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_data(struct rds_ib_connection *ic,
  71                                    struct rm_data_op *op,
  72                                    int wc_status)
  73 {
  74         if (op->op_nents)
  75                 ib_dma_unmap_sg(ic->i_cm_id->device,
  76                                 op->op_sg, op->op_nents,
  77                                 DMA_TO_DEVICE);
  78 }
  79
  80 static void rds_ib_send_unmap_rdma(struct rds_ib_connection *ic,
  81                                    struct rm_rdma_op *op,
  82                                    int wc_status)
  83 {
  84         if (op->op_mapped) {
  85                 ib_dma_unmap_sg(ic->i_cm_id->device,
  86                                 op->op_sg, op->op_nents,
  87                                 op->op_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
  88                 op->op_mapped = 0;
  89         }
  90
  91         /* If the user asked for a completion notification on this
  92          * message, we can implement three different semantics:
  93          *  1.  Notify when we received the ACK on the RDS message
  94          *      that was queued with the RDMA. This provides reliable
  95          *      notification of RDMA status at the expense of a one-way
  96          *      packet delay.
  97          *  2.  Notify when the IB stack gives us the completion event for
  98          *      the RDMA operation.
  99          *  3.  Notify when the IB stack gives us the completion event for
 100          *      the accompanying RDS messages.
 101          * Here, we implement approach #3. To implement approach #2,
 102          * we would need to take an event for the rdma WR. To implement #1,
 103          * don't call rds_rdma_send_complete at all, and fall back to the notify
 104          * handling in the ACK processing code.
 105          *
 106          * Note: There's no need to explicitly sync any RDMA buffers using
 107          * ib_dma_sync_sg_for_cpu - the completion for the RDMA
 108          * operation itself unmapped the RDMA buffers, which takes care
 109          * of synching.
 110          */
 111         rds_ib_send_complete(container_of(op, struct rds_message, rdma),
 112                              wc_status, rds_rdma_send_complete);
 113
 114         if (op->op_write)
 115                 rds_stats_add(s_send_rdma_bytes, op->op_bytes);
 116         else
 117                 rds_stats_add(s_recv_rdma_bytes, op->op_bytes);
 118 }
 119
 120 static void rds_ib_send_unmap_atomic(struct rds_ib_connection *ic,
 121                                      struct rm_atomic_op *op,
 122                                      int wc_status)
 123 {
 124         /* unmap atomic recvbuf */
 125         if (op->op_mapped) {
 126                 ib_dma_unmap_sg(ic->i_cm_id->device, op->op_sg, 1,
 127                                 DMA_FROM_DEVICE);
 128                 op->op_mapped = 0;
 129         }
 130
 131         rds_ib_send_complete(container_of(op, struct rds_message, atomic),
 132                              wc_status, rds_atomic_send_complete);
 133
 134         if (op->op_type == RDS_ATOMIC_TYPE_CSWP)
 135                 rds_ib_stats_inc(s_ib_atomic_cswp);
 136         else
 137                 rds_ib_stats_inc(s_ib_atomic_fadd);
 138 }
 139
 140 /*
 141  * Unmap the resources associated with a struct send_work.
 142  *
 143  * Returns the rm for no good reason other than it is unobtainable
 144  * other than by switching on wr.opcode, currently, and the caller,
 145  * the event handler, needs it.
