Re: [PATCH v3 8/8] examples/vhost:support TX offload in vhost sample

[Yuanhan Liu <yuanhan.liu@linux.intel.com>]

On Wed, Nov 04, 2015 at 06:54:16PM +0800, Jijiang Liu wrote:
> Change the vhost sample to support and test TX offload.
> 
> Signed-off-by: Jijiang Liu <jijiang.liu@intel.com>
> ---
>  examples/vhost/main.c |  128 ++++++++++++++++++++++++++++++++++++++++++-------
>  1 files changed, 111 insertions(+), 17 deletions(-)
> 
> diff --git a/examples/vhost/main.c b/examples/vhost/main.c
> index 9eac2d0..06e1e8b 100644
> --- a/examples/vhost/main.c
> +++ b/examples/vhost/main.c
> @@ -50,6 +50,10 @@
>  #include <rte_string_fns.h>
>  #include <rte_malloc.h>
>  #include <rte_virtio_net.h>
> +#include <rte_tcp.h>
> +#include <rte_ip.h>
> +#include <rte_udp.h>
> +#include <rte_sctp.h>
>  
>  #include "main.h"
>  
> @@ -140,6 +144,8 @@
>  
>  #define MBUF_EXT_MEM(mb)   (rte_mbuf_from_indirect(mb) != (mb))
>  
> +#define VIRTIO_TX_CKSUM_OFFLOAD_MASK (PKT_TX_IP_CKSUM | PKT_TX_L4_MASK)
> +
>  /* mask of enabled ports */
>  static uint32_t enabled_port_mask = 0;
>  
> @@ -197,6 +203,13 @@ typedef enum {
>  static uint32_t enable_stats = 0;
>  /* Enable retries on RX. */
>  static uint32_t enable_retry = 1;
> +
> +/* Disable TX checksum offload */
      ^^^^^^^
You meant to "Enable"?

> +static uint32_t enable_tx_csum;
> +
> +/* Disable TSO offload */
> +static uint32_t enable_tso;

Actually, I'd like to see TSO/CSUM offloading is enabled by default:
they are so common, and they are enabled in a lot places by default,
say, kernel, qemu. There is no reason to make it "disable" here.

> +
>  /* Specify timeout (in useconds) between retries on RX. */
>  static uint32_t burst_rx_delay_time = BURST_RX_WAIT_US;
>  /* Specify the number of retries on RX. */
> @@ -292,20 +305,6 @@ struct vlan_ethhdr {
>  	__be16          h_vlan_encapsulated_proto;
>  };
>  
> -/* IPv4 Header */
> -struct ipv4_hdr {
> -	uint8_t  version_ihl;		/**< version and header length */
> -	uint8_t  type_of_service;	/**< type of service */
> -	uint16_t total_length;		/**< length of packet */
> -	uint16_t packet_id;		/**< packet ID */
> -	uint16_t fragment_offset;	/**< fragmentation offset */
> -	uint8_t  time_to_live;		/**< time to live */
> -	uint8_t  next_proto_id;		/**< protocol ID */
> -	uint16_t hdr_checksum;		/**< header checksum */
> -	uint32_t src_addr;		/**< source address */
> -	uint32_t dst_addr;		/**< destination address */
> -} __attribute__((__packed__));


Minor nit: it's a cleanup, having nothing to do with this patch (to
demonstrate TSO/CSUM). It belongs to another patch.

