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