rdma: add support for RSS configuration

Type: feature

Change-Id: I911ee94b624ef5edf46b34666597eff93df1d1ed
Signed-off-by: Benoît Ganne <bganne@cisco.com>

7 files changed, 338 insertions, 14 deletions

diff --git a/src/plugins/rdma/api.c b/src/plugins/rdma/api.c
index 06e7385136f..7fe77105596 100644
--- a/src/plugins/rdma/api.c
+++ b/src/plugins/rdma/api.c
@@ -44,6 +44,69 @@ rdma_api_mode (vl_api_rdma_mode_t mode)
   return RDMA_MODE_AUTO;
 }
 
+static rdma_rss4_t
+rdma_api_rss4 (const vl_api_rdma_rss4_t rss4)
+{
+  switch (rss4)
+    {
+    case RDMA_API_RSS4_AUTO:
+      return RDMA_RSS4_AUTO;
+    case RDMA_API_RSS4_IP:
+      return RDMA_RSS4_IP;
+    case RDMA_API_RSS4_IP_UDP:
+      return RDMA_RSS4_IP_UDP;
+    case RDMA_API_RSS4_IP_TCP:
+      return RDMA_RSS4_IP_TCP;
+    }
+  return RDMA_RSS4_AUTO;
+}
+
+static rdma_rss6_t
+rdma_api_rss6 (const vl_api_rdma_rss6_t rss6)
+{
+  switch (rss6)
+    {
+    case RDMA_API_RSS6_AUTO:
+      return RDMA_RSS6_AUTO;
+    case RDMA_API_RSS6_IP:
+      return RDMA_RSS6_IP;
+    case RDMA_API_RSS6_IP_UDP:
+      return RDMA_RSS6_IP_UDP;
+    case RDMA_API_RSS6_IP_TCP:
+      return RDMA_RSS6_IP_TCP;
+    }
+  return RDMA_RSS6_AUTO;
+}
+
+static void
+vl_api_rdma_create_v3_t_handler (vl_api_rdma_create_v3_t *mp)
+{
+  vlib_main_t *vm = vlib_get_main ();
+  rdma_main_t *rm = &rdma_main;
+  vl_api_rdma_create_v3_reply_t *rmp;
+  rdma_create_if_args_t args;
+  int rv;
+
+  clib_memset (&args, 0, sizeof (rdma_create_if_args_t));
+
+  args.ifname = mp->host_if;
+  args.name = mp->name;
+  args.rxq_num = ntohs (mp->rxq_num);
+  args.rxq_size = ntohs (mp->rxq_size);
+  args.txq_size = ntohs (mp->txq_size);
+  args.mode = rdma_api_mode (mp->mode);
+  args.disable_striding_rq = 0;
+  args.no_multi_seg = mp->no_multi_seg;
+  args.max_pktlen = ntohs (mp->max_pktlen);
+  args.rss4 = rdma_api_rss4 (mp->rss4);
+  args.rss6 = rdma_api_rss6 (mp->rss6);
+  rdma_create_if (vm, &args);
+  rv = args.rv;
+
+  REPLY_MACRO2 (VL_API_RDMA_CREATE_V3_REPLY + rm->msg_id_base,
+		({ rmp->sw_if_index = ntohl (args.sw_if_index); }));
+}
+
 static void
 vl_api_rdma_create_v2_t_handler (vl_api_rdma_create_v2_t * mp)
 {
diff --git a/src/plugins/rdma/device.c b/src/plugins/rdma/device.c
index 14262f597b7..1198d99b14e 100644
--- a/src/plugins/rdma/device.c
+++ b/src/plugins/rdma/device.c
@@ -311,7 +311,7 @@ rdma_async_event_read_ready (clib_file_t * f)
       vlib_log_emerg (rm->log_class, "%s: fatal error", rd->name);
       break;
     default:
-      rdma_log__ (VLIB_LOG_LEVEL_ERR, rd, "unhandeld RDMA async event %i",
+      rdma_log__ (VLIB_LOG_LEVEL_ERR, rd, "unhandeld RDMA async event %d",
 		  event.event_type);
       break;
     }
@@ -611,8 +611,46 @@ rdma_rxq_init (vlib_main_t * vm, rdma_device_t * rd, u16 qid, u32 n_desc,
   return 0;
 }
 
