Merge tag 'nfs-for-3.17-1' of git://git.linux-nfs.org/projects/trondmy/linux-nfs

Pull NFS client updates from Trond Myklebust:
 "Highlights include:

   - stable fix for a bug in nfs3_list_one_acl()
   - speed up NFS path walks by supporting LOOKUP_RCU
   - more read/write code cleanups
   - pNFS fixes for layout return on close
   - fixes for the RCU handling in the rpcsec_gss code
   - more NFS/RDMA fixes"

* tag 'nfs-for-3.17-1' of git://git.linux-nfs.org/projects/trondmy/linux-nfs: (79 commits)
  nfs: reject changes to resvport and sharecache during remount
  NFS: Avoid infinite loop when RELEASE_LOCKOWNER getting expired error
  SUNRPC: remove all refcounting of groupinfo from rpcauth_lookupcred
  NFS: fix two problems in lookup_revalidate in RCU-walk
  NFS: allow lockless access to access_cache
  NFS: teach nfs_lookup_verify_inode to handle LOOKUP_RCU
  NFS: teach nfs_neg_need_reval to understand LOOKUP_RCU
  NFS: support RCU_WALK in nfs_permission()
  sunrpc/auth: allow lockless (rcu) lookup of credential cache.
  NFS: prepare for RCU-walk support but pushing tests later in code.
  NFS: nfs4_lookup_revalidate: only evaluate parent if it will be used.
  NFS: add checks for returned value of try_module_get()
  nfs: clear_request_commit while holding i_lock
  pnfs: add pnfs_put_lseg_async
  pnfs: find swapped pages on pnfs commit lists too
  nfs: fix comment and add warn_on for PG_INODE_REF
  nfs: check wait_on_bit_lock err in page_group_lock
  sunrpc: remove "ec" argument from encrypt_v2 operation
  sunrpc: clean up sparse endianness warnings in gss_krb5_wrap.c
  sunrpc: clean up sparse endianness warnings in gss_krb5_seal.c
  ...

4 files changed, 604 insertions, 296 deletions

diff --git a/net/sunrpc/xprtrdma/rpc_rdma.c b/net/sunrpc/xprtrdma/rpc_rdma.c
index 693966d3f33b..6166c985fe24 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 */
@@ -279,13 +271,37 @@ rpcrdma_create_chunks(struct rpc_rqst *rqst, struct xdr_buf *target,
 	return (unsigned char *)iptr - (unsigned char *)headerp;
 
 out:
-	for (pos = 0; nchunks--;)
-		pos += rpcrdma_deregister_external(
-				&req->rl_segments[pos], r_xprt);
+	if (r_xprt->rx_ia.ri_memreg_strategy != RPCRDMA_FRMR) {
+		for (pos = 0; nchunks--;)
+			pos += rpcrdma_deregister_external(
+					&req->rl_segments[pos], r_xprt);
+	}
 	return n;
 }
 
 /*
+ * 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 66f91f0d071a..2faac4940563 100644
--- a/net/sunrpc/xprtrdma/transport.c
+++ b/net/sunrpc/xprtrdma/transport.c
@@ -296,7 +296,6 @@ xprt_setup_rdma(struct xprt_create *args)
 
 	xprt->resvport = 0;		/* privileged port not needed */
 	xprt->tsh_size = 0;		/* RPC-RDMA handles framing */
-	xprt->max_payload = RPCRDMA_MAX_DATA_SEGS * PAGE_SIZE;
 	xprt->ops = &xprt_rdma_procs;
 
 	/*
@@ -382,6 +381,9 @@ xprt_setup_rdma(struct xprt_create *args)
 	new_ep->rep_xprt = xprt;
 
 	xprt_rdma_format_addresses(xprt);
+	xprt->max_payload = rpcrdma_max_payload(new_xprt);
+	dprintk("RPC:       %s: transport data payload maximum: %zu bytes\n",
+		__func__, xprt->max_payload);
 
 	if (!try_module_get(THIS_MODULE))
 		goto out4;
@@ -412,7 +414,7 @@ xprt_rdma_close(struct rpc_xprt *xprt)
 	if (r_xprt->rx_ep.rep_connected > 0)
 		xprt->reestablish_timeout = 0;
 	xprt_disconnect_done(xprt);
-	(void) rpcrdma_ep_disconnect(&r_xprt->rx_ep, &r_xprt->rx_ia);
+	rpcrdma_ep_disconnect(&r_xprt->rx_ep, &r_xprt->rx_ia);
 }
 
 static void
@@ -595,13 +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/verbs.c b/net/sunrpc/xprtrdma/verbs.c
index 13dbd1c389ff..61c41298b4ea 100644
--- a/net/sunrpc/xprtrdma/verbs.c
+++ b/net/sunrpc/xprtrdma/verbs.c
@@ -61,6 +61,8 @@
 # define RPCDBG_FACILITY	RPCDBG_TRANS
 #endif
 
+static void rpcrdma_reset_frmrs(struct rpcrdma_ia *);
+
 /*
  * internal functions
  */
@@ -103,17 +105,6 @@ rpcrdma_run_tasklet(unsigned long data)
 
 static DECLARE_TASKLET(rpcrdma_tasklet_g, rpcrdma_run_tasklet, 0UL);
 
-static inline void
-rpcrdma_schedule_tasklet(struct rpcrdma_rep *rep)
-{
-	unsigned long flags;
-
-	spin_lock_irqsave(&rpcrdma_tk_lock_g, flags);
-	list_add_tail(&rep->rr_list, &rpcrdma_tasklets_g);
-	spin_unlock_irqrestore(&rpcrdma_tk_lock_g, flags);
-	tasklet_schedule(&rpcrdma_tasklet_g);
-}
-
 static void
 rpcrdma_qp_async_error_upcall(struct ib_event *event, void *context)
 {
@@ -153,12 +144,7 @@ rpcrdma_sendcq_process_wc(struct ib_wc *wc)
 	if (wc->wr_id == 0ULL)
 		return;
 	if (wc->status != IB_WC_SUCCESS)
-		return;
-
-	if (wc->opcode == IB_WC_FAST_REG_MR)
-		frmr->r.frmr.state = FRMR_IS_VALID;
-	else if (wc->opcode == IB_WC_LOCAL_INV)
-		frmr->r.frmr.state = FRMR_IS_INVALID;
+		frmr->r.frmr.fr_state = FRMR_IS_STALE;
 }
 
 static int
@@ -217,7 +203,7 @@ rpcrdma_sendcq_upcall(struct ib_cq *cq, void *cq_context)
 }
 
 static void
-rpcrdma_recvcq_process_wc(struct ib_wc *wc)
+rpcrdma_recvcq_process_wc(struct ib_wc *wc, struct list_head *sched_list)
 {
 	struct rpcrdma_rep *rep =
 			(struct rpcrdma_rep *)(unsigned long)wc->wr_id;
@@ -248,28 +234,38 @@ rpcrdma_recvcq_process_wc(struct ib_wc *wc)
 	}
 
 out_schedule:
-	rpcrdma_schedule_tasklet(rep);
+	list_add_tail(&rep->rr_list, sched_list);
 }
 
 static int
 rpcrdma_recvcq_poll(struct ib_cq *cq, struct rpcrdma_ep *ep)
 {
+	struct list_head sched_list;
 	struct ib_wc *wcs;
 	int budget, count, rc;
+	unsigned long flags;
 
