net/rds/ib_send.c

   1 /*
   2  * Copyright (c) 2006, 2017 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
  43 /*
  44  * Convert IB-specific error message to RDS error message and call core
  45  * completion handler.
  46  */
  47 static void rds_ib_send_complete(struct rds_message *rm,
  48                                  int wc_status,
  49                                  void (*complete)(struct rds_message *rm, int status))
  50 {
  51         int notify_status;
  52
  53         switch (wc_status) {
  54         case IB_WC_WR_FLUSH_ERR:
  55                 return;
  56
  57         case IB_WC_SUCCESS:
  58                 notify_status = RDS_RDMA_SUCCESS;
  59                 break;
  60
  61         case IB_WC_REM_ACCESS_ERR:
  62                 notify_status = RDS_RDMA_REMOTE_ERROR;
  63                 break;
  64
  65         default:
  66                 notify_status = RDS_RDMA_OTHER_ERROR;
  67                 break;
  68         }
  69         complete(rm, notify_status);
  70 }
  71
  72 static void rds_ib_send_unmap_data(struct rds_ib_connection *ic,
  73                                    struct rm_data_op *op,
  74                                    int wc_status)
  75 {
  76         if (op->op_nents)
  77                 ib_dma_unmap_sg(ic->i_cm_id->device,
  78                                 op->op_sg, op->op_nents,
  79                                 DMA_TO_DEVICE);
  80 }
  81
  82 static void rds_ib_send_unmap_rdma(struct rds_ib_connection *ic,
  83                                    struct rm_rdma_op *op,
  84                                    int wc_status)
  85 {
  86         if (op->op_mapped) {
  87                 ib_dma_unmap_sg(ic->i_cm_id->device,
  88                                 op->op_sg, op->op_nents,
  89                                 op->op_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
  90                 op->op_mapped = 0;
  91         }
  92
  93         /* If the user asked for a completion notification on this
  94          * message, we can implement three different semantics:
  95          *  1.  Notify when we received the ACK on the RDS message
  96          *      that was queued with the RDMA. This provides reliable
  97          *      notification of RDMA status at the expense of a one-way
  98          *      packet delay.
  99          *  2.  Notify when the IB stack gives us the completion event for
 100          *      the RDMA operation.
 101          *  3.  Notify when the IB stack gives us the completion event for
 102          *      the accompanying RDS messages.
 103          * Here, we implement approach #3. To implement approach #2,
 104          * we would need to take an event for the rdma WR. To implement #1,
 105          * don't call rds_rdma_send_complete at all, and fall back to the notify
 106          * handling in the ACK processing code.
 107          *
 108          * Note: There's no need to explicitly sync any RDMA buffers using
 109          * ib_dma_sync_sg_for_cpu - the completion for the RDMA
 110          * operation itself unmapped the RDMA buffers, which takes care
 111          * of synching.
 112          */
 113         rds_ib_send_complete(container_of(op, struct rds_message, rdma),
 114                              wc_status, rds_rdma_send_complete);
 115
 116         if (op->op_write)
 117                 rds_stats_add(s_send_rdma_bytes, op->op_bytes);
 118         else
 119                 rds_stats_add(s_recv_rdma_bytes, op->op_bytes);
 120 }
 121
 122 static void rds_ib_send_unmap_atomic(struct rds_ib_connection *ic,
 123                                      struct rm_atomic_op *op,
 124                                      int wc_status)
 125 {
 126         /* unmap atomic recvbuf */
 127         if (op->op_mapped) {
 128                 ib_dma_unmap_sg(ic->i_cm_id->device, op->op_sg, 1,
 129                                 DMA_FROM_DEVICE);
 130                 op->op_mapped = 0;
 131         }
 132
 133         rds_ib_send_complete(container_of(op, struct rds_message, atomic),
 134                              wc_status, rds_atomic_send_complete);
 135
 136         if (op->op_type == RDS_ATOMIC_TYPE_CSWP)
 137                 rds_ib_stats_inc(s_ib_atomic_cswp);
 138         else
 139                 rds_ib_stats_inc(s_ib_atomic_fadd);
 140 }
 141
 142 /*
 143  * Unmap the resources associated with a struct send_work.
 144  *
 145  * Returns the rm for no good reason other than it is unobtainable
 146  * other than by switching on wr.opcode, currently, and the caller,
 147  * the event handler, needs it.
