[v8,2/2] net/af_xdp: Refactor af_xdp_tx_zc

Both legs of the loop share the same logic: the common parts are about
reserving and filling both address and length into the description.

This is moved into reserve_and_fill().

Bugzilla ID: 1440
Suggested-by: Maryam Tahhan <mtahhan@redhat.com>
Signed-off-by: Ariel Otilibili <ariel.otilibili@6wind.com>
---
 drivers/net/af_xdp/rte_eth_af_xdp.c | 75 ++++++++++++++++-------------
 1 file changed, 41 insertions(+), 34 deletions(-)
  

On Thu,  6 Feb 2025 21:46:45 +0100
Ariel Otilibili <ariel.otilibili@6wind.com> wrote:

>  
> +static inline struct xdp_desc *
> +reserve_and_fill(struct pkt_tx_queue *txq, struct rte_mbuf *mbuf,
> +		 struct xsk_umem_info *umem, void **pkt_ptr)
> +{
> +	struct xdp_desc *desc = NULL;
> +	uint64_t addr, offset;
> +	uint32_t idx_tx;
> +
> +	if (!xsk_ring_prod__reserve(&txq->tx, 1, &idx_tx))
> +		goto out;
> +
> +	desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx);
> +	desc->len = mbuf->pkt_len;
> +
> +	addr = (uint64_t)mbuf - (uint64_t)umem->buffer
> +		- umem->mb_pool->header_size;

addr (and the cast of mbuf) should probably be uintptr_t since the
intent is to do calculations with pointers.
  

>
> <snip>
>

@@ -559,51 +587,30 @@ af_xdp_tx_zc(void *queue, struct rte_mbuf **bufs,
> uint16_t nb_pkts)
>                 mbuf = bufs[i];
>
>                 if (mbuf->pool == umem->mb_pool) {
> -                       if (!xsk_ring_prod__reserve(&txq->tx, 1, &idx_tx))
> {
> +                 if (!(desc = reserve_and_fill(txq, mbuf, umem, NULL))) {
>                                 kick_tx(txq, cq);
> -                               if (!xsk_ring_prod__reserve(&txq->tx, 1,
> -                                                           &idx_tx))
> +                               desc = reserve_and_fill(txq, mbuf, umem,
> NULL);
> +                               if (!desc)
>                                         goto out;
>                         }
> -                       desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx);
> -                       desc->len = mbuf->pkt_len;
> -                       addr = (uint64_t)mbuf - (uint64_t)umem->buffer -
> -                                       umem->mb_pool->header_size;
> -                       offset = rte_pktmbuf_mtod(mbuf, uint64_t) -
> -                                       (uint64_t)mbuf +
> -                                       umem->mb_pool->header_size;
> -                       offset = offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT;
> -                       desc->addr = addr | offset;
> +
>                         tx_bytes += desc->len;
>                         count++;
>                 } else {
> -                       struct rte_mbuf *local_mbuf =
> -                                       rte_pktmbuf_alloc(umem->mb_pool);
> -                       void *pkt;
> -
> -                       if (local_mbuf == NULL)
> +                       if (!(local_mbuf =
> rte_pktmbuf_alloc(umem->mb_pool)))
>                                 goto out;
>
> -                       if (!xsk_ring_prod__reserve(&txq->tx, 1, &idx_tx))
> {
> +                       desc = reserve_and_fill(txq, local_mbuf, umem,
> &pkt);
> +                       if (!desc) {
>                                 rte_pktmbuf_free(local_mbuf);
>                                 goto out;
>                         }
>
> -                       desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx);
> -                       desc->len = mbuf->pkt_len;
> -
> -                       addr = (uint64_t)local_mbuf -
> (uint64_t)umem->buffer -
> -                                       umem->mb_pool->header_size;
> -                       offset = rte_pktmbuf_mtod(local_mbuf, uint64_t) -
> -                                       (uint64_t)local_mbuf +
> -                                       umem->mb_pool->header_size;
> -                       pkt = xsk_umem__get_data(umem->buffer, addr +
> offset);
> -                       offset = offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT;
> -                       desc->addr = addr | offset;
> +                       desc->len = local_mbuf->pkt_len;
>

Sorry if my remarks were confusing, it was just missing from the previous
patch and it needs to be:
desc->len = mbuf->pkt_len;

We need to keep this the same as the original code. This is a scenario
where we need to copy the data from an mbuf that isn't in from the pool of
buffers allocated for the umem. So the desc->len needs to be set to that of
the (non umem) mbuf.

The other changes look good. Nearly there,

Thanks again

<snip>
  

On 06/02/2025 21:42, Stephen Hemminger wrote:
> On Thu,  6 Feb 2025 21:46:45 +0100
> Ariel Otilibili<ariel.otilibili@6wind.com> wrote:
>
>>   
>> +static inline struct xdp_desc *
>> +reserve_and_fill(struct pkt_tx_queue *txq, struct rte_mbuf *mbuf,
>> +		 struct xsk_umem_info *umem, void **pkt_ptr)
>> +{
>> +	struct xdp_desc *desc = NULL;
>> +	uint64_t addr, offset;
>> +	uint32_t idx_tx;
>> +
>> +	if (!xsk_ring_prod__reserve(&txq->tx, 1, &idx_tx))
>> +		goto out;
>> +
>> +	desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx);
>> +	desc->len = mbuf->pkt_len;
>> +
>> +	addr = (uint64_t)mbuf - (uint64_t)umem->buffer
>> +		- umem->mb_pool->header_size;
> addr (and the cast of mbuf) should probably be uintptr_t since the
> intent is to do calculations with pointers.
>
I think it's ok as we would end up casting it anyway for the `struct xdp_desc`

/* Rx/Tx descriptor */
struct xdp_desc {
     __u64 addr;
     __u32 len;
     __u32 options;
};
  

Hi Maryam, hi Stephen;

On Fri, Feb 7, 2025 at 10:14 AM Maryam Tahhan <mtahhan@redhat.com> wrote:

> Sorry if my remarks were confusing, it was just missing from the previous
> patch and it needs to be:
> desc->len = mbuf->pkt_len;
>
> We need to keep this the same as the original code. This is a scenario
> where we need to copy the data from an mbuf that isn't in from the pool of
> buffers allocated for the umem. So the desc->len needs to be set to that of
> the (non umem) mbuf.
>
> The other changes look good. Nearly there,
>
> Thanks again
>

Thanks for the feedback. There is it,
http://inbox.dpdk.org/dev/20250207104552.1663519-1-ariel.otilibili@6wind.com/

>
> <snip>
>