af_xdp: introduce to netns api

In some situation, we support to deploy vpp as per host
and handler packet in container, so we use xdp to redirect
the flow.

Type: improvement

Signed-off-by: arikachen <eaglesora@gmail.com>
Change-Id: Iab42d6a0abb2b330a284d519018a90aff2fa4371

7 files changed, 227 insertions, 10 deletions

diff --git a/src/plugins/af_xdp/af_xdp.api b/src/plugins/af_xdp/af_xdp.api
index c6716123703..9fe6bc04939 100644
--- a/src/plugins/af_xdp/af_xdp.api
+++ b/src/plugins/af_xdp/af_xdp.api
@@ -61,6 +61,38 @@ define af_xdp_create
 };
 
 /** \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 af_xdp interface name (optional)
+    @param rxq_num - number of receive queues. 65535 can be used as special value to request all available queues (optional)
+    @param rxq_size - receive queue size (optional)
+    @param txq_size - transmit queue size (optional)
+    @param mode - operation mode (optional)
+    @param flags - flags (optional)
+    @param prog - eBPF program path (optional)
+    @param namespace - netns of nic (optional)
+*/
+
+define af_xdp_create_v2
+{
+  u32 client_index;
+  u32 context;
+
+  string host_if[64];
+  string name[64];
+  u16 rxq_num [default=1];
+  u16 rxq_size [default=0];
+  u16 txq_size [default=0];
+  vl_api_af_xdp_mode_t mode [default=0];
+  vl_api_af_xdp_flag_t flags [default=0];
+  string prog[256];
+  string namespace[64];
+  option vat_help = "<host-if linux-ifname> [name ifname] [rx-queue-size size] [tx-queue-size size] [num-rx-queues <num|all>] [prog pathname] [netns ns] [zero-copy|no-zero-copy] [no-syscall-lock]";
+  option status="in_progress";
+};
+
+/** \brief
     @param context - sender context, to match reply w/ request
     @param retval - return value for request
     @param sw_if_index - software index for the new af_xdp interface
@@ -75,6 +107,20 @@ define af_xdp_create_reply
 };
 
 /** \brief
+    @param context - sender context, to match reply w/ request
+    @param retval - return value for request
+    @param sw_if_index - software index for the new af_xdp interface
+*/
+
+define af_xdp_create_v2_reply
+{
+  u32 context;
+  i32 retval;
+  vl_api_interface_index_t sw_if_index;
+  option status="in_progress";
+};
+
+/** \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
diff --git a/src/plugins/af_xdp/af_xdp.h b/src/plugins/af_xdp/af_xdp.h
index 825a3fb29fd..85c8fee545d 100644
--- a/src/plugins/af_xdp/af_xdp.h
+++ b/src/plugins/af_xdp/af_xdp.h
@@ -113,6 +113,8 @@ typedef struct
 
   u8 rxq_num;
 
+  char *netns;
+
   struct xsk_umem **umem;
   struct xsk_socket **xsk;
 
@@ -149,6 +151,7 @@ typedef struct
   char *linux_ifname;
   char *name;
   char *prog;
+  char *netns;
   af_xdp_mode_t mode;
   af_xdp_create_flag_t flags;
   u32 rxq_size;
diff --git a/src/plugins/af_xdp/api.c b/src/plugins/af_xdp/api.c
index 1864c4c2ee9..45dab284e4c 100644
--- a/src/plugins/af_xdp/api.c
+++ b/src/plugins/af_xdp/api.c
@@ -87,6 +87,38 @@ vl_api_af_xdp_create_t_handler (vl_api_af_xdp_create_t * mp)
 }
 
