net/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(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         int same_rm = 0;
 149
 150 restart:
 151         batch_count = 0;
 152
 153         /*
 154          * sendmsg calls here after having queued its message on the send
 155          * queue.  We only have one task feeding the connection at a time.  If
 156          * another thread is already feeding the queue then we back off.  This
 157          * avoids blocking the caller and trading per-connection data between
 158          * caches per message.
 159          */
 160         if (!acquire_in_xmit(cp)) {
 161                 rds_stats_inc(s_send_lock_contention);
 162                 ret = -ENOMEM;
 163                 goto out;
 164         }
 165
 166         if (rds_destroy_pending(cp->cp_conn)) {
 167                 release_in_xmit(cp);
 168                 ret = -ENETUNREACH; /* dont requeue send work */
 169                 goto out;
 170         }
 171
 172         /*
 173          * we record the send generation after doing the xmit acquire.
 174          * if someone else manages to jump in and do some work, we'll use
 175          * this to avoid a goto restart farther down.
 176          *
 177          * The acquire_in_xmit() check above ensures that only one
 178          * caller can increment c_send_gen at any time.
 179          */
 180         send_gen = READ_ONCE(cp->cp_send_gen) + 1;
 181         WRITE_ONCE(cp->cp_send_gen, send_gen);
 182
 183         /*
 184          * rds_conn_shutdown() sets the conn state and then tests RDS_IN_XMIT,
 185          * we do the opposite to avoid races.
 186          */
 187         if (!rds_conn_path_up(cp)) {
 188                 release_in_xmit(cp);
 189                 ret = 0;
 190                 goto out;
 191         }
 192
 193         if (conn->c_trans->xmit_path_prepare)
 194                 conn->c_trans->xmit_path_prepare(cp);
 195
 196         /*
 197          * spin trying to push headers and data down the connection until
 198          * the connection doesn't make forward progress.
 199          */
 200         while (1) {
 201
 202                 rm = cp->cp_xmit_rm;
 203
 204                 if (!rm) {
 205                         same_rm = 0;
 206                 } else {
 207                         same_rm++;
 208                         if (same_rm >= 4096) {
 209                                 rds_stats_inc(s_send_stuck_rm);
 210                                 ret = -EAGAIN;
 211                                 break;
 212                         }
 213                 }
 214
 215                 /*
 216                  * If between sending messages, we can send a pending congestion
 217                  * map update.
 218                  */
 219                 if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) {
 220                         rm = rds_cong_update_alloc(conn);
 221                         if (IS_ERR(rm)) {
 222                                 ret = PTR_ERR(rm);
 223                                 break;
 224                         }
 225                         rm->data.op_active = 1;
 226                         rm->m_inc.i_conn_path = cp;
 227                         rm->m_inc.i_conn = cp->cp_conn;
 228
 229                         cp->cp_xmit_rm = rm;
 230                 }
 231
 232                 /*
 233                  * If not already working on one, grab the next message.
 234                  *
 235                  * cp_xmit_rm holds a ref while we're sending this message down
 236                  * the connction.  We can use this ref while holding the
 237                  * send_sem.. rds_send_reset() is serialized with it.
 238                  */
 239                 if (!rm) {
 240                         unsigned int len;
 241
 242                         batch_count++;
 243
 244                         /* we want to process as big a batch as we can, but
 245                          * we also want to avoid softlockups.  If we've been
 246                          * through a lot of messages, lets back off and see
 247                          * if anyone else jumps in
 248                          */
 249                         if (batch_count >= send_batch_count)
 250                                 goto over_batch;
 251
 252                         spin_lock_irqsave(&cp->cp_lock, flags);
 253
 254                         if (!list_empty(&cp->cp_send_queue)) {
 255                                 rm = list_entry(cp->cp_send_queue.next,
 256                                                 struct rds_message,
 257                                                 m_conn_item);
 258                                 rds_message_addref(rm);
 259
 260                                 /*
 261                                  * Move the message from the send queue to the retransmit
 262                                  * list right away.
 263                                  */
 264                                 list_move_tail(&rm->m_conn_item,
 265                                                &cp->cp_retrans);
 266                         }
 267
 268                         spin_unlock_irqrestore(&cp->cp_lock, flags);
 269
 270                         if (!rm)
 271                                 break;
 272
 273                         /* Unfortunately, the way Infiniband deals with
 274                          * RDMA to a bad MR key is by moving the entire
 275                          * queue pair to error state. We cold possibly
 276                          * recover from that, but right now we drop the
 277                          * connection.
 278                          * Therefore, we never retransmit messages with RDMA ops.
 279                          */
 280                         if (test_bit(RDS_MSG_FLUSH, &rm->m_flags) ||
 281                             (rm->rdma.op_active &&
 282                             test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))) {
 283                                 spin_lock_irqsave(&cp->cp_lock, flags);
 284                                 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags))
 285                                         list_move(&rm->m_conn_item, &to_be_dropped);
 286                                 spin_unlock_irqrestore(&cp->cp_lock, flags);
 287                                 continue;
 288                         }
 289
 290                         /* Require an ACK every once in a while */
 291                         len = ntohl(rm->m_inc.i_hdr.h_len);
 292                         if (cp->cp_unacked_packets == 0 ||
 293                             cp->cp_unacked_bytes < len) {
 294                                 set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
 295
 296                                 cp->cp_unacked_packets =
 297                                         rds_sysctl_max_unacked_packets;
 298                                 cp->cp_unacked_bytes =
 299                                         rds_sysctl_max_unacked_bytes;
 300                                 rds_stats_inc(s_send_ack_required);
 301                         } else {
 302                                 cp->cp_unacked_bytes -= len;
 303                                 cp->cp_unacked_packets--;
 304                         }
 305
 306                         cp->cp_xmit_rm = rm;
 307                 }
 308
 309                 /* The transport either sends the whole rdma or none of it */
 310                 if (rm->rdma.op_active && !cp->cp_xmit_rdma_sent) {
 311                         rm->m_final_op = &rm->rdma;
 312                         /* The transport owns the mapped memory for now.
 313                          * You can't unmap it while it's on the send queue
 314                          */
 315                         set_bit(RDS_MSG_MAPPED, &rm->m_flags);
 316                         ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
 317                         if (ret) {
 318                                 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
 319                                 wake_up_interruptible(&rm->m_flush_wait);
 320                                 break;
 321                         }
 322                         cp->cp_xmit_rdma_sent = 1;
 323
 324                 }
 325
 326                 if (rm->atomic.op_active && !cp->cp_xmit_atomic_sent) {
 327                         rm->m_final_op = &rm->atomic;
 328                         /* The transport owns the mapped memory for now.
 329                          * You can't unmap it while it's on the send queue
 330                          */
 331                         set_bit(RDS_MSG_MAPPED, &rm->m_flags);
 332                         ret = conn->c_trans->xmit_atomic(conn, &rm->atomic);
 333                         if (ret) {
 334                                 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
 335                                 wake_up_interruptible(&rm->m_flush_wait);
 336                                 break;
 337                         }
 338                         cp->cp_xmit_atomic_sent = 1;
 339
 340                 }
 341
 342                 /*
 343                  * A number of cases require an RDS header to be sent
 344                  * even if there is no data.
 345                  * We permit 0-byte sends; rds-ping depends on this.
 346                  * However, if there are exclusively attached silent ops,
 347                  * we skip the hdr/data send, to enable silent operation.
 348                  */
 349                 if (rm->data.op_nents == 0) {
 350                         int ops_present;
 351                         int all_ops_are_silent = 1;
 352
 353                         ops_present = (rm->atomic.op_active || rm->rdma.op_active);
 354                         if (rm->atomic.op_active && !rm->atomic.op_silent)
 355                                 all_ops_are_silent = 0;
 356                         if (rm->rdma.op_active && !rm->rdma.op_silent)
 357                                 all_ops_are_silent = 0;
 358
 359                         if (ops_present && all_ops_are_silent
 360                             && !rm->m_rdma_cookie)
 361                                 rm->data.op_active = 0;
 362                 }
 363
 364                 if (rm->data.op_active && !cp->cp_xmit_data_sent) {
 365                         rm->m_final_op = &rm->data;
 366
 367                         ret = conn->c_trans->xmit(conn, rm,
 368                                                   cp->cp_xmit_hdr_off,
 369                                                   cp->cp_xmit_sg,
 370                                                   cp->cp_xmit_data_off);
 371                         if (ret <= 0)
 372                                 break;
 373
 374                         if (cp->cp_xmit_hdr_off < sizeof(struct rds_header)) {
 375                                 tmp = min_t(int, ret,
 376                                             sizeof(struct rds_header) -
 377                                             cp->cp_xmit_hdr_off);
 378                                 cp->cp_xmit_hdr_off += tmp;
 379                                 ret -= tmp;
 380                         }
 381
 382                         sg = &rm->data.op_sg[cp->cp_xmit_sg];
 383                         while (ret) {
 384                                 tmp = min_t(int, ret, sg->length -
 385                                                       cp->cp_xmit_data_off);
 386                                 cp->cp_xmit_data_off += tmp;
 387                                 ret -= tmp;
 388                                 if (cp->cp_xmit_data_off == sg->length) {
 389                                         cp->cp_xmit_data_off = 0;
 390                                         sg++;
 391                                         cp->cp_xmit_sg++;
 392                                         BUG_ON(ret != 0 && cp->cp_xmit_sg ==
 393                                                rm->data.op_nents);
 394                                 }
 395                         }
 396
 397                         if (cp->cp_xmit_hdr_off == sizeof(struct rds_header) &&
 398                             (cp->cp_xmit_sg == rm->data.op_nents))
 399                                 cp->cp_xmit_data_sent = 1;
 400                 }
 401
 402                 /*
 403                  * A rm will only take multiple times through this loop
 404                  * if there is a data op. Thus, if the data is sent (or there was
 405                  * none), then we're done with the rm.
 406                  */
 407                 if (!rm->data.op_active || cp->cp_xmit_data_sent) {
 408                         cp->cp_xmit_rm = NULL;
 409                         cp->cp_xmit_sg = 0;
 410                         cp->cp_xmit_hdr_off = 0;
 411                         cp->cp_xmit_data_off = 0;
 412                         cp->cp_xmit_rdma_sent = 0;
 413                         cp->cp_xmit_atomic_sent = 0;
 414                         cp->cp_xmit_data_sent = 0;
 415
 416                         rds_message_put(rm);
 417                 }
 418         }
 419
 420 over_batch:
 421         if (conn->c_trans->xmit_path_complete)
 422                 conn->c_trans->xmit_path_complete(cp);
 423         release_in_xmit(cp);
 424
 425         /* Nuke any messages we decided not to retransmit. */
 426         if (!list_empty(&to_be_dropped)) {
 427                 /* irqs on here, so we can put(), unlike above */
 428                 list_for_each_entry(rm, &to_be_dropped, m_conn_item)
 429                         rds_message_put(rm);
 430                 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
 431         }
 432
 433         /*
 434          * Other senders can queue a message after we last test the send queue
 435          * but before we clear RDS_IN_XMIT.  In that case they'd back off and
 436          * not try and send their newly queued message.  We need to check the
 437          * send queue after having cleared RDS_IN_XMIT so that their message
 438          * doesn't get stuck on the send queue.
 439          *
 440          * If the transport cannot continue (i.e ret != 0), then it must
 441          * call us when more room is available, such as from the tx
 442          * completion handler.
