net/rds/ib_send.c

   1 /*
   2  * Copyright (c) 2006, 2019 Oracle and/or its affiliates. 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_single_path.h"
  40 #include "rds.h"
  41 #include "ib.h"
  42 #include "ib_mr.h"
  43
  44 /*
  45  * Convert IB-specific error message to RDS error message and call core
  46  * completion handler.
  47  */
  48 static void rds_ib_send_complete(struct rds_message *rm,
  49                                  int wc_status,
  50                                  void (*complete)(struct rds_message *rm, int status))
  51 {
  52         int notify_status;
  53
  54         switch (wc_status) {
  55         case IB_WC_WR_FLUSH_ERR:
  56                 return;
  57
  58         case IB_WC_SUCCESS:
  59                 notify_status = RDS_RDMA_SUCCESS;
  60                 break;
  61
  62         case IB_WC_REM_ACCESS_ERR:
  63                 notify_status = RDS_RDMA_REMOTE_ERROR;
  64                 break;
  65
  66         default:
  67                 notify_status = RDS_RDMA_OTHER_ERROR;
  68                 break;
  69         }
  70         complete(rm, notify_status);
  71 }
  72
  73 static void rds_ib_send_unmap_data(struct rds_ib_connection *ic,
  74                                    struct rm_data_op *op,
  75                                    int wc_status)
  76 {
  77         if (op->op_nents)
  78                 ib_dma_unmap_sg(ic->i_cm_id->device,
  79                                 op->op_sg, op->op_nents,
  80                                 DMA_TO_DEVICE);
  81 }
  82
  83 static void rds_ib_send_unmap_rdma(struct rds_ib_connection *ic,
  84                                    struct rm_rdma_op *op,
  85                                    int wc_status)
  86 {
  87         if (op->op_mapped) {
  88                 ib_dma_unmap_sg(ic->i_cm_id->device,
  89                                 op->op_sg, op->op_nents,
  90                                 op->op_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
  91                 op->op_mapped = 0;
  92         }
  93
  94         /* If the user asked for a completion notification on this
  95          * message, we can implement three different semantics:
  96          *  1.  Notify when we received the ACK on the RDS message
  97          *      that was queued with the RDMA. This provides reliable
  98          *      notification of RDMA status at the expense of a one-way
  99          *      packet delay.
 100          *  2.  Notify when the IB stack gives us the completion event for
 101          *      the RDMA operation.
 102          *  3.  Notify when the IB stack gives us the completion event for
 103          *      the accompanying RDS messages.
 104          * Here, we implement approach #3. To implement approach #2,
 105          * we would need to take an event for the rdma WR. To implement #1,
 106          * don't call rds_rdma_send_complete at all, and fall back to the notify
 107          * handling in the ACK processing code.
 108          *
 109          * Note: There's no need to explicitly sync any RDMA buffers using
 110          * ib_dma_sync_sg_for_cpu - the completion for the RDMA
 111          * operation itself unmapped the RDMA buffers, which takes care
 112          * of synching.
 113          */
 114         rds_ib_send_complete(container_of(op, struct rds_message, rdma),
 115                              wc_status, rds_rdma_send_complete);
 116
 117         if (op->op_write)
 118                 rds_stats_add(s_send_rdma_bytes, op->op_bytes);
 119         else
 120                 rds_stats_add(s_recv_rdma_bytes, op->op_bytes);
 121 }
 122
 123 static void rds_ib_send_unmap_atomic(struct rds_ib_connection *ic,
 124                                      struct rm_atomic_op *op,
 125                                      int wc_status)
 126 {
 127         /* unmap atomic recvbuf */
 128         if (op->op_mapped) {
 129                 ib_dma_unmap_sg(ic->i_cm_id->device, op->op_sg, 1,
 130                                 DMA_FROM_DEVICE);
 131                 op->op_mapped = 0;
 132         }
 133
 134         rds_ib_send_complete(container_of(op, struct rds_message, atomic),
 135                              wc_status, rds_atomic_send_complete);
 136
 137         if (op->op_type == RDS_ATOMIC_TYPE_CSWP)
 138                 rds_ib_stats_inc(s_ib_atomic_cswp);
 139         else
 140                 rds_ib_stats_inc(s_ib_atomic_fadd);
 141 }
 142
 143 /*
 144  * Unmap the resources associated with a struct send_work.
 145  *
 146  * Returns the rm for no good reason other than it is unobtainable
 147  * other than by switching on wr.opcode, currently, and the caller,
 148  * the event handler, needs it.
 149  */
 150 static struct rds_message *rds_ib_send_unmap_op(struct rds_ib_connection *ic,
 151                                                 struct rds_ib_send_work *send,
 152                                                 int wc_status)
 153 {
 154         struct rds_message *rm = NULL;
 155
 156         /* In the error case, wc.opcode sometimes contains garbage */
 157         switch (send->s_wr.opcode) {
 158         case IB_WR_SEND:
 159                 if (send->s_op) {
 160                         rm = container_of(send->s_op, struct rds_message, data);
 161                         rds_ib_send_unmap_data(ic, send->s_op, wc_status);
 162                 }
 163                 break;
 164         case IB_WR_RDMA_WRITE:
 165         case IB_WR_RDMA_READ:
 166                 if (send->s_op) {
 167                         rm = container_of(send->s_op, struct rds_message, rdma);
 168                         rds_ib_send_unmap_rdma(ic, send->s_op, wc_status);
 169                 }
 170                 break;
 171         case IB_WR_ATOMIC_FETCH_AND_ADD:
 172         case IB_WR_ATOMIC_CMP_AND_SWP:
 173                 if (send->s_op) {
 174                         rm = container_of(send->s_op, struct rds_message, atomic);
 175                         rds_ib_send_unmap_atomic(ic, send->s_op, wc_status);
 176                 }
 177                 break;
 178         default:
 179                 printk_ratelimited(KERN_NOTICE
 180                                "RDS/IB: %s: unexpected opcode 0x%x in WR!\n",
 181                                __func__, send->s_wr.opcode);
 182                 break;
 183         }
 184
 185         send->s_wr.opcode = 0xdead;
 186
 187         return rm;
 188 }
 189
 190 void rds_ib_send_init_ring(struct rds_ib_connection *ic)
 191 {
 192         struct rds_ib_send_work *send;
 193         u32 i;
 194
 195         for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
 196                 struct ib_sge *sge;
 197
 198                 send->s_op = NULL;
 199
 200                 send->s_wr.wr_id = i;
 201                 send->s_wr.sg_list = send->s_sge;
 202                 send->s_wr.ex.imm_data = 0;
 203
 204                 sge = &send->s_sge[0];
 205                 sge->addr = ic->i_send_hdrs_dma[i];
 206
 207                 sge->length = sizeof(struct rds_header);
 208                 sge->lkey = ic->i_pd->local_dma_lkey;
 209
 210                 send->s_sge[1].lkey = ic->i_pd->local_dma_lkey;
 211         }
 212 }
 213
 214 void rds_ib_send_clear_ring(struct rds_ib_connection *ic)
 215 {
 216         struct rds_ib_send_work *send;
 217         u32 i;
 218
 219         for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
 220                 if (send->s_op && send->s_wr.opcode != 0xdead)
 221                         rds_ib_send_unmap_op(ic, send, IB_WC_WR_FLUSH_ERR);
 222         }
 223 }
 224
 225 /*
 226  * The only fast path caller always has a non-zero nr, so we don't
 227  * bother testing nr before performing the atomic sub.
 228  */
 229 static void rds_ib_sub_signaled(struct rds_ib_connection *ic, int nr)
 230 {
 231         if ((atomic_sub_return(nr, &ic->i_signaled_sends) == 0) &&
 232             waitqueue_active(&rds_ib_ring_empty_wait))
 233                 wake_up(&rds_ib_ring_empty_wait);
 234         BUG_ON(atomic_read(&ic->i_signaled_sends) < 0);
 235 }
 236
 237 /*
 238  * The _oldest/_free ring operations here race cleanly with the alloc/unalloc
 239  * operations performed in the send path.  As the sender allocs and potentially
 240  * unallocs the next free entry in the ring it doesn't alter which is
 241  * the next to be freed, which is what this is concerned with.
 242  */
 243 void rds_ib_send_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc)
 244 {
 245         struct rds_message *rm = NULL;
 246         struct rds_connection *conn = ic->conn;
 247         struct rds_ib_send_work *send;
 248         u32 completed;
 249         u32 oldest;
 250         u32 i = 0;
 251         int nr_sig = 0;
 252
 253
 254         rdsdebug("wc wr_id 0x%llx status %u (%s) byte_len %u imm_data %u\n",
 255                  (unsigned long long)wc->wr_id, wc->status,
 256                  ib_wc_status_msg(wc->status), wc->byte_len,
 257                  be32_to_cpu(wc->ex.imm_data));
 258         rds_ib_stats_inc(s_ib_tx_cq_event);
 259
 260         if (wc->wr_id == RDS_IB_ACK_WR_ID) {
 261                 if (time_after(jiffies, ic->i_ack_queued + HZ / 2))
 262                         rds_ib_stats_inc(s_ib_tx_stalled);
 263                 rds_ib_ack_send_complete(ic);
 264                 return;
 265         }
 266
 267         oldest = rds_ib_ring_oldest(&ic->i_send_ring);
 268
 269         completed = rds_ib_ring_completed(&ic->i_send_ring, wc->wr_id, oldest);
 270
 271         for (i = 0; i < completed; i++) {
 272                 send = &ic->i_sends[oldest];
 273                 if (send->s_wr.send_flags & IB_SEND_SIGNALED)
 274                         nr_sig++;
 275
 276                 rm = rds_ib_send_unmap_op(ic, send, wc->status);
 277
 278                 if (time_after(jiffies, send->s_queued + HZ / 2))
 279                         rds_ib_stats_inc(s_ib_tx_stalled);
 280
 281                 if (send->s_op) {
 282                         if (send->s_op == rm->m_final_op) {
 283                                 /* If anyone waited for this message to get
 284                                  * flushed out, wake them up now
 285                                  */
 286                                 rds_message_unmapped(rm);
 287                         }
 288                         rds_message_put(rm);
 289                         send->s_op = NULL;
 290                 }
 291
 292                 oldest = (oldest + 1) % ic->i_send_ring.w_nr;
 293         }
 294
 295         rds_ib_ring_free(&ic->i_send_ring, completed);
 296         rds_ib_sub_signaled(ic, nr_sig);
 297         nr_sig = 0;
 298
 299         if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
 300             test_bit(0, &conn->c_map_queued))
 301                 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
 302
 303         /* We expect errors as the qp is drained during shutdown */
 304         if (wc->status != IB_WC_SUCCESS && rds_conn_up(conn)) {
 305                 rds_ib_conn_error(conn, "send completion on <%pI6c,%pI6c,%d> had status %u (%s), vendor err 0x%x, disconnecting and reconnecting\n",
 306                                   &conn->c_laddr, &conn->c_faddr,
 307                                   conn->c_tos, wc->status,
 308                                   ib_wc_status_msg(wc->status), wc->vendor_err);
 309         }
 310 }
 311
 312 /*
 313  * This is the main function for allocating credits when sending
 314  * messages.
 315  *
 316  * Conceptually, we have two counters:
 317  *  -   send credits: this tells us how many WRs we're allowed
 318  *      to submit without overruning the receiver's queue. For
 319  *      each SEND WR we post, we decrement this by one.
 320  *
 321  *  -   posted credits: this tells us how many WRs we recently
 322  *      posted to the receive queue. This value is transferred
 323  *      to the peer as a "credit update" in a RDS header field.
 324  *      Every time we transmit credits to the peer, we subtract
 325  *      the amount of transferred credits from this counter.
 326  *
 327  * It is essential that we avoid situations where both sides have
 328  * exhausted their send credits, and are unable to send new credits
 329  * to the peer. We achieve this by requiring that we send at least
 330  * one credit update to the peer before exhausting our credits.
 331  * When new credits arrive, we subtract one credit that is withheld
 332  * until we've posted new buffers and are ready to transmit these
 333  * credits (see rds_ib_send_add_credits below).
 334  *
 335  * The RDS send code is essentially single-threaded; rds_send_xmit
 336  * sets RDS_IN_XMIT to ensure exclusive access to the send ring.
 337  * However, the ACK sending code is independent and can race with
 338  * message SENDs.
 339  *
 340  * In the send path, we need to update the counters for send credits
 341  * and the counter of posted buffers atomically - when we use the
 342  * last available credit, we cannot allow another thread to race us
 343  * and grab the posted credits counter.  Hence, we have to use a
 344  * spinlock to protect the credit counter, or use atomics.
 345  *
 346  * Spinlocks shared between the send and the receive path are bad,
 347  * because they create unnecessary delays. An early implementation
 348  * using a spinlock showed a 5% degradation in throughput at some
 349  * loads.
 350  *
 351  * This implementation avoids spinlocks completely, putting both
 352  * counters into a single atomic, and updating that atomic using
 353  * atomic_add (in the receive path, when receiving fresh credits),
 354  * and using atomic_cmpxchg when updating the two counters.
 355  */
 356 int rds_ib_send_grab_credits(struct rds_ib_connection *ic,
 357                              u32 wanted, u32 *adv_credits, int need_posted, int max_posted)
 358 {
 359         unsigned int avail, posted, got = 0, advertise;
 360         long oldval, newval;
 361
 362         *adv_credits = 0;
 363         if (!ic->i_flowctl)
 364                 return wanted;
 365
 366 try_again:
 367         advertise = 0;
 368         oldval = newval = atomic_read(&ic->i_credits);
 369         posted = IB_GET_POST_CREDITS(oldval);
 370         avail = IB_GET_SEND_CREDITS(oldval);
 371
 372         rdsdebug("wanted=%u credits=%u posted=%u\n",
 373                         wanted, avail, posted);
 374
 375         /* The last credit must be used to send a credit update. */
 376         if (avail && !posted)
 377                 avail--;
 378
 379         if (avail < wanted) {
 380                 struct rds_connection *conn = ic->i_cm_id->context;
 381
 382                 /* Oops, there aren't that many credits left! */
 383                 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 384                 got = avail;
 385         } else {
 386                 /* Sometimes you get what you want, lalala. */
 387                 got = wanted;
 388         }
 389         newval -= IB_SET_SEND_CREDITS(got);
 390
 391         /*
 392          * If need_posted is non-zero, then the caller wants
 393          * the posted regardless of whether any send credits are
 394          * available.
 395          */
 396         if (posted && (got || need_posted)) {
 397                 advertise = min_t(unsigned int, posted, max_posted);
 398                 newval -= IB_SET_POST_CREDITS(advertise);
 399         }
 400
 401         /* Finally bill everything */
 402         if (atomic_cmpxchg(&ic->i_credits, oldval, newval) != oldval)
 403                 goto try_again;
 404
 405         *adv_credits = advertise;
 406         return got;
 407 }
 408
 409 void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits)
 410 {
 411         struct rds_ib_connection *ic = conn->c_transport_data;
 412
 413         if (credits == 0)
 414                 return;
 415
 416         rdsdebug("credits=%u current=%u%s\n",
 417                         credits,
 418                         IB_GET_SEND_CREDITS(atomic_read(&ic->i_credits)),
 419                         test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ? ", ll_send_full" : "");
 420
 421         atomic_add(IB_SET_SEND_CREDITS(credits), &ic->i_credits);
 422         if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags))
 423                 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
 424
 425         WARN_ON(IB_GET_SEND_CREDITS(credits) >= 16384);
 426
 427         rds_ib_stats_inc(s_ib_rx_credit_updates);
 428 }
 429
 430 void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted)
 431 {
 432         struct rds_ib_connection *ic = conn->c_transport_data;
 433
 434         if (posted == 0)
 435                 return;
 436
 437         atomic_add(IB_SET_POST_CREDITS(posted), &ic->i_credits);
 438
 439         /* Decide whether to send an update to the peer now.
 440          * If we would send a credit update for every single buffer we
 441          * post, we would end up with an ACK storm (ACK arrives,
 442          * consumes buffer, we refill the ring, send ACK to remote
 443          * advertising the newly posted buffer... ad inf)
 444          *
 445          * Performance pretty much depends on how often we send
 446          * credit updates - too frequent updates mean lots of ACKs.
 447          * Too infrequent updates, and the peer will run out of
 448          * credits and has to throttle.
 449          * For the time being, 16 seems to be a good compromise.
 450          */
 451         if (IB_GET_POST_CREDITS(atomic_read(&ic->i_credits)) >= 16)
 452                 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
 453 }
 454
 455 static inline int rds_ib_set_wr_signal_state(struct rds_ib_connection *ic,
 456                                              struct rds_ib_send_work *send,
 457                                              bool notify)
 458 {
 459         /*
 460          * We want to delay signaling completions just enough to get
 461          * the batching benefits but not so much that we create dead time
 462          * on the wire.
 463          */
 464         if (ic->i_unsignaled_wrs-- == 0 || notify) {
 465                 ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
 466                 send->s_wr.send_flags |= IB_SEND_SIGNALED;
 467                 return 1;
 468         }
 469         return 0;
 470 }
 471
 472 /*
 473  * This can be called multiple times for a given message.  The first time
 474  * we see a message we map its scatterlist into the IB device so that
 475  * we can provide that mapped address to the IB scatter gather entries
 476  * in the IB work requests.  We translate the scatterlist into a series
 477  * of work requests that fragment the message.  These work requests complete
 478  * in order so we pass ownership of the message to the completion handler
 479  * once we send the final fragment.
 480  *
 481  * The RDS core uses the c_send_lock to only enter this function once
 482  * per connection.  This makes sure that the tx ring alloc/unalloc pairs
 483  * don't get out of sync and confuse the ring.
 484  */
 485 int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
 486                 unsigned int hdr_off, unsigned int sg, unsigned int off)
 487 {
 488         struct rds_ib_connection *ic = conn->c_transport_data;
 489         struct ib_device *dev = ic->i_cm_id->device;
 490         struct rds_ib_send_work *send = NULL;
 491         struct rds_ib_send_work *first;
 492         struct rds_ib_send_work *prev;
 493         const struct ib_send_wr *failed_wr;
 494         struct scatterlist *scat;
 495         u32 pos;
 496         u32 i;
 497         u32 work_alloc;
 498         u32 credit_alloc = 0;
 499         u32 posted;
 500         u32 adv_credits = 0;
 501         int send_flags = 0;
 502         int bytes_sent = 0;
 503         int ret;
 504         int flow_controlled = 0;
 505         int nr_sig = 0;
 506
 507         BUG_ON(off % RDS_FRAG_SIZE);
 508         BUG_ON(hdr_off != 0 && hdr_off != sizeof(struct rds_header));
 509
 510         /* Do not send cong updates to IB loopback */
 511         if (conn->c_loopback
 512             && rm->m_inc.i_hdr.h_flags & RDS_FLAG_CONG_BITMAP) {
 513                 rds_cong_map_updated(conn->c_fcong, ~(u64) 0);
 514                 scat = &rm->data.op_sg[sg];
 515                 ret = max_t(int, RDS_CONG_MAP_BYTES, scat->length);
 516                 return sizeof(struct rds_header) + ret;
 517         }
 518
 519         /* FIXME we may overallocate here */
 520         if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0)
 521                 i = 1;
 522         else
 523                 i = DIV_ROUND_UP(be32_to_cpu(rm->m_inc.i_hdr.h_len), RDS_FRAG_SIZE);
 524
 525         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
 526         if (work_alloc == 0) {
 527                 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 528                 rds_ib_stats_inc(s_ib_tx_ring_full);
 529                 ret = -ENOMEM;
 530                 goto out;
 531         }
 532
 533         if (ic->i_flowctl) {
 534                 credit_alloc = rds_ib_send_grab_credits(ic, work_alloc, &posted, 0, RDS_MAX_ADV_CREDIT);
 535                 adv_credits += posted;
 536                 if (credit_alloc < work_alloc) {
 537                         rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - credit_alloc);
 538                         work_alloc = credit_alloc;
 539                         flow_controlled = 1;
 540                 }
 541                 if (work_alloc == 0) {
 542                         set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 543                         rds_ib_stats_inc(s_ib_tx_throttle);
 544                         ret = -ENOMEM;
 545                         goto out;
 546                 }
 547         }
 548
 549         /* map the message the first time we see it */
 550         if (!ic->i_data_op) {
 551                 if (rm->data.op_nents) {
 552                         rm->data.op_count = ib_dma_map_sg(dev,
 553                                                           rm->data.op_sg,
 554                                                           rm->data.op_nents,
 555                                                           DMA_TO_DEVICE);
 556                         rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->data.op_count);
 557                         if (rm->data.op_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.op_count = 0;
 565                 }
 566
 567                 rds_message_addref(rm);
 568                 rm->data.op_dmasg = 0;
 569                 rm->data.op_dmaoff = 0;
 570                 ic->i_data_op = &rm->data;
 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.op_active) {
 581                         struct rds_ext_header_rdma ext_hdr;
 582
 583                         ext_hdr.h_rdma_rkey = cpu_to_be32(rm->rdma.op_rkey);
 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                 if (ic->i_flowctl) {
 604                         rds_ib_send_grab_credits(ic, 0, &posted, 1, RDS_MAX_ADV_CREDIT - adv_credits);
 605                         adv_credits += posted;
 606                         BUG_ON(adv_credits > 255);
 607                 }
 608         }
 609
 610         /* Sometimes you want to put a fence between an RDMA
 611          * READ and the following SEND.
 612          * We could either do this all the time
 613          * or when requested by the user. Right now, we let
 614          * the application choose.
 615          */
 616         if (rm->rdma.op_active && rm->rdma.op_fence)
 617                 send_flags = IB_SEND_FENCE;
 618
 619         /* Each frag gets a header. Msgs may be 0 bytes */
 620         send = &ic->i_sends[pos];
 621         first = send;
 622         prev = NULL;
 623         scat = &ic->i_data_op->op_sg[rm->data.op_dmasg];
 624         i = 0;
 625         do {
 626                 unsigned int len = 0;
 627
 628                 /* Set up the header */
 629                 send->s_wr.send_flags = send_flags;
 630                 send->s_wr.opcode = IB_WR_SEND;
 631                 send->s_wr.num_sge = 1;
 632                 send->s_wr.next = NULL;
 633                 send->s_queued = jiffies;
 634                 send->s_op = NULL;
 635
 636                 send->s_sge[0].addr = ic->i_send_hdrs_dma[pos];
 637
 638                 send->s_sge[0].length = sizeof(struct rds_header);
 639                 send->s_sge[0].lkey = ic->i_pd->local_dma_lkey;
 640
 641                 ib_dma_sync_single_for_cpu(ic->rds_ibdev->dev,
 642                                            ic->i_send_hdrs_dma[pos],
 643                                            sizeof(struct rds_header),
 644                                            DMA_TO_DEVICE);
 645                 memcpy(ic->i_send_hdrs[pos], &rm->m_inc.i_hdr,
 646                        sizeof(struct rds_header));
 647
 648
 649                 /* Set up the data, if present */
 650                 if (i < work_alloc
 651                     && scat != &rm->data.op_sg[rm->data.op_count]) {
 652                         len = min(RDS_FRAG_SIZE,
 653                                   sg_dma_len(scat) - rm->data.op_dmaoff);
 654                         send->s_wr.num_sge = 2;
 655
 656                         send->s_sge[1].addr = sg_dma_address(scat);
 657                         send->s_sge[1].addr += rm->data.op_dmaoff;
 658                         send->s_sge[1].length = len;
 659                         send->s_sge[1].lkey = ic->i_pd->local_dma_lkey;
 660
 661                         bytes_sent += len;
 662                         rm->data.op_dmaoff += len;
 663                         if (rm->data.op_dmaoff == sg_dma_len(scat)) {
 664                                 scat++;
 665                                 rm->data.op_dmasg++;
 666                                 rm->data.op_dmaoff = 0;
 667                         }
 668                 }
 669
 670                 rds_ib_set_wr_signal_state(ic, send, false);
 671
 672                 /*
 673                  * Always signal the last one if we're stopping due to flow control.
 674                  */
 675                 if (ic->i_flowctl && flow_controlled && i == (work_alloc - 1)) {
 676                         rds_ib_set_wr_signal_state(ic, send, true);
 677                         send->s_wr.send_flags |= IB_SEND_SOLICITED;
 678                 }
 679
 680                 if (send->s_wr.send_flags & IB_SEND_SIGNALED)
 681                         nr_sig++;
 682
 683                 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
 684                          &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
 685
 686                 if (ic->i_flowctl && adv_credits) {
 687                         struct rds_header *hdr = ic->i_send_hdrs[pos];
 688
 689                         /* add credit and redo the header checksum */
 690                         hdr->h_credit = adv_credits;
 691                         rds_message_make_checksum(hdr);
 692                         adv_credits = 0;
 693                         rds_ib_stats_inc(s_ib_tx_credit_updates);
 694                 }
 695                 ib_dma_sync_single_for_device(ic->rds_ibdev->dev,
 696                                               ic->i_send_hdrs_dma[pos],
 697                                               sizeof(struct rds_header),
 698                                               DMA_TO_DEVICE);
 699
 700                 if (prev)
 701                         prev->s_wr.next = &send->s_wr;
 702                 prev = send;
 703
 704                 pos = (pos + 1) % ic->i_send_ring.w_nr;
 705                 send = &ic->i_sends[pos];
 706                 i++;
 707
 708         } while (i < work_alloc
 709                  && scat != &rm->data.op_sg[rm->data.op_count]);
 710
 711         /* Account the RDS header in the number of bytes we sent, but just once.
 712          * The caller has no concept of fragmentation. */
 713         if (hdr_off == 0)
 714                 bytes_sent += sizeof(struct rds_header);
 715
 716         /* if we finished the message then send completion owns it */
 717         if (scat == &rm->data.op_sg[rm->data.op_count]) {
 718                 prev->s_op = ic->i_data_op;
 719                 prev->s_wr.send_flags |= IB_SEND_SOLICITED;
 720                 if (!(prev->s_wr.send_flags & IB_SEND_SIGNALED))
 721                         nr_sig += rds_ib_set_wr_signal_state(ic, prev, true);
 722                 ic->i_data_op = NULL;
 723         }
 724
 725         /* Put back wrs & credits we didn't use */
 726         if (i < work_alloc) {
 727                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
 728                 work_alloc = i;
 729         }
 730         if (ic->i_flowctl && i < credit_alloc)
 731                 rds_ib_send_add_credits(conn, credit_alloc - i);
 732
 733         if (nr_sig)
 734                 atomic_add(nr_sig, &ic->i_signaled_sends);
 735
 736         /* XXX need to worry about failed_wr and partial sends. */
 737         failed_wr = &first->s_wr;
 738         ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
 739         rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
 740                  first, &first->s_wr, ret, failed_wr);
 741         BUG_ON(failed_wr != &first->s_wr);
 742         if (ret) {
 743                 printk(KERN_WARNING "RDS/IB: ib_post_send to %pI6c "
 744                        "returned %d\n", &conn->c_faddr, ret);
 745                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 746                 rds_ib_sub_signaled(ic, nr_sig);
 747                 if (prev->s_op) {
 748                         ic->i_data_op = prev->s_op;
 749                         prev->s_op = NULL;
 750                 }
 751
 752                 rds_ib_conn_error(ic->conn, "ib_post_send failed\n");
 753                 goto out;
 754         }
 755
 756         ret = bytes_sent;
 757 out:
 758         BUG_ON(adv_credits);
 759         return ret;
 760 }
 761
 762 /*
 763  * Issue atomic operation.
 764  * A simplified version of the rdma case, we always map 1 SG, and
 765  * only 8 bytes, for the return value from the atomic operation.
 766  */
 767 int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op)
 768 {
 769         struct rds_ib_connection *ic = conn->c_transport_data;
 770         struct rds_ib_send_work *send = NULL;
 771         const struct ib_send_wr *failed_wr;
 772         u32 pos;
 773         u32 work_alloc;
 774         int ret;
 775         int nr_sig = 0;
 776
 777         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, 1, &pos);
 778         if (work_alloc != 1) {
 779                 rds_ib_stats_inc(s_ib_tx_ring_full);
 780                 ret = -ENOMEM;
 781                 goto out;
 782         }
 783
 784         /* address of send request in ring */
 785         send = &ic->i_sends[pos];
 786         send->s_queued = jiffies;
 787
 788         if (op->op_type == RDS_ATOMIC_TYPE_CSWP) {
 789                 send->s_atomic_wr.wr.opcode = IB_WR_MASKED_ATOMIC_CMP_AND_SWP;
 790                 send->s_atomic_wr.compare_add = op->op_m_cswp.compare;
 791                 send->s_atomic_wr.swap = op->op_m_cswp.swap;
 792                 send->s_atomic_wr.compare_add_mask = op->op_m_cswp.compare_mask;
 793                 send->s_atomic_wr.swap_mask = op->op_m_cswp.swap_mask;
 794         } else { /* FADD */
 795                 send->s_atomic_wr.wr.opcode = IB_WR_MASKED_ATOMIC_FETCH_AND_ADD;
 796                 send->s_atomic_wr.compare_add = op->op_m_fadd.add;
 797                 send->s_atomic_wr.swap = 0;
 798                 send->s_atomic_wr.compare_add_mask = op->op_m_fadd.nocarry_mask;
 799                 send->s_atomic_wr.swap_mask = 0;
 800         }
 801         send->s_wr.send_flags = 0;
 802         nr_sig = rds_ib_set_wr_signal_state(ic, send, op->op_notify);
 803         send->s_atomic_wr.wr.num_sge = 1;
 804         send->s_atomic_wr.wr.next = NULL;
 805         send->s_atomic_wr.remote_addr = op->op_remote_addr;
 806         send->s_atomic_wr.rkey = op->op_rkey;
 807         send->s_op = op;
 808         rds_message_addref(container_of(send->s_op, struct rds_message, atomic));
 809
 810         /* map 8 byte retval buffer to the device */
 811         ret = ib_dma_map_sg(ic->i_cm_id->device, op->op_sg, 1, DMA_FROM_DEVICE);
 812         rdsdebug("ic %p mapping atomic op %p. mapped %d pg\n", ic, op, ret);
 813         if (ret != 1) {
 814                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 815                 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
 816                 ret = -ENOMEM; /* XXX ? */
 817                 goto out;
 818         }
 819
 820         /* Convert our struct scatterlist to struct ib_sge */
 821         send->s_sge[0].addr = sg_dma_address(op->op_sg);
 822         send->s_sge[0].length = sg_dma_len(op->op_sg);
 823         send->s_sge[0].lkey = ic->i_pd->local_dma_lkey;
 824
 825         rdsdebug("rva %Lx rpa %Lx len %u\n", op->op_remote_addr,
 826                  send->s_sge[0].addr, send->s_sge[0].length);
 827
 828         if (nr_sig)
 829                 atomic_add(nr_sig, &ic->i_signaled_sends);
 830
 831         failed_wr = &send->s_atomic_wr.wr;
 832         ret = ib_post_send(ic->i_cm_id->qp, &send->s_atomic_wr.wr, &failed_wr);
 833         rdsdebug("ic %p send %p (wr %p) ret %d wr %p\n", ic,
 834                  send, &send->s_atomic_wr, ret, failed_wr);
 835         BUG_ON(failed_wr != &send->s_atomic_wr.wr);
 836         if (ret) {
 837                 printk(KERN_WARNING "RDS/IB: atomic ib_post_send to %pI6c "
 838                        "returned %d\n", &conn->c_faddr, ret);
 839                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 840                 rds_ib_sub_signaled(ic, nr_sig);
 841                 goto out;
 842         }
 843
 844         if (unlikely(failed_wr != &send->s_atomic_wr.wr)) {
 845                 printk(KERN_WARNING "RDS/IB: atomic ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
 846                 BUG_ON(failed_wr != &send->s_atomic_wr.wr);
 847         }
 848
 849 out:
 850         return ret;
 851 }
 852
 853 int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op)
 854 {
 855         struct rds_ib_connection *ic = conn->c_transport_data;
 856         struct rds_ib_send_work *send = NULL;
 857         struct rds_ib_send_work *first;
 858         struct rds_ib_send_work *prev;
 859         const struct ib_send_wr *failed_wr;
 860         struct scatterlist *scat;
 861         unsigned long len;
 862         u64 remote_addr = op->op_remote_addr;
 863         u32 max_sge = ic->rds_ibdev->max_sge;
 864         u32 pos;
 865         u32 work_alloc;
 866         u32 i;
 867         u32 j;
 868         int sent;
 869         int ret;
 870         int num_sge;
 871         int nr_sig = 0;
 872         u64 odp_addr = op->op_odp_addr;
 873         u32 odp_lkey = 0;
 874
 875         /* map the op the first time we see it */
 876         if (!op->op_odp_mr) {
 877                 if (!op->op_mapped) {
 878                         op->op_count =
 879                                 ib_dma_map_sg(ic->i_cm_id->device, op->op_sg,
 880                                               op->op_nents,
 881                                               (op->op_write) ? DMA_TO_DEVICE :
 882                                                                DMA_FROM_DEVICE);
 883                         rdsdebug("ic %p mapping op %p: %d\n", ic, op,
 884                                  op->op_count);
 885                         if (op->op_count == 0) {
 886                                 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
 887                                 ret = -ENOMEM; /* XXX ? */
 888                                 goto out;
 889                         }
 890                         op->op_mapped = 1;
 891                 }
 892         } else {
 893                 op->op_count = op->op_nents;
 894                 odp_lkey = rds_ib_get_lkey(op->op_odp_mr->r_trans_private);
 895         }
 896
 897         /*
 898          * Instead of knowing how to return a partial rdma read/write we insist that there
 899          * be enough work requests to send the entire message.
