samples/bpf: Add new sample xsk_fwd.c

This sample code illustrates the packet forwarding between multiple
AF_XDP sockets in multi-threading environment. All the threads and
sockets are sharing a common buffer pool, with each socket having
its own private buffer cache. The sockets are created with the
xsk_socket__create_shared() function, which allows multiple AF_XDP
sockets to share the same UMEM object.

Example 1: Single thread handling two sockets. Packets received
from socket A (on top of interface IFA, queue QA) are forwarded
to socket B (on top of interface IFB, queue QB) and vice-versa.
The thread is affinitized to CPU core C:

./xsk_fwd -i IFA -q QA -i IFB -q QB -c C

Example 2: Two threads, each handling two sockets. Packets from
socket A are sent to socket B (by thread X), packets
from socket B are sent to socket A (by thread X); packets from
socket C are sent to socket D (by thread Y), packets from socket
D are sent to socket C (by thread Y). The two threads are bound
to CPU cores CX and CY:

./xdp_fwd -i IFA -q QA -i IFB -q QB -i IFC -q QC -i IFD -q QD -c CX -c CY
Signed-off-by: Cristian Dumitrescu <cristian.dumitrescu@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Björn Töpel <bjorn.topel@intel.com>
Link: https://lore.kernel.org/bpf/1598603189-32145-15-git-send-email-magnus.karlsson@intel.com

samples/bpf/Makefile

@@ -48,6 +48,7 @@ tprogs-y += syscall_tp
 tprogs-y += cpustat
 tprogs-y += xdp_adjust_tail
 tprogs-y += xdpsock
+tprogs-y += xsk_fwd
 tprogs-y += xdp_fwd
 tprogs-y += task_fd_query
 tprogs-y += xdp_sample_pkts
@@ -104,6 +105,7 @@ syscall_tp-objs := syscall_tp_user.o
 cpustat-objs := cpustat_user.o
 xdp_adjust_tail-objs := xdp_adjust_tail_user.o
 xdpsock-objs := xdpsock_user.o
+xsk_fwd-objs := xsk_fwd.o
 xdp_fwd-objs := xdp_fwd_user.o
 task_fd_query-objs := bpf_load.o task_fd_query_user.o $(TRACE_HELPERS)
 xdp_sample_pkts-objs := xdp_sample_pkts_user.o $(TRACE_HELPERS)
@@ -203,6 +205,7 @@ TPROGLDLIBS_trace_output	+= -lrt
 TPROGLDLIBS_map_perf_test	+= -lrt
 TPROGLDLIBS_test_overhead	+= -lrt
 TPROGLDLIBS_xdpsock		+= -pthread
+TPROGLDLIBS_xsk_fwd		+= -pthread
 
 # Allows pointing LLC/CLANG to a LLVM backend with bpf support, redefine on cmdline:
 #  make M=samples/bpf/ LLC=~/git/llvm/build/bin/llc CLANG=~/git/llvm/build/bin/clang

samples/bpf/xsk_fwd.c

+// SPDX-License-Identifier: GPL-2.0
+/* Copyright(c) 2020 Intel Corporation. */
+
+#define _GNU_SOURCE
+#include <poll.h>
+#include <pthread.h>
+#include <signal.h>
+#include <sched.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/mman.h>
+#include <sys/resource.h>
+#include <sys/socket.h>
+#include <sys/types.h>
+#include <time.h>
+#include <unistd.h>
+#include <getopt.h>
+#include <netinet/ether.h>
+#include <net/if.h>
+
+#include <linux/bpf.h>
+#include <linux/if_link.h>
+#include <linux/if_xdp.h>
+
+#include <bpf/libbpf.h>
+#include <bpf/xsk.h>
+#include <bpf/bpf.h>
+
+#define ARRAY_SIZE(x) (sizeof(x) / sizeof((x)[0]))
+
+typedef __u64 u64;
+typedef __u32 u32;
+typedef __u16 u16;
+typedef __u8  u8;
+
+/* This program illustrates the packet forwarding between multiple AF_XDP
+ * sockets in multi-threaded environment. All threads are sharing a common
+ * buffer pool, with each socket having its own private buffer cache.
+ *
+ * Example 1: Single thread handling two sockets. The packets received by socket
+ * A (interface IFA, queue QA) are forwarded to socket B (interface IFB, queue
+ * QB), while the packets received by socket B are forwarded to socket A. The
+ * thread is running on CPU core X:
+ *
+ *         ./xsk_fwd -i IFA -q QA -i IFB -q QB -c X
+ *
+ * Example 2: Two threads, each handling two sockets. The thread running on CPU
+ * core X forwards all the packets received by socket A to socket B, and all the
+ * packets received by socket B to socket A. The thread running on CPU core Y is
+ * performing the same packet forwarding between sockets C and D:
+ *
+ *         ./xsk_fwd -i IFA -q QA -i IFB -q QB -i IFC -q QC -i IFD -q QD
+ *         -c CX -c CY
+ */
+
+/*
+ * Buffer pool and buffer cache
+ *
+ * For packet forwarding, the packet buffers are typically allocated from the
+ * pool for packet reception and freed back to the pool for further reuse once
+ * the packet transmission is completed.
