kernel/net/rds/iw_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 #include <linux/ratelimit.h>
  38
  39 #include "rds.h"
  40 #include "iw.h"
  41
  42 static void rds_iw_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_iw_send_unmap_rdma(struct rds_iw_connection *ic,
  67                                    struct rm_rdma_op *op)
  68 {
  69         if (op->op_mapped) {
  70                 ib_dma_unmap_sg(ic->i_cm_id->device,
  71                         op->op_sg, op->op_nents,
  72                         op->op_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
  73                 op->op_mapped = 0;
  74         }
  75 }
  76
  77 static void rds_iw_send_unmap_rm(struct rds_iw_connection *ic,
  78                           struct rds_iw_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->data.op_sg, rm->data.op_nents,
  87                      DMA_TO_DEVICE);
  88
  89         if (rm->rdma.op_active) {
  90                 rds_iw_send_unmap_rdma(ic, &rm->rdma);
  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_iw_send_rdma_complete(rm, wc_status);
 113
 114                 if (rm->rdma.op_write)
 115                         rds_stats_add(s_send_rdma_bytes, rm->rdma.op_bytes);
 116                 else
 117                         rds_stats_add(s_recv_rdma_bytes, rm->rdma.op_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_iw_send_init_ring(struct rds_iw_connection *ic)
 129 {
 130         struct rds_iw_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                 send->s_mapping = NULL;
 139
 140                 send->s_send_wr.next = NULL;
 141                 send->s_send_wr.wr_id = i;
 142                 send->s_send_wr.sg_list = send->s_sge;
 143                 send->s_send_wr.num_sge = 1;
 144                 send->s_send_wr.opcode = IB_WR_SEND;
 145                 send->s_send_wr.send_flags = 0;
 146                 send->s_send_wr.ex.imm_data = 0;
 147
 148                 sge = rds_iw_data_sge(ic, send->s_sge);
 149                 sge->lkey = 0;
 150
 151                 sge = rds_iw_header_sge(ic, send->s_sge);
 152                 sge->addr = ic->i_send_hdrs_dma + (i * sizeof(struct rds_header));
 153                 sge->length = sizeof(struct rds_header);
 154                 sge->lkey = 0;
 155
 156                 send->s_mr = ib_alloc_mr(ic->i_pd, IB_MR_TYPE_MEM_REG,
 157                                          fastreg_message_size);
 158                 if (IS_ERR(send->s_mr)) {
 159                         printk(KERN_WARNING "RDS/IW: ib_alloc_mr failed\n");
 160                         break;
 161                 }
 162         }
 163 }
 164
 165 void rds_iw_send_clear_ring(struct rds_iw_connection *ic)
 166 {
 167         struct rds_iw_send_work *send;
 168         u32 i;
 169
 170         for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
 171                 BUG_ON(!send->s_mr);
 172                 ib_dereg_mr(send->s_mr);
 173                 if (send->s_send_wr.opcode == 0xdead)
 174                         continue;
 175                 if (send->s_rm)
 176                         rds_iw_send_unmap_rm(ic, send, IB_WC_WR_FLUSH_ERR);
 177                 if (send->s_op)
 178                         rds_iw_send_unmap_rdma(ic, send->s_op);
 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_iw_send_cq_comp_handler(struct ib_cq *cq, void *context)
 189 {
 190         struct rds_connection *conn = context;
 191         struct rds_iw_connection *ic = conn->c_transport_data;
 192         struct ib_wc wc;
 193         struct rds_iw_send_work *send;
 194         u32 completed;
 195         u32 oldest;
 196         u32 i;
 197         int ret;
 198
 199         rdsdebug("cq %p conn %p\n", cq, conn);
 200         rds_iw_stats_inc(s_iw_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_iw_stats_inc(s_iw_tx_cq_event);
 210
 211                 if (wc.status != IB_WC_SUCCESS) {
 212                         printk(KERN_ERR "WC Error:  status = %d opcode = %d\n", wc.status, wc.opcode);
 213                         break;
 214                 }
 215
 216                 if (wc.opcode == IB_WC_LOCAL_INV && wc.wr_id == RDS_IW_LOCAL_INV_WR_ID) {
 217                         ic->i_fastreg_posted = 0;
 218                         continue;
 219                 }
 220
 221                 if (wc.opcode == IB_WC_REG_MR && wc.wr_id == RDS_IW_REG_WR_ID) {
 222                         ic->i_fastreg_posted = 1;
 223                         continue;
 224                 }
 225
 226                 if (wc.wr_id == RDS_IW_ACK_WR_ID) {
 227                         if (time_after(jiffies, ic->i_ack_queued + HZ/2))
 228                                 rds_iw_stats_inc(s_iw_tx_stalled);
 229                         rds_iw_ack_send_complete(ic);
 230                         continue;
 231                 }
 232
 233                 oldest = rds_iw_ring_oldest(&ic->i_send_ring);
 234
 235                 completed = rds_iw_ring_completed(&ic->i_send_ring, wc.wr_id, oldest);
