[dpdk-dev] [PATCH 1/2] net/sfc: free mbufs in bulks on EF10 native Tx datapath reap

[dpdk-dev] [PATCH 1/2] net/sfc: free mbufs in bulks on EF10 native Tx datapath reap

[Andrew Rybchenko]

From: Ivan Malov <ivan.malov@oktetlabs.ru>

Signed-off-by: Ivan Malov <ivan.malov@oktetlabs.ru>
Signed-off-by: Andrew Rybchenko <arybchenko@solarflare.com>
---
 drivers/net/sfc/sfc_ef10_tx.c | 48 ++++++++++++++++++++++++++++++++++++-------
 drivers/net/sfc/sfc_tweak.h   |  3 +++
 2 files changed, 44 insertions(+), 7 deletions(-)

diff --git a/drivers/net/sfc/sfc_ef10_tx.c b/drivers/net/sfc/sfc_ef10_tx.c
index 182fc23..5127a7a 100644
--- a/drivers/net/sfc/sfc_ef10_tx.c
+++ b/drivers/net/sfc/sfc_ef10_tx.c
@@ -158,17 +158,35 @@ struct sfc_ef10_txq {
 	pending += sfc_ef10_tx_process_events(txq);
 
 	if (pending != completed) {
+		struct rte_mbuf *bulk[SFC_TX_REAP_BULK_SIZE];
+		unsigned int nb = 0;
+
 		do {
 			struct sfc_ef10_tx_sw_desc *txd;
+			struct rte_mbuf *m;
 
 			txd = &txq->sw_ring[completed & ptr_mask];
+			if (txd->mbuf == NULL)
+				continue;
 
-			if (txd->mbuf != NULL) {
-				rte_pktmbuf_free(txd->mbuf);
-				txd->mbuf = NULL;
+			m = rte_pktmbuf_prefree_seg(txd->mbuf);
+			txd->mbuf = NULL;
+			if (m == NULL)
+				continue;
+
+			if ((nb == RTE_DIM(bulk)) ||
+			    ((nb != 0) && (m->pool != bulk[0]->pool))) {
+				rte_mempool_put_bulk(bulk[0]->pool,
+						     (void *)bulk, nb);
+				nb = 0;
 			}
+
+			bulk[nb++] = m;
 		} while (++completed != pending);
 
+		if (nb != 0)
+			rte_mempool_put_bulk(bulk[0]->pool, (void *)bulk, nb);
+
 		txq->completed = completed;
 	}
 
@@ -325,6 +343,7 @@ struct sfc_ef10_txq {
 		do {
 			phys_addr_t seg_addr = rte_mbuf_data_dma_addr(m_seg);
 			unsigned int seg_len = rte_pktmbuf_data_len(m_seg);
+			unsigned int id = added & ptr_mask;
 
 			SFC_ASSERT(seg_len <= SFC_EF10_TX_DMA_DESC_LEN_MAX);
 
@@ -332,15 +351,30 @@ struct sfc_ef10_txq {
 
 			sfc_ef10_tx_qdesc_dma_create(seg_addr,
 				seg_len, (pkt_len == 0),
-				&txq->txq_hw_ring[added & ptr_mask]);
+				&txq->txq_hw_ring[id]);
+
+			/*
+			 * rte_pktmbuf_free() is commonly used in DPDK for
+			 * recycling packets - the function checks every
+			 * segment's reference counter and returns the
+			 * buffer to its pool whenever possible;
+			 * nevertheless, freeing mbuf segments one by one
+			 * may entail some performance decline;
+			 * from this point, sfc_efx_tx_reap() does the same job
+			 * on its own and frees buffers in bulks (all mbufs
+			 * within a bulk belong to the same pool);
+			 * from this perspective, individual segment pointers
+			 * must be associated with the corresponding SW
+			 * descriptors independently so that only one loop
+			 * is sufficient on reap to inspect all the buffers
+			 */
+			txq->sw_ring[id].mbuf = m_seg;
+
 			++added;
 
 		} while ((m_seg = m_seg->next) != 0);
 
 		dma_desc_space -= (added - pkt_start);
-
-		/* Assign mbuf to the last used desc */
-		txq->sw_ring[(added - 1) & ptr_mask].mbuf = *pktp;
 	}
 
 	if (likely(added != txq->added)) {
diff --git a/drivers/net/sfc/sfc_tweak.h b/drivers/net/sfc/sfc_tweak.h
index 4ef7fc8..fd2f75c 100644
--- a/drivers/net/sfc/sfc_tweak.h
+++ b/drivers/net/sfc/sfc_tweak.h
@@ -53,4 +53,7 @@
 /** Default free threshold follows recommendations from DPDK documentation */
 #define SFC_TX_DEFAULT_FREE_THRESH	32
 
+/** Number of mbufs to be freed in bulk in a single call */
+#define SFC_TX_REAP_BULK_SIZE		32
+
 #endif /* _SFC_TWEAK_H_ */
-- 
1.8.2.3

[dpdk-dev] [PATCH 2/2] net/sfc: free mbufs in bulks on simple EF10 Tx datapath reap

[Andrew Rybchenko]

