net/rds/ib_send.c

   1 /*
   2  * Copyright (c) 2006 Oracle.  All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  *
  32  */
  33 #include <linux/kernel.h>
  34 #include <linux/in.h>
  35 #include <linux/device.h>
  36 #include <linux/dmapool.h>
  37
  38 #include "rds.h"
  39 #include "ib.h"
  40
  41 /*
  42  * Convert IB-specific error message to RDS error message and call core
  43  * completion handler.
  44  */
  45 static void rds_ib_send_complete(struct rds_message *rm,
  46                                  int wc_status,
  47                                  void (*complete)(struct rds_message *rm, int status))
  48 {
  49         int notify_status;
  50
  51         switch (wc_status) {
  52         case IB_WC_WR_FLUSH_ERR:
  53                 return;
  54
  55         case IB_WC_SUCCESS:
  56                 notify_status = RDS_RDMA_SUCCESS;
  57                 break;
  58
  59         case IB_WC_REM_ACCESS_ERR:
  60                 notify_status = RDS_RDMA_REMOTE_ERROR;
  61                 break;
  62
  63         default:
  64                 notify_status = RDS_RDMA_OTHER_ERROR;
  65                 break;
  66         }
  67         complete(rm, notify_status);
  68 }
  69
  70 static void rds_ib_send_unmap_rm(struct rds_ib_connection *ic,
  71                           struct rds_ib_send_work *send,
  72                           int wc_status)
  73 {
  74         struct rds_message *rm = send->s_rm;
  75
  76         rdsdebug("ic %p send %p rm %p\n", ic, send, rm);
  77
  78         ib_dma_unmap_sg(ic->i_cm_id->device,
  79                         rm->data.m_sg, rm->data.m_nents,
  80                         DMA_TO_DEVICE);
  81
  82         if (rm->rdma.m_rdma_op.r_active) {
  83                 struct rds_rdma_op *op = &rm->rdma.m_rdma_op;
  84
  85                 if (op->r_mapped) {
  86                         ib_dma_unmap_sg(ic->i_cm_id->device,
  87                                         op->r_sg, op->r_nents,
  88                                         op->r_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
  89                         op->r_mapped = 0;
  90                 }
  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_ib_send_complete(rm, wc_status, rds_rdma_send_complete);
 113
 114                 if (rm->rdma.m_rdma_op.r_write)
 115                         rds_stats_add(s_send_rdma_bytes, rm->rdma.m_rdma_op.r_bytes);
 116                 else
 117                         rds_stats_add(s_recv_rdma_bytes, rm->rdma.m_rdma_op.r_bytes);
 118         }
 119
 120         if (rm->atomic.op_active) {
 121                 struct rm_atomic_op *op = &rm->atomic;
 122
 123                 /* unmap atomic recvbuf */
 124                 if (op->op_mapped) {
 125                         ib_dma_unmap_sg(ic->i_cm_id->device, op->op_sg, 1,
 126                                         DMA_FROM_DEVICE);
 127                         op->op_mapped = 0;
 128                 }
 129
 130                 rds_ib_send_complete(rm, wc_status, rds_atomic_send_complete);
 131
 132                 if (rm->atomic.op_type == RDS_ATOMIC_TYPE_CSWP)
 133                         rds_stats_inc(s_atomic_cswp);
 134                 else
 135                         rds_stats_inc(s_atomic_fadd);
 136         }
 137
 138         /* If anyone waited for this message to get flushed out, wake
 139          * them up now */
 140         rds_message_unmapped(rm);
 141
 142         rds_message_put(rm);
 143         send->s_rm = NULL;
 144 }
 145
 146 void rds_ib_send_init_ring(struct rds_ib_connection *ic)
 147 {
 148         struct rds_ib_send_work *send;
 149         u32 i;
 150
 151         for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
 152                 struct ib_sge *sge;
 153
 154                 send->s_rm = NULL;
 155                 send->s_op = NULL;
 156
 157                 send->s_wr.wr_id = i;
 158                 send->s_wr.sg_list = send->s_sge;
 159                 send->s_wr.num_sge = 1;
 160                 send->s_wr.opcode = IB_WR_SEND;
 161                 send->s_wr.send_flags = 0;
 162                 send->s_wr.ex.imm_data = 0;
 163
 164                 sge = rds_ib_data_sge(ic, send->s_sge);
 165                 sge->lkey = ic->i_mr->lkey;
 166
 167                 sge = rds_ib_header_sge(ic, send->s_sge);
 168                 sge->addr = ic->i_send_hdrs_dma + (i * sizeof(struct rds_header));
 169                 sge->length = sizeof(struct rds_header);
 170                 sge->lkey = ic->i_mr->lkey;
 171         }
 172 }
 173
 174 void rds_ib_send_clear_ring(struct rds_ib_connection *ic)
 175 {
 176         struct rds_ib_send_work *send;
 177         u32 i;
 178
 179         for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
 180                 if (!send->s_rm || send->s_wr.opcode == 0xdead)
 181                         continue;
 182                 rds_ib_send_unmap_rm(ic, send, IB_WC_WR_FLUSH_ERR);
 183         }
 184 }
 185
 186 /*
 187  * The _oldest/_free ring operations here race cleanly with the alloc/unalloc
 188  * operations performed in the send path.  As the sender allocs and potentially
 189  * unallocs the next free entry in the ring it doesn't alter which is
 190  * the next to be freed, which is what this is concerned with.