 146  */
 147 static struct rds_message *rds_ib_send_unmap_op(struct rds_ib_connection *ic,
 148                                                 struct rds_ib_send_work *send,
 149                                                 int wc_status)
 150 {
 151         struct rds_message *rm = NULL;
 152
 153         /* In the error case, wc.opcode sometimes contains garbage */
 154         switch (send->s_wr.opcode) {
 155         case IB_WR_SEND:
 156                 if (send->s_op) {
 157                         rm = container_of(send->s_op, struct rds_message, data);
 158                         rds_ib_send_unmap_data(ic, send->s_op, wc_status);
 159                 }
 160                 break;
 161         case IB_WR_RDMA_WRITE:
 162         case IB_WR_RDMA_READ:
 163                 if (send->s_op) {
 164                         rm = container_of(send->s_op, struct rds_message, rdma);
 165                         rds_ib_send_unmap_rdma(ic, send->s_op, wc_status);
 166                 }
 167                 break;
 168         case IB_WR_ATOMIC_FETCH_AND_ADD:
 169         case IB_WR_ATOMIC_CMP_AND_SWP:
 170                 if (send->s_op) {
 171                         rm = container_of(send->s_op, struct rds_message, atomic);
 172                         rds_ib_send_unmap_atomic(ic, send->s_op, wc_status);
 173                 }
 174                 break;
 175         default:
 176                 if (printk_ratelimit())
 177                         printk(KERN_NOTICE
 178                                "RDS/IB: %s: unexpected opcode 0x%x in WR!\n",
 179                                __func__, send->s_wr.opcode);
 180                 break;
 181         }
 182
 183         send->s_wr.opcode = 0xdead;
 184
 185         return rm;
 186 }
 187
 188 void rds_ib_send_init_ring(struct rds_ib_connection *ic)
 189 {
 190         struct rds_ib_send_work *send;
 191         u32 i;
 192
 193         for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
 194                 struct ib_sge *sge;
 195
 196                 send->s_op = NULL;
 197
 198                 send->s_wr.wr_id = i;
 199                 send->s_wr.sg_list = send->s_sge;
 200                 send->s_wr.ex.imm_data = 0;
 201
 202                 sge = &send->s_sge[0];
 203                 sge->addr = ic->i_send_hdrs_dma + (i * sizeof(struct rds_header));
 204                 sge->length = sizeof(struct rds_header);
 205                 sge->lkey = ic->i_mr->lkey;
 206
 207                 send->s_sge[1].lkey = ic->i_mr->lkey;
 208         }
 209 }
 210
 211 void rds_ib_send_clear_ring(struct rds_ib_connection *ic)
 212 {
 213         struct rds_ib_send_work *send;
 214         u32 i;
 215
 216         for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
 217                 if (send->s_op && send->s_wr.opcode != 0xdead)
 218                         rds_ib_send_unmap_op(ic, send, IB_WC_WR_FLUSH_ERR);
 219         }
 220 }
 221
 222 /*
 223  * The _oldest/_free ring operations here race cleanly with the alloc/unalloc
 224  * operations performed in the send path.  As the sender allocs and potentially
 225  * unallocs the next free entry in the ring it doesn't alter which is
 226  * the next to be freed, which is what this is concerned with.
 227  */
 228 void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context)
 229 {
 230         struct rds_connection *conn = context;
 231         struct rds_ib_connection *ic = conn->c_transport_data;
 232         struct rds_message *rm = NULL;
 233         struct ib_wc wc;
 234         struct rds_ib_send_work *send;
 235         u32 completed;
 236         u32 oldest;
 237         u32 i = 0;
 238         int ret;
 239
 240         rdsdebug("cq %p conn %p\n", cq, conn);
 241         rds_ib_stats_inc(s_ib_tx_cq_call);
 242         ret = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 243         if (ret)
 244                 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
 245
 246         while (ib_poll_cq(cq, 1, &wc) > 0) {
 247                 rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
 248                          (unsigned long long)wc.wr_id, wc.status, wc.byte_len,
 249                          be32_to_cpu(wc.ex.imm_data));
 250                 rds_ib_stats_inc(s_ib_tx_cq_event);
 251
 252                 if (wc.wr_id == RDS_IB_ACK_WR_ID) {
 253                         if (ic->i_ack_queued + HZ/2 < jiffies)
 254                                 rds_ib_stats_inc(s_ib_tx_stalled);
 255                         rds_ib_ack_send_complete(ic);
 256                         continue;
 257                 }
 258
 259                 oldest = rds_ib_ring_oldest(&ic->i_send_ring);
 260
 261                 completed = rds_ib_ring_completed(&ic->i_send_ring, wc.wr_id, oldest);
 262
 263                 for (i = 0; i < completed; i++) {
 264                         send = &ic->i_sends[oldest];
 265
 266                         rm = rds_ib_send_unmap_op(ic, send, wc.status);
 267
 268                         if (send->s_queued + HZ/2 < jiffies)
 269                                 rds_ib_stats_inc(s_ib_tx_stalled);
 270
 271                         if (send->s_op) {
 272                                 if (send->s_op == rm->m_final_op) {
 273                                         /* If anyone waited for this message to get flushed out, wake
 274                                          * them up now */
 275                                         rds_message_unmapped(rm);
 276                                 }
 277                                 rds_message_put(rm);
 278                                 send->s_op = NULL;
 279                         }
 280
 281                         oldest = (oldest + 1) % ic->i_send_ring.w_nr;
 282                 }
 283
 284                 rds_ib_ring_free(&ic->i_send_ring, completed);
 285
 286                 if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
 287                     test_bit(0, &conn->c_map_queued))
 288                         queue_delayed_work(rds_wq, &conn->c_send_w, 0);
 289
 290                 /* We expect errors as the qp is drained during shutdown */
 291                 if (wc.status != IB_WC_SUCCESS && rds_conn_up(conn)) {
 292                         rds_ib_conn_error(conn,
 293                                 "send completion on %pI4 "
 294                                 "had status %u, disconnecting and reconnecting\n",
 295                                 &conn->c_faddr, wc.status);
 296                 }
 297         }
 298 }
 299
 300 /*
 301  * This is the main function for allocating credits when sending
 302  * messages.