 236
 237                 for (i = 0; i < completed; i++) {
 238                         send = &ic->i_sends[oldest];
 239
 240                         /* In the error case, wc.opcode sometimes contains garbage */
 241                         switch (send->s_send_wr.opcode) {
 242                         case IB_WR_SEND:
 243                                 if (send->s_rm)
 244                                         rds_iw_send_unmap_rm(ic, send, wc.status);
 245                                 break;
 246                         case IB_WR_REG_MR:
 247                         case IB_WR_RDMA_WRITE:
 248                         case IB_WR_RDMA_READ:
 249                         case IB_WR_RDMA_READ_WITH_INV:
 250                                 /* Nothing to be done - the SG list will be unmapped
 251                                  * when the SEND completes. */
 252                                 break;
 253                         default:
 254                                 printk_ratelimited(KERN_NOTICE
 255                                                 "RDS/IW: %s: unexpected opcode 0x%x in WR!\n",
 256                                                 __func__, send->s_send_wr.opcode);
 257                                 break;
 258                         }
 259
 260                         send->s_send_wr.opcode = 0xdead;
 261                         send->s_send_wr.num_sge = 1;
 262                         if (time_after(jiffies, send->s_queued + HZ/2))
 263                                 rds_iw_stats_inc(s_iw_tx_stalled);
 264
 265                         /* If a RDMA operation produced an error, signal this right
 266                          * away. If we don't, the subsequent SEND that goes with this
 267                          * RDMA will be canceled with ERR_WFLUSH, and the application
 268                          * never learn that the RDMA failed. */
 269                         if (unlikely(wc.status == IB_WC_REM_ACCESS_ERR && send->s_op)) {
 270                                 struct rds_message *rm;
 271
 272                                 rm = rds_send_get_message(conn, send->s_op);
 273                                 if (rm)
 274                                         rds_iw_send_rdma_complete(rm, wc.status);
 275                         }
 276
 277                         oldest = (oldest + 1) % ic->i_send_ring.w_nr;
 278                 }
 279
 280                 rds_iw_ring_free(&ic->i_send_ring, completed);
 281
 282                 if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
 283                     test_bit(0, &conn->c_map_queued))
 284                         queue_delayed_work(rds_wq, &conn->c_send_w, 0);
 285
 286                 /* We expect errors as the qp is drained during shutdown */
 287                 if (wc.status != IB_WC_SUCCESS && rds_conn_up(conn)) {
 288                         rds_iw_conn_error(conn,
 289                                 "send completion on %pI4 "
 290                                 "had status %u, disconnecting and reconnecting\n",
 291                                 &conn->c_faddr, wc.status);
 292                 }
 293         }
 294 }
 295
 296 /*
 297  * This is the main function for allocating credits when sending
 298  * messages.
 299  *
 300  * Conceptually, we have two counters:
 301  *  -   send credits: this tells us how many WRs we're allowed
 302  *      to submit without overruning the receiver's queue. For
 303  *      each SEND WR we post, we decrement this by one.
 304  *
 305  *  -   posted credits: this tells us how many WRs we recently
 306  *      posted to the receive queue. This value is transferred
 307  *      to the peer as a "credit update" in a RDS header field.
 308  *      Every time we transmit credits to the peer, we subtract
 309  *      the amount of transferred credits from this counter.
 310  *
 311  * It is essential that we avoid situations where both sides have
 312  * exhausted their send credits, and are unable to send new credits
 313  * to the peer. We achieve this by requiring that we send at least
 314  * one credit update to the peer before exhausting our credits.
 315  * When new credits arrive, we subtract one credit that is withheld
 316  * until we've posted new buffers and are ready to transmit these
 317  * credits (see rds_iw_send_add_credits below).
 318  *
 319  * The RDS send code is essentially single-threaded; rds_send_xmit
 320  * grabs c_send_lock to ensure exclusive access to the send ring.
 321  * However, the ACK sending code is independent and can race with
 322  * message SENDs.
 323  *
 324  * In the send path, we need to update the counters for send credits
 325  * and the counter of posted buffers atomically - when we use the
 326  * last available credit, we cannot allow another thread to race us
 327  * and grab the posted credits counter.  Hence, we have to use a
 328  * spinlock to protect the credit counter, or use atomics.