From: Ivan Malov <ivan.malov@oktetlabs.ru>

Signed-off-by: Ivan Malov <ivan.malov@oktetlabs.ru>
Signed-off-by: Andrew Rybchenko <arybchenko@solarflare.com>
---
 doc/guides/nics/sfc_efx.rst   |  4 +++-
 drivers/net/sfc/sfc_dp_tx.h   |  2 ++
 drivers/net/sfc/sfc_ef10_tx.c | 15 ++++++++++++++-
 drivers/net/sfc/sfc_ethdev.c  |  6 ++++++
 drivers/net/sfc/sfc_tx.c      | 17 +++++++++++++++++
 5 files changed, 42 insertions(+), 2 deletions(-)

diff --git a/doc/guides/nics/sfc_efx.rst b/doc/guides/nics/sfc_efx.rst
index 973a4a0..028b980 100644
--- a/doc/guides/nics/sfc_efx.rst
+++ b/doc/guides/nics/sfc_efx.rst
@@ -245,12 +245,14 @@ boolean parameters value.
   features available and required by the datapath implementation.
   **efx** chooses libefx-based datapath which supports VLAN insertion
   (full-feature firmware variant only), TSO and multi-segment mbufs.
+  Mbuf segments may come from different mempools, and mbuf reference
+  counters are treated responsibly.
   **ef10** chooses EF10 (SFN7xxx, SFN8xxx) native datapath which is
   more efficient than libefx-based but has no VLAN insertion and TSO
   support yet.
   **ef10_simple** chooses EF10 (SFN7xxx, SFN8xxx) native datapath which
   is even more faster then **ef10** but does not support multi-segment
-  mbufs.
+  mbufs, disallows multiple mempools and neglects mbuf reference counters.
 
 - ``perf_profile`` [auto|throughput|low-latency] (default **throughput**)
 
diff --git a/drivers/net/sfc/sfc_dp_tx.h b/drivers/net/sfc/sfc_dp_tx.h
index db2a70b..94d1b10 100644
--- a/drivers/net/sfc/sfc_dp_tx.h
+++ b/drivers/net/sfc/sfc_dp_tx.h
@@ -142,6 +142,8 @@ struct sfc_dp_tx {
 #define SFC_DP_TX_FEAT_TSO		0x2
 #define SFC_DP_TX_FEAT_MULTI_SEG	0x4
 #define SFC_DP_TX_FEAT_MULTI_PROCESS	0x8
+#define SFC_DP_TX_FEAT_MULTI_POOL	0x10
+#define SFC_DP_TX_FEAT_REFCNT		0x20
 	sfc_dp_tx_qcreate_t		*qcreate;
 	sfc_dp_tx_qdestroy_t		*qdestroy;
 	sfc_dp_tx_qstart_t		*qstart;
diff --git a/drivers/net/sfc/sfc_ef10_tx.c b/drivers/net/sfc/sfc_ef10_tx.c
index 5127a7a..9047b3e 100644
--- a/drivers/net/sfc/sfc_ef10_tx.c
+++ b/drivers/net/sfc/sfc_ef10_tx.c
@@ -401,14 +401,25 @@ struct sfc_ef10_txq {
 	pending += sfc_ef10_tx_process_events(txq);
 
 	if (pending != completed) {
+		struct rte_mbuf *bulk[SFC_TX_REAP_BULK_SIZE];
+		unsigned int nb = 0;
+
 		do {
 			struct sfc_ef10_tx_sw_desc *txd;
 
 			txd = &txq->sw_ring[completed & ptr_mask];
 
-			rte_pktmbuf_free_seg(txd->mbuf);
+			if (nb == RTE_DIM(bulk)) {
+				rte_mempool_put_bulk(bulk[0]->pool,
+						     (void *)bulk, nb);
+				nb = 0;
+			}
+
+			bulk[nb++] = txd->mbuf;
 		} while (++completed != pending);
 
+		rte_mempool_put_bulk(bulk[0]->pool, (void *)bulk, nb);
+
 		txq->completed = completed;
 	}
 
@@ -614,6 +625,8 @@ struct sfc_dp_tx sfc_ef10_tx = {
 		.hw_fw_caps	= SFC_DP_HW_FW_CAP_EF10,
 	},
 	.features		= SFC_DP_TX_FEAT_MULTI_SEG |
+				  SFC_DP_TX_FEAT_MULTI_POOL |
+				  SFC_DP_TX_FEAT_REFCNT |
 				  SFC_DP_TX_FEAT_MULTI_PROCESS,
 	.qcreate		= sfc_ef10_tx_qcreate,
 	.qdestroy		= sfc_ef10_tx_qdestroy,
diff --git a/drivers/net/sfc/sfc_ethdev.c b/drivers/net/sfc/sfc_ethdev.c
index 2b037d8..26e8c93 100644
--- a/drivers/net/sfc/sfc_ethdev.c
+++ b/drivers/net/sfc/sfc_ethdev.c
@@ -145,6 +145,12 @@
 	if (~sa->dp_tx->features & SFC_DP_TX_FEAT_MULTI_SEG)
 		dev_info->default_txconf.txq_flags |= ETH_TXQ_FLAGS_NOMULTSEGS;
 
+	if (~sa->dp_tx->features & SFC_DP_TX_FEAT_MULTI_POOL)
+		dev_info->default_txconf.txq_flags |= ETH_TXQ_FLAGS_NOMULTMEMP;
+
+	if (~sa->dp_tx->features & SFC_DP_TX_FEAT_REFCNT)
+		dev_info->default_txconf.txq_flags |= ETH_TXQ_FLAGS_NOREFCOUNT;
+
 #if EFSYS_OPT_RX_SCALE
 	if (sa->rss_support != EFX_RX_SCALE_UNAVAILABLE) {
 		dev_info->reta_size = EFX_RSS_TBL_SIZE;
diff --git a/drivers/net/sfc/sfc_tx.c b/drivers/net/sfc/sfc_tx.c
index bf59601..4ea7bd7 100644
--- a/drivers/net/sfc/sfc_tx.c
+++ b/drivers/net/sfc/sfc_tx.c
@@ -91,6 +91,21 @@
 		rc = EINVAL;
 	}
 