+static uint64_t
+rdma_rss42ibv (const rdma_rss4_t rss4)
+{
+  switch (rss4)
+    {
+    case RDMA_RSS4_IP:
+      return IBV_RX_HASH_SRC_IPV4 | IBV_RX_HASH_DST_IPV4;
+    case RDMA_RSS4_IP_UDP:
+      return IBV_RX_HASH_SRC_IPV4 | IBV_RX_HASH_DST_IPV4 |
+	     IBV_RX_HASH_SRC_PORT_UDP | IBV_RX_HASH_DST_PORT_UDP;
+    case RDMA_RSS4_AUTO: /* fallthrough */
+    case RDMA_RSS4_IP_TCP:
+      return IBV_RX_HASH_SRC_IPV4 | IBV_RX_HASH_DST_IPV4 |
+	     IBV_RX_HASH_SRC_PORT_TCP | IBV_RX_HASH_DST_PORT_TCP;
+    }
+  ASSERT (0);
+  return 0;
+}
+
+static uint64_t
+rdma_rss62ibv (const rdma_rss6_t rss6)
+{
+  switch (rss6)
+    {
+    case RDMA_RSS6_IP:
+      return IBV_RX_HASH_SRC_IPV6 | IBV_RX_HASH_DST_IPV6;
+    case RDMA_RSS6_IP_UDP:
+      return IBV_RX_HASH_SRC_IPV6 | IBV_RX_HASH_DST_IPV6 |
+	     IBV_RX_HASH_SRC_PORT_UDP | IBV_RX_HASH_DST_PORT_UDP;
+    case RDMA_RSS6_AUTO: /* fallthrough */
+    case RDMA_RSS6_IP_TCP:
+      return IBV_RX_HASH_SRC_IPV6 | IBV_RX_HASH_DST_IPV6 |
+	     IBV_RX_HASH_SRC_PORT_TCP | IBV_RX_HASH_DST_PORT_TCP;
+    }
+  ASSERT (0);
+  return 0;
+}
+
 static clib_error_t *
-rdma_rxq_finalize (vlib_main_t * vm, rdma_device_t * rd)
+rdma_rxq_finalize (vlib_main_t *vm, rdma_device_t *rd)
 {
   struct ibv_rwq_ind_table_init_attr rwqia;
   struct ibv_qp_init_attr_ex qpia;
@@ -658,15 +696,11 @@ rdma_rxq_finalize (vlib_main_t * vm, rdma_device_t * rd)
   qpia.rx_hash_conf.rx_hash_key = rdma_rss_hash_key;
   qpia.rx_hash_conf.rx_hash_function = IBV_RX_HASH_FUNC_TOEPLITZ;
 
-  qpia.rx_hash_conf.rx_hash_fields_mask =
-    IBV_RX_HASH_SRC_IPV4 | IBV_RX_HASH_DST_IPV4 | IBV_RX_HASH_SRC_PORT_TCP |
-    IBV_RX_HASH_DST_PORT_TCP;
+  qpia.rx_hash_conf.rx_hash_fields_mask = rdma_rss42ibv (rd->rss4);
   if ((rd->rx_qp4 = ibv_create_qp_ex (rd->ctx, &qpia)) == 0)
     return clib_error_return_unix (0, "IPv4 Queue Pair create failed");
 
-  qpia.rx_hash_conf.rx_hash_fields_mask =
-    IBV_RX_HASH_SRC_IPV6 | IBV_RX_HASH_DST_IPV6 | IBV_RX_HASH_SRC_PORT_TCP |
-    IBV_RX_HASH_DST_PORT_TCP;
+  qpia.rx_hash_conf.rx_hash_fields_mask = rdma_rss62ibv (rd->rss6);
   if ((rd->rx_qp6 = ibv_create_qp_ex (rd->ctx, &qpia)) == 0)
     return clib_error_return_unix (0, "IPv6 Queue Pair create failed");
 