 329  *
 330  * Spinlocks shared between the send and the receive path are bad,
 331  * because they create unnecessary delays. An early implementation
 332  * using a spinlock showed a 5% degradation in throughput at some
 333  * loads.
 334  *
 335  * This implementation avoids spinlocks completely, putting both
 336  * counters into a single atomic, and updating that atomic using
 337  * atomic_add (in the receive path, when receiving fresh credits),
 338  * and using atomic_cmpxchg when updating the two counters.
 339  */
 340 int rds_iw_send_grab_credits(struct rds_iw_connection *ic,
 341                              u32 wanted, u32 *adv_credits, int need_posted, int max_posted)
 342 {
 343         unsigned int avail, posted, got = 0, advertise;
 344         long oldval, newval;
 345
 346         *adv_credits = 0;
 347         if (!ic->i_flowctl)
 348                 return wanted;
 349
 350 try_again:
 351         advertise = 0;
 352         oldval = newval = atomic_read(&ic->i_credits);
 353         posted = IB_GET_POST_CREDITS(oldval);
 354         avail = IB_GET_SEND_CREDITS(oldval);
 355
 356         rdsdebug("wanted=%u credits=%u posted=%u\n",
 357                         wanted, avail, posted);
 358
 359         /* The last credit must be used to send a credit update. */
 360         if (avail && !posted)
 361                 avail--;
 362
 363         if (avail < wanted) {
 364                 struct rds_connection *conn = ic->i_cm_id->context;
 365
 366                 /* Oops, there aren't that many credits left! */
 367                 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 368                 got = avail;
 369         } else {
 370                 /* Sometimes you get what you want, lalala. */
 371                 got = wanted;
 372         }
 373         newval -= IB_SET_SEND_CREDITS(got);
 374
 375         /*
 376          * If need_posted is non-zero, then the caller wants
 377          * the posted regardless of whether any send credits are
 378          * available.
 379          */
 380         if (posted && (got || need_posted)) {
 381                 advertise = min_t(unsigned int, posted, max_posted);
 382                 newval -= IB_SET_POST_CREDITS(advertise);
 383         }
 384
 385         /* Finally bill everything */
 386         if (atomic_cmpxchg(&ic->i_credits, oldval, newval) != oldval)
 387                 goto try_again;
 388
 389         *adv_credits = advertise;
 390         return got;
 391 }
 392
 393 void rds_iw_send_add_credits(struct rds_connection *conn, unsigned int credits)
 394 {
 395         struct rds_iw_connection *ic = conn->c_transport_data;
 396
 397         if (credits == 0)
 398                 return;
 399
 400         rdsdebug("credits=%u current=%u%s\n",
 401                         credits,
 402                         IB_GET_SEND_CREDITS(atomic_read(&ic->i_credits)),
 403                         test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ? ", ll_send_full" : "");
 404
 405         atomic_add(IB_SET_SEND_CREDITS(credits), &ic->i_credits);
 406         if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags))
 407                 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
 408
 409         WARN_ON(IB_GET_SEND_CREDITS(credits) >= 16384);
 410
 411         rds_iw_stats_inc(s_iw_rx_credit_updates);
 412 }
 413
 414 void rds_iw_advertise_credits(struct rds_connection *conn, unsigned int posted)
 415 {
 416         struct rds_iw_connection *ic = conn->c_transport_data;
 417
 418         if (posted == 0)
 419                 return;
 420
 421         atomic_add(IB_SET_POST_CREDITS(posted), &ic->i_credits);
 422
 423         /* Decide whether to send an update to the peer now.
 424          * If we would send a credit update for every single buffer we
 425          * post, we would end up with an ACK storm (ACK arrives,
 426          * consumes buffer, we refill the ring, send ACK to remote
 427          * advertising the newly posted buffer... ad inf)
 428          *
 429          * Performance pretty much depends on how often we send
 430          * credit updates - too frequent updates mean lots of ACKs.
 431          * Too infrequent updates, and the peer will run out of
 432          * credits and has to throttle.
 433          * For the time being, 16 seems to be a good compromise.
