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