Re: [PATCHv2 1/1] IB/rxe: optimize the recv process

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Sun, Mar 18, 2018 at 11:20:52PM -0400, Zhu Yanjun wrote:
> In mcast recv process, the function skb_clone is used. In fact,
> the refcount can be increased to replace cloning a new skb since
> the original skb will not be modified before it is freed.

Even in the recv flow?

> 
> This can make the performance better and save the memory.
> 
> CC: Srinivas Eeda <srinivas.eeda@xxxxxxxxxx>
> CC: Junxiao Bi <junxiao.bi@xxxxxxxxxx>
> Signed-off-by: Zhu Yanjun <yanjun.zhu@xxxxxxxxxx>
> ---
> V1->V2: Following Yuval's advice, skb test is removed.
> ---
>  drivers/infiniband/sw/rxe/rxe_recv.c | 16 +++++++++-------
>  1 file changed, 9 insertions(+), 7 deletions(-)
> 
> diff --git a/drivers/infiniband/sw/rxe/rxe_recv.c b/drivers/infiniband/sw/rxe/rxe_recv.c
> index 4c3f899..e0e4202 100644
> --- a/drivers/infiniband/sw/rxe/rxe_recv.c
> +++ b/drivers/infiniband/sw/rxe/rxe_recv.c
> @@ -276,7 +276,7 @@ static void rxe_rcv_mcast_pkt(struct rxe_dev *rxe, struct sk_buff *skb)
>  {
>  	struct rxe_pkt_info *pkt = SKB_TO_PKT(skb);
>  	struct rxe_mc_grp *mcg;
> -	struct sk_buff *skb_copy;
> +	struct sk_buff *skb_reference;
>  	struct rxe_mc_elem *mce;
>  	struct rxe_qp *qp;
>  	union ib_gid dgid;
> @@ -309,16 +309,19 @@ static void rxe_rcv_mcast_pkt(struct rxe_dev *rxe, struct sk_buff *skb)
>  			continue;
>  
>  		/* if *not* the last qp in the list
> -		 * make a copy of the skb to post to the next qp
> +		 * make a reference of the skb to post to the next qp
>  		 */
> -		skb_copy = (mce->qp_list.next != &mcg->qp_list) ?
> -				skb_clone(skb, GFP_ATOMIC) : NULL;
> +		skb_reference = NULL;
> +		if (mce->qp_list.next != &mcg->qp_list) {
> +			skb_reference = skb;
> +			refcount_inc(&skb->users);
> +		}

My question from v1 remains, i don't see why we need this temp variable
(skb_clone or skb_reference).
As i see it (unless i'm wrong), code can be something like:

	if (mce->qp_list.next != &mcg->qp_list)
		refcount_inc(&skb->users);

Then do the recv and skb remains as it was since we didn't cloned, just inc
ref.

>  
>  		pkt->qp = qp;
>  		rxe_add_ref(qp);
>  		rxe_rcv_pkt(rxe, pkt, skb);
>  
> -		skb = skb_copy;
> +		skb = skb_reference;
>  		if (!skb)
>  			break;

Also, this check i believe covers only the case where skb_clone fails to
allocate memory since the exit criteria (loop on QP list) is enough and
because we no longer have skb_clone then skb **can never** be NULL.

>  	}
> @@ -328,8 +331,7 @@ static void rxe_rcv_mcast_pkt(struct rxe_dev *rxe, struct sk_buff *skb)
>  	rxe_drop_ref(mcg);	/* drop ref from rxe_pool_get_key. */
>  
>  err1:
> -	if (skb)
> -		kfree_skb(skb);
> +	kfree_skb(skb);
>  }
>  
>  static int rxe_match_dgid(struct rxe_dev *rxe, struct sk_buff *skb)
> -- 
> 2.7.4
> 
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html



[Index of Archives]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Photo]     [Yosemite News]     [Yosemite Photos]     [Linux Kernel]     [Linux SCSI]     [XFree86]

  Powered by Linux