+ *
+ * The buffer pool is shared between multiple threads. In order to minimize the
+ * access latency to the shared buffer pool, each thread creates one (or
+ * several) buffer caches, which, unlike the buffer pool, are private to the
+ * thread that creates them and therefore cannot be shared with other threads.
+ * The access to the shared pool is only needed either (A) when the cache gets
+ * empty due to repeated buffer allocations and it needs to be replenished from
+ * the pool, or (B) when the cache gets full due to repeated buffer free and it
+ * needs to be flushed back to the pull.
+ *
+ * In a packet forwarding system, a packet received on any input port can
+ * potentially be transmitted on any output port, depending on the forwarding
+ * configuration. For AF_XDP sockets, for this to work with zero-copy of the
+ * packet buffers when, it is required that the buffer pool memory fits into the
+ * UMEM area shared by all the sockets.
+ */
+
+struct bpool_params {
+	u32 n_buffers;
+	u32 buffer_size;
+	int mmap_flags;
+
+	u32 n_users_max;
+	u32 n_buffers_per_slab;
+};
+
+/* This buffer pool implementation organizes the buffers into equally sized
+ * slabs of *n_buffers_per_slab*. Initially, there are *n_slabs* slabs in the
+ * pool that are completely filled with buffer pointers (full slabs).
+ *
+ * Each buffer cache has a slab for buffer allocation and a slab for buffer
+ * free, with both of these slabs initially empty. When the cache's allocation
+ * slab goes empty, it is swapped with one of the available full slabs from the
+ * pool, if any is available. When the cache's free slab goes full, it is
+ * swapped for one of the empty slabs from the pool, which is guaranteed to
+ * succeed.
+ *
+ * Partially filled slabs never get traded between the cache and the pool
+ * (except when the cache itself is destroyed), which enables fast operation
+ * through pointer swapping.
+ */
+struct bpool {
+	struct bpool_params params;
+	pthread_mutex_t lock;
+	void *addr;
+
+	u64 **slabs;
+	u64 **slabs_reserved;
+	u64 *buffers;
+	u64 *buffers_reserved;
+
+	u64 n_slabs;
+	u64 n_slabs_reserved;
+	u64 n_buffers;
+
+	u64 n_slabs_available;
+	u64 n_slabs_reserved_available;
+
+	struct xsk_umem_config umem_cfg;
+	struct xsk_ring_prod umem_fq;
+	struct xsk_ring_cons umem_cq;
+	struct xsk_umem *umem;
+};
+
+static struct bpool *
+bpool_init(struct bpool_params *params,
+	   struct xsk_umem_config *umem_cfg)
+{
+	struct rlimit r = {RLIM_INFINITY, RLIM_INFINITY};
+	u64 n_slabs, n_slabs_reserved, n_buffers, n_buffers_reserved;
+	u64 slabs_size, slabs_reserved_size;
+	u64 buffers_size, buffers_reserved_size;
+	u64 total_size, i;
+	struct bpool *bp;
+	u8 *p;
+	int status;
+
+	/* mmap prep. */
+	if (setrlimit(RLIMIT_MEMLOCK, &r))
+		return NULL;
+
+	/* bpool internals dimensioning. */
+	n_slabs = (params->n_buffers + params->n_buffers_per_slab - 1) /
+		params->n_buffers_per_slab;
+	n_slabs_reserved = params->n_users_max * 2;
+	n_buffers = n_slabs * params->n_buffers_per_slab;
+	n_buffers_reserved = n_slabs_reserved * params->n_buffers_per_slab;
+
+	slabs_size = n_slabs * sizeof(u64 *);
+	slabs_reserved_size = n_slabs_reserved * sizeof(u64 *);
+	buffers_size = n_buffers * sizeof(u64);
+	buffers_reserved_size = n_buffers_reserved * sizeof(u64);
+
+	total_size = sizeof(struct bpool) +
+		slabs_size + slabs_reserved_size +
+		buffers_size + buffers_reserved_size;
+
+	/* bpool memory allocation. */
+	p = calloc(total_size, sizeof(u8));
+	if (!p)
+		return NULL;
+
+	/* bpool memory initialization. */
+	bp = (struct bpool *)p;
+	memcpy(&bp->params, params, sizeof(*params));
+	bp->params.n_buffers = n_buffers;
+
+	bp->slabs = (u64 **)&p[sizeof(struct bpool)];
+	bp->slabs_reserved = (u64 **)&p[sizeof(struct bpool) +
+		slabs_size];
+	bp->buffers = (u64 *)&p[sizeof(struct bpool) +
