Merge tag 'nfs-for-5.5-1' of git://git.linux-nfs.org/projects/trondmy/linux-nfs

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

  Features:

   - NFSv4.2 now supports cross device offloaded copy (i.e. offloaded
     copy of a file from one source server to a different target
     server).

   - New RDMA tracepoints for debugging congestion control and Local
     Invalidate WRs.

  Bugfixes and cleanups

   - Drop the NFSv4.1 session slot if nfs4_delegreturn_prepare waits for
     layoutreturn

   - Handle bad/dead sessions correctly in nfs41_sequence_process()

   - Various bugfixes to the delegation return operation.

   - Various bugfixes pertaining to delegations that have been revoked.

   - Cleanups to the NFS timespec code to avoid unnecessary conversions
     between timespec and timespec64.

   - Fix unstable RDMA connections after a reconnect

   - Close race between waking an RDMA sender and posting a receive

   - Wake pending RDMA tasks if connection fails

   - Fix MR list corruption, and clean up MR usage

   - Fix another RPCSEC_GSS issue with MIC buffer space"

* tag 'nfs-for-5.5-1' of git://git.linux-nfs.org/projects/trondmy/linux-nfs: (79 commits)
  SUNRPC: Capture completion of all RPC tasks
  SUNRPC: Fix another issue with MIC buffer space
  NFS4: Trace lock reclaims
  NFS4: Trace state recovery operation
  NFSv4.2 fix memory leak in nfs42_ssc_open
  NFSv4.2 fix kfree in __nfs42_copy_file_range
  NFS: remove duplicated include from nfs4file.c
  NFSv4: Make _nfs42_proc_copy_notify() static
  NFS: Fallocate should use the nfs4_fattr_bitmap
  NFS: Return -ETXTBSY when attempting to write to a swapfile
  fs: nfs: sysfs: Remove NULL check before kfree
  NFS: remove unneeded semicolon
  NFSv4: add declaration of current_stateid
  NFSv4.x: Drop the slot if nfs4_delegreturn_prepare waits for layoutreturn
  NFSv4.x: Handle bad/dead sessions correctly in nfs41_sequence_process()
  nfsv4: Move NFSPROC4_CLNT_COPY_NOTIFY to end of list
  SUNRPC: Avoid RPC delays when exiting suspend
  NFS: Add a tracepoint in nfs_fh_to_dentry()
  NFSv4: Don't retry the GETATTR on old stateid in nfs4_delegreturn_done()
  NFSv4: Handle NFS4ERR_OLD_STATEID in delegreturn
  ...

11 files changed, 435 insertions, 337 deletions

diff --git a/net/sunrpc/clnt.c b/net/sunrpc/clnt.c
index f7f78566be46..a3379765605d 100644
--- a/net/sunrpc/clnt.c
+++ b/net/sunrpc/clnt.c
@@ -591,6 +591,9 @@ struct rpc_clnt *rpc_create(struct rpc_create_args *args)
 	xprt->resvport = 1;
 	if (args->flags & RPC_CLNT_CREATE_NONPRIVPORT)
 		xprt->resvport = 0;
+	xprt->reuseport = 0;
+	if (args->flags & RPC_CLNT_CREATE_REUSEPORT)
+		xprt->reuseport = 1;
 
 	clnt = rpc_create_xprt(args, xprt);
 	if (IS_ERR(clnt) || args->nconnect <= 1)
@@ -1676,8 +1679,6 @@ call_reserveresult(struct rpc_task *task)
 			return;
 		}
 
-		printk(KERN_ERR "%s: status=%d, but no request slot, exiting\n",
-				__func__, status);
 		rpc_call_rpcerror(task, -EIO);
 		return;
 	}
@@ -1686,11 +1687,8 @@ call_reserveresult(struct rpc_task *task)
 	 * Even though there was an error, we may have acquired
 	 * a request slot somehow.  Make sure not to leak it.
 	 */
-	if (task->tk_rqstp) {
-		printk(KERN_ERR "%s: status=%d, request allocated anyway\n",
-				__func__, status);
+	if (task->tk_rqstp)
 		xprt_release(task);
-	}
 
 	switch (status) {
 	case -ENOMEM:
@@ -1699,14 +1697,9 @@ call_reserveresult(struct rpc_task *task)
 	case -EAGAIN:	/* woken up; retry */
 		task->tk_action = call_retry_reserve;
 		return;
-	case -EIO:	/* probably a shutdown */
-		break;
 	default:
-		printk(KERN_ERR "%s: unrecognized error %d, exiting\n",
-				__func__, status);
-		break;
+		rpc_call_rpcerror(task, status);
 	}
-	rpc_call_rpcerror(task, status);
 }
 
