net/rds/send.c

   1 /*
   2  * Copyright (c) 2006 Oracle.  All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  *
  32  */
  33 #include <linux/kernel.h>
  34 #include <linux/gfp.h>
  35 #include <net/sock.h>
  36 #include <linux/in.h>
  37 #include <linux/list.h>
  38
  39 #include "rds.h"
  40
  41 /* When transmitting messages in rds_send_xmit, we need to emerge from
  42  * time to time and briefly release the CPU. Otherwise the softlock watchdog
  43  * will kick our shin.
  44  * Also, it seems fairer to not let one busy connection stall all the
  45  * others.
  46  *
  47  * send_batch_count is the number of times we'll loop in send_xmit. Setting
  48  * it to 0 will restore the old behavior (where we looped until we had
  49  * drained the queue).
  50  */
  51 static int send_batch_count = 64;
  52 module_param(send_batch_count, int, 0444);
  53 MODULE_PARM_DESC(send_batch_count, " batch factor when working the send queue");
  54
  55 /*
  56  * Reset the send state. Caller must hold c_send_lock when calling here.
  57  */
  58 void rds_send_reset(struct rds_connection *conn)
  59 {
  60         struct rds_message *rm, *tmp;
  61         unsigned long flags;
  62
  63         if (conn->c_xmit_rm) {
  64                 /* Tell the user the RDMA op is no longer mapped by the
  65                  * transport. This isn't entirely true (it's flushed out
  66                  * independently) but as the connection is down, there's
  67                  * no ongoing RDMA to/from that memory */
  68                 rds_message_unmapped(conn->c_xmit_rm);
  69                 rds_message_put(conn->c_xmit_rm);
  70                 conn->c_xmit_rm = NULL;
  71         }
  72         conn->c_xmit_sg = 0;
  73         conn->c_xmit_hdr_off = 0;
  74         conn->c_xmit_data_off = 0;
  75         conn->c_xmit_atomic_sent = 0;
  76         conn->c_xmit_rdma_sent = 0;
  77         conn->c_xmit_data_sent = 0;
  78
  79         conn->c_map_queued = 0;
  80
  81         conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
  82         conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
  83
  84         /* Mark messages as retransmissions, and move them to the send q */
  85         spin_lock_irqsave(&conn->c_lock, flags);
  86         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
  87                 set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
  88                 set_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags);
  89         }
  90         list_splice_init(&conn->c_retrans, &conn->c_send_queue);
  91         spin_unlock_irqrestore(&conn->c_lock, flags);
  92 }
  93
  94 /*
  95  * We're making the concious trade-off here to only send one message
  96  * down the connection at a time.
  97  *   Pro:
  98  *      - tx queueing is a simple fifo list
  99  *      - reassembly is optional and easily done by transports per conn
 100  *      - no per flow rx lookup at all, straight to the socket
 101  *      - less per-frag memory and wire overhead
 102  *   Con:
 103  *      - queued acks can be delayed behind large messages
 104  *   Depends:
 105  *      - small message latency is higher behind queued large messages
 106  *      - large message latency isn't starved by intervening small sends
 107  */
 108 int rds_send_xmit(struct rds_connection *conn)
 109 {
 110         struct rds_message *rm;
 111         unsigned long flags;
 112         unsigned int tmp;
 113         unsigned int send_quota = send_batch_count;
 114         struct scatterlist *sg;
 115         int ret = 0;
 116         int was_empty = 0;
 117         LIST_HEAD(to_be_dropped);
 118
 119         /*
 120          * sendmsg calls here after having queued its message on the send
 121          * queue.  We only have one task feeding the connection at a time.  If
 122          * another thread is already feeding the queue then we back off.  This
 123          * avoids blocking the caller and trading per-connection data between
 124          * caches per message.
 125          *
 126          * The sem holder will issue a retry if they notice that someone queued
 127          * a message after they stopped walking the send queue but before they
 128          * dropped the sem.
 129          */
 130         if (!mutex_trylock(&conn->c_send_lock)) {
 131                 rds_stats_inc(s_send_sem_contention);
 132                 ret = -ENOMEM;
 133                 goto out;
 134         }
 135
 136         if (conn->c_trans->xmit_prepare)
 137                 conn->c_trans->xmit_prepare(conn);
 138
 139         /*
 140          * spin trying to push headers and data down the connection until
 141          * the connection doesn't make forward progress.
 142          */
 143         while (--send_quota) {
 144
 145                 rm = conn->c_xmit_rm;
 146
 147                 /*
 148                  * If between sending messages, we can send a pending congestion
 149                  * map update.
 150                  *
 151                  * Transports either define a special xmit_cong_map function,
 152                  * or we allocate a cong_map message and treat it just like any
 153                  * other send.
