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_fast_reg_mr(ic->i_pd, fastreg_message_size);
 157                 if (IS_ERR(send->s_mr)) {
 158                         printk(KERN_WARNING "RDS/IW: ib_alloc_fast_reg_mr failed\n");
 159                         break;
 160                 }
 161
 162                 send->s_page_list = ib_alloc_fast_reg_page_list(
 163                         ic->i_cm_id->device, fastreg_message_size);
 164                 if (IS_ERR(send->s_page_list)) {
 165                         printk(KERN_WARNING "RDS/IW: ib_alloc_fast_reg_page_list failed\n");
 166                         break;
 167                 }
 168         }
 169 }
 170
 171 void rds_iw_send_clear_ring(struct rds_iw_connection *ic)
 172 {
 173         struct rds_iw_send_work *send;
 174         u32 i;
 175
 176         for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
 177                 BUG_ON(!send->s_mr);
 178                 ib_dereg_mr(send->s_mr);
 179                 BUG_ON(!send->s_page_list);
 180                 ib_free_fast_reg_page_list(send->s_page_list);
 181                 if (send->s_wr.opcode == 0xdead)
 182                         continue;
 183                 if (send->s_rm)
 184                         rds_iw_send_unmap_rm(ic, send, IB_WC_WR_FLUSH_ERR);
 185                 if (send->s_op)
 186                         rds_iw_send_unmap_rdma(ic, send->s_op);
 187         }
 188 }
 189
 190 /*
 191  * The _oldest/_free ring operations here race cleanly with the alloc/unalloc
 192  * operations performed in the send path.  As the sender allocs and potentially
 193  * unallocs the next free entry in the ring it doesn't alter which is
 194  * the next to be freed, which is what this is concerned with.
 195  */
 196 void rds_iw_send_cq_comp_handler(struct ib_cq *cq, void *context)
 197 {
 198         struct rds_connection *conn = context;
 199         struct rds_iw_connection *ic = conn->c_transport_data;
 200         struct ib_wc wc;
 201         struct rds_iw_send_work *send;
 202         u32 completed;
 203         u32 oldest;
 204         u32 i;
 205         int ret;
 206
 207         rdsdebug("cq %p conn %p\n", cq, conn);
 208         rds_iw_stats_inc(s_iw_tx_cq_call);
 209         ret = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 210         if (ret)
 211                 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
 212
 213         while (ib_poll_cq(cq, 1, &wc) > 0) {
 214                 rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
 215                          (unsigned long long)wc.wr_id, wc.status, wc.byte_len,
 216                          be32_to_cpu(wc.ex.imm_data));
 217                 rds_iw_stats_inc(s_iw_tx_cq_event);
 218
 219                 if (wc.status != IB_WC_SUCCESS) {
 220                         printk(KERN_ERR "WC Error:  status = %d opcode = %d\n", wc.status, wc.opcode);
 221                         break;
 222                 }
 223
 224                 if (wc.opcode == IB_WC_LOCAL_INV && wc.wr_id == RDS_IW_LOCAL_INV_WR_ID) {
 225                         ic->i_fastreg_posted = 0;
 226                         continue;
 227                 }
 228
 229                 if (wc.opcode == IB_WC_FAST_REG_MR && wc.wr_id == RDS_IW_FAST_REG_WR_ID) {
 230                         ic->i_fastreg_posted = 1;
 231                         continue;
 232                 }
 233
 234                 if (wc.wr_id == RDS_IW_ACK_WR_ID) {
 235                         if (ic->i_ack_queued + HZ/2 < jiffies)
 236                                 rds_iw_stats_inc(s_iw_tx_stalled);
 237                         rds_iw_ack_send_complete(ic);
 238                         continue;
 239                 }
 240
 241                 oldest = rds_iw_ring_oldest(&ic->i_send_ring);
 242
 243                 completed = rds_iw_ring_completed(&ic->i_send_ring, wc.wr_id, oldest);
 244
 245                 for (i = 0; i < completed; i++) {
 246                         send = &ic->i_sends[oldest];
 247
 248                         /* In the error case, wc.opcode sometimes contains garbage */
 249                         switch (send->s_wr.opcode) {
 250                         case IB_WR_SEND:
 251                                 if (send->s_rm)
 252                                         rds_iw_send_unmap_rm(ic, send, wc.status);
 253                                 break;
 254                         case IB_WR_FAST_REG_MR:
 255                         case IB_WR_RDMA_WRITE:
 256                         case IB_WR_RDMA_READ:
 257                         case IB_WR_RDMA_READ_WITH_INV:
 258                                 /* Nothing to be done - the SG list will be unmapped
