Re: procfs difference between NetBSD and Linux

[David Holland <dholland%netbsd.org@localhost>]

On Sat, Jun 05, 2021 at 11:03:05PM +0700, Robert Elz wrote:
 >   | It applies, compiled, and builds a release with no problems, running
 >   | tests now.
 > 
 > Unfortunately, it doesn't work, kernel segv in vn_open().

If it had worked first go I would have been very nervous :-)

 > I believe the cause is this code (in namei()):
 > 
 >                 if (cnp->cn_nameiop != LOOKUP &&
 >                     (searchdir == NULL ||
 >                      searchdir->v_mount != foundobj->v_mount)) {
 >                         if (searchdir) {
 > 				/*... irrelevant for now */
 >                         }
 >                         vrele(foundobj);
 >                         foundobj = NULL;
 >                         ndp->ni_dvp = NULL;
 >                         ndp->ni_vp = NULL;
 >      			state->attempt_retry = 1;

Right, it would need to keep foundobj on that path. Missed tht, I
guess...

 > I am trying a fix for this by making the initial test shown above be:
 > 
 >                 if (cnp->cn_nameiop != LOOKUP &&
 >                     (cnp->cn_flags & NONEXCLHACK) == 0 &&
 >                     (searchdir == NULL ||
 >                      searchdir->v_mount != foundobj->v_mount)) {
 > 
 > which of course then makes the test of NONEXCLHACK inside "case CREATE:"
 > meaningless, but harmless, so I just left that for now. 

Yeah, but please revert it before committing.

 > I added
 > 
 >                 if (foundobj != NULL && cnp->cn_flags & NONEXCLHACK) {
 >                         if (searchdir != NULL) { 
 >                                 if (searchdir_locked) {
 >                                         VOP_UNLOCK(searchdir);
 >                                         searchdir_locked = false;
 >                                 }
 >                                 vrele(searchdir);
 >                         }
 >                         searchdir = NULL;
 >                 }
 > 
 > which might be overly complicated, but seems to fit with what is needed
 > (or done anyway) in what comes later when searchdir != NULL.
 > (searchdir is later placed into ndp->ni_dvp).

It's not complicated enough, I think it probably needs a test that
foundobj != searchdir. :-|

also slap a comment on that that it's only actually needed for the
case where foundobj is a mount point, or maybe explicitly include that
test, or otherwise we'll forget the context and a few years down the
line be confused about what's actually going on.

-- 
David A. Holland
dholland%netbsd.org@localhost