net/rds/send.c

   1 /*
   2  * Copyright (c) 2006 Oracle.  All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  *
  32  */
  33 #include <linux/kernel.h>
  34 #include <linux/gfp.h>
  35 #include <net/sock.h>
  36 #include <linux/in.h>
  37 #include <linux/list.h>
  38
  39 #include "rds.h"
  40
  41 /* When transmitting messages in rds_send_xmit, we need to emerge from
  42  * time to time and briefly release the CPU. Otherwise the softlock watchdog
  43  * will kick our shin.
  44  * Also, it seems fairer to not let one busy connection stall all the
  45  * others.
  46  *
  47  * send_batch_count is the number of times we'll loop in send_xmit. Setting
  48  * it to 0 will restore the old behavior (where we looped until we had
  49  * drained the queue).
  50  */
  51 static int send_batch_count = 64;
  52 module_param(send_batch_count, int, 0444);
  53 MODULE_PARM_DESC(send_batch_count, " batch factor when working the send queue");
  54
  55 /*
  56  * Reset the send state. Caller must hold c_send_lock when calling here.
  57  */
  58 void rds_send_reset(struct rds_connection *conn)
  59 {
  60         struct rds_message *rm, *tmp;
  61         unsigned long flags;
  62
  63         if (conn->c_xmit_rm) {
  64                 /* Tell the user the RDMA op is no longer mapped by the
  65                  * transport. This isn't entirely true (it's flushed out
  66                  * independently) but as the connection is down, there's
  67                  * no ongoing RDMA to/from that memory */
  68                 rds_message_unmapped(conn->c_xmit_rm);
  69                 rds_message_put(conn->c_xmit_rm);
  70                 conn->c_xmit_rm = NULL;
  71         }
  72         conn->c_xmit_sg = 0;
  73         conn->c_xmit_hdr_off = 0;
  74         conn->c_xmit_data_off = 0;
  75         conn->c_xmit_atomic_sent = 0;
  76         conn->c_xmit_rdma_sent = 0;
  77         conn->c_xmit_data_sent = 0;
  78
  79         conn->c_map_queued = 0;
  80
  81         conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
  82         conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
  83
  84         /* Mark messages as retransmissions, and move them to the send q */
  85         spin_lock_irqsave(&conn->c_lock, flags);
  86         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
  87                 set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
  88                 set_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags);
  89         }
  90         list_splice_init(&conn->c_retrans, &conn->c_send_queue);
  91         spin_unlock_irqrestore(&conn->c_lock, flags);
  92 }
  93
  94 /*
  95  * We're making the concious trade-off here to only send one message
  96  * down the connection at a time.
  97  *   Pro:
  98  *      - tx queueing is a simple fifo list
  99  *      - reassembly is optional and easily done by transports per conn
 100  *      - no per flow rx lookup at all, straight to the socket
 101  *      - less per-frag memory and wire overhead
 102  *   Con:
 103  *      - queued acks can be delayed behind large messages
 104  *   Depends:
 105  *      - small message latency is higher behind queued large messages
 106  *      - large message latency isn't starved by intervening small sends
 107  */
 108 int rds_send_xmit(struct rds_connection *conn)
 109 {
 110         struct rds_message *rm;
 111         unsigned long flags;
 112         unsigned int tmp;
 113         unsigned int send_quota = send_batch_count;
 114         struct scatterlist *sg;
 115         int ret = 0;
 116         int was_empty = 0;
 117         LIST_HEAD(to_be_dropped);
 118
 119         /*
 120          * sendmsg calls here after having queued its message on the send
 121          * queue.  We only have one task feeding the connection at a time.  If
 122          * another thread is already feeding the queue then we back off.  This
 123          * avoids blocking the caller and trading per-connection data between
 124          * caches per message.
 125          *
 126          * The sem holder will issue a retry if they notice that someone queued
 127          * a message after they stopped walking the send queue but before they
 128          * dropped the sem.
