net/rds/ib_send.c

   1 /*
   2  * Copyright (c) 2006 Oracle.  All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  *
  32  */
  33 #include <linux/kernel.h>
  34 #include <linux/in.h>
  35 #include <linux/device.h>
  36 #include <linux/dmapool.h>
  37
  38 #include "rds.h"
  39 #include "ib.h"
  40
  41 static void rds_ib_send_rdma_complete(struct rds_message *rm,
  42                                       int wc_status)
  43 {
  44         int notify_status;
  45
  46         switch (wc_status) {
  47         case IB_WC_WR_FLUSH_ERR:
  48                 return;
  49
  50         case IB_WC_SUCCESS:
  51                 notify_status = RDS_RDMA_SUCCESS;
  52                 break;
  53
  54         case IB_WC_REM_ACCESS_ERR:
  55                 notify_status = RDS_RDMA_REMOTE_ERROR;
  56                 break;
  57
  58         default:
  59                 notify_status = RDS_RDMA_OTHER_ERROR;
  60                 break;
  61         }
  62         rds_rdma_send_complete(rm, notify_status);
  63 }
  64
  65 static void rds_ib_send_atomic_complete(struct rds_message *rm,
  66                                       int wc_status)
  67 {
  68         int notify_status;
  69
  70         if (wc_status != IB_WC_SUCCESS)
  71                 notify_status = RDS_RDMA_OTHER_ERROR;
  72         else
  73                 notify_status = RDS_RDMA_SUCCESS;
  74
  75         rds_atomic_send_complete(rm, notify_status);
  76 }
  77
  78 static void rds_ib_send_unmap_rm(struct rds_ib_connection *ic,
  79                           struct rds_ib_send_work *send,
  80                           int wc_status)
  81 {
  82         struct rds_message *rm = send->s_rm;
  83
  84         rdsdebug("ic %p send %p rm %p\n", ic, send, rm);
  85
  86         ib_dma_unmap_sg(ic->i_cm_id->device,
  87                         rm->data.m_sg, rm->data.m_nents,
  88                         DMA_TO_DEVICE);
  89
  90         if (rm->rdma.m_rdma_op.r_active) {
  91                 struct rds_rdma_op *op = &rm->rdma.m_rdma_op;
  92
  93                 if (op->r_mapped) {
  94                         ib_dma_unmap_sg(ic->i_cm_id->device,
  95                                         op->r_sg, op->r_nents,
  96                                         op->r_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
  97                         op->r_mapped = 0;
  98                 }
  99
 100                 /* If the user asked for a completion notification on this
 101                  * message, we can implement three different semantics:
 102                  *  1.  Notify when we received the ACK on the RDS message
 103                  *      that was queued with the RDMA. This provides reliable
 104                  *      notification of RDMA status at the expense of a one-way
 105                  *      packet delay.
 106                  *  2.  Notify when the IB stack gives us the completion event for
 107                  *      the RDMA operation.
 108                  *  3.  Notify when the IB stack gives us the completion event for
 109                  *      the accompanying RDS messages.
 110                  * Here, we implement approach #3. To implement approach #2,
 111                  * call rds_rdma_send_complete from the cq_handler. To implement #1,
 112                  * don't call rds_rdma_send_complete at all, and fall back to the notify
 113                  * handling in the ACK processing code.
 114                  *
 115                  * Note: There's no need to explicitly sync any RDMA buffers using
 116                  * ib_dma_sync_sg_for_cpu - the completion for the RDMA
 117                  * operation itself unmapped the RDMA buffers, which takes care
 118                  * of synching.