 154                  */
 155                 if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) {
 156                         if (conn->c_trans->xmit_cong_map) {
 157                                 unsigned long map_offset = 0;
 158                                 unsigned long map_bytes = sizeof(struct rds_header) +
 159                                         RDS_CONG_MAP_BYTES;
 160
 161                                 while (map_bytes) {
 162                                         ret = conn->c_trans->xmit_cong_map(conn, conn->c_lcong,
 163                                                                            map_offset);
 164                                         if (ret <= 0) {
 165                                                 /* too far down the rabbithole! */
 166                                                 mutex_unlock(&conn->c_send_lock);
 167                                                 rds_conn_error(conn, "Cong map xmit failed\n");
 168                                                 goto out;
 169                                         }
 170
 171                                         map_offset += ret;
 172                                         map_bytes -= ret;
 173                                 }
 174                         } else {
 175                                 /* send cong update like a normal rm */
 176                                 rm = rds_cong_update_alloc(conn);
 177                                 if (IS_ERR(rm)) {
 178                                         ret = PTR_ERR(rm);
 179                                         break;
 180                                 }
 181                                 rm->data.op_active = 1;
 182
 183                                 conn->c_xmit_rm = rm;
 184                         }
 185                 }
 186
 187                 /*
 188                  * If not already working on one, grab the next message.
 189                  *
 190                  * c_xmit_rm holds a ref while we're sending this message down
 191                  * the connction.  We can use this ref while holding the
 192                  * send_sem.. rds_send_reset() is serialized with it.
 193                  */
 194                 if (!rm) {
 195                         unsigned int len;
 196
 197                         spin_lock_irqsave(&conn->c_lock, flags);
 198
 199                         if (!list_empty(&conn->c_send_queue)) {
 200                                 rm = list_entry(conn->c_send_queue.next,
 201                                                 struct rds_message,
 202                                                 m_conn_item);
 203                                 rds_message_addref(rm);
 204
 205                                 /*
 206                                  * Move the message from the send queue to the retransmit
 207                                  * list right away.
 208                                  */
 209                                 list_move_tail(&rm->m_conn_item, &conn->c_retrans);
 210                         }
 211
 212                         spin_unlock_irqrestore(&conn->c_lock, flags);
 213
 214                         if (!rm) {
 215                                 was_empty = 1;
 216                                 break;
 217                         }
 218
 219                         /* Unfortunately, the way Infiniband deals with
 220                          * RDMA to a bad MR key is by moving the entire
 221                          * queue pair to error state. We cold possibly
 222                          * recover from that, but right now we drop the
 223                          * connection.
 224                          * Therefore, we never retransmit messages with RDMA ops.
 225                          */
 226                         if (rm->rdma.op_active &&
 227                             test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags)) {
 228                                 spin_lock_irqsave(&conn->c_lock, flags);
 229                                 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags))
 230                                         list_move(&rm->m_conn_item, &to_be_dropped);
 231                                 spin_unlock_irqrestore(&conn->c_lock, flags);
 232                                 rds_message_put(rm);
 233                                 continue;
 234                         }
 235
 236                         /* Require an ACK every once in a while */
 237                         len = ntohl(rm->m_inc.i_hdr.h_len);
 238                         if (conn->c_unacked_packets == 0 ||
 239                             conn->c_unacked_bytes < len) {
 240                                 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
 241
 242                                 conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
 243                                 conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
 244                                 rds_stats_inc(s_send_ack_required);
 245                         } else {
 246                                 conn->c_unacked_bytes -= len;
 247                                 conn->c_unacked_packets--;
 248                         }
 249
 250                         conn->c_xmit_rm = rm;
 251                 }
 252
 253                 /* The transport either sends the whole rdma or none of it */
 254                 if (rm->rdma.op_active && !conn->c_xmit_rdma_sent) {
 255                         rm->m_final_op = &rm->rdma;
 256                         ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
 257                         if (ret)
 258                                 break;
 259                         conn->c_xmit_rdma_sent = 1;
 260
 261                         /* The transport owns the mapped memory for now.
 262                          * You can't unmap it while it's on the send queue */
 263                         set_bit(RDS_MSG_MAPPED, &rm->m_flags);
 264                 }
 265
 266                 if (rm->atomic.op_active && !conn->c_xmit_atomic_sent) {
 267                         rm->m_final_op = &rm->atomic;
 268                         ret = conn->c_trans->xmit_atomic(conn, &rm->atomic);
 269                         if (ret)
 270                                 break;
 271                         conn->c_xmit_atomic_sent = 1;
 272
 273                         /* The transport owns the mapped memory for now.
 274                          * You can't unmap it while it's on the send queue */
 275                         set_bit(RDS_MSG_MAPPED, &rm->m_flags);
 276                 }
 277
 278                 /*
 279                  * A number of cases require an RDS header to be sent
 280                  * even if there is no data.
 281                  * We permit 0-byte sends; rds-ping depends on this.
 282                  * However, if there are exclusively attached silent ops,
 283                  * we skip the hdr/data send, to enable silent operation.