+	INIT_LIST_HEAD(&sched_list);
 	budget = RPCRDMA_WC_BUDGET / RPCRDMA_POLLSIZE;
 	do {
 		wcs = ep->rep_recv_wcs;
 
 		rc = ib_poll_cq(cq, RPCRDMA_POLLSIZE, wcs);
 		if (rc <= 0)
-			return rc;
+			goto out_schedule;
 
 		count = rc;
 		while (count-- > 0)
-			rpcrdma_recvcq_process_wc(wcs++);
+			rpcrdma_recvcq_process_wc(wcs++, &sched_list);
 	} while (rc == RPCRDMA_POLLSIZE && --budget);
-	return 0;
+	rc = 0;
+
+out_schedule:
+	spin_lock_irqsave(&rpcrdma_tk_lock_g, flags);
+	list_splice_tail(&sched_list, &rpcrdma_tasklets_g);
+	spin_unlock_irqrestore(&rpcrdma_tk_lock_g, flags);
+	tasklet_schedule(&rpcrdma_tasklet_g);
+	return rc;
 }
 
 /*
@@ -310,6 +306,13 @@ rpcrdma_recvcq_upcall(struct ib_cq *cq, void *cq_context)
 	rpcrdma_recvcq_poll(cq, ep);
 }
 
+static void
+rpcrdma_flush_cqs(struct rpcrdma_ep *ep)
+{
+	rpcrdma_recvcq_upcall(ep->rep_attr.recv_cq, ep);
+	rpcrdma_sendcq_upcall(ep->rep_attr.send_cq, ep);
+}
+
 #ifdef RPC_DEBUG
 static const char * const conn[] = {
 	"address resolved",
@@ -323,8 +326,16 @@ static const char * const conn[] = {
 	"rejected",
 	"established",
 	"disconnected",
-	"device removal"
+	"device removal",
+	"multicast join",
+	"multicast error",
+	"address change",
+	"timewait exit",
 };
+
+#define CONNECTION_MSG(status)						\
+	((status) < ARRAY_SIZE(conn) ?					\
+		conn[(status)] : "unrecognized connection error")
 #endif
 
 static int
@@ -382,23 +393,18 @@ rpcrdma_conn_upcall(struct rdma_cm_id *id, struct rdma_cm_event *event)
 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
 		connstate = -ENODEV;
 connected:
-		dprintk("RPC:       %s: %s: %pI4:%u (ep 0x%p event 0x%x)\n",
-			__func__,
-			(event->event <= 11) ? conn[event->event] :
-						"unknown connection error",
-			&addr->sin_addr.s_addr,
-			ntohs(addr->sin_port),
-			ep, event->event);
 		atomic_set(&rpcx_to_rdmax(ep->rep_xprt)->rx_buf.rb_credits, 1);
 		dprintk("RPC:       %s: %sconnected\n",
 					__func__, connstate > 0 ? "" : "dis");
 		ep->rep_connected = connstate;
 		ep->rep_func(ep);
 		wake_up_all(&ep->rep_connect_wait);
-		break;
+		/*FALLTHROUGH*/
 	default:
-		dprintk("RPC:       %s: unexpected CM event %d\n",
-			__func__, event->event);
+		dprintk("RPC:       %s: %pI4:%u (ep 0x%p): %s\n",
+			__func__, &addr->sin_addr.s_addr,
+			ntohs(addr->sin_port), ep,
+			CONNECTION_MSG(event->event));
 		break;
 	}
 
@@ -558,12 +564,7 @@ rpcrdma_ia_open(struct rpcrdma_xprt *xprt, struct sockaddr *addr, int memreg)
 		if (!ia->ri_id->device->alloc_fmr) {
 			dprintk("RPC:       %s: MTHCAFMR registration "
 				"not supported by HCA\n", __func__);
-#if RPCRDMA_PERSISTENT_REGISTRATION
 			memreg = RPCRDMA_ALLPHYSICAL;
-#else
-			rc = -ENOMEM;
-			goto out2;
-#endif
 		}
 	}
 
@@ -578,20 +579,16 @@ rpcrdma_ia_open(struct rpcrdma_xprt *xprt, struct sockaddr *addr, int memreg)
 	switch (memreg) {
 	case RPCRDMA_FRMR:
 		break;
-#if RPCRDMA_PERSISTENT_REGISTRATION
 	case RPCRDMA_ALLPHYSICAL:
 		mem_priv = IB_ACCESS_LOCAL_WRITE |
 				IB_ACCESS_REMOTE_WRITE |
 				IB_ACCESS_REMOTE_READ;
 		goto register_setup;
-#endif
 	case RPCRDMA_MTHCAFMR:
 		if (ia->ri_have_dma_lkey)
 			break;
 		mem_priv = IB_ACCESS_LOCAL_WRITE;
-#if RPCRDMA_PERSISTENT_REGISTRATION
 	register_setup:
-#endif
 		ia->ri_bind_mem = ib_get_dma_mr(ia->ri_pd, mem_priv);
 		if (IS_ERR(ia->ri_bind_mem)) {
 			printk(KERN_ALERT "%s: ib_get_dma_mr for "
@@ -613,6 +610,7 @@ rpcrdma_ia_open(struct rpcrdma_xprt *xprt, struct sockaddr *addr, int memreg)
 	/* Else will do memory reg/dereg for each chunk */
 	ia->ri_memreg_strategy = memreg;
 
+	rwlock_init(&ia->ri_qplock);
 	return 0;
 out2:
 	rdma_destroy_id(ia->ri_id);
@@ -826,10 +824,7 @@ rpcrdma_ep_destroy(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
 	cancel_delayed_work_sync(&ep->rep_connect_worker);
 
 	if (ia->ri_id->qp) {
-		rc = rpcrdma_ep_disconnect(ep, ia);
-		if (rc)
-			dprintk("RPC:       %s: rpcrdma_ep_disconnect"
-				" returned %i\n", __func__, rc);
+		rpcrdma_ep_disconnect(ep, ia);
 		rdma_destroy_qp(ia->ri_id);
 		ia->ri_id->qp = NULL;
 	}
@@ -859,7 +854,7 @@ rpcrdma_ep_destroy(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
 int
 rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
 {
-	struct rdma_cm_id *id;
+	struct rdma_cm_id *id, *old;
 	int rc = 0;
 	int retry_count = 0;
 
@@ -867,13 +862,12 @@ rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
 		struct rpcrdma_xprt *xprt;
 retry:
 		dprintk("RPC:       %s: reconnecting...\n", __func__);
-		rc = rpcrdma_ep_disconnect(ep, ia);
-		if (rc && rc != -ENOTCONN)
-			dprintk("RPC:       %s: rpcrdma_ep_disconnect"
-				" status %i\n", __func__, rc);
 