 303  *
 304  * Conceptually, we have two counters:
 305  *  -   send credits: this tells us how many WRs we're allowed
 306  *      to submit without overruning the reciever's queue. For
 307  *      each SEND WR we post, we decrement this by one.
 308  *
 309  *  -   posted credits: this tells us how many WRs we recently
 310  *      posted to the receive queue. This value is transferred
 311  *      to the peer as a "credit update" in a RDS header field.
 312  *      Every time we transmit credits to the peer, we subtract
 313  *      the amount of transferred credits from this counter.
 314  *
 315  * It is essential that we avoid situations where both sides have
 316  * exhausted their send credits, and are unable to send new credits
 317  * to the peer. We achieve this by requiring that we send at least
 318  * one credit update to the peer before exhausting our credits.
 319  * When new credits arrive, we subtract one credit that is withheld
 320  * until we've posted new buffers and are ready to transmit these
 321  * credits (see rds_ib_send_add_credits below).
 322  *
 323  * The RDS send code is essentially single-threaded; rds_send_xmit
 324  * grabs c_send_lock to ensure exclusive access to the send ring.
 325  * However, the ACK sending code is independent and can race with
 326  * message SENDs.
 327  *
 328  * In the send path, we need to update the counters for send credits
 329  * and the counter of posted buffers atomically - when we use the
 330  * last available credit, we cannot allow another thread to race us
 331  * and grab the posted credits counter.  Hence, we have to use a
 332  * spinlock to protect the credit counter, or use atomics.
 333  *
 334  * Spinlocks shared between the send and the receive path are bad,
 335  * because they create unnecessary delays. An early implementation
 336  * using a spinlock showed a 5% degradation in throughput at some
 337  * loads.
 338  *
 339  * This implementation avoids spinlocks completely, putting both
 340  * counters into a single atomic, and updating that atomic using
 341  * atomic_add (in the receive path, when receiving fresh credits),
 342  * and using atomic_cmpxchg when updating the two counters.
 343  */
 344 int rds_ib_send_grab_credits(struct rds_ib_connection *ic,
 345                              u32 wanted, u32 *adv_credits, int need_posted, int max_posted)
 346 {
 347         unsigned int avail, posted, got = 0, advertise;
 348         long oldval, newval;
 349
 350         *adv_credits = 0;
 351         if (!ic->i_flowctl)
 352                 return wanted;
 353
 354 try_again:
 355         advertise = 0;
 356         oldval = newval = atomic_read(&ic->i_credits);
 357         posted = IB_GET_POST_CREDITS(oldval);
 358         avail = IB_GET_SEND_CREDITS(oldval);
 359
 360         rdsdebug("rds_ib_send_grab_credits(%u): credits=%u posted=%u\n",
 361                         wanted, avail, posted);
 362
 363         /* The last credit must be used to send a credit update. */
 364         if (avail && !posted)
 365                 avail--;
 366
 367         if (avail < wanted) {
 368                 struct rds_connection *conn = ic->i_cm_id->context;
 369
 370                 /* Oops, there aren't that many credits left! */
 371                 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 372                 got = avail;
 373         } else {
 374                 /* Sometimes you get what you want, lalala. */
 375                 got = wanted;
 376         }
 377         newval -= IB_SET_SEND_CREDITS(got);
 378
 379         /*
 380          * If need_posted is non-zero, then the caller wants
 381          * the posted regardless of whether any send credits are
 382          * available.
