Merge tag 'nfsd-4.12' of git://linux-nfs.org/~bfields/linux

Pull nfsd updates from Bruce Fields:
 "Another RDMA update from Chuck Lever, and a bunch of miscellaneous
  bugfixes"

* tag 'nfsd-4.12' of git://linux-nfs.org/~bfields/linux: (26 commits)
  nfsd: Fix up the "supattr_exclcreat" attributes
  nfsd: encoders mustn't use unitialized values in error cases
  nfsd: fix undefined behavior in nfsd4_layout_verify
  lockd: fix lockd shutdown race
  NFSv4: Fix callback server shutdown
  SUNRPC: Refactor svc_set_num_threads()
  NFSv4.x/callback: Create the callback service through svc_create_pooled
  lockd: remove redundant check on block
  svcrdma: Clean out old XDR encoders
  svcrdma: Remove the req_map cache
  svcrdma: Remove unused RDMA Write completion handler
  svcrdma: Reduce size of sge array in struct svc_rdma_op_ctxt
  svcrdma: Clean up RPC-over-RDMA backchannel reply processing
  svcrdma: Report Write/Reply chunk overruns
  svcrdma: Clean up RDMA_ERROR path
  svcrdma: Use rdma_rw API in RPC reply path
  svcrdma: Introduce local rdma_rw API helpers
  svcrdma: Clean up svc_rdma_get_inv_rkey()
  svcrdma: Add helper to save pages under I/O
  svcrdma: Eliminate RPCRDMA_SQ_DEPTH_MULT
  ...

8 files changed, 1100 insertions, 749 deletions

diff --git a/net/sunrpc/xprtrdma/Makefile b/net/sunrpc/xprtrdma/Makefile
index ef19fa42c50f..c1ae8142ab73 100644
--- a/net/sunrpc/xprtrdma/Makefile
+++ b/net/sunrpc/xprtrdma/Makefile
@@ -4,5 +4,5 @@ rpcrdma-y := transport.o rpc_rdma.o verbs.o \
 	fmr_ops.o frwr_ops.o \
 	svc_rdma.o svc_rdma_backchannel.o svc_rdma_transport.o \
 	svc_rdma_marshal.o svc_rdma_sendto.o svc_rdma_recvfrom.o \
-	module.o
+	svc_rdma_rw.o module.o
 rpcrdma-$(CONFIG_SUNRPC_BACKCHANNEL) += backchannel.o
diff --git a/net/sunrpc/xprtrdma/svc_rdma.c b/net/sunrpc/xprtrdma/svc_rdma.c
index c846ca9f1eba..a4a8f6989ee7 100644
--- a/net/sunrpc/xprtrdma/svc_rdma.c
+++ b/net/sunrpc/xprtrdma/svc_rdma.c
@@ -58,9 +58,9 @@ unsigned int svcrdma_max_requests = RPCRDMA_MAX_REQUESTS;
 unsigned int svcrdma_max_bc_requests = RPCRDMA_MAX_BC_REQUESTS;
 static unsigned int min_max_requests = 4;
 static unsigned int max_max_requests = 16384;
-unsigned int svcrdma_max_req_size = RPCRDMA_MAX_REQ_SIZE;
-static unsigned int min_max_inline = 4096;
-static unsigned int max_max_inline = 65536;
+unsigned int svcrdma_max_req_size = RPCRDMA_DEF_INLINE_THRESH;
+static unsigned int min_max_inline = RPCRDMA_DEF_INLINE_THRESH;
+static unsigned int max_max_inline = RPCRDMA_MAX_INLINE_THRESH;
 
 atomic_t rdma_stat_recv;
 atomic_t rdma_stat_read;
@@ -247,8 +247,6 @@ int svc_rdma_init(void)
 	dprintk("SVCRDMA Module Init, register RPC RDMA transport\n");
 	dprintk("\tsvcrdma_ord      : %d\n", svcrdma_ord);
 	dprintk("\tmax_requests     : %u\n", svcrdma_max_requests);
-	dprintk("\tsq_depth         : %u\n",
-		svcrdma_max_requests * RPCRDMA_SQ_DEPTH_MULT);
 	dprintk("\tmax_bc_requests  : %u\n", svcrdma_max_bc_requests);
 	dprintk("\tmax_inline       : %d\n", svcrdma_max_req_size);
 
diff --git a/net/sunrpc/xprtrdma/svc_rdma_backchannel.c b/net/sunrpc/xprtrdma/svc_rdma_backchannel.c
index ff1df40f0d26..c676ed0efb5a 100644
--- a/net/sunrpc/xprtrdma/svc_rdma_backchannel.c
+++ b/net/sunrpc/xprtrdma/svc_rdma_backchannel.c
@@ -12,7 +12,17 @@
 
 #undef SVCRDMA_BACKCHANNEL_DEBUG
 
-int svc_rdma_handle_bc_reply(struct rpc_xprt *xprt, struct rpcrdma_msg *rmsgp,
+/**
+ * svc_rdma_handle_bc_reply - Process incoming backchannel reply
+ * @xprt: controlling backchannel transport
+ * @rdma_resp: pointer to incoming transport header
+ * @rcvbuf: XDR buffer into which to decode the reply
+ *
+ * Returns:
+ *	%0 if @rcvbuf is filled in, xprt_complete_rqst called,
+ *	%-EAGAIN if server should call ->recvfrom again.
+ */
+int svc_rdma_handle_bc_reply(struct rpc_xprt *xprt, __be32 *rdma_resp,
 			     struct xdr_buf *rcvbuf)
 {
 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
@@ -27,13 +37,13 @@ int svc_rdma_handle_bc_reply(struct rpc_xprt *xprt, struct rpcrdma_msg *rmsgp,
 
 	p = (__be32 *)src->iov_base;
 	len = src->iov_len;
-	xid = rmsgp->rm_xid;
+	xid = *rdma_resp;
 