-		rpcrdma_clean_cq(ep->rep_attr.recv_cq);
-		rpcrdma_clean_cq(ep->rep_attr.send_cq);
+		rpcrdma_ep_disconnect(ep, ia);
+		rpcrdma_flush_cqs(ep);
+
+		if (ia->ri_memreg_strategy == RPCRDMA_FRMR)
+			rpcrdma_reset_frmrs(ia);
 
 		xprt = container_of(ia, struct rpcrdma_xprt, rx_ia);
 		id = rpcrdma_create_id(xprt, ia,
@@ -905,9 +899,14 @@ retry:
 			rc = -ENETUNREACH;
 			goto out;
 		}
-		rdma_destroy_qp(ia->ri_id);
-		rdma_destroy_id(ia->ri_id);
+
+		write_lock(&ia->ri_qplock);
+		old = ia->ri_id;
 		ia->ri_id = id;
+		write_unlock(&ia->ri_qplock);
+
+		rdma_destroy_qp(old);
+		rdma_destroy_id(old);
 	} else {
 		dprintk("RPC:       %s: connecting...\n", __func__);
 		rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
@@ -974,13 +973,12 @@ out:
  * This call is not reentrant, and must not be made in parallel
  * on the same endpoint.
  */
-int
+void
 rpcrdma_ep_disconnect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
 {
 	int rc;
 
-	rpcrdma_clean_cq(ep->rep_attr.recv_cq);
-	rpcrdma_clean_cq(ep->rep_attr.send_cq);
+	rpcrdma_flush_cqs(ep);
 	rc = rdma_disconnect(ia->ri_id);
 	if (!rc) {
 		/* returns without wait if not connected */
@@ -992,12 +990,93 @@ rpcrdma_ep_disconnect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
 		dprintk("RPC:       %s: rdma_disconnect %i\n", __func__, rc);
 		ep->rep_connected = rc;
 	}
+}
+
+static int
+rpcrdma_init_fmrs(struct rpcrdma_ia *ia, struct rpcrdma_buffer *buf)
+{
+	int mr_access_flags = IB_ACCESS_REMOTE_WRITE | IB_ACCESS_REMOTE_READ;
+	struct ib_fmr_attr fmr_attr = {
+		.max_pages	= RPCRDMA_MAX_DATA_SEGS,
+		.max_maps	= 1,
+		.page_shift	= PAGE_SHIFT
+	};
+	struct rpcrdma_mw *r;
+	int i, rc;
+
+	i = (buf->rb_max_requests + 1) * RPCRDMA_MAX_SEGS;
+	dprintk("RPC:       %s: initalizing %d FMRs\n", __func__, i);
+
+	while (i--) {
+		r = kzalloc(sizeof(*r), GFP_KERNEL);
+		if (r == NULL)
+			return -ENOMEM;
+
+		r->r.fmr = ib_alloc_fmr(ia->ri_pd, mr_access_flags, &fmr_attr);
+		if (IS_ERR(r->r.fmr)) {
+			rc = PTR_ERR(r->r.fmr);
+			dprintk("RPC:       %s: ib_alloc_fmr failed %i\n",
+				__func__, rc);
+			goto out_free;
+		}
+
+		list_add(&r->mw_list, &buf->rb_mws);
+		list_add(&r->mw_all, &buf->rb_all);
+	}
+	return 0;
+
+out_free:
+	kfree(r);
+	return rc;
+}
+
+static int
+rpcrdma_init_frmrs(struct rpcrdma_ia *ia, struct rpcrdma_buffer *buf)
+{
+	struct rpcrdma_frmr *f;
+	struct rpcrdma_mw *r;
+	int i, rc;
+
+	i = (buf->rb_max_requests + 1) * RPCRDMA_MAX_SEGS;
+	dprintk("RPC:       %s: initalizing %d FRMRs\n", __func__, i);
+
+	while (i--) {
+		r = kzalloc(sizeof(*r), GFP_KERNEL);
+		if (r == NULL)
+			return -ENOMEM;
+		f = &r->r.frmr;
+
+		f->fr_mr = ib_alloc_fast_reg_mr(ia->ri_pd,
+						ia->ri_max_frmr_depth);
+		if (IS_ERR(f->fr_mr)) {
+			rc = PTR_ERR(f->fr_mr);
+			dprintk("RPC:       %s: ib_alloc_fast_reg_mr "
+				"failed %i\n", __func__, rc);
+			goto out_free;
+		}
+
+		f->fr_pgl = ib_alloc_fast_reg_page_list(ia->ri_id->device,
+							ia->ri_max_frmr_depth);
+		if (IS_ERR(f->fr_pgl)) {
+			rc = PTR_ERR(f->fr_pgl);
+			dprintk("RPC:       %s: ib_alloc_fast_reg_page_list "
+				"failed %i\n", __func__, rc);
+
+			ib_dereg_mr(f->fr_mr);
+			goto out_free;
+		}
+
+		list_add(&r->mw_list, &buf->rb_mws);
+		list_add(&r->mw_all, &buf->rb_all);
+	}
+
+	return 0;
+
+out_free:
+	kfree(r);
 	return rc;
 }
 
-/*
- * Initialize buffer memory
- */
 int
 rpcrdma_buffer_create(struct rpcrdma_buffer *buf, struct rpcrdma_ep *ep,
 	struct rpcrdma_ia *ia, struct rpcrdma_create_data_internal *cdata)
@@ -1005,7 +1084,6 @@ rpcrdma_buffer_create(struct rpcrdma_buffer *buf, struct rpcrdma_ep *ep,
 	char *p;
 	size_t len, rlen, wlen;
 	int i, rc;
-	struct rpcrdma_mw *r;
 
 	buf->rb_max_requests = cdata->max_requests;
 	spin_lock_init(&buf->rb_lock);
@@ -1016,28 +1094,12 @@ rpcrdma_buffer_create(struct rpcrdma_buffer *buf, struct rpcrdma_ep *ep,
 	 *   2.  arrays of struct rpcrdma_req to fill in pointers
 	 *   3.  array of struct rpcrdma_rep for replies
 	 *   4.  padding, if any
-	 *   5.  mw's, fmr's or frmr's, if any
 	 * Send/recv buffers in req/rep need to be registered
 	 */
-
 	len = buf->rb_max_requests *
 		(sizeof(struct rpcrdma_req *) + sizeof(struct rpcrdma_rep *));
 	len += cdata->padding;
-	switch (ia->ri_memreg_strategy) {
-	case RPCRDMA_FRMR:
-		len += buf->rb_max_requests * RPCRDMA_MAX_SEGS *
-				sizeof(struct rpcrdma_mw);
-		break;
-	case RPCRDMA_MTHCAFMR:
-		/* TBD we are perhaps overallocating here */
-		len += (buf->rb_max_requests + 1) * RPCRDMA_MAX_SEGS *
-				sizeof(struct rpcrdma_mw);
-		break;
-	default:
-		break;
-	}
 