 129          */
 130         if (!mutex_trylock(&conn->c_send_lock)) {
 131                 rds_stats_inc(s_send_sem_contention);
 132                 ret = -ENOMEM;
 133                 goto out;
 134         }
 135
 136         if (conn->c_trans->xmit_prepare)
 137                 conn->c_trans->xmit_prepare(conn);
 138
 139         /*
 140          * spin trying to push headers and data down the connection until
 141          * the connection doesn't make forward progress.
 142          */
 143         while (--send_quota) {
 144
 145                 rm = conn->c_xmit_rm;
 146
 147                 /*
 148                  * If between sending messages, we can send a pending congestion
 149                  * map update.
 150                  *
 151                  * Transports either define a special xmit_cong_map function,
 152                  * or we allocate a cong_map message and treat it just like any
 153                  * other send.
 154                  */
 155                 if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) {
 156                         if (conn->c_trans->xmit_cong_map) {
 157                                 unsigned long map_offset = 0;
 158                                 unsigned long map_bytes = sizeof(struct rds_header) +
 159                                         RDS_CONG_MAP_BYTES;
 160
 161                                 while (map_bytes) {
 162                                         ret = conn->c_trans->xmit_cong_map(conn, conn->c_lcong,
 163                                                                            map_offset);
 164                                         if (ret <= 0) {
 165                                                 /* too far down the rabbithole! */
 166                                                 mutex_unlock(&conn->c_send_lock);
 167                                                 rds_conn_error(conn, "Cong map xmit failed\n");
 168                                                 goto out;
 169                                         }
 170
 171                                         map_offset += ret;
 172                                         map_bytes -= ret;
 173                                 }
 174                         } else {
 175                                 /* send cong update like a normal rm */
 176                                 rm = rds_cong_update_alloc(conn);
 177                                 if (IS_ERR(rm)) {
 178                                         ret = PTR_ERR(rm);
 179                                         break;
 180                                 }
 181                                 rm->data.op_active = 1;
 182
 183                                 conn->c_xmit_rm = rm;
 184                         }
 185                 }
 186
 187                 /*
 188                  * If not already working on one, grab the next message.
 189                  *
 190                  * c_xmit_rm holds a ref while we're sending this message down
 191                  * the connction.  We can use this ref while holding the
 192                  * send_sem.. rds_send_reset() is serialized with it.
 193                  */
 194                 if (!rm) {
 195                         unsigned int len;
 196
 197                         spin_lock_irqsave(&conn->c_lock, flags);
 198
 199                         if (!list_empty(&conn->c_send_queue)) {
 200                                 rm = list_entry(conn->c_send_queue.next,
 201                                                 struct rds_message,
 202                                                 m_conn_item);
 203                                 rds_message_addref(rm);
 204
 205                                 /*
 206                                  * Move the message from the send queue to the retransmit
 207                                  * list right away.
 208                                  */
 209                                 list_move_tail(&rm->m_conn_item, &conn->c_retrans);
 210                         }
 211
 212                         spin_unlock_irqrestore(&conn->c_lock, flags);
 213
 214                         if (!rm) {
 215                                 was_empty = 1;
 216                                 break;
 217                         }
 218
 219                         /* Unfortunately, the way Infiniband deals with
 220                          * RDMA to a bad MR key is by moving the entire
 221                          * queue pair to error state. We cold possibly
 222                          * recover from that, but right now we drop the
 223                          * connection.
 224                          * Therefore, we never retransmit messages with RDMA ops.