 #ifdef SVCRDMA_BACKCHANNEL_DEBUG
 	pr_info("%s: xid=%08x, length=%zu\n",
 		__func__, be32_to_cpu(xid), len);
 	pr_info("%s: RPC/RDMA: %*ph\n",
-		__func__, (int)RPCRDMA_HDRLEN_MIN, rmsgp);
+		__func__, (int)RPCRDMA_HDRLEN_MIN, rdma_resp);
 	pr_info("%s:      RPC: %*ph\n",
 		__func__, (int)len, p);
 #endif
@@ -53,7 +63,7 @@ int svc_rdma_handle_bc_reply(struct rpc_xprt *xprt, struct rpcrdma_msg *rmsgp,
 		goto out_unlock;
 	memcpy(dst->iov_base, p, len);
 
-	credits = be32_to_cpu(rmsgp->rm_credit);
+	credits = be32_to_cpup(rdma_resp + 2);
 	if (credits == 0)
 		credits = 1;	/* don't deadlock */
 	else if (credits > r_xprt->rx_buf.rb_bc_max_requests)
@@ -90,9 +100,9 @@ out_notfound:
  * Caller holds the connection's mutex and has already marshaled
  * the RPC/RDMA request.
  *
- * This is similar to svc_rdma_reply, but takes an rpc_rqst
- * instead, does not support chunks, and avoids blocking memory
- * allocation.
+ * This is similar to svc_rdma_send_reply_msg, but takes a struct
+ * rpc_rqst instead, does not support chunks, and avoids blocking
+ * memory allocation.
  *
  * XXX: There is still an opportunity to block in svc_rdma_send()
  * if there are no SQ entries to post the Send. This may occur if
@@ -101,59 +111,36 @@ out_notfound:
 static int svc_rdma_bc_sendto(struct svcxprt_rdma *rdma,
 			      struct rpc_rqst *rqst)
 {
-	struct xdr_buf *sndbuf = &rqst->rq_snd_buf;
 	struct svc_rdma_op_ctxt *ctxt;
-	struct svc_rdma_req_map *vec;
-	struct ib_send_wr send_wr;
 	int ret;
 
-	vec = svc_rdma_get_req_map(rdma);
-	ret = svc_rdma_map_xdr(rdma, sndbuf, vec, false);
-	if (ret)
+	ctxt = svc_rdma_get_context(rdma);
+
+	/* rpcrdma_bc_send_request builds the transport header and
+	 * the backchannel RPC message in the same buffer. Thus only
+	 * one SGE is needed to send both.
+	 */
+	ret = svc_rdma_map_reply_hdr(rdma, ctxt, rqst->rq_buffer,
+				     rqst->rq_snd_buf.len);
+	if (ret < 0)
 		goto out_err;
 
 	ret = svc_rdma_repost_recv(rdma, GFP_NOIO);
 	if (ret)
 		goto out_err;
 
-	ctxt = svc_rdma_get_context(rdma);
-	ctxt->pages[0] = virt_to_page(rqst->rq_buffer);
-	ctxt->count = 1;
-
-	ctxt->direction = DMA_TO_DEVICE;
-	ctxt->sge[0].lkey = rdma->sc_pd->local_dma_lkey;
-	ctxt->sge[0].length = sndbuf->len;
-	ctxt->sge[0].addr =
-	    ib_dma_map_page(rdma->sc_cm_id->device, ctxt->pages[0], 0,
-			    sndbuf->len, DMA_TO_DEVICE);
-	if (ib_dma_mapping_error(rdma->sc_cm_id->device, ctxt->sge[0].addr)) {
-		ret = -EIO;
-		goto out_unmap;
-	}
-	svc_rdma_count_mappings(rdma, ctxt);
-
-	memset(&send_wr, 0, sizeof(send_wr));
-	ctxt->cqe.done = svc_rdma_wc_send;
-	send_wr.wr_cqe = &ctxt->cqe;
-	send_wr.sg_list = ctxt->sge;
-	send_wr.num_sge = 1;
-	send_wr.opcode = IB_WR_SEND;
-	send_wr.send_flags = IB_SEND_SIGNALED;
-
-	ret = svc_rdma_send(rdma, &send_wr);
-	if (ret) {
-		ret = -EIO;
+	ret = svc_rdma_post_send_wr(rdma, ctxt, 1, 0);
+	if (ret)
 		goto out_unmap;
-	}
 
 out_err:
-	svc_rdma_put_req_map(rdma, vec);
 	dprintk("svcrdma: %s returns %d\n", __func__, ret);
 	return ret;
 
 out_unmap:
 	svc_rdma_unmap_dma(ctxt);
 	svc_rdma_put_context(ctxt, 1);
+	ret = -EIO;
 	goto out_err;
 }
 