 434          */
 435         if (IB_GET_POST_CREDITS(atomic_read(&ic->i_credits)) >= 16)
 436                 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
 437 }
 438
 439 static inline void
 440 rds_iw_xmit_populate_wr(struct rds_iw_connection *ic,
 441                 struct rds_iw_send_work *send, unsigned int pos,
 442                 unsigned long buffer, unsigned int length,
 443                 int send_flags)
 444 {
 445         struct ib_sge *sge;
 446
 447         WARN_ON(pos != send - ic->i_sends);
 448
 449         send->s_send_wr.send_flags = send_flags;
 450         send->s_send_wr.opcode = IB_WR_SEND;
 451         send->s_send_wr.num_sge = 2;
 452         send->s_send_wr.next = NULL;
 453         send->s_queued = jiffies;
 454         send->s_op = NULL;
 455
 456         if (length != 0) {
 457                 sge = rds_iw_data_sge(ic, send->s_sge);
 458                 sge->addr = buffer;
 459                 sge->length = length;
 460                 sge->lkey = rds_iw_local_dma_lkey(ic);
 461
 462                 sge = rds_iw_header_sge(ic, send->s_sge);
 463         } else {
 464                 /* We're sending a packet with no payload. There is only
 465                  * one SGE */
 466                 send->s_send_wr.num_sge = 1;
 467                 sge = &send->s_sge[0];
 468         }
 469
 470         sge->addr = ic->i_send_hdrs_dma + (pos * sizeof(struct rds_header));
 471         sge->length = sizeof(struct rds_header);
 472         sge->lkey = rds_iw_local_dma_lkey(ic);
 473 }
 474
 475 /*
 476  * This can be called multiple times for a given message.  The first time
 477  * we see a message we map its scatterlist into the IB device so that
 478  * we can provide that mapped address to the IB scatter gather entries
 479  * in the IB work requests.  We translate the scatterlist into a series
 480  * of work requests that fragment the message.  These work requests complete
 481  * in order so we pass ownership of the message to the completion handler
 482  * once we send the final fragment.
 483  *
 484  * The RDS core uses the c_send_lock to only enter this function once
 485  * per connection.  This makes sure that the tx ring alloc/unalloc pairs
 486  * don't get out of sync and confuse the ring.
 487  */
 488 int rds_iw_xmit(struct rds_connection *conn, struct rds_message *rm,
 489                 unsigned int hdr_off, unsigned int sg, unsigned int off)
 490 {
 491         struct rds_iw_connection *ic = conn->c_transport_data;
 492         struct ib_device *dev = ic->i_cm_id->device;
 493         struct rds_iw_send_work *send = NULL;
 494         struct rds_iw_send_work *first;
 495         struct rds_iw_send_work *prev;
 496         struct ib_send_wr *failed_wr;
 497         struct scatterlist *scat;
 498         u32 pos;
 499         u32 i;
 500         u32 work_alloc;
 501         u32 credit_alloc;
 502         u32 posted;
 503         u32 adv_credits = 0;
 504         int send_flags = 0;
 505         int sent;
 506         int ret;
 507         int flow_controlled = 0;
 508
 509         BUG_ON(off % RDS_FRAG_SIZE);
 510         BUG_ON(hdr_off != 0 && hdr_off != sizeof(struct rds_header));
 511
 512         /* Fastreg support */
 513         if (rds_rdma_cookie_key(rm->m_rdma_cookie) && !ic->i_fastreg_posted) {
 514                 ret = -EAGAIN;
 515                 goto out;
 516         }
 517
 518         /* FIXME we may overallocate here */
 519         if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0)
 520                 i = 1;
 521         else
 522                 i = ceil(be32_to_cpu(rm->m_inc.i_hdr.h_len), RDS_FRAG_SIZE);
 523
 524         work_alloc = rds_iw_ring_alloc(&ic->i_send_ring, i, &pos);
 525         if (work_alloc == 0) {
 526                 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 527                 rds_iw_stats_inc(s_iw_tx_ring_full);
 528                 ret = -ENOMEM;
 529                 goto out;
 530         }
 531
 532         credit_alloc = work_alloc;
 533         if (ic->i_flowctl) {
 534                 credit_alloc = rds_iw_send_grab_credits(ic, work_alloc, &posted, 0, RDS_MAX_ADV_CREDIT);
 535                 adv_credits += posted;
 536                 if (credit_alloc < work_alloc) {
 537                         rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc - credit_alloc);
 538                         work_alloc = credit_alloc;
 539                         flow_controlled++;
 540                 }
 541                 if (work_alloc == 0) {
 542                         set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 543                         rds_iw_stats_inc(s_iw_tx_throttle);
 544                         ret = -ENOMEM;
 545                         goto out;
 546                 }
 547         }
 548
 549         /* map the message the first time we see it */
 550         if (!ic->i_rm) {
 551                 /*
 552                 printk(KERN_NOTICE "rds_iw_xmit prep msg dport=%u flags=0x%x len=%d\n",
 553                                 be16_to_cpu(rm->m_inc.i_hdr.h_dport),
 554                                 rm->m_inc.i_hdr.h_flags,
 555                                 be32_to_cpu(rm->m_inc.i_hdr.h_len));
 556                    */
 557                 if (rm->data.op_nents) {
 558                         rm->data.op_count = ib_dma_map_sg(dev,
