Re: [dpdk-dev] [RFC PATCH] ixgbe: ixgbe_recv_pkts_vec shouldn't override mbuf buffer length

["Ananyev, Konstantin" <konstantin.ananyev@intel.com>]

> -----Original Message-----
> From: Jean-Mickael Guerin [mailto:jean-mickael.guerin@6wind.com]
> Sent: Friday, December 05, 2014 4:59 PM
> To: Ananyev, Konstantin
> Cc: dev@dpdk.org
> Subject: Re: [dpdk-dev] [RFC PATCH] ixgbe: ixgbe_recv_pkts_vec shouldn't override mbuf buffer length
> 
> On 05/12/2014 16:20, Konstantin Ananyev wrote:
> > That's an alternative way to fix the problem described in the patch:
> > http://dpdk.org/ml/archives/dev/2014-December/009394.html.
> > The main difference is:
> > - move buf_len fields out of rearm_data marker.
> > - make ixgbe_recv_pkts_vec() not touch buf_len field at all
> > (as all other RX functions behave).
> >
> > Signed-off-by: Konstantin Ananyev <konstantin.ananyev@intel.com>
> > ---
> >   lib/librte_mbuf/rte_mbuf.h            |  7 +++++--
> >   lib/librte_pmd_ixgbe/ixgbe_rxtx_vec.c | 20 +++++++++++++++-----
> >   2 files changed, 20 insertions(+), 7 deletions(-)
> >
> > diff --git a/lib/librte_mbuf/rte_mbuf.h b/lib/librte_mbuf/rte_mbuf.h
> > index 2e5fce5..bb88318 100644
> > --- a/lib/librte_mbuf/rte_mbuf.h
> > +++ b/lib/librte_mbuf/rte_mbuf.h
> > @@ -179,6 +179,8 @@ const char *rte_get_tx_ol_flag_name(uint64_t mask);
> >   typedef void    *MARKER[0];   /**< generic marker for a point in a structure */
> >   typedef uint64_t MARKER64[0]; /**< marker that allows us to overwrite 8 bytes
> >                                  * with a single assignment */
> > +typedef uint8_t MARKER8[0];   /**< generic marker with 1B alignment */
> > +
> >   /**
> >    * The generic rte_mbuf, containing a packet mbuf.
> >    */
> > @@ -188,9 +190,10 @@ struct rte_mbuf {
> >   	void *buf_addr;           /**< Virtual address of segment buffer. */
> >   	phys_addr_t buf_physaddr; /**< Physical address of segment buffer. */
> >
> > -	/* next 8 bytes are initialised on RX descriptor rearm */
> > -	MARKER64 rearm_data;
> >   	uint16_t buf_len;         /**< Length of segment buffer. */
> > +
> > +	/* next 6 bytes are initialised on RX descriptor rearm */
> > +	MARKER8 rearm_data;
> >   	uint16_t data_off;
> >
> >   	/**
> > diff --git a/lib/librte_pmd_ixgbe/ixgbe_rxtx_vec.c b/lib/librte_pmd_ixgbe/ixgbe_rxtx_vec.c
> > index 579bc46..d5fc0cc 100644
> > --- a/lib/librte_pmd_ixgbe/ixgbe_rxtx_vec.c
> > +++ b/lib/librte_pmd_ixgbe/ixgbe_rxtx_vec.c
> > @@ -79,13 +79,22 @@ ixgbe_rxq_rearm(struct igb_rx_queue *rxq)
> >   	/* Initialize the mbufs in vector, process 2 mbufs in one loop */
> >   	for (i = 0; i < RTE_IXGBE_RXQ_REARM_THRESH; i += 2, rxep += 2) {
> >   		__m128i vaddr0, vaddr1;
> > +		uintptr_t p0, p1;
> >
> >   		mb0 = rxep[0].mbuf;
> >   		mb1 = rxep[1].mbuf;
> >
> > -		/* flush mbuf with pkt template */
> > -		mb0->rearm_data[0] = rxq->mbuf_initializer;
> > -		mb1->rearm_data[0] = rxq->mbuf_initializer;
> > +		/*
> > +		 * Flush mbuf with pkt template.
> > +		 * Data to be rearmed is 6 bytes long.
> > +		 * Though, RX will overwrite ol_flags that are coming next
> > +		 * anyway. So overwrite whole 8 bytes with one load:
> > +		 * 6 bytes of rearm_data plus first 2 bytes of ol_flags.
> > +		 */
> > +		p0 = (uintptr_t)&mb0->rearm_data;
> > +		*(uint64_t *)p0 = rxq->mbuf_initializer;
> > +		p1 = (uintptr_t)&mb1->rearm_data;
> > +		*(uint64_t *)p1 = rxq->mbuf_initializer;
> >
> >   		/* load buf_addr(lo 64bit) and buf_physaddr(hi 64bit) */
> >   		vaddr0 = _mm_loadu_si128((__m128i *)&(mb0->buf_addr));
> > @@ -732,14 +741,15 @@ static struct ixgbe_txq_ops vec_txq_ops = {
> >   int
> >   ixgbe_rxq_vec_setup(struct igb_rx_queue *rxq)
> >   {
> > +	uintptr_t p;
> >   	struct rte_mbuf mb_def = { .buf_addr = 0 }; /* zeroed mbuf */
> >
> >   	mb_def.nb_segs = 1;
> >   	mb_def.data_off = RTE_PKTMBUF_HEADROOM;
> > -	mb_def.buf_len = rxq->mb_pool->elt_size - sizeof(struct rte_mbuf);
> >   	mb_def.port = rxq->port_id;
> >   	rte_mbuf_refcnt_set(&mb_def, 1);
> > -	rxq->mbuf_initializer = *((uint64_t *)&mb_def.rearm_data);
> > +	p = (uintptr_t)&mb_def.rearm_data;
> > +	rxq->mbuf_initializer = *(uint64_t *)p;
> >   	return 0;
> >   }
> >
> >
> 
> The patch introduces writes on unaligned data, but we can assume no
> performance penalty on intel hw, correct?
> 

Yes to both:
it introduces 64bit unaligned store.
I run performance test on IVB board, didn't see any degradation.
Konstantin