Re: [dpdk-dev] [PATCH v9 3/4] net/ixgbe: cleanup Tx buffers

[Ferruh Yigit <ferruh.yigit@intel.com>]

On 1/16/2020 2:47 PM, Ferruh Yigit wrote:
> On 1/13/2020 9:57 AM, Chenxu Di wrote:
>> Add support to the ixgbe driver for the API rte_eth_tx_done_cleanup
>> to force free consumed buffers on Tx ring.
>>
>> Signed-off-by: Chenxu Di <chenxux.di@intel.com>
>> ---
>>  drivers/net/ixgbe/ixgbe_ethdev.c |   2 +
>>  drivers/net/ixgbe/ixgbe_rxtx.c   | 109 +++++++++++++++++++++++++++++++
>>  drivers/net/ixgbe/ixgbe_rxtx.h   |   2 +-
>>  3 files changed, 112 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/net/ixgbe/ixgbe_ethdev.c b/drivers/net/ixgbe/ixgbe_ethdev.c
>> index 2c6fd0f13..75bdd391a 100644
>> --- a/drivers/net/ixgbe/ixgbe_ethdev.c
>> +++ b/drivers/net/ixgbe/ixgbe_ethdev.c
>> @@ -601,6 +601,7 @@ static const struct eth_dev_ops ixgbe_eth_dev_ops = {
>>  	.udp_tunnel_port_add  = ixgbe_dev_udp_tunnel_port_add,
>>  	.udp_tunnel_port_del  = ixgbe_dev_udp_tunnel_port_del,
>>  	.tm_ops_get           = ixgbe_tm_ops_get,
>> +	.tx_done_cleanup      = ixgbe_dev_tx_done_cleanup,
>>  };
>>  
>>  /*
>> @@ -649,6 +650,7 @@ static const struct eth_dev_ops ixgbevf_eth_dev_ops = {
>>  	.reta_query           = ixgbe_dev_rss_reta_query,
>>  	.rss_hash_update      = ixgbe_dev_rss_hash_update,
>>  	.rss_hash_conf_get    = ixgbe_dev_rss_hash_conf_get,
>> +	.tx_done_cleanup      = ixgbe_dev_tx_done_cleanup,
>>  };
>>  
>>  /* store statistics names and its offset in stats structure */
>> diff --git a/drivers/net/ixgbe/ixgbe_rxtx.c b/drivers/net/ixgbe/ixgbe_rxtx.c
>> index fa572d184..a2e85ed5b 100644
>> --- a/drivers/net/ixgbe/ixgbe_rxtx.c
>> +++ b/drivers/net/ixgbe/ixgbe_rxtx.c
>> @@ -2306,6 +2306,115 @@ ixgbe_tx_queue_release_mbufs(struct ixgbe_tx_queue *txq)
>>  	}
>>  }
>>  
>> +static int
>> +ixgbe_tx_done_cleanup_full(struct ixgbe_tx_queue *txq, uint32_t free_cnt)
>> +{
>> +	struct ixgbe_tx_entry *swr_ring = txq->sw_ring;
>> +	uint16_t i, tx_last, tx_id;
>> +	uint16_t nb_tx_free_last;
>> +	uint16_t nb_tx_to_clean;
>> +	uint32_t pkt_cnt;
>> +
>> +	/* Start free mbuf from the next of tx_tail */
>> +	tx_last = txq->tx_tail;
>> +	tx_id  = swr_ring[tx_last].next_id;
>> +
>> +	if (txq->nb_tx_free == 0 && ixgbe_xmit_cleanup(txq))
>> +		return 0;
>> +
>> +	nb_tx_to_clean = txq->nb_tx_free;
>> +	nb_tx_free_last = txq->nb_tx_free;
>> +	if (!free_cnt)
>> +		free_cnt = txq->nb_tx_desc;
>> +
>> +	/* Loop through swr_ring to count the amount of
>> +	 * freeable mubfs and packets.
>> +	 */
>> +	for (pkt_cnt = 0; pkt_cnt < free_cnt; ) {
>> +		for (i = 0; i < nb_tx_to_clean &&
>> +			pkt_cnt < free_cnt &&
>> +			tx_id != tx_last; i++) {
>> +			if (swr_ring[tx_id].mbuf != NULL) {
>> +				rte_pktmbuf_free_seg(swr_ring[tx_id].mbuf);
>> +				swr_ring[tx_id].mbuf = NULL;
>> +
>> +				/*
>> +				 * last segment in the packet,
>> +				 * increment packet count
>> +				 */
>> +				pkt_cnt += (swr_ring[tx_id].last_id == tx_id);
>> +			}
>> +
>> +			tx_id = swr_ring[tx_id].next_id;
>> +		}
>> +
>> +		if (txq->tx_rs_thresh > txq->nb_tx_desc -
>> +			txq->nb_tx_free || tx_id == tx_last)
>> +			break;
>> +
>> +		if (pkt_cnt < free_cnt) {
>> +			if (ixgbe_xmit_cleanup(txq))
>> +				break;
>> +
>> +			nb_tx_to_clean = txq->nb_tx_free - nb_tx_free_last;
>> +			nb_tx_free_last = txq->nb_tx_free;
>> +		}
>> +	}
>> +
>> +	return (int)pkt_cnt;
>> +}
>> +
>> +static int
>> +ixgbe_tx_done_cleanup_simple(struct ixgbe_tx_queue *txq,
>> +			uint32_t free_cnt)
>> +{
>> +	int i, n, cnt;
>> +
>> +	if (free_cnt == 0 || free_cnt > txq->nb_tx_desc)
>> +		free_cnt = txq->nb_tx_desc;
>> +
>> +	cnt = free_cnt - free_cnt % txq->tx_rs_thresh;
>> +
>> +	for (i = 0; i < cnt; i += n) {
>> +		if (txq->nb_tx_desc - txq->nb_tx_free < txq->tx_rs_thresh)
>> +			break;
>> +
>> +		n = ixgbe_tx_free_bufs(txq);
>> +
>> +		if (n == 0)
>> +			break;
>> +	}
>> +
>> +	return i;
>> +}
>> +
>> +static int
>> +ixgbe_tx_done_cleanup_vec(struct ixgbe_tx_queue *txq __rte_unused,
>> +			uint32_t free_cnt __rte_unused)
>> +{
>> +	return -ENOTSUP;
>> +}
>> +
>> +int
>> +ixgbe_dev_tx_done_cleanup(void *tx_queue, uint32_t free_cnt)
>> +{
>> +	struct ixgbe_tx_queue *txq = (struct ixgbe_tx_queue *)tx_queue;
>> +	if (txq->offloads == 0 &&
>> +#ifdef RTE_LIBRTE_SECURITY
>> +			!(txq->using_ipsec) &&
>> +#endif
>> +			txq->tx_rs_thresh >= RTE_PMD_IXGBE_TX_MAX_BURST)
>> +#ifdef RTE_IXGBE_INC_VECTOR
>> +		if (txq->tx_rs_thresh <= RTE_IXGBE_TX_MAX_FREE_BUF_SZ &&
>> +				(rte_eal_process_type() != RTE_PROC_PRIMARY ||
>> +					txq->sw_ring_v != NULL))
>> +			return ixgbe_tx_done_cleanup_vec(txq, free_cnt);
>> +#endif
>> +		return ixgbe_tx_done_cleanup_simple(txq, free_cnt);
>> +
>> +	return ixgbe_tx_done_cleanup_full(txq, free_cnt);
>> +}
>> +
> 
> Missing curly parantheses in the 'if' blocks are causing confusion on which
> return patch to take.
> 
> the above code is like this:
> if (txq->offloads == 0 && ...)
>   if (txq->tx_rs_thresh <= RTE_IXGBE_TX_MAX_FREE_BUF_SZ && ...)
>     return ixgbe_tx_done_cleanup_vec(txq, free_cnt);
>   return ixgbe_tx_done_cleanup_simple(txq, free_cnt); <----- [*]
> return ixgbe_tx_done_cleanup_full(txq, free_cnt);
> 
> It is not clear, and looks like wrong based on indentation, when to get the [*]
> path above.
> 
> I will add curly parantheses while merging.
> 
> Btw, why we need "#ifdef RTE_IXGBE_INC_VECTOR" here, can't we remove it?
> 

