net/rds/ib_send.c

   1 /*
   2  * Copyright (c) 2006 Oracle.  All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  *
  32  */
  33 #include <linux/kernel.h>
  34 #include <linux/in.h>
  35 #include <linux/device.h>
  36 #include <linux/dmapool.h>
  37
  38 #include "rds.h"
  39 #include "rdma.h"
  40 #include "ib.h"
  41
  42 static void rds_ib_send_rdma_complete(struct rds_message *rm,
  43                                       int wc_status)
  44 {
  45         int notify_status;
  46
  47         switch (wc_status) {
  48         case IB_WC_WR_FLUSH_ERR:
  49                 return;
  50
  51         case IB_WC_SUCCESS:
  52                 notify_status = RDS_RDMA_SUCCESS;
  53                 break;
  54
  55         case IB_WC_REM_ACCESS_ERR:
  56                 notify_status = RDS_RDMA_REMOTE_ERROR;
  57                 break;
  58
  59         default:
  60                 notify_status = RDS_RDMA_OTHER_ERROR;
  61                 break;
  62         }
  63         rds_rdma_send_complete(rm, notify_status);
  64 }
  65
  66 static void rds_ib_send_unmap_rdma(struct rds_ib_connection *ic,
  67                                    struct rds_rdma_op *op)
  68 {
  69         if (op->r_mapped) {
  70                 ib_dma_unmap_sg(ic->i_cm_id->device,
  71                         op->r_sg, op->r_nents,
  72                         op->r_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
  73                 op->r_mapped = 0;
  74         }
  75 }
  76
  77 static void rds_ib_send_unmap_rm(struct rds_ib_connection *ic,
  78                           struct rds_ib_send_work *send,
  79                           int wc_status)
  80 {
  81         struct rds_message *rm = send->s_rm;
  82
  83         rdsdebug("ic %p send %p rm %p\n", ic, send, rm);
  84
  85         ib_dma_unmap_sg(ic->i_cm_id->device,
  86                      rm->m_sg, rm->m_nents,
  87                      DMA_TO_DEVICE);
  88
  89         if (rm->m_rdma_op != NULL) {
  90                 rds_ib_send_unmap_rdma(ic, rm->m_rdma_op);
  91
  92                 /* If the user asked for a completion notification on this
  93                  * message, we can implement three different semantics:
  94                  *  1.  Notify when we received the ACK on the RDS message
  95                  *      that was queued with the RDMA. This provides reliable
  96                  *      notification of RDMA status at the expense of a one-way
  97                  *      packet delay.
  98                  *  2.  Notify when the IB stack gives us the completion event for
  99                  *      the RDMA operation.
 100                  *  3.  Notify when the IB stack gives us the completion event for
 101                  *      the accompanying RDS messages.
 102                  * Here, we implement approach #3. To implement approach #2,
 103                  * call rds_rdma_send_complete from the cq_handler. To implement #1,
 104                  * don't call rds_rdma_send_complete at all, and fall back to the notify
 105                  * handling in the ACK processing code.
 106                  *
 107                  * Note: There's no need to explicitly sync any RDMA buffers using
 108                  * ib_dma_sync_sg_for_cpu - the completion for the RDMA
 109                  * operation itself unmapped the RDMA buffers, which takes care
 110                  * of synching.
