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 "ib.h"
  40
  41 /*
  42  * Convert IB-specific error message to RDS error message and call core
  43  * completion handler.
  44  */
  45 static void rds_ib_send_complete(struct rds_message *rm,
  46                                  int wc_status,
  47                                  void (*complete)(struct rds_message *rm, int status))
  48 {
  49         int notify_status;
  50
  51         switch (wc_status) {
  52         case IB_WC_WR_FLUSH_ERR:
  53                 return;
  54
  55         case IB_WC_SUCCESS:
  56                 notify_status = RDS_RDMA_SUCCESS;
  57                 break;
  58
  59         case IB_WC_REM_ACCESS_ERR:
  60                 notify_status = RDS_RDMA_REMOTE_ERROR;
  61                 break;
  62
  63         default:
  64                 notify_status = RDS_RDMA_OTHER_ERROR;
  65                 break;
  66         }
  67         complete(rm, notify_status);
  68 }
  69
  70 static void rds_ib_send_unmap_rm(struct rds_ib_connection *ic,
  71                           struct rds_ib_send_work *send,
  72                           int wc_status)
  73 {
  74         struct rds_message *rm = send->s_rm;
  75
  76         rdsdebug("ic %p send %p rm %p\n", ic, send, rm);
  77
  78         ib_dma_unmap_sg(ic->i_cm_id->device,
  79                         rm->data.m_sg, rm->data.m_nents,
  80                         DMA_TO_DEVICE);
  81
  82         if (rm->rdma.m_rdma_op.r_active) {
  83                 struct rds_rdma_op *op = &rm->rdma.m_rdma_op;
  84
  85                 if (op->r_mapped) {
  86                         ib_dma_unmap_sg(ic->i_cm_id->device,
  87                                         op->r_sg, op->r_nents,
  88                                         op->r_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
  89                         op->r_mapped = 0;
  90                 }
  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_complete(rm, wc_status, rds_rdma_send_complete);
 113
 114                 if (rm->rdma.m_rdma_op.r_write)
 115                         rds_stats_add(s_send_rdma_bytes, rm->rdma.m_rdma_op.r_bytes);
 116                 else
 117                         rds_stats_add(s_recv_rdma_bytes, rm->rdma.m_rdma_op.r_bytes);
 118         }
 119
 120         if (rm->atomic.op_active) {
 121                 struct rm_atomic_op *op = &rm->atomic;
 122
 123                 /* unmap atomic recvbuf */
 124                 if (op->op_mapped) {
 125                         ib_dma_unmap_sg(ic->i_cm_id->device, op->op_sg, 1,
 126                                         DMA_FROM_DEVICE);
 127                         op->op_mapped = 0;
 128                 }
 129
 130                 rds_ib_send_complete(rm, wc_status, rds_atomic_send_complete);
 131
 132                 if (rm->atomic.op_type == RDS_ATOMIC_TYPE_CSWP)
 133                         rds_stats_inc(s_atomic_cswp);
 134                 else
 135                         rds_stats_inc(s_atomic_fadd);
 136         }
 137
 138         /* If anyone waited for this message to get flushed out, wake
 139          * them up now */
 140         rds_message_unmapped(rm);
 141
 142         rds_message_put(rm);
 143         send->s_rm = NULL;
 144 }
 145
 146 void rds_ib_send_init_ring(struct rds_ib_connection *ic)
 147 {
 148         struct rds_ib_send_work *send;
 149         u32 i;
 150
 151         for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
 152                 struct ib_sge *sge;
 153
 154                 send->s_rm = NULL;
 155                 send->s_op = NULL;
 156
 157                 send->s_wr.wr_id = i;
 158                 send->s_wr.sg_list = send->s_sge;
 159                 send->s_wr.ex.imm_data = 0;
 160
 161                 sge = &send->s_sge[0];
 162                 sge->addr = ic->i_send_hdrs_dma + (i * sizeof(struct rds_header));
 163                 sge->length = sizeof(struct rds_header);
 164                 sge->lkey = ic->i_mr->lkey;
 165
 166                 send->s_sge[1].lkey = ic->i_mr->lkey;
 167         }
 168 }
 169
 170 void rds_ib_send_clear_ring(struct rds_ib_connection *ic)
 171 {
 172         struct rds_ib_send_work *send;
 173         u32 i;
 174
 175         for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
 176                 if (!send->s_rm || send->s_wr.opcode == 0xdead)
 177                         continue;
 178                 rds_ib_send_unmap_rm(ic, send, IB_WC_WR_FLUSH_ERR);
 179         }
 180 }
 181
 182 /*
 183  * The _oldest/_free ring operations here race cleanly with the alloc/unalloc
 184  * operations performed in the send path.  As the sender allocs and potentially
 185  * unallocs the next free entry in the ring it doesn't alter which is
 186  * the next to be freed, which is what this is concerned with.