+	if (((flags & ETH_TXQ_FLAGS_NOMULTMEMP) == 0) &&
+	    (~sa->dp_tx->features & SFC_DP_TX_FEAT_MULTI_POOL)) {
+		sfc_err(sa, "multi-mempool is not supported by %s datapath",
+			sa->dp_tx->dp.name);
+		rc = EINVAL;
+	}
+
+	if (((flags & ETH_TXQ_FLAGS_NOREFCOUNT) == 0) &&
+	    (~sa->dp_tx->features & SFC_DP_TX_FEAT_REFCNT)) {
+		sfc_err(sa,
+			"mbuf reference counters are neglected by %s datapath",
+			sa->dp_tx->dp.name);
+		rc = EINVAL;
+	}
+
 	if ((flags & ETH_TXQ_FLAGS_NOVLANOFFL) == 0) {
 		if (!encp->enc_hw_tx_insert_vlan_enabled) {
 			sfc_err(sa, "VLAN offload is not supported");
@@ -1023,6 +1038,8 @@ struct sfc_dp_tx sfc_efx_tx = {
 	},
 	.features		= SFC_DP_TX_FEAT_VLAN_INSERT |
 				  SFC_DP_TX_FEAT_TSO |
+				  SFC_DP_TX_FEAT_MULTI_POOL |
+				  SFC_DP_TX_FEAT_REFCNT |
 				  SFC_DP_TX_FEAT_MULTI_SEG,
 	.qcreate		= sfc_efx_tx_qcreate,
 	.qdestroy		= sfc_efx_tx_qdestroy,
-- 
1.8.2.3

Re: [dpdk-dev] [PATCH 1/2] net/sfc: free mbufs in bulks on EF10 native Tx datapath reap

[Ferruh Yigit]

On 9/8/2017 3:15 PM, Andrew Rybchenko wrote:
> From: Ivan Malov <ivan.malov@oktetlabs.ru>
> 
> Signed-off-by: Ivan Malov <ivan.malov@oktetlabs.ru>
> Signed-off-by: Andrew Rybchenko <arybchenko@solarflare.com>
> ---
>  drivers/net/sfc/sfc_ef10_tx.c | 48 ++++++++++++++++++++++++++++++++++++-------
>  drivers/net/sfc/sfc_tweak.h   |  3 +++
>  2 files changed, 44 insertions(+), 7 deletions(-)
> 
> diff --git a/drivers/net/sfc/sfc_ef10_tx.c b/drivers/net/sfc/sfc_ef10_tx.c
> index 182fc23..5127a7a 100644
> --- a/drivers/net/sfc/sfc_ef10_tx.c
> +++ b/drivers/net/sfc/sfc_ef10_tx.c
> @@ -158,17 +158,35 @@ struct sfc_ef10_txq {
>  	pending += sfc_ef10_tx_process_events(txq);
>  
>  	if (pending != completed) {
> +		struct rte_mbuf *bulk[SFC_TX_REAP_BULK_SIZE];
> +		unsigned int nb = 0;
> +
>  		do {
>  			struct sfc_ef10_tx_sw_desc *txd;
> +			struct rte_mbuf *m;
>  
>  			txd = &txq->sw_ring[completed & ptr_mask];
> +			if (txd->mbuf == NULL)
> +				continue;
>  
> -			if (txd->mbuf != NULL) {
> -				rte_pktmbuf_free(txd->mbuf);
> -				txd->mbuf = NULL;
> +			m = rte_pktmbuf_prefree_seg(txd->mbuf);
> +			txd->mbuf = NULL;
> +			if (m == NULL)
> +				continue;
> +
> +			if ((nb == RTE_DIM(bulk)) ||
> +			    ((nb != 0) && (m->pool != bulk[0]->pool))) {

ICC is giving warning [1] here, as far as I can see this is false
positive but can you please double check in case I am missing something?

And unless if you don't see a way to convince icc without effecting
performance, would you mind updating patch to ignore warning [2] ?


[1]
error #3656: variable "bulk" may be used before its value is set

[2]
  diff --git a/drivers/net/sfc/Makefile b/drivers/net/sfc/Makefile
  index 57aa963ba..359314809 100644
  --- a/drivers/net/sfc/Makefile
  +++ b/drivers/net/sfc/Makefile
  @@ -65,6 +65,7 @@ CFLAGS += -Wbad-function-cast
   CFLAGS_BASE_DRIVER += -Wno-empty-body
   else ifeq ($(CONFIG_RTE_TOOLCHAIN_ICC),y)
   CFLAGS_BASE_DRIVER += -Wno-unused-but-set-variable
  +CFLAGS_sfc_ef10_tx.o += -wd3656
   endif

   #

> +				rte_mempool_put_bulk(bulk[0]->pool,
> +						     (void *)bulk, nb);
> +				nb = 0;
>  			}
> +
> +			bulk[nb++] = m;
>  		} while (++completed != pending);
>  
> +		if (nb != 0)
> +			rte_mempool_put_bulk(bulk[0]->pool, (void *)bulk, nb);
> +
>  		txq->completed = completed;
>  	}

