think pipe should be the preferred
> driver for anything else than regular files, an extra poll() won't
> hurt. What do you think about this patch?
> 
> --- affile.c    2011-02-13 14:34:38.000000000 +0100
> +++ affile.c-new        2011-04-08 11:13:52.000000000 +0200
> @@ -46,6 +46,7 @@
>  {
>    cap_t saved_caps;
>    struct stat st;
> +  gboolean is_regular;
> 
>    if (strstr(name, "../") || strstr(name, "/.."))
>      {
> @@ -75,20 +76,21 @@
>    *fd = -1;
>    if (stat(name, &st) >= 0)
>      {
> -      if (is_pipe && !S_ISFIFO(st.st_mode))
> +      is_regular = !!S_ISREG(st.st_mode);
> +      if (!is_pipe && !is_regular)
>          {
> -          msg_warning("WARNING: you are using the pipe driver,
> underlying file is not a FIFO, it should be used by file()",
> +          msg_warning("WARNING: you are using the file driver,
> underlying file is not a regular file, it should be used by pipe()",
>                      evt_tag_str("filename", name),
>                      NULL);
>          }
> -      else if (!is_pipe && S_ISFIFO(st.st_mode))
> +      else if (is_pipe && is_regular)
>          {
> -          msg_warning("WARNING: you are using the file driver,
> underlying file is a FIFO, it should be used by pipe()",
> +          msg_warning("WARNING: you are using the pipe driver,
> underlying file is a regular file, it should be used by file()",
>                        evt_tag_str("filename", name),
>                        NULL);
>          }
>        if (regular)
> -        *regular = !!S_ISREG(st.st_mode);
> +        *regular = is_regular;
>      }
>    else if (regular)
>      *regular = TRUE;
> 
> Regards,
> 
> Sandor
> 

The intent has always been to use pipe() for named pipes, and file() for
everything else.

I'm now thinking about introducing a device() destination/source which
would be equivalent to file(), but follow_freq() would not be allowed.

What do you think?

-- 
Bazsi