 /*
@@ -2906,7 +2899,7 @@ int rpc_clnt_add_xprt(struct rpc_clnt *clnt,
 	struct rpc_xprt *xprt;
 	unsigned long connect_timeout;
 	unsigned long reconnect_timeout;
-	unsigned char resvport;
+	unsigned char resvport, reuseport;
 	int ret = 0;
 
 	rcu_read_lock();
@@ -2918,6 +2911,7 @@ int rpc_clnt_add_xprt(struct rpc_clnt *clnt,
 		return -EAGAIN;
 	}
 	resvport = xprt->resvport;
+	reuseport = xprt->reuseport;
 	connect_timeout = xprt->connect_timeout;
 	reconnect_timeout = xprt->max_reconnect_timeout;
 	rcu_read_unlock();
@@ -2928,6 +2922,7 @@ int rpc_clnt_add_xprt(struct rpc_clnt *clnt,
 		goto out_put_switch;
 	}
 	xprt->resvport = resvport;
+	xprt->reuseport = reuseport;
 	if (xprt->ops->set_connect_timeout != NULL)
 		xprt->ops->set_connect_timeout(xprt,
 				connect_timeout,
diff --git a/net/sunrpc/sched.c b/net/sunrpc/sched.c
index 360afe153193..9c79548c6847 100644
--- a/net/sunrpc/sched.c
+++ b/net/sunrpc/sched.c
@@ -260,7 +260,7 @@ static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const c
 	rpc_reset_waitqueue_priority(queue);
 	queue->qlen = 0;
 	queue->timer_list.expires = 0;
-	INIT_DEFERRABLE_WORK(&queue->timer_list.dwork, __rpc_queue_timer_fn);
+	INIT_DELAYED_WORK(&queue->timer_list.dwork, __rpc_queue_timer_fn);
 	INIT_LIST_HEAD(&queue->timer_list.list);
 	rpc_assign_waitqueue_name(queue, qname);
 }
@@ -824,6 +824,7 @@ rpc_reset_task_statistics(struct rpc_task *task)
  */
 void rpc_exit_task(struct rpc_task *task)
 {
+	trace_rpc_task_end(task, task->tk_action);
 	task->tk_action = NULL;
 	if (task->tk_ops->rpc_count_stats)
 		task->tk_ops->rpc_count_stats(task, task->tk_calldata);
diff --git a/net/sunrpc/xdr.c b/net/sunrpc/xdr.c
index 14ba9e72a204..f3104be8ff5d 100644
--- a/net/sunrpc/xdr.c
+++ b/net/sunrpc/xdr.c
@@ -436,13 +436,12 @@ xdr_shrink_bufhead(struct xdr_buf *buf, size_t len)
 }
 
 /**
- * xdr_shrink_pagelen
+ * xdr_shrink_pagelen - shrinks buf->pages by up to @len bytes
  * @buf: xdr_buf
  * @len: bytes to remove from buf->pages
  *
- * Shrinks XDR buffer's page array buf->pages by
- * 'len' bytes. The extra data is not lost, but is instead
- * moved into the tail.
+ * The extra data is not lost, but is instead moved into buf->tail.
+ * Returns the actual number of bytes moved.
  */
 static unsigned int
 xdr_shrink_pagelen(struct xdr_buf *buf, size_t len)
@@ -455,8 +454,8 @@ xdr_shrink_pagelen(struct xdr_buf *buf, size_t len)
 
 	result = 0;
 	tail = buf->tail;
-	BUG_ON (len > pglen);
-
+	if (len > buf->page_len)
+		len = buf-> page_len;
 	tailbuf_len = buf->buflen - buf->head->iov_len - buf->page_len;
 
 	/* Shift the tail first */
diff --git a/net/sunrpc/xprt.c b/net/sunrpc/xprt.c
index 41df4c507193..1aafe8d3f3f4 100644
--- a/net/sunrpc/xprt.c
+++ b/net/sunrpc/xprt.c
@@ -205,20 +205,20 @@ int xprt_reserve_xprt(struct rpc_xprt *xprt, struct rpc_task *task)
 
 	if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) {
 		if (task == xprt->snd_task)
-			return 1;
+			goto out_locked;
 		goto out_sleep;
 	}
 	if (test_bit(XPRT_WRITE_SPACE, &xprt->state))
 		goto out_unlock;
 	xprt->snd_task = task;
 
+out_locked:
+	trace_xprt_reserve_xprt(xprt, task);
 	return 1;
 
 out_unlock:
 	xprt_clear_locked(xprt);
 out_sleep:
-	dprintk("RPC: %5u failed to lock transport %p\n",
-			task->tk_pid, xprt);
 	task->tk_status = -EAGAIN;
 	if  (RPC_IS_SOFT(task))
 		rpc_sleep_on_timeout(&xprt->sending, task, NULL,
@@ -269,23 +269,22 @@ int xprt_reserve_xprt_cong(struct rpc_xprt *xprt, struct rpc_task *task)
 
 	if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) {
 		if (task == xprt->snd_task)
-			return 1;
+			goto out_locked;
 		goto out_sleep;
 	}
 	if (req == NULL) {
 		xprt->snd_task = task;
-		return 1;
+		goto out_locked;
 	}
 	if (test_bit(XPRT_WRITE_SPACE, &xprt->state))
 		goto out_unlock;
 	if (!xprt_need_congestion_window_wait(xprt)) {
 		xprt->snd_task = task;
-		return 1;
+		goto out_locked;
 	}
 out_unlock:
 	xprt_clear_locked(xprt);
 out_sleep:
-	dprintk("RPC: %5u failed to lock transport %p\n", task->tk_pid, xprt);
 	task->tk_status = -EAGAIN;
 	if (RPC_IS_SOFT(task))
 		rpc_sleep_on_timeout(&xprt->sending, task, NULL,
@@ -293,6 +292,9 @@ out_sleep:
 	else
 		rpc_sleep_on(&xprt->sending, task, NULL);
 	return 0;
+out_locked:
+	trace_xprt_reserve_cong(xprt, task);
+	return 1;
 }
 EXPORT_SYMBOL_GPL(xprt_reserve_xprt_cong);
 