 119                  */
 120                 rds_ib_send_rdma_complete(rm, wc_status);
 121
 122                 if (rm->rdma.m_rdma_op.r_write)
 123                         rds_stats_add(s_send_rdma_bytes, rm->rdma.m_rdma_op.r_bytes);
 124                 else
 125                         rds_stats_add(s_recv_rdma_bytes, rm->rdma.m_rdma_op.r_bytes);
 126         }
 127
 128         if (rm->atomic.op_active) {
 129                 struct rm_atomic_op *op = &rm->atomic;
 130
 131                 /* unmap atomic recvbuf */
 132                 if (op->op_mapped) {
 133                         ib_dma_unmap_sg(ic->i_cm_id->device, op->op_sg, 1,
 134                                         DMA_FROM_DEVICE);
 135                         op->op_mapped = 0;
 136                 }
 137
 138                 rds_ib_send_atomic_complete(rm, wc_status);
 139
 140                 if (rm->atomic.op_type == RDS_ATOMIC_TYPE_CSWP)
 141                         rds_stats_inc(s_atomic_cswp);
 142                 else
 143                         rds_stats_inc(s_atomic_fadd);
 144         }
 145
 146         /* If anyone waited for this message to get flushed out, wake
 147          * them up now */
 148         rds_message_unmapped(rm);
 149
 150         rds_message_put(rm);
 151         send->s_rm = NULL;
 152 }
 153
 154 void rds_ib_send_init_ring(struct rds_ib_connection *ic)
 155 {
 156         struct rds_ib_send_work *send;
 157         u32 i;
 158
 159         for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
 160                 struct ib_sge *sge;
 161
 162                 send->s_rm = NULL;
 163                 send->s_op = NULL;
 164
 165                 send->s_wr.wr_id = i;
 166                 send->s_wr.sg_list = send->s_sge;
 167                 send->s_wr.num_sge = 1;
 168                 send->s_wr.opcode = IB_WR_SEND;
 169                 send->s_wr.send_flags = 0;
 170                 send->s_wr.ex.imm_data = 0;
 171
 172                 sge = rds_ib_data_sge(ic, send->s_sge);
 173                 sge->lkey = ic->i_mr->lkey;
 174
 175                 sge = rds_ib_header_sge(ic, send->s_sge);
 176                 sge->addr = ic->i_send_hdrs_dma + (i * sizeof(struct rds_header));
 177                 sge->length = sizeof(struct rds_header);
 178                 sge->lkey = ic->i_mr->lkey;
 179         }
 180 }
 181
 182 void rds_ib_send_clear_ring(struct rds_ib_connection *ic)
 183 {
 184         struct rds_ib_send_work *send;
 185         u32 i;
 186
 187         for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
 188                 if (!send->s_rm || send->s_wr.opcode == 0xdead)
 189                         continue;
 190                 rds_ib_send_unmap_rm(ic, send, IB_WC_WR_FLUSH_ERR);
 191         }
 192 }
 193
 194 /*
 195  * The _oldest/_free ring operations here race cleanly with the alloc/unalloc
 196  * operations performed in the send path.  As the sender allocs and potentially
 197  * unallocs the next free entry in the ring it doesn't alter which is
 198  * the next to be freed, which is what this is concerned with.
 199  */
 200 void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context)
 201 {
 202         struct rds_connection *conn = context;
 203         struct rds_ib_connection *ic = conn->c_transport_data;
 204         struct ib_wc wc;
 205         struct rds_ib_send_work *send;
 206         u32 completed;
 207         u32 oldest;
 208         u32 i = 0;
 209         int ret;
 210
 211         rdsdebug("cq %p conn %p\n", cq, conn);
 212         rds_ib_stats_inc(s_ib_tx_cq_call);
 213         ret = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 214         if (ret)
 215                 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
 216
 217         while (ib_poll_cq(cq, 1, &wc) > 0) {
 218                 rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
 219                          (unsigned long long)wc.wr_id, wc.status, wc.byte_len,
 220                          be32_to_cpu(wc.ex.imm_data));
 221                 rds_ib_stats_inc(s_ib_tx_cq_event);
 222
 223                 if (wc.wr_id == RDS_IB_ACK_WR_ID) {
 224                         if (ic->i_ack_queued + HZ/2 < jiffies)
 225                                 rds_ib_stats_inc(s_ib_tx_stalled);
 226                         rds_ib_ack_send_complete(ic);
 227                         continue;
 228                 }
 229
 230                 oldest = rds_ib_ring_oldest(&ic->i_send_ring);
 231
 232                 completed = rds_ib_ring_completed(&ic->i_send_ring, wc.wr_id, oldest);
 233
 234                 for (i = 0; i < completed; i++) {
 235                         send = &ic->i_sends[oldest];
 236
 237                         /* In the error case, wc.opcode sometimes contains garbage */
 238                         switch (send->s_wr.opcode) {
 239                         case IB_WR_SEND:
 240                                 if (send->s_rm)
 241                                         rds_ib_send_unmap_rm(ic, send, wc.status);
 242                                 break;
 243                         case IB_WR_RDMA_WRITE:
 244                         case IB_WR_RDMA_READ:
 245                         case IB_WR_ATOMIC_FETCH_AND_ADD:
 246                         case IB_WR_ATOMIC_CMP_AND_SWP:
 247                                 /* Nothing to be done - the SG list will be unmapped
 248                                  * when the SEND completes. */
 249                                 break;
 250                         default:
 251                                 if (printk_ratelimit())
 252                                         printk(KERN_NOTICE
 253                                                 "RDS/IB: %s: unexpected opcode 0x%x in WR!\n",
 254                                                 __func__, send->s_wr.opcode);
 255                                 break;
 256                         }
 257
 258                         send->s_wr.opcode = 0xdead;
 259                         send->s_wr.num_sge = 1;
 260                         if (send->s_queued + HZ/2 < jiffies)
 261                                 rds_ib_stats_inc(s_ib_tx_stalled);
 262
 263                         /* If a RDMA operation produced an error, signal this right
 264                          * away. If we don't, the subsequent SEND that goes with this
 265                          * RDMA will be canceled with ERR_WFLUSH, and the application
 266                          * never learn that the RDMA failed. */
 267                         if (unlikely(wc.status == IB_WC_REM_ACCESS_ERR && send->s_op)) {
 268                                 struct rds_message *rm;
 269
 270                                 rm = rds_send_get_message(conn, send->s_op);
 271                                 if (rm) {
 272                                         rds_ib_send_unmap_rm(ic, send, wc.status);
 273                                         rds_ib_send_rdma_complete(rm, wc.status);
 274                                         rds_message_put(rm);
 275                                 }
 276                         }
 277
 278                         oldest = (oldest + 1) % ic->i_send_ring.w_nr;
 279                 }
 280
 281                 rds_ib_ring_free(&ic->i_send_ring, completed);
 282
 283                 if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
 284                     test_bit(0, &conn->c_map_queued))
 285                         queue_delayed_work(rds_wq, &conn->c_send_w, 0);
 286
 287                 /* We expect errors as the qp is drained during shutdown */
 288                 if (wc.status != IB_WC_SUCCESS && rds_conn_up(conn)) {
 289                         rds_ib_conn_error(conn,
 290                                 "send completion on %pI4 "
 291                                 "had status %u, disconnecting and reconnecting\n",
 292                                 &conn->c_faddr, wc.status);
 293                 }
 294         }
 295 }
 296
 297 /*
 298  * This is the main function for allocating credits when sending
 299  * messages.
 300  *
 301  * Conceptually, we have two counters:
 302  *  -   send credits: this tells us how many WRs we're allowed
 303  *      to submit without overruning the reciever's queue. For
 304  *      each SEND WR we post, we decrement this by one.
 305  *
 306  *  -   posted credits: this tells us how many WRs we recently
 307  *      posted to the receive queue. This value is transferred
 308  *      to the peer as a "credit update" in a RDS header field.
 309  *      Every time we transmit credits to the peer, we subtract
 310  *      the amount of transferred credits from this counter.
 311  *
 312  * It is essential that we avoid situations where both sides have
 313  * exhausted their send credits, and are unable to send new credits
 314  * to the peer. We achieve this by requiring that we send at least
 315  * one credit update to the peer before exhausting our credits.
 316  * When new credits arrive, we subtract one credit that is withheld
 317  * until we've posted new buffers and are ready to transmit these
 318  * credits (see rds_ib_send_add_credits below).
 319  *
 320  * The RDS send code is essentially single-threaded; rds_send_xmit
 321  * grabs c_send_lock to ensure exclusive access to the send ring.
 322  * However, the ACK sending code is independent and can race with
 323  * message SENDs.
 324  *
 325  * In the send path, we need to update the counters for send credits
 326  * and the counter of posted buffers atomically - when we use the
 327  * last available credit, we cannot allow another thread to race us
 328  * and grab the posted credits counter.  Hence, we have to use a
 329  * spinlock to protect the credit counter, or use atomics.
 330  *
 331  * Spinlocks shared between the send and the receive path are bad,
 332  * because they create unnecessary delays. An early implementation
 333  * using a spinlock showed a 5% degradation in throughput at some
 334  * loads.
 335  *
 336  * This implementation avoids spinlocks completely, putting both
 337  * counters into a single atomic, and updating that atomic using
 338  * atomic_add (in the receive path, when receiving fresh credits),
 339  * and using atomic_cmpxchg when updating the two counters.