 187  */
 188 void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context)
 189 {
 190         struct rds_connection *conn = context;
 191         struct rds_ib_connection *ic = conn->c_transport_data;
 192         struct ib_wc wc;
 193         struct rds_ib_send_work *send;
 194         u32 completed;
 195         u32 oldest;
 196         u32 i = 0;
 197         int ret;
 198
 199         rdsdebug("cq %p conn %p\n", cq, conn);
 200         rds_ib_stats_inc(s_ib_tx_cq_call);
 201         ret = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 202         if (ret)
 203                 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
 204
 205         while (ib_poll_cq(cq, 1, &wc) > 0) {
 206                 rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
 207                          (unsigned long long)wc.wr_id, wc.status, wc.byte_len,
 208                          be32_to_cpu(wc.ex.imm_data));
 209                 rds_ib_stats_inc(s_ib_tx_cq_event);
 210
 211                 if (wc.wr_id == RDS_IB_ACK_WR_ID) {
 212                         if (ic->i_ack_queued + HZ/2 < jiffies)
 213                                 rds_ib_stats_inc(s_ib_tx_stalled);
 214                         rds_ib_ack_send_complete(ic);
 215                         continue;
 216                 }
 217
 218                 oldest = rds_ib_ring_oldest(&ic->i_send_ring);
 219
 220                 completed = rds_ib_ring_completed(&ic->i_send_ring, wc.wr_id, oldest);
 221
 222                 for (i = 0; i < completed; i++) {
 223                         send = &ic->i_sends[oldest];
 224
 225                         /* In the error case, wc.opcode sometimes contains garbage */
 226                         switch (send->s_wr.opcode) {
 227                         case IB_WR_SEND:
 228                                 if (send->s_rm)
 229                                         rds_ib_send_unmap_rm(ic, send, wc.status);
 230                                 break;
 231                         case IB_WR_RDMA_WRITE:
 232                         case IB_WR_RDMA_READ:
 233                         case IB_WR_ATOMIC_FETCH_AND_ADD:
 234                         case IB_WR_ATOMIC_CMP_AND_SWP:
 235                                 /* Nothing to be done - the SG list will be unmapped
 236                                  * when the SEND completes. */
 237                                 break;
 238                         default:
 239                                 if (printk_ratelimit())
 240                                         printk(KERN_NOTICE
 241                                                 "RDS/IB: %s: unexpected opcode 0x%x in WR!\n",
 242                                                 __func__, send->s_wr.opcode);
 243                                 break;
 244                         }
 245
 246                         send->s_wr.opcode = 0xdead;
 247                         send->s_wr.num_sge = 1;
 248                         if (send->s_queued + HZ/2 < jiffies)
 249                                 rds_ib_stats_inc(s_ib_tx_stalled);
 250
 251                         /* If a RDMA operation produced an error, signal this right
 252                          * away. If we don't, the subsequent SEND that goes with this
 253                          * RDMA will be canceled with ERR_WFLUSH, and the application
 254                          * never learn that the RDMA failed. */
 255                         if (unlikely(wc.status == IB_WC_REM_ACCESS_ERR && send->s_op)) {
 256                                 struct rds_message *rm;
 257
 258                                 rm = rds_send_get_message(conn, send->s_op);
 259                                 if (rm) {
 260                                         rds_ib_send_unmap_rm(ic, send, wc.status);
 261                                         rds_ib_send_complete(rm, wc.status, rds_rdma_send_complete);
 262                                         rds_message_put(rm);
 263                                 }
 264                         }
 265
 266                         oldest = (oldest + 1) % ic->i_send_ring.w_nr;
 267                 }
 268
 269                 rds_ib_ring_free(&ic->i_send_ring, completed);
 270
 271                 if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
 272                     test_bit(0, &conn->c_map_queued))
 273                         queue_delayed_work(rds_wq, &conn->c_send_w, 0);
 274
 275                 /* We expect errors as the qp is drained during shutdown */
 276                 if (wc.status != IB_WC_SUCCESS && rds_conn_up(conn)) {
 277                         rds_ib_conn_error(conn,
 278                                 "send completion on %pI4 "
 279                                 "had status %u, disconnecting and reconnecting\n",
 280                                 &conn->c_faddr, wc.status);
 281                 }
 282         }
 283 }
 284
 285 /*
 286  * This is the main function for allocating credits when sending
 287  * messages.
