net/rds/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/moduleparam.h>
  35 #include <linux/gfp.h>
  36 #include <net/sock.h>
  37 #include <linux/in.h>
  38 #include <linux/list.h>
  39 #include <linux/ratelimit.h>
  40 #include <linux/export.h>
  41 #include <linux/sizes.h>
  42
  43 #include "rds.h"
  44
  45 /* When transmitting messages in rds_send_xmit, we need to emerge from
  46  * time to time and briefly release the CPU. Otherwise the softlock watchdog
  47  * will kick our shin.
  48  * Also, it seems fairer to not let one busy connection stall all the
  49  * others.
  50  *
  51  * send_batch_count is the number of times we'll loop in send_xmit. Setting
  52  * it to 0 will restore the old behavior (where we looped until we had
  53  * drained the queue).
  54  */
  55 static int send_batch_count = SZ_1K;
  56 module_param(send_batch_count, int, 0444);
  57 MODULE_PARM_DESC(send_batch_count, " batch factor when working the send queue");
  58
  59 static void rds_send_remove_from_sock(struct list_head *messages, int status);
  60
  61 /*
  62  * Reset the send state.  Callers must ensure that this doesn't race with
  63  * rds_send_xmit().
  64  */
  65 void rds_send_path_reset(struct rds_conn_path *cp)
  66 {
  67         struct rds_message *rm, *tmp;
  68         unsigned long flags;
  69
  70         if (cp->cp_xmit_rm) {
  71                 rm = cp->cp_xmit_rm;
  72                 cp->cp_xmit_rm = NULL;
  73                 /* Tell the user the RDMA op is no longer mapped by the
  74                  * transport. This isn't entirely true (it's flushed out
  75                  * independently) but as the connection is down, there's
  76                  * no ongoing RDMA to/from that memory */
  77                 rds_message_unmapped(rm);
  78                 rds_message_put(rm);
  79         }
  80
  81         cp->cp_xmit_sg = 0;
  82         cp->cp_xmit_hdr_off = 0;
  83         cp->cp_xmit_data_off = 0;
  84         cp->cp_xmit_atomic_sent = 0;
  85         cp->cp_xmit_rdma_sent = 0;
  86         cp->cp_xmit_data_sent = 0;
  87
  88         cp->cp_conn->c_map_queued = 0;
  89
  90         cp->cp_unacked_packets = rds_sysctl_max_unacked_packets;
  91         cp->cp_unacked_bytes = rds_sysctl_max_unacked_bytes;
  92
  93         /* Mark messages as retransmissions, and move them to the send q */
  94         spin_lock_irqsave(&cp->cp_lock, flags);
  95         list_for_each_entry_safe(rm, tmp, &cp->cp_retrans, m_conn_item) {
  96                 set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
  97                 set_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags);
  98         }
  99         list_splice_init(&cp->cp_retrans, &cp->cp_send_queue);
 100         spin_unlock_irqrestore(&cp->cp_lock, flags);
 101 }
 102 EXPORT_SYMBOL_GPL(rds_send_path_reset);
 103
 104 static int acquire_in_xmit(struct rds_conn_path *cp)
 105 {
 106         return test_and_set_bit(RDS_IN_XMIT, &cp->cp_flags) == 0;
 107 }
 108
 109 static void release_in_xmit(struct rds_conn_path *cp)
 110 {
 111         clear_bit(RDS_IN_XMIT, &cp->cp_flags);
 112         smp_mb__after_atomic();
 113         /*
 114          * We don't use wait_on_bit()/wake_up_bit() because our waking is in a
 115          * hot path and finding waiters is very rare.  We don't want to walk
 116          * the system-wide hashed waitqueue buckets in the fast path only to
 117          * almost never find waiters.
 118          */
 119         if (waitqueue_active(&cp->cp_waitq))
 120                 wake_up_all(&cp->cp_waitq);
 121 }
 122
 123 /*
 124  * We're making the conscious trade-off here to only send one message
 125  * down the connection at a time.
 126  *   Pro:
 127  *      - tx queueing is a simple fifo list
 128  *      - reassembly is optional and easily done by transports per conn
 129  *      - no per flow rx lookup at all, straight to the socket
 130  *      - less per-frag memory and wire overhead
 131  *   Con:
 132  *      - queued acks can be delayed behind large messages
 133  *   Depends:
 134  *      - small message latency is higher behind queued large messages
 135  *      - large message latency isn't starved by intervening small sends
 136  */
 137 int rds_send_xmit(struct rds_conn_path *cp)
 138 {
 139         struct rds_connection *conn = cp->cp_conn;
 140         struct rds_message *rm;
 141         unsigned long flags;
 142         unsigned int tmp;
 143         struct scatterlist *sg;
 144         int ret = 0;
 145         LIST_HEAD(to_be_dropped);
 146         int batch_count;
 147         unsigned long send_gen = 0;
 148
 149 restart:
 150         batch_count = 0;
 151
 152         /*
 153          * sendmsg calls here after having queued its message on the send
 154          * queue.  We only have one task feeding the connection at a time.  If
 155          * another thread is already feeding the queue then we back off.  This
 156          * avoids blocking the caller and trading per-connection data between
 157          * caches per message.
 158          */
 159         if (!acquire_in_xmit(cp)) {
 160                 rds_stats_inc(s_send_lock_contention);
 161                 ret = -ENOMEM;
 162                 goto out;
 163         }
 164
 165         if (test_bit(RDS_DESTROY_PENDING, &cp->cp_flags)) {
 166                 release_in_xmit(cp);
 167                 ret = -ENETUNREACH; /* dont requeue send work */
 168                 goto out;
 169         }
 170
 171         /*
 172          * we record the send generation after doing the xmit acquire.
 173          * if someone else manages to jump in and do some work, we'll use
 174          * this to avoid a goto restart farther down.
 175          *
 176          * The acquire_in_xmit() check above ensures that only one
 177          * caller can increment c_send_gen at any time.