 443          *
 444          * We have an extra generation check here so that if someone manages
 445          * to jump in after our release_in_xmit, we'll see that they have done
 446          * some work and we will skip our goto
 447          */
 448         if (ret == 0) {
 449                 bool raced;
 450
 451                 smp_mb();
 452                 raced = send_gen != READ_ONCE(cp->cp_send_gen);
 453
 454                 if ((test_bit(0, &conn->c_map_queued) ||
 455                     !list_empty(&cp->cp_send_queue)) && !raced) {
 456                         if (batch_count < send_batch_count)
 457                                 goto restart;
 458                         rcu_read_lock();
 459                         if (rds_destroy_pending(cp->cp_conn))
 460                                 ret = -ENETUNREACH;
 461                         else
 462                                 queue_delayed_work(rds_wq, &cp->cp_send_w, 1);
 463                         rcu_read_unlock();
 464                 } else if (raced) {
 465                         rds_stats_inc(s_send_lock_queue_raced);
 466                 }
 467         }
 468 out:
 469         return ret;
 470 }
 471 EXPORT_SYMBOL_GPL(rds_send_xmit);
 472
 473 static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
 474 {
 475         u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
 476
 477         assert_spin_locked(&rs->rs_lock);
 478
 479         BUG_ON(rs->rs_snd_bytes < len);
 480         rs->rs_snd_bytes -= len;
 481
 482         if (rs->rs_snd_bytes == 0)
 483                 rds_stats_inc(s_send_queue_empty);
 484 }
 485
 486 static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
 487                                     is_acked_func is_acked)
 488 {
 489         if (is_acked)
 490                 return is_acked(rm, ack);
 491         return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
 492 }
 493
 494 /*
 495  * This is pretty similar to what happens below in the ACK
 496  * handling code - except that we call here as soon as we get
 497  * the IB send completion on the RDMA op and the accompanying
 498  * message.
 499  */
 500 void rds_rdma_send_complete(struct rds_message *rm, int status)
 501 {
 502         struct rds_sock *rs = NULL;
 503         struct rm_rdma_op *ro;
 504         struct rds_notifier *notifier;
 505         unsigned long flags;
 506
 507         spin_lock_irqsave(&rm->m_rs_lock, flags);
 508
 509         ro = &rm->rdma;
 510         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
 511             ro->op_active && ro->op_notify && ro->op_notifier) {
 512                 notifier = ro->op_notifier;
 513                 rs = rm->m_rs;
 514                 sock_hold(rds_rs_to_sk(rs));
 515
 516                 notifier->n_status = status;
 517                 spin_lock(&rs->rs_lock);
 518                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
 519                 spin_unlock(&rs->rs_lock);
 520
 521                 ro->op_notifier = NULL;
 522         }
 523
 524         spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 525
 526         if (rs) {
 527                 rds_wake_sk_sleep(rs);
 528                 sock_put(rds_rs_to_sk(rs));
 529         }
 530 }
 531 EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
 532
 533 /*
 534  * Just like above, except looks at atomic op
 535  */
 536 void rds_atomic_send_complete(struct rds_message *rm, int status)
 537 {
 538         struct rds_sock *rs = NULL;
 539         struct rm_atomic_op *ao;
 540         struct rds_notifier *notifier;
 541         unsigned long flags;
 542
 543         spin_lock_irqsave(&rm->m_rs_lock, flags);
 544
 545         ao = &rm->atomic;
 546         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
 547             && ao->op_active && ao->op_notify && ao->op_notifier) {
 548                 notifier = ao->op_notifier;
 549                 rs = rm->m_rs;
 550                 sock_hold(rds_rs_to_sk(rs));
 551
 552                 notifier->n_status = status;
 553                 spin_lock(&rs->rs_lock);
 554                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
 555                 spin_unlock(&rs->rs_lock);
 556
 557                 ao->op_notifier = NULL;
 558         }
 559
 560         spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 561
 562         if (rs) {
 563                 rds_wake_sk_sleep(rs);
 564                 sock_put(rds_rs_to_sk(rs));
 565         }
 566 }
 567 EXPORT_SYMBOL_GPL(rds_atomic_send_complete);
 568
 569 /*
 570  * This is the same as rds_rdma_send_complete except we
 571  * don't do any locking - we have all the ingredients (message,
 572  * socket, socket lock) and can just move the notifier.
 573  */
 574 static inline void
 575 __rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
 576 {
 577         struct rm_rdma_op *ro;
 578         struct rm_atomic_op *ao;
 579
 580         ro = &rm->rdma;
 581         if (ro->op_active && ro->op_notify && ro->op_notifier) {
 582                 ro->op_notifier->n_status = status;
 583                 list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue);
 584                 ro->op_notifier = NULL;
 585         }
 586
 587         ao = &rm->atomic;
 588         if (ao->op_active && ao->op_notify && ao->op_notifier) {
 589                 ao->op_notifier->n_status = status;
 590                 list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue);
 591                 ao->op_notifier = NULL;
 592         }
 593
 594         /* No need to wake the app - caller does this */
 595 }
 596
 597 /*
 598  * This removes messages from the socket's list if they're on it.  The list
 599  * argument must be private to the caller, we must be able to modify it
 600  * without locks.  The messages must have a reference held for their
 601  * position on the list.  This function will drop that reference after
 602  * removing the messages from the 'messages' list regardless of if it found
 603  * the messages on the socket list or not.
 604  */
 605 static void rds_send_remove_from_sock(struct list_head *messages, int status)
 606 {
 607         unsigned long flags;
 608         struct rds_sock *rs = NULL;
 609         struct rds_message *rm;
 610
 611         while (!list_empty(messages)) {
 612                 int was_on_sock = 0;
 613
 614                 rm = list_entry(messages->next, struct rds_message,
 615                                 m_conn_item);
 616                 list_del_init(&rm->m_conn_item);
 617
 618                 /*
 619                  * If we see this flag cleared then we're *sure* that someone
 620                  * else beat us to removing it from the sock.  If we race
 621                  * with their flag update we'll get the lock and then really
 622                  * see that the flag has been cleared.
 623                  *
 624                  * The message spinlock makes sure nobody clears rm->m_rs
 625                  * while we're messing with it. It does not prevent the
 626                  * message from being removed from the socket, though.
 627                  */
 628                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 629                 if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags))
 630                         goto unlock_and_drop;
 631
 632                 if (rs != rm->m_rs) {
 633                         if (rs) {
 634                                 rds_wake_sk_sleep(rs);
 635                                 sock_put(rds_rs_to_sk(rs));
 636                         }
 637                         rs = rm->m_rs;
 638                         if (rs)
 639                                 sock_hold(rds_rs_to_sk(rs));
 640                 }
 641                 if (!rs)
 642                         goto unlock_and_drop;
 643                 spin_lock(&rs->rs_lock);
 644
 645                 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
 646                         struct rm_rdma_op *ro = &rm->rdma;
 647                         struct rds_notifier *notifier;
 648
 649                         list_del_init(&rm->m_sock_item);
 650                         rds_send_sndbuf_remove(rs, rm);
 651
 652                         if (ro->op_active && ro->op_notifier &&
 653                                (ro->op_notify || (ro->op_recverr && status))) {
 654                                 notifier = ro->op_notifier;
 655                                 list_add_tail(&notifier->n_list,
 656                                                 &rs->rs_notify_queue);
 657                                 if (!notifier->n_status)
 658                                         notifier->n_status = status;
 659                                 rm->rdma.op_notifier = NULL;
 660                         }
 661                         was_on_sock = 1;
 662                 }
 663                 spin_unlock(&rs->rs_lock);
 664
 665 unlock_and_drop:
 666                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 667                 rds_message_put(rm);
 668                 if (was_on_sock)
 669                         rds_message_put(rm);
 670         }
 671
 672         if (rs) {
 673                 rds_wake_sk_sleep(rs);
 674                 sock_put(rds_rs_to_sk(rs));
 675         }
 676 }
 677
 678 /*
 679  * Transports call here when they've determined that the receiver queued
 680  * messages up to, and including, the given sequence number.  Messages are
 681  * moved to the retrans queue when rds_send_xmit picks them off the send
 682  * queue. This means that in the TCP case, the message may not have been
 683  * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
 684  * checks the RDS_MSG_HAS_ACK_SEQ bit.
 685  */
 686 void rds_send_path_drop_acked(struct rds_conn_path *cp, u64 ack,
 687                               is_acked_func is_acked)
 688 {
 689         struct rds_message *rm, *tmp;
 690         unsigned long flags;
 691         LIST_HEAD(list);
 692
 693         spin_lock_irqsave(&cp->cp_lock, flags);
 694
 695         list_for_each_entry_safe(rm, tmp, &cp->cp_retrans, m_conn_item) {
 696                 if (!rds_send_is_acked(rm, ack, is_acked))
 697                         break;
 698
 699                 list_move(&rm->m_conn_item, &list);
 700                 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
 701         }
 702
 703         /* order flag updates with spin locks */
 704         if (!list_empty(&list))
 705                 smp_mb__after_atomic();
 706
 707         spin_unlock_irqrestore(&cp->cp_lock, flags);
 708
 709         /* now remove the messages from the sock list as needed */
 710         rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
 711 }
 712 EXPORT_SYMBOL_GPL(rds_send_path_drop_acked);
 713
 714 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
 715                          is_acked_func is_acked)
 716 {
 717         WARN_ON(conn->c_trans->t_mp_capable);
 718         rds_send_path_drop_acked(&conn->c_path[0], ack, is_acked);
 719 }
 720 EXPORT_SYMBOL_GPL(rds_send_drop_acked);
 721
 722 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in6 *dest)
 723 {
 724         struct rds_message *rm, *tmp;
 725         struct rds_connection *conn;
 726         struct rds_conn_path *cp;
 727         unsigned long flags;
 728         LIST_HEAD(list);
 729
 730         /* get all the messages we're dropping under the rs lock */
 731         spin_lock_irqsave(&rs->rs_lock, flags);
 732
 733         list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) {
 734                 if (dest &&
 735                     (!ipv6_addr_equal(&dest->sin6_addr, &rm->m_daddr) ||
 736                      dest->sin6_port != rm->m_inc.i_hdr.h_dport))
 737                         continue;
 738
 739                 list_move(&rm->m_sock_item, &list);
 740                 rds_send_sndbuf_remove(rs, rm);
 741                 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
 742         }
 743
 744         /* order flag updates with the rs lock */
 745         smp_mb__after_atomic();
 746
 747         spin_unlock_irqrestore(&rs->rs_lock, flags);
 748
 749         if (list_empty(&list))
 750                 return;
 751
 752         /* Remove the messages from the conn */
 753         list_for_each_entry(rm, &list, m_sock_item) {
 754
 755                 conn = rm->m_inc.i_conn;
 756                 if (conn->c_trans->t_mp_capable)
 757                         cp = rm->m_inc.i_conn_path;
 758                 else
 759                         cp = &conn->c_path[0];
 760
 761                 spin_lock_irqsave(&cp->cp_lock, flags);
 762                 /*
 763                  * Maybe someone else beat us to removing rm from the conn.
 764                  * If we race with their flag update we'll get the lock and
 765                  * then really see that the flag has been cleared.
 766                  */
 767                 if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
 768                         spin_unlock_irqrestore(&cp->cp_lock, flags);
 769                         continue;
 770                 }
 771                 list_del_init(&rm->m_conn_item);
 772                 spin_unlock_irqrestore(&cp->cp_lock, flags);
 773
 774                 /*
 775                  * Couldn't grab m_rs_lock in top loop (lock ordering),
 776                  * but we can now.
 777                  */
 778                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 779
 780                 spin_lock(&rs->rs_lock);
 781                 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
 782                 spin_unlock(&rs->rs_lock);
 783
 784                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 785
 786                 rds_message_put(rm);
 787         }
 788
 789         rds_wake_sk_sleep(rs);
 790
 791         while (!list_empty(&list)) {
 792                 rm = list_entry(list.next, struct rds_message, m_sock_item);
 793                 list_del_init(&rm->m_sock_item);
 794                 rds_message_wait(rm);
 795
 796                 /* just in case the code above skipped this message
 797                  * because RDS_MSG_ON_CONN wasn't set, run it again here
 798                  * taking m_rs_lock is the only thing that keeps us
 799                  * from racing with ack processing.
 800                  */
 801                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 802
 803                 spin_lock(&rs->rs_lock);
 804                 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
 805                 spin_unlock(&rs->rs_lock);
 806
 807                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 808
 809                 rds_message_put(rm);
 810         }
 811 }
 812
 813 /*
 814  * we only want this to fire once so we use the callers 'queued'.  It's
 815  * possible that another thread can race with us and remove the
 816  * message from the flow with RDS_CANCEL_SENT_TO.
 817  */
 818 static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
 819                              struct rds_conn_path *cp,
 820                              struct rds_message *rm, __be16 sport,
 821                              __be16 dport, int *queued)
 822 {
 823         unsigned long flags;
 824         u32 len;
 825
 826         if (*queued)
 827                 goto out;
 828
 829         len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
 830
 831         /* this is the only place which holds both the socket's rs_lock
 832          * and the connection's c_lock */
 833         spin_lock_irqsave(&rs->rs_lock, flags);
 834
 835         /*
 836          * If there is a little space in sndbuf, we don't queue anything,
 837          * and userspace gets -EAGAIN. But poll() indicates there's send
 838          * room. This can lead to bad behavior (spinning) if snd_bytes isn't
 839          * freed up by incoming acks. So we check the *old* value of
 840          * rs_snd_bytes here to allow the last msg to exceed the buffer,
 841          * and poll() now knows no more data can be sent.
 842          */
 843         if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
 844                 rs->rs_snd_bytes += len;
 845
 846                 /* let recv side know we are close to send space exhaustion.
 847                  * This is probably not the optimal way to do it, as this
 848                  * means we set the flag on *all* messages as soon as our
 849                  * throughput hits a certain threshold.
 850                  */
 851                 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
 852                         set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
 853
 854                 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
 855                 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
 856                 rds_message_addref(rm);
 857                 sock_hold(rds_rs_to_sk(rs));
 858                 rm->m_rs = rs;
 859
 860                 /* The code ordering is a little weird, but we're
 861                    trying to minimize the time we hold c_lock */
 862                 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
 863                 rm->m_inc.i_conn = conn;
 864                 rm->m_inc.i_conn_path = cp;
 865                 rds_message_addref(rm);
 866
 867                 spin_lock(&cp->cp_lock);
 868                 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(cp->cp_next_tx_seq++);
 869                 list_add_tail(&rm->m_conn_item, &cp->cp_send_queue);
 870                 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
 871                 spin_unlock(&cp->cp_lock);
 872
 873                 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
 874                          rm, len, rs, rs->rs_snd_bytes,
 875                          (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
 876
 877                 *queued = 1;
 878         }
 879
 880         spin_unlock_irqrestore(&rs->rs_lock, flags);
 881 out:
 882         return *queued;
 883 }
 884
 885 /*
 886  * rds_message is getting to be quite complicated, and we'd like to allocate
 887  * it all in one go. This figures out how big it needs to be up front.
 888  */
 889 static int rds_rm_size(struct msghdr *msg, int num_sgs,
 890                        struct rds_iov_vector_arr *vct)
 891 {
 892         struct cmsghdr *cmsg;
 893         int size = 0;
 894         int cmsg_groups = 0;
 895         int retval;
 896         bool zcopy_cookie = false;
 897         struct rds_iov_vector *iov, *tmp_iov;
 898
 899         if (num_sgs < 0)
 900                 return -EINVAL;
 901
 902         for_each_cmsghdr(cmsg, msg) {
 903                 if (!CMSG_OK(msg, cmsg))
 904                         return -EINVAL;
 905
 906                 if (cmsg->cmsg_level != SOL_RDS)
 907                         continue;
 908
 909                 switch (cmsg->cmsg_type) {
 910                 case RDS_CMSG_RDMA_ARGS:
 911                         if (vct->indx >= vct->len) {
 912                                 vct->len += vct->incr;
 913                                 tmp_iov =
 914                                         krealloc(vct->vec,
 915                                                  vct->len *
 916                                                  sizeof(struct rds_iov_vector),
 917                                                  GFP_KERNEL);
 918                                 if (!tmp_iov) {
 919                                         vct->len -= vct->incr;
 920                                         return -ENOMEM;
 921                                 }
 922                                 vct->vec = tmp_iov;
 923                         }
 924                         iov = &vct->vec[vct->indx];
 925                         memset(iov, 0, sizeof(struct rds_iov_vector));
 926                         vct->indx++;
 927                         cmsg_groups |= 1;
 928                         retval = rds_rdma_extra_size(CMSG_DATA(cmsg), iov);
 929                         if (retval < 0)
 930                                 return retval;
 931                         size += retval;
 932
 933                         break;
 934
 935                 case RDS_CMSG_ZCOPY_COOKIE:
 936                         zcopy_cookie = true;
 937                         /* fall through */
 938
 939                 case RDS_CMSG_RDMA_DEST:
 940                 case RDS_CMSG_RDMA_MAP:
 941                         cmsg_groups |= 2;
 942                         /* these are valid but do no add any size */
 943                         break;
 944
 945                 case RDS_CMSG_ATOMIC_CSWP:
 946                 case RDS_CMSG_ATOMIC_FADD:
 947                 case RDS_CMSG_MASKED_ATOMIC_CSWP:
 948                 case RDS_CMSG_MASKED_ATOMIC_FADD:
 949                         cmsg_groups |= 1;
 950                         size += sizeof(struct scatterlist);
 951                         break;
 952
 953                 default:
 954                         return -EINVAL;
 955                 }
 956
 957         }
 958
 959         if ((msg->msg_flags & MSG_ZEROCOPY) && !zcopy_cookie)
 960                 return -EINVAL;
 961
 962         size += num_sgs * sizeof(struct scatterlist);
 963
 964         /* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */
 965         if (cmsg_groups == 3)
 966                 return -EINVAL;
 967
 968         return size;
 969 }
 970
 971 static int rds_cmsg_zcopy(struct rds_sock *rs, struct rds_message *rm,
 972                           struct cmsghdr *cmsg)
 973 {
 974         u32 *cookie;
 975
 976         if (cmsg->cmsg_len < CMSG_LEN(sizeof(*cookie)) ||
 977             !rm->data.op_mmp_znotifier)
 978                 return -EINVAL;
 979         cookie = CMSG_DATA(cmsg);
 980         rm->data.op_mmp_znotifier->z_cookie = *cookie;
 981         return 0;
 982 }
 983
 984 static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
 985                          struct msghdr *msg, int *allocated_mr,
 986                          struct rds_iov_vector_arr *vct)
 987 {
 988         struct cmsghdr *cmsg;
 989         int ret = 0, ind = 0;
 990
 991         for_each_cmsghdr(cmsg, msg) {
 992                 if (!CMSG_OK(msg, cmsg))
 993                         return -EINVAL;
 994
 995                 if (cmsg->cmsg_level != SOL_RDS)
 996                         continue;
 997
 998                 /* As a side effect, RDMA_DEST and RDMA_MAP will set
 999                  * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
1000                  */
1001                 switch (cmsg->cmsg_type) {
1002                 case RDS_CMSG_RDMA_ARGS:
1003                         if (ind >= vct->indx)
1004                                 return -ENOMEM;
1005                         ret = rds_cmsg_rdma_args(rs, rm, cmsg, &vct->vec[ind]);
1006                         ind++;
1007                         break;
1008
1009                 case RDS_CMSG_RDMA_DEST:
1010                         ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
1011                         break;
1012
1013                 case RDS_CMSG_RDMA_MAP:
1014                         ret = rds_cmsg_rdma_map(rs, rm, cmsg);
1015                         if (!ret)
1016                                 *allocated_mr = 1;
1017                         else if (ret == -ENODEV)
1018                                 /* Accommodate the get_mr() case which can fail
1019                                  * if connection isn't established yet.
