Re: [dpdk-dev] [PATCH] mbuf: extend pktmbuf pool private structure

[Shahaf Shuler <shahafs@mellanox.com>]

Wednesday, November 20, 2019 1:51 AM, Stephen Hemminger:
> Subject: Re: [dpdk-dev] [PATCH] mbuf: extend pktmbuf pool private
> structure
> 
> On Tue, 19 Nov 2019 23:30:15 +0100
> Thomas Monjalon <thomas@monjalon.net> wrote:
> 
> > 19/11/2019 17:25, Stephen Hemminger:
> > > On Tue, 19 Nov 2019 15:23:50 +0000
> > > Shahaf Shuler <shahafs@mellanox.com> wrote:
> > >
> > > > Tuesday, November 19, 2019 11:33 AM, Thomas Monjalon:
> > > > > Subject: Re: [PATCH] mbuf: extend pktmbuf pool private structure
> > > > >
> > > > > 18/11/2019 11:02, Shahaf Shuler:
> > > > > >  struct rte_pktmbuf_pool_private {
> > > > > >  	uint16_t mbuf_data_room_size; /**< Size of data space in
> each
> > > > > mbuf. */
> > > > > >  	uint16_t mbuf_priv_size;      /**< Size of private area in each
> mbuf.
> > > > > */
> > > > > > +	uint32_t reserved; /**< reserved for future use. */
> > > > >
> > > > > Maybe simpler to give the future name "flags" and keep the
> comment
> > > > > "reserved for future use".
> > > >
> > > > I'm am OK w/ changing to flags.
> > > > If Olivier accepts maybe you can change while applying?
> > >
> > > After the Linux openat experience if you want to add flags.
> > > Then all usage of API needs to validate that flags is 0.
> >
> > Sorry Stephen, I don't understand what you mean.
> > Please could you explain?
> >
> >
> 
> Any time a new field is added that maybe used later you can not guarantee
> that existing code correctly initializes the value to zero. What happened with
> openat() was that there was a flag value that was originally unused, but since
> kernel did not enforce that it was zero; it could not later be used for
> extensions.
> 
> You need to make sure that all reserved fields are initialized.
> That means when a private pool is created it is zeroed. And if a flag is new
> argument to an API, check for zero at create time.

I guess we can hard code the value for 0 on the rte_pktmbuf_pool_create function and have some assert on the rte_pktmbuf_pool_init callback (we cannot fail as this function returns void).
Any other places you find problematic? 

> 
> An example of how DPDK failed at this is the filter field in rte_pdump. Since it
> is not checked for NULL, it can't safely be used now (and still claim API/ABI
> compatiablity).