 284                  */
 285                 if (rm->data.op_nents == 0) {
 286                         int ops_present;
 287                         int all_ops_are_silent = 1;
 288
 289                         ops_present = (rm->atomic.op_active || rm->rdma.op_active);
 290                         if (rm->atomic.op_active && !rm->atomic.op_silent)
 291                                 all_ops_are_silent = 0;
 292                         if (rm->rdma.op_active && !rm->rdma.op_silent)
 293                                 all_ops_are_silent = 0;
 294
 295                         if (ops_present && all_ops_are_silent
 296                             && !rm->m_rdma_cookie)
 297                                 rm->data.op_active = 0;
 298                 }
 299
 300                 if (rm->data.op_active && !conn->c_xmit_data_sent) {
 301                         rm->m_final_op = &rm->data;
 302                         ret = conn->c_trans->xmit(conn, rm,
 303                                                   conn->c_xmit_hdr_off,
 304                                                   conn->c_xmit_sg,
 305                                                   conn->c_xmit_data_off);
 306                         if (ret <= 0)
 307                                 break;
 308
 309                         if (conn->c_xmit_hdr_off < sizeof(struct rds_header)) {
 310                                 tmp = min_t(int, ret,
 311                                             sizeof(struct rds_header) -
 312                                             conn->c_xmit_hdr_off);
 313                                 conn->c_xmit_hdr_off += tmp;
 314                                 ret -= tmp;
 315                         }
 316
 317                         sg = &rm->data.op_sg[conn->c_xmit_sg];
 318                         while (ret) {
 319                                 tmp = min_t(int, ret, sg->length -
 320                                                       conn->c_xmit_data_off);
 321                                 conn->c_xmit_data_off += tmp;
 322                                 ret -= tmp;
 323                                 if (conn->c_xmit_data_off == sg->length) {
 324                                         conn->c_xmit_data_off = 0;
 325                                         sg++;
 326                                         conn->c_xmit_sg++;
 327                                         BUG_ON(ret != 0 &&
 328                                                conn->c_xmit_sg == rm->data.op_nents);
 329                                 }
 330                         }
 331
 332                         if (conn->c_xmit_hdr_off == sizeof(struct rds_header) &&
 333                             (conn->c_xmit_sg == rm->data.op_nents))
 334                                 conn->c_xmit_data_sent = 1;
 335                 }
 336
 337                 /*
 338                  * A rm will only take multiple times through this loop
 339                  * if there is a data op. Thus, if the data is sent (or there was
 340                  * none), then we're done with the rm.
 341                  */
 342                 if (!rm->data.op_active || conn->c_xmit_data_sent) {
 343                         conn->c_xmit_rm = NULL;
 344                         conn->c_xmit_sg = 0;
 345                         conn->c_xmit_hdr_off = 0;
 346                         conn->c_xmit_data_off = 0;
 347                         conn->c_xmit_rdma_sent = 0;
 348                         conn->c_xmit_atomic_sent = 0;
 349                         conn->c_xmit_data_sent = 0;
 350
 351                         rds_message_put(rm);
 352                 }
 353         }
 354
 355         /* Nuke any messages we decided not to retransmit. */
 356         if (!list_empty(&to_be_dropped))
 357                 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
 358
 359         if (conn->c_trans->xmit_complete)
 360                 conn->c_trans->xmit_complete(conn);
 361
 362         /*
 363          * We might be racing with another sender who queued a message but
 364          * backed off on noticing that we held the c_send_lock.  If we check
 365          * for queued messages after dropping the sem then either we'll
 366          * see the queued message or the queuer will get the sem.  If we
 367          * notice the queued message then we trigger an immediate retry.
 368          *
 369          * We need to be careful only to do this when we stopped processing
 370          * the send queue because it was empty.  It's the only way we
 371          * stop processing the loop when the transport hasn't taken
 372          * responsibility for forward progress.
 373          */
 374         mutex_unlock(&conn->c_send_lock);
 375
 376         if (send_quota == 0 && !was_empty) {
 377                 /* We exhausted the send quota, but there's work left to
 378                  * do. Return and (re-)schedule the send worker.
 379                  */
 380                 ret = -EAGAIN;
 381         }
 382
 383         if (ret == 0 && was_empty) {
 384                 /* A simple bit test would be way faster than taking the
 385                  * spin lock */
 386                 spin_lock_irqsave(&conn->c_lock, flags);
 387                 if (!list_empty(&conn->c_send_queue)) {
 388                         rds_stats_inc(s_send_sem_queue_raced);
 389                         ret = -EAGAIN;
 390                 }
 391                 spin_unlock_irqrestore(&conn->c_lock, flags);
 392         }
 393 out:
 394         return ret;
 395 }
 396
 397 static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
 398 {
 399         u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
 400
 401         assert_spin_locked(&rs->rs_lock);
 402
 403         BUG_ON(rs->rs_snd_bytes < len);
 404         rs->rs_snd_bytes -= len;
 405
 406         if (rs->rs_snd_bytes == 0)
 407                 rds_stats_inc(s_send_queue_empty);
 408 }
 409
 410 static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
 411                                     is_acked_func is_acked)
 412 {
 413         if (is_acked)
 414                 return is_acked(rm, ack);
 415         return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
 416 }
 417
 418 /*
 419  * Returns true if there are no messages on the send and retransmit queues
 420  * which have a sequence number greater than or equal to the given sequence
 421  * number.
