[netperf-dev] netperf2 commit notice r349 - in trunk: . src

[raj at netperf.org]

Author: raj
Date: 2010-05-21 10:53:14 -0700 (Fri, 21 May 2010)
New Revision: 349

Modified:
   trunk/AUTHORS
   trunk/Release_Notes
   trunk/src/nettest_bsd.c
Log:
workaround to get ENOBUFS on tx overflow on linux provided by Andrew Gallatin

Modified: trunk/AUTHORS
===================================================================
--- trunk/AUTHORS	2010-04-28 00:19:42 UTC (rev 348)
+++ trunk/AUTHORS	2010-05-21 17:53:14 UTC (rev 349)
@@ -105,6 +105,7 @@
 calibration-free
 fixes to configure to recognize Solaris 11
 fixes to netcpu_procstat.c for later linux kernels
+workarounds to get Linux to report ENOBUFS on TX queue overflows
 
 Mark Cooper
 pointing-out the need for -lresolv when compiling -DDO_DNS on RedHat

Modified: trunk/Release_Notes
===================================================================
--- trunk/Release_Notes	2010-04-28 00:19:42 UTC (rev 348)
+++ trunk/Release_Notes	2010-05-21 17:53:14 UTC (rev 349)
@@ -1,5 +1,9 @@
 These are the Release Notes for post-revision 2.4.5 of netperf:
 
+*) Add a workaround to get Linux to report TX queue drops in a
+   UDP_STREAM test when the socket buffer size is larger than the
+   TX queue.  Provided by Andrew Gallatin.
+
 *) Fix the configure script to know it does not have to look for an 
    SCTP library on FreeBSD 8.X
 

Modified: trunk/src/nettest_bsd.c
===================================================================
--- trunk/src/nettest_bsd.c	2010-04-28 00:19:42 UTC (rev 348)
+++ trunk/src/nettest_bsd.c	2010-05-21 17:53:14 UTC (rev 349)
@@ -6323,6 +6323,35 @@
   
 }
 
+#if defined(__linux)
+/*
+ * Linux has this odd behavior where if the socket buffers are larger than
+ * a device's txqueuelen, the kernel will silently drop transmits which would
+ * not fit into the tx queue, and not  pass an ENOBUFS error back to the 
+ * application.  As a result, a UDP stream test can report absurd transmit
+ * bandwidths (like 20Gb/s on a 1GbE NIC).  This behavior can be avoided if 
+ * you  request extended error reporting on the socket.  This is done by 
+ * setting the IP_RECVERR socket option at the IP level.
+ */
+static void
+enable_enobufs(int s)
+{
+  struct protoent *pr;
+  int on = 1;
+
+  if ((pr = getprotobyname("ip")) == NULL) {
+    fprintf(where, "enable_enobufs failed: getprotobyname\n");
+    fflush(where);
+    return;
+  }
+  if (setsockopt(s, pr->p_proto, IP_RECVERR, (char *)&on, sizeof(on)) < 0) {
+    fprintf(where, "enable_enobufs failed: setsockopt\n");
+    fflush(where);
+    return;
+  }
+}
+#endif
+
 void
 send_udp_stream(char remote_host[])
 {
@@ -6554,6 +6583,10 @@
           fflush(where);
        }
     }
+
+#if defined (__linux)
+  enable_enobufs(data_socket);
+#endif
     
 #ifdef WIN32
   /* this is used so the timer thread can close the socket out from */