 191  */
 192 void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context)
 193 {
 194         struct rds_connection *conn = context;
 195         struct rds_ib_connection *ic = conn->c_transport_data;
 196         struct ib_wc wc;
 197         struct rds_ib_send_work *send;
 198         u32 completed;
 199         u32 oldest;
 200         u32 i = 0;
 201         int ret;
 202
 203         rdsdebug("cq %p conn %p\n", cq, conn);
 204         rds_ib_stats_inc(s_ib_tx_cq_call);
 205         ret = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 206         if (ret)
 207                 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
 208
 209         while (ib_poll_cq(cq, 1, &wc) > 0) {
 210                 rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
 211                          (unsigned long long)wc.wr_id, wc.status, wc.byte_len,
 212                          be32_to_cpu(wc.ex.imm_data));
 213                 rds_ib_stats_inc(s_ib_tx_cq_event);
 214
 215                 if (wc.wr_id == RDS_IB_ACK_WR_ID) {
 216                         if (ic->i_ack_queued + HZ/2 < jiffies)
 217                                 rds_ib_stats_inc(s_ib_tx_stalled);
 218                         rds_ib_ack_send_complete(ic);
 219                         continue;
 220                 }
 221
 222                 oldest = rds_ib_ring_oldest(&ic->i_send_ring);
 223
 224                 completed = rds_ib_ring_completed(&ic->i_send_ring, wc.wr_id, oldest);
 225
 226                 for (i = 0; i < completed; i++) {
 227                         send = &ic->i_sends[oldest];
 228
 229                         /* In the error case, wc.opcode sometimes contains garbage */
 230                         switch (send->s_wr.opcode) {
 231                         case IB_WR_SEND:
 232                                 if (send->s_rm)
 233                                         rds_ib_send_unmap_rm(ic, send, wc.status);
 234                                 break;
 235                         case IB_WR_RDMA_WRITE:
 236                         case IB_WR_RDMA_READ:
 237                         case IB_WR_ATOMIC_FETCH_AND_ADD:
 238                         case IB_WR_ATOMIC_CMP_AND_SWP:
 239                                 /* Nothing to be done - the SG list will be unmapped
 240                                  * when the SEND completes. */
 241                                 break;
 242                         default:
 243                                 if (printk_ratelimit())
 244                                         printk(KERN_NOTICE
 245                                                 "RDS/IB: %s: unexpected opcode 0x%x in WR!\n",
 246                                                 __func__, send->s_wr.opcode);
 247                                 break;
 248                         }
 249
 250                         send->s_wr.opcode = 0xdead;
 251                         send->s_wr.num_sge = 1;
 252                         if (send->s_queued + HZ/2 < jiffies)
 253                                 rds_ib_stats_inc(s_ib_tx_stalled);
 254
 255                         /* If a RDMA operation produced an error, signal this right
 256                          * away. If we don't, the subsequent SEND that goes with this
 257                          * RDMA will be canceled with ERR_WFLUSH, and the application
 258                          * never learn that the RDMA failed. */
 259                         if (unlikely(wc.status == IB_WC_REM_ACCESS_ERR && send->s_op)) {
 260                                 struct rds_message *rm;
 261
 262                                 rm = rds_send_get_message(conn, send->s_op);
 263                                 if (rm) {
 264                                         rds_ib_send_unmap_rm(ic, send, wc.status);
 265                                         rds_ib_send_complete(rm, wc.status, rds_rdma_send_complete);
 266                                         rds_message_put(rm);
 267                                 }
 268                         }
 269
 270                         oldest = (oldest + 1) % ic->i_send_ring.w_nr;
 271                 }
 272
 273                 rds_ib_ring_free(&ic->i_send_ring, completed);
 274
 275                 if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
 276                     test_bit(0, &conn->c_map_queued))
 277                         queue_delayed_work(rds_wq, &conn->c_send_w, 0);
 278
 279                 /* We expect errors as the qp is drained during shutdown */
 280                 if (wc.status != IB_WC_SUCCESS && rds_conn_up(conn)) {
 281                         rds_ib_conn_error(conn,
 282                                 "send completion on %pI4 "
 283                                 "had status %u, disconnecting and reconnecting\n",
 284                                 &conn->c_faddr, wc.status);
 285                 }
 286         }
 287 }
 288
 289 /*
 290  * This is the main function for allocating credits when sending
 291  * messages.