 static void
+vl_api_af_xdp_create_v2_t_handler (vl_api_af_xdp_create_v2_t *mp)
+{
+  vlib_main_t *vm = vlib_get_main ();
+  af_xdp_main_t *rm = &af_xdp_main;
+  vl_api_af_xdp_create_v2_reply_t *rmp;
+  af_xdp_create_if_args_t args;
+  int rv;
+
+  clib_memset (&args, 0, sizeof (af_xdp_create_if_args_t));
+
+  args.linux_ifname = mp->host_if[0] ? (char *) mp->host_if : 0;
+  args.name = mp->name[0] ? (char *) mp->name : 0;
+  args.prog = mp->prog[0] ? (char *) mp->prog : 0;
+  args.netns = mp->namespace[0] ? (char *) mp->namespace : 0;
+  args.mode = af_xdp_api_mode (mp->mode);
+  args.flags = af_xdp_api_flags (mp->flags);
+  args.rxq_size = ntohs (mp->rxq_size);
+  args.txq_size = ntohs (mp->txq_size);
+  args.rxq_num = ntohs (mp->rxq_num);
+
+  af_xdp_create_if (vm, &args);
+  rv = args.rv;
+
+  /* clang-format off */
+  REPLY_MACRO2 (VL_API_AF_XDP_CREATE_V2_REPLY + rm->msg_id_base,
+    ({
+      rmp->sw_if_index = ntohl (args.sw_if_index);
+    }));
+  /* clang-format on */
+}
+
+static void
 vl_api_af_xdp_delete_t_handler (vl_api_af_xdp_delete_t * mp)
 {
   vlib_main_t *vm = vlib_get_main ();
diff --git a/src/plugins/af_xdp/cli.c b/src/plugins/af_xdp/cli.c
index 2f3deffaaee..660725813aa 100644
--- a/src/plugins/af_xdp/cli.c
+++ b/src/plugins/af_xdp/cli.c
@@ -40,6 +40,8 @@ af_xdp_create_command_fn (vlib_main_t * vm, unformat_input_t * input,
 
   vec_free (args.linux_ifname);
   vec_free (args.name);
+  vec_free (args.prog);
+  vec_free (args.netns);
 
   return args.error;
 }
@@ -50,7 +52,7 @@ VLIB_CLI_COMMAND (af_xdp_create_command, static) = {
   .short_help =
     "create interface af_xdp <host-if linux-ifname> [name ifname] "
     "[rx-queue-size size] [tx-queue-size size] [num-rx-queues <num|all>] "
-    "[prog pathname] [zero-copy|no-zero-copy] [no-syscall-lock]",
+    "[prog pathname] [netns ns] [zero-copy|no-zero-copy] [no-syscall-lock]",
   .function = af_xdp_create_command_fn,
 };
 /* *INDENT-ON* */
diff --git a/src/plugins/af_xdp/device.c b/src/plugins/af_xdp/device.c
index 8365a716e97..0b39280c569 100644
--- a/src/plugins/af_xdp/device.c
+++ b/src/plugins/af_xdp/device.c
@@ -25,6 +25,7 @@
 #include <vlib/vlib.h>
 #include <vlib/unix/unix.h>
 #include <vlib/pci/pci.h>
+#include <vppinfra/linux/netns.h>
 #include <vppinfra/linux/sysfs.h>
 #include <vppinfra/unix.h>
 #include <vnet/ethernet/ethernet.h>
@@ -89,6 +90,48 @@ af_xdp_flag_change (vnet_main_t * vnm, vnet_hw_interface_t * hw, u32 flags)
   return ~0;
 }
 
+int
+af_xdp_enter_netns (char *netns, int *fds)
+{
+  *fds = *(fds + 1) = -1;
+  if (netns != NULL)
+    {
+      *fds = clib_netns_open (NULL /* self */);
+      if ((*(fds + 1) = clib_netns_open ((u8 *) netns)) == -1)
+	return VNET_API_ERROR_SYSCALL_ERROR_8;
+      if (clib_setns (*(fds + 1)) == -1)
+	return VNET_API_ERROR_SYSCALL_ERROR_9;
+    }
+  return 0;
+}
+
+void
+af_xdp_cleanup_netns (int *fds)
+{
+  if (*fds != -1)
+    close (*fds);
+
+  if (*(fds + 1) != -1)
+    close (*(fds + 1));
+
+  *fds = *(fds + 1) = -1;
+}
+
+int
+af_xdp_exit_netns (char *netns, int *fds)
+{
+  int ret = 0;
+  if (netns != NULL)
+    {
+      if (*fds != -1)
+	ret = clib_setns (*fds);
+
+      af_xdp_cleanup_netns (fds);
+    }
+
+  return ret;
+}
+
 void
 af_xdp_delete_if (vlib_main_t * vm, af_xdp_device_t * ad)
 {
@@ -118,7 +161,11 @@ af_xdp_delete_if (vlib_main_t * vm, af_xdp_device_t * ad)
 
   if (ad->bpf_obj)
     {
+      int ns_fds[2];
+      af_xdp_enter_netns (ad->netns, ns_fds);
       bpf_set_link_xdp_fd (ad->linux_ifindex, -1, 0);
+      af_xdp_exit_netns (ad->netns, ns_fds);
+
       bpf_object__unload (ad->bpf_obj);
     }
 
