[rds-commits] zab commits r137 - trunk/linux/net/rds

[svn-commits@oss.oracle.com]

Author: zab
Date: 2006-07-26 19:09:12 -0500 (Wed, 26 Jul 2006)
New Revision: 137

Modified:
   trunk/linux/net/rds/ib_recv.c
Log:
Correctly copy IB headers that straddle the posted fragment and header regions.


Modified: trunk/linux/net/rds/ib_recv.c
===================================================================
--- trunk/linux/net/rds/ib_recv.c	2006-07-26 23:01:15 UTC (rev 136)
+++ trunk/linux/net/rds/ib_recv.c	2006-07-27 00:09:12 UTC (rev 137)
@@ -264,6 +264,41 @@
 }
 
 /*
+ * Work is posted as a RDS_FRAG_SIZE payload and then a header.  This is
+ * done so that we can send fragments without headers and keep the fragments
+ * large and aligned.  The sender doesn't pad their fragments so the header
+ * will spill into the posted regions just after the fragment.
+ *
+ * XXX If we were to flip r_page into userspace or the page cache then we'd
+ * have to zero the header and possibly the rest of the page.
+ */
+static void rds_ib_copy_header(struct rds_ib_connection *ic,
+			       struct rds_ib_incoming *ibinc,
+			       struct rds_ib_recv_work *recv, u32 byte_len)
+{
+	void *addr;
+	u32 start;
+
+	BUG_ON(byte_len < sizeof(struct rds_header));
+
+	/* get the start of the header from the tail of the fragment */
+	start = byte_len - sizeof(struct rds_header);
+	if (start < RDS_FRAG_SIZE) {
+		addr = kmap_atomic(recv->r_page, KM_SOFTIRQ0);
+		memcpy(&ibinc->ii_inc.i_hdr, addr + start, 
+		       min(RDS_FRAG_SIZE - start, sizeof(struct rds_header)));
+		kunmap_atomic(addr, KM_SOFTIRQ0);
+	}
+
+	/* and the rest that might have spilled into the posted header space */
+	if (byte_len > RDS_FRAG_SIZE) {
+		memcpy(&ibinc->ii_inc.i_hdr,
+		       &ic->i_recv_hdrs[recv - ic->i_recvs],
+		       byte_len - RDS_FRAG_SIZE);
+	}
+}
+
+/*
  * Rings are posted with all the allocations they'll need to queue the
  * incoming message to the receiving socket so this can't fail.  It relies
  * on being called in the order that the sender sent in to infer which
@@ -287,25 +322,24 @@
 	 * off its list.
 	 */ 
 	if (ibinc == NULL) {
+		if (byte_len < sizeof(struct rds_header)) {
+			rds_ib_recv_unmap_page(ic, recv);
+			if (!ic->i_wc_err) {
+				ic->i_wc_err = 1;
+				printk(KERN_WARNING "RDS/IB: incoming message "
+				       "from %u.%u.%u.%u didn't inclue a "
+				       "header, disconnecting and "
+				       "reconnecting\n",
+				       NIPQUAD(conn->c_faddr));
+				queue_work(rds_wq, &conn->c_shutdown_work);
+			}
+		}
+
 		ibinc = recv->r_ibinc;
 		recv->r_ibinc = NULL;
 
-		if (byte_len <= RDS_FRAG_SIZE) {
-			/* 
-			 * XXX The remainter of the page will need to be zeroed
-			 * if we map it to userspace or flip it into the page
-			 * cache.
-			 */
-			void *addr = kmap_atomic(recv->r_page, KM_SOFTIRQ0);
-			memcpy(&ibinc->ii_inc.i_hdr,
-			       addr + byte_len - sizeof(struct rds_header),
-			       sizeof(struct rds_header));
-			kunmap_atomic(addr, KM_SOFTIRQ0);
-		} else {
-			int i = recv - ic->i_recvs;
-			memcpy(&ibinc->ii_inc.i_hdr, &ic->i_recv_hdrs[i],
-			       sizeof(struct rds_header));
-		}
+		rds_ib_copy_header(ic, ibinc, recv, byte_len);
+
 		ic->i_ibinc = ibinc;
 		ic->i_recv_data_rem = be32_to_cpu(ibinc->ii_inc.i_hdr.h_len);
 		rdsdebug("ic %p ibinc %p rem %u\n", ic, ibinc,