1020                                  */
1021                                 ret = -EAGAIN;
1022                         break;
1023                 case RDS_CMSG_ATOMIC_CSWP:
1024                 case RDS_CMSG_ATOMIC_FADD:
1025                 case RDS_CMSG_MASKED_ATOMIC_CSWP:
1026                 case RDS_CMSG_MASKED_ATOMIC_FADD:
1027                         ret = rds_cmsg_atomic(rs, rm, cmsg);
1028                         break;
1029
1030                 case RDS_CMSG_ZCOPY_COOKIE:
1031                         ret = rds_cmsg_zcopy(rs, rm, cmsg);
1032                         break;
1033
1034                 default:
1035                         return -EINVAL;
1036                 }
1037
1038                 if (ret)
1039                         break;
1040         }
1041
1042         return ret;
1043 }
1044
1045 static int rds_send_mprds_hash(struct rds_sock *rs,
1046                                struct rds_connection *conn, int nonblock)
1047 {
1048         int hash;
1049
1050         if (conn->c_npaths == 0)
1051                 hash = RDS_MPATH_HASH(rs, RDS_MPATH_WORKERS);
1052         else
1053                 hash = RDS_MPATH_HASH(rs, conn->c_npaths);
1054         if (conn->c_npaths == 0 && hash != 0) {
1055                 rds_send_ping(conn, 0);
1056
1057                 /* The underlying connection is not up yet.  Need to wait
1058                  * until it is up to be sure that the non-zero c_path can be
1059                  * used.  But if we are interrupted, we have to use the zero
1060                  * c_path in case the connection ends up being non-MP capable.
1061                  */
1062                 if (conn->c_npaths == 0) {
1063                         /* Cannot wait for the connection be made, so just use
1064                          * the base c_path.
1065                          */
1066                         if (nonblock)
1067                                 return 0;
1068                         if (wait_event_interruptible(conn->c_hs_waitq,
1069                                                      conn->c_npaths != 0))
1070                                 hash = 0;
1071                 }
1072                 if (conn->c_npaths == 1)
1073                         hash = 0;
1074         }
1075         return hash;
1076 }
1077
1078 static int rds_rdma_bytes(struct msghdr *msg, size_t *rdma_bytes)
1079 {
1080         struct rds_rdma_args *args;
1081         struct cmsghdr *cmsg;
1082
1083         for_each_cmsghdr(cmsg, msg) {
1084                 if (!CMSG_OK(msg, cmsg))
1085                         return -EINVAL;
1086
1087                 if (cmsg->cmsg_level != SOL_RDS)
1088                         continue;
1089
1090                 if (cmsg->cmsg_type == RDS_CMSG_RDMA_ARGS) {
1091                         if (cmsg->cmsg_len <
1092                             CMSG_LEN(sizeof(struct rds_rdma_args)))
1093                                 return -EINVAL;
1094                         args = CMSG_DATA(cmsg);
1095                         *rdma_bytes += args->remote_vec.bytes;
1096                 }
1097         }
1098         return 0;
1099 }
1100
1101 int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len)
1102 {
1103         struct sock *sk = sock->sk;
1104         struct rds_sock *rs = rds_sk_to_rs(sk);
1105         DECLARE_SOCKADDR(struct sockaddr_in6 *, sin6, msg->msg_name);
1106         DECLARE_SOCKADDR(struct sockaddr_in *, usin, msg->msg_name);
1107         __be16 dport;
1108         struct rds_message *rm = NULL;
1109         struct rds_connection *conn;
1110         int ret = 0;
1111         int queued = 0, allocated_mr = 0;
1112         int nonblock = msg->msg_flags & MSG_DONTWAIT;
1113         long timeo = sock_sndtimeo(sk, nonblock);
1114         struct rds_conn_path *cpath;
1115         struct in6_addr daddr;
1116         __u32 scope_id = 0;
1117         size_t total_payload_len = payload_len, rdma_payload_len = 0;
1118         bool zcopy = ((msg->msg_flags & MSG_ZEROCOPY) &&
1119                       sock_flag(rds_rs_to_sk(rs), SOCK_ZEROCOPY));
1120         int num_sgs = DIV_ROUND_UP(payload_len, PAGE_SIZE);
1121         int namelen;
1122         struct rds_iov_vector_arr vct;
1123         int ind;
1124
1125         memset(&vct, 0, sizeof(vct));
1126
1127         /* expect 1 RDMA CMSG per rds_sendmsg. can still grow if more needed. */
1128         vct.incr = 1;
1129
1130         /* Mirror Linux UDP mirror of BSD error message compatibility */
1131         /* XXX: Perhaps MSG_MORE someday */
1132         if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT | MSG_ZEROCOPY)) {
1133                 ret = -EOPNOTSUPP;
1134                 goto out;
1135         }
1136
1137         namelen = msg->msg_namelen;
1138         if (namelen != 0) {
1139                 if (namelen < sizeof(*usin)) {
1140                         ret = -EINVAL;
1141                         goto out;
1142                 }
1143                 switch (usin->sin_family) {
1144                 case AF_INET:
1145                         if (usin->sin_addr.s_addr == htonl(INADDR_ANY) ||
1146                             usin->sin_addr.s_addr == htonl(INADDR_BROADCAST) ||
1147                             ipv4_is_multicast(usin->sin_addr.s_addr)) {
1148                                 ret = -EINVAL;
1149                                 goto out;
1150                         }
1151                         ipv6_addr_set_v4mapped(usin->sin_addr.s_addr, &daddr);
1152                         dport = usin->sin_port;
1153                         break;
1154
1155 #if IS_ENABLED(CONFIG_IPV6)
1156                 case AF_INET6: {
1157                         int addr_type;
1158
1159                         if (namelen < sizeof(*sin6)) {
1160                                 ret = -EINVAL;
1161                                 goto out;
1162                         }
1163                         addr_type = ipv6_addr_type(&sin6->sin6_addr);
1164                         if (!(addr_type & IPV6_ADDR_UNICAST)) {
1165                                 __be32 addr4;
1166
1167                                 if (!(addr_type & IPV6_ADDR_MAPPED)) {
1168                                         ret = -EINVAL;
1169                                         goto out;
1170                                 }
1171
1172                                 /* It is a mapped address.  Need to do some
1173                                  * sanity checks.
1174                                  */
1175                                 addr4 = sin6->sin6_addr.s6_addr32[3];
1176                                 if (addr4 == htonl(INADDR_ANY) ||
1177                                     addr4 == htonl(INADDR_BROADCAST) ||
1178                                     ipv4_is_multicast(addr4)) {
1179                                         ret = -EINVAL;
1180                                         goto out;
1181                                 }
1182                         }
1183                         if (addr_type & IPV6_ADDR_LINKLOCAL) {
1184                                 if (sin6->sin6_scope_id == 0) {
1185                                         ret = -EINVAL;
1186                                         goto out;
1187                                 }
1188                                 scope_id = sin6->sin6_scope_id;
1189                         }
1190
1191                         daddr = sin6->sin6_addr;
1192                         dport = sin6->sin6_port;
1193                         break;
1194                 }
1195 #endif
1196
1197                 default:
1198                         ret = -EINVAL;
1199                         goto out;
1200                 }
1201         } else {
1202                 /* We only care about consistency with ->connect() */
1203                 lock_sock(sk);
1204                 daddr = rs->rs_conn_addr;
1205                 dport = rs->rs_conn_port;
1206                 scope_id = rs->rs_bound_scope_id;
1207                 release_sock(sk);
1208         }
1209
1210         lock_sock(sk);
1211         if (ipv6_addr_any(&rs->rs_bound_addr) || ipv6_addr_any(&daddr)) {
1212                 release_sock(sk);
1213                 ret = -ENOTCONN;
1214                 goto out;
1215         } else if (namelen != 0) {
1216                 /* Cannot send to an IPv4 address using an IPv6 source
1217                  * address and cannot send to an IPv6 address using an
1218                  * IPv4 source address.
1219                  */
1220                 if (ipv6_addr_v4mapped(&daddr) ^
1221                     ipv6_addr_v4mapped(&rs->rs_bound_addr)) {
1222                         release_sock(sk);
1223                         ret = -EOPNOTSUPP;
1224                         goto out;
1225                 }
1226                 /* If the socket is already bound to a link local address,
1227                  * it can only send to peers on the same link.  But allow
1228                  * communicating beween link local and non-link local address.
1229                  */
1230                 if (scope_id != rs->rs_bound_scope_id) {
1231                         if (!scope_id) {
1232                                 scope_id = rs->rs_bound_scope_id;
1233                         } else if (rs->rs_bound_scope_id) {
1234                                 release_sock(sk);
1235                                 ret = -EINVAL;
1236                                 goto out;
1237                         }
1238                 }
1239         }
1240         release_sock(sk);
1241
1242         ret = rds_rdma_bytes(msg, &rdma_payload_len);
1243         if (ret)
1244                 goto out;
1245
1246         total_payload_len += rdma_payload_len;
1247         if (max_t(size_t, payload_len, rdma_payload_len) > RDS_MAX_MSG_SIZE) {
1248                 ret = -EMSGSIZE;
1249                 goto out;
1250         }
1251
1252         if (payload_len > rds_sk_sndbuf(rs)) {
1253                 ret = -EMSGSIZE;
1254                 goto out;
1255         }
1256
1257         if (zcopy) {
1258                 if (rs->rs_transport->t_type != RDS_TRANS_TCP) {
1259                         ret = -EOPNOTSUPP;
1260                         goto out;
1261                 }
1262                 num_sgs = iov_iter_npages(&msg->msg_iter, INT_MAX);
1263         }
1264         /* size of rm including all sgs */
1265         ret = rds_rm_size(msg, num_sgs, &vct);
1266         if (ret < 0)
1267                 goto out;
1268
1269         rm = rds_message_alloc(ret, GFP_KERNEL);
1270         if (!rm) {
1271                 ret = -ENOMEM;
1272                 goto out;
1273         }
1274
1275         /* Attach data to the rm */
1276         if (payload_len) {
1277                 rm->data.op_sg = rds_message_alloc_sgs(rm, num_sgs, &ret);
1278                 if (!rm->data.op_sg)
1279                         goto out;
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                 /* Trigger connection so that its ready for the next retry */
1317                 if (ret ==  -EAGAIN)
1318                         rds_conn_connect_if_down(conn);
1319                 goto out;
1320         }
1321
1322         if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) {
1323                 printk_ratelimited(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
1324                                &rm->rdma, conn->c_trans->xmit_rdma);
1325                 ret = -EOPNOTSUPP;
1326                 goto out;
1327         }
1328
1329         if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
1330                 printk_ratelimited(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
1331                                &rm->atomic, conn->c_trans->xmit_atomic);
1332                 ret = -EOPNOTSUPP;
1333                 goto out;
1334         }
1335
1336         if (rds_destroy_pending(conn)) {
1337                 ret = -EAGAIN;
1338                 goto out;
1339         }
1340
1341         rds_conn_path_connect_if_down(cpath);
1342
1343         ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
1344         if (ret) {
1345                 rs->rs_seen_congestion = 1;
1346                 goto out;
1347         }
1348         while (!rds_send_queue_rm(rs, conn, cpath, rm, rs->rs_bound_port,
1349                                   dport, &queued)) {
1350                 rds_stats_inc(s_send_queue_full);
1351
1352                 if (nonblock) {
1353                         ret = -EAGAIN;
1354                         goto out;
1355                 }
1356
1357                 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
1358                                         rds_send_queue_rm(rs, conn, cpath, rm,
1359                                                           rs->rs_bound_port,
1360                                                           dport,
1361                                                           &queued),
1362                                         timeo);
1363                 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
1364                 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
1365                         continue;
1366
1367                 ret = timeo;
1368                 if (ret == 0)
1369                         ret = -ETIMEDOUT;
1370                 goto out;
1371         }
1372
1373         /*
1374          * By now we've committed to the send.  We reuse rds_send_worker()
1375          * to retry sends in the rds thread if the transport asks us to.
1376          */
1377         rds_stats_inc(s_send_queued);
1378
1379         ret = rds_send_xmit(cpath);
1380         if (ret == -ENOMEM || ret == -EAGAIN) {
1381                 ret = 0;
1382                 rcu_read_lock();
1383                 if (rds_destroy_pending(cpath->cp_conn))
1384                         ret = -ENETUNREACH;
1385                 else
1386                         queue_delayed_work(rds_wq, &cpath->cp_send_w, 1);
1387                 rcu_read_unlock();
1388         }
1389         if (ret)
1390                 goto out;
1391         rds_message_put(rm);
1392
1393         for (ind = 0; ind < vct.indx; ind++)
1394                 kfree(vct.vec[ind].iov);
1395         kfree(vct.vec);
1396
1397         return payload_len;
1398
1399 out:
1400         for (ind = 0; ind < vct.indx; ind++)
1401                 kfree(vct.vec[ind].iov);
1402         kfree(vct.vec);
1403
1404         /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1405          * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1406          * or in any other way, we need to destroy the MR again */
1407         if (allocated_mr)
1408                 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
1409
1410         if (rm)
1411                 rds_message_put(rm);
1412         return ret;
1413 }
1414
1415 /*
1416  * send out a probe. Can be shared by rds_send_ping,
1417  * rds_send_pong, rds_send_hb.
1418  * rds_send_hb should use h_flags
1419  *   RDS_FLAG_HB_PING|RDS_FLAG_ACK_REQUIRED
1420  * or
1421  *   RDS_FLAG_HB_PONG|RDS_FLAG_ACK_REQUIRED
1422  */
1423 static int
1424 rds_send_probe(struct rds_conn_path *cp, __be16 sport,
1425                __be16 dport, u8 h_flags)
1426 {
1427         struct rds_message *rm;
1428         unsigned long flags;
1429         int ret = 0;
1430
1431         rm = rds_message_alloc(0, GFP_ATOMIC);
1432         if (!rm) {
1433                 ret = -ENOMEM;
1434                 goto out;
1435         }
1436
1437         rm->m_daddr = cp->cp_conn->c_faddr;
1438         rm->data.op_active = 1;
1439
1440         rds_conn_path_connect_if_down(cp);
1441
1442         ret = rds_cong_wait(cp->cp_conn->c_fcong, dport, 1, NULL);
1443         if (ret)
1444                 goto out;
1445
1446         spin_lock_irqsave(&cp->cp_lock, flags);
1447         list_add_tail(&rm->m_conn_item, &cp->cp_send_queue);
1448         set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
1449         rds_message_addref(rm);
1450         rm->m_inc.i_conn = cp->cp_conn;
1451         rm->m_inc.i_conn_path = cp;
1452
1453         rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport,
1454                                     cp->cp_next_tx_seq);
1455         rm->m_inc.i_hdr.h_flags |= h_flags;
1456         cp->cp_next_tx_seq++;
1457
1458         if (RDS_HS_PROBE(be16_to_cpu(sport), be16_to_cpu(dport)) &&
1459             cp->cp_conn->c_trans->t_mp_capable) {
1460                 u16 npaths = cpu_to_be16(RDS_MPATH_WORKERS);
1461                 u32 my_gen_num = cpu_to_be32(cp->cp_conn->c_my_gen_num);
1462
1463                 rds_message_add_extension(&rm->m_inc.i_hdr,
1464                                           RDS_EXTHDR_NPATHS, &npaths,
1465                                           sizeof(npaths));
1466                 rds_message_add_extension(&rm->m_inc.i_hdr,
1467                                           RDS_EXTHDR_GEN_NUM,
1468                                           &my_gen_num,
1469                                           sizeof(u32));
1470         }
1471         spin_unlock_irqrestore(&cp->cp_lock, flags);
1472
1473         rds_stats_inc(s_send_queued);
1474         rds_stats_inc(s_send_pong);
1475
1476         /* schedule the send work on rds_wq */
1477         rcu_read_lock();
1478         if (!rds_destroy_pending(cp->cp_conn))
1479                 queue_delayed_work(rds_wq, &cp->cp_send_w, 1);
1480         rcu_read_unlock();
1481
1482         rds_message_put(rm);
1483         return 0;
1484
1485 out:
1486         if (rm)
1487                 rds_message_put(rm);
1488         return ret;
1489 }
1490
1491 int
1492 rds_send_pong(struct rds_conn_path *cp, __be16 dport)
1493 {
1494         return rds_send_probe(cp, 0, dport, 0);
1495 }
1496
1497 void
1498 rds_send_ping(struct rds_connection *conn, int cp_index)
1499 {
1500         unsigned long flags;
1501         struct rds_conn_path *cp = &conn->c_path[cp_index];
1502
1503         spin_lock_irqsave(&cp->cp_lock, flags);
1504         if (conn->c_ping_triggered) {
1505                 spin_unlock_irqrestore(&cp->cp_lock, flags);
1506                 return;
1507         }
1508         conn->c_ping_triggered = 1;
1509         spin_unlock_irqrestore(&cp->cp_lock, flags);
1510         rds_send_probe(cp, cpu_to_be16(RDS_FLAG_PROBE_PORT), 0, 0);
1511 }
1512 EXPORT_SYMBOL_GPL(rds_send_ping);