rds/send.c

   1 /*
   2  * Copyright (c) 2006, 2018 Oracle and/or its affiliates. All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  *
  32  */
  33 #include <linux/kernel.h>
  34 #include <linux/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_lock(RDS_IN_XMIT, &cp->cp_flags) == 0;
 107 }
 108
 109 static void release_in_xmit(struct rds_conn_path *cp)
 110 {
 111         clear_bit_unlock(RDS_IN_XMIT, &cp->cp_flags);
 112         /*
 113          * We don't use wait_on_bit()/wake_up_bit() because our waking is in a
 114          * hot path and finding waiters is very rare.  We don't want to walk
 115          * the system-wide hashed waitqueue buckets in the fast path only to
 116          * almost never find waiters.
 117          */
 118         if (waitqueue_active(&cp->cp_waitq))
 119                 wake_up_all(&cp->cp_waitq);
 120 }
 121
 122 /*
 123  * We're making the conscious trade-off here to only send one message
 124  * down the connection at a time.
 125  *   Pro:
 126  *      - tx queueing is a simple fifo list
 127  *      - reassembly is optional and easily done by transports per conn
 128  *      - no per flow rx lookup at all, straight to the socket
 129  *      - less per-frag memory and wire overhead
 130  *   Con:
 131  *      - queued acks can be delayed behind large messages
 132  *   Depends:
 133  *      - small message latency is higher behind queued large messages
 134  *      - large message latency isn't starved by intervening small sends
 135  */
 136 int rds_send_xmit(struct rds_conn_path *cp)
 137 {
 138         struct rds_connection *conn = cp->cp_conn;
 139         struct rds_message *rm;
 140         unsigned long flags;
 141         unsigned int tmp;
 142         struct scatterlist *sg;
 143         int ret = 0;
 144         LIST_HEAD(to_be_dropped);
 145         int batch_count;
 146         unsigned long send_gen = 0;
 147         int same_rm = 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 (rds_destroy_pending(cp->cp_conn)) {
 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                 if (!rm) {
 204                         same_rm = 0;
 205                 } else {
 206                         same_rm++;
 207                         if (same_rm >= 4096) {
 208                                 rds_stats_inc(s_send_stuck_rm);
 209                                 ret = -EAGAIN;
 210                                 break;
 211                         }
 212                 }
 213
 214                 /*
 215                  * If between sending messages, we can send a pending congestion
 216                  * map update.
 217                  */
 218                 if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) {
 219                         rm = rds_cong_update_alloc(conn);
 220                         if (IS_ERR(rm)) {
 221                                 ret = PTR_ERR(rm);
 222                                 break;
 223                         }
 224                         rm->data.op_active = 1;
 225                         rm->m_inc.i_conn_path = cp;
 226                         rm->m_inc.i_conn = cp->cp_conn;
 227
 228                         cp->cp_xmit_rm = rm;
 229                 }
 230
 231                 /*
 232                  * If not already working on one, grab the next message.
 233                  *
 234                  * cp_xmit_rm holds a ref while we're sending this message down
 235                  * the connction.  We can use this ref while holding the
 236                  * send_sem.. rds_send_reset() is serialized with it.
 237                  */
 238                 if (!rm) {
 239                         unsigned int len;
 240
 241                         batch_count++;
 242
 243                         /* we want to process as big a batch as we can, but
 244                          * we also want to avoid softlockups.  If we've been
 245                          * through a lot of messages, lets back off and see
 246                          * if anyone else jumps in
 247                          */
 248                         if (batch_count >= send_batch_count)
 249                                 goto over_batch;
 250
 251                         spin_lock_irqsave(&cp->cp_lock, flags);
 252
 253                         if (!list_empty(&cp->cp_send_queue)) {
 254                                 rm = list_entry(cp->cp_send_queue.next,
 255                                                 struct rds_message,
 256                                                 m_conn_item);
 257                                 rds_message_addref(rm);
 258
 259                                 /*
 260                                  * Move the message from the send queue to the retransmit
 261                                  * list right away.
 262                                  */
 263                                 list_move_tail(&rm->m_conn_item,
 264                                                &cp->cp_retrans);
 265                         }
 266
 267                         spin_unlock_irqrestore(&cp->cp_lock, flags);
 268
 269                         if (!rm)
 270                                 break;
 271
 272                         /* Unfortunately, the way Infiniband deals with
 273                          * RDMA to a bad MR key is by moving the entire
 274                          * queue pair to error state. We could possibly
 275                          * recover from that, but right now we drop the
 276                          * connection.
 277                          * Therefore, we never retransmit messages with RDMA ops.
 278                          */
 279                         if (test_bit(RDS_MSG_FLUSH, &rm->m_flags) ||
 280                             (rm->rdma.op_active &&
 281                             test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))) {
 282                                 spin_lock_irqsave(&cp->cp_lock, flags);
 283                                 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags))
 284                                         list_move(&rm->m_conn_item, &to_be_dropped);
 285                                 spin_unlock_irqrestore(&cp->cp_lock, flags);
 286                                 continue;
 287                         }
 288
 289                         /* Require an ACK every once in a while */
 290                         len = ntohl(rm->m_inc.i_hdr.h_len);
 291                         if (cp->cp_unacked_packets == 0 ||
 292                             cp->cp_unacked_bytes < len) {
 293                                 set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
 294
 295                                 cp->cp_unacked_packets =
 296                                         rds_sysctl_max_unacked_packets;
 297                                 cp->cp_unacked_bytes =
 298                                         rds_sysctl_max_unacked_bytes;
 299                                 rds_stats_inc(s_send_ack_required);
 300                         } else {
 301                                 cp->cp_unacked_bytes -= len;
 302                                 cp->cp_unacked_packets--;
 303                         }
 304
 305                         cp->cp_xmit_rm = rm;
 306                 }
 307
 308                 /* The transport either sends the whole rdma or none of it */
 309                 if (rm->rdma.op_active && !cp->cp_xmit_rdma_sent) {
 310                         rm->m_final_op = &rm->rdma;
 311                         /* The transport owns the mapped memory for now.