 259                                  * when the SEND completes. */
 260                                 break;
 261                         default:
 262                                 printk_ratelimited(KERN_NOTICE
 263                                                 "RDS/IW: %s: unexpected opcode 0x%x in WR!\n",
 264                                                 __func__, send->s_wr.opcode);
 265                                 break;
 266                         }
 267
 268                         send->s_wr.opcode = 0xdead;
 269                         send->s_wr.num_sge = 1;
 270                         if (send->s_queued + HZ/2 < jiffies)
 271                                 rds_iw_stats_inc(s_iw_tx_stalled);
 272
 273                         /* If a RDMA operation produced an error, signal this right
 274                          * away. If we don't, the subsequent SEND that goes with this
 275                          * RDMA will be canceled with ERR_WFLUSH, and the application
 276                          * never learn that the RDMA failed. */
 277                         if (unlikely(wc.status == IB_WC_REM_ACCESS_ERR && send->s_op)) {
 278                                 struct rds_message *rm;
 279
 280                                 rm = rds_send_get_message(conn, send->s_op);
 281                                 if (rm)
 282                                         rds_iw_send_rdma_complete(rm, wc.status);
 283                         }
 284
 285                         oldest = (oldest + 1) % ic->i_send_ring.w_nr;
 286                 }
 287
 288                 rds_iw_ring_free(&ic->i_send_ring, completed);
 289
 290                 if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
 291                     test_bit(0, &conn->c_map_queued))
 292                         queue_delayed_work(rds_wq, &conn->c_send_w, 0);
 293
 294                 /* We expect errors as the qp is drained during shutdown */
 295                 if (wc.status != IB_WC_SUCCESS && rds_conn_up(conn)) {
 296                         rds_iw_conn_error(conn,
 297                                 "send completion on %pI4 "
 298                                 "had status %u, disconnecting and reconnecting\n",
 299                                 &conn->c_faddr, wc.status);
 300                 }
 301         }
 302 }
 303
 304 /*
 305  * This is the main function for allocating credits when sending
 306  * messages.
 307  *
 308  * Conceptually, we have two counters:
 309  *  -   send credits: this tells us how many WRs we're allowed
 310  *      to submit without overruning the receiver's queue. For
 311  *      each SEND WR we post, we decrement this by one.
 312  *
 313  *  -   posted credits: this tells us how many WRs we recently
 314  *      posted to the receive queue. This value is transferred
 315  *      to the peer as a "credit update" in a RDS header field.
 316  *      Every time we transmit credits to the peer, we subtract
 317  *      the amount of transferred credits from this counter.
 318  *
 319  * It is essential that we avoid situations where both sides have
 320  * exhausted their send credits, and are unable to send new credits
 321  * to the peer. We achieve this by requiring that we send at least
 322  * one credit update to the peer before exhausting our credits.
 323  * When new credits arrive, we subtract one credit that is withheld
 324  * until we've posted new buffers and are ready to transmit these
 325  * credits (see rds_iw_send_add_credits below).
 326  *
 327  * The RDS send code is essentially single-threaded; rds_send_xmit
 328  * grabs c_send_lock to ensure exclusive access to the send ring.
 329  * However, the ACK sending code is independent and can race with
 330  * message SENDs.
 331  *
 332  * In the send path, we need to update the counters for send credits
 333  * and the counter of posted buffers atomically - when we use the
 334  * last available credit, we cannot allow another thread to race us
 335  * and grab the posted credits counter.  Hence, we have to use a
 336  * spinlock to protect the credit counter, or use atomics.
 337  *
 338  * Spinlocks shared between the send and the receive path are bad,
 339  * because they create unnecessary delays. An early implementation
 340  * using a spinlock showed a 5% degradation in throughput at some
 341  * loads.
 342  *
 343  * This implementation avoids spinlocks completely, putting both
 344  * counters into a single atomic, and updating that atomic using
 345  * atomic_add (in the receive path, when receiving fresh credits),
 346  * and using atomic_cmpxchg when updating the two counters.
