RDS/IB: Receive datagrams via IB
authorAndy Grover <andy.grover@oracle.com>
Tue, 24 Feb 2009 15:30:34 +0000 (15:30 +0000)
committerDavid S. Miller <davem@davemloft.net>
Fri, 27 Feb 2009 07:39:32 +0000 (23:39 -0800)
Header parsing, ring refill. It puts the incoming data into an
rds_incoming struct, which is passed up to rds-core.

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

diff --git a/net/rds/ib_recv.c b/net/rds/ib_recv.c
new file mode 100644 (file)
index 0000000..5061b55
--- /dev/null
@@ -0,0 +1,869 @@
+/*
+ * 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/pci.h>
+#include <linux/dma-mapping.h>
+#include <rdma/rdma_cm.h>
+
+#include "rds.h"
+#include "ib.h"
+
+static struct kmem_cache *rds_ib_incoming_slab;
+static struct kmem_cache *rds_ib_frag_slab;
+static atomic_t        rds_ib_allocation = ATOMIC_INIT(0);
+
+static void rds_ib_frag_drop_page(struct rds_page_frag *frag)
+{
+       rdsdebug("frag %p page %p\n", frag, frag->f_page);
+       __free_page(frag->f_page);
+       frag->f_page = NULL;
+}
+
+static void rds_ib_frag_free(struct rds_page_frag *frag)
+{
+       rdsdebug("frag %p page %p\n", frag, frag->f_page);
+       BUG_ON(frag->f_page != NULL);
+       kmem_cache_free(rds_ib_frag_slab, frag);
+}
+
+/*
+ * We map a page at a time.  Its fragments are posted in order.  This
+ * is called in fragment order as the fragments get send completion events.
+ * Only the last frag in the page performs the unmapping.
+ *
+ * It's OK for ring cleanup to call this in whatever order it likes because
+ * DMA is not in flight and so we can unmap while other ring entries still
+ * hold page references in their frags.
+ */
+static void rds_ib_recv_unmap_page(struct rds_ib_connection *ic,
+                                  struct rds_ib_recv_work *recv)
+{
+       struct rds_page_frag *frag = recv->r_frag;
+
+       rdsdebug("recv %p frag %p page %p\n", recv, frag, frag->f_page);
+       if (frag->f_mapped)
+               ib_dma_unmap_page(ic->i_cm_id->device,
+                              frag->f_mapped,
+                              RDS_FRAG_SIZE, DMA_FROM_DEVICE);
+       frag->f_mapped = 0;
+}
+
+void rds_ib_recv_init_ring(struct rds_ib_connection *ic)
+{
+       struct rds_ib_recv_work *recv;
+       u32 i;
+
+       for (i = 0, recv = ic->i_recvs; i < ic->i_recv_ring.w_nr; i++, recv++) {
+               struct ib_sge *sge;
+
+               recv->r_ibinc = NULL;
+               recv->r_frag = NULL;
+
+               recv->r_wr.next = NULL;
+               recv->r_wr.wr_id = i;
+               recv->r_wr.sg_list = recv->r_sge;
+               recv->r_wr.num_sge = RDS_IB_RECV_SGE;
+
+               sge = rds_ib_data_sge(ic, recv->r_sge);
+               sge->addr = 0;
+               sge->length = RDS_FRAG_SIZE;
+               sge->lkey = ic->i_mr->lkey;
+
+               sge = rds_ib_header_sge(ic, recv->r_sge);
+               sge->addr = ic->i_recv_hdrs_dma + (i * sizeof(struct rds_header));
+               sge->length = sizeof(struct rds_header);
+               sge->lkey = ic->i_mr->lkey;
+       }
+}
+
+static void rds_ib_recv_clear_one(struct rds_ib_connection *ic,
+                                 struct rds_ib_recv_work *recv)
+{
+       if (recv->r_ibinc) {
+               rds_inc_put(&recv->r_ibinc->ii_inc);
+               recv->r_ibinc = NULL;
+       }
+       if (recv->r_frag) {
+               rds_ib_recv_unmap_page(ic, recv);
+               if (recv->r_frag->f_page)
+                       rds_ib_frag_drop_page(recv->r_frag);