+		slabs_size + slabs_reserved_size];
+	bp->buffers_reserved = (u64 *)&p[sizeof(struct bpool) +
+		slabs_size + slabs_reserved_size + buffers_size];
+
+	bp->n_slabs = n_slabs;
+	bp->n_slabs_reserved = n_slabs_reserved;
+	bp->n_buffers = n_buffers;
+
+	for (i = 0; i < n_slabs; i++)
+		bp->slabs[i] = &bp->buffers[i * params->n_buffers_per_slab];
+	bp->n_slabs_available = n_slabs;
+
+	for (i = 0; i < n_slabs_reserved; i++)
+		bp->slabs_reserved[i] = &bp->buffers_reserved[i *
+			params->n_buffers_per_slab];
+	bp->n_slabs_reserved_available = n_slabs_reserved;
+
+	for (i = 0; i < n_buffers; i++)
+		bp->buffers[i] = i * params->buffer_size;
+
+	/* lock. */
+	status = pthread_mutex_init(&bp->lock, NULL);
+	if (status) {
+		free(p);
+		return NULL;
+	}
+
+	/* mmap. */
+	bp->addr = mmap(NULL,
+			n_buffers * params->buffer_size,
+			PROT_READ | PROT_WRITE,
+			MAP_PRIVATE | MAP_ANONYMOUS | params->mmap_flags,
+			-1,
+			0);
+	if (bp->addr == MAP_FAILED) {
+		pthread_mutex_destroy(&bp->lock);
+		free(p);
+		return NULL;
+	}
+
+	/* umem. */
+	status = xsk_umem__create(&bp->umem,
+				  bp->addr,
+				  bp->params.n_buffers * bp->params.buffer_size,
+				  &bp->umem_fq,
+				  &bp->umem_cq,
+				  umem_cfg);
+	if (status) {
+		munmap(bp->addr, bp->params.n_buffers * bp->params.buffer_size);
+		pthread_mutex_destroy(&bp->lock);
+		free(p);
+		return NULL;
+	}
+	memcpy(&bp->umem_cfg, umem_cfg, sizeof(*umem_cfg));
+
+	return bp;
+}
+
+static void
+bpool_free(struct bpool *bp)
+{
+	if (!bp)
+		return;
+
+	xsk_umem__delete(bp->umem);
+	munmap(bp->addr, bp->params.n_buffers * bp->params.buffer_size);
+	pthread_mutex_destroy(&bp->lock);
+	free(bp);
+}
+
+struct bcache {
+	struct bpool *bp;
+
+	u64 *slab_cons;
+	u64 *slab_prod;
+
+	u64 n_buffers_cons;
+	u64 n_buffers_prod;
+};
+
+static u32
+bcache_slab_size(struct bcache *bc)
+{
+	struct bpool *bp = bc->bp;
+
+	return bp->params.n_buffers_per_slab;
+}
+
+static struct bcache *
+bcache_init(struct bpool *bp)
+{
+	struct bcache *bc;
+
+	bc = calloc(1, sizeof(struct bcache));
+	if (!bc)
+		return NULL;
+
+	bc->bp = bp;
+	bc->n_buffers_cons = 0;
+	bc->n_buffers_prod = 0;
+
+	pthread_mutex_lock(&bp->lock);
+	if (bp->n_slabs_reserved_available == 0) {
+		pthread_mutex_unlock(&bp->lock);
+		free(bc);
+		return NULL;
+	}
+
+	bc->slab_cons = bp->slabs_reserved[bp->n_slabs_reserved_available - 1];
+	bc->slab_prod = bp->slabs_reserved[bp->n_slabs_reserved_available - 2];
+	bp->n_slabs_reserved_available -= 2;
+	pthread_mutex_unlock(&bp->lock);
+
+	return bc;
+}
+
+static void
+bcache_free(struct bcache *bc)
+{
+	struct bpool *bp;
+
+	if (!bc)
+		return;
+
+	/* In order to keep this example simple, the case of freeing any
+	 * existing buffers from the cache back to the pool is ignored.
+	 */
+
+	bp = bc->bp;
+	pthread_mutex_lock(&bp->lock);
+	bp->slabs_reserved[bp->n_slabs_reserved_available] = bc->slab_prod;
+	bp->slabs_reserved[bp->n_slabs_reserved_available + 1] = bc->slab_cons;
+	bp->n_slabs_reserved_available += 2;
+	pthread_mutex_unlock(&bp->lock);
+
+	free(bc);
+}
+
+/* To work correctly, the implementation requires that the *n_buffers* input
+ * argument is never greater than the buffer pool's *n_buffers_per_slab*. This
+ * is typically the case, with one exception taking place when large number of
+ * buffers are allocated at init time (e.g. for the UMEM fill queue setup).
+ */
+static inline u32
+bcache_cons_check(struct bcache *bc, u32 n_buffers)
+{
+	struct bpool *bp = bc->bp;
+	u64 n_buffers_per_slab = bp->params.n_buffers_per_slab;
+	u64 n_buffers_cons = bc->n_buffers_cons;
+	u64 n_slabs_available;
+	u64 *slab_full;
+
+	/*
+	 * Consumer slab is not empty: Use what's available locally. Do not
+	 * look for more buffers from the pool when the ask can only be
+	 * partially satisfied.
