RDS/IB: Implement IB-specific datagram send.
authorAndy Grover <andy.grover@oracle.com>
Tue, 24 Feb 2009 15:30:33 +0000 (15:30 +0000)
committerDavid S. Miller <davem@davemloft.net>
Fri, 27 Feb 2009 07:39:31 +0000 (23:39 -0800)
Specific to IB is a credits-based flow control mechanism, in
addition to the expected usage of the IB API to package outgoing
data into work requests.

Signed-off-by: Andy Grover <andy.grover@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/rds/ib_send.c [new file with mode: 0644]

diff --git a/net/rds/ib_send.c b/net/rds/ib_send.c
new file mode 100644 (file)
index 0000000..cb6c52c
--- /dev/null
@@ -0,0 +1,874 @@
+/*
+ * Copyright (c) 2006 Oracle.  All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+#include <linux/kernel.h>
+#include <linux/in.h>
+#include <linux/device.h>
+#include <linux/dmapool.h>
+
+#include "rds.h"
+#include "rdma.h"
+#include "ib.h"
+
+static void rds_ib_send_rdma_complete(struct rds_message *rm,
+                                     int wc_status)
+{
+       int notify_status;
+
+       switch (wc_status) {
+       case IB_WC_WR_FLUSH_ERR:
+               return;
+
+       case IB_WC_SUCCESS:
+               notify_status = RDS_RDMA_SUCCESS;
+               break;
+
+       case IB_WC_REM_ACCESS_ERR:
+               notify_status = RDS_RDMA_REMOTE_ERROR;
+               break;
+
+       default:
+               notify_status = RDS_RDMA_OTHER_ERROR;
+               break;
+       }
+       rds_rdma_send_complete(rm, notify_status);
+}
+
+static void rds_ib_send_unmap_rdma(struct rds_ib_connection *ic,
+                                  struct rds_rdma_op *op)
+{
+       if (op->r_mapped) {
+               ib_dma_unmap_sg(ic->i_cm_id->device,
+                       op->r_sg, op->r_nents,
+                       op->r_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
+               op->r_mapped = 0;
+       }
+}
+
+static void rds_ib_send_unmap_rm(struct rds_ib_connection *ic,
+                         struct rds_ib_send_work *send,
+                         int wc_status)
+{
+       struct rds_message *rm = send->s_rm;
+
+       rdsdebug("ic %p send %p rm %p\n", ic, send, rm);
+
+       ib_dma_unmap_sg(ic->i_cm_id->device,
+                    rm->m_sg, rm->m_nents,
+                    DMA_TO_DEVICE);
+
+       if (rm->m_rdma_op != NULL) {
+               rds_ib_send_unmap_rdma(ic, rm->m_rdma_op);
+
+               /* If the user asked for a completion notification on this
+                * message, we can implement three different semantics:
+                *  1.  Notify when we received the ACK on the RDS message
+                *      that was queued with the RDMA. This provides reliable
+                *      notification of RDMA status at the expense of a one-way
+                *      packet delay.
+                *  2.  Notify when the IB stack gives us the completion event for
+                *      the RDMA operation.
+                *  3.  Notify when the IB stack gives us the completion event for
+                *      the accompanying RDS messages.
+                * Here, we implement approach #3. To implement approach #2,
+                * call rds_rdma_send_complete from the cq_handler. To implement #1,
+                * don't call rds_rdma_send_complete at all, and fall back to the notify
+                * handling in the ACK processing code.
+                *
+                * Note: There's no need to explicitly sync any RDMA buffers using
+                * ib_dma_sync_sg_for_cpu - the completion for the RDMA
+                * operation itself unmapped the RDMA buffers, which takes care
+                * of synching.
