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
 298         if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
 299             test_bit(0, &conn->c_map_queued))
 300                 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
 301
 302         /* We expect errors as the qp is drained during shutdown */
 303         if (wc->status != IB_WC_SUCCESS && rds_conn_up(conn)) {
 304                 rds_ib_conn_error(conn, "send completion on <%pI6c,%pI6c,%d> had status %u (%s), vendor err 0x%x, disconnecting and reconnecting\n",
 305                                   &conn->c_laddr, &conn->c_faddr,
 306                                   conn->c_tos, wc->status,
 307                                   ib_wc_status_msg(wc->status), wc->vendor_err);
 308         }
 309 }
 310
 311 /*
 312  * This is the main function for allocating credits when sending
 313  * messages.
 314  *
 315  * Conceptually, we have two counters:
 316  *  -   send credits: this tells us how many WRs we're allowed
 317  *      to submit without overruning the receiver's queue. For
 318  *      each SEND WR we post, we decrement this by one.
 319  *
 320  *  -   posted credits: this tells us how many WRs we recently
 321  *      posted to the receive queue. This value is transferred
 322  *      to the peer as a "credit update" in a RDS header field.
 323  *      Every time we transmit credits to the peer, we subtract
 324  *      the amount of transferred credits from this counter.
 325  *
 326  * It is essential that we avoid situations where both sides have
 327  * exhausted their send credits, and are unable to send new credits
 328  * to the peer. We achieve this by requiring that we send at least
 329  * one credit update to the peer before exhausting our credits.
 330  * When new credits arrive, we subtract one credit that is withheld
 331  * until we've posted new buffers and are ready to transmit these
 332  * credits (see rds_ib_send_add_credits below).
 333  *
 334  * The RDS send code is essentially single-threaded; rds_send_xmit
 335  * sets RDS_IN_XMIT to ensure exclusive access to the send ring.
 336  * However, the ACK sending code is independent and can race with
 337  * message SENDs.
 338  *
 339  * In the send path, we need to update the counters for send credits
 340  * and the counter of posted buffers atomically - when we use the
 341  * last available credit, we cannot allow another thread to race us
 342  * and grab the posted credits counter.  Hence, we have to use a
 343  * spinlock to protect the credit counter, or use atomics.
 344  *
 345  * Spinlocks shared between the send and the receive path are bad,
 346  * because they create unnecessary delays. An early implementation
 347  * using a spinlock showed a 5% degradation in throughput at some
 348  * loads.
 349  *
 350  * This implementation avoids spinlocks completely, putting both
 351  * counters into a single atomic, and updating that atomic using
 352  * atomic_add (in the receive path, when receiving fresh credits),
 353  * and using atomic_cmpxchg when updating the two counters.
 354  */
 355 int rds_ib_send_grab_credits(struct rds_ib_connection *ic,
 356                              u32 wanted, u32 *adv_credits, int need_posted, int max_posted)
 357 {
 358         unsigned int avail, posted, got = 0, advertise;
 359         long oldval, newval;
 360
 361         *adv_credits = 0;
 362         if (!ic->i_flowctl)
 363                 return wanted;
 364
 365 try_again:
 366         advertise = 0;
 367         oldval = newval = atomic_read(&ic->i_credits);
 368         posted = IB_GET_POST_CREDITS(oldval);
 369         avail = IB_GET_SEND_CREDITS(oldval);
 370
 371         rdsdebug("wanted=%u credits=%u posted=%u\n",
 372                         wanted, avail, posted);
 373
 374         /* The last credit must be used to send a credit update. */
 375         if (avail && !posted)
 376                 avail--;
 377
 378         if (avail < wanted) {
 379                 struct rds_connection *conn = ic->i_cm_id->context;
 380
 381                 /* Oops, there aren't that many credits left! */
 382                 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 383                 got = avail;
 384         } else {
 385                 /* Sometimes you get what you want, lalala. */
 386                 got = wanted;
 387         }
 388         newval -= IB_SET_SEND_CREDITS(got);
 389
 390         /*
 391          * If need_posted is non-zero, then the caller wants
 392          * the posted regardless of whether any send credits are
 393          * available.