+	 */
+	if (n_buffers_cons)
+		return (n_buffers_cons < n_buffers) ?
+			n_buffers_cons :
+			n_buffers;
+
+	/*
+	 * Consumer slab is empty: look to trade the current consumer slab
+	 * (full) for a full slab from the pool, if any is available.
+	 */
+	pthread_mutex_lock(&bp->lock);
+	n_slabs_available = bp->n_slabs_available;
+	if (!n_slabs_available) {
+		pthread_mutex_unlock(&bp->lock);
+		return 0;
+	}
+
+	n_slabs_available--;
+	slab_full = bp->slabs[n_slabs_available];
+	bp->slabs[n_slabs_available] = bc->slab_cons;
+	bp->n_slabs_available = n_slabs_available;
+	pthread_mutex_unlock(&bp->lock);
+
+	bc->slab_cons = slab_full;
+	bc->n_buffers_cons = n_buffers_per_slab;
+	return n_buffers;
+}
+
+static inline u64
+bcache_cons(struct bcache *bc)
+{
+	u64 n_buffers_cons = bc->n_buffers_cons - 1;
+	u64 buffer;
+
+	buffer = bc->slab_cons[n_buffers_cons];
+	bc->n_buffers_cons = n_buffers_cons;
+	return buffer;
+}
+
+static inline void
+bcache_prod(struct bcache *bc, u64 buffer)
+{
+	struct bpool *bp = bc->bp;
+	u64 n_buffers_per_slab = bp->params.n_buffers_per_slab;
+	u64 n_buffers_prod = bc->n_buffers_prod;
+	u64 n_slabs_available;
+	u64 *slab_empty;
+
+	/*
+	 * Producer slab is not yet full: store the current buffer to it.
+	 */
+	if (n_buffers_prod < n_buffers_per_slab) {
+		bc->slab_prod[n_buffers_prod] = buffer;
+		bc->n_buffers_prod = n_buffers_prod + 1;
+		return;
+	}
+
+	/*
+	 * Producer slab is full: trade the cache's current producer slab
+	 * (full) for an empty slab from the pool, then store the current
+	 * buffer to the new producer slab. As one full slab exists in the
+	 * cache, it is guaranteed that there is at least one empty slab
+	 * available in the pool.
+	 */
+	pthread_mutex_lock(&bp->lock);
+	n_slabs_available = bp->n_slabs_available;
+	slab_empty = bp->slabs[n_slabs_available];
+	bp->slabs[n_slabs_available] = bc->slab_prod;
+	bp->n_slabs_available = n_slabs_available + 1;
+	pthread_mutex_unlock(&bp->lock);
+
+	slab_empty[0] = buffer;
+	bc->slab_prod = slab_empty;
+	bc->n_buffers_prod = 1;
+}
+
+/*
+ * Port
+ *
+ * Each of the forwarding ports sits on top of an AF_XDP socket. In order for
+ * packet forwarding to happen with no packet buffer copy, all the sockets need
+ * to share the same UMEM area, which is used as the buffer pool memory.
+ */
+#ifndef MAX_BURST_RX
+#define MAX_BURST_RX 64
+#endif
+
+#ifndef MAX_BURST_TX
+#define MAX_BURST_TX 64
+#endif
+
+struct burst_rx {
+	u64 addr[MAX_BURST_RX];
+	u32 len[MAX_BURST_RX];
+};
+
+struct burst_tx {
+	u64 addr[MAX_BURST_TX];
+	u32 len[MAX_BURST_TX];
+	u32 n_pkts;
+};
+
+struct port_params {
+	struct xsk_socket_config xsk_cfg;
+	struct bpool *bp;
+	const char *iface;
+	u32 iface_queue;
+};
+
+struct port {
+	struct port_params params;
+
+	struct bcache *bc;
+
+	struct xsk_ring_cons rxq;
+	struct xsk_ring_prod txq;
+	struct xsk_ring_prod umem_fq;
+	struct xsk_ring_cons umem_cq;
+	struct xsk_socket *xsk;
+	int umem_fq_initialized;
+
+	u64 n_pkts_rx;
+	u64 n_pkts_tx;
+};
+
+static void
+port_free(struct port *p)
+{
+	if (!p)
+		return;
+
+	/* To keep this example simple, the code to free the buffers from the
+	 * socket's receive and transmit queues, as well as from the UMEM fill
+	 * and completion queues, is not included.
