[PATCH v1 04/13] xprtrdma: Update rkeys after transport reconnect

[Chuck Lever <chuck.lever@oracle.com>]

Various reports of:

  rpcrdma_qp_async_error_upcall: QP error 3 on device mlx4_0
		ep ffff8800bfd3e848

Ensure that rkeys in already-marshalled RPC/RDMA headers are
refreshed after the QP has been replaced by a reconnect.

BugLink: https://bugzilla.linux-nfs.org/show_bug.cgi?id=249
Suggested-by: Selvin Xavier <Selvin.Xavier@Emulex.Com>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
---
 net/sunrpc/xprtrdma/rpc_rdma.c  |   77 ++++++++++++++++++++-------------------
 net/sunrpc/xprtrdma/transport.c |   11 +++---
 net/sunrpc/xprtrdma/xprt_rdma.h |   10 +++++
 3 files changed, 56 insertions(+), 42 deletions(-)

diff --git a/net/sunrpc/xprtrdma/rpc_rdma.c b/net/sunrpc/xprtrdma/rpc_rdma.c
index 693966d..6eeb6d2 100644
--- a/net/sunrpc/xprtrdma/rpc_rdma.c
+++ b/net/sunrpc/xprtrdma/rpc_rdma.c
@@ -53,14 +53,6 @@
 # define RPCDBG_FACILITY	RPCDBG_TRANS
 #endif
 
-enum rpcrdma_chunktype {
-	rpcrdma_noch = 0,
-	rpcrdma_readch,
-	rpcrdma_areadch,
-	rpcrdma_writech,
-	rpcrdma_replych
-};
-
 #ifdef RPC_DEBUG
 static const char transfertypes[][12] = {
 	"pure inline",	/* no chunks */
@@ -286,6 +278,30 @@ out:
 }
 
 /*
+ * Marshal chunks. This routine returns the header length
+ * consumed by marshaling.
+ *
+ * Returns positive RPC/RDMA header size, or negative errno.
+ */
+
+ssize_t
+rpcrdma_marshal_chunks(struct rpc_rqst *rqst, ssize_t result)
+{
+	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
+	struct rpcrdma_msg *headerp = (struct rpcrdma_msg *)req->rl_base;
+
+	if (req->rl_rtype != rpcrdma_noch) {
+		result = rpcrdma_create_chunks(rqst,
+				&rqst->rq_snd_buf, headerp, req->rl_rtype);
+
+	} else if (req->rl_wtype != rpcrdma_noch) {
+		result = rpcrdma_create_chunks(rqst,
+				&rqst->rq_rcv_buf, headerp, req->rl_wtype);
+	}
+	return result;
+}
+
+/*
  * Copy write data inline.
  * This function is used for "small" requests. Data which is passed
  * to RPC via iovecs (or page list) is copied directly into the
@@ -377,7 +393,6 @@ rpcrdma_marshal_req(struct rpc_rqst *rqst)
 	char *base;
 	size_t rpclen, padlen;
 	ssize_t hdrlen;
-	enum rpcrdma_chunktype rtype, wtype;
 	struct rpcrdma_msg *headerp;
 
 	/*
@@ -415,13 +430,13 @@ rpcrdma_marshal_req(struct rpc_rqst *rqst)
 	 * into pages; otherwise use reply chunks.
 	 */
 	if (rqst->rq_rcv_buf.buflen <= RPCRDMA_INLINE_READ_THRESHOLD(rqst))
-		wtype = rpcrdma_noch;
+		req->rl_wtype = rpcrdma_noch;
 	else if (rqst->rq_rcv_buf.page_len == 0)
-		wtype = rpcrdma_replych;
+		req->rl_wtype = rpcrdma_replych;
 	else if (rqst->rq_rcv_buf.flags & XDRBUF_READ)
-		wtype = rpcrdma_writech;
+		req->rl_wtype = rpcrdma_writech;
 	else
-		wtype = rpcrdma_replych;
+		req->rl_wtype = rpcrdma_replych;
 
 	/*
 	 * Chunks needed for arguments?
@@ -438,16 +453,16 @@ rpcrdma_marshal_req(struct rpc_rqst *rqst)
 	 * TBD check NFSv4 setacl
 	 */
 	if (rqst->rq_snd_buf.len <= RPCRDMA_INLINE_WRITE_THRESHOLD(rqst))
-		rtype = rpcrdma_noch;
+		req->rl_rtype = rpcrdma_noch;
 	else if (rqst->rq_snd_buf.page_len == 0)
-		rtype = rpcrdma_areadch;
+		req->rl_rtype = rpcrdma_areadch;
 	else
-		rtype = rpcrdma_readch;
+		req->rl_rtype = rpcrdma_readch;
 
 	/* The following simplification is not true forever */
-	if (rtype != rpcrdma_noch && wtype == rpcrdma_replych)
-		wtype = rpcrdma_noch;
-	if (rtype != rpcrdma_noch && wtype != rpcrdma_noch) {
+	if (req->rl_rtype != rpcrdma_noch && req->rl_wtype == rpcrdma_replych)
+		req->rl_wtype = rpcrdma_noch;
+	if (req->rl_rtype != rpcrdma_noch && req->rl_wtype != rpcrdma_noch) {
 		dprintk("RPC:       %s: cannot marshal multiple chunk lists\n",
 			__func__);
 		return -EIO;
@@ -461,7 +476,7 @@ rpcrdma_marshal_req(struct rpc_rqst *rqst)
 	 * When padding is in use and applies to the transfer, insert
 	 * it and change the message type.
 	 */
-	if (rtype == rpcrdma_noch) {
+	if (req->rl_rtype == rpcrdma_noch) {
 