 559                                                           rm->data.op_sg,
 560                                                           rm->data.op_nents,
 561                                                           DMA_TO_DEVICE);
 562                         rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->data.op_count);
 563                         if (rm->data.op_count == 0) {
 564                                 rds_iw_stats_inc(s_iw_tx_sg_mapping_failure);
 565                                 rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
 566                                 ret = -ENOMEM; /* XXX ? */
 567                                 goto out;
 568                         }
 569                 } else {
 570                         rm->data.op_count = 0;
 571                 }
 572
 573                 ic->i_unsignaled_wrs = rds_iw_sysctl_max_unsig_wrs;
 574                 ic->i_unsignaled_bytes = rds_iw_sysctl_max_unsig_bytes;
 575                 rds_message_addref(rm);
 576                 rm->data.op_dmasg = 0;
 577                 rm->data.op_dmaoff = 0;
 578                 ic->i_rm = rm;
 579
 580                 /* Finalize the header */
 581                 if (test_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags))
 582                         rm->m_inc.i_hdr.h_flags |= RDS_FLAG_ACK_REQUIRED;
 583                 if (test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))
 584                         rm->m_inc.i_hdr.h_flags |= RDS_FLAG_RETRANSMITTED;
 585
 586                 /* If it has a RDMA op, tell the peer we did it. This is
 587                  * used by the peer to release use-once RDMA MRs. */
 588                 if (rm->rdma.op_active) {
 589                         struct rds_ext_header_rdma ext_hdr;
 590
 591                         ext_hdr.h_rdma_rkey = cpu_to_be32(rm->rdma.op_rkey);
 592                         rds_message_add_extension(&rm->m_inc.i_hdr,
 593                                         RDS_EXTHDR_RDMA, &ext_hdr, sizeof(ext_hdr));
 594                 }
 595                 if (rm->m_rdma_cookie) {
 596                         rds_message_add_rdma_dest_extension(&rm->m_inc.i_hdr,
 597                                         rds_rdma_cookie_key(rm->m_rdma_cookie),
 598                                         rds_rdma_cookie_offset(rm->m_rdma_cookie));
 599                 }
 600
 601                 /* Note - rds_iw_piggyb_ack clears the ACK_REQUIRED bit, so
 602                  * we should not do this unless we have a chance of at least
 603                  * sticking the header into the send ring. Which is why we
 604                  * should call rds_iw_ring_alloc first. */
 605                 rm->m_inc.i_hdr.h_ack = cpu_to_be64(rds_iw_piggyb_ack(ic));
 606                 rds_message_make_checksum(&rm->m_inc.i_hdr);
 607
 608                 /*
 609                  * Update adv_credits since we reset the ACK_REQUIRED bit.
 610                  */
 611                 rds_iw_send_grab_credits(ic, 0, &posted, 1, RDS_MAX_ADV_CREDIT - adv_credits);
 612                 adv_credits += posted;
 613                 BUG_ON(adv_credits > 255);
 614         }
 615
 616         send = &ic->i_sends[pos];
 617         first = send;
 618         prev = NULL;
 619         scat = &rm->data.op_sg[rm->data.op_dmasg];
 620         sent = 0;
 621         i = 0;
 622
 623         /* Sometimes you want to put a fence between an RDMA
 624          * READ and the following SEND.
 625          * We could either do this all the time
 626          * or when requested by the user. Right now, we let
 627          * the application choose.
 628          */
 629         if (rm->rdma.op_active && rm->rdma.op_fence)
 630                 send_flags = IB_SEND_FENCE;
 631
 632         /*
 633          * We could be copying the header into the unused tail of the page.
 634          * That would need to be changed in the future when those pages might
 635          * be mapped userspace pages or page cache pages.  So instead we always
 636          * use a second sge and our long-lived ring of mapped headers.  We send
 637          * the header after the data so that the data payload can be aligned on
 638          * the receiver.
 639          */
 640
 641         /* handle a 0-len message */
 642         if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0) {
 643                 rds_iw_xmit_populate_wr(ic, send, pos, 0, 0, send_flags);
 644                 goto add_header;
 645         }
 646
 647         /* if there's data reference it with a chain of work reqs */
 648         for (; i < work_alloc && scat != &rm->data.op_sg[rm->data.op_count]; i++) {
 649                 unsigned int len;
 650
 651                 send = &ic->i_sends[pos];
 652
 653                 len = min(RDS_FRAG_SIZE,
 654                           ib_sg_dma_len(dev, scat) - rm->data.op_dmaoff);
 655                 rds_iw_xmit_populate_wr(ic, send, pos,
 656                         ib_sg_dma_address(dev, scat) + rm->data.op_dmaoff, len,
 657                         send_flags);
 658
 659                 /*
 660                  * We want to delay signaling completions just enough to get
 661                  * the batching benefits but not so much that we create dead time
 662                  * on the wire.
