Re: [RFC net-next] netmem: replace __netmem_clear_lsb() with netmem_to_nmdesc()

[Byungchul Park <byungchul@sk.com>]

On Mon, Jul 28, 2025 at 07:58:13PM +0100, Pavel Begunkov wrote:
> On 7/28/25 19:46, Pavel Begunkov wrote:
> > On 7/28/25 18:44, Mina Almasry wrote:
> > > On Sun, Jul 27, 2025 at 9:21 PM Byungchul Park <byungchul@sk.com> wrote:
> > > > 
> > 
> > ...>> - * Return: the netmem_ref cast to net_iov* regardless of its underlying type.
> > > > + * Return: the pointer to struct netmem_desc * regardless of its
> > > > + * underlying type.
> > > >    */
> > > > -static inline struct net_iov *__netmem_clear_lsb(netmem_ref netmem)
> > > > +static inline struct netmem_desc *netmem_to_nmdesc(netmem_ref netmem)
> > > >   {
> > > > -       return (struct net_iov *)((__force unsigned long)netmem & ~NET_IOV);
> > > > +       if (netmem_is_net_iov(netmem))
> > > > +               return &((struct net_iov *)((__force unsigned long)netmem &
> > > > +                                           ~NET_IOV))->desc;
> > > > +
> > > > +       return __netmem_to_nmdesc(netmem);
> > > 
> > > The if statement generates overhead. I'd rather avoid it. We can
> > > implement netmem_to_nmdesc like this, no?
> > > 
> > > netmem_to_nmdesc(netmem_ref netmem)
> > > {
> > >    return (struct netmem_desc)((__force unsigned long)netmem & ~NET_IOV);
> > > }
> > > 
> > > Because netmem_desc is the first element in both net_iov and page for
> > > the moment. (yes I know that will change eventually, but we don't have
> > > to incur overhead of an extra if statement until netmem_desc is
> > > removed from page, right?)
> > 
> > Same concern, but I think the goal here should be to make enough
> 
> s/make/give/
> 
> 
> > info to the compiler to optimise it out without assumptions on
> > the layouts nor NET_IOV_ASSERT_OFFSET. Currently it's not so bad,
> > but we should be able to remove this test+cmove.
> > 
> >      movq    %rdi, %rax    # netmem, tmp105
> >      andq    $-2, %rax    #, tmp105
> >      testb    $1, %dil    #, netmem
> >      cmove    %rdi, %rax    # tmp105,, netmem, <retval>
> >      jmp    __x86_return_thunk
> 
> struct netmem_desc *netmem_to_nmdesc(netmem_ref netmem)
> {
>        void *p = (void *)((__force unsigned long)netmem & ~NET_IOV);
> 
>        if (netmem_is_net_iov(netmem))
>                return &((struct net_iov *)p)->desc;
>        return __pp_page_to_nmdesc((struct page *)p);
> }

I wanted to remove constraints that can be removed, but Mina want not to
add additional overhead more.  So I'm thinking to keep the constraint,
'netmem_desc is the first member of net_iov'.

Thoughts?

	Byungchul

> movq    %rdi, %rax      # netmem, netmem
> andq    $-2, %rax       #, netmem
> jmp     __x86_return_thunk
> 
> 
> This should do it, and if the layouts change, it'd still
> remain correct.
> 
> --
> Pavel Begunkov