 288  *
 289  * Conceptually, we have two counters:
 290  *  -   send credits: this tells us how many WRs we're allowed
 291  *      to submit without overruning the reciever's queue. For
 292  *      each SEND WR we post, we decrement this by one.
 293  *
 294  *  -   posted credits: this tells us how many WRs we recently
 295  *      posted to the receive queue. This value is transferred
 296  *      to the peer as a "credit update" in a RDS header field.
 297  *      Every time we transmit credits to the peer, we subtract
 298  *      the amount of transferred credits from this counter.
 299  *
 300  * It is essential that we avoid situations where both sides have
 301  * exhausted their send credits, and are unable to send new credits
 302  * to the peer. We achieve this by requiring that we send at least
 303  * one credit update to the peer before exhausting our credits.
 304  * When new credits arrive, we subtract one credit that is withheld
 305  * until we've posted new buffers and are ready to transmit these
 306  * credits (see rds_ib_send_add_credits below).
 307  *
 308  * The RDS send code is essentially single-threaded; rds_send_xmit
 309  * grabs c_send_lock to ensure exclusive access to the send ring.
 310  * However, the ACK sending code is independent and can race with
 311  * message SENDs.
 312  *
 313  * In the send path, we need to update the counters for send credits
 314  * and the counter of posted buffers atomically - when we use the
 315  * last available credit, we cannot allow another thread to race us
 316  * and grab the posted credits counter.  Hence, we have to use a
 317  * spinlock to protect the credit counter, or use atomics.
 318  *
 319  * Spinlocks shared between the send and the receive path are bad,
 320  * because they create unnecessary delays. An early implementation
 321  * using a spinlock showed a 5% degradation in throughput at some
 322  * loads.
 323  *
 324  * This implementation avoids spinlocks completely, putting both
 325  * counters into a single atomic, and updating that atomic using
 326  * atomic_add (in the receive path, when receiving fresh credits),
 327  * and using atomic_cmpxchg when updating the two counters.
 328  */
 329 int rds_ib_send_grab_credits(struct rds_ib_connection *ic,
 330                              u32 wanted, u32 *adv_credits, int need_posted, int max_posted)
 331 {
 332         unsigned int avail, posted, got = 0, advertise;
 333         long oldval, newval;
 334
 335         *adv_credits = 0;
 336         if (!ic->i_flowctl)
 337                 return wanted;
 338
 339 try_again:
 340         advertise = 0;
 341         oldval = newval = atomic_read(&ic->i_credits);
 342         posted = IB_GET_POST_CREDITS(oldval);
 343         avail = IB_GET_SEND_CREDITS(oldval);
 344
 345         rdsdebug("rds_ib_send_grab_credits(%u): credits=%u posted=%u\n",
 346                         wanted, avail, posted);
 347
 348         /* The last credit must be used to send a credit update. */
 349         if (avail && !posted)
 350                 avail--;
 351
 352         if (avail < wanted) {
 353                 struct rds_connection *conn = ic->i_cm_id->context;
 354
 355                 /* Oops, there aren't that many credits left! */
 356                 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 357                 got = avail;
 358         } else {
 359                 /* Sometimes you get what you want, lalala. */
 360                 got = wanted;
 361         }
 362         newval -= IB_SET_SEND_CREDITS(got);
 363
 364         /*
 365          * If need_posted is non-zero, then the caller wants
 366          * the posted regardless of whether any send credits are
 367          * available.
 368          */
 369         if (posted && (got || need_posted)) {
 370                 advertise = min_t(unsigned int, posted, max_posted);
 371                 newval -= IB_SET_POST_CREDITS(advertise);
 372         }
 373
 374         /* Finally bill everything */
 375         if (atomic_cmpxchg(&ic->i_credits, oldval, newval) != oldval)
 376                 goto try_again;
 377
 378         *adv_credits = advertise;
 379         return got;
 380 }
 381
 382 void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits)
 383 {
 384         struct rds_ib_connection *ic = conn->c_transport_data;
 385
 386         if (credits == 0)
 387                 return;
 388
 389         rdsdebug("rds_ib_send_add_credits(%u): current=%u%s\n",
 390                         credits,
 391                         IB_GET_SEND_CREDITS(atomic_read(&ic->i_credits)),
 392                         test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ? ", ll_send_full" : "");
 393
 394         atomic_add(IB_SET_SEND_CREDITS(credits), &ic->i_credits);
 395         if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags))
 396                 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
 397
 398         WARN_ON(IB_GET_SEND_CREDITS(credits) >= 16384);
 399
 400         rds_ib_stats_inc(s_ib_rx_credit_updates);
 401 }
 402
 403 void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted)
 404 {
 405         struct rds_ib_connection *ic = conn->c_transport_data;
 406
 407         if (posted == 0)
 408                 return;
 409
 410         atomic_add(IB_SET_POST_CREDITS(posted), &ic->i_credits);
 411
 412         /* Decide whether to send an update to the peer now.