 312                          * You can't unmap it while it's on the send queue
 313                          */
 314                         set_bit(RDS_MSG_MAPPED, &rm->m_flags);
 315                         ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
 316                         if (ret) {
 317                                 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
 318                                 wake_up_interruptible(&rm->m_flush_wait);
 319                                 break;
 320                         }
 321                         cp->cp_xmit_rdma_sent = 1;
 322
 323                 }
 324
 325                 if (rm->atomic.op_active && !cp->cp_xmit_atomic_sent) {
 326                         rm->m_final_op = &rm->atomic;
 327                         /* The transport owns the mapped memory for now.
 328                          * You can't unmap it while it's on the send queue
 329                          */
 330                         set_bit(RDS_MSG_MAPPED, &rm->m_flags);
 331                         ret = conn->c_trans->xmit_atomic(conn, &rm->atomic);
 332                         if (ret) {
 333                                 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
 334                                 wake_up_interruptible(&rm->m_flush_wait);
 335                                 break;
 336                         }
 337                         cp->cp_xmit_atomic_sent = 1;
 338
 339                 }
 340
 341                 /*
 342                  * A number of cases require an RDS header to be sent
 343                  * even if there is no data.
 344                  * We permit 0-byte sends; rds-ping depends on this.
 345                  * However, if there are exclusively attached silent ops,
 346                  * we skip the hdr/data send, to enable silent operation.
 347                  */
 348                 if (rm->data.op_nents == 0) {
 349                         int ops_present;
 350                         int all_ops_are_silent = 1;
 351
 352                         ops_present = (rm->atomic.op_active || rm->rdma.op_active);
 353                         if (rm->atomic.op_active && !rm->atomic.op_silent)
 354                                 all_ops_are_silent = 0;
 355                         if (rm->rdma.op_active && !rm->rdma.op_silent)
 356                                 all_ops_are_silent = 0;
 357
 358                         if (ops_present && all_ops_are_silent
 359                             && !rm->m_rdma_cookie)
 360                                 rm->data.op_active = 0;
 361                 }
 362
 363                 if (rm->data.op_active && !cp->cp_xmit_data_sent) {
 364                         rm->m_final_op = &rm->data;
 365
 366                         ret = conn->c_trans->xmit(conn, rm,
 367                                                   cp->cp_xmit_hdr_off,
 368                                                   cp->cp_xmit_sg,
 369                                                   cp->cp_xmit_data_off);
 370                         if (ret <= 0)
 371                                 break;
 372
 373                         if (cp->cp_xmit_hdr_off < sizeof(struct rds_header)) {
 374                                 tmp = min_t(int, ret,
 375                                             sizeof(struct rds_header) -
 376                                             cp->cp_xmit_hdr_off);
 377                                 cp->cp_xmit_hdr_off += tmp;
 378                                 ret -= tmp;
 379                         }
 380
 381                         sg = &rm->data.op_sg[cp->cp_xmit_sg];
 382                         while (ret) {
 383                                 tmp = min_t(int, ret, sg->length -
 384                                                       cp->cp_xmit_data_off);
 385                                 cp->cp_xmit_data_off += tmp;
 386                                 ret -= tmp;
 387                                 if (cp->cp_xmit_data_off == sg->length) {
 388                                         cp->cp_xmit_data_off = 0;
 389                                         sg++;
 390                                         cp->cp_xmit_sg++;
 391                                         BUG_ON(ret != 0 && cp->cp_xmit_sg ==
 392                                                rm->data.op_nents);
 393                                 }
 394                         }
 395
 396                         if (cp->cp_xmit_hdr_off == sizeof(struct rds_header) &&
 397                             (cp->cp_xmit_sg == rm->data.op_nents))
 398                                 cp->cp_xmit_data_sent = 1;
 399                 }
 400
 401                 /*
 402                  * A rm will only take multiple times through this loop
 403                  * if there is a data op. Thus, if the data is sent (or there was
 404                  * none), then we're done with the rm.
 405                  */
 406                 if (!rm->data.op_active || cp->cp_xmit_data_sent) {
 407                         cp->cp_xmit_rm = NULL;
 408                         cp->cp_xmit_sg = 0;
 409                         cp->cp_xmit_hdr_off = 0;
 410                         cp->cp_xmit_data_off = 0;
 411                         cp->cp_xmit_rdma_sent = 0;
 412                         cp->cp_xmit_atomic_sent = 0;
 413                         cp->cp_xmit_data_sent = 0;
 414
 415                         rds_message_put(rm);
 416                 }
 417         }
 418
 419 over_batch:
 420         if (conn->c_trans->xmit_path_complete)
 421                 conn->c_trans->xmit_path_complete(cp);
 422         release_in_xmit(cp);
 423
 424         /* Nuke any messages we decided not to retransmit. */
 425         if (!list_empty(&to_be_dropped)) {
 426                 /* irqs on here, so we can put(), unlike above */
 427                 list_for_each_entry(rm, &to_be_dropped, m_conn_item)
 428                         rds_message_put(rm);
 429                 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
 430         }
 431
 432         /*
 433          * Other senders can queue a message after we last test the send queue
 434          * but before we clear RDS_IN_XMIT.  In that case they'd back off and
 435          * not try and send their newly queued message.  We need to check the
 436          * send queue after having cleared RDS_IN_XMIT so that their message
 437          * doesn't get stuck on the send queue.
 438          *
 439          * If the transport cannot continue (i.e ret != 0), then it must
 440          * call us when more room is available, such as from the tx
 441          * completion handler.
 442          *
 443          * We have an extra generation check here so that if someone manages
 444          * to jump in after our release_in_xmit, we'll see that they have done
 445          * some work and we will skip our goto
 446          */
 447         if (ret == 0) {
 448                 bool raced;
 449
 450                 smp_mb();
 451                 raced = send_gen != READ_ONCE(cp->cp_send_gen);
 452
 453                 if ((test_bit(0, &conn->c_map_queued) ||
 454                     !list_empty(&cp->cp_send_queue)) && !raced) {
 455                         if (batch_count < send_batch_count)
 456                                 goto restart;
 457                         rcu_read_lock();
 458                         if (rds_destroy_pending(cp->cp_conn))
 459                                 ret = -ENETUNREACH;
 460                         else
 461                                 queue_delayed_work(rds_wq, &cp->cp_send_w, 1);
 462                         rcu_read_unlock();
 463                 } else if (raced) {
 464                         rds_stats_inc(s_send_lock_queue_raced);
 465                 }
 466         }
 467 out:
 468         return ret;
 469 }
 470 EXPORT_SYMBOL_GPL(rds_send_xmit);
 471
 472 static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
 473 {
 474         u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
 475
 476         assert_spin_locked(&rs->rs_lock);
 477
 478         BUG_ON(rs->rs_snd_bytes < len);
 479         rs->rs_snd_bytes -= len;
 480
 481         if (rs->rs_snd_bytes == 0)
 482                 rds_stats_inc(s_send_queue_empty);
 483 }
 484
 485 static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
 486                                     is_acked_func is_acked)
 487 {
 488         if (is_acked)
 489                 return is_acked(rm, ack);
 490         return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
 491 }
 492
 493 /*
 494  * This is pretty similar to what happens below in the ACK
 495  * handling code - except that we call here as soon as we get
 496  * the IB send completion on the RDMA op and the accompanying
 497  * message.
 498  */
 499 void rds_rdma_send_complete(struct rds_message *rm, int status)
 500 {
 501         struct rds_sock *rs = NULL;
 502         struct rm_rdma_op *ro;
 503         struct rds_notifier *notifier;
 504         unsigned long flags;
 505
 506         spin_lock_irqsave(&rm->m_rs_lock, flags);
 507
 508         ro = &rm->rdma;
 509         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
 510             ro->op_active && ro->op_notify && ro->op_notifier) {
 511                 notifier = ro->op_notifier;
 512                 rs = rm->m_rs;
 513                 sock_hold(rds_rs_to_sk(rs));
 514
 515                 notifier->n_status = status;
 516                 spin_lock(&rs->rs_lock);
 517                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
 518                 spin_unlock(&rs->rs_lock);
 519
 520                 ro->op_notifier = NULL;
 521         }
 522
 523         spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 524
 525         if (rs) {
 526                 rds_wake_sk_sleep(rs);
 527                 sock_put(rds_rs_to_sk(rs));
 528         }
 529 }
 530 EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
 531
 532 /*
 533  * Just like above, except looks at atomic op
 534  */
 535 void rds_atomic_send_complete(struct rds_message *rm, int status)
 536 {
 537         struct rds_sock *rs = NULL;
 538         struct rm_atomic_op *ao;
 539         struct rds_notifier *notifier;
 540         unsigned long flags;
 541
 542         spin_lock_irqsave(&rm->m_rs_lock, flags);
 543
 544         ao = &rm->atomic;
 545         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
 546             && ao->op_active && ao->op_notify && ao->op_notifier) {
 547                 notifier = ao->op_notifier;
 548                 rs = rm->m_rs;
 549                 sock_hold(rds_rs_to_sk(rs));
 550
 551                 notifier->n_status = status;
 552                 spin_lock(&rs->rs_lock);
 553                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
 554                 spin_unlock(&rs->rs_lock);
 555
 556                 ao->op_notifier = NULL;
 557         }
 558
 559         spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 560
 561         if (rs) {
 562                 rds_wake_sk_sleep(rs);
 563                 sock_put(rds_rs_to_sk(rs));
 564         }
 565 }
 566 EXPORT_SYMBOL_GPL(rds_atomic_send_complete);
 567
 568 /*
 569  * This is the same as rds_rdma_send_complete except we
 570  * don't do any locking - we have all the ingredients (message,
 571  * socket, socket lock) and can just move the notifier.
 572  */
 573 static inline void
 574 __rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
 575 {
 576         struct rm_rdma_op *ro;
 577         struct rm_atomic_op *ao;
 578
 579         ro = &rm->rdma;
 580         if (ro->op_active && ro->op_notify && ro->op_notifier) {
 581                 ro->op_notifier->n_status = status;
 582                 list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue);
 583                 ro->op_notifier = NULL;
 584         }
 585
 586         ao = &rm->atomic;
 587         if (ao->op_active && ao->op_notify && ao->op_notifier) {
 588                 ao->op_notifier->n_status = status;
 589                 list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue);
 590                 ao->op_notifier = NULL;
 591         }
 592
 593         /* No need to wake the app - caller does this */
 594 }
 595
 596 /*
 597  * This removes messages from the socket's list if they're on it.  The list
 598  * argument must be private to the caller, we must be able to modify it
 599  * without locks.  The messages must have a reference held for their
 600  * position on the list.  This function will drop that reference after
 601  * removing the messages from the 'messages' list regardless of if it found
 602  * the messages on the socket list or not.
 603  */
 604 static void rds_send_remove_from_sock(struct list_head *messages, int status)
 605 {
 606         unsigned long flags;
 607         struct rds_sock *rs = NULL;
 608         struct rds_message *rm;
 609
 610         while (!list_empty(messages)) {
 611                 int was_on_sock = 0;
 612
 613                 rm = list_entry(messages->next, struct rds_message,
 614                                 m_conn_item);
 615                 list_del_init(&rm->m_conn_item);
 616
 617                 /*
 618                  * If we see this flag cleared then we're *sure* that someone
 619                  * else beat us to removing it from the sock.  If we race
 620                  * with their flag update we'll get the lock and then really
 621                  * see that the flag has been cleared.
 622                  *
 623                  * The message spinlock makes sure nobody clears rm->m_rs
 624                  * while we're messing with it. It does not prevent the
 625                  * message from being removed from the socket, though.
 626                  */
 627                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 628                 if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags))
 629                         goto unlock_and_drop;
 630
 631                 if (rs != rm->m_rs) {
 632                         if (rs) {
 633                                 rds_wake_sk_sleep(rs);
 634                                 sock_put(rds_rs_to_sk(rs));
 635                         }
 636                         rs = rm->m_rs;
 637                         if (rs)
 638                                 sock_hold(rds_rs_to_sk(rs));
 639                 }
 640                 if (!rs)
 641                         goto unlock_and_drop;
 642                 spin_lock(&rs->rs_lock);
 643
 644                 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
 645                         struct rm_rdma_op *ro = &rm->rdma;
 646                         struct rds_notifier *notifier;
 647
 648                         list_del_init(&rm->m_sock_item);
 649                         rds_send_sndbuf_remove(rs, rm);
 650
 651                         if (ro->op_active && ro->op_notifier &&
 652                                (ro->op_notify || (ro->op_recverr && status))) {
 653                                 notifier = ro->op_notifier;
 654                                 list_add_tail(&notifier->n_list,
 655                                                 &rs->rs_notify_queue);
 656                                 if (!notifier->n_status)
 657                                         notifier->n_status = status;
 658                                 rm->rdma.op_notifier = NULL;
 659                         }
 660                         was_on_sock = 1;
 661                 }
 662                 spin_unlock(&rs->rs_lock);
 663
 664 unlock_and_drop:
 665                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 666                 rds_message_put(rm);
 667                 if (was_on_sock)
 668                         rds_message_put(rm);
 669         }
 670
 671         if (rs) {
 672                 rds_wake_sk_sleep(rs);
 673                 sock_put(rds_rs_to_sk(rs));
 674         }
 675 }
 676
 677 /*
 678  * Transports call here when they've determined that the receiver queued
 679  * messages up to, and including, the given sequence number.  Messages are
 680  * moved to the retrans queue when rds_send_xmit picks them off the send
 681  * queue. This means that in the TCP case, the message may not have been
 682  * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
 683  * checks the RDS_MSG_HAS_ACK_SEQ bit.