+               rds_ib_frag_free(recv->r_frag);
+               recv->r_frag = NULL;
+       }
+}
+
+void rds_ib_recv_clear_ring(struct rds_ib_connection *ic)
+{
+       u32 i;
+
+       for (i = 0; i < ic->i_recv_ring.w_nr; i++)
+               rds_ib_recv_clear_one(ic, &ic->i_recvs[i]);
+
+       if (ic->i_frag.f_page)
+               rds_ib_frag_drop_page(&ic->i_frag);
+}
+
+static int rds_ib_recv_refill_one(struct rds_connection *conn,
+                                 struct rds_ib_recv_work *recv,
+                                 gfp_t kptr_gfp, gfp_t page_gfp)
+{
+       struct rds_ib_connection *ic = conn->c_transport_data;
+       dma_addr_t dma_addr;
+       struct ib_sge *sge;
+       int ret = -ENOMEM;
+
+       if (recv->r_ibinc == NULL) {
+               if (atomic_read(&rds_ib_allocation) >= rds_ib_sysctl_max_recv_allocation) {
+                       rds_ib_stats_inc(s_ib_rx_alloc_limit);
+                       goto out;
+               }
+               recv->r_ibinc = kmem_cache_alloc(rds_ib_incoming_slab,
+                                                kptr_gfp);
+               if (recv->r_ibinc == NULL)
+                       goto out;
+               atomic_inc(&rds_ib_allocation);
+               INIT_LIST_HEAD(&recv->r_ibinc->ii_frags);
+               rds_inc_init(&recv->r_ibinc->ii_inc, conn, conn->c_faddr);
+       }
+
+       if (recv->r_frag == NULL) {
+               recv->r_frag = kmem_cache_alloc(rds_ib_frag_slab, kptr_gfp);
+               if (recv->r_frag == NULL)
+                       goto out;
+               INIT_LIST_HEAD(&recv->r_frag->f_item);
+               recv->r_frag->f_page = NULL;
+       }
+
+       if (ic->i_frag.f_page == NULL) {
+               ic->i_frag.f_page = alloc_page(page_gfp);
+               if (ic->i_frag.f_page == NULL)
+                       goto out;
+               ic->i_frag.f_offset = 0;
+       }
+
+       dma_addr = ib_dma_map_page(ic->i_cm_id->device,
+                                 ic->i_frag.f_page,
+                                 ic->i_frag.f_offset,
+                                 RDS_FRAG_SIZE,
+                                 DMA_FROM_DEVICE);
+       if (ib_dma_mapping_error(ic->i_cm_id->device, dma_addr))
+               goto out;
+
+       /*
+        * Once we get the RDS_PAGE_LAST_OFF frag then rds_ib_frag_unmap()
+        * must be called on this recv.  This happens as completions hit
+        * in order or on connection shutdown.
+        */
+       recv->r_frag->f_page = ic->i_frag.f_page;
+       recv->r_frag->f_offset = ic->i_frag.f_offset;
+       recv->r_frag->f_mapped = dma_addr;
+
+       sge = rds_ib_data_sge(ic, recv->r_sge);
+       sge->addr = dma_addr;
+       sge->length = RDS_FRAG_SIZE;
+
+       sge = rds_ib_header_sge(ic, recv->r_sge);
+       sge->addr = ic->i_recv_hdrs_dma + (recv - ic->i_recvs) * sizeof(struct rds_header);
+       sge->length = sizeof(struct rds_header);
+
+       get_page(recv->r_frag->f_page);
+
+       if (ic->i_frag.f_offset < RDS_PAGE_LAST_OFF) {
+               ic->i_frag.f_offset += RDS_FRAG_SIZE;
+       } else {
+               put_page(ic->i_frag.f_page);
+               ic->i_frag.f_page = NULL;
+               ic->i_frag.f_offset = 0;
+       }
+
+       ret = 0;
+out:
+       return ret;
+}
+
+/*
+ * This tries to allocate and post unused work requests after making sure that
+ * they have all the allocations they need to queue received fragments into
+ * sockets.  The i_recv_mutex is held here so that ring_alloc and _unalloc
+ * pairs don't go unmatched.