 383          */
 384         if (posted && (got || need_posted)) {
 385                 advertise = min_t(unsigned int, posted, max_posted);
 386                 newval -= IB_SET_POST_CREDITS(advertise);
 387         }
 388
 389         /* Finally bill everything */
 390         if (atomic_cmpxchg(&ic->i_credits, oldval, newval) != oldval)
 391                 goto try_again;
 392
 393         *adv_credits = advertise;
 394         return got;
 395 }
 396
 397 void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits)
 398 {
 399         struct rds_ib_connection *ic = conn->c_transport_data;
 400
 401         if (credits == 0)
 402                 return;
 403
 404         rdsdebug("rds_ib_send_add_credits(%u): current=%u%s\n",
 405                         credits,
 406                         IB_GET_SEND_CREDITS(atomic_read(&ic->i_credits)),
 407                         test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ? ", ll_send_full" : "");
 408
 409         atomic_add(IB_SET_SEND_CREDITS(credits), &ic->i_credits);
 410         if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags))
 411                 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
 412
 413         WARN_ON(IB_GET_SEND_CREDITS(credits) >= 16384);
 414
 415         rds_ib_stats_inc(s_ib_rx_credit_updates);
 416 }
 417
 418 void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted)
 419 {
 420         struct rds_ib_connection *ic = conn->c_transport_data;
 421
 422         if (posted == 0)
 423                 return;
 424
 425         atomic_add(IB_SET_POST_CREDITS(posted), &ic->i_credits);
 426
 427         /* Decide whether to send an update to the peer now.
 428          * If we would send a credit update for every single buffer we
 429          * post, we would end up with an ACK storm (ACK arrives,
 430          * consumes buffer, we refill the ring, send ACK to remote
 431          * advertising the newly posted buffer... ad inf)
 432          *
 433          * Performance pretty much depends on how often we send
 434          * credit updates - too frequent updates mean lots of ACKs.
 435          * Too infrequent updates, and the peer will run out of
 436          * credits and has to throttle.
 437          * For the time being, 16 seems to be a good compromise.
 438          */
 439         if (IB_GET_POST_CREDITS(atomic_read(&ic->i_credits)) >= 16)
 440                 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
 441 }
 442
 443 static inline void rds_ib_set_wr_signal_state(struct rds_ib_connection *ic,
 444                                               struct rds_ib_send_work *send,
 445                                               bool notify)
 446 {
 447         /*
 448          * We want to delay signaling completions just enough to get
 449          * the batching benefits but not so much that we create dead time
 450          * on the wire.
 451          */
 452         if (ic->i_unsignaled_wrs-- == 0 || notify) {
 453                 ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
 454                 send->s_wr.send_flags |= IB_SEND_SIGNALED;
 455         }
 456 }
 457
 458 /*
 459  * This can be called multiple times for a given message.  The first time
 460  * we see a message we map its scatterlist into the IB device so that
 461  * we can provide that mapped address to the IB scatter gather entries
 462  * in the IB work requests.  We translate the scatterlist into a series
 463  * of work requests that fragment the message.  These work requests complete
 464  * in order so we pass ownership of the message to the completion handler
 465  * once we send the final fragment.
 466  *
 467  * The RDS core uses the c_send_lock to only enter this function once
 468  * per connection.  This makes sure that the tx ring alloc/unalloc pairs
 469  * don't get out of sync and confuse the ring.
 470  */
 471 int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
 472                 unsigned int hdr_off, unsigned int sg, unsigned int off)
 473 {
 474         struct rds_ib_connection *ic = conn->c_transport_data;
 475         struct ib_device *dev = ic->i_cm_id->device;
 476         struct rds_ib_send_work *send = NULL;
 477         struct rds_ib_send_work *first;
 478         struct rds_ib_send_work *prev;
 479         struct ib_send_wr *failed_wr;
 480         struct scatterlist *scat;
 481         u32 pos;
 482         u32 i;
 483         u32 work_alloc;
 484         u32 credit_alloc = 0;
 485         u32 posted;
 486         u32 adv_credits = 0;
 487         int send_flags = 0;
 488         int bytes_sent = 0;
 489         int ret;
 490         int flow_controlled = 0;
 491
 492         BUG_ON(off % RDS_FRAG_SIZE);
 493         BUG_ON(hdr_off != 0 && hdr_off != sizeof(struct rds_header));
 494
 495         /* Do not send cong updates to IB loopback */
 496         if (conn->c_loopback
