The BCE chips appear to have an undocumented requirement that RX frames be

aligned on an 8 byte boundary.  Prior to rev 1.36 this wasn't a problem
because mbuf clusters tend be naturally aligned.  The switch to using
split buffers with the first buffer being the embedded data area of the
mbuf has broken this assumption, at least on i386, causing a complete
failure of RX functionality.  Fix this for now by using a full cluster for
the first RX buffer.  A more sophisticated approach could be done with the
old buffer scheme to realign the m_data pointer with m_adj(), but I'm also
not clear on performance benefits of this old scheme or the performance
implications of adding an m_adj() call to every allocation.

1 files changed, 8 insertions, 0 deletions

diff --git a/sys/dev/bce/if_bce.c b/sys/dev/bce/if_bce.c
index 859d74f..14bfcb9 100644
--- a/sys/dev/bce/if_bce.c
+++ b/sys/dev/bce/if_bce.c
@@ -776,7 +776,11 @@ bce_attach(device_t dev)
 	ifp->if_capenable    = ifp->if_capabilities;
 
 	/* Use standard mbuf sizes for buffer allocation. */
+#ifdef BCE_USE_SPLIT_HEADER
 	sc->rx_bd_mbuf_alloc_size = MHLEN;
+#else
+	sc->rx_bd_mbuf_alloc_size = MCLBYTES;;
+#endif
 	sc->pg_bd_mbuf_alloc_size = MCLBYTES;
 
 	ifp->if_snd.ifq_drv_maxlen = USABLE_TX_BD;
@@ -3804,7 +3808,11 @@ bce_get_rx_buf(struct bce_softc *sc, struct mbuf *m, u16 *prod,
 			goto bce_get_rx_buf_exit);
 
 		/* This is a new mbuf allocation. */
+#ifdef BCE_USE_SPLIT_HEADER
 		MGETHDR(m_new, M_DONTWAIT, MT_DATA);
+#else
+		m_new = m_getcl(M_DONTWAIT, MT_DATA, M_PKTHDR);
+#endif
 		if (m_new == NULL) {
 			sc->mbuf_alloc_failed++;
 			rc = ENOBUFS;