[dpdk-dev] kni: optimizing the rte_kni_rx_burst

  The current implementation of rte_kni_rx_burst polls the fifo for buffers.
Irrespective of success or failure, it allocates the mbuf and try to put them into the alloc_q
if the buffers are not added to alloc_q, it frees them.
This waste lots of cpu cycles in allocating and freeing the buffers if alloc_q is full.

The logic has been changed to:
1. Initially allocand add buffer(burstsize) to alloc_q
2. Add buffers to alloc_q only when you are pulling out the buffers.

Signed-off-by: Hemant Agrawal <Hemant@freescale.com>
---
 lib/librte_kni/rte_kni.c |    8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)
  

Is there anyone interested in KNI to review this patch please?


2014-07-23 12:15, Hemant Agrawal:
> The current implementation of rte_kni_rx_burst polls the fifo for buffers.
> Irrespective of success or failure, it allocates the mbuf and try to put them into the alloc_q
> if the buffers are not added to alloc_q, it frees them.
> This waste lots of cpu cycles in allocating and freeing the buffers if alloc_q is full.
> 
> The logic has been changed to:
> 1. Initially allocand add buffer(burstsize) to alloc_q
> 2. Add buffers to alloc_q only when you are pulling out the buffers.
> 
> Signed-off-by: Hemant Agrawal <Hemant@freescale.com>
> ---
>  lib/librte_kni/rte_kni.c |    8 ++++++--
>  1 file changed, 6 insertions(+), 2 deletions(-)
> 
> diff --git a/lib/librte_kni/rte_kni.c b/lib/librte_kni/rte_kni.c
> index 76feef4..01e85f8 100644
> --- a/lib/librte_kni/rte_kni.c
> +++ b/lib/librte_kni/rte_kni.c
> @@ -263,6 +263,9 @@ rte_kni_alloc(struct rte_mempool *pktmbuf_pool,
>  
>  	ctx->in_use = 1;
>  
> +	/* Allocate mbufs and then put them into alloc_q */
> +	kni_allocate_mbufs(ctx);
> +
>  	return ctx;
>  
>  fail:
> @@ -369,8 +372,9 @@ rte_kni_rx_burst(struct rte_kni *kni, struct rte_mbuf **mbufs, unsigned num)
>  {
>  	unsigned ret = kni_fifo_get(kni->tx_q, (void **)mbufs, num);
>  
> -	/* Allocate mbufs and then put them into alloc_q */
> -	kni_allocate_mbufs(kni);
> +	/* If buffers removed, allocate mbufs and then put them into alloc_q */
> +	if(ret)
> +		kni_allocate_mbufs(kni);
>  
>  	return ret;
>  }
  

Ping

2014-11-11 23:58, Thomas Monjalon:
> Is there anyone interested in KNI to review this patch please?
> 
> 
> 2014-07-23 12:15, Hemant Agrawal:
> > The current implementation of rte_kni_rx_burst polls the fifo for buffers.
> > Irrespective of success or failure, it allocates the mbuf and try to put them into the alloc_q
> > if the buffers are not added to alloc_q, it frees them.
> > This waste lots of cpu cycles in allocating and freeing the buffers if alloc_q is full.
> > 
> > The logic has been changed to:
> > 1. Initially allocand add buffer(burstsize) to alloc_q
> > 2. Add buffers to alloc_q only when you are pulling out the buffers.
> > 
> > Signed-off-by: Hemant Agrawal <Hemant@freescale.com>
> > ---
> >  lib/librte_kni/rte_kni.c |    8 ++++++--
> >  1 file changed, 6 insertions(+), 2 deletions(-)
> > 
> > diff --git a/lib/librte_kni/rte_kni.c b/lib/librte_kni/rte_kni.c
> > index 76feef4..01e85f8 100644
> > --- a/lib/librte_kni/rte_kni.c
> > +++ b/lib/librte_kni/rte_kni.c
> > @@ -263,6 +263,9 @@ rte_kni_alloc(struct rte_mempool *pktmbuf_pool,
> >  
> >  	ctx->in_use = 1;
> >  
> > +	/* Allocate mbufs and then put them into alloc_q */
> > +	kni_allocate_mbufs(ctx);
> > +
> >  	return ctx;
> >  
> >  fail:
> > @@ -369,8 +372,9 @@ rte_kni_rx_burst(struct rte_kni *kni, struct rte_mbuf **mbufs, unsigned num)
> >  {
> >  	unsigned ret = kni_fifo_get(kni->tx_q, (void **)mbufs, num);
> >  
> > -	/* Allocate mbufs and then put them into alloc_q */
> > -	kni_allocate_mbufs(kni);
> > +	/* If buffers removed, allocate mbufs and then put them into alloc_q */
> > +	if(ret)
> > +		kni_allocate_mbufs(kni);
> >  
> >  	return ret;
> >  }
  

2014-11-26 22:20, Thomas Monjalon:
> Ping
> 
> 2014-11-11 23:58, Thomas Monjalon:
> > Is there anyone interested in KNI to review this patch please?
> > 
> > 
> > 2014-07-23 12:15, Hemant Agrawal:
> > > The current implementation of rte_kni_rx_burst polls the fifo for buffers.
> > > Irrespective of success or failure, it allocates the mbuf and try to put them into the alloc_q
> > > if the buffers are not added to alloc_q, it frees them.
> > > This waste lots of cpu cycles in allocating and freeing the buffers if alloc_q is full.
> > > 
> > > The logic has been changed to:
> > > 1. Initially allocand add buffer(burstsize) to alloc_q
> > > 2. Add buffers to alloc_q only when you are pulling out the buffers.
> > > 
> > > Signed-off-by: Hemant Agrawal <Hemant@freescale.com>

From http://dpdk.org/ml/archives/dev/2015-January/011771.html, Jay said
"The patch looks good from a DPDK 1.6r2 viewpoint. We saw the same behavior
in our app and ended up avoiding it higher in the stack (in our code)."
Reviewed-by: Jay Rolette <rolette@infiniteio.com>
  

> > > 2014-07-23 12:15, Hemant Agrawal:
> > > > The current implementation of rte_kni_rx_burst polls the fifo for buffers.
> > > > Irrespective of success or failure, it allocates the mbuf and try to put them into the alloc_q
> > > > if the buffers are not added to alloc_q, it frees them.
> > > > This waste lots of cpu cycles in allocating and freeing the buffers if alloc_q is full.
> > > > 
> > > > The logic has been changed to:
> > > > 1. Initially allocand add buffer(burstsize) to alloc_q
> > > > 2. Add buffers to alloc_q only when you are pulling out the buffers.
> > > > 
> > > > Signed-off-by: Hemant Agrawal <Hemant@freescale.com>
> 
> From http://dpdk.org/ml/archives/dev/2015-January/011771.html, Jay said
> "The patch looks good from a DPDK 1.6r2 viewpoint. We saw the same behavior
> in our app and ended up avoiding it higher in the stack (in our code)."
> Reviewed-by: Jay Rolette <rolette@infiniteio.com>

Applied, thanks