+                */
+               rds_ib_send_rdma_complete(rm, wc_status);
+
+               if (rm->m_rdma_op->r_write)
+                       rds_stats_add(s_send_rdma_bytes, rm->m_rdma_op->r_bytes);
+               else
+                       rds_stats_add(s_recv_rdma_bytes, rm->m_rdma_op->r_bytes);
+       }
+
+       /* If anyone waited for this message to get flushed out, wake
+        * them up now */
+       rds_message_unmapped(rm);
+
+       rds_message_put(rm);
+       send->s_rm = NULL;
+}
+
+void rds_ib_send_init_ring(struct rds_ib_connection *ic)
+{
+       struct rds_ib_send_work *send;
+       u32 i;
+
+       for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
+               struct ib_sge *sge;
+
+               send->s_rm = NULL;
+               send->s_op = NULL;
+
+               send->s_wr.wr_id = i;
+               send->s_wr.sg_list = send->s_sge;
+               send->s_wr.num_sge = 1;
+               send->s_wr.opcode = IB_WR_SEND;
+               send->s_wr.send_flags = 0;
+               send->s_wr.ex.imm_data = 0;
+
+               sge = rds_ib_data_sge(ic, send->s_sge);
+               sge->lkey = ic->i_mr->lkey;
+
+               sge = rds_ib_header_sge(ic, send->s_sge);
+               sge->addr = ic->i_send_hdrs_dma + (i * sizeof(struct rds_header));
+               sge->length = sizeof(struct rds_header);
+               sge->lkey = ic->i_mr->lkey;
+       }
+}
+
+void rds_ib_send_clear_ring(struct rds_ib_connection *ic)
+{
+       struct rds_ib_send_work *send;
+       u32 i;
+
+       for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
+               if (send->s_wr.opcode == 0xdead)
+                       continue;
+               if (send->s_rm)
+                       rds_ib_send_unmap_rm(ic, send, IB_WC_WR_FLUSH_ERR);
+               if (send->s_op)
+                       rds_ib_send_unmap_rdma(ic, send->s_op);
+       }
+}
+
+/*
+ * The _oldest/_free ring operations here race cleanly with the alloc/unalloc
+ * operations performed in the send path.  As the sender allocs and potentially
+ * unallocs the next free entry in the ring it doesn't alter which is
+ * the next to be freed, which is what this is concerned with.
+ */
+void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context)
+{
+       struct rds_connection *conn = context;
+       struct rds_ib_connection *ic = conn->c_transport_data;
+       struct ib_wc wc;
+       struct rds_ib_send_work *send;
+       u32 completed;
+       u32 oldest;
+       u32 i = 0;
+       int ret;
+
+       rdsdebug("cq %p conn %p\n", cq, conn);
+       rds_ib_stats_inc(s_ib_tx_cq_call);
+       ret = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
+       if (ret)
+               rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
+
+       while (ib_poll_cq(cq, 1, &wc) > 0) {
+               rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
+                        (unsigned long long)wc.wr_id, wc.status, wc.byte_len,
+                        be32_to_cpu(wc.ex.imm_data));
+               rds_ib_stats_inc(s_ib_tx_cq_event);
+
+               if (wc.wr_id == RDS_IB_ACK_WR_ID) {
+                       if (ic->i_ack_queued + HZ/2 < jiffies)
+                               rds_ib_stats_inc(s_ib_tx_stalled);
+                       rds_ib_ack_send_complete(ic);
+                       continue;
+               }
+
+               oldest = rds_ib_ring_oldest(&ic->i_send_ring);
+
+               completed = rds_ib_ring_completed(&ic->i_send_ring, wc.wr_id, oldest);
+
+               for (i = 0; i < completed; i++) {
+                       send = &ic->i_sends[oldest];
+
+                       /* In the error case, wc.opcode sometimes contains garbage */
+                       switch (send->s_wr.opcode) {
+                       case IB_WR_SEND:
+                               if (send->s_rm)
+                                       rds_ib_send_unmap_rm(ic, send, wc.status);
+                               break;
+                       case IB_WR_RDMA_WRITE:
+                       case IB_WR_RDMA_READ:
+                               /* Nothing to be done - the SG list will be unmapped
+                                * when the SEND completes. */
