net/rds/ib_send.c

   1 /*
   2  * Copyright (c) 2006 Oracle.  All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  *
  32  */
  33 #include <linux/kernel.h>
  34 #include <linux/in.h>
  35 #include <linux/device.h>
  36 #include <linux/dmapool.h>
  37 #include <linux/ratelimit.h>
  38
  39 #include "rds.h"
  40 #include "ib.h"
  41
  42 /*
  43  * Convert IB-specific error message to RDS error message and call core
  44  * completion handler.
  45  */
  46 static void rds_ib_send_complete(struct rds_message *rm,
  47                                  int wc_status,
  48                                  void (*complete)(struct rds_message *rm, int status))
  49 {
  50         int notify_status;
  51
  52         switch (wc_status) {
  53         case IB_WC_WR_FLUSH_ERR:
  54                 return;
  55
  56         case IB_WC_SUCCESS:
  57                 notify_status = RDS_RDMA_SUCCESS;
  58                 break;
  59
  60         case IB_WC_REM_ACCESS_ERR:
  61                 notify_status = RDS_RDMA_REMOTE_ERROR;
  62                 break;
  63
  64         default:
  65                 notify_status = RDS_RDMA_OTHER_ERROR;
  66                 break;
  67         }
  68         complete(rm, notify_status);
  69 }
  70
  71 static void rds_ib_send_unmap_data(struct rds_ib_connection *ic,
  72                                    struct rm_data_op *op,
  73                                    int wc_status)
  74 {
  75         if (op->op_nents)
  76                 ib_dma_unmap_sg(ic->i_cm_id->device,
  77                                 op->op_sg, op->op_nents,
  78                                 DMA_TO_DEVICE);
  79 }
  80
  81 static void rds_ib_send_unmap_rdma(struct rds_ib_connection *ic,
  82                                    struct rm_rdma_op *op,
  83                                    int wc_status)
  84 {
  85         if (op->op_mapped) {
  86                 ib_dma_unmap_sg(ic->i_cm_id->device,
  87                                 op->op_sg, op->op_nents,
  88                                 op->op_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
  89                 op->op_mapped = 0;
  90         }
  91
  92         /* If the user asked for a completion notification on this
  93          * message, we can implement three different semantics:
  94          *  1.  Notify when we received the ACK on the RDS message
  95          *      that was queued with the RDMA. This provides reliable
  96          *      notification of RDMA status at the expense of a one-way
  97          *      packet delay.
  98          *  2.  Notify when the IB stack gives us the completion event for
  99          *      the RDMA operation.
 100          *  3.  Notify when the IB stack gives us the completion event for
 101          *      the accompanying RDS messages.
 102          * Here, we implement approach #3. To implement approach #2,
 103          * we would need to take an event for the rdma WR. To implement #1,
 104          * don't call rds_rdma_send_complete at all, and fall back to the notify
 105          * handling in the ACK processing code.
 106          *
 107          * Note: There's no need to explicitly sync any RDMA buffers using
 108          * ib_dma_sync_sg_for_cpu - the completion for the RDMA
 109          * operation itself unmapped the RDMA buffers, which takes care
 110          * of synching.
 111          */
 112         rds_ib_send_complete(container_of(op, struct rds_message, rdma),
 113                              wc_status, rds_rdma_send_complete);
 114
 115         if (op->op_write)
 116                 rds_stats_add(s_send_rdma_bytes, op->op_bytes);
 117         else
 118                 rds_stats_add(s_recv_rdma_bytes, op->op_bytes);
 119 }
 120
 121 static void rds_ib_send_unmap_atomic(struct rds_ib_connection *ic,
 122                                      struct rm_atomic_op *op,
 123                                      int wc_status)
 124 {
 125         /* unmap atomic recvbuf */
 126         if (op->op_mapped) {
 127                 ib_dma_unmap_sg(ic->i_cm_id->device, op->op_sg, 1,
 128                                 DMA_FROM_DEVICE);
 129                 op->op_mapped = 0;
 130         }
 131
 132         rds_ib_send_complete(container_of(op, struct rds_message, atomic),
 133                              wc_status, rds_atomic_send_complete);
 134
 135         if (op->op_type == RDS_ATOMIC_TYPE_CSWP)
 136                 rds_ib_stats_inc(s_ib_atomic_cswp);
 137         else
 138                 rds_ib_stats_inc(s_ib_atomic_fadd);
 139 }
 140
 141 /*
 142  * Unmap the resources associated with a struct send_work.
 143  *
 144  * Returns the rm for no good reason other than it is unobtainable
 145  * other than by switching on wr.opcode, currently, and the caller,
 146  * the event handler, needs it.
 147  */
 148 static struct rds_message *rds_ib_send_unmap_op(struct rds_ib_connection *ic,
 149                                                 struct rds_ib_send_work *send,
 150                                                 int wc_status)
 151 {
 152         struct rds_message *rm = NULL;
 153
 154         /* In the error case, wc.opcode sometimes contains garbage */
 155         switch (send->s_wr.opcode) {
 156         case IB_WR_SEND:
 157                 if (send->s_op) {
 158                         rm = container_of(send->s_op, struct rds_message, data);
 159                         rds_ib_send_unmap_data(ic, send->s_op, wc_status);
 160                 }
 161                 break;
 162         case IB_WR_RDMA_WRITE:
 163         case IB_WR_RDMA_READ:
 164                 if (send->s_op) {
 165                         rm = container_of(send->s_op, struct rds_message, rdma);
 166                         rds_ib_send_unmap_rdma(ic, send->s_op, wc_status);
 167                 }
 168                 break;
 169         case IB_WR_ATOMIC_FETCH_AND_ADD:
 170         case IB_WR_ATOMIC_CMP_AND_SWP:
 171                 if (send->s_op) {
 172                         rm = container_of(send->s_op, struct rds_message, atomic);
 173                         rds_ib_send_unmap_atomic(ic, send->s_op, wc_status);
 174                 }
 175                 break;
 176         default:
 177                 printk_ratelimited(KERN_NOTICE
 178                                "RDS/IB: %s: unexpected opcode 0x%x in WR!\n",
 179                                __func__, send->s_wr.opcode);
 180                 break;
 181         }
 182
 183         send->s_wr.opcode = 0xdead;
 184
 185         return rm;
 186 }
 187
 188 void rds_ib_send_init_ring(struct rds_ib_connection *ic)
 189 {
 190         struct rds_ib_send_work *send;
 191         u32 i;
 192
 193         for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
 194                 struct ib_sge *sge;
 195
 196                 send->s_op = NULL;
 197
 198                 send->s_wr.wr_id = i;
 199                 send->s_wr.sg_list = send->s_sge;
 200                 send->s_wr.ex.imm_data = 0;
 201
 202                 sge = &send->s_sge[0];
 203                 sge->addr = ic->i_send_hdrs_dma + (i * sizeof(struct rds_header));
 204                 sge->length = sizeof(struct rds_header);
 205                 sge->lkey = ic->i_pd->local_dma_lkey;
 206
 207                 send->s_sge[1].lkey = ic->i_pd->local_dma_lkey;
 208         }
 209 }
 210
 211 void rds_ib_send_clear_ring(struct rds_ib_connection *ic)
 212 {
 213         struct rds_ib_send_work *send;
 214         u32 i;
 215
 216         for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
 217                 if (send->s_op && send->s_wr.opcode != 0xdead)
 218                         rds_ib_send_unmap_op(ic, send, IB_WC_WR_FLUSH_ERR);
 219         }
 220 }
 221
 222 /*
 223  * The only fast path caller always has a non-zero nr, so we don't
 224  * bother testing nr before performing the atomic sub.
 225  */
 226 static void rds_ib_sub_signaled(struct rds_ib_connection *ic, int nr)
 227 {
 228         if ((atomic_sub_return(nr, &ic->i_signaled_sends) == 0) &&
 229             waitqueue_active(&rds_ib_ring_empty_wait))
 230                 wake_up(&rds_ib_ring_empty_wait);
 231         BUG_ON(atomic_read(&ic->i_signaled_sends) < 0);
 232 }
 233
 234 /*
 235  * The _oldest/_free ring operations here race cleanly with the alloc/unalloc
 236  * operations performed in the send path.  As the sender allocs and potentially
 237  * unallocs the next free entry in the ring it doesn't alter which is
 238  * the next to be freed, which is what this is concerned with.
 239  */
 240 void rds_ib_send_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc)
 241 {
 242         struct rds_message *rm = NULL;
 243         struct rds_connection *conn = ic->conn;
 244         struct rds_ib_send_work *send;
 245         u32 completed;
 246         u32 oldest;
 247         u32 i = 0;
 248         int nr_sig = 0;
 249
 250
 251         rdsdebug("wc wr_id 0x%llx status %u (%s) byte_len %u imm_data %u\n",
 252                  (unsigned long long)wc->wr_id, wc->status,
 253                  ib_wc_status_msg(wc->status), wc->byte_len,
 254                  be32_to_cpu(wc->ex.imm_data));
 255         rds_ib_stats_inc(s_ib_tx_cq_event);
 256
 257         if (wc->wr_id == RDS_IB_ACK_WR_ID) {
 258                 if (time_after(jiffies, ic->i_ack_queued + HZ / 2))
 259                         rds_ib_stats_inc(s_ib_tx_stalled);
 260                 rds_ib_ack_send_complete(ic);
 261                 return;
 262         }
 263
 264         oldest = rds_ib_ring_oldest(&ic->i_send_ring);
 265
 266         completed = rds_ib_ring_completed(&ic->i_send_ring, wc->wr_id, oldest);
 267
 268         for (i = 0; i < completed; i++) {
 269                 send = &ic->i_sends[oldest];
 270                 if (send->s_wr.send_flags & IB_SEND_SIGNALED)
 271                         nr_sig++;
 272
 273                 rm = rds_ib_send_unmap_op(ic, send, wc->status);
 274
 275                 if (time_after(jiffies, send->s_queued + HZ / 2))
 276                         rds_ib_stats_inc(s_ib_tx_stalled);
 277
 278                 if (send->s_op) {
 279                         if (send->s_op == rm->m_final_op) {
 280                                 /* If anyone waited for this message to get
 281                                  * flushed out, wake them up now
 282                                  */
 283                                 rds_message_unmapped(rm);
 284                         }
 285                         rds_message_put(rm);
 286                         send->s_op = NULL;
 287                 }
 288
 289                 oldest = (oldest + 1) % ic->i_send_ring.w_nr;
 290         }
 291
 292         rds_ib_ring_free(&ic->i_send_ring, completed);
 293         rds_ib_sub_signaled(ic, nr_sig);
 294         nr_sig = 0;
 295
 296         if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
 297             test_bit(0, &conn->c_map_queued))
 298                 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
 299
 300         /* We expect errors as the qp is drained during shutdown */
 301         if (wc->status != IB_WC_SUCCESS && rds_conn_up(conn)) {
 302                 rds_ib_conn_error(conn, "send completion on %pI4 had status %u (%s), disconnecting and reconnecting\n",
 303                                   &conn->c_faddr, wc->status,
 304                                   ib_wc_status_msg(wc->status));
 305         }
 306 }
 307
 308 /*
 309  * This is the main function for allocating credits when sending
 310  * messages.
 311  *
 312  * Conceptually, we have two counters:
 313  *  -   send credits: this tells us how many WRs we're allowed
 314  *      to submit without overruning the receiver's queue. For
 315  *      each SEND WR we post, we decrement this by one.
 316  *
 317  *  -   posted credits: this tells us how many WRs we recently
 318  *      posted to the receive queue. This value is transferred
 319  *      to the peer as a "credit update" in a RDS header field.
 320  *      Every time we transmit credits to the peer, we subtract
 321  *      the amount of transferred credits from this counter.
 322  *
 323  * It is essential that we avoid situations where both sides have
 324  * exhausted their send credits, and are unable to send new credits
 325  * to the peer. We achieve this by requiring that we send at least
 326  * one credit update to the peer before exhausting our credits.
 327  * When new credits arrive, we subtract one credit that is withheld
 328  * until we've posted new buffers and are ready to transmit these
 329  * credits (see rds_ib_send_add_credits below).
 330  *
 331  * The RDS send code is essentially single-threaded; rds_send_xmit
 332  * sets RDS_IN_XMIT to ensure exclusive access to the send ring.
 333  * However, the ACK sending code is independent and can race with
 334  * message SENDs.
 335  *
 336  * In the send path, we need to update the counters for send credits
 337  * and the counter of posted buffers atomically - when we use the
 338  * last available credit, we cannot allow another thread to race us
 339  * and grab the posted credits counter.  Hence, we have to use a
 340  * spinlock to protect the credit counter, or use atomics.
 341  *
 342  * Spinlocks shared between the send and the receive path are bad,
 343  * because they create unnecessary delays. An early implementation
 344  * using a spinlock showed a 5% degradation in throughput at some
 345  * loads.
 346  *
 347  * This implementation avoids spinlocks completely, putting both
 348  * counters into a single atomic, and updating that atomic using
 349  * atomic_add (in the receive path, when receiving fresh credits),
 350  * and using atomic_cmpxchg when updating the two counters.
 351  */
 352 int rds_ib_send_grab_credits(struct rds_ib_connection *ic,
 353                              u32 wanted, u32 *adv_credits, int need_posted, int max_posted)
 354 {
 355         unsigned int avail, posted, got = 0, advertise;
 356         long oldval, newval;
 357
 358         *adv_credits = 0;
 359         if (!ic->i_flowctl)
 360                 return wanted;
 361
 362 try_again:
 363         advertise = 0;
 364         oldval = newval = atomic_read(&ic->i_credits);
 365         posted = IB_GET_POST_CREDITS(oldval);
 366         avail = IB_GET_SEND_CREDITS(oldval);
 367
 368         rdsdebug("wanted=%u credits=%u posted=%u\n",
 369                         wanted, avail, posted);
 370
 371         /* The last credit must be used to send a credit update. */
 372         if (avail && !posted)
 373                 avail--;
 374
 375         if (avail < wanted) {
 376                 struct rds_connection *conn = ic->i_cm_id->context;
 377
 378                 /* Oops, there aren't that many credits left! */
 379                 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 380                 got = avail;
 381         } else {
 382                 /* Sometimes you get what you want, lalala. */
 383                 got = wanted;
 384         }
 385         newval -= IB_SET_SEND_CREDITS(got);
 386
 387         /*
 388          * If need_posted is non-zero, then the caller wants
 389          * the posted regardless of whether any send credits are
 390          * available.
 391          */
 392         if (posted && (got || need_posted)) {
 393                 advertise = min_t(unsigned int, posted, max_posted);
 394                 newval -= IB_SET_POST_CREDITS(advertise);
 395         }
 396
 397         /* Finally bill everything */
 398         if (atomic_cmpxchg(&ic->i_credits, oldval, newval) != oldval)
 399                 goto try_again;
 400
 401         *adv_credits = advertise;
 402         return got;
 403 }
 404
 405 void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits)
 406 {
 407         struct rds_ib_connection *ic = conn->c_transport_data;
 408
 409         if (credits == 0)
 410                 return;
 411
 412         rdsdebug("credits=%u current=%u%s\n",
 413                         credits,
 414                         IB_GET_SEND_CREDITS(atomic_read(&ic->i_credits)),
 415                         test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ? ", ll_send_full" : "");
 416
 417         atomic_add(IB_SET_SEND_CREDITS(credits), &ic->i_credits);
 418         if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags))
 419                 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
 420
 421         WARN_ON(IB_GET_SEND_CREDITS(credits) >= 16384);
 422
 423         rds_ib_stats_inc(s_ib_rx_credit_updates);
 424 }
 425
 426 void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted)
 427 {
 428         struct rds_ib_connection *ic = conn->c_transport_data;
 429
 430         if (posted == 0)
 431                 return;
 432
 433         atomic_add(IB_SET_POST_CREDITS(posted), &ic->i_credits);
 434
 435         /* Decide whether to send an update to the peer now.
 436          * If we would send a credit update for every single buffer we
 437          * post, we would end up with an ACK storm (ACK arrives,
 438          * consumes buffer, we refill the ring, send ACK to remote
 439          * advertising the newly posted buffer... ad inf)
 440          *
 441          * Performance pretty much depends on how often we send
 442          * credit updates - too frequent updates mean lots of ACKs.
 443          * Too infrequent updates, and the peer will run out of
 444          * credits and has to throttle.
 445          * For the time being, 16 seems to be a good compromise.
 446          */
 447         if (IB_GET_POST_CREDITS(atomic_read(&ic->i_credits)) >= 16)
 448                 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
 449 }
 450
 451 static inline int rds_ib_set_wr_signal_state(struct rds_ib_connection *ic,
 452                                              struct rds_ib_send_work *send,
 453                                              bool notify)
 454 {
 455         /*
 456          * We want to delay signaling completions just enough to get
 457          * the batching benefits but not so much that we create dead time
 458          * on the wire.
 459          */
 460         if (ic->i_unsignaled_wrs-- == 0 || notify) {
 461                 ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
 462                 send->s_wr.send_flags |= IB_SEND_SIGNALED;
 463                 return 1;
 464         }
 465         return 0;
 466 }
 467
 468 /*
 469  * This can be called multiple times for a given message.  The first time
 470  * we see a message we map its scatterlist into the IB device so that
 471  * we can provide that mapped address to the IB scatter gather entries
 472  * in the IB work requests.  We translate the scatterlist into a series
 473  * of work requests that fragment the message.  These work requests complete
 474  * in order so we pass ownership of the message to the completion handler
 475  * once we send the final fragment.
 476  *
 477  * The RDS core uses the c_send_lock to only enter this function once
 478  * per connection.  This makes sure that the tx ring alloc/unalloc pairs
 479  * don't get out of sync and confuse the ring.
 480  */
 481 int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
 482                 unsigned int hdr_off, unsigned int sg, unsigned int off)
 483 {
 484         struct rds_ib_connection *ic = conn->c_transport_data;
 485         struct ib_device *dev = ic->i_cm_id->device;
 486         struct rds_ib_send_work *send = NULL;
 487         struct rds_ib_send_work *first;
 488         struct rds_ib_send_work *prev;
 489         struct ib_send_wr *failed_wr;
 490         struct scatterlist *scat;
 491         u32 pos;
 492         u32 i;
 493         u32 work_alloc;
 494         u32 credit_alloc = 0;
 495         u32 posted;
 496         u32 adv_credits = 0;
 497         int send_flags = 0;
 498         int bytes_sent = 0;
 499         int ret;
 500         int flow_controlled = 0;
 501         int nr_sig = 0;
 502
 503         BUG_ON(off % RDS_FRAG_SIZE);
 504         BUG_ON(hdr_off != 0 && hdr_off != sizeof(struct rds_header));
 505
 506         /* Do not send cong updates to IB loopback */
 507         if (conn->c_loopback
 508             && rm->m_inc.i_hdr.h_flags & RDS_FLAG_CONG_BITMAP) {
 509                 rds_cong_map_updated(conn->c_fcong, ~(u64) 0);
 510                 scat = &rm->data.op_sg[sg];
 511                 ret = max_t(int, RDS_CONG_MAP_BYTES, scat->length);
 512                 return sizeof(struct rds_header) + ret;
 513         }
 514
 515         /* FIXME we may overallocate here */
 516         if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0)
 517                 i = 1;
 518         else
 519                 i = ceil(be32_to_cpu(rm->m_inc.i_hdr.h_len), RDS_FRAG_SIZE);
 520
 521         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
 522         if (work_alloc == 0) {
 523                 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 524                 rds_ib_stats_inc(s_ib_tx_ring_full);
 525                 ret = -ENOMEM;
 526                 goto out;
 527         }
 528
 529         if (ic->i_flowctl) {
 530                 credit_alloc = rds_ib_send_grab_credits(ic, work_alloc, &posted, 0, RDS_MAX_ADV_CREDIT);
 531                 adv_credits += posted;
 532                 if (credit_alloc < work_alloc) {
 533                         rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - credit_alloc);
 534                         work_alloc = credit_alloc;
 535                         flow_controlled = 1;
 536                 }
 537                 if (work_alloc == 0) {
 538                         set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
 539                         rds_ib_stats_inc(s_ib_tx_throttle);
 540                         ret = -ENOMEM;
 541                         goto out;
 542                 }
 543         }
 544
 545         /* map the message the first time we see it */
 546         if (!ic->i_data_op) {
 547                 if (rm->data.op_nents) {
 548                         rm->data.op_count = ib_dma_map_sg(dev,
 549                                                           rm->data.op_sg,
 550                                                           rm->data.op_nents,
 551                                                           DMA_TO_DEVICE);
 552                         rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->data.op_count);
 553                         if (rm->data.op_count == 0) {
 554                                 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
 555                                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 556                                 ret = -ENOMEM; /* XXX ? */
 557                                 goto out;
 558                         }
 559                 } else {
 560                         rm->data.op_count = 0;
 561                 }
 562
 563                 rds_message_addref(rm);
 564                 rm->data.op_dmasg = 0;
 565                 rm->data.op_dmaoff = 0;
 566                 ic->i_data_op = &rm->data;
 567
 568                 /* Finalize the header */
 569                 if (test_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags))
 570                         rm->m_inc.i_hdr.h_flags |= RDS_FLAG_ACK_REQUIRED;
 571                 if (test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))
 572                         rm->m_inc.i_hdr.h_flags |= RDS_FLAG_RETRANSMITTED;
 573
 574                 /* If it has a RDMA op, tell the peer we did it. This is
 575                  * used by the peer to release use-once RDMA MRs. */
 576                 if (rm->rdma.op_active) {
 577                         struct rds_ext_header_rdma ext_hdr;
 578
 579                         ext_hdr.h_rdma_rkey = cpu_to_be32(rm->rdma.op_rkey);
 580                         rds_message_add_extension(&rm->m_inc.i_hdr,
 581                                         RDS_EXTHDR_RDMA, &ext_hdr, sizeof(ext_hdr));
 582                 }
 583                 if (rm->m_rdma_cookie) {
 584                         rds_message_add_rdma_dest_extension(&rm->m_inc.i_hdr,
 585                                         rds_rdma_cookie_key(rm->m_rdma_cookie),
 586                                         rds_rdma_cookie_offset(rm->m_rdma_cookie));
 587                 }
 588
 589                 /* Note - rds_ib_piggyb_ack clears the ACK_REQUIRED bit, so
 590                  * we should not do this unless we have a chance of at least
 591                  * sticking the header into the send ring. Which is why we
 592                  * should call rds_ib_ring_alloc first. */
 593                 rm->m_inc.i_hdr.h_ack = cpu_to_be64(rds_ib_piggyb_ack(ic));
 594                 rds_message_make_checksum(&rm->m_inc.i_hdr);
 595
 596                 /*
 597                  * Update adv_credits since we reset the ACK_REQUIRED bit.
 598                  */
 599                 if (ic->i_flowctl) {
 600                         rds_ib_send_grab_credits(ic, 0, &posted, 1, RDS_MAX_ADV_CREDIT - adv_credits);
 601                         adv_credits += posted;
 602                         BUG_ON(adv_credits > 255);
 603                 }
 604         }
 605
 606         /* Sometimes you want to put a fence between an RDMA
 607          * READ and the following SEND.
 608          * We could either do this all the time
 609          * or when requested by the user. Right now, we let
 610          * the application choose.
 611          */
 612         if (rm->rdma.op_active && rm->rdma.op_fence)
 613                 send_flags = IB_SEND_FENCE;
 614
 615         /* Each frag gets a header. Msgs may be 0 bytes */
 616         send = &ic->i_sends[pos];
 617         first = send;
 618         prev = NULL;
 619         scat = &ic->i_data_op->op_sg[rm->data.op_dmasg];
 620         i = 0;
 621         do {
 622                 unsigned int len = 0;
 623
 624                 /* Set up the header */
 625                 send->s_wr.send_flags = send_flags;
 626                 send->s_wr.opcode = IB_WR_SEND;
 627                 send->s_wr.num_sge = 1;
 628                 send->s_wr.next = NULL;
 629                 send->s_queued = jiffies;
 630                 send->s_op = NULL;
 631
 632                 send->s_sge[0].addr = ic->i_send_hdrs_dma
 633                         + (pos * sizeof(struct rds_header));
 634                 send->s_sge[0].length = sizeof(struct rds_header);
 635
 636                 memcpy(&ic->i_send_hdrs[pos], &rm->m_inc.i_hdr, sizeof(struct rds_header));
 637
 638                 /* Set up the data, if present */
 639                 if (i < work_alloc
 640                     && scat != &rm->data.op_sg[rm->data.op_count]) {
 641                         len = min(RDS_FRAG_SIZE,
 642                                 ib_sg_dma_len(dev, scat) - rm->data.op_dmaoff);
 643                         send->s_wr.num_sge = 2;
 644
 645                         send->s_sge[1].addr = ib_sg_dma_address(dev, scat);
 646                         send->s_sge[1].addr += rm->data.op_dmaoff;
 647                         send->s_sge[1].length = len;
 648
 649                         bytes_sent += len;
 650                         rm->data.op_dmaoff += len;
 651                         if (rm->data.op_dmaoff == ib_sg_dma_len(dev, scat)) {
 652                                 scat++;
 653                                 rm->data.op_dmasg++;
 654                                 rm->data.op_dmaoff = 0;
 655                         }
 656                 }
 657
 658                 rds_ib_set_wr_signal_state(ic, send, 0);
 659
 660                 /*
 661                  * Always signal the last one if we're stopping due to flow control.
 662                  */
 663                 if (ic->i_flowctl && flow_controlled && i == (work_alloc-1))
 664                         send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
 665
 666                 if (send->s_wr.send_flags & IB_SEND_SIGNALED)
 667                         nr_sig++;
 668
 669                 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
 670                          &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
 671
 672                 if (ic->i_flowctl && adv_credits) {
 673                         struct rds_header *hdr = &ic->i_send_hdrs[pos];
 674
 675                         /* add credit and redo the header checksum */
 676                         hdr->h_credit = adv_credits;
 677                         rds_message_make_checksum(hdr);
 678                         adv_credits = 0;
 679                         rds_ib_stats_inc(s_ib_tx_credit_updates);
 680                 }
 681
 682                 if (prev)
 683                         prev->s_wr.next = &send->s_wr;
 684                 prev = send;
 685
 686                 pos = (pos + 1) % ic->i_send_ring.w_nr;
 687                 send = &ic->i_sends[pos];
 688                 i++;
 689
 690         } while (i < work_alloc
 691                  && scat != &rm->data.op_sg[rm->data.op_count]);
 692
 693         /* Account the RDS header in the number of bytes we sent, but just once.
 694          * The caller has no concept of fragmentation. */
 695         if (hdr_off == 0)
 696                 bytes_sent += sizeof(struct rds_header);
 697
 698         /* if we finished the message then send completion owns it */
 699         if (scat == &rm->data.op_sg[rm->data.op_count]) {
 700                 prev->s_op = ic->i_data_op;
 701                 prev->s_wr.send_flags |= IB_SEND_SOLICITED;
 702                 if (!(prev->s_wr.send_flags & IB_SEND_SIGNALED)) {
 703                         ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
 704                         prev->s_wr.send_flags |= IB_SEND_SIGNALED;
 705                         nr_sig++;
 706                 }
 707                 ic->i_data_op = NULL;
 708         }
 709
 710         /* Put back wrs & credits we didn't use */
 711         if (i < work_alloc) {
 712                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
 713                 work_alloc = i;
 714         }
 715         if (ic->i_flowctl && i < credit_alloc)
 716                 rds_ib_send_add_credits(conn, credit_alloc - i);
 717
 718         if (nr_sig)
 719                 atomic_add(nr_sig, &ic->i_signaled_sends);
 720
 721         /* XXX need to worry about failed_wr and partial sends. */
 722         failed_wr = &first->s_wr;
 723         ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
 724         rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
 725                  first, &first->s_wr, ret, failed_wr);
 726         BUG_ON(failed_wr != &first->s_wr);
 727         if (ret) {
 728                 printk(KERN_WARNING "RDS/IB: ib_post_send to %pI4 "
 729                        "returned %d\n", &conn->c_faddr, ret);
 730                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 731                 rds_ib_sub_signaled(ic, nr_sig);
 732                 if (prev->s_op) {
 733                         ic->i_data_op = prev->s_op;
 734                         prev->s_op = NULL;
 735                 }
 736
 737                 rds_ib_conn_error(ic->conn, "ib_post_send failed\n");
 738                 goto out;
 739         }
 740
 741         ret = bytes_sent;
 742 out:
 743         BUG_ON(adv_credits);
 744         return ret;
 745 }
 746
 747 /*
 748  * Issue atomic operation.