+ *
+ * -1 is returned if posting fails due to temporary resource exhaustion.
+ */
+int rds_ib_recv_refill(struct rds_connection *conn, gfp_t kptr_gfp,
+                      gfp_t page_gfp, int prefill)
+{
+       struct rds_ib_connection *ic = conn->c_transport_data;
+       struct rds_ib_recv_work *recv;
+       struct ib_recv_wr *failed_wr;
+       unsigned int posted = 0;
+       int ret = 0;
+       u32 pos;
+
+       while ((prefill || rds_conn_up(conn))
+                       && rds_ib_ring_alloc(&ic->i_recv_ring, 1, &pos)) {
+               if (pos >= ic->i_recv_ring.w_nr) {
+                       printk(KERN_NOTICE "Argh - ring alloc returned pos=%u\n",
+                                       pos);
+                       ret = -EINVAL;
+                       break;
+               }
+
+               recv = &ic->i_recvs[pos];
+               ret = rds_ib_recv_refill_one(conn, recv, kptr_gfp, page_gfp);
+               if (ret) {
+                       ret = -1;
+                       break;
+               }
+
+               /* XXX when can this fail? */
+               ret = ib_post_recv(ic->i_cm_id->qp, &recv->r_wr, &failed_wr);
+               rdsdebug("recv %p ibinc %p page %p addr %lu ret %d\n", recv,
+                        recv->r_ibinc, recv->r_frag->f_page,
+                        (long) recv->r_frag->f_mapped, ret);
+               if (ret) {
+                       rds_ib_conn_error(conn, "recv post on "
+                              "%pI4 returned %d, disconnecting and "
+                              "reconnecting\n", &conn->c_faddr,
+                              ret);
+                       ret = -1;
+                       break;
+               }
+
+               posted++;
+       }
+
+       /* We're doing flow control - update the window. */
+       if (ic->i_flowctl && posted)
+               rds_ib_advertise_credits(conn, posted);
+
+       if (ret)
+               rds_ib_ring_unalloc(&ic->i_recv_ring, 1);
+       return ret;
+}
+
+void rds_ib_inc_purge(struct rds_incoming *inc)
+{
+       struct rds_ib_incoming *ibinc;
+       struct rds_page_frag *frag;
+       struct rds_page_frag *pos;
+
+       ibinc = container_of(inc, struct rds_ib_incoming, ii_inc);
+       rdsdebug("purging ibinc %p inc %p\n", ibinc, inc);
+
+       list_for_each_entry_safe(frag, pos, &ibinc->ii_frags, f_item) {
+               list_del_init(&frag->f_item);
+               rds_ib_frag_drop_page(frag);
+               rds_ib_frag_free(frag);
+       }
+}
+
+void rds_ib_inc_free(struct rds_incoming *inc)
+{
+       struct rds_ib_incoming *ibinc;
+
+       ibinc = container_of(inc, struct rds_ib_incoming, ii_inc);
+
+       rds_ib_inc_purge(inc);
+       rdsdebug("freeing ibinc %p inc %p\n", ibinc, inc);
+       BUG_ON(!list_empty(&ibinc->ii_frags));
+       kmem_cache_free(rds_ib_incoming_slab, ibinc);
+       atomic_dec(&rds_ib_allocation);
+       BUG_ON(atomic_read(&rds_ib_allocation) < 0);
+}
+
+int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iovec *first_iov,
+                           size_t size)
+{
+       struct rds_ib_incoming *ibinc;
+       struct rds_page_frag *frag;
+       struct iovec *iov = first_iov;
+       unsigned long to_copy;
+       unsigned long frag_off = 0;
+       unsigned long iov_off = 0;
+       int copied = 0;
+       int ret;
+       u32 len;
+
+       ibinc = container_of(inc, struct rds_ib_incoming, ii_inc);
+       frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item);
+       len = be32_to_cpu(inc->i_hdr.h_len);
+
+       while (copied < size && copied < len) {
+               if (frag_off == RDS_FRAG_SIZE) {
+                       frag = list_entry(frag->f_item.next,
+                                         struct rds_page_frag, f_item);
+                       frag_off = 0;
+               }
+               while (iov_off == iov->iov_len) {
+                       iov_off = 0;
+                       iov++;
+               }
+
+               to_copy = min(iov->iov_len - iov_off, RDS_FRAG_SIZE - frag_off);