diff --git a/net/sunrpc/xprtrdma/svc_rdma_marshal.c b/net/sunrpc/xprtrdma/svc_rdma_marshal.c
index 1c4aabf0f657..bdcf7d85a3dc 100644
--- a/net/sunrpc/xprtrdma/svc_rdma_marshal.c
+++ b/net/sunrpc/xprtrdma/svc_rdma_marshal.c
@@ -166,92 +166,3 @@ out_inval:
 	dprintk("svcrdma: failed to parse transport header\n");
 	return -EINVAL;
 }
-
-int svc_rdma_xdr_encode_error(struct svcxprt_rdma *xprt,
-			      struct rpcrdma_msg *rmsgp,
-			      enum rpcrdma_errcode err, __be32 *va)
-{
-	__be32 *startp = va;
-
-	*va++ = rmsgp->rm_xid;
-	*va++ = rmsgp->rm_vers;
-	*va++ = xprt->sc_fc_credits;
-	*va++ = rdma_error;
-	*va++ = cpu_to_be32(err);
-	if (err == ERR_VERS) {
-		*va++ = rpcrdma_version;
-		*va++ = rpcrdma_version;
-	}
-
-	return (int)((unsigned long)va - (unsigned long)startp);
-}
-
-/**
- * svc_rdma_xdr_get_reply_hdr_length - Get length of Reply transport header
- * @rdma_resp: buffer containing Reply transport header
- *
- * Returns length of transport header, in bytes.
- */
-unsigned int svc_rdma_xdr_get_reply_hdr_len(__be32 *rdma_resp)
-{
-	unsigned int nsegs;
-	__be32 *p;
-
-	p = rdma_resp;
-
-	/* RPC-over-RDMA V1 replies never have a Read list. */
-	p += rpcrdma_fixed_maxsz + 1;
-
-	/* Skip Write list. */
-	while (*p++ != xdr_zero) {
-		nsegs = be32_to_cpup(p++);
-		p += nsegs * rpcrdma_segment_maxsz;
-	}
-
-	/* Skip Reply chunk. */
-	if (*p++ != xdr_zero) {
-		nsegs = be32_to_cpup(p++);
-		p += nsegs * rpcrdma_segment_maxsz;
-	}
-
-	return (unsigned long)p - (unsigned long)rdma_resp;
-}
-
-void svc_rdma_xdr_encode_write_list(struct rpcrdma_msg *rmsgp, int chunks)
-{
-	struct rpcrdma_write_array *ary;
-
-	/* no read-list */
-	rmsgp->rm_body.rm_chunks[0] = xdr_zero;
-
-	/* write-array discrim */
-	ary = (struct rpcrdma_write_array *)
-		&rmsgp->rm_body.rm_chunks[1];
-	ary->wc_discrim = xdr_one;
-	ary->wc_nchunks = cpu_to_be32(chunks);
-
-	/* write-list terminator */
-	ary->wc_array[chunks].wc_target.rs_handle = xdr_zero;
-
-	/* reply-array discriminator */
-	ary->wc_array[chunks].wc_target.rs_length = xdr_zero;
-}
-
-void svc_rdma_xdr_encode_reply_array(struct rpcrdma_write_array *ary,
-				 int chunks)
-{
-	ary->wc_discrim = xdr_one;
-	ary->wc_nchunks = cpu_to_be32(chunks);
-}
-
-void svc_rdma_xdr_encode_array_chunk(struct rpcrdma_write_array *ary,
-				     int chunk_no,
-				     __be32 rs_handle,
-				     __be64 rs_offset,
-				     u32 write_len)
-{
-	struct rpcrdma_segment *seg = &ary->wc_array[chunk_no].wc_target;
-	seg->rs_handle = rs_handle;
-	seg->rs_offset = rs_offset;
-	seg->rs_length = cpu_to_be32(write_len);
-}
diff --git a/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c b/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c
index f7b2daf72a86..27a99bf5b1a6 100644
--- a/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c
+++ b/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c
@@ -558,33 +558,85 @@ static void rdma_read_complete(struct svc_rqst *rqstp,
 	rqstp->rq_arg.buflen = head->arg.buflen;
 }
 
+static void svc_rdma_send_error(struct svcxprt_rdma *xprt,
+				__be32 *rdma_argp, int status)
+{
+	struct svc_rdma_op_ctxt *ctxt;
+	__be32 *p, *err_msgp;
+	unsigned int length;
+	struct page *page;
+	int ret;
+
+	ret = svc_rdma_repost_recv(xprt, GFP_KERNEL);
+	if (ret)
+		return;
+
+	page = alloc_page(GFP_KERNEL);
+	if (!page)
+		return;
+	err_msgp = page_address(page);
+
+	p = err_msgp;
+	*p++ = *rdma_argp;
+	*p++ = *(rdma_argp + 1);
+	*p++ = xprt->sc_fc_credits;
+	*p++ = rdma_error;
+	if (status == -EPROTONOSUPPORT) {
+		*p++ = err_vers;
+		*p++ = rpcrdma_version;
+		*p++ = rpcrdma_version;
+	} else {
+		*p++ = err_chunk;
+	}
+	length = (unsigned long)p - (unsigned long)err_msgp;
+
+	/* Map transport header; no RPC message payload */
+	ctxt = svc_rdma_get_context(xprt);
+	ret = svc_rdma_map_reply_hdr(xprt, ctxt, err_msgp, length);
+	if (ret) {
+		dprintk("svcrdma: Error %d mapping send for protocol error\n",
+			ret);
+		return;
+	}
+
+	ret = svc_rdma_post_send_wr(xprt, ctxt, 1, 0);
+	if (ret) {
+		dprintk("svcrdma: Error %d posting send for protocol error\n",
+			ret);
+		svc_rdma_unmap_dma(ctxt);
+		svc_rdma_put_context(ctxt, 1);
+	}
+}
+
 /* By convention, backchannel calls arrive via rdma_msg type
  * messages, and never populate the chunk lists. This makes
  * the RPC/RDMA header small and fixed in size, so it is
  * straightforward to check the RPC header's direction field.
  */
-static bool
-svc_rdma_is_backchannel_reply(struct svc_xprt *xprt, struct rpcrdma_msg *rmsgp)
+static bool svc_rdma_is_backchannel_reply(struct svc_xprt *xprt,
+					  __be32 *rdma_resp)
 {
-	__be32 *p = (__be32 *)rmsgp;
+	__be32 *p;
 
 	if (!xprt->xpt_bc_xprt)
 		return false;
 
