Re: CVS commit: src/sys/kern

[Robert Elz <kre%munnari.OZ.AU@localhost>]

    Date:        Tue, 04 Oct 2022 10:09:35 -0400
    From:        Christos Zoulas <christos%zoulas.com@localhost>
    Message-ID:  <8dd220d16861eb3a890461bdf02d1796%zoulas.com@localhost>

  | I always forget the O_CLOEXEC is special
  | in that regard. I wish it was not, but it is difficult to fix.

POSIX is adding O_CLOFORK in the next version (no guarantees I remembered
the symbol name spelling correctly here) which will have essentially the
same (open time) semantics (similar long term sematics as well, just
applied at a different time).

I assume we will need to add that at some point or other.

  | The question is how to find the vnode?

Not really, I assume that part will be fairly easy (probably trivial),
I just didn't have the energy to go work it out when sending that mail.
We have the file descriptor, and I suspect the file* (need to check to
make sure the right one is immediately available, but we can get it from
the fd if not), we know it refers to a vnode (it came from vn_open()),
so getting the vnode* from the file* is not something difficult, I think.

  | Perhaps it is easiest to fail the open call if O_EXLOCK or
  | O_SHLOCK are specified in a cloning open?

That would be an option, and is better than just ignoring them, but
better still would be to make them work.

Since open_setfp() does nothing (much) when none of the relevant O_xxx
flags that it tests are set (the fd open flags, as distinct from the
fp ones), and the code calls VOP_UNLOCK(vp) after it, we know that vp
is intended to be locked when open_setfp() is called (further confirmed
as when any of the O_??LOCK flags is set, open_setfp() does a VOP_UNLOCK()
and later a vn_lock() (which I am guessing is the inverse).

Maybe all that's needed is a vn_lock() call (on the vp that we still need
to fetch) and then call open_setfp() ?   But this is all beyond what I
know enough about to be sure, particularly to avoid doing anything which
might deadlock, etc.

kre

ps: if this gets done properly, then special case code to handle O_NONBLOCK
(and O_NOSIGPIPE, ...) in cloning device drivers won't be needed either,
open_setfp() is where all of that is normally added to the file* for the
fd being returned, it was not "simply happening" because that call is
missing in the cloned device case.