Hi tech@,

Lately I've been trying to port nullfs from 4.4BSD-Lite2 back to
OpenBSD.

4.4BSD used `struct vnodeopv_*` to store and handle vnode operations,
while we have `struct vops`, so I'm translating each operation one
by one. Most of them are just calling the aliased vnode's VOP_*
operation, some are a bit trickier.

One thing I've been struggling with is vnode reference counting.
When syscalls call VOP_* on nullfs vnodes, their handling of
refcounting is not automatically passed through to the aliased
vnodes, so we have to replicate the behavior they expect before and
after calling VOP_* in each nullfs vop (in 4.4BSD, how each operation
handles refcounting is stored in the vnodeopv_* structures so this
can be done automatically, but obviously we can't do that).

At first I looked at VOP_LOOKUP(9) that describes how vnodes'
refcounting should be handled before and after calling each operation.
Then, out of curiosity, I tried to print some v_usecounts before
and after calling VOP_* operations inside syscalls (on a ufs), and
found some discrepancies with the information in VOP_LOOKUP(9). For
example, the call to VOP_MKDIR in domkdirat (sys/kern/vfs_syscalls.c)
decrements its dvp's v_usecount, which is not mentionned in the man
page. Even weirder, when doing the same thing for the VOP_LOOKUP
call inside vfs_lookup (sys/kern/vfs_lookup.c), sometimes dp's
v_usecount is decremented, sometimes it stays the same.

So I was wondering, is the man page up-to-date? If not, should I
test every VOP_* call inside syscalls and fix the docs? Also, does
someone understand how VOP_LOOKUP should handle refcounting?

Thanks