 422  */
 423 int rds_send_acked_before(struct rds_connection *conn, u64 seq)
 424 {
 425         struct rds_message *rm, *tmp;
 426         int ret = 1;
 427
 428         spin_lock(&conn->c_lock);
 429
 430         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
 431                 if (be64_to_cpu(rm->m_inc.i_hdr.h_sequence) < seq)
 432                         ret = 0;
 433                 break;
 434         }
 435
 436         list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
 437                 if (be64_to_cpu(rm->m_inc.i_hdr.h_sequence) < seq)
 438                         ret = 0;
 439                 break;
 440         }
 441
 442         spin_unlock(&conn->c_lock);
 443
 444         return ret;
 445 }
 446
 447 /*
 448  * This is pretty similar to what happens below in the ACK
 449  * handling code - except that we call here as soon as we get
 450  * the IB send completion on the RDMA op and the accompanying
 451  * message.
 452  */
 453 void rds_rdma_send_complete(struct rds_message *rm, int status)
 454 {
 455         struct rds_sock *rs = NULL;
 456         struct rm_rdma_op *ro;
 457         struct rds_notifier *notifier;
 458         unsigned long flags;
 459
 460         spin_lock_irqsave(&rm->m_rs_lock, flags);
 461
 462         ro = &rm->rdma;
 463         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
 464             ro->op_active && ro->op_notify && ro->op_notifier) {
 465                 notifier = ro->op_notifier;
 466                 rs = rm->m_rs;
 467                 sock_hold(rds_rs_to_sk(rs));
 468
 469                 notifier->n_status = status;
 470                 spin_lock(&rs->rs_lock);
 471                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
 472                 spin_unlock(&rs->rs_lock);
 473
 474                 ro->op_notifier = NULL;
 475         }
 476
 477         spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 478
 479         if (rs) {
 480                 rds_wake_sk_sleep(rs);
 481                 sock_put(rds_rs_to_sk(rs));
 482         }
 483 }
 484 EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
 485
 486 /*
 487  * Just like above, except looks at atomic op
 488  */
 489 void rds_atomic_send_complete(struct rds_message *rm, int status)
 490 {
 491         struct rds_sock *rs = NULL;
 492         struct rm_atomic_op *ao;
 493         struct rds_notifier *notifier;
 494
 495         spin_lock(&rm->m_rs_lock);
 496
 497         ao = &rm->atomic;
 498         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
 499             && ao->op_active && ao->op_notify && ao->op_notifier) {
 500                 notifier = ao->op_notifier;
 501                 rs = rm->m_rs;
 502                 sock_hold(rds_rs_to_sk(rs));
 503
 504                 notifier->n_status = status;
 505                 spin_lock(&rs->rs_lock);
 506                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
 507                 spin_unlock(&rs->rs_lock);
 508
 509                 ao->op_notifier = NULL;
 510         }
 511
 512         spin_unlock(&rm->m_rs_lock);
 513
 514         if (rs) {
 515                 rds_wake_sk_sleep(rs);
 516                 sock_put(rds_rs_to_sk(rs));
 517         }
 518 }
 519 EXPORT_SYMBOL_GPL(rds_atomic_send_complete);
 520
 521 /*
 522  * This is the same as rds_rdma_send_complete except we
 523  * don't do any locking - we have all the ingredients (message,
 524  * socket, socket lock) and can just move the notifier.
 525  */
 526 static inline void
 527 __rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
 528 {
 529         struct rm_rdma_op *ro;
 530         struct rm_atomic_op *ao;
 531
 532         ro = &rm->rdma;
 533         if (ro->op_active && ro->op_notify && ro->op_notifier) {
 534                 ro->op_notifier->n_status = status;
 535                 list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue);
 536                 ro->op_notifier = NULL;
 537         }
 538
 539         ao = &rm->atomic;
 540         if (ao->op_active && ao->op_notify && ao->op_notifier) {
 541                 ao->op_notifier->n_status = status;
 542                 list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue);
 543                 ao->op_notifier = NULL;
 544         }
 545
 546         /* No need to wake the app - caller does this */
 547 }
 548
 549 /*
 550  * This is called from the IB send completion when we detect
 551  * a RDMA operation that failed with remote access error.
 552  * So speed is not an issue here.
