rds: fix rds_send_xmit() serialization
[linux-2.6.git] / net / rds / ib_send.c
index 5375020..e88cb4a 100644 (file)
 #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)
+/*
+ * Convert IB-specific error message to RDS error message and call core
+ * completion handler.
+ */
+static void rds_ib_send_complete(struct rds_message *rm,
+                                int wc_status,
+                                void (*complete)(struct rds_message *rm, int status))
 {
        int notify_status;
 
@@ -60,69 +64,125 @@ static void rds_ib_send_rdma_complete(struct rds_message *rm,
                notify_status = RDS_RDMA_OTHER_ERROR;
                break;
        }
-       rds_rdma_send_complete(rm, notify_status);
+       complete(rm, notify_status);
+}
+
+static void rds_ib_send_unmap_data(struct rds_ib_connection *ic,
+                                  struct rm_data_op *op,
+                                  int wc_status)
+{
+       if (op->op_nents)
+               ib_dma_unmap_sg(ic->i_cm_id->device,
+                               op->op_sg, op->op_nents,
+                               DMA_TO_DEVICE);
 }
 
 static void rds_ib_send_unmap_rdma(struct rds_ib_connection *ic,
-                                  struct rds_rdma_op *op)
+                                  struct rm_rdma_op *op,
+                                  int wc_status)
 {
-       if (op->r_mapped) {
+       if (op->op_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;
+                               op->op_sg, op->op_nents,
+                               op->op_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
+               op->op_mapped = 0;
        }
+
+       /* 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,
+        * we would need to take an event for the rdma WR. 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_complete(container_of(op, struct rds_message, rdma),
+                            wc_status, rds_rdma_send_complete);
+
+       if (op->op_write)
+               rds_stats_add(s_send_rdma_bytes, op->op_bytes);
+       else
+               rds_stats_add(s_recv_rdma_bytes, op->op_bytes);
 }
 
-static void rds_ib_send_unmap_rm(struct rds_ib_connection *ic,
-                         struct rds_ib_send_work *send,
-                         int wc_status)
+static void rds_ib_send_unmap_atomic(struct rds_ib_connection *ic,
+                                    struct rm_atomic_op *op,
+                                    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->data.m_sg, rm->data.m_nents,
-                       DMA_TO_DEVICE);
-
-       if (rm->rdma.m_rdma_op) {
-               rds_ib_send_unmap_rdma(ic, rm->rdma.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);
+       /* unmap atomic recvbuf */
+       if (op->op_mapped) {
+               ib_dma_unmap_sg(ic->i_cm_id->device, op->op_sg, 1,
+                               DMA_FROM_DEVICE);
+               op->op_mapped = 0;
+       }
 
-               if (rm->rdma.m_rdma_op->r_write)
-                       rds_stats_add(s_send_rdma_bytes, rm->rdma.m_rdma_op->r_bytes);
-               else
-                       rds_stats_add(s_recv_rdma_bytes, rm->rdma.m_rdma_op->r_bytes);
+       rds_ib_send_complete(container_of(op, struct rds_message, atomic),
+                            wc_status, rds_atomic_send_complete);
+
+       if (op->op_type == RDS_ATOMIC_TYPE_CSWP)
+               rds_ib_stats_inc(s_ib_atomic_cswp);
+       else
+               rds_ib_stats_inc(s_ib_atomic_fadd);
+}
+
+/*
+ * Unmap the resources associated with a struct send_work.
+ *
+ * Returns the rm for no good reason other than it is unobtainable
+ * other than by switching on wr.opcode, currently, and the caller,
+ * the event handler, needs it.
+ */
+static struct rds_message *rds_ib_send_unmap_op(struct rds_ib_connection *ic,
+                                               struct rds_ib_send_work *send,
+                                               int wc_status)
+{
+       struct rds_message *rm = NULL;
+
+       /* In the error case, wc.opcode sometimes contains garbage */
+       switch (send->s_wr.opcode) {
+       case IB_WR_SEND:
+               if (send->s_op) {
+                       rm = container_of(send->s_op, struct rds_message, data);
+                       rds_ib_send_unmap_data(ic, send->s_op, wc_status);
+               }
+               break;
+       case IB_WR_RDMA_WRITE:
+       case IB_WR_RDMA_READ:
+               if (send->s_op) {
+                       rm = container_of(send->s_op, struct rds_message, rdma);
+                       rds_ib_send_unmap_rdma(ic, send->s_op, wc_status);
+               }
+               break;
+       case IB_WR_ATOMIC_FETCH_AND_ADD:
+       case IB_WR_ATOMIC_CMP_AND_SWP:
+               if (send->s_op) {
+                       rm = container_of(send->s_op, struct rds_message, atomic);
+                       rds_ib_send_unmap_atomic(ic, send->s_op, wc_status);
+               }
+               break;
+       default:
+               if (printk_ratelimit())
+                       printk(KERN_NOTICE
+                              "RDS/IB: %s: unexpected opcode 0x%x in WR!\n",
+                              __func__, send->s_wr.opcode);
+               break;
        }
 