 749  * A simplified version of the rdma case, we always map 1 SG, and
 750  * only 8 bytes, for the return value from the atomic operation.
 751  */
 752 int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op)
 753 {
 754         struct rds_ib_connection *ic = conn->c_transport_data;
 755         struct rds_ib_send_work *send = NULL;
 756         struct ib_send_wr *failed_wr;
 757         struct rds_ib_device *rds_ibdev;
 758         u32 pos;
 759         u32 work_alloc;
 760         int ret;
 761         int nr_sig = 0;
 762
 763         rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
 764
 765         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, 1, &pos);
 766         if (work_alloc != 1) {
 767                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 768                 rds_ib_stats_inc(s_ib_tx_ring_full);
 769                 ret = -ENOMEM;
 770                 goto out;
 771         }
 772
 773         /* address of send request in ring */
 774         send = &ic->i_sends[pos];
 775         send->s_queued = jiffies;
 776
 777         if (op->op_type == RDS_ATOMIC_TYPE_CSWP) {
 778                 send->s_atomic_wr.wr.opcode = IB_WR_MASKED_ATOMIC_CMP_AND_SWP;
 779                 send->s_atomic_wr.compare_add = op->op_m_cswp.compare;
 780                 send->s_atomic_wr.swap = op->op_m_cswp.swap;
 781                 send->s_atomic_wr.compare_add_mask = op->op_m_cswp.compare_mask;
 782                 send->s_atomic_wr.swap_mask = op->op_m_cswp.swap_mask;
 783         } else { /* FADD */
 784                 send->s_atomic_wr.wr.opcode = IB_WR_MASKED_ATOMIC_FETCH_AND_ADD;
 785                 send->s_atomic_wr.compare_add = op->op_m_fadd.add;
 786                 send->s_atomic_wr.swap = 0;
 787                 send->s_atomic_wr.compare_add_mask = op->op_m_fadd.nocarry_mask;
 788                 send->s_atomic_wr.swap_mask = 0;
 789         }
 790         nr_sig = rds_ib_set_wr_signal_state(ic, send, op->op_notify);
 791         send->s_atomic_wr.wr.num_sge = 1;
 792         send->s_atomic_wr.wr.next = NULL;
 793         send->s_atomic_wr.remote_addr = op->op_remote_addr;
 794         send->s_atomic_wr.rkey = op->op_rkey;
 795         send->s_op = op;
 796         rds_message_addref(container_of(send->s_op, struct rds_message, atomic));
 797
 798         /* map 8 byte retval buffer to the device */
 799         ret = ib_dma_map_sg(ic->i_cm_id->device, op->op_sg, 1, DMA_FROM_DEVICE);
 800         rdsdebug("ic %p mapping atomic op %p. mapped %d pg\n", ic, op, ret);
 801         if (ret != 1) {
 802                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 803                 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
 804                 ret = -ENOMEM; /* XXX ? */
 805                 goto out;
 806         }
 807
 808         /* Convert our struct scatterlist to struct ib_sge */
 809         send->s_sge[0].addr = ib_sg_dma_address(ic->i_cm_id->device, op->op_sg);
 810         send->s_sge[0].length = ib_sg_dma_len(ic->i_cm_id->device, op->op_sg);
 811         send->s_sge[0].lkey = ic->i_pd->local_dma_lkey;
 812
 813         rdsdebug("rva %Lx rpa %Lx len %u\n", op->op_remote_addr,
 814                  send->s_sge[0].addr, send->s_sge[0].length);
 815
 816         if (nr_sig)
 817                 atomic_add(nr_sig, &ic->i_signaled_sends);
 818
 819         failed_wr = &send->s_atomic_wr.wr;
 820         ret = ib_post_send(ic->i_cm_id->qp, &send->s_atomic_wr.wr, &failed_wr);
 821         rdsdebug("ic %p send %p (wr %p) ret %d wr %p\n", ic,
 822                  send, &send->s_atomic_wr, ret, failed_wr);
 823         BUG_ON(failed_wr != &send->s_atomic_wr.wr);
 824         if (ret) {
 825                 printk(KERN_WARNING "RDS/IB: atomic ib_post_send to %pI4 "
 826                        "returned %d\n", &conn->c_faddr, ret);
 827                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 828                 rds_ib_sub_signaled(ic, nr_sig);
 829                 goto out;
 830         }
 831
 832         if (unlikely(failed_wr != &send->s_atomic_wr.wr)) {
 833                 printk(KERN_WARNING "RDS/IB: atomic ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
 834                 BUG_ON(failed_wr != &send->s_atomic_wr.wr);
 835         }
 836
 837 out:
 838         return ret;
 839 }
 840
 841 int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op)
 842 {
 843         struct rds_ib_connection *ic = conn->c_transport_data;
 844         struct rds_ib_send_work *send = NULL;
 845         struct rds_ib_send_work *first;
 846         struct rds_ib_send_work *prev;
 847         struct ib_send_wr *failed_wr;
 848         struct scatterlist *scat;
 849         unsigned long len;
 850         u64 remote_addr = op->op_remote_addr;
 851         u32 max_sge = ic->rds_ibdev->max_sge;
 852         u32 pos;
 853         u32 work_alloc;
 854         u32 i;
 855         u32 j;
 856         int sent;
 857         int ret;
 858         int num_sge;
 859         int nr_sig = 0;
 860
 861         /* map the op the first time we see it */
 862         if (!op->op_mapped) {
 863                 op->op_count = ib_dma_map_sg(ic->i_cm_id->device,
 864                                              op->op_sg, op->op_nents, (op->op_write) ?