 394          */
 395         if (posted && (got || need_posted)) {
 396                 advertise = min_t(unsigned int, posted, max_posted);
 397                 newval -= IB_SET_POST_CREDITS(advertise);
 398         }
 399
 400         /* Finally bill everything */
 401         if (atomic_cmpxchg(&ic->i_credits, oldval, newval) != oldval)
 402                 goto try_again;
 403
 404         *adv_credits = advertise;
 405         return got;
 406 }
 407
 408 void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits)
 409 {
 410         struct rds_ib_connection *ic = conn->c_transport_data;
 411
 412         if (credits == 0)
 413                 return;
 414
 415         rdsdebug("credits=%u current=%u%s\n",
 416                         credits,
 417                         IB_GET_SEND_CREDITS(atomic_read(&ic->i_credits)),
 418                         test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ? ", ll_send_full" : "");
 419
 420         atomic_add(IB_SET_SEND_CREDITS(credits), &ic->i_credits);
 421         if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags))
 422                 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
 423
 424         WARN_ON(IB_GET_SEND_CREDITS(credits) >= 16384);
 425
 426         rds_ib_stats_inc(s_ib_rx_credit_updates);
 427 }
 428
 429 void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted)
 430 {
 431         struct rds_ib_connection *ic = conn->c_transport_data;
 432
 433         if (posted == 0)
 434                 return;
 435
 436         atomic_add(IB_SET_POST_CREDITS(posted), &ic->i_credits);
 437
 438         /* Decide whether to send an update to the peer now.
 439          * If we would send a credit update for every single buffer we
 440          * post, we would end up with an ACK storm (ACK arrives,
 441          * consumes buffer, we refill the ring, send ACK to remote
 442          * advertising the newly posted buffer... ad inf)
 443          *
 444          * Performance pretty much depends on how often we send
 445          * credit updates - too frequent updates mean lots of ACKs.
 446          * Too infrequent updates, and the peer will run out of
 447          * credits and has to throttle.
 448          * For the time being, 16 seems to be a good compromise.
 449          */
 450         if (IB_GET_POST_CREDITS(atomic_read(&ic->i_credits)) >= 16)
 451                 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
 452 }
 453
 454 static inline int rds_ib_set_wr_signal_state(struct rds_ib_connection *ic,
 455                                              struct rds_ib_send_work *send,
 456                                              bool notify)
 457 {
 458         /*
 459          * We want to delay signaling completions just enough to get
 460          * the batching benefits but not so much that we create dead time
 461          * on the wire.
 462          */
 463         if (ic->i_unsignaled_wrs-- == 0 || notify) {
 464                 ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
 465                 send->s_wr.send_flags |= IB_SEND_SIGNALED;
 466                 return 1;
 467         }
 468         return 0;
 469 }
 470
 471 /*
 472  * This can be called multiple times for a given message.  The first time
 473  * we see a message we map its scatterlist into the IB device so that
 474  * we can provide that mapped address to the IB scatter gather entries
 475  * in the IB work requests.  We translate the scatterlist into a series
 476  * of work requests that fragment the message.  These work requests complete
 477  * in order so we pass ownership of the message to the completion handler
 478  * once we send the final fragment.
 479  *
 480  * The RDS core uses the c_send_lock to only enter this function once
 481  * per connection.  This makes sure that the tx ring alloc/unalloc pairs
 482  * don't get out of sync and confuse the ring.