 663                  */
 664                 if (ic->i_unsignaled_wrs-- == 0) {
 665                         ic->i_unsignaled_wrs = rds_iw_sysctl_max_unsig_wrs;
 666                         send->s_send_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
 667                 }
 668
 669                 ic->i_unsignaled_bytes -= len;
 670                 if (ic->i_unsignaled_bytes <= 0) {
 671                         ic->i_unsignaled_bytes = rds_iw_sysctl_max_unsig_bytes;
 672                         send->s_send_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
 673                 }
 674
 675                 /*
 676                  * Always signal the last one if we're stopping due to flow control.
 677                  */
 678                 if (flow_controlled && i == (work_alloc-1))
 679                         send->s_send_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
 680
 681                 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
 682                          &send->s_send_wr, send->s_send_wr.num_sge, send->s_send_wr.next);
 683
 684                 sent += len;
 685                 rm->data.op_dmaoff += len;
 686                 if (rm->data.op_dmaoff == ib_sg_dma_len(dev, scat)) {
 687                         scat++;
 688                         rm->data.op_dmaoff = 0;
 689                         rm->data.op_dmasg++;
 690                 }
 691
 692 add_header:
 693                 /* Tack on the header after the data. The header SGE should already
 694                  * have been set up to point to the right header buffer. */
 695                 memcpy(&ic->i_send_hdrs[pos], &rm->m_inc.i_hdr, sizeof(struct rds_header));
 696
 697                 if (0) {
 698                         struct rds_header *hdr = &ic->i_send_hdrs[pos];
 699
 700                         printk(KERN_NOTICE "send WR dport=%u flags=0x%x len=%d\n",
 701                                 be16_to_cpu(hdr->h_dport),
 702                                 hdr->h_flags,
 703                                 be32_to_cpu(hdr->h_len));
 704                 }
 705                 if (adv_credits) {
 706                         struct rds_header *hdr = &ic->i_send_hdrs[pos];
 707
 708                         /* add credit and redo the header checksum */
 709                         hdr->h_credit = adv_credits;
 710                         rds_message_make_checksum(hdr);
 711                         adv_credits = 0;
 712                         rds_iw_stats_inc(s_iw_tx_credit_updates);
 713                 }
 714
 715                 if (prev)
 716                         prev->s_send_wr.next = &send->s_send_wr;
 717                 prev = send;
 718
 719                 pos = (pos + 1) % ic->i_send_ring.w_nr;
 720         }
 721
 722         /* Account the RDS header in the number of bytes we sent, but just once.
 723          * The caller has no concept of fragmentation. */
 724         if (hdr_off == 0)
 725                 sent += sizeof(struct rds_header);
 726
 727         /* if we finished the message then send completion owns it */
 728         if (scat == &rm->data.op_sg[rm->data.op_count]) {
 729                 prev->s_rm = ic->i_rm;
 730                 prev->s_send_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
 731                 ic->i_rm = NULL;
 732         }
 733
 734         if (i < work_alloc) {
 735                 rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc - i);
 736                 work_alloc = i;
 737         }
 738         if (ic->i_flowctl && i < credit_alloc)
 739                 rds_iw_send_add_credits(conn, credit_alloc - i);
 740
 741         /* XXX need to worry about failed_wr and partial sends. */
 742         failed_wr = &first->s_send_wr;
 743         ret = ib_post_send(ic->i_cm_id->qp, &first->s_send_wr, &failed_wr);
 744         rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
 745                  first, &first->s_send_wr, ret, failed_wr);
 746         BUG_ON(failed_wr != &first->s_send_wr);
 747         if (ret) {
 748                 printk(KERN_WARNING "RDS/IW: ib_post_send to %pI4 "
 749                        "returned %d\n", &conn->c_faddr, ret);
 750                 rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
 751                 if (prev->s_rm) {
 752                         ic->i_rm = prev->s_rm;
 753                         prev->s_rm = NULL;
 754                 }
 755                 goto out;
 756         }
 757
 758         ret = sent;
 759 out:
 760         BUG_ON(adv_credits);
 761         return ret;
 762 }
 763
 764 static int rds_iw_build_send_reg(struct rds_iw_send_work *send,
