devices: af_packet, fix tx stall by retrying failed sendto
Change-Id: I6bed66f740b34673a4883eda1c7f7310c57e131b
Type: fix
Signed-off-by: Mohammed Hawari <mohammed@hawari.fr>
Signed-off-by: Benoît Ganne <bganne@cisco.com>
diff --git a/src/vnet/devices/af_packet/af_packet.c b/src/vnet/devices/af_packet/af_packet.c
index 83c1068..ec65bf6 100644
--- a/src/vnet/devices/af_packet/af_packet.c
+++ b/src/vnet/devices/af_packet/af_packet.c
@@ -437,6 +437,7 @@
tx_queue->next_tx_frame = 0;
tx_queue->queue_id = queue_id;
+ tx_queue->is_tx_pending = 0;
clib_spinlock_init (&tx_queue->lockp);
}
diff --git a/src/vnet/devices/af_packet/af_packet.h b/src/vnet/devices/af_packet/af_packet.h
index 10a7aaf..e5dc615 100644
--- a/src/vnet/devices/af_packet/af_packet.h
+++ b/src/vnet/devices/af_packet/af_packet.h
@@ -82,6 +82,7 @@
u32 rx_frame_offset;
u16 num_rx_pkts;
u8 is_rx_pending;
+ u8 is_tx_pending;
vnet_hw_if_rx_mode mode;
} af_packet_queue_t;
diff --git a/src/vnet/devices/af_packet/device.c b/src/vnet/devices/af_packet/device.c
index 1e177f6..e1eb46a 100644
--- a/src/vnet/devices/af_packet/device.c
+++ b/src/vnet/devices/af_packet/device.c
@@ -418,9 +418,10 @@
CLIB_MEMORY_BARRIER ();
- if (PREDICT_TRUE (n_sent))
+ if (PREDICT_TRUE (n_sent || tx_queue->is_tx_pending))
{
tx_queue->next_tx_frame = tx_frame;
+ tx_queue->is_tx_pending = 0;
if (PREDICT_FALSE (
sendto (tx_queue->fd, NULL, 0, MSG_DONTWAIT, NULL, 0) == -1))
@@ -429,11 +430,26 @@
* Note that we have no reliable way to properly determine the
* disposition of the packets we just enqueued for delivery.
*/
- vlib_error_count (vm, node->node_index,
- unix_error_is_fatal (errno) ?
- AF_PACKET_TX_ERROR_TXRING_FATAL :
- AF_PACKET_TX_ERROR_TXRING_EAGAIN,
- n_sent);
+ uword counter;
+
+ if (unix_error_is_fatal (errno))
+ {
+ counter = AF_PACKET_TX_ERROR_TXRING_FATAL;
+ }
+ else
+ {
+ counter = AF_PACKET_TX_ERROR_TXRING_EAGAIN;
+ /* non-fatal error: kick again next time
+ * note that you could still end up in a deadlock: if you do not
+ * try to send new packets (ie reschedule this tx node), eg.
+ * because your peer is waiting for the unsent packets to reply
+ * to you but your waiting for its reply etc., you are not going
+ * to kick again, and everybody is waiting for the other to talk
+ * 1st... */
+ tx_queue->is_tx_pending = 1;
+ }
+
+ vlib_error_count (vm, node->node_index, counter, 1);
}
}