Re: [PATCH V2] net/thunderx: add support for hardware first skip feature

[Jerin Jacob <jerin.jacob@caviumnetworks.com>]

-----Original Message-----
> Date: Thu, 14 Jun 2018 15:06:19 +0530
> From: rkudurumalla <rakesh.kudurumalla@cavium.com>
> To: dev@dpdk.org
> Cc: ferruh.yigit@intel.com, jerin.jacob@caviumnetworks.com, Rakesh
>  Kudurumalla <rkudurumalla@caviumnetworks.com>
> Subject: [PATCH V2] net/thunderx: add support for hardware first skip

Add [dpdk-dev] in subject.
[dpdk-dev] [PATCH v2] net/thunderx: add support for hardware first skip

>  feature
> X-Mailer: git-send-email 2.7.4
> 
> This feature is used to create a hole between HEADROOM
> and actual data.Size of hole is specified in bytes as
> module param to pmd
> 
> Signed-off-by: Rakesh Kudurumalla <rkudurumalla@caviumnetworks.com>
> ---
> V1: https://dpdk.org/ml/archives/dev/2018-May/103058.html
> V2: Includes changes in thunder.rst to render outupt correctly in html
> 
>  doc/guides/nics/thunderx.rst         | 18 +++++++++
>  drivers/net/thunderx/base/nicvf_hw.c | 12 ++++++
>  drivers/net/thunderx/base/nicvf_hw.h |  1 +
>  drivers/net/thunderx/nicvf_ethdev.c  | 76 +++++++++++++++++++++++++++++++++++-
>  drivers/net/thunderx/nicvf_ethdev.h  |  1 +
>  drivers/net/thunderx/nicvf_struct.h  |  1 +
>  6 files changed, 108 insertions(+), 1 deletion(-)
> 
> diff --git a/doc/guides/nics/thunderx.rst b/doc/guides/nics/thunderx.rst
> index 2642374..7825b43 100644
> --- a/doc/guides/nics/thunderx.rst
> +++ b/doc/guides/nics/thunderx.rst
> @@ -30,6 +30,7 @@ Features of the ThunderX PMD are:
>  - SR-IOV VF
>  - NUMA support
>  - Multi queue set support (up to 96 queues (12 queue sets)) per port
> +- First skip support
>  
>  Supported ThunderX SoCs
>  -----------------------
> @@ -312,6 +313,18 @@ We will choose four secondary queue sets from the ending of the list (0002:01:01
>  
>  The nicvf thunderx driver will make use of attached secondary VFs automatically during the interface configuration stage.
>  
> +
> +Module params
> +--------------
> +

Create a sub section called skip_data_bytes. Since we are referecing this
fearture as skip_data_bytes, please remove first skip instance. I have
reformatted the documention. Here is the diff. Please consider using it.

➜ [master]laptop [dpdk.org] $ git diff
diff --git a/doc/guides/nics/thunderx.rst b/doc/guides/nics/thunderx.rst
index 5b00dd5e8..e84eaafe5 100644
--- a/doc/guides/nics/thunderx.rst
+++ b/doc/guides/nics/thunderx.rst
@@ -30,7 +30,7 @@ Features of the ThunderX PMD are:
 - SR-IOV VF
 - NUMA support
 - Multi queue set support (up to 96 queues (12 queue sets)) per port
-- First skip support
+- Skip data bytes
 
 Supported ThunderX SoCs
 -----------------------
@@ -317,13 +317,16 @@ The nicvf thunderx driver will make use of
attached secondary VFs automatically
 Module params
 --------------
 
-    This feature is used to create a hole between HEADROOM and actual
     data.Size of hole is specified
-    in bytes as module param("skip_data_bytes") to pmd.
-    This scheme is useful when application would like to insert vlan
     header without disturbing HEADROOM
+skip_data_bytes
+~~~~~~~~~~~~~~~
+This feature is used to create a hole between HEADROOM and actual data.
Size of hole is specified
+in bytes as module param("skip_data_bytes") to pmd.
+This scheme is useful when application would like to insert vlan header
without disturbing HEADROOM.
 