 497             && rm->m_inc.i_hdr.h_flags & RDS_FLAG_CONG_BITMAP) {
 498                 rds_cong_map_updated(conn->c_fcong, ~(u64) 0);
 499                 return sizeof(struct rds_header) + RDS_CONG_MAP_BYTES;
 500         }
 501
 502         /* FIXME we may overallocate here */
 503         if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0)
 504                 i = 1;
 505         else
 506                 i = ceil(be32_to_cpu(rm->m_inc.i_hdr.h_len), RDS_FRAG_SIZE);
 507
 508         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
 509         if (work_alloc == 0) {
 510                 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 511                 rds_ib_stats_inc(s_ib_tx_ring_full);
 512                 ret = -ENOMEM;
 513                 goto out;
 514         }
 515
 516         if (ic->i_flowctl) {
 517                 credit_alloc = rds_ib_send_grab_credits(ic, work_alloc, &posted, 0, RDS_MAX_ADV_CREDIT);
 518                 adv_credits += posted;
 519                 if (credit_alloc < work_alloc) {
 520                         rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - credit_alloc);
 521                         work_alloc = credit_alloc;
 522                         flow_controlled = 1;
 523                 }
 524                 if (work_alloc == 0) {
 525                         set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 526                         rds_ib_stats_inc(s_ib_tx_throttle);
 527                         ret = -ENOMEM;
 528                         goto out;
 529                 }
 530         }
 531
 532         /* map the message the first time we see it */
 533         if (!ic->i_data_op) {
 534                 if (rm->data.op_nents) {
 535                         rm->data.op_count = ib_dma_map_sg(dev,
 536                                                           rm->data.op_sg,
 537                                                           rm->data.op_nents,
 538                                                           DMA_TO_DEVICE);
 539                         rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->data.op_count);
 540                         if (rm->data.op_count == 0) {
 541                                 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
 542                                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 543                                 ret = -ENOMEM; /* XXX ? */
 544                                 goto out;
 545                         }
 546                 } else {
 547                         rm->data.op_count = 0;
 548                 }
 549
 550                 rds_message_addref(rm);
 551                 ic->i_data_op = &rm->data;
 552
 553                 /* Finalize the header */
 554                 if (test_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags))
 555                         rm->m_inc.i_hdr.h_flags |= RDS_FLAG_ACK_REQUIRED;
 556                 if (test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))
 557                         rm->m_inc.i_hdr.h_flags |= RDS_FLAG_RETRANSMITTED;
 558
 559                 /* If it has a RDMA op, tell the peer we did it. This is
 560                  * used by the peer to release use-once RDMA MRs. */
 561                 if (rm->rdma.op_active) {
 562                         struct rds_ext_header_rdma ext_hdr;
 563
 564                         ext_hdr.h_rdma_rkey = cpu_to_be32(rm->rdma.op_rkey);
 565                         rds_message_add_extension(&rm->m_inc.i_hdr,
 566                                         RDS_EXTHDR_RDMA, &ext_hdr, sizeof(ext_hdr));
 567                 }
 568                 if (rm->m_rdma_cookie) {
 569                         rds_message_add_rdma_dest_extension(&rm->m_inc.i_hdr,
 570                                         rds_rdma_cookie_key(rm->m_rdma_cookie),
 571                                         rds_rdma_cookie_offset(rm->m_rdma_cookie));
 572                 }
 573
 574                 /* Note - rds_ib_piggyb_ack clears the ACK_REQUIRED bit, so
 575                  * we should not do this unless we have a chance of at least
 576                  * sticking the header into the send ring. Which is why we
 577                  * should call rds_ib_ring_alloc first. */
 578                 rm->m_inc.i_hdr.h_ack = cpu_to_be64(rds_ib_piggyb_ack(ic));
 579                 rds_message_make_checksum(&rm->m_inc.i_hdr);
 580
 581                 /*
 582                  * Update adv_credits since we reset the ACK_REQUIRED bit.
 583                  */
 584                 if (ic->i_flowctl) {
 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
 591         /* Sometimes you want to put a fence between an RDMA
 592          * READ and the following SEND.
 593          * We could either do this all the time
 594          * or when requested by the user. Right now, we let
 595          * the application choose.
