Move updating TX packet counter to the inside of send loop. axe(4)

controllers combine multiple TX requests into single one if there
is room in TX buffer of controller. Updating TX packet counter at
the end of TX completion resulted in incorrect TX packet counter as
axe(4) thought it sent 1 packet. There is no easy way to know how
many combined TX were completed in the callback.
Because this change updates TX packet counter before actual
transmission, it may not be ideal one. But I believe it's better
than showing fake 8kpps under high TX load. With this change, TX
shows 221kpps on Linksus USB200M.

1 files changed, 11 insertions, 1 deletions

diff --git a/sys/dev/usb/net/if_axe.c b/sys/dev/usb/net/if_axe.c
index 8f32dca..5e62f9a 100644
--- a/sys/dev/usb/net/if_axe.c
+++ b/sys/dev/usb/net/if_axe.c
@@ -912,7 +912,6 @@ axe_bulk_write_callback(struct usb_xfer *xfer, usb_error_t error)
 	switch (USB_GET_STATE(xfer)) {
 	case USB_ST_TRANSFERRED:
 		DPRINTFN(11, "transfer complete\n");
-		ifp->if_opackets++;
 		/* FALLTHROUGH */
 	case USB_ST_SETUP:
 tr_setup:
@@ -958,6 +957,17 @@ tr_setup:
 			pos += m->m_pkthdr.len;
 
 			/*
+			 * XXX
+			 * Update TX packet counter here. This is not
+			 * correct way but it seems that there is no way
+			 * to know how many packets are sent at the end
+			 * of transfer because controller combines
+			 * multiple writes into single one if there is
+			 * room in TX buffer of controller.
+			 */
+			ifp->if_opackets++;
+
+			/*
 			 * if there's a BPF listener, bounce a copy
 			 * of this frame to him:
 			 */