autofs not mounting directories

autofs not mounting directories

[Noel Hunt]

I have been given reports by users that after a reboot,
cron jobs don't run, and also sudo su - ..., while not
failing, is not cd'ing to the user's home directory. In
the cron case, there were commands of the form

/a/b/cmd > /a/b/log 2>&1

(/a/b being an automount point) where one would expect
the modify time of /a/b/log to change even if the command
fails, but the files modification times were unchanged from
the last access.

I was also told that a cd from the shell to the directory
in question solves the problem.

The machines on which this is reported had no cron logging
which I have now remedied, so I can't try to reproduce this
until I can reboot these machines.

If `directory' is an automount mountpoint, is there some
difference between `cd directory', and say cron running
`/directory/bin/cmd', in the way that `directory' is
accessed? I understand the former will cause a chdir()
which will invoke autofs_root_lookup(); in the latter case,
I don't know what goes on.

Noel Hunt

Re: autofs not mounting directories

[Ian Kent]

On Tue, 16 Mar 2004, Noel Hunt wrote:

> 
> If `directory' is an automount mountpoint, is there some
> difference between `cd directory', and say cron running
> `/directory/bin/cmd', in the way that `directory' is
> accessed? I understand the former will cause a chdir()
> which will invoke autofs_root_lookup(); in the latter case,
> I don't know what goes on.

Yes.

Sounds like you may be suffering from the same problem that was reported 
by Michael Blandford. I'm mystified I haven't seen the problem before as 
it is so fundamental to proper operation.

I'll be looking to get a test group together from the list in a couple of 
days.

Ian

Re: autofs not mounting directories

[Robert Au]

On Wed, Mar 17, 2004 at 11:24:36AM +0800, Ian Kent wrote:
> On Tue, 16 Mar 2004, Noel Hunt wrote:
> > If `directory' is an automount mountpoint, is there some
> > difference between `cd directory', and say cron running
> > `/directory/bin/cmd', in the way that `directory' is
> > accessed? I understand the former will cause a chdir()
> > which will invoke autofs_root_lookup(); in the latter case,
> > I don't know what goes on.
> 
> Yes.
> 
> Sounds like you may be suffering from the same problem that was reported 
> by Michael Blandford. I'm mystified I haven't seen the problem before as 
> it is so fundamental to proper operation.

I don't see Michael's original report in the list archives. Did he send it
to the list?
 
> I'll be looking to get a test group together from the list in a couple of 
> days.

I think we are seeing something similar here; sporadically, listing a file
with ls will fail until a process changes working directory to that file's
directory. I'd like to help out in the testing. We are seeing this problem
with Red Hat 7.2 (kernel 2.4.7-10), autofs 4.1.0, and kernel module
autofs4-2.4-module-20031201.

-- 
Robert Au                     rau@lsil.com
                              LSI Logic

Re: autofs not mounting directories

[Michael Blandford]

Robert Au wrote:

>I think we are seeing something similar here; sporadically, listing a file
>with ls will fail until a process changes working directory to that file's
>directory. I'd like to help out in the testing. We are seeing this problem
>with Red Hat 7.2 (kernel 2.4.7-10), autofs 4.1.0, and kernel module
>autofs4-2.4-module-20031201.
>  
>
I could replicate the problem by cd'ing into a directory ( that caused 
it to get automounted ) and try to touch a file.  It would fail.  cd out 
of the directory and back in and it would now work.  I only see this 
problem on 4.1.x.

Michael

Re: autofs not mounting directories

[Ian Kent]

On Wed, 17 Mar 2004, Robert Au wrote:

> > 
> > Sounds like you may be suffering from the same problem that was reported 
> > by Michael Blandford. I'm mystified I haven't seen the problem before as 
> > it is so fundamental to proper operation.
> 
> I don't see Michael's original report in the list archives. Did he send it
> to the list?

No.

It was reported when he was testing changes he had requested for 4.1.1. 

>  
> > I'll be looking to get a test group together from the list in a couple of 
> > days.
> 
> I think we are seeing something similar here; sporadically, listing a file
> with ls will fail until a process changes working directory to that file's
> directory. I'd like to help out in the testing. We are seeing this problem
> with Red Hat 7.2 (kernel 2.4.7-10), autofs 4.1.0, and kernel module
> autofs4-2.4-module-20031201.

2.4.7. I wonder if the patch I'm creating against 2.4.20 will work for 
you.

Strange this has gone un-noticed for over 18 months and then suddenly 
reports are popping up everywhere. Maybe it's just the focus at the moment 
so everything seems to be caused by it.

Ian

Re: autofs not mounting directories

[Michael Blandford]

Ian Kent wrote:

>2.4.7. I wonder if the patch I'm creating against 2.4.20 will work for 
>you.
>
>Strange this has gone un-noticed for over 18 months and then suddenly 
>reports are popping up everywhere. Maybe it's just the focus at the moment 
>so everything seems to be caused by it.
>  
>

After we get it fixed for 2.4.20, I will take a stab at backporting it 
to 2.4.9 which will probably be close enough for 2.4.7

Michael

Re: autofs not mounting directories

[Ian Kent]

On Wed, 17 Mar 2004, Michael Blandford wrote:

> Ian Kent wrote:
> 
> >2.4.7. I wonder if the patch I'm creating against 2.4.20 will work for 
> >you.
> >
> >Strange this has gone un-noticed for over 18 months and then suddenly 
> >reports are popping up everywhere. Maybe it's just the focus at the moment 
> >so everything seems to be caused by it.
> >  
> >
> 
> After we get it fixed for 2.4.20, I will take a stab at backporting it 
> to 2.4.9 which will probably be close enough for 2.4.7
> 

Shouldn't be to bad. It needed practically no changes until about 2.4.18 
or 20.