<...>

Re: [dpdk-dev] [PATCH 2/2] net/sfc: free mbufs in bulks on simple EF10 Tx datapath reap

[Ferruh Yigit]

On 9/8/2017 3:15 PM, Andrew Rybchenko wrote:
> From: Ivan Malov <ivan.malov@oktetlabs.ru>
> 
> Signed-off-by: Ivan Malov <ivan.malov@oktetlabs.ru>
> Signed-off-by: Andrew Rybchenko <arybchenko@solarflare.com>
> ---
>  doc/guides/nics/sfc_efx.rst   |  4 +++-
>  drivers/net/sfc/sfc_dp_tx.h   |  2 ++
>  drivers/net/sfc/sfc_ef10_tx.c | 15 ++++++++++++++-
>  drivers/net/sfc/sfc_ethdev.c  |  6 ++++++
>  drivers/net/sfc/sfc_tx.c      | 17 +++++++++++++++++
>  5 files changed, 42 insertions(+), 2 deletions(-)
> 
> diff --git a/doc/guides/nics/sfc_efx.rst b/doc/guides/nics/sfc_efx.rst
> index 973a4a0..028b980 100644
> --- a/doc/guides/nics/sfc_efx.rst
> +++ b/doc/guides/nics/sfc_efx.rst
> @@ -245,12 +245,14 @@ boolean parameters value.
>    features available and required by the datapath implementation.
>    **efx** chooses libefx-based datapath which supports VLAN insertion
>    (full-feature firmware variant only), TSO and multi-segment mbufs.
> +  Mbuf segments may come from different mempools, and mbuf reference
> +  counters are treated responsibly.

This is also the case for ef10 native, right? Does it make sense to
document it in below too?

>    **ef10** chooses EF10 (SFN7xxx, SFN8xxx) native datapath which is
>    more efficient than libefx-based but has no VLAN insertion and TSO
>    support yet.
>    **ef10_simple** chooses EF10 (SFN7xxx, SFN8xxx) native datapath which
>    is even more faster then **ef10** but does not support multi-segment
> -  mbufs.
> +  mbufs, disallows multiple mempools and neglects mbuf reference counters.
>  
>  - ``perf_profile`` [auto|throughput|low-latency] (default **throughput**)
>  

<...>

> --- a/drivers/net/sfc/sfc_ef10_tx.c
> +++ b/drivers/net/sfc/sfc_ef10_tx.c
> @@ -401,14 +401,25 @@ struct sfc_ef10_txq {
>  	pending += sfc_ef10_tx_process_events(txq);
>  
>  	if (pending != completed) {
> +		struct rte_mbuf *bulk[SFC_TX_REAP_BULK_SIZE];
> +		unsigned int nb = 0;
> +
>  		do {
>  			struct sfc_ef10_tx_sw_desc *txd;
>  
>  			txd = &txq->sw_ring[completed & ptr_mask];
>  
> -			rte_pktmbuf_free_seg(txd->mbuf);
> +			if (nb == RTE_DIM(bulk)) {
> +				rte_mempool_put_bulk(bulk[0]->pool,
> +						     (void *)bulk, nb);

same warning here, again false positive I think:
error #3656: variable "bulk" may be used before its value is set

The patch to ignore the warning will take care of this one too.

> +				nb = 0;
> +			}
> +
> +			bulk[nb++] = txd->mbuf;
>  		} while (++completed != pending);
>  
> +		rte_mempool_put_bulk(bulk[0]->pool, (void *)bulk, nb);
> +
>  		txq->completed = completed;
>  	}

<...>

Re: [dpdk-dev] [PATCH 1/2] net/sfc: free mbufs in bulks on EF10 native Tx datapath reap

[Stephen Hemminger]

On Fri, 8 Sep 2017 15:15:50 +0100
Andrew Rybchenko <arybchenko@solarflare.com> wrote:

> +			txd->mbuf = NULL;
> +			if (m == NULL)
> +				continue;
> +
> +			if ((nb == RTE_DIM(bulk)) ||
> +			    ((nb != 0) && (m->pool != bulk[0]->pool))) {
> +				rte_mempool_put_bulk(bulk[0]->pool,
> +						     (void *)bulk, nb);
> +				nb = 0;
>  			}
> +

Why not add rte_mbuf_free_bulk (inline) to base code, rather than recoding
everywhere?

Re: [dpdk-dev] [PATCH 1/2] net/sfc: free mbufs in bulks on EF10 native Tx datapath reap

[Andrew Rybchenko]

On 09/13/2017 12:51 AM, Stephen Hemminger wrote:
> On Fri, 8 Sep 2017 15:15:50 +0100
> Andrew Rybchenko <arybchenko@solarflare.com> wrote:
>
>> +			txd->mbuf = NULL;
>> +			if (m == NULL)
>> +				continue;
>> +
>> +			if ((nb == RTE_DIM(bulk)) ||
>> +			    ((nb != 0) && (m->pool != bulk[0]->pool))) {
>> +				rte_mempool_put_bulk(bulk[0]->pool,
>> +						     (void *)bulk, nb);
>> +				nb = 0;
>>   			}
>> +
> Why not add rte_mbuf_free_bulk (inline) to base code, rather than recoding
> everywhere?