Since 'ixgbe_tx_done_cleanup_vec()' already implemented in this file, instead of
vector specific files, I am removing the ifdef.

So making changes [1] and function becomes [2]. Please validate it in next-net.

[1]
 diff --git a/drivers/net/ixgbe/ixgbe_rxtx.c b/drivers/net/ixgbe/ixgbe_rxtx.c
 index a2e85ed5b..f41dc13d5 100644
 --- a/drivers/net/ixgbe/ixgbe_rxtx.c
 +++ b/drivers/net/ixgbe/ixgbe_rxtx.c
 @@ -2403,14 +2403,15 @@ ixgbe_dev_tx_done_cleanup(void *tx_queue, uint32_t
free_cnt)
  #ifdef RTE_LIBRTE_SECURITY
                         !(txq->using_ipsec) &&
  #endif
 -                       txq->tx_rs_thresh >= RTE_PMD_IXGBE_TX_MAX_BURST)
 -#ifdef RTE_IXGBE_INC_VECTOR
 +                       txq->tx_rs_thresh >= RTE_PMD_IXGBE_TX_MAX_BURST) {
                 if (txq->tx_rs_thresh <= RTE_IXGBE_TX_MAX_FREE_BUF_SZ &&
                                 (rte_eal_process_type() != RTE_PROC_PRIMARY ||
 -                                       txq->sw_ring_v != NULL))
 +                                       txq->sw_ring_v != NULL)) {
                         return ixgbe_tx_done_cleanup_vec(txq, free_cnt);
 -#endif
 -               return ixgbe_tx_done_cleanup_simple(txq, free_cnt);
 +               } else {
 +                       return ixgbe_tx_done_cleanup_simple(txq, free_cnt);
 +               }
 +       }

         return ixgbe_tx_done_cleanup_full(txq, free_cnt);
  }


[2]
 int
 ixgbe_dev_tx_done_cleanup(void *tx_queue, uint32_t free_cnt)
 {
         struct ixgbe_tx_queue *txq = (struct ixgbe_tx_queue *)tx_queue;
         if (txq->offloads == 0 &&
 #ifdef RTE_LIBRTE_SECURITY
                         !(txq->using_ipsec) &&
 #endif
                         txq->tx_rs_thresh >= RTE_PMD_IXGBE_TX_MAX_BURST) {
                 if (txq->tx_rs_thresh <= RTE_IXGBE_TX_MAX_FREE_BUF_SZ &&
                                 (rte_eal_process_type() != RTE_PROC_PRIMARY ||
                                         txq->sw_ring_v != NULL)) {
                         return ixgbe_tx_done_cleanup_vec(txq, free_cnt);
                 } else {
                         return ixgbe_tx_done_cleanup_simple(txq, free_cnt);
                 }
         }

         return ixgbe_tx_done_cleanup_full(txq, free_cnt);
 }