[openib-general] [PATCH 9 of 10] ipath - fix null deref during rdma ops

[Bryan O'Sullivan]

The problem was that node A's sending thread, which handles sending RDMA
read response data, would write the trigger word, the last packet would
be sent, node B would send a new RDMA read request, node A's interrupt
handler would initialize s_rdma_sge, then node A's sending thread would
update s_rdma_sge.  This didn't happen very often naturally but was more
frequent with 1 byte RDMA reads.  Rather than adding more locking or
increasing the QP structure size and copying sge data, I modified the
copy routine to update the pointers before writing the trigger word to
avoid the update race.

Signed-off-by: Ralph Campbell <ralphc at pathscale.com>
Signed-off-by: Bryan O'Sullivan <bos at pathscale.com>

diff -r 551717ecc3db -r 3d844dee2f61 drivers/infiniband/hw/ipath/ipath_layer.c
--- a/drivers/infiniband/hw/ipath/ipath_layer.c	Tue May 23 11:29:16 2006 -0700
+++ b/drivers/infiniband/hw/ipath/ipath_layer.c	Tue May 23 11:29:16 2006 -0700
@@ -872,12 +872,13 @@ static void copy_io(u32 __iomem *piobuf,
 		update_sge(ss, len);
 		length -= len;
 	}
+	/* Update address before sending packet. */
+	update_sge(ss, length);
 	/* must flush early everything before trigger word */
 	ipath_flush_wc();
 	__raw_writel(last, piobuf);
 	/* be sure trigger word is written */
 	ipath_flush_wc();
-	update_sge(ss, length);
 }
 
 /**
@@ -943,17 +944,18 @@ int ipath_verbs_send(struct ipath_devdat
 	if (likely(ss->num_sge == 1 && len <= ss->sge.length &&
 		   !((unsigned long)ss->sge.vaddr & (sizeof(u32) - 1)))) {
 		u32 w;
-
+		u32 *addr = (u32 *) ss->sge.vaddr;
+
+		/* Update address before sending packet. */
+		update_sge(ss, len);
 		/* Need to round up for the last dword in the packet. */
 		w = (len + 3) >> 2;
-		__iowrite32_copy(piobuf, ss->sge.vaddr, w - 1);
+		__iowrite32_copy(piobuf, addr, w - 1);
 		/* must flush early everything before trigger word */
 		ipath_flush_wc();
-		__raw_writel(((u32 *) ss->sge.vaddr)[w - 1],
-			     piobuf + w - 1);
+		__raw_writel(addr[w - 1], piobuf + w - 1);
 		/* be sure trigger word is written */
 		ipath_flush_wc();
-		update_sge(ss, len);
 		ret = 0;
 		goto bail;
 	}