 684  */
 685 void rds_send_path_drop_acked(struct rds_conn_path *cp, u64 ack,
 686                               is_acked_func is_acked)
 687 {
 688         struct rds_message *rm, *tmp;
 689         unsigned long flags;
 690         LIST_HEAD(list);
 691
 692         spin_lock_irqsave(&cp->cp_lock, flags);
 693
 694         list_for_each_entry_safe(rm, tmp, &cp->cp_retrans, m_conn_item) {
 695                 if (!rds_send_is_acked(rm, ack, is_acked))
 696                         break;
 697
 698                 list_move(&rm->m_conn_item, &list);
 699                 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
 700         }
 701
 702         /* order flag updates with spin locks */
 703         if (!list_empty(&list))
 704                 smp_mb__after_atomic();
 705
 706         spin_unlock_irqrestore(&cp->cp_lock, flags);
 707
 708         /* now remove the messages from the sock list as needed */
 709         rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
 710 }
 711 EXPORT_SYMBOL_GPL(rds_send_path_drop_acked);
 712
 713 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
 714                          is_acked_func is_acked)
 715 {
 716         WARN_ON(conn->c_trans->t_mp_capable);
 717         rds_send_path_drop_acked(&conn->c_path[0], ack, is_acked);
 718 }
 719 EXPORT_SYMBOL_GPL(rds_send_drop_acked);
 720
 721 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in6 *dest)
 722 {
 723         struct rds_message *rm, *tmp;
 724         struct rds_connection *conn;
 725         struct rds_conn_path *cp;
 726         unsigned long flags;
 727         LIST_HEAD(list);
 728
 729         /* get all the messages we're dropping under the rs lock */
 730         spin_lock_irqsave(&rs->rs_lock, flags);
 731
 732         list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) {
 733                 if (dest &&
 734                     (!ipv6_addr_equal(&dest->sin6_addr, &rm->m_daddr) ||
 735                      dest->sin6_port != rm->m_inc.i_hdr.h_dport))
 736                         continue;
 737
 738                 list_move(&rm->m_sock_item, &list);
 739                 rds_send_sndbuf_remove(rs, rm);
 740                 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
 741         }
 742
 743         /* order flag updates with the rs lock */
 744         smp_mb__after_atomic();
 745
 746         spin_unlock_irqrestore(&rs->rs_lock, flags);
 747
 748         if (list_empty(&list))
 749                 return;
 750
 751         /* Remove the messages from the conn */
 752         list_for_each_entry(rm, &list, m_sock_item) {
 753
 754                 conn = rm->m_inc.i_conn;
 755                 if (conn->c_trans->t_mp_capable)
 756                         cp = rm->m_inc.i_conn_path;
 757                 else
 758                         cp = &conn->c_path[0];
 759
 760                 spin_lock_irqsave(&cp->cp_lock, flags);
 761                 /*
 762                  * Maybe someone else beat us to removing rm from the conn.
 763                  * If we race with their flag update we'll get the lock and
 764                  * then really see that the flag has been cleared.
 765                  */
 766                 if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
 767                         spin_unlock_irqrestore(&cp->cp_lock, flags);
 768                         continue;
 769                 }
 770                 list_del_init(&rm->m_conn_item);
 771                 spin_unlock_irqrestore(&cp->cp_lock, flags);
 772
 773                 /*
 774                  * Couldn't grab m_rs_lock in top loop (lock ordering),
 775                  * but we can now.
 776                  */
 777                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 778
 779                 spin_lock(&rs->rs_lock);
 780                 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
 781                 spin_unlock(&rs->rs_lock);
 782
 783                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 784
 785                 rds_message_put(rm);
 786         }
 787
 788         rds_wake_sk_sleep(rs);
 789
 790         while (!list_empty(&list)) {
 791                 rm = list_entry(list.next, struct rds_message, m_sock_item);
 792                 list_del_init(&rm->m_sock_item);
 793                 rds_message_wait(rm);
 794
 795                 /* just in case the code above skipped this message
 796                  * because RDS_MSG_ON_CONN wasn't set, run it again here
 797                  * taking m_rs_lock is the only thing that keeps us
 798                  * from racing with ack processing.
 799                  */
 800                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 801
 802                 spin_lock(&rs->rs_lock);
 803                 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
 804                 spin_unlock(&rs->rs_lock);
 805
 806                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 807
 808                 rds_message_put(rm);
 809         }
 810 }
 811
 812 /*
 813  * we only want this to fire once so we use the callers 'queued'.  It's
 814  * possible that another thread can race with us and remove the
 815  * message from the flow with RDS_CANCEL_SENT_TO.
 816  */
 817 static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
 818                              struct rds_conn_path *cp,
 819                              struct rds_message *rm, __be16 sport,
 820                              __be16 dport, int *queued)
 821 {
 822         unsigned long flags;
 823         u32 len;
 824
 825         if (*queued)
 826                 goto out;
 827
 828         len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
 829
 830         /* this is the only place which holds both the socket's rs_lock
 831          * and the connection's c_lock */
 832         spin_lock_irqsave(&rs->rs_lock, flags);
 833
 834         /*
 835          * If there is a little space in sndbuf, we don't queue anything,
 836          * and userspace gets -EAGAIN. But poll() indicates there's send
 837          * room. This can lead to bad behavior (spinning) if snd_bytes isn't
 838          * freed up by incoming acks. So we check the *old* value of
 839          * rs_snd_bytes here to allow the last msg to exceed the buffer,
 840          * and poll() now knows no more data can be sent.
 841          */
 842         if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
 843                 rs->rs_snd_bytes += len;
 844
 845                 /* let recv side know we are close to send space exhaustion.
 846                  * This is probably not the optimal way to do it, as this
 847                  * means we set the flag on *all* messages as soon as our
 848                  * throughput hits a certain threshold.
 849                  */
 850                 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
 851                         set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
 852
 853                 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
 854                 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
 855                 rds_message_addref(rm);
 856                 sock_hold(rds_rs_to_sk(rs));
 857                 rm->m_rs = rs;
 858
 859                 /* The code ordering is a little weird, but we're
 860                    trying to minimize the time we hold c_lock */
 861                 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
 862                 rm->m_inc.i_conn = conn;
 863                 rm->m_inc.i_conn_path = cp;
 864                 rds_message_addref(rm);
 865
 866                 spin_lock(&cp->cp_lock);
 867                 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(cp->cp_next_tx_seq++);
 868                 list_add_tail(&rm->m_conn_item, &cp->cp_send_queue);
 869                 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
 870                 spin_unlock(&cp->cp_lock);
 871
 872                 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
 873                          rm, len, rs, rs->rs_snd_bytes,
 874                          (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
 875
 876                 *queued = 1;
 877         }
 878
 879         spin_unlock_irqrestore(&rs->rs_lock, flags);
 880 out:
 881         return *queued;
 882 }
 883
 884 /*
 885  * rds_message is getting to be quite complicated, and we'd like to allocate
 886  * it all in one go. This figures out how big it needs to be up front.
 887  */
 888 static int rds_rm_size(struct msghdr *msg, int num_sgs,
 889                        struct rds_iov_vector_arr *vct)
 890 {
 891         struct cmsghdr *cmsg;
 892         int size = 0;
 893         int cmsg_groups = 0;
 894         int retval;
 895         bool zcopy_cookie = false;
 896         struct rds_iov_vector *iov, *tmp_iov;
 897
 898         if (num_sgs < 0)
 899                 return -EINVAL;
 900
 901         for_each_cmsghdr(cmsg, msg) {
 902                 if (!CMSG_OK(msg, cmsg))
 903                         return -EINVAL;
 904
 905                 if (cmsg->cmsg_level != SOL_RDS)
 906                         continue;
 907
 908                 switch (cmsg->cmsg_type) {
 909                 case RDS_CMSG_RDMA_ARGS:
 910                         if (vct->indx >= vct->len) {
 911                                 vct->len += vct->incr;
 912                                 tmp_iov =
 913                                         krealloc(vct->vec,
 914                                                  vct->len *
 915                                                  sizeof(struct rds_iov_vector),
 916                                                  GFP_KERNEL);
 917                                 if (!tmp_iov) {
 918                                         vct->len -= vct->incr;
 919                                         return -ENOMEM;
 920                                 }
 921                                 vct->vec = tmp_iov;
 922                         }
 923                         iov = &vct->vec[vct->indx];
 924                         memset(iov, 0, sizeof(struct rds_iov_vector));
 925                         vct->indx++;
 926                         cmsg_groups |= 1;
 927                         retval = rds_rdma_extra_size(CMSG_DATA(cmsg), iov);
 928                         if (retval < 0)
 929                                 return retval;
 930                         size += retval;
 931
 932                         break;
 933
 934                 case RDS_CMSG_ZCOPY_COOKIE:
 935                         zcopy_cookie = true;
 936                         fallthrough;
 937
 938                 case RDS_CMSG_RDMA_DEST:
 939                 case RDS_CMSG_RDMA_MAP:
 940                         cmsg_groups |= 2;
 941                         /* these are valid but do no add any size */
 942                         break;
 943
 944                 case RDS_CMSG_ATOMIC_CSWP:
 945                 case RDS_CMSG_ATOMIC_FADD:
 946                 case RDS_CMSG_MASKED_ATOMIC_CSWP:
 947                 case RDS_CMSG_MASKED_ATOMIC_FADD:
 948                         cmsg_groups |= 1;
 949                         size += sizeof(struct scatterlist);
 950                         break;
 951
 952                 default:
 953                         return -EINVAL;
 954                 }
 955
 956         }
 957
 958         if ((msg->msg_flags & MSG_ZEROCOPY) && !zcopy_cookie)
 959                 return -EINVAL;
 960
 961         size += num_sgs * sizeof(struct scatterlist);
 962
 963         /* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */
 964         if (cmsg_groups == 3)
 965                 return -EINVAL;
 966
 967         return size;
 968 }
 969
 970 static int rds_cmsg_zcopy(struct rds_sock *rs, struct rds_message *rm,
 971                           struct cmsghdr *cmsg)
 972 {
 973         u32 *cookie;
 974
 975         if (cmsg->cmsg_len < CMSG_LEN(sizeof(*cookie)) ||
 976             !rm->data.op_mmp_znotifier)
 977                 return -EINVAL;
 978         cookie = CMSG_DATA(cmsg);
 979         rm->data.op_mmp_znotifier->z_cookie = *cookie;
 980         return 0;
 981 }
 982
 983 static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
 984                          struct msghdr *msg, int *allocated_mr,
 985                          struct rds_iov_vector_arr *vct)
 986 {
 987         struct cmsghdr *cmsg;
 988         int ret = 0, ind = 0;
 989
 990         for_each_cmsghdr(cmsg, msg) {
 991                 if (!CMSG_OK(msg, cmsg))
 992                         return -EINVAL;
 993
 994                 if (cmsg->cmsg_level != SOL_RDS)
 995                         continue;
 996
 997                 /* As a side effect, RDMA_DEST and RDMA_MAP will set
 998                  * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
 999                  */
1000                 switch (cmsg->cmsg_type) {
1001                 case RDS_CMSG_RDMA_ARGS:
1002                         if (ind >= vct->indx)
1003                                 return -ENOMEM;
1004                         ret = rds_cmsg_rdma_args(rs, rm, cmsg, &vct->vec[ind]);
1005                         ind++;
1006                         break;
1007
1008                 case RDS_CMSG_RDMA_DEST:
1009                         ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
1010                         break;
1011
1012                 case RDS_CMSG_RDMA_MAP:
1013                         ret = rds_cmsg_rdma_map(rs, rm, cmsg);
1014                         if (!ret)
1015                                 *allocated_mr = 1;
1016                         else if (ret == -ENODEV)
1017                                 /* Accommodate the get_mr() case which can fail
1018                                  * if connection isn't established yet.