 178          */
 179         send_gen = READ_ONCE(cp->cp_send_gen) + 1;
 180         WRITE_ONCE(cp->cp_send_gen, send_gen);
 181
 182         /*
 183          * rds_conn_shutdown() sets the conn state and then tests RDS_IN_XMIT,
 184          * we do the opposite to avoid races.
 185          */
 186         if (!rds_conn_path_up(cp)) {
 187                 release_in_xmit(cp);
 188                 ret = 0;
 189                 goto out;
 190         }
 191
 192         if (conn->c_trans->xmit_path_prepare)
 193                 conn->c_trans->xmit_path_prepare(cp);
 194
 195         /*
 196          * spin trying to push headers and data down the connection until
 197          * the connection doesn't make forward progress.
 198          */
 199         while (1) {
 200
 201                 rm = cp->cp_xmit_rm;
 202
 203                 /*
 204                  * If between sending messages, we can send a pending congestion
 205                  * map update.
 206                  */
 207                 if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) {
 208                         rm = rds_cong_update_alloc(conn);
 209                         if (IS_ERR(rm)) {
 210                                 ret = PTR_ERR(rm);
 211                                 break;
 212                         }
 213                         rm->data.op_active = 1;
 214                         rm->m_inc.i_conn_path = cp;
 215                         rm->m_inc.i_conn = cp->cp_conn;
 216
 217                         cp->cp_xmit_rm = rm;
 218                 }
 219
 220                 /*
 221                  * If not already working on one, grab the next message.
 222                  *
 223                  * cp_xmit_rm holds a ref while we're sending this message down
 224                  * the connction.  We can use this ref while holding the
 225                  * send_sem.. rds_send_reset() is serialized with it.
 226                  */
 227                 if (!rm) {
 228                         unsigned int len;
 229
 230                         batch_count++;
 231
 232                         /* we want to process as big a batch as we can, but
 233                          * we also want to avoid softlockups.  If we've been
 234                          * through a lot of messages, lets back off and see
 235                          * if anyone else jumps in
 236                          */
 237                         if (batch_count >= send_batch_count)
 238                                 goto over_batch;
 239
 240                         spin_lock_irqsave(&cp->cp_lock, flags);
 241
 242                         if (!list_empty(&cp->cp_send_queue)) {
 243                                 rm = list_entry(cp->cp_send_queue.next,
 244                                                 struct rds_message,
 245                                                 m_conn_item);
 246                                 rds_message_addref(rm);
 247
 248                                 /*
 249                                  * Move the message from the send queue to the retransmit
 250                                  * list right away.
 251                                  */
 252                                 list_move_tail(&rm->m_conn_item,
 253                                                &cp->cp_retrans);
 254                         }
 255
 256                         spin_unlock_irqrestore(&cp->cp_lock, flags);
 257
 258                         if (!rm)
 259                                 break;
 260
 261                         /* Unfortunately, the way Infiniband deals with
 262                          * RDMA to a bad MR key is by moving the entire
 263                          * queue pair to error state. We cold possibly
 264                          * recover from that, but right now we drop the
 265                          * connection.
 266                          * Therefore, we never retransmit messages with RDMA ops.
 267                          */
 268                         if (test_bit(RDS_MSG_FLUSH, &rm->m_flags) ||
 269                             (rm->rdma.op_active &&
 270                             test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))) {
 271                                 spin_lock_irqsave(&cp->cp_lock, flags);
 272                                 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags))
 273                                         list_move(&rm->m_conn_item, &to_be_dropped);
 274                                 spin_unlock_irqrestore(&cp->cp_lock, flags);
 275                                 continue;
 276                         }
 277
 278                         /* Require an ACK every once in a while */
 279                         len = ntohl(rm->m_inc.i_hdr.h_len);
 280                         if (cp->cp_unacked_packets == 0 ||
 281                             cp->cp_unacked_bytes < len) {
 282                                 set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
 283
 284                                 cp->cp_unacked_packets =
 285                                         rds_sysctl_max_unacked_packets;
 286                                 cp->cp_unacked_bytes =
 287                                         rds_sysctl_max_unacked_bytes;
 288                                 rds_stats_inc(s_send_ack_required);
 289                         } else {
 290                                 cp->cp_unacked_bytes -= len;
 291                                 cp->cp_unacked_packets--;
 292                         }
 293
 294                         cp->cp_xmit_rm = rm;
 295                 }
 296
 297                 /* The transport either sends the whole rdma or none of it */
 298                 if (rm->rdma.op_active && !cp->cp_xmit_rdma_sent) {
 299                         rm->m_final_op = &rm->rdma;
 300                         /* The transport owns the mapped memory for now.
 301                          * You can't unmap it while it's on the send queue
 302                          */
 303                         set_bit(RDS_MSG_MAPPED, &rm->m_flags);
 304                         ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
 305                         if (ret) {
 306                                 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
 307                                 wake_up_interruptible(&rm->m_flush_wait);
 308                                 break;
 309                         }
 310                         cp->cp_xmit_rdma_sent = 1;
 311
 312                 }
 313
 314                 if (rm->atomic.op_active && !cp->cp_xmit_atomic_sent) {
 315                         rm->m_final_op = &rm->atomic;
 316                         /* The transport owns the mapped memory for now.
 317                          * You can't unmap it while it's on the send queue
 318                          */
 319                         set_bit(RDS_MSG_MAPPED, &rm->m_flags);
 320                         ret = conn->c_trans->xmit_atomic(conn, &rm->atomic);
 321                         if (ret) {
 322                                 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
 323                                 wake_up_interruptible(&rm->m_flush_wait);
 324                                 break;
 325                         }
 326                         cp->cp_xmit_atomic_sent = 1;
 327
 328                 }
 329
 330                 /*
 331                  * A number of cases require an RDS header to be sent
 332                  * even if there is no data.
 333                  * We permit 0-byte sends; rds-ping depends on this.
 334                  * However, if there are exclusively attached silent ops,
 335                  * we skip the hdr/data send, to enable silent operation.