 340  */
 341 int rds_ib_send_grab_credits(struct rds_ib_connection *ic,
 342                              u32 wanted, u32 *adv_credits, int need_posted, int max_posted)
 343 {
 344         unsigned int avail, posted, got = 0, advertise;
 345         long oldval, newval;
 346
 347         *adv_credits = 0;
 348         if (!ic->i_flowctl)
 349                 return wanted;
 350
 351 try_again:
 352         advertise = 0;
 353         oldval = newval = atomic_read(&ic->i_credits);
 354         posted = IB_GET_POST_CREDITS(oldval);
 355         avail = IB_GET_SEND_CREDITS(oldval);
 356
 357         rdsdebug("rds_ib_send_grab_credits(%u): credits=%u posted=%u\n",
 358                         wanted, avail, posted);
 359
 360         /* The last credit must be used to send a credit update. */
 361         if (avail && !posted)
 362                 avail--;
 363
 364         if (avail < wanted) {
 365                 struct rds_connection *conn = ic->i_cm_id->context;
 366
 367                 /* Oops, there aren't that many credits left! */
 368                 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 369                 got = avail;
 370         } else {
 371                 /* Sometimes you get what you want, lalala. */
 372                 got = wanted;
 373         }
 374         newval -= IB_SET_SEND_CREDITS(got);
 375
 376         /*
 377          * If need_posted is non-zero, then the caller wants
 378          * the posted regardless of whether any send credits are
 379          * available.
 380          */
 381         if (posted && (got || need_posted)) {
 382                 advertise = min_t(unsigned int, posted, max_posted);
 383                 newval -= IB_SET_POST_CREDITS(advertise);
 384         }
 385
 386         /* Finally bill everything */
 387         if (atomic_cmpxchg(&ic->i_credits, oldval, newval) != oldval)
 388                 goto try_again;
 389
 390         *adv_credits = advertise;
 391         return got;
 392 }
 393
 394 void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits)
 395 {
 396         struct rds_ib_connection *ic = conn->c_transport_data;
 397
 398         if (credits == 0)
 399                 return;
 400
 401         rdsdebug("rds_ib_send_add_credits(%u): current=%u%s\n",
 402                         credits,
 403                         IB_GET_SEND_CREDITS(atomic_read(&ic->i_credits)),
 404                         test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ? ", ll_send_full" : "");
 405
 406         atomic_add(IB_SET_SEND_CREDITS(credits), &ic->i_credits);
 407         if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags))
 408                 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
 409
 410         WARN_ON(IB_GET_SEND_CREDITS(credits) >= 16384);
 411
 412         rds_ib_stats_inc(s_ib_rx_credit_updates);
 413 }
 414
 415 void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted)
 416 {
 417         struct rds_ib_connection *ic = conn->c_transport_data;
 418
 419         if (posted == 0)
 420                 return;
 421
 422         atomic_add(IB_SET_POST_CREDITS(posted), &ic->i_credits);
 423
 424         /* Decide whether to send an update to the peer now.
 425          * If we would send a credit update for every single buffer we
 426          * post, we would end up with an ACK storm (ACK arrives,
 427          * consumes buffer, we refill the ring, send ACK to remote
 428          * advertising the newly posted buffer... ad inf)
 429          *
 430          * Performance pretty much depends on how often we send
 431          * credit updates - too frequent updates mean lots of ACKs.
 432          * Too infrequent updates, and the peer will run out of
 433          * credits and has to throttle.
 434          * For the time being, 16 seems to be a good compromise.
 435          */
 436         if (IB_GET_POST_CREDITS(atomic_read(&ic->i_credits)) >= 16)
 437                 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
 438 }
 439
 440 static inline void
 441 rds_ib_xmit_populate_wr(struct rds_ib_connection *ic,
 442                 struct rds_ib_send_work *send, unsigned int pos,
 443                 unsigned long buffer, unsigned int length,
 444                 int send_flags)
 445 {
 446         struct ib_sge *sge;
 447
 448         WARN_ON(pos != send - ic->i_sends);
 449
 450         send->s_wr.send_flags = send_flags;
 451         send->s_wr.opcode = IB_WR_SEND;
 452         send->s_wr.num_sge = 2;
 453         send->s_wr.next = NULL;
 454         send->s_queued = jiffies;
 455         send->s_op = NULL;
 456
 457         if (length != 0) {
 458                 sge = rds_ib_data_sge(ic, send->s_sge);
 459                 sge->addr = buffer;
 460                 sge->length = length;
 461                 sge->lkey = ic->i_mr->lkey;
 462
 463                 sge = rds_ib_header_sge(ic, send->s_sge);
 464         } else {
 465                 /* We're sending a packet with no payload. There is only
 466                  * one SGE */
 467                 send->s_wr.num_sge = 1;
 468                 sge = &send->s_sge[0];
 469         }
 470
 471         sge->addr = ic->i_send_hdrs_dma + (pos * sizeof(struct rds_header));
 472         sge->length = sizeof(struct rds_header);
 473         sge->lkey = ic->i_mr->lkey;
 474 }
 475
 476 /*
 477  * This can be called multiple times for a given message.  The first time
 478  * we see a message we map its scatterlist into the IB device so that
 479  * we can provide that mapped address to the IB scatter gather entries
 480  * in the IB work requests.  We translate the scatterlist into a series
 481  * of work requests that fragment the message.  These work requests complete
 482  * in order so we pass ownership of the message to the completion handler
 483  * once we send the final fragment.
