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_wr.next = NULL;
 141                 send->s_wr.wr_id = i;
 142                 send->s_wr.sg_list = send->s_sge;
 143                 send->s_wr.num_sge = 1;
 144                 send->s_wr.opcode = IB_WR_SEND;
 145                 send->s_wr.send_flags = 0;
 146                 send->s_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                 send->s_page_list = ib_alloc_fast_reg_page_list(
 164                         ic->i_cm_id->device, fastreg_message_size);
 165                 if (IS_ERR(send->s_page_list)) {
 166                         printk(KERN_WARNING "RDS/IW: ib_alloc_fast_reg_page_list failed\n");
 167                         break;
 168                 }
 169         }
 170 }
 171
 172 void rds_iw_send_clear_ring(struct rds_iw_connection *ic)
 173 {
 174         struct rds_iw_send_work *send;
 175         u32 i;
 176
 177         for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
 178                 BUG_ON(!send->s_mr);
 179                 ib_dereg_mr(send->s_mr);
 180                 BUG_ON(!send->s_page_list);
 181                 ib_free_fast_reg_page_list(send->s_page_list);
 182                 if (send->s_wr.opcode == 0xdead)
 183                         continue;
 184                 if (send->s_rm)
 185                         rds_iw_send_unmap_rm(ic, send, IB_WC_WR_FLUSH_ERR);
 186                 if (send->s_op)
 187                         rds_iw_send_unmap_rdma(ic, send->s_op);
 188         }
 189 }
 190
 191 /*
 192  * The _oldest/_free ring operations here race cleanly with the alloc/unalloc
 193  * operations performed in the send path.  As the sender allocs and potentially
 194  * unallocs the next free entry in the ring it doesn't alter which is
 195  * the next to be freed, which is what this is concerned with.
 196  */
 197 void rds_iw_send_cq_comp_handler(struct ib_cq *cq, void *context)
 198 {
 199         struct rds_connection *conn = context;
 200         struct rds_iw_connection *ic = conn->c_transport_data;
 201         struct ib_wc wc;
 202         struct rds_iw_send_work *send;
 203         u32 completed;
 204         u32 oldest;
 205         u32 i;
 206         int ret;
 207
 208         rdsdebug("cq %p conn %p\n", cq, conn);
 209         rds_iw_stats_inc(s_iw_tx_cq_call);
 210         ret = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 211         if (ret)
 212                 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
 213
 214         while (ib_poll_cq(cq, 1, &wc) > 0) {
 215                 rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
 216                          (unsigned long long)wc.wr_id, wc.status, wc.byte_len,
 217                          be32_to_cpu(wc.ex.imm_data));
 218                 rds_iw_stats_inc(s_iw_tx_cq_event);
 219
 220                 if (wc.status != IB_WC_SUCCESS) {
 221                         printk(KERN_ERR "WC Error:  status = %d opcode = %d\n", wc.status, wc.opcode);
 222                         break;
 223                 }
 224
 225                 if (wc.opcode == IB_WC_LOCAL_INV && wc.wr_id == RDS_IW_LOCAL_INV_WR_ID) {
 226                         ic->i_fastreg_posted = 0;
 227                         continue;
 228                 }
 229
 230                 if (wc.opcode == IB_WC_FAST_REG_MR && wc.wr_id == RDS_IW_FAST_REG_WR_ID) {
 231                         ic->i_fastreg_posted = 1;
 232                         continue;
 233                 }
 234
 235                 if (wc.wr_id == RDS_IW_ACK_WR_ID) {
 236                         if (time_after(jiffies, ic->i_ack_queued + HZ/2))
 237                                 rds_iw_stats_inc(s_iw_tx_stalled);
 238                         rds_iw_ack_send_complete(ic);
 239                         continue;
 240                 }
 241
 242                 oldest = rds_iw_ring_oldest(&ic->i_send_ring);
 243
 244                 completed = rds_iw_ring_completed(&ic->i_send_ring, wc.wr_id, oldest);
 245
 246                 for (i = 0; i < completed; i++) {
 247                         send = &ic->i_sends[oldest];
 248
 249                         /* In the error case, wc.opcode sometimes contains garbage */
 250                         switch (send->s_wr.opcode) {
 251                         case IB_WR_SEND:
 252                                 if (send->s_rm)
 253                                         rds_iw_send_unmap_rm(ic, send, wc.status);
 254                                 break;
 255                         case IB_WR_FAST_REG_MR:
 256                         case IB_WR_RDMA_WRITE:
 257                         case IB_WR_RDMA_READ:
 258                         case IB_WR_RDMA_READ_WITH_INV:
 259                                 /* Nothing to be done - the SG list will be unmapped
 260                                  * when the SEND completes. */
 261                                 break;
 262                         default:
 263                                 printk_ratelimited(KERN_NOTICE
 264                                                 "RDS/IW: %s: unexpected opcode 0x%x in WR!\n",
 265                                                 __func__, send->s_wr.opcode);
 266                                 break;
 267                         }
 268
 269                         send->s_wr.opcode = 0xdead;
 270                         send->s_wr.num_sge = 1;
 271                         if (time_after(jiffies, send->s_queued + HZ/2))
 272                                 rds_iw_stats_inc(s_iw_tx_stalled);
 273
 274                         /* If a RDMA operation produced an error, signal this right
 275                          * away. If we don't, the subsequent SEND that goes with this
 276                          * RDMA will be canceled with ERR_WFLUSH, and the application
 277                          * never learn that the RDMA failed. */
 278                         if (unlikely(wc.status == IB_WC_REM_ACCESS_ERR && send->s_op)) {
 279                                 struct rds_message *rm;
 280
 281                                 rm = rds_send_get_message(conn, send->s_op);
 282                                 if (rm)
 283                                         rds_iw_send_rdma_complete(rm, wc.status);
 284                         }
 285
 286                         oldest = (oldest + 1) % ic->i_send_ring.w_nr;
 287                 }
 288
 289                 rds_iw_ring_free(&ic->i_send_ring, completed);
 290
 291                 if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
 292                     test_bit(0, &conn->c_map_queued))
 293                         queue_delayed_work(rds_wq, &conn->c_send_w, 0);
 294
 295                 /* We expect errors as the qp is drained during shutdown */
 296                 if (wc.status != IB_WC_SUCCESS && rds_conn_up(conn)) {
 297                         rds_iw_conn_error(conn,
 298                                 "send completion on %pI4 "
 299                                 "had status %u, disconnecting and reconnecting\n",
 300                                 &conn->c_faddr, wc.status);
 301                 }
 302         }
 303 }
 304
 305 /*
 306  * This is the main function for allocating credits when sending
 307  * messages.
 308  *
 309  * Conceptually, we have two counters:
 310  *  -   send credits: this tells us how many WRs we're allowed
 311  *      to submit without overruning the receiver's queue. For
 312  *      each SEND WR we post, we decrement this by one.
 313  *
 314  *  -   posted credits: this tells us how many WRs we recently
 315  *      posted to the receive queue. This value is transferred
 316  *      to the peer as a "credit update" in a RDS header field.
 317  *      Every time we transmit credits to the peer, we subtract
 318  *      the amount of transferred credits from this counter.
 319  *
 320  * It is essential that we avoid situations where both sides have
 321  * exhausted their send credits, and are unable to send new credits
 322  * to the peer. We achieve this by requiring that we send at least
 323  * one credit update to the peer before exhausting our credits.
 324  * When new credits arrive, we subtract one credit that is withheld
 325  * until we've posted new buffers and are ready to transmit these
 326  * credits (see rds_iw_send_add_credits below).
 327  *
 328  * The RDS send code is essentially single-threaded; rds_send_xmit
 329  * grabs c_send_lock to ensure exclusive access to the send ring.
 330  * However, the ACK sending code is independent and can race with
 331  * message SENDs.
 332  *
 333  * In the send path, we need to update the counters for send credits
 334  * and the counter of posted buffers atomically - when we use the
 335  * last available credit, we cannot allow another thread to race us
 336  * and grab the posted credits counter.  Hence, we have to use a
 337  * spinlock to protect the credit counter, or use atomics.
 338  *
 339  * Spinlocks shared between the send and the receive path are bad,
 340  * because they create unnecessary delays. An early implementation
 341  * using a spinlock showed a 5% degradation in throughput at some
 342  * loads.
 343  *
 344  * This implementation avoids spinlocks completely, putting both
 345  * counters into a single atomic, and updating that atomic using
 346  * atomic_add (in the receive path, when receiving fresh credits),
 347  * and using atomic_cmpxchg when updating the two counters.