+                               break;
+                       default:
+                               if (printk_ratelimit())
+                                       printk(KERN_NOTICE
+                                               "RDS/IB: %s: unexpected opcode 0x%x in WR!\n",
+                                               __func__, send->s_wr.opcode);
+                               break;
+                       }
+
+                       send->s_wr.opcode = 0xdead;
+                       send->s_wr.num_sge = 1;
+                       if (send->s_queued + HZ/2 < jiffies)
+                               rds_ib_stats_inc(s_ib_tx_stalled);
+
+                       /* If a RDMA operation produced an error, signal this right
+                        * away. If we don't, the subsequent SEND that goes with this
+                        * RDMA will be canceled with ERR_WFLUSH, and the application
+                        * never learn that the RDMA failed. */
+                       if (unlikely(wc.status == IB_WC_REM_ACCESS_ERR && send->s_op)) {
+                               struct rds_message *rm;
+
+                               rm = rds_send_get_message(conn, send->s_op);
+                               if (rm)
+                                       rds_ib_send_rdma_complete(rm, wc.status);
+                       }
+
+                       oldest = (oldest + 1) % ic->i_send_ring.w_nr;
+               }
+
+               rds_ib_ring_free(&ic->i_send_ring, completed);
+
+               if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags)
+                || test_bit(0, &conn->c_map_queued))
+                       queue_delayed_work(rds_wq, &conn->c_send_w, 0);
+
+               /* We expect errors as the qp is drained during shutdown */
+               if (wc.status != IB_WC_SUCCESS && rds_conn_up(conn)) {
+                       rds_ib_conn_error(conn,
+                               "send completion on %pI4 "
+                               "had status %u, disconnecting and reconnecting\n",
+                               &conn->c_faddr, wc.status);
+               }
+       }
+}
+
+/*
+ * This is the main function for allocating credits when sending
+ * messages.
+ *
+ * Conceptually, we have two counters:
+ *  -  send credits: this tells us how many WRs we're allowed
+ *     to submit without overruning the reciever's queue. For
+ *     each SEND WR we post, we decrement this by one.
+ *
+ *  -  posted credits: this tells us how many WRs we recently
+ *     posted to the receive queue. This value is transferred
+ *     to the peer as a "credit update" in a RDS header field.
+ *     Every time we transmit credits to the peer, we subtract
+ *     the amount of transferred credits from this counter.
+ *
+ * It is essential that we avoid situations where both sides have
+ * exhausted their send credits, and are unable to send new credits
+ * to the peer. We achieve this by requiring that we send at least
+ * one credit update to the peer before exhausting our credits.
+ * When new credits arrive, we subtract one credit that is withheld
+ * until we've posted new buffers and are ready to transmit these
+ * credits (see rds_ib_send_add_credits below).
+ *
+ * The RDS send code is essentially single-threaded; rds_send_xmit
+ * grabs c_send_lock to ensure exclusive access to the send ring.
+ * However, the ACK sending code is independent and can race with
+ * message SENDs.
+ *
+ * In the send path, we need to update the counters for send credits
+ * and the counter of posted buffers atomically - when we use the
+ * last available credit, we cannot allow another thread to race us
+ * and grab the posted credits counter.  Hence, we have to use a
+ * spinlock to protect the credit counter, or use atomics.
+ *
+ * Spinlocks shared between the send and the receive path are bad,
+ * because they create unnecessary delays. An early implementation
+ * using a spinlock showed a 5% degradation in throughput at some
+ * loads.
+ *
+ * This implementation avoids spinlocks completely, putting both
+ * counters into a single atomic, and updating that atomic using
+ * atomic_add (in the receive path, when receiving fresh credits),
+ * and using atomic_cmpxchg when updating the two counters.
