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_rdma_sent = 0;
  76         conn->c_xmit_atomic_sent = 0;
  77
  78         conn->c_map_queued = 0;
  79
  80         conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
  81         conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
  82
  83         /* Mark messages as retransmissions, and move them to the send q */
  84         spin_lock_irqsave(&conn->c_lock, flags);
  85         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
  86                 set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
  87                 set_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags);
  88         }
  89         list_splice_init(&conn->c_retrans, &conn->c_send_queue);
  90         spin_unlock_irqrestore(&conn->c_lock, flags);
  91 }
  92
  93 /*
  94  * We're making the concious trade-off here to only send one message
  95  * down the connection at a time.
  96  *   Pro:
  97  *      - tx queueing is a simple fifo list
  98  *      - reassembly is optional and easily done by transports per conn
  99  *      - no per flow rx lookup at all, straight to the socket
 100  *      - less per-frag memory and wire overhead
 101  *   Con:
 102  *      - queued acks can be delayed behind large messages
 103  *   Depends:
 104  *      - small message latency is higher behind queued large messages
 105  *      - large message latency isn't starved by intervening small sends
 106  */
 107 int rds_send_xmit(struct rds_connection *conn)
 108 {
 109         struct rds_message *rm;
 110         unsigned long flags;
 111         unsigned int tmp;
 112         unsigned int send_quota = send_batch_count;
 113         struct scatterlist *sg;
 114         int ret = 0;
 115         int was_empty = 0;
 116         LIST_HEAD(to_be_dropped);
 117
 118         /*
 119          * sendmsg calls here after having queued its message on the send
 120          * queue.  We only have one task feeding the connection at a time.  If
 121          * another thread is already feeding the queue then we back off.  This
 122          * avoids blocking the caller and trading per-connection data between
 123          * caches per message.
 124          *
 125          * The sem holder will issue a retry if they notice that someone queued
 126          * a message after they stopped walking the send queue but before they
 127          * dropped the sem.
 128          */
 129         if (!mutex_trylock(&conn->c_send_lock)) {
 130                 rds_stats_inc(s_send_sem_contention);
 131                 ret = -ENOMEM;
 132                 goto out;
 133         }
 134
 135         if (conn->c_trans->xmit_prepare)
 136                 conn->c_trans->xmit_prepare(conn);
 137
 138         /*
 139          * spin trying to push headers and data down the connection until
 140          * the connection doens't make forward progress.
 141          */
 142         while (--send_quota) {
 143                 /*
 144                  * See if need to send a congestion map update if we're
 145                  * between sending messages.  The send_sem protects our sole
 146                  * use of c_map_offset and _bytes.
 147                  * Note this is used only by transports that define a special
 148                  * xmit_cong_map function. For all others, we create allocate
 149                  * a cong_map message and treat it just like any other send.
 150                  */
 151                 if (conn->c_map_bytes) {
 152                         ret = conn->c_trans->xmit_cong_map(conn, conn->c_lcong,
 153                                                 conn->c_map_offset);
 154                         if (ret <= 0)
 155                                 break;
 156
 157                         conn->c_map_offset += ret;
 158                         conn->c_map_bytes -= ret;
 159                         if (conn->c_map_bytes)
 160                                 continue;
 161                 }
 162
 163                 /* If we're done sending the current message, clear the
 164                  * offset and S/G temporaries.
 165                  */
 166                 rm = conn->c_xmit_rm;
 167                 if (rm &&
 168                     conn->c_xmit_hdr_off == sizeof(struct rds_header) &&
 169                     conn->c_xmit_sg == rm->data.m_nents) {
 170                         conn->c_xmit_rm = NULL;
 171                         conn->c_xmit_sg = 0;
 172                         conn->c_xmit_hdr_off = 0;
 173                         conn->c_xmit_data_off = 0;
 174                         conn->c_xmit_rdma_sent = 0;
 175                         conn->c_xmit_atomic_sent = 0;
 176
 177                         /* Release the reference to the previous message. */
 178                         rds_message_put(rm);
 179                         rm = NULL;
 180                 }
 181
 182                 /* If we're asked to send a cong map update, do so.
 183                  */
 184                 if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) {
 185                         if (conn->c_trans->xmit_cong_map) {
 186                                 conn->c_map_offset = 0;
 187                                 conn->c_map_bytes = sizeof(struct rds_header) +
 188                                         RDS_CONG_MAP_BYTES;
 189                                 continue;
 190                         }
 191
 192                         rm = rds_cong_update_alloc(conn);
 193                         if (IS_ERR(rm)) {
 194                                 ret = PTR_ERR(rm);
 195                                 break;
 196                         }
 197
 198                         conn->c_xmit_rm = rm;
 199                 }
 200
 201                 /*
 202                  * Grab the next message from the send queue, if there is one.