 765                                  struct scatterlist *sg,
 766                                  int sg_nents)
 767 {
 768         int n;
 769
 770         n = ib_map_mr_sg(send->s_mr, sg, sg_nents, PAGE_SIZE);
 771         if (unlikely(n != sg_nents))
 772                 return n < 0 ? n : -EINVAL;
 773
 774         send->s_reg_wr.wr.opcode = IB_WR_REG_MR;
 775         send->s_reg_wr.wr.wr_id = 0;
 776         send->s_reg_wr.wr.num_sge = 0;
 777         send->s_reg_wr.mr = send->s_mr;
 778         send->s_reg_wr.key = send->s_mr->rkey;
 779         send->s_reg_wr.access = IB_ACCESS_REMOTE_WRITE;
 780
 781         ib_update_fast_reg_key(send->s_mr, send->s_remap_count++);
 782
 783         return 0;
 784 }
 785
 786 int rds_iw_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op)
 787 {
 788         struct rds_iw_connection *ic = conn->c_transport_data;
 789         struct rds_iw_send_work *send = NULL;
 790         struct rds_iw_send_work *first;
 791         struct rds_iw_send_work *prev;
 792         struct ib_send_wr *failed_wr;
 793         struct rds_iw_device *rds_iwdev;
 794         struct scatterlist *scat;
 795         unsigned long len;
 796         u64 remote_addr = op->op_remote_addr;
 797         u32 pos, fr_pos;
 798         u32 work_alloc;
 799         u32 i;
 800         u32 j;
 801         int sent;
 802         int ret;
 803         int num_sge;
 804         int sg_nents;
 805
 806         rds_iwdev = ib_get_client_data(ic->i_cm_id->device, &rds_iw_client);
 807
 808         /* map the message the first time we see it */
 809         if (!op->op_mapped) {
 810                 op->op_count = ib_dma_map_sg(ic->i_cm_id->device,
 811                                              op->op_sg, op->op_nents, (op->op_write) ?
 812                                              DMA_TO_DEVICE : DMA_FROM_DEVICE);
 813                 rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->op_count);
 814                 if (op->op_count == 0) {
 815                         rds_iw_stats_inc(s_iw_tx_sg_mapping_failure);
 816                         ret = -ENOMEM; /* XXX ? */
 817                         goto out;
 818                 }
 819
 820                 op->op_mapped = 1;
 821         }
 822
 823         if (!op->op_write) {
 824                 /* Alloc space on the send queue for the fastreg */
 825                 work_alloc = rds_iw_ring_alloc(&ic->i_send_ring, 1, &fr_pos);
 826                 if (work_alloc != 1) {
 827                         rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
 828                         rds_iw_stats_inc(s_iw_tx_ring_full);
 829                         ret = -ENOMEM;
 830                         goto out;
 831                 }
 832         }
 833
 834         /*
 835          * Instead of knowing how to return a partial rdma read/write we insist that there
 836          * be enough work requests to send the entire message.
 837          */
 838         i = ceil(op->op_count, rds_iwdev->max_sge);
 839
 840         work_alloc = rds_iw_ring_alloc(&ic->i_send_ring, i, &pos);
 841         if (work_alloc != i) {
 842                 rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
 843                 rds_iw_stats_inc(s_iw_tx_ring_full);
 844                 ret = -ENOMEM;
 845                 goto out;
 846         }
 847
 848         send = &ic->i_sends[pos];
 849         if (!op->op_write) {
 850                 first = prev = &ic->i_sends[fr_pos];
 851         } else {
 852                 first = send;
 853                 prev = NULL;
 854         }
 855         scat = &op->op_sg[0];
 856         sent = 0;
 857         num_sge = op->op_count;
 858         sg_nents = 0;
 859
 860         for (i = 0; i < work_alloc && scat != &op->op_sg[op->op_count]; i++) {
 861                 send->s_rdma_wr.wr.send_flags = 0;
 862                 send->s_queued = jiffies;
 863
 864                 /*
 865                  * We want to delay signaling completions just enough to get
 866                  * the batching benefits but not so much that we create dead time on the wire.
 867                  */
 868                 if (ic->i_unsignaled_wrs-- == 0) {
 869                         ic->i_unsignaled_wrs = rds_iw_sysctl_max_unsig_wrs;
 870                         send->s_rdma_wr.wr.send_flags = IB_SEND_SIGNALED;
 871                 }
 872
 873                 /* To avoid the need to have the plumbing to invalidate the fastreg_mr used
 874                  * for local access after RDS is finished with it, using
 875                  * IB_WR_RDMA_READ_WITH_INV will invalidate it after the read has completed.