-	/* allocate 1, 4 and 5 in one shot */
 	p = kzalloc(len, GFP_KERNEL);
 	if (p == NULL) {
 		dprintk("RPC:       %s: req_t/rep_t/pad kzalloc(%zd) failed\n",
@@ -1064,51 +1126,17 @@ rpcrdma_buffer_create(struct rpcrdma_buffer *buf, struct rpcrdma_ep *ep,
 	p += cdata->padding;
 
 	INIT_LIST_HEAD(&buf->rb_mws);
-	r = (struct rpcrdma_mw *)p;
+	INIT_LIST_HEAD(&buf->rb_all);
 	switch (ia->ri_memreg_strategy) {
 	case RPCRDMA_FRMR:
-		for (i = buf->rb_max_requests * RPCRDMA_MAX_SEGS; i; i--) {
-			r->r.frmr.fr_mr = ib_alloc_fast_reg_mr(ia->ri_pd,
-						ia->ri_max_frmr_depth);
-			if (IS_ERR(r->r.frmr.fr_mr)) {
-				rc = PTR_ERR(r->r.frmr.fr_mr);
-				dprintk("RPC:       %s: ib_alloc_fast_reg_mr"
-					" failed %i\n", __func__, rc);
-				goto out;
-			}
-			r->r.frmr.fr_pgl = ib_alloc_fast_reg_page_list(
-						ia->ri_id->device,
-						ia->ri_max_frmr_depth);
-			if (IS_ERR(r->r.frmr.fr_pgl)) {
-				rc = PTR_ERR(r->r.frmr.fr_pgl);
-				dprintk("RPC:       %s: "
-					"ib_alloc_fast_reg_page_list "
-					"failed %i\n", __func__, rc);
-
-				ib_dereg_mr(r->r.frmr.fr_mr);
-				goto out;
-			}
-			list_add(&r->mw_list, &buf->rb_mws);
-			++r;
-		}
+		rc = rpcrdma_init_frmrs(ia, buf);
+		if (rc)
+			goto out;
 		break;
 	case RPCRDMA_MTHCAFMR:
-		/* TBD we are perhaps overallocating here */
-		for (i = (buf->rb_max_requests+1) * RPCRDMA_MAX_SEGS; i; i--) {
-			static struct ib_fmr_attr fa =
-				{ RPCRDMA_MAX_DATA_SEGS, 1, PAGE_SHIFT };
-			r->r.fmr = ib_alloc_fmr(ia->ri_pd,
-				IB_ACCESS_REMOTE_WRITE | IB_ACCESS_REMOTE_READ,
-				&fa);
-			if (IS_ERR(r->r.fmr)) {
-				rc = PTR_ERR(r->r.fmr);
-				dprintk("RPC:       %s: ib_alloc_fmr"
-					" failed %i\n", __func__, rc);
-				goto out;
-			}
-			list_add(&r->mw_list, &buf->rb_mws);
-			++r;
-		}
+		rc = rpcrdma_init_fmrs(ia, buf);
+		if (rc)
+			goto out;
 		break;
 	default:
 		break;
@@ -1176,24 +1204,57 @@ out:
 	return rc;
 }
 
-/*
- * Unregister and destroy buffer memory. Need to deal with
- * partial initialization, so it's callable from failed create.
- * Must be called before destroying endpoint, as registrations
- * reference it.
- */
+static void
+rpcrdma_destroy_fmrs(struct rpcrdma_buffer *buf)
+{
+	struct rpcrdma_mw *r;
+	int rc;
+
+	while (!list_empty(&buf->rb_all)) {
+		r = list_entry(buf->rb_all.next, struct rpcrdma_mw, mw_all);
+		list_del(&r->mw_all);
+		list_del(&r->mw_list);
+
+		rc = ib_dealloc_fmr(r->r.fmr);
+		if (rc)
+			dprintk("RPC:       %s: ib_dealloc_fmr failed %i\n",
+				__func__, rc);
+
+		kfree(r);
+	}
+}
+
+static void
+rpcrdma_destroy_frmrs(struct rpcrdma_buffer *buf)
+{
+	struct rpcrdma_mw *r;
+	int rc;
+
+	while (!list_empty(&buf->rb_all)) {
+		r = list_entry(buf->rb_all.next, struct rpcrdma_mw, mw_all);
+		list_del(&r->mw_all);
+		list_del(&r->mw_list);
+
+		rc = ib_dereg_mr(r->r.frmr.fr_mr);
+		if (rc)
+			dprintk("RPC:       %s: ib_dereg_mr failed %i\n",
+				__func__, rc);
+		ib_free_fast_reg_page_list(r->r.frmr.fr_pgl);
+
+		kfree(r);
+	}
+}
+
 void
 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
 {
-	int rc, i;
 	struct rpcrdma_ia *ia = rdmab_to_ia(buf);
-	struct rpcrdma_mw *r;
+	int i;
 
 	/* clean up in reverse order from create
 	 *   1.  recv mr memory (mr free, then kfree)
 	 *   2.  send mr memory (mr free, then kfree)
-	 *   3.  padding (if any) [moved to rpcrdma_ep_destroy]
-	 *   4.  arrays
+	 *   3.  MWs
 	 */
 	dprintk("RPC:       %s: entering\n", __func__);
 
@@ -1212,34 +1273,217 @@ rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
 		}
 	}
 