-	if (rmsgp->rm_type != rdma_msg)
+	p = rdma_resp + 3;
+	if (*p++ != rdma_msg)
 		return false;
-	if (rmsgp->rm_body.rm_chunks[0] != xdr_zero)
+
+	if (*p++ != xdr_zero)
 		return false;
-	if (rmsgp->rm_body.rm_chunks[1] != xdr_zero)
+	if (*p++ != xdr_zero)
 		return false;
-	if (rmsgp->rm_body.rm_chunks[2] != xdr_zero)
+	if (*p++ != xdr_zero)
 		return false;
 
-	/* sanity */
-	if (p[7] != rmsgp->rm_xid)
+	/* XID sanity */
+	if (*p++ != *rdma_resp)
 		return false;
 	/* call direction */
-	if (p[8] == cpu_to_be32(RPC_CALL))
+	if (*p == cpu_to_be32(RPC_CALL))
 		return false;
 
 	return true;
@@ -650,8 +702,9 @@ int svc_rdma_recvfrom(struct svc_rqst *rqstp)
 		goto out_drop;
 	rqstp->rq_xprt_hlen = ret;
 
-	if (svc_rdma_is_backchannel_reply(xprt, rmsgp)) {
-		ret = svc_rdma_handle_bc_reply(xprt->xpt_bc_xprt, rmsgp,
+	if (svc_rdma_is_backchannel_reply(xprt, &rmsgp->rm_xid)) {
+		ret = svc_rdma_handle_bc_reply(xprt->xpt_bc_xprt,
+					       &rmsgp->rm_xid,
 					       &rqstp->rq_arg);
 		svc_rdma_put_context(ctxt, 0);
 		if (ret)
@@ -686,7 +739,7 @@ complete:
 	return ret;
 
 out_err:
-	svc_rdma_send_error(rdma_xprt, rmsgp, ret);
+	svc_rdma_send_error(rdma_xprt, &rmsgp->rm_xid, ret);
 	svc_rdma_put_context(ctxt, 0);
 	return 0;
 