-       /* If anyone waited for this message to get flushed out, wake
-        * them up now */
-       rds_message_unmapped(rm);
+       send->s_wr.opcode = 0xdead;
 
-       rds_message_put(rm);
-       send->s_rm = NULL;
+       return rm;
 }
 
 void rds_ib_send_init_ring(struct rds_ib_connection *ic)
@@ -133,23 +193,18 @@ void rds_ib_send_init_ring(struct rds_ib_connection *ic)
        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 = &send->s_sge[0];
                sge->addr = ic->i_send_hdrs_dma + (i * sizeof(struct rds_header));
                sge->length = sizeof(struct rds_header);
                sge->lkey = ic->i_mr->lkey;
+
+               send->s_sge[1].lkey = ic->i_mr->lkey;
        }
 }
 
@@ -159,12 +214,8 @@ void rds_ib_send_clear_ring(struct rds_ib_connection *ic)
        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);
+               if (send->s_op && send->s_wr.opcode != 0xdead)
+                       rds_ib_send_unmap_op(ic, send, IB_WC_WR_FLUSH_ERR);
        }
 }
 
@@ -178,6 +229,7 @@ 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 rds_message *rm = NULL;
        struct ib_wc wc;
        struct rds_ib_send_work *send;
        u32 completed;
@@ -211,44 +263,19 @@ void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context)
                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;
-                       }
+                       rm = rds_ib_send_unmap_op(ic, send, wc.status);
 
-                       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) {
-                                       if (rm->rdma.m_rdma_op)
-                                               rds_ib_send_unmap_rdma(ic, rm->rdma.m_rdma_op);
-                                       rds_ib_send_rdma_complete(rm, wc.status);
-                                       rds_message_put(rm);
+                       if (send->s_op) {
+                               if (send->s_op == rm->m_final_op) {
+                                       /* If anyone waited for this message to get flushed out, wake
+                                        * them up now */
+                                       rds_message_unmapped(rm);
                                }
+                               rds_message_put(rm);
+                               send->s_op = NULL;
                        }
 
                        oldest = (oldest + 1) % ic->i_send_ring.w_nr;
@@ -294,7 +321,7 @@ void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context)
  * 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.
+ * sets RDS_IN_XMIT to ensure exclusive access to the send ring.
  * However, the ACK sending code is independent and can race with
  * message SENDs.
  *
@@ -413,40 +440,19 @@ void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted)
                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)
+static inline void rds_ib_set_wr_signal_state(struct rds_ib_connection *ic,
+                                             struct rds_ib_send_work *send,
+                                             bool notify)
 {
-       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];
+       /*
+        * 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 || notify) {
+               ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
+               send->s_wr.send_flags |= IB_SEND_SIGNALED;
        }
-
-       sge->addr = ic->i_send_hdrs_dma + (pos * sizeof(struct rds_header));
-       sge->length = sizeof(struct rds_header);
-       sge->lkey = ic->i_mr->lkey;
 }
 
 /*
@@ -475,11 +481,11 @@ int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
        u32 pos;
        u32 i;
        u32 work_alloc;
-       u32 credit_alloc;
+       u32 credit_alloc = 0;
        u32 posted;
        u32 adv_credits = 0;
        int send_flags = 0;
-       int sent;
+       int bytes_sent = 0;
        int ret;
        int flow_controlled = 0;
 
@@ -507,14 +513,13 @@ int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
                goto out;
        }
 