 876                  */
 877                 if (op->op_write)
 878                         send->s_rdma_wr.wr.opcode = IB_WR_RDMA_WRITE;
 879                 else
 880                         send->s_rdma_wr.wr.opcode = IB_WR_RDMA_READ_WITH_INV;
 881
 882                 send->s_rdma_wr.remote_addr = remote_addr;
 883                 send->s_rdma_wr.rkey = op->op_rkey;
 884                 send->s_op = op;
 885
 886                 if (num_sge > rds_iwdev->max_sge) {
 887                         send->s_rdma_wr.wr.num_sge = rds_iwdev->max_sge;
 888                         num_sge -= rds_iwdev->max_sge;
 889                 } else
 890                         send->s_rdma_wr.wr.num_sge = num_sge;
 891
 892                 send->s_rdma_wr.wr.next = NULL;
 893
 894                 if (prev)
 895                         prev->s_send_wr.next = &send->s_rdma_wr.wr;
 896
 897                 for (j = 0; j < send->s_rdma_wr.wr.num_sge &&
 898                      scat != &op->op_sg[op->op_count]; j++) {
 899                         len = ib_sg_dma_len(ic->i_cm_id->device, scat);
 900
 901                         if (send->s_rdma_wr.wr.opcode == IB_WR_RDMA_READ_WITH_INV)
 902                                 sg_nents++;
 903                         else {
 904                                 send->s_sge[j].addr = ib_sg_dma_address(ic->i_cm_id->device, scat);
 905                                 send->s_sge[j].length = len;
 906                                 send->s_sge[j].lkey = rds_iw_local_dma_lkey(ic);
 907                         }
 908
 909                         sent += len;
 910                         rdsdebug("ic %p sent %d remote_addr %llu\n", ic, sent, remote_addr);
 911                         remote_addr += len;
 912
 913                         scat++;
 914                 }
 915
 916                 if (send->s_rdma_wr.wr.opcode == IB_WR_RDMA_READ_WITH_INV) {
 917                         send->s_rdma_wr.wr.num_sge = 1;
 918                         send->s_sge[0].addr = conn->c_xmit_rm->m_rs->rs_user_addr;
 919                         send->s_sge[0].length = conn->c_xmit_rm->m_rs->rs_user_bytes;
 920                         send->s_sge[0].lkey = ic->i_sends[fr_pos].s_mr->lkey;
 921                 }
 922
 923                 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
 924                         &send->s_rdma_wr,
 925                         send->s_rdma_wr.wr.num_sge,
 926                         send->s_rdma_wr.wr.next);
 927
 928                 prev = send;
 929                 if (++send == &ic->i_sends[ic->i_send_ring.w_nr])
 930                         send = ic->i_sends;
 931         }
 932
 933         /* if we finished the message then send completion owns it */
 934         if (scat == &op->op_sg[op->op_count])
 935                 first->s_rdma_wr.wr.send_flags = IB_SEND_SIGNALED;
 936
 937         if (i < work_alloc) {
 938                 rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc - i);
 939                 work_alloc = i;
 940         }
 941
 942         /* On iWARP, local memory access by a remote system (ie, RDMA Read) is not
 943          * recommended.  Putting the lkey on the wire is a security hole, as it can
 944          * allow for memory access to all of memory on the remote system.  Some
 945          * adapters do not allow using the lkey for this at all.  To bypass this use a
 946          * fastreg_mr (or possibly a dma_mr)
 947          */
 948         if (!op->op_write) {
 949                 ret = rds_iw_build_send_reg(&ic->i_sends[fr_pos],
 950                                             &op->op_sg[0], sg_nents);
 951                 if (ret) {
 952                         printk(KERN_WARNING "RDS/IW: failed to reg send mem\n");
 953                         goto out;
 954                 }
 955                 work_alloc++;
 956         }
 957
 958         failed_wr = &first->s_rdma_wr.wr;
 959         ret = ib_post_send(ic->i_cm_id->qp, &first->s_rdma_wr.wr, &failed_wr);
 960         rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
 961                  first, &first->s_rdma_wr, ret, failed_wr);
 962         BUG_ON(failed_wr != &first->s_rdma_wr.wr);
 963         if (ret) {
 964                 printk(KERN_WARNING "RDS/IW: rdma ib_post_send to %pI4 "
 965                        "returned %d\n", &conn->c_faddr, ret);
 966                 rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
 967                 goto out;
 968         }
 969
 970 out:
 971         return ret;
 972 }
 973
 974 void rds_iw_xmit_complete(struct rds_connection *conn)
 975 {
 976         struct rds_iw_connection *ic = conn->c_transport_data;
 977
 978         /* We may have a pending ACK or window update we were unable
 979          * to send previously (due to flow control). Try again. */
 980         rds_iw_attempt_ack(ic);
 981 }