diff --git a/net/sunrpc/xprtrdma/svc_rdma_rw.c b/net/sunrpc/xprtrdma/svc_rdma_rw.c
new file mode 100644
index 000000000000..0cf620277693
--- /dev/null
+++ b/net/sunrpc/xprtrdma/svc_rdma_rw.c
@@ -0,0 +1,512 @@
+/*
+ * Copyright (c) 2016 Oracle.  All rights reserved.
+ *
+ * Use the core R/W API to move RPC-over-RDMA Read and Write chunks.
+ */
+
+#include <linux/sunrpc/rpc_rdma.h>
+#include <linux/sunrpc/svc_rdma.h>
+#include <linux/sunrpc/debug.h>
+
+#include <rdma/rw.h>
+
+#define RPCDBG_FACILITY	RPCDBG_SVCXPRT
+
+/* Each R/W context contains state for one chain of RDMA Read or
+ * Write Work Requests.
+ *
+ * Each WR chain handles a single contiguous server-side buffer,
+ * because scatterlist entries after the first have to start on
+ * page alignment. xdr_buf iovecs cannot guarantee alignment.
+ *
+ * Each WR chain handles only one R_key. Each RPC-over-RDMA segment
+ * from a client may contain a unique R_key, so each WR chain moves
+ * up to one segment at a time.
+ *
+ * The scatterlist makes this data structure over 4KB in size. To
+ * make it less likely to fail, and to handle the allocation for
+ * smaller I/O requests without disabling bottom-halves, these
+ * contexts are created on demand, but cached and reused until the
+ * controlling svcxprt_rdma is destroyed.
+ */
+struct svc_rdma_rw_ctxt {
+	struct list_head	rw_list;
+	struct rdma_rw_ctx	rw_ctx;
+	int			rw_nents;
+	struct sg_table		rw_sg_table;
+	struct scatterlist	rw_first_sgl[0];
+};
+
+static inline struct svc_rdma_rw_ctxt *
+svc_rdma_next_ctxt(struct list_head *list)
+{
+	return list_first_entry_or_null(list, struct svc_rdma_rw_ctxt,
+					rw_list);
+}
+
+static struct svc_rdma_rw_ctxt *
+svc_rdma_get_rw_ctxt(struct svcxprt_rdma *rdma, unsigned int sges)
+{
+	struct svc_rdma_rw_ctxt *ctxt;
+
+	spin_lock(&rdma->sc_rw_ctxt_lock);
+
+	ctxt = svc_rdma_next_ctxt(&rdma->sc_rw_ctxts);
+	if (ctxt) {
+		list_del(&ctxt->rw_list);
+		spin_unlock(&rdma->sc_rw_ctxt_lock);
+	} else {
+		spin_unlock(&rdma->sc_rw_ctxt_lock);
+		ctxt = kmalloc(sizeof(*ctxt) +
+			       SG_CHUNK_SIZE * sizeof(struct scatterlist),
+			       GFP_KERNEL);
+		if (!ctxt)
+			goto out;
+		INIT_LIST_HEAD(&ctxt->rw_list);
+	}
+
+	ctxt->rw_sg_table.sgl = ctxt->rw_first_sgl;
+	if (sg_alloc_table_chained(&ctxt->rw_sg_table, sges,
+				   ctxt->rw_sg_table.sgl)) {
+		kfree(ctxt);
+		ctxt = NULL;
+	}
+out:
+	return ctxt;
+}
+
+static void svc_rdma_put_rw_ctxt(struct svcxprt_rdma *rdma,
+				 struct svc_rdma_rw_ctxt *ctxt)
+{
+	sg_free_table_chained(&ctxt->rw_sg_table, true);
+
+	spin_lock(&rdma->sc_rw_ctxt_lock);
+	list_add(&ctxt->rw_list, &rdma->sc_rw_ctxts);
+	spin_unlock(&rdma->sc_rw_ctxt_lock);
+}
+
+/**
+ * svc_rdma_destroy_rw_ctxts - Free accumulated R/W contexts
+ * @rdma: transport about to be destroyed
+ *
+ */
+void svc_rdma_destroy_rw_ctxts(struct svcxprt_rdma *rdma)
+{
+	struct svc_rdma_rw_ctxt *ctxt;
+
+	while ((ctxt = svc_rdma_next_ctxt(&rdma->sc_rw_ctxts)) != NULL) {
+		list_del(&ctxt->rw_list);
+		kfree(ctxt);
+	}
+}
+
+/* A chunk context tracks all I/O for moving one Read or Write
+ * chunk. This is a a set of rdma_rw's that handle data movement
+ * for all segments of one chunk.
+ *
+ * These are small, acquired with a single allocator call, and
+ * no more than one is needed per chunk. They are allocated on
+ * demand, and not cached.
+ */
+struct svc_rdma_chunk_ctxt {
+	struct ib_cqe		cc_cqe;
+	struct svcxprt_rdma	*cc_rdma;
+	struct list_head	cc_rwctxts;
+	int			cc_sqecount;
+	enum dma_data_direction cc_dir;
+};
+
+static void svc_rdma_cc_init(struct svcxprt_rdma *rdma,
+			     struct svc_rdma_chunk_ctxt *cc,
+			     enum dma_data_direction dir)
+{
+	cc->cc_rdma = rdma;
+	svc_xprt_get(&rdma->sc_xprt);
+
+	INIT_LIST_HEAD(&cc->cc_rwctxts);
+	cc->cc_sqecount = 0;
+	cc->cc_dir = dir;
+}
+
+static void svc_rdma_cc_release(struct svc_rdma_chunk_ctxt *cc)
+{
+	struct svcxprt_rdma *rdma = cc->cc_rdma;
+	struct svc_rdma_rw_ctxt *ctxt;
+
+	while ((ctxt = svc_rdma_next_ctxt(&cc->cc_rwctxts)) != NULL) {
+		list_del(&ctxt->rw_list);
+
+		rdma_rw_ctx_destroy(&ctxt->rw_ctx, rdma->sc_qp,
+				    rdma->sc_port_num, ctxt->rw_sg_table.sgl,
+				    ctxt->rw_nents, cc->cc_dir);
+		svc_rdma_put_rw_ctxt(rdma, ctxt);
+	}
+	svc_xprt_put(&rdma->sc_xprt);
+}
+
+/* State for sending a Write or Reply chunk.
+ *  - Tracks progress of writing one chunk over all its segments
+ *  - Stores arguments for the SGL constructor functions
+ */
+struct svc_rdma_write_info {
+	/* write state of this chunk */
+	unsigned int		wi_seg_off;
+	unsigned int		wi_seg_no;
+	unsigned int		wi_nsegs;
+	__be32			*wi_segs;
+
+	/* SGL constructor arguments */
+	struct xdr_buf		*wi_xdr;
+	unsigned char		*wi_base;
+	unsigned int		wi_next_off;
+
+	struct svc_rdma_chunk_ctxt	wi_cc;
+};
+
+static struct svc_rdma_write_info *
+svc_rdma_write_info_alloc(struct svcxprt_rdma *rdma, __be32 *chunk)
+{