@@ -802,6 +836,9 @@ rdma_dev_init (vlib_main_t * vm, rdma_device_t * rd,
 
   ethernet_mac_address_generate (rd->hwaddr.bytes);
 
+  rd->rss4 = args->rss4;
+  rd->rss6 = args->rss6;
+
   /*
    * /!\ WARNING /!\ creation order is important
    * We *must* create TX queues *before* RX queues, otherwise we will receive
@@ -862,8 +899,9 @@ rdma_create_if (vlib_main_t * vm, rdma_create_if_args_t * args)
       !is_pow2 (args->rxq_size) || !is_pow2 (args->txq_size))
     {
       args->rv = VNET_API_ERROR_INVALID_VALUE;
-      args->error = clib_error_return (0, "queue size must be a power of two "
-				       "between %i and 65535",
+      args->error = clib_error_return (0,
+				       "queue size must be a power of two "
+				       "between %d and 65535",
 				       VLIB_FRAME_SIZE);
       goto err0;
     }
diff --git a/src/plugins/rdma/format.c b/src/plugins/rdma/format.c
index 84b6439a2fd..aada52a1ec3 100644
--- a/src/plugins/rdma/format.c
+++ b/src/plugins/rdma/format.c
@@ -73,6 +73,42 @@ format_rdma_bit_flag (u8 * s, va_list * args)
 }
 
 u8 *
+format_rdma_rss4 (u8 *s, va_list *args)
+{
+  const rdma_rss4_t *rss4 = va_arg (*args, const rdma_rss4_t *);
+  switch (*rss4)
+    {
+    case RDMA_RSS4_IP:
+      return format (s, "ipv4");
+    case RDMA_RSS4_IP_UDP:
+      return format (s, "ipv4-udp");
+    case RDMA_RSS4_AUTO: /* fallthrough */
+    case RDMA_RSS4_IP_TCP:
+      return format (s, "ipv4-tcp");
+    }
+  ASSERT (0);
+  return format (s, "unknown(%x)", *rss4);
+}
+
+u8 *
+format_rdma_rss6 (u8 *s, va_list *args)
+{
+  const rdma_rss6_t *rss6 = va_arg (*args, const rdma_rss6_t *);
+  switch (*rss6)
+    {
+    case RDMA_RSS6_IP:
+      return format (s, "ipv6");
+    case RDMA_RSS6_IP_UDP:
+      return format (s, "ipv6-udp");
+    case RDMA_RSS6_AUTO: /* fallthrough */
+    case RDMA_RSS6_IP_TCP:
+      return format (s, "ipv6-tcp");
+    }
+  ASSERT (0);
+  return format (s, "unknown(%x)", *rss6);
+}
+
+u8 *
 format_rdma_device (u8 * s, va_list * args)
 {
   vlib_main_t *vm = vlib_get_main ();
@@ -96,8 +132,10 @@ format_rdma_device (u8 * s, va_list * args)
 		  format_vlib_pci_vpd, d->vpd_r, "SN");
       vlib_pci_free_device_info (d);
     }
-  s = format (s, "%Uflags: %U", format_white_space, indent,
+  s = format (s, "%Uflags: %U\n", format_white_space, indent,
 	      format_rdma_device_flags, rd);
+  s = format (s, "%Urss: %U %U", format_white_space, indent, format_rdma_rss4,
+	      &rd->rss4, format_rdma_rss6, &rd->rss6);
   if (rd->error)
     s = format (s, "\n%Uerror %U", format_white_space, indent,
 		format_clib_error, rd->error);
diff --git a/src/plugins/rdma/rdma.api b/src/plugins/rdma/rdma.api
index 4519e2316d3..f2c70c7e514 100644
--- a/src/plugins/rdma/rdma.api
+++ b/src/plugins/rdma/rdma.api
@@ -15,7 +15,7 @@
  *------------------------------------------------------------------
  */
 