+ */
+int rds_ib_send_grab_credits(struct rds_ib_connection *ic,
+                            u32 wanted, u32 *adv_credits, int need_posted)
+{
+       unsigned int avail, posted, got = 0, advertise;
+       long oldval, newval;
+
+       *adv_credits = 0;
+       if (!ic->i_flowctl)
+               return wanted;
+
+try_again:
+       advertise = 0;
+       oldval = newval = atomic_read(&ic->i_credits);
+       posted = IB_GET_POST_CREDITS(oldval);
+       avail = IB_GET_SEND_CREDITS(oldval);
+
+       rdsdebug("rds_ib_send_grab_credits(%u): credits=%u posted=%u\n",
+                       wanted, avail, posted);
+
+       /* The last credit must be used to send a credit update. */
+       if (avail && !posted)
+               avail--;
+
+       if (avail < wanted) {
+               struct rds_connection *conn = ic->i_cm_id->context;
+
+               /* Oops, there aren't that many credits left! */
+               set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
+               got = avail;
+       } else {
+               /* Sometimes you get what you want, lalala. */
+               got = wanted;
+       }
+       newval -= IB_SET_SEND_CREDITS(got);
+
+       /*
+        * If need_posted is non-zero, then the caller wants
+        * the posted regardless of whether any send credits are
+        * available.
+        */
+       if (posted && (got || need_posted)) {
+               advertise = min_t(unsigned int, posted, RDS_MAX_ADV_CREDIT);
+               newval -= IB_SET_POST_CREDITS(advertise);
+       }
+
+       /* Finally bill everything */
+       if (atomic_cmpxchg(&ic->i_credits, oldval, newval) != oldval)
+               goto try_again;
+
+       *adv_credits = advertise;
+       return got;
+}
+
+void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits)
+{
+       struct rds_ib_connection *ic = conn->c_transport_data;
+
+       if (credits == 0)
+               return;
+
+       rdsdebug("rds_ib_send_add_credits(%u): current=%u%s\n",
+                       credits,
+                       IB_GET_SEND_CREDITS(atomic_read(&ic->i_credits)),
+                       test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ? ", ll_send_full" : "");
+
+       atomic_add(IB_SET_SEND_CREDITS(credits), &ic->i_credits);
+       if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags))
+               queue_delayed_work(rds_wq, &conn->c_send_w, 0);
+
+       WARN_ON(IB_GET_SEND_CREDITS(credits) >= 16384);
+
+       rds_ib_stats_inc(s_ib_rx_credit_updates);
+}
+
+void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted)
+{
+       struct rds_ib_connection *ic = conn->c_transport_data;
+
+       if (posted == 0)
+               return;
+
+       atomic_add(IB_SET_POST_CREDITS(posted), &ic->i_credits);
+
+       /* Decide whether to send an update to the peer now.
+        * If we would send a credit update for every single buffer we
+        * post, we would end up with an ACK storm (ACK arrives,
+        * consumes buffer, we refill the ring, send ACK to remote
+        * advertising the newly posted buffer... ad inf)
+        *
+        * Performance pretty much depends on how often we send
+        * credit updates - too frequent updates mean lots of ACKs.
+        * Too infrequent updates, and the peer will run out of
+        * credits and has to throttle.
+        * For the time being, 16 seems to be a good compromise.
+        */
+       if (IB_GET_POST_CREDITS(atomic_read(&ic->i_credits)) >= 16)
+               set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
+}
+
+static inline void
+rds_ib_xmit_populate_wr(struct rds_ib_connection *ic,
+               struct rds_ib_send_work *send, unsigned int pos,
+               unsigned long buffer, unsigned int length,
+               int send_flags)
+{
+       struct ib_sge *sge;
+
+       WARN_ON(pos != send - ic->i_sends);
+
+       send->s_wr.send_flags = send_flags;
+       send->s_wr.opcode = IB_WR_SEND;
+       send->s_wr.num_sge = 2;
+       send->s_wr.next = NULL;
+       send->s_queued = jiffies;
+       send->s_op = NULL;
+
+       if (length != 0) {
+               sge = rds_ib_data_sge(ic, send->s_sge);
+               sge->addr = buffer;
+               sge->length = length;
+               sge->lkey = ic->i_mr->lkey;
+
+               sge = rds_ib_header_sge(ic, send->s_sge);
+       } else {
+               /* We're sending a packet with no payload. There is only
+                * one SGE */
+               send->s_wr.num_sge = 1;
+               sge = &send->s_sge[0];
+       }
+
+       sge->addr = ic->i_send_hdrs_dma + (pos * sizeof(struct rds_header));
+       sge->length = sizeof(struct rds_header);
+       sge->lkey = ic->i_mr->lkey;
+}
+
+/*
+ * This can be called multiple times for a given message.  The first time
+ * we see a message we map its scatterlist into the IB device so that
+ * we can provide that mapped address to the IB scatter gather entries
+ * in the IB work requests.  We translate the scatterlist into a series
+ * of work requests that fragment the message.  These work requests complete
+ * in order so we pass ownership of the message to the completion handler
+ * once we send the final fragment.