+               to_copy = min_t(size_t, to_copy, size - copied);
+               to_copy = min_t(unsigned long, to_copy, len - copied);
+
+               rdsdebug("%lu bytes to user [%p, %zu] + %lu from frag "
+                        "[%p, %lu] + %lu\n",
+                        to_copy, iov->iov_base, iov->iov_len, iov_off,
+                        frag->f_page, frag->f_offset, frag_off);
+
+               /* XXX needs + offset for multiple recvs per page */
+               ret = rds_page_copy_to_user(frag->f_page,
+                                           frag->f_offset + frag_off,
+                                           iov->iov_base + iov_off,
+                                           to_copy);
+               if (ret) {
+                       copied = ret;
+                       break;
+               }
+
+               iov_off += to_copy;
+               frag_off += to_copy;
+               copied += to_copy;
+       }
+
+       return copied;
+}
+
+/* ic starts out kzalloc()ed */
+void rds_ib_recv_init_ack(struct rds_ib_connection *ic)
+{
+       struct ib_send_wr *wr = &ic->i_ack_wr;
+       struct ib_sge *sge = &ic->i_ack_sge;
+
+       sge->addr = ic->i_ack_dma;
+       sge->length = sizeof(struct rds_header);
+       sge->lkey = ic->i_mr->lkey;
+
+       wr->sg_list = sge;
+       wr->num_sge = 1;
+       wr->opcode = IB_WR_SEND;
+       wr->wr_id = RDS_IB_ACK_WR_ID;
+       wr->send_flags = IB_SEND_SIGNALED | IB_SEND_SOLICITED;
+}
+
+/*
+ * You'd think that with reliable IB connections you wouldn't need to ack
+ * messages that have been received.  The problem is that IB hardware generates
+ * an ack message before it has DMAed the message into memory.  This creates a
+ * potential message loss if the HCA is disabled for any reason between when it
+ * sends the ack and before the message is DMAed and processed.  This is only a
+ * potential issue if another HCA is available for fail-over.
+ *
+ * When the remote host receives our ack they'll free the sent message from
+ * their send queue.  To decrease the latency of this we always send an ack
+ * immediately after we've received messages.
+ *
+ * For simplicity, we only have one ack in flight at a time.  This puts
+ * pressure on senders to have deep enough send queues to absorb the latency of
+ * a single ack frame being in flight.  This might not be good enough.
+ *
+ * This is implemented by have a long-lived send_wr and sge which point to a
+ * statically allocated ack frame.  This ack wr does not fall under the ring
+ * accounting that the tx and rx wrs do.  The QP attribute specifically makes
+ * room for it beyond the ring size.  Send completion notices its special
+ * wr_id and avoids working with the ring in that case.
+ */
+static void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq,
+                               int ack_required)
+{
+       rds_ib_set_64bit(&ic->i_ack_next, seq);
+       if (ack_required) {
+               smp_mb__before_clear_bit();
+               set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
+       }
+}
+
+static u64 rds_ib_get_ack(struct rds_ib_connection *ic)
+{
+       clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
+       smp_mb__after_clear_bit();
+
+       return ic->i_ack_next;
+}
+
+static void rds_ib_send_ack(struct rds_ib_connection *ic, unsigned int adv_credits)
+{
+       struct rds_header *hdr = ic->i_ack;
+       struct ib_send_wr *failed_wr;
+       u64 seq;
+       int ret;
+
+       seq = rds_ib_get_ack(ic);
+
+       rdsdebug("send_ack: ic %p ack %llu\n", ic, (unsigned long long) seq);
+       rds_message_populate_header(hdr, 0, 0, 0);
+       hdr->h_ack = cpu_to_be64(seq);
+       hdr->h_credit = adv_credits;
+       rds_message_make_checksum(hdr);
+       ic->i_ack_queued = jiffies;
+
+       ret = ib_post_send(ic->i_cm_id->qp, &ic->i_ack_wr, &failed_wr);
+       if (unlikely(ret)) {
+               /* Failed to send. Release the WR, and
+                * force another ACK.
