Re: [PATCH 3/9 v1 RFC] skbuff: replace sock_zerocopy_put() with skb_zcopy_put()

[Willem de Bruijn <willemdebruijn.kernel@gmail.com>]

> > > All uargs should have a callback function, (unless nouarg
> > > is set), so push all special case logic handling down into
> > > the callbacks.  This slightly pessimizes the refcounted cases,
> >
> > What does this mean?
>
> The current zerocopy_put() code does:
>   1) if uarg, dec refcount, if refcount == 0:
>      if callback, run callback, else consume skb.
>
> This is called from the main TCP/UDP send path.  These would be called
> for the zctap case as well, so it should be made generic - not specific
> to the current zerocopy implementation.  The patch changes this into:
>
>   1) if uarg, run callback.
>
> Then, the msg_zerocopy code does:
>
>   1) save state,
>   2) dec refcount, run rest of callback on 0.
>
> Which is the same as before.  The !uarg case is never handled here.
> The zctap cases switch to their own callbacks.
>
>
> The current zerocopy clear code does:
>   1) if no_uarg, skip
>   2) if msg_zerocopy, save state, dec refcount, run callback when 0.
>   3) otherwise just run callback.
>   4) clear flags
>
> I would like to remove the msg_zerocopy specific logic from the function,
> so this becomes:
>
>   1) if uarg, run callback.
>   2) clear flags

That sounds fine. Especially since we can simplify the logic after the
commit I mentioned. I just didn't understand what you meant by
pessimize.

> > > -void sock_zerocopy_callback(struct ubuf_info *uarg, bool success)
> > > +static void __sock_zerocopy_callback(struct ubuf_info *uarg)
> > >  {
> > >         struct sk_buff *tail, *skb = skb_from_uarg(uarg);
> > >         struct sock_exterr_skb *serr;
> > > @@ -1222,7 +1222,7 @@ void sock_zerocopy_callback(struct ubuf_info *uarg, bool success)
> > >         serr->ee.ee_origin = SO_EE_ORIGIN_ZEROCOPY;
> > >         serr->ee.ee_data = hi;
> > >         serr->ee.ee_info = lo;
> > > -       if (!success)
> > > +       if (!uarg->zerocopy)
> > >                 serr->ee.ee_code |= SO_EE_CODE_ZEROCOPY_COPIED;
> > >
> > >         q = &sk->sk_error_queue;
> > > @@ -1241,18 +1241,15 @@ void sock_zerocopy_callback(struct ubuf_info *uarg, bool success)
> > >         consume_skb(skb);
> > >         sock_put(sk);
> > >  }
> > > -EXPORT_SYMBOL_GPL(sock_zerocopy_callback);
> > >
> > > -void sock_zerocopy_put(struct ubuf_info *uarg)
> > > +void sock_zerocopy_callback(struct ubuf_info *uarg, bool success)
> > >  {
> > > -       if (uarg && refcount_dec_and_test(&uarg->refcnt)) {
> > > -               if (uarg->callback)
> > > -                       uarg->callback(uarg, uarg->zerocopy);
> > > -               else
> > > -                       consume_skb(skb_from_uarg(uarg));
> >
> > I suppose this can be removed after commit 0a4a060bb204 ("sock: fix
> > zerocopy_success regression with msg_zerocopy"). Cleaning that up
> > would better be a separate patch that explains why the removal is
> > safe.
>
> I'll split the patches out.

Thanks. Yes, splitting that patch in two will help (me) follow it better.
>
> > It's also fine to bundle with moving refcount_dec_and_test into
> > sock_zerocopy_callback, which indeed follows from it.
> >
> > > -       }
> > > +       uarg->zerocopy = uarg->zerocopy & success;
> > > +
> > > +       if (refcount_dec_and_test(&uarg->refcnt))
> > > +               __sock_zerocopy_callback(uarg);
> >
> > This can be wrapped in existing sock_zerocopy_callback. No need for a
> > __sock_zerocopy_callback.
>
> The compiler will inline the helper anyway, since it's a single
> callsite.

True. I just don't think the wrapper adds much value here.