 413          * If we would send a credit update for every single buffer we
 414          * post, we would end up with an ACK storm (ACK arrives,
 415          * consumes buffer, we refill the ring, send ACK to remote
 416          * advertising the newly posted buffer... ad inf)
 417          *
 418          * Performance pretty much depends on how often we send
 419          * credit updates - too frequent updates mean lots of ACKs.
 420          * Too infrequent updates, and the peer will run out of
 421          * credits and has to throttle.
 422          * For the time being, 16 seems to be a good compromise.
 423          */
 424         if (IB_GET_POST_CREDITS(atomic_read(&ic->i_credits)) >= 16)
 425                 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
 426 }
 427
 428 static inline void
 429 rds_ib_xmit_populate_wr(struct rds_ib_connection *ic,
 430                 struct rds_ib_send_work *send, unsigned int pos,
 431                 unsigned long buffer, unsigned int length,
 432                 int send_flags)
 433 {
 434         struct ib_sge *sge;
 435
 436         WARN_ON(pos != send - ic->i_sends);
 437
 438         send->s_wr.send_flags = send_flags;
 439         send->s_wr.opcode = IB_WR_SEND;
 440         send->s_wr.num_sge = 1;
 441         send->s_wr.next = NULL;
 442         send->s_queued = jiffies;
 443         send->s_op = NULL;
 444
 445         sge = &send->s_sge[0];
 446         sge->addr = ic->i_send_hdrs_dma + (pos * sizeof(struct rds_header));
 447         sge->length = sizeof(struct rds_header);
 448         sge->lkey = ic->i_mr->lkey;
 449
 450         if (length != 0) {
 451                 send->s_wr.num_sge = 2;
 452
 453                 sge = &send->s_sge[1];
 454                 sge->addr = buffer;
 455                 sge->length = length;
 456                 sge->lkey = ic->i_mr->lkey;
 457         }
 458 }
 459
 460 /*
 461  * This can be called multiple times for a given message.  The first time
 462  * we see a message we map its scatterlist into the IB device so that
 463  * we can provide that mapped address to the IB scatter gather entries
 464  * in the IB work requests.  We translate the scatterlist into a series
 465  * of work requests that fragment the message.  These work requests complete
 466  * in order so we pass ownership of the message to the completion handler
 467  * once we send the final fragment.
 468  *
 469  * The RDS core uses the c_send_lock to only enter this function once
 470  * per connection.  This makes sure that the tx ring alloc/unalloc pairs
 471  * don't get out of sync and confuse the ring.
 472  */
 473 int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
 474                 unsigned int hdr_off, unsigned int sg, unsigned int off)
 475 {
 476         struct rds_ib_connection *ic = conn->c_transport_data;
 477         struct ib_device *dev = ic->i_cm_id->device;
 478         struct rds_ib_send_work *send = NULL;
 479         struct rds_ib_send_work *first;
 480         struct rds_ib_send_work *prev;
 481         struct ib_send_wr *failed_wr;
 482         struct scatterlist *scat;
 483         u32 pos;
 484         u32 i;
 485         u32 work_alloc;
 486         u32 credit_alloc;
 487         u32 posted;
 488         u32 adv_credits = 0;
 489         int send_flags = 0;
 490         int sent;
 491         int ret;
 492         int flow_controlled = 0;
 493
 494         BUG_ON(off % RDS_FRAG_SIZE);
 495         BUG_ON(hdr_off != 0 && hdr_off != sizeof(struct rds_header));
 496
 497         /* Do not send cong updates to IB loopback */
 498         if (conn->c_loopback
 499             && rm->m_inc.i_hdr.h_flags & RDS_FLAG_CONG_BITMAP) {
 500                 rds_cong_map_updated(conn->c_fcong, ~(u64) 0);
 501                 return sizeof(struct rds_header) + RDS_CONG_MAP_BYTES;
 502         }
 503
 504         /* FIXME we may overallocate here */
 505         if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0)
 506                 i = 1;
 507         else
 508                 i = ceil(be32_to_cpu(rm->m_inc.i_hdr.h_len), RDS_FRAG_SIZE);
 509
 510         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
 511         if (work_alloc == 0) {
 512                 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 513                 rds_ib_stats_inc(s_ib_tx_ring_full);
 514                 ret = -ENOMEM;
 515                 goto out;
 516         }
 517
 518         credit_alloc = work_alloc;
 519         if (ic->i_flowctl) {
 520                 credit_alloc = rds_ib_send_grab_credits(ic, work_alloc, &posted, 0, RDS_MAX_ADV_CREDIT);
 521                 adv_credits += posted;