 596          */
 597         if (rm->rdma.op_active && rm->rdma.op_fence)
 598                 send_flags = IB_SEND_FENCE;
 599
 600         /* Each frag gets a header. Msgs may be 0 bytes */
 601         send = &ic->i_sends[pos];
 602         first = send;
 603         prev = NULL;
 604         scat = &ic->i_data_op->op_sg[sg];
 605         i = 0;
 606         do {
 607                 unsigned int len = 0;
 608
 609                 /* Set up the header */
 610                 send->s_wr.send_flags = send_flags;
 611                 send->s_wr.opcode = IB_WR_SEND;
 612                 send->s_wr.num_sge = 1;
 613                 send->s_wr.next = NULL;
 614                 send->s_queued = jiffies;
 615                 send->s_op = NULL;
 616
 617                 send->s_sge[0].addr = ic->i_send_hdrs_dma
 618                         + (pos * sizeof(struct rds_header));
 619                 send->s_sge[0].length = sizeof(struct rds_header);
 620
 621                 memcpy(&ic->i_send_hdrs[pos], &rm->m_inc.i_hdr, sizeof(struct rds_header));
 622
 623                 /* Set up the data, if present */
 624                 if (i < work_alloc
 625                     && scat != &rm->data.op_sg[rm->data.op_count]) {
 626                         len = min(RDS_FRAG_SIZE, ib_sg_dma_len(dev, scat) - off);
 627                         send->s_wr.num_sge = 2;
 628
 629                         send->s_sge[1].addr = ib_sg_dma_address(dev, scat) + off;
 630                         send->s_sge[1].length = len;
 631
 632                         bytes_sent += len;
 633                         off += len;
 634                         if (off == ib_sg_dma_len(dev, scat)) {
 635                                 scat++;
 636                                 off = 0;
 637                         }
 638                 }
 639
 640                 rds_ib_set_wr_signal_state(ic, send, 0);
 641
 642                 /*
 643                  * Always signal the last one if we're stopping due to flow control.
 644                  */
 645                 if (ic->i_flowctl && flow_controlled && i == (work_alloc-1))
 646                         send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
 647
 648                 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
 649                          &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
 650
 651                 if (ic->i_flowctl && adv_credits) {
 652                         struct rds_header *hdr = &ic->i_send_hdrs[pos];
 653
 654                         /* add credit and redo the header checksum */
 655                         hdr->h_credit = adv_credits;
 656                         rds_message_make_checksum(hdr);
 657                         adv_credits = 0;
 658                         rds_ib_stats_inc(s_ib_tx_credit_updates);
 659                 }
 660
 661                 if (prev)
 662                         prev->s_wr.next = &send->s_wr;
 663                 prev = send;
 664
 665                 pos = (pos + 1) % ic->i_send_ring.w_nr;
 666                 send = &ic->i_sends[pos];
 667                 i++;
 668
 669         } while (i < work_alloc
 670                  && scat != &rm->data.op_sg[rm->data.op_count]);
 671
 672         /* Account the RDS header in the number of bytes we sent, but just once.
 673          * The caller has no concept of fragmentation. */
 674         if (hdr_off == 0)
 675                 bytes_sent += sizeof(struct rds_header);
 676
 677         /* if we finished the message then send completion owns it */
 678         if (scat == &rm->data.op_sg[rm->data.op_count]) {
 679                 prev->s_op = ic->i_data_op;
 680                 prev->s_wr.send_flags |= IB_SEND_SOLICITED;
 681                 ic->i_data_op = NULL;
 682         }
 683
 684         /* Put back wrs & credits we didn't use */
 685         if (i < work_alloc) {
 686                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
 687                 work_alloc = i;
 688         }
 689         if (ic->i_flowctl && i < credit_alloc)
 690                 rds_ib_send_add_credits(conn, credit_alloc - i);
 691
 692         /* XXX need to worry about failed_wr and partial sends. */
 693         failed_wr = &first->s_wr;
 694         ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
 695         rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
 696                  first, &first->s_wr, ret, failed_wr);
 697         BUG_ON(failed_wr != &first->s_wr);
 698         if (ret) {
 699                 printk(KERN_WARNING "RDS/IB: ib_post_send to %pI4 "
 700                        "returned %d\n", &conn->c_faddr, ret);
 701                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 702                 if (prev->s_op) {
 703                         ic->i_data_op = prev->s_op;
 704                         prev->s_op = NULL;
 705                 }
 706
 707                 rds_ib_conn_error(ic->conn, "ib_post_send failed\n");
 708                 goto out;
 709         }
 710
 711         ret = bytes_sent;
 712 out:
 713         BUG_ON(adv_credits);
 714         return ret;
 715 }
 716
 717 /*
 718  * Issue atomic operation.
 719  * A simplified version of the rdma case, we always map 1 SG, and
 720  * only 8 bytes, for the return value from the atomic operation.