 111                  */
 112                 rds_ib_send_rdma_complete(rm, wc_status);
 113
 114                 if (rm->m_rdma_op->r_write)
 115                         rds_stats_add(s_send_rdma_bytes, rm->m_rdma_op->r_bytes);
 116                 else
 117                         rds_stats_add(s_recv_rdma_bytes, rm->m_rdma_op->r_bytes);
 118         }
 119
 120         /* If anyone waited for this message to get flushed out, wake
 121          * them up now */
 122         rds_message_unmapped(rm);
 123
 124         rds_message_put(rm);
 125         send->s_rm = NULL;
 126 }
 127
 128 void rds_ib_send_init_ring(struct rds_ib_connection *ic)
 129 {
 130         struct rds_ib_send_work *send;
 131         u32 i;
 132
 133         for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
 134                 struct ib_sge *sge;
 135
 136                 send->s_rm = NULL;
 137                 send->s_op = NULL;
 138
 139                 send->s_wr.wr_id = i;
 140                 send->s_wr.sg_list = send->s_sge;
 141                 send->s_wr.num_sge = 1;
 142                 send->s_wr.opcode = IB_WR_SEND;
 143                 send->s_wr.send_flags = 0;
 144                 send->s_wr.ex.imm_data = 0;
 145
 146                 sge = rds_ib_data_sge(ic, send->s_sge);
 147                 sge->lkey = ic->i_mr->lkey;
 148
 149                 sge = rds_ib_header_sge(ic, send->s_sge);
 150                 sge->addr = ic->i_send_hdrs_dma + (i * sizeof(struct rds_header));
 151                 sge->length = sizeof(struct rds_header);
 152                 sge->lkey = ic->i_mr->lkey;
 153         }
 154 }
 155
 156 void rds_ib_send_clear_ring(struct rds_ib_connection *ic)
 157 {
 158         struct rds_ib_send_work *send;
 159         u32 i;
 160
 161         for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
 162                 if (send->s_wr.opcode == 0xdead)
 163                         continue;
 164                 if (send->s_rm)
 165                         rds_ib_send_unmap_rm(ic, send, IB_WC_WR_FLUSH_ERR);
 166                 if (send->s_op)
 167                         rds_ib_send_unmap_rdma(ic, send->s_op);
 168         }
 169 }
 170
 171 /*
 172  * The _oldest/_free ring operations here race cleanly with the alloc/unalloc
 173  * operations performed in the send path.  As the sender allocs and potentially
 174  * unallocs the next free entry in the ring it doesn't alter which is
 175  * the next to be freed, which is what this is concerned with.
 176  */
 177 void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context)
 178 {
 179         struct rds_connection *conn = context;
 180         struct rds_ib_connection *ic = conn->c_transport_data;
 181         struct ib_wc wc;
 182         struct rds_ib_send_work *send;
 183         u32 completed;
 184         u32 oldest;
 185         u32 i = 0;
 186         int ret;
 187
 188         rdsdebug("cq %p conn %p\n", cq, conn);
 189         rds_ib_stats_inc(s_ib_tx_cq_call);
 190         ret = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 191         if (ret)
 192                 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
 193
 194         while (ib_poll_cq(cq, 1, &wc) > 0) {
 195                 rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
 196                          (unsigned long long)wc.wr_id, wc.status, wc.byte_len,
 197                          be32_to_cpu(wc.ex.imm_data));
 198                 rds_ib_stats_inc(s_ib_tx_cq_event);
 199
 200                 if (wc.wr_id == RDS_IB_ACK_WR_ID) {
 201                         if (ic->i_ack_queued + HZ/2 < jiffies)
 202                                 rds_ib_stats_inc(s_ib_tx_stalled);
 203                         rds_ib_ack_send_complete(ic);
 204                         continue;
 205                 }
 206
 207                 oldest = rds_ib_ring_oldest(&ic->i_send_ring);
 208
 209                 completed = rds_ib_ring_completed(&ic->i_send_ring, wc.wr_id, oldest);
 210
 211                 for (i = 0; i < completed; i++) {
 212                         send = &ic->i_sends[oldest];
 213
 214                         /* In the error case, wc.opcode sometimes contains garbage */
 215                         switch (send->s_wr.opcode) {
 216                         case IB_WR_SEND:
 217                                 if (send->s_rm)
 218                                         rds_ib_send_unmap_rm(ic, send, wc.status);
 219                                 break;
 220                         case IB_WR_RDMA_WRITE:
 221                         case IB_WR_RDMA_READ:
 222                                 /* Nothing to be done - the SG list will be unmapped
 223                                  * when the SEND completes. */
 224                                 break;
 225                         default:
 226                                 if (printk_ratelimit())
 227                                         printk(KERN_NOTICE
 228                                                 "RDS/IB: %s: unexpected opcode 0x%x in WR!\n",
 229                                                 __func__, send->s_wr.opcode);
 230                                 break;
 231                         }
 232
 233                         send->s_wr.opcode = 0xdead;
 234                         send->s_wr.num_sge = 1;
 235                         if (send->s_queued + HZ/2 < jiffies)
 236                                 rds_ib_stats_inc(s_ib_tx_stalled);
 237
 238                         /* If a RDMA operation produced an error, signal this right