 483  */
 484 int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
 485                 unsigned int hdr_off, unsigned int sg, unsigned int off)
 486 {
 487         struct rds_ib_connection *ic = conn->c_transport_data;
 488         struct ib_device *dev = ic->i_cm_id->device;
 489         struct rds_ib_send_work *send = NULL;
 490         struct rds_ib_send_work *first;
 491         struct rds_ib_send_work *prev;
 492         const struct ib_send_wr *failed_wr;
 493         struct scatterlist *scat;
 494         u32 pos;
 495         u32 i;
 496         u32 work_alloc;
 497         u32 credit_alloc = 0;
 498         u32 posted;
 499         u32 adv_credits = 0;
 500         int send_flags = 0;
 501         int bytes_sent = 0;
 502         int ret;
 503         int flow_controlled = 0;
 504         int nr_sig = 0;
 505
 506         BUG_ON(off % RDS_FRAG_SIZE);
 507         BUG_ON(hdr_off != 0 && hdr_off != sizeof(struct rds_header));
 508
 509         /* Do not send cong updates to IB loopback */
 510         if (conn->c_loopback
 511             && rm->m_inc.i_hdr.h_flags & RDS_FLAG_CONG_BITMAP) {
 512                 rds_cong_map_updated(conn->c_fcong, ~(u64) 0);
 513                 scat = &rm->data.op_sg[sg];
 514                 ret = max_t(int, RDS_CONG_MAP_BYTES, scat->length);
 515                 return sizeof(struct rds_header) + ret;
 516         }
 517
 518         /* FIXME we may overallocate here */
 519         if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0)
 520                 i = 1;
 521         else
 522                 i = DIV_ROUND_UP(be32_to_cpu(rm->m_inc.i_hdr.h_len), RDS_FRAG_SIZE);
 523
 524         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
 525         if (work_alloc == 0) {
 526                 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 527                 rds_ib_stats_inc(s_ib_tx_ring_full);
 528                 ret = -ENOMEM;
 529                 goto out;
 530         }
 531
 532         if (ic->i_flowctl) {
 533                 credit_alloc = rds_ib_send_grab_credits(ic, work_alloc, &posted, 0, RDS_MAX_ADV_CREDIT);
 534                 adv_credits += posted;
 535                 if (credit_alloc < work_alloc) {
 536                         rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - credit_alloc);
 537                         work_alloc = credit_alloc;
 538                         flow_controlled = 1;
 539                 }
 540                 if (work_alloc == 0) {
 541                         set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 542                         rds_ib_stats_inc(s_ib_tx_throttle);
 543                         ret = -ENOMEM;
 544                         goto out;
 545                 }
 546         }
 547
 548         /* map the message the first time we see it */
 549         if (!ic->i_data_op) {
 550                 if (rm->data.op_nents) {
 551                         rm->data.op_count = ib_dma_map_sg(dev,
 552                                                           rm->data.op_sg,
 553                                                           rm->data.op_nents,
 554                                                           DMA_TO_DEVICE);
 555                         rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->data.op_count);
 556                         if (rm->data.op_count == 0) {
 557                                 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
 558                                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 559                                 ret = -ENOMEM; /* XXX ? */
 560                                 goto out;
 561                         }
 562                 } else {
 563                         rm->data.op_count = 0;
 564                 }
 565
 566                 rds_message_addref(rm);
 567                 rm->data.op_dmasg = 0;
 568                 rm->data.op_dmaoff = 0;
 569                 ic->i_data_op = &rm->data;
 570
 571                 /* Finalize the header */
 572                 if (test_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags))
 573                         rm->m_inc.i_hdr.h_flags |= RDS_FLAG_ACK_REQUIRED;
 574                 if (test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))
 575                         rm->m_inc.i_hdr.h_flags |= RDS_FLAG_RETRANSMITTED;
 576
 577                 /* If it has a RDMA op, tell the peer we did it. This is
 578                  * used by the peer to release use-once RDMA MRs. */
 579                 if (rm->rdma.op_active) {
 580                         struct rds_ext_header_rdma ext_hdr;
 581
 582                         ext_hdr.h_rdma_rkey = cpu_to_be32(rm->rdma.op_rkey);
 583                         rds_message_add_extension(&rm->m_inc.i_hdr,
 584                                         RDS_EXTHDR_RDMA, &ext_hdr, sizeof(ext_hdr));
 585                 }
 586                 if (rm->m_rdma_cookie) {
 587                         rds_message_add_rdma_dest_extension(&rm->m_inc.i_hdr,
 588                                         rds_rdma_cookie_key(rm->m_rdma_cookie),
 589                                         rds_rdma_cookie_offset(rm->m_rdma_cookie));
 590                 }
 591
 592                 /* Note - rds_ib_piggyb_ack clears the ACK_REQUIRED bit, so
 593                  * we should not do this unless we have a chance of at least
 594                  * sticking the header into the send ring. Which is why we
 595                  * should call rds_ib_ring_alloc first. */
 596                 rm->m_inc.i_hdr.h_ack = cpu_to_be64(rds_ib_piggyb_ack(ic));
 597                 rds_message_make_checksum(&rm->m_inc.i_hdr);
 598
 599                 /*
 600                  * Update adv_credits since we reset the ACK_REQUIRED bit.
 601                  */
 602                 if (ic->i_flowctl) {
 603                         rds_ib_send_grab_credits(ic, 0, &posted, 1, RDS_MAX_ADV_CREDIT - adv_credits);
 604                         adv_credits += posted;
 605                         BUG_ON(adv_credits > 255);
 606                 }
 607         }
 608
 609         /* Sometimes you want to put a fence between an RDMA
 610          * READ and the following SEND.
 611          * We could either do this all the time
 612          * or when requested by the user. Right now, we let
 613          * the application choose.
 614          */
 615         if (rm->rdma.op_active && rm->rdma.op_fence)
 616                 send_flags = IB_SEND_FENCE;
 617
 618         /* Each frag gets a header. Msgs may be 0 bytes */
 619         send = &ic->i_sends[pos];
 620         first = send;
 621         prev = NULL;
 622         scat = &ic->i_data_op->op_sg[rm->data.op_dmasg];
 623         i = 0;
 624         do {
 625                 unsigned int len = 0;
 626
 627                 /* Set up the header */
 628                 send->s_wr.send_flags = send_flags;
 629                 send->s_wr.opcode = IB_WR_SEND;
 630                 send->s_wr.num_sge = 1;
 631                 send->s_wr.next = NULL;
 632                 send->s_queued = jiffies;
 633                 send->s_op = NULL;
 634
 635                 send->s_sge[0].addr = ic->i_send_hdrs_dma[pos];
 636
 637                 send->s_sge[0].length = sizeof(struct rds_header);
 638                 send->s_sge[0].lkey = ic->i_pd->local_dma_lkey;
 639
 640                 ib_dma_sync_single_for_cpu(ic->rds_ibdev->dev,
 641                                            ic->i_send_hdrs_dma[pos],
 642                                            sizeof(struct rds_header),
 643                                            DMA_TO_DEVICE);
 644                 memcpy(ic->i_send_hdrs[pos], &rm->m_inc.i_hdr,
 645                        sizeof(struct rds_header));
 646
 647
 648                 /* Set up the data, if present */
 649                 if (i < work_alloc
 650                     && scat != &rm->data.op_sg[rm->data.op_count]) {
 651                         len = min(RDS_FRAG_SIZE,
 652                                   sg_dma_len(scat) - rm->data.op_dmaoff);
 653                         send->s_wr.num_sge = 2;
 654
 655                         send->s_sge[1].addr = sg_dma_address(scat);
 656                         send->s_sge[1].addr += rm->data.op_dmaoff;
 657                         send->s_sge[1].length = len;
 658                         send->s_sge[1].lkey = ic->i_pd->local_dma_lkey;
 659
 660                         bytes_sent += len;
 661                         rm->data.op_dmaoff += len;
 662                         if (rm->data.op_dmaoff == sg_dma_len(scat)) {
 663                                 scat++;
 664                                 rm->data.op_dmasg++;
 665                                 rm->data.op_dmaoff = 0;
 666                         }
 667                 }
 668
 669                 rds_ib_set_wr_signal_state(ic, send, false);
 670
 671                 /*
 672                  * Always signal the last one if we're stopping due to flow control.
 673                  */
 674                 if (ic->i_flowctl && flow_controlled && i == (work_alloc - 1)) {
 675                         rds_ib_set_wr_signal_state(ic, send, true);
 676                         send->s_wr.send_flags |= IB_SEND_SOLICITED;
 677                 }
 678
 679                 if (send->s_wr.send_flags & IB_SEND_SIGNALED)
 680                         nr_sig++;
 681
 682                 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
 683                          &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
 684
 685                 if (ic->i_flowctl && adv_credits) {
 686                         struct rds_header *hdr = ic->i_send_hdrs[pos];
 687
 688                         /* add credit and redo the header checksum */
 689                         hdr->h_credit = adv_credits;
 690                         rds_message_make_checksum(hdr);
 691                         adv_credits = 0;
 692                         rds_ib_stats_inc(s_ib_tx_credit_updates);
 693                 }
 694                 ib_dma_sync_single_for_device(ic->rds_ibdev->dev,
 695                                               ic->i_send_hdrs_dma[pos],
 696                                               sizeof(struct rds_header),
 697                                               DMA_TO_DEVICE);
 698
 699                 if (prev)
 700                         prev->s_wr.next = &send->s_wr;
 701                 prev = send;
 702
 703                 pos = (pos + 1) % ic->i_send_ring.w_nr;
 704                 send = &ic->i_sends[pos];
 705                 i++;
 706
 707         } while (i < work_alloc
 708                  && scat != &rm->data.op_sg[rm->data.op_count]);
 709
 710         /* Account the RDS header in the number of bytes we sent, but just once.
 711          * The caller has no concept of fragmentation. */
 712         if (hdr_off == 0)
 713                 bytes_sent += sizeof(struct rds_header);
 714
 715         /* if we finished the message then send completion owns it */
 716         if (scat == &rm->data.op_sg[rm->data.op_count]) {
 717                 prev->s_op = ic->i_data_op;
 718                 prev->s_wr.send_flags |= IB_SEND_SOLICITED;
 719                 if (!(prev->s_wr.send_flags & IB_SEND_SIGNALED))
 720                         nr_sig += rds_ib_set_wr_signal_state(ic, prev, true);
 721                 ic->i_data_op = NULL;
 722         }
 723
 724         /* Put back wrs & credits we didn't use */
 725         if (i < work_alloc) {
 726                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
 727                 work_alloc = i;
 728         }
 729         if (ic->i_flowctl && i < credit_alloc)
 730                 rds_ib_send_add_credits(conn, credit_alloc - i);
 731
 732         if (nr_sig)
 733                 atomic_add(nr_sig, &ic->i_signaled_sends);
 734
 735         /* XXX need to worry about failed_wr and partial sends. */
 736         failed_wr = &first->s_wr;
 737         ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
 738         rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
 739                  first, &first->s_wr, ret, failed_wr);
 740         BUG_ON(failed_wr != &first->s_wr);
 741         if (ret) {
 742                 printk(KERN_WARNING "RDS/IB: ib_post_send to %pI6c "
 743                        "returned %d\n", &conn->c_faddr, ret);
 744                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 745                 rds_ib_sub_signaled(ic, nr_sig);
 746                 if (prev->s_op) {
 747                         ic->i_data_op = prev->s_op;
 748                         prev->s_op = NULL;
 749                 }
 750
 751                 rds_ib_conn_error(ic->conn, "ib_post_send failed\n");
 752                 goto out;
 753         }
 754
 755         ret = bytes_sent;
 756 out:
 757         BUG_ON(adv_credits);
 758         return ret;
 759 }
 760
 761 /*
 762  * Issue atomic operation.
 763  * A simplified version of the rdma case, we always map 1 SG, and
 764  * only 8 bytes, for the return value from the atomic operation.
 765  */
 766 int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op)
 767 {
 768         struct rds_ib_connection *ic = conn->c_transport_data;
 769         struct rds_ib_send_work *send = NULL;
 770         const struct ib_send_wr *failed_wr;
 771         u32 pos;
 772         u32 work_alloc;
 773         int ret;
 774         int nr_sig = 0;
 775
 776         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, 1, &pos);
 777         if (work_alloc != 1) {
 778                 rds_ib_stats_inc(s_ib_tx_ring_full);
 779                 ret = -ENOMEM;
 780                 goto out;
 781         }
 782
 783         /* address of send request in ring */
 784         send = &ic->i_sends[pos];
 785         send->s_queued = jiffies;
 786
 787         if (op->op_type == RDS_ATOMIC_TYPE_CSWP) {
 788                 send->s_atomic_wr.wr.opcode = IB_WR_MASKED_ATOMIC_CMP_AND_SWP;
 789                 send->s_atomic_wr.compare_add = op->op_m_cswp.compare;
 790                 send->s_atomic_wr.swap = op->op_m_cswp.swap;
 791                 send->s_atomic_wr.compare_add_mask = op->op_m_cswp.compare_mask;
 792                 send->s_atomic_wr.swap_mask = op->op_m_cswp.swap_mask;
 793         } else { /* FADD */
 794                 send->s_atomic_wr.wr.opcode = IB_WR_MASKED_ATOMIC_FETCH_AND_ADD;
 795                 send->s_atomic_wr.compare_add = op->op_m_fadd.add;
 796                 send->s_atomic_wr.swap = 0;
 797                 send->s_atomic_wr.compare_add_mask = op->op_m_fadd.nocarry_mask;
 798                 send->s_atomic_wr.swap_mask = 0;
 799         }
 800         send->s_wr.send_flags = 0;
 801         nr_sig = rds_ib_set_wr_signal_state(ic, send, op->op_notify);
 802         send->s_atomic_wr.wr.num_sge = 1;
 803         send->s_atomic_wr.wr.next = NULL;
 804         send->s_atomic_wr.remote_addr = op->op_remote_addr;
 805         send->s_atomic_wr.rkey = op->op_rkey;
 806         send->s_op = op;
 807         rds_message_addref(container_of(send->s_op, struct rds_message, atomic));
 808
 809         /* map 8 byte retval buffer to the device */
 810         ret = ib_dma_map_sg(ic->i_cm_id->device, op->op_sg, 1, DMA_FROM_DEVICE);
 811         rdsdebug("ic %p mapping atomic op %p. mapped %d pg\n", ic, op, ret);
 812         if (ret != 1) {
 813                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 814                 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
 815                 ret = -ENOMEM; /* XXX ? */
 816                 goto out;
 817         }
 818
 819         /* Convert our struct scatterlist to struct ib_sge */
 820         send->s_sge[0].addr = sg_dma_address(op->op_sg);
 821         send->s_sge[0].length = sg_dma_len(op->op_sg);
 822         send->s_sge[0].lkey = ic->i_pd->local_dma_lkey;
 823
 824         rdsdebug("rva %Lx rpa %Lx len %u\n", op->op_remote_addr,
 825                  send->s_sge[0].addr, send->s_sge[0].length);
 826
 827         if (nr_sig)
 828                 atomic_add(nr_sig, &ic->i_signaled_sends);
 829
 830         failed_wr = &send->s_atomic_wr.wr;
 831         ret = ib_post_send(ic->i_cm_id->qp, &send->s_atomic_wr.wr, &failed_wr);
 832         rdsdebug("ic %p send %p (wr %p) ret %d wr %p\n", ic,
 833                  send, &send->s_atomic_wr, ret, failed_wr);
 834         BUG_ON(failed_wr != &send->s_atomic_wr.wr);
 835         if (ret) {
 836                 printk(KERN_WARNING "RDS/IB: atomic ib_post_send to %pI6c "
 837                        "returned %d\n", &conn->c_faddr, ret);
 838                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 839                 rds_ib_sub_signaled(ic, nr_sig);
 840                 goto out;
 841         }
 842
 843         if (unlikely(failed_wr != &send->s_atomic_wr.wr)) {
 844                 printk(KERN_WARNING "RDS/IB: atomic ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
 845                 BUG_ON(failed_wr != &send->s_atomic_wr.wr);
 846         }
 847
 848 out:
 849         return ret;
 850 }
 851
 852 int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op)
 853 {
 854         struct rds_ib_connection *ic = conn->c_transport_data;
 855         struct rds_ib_send_work *send = NULL;
 856         struct rds_ib_send_work *first;
 857         struct rds_ib_send_work *prev;
 858         const struct ib_send_wr *failed_wr;
 859         struct scatterlist *scat;
 860         unsigned long len;
 861         u64 remote_addr = op->op_remote_addr;
 862         u32 max_sge = ic->rds_ibdev->max_sge;
 863         u32 pos;
 864         u32 work_alloc;
 865         u32 i;
 866         u32 j;
 867         int sent;
 868         int ret;
 869         int num_sge;
 870         int nr_sig = 0;
 871         u64 odp_addr = op->op_odp_addr;
 872         u32 odp_lkey = 0;
 873
 874         /* map the op the first time we see it */
 875         if (!op->op_odp_mr) {
 876                 if (!op->op_mapped) {
 877                         op->op_count =
 878                                 ib_dma_map_sg(ic->i_cm_id->device, op->op_sg,
 879                                               op->op_nents,
 880                                               (op->op_write) ? DMA_TO_DEVICE :
 881                                                                DMA_FROM_DEVICE);
 882                         rdsdebug("ic %p mapping op %p: %d\n", ic, op,
 883                                  op->op_count);
 884                         if (op->op_count == 0) {
 885                                 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
 886                                 ret = -ENOMEM; /* XXX ? */
 887                                 goto out;
 888                         }
 889                         op->op_mapped = 1;
 890                 }
 891         } else {
 892                 op->op_count = op->op_nents;
 893                 odp_lkey = rds_ib_get_lkey(op->op_odp_mr->r_trans_private);
 894         }
 895
 896         /*
 897          * Instead of knowing how to return a partial rdma read/write we insist that there
 898          * be enough work requests to send the entire message.