 721  */
 722 int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op)
 723 {
 724         struct rds_ib_connection *ic = conn->c_transport_data;
 725         struct rds_ib_send_work *send = NULL;
 726         struct ib_send_wr *failed_wr;
 727         struct rds_ib_device *rds_ibdev;
 728         u32 pos;
 729         u32 work_alloc;
 730         int ret;
 731
 732         rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
 733
 734         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, 1, &pos);
 735         if (work_alloc != 1) {
 736                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 737                 rds_ib_stats_inc(s_ib_tx_ring_full);
 738                 ret = -ENOMEM;
 739                 goto out;
 740         }
 741
 742         /* address of send request in ring */
 743         send = &ic->i_sends[pos];
 744         send->s_queued = jiffies;
 745
 746         if (op->op_type == RDS_ATOMIC_TYPE_CSWP) {
 747                 send->s_wr.opcode = IB_WR_ATOMIC_CMP_AND_SWP;
 748                 send->s_wr.wr.atomic.compare_add = op->op_compare;
 749                 send->s_wr.wr.atomic.swap = op->op_swap_add;
 750         } else { /* FADD */
 751                 send->s_wr.opcode = IB_WR_ATOMIC_FETCH_AND_ADD;
 752                 send->s_wr.wr.atomic.compare_add = op->op_swap_add;
 753                 send->s_wr.wr.atomic.swap = 0;
 754         }
 755         rds_ib_set_wr_signal_state(ic, send, op->op_notify);
 756         send->s_wr.num_sge = 1;
 757         send->s_wr.next = NULL;
 758         send->s_wr.wr.atomic.remote_addr = op->op_remote_addr;
 759         send->s_wr.wr.atomic.rkey = op->op_rkey;
 760         send->s_op = op;
 761         rds_message_addref(container_of(send->s_op, struct rds_message, atomic));
 762
 763         /* map 8 byte retval buffer to the device */
 764         ret = ib_dma_map_sg(ic->i_cm_id->device, op->op_sg, 1, DMA_FROM_DEVICE);
 765         rdsdebug("ic %p mapping atomic op %p. mapped %d pg\n", ic, op, ret);
 766         if (ret != 1) {
 767                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 768                 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
 769                 ret = -ENOMEM; /* XXX ? */
 770                 goto out;
 771         }
 772
 773         /* Convert our struct scatterlist to struct ib_sge */
 774         send->s_sge[0].addr = ib_sg_dma_address(ic->i_cm_id->device, op->op_sg);
 775         send->s_sge[0].length = ib_sg_dma_len(ic->i_cm_id->device, op->op_sg);
 776         send->s_sge[0].lkey = ic->i_mr->lkey;
 777
 778         rdsdebug("rva %Lx rpa %Lx len %u\n", op->op_remote_addr,
 779                  send->s_sge[0].addr, send->s_sge[0].length);
 780
 781         failed_wr = &send->s_wr;
 782         ret = ib_post_send(ic->i_cm_id->qp, &send->s_wr, &failed_wr);
 783         rdsdebug("ic %p send %p (wr %p) ret %d wr %p\n", ic,
 784                  send, &send->s_wr, ret, failed_wr);
 785         BUG_ON(failed_wr != &send->s_wr);
 786         if (ret) {
 787                 printk(KERN_WARNING "RDS/IB: atomic ib_post_send to %pI4 "
 788                        "returned %d\n", &conn->c_faddr, ret);
 789                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 790                 goto out;
 791         }
 792
 793         if (unlikely(failed_wr != &send->s_wr)) {
 794                 printk(KERN_WARNING "RDS/IB: atomic ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
 795                 BUG_ON(failed_wr != &send->s_wr);
 796         }
 797
 798 out:
 799         return ret;
 800 }
 801
 802 int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op)
 803 {
 804         struct rds_ib_connection *ic = conn->c_transport_data;
 805         struct rds_ib_send_work *send = NULL;
 806         struct rds_ib_send_work *first;
 807         struct rds_ib_send_work *prev;
 808         struct ib_send_wr *failed_wr;
 809         struct rds_ib_device *rds_ibdev;
 810         struct scatterlist *scat;
 811         unsigned long len;
 812         u64 remote_addr = op->op_remote_addr;
 813         u32 pos;
 814         u32 work_alloc;
 815         u32 i;
 816         u32 j;
 817         int sent;
 818         int ret;
 819         int num_sge;
 820
 821         rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
 822
 823         /* map the op the first time we see it */
 824         if (!op->op_mapped) {
 825                 op->op_count = ib_dma_map_sg(ic->i_cm_id->device,
 826                                              op->op_sg, op->op_nents, (op->op_write) ?