+	switch (ia->ri_memreg_strategy) {
+	case RPCRDMA_FRMR:
+		rpcrdma_destroy_frmrs(buf);
+		break;
+	case RPCRDMA_MTHCAFMR:
+		rpcrdma_destroy_fmrs(buf);
+		break;
+	default:
+		break;
+	}
+
+	kfree(buf->rb_pool);
+}
+
+/* After a disconnect, a flushed FAST_REG_MR can leave an FRMR in
+ * an unusable state. Find FRMRs in this state and dereg / reg
+ * each.  FRMRs that are VALID and attached to an rpcrdma_req are
+ * also torn down.
+ *
+ * This gives all in-use FRMRs a fresh rkey and leaves them INVALID.
+ *
+ * This is invoked only in the transport connect worker in order
+ * to serialize with rpcrdma_register_frmr_external().
+ */
+static void
+rpcrdma_reset_frmrs(struct rpcrdma_ia *ia)
+{
+	struct rpcrdma_xprt *r_xprt =
+				container_of(ia, struct rpcrdma_xprt, rx_ia);
+	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
+	struct list_head *pos;
+	struct rpcrdma_mw *r;
+	int rc;
+
+	list_for_each(pos, &buf->rb_all) {
+		r = list_entry(pos, struct rpcrdma_mw, mw_all);
+
+		if (r->r.frmr.fr_state == FRMR_IS_INVALID)
+			continue;
+
+		rc = ib_dereg_mr(r->r.frmr.fr_mr);
+		if (rc)
+			dprintk("RPC:       %s: ib_dereg_mr failed %i\n",
+				__func__, rc);
+		ib_free_fast_reg_page_list(r->r.frmr.fr_pgl);
+
+		r->r.frmr.fr_mr = ib_alloc_fast_reg_mr(ia->ri_pd,
+					ia->ri_max_frmr_depth);
+		if (IS_ERR(r->r.frmr.fr_mr)) {
+			rc = PTR_ERR(r->r.frmr.fr_mr);
+			dprintk("RPC:       %s: ib_alloc_fast_reg_mr"
+				" failed %i\n", __func__, rc);
+			continue;
+		}
+		r->r.frmr.fr_pgl = ib_alloc_fast_reg_page_list(
+					ia->ri_id->device,
+					ia->ri_max_frmr_depth);
+		if (IS_ERR(r->r.frmr.fr_pgl)) {
+			rc = PTR_ERR(r->r.frmr.fr_pgl);
+			dprintk("RPC:       %s: "
+				"ib_alloc_fast_reg_page_list "
+				"failed %i\n", __func__, rc);
+
+			ib_dereg_mr(r->r.frmr.fr_mr);
+			continue;
+		}
+		r->r.frmr.fr_state = FRMR_IS_INVALID;
+	}
+}
+
+/* "*mw" can be NULL when rpcrdma_buffer_get_mrs() fails, leaving
+ * some req segments uninitialized.
+ */
+static void
+rpcrdma_buffer_put_mr(struct rpcrdma_mw **mw, struct rpcrdma_buffer *buf)
+{
+	if (*mw) {
+		list_add_tail(&(*mw)->mw_list, &buf->rb_mws);
+		*mw = NULL;
+	}
+}
+
+/* Cycle mw's back in reverse order, and "spin" them.
+ * This delays and scrambles reuse as much as possible.
+ */
+static void
+rpcrdma_buffer_put_mrs(struct rpcrdma_req *req, struct rpcrdma_buffer *buf)
+{
+	struct rpcrdma_mr_seg *seg = req->rl_segments;
+	struct rpcrdma_mr_seg *seg1 = seg;
+	int i;
+
+	for (i = 1, seg++; i < RPCRDMA_MAX_SEGS; seg++, i++)
+		rpcrdma_buffer_put_mr(&seg->mr_chunk.rl_mw, buf);
+	rpcrdma_buffer_put_mr(&seg1->mr_chunk.rl_mw, buf);
+}
+
+static void
+rpcrdma_buffer_put_sendbuf(struct rpcrdma_req *req, struct rpcrdma_buffer *buf)
+{
+	buf->rb_send_bufs[--buf->rb_send_index] = req;
+	req->rl_niovs = 0;
+	if (req->rl_reply) {
+		buf->rb_recv_bufs[--buf->rb_recv_index] = req->rl_reply;
+		req->rl_reply->rr_func = NULL;
+		req->rl_reply = NULL;
+	}
+}
+
+/* rpcrdma_unmap_one() was already done by rpcrdma_deregister_frmr_external().
+ * Redo only the ib_post_send().
+ */
+static void
+rpcrdma_retry_local_inv(struct rpcrdma_mw *r, struct rpcrdma_ia *ia)
+{
+	struct rpcrdma_xprt *r_xprt =
+				container_of(ia, struct rpcrdma_xprt, rx_ia);
+	struct ib_send_wr invalidate_wr, *bad_wr;
+	int rc;
+
+	dprintk("RPC:       %s: FRMR %p is stale\n", __func__, r);
+
+	/* When this FRMR is re-inserted into rb_mws, it is no longer stale */
+	r->r.frmr.fr_state = FRMR_IS_INVALID;
+
+	memset(&invalidate_wr, 0, sizeof(invalidate_wr));
+	invalidate_wr.wr_id = (unsigned long)(void *)r;
+	invalidate_wr.opcode = IB_WR_LOCAL_INV;
+	invalidate_wr.ex.invalidate_rkey = r->r.frmr.fr_mr->rkey;
+	DECR_CQCOUNT(&r_xprt->rx_ep);
+
+	dprintk("RPC:       %s: frmr %p invalidating rkey %08x\n",
+		__func__, r, r->r.frmr.fr_mr->rkey);
+
+	read_lock(&ia->ri_qplock);
+	rc = ib_post_send(ia->ri_id->qp, &invalidate_wr, &bad_wr);
+	read_unlock(&ia->ri_qplock);
+	if (rc) {
+		/* Force rpcrdma_buffer_get() to retry */
+		r->r.frmr.fr_state = FRMR_IS_STALE;
+		dprintk("RPC:       %s: ib_post_send failed, %i\n",
+			__func__, rc);
+	}
+}
+
+static void
+rpcrdma_retry_flushed_linv(struct list_head *stale,
+			   struct rpcrdma_buffer *buf)
+{
+	struct rpcrdma_ia *ia = rdmab_to_ia(buf);
+	struct list_head *pos;
+	struct rpcrdma_mw *r;
+	unsigned long flags;
+
+	list_for_each(pos, stale) {
+		r = list_entry(pos, struct rpcrdma_mw, mw_list);
+		rpcrdma_retry_local_inv(r, ia);
+	}
+
+	spin_lock_irqsave(&buf->rb_lock, flags);
+	list_splice_tail(stale, &buf->rb_mws);
+	spin_unlock_irqrestore(&buf->rb_lock, flags);
+}
+
+static struct rpcrdma_req *
+rpcrdma_buffer_get_frmrs(struct rpcrdma_req *req, struct rpcrdma_buffer *buf,
+			 struct list_head *stale)
+{
+	struct rpcrdma_mw *r;
+	int i;
+
+	i = RPCRDMA_MAX_SEGS - 1;
 	while (!list_empty(&buf->rb_mws)) {
 		r = list_entry(buf->rb_mws.next,
-			struct rpcrdma_mw, mw_list);
+			       struct rpcrdma_mw, mw_list);
 		list_del(&r->mw_list);
-		switch (ia->ri_memreg_strategy) {
-		case RPCRDMA_FRMR:
-			rc = ib_dereg_mr(r->r.frmr.fr_mr);
-			if (rc)
-				dprintk("RPC:       %s:"
-					" ib_dereg_mr"
-					" failed %i\n",
-					__func__, rc);
-			ib_free_fast_reg_page_list(r->r.frmr.fr_pgl);
-			break;
-		case RPCRDMA_MTHCAFMR:
-			rc = ib_dealloc_fmr(r->r.fmr);
-			if (rc)
-				dprintk("RPC:       %s:"
-					" ib_dealloc_fmr"
-					" failed %i\n",
-					__func__, rc);
-			break;
-		default:
-			break;
+		if (r->r.frmr.fr_state == FRMR_IS_STALE) {
+			list_add(&r->mw_list, stale);
+			continue;
 		}
+		req->rl_segments[i].mr_chunk.rl_mw = r;
+		if (unlikely(i-- == 0))
+			return req;	/* Success */
 	}
 