 239                          * away. If we don't, the subsequent SEND that goes with this
 240                          * RDMA will be canceled with ERR_WFLUSH, and the application
 241                          * never learn that the RDMA failed. */
 242                         if (unlikely(wc.status == IB_WC_REM_ACCESS_ERR && send->s_op)) {
 243                                 struct rds_message *rm;
 244
 245                                 rm = rds_send_get_message(conn, send->s_op);
 246                                 if (rm) {
 247                                         if (rm->m_rdma_op)
 248                                                 rds_ib_send_unmap_rdma(ic, rm->m_rdma_op);
 249                                         rds_ib_send_rdma_complete(rm, wc.status);
 250                                         rds_message_put(rm);
 251                                 }
 252                         }
 253
 254                         oldest = (oldest + 1) % ic->i_send_ring.w_nr;
 255                 }
 256
 257                 rds_ib_ring_free(&ic->i_send_ring, completed);
 258
 259                 if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
 260                     test_bit(0, &conn->c_map_queued))
 261                         queue_delayed_work(rds_wq, &conn->c_send_w, 0);
 262
 263                 /* We expect errors as the qp is drained during shutdown */
 264                 if (wc.status != IB_WC_SUCCESS && rds_conn_up(conn)) {
 265                         rds_ib_conn_error(conn,
 266                                 "send completion on %pI4 "
 267                                 "had status %u, disconnecting and reconnecting\n",
 268                                 &conn->c_faddr, wc.status);
 269                 }
 270         }
 271 }
 272
 273 /*
 274  * This is the main function for allocating credits when sending
 275  * messages.
 276  *
 277  * Conceptually, we have two counters:
 278  *  -   send credits: this tells us how many WRs we're allowed
 279  *      to submit without overruning the reciever's queue. For
 280  *      each SEND WR we post, we decrement this by one.
 281  *
 282  *  -   posted credits: this tells us how many WRs we recently
 283  *      posted to the receive queue. This value is transferred
 284  *      to the peer as a "credit update" in a RDS header field.
 285  *      Every time we transmit credits to the peer, we subtract
 286  *      the amount of transferred credits from this counter.
 287  *
 288  * It is essential that we avoid situations where both sides have
 289  * exhausted their send credits, and are unable to send new credits
 290  * to the peer. We achieve this by requiring that we send at least
 291  * one credit update to the peer before exhausting our credits.
 292  * When new credits arrive, we subtract one credit that is withheld
 293  * until we've posted new buffers and are ready to transmit these
 294  * credits (see rds_ib_send_add_credits below).
 295  *
 296  * The RDS send code is essentially single-threaded; rds_send_xmit
 297  * grabs c_send_lock to ensure exclusive access to the send ring.
 298  * However, the ACK sending code is independent and can race with
 299  * message SENDs.
 300  *
 301  * In the send path, we need to update the counters for send credits
 302  * and the counter of posted buffers atomically - when we use the
 303  * last available credit, we cannot allow another thread to race us
 304  * and grab the posted credits counter.  Hence, we have to use a
 305  * spinlock to protect the credit counter, or use atomics.
 306  *
 307  * Spinlocks shared between the send and the receive path are bad,
 308  * because they create unnecessary delays. An early implementation
 309  * using a spinlock showed a 5% degradation in throughput at some
 310  * loads.
 311  *
 312  * This implementation avoids spinlocks completely, putting both
 313  * counters into a single atomic, and updating that atomic using
 314  * atomic_add (in the receive path, when receiving fresh credits),
 315  * and using atomic_cmpxchg when updating the two counters.
 316  */
 317 int rds_ib_send_grab_credits(struct rds_ib_connection *ic,
 318                              u32 wanted, u32 *adv_credits, int need_posted, int max_posted)
 319 {
 320         unsigned int avail, posted, got = 0, advertise;
 321         long oldval, newval;
 322
 323         *adv_credits = 0;
 324         if (!ic->i_flowctl)
 325                 return wanted;
 326
 327 try_again:
 328         advertise = 0;
 329         oldval = newval = atomic_read(&ic->i_credits);
 330         posted = IB_GET_POST_CREDITS(oldval);
 331         avail = IB_GET_SEND_CREDITS(oldval);
 332
 333         rdsdebug("rds_ib_send_grab_credits(%u): credits=%u posted=%u\n",
 334                         wanted, avail, posted);
 335
 336         /* The last credit must be used to send a credit update. */
 337         if (avail && !posted)
 338                 avail--;
 339
 340         if (avail < wanted) {
 341                 struct rds_connection *conn = ic->i_cm_id->context;
 342
 343                 /* Oops, there aren't that many credits left! */
 344                 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 345                 got = avail;
 346         } else {
 347                 /* Sometimes you get what you want, lalala. */
 348                 got = wanted;
 349         }
 350         newval -= IB_SET_SEND_CREDITS(got);
 351
 352         /*
 353          * If need_posted is non-zero, then the caller wants
 354          * the posted regardless of whether any send credits are
 355          * available.