+	 */
+
+	if (p->xsk)
+		xsk_socket__delete(p->xsk);
+
+	bcache_free(p->bc);
+
+	free(p);
+}
+
+static struct port *
+port_init(struct port_params *params)
+{
+	struct port *p;
+	u32 umem_fq_size, pos = 0;
+	int status, i;
+
+	/* Memory allocation and initialization. */
+	p = calloc(sizeof(struct port), 1);
+	if (!p)
+		return NULL;
+
+	memcpy(&p->params, params, sizeof(p->params));
+	umem_fq_size = params->bp->umem_cfg.fill_size;
+
+	/* bcache. */
+	p->bc = bcache_init(params->bp);
+	if (!p->bc ||
+	    (bcache_slab_size(p->bc) < umem_fq_size) ||
+	    (bcache_cons_check(p->bc, umem_fq_size) < umem_fq_size)) {
+		port_free(p);
+		return NULL;
+	}
+
+	/* xsk socket. */
+	status = xsk_socket__create_shared(&p->xsk,
+					   params->iface,
+					   params->iface_queue,
+					   params->bp->umem,
+					   &p->rxq,
+					   &p->txq,
+					   &p->umem_fq,
+					   &p->umem_cq,
+					   &params->xsk_cfg);
+	if (status) {
+		port_free(p);
+		return NULL;
+	}
+
+	/* umem fq. */
+	xsk_ring_prod__reserve(&p->umem_fq, umem_fq_size, &pos);
+
+	for (i = 0; i < umem_fq_size; i++)
+		*xsk_ring_prod__fill_addr(&p->umem_fq, pos + i) =
+			bcache_cons(p->bc);
+
+	xsk_ring_prod__submit(&p->umem_fq, umem_fq_size);
+	p->umem_fq_initialized = 1;
+
+	return p;
+}
+
+static inline u32
+port_rx_burst(struct port *p, struct burst_rx *b)
+{
+	u32 n_pkts, pos, i;
+
+	/* Free buffers for FQ replenish. */
+	n_pkts = ARRAY_SIZE(b->addr);
+
+	n_pkts = bcache_cons_check(p->bc, n_pkts);
+	if (!n_pkts)
+		return 0;
+
+	/* RXQ. */
+	n_pkts = xsk_ring_cons__peek(&p->rxq, n_pkts, &pos);
+	if (!n_pkts) {
+		if (xsk_ring_prod__needs_wakeup(&p->umem_fq)) {
+			struct pollfd pollfd = {
+				.fd = xsk_socket__fd(p->xsk),
+				.events = POLLIN,
+			};
+
+			poll(&pollfd, 1, 0);
+		}
+		return 0;
+	}
+
+	for (i = 0; i < n_pkts; i++) {
+		b->addr[i] = xsk_ring_cons__rx_desc(&p->rxq, pos + i)->addr;
+		b->len[i] = xsk_ring_cons__rx_desc(&p->rxq, pos + i)->len;
+	}
+
+	xsk_ring_cons__release(&p->rxq, n_pkts);
+	p->n_pkts_rx += n_pkts;
+
+	/* UMEM FQ. */
+	for ( ; ; ) {
+		int status;
+
+		status = xsk_ring_prod__reserve(&p->umem_fq, n_pkts, &pos);
+		if (status == n_pkts)
+			break;
+
+		if (xsk_ring_prod__needs_wakeup(&p->umem_fq)) {
+			struct pollfd pollfd = {
+				.fd = xsk_socket__fd(p->xsk),
+				.events = POLLIN,
+			};
+
+			poll(&pollfd, 1, 0);
+		}
+	}
+
+	for (i = 0; i < n_pkts; i++)
+		*xsk_ring_prod__fill_addr(&p->umem_fq, pos + i) =
+			bcache_cons(p->bc);
+
+	xsk_ring_prod__submit(&p->umem_fq, n_pkts);
+
+	return n_pkts;
+}
+
+static inline void
+port_tx_burst(struct port *p, struct burst_tx *b)
+{
+	u32 n_pkts, pos, i;
+	int status;
+
+	/* UMEM CQ. */
+	n_pkts = p->params.bp->umem_cfg.comp_size;
+
+	n_pkts = xsk_ring_cons__peek(&p->umem_cq, n_pkts, &pos);
+
+	for (i = 0; i < n_pkts; i++) {
+		u64 addr = *xsk_ring_cons__comp_addr(&p->umem_cq, pos + i);
+
+		bcache_prod(p->bc, addr);
+	}
+
+	xsk_ring_cons__release(&p->umem_cq, n_pkts);
+
+	/* TXQ. */
+	n_pkts = b->n_pkts;
+
+	for ( ; ; ) {
+		status = xsk_ring_prod__reserve(&p->txq, n_pkts, &pos);
+		if (status == n_pkts)
+			break;
+
+		if (xsk_ring_prod__needs_wakeup(&p->txq))
+			sendto(xsk_socket__fd(p->xsk), NULL, 0, MSG_DONTWAIT,
+			       NULL, 0);
+	}
+
+	for (i = 0; i < n_pkts; i++) {
+		xsk_ring_prod__tx_desc(&p->txq, pos + i)->addr = b->addr[i];
+		xsk_ring_prod__tx_desc(&p->txq, pos + i)->len = b->len[i];
+	}
+
+	xsk_ring_prod__submit(&p->txq, n_pkts);
+	if (xsk_ring_prod__needs_wakeup(&p->txq))
+		sendto(xsk_socket__fd(p->xsk), NULL, 0, MSG_DONTWAIT, NULL, 0);
+	p->n_pkts_tx += n_pkts;
+}
+
+/*
+ * Thread
+ *
+ * Packet forwarding threads.