 336                  */
 337                 if (rm->data.op_nents == 0) {
 338                         int ops_present;
 339                         int all_ops_are_silent = 1;
 340
 341                         ops_present = (rm->atomic.op_active || rm->rdma.op_active);
 342                         if (rm->atomic.op_active && !rm->atomic.op_silent)
 343                                 all_ops_are_silent = 0;
 344                         if (rm->rdma.op_active && !rm->rdma.op_silent)
 345                                 all_ops_are_silent = 0;
 346
 347                         if (ops_present && all_ops_are_silent
 348                             && !rm->m_rdma_cookie)
 349                                 rm->data.op_active = 0;
 350                 }
 351
 352                 if (rm->data.op_active && !cp->cp_xmit_data_sent) {
 353                         rm->m_final_op = &rm->data;
 354
 355                         ret = conn->c_trans->xmit(conn, rm,
 356                                                   cp->cp_xmit_hdr_off,
 357                                                   cp->cp_xmit_sg,
 358                                                   cp->cp_xmit_data_off);
 359                         if (ret <= 0)
 360                                 break;
 361
 362                         if (cp->cp_xmit_hdr_off < sizeof(struct rds_header)) {
 363                                 tmp = min_t(int, ret,
 364                                             sizeof(struct rds_header) -
 365                                             cp->cp_xmit_hdr_off);
 366                                 cp->cp_xmit_hdr_off += tmp;
 367                                 ret -= tmp;
 368                         }
 369
 370                         sg = &rm->data.op_sg[cp->cp_xmit_sg];
 371                         while (ret) {
 372                                 tmp = min_t(int, ret, sg->length -
 373                                                       cp->cp_xmit_data_off);
 374                                 cp->cp_xmit_data_off += tmp;
 375                                 ret -= tmp;
 376                                 if (cp->cp_xmit_data_off == sg->length) {
 377                                         cp->cp_xmit_data_off = 0;
 378                                         sg++;
 379                                         cp->cp_xmit_sg++;
 380                                         BUG_ON(ret != 0 && cp->cp_xmit_sg ==
 381                                                rm->data.op_nents);
 382                                 }
 383                         }
 384
 385                         if (cp->cp_xmit_hdr_off == sizeof(struct rds_header) &&
 386                             (cp->cp_xmit_sg == rm->data.op_nents))
 387                                 cp->cp_xmit_data_sent = 1;
 388                 }
 389
 390                 /*
 391                  * A rm will only take multiple times through this loop
 392                  * if there is a data op. Thus, if the data is sent (or there was
 393                  * none), then we're done with the rm.
 394                  */
 395                 if (!rm->data.op_active || cp->cp_xmit_data_sent) {
 396                         cp->cp_xmit_rm = NULL;
 397                         cp->cp_xmit_sg = 0;
 398                         cp->cp_xmit_hdr_off = 0;
 399                         cp->cp_xmit_data_off = 0;
 400                         cp->cp_xmit_rdma_sent = 0;
 401                         cp->cp_xmit_atomic_sent = 0;
 402                         cp->cp_xmit_data_sent = 0;
 403
 404                         rds_message_put(rm);
 405                 }
 406         }
 407
 408 over_batch:
 409         if (conn->c_trans->xmit_path_complete)
 410                 conn->c_trans->xmit_path_complete(cp);
 411         release_in_xmit(cp);
 412
 413         /* Nuke any messages we decided not to retransmit. */
 414         if (!list_empty(&to_be_dropped)) {
 415                 /* irqs on here, so we can put(), unlike above */
 416                 list_for_each_entry(rm, &to_be_dropped, m_conn_item)
 417                         rds_message_put(rm);
 418                 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
 419         }
 420
 421         /*
 422          * Other senders can queue a message after we last test the send queue
 423          * but before we clear RDS_IN_XMIT.  In that case they'd back off and
 424          * not try and send their newly queued message.  We need to check the
 425          * send queue after having cleared RDS_IN_XMIT so that their message
 426          * doesn't get stuck on the send queue.
 427          *
 428          * If the transport cannot continue (i.e ret != 0), then it must
 429          * call us when more room is available, such as from the tx
 430          * completion handler.
 431          *
 432          * We have an extra generation check here so that if someone manages
 433          * to jump in after our release_in_xmit, we'll see that they have done
 434          * some work and we will skip our goto
 435          */
 436         if (ret == 0) {
 437                 bool raced;
 438
 439                 smp_mb();
 440                 raced = send_gen != READ_ONCE(cp->cp_send_gen);
 441
 442                 if ((test_bit(0, &conn->c_map_queued) ||
 443                     !list_empty(&cp->cp_send_queue)) && !raced) {
 444                         if (batch_count < send_batch_count)
 445                                 goto restart;
 446                         rcu_read_lock();
 447                         if (test_bit(RDS_DESTROY_PENDING, &cp->cp_flags))
 448                                 ret = -ENETUNREACH;
 449                         else
 450                                 queue_delayed_work(rds_wq, &cp->cp_send_w, 1);
 451                         rcu_read_unlock();
 452                 } else if (raced) {
 453                         rds_stats_inc(s_send_lock_queue_raced);
 454                 }
 455         }
 456 out:
 457         return ret;
 458 }
 459 EXPORT_SYMBOL_GPL(rds_send_xmit);
 460
 461 static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
 462 {
 463         u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
 464
 465         assert_spin_locked(&rs->rs_lock);
 466
 467         BUG_ON(rs->rs_snd_bytes < len);
 468         rs->rs_snd_bytes -= len;
 469
 470         if (rs->rs_snd_bytes == 0)
 471                 rds_stats_inc(s_send_queue_empty);
 472 }
 473
 474 static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
 475                                     is_acked_func is_acked)
 476 {
 477         if (is_acked)
 478                 return is_acked(rm, ack);
 479         return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
 480 }
 481
 482 /*
 483  * This is pretty similar to what happens below in the ACK
 484  * handling code - except that we call here as soon as we get
 485  * the IB send completion on the RDMA op and the accompanying
 486  * message.
 487  */
 488 void rds_rdma_send_complete(struct rds_message *rm, int status)
 489 {
 490         struct rds_sock *rs = NULL;
 491         struct rm_rdma_op *ro;
 492         struct rds_notifier *notifier;
 493         unsigned long flags;
 494         unsigned int notify = 0;
 495
 496         spin_lock_irqsave(&rm->m_rs_lock, flags);
 497
 498         notify =  rm->rdma.op_notify | rm->data.op_notify;
 499         ro = &rm->rdma;
 500         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
 501             ro->op_active && notify && ro->op_notifier) {
 502                 notifier = ro->op_notifier;
 503                 rs = rm->m_rs;
 504                 sock_hold(rds_rs_to_sk(rs));
 505
 506                 notifier->n_status = status;
 507                 spin_lock(&rs->rs_lock);
 508                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
 509                 spin_unlock(&rs->rs_lock);
 510
 511                 ro->op_notifier = NULL;
 512         }
 513
 514         spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 515
 516         if (rs) {
 517                 rds_wake_sk_sleep(rs);
 518                 sock_put(rds_rs_to_sk(rs));
 519         }
 520 }
 521 EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
 522
 523 /*
 524  * Just like above, except looks at atomic op
 525  */
 526 void rds_atomic_send_complete(struct rds_message *rm, int status)
 527 {
 528         struct rds_sock *rs = NULL;
 529         struct rm_atomic_op *ao;
 530         struct rds_notifier *notifier;
 531         unsigned long flags;
 532
 533         spin_lock_irqsave(&rm->m_rs_lock, flags);
 534
 535         ao = &rm->atomic;
 536         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
 537             && ao->op_active && ao->op_notify && ao->op_notifier) {
 538                 notifier = ao->op_notifier;
 539                 rs = rm->m_rs;
 540                 sock_hold(rds_rs_to_sk(rs));
 541
 542                 notifier->n_status = status;
 543                 spin_lock(&rs->rs_lock);
 544                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
 545                 spin_unlock(&rs->rs_lock);
 546
 547                 ao->op_notifier = NULL;
 548         }
 549
 550         spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 551
 552         if (rs) {
 553                 rds_wake_sk_sleep(rs);
 554                 sock_put(rds_rs_to_sk(rs));
 555         }
 556 }
 557 EXPORT_SYMBOL_GPL(rds_atomic_send_complete);
 558
 559 /*
 560  * This is the same as rds_rdma_send_complete except we
 561  * don't do any locking - we have all the ingredients (message,
 562  * socket, socket lock) and can just move the notifier.
 563  */
 564 static inline void
 565 __rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
 566 {
 567         struct rm_rdma_op *ro;
 568         struct rm_atomic_op *ao;
 569
 570         ro = &rm->rdma;
 571         if (ro->op_active && ro->op_notify && ro->op_notifier) {
 572                 ro->op_notifier->n_status = status;
 573                 list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue);
 574                 ro->op_notifier = NULL;
 575         }
 576
 577         ao = &rm->atomic;
 578         if (ao->op_active && ao->op_notify && ao->op_notifier) {
 579                 ao->op_notifier->n_status = status;
 580                 list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue);
 581                 ao->op_notifier = NULL;
 582         }
 583
 584         /* No need to wake the app - caller does this */
 585 }
 586
 587 /*
 588  * This removes messages from the socket's list if they're on it.  The list
 589  * argument must be private to the caller, we must be able to modify it
 590  * without locks.  The messages must have a reference held for their
 591  * position on the list.  This function will drop that reference after
 592  * removing the messages from the 'messages' list regardless of if it found
 593  * the messages on the socket list or not.
 594  */
 595 static void rds_send_remove_from_sock(struct list_head *messages, int status)
 596 {
 597         unsigned long flags;
 598         struct rds_sock *rs = NULL;
 599         struct rds_message *rm;
 600
 601         while (!list_empty(messages)) {
 602                 int was_on_sock = 0;
 603
 604                 rm = list_entry(messages->next, struct rds_message,
 605                                 m_conn_item);
 606                 list_del_init(&rm->m_conn_item);
 607
 608                 /*
 609                  * If we see this flag cleared then we're *sure* that someone
 610                  * else beat us to removing it from the sock.  If we race
 611                  * with their flag update we'll get the lock and then really
 612                  * see that the flag has been cleared.
 613                  *
 614                  * The message spinlock makes sure nobody clears rm->m_rs
 615                  * while we're messing with it. It does not prevent the
 616                  * message from being removed from the socket, though.
 617                  */
 618                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 619                 if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags))
 620                         goto unlock_and_drop;
 621
 622                 if (rs != rm->m_rs) {
 623                         if (rs) {
 624                                 rds_wake_sk_sleep(rs);
 625                                 sock_put(rds_rs_to_sk(rs));
 626                         }
 627                         rs = rm->m_rs;
 628                         if (rs)
 629                                 sock_hold(rds_rs_to_sk(rs));
 630                 }
 631                 if (!rs)
 632                         goto unlock_and_drop;
 633                 spin_lock(&rs->rs_lock);
 634
 635                 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
 636                         struct rm_rdma_op *ro = &rm->rdma;
 637                         struct rds_notifier *notifier;
 638
 639                         list_del_init(&rm->m_sock_item);
 640                         rds_send_sndbuf_remove(rs, rm);
 641
 642                         if (ro->op_active && ro->op_notifier &&
 643                                (ro->op_notify || (ro->op_recverr && status))) {
 644                                 notifier = ro->op_notifier;
 645                                 list_add_tail(&notifier->n_list,
 646                                                 &rs->rs_notify_queue);
 647                                 if (!notifier->n_status)
 648                                         notifier->n_status = status;
 649                                 rm->rdma.op_notifier = NULL;
 650                         }
 651                         was_on_sock = 1;
 652                         rm->m_rs = NULL;
 653                 }
 654                 spin_unlock(&rs->rs_lock);
 655
 656 unlock_and_drop:
 657                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 658                 rds_message_put(rm);
 659                 if (was_on_sock)
 660                         rds_message_put(rm);
 661         }
 662
 663         if (rs) {
 664                 rds_wake_sk_sleep(rs);
 665                 sock_put(rds_rs_to_sk(rs));
 666         }
 667 }
 668
 669 /*
 670  * Transports call here when they've determined that the receiver queued
 671  * messages up to, and including, the given sequence number.  Messages are
 672  * moved to the retrans queue when rds_send_xmit picks them off the send
 673  * queue. This means that in the TCP case, the message may not have been
 674  * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
 675  * checks the RDS_MSG_HAS_ACK_SEQ bit.