-option version = "2.0.0";
+option version = "3.0.0";
 import "vnet/interface_types.api";
 
 enum rdma_mode
@@ -25,7 +25,6 @@ enum rdma_mode
   RDMA_API_MODE_DV = 2,
 };
 
-
 /** \brief
     @param client_index - opaque cookie to identify the sender
     @param context - sender context, to match reply w/ request
@@ -67,6 +66,7 @@ define rdma_create
 
 define rdma_create_v2
 {
+  option deprecated;
   u32 client_index;
   u32 context;
 
@@ -81,6 +81,55 @@ define rdma_create_v2
   option vat_help = "<host-if ifname> [name <name>] [rx-queue-size <size>] [tx-queue-size <size>] [num-rx-queues <size>] [mode <auto|ibv|dv>] [no-multi-seg] [max-pktlen <size>]";
 };
 
+enum rdma_rss4
+{
+  RDMA_API_RSS4_AUTO = 0,
+  RDMA_API_RSS4_IP = 1,
+  RDMA_API_RSS4_IP_UDP = 2,
+  RDMA_API_RSS4_IP_TCP = 3,
+};
+
+enum rdma_rss6
+{
+  RDMA_API_RSS6_AUTO = 0,
+  RDMA_API_RSS6_IP = 1,
+  RDMA_API_RSS6_IP_UDP = 2,
+  RDMA_API_RSS6_IP_TCP = 3,
+};
+
+/** \brief
+    @param client_index - opaque cookie to identify the sender
+    @param context - sender context, to match reply w/ request
+    @param host_if - Linux netdev interface name
+    @param name - new rdma interface name
+    @param rxq_num - number of receive queues (optional)
+    @param rxq_size - receive queue size (optional)
+    @param txq_size - transmit queue size (optional)
+    @param mode - operation mode (optional)
+    @param no_multi_seg (optional) - disable chained buffer RX
+    @param max_pktlen (optional) - maximal RX packet size.
+    @param rss4 (optional) - IPv4 RSS
+    @param rss6 (optional) - IPv6 RSS
+*/
+
+define rdma_create_v3
+{
+  u32 client_index;
+  u32 context;
+
+  string host_if[64];
+  string name[64];
+  u16 rxq_num [default=1];
+  u16 rxq_size [default=1024];
+  u16 txq_size [default=1024];
+  vl_api_rdma_mode_t mode [default=0];
+  bool no_multi_seg [default=0];
+  u16 max_pktlen [default=0];
+  vl_api_rdma_rss4_t rss4 [default=0];
+  vl_api_rdma_rss6_t rss6 [default=0];
+  option vat_help = "<host-if ifname> [name <name>] [rx-queue-size <size>] [tx-queue-size <size>] [num-rx-queues <size>] [mode <auto|ibv|dv>] [no-multi-seg] [max-pktlen <size>] [rss <ipv4|ipv4-udp|ipv4-tcp>] [rss <ipv6|ipv6-udp|ipv6-tcp>]";
+};
+
 
 /** \brief
     @param context - sender context, to match reply w/ request
@@ -114,6 +163,19 @@ define rdma_create_v2_reply
     @param sw_if_index - interface index
 */
 
