Re: [dpdk-dev] [PATCH] mbuf: move headers not fragmented check to checksum

[Olivier Matz <olivier.matz@6wind.com>]

Hi Andrew,

On Thu, Mar 28, 2019 at 08:04:31PM +0300, Andrew Rybchenko wrote:
> Ping? (I have a number of net/sfc patches which heavily depend on this
> one and must not be applied without this one)
> 
> Andrew.
> 
> On 2/19/19 9:30 AM, Andrew Rybchenko wrote:
> > rte_validate_tx_offload() is used in Tx prepare callbacks
> > (RTE_LIBRTE_ETHDEV_DEBUG only) to check Tx offloads consistency.
> > Requirement that packet headers should not be fragmented is not
> > documented and unclear where it comes from except
> > rte_net_intel_cksum_prepare() functions which relies on it.
> > 
> > It could be NIC vendor specific driver or hardware limitation, but,
> > if so, it should be documented and checked in corresponding Tx
> > prepare callbacks.
> > 
> > Signed-off-by: Andrew Rybchenko <arybchenko@solarflare.com>
> > Acked-by: Konstantin Ananyev <konstantin.ananyev@intel.com>
> > ---
> > Looks good to me, though extra-testing would be needed.
> > Konstantin Ananyev <konstantin.ananyev@intel.com>
> > 
> >   lib/librte_mbuf/rte_mbuf.h | 12 ------------
> >   lib/librte_net/rte_net.h   | 17 +++++++++++++++++
> >   2 files changed, 17 insertions(+), 12 deletions(-)
> > 
> > diff --git a/lib/librte_mbuf/rte_mbuf.h b/lib/librte_mbuf/rte_mbuf.h
> > index d961cca..73daa81 100644
> > --- a/lib/librte_mbuf/rte_mbuf.h
> > +++ b/lib/librte_mbuf/rte_mbuf.h
> > @@ -2257,23 +2257,11 @@ static inline int rte_pktmbuf_chain(struct rte_mbuf *head, struct rte_mbuf *tail
> >   rte_validate_tx_offload(const struct rte_mbuf *m)
> >   {
> >   	uint64_t ol_flags = m->ol_flags;
> > -	uint64_t inner_l3_offset = m->l2_len;
> >   	/* Does packet set any of available offloads? */
> >   	if (!(ol_flags & PKT_TX_OFFLOAD_MASK))
> >   		return 0;
> > -	if (ol_flags & PKT_TX_OUTER_IP_CKSUM)
> > -		/* NB: elaborating the addition like this instead of using
> > -		 *     += gives the result uint64_t type instead of int,
> > -		 *     avoiding compiler warnings on gcc 8.1 at least */
> > -		inner_l3_offset = inner_l3_offset + m->outer_l2_len +
> > -				  m->outer_l3_len;
> > -
> > -	/* Headers are fragmented */
> > -	if (rte_pktmbuf_data_len(m) < inner_l3_offset + m->l3_len + m->l4_len)
> > -		return -ENOTSUP;
> > -
> >   	/* IP checksum can be counted only for IPv4 packet */
> >   	if ((ol_flags & PKT_TX_IP_CKSUM) && (ol_flags & PKT_TX_IPV6))
> >   		return -EINVAL;
> > diff --git a/lib/librte_net/rte_net.h b/lib/librte_net/rte_net.h
> > index e59760a..bd75aea 100644
> > --- a/lib/librte_net/rte_net.h
> > +++ b/lib/librte_net/rte_net.h
> > @@ -118,10 +118,27 @@ uint32_t rte_net_get_ptype(const struct rte_mbuf *m,
> >   	struct udp_hdr *udp_hdr;
> >   	uint64_t inner_l3_offset = m->l2_len;
> > +	/*
> > +	 * Does packet set any of available offloads?
> > +	 * Mainly it is required to avoid fragmented headers check if
> > +	 * no offloads are requested.
> > +	 */
> > +	if (!(ol_flags & PKT_TX_OFFLOAD_MASK))
> > +		return 0;
> > +
> >   	if ((ol_flags & PKT_TX_OUTER_IP_CKSUM) ||
> >   		(ol_flags & PKT_TX_OUTER_IPV6))
> >   		inner_l3_offset += m->outer_l2_len + m->outer_l3_len;
> > +	/*
> > +	 * Check if headers are fragmented.
> > +	 * The check could be less strict depending on which offloads are
> > +	 * requested and headers to be used, but let's keep it simple.
> > +	 */
> > +	if (unlikely(rte_pktmbuf_data_len(m) <
> > +		     inner_l3_offset + m->l3_len + m->l4_len))
> > +		return -ENOTSUP;
> > +
> >   	if (ol_flags & PKT_TX_IPV4) {
> >   		ipv4_hdr = rte_pktmbuf_mtod_offset(m, struct ipv4_hdr *,
> >   				inner_l3_offset);
> 


To summarize, the previous code was in a generic part, only enabled if
RTE_LIBRTE_ETHDEV_DEBUG is set, and it is moved in an intel-specific part,
but always enabled. Am I correct?

So it may have a performance impact on intel NICs. Shouldn't it be under
a debug option?

Regards,
Olivier