I may have the patch done tonight.

One consequence of this is that when cd'ing to a directory we will have to 
trigger the mount. To do it any other way will mean, essentially, 
rewriting autofs4 from the ground up. Maybe not a bad idea, but not now.
At least that is my impression atm.


Ian

Re: autofs not mounting directories

[Noel Hunt]

> One consequence of this is that when cd'ing to a directory we will have to 
> trigger the mount.

Isn't this the normal behaviour?

Re: autofs not mounting directories

[Jim Carter]

On Thu, 18 Mar 2004, Ian Kent wrote:
> One consequence of this is that when cd'ing to a directory we will have to 
> trigger the mount. To do it any other way will mean, essentially, 
> rewriting autofs4 from the ground up....

Am I missing something here?  If your process "cd"s to a directory, you
aren't changing to the mount point, you're changing to the referent.  That
is only allowed if you have X permission, and the only way the local kernel
can know that is to first mount the referent, and subsequently stat the
mount-root.

So it doesn't bother me that "cd" triggers mounting, even though in 0.1% of
the cases the process might "cd" elsewhere immediately without looking at
any files within the mount-root.  In most cases the very next system call
will be either to list the directory content or to open a file already
known to be in that directory, using a relative path.  So the mount is
hardly ever wasted.

I wonder if "cd"ing to the mount point, then mounting something over it, 
might trigger subtle bugs.  Example: the (nonexistent) content of the mount 
point dir is cached, and when you enumerate the directory you get the 
cached nothingness rather than the content of the referent, which you 
aren't even "cd"d to.   Hmmm, does this sound familiar?

James F. Carter          Voice 310 825 2897    FAX 310 206 6673
UCLA-Mathnet;  6115 MSA; 405 Hilgard Ave.; Los Angeles, CA, USA  90095-1555
Email: jimc@math.ucla.edu    http://www.math.ucla.edu/~jimc (q.v. for PGP key)

Re: autofs not mounting directories

[Ian Kent]

On Thu, 18 Mar 2004, Jim Carter wrote:

> On Thu, 18 Mar 2004, Ian Kent wrote:
> > One consequence of this is that when cd'ing to a directory we will have to 
> > trigger the mount. To do it any other way will mean, essentially, 
> > rewriting autofs4 from the ground up....
> 
> Am I missing something here?  If your process "cd"s to a directory, you
> aren't changing to the mount point, you're changing to the referent.  That
> is only allowed if you have X permission, and the only way the local kernel
> can know that is to first mount the referent, and subsequently stat the
> mount-root.
> 
> So it doesn't bother me that "cd" triggers mounting, even though in 0.1% of
> the cases the process might "cd" elsewhere immediately without looking at
> any files within the mount-root.  In most cases the very next system call
> will be either to list the directory content or to open a file already
> known to be in that directory, using a relative path.  So the mount is
> hardly ever wasted.

Yes. It seemed like a fair policy to me.

I've stubonly resisted changing anything outside of the autofs4 module but 
in this case the increase in complexity for the module, if it can be done 
at all, is not worth it.

Ian

Re: autofs not mounting directories

[Ian Kent]

On Thu, 18 Mar 2004, Noel Hunt wrote:

> > One consequence of this is that when cd'ing to a directory we will have to 
> > trigger the mount.
> 
> Isn't this the normal behaviour?

Not previously.

Check my reply to Jim Carter for my comment on this.

Ian

Re: autofs not mounting directories

[mmarion]

On 17 Mar, Ian Kent wrote:

> Sounds like you may be suffering from the same problem that was reported 
> by Michael Blandford. I'm mystified I haven't seen the problem before as 
> it is so fundamental to proper operation.

Reading over this thread, it sounds like it might be related to something we
patched for a long time ago (way before you took over).  Jim Barbour here at
Qualcomm figured out a one line patch that fixed similar issues we were having
with our program map mounted trees.  He sent it to the old list way back in 
early '01 and noone seemed to care then.  

I still have to add this one line in for our direct maps (via program) to
work:

*** autofs-4.1.1-orig/daemon/automount.c	Thu Jan 29 08:01:22 2004
--- autofs-4.1.1/daemon/automount.c	Fri Mar  5 12:57:47 2004
***************
*** 1066,1071 ****
--- 1066,1072 ----
  			close(ap.ioctlfd);
  			close(ap.state_pipe[0]);
  			close(ap.state_pipe[1]);
+ 			chdir(ap.path);
  
  			err = ap.lookup->lookup_mount(ap.path,
  						      pkt->name, pkt->len,

This is the pre-release 4.1.1 btw.. in case it's slightly different in the
official 4.1.1.

-- 
Mike Marion-Unix SysAdmin/Staff Engineer-http://www.qualcomm.com
Alcohol and Calculus don't mix. Never drink and derive.
==> Another /. gem