> -
>  /* Header lengths. */
>  #define VLAN_HLEN       4
>  #define VLAN_ETH_HLEN   18
> @@ -441,6 +440,14 @@ port_init(uint8_t port)
>  
>  	if (port >= rte_eth_dev_count()) return -1;
>  
> +	if (enable_tx_csum == 0)
> +		rte_vhost_feature_disable(1ULL << VIRTIO_NET_F_CSUM);
> +
> +	if (enable_tso == 0) {
> +		rte_vhost_feature_disable(1ULL << VIRTIO_NET_F_HOST_TSO4);
> +		rte_vhost_feature_disable(1ULL << VIRTIO_NET_F_HOST_TSO6);
> +	}
> +
>  	rx_rings = (uint16_t)dev_info.max_rx_queues;
>  	/* Configure ethernet device. */
>  	retval = rte_eth_dev_configure(port, rx_rings, tx_rings, &port_conf);
> @@ -576,7 +583,9 @@ us_vhost_usage(const char *prgname)
>  	"		--rx-desc-num [0-N]: the number of descriptors on rx, "
>  			"used only when zero copy is enabled.\n"
>  	"		--tx-desc-num [0-N]: the number of descriptors on tx, "
> -			"used only when zero copy is enabled.\n",
> +			"used only when zero copy is enabled.\n"
> +	"		--tx-csum [0|1] disable/enable TX checksum offload.\n"
> +	"		--tso [0|1] disable/enable TCP segement offload.\n",
>  	       prgname);
>  }
>  
> @@ -602,6 +611,8 @@ us_vhost_parse_args(int argc, char **argv)
>  		{"zero-copy", required_argument, NULL, 0},
>  		{"rx-desc-num", required_argument, NULL, 0},
>  		{"tx-desc-num", required_argument, NULL, 0},
> +		{"tx-csum", required_argument, NULL, 0},
> +		{"tso", required_argument, NULL, 0},
>  		{NULL, 0, 0, 0},
>  	};
>  
> @@ -656,6 +667,28 @@ us_vhost_parse_args(int argc, char **argv)
>  				}
>  			}
>  
> +			/* Enable/disable TX checksum offload. */
> +			if (!strncmp(long_option[option_index].name, "tx-csum", MAX_LONG_OPT_SZ)) {
> +				ret = parse_num_opt(optarg, 1);
> +				if (ret == -1) {
> +					RTE_LOG(INFO, VHOST_CONFIG, "Invalid argument for tx-csum [0|1]\n");
> +					us_vhost_usage(prgname);
> +					return -1;
> +				} else
> +					enable_tx_csum = ret;
> +			}
> +
> +			/* Enable/disable TSO offload. */
> +			if (!strncmp(long_option[option_index].name, "tso", MAX_LONG_OPT_SZ)) {
> +				ret = parse_num_opt(optarg, 1);
> +				if (ret == -1) {
> +					RTE_LOG(INFO, VHOST_CONFIG, "Invalid argument for tso [0|1]\n");
> +					us_vhost_usage(prgname);
> +					return -1;
> +				} else
> +					enable_tso = ret;
> +			}
> +
>  			/* Specify the retries delay time (in useconds) on RX. */
>  			if (!strncmp(long_option[option_index].name, "rx-retry-delay", MAX_LONG_OPT_SZ)) {
>  				ret = parse_num_opt(optarg, INT32_MAX);
> @@ -1114,6 +1147,63 @@ find_local_dest(struct virtio_net *dev, struct rte_mbuf *m,
>  	return 0;
>  }
>  
> +static uint16_t
> +get_psd_sum(void *l3_hdr, uint64_t ol_flags)
> +{
> +	if (ol_flags & PKT_TX_IPV4)
> +		return rte_ipv4_phdr_cksum(l3_hdr, ol_flags);
> +	else /* assume ethertype == ETHER_TYPE_IPv6 */
> +		return rte_ipv6_phdr_cksum(l3_hdr, ol_flags);
> +}
> +
> +static void virtio_tx_offload(struct rte_mbuf *m)
> +{
> +	void *l3_hdr;
> +	struct ipv4_hdr *ipv4_hdr = NULL;
> +	struct tcp_hdr *tcp_hdr = NULL;
> +	struct udp_hdr *udp_hdr = NULL;
> +	struct sctp_hdr *sctp_hdr = NULL;
> +	struct ether_hdr *eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
> +
> +	l3_hdr = (char *)eth_hdr + m->l2_len;
> +
> +	if (m->ol_flags & PKT_TX_IPV4) {
> +		ipv4_hdr = (struct ipv4_hdr *)l3_hdr;
> +		if (m->ol_flags & PKT_TX_IP_CKSUM)
> +			ipv4_hdr->hdr_checksum = 0;
> +	}
> +
> +	if (m->ol_flags & PKT_TX_L4_MASK) {
> +		switch (m->ol_flags & PKT_TX_L4_MASK) {
> +		case PKT_TX_TCP_CKSUM:
> +			tcp_hdr = (struct tcp_hdr *)
> +					((char *)l3_hdr + m->l3_len);
> +			tcp_hdr->cksum = get_psd_sum(l3_hdr, m->ol_flags);

I'm wondering that's necessary here (even for the data going through
nic). AFAIK, the kernel sending the data will calculate pseudo checksum.

(I may be wrong; a simple validation could prove that) 

	--yliu