-	kfree(buf->rb_pool);
+	/* Not enough entries on rb_mws for this req */
+	rpcrdma_buffer_put_sendbuf(req, buf);
+	rpcrdma_buffer_put_mrs(req, buf);
+	return NULL;
+}
+
+static struct rpcrdma_req *
+rpcrdma_buffer_get_fmrs(struct rpcrdma_req *req, struct rpcrdma_buffer *buf)
+{
+	struct rpcrdma_mw *r;
+	int i;
+
+	i = RPCRDMA_MAX_SEGS - 1;
+	while (!list_empty(&buf->rb_mws)) {
+		r = list_entry(buf->rb_mws.next,
+			       struct rpcrdma_mw, mw_list);
+		list_del(&r->mw_list);
+		req->rl_segments[i].mr_chunk.rl_mw = r;
+		if (unlikely(i-- == 0))
+			return req;	/* Success */
+	}
+
+	/* Not enough entries on rb_mws for this req */
+	rpcrdma_buffer_put_sendbuf(req, buf);
+	rpcrdma_buffer_put_mrs(req, buf);
+	return NULL;
 }
 
 /*
@@ -1254,10 +1498,10 @@ rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
 struct rpcrdma_req *
 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
 {
+	struct rpcrdma_ia *ia = rdmab_to_ia(buffers);
+	struct list_head stale;
 	struct rpcrdma_req *req;
 	unsigned long flags;
-	int i;
-	struct rpcrdma_mw *r;
 
 	spin_lock_irqsave(&buffers->rb_lock, flags);
 	if (buffers->rb_send_index == buffers->rb_max_requests) {
@@ -1277,16 +1521,21 @@ rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
 		buffers->rb_recv_bufs[buffers->rb_recv_index++] = NULL;
 	}
 	buffers->rb_send_bufs[buffers->rb_send_index++] = NULL;
-	if (!list_empty(&buffers->rb_mws)) {
-		i = RPCRDMA_MAX_SEGS - 1;
-		do {
-			r = list_entry(buffers->rb_mws.next,
-					struct rpcrdma_mw, mw_list);
-			list_del(&r->mw_list);
-			req->rl_segments[i].mr_chunk.rl_mw = r;
-		} while (--i >= 0);
+
+	INIT_LIST_HEAD(&stale);
+	switch (ia->ri_memreg_strategy) {
+	case RPCRDMA_FRMR:
+		req = rpcrdma_buffer_get_frmrs(req, buffers, &stale);
+		break;
+	case RPCRDMA_MTHCAFMR:
+		req = rpcrdma_buffer_get_fmrs(req, buffers);
+		break;
+	default:
+		break;
 	}
 	spin_unlock_irqrestore(&buffers->rb_lock, flags);
+	if (!list_empty(&stale))
+		rpcrdma_retry_flushed_linv(&stale, buffers);
 	return req;
 }
 
@@ -1299,34 +1548,14 @@ rpcrdma_buffer_put(struct rpcrdma_req *req)
 {
 	struct rpcrdma_buffer *buffers = req->rl_buffer;
 	struct rpcrdma_ia *ia = rdmab_to_ia(buffers);
-	int i;
 	unsigned long flags;
 
 	spin_lock_irqsave(&buffers->rb_lock, flags);
-	buffers->rb_send_bufs[--buffers->rb_send_index] = req;
-	req->rl_niovs = 0;
-	if (req->rl_reply) {
-		buffers->rb_recv_bufs[--buffers->rb_recv_index] = req->rl_reply;
-		req->rl_reply->rr_func = NULL;
-		req->rl_reply = NULL;
-	}
+	rpcrdma_buffer_put_sendbuf(req, buffers);
 	switch (ia->ri_memreg_strategy) {
 	case RPCRDMA_FRMR:
 	case RPCRDMA_MTHCAFMR:
-		/*
-		 * Cycle mw's back in reverse order, and "spin" them.
-		 * This delays and scrambles reuse as much as possible.
-		 */
-		i = 1;
-		do {
-			struct rpcrdma_mw **mw;
-			mw = &req->rl_segments[i].mr_chunk.rl_mw;
-			list_add_tail(&(*mw)->mw_list, &buffers->rb_mws);
-			*mw = NULL;
-		} while (++i < RPCRDMA_MAX_SEGS);
-		list_add_tail(&req->rl_segments[0].mr_chunk.rl_mw->mw_list,
-					&buffers->rb_mws);
-		req->rl_segments[0].mr_chunk.rl_mw = NULL;
+		rpcrdma_buffer_put_mrs(req, buffers);
 		break;
 	default:
 		break;
@@ -1388,6 +1617,9 @@ rpcrdma_register_internal(struct rpcrdma_ia *ia, void *va, int len,
 	 */
 	iov->addr = ib_dma_map_single(ia->ri_id->device,
 			va, len, DMA_BIDIRECTIONAL);
+	if (ib_dma_mapping_error(ia->ri_id->device, iov->addr))
+		return -ENOMEM;
+
 	iov->length = len;
 