+                */
+               clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
+               set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
+
+               rds_ib_stats_inc(s_ib_ack_send_failure);
+               /* Need to finesse this later. */
+               BUG();
+       } else
+               rds_ib_stats_inc(s_ib_ack_sent);
+}
+
+/*
+ * There are 3 ways of getting acknowledgements to the peer:
+ *  1. We call rds_ib_attempt_ack from the recv completion handler
+ *     to send an ACK-only frame.
+ *     However, there can be only one such frame in the send queue
+ *     at any time, so we may have to postpone it.
+ *  2. When another (data) packet is transmitted while there's
+ *     an ACK in the queue, we piggyback the ACK sequence number
+ *     on the data packet.
+ *  3. If the ACK WR is done sending, we get called from the
+ *     send queue completion handler, and check whether there's
+ *     another ACK pending (postponed because the WR was on the
+ *     queue). If so, we transmit it.
+ *
+ * We maintain 2 variables:
+ *  -  i_ack_flags, which keeps track of whether the ACK WR
+ *     is currently in the send queue or not (IB_ACK_IN_FLIGHT)
+ *  -  i_ack_next, which is the last sequence number we received
+ *
+ * Potentially, send queue and receive queue handlers can run concurrently.
+ *
+ * Reconnecting complicates this picture just slightly. When we
+ * reconnect, we may be seeing duplicate packets. The peer
+ * is retransmitting them, because it hasn't seen an ACK for
+ * them. It is important that we ACK these.
+ *
+ * ACK mitigation adds a header flag "ACK_REQUIRED"; any packet with
+ * this flag set *MUST* be acknowledged immediately.
+ */
+
+/*
+ * When we get here, we're called from the recv queue handler.
+ * Check whether we ought to transmit an ACK.
+ */
+void rds_ib_attempt_ack(struct rds_ib_connection *ic)
+{
+       unsigned int adv_credits;
+
+       if (!test_bit(IB_ACK_REQUESTED, &ic->i_ack_flags))
+               return;
+
+       if (test_and_set_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags)) {
+               rds_ib_stats_inc(s_ib_ack_send_delayed);
+               return;
+       }
+
+       /* Can we get a send credit? */
+       if (!rds_ib_send_grab_credits(ic, 1, &adv_credits, 0)) {
+               rds_ib_stats_inc(s_ib_tx_throttle);
+               clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
+               return;
+       }
+
+       clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
+       rds_ib_send_ack(ic, adv_credits);
+}
+
+/*
+ * We get here from the send completion handler, when the
+ * adapter tells us the ACK frame was sent.
+ */
+void rds_ib_ack_send_complete(struct rds_ib_connection *ic)
+{
+       clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
+       rds_ib_attempt_ack(ic);
+}
+
+/*
+ * This is called by the regular xmit code when it wants to piggyback
+ * an ACK on an outgoing frame.
+ */
+u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic)
+{
+       if (test_and_clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags))
+               rds_ib_stats_inc(s_ib_ack_send_piggybacked);
+       return rds_ib_get_ack(ic);
+}
+
+/*
+ * It's kind of lame that we're copying from the posted receive pages into
+ * long-lived bitmaps.  We could have posted the bitmaps and rdma written into
+ * them.  But receiving new congestion bitmaps should be a *rare* event, so
+ * hopefully we won't need to invest that complexity in making it more
+ * efficient.  By copying we can share a simpler core with TCP which has to
+ * copy.