 148  */
 149 static struct rds_message *rds_ib_send_unmap_op(struct rds_ib_connection *ic,
 150                                                 struct rds_ib_send_work *send,
 151                                                 int wc_status)
 152 {
 153         struct rds_message *rm = NULL;
 154
 155         /* In the error case, wc.opcode sometimes contains garbage */
 156         switch (send->s_wr.opcode) {
 157         case IB_WR_SEND:
 158                 if (send->s_op) {
 159                         rm = container_of(send->s_op, struct rds_message, data);
 160                         rds_ib_send_unmap_data(ic, send->s_op, wc_status);
 161                 }
 162                 break;
 163         case IB_WR_RDMA_WRITE:
 164         case IB_WR_RDMA_READ:
 165                 if (send->s_op) {
 166                         rm = container_of(send->s_op, struct rds_message, rdma);
 167                         rds_ib_send_unmap_rdma(ic, send->s_op, wc_status);
 168                 }
 169                 break;
 170         case IB_WR_ATOMIC_FETCH_AND_ADD:
 171         case IB_WR_ATOMIC_CMP_AND_SWP:
 172                 if (send->s_op) {
 173                         rm = container_of(send->s_op, struct rds_message, atomic);
 174                         rds_ib_send_unmap_atomic(ic, send->s_op, wc_status);
 175                 }
 176                 break;
 177         default:
 178                 printk_ratelimited(KERN_NOTICE
 179                                "RDS/IB: %s: unexpected opcode 0x%x in WR!\n",
 180                                __func__, send->s_wr.opcode);
 181                 break;
 182         }
 183
 184         send->s_wr.opcode = 0xdead;
 185
 186         return rm;
 187 }
 188
 189 void rds_ib_send_init_ring(struct rds_ib_connection *ic)
 190 {
 191         struct rds_ib_send_work *send;
 192         u32 i;
 193
 194         for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
 195                 struct ib_sge *sge;
 196
 197                 send->s_op = NULL;
 198
 199                 send->s_wr.wr_id = i;
 200                 send->s_wr.sg_list = send->s_sge;
 201                 send->s_wr.ex.imm_data = 0;
 202
 203                 sge = &send->s_sge[0];
 204                 sge->addr = ic->i_send_hdrs_dma + (i * sizeof(struct rds_header));
 205                 sge->length = sizeof(struct rds_header);
 206                 sge->lkey = ic->i_pd->local_dma_lkey;
 207
 208                 send->s_sge[1].lkey = ic->i_pd->local_dma_lkey;
 209         }
 210 }
 211
 212 void rds_ib_send_clear_ring(struct rds_ib_connection *ic)
 213 {
 214         struct rds_ib_send_work *send;
 215         u32 i;
 216
 217         for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
 218                 if (send->s_op && send->s_wr.opcode != 0xdead)
 219                         rds_ib_send_unmap_op(ic, send, IB_WC_WR_FLUSH_ERR);
 220         }
 221 }
 222
 223 /*
 224  * The only fast path caller always has a non-zero nr, so we don't
 225  * bother testing nr before performing the atomic sub.
 226  */
 227 static void rds_ib_sub_signaled(struct rds_ib_connection *ic, int nr)
 228 {
 229         if ((atomic_sub_return(nr, &ic->i_signaled_sends) == 0) &&
 230             waitqueue_active(&rds_ib_ring_empty_wait))
 231                 wake_up(&rds_ib_ring_empty_wait);
 232         BUG_ON(atomic_read(&ic->i_signaled_sends) < 0);
 233 }
 234
 235 /*
 236  * The _oldest/_free ring operations here race cleanly with the alloc/unalloc
 237  * operations performed in the send path.  As the sender allocs and potentially
 238  * unallocs the next free entry in the ring it doesn't alter which is
 239  * the next to be freed, which is what this is concerned with.
 240  */
 241 void rds_ib_send_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc)
 242 {
 243         struct rds_message *rm = NULL;
 244         struct rds_connection *conn = ic->conn;
 245         struct rds_ib_send_work *send;
 246         u32 completed;
 247         u32 oldest;
 248         u32 i = 0;
 249         int nr_sig = 0;
 250
 251
 252         rdsdebug("wc wr_id 0x%llx status %u (%s) byte_len %u imm_data %u\n",
 253                  (unsigned long long)wc->wr_id, wc->status,
 254                  ib_wc_status_msg(wc->status), wc->byte_len,
 255                  be32_to_cpu(wc->ex.imm_data));
 256         rds_ib_stats_inc(s_ib_tx_cq_event);
 257
 258         if (wc->wr_id == RDS_IB_ACK_WR_ID) {
 259                 if (time_after(jiffies, ic->i_ack_queued + HZ / 2))
 260                         rds_ib_stats_inc(s_ib_tx_stalled);
 261                 rds_ib_ack_send_complete(ic);
 262                 return;
 263         }
 264
 265         oldest = rds_ib_ring_oldest(&ic->i_send_ring);
 266
 267         completed = rds_ib_ring_completed(&ic->i_send_ring, wc->wr_id, oldest);
 268
 269         for (i = 0; i < completed; i++) {
 270                 send = &ic->i_sends[oldest];
 271                 if (send->s_wr.send_flags & IB_SEND_SIGNALED)
 272                         nr_sig++;
 273
 274                 rm = rds_ib_send_unmap_op(ic, send, wc->status);
 275
 276                 if (time_after(jiffies, send->s_queued + HZ / 2))
 277                         rds_ib_stats_inc(s_ib_tx_stalled);
 278
 279                 if (send->s_op) {
 280                         if (send->s_op == rm->m_final_op) {
 281                                 /* If anyone waited for this message to get
 282                                  * flushed out, wake them up now
 283                                  */
 284                                 rds_message_unmapped(rm);
 285                         }
 286                         rds_message_put(rm);
 287                         send->s_op = NULL;
 288                 }
 289
 290                 oldest = (oldest + 1) % ic->i_send_ring.w_nr;
 291         }
 292
 293         rds_ib_ring_free(&ic->i_send_ring, completed);
 294         rds_ib_sub_signaled(ic, nr_sig);
 295         nr_sig = 0;
 296
 297         if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
 298             test_bit(0, &conn->c_map_queued))
 299                 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
 300
 301         /* We expect errors as the qp is drained during shutdown */
 302         if (wc->status != IB_WC_SUCCESS && rds_conn_up(conn)) {
 303                 rds_ib_conn_error(conn, "send completion on <%pI6c,%pI6c,%d> had status %u (%s), disconnecting and reconnecting\n",
 304                                   &conn->c_laddr, &conn->c_faddr,
 305                                   conn->c_tos, wc->status,
 306                                   ib_wc_status_msg(wc->status));
 307         }
 308 }
 309
 310 /*
 311  * This is the main function for allocating credits when sending
 312  * messages.