 676  */
 677 void rds_send_path_drop_acked(struct rds_conn_path *cp, u64 ack,
 678                               is_acked_func is_acked)
 679 {
 680         struct rds_message *rm, *tmp;
 681         unsigned long flags;
 682         LIST_HEAD(list);
 683
 684         spin_lock_irqsave(&cp->cp_lock, flags);
 685
 686         list_for_each_entry_safe(rm, tmp, &cp->cp_retrans, m_conn_item) {
 687                 if (!rds_send_is_acked(rm, ack, is_acked))
 688                         break;
 689
 690                 list_move(&rm->m_conn_item, &list);
 691                 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
 692         }
 693
 694         /* order flag updates with spin locks */
 695         if (!list_empty(&list))
 696                 smp_mb__after_atomic();
 697
 698         spin_unlock_irqrestore(&cp->cp_lock, flags);
 699
 700         /* now remove the messages from the sock list as needed */
 701         rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
 702 }
 703 EXPORT_SYMBOL_GPL(rds_send_path_drop_acked);
 704
 705 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
 706                          is_acked_func is_acked)
 707 {
 708         WARN_ON(conn->c_trans->t_mp_capable);
 709         rds_send_path_drop_acked(&conn->c_path[0], ack, is_acked);
 710 }
 711 EXPORT_SYMBOL_GPL(rds_send_drop_acked);
 712
 713 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
 714 {
 715         struct rds_message *rm, *tmp;
 716         struct rds_connection *conn;
 717         struct rds_conn_path *cp;
 718         unsigned long flags;
 719         LIST_HEAD(list);
 720
 721         /* get all the messages we're dropping under the rs lock */
 722         spin_lock_irqsave(&rs->rs_lock, flags);
 723
 724         list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) {
 725                 if (dest && (dest->sin_addr.s_addr != rm->m_daddr ||
 726                              dest->sin_port != rm->m_inc.i_hdr.h_dport))
 727                         continue;
 728
 729                 list_move(&rm->m_sock_item, &list);
 730                 rds_send_sndbuf_remove(rs, rm);
 731                 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
 732         }
 733
 734         /* order flag updates with the rs lock */
 735         smp_mb__after_atomic();
 736
 737         spin_unlock_irqrestore(&rs->rs_lock, flags);
 738
 739         if (list_empty(&list))
 740                 return;
 741
 742         /* Remove the messages from the conn */
 743         list_for_each_entry(rm, &list, m_sock_item) {
 744
 745                 conn = rm->m_inc.i_conn;
 746                 if (conn->c_trans->t_mp_capable)
 747                         cp = rm->m_inc.i_conn_path;
 748                 else
 749                         cp = &conn->c_path[0];
 750
 751                 spin_lock_irqsave(&cp->cp_lock, flags);
 752                 /*
 753                  * Maybe someone else beat us to removing rm from the conn.
 754                  * If we race with their flag update we'll get the lock and
 755                  * then really see that the flag has been cleared.
 756                  */
 757                 if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
 758                         spin_unlock_irqrestore(&cp->cp_lock, flags);
 759                         spin_lock_irqsave(&rm->m_rs_lock, flags);
 760                         rm->m_rs = NULL;
 761                         spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 762                         continue;
 763                 }
 764                 list_del_init(&rm->m_conn_item);
 765                 spin_unlock_irqrestore(&cp->cp_lock, flags);
 766
 767                 /*
 768                  * Couldn't grab m_rs_lock in top loop (lock ordering),
 769                  * but we can now.
 770                  */
 771                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 772
 773                 spin_lock(&rs->rs_lock);
 774                 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
 775                 spin_unlock(&rs->rs_lock);
 776
 777                 rm->m_rs = NULL;
 778                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 779
 780                 rds_message_put(rm);
 781         }
 782
 783         rds_wake_sk_sleep(rs);
 784
 785         while (!list_empty(&list)) {
 786                 rm = list_entry(list.next, struct rds_message, m_sock_item);
 787                 list_del_init(&rm->m_sock_item);
 788                 rds_message_wait(rm);
 789
 790                 /* just in case the code above skipped this message
 791                  * because RDS_MSG_ON_CONN wasn't set, run it again here
 792                  * taking m_rs_lock is the only thing that keeps us
 793                  * from racing with ack processing.
 794                  */
 795                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 796
 797                 spin_lock(&rs->rs_lock);
 798                 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
 799                 spin_unlock(&rs->rs_lock);
 800
 801                 rm->m_rs = NULL;
 802                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 803
 804                 rds_message_put(rm);
 805         }
 806 }
 807
 808 /*
 809  * we only want this to fire once so we use the callers 'queued'.  It's
 810  * possible that another thread can race with us and remove the
 811  * message from the flow with RDS_CANCEL_SENT_TO.
 812  */
 813 static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
 814                              struct rds_conn_path *cp,
 815                              struct rds_message *rm, __be16 sport,
 816                              __be16 dport, int *queued)
 817 {
 818         unsigned long flags;
 819         u32 len;
 820
 821         if (*queued)
 822                 goto out;
 823
 824         len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
 825
 826         /* this is the only place which holds both the socket's rs_lock
 827          * and the connection's c_lock */
 828         spin_lock_irqsave(&rs->rs_lock, flags);
 829
 830         /*
 831          * If there is a little space in sndbuf, we don't queue anything,
 832          * and userspace gets -EAGAIN. But poll() indicates there's send
 833          * room. This can lead to bad behavior (spinning) if snd_bytes isn't
 834          * freed up by incoming acks. So we check the *old* value of
 835          * rs_snd_bytes here to allow the last msg to exceed the buffer,
 836          * and poll() now knows no more data can be sent.