I'm not 100% sure that I understand the question in a right way, but if 
you're
talking about base driver code, it is not used in native datapath 
implementations
at all (just header files with HW/SW interface definition). In fact 
patches 1 and 2
of the series are slightly different and the difference is proved by 
performance
measurements.

Re: [dpdk-dev] [PATCH 1/2] net/sfc: free mbufs in bulks on EF10 native Tx datapath reap

[Andrew Rybchenko]

On 09/12/2017 09:26 PM, Ferruh Yigit wrote:
> On 9/8/2017 3:15 PM, Andrew Rybchenko wrote:
>> From: Ivan Malov <ivan.malov@oktetlabs.ru>
>>
>> Signed-off-by: Ivan Malov <ivan.malov@oktetlabs.ru>
>> Signed-off-by: Andrew Rybchenko <arybchenko@solarflare.com>
>> ---
>>   drivers/net/sfc/sfc_ef10_tx.c | 48 ++++++++++++++++++++++++++++++++++++-------
>>   drivers/net/sfc/sfc_tweak.h   |  3 +++
>>   2 files changed, 44 insertions(+), 7 deletions(-)
>>
>> diff --git a/drivers/net/sfc/sfc_ef10_tx.c b/drivers/net/sfc/sfc_ef10_tx.c
>> index 182fc23..5127a7a 100644
>> --- a/drivers/net/sfc/sfc_ef10_tx.c
>> +++ b/drivers/net/sfc/sfc_ef10_tx.c
>> @@ -158,17 +158,35 @@ struct sfc_ef10_txq {
>>   	pending += sfc_ef10_tx_process_events(txq);
>>   
>>   	if (pending != completed) {
>> +		struct rte_mbuf *bulk[SFC_TX_REAP_BULK_SIZE];
>> +		unsigned int nb = 0;
>> +
>>   		do {
>>   			struct sfc_ef10_tx_sw_desc *txd;
>> +			struct rte_mbuf *m;
>>   
>>   			txd = &txq->sw_ring[completed & ptr_mask];
>> +			if (txd->mbuf == NULL)
>> +				continue;
>>   
>> -			if (txd->mbuf != NULL) {
>> -				rte_pktmbuf_free(txd->mbuf);
>> -				txd->mbuf = NULL;
>> +			m = rte_pktmbuf_prefree_seg(txd->mbuf);
>> +			txd->mbuf = NULL;
>> +			if (m == NULL)
>> +				continue;
>> +
>> +			if ((nb == RTE_DIM(bulk)) ||
>> +			    ((nb != 0) && (m->pool != bulk[0]->pool))) {
> ICC is giving warning [1] here, as far as I can see this is false
> positive but can you please double check in case I am missing something?

Yes, I think it is false positive.

> And unless if you don't see a way to convince icc without effecting
> performance, would you mind updating patch to ignore warning [2] ?

Thanks, done.

>
>
> [1]
> error #3656: variable "bulk" may be used before its value is set
>
> [2]
>    diff --git a/drivers/net/sfc/Makefile b/drivers/net/sfc/Makefile
>    index 57aa963ba..359314809 100644
>    --- a/drivers/net/sfc/Makefile
>    +++ b/drivers/net/sfc/Makefile
>    @@ -65,6 +65,7 @@ CFLAGS += -Wbad-function-cast
>     CFLAGS_BASE_DRIVER += -Wno-empty-body
>     else ifeq ($(CONFIG_RTE_TOOLCHAIN_ICC),y)
>     CFLAGS_BASE_DRIVER += -Wno-unused-but-set-variable
>    +CFLAGS_sfc_ef10_tx.o += -wd3656
>     endif
>
>     #

<...>

Re: [dpdk-dev] [PATCH 2/2] net/sfc: free mbufs in bulks on simple EF10 Tx datapath reap

[Andrew Rybchenko]

On 09/12/2017 09:28 PM, Ferruh Yigit wrote:
> On 9/8/2017 3:15 PM, Andrew Rybchenko wrote:
>> From: Ivan Malov <ivan.malov@oktetlabs.ru>
>>
>> Signed-off-by: Ivan Malov <ivan.malov@oktetlabs.ru>
>> Signed-off-by: Andrew Rybchenko <arybchenko@solarflare.com>
>> ---
>>   doc/guides/nics/sfc_efx.rst   |  4 +++-
>>   drivers/net/sfc/sfc_dp_tx.h   |  2 ++
>>   drivers/net/sfc/sfc_ef10_tx.c | 15 ++++++++++++++-
>>   drivers/net/sfc/sfc_ethdev.c  |  6 ++++++
>>   drivers/net/sfc/sfc_tx.c      | 17 +++++++++++++++++
>>   5 files changed, 42 insertions(+), 2 deletions(-)
>>
>> diff --git a/doc/guides/nics/sfc_efx.rst b/doc/guides/nics/sfc_efx.rst
>> index 973a4a0..028b980 100644
>> --- a/doc/guides/nics/sfc_efx.rst
>> +++ b/doc/guides/nics/sfc_efx.rst
>> @@ -245,12 +245,14 @@ boolean parameters value.
>>     features available and required by the datapath implementation.
>>     **efx** chooses libefx-based datapath which supports VLAN insertion
>>     (full-feature firmware variant only), TSO and multi-segment mbufs.
>> +  Mbuf segments may come from different mempools, and mbuf reference
>> +  counters are treated responsibly.
> This is also the case for ef10 native, right? Does it make sense to
> document it in below too?