 348  */
 349 int rds_iw_send_grab_credits(struct rds_iw_connection *ic,
 350                              u32 wanted, u32 *adv_credits, int need_posted, int max_posted)
 351 {
 352         unsigned int avail, posted, got = 0, advertise;
 353         long oldval, newval;
 354
 355         *adv_credits = 0;
 356         if (!ic->i_flowctl)
 357                 return wanted;
 358
 359 try_again:
 360         advertise = 0;
 361         oldval = newval = atomic_read(&ic->i_credits);
 362         posted = IB_GET_POST_CREDITS(oldval);
 363         avail = IB_GET_SEND_CREDITS(oldval);
 364
 365         rdsdebug("wanted=%u credits=%u posted=%u\n",
 366                         wanted, avail, posted);
 367
 368         /* The last credit must be used to send a credit update. */
 369         if (avail && !posted)
 370                 avail--;
 371
 372         if (avail < wanted) {
 373                 struct rds_connection *conn = ic->i_cm_id->context;
 374
 375                 /* Oops, there aren't that many credits left! */
 376                 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 377                 got = avail;
 378         } else {
 379                 /* Sometimes you get what you want, lalala. */
 380                 got = wanted;
 381         }
 382         newval -= IB_SET_SEND_CREDITS(got);
 383
 384         /*
 385          * If need_posted is non-zero, then the caller wants
 386          * the posted regardless of whether any send credits are
 387          * available.
 388          */
 389         if (posted && (got || need_posted)) {
 390                 advertise = min_t(unsigned int, posted, max_posted);
 391                 newval -= IB_SET_POST_CREDITS(advertise);
 392         }
 393
 394         /* Finally bill everything */
 395         if (atomic_cmpxchg(&ic->i_credits, oldval, newval) != oldval)
 396                 goto try_again;
 397
 398         *adv_credits = advertise;
 399         return got;
 400 }
 401
 402 void rds_iw_send_add_credits(struct rds_connection *conn, unsigned int credits)
 403 {
 404         struct rds_iw_connection *ic = conn->c_transport_data;
 405
 406         if (credits == 0)
 407                 return;
 408
 409         rdsdebug("credits=%u current=%u%s\n",
 410                         credits,
 411                         IB_GET_SEND_CREDITS(atomic_read(&ic->i_credits)),
 412                         test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ? ", ll_send_full" : "");
 413
 414         atomic_add(IB_SET_SEND_CREDITS(credits), &ic->i_credits);
 415         if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags))
 416                 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
 417
 418         WARN_ON(IB_GET_SEND_CREDITS(credits) >= 16384);
 419
 420         rds_iw_stats_inc(s_iw_rx_credit_updates);
 421 }
 422
 423 void rds_iw_advertise_credits(struct rds_connection *conn, unsigned int posted)
 424 {
 425         struct rds_iw_connection *ic = conn->c_transport_data;
 426
 427         if (posted == 0)
 428                 return;
 429
 430         atomic_add(IB_SET_POST_CREDITS(posted), &ic->i_credits);
 431
 432         /* Decide whether to send an update to the peer now.
 433          * If we would send a credit update for every single buffer we
 434          * post, we would end up with an ACK storm (ACK arrives,
 435          * consumes buffer, we refill the ring, send ACK to remote
 436          * advertising the newly posted buffer... ad inf)
 437          *
 438          * Performance pretty much depends on how often we send
 439          * credit updates - too frequent updates mean lots of ACKs.
 440          * Too infrequent updates, and the peer will run out of
 441          * credits and has to throttle.
 442          * For the time being, 16 seems to be a good compromise.
 443          */
 444         if (IB_GET_POST_CREDITS(atomic_read(&ic->i_credits)) >= 16)
 445                 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
 446 }
 447
 448 static inline void
 449 rds_iw_xmit_populate_wr(struct rds_iw_connection *ic,
 450                 struct rds_iw_send_work *send, unsigned int pos,
 451                 unsigned long buffer, unsigned int length,
 452                 int send_flags)
 453 {
 454         struct ib_sge *sge;
 455
 456         WARN_ON(pos != send - ic->i_sends);
 457
 458         send->s_wr.send_flags = send_flags;
 459         send->s_wr.opcode = IB_WR_SEND;
 460         send->s_wr.num_sge = 2;
 461         send->s_wr.next = NULL;
 462         send->s_queued = jiffies;
 463         send->s_op = NULL;
 464
 465         if (length != 0) {
 466                 sge = rds_iw_data_sge(ic, send->s_sge);
 467                 sge->addr = buffer;
 468                 sge->length = length;
 469                 sge->lkey = rds_iw_local_dma_lkey(ic);
 470
 471                 sge = rds_iw_header_sge(ic, send->s_sge);
 472         } else {
 473                 /* We're sending a packet with no payload. There is only
 474                  * one SGE */
 475                 send->s_wr.num_sge = 1;
 476                 sge = &send->s_sge[0];
 477         }
 478
 479         sge->addr = ic->i_send_hdrs_dma + (pos * sizeof(struct rds_header));
 480         sge->length = sizeof(struct rds_header);
 481         sge->lkey = rds_iw_local_dma_lkey(ic);
 482 }
 483
 484 /*
 485  * This can be called multiple times for a given message.  The first time
 486  * we see a message we map its scatterlist into the IB device so that
 487  * we can provide that mapped address to the IB scatter gather entries
 488  * in the IB work requests.  We translate the scatterlist into a series
 489  * of work requests that fragment the message.  These work requests complete
 490  * in order so we pass ownership of the message to the completion handler
 491  * once we send the final fragment.