 837          */
 838         if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
 839                 rs->rs_snd_bytes += len;
 840
 841                 /* let recv side know we are close to send space exhaustion.
 842                  * This is probably not the optimal way to do it, as this
 843                  * means we set the flag on *all* messages as soon as our
 844                  * throughput hits a certain threshold.
 845                  */
 846                 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
 847                         set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
 848
 849                 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
 850                 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
 851                 rds_message_addref(rm);
 852                 rm->m_rs = rs;
 853
 854                 /* The code ordering is a little weird, but we're
 855                    trying to minimize the time we hold c_lock */
 856                 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
 857                 rm->m_inc.i_conn = conn;
 858                 rm->m_inc.i_conn_path = cp;
 859                 rds_message_addref(rm);
 860
 861                 spin_lock(&cp->cp_lock);
 862                 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(cp->cp_next_tx_seq++);
 863                 list_add_tail(&rm->m_conn_item, &cp->cp_send_queue);
 864                 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
 865                 spin_unlock(&cp->cp_lock);
 866
 867                 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
 868                          rm, len, rs, rs->rs_snd_bytes,
 869                          (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
 870
 871                 *queued = 1;
 872         }
 873
 874         spin_unlock_irqrestore(&rs->rs_lock, flags);
 875 out:
 876         return *queued;
 877 }
 878
 879 /*
 880  * rds_message is getting to be quite complicated, and we'd like to allocate
 881  * it all in one go. This figures out how big it needs to be up front.
 882  */
 883 static int rds_rm_size(struct msghdr *msg, int data_len)
 884 {
 885         struct cmsghdr *cmsg;
 886         int size = 0;
 887         int cmsg_groups = 0;
 888         int retval;
 889
 890         for_each_cmsghdr(cmsg, msg) {
 891                 if (!CMSG_OK(msg, cmsg))
 892                         return -EINVAL;
 893
 894                 if (cmsg->cmsg_level != SOL_RDS)
 895                         continue;
 896
 897                 switch (cmsg->cmsg_type) {
 898                 case RDS_CMSG_RDMA_ARGS:
 899                         cmsg_groups |= 1;
 900                         retval = rds_rdma_extra_size(CMSG_DATA(cmsg));
 901                         if (retval < 0)
 902                                 return retval;
 903                         size += retval;
 904
 905                         break;
 906
 907                 case RDS_CMSG_RDMA_DEST:
 908                 case RDS_CMSG_RDMA_MAP:
 909                         cmsg_groups |= 2;
 910                         /* these are valid but do no add any size */
 911                         break;
 912
 913                 case RDS_CMSG_ATOMIC_CSWP:
 914                 case RDS_CMSG_ATOMIC_FADD:
 915                 case RDS_CMSG_MASKED_ATOMIC_CSWP:
 916                 case RDS_CMSG_MASKED_ATOMIC_FADD:
 917                         cmsg_groups |= 1;
 918                         size += sizeof(struct scatterlist);
 919                         break;
 920
 921                 default:
 922                         return -EINVAL;
 923                 }
 924
 925         }
 926
 927         size += ceil(data_len, PAGE_SIZE) * sizeof(struct scatterlist);
 928
 929         /* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */
 930         if (cmsg_groups == 3)
 931                 return -EINVAL;
 932
 933         return size;
 934 }
 935
 936 static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
 937                          struct msghdr *msg, int *allocated_mr)
 938 {
 939         struct cmsghdr *cmsg;
 940         int ret = 0;
 941
 942         for_each_cmsghdr(cmsg, msg) {
 943                 if (!CMSG_OK(msg, cmsg))
 944                         return -EINVAL;
 945
 946                 if (cmsg->cmsg_level != SOL_RDS)
 947                         continue;
 948
 949                 /* As a side effect, RDMA_DEST and RDMA_MAP will set
 950                  * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
 951                  */
 952                 switch (cmsg->cmsg_type) {
 953                 case RDS_CMSG_RDMA_ARGS:
 954                         ret = rds_cmsg_rdma_args(rs, rm, cmsg);
 955                         break;
 956
 957                 case RDS_CMSG_RDMA_DEST:
 958                         ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
 959                         break;
 960
 961                 case RDS_CMSG_RDMA_MAP:
 962                         ret = rds_cmsg_rdma_map(rs, rm, cmsg);
 963                         if (!ret)
 964                                 *allocated_mr = 1;
 965                         else if (ret == -ENODEV)
 966                                 /* Accommodate the get_mr() case which can fail
 967                                  * if connection isn't established yet.
 968                                  */
 969                                 ret = -EAGAIN;
 970                         break;
 971                 case RDS_CMSG_ATOMIC_CSWP:
 972                 case RDS_CMSG_ATOMIC_FADD:
 973                 case RDS_CMSG_MASKED_ATOMIC_CSWP:
 974                 case RDS_CMSG_MASKED_ATOMIC_FADD:
 975                         ret = rds_cmsg_atomic(rs, rm, cmsg);
 976                         break;
 977
 978                 default:
 979                         return -EINVAL;
 980                 }
 981
 982                 if (ret)
 983                         break;
 984         }
 985
 986         return ret;
 987 }
 988
 989 static int rds_send_mprds_hash(struct rds_sock *rs, struct rds_connection *conn)
 990 {
 991         int hash;
 992
 993         if (conn->c_npaths == 0)
 994                 hash = RDS_MPATH_HASH(rs, RDS_MPATH_WORKERS);
 995         else
 996                 hash = RDS_MPATH_HASH(rs, conn->c_npaths);
 997         if (conn->c_npaths == 0 && hash != 0) {
 998                 rds_send_ping(conn, 0);
 999
1000                 if (conn->c_npaths == 0) {
1001                         wait_event_interruptible(conn->c_hs_waitq,
1002                                                  (conn->c_npaths != 0));
1003                 }
1004                 if (conn->c_npaths == 1)
1005                         hash = 0;
1006         }
1007         return hash;
1008 }
1009
1010 static int rds_rdma_bytes(struct msghdr *msg, size_t *rdma_bytes)
1011 {
1012         struct rds_rdma_args *args;
1013         struct cmsghdr *cmsg;
1014
1015         for_each_cmsghdr(cmsg, msg) {
1016                 if (!CMSG_OK(msg, cmsg))
1017                         return -EINVAL;
1018
1019                 if (cmsg->cmsg_level != SOL_RDS)
1020                         continue;
1021
1022                 if (cmsg->cmsg_type == RDS_CMSG_RDMA_ARGS) {
1023                         if (cmsg->cmsg_len <
1024                             CMSG_LEN(sizeof(struct rds_rdma_args)))
1025                                 return -EINVAL;
1026                         args = CMSG_DATA(cmsg);
1027                         *rdma_bytes += args->remote_vec.bytes;
1028                 }
1029         }
1030         return 0;
1031 }
1032
1033 int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len)
1034 {
1035         struct sock *sk = sock->sk;
1036         struct rds_sock *rs = rds_sk_to_rs(sk);
1037         DECLARE_SOCKADDR(struct sockaddr_in *, usin, msg->msg_name);
1038         __be32 daddr;
1039         __be16 dport;
1040         struct rds_message *rm = NULL;
1041         struct rds_connection *conn;
1042         int ret = 0;
1043         int queued = 0, allocated_mr = 0;
1044         int nonblock = msg->msg_flags & MSG_DONTWAIT;
1045         long timeo = sock_sndtimeo(sk, nonblock);
1046         struct rds_conn_path *cpath;
1047         size_t total_payload_len = payload_len, rdma_payload_len = 0;
1048
1049         /* Mirror Linux UDP mirror of BSD error message compatibility */
1050         /* XXX: Perhaps MSG_MORE someday */
1051         if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT)) {
1052                 ret = -EOPNOTSUPP;
1053                 goto out;
1054         }
1055
1056         if (msg->msg_namelen) {
1057                 /* XXX fail non-unicast destination IPs? */
1058                 if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) {
1059                         ret = -EINVAL;
1060                         goto out;
1061                 }
1062                 daddr = usin->sin_addr.s_addr;
1063                 dport = usin->sin_port;
1064         } else {
1065                 /* We only care about consistency with ->connect() */
1066                 lock_sock(sk);
1067                 daddr = rs->rs_conn_addr;
1068                 dport = rs->rs_conn_port;
1069                 release_sock(sk);
1070         }
1071
1072         lock_sock(sk);
1073         if (daddr == 0 || rs->rs_bound_addr == 0) {
1074                 release_sock(sk);
1075                 ret = -ENOTCONN; /* XXX not a great errno */
1076                 goto out;
1077         }
1078         release_sock(sk);
1079
1080         ret = rds_rdma_bytes(msg, &rdma_payload_len);
1081         if (ret)
1082                 goto out;
1083
1084         total_payload_len += rdma_payload_len;
1085         if (max_t(size_t, payload_len, rdma_payload_len) > RDS_MAX_MSG_SIZE) {
1086                 ret = -EMSGSIZE;
1087                 goto out;
1088         }
1089
1090         if (payload_len > rds_sk_sndbuf(rs)) {
1091                 ret = -EMSGSIZE;
1092                 goto out;
1093         }
1094
1095         /* size of rm including all sgs */
1096         ret = rds_rm_size(msg, payload_len);
1097         if (ret < 0)
1098                 goto out;
1099
1100         rm = rds_message_alloc(ret, GFP_KERNEL);
1101         if (!rm) {
1102                 ret = -ENOMEM;
1103                 goto out;
1104         }
1105
1106         /* Attach data to the rm */
1107         if (payload_len) {
1108                 rm->data.op_sg = rds_message_alloc_sgs(rm, ceil(payload_len, PAGE_SIZE));
1109                 if (!rm->data.op_sg) {
1110                         ret = -ENOMEM;
1111                         goto out;
1112                 }
1113                 ret = rds_message_copy_from_user(rm, &msg->msg_iter);
1114                 if (ret)
1115                         goto out;
1116         }
1117         rm->data.op_active = 1;
1118
1119         rm->m_daddr = daddr;
1120
1121         /* rds_conn_create has a spinlock that runs with IRQ off.
1122          * Caching the conn in the socket helps a lot. */
1123         if (rs->rs_conn && rs->rs_conn->c_faddr == daddr)
1124                 conn = rs->rs_conn;
1125         else {
1126                 conn = rds_conn_create_outgoing(sock_net(sock->sk),
1127                                                 rs->rs_bound_addr, daddr,
1128                                         rs->rs_transport,
1129                                         sock->sk->sk_allocation);
1130                 if (IS_ERR(conn)) {
1131                         ret = PTR_ERR(conn);
1132                         goto out;
1133                 }
1134                 rs->rs_conn = conn;
1135         }
1136
1137         /* Parse any control messages the user may have included. */
1138         ret = rds_cmsg_send(rs, rm, msg, &allocated_mr);
1139         if (ret) {
1140                 /* Trigger connection so that its ready for the next retry */
1141                 if (ret ==  -EAGAIN)
1142                         rds_conn_connect_if_down(conn);
1143                 goto out;
1144         }
1145
1146         if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) {
1147                 printk_ratelimited(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
1148                                &rm->rdma, conn->c_trans->xmit_rdma);
1149                 ret = -EOPNOTSUPP;
1150                 goto out;
1151         }
1152
1153         if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
1154                 printk_ratelimited(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
1155                                &rm->atomic, conn->c_trans->xmit_atomic);
1156                 ret = -EOPNOTSUPP;
1157                 goto out;
1158         }
1159
1160         if (conn->c_trans->t_mp_capable)
1161                 cpath = &conn->c_path[rds_send_mprds_hash(rs, conn)];
1162         else
1163                 cpath = &conn->c_path[0];
1164
1165         if (test_bit(RDS_DESTROY_PENDING, &cpath->cp_flags)) {
1166                 ret = -EAGAIN;
1167                 goto out;
1168         }
1169
1170         rds_conn_path_connect_if_down(cpath);
1171
1172         ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
1173         if (ret) {
1174                 rs->rs_seen_congestion = 1;
1175                 goto out;
1176         }
1177         while (!rds_send_queue_rm(rs, conn, cpath, rm, rs->rs_bound_port,
1178                                   dport, &queued)) {
1179                 rds_stats_inc(s_send_queue_full);
1180
1181                 if (nonblock) {
1182                         ret = -EAGAIN;
1183                         goto out;
1184                 }
1185
1186                 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
1187                                         rds_send_queue_rm(rs, conn, cpath, rm,
1188                                                           rs->rs_bound_port,
1189                                                           dport,
1190                                                           &queued),
1191                                         timeo);
1192                 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
1193                 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
1194                         continue;
1195
1196                 ret = timeo;
1197                 if (ret == 0)
1198                         ret = -ETIMEDOUT;
1199                 goto out;
1200         }
1201
1202         /*
1203          * By now we've committed to the send.  We reuse rds_send_worker()
1204          * to retry sends in the rds thread if the transport asks us to.