1019                                  */
1020                                 ret = -EAGAIN;
1021                         break;
1022                 case RDS_CMSG_ATOMIC_CSWP:
1023                 case RDS_CMSG_ATOMIC_FADD:
1024                 case RDS_CMSG_MASKED_ATOMIC_CSWP:
1025                 case RDS_CMSG_MASKED_ATOMIC_FADD:
1026                         ret = rds_cmsg_atomic(rs, rm, cmsg);
1027                         break;
1028
1029                 case RDS_CMSG_ZCOPY_COOKIE:
1030                         ret = rds_cmsg_zcopy(rs, rm, cmsg);
1031                         break;
1032
1033                 default:
1034                         return -EINVAL;
1035                 }
1036
1037                 if (ret)
1038                         break;
1039         }
1040
1041         return ret;
1042 }
1043
1044 static int rds_send_mprds_hash(struct rds_sock *rs,
1045                                struct rds_connection *conn, int nonblock)
1046 {
1047         int hash;
1048
1049         if (conn->c_npaths == 0)
1050                 hash = RDS_MPATH_HASH(rs, RDS_MPATH_WORKERS);
1051         else
1052                 hash = RDS_MPATH_HASH(rs, conn->c_npaths);
1053         if (conn->c_npaths == 0 && hash != 0) {
1054                 rds_send_ping(conn, 0);
1055
1056                 /* The underlying connection is not up yet.  Need to wait
1057                  * until it is up to be sure that the non-zero c_path can be
1058                  * used.  But if we are interrupted, we have to use the zero
1059                  * c_path in case the connection ends up being non-MP capable.
1060                  */
1061                 if (conn->c_npaths == 0) {
1062                         /* Cannot wait for the connection be made, so just use
1063                          * the base c_path.
1064                          */
1065                         if (nonblock)
1066                                 return 0;
1067                         if (wait_event_interruptible(conn->c_hs_waitq,
1068                                                      conn->c_npaths != 0))
1069                                 hash = 0;
1070                 }
1071                 if (conn->c_npaths == 1)
1072                         hash = 0;
1073         }
1074         return hash;
1075 }
1076
1077 static int rds_rdma_bytes(struct msghdr *msg, size_t *rdma_bytes)
1078 {
1079         struct rds_rdma_args *args;
1080         struct cmsghdr *cmsg;
1081
1082         for_each_cmsghdr(cmsg, msg) {
1083                 if (!CMSG_OK(msg, cmsg))
1084                         return -EINVAL;
1085
1086                 if (cmsg->cmsg_level != SOL_RDS)
1087                         continue;
1088
1089                 if (cmsg->cmsg_type == RDS_CMSG_RDMA_ARGS) {
1090                         if (cmsg->cmsg_len <
1091                             CMSG_LEN(sizeof(struct rds_rdma_args)))
1092                                 return -EINVAL;
1093                         args = CMSG_DATA(cmsg);
1094                         *rdma_bytes += args->remote_vec.bytes;
1095                 }
1096         }
1097         return 0;
1098 }
1099
1100 int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len)
1101 {
1102         struct sock *sk = sock->sk;
1103         struct rds_sock *rs = rds_sk_to_rs(sk);
1104         DECLARE_SOCKADDR(struct sockaddr_in6 *, sin6, msg->msg_name);
1105         DECLARE_SOCKADDR(struct sockaddr_in *, usin, msg->msg_name);
1106         __be16 dport;
1107         struct rds_message *rm = NULL;
1108         struct rds_connection *conn;
1109         int ret = 0;
1110         int queued = 0, allocated_mr = 0;
1111         int nonblock = msg->msg_flags & MSG_DONTWAIT;
1112         long timeo = sock_sndtimeo(sk, nonblock);
1113         struct rds_conn_path *cpath;
1114         struct in6_addr daddr;
1115         __u32 scope_id = 0;
1116         size_t rdma_payload_len = 0;
1117         bool zcopy = ((msg->msg_flags & MSG_ZEROCOPY) &&
1118                       sock_flag(rds_rs_to_sk(rs), SOCK_ZEROCOPY));
1119         int num_sgs = DIV_ROUND_UP(payload_len, PAGE_SIZE);
1120         int namelen;
1121         struct rds_iov_vector_arr vct;
1122         int ind;
1123
1124         memset(&vct, 0, sizeof(vct));
1125
1126         /* expect 1 RDMA CMSG per rds_sendmsg. can still grow if more needed. */
1127         vct.incr = 1;
1128
1129         /* Mirror Linux UDP mirror of BSD error message compatibility */
1130         /* XXX: Perhaps MSG_MORE someday */
1131         if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT | MSG_ZEROCOPY)) {
1132                 ret = -EOPNOTSUPP;
1133                 goto out;
1134         }
1135
1136         namelen = msg->msg_namelen;
1137         if (namelen != 0) {
1138                 if (namelen < sizeof(*usin)) {
1139                         ret = -EINVAL;
1140                         goto out;
1141                 }
1142                 switch (usin->sin_family) {
1143                 case AF_INET:
1144                         if (usin->sin_addr.s_addr == htonl(INADDR_ANY) ||
1145                             usin->sin_addr.s_addr == htonl(INADDR_BROADCAST) ||
1146                             ipv4_is_multicast(usin->sin_addr.s_addr)) {
1147                                 ret = -EINVAL;
1148                                 goto out;
1149                         }
1150                         ipv6_addr_set_v4mapped(usin->sin_addr.s_addr, &daddr);
1151                         dport = usin->sin_port;
1152                         break;
1153
1154 #if IS_ENABLED(CONFIG_IPV6)
1155                 case AF_INET6: {
1156                         int addr_type;
1157
1158                         if (namelen < sizeof(*sin6)) {
1159                                 ret = -EINVAL;
1160                                 goto out;
1161                         }
1162                         addr_type = ipv6_addr_type(&sin6->sin6_addr);
1163                         if (!(addr_type & IPV6_ADDR_UNICAST)) {
1164                                 __be32 addr4;
1165
1166                                 if (!(addr_type & IPV6_ADDR_MAPPED)) {
1167                                         ret = -EINVAL;
1168                                         goto out;
1169                                 }
1170
1171                                 /* It is a mapped address.  Need to do some
1172                                  * sanity checks.