 	if (ia->ri_have_dma_lkey) {
@@ -1483,8 +1715,10 @@ rpcrdma_register_frmr_external(struct rpcrdma_mr_seg *seg,
 			struct rpcrdma_xprt *r_xprt)
 {
 	struct rpcrdma_mr_seg *seg1 = seg;
-	struct ib_send_wr invalidate_wr, frmr_wr, *bad_wr, *post_wr;
-
+	struct rpcrdma_mw *mw = seg1->mr_chunk.rl_mw;
+	struct rpcrdma_frmr *frmr = &mw->r.frmr;
+	struct ib_mr *mr = frmr->fr_mr;
+	struct ib_send_wr fastreg_wr, *bad_wr;
 	u8 key;
 	int len, pageoff;
 	int i, rc;
@@ -1502,8 +1736,7 @@ rpcrdma_register_frmr_external(struct rpcrdma_mr_seg *seg,
 		rpcrdma_map_one(ia, seg, writing);
 		pa = seg->mr_dma;
 		for (seg_len = seg->mr_len; seg_len > 0; seg_len -= PAGE_SIZE) {
-			seg1->mr_chunk.rl_mw->r.frmr.fr_pgl->
-				page_list[page_no++] = pa;
+			frmr->fr_pgl->page_list[page_no++] = pa;
 			pa += PAGE_SIZE;
 		}
 		len += seg->mr_len;
@@ -1515,65 +1748,51 @@ rpcrdma_register_frmr_external(struct rpcrdma_mr_seg *seg,
 			break;
 	}
 	dprintk("RPC:       %s: Using frmr %p to map %d segments\n",
-		__func__, seg1->mr_chunk.rl_mw, i);
-
-	if (unlikely(seg1->mr_chunk.rl_mw->r.frmr.state == FRMR_IS_VALID)) {
-		dprintk("RPC:       %s: frmr %x left valid, posting invalidate.\n",
-			__func__,
-			seg1->mr_chunk.rl_mw->r.frmr.fr_mr->rkey);
-		/* Invalidate before using. */
-		memset(&invalidate_wr, 0, sizeof invalidate_wr);
-		invalidate_wr.wr_id = (unsigned long)(void *)seg1->mr_chunk.rl_mw;
-		invalidate_wr.next = &frmr_wr;
-		invalidate_wr.opcode = IB_WR_LOCAL_INV;
-		invalidate_wr.send_flags = IB_SEND_SIGNALED;
-		invalidate_wr.ex.invalidate_rkey =
-			seg1->mr_chunk.rl_mw->r.frmr.fr_mr->rkey;
-		DECR_CQCOUNT(&r_xprt->rx_ep);
-		post_wr = &invalidate_wr;
-	} else
-		post_wr = &frmr_wr;
-
-	/* Prepare FRMR WR */
-	memset(&frmr_wr, 0, sizeof frmr_wr);
-	frmr_wr.wr_id = (unsigned long)(void *)seg1->mr_chunk.rl_mw;
-	frmr_wr.opcode = IB_WR_FAST_REG_MR;
-	frmr_wr.send_flags = IB_SEND_SIGNALED;
-	frmr_wr.wr.fast_reg.iova_start = seg1->mr_dma;
-	frmr_wr.wr.fast_reg.page_list = seg1->mr_chunk.rl_mw->r.frmr.fr_pgl;
-	frmr_wr.wr.fast_reg.page_list_len = page_no;
-	frmr_wr.wr.fast_reg.page_shift = PAGE_SHIFT;
-	frmr_wr.wr.fast_reg.length = page_no << PAGE_SHIFT;
-	if (frmr_wr.wr.fast_reg.length < len) {
-		while (seg1->mr_nsegs--)
-			rpcrdma_unmap_one(ia, seg++);
-		return -EIO;
+		__func__, mw, i);
+
+	frmr->fr_state = FRMR_IS_VALID;
+
+	memset(&fastreg_wr, 0, sizeof(fastreg_wr));
+	fastreg_wr.wr_id = (unsigned long)(void *)mw;
+	fastreg_wr.opcode = IB_WR_FAST_REG_MR;
+	fastreg_wr.wr.fast_reg.iova_start = seg1->mr_dma;
+	fastreg_wr.wr.fast_reg.page_list = frmr->fr_pgl;
+	fastreg_wr.wr.fast_reg.page_list_len = page_no;
+	fastreg_wr.wr.fast_reg.page_shift = PAGE_SHIFT;
+	fastreg_wr.wr.fast_reg.length = page_no << PAGE_SHIFT;
+	if (fastreg_wr.wr.fast_reg.length < len) {
+		rc = -EIO;
+		goto out_err;
 	}
 
 	/* Bump the key */
-	key = (u8)(seg1->mr_chunk.rl_mw->r.frmr.fr_mr->rkey & 0x000000FF);
-	ib_update_fast_reg_key(seg1->mr_chunk.rl_mw->r.frmr.fr_mr, ++key);
+	key = (u8)(mr->rkey & 0x000000FF);
+	ib_update_fast_reg_key(mr, ++key);
 
-	frmr_wr.wr.fast_reg.access_flags = (writing ?
+	fastreg_wr.wr.fast_reg.access_flags = (writing ?
 				IB_ACCESS_REMOTE_WRITE | IB_ACCESS_LOCAL_WRITE :
 				IB_ACCESS_REMOTE_READ);
-	frmr_wr.wr.fast_reg.rkey = seg1->mr_chunk.rl_mw->r.frmr.fr_mr->rkey;
+	fastreg_wr.wr.fast_reg.rkey = mr->rkey;
 	DECR_CQCOUNT(&r_xprt->rx_ep);
 
-	rc = ib_post_send(ia->ri_id->qp, post_wr, &bad_wr);
-
+	rc = ib_post_send(ia->ri_id->qp, &fastreg_wr, &bad_wr);
 	if (rc) {
 		dprintk("RPC:       %s: failed ib_post_send for register,"
 			" status %i\n", __func__, rc);
-		while (i--)
-			rpcrdma_unmap_one(ia, --seg);
+		ib_update_fast_reg_key(mr, --key);
+		goto out_err;
 	} else {
-		seg1->mr_rkey = seg1->mr_chunk.rl_mw->r.frmr.fr_mr->rkey;
+		seg1->mr_rkey = mr->rkey;
 		seg1->mr_base = seg1->mr_dma + pageoff;
 		seg1->mr_nsegs = i;
 		seg1->mr_len = len;
 	}
 	*nsegs = i;
+	return 0;
+out_err:
+	frmr->fr_state = FRMR_IS_INVALID;
+	while (i--)
+		rpcrdma_unmap_one(ia, --seg);
 	return rc;
 }
 
@@ -1585,20 +1804,25 @@ rpcrdma_deregister_frmr_external(struct rpcrdma_mr_seg *seg,
 	struct ib_send_wr invalidate_wr, *bad_wr;
 	int rc;
 
-	while (seg1->mr_nsegs--)
-		rpcrdma_unmap_one(ia, seg++);
+	seg1->mr_chunk.rl_mw->r.frmr.fr_state = FRMR_IS_INVALID;
 
 	memset(&invalidate_wr, 0, sizeof invalidate_wr);
 	invalidate_wr.wr_id = (unsigned long)(void *)seg1->mr_chunk.rl_mw;
 	invalidate_wr.opcode = IB_WR_LOCAL_INV;
-	invalidate_wr.send_flags = IB_SEND_SIGNALED;
 	invalidate_wr.ex.invalidate_rkey = seg1->mr_chunk.rl_mw->r.frmr.fr_mr->rkey;
 	DECR_CQCOUNT(&r_xprt->rx_ep);
 
+	read_lock(&ia->ri_qplock);
+	while (seg1->mr_nsegs--)
+		rpcrdma_unmap_one(ia, seg++);
 	rc = ib_post_send(ia->ri_id->qp, &invalidate_wr, &bad_wr);
-	if (rc)
+	read_unlock(&ia->ri_qplock);
+	if (rc) {
+		/* Force rpcrdma_buffer_get() to retry */
+		seg1->mr_chunk.rl_mw->r.frmr.fr_state = FRMR_IS_STALE;
 		dprintk("RPC:       %s: failed ib_post_send for invalidate,"
 			" status %i\n", __func__, rc);
+	}
 	return rc;
 }
 
@@ -1656,8 +1880,10 @@ rpcrdma_deregister_fmr_external(struct rpcrdma_mr_seg *seg,
 
 	list_add(&seg1->mr_chunk.rl_mw->r.fmr->list, &l);
 	rc = ib_unmap_fmr(&l);
+	read_lock(&ia->ri_qplock);
 	while (seg1->mr_nsegs--)
 		rpcrdma_unmap_one(ia, seg++);
+	read_unlock(&ia->ri_qplock);
 	if (rc)
 		dprintk("RPC:       %s: failed ib_unmap_fmr,"
 			" status %i\n", __func__, rc);
@@ -1673,7 +1899,6 @@ rpcrdma_register_external(struct rpcrdma_mr_seg *seg,
 
 	switch (ia->ri_memreg_strategy) {
 
-#if RPCRDMA_PERSISTENT_REGISTRATION
 	case RPCRDMA_ALLPHYSICAL:
 		rpcrdma_map_one(ia, seg, writing);
 		seg->mr_rkey = ia->ri_bind_mem->rkey;
@@ -1681,7 +1906,6 @@ rpcrdma_register_external(struct rpcrdma_mr_seg *seg,
 		seg->mr_nsegs = 1;
 		nsegs = 1;
 		break;
-#endif
 