@@ -357,6 +359,7 @@ void xprt_release_xprt(struct rpc_xprt *xprt, struct rpc_task *task)
 		xprt_clear_locked(xprt);
 		__xprt_lock_write_next(xprt);
 	}
+	trace_xprt_release_xprt(xprt, task);
 }
 EXPORT_SYMBOL_GPL(xprt_release_xprt);
 
@@ -374,6 +377,7 @@ void xprt_release_xprt_cong(struct rpc_xprt *xprt, struct rpc_task *task)
 		xprt_clear_locked(xprt);
 		__xprt_lock_write_next_cong(xprt);
 	}
+	trace_xprt_release_cong(xprt, task);
 }
 EXPORT_SYMBOL_GPL(xprt_release_xprt_cong);
 
@@ -395,8 +399,7 @@ __xprt_get_cong(struct rpc_xprt *xprt, struct rpc_rqst *req)
 {
 	if (req->rq_cong)
 		return 1;
-	dprintk("RPC: %5u xprt_cwnd_limited cong = %lu cwnd = %lu\n",
-			req->rq_task->tk_pid, xprt->cong, xprt->cwnd);
+	trace_xprt_get_cong(xprt, req->rq_task);
 	if (RPCXPRT_CONGESTED(xprt)) {
 		xprt_set_congestion_window_wait(xprt);
 		return 0;
@@ -418,6 +421,7 @@ __xprt_put_cong(struct rpc_xprt *xprt, struct rpc_rqst *req)
 	req->rq_cong = 0;
 	xprt->cong -= RPC_CWNDSCALE;
 	xprt_test_and_clear_congestion_window_wait(xprt);
+	trace_xprt_put_cong(xprt, req->rq_task);
 	__xprt_lock_write_next_cong(xprt);
 }
 
diff --git a/net/sunrpc/xprtrdma/backchannel.c b/net/sunrpc/xprtrdma/backchannel.c
index b458bf53ca69..9d02eae353c6 100644
--- a/net/sunrpc/xprtrdma/backchannel.c
+++ b/net/sunrpc/xprtrdma/backchannel.c
@@ -79,7 +79,7 @@ static int rpcrdma_bc_marshal_reply(struct rpc_rqst *rqst)
 	*p = xdr_zero;
 
 	if (rpcrdma_prepare_send_sges(r_xprt, req, RPCRDMA_HDRLEN_MIN,
-				      &rqst->rq_snd_buf, rpcrdma_noch))
+				      &rqst->rq_snd_buf, rpcrdma_noch_pullup))
 		return -EIO;
 
 	trace_xprtrdma_cb_reply(rqst);
diff --git a/net/sunrpc/xprtrdma/frwr_ops.c b/net/sunrpc/xprtrdma/frwr_ops.c
index 30065a28628c..523722be6a16 100644
--- a/net/sunrpc/xprtrdma/frwr_ops.c
+++ b/net/sunrpc/xprtrdma/frwr_ops.c
@@ -36,8 +36,8 @@
  * connect worker from running concurrently.
  *
  * When the underlying transport disconnects, MRs that are in flight
- * are flushed and are likely unusable. Thus all flushed MRs are
- * destroyed. New MRs are created on demand.
+ * are flushed and are likely unusable. Thus all MRs are destroyed.
+ * New MRs are created on demand.
  */
 
 #include <linux/sunrpc/rpc_rdma.h>
@@ -88,8 +88,10 @@ void frwr_release_mr(struct rpcrdma_mr *mr)
 	kfree(mr);
 }
 
-static void frwr_mr_recycle(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mr *mr)
+static void frwr_mr_recycle(struct rpcrdma_mr *mr)
 {
+	struct rpcrdma_xprt *r_xprt = mr->mr_xprt;
+
 	trace_xprtrdma_mr_recycle(mr);
 
 	if (mr->mr_dir != DMA_NONE) {
@@ -107,32 +109,6 @@ static void frwr_mr_recycle(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mr *mr)
 	frwr_release_mr(mr);
 }
 
-/* MRs are dynamically allocated, so simply clean up and release the MR.
- * A replacement MR will subsequently be allocated on demand.
- */
-static void
-frwr_mr_recycle_worker(struct work_struct *work)
-{
-	struct rpcrdma_mr *mr = container_of(work, struct rpcrdma_mr,
-					     mr_recycle);
-
-	frwr_mr_recycle(mr->mr_xprt, mr);
-}
-
-/* frwr_recycle - Discard MRs
- * @req: request to reset
- *
- * Used after a reconnect. These MRs could be in flight, we can't
- * tell. Safe thing to do is release them.
- */
-void frwr_recycle(struct rpcrdma_req *req)
-{
-	struct rpcrdma_mr *mr;
-
-	while ((mr = rpcrdma_mr_pop(&req->rl_registered)))
-		frwr_mr_recycle(mr->mr_xprt, mr);
-}
-
 /* frwr_reset - Place MRs back on the free list
  * @req: request to reset
  *
@@ -166,9 +142,6 @@ int frwr_init_mr(struct rpcrdma_ia *ia, struct rpcrdma_mr *mr)
 	struct ib_mr *frmr;
 	int rc;
 