 292  *
 293  * Conceptually, we have two counters:
 294  *  -   send credits: this tells us how many WRs we're allowed
 295  *      to submit without overruning the reciever's queue. For
 296  *      each SEND WR we post, we decrement this by one.
 297  *
 298  *  -   posted credits: this tells us how many WRs we recently
 299  *      posted to the receive queue. This value is transferred
 300  *      to the peer as a "credit update" in a RDS header field.
 301  *      Every time we transmit credits to the peer, we subtract
 302  *      the amount of transferred credits from this counter.
 303  *
 304  * It is essential that we avoid situations where both sides have
 305  * exhausted their send credits, and are unable to send new credits
 306  * to the peer. We achieve this by requiring that we send at least
 307  * one credit update to the peer before exhausting our credits.
 308  * When new credits arrive, we subtract one credit that is withheld
 309  * until we've posted new buffers and are ready to transmit these
 310  * credits (see rds_ib_send_add_credits below).
 311  *
 312  * The RDS send code is essentially single-threaded; rds_send_xmit
 313  * grabs c_send_lock to ensure exclusive access to the send ring.
 314  * However, the ACK sending code is independent and can race with
 315  * message SENDs.
 316  *
 317  * In the send path, we need to update the counters for send credits
 318  * and the counter of posted buffers atomically - when we use the
 319  * last available credit, we cannot allow another thread to race us
 320  * and grab the posted credits counter.  Hence, we have to use a
 321  * spinlock to protect the credit counter, or use atomics.
 322  *
 323  * Spinlocks shared between the send and the receive path are bad,
 324  * because they create unnecessary delays. An early implementation
 325  * using a spinlock showed a 5% degradation in throughput at some
 326  * loads.
 327  *
 328  * This implementation avoids spinlocks completely, putting both
 329  * counters into a single atomic, and updating that atomic using
 330  * atomic_add (in the receive path, when receiving fresh credits),
 331  * and using atomic_cmpxchg when updating the two counters.
 332  */
 333 int rds_ib_send_grab_credits(struct rds_ib_connection *ic,
 334                              u32 wanted, u32 *adv_credits, int need_posted, int max_posted)
 335 {
 336         unsigned int avail, posted, got = 0, advertise;
 337         long oldval, newval;
 338
 339         *adv_credits = 0;
 340         if (!ic->i_flowctl)
 341                 return wanted;
 342
 343 try_again:
 344         advertise = 0;
 345         oldval = newval = atomic_read(&ic->i_credits);
 346         posted = IB_GET_POST_CREDITS(oldval);
 347         avail = IB_GET_SEND_CREDITS(oldval);
 348
 349         rdsdebug("rds_ib_send_grab_credits(%u): credits=%u posted=%u\n",
 350                         wanted, avail, posted);
 351
 352         /* The last credit must be used to send a credit update. */
 353         if (avail && !posted)
 354                 avail--;
 355
 356         if (avail < wanted) {
 357                 struct rds_connection *conn = ic->i_cm_id->context;
 358
 359                 /* Oops, there aren't that many credits left! */
 360                 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 361                 got = avail;
 362         } else {
 363                 /* Sometimes you get what you want, lalala. */
 364                 got = wanted;
 365         }
 366         newval -= IB_SET_SEND_CREDITS(got);
 367
 368         /*
 369          * If need_posted is non-zero, then the caller wants
 370          * the posted regardless of whether any send credits are
 371          * available.
 372          */
 373         if (posted && (got || need_posted)) {
 374                 advertise = min_t(unsigned int, posted, max_posted);
 375                 newval -= IB_SET_POST_CREDITS(advertise);
 376         }
 377
 378         /* Finally bill everything */
 379         if (atomic_cmpxchg(&ic->i_credits, oldval, newval) != oldval)
 380                 goto try_again;
 381
 382         *adv_credits = advertise;
 383         return got;
 384 }
 385
 386 void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits)
 387 {
 388         struct rds_ib_connection *ic = conn->c_transport_data;
 389
 390         if (credits == 0)
 391                 return;
 392
 393         rdsdebug("rds_ib_send_add_credits(%u): current=%u%s\n",
 394                         credits,
 395                         IB_GET_SEND_CREDITS(atomic_read(&ic->i_credits)),
 396                         test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ? ", ll_send_full" : "");
 397
 398         atomic_add(IB_SET_SEND_CREDITS(credits), &ic->i_credits);
 399         if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags))
 400                 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
 401
 402         WARN_ON(IB_GET_SEND_CREDITS(credits) >= 16384);
 403
 404         rds_ib_stats_inc(s_ib_rx_credit_updates);
 405 }
 406
 407 void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted)
 408 {
 409         struct rds_ib_connection *ic = conn->c_transport_data;
 410
 411         if (posted == 0)
 412                 return;
 413
 414         atomic_add(IB_SET_POST_CREDITS(posted), &ic->i_credits);
 415
 416         /* Decide whether to send an update to the peer now.
 417          * If we would send a credit update for every single buffer we
 418          * post, we would end up with an ACK storm (ACK arrives,
 419          * consumes buffer, we refill the ring, send ACK to remote
 420          * advertising the newly posted buffer... ad inf)
 421          *
 422          * Performance pretty much depends on how often we send
 423          * credit updates - too frequent updates mean lots of ACKs.
 424          * Too infrequent updates, and the peer will run out of
 425          * credits and has to throttle.
 426          * For the time being, 16 seems to be a good compromise.
 427          */
 428         if (IB_GET_POST_CREDITS(atomic_read(&ic->i_credits)) >= 16)
 429                 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
 430 }
 431
 432 static inline void
 433 rds_ib_xmit_populate_wr(struct rds_ib_connection *ic,
 434                 struct rds_ib_send_work *send, unsigned int pos,
 435                 unsigned long buffer, unsigned int length,
 436                 int send_flags)
 437 {
 438         struct ib_sge *sge;
 439
 440         WARN_ON(pos != send - ic->i_sends);
 441
 442         send->s_wr.send_flags = send_flags;
 443         send->s_wr.opcode = IB_WR_SEND;
 444         send->s_wr.num_sge = 2;
 445         send->s_wr.next = NULL;
 446         send->s_queued = jiffies;
 447         send->s_op = NULL;
 448
 449         if (length != 0) {
 450                 sge = rds_ib_data_sge(ic, send->s_sge);
 451                 sge->addr = buffer;
 452                 sge->length = length;
 453                 sge->lkey = ic->i_mr->lkey;
 454
 455                 sge = rds_ib_header_sge(ic, send->s_sge);
 456         } else {
 457                 /* We're sending a packet with no payload. There is only
 458                  * one SGE */
 459                 send->s_wr.num_sge = 1;
 460                 sge = &send->s_sge[0];
 461         }
 462
 463         sge->addr = ic->i_send_hdrs_dma + (pos * sizeof(struct rds_header));
 464         sge->length = sizeof(struct rds_header);
 465         sge->lkey = ic->i_mr->lkey;
 466 }
 467
 468 /*
 469  * This can be called multiple times for a given message.  The first time
 470  * we see a message we map its scatterlist into the IB device so that
 471  * we can provide that mapped address to the IB scatter gather entries
 472  * in the IB work requests.  We translate the scatterlist into a series
 473  * of work requests that fragment the message.  These work requests complete
 474  * in order so we pass ownership of the message to the completion handler
 475  * once we send the final fragment.
 476  *
 477  * The RDS core uses the c_send_lock to only enter this function once
 478  * per connection.  This makes sure that the tx ring alloc/unalloc pairs
 479  * don't get out of sync and confuse the ring.
 480  */
 481 int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
 482                 unsigned int hdr_off, unsigned int sg, unsigned int off)
 483 {
 484         struct rds_ib_connection *ic = conn->c_transport_data;
 485         struct ib_device *dev = ic->i_cm_id->device;
 486         struct rds_ib_send_work *send = NULL;
 487         struct rds_ib_send_work *first;
 488         struct rds_ib_send_work *prev;
 489         struct ib_send_wr *failed_wr;
 490         struct scatterlist *scat;
 491         u32 pos;
 492         u32 i;
 493         u32 work_alloc;
 494         u32 credit_alloc;
 495         u32 posted;
 496         u32 adv_credits = 0;
 497         int send_flags = 0;
 498         int sent;
 499         int ret;
 500         int flow_controlled = 0;
 501
 502         BUG_ON(off % RDS_FRAG_SIZE);
 503         BUG_ON(hdr_off != 0 && hdr_off != sizeof(struct rds_header));
 504
 505         /* Do not send cong updates to IB loopback */
 506         if (conn->c_loopback
 507             && rm->m_inc.i_hdr.h_flags & RDS_FLAG_CONG_BITMAP) {
 508                 rds_cong_map_updated(conn->c_fcong, ~(u64) 0);
 509                 return sizeof(struct rds_header) + RDS_CONG_MAP_BYTES;
 510         }
 511
 512         /* FIXME we may overallocate here */
 513         if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0)
 514                 i = 1;
 515         else
 516                 i = ceil(be32_to_cpu(rm->m_inc.i_hdr.h_len), RDS_FRAG_SIZE);
 517
 518         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
 519         if (work_alloc == 0) {
 520                 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 521                 rds_ib_stats_inc(s_ib_tx_ring_full);
 522                 ret = -ENOMEM;
 523                 goto out;
 524         }
 525
 526         credit_alloc = work_alloc;
 527         if (ic->i_flowctl) {
 528                 credit_alloc = rds_ib_send_grab_credits(ic, work_alloc, &posted, 0, RDS_MAX_ADV_CREDIT);
 529                 adv_credits += posted;
 530                 if (credit_alloc < work_alloc) {
 531                         rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - credit_alloc);
 532                         work_alloc = credit_alloc;
 533                         flow_controlled++;
 534                 }
 535                 if (work_alloc == 0) {
 536                         set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 537                         rds_ib_stats_inc(s_ib_tx_throttle);
 538                         ret = -ENOMEM;
 539                         goto out;
 540                 }
 541         }
 542
 543         /* map the message the first time we see it */
 544         if (!ic->i_rm) {
 545                 /*
 546                 printk(KERN_NOTICE "rds_ib_xmit prep msg dport=%u flags=0x%x len=%d\n",
 547                                 be16_to_cpu(rm->m_inc.i_hdr.h_dport),
 548                                 rm->m_inc.i_hdr.h_flags,
 549                                 be32_to_cpu(rm->m_inc.i_hdr.h_len));
 550                    */
 551                 if (rm->data.m_nents) {
 552                         rm->data.m_count = ib_dma_map_sg(dev,
 553                                                             rm->data.m_sg,
 554                                                             rm->data.m_nents,
 555                                                             DMA_TO_DEVICE);
 556                         rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->data.m_count);
 557                         if (rm->data.m_count == 0) {
 558                                 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
 559                                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 560                                 ret = -ENOMEM; /* XXX ? */
 561                                 goto out;
 562                         }
 563                 } else {
 564                         rm->data.m_count = 0;
 565                 }
 566
 567                 ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
 568                 ic->i_unsignaled_bytes = rds_ib_sysctl_max_unsig_bytes;
 569                 rds_message_addref(rm);
 570                 ic->i_rm = rm;
 571
 572                 /* Finalize the header */
 573                 if (test_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags))
 574                         rm->m_inc.i_hdr.h_flags |= RDS_FLAG_ACK_REQUIRED;
 575                 if (test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))
 576                         rm->m_inc.i_hdr.h_flags |= RDS_FLAG_RETRANSMITTED;
 577
 578                 /* If it has a RDMA op, tell the peer we did it. This is
 579                  * used by the peer to release use-once RDMA MRs. */
 580                 if (rm->rdma.m_rdma_op.r_active) {
 581                         struct rds_ext_header_rdma ext_hdr;
 582
 583                         ext_hdr.h_rdma_rkey = cpu_to_be32(rm->rdma.m_rdma_op.r_key);
 584                         rds_message_add_extension(&rm->m_inc.i_hdr,
 585                                         RDS_EXTHDR_RDMA, &ext_hdr, sizeof(ext_hdr));
 586                 }
 587                 if (rm->m_rdma_cookie) {
 588                         rds_message_add_rdma_dest_extension(&rm->m_inc.i_hdr,
 589                                         rds_rdma_cookie_key(rm->m_rdma_cookie),
 590                                         rds_rdma_cookie_offset(rm->m_rdma_cookie));
 591                 }
 592
 593                 /* Note - rds_ib_piggyb_ack clears the ACK_REQUIRED bit, so
 594                  * we should not do this unless we have a chance of at least
 595                  * sticking the header into the send ring. Which is why we
 596                  * should call rds_ib_ring_alloc first. */
 597                 rm->m_inc.i_hdr.h_ack = cpu_to_be64(rds_ib_piggyb_ack(ic));
 598                 rds_message_make_checksum(&rm->m_inc.i_hdr);
 599
 600                 /*
 601                  * Update adv_credits since we reset the ACK_REQUIRED bit.
 602                  */
 603                 rds_ib_send_grab_credits(ic, 0, &posted, 1, RDS_MAX_ADV_CREDIT - adv_credits);
 604                 adv_credits += posted;
 605                 BUG_ON(adv_credits > 255);
 606         }
 607
 608         send = &ic->i_sends[pos];
 609         first = send;
 610         prev = NULL;
 611         scat = &rm->data.m_sg[sg];
 612         sent = 0;
 613         i = 0;
 614
 615         /* Sometimes you want to put a fence between an RDMA
 616          * READ and the following SEND.