Thanks, will add.

>>     **ef10** chooses EF10 (SFN7xxx, SFN8xxx) native datapath which is
>>     more efficient than libefx-based but has no VLAN insertion and TSO
>>     support yet.
>>     **ef10_simple** chooses EF10 (SFN7xxx, SFN8xxx) native datapath which
>>     is even more faster then **ef10** but does not support multi-segment
>> -  mbufs.
>> +  mbufs, disallows multiple mempools and neglects mbuf reference counters.
>>   
>>   - ``perf_profile`` [auto|throughput|low-latency] (default **throughput**)
>>   
> <...>
>
>> --- a/drivers/net/sfc/sfc_ef10_tx.c
>> +++ b/drivers/net/sfc/sfc_ef10_tx.c
>> @@ -401,14 +401,25 @@ struct sfc_ef10_txq {
>>   	pending += sfc_ef10_tx_process_events(txq);
>>   
>>   	if (pending != completed) {
>> +		struct rte_mbuf *bulk[SFC_TX_REAP_BULK_SIZE];
>> +		unsigned int nb = 0;
>> +
>>   		do {
>>   			struct sfc_ef10_tx_sw_desc *txd;
>>   
>>   			txd = &txq->sw_ring[completed & ptr_mask];
>>   
>> -			rte_pktmbuf_free_seg(txd->mbuf);
>> +			if (nb == RTE_DIM(bulk)) {
>> +				rte_mempool_put_bulk(bulk[0]->pool,
>> +						     (void *)bulk, nb);
> same warning here, again false positive I think:
> error #3656: variable "bulk" may be used before its value is set

I think this one is false positive as well.

> The patch to ignore the warning will take care of this one too.
>
>> +				nb = 0;
>> +			}
>> +
>> +			bulk[nb++] = txd->mbuf;
>>   		} while (++completed != pending);
>>   
>> +		rte_mempool_put_bulk(bulk[0]->pool, (void *)bulk, nb);
>> +
>>   		txq->completed = completed;
>>   	}
> <...>
>

[dpdk-dev] [PATCH v2 1/2] net/sfc: free mbufs in bulks on EF10 native Tx datapath reap

[Andrew Rybchenko]

From: Ivan Malov <ivan.malov@oktetlabs.ru>

Signed-off-by: Ivan Malov <ivan.malov@oktetlabs.ru>
Signed-off-by: Andrew Rybchenko <arybchenko@solarflare.com>
---
 drivers/net/sfc/Makefile      |  3 +++
 drivers/net/sfc/sfc_ef10_tx.c | 48 ++++++++++++++++++++++++++++++++++++-------
 drivers/net/sfc/sfc_tweak.h   |  3 +++
 3 files changed, 47 insertions(+), 7 deletions(-)

diff --git a/drivers/net/sfc/Makefile b/drivers/net/sfc/Makefile
index 57aa963..0adb786 100644
--- a/drivers/net/sfc/Makefile
+++ b/drivers/net/sfc/Makefile
@@ -65,6 +65,9 @@ CFLAGS += -Wbad-function-cast
 CFLAGS_BASE_DRIVER += -Wno-empty-body
 else ifeq ($(CONFIG_RTE_TOOLCHAIN_ICC),y)
 CFLAGS_BASE_DRIVER += -Wno-unused-but-set-variable
+# Suppress ICC false positive warning on 'bulk' may be used before its
+# value is set
+CFLAGS_sfc_ef10_tx.o += -wd3656
 endif
 
 #
diff --git a/drivers/net/sfc/sfc_ef10_tx.c b/drivers/net/sfc/sfc_ef10_tx.c
index 182fc23..5127a7a 100644
--- a/drivers/net/sfc/sfc_ef10_tx.c
+++ b/drivers/net/sfc/sfc_ef10_tx.c
@@ -158,17 +158,35 @@ struct sfc_ef10_txq {
 	pending += sfc_ef10_tx_process_events(txq);
 
 	if (pending != completed) {
+		struct rte_mbuf *bulk[SFC_TX_REAP_BULK_SIZE];
+		unsigned int nb = 0;
+
 		do {
 			struct sfc_ef10_tx_sw_desc *txd;
+			struct rte_mbuf *m;
 
 			txd = &txq->sw_ring[completed & ptr_mask];
+			if (txd->mbuf == NULL)
+				continue;
 
-			if (txd->mbuf != NULL) {
-				rte_pktmbuf_free(txd->mbuf);
-				txd->mbuf = NULL;
+			m = rte_pktmbuf_prefree_seg(txd->mbuf);
+			txd->mbuf = NULL;
+			if (m == NULL)
+				continue;
+
+			if ((nb == RTE_DIM(bulk)) ||
+			    ((nb != 0) && (m->pool != bulk[0]->pool))) {
+				rte_mempool_put_bulk(bulk[0]->pool,
+						     (void *)bulk, nb);
+				nb = 0;
 			}
+
+			bulk[nb++] = m;
 		} while (++completed != pending);
 
+		if (nb != 0)
+			rte_mempool_put_bulk(bulk[0]->pool, (void *)bulk, nb);
+
 		txq->completed = completed;
 	}
 
@@ -325,6 +343,7 @@ struct sfc_ef10_txq {
 		do {
 			phys_addr_t seg_addr = rte_mbuf_data_dma_addr(m_seg);
 			unsigned int seg_len = rte_pktmbuf_data_len(m_seg);
+			unsigned int id = added & ptr_mask;
 
