Re: [PATCH v4 4/8] NFSD add COPY_NOTIFY operation

[Olga Kornievskaia <olga.kornievskaia@gmail.com>]

On Tue, Jul 30, 2019 at 11:55 AM J. Bruce Fields <bfields@fieldses.org> wrote:
>
> On Tue, Jul 30, 2019 at 11:48:27AM -0400, Olga Kornievskaia wrote:
> > On Tue, Jul 23, 2019 at 4:46 PM J. Bruce Fields <bfields@fieldses.org> wrote:
> > >
> > > On Mon, Jul 22, 2019 at 04:17:44PM -0400, Olga Kornievskaia wrote:
> > > > Let me see if I understand your suspicion and ask for guidance how to
> > > > resolve it as perhaps I'm misusing the function. idr_alloc_cyclic()
> > > > keeps track of the structure of the 2nd arguments with a value it
> > > > returns. How do I initiate the structure with the value of the
> > > > function without knowing the value which can only be returned when I
> > > > call the function to add it to the list? what you are suggesting is to
> > > > somehow get the value for the new_id but not associate anything then
> > > > update the copy structure with that value and then call
> > > > idr_alloc_cyclic() (or something else) to create that association of
> > > > the new_id and the structure? I don't know how to do that.
> > >
> > > You could move the initialization under the s2s_cp_lock.  But there's
> > > additional initialization that's done in the caller.
> >
> > I still don't understand what you are looking for here and why. I'm
> > following what the normal stid allocation does.  There is no extra code
> > there to see if it initiated or not. nfs4_alloc_stid() calls
> > idr_alloc_cyclic() creates an association between the stid pointer and
> > at the time uninitialized nfs4_stid structure which is then filled in
> > with the return of the idr_alloc_cyclic(). That's exactly what the new
> > code is doing (well accept that i'll change it to store the
> > stateid_t).
>
> Yes, I'm a little worried about normal stid allocation too.  It's got
> one extra safeguard because of the check for 0 sc_type in the lookup,
> I haven't yet convinced myself that's enough.
>
> The race I'm worried about is: one task does the idr allocation and
> drops locks.  Before it has the chance to finish initializing the
> object, a second task looks it up in the idr and does something with it.
> It sees the not-yet-initialized fields.

Can the spin_lock() that we call before the idr_alloc_cyclic() be held
thru the initialization of the stid then? I'm just not sure what this
idr_preload_end() with a spin_lock but otherwise I don't see why we
can't and since idr_find() takes the same spin lock before the call,
it would solve the problem.

>
> --b.