+Example:
+   .. code-block:: console
 
-    use ``-w pci_id,skip_data_bytes="number of bytes to skip"``
-
+      -w 0002:01:00.2,skip_data_bytes=8
 
 Limitations
 -----------
@@ -349,7 +352,7 @@ The ThunderX SoC family NICs support up to 12
segments per packet when working
 in scatter/gather mode. So, setting MTU will result with ``EINVAL``
when the
 frame size does not fit in the maximum number of segments.
 
-First_skip
-~~~~~~~~~~
+skip_data_bytes
+~~~~~~~~~~~~~~~
 
-Maximum limit on first_skip is 128 bytes and number of bytes should be
multiple of 8.
+Maximum limit of skip_data_bytes is 128 bytes and number of bytes
should be multiple of 8.



> +    This feature is used to create a hole between HEADROOM and actual data.Size of hole is specified
> +    in bytes as module param("skip_data_bytes") to pmd.
> +    This scheme is useful when application would like to insert vlan header without disturbing HEADROOM
> +
> +
> +    use ``-w pci_id,skip_data_bytes="number of bytes to skip"``
> +
> +
>  Limitations
>  -----------
>  
> @@ -335,3 +348,8 @@ Maximum packet segments
>  The ThunderX SoC family NICs support up to 12 segments per packet when working
>  in scatter/gather mode. So, setting MTU will result with ``EINVAL`` when the
>  frame size does not fit in the maximum number of segments.
> +
> +First_skip
> +~~~~~~~~~~
> +
> +Maximum limit on first_skip is 128 bytes and number of bytes should be multiple of 8.
> diff --git a/drivers/net/thunderx/base/nicvf_hw.c b/drivers/net/thunderx/base/nicvf_hw.c
> index ea8092c..b07a293 100644
> --- a/drivers/net/thunderx/base/nicvf_hw.c
> +++ b/drivers/net/thunderx/base/nicvf_hw.c
> @@ -703,6 +703,18 @@ nicvf_vlan_hw_strip(struct nicvf *nic, bool enable)
>  }
>  
>  void
> +nicvf_first_skip_config(struct nicvf *nic, uint8_t num_dwords)
> +{
> +	uint64_t val;
> +
> +	val = nicvf_reg_read(nic, NIC_VNIC_RQ_GEN_CFG);
> +	val &= ~(0xfULL);
> +	val |= (num_dwords & 0xf);
> +
> +	nicvf_reg_write(nic, NIC_VNIC_RQ_GEN_CFG, val);
> +}
> +
> +void
>  nicvf_apad_config(struct nicvf *nic, bool enable)
>  {
>  	uint64_t val;
> diff --git a/drivers/net/thunderx/base/nicvf_hw.h b/drivers/net/thunderx/base/nicvf_hw.h
> index 284d0bd..fd13ea8 100644
> --- a/drivers/net/thunderx/base/nicvf_hw.h
> +++ b/drivers/net/thunderx/base/nicvf_hw.h
> @@ -193,6 +193,7 @@ uint32_t nicvf_qsize_sq_roundup(uint32_t val);
>  void nicvf_vlan_hw_strip(struct nicvf *nic, bool enable);
>  
>  void nicvf_apad_config(struct nicvf *nic, bool enable);
> +void nicvf_first_skip_config(struct nicvf *nic, uint8_t dwords);
>  
>  int nicvf_rss_config(struct nicvf *nic, uint32_t  qcnt, uint64_t cfg);
>  int nicvf_rss_term(struct nicvf *nic);
> diff --git a/drivers/net/thunderx/nicvf_ethdev.c b/drivers/net/thunderx/nicvf_ethdev.c
> index 99fcd51..466cb86 100644
> --- a/drivers/net/thunderx/nicvf_ethdev.c
> +++ b/drivers/net/thunderx/nicvf_ethdev.c
> @@ -34,6 +34,8 @@
>  #include <rte_pci.h>
>  #include <rte_bus_pci.h>
>  #include <rte_tailq.h>