-	/* NB: ib_alloc_mr and device drivers typically allocate
-	 *     memory with GFP_KERNEL.
-	 */
 	frmr = ib_alloc_mr(ia->ri_pd, ia->ri_mrtype, depth);
 	if (IS_ERR(frmr))
 		goto out_mr_err;
@@ -180,7 +153,6 @@ int frwr_init_mr(struct rpcrdma_ia *ia, struct rpcrdma_mr *mr)
 	mr->frwr.fr_mr = frmr;
 	mr->mr_dir = DMA_NONE;
 	INIT_LIST_HEAD(&mr->mr_list);
-	INIT_WORK(&mr->mr_recycle, frwr_mr_recycle_worker);
 	init_completion(&mr->frwr.fr_linv_done);
 
 	sg_init_table(sg, depth);
@@ -424,7 +396,7 @@ int frwr_send(struct rpcrdma_ia *ia, struct rpcrdma_req *req)
 	struct ib_send_wr *post_wr;
 	struct rpcrdma_mr *mr;
 
-	post_wr = &req->rl_sendctx->sc_wr;
+	post_wr = &req->rl_wr;
 	list_for_each_entry(mr, &req->rl_registered, mr_list) {
 		struct rpcrdma_frwr *frwr;
 
@@ -440,9 +412,6 @@ int frwr_send(struct rpcrdma_ia *ia, struct rpcrdma_req *req)
 		post_wr = &frwr->fr_regwr.wr;
 	}
 
-	/* If ib_post_send fails, the next ->send_request for
-	 * @req will queue these MRs for recovery.
-	 */
 	return ib_post_send(ia->ri_id->qp, post_wr, NULL);
 }
 
@@ -468,7 +437,7 @@ void frwr_reminv(struct rpcrdma_rep *rep, struct list_head *mrs)
 static void __frwr_release_mr(struct ib_wc *wc, struct rpcrdma_mr *mr)
 {
 	if (wc->status != IB_WC_SUCCESS)
-		rpcrdma_mr_recycle(mr);
+		frwr_mr_recycle(mr);
 	else
 		rpcrdma_mr_put(mr);
 }
@@ -570,7 +539,6 @@ void frwr_unmap_sync(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
 	 */
 	bad_wr = NULL;
 	rc = ib_post_send(r_xprt->rx_ia.ri_id->qp, first, &bad_wr);
-	trace_xprtrdma_post_send(req, rc);
 
 	/* The final LOCAL_INV WR in the chain is supposed to
 	 * do the wake. If it was never posted, the wake will
@@ -583,6 +551,7 @@ void frwr_unmap_sync(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
 
 	/* Recycle MRs in the LOCAL_INV chain that did not get posted.
 	 */
+	trace_xprtrdma_post_linv(req, rc);
 	while (bad_wr) {
 		frwr = container_of(bad_wr, struct rpcrdma_frwr,
 				    fr_invwr);
@@ -590,7 +559,7 @@ void frwr_unmap_sync(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
 		bad_wr = bad_wr->next;
 
 		list_del_init(&mr->mr_list);
-		rpcrdma_mr_recycle(mr);
+		frwr_mr_recycle(mr);
 	}
 }
 
@@ -673,18 +642,18 @@ void frwr_unmap_async(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
 	 */
 	bad_wr = NULL;
 	rc = ib_post_send(r_xprt->rx_ia.ri_id->qp, first, &bad_wr);
-	trace_xprtrdma_post_send(req, rc);
 	if (!rc)
 		return;
 
 	/* Recycle MRs in the LOCAL_INV chain that did not get posted.
 	 */
+	trace_xprtrdma_post_linv(req, rc);
 	while (bad_wr) {
 		frwr = container_of(bad_wr, struct rpcrdma_frwr, fr_invwr);
 		mr = container_of(frwr, struct rpcrdma_mr, frwr);
 		bad_wr = bad_wr->next;
 
-		rpcrdma_mr_recycle(mr);
+		frwr_mr_recycle(mr);
 	}
 
 	/* The final LOCAL_INV WR in the chain is supposed to
diff --git a/net/sunrpc/xprtrdma/rpc_rdma.c b/net/sunrpc/xprtrdma/rpc_rdma.c
index b86b5fd62d9f..aec3beb93b25 100644
--- a/net/sunrpc/xprtrdma/rpc_rdma.c
+++ b/net/sunrpc/xprtrdma/rpc_rdma.c
@@ -78,8 +78,6 @@ static unsigned int rpcrdma_max_call_header_size(unsigned int maxsegs)
 	size += rpcrdma_segment_maxsz * sizeof(__be32);
 	size += sizeof(__be32);	/* list discriminator */
 
-	dprintk("RPC:       %s: max call header size = %u\n",
-		__func__, size);
 	return size;
 }
 
@@ -100,8 +98,6 @@ static unsigned int rpcrdma_max_reply_header_size(unsigned int maxsegs)
 	size += maxsegs * rpcrdma_segment_maxsz * sizeof(__be32);
 	size += sizeof(__be32);	/* list discriminator */
 
-	dprintk("RPC:       %s: max reply header size = %u\n",
-		__func__, size);
 	return size;
 }
 
@@ -363,8 +359,7 @@ static struct rpcrdma_mr_seg *rpcrdma_mr_prepare(struct rpcrdma_xprt *r_xprt,
 out_getmr_err:
 	trace_xprtrdma_nomrs(req);
 	xprt_wait_for_buffer_space(&r_xprt->rx_xprt);
-	if (r_xprt->rx_ep.rep_connected != -ENODEV)
-		schedule_work(&r_xprt->rx_buf.rb_refresh_worker);
+	rpcrdma_mrs_refresh(r_xprt);
 	return ERR_PTR(-EAGAIN);
 }
 