 617          * We could either do this all the time
 618          * or when requested by the user. Right now, we let
 619          * the application choose.
 620          */
 621         if (rm->rdma.m_rdma_op.r_active && rm->rdma.m_rdma_op.r_fence)
 622                 send_flags = IB_SEND_FENCE;
 623
 624         /*
 625          * We could be copying the header into the unused tail of the page.
 626          * That would need to be changed in the future when those pages might
 627          * be mapped userspace pages or page cache pages.  So instead we always
 628          * use a second sge and our long-lived ring of mapped headers.  We send
 629          * the header after the data so that the data payload can be aligned on
 630          * the receiver.
 631          */
 632
 633         /* handle a 0-len message */
 634         if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0) {
 635                 rds_ib_xmit_populate_wr(ic, send, pos, 0, 0, send_flags);
 636                 goto add_header;
 637         }
 638
 639         /* if there's data reference it with a chain of work reqs */
 640         for (; i < work_alloc && scat != &rm->data.m_sg[rm->data.m_count]; i++) {
 641                 unsigned int len;
 642
 643                 send = &ic->i_sends[pos];
 644
 645                 len = min(RDS_FRAG_SIZE, ib_sg_dma_len(dev, scat) - off);
 646                 rds_ib_xmit_populate_wr(ic, send, pos,
 647                                 ib_sg_dma_address(dev, scat) + off, len,
 648                                 send_flags);
 649
 650                 /*
 651                  * We want to delay signaling completions just enough to get
 652                  * the batching benefits but not so much that we create dead time
 653                  * on the wire.
 654                  */
 655                 if (ic->i_unsignaled_wrs-- == 0) {
 656                         ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
 657                         send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
 658                 }
 659
 660                 ic->i_unsignaled_bytes -= len;
 661                 if (ic->i_unsignaled_bytes <= 0) {
 662                         ic->i_unsignaled_bytes = rds_ib_sysctl_max_unsig_bytes;
 663                         send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
 664                 }
 665
 666                 /*
 667                  * Always signal the last one if we're stopping due to flow control.
 668                  */
 669                 if (flow_controlled && i == (work_alloc-1))
 670                         send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
 671
 672                 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
 673                          &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
 674
 675                 sent += len;
 676                 off += len;
 677                 if (off == ib_sg_dma_len(dev, scat)) {
 678                         scat++;
 679                         off = 0;
 680                 }
 681
 682 add_header:
 683                 /* Tack on the header after the data. The header SGE should already
 684                  * have been set up to point to the right header buffer. */
 685                 memcpy(&ic->i_send_hdrs[pos], &rm->m_inc.i_hdr, sizeof(struct rds_header));
 686
 687                 if (0) {
 688                         struct rds_header *hdr = &ic->i_send_hdrs[pos];
 689
 690                         printk(KERN_NOTICE "send WR dport=%u flags=0x%x len=%d\n",
 691                                 be16_to_cpu(hdr->h_dport),
 692                                 hdr->h_flags,
 693                                 be32_to_cpu(hdr->h_len));
 694                 }
 695                 if (adv_credits) {
 696                         struct rds_header *hdr = &ic->i_send_hdrs[pos];
 697
 698                         /* add credit and redo the header checksum */
 699                         hdr->h_credit = adv_credits;
 700                         rds_message_make_checksum(hdr);
 701                         adv_credits = 0;
 702                         rds_ib_stats_inc(s_ib_tx_credit_updates);
 703                 }
 704
 705                 if (prev)
 706                         prev->s_wr.next = &send->s_wr;
 707                 prev = send;
 708
 709                 pos = (pos + 1) % ic->i_send_ring.w_nr;
 710         }
 711
 712         /* Account the RDS header in the number of bytes we sent, but just once.
 713          * The caller has no concept of fragmentation. */
 714         if (hdr_off == 0)
 715                 sent += sizeof(struct rds_header);
 716
 717         /* if we finished the message then send completion owns it */
 718         if (scat == &rm->data.m_sg[rm->data.m_count]) {
 719                 prev->s_rm = ic->i_rm;
 720                 prev->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
 721                 ic->i_rm = NULL;
 722         }
 723
 724         if (i < work_alloc) {
 725                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
 726                 work_alloc = i;
 727         }
 728         if (ic->i_flowctl && i < credit_alloc)
 729                 rds_ib_send_add_credits(conn, credit_alloc - i);
 730
 731         /* XXX need to worry about failed_wr and partial sends. */
 732         failed_wr = &first->s_wr;
 733         ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
 734         rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
 735                  first, &first->s_wr, ret, failed_wr);
 736         BUG_ON(failed_wr != &first->s_wr);
 737         if (ret) {
 738                 printk(KERN_WARNING "RDS/IB: ib_post_send to %pI4 "
 739                        "returned %d\n", &conn->c_faddr, ret);
 740                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 741                 if (prev->s_rm) {
 742                         ic->i_rm = prev->s_rm;
 743                         prev->s_rm = NULL;
 744                 }
 745
 746                 rds_ib_conn_error(ic->conn, "ib_post_send failed\n");
 747                 goto out;
 748         }
 749
 750         ret = sent;
 751 out:
 752         BUG_ON(adv_credits);
 753         return ret;
 754 }
 755
 756 /*
 757  * Issue atomic operation.