 356          */
 357         if (posted && (got || need_posted)) {
 358                 advertise = min_t(unsigned int, posted, max_posted);
 359                 newval -= IB_SET_POST_CREDITS(advertise);
 360         }
 361
 362         /* Finally bill everything */
 363         if (atomic_cmpxchg(&ic->i_credits, oldval, newval) != oldval)
 364                 goto try_again;
 365
 366         *adv_credits = advertise;
 367         return got;
 368 }
 369
 370 void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits)
 371 {
 372         struct rds_ib_connection *ic = conn->c_transport_data;
 373
 374         if (credits == 0)
 375                 return;
 376
 377         rdsdebug("rds_ib_send_add_credits(%u): current=%u%s\n",
 378                         credits,
 379                         IB_GET_SEND_CREDITS(atomic_read(&ic->i_credits)),
 380                         test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ? ", ll_send_full" : "");
 381
 382         atomic_add(IB_SET_SEND_CREDITS(credits), &ic->i_credits);
 383         if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags))
 384                 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
 385
 386         WARN_ON(IB_GET_SEND_CREDITS(credits) >= 16384);
 387
 388         rds_ib_stats_inc(s_ib_rx_credit_updates);
 389 }
 390
 391 void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted)
 392 {
 393         struct rds_ib_connection *ic = conn->c_transport_data;
 394
 395         if (posted == 0)
 396                 return;
 397
 398         atomic_add(IB_SET_POST_CREDITS(posted), &ic->i_credits);
 399
 400         /* Decide whether to send an update to the peer now.
 401          * If we would send a credit update for every single buffer we
 402          * post, we would end up with an ACK storm (ACK arrives,
 403          * consumes buffer, we refill the ring, send ACK to remote
 404          * advertising the newly posted buffer... ad inf)
 405          *
 406          * Performance pretty much depends on how often we send
 407          * credit updates - too frequent updates mean lots of ACKs.
 408          * Too infrequent updates, and the peer will run out of
 409          * credits and has to throttle.
 410          * For the time being, 16 seems to be a good compromise.
 411          */
 412         if (IB_GET_POST_CREDITS(atomic_read(&ic->i_credits)) >= 16)
 413                 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
 414 }
 415
 416 static inline void
 417 rds_ib_xmit_populate_wr(struct rds_ib_connection *ic,
 418                 struct rds_ib_send_work *send, unsigned int pos,
 419                 unsigned long buffer, unsigned int length,
 420                 int send_flags)
 421 {
 422         struct ib_sge *sge;
 423
 424         WARN_ON(pos != send - ic->i_sends);
 425
 426         send->s_wr.send_flags = send_flags;
 427         send->s_wr.opcode = IB_WR_SEND;
 428         send->s_wr.num_sge = 2;
 429         send->s_wr.next = NULL;
 430         send->s_queued = jiffies;
 431         send->s_op = NULL;
 432
 433         if (length != 0) {
 434                 sge = rds_ib_data_sge(ic, send->s_sge);
 435                 sge->addr = buffer;
 436                 sge->length = length;
 437                 sge->lkey = ic->i_mr->lkey;
 438
 439                 sge = rds_ib_header_sge(ic, send->s_sge);
 440         } else {
 441                 /* We're sending a packet with no payload. There is only
 442                  * one SGE */
 443                 send->s_wr.num_sge = 1;
 444                 sge = &send->s_sge[0];
 445         }
 446
 447         sge->addr = ic->i_send_hdrs_dma + (pos * sizeof(struct rds_header));
 448         sge->length = sizeof(struct rds_header);
 449         sge->lkey = ic->i_mr->lkey;
 450 }
 451
 452 /*
 453  * This can be called multiple times for a given message.  The first time
 454  * we see a message we map its scatterlist into the IB device so that
 455  * we can provide that mapped address to the IB scatter gather entries
 456  * in the IB work requests.  We translate the scatterlist into a series
 457  * of work requests that fragment the message.  These work requests complete
 458  * in order so we pass ownership of the message to the completion handler
 459  * once we send the final fragment.
