The vnode op should behave as described in vnode_if.src. For symlink, it
  is defined to return with the parent unlocked, not with the parent's state
  as set by the absense of LOCKPARENT. VOP_LOOKUP is the only call AFAIK
  that is defined to respect LOCKPARENT.

Good point; I should have realized that.

For sys_symlink, the componentname parameter is defined as IN, and
thus it seems it should not be modified.  So, while the patch I sent
seems OK, it would seem that consing up a fresh struct componentname
for the VOP_LOOKUP is better.  Either that or restoring the LOCKPARENT
flag after the call.  But I don't see it written anywhere that the
componentname struct is not modified :-)

  I think that the coda_symlink change needs to happen, and that sys_symlink
  is not wrong.
 
You have convinced me.  Should I be filing a PR, or are you or Jaromir
going to commit it (presumably to -current first, with a pullup
request to follow)?