 865                                              DMA_TO_DEVICE : DMA_FROM_DEVICE);
 866                 rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->op_count);
 867                 if (op->op_count == 0) {
 868                         rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
 869                         ret = -ENOMEM; /* XXX ? */
 870                         goto out;
 871                 }
 872
 873                 op->op_mapped = 1;
 874         }
 875
 876         /*
 877          * Instead of knowing how to return a partial rdma read/write we insist that there
 878          * be enough work requests to send the entire message.
 879          */
 880         i = ceil(op->op_count, max_sge);
 881
 882         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
 883         if (work_alloc != i) {
 884                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 885                 rds_ib_stats_inc(s_ib_tx_ring_full);
 886                 ret = -ENOMEM;
 887                 goto out;
 888         }
 889
 890         send = &ic->i_sends[pos];
 891         first = send;
 892         prev = NULL;
 893         scat = &op->op_sg[0];
 894         sent = 0;
 895         num_sge = op->op_count;
 896
 897         for (i = 0; i < work_alloc && scat != &op->op_sg[op->op_count]; i++) {
 898                 send->s_wr.send_flags = 0;
 899                 send->s_queued = jiffies;
 900                 send->s_op = NULL;
 901
 902                 nr_sig += rds_ib_set_wr_signal_state(ic, send, 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 = ib_sg_dma_len(ic->i_cm_id->device, scat);
 923                         send->s_sge[j].addr =
 924                                  ib_sg_dma_address(ic->i_cm_id->device, scat);
 925                         send->s_sge[j].length = len;
 926                         send->s_sge[j].lkey = ic->i_pd->local_dma_lkey;
 927
 928                         sent += len;
 929                         rdsdebug("ic %p sent %d remote_addr %llu\n", ic, sent, remote_addr);
 930
 931                         remote_addr += len;
 932                         scat++;
 933                 }
 934
 935                 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
 936                         &send->s_rdma_wr.wr,
 937                         send->s_rdma_wr.wr.num_sge,
 938                         send->s_rdma_wr.wr.next);
 939
 940                 prev = send;
 941                 if (++send == &ic->i_sends[ic->i_send_ring.w_nr])
 942                         send = ic->i_sends;
 943         }
 944
 945         /* give a reference to the last op */
 946         if (scat == &op->op_sg[op->op_count]) {
 947                 prev->s_op = op;
 948                 rds_message_addref(container_of(op, struct rds_message, rdma));
 949         }
 950
 951         if (i < work_alloc) {
 952                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
 953                 work_alloc = i;
 954         }
 955
 956         if (nr_sig)
 957                 atomic_add(nr_sig, &ic->i_signaled_sends);
 958
 959         failed_wr = &first->s_rdma_wr.wr;
 960         ret = ib_post_send(ic->i_cm_id->qp, &first->s_rdma_wr.wr, &failed_wr);
 961         rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
 962                  first, &first->s_rdma_wr.wr, ret, failed_wr);
 963         BUG_ON(failed_wr != &first->s_rdma_wr.wr);
 964         if (ret) {
 965                 printk(KERN_WARNING "RDS/IB: rdma ib_post_send to %pI4 "
 966                        "returned %d\n", &conn->c_faddr, ret);
 967                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 968                 rds_ib_sub_signaled(ic, nr_sig);
 969                 goto out;
 970         }
 971
 972         if (unlikely(failed_wr != &first->s_rdma_wr.wr)) {
 973                 printk(KERN_WARNING "RDS/IB: ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
 974                 BUG_ON(failed_wr != &first->s_rdma_wr.wr);
 975         }
 976
 977
 978 out:
 979         return ret;
 980 }
 981
 982 void rds_ib_xmit_complete(struct rds_connection *conn)
 983 {
 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 }