 492  *
 493  * The RDS core uses the c_send_lock to only enter this function once
 494  * per connection.  This makes sure that the tx ring alloc/unalloc pairs
 495  * don't get out of sync and confuse the ring.
 496  */
 497 int rds_iw_xmit(struct rds_connection *conn, struct rds_message *rm,
 498                 unsigned int hdr_off, unsigned int sg, unsigned int off)
 499 {
 500         struct rds_iw_connection *ic = conn->c_transport_data;
 501         struct ib_device *dev = ic->i_cm_id->device;
 502         struct rds_iw_send_work *send = NULL;
 503         struct rds_iw_send_work *first;
 504         struct rds_iw_send_work *prev;
 505         struct ib_send_wr *failed_wr;
 506         struct scatterlist *scat;
 507         u32 pos;
 508         u32 i;
 509         u32 work_alloc;
 510         u32 credit_alloc;
 511         u32 posted;
 512         u32 adv_credits = 0;
 513         int send_flags = 0;
 514         int sent;
 515         int ret;
 516         int flow_controlled = 0;
 517
 518         BUG_ON(off % RDS_FRAG_SIZE);
 519         BUG_ON(hdr_off != 0 && hdr_off != sizeof(struct rds_header));
 520
 521         /* Fastreg support */
 522         if (rds_rdma_cookie_key(rm->m_rdma_cookie) && !ic->i_fastreg_posted) {
 523                 ret = -EAGAIN;
 524                 goto out;
 525         }
 526
 527         /* FIXME we may overallocate here */
 528         if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0)
 529                 i = 1;
 530         else
 531                 i = ceil(be32_to_cpu(rm->m_inc.i_hdr.h_len), RDS_FRAG_SIZE);
 532
 533         work_alloc = rds_iw_ring_alloc(&ic->i_send_ring, i, &pos);
 534         if (work_alloc == 0) {
 535                 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 536                 rds_iw_stats_inc(s_iw_tx_ring_full);
 537                 ret = -ENOMEM;
 538                 goto out;
 539         }
 540
 541         credit_alloc = work_alloc;
 542         if (ic->i_flowctl) {
 543                 credit_alloc = rds_iw_send_grab_credits(ic, work_alloc, &posted, 0, RDS_MAX_ADV_CREDIT);
 544                 adv_credits += posted;
 545                 if (credit_alloc < work_alloc) {
 546                         rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc - credit_alloc);
 547                         work_alloc = credit_alloc;
 548                         flow_controlled++;
 549                 }
 550                 if (work_alloc == 0) {
 551                         set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 552                         rds_iw_stats_inc(s_iw_tx_throttle);
 553                         ret = -ENOMEM;
 554                         goto out;
 555                 }
 556         }
 557
 558         /* map the message the first time we see it */
 559         if (!ic->i_rm) {
 560                 /*
 561                 printk(KERN_NOTICE "rds_iw_xmit prep msg dport=%u flags=0x%x len=%d\n",
 562                                 be16_to_cpu(rm->m_inc.i_hdr.h_dport),
 563                                 rm->m_inc.i_hdr.h_flags,
 564                                 be32_to_cpu(rm->m_inc.i_hdr.h_len));
 565                    */
 566                 if (rm->data.op_nents) {
 567                         rm->data.op_count = ib_dma_map_sg(dev,
 568                                                           rm->data.op_sg,
 569                                                           rm->data.op_nents,
 570                                                           DMA_TO_DEVICE);
 571                         rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->data.op_count);
 572                         if (rm->data.op_count == 0) {
 573                                 rds_iw_stats_inc(s_iw_tx_sg_mapping_failure);
 574                                 rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
 575                                 ret = -ENOMEM; /* XXX ? */
 576                                 goto out;
 577                         }
 578                 } else {
 579                         rm->data.op_count = 0;
 580                 }
 581
 582                 ic->i_unsignaled_wrs = rds_iw_sysctl_max_unsig_wrs;
 583                 ic->i_unsignaled_bytes = rds_iw_sysctl_max_unsig_bytes;
 584                 rds_message_addref(rm);
