Re: [PATCH v1 02/12] node: add IP4 lookup FIB node

[Nitin Saxena <nsaxena16@gmail.com>]

Hi Ankur,

Please see my comments inline below

Thanks,
Nitin

On Tue, Apr 15, 2025 at 5:41 PM Ankur Dwivedi <adwivedi@marvell.com> wrote:
>
> Adds a lookup FIB node for IP4.
>
> Signed-off-by: Ankur Dwivedi <adwivedi@marvell.com>
> ---
>  lib/node/ip4_lookup_fib.c | 127 ++++++++++++++++++++++++++++++++++++++
>  lib/node/meson.build      |   3 +-
>  2 files changed, 129 insertions(+), 1 deletion(-)
>  create mode 100644 lib/node/ip4_lookup_fib.c
>
> diff --git a/lib/node/ip4_lookup_fib.c b/lib/node/ip4_lookup_fib.c
> new file mode 100644
> index 0000000000..9c71610718
> --- /dev/null
> +++ b/lib/node/ip4_lookup_fib.c
> @@ -0,0 +1,127 @@
> +/* SPDX-License-Identifier: BSD-3-Clause
> + * Copyright(C) 2025 Marvell.
> + */
> +
> +#include <rte_errno.h>
> +#include <rte_ether.h>
> +#include <rte_fib.h>
> +#include <rte_graph.h>
> +#include <rte_graph_worker.h>
> +#include <rte_ip.h>
> +
> +#include "rte_node_ip4_api.h"
> +
> +#include "node_private.h"
> +
> +/* IP4 Lookup global data struct */
> +struct ip4_lookup_fib_node_main {
> +       struct rte_fib *fib[RTE_MAX_NUMA_NODES];
> +};
> +
> +struct ip4_lookup_fib_node_ctx {
> +       /* Socket's FIB */
> +       struct rte_fib *fib;
> +       /* Dynamic offset to mbuf priv1 */
> +       int mbuf_priv1_off;
> +};
> +
> +static struct ip4_lookup_fib_node_main ip4_lookup_fib_nm;
> +
> +#define FIB_MAX_ROUTES (1 << 16)
> +#define FIB_NUM_TBL8   (1 << 15)
> +#define FIB_DEFAULT_NH 999

These macros may not be required if we expose public setup_api() with
arguments. See below

> +
> +#define IP4_LOOKUP_NODE_FIB(ctx) \
> +       (((struct ip4_lookup_fib_node_ctx *)ctx)->fib)
> +
> +#define IP4_LOOKUP_NODE_PRIV1_OFF(ctx) \
> +       (((struct ip4_lookup_fib_node_ctx *)ctx)->mbuf_priv1_off)
> +
> +static int
> +setup_fib(unsigned int socket)

Should we add public API to allow applications to control MAX_ROUTES?
In a typical stack multiple fibs can be set up for each VRF (~~ port_id).

A public API:

int rte_ip4_lookup_fib_setup(int fib_index, int port_id, uint32_t
max_routes) where

For now we can assume fib_index == 0, is global fib table (can be
extended to VRF later). Socket_id can be determined from port_Id

> +{
> +       struct ip4_lookup_fib_node_main *nm = &ip4_lookup_fib_nm;
> +       struct rte_fib_conf conf;
> +       char s[RTE_FIB_NAMESIZE];
> +
> +       /* One fib per socket */
> +       if (nm->fib[socket])
> +               return 0;
> +
> +       conf.type = RTE_FIB_DIR24_8;
> +       conf.default_nh = FIB_DEFAULT_NH;

FIB_DEFAULT_NH can be defined in such a way, fast path can decode
next_edge from return of rte_fib_lookup_bulk()
Like
union {
   struct u64;
   struct {
       uint32_t next_hop_id;
       uint16_t  next_edge;
       uint16_t. reserved;
   };
} rte_ip4_lookup_fib_nexthop_t;

FIB_DEFAULT_NH should be set as

rte_ip4_lookup_fib_nexthop_t default_next_hop = {.next_edge =
IP4_FIB_LOOKUP_NEXT_DROP}

This way in fast path a return from fib_bulk() can be directly used to
send packet to pkt_drop node

