xen/netfront: Add 2 bytes padding in the rx mbuf

The ethernet header size is not word aligned. Therefore the IP packet and so
on won't be align. On some architecture (such as ARM) unaligned access may
be slower and/or defined. Therefore we might reveice an alignement fault.
To void this case, we need to pull-up the data of ETHER_ALIGN bytes.

I'm not sure how this patch will impact x86, we need to do some benchmarking
without and with it.

I'mi also not sure m_copyup is the right function to any. Can any expert to the
network stack can tell me if there is a better solution?

diff --git a/sys/dev/xen/netfront/netfront.c b/sys/dev/xen/netfront/netfront.c
index 835c6ae09462975fc6791a8d92b6ad6c349e800a..9f6ee017da779b508311ef7010ce60d9f4068296 100644 (file)
--- a/sys/dev/xen/netfront/netfront.c
+++ b/sys/dev/xen/netfront/netfront.c
@@ -88,6 +88,8 @@ __FBSDID("$FreeBSD$");
 
 #include "xenbus_if.h"
 
+#define ETHER_ALIGN    2
+
 /* Features supported by all backends.  TSO and LRO can be negotiated */
 #define XN_CSUM_FEATURES       (CSUM_TCP | CSUM_UDP)
 
@@ -1357,6 +1359,21 @@ next_skip_queue:
                ref_cons = *cons + frags;
                frags++;
        }
+
+       if (!err) {
+               /* The ethernet header size is not word aligned. Therefore,
+                * the IP packet and so on won't be align. On some architecture
+                * (such as ARM) unaligned access may be slower and/or denied.
+                * Therefore we might receive an alignment fault.
+                * To avoid this case, we need to pull-up the data of
+                * ETHER_ALIGN bytes.
+                */
+               m0 = m_copyup(m0, min(m0->m_pkthdr.len, max_protohdr),
+                             ETHER_ALIGN);
+               if (!m0)
+                       err = ENOMEM;
+       }
+
        *list = m0;
        *cons += frags;
        *pages_flipped_p = pages_flipped;