 585                 rm->data.op_dmasg = 0;
 586                 rm->data.op_dmaoff = 0;
 587                 ic->i_rm = rm;
 588
 589                 /* Finalize the header */
 590                 if (test_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags))
 591                         rm->m_inc.i_hdr.h_flags |= RDS_FLAG_ACK_REQUIRED;
 592                 if (test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))
 593                         rm->m_inc.i_hdr.h_flags |= RDS_FLAG_RETRANSMITTED;
 594
 595                 /* If it has a RDMA op, tell the peer we did it. This is
 596                  * used by the peer to release use-once RDMA MRs. */
 597                 if (rm->rdma.op_active) {
 598                         struct rds_ext_header_rdma ext_hdr;
 599
 600                         ext_hdr.h_rdma_rkey = cpu_to_be32(rm->rdma.op_rkey);
 601                         rds_message_add_extension(&rm->m_inc.i_hdr,
 602                                         RDS_EXTHDR_RDMA, &ext_hdr, sizeof(ext_hdr));
 603                 }
 604                 if (rm->m_rdma_cookie) {
 605                         rds_message_add_rdma_dest_extension(&rm->m_inc.i_hdr,
 606                                         rds_rdma_cookie_key(rm->m_rdma_cookie),
 607                                         rds_rdma_cookie_offset(rm->m_rdma_cookie));
 608                 }
 609
 610                 /* Note - rds_iw_piggyb_ack clears the ACK_REQUIRED bit, so
 611                  * we should not do this unless we have a chance of at least
 612                  * sticking the header into the send ring. Which is why we
 613                  * should call rds_iw_ring_alloc first. */
 614                 rm->m_inc.i_hdr.h_ack = cpu_to_be64(rds_iw_piggyb_ack(ic));
 615                 rds_message_make_checksum(&rm->m_inc.i_hdr);
 616
 617                 /*
 618                  * Update adv_credits since we reset the ACK_REQUIRED bit.
 619                  */
 620                 rds_iw_send_grab_credits(ic, 0, &posted, 1, RDS_MAX_ADV_CREDIT - adv_credits);
 621                 adv_credits += posted;
 622                 BUG_ON(adv_credits > 255);
 623         }
 624
 625         send = &ic->i_sends[pos];
 626         first = send;
 627         prev = NULL;
 628         scat = &rm->data.op_sg[rm->data.op_dmasg];
 629         sent = 0;
 630         i = 0;
 631
 632         /* Sometimes you want to put a fence between an RDMA
 633          * READ and the following SEND.
 634          * We could either do this all the time
 635          * or when requested by the user. Right now, we let
 636          * the application choose.
 637          */
 638         if (rm->rdma.op_active && rm->rdma.op_fence)
 639                 send_flags = IB_SEND_FENCE;
 640
 641         /*
 642          * We could be copying the header into the unused tail of the page.
 643          * That would need to be changed in the future when those pages might
 644          * be mapped userspace pages or page cache pages.  So instead we always
 645          * use a second sge and our long-lived ring of mapped headers.  We send
 646          * the header after the data so that the data payload can be aligned on
 647          * the receiver.
 648          */
 649
 650         /* handle a 0-len message */
 651         if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0) {
 652                 rds_iw_xmit_populate_wr(ic, send, pos, 0, 0, send_flags);
 653                 goto add_header;
 654         }
 655
 656         /* if there's data reference it with a chain of work reqs */
 657         for (; i < work_alloc && scat != &rm->data.op_sg[rm->data.op_count]; i++) {
 658                 unsigned int len;
 659
 660                 send = &ic->i_sends[pos];
 661
 662                 len = min(RDS_FRAG_SIZE,
 663                           ib_sg_dma_len(dev, scat) - rm->data.op_dmaoff);
 664                 rds_iw_xmit_populate_wr(ic, send, pos,
 665                         ib_sg_dma_address(dev, scat) + rm->data.op_dmaoff, len,
 666                         send_flags);
 667
 668                 /*
 669                  * We want to delay signaling completions just enough to get
 670                  * the batching benefits but not so much that we create dead time
 671                  * on the wire.