+ */
+static void rds_ib_cong_recv(struct rds_connection *conn,
+                             struct rds_ib_incoming *ibinc)
+{
+       struct rds_cong_map *map;
+       unsigned int map_off;
+       unsigned int map_page;
+       struct rds_page_frag *frag;
+       unsigned long frag_off;
+       unsigned long to_copy;
+       unsigned long copied;
+       uint64_t uncongested = 0;
+       void *addr;
+
+       /* catch completely corrupt packets */
+       if (be32_to_cpu(ibinc->ii_inc.i_hdr.h_len) != RDS_CONG_MAP_BYTES)
+               return;
+
+       map = conn->c_fcong;
+       map_page = 0;
+       map_off = 0;
+
+       frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item);
+       frag_off = 0;
+
+       copied = 0;
+
+       while (copied < RDS_CONG_MAP_BYTES) {
+               uint64_t *src, *dst;
+               unsigned int k;
+
+               to_copy = min(RDS_FRAG_SIZE - frag_off, PAGE_SIZE - map_off);
+               BUG_ON(to_copy & 7); /* Must be 64bit aligned. */
+
+               addr = kmap_atomic(frag->f_page, KM_SOFTIRQ0);
+
+               src = addr + frag_off;
+               dst = (void *)map->m_page_addrs[map_page] + map_off;
+               for (k = 0; k < to_copy; k += 8) {
+                       /* Record ports that became uncongested, ie
+                        * bits that changed from 0 to 1. */
+                       uncongested |= ~(*src) & *dst;
+                       *dst++ = *src++;
+               }
+               kunmap_atomic(addr, KM_SOFTIRQ0);
+
+               copied += to_copy;
+
+               map_off += to_copy;
+               if (map_off == PAGE_SIZE) {
+                       map_off = 0;
+                       map_page++;
+               }
+
+               frag_off += to_copy;
+               if (frag_off == RDS_FRAG_SIZE) {
+                       frag = list_entry(frag->f_item.next,
+                                         struct rds_page_frag, f_item);
+                       frag_off = 0;
+               }
+       }
+
+       /* the congestion map is in little endian order */
+       uncongested = le64_to_cpu(uncongested);
+
+       rds_cong_map_updated(map, uncongested);
+}
+
+/*
+ * Rings are posted with all the allocations they'll need to queue the
+ * incoming message to the receiving socket so this can't fail.
+ * All fragments start with a header, so we can make sure we're not receiving
+ * garbage, and we can tell a small 8 byte fragment from an ACK frame.
+ */
+struct rds_ib_ack_state {
+       u64             ack_next;
+       u64             ack_recv;
+       unsigned int    ack_required:1;
+       unsigned int    ack_next_valid:1;
+       unsigned int    ack_recv_valid:1;
+};
+
+static void rds_ib_process_recv(struct rds_connection *conn,
+                               struct rds_ib_recv_work *recv, u32 byte_len,
+                               struct rds_ib_ack_state *state)
+{
+       struct rds_ib_connection *ic = conn->c_transport_data;
+       struct rds_ib_incoming *ibinc = ic->i_ibinc;
+       struct rds_header *ihdr, *hdr;
+
+       /* XXX shut down the connection if port 0,0 are seen? */
+
+       rdsdebug("ic %p ibinc %p recv %p byte len %u\n", ic, ibinc, recv,
+                byte_len);
+
+       if (byte_len < sizeof(struct rds_header)) {
+               rds_ib_conn_error(conn, "incoming message "
+                      "from %pI4 didn't inclue a "
+                      "header, disconnecting and "
+                      "reconnecting\n",
+                      &conn->c_faddr);
+               return;
+       }
+       byte_len -= sizeof(struct rds_header);
+
+       ihdr = &ic->i_recv_hdrs[recv - ic->i_recvs];
+
+       /* Validate the checksum. */
+       if (!rds_message_verify_checksum(ihdr)) {
+               rds_ib_conn_error(conn, "incoming message "
+                      "from %pI4 has corrupted header - "
+                      "forcing a reconnect\n",
+                      &conn->c_faddr);
+               rds_stats_inc(s_recv_drop_bad_checksum);
+               return;
+       }
+
+       /* Process the ACK sequence which comes with every packet */
+       state->ack_recv = be64_to_cpu(ihdr->h_ack);
+       state->ack_recv_valid = 1;
+
+       /* Process the credits update if there was one */
+       if (ihdr->h_credit)
+               rds_ib_send_add_credits(conn, ihdr->h_credit);
+
+       if (ihdr->h_sport == 0 && ihdr->h_dport == 0 && byte_len == 0) {
+               /* This is an ACK-only packet. The fact that it gets
+                * special treatment here is that historically, ACKs
+                * were rather special beasts.