+ *
+ * The RDS core uses the c_send_lock to only enter this function once
+ * per connection.  This makes sure that the tx ring alloc/unalloc pairs
+ * don't get out of sync and confuse the ring.
+ */
+int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
+               unsigned int hdr_off, unsigned int sg, unsigned int off)
+{
+       struct rds_ib_connection *ic = conn->c_transport_data;
+       struct ib_device *dev = ic->i_cm_id->device;
+       struct rds_ib_send_work *send = NULL;
+       struct rds_ib_send_work *first;
+       struct rds_ib_send_work *prev;
+       struct ib_send_wr *failed_wr;
+       struct scatterlist *scat;
+       u32 pos;
+       u32 i;
+       u32 work_alloc;
+       u32 credit_alloc;
+       u32 posted;
+       u32 adv_credits = 0;
+       int send_flags = 0;
+       int sent;
+       int ret;
+       int flow_controlled = 0;
+
+       BUG_ON(off % RDS_FRAG_SIZE);
+       BUG_ON(hdr_off != 0 && hdr_off != sizeof(struct rds_header));
+
+       /* FIXME we may overallocate here */
+       if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0)
+               i = 1;
+       else
+               i = ceil(be32_to_cpu(rm->m_inc.i_hdr.h_len), RDS_FRAG_SIZE);
+
+       work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
+       if (work_alloc == 0) {
+               set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
+               rds_ib_stats_inc(s_ib_tx_ring_full);
+               ret = -ENOMEM;
+               goto out;
+       }
+
+       credit_alloc = work_alloc;
+       if (ic->i_flowctl) {
+               credit_alloc = rds_ib_send_grab_credits(ic, work_alloc, &posted, 0);
+               adv_credits += posted;
+               if (credit_alloc < work_alloc) {
+                       rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - credit_alloc);
+                       work_alloc = credit_alloc;
+                       flow_controlled++;
+               }
+               if (work_alloc == 0) {
+                       rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
+                       rds_ib_stats_inc(s_ib_tx_throttle);
+                       ret = -ENOMEM;
+                       goto out;
+               }
+       }
+
+       /* map the message the first time we see it */
+       if (ic->i_rm == NULL) {
+               /*
+               printk(KERN_NOTICE "rds_ib_xmit prep msg dport=%u flags=0x%x len=%d\n",
+                               be16_to_cpu(rm->m_inc.i_hdr.h_dport),
+                               rm->m_inc.i_hdr.h_flags,
+                               be32_to_cpu(rm->m_inc.i_hdr.h_len));
+                  */
+               if (rm->m_nents) {
+                       rm->m_count = ib_dma_map_sg(dev,
+                                        rm->m_sg, rm->m_nents, DMA_TO_DEVICE);
+                       rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->m_count);
+                       if (rm->m_count == 0) {
+                               rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
+                               rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
+                               ret = -ENOMEM; /* XXX ? */
+                               goto out;
+                       }
+               } else {
+                       rm->m_count = 0;
+               }
+
+               ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
+               ic->i_unsignaled_bytes = rds_ib_sysctl_max_unsig_bytes;
+               rds_message_addref(rm);
+               ic->i_rm = rm;
+
+               /* Finalize the header */
+               if (test_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags))
+                       rm->m_inc.i_hdr.h_flags |= RDS_FLAG_ACK_REQUIRED;
+               if (test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))
+                       rm->m_inc.i_hdr.h_flags |= RDS_FLAG_RETRANSMITTED;
+
+               /* If it has a RDMA op, tell the peer we did it. This is