1173                                  */
1174                                 addr4 = sin6->sin6_addr.s6_addr32[3];
1175                                 if (addr4 == htonl(INADDR_ANY) ||
1176                                     addr4 == htonl(INADDR_BROADCAST) ||
1177                                     ipv4_is_multicast(addr4)) {
1178                                         ret = -EINVAL;
1179                                         goto out;
1180                                 }
1181                         }
1182                         if (addr_type & IPV6_ADDR_LINKLOCAL) {
1183                                 if (sin6->sin6_scope_id == 0) {
1184                                         ret = -EINVAL;
1185                                         goto out;
1186                                 }
1187                                 scope_id = sin6->sin6_scope_id;
1188                         }
1189
1190                         daddr = sin6->sin6_addr;
1191                         dport = sin6->sin6_port;
1192                         break;
1193                 }
1194 #endif
1195
1196                 default:
1197                         ret = -EINVAL;
1198                         goto out;
1199                 }
1200         } else {
1201                 /* We only care about consistency with ->connect() */
1202                 lock_sock(sk);
1203                 daddr = rs->rs_conn_addr;
1204                 dport = rs->rs_conn_port;
1205                 scope_id = rs->rs_bound_scope_id;
1206                 release_sock(sk);
1207         }
1208
1209         lock_sock(sk);
1210         if (ipv6_addr_any(&rs->rs_bound_addr) || ipv6_addr_any(&daddr)) {
1211                 release_sock(sk);
1212                 ret = -ENOTCONN;
1213                 goto out;
1214         } else if (namelen != 0) {
1215                 /* Cannot send to an IPv4 address using an IPv6 source
1216                  * address and cannot send to an IPv6 address using an
1217                  * IPv4 source address.
1218                  */
1219                 if (ipv6_addr_v4mapped(&daddr) ^
1220                     ipv6_addr_v4mapped(&rs->rs_bound_addr)) {
1221                         release_sock(sk);
1222                         ret = -EOPNOTSUPP;
1223                         goto out;
1224                 }
1225                 /* If the socket is already bound to a link local address,
1226                  * it can only send to peers on the same link.  But allow
1227                  * communicating between link local and non-link local address.
1228                  */
1229                 if (scope_id != rs->rs_bound_scope_id) {
1230                         if (!scope_id) {
1231                                 scope_id = rs->rs_bound_scope_id;
1232                         } else if (rs->rs_bound_scope_id) {
1233                                 release_sock(sk);
1234                                 ret = -EINVAL;
1235                                 goto out;
1236                         }
1237                 }
1238         }
1239         release_sock(sk);
1240
1241         ret = rds_rdma_bytes(msg, &rdma_payload_len);
1242         if (ret)
1243                 goto out;
1244
1245         if (max_t(size_t, payload_len, rdma_payload_len) > RDS_MAX_MSG_SIZE) {
1246                 ret = -EMSGSIZE;
1247                 goto out;
1248         }
1249
1250         if (payload_len > rds_sk_sndbuf(rs)) {
1251                 ret = -EMSGSIZE;
1252                 goto out;
1253         }
1254
1255         if (zcopy) {
1256                 if (rs->rs_transport->t_type != RDS_TRANS_TCP) {
1257                         ret = -EOPNOTSUPP;
1258                         goto out;
1259                 }
1260                 num_sgs = iov_iter_npages(&msg->msg_iter, INT_MAX);
1261         }
1262         /* size of rm including all sgs */
1263         ret = rds_rm_size(msg, num_sgs, &vct);
1264         if (ret < 0)
1265                 goto out;
1266
1267         rm = rds_message_alloc(ret, GFP_KERNEL);
1268         if (!rm) {
1269                 ret = -ENOMEM;
1270                 goto out;
1271         }
1272
1273         /* Attach data to the rm */
1274         if (payload_len) {
1275                 rm->data.op_sg = rds_message_alloc_sgs(rm, num_sgs);
1276                 if (IS_ERR(rm->data.op_sg)) {
1277                         ret = PTR_ERR(rm->data.op_sg);
1278                         goto out;
1279                 }
1280                 ret = rds_message_copy_from_user(rm, &msg->msg_iter, zcopy);
1281                 if (ret)
1282                         goto out;
1283         }
1284         rm->data.op_active = 1;
1285
1286         rm->m_daddr = daddr;
1287
1288         /* rds_conn_create has a spinlock that runs with IRQ off.
1289          * Caching the conn in the socket helps a lot. */
1290         if (rs->rs_conn && ipv6_addr_equal(&rs->rs_conn->c_faddr, &daddr) &&
1291             rs->rs_tos == rs->rs_conn->c_tos) {
1292                 conn = rs->rs_conn;
1293         } else {
1294                 conn = rds_conn_create_outgoing(sock_net(sock->sk),
1295                                                 &rs->rs_bound_addr, &daddr,
1296                                                 rs->rs_transport, rs->rs_tos,
1297                                                 sock->sk->sk_allocation,
1298                                                 scope_id);
1299                 if (IS_ERR(conn)) {
1300                         ret = PTR_ERR(conn);
1301                         goto out;
1302                 }
1303                 rs->rs_conn = conn;
1304         }
1305
1306         if (conn->c_trans->t_mp_capable)
1307                 cpath = &conn->c_path[rds_send_mprds_hash(rs, conn, nonblock)];
1308         else
1309                 cpath = &conn->c_path[0];
1310
1311         rm->m_conn_path = cpath;
1312
1313         /* Parse any control messages the user may have included. */
1314         ret = rds_cmsg_send(rs, rm, msg, &allocated_mr, &vct);
1315         if (ret)
1316                 goto out;
1317
1318         if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) {
1319                 printk_ratelimited(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
1320                                &rm->rdma, conn->c_trans->xmit_rdma);
1321                 ret = -EOPNOTSUPP;
1322                 goto out;
1323         }
1324
1325         if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
1326                 printk_ratelimited(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
1327                                &rm->atomic, conn->c_trans->xmit_atomic);
1328                 ret = -EOPNOTSUPP;
1329                 goto out;
1330         }
1331
1332         if (rds_destroy_pending(conn)) {
1333                 ret = -EAGAIN;
1334                 goto out;
1335         }
1336
1337         if (rds_conn_path_down(cpath))
1338                 rds_check_all_paths(conn);
1339
1340         ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
1341         if (ret) {
1342                 rs->rs_seen_congestion = 1;
1343                 goto out;
1344         }
1345         while (!rds_send_queue_rm(rs, conn, cpath, rm, rs->rs_bound_port,
1346                                   dport, &queued)) {
1347                 rds_stats_inc(s_send_queue_full);
1348
1349                 if (nonblock) {
1350                         ret = -EAGAIN;
1351                         goto out;
1352                 }
1353
1354                 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
1355                                         rds_send_queue_rm(rs, conn, cpath, rm,
1356                                                           rs->rs_bound_port,
1357                                                           dport,
1358                                                           &queued),
1359                                         timeo);
1360                 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
1361                 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
1362                         continue;
1363
1364                 ret = timeo;
1365                 if (ret == 0)
1366                         ret = -ETIMEDOUT;
1367                 goto out;
1368         }
1369
1370         /*
1371          * By now we've committed to the send.  We reuse rds_send_worker()
1372          * to retry sends in the rds thread if the transport asks us to.
