[dpdk-dev] [PATCH 1/4] rte_mbuf: add rte_pktmbuf_coalesce

Kulasek, TomaszX tomaszx.kulasek at intel.com
Thu Dec 29 16:58:01 CET 2016


Hi Olivier,

> -----Original Message-----
> From: Olivier Matz [mailto:olivier.matz at 6wind.com]
> Sent: Friday, December 16, 2016 11:06
> To: Kulasek, TomaszX <tomaszx.kulasek at intel.com>
> Cc: dev at dpdk.org
> Subject: Re: [dpdk-dev] [PATCH 1/4] rte_mbuf: add rte_pktmbuf_coalesce
> 
> Hi Tomasz,
> 
> On Fri,  2 Dec 2016 18:07:43 +0100, Tomasz Kulasek
> <tomaszx.kulasek at intel.com> wrote:
> > This patch adds function rte_pktmbuf_coalesce to let crypto PMD
> > coalesce chained mbuf before crypto operation and extend their
> > capabilities to support segmented mbufs when device cannot handle
> > them natively.
> >
> >
> > Signed-off-by: Tomasz Kulasek <tomaszx.kulasek at intel.com>
> > ---
> >  lib/librte_mbuf/rte_mbuf.h |   34 ++++++++++++++++++++++++++++++++++
> >  1 file changed, 34 insertions(+)
> >
> > diff --git a/lib/librte_mbuf/rte_mbuf.h b/lib/librte_mbuf/rte_mbuf.h
> > index ead7c6e..f048681 100644
> > --- a/lib/librte_mbuf/rte_mbuf.h
> > +++ b/lib/librte_mbuf/rte_mbuf.h
> > @@ -1647,6 +1647,40 @@ static inline int rte_pktmbuf_chain(struct
> > rte_mbuf *head, struct rte_mbuf *tail }
> >
> >  /**
> > + * Coalesce data from mbuf to the continuous buffer.
> > + *
> > + * @param mbuf_dst
> > + *   Contiguous destination mbuf
> > + * @param mbuf_src
> > + *   Uncontiguous source mbuf
> > + *
> > + * @return
> > + *   - 0, on success
> > + *   - -EINVAL, on error
> > + */
> 
> I think the API should be clarified. In your case, it is expected that the
> destination mbuf is already filled with uninitialized data (i.e. that
> rte_pktmbuf_append() has been called).
> 
> We could wonder if a better API wouldn't be to allocate the dst mbuf in
> the function, call append()/prepend(), and do the copy.
> 
> Even better, we could have:
> 
>   int rte_pktmbuf_linearize(struct rte_mbuf *m)
> 
> It will reuse the same mbuf (maybe moving the data).
> 
> 
> > +
> > +#include <rte_hexdump.h>
> 
> This should be removed.
> 
> > +
> > +static inline int
> > +rte_pktmbuf_coalesce(struct rte_mbuf *mbuf_dst, struct rte_mbuf
> *mbuf_src) {
> 
> Source mbuf should be const.
> 
> > +	char *dst;
> > +
> > +	if (!rte_pktmbuf_is_contiguous(mbuf_dst) ||
> > +			rte_pktmbuf_data_len(mbuf_dst) >=
> > +			rte_pktmbuf_pkt_len(mbuf_src))
> > +		return -EINVAL;
> 
> Why >= ?
> 
> > +
> > +	dst = rte_pktmbuf_mtod(mbuf_dst, char *);
> > +
> > +	if (!__rte_pktmbuf_read(mbuf_src, 0, rte_pktmbuf_pkt_len(mbuf_src),
> > +			dst))
> 
> When a function returns a pointer, I think it is clearer to do:
>   if (func() == NULL)
> than:
>   if (!func())
> 
> 
> > +		return -EINVAL;
> > +
> > +	return 0;
> > +}
> > +
> > +/**
> >   * Dump an mbuf structure to a file.
> >   *
> >   * Dump all fields for the given packet mbuf and all its associated
> 
> 
> One more question, I don't see where this function is used in your
> patchset. What is your use-case?
> 
> Regards,
> Olivier

This function is needed for crypto-perf application: http://dpdk.org/dev/patchwork/patch/17492/ to compare performance of crypto operations on segmented mbufs, when scatter gather is or is not supported by crypto PMD. It will be introduced with v2.

When device doesn't support scatter-gather, we want to know an overhead of manual coalescing mbuf.


	struct rte_cryptodev_info dev_info;
	int linearize = 0;

	/* Check if source mbufs require coalescing */
	if (ctx->options->segments_nb > 1) {
		rte_cryptodev_info_get(ctx->dev_id, &dev_info);
		if ((dev_info.feature_flags &
				RTE_CRYPTODEV_FF_MBUF_SCATTER_GATHER) == 0)
			linearize = 1;
	}

	// ...
 
	if (linearize) {
		/* PMD doesn't support scatter-gather and source buffer
		 * is segmented.
		 * We need to linearize it before enqueuing.
		 */
		for (i = 0; i < burst_size; i++)
			rte_pktmbuf_linearize(ops[i]->sym->m_src);
	}

	/* Enqueue burst of ops on crypto device */
	ops_enqd = rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id,
			ops, burst_size);


I have checked this use case (in crypto-perf application), and you're right, using rte_pktmbuf_linearize() function here is better and more straightforward. I will change it in v2.

Thanks for suggestions
Tomasz.


More information about the dev mailing list