 	/* Registration using frmr registration */
 	case RPCRDMA_FRMR:
@@ -1711,11 +1935,11 @@ rpcrdma_deregister_external(struct rpcrdma_mr_seg *seg,
 
 	switch (ia->ri_memreg_strategy) {
 
-#if RPCRDMA_PERSISTENT_REGISTRATION
 	case RPCRDMA_ALLPHYSICAL:
+		read_lock(&ia->ri_qplock);
 		rpcrdma_unmap_one(ia, seg);
+		read_unlock(&ia->ri_qplock);
 		break;
-#endif
 
 	case RPCRDMA_FRMR:
 		rc = rpcrdma_deregister_frmr_external(seg, ia, r_xprt);
@@ -1809,3 +2033,44 @@ rpcrdma_ep_post_recv(struct rpcrdma_ia *ia,
 			rc);
 	return rc;
 }
+
+/* Physical mapping means one Read/Write list entry per-page.
+ * All list entries must fit within an inline buffer
+ *
+ * NB: The server must return a Write list for NFS READ,
+ *     which has the same constraint. Factor in the inline
+ *     rsize as well.
+ */
+static size_t
+rpcrdma_physical_max_payload(struct rpcrdma_xprt *r_xprt)
+{
+	struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
+	unsigned int inline_size, pages;
+
+	inline_size = min_t(unsigned int,
+			    cdata->inline_wsize, cdata->inline_rsize);
+	inline_size -= RPCRDMA_HDRLEN_MIN;
+	pages = inline_size / sizeof(struct rpcrdma_segment);
+	return pages << PAGE_SHIFT;
+}
+
+static size_t
+rpcrdma_mr_max_payload(struct rpcrdma_xprt *r_xprt)
+{
+	return RPCRDMA_MAX_DATA_SEGS << PAGE_SHIFT;
+}
+
+size_t
+rpcrdma_max_payload(struct rpcrdma_xprt *r_xprt)
+{
+	size_t result;
+
+	switch (r_xprt->rx_ia.ri_memreg_strategy) {
+	case RPCRDMA_ALLPHYSICAL:
+		result = rpcrdma_physical_max_payload(r_xprt);
+		break;
+	default:
+		result = rpcrdma_mr_max_payload(r_xprt);
+	}
+	return result;
+}
diff --git a/net/sunrpc/xprtrdma/xprt_rdma.h b/net/sunrpc/xprtrdma/xprt_rdma.h
index 89e7cd479705..c419498b8f46 100644
--- a/net/sunrpc/xprtrdma/xprt_rdma.h
+++ b/net/sunrpc/xprtrdma/xprt_rdma.h
@@ -59,6 +59,7 @@
  * Interface Adapter -- one per transport instance
  */
 struct rpcrdma_ia {
+	rwlock_t		ri_qplock;
 	struct rdma_cm_id 	*ri_id;
 	struct ib_pd		*ri_pd;
 	struct ib_mr		*ri_bind_mem;
@@ -98,6 +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
@@ -137,6 +146,40 @@ struct rpcrdma_rep {
 };
 
 /*
+ * struct rpcrdma_mw - external memory region metadata
+ *
+ * An external memory region is any buffer or page that is registered
+ * on the fly (ie, not pre-registered).
+ *
+ * Each rpcrdma_buffer has a list of free MWs anchored in rb_mws. During
+ * call_allocate, rpcrdma_buffer_get() assigns one to each segment in
+ * an rpcrdma_req. Then rpcrdma_register_external() grabs these to keep
+ * track of registration metadata while each RPC is pending.
+ * rpcrdma_deregister_external() uses this metadata to unmap and
+ * release these resources when an RPC is complete.
+ */
+enum rpcrdma_frmr_state {
+	FRMR_IS_INVALID,	/* ready to be used */
+	FRMR_IS_VALID,		/* in use */
+	FRMR_IS_STALE,		/* failed completion */
+};
+
+struct rpcrdma_frmr {
+	struct ib_fast_reg_page_list	*fr_pgl;
+	struct ib_mr			*fr_mr;
+	enum rpcrdma_frmr_state		fr_state;
+};
+
+struct rpcrdma_mw {
+	union {
+		struct ib_fmr		*fmr;
+		struct rpcrdma_frmr	frmr;
+	} r;
+	struct list_head	mw_list;
+	struct list_head	mw_all;
+};
+
+/*
  * struct rpcrdma_req -- structure central to the request/reply sequence.
  *
  * N of these are associated with a transport instance, and stored in
@@ -163,17 +206,7 @@ struct rpcrdma_rep {
 struct rpcrdma_mr_seg {		/* chunk descriptors */
 	union {				/* chunk memory handles */
 		struct ib_mr	*rl_mr;		/* if registered directly */
-		struct rpcrdma_mw {		/* if registered from region */
-			union {
-				struct ib_fmr	*fmr;
-				struct {
-					struct ib_fast_reg_page_list *fr_pgl;
-					struct ib_mr *fr_mr;
-					enum { FRMR_IS_INVALID, FRMR_IS_VALID  } state;
-				} frmr;
-			} r;
-			struct list_head mw_list;
-		} *rl_mw;
+		struct rpcrdma_mw *rl_mw;	/* if registered from region */
 	} mr_chunk;
 	u64		mr_base;	/* registration result */
 	u32		mr_rkey;	/* registration result */
@@ -191,6 +224,7 @@ 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 */
@@ -214,6 +248,7 @@ struct rpcrdma_buffer {
 	atomic_t	rb_credits;	/* most recent server credits */
 	int		rb_max_requests;/* client max requests */
 	struct list_head rb_mws;	/* optional memory windows/fmrs/frmrs */
+	struct list_head rb_all;
 	int		rb_send_index;
 	struct rpcrdma_req	**rb_send_bufs;
 	int		rb_recv_index;
@@ -306,7 +341,7 @@ int rpcrdma_ep_create(struct rpcrdma_ep *, struct rpcrdma_ia *,
 				struct rpcrdma_create_data_internal *);
 void rpcrdma_ep_destroy(struct rpcrdma_ep *, struct rpcrdma_ia *);
 int rpcrdma_ep_connect(struct rpcrdma_ep *, struct rpcrdma_ia *);
-int rpcrdma_ep_disconnect(struct rpcrdma_ep *, struct rpcrdma_ia *);
+void rpcrdma_ep_disconnect(struct rpcrdma_ep *, struct rpcrdma_ia *);
 
 int rpcrdma_ep_post(struct rpcrdma_ia *, struct rpcrdma_ep *,
 				struct rpcrdma_req *);
@@ -346,7 +381,9 @@ 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 *);
 
 /* Temporary NFS request map cache. Created in svc_rdma.c  */
 extern struct kmem_cache *svc_rdma_map_cachep;