[dpdk-dev] [PATCH 1/1] mbuf: move pool pointer in first half

[Thomas Monjalon]

09/11/2020 15:08, Jerin Jacob:
> On Mon, Nov 9, 2020 at 7:32 PM Thomas Monjalon <thomas at monjalon.net> wrote:
> > 09/11/2020 14:35, Jerin Jacob:
> > > Are you facing the issue with 32bit? Could you share the steps to
> > > reproduce and gcc version?
> >
> > Oh you're right, the issue was with 32-bit build,
> 
> Thanks
> 
> > sorry for the confusion.
> 
> >
> > > --- a/drivers/net/octeontx2/otx2_ethdev.c
> > > +++ b/drivers/net/octeontx2/otx2_ethdev.c
> > > @@ -749,7 +749,7 @@ nix_tx_offload_flags(struct rte_eth_dev *eth_dev)
> > >         RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, pkt_len) !=
> > >                          offsetof(struct rte_mbuf, ol_flags) + 12);
> > >         RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, tx_offload) !=
> > > -                        offsetof(struct rte_mbuf, pool) + 2 * sizeof(void *));
> > > +                        offsetof(struct rte_mbuf, pool) + 2 * sizeof(uint64_t));
> >
> > The actual "fix" is
> > offsetof(struct rte_mbuf, pool) + sizeof(uint64_t) + sizeof(void *)
> >
> > I don't understand the octeontx2 vector code.
> > Please check what is the impact of this offset change.
> 
> Tested the changes, No issue seen. All the expectation of vector code
> is expressed with RTE_BUILD_BUG_ON.
> 
> > BTW, is 32-bit build really supported with octeontx2?
> 
> No. I think, keeping assert as "sizeof(void *)"(Same as now) and remove build
> support for 32bit works too for octeontx2.

OK, I think it's better than tweaking RTE_BUILD_BUG_ON for
something not really supported.

> We will add it when really required.

Then I'm going to send a patch to disable octeontx2 drivers on 32-bit.

Note there is another build issue with octeontx2 drivers on CentOS/RHEL 7
with Arm GGC 4.8.