@@ -393,7 +388,7 @@ static int rpcrdma_encode_read_list(struct rpcrdma_xprt *r_xprt,
 	unsigned int pos;
 	int nsegs;
 
-	if (rtype == rpcrdma_noch)
+	if (rtype == rpcrdma_noch_pullup || rtype == rpcrdma_noch_mapped)
 		goto done;
 
 	pos = rqst->rq_snd_buf.head[0].iov_len;
@@ -565,6 +560,7 @@ static void rpcrdma_sendctx_done(struct kref *kref)
  */
 void rpcrdma_sendctx_unmap(struct rpcrdma_sendctx *sc)
 {
+	struct rpcrdma_regbuf *rb = sc->sc_req->rl_sendbuf;
 	struct ib_sge *sge;
 
 	if (!sc->sc_unmap_count)
@@ -576,7 +572,7 @@ void rpcrdma_sendctx_unmap(struct rpcrdma_sendctx *sc)
 	 */
 	for (sge = &sc->sc_sges[2]; sc->sc_unmap_count;
 	     ++sge, --sc->sc_unmap_count)
-		ib_dma_unmap_page(sc->sc_device, sge->addr, sge->length,
+		ib_dma_unmap_page(rdmab_device(rb), sge->addr, sge->length,
 				  DMA_TO_DEVICE);
 
 	kref_put(&sc->sc_req->rl_kref, rpcrdma_sendctx_done);
@@ -589,149 +585,228 @@ static bool rpcrdma_prepare_hdr_sge(struct rpcrdma_xprt *r_xprt,
 {
 	struct rpcrdma_sendctx *sc = req->rl_sendctx;
 	struct rpcrdma_regbuf *rb = req->rl_rdmabuf;
-	struct ib_sge *sge = sc->sc_sges;
+	struct ib_sge *sge = &sc->sc_sges[req->rl_wr.num_sge++];
 
 	if (!rpcrdma_regbuf_dma_map(r_xprt, rb))
-		goto out_regbuf;
+		return false;
 	sge->addr = rdmab_addr(rb);
 	sge->length = len;
 	sge->lkey = rdmab_lkey(rb);
 
 	ib_dma_sync_single_for_device(rdmab_device(rb), sge->addr, sge->length,
 				      DMA_TO_DEVICE);
-	sc->sc_wr.num_sge++;
 	return true;
-
-out_regbuf:
-	pr_err("rpcrdma: failed to DMA map a Send buffer\n");
-	return false;
 }
 
-/* Prepare the Send SGEs. The head and tail iovec, and each entry
- * in the page list, gets its own SGE.
+/* The head iovec is straightforward, as it is usually already
+ * DMA-mapped. Sync the content that has changed.
  */
-static bool rpcrdma_prepare_msg_sges(struct rpcrdma_xprt *r_xprt,
-				     struct rpcrdma_req *req,
-				     struct xdr_buf *xdr,
-				     enum rpcrdma_chunktype rtype)
+static bool rpcrdma_prepare_head_iov(struct rpcrdma_xprt *r_xprt,
+				     struct rpcrdma_req *req, unsigned int len)
 {
 	struct rpcrdma_sendctx *sc = req->rl_sendctx;
-	unsigned int sge_no, page_base, len, remaining;
+	struct ib_sge *sge = &sc->sc_sges[req->rl_wr.num_sge++];
 	struct rpcrdma_regbuf *rb = req->rl_sendbuf;
-	struct ib_sge *sge = sc->sc_sges;
-	struct page *page, **ppages;
 
-	/* The head iovec is straightforward, as it is already
-	 * DMA-mapped. Sync the content that has changed.
-	 */
 	if (!rpcrdma_regbuf_dma_map(r_xprt, rb))
-		goto out_regbuf;
-	sc->sc_device = rdmab_device(rb);
-	sge_no = 1;
-	sge[sge_no].addr = rdmab_addr(rb);
-	sge[sge_no].length = xdr->head[0].iov_len;
-	sge[sge_no].lkey = rdmab_lkey(rb);
-	ib_dma_sync_single_for_device(rdmab_device(rb), sge[sge_no].addr,
-				      sge[sge_no].length, DMA_TO_DEVICE);
-
-	/* If there is a Read chunk, the page list is being handled
-	 * via explicit RDMA, and thus is skipped here. However, the
-	 * tail iovec may include an XDR pad for the page list, as
-	 * well as additional content, and may not reside in the
-	 * same page as the head iovec.
-	 */
-	if (rtype == rpcrdma_readch) {
-		len = xdr->tail[0].iov_len;
+		return false;
 
-		/* Do not include the tail if it is only an XDR pad */
-		if (len < 4)
-			goto out;
+	sge->addr = rdmab_addr(rb);
+	sge->length = len;
+	sge->lkey = rdmab_lkey(rb);
 
-		page = virt_to_page(xdr->tail[0].iov_base);
-		page_base = offset_in_page(xdr->tail[0].iov_base);
+	ib_dma_sync_single_for_device(rdmab_device(rb), sge->addr, sge->length,
+				      DMA_TO_DEVICE);
+	return true;
+}
 