 313  *
 314  * Conceptually, we have two counters:
 315  *  -   send credits: this tells us how many WRs we're allowed
 316  *      to submit without overruning the receiver's queue. For
 317  *      each SEND WR we post, we decrement this by one.
 318  *
 319  *  -   posted credits: this tells us how many WRs we recently
 320  *      posted to the receive queue. This value is transferred
 321  *      to the peer as a "credit update" in a RDS header field.
 322  *      Every time we transmit credits to the peer, we subtract
 323  *      the amount of transferred credits from this counter.
 324  *
 325  * It is essential that we avoid situations where both sides have
 326  * exhausted their send credits, and are unable to send new credits
 327  * to the peer. We achieve this by requiring that we send at least
 328  * one credit update to the peer before exhausting our credits.
 329  * When new credits arrive, we subtract one credit that is withheld
 330  * until we've posted new buffers and are ready to transmit these
 331  * credits (see rds_ib_send_add_credits below).
 332  *
 333  * The RDS send code is essentially single-threaded; rds_send_xmit
 334  * sets RDS_IN_XMIT to ensure exclusive access to the send ring.
 335  * However, the ACK sending code is independent and can race with
 336  * message SENDs.
 337  *
 338  * In the send path, we need to update the counters for send credits
 339  * and the counter of posted buffers atomically - when we use the
 340  * last available credit, we cannot allow another thread to race us
 341  * and grab the posted credits counter.  Hence, we have to use a
 342  * spinlock to protect the credit counter, or use atomics.
 343  *
 344  * Spinlocks shared between the send and the receive path are bad,
 345  * because they create unnecessary delays. An early implementation
 346  * using a spinlock showed a 5% degradation in throughput at some
 347  * loads.
 348  *
 349  * This implementation avoids spinlocks completely, putting both
 350  * counters into a single atomic, and updating that atomic using
 351  * atomic_add (in the receive path, when receiving fresh credits),
 352  * and using atomic_cmpxchg when updating the two counters.
 353  */
 354 int rds_ib_send_grab_credits(struct rds_ib_connection *ic,
 355                              u32 wanted, u32 *adv_credits, int need_posted, int max_posted)
 356 {
 357         unsigned int avail, posted, got = 0, advertise;
 358         long oldval, newval;
 359
 360         *adv_credits = 0;
 361         if (!ic->i_flowctl)
 362                 return wanted;
 363
 364 try_again:
 365         advertise = 0;
 366         oldval = newval = atomic_read(&ic->i_credits);
 367         posted = IB_GET_POST_CREDITS(oldval);
 368         avail = IB_GET_SEND_CREDITS(oldval);
 369
 370         rdsdebug("wanted=%u credits=%u posted=%u\n",
 371                         wanted, avail, posted);
 372
 373         /* The last credit must be used to send a credit update. */
 374         if (avail && !posted)
 375                 avail--;
 376
 377         if (avail < wanted) {
 378                 struct rds_connection *conn = ic->i_cm_id->context;
 379
 380                 /* Oops, there aren't that many credits left! */
 381                 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 382                 got = avail;
 383         } else {
 384                 /* Sometimes you get what you want, lalala. */
 385                 got = wanted;
 386         }
 387         newval -= IB_SET_SEND_CREDITS(got);
 388
 389         /*
 390          * If need_posted is non-zero, then the caller wants
 391          * the posted regardless of whether any send credits are
 392          * available.
 393          */
 394         if (posted && (got || need_posted)) {
 395                 advertise = min_t(unsigned int, posted, max_posted);
 396                 newval -= IB_SET_POST_CREDITS(advertise);
 397         }
 398
 399         /* Finally bill everything */
 400         if (atomic_cmpxchg(&ic->i_credits, oldval, newval) != oldval)
 401                 goto try_again;
 402
 403         *adv_credits = advertise;
 404         return got;
 405 }
 406
 407 void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits)
 408 {
 409         struct rds_ib_connection *ic = conn->c_transport_data;
 410
 411         if (credits == 0)
 412                 return;
 413
 414         rdsdebug("credits=%u current=%u%s\n",
 415                         credits,
 416                         IB_GET_SEND_CREDITS(atomic_read(&ic->i_credits)),
 417                         test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ? ", ll_send_full" : "");
 418
 419         atomic_add(IB_SET_SEND_CREDITS(credits), &ic->i_credits);
 420         if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags))
 421                 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
 422
 423         WARN_ON(IB_GET_SEND_CREDITS(credits) >= 16384);
 424
 425         rds_ib_stats_inc(s_ib_rx_credit_updates);
 426 }
 427
 428 void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted)
 429 {
 430         struct rds_ib_connection *ic = conn->c_transport_data;
 431
 432         if (posted == 0)
 433                 return;
 434
 435         atomic_add(IB_SET_POST_CREDITS(posted), &ic->i_credits);
 436
 437         /* Decide whether to send an update to the peer now.
 438          * If we would send a credit update for every single buffer we
 439          * post, we would end up with an ACK storm (ACK arrives,
 440          * consumes buffer, we refill the ring, send ACK to remote
 441          * advertising the newly posted buffer... ad inf)
 442          *
 443          * Performance pretty much depends on how often we send
 444          * credit updates - too frequent updates mean lots of ACKs.
 445          * Too infrequent updates, and the peer will run out of
 446          * credits and has to throttle.
 447          * For the time being, 16 seems to be a good compromise.
 448          */
 449         if (IB_GET_POST_CREDITS(atomic_read(&ic->i_credits)) >= 16)
 450                 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
 451 }
 452
 453 static inline int rds_ib_set_wr_signal_state(struct rds_ib_connection *ic,
 454                                              struct rds_ib_send_work *send,
 455                                              bool notify)
 456 {
 457         /*
 458          * We want to delay signaling completions just enough to get
 459          * the batching benefits but not so much that we create dead time
 460          * on the wire.
 461          */
 462         if (ic->i_unsignaled_wrs-- == 0 || notify) {
 463                 ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
 464                 send->s_wr.send_flags |= IB_SEND_SIGNALED;
 465                 return 1;
 466         }
 467         return 0;
 468 }
 469
 470 /*
 471  * This can be called multiple times for a given message.  The first time
 472  * we see a message we map its scatterlist into the IB device so that
 473  * we can provide that mapped address to the IB scatter gather entries
 474  * in the IB work requests.  We translate the scatterlist into a series
 475  * of work requests that fragment the message.  These work requests complete
 476  * in order so we pass ownership of the message to the completion handler
 477  * once we send the final fragment.
 478  *
 479  * The RDS core uses the c_send_lock to only enter this function once
 480  * per connection.  This makes sure that the tx ring alloc/unalloc pairs
 481  * don't get out of sync and confuse the ring.
 482  */
 483 int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
 484                 unsigned int hdr_off, unsigned int sg, unsigned int off)
 485 {
 486         struct rds_ib_connection *ic = conn->c_transport_data;
 487         struct ib_device *dev = ic->i_cm_id->device;
 488         struct rds_ib_send_work *send = NULL;
 489         struct rds_ib_send_work *first;
 490         struct rds_ib_send_work *prev;
 491         const struct ib_send_wr *failed_wr;
 492         struct scatterlist *scat;
 493         u32 pos;
 494         u32 i;
 495         u32 work_alloc;
 496         u32 credit_alloc = 0;
 497         u32 posted;
 498         u32 adv_credits = 0;
 499         int send_flags = 0;
 500         int bytes_sent = 0;
 501         int ret;
 502         int flow_controlled = 0;
 503         int nr_sig = 0;
 504
 505         BUG_ON(off % RDS_FRAG_SIZE);
 506         BUG_ON(hdr_off != 0 && hdr_off != sizeof(struct rds_header));
 507
 508         /* Do not send cong updates to IB loopback */
 509         if (conn->c_loopback
 510             && rm->m_inc.i_hdr.h_flags & RDS_FLAG_CONG_BITMAP) {
 511                 rds_cong_map_updated(conn->c_fcong, ~(u64) 0);
 512                 scat = &rm->data.op_sg[sg];
 513                 ret = max_t(int, RDS_CONG_MAP_BYTES, scat->length);
 514                 return sizeof(struct rds_header) + ret;
 515         }
 516
 517         /* FIXME we may overallocate here */
 518         if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0)
 519                 i = 1;
 520         else
 521                 i = DIV_ROUND_UP(be32_to_cpu(rm->m_inc.i_hdr.h_len), RDS_FRAG_SIZE);
 522
 523         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
 524         if (work_alloc == 0) {
 525                 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 526                 rds_ib_stats_inc(s_ib_tx_ring_full);
 527                 ret = -ENOMEM;
 528                 goto out;
 529         }
 530
 531         if (ic->i_flowctl) {
 532                 credit_alloc = rds_ib_send_grab_credits(ic, work_alloc, &posted, 0, RDS_MAX_ADV_CREDIT);
 533                 adv_credits += posted;
 534                 if (credit_alloc < work_alloc) {
 535                         rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - credit_alloc);
 536                         work_alloc = credit_alloc;
 537                         flow_controlled = 1;
 538                 }
 539                 if (work_alloc == 0) {
 540                         set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 541                         rds_ib_stats_inc(s_ib_tx_throttle);
 542                         ret = -ENOMEM;
 543                         goto out;
 544                 }
 545         }
 546
 547         /* map the message the first time we see it */
 548         if (!ic->i_data_op) {
 549                 if (rm->data.op_nents) {
 550                         rm->data.op_count = ib_dma_map_sg(dev,
 551                                                           rm->data.op_sg,
 552                                                           rm->data.op_nents,
 553                                                           DMA_TO_DEVICE);
 554                         rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->data.op_count);
 555                         if (rm->data.op_count == 0) {
 556                                 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
 557                                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 558                                 ret = -ENOMEM; /* XXX ? */
 559                                 goto out;
 560                         }
 561                 } else {
 562                         rm->data.op_count = 0;
 563                 }
 564
 565                 rds_message_addref(rm);
 566                 rm->data.op_dmasg = 0;
 567                 rm->data.op_dmaoff = 0;
 568                 ic->i_data_op = &rm->data;
 569
 570                 /* Finalize the header */
 571                 if (test_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags))
 572                         rm->m_inc.i_hdr.h_flags |= RDS_FLAG_ACK_REQUIRED;
 573                 if (test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))
 574                         rm->m_inc.i_hdr.h_flags |= RDS_FLAG_RETRANSMITTED;
 575
 576                 /* If it has a RDMA op, tell the peer we did it. This is
 577                  * used by the peer to release use-once RDMA MRs. */
 578                 if (rm->rdma.op_active) {
 579                         struct rds_ext_header_rdma ext_hdr;
 580
 581                         ext_hdr.h_rdma_rkey = cpu_to_be32(rm->rdma.op_rkey);
 582                         rds_message_add_extension(&rm->m_inc.i_hdr,
 583                                         RDS_EXTHDR_RDMA, &ext_hdr, sizeof(ext_hdr));
 584                 }
 585                 if (rm->m_rdma_cookie) {
 586                         rds_message_add_rdma_dest_extension(&rm->m_inc.i_hdr,
 587                                         rds_rdma_cookie_key(rm->m_rdma_cookie),
 588                                         rds_rdma_cookie_offset(rm->m_rdma_cookie));
 589                 }
 590
 591                 /* Note - rds_ib_piggyb_ack clears the ACK_REQUIRED bit, so
 592                  * we should not do this unless we have a chance of at least
 593                  * sticking the header into the send ring. Which is why we
 594                  * should call rds_ib_ring_alloc first. */
 595                 rm->m_inc.i_hdr.h_ack = cpu_to_be64(rds_ib_piggyb_ack(ic));
 596                 rds_message_make_checksum(&rm->m_inc.i_hdr);
 597
 598                 /*
 599                  * Update adv_credits since we reset the ACK_REQUIRED bit.
 600                  */
 601                 if (ic->i_flowctl) {
 602                         rds_ib_send_grab_credits(ic, 0, &posted, 1, RDS_MAX_ADV_CREDIT - adv_credits);
 603                         adv_credits += posted;
 604                         BUG_ON(adv_credits > 255);
 605                 }
 606         }
 607
 608         /* Sometimes you want to put a fence between an RDMA
 609          * READ and the following SEND.
 610          * We could either do this all the time
 611          * or when requested by the user. Right now, we let
 612          * the application choose.
 613          */
 614         if (rm->rdma.op_active && rm->rdma.op_fence)
 615                 send_flags = IB_SEND_FENCE;
 616
 617         /* Each frag gets a header. Msgs may be 0 bytes */
 618         send = &ic->i_sends[pos];
 619         first = send;
 620         prev = NULL;
 621         scat = &ic->i_data_op->op_sg[rm->data.op_dmasg];
 622         i = 0;
 623         do {
 624                 unsigned int len = 0;
 625
 626                 /* Set up the header */
 627                 send->s_wr.send_flags = send_flags;
 628                 send->s_wr.opcode = IB_WR_SEND;
 629                 send->s_wr.num_sge = 1;
 630                 send->s_wr.next = NULL;
 631                 send->s_queued = jiffies;
 632                 send->s_op = NULL;
 633
 634                 send->s_sge[0].addr = ic->i_send_hdrs_dma
 635                         + (pos * sizeof(struct rds_header));
 636                 send->s_sge[0].length = sizeof(struct rds_header);
 637
 638                 memcpy(&ic->i_send_hdrs[pos], &rm->m_inc.i_hdr, sizeof(struct rds_header));
 639
 640                 /* Set up the data, if present */
 641                 if (i < work_alloc
 642                     && scat != &rm->data.op_sg[rm->data.op_count]) {
 643                         len = min(RDS_FRAG_SIZE,
 644                                   sg_dma_len(scat) - rm->data.op_dmaoff);
 645                         send->s_wr.num_sge = 2;
 646
 647                         send->s_sge[1].addr = sg_dma_address(scat);
 648                         send->s_sge[1].addr += rm->data.op_dmaoff;
 649                         send->s_sge[1].length = len;
 650
 651                         bytes_sent += len;
 652                         rm->data.op_dmaoff += len;
 653                         if (rm->data.op_dmaoff == sg_dma_len(scat)) {
 654                                 scat++;
 655                                 rm->data.op_dmasg++;
 656                                 rm->data.op_dmaoff = 0;
 657                         }
 658                 }
 659
 660                 rds_ib_set_wr_signal_state(ic, send, false);
 661
 662                 /*
 663                  * Always signal the last one if we're stopping due to flow control.
 664                  */
 665                 if (ic->i_flowctl && flow_controlled && i == (work_alloc - 1)) {
 666                         rds_ib_set_wr_signal_state(ic, send, true);
 667                         send->s_wr.send_flags |= IB_SEND_SOLICITED;
 668                 }
 669
 670                 if (send->s_wr.send_flags & IB_SEND_SIGNALED)
 671                         nr_sig++;
 672
 673                 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
 674                          &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
 675
 676                 if (ic->i_flowctl && adv_credits) {
 677                         struct rds_header *hdr = &ic->i_send_hdrs[pos];
 678
 679                         /* add credit and redo the header checksum */
 680                         hdr->h_credit = adv_credits;
 681                         rds_message_make_checksum(hdr);
 682                         adv_credits = 0;
 683                         rds_ib_stats_inc(s_ib_tx_credit_updates);
 684                 }
 685
 686                 if (prev)
 687                         prev->s_wr.next = &send->s_wr;
 688                 prev = send;
 689
 690                 pos = (pos + 1) % ic->i_send_ring.w_nr;
 691                 send = &ic->i_sends[pos];
 692                 i++;
 693
 694         } while (i < work_alloc
 695                  && scat != &rm->data.op_sg[rm->data.op_count]);
 696
 697         /* Account the RDS header in the number of bytes we sent, but just once.
 698          * The caller has no concept of fragmentation. */
 699         if (hdr_off == 0)
 700                 bytes_sent += sizeof(struct rds_header);
 701
 702         /* if we finished the message then send completion owns it */
 703         if (scat == &rm->data.op_sg[rm->data.op_count]) {
 704                 prev->s_op = ic->i_data_op;
 705                 prev->s_wr.send_flags |= IB_SEND_SOLICITED;
 706                 if (!(prev->s_wr.send_flags & IB_SEND_SIGNALED))
 707                         nr_sig += rds_ib_set_wr_signal_state(ic, prev, true);
 708                 ic->i_data_op = NULL;
 709         }
 710
 711         /* Put back wrs & credits we didn't use */
 712         if (i < work_alloc) {
 713                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
 714                 work_alloc = i;
 715         }
 716         if (ic->i_flowctl && i < credit_alloc)
 717                 rds_ib_send_add_credits(conn, credit_alloc - i);
 718
 719         if (nr_sig)
 720                 atomic_add(nr_sig, &ic->i_signaled_sends);
 721
 722         /* XXX need to worry about failed_wr and partial sends. */
 723         failed_wr = &first->s_wr;
 724         ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
 725         rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
 726                  first, &first->s_wr, ret, failed_wr);
 727         BUG_ON(failed_wr != &first->s_wr);
 728         if (ret) {
 729                 printk(KERN_WARNING "RDS/IB: ib_post_send to %pI6c "
 730                        "returned %d\n", &conn->c_faddr, ret);
 731                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 732                 rds_ib_sub_signaled(ic, nr_sig);
 733                 if (prev->s_op) {
 734                         ic->i_data_op = prev->s_op;
 735                         prev->s_op = NULL;
 736                 }
 737
 738                 rds_ib_conn_error(ic->conn, "ib_post_send failed\n");
 739                 goto out;
 740         }
 741
 742         ret = bytes_sent;
 743 out:
 744         BUG_ON(adv_credits);
 745         return ret;
 746 }
 747
 748 /*
 749  * Issue atomic operation.
 750  * A simplified version of the rdma case, we always map 1 SG, and
 751  * only 8 bytes, for the return value from the atomic operation.
