svn commit: r274167 - in head/sys: kern sys

[Mateusz Guzik]

On Thu, Nov 20, 2014 at 09:24:37PM -0800, Rui Paulo wrote:
> On Nov 6, 2014, at 00:12, Mateusz Guzik <mjg at freebsd.org> wrote:
> > 
> > Author: mjg
> > Date: Thu Nov  6 08:12:34 2014
> > New Revision: 274167
> > URL: https://svnweb.freebsd.org/changeset/base/274167
> > 
> > Log:
> >  Add sysctl kern.proc.cwd
> > 
> >  It returns only current working directory of given process which saves a lot of
> >  overhead over kern.proc.filedesc if given proc has a lot of open fds.
> > 
> >  Submitted by:	Tiwei Bie <btw mail.ustc.edu.cn> (slightly modified)
> >  X-Additional:	JuniorJobs project
> > 
> > Modified:
> >  head/sys/kern/kern_descrip.c
> >  head/sys/sys/sysctl.h
> >  head/sys/sys/user.h
> > 
> > Modified: head/sys/kern/kern_descrip.c
> > ==============================================================================
> > --- head/sys/kern/kern_descrip.c	Thu Nov  6 07:44:10 2014	(r274166)
> > +++ head/sys/kern/kern_descrip.c	Thu Nov  6 08:12:34 2014	(r274167)
> > @@ -3406,6 +3406,73 @@ static SYSCTL_NODE(_kern_proc, KERN_PROC
> >     CTLFLAG_RD|CTLFLAG_MPSAFE, sysctl_kern_proc_filedesc,
> >     "Process filedesc entries");
> > 
> > +/*
> > + * Store a process current working directory information to sbuf.
> > + *
> > + * Takes a locked proc as argument, and returns with the proc unlocked.
> 
> I never really liked this approach because it makes it harder to see the locking pattern.
> 
> Why not call fdhold(p) in the sysctl handler, unlock the process, call kern_proc_cwd_out() with an additional fdp argument, and then call fddrop()?
> 

This code just mimics what was done with kern_proc_filedesc_out, which
happens to have semi-valid usecase when fdhold returns NULL.

In general I agree and this scheme is going away.

I have a patch in my backlog which "pins" the process (i.e. prevents exec
and exit) for the duration of kern_proc_filedesc_out/whatever, so it will
effectively mean such functions will no longer play with fdhold/fddrop.

Interestingly enough this requires some changes in devfs and I didn't have
the time to ensure they work properly.

-- 
Mateusz Guzik <mjguzik gmail.com>