 460  *
 461  * The RDS core uses the c_send_lock to only enter this function once
 462  * per connection.  This makes sure that the tx ring alloc/unalloc pairs
 463  * don't get out of sync and confuse the ring.
 464  */
 465 int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
 466                 unsigned int hdr_off, unsigned int sg, unsigned int off)
 467 {
 468         struct rds_ib_connection *ic = conn->c_transport_data;
 469         struct ib_device *dev = ic->i_cm_id->device;
 470         struct rds_ib_send_work *send = NULL;
 471         struct rds_ib_send_work *first;
 472         struct rds_ib_send_work *prev;
 473         struct ib_send_wr *failed_wr;
 474         struct scatterlist *scat;
 475         u32 pos;
 476         u32 i;
 477         u32 work_alloc;
 478         u32 credit_alloc;
 479         u32 posted;
 480         u32 adv_credits = 0;
 481         int send_flags = 0;
 482         int sent;
 483         int ret;
 484         int flow_controlled = 0;
 485
 486         BUG_ON(off % RDS_FRAG_SIZE);
 487         BUG_ON(hdr_off != 0 && hdr_off != sizeof(struct rds_header));
 488
 489         /* Do not send cong updates to IB loopback */
 490         if (conn->c_loopback
 491             && rm->m_inc.i_hdr.h_flags & RDS_FLAG_CONG_BITMAP) {
 492                 rds_cong_map_updated(conn->c_fcong, ~(u64) 0);
 493                 return sizeof(struct rds_header) + RDS_CONG_MAP_BYTES;
 494         }
 495
 496         /* FIXME we may overallocate here */
 497         if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0)
 498                 i = 1;
 499         else
 500                 i = ceil(be32_to_cpu(rm->m_inc.i_hdr.h_len), RDS_FRAG_SIZE);
 501
 502         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
 503         if (work_alloc == 0) {
 504                 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 505                 rds_ib_stats_inc(s_ib_tx_ring_full);
 506                 ret = -ENOMEM;
 507                 goto out;
 508         }
 509
 510         credit_alloc = work_alloc;
 511         if (ic->i_flowctl) {
 512                 credit_alloc = rds_ib_send_grab_credits(ic, work_alloc, &posted, 0, RDS_MAX_ADV_CREDIT);
 513                 adv_credits += posted;
 514                 if (credit_alloc < work_alloc) {
 515                         rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - credit_alloc);
 516                         work_alloc = credit_alloc;
 517                         flow_controlled++;
 518                 }
 519                 if (work_alloc == 0) {
 520                         set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 521                         rds_ib_stats_inc(s_ib_tx_throttle);
 522                         ret = -ENOMEM;
 523                         goto out;
 524                 }
 525         }
 526
 527         /* map the message the first time we see it */
 528         if (ic->i_rm == NULL) {
 529                 /*
 530                 printk(KERN_NOTICE "rds_ib_xmit prep msg dport=%u flags=0x%x len=%d\n",
 531                                 be16_to_cpu(rm->m_inc.i_hdr.h_dport),
 532                                 rm->m_inc.i_hdr.h_flags,
 533                                 be32_to_cpu(rm->m_inc.i_hdr.h_len));
 534                    */
 535                 if (rm->m_nents) {
 536                         rm->m_count = ib_dma_map_sg(dev,
 537                                          rm->m_sg, rm->m_nents, DMA_TO_DEVICE);
 538                         rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->m_count);
 539                         if (rm->m_count == 0) {
 540                                 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
 541                                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 542                                 ret = -ENOMEM; /* XXX ? */
 543                                 goto out;
 544                         }
 545                 } else {
 546                         rm->m_count = 0;
 547                 }
 548
 549                 ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
 550                 ic->i_unsignaled_bytes = rds_ib_sysctl_max_unsig_bytes;
 551                 rds_message_addref(rm);
 552                 ic->i_rm = rm;
 553
 554                 /* Finalize the header */
 555                 if (test_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags))
 556                         rm->m_inc.i_hdr.h_flags |= RDS_FLAG_ACK_REQUIRED;