-       credit_alloc = work_alloc;
        if (ic->i_flowctl) {
                credit_alloc = rds_ib_send_grab_credits(ic, work_alloc, &posted, 0, RDS_MAX_ADV_CREDIT);
                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++;
+                       flow_controlled = 1;
                }
                if (work_alloc == 0) {
                        set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
@@ -525,33 +530,25 @@ int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
        }
 
        /* map the message the first time we see it */
-       if (!ic->i_rm) {
-               /*
-               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->data.m_nents) {
-                       rm->data.m_count = ib_dma_map_sg(dev,
-                                                           rm->data.m_sg,
-                                                           rm->data.m_nents,
-                                                           DMA_TO_DEVICE);
-                       rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->data.m_count);
-                       if (rm->data.m_count == 0) {
+       if (!ic->i_data_op) {
+               if (rm->data.op_nents) {
+                       rm->data.op_count = ib_dma_map_sg(dev,
+                                                         rm->data.op_sg,
+                                                         rm->data.op_nents,
+                                                         DMA_TO_DEVICE);
+                       rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->data.op_count);
+                       if (rm->data.op_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->data.m_count = 0;
+                       rm->data.op_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;
+               ic->i_data_op = &rm->data;
 
                /* Finalize the header */
                if (test_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags))
@@ -561,10 +558,10 @@ int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
 
                /* 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->rdma.m_rdma_op) {
+               if (rm->rdma.op_active) {
                        struct rds_ext_header_rdma ext_hdr;
 
-                       ext_hdr.h_rdma_rkey = cpu_to_be32(rm->rdma.m_rdma_op->r_key);
+                       ext_hdr.h_rdma_rkey = cpu_to_be32(rm->rdma.op_rkey);
                        rds_message_add_extension(&rm->m_inc.i_hdr,
                                        RDS_EXTHDR_RDMA, &ext_hdr, sizeof(ext_hdr));
                }
@@ -584,99 +581,74 @@ int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
                /*
                 * Update adv_credits since we reset the ACK_REQUIRED bit.
                 */
-               rds_ib_send_grab_credits(ic, 0, &posted, 1, RDS_MAX_ADV_CREDIT - adv_credits);
-               adv_credits += posted;
-               BUG_ON(adv_credits > 255);
+               if (ic->i_flowctl) {
+                       rds_ib_send_grab_credits(ic, 0, &posted, 1, RDS_MAX_ADV_CREDIT - adv_credits);
+                       adv_credits += posted;
+                       BUG_ON(adv_credits > 255);
+               }
        }
 
-       send = &ic->i_sends[pos];
-       first = send;
-       prev = NULL;
-       scat = &rm->data.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->rdma.m_rdma_op && rm->rdma.m_rdma_op->r_fence)
+       if (rm->rdma.op_active && rm->rdma.op_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.
-        */
+       /* Each frag gets a header. Msgs may be 0 bytes */
+       send = &ic->i_sends[pos];
+       first = send;
+       prev = NULL;
+       scat = &ic->i_data_op->op_sg[sg];
+       i = 0;
+       do {
+               unsigned int len = 0;
 
-       /* 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;
-       }
+               /* Set up the header */
+               send->s_wr.send_flags = send_flags;
+               send->s_wr.opcode = IB_WR_SEND;
+               send->s_wr.num_sge = 1;
+               send->s_wr.next = NULL;
+               send->s_queued = jiffies;
+               send->s_op = NULL;
 
-       /* if there's data reference it with a chain of work reqs */
-       for (; i < work_alloc && scat != &rm->data.m_sg[rm->data.m_count]; i++) {
-               unsigned int len;
+               send->s_sge[0].addr = ic->i_send_hdrs_dma
+                       + (pos * sizeof(struct rds_header));
+               send->s_sge[0].length = sizeof(struct rds_header);
 
-               send = &ic->i_sends[pos];
+               memcpy(&ic->i_send_hdrs[pos], &rm->m_inc.i_hdr, sizeof(struct rds_header));
 
-               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);
+               /* Set up the data, if present */
+               if (i < work_alloc
+                   && scat != &rm->data.op_sg[rm->data.op_count]) {
+                       len = min(RDS_FRAG_SIZE, ib_sg_dma_len(dev, scat) - off);
+                       send->s_wr.num_sge = 2;
 
