DTrace script to trace processes entering vfs::vop_remove

[dteske at FreeBSD.org]

> -----Original Message-----
> From: Mark Johnston [mailto:markjdb at gmail.com] On Behalf Of Mark
> Johnston
> Sent: Wednesday, December 3, 2014 4:45 PM
> To: dteske at FreeBSD.org
> Cc: freebsd-dtrace at freebsd.org; 'Julian Elischer'
> Subject: Re: DTrace script to trace processes entering vfs::vop_remove
> 
> On Wed, Dec 03, 2014 at 03:19:31PM -0800, dteske at FreeBSD.org wrote:
> > Hi markj, list,
> >
> > I wrote a script for $work to help me find out "who on Earth
> > keeps deleting files XYZ?" from a particular storage server.
> >
> > Please find attached a copy of watch_vop_remove.d which
> > has the following sample output:
> >
> > 2014 Dec  3 11:58:52 rm[75596]: /tmp/foo
> >  -+= 72846 0.0 -bash
> >   \-+= 75589 0.0 /bin/bash /usr/home/support/bash_script
> >     \-+= 75596 0.0 rm -f /tmp/foo
> >
> > The above sample output was displayed when executing the following shell
> > script:
> >
> > #!/bin/bash
> > touch /tmp/foo
> > rm -f /tmp/foo
> >
> > The output format displayed for each vop_remove() call is as follows:
> >
> > DATE process[PID]: PATH_TO_DELETE
> >  -+= GPID UID.GID grandparent_process [arguments (up to 3)]
> >   \-+= PPID UID.GID parent_process [arguments (up to 3)]
> >     \-+= PID UID.GID process [arguments (up to 3)]
> 
> This is neat. I just had a few comments:
> - You can use walltimestamp when printing the date and time, instead of
>   timestamp + blah.

I read that online as well, however:
walltimestamp appears to _always_ be zero.


> - It's possible to get the full argv of the current process with
>   curpsinfo->pr_psargs. It can be done for other processes too; see
>   /usr/lib/dtrace/psinfo.d. (This might not be true depending on the
>   FreeBSD version you're on.)

Thanks! I'll have a look.

> - Running this script with a make -j4 buildkernel causes dtrace to run
>   out of dynamic variable space.
> 

Any recommendation on how to fix that?

#pragma D option dynvarsize=what_exactly?
(16m causes a warning that it's lowering the dynamic variable memory)


> I'd really really like to fix name resolution so that we don't have to
> jump through so many hoops to write scripts like this, though. One
> approach is to do what Solaris does, which is keep a cached path in the
> vnode itself (v_path).
> 

Yes, that would be great. But perhaps not something we should
do solely for dtrace's benefit.
-- 
Devin