 672                  */
 673                 if (ic->i_unsignaled_wrs-- == 0) {
 674                         ic->i_unsignaled_wrs = rds_iw_sysctl_max_unsig_wrs;
 675                         send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
 676                 }
 677
 678                 ic->i_unsignaled_bytes -= len;
 679                 if (ic->i_unsignaled_bytes <= 0) {
 680                         ic->i_unsignaled_bytes = rds_iw_sysctl_max_unsig_bytes;
 681                         send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
 682                 }
 683
 684                 /*
 685                  * Always signal the last one if we're stopping due to flow control.
 686                  */
 687                 if (flow_controlled && i == (work_alloc-1))
 688                         send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
 689
 690                 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
 691                          &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
 692
 693                 sent += len;
 694                 rm->data.op_dmaoff += len;
 695                 if (rm->data.op_dmaoff == ib_sg_dma_len(dev, scat)) {
 696                         scat++;
 697                         rm->data.op_dmaoff = 0;
 698                         rm->data.op_dmasg++;
 699                 }
 700
 701 add_header:
 702                 /* Tack on the header after the data. The header SGE should already
 703                  * have been set up to point to the right header buffer. */
 704                 memcpy(&ic->i_send_hdrs[pos], &rm->m_inc.i_hdr, sizeof(struct rds_header));
 705
 706                 if (0) {
 707                         struct rds_header *hdr = &ic->i_send_hdrs[pos];
 708
 709                         printk(KERN_NOTICE "send WR dport=%u flags=0x%x len=%d\n",
 710                                 be16_to_cpu(hdr->h_dport),
 711                                 hdr->h_flags,
 712                                 be32_to_cpu(hdr->h_len));
 713                 }
 714                 if (adv_credits) {
 715                         struct rds_header *hdr = &ic->i_send_hdrs[pos];
 716
 717                         /* add credit and redo the header checksum */
 718                         hdr->h_credit = adv_credits;
 719                         rds_message_make_checksum(hdr);
 720                         adv_credits = 0;
 721                         rds_iw_stats_inc(s_iw_tx_credit_updates);
 722                 }
 723
 724                 if (prev)
 725                         prev->s_wr.next = &send->s_wr;
 726                 prev = send;
 727
 728                 pos = (pos + 1) % ic->i_send_ring.w_nr;
 729         }
 730
 731         /* Account the RDS header in the number of bytes we sent, but just once.
 732          * The caller has no concept of fragmentation. */
 733         if (hdr_off == 0)
 734                 sent += sizeof(struct rds_header);
 735
 736         /* if we finished the message then send completion owns it */
 737         if (scat == &rm->data.op_sg[rm->data.op_count]) {
 738                 prev->s_rm = ic->i_rm;
 739                 prev->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
 740                 ic->i_rm = NULL;
 741         }
 742
 743         if (i < work_alloc) {
 744                 rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc - i);
 745                 work_alloc = i;
 746         }
 747         if (ic->i_flowctl && i < credit_alloc)
 748                 rds_iw_send_add_credits(conn, credit_alloc - i);
 749
 750         /* XXX need to worry about failed_wr and partial sends. */
 751         failed_wr = &first->s_wr;
 752         ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
 753         rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
 754                  first, &first->s_wr, ret, failed_wr);
 755         BUG_ON(failed_wr != &first->s_wr);
 756         if (ret) {
 757                 printk(KERN_WARNING "RDS/IW: ib_post_send to %pI4 "
 758                        "returned %d\n", &conn->c_faddr, ret);
 759                 rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
 760                 if (prev->s_rm) {
 761                         ic->i_rm = prev->s_rm;
 762                         prev->s_rm = NULL;
 763                 }
 764                 goto out;
 765         }
 766
 767         ret = sent;
 768 out:
 769         BUG_ON(adv_credits);
 770         return ret;
 771 }
 772
 773 static void rds_iw_build_send_fastreg(struct rds_iw_device *rds_iwdev, struct rds_iw_connection *ic, struct rds_iw_send_work *send, int nent, int len, u64 sg_addr)
 774 {
 775         BUG_ON(nent > send->s_page_list->max_page_list_len);
 776         /*
 777          * Perform a WR for the fast_reg_mr. Each individual page
 778          * in the sg list is added to the fast reg page list and placed
 779          * inside the fast_reg_mr WR.