 347  */
 348 int rds_iw_send_grab_credits(struct rds_iw_connection *ic,
 349                              u32 wanted, u32 *adv_credits, int need_posted, int max_posted)
 350 {
 351         unsigned int avail, posted, got = 0, advertise;
 352         long oldval, newval;
 353
 354         *adv_credits = 0;
 355         if (!ic->i_flowctl)
 356                 return wanted;
 357
 358 try_again:
 359         advertise = 0;
 360         oldval = newval = atomic_read(&ic->i_credits);
 361         posted = IB_GET_POST_CREDITS(oldval);
 362         avail = IB_GET_SEND_CREDITS(oldval);
 363
 364         rdsdebug("rds_iw_send_grab_credits(%u): credits=%u posted=%u\n",
 365                         wanted, avail, posted);
 366
 367         /* The last credit must be used to send a credit update. */
 368         if (avail && !posted)
 369                 avail--;
 370
 371         if (avail < wanted) {
 372                 struct rds_connection *conn = ic->i_cm_id->context;
 373
 374                 /* Oops, there aren't that many credits left! */
 375                 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 376                 got = avail;
 377         } else {
 378                 /* Sometimes you get what you want, lalala. */
 379                 got = wanted;
 380         }
 381         newval -= IB_SET_SEND_CREDITS(got);
 382
 383         /*
 384          * If need_posted is non-zero, then the caller wants
 385          * the posted regardless of whether any send credits are
 386          * available.
 387          */
 388         if (posted && (got || need_posted)) {
 389                 advertise = min_t(unsigned int, posted, max_posted);
 390                 newval -= IB_SET_POST_CREDITS(advertise);
 391         }
 392
 393         /* Finally bill everything */
 394         if (atomic_cmpxchg(&ic->i_credits, oldval, newval) != oldval)
 395                 goto try_again;
 396
 397         *adv_credits = advertise;
 398         return got;
 399 }
 400
 401 void rds_iw_send_add_credits(struct rds_connection *conn, unsigned int credits)
 402 {
 403         struct rds_iw_connection *ic = conn->c_transport_data;
 404
 405         if (credits == 0)
 406                 return;
 407
 408         rdsdebug("rds_iw_send_add_credits(%u): current=%u%s\n",
 409                         credits,
 410                         IB_GET_SEND_CREDITS(atomic_read(&ic->i_credits)),
 411                         test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ? ", ll_send_full" : "");
 412
 413         atomic_add(IB_SET_SEND_CREDITS(credits), &ic->i_credits);
 414         if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags))
 415                 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
 416
 417         WARN_ON(IB_GET_SEND_CREDITS(credits) >= 16384);
 418
 419         rds_iw_stats_inc(s_iw_rx_credit_updates);
 420 }
 421
 422 void rds_iw_advertise_credits(struct rds_connection *conn, unsigned int posted)
 423 {
 424         struct rds_iw_connection *ic = conn->c_transport_data;
 425
 426         if (posted == 0)
 427                 return;
 428
 429         atomic_add(IB_SET_POST_CREDITS(posted), &ic->i_credits);
 430
 431         /* Decide whether to send an update to the peer now.
 432          * If we would send a credit update for every single buffer we
 433          * post, we would end up with an ACK storm (ACK arrives,
 434          * consumes buffer, we refill the ring, send ACK to remote
 435          * advertising the newly posted buffer... ad inf)
 436          *
 437          * Performance pretty much depends on how often we send
 438          * credit updates - too frequent updates mean lots of ACKs.
 439          * Too infrequent updates, and the peer will run out of
 440          * credits and has to throttle.
 441          * For the time being, 16 seems to be a good compromise.