-		/* If the content in the page list is an odd length,
-		 * xdr_write_pages() has added a pad at the beginning
-		 * of the tail iovec. Force the tail's non-pad content
-		 * to land at the next XDR position in the Send message.
-		 */
-		page_base += len & 3;
-		len -= len & 3;
-		goto map_tail;
-	}
+/* If there is a page list present, DMA map and prepare an
+ * SGE for each page to be sent.
+ */
+static bool rpcrdma_prepare_pagelist(struct rpcrdma_req *req,
+				     struct xdr_buf *xdr)
+{
+	struct rpcrdma_sendctx *sc = req->rl_sendctx;
+	struct rpcrdma_regbuf *rb = req->rl_sendbuf;
+	unsigned int page_base, len, remaining;
+	struct page **ppages;
+	struct ib_sge *sge;
 
-	/* If there is a page list present, temporarily DMA map
-	 * and prepare an SGE for each page to be sent.
-	 */
-	if (xdr->page_len) {
-		ppages = xdr->pages + (xdr->page_base >> PAGE_SHIFT);
-		page_base = offset_in_page(xdr->page_base);
-		remaining = xdr->page_len;
-		while (remaining) {
-			sge_no++;
-			if (sge_no > RPCRDMA_MAX_SEND_SGES - 2)
-				goto out_mapping_overflow;
-
-			len = min_t(u32, PAGE_SIZE - page_base, remaining);
-			sge[sge_no].addr =
-				ib_dma_map_page(rdmab_device(rb), *ppages,
-						page_base, len, DMA_TO_DEVICE);
-			if (ib_dma_mapping_error(rdmab_device(rb),
-						 sge[sge_no].addr))
-				goto out_mapping_err;
-			sge[sge_no].length = len;
-			sge[sge_no].lkey = rdmab_lkey(rb);
-
-			sc->sc_unmap_count++;
-			ppages++;
-			remaining -= len;
-			page_base = 0;
-		}
-	}
+	ppages = xdr->pages + (xdr->page_base >> PAGE_SHIFT);
+	page_base = offset_in_page(xdr->page_base);
+	remaining = xdr->page_len;
+	while (remaining) {
+		sge = &sc->sc_sges[req->rl_wr.num_sge++];
+		len = min_t(unsigned int, PAGE_SIZE - page_base, remaining);
+		sge->addr = ib_dma_map_page(rdmab_device(rb), *ppages,
+					    page_base, len, DMA_TO_DEVICE);
+		if (ib_dma_mapping_error(rdmab_device(rb), sge->addr))
+			goto out_mapping_err;
 
-	/* The tail iovec is not always constructed in the same
-	 * page where the head iovec resides (see, for example,
-	 * gss_wrap_req_priv). To neatly accommodate that case,
-	 * DMA map it separately.
-	 */
-	if (xdr->tail[0].iov_len) {
-		page = virt_to_page(xdr->tail[0].iov_base);
-		page_base = offset_in_page(xdr->tail[0].iov_base);
-		len = xdr->tail[0].iov_len;
+		sge->length = len;
+		sge->lkey = rdmab_lkey(rb);
 
-map_tail:
-		sge_no++;
-		sge[sge_no].addr =
-			ib_dma_map_page(rdmab_device(rb), page, page_base, len,
-					DMA_TO_DEVICE);
-		if (ib_dma_mapping_error(rdmab_device(rb), sge[sge_no].addr))
-			goto out_mapping_err;
-		sge[sge_no].length = len;
-		sge[sge_no].lkey = rdmab_lkey(rb);
 		sc->sc_unmap_count++;
+		ppages++;
+		remaining -= len;
+		page_base = 0;
 	}
 
-out:
-	sc->sc_wr.num_sge += sge_no;
-	if (sc->sc_unmap_count)
-		kref_get(&req->rl_kref);
 	return true;
 
-out_regbuf:
-	pr_err("rpcrdma: failed to DMA map a Send buffer\n");
+out_mapping_err:
+	trace_xprtrdma_dma_maperr(sge->addr);
 	return false;
+}
 
-out_mapping_overflow:
-	rpcrdma_sendctx_unmap(sc);
-	pr_err("rpcrdma: too many Send SGEs (%u)\n", sge_no);
-	return false;
+/* The tail iovec may include an XDR pad for the page list,
+ * as well as additional content, and may not reside in the
+ * same page as the head iovec.
+ */
+static bool rpcrdma_prepare_tail_iov(struct rpcrdma_req *req,
+				     struct xdr_buf *xdr,
+				     unsigned int page_base, unsigned int len)
+{
+	struct rpcrdma_sendctx *sc = req->rl_sendctx;
+	struct ib_sge *sge = &sc->sc_sges[req->rl_wr.num_sge++];
+	struct rpcrdma_regbuf *rb = req->rl_sendbuf;
+	struct page *page = virt_to_page(xdr->tail[0].iov_base);
+
+	sge->addr = ib_dma_map_page(rdmab_device(rb), page, page_base, len,
+				    DMA_TO_DEVICE);
+	if (ib_dma_mapping_error(rdmab_device(rb), sge->addr))
+		goto out_mapping_err;
+
+	sge->length = len;
+	sge->lkey = rdmab_lkey(rb);
+	++sc->sc_unmap_count;
+	return true;
 
 out_mapping_err:
-	rpcrdma_sendctx_unmap(sc);
-	trace_xprtrdma_dma_maperr(sge[sge_no].addr);
+	trace_xprtrdma_dma_maperr(sge->addr);
 	return false;
 }
 