1373          */
1374         rds_stats_inc(s_send_queued);
1375
1376         ret = rds_send_xmit(cpath);
1377         if (ret == -ENOMEM || ret == -EAGAIN) {
1378                 ret = 0;
1379                 rcu_read_lock();
1380                 if (rds_destroy_pending(cpath->cp_conn))
1381                         ret = -ENETUNREACH;
1382                 else
1383                         queue_delayed_work(rds_wq, &cpath->cp_send_w, 1);
1384                 rcu_read_unlock();
1385         }
1386         if (ret)
1387                 goto out;
1388         rds_message_put(rm);
1389
1390         for (ind = 0; ind < vct.indx; ind++)
1391                 kfree(vct.vec[ind].iov);
1392         kfree(vct.vec);
1393
1394         return payload_len;
1395
1396 out:
1397         for (ind = 0; ind < vct.indx; ind++)
1398                 kfree(vct.vec[ind].iov);
1399         kfree(vct.vec);
1400
1401         /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1402          * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1403          * or in any other way, we need to destroy the MR again */
1404         if (allocated_mr)
1405                 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
1406
1407         if (rm)
1408                 rds_message_put(rm);
1409         return ret;
1410 }
1411
1412 /*
1413  * send out a probe. Can be shared by rds_send_ping,
1414  * rds_send_pong, rds_send_hb.
1415  * rds_send_hb should use h_flags
1416  *   RDS_FLAG_HB_PING|RDS_FLAG_ACK_REQUIRED
1417  * or
1418  *   RDS_FLAG_HB_PONG|RDS_FLAG_ACK_REQUIRED
1419  */
1420 static int
1421 rds_send_probe(struct rds_conn_path *cp, __be16 sport,
1422                __be16 dport, u8 h_flags)
1423 {
1424         struct rds_message *rm;
1425         unsigned long flags;
1426         int ret = 0;
1427
1428         rm = rds_message_alloc(0, GFP_ATOMIC);
1429         if (!rm) {
1430                 ret = -ENOMEM;
1431                 goto out;
1432         }
1433
1434         rm->m_daddr = cp->cp_conn->c_faddr;
1435         rm->data.op_active = 1;
1436
1437         rds_conn_path_connect_if_down(cp);
1438
1439         ret = rds_cong_wait(cp->cp_conn->c_fcong, dport, 1, NULL);
1440         if (ret)
1441                 goto out;
1442
1443         spin_lock_irqsave(&cp->cp_lock, flags);
1444         list_add_tail(&rm->m_conn_item, &cp->cp_send_queue);
1445         set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
1446         rds_message_addref(rm);
1447         rm->m_inc.i_conn = cp->cp_conn;
1448         rm->m_inc.i_conn_path = cp;
1449
1450         rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport,
1451                                     cp->cp_next_tx_seq);
1452         rm->m_inc.i_hdr.h_flags |= h_flags;
1453         cp->cp_next_tx_seq++;
1454
1455         if (RDS_HS_PROBE(be16_to_cpu(sport), be16_to_cpu(dport)) &&
1456             cp->cp_conn->c_trans->t_mp_capable) {
1457                 u16 npaths = cpu_to_be16(RDS_MPATH_WORKERS);
1458                 u32 my_gen_num = cpu_to_be32(cp->cp_conn->c_my_gen_num);
1459
1460                 rds_message_add_extension(&rm->m_inc.i_hdr,
1461                                           RDS_EXTHDR_NPATHS, &npaths,
1462                                           sizeof(npaths));
1463                 rds_message_add_extension(&rm->m_inc.i_hdr,
1464                                           RDS_EXTHDR_GEN_NUM,
1465                                           &my_gen_num,
1466                                           sizeof(u32));
1467         }
1468         spin_unlock_irqrestore(&cp->cp_lock, flags);
1469
1470         rds_stats_inc(s_send_queued);
1471         rds_stats_inc(s_send_pong);
1472
1473         /* schedule the send work on rds_wq */
1474         rcu_read_lock();
1475         if (!rds_destroy_pending(cp->cp_conn))
1476                 queue_delayed_work(rds_wq, &cp->cp_send_w, 1);
1477         rcu_read_unlock();
1478
1479         rds_message_put(rm);
1480         return 0;
1481
1482 out:
1483         if (rm)
1484                 rds_message_put(rm);
1485         return ret;
1486 }
1487
1488 int
1489 rds_send_pong(struct rds_conn_path *cp, __be16 dport)
1490 {
1491         return rds_send_probe(cp, 0, dport, 0);
1492 }
1493
1494 void
1495 rds_send_ping(struct rds_connection *conn, int cp_index)
1496 {
1497         unsigned long flags;
1498         struct rds_conn_path *cp = &conn->c_path[cp_index];
1499
1500         spin_lock_irqsave(&cp->cp_lock, flags);
1501         if (conn->c_ping_triggered) {
1502                 spin_unlock_irqrestore(&cp->cp_lock, flags);
1503                 return;
1504         }
1505         conn->c_ping_triggered = 1;
1506         spin_unlock_irqrestore(&cp->cp_lock, flags);
1507         rds_send_probe(cp, cpu_to_be16(RDS_FLAG_PROBE_PORT), 0, 0);
1508 }
1509 EXPORT_SYMBOL_GPL(rds_send_ping);