+                */
+               rds_ib_stats_inc(s_ib_ack_received);
+
+               /*
+                * Usually the frags make their way on to incs and are then freed as
+                * the inc is freed.  We don't go that route, so we have to drop the
+                * page ref ourselves.  We can't just leave the page on the recv
+                * because that confuses the dma mapping of pages and each recv's use
+                * of a partial page.  We can leave the frag, though, it will be
+                * reused.
+                *
+                * FIXME: Fold this into the code path below.
+                */
+               rds_ib_frag_drop_page(recv->r_frag);
+               return;
+       }
+
+       /*
+        * If we don't already have an inc on the connection then this
+        * fragment has a header and starts a message.. copy its header
+        * into the inc and save the inc so we can hang upcoming fragments
+        * off its list.
+        */
+       if (ibinc == NULL) {
+               ibinc = recv->r_ibinc;
+               recv->r_ibinc = NULL;
+               ic->i_ibinc = ibinc;
+
+               hdr = &ibinc->ii_inc.i_hdr;
+               memcpy(hdr, ihdr, sizeof(*hdr));
+               ic->i_recv_data_rem = be32_to_cpu(hdr->h_len);
+
+               rdsdebug("ic %p ibinc %p rem %u flag 0x%x\n", ic, ibinc,
+                        ic->i_recv_data_rem, hdr->h_flags);
+       } else {
+               hdr = &ibinc->ii_inc.i_hdr;
+               /* We can't just use memcmp here; fragments of a
+                * single message may carry different ACKs */
+               if (hdr->h_sequence != ihdr->h_sequence
+                || hdr->h_len != ihdr->h_len
+                || hdr->h_sport != ihdr->h_sport
+                || hdr->h_dport != ihdr->h_dport) {
+                       rds_ib_conn_error(conn,
+                               "fragment header mismatch; forcing reconnect\n");
+                       return;
+               }
+       }
+
+       list_add_tail(&recv->r_frag->f_item, &ibinc->ii_frags);
+       recv->r_frag = NULL;
+
+       if (ic->i_recv_data_rem > RDS_FRAG_SIZE)
+               ic->i_recv_data_rem -= RDS_FRAG_SIZE;
+       else {
+               ic->i_recv_data_rem = 0;
+               ic->i_ibinc = NULL;
+
+               if (ibinc->ii_inc.i_hdr.h_flags == RDS_FLAG_CONG_BITMAP)
+                       rds_ib_cong_recv(conn, ibinc);
+               else {
+                       rds_recv_incoming(conn, conn->c_faddr, conn->c_laddr,
+                                         &ibinc->ii_inc, GFP_ATOMIC,
+                                         KM_SOFTIRQ0);
+                       state->ack_next = be64_to_cpu(hdr->h_sequence);
+                       state->ack_next_valid = 1;
+               }
+
+               /* Evaluate the ACK_REQUIRED flag *after* we received
+                * the complete frame, and after bumping the next_rx
+                * sequence. */
+               if (hdr->h_flags & RDS_FLAG_ACK_REQUIRED) {
+                       rds_stats_inc(s_recv_ack_required);
+                       state->ack_required = 1;
+               }
+
+               rds_inc_put(&ibinc->ii_inc);
+       }
+}
+
+/*
+ * Plucking the oldest entry from the ring can be done concurrently with
+ * the thread refilling the ring.  Each ring operation is protected by
+ * spinlocks and the transient state of refilling doesn't change the
+ * recording of which entry is oldest.
+ *
+ * This relies on IB only calling one cq comp_handler for each cq so that
+ * there will only be one caller of rds_recv_incoming() per RDS connection.