 752  */
 753 int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op)
 754 {
 755         struct rds_ib_connection *ic = conn->c_transport_data;
 756         struct rds_ib_send_work *send = NULL;
 757         const struct ib_send_wr *failed_wr;
 758         u32 pos;
 759         u32 work_alloc;
 760         int ret;
 761         int nr_sig = 0;
 762
 763         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, 1, &pos);
 764         if (work_alloc != 1) {
 765                 rds_ib_stats_inc(s_ib_tx_ring_full);
 766                 ret = -ENOMEM;
 767                 goto out;
 768         }
 769
 770         /* address of send request in ring */
 771         send = &ic->i_sends[pos];
 772         send->s_queued = jiffies;
 773
 774         if (op->op_type == RDS_ATOMIC_TYPE_CSWP) {
 775                 send->s_atomic_wr.wr.opcode = IB_WR_MASKED_ATOMIC_CMP_AND_SWP;
 776                 send->s_atomic_wr.compare_add = op->op_m_cswp.compare;
 777                 send->s_atomic_wr.swap = op->op_m_cswp.swap;
 778                 send->s_atomic_wr.compare_add_mask = op->op_m_cswp.compare_mask;
 779                 send->s_atomic_wr.swap_mask = op->op_m_cswp.swap_mask;
 780         } else { /* FADD */
 781                 send->s_atomic_wr.wr.opcode = IB_WR_MASKED_ATOMIC_FETCH_AND_ADD;
 782                 send->s_atomic_wr.compare_add = op->op_m_fadd.add;
 783                 send->s_atomic_wr.swap = 0;
 784                 send->s_atomic_wr.compare_add_mask = op->op_m_fadd.nocarry_mask;
 785                 send->s_atomic_wr.swap_mask = 0;
 786         }
 787         send->s_wr.send_flags = 0;
 788         nr_sig = rds_ib_set_wr_signal_state(ic, send, op->op_notify);
 789         send->s_atomic_wr.wr.num_sge = 1;
 790         send->s_atomic_wr.wr.next = NULL;
 791         send->s_atomic_wr.remote_addr = op->op_remote_addr;
 792         send->s_atomic_wr.rkey = op->op_rkey;
 793         send->s_op = op;
 794         rds_message_addref(container_of(send->s_op, struct rds_message, atomic));
 795
 796         /* map 8 byte retval buffer to the device */
 797         ret = ib_dma_map_sg(ic->i_cm_id->device, op->op_sg, 1, DMA_FROM_DEVICE);
 798         rdsdebug("ic %p mapping atomic op %p. mapped %d pg\n", ic, op, ret);
 799         if (ret != 1) {
 800                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 801                 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
 802                 ret = -ENOMEM; /* XXX ? */
 803                 goto out;
 804         }
 805
 806         /* Convert our struct scatterlist to struct ib_sge */
 807         send->s_sge[0].addr = sg_dma_address(op->op_sg);
 808         send->s_sge[0].length = sg_dma_len(op->op_sg);
 809         send->s_sge[0].lkey = ic->i_pd->local_dma_lkey;
 810
 811         rdsdebug("rva %Lx rpa %Lx len %u\n", op->op_remote_addr,
 812                  send->s_sge[0].addr, send->s_sge[0].length);
 813
 814         if (nr_sig)
 815                 atomic_add(nr_sig, &ic->i_signaled_sends);
 816
 817         failed_wr = &send->s_atomic_wr.wr;
 818         ret = ib_post_send(ic->i_cm_id->qp, &send->s_atomic_wr.wr, &failed_wr);
 819         rdsdebug("ic %p send %p (wr %p) ret %d wr %p\n", ic,
 820                  send, &send->s_atomic_wr, ret, failed_wr);
 821         BUG_ON(failed_wr != &send->s_atomic_wr.wr);
 822         if (ret) {
 823                 printk(KERN_WARNING "RDS/IB: atomic ib_post_send to %pI6c "
 824                        "returned %d\n", &conn->c_faddr, ret);
 825                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 826                 rds_ib_sub_signaled(ic, nr_sig);
 827                 goto out;
 828         }
 829
 830         if (unlikely(failed_wr != &send->s_atomic_wr.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_atomic_wr.wr);
 833         }
 834
 835 out:
 836         return ret;
 837 }
 838
 839 int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_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         const struct ib_send_wr *failed_wr;
 846         struct scatterlist *scat;
 847         unsigned long len;
 848         u64 remote_addr = op->op_remote_addr;
 849         u32 max_sge = ic->rds_ibdev->max_sge;
 850         u32 pos;
 851         u32 work_alloc;
 852         u32 i;
 853         u32 j;
 854         int sent;
 855         int ret;
 856         int num_sge;
 857         int nr_sig = 0;
 858
 859         /* map the op the first time we see it */
 860         if (!op->op_mapped) {
 861                 op->op_count = ib_dma_map_sg(ic->i_cm_id->device,
 862                                              op->op_sg, op->op_nents, (op->op_write) ?