-               /*
-                * 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;
-               }
+                       send->s_sge[1].addr = ib_sg_dma_address(dev, scat) + off;
+                       send->s_sge[1].length = len;
 
-               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;
+                       bytes_sent += len;
+                       off += len;
+                       if (off == ib_sg_dma_len(dev, scat)) {
+                               scat++;
+                               off = 0;
+                       }
                }
 
+               rds_ib_set_wr_signal_state(ic, send, 0);
+
                /*
                 * Always signal the last one if we're stopping due to flow control.
                 */
-               if (flow_controlled && i == (work_alloc-1))
+               if (ic->i_flowctl && 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) {
+               if (ic->i_flowctl && adv_credits) {
                        struct rds_header *hdr = &ic->i_send_hdrs[pos];
 
                        /* add credit and redo the header checksum */
@@ -691,20 +663,25 @@ add_header:
                prev = send;
 
                pos = (pos + 1) % ic->i_send_ring.w_nr;
-       }
+               send = &ic->i_sends[pos];
+               i++;
+
+       } while (i < work_alloc
+                && scat != &rm->data.op_sg[rm->data.op_count]);
 
        /* 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);
+               bytes_sent += sizeof(struct rds_header);
 
        /* if we finished the message then send completion owns it */
-       if (scat == &rm->data.m_sg[rm->data.m_count]) {
-               prev->s_rm = ic->i_rm;
-               prev->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
-               ic->i_rm = NULL;
+       if (scat == &rm->data.op_sg[rm->data.op_count]) {
+               prev->s_op = ic->i_data_op;
+               prev->s_wr.send_flags |= IB_SEND_SOLICITED;
+               ic->i_data_op = NULL;
        }
 
+       /* Put back wrs & credits we didn't use */
        if (i < work_alloc) {
                rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
                work_alloc = i;
@@ -722,32 +699,117 @@ add_header:
                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;
+               if (prev->s_op) {
+                       ic->i_data_op = prev->s_op;
+                       prev->s_op = NULL;
                }
 
                rds_ib_conn_error(ic->conn, "ib_post_send failed\n");
                goto out;
        }
 
-       ret = sent;
+       ret = bytes_sent;
 out:
        BUG_ON(adv_credits);
        return ret;
 }
 