 		padlen = rpcrdma_inline_pullup(rqst,
 						RPCRDMA_INLINE_PAD_VALUE(rqst));
@@ -476,7 +491,7 @@ rpcrdma_marshal_req(struct rpc_rqst *rqst)
 			headerp->rm_body.rm_padded.rm_pempty[1] = xdr_zero;
 			headerp->rm_body.rm_padded.rm_pempty[2] = xdr_zero;
 			hdrlen += 2 * sizeof(u32); /* extra words in padhdr */
-			if (wtype != rpcrdma_noch) {
+			if (req->rl_wtype != rpcrdma_noch) {
 				dprintk("RPC:       %s: invalid chunk list\n",
 					__func__);
 				return -EIO;
@@ -497,30 +512,18 @@ rpcrdma_marshal_req(struct rpc_rqst *rqst)
 			 * on receive. Therefore, we request a reply chunk
 			 * for non-writes wherever feasible and efficient.
 			 */
-			if (wtype == rpcrdma_noch)
-				wtype = rpcrdma_replych;
+			if (req->rl_wtype == rpcrdma_noch)
+				req->rl_wtype = rpcrdma_replych;
 		}
 	}
 
-	/*
-	 * Marshal chunks. This routine will return the header length
-	 * consumed by marshaling.
-	 */
-	if (rtype != rpcrdma_noch) {
-		hdrlen = rpcrdma_create_chunks(rqst,
-					&rqst->rq_snd_buf, headerp, rtype);
-		wtype = rtype;	/* simplify dprintk */
-
-	} else if (wtype != rpcrdma_noch) {
-		hdrlen = rpcrdma_create_chunks(rqst,
-					&rqst->rq_rcv_buf, headerp, wtype);
-	}
+	hdrlen = rpcrdma_marshal_chunks(rqst, hdrlen);
 	if (hdrlen < 0)
 		return hdrlen;
 
 	dprintk("RPC:       %s: %s: hdrlen %zd rpclen %zd padlen %zd"
 		" headerp 0x%p base 0x%p lkey 0x%x\n",
-		__func__, transfertypes[wtype], hdrlen, rpclen, padlen,
+		__func__, transfertypes[req->rl_wtype], hdrlen, rpclen, padlen,
 		headerp, base, req->rl_iov.lkey);
 
 	/*
diff --git a/net/sunrpc/xprtrdma/transport.c b/net/sunrpc/xprtrdma/transport.c
index 4185102..f6d280b 100644
--- a/net/sunrpc/xprtrdma/transport.c
+++ b/net/sunrpc/xprtrdma/transport.c
@@ -597,13 +597,14 @@ xprt_rdma_send_request(struct rpc_task *task)
 	struct rpc_xprt *xprt = rqst->rq_xprt;
 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
-	int rc;
+	int rc = 0;
 
-	if (req->rl_niovs == 0) {
+	if (req->rl_niovs == 0)
 		rc = rpcrdma_marshal_req(rqst);
-		if (rc < 0)
-			goto failed_marshal;
-	}
+	else if (r_xprt->rx_ia.ri_memreg_strategy == RPCRDMA_FRMR)
+		rc = rpcrdma_marshal_chunks(rqst, 0);
+	if (rc < 0)
+		goto failed_marshal;
 
 	if (req->rl_reply == NULL) 		/* e.g. reconnection */
 		rpcrdma_recv_buffer_get(req);
diff --git a/net/sunrpc/xprtrdma/xprt_rdma.h b/net/sunrpc/xprtrdma/xprt_rdma.h
index f3d86b2..c270e59 100644
--- a/net/sunrpc/xprtrdma/xprt_rdma.h
+++ b/net/sunrpc/xprtrdma/xprt_rdma.h
@@ -99,6 +99,14 @@ struct rpcrdma_ep {
 #define INIT_CQCOUNT(ep) atomic_set(&(ep)->rep_cqcount, (ep)->rep_cqinit)
 #define DECR_CQCOUNT(ep) atomic_sub_return(1, &(ep)->rep_cqcount)
 
+enum rpcrdma_chunktype {
+	rpcrdma_noch = 0,
+	rpcrdma_readch,
+	rpcrdma_areadch,
+	rpcrdma_writech,
+	rpcrdma_replych
+};
+
 /*
  * struct rpcrdma_rep -- this structure encapsulates state required to recv
  * and complete a reply, asychronously. It needs several pieces of
@@ -192,6 +200,7 @@ struct rpcrdma_req {
 	unsigned int	rl_niovs;	/* 0, 2 or 4 */
 	unsigned int	rl_nchunks;	/* non-zero if chunks */
 	unsigned int	rl_connect_cookie;	/* retry detection */
+	enum rpcrdma_chunktype	rl_rtype, rl_wtype;
 	struct rpcrdma_buffer *rl_buffer; /* home base for this structure */
 	struct rpcrdma_rep	*rl_reply;/* holder for reply buffer */
 	struct rpcrdma_mr_seg rl_segments[RPCRDMA_MAX_SEGS];/* chunk segments */
@@ -347,6 +356,7 @@ void rpcrdma_reply_handler(struct rpcrdma_rep *);
 /*
  * RPC/RDMA protocol calls - xprtrdma/rpc_rdma.c
  */
+ssize_t rpcrdma_marshal_chunks(struct rpc_rqst *, ssize_t);
 int rpcrdma_marshal_req(struct rpc_rqst *);
 size_t rpcrdma_max_payload(struct rpcrdma_xprt *);