@@ -127,6 +174,9 @@ af_xdp_delete_if (vlib_main_t * vm, af_xdp_device_t * ad)
   vec_free (ad->buffer_template);
   vec_free (ad->rxqs);
   vec_free (ad->txqs);
+  vec_free (ad->name);
+  vec_free (ad->linux_ifname);
+  vec_free (ad->netns);
   clib_error_free (ad->error);
   pool_put (axm->devices, ad);
 }
@@ -392,7 +442,8 @@ af_xdp_create_if (vlib_main_t * vm, af_xdp_create_if_args_t * args)
   vnet_sw_interface_t *sw;
   vnet_hw_interface_t *hw;
   int rxq_num, txq_num, q_num;
-  int i;
+  int ns_fds[2];
+  int i, ret;
 
   args->rxq_size = args->rxq_size ? args->rxq_size : 2 * VLIB_FRAME_SIZE;
   args->txq_size = args->txq_size ? args->txq_size : 2 * VLIB_FRAME_SIZE;
@@ -417,13 +468,22 @@ af_xdp_create_if (vlib_main_t * vm, af_xdp_create_if_args_t * args)
       goto err0;
     }
 
+  ret = af_xdp_enter_netns (args->netns, ns_fds);
+  if (ret)
+    {
+      args->rv = ret;
+      args->error = clib_error_return (0, "enter netns %s failed, ret %d",
+				       args->netns, args->rv);
+      goto err0;
+    }
+
   af_xdp_get_q_count (args->linux_ifname, &rxq_num, &txq_num);
   if (args->rxq_num > rxq_num && AF_XDP_NUM_RX_QUEUES_ALL != args->rxq_num)
     {
       args->rv = VNET_API_ERROR_INVALID_VALUE;
       args->error = clib_error_create ("too many rxq requested (%d > %d)",
 				       args->rxq_num, rxq_num);
-      goto err0;
+      goto err1;
     }
   rxq_num = clib_min (rxq_num, args->rxq_num);
   txq_num = clib_min (txq_num, tm->n_vlib_mains);
@@ -437,8 +497,10 @@ af_xdp_create_if (vlib_main_t * vm, af_xdp_create_if_args_t * args)
   ad->linux_ifname = (char *) format (0, "%s", args->linux_ifname);
   vec_validate (ad->linux_ifname, IFNAMSIZ - 1);	/* libbpf expects ifname to be at least IFNAMSIZ */
 
+  ad->netns = (char *) format (0, "%s", args->netns);
+
   if (args->prog && af_xdp_load_program (args, ad))
-    goto err1;
+    goto err2;
 
   q_num = clib_max (rxq_num, txq_num);
   ad->rxq_num = rxq_num;
@@ -476,7 +538,7 @@ af_xdp_create_if (vlib_main_t * vm, af_xdp_create_if_args_t * args)
 	      (i < rxq_num && AF_XDP_NUM_RX_QUEUES_ALL != args->rxq_num))
 	    {
 	      ad->rxq_num = ad->txq_num = 0;
-	      goto err1; /* failed creating requested rxq: fatal error, bailing
+	      goto err2; /* failed creating requested rxq: fatal error, bailing
 			    out */
 	    }
 
@@ -487,6 +549,13 @@ af_xdp_create_if (vlib_main_t * vm, af_xdp_create_if_args_t * args)
 	}
     }
 
+  if (af_xdp_exit_netns (args->netns, ns_fds))
+    {
+      args->rv = VNET_API_ERROR_SYSCALL_ERROR_10;
+      args->error = clib_error_return (0, "exit netns failed");
+      goto err2;
+    }
+
   if (ad->txq_num < tm->n_vlib_mains)
     {
       /* initialize lock for shared txq */
@@ -501,8 +570,16 @@ af_xdp_create_if (vlib_main_t * vm, af_xdp_create_if_args_t * args)
 					   af_xdp_get_numa
 					   (ad->linux_ifname));
   if (!args->name)
-    ad->name =
-      (char *) format (0, "%s/%d", ad->linux_ifname, ad->dev_instance);
+    {
+      char *ifname = ad->linux_ifname;
+      if (args->netns != NULL && strncmp (args->netns, "pid:", 4) == 0)
+	{
+	  ad->name =
+	    (char *) format (0, "%s/%u", ifname, atoi (args->netns + 4));
+	}
+      else
+	ad->name = (char *) format (0, "%s/%d", ifname, ad->dev_instance);
+    }
   else
     ad->name = (char *) format (0, "%s", args->name);
 