 553  */
 554 struct rds_message *rds_send_get_message(struct rds_connection *conn,
 555                                          struct rm_rdma_op *op)
 556 {
 557         struct rds_message *rm, *tmp, *found = NULL;
 558         unsigned long flags;
 559
 560         spin_lock_irqsave(&conn->c_lock, flags);
 561
 562         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
 563                 if (&rm->rdma == op) {
 564                         atomic_inc(&rm->m_refcount);
 565                         found = rm;
 566                         goto out;
 567                 }
 568         }
 569
 570         list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
 571                 if (&rm->rdma == op) {
 572                         atomic_inc(&rm->m_refcount);
 573                         found = rm;
 574                         break;
 575                 }
 576         }
 577
 578 out:
 579         spin_unlock_irqrestore(&conn->c_lock, flags);
 580
 581         return found;
 582 }
 583 EXPORT_SYMBOL_GPL(rds_send_get_message);
 584
 585 /*
 586  * This removes messages from the socket's list if they're on it.  The list
 587  * argument must be private to the caller, we must be able to modify it
 588  * without locks.  The messages must have a reference held for their
 589  * position on the list.  This function will drop that reference after
 590  * removing the messages from the 'messages' list regardless of if it found
 591  * the messages on the socket list or not.
 592  */
 593 void rds_send_remove_from_sock(struct list_head *messages, int status)
 594 {
 595         unsigned long flags;
 596         struct rds_sock *rs = NULL;
 597         struct rds_message *rm;
 598
 599         while (!list_empty(messages)) {
 600                 int was_on_sock = 0;
 601
 602                 rm = list_entry(messages->next, struct rds_message,
 603                                 m_conn_item);
 604                 list_del_init(&rm->m_conn_item);
 605
 606                 /*
 607                  * If we see this flag cleared then we're *sure* that someone
 608                  * else beat us to removing it from the sock.  If we race
 609                  * with their flag update we'll get the lock and then really
 610                  * see that the flag has been cleared.
 611                  *
 612                  * The message spinlock makes sure nobody clears rm->m_rs
 613                  * while we're messing with it. It does not prevent the
 614                  * message from being removed from the socket, though.
 615                  */
 616                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 617                 if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags))
 618                         goto unlock_and_drop;
 619
 620                 if (rs != rm->m_rs) {
 621                         if (rs) {
 622                                 rds_wake_sk_sleep(rs);
 623                                 sock_put(rds_rs_to_sk(rs));
 624                         }
 625                         rs = rm->m_rs;
 626                         sock_hold(rds_rs_to_sk(rs));
 627                 }
 628                 spin_lock(&rs->rs_lock);
 629
 630                 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
 631                         struct rm_rdma_op *ro = &rm->rdma;
 632                         struct rds_notifier *notifier;
 633
 634                         list_del_init(&rm->m_sock_item);
 635                         rds_send_sndbuf_remove(rs, rm);
 636
 637                         if (ro->op_active && ro->op_notifier &&
 638                                (ro->op_notify || (ro->op_recverr && status))) {
 639                                 notifier = ro->op_notifier;
 640                                 list_add_tail(&notifier->n_list,
 641                                                 &rs->rs_notify_queue);
 642                                 if (!notifier->n_status)
 643                                         notifier->n_status = status;
 644                                 rm->rdma.op_notifier = NULL;
 645                         }
 646                         was_on_sock = 1;
 647                         rm->m_rs = NULL;
 648                 }
 649                 spin_unlock(&rs->rs_lock);
 650
 651 unlock_and_drop:
 652                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 653                 rds_message_put(rm);
 654                 if (was_on_sock)
 655                         rds_message_put(rm);
 656         }
 657
 658         if (rs) {
 659                 rds_wake_sk_sleep(rs);
 660                 sock_put(rds_rs_to_sk(rs));
 661         }
 662 }
 663
 664 /*
 665  * Transports call here when they've determined that the receiver queued
 666  * messages up to, and including, the given sequence number.  Messages are
 667  * moved to the retrans queue when rds_send_xmit picks them off the send
 668  * queue. This means that in the TCP case, the message may not have been
 669  * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
 670  * checks the RDS_MSG_HAS_ACK_SEQ bit.
 671  *
 672  * XXX It's not clear to me how this is safely serialized with socket
 673  * destruction.  Maybe it should bail if it sees SOCK_DEAD.
 674  */
 675 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
 676                          is_acked_func is_acked)
 677 {
 678         struct rds_message *rm, *tmp;
 679         unsigned long flags;
 680         LIST_HEAD(list);
 681
 682         spin_lock_irqsave(&conn->c_lock, flags);
 683
 684         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
 685                 if (!rds_send_is_acked(rm, ack, is_acked))
 686                         break;
 687
 688                 list_move(&rm->m_conn_item, &list);
 689                 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
 690         }
 691
 692         /* order flag updates with spin locks */
 693         if (!list_empty(&list))
 694                 smp_mb__after_clear_bit();
 695
 696         spin_unlock_irqrestore(&conn->c_lock, flags);
 697
 698         /* now remove the messages from the sock list as needed */
 699         rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
 700 }
 701 EXPORT_SYMBOL_GPL(rds_send_drop_acked);
 702
 703 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
 704 {
 705         struct rds_message *rm, *tmp;
 706         struct rds_connection *conn;
 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_clear_bit();
 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
 736                 spin_lock_irqsave(&conn->c_lock, flags);
 737                 /*
 738                  * Maybe someone else beat us to removing rm from the conn.
 739                  * If we race with their flag update we'll get the lock and
 740                  * then really see that the flag has been cleared.
 741                  */
 742                 if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
 743                         spin_unlock_irqrestore(&conn->c_lock, flags);
 744                         continue;
 745                 }
 746                 list_del_init(&rm->m_conn_item);
 747                 spin_unlock_irqrestore(&conn->c_lock, flags);
 748
 749                 /*
 750                  * Couldn't grab m_rs_lock in top loop (lock ordering),
 751                  * but we can now.
 752                  */
 753                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 754
 755                 spin_lock(&rs->rs_lock);
 756                 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
 757                 spin_unlock(&rs->rs_lock);
 758
 759                 rm->m_rs = NULL;
 760                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 761
 762                 rds_message_put(rm);
 763         }
 764
 765         rds_wake_sk_sleep(rs);
 766
 767         while (!list_empty(&list)) {
 768                 rm = list_entry(list.next, struct rds_message, m_sock_item);
 769                 list_del_init(&rm->m_sock_item);
 770
 771                 rds_message_wait(rm);
 772                 rds_message_put(rm);
 773         }
 774 }
 775
 776 /*
 777  * we only want this to fire once so we use the callers 'queued'.  It's
 778  * possible that another thread can race with us and remove the
 779  * message from the flow with RDS_CANCEL_SENT_TO.
 780  */
 781 static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
 782                              struct rds_message *rm, __be16 sport,
 783                              __be16 dport, int *queued)
 784 {
 785         unsigned long flags;
 786         u32 len;
 787
 788         if (*queued)
 789                 goto out;
 790
 791         len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
 792
 793         /* this is the only place which holds both the socket's rs_lock
 794          * and the connection's c_lock */
 795         spin_lock_irqsave(&rs->rs_lock, flags);
 796
 797         /*
 798          * If there is a little space in sndbuf, we don't queue anything,
 799          * and userspace gets -EAGAIN. But poll() indicates there's send
 800          * room. This can lead to bad behavior (spinning) if snd_bytes isn't
 801          * freed up by incoming acks. So we check the *old* value of
 802          * rs_snd_bytes here to allow the last msg to exceed the buffer,
 803          * and poll() now knows no more data can be sent.
 804          */
 805         if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
 806                 rs->rs_snd_bytes += len;
 807
 808                 /* let recv side know we are close to send space exhaustion.
 809                  * This is probably not the optimal way to do it, as this
 810                  * means we set the flag on *all* messages as soon as our
 811                  * throughput hits a certain threshold.
 812                  */
 813                 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
 814                         __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
 815
 816                 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
 817                 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
 818                 rds_message_addref(rm);
 819                 rm->m_rs = rs;
 820
 821                 /* The code ordering is a little weird, but we're
 822                    trying to minimize the time we hold c_lock */
 823                 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
 824                 rm->m_inc.i_conn = conn;
 825                 rds_message_addref(rm);
 826
 827                 spin_lock(&conn->c_lock);
 828                 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(conn->c_next_tx_seq++);
 829                 list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
 830                 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
 831                 spin_unlock(&conn->c_lock);
 832
 833                 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
 834                          rm, len, rs, rs->rs_snd_bytes,
 835                          (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
 836
 837                 *queued = 1;
 838         }
 839
 840         spin_unlock_irqrestore(&rs->rs_lock, flags);
 841 out:
 842         return *queued;
 843 }
 844
 845 /*
 846  * rds_message is getting to be quite complicated, and we'd like to allocate
 847  * it all in one go. This figures out how big it needs to be up front.