> +       conf.max_routes = FIB_MAX_ROUTES;
> +       conf.rib_ext_sz = 0;
> +       conf.dir24_8.nh_sz = RTE_FIB_DIR24_8_4B;
> +       conf.dir24_8.num_tbl8 = FIB_NUM_TBL8;
> +       conf.flags = 0;
> +       snprintf(s, sizeof(s), "IPV4_LOOKUP_FIB_%d", socket);
> +       nm->fib[socket] = rte_fib_create(s, socket, &conf);
> +       if (nm->fib[socket] == NULL)
> +               return -rte_errno;
> +
> +       return 0;
> +}
> +
> +static int
> +ip4_lookup_fib_node_init(const struct rte_graph *graph, struct rte_node *node)
> +{
> +       static uint8_t init_once;
> +       unsigned int socket;
> +       uint16_t lcore_id;
> +       int rc;
> +
> +       RTE_BUILD_BUG_ON(sizeof(struct ip4_lookup_fib_node_ctx) > RTE_NODE_CTX_SZ);
> +
> +       if (!init_once) {
> +               node_mbuf_priv1_dynfield_offset = rte_mbuf_dynfield_register(
> +                               &node_mbuf_priv1_dynfield_desc);
> +               if (node_mbuf_priv1_dynfield_offset < 0)
> +                       return -rte_errno;

You may need to rebase this patch on top of
https://patches.dpdk.org/project/dpdk/patch/20250409135554.2180390-2-nsaxena@marvell.com/
for using global mbuf field

> +
> +               /* Setup FIB for all sockets */
> +               RTE_LCORE_FOREACH(lcore_id)

Instead can be use rte_socket_count() which allow to loop for socket
instead of port? Ideally we may need to add fib for virtual interfaces
or VRF (later)

> +               {
> +                       socket = rte_lcore_to_socket_id(lcore_id);
> +                       rc = setup_fib(socket);
> +                       if (rc) {
> +                               node_err("ip4_lookup_fib",
> +                                        "Failed to setup fib for sock %u, rc=%d",
> +                                        socket, rc);
> +                               return rc;
> +                       }
> +               }
> +               init_once = 1;
> +       }
> +
> +       /* Update socket's FIB and mbuf dyn priv1 offset in node ctx */
> +       IP4_LOOKUP_NODE_FIB(node->ctx) = ip4_lookup_fib_nm.fib[graph->socket];
> +       IP4_LOOKUP_NODE_PRIV1_OFF(node->ctx) = node_mbuf_priv1_dynfield_offset;
> +
> +       node_dbg("ip4_lookup_fib", "Initialized ip4_lookup_fib node");
> +
> +       return 0;
> +}
> +
> +static struct rte_node_xstats ip4_lookup_fib_xstats = {
> +       .nb_xstats = 1,
> +       .xstat_desc = {
> +               [0] = "ip4_lookup_fib_error",
> +       },
> +};
> +
> +static struct rte_node_register ip4_lookup_fib_node = {
> +       .name = "ip4_lookup_fib",
> +
> +       .init = ip4_lookup_fib_node_init,
> +       .xstats = &ip4_lookup_fib_xstats,
> +
> +       .nb_edges = RTE_NODE_IP4_LOOKUP_NEXT_PKT_DROP + 1,
> +       .next_nodes = {
> +               [RTE_NODE_IP4_LOOKUP_NEXT_IP4_LOCAL] = "ip4_local",
> +               [RTE_NODE_IP4_LOOKUP_NEXT_REWRITE] = "ip4_rewrite",
> +               [RTE_NODE_IP4_LOOKUP_NEXT_PKT_DROP] = "pkt_drop",
> +       },
> +};
> +
> +RTE_NODE_REGISTER(ip4_lookup_fib_node);
> diff --git a/lib/node/meson.build b/lib/node/meson.build
> index 0bed97a96c..d2011c8f56 100644
> --- a/lib/node/meson.build
> +++ b/lib/node/meson.build
> @@ -13,6 +13,7 @@ sources = files(
>          'ethdev_tx.c',
>          'ip4_local.c',
>          'ip4_lookup.c',
> +        'ip4_lookup_fib.c',
>          'ip4_reassembly.c',
>          'ip4_rewrite.c',
>          'ip6_lookup.c',
> @@ -34,4 +35,4 @@ headers = files(
>
>  # Strict-aliasing rules are violated by uint8_t[] to context size casts.
>  cflags += '-fno-strict-aliasing'
> -deps += ['graph', 'mbuf', 'lpm', 'ethdev', 'mempool', 'cryptodev', 'ip_frag']
> +deps += ['graph', 'mbuf', 'lpm', 'ethdev', 'mempool', 'cryptodev', 'ip_frag', 'fib']
> --
> 2.25.1
>