 557                 if (test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))
 558                         rm->m_inc.i_hdr.h_flags |= RDS_FLAG_RETRANSMITTED;
 559
 560                 /* If it has a RDMA op, tell the peer we did it. This is
 561                  * used by the peer to release use-once RDMA MRs. */
 562                 if (rm->m_rdma_op) {
 563                         struct rds_ext_header_rdma ext_hdr;
 564
 565                         ext_hdr.h_rdma_rkey = cpu_to_be32(rm->m_rdma_op->r_key);
 566                         rds_message_add_extension(&rm->m_inc.i_hdr,
 567                                         RDS_EXTHDR_RDMA, &ext_hdr, sizeof(ext_hdr));
 568                 }
 569                 if (rm->m_rdma_cookie) {
 570                         rds_message_add_rdma_dest_extension(&rm->m_inc.i_hdr,
 571                                         rds_rdma_cookie_key(rm->m_rdma_cookie),
 572                                         rds_rdma_cookie_offset(rm->m_rdma_cookie));
 573                 }
 574
 575                 /* Note - rds_ib_piggyb_ack clears the ACK_REQUIRED bit, so
 576                  * we should not do this unless we have a chance of at least
 577                  * sticking the header into the send ring. Which is why we
 578                  * should call rds_ib_ring_alloc first. */
 579                 rm->m_inc.i_hdr.h_ack = cpu_to_be64(rds_ib_piggyb_ack(ic));
 580                 rds_message_make_checksum(&rm->m_inc.i_hdr);
 581
 582                 /*
 583                  * Update adv_credits since we reset the ACK_REQUIRED bit.
 584                  */
 585                 rds_ib_send_grab_credits(ic, 0, &posted, 1, RDS_MAX_ADV_CREDIT - adv_credits);
 586                 adv_credits += posted;
 587                 BUG_ON(adv_credits > 255);
 588         }
 589
 590         send = &ic->i_sends[pos];
 591         first = send;
 592         prev = NULL;
 593         scat = &rm->m_sg[sg];
 594         sent = 0;
 595         i = 0;
 596
 597         /* Sometimes you want to put a fence between an RDMA
 598          * READ and the following SEND.
 599          * We could either do this all the time
 600          * or when requested by the user. Right now, we let
 601          * the application choose.
 602          */
 603         if (rm->m_rdma_op && rm->m_rdma_op->r_fence)
 604                 send_flags = IB_SEND_FENCE;
 605
 606         /*
 607          * We could be copying the header into the unused tail of the page.
 608          * That would need to be changed in the future when those pages might
 609          * be mapped userspace pages or page cache pages.  So instead we always
 610          * use a second sge and our long-lived ring of mapped headers.  We send
 611          * the header after the data so that the data payload can be aligned on
 612          * the receiver.
 613          */
 614
 615         /* handle a 0-len message */
 616         if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0) {
 617                 rds_ib_xmit_populate_wr(ic, send, pos, 0, 0, send_flags);
 618                 goto add_header;
 619         }
 620
 621         /* if there's data reference it with a chain of work reqs */
 622         for (; i < work_alloc && scat != &rm->m_sg[rm->m_count]; i++) {
 623                 unsigned int len;
 624
 625                 send = &ic->i_sends[pos];
 626
 627                 len = min(RDS_FRAG_SIZE, ib_sg_dma_len(dev, scat) - off);
 628                 rds_ib_xmit_populate_wr(ic, send, pos,
 629                                 ib_sg_dma_address(dev, scat) + off, len,
 630                                 send_flags);
 631
 632                 /*
 633                  * We want to delay signaling completions just enough to get
 634                  * the batching benefits but not so much that we create dead time
 635                  * on the wire.
 636                  */
 637                 if (ic->i_unsignaled_wrs-- == 0) {
 638                         ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
 639                         send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
 640                 }
 641
 642                 ic->i_unsignaled_bytes -= len;
 643                 if (ic->i_unsignaled_bytes <= 0) {
 644                         ic->i_unsignaled_bytes = rds_ib_sysctl_max_unsig_bytes;
 645                         send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
 646                 }
 647
 648                 /*
 649                  * Always signal the last one if we're stopping due to flow control.
