>Number:         4004
>Category:       kern
>Synopsis:       Smarter polling of scsi ethernet devices
>Confidential:   no
>Severity:       non-critical
>Priority:       low
>Responsible:    kern-bug-people (Kernel Bug People)
>State:          open
>Class:          change-request
>Submitter-Id:   net
>Arrival-Date:   Mon Aug 18 02:50:02 1997
>Last-Modified:
>Originator:     Ian Dall
>Organization:
	
>Release:        1.2
>Environment:
	
System: NetBSD sibyl.chez-dall.org.au 1.2G NetBSD 1.2G (SIBYL) #0: Thu Jul 3 20:33:43 CST 1997 ian@sibyl.chez-dall.org.au:/usr3/netbsd-1.2/lsrc/sys/arch/pc532/compile/SIBYL pc532


>Description:
	
The code for the scsi ethernet adaptor (sys/scsi/if_se.c) could be smarter
in polling the device.
>How-To-Repeat:
	
Flood a machine running the scsi/ethernet code with packets from a fast
machine. An easy way to do this is with 'spray -l 1500 ...' Note the number
of packets dropped.

>Fix:
	
The following patch employs a smarter adaptive polling scheme. It also
improves the comments, in particular giving due credit to Phil Budne for
his efforts in nunderstanding the device. It also relaxes the pattern
for auto config so more devices match.

--- ../../../src/sys/scsi/if_se.c	Tue Apr 29 20:55:43 1997
+++ if_se.c	Mon Jun 30 21:08:51 1997
@@ -31,11 +31,15 @@
  */
 
 /*
- * Written by Ian Dall <ian.dall@dsto.defence.gov.au> Feb 3, 1997
- */
-/*
  * Driver for Cabletron EA41x scsi ethernet adaptor.
  *
+ * Written by Ian Dall <ian.dall@dsto.defence.gov.au> Feb 3, 1997
+ *
+ * Acknowledgement: Thanks are due to Philip L. Budne <budd@cs.bu.edu>
+ * who reverse engineered the the EA41x. In developing this code,
+ * Phil's userland daemon "etherd", was refered to extensively in lieu
+ * of accurate documentation for the device.
+ *
  * This is a weird device! It doesn't conform to the scsi spec in much
  * at all. About the only standard command supported in inquiry. Most
  * commands are 6 bytes long, but the recv data is only 1 byte.  Data
@@ -123,19 +127,31 @@
 /* Make this big enough for an ETHERMTU packet in promiscuous mode. */
 #define MAX_SNAP	(ETHERMTU + sizeof(struct ether_header) + \
 			 SE_PREFIX + ETHER_CRC)
-#define RBUF_LEN     (16 * 1024)
 
-/*
- * se_poll and se_poll0 are the normal polling rate and the
- * minimum polling rate respectively. If se_poll0 should be
- * chosen so that at maximum ethernet speed, we will read nearly
- * full buffers. se_poll should be chosen for reasonable maximum
- * latency.
+/* 10 full length packets appears to be the max ever returned. 16k is OK */
+#define RBUF_LEN	(16 * 1024)
+
+/* Tuning parameters: 
+ * The EA41x only returns a maximum of 10 packets (regardless of size).
+ * We will attempt to adapt to polling fast enough to get RDATA_GOAL packets
+ * per read
+ */
+#define RDATA_MAX 10
+#define RDATA_GOAL 8
+
+/* se_poll and se_poll0 are the normal polling rate and the minimum
+ * polling rate respectively. se_poll0 should be chosen so that at
+ * maximum ethernet speed, we will read nearly RDATA_MAX packets. se_poll
+ * should be chosen for reasonable maximum latency.
+ * In practice, if we are being saturated with min length packets, we
+ * can't poll fast enough. Polling with zero delay actually
+ * worsens performance. se_poll0 is enforced to be always at least 1
  */
 #define SE_POLL 40		/* default in milliseconds */
-#define SE_POLL0 20		/* Default in milliseconds */
+#define SE_POLL0 10		/* default in milliseconds */
 int se_poll = 0; /* Delay in ticks set at attach time */
 int se_poll0 = 0;
+int se_max_received = 0;	/* Instrumentation */
 
 #define	PROTOCMD(p, d) \
 	((d) = (p))
@@ -174,6 +190,7 @@
 	int sc_debug;
 	int sc_flags;
 #define SE_NEED_RECV 0x1
+	int sc_last_timeout;
 };
 
 cdev_decl(se);
@@ -232,9 +249,9 @@
 
 struct scsi_inquiry_pattern se_patterns[] = {
 	{T_PROCESSOR, T_FIXED,
-	 "CABLETRN",	     "EA412/14/19",		    ""},
+	 "CABLETRN",         "EA412",                 ""},
 	{T_PROCESSOR, T_FIXED,
-	 "Cabletrn",	     "EA412/",			    ""},
+	 "Cabletrn",         "EA412",                 ""},
 };
 
 /*
@@ -497,32 +514,36 @@
 			/* Reschedule after a delay */
 			timeout(se_recv, (void *)sc, se_poll);
 		} else {
-			int n;
+			int n, ntimeo;
 			n = se_read(sc, xs->data, xs->datalen - xs->resid);
-
-			if(n > 0) {
-#ifdef SEDEBUG
-				if (sc->sc_debug)
-					printf("sedone: received %d packets \n", n);
-#endif
-				if(ifp->if_snd.ifq_head)
-					/* Output is
-					 * pending. Do next
-					 * recv after the next
-					 * send. */
-					sc->sc_flags |= SE_NEED_RECV;
-				else {
-					timeout(se_recv, (void *)sc, se_poll0);
-				}
-			} else {
-				/* Reschedule after a delay */
-				timeout(se_recv, (void *)sc, se_poll);
+			if (n > se_max_received)
+				se_max_received = n;
+			if (n == 0)
+				ntimeo = se_poll;
+			else if (n >= RDATA_MAX)
+				ntimeo = se_poll0;
+			else {
+				ntimeo = sc->sc_last_timeout;
+				ntimeo = (ntimeo * RDATA_GOAL)/n;
+				ntimeo = (ntimeo < se_poll0?
+					  se_poll0: ntimeo);
+				ntimeo = (ntimeo > se_poll?
+					  se_poll: ntimeo);
+			}
+			sc->sc_last_timeout = ntimeo;
+			if (ntimeo == se_poll0  &&
+			    ifp->if_snd.ifq_head)
+				/* Output is pending. Do next recv
+				 * after the next send.  */
+				sc->sc_flags |= SE_NEED_RECV;
+			else {
+				timeout(se_recv, (void *)sc, ntimeo);
 			}
 		}
 	}
 	splx(s);
 }
-  
+
 static void
 se_recv(v)
 	void *v;


	
>Audit-Trail:
>Unformatted: