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 #include <linux/ratelimit.h>
  38
  39 #include "rds.h"
  40 #include "ib.h"
  41
  42 static char *rds_ib_wc_status_strings[] = {
  43 #define RDS_IB_WC_STATUS_STR(foo) \
  44                 [IB_WC_##foo] = __stringify(IB_WC_##foo)
  45         RDS_IB_WC_STATUS_STR(SUCCESS),
  46         RDS_IB_WC_STATUS_STR(LOC_LEN_ERR),
  47         RDS_IB_WC_STATUS_STR(LOC_QP_OP_ERR),
  48         RDS_IB_WC_STATUS_STR(LOC_EEC_OP_ERR),
  49         RDS_IB_WC_STATUS_STR(LOC_PROT_ERR),
  50         RDS_IB_WC_STATUS_STR(WR_FLUSH_ERR),
  51         RDS_IB_WC_STATUS_STR(MW_BIND_ERR),
  52         RDS_IB_WC_STATUS_STR(BAD_RESP_ERR),
  53         RDS_IB_WC_STATUS_STR(LOC_ACCESS_ERR),
  54         RDS_IB_WC_STATUS_STR(REM_INV_REQ_ERR),
  55         RDS_IB_WC_STATUS_STR(REM_ACCESS_ERR),
  56         RDS_IB_WC_STATUS_STR(REM_OP_ERR),
  57         RDS_IB_WC_STATUS_STR(RETRY_EXC_ERR),
  58         RDS_IB_WC_STATUS_STR(RNR_RETRY_EXC_ERR),
  59         RDS_IB_WC_STATUS_STR(LOC_RDD_VIOL_ERR),
  60         RDS_IB_WC_STATUS_STR(REM_INV_RD_REQ_ERR),
  61         RDS_IB_WC_STATUS_STR(REM_ABORT_ERR),
  62         RDS_IB_WC_STATUS_STR(INV_EECN_ERR),
  63         RDS_IB_WC_STATUS_STR(INV_EEC_STATE_ERR),
  64         RDS_IB_WC_STATUS_STR(FATAL_ERR),
  65         RDS_IB_WC_STATUS_STR(RESP_TIMEOUT_ERR),
  66         RDS_IB_WC_STATUS_STR(GENERAL_ERR),
  67 #undef RDS_IB_WC_STATUS_STR
  68 };
  69
  70 char *rds_ib_wc_status_str(enum ib_wc_status status)
  71 {
  72         return rds_str_array(rds_ib_wc_status_strings,
  73                              ARRAY_SIZE(rds_ib_wc_status_strings), status);
  74 }
  75
  76 /*
  77  * Convert IB-specific error message to RDS error message and call core
  78  * completion handler.
  79  */
  80 static void rds_ib_send_complete(struct rds_message *rm,
  81                                  int wc_status,
  82                                  void (*complete)(struct rds_message *rm, int status))
  83 {
  84         int notify_status;
  85
  86         switch (wc_status) {
  87         case IB_WC_WR_FLUSH_ERR:
  88                 return;
  89
  90         case IB_WC_SUCCESS:
  91                 notify_status = RDS_RDMA_SUCCESS;
  92                 break;
  93
  94         case IB_WC_REM_ACCESS_ERR:
  95                 notify_status = RDS_RDMA_REMOTE_ERROR;
  96                 break;
  97
  98         default:
  99                 notify_status = RDS_RDMA_OTHER_ERROR;
 100                 break;
 101         }
 102         complete(rm, notify_status);
 103 }
 104
 105 static void rds_ib_send_unmap_data(struct rds_ib_connection *ic,
 106                                    struct rm_data_op *op,
 107                                    int wc_status)
 108 {
 109         if (op->op_nents)
 110                 ib_dma_unmap_sg(ic->i_cm_id->device,
 111                                 op->op_sg, op->op_nents,
 112                                 DMA_TO_DEVICE);
 113 }
 114
 115 static void rds_ib_send_unmap_rdma(struct rds_ib_connection *ic,
 116                                    struct rm_rdma_op *op,
 117                                    int wc_status)
 118 {
 119         if (op->op_mapped) {
 120                 ib_dma_unmap_sg(ic->i_cm_id->device,
 121                                 op->op_sg, op->op_nents,
 122                                 op->op_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
 123                 op->op_mapped = 0;
 124         }
 125
 126         /* If the user asked for a completion notification on this
 127          * message, we can implement three different semantics:
 128          *  1.  Notify when we received the ACK on the RDS message
 129          *      that was queued with the RDMA. This provides reliable
 130          *      notification of RDMA status at the expense of a one-way
 131          *      packet delay.
 132          *  2.  Notify when the IB stack gives us the completion event for
 133          *      the RDMA operation.
 134          *  3.  Notify when the IB stack gives us the completion event for
 135          *      the accompanying RDS messages.
 136          * Here, we implement approach #3. To implement approach #2,
 137          * we would need to take an event for the rdma WR. To implement #1,
 138          * don't call rds_rdma_send_complete at all, and fall back to the notify
 139          * handling in the ACK processing code.
 140          *
 141          * Note: There's no need to explicitly sync any RDMA buffers using
 142          * ib_dma_sync_sg_for_cpu - the completion for the RDMA
 143          * operation itself unmapped the RDMA buffers, which takes care
 144          * of synching.
 145          */
 146         rds_ib_send_complete(container_of(op, struct rds_message, rdma),
 147                              wc_status, rds_rdma_send_complete);
 148
 149         if (op->op_write)
 150                 rds_stats_add(s_send_rdma_bytes, op->op_bytes);
 151         else
 152                 rds_stats_add(s_recv_rdma_bytes, op->op_bytes);
 153 }
 154
 155 static void rds_ib_send_unmap_atomic(struct rds_ib_connection *ic,
 156                                      struct rm_atomic_op *op,
 157                                      int wc_status)
 158 {
 159         /* unmap atomic recvbuf */
 160         if (op->op_mapped) {
 161                 ib_dma_unmap_sg(ic->i_cm_id->device, op->op_sg, 1,
 162                                 DMA_FROM_DEVICE);
 163                 op->op_mapped = 0;
 164         }
 165
 166         rds_ib_send_complete(container_of(op, struct rds_message, atomic),
 167                              wc_status, rds_atomic_send_complete);
 168
 169         if (op->op_type == RDS_ATOMIC_TYPE_CSWP)
 170                 rds_ib_stats_inc(s_ib_atomic_cswp);
 171         else
 172                 rds_ib_stats_inc(s_ib_atomic_fadd);
 173 }
 174
 175 /*
 176  * Unmap the resources associated with a struct send_work.
 177  *
 178  * Returns the rm for no good reason other than it is unobtainable
 179  * other than by switching on wr.opcode, currently, and the caller,
 180  * the event handler, needs it.
 181  */
 182 static struct rds_message *rds_ib_send_unmap_op(struct rds_ib_connection *ic,
 183                                                 struct rds_ib_send_work *send,
 184                                                 int wc_status)
 185 {
 186         struct rds_message *rm = NULL;
 187
 188         /* In the error case, wc.opcode sometimes contains garbage */
 189         switch (send->s_wr.opcode) {
 190         case IB_WR_SEND:
 191                 if (send->s_op) {
 192                         rm = container_of(send->s_op, struct rds_message, data);
 193                         rds_ib_send_unmap_data(ic, send->s_op, wc_status);
 194                 }
 195                 break;
 196         case IB_WR_RDMA_WRITE:
 197         case IB_WR_RDMA_READ:
 198                 if (send->s_op) {
 199                         rm = container_of(send->s_op, struct rds_message, rdma);
 200                         rds_ib_send_unmap_rdma(ic, send->s_op, wc_status);
 201                 }
 202                 break;
 203         case IB_WR_ATOMIC_FETCH_AND_ADD:
 204         case IB_WR_ATOMIC_CMP_AND_SWP:
 205                 if (send->s_op) {
 206                         rm = container_of(send->s_op, struct rds_message, atomic);
 207                         rds_ib_send_unmap_atomic(ic, send->s_op, wc_status);
 208                 }
 209                 break;
 210         default:
 211                 printk_ratelimited(KERN_NOTICE
 212                                "RDS/IB: %s: unexpected opcode 0x%x in WR!\n",
 213                                __func__, send->s_wr.opcode);
 214                 break;
 215         }
 216
 217         send->s_wr.opcode = 0xdead;
 218
 219         return rm;
 220 }
 221
 222 void rds_ib_send_init_ring(struct rds_ib_connection *ic)
 223 {
 224         struct rds_ib_send_work *send;
 225         u32 i;
 226
 227         for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
 228                 struct ib_sge *sge;
 229
 230                 send->s_op = NULL;
 231
 232                 send->s_wr.wr_id = i;
 233                 send->s_wr.sg_list = send->s_sge;
 234                 send->s_wr.ex.imm_data = 0;
 235
 236                 sge = &send->s_sge[0];
 237                 sge->addr = ic->i_send_hdrs_dma + (i * sizeof(struct rds_header));
 238                 sge->length = sizeof(struct rds_header);
 239                 sge->lkey = ic->i_mr->lkey;
 240
 241                 send->s_sge[1].lkey = ic->i_mr->lkey;
 242         }
 243 }
 244
 245 void rds_ib_send_clear_ring(struct rds_ib_connection *ic)
 246 {
 247         struct rds_ib_send_work *send;
 248         u32 i;
 249
 250         for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
 251                 if (send->s_op && send->s_wr.opcode != 0xdead)
 252                         rds_ib_send_unmap_op(ic, send, IB_WC_WR_FLUSH_ERR);
 253         }
 254 }
 255
 256 /*
 257  * The only fast path caller always has a non-zero nr, so we don't
 258  * bother testing nr before performing the atomic sub.
 259  */
 260 static void rds_ib_sub_signaled(struct rds_ib_connection *ic, int nr)
 261 {
 262         if ((atomic_sub_return(nr, &ic->i_signaled_sends) == 0) &&
 263             waitqueue_active(&rds_ib_ring_empty_wait))
 264                 wake_up(&rds_ib_ring_empty_wait);
 265         BUG_ON(atomic_read(&ic->i_signaled_sends) < 0);
 266 }
 267
 268 /*
 269  * The _oldest/_free ring operations here race cleanly with the alloc/unalloc
 270  * operations performed in the send path.  As the sender allocs and potentially
 271  * unallocs the next free entry in the ring it doesn't alter which is
 272  * the next to be freed, which is what this is concerned with.
 273  */
 274 void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context)
 275 {
 276         struct rds_connection *conn = context;
 277         struct rds_ib_connection *ic = conn->c_transport_data;
 278         struct rds_message *rm = NULL;
 279         struct ib_wc wc;
 280         struct rds_ib_send_work *send;
 281         u32 completed;
 282         u32 oldest;
 283         u32 i = 0;
 284         int ret;
 285         int nr_sig = 0;
 286
 287         rdsdebug("cq %p conn %p\n", cq, conn);
 288         rds_ib_stats_inc(s_ib_tx_cq_call);
 289         ret = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 290         if (ret)
 291                 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
 292
 293         while (ib_poll_cq(cq, 1, &wc) > 0) {
 294                 rdsdebug("wc wr_id 0x%llx status %u (%s) byte_len %u imm_data %u\n",
 295                          (unsigned long long)wc.wr_id, wc.status,
 296                          rds_ib_wc_status_str(wc.status), wc.byte_len,
 297                          be32_to_cpu(wc.ex.imm_data));
 298                 rds_ib_stats_inc(s_ib_tx_cq_event);
 299
 300                 if (wc.wr_id == RDS_IB_ACK_WR_ID) {
 301                         if (ic->i_ack_queued + HZ/2 < jiffies)
 302                                 rds_ib_stats_inc(s_ib_tx_stalled);
 303                         rds_ib_ack_send_complete(ic);
 304                         continue;
 305                 }
 306
 307                 oldest = rds_ib_ring_oldest(&ic->i_send_ring);
 308
 309                 completed = rds_ib_ring_completed(&ic->i_send_ring, wc.wr_id, oldest);
 310
 311                 for (i = 0; i < completed; i++) {
 312                         send = &ic->i_sends[oldest];
 313                         if (send->s_wr.send_flags & IB_SEND_SIGNALED)
 314                                 nr_sig++;
 315
 316                         rm = rds_ib_send_unmap_op(ic, send, wc.status);
 317
 318                         if (send->s_queued + HZ/2 < jiffies)
 319                                 rds_ib_stats_inc(s_ib_tx_stalled);
 320
 321                         if (send->s_op) {
 322                                 if (send->s_op == rm->m_final_op) {
 323                                         /* If anyone waited for this message to get flushed out, wake
 324                                          * them up now */
 325                                         rds_message_unmapped(rm);
 326                                 }
 327                                 rds_message_put(rm);
 328                                 send->s_op = NULL;
 329                         }
 330
 331                         oldest = (oldest + 1) % ic->i_send_ring.w_nr;
 332                 }
 333
 334                 rds_ib_ring_free(&ic->i_send_ring, completed);
 335                 rds_ib_sub_signaled(ic, nr_sig);
 336                 nr_sig = 0;
 337
 338                 if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
 339                     test_bit(0, &conn->c_map_queued))
 340                         queue_delayed_work(rds_wq, &conn->c_send_w, 0);
 341
 342                 /* We expect errors as the qp is drained during shutdown */
 343                 if (wc.status != IB_WC_SUCCESS && rds_conn_up(conn)) {
 344                         rds_ib_conn_error(conn, "send completion on %pI4 had status "
 345                                           "%u (%s), disconnecting and reconnecting\n",
 346                                           &conn->c_faddr, wc.status,
 347                                           rds_ib_wc_status_str(wc.status));
 348                 }
 349         }
 350 }
 351
 352 /*
 353  * This is the main function for allocating credits when sending
 354  * messages.
 355  *
 356  * Conceptually, we have two counters:
 357  *  -   send credits: this tells us how many WRs we're allowed
 358  *      to submit without overruning the receiver's queue. For
 359  *      each SEND WR we post, we decrement this by one.
 360  *
 361  *  -   posted credits: this tells us how many WRs we recently
 362  *      posted to the receive queue. This value is transferred
 363  *      to the peer as a "credit update" in a RDS header field.
 364  *      Every time we transmit credits to the peer, we subtract
 365  *      the amount of transferred credits from this counter.
 366  *
 367  * It is essential that we avoid situations where both sides have
 368  * exhausted their send credits, and are unable to send new credits
 369  * to the peer. We achieve this by requiring that we send at least
 370  * one credit update to the peer before exhausting our credits.
 371  * When new credits arrive, we subtract one credit that is withheld
 372  * until we've posted new buffers and are ready to transmit these
 373  * credits (see rds_ib_send_add_credits below).
 374  *
 375  * The RDS send code is essentially single-threaded; rds_send_xmit
 376  * sets RDS_IN_XMIT to ensure exclusive access to the send ring.
 377  * However, the ACK sending code is independent and can race with
 378  * message SENDs.
 379  *
 380  * In the send path, we need to update the counters for send credits
 381  * and the counter of posted buffers atomically - when we use the
 382  * last available credit, we cannot allow another thread to race us
 383  * and grab the posted credits counter.  Hence, we have to use a
 384  * spinlock to protect the credit counter, or use atomics.
 385  *
 386  * Spinlocks shared between the send and the receive path are bad,
 387  * because they create unnecessary delays. An early implementation
 388  * using a spinlock showed a 5% degradation in throughput at some
 389  * loads.
 390  *
 391  * This implementation avoids spinlocks completely, putting both
 392  * counters into a single atomic, and updating that atomic using
 393  * atomic_add (in the receive path, when receiving fresh credits),
 394  * and using atomic_cmpxchg when updating the two counters.
 395  */
 396 int rds_ib_send_grab_credits(struct rds_ib_connection *ic,
 397                              u32 wanted, u32 *adv_credits, int need_posted, int max_posted)
 398 {
 399         unsigned int avail, posted, got = 0, advertise;
 400         long oldval, newval;
 401
 402         *adv_credits = 0;
 403         if (!ic->i_flowctl)
 404                 return wanted;
 405
 406 try_again:
 407         advertise = 0;
 408         oldval = newval = atomic_read(&ic->i_credits);
 409         posted = IB_GET_POST_CREDITS(oldval);
 410         avail = IB_GET_SEND_CREDITS(oldval);
 411
 412         rdsdebug("rds_ib_send_grab_credits(%u): credits=%u posted=%u\n",
 413                         wanted, avail, posted);
 414
 415         /* The last credit must be used to send a credit update. */
 416         if (avail && !posted)
 417                 avail--;
 418
 419         if (avail < wanted) {
 420                 struct rds_connection *conn = ic->i_cm_id->context;
 421
 422                 /* Oops, there aren't that many credits left! */
 423                 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 424                 got = avail;
 425         } else {
 426                 /* Sometimes you get what you want, lalala. */
 427                 got = wanted;
 428         }
 429         newval -= IB_SET_SEND_CREDITS(got);
 430
 431         /*
 432          * If need_posted is non-zero, then the caller wants
 433          * the posted regardless of whether any send credits are
 434          * available.
 435          */
 436         if (posted && (got || need_posted)) {
 437                 advertise = min_t(unsigned int, posted, max_posted);
 438                 newval -= IB_SET_POST_CREDITS(advertise);
 439         }
 440
 441         /* Finally bill everything */
 442         if (atomic_cmpxchg(&ic->i_credits, oldval, newval) != oldval)
 443                 goto try_again;
 444
 445         *adv_credits = advertise;
 446         return got;
 447 }
 448
 449 void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits)
 450 {
 451         struct rds_ib_connection *ic = conn->c_transport_data;
 452
 453         if (credits == 0)
 454                 return;
 455
 456         rdsdebug("rds_ib_send_add_credits(%u): current=%u%s\n",
 457                         credits,
 458                         IB_GET_SEND_CREDITS(atomic_read(&ic->i_credits)),
 459                         test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ? ", ll_send_full" : "");
 460
 461         atomic_add(IB_SET_SEND_CREDITS(credits), &ic->i_credits);
 462         if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags))
 463                 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
 464
 465         WARN_ON(IB_GET_SEND_CREDITS(credits) >= 16384);
 466
 467         rds_ib_stats_inc(s_ib_rx_credit_updates);
 468 }
 469
 470 void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted)
 471 {
 472         struct rds_ib_connection *ic = conn->c_transport_data;
 473
 474         if (posted == 0)
 475                 return;
 476
 477         atomic_add(IB_SET_POST_CREDITS(posted), &ic->i_credits);
 478
 479         /* Decide whether to send an update to the peer now.
 480          * If we would send a credit update for every single buffer we
 481          * post, we would end up with an ACK storm (ACK arrives,
 482          * consumes buffer, we refill the ring, send ACK to remote
 483          * advertising the newly posted buffer... ad inf)
 484          *
 485          * Performance pretty much depends on how often we send
 486          * credit updates - too frequent updates mean lots of ACKs.
 487          * Too infrequent updates, and the peer will run out of
 488          * credits and has to throttle.
 489          * For the time being, 16 seems to be a good compromise.
 490          */
 491         if (IB_GET_POST_CREDITS(atomic_read(&ic->i_credits)) >= 16)
 492                 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
 493 }
 494
 495 static inline int rds_ib_set_wr_signal_state(struct rds_ib_connection *ic,
 496                                              struct rds_ib_send_work *send,
 497                                              bool notify)
 498 {
 499         /*
 500          * We want to delay signaling completions just enough to get
 501          * the batching benefits but not so much that we create dead time
 502          * on the wire.
 503          */
 504         if (ic->i_unsignaled_wrs-- == 0 || notify) {
 505                 ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
 506                 send->s_wr.send_flags |= IB_SEND_SIGNALED;
 507                 return 1;
 508         }
 509         return 0;
 510 }
 511
 512 /*
 513  * This can be called multiple times for a given message.  The first time
 514  * we see a message we map its scatterlist into the IB device so that
 515  * we can provide that mapped address to the IB scatter gather entries
 516  * in the IB work requests.  We translate the scatterlist into a series
 517  * of work requests that fragment the message.  These work requests complete
 518  * in order so we pass ownership of the message to the completion handler
 519  * once we send the final fragment.
 520  *
 521  * The RDS core uses the c_send_lock to only enter this function once
 522  * per connection.  This makes sure that the tx ring alloc/unalloc pairs
 523  * don't get out of sync and confuse the ring.
 524  */
 525 int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
 526                 unsigned int hdr_off, unsigned int sg, unsigned int off)
 527 {
 528         struct rds_ib_connection *ic = conn->c_transport_data;
 529         struct ib_device *dev = ic->i_cm_id->device;
 530         struct rds_ib_send_work *send = NULL;
 531         struct rds_ib_send_work *first;
 532         struct rds_ib_send_work *prev;
 533         struct ib_send_wr *failed_wr;
 534         struct scatterlist *scat;
 535         u32 pos;
 536         u32 i;
 537         u32 work_alloc;
 538         u32 credit_alloc = 0;
 539         u32 posted;
 540         u32 adv_credits = 0;
 541         int send_flags = 0;
 542         int bytes_sent = 0;
 543         int ret;
 544         int flow_controlled = 0;
 545         int nr_sig = 0;
 546
 547         BUG_ON(off % RDS_FRAG_SIZE);
 548         BUG_ON(hdr_off != 0 && hdr_off != sizeof(struct rds_header));
 549
 550         /* Do not send cong updates to IB loopback */
 551         if (conn->c_loopback
 552             && rm->m_inc.i_hdr.h_flags & RDS_FLAG_CONG_BITMAP) {
 553                 rds_cong_map_updated(conn->c_fcong, ~(u64) 0);
 554                 scat = &rm->data.op_sg[sg];
 555                 ret = sizeof(struct rds_header) + RDS_CONG_MAP_BYTES;
 556                 ret = min_t(int, ret, scat->length - conn->c_xmit_data_off);
 557                 return ret;
 558         }
 559
 560         /* FIXME we may overallocate here */
 561         if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0)
 562                 i = 1;
 563         else
 564                 i = ceil(be32_to_cpu(rm->m_inc.i_hdr.h_len), RDS_FRAG_SIZE);
 565
 566         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
 567         if (work_alloc == 0) {
 568                 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 569                 rds_ib_stats_inc(s_ib_tx_ring_full);
 570                 ret = -ENOMEM;
 571                 goto out;
 572         }
 573
 574         if (ic->i_flowctl) {
 575                 credit_alloc = rds_ib_send_grab_credits(ic, work_alloc, &posted, 0, RDS_MAX_ADV_CREDIT);
 576                 adv_credits += posted;
 577                 if (credit_alloc < work_alloc) {
 578                         rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - credit_alloc);
 579                         work_alloc = credit_alloc;
 580                         flow_controlled = 1;
 581                 }
 582                 if (work_alloc == 0) {
 583                         set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 584                         rds_ib_stats_inc(s_ib_tx_throttle);
 585                         ret = -ENOMEM;
 586                         goto out;
 587                 }
 588         }
 589
 590         /* map the message the first time we see it */
 591         if (!ic->i_data_op) {
 592                 if (rm->data.op_nents) {
 593                         rm->data.op_count = ib_dma_map_sg(dev,
 594                                                           rm->data.op_sg,
 595                                                           rm->data.op_nents,
 596                                                           DMA_TO_DEVICE);
 597                         rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->data.op_count);
 598                         if (rm->data.op_count == 0) {
 599                                 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
 600                                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 601                                 ret = -ENOMEM; /* XXX ? */
 602                                 goto out;
 603                         }
 604                 } else {
 605                         rm->data.op_count = 0;
 606                 }
 607
 608                 rds_message_addref(rm);
 609                 ic->i_data_op = &rm->data;
 610
 611                 /* Finalize the header */
 612                 if (test_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags))
 613                         rm->m_inc.i_hdr.h_flags |= RDS_FLAG_ACK_REQUIRED;
 614                 if (test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))
 615                         rm->m_inc.i_hdr.h_flags |= RDS_FLAG_RETRANSMITTED;
 616
 617                 /* If it has a RDMA op, tell the peer we did it. This is
 618                  * used by the peer to release use-once RDMA MRs. */
 619                 if (rm->rdma.op_active) {
 620                         struct rds_ext_header_rdma ext_hdr;
 621
 622                         ext_hdr.h_rdma_rkey = cpu_to_be32(rm->rdma.op_rkey);
 623                         rds_message_add_extension(&rm->m_inc.i_hdr,
 624                                         RDS_EXTHDR_RDMA, &ext_hdr, sizeof(ext_hdr));
 625                 }
 626                 if (rm->m_rdma_cookie) {
 627                         rds_message_add_rdma_dest_extension(&rm->m_inc.i_hdr,
 628                                         rds_rdma_cookie_key(rm->m_rdma_cookie),
 629                                         rds_rdma_cookie_offset(rm->m_rdma_cookie));
 630                 }
 631
 632                 /* Note - rds_ib_piggyb_ack clears the ACK_REQUIRED bit, so
 633                  * we should not do this unless we have a chance of at least
 634                  * sticking the header into the send ring. Which is why we
 635                  * should call rds_ib_ring_alloc first. */
 636                 rm->m_inc.i_hdr.h_ack = cpu_to_be64(rds_ib_piggyb_ack(ic));
 637                 rds_message_make_checksum(&rm->m_inc.i_hdr);
 638
 639                 /*
 640                  * Update adv_credits since we reset the ACK_REQUIRED bit.
 641                  */
 642                 if (ic->i_flowctl) {
 643                         rds_ib_send_grab_credits(ic, 0, &posted, 1, RDS_MAX_ADV_CREDIT - adv_credits);
 644                         adv_credits += posted;
 645                         BUG_ON(adv_credits > 255);
 646                 }
 647         }
 648
 649         /* Sometimes you want to put a fence between an RDMA
 650          * READ and the following SEND.
 651          * We could either do this all the time
 652          * or when requested by the user. Right now, we let
 653          * the application choose.
 654          */
 655         if (rm->rdma.op_active && rm->rdma.op_fence)
 656                 send_flags = IB_SEND_FENCE;
 657
 658         /* Each frag gets a header. Msgs may be 0 bytes */
 659         send = &ic->i_sends[pos];
 660         first = send;
 661         prev = NULL;
 662         scat = &ic->i_data_op->op_sg[sg];
 663         i = 0;
 664         do {
 665                 unsigned int len = 0;
 666
 667                 /* Set up the header */
 668                 send->s_wr.send_flags = send_flags;
 669                 send->s_wr.opcode = IB_WR_SEND;
 670                 send->s_wr.num_sge = 1;
 671                 send->s_wr.next = NULL;
 672                 send->s_queued = jiffies;
 673                 send->s_op = NULL;
 674
 675                 send->s_sge[0].addr = ic->i_send_hdrs_dma
 676                         + (pos * sizeof(struct rds_header));
 677                 send->s_sge[0].length = sizeof(struct rds_header);
 678
 679                 memcpy(&ic->i_send_hdrs[pos], &rm->m_inc.i_hdr, sizeof(struct rds_header));
 680
 681                 /* Set up the data, if present */
 682                 if (i < work_alloc
 683                     && scat != &rm->data.op_sg[rm->data.op_count]) {
 684                         len = min(RDS_FRAG_SIZE, ib_sg_dma_len(dev, scat) - off);
 685                         send->s_wr.num_sge = 2;
 686
 687                         send->s_sge[1].addr = ib_sg_dma_address(dev, scat) + off;
 688                         send->s_sge[1].length = len;
 689
 690                         bytes_sent += len;
 691                         off += len;
 692                         if (off == ib_sg_dma_len(dev, scat)) {
 693                                 scat++;
 694                                 off = 0;
 695                         }
 696                 }
 697
 698                 rds_ib_set_wr_signal_state(ic, send, 0);
 699
 700                 /*
 701                  * Always signal the last one if we're stopping due to flow control.
 702                  */
 703                 if (ic->i_flowctl && flow_controlled && i == (work_alloc-1))
 704                         send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
 705
 706                 if (send->s_wr.send_flags & IB_SEND_SIGNALED)
 707                         nr_sig++;
 708
 709                 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
 710                          &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
 711
 712                 if (ic->i_flowctl && adv_credits) {
 713                         struct rds_header *hdr = &ic->i_send_hdrs[pos];
 714
 715                         /* add credit and redo the header checksum */
 716                         hdr->h_credit = adv_credits;
 717                         rds_message_make_checksum(hdr);
 718                         adv_credits = 0;
 719                         rds_ib_stats_inc(s_ib_tx_credit_updates);
 720                 }
 721
 722                 if (prev)
 723                         prev->s_wr.next = &send->s_wr;
 724                 prev = send;
 725
 726                 pos = (pos + 1) % ic->i_send_ring.w_nr;
 727                 send = &ic->i_sends[pos];
 728                 i++;
 729
 730         } while (i < work_alloc
 731                  && scat != &rm->data.op_sg[rm->data.op_count]);
 732
 733         /* Account the RDS header in the number of bytes we sent, but just once.
 734          * The caller has no concept of fragmentation. */
 735         if (hdr_off == 0)
 736                 bytes_sent += sizeof(struct rds_header);
 737
 738         /* if we finished the message then send completion owns it */
 739         if (scat == &rm->data.op_sg[rm->data.op_count]) {
 740                 prev->s_op = ic->i_data_op;
 741                 prev->s_wr.send_flags |= IB_SEND_SOLICITED;
 742                 ic->i_data_op = NULL;
 743         }
 744
 745         /* Put back wrs & credits we didn't use */
 746         if (i < work_alloc) {
 747                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
 748                 work_alloc = i;
 749         }
 750         if (ic->i_flowctl && i < credit_alloc)
 751                 rds_ib_send_add_credits(conn, credit_alloc - i);
 752
 753         if (nr_sig)
 754                 atomic_add(nr_sig, &ic->i_signaled_sends);
 755
 756         /* XXX need to worry about failed_wr and partial sends. */
 757         failed_wr = &first->s_wr;
 758         ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
 759         rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
 760                  first, &first->s_wr, ret, failed_wr);
 761         BUG_ON(failed_wr != &first->s_wr);
 762         if (ret) {
 763                 printk(KERN_WARNING "RDS/IB: ib_post_send to %pI4 "
 764                        "returned %d\n", &conn->c_faddr, ret);
 765                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 766                 rds_ib_sub_signaled(ic, nr_sig);
 767                 if (prev->s_op) {
 768                         ic->i_data_op = prev->s_op;
 769                         prev->s_op = NULL;
 770                 }
 771
 772                 rds_ib_conn_error(ic->conn, "ib_post_send failed\n");
 773                 goto out;
 774         }
 775
 776         ret = bytes_sent;
 777 out:
 778         BUG_ON(adv_credits);
 779         return ret;
 780 }
 781
 782 /*
 783  * Issue atomic operation.
 784  * A simplified version of the rdma case, we always map 1 SG, and
 785  * only 8 bytes, for the return value from the atomic operation.
 786  */
 787 int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op)
 788 {
 789         struct rds_ib_connection *ic = conn->c_transport_data;
 790         struct rds_ib_send_work *send = NULL;
 791         struct ib_send_wr *failed_wr;
 792         struct rds_ib_device *rds_ibdev;
 793         u32 pos;
 794         u32 work_alloc;
 795         int ret;
 796         int nr_sig = 0;
 797
 798         rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
 799
 800         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, 1, &pos);
 801         if (work_alloc != 1) {
 802                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 803                 rds_ib_stats_inc(s_ib_tx_ring_full);
 804                 ret = -ENOMEM;
 805                 goto out;
 806         }
 807
 808         /* address of send request in ring */
 809         send = &ic->i_sends[pos];
 810         send->s_queued = jiffies;
 811
 812         if (op->op_type == RDS_ATOMIC_TYPE_CSWP) {
 813                 send->s_wr.opcode = IB_WR_MASKED_ATOMIC_CMP_AND_SWP;
 814                 send->s_wr.wr.atomic.compare_add = op->op_m_cswp.compare;
 815                 send->s_wr.wr.atomic.swap = op->op_m_cswp.swap;
 816                 send->s_wr.wr.atomic.compare_add_mask = op->op_m_cswp.compare_mask;
 817                 send->s_wr.wr.atomic.swap_mask = op->op_m_cswp.swap_mask;
 818         } else { /* FADD */
 819                 send->s_wr.opcode = IB_WR_MASKED_ATOMIC_FETCH_AND_ADD;
 820                 send->s_wr.wr.atomic.compare_add = op->op_m_fadd.add;
 821                 send->s_wr.wr.atomic.swap = 0;
 822                 send->s_wr.wr.atomic.compare_add_mask = op->op_m_fadd.nocarry_mask;
 823                 send->s_wr.wr.atomic.swap_mask = 0;
 824         }
 825         nr_sig = rds_ib_set_wr_signal_state(ic, send, op->op_notify);
 826         send->s_wr.num_sge = 1;
 827         send->s_wr.next = NULL;
 828         send->s_wr.wr.atomic.remote_addr = op->op_remote_addr;
 829         send->s_wr.wr.atomic.rkey = op->op_rkey;
 830         send->s_op = op;
 831         rds_message_addref(container_of(send->s_op, struct rds_message, atomic));
 832
 833         /* map 8 byte retval buffer to the device */
 834         ret = ib_dma_map_sg(ic->i_cm_id->device, op->op_sg, 1, DMA_FROM_DEVICE);
 835         rdsdebug("ic %p mapping atomic op %p. mapped %d pg\n", ic, op, ret);
 836         if (ret != 1) {
 837                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 838                 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
 839                 ret = -ENOMEM; /* XXX ? */
 840                 goto out;
 841         }
 842
 843         /* Convert our struct scatterlist to struct ib_sge */
 844         send->s_sge[0].addr = ib_sg_dma_address(ic->i_cm_id->device, op->op_sg);
 845         send->s_sge[0].length = ib_sg_dma_len(ic->i_cm_id->device, op->op_sg);
 846         send->s_sge[0].lkey = ic->i_mr->lkey;
 847
 848         rdsdebug("rva %Lx rpa %Lx len %u\n", op->op_remote_addr,
 849                  send->s_sge[0].addr, send->s_sge[0].length);
 850
 851         if (nr_sig)
 852                 atomic_add(nr_sig, &ic->i_signaled_sends);
 853
 854         failed_wr = &send->s_wr;
 855         ret = ib_post_send(ic->i_cm_id->qp, &send->s_wr, &failed_wr);
 856         rdsdebug("ic %p send %p (wr %p) ret %d wr %p\n", ic,
 857                  send, &send->s_wr, ret, failed_wr);
 858         BUG_ON(failed_wr != &send->s_wr);
 859         if (ret) {
 860                 printk(KERN_WARNING "RDS/IB: atomic ib_post_send to %pI4 "
 861                        "returned %d\n", &conn->c_faddr, ret);
 862                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 863                 rds_ib_sub_signaled(ic, nr_sig);
 864                 goto out;
 865         }
 866
 867         if (unlikely(failed_wr != &send->s_wr)) {
 868                 printk(KERN_WARNING "RDS/IB: atomic ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
 869                 BUG_ON(failed_wr != &send->s_wr);
 870         }
 871
 872 out:
 873         return ret;
 874 }
 875
 876 int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op)
 877 {
 878         struct rds_ib_connection *ic = conn->c_transport_data;
 879         struct rds_ib_send_work *send = NULL;
 880         struct rds_ib_send_work *first;
 881         struct rds_ib_send_work *prev;
 882         struct ib_send_wr *failed_wr;
 883         struct scatterlist *scat;
 884         unsigned long len;
 885         u64 remote_addr = op->op_remote_addr;
 886         u32 max_sge = ic->rds_ibdev->max_sge;
 887         u32 pos;
 888         u32 work_alloc;
 889         u32 i;
 890         u32 j;
 891         int sent;
 892         int ret;
 893         int num_sge;
 894         int nr_sig = 0;
 895
 896         /* map the op the first time we see it */
 897         if (!op->op_mapped) {
 898                 op->op_count = ib_dma_map_sg(ic->i_cm_id->device,
 899                                              op->op_sg, op->op_nents, (op->op_write) ?