 900          */
 901         i = DIV_ROUND_UP(op->op_count, max_sge);
 902
 903         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
 904         if (work_alloc != i) {
 905                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 906                 rds_ib_stats_inc(s_ib_tx_ring_full);
 907                 ret = -ENOMEM;
 908                 goto out;
 909         }
 910
 911         send = &ic->i_sends[pos];
 912         first = send;
 913         prev = NULL;
 914         scat = &op->op_sg[0];
 915         sent = 0;
 916         num_sge = op->op_count;
 917
 918         for (i = 0; i < work_alloc && scat != &op->op_sg[op->op_count]; i++) {
 919                 send->s_wr.send_flags = 0;
 920                 send->s_queued = jiffies;
 921                 send->s_op = NULL;
 922
 923                 if (!op->op_notify)
 924                         nr_sig += rds_ib_set_wr_signal_state(ic, send,
 925                                                              op->op_notify);
 926
 927                 send->s_wr.opcode = op->op_write ? IB_WR_RDMA_WRITE : IB_WR_RDMA_READ;
 928                 send->s_rdma_wr.remote_addr = remote_addr;
 929                 send->s_rdma_wr.rkey = op->op_rkey;
 930
 931                 if (num_sge > max_sge) {
 932                         send->s_rdma_wr.wr.num_sge = max_sge;
 933                         num_sge -= max_sge;
 934                 } else {
 935                         send->s_rdma_wr.wr.num_sge = num_sge;
 936                 }
 937
 938                 send->s_rdma_wr.wr.next = NULL;
 939
 940                 if (prev)
 941                         prev->s_rdma_wr.wr.next = &send->s_rdma_wr.wr;
 942
 943                 for (j = 0; j < send->s_rdma_wr.wr.num_sge &&
 944                      scat != &op->op_sg[op->op_count]; j++) {
 945                         len = sg_dma_len(scat);
 946                         if (!op->op_odp_mr) {
 947                                 send->s_sge[j].addr = sg_dma_address(scat);
 948                                 send->s_sge[j].lkey = ic->i_pd->local_dma_lkey;
 949                         } else {
 950                                 send->s_sge[j].addr = odp_addr;
 951                                 send->s_sge[j].lkey = odp_lkey;
 952                         }
 953                         send->s_sge[j].length = len;
 954
 955                         sent += len;
 956                         rdsdebug("ic %p sent %d remote_addr %llu\n", ic, sent, remote_addr);
 957
 958                         remote_addr += len;
 959                         odp_addr += len;
 960                         scat++;
 961                 }
 962
 963                 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
 964                         &send->s_rdma_wr.wr,
 965                         send->s_rdma_wr.wr.num_sge,
 966                         send->s_rdma_wr.wr.next);
 967
 968                 prev = send;
 969                 if (++send == &ic->i_sends[ic->i_send_ring.w_nr])
 970                         send = ic->i_sends;
 971         }
 972
 973         /* give a reference to the last op */
 974         if (scat == &op->op_sg[op->op_count]) {
 975                 prev->s_op = op;
 976                 rds_message_addref(container_of(op, struct rds_message, rdma));
 977         }
 978
 979         if (i < work_alloc) {
 980                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
 981                 work_alloc = i;
 982         }
 983
 984         if (nr_sig)
 985                 atomic_add(nr_sig, &ic->i_signaled_sends);
 986
 987         failed_wr = &first->s_rdma_wr.wr;
 988         ret = ib_post_send(ic->i_cm_id->qp, &first->s_rdma_wr.wr, &failed_wr);
 989         rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
 990                  first, &first->s_rdma_wr.wr, ret, failed_wr);
 991         BUG_ON(failed_wr != &first->s_rdma_wr.wr);
 992         if (ret) {
 993                 printk(KERN_WARNING "RDS/IB: rdma ib_post_send to %pI6c "
 994                        "returned %d\n", &conn->c_faddr, ret);
 995                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 996                 rds_ib_sub_signaled(ic, nr_sig);
 997                 goto out;
 998         }
 999
1000         if (unlikely(failed_wr != &first->s_rdma_wr.wr)) {
1001                 printk(KERN_WARNING "RDS/IB: ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
1002                 BUG_ON(failed_wr != &first->s_rdma_wr.wr);
1003         }
1004
1005
1006 out:
1007         return ret;
1008 }
1009
1010 void rds_ib_xmit_path_complete(struct rds_conn_path *cp)
1011 {
1012         struct rds_connection *conn = cp->cp_conn;
1013         struct rds_ib_connection *ic = conn->c_transport_data;
1014
1015         /* We may have a pending ACK or window update we were unable
1016          * to send previously (due to flow control). Try again. */
1017         rds_ib_attempt_ack(ic);
1018 }