1205          */
1206         rds_stats_inc(s_send_queued);
1207
1208         ret = rds_send_xmit(cpath);
1209         if (ret == -ENOMEM || ret == -EAGAIN) {
1210                 ret = 0;
1211                 rcu_read_lock();
1212                 if (test_bit(RDS_DESTROY_PENDING, &cpath->cp_flags))
1213                         ret = -ENETUNREACH;
1214                 else
1215                         queue_delayed_work(rds_wq, &cpath->cp_send_w, 1);
1216                 rcu_read_unlock();
1217         }
1218         if (ret)
1219                 goto out;
1220         rds_message_put(rm);
1221         return payload_len;
1222
1223 out:
1224         /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1225          * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1226          * or in any other way, we need to destroy the MR again */
1227         if (allocated_mr)
1228                 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
1229
1230         if (rm)
1231                 rds_message_put(rm);
1232         return ret;
1233 }
1234
1235 /*
1236  * send out a probe. Can be shared by rds_send_ping,
1237  * rds_send_pong, rds_send_hb.
1238  * rds_send_hb should use h_flags
1239  *   RDS_FLAG_HB_PING|RDS_FLAG_ACK_REQUIRED
1240  * or
1241  *   RDS_FLAG_HB_PONG|RDS_FLAG_ACK_REQUIRED
1242  */
1243 static int
1244 rds_send_probe(struct rds_conn_path *cp, __be16 sport,
1245                __be16 dport, u8 h_flags)
1246 {
1247         struct rds_message *rm;
1248         unsigned long flags;
1249         int ret = 0;
1250
1251         rm = rds_message_alloc(0, GFP_ATOMIC);
1252         if (!rm) {
1253                 ret = -ENOMEM;
1254                 goto out;
1255         }
1256
1257         rm->m_daddr = cp->cp_conn->c_faddr;
1258         rm->data.op_active = 1;
1259
1260         rds_conn_path_connect_if_down(cp);
1261
1262         ret = rds_cong_wait(cp->cp_conn->c_fcong, dport, 1, NULL);
1263         if (ret)
1264                 goto out;
1265
1266         spin_lock_irqsave(&cp->cp_lock, flags);
1267         list_add_tail(&rm->m_conn_item, &cp->cp_send_queue);
1268         set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
1269         rds_message_addref(rm);
1270         rm->m_inc.i_conn = cp->cp_conn;
1271         rm->m_inc.i_conn_path = cp;
1272
1273         rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport,
1274                                     cp->cp_next_tx_seq);
1275         rm->m_inc.i_hdr.h_flags |= h_flags;
1276         cp->cp_next_tx_seq++;
1277
1278         if (RDS_HS_PROBE(be16_to_cpu(sport), be16_to_cpu(dport)) &&
1279             cp->cp_conn->c_trans->t_mp_capable) {
1280                 u16 npaths = cpu_to_be16(RDS_MPATH_WORKERS);
1281                 u32 my_gen_num = cpu_to_be32(cp->cp_conn->c_my_gen_num);
1282
1283                 rds_message_add_extension(&rm->m_inc.i_hdr,
1284                                           RDS_EXTHDR_NPATHS, &npaths,
1285                                           sizeof(npaths));
1286                 rds_message_add_extension(&rm->m_inc.i_hdr,
1287                                           RDS_EXTHDR_GEN_NUM,
1288                                           &my_gen_num,
1289                                           sizeof(u32));
1290         }
1291         spin_unlock_irqrestore(&cp->cp_lock, flags);
1292
1293         rds_stats_inc(s_send_queued);
1294         rds_stats_inc(s_send_pong);
1295
1296         /* schedule the send work on rds_wq */
1297         rcu_read_lock();
1298         if (!test_bit(RDS_DESTROY_PENDING, &cp->cp_flags))
1299                 queue_delayed_work(rds_wq, &cp->cp_send_w, 1);
1300         rcu_read_unlock();
1301
1302         rds_message_put(rm);
1303         return 0;
1304
1305 out:
1306         if (rm)
1307                 rds_message_put(rm);
1308         return ret;
1309 }
1310
1311 int
1312 rds_send_pong(struct rds_conn_path *cp, __be16 dport)
1313 {
1314         return rds_send_probe(cp, 0, dport, 0);
1315 }
1316
1317 void
1318 rds_send_ping(struct rds_connection *conn, int cp_index)
1319 {
1320         unsigned long flags;
1321         struct rds_conn_path *cp = &conn->c_path[cp_index];
1322
1323         spin_lock_irqsave(&cp->cp_lock, flags);
1324         if (conn->c_ping_triggered) {
1325                 spin_unlock_irqrestore(&cp->cp_lock, flags);
1326                 return;
1327         }
1328         conn->c_ping_triggered = 1;
1329         spin_unlock_irqrestore(&cp->cp_lock, flags);
1330         rds_send_probe(cp, cpu_to_be16(RDS_FLAG_PROBE_PORT), 0, 0);
1331 }
1332 EXPORT_SYMBOL_GPL(rds_send_ping);