+ */
+#ifndef MAX_PORTS_PER_THREAD
+#define MAX_PORTS_PER_THREAD 16
+#endif
+
+struct thread_data {
+	struct port *ports_rx[MAX_PORTS_PER_THREAD];
+	struct port *ports_tx[MAX_PORTS_PER_THREAD];
+	u32 n_ports_rx;
+	struct burst_rx burst_rx;
+	struct burst_tx burst_tx[MAX_PORTS_PER_THREAD];
+	u32 cpu_core_id;
+	int quit;
+};
+
+static void swap_mac_addresses(void *data)
+{
+	struct ether_header *eth = (struct ether_header *)data;
+	struct ether_addr *src_addr = (struct ether_addr *)&eth->ether_shost;
+	struct ether_addr *dst_addr = (struct ether_addr *)&eth->ether_dhost;
+	struct ether_addr tmp;
+
+	tmp = *src_addr;
+	*src_addr = *dst_addr;
+	*dst_addr = tmp;
+}
+
+static void *
+thread_func(void *arg)
+{
+	struct thread_data *t = arg;
+	cpu_set_t cpu_cores;
+	u32 i;
+
+	CPU_ZERO(&cpu_cores);
+	CPU_SET(t->cpu_core_id, &cpu_cores);
+	pthread_setaffinity_np(pthread_self(), sizeof(cpu_set_t), &cpu_cores);
+
+	for (i = 0; !t->quit; i = (i + 1) & (t->n_ports_rx - 1)) {
+		struct port *port_rx = t->ports_rx[i];
+		struct port *port_tx = t->ports_tx[i];
+		struct burst_rx *brx = &t->burst_rx;
+		struct burst_tx *btx = &t->burst_tx[i];
+		u32 n_pkts, j;
+
+		/* RX. */
+		n_pkts = port_rx_burst(port_rx, brx);
+		if (!n_pkts)
+			continue;
+
+		/* Process & TX. */
+		for (j = 0; j < n_pkts; j++) {
+			u64 addr = xsk_umem__add_offset_to_addr(brx->addr[j]);
+			u8 *pkt = xsk_umem__get_data(port_rx->params.bp->addr,
+						     addr);
+
+			swap_mac_addresses(pkt);
+
+			btx->addr[btx->n_pkts] = brx->addr[j];
+			btx->len[btx->n_pkts] = brx->len[j];
+			btx->n_pkts++;
+
+			if (btx->n_pkts == MAX_BURST_TX) {
+				port_tx_burst(port_tx, btx);
+				btx->n_pkts = 0;
+			}
+		}
+	}
+
+	return NULL;
+}
+
+/*
+ * Process
+ */
+static const struct bpool_params bpool_params_default = {
+	.n_buffers = 64 * 1024,
+	.buffer_size = XSK_UMEM__DEFAULT_FRAME_SIZE,
+	.mmap_flags = 0,
+
+	.n_users_max = 16,
+	.n_buffers_per_slab = XSK_RING_PROD__DEFAULT_NUM_DESCS * 2,
+};
+
+static const struct xsk_umem_config umem_cfg_default = {
+	.fill_size = XSK_RING_PROD__DEFAULT_NUM_DESCS * 2,
+	.comp_size = XSK_RING_CONS__DEFAULT_NUM_DESCS,
+	.frame_size = XSK_UMEM__DEFAULT_FRAME_SIZE,
+	.frame_headroom = XSK_UMEM__DEFAULT_FRAME_HEADROOM,
+	.flags = 0,
+};
+
+static const struct port_params port_params_default = {
+	.xsk_cfg = {
+		.rx_size = XSK_RING_CONS__DEFAULT_NUM_DESCS,
+		.tx_size = XSK_RING_PROD__DEFAULT_NUM_DESCS,
+		.libbpf_flags = 0,
+		.xdp_flags = XDP_FLAGS_DRV_MODE,
+		.bind_flags = XDP_USE_NEED_WAKEUP | XDP_ZEROCOPY,
+	},
+
+	.bp = NULL,
+	.iface = NULL,
+	.iface_queue = 0,
+};
+
+#ifndef MAX_PORTS
+#define MAX_PORTS 64
+#endif
+
+#ifndef MAX_THREADS
+#define MAX_THREADS 64
+#endif
+
+static struct bpool_params bpool_params;
+static struct xsk_umem_config umem_cfg;
+static struct bpool *bp;
+
+static struct port_params port_params[MAX_PORTS];
+static struct port *ports[MAX_PORTS];
+static u64 n_pkts_rx[MAX_PORTS];
+static u64 n_pkts_tx[MAX_PORTS];
+static int n_ports;
+
+static pthread_t threads[MAX_THREADS];
+static struct thread_data thread_data[MAX_THREADS];
+static int n_threads;
+
+static void
+print_usage(char *prog_name)
+{
+	const char *usage =
+		"Usage:\n"
+		"\t%s [ -b SIZE ] -c CORE -i INTERFACE [ -q QUEUE ]\n"
+		"\n"
+		"-c CORE        CPU core to run a packet forwarding thread\n"
+		"               on. May be invoked multiple times.\n"
+		"\n"
+		"-b SIZE        Number of buffers in the buffer pool shared\n"