+ */
+void rds_ib_recv_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_ack_state state = { 0, };
+       struct rds_ib_recv_work *recv;
+
+       rdsdebug("conn %p cq %p\n", conn, cq);
+
+       rds_ib_stats_inc(s_ib_rx_cq_call);
+
+       ib_req_notify_cq(cq, IB_CQ_SOLICITED);
+
+       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_rx_cq_event);
+
+               recv = &ic->i_recvs[rds_ib_ring_oldest(&ic->i_recv_ring)];
+
+               rds_ib_recv_unmap_page(ic, recv);
+
+               /*
+                * Also process recvs in connecting state because it is possible
+                * to get a recv completion _before_ the rdmacm ESTABLISHED
+                * event is processed.
+                */
+               if (rds_conn_up(conn) || rds_conn_connecting(conn)) {
+                       /* We expect errors as the qp is drained during shutdown */
+                       if (wc.status == IB_WC_SUCCESS) {
+                               rds_ib_process_recv(conn, recv, wc.byte_len, &state);
+                       } else {
+                               rds_ib_conn_error(conn, "recv completion on "
+                                      "%pI4 had status %u, disconnecting and "
+                                      "reconnecting\n", &conn->c_faddr,
+                                      wc.status);
+                       }
+               }
+
+               rds_ib_ring_free(&ic->i_recv_ring, 1);
+       }
+
+       if (state.ack_next_valid)
+               rds_ib_set_ack(ic, state.ack_next, state.ack_required);
+       if (state.ack_recv_valid && state.ack_recv > ic->i_ack_recv) {
+               rds_send_drop_acked(conn, state.ack_recv, NULL);
+               ic->i_ack_recv = state.ack_recv;
+       }
+       if (rds_conn_up(conn))
+               rds_ib_attempt_ack(ic);
+
+       /* If we ever end up with a really empty receive ring, we're
+        * in deep trouble, as the sender will definitely see RNR
+        * timeouts. */
+       if (rds_ib_ring_empty(&ic->i_recv_ring))
+               rds_ib_stats_inc(s_ib_rx_ring_empty);
+
+       /*
+        * If the ring is running low, then schedule the thread to refill.
+        */
+       if (rds_ib_ring_low(&ic->i_recv_ring))
+               queue_delayed_work(rds_wq, &conn->c_recv_w, 0);
+}
+
+int rds_ib_recv(struct rds_connection *conn)
+{
+       struct rds_ib_connection *ic = conn->c_transport_data;
+       int ret = 0;
+
+       rdsdebug("conn %p\n", conn);
+
+       /*
+        * If we get a temporary posting failure in this context then
+        * we're really low and we want the caller to back off for a bit.
+        */
+       mutex_lock(&ic->i_recv_mutex);
+       if (rds_ib_recv_refill(conn, GFP_KERNEL, GFP_HIGHUSER, 0))
+               ret = -ENOMEM;
+       else
+               rds_ib_stats_inc(s_ib_rx_refill_from_thread);
+       mutex_unlock(&ic->i_recv_mutex);
+
+       if (rds_conn_up(conn))
+               rds_ib_attempt_ack(ic);
+
+       return ret;
+}
+
+int __init rds_ib_recv_init(void)
+{
+       struct sysinfo si;
+       int ret = -ENOMEM;
+
+       /* Default to 30% of all available RAM for recv memory */
+       si_meminfo(&si);
+       rds_ib_sysctl_max_recv_allocation = si.totalram / 3 * PAGE_SIZE / RDS_FRAG_SIZE;
+
+       rds_ib_incoming_slab = kmem_cache_create("rds_ib_incoming",
+                                       sizeof(struct rds_ib_incoming),
+                                       0, 0, NULL);
+       if (rds_ib_incoming_slab == NULL)
+               goto out;
+
+       rds_ib_frag_slab = kmem_cache_create("rds_ib_frag",
+                                       sizeof(struct rds_page_frag),
+                                       0, 0, NULL);
+       if (rds_ib_frag_slab == NULL)
+               kmem_cache_destroy(rds_ib_incoming_slab);
+       else
+               ret = 0;
+out:
+       return ret;
+}
+
+void rds_ib_recv_exit(void)
+{
+       kmem_cache_destroy(rds_ib_incoming_slab);
+       kmem_cache_destroy(rds_ib_frag_slab);
+}