> +#include <rte_devargs.h>
> +#include <rte_kvargs.h>
>  
>  #include "base/nicvf_plat.h"
>  
> @@ -1230,6 +1232,7 @@ nicvf_rxq_mbuf_setup(struct nicvf_rxq *rxq)
>  {
>  	uintptr_t p;
>  	struct rte_mbuf mb_def;
> +	struct nicvf *nic = rxq->nic;
>  
>  	RTE_BUILD_BUG_ON(sizeof(union mbuf_initializer) != 8);
>  	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, data_off) % 8 != 0);
> @@ -1240,7 +1243,7 @@ nicvf_rxq_mbuf_setup(struct nicvf_rxq *rxq)
>  	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, port) -
>  				offsetof(struct rte_mbuf, data_off) != 6);
>  	mb_def.nb_segs = 1;
> -	mb_def.data_off = RTE_PKTMBUF_HEADROOM;
> +	mb_def.data_off = RTE_PKTMBUF_HEADROOM + (nic->skip_bytes);
>  	mb_def.port = rxq->port_id;
>  	rte_mbuf_refcnt_set(&mb_def, 1);
>  
> @@ -1260,9 +1263,19 @@ nicvf_dev_rx_queue_setup(struct rte_eth_dev *dev, uint16_t qidx,
>  	struct nicvf_rxq *rxq;
>  	struct nicvf *nic = nicvf_pmd_priv(dev);
>  	uint64_t offloads;
> +	uint32_t buffsz;
> +	struct rte_pktmbuf_pool_private *mbp_priv;
>  
>  	PMD_INIT_FUNC_TRACE();
>  
> +	/* First skip check */
> +	mbp_priv = rte_mempool_get_priv(mp);
> +	buffsz = mbp_priv->mbuf_data_room_size - RTE_PKTMBUF_HEADROOM;
> +	if (buffsz < (uint32_t)(nic->skip_bytes)) {
> +		PMD_INIT_LOG(ERR, "First skip is more than configured buffer size");
> +		return -EINVAL;
> +	}
> +
>  	if (qidx >= MAX_RCV_QUEUES_PER_QS)
>  		nic = nic->snicvf[qidx / MAX_RCV_QUEUES_PER_QS - 1];
>  
> @@ -1298,6 +1311,7 @@ nicvf_dev_rx_queue_setup(struct rte_eth_dev *dev, uint16_t qidx,
>  		return -EINVAL;
>  	}
>  
> +
>  	/* Check rx_free_thresh upper bound */
>  	rx_free_thresh = (uint16_t)((rx_conf->rx_free_thresh) ?
>  				rx_conf->rx_free_thresh :
> @@ -1498,6 +1512,7 @@ nicvf_vf_start(struct rte_eth_dev *dev, struct nicvf *nic, uint32_t rbdrsz)
>  			return -EINVAL;
>  		}
>  		rxq->mbuf_phys_off -= data_off;
> +		rxq->mbuf_phys_off -= nic->skip_bytes;
>  
>  		if (mbuf_phys_off == 0)
>  			mbuf_phys_off = rxq->mbuf_phys_off;
> @@ -1978,6 +1993,59 @@ static const struct eth_dev_ops nicvf_eth_dev_ops = {
>  	.get_reg                  = nicvf_dev_get_regs,
>  };
>  
> +static inline int
> +nicvf_set_first_skip(struct rte_eth_dev *dev)
> +{
> +	int bytes_to_skip = 0;
> +	int ret = 0;
> +	unsigned int i;
> +	struct rte_kvargs *kvlist;
> +	static const char *const skip[] = {
> +		SKIP_DATA_BYTES,
> +		NULL};
> +	struct nicvf *nic = nicvf_pmd_priv(dev);
> +
> +	if (!dev->device->devargs) {
> +		nicvf_first_skip_config(nic, 0);
> +		return ret;
> +	}
> +
> +	kvlist = rte_kvargs_parse(dev->device->devargs->args, skip);
> +	if (!kvlist)
> +		return -EINVAL;
> +
> +	if (kvlist->count == 0)
> +		goto exit;
> +
> +	for (i = 0; i != kvlist->count; ++i) {
> +		const struct rte_kvargs_pair *pair = &kvlist->pairs[i];
> +
> +		if (!strcmp(pair->key, SKIP_DATA_BYTES))
> +			bytes_to_skip = atoi(pair->value);
> +	}
> +
> +		/*128 bytes amounts to one cache line*/

Please align the comment.

Rest looks good to me.