 780          */
 781         send->s_wr.opcode = IB_WR_FAST_REG_MR;
 782         send->s_wr.wr.fast_reg.length = len;
 783         send->s_wr.wr.fast_reg.rkey = send->s_mr->rkey;
 784         send->s_wr.wr.fast_reg.page_list = send->s_page_list;
 785         send->s_wr.wr.fast_reg.page_list_len = nent;
 786         send->s_wr.wr.fast_reg.page_shift = PAGE_SHIFT;
 787         send->s_wr.wr.fast_reg.access_flags = IB_ACCESS_REMOTE_WRITE;
 788         send->s_wr.wr.fast_reg.iova_start = sg_addr;
 789
 790         ib_update_fast_reg_key(send->s_mr, send->s_remap_count++);
 791 }
 792
 793 int rds_iw_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op)
 794 {
 795         struct rds_iw_connection *ic = conn->c_transport_data;
 796         struct rds_iw_send_work *send = NULL;
 797         struct rds_iw_send_work *first;
 798         struct rds_iw_send_work *prev;
 799         struct ib_send_wr *failed_wr;
 800         struct rds_iw_device *rds_iwdev;
 801         struct scatterlist *scat;
 802         unsigned long len;
 803         u64 remote_addr = op->op_remote_addr;
 804         u32 pos, fr_pos;
 805         u32 work_alloc;
 806         u32 i;
 807         u32 j;
 808         int sent;
 809         int ret;
 810         int num_sge;
 811
 812         rds_iwdev = ib_get_client_data(ic->i_cm_id->device, &rds_iw_client);
 813
 814         /* map the message the first time we see it */
 815         if (!op->op_mapped) {
 816                 op->op_count = ib_dma_map_sg(ic->i_cm_id->device,
 817                                              op->op_sg, op->op_nents, (op->op_write) ?
 818                                              DMA_TO_DEVICE : DMA_FROM_DEVICE);
 819                 rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->op_count);
 820                 if (op->op_count == 0) {
 821                         rds_iw_stats_inc(s_iw_tx_sg_mapping_failure);
 822                         ret = -ENOMEM; /* XXX ? */
 823                         goto out;
 824                 }
 825
 826                 op->op_mapped = 1;
 827         }
 828
 829         if (!op->op_write) {
 830                 /* Alloc space on the send queue for the fastreg */
 831                 work_alloc = rds_iw_ring_alloc(&ic->i_send_ring, 1, &fr_pos);
 832                 if (work_alloc != 1) {
 833                         rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
 834                         rds_iw_stats_inc(s_iw_tx_ring_full);
 835                         ret = -ENOMEM;
 836                         goto out;
 837                 }
 838         }
 839
 840         /*
 841          * Instead of knowing how to return a partial rdma read/write we insist that there
 842          * be enough work requests to send the entire message.
 843          */
 844         i = ceil(op->op_count, rds_iwdev->max_sge);
 845
 846         work_alloc = rds_iw_ring_alloc(&ic->i_send_ring, i, &pos);
 847         if (work_alloc != i) {
 848                 rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
 849                 rds_iw_stats_inc(s_iw_tx_ring_full);
 850                 ret = -ENOMEM;
 851                 goto out;
 852         }
 853
 854         send = &ic->i_sends[pos];
 855         if (!op->op_write) {
 856                 first = prev = &ic->i_sends[fr_pos];
 857         } else {
 858                 first = send;
 859                 prev = NULL;
 860         }
 861         scat = &op->op_sg[0];
 862         sent = 0;
 863         num_sge = op->op_count;
 864
 865         for (i = 0; i < work_alloc && scat != &op->op_sg[op->op_count]; i++) {
 866                 send->s_wr.send_flags = 0;
 867                 send->s_queued = jiffies;
 868
 869                 /*
 870                  * We want to delay signaling completions just enough to get
 871                  * the batching benefits but not so much that we create dead time on the wire.
 872                  */
 873                 if (ic->i_unsignaled_wrs-- == 0) {
 874                         ic->i_unsignaled_wrs = rds_iw_sysctl_max_unsig_wrs;
 875                         send->s_wr.send_flags = IB_SEND_SIGNALED;
 876                 }
 877
 878                 /* To avoid the need to have the plumbing to invalidate the fastreg_mr used
 879                  * for local access after RDS is finished with it, using
 880                  * IB_WR_RDMA_READ_WITH_INV will invalidate it after the read has completed.