 203                  *
 204                  * c_xmit_rm holds a ref while we're sending this message down
 205                  * the connction.  We can use this ref while holding the
 206                  * send_sem.. rds_send_reset() is serialized with it.
 207                  */
 208                 if (!rm) {
 209                         unsigned int len;
 210
 211                         spin_lock_irqsave(&conn->c_lock, flags);
 212
 213                         if (!list_empty(&conn->c_send_queue)) {
 214                                 rm = list_entry(conn->c_send_queue.next,
 215                                                 struct rds_message,
 216                                                 m_conn_item);
 217                                 rds_message_addref(rm);
 218
 219                                 /*
 220                                  * Move the message from the send queue to the retransmit
 221                                  * list right away.
 222                                  */
 223                                 list_move_tail(&rm->m_conn_item, &conn->c_retrans);
 224                         }
 225
 226                         spin_unlock_irqrestore(&conn->c_lock, flags);
 227
 228                         if (!rm) {
 229                                 was_empty = 1;
 230                                 break;
 231                         }
 232
 233                         /* Unfortunately, the way Infiniband deals with
 234                          * RDMA to a bad MR key is by moving the entire
 235                          * queue pair to error state. We cold possibly
 236                          * recover from that, but right now we drop the
 237                          * connection.
 238                          * Therefore, we never retransmit messages with RDMA ops.
 239                          */
 240                         if (rm->rdma.m_rdma_op.r_active &&
 241                             test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags)) {
 242                                 spin_lock_irqsave(&conn->c_lock, flags);
 243                                 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags))
 244                                         list_move(&rm->m_conn_item, &to_be_dropped);
 245                                 spin_unlock_irqrestore(&conn->c_lock, flags);
 246                                 rds_message_put(rm);
 247                                 continue;
 248                         }
 249
 250                         /* Require an ACK every once in a while */
 251                         len = ntohl(rm->m_inc.i_hdr.h_len);
 252                         if (conn->c_unacked_packets == 0 ||
 253                             conn->c_unacked_bytes < len) {
 254                                 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
 255
 256                                 conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
 257                                 conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
 258                                 rds_stats_inc(s_send_ack_required);
 259                         } else {
 260                                 conn->c_unacked_bytes -= len;
 261                                 conn->c_unacked_packets--;
 262                         }
 263
 264                         conn->c_xmit_rm = rm;
 265                 }
 266
 267
 268                 if (rm->atomic.op_active && !conn->c_xmit_atomic_sent) {
 269                         ret = conn->c_trans->xmit_atomic(conn, &rm->atomic);
 270                         if (ret)
 271                                 break;
 272                         conn->c_xmit_atomic_sent = 1;
 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                  * Try and send an rdma message.  Let's see if we can
 280                  * keep this simple and require that the transport either
 281                  * send the whole rdma or none of it.
 282                  */
 283                 if (rm->rdma.m_rdma_op.r_active && !conn->c_xmit_rdma_sent) {
 284                         ret = conn->c_trans->xmit_rdma(conn, &rm->rdma.m_rdma_op);
 285                         if (ret)
 286                                 break;
 287                         conn->c_xmit_rdma_sent = 1;
 288                         /* The transport owns the mapped memory for now.
 289                          * You can't unmap it while it's on the send queue */
 290                         set_bit(RDS_MSG_MAPPED, &rm->m_flags);
 291                 }
 292
 293                 if (conn->c_xmit_hdr_off < sizeof(struct rds_header) ||
 294                     conn->c_xmit_sg < rm->data.m_nents) {
 295                         ret = conn->c_trans->xmit(conn, rm,
 296                                                   conn->c_xmit_hdr_off,
 297                                                   conn->c_xmit_sg,
 298                                                   conn->c_xmit_data_off);
 299                         if (ret <= 0)
 300                                 break;
 301
 302                         if (conn->c_xmit_hdr_off < sizeof(struct rds_header)) {
 303                                 tmp = min_t(int, ret,
 304                                             sizeof(struct rds_header) -
 305                                             conn->c_xmit_hdr_off);
 306                                 conn->c_xmit_hdr_off += tmp;
 307                                 ret -= tmp;
 308                         }
 309
 310                         sg = &rm->data.m_sg[conn->c_xmit_sg];
 311                         while (ret) {
 312                                 tmp = min_t(int, ret, sg->length -
 313                                                       conn->c_xmit_data_off);
 314                                 conn->c_xmit_data_off += tmp;
 315                                 ret -= tmp;
 316                                 if (conn->c_xmit_data_off == sg->length) {
 317                                         conn->c_xmit_data_off = 0;
 318                                         sg++;
 319                                         conn->c_xmit_sg++;
 320                                         BUG_ON(ret != 0 &&
 321                                                conn->c_xmit_sg == rm->data.m_nents);
 322                                 }
 323                         }
 324                 }
 325         }
 326
 327         /* Nuke any messages we decided not to retransmit. */
 328         if (!list_empty(&to_be_dropped))
 329                 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
 330
 331         if (conn->c_trans->xmit_complete)
 332                 conn->c_trans->xmit_complete(conn);
 333
 334         /*
 335          * We might be racing with another sender who queued a message but
 336          * backed off on noticing that we held the c_send_lock.  If we check
 337          * for queued messages after dropping the sem then either we'll
 338          * see the queued message or the queuer will get the sem.  If we
 339          * notice the queued message then we trigger an immediate retry.
 340          *
 341          * We need to be careful only to do this when we stopped processing
 342          * the send queue because it was empty.  It's the only way we
 343          * stop processing the loop when the transport hasn't taken
 344          * responsibility for forward progress.
 345          */
 346         mutex_unlock(&conn->c_send_lock);
 347
 348         if (conn->c_map_bytes || (send_quota == 0 && !was_empty)) {
 349                 /* We exhausted the send quota, but there's work left to
 350                  * do. Return and (re-)schedule the send worker.
 351                  */
 352                 ret = -EAGAIN;
 353         }
 354
 355         if (ret == 0 && was_empty) {
 356                 /* A simple bit test would be way faster than taking the
 357                  * spin lock */
 358                 spin_lock_irqsave(&conn->c_lock, flags);
 359                 if (!list_empty(&conn->c_send_queue)) {
 360                         rds_stats_inc(s_send_sem_queue_raced);
 361                         ret = -EAGAIN;
 362                 }
 363                 spin_unlock_irqrestore(&conn->c_lock, flags);
 364         }
 365 out:
 366         return ret;
 367 }
 368
 369 static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
 370 {
 371         u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
 372
 373         assert_spin_locked(&rs->rs_lock);
 374
 375         BUG_ON(rs->rs_snd_bytes < len);
 376         rs->rs_snd_bytes -= len;
 377
 378         if (rs->rs_snd_bytes == 0)
 379                 rds_stats_inc(s_send_queue_empty);
 380 }
 381
 382 static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
 383                                     is_acked_func is_acked)
 384 {
 385         if (is_acked)
 386                 return is_acked(rm, ack);
 387         return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
 388 }
 389
 390 /*
 391  * Returns true if there are no messages on the send and retransmit queues
 392  * which have a sequence number greater than or equal to the given sequence
 393  * number.
 394  */
 395 int rds_send_acked_before(struct rds_connection *conn, u64 seq)
 396 {
 397         struct rds_message *rm, *tmp;
 398         int ret = 1;
 399
 400         spin_lock(&conn->c_lock);
 401
 402         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
 403                 if (be64_to_cpu(rm->m_inc.i_hdr.h_sequence) < seq)
 404                         ret = 0;
 405                 break;
 406         }
 407
 408         list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
 409                 if (be64_to_cpu(rm->m_inc.i_hdr.h_sequence) < seq)
 410                         ret = 0;
 411                 break;
 412         }
 413
 414         spin_unlock(&conn->c_lock);
 415
 416         return ret;
 417 }
 418
 419 /*
 420  * This is pretty similar to what happens below in the ACK
 421  * handling code - except that we call here as soon as we get
 422  * the IB send completion on the RDMA op and the accompanying
 423  * message.
 424  */
 425 void rds_rdma_send_complete(struct rds_message *rm, int status)
 426 {
 427         struct rds_sock *rs = NULL;
 428         struct rds_rdma_op *ro;
 429         struct rds_notifier *notifier;
 430         unsigned long flags;
 431
 432         spin_lock_irqsave(&rm->m_rs_lock, flags);
 433
 434         ro = &rm->rdma.m_rdma_op;
 435         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
 436             ro->r_active && ro->r_notify && ro->r_notifier) {
 437                 notifier = ro->r_notifier;
 438                 rs = rm->m_rs;
 439                 sock_hold(rds_rs_to_sk(rs));
 440
 441                 notifier->n_status = status;
 442                 spin_lock(&rs->rs_lock);
 443                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
 444                 spin_unlock(&rs->rs_lock);
 445
 446                 ro->r_notifier = NULL;
 447         }
 448
 449         spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 450
 451         if (rs) {
 452                 rds_wake_sk_sleep(rs);
 453                 sock_put(rds_rs_to_sk(rs));
 454         }
 455 }
 456 EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
 457
 458 /*
 459  * Just like above, except looks at atomic op
 460  */
 461 void rds_atomic_send_complete(struct rds_message *rm, int status)
 462 {
 463         struct rds_sock *rs = NULL;
 464         struct rm_atomic_op *ao;
 465         struct rds_notifier *notifier;
 466
 467         spin_lock(&rm->m_rs_lock);
 468
 469         ao = &rm->atomic;
 470         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
 471             && ao->op_active && ao->op_notify && ao->op_notifier) {
 472                 notifier = ao->op_notifier;
 473                 rs = rm->m_rs;
 474                 sock_hold(rds_rs_to_sk(rs));
 475
 476                 notifier->n_status = status;
 477                 spin_lock(&rs->rs_lock);
 478                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
 479                 spin_unlock(&rs->rs_lock);
 480
 481                 ao->op_notifier = NULL;
 482         }
 483
 484         spin_unlock(&rm->m_rs_lock);
 485
 486         if (rs) {
 487                 rds_wake_sk_sleep(rs);
 488                 sock_put(rds_rs_to_sk(rs));
 489         }
 490 }
 491 EXPORT_SYMBOL_GPL(rds_atomic_send_complete);
 492
 493 /*
 494  * This is the same as rds_rdma_send_complete except we
 495  * don't do any locking - we have all the ingredients (message,
 496  * socket, socket lock) and can just move the notifier.
 497  */
 498 static inline void
 499 __rds_rdma_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
 500 {
 501         struct rds_rdma_op *ro;
 502
 503         ro = &rm->rdma.m_rdma_op;
 504         if (ro->r_active && ro->r_notify && ro->r_notifier) {
 505                 ro->r_notifier->n_status = status;
 506                 list_add_tail(&ro->r_notifier->n_list, &rs->rs_notify_queue);
 507                 ro->r_notifier = NULL;
 508         }
 509
 510         /* No need to wake the app - caller does this */
 511 }
 512
 513 /*
 514  * This is called from the IB send completion when we detect
 515  * a RDMA operation that failed with remote access error.
 516  * So speed is not an issue here.
 517  */
 518 struct rds_message *rds_send_get_message(struct rds_connection *conn,
 519                                          struct rds_rdma_op *op)
 520 {
 521         struct rds_message *rm, *tmp, *found = NULL;
 522         unsigned long flags;
 523
 524         spin_lock_irqsave(&conn->c_lock, flags);
 525
 526         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
 527                 if (&rm->rdma.m_rdma_op == op) {
 528                         atomic_inc(&rm->m_refcount);
 529                         found = rm;
 530                         goto out;
 531                 }
 532         }
 533
 534         list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
 535                 if (&rm->rdma.m_rdma_op == op) {
 536                         atomic_inc(&rm->m_refcount);
 537                         found = rm;
 538                         break;
 539                 }
 540         }
 541
 542 out:
 543         spin_unlock_irqrestore(&conn->c_lock, flags);
 544
 545         return found;
 546 }
 547 EXPORT_SYMBOL_GPL(rds_send_get_message);
 548
 549 /*
 550  * This removes messages from the socket's list if they're on it.  The list
 551  * argument must be private to the caller, we must be able to modify it
 552  * without locks.  The messages must have a reference held for their
 553  * position on the list.  This function will drop that reference after
 554  * removing the messages from the 'messages' list regardless of if it found
 555  * the messages on the socket list or not.
 556  */
 557 void rds_send_remove_from_sock(struct list_head *messages, int status)
 558 {
 559         unsigned long flags;
 560         struct rds_sock *rs = NULL;
 561         struct rds_message *rm;
 562
 563         while (!list_empty(messages)) {
 564                 int was_on_sock = 0;
 565
 566                 rm = list_entry(messages->next, struct rds_message,
 567                                 m_conn_item);
 568                 list_del_init(&rm->m_conn_item);
 569
 570                 /*
 571                  * If we see this flag cleared then we're *sure* that someone
 572                  * else beat us to removing it from the sock.  If we race
 573                  * with their flag update we'll get the lock and then really
 574                  * see that the flag has been cleared.
 575                  *
 576                  * The message spinlock makes sure nobody clears rm->m_rs
 577                  * while we're messing with it. It does not prevent the
 578                  * message from being removed from the socket, though.
 579                  */
 580                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 581                 if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags))
 582                         goto unlock_and_drop;
 583
 584                 if (rs != rm->m_rs) {
 585                         if (rs) {
 586                                 rds_wake_sk_sleep(rs);
 587                                 sock_put(rds_rs_to_sk(rs));
 588                         }
 589                         rs = rm->m_rs;
 590                         sock_hold(rds_rs_to_sk(rs));
 591                 }
 592                 spin_lock(&rs->rs_lock);
 593
 594                 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
 595                         struct rds_rdma_op *ro = &rm->rdma.m_rdma_op;
 596                         struct rds_notifier *notifier;
 597
 598                         list_del_init(&rm->m_sock_item);
 599                         rds_send_sndbuf_remove(rs, rm);
 600
 601                         if (ro->r_active && ro->r_notifier &&
 602                             (ro->r_notify || (ro->r_recverr && status))) {
 603                                 notifier = ro->r_notifier;
 604                                 list_add_tail(&notifier->n_list,
 605                                                 &rs->rs_notify_queue);
 606                                 if (!notifier->n_status)
 607                                         notifier->n_status = status;
 608                                 rm->rdma.m_rdma_op.r_notifier = NULL;
 609                         }
 610                         was_on_sock = 1;
 611                         rm->m_rs = NULL;
 612                 }
 613                 spin_unlock(&rs->rs_lock);
 614
 615 unlock_and_drop:
 616                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 617                 rds_message_put(rm);
 618                 if (was_on_sock)
 619                         rds_message_put(rm);
 620         }
 621
 622         if (rs) {
 623                 rds_wake_sk_sleep(rs);
 624                 sock_put(rds_rs_to_sk(rs));
 625         }
 626 }
 627
 628 /*
 629  * Transports call here when they've determined that the receiver queued
 630  * messages up to, and including, the given sequence number.  Messages are
 631  * moved to the retrans queue when rds_send_xmit picks them off the send
 632  * queue. This means that in the TCP case, the message may not have been
 633  * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
 634  * checks the RDS_MSG_HAS_ACK_SEQ bit.
 635  *
 636  * XXX It's not clear to me how this is safely serialized with socket
 637  * destruction.  Maybe it should bail if it sees SOCK_DEAD.
 638  */
 639 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
 640                          is_acked_func is_acked)
 641 {
 642         struct rds_message *rm, *tmp;
 643         unsigned long flags;
 644         LIST_HEAD(list);
 645
 646         spin_lock_irqsave(&conn->c_lock, flags);
 647
 648         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
 649                 if (!rds_send_is_acked(rm, ack, is_acked))
 650                         break;
 651
 652                 list_move(&rm->m_conn_item, &list);
 653                 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
 654         }
 655
 656         /* order flag updates with spin locks */
 657         if (!list_empty(&list))
 658                 smp_mb__after_clear_bit();
 659
 660         spin_unlock_irqrestore(&conn->c_lock, flags);
 661
 662         /* now remove the messages from the sock list as needed */
 663         rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
 664 }
 665 EXPORT_SYMBOL_GPL(rds_send_drop_acked);
 666
 667 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
 668 {
 669         struct rds_message *rm, *tmp;
 670         struct rds_connection *conn;
 671         unsigned long flags;
 672         LIST_HEAD(list);
 673
 674         /* get all the messages we're dropping under the rs lock */
 675         spin_lock_irqsave(&rs->rs_lock, flags);
 676
 677         list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) {
 678                 if (dest && (dest->sin_addr.s_addr != rm->m_daddr ||
 679                              dest->sin_port != rm->m_inc.i_hdr.h_dport))
 680                         continue;
 681
 682                 list_move(&rm->m_sock_item, &list);
 683                 rds_send_sndbuf_remove(rs, rm);
 684                 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
 685         }
 686
 687         /* order flag updates with the rs lock */
 688         smp_mb__after_clear_bit();
 689
 690         spin_unlock_irqrestore(&rs->rs_lock, flags);
 691
 692         if (list_empty(&list))
 693                 return;
 694
 695         /* Remove the messages from the conn */
 696         list_for_each_entry(rm, &list, m_sock_item) {
 697
 698                 conn = rm->m_inc.i_conn;
 699
 700                 spin_lock_irqsave(&conn->c_lock, flags);
 701                 /*
 702                  * Maybe someone else beat us to removing rm from the conn.
 703                  * If we race with their flag update we'll get the lock and
 704                  * then really see that the flag has been cleared.
 705                  */
 706                 if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
 707                         spin_unlock_irqrestore(&conn->c_lock, flags);
 708                         continue;
 709                 }
 710                 list_del_init(&rm->m_conn_item);
 711                 spin_unlock_irqrestore(&conn->c_lock, flags);
 712
 713                 /*
 714                  * Couldn't grab m_rs_lock in top loop (lock ordering),
 715                  * but we can now.
 716                  */
 717                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 718
 719                 spin_lock(&rs->rs_lock);
 720                 __rds_rdma_send_complete(rs, rm, RDS_RDMA_CANCELED);
 721                 spin_unlock(&rs->rs_lock);
 722
 723                 rm->m_rs = NULL;
 724                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 725
 726                 rds_message_put(rm);
 727         }
 728
 729         rds_wake_sk_sleep(rs);
 730
 731         while (!list_empty(&list)) {
 732                 rm = list_entry(list.next, struct rds_message, m_sock_item);
 733                 list_del_init(&rm->m_sock_item);
 734
 735                 rds_message_wait(rm);
 736                 rds_message_put(rm);
 737         }
 738 }
 739
 740 /*
 741  * we only want this to fire once so we use the callers 'queued'.  It's
 742  * possible that another thread can race with us and remove the
 743  * message from the flow with RDS_CANCEL_SENT_TO.
 744  */
 745 static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
 746                              struct rds_message *rm, __be16 sport,
 747                              __be16 dport, int *queued)
 748 {
 749         unsigned long flags;
 750         u32 len;
 751
 752         if (*queued)
 753                 goto out;
 754
 755         len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
 756
 757         /* this is the only place which holds both the socket's rs_lock
 758          * and the connection's c_lock */
 759         spin_lock_irqsave(&rs->rs_lock, flags);
 760
 761         /*
 762          * If there is a little space in sndbuf, we don't queue anything,
 763          * and userspace gets -EAGAIN. But poll() indicates there's send
 764          * room. This can lead to bad behavior (spinning) if snd_bytes isn't
 765          * freed up by incoming acks. So we check the *old* value of
 766          * rs_snd_bytes here to allow the last msg to exceed the buffer,
 767          * and poll() now knows no more data can be sent.
 768          */
 769         if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
 770                 rs->rs_snd_bytes += len;
 771
 772                 /* let recv side know we are close to send space exhaustion.
 773                  * This is probably not the optimal way to do it, as this
 774                  * means we set the flag on *all* messages as soon as our
 775                  * throughput hits a certain threshold.
 776                  */
 777                 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
 778                         __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
 779
 780                 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
 781                 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
 782                 rds_message_addref(rm);
 783                 rm->m_rs = rs;
 784
 785                 /* The code ordering is a little weird, but we're
 786                    trying to minimize the time we hold c_lock */
 787                 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
 788                 rm->m_inc.i_conn = conn;
 789                 rds_message_addref(rm);
 790
 791                 spin_lock(&conn->c_lock);
 792                 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(conn->c_next_tx_seq++);
 793                 list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
 794                 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
 795                 spin_unlock(&conn->c_lock);
 796
 797                 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
 798                          rm, len, rs, rs->rs_snd_bytes,
 799                          (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
 800
 801                 *queued = 1;
 802         }
 803
 804         spin_unlock_irqrestore(&rs->rs_lock, flags);
 805 out:
 806         return *queued;
 807 }
 808
 809 /*
 810  * rds_message is getting to be quite complicated, and we'd like to allocate
 811  * it all in one go. This figures out how big it needs to be up front.
 812  */
 813 static int rds_rm_size(struct msghdr *msg, int data_len)
 814 {
 815         struct cmsghdr *cmsg;
 816         int size = 0;
 817         int retval;
 818
 819         for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
 820                 if (!CMSG_OK(msg, cmsg))
 821                         return -EINVAL;
 822
 823                 if (cmsg->cmsg_level != SOL_RDS)
 824                         continue;
 825
 826                 switch (cmsg->cmsg_type) {
 827                 case RDS_CMSG_RDMA_ARGS:
 828                         retval = rds_rdma_extra_size(CMSG_DATA(cmsg));
 829                         if (retval < 0)
 830                                 return retval;
 831                         size += retval;
 832                         break;
 833
 834                 case RDS_CMSG_RDMA_DEST:
 835                 case RDS_CMSG_RDMA_MAP:
 836                         /* these are valid but do no add any size */
 837                         break;
 838
 839                 case RDS_CMSG_ATOMIC_CSWP:
 840                 case RDS_CMSG_ATOMIC_FADD:
 841                         size += sizeof(struct scatterlist);
 842                         break;
 843
 844                 default:
 845                         return -EINVAL;
 846                 }
 847
 848         }
 849
 850         size += ceil(data_len, PAGE_SIZE) * sizeof(struct scatterlist);
 851
 852         return size;
 853 }
 854
 855 static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
 856                          struct msghdr *msg, int *allocated_mr)
 857 {
 858         struct cmsghdr *cmsg;
 859         int ret = 0;
 860
 861         for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
 862                 if (!CMSG_OK(msg, cmsg))
 863                         return -EINVAL;
 864
 865                 if (cmsg->cmsg_level != SOL_RDS)
 866                         continue;
 867
 868                 /* As a side effect, RDMA_DEST and RDMA_MAP will set
 869                  * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
 870                  */
 871                 switch (cmsg->cmsg_type) {
 872                 case RDS_CMSG_RDMA_ARGS:
 873                         ret = rds_cmsg_rdma_args(rs, rm, cmsg);
 874                         break;
 875
 876                 case RDS_CMSG_RDMA_DEST:
 877                         ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
 878                         break;
 879
 880                 case RDS_CMSG_RDMA_MAP:
 881                         ret = rds_cmsg_rdma_map(rs, rm, cmsg);
 882                         if (!ret)
 883                                 *allocated_mr = 1;
 884                         break;
 885                 case RDS_CMSG_ATOMIC_CSWP:
 886                 case RDS_CMSG_ATOMIC_FADD:
 887                         ret = rds_cmsg_atomic(rs, rm, cmsg);
 888                         break;
 889
 890                 default:
 891                         return -EINVAL;
 892                 }
 893
 894                 if (ret)
 895                         break;
 896         }
 897
 898         return ret;
 899 }
 900
 901 int rds_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
 902                 size_t payload_len)
 903 {
 904         struct sock *sk = sock->sk;
 905         struct rds_sock *rs = rds_sk_to_rs(sk);
 906         struct sockaddr_in *usin = (struct sockaddr_in *)msg->msg_name;
 907         __be32 daddr;
 908         __be16 dport;
 909         struct rds_message *rm = NULL;
 910         struct rds_connection *conn;
 911         int ret = 0;
 912         int queued = 0, allocated_mr = 0;
 913         int nonblock = msg->msg_flags & MSG_DONTWAIT;
 914         long timeo = sock_sndtimeo(sk, nonblock);
 915
 916         /* Mirror Linux UDP mirror of BSD error message compatibility */
 917         /* XXX: Perhaps MSG_MORE someday */
 918         if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT)) {
 919                 printk(KERN_INFO "msg_flags 0x%08X\n", msg->msg_flags);
 920                 ret = -EOPNOTSUPP;
 921                 goto out;
 922         }
 923
 924         if (msg->msg_namelen) {
 925                 /* XXX fail non-unicast destination IPs? */
 926                 if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) {
 927                         ret = -EINVAL;
 928                         goto out;
 929                 }
 930                 daddr = usin->sin_addr.s_addr;
 931                 dport = usin->sin_port;
 932         } else {
 933                 /* We only care about consistency with ->connect() */
 934                 lock_sock(sk);
 935                 daddr = rs->rs_conn_addr;
 936                 dport = rs->rs_conn_port;
 937                 release_sock(sk);
 938         }
 939
 940         /* racing with another thread binding seems ok here */
 941         if (daddr == 0 || rs->rs_bound_addr == 0) {
 942                 ret = -ENOTCONN; /* XXX not a great errno */
 943                 goto out;
 944         }
 945
 946         /* size of rm including all sgs */
 947         ret = rds_rm_size(msg, payload_len);
 948         if (ret < 0)
 949                 goto out;
 950
 951         rm = rds_message_alloc(ret, GFP_KERNEL);
 952         if (!rm) {
 953                 ret = -ENOMEM;
 954                 goto out;
 955         }
 956
 957         rm->data.m_sg = rds_message_alloc_sgs(rm, ceil(payload_len, PAGE_SIZE));
 958         /* XXX fix this to not allocate memory */
 959         ret = rds_message_copy_from_user(rm, msg->msg_iov, payload_len);
 960         if (ret)
 961                 goto out;
 962
 963         rm->m_daddr = daddr;
 964
 965         /* rds_conn_create has a spinlock that runs with IRQ off.