+	struct svc_rdma_write_info *info;
+
+	info = kmalloc(sizeof(*info), GFP_KERNEL);
+	if (!info)
+		return info;
+
+	info->wi_seg_off = 0;
+	info->wi_seg_no = 0;
+	info->wi_nsegs = be32_to_cpup(++chunk);
+	info->wi_segs = ++chunk;
+	svc_rdma_cc_init(rdma, &info->wi_cc, DMA_TO_DEVICE);
+	return info;
+}
+
+static void svc_rdma_write_info_free(struct svc_rdma_write_info *info)
+{
+	svc_rdma_cc_release(&info->wi_cc);
+	kfree(info);
+}
+
+/**
+ * svc_rdma_write_done - Write chunk completion
+ * @cq: controlling Completion Queue
+ * @wc: Work Completion
+ *
+ * Pages under I/O are freed by a subsequent Send completion.
+ */
+static void svc_rdma_write_done(struct ib_cq *cq, struct ib_wc *wc)
+{
+	struct ib_cqe *cqe = wc->wr_cqe;
+	struct svc_rdma_chunk_ctxt *cc =
+			container_of(cqe, struct svc_rdma_chunk_ctxt, cc_cqe);
+	struct svcxprt_rdma *rdma = cc->cc_rdma;
+	struct svc_rdma_write_info *info =
+			container_of(cc, struct svc_rdma_write_info, wi_cc);
+
+	atomic_add(cc->cc_sqecount, &rdma->sc_sq_avail);
+	wake_up(&rdma->sc_send_wait);
+
+	if (unlikely(wc->status != IB_WC_SUCCESS)) {
+		set_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags);
+		if (wc->status != IB_WC_WR_FLUSH_ERR)
+			pr_err("svcrdma: write ctx: %s (%u/0x%x)\n",
+			       ib_wc_status_msg(wc->status),
+			       wc->status, wc->vendor_err);
+	}
+
+	svc_rdma_write_info_free(info);
+}
+
+/* This function sleeps when the transport's Send Queue is congested.
+ *
+ * Assumptions:
+ * - If ib_post_send() succeeds, only one completion is expected,
+ *   even if one or more WRs are flushed. This is true when posting
+ *   an rdma_rw_ctx or when posting a single signaled WR.
+ */
+static int svc_rdma_post_chunk_ctxt(struct svc_rdma_chunk_ctxt *cc)
+{
+	struct svcxprt_rdma *rdma = cc->cc_rdma;
+	struct svc_xprt *xprt = &rdma->sc_xprt;
+	struct ib_send_wr *first_wr, *bad_wr;
+	struct list_head *tmp;
+	struct ib_cqe *cqe;
+	int ret;
+
+	first_wr = NULL;
+	cqe = &cc->cc_cqe;
+	list_for_each(tmp, &cc->cc_rwctxts) {
+		struct svc_rdma_rw_ctxt *ctxt;
+
+		ctxt = list_entry(tmp, struct svc_rdma_rw_ctxt, rw_list);
+		first_wr = rdma_rw_ctx_wrs(&ctxt->rw_ctx, rdma->sc_qp,
+					   rdma->sc_port_num, cqe, first_wr);
+		cqe = NULL;
+	}
+
+	do {
+		if (atomic_sub_return(cc->cc_sqecount,
+				      &rdma->sc_sq_avail) > 0) {
+			ret = ib_post_send(rdma->sc_qp, first_wr, &bad_wr);
+			if (ret)
+				break;
+			return 0;
+		}
+
+		atomic_inc(&rdma_stat_sq_starve);
+		atomic_add(cc->cc_sqecount, &rdma->sc_sq_avail);
+		wait_event(rdma->sc_send_wait,
+			   atomic_read(&rdma->sc_sq_avail) > cc->cc_sqecount);
+	} while (1);
+
+	pr_err("svcrdma: ib_post_send failed (%d)\n", ret);
+	set_bit(XPT_CLOSE, &xprt->xpt_flags);
+
+	/* If even one was posted, there will be a completion. */
+	if (bad_wr != first_wr)
+		return 0;
+
+	atomic_add(cc->cc_sqecount, &rdma->sc_sq_avail);
+	wake_up(&rdma->sc_send_wait);
+	return -ENOTCONN;
+}
+
+/* Build and DMA-map an SGL that covers one kvec in an xdr_buf
+ */
+static void svc_rdma_vec_to_sg(struct svc_rdma_write_info *info,
+			       unsigned int len,
+			       struct svc_rdma_rw_ctxt *ctxt)
+{
+	struct scatterlist *sg = ctxt->rw_sg_table.sgl;
+
+	sg_set_buf(&sg[0], info->wi_base, len);
+	info->wi_base += len;
+
+	ctxt->rw_nents = 1;
+}
+
+/* Build and DMA-map an SGL that covers part of an xdr_buf's pagelist.
+ */
+static void svc_rdma_pagelist_to_sg(struct svc_rdma_write_info *info,
+				    unsigned int remaining,
+				    struct svc_rdma_rw_ctxt *ctxt)
+{
+	unsigned int sge_no, sge_bytes, page_off, page_no;
+	struct xdr_buf *xdr = info->wi_xdr;
+	struct scatterlist *sg;
+	struct page **page;
+
+	page_off = (info->wi_next_off + xdr->page_base) & ~PAGE_MASK;
+	page_no = (info->wi_next_off + xdr->page_base) >> PAGE_SHIFT;
+	page = xdr->pages + page_no;
+	info->wi_next_off += remaining;
+	sg = ctxt->rw_sg_table.sgl;
+	sge_no = 0;
+	do {
+		sge_bytes = min_t(unsigned int, remaining,
+				  PAGE_SIZE - page_off);
+		sg_set_page(sg, *page, sge_bytes, page_off);
+
+		remaining -= sge_bytes;
+		sg = sg_next(sg);
+		page_off = 0;
+		sge_no++;
+		page++;
+	} while (remaining);
+
+	ctxt->rw_nents = sge_no;
+}
+
+/* Construct RDMA Write WRs to send a portion of an xdr_buf containing
+ * an RPC Reply.
+ */
+static int
+svc_rdma_build_writes(struct svc_rdma_write_info *info,
+		      void (*constructor)(struct svc_rdma_write_info *info,
+					  unsigned int len,
+					  struct svc_rdma_rw_ctxt *ctxt),
+		      unsigned int remaining)
+{
+	struct svc_rdma_chunk_ctxt *cc = &info->wi_cc;
+	struct svcxprt_rdma *rdma = cc->cc_rdma;
+	struct svc_rdma_rw_ctxt *ctxt;
+	__be32 *seg;
+	int ret;
+
+	cc->cc_cqe.done = svc_rdma_write_done;
+	seg = info->wi_segs + info->wi_seg_no * rpcrdma_segment_maxsz;
+	do {
+		unsigned int write_len;
+		u32 seg_length, seg_handle;
+		u64 seg_offset;
+
+		if (info->wi_seg_no >= info->wi_nsegs)
+			goto out_overflow;
+