+                * used by the peer to release use-once RDMA MRs. */
+               if (rm->m_rdma_op) {
+                       struct rds_ext_header_rdma ext_hdr;
+
+                       ext_hdr.h_rdma_rkey = cpu_to_be32(rm->m_rdma_op->r_key);
+                       rds_message_add_extension(&rm->m_inc.i_hdr,
+                                       RDS_EXTHDR_RDMA, &ext_hdr, sizeof(ext_hdr));
+               }
+               if (rm->m_rdma_cookie) {
+                       rds_message_add_rdma_dest_extension(&rm->m_inc.i_hdr,
+                                       rds_rdma_cookie_key(rm->m_rdma_cookie),
+                                       rds_rdma_cookie_offset(rm->m_rdma_cookie));
+               }
+
+               /* Note - rds_ib_piggyb_ack clears the ACK_REQUIRED bit, so
+                * we should not do this unless we have a chance of at least
+                * sticking the header into the send ring. Which is why we
+                * should call rds_ib_ring_alloc first. */
+               rm->m_inc.i_hdr.h_ack = cpu_to_be64(rds_ib_piggyb_ack(ic));
+               rds_message_make_checksum(&rm->m_inc.i_hdr);
+
+               /*
+                * Update adv_credits since we reset the ACK_REQUIRED bit.
+                */
+               rds_ib_send_grab_credits(ic, 0, &posted, 1);
+               adv_credits += posted;
+               BUG_ON(adv_credits > 255);
+       } else if (ic->i_rm != rm)
+               BUG();
+
+       send = &ic->i_sends[pos];
+       first = send;
+       prev = NULL;
+       scat = &rm->m_sg[sg];
+       sent = 0;
+       i = 0;
+
+       /* Sometimes you want to put a fence between an RDMA
+        * READ and the following SEND.
+        * We could either do this all the time
+        * or when requested by the user. Right now, we let
+        * the application choose.
+        */
+       if (rm->m_rdma_op && rm->m_rdma_op->r_fence)
+               send_flags = IB_SEND_FENCE;
+
+       /*
+        * We could be copying the header into the unused tail of the page.
+        * That would need to be changed in the future when those pages might
+        * be mapped userspace pages or page cache pages.  So instead we always
+        * use a second sge and our long-lived ring of mapped headers.  We send
+        * the header after the data so that the data payload can be aligned on
+        * the receiver.
+        */
+
+       /* handle a 0-len message */
+       if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0) {
+               rds_ib_xmit_populate_wr(ic, send, pos, 0, 0, send_flags);
+               goto add_header;
+       }
+
+       /* if there's data reference it with a chain of work reqs */
+       for (; i < work_alloc && scat != &rm->m_sg[rm->m_count]; i++) {
+               unsigned int len;
+
+               send = &ic->i_sends[pos];
+
+               len = min(RDS_FRAG_SIZE, ib_sg_dma_len(dev, scat) - off);
+               rds_ib_xmit_populate_wr(ic, send, pos,
+                               ib_sg_dma_address(dev, scat) + off, len,
+                               send_flags);
+
+               /*
+                * We want to delay signaling completions just enough to get
+                * the batching benefits but not so much that we create dead time
+                * on the wire.
+                */
+               if (ic->i_unsignaled_wrs-- == 0) {
+                       ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
+                       send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
+               }
+
+               ic->i_unsignaled_bytes -= len;
+               if (ic->i_unsignaled_bytes <= 0) {
+                       ic->i_unsignaled_bytes = rds_ib_sysctl_max_unsig_bytes;
+                       send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
+               }
+
+               /*
+                * Always signal the last one if we're stopping due to flow control.