 442          */
 443         if (IB_GET_POST_CREDITS(atomic_read(&ic->i_credits)) >= 16)
 444                 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
 445 }
 446
 447 static inline void
 448 rds_iw_xmit_populate_wr(struct rds_iw_connection *ic,
 449                 struct rds_iw_send_work *send, unsigned int pos,
 450                 unsigned long buffer, unsigned int length,
 451                 int send_flags)
 452 {
 453         struct ib_sge *sge;
 454
 455         WARN_ON(pos != send - ic->i_sends);
 456
 457         send->s_wr.send_flags = send_flags;
 458         send->s_wr.opcode = IB_WR_SEND;
 459         send->s_wr.num_sge = 2;
 460         send->s_wr.next = NULL;
 461         send->s_queued = jiffies;
 462         send->s_op = NULL;
 463
 464         if (length != 0) {
 465                 sge = rds_iw_data_sge(ic, send->s_sge);
 466                 sge->addr = buffer;
 467                 sge->length = length;
 468                 sge->lkey = rds_iw_local_dma_lkey(ic);
 469
 470                 sge = rds_iw_header_sge(ic, send->s_sge);
 471         } else {
 472                 /* We're sending a packet with no payload. There is only
 473                  * one SGE */
 474                 send->s_wr.num_sge = 1;
 475                 sge = &send->s_sge[0];
 476         }
 477
 478         sge->addr = ic->i_send_hdrs_dma + (pos * sizeof(struct rds_header));
 479         sge->length = sizeof(struct rds_header);
 480         sge->lkey = rds_iw_local_dma_lkey(ic);
 481 }
 482
 483 /*
 484  * This can be called multiple times for a given message.  The first time
 485  * we see a message we map its scatterlist into the IB device so that
 486  * we can provide that mapped address to the IB scatter gather entries
 487  * in the IB work requests.  We translate the scatterlist into a series
 488  * of work requests that fragment the message.  These work requests complete
 489  * in order so we pass ownership of the message to the completion handler
 490  * once we send the final fragment.
 491  *
 492  * The RDS core uses the c_send_lock to only enter this function once
 493  * per connection.  This makes sure that the tx ring alloc/unalloc pairs
 494  * don't get out of sync and confuse the ring.
 495  */
 496 int rds_iw_xmit(struct rds_connection *conn, struct rds_message *rm,
 497                 unsigned int hdr_off, unsigned int sg, unsigned int off)
 498 {
 499         struct rds_iw_connection *ic = conn->c_transport_data;
 500         struct ib_device *dev = ic->i_cm_id->device;
 501         struct rds_iw_send_work *send = NULL;
 502         struct rds_iw_send_work *first;
 503         struct rds_iw_send_work *prev;
 504         struct ib_send_wr *failed_wr;
 505         struct scatterlist *scat;
 506         u32 pos;
 507         u32 i;
 508         u32 work_alloc;
 509         u32 credit_alloc;
 510         u32 posted;
 511         u32 adv_credits = 0;
 512         int send_flags = 0;
 513         int sent;
 514         int ret;
 515         int flow_controlled = 0;
 516
 517         BUG_ON(off % RDS_FRAG_SIZE);
 518         BUG_ON(hdr_off != 0 && hdr_off != sizeof(struct rds_header));
 519
 520         /* Fastreg support */
 521         if (rds_rdma_cookie_key(rm->m_rdma_cookie) && !ic->i_fastreg_posted) {
 522                 ret = -EAGAIN;
 523                 goto out;
 524         }
 525
 526         /* FIXME we may overallocate here */
 527         if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0)
 528                 i = 1;
 529         else
 530                 i = ceil(be32_to_cpu(rm->m_inc.i_hdr.h_len), RDS_FRAG_SIZE);
 531
 532         work_alloc = rds_iw_ring_alloc(&ic->i_send_ring, i, &pos);
 533         if (work_alloc == 0) {
 534                 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 535                 rds_iw_stats_inc(s_iw_tx_ring_full);
 536                 ret = -ENOMEM;
 537                 goto out;
 538         }
 539
 540         credit_alloc = work_alloc;
 541         if (ic->i_flowctl) {
 542                 credit_alloc = rds_iw_send_grab_credits(ic, work_alloc, &posted, 0, RDS_MAX_ADV_CREDIT);
 543                 adv_credits += posted;
 544                 if (credit_alloc < work_alloc) {
 545                         rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc - credit_alloc);
 546                         work_alloc = credit_alloc;
 547                         flow_controlled++;
 548                 }
 549                 if (work_alloc == 0) {
 550                         set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 551                         rds_iw_stats_inc(s_iw_tx_throttle);
 552                         ret = -ENOMEM;
 553                         goto out;