+		"               by all the forwarding threads. Default: %u.\n"
+		"\n"
+		"-i INTERFACE   Network interface. Each (INTERFACE, QUEUE)\n"
+		"               pair specifies one forwarding port. May be\n"
+		"               invoked multiple times.\n"
+		"\n"
+		"-q QUEUE       Network interface queue for RX and TX. Each\n"
+		"               (INTERFACE, QUEUE) pair specified one\n"
+		"               forwarding port. Default: %u. May be invoked\n"
+		"               multiple times.\n"
+		"\n";
+	printf(usage,
+	       prog_name,
+	       bpool_params_default.n_buffers,
+	       port_params_default.iface_queue);
+}
+
+static int
+parse_args(int argc, char **argv)
+{
+	struct option lgopts[] = {
+		{ NULL,  0, 0, 0 }
+	};
+	int opt, option_index;
+
+	/* Parse the input arguments. */
+	for ( ; ;) {
+		opt = getopt_long(argc, argv, "c:i:q:", lgopts, &option_index);
+		if (opt == EOF)
+			break;
+
+		switch (opt) {
+		case 'b':
+			bpool_params.n_buffers = atoi(optarg);
+			break;
+
+		case 'c':
+			if (n_threads == MAX_THREADS) {
+				printf("Max number of threads (%d) reached.\n",
+				       MAX_THREADS);
+				return -1;
+			}
+
+			thread_data[n_threads].cpu_core_id = atoi(optarg);
+			n_threads++;
+			break;
+
+		case 'i':
+			if (n_ports == MAX_PORTS) {
+				printf("Max number of ports (%d) reached.\n",
+				       MAX_PORTS);
+				return -1;
+			}
+
+			port_params[n_ports].iface = optarg;
+			port_params[n_ports].iface_queue = 0;
+			n_ports++;
+			break;
+
+		case 'q':
+			if (n_ports == 0) {
+				printf("No port specified for queue.\n");
+				return -1;
+			}
+			port_params[n_ports - 1].iface_queue = atoi(optarg);
+			break;
+
+		default:
+			printf("Illegal argument.\n");
+			return -1;
+		}
+	}
+
+	optind = 1; /* reset getopt lib */
+
+	/* Check the input arguments. */
+	if (!n_ports) {
+		printf("No ports specified.\n");
+		return -1;
+	}
+
+	if (!n_threads) {
+		printf("No threads specified.\n");
+		return -1;
+	}
+
+	if (n_ports % n_threads) {
+		printf("Ports cannot be evenly distributed to threads.\n");
+		return -1;
+	}
+
+	return 0;
+}
+
+static void
+print_port(u32 port_id)
+{
+	struct port *port = ports[port_id];
+
+	printf("Port %u: interface = %s, queue = %u\n",
+	       port_id, port->params.iface, port->params.iface_queue);
+}
+
+static void
+print_thread(u32 thread_id)
+{
+	struct thread_data *t = &thread_data[thread_id];
+	u32 i;
+
+	printf("Thread %u (CPU core %u): ",
+	       thread_id, t->cpu_core_id);
+
+	for (i = 0; i < t->n_ports_rx; i++) {
+		struct port *port_rx = t->ports_rx[i];
+		struct port *port_tx = t->ports_tx[i];
+
+		printf("(%s, %u) -> (%s, %u), ",
+		       port_rx->params.iface,
+		       port_rx->params.iface_queue,
+		       port_tx->params.iface,
+		       port_tx->params.iface_queue);
+	}
+
+	printf("\n");
+}
+
+static void
+print_port_stats_separator(void)
+{
+	printf("+-%4s-+-%12s-+-%13s-+-%12s-+-%13s-+\n",
+	       "----",
+	       "------------",
+	       "-------------",
+	       "------------",
+	       "-------------");
+}
+
+static void
+print_port_stats_header(void)
+{
+	print_port_stats_separator();
+	printf("| %4s | %12s | %13s | %12s | %13s |\n",
+	       "Port",
+	       "RX packets",
+	       "RX rate (pps)",
+	       "TX packets",
+	       "TX_rate (pps)");
+	print_port_stats_separator();
+}
+
+static void
+print_port_stats_trailer(void)
+{
+	print_port_stats_separator();
+	printf("\n");
+}
+
+static void
