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         spin_lock_irqsave(&conn->c_send_lock, flags);
  64         if (conn->c_xmit_rm) {
  65                 rm = conn->c_xmit_rm;
  66                 conn->c_xmit_rm = NULL;
  67                 /* Tell the user the RDMA op is no longer mapped by the
  68                  * transport. This isn't entirely true (it's flushed out
  69                  * independently) but as the connection is down, there's
  70                  * no ongoing RDMA to/from that memory */
  71                 rds_message_unmapped(rm);
  72                 spin_unlock_irqrestore(&conn->c_send_lock, flags);
  73
  74                 rds_message_put(rm);
  75         } else {
  76                 spin_unlock_irqrestore(&conn->c_send_lock, flags);
  77         }
  78
  79         conn->c_xmit_sg = 0;
  80         conn->c_xmit_hdr_off = 0;
  81         conn->c_xmit_data_off = 0;
  82         conn->c_xmit_atomic_sent = 0;
  83         conn->c_xmit_rdma_sent = 0;
  84         conn->c_xmit_data_sent = 0;
  85
  86         conn->c_map_queued = 0;
  87
  88         conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
  89         conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
  90
  91         /* Mark messages as retransmissions, and move them to the send q */
  92         spin_lock_irqsave(&conn->c_lock, flags);
  93         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
  94                 set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
  95                 set_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags);
  96         }
  97         list_splice_init(&conn->c_retrans, &conn->c_send_queue);
  98         spin_unlock_irqrestore(&conn->c_lock, flags);
  99 }
 100
 101 /*
 102  * We're making the concious trade-off here to only send one message
 103  * down the connection at a time.
 104  *   Pro:
 105  *      - tx queueing is a simple fifo list
 106  *      - reassembly is optional and easily done by transports per conn
 107  *      - no per flow rx lookup at all, straight to the socket
 108  *      - less per-frag memory and wire overhead
 109  *   Con:
 110  *      - queued acks can be delayed behind large messages
 111  *   Depends:
 112  *      - small message latency is higher behind queued large messages
 113  *      - large message latency isn't starved by intervening small sends
 114  */
 115 int rds_send_xmit(struct rds_connection *conn)
 116 {
 117         struct rds_message *rm;
 118         unsigned long flags;
 119         unsigned int tmp;
 120         struct scatterlist *sg;
 121         int ret = 0;
 122         int gen = 0;
 123         LIST_HEAD(to_be_dropped);
 124
 125 restart:
 126         if (!rds_conn_up(conn))
 127                 goto out;
 128
 129         /*
 130          * sendmsg calls here after having queued its message on the send
 131          * queue.  We only have one task feeding the connection at a time.  If
 132          * another thread is already feeding the queue then we back off.  This
 133          * avoids blocking the caller and trading per-connection data between
 134          * caches per message.
 135          */
 136         if (!spin_trylock_irqsave(&conn->c_send_lock, flags)) {
 137                 rds_stats_inc(s_send_lock_contention);
 138                 ret = -ENOMEM;
 139                 goto out;
 140         }
 141         atomic_inc(&conn->c_senders);
 142
 143         if (conn->c_trans->xmit_prepare)
 144                 conn->c_trans->xmit_prepare(conn);
 145
 146         gen = atomic_inc_return(&conn->c_send_generation);
 147
 148         /*
 149          * spin trying to push headers and data down the connection until
 150          * the connection doesn't make forward progress.
 151          */
 152         while (1) {
 153
 154                 rm = conn->c_xmit_rm;
 155
 156                 /*
 157                  * If between sending messages, we can send a pending congestion
 158                  * map update.
 159                  */
 160                 if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) {
 161                         rm = rds_cong_update_alloc(conn);
 162                         if (IS_ERR(rm)) {
 163                                 ret = PTR_ERR(rm);
 164                                 break;
 165                         }
 166                         rm->data.op_active = 1;
 167
 168                         conn->c_xmit_rm = rm;
 169                 }
 170
 171                 /*
 172                  * If not already working on one, grab the next message.
 173                  *
 174                  * c_xmit_rm holds a ref while we're sending this message down
 175                  * the connction.  We can use this ref while holding the
 176                  * send_sem.. rds_send_reset() is serialized with it.
 177                  */
 178                 if (!rm) {
 179                         unsigned int len;
 180
 181                         spin_lock(&conn->c_lock);
 182
 183                         if (!list_empty(&conn->c_send_queue)) {
 184                                 rm = list_entry(conn->c_send_queue.next,
 185                                                 struct rds_message,
 186                                                 m_conn_item);
 187                                 rds_message_addref(rm);
 188
 189                                 /*
 190                                  * Move the message from the send queue to the retransmit
 191                                  * list right away.
 192                                  */
 193                                 list_move_tail(&rm->m_conn_item, &conn->c_retrans);
 194                         }
 195
 196                         spin_unlock(&conn->c_lock);
 197
 198                         if (!rm)
 199                                 break;
 200
 201                         /* Unfortunately, the way Infiniband deals with
 202                          * RDMA to a bad MR key is by moving the entire
 203                          * queue pair to error state. We cold possibly
 204                          * recover from that, but right now we drop the
 205                          * connection.
 206                          * Therefore, we never retransmit messages with RDMA ops.
 207                          */
 208                         if (rm->rdma.op_active &&
 209                             test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags)) {
 210                                 spin_lock(&conn->c_lock);
 211                                 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags))
 212                                         list_move(&rm->m_conn_item, &to_be_dropped);
 213                                 spin_unlock(&conn->c_lock);
 214                                 continue;
 215                         }
 216
 217                         /* Require an ACK every once in a while */
 218                         len = ntohl(rm->m_inc.i_hdr.h_len);
 219                         if (conn->c_unacked_packets == 0 ||
 220                             conn->c_unacked_bytes < len) {
 221                                 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
 222
 223                                 conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
 224                                 conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
 225                                 rds_stats_inc(s_send_ack_required);
 226                         } else {
 227                                 conn->c_unacked_bytes -= len;
 228                                 conn->c_unacked_packets--;
 229                         }
 230
 231                         conn->c_xmit_rm = rm;
 232                 }
 233
 234                 /* The transport either sends the whole rdma or none of it */
 235                 if (rm->rdma.op_active && !conn->c_xmit_rdma_sent) {
 236                         rm->m_final_op = &rm->rdma;
 237                         ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
 238                         if (ret)
 239                                 break;
 240                         conn->c_xmit_rdma_sent = 1;
 241
 242                         /* The transport owns the mapped memory for now.
 243                          * You can't unmap it while it's on the send queue */
 244                         set_bit(RDS_MSG_MAPPED, &rm->m_flags);
 245                 }
 246
 247                 if (rm->atomic.op_active && !conn->c_xmit_atomic_sent) {
 248                         rm->m_final_op = &rm->atomic;
 249                         ret = conn->c_trans->xmit_atomic(conn, &rm->atomic);
 250                         if (ret)
 251                                 break;
 252                         conn->c_xmit_atomic_sent = 1;
 253
 254                         /* The transport owns the mapped memory for now.
 255                          * You can't unmap it while it's on the send queue */
 256                         set_bit(RDS_MSG_MAPPED, &rm->m_flags);
 257                 }
 258
 259                 /*
 260                  * A number of cases require an RDS header to be sent
 261                  * even if there is no data.
 262                  * We permit 0-byte sends; rds-ping depends on this.
 263                  * However, if there are exclusively attached silent ops,
 264                  * we skip the hdr/data send, to enable silent operation.
 265                  */
 266                 if (rm->data.op_nents == 0) {
 267                         int ops_present;
 268                         int all_ops_are_silent = 1;
 269
 270                         ops_present = (rm->atomic.op_active || rm->rdma.op_active);
 271                         if (rm->atomic.op_active && !rm->atomic.op_silent)
 272                                 all_ops_are_silent = 0;
 273                         if (rm->rdma.op_active && !rm->rdma.op_silent)
 274                                 all_ops_are_silent = 0;
 275
 276                         if (ops_present && all_ops_are_silent
 277                             && !rm->m_rdma_cookie)
 278                                 rm->data.op_active = 0;
 279                 }
 280
 281                 if (rm->data.op_active && !conn->c_xmit_data_sent) {
 282                         rm->m_final_op = &rm->data;
 283                         ret = conn->c_trans->xmit(conn, rm,
 284                                                   conn->c_xmit_hdr_off,
 285                                                   conn->c_xmit_sg,
 286                                                   conn->c_xmit_data_off);
 287                         if (ret <= 0)
 288                                 break;
 289
 290                         if (conn->c_xmit_hdr_off < sizeof(struct rds_header)) {
 291                                 tmp = min_t(int, ret,
 292                                             sizeof(struct rds_header) -
 293                                             conn->c_xmit_hdr_off);
 294                                 conn->c_xmit_hdr_off += tmp;
 295                                 ret -= tmp;
 296                         }
 297
 298                         sg = &rm->data.op_sg[conn->c_xmit_sg];
 299                         while (ret) {
 300                                 tmp = min_t(int, ret, sg->length -
 301                                                       conn->c_xmit_data_off);
 302                                 conn->c_xmit_data_off += tmp;
 303                                 ret -= tmp;
 304                                 if (conn->c_xmit_data_off == sg->length) {
 305                                         conn->c_xmit_data_off = 0;
 306                                         sg++;
 307                                         conn->c_xmit_sg++;
 308                                         BUG_ON(ret != 0 &&
 309                                                conn->c_xmit_sg == rm->data.op_nents);
 310                                 }
 311                         }
 312
 313                         if (conn->c_xmit_hdr_off == sizeof(struct rds_header) &&
 314                             (conn->c_xmit_sg == rm->data.op_nents))
 315                                 conn->c_xmit_data_sent = 1;
 316                 }
 317
 318                 /*
 319                  * A rm will only take multiple times through this loop
 320                  * if there is a data op. Thus, if the data is sent (or there was
 321                  * none), then we're done with the rm.
 322                  */
 323                 if (!rm->data.op_active || conn->c_xmit_data_sent) {
 324                         conn->c_xmit_rm = NULL;
 325                         conn->c_xmit_sg = 0;
 326                         conn->c_xmit_hdr_off = 0;
 327                         conn->c_xmit_data_off = 0;
 328                         conn->c_xmit_rdma_sent = 0;
 329                         conn->c_xmit_atomic_sent = 0;
 330                         conn->c_xmit_data_sent = 0;
 331
 332                         rds_message_put(rm);
 333                 }
 334         }
 335
 336         if (conn->c_trans->xmit_complete)
 337                 conn->c_trans->xmit_complete(conn);
 338
 339         /*
 340          * We might be racing with another sender who queued a message but
 341          * backed off on noticing that we held the c_send_lock.  If we check
 342          * for queued messages after dropping the sem then either we'll
 343          * see the queued message or the queuer will get the sem.  If we
 344          * notice the queued message then we trigger an immediate retry.
 345          *
 346          * We need to be careful only to do this when we stopped processing
 347          * the send queue because it was empty.  It's the only way we
 348          * stop processing the loop when the transport hasn't taken
 349          * responsibility for forward progress.
 350          */
 351         spin_unlock_irqrestore(&conn->c_send_lock, flags);
 352
 353         /* Nuke any messages we decided not to retransmit. */
 354         if (!list_empty(&to_be_dropped)) {
 355                 /* irqs on here, so we can put(), unlike above */
 356                 list_for_each_entry(rm, &to_be_dropped, m_conn_item)
 357                         rds_message_put(rm);
 358                 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
 359         }
 360
 361         atomic_dec(&conn->c_senders);
 362
 363         /*
 364          * Other senders will see we have c_send_lock and exit. We
 365          * need to recheck the send queue and race again for c_send_lock
 366          * to make sure messages don't just sit on the send queue, if
 367          * somebody hasn't already beat us into the loop.
 368          *
 369          * If the transport cannot continue (i.e ret != 0), then it must
 370          * call us when more room is available, such as from the tx
 371          * completion handler.
 372          */
 373         if (ret == 0) {
 374                 smp_mb();
 375                 if (!list_empty(&conn->c_send_queue)) {
 376                         rds_stats_inc(s_send_lock_queue_raced);
 377                         if (gen == atomic_read(&conn->c_send_generation)) {
 378                                 goto restart;
 379                         }
 380                 }
 381         }
 382 out:
 383         return ret;
 384 }
 385
 386 static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
 387 {
 388         u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
 389
 390         assert_spin_locked(&rs->rs_lock);
 391
 392         BUG_ON(rs->rs_snd_bytes < len);
 393         rs->rs_snd_bytes -= len;
 394
 395         if (rs->rs_snd_bytes == 0)
 396                 rds_stats_inc(s_send_queue_empty);
 397 }
 398
 399 static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
 400                                     is_acked_func is_acked)
 401 {
 402         if (is_acked)
 403                 return is_acked(rm, ack);
 404         return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
 405 }
 406
 407 /*
 408  * This is pretty similar to what happens below in the ACK
 409  * handling code - except that we call here as soon as we get
 410  * the IB send completion on the RDMA op and the accompanying
 411  * message.
 412  */
 413 void rds_rdma_send_complete(struct rds_message *rm, int status)
 414 {
 415         struct rds_sock *rs = NULL;
 416         struct rm_rdma_op *ro;
 417         struct rds_notifier *notifier;
 418         unsigned long flags;
 419
 420         spin_lock_irqsave(&rm->m_rs_lock, flags);
 421
 422         ro = &rm->rdma;
 423         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
 424             ro->op_active && ro->op_notify && ro->op_notifier) {
 425                 notifier = ro->op_notifier;
 426                 rs = rm->m_rs;
 427                 sock_hold(rds_rs_to_sk(rs));
 428
 429                 notifier->n_status = status;
 430                 spin_lock(&rs->rs_lock);
 431                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
 432                 spin_unlock(&rs->rs_lock);
 433
 434                 ro->op_notifier = NULL;
 435         }
 436
 437         spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 438
 439         if (rs) {
 440                 rds_wake_sk_sleep(rs);
 441                 sock_put(rds_rs_to_sk(rs));
 442         }
 443 }
 444 EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
 445
 446 /*
 447  * Just like above, except looks at atomic op
 448  */
 449 void rds_atomic_send_complete(struct rds_message *rm, int status)
 450 {
 451         struct rds_sock *rs = NULL;
 452         struct rm_atomic_op *ao;
 453         struct rds_notifier *notifier;
 454         unsigned long flags;
 455
 456         spin_lock_irqsave(&rm->m_rs_lock, flags);
 457
 458         ao = &rm->atomic;
 459         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
 460             && ao->op_active && ao->op_notify && ao->op_notifier) {
 461                 notifier = ao->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                 ao->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_atomic_send_complete);
 481
 482 /*
 483  * This is the same as rds_rdma_send_complete except we
 484  * don't do any locking - we have all the ingredients (message,
 485  * socket, socket lock) and can just move the notifier.
 486  */
 487 static inline void
 488 __rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
 489 {
 490         struct rm_rdma_op *ro;
 491         struct rm_atomic_op *ao;
 492
 493         ro = &rm->rdma;
 494         if (ro->op_active && ro->op_notify && ro->op_notifier) {
 495                 ro->op_notifier->n_status = status;
 496                 list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue);
 497                 ro->op_notifier = NULL;
 498         }
 499
 500         ao = &rm->atomic;
 501         if (ao->op_active && ao->op_notify && ao->op_notifier) {
 502                 ao->op_notifier->n_status = status;
 503                 list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue);
 504                 ao->op_notifier = NULL;
 505         }
 506
 507         /* No need to wake the app - caller does this */
 508 }
 509
 510 /*
 511  * This is called from the IB send completion when we detect
 512  * a RDMA operation that failed with remote access error.
 513  * So speed is not an issue here.
 514  */
 515 struct rds_message *rds_send_get_message(struct rds_connection *conn,
 516                                          struct rm_rdma_op *op)
 517 {
 518         struct rds_message *rm, *tmp, *found = NULL;
 519         unsigned long flags;
 520
 521         spin_lock_irqsave(&conn->c_lock, flags);
 522
 523         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
 524                 if (&rm->rdma == op) {
 525                         atomic_inc(&rm->m_refcount);
 526                         found = rm;
 527                         goto out;
 528                 }
 529         }
 530
 531         list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
 532                 if (&rm->rdma == op) {
 533                         atomic_inc(&rm->m_refcount);
 534                         found = rm;
 535                         break;
 536                 }
 537         }
 538
 539 out:
 540         spin_unlock_irqrestore(&conn->c_lock, flags);
 541
 542         return found;
 543 }
 544 EXPORT_SYMBOL_GPL(rds_send_get_message);
 545
 546 /*
 547  * This removes messages from the socket's list if they're on it.  The list
 548  * argument must be private to the caller, we must be able to modify it
 549  * without locks.  The messages must have a reference held for their
 550  * position on the list.  This function will drop that reference after
 551  * removing the messages from the 'messages' list regardless of if it found
 552  * the messages on the socket list or not.
 553  */
 554 void rds_send_remove_from_sock(struct list_head *messages, int status)
 555 {
 556         unsigned long flags;
 557         struct rds_sock *rs = NULL;
 558         struct rds_message *rm;
 559
 560         while (!list_empty(messages)) {
 561                 int was_on_sock = 0;
 562
 563                 rm = list_entry(messages->next, struct rds_message,
 564                                 m_conn_item);
 565                 list_del_init(&rm->m_conn_item);
 566
 567                 /*
 568                  * If we see this flag cleared then we're *sure* that someone
 569                  * else beat us to removing it from the sock.  If we race
 570                  * with their flag update we'll get the lock and then really
 571                  * see that the flag has been cleared.
 572                  *
 573                  * The message spinlock makes sure nobody clears rm->m_rs
 574                  * while we're messing with it. It does not prevent the
 575                  * message from being removed from the socket, though.
 576                  */
 577                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 578                 if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags))
 579                         goto unlock_and_drop;
 580
 581                 if (rs != rm->m_rs) {
 582                         if (rs) {
 583                                 rds_wake_sk_sleep(rs);
 584                                 sock_put(rds_rs_to_sk(rs));
 585                         }
 586                         rs = rm->m_rs;
 587                         sock_hold(rds_rs_to_sk(rs));
 588                 }
 589                 spin_lock(&rs->rs_lock);
 590
 591                 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
 592                         struct rm_rdma_op *ro = &rm->rdma;
 593                         struct rds_notifier *notifier;
 594
 595                         list_del_init(&rm->m_sock_item);
 596                         rds_send_sndbuf_remove(rs, rm);
 597
 598                         if (ro->op_active && ro->op_notifier &&
 599                                (ro->op_notify || (ro->op_recverr && status))) {
 600                                 notifier = ro->op_notifier;
 601                                 list_add_tail(&notifier->n_list,
 602                                                 &rs->rs_notify_queue);
 603                                 if (!notifier->n_status)
 604                                         notifier->n_status = status;
 605                                 rm->rdma.op_notifier = NULL;
 606                         }
 607                         was_on_sock = 1;
 608                         rm->m_rs = NULL;
 609                 }
 610                 spin_unlock(&rs->rs_lock);
 611
 612 unlock_and_drop:
 613                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 614                 rds_message_put(rm);
 615                 if (was_on_sock)
 616                         rds_message_put(rm);
 617         }
 618
 619         if (rs) {
 620                 rds_wake_sk_sleep(rs);
 621                 sock_put(rds_rs_to_sk(rs));
 622         }
 623 }
 624
 625 /*
 626  * Transports call here when they've determined that the receiver queued
 627  * messages up to, and including, the given sequence number.  Messages are
 628  * moved to the retrans queue when rds_send_xmit picks them off the send
 629  * queue. This means that in the TCP case, the message may not have been
 630  * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
 631  * checks the RDS_MSG_HAS_ACK_SEQ bit.
 632  *
 633  * XXX It's not clear to me how this is safely serialized with socket
 634  * destruction.  Maybe it should bail if it sees SOCK_DEAD.
 635  */
 636 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
 637                          is_acked_func is_acked)
 638 {
 639         struct rds_message *rm, *tmp;
 640         unsigned long flags;
 641         LIST_HEAD(list);
 642
 643         spin_lock_irqsave(&conn->c_lock, flags);
 644
 645         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
 646                 if (!rds_send_is_acked(rm, ack, is_acked))
 647                         break;
 648
 649                 list_move(&rm->m_conn_item, &list);
 650                 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
 651         }
 652
 653         /* order flag updates with spin locks */
 654         if (!list_empty(&list))
 655                 smp_mb__after_clear_bit();
 656
 657         spin_unlock_irqrestore(&conn->c_lock, flags);
 658
 659         /* now remove the messages from the sock list as needed */
 660         rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
 661 }
 662 EXPORT_SYMBOL_GPL(rds_send_drop_acked);
 663
 664 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
 665 {
 666         struct rds_message *rm, *tmp;
 667         struct rds_connection *conn;
 668         unsigned long flags;
 669         LIST_HEAD(list);
 670
 671         /* get all the messages we're dropping under the rs lock */
 672         spin_lock_irqsave(&rs->rs_lock, flags);
 673
 674         list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) {
 675                 if (dest && (dest->sin_addr.s_addr != rm->m_daddr ||
 676                              dest->sin_port != rm->m_inc.i_hdr.h_dport))
 677                         continue;
 678
 679                 list_move(&rm->m_sock_item, &list);
 680                 rds_send_sndbuf_remove(rs, rm);
 681                 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
 682         }
 683
 684         /* order flag updates with the rs lock */
 685         smp_mb__after_clear_bit();
 686
 687         spin_unlock_irqrestore(&rs->rs_lock, flags);
 688
 689         if (list_empty(&list))
 690                 return;
 691
 692         /* Remove the messages from the conn */
 693         list_for_each_entry(rm, &list, m_sock_item) {
 694
 695                 conn = rm->m_inc.i_conn;
 696
 697                 spin_lock_irqsave(&conn->c_lock, flags);
 698                 /*
 699                  * Maybe someone else beat us to removing rm from the conn.
 700                  * If we race with their flag update we'll get the lock and
 701                  * then really see that the flag has been cleared.
 702                  */
 703                 if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
 704                         spin_unlock_irqrestore(&conn->c_lock, flags);
 705                         continue;
 706                 }
 707                 list_del_init(&rm->m_conn_item);
 708                 spin_unlock_irqrestore(&conn->c_lock, flags);
 709
 710                 /*
 711                  * Couldn't grab m_rs_lock in top loop (lock ordering),
 712                  * but we can now.
 713                  */
 714                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 715
 716                 spin_lock(&rs->rs_lock);
 717                 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
 718                 spin_unlock(&rs->rs_lock);
 719
 720                 rm->m_rs = NULL;
 721                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 722
 723                 rds_message_put(rm);
 724         }
 725
 726         rds_wake_sk_sleep(rs);
 727
 728         while (!list_empty(&list)) {
 729                 rm = list_entry(list.next, struct rds_message, m_sock_item);
 730                 list_del_init(&rm->m_sock_item);
 731
 732                 rds_message_wait(rm);
 733                 rds_message_put(rm);
 734         }
 735 }
 736
 737 /*
 738  * we only want this to fire once so we use the callers 'queued'.  It's
 739  * possible that another thread can race with us and remove the
 740  * message from the flow with RDS_CANCEL_SENT_TO.
 741  */
 742 static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
 743                              struct rds_message *rm, __be16 sport,
 744                              __be16 dport, int *queued)
 745 {
 746         unsigned long flags;
 747         u32 len;
 748
 749         if (*queued)
 750                 goto out;
 751
 752         len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
 753
 754         /* this is the only place which holds both the socket's rs_lock
 755          * and the connection's c_lock */
 756         spin_lock_irqsave(&rs->rs_lock, flags);
 757
 758         /*
 759          * If there is a little space in sndbuf, we don't queue anything,
 760          * and userspace gets -EAGAIN. But poll() indicates there's send
 761          * room. This can lead to bad behavior (spinning) if snd_bytes isn't
 762          * freed up by incoming acks. So we check the *old* value of
 763          * rs_snd_bytes here to allow the last msg to exceed the buffer,
 764          * and poll() now knows no more data can be sent.
 765          */
 766         if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
 767                 rs->rs_snd_bytes += len;
 768
 769                 /* let recv side know we are close to send space exhaustion.
 770                  * This is probably not the optimal way to do it, as this
 771                  * means we set the flag on *all* messages as soon as our
 772                  * throughput hits a certain threshold.
 773                  */
 774                 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
 775                         __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
 776
 777                 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
 778                 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
 779                 rds_message_addref(rm);
 780                 rm->m_rs = rs;
 781
 782                 /* The code ordering is a little weird, but we're
 783                    trying to minimize the time we hold c_lock */
 784                 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
 785                 rm->m_inc.i_conn = conn;
 786                 rds_message_addref(rm);
 787
 788                 spin_lock(&conn->c_lock);
 789                 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(conn->c_next_tx_seq++);
 790                 list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
 791                 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
 792                 spin_unlock(&conn->c_lock);
 793
 794                 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
 795                          rm, len, rs, rs->rs_snd_bytes,
 796                          (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
 797
 798                 *queued = 1;
 799         }
 800
 801         spin_unlock_irqrestore(&rs->rs_lock, flags);
 802 out:
 803         return *queued;
 804 }
 805
 806 /*
 807  * rds_message is getting to be quite complicated, and we'd like to allocate
 808  * it all in one go. This figures out how big it needs to be up front.
 809  */
 810 static int rds_rm_size(struct msghdr *msg, int data_len)
 811 {
 812         struct cmsghdr *cmsg;
 813         int size = 0;
 814         int cmsg_groups = 0;
 815         int retval;
 816
 817         for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
 818                 if (!CMSG_OK(msg, cmsg))
 819                         return -EINVAL;
 820
 821                 if (cmsg->cmsg_level != SOL_RDS)
 822                         continue;
 823
 824                 switch (cmsg->cmsg_type) {
 825                 case RDS_CMSG_RDMA_ARGS:
 826                         cmsg_groups |= 1;
 827                         retval = rds_rdma_extra_size(CMSG_DATA(cmsg));
 828                         if (retval < 0)
 829                                 return retval;
 830                         size += retval;
 831
 832                         break;
 833
 834                 case RDS_CMSG_RDMA_DEST:
 835                 case RDS_CMSG_RDMA_MAP:
 836                         cmsg_groups |= 2;
 837                         /* these are valid but do no add any size */
 838                         break;
 839
 840                 case RDS_CMSG_ATOMIC_CSWP:
 841                 case RDS_CMSG_ATOMIC_FADD:
 842                         cmsg_groups |= 1;
 843                         size += sizeof(struct scatterlist);
 844                         break;
 845
 846                 default:
 847                         return -EINVAL;
 848                 }
 849
 850         }
 851
 852         size += ceil(data_len, PAGE_SIZE) * sizeof(struct scatterlist);
 853
 854         /* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */
 855         if (cmsg_groups == 3)
 856                 return -EINVAL;
 857
 858         return size;
 859 }
 860
 861 static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
 862                          struct msghdr *msg, int *allocated_mr)
 863 {
 864         struct cmsghdr *cmsg;
 865         int ret = 0;
 866
 867         for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
 868                 if (!CMSG_OK(msg, cmsg))
 869                         return -EINVAL;
 870
 871                 if (cmsg->cmsg_level != SOL_RDS)
 872                         continue;
 873
 874                 /* As a side effect, RDMA_DEST and RDMA_MAP will set
 875                  * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
 876                  */
 877                 switch (cmsg->cmsg_type) {
 878                 case RDS_CMSG_RDMA_ARGS:
 879                         ret = rds_cmsg_rdma_args(rs, rm, cmsg);
 880                         break;
 881
 882                 case RDS_CMSG_RDMA_DEST:
 883                         ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
 884                         break;
 885
 886                 case RDS_CMSG_RDMA_MAP:
 887                         ret = rds_cmsg_rdma_map(rs, rm, cmsg);
 888                         if (!ret)
 889                                 *allocated_mr = 1;
 890                         break;
 891                 case RDS_CMSG_ATOMIC_CSWP:
 892                 case RDS_CMSG_ATOMIC_FADD:
 893                         ret = rds_cmsg_atomic(rs, rm, cmsg);
 894                         break;
 895
 896                 default:
 897                         return -EINVAL;
 898                 }
 899
 900                 if (ret)
 901                         break;
 902         }
 903
 904         return ret;
 905 }
 906
 907 int rds_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
 908                 size_t payload_len)
 909 {
 910         struct sock *sk = sock->sk;
 911         struct rds_sock *rs = rds_sk_to_rs(sk);
 912         struct sockaddr_in *usin = (struct sockaddr_in *)msg->msg_name;
 913         __be32 daddr;
 914         __be16 dport;
 915         struct rds_message *rm = NULL;
 916         struct rds_connection *conn;
 917         int ret = 0;
 918         int queued = 0, allocated_mr = 0;
 919         int nonblock = msg->msg_flags & MSG_DONTWAIT;
 920         long timeo = sock_sndtimeo(sk, nonblock);
 921
 922         /* Mirror Linux UDP mirror of BSD error message compatibility */
 923         /* XXX: Perhaps MSG_MORE someday */
 924         if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT)) {
 925                 printk(KERN_INFO "msg_flags 0x%08X\n", msg->msg_flags);
 926                 ret = -EOPNOTSUPP;
 927                 goto out;
 928         }
 929
 930         if (msg->msg_namelen) {
 931                 /* XXX fail non-unicast destination IPs? */
 932                 if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) {
 933                         ret = -EINVAL;
 934                         goto out;
 935                 }
 936                 daddr = usin->sin_addr.s_addr;
 937                 dport = usin->sin_port;
 938         } else {
 939                 /* We only care about consistency with ->connect() */
 940                 lock_sock(sk);
 941                 daddr = rs->rs_conn_addr;
 942                 dport = rs->rs_conn_port;
 943                 release_sock(sk);
 944         }
 945
 946         /* racing with another thread binding seems ok here */
 947         if (daddr == 0 || rs->rs_bound_addr == 0) {
 948                 ret = -ENOTCONN; /* XXX not a great errno */
 949                 goto out;
 950         }
 951
 952         /* size of rm including all sgs */
 953         ret = rds_rm_size(msg, payload_len);
 954         if (ret < 0)
 955                 goto out;
 956
 957         rm = rds_message_alloc(ret, GFP_KERNEL);
 958         if (!rm) {
 959                 ret = -ENOMEM;
 960                 goto out;
 961         }
 962
 963         /* Attach data to the rm */
 964         if (payload_len) {
 965                 rm->data.op_sg = rds_message_alloc_sgs(rm, ceil(payload_len, PAGE_SIZE));
 966                 ret = rds_message_copy_from_user(rm, msg->msg_iov, payload_len);
 967                 if (ret)
 968                         goto out;
 969         }
 970         rm->data.op_active = 1;
 971
 972         rm->m_daddr = daddr;
 973
 974         /* rds_conn_create has a spinlock that runs with IRQ off.
 975          * Caching the conn in the socket helps a lot. */
 976         if (rs->rs_conn && rs->rs_conn->c_faddr == daddr)
 977                 conn = rs->rs_conn;
 978         else {
 979                 conn = rds_conn_create_outgoing(rs->rs_bound_addr, daddr,
 980                                         rs->rs_transport,
 981                                         sock->sk->sk_allocation);
 982                 if (IS_ERR(conn)) {
 983                         ret = PTR_ERR(conn);
 984                         goto out;
 985                 }
 986                 rs->rs_conn = conn;
 987         }
 988
 989         /* Parse any control messages the user may have included. */
 990         ret = rds_cmsg_send(rs, rm, msg, &allocated_mr);
 991         if (ret)
 992                 goto out;
 993
 994         if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) {
 995                 if (printk_ratelimit())
 996                         printk(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
 997                                &rm->rdma, conn->c_trans->xmit_rdma);
 998                 ret = -EOPNOTSUPP;
 999                 goto out;
1000         }
1001
1002         if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
1003                 if (printk_ratelimit())
1004                         printk(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
1005                                &rm->atomic, conn->c_trans->xmit_atomic);
1006                 ret = -EOPNOTSUPP;
1007                 goto out;
1008         }
1009
1010         rds_conn_connect_if_down(conn);
1011
1012         ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
1013         if (ret) {
1014                 rs->rs_seen_congestion = 1;
1015                 goto out;
1016         }
1017
1018         while (!rds_send_queue_rm(rs, conn, rm, rs->rs_bound_port,
1019                                   dport, &queued)) {
1020                 rds_stats_inc(s_send_queue_full);
1021                 /* XXX make sure this is reasonable */
1022                 if (payload_len > rds_sk_sndbuf(rs)) {
1023                         ret = -EMSGSIZE;
1024                         goto out;
1025                 }
1026                 if (nonblock) {
1027                         ret = -EAGAIN;
1028                         goto out;
1029                 }
1030
1031                 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
1032                                         rds_send_queue_rm(rs, conn, rm,
1033                                                           rs->rs_bound_port,
1034                                                           dport,
1035                                                           &queued),
1036                                         timeo);
1037                 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
1038                 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
1039                         continue;
1040
1041                 ret = timeo;
1042                 if (ret == 0)
1043                         ret = -ETIMEDOUT;
1044                 goto out;
1045         }
1046
1047         /*
1048          * By now we've committed to the send.  We reuse rds_send_worker()
1049          * to retry sends in the rds thread if the transport asks us to.
1050          */
1051         rds_stats_inc(s_send_queued);
1052
1053         if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
1054                 rds_send_xmit(conn);
1055
1056         rds_message_put(rm);
1057         return payload_len;
1058
1059 out:
1060         /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1061          * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1062          * or in any other way, we need to destroy the MR again */
1063         if (allocated_mr)
1064                 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
1065
1066         if (rm)
1067                 rds_message_put(rm);
1068         return ret;
1069 }
1070
1071 /*
1072  * Reply to a ping packet.
1073  */
1074 int
1075 rds_send_pong(struct rds_connection *conn, __be16 dport)
1076 {
1077         struct rds_message *rm;
1078         unsigned long flags;
1079         int ret = 0;
1080
1081         rm = rds_message_alloc(0, GFP_ATOMIC);
1082         if (!rm) {
1083                 ret = -ENOMEM;
1084                 goto out;
1085         }
1086
1087         rm->m_daddr = conn->c_faddr;
1088         rm->data.op_active = 1;
1089
1090         rds_conn_connect_if_down(conn);
1091
1092         ret = rds_cong_wait(conn->c_fcong, dport, 1, NULL);
1093         if (ret)
1094                 goto out;
1095
1096         spin_lock_irqsave(&conn->c_lock, flags);
1097         list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
1098         set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
1099         rds_message_addref(rm);
1100         rm->m_inc.i_conn = conn;
1101
1102         rds_message_populate_header(&rm->m_inc.i_hdr, 0, dport,
1103                                     conn->c_next_tx_seq);
1104         conn->c_next_tx_seq++;
1105         spin_unlock_irqrestore(&conn->c_lock, flags);
1106
1107         rds_stats_inc(s_send_queued);
1108         rds_stats_inc(s_send_pong);
1109
1110         if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
1111                 rds_send_xmit(conn);
1112
1113         rds_message_put(rm);
1114         return 0;
1115
1116 out:
1117         if (rm)
1118                 rds_message_put(rm);
1119         return ret;
1120 }