 484  *
 485  * The RDS core uses the c_send_lock to only enter this function once
 486  * per connection.  This makes sure that the tx ring alloc/unalloc pairs
 487  * don't get out of sync and confuse the ring.
 488  */
 489 int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
 490                 unsigned int hdr_off, unsigned int sg, unsigned int off)
 491 {
 492         struct rds_ib_connection *ic = conn->c_transport_data;
 493         struct ib_device *dev = ic->i_cm_id->device;
 494         struct rds_ib_send_work *send = NULL;
 495         struct rds_ib_send_work *first;
 496         struct rds_ib_send_work *prev;
 497         struct ib_send_wr *failed_wr;
 498         struct scatterlist *scat;
 499         u32 pos;
 500         u32 i;
 501         u32 work_alloc;
 502         u32 credit_alloc;
 503         u32 posted;
 504         u32 adv_credits = 0;
 505         int send_flags = 0;
 506         int sent;
 507         int ret;
 508         int flow_controlled = 0;
 509
 510         BUG_ON(off % RDS_FRAG_SIZE);
 511         BUG_ON(hdr_off != 0 && hdr_off != sizeof(struct rds_header));
 512
 513         /* Do not send cong updates to IB loopback */
 514         if (conn->c_loopback
 515             && rm->m_inc.i_hdr.h_flags & RDS_FLAG_CONG_BITMAP) {
 516                 rds_cong_map_updated(conn->c_fcong, ~(u64) 0);
 517                 return sizeof(struct rds_header) + RDS_CONG_MAP_BYTES;
 518         }
 519
 520         /* FIXME we may overallocate here */
 521         if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0)
 522                 i = 1;
 523         else
 524                 i = ceil(be32_to_cpu(rm->m_inc.i_hdr.h_len), RDS_FRAG_SIZE);
 525
 526         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
 527         if (work_alloc == 0) {
 528                 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 529                 rds_ib_stats_inc(s_ib_tx_ring_full);
 530                 ret = -ENOMEM;
 531                 goto out;
 532         }
 533
 534         credit_alloc = work_alloc;
 535         if (ic->i_flowctl) {
 536                 credit_alloc = rds_ib_send_grab_credits(ic, work_alloc, &posted, 0, RDS_MAX_ADV_CREDIT);
 537                 adv_credits += posted;
 538                 if (credit_alloc < work_alloc) {
 539                         rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - credit_alloc);
 540                         work_alloc = credit_alloc;
 541                         flow_controlled++;
 542                 }
 543                 if (work_alloc == 0) {
 544                         set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 545                         rds_ib_stats_inc(s_ib_tx_throttle);
 546                         ret = -ENOMEM;
 547                         goto out;
 548                 }
 549         }
 550
 551         /* map the message the first time we see it */
 552         if (!ic->i_rm) {
 553                 /*
 554                 printk(KERN_NOTICE "rds_ib_xmit prep msg dport=%u flags=0x%x len=%d\n",
 555                                 be16_to_cpu(rm->m_inc.i_hdr.h_dport),
 556                                 rm->m_inc.i_hdr.h_flags,
 557                                 be32_to_cpu(rm->m_inc.i_hdr.h_len));
 558                    */
 559                 if (rm->data.m_nents) {
 560                         rm->data.m_count = ib_dma_map_sg(dev,
 561                                                             rm->data.m_sg,
 562                                                             rm->data.m_nents,
 563                                                             DMA_TO_DEVICE);
 564                         rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->data.m_count);
 565                         if (rm->data.m_count == 0) {
 566                                 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
 567                                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 568                                 ret = -ENOMEM; /* XXX ? */
 569                                 goto out;
 570                         }
 571                 } else {
 572                         rm->data.m_count = 0;
 573                 }
 574
 575                 ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
 576                 ic->i_unsignaled_bytes = rds_ib_sysctl_max_unsig_bytes;
 577                 rds_message_addref(rm);
 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.m_rdma_op.r_active) {
 589                         struct rds_ext_header_rdma ext_hdr;
 590
 591                         ext_hdr.h_rdma_rkey = cpu_to_be32(rm->rdma.m_rdma_op.r_key);
 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_ib_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_ib_ring_alloc first. */
 605                 rm->m_inc.i_hdr.h_ack = cpu_to_be64(rds_ib_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_ib_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.m_sg[sg];
 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.m_rdma_op.r_active && rm->rdma.m_rdma_op.r_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_ib_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.m_sg[rm->data.m_count]; i++) {
 649                 unsigned int len;
 650
 651                 send = &ic->i_sends[pos];
 652
 653                 len = min(RDS_FRAG_SIZE, ib_sg_dma_len(dev, scat) - off);
 654                 rds_ib_xmit_populate_wr(ic, send, pos,
 655                                 ib_sg_dma_address(dev, scat) + off, len,
 656                                 send_flags);
 657
 658                 /*
 659                  * We want to delay signaling completions just enough to get
 660                  * the batching benefits but not so much that we create dead time
 661                  * on the wire.