 554                 }
 555         }
 556
 557         /* map the message the first time we see it */
 558         if (!ic->i_rm) {
 559                 /*
 560                 printk(KERN_NOTICE "rds_iw_xmit prep msg dport=%u flags=0x%x len=%d\n",
 561                                 be16_to_cpu(rm->m_inc.i_hdr.h_dport),
 562                                 rm->m_inc.i_hdr.h_flags,
 563                                 be32_to_cpu(rm->m_inc.i_hdr.h_len));
 564                    */
 565                 if (rm->data.op_nents) {
 566                         rm->data.op_count = ib_dma_map_sg(dev,
 567                                                           rm->data.op_sg,
 568                                                           rm->data.op_nents,
 569                                                           DMA_TO_DEVICE);
 570                         rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->data.op_count);
 571                         if (rm->data.op_count == 0) {
 572                                 rds_iw_stats_inc(s_iw_tx_sg_mapping_failure);
 573                                 rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
 574                                 ret = -ENOMEM; /* XXX ? */
 575                                 goto out;
 576                         }
 577                 } else {
 578                         rm->data.op_count = 0;
 579                 }
 580
 581                 ic->i_unsignaled_wrs = rds_iw_sysctl_max_unsig_wrs;
 582                 ic->i_unsignaled_bytes = rds_iw_sysctl_max_unsig_bytes;
 583                 rds_message_addref(rm);
 584                 ic->i_rm = rm;
 585
 586                 /* Finalize the header */
 587                 if (test_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags))
 588                         rm->m_inc.i_hdr.h_flags |= RDS_FLAG_ACK_REQUIRED;
 589                 if (test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))
 590                         rm->m_inc.i_hdr.h_flags |= RDS_FLAG_RETRANSMITTED;
 591
 592                 /* If it has a RDMA op, tell the peer we did it. This is
 593                  * used by the peer to release use-once RDMA MRs. */
 594                 if (rm->rdma.op_active) {
 595                         struct rds_ext_header_rdma ext_hdr;
 596
 597                         ext_hdr.h_rdma_rkey = cpu_to_be32(rm->rdma.op_rkey);
 598                         rds_message_add_extension(&rm->m_inc.i_hdr,
 599                                         RDS_EXTHDR_RDMA, &ext_hdr, sizeof(ext_hdr));
 600                 }
 601                 if (rm->m_rdma_cookie) {
 602                         rds_message_add_rdma_dest_extension(&rm->m_inc.i_hdr,
 603                                         rds_rdma_cookie_key(rm->m_rdma_cookie),
 604                                         rds_rdma_cookie_offset(rm->m_rdma_cookie));
 605                 }
 606
 607                 /* Note - rds_iw_piggyb_ack clears the ACK_REQUIRED bit, so
 608                  * we should not do this unless we have a chance of at least
 609                  * sticking the header into the send ring. Which is why we
 610                  * should call rds_iw_ring_alloc first. */
 611                 rm->m_inc.i_hdr.h_ack = cpu_to_be64(rds_iw_piggyb_ack(ic));
 612                 rds_message_make_checksum(&rm->m_inc.i_hdr);
 613
 614                 /*
 615                  * Update adv_credits since we reset the ACK_REQUIRED bit.
 616                  */
 617                 rds_iw_send_grab_credits(ic, 0, &posted, 1, RDS_MAX_ADV_CREDIT - adv_credits);
 618                 adv_credits += posted;
 619                 BUG_ON(adv_credits > 255);
 620         }
 621
 622         send = &ic->i_sends[pos];
 623         first = send;
 624         prev = NULL;
 625         scat = &rm->data.op_sg[sg];
 626         sent = 0;
 627         i = 0;
 628
 629         /* Sometimes you want to put a fence between an RDMA
 630          * READ and the following SEND.
 631          * We could either do this all the time
 632          * or when requested by the user. Right now, we let
 633          * the application choose.
 634          */
 635         if (rm->rdma.op_active && rm->rdma.op_fence)
 636                 send_flags = IB_SEND_FENCE;
 637
 638         /*
 639          * We could be copying the header into the unused tail of the page.
 640          * That would need to be changed in the future when those pages might
 641          * be mapped userspace pages or page cache pages.  So instead we always
 642          * use a second sge and our long-lived ring of mapped headers.  We send
 643          * the header after the data so that the data payload can be aligned on
 644          * the receiver.
