svn commit: r334708 - head/sys/kern

[Konstantin Belousov]

On Fri, Jun 08, 2018 at 02:30:10PM -0400, Mark Johnston wrote:
> On Fri, Jun 08, 2018 at 08:37:55PM +0300, Konstantin Belousov wrote:
> > On Thu, Jun 07, 2018 at 11:02:29PM -0700, Ryan Libby wrote:
> > > On Thu, Jun 7, 2018 at 10:03 PM, Mateusz Guzik <mjguzik at gmail.com> wrote:
> > > > Checking it without any locks is perfectly valid in this case. It is done
> > > > after v_holdcnt gets bumped from a non-zero value. So at that time it
> > > > is at least two. Of course that result is stale as an arbitrary number of
> > > > other threads could have bumped and dropped the ref past that point.
> > > > The minimum value is 1 since we hold the ref. But this means the
> > > > vnode must not be on the free list and that's what the assertion is
> > > > verifying.
> > > >
> > > > The problem is indeed lack of ordering against the code clearing the
> > > > flag for the case where 2 threads to vhold and one does the 0->1
> > > > transition.
> > > >
> > > > That said, the fence is required for the assertion to work.
> > > >
> > > 
> > > Yeah, I agree with this logic.  What I mean is that reordering between
> > > v_holdcnt 0->1 and v_iflag is normally settled by the release and
> > > acquisition of the vnode interlock, which we are supposed to hold for
> > > v_*i*flag.  A quick scan seems to show all of the checks of VI_FREE that
> > > are not asserts do hold the vnode interlock.  So, I'm just saying that I
> > > don't think the possible reordering affects them.
> > But do we know that only VI_FREE checks are affected ?
> > 
> > My concern is that users of _vhold() rely on seeing up to date state of the
> > vnode, and VI_FREE is only an example of the problem.  Most likely, the
> > code which fetched the vnode pointer before _vhold() call, should guarantee
> > visibility.
> 
> Wouldn't this be a problem only if we permit lockless accesses of vnode
> state outside of _vhold() and other vnode subroutines? The current
> protocol requires that the interlock be held, and this synchronizes with
> code which performs 0->1 and 1->0 transitions of the hold count. If this
> requirement is relaxed in the future, then fences would indeed be
> needed.

I do not claim that my concern is a real problem. I stated it as a
thing to look at when deciding whether the fences should be added
(unconditionally ?).

If you argument is that the only current lock-less protocol for the
struct vnode state is the v_holdcnt transitions for > 1, then I can
agree with it.