[dpdk-dev] Could DPDK ring provide peek interface or any comments for this function?

Yan, Liming (Nokia - CN/Hangzhou) liming.yan at nokia.com
Wed Aug 5 05:08:44 CEST 2015


> -----Original Message-----
> From: ext Olivier MATZ [mailto:olivier.matz at 6wind.com]
> Sent: Tuesday, August 04, 2015 8:57 PM
> To: Yan, Liming (Nokia - CN/Hangzhou); dev at dpdk.org
> Subject: Re: [dpdk-dev] Could DPDK ring provide peek interface or any
> comments for this function?
> 
> Hi,
> 
> 
> On 08/04/2015 04:20 AM, Yan, Liming (Nokia - CN/Hangzhou) wrote:
> > Hi,
> >     As we see, DPDK ring has mainly the enqueue/dequeue APIs for
> SINGLE/MULTI producer/consumer.  I have a requirement in work to take a
> peek on the top object in the ring. I don't want to consume it, just
> check the reference of the object.  How could this be supported?   Or
> any alternative solution for my requirement?  Thanks.
> >     I wrote my own simple peek interface but it's not MC-safe. I
> think it's better if DPDK can support it officially.
> 
> I think it's a good idea. I think we could avoid a test by using
> the same naming convention than other functions applying to single
> consumer rings.
> 
> I'm not sure the function is safe on mc rings: the pointer returned
> via peek() should not be accessed while another core dequeues the
> object and modify/free it.
> 
[limiyan wrote] 
Seems it's good for sc-ring.  Could this be promoted into formal DPDK?  
For mc-ring peek, at least I think it's quite complicated even if we use something like ref_count and delayed free.

> 
> A simplified version would be something like that:
> 
> 
> /**
>   * Peek one object from a single-consumer ring
>   *
>   * The returned object is not dequeued. This function does not work
>   * on a multi-consumers ring.
>   *
>   * @param r
>   *   A pointer to the ring structure.
>   * @param obj_p
>   *   A pointer to a void * pointer (object) that will be filled.
>   * @return
>   *   - 0: Success; objects dequeued.
>   *   - -ENOENT: Not enough entries in the ring to peek
>   */
> static inline int
> rte_ring_sc_peek(struct rte_ring *r, void **obj_p)
> {
> 	if (unlikely(rte_ring_empty(r)))
> 		return -ENOENT;
> 	*obj_p = r->ring[r->cons.head & r->prod.mask];
> 	return 0;
> }
> 
> 
> Regards,
> Olivier
> 
> 
> >
> >
> >   /**
> > + * Just take a peek on the top object from a ring but not consume it.
> > + * Note: This interface is MC and multi-thread not safe.
> > + *       It can only be used for ring with RING_F_SC_DEQ attribute.
> > + *
> > + * @param r
> > + *   A pointer to the ring structure.
> > + * @param obj_p
> > + *   A pointer to a void * pointer (object) that will be filled.
> > + * @return
> > + *   - 0: Success, object is peeked.
> > + *   - -ENOENT: Not entries in the ring.
> > + *   - - EPERM: Operation not permitted/supported
> > + */
> > +static inline int __attribute__((always_inline))
> > +rte_ring_peek(const struct rte_ring *r, void **obj_p)
> > +{
> > +	uint32_t cons_head, prod_tail;
> > +
> > +	if (r->cons.sc_dequeue) {
> > +		cons_head = r->cons.head;
> > +		prod_tail = r->prod.tail;
> > +
> > +		if (prod_tail - cons_head == 0) {
> > +			return -ENOENT;
> > +		}
> > +		*obj_p = r->ring[cons_head & r->prod.mask];
> > +	} else {
> > +		return -EPERM;
> > +	}
> > +	return 0;
> > +}
> > +
> > +/**
> >
> >
> > --------^_^--------
> > Best Regards
> > Yan Limin
> >
> >



More information about the dev mailing list