+                */
+               if (flow_controlled && i == (work_alloc-1))
+                       send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
+
+               rdsdebug("send %p wr %p num_sge %u next %p\n", send,
+                        &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
+
+               sent += len;
+               off += len;
+               if (off == ib_sg_dma_len(dev, scat)) {
+                       scat++;
+                       off = 0;
+               }
+
+add_header:
+               /* Tack on the header after the data. The header SGE should already
+                * have been set up to point to the right header buffer. */
+               memcpy(&ic->i_send_hdrs[pos], &rm->m_inc.i_hdr, sizeof(struct rds_header));
+
+               if (0) {
+                       struct rds_header *hdr = &ic->i_send_hdrs[pos];
+
+                       printk(KERN_NOTICE "send WR dport=%u flags=0x%x len=%d\n",
+                               be16_to_cpu(hdr->h_dport),
+                               hdr->h_flags,
+                               be32_to_cpu(hdr->h_len));
+               }
+               if (adv_credits) {
+                       struct rds_header *hdr = &ic->i_send_hdrs[pos];
+
+                       /* add credit and redo the header checksum */
+                       hdr->h_credit = adv_credits;
+                       rds_message_make_checksum(hdr);
+                       adv_credits = 0;
+                       rds_ib_stats_inc(s_ib_tx_credit_updates);
+               }
+
+               if (prev)
+                       prev->s_wr.next = &send->s_wr;
+               prev = send;
+
+               pos = (pos + 1) % ic->i_send_ring.w_nr;
+       }
+
+       /* Account the RDS header in the number of bytes we sent, but just once.
+        * The caller has no concept of fragmentation. */
+       if (hdr_off == 0)
+               sent += sizeof(struct rds_header);
+
+       /* if we finished the message then send completion owns it */
+       if (scat == &rm->m_sg[rm->m_count]) {
+               prev->s_rm = ic->i_rm;
+               prev->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
+               ic->i_rm = NULL;
+       }
+
+       if (i < work_alloc) {
+               rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
+               work_alloc = i;
+       }
+       if (ic->i_flowctl && i < credit_alloc)
+               rds_ib_send_add_credits(conn, credit_alloc - i);
+
+       /* XXX need to worry about failed_wr and partial sends. */
+       failed_wr = &first->s_wr;
+       ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
+       rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
+                first, &first->s_wr, ret, failed_wr);
+       BUG_ON(failed_wr != &first->s_wr);
+       if (ret) {
+               printk(KERN_WARNING "RDS/IB: ib_post_send to %pI4 "
+                      "returned %d\n", &conn->c_faddr, ret);
+               rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
+               if (prev->s_rm) {
+                       ic->i_rm = prev->s_rm;
+                       prev->s_rm = NULL;
+               }
+               /* Finesse this later */
+               BUG();
+               goto out;
+       }
+
+       ret = sent;
+out:
+       BUG_ON(adv_credits);
+       return ret;
+}
+
+int rds_ib_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
+{
+       struct rds_ib_connection *ic = conn->c_transport_data;
+       struct rds_ib_send_work *send = NULL;
+       struct rds_ib_send_work *first;
+       struct rds_ib_send_work *prev;
+       struct ib_send_wr *failed_wr;
+       struct rds_ib_device *rds_ibdev;
+       struct scatterlist *scat;
+       unsigned long len;
+       u64 remote_addr = op->r_remote_addr;
+       u32 pos;
+       u32 work_alloc;
+       u32 i;
+       u32 j;
+       int sent;
+       int ret;
+       int num_sge;
+
+       rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
+
+       /* map the message the first time we see it */
+       if (!op->r_mapped) {
+               op->r_count = ib_dma_map_sg(ic->i_cm_id->device,
+                                       op->r_sg, op->r_nents, (op->r_write) ?
