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