 522                 if (credit_alloc < work_alloc) {
 523                         rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - credit_alloc);
 524                         work_alloc = credit_alloc;
 525                         flow_controlled++;
 526                 }
 527                 if (work_alloc == 0) {
 528                         set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 529                         rds_ib_stats_inc(s_ib_tx_throttle);
 530                         ret = -ENOMEM;
 531                         goto out;
 532                 }
 533         }
 534
 535         /* map the message the first time we see it */
 536         if (!ic->i_rm) {
 537                 if (rm->data.m_nents) {
 538                         rm->data.m_count = ib_dma_map_sg(dev,
 539                                                             rm->data.m_sg,
 540                                                             rm->data.m_nents,
 541                                                             DMA_TO_DEVICE);
 542                         rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->data.m_count);
 543                         if (rm->data.m_count == 0) {
 544                                 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
 545                                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 546                                 ret = -ENOMEM; /* XXX ? */
 547                                 goto out;
 548                         }
 549                 } else {
 550                         rm->data.m_count = 0;
 551                 }
 552
 553                 ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
 554                 ic->i_unsignaled_bytes = rds_ib_sysctl_max_unsig_bytes;
 555                 rds_message_addref(rm);
 556                 ic->i_rm = rm;
 557
 558                 /* Finalize the header */
 559                 if (test_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags))
 560                         rm->m_inc.i_hdr.h_flags |= RDS_FLAG_ACK_REQUIRED;
 561                 if (test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))
 562                         rm->m_inc.i_hdr.h_flags |= RDS_FLAG_RETRANSMITTED;
 563
 564                 /* If it has a RDMA op, tell the peer we did it. This is
 565                  * used by the peer to release use-once RDMA MRs. */
 566                 if (rm->rdma.m_rdma_op.r_active) {
 567                         struct rds_ext_header_rdma ext_hdr;
 568
 569                         ext_hdr.h_rdma_rkey = cpu_to_be32(rm->rdma.m_rdma_op.r_key);
 570                         rds_message_add_extension(&rm->m_inc.i_hdr,
 571                                         RDS_EXTHDR_RDMA, &ext_hdr, sizeof(ext_hdr));
 572                 }
 573                 if (rm->m_rdma_cookie) {
 574                         rds_message_add_rdma_dest_extension(&rm->m_inc.i_hdr,
 575                                         rds_rdma_cookie_key(rm->m_rdma_cookie),
 576                                         rds_rdma_cookie_offset(rm->m_rdma_cookie));
 577                 }
 578
 579                 /* Note - rds_ib_piggyb_ack clears the ACK_REQUIRED bit, so
 580                  * we should not do this unless we have a chance of at least
 581                  * sticking the header into the send ring. Which is why we
 582                  * should call rds_ib_ring_alloc first. */
 583                 rm->m_inc.i_hdr.h_ack = cpu_to_be64(rds_ib_piggyb_ack(ic));
 584                 rds_message_make_checksum(&rm->m_inc.i_hdr);
 585
 586                 /*
 587                  * Update adv_credits since we reset the ACK_REQUIRED bit.
 588                  */
 589                 rds_ib_send_grab_credits(ic, 0, &posted, 1, RDS_MAX_ADV_CREDIT - adv_credits);
 590                 adv_credits += posted;
 591                 BUG_ON(adv_credits > 255);
 592         }
 593
 594         send = &ic->i_sends[pos];
 595         first = send;
 596         prev = NULL;
 597         scat = &rm->data.m_sg[sg];
 598         sent = 0;
 599         i = 0;
 600
 601         /* Sometimes you want to put a fence between an RDMA
 602          * READ and the following SEND.
 603          * We could either do this all the time
 604          * or when requested by the user. Right now, we let
 605          * the application choose.
 606          */
 607         if (rm->rdma.m_rdma_op.r_active && rm->rdma.m_rdma_op.r_fence)
 608                 send_flags = IB_SEND_FENCE;
 609
 610         /*
 611          * We could be copying the header into the unused tail of the page.