+                                       DMA_TO_DEVICE : DMA_FROM_DEVICE);
+               rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->r_count);
+               if (op->r_count == 0) {
+                       rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
+                       ret = -ENOMEM; /* XXX ? */
+                       goto out;
+               }
+
+               op->r_mapped = 1;
+       }
+
+       /*
+        * Instead of knowing how to return a partial rdma read/write we insist that there
+        * be enough work requests to send the entire message.
+        */
+       i = ceil(op->r_count, rds_ibdev->max_sge);
+
+       work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
+       if (work_alloc != i) {
+               rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
+               rds_ib_stats_inc(s_ib_tx_ring_full);
+               ret = -ENOMEM;
+               goto out;
+       }
+
+       send = &ic->i_sends[pos];
+       first = send;
+       prev = NULL;
+       scat = &op->r_sg[0];
+       sent = 0;
+       num_sge = op->r_count;
+
+       for (i = 0; i < work_alloc && scat != &op->r_sg[op->r_count]; i++) {
+               send->s_wr.send_flags = 0;
+               send->s_queued = jiffies;
+               /*
+                * We want to delay signaling completions just enough to get
+                * the batching benefits but not so much that we create dead time on the wire.
+                */
+               if (ic->i_unsignaled_wrs-- == 0) {
+                       ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
+                       send->s_wr.send_flags = IB_SEND_SIGNALED;
+               }
+
+               send->s_wr.opcode = op->r_write ? IB_WR_RDMA_WRITE : IB_WR_RDMA_READ;
+               send->s_wr.wr.rdma.remote_addr = remote_addr;
+               send->s_wr.wr.rdma.rkey = op->r_key;
+               send->s_op = op;
+
+               if (num_sge > rds_ibdev->max_sge) {
+                       send->s_wr.num_sge = rds_ibdev->max_sge;
+                       num_sge -= rds_ibdev->max_sge;
+               } else {
+                       send->s_wr.num_sge = num_sge;
+               }
+
+               send->s_wr.next = NULL;
+
+               if (prev)
+                       prev->s_wr.next = &send->s_wr;
+
+               for (j = 0; j < send->s_wr.num_sge && scat != &op->r_sg[op->r_count]; j++) {
+                       len = ib_sg_dma_len(ic->i_cm_id->device, scat);
+                       send->s_sge[j].addr =
+                                ib_sg_dma_address(ic->i_cm_id->device, scat);
+                       send->s_sge[j].length = len;
+                       send->s_sge[j].lkey = ic->i_mr->lkey;
+
+                       sent += len;
+                       rdsdebug("ic %p sent %d remote_addr %llu\n", ic, sent, remote_addr);
+
+                       remote_addr += len;
+                       scat++;
+               }
+
+               rdsdebug("send %p wr %p num_sge %u next %p\n", send,
+                       &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
+
+               prev = send;
+               if (++send == &ic->i_sends[ic->i_send_ring.w_nr])
+                       send = ic->i_sends;
+       }
+
+       /* if we finished the message then send completion owns it */
+       if (scat == &op->r_sg[op->r_count])
+               prev->s_wr.send_flags = IB_SEND_SIGNALED;
+
+       if (i < work_alloc) {
+               rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
+               work_alloc = i;
+       }
+
+       failed_wr = &first->s_wr;
+       ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
+       rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
+                first, &first->s_wr, ret, failed_wr);
+       BUG_ON(failed_wr != &first->s_wr);
+       if (ret) {
+               printk(KERN_WARNING "RDS/IB: rdma ib_post_send to %pI4 "
+                      "returned %d\n", &conn->c_faddr, ret);
+               rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
+               goto out;
+       }
+
+       if (unlikely(failed_wr != &first->s_wr)) {
+               printk(KERN_WARNING "RDS/IB: ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
+               BUG_ON(failed_wr != &first->s_wr);
+       }
+
+
+out:
+       return ret;
+}
+
+void rds_ib_xmit_complete(struct rds_connection *conn)
+{
+       struct rds_ib_connection *ic = conn->c_transport_data;
+
+       /* We may have a pending ACK or window update we were unable
+        * to send previously (due to flow control). Try again. */
+       rds_ib_attempt_ack(ic);
+}