 758  * A simplified version of the rdma case, we always map 1 SG, and
 759  * only 8 bytes, for the return value from the atomic operation.
 760  */
 761 int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op)
 762 {
 763         struct rds_ib_connection *ic = conn->c_transport_data;
 764         struct rds_ib_send_work *send = NULL;
 765         struct ib_send_wr *failed_wr;
 766         struct rds_ib_device *rds_ibdev;
 767         u32 pos;
 768         u32 work_alloc;
 769         int ret;
 770
 771         rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
 772
 773         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, 1, &pos);
 774         if (work_alloc != 1) {
 775                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 776                 rds_ib_stats_inc(s_ib_tx_ring_full);
 777                 ret = -ENOMEM;
 778                 goto out;
 779         }
 780
 781         /* address of send request in ring */
 782         send = &ic->i_sends[pos];
 783         send->s_queued = jiffies;
 784
 785         if (op->op_type == RDS_ATOMIC_TYPE_CSWP) {
 786                 send->s_wr.opcode = IB_WR_ATOMIC_CMP_AND_SWP;
 787                 send->s_wr.wr.atomic.compare_add = op->op_compare;
 788                 send->s_wr.wr.atomic.swap = op->op_swap_add;
 789         } else { /* FADD */
 790                 send->s_wr.opcode = IB_WR_ATOMIC_FETCH_AND_ADD;
 791                 send->s_wr.wr.atomic.compare_add = op->op_swap_add;
 792                 send->s_wr.wr.atomic.swap = 0;
 793         }
 794         send->s_wr.send_flags = IB_SEND_SIGNALED;
 795         send->s_wr.num_sge = 1;
 796         send->s_wr.next = NULL;
 797         send->s_wr.wr.atomic.remote_addr = op->op_remote_addr;
 798         send->s_wr.wr.atomic.rkey = op->op_rkey;
 799
 800         /* map 8 byte retval buffer to the device */
 801         ret = ib_dma_map_sg(ic->i_cm_id->device, op->op_sg, 1, DMA_FROM_DEVICE);
 802         rdsdebug("ic %p mapping atomic op %p. mapped %d pg\n", ic, op, ret);
 803         if (ret != 1) {
 804                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 805                 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
 806                 ret = -ENOMEM; /* XXX ? */
 807                 goto out;
 808         }
 809
 810         /* Convert our struct scatterlist to struct ib_sge */
 811         send->s_sge[0].addr = ib_sg_dma_address(ic->i_cm_id->device, op->op_sg);
 812         send->s_sge[0].length = ib_sg_dma_len(ic->i_cm_id->device, op->op_sg);
 813         send->s_sge[0].lkey = ic->i_mr->lkey;
 814
 815         rdsdebug("rva %Lx rpa %Lx len %u\n", op->op_remote_addr,
 816                  send->s_sge[0].addr, send->s_sge[0].length);
 817
 818         failed_wr = &send->s_wr;
 819         ret = ib_post_send(ic->i_cm_id->qp, &send->s_wr, &failed_wr);
 820         rdsdebug("ic %p send %p (wr %p) ret %d wr %p\n", ic,
 821                  send, &send->s_wr, ret, failed_wr);
 822         BUG_ON(failed_wr != &send->s_wr);
 823         if (ret) {
 824                 printk(KERN_WARNING "RDS/IB: atomic ib_post_send to %pI4 "
 825                        "returned %d\n", &conn->c_faddr, ret);
 826                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 827                 goto out;
 828         }
 829
 830         if (unlikely(failed_wr != &send->s_wr)) {
 831                 printk(KERN_WARNING "RDS/IB: atomic ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
 832                 BUG_ON(failed_wr != &send->s_wr);
 833         }
 834
 835 out:
 836         return ret;
 837 }
 838
 839 int rds_ib_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
 840 {
 841         struct rds_ib_connection *ic = conn->c_transport_data;
 842         struct rds_ib_send_work *send = NULL;
 843         struct rds_ib_send_work *first;
 844         struct rds_ib_send_work *prev;
 845         struct ib_send_wr *failed_wr;
 846         struct rds_ib_device *rds_ibdev;
 847         struct scatterlist *scat;
 848         unsigned long len;
 849         u64 remote_addr = op->r_remote_addr;
 850         u32 pos;
 851         u32 work_alloc;
 852         u32 i;
 853         u32 j;
 854         int sent;
 855         int ret;
 856         int num_sge;
 857
 858         rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
 859
 860         /* map the message the first time we see it */
 861         if (!op->r_mapped) {
 862                 op->r_count = ib_dma_map_sg(ic->i_cm_id->device,
 863                                         op->r_sg, op->r_nents, (op->r_write) ?
