1 files changed, 40 insertions, 21 deletions

diff --git a/net/sunrpc/xprtrdma/svc_rdma_rw.c b/net/sunrpc/xprtrdma/svc_rdma_rw.c
index 2121c9b4d275..04ec2a83af69 100644
--- a/net/sunrpc/xprtrdma/svc_rdma_rw.c
+++ b/net/sunrpc/xprtrdma/svc_rdma_rw.c
@@ -322,8 +322,6 @@ static int svc_rdma_post_chunk_ctxt(struct svc_rdma_chunk_ctxt *cc)
 		if (atomic_sub_return(cc->cc_sqecount,
 				      &rdma->sc_sq_avail) > 0) {
 			ret = ib_post_send(rdma->sc_qp, first_wr, &bad_wr);
-			trace_svcrdma_post_rw(&cc->cc_cqe,
-					      cc->cc_sqecount, ret);
 			if (ret)
 				break;
 			return 0;
@@ -336,6 +334,7 @@ static int svc_rdma_post_chunk_ctxt(struct svc_rdma_chunk_ctxt *cc)
 		trace_svcrdma_sq_retry(rdma);
 	} while (1);
 
+	trace_svcrdma_sq_post_err(rdma, ret);
 	set_bit(XPT_CLOSE, &xprt->xpt_flags);
 
 	/* If even one was posted, there will be a completion. */
@@ -481,18 +480,19 @@ static int svc_rdma_send_xdr_kvec(struct svc_rdma_write_info *info,
 				     vec->iov_len);
 }
 
-/* Send an xdr_buf's page list by itself. A Write chunk is
- * just the page list. a Reply chunk is the head, page list,
- * and tail. This function is shared between the two types
- * of chunk.
+/* Send an xdr_buf's page list by itself. A Write chunk is just
+ * the page list. A Reply chunk is @xdr's head, page list, and
+ * tail. This function is shared between the two types of chunk.
  */
 static int svc_rdma_send_xdr_pagelist(struct svc_rdma_write_info *info,
-				      struct xdr_buf *xdr)
+				      struct xdr_buf *xdr,
+				      unsigned int offset,
+				      unsigned long length)
 {
 	info->wi_xdr = xdr;
-	info->wi_next_off = 0;
+	info->wi_next_off = offset - xdr->head[0].iov_len;
 	return svc_rdma_build_writes(info, svc_rdma_pagelist_to_sg,
-				     xdr->page_len);
+				     length);
 }
 
 /**
@@ -500,6 +500,8 @@ static int svc_rdma_send_xdr_pagelist(struct svc_rdma_write_info *info,
  * @rdma: controlling RDMA transport
  * @wr_ch: Write chunk provided by client
  * @xdr: xdr_buf containing the data payload
+ * @offset: payload's byte offset in @xdr
+ * @length: size of payload, in bytes
  *
  * Returns a non-negative number of bytes the chunk consumed, or
  *	%-E2BIG if the payload was larger than the Write chunk,
@@ -509,19 +511,20 @@ static int svc_rdma_send_xdr_pagelist(struct svc_rdma_write_info *info,
  *	%-EIO if rdma_rw initialization failed (DMA mapping, etc).
  */
 int svc_rdma_send_write_chunk(struct svcxprt_rdma *rdma, __be32 *wr_ch,
-			      struct xdr_buf *xdr)
+			      struct xdr_buf *xdr,
+			      unsigned int offset, unsigned long length)
 {
 	struct svc_rdma_write_info *info;
 	int ret;
 
-	if (!xdr->page_len)
+	if (!length)
 		return 0;
 
 	info = svc_rdma_write_info_alloc(rdma, wr_ch);
 	if (!info)
 		return -ENOMEM;
 
-	ret = svc_rdma_send_xdr_pagelist(info, xdr);
+	ret = svc_rdma_send_xdr_pagelist(info, xdr, offset, length);
 	if (ret < 0)
 		goto out_err;
 
@@ -530,7 +533,7 @@ int svc_rdma_send_write_chunk(struct svcxprt_rdma *rdma, __be32 *wr_ch,
 		goto out_err;
 
 	trace_svcrdma_encode_write(xdr->page_len);
-	return xdr->page_len;
+	return length;
 
 out_err:
 	svc_rdma_write_info_free(info);
@@ -570,7 +573,9 @@ int svc_rdma_send_reply_chunk(struct svcxprt_rdma *rdma, __be32 *rp_ch,
 	 * client did not provide Write chunks.
 	 */
 	if (!writelist && xdr->page_len) {
-		ret = svc_rdma_send_xdr_pagelist(info, xdr);
+		ret = svc_rdma_send_xdr_pagelist(info, xdr,
+						 xdr->head[0].iov_len,
+						 xdr->page_len);
 		if (ret < 0)
 			goto out_err;
 		consumed += xdr->page_len;
@@ -671,7 +676,6 @@ static int svc_rdma_build_read_chunk(struct svc_rqst *rqstp,
 				     struct svc_rdma_read_info *info,
 				     __be32 *p)
 {
-	unsigned int i;
 	int ret;
 
 	ret = -EINVAL;
@@ -694,12 +698,6 @@ static int svc_rdma_build_read_chunk(struct svc_rqst *rqstp,
 		info->ri_chunklen += rs_length;
 	}
 
-	/* Pages under I/O have been copied to head->rc_pages.
-	 * Prevent their premature release by svc_xprt_release() .
-	 */
-	for (i = 0; i < info->ri_readctxt->rc_page_count; i++)
-		rqstp->rq_pages[i] = NULL;
-
 	return ret;
 }
 
@@ -794,6 +792,26 @@ out:
 	return ret;
 }
 
+/* Pages under I/O have been copied to head->rc_pages. Ensure they
+ * are not released by svc_xprt_release() until the I/O is complete.
+ *
+ * This has to be done after all Read WRs are constructed to properly
+ * handle a page that is part of I/O on behalf of two different RDMA
+ * segments.
+ *
+ * Do this only if I/O has been posted. Otherwise, we do indeed want
+ * svc_xprt_release() to clean things up properly.
+ */
+static void svc_rdma_save_io_pages(struct svc_rqst *rqstp,
+				   const unsigned int start,
+				   const unsigned int num_pages)
+{
+	unsigned int i;
+
+	for (i = start; i < num_pages + start; i++)
+		rqstp->rq_pages[i] = NULL;
+}
+
 /**
  * svc_rdma_recv_read_chunk - Pull a Read chunk from the client
  * @rdma: controlling RDMA transport
@@ -847,6 +865,7 @@ int svc_rdma_recv_read_chunk(struct svcxprt_rdma *rdma, struct svc_rqst *rqstp,
 	ret = svc_rdma_post_chunk_ctxt(&info->ri_cc);
 	if (ret < 0)
 		goto out_err;
+	svc_rdma_save_io_pages(rqstp, 0, head->rc_page_count);
 	return 0;
 
 out_err: