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                         ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
 256                         if (ret)
 257                                 break;
 258                         conn->c_xmit_rdma_sent = 1;
 259
 260                         /* The transport owns the mapped memory for now.
 261                          * You can't unmap it while it's on the send queue */
 262                         set_bit(RDS_MSG_MAPPED, &rm->m_flags);
 263                 }
 264
 265                 if (rm->atomic.op_active && !conn->c_xmit_atomic_sent) {
 266                         ret = conn->c_trans->xmit_atomic(conn, rm);
 267                         if (ret)
 268                                 break;
 269                         conn->c_xmit_atomic_sent = 1;
 270                         /* The transport owns the mapped memory for now.
 271                          * You can't unmap it while it's on the send queue */
 272                         set_bit(RDS_MSG_MAPPED, &rm->m_flags);
 273                 }
 274
 275                 /*
 276                  * A number of cases require an RDS header to be sent
 277                  * even if there is no data.
 278                  * We permit 0-byte sends; rds-ping depends on this.
 279                  * However, if there are exclusively attached silent ops,
 280                  * we skip the hdr/data send, to enable silent operation.
 281                  */
 282                 if (rm->data.op_nents == 0) {
 283                         int ops_present;
 284                         int all_ops_are_silent = 1;
 285
 286                         ops_present = (rm->atomic.op_active || rm->rdma.op_active);
 287                         if (rm->atomic.op_active && !rm->atomic.op_silent)
 288                                 all_ops_are_silent = 0;
 289                         if (rm->rdma.op_active && !rm->rdma.op_silent)
 290                                 all_ops_are_silent = 0;
 291
 292                         if (ops_present && all_ops_are_silent
 293                             && !rm->m_rdma_cookie)
 294                                 rm->data.op_active = 0;
 295                 }
 296
 297                 if (rm->data.op_active && !conn->c_xmit_data_sent) {
 298                         ret = conn->c_trans->xmit(conn, rm,
 299                                                   conn->c_xmit_hdr_off,
 300                                                   conn->c_xmit_sg,
 301                                                   conn->c_xmit_data_off);
 302                         if (ret <= 0)
 303                                 break;
 304
 305                         if (conn->c_xmit_hdr_off < sizeof(struct rds_header)) {
 306                                 tmp = min_t(int, ret,
 307                                             sizeof(struct rds_header) -
 308                                             conn->c_xmit_hdr_off);
 309                                 conn->c_xmit_hdr_off += tmp;
 310                                 ret -= tmp;
 311                         }
 312
 313                         sg = &rm->data.op_sg[conn->c_xmit_sg];
 314                         while (ret) {
 315                                 tmp = min_t(int, ret, sg->length -
 316                                                       conn->c_xmit_data_off);
 317                                 conn->c_xmit_data_off += tmp;
 318                                 ret -= tmp;
 319                                 if (conn->c_xmit_data_off == sg->length) {
 320                                         conn->c_xmit_data_off = 0;
 321                                         sg++;
 322                                         conn->c_xmit_sg++;
 323                                         BUG_ON(ret != 0 &&
 324                                                conn->c_xmit_sg == rm->data.op_nents);
 325                                 }
 326                         }
 327
 328                         if (conn->c_xmit_hdr_off == sizeof(struct rds_header) &&
 329                             (conn->c_xmit_sg == rm->data.op_nents))
 330                                 conn->c_xmit_data_sent = 1;
 331                 }
 332
 333                 /*
 334                  * A rm will only take multiple times through this loop
 335                  * if there is a data op. Thus, if the data is sent (or there was
 336                  * none), then we're done with the rm.
 337                  */
 338                 if (!rm->data.op_active || conn->c_xmit_data_sent) {
 339                         conn->c_xmit_rm = NULL;
 340                         conn->c_xmit_sg = 0;
 341                         conn->c_xmit_hdr_off = 0;
 342                         conn->c_xmit_data_off = 0;
 343                         conn->c_xmit_rdma_sent = 0;
 344                         conn->c_xmit_atomic_sent = 0;
 345                         conn->c_xmit_data_sent = 0;
 346
 347                         rds_message_put(rm);
 348                 }
 349         }
 350
 351         /* Nuke any messages we decided not to retransmit. */
 352         if (!list_empty(&to_be_dropped))
 353                 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
 354
 355         if (conn->c_trans->xmit_complete)
 356                 conn->c_trans->xmit_complete(conn);
 357
 358         /*
 359          * We might be racing with another sender who queued a message but
 360          * backed off on noticing that we held the c_send_lock.  If we check
 361          * for queued messages after dropping the sem then either we'll
 362          * see the queued message or the queuer will get the sem.  If we
 363          * notice the queued message then we trigger an immediate retry.
 364          *
 365          * We need to be careful only to do this when we stopped processing
 366          * the send queue because it was empty.  It's the only way we
 367          * stop processing the loop when the transport hasn't taken
 368          * responsibility for forward progress.
 369          */
 370         mutex_unlock(&conn->c_send_lock);
 371
 372         if (send_quota == 0 && !was_empty) {
 373                 /* We exhausted the send quota, but there's work left to
 374                  * do. Return and (re-)schedule the send worker.
 375                  */
 376                 ret = -EAGAIN;
 377         }
 378
 379         if (ret == 0 && was_empty) {
 380                 /* A simple bit test would be way faster than taking the
 381                  * spin lock */
 382                 spin_lock_irqsave(&conn->c_lock, flags);
 383                 if (!list_empty(&conn->c_send_queue)) {
 384                         rds_stats_inc(s_send_sem_queue_raced);
 385                         ret = -EAGAIN;
 386                 }
 387                 spin_unlock_irqrestore(&conn->c_lock, flags);
 388         }
 389 out:
 390         return ret;
 391 }
 392
 393 static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
 394 {
 395         u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
 396
 397         assert_spin_locked(&rs->rs_lock);
 398
 399         BUG_ON(rs->rs_snd_bytes < len);
 400         rs->rs_snd_bytes -= len;
 401
 402         if (rs->rs_snd_bytes == 0)
 403                 rds_stats_inc(s_send_queue_empty);
 404 }
 405
 406 static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
 407                                     is_acked_func is_acked)
 408 {
 409         if (is_acked)
 410                 return is_acked(rm, ack);
 411         return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
 412 }
 413
 414 /*
 415  * Returns true if there are no messages on the send and retransmit queues
 416  * which have a sequence number greater than or equal to the given sequence
 417  * number.
 418  */
 419 int rds_send_acked_before(struct rds_connection *conn, u64 seq)
 420 {
 421         struct rds_message *rm, *tmp;
 422         int ret = 1;
 423
 424         spin_lock(&conn->c_lock);
 425
 426         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
 427                 if (be64_to_cpu(rm->m_inc.i_hdr.h_sequence) < seq)
 428                         ret = 0;
 429                 break;
 430         }
 431
 432         list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
 433                 if (be64_to_cpu(rm->m_inc.i_hdr.h_sequence) < seq)
 434                         ret = 0;
 435                 break;
 436         }
 437
 438         spin_unlock(&conn->c_lock);
 439
 440         return ret;
 441 }
 442
 443 /*
 444  * This is pretty similar to what happens below in the ACK
 445  * handling code - except that we call here as soon as we get
 446  * the IB send completion on the RDMA op and the accompanying
 447  * message.
 448  */
 449 void rds_rdma_send_complete(struct rds_message *rm, int status)
 450 {
 451         struct rds_sock *rs = NULL;
 452         struct rm_rdma_op *ro;
 453         struct rds_notifier *notifier;
 454         unsigned long flags;
 455
 456         spin_lock_irqsave(&rm->m_rs_lock, flags);
 457
 458         ro = &rm->rdma;
 459         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
 460             ro->op_active && ro->op_notify && ro->op_notifier) {
 461                 notifier = ro->op_notifier;
 462                 rs = rm->m_rs;
 463                 sock_hold(rds_rs_to_sk(rs));
 464
 465                 notifier->n_status = status;
 466                 spin_lock(&rs->rs_lock);
 467                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
 468                 spin_unlock(&rs->rs_lock);
 469
 470                 ro->op_notifier = NULL;
 471         }
 472
 473         spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 474
 475         if (rs) {
 476                 rds_wake_sk_sleep(rs);
 477                 sock_put(rds_rs_to_sk(rs));
 478         }
 479 }
 480 EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
 481
 482 /*
 483  * Just like above, except looks at atomic op
 484  */
 485 void rds_atomic_send_complete(struct rds_message *rm, int status)
 486 {
 487         struct rds_sock *rs = NULL;
 488         struct rm_atomic_op *ao;
 489         struct rds_notifier *notifier;
 490
 491         spin_lock(&rm->m_rs_lock);
 492
 493         ao = &rm->atomic;
 494         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
 495             && ao->op_active && ao->op_notify && ao->op_notifier) {
 496                 notifier = ao->op_notifier;
 497                 rs = rm->m_rs;
 498                 sock_hold(rds_rs_to_sk(rs));
 499
 500                 notifier->n_status = status;
 501                 spin_lock(&rs->rs_lock);
 502                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
 503                 spin_unlock(&rs->rs_lock);
 504
 505                 ao->op_notifier = NULL;
 506         }
 507
 508         spin_unlock(&rm->m_rs_lock);
 509
 510         if (rs) {
 511                 rds_wake_sk_sleep(rs);
 512                 sock_put(rds_rs_to_sk(rs));
 513         }
 514 }
 515 EXPORT_SYMBOL_GPL(rds_atomic_send_complete);
 516
 517 /*
 518  * This is the same as rds_rdma_send_complete except we
 519  * don't do any locking - we have all the ingredients (message,
 520  * socket, socket lock) and can just move the notifier.
 521  */
 522 static inline void
 523 __rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
 524 {
 525         struct rm_rdma_op *ro;
 526         struct rm_atomic_op *ao;
 527
 528         ro = &rm->rdma;
 529         if (ro->op_active && ro->op_notify && ro->op_notifier) {
 530                 ro->op_notifier->n_status = status;
 531                 list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue);
 532                 ro->op_notifier = NULL;
 533         }
 534
 535         ao = &rm->atomic;
 536         if (ao->op_active && ao->op_notify && ao->op_notifier) {
 537                 ao->op_notifier->n_status = status;
 538                 list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue);
 539                 ao->op_notifier = NULL;
 540         }
 541
 542         /* No need to wake the app - caller does this */
 543 }
 544
 545 /*
 546  * This is called from the IB send completion when we detect
 547  * a RDMA operation that failed with remote access error.
 548  * So speed is not an issue here.
 549  */
 550 struct rds_message *rds_send_get_message(struct rds_connection *conn,
 551                                          struct rm_rdma_op *op)
 552 {
 553         struct rds_message *rm, *tmp, *found = NULL;
 554         unsigned long flags;
 555
 556         spin_lock_irqsave(&conn->c_lock, flags);
 557
 558         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
 559                 if (&rm->rdma == op) {
 560                         atomic_inc(&rm->m_refcount);
 561                         found = rm;
 562                         goto out;
 563                 }
 564         }
 565
 566         list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
 567                 if (&rm->rdma == op) {
 568                         atomic_inc(&rm->m_refcount);
 569                         found = rm;
 570                         break;
 571                 }
 572         }
 573
 574 out:
 575         spin_unlock_irqrestore(&conn->c_lock, flags);
 576
 577         return found;
 578 }
 579 EXPORT_SYMBOL_GPL(rds_send_get_message);
 580
 581 /*
 582  * This removes messages from the socket's list if they're on it.  The list
 583  * argument must be private to the caller, we must be able to modify it
 584  * without locks.  The messages must have a reference held for their
 585  * position on the list.  This function will drop that reference after
 586  * removing the messages from the 'messages' list regardless of if it found
 587  * the messages on the socket list or not.
 588  */
 589 void rds_send_remove_from_sock(struct list_head *messages, int status)
 590 {
 591         unsigned long flags;
 592         struct rds_sock *rs = NULL;
 593         struct rds_message *rm;
 594
 595         while (!list_empty(messages)) {
 596                 int was_on_sock = 0;
 597
 598                 rm = list_entry(messages->next, struct rds_message,
 599                                 m_conn_item);
 600                 list_del_init(&rm->m_conn_item);
 601
 602                 /*
 603                  * If we see this flag cleared then we're *sure* that someone
 604                  * else beat us to removing it from the sock.  If we race
 605                  * with their flag update we'll get the lock and then really
 606                  * see that the flag has been cleared.
 607                  *
 608                  * The message spinlock makes sure nobody clears rm->m_rs
 609                  * while we're messing with it. It does not prevent the
 610                  * message from being removed from the socket, though.
 611                  */
 612                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 613                 if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags))
 614                         goto unlock_and_drop;
 615
 616                 if (rs != rm->m_rs) {
 617                         if (rs) {
 618                                 rds_wake_sk_sleep(rs);
 619                                 sock_put(rds_rs_to_sk(rs));
 620                         }
 621                         rs = rm->m_rs;
 622                         sock_hold(rds_rs_to_sk(rs));
 623                 }
 624                 spin_lock(&rs->rs_lock);
 625
 626                 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
 627                         struct rm_rdma_op *ro = &rm->rdma;
 628                         struct rds_notifier *notifier;
 629
 630                         list_del_init(&rm->m_sock_item);
 631                         rds_send_sndbuf_remove(rs, rm);
 632
 633                         if (ro->op_active && ro->op_notifier &&
 634                                (ro->op_notify || (ro->op_recverr && status))) {
 635                                 notifier = ro->op_notifier;
 636                                 list_add_tail(&notifier->n_list,
 637                                                 &rs->rs_notify_queue);
 638                                 if (!notifier->n_status)
 639                                         notifier->n_status = status;
 640                                 rm->rdma.op_notifier = NULL;
 641                         }
 642                         was_on_sock = 1;
 643                         rm->m_rs = NULL;
 644                 }
 645                 spin_unlock(&rs->rs_lock);
 646
 647 unlock_and_drop:
 648                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 649                 rds_message_put(rm);
 650                 if (was_on_sock)
 651                         rds_message_put(rm);
 652         }
 653
 654         if (rs) {
 655                 rds_wake_sk_sleep(rs);
 656                 sock_put(rds_rs_to_sk(rs));
 657         }
 658 }
 659
 660 /*
 661  * Transports call here when they've determined that the receiver queued
 662  * messages up to, and including, the given sequence number.  Messages are
 663  * moved to the retrans queue when rds_send_xmit picks them off the send
 664  * queue. This means that in the TCP case, the message may not have been
 665  * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
 666  * checks the RDS_MSG_HAS_ACK_SEQ bit.
 667  *
 668  * XXX It's not clear to me how this is safely serialized with socket
 669  * destruction.  Maybe it should bail if it sees SOCK_DEAD.
 670  */
 671 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
 672                          is_acked_func is_acked)
 673 {
 674         struct rds_message *rm, *tmp;
 675         unsigned long flags;
 676         LIST_HEAD(list);
 677
 678         spin_lock_irqsave(&conn->c_lock, flags);
 679
 680         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
 681                 if (!rds_send_is_acked(rm, ack, is_acked))
 682                         break;
 683
 684                 list_move(&rm->m_conn_item, &list);
 685                 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
 686         }
 687
 688         /* order flag updates with spin locks */
 689         if (!list_empty(&list))
 690                 smp_mb__after_clear_bit();
 691
 692         spin_unlock_irqrestore(&conn->c_lock, flags);
 693
 694         /* now remove the messages from the sock list as needed */
 695         rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
 696 }
 697 EXPORT_SYMBOL_GPL(rds_send_drop_acked);
 698
 699 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
 700 {
 701         struct rds_message *rm, *tmp;
 702         struct rds_connection *conn;
 703         unsigned long flags;
 704         LIST_HEAD(list);
 705
 706         /* get all the messages we're dropping under the rs lock */
 707         spin_lock_irqsave(&rs->rs_lock, flags);
 708
 709         list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) {
 710                 if (dest && (dest->sin_addr.s_addr != rm->m_daddr ||
 711                              dest->sin_port != rm->m_inc.i_hdr.h_dport))
 712                         continue;
 713
 714                 list_move(&rm->m_sock_item, &list);
 715                 rds_send_sndbuf_remove(rs, rm);
 716                 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
 717         }
 718
 719         /* order flag updates with the rs lock */
 720         smp_mb__after_clear_bit();
 721
 722         spin_unlock_irqrestore(&rs->rs_lock, flags);
 723
 724         if (list_empty(&list))
 725                 return;
 726
 727         /* Remove the messages from the conn */
 728         list_for_each_entry(rm, &list, m_sock_item) {
 729
 730                 conn = rm->m_inc.i_conn;
 731
 732                 spin_lock_irqsave(&conn->c_lock, flags);
 733                 /*
 734                  * Maybe someone else beat us to removing rm from the conn.
 735                  * If we race with their flag update we'll get the lock and
 736                  * then really see that the flag has been cleared.
 737                  */
 738                 if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
 739                         spin_unlock_irqrestore(&conn->c_lock, flags);
 740                         continue;
 741                 }
 742                 list_del_init(&rm->m_conn_item);
 743                 spin_unlock_irqrestore(&conn->c_lock, flags);
 744
 745                 /*
 746                  * Couldn't grab m_rs_lock in top loop (lock ordering),
 747                  * but we can now.
 748                  */
 749                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 750
 751                 spin_lock(&rs->rs_lock);
 752                 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
 753                 spin_unlock(&rs->rs_lock);
 754
 755                 rm->m_rs = NULL;
 756                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 757
 758                 rds_message_put(rm);
 759         }
 760
 761         rds_wake_sk_sleep(rs);
 762
 763         while (!list_empty(&list)) {
 764                 rm = list_entry(list.next, struct rds_message, m_sock_item);
 765                 list_del_init(&rm->m_sock_item);
 766
 767                 rds_message_wait(rm);
 768                 rds_message_put(rm);
 769         }
 770 }
 771
 772 /*
 773  * we only want this to fire once so we use the callers 'queued'.  It's
 774  * possible that another thread can race with us and remove the
 775  * message from the flow with RDS_CANCEL_SENT_TO.
 776  */
 777 static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
 778                              struct rds_message *rm, __be16 sport,
 779                              __be16 dport, int *queued)
 780 {
 781         unsigned long flags;
 782         u32 len;
 783
 784         if (*queued)
 785                 goto out;
 786
 787         len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
 788
 789         /* this is the only place which holds both the socket's rs_lock
 790          * and the connection's c_lock */
 791         spin_lock_irqsave(&rs->rs_lock, flags);
 792
 793         /*
 794          * If there is a little space in sndbuf, we don't queue anything,
 795          * and userspace gets -EAGAIN. But poll() indicates there's send
 796          * room. This can lead to bad behavior (spinning) if snd_bytes isn't
 797          * freed up by incoming acks. So we check the *old* value of
 798          * rs_snd_bytes here to allow the last msg to exceed the buffer,
 799          * and poll() now knows no more data can be sent.
 800          */
 801         if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
 802                 rs->rs_snd_bytes += len;
 803
 804                 /* let recv side know we are close to send space exhaustion.
 805                  * This is probably not the optimal way to do it, as this
 806                  * means we set the flag on *all* messages as soon as our
 807                  * throughput hits a certain threshold.
 808                  */
 809                 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
 810                         __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
 811
 812                 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
 813                 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
 814                 rds_message_addref(rm);
 815                 rm->m_rs = rs;
 816
 817                 /* The code ordering is a little weird, but we're
 818                    trying to minimize the time we hold c_lock */
 819                 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
 820                 rm->m_inc.i_conn = conn;
 821                 rds_message_addref(rm);
 822
 823                 spin_lock(&conn->c_lock);
 824                 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(conn->c_next_tx_seq++);
 825                 list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
 826                 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
 827                 spin_unlock(&conn->c_lock);
 828
 829                 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
 830                          rm, len, rs, rs->rs_snd_bytes,
 831                          (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
 832
 833                 *queued = 1;
 834         }
 835
 836         spin_unlock_irqrestore(&rs->rs_lock, flags);
 837 out:
 838         return *queued;
 839 }
 840
 841 /*
 842  * rds_message is getting to be quite complicated, and we'd like to allocate
 843  * it all in one go. This figures out how big it needs to be up front.
 844  */
 845 static int rds_rm_size(struct msghdr *msg, int data_len)
 846 {
 847         struct cmsghdr *cmsg;
 848         int size = 0;
 849         int retval;
 850
 851         for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
 852                 if (!CMSG_OK(msg, cmsg))
 853                         return -EINVAL;
 854
 855                 if (cmsg->cmsg_level != SOL_RDS)
 856                         continue;
 857
 858                 switch (cmsg->cmsg_type) {
 859                 case RDS_CMSG_RDMA_ARGS:
 860                         retval = rds_rdma_extra_size(CMSG_DATA(cmsg));
 861                         if (retval < 0)
 862                                 return retval;
 863                         size += retval;
 864                         break;
 865
 866                 case RDS_CMSG_RDMA_DEST:
 867                 case RDS_CMSG_RDMA_MAP:
 868                         /* these are valid but do no add any size */
 869                         break;
 870
 871                 case RDS_CMSG_ATOMIC_CSWP:
 872                 case RDS_CMSG_ATOMIC_FADD:
 873                         size += sizeof(struct scatterlist);
 874                         break;
 875
 876                 default:
 877                         return -EINVAL;
 878                 }
 879
 880         }
 881
 882         size += ceil(data_len, PAGE_SIZE) * sizeof(struct scatterlist);
 883
 884         return size;
 885 }
 886
 887 static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
 888                          struct msghdr *msg, int *allocated_mr)
 889 {
 890         struct cmsghdr *cmsg;
 891         int ret = 0;
 892
 893         for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
 894                 if (!CMSG_OK(msg, cmsg))
 895                         return -EINVAL;
 896
 897                 if (cmsg->cmsg_level != SOL_RDS)
 898                         continue;
 899
 900                 /* As a side effect, RDMA_DEST and RDMA_MAP will set
 901                  * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
 902                  */
 903                 switch (cmsg->cmsg_type) {
 904                 case RDS_CMSG_RDMA_ARGS:
 905                         ret = rds_cmsg_rdma_args(rs, rm, cmsg);
 906                         break;
 907
 908                 case RDS_CMSG_RDMA_DEST:
 909                         ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
 910                         break;
 911
 912                 case RDS_CMSG_RDMA_MAP:
 913                         ret = rds_cmsg_rdma_map(rs, rm, cmsg);
 914                         if (!ret)
 915                                 *allocated_mr = 1;
 916                         break;
 917                 case RDS_CMSG_ATOMIC_CSWP:
 918                 case RDS_CMSG_ATOMIC_FADD:
 919                         ret = rds_cmsg_atomic(rs, rm, cmsg);
 920                         break;
 921
 922                 default:
 923                         return -EINVAL;
 924                 }
 925
 926                 if (ret)
 927                         break;
 928         }
 929
 930         return ret;
 931 }
 932
 933 int rds_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
 934                 size_t payload_len)
 935 {
 936         struct sock *sk = sock->sk;
 937         struct rds_sock *rs = rds_sk_to_rs(sk);
 938         struct sockaddr_in *usin = (struct sockaddr_in *)msg->msg_name;
 939         __be32 daddr;
 940         __be16 dport;
 941         struct rds_message *rm = NULL;
 942         struct rds_connection *conn;
 943         int ret = 0;
 944         int queued = 0, allocated_mr = 0;
 945         int nonblock = msg->msg_flags & MSG_DONTWAIT;
 946         long timeo = sock_sndtimeo(sk, nonblock);
 947
 948         /* Mirror Linux UDP mirror of BSD error message compatibility */
 949         /* XXX: Perhaps MSG_MORE someday */
 950         if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT)) {
 951                 printk(KERN_INFO "msg_flags 0x%08X\n", msg->msg_flags);
 952                 ret = -EOPNOTSUPP;
 953                 goto out;
 954         }
 955
 956         if (msg->msg_namelen) {
 957                 /* XXX fail non-unicast destination IPs? */
 958                 if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) {
 959                         ret = -EINVAL;
 960                         goto out;
 961                 }
 962                 daddr = usin->sin_addr.s_addr;
 963                 dport = usin->sin_port;
 964         } else {
 965                 /* We only care about consistency with ->connect() */
 966                 lock_sock(sk);
 967                 daddr = rs->rs_conn_addr;
 968                 dport = rs->rs_conn_port;
 969                 release_sock(sk);
 970         }
 971
 972         /* racing with another thread binding seems ok here */
 973         if (daddr == 0 || rs->rs_bound_addr == 0) {
 974                 ret = -ENOTCONN; /* XXX not a great errno */
 975                 goto out;
 976         }
 977
 978         /* size of rm including all sgs */
 979         ret = rds_rm_size(msg, payload_len);
 980         if (ret < 0)
 981                 goto out;
 982
 983         rm = rds_message_alloc(ret, GFP_KERNEL);
 984         if (!rm) {
 985                 ret = -ENOMEM;
 986                 goto out;
 987         }
 988
 989         /* Attach data to the rm */
 990         if (payload_len) {
 991                 rm->data.op_sg = rds_message_alloc_sgs(rm, ceil(payload_len, PAGE_SIZE));
 992                 ret = rds_message_copy_from_user(rm, msg->msg_iov, payload_len);
 993                 if (ret)
 994                         goto out;
 995         }
 996         rm->data.op_active = 1;
 997
 998         rm->m_daddr = daddr;
 999
1000         /* rds_conn_create has a spinlock that runs with IRQ off.
1001          * Caching the conn in the socket helps a lot. */
1002         if (rs->rs_conn && rs->rs_conn->c_faddr == daddr)
1003                 conn = rs->rs_conn;
1004         else {
1005                 conn = rds_conn_create_outgoing(rs->rs_bound_addr, daddr,
1006                                         rs->rs_transport,
1007                                         sock->sk->sk_allocation);
1008                 if (IS_ERR(conn)) {
1009                         ret = PTR_ERR(conn);
1010                         goto out;
1011                 }
1012                 rs->rs_conn = conn;
1013         }
1014
1015         /* Parse any control messages the user may have included. */
1016         ret = rds_cmsg_send(rs, rm, msg, &allocated_mr);
1017         if (ret)
1018                 goto out;
1019
1020         if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) {
1021                 if (printk_ratelimit())
1022                         printk(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
1023                                &rm->rdma, conn->c_trans->xmit_rdma);
1024                 ret = -EOPNOTSUPP;
1025                 goto out;
1026         }
1027
1028         if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
1029                 if (printk_ratelimit())
1030                         printk(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
1031                                &rm->atomic, conn->c_trans->xmit_atomic);
1032                 ret = -EOPNOTSUPP;
1033                 goto out;
1034         }
1035
1036         /* If the connection is down, trigger a connect. We may
1037          * have scheduled a delayed reconnect however - in this case
1038          * we should not interfere.
1039          */
1040         if (rds_conn_state(conn) == RDS_CONN_DOWN &&
1041             !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
1042                 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
1043
1044         ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
1045         if (ret) {
1046                 rs->rs_seen_congestion = 1;
1047                 goto out;
1048         }
1049
1050         while (!rds_send_queue_rm(rs, conn, rm, rs->rs_bound_port,
1051                                   dport, &queued)) {
1052                 rds_stats_inc(s_send_queue_full);
1053                 /* XXX make sure this is reasonable */
1054                 if (payload_len > rds_sk_sndbuf(rs)) {
1055                         ret = -EMSGSIZE;
1056                         goto out;
1057                 }
1058                 if (nonblock) {
1059                         ret = -EAGAIN;
1060                         goto out;
1061                 }
1062
1063                 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
1064                                         rds_send_queue_rm(rs, conn, rm,
1065                                                           rs->rs_bound_port,
1066                                                           dport,
1067                                                           &queued),
1068                                         timeo);
1069                 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
1070                 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
1071                         continue;
1072
1073                 ret = timeo;
1074                 if (ret == 0)
1075                         ret = -ETIMEDOUT;
1076                 goto out;
1077         }
1078
1079         /*
1080          * By now we've committed to the send.  We reuse rds_send_worker()
1081          * to retry sends in the rds thread if the transport asks us to.
1082          */
1083         rds_stats_inc(s_send_queued);
1084
1085         if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
1086                 rds_send_worker(&conn->c_send_w.work);
1087
1088         rds_message_put(rm);
1089         return payload_len;
1090
1091 out:
1092         /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1093          * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1094          * or in any other way, we need to destroy the MR again */
1095         if (allocated_mr)
1096                 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
1097
1098         if (rm)
1099                 rds_message_put(rm);
1100         return ret;
1101 }
1102
1103 /*
1104  * Reply to a ping packet.
1105  */
1106 int
1107 rds_send_pong(struct rds_connection *conn, __be16 dport)
1108 {
1109         struct rds_message *rm;
1110         unsigned long flags;
1111         int ret = 0;
1112
1113         rm = rds_message_alloc(0, GFP_ATOMIC);
1114         if (!rm) {
1115                 ret = -ENOMEM;
1116                 goto out;
1117         }
1118
1119         rm->m_daddr = conn->c_faddr;
1120
1121         /* If the connection is down, trigger a connect. We may
1122          * have scheduled a delayed reconnect however - in this case
1123          * we should not interfere.
1124          */
1125         if (rds_conn_state(conn) == RDS_CONN_DOWN &&
1126             !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
1127                 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
1128
1129         ret = rds_cong_wait(conn->c_fcong, dport, 1, NULL);
1130         if (ret)
1131                 goto out;
1132
1133         spin_lock_irqsave(&conn->c_lock, flags);
1134         list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
1135         set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
1136         rds_message_addref(rm);
1137         rm->m_inc.i_conn = conn;
1138
1139         rds_message_populate_header(&rm->m_inc.i_hdr, 0, dport,
1140                                     conn->c_next_tx_seq);
1141         conn->c_next_tx_seq++;
1142         spin_unlock_irqrestore(&conn->c_lock, flags);
1143
1144         rds_stats_inc(s_send_queued);
1145         rds_stats_inc(s_send_pong);
1146
1147         queue_delayed_work(rds_wq, &conn->c_send_w, 0);
1148         rds_message_put(rm);
1149         return 0;
1150
1151 out:
1152         if (rm)
1153                 rds_message_put(rm);
1154         return ret;
1155 }