+		seg_handle = be32_to_cpup(seg);
+		seg_length = be32_to_cpup(seg + 1);
+		xdr_decode_hyper(seg + 2, &seg_offset);
+		seg_offset += info->wi_seg_off;
+
+		write_len = min(remaining, seg_length - info->wi_seg_off);
+		ctxt = svc_rdma_get_rw_ctxt(rdma,
+					    (write_len >> PAGE_SHIFT) + 2);
+		if (!ctxt)
+			goto out_noctx;
+
+		constructor(info, write_len, ctxt);
+		ret = rdma_rw_ctx_init(&ctxt->rw_ctx, rdma->sc_qp,
+				       rdma->sc_port_num, ctxt->rw_sg_table.sgl,
+				       ctxt->rw_nents, 0, seg_offset,
+				       seg_handle, DMA_TO_DEVICE);
+		if (ret < 0)
+			goto out_initerr;
+
+		list_add(&ctxt->rw_list, &cc->cc_rwctxts);
+		cc->cc_sqecount += ret;
+		if (write_len == seg_length - info->wi_seg_off) {
+			seg += 4;
+			info->wi_seg_no++;
+			info->wi_seg_off = 0;
+		} else {
+			info->wi_seg_off += write_len;
+		}
+		remaining -= write_len;
+	} while (remaining);
+
+	return 0;
+
+out_overflow:
+	dprintk("svcrdma: inadequate space in Write chunk (%u)\n",
+		info->wi_nsegs);
+	return -E2BIG;
+
+out_noctx:
+	dprintk("svcrdma: no R/W ctxs available\n");
+	return -ENOMEM;
+
+out_initerr:
+	svc_rdma_put_rw_ctxt(rdma, ctxt);
+	pr_err("svcrdma: failed to map pagelist (%d)\n", ret);
+	return -EIO;
+}
+
+/* Send one of an xdr_buf's kvecs by itself. To send a Reply
+ * chunk, the whole RPC Reply is written back to the client.
+ * This function writes either the head or tail of the xdr_buf
+ * containing the Reply.
+ */
+static int svc_rdma_send_xdr_kvec(struct svc_rdma_write_info *info,
+				  struct kvec *vec)
+{
+	info->wi_base = vec->iov_base;
+	return svc_rdma_build_writes(info, svc_rdma_vec_to_sg,
+				     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.
+ */
+static int svc_rdma_send_xdr_pagelist(struct svc_rdma_write_info *info,
+				      struct xdr_buf *xdr)
+{
+	info->wi_xdr = xdr;
+	info->wi_next_off = 0;
+	return svc_rdma_build_writes(info, svc_rdma_pagelist_to_sg,
+				     xdr->page_len);
+}
+
+/**
+ * svc_rdma_send_write_chunk - Write all segments in a Write chunk
+ * @rdma: controlling RDMA transport
+ * @wr_ch: Write chunk provided by client
+ * @xdr: xdr_buf containing the data payload
+ *
+ * Returns a non-negative number of bytes the chunk consumed, or
+ *	%-E2BIG if the payload was larger than the Write chunk,
+ *	%-ENOMEM if rdma_rw context pool was exhausted,
+ *	%-ENOTCONN if posting failed (connection is lost),
+ *	%-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 svc_rdma_write_info *info;
+	int ret;
+
+	if (!xdr->page_len)
+		return 0;
+
+	info = svc_rdma_write_info_alloc(rdma, wr_ch);
+	if (!info)
+		return -ENOMEM;
+
+	ret = svc_rdma_send_xdr_pagelist(info, xdr);
+	if (ret < 0)
+		goto out_err;
+
+	ret = svc_rdma_post_chunk_ctxt(&info->wi_cc);
+	if (ret < 0)
+		goto out_err;
+	return xdr->page_len;
+
+out_err:
+	svc_rdma_write_info_free(info);
+	return ret;
+}
+
+/**
+ * svc_rdma_send_reply_chunk - Write all segments in the Reply chunk
+ * @rdma: controlling RDMA transport
+ * @rp_ch: Reply chunk provided by client
+ * @writelist: true if client provided a Write list
+ * @xdr: xdr_buf containing an RPC Reply
+ *
+ * Returns a non-negative number of bytes the chunk consumed, or
+ *	%-E2BIG if the payload was larger than the Reply chunk,
+ *	%-ENOMEM if rdma_rw context pool was exhausted,
+ *	%-ENOTCONN if posting failed (connection is lost),
+ *	%-EIO if rdma_rw initialization failed (DMA mapping, etc).
+ */
+int svc_rdma_send_reply_chunk(struct svcxprt_rdma *rdma, __be32 *rp_ch,
+			      bool writelist, struct xdr_buf *xdr)
+{
+	struct svc_rdma_write_info *info;
+	int consumed, ret;
+
+	info = svc_rdma_write_info_alloc(rdma, rp_ch);
+	if (!info)
+		return -ENOMEM;
+
+	ret = svc_rdma_send_xdr_kvec(info, &xdr->head[0]);
+	if (ret < 0)
+		goto out_err;
+	consumed = xdr->head[0].iov_len;
+
+	/* Send the page list in the Reply chunk only if the
+	 * client did not provide Write chunks.
+	 */
+	if (!writelist && xdr->page_len) {
+		ret = svc_rdma_send_xdr_pagelist(info, xdr);
+		if (ret < 0)
+			goto out_err;
+		consumed += xdr->page_len;
+	}
+
+	if (xdr->tail[0].iov_len) {
+		ret = svc_rdma_send_xdr_kvec(info, &xdr->tail[0]);
+		if (ret < 0)
+			goto out_err;
+		consumed += xdr->tail[0].iov_len;
+	}
+
+	ret = svc_rdma_post_chunk_ctxt(&info->wi_cc);
+	if (ret < 0)
+		goto out_err;
+	return consumed;
+
+out_err:
+	svc_rdma_write_info_free(info);
+	return ret;
+}
diff --git a/net/sunrpc/xprtrdma/svc_rdma_sendto.c b/net/sunrpc/xprtrdma/svc_rdma_sendto.c
index 515221b16d09..1736337f3a55 100644
--- a/net/sunrpc/xprtrdma/svc_rdma_sendto.c
+++ b/net/sunrpc/xprtrdma/svc_rdma_sendto.c
@@ -1,4 +1,5 @@
 /*
+ * Copyright (c) 2016 Oracle. All rights reserved.
  * Copyright (c) 2014 Open Grid Computing, Inc. All rights reserved.
  * Copyright (c) 2005-2006 Network Appliance, Inc. All rights reserved.
  *
@@ -40,6 +41,63 @@
  * Author: Tom Tucker <tom@opengridcomputing.com>
  */
 