 900                                              DMA_TO_DEVICE : DMA_FROM_DEVICE);
 901                 rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->op_count);
 902                 if (op->op_count == 0) {
 903                         rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
 904                         ret = -ENOMEM; /* XXX ? */
 905                         goto out;
 906                 }
 907
 908                 op->op_mapped = 1;
 909         }
 910
 911         /*
 912          * Instead of knowing how to return a partial rdma read/write we insist that there
 913          * be enough work requests to send the entire message.
 914          */
 915         i = ceil(op->op_count, max_sge);
 916
 917         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
 918         if (work_alloc != i) {
 919                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 920                 rds_ib_stats_inc(s_ib_tx_ring_full);
 921                 ret = -ENOMEM;
 922                 goto out;
 923         }
 924
 925         send = &ic->i_sends[pos];
 926         first = send;
 927         prev = NULL;
 928         scat = &op->op_sg[0];
 929         sent = 0;
 930         num_sge = op->op_count;
 931
 932         for (i = 0; i < work_alloc && scat != &op->op_sg[op->op_count]; i++) {
 933                 send->s_wr.send_flags = 0;
 934                 send->s_queued = jiffies;
 935                 send->s_op = NULL;
 936
 937                 nr_sig += rds_ib_set_wr_signal_state(ic, send, op->op_notify);
 938
 939                 send->s_wr.opcode = op->op_write ? IB_WR_RDMA_WRITE : IB_WR_RDMA_READ;
 940                 send->s_wr.wr.rdma.remote_addr = remote_addr;
 941                 send->s_wr.wr.rdma.rkey = op->op_rkey;
 942
 943                 if (num_sge > max_sge) {
 944                         send->s_wr.num_sge = max_sge;
 945                         num_sge -= max_sge;
 946                 } else {
 947                         send->s_wr.num_sge = num_sge;
 948                 }
 949
 950                 send->s_wr.next = NULL;
 951
 952                 if (prev)
 953                         prev->s_wr.next = &send->s_wr;
 954
 955                 for (j = 0; j < send->s_wr.num_sge && scat != &op->op_sg[op->op_count]; j++) {
 956                         len = ib_sg_dma_len(ic->i_cm_id->device, scat);
 957                         send->s_sge[j].addr =
 958                                  ib_sg_dma_address(ic->i_cm_id->device, scat);
 959                         send->s_sge[j].length = len;
 960                         send->s_sge[j].lkey = ic->i_mr->lkey;
 961
 962                         sent += len;
 963                         rdsdebug("ic %p sent %d remote_addr %llu\n", ic, sent, remote_addr);
 964
 965                         remote_addr += len;
 966                         scat++;
 967                 }
 968
 969                 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
 970                         &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
 971
 972                 prev = send;
 973                 if (++send == &ic->i_sends[ic->i_send_ring.w_nr])
 974                         send = ic->i_sends;
 975         }
 976
 977         /* give a reference to the last op */
 978         if (scat == &op->op_sg[op->op_count]) {
 979                 prev->s_op = op;
 980                 rds_message_addref(container_of(op, struct rds_message, rdma));
 981         }
 982
 983         if (i < work_alloc) {
 984                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
 985                 work_alloc = i;
 986         }
 987
 988         if (nr_sig)
 989                 atomic_add(nr_sig, &ic->i_signaled_sends);
 990
 991         failed_wr = &first->s_wr;
 992         ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
 993         rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
 994                  first, &first->s_wr, ret, failed_wr);
 995         BUG_ON(failed_wr != &first->s_wr);
 996         if (ret) {
 997                 printk(KERN_WARNING "RDS/IB: rdma ib_post_send to %pI4 "
 998                        "returned %d\n", &conn->c_faddr, ret);
 999                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
1000                 rds_ib_sub_signaled(ic, nr_sig);
1001                 goto out;
1002         }
1003
1004         if (unlikely(failed_wr != &first->s_wr)) {
1005                 printk(KERN_WARNING "RDS/IB: ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
1006                 BUG_ON(failed_wr != &first->s_wr);
1007         }
1008
1009
1010 out:
1011         return ret;
1012 }
1013
1014 void rds_ib_xmit_complete(struct rds_connection *conn)
1015 {
1016         struct rds_ib_connection *ic = conn->c_transport_data;
1017
1018         /* We may have a pending ACK or window update we were unable
1019          * to send previously (due to flow control). Try again. */
1020         rds_ib_attempt_ack(ic);
1021 }