 848  */
 849 static int rds_rm_size(struct msghdr *msg, int data_len)
 850 {
 851         struct cmsghdr *cmsg;
 852         int size = 0;
 853         int cmsg_groups = 0;
 854         int retval;
 855
 856         for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
 857                 if (!CMSG_OK(msg, cmsg))
 858                         return -EINVAL;
 859
 860                 if (cmsg->cmsg_level != SOL_RDS)
 861                         continue;
 862
 863                 switch (cmsg->cmsg_type) {
 864                 case RDS_CMSG_RDMA_ARGS:
 865                         cmsg_groups |= 1;
 866                         retval = rds_rdma_extra_size(CMSG_DATA(cmsg));
 867                         if (retval < 0)
 868                                 return retval;
 869                         size += retval;
 870
 871                         break;
 872
 873                 case RDS_CMSG_RDMA_DEST:
 874                 case RDS_CMSG_RDMA_MAP:
 875                         cmsg_groups |= 2;
 876                         /* these are valid but do no add any size */
 877                         break;
 878
 879                 case RDS_CMSG_ATOMIC_CSWP:
 880                 case RDS_CMSG_ATOMIC_FADD:
 881                         cmsg_groups |= 1;
 882                         size += sizeof(struct scatterlist);
 883                         break;
 884
 885                 default:
 886                         return -EINVAL;
 887                 }
 888
 889         }
 890
 891         size += ceil(data_len, PAGE_SIZE) * sizeof(struct scatterlist);
 892
 893         /* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */
 894         if (cmsg_groups == 3)
 895                 return -EINVAL;
 896
 897         return size;
 898 }
 899
 900 static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
 901                          struct msghdr *msg, int *allocated_mr)
 902 {
 903         struct cmsghdr *cmsg;
 904         int ret = 0;
 905
 906         for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
 907                 if (!CMSG_OK(msg, cmsg))
 908                         return -EINVAL;
 909
 910                 if (cmsg->cmsg_level != SOL_RDS)
 911                         continue;
 912
 913                 /* As a side effect, RDMA_DEST and RDMA_MAP will set
 914                  * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
 915                  */
 916                 switch (cmsg->cmsg_type) {
 917                 case RDS_CMSG_RDMA_ARGS:
 918                         ret = rds_cmsg_rdma_args(rs, rm, cmsg);
 919                         break;
 920
 921                 case RDS_CMSG_RDMA_DEST:
 922                         ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
 923                         break;
 924
 925                 case RDS_CMSG_RDMA_MAP:
 926                         ret = rds_cmsg_rdma_map(rs, rm, cmsg);
 927                         if (!ret)
 928                                 *allocated_mr = 1;
 929                         break;
 930                 case RDS_CMSG_ATOMIC_CSWP:
 931                 case RDS_CMSG_ATOMIC_FADD:
 932                         ret = rds_cmsg_atomic(rs, rm, cmsg);
 933                         break;
 934
 935                 default:
 936                         return -EINVAL;
 937                 }
 938
 939                 if (ret)
 940                         break;
 941         }
 942
 943         return ret;
 944 }
 945
 946 int rds_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
 947                 size_t payload_len)
 948 {
 949         struct sock *sk = sock->sk;
 950         struct rds_sock *rs = rds_sk_to_rs(sk);
 951         struct sockaddr_in *usin = (struct sockaddr_in *)msg->msg_name;
 952         __be32 daddr;
 953         __be16 dport;
 954         struct rds_message *rm = NULL;
 955         struct rds_connection *conn;
 956         int ret = 0;
 957         int queued = 0, allocated_mr = 0;
 958         int nonblock = msg->msg_flags & MSG_DONTWAIT;
 959         long timeo = sock_sndtimeo(sk, nonblock);
 960
 961         /* Mirror Linux UDP mirror of BSD error message compatibility */
 962         /* XXX: Perhaps MSG_MORE someday */
 963         if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT)) {
 964                 printk(KERN_INFO "msg_flags 0x%08X\n", msg->msg_flags);
 965                 ret = -EOPNOTSUPP;
 966                 goto out;
 967         }
 968
 969         if (msg->msg_namelen) {
 970                 /* XXX fail non-unicast destination IPs? */
 971                 if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) {
 972                         ret = -EINVAL;
 973                         goto out;
 974                 }
 975                 daddr = usin->sin_addr.s_addr;
 976                 dport = usin->sin_port;
 977         } else {
 978                 /* We only care about consistency with ->connect() */
 979                 lock_sock(sk);
 980                 daddr = rs->rs_conn_addr;
 981                 dport = rs->rs_conn_port;
 982                 release_sock(sk);
 983         }
 984
 985         /* racing with another thread binding seems ok here */
 986         if (daddr == 0 || rs->rs_bound_addr == 0) {
 987                 ret = -ENOTCONN; /* XXX not a great errno */
 988                 goto out;
 989         }
 990
 991         /* size of rm including all sgs */
 992         ret = rds_rm_size(msg, payload_len);
 993         if (ret < 0)
 994                 goto out;
 995
 996         rm = rds_message_alloc(ret, GFP_KERNEL);
 997         if (!rm) {
 998                 ret = -ENOMEM;
 999                 goto out;
1000         }
1001
1002         /* Attach data to the rm */
1003         if (payload_len) {
1004                 rm->data.op_sg = rds_message_alloc_sgs(rm, ceil(payload_len, PAGE_SIZE));
1005                 ret = rds_message_copy_from_user(rm, msg->msg_iov, payload_len);
1006                 if (ret)
1007                         goto out;
1008         }
1009         rm->data.op_active = 1;
1010
1011         rm->m_daddr = daddr;
1012
1013         /* rds_conn_create has a spinlock that runs with IRQ off.
