A small fix for if_em.c, if_igb.c, if_ixgbe.c

[Michael Tuexen]

On Dec 5, 2013, at 7:29 PM, Adrian Chadd <adrian at freebsd.org> wrote:

> Hi,
> 
> Yes. Looking at the ixgbe code, ixgbe_mq_start_locked() returns an
> error from ixgbe_xmit() but if it fails, it puts the buffer back. But
> it's already successfully queued a frame to the driver, so in this
> instance it shouldn't return the error from ixgbe_mq_start_locked().
> 
> The same deal in if_em.c and igb.c
> 
> Now, drbr_putback() used to fail and now it doesn't, as you've said.
> So we should change the xxx_mq_start_locked() to set err=0 if we go
> via the drbr_putback() routine, as it hasn't actually failed to
> transmit.
> 
> Now the very dirty thing is this - the error from xxx_transmit() is
> for the mbuf being queued at the end; but xxx_mq_start_locked()
> failures are for transmitting from the front. If there's only packet
> in the queue and that fails then they're the same thing and returning
> the error from xxx_mq_start_locked() matches the current mbuf being
> queued. But otherwise, they're referring to totally different packets.
> For TCP this may hurt; the TCP stack treats ENOBUFS a certain way and
> kicks off a timer to schedule a retransmit. I don't think we can fix
> _this_ right now.
> 
> So Michael - can you redo your patch to set err=0 if drbr_putback() is
> called, and retest?
Hi Adrian,

I guess you are talking about a patch like:

[bsd5:~/head/sys/dev] tuexen% svn diff -x -p
Index: e1000/if_em.c
===================================================================
--- e1000/if_em.c	(revision 259039)
+++ e1000/if_em.c	(working copy)
@@ -935,6 +935,7 @@ em_mq_start_locked(struct ifnet *ifp, struct tx_ri
 				drbr_advance(ifp, txr->br);
 			else 
 				drbr_putback(ifp, txr->br, next);
+				err = 0;
 			break;
 		}
 		drbr_advance(ifp, txr->br);
Index: e1000/if_igb.c
===================================================================
--- e1000/if_igb.c	(revision 259039)
+++ e1000/if_igb.c	(working copy)
@@ -1024,6 +1024,7 @@ igb_mq_start_locked(struct ifnet *ifp, struct tx_r
 				 * may have changed it.
 				 */
 				drbr_putback(ifp, txr->br, next);
+				err = 0;
 			}
 			break;
 		}
Index: ixgbe/ixgbe.c
===================================================================
--- ixgbe/ixgbe.c	(revision 259039)
+++ ixgbe/ixgbe.c	(working copy)
@@ -864,6 +864,7 @@ ixgbe_mq_start_locked(struct ifnet *ifp, struct tx
 				drbr_advance(ifp, txr->br);
 			} else {
 				drbr_putback(ifp, txr->br, next);
+				err = 0;
 			}
 #endif
 			break;
Index: ixgbe/ixv.c
===================================================================
--- ixgbe/ixv.c	(revision 259039)
+++ ixgbe/ixv.c	(working copy)
@@ -629,6 +629,7 @@ ixv_mq_start_locked(struct ifnet *ifp, struct tx_r
 				drbr_advance(ifp, txr->br);
 			} else {
 				drbr_putback(ifp, txr->br, next);
+				err = 0;
 			}
 			break;
 		}
Index: virtio/network/if_vtnet.c
===================================================================
--- virtio/network/if_vtnet.c	(revision 259039)
+++ virtio/network/if_vtnet.c	(working copy)
@@ -2242,9 +2242,10 @@ vtnet_txq_mq_start_locked(struct vtnet_txq *txq, s
 	while ((m = drbr_peek(ifp, br)) != NULL) {
 		error = vtnet_txq_encap(txq, &m);
 		if (error) {
-			if (m != NULL)
+			if (m != NULL) {
 				drbr_putback(ifp, br, m);
-			else
+				error = 0;
+			} else
 				drbr_advance(ifp, br);
 			break;
 		}

I looked for drivers using drbr_putback() and used a similar fix. Please note
that sys/dev/oce/oce_if.c seems strange. It uses drbr_putback() and drbr_enqueue(),
so I left it out for now.

I tested the igb driver and the above patch fixes the problem I saw.

From your above description I think the above patch is a valid patch.
However, xxx_transmit() can still return an error, even if there is no
problem with the provided packet. This is an issue for transport protocols...

Best regards
Michael



> 
> Thanks!
> 
> 
> 
> 
> -adrian
>