 662                  */
 663                 if (ic->i_unsignaled_wrs-- == 0) {
 664                         ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
 665                         send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
 666                 }
 667
 668                 ic->i_unsignaled_bytes -= len;
 669                 if (ic->i_unsignaled_bytes <= 0) {
 670                         ic->i_unsignaled_bytes = rds_ib_sysctl_max_unsig_bytes;
 671                         send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
 672                 }
 673
 674                 /*
 675                  * Always signal the last one if we're stopping due to flow control.
 676                  */
 677                 if (flow_controlled && i == (work_alloc-1))
 678                         send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
 679
 680                 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
 681                          &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
 682
 683                 sent += len;
 684                 off += len;
 685                 if (off == ib_sg_dma_len(dev, scat)) {
 686                         scat++;
 687                         off = 0;
 688                 }
 689
 690 add_header:
 691                 /* Tack on the header after the data. The header SGE should already
 692                  * have been set up to point to the right header buffer. */
 693                 memcpy(&ic->i_send_hdrs[pos], &rm->m_inc.i_hdr, sizeof(struct rds_header));
 694
 695                 if (0) {
 696                         struct rds_header *hdr = &ic->i_send_hdrs[pos];
 697
 698                         printk(KERN_NOTICE "send WR dport=%u flags=0x%x len=%d\n",
 699                                 be16_to_cpu(hdr->h_dport),
 700                                 hdr->h_flags,
 701                                 be32_to_cpu(hdr->h_len));
 702                 }
 703                 if (adv_credits) {
 704                         struct rds_header *hdr = &ic->i_send_hdrs[pos];
 705
 706                         /* add credit and redo the header checksum */
 707                         hdr->h_credit = adv_credits;
 708                         rds_message_make_checksum(hdr);
 709                         adv_credits = 0;
 710                         rds_ib_stats_inc(s_ib_tx_credit_updates);
 711                 }
 712
 713                 if (prev)
 714                         prev->s_wr.next = &send->s_wr;
 715                 prev = send;
 716
 717                 pos = (pos + 1) % ic->i_send_ring.w_nr;
 718         }
 719
 720         /* Account the RDS header in the number of bytes we sent, but just once.
 721          * The caller has no concept of fragmentation. */
 722         if (hdr_off == 0)
 723                 sent += sizeof(struct rds_header);
 724
 725         /* if we finished the message then send completion owns it */
 726         if (scat == &rm->data.m_sg[rm->data.m_count]) {
 727                 prev->s_rm = ic->i_rm;
 728                 prev->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
 729                 ic->i_rm = NULL;
 730         }
 731
 732         if (i < work_alloc) {
 733                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
 734                 work_alloc = i;
 735         }
 736         if (ic->i_flowctl && i < credit_alloc)
 737                 rds_ib_send_add_credits(conn, credit_alloc - i);
 738
 739         /* XXX need to worry about failed_wr and partial sends. */
 740         failed_wr = &first->s_wr;
 741         ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
 742         rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
 743                  first, &first->s_wr, ret, failed_wr);
 744         BUG_ON(failed_wr != &first->s_wr);
 745         if (ret) {
 746                 printk(KERN_WARNING "RDS/IB: ib_post_send to %pI4 "
 747                        "returned %d\n", &conn->c_faddr, ret);
 748                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 749                 if (prev->s_rm) {
 750                         ic->i_rm = prev->s_rm;
 751                         prev->s_rm = NULL;
 752                 }
 753
 754                 rds_ib_conn_error(ic->conn, "ib_post_send failed\n");
 755                 goto out;
 756         }
 757
 758         ret = sent;
 759 out:
 760         BUG_ON(adv_credits);
 761         return ret;
 762 }
 763
 764 /*
 765  * Issue atomic operation.
 766  * A simplified version of the rdma case, we always map 1 SG, and
 767  * only 8 bytes, for the return value from the atomic operation.