+define rdma_create_v3_reply
+{
+  u32 context;
+  i32 retval;
+  vl_api_interface_index_t sw_if_index;
+};
+
+/** \brief
+    @param client_index - opaque cookie to identify the sender
+    @param context - sender context, to match reply w/ request
+    @param sw_if_index - interface index
+*/
+
 autoreply define rdma_delete
 {
   u32 client_index;
diff --git a/src/plugins/rdma/rdma.h b/src/plugins/rdma/rdma.h
index a8ab07012ee..323275c9f50 100644
--- a/src/plugins/rdma/rdma.h
+++ b/src/plugins/rdma/rdma.h
@@ -174,6 +174,23 @@ STATIC_ASSERT_OFFSET_OF (rdma_txq_t, cacheline2, 128);
 #define RDMA_RXQ_MAX_CHAIN_LOG_SZ 3	/* This should NOT be lower than 3! */
 #define RDMA_RXQ_MAX_CHAIN_SZ (1U << RDMA_RXQ_MAX_CHAIN_LOG_SZ)
 #define RDMA_RXQ_LEGACY_MODE_MAX_CHAIN_SZ 5
+
+typedef enum
+{
+  RDMA_RSS4_AUTO = 0,
+  RDMA_RSS4_IP,
+  RDMA_RSS4_IP_UDP,
+  RDMA_RSS4_IP_TCP,
+} rdma_rss4_t;
+
+typedef enum
+{
+  RDMA_RSS6_AUTO = 0,
+  RDMA_RSS6_IP,
+  RDMA_RSS6_IP_UDP,
+  RDMA_RSS6_IP_TCP,
+} rdma_rss6_t;
+
 typedef struct
 {
   CLIB_CACHE_LINE_ALIGN_MARK (cacheline0);
@@ -195,6 +212,8 @@ typedef struct
   mac_address_t hwaddr;
   u32 async_event_clib_file_index;
   u32 dev_instance;
+  rdma_rss4_t rss4;
+  rdma_rss6_t rss6;
 
   struct ibv_context *ctx;
   struct ibv_pd *pd;
@@ -264,6 +283,8 @@ typedef struct
   u8 no_multi_seg;
   u8 disable_striding_rq;
   u16 max_pktlen;
+  rdma_rss4_t rss4;
+  rdma_rss6_t rss6;
 
   /* return */
   int rv;
diff --git a/src/plugins/rdma/test_api.c b/src/plugins/rdma/test_api.c
index ff5dec1ad95..e9d5fcaad98 100644
--- a/src/plugins/rdma/test_api.c
+++ b/src/plugins/rdma/test_api.c
@@ -58,6 +58,40 @@ api_rdma_mode (rdma_mode_t mode)
   return ~0;
 }
 
+static vl_api_rdma_rss4_t
+api_rdma_rss4 (rdma_rss4_t rss4)
+{
+  switch (rss4)
+    {
+    case RDMA_RSS4_AUTO:
+      return RDMA_API_RSS4_AUTO;
+    case RDMA_RSS4_IP:
+      return RDMA_API_RSS4_IP;
+    case RDMA_RSS4_IP_UDP:
+      return RDMA_API_RSS4_IP_UDP;
+    case RDMA_RSS4_IP_TCP:
+      return RDMA_API_RSS4_IP_TCP;
+    }
+  return ~0;
+}
+
+static vl_api_rdma_rss6_t
+api_rdma_rss6 (rdma_rss6_t rss6)
+{
+  switch (rss6)
+    {
+    case RDMA_RSS6_AUTO:
+      return RDMA_API_RSS6_AUTO;
+    case RDMA_RSS6_IP:
+      return RDMA_API_RSS6_IP;
+    case RDMA_RSS6_IP_UDP:
+      return RDMA_API_RSS6_IP_UDP;
+    case RDMA_RSS6_IP_TCP:
+      return RDMA_API_RSS6_IP_TCP;
+    }
+  return ~0;
+}
+
 /* rdma create API */
 static int
 api_rdma_create (vat_main_t * vam)
@@ -103,7 +137,10 @@ api_rdma_create_v2 (vat_main_t * vam)
   M (RDMA_CREATE_V2, mp);
 
   snprintf ((char *) mp->host_if, sizeof (mp->host_if), "%s", args.ifname);
-  snprintf ((char *) mp->name, sizeof (mp->name), "%s", args.name);
+  if (args.name)
+    snprintf ((char *) mp->name, sizeof (mp->name), "%s", args.name);
+  else
+    mp->name[0] = 0;
   mp->rxq_num = clib_host_to_net_u16 (args.rxq_num);
   mp->rxq_size = clib_host_to_net_u16 (args.rxq_size);
   mp->txq_size = clib_host_to_net_u16 (args.txq_size);
@@ -117,6 +154,41 @@ api_rdma_create_v2 (vat_main_t * vam)
   return ret;
 }
 
+static int
+api_rdma_create_v3 (vat_main_t *vam)
+{
+  vl_api_rdma_create_v3_t *mp;
+  rdma_create_if_args_t args;
+  int ret;
+
+  if (!unformat_user (vam->input, unformat_rdma_create_if_args, &args))
+    {
+      clib_warning ("unknown input `%U'", format_unformat_error, vam->input);
+      return -99;
+    }
+
+  M (RDMA_CREATE_V3, mp);
+
+  snprintf ((char *) mp->host_if, sizeof (mp->host_if), "%s", args.ifname);
+  if (args.name)
+    snprintf ((char *) mp->name, sizeof (mp->name), "%s", args.name);
+  else
+    mp->name[0] = 0;
+  mp->rxq_num = clib_host_to_net_u16 (args.rxq_num);
+  mp->rxq_size = clib_host_to_net_u16 (args.rxq_size);
+  mp->txq_size = clib_host_to_net_u16 (args.txq_size);
+  mp->mode = api_rdma_mode (args.mode);
+  mp->no_multi_seg = args.no_multi_seg;
+  mp->max_pktlen = clib_host_to_net_u16 (args.max_pktlen);
+  mp->rss4 = api_rdma_rss4 (args.rss4);
+  mp->rss6 = api_rdma_rss6 (args.rss6);
+
+  S (mp);
+  W (ret);
+
+  return ret;
+}
+
 /* rdma-create reply handler */
 static void
 vl_api_rdma_create_reply_t_handler (vl_api_rdma_create_reply_t * mp)
@@ -153,6 +225,24 @@ vl_api_rdma_create_v2_reply_t_handler (vl_api_rdma_create_v2_reply_t * mp)
   vam->regenerate_interface_table = 1;
 }
 
+/* rdma-create reply handler v3 */
+static void
+vl_api_rdma_create_v3_reply_t_handler (vl_api_rdma_create_v3_reply_t *mp)
+{
+  vat_main_t *vam = rdma_test_main.vat_main;
+  i32 retval = ntohl (mp->retval);
+
+  if (retval == 0)
+    {
+      fformat (vam->ofp, "created rdma with sw_if_index %d\n",
+	       ntohl (mp->sw_if_index));
+    }
+
+  vam->retval = retval;
+  vam->result_ready = 1;
+  vam->regenerate_interface_table = 1;
+}
+
 /* rdma delete API */
 static int
 api_rdma_delete (vat_main_t * vam)
diff --git a/src/plugins/rdma/unformat.c b/src/plugins/rdma/unformat.c
index 26c184793bd..c529958b585 100644
--- a/src/plugins/rdma/unformat.c
+++ b/src/plugins/rdma/unformat.c
@@ -54,6 +54,18 @@ unformat_rdma_create_if_args (unformat_input_t * input, va_list * vargs)
 	args->no_multi_seg = 1;
       else if (unformat (line_input, "max-pktlen %u", &tmp))
 	args->max_pktlen = tmp;
+      else if (unformat (line_input, "rss ipv4-udp"))
+	args->rss4 = RDMA_RSS4_IP_UDP;
+      else if (unformat (line_input, "rss ipv4-tcp"))
+	args->rss4 = RDMA_RSS4_IP_TCP;
+      else if (unformat (line_input, "rss ipv4"))
+	args->rss4 = RDMA_RSS4_IP;
+      else if (unformat (line_input, "rss ipv6-udp"))
+	args->rss6 = RDMA_RSS6_IP_UDP;
+      else if (unformat (line_input, "rss ipv6-tcp"))
+	args->rss6 = RDMA_RSS6_IP_TCP;
+      else if (unformat (line_input, "rss ipv6"))
+	args->rss6 = RDMA_RSS6_IP;
       else
 	{
 	  /* return failure on unknown input */