 650                  */
 651                 if (flow_controlled && i == (work_alloc-1))
 652                         send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
 653
 654                 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
 655                          &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
 656
 657                 sent += len;
 658                 off += len;
 659                 if (off == ib_sg_dma_len(dev, scat)) {
 660                         scat++;
 661                         off = 0;
 662                 }
 663
 664 add_header:
 665                 /* Tack on the header after the data. The header SGE should already
 666                  * have been set up to point to the right header buffer. */
 667                 memcpy(&ic->i_send_hdrs[pos], &rm->m_inc.i_hdr, sizeof(struct rds_header));
 668
 669                 if (0) {
 670                         struct rds_header *hdr = &ic->i_send_hdrs[pos];
 671
 672                         printk(KERN_NOTICE "send WR dport=%u flags=0x%x len=%d\n",
 673                                 be16_to_cpu(hdr->h_dport),
 674                                 hdr->h_flags,
 675                                 be32_to_cpu(hdr->h_len));
 676                 }
 677                 if (adv_credits) {
 678                         struct rds_header *hdr = &ic->i_send_hdrs[pos];
 679
 680                         /* add credit and redo the header checksum */
 681                         hdr->h_credit = adv_credits;
 682                         rds_message_make_checksum(hdr);
 683                         adv_credits = 0;
 684                         rds_ib_stats_inc(s_ib_tx_credit_updates);
 685                 }
 686
 687                 if (prev)
 688                         prev->s_wr.next = &send->s_wr;
 689                 prev = send;
 690
 691                 pos = (pos + 1) % ic->i_send_ring.w_nr;
 692         }
 693
 694         /* Account the RDS header in the number of bytes we sent, but just once.
 695          * The caller has no concept of fragmentation. */
 696         if (hdr_off == 0)
 697                 sent += sizeof(struct rds_header);
 698
 699         /* if we finished the message then send completion owns it */
 700         if (scat == &rm->m_sg[rm->m_count]) {
 701                 prev->s_rm = ic->i_rm;
 702                 prev->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
 703                 ic->i_rm = NULL;
 704         }
 705
 706         if (i < work_alloc) {
 707                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
 708                 work_alloc = i;
 709         }
 710         if (ic->i_flowctl && i < credit_alloc)
 711                 rds_ib_send_add_credits(conn, credit_alloc - i);
 712
 713         /* XXX need to worry about failed_wr and partial sends. */
 714         failed_wr = &first->s_wr;
 715         ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
 716         rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
 717                  first, &first->s_wr, ret, failed_wr);
 718         BUG_ON(failed_wr != &first->s_wr);
 719         if (ret) {
 720                 printk(KERN_WARNING "RDS/IB: ib_post_send to %pI4 "
 721                        "returned %d\n", &conn->c_faddr, ret);
 722                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 723                 if (prev->s_rm) {
 724                         ic->i_rm = prev->s_rm;
 725                         prev->s_rm = NULL;
 726                 }
 727
 728                 rds_ib_conn_error(ic->conn, "ib_post_send failed\n");
 729                 goto out;
 730         }
 731
 732         ret = sent;
 733 out:
 734         BUG_ON(adv_credits);
 735         return ret;
 736 }
 737
 738 int rds_ib_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
 739 {
 740         struct rds_ib_connection *ic = conn->c_transport_data;
 741         struct rds_ib_send_work *send = NULL;
 742         struct rds_ib_send_work *first;
 743         struct rds_ib_send_work *prev;
 744         struct ib_send_wr *failed_wr;
 745         struct rds_ib_device *rds_ibdev;
 746         struct scatterlist *scat;
 747         unsigned long len;
 748         u64 remote_addr = op->r_remote_addr;
 749         u32 pos;
 750         u32 work_alloc;
 751         u32 i;
 752         u32 j;
 753         int sent;
 754         int ret;
 755         int num_sge;
 756
 757         rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
 758
 759         /* map the message the first time we see it */
 760         if (!op->r_mapped) {
 761                 op->r_count = ib_dma_map_sg(ic->i_cm_id->device,
 762                                         op->r_sg, op->r_nents, (op->r_write) ?