 			SFC_ASSERT(seg_len <= SFC_EF10_TX_DMA_DESC_LEN_MAX);
 
@@ -332,15 +351,30 @@ struct sfc_ef10_txq {
 
 			sfc_ef10_tx_qdesc_dma_create(seg_addr,
 				seg_len, (pkt_len == 0),
-				&txq->txq_hw_ring[added & ptr_mask]);
+				&txq->txq_hw_ring[id]);
+
+			/*
+			 * rte_pktmbuf_free() is commonly used in DPDK for
+			 * recycling packets - the function checks every
+			 * segment's reference counter and returns the
+			 * buffer to its pool whenever possible;
+			 * nevertheless, freeing mbuf segments one by one
+			 * may entail some performance decline;
+			 * from this point, sfc_efx_tx_reap() does the same job
+			 * on its own and frees buffers in bulks (all mbufs
+			 * within a bulk belong to the same pool);
+			 * from this perspective, individual segment pointers
+			 * must be associated with the corresponding SW
+			 * descriptors independently so that only one loop
+			 * is sufficient on reap to inspect all the buffers
+			 */
+			txq->sw_ring[id].mbuf = m_seg;
+
 			++added;
 
 		} while ((m_seg = m_seg->next) != 0);
 
 		dma_desc_space -= (added - pkt_start);
-
-		/* Assign mbuf to the last used desc */
-		txq->sw_ring[(added - 1) & ptr_mask].mbuf = *pktp;
 	}
 
 	if (likely(added != txq->added)) {
diff --git a/drivers/net/sfc/sfc_tweak.h b/drivers/net/sfc/sfc_tweak.h
index 4ef7fc8..fd2f75c 100644
--- a/drivers/net/sfc/sfc_tweak.h
+++ b/drivers/net/sfc/sfc_tweak.h
@@ -53,4 +53,7 @@
 /** Default free threshold follows recommendations from DPDK documentation */
 #define SFC_TX_DEFAULT_FREE_THRESH	32
 
+/** Number of mbufs to be freed in bulk in a single call */
+#define SFC_TX_REAP_BULK_SIZE		32
+
 #endif /* _SFC_TWEAK_H_ */
-- 
1.8.2.3

[dpdk-dev] [PATCH v2 2/2] net/sfc: free mbufs in bulks on simple EF10 Tx datapath reap

[Andrew Rybchenko]

From: Ivan Malov <ivan.malov@oktetlabs.ru>

Signed-off-by: Ivan Malov <ivan.malov@oktetlabs.ru>
Signed-off-by: Andrew Rybchenko <arybchenko@solarflare.com>
---
 doc/guides/nics/sfc_efx.rst   |  6 +++++-
 drivers/net/sfc/sfc_dp_tx.h   |  2 ++
 drivers/net/sfc/sfc_ef10_tx.c | 15 ++++++++++++++-
 drivers/net/sfc/sfc_ethdev.c  |  6 ++++++
 drivers/net/sfc/sfc_tx.c      | 17 +++++++++++++++++
 5 files changed, 44 insertions(+), 2 deletions(-)

diff --git a/doc/guides/nics/sfc_efx.rst b/doc/guides/nics/sfc_efx.rst
index 973a4a0..ae2b54a 100644
--- a/doc/guides/nics/sfc_efx.rst
+++ b/doc/guides/nics/sfc_efx.rst
@@ -245,12 +245,16 @@ boolean parameters value.
   features available and required by the datapath implementation.
   **efx** chooses libefx-based datapath which supports VLAN insertion
   (full-feature firmware variant only), TSO and multi-segment mbufs.
+  Mbuf segments may come from different mempools, and mbuf reference
+  counters are treated responsibly.
   **ef10** chooses EF10 (SFN7xxx, SFN8xxx) native datapath which is
   more efficient than libefx-based but has no VLAN insertion and TSO
   support yet.
+  Mbuf segments may come from different mempools, and mbuf reference
+  counters are treated responsibly.
   **ef10_simple** chooses EF10 (SFN7xxx, SFN8xxx) native datapath which
   is even more faster then **ef10** but does not support multi-segment
-  mbufs.
+  mbufs, disallows multiple mempools and neglects mbuf reference counters.
 
 - ``perf_profile`` [auto|throughput|low-latency] (default **throughput**)
 
diff --git a/drivers/net/sfc/sfc_dp_tx.h b/drivers/net/sfc/sfc_dp_tx.h
index db2a70b..94d1b10 100644
--- a/drivers/net/sfc/sfc_dp_tx.h
+++ b/drivers/net/sfc/sfc_dp_tx.h
@@ -142,6 +142,8 @@ struct sfc_dp_tx {
 #define SFC_DP_TX_FEAT_TSO		0x2
 #define SFC_DP_TX_FEAT_MULTI_SEG	0x4
 #define SFC_DP_TX_FEAT_MULTI_PROCESS	0x8
+#define SFC_DP_TX_FEAT_MULTI_POOL	0x10
+#define SFC_DP_TX_FEAT_REFCNT		0x20
 	sfc_dp_tx_qcreate_t		*qcreate;
 	sfc_dp_tx_qdestroy_t		*qdestroy;
 	sfc_dp_tx_qstart_t		*qstart;
diff --git a/drivers/net/sfc/sfc_ef10_tx.c b/drivers/net/sfc/sfc_ef10_tx.c
index 5127a7a..9047b3e 100644
--- a/drivers/net/sfc/sfc_ef10_tx.c
+++ b/drivers/net/sfc/sfc_ef10_tx.c
@@ -401,14 +401,25 @@ struct sfc_ef10_txq {
 	pending += sfc_ef10_tx_process_events(txq);
 
 	if (pending != completed) {
+		struct rte_mbuf *bulk[SFC_TX_REAP_BULK_SIZE];
+		unsigned int nb = 0;
+
 		do {
 			struct sfc_ef10_tx_sw_desc *txd;
 
 			txd = &txq->sw_ring[completed & ptr_mask];
 