 827                                              DMA_TO_DEVICE : DMA_FROM_DEVICE);
 828                 rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->op_count);
 829                 if (op->op_count == 0) {
 830                         rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
 831                         ret = -ENOMEM; /* XXX ? */
 832                         goto out;
 833                 }
 834
 835                 op->op_mapped = 1;
 836         }
 837
 838         /*
 839          * Instead of knowing how to return a partial rdma read/write we insist that there
 840          * be enough work requests to send the entire message.
 841          */
 842         i = ceil(op->op_count, rds_ibdev->max_sge);
 843
 844         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
 845         if (work_alloc != i) {
 846                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 847                 rds_ib_stats_inc(s_ib_tx_ring_full);
 848                 ret = -ENOMEM;
 849                 goto out;
 850         }
 851
 852         send = &ic->i_sends[pos];
 853         first = send;
 854         prev = NULL;
 855         scat = &op->op_sg[0];
 856         sent = 0;
 857         num_sge = op->op_count;
 858
 859         for (i = 0; i < work_alloc && scat != &op->op_sg[op->op_count]; i++) {
 860                 send->s_wr.send_flags = 0;
 861                 send->s_queued = jiffies;
 862                 send->s_op = NULL;
 863
 864                 rds_ib_set_wr_signal_state(ic, send, op->op_notify);
 865
 866                 send->s_wr.opcode = op->op_write ? IB_WR_RDMA_WRITE : IB_WR_RDMA_READ;
 867                 send->s_wr.wr.rdma.remote_addr = remote_addr;
 868                 send->s_wr.wr.rdma.rkey = op->op_rkey;
 869
 870                 if (num_sge > rds_ibdev->max_sge) {
 871                         send->s_wr.num_sge = rds_ibdev->max_sge;
 872                         num_sge -= rds_ibdev->max_sge;
 873                 } else {
 874                         send->s_wr.num_sge = num_sge;
 875                 }
 876
 877                 send->s_wr.next = NULL;
 878
 879                 if (prev)
 880                         prev->s_wr.next = &send->s_wr;
 881
 882                 for (j = 0; j < send->s_wr.num_sge && scat != &op->op_sg[op->op_count]; j++) {
 883                         len = ib_sg_dma_len(ic->i_cm_id->device, scat);
 884                         send->s_sge[j].addr =
 885                                  ib_sg_dma_address(ic->i_cm_id->device, scat);
 886                         send->s_sge[j].length = len;
 887                         send->s_sge[j].lkey = ic->i_mr->lkey;
 888
 889                         sent += len;
 890                         rdsdebug("ic %p sent %d remote_addr %llu\n", ic, sent, remote_addr);
 891
 892                         remote_addr += len;
 893                         scat++;
 894                 }
 895
 896                 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
 897                         &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
 898
 899                 prev = send;
 900                 if (++send == &ic->i_sends[ic->i_send_ring.w_nr])
 901                         send = ic->i_sends;
 902         }
 903
 904         /* give a reference to the last op */
 905         if (scat == &op->op_sg[op->op_count]) {
 906                 prev->s_op = op;
 907                 rds_message_addref(container_of(op, struct rds_message, rdma));
 908         }
 909
 910         if (i < work_alloc) {
 911                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
 912                 work_alloc = i;
 913         }
 914
 915         failed_wr = &first->s_wr;
 916         ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
 917         rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
 918                  first, &first->s_wr, ret, failed_wr);
 919         BUG_ON(failed_wr != &first->s_wr);
 920         if (ret) {
 921                 printk(KERN_WARNING "RDS/IB: rdma ib_post_send to %pI4 "
 922                        "returned %d\n", &conn->c_faddr, ret);
 923                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 924                 goto out;
 925         }
 926
 927         if (unlikely(failed_wr != &first->s_wr)) {
 928                 printk(KERN_WARNING "RDS/IB: ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
 929                 BUG_ON(failed_wr != &first->s_wr);
 930         }
 931
 932
 933 out:
 934         return ret;
 935 }
 936
 937 void rds_ib_xmit_complete(struct rds_connection *conn)
 938 {
 939         struct rds_ib_connection *ic = conn->c_transport_data;
 940
 941         /* We may have a pending ACK or window update we were unable
 942          * to send previously (due to flow control). Try again. */
 943         rds_ib_attempt_ack(ic);
 944 }