1014          * Caching the conn in the socket helps a lot. */
1015         if (rs->rs_conn && rs->rs_conn->c_faddr == daddr)
1016                 conn = rs->rs_conn;
1017         else {
1018                 conn = rds_conn_create_outgoing(rs->rs_bound_addr, daddr,
1019                                         rs->rs_transport,
1020                                         sock->sk->sk_allocation);
1021                 if (IS_ERR(conn)) {
1022                         ret = PTR_ERR(conn);
1023                         goto out;
1024                 }
1025                 rs->rs_conn = conn;
1026         }
1027
1028         /* Parse any control messages the user may have included. */
1029         ret = rds_cmsg_send(rs, rm, msg, &allocated_mr);
1030         if (ret)
1031                 goto out;
1032
1033         if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) {
1034                 if (printk_ratelimit())
1035                         printk(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
1036                                &rm->rdma, conn->c_trans->xmit_rdma);
1037                 ret = -EOPNOTSUPP;
1038                 goto out;
1039         }
1040
1041         if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
1042                 if (printk_ratelimit())
1043                         printk(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
1044                                &rm->atomic, conn->c_trans->xmit_atomic);
1045                 ret = -EOPNOTSUPP;
1046                 goto out;
1047         }
1048
1049         /* If the connection is down, trigger a connect. We may
1050          * have scheduled a delayed reconnect however - in this case
1051          * we should not interfere.
1052          */
1053         if (rds_conn_state(conn) == RDS_CONN_DOWN &&
1054             !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
1055                 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
1056
1057         ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
1058         if (ret) {
1059                 rs->rs_seen_congestion = 1;
1060                 goto out;
1061         }
1062
1063         while (!rds_send_queue_rm(rs, conn, rm, rs->rs_bound_port,
1064                                   dport, &queued)) {
1065                 rds_stats_inc(s_send_queue_full);
1066                 /* XXX make sure this is reasonable */
1067                 if (payload_len > rds_sk_sndbuf(rs)) {
1068                         ret = -EMSGSIZE;
1069                         goto out;
1070                 }
1071                 if (nonblock) {
1072                         ret = -EAGAIN;
1073                         goto out;
1074                 }
1075
1076                 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
1077                                         rds_send_queue_rm(rs, conn, rm,
1078                                                           rs->rs_bound_port,
1079                                                           dport,
1080                                                           &queued),
1081                                         timeo);
1082                 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
1083                 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
1084                         continue;
1085
1086                 ret = timeo;
1087                 if (ret == 0)
1088                         ret = -ETIMEDOUT;
1089                 goto out;
1090         }
1091
1092         /*
1093          * By now we've committed to the send.  We reuse rds_send_worker()
1094          * to retry sends in the rds thread if the transport asks us to.
1095          */
1096         rds_stats_inc(s_send_queued);
1097
1098         if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
1099                 rds_send_worker(&conn->c_send_w.work);
1100
1101         rds_message_put(rm);
1102         return payload_len;
1103
1104 out:
1105         /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1106          * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1107          * or in any other way, we need to destroy the MR again */
1108         if (allocated_mr)
1109                 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
1110
1111         if (rm)
1112                 rds_message_put(rm);
1113         return ret;
1114 }
1115
1116 /*
1117  * Reply to a ping packet.
1118  */
1119 int
1120 rds_send_pong(struct rds_connection *conn, __be16 dport)
1121 {
1122         struct rds_message *rm;
1123         unsigned long flags;
1124         int ret = 0;
1125
1126         rm = rds_message_alloc(0, GFP_ATOMIC);
1127         if (!rm) {
1128                 ret = -ENOMEM;
1129                 goto out;
1130         }
1131
1132         rm->m_daddr = conn->c_faddr;
1133
1134         /* If the connection is down, trigger a connect. We may
1135          * have scheduled a delayed reconnect however - in this case
1136          * we should not interfere.
1137          */
1138         if (rds_conn_state(conn) == RDS_CONN_DOWN &&
1139             !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
1140                 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
1141
1142         ret = rds_cong_wait(conn->c_fcong, dport, 1, NULL);
1143         if (ret)
1144                 goto out;
1145
1146         spin_lock_irqsave(&conn->c_lock, flags);
1147         list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
1148         set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
1149         rds_message_addref(rm);
1150         rm->m_inc.i_conn = conn;
1151
1152         rds_message_populate_header(&rm->m_inc.i_hdr, 0, dport,
1153                                     conn->c_next_tx_seq);
1154         conn->c_next_tx_seq++;
1155         spin_unlock_irqrestore(&conn->c_lock, flags);
1156
1157         rds_stats_inc(s_send_queued);
1158         rds_stats_inc(s_send_pong);
1159
1160         queue_delayed_work(rds_wq, &conn->c_send_w, 0);
1161         rds_message_put(rm);
1162         return 0;
1163
1164 out:
1165         if (rm)
1166                 rds_message_put(rm);
1167         return ret;
1168 }