 881                  */
 882                 if (op->op_write)
 883                         send->s_wr.opcode = IB_WR_RDMA_WRITE;
 884                 else
 885                         send->s_wr.opcode = IB_WR_RDMA_READ_WITH_INV;
 886
 887                 send->s_wr.wr.rdma.remote_addr = remote_addr;
 888                 send->s_wr.wr.rdma.rkey = op->op_rkey;
 889                 send->s_op = op;
 890
 891                 if (num_sge > rds_iwdev->max_sge) {
 892                         send->s_wr.num_sge = rds_iwdev->max_sge;
 893                         num_sge -= rds_iwdev->max_sge;
 894                 } else
 895                         send->s_wr.num_sge = num_sge;
 896
 897                 send->s_wr.next = NULL;
 898
 899                 if (prev)
 900                         prev->s_wr.next = &send->s_wr;
 901
 902                 for (j = 0; j < send->s_wr.num_sge && scat != &op->op_sg[op->op_count]; j++) {
 903                         len = ib_sg_dma_len(ic->i_cm_id->device, scat);
 904
 905                         if (send->s_wr.opcode == IB_WR_RDMA_READ_WITH_INV)
 906                                 send->s_page_list->page_list[j] = ib_sg_dma_address(ic->i_cm_id->device, scat);
 907                         else {
 908                                 send->s_sge[j].addr = ib_sg_dma_address(ic->i_cm_id->device, scat);
 909                                 send->s_sge[j].length = len;
 910                                 send->s_sge[j].lkey = rds_iw_local_dma_lkey(ic);
 911                         }
 912
 913                         sent += len;
 914                         rdsdebug("ic %p sent %d remote_addr %llu\n", ic, sent, remote_addr);
 915                         remote_addr += len;
 916
 917                         scat++;
 918                 }
 919
 920                 if (send->s_wr.opcode == IB_WR_RDMA_READ_WITH_INV) {
 921                         send->s_wr.num_sge = 1;
 922                         send->s_sge[0].addr = conn->c_xmit_rm->m_rs->rs_user_addr;
 923                         send->s_sge[0].length = conn->c_xmit_rm->m_rs->rs_user_bytes;
 924                         send->s_sge[0].lkey = ic->i_sends[fr_pos].s_mr->lkey;
 925                 }
 926
 927                 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
 928                         &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
 929
 930                 prev = send;
 931                 if (++send == &ic->i_sends[ic->i_send_ring.w_nr])
 932                         send = ic->i_sends;
 933         }
 934
 935         /* if we finished the message then send completion owns it */
 936         if (scat == &op->op_sg[op->op_count])
 937                 first->s_wr.send_flags = IB_SEND_SIGNALED;
 938
 939         if (i < work_alloc) {
 940                 rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc - i);
 941                 work_alloc = i;
 942         }
 943
 944         /* On iWARP, local memory access by a remote system (ie, RDMA Read) is not
 945          * recommended.  Putting the lkey on the wire is a security hole, as it can
 946          * allow for memory access to all of memory on the remote system.  Some
 947          * adapters do not allow using the lkey for this at all.  To bypass this use a
 948          * fastreg_mr (or possibly a dma_mr)
 949          */
 950         if (!op->op_write) {
 951                 rds_iw_build_send_fastreg(rds_iwdev, ic, &ic->i_sends[fr_pos],
 952                         op->op_count, sent, conn->c_xmit_rm->m_rs->rs_user_addr);
 953                 work_alloc++;
 954         }
 955
 956         failed_wr = &first->s_wr;
 957         ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
 958         rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
 959                  first, &first->s_wr, ret, failed_wr);
 960         BUG_ON(failed_wr != &first->s_wr);
 961         if (ret) {
 962                 printk(KERN_WARNING "RDS/IW: rdma ib_post_send to %pI4 "
 963                        "returned %d\n", &conn->c_faddr, ret);
 964                 rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
 965                 goto out;
 966         }
 967
 968 out:
 969         return ret;
 970 }
 971
 972 void rds_iw_xmit_complete(struct rds_connection *conn)
 973 {
 974         struct rds_iw_connection *ic = conn->c_transport_data;
 975
 976         /* We may have a pending ACK or window update we were unable
 977          * to send previously (due to flow control). Try again. */
 978         rds_iw_attempt_ack(ic);
 979 }