 645          */
 646
 647         /* handle a 0-len message */
 648         if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0) {
 649                 rds_iw_xmit_populate_wr(ic, send, pos, 0, 0, send_flags);
 650                 goto add_header;
 651         }
 652
 653         /* if there's data reference it with a chain of work reqs */
 654         for (; i < work_alloc && scat != &rm->data.op_sg[rm->data.op_count]; i++) {
 655                 unsigned int len;
 656
 657                 send = &ic->i_sends[pos];
 658
 659                 len = min(RDS_FRAG_SIZE, ib_sg_dma_len(dev, scat) - off);
 660                 rds_iw_xmit_populate_wr(ic, send, pos,
 661                                 ib_sg_dma_address(dev, scat) + off, len,
 662                                 send_flags);
 663
 664                 /*
 665                  * We want to delay signaling completions just enough to get
 666                  * the batching benefits but not so much that we create dead time
 667                  * on the wire.
 668                  */
 669                 if (ic->i_unsignaled_wrs-- == 0) {
 670                         ic->i_unsignaled_wrs = rds_iw_sysctl_max_unsig_wrs;
 671                         send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
 672                 }
 673
 674                 ic->i_unsignaled_bytes -= len;
 675                 if (ic->i_unsignaled_bytes <= 0) {
 676                         ic->i_unsignaled_bytes = rds_iw_sysctl_max_unsig_bytes;
 677                         send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
 678                 }
 679
 680                 /*
 681                  * Always signal the last one if we're stopping due to flow control.
 682                  */
 683                 if (flow_controlled && i == (work_alloc-1))
 684                         send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
 685
 686                 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
 687                          &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
 688
 689                 sent += len;
 690                 off += len;
 691                 if (off == ib_sg_dma_len(dev, scat)) {
 692                         scat++;
 693                         off = 0;
 694                 }
 695
 696 add_header:
 697                 /* Tack on the header after the data. The header SGE should already
 698                  * have been set up to point to the right header buffer. */
 699                 memcpy(&ic->i_send_hdrs[pos], &rm->m_inc.i_hdr, sizeof(struct rds_header));
 700
 701                 if (0) {
 702                         struct rds_header *hdr = &ic->i_send_hdrs[pos];
 703
 704                         printk(KERN_NOTICE "send WR dport=%u flags=0x%x len=%d\n",
 705                                 be16_to_cpu(hdr->h_dport),
 706                                 hdr->h_flags,
 707                                 be32_to_cpu(hdr->h_len));
 708                 }
 709                 if (adv_credits) {
 710                         struct rds_header *hdr = &ic->i_send_hdrs[pos];
 711
 712                         /* add credit and redo the header checksum */
 713                         hdr->h_credit = adv_credits;
 714                         rds_message_make_checksum(hdr);
 715                         adv_credits = 0;
 716                         rds_iw_stats_inc(s_iw_tx_credit_updates);
 717                 }
 718
 719                 if (prev)
 720                         prev->s_wr.next = &send->s_wr;
 721                 prev = send;
 722
 723                 pos = (pos + 1) % ic->i_send_ring.w_nr;
 724         }
 725
 726         /* Account the RDS header in the number of bytes we sent, but just once.
 727          * The caller has no concept of fragmentation. */
 728         if (hdr_off == 0)
 729                 sent += sizeof(struct rds_header);
 730
 731         /* if we finished the message then send completion owns it */
 732         if (scat == &rm->data.op_sg[rm->data.op_count]) {
 733                 prev->s_rm = ic->i_rm;
 734                 prev->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
 735                 ic->i_rm = NULL;
 736         }
 737
 738         if (i < work_alloc) {
 739                 rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc - i);
 740                 work_alloc = i;
 741         }
 742         if (ic->i_flowctl && i < credit_alloc)
 743                 rds_iw_send_add_credits(conn, credit_alloc - i);
 744
 745         /* XXX need to worry about failed_wr and partial sends. */
 746         failed_wr = &first->s_wr;
 747         ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
 748         rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
 749                  first, &first->s_wr, ret, failed_wr);
 750         BUG_ON(failed_wr != &first->s_wr);
 751         if (ret) {
 752                 printk(KERN_WARNING "RDS/IW: ib_post_send to %pI4 "
 753                        "returned %d\n", &conn->c_faddr, ret);
 754                 rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
 755                 if (prev->s_rm) {
 756                         ic->i_rm = prev->s_rm;
 757                         prev->s_rm = NULL;
 758                 }
 759                 goto out;
 760         }
 761
 762         ret = sent;
 763 out:
 764         BUG_ON(adv_credits);
 765         return ret;
 766 }
 767
 768 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)
 769 {
 770         BUG_ON(nent > send->s_page_list->max_page_list_len);
 771         /*
 772          * Perform a WR for the fast_reg_mr. Each individual page
 773          * in the sg list is added to the fast reg page list and placed
 774          * inside the fast_reg_mr WR.
