selftests/xsk: adjust packet pacing for multi-buffer support

Modify the packet pacing algorithm so that it works with multi-buffer packets. This algorithm makes sure we do not send too many buffers to the receiving thread so that packets have to be dropped. The previous algorithm made the assumption that each packet only consumes one buffer, but that is not true anymore when multi-buffer support gets added. Instead, we find out what the largest packet size is in the packet stream and assume that each packet will consume this many buffers. This is conservative and overly cautious as there might be smaller packets in the stream that need fewer buffers per packet. But it keeps the algorithm simple. Also simplify it by removing the pthread conditional and just test if there is enough space in the Rx thread before trying to send one more batch. Also makes the tests run faster. Signed-off-by: Magnus Karlsson <[email protected]> Link: https://lore.kernel.org/r/[email protected] Signed-off-by: Alexei Starovoitov <[email protected]>
niw · May 17, 2023 · 7cd6df4 · 7cd6df4
1 parent 2f6eae0
commit 7cd6df4
Show file tree

Hide file tree

Showing 2 changed files with 30 additions and 20 deletions.
diff --git a/tools/testing/selftests/bpf/xskxceiver.c b/tools/testing/selftests/bpf/xskxceiver.c
@@ -555,6 +555,11 @@ static void pkt_set(struct xsk_umem_info *umem, struct pkt *pkt, int offset, u32
 		pkt->valid = true;
 }
 
+static u32 pkt_get_buffer_len(struct xsk_umem_info *umem, u32 len)
+{
+	return ceil_u32(len, umem->frame_size) * umem->frame_size;
+}
+
 static struct pkt_stream *pkt_stream_generate(struct xsk_umem_info *umem, u32 nb_pkts, u32 pkt_len)
 {
 	struct pkt_stream *pkt_stream;
@@ -564,6 +569,8 @@ static struct pkt_stream *pkt_stream_generate(struct xsk_umem_info *umem, u32 nb
 	if (!pkt_stream)
 		exit_with_error(ENOMEM);
 
+	pkt_stream->nb_pkts = nb_pkts;
+	pkt_stream->max_pkt_len = pkt_len;
 	for (i = 0; i < nb_pkts; i++) {
 		struct pkt *pkt = &pkt_stream->pkts[i];
 
@@ -661,10 +668,14 @@ static void __pkt_stream_generate_custom(struct ifobject *ifobj,
 		exit_with_error(ENOMEM);
 
 	for (i = 0; i < nb_pkts; i++) {
-		pkt_stream->pkts[i].offset = pkts[i].offset;
-		pkt_stream->pkts[i].len = pkts[i].len;
-		pkt_stream->pkts[i].pkt_nb = i;
-		pkt_stream->pkts[i].valid = pkts[i].valid;
+		struct pkt *pkt = &pkt_stream->pkts[i];
+
+		pkt->offset = pkts[i].offset;
+		pkt->len = pkts[i].len;
+		pkt->pkt_nb = i;
+		pkt->valid = pkts[i].valid;
+		if (pkt->len > pkt_stream->max_pkt_len)
+			pkt_stream->max_pkt_len = pkt->len;
 	}
 
 	ifobj->pkt_stream = pkt_stream;
@@ -926,8 +937,6 @@ static int receive_pkts(struct test_spec *test, struct pollfd *fds)
 
 		pthread_mutex_lock(&pacing_mutex);
 		pkts_in_flight -= pkts_sent;
-		if (pkts_in_flight < umem->num_frames)
-			pthread_cond_signal(&pacing_cond);
 		pthread_mutex_unlock(&pacing_mutex);
 		pkts_sent = 0;
 	}
@@ -938,10 +947,18 @@ static int receive_pkts(struct test_spec *test, struct pollfd *fds)
 static int __send_pkts(struct ifobject *ifobject, struct pollfd *fds, bool timeout)
 {
 	struct xsk_socket_info *xsk = ifobject->xsk;
+	struct xsk_umem_info *umem = ifobject->umem;
+	u32 i, idx = 0, valid_pkts = 0, buffer_len;
 	bool use_poll = ifobject->use_poll;
-	u32 i, idx = 0, valid_pkts = 0;
 	int ret;
 
+	buffer_len = pkt_get_buffer_len(umem, ifobject->pkt_stream->max_pkt_len);
+	/* pkts_in_flight might be negative if many invalid packets are sent */
+	if (pkts_in_flight >= (int)((umem_size(umem) - BATCH_SIZE * buffer_len) / buffer_len)) {
+		kick_tx(xsk);
+		return TEST_CONTINUE;
+	}
+
 	while (xsk_ring_prod__reserve(&xsk->tx, BATCH_SIZE, &idx) < BATCH_SIZE) {
 		if (use_poll) {
 			ret = poll(fds, 1, POLL_TMOUT);
@@ -972,7 +989,7 @@ static int __send_pkts(struct ifobject *ifobject, struct pollfd *fds, bool timeo
 		if (!pkt)
 			break;
 
-		tx_desc->addr = pkt_get_addr(pkt, ifobject->umem);
+		tx_desc->addr = pkt_get_addr(pkt, umem);
 		tx_desc->len = pkt->len;
 		if (pkt->valid) {
 			valid_pkts++;
@@ -982,11 +999,6 @@ static int __send_pkts(struct ifobject *ifobject, struct pollfd *fds, bool timeo
 
 	pthread_mutex_lock(&pacing_mutex);
 	pkts_in_flight += valid_pkts;
-	/* pkts_in_flight might be negative if many invalid packets are sent */
-	if (pkts_in_flight >= (int)(ifobject->umem->num_frames - BATCH_SIZE)) {
-		kick_tx(xsk);
-		pthread_cond_wait(&pacing_cond, &pacing_mutex);
-	}
 	pthread_mutex_unlock(&pacing_mutex);
 
 	xsk_ring_prod__submit(&xsk->tx, i);
@@ -1032,9 +1044,11 @@ static int send_pkts(struct test_spec *test, struct ifobject *ifobject)
 
 	while (pkt_stream->current_pkt_nb < pkt_stream->nb_pkts) {
 		ret = __send_pkts(ifobject, &fds, timeout);
+		if (ret == TEST_CONTINUE && !test->fail)
+			continue;
 		if ((ret || test->fail) && !timeout)
 			return TEST_FAILURE;
-		else if (ret == TEST_PASS && timeout)
+		if (ret == TEST_PASS && timeout)
 			return ret;
 	}
 
@@ -1319,12 +1333,8 @@ static void *worker_testapp_validate_rx(void *arg)
 
 	if (!err && ifobject->validation_func)
 		err = ifobject->validation_func(ifobject);
-	if (err) {
+	if (err)
 		report_failure(test);
-		pthread_mutex_lock(&pacing_mutex);
-		pthread_cond_signal(&pacing_cond);
-		pthread_mutex_unlock(&pacing_mutex);
-	}
 
 	pthread_exit(NULL);
 }

diff --git a/tools/testing/selftests/bpf/xskxceiver.h b/tools/testing/selftests/bpf/xskxceiver.h
@@ -121,6 +121,7 @@ struct pkt_stream {
 	u32 nb_pkts;
 	u32 current_pkt_nb;
 	struct pkt *pkts;
+	u32 max_pkt_len;
 };
 
 struct ifobject;
@@ -173,7 +174,6 @@ struct test_spec {
 
 pthread_barrier_t barr;
 pthread_mutex_t pacing_mutex = PTHREAD_MUTEX_INITIALIZER;
-pthread_cond_t pacing_cond = PTHREAD_COND_INITIALIZER;
 
 int pkts_in_flight;