+/* Copy the tail to the end of the head buffer.
+ */
+static void rpcrdma_pullup_tail_iov(struct rpcrdma_xprt *r_xprt,
+				    struct rpcrdma_req *req,
+				    struct xdr_buf *xdr)
+{
+	unsigned char *dst;
+
+	dst = (unsigned char *)xdr->head[0].iov_base;
+	dst += xdr->head[0].iov_len + xdr->page_len;
+	memmove(dst, xdr->tail[0].iov_base, xdr->tail[0].iov_len);
+	r_xprt->rx_stats.pullup_copy_count += xdr->tail[0].iov_len;
+}
+
+/* Copy pagelist content into the head buffer.
+ */
+static void rpcrdma_pullup_pagelist(struct rpcrdma_xprt *r_xprt,
+				    struct rpcrdma_req *req,
+				    struct xdr_buf *xdr)
+{
+	unsigned int len, page_base, remaining;
+	struct page **ppages;
+	unsigned char *src, *dst;
+
+	dst = (unsigned char *)xdr->head[0].iov_base;
+	dst += xdr->head[0].iov_len;
+	ppages = xdr->pages + (xdr->page_base >> PAGE_SHIFT);
+	page_base = offset_in_page(xdr->page_base);
+	remaining = xdr->page_len;
+	while (remaining) {
+		src = page_address(*ppages);
+		src += page_base;
+		len = min_t(unsigned int, PAGE_SIZE - page_base, remaining);
+		memcpy(dst, src, len);
+		r_xprt->rx_stats.pullup_copy_count += len;
+
+		ppages++;
+		dst += len;
+		remaining -= len;
+		page_base = 0;
+	}
+}
+
+/* Copy the contents of @xdr into @rl_sendbuf and DMA sync it.
+ * When the head, pagelist, and tail are small, a pull-up copy
+ * is considerably less costly than DMA mapping the components
+ * of @xdr.
+ *
+ * Assumptions:
+ *  - the caller has already verified that the total length
+ *    of the RPC Call body will fit into @rl_sendbuf.
+ */
+static bool rpcrdma_prepare_noch_pullup(struct rpcrdma_xprt *r_xprt,
+					struct rpcrdma_req *req,
+					struct xdr_buf *xdr)
+{
+	if (unlikely(xdr->tail[0].iov_len))
+		rpcrdma_pullup_tail_iov(r_xprt, req, xdr);
+
+	if (unlikely(xdr->page_len))
+		rpcrdma_pullup_pagelist(r_xprt, req, xdr);
+
+	/* The whole RPC message resides in the head iovec now */
+	return rpcrdma_prepare_head_iov(r_xprt, req, xdr->len);
+}
+
+static bool rpcrdma_prepare_noch_mapped(struct rpcrdma_xprt *r_xprt,
+					struct rpcrdma_req *req,
+					struct xdr_buf *xdr)
+{
+	struct kvec *tail = &xdr->tail[0];
+
+	if (!rpcrdma_prepare_head_iov(r_xprt, req, xdr->head[0].iov_len))
+		return false;
+	if (xdr->page_len)
+		if (!rpcrdma_prepare_pagelist(req, xdr))
+			return false;
+	if (tail->iov_len)
+		if (!rpcrdma_prepare_tail_iov(req, xdr,
+					      offset_in_page(tail->iov_base),
+					      tail->iov_len))
+			return false;
+
+	if (req->rl_sendctx->sc_unmap_count)
+		kref_get(&req->rl_kref);
+	return true;
+}
+
+static bool rpcrdma_prepare_readch(struct rpcrdma_xprt *r_xprt,
+				   struct rpcrdma_req *req,
+				   struct xdr_buf *xdr)
+{
+	if (!rpcrdma_prepare_head_iov(r_xprt, req, xdr->head[0].iov_len))
+		return false;
+
+	/* If there is a Read chunk, the page list is being handled
+	 * via explicit RDMA, and thus is skipped here.
+	 */
+
+	/* Do not include the tail if it is only an XDR pad */
+	if (xdr->tail[0].iov_len > 3) {
+		unsigned int page_base, len;
+
+		/* If the content in the page list is an odd length,
+		 * xdr_write_pages() adds a pad at the beginning of
+		 * the tail iovec. Force the tail's non-pad content to
+		 * land at the next XDR position in the Send message.
+		 */
+		page_base = offset_in_page(xdr->tail[0].iov_base);
+		len = xdr->tail[0].iov_len;
+		page_base += len & 3;
+		len -= len & 3;
+		if (!rpcrdma_prepare_tail_iov(req, xdr, page_base, len))
+			return false;
+		kref_get(&req->rl_kref);
+	}
+
+	return true;
+}
+
 /**
  * rpcrdma_prepare_send_sges - Construct SGEs for a Send WR
  * @r_xprt: controlling transport
@@ -742,31 +817,53 @@ out_mapping_err:
  *
  * Returns 0 on success; otherwise a negative errno is returned.
  */
-int
-rpcrdma_prepare_send_sges(struct rpcrdma_xprt *r_xprt,
-			  struct rpcrdma_req *req, u32 hdrlen,
-			  struct xdr_buf *xdr, enum rpcrdma_chunktype rtype)
+inline int rpcrdma_prepare_send_sges(struct rpcrdma_xprt *r_xprt,
+				     struct rpcrdma_req *req, u32 hdrlen,
+				     struct xdr_buf *xdr,
+				     enum rpcrdma_chunktype rtype)
 {
 	int ret;
 