+/* Operation
+ *
+ * The main entry point is svc_rdma_sendto. This is called by the
+ * RPC server when an RPC Reply is ready to be transmitted to a client.
+ *
+ * The passed-in svc_rqst contains a struct xdr_buf which holds an
+ * XDR-encoded RPC Reply message. sendto must construct the RPC-over-RDMA
+ * transport header, post all Write WRs needed for this Reply, then post
+ * a Send WR conveying the transport header and the RPC message itself to
+ * the client.
+ *
+ * svc_rdma_sendto must fully transmit the Reply before returning, as
+ * the svc_rqst will be recycled as soon as sendto returns. Remaining
+ * resources referred to by the svc_rqst are also recycled at that time.
+ * Therefore any resources that must remain longer must be detached
+ * from the svc_rqst and released later.
+ *
+ * Page Management
+ *
+ * The I/O that performs Reply transmission is asynchronous, and may
+ * complete well after sendto returns. Thus pages under I/O must be
+ * removed from the svc_rqst before sendto returns.
+ *
+ * The logic here depends on Send Queue and completion ordering. Since
+ * the Send WR is always posted last, it will always complete last. Thus
+ * when it completes, it is guaranteed that all previous Write WRs have
+ * also completed.
+ *
+ * Write WRs are constructed and posted. Each Write segment gets its own
+ * svc_rdma_rw_ctxt, allowing the Write completion handler to find and
+ * DMA-unmap the pages under I/O for that Write segment. The Write
+ * completion handler does not release any pages.
+ *
+ * When the Send WR is constructed, it also gets its own svc_rdma_op_ctxt.
+ * The ownership of all of the Reply's pages are transferred into that
+ * ctxt, the Send WR is posted, and sendto returns.
+ *
+ * The svc_rdma_op_ctxt is presented when the Send WR completes. The
+ * Send completion handler finally releases the Reply's pages.
+ *
+ * This mechanism also assumes that completions on the transport's Send
+ * Completion Queue do not run in parallel. Otherwise a Write completion
+ * and Send completion running at the same time could release pages that
+ * are still DMA-mapped.
+ *
+ * Error Handling
+ *
+ * - If the Send WR is posted successfully, it will either complete
+ *   successfully, or get flushed. Either way, the Send completion
+ *   handler releases the Reply's pages.
+ * - If the Send WR cannot be not posted, the forward path releases
+ *   the Reply's pages.
+ *
+ * This handles the case, without the use of page reference counting,
+ * where two different Write segments send portions of the same page.
+ */
+
 #include <linux/sunrpc/debug.h>
 #include <linux/sunrpc/rpc_rdma.h>
 #include <linux/spinlock.h>
@@ -55,113 +113,141 @@ static u32 xdr_padsize(u32 len)
 	return (len & 3) ? (4 - (len & 3)) : 0;
 }
 
-int svc_rdma_map_xdr(struct svcxprt_rdma *xprt,
-		     struct xdr_buf *xdr,
-		     struct svc_rdma_req_map *vec,
-		     bool write_chunk_present)
+/* Returns length of transport header, in bytes.
+ */
+static unsigned int svc_rdma_reply_hdr_len(__be32 *rdma_resp)
 {
-	int sge_no;
-	u32 sge_bytes;
-	u32 page_bytes;
-	u32 page_off;
-	int page_no;
-
-	if (xdr->len !=
-	    (xdr->head[0].iov_len + xdr->page_len + xdr->tail[0].iov_len)) {
-		pr_err("svcrdma: %s: XDR buffer length error\n", __func__);
-		return -EIO;
-	}
+	unsigned int nsegs;
+	__be32 *p;
 
-	/* Skip the first sge, this is for the RPCRDMA header */
-	sge_no = 1;
+	p = rdma_resp;
+
+	/* RPC-over-RDMA V1 replies never have a Read list. */
+	p += rpcrdma_fixed_maxsz + 1;