-int rds_ib_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
+/*
+ * Issue atomic operation.
+ * A simplified version of the rdma case, we always map 1 SG, and
+ * only 8 bytes, for the return value from the atomic operation.
+ */
+int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op)
+{
+       struct rds_ib_connection *ic = conn->c_transport_data;
+       struct rds_ib_send_work *send = NULL;
+       struct ib_send_wr *failed_wr;
+       struct rds_ib_device *rds_ibdev;
+       u32 pos;
+       u32 work_alloc;
+       int ret;
+
+       rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
+
+       work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, 1, &pos);
+       if (work_alloc != 1) {
+               rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
+               rds_ib_stats_inc(s_ib_tx_ring_full);
+               ret = -ENOMEM;
+               goto out;
+       }
+
+       /* address of send request in ring */
+       send = &ic->i_sends[pos];
+       send->s_queued = jiffies;
+
+       if (op->op_type == RDS_ATOMIC_TYPE_CSWP) {
+               send->s_wr.opcode = IB_WR_ATOMIC_CMP_AND_SWP;
+               send->s_wr.wr.atomic.compare_add = op->op_compare;
+               send->s_wr.wr.atomic.swap = op->op_swap_add;
+       } else { /* FADD */
+               send->s_wr.opcode = IB_WR_ATOMIC_FETCH_AND_ADD;
+               send->s_wr.wr.atomic.compare_add = op->op_swap_add;
+               send->s_wr.wr.atomic.swap = 0;
+       }
+       rds_ib_set_wr_signal_state(ic, send, op->op_notify);
+       send->s_wr.num_sge = 1;
+       send->s_wr.next = NULL;
+       send->s_wr.wr.atomic.remote_addr = op->op_remote_addr;
+       send->s_wr.wr.atomic.rkey = op->op_rkey;
+       send->s_op = op;
+       rds_message_addref(container_of(send->s_op, struct rds_message, atomic));
+
+       /* map 8 byte retval buffer to the device */
+       ret = ib_dma_map_sg(ic->i_cm_id->device, op->op_sg, 1, DMA_FROM_DEVICE);
+       rdsdebug("ic %p mapping atomic op %p. mapped %d pg\n", ic, op, ret);
+       if (ret != 1) {
+               rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
+               rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
+               ret = -ENOMEM; /* XXX ? */
+               goto out;
+       }
+
+       /* Convert our struct scatterlist to struct ib_sge */
+       send->s_sge[0].addr = ib_sg_dma_address(ic->i_cm_id->device, op->op_sg);
+       send->s_sge[0].length = ib_sg_dma_len(ic->i_cm_id->device, op->op_sg);
+       send->s_sge[0].lkey = ic->i_mr->lkey;
+
+       rdsdebug("rva %Lx rpa %Lx len %u\n", op->op_remote_addr,
+                send->s_sge[0].addr, send->s_sge[0].length);
+
+       failed_wr = &send->s_wr;
+       ret = ib_post_send(ic->i_cm_id->qp, &send->s_wr, &failed_wr);
+       rdsdebug("ic %p send %p (wr %p) ret %d wr %p\n", ic,
+                send, &send->s_wr, ret, failed_wr);
+       BUG_ON(failed_wr != &send->s_wr);
+       if (ret) {
+               printk(KERN_WARNING "RDS/IB: atomic 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 != &send->s_wr)) {
+               printk(KERN_WARNING "RDS/IB: atomic ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
+               BUG_ON(failed_wr != &send->s_wr);
+       }
+
+out:
+       return ret;
+}
+
+int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_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;
+       u64 remote_addr = op->op_remote_addr;
+       u32 max_sge = ic->rds_ibdev->max_sge;
        u32 pos;
        u32 work_alloc;
        u32 i;
@@ -756,28 +818,26 @@ int rds_ib_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
        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) {
+       /* map the op the first time we see it */
+       if (!op->op_mapped) {
+               op->op_count = ib_dma_map_sg(ic->i_cm_id->device,
+                                            op->op_sg, op->op_nents, (op->op_write) ?
+                                            DMA_TO_DEVICE : DMA_FROM_DEVICE);
+               rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->op_count);
+               if (op->op_count == 0) {
                        rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
                        ret = -ENOMEM; /* XXX ? */
                        goto out;
                }
 
-               op->r_mapped = 1;
+               op->op_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);
+       i = ceil(op->op_count, max_sge);
 
        work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
        if (work_alloc != i) {
@@ -790,30 +850,24 @@ int rds_ib_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
        send = &ic->i_sends[pos];
        first = send;
        prev = NULL;
-       scat = &op->r_sg[0];
+       scat = &op->op_sg[0];
        sent = 0;
-       num_sge = op->r_count;
+       num_sge = op->op_count;
 
-       for (i = 0; i < work_alloc && scat != &op->r_sg[op->r_count]; i++) {
+       for (i = 0; i < work_alloc && scat != &op->op_sg[op->op_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_op = NULL;
+
+               rds_ib_set_wr_signal_state(ic, send, op->op_notify);
 
-               send->s_wr.opcode = op->r_write ? IB_WR_RDMA_WRITE : IB_WR_RDMA_READ;
+               send->s_wr.opcode = op->op_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;
+               send->s_wr.wr.rdma.rkey = op->op_rkey;
 
-               if (num_sge > rds_ibdev->max_sge) {
-                       send->s_wr.num_sge = rds_ibdev->max_sge;
-                       num_sge -= rds_ibdev->max_sge;
+               if (num_sge > max_sge) {
+                       send->s_wr.num_sge = max_sge;
+                       num_sge -= max_sge;
                } else {
                        send->s_wr.num_sge = num_sge;
                }
@@ -823,7 +877,7 @@ int rds_ib_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
                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++) {
+               for (j = 0; j < send->s_wr.num_sge && scat != &op->op_sg[op->op_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);
@@ -845,9 +899,11 @@ int rds_ib_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
                        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;
+       /* give a reference to the last op */
+       if (scat == &op->op_sg[op->op_count]) {
+               prev->s_op = op;
+               rds_message_addref(container_of(op, struct rds_message, rdma));
+       }
 
        if (i < work_alloc) {
                rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);