 899          */
 900         i = DIV_ROUND_UP(op->op_count, max_sge);
 901
 902         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
 903         if (work_alloc != i) {
 904                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 905                 rds_ib_stats_inc(s_ib_tx_ring_full);
 906                 ret = -ENOMEM;
 907                 goto out;
 908         }
 909
 910         send = &ic->i_sends[pos];
 911         first = send;
 912         prev = NULL;
 913         scat = &op->op_sg[0];
 914         sent = 0;
 915         num_sge = op->op_count;
 916
 917         for (i = 0; i < work_alloc && scat != &op->op_sg[op->op_count]; i++) {
 918                 send->s_wr.send_flags = 0;
 919                 send->s_queued = jiffies;
 920                 send->s_op = NULL;
 921
 922                 if (!op->op_notify)
 923                         nr_sig += rds_ib_set_wr_signal_state(ic, send,
 924                                                              op->op_notify);
 925
 926                 send->s_wr.opcode = op->op_write ? IB_WR_RDMA_WRITE : IB_WR_RDMA_READ;
 927                 send->s_rdma_wr.remote_addr = remote_addr;
 928                 send->s_rdma_wr.rkey = op->op_rkey;
 929
 930                 if (num_sge > max_sge) {
 931                         send->s_rdma_wr.wr.num_sge = max_sge;
 932                         num_sge -= max_sge;
 933                 } else {
 934                         send->s_rdma_wr.wr.num_sge = num_sge;
 935                 }
 936
 937                 send->s_rdma_wr.wr.next = NULL;
 938
 939                 if (prev)
 940                         prev->s_rdma_wr.wr.next = &send->s_rdma_wr.wr;
 941
 942                 for (j = 0; j < send->s_rdma_wr.wr.num_sge &&
 943                      scat != &op->op_sg[op->op_count]; j++) {
 944                         len = sg_dma_len(scat);
 945                         if (!op->op_odp_mr) {
 946                                 send->s_sge[j].addr = sg_dma_address(scat);
 947                                 send->s_sge[j].lkey = ic->i_pd->local_dma_lkey;
 948                         } else {
 949                                 send->s_sge[j].addr = odp_addr;
 950                                 send->s_sge[j].lkey = odp_lkey;
 951                         }
 952                         send->s_sge[j].length = len;
 953
 954                         sent += len;
 955                         rdsdebug("ic %p sent %d remote_addr %llu\n", ic, sent, remote_addr);
 956
 957                         remote_addr += len;
 958                         odp_addr += len;
 959                         scat++;
 960                 }
 961
 962                 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
 963                         &send->s_rdma_wr.wr,
 964                         send->s_rdma_wr.wr.num_sge,
 965                         send->s_rdma_wr.wr.next);
 966
 967                 prev = send;
 968                 if (++send == &ic->i_sends[ic->i_send_ring.w_nr])
 969                         send = ic->i_sends;
 970         }
 971
 972         /* give a reference to the last op */
 973         if (scat == &op->op_sg[op->op_count]) {
 974                 prev->s_op = op;
 975                 rds_message_addref(container_of(op, struct rds_message, rdma));
 976         }
 977
 978         if (i < work_alloc) {
 979                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
 980                 work_alloc = i;
 981         }
 982
 983         if (nr_sig)
 984                 atomic_add(nr_sig, &ic->i_signaled_sends);
 985
 986         failed_wr = &first->s_rdma_wr.wr;
 987         ret = ib_post_send(ic->i_cm_id->qp, &first->s_rdma_wr.wr, &failed_wr);
 988         rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
 989                  first, &first->s_rdma_wr.wr, ret, failed_wr);
 990         BUG_ON(failed_wr != &first->s_rdma_wr.wr);
 991         if (ret) {
 992                 printk(KERN_WARNING "RDS/IB: rdma ib_post_send to %pI6c "
 993                        "returned %d\n", &conn->c_faddr, ret);
 994                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 995                 rds_ib_sub_signaled(ic, nr_sig);
 996                 goto out;
 997         }
 998
 999         if (unlikely(failed_wr != &first->s_rdma_wr.wr)) {
1000                 printk(KERN_WARNING "RDS/IB: ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
1001                 BUG_ON(failed_wr != &first->s_rdma_wr.wr);
1002         }
1003
1004
1005 out:
1006         return ret;
1007 }
1008
1009 void rds_ib_xmit_path_complete(struct rds_conn_path *cp)
1010 {
1011         struct rds_connection *conn = cp->cp_conn;
1012         struct rds_ib_connection *ic = conn->c_transport_data;
1013
1014         /* We may have a pending ACK or window update we were unable
1015          * to send previously (due to flow control). Try again. */
1016         rds_ib_attempt_ack(ic);
1017 }