 768  */
 769 int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op)
 770 {
 771         struct rds_ib_connection *ic = conn->c_transport_data;
 772         struct rds_ib_send_work *send = NULL;
 773         struct ib_send_wr *failed_wr;
 774         struct rds_ib_device *rds_ibdev;
 775         u32 pos;
 776         u32 work_alloc;
 777         int ret;
 778
 779         rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
 780
 781         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, 1, &pos);
 782         if (work_alloc != 1) {
 783                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 784                 rds_ib_stats_inc(s_ib_tx_ring_full);
 785                 ret = -ENOMEM;
 786                 goto out;
 787         }
 788
 789         /* address of send request in ring */
 790         send = &ic->i_sends[pos];
 791         send->s_queued = jiffies;
 792
 793         if (op->op_type == RDS_ATOMIC_TYPE_CSWP) {
 794                 send->s_wr.opcode = IB_WR_ATOMIC_CMP_AND_SWP;
 795                 send->s_wr.wr.atomic.compare_add = op->op_compare;
 796                 send->s_wr.wr.atomic.swap = op->op_swap_add;
 797         } else { /* FADD */
 798                 send->s_wr.opcode = IB_WR_ATOMIC_FETCH_AND_ADD;
 799                 send->s_wr.wr.atomic.compare_add = op->op_swap_add;
 800                 send->s_wr.wr.atomic.swap = 0;
 801         }
 802         send->s_wr.send_flags = IB_SEND_SIGNALED;
 803         send->s_wr.num_sge = 1;
 804         send->s_wr.next = NULL;
 805         send->s_wr.wr.atomic.remote_addr = op->op_remote_addr;
 806         send->s_wr.wr.atomic.rkey = op->op_rkey;
 807
 808         /* map 8 byte retval buffer to the device */
 809         ret = ib_dma_map_sg(ic->i_cm_id->device, op->op_sg, 1, DMA_FROM_DEVICE);
 810         rdsdebug("ic %p mapping atomic op %p. mapped %d pg\n", ic, op, ret);
 811         if (ret != 1) {
 812                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 813                 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
 814                 ret = -ENOMEM; /* XXX ? */
 815                 goto out;
 816         }
 817
 818         /* Convert our struct scatterlist to struct ib_sge */
 819         send->s_sge[0].addr = ib_sg_dma_address(ic->i_cm_id->device, op->op_sg);
 820         send->s_sge[0].length = ib_sg_dma_len(ic->i_cm_id->device, op->op_sg);
 821         send->s_sge[0].lkey = ic->i_mr->lkey;
 822
 823         rdsdebug("rva %Lx rpa %Lx len %u\n", op->op_remote_addr,
 824                  send->s_sge[0].addr, send->s_sge[0].length);
 825
 826         failed_wr = &send->s_wr;
 827         ret = ib_post_send(ic->i_cm_id->qp, &send->s_wr, &failed_wr);
 828         rdsdebug("ic %p send %p (wr %p) ret %d wr %p\n", ic,
 829                  send, &send->s_wr, ret, failed_wr);
 830         BUG_ON(failed_wr != &send->s_wr);
 831         if (ret) {
 832                 printk(KERN_WARNING "RDS/IB: atomic ib_post_send to %pI4 "
 833                        "returned %d\n", &conn->c_faddr, ret);
 834                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 835                 goto out;
 836         }
 837
 838         if (unlikely(failed_wr != &send->s_wr)) {
 839                 printk(KERN_WARNING "RDS/IB: atomic ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
 840                 BUG_ON(failed_wr != &send->s_wr);
 841         }
 842
 843 out:
 844         return ret;
 845 }
 846
 847 int rds_ib_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
 848 {
 849         struct rds_ib_connection *ic = conn->c_transport_data;
 850         struct rds_ib_send_work *send = NULL;
 851         struct rds_ib_send_work *first;
 852         struct rds_ib_send_work *prev;
 853         struct ib_send_wr *failed_wr;
 854         struct rds_ib_device *rds_ibdev;
 855         struct scatterlist *scat;
 856         unsigned long len;
 857         u64 remote_addr = op->r_remote_addr;
 858         u32 pos;
 859         u32 work_alloc;
 860         u32 i;
 861         u32 j;
 862         int sent;
 863         int ret;
 864         int num_sge;
 865
 866         rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
 867
 868         /* map the message the first time we see it */
 869         if (!op->r_mapped) {
 870                 op->r_count = ib_dma_map_sg(ic->i_cm_id->device,
 871                                         op->r_sg, op->r_nents, (op->r_write) ?