+print_port_stats(int port_id, u64 ns_diff)
+{
+	struct port *p = ports[port_id];
+	double rx_pps, tx_pps;
+
+	rx_pps = (p->n_pkts_rx - n_pkts_rx[port_id]) * 1000000000. / ns_diff;
+	tx_pps = (p->n_pkts_tx - n_pkts_tx[port_id]) * 1000000000. / ns_diff;
+
+	printf("| %4d | %12llu | %13.0f | %12llu | %13.0f |\n",
+	       port_id,
+	       p->n_pkts_rx,
+	       rx_pps,
+	       p->n_pkts_tx,
+	       tx_pps);
+
+	n_pkts_rx[port_id] = p->n_pkts_rx;
+	n_pkts_tx[port_id] = p->n_pkts_tx;
+}
+
+static void
+print_port_stats_all(u64 ns_diff)
+{
+	int i;
+
+	print_port_stats_header();
+	for (i = 0; i < n_ports; i++)
+		print_port_stats(i, ns_diff);
+	print_port_stats_trailer();
+}
+
+static int quit;
+
+static void
+signal_handler(int sig)
+{
+	quit = 1;
+}
+
+static void remove_xdp_program(void)
+{
+	int i;
+
+	for (i = 0 ; i < n_ports; i++)
+		bpf_set_link_xdp_fd(if_nametoindex(port_params[i].iface), -1,
+				    port_params[i].xsk_cfg.xdp_flags);
+}
+
+int main(int argc, char **argv)
+{
+	struct timespec time;
+	u64 ns0;
+	int i;
+
+	/* Parse args. */
+	memcpy(&bpool_params, &bpool_params_default,
+	       sizeof(struct bpool_params));
+	memcpy(&umem_cfg, &umem_cfg_default,
+	       sizeof(struct xsk_umem_config));
+	for (i = 0; i < MAX_PORTS; i++)
+		memcpy(&port_params[i], &port_params_default,
+		       sizeof(struct port_params));
+
+	if (parse_args(argc, argv)) {
+		print_usage(argv[0]);
+		return -1;
+	}
+
+	/* Buffer pool initialization. */
+	bp = bpool_init(&bpool_params, &umem_cfg);
+	if (!bp) {
+		printf("Buffer pool initialization failed.\n");
+		return -1;
+	}
+	printf("Buffer pool created successfully.\n");
+
+	/* Ports initialization. */
+	for (i = 0; i < MAX_PORTS; i++)
+		port_params[i].bp = bp;
+
+	for (i = 0; i < n_ports; i++) {
+		ports[i] = port_init(&port_params[i]);
+		if (!ports[i]) {
+			printf("Port %d initialization failed.\n", i);
+			return -1;
+		}
+		print_port(i);
+	}
+	printf("All ports created successfully.\n");
+
+	/* Threads. */
+	for (i = 0; i < n_threads; i++) {
+		struct thread_data *t = &thread_data[i];
+		u32 n_ports_per_thread = n_ports / n_threads, j;
+
+		for (j = 0; j < n_ports_per_thread; j++) {
+			t->ports_rx[j] = ports[i * n_ports_per_thread + j];
+			t->ports_tx[j] = ports[i * n_ports_per_thread +
+				(j + 1) % n_ports_per_thread];
+		}
+
+		t->n_ports_rx = n_ports_per_thread;
+
+		print_thread(i);
+	}
+
+	for (i = 0; i < n_threads; i++) {
+		int status;
+
+		status = pthread_create(&threads[i],
+					NULL,
+					thread_func,
+					&thread_data[i]);
+		if (status) {
+			printf("Thread %d creation failed.\n", i);
+			return -1;
+		}
+	}
+	printf("All threads created successfully.\n");
+
+	/* Print statistics. */
+	signal(SIGINT, signal_handler);
+	signal(SIGTERM, signal_handler);
+	signal(SIGABRT, signal_handler);
+
+	clock_gettime(CLOCK_MONOTONIC, &time);
+	ns0 = time.tv_sec * 1000000000UL + time.tv_nsec;
+	for ( ; !quit; ) {
+		u64 ns1, ns_diff;
+
+		sleep(1);
+		clock_gettime(CLOCK_MONOTONIC, &time);
+		ns1 = time.tv_sec * 1000000000UL + time.tv_nsec;
+		ns_diff = ns1 - ns0;
+		ns0 = ns1;
+
+		print_port_stats_all(ns_diff);
+	}
+
+	/* Threads completion. */
+	printf("Quit.\n");
+	for (i = 0; i < n_threads; i++)
+		thread_data[i].quit = 1;
+
+	for (i = 0; i < n_threads; i++)
+		pthread_join(threads[i], NULL);
+
+	for (i = 0; i < n_ports; i++)
+		port_free(ports[i]);
+
+	bpool_free(bp);
+
+	remove_xdp_program();
+
+	return 0;
+}