NDFREE(): Fix unlocking for LOCKPARENT|LOCKLEAF and ndp->ni_dvp == ndp->ni_vp

[kib]

Summary

NDFREE() calculates unlock_dvp after ndp->ni_vp is unlocked and zeroed out. This makes the comparision of ni_dvp with ni_vp always fail.

Move the calculation of unlock_dvp right after unlock_vp, so that the code sees correct ni_vp value.

[markj]

So if ni_dvp == ni_vp, does the caller hold two references on vp or one? It was not clear to me from some quick reading of the code.

[kib]

In D20304#437658, @markj wrote:

So if ni_dvp == ni_vp, does the caller hold two references on vp or one? It was not clear to me from some quick reading of the code.

dvp == vp means that this was either dot lookup, or dotdot with vp == rootvp. For dotdot case, lookup() itself does the following:

				ndp->ni_dvp = dp;
				ndp->ni_vp = dp;
				VREF(dp);

for the root directory case (also chroot, jail pseudo-roots).

For UFS, ufs_lookup() takes a reference on the dvp directly in case of dot, see ufs_lookup().

So vp holds two references, but In either case, the code does not recurse on lock.

[mckusick]

This change looks correct to me.

[pho]

I reproduced the problem and verified that the patch fixes it.
I ran all of the stress2 tests on both amd64 and i386.
No problems seen.