 872                                         DMA_TO_DEVICE : DMA_FROM_DEVICE);
 873                 rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->r_count);
 874                 if (op->r_count == 0) {
 875                         rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
 876                         ret = -ENOMEM; /* XXX ? */
 877                         goto out;
 878                 }
 879
 880                 op->r_mapped = 1;
 881         }
 882
 883         /*
 884          * Instead of knowing how to return a partial rdma read/write we insist that there
 885          * be enough work requests to send the entire message.
 886          */
 887         i = ceil(op->r_count, rds_ibdev->max_sge);
 888
 889         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
 890         if (work_alloc != i) {
 891                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 892                 rds_ib_stats_inc(s_ib_tx_ring_full);
 893                 ret = -ENOMEM;
 894                 goto out;
 895         }
 896
 897         send = &ic->i_sends[pos];
 898         first = send;
 899         prev = NULL;
 900         scat = &op->r_sg[0];
 901         sent = 0;
 902         num_sge = op->r_count;
 903
 904         for (i = 0; i < work_alloc && scat != &op->r_sg[op->r_count]; i++) {
 905                 send->s_wr.send_flags = 0;
 906                 send->s_queued = jiffies;
 907                 /*
 908                  * We want to delay signaling completions just enough to get
 909                  * the batching benefits but not so much that we create dead time on the wire.
 910                  */
 911                 if (ic->i_unsignaled_wrs-- == 0) {
 912                         ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
 913                         send->s_wr.send_flags = IB_SEND_SIGNALED;
 914                 }
 915
 916                 send->s_wr.opcode = op->r_write ? IB_WR_RDMA_WRITE : IB_WR_RDMA_READ;
 917                 send->s_wr.wr.rdma.remote_addr = remote_addr;
 918                 send->s_wr.wr.rdma.rkey = op->r_key;
 919                 send->s_op = op;
 920
 921                 if (num_sge > rds_ibdev->max_sge) {
 922                         send->s_wr.num_sge = rds_ibdev->max_sge;
 923                         num_sge -= rds_ibdev->max_sge;
 924                 } else {
 925                         send->s_wr.num_sge = num_sge;
 926                 }
 927
 928                 send->s_wr.next = NULL;
 929
 930                 if (prev)
 931                         prev->s_wr.next = &send->s_wr;
 932
 933                 for (j = 0; j < send->s_wr.num_sge && scat != &op->r_sg[op->r_count]; j++) {
 934                         len = ib_sg_dma_len(ic->i_cm_id->device, scat);
 935                         send->s_sge[j].addr =
 936                                  ib_sg_dma_address(ic->i_cm_id->device, scat);
 937                         send->s_sge[j].length = len;
 938                         send->s_sge[j].lkey = ic->i_mr->lkey;
 939
 940                         sent += len;
 941                         rdsdebug("ic %p sent %d remote_addr %llu\n", ic, sent, remote_addr);
 942
 943                         remote_addr += len;
 944                         scat++;
 945                 }
 946
 947                 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
 948                         &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
 949
 950                 prev = send;
 951                 if (++send == &ic->i_sends[ic->i_send_ring.w_nr])
 952                         send = ic->i_sends;
 953         }
 954
 955         /* if we finished the message then send completion owns it */
 956         if (scat == &op->r_sg[op->r_count])
 957                 prev->s_wr.send_flags = IB_SEND_SIGNALED;
 958
 959         if (i < work_alloc) {
 960                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
 961                 work_alloc = i;
 962         }
 963
 964         failed_wr = &first->s_wr;
 965         ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
 966         rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
 967                  first, &first->s_wr, ret, failed_wr);
 968         BUG_ON(failed_wr != &first->s_wr);
 969         if (ret) {
 970                 printk(KERN_WARNING "RDS/IB: rdma ib_post_send to %pI4 "
 971                        "returned %d\n", &conn->c_faddr, ret);
 972                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 973                 goto out;
 974         }
 975
 976         if (unlikely(failed_wr != &first->s_wr)) {
 977                 printk(KERN_WARNING "RDS/IB: ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
 978                 BUG_ON(failed_wr != &first->s_wr);
 979         }
 980
 981
 982 out:
 983         return ret;
 984 }
 985
 986 void rds_ib_xmit_complete(struct rds_connection *conn)
 987 {
 988         struct rds_ib_connection *ic = conn->c_transport_data;
 989
 990         /* We may have a pending ACK or window update we were unable
 991          * to send previously (due to flow control). Try again. */
 992         rds_ib_attempt_ack(ic);
 993 }