 612          * That would need to be changed in the future when those pages might
 613          * be mapped userspace pages or page cache pages.  So instead we always
 614          * use a second sge and our long-lived ring of mapped headers.  We send
 615          * the header after the data so that the data payload can be aligned on
 616          * the receiver.
 617          */
 618
 619         /* handle a 0-len message */
 620         if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0) {
 621                 rds_ib_xmit_populate_wr(ic, send, pos, 0, 0, send_flags);
 622                 goto add_header;
 623         }
 624
 625         /* if there's data reference it with a chain of work reqs */
 626         for (; i < work_alloc && scat != &rm->data.m_sg[rm->data.m_count]; i++) {
 627                 unsigned int len;
 628
 629                 send = &ic->i_sends[pos];
 630
 631                 len = min(RDS_FRAG_SIZE, ib_sg_dma_len(dev, scat) - off);
 632                 rds_ib_xmit_populate_wr(ic, send, pos,
 633                                 ib_sg_dma_address(dev, scat) + off, len,
 634                                 send_flags);
 635
 636                 /*
 637                  * We want to delay signaling completions just enough to get
 638                  * the batching benefits but not so much that we create dead time
 639                  * on the wire.
 640                  */
 641                 if (ic->i_unsignaled_wrs-- == 0) {
 642                         ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
 643                         send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
 644                 }
 645
 646                 ic->i_unsignaled_bytes -= len;
 647                 if (ic->i_unsignaled_bytes <= 0) {
 648                         ic->i_unsignaled_bytes = rds_ib_sysctl_max_unsig_bytes;
 649                         send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
 650                 }
 651
 652                 /*
 653                  * Always signal the last one if we're stopping due to flow control.
 654                  */
 655                 if (flow_controlled && i == (work_alloc-1))
 656                         send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
 657
 658                 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
 659                          &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
 660
 661                 sent += len;
 662                 off += len;
 663                 if (off == ib_sg_dma_len(dev, scat)) {
 664                         scat++;
 665                         off = 0;
 666                 }
 667
 668 add_header:
 669                 /* Tack on the header after the data. The header SGE should already
 670                  * have been set up to point to the right header buffer. */
 671                 memcpy(&ic->i_send_hdrs[pos], &rm->m_inc.i_hdr, sizeof(struct rds_header));
 672
 673                 if (adv_credits) {
 674                         struct rds_header *hdr = &ic->i_send_hdrs[pos];
 675
 676                         /* add credit and redo the header checksum */
 677                         hdr->h_credit = adv_credits;
 678                         rds_message_make_checksum(hdr);
 679                         adv_credits = 0;
 680                         rds_ib_stats_inc(s_ib_tx_credit_updates);
 681                 }
 682
 683                 if (prev)
 684                         prev->s_wr.next = &send->s_wr;
 685                 prev = send;
 686
 687                 pos = (pos + 1) % ic->i_send_ring.w_nr;
 688         }
 689
 690         /* Account the RDS header in the number of bytes we sent, but just once.
 691          * The caller has no concept of fragmentation. */
 692         if (hdr_off == 0)
 693                 sent += sizeof(struct rds_header);
 694
 695         /* if we finished the message then send completion owns it */
 696         if (scat == &rm->data.m_sg[rm->data.m_count]) {
 697                 prev->s_rm = ic->i_rm;
 698                 prev->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
 699                 ic->i_rm = NULL;
 700         }
 701
 702         if (i < work_alloc) {
 703                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
 704                 work_alloc = i;
 705         }
 706         if (ic->i_flowctl && i < credit_alloc)
 707                 rds_ib_send_add_credits(conn, credit_alloc - i);
 708
 709         /* XXX need to worry about failed_wr and partial sends. */
 710         failed_wr = &first->s_wr;
 711         ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
 712         rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
 713                  first, &first->s_wr, ret, failed_wr);
 714         BUG_ON(failed_wr != &first->s_wr);
 715         if (ret) {
 716                 printk(KERN_WARNING "RDS/IB: ib_post_send to %pI4 "
 717                        "returned %d\n", &conn->c_faddr, ret);
 718                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 719                 if (prev->s_rm) {
 720                         ic->i_rm = prev->s_rm;
 721                         prev->s_rm = NULL;
 722                 }
 723
 724                 rds_ib_conn_error(ic->conn, "ib_post_send failed\n");
 725                 goto out;
 726         }
 727
 728         ret = sent;
 729 out:
 730         BUG_ON(adv_credits);
 731         return ret;
 732 }
 733
 734 /*
 735  * Issue atomic operation.
 736  * A simplified version of the rdma case, we always map 1 SG, and
 737  * only 8 bytes, for the return value from the atomic operation.