 966          * Caching the conn in the socket helps a lot. */
 967         if (rs->rs_conn && rs->rs_conn->c_faddr == daddr)
 968                 conn = rs->rs_conn;
 969         else {
 970                 conn = rds_conn_create_outgoing(rs->rs_bound_addr, daddr,
 971                                         rs->rs_transport,
 972                                         sock->sk->sk_allocation);
 973                 if (IS_ERR(conn)) {
 974                         ret = PTR_ERR(conn);
 975                         goto out;
 976                 }
 977                 rs->rs_conn = conn;
 978         }
 979
 980         /* Parse any control messages the user may have included. */
 981         ret = rds_cmsg_send(rs, rm, msg, &allocated_mr);
 982         if (ret)
 983                 goto out;
 984
 985         if ((rm->m_rdma_cookie || rm->rdma.m_rdma_op.r_active) &&
 986                !conn->c_trans->xmit_rdma) {
 987                 if (printk_ratelimit())
 988                         printk(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
 989                                &rm->rdma.m_rdma_op, conn->c_trans->xmit_rdma);
 990                 ret = -EOPNOTSUPP;
 991                 goto out;
 992         }
 993
 994         if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
 995                 if (printk_ratelimit())
 996                         printk(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
 997                                &rm->atomic, conn->c_trans->xmit_atomic);
 998                 ret = -EOPNOTSUPP;
 999                 goto out;
1000         }
1001
1002         /* If the connection is down, trigger a connect. We may
1003          * have scheduled a delayed reconnect however - in this case
1004          * we should not interfere.
1005          */
1006         if (rds_conn_state(conn) == RDS_CONN_DOWN &&
1007             !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
1008                 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
1009
1010         ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
1011         if (ret) {
1012                 rs->rs_seen_congestion = 1;
1013                 goto out;
1014         }
1015
1016         while (!rds_send_queue_rm(rs, conn, rm, rs->rs_bound_port,
1017                                   dport, &queued)) {
1018                 rds_stats_inc(s_send_queue_full);
1019                 /* XXX make sure this is reasonable */
1020                 if (payload_len > rds_sk_sndbuf(rs)) {
1021                         ret = -EMSGSIZE;
1022                         goto out;
1023                 }
1024                 if (nonblock) {
1025                         ret = -EAGAIN;
1026                         goto out;
1027                 }
1028
1029                 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
1030                                         rds_send_queue_rm(rs, conn, rm,
1031                                                           rs->rs_bound_port,
1032                                                           dport,
1033                                                           &queued),
1034                                         timeo);
1035                 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
1036                 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
1037                         continue;
1038
1039                 ret = timeo;
1040                 if (ret == 0)
1041                         ret = -ETIMEDOUT;
1042                 goto out;
1043         }
1044
1045         /*
1046          * By now we've committed to the send.  We reuse rds_send_worker()
1047          * to retry sends in the rds thread if the transport asks us to.
1048          */
1049         rds_stats_inc(s_send_queued);
1050
1051         if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
1052                 rds_send_worker(&conn->c_send_w.work);
1053
1054         rds_message_put(rm);
1055         return payload_len;
1056
1057 out:
1058         /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1059          * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1060          * or in any other way, we need to destroy the MR again */
1061         if (allocated_mr)
1062                 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
1063
1064         if (rm)
1065                 rds_message_put(rm);
1066         return ret;
1067 }
1068
1069 /*
1070  * Reply to a ping packet.
1071  */
1072 int
1073 rds_send_pong(struct rds_connection *conn, __be16 dport)
1074 {
1075         struct rds_message *rm;
1076         unsigned long flags;
1077         int ret = 0;
1078
1079         rm = rds_message_alloc(0, GFP_ATOMIC);
1080         if (!rm) {
1081                 ret = -ENOMEM;
1082                 goto out;
1083         }
1084
1085         rm->m_daddr = conn->c_faddr;
1086
1087         /* If the connection is down, trigger a connect. We may
1088          * have scheduled a delayed reconnect however - in this case
1089          * we should not interfere.
1090          */
1091         if (rds_conn_state(conn) == RDS_CONN_DOWN &&
1092             !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
1093                 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
1094
1095         ret = rds_cong_wait(conn->c_fcong, dport, 1, NULL);
1096         if (ret)
1097                 goto out;
1098
1099         spin_lock_irqsave(&conn->c_lock, flags);
1100         list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
1101         set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
1102         rds_message_addref(rm);
1103         rm->m_inc.i_conn = conn;
1104
1105         rds_message_populate_header(&rm->m_inc.i_hdr, 0, dport,
1106                                     conn->c_next_tx_seq);
1107         conn->c_next_tx_seq++;
1108         spin_unlock_irqrestore(&conn->c_lock, flags);
1109
1110         rds_stats_inc(s_send_queued);
1111         rds_stats_inc(s_send_pong);
1112
1113         queue_delayed_work(rds_wq, &conn->c_send_w, 0);
1114         rds_message_put(rm);
1115         return 0;
1116
1117 out:
1118         if (rm)
1119                 rds_message_put(rm);
1120         return ret;
1121 }