@@ -516,7 +593,7 @@ af_xdp_create_if (vlib_main_t * vm, af_xdp_create_if_args_t * args)
       args->rv = VNET_API_ERROR_INVALID_INTERFACE;
       args->error =
 	clib_error_return (0, "ethernet_register_interface() failed");
-      goto err1;
+      goto err2;
     }
 
   sw = vnet_get_hw_sw_interface (vnm, ad->hw_if_index);
@@ -543,7 +620,7 @@ af_xdp_create_if (vlib_main_t * vm, af_xdp_create_if_args_t * args)
       vnet_hw_if_set_rx_queue_file_index (vnm, rxq->queue_index,
 					  rxq->file_index);
       if (af_xdp_device_set_rxq_mode (ad, rxq, AF_XDP_RXQ_MODE_POLLING))
-	goto err1;
+	goto err2;
     }
 
   vnet_hw_if_update_runtime_data (vnm, ad->hw_if_index);
@@ -558,8 +635,10 @@ af_xdp_create_if (vlib_main_t * vm, af_xdp_create_if_args_t * args)
 
   return;
 
-err1:
+err2:
   af_xdp_delete_if (vm, ad);
+err1:
+  af_xdp_cleanup_netns (ns_fds);
 err0:
   vlib_log_err (am->log_class, "%U", format_clib_error, args->error);
 }
diff --git a/src/plugins/af_xdp/test_api.c b/src/plugins/af_xdp/test_api.c
index 6dffa29bdd1..46ba6f100ee 100644
--- a/src/plugins/af_xdp/test_api.c
+++ b/src/plugins/af_xdp/test_api.c
@@ -91,6 +91,41 @@ api_af_xdp_create (vat_main_t * vam)
   return ret;
 }
 
+/* af_xdp create v2 API */
+static int
+api_af_xdp_create_v2 (vat_main_t *vam)
+{
+  vl_api_af_xdp_create_v2_t *mp;
+  af_xdp_create_if_args_t args;
+  int ret;
+
+  if (!unformat_user (vam->input, unformat_af_xdp_create_if_args, &args))
+    {
+      clib_warning ("unknown input `%U'", format_unformat_error, vam->input);
+      return -99;
+    }
+
+  M (AF_XDP_CREATE, mp);
+
+  snprintf ((char *) mp->host_if, sizeof (mp->host_if), "%s",
+	    args.linux_ifname ?: "");
+  snprintf ((char *) mp->name, sizeof (mp->name), "%s", args.name ?: "");
+  snprintf ((char *) mp->namespace, sizeof (mp->namespace), "%s",
+	    args.netns ?: "");
+  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_af_xdp_mode (args.mode);
+  if (args.flags & AF_XDP_CREATE_FLAGS_NO_SYSCALL_LOCK)
+    mp->flags |= AF_XDP_API_FLAGS_NO_SYSCALL_LOCK;
+  snprintf ((char *) mp->prog, sizeof (mp->prog), "%s", args.prog ?: "");
+
+  S (mp);
+  W (ret);
+
+  return ret;
+}
+
 /* af_xdp-create reply handler */
 static void
 vl_api_af_xdp_create_reply_t_handler (vl_api_af_xdp_create_reply_t * mp)
@@ -109,6 +144,24 @@ vl_api_af_xdp_create_reply_t_handler (vl_api_af_xdp_create_reply_t * mp)
   vam->regenerate_interface_table = 1;
 }
 
+/* af_xdp-create v2 reply handler */
+static void
+vl_api_af_xdp_create_v2_reply_t_handler (vl_api_af_xdp_create_v2_reply_t *mp)
+{
+  vat_main_t *vam = af_xdp_test_main.vat_main;
+  i32 retval = ntohl (mp->retval);
+
+  if (retval == 0)
+    {
+      fformat (vam->ofp, "created af_xdp with sw_if_index %d\n",
+	       ntohl (mp->sw_if_index));
+    }
+
+  vam->retval = retval;
+  vam->result_ready = 1;
+  vam->regenerate_interface_table = 1;
+}
+
 /* af_xdp delete API */
 static int
 api_af_xdp_delete (vat_main_t * vam)
diff --git a/src/plugins/af_xdp/unformat.c b/src/plugins/af_xdp/unformat.c
index bb4c3048d23..8c0482d83ff 100644
--- a/src/plugins/af_xdp/unformat.c
+++ b/src/plugins/af_xdp/unformat.c
@@ -46,6 +46,8 @@ unformat_af_xdp_create_if_args (unformat_input_t * input, va_list * vargs)
 	;
       else if (unformat (line_input, "prog %s", &args->prog))
 	;
+      else if (unformat (line_input, "netns %s", &args->netns))
+	;
       else if (unformat (line_input, "no-zero-copy"))
 	args->mode = AF_XDP_MODE_COPY;
       else if (unformat (line_input, "zero-copy"))