 863                                              DMA_TO_DEVICE : DMA_FROM_DEVICE);
 864                 rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->op_count);
 865                 if (op->op_count == 0) {
 866                         rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
 867                         ret = -ENOMEM; /* XXX ? */
 868                         goto out;
 869                 }
 870
 871                 op->op_mapped = 1;
 872         }
 873
 874         /*
 875          * Instead of knowing how to return a partial rdma read/write we insist that there
 876          * be enough work requests to send the entire message.
 877          */
 878         i = DIV_ROUND_UP(op->op_count, max_sge);
 879
 880         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
 881         if (work_alloc != i) {
 882                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 883                 rds_ib_stats_inc(s_ib_tx_ring_full);
 884                 ret = -ENOMEM;
 885                 goto out;
 886         }
 887
 888         send = &ic->i_sends[pos];
 889         first = send;
 890         prev = NULL;
 891         scat = &op->op_sg[0];
 892         sent = 0;
 893         num_sge = op->op_count;
 894
 895         for (i = 0; i < work_alloc && scat != &op->op_sg[op->op_count]; i++) {
 896                 send->s_wr.send_flags = 0;
 897                 send->s_queued = jiffies;
 898                 send->s_op = NULL;
 899
 900                 if (!op->op_notify)
 901                         nr_sig += rds_ib_set_wr_signal_state(ic, send,
 902                                                              op->op_notify);
 903
 904                 send->s_wr.opcode = op->op_write ? IB_WR_RDMA_WRITE : IB_WR_RDMA_READ;
 905                 send->s_rdma_wr.remote_addr = remote_addr;
 906                 send->s_rdma_wr.rkey = op->op_rkey;
 907
 908                 if (num_sge > max_sge) {
 909                         send->s_rdma_wr.wr.num_sge = max_sge;
 910                         num_sge -= max_sge;
 911                 } else {
 912                         send->s_rdma_wr.wr.num_sge = num_sge;
 913                 }
 914
 915                 send->s_rdma_wr.wr.next = NULL;
 916
 917                 if (prev)
 918                         prev->s_rdma_wr.wr.next = &send->s_rdma_wr.wr;
 919
 920                 for (j = 0; j < send->s_rdma_wr.wr.num_sge &&
 921                      scat != &op->op_sg[op->op_count]; j++) {
 922                         len = sg_dma_len(scat);
 923                         send->s_sge[j].addr = sg_dma_address(scat);
 924                         send->s_sge[j].length = len;
 925                         send->s_sge[j].lkey = ic->i_pd->local_dma_lkey;
 926
 927                         sent += len;
 928                         rdsdebug("ic %p sent %d remote_addr %llu\n", ic, sent, remote_addr);
 929
 930                         remote_addr += len;
 931                         scat++;
 932                 }
 933
 934                 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
 935                         &send->s_rdma_wr.wr,
 936                         send->s_rdma_wr.wr.num_sge,
 937                         send->s_rdma_wr.wr.next);
 938
 939                 prev = send;
 940                 if (++send == &ic->i_sends[ic->i_send_ring.w_nr])
 941                         send = ic->i_sends;
 942         }
 943
 944         /* give a reference to the last op */
 945         if (scat == &op->op_sg[op->op_count]) {
 946                 prev->s_op = op;
 947                 rds_message_addref(container_of(op, struct rds_message, rdma));
 948         }
 949
 950         if (i < work_alloc) {
 951                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
 952                 work_alloc = i;
 953         }
 954
 955         if (nr_sig)
 956                 atomic_add(nr_sig, &ic->i_signaled_sends);
 957
 958         failed_wr = &first->s_rdma_wr.wr;
 959         ret = ib_post_send(ic->i_cm_id->qp, &first->s_rdma_wr.wr, &failed_wr);
 960         rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
 961                  first, &first->s_rdma_wr.wr, ret, failed_wr);
 962         BUG_ON(failed_wr != &first->s_rdma_wr.wr);
 963         if (ret) {
 964                 printk(KERN_WARNING "RDS/IB: rdma ib_post_send to %pI6c "
 965                        "returned %d\n", &conn->c_faddr, ret);
 966                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 967                 rds_ib_sub_signaled(ic, nr_sig);
 968                 goto out;
 969         }
 970
 971         if (unlikely(failed_wr != &first->s_rdma_wr.wr)) {
 972                 printk(KERN_WARNING "RDS/IB: ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
 973                 BUG_ON(failed_wr != &first->s_rdma_wr.wr);
 974         }
 975
 976
 977 out:
 978         return ret;
 979 }
 980
 981 void rds_ib_xmit_path_complete(struct rds_conn_path *cp)
 982 {
 983         struct rds_connection *conn = cp->cp_conn;
 984         struct rds_ib_connection *ic = conn->c_transport_data;
 985
 986         /* We may have a pending ACK or window update we were unable
 987          * to send previously (due to flow control). Try again. */
 988         rds_ib_attempt_ack(ic);
 989 }