 864                                         DMA_TO_DEVICE : DMA_FROM_DEVICE);
 865                 rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->r_count);
 866                 if (op->r_count == 0) {
 867                         rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
 868                         ret = -ENOMEM; /* XXX ? */
 869                         goto out;
 870                 }
 871
 872                 op->r_mapped = 1;
 873         }
 874
 875         /*
 876          * Instead of knowing how to return a partial rdma read/write we insist that there
 877          * be enough work requests to send the entire message.
 878          */
 879         i = ceil(op->r_count, rds_ibdev->max_sge);
 880
 881         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
 882         if (work_alloc != i) {
 883                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 884                 rds_ib_stats_inc(s_ib_tx_ring_full);
 885                 ret = -ENOMEM;
 886                 goto out;
 887         }
 888
 889         send = &ic->i_sends[pos];
 890         first = send;
 891         prev = NULL;
 892         scat = &op->r_sg[0];
 893         sent = 0;
 894         num_sge = op->r_count;
 895
 896         for (i = 0; i < work_alloc && scat != &op->r_sg[op->r_count]; i++) {
 897                 send->s_wr.send_flags = 0;
 898                 send->s_queued = jiffies;
 899                 /*
 900                  * We want to delay signaling completions just enough to get
 901                  * the batching benefits but not so much that we create dead time on the wire.
 902                  */
 903                 if (ic->i_unsignaled_wrs-- == 0) {
 904                         ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
 905                         send->s_wr.send_flags = IB_SEND_SIGNALED;
 906                 }
 907
 908                 send->s_wr.opcode = op->r_write ? IB_WR_RDMA_WRITE : IB_WR_RDMA_READ;
 909                 send->s_wr.wr.rdma.remote_addr = remote_addr;
 910                 send->s_wr.wr.rdma.rkey = op->r_key;
 911                 send->s_op = op;
 912
 913                 if (num_sge > rds_ibdev->max_sge) {
 914                         send->s_wr.num_sge = rds_ibdev->max_sge;
 915                         num_sge -= rds_ibdev->max_sge;
 916                 } else {
 917                         send->s_wr.num_sge = num_sge;
 918                 }
 919
 920                 send->s_wr.next = NULL;
 921
 922                 if (prev)
 923                         prev->s_wr.next = &send->s_wr;
 924
 925                 for (j = 0; j < send->s_wr.num_sge && scat != &op->r_sg[op->r_count]; j++) {
 926                         len = ib_sg_dma_len(ic->i_cm_id->device, scat);
 927                         send->s_sge[j].addr =
 928                                  ib_sg_dma_address(ic->i_cm_id->device, scat);
 929                         send->s_sge[j].length = len;
 930                         send->s_sge[j].lkey = ic->i_mr->lkey;
 931
 932                         sent += len;
 933                         rdsdebug("ic %p sent %d remote_addr %llu\n", ic, sent, remote_addr);
 934
 935                         remote_addr += len;
 936                         scat++;
 937                 }
 938
 939                 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
 940                         &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
 941
 942                 prev = send;
 943                 if (++send == &ic->i_sends[ic->i_send_ring.w_nr])
 944                         send = ic->i_sends;
 945         }
 946
 947         /* if we finished the message then send completion owns it */
 948         if (scat == &op->r_sg[op->r_count])
 949                 prev->s_wr.send_flags = IB_SEND_SIGNALED;
 950
 951         if (i < work_alloc) {
 952                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
 953                 work_alloc = i;
 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/IB: rdma ib_post_send to %pI4 "
 963                        "returned %d\n", &conn->c_faddr, ret);
 964                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 965                 goto out;
 966         }
 967
 968         if (unlikely(failed_wr != &first->s_wr)) {
 969                 printk(KERN_WARNING "RDS/IB: ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
 970                 BUG_ON(failed_wr != &first->s_wr);
 971         }
 972
 973
 974 out:
 975         return ret;
 976 }
 977
 978 void rds_ib_xmit_complete(struct rds_connection *conn)
 979 {
 980         struct rds_ib_connection *ic = conn->c_transport_data;
 981
 982         /* We may have a pending ACK or window update we were unable
 983          * to send previously (due to flow control). Try again. */
 984         rds_ib_attempt_ack(ic);
 985 }