 	ret = -EAGAIN;
 	req->rl_sendctx = rpcrdma_sendctx_get_locked(r_xprt);
 	if (!req->rl_sendctx)
-		goto err;
-	req->rl_sendctx->sc_wr.num_sge = 0;
+		goto out_nosc;
 	req->rl_sendctx->sc_unmap_count = 0;
 	req->rl_sendctx->sc_req = req;
 	kref_init(&req->rl_kref);
+	req->rl_wr.wr_cqe = &req->rl_sendctx->sc_cqe;
+	req->rl_wr.sg_list = req->rl_sendctx->sc_sges;
+	req->rl_wr.num_sge = 0;
+	req->rl_wr.opcode = IB_WR_SEND;
 
 	ret = -EIO;
 	if (!rpcrdma_prepare_hdr_sge(r_xprt, req, hdrlen))
-		goto err;
-	if (rtype != rpcrdma_areadch)
-		if (!rpcrdma_prepare_msg_sges(r_xprt, req, xdr, rtype))
-			goto err;
+		goto out_unmap;
+
+	switch (rtype) {
+	case rpcrdma_noch_pullup:
+		if (!rpcrdma_prepare_noch_pullup(r_xprt, req, xdr))
+			goto out_unmap;
+		break;
+	case rpcrdma_noch_mapped:
+		if (!rpcrdma_prepare_noch_mapped(r_xprt, req, xdr))
+			goto out_unmap;
+		break;
+	case rpcrdma_readch:
+		if (!rpcrdma_prepare_readch(r_xprt, req, xdr))
+			goto out_unmap;
+		break;
+	case rpcrdma_areadch:
+		break;
+	default:
+		goto out_unmap;
+	}
+
 	return 0;
 
-err:
+out_unmap:
+	rpcrdma_sendctx_unmap(req->rl_sendctx);
+out_nosc:
 	trace_xprtrdma_prepsend_failed(&req->rl_slot, ret);
 	return ret;
 }
@@ -796,6 +893,7 @@ rpcrdma_marshal_req(struct rpcrdma_xprt *r_xprt, struct rpc_rqst *rqst)
 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 	struct xdr_stream *xdr = &req->rl_stream;
 	enum rpcrdma_chunktype rtype, wtype;
+	struct xdr_buf *buf = &rqst->rq_snd_buf;
 	bool ddp_allowed;
 	__be32 *p;
 	int ret;
@@ -853,8 +951,9 @@ rpcrdma_marshal_req(struct rpcrdma_xprt *r_xprt, struct rpc_rqst *rqst)
 	 */
 	if (rpcrdma_args_inline(r_xprt, rqst)) {
 		*p++ = rdma_msg;
-		rtype = rpcrdma_noch;
-	} else if (ddp_allowed && rqst->rq_snd_buf.flags & XDRBUF_WRITE) {
+		rtype = buf->len < rdmab_length(req->rl_sendbuf) ?
+			rpcrdma_noch_pullup : rpcrdma_noch_mapped;
+	} else if (ddp_allowed && buf->flags & XDRBUF_WRITE) {
 		*p++ = rdma_msg;
 		rtype = rpcrdma_readch;
 	} else {
@@ -863,12 +962,6 @@ rpcrdma_marshal_req(struct rpcrdma_xprt *r_xprt, struct rpc_rqst *rqst)
 		rtype = rpcrdma_areadch;
 	}
 
-	/* If this is a retransmit, discard previously registered
-	 * chunks. Very likely the connection has been replaced,
-	 * so these registrations are invalid and unusable.
-	 */
-	frwr_recycle(req);
-
 	/* This implementation supports the following combinations
 	 * of chunk lists in one RPC-over-RDMA Call message:
 	 *
@@ -902,7 +995,7 @@ rpcrdma_marshal_req(struct rpcrdma_xprt *r_xprt, struct rpc_rqst *rqst)
 		goto out_err;
 
 	ret = rpcrdma_prepare_send_sges(r_xprt, req, req->rl_hdrbuf.len,
-					&rqst->rq_snd_buf, rtype);
+					buf, rtype);
 	if (ret)
 		goto out_err;
 
@@ -916,6 +1009,40 @@ out_err:
 	return ret;
 }
 
+static void __rpcrdma_update_cwnd_locked(struct rpc_xprt *xprt,
+					 struct rpcrdma_buffer *buf,
+					 u32 grant)
+{
+	buf->rb_credits = grant;
+	xprt->cwnd = grant << RPC_CWNDSHIFT;
+}
+
+static void rpcrdma_update_cwnd(struct rpcrdma_xprt *r_xprt, u32 grant)
+{
+	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
+
+	spin_lock(&xprt->transport_lock);
+	__rpcrdma_update_cwnd_locked(xprt, &r_xprt->rx_buf, grant);
+	spin_unlock(&xprt->transport_lock);
+}
+
+/**
+ * rpcrdma_reset_cwnd - Reset the xprt's congestion window
+ * @r_xprt: controlling transport instance
+ *
+ * Prepare @r_xprt for the next connection by reinitializing
+ * its credit grant to one (see RFC 8166, Section 3.3.3).
+ */
+void rpcrdma_reset_cwnd(struct rpcrdma_xprt *r_xprt)
+{
+	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
+