 738  */
 739 int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op)
 740 {
 741         struct rds_ib_connection *ic = conn->c_transport_data;
 742         struct rds_ib_send_work *send = NULL;
 743         struct ib_send_wr *failed_wr;
 744         struct rds_ib_device *rds_ibdev;
 745         u32 pos;
 746         u32 work_alloc;
 747         int ret;
 748
 749         rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
 750
 751         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, 1, &pos);
 752         if (work_alloc != 1) {
 753                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 754                 rds_ib_stats_inc(s_ib_tx_ring_full);
 755                 ret = -ENOMEM;
 756                 goto out;
 757         }
 758
 759         /* address of send request in ring */
 760         send = &ic->i_sends[pos];
 761         send->s_queued = jiffies;
 762
 763         if (op->op_type == RDS_ATOMIC_TYPE_CSWP) {
 764                 send->s_wr.opcode = IB_WR_ATOMIC_CMP_AND_SWP;
 765                 send->s_wr.wr.atomic.compare_add = op->op_compare;
 766                 send->s_wr.wr.atomic.swap = op->op_swap_add;
 767         } else { /* FADD */
 768                 send->s_wr.opcode = IB_WR_ATOMIC_FETCH_AND_ADD;
 769                 send->s_wr.wr.atomic.compare_add = op->op_swap_add;
 770                 send->s_wr.wr.atomic.swap = 0;
 771         }
 772         send->s_wr.send_flags = IB_SEND_SIGNALED;
 773         send->s_wr.num_sge = 1;
 774         send->s_wr.next = NULL;
 775         send->s_wr.wr.atomic.remote_addr = op->op_remote_addr;
 776         send->s_wr.wr.atomic.rkey = op->op_rkey;
 777
 778         /* map 8 byte retval buffer to the device */
 779         ret = ib_dma_map_sg(ic->i_cm_id->device, op->op_sg, 1, DMA_FROM_DEVICE);
 780         rdsdebug("ic %p mapping atomic op %p. mapped %d pg\n", ic, op, ret);
 781         if (ret != 1) {
 782                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 783                 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
 784                 ret = -ENOMEM; /* XXX ? */
 785                 goto out;
 786         }
 787
 788         /* Convert our struct scatterlist to struct ib_sge */
 789         send->s_sge[0].addr = ib_sg_dma_address(ic->i_cm_id->device, op->op_sg);
 790         send->s_sge[0].length = ib_sg_dma_len(ic->i_cm_id->device, op->op_sg);
 791         send->s_sge[0].lkey = ic->i_mr->lkey;
 792
 793         rdsdebug("rva %Lx rpa %Lx len %u\n", op->op_remote_addr,
 794                  send->s_sge[0].addr, send->s_sge[0].length);
 795
 796         failed_wr = &send->s_wr;
 797         ret = ib_post_send(ic->i_cm_id->qp, &send->s_wr, &failed_wr);
 798         rdsdebug("ic %p send %p (wr %p) ret %d wr %p\n", ic,
 799                  send, &send->s_wr, ret, failed_wr);
 800         BUG_ON(failed_wr != &send->s_wr);
 801         if (ret) {
 802                 printk(KERN_WARNING "RDS/IB: atomic ib_post_send to %pI4 "
 803                        "returned %d\n", &conn->c_faddr, ret);
 804                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 805                 goto out;
 806         }
 807
 808         if (unlikely(failed_wr != &send->s_wr)) {
 809                 printk(KERN_WARNING "RDS/IB: atomic ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
 810                 BUG_ON(failed_wr != &send->s_wr);
 811         }
 812
 813 out:
 814         return ret;
 815 }
 816
 817 int rds_ib_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
 818 {
 819         struct rds_ib_connection *ic = conn->c_transport_data;
 820         struct rds_ib_send_work *send = NULL;
 821         struct rds_ib_send_work *first;
 822         struct rds_ib_send_work *prev;
 823         struct ib_send_wr *failed_wr;
 824         struct rds_ib_device *rds_ibdev;
 825         struct scatterlist *scat;
 826         unsigned long len;
 827         u64 remote_addr = op->r_remote_addr;
 828         u32 pos;
 829         u32 work_alloc;
 830         u32 i;
 831         u32 j;
 832         int sent;
 833         int ret;
 834         int num_sge;
 835
 836         rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
 837
 838         /* map the message the first time we see it */
 839         if (!op->r_mapped) {
 840                 op->r_count = ib_dma_map_sg(ic->i_cm_id->device,
 841                                         op->r_sg, op->r_nents, (op->r_write) ?