 763                                         DMA_TO_DEVICE : DMA_FROM_DEVICE);
 764                 rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->r_count);
 765                 if (op->r_count == 0) {
 766                         rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
 767                         ret = -ENOMEM; /* XXX ? */
 768                         goto out;
 769                 }
 770
 771                 op->r_mapped = 1;
 772         }
 773
 774         /*
 775          * Instead of knowing how to return a partial rdma read/write we insist that there
 776          * be enough work requests to send the entire message.
 777          */
 778         i = ceil(op->r_count, rds_ibdev->max_sge);
 779
 780         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
 781         if (work_alloc != i) {
 782                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 783                 rds_ib_stats_inc(s_ib_tx_ring_full);
 784                 ret = -ENOMEM;
 785                 goto out;
 786         }
 787
 788         send = &ic->i_sends[pos];
 789         first = send;
 790         prev = NULL;
 791         scat = &op->r_sg[0];
 792         sent = 0;
 793         num_sge = op->r_count;
 794
 795         for (i = 0; i < work_alloc && scat != &op->r_sg[op->r_count]; i++) {
 796                 send->s_wr.send_flags = 0;
 797                 send->s_queued = jiffies;
 798                 /*
 799                  * We want to delay signaling completions just enough to get
 800                  * the batching benefits but not so much that we create dead time on the wire.
 801                  */
 802                 if (ic->i_unsignaled_wrs-- == 0) {
 803                         ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
 804                         send->s_wr.send_flags = IB_SEND_SIGNALED;
 805                 }
 806
 807                 send->s_wr.opcode = op->r_write ? IB_WR_RDMA_WRITE : IB_WR_RDMA_READ;
 808                 send->s_wr.wr.rdma.remote_addr = remote_addr;
 809                 send->s_wr.wr.rdma.rkey = op->r_key;
 810                 send->s_op = op;
 811
 812                 if (num_sge > rds_ibdev->max_sge) {
 813                         send->s_wr.num_sge = rds_ibdev->max_sge;
 814                         num_sge -= rds_ibdev->max_sge;
 815                 } else {
 816                         send->s_wr.num_sge = num_sge;
 817                 }
 818
 819                 send->s_wr.next = NULL;
 820
 821                 if (prev)
 822                         prev->s_wr.next = &send->s_wr;
 823
 824                 for (j = 0; j < send->s_wr.num_sge && scat != &op->r_sg[op->r_count]; j++) {
 825                         len = ib_sg_dma_len(ic->i_cm_id->device, scat);
 826                         send->s_sge[j].addr =
 827                                  ib_sg_dma_address(ic->i_cm_id->device, scat);
 828                         send->s_sge[j].length = len;
 829                         send->s_sge[j].lkey = ic->i_mr->lkey;
 830
 831                         sent += len;
 832                         rdsdebug("ic %p sent %d remote_addr %llu\n", ic, sent, remote_addr);
 833
 834                         remote_addr += len;
 835                         scat++;
 836                 }
 837
 838                 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
 839                         &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
 840
 841                 prev = send;
 842                 if (++send == &ic->i_sends[ic->i_send_ring.w_nr])
 843                         send = ic->i_sends;
 844         }
 845
 846         /* if we finished the message then send completion owns it */
 847         if (scat == &op->r_sg[op->r_count])
 848                 prev->s_wr.send_flags = IB_SEND_SIGNALED;
 849
 850         if (i < work_alloc) {
 851                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
 852                 work_alloc = i;
 853         }
 854
 855         failed_wr = &first->s_wr;
 856         ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
 857         rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
 858                  first, &first->s_wr, ret, failed_wr);
 859         BUG_ON(failed_wr != &first->s_wr);
 860         if (ret) {
 861                 printk(KERN_WARNING "RDS/IB: rdma ib_post_send to %pI4 "
 862                        "returned %d\n", &conn->c_faddr, ret);
 863                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 864                 goto out;
 865         }
 866
 867         if (unlikely(failed_wr != &first->s_wr)) {
 868                 printk(KERN_WARNING "RDS/IB: ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
 869                 BUG_ON(failed_wr != &first->s_wr);
 870         }
 871
 872
 873 out:
 874         return ret;
 875 }
 876
 877 void rds_ib_xmit_complete(struct rds_connection *conn)
 878 {
 879         struct rds_ib_connection *ic = conn->c_transport_data;
 880
 881         /* We may have a pending ACK or window update we were unable
 882          * to send previously (due to flow control). Try again. */
 883         rds_ib_attempt_ack(ic);
 884 }