 775          */
 776         send->s_wr.opcode = IB_WR_FAST_REG_MR;
 777         send->s_wr.wr.fast_reg.length = len;
 778         send->s_wr.wr.fast_reg.rkey = send->s_mr->rkey;
 779         send->s_wr.wr.fast_reg.page_list = send->s_page_list;
 780         send->s_wr.wr.fast_reg.page_list_len = nent;
 781         send->s_wr.wr.fast_reg.page_shift = PAGE_SHIFT;
 782         send->s_wr.wr.fast_reg.access_flags = IB_ACCESS_REMOTE_WRITE;
 783         send->s_wr.wr.fast_reg.iova_start = sg_addr;
 784
 785         ib_update_fast_reg_key(send->s_mr, send->s_remap_count++);
 786 }
 787
 788 int rds_iw_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op)
 789 {
 790         struct rds_iw_connection *ic = conn->c_transport_data;
 791         struct rds_iw_send_work *send = NULL;
 792         struct rds_iw_send_work *first;
 793         struct rds_iw_send_work *prev;
 794         struct ib_send_wr *failed_wr;
 795         struct rds_iw_device *rds_iwdev;
 796         struct scatterlist *scat;
 797         unsigned long len;
 798         u64 remote_addr = op->op_remote_addr;
 799         u32 pos, fr_pos;
 800         u32 work_alloc;
 801         u32 i;
 802         u32 j;
 803         int sent;
 804         int ret;
 805         int num_sge;
 806
 807         rds_iwdev = ib_get_client_data(ic->i_cm_id->device, &rds_iw_client);
 808
 809         /* map the message the first time we see it */
 810         if (!op->op_mapped) {
 811                 op->op_count = ib_dma_map_sg(ic->i_cm_id->device,
 812                                              op->op_sg, op->op_nents, (op->op_write) ?
 813                                              DMA_TO_DEVICE : DMA_FROM_DEVICE);
 814                 rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->op_count);
 815                 if (op->op_count == 0) {
 816                         rds_iw_stats_inc(s_iw_tx_sg_mapping_failure);
 817                         ret = -ENOMEM; /* XXX ? */
 818                         goto out;
 819                 }
 820
 821                 op->op_mapped = 1;
 822         }
 823
 824         if (!op->op_write) {
 825                 /* Alloc space on the send queue for the fastreg */
 826                 work_alloc = rds_iw_ring_alloc(&ic->i_send_ring, 1, &fr_pos);
 827                 if (work_alloc != 1) {
 828                         rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
 829                         rds_iw_stats_inc(s_iw_tx_ring_full);
 830                         ret = -ENOMEM;
 831                         goto out;
 832                 }
 833         }
 834
 835         /*
 836          * Instead of knowing how to return a partial rdma read/write we insist that there
 837          * be enough work requests to send the entire message.