-			rte_pktmbuf_free_seg(txd->mbuf);
+			if (nb == RTE_DIM(bulk)) {
+				rte_mempool_put_bulk(bulk[0]->pool,
+						     (void *)bulk, nb);
+				nb = 0;
+			}
+
+			bulk[nb++] = txd->mbuf;
 		} while (++completed != pending);
 
+		rte_mempool_put_bulk(bulk[0]->pool, (void *)bulk, nb);
+
 		txq->completed = completed;
 	}
 
@@ -614,6 +625,8 @@ struct sfc_dp_tx sfc_ef10_tx = {
 		.hw_fw_caps	= SFC_DP_HW_FW_CAP_EF10,
 	},
 	.features		= SFC_DP_TX_FEAT_MULTI_SEG |
+				  SFC_DP_TX_FEAT_MULTI_POOL |
+				  SFC_DP_TX_FEAT_REFCNT |
 				  SFC_DP_TX_FEAT_MULTI_PROCESS,
 	.qcreate		= sfc_ef10_tx_qcreate,
 	.qdestroy		= sfc_ef10_tx_qdestroy,
diff --git a/drivers/net/sfc/sfc_ethdev.c b/drivers/net/sfc/sfc_ethdev.c
index 08299fe..9e65b6a 100644
--- a/drivers/net/sfc/sfc_ethdev.c
+++ b/drivers/net/sfc/sfc_ethdev.c
@@ -145,6 +145,12 @@
 	if (~sa->dp_tx->features & SFC_DP_TX_FEAT_MULTI_SEG)
 		dev_info->default_txconf.txq_flags |= ETH_TXQ_FLAGS_NOMULTSEGS;
 
+	if (~sa->dp_tx->features & SFC_DP_TX_FEAT_MULTI_POOL)
+		dev_info->default_txconf.txq_flags |= ETH_TXQ_FLAGS_NOMULTMEMP;
+
+	if (~sa->dp_tx->features & SFC_DP_TX_FEAT_REFCNT)
+		dev_info->default_txconf.txq_flags |= ETH_TXQ_FLAGS_NOREFCOUNT;
+
 #if EFSYS_OPT_RX_SCALE
 	if (sa->rss_support != EFX_RX_SCALE_UNAVAILABLE) {
 		dev_info->reta_size = EFX_RSS_TBL_SIZE;
diff --git a/drivers/net/sfc/sfc_tx.c b/drivers/net/sfc/sfc_tx.c
index bf59601..4ea7bd7 100644
--- a/drivers/net/sfc/sfc_tx.c
+++ b/drivers/net/sfc/sfc_tx.c
@@ -91,6 +91,21 @@
 		rc = EINVAL;
 	}
 
+	if (((flags & ETH_TXQ_FLAGS_NOMULTMEMP) == 0) &&
+	    (~sa->dp_tx->features & SFC_DP_TX_FEAT_MULTI_POOL)) {
+		sfc_err(sa, "multi-mempool is not supported by %s datapath",
+			sa->dp_tx->dp.name);
+		rc = EINVAL;
+	}
+
+	if (((flags & ETH_TXQ_FLAGS_NOREFCOUNT) == 0) &&
+	    (~sa->dp_tx->features & SFC_DP_TX_FEAT_REFCNT)) {
+		sfc_err(sa,
+			"mbuf reference counters are neglected by %s datapath",
+			sa->dp_tx->dp.name);
+		rc = EINVAL;
+	}
+
 	if ((flags & ETH_TXQ_FLAGS_NOVLANOFFL) == 0) {
 		if (!encp->enc_hw_tx_insert_vlan_enabled) {
 			sfc_err(sa, "VLAN offload is not supported");
@@ -1023,6 +1038,8 @@ struct sfc_dp_tx sfc_efx_tx = {
 	},
 	.features		= SFC_DP_TX_FEAT_VLAN_INSERT |
 				  SFC_DP_TX_FEAT_TSO |
+				  SFC_DP_TX_FEAT_MULTI_POOL |
+				  SFC_DP_TX_FEAT_REFCNT |
 				  SFC_DP_TX_FEAT_MULTI_SEG,
 	.qcreate		= sfc_efx_tx_qcreate,
 	.qdestroy		= sfc_efx_tx_qdestroy,
-- 
1.8.2.3

Re: [dpdk-dev] [PATCH v2 1/2] net/sfc: free mbufs in bulks on EF10 native Tx datapath reap

[Ferruh Yigit]

On 9/13/2017 10:33 AM, Andrew Rybchenko wrote:
> From: Ivan Malov <ivan.malov@oktetlabs.ru>
> 
> Signed-off-by: Ivan Malov <ivan.malov@oktetlabs.ru>
> Signed-off-by: Andrew Rybchenko <arybchenko@solarflare.com>

Series applied to dpdk-next-net/master, thanks.