 842                                         DMA_TO_DEVICE : DMA_FROM_DEVICE);
 843                 rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->r_count);
 844                 if (op->r_count == 0) {
 845                         rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
 846                         ret = -ENOMEM; /* XXX ? */
 847                         goto out;
 848                 }
 849
 850                 op->r_mapped = 1;
 851         }
 852
 853         /*
 854          * Instead of knowing how to return a partial rdma read/write we insist that there
 855          * be enough work requests to send the entire message.
 856          */
 857         i = ceil(op->r_count, rds_ibdev->max_sge);
 858
 859         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
 860         if (work_alloc != i) {
 861                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 862                 rds_ib_stats_inc(s_ib_tx_ring_full);
 863                 ret = -ENOMEM;
 864                 goto out;
 865         }
 866
 867         send = &ic->i_sends[pos];
 868         first = send;
 869         prev = NULL;
 870         scat = &op->r_sg[0];
 871         sent = 0;
 872         num_sge = op->r_count;
 873
 874         for (i = 0; i < work_alloc && scat != &op->r_sg[op->r_count]; i++) {
 875                 send->s_wr.send_flags = 0;
 876                 send->s_queued = jiffies;
 877                 /*
 878                  * We want to delay signaling completions just enough to get
 879                  * the batching benefits but not so much that we create dead time on the wire.
 880                  */
 881                 if (ic->i_unsignaled_wrs-- == 0) {
 882                         ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
 883                         send->s_wr.send_flags = IB_SEND_SIGNALED;
 884                 }
 885
 886                 send->s_wr.opcode = op->r_write ? IB_WR_RDMA_WRITE : IB_WR_RDMA_READ;
 887                 send->s_wr.wr.rdma.remote_addr = remote_addr;
 888                 send->s_wr.wr.rdma.rkey = op->r_key;
 889                 send->s_op = op;
 890
 891                 if (num_sge > rds_ibdev->max_sge) {
 892                         send->s_wr.num_sge = rds_ibdev->max_sge;
 893                         num_sge -= rds_ibdev->max_sge;
 894                 } else {
 895                         send->s_wr.num_sge = num_sge;
 896                 }
 897
 898                 send->s_wr.next = NULL;
 899
 900                 if (prev)
 901                         prev->s_wr.next = &send->s_wr;
 902
 903                 for (j = 0; j < send->s_wr.num_sge && scat != &op->r_sg[op->r_count]; j++) {
 904                         len = ib_sg_dma_len(ic->i_cm_id->device, scat);
 905                         send->s_sge[j].addr =
 906                                  ib_sg_dma_address(ic->i_cm_id->device, scat);
 907                         send->s_sge[j].length = len;
 908                         send->s_sge[j].lkey = ic->i_mr->lkey;
 909
 910                         sent += len;
 911                         rdsdebug("ic %p sent %d remote_addr %llu\n", ic, sent, remote_addr);
 912
 913                         remote_addr += len;
 914                         scat++;
 915                 }
 916
 917                 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
 918                         &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
 919
 920                 prev = send;
 921                 if (++send == &ic->i_sends[ic->i_send_ring.w_nr])
 922                         send = ic->i_sends;
 923         }
 924
 925         /* if we finished the message then send completion owns it */
 926         if (scat == &op->r_sg[op->r_count])
 927                 prev->s_wr.send_flags = IB_SEND_SIGNALED;
 928
 929         if (i < work_alloc) {
 930                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
 931                 work_alloc = i;
 932         }
 933
 934         failed_wr = &first->s_wr;
 935         ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
 936         rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
 937                  first, &first->s_wr, ret, failed_wr);
 938         BUG_ON(failed_wr != &first->s_wr);
 939         if (ret) {
 940                 printk(KERN_WARNING "RDS/IB: rdma ib_post_send to %pI4 "
 941                        "returned %d\n", &conn->c_faddr, ret);
 942                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 943                 goto out;
 944         }
 945
 946         if (unlikely(failed_wr != &first->s_wr)) {
 947                 printk(KERN_WARNING "RDS/IB: ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
 948                 BUG_ON(failed_wr != &first->s_wr);
 949         }
 950
 951
 952 out:
 953         return ret;
 954 }
 955
 956 void rds_ib_xmit_complete(struct rds_connection *conn)
 957 {
 958         struct rds_ib_connection *ic = conn->c_transport_data;
 959
 960         /* We may have a pending ACK or window update we were unable
 961          * to send previously (due to flow control). Try again. */
 962         rds_ib_attempt_ack(ic);
 963 }