 838          */
 839         i = ceil(op->op_count, rds_iwdev->max_sge);
 840
 841         work_alloc = rds_iw_ring_alloc(&ic->i_send_ring, i, &pos);
 842         if (work_alloc != i) {
 843                 rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
 844                 rds_iw_stats_inc(s_iw_tx_ring_full);
 845                 ret = -ENOMEM;
 846                 goto out;
 847         }
 848
 849         send = &ic->i_sends[pos];
 850         if (!op->op_write) {
 851                 first = prev = &ic->i_sends[fr_pos];
 852         } else {
 853                 first = send;
 854                 prev = NULL;
 855         }
 856         scat = &op->op_sg[0];
 857         sent = 0;
 858         num_sge = op->op_count;
 859
 860         for (i = 0; i < work_alloc && scat != &op->op_sg[op->op_count]; i++) {
 861                 send->s_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_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_wr.opcode = IB_WR_RDMA_WRITE;
 879                 else
 880                         send->s_wr.opcode = IB_WR_RDMA_READ_WITH_INV;
 881
 882                 send->s_wr.wr.rdma.remote_addr = remote_addr;
 883                 send->s_wr.wr.rdma.rkey = op->op_rkey;
 884                 send->s_op = op;
 885
 886                 if (num_sge > rds_iwdev->max_sge) {
 887                         send->s_wr.num_sge = rds_iwdev->max_sge;
 888                         num_sge -= rds_iwdev->max_sge;
 889                 } else
 890                         send->s_wr.num_sge = num_sge;
 891
 892                 send->s_wr.next = NULL;
 893
 894                 if (prev)
 895                         prev->s_wr.next = &send->s_wr;
 896
 897                 for (j = 0; j < send->s_wr.num_sge && scat != &op->op_sg[op->op_count]; j++) {
 898                         len = ib_sg_dma_len(ic->i_cm_id->device, scat);
 899
 900                         if (send->s_wr.opcode == IB_WR_RDMA_READ_WITH_INV)
 901                                 send->s_page_list->page_list[j] = ib_sg_dma_address(ic->i_cm_id->device, scat);
 902                         else {
 903                                 send->s_sge[j].addr = ib_sg_dma_address(ic->i_cm_id->device, scat);
 904                                 send->s_sge[j].length = len;
 905                                 send->s_sge[j].lkey = rds_iw_local_dma_lkey(ic);
 906                         }
 907
 908                         sent += len;
 909                         rdsdebug("ic %p sent %d remote_addr %llu\n", ic, sent, remote_addr);
 910                         remote_addr += len;
 911
 912                         scat++;
 913                 }
 914
 915                 if (send->s_wr.opcode == IB_WR_RDMA_READ_WITH_INV) {
 916                         send->s_wr.num_sge = 1;
 917                         send->s_sge[0].addr = conn->c_xmit_rm->m_rs->rs_user_addr;
 918                         send->s_sge[0].length = conn->c_xmit_rm->m_rs->rs_user_bytes;
 919                         send->s_sge[0].lkey = ic->i_sends[fr_pos].s_mr->lkey;
 920                 }
 921
 922                 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
 923                         &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
 924
 925                 prev = send;
 926                 if (++send == &ic->i_sends[ic->i_send_ring.w_nr])
 927                         send = ic->i_sends;
 928         }
 929
 930         /* if we finished the message then send completion owns it */
 931         if (scat == &op->op_sg[op->op_count])
 932                 first->s_wr.send_flags = IB_SEND_SIGNALED;
 933
 934         if (i < work_alloc) {
 935                 rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc - i);
 936                 work_alloc = i;
 937         }
 938
 939         /* On iWARP, local memory access by a remote system (ie, RDMA Read) is not
 940          * recommended.  Putting the lkey on the wire is a security hole, as it can
 941          * allow for memory access to all of memory on the remote system.  Some
 942          * adapters do not allow using the lkey for this at all.  To bypass this use a
 943          * fastreg_mr (or possibly a dma_mr)
 944          */
 945         if (!op->op_write) {
 946                 rds_iw_build_send_fastreg(rds_iwdev, ic, &ic->i_sends[fr_pos],
 947                         op->op_count, sent, conn->c_xmit_rm->m_rs->rs_user_addr);
 948                 work_alloc++;
 949         }
 950
 951         failed_wr = &first->s_wr;
 952         ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
 953         rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
 954                  first, &first->s_wr, ret, failed_wr);
 955         BUG_ON(failed_wr != &first->s_wr);
 956         if (ret) {
 957                 printk(KERN_WARNING "RDS/IW: rdma ib_post_send to %pI4 "
 958                        "returned %d\n", &conn->c_faddr, ret);
 959                 rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
 960                 goto out;
 961         }
 962
 963 out:
 964         return ret;
 965 }
 966
 967 void rds_iw_xmit_complete(struct rds_connection *conn)
 968 {
 969         struct rds_iw_connection *ic = conn->c_transport_data;
 970
 971         /* We may have a pending ACK or window update we were unable
 972          * to send previously (due to flow control). Try again. */
 973         rds_iw_attempt_ack(ic);
 974 }