 225                          */
 226                         if (rm->rdma.op_active &&
 227                             test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags)) {
 228                                 spin_lock_irqsave(&conn->c_lock, flags);
 229                                 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags))
 230                                         list_move(&rm->m_conn_item, &to_be_dropped);
 231                                 spin_unlock_irqrestore(&conn->c_lock, flags);
 232                                 rds_message_put(rm);
 233                                 continue;
 234                         }
 235
 236                         /* Require an ACK every once in a while */
 237                         len = ntohl(rm->m_inc.i_hdr.h_len);
 238                         if (conn->c_unacked_packets == 0 ||
 239                             conn->c_unacked_bytes < len) {
 240                                 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
 241
 242                                 conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
 243                                 conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
 244                                 rds_stats_inc(s_send_ack_required);
 245                         } else {
 246                                 conn->c_unacked_bytes -= len;
 247                                 conn->c_unacked_packets--;
 248                         }
 249
 250                         conn->c_xmit_rm = rm;
 251                 }
 252
 253                 if (rm->atomic.op_active && !conn->c_xmit_atomic_sent) {
 254                         ret = conn->c_trans->xmit_atomic(conn, rm);
 255                         if (ret)
 256                                 break;
 257                         conn->c_xmit_atomic_sent = 1;
 258                         /* The transport owns the mapped memory for now.
 259                          * You can't unmap it while it's on the send queue */
 260                         set_bit(RDS_MSG_MAPPED, &rm->m_flags);
 261
 262                         /*
 263                          * This is evil, muahaha.
 264                          * We permit 0-byte sends. (rds-ping depends on this.)
 265                          * BUT if there is an atomic op and no sent data,
 266                          * we turn off sending the header, to achieve
 267                          * "silent" atomics.
 268                          * But see below; RDMA op might toggle this back on!
 269                          */
 270                         if (rm->data.op_nents == 0)
 271                                 rm->data.op_active = 0;
 272                 }
 273
 274                 /* The transport either sends the whole rdma or none of it */
 275                 if (rm->rdma.op_active && !conn->c_xmit_rdma_sent) {
 276                         ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
 277                         if (ret)
 278                                 break;
 279                         conn->c_xmit_rdma_sent = 1;
 280
 281                         /* rdmas need data sent, even if just the header */
 282                         rm->data.op_active = 1;
 283
 284                         /* The transport owns the mapped memory for now.
 285                          * You can't unmap it while it's on the send queue */
 286                         set_bit(RDS_MSG_MAPPED, &rm->m_flags);
 287                 }
 288
 289                 if (rm->data.op_active && !conn->c_xmit_data_sent) {
 290                         ret = conn->c_trans->xmit(conn, rm,
 291                                                   conn->c_xmit_hdr_off,
 292                                                   conn->c_xmit_sg,
 293                                                   conn->c_xmit_data_off);
 294                         if (ret <= 0)
 295                                 break;
 296
 297                         if (conn->c_xmit_hdr_off < sizeof(struct rds_header)) {
 298                                 tmp = min_t(int, ret,
 299                                             sizeof(struct rds_header) -
 300                                             conn->c_xmit_hdr_off);
 301                                 conn->c_xmit_hdr_off += tmp;
 302                                 ret -= tmp;
 303                         }
 304
 305                         sg = &rm->data.op_sg[conn->c_xmit_sg];
 306                         while (ret) {
 307                                 tmp = min_t(int, ret, sg->length -
 308                                                       conn->c_xmit_data_off);
 309                                 conn->c_xmit_data_off += tmp;
 310                                 ret -= tmp;
 311                                 if (conn->c_xmit_data_off == sg->length) {
 312                                         conn->c_xmit_data_off = 0;
 313                                         sg++;
 314                                         conn->c_xmit_sg++;
 315                                         BUG_ON(ret != 0 &&
 316                                                conn->c_xmit_sg == rm->data.op_nents);
 317                                 }
 318                         }
 319
 320                         if (conn->c_xmit_hdr_off == sizeof(struct rds_header) &&
 321                             (conn->c_xmit_sg == rm->data.op_nents))
 322                                 conn->c_xmit_data_sent = 1;
 323                 }
 324
 325                 /*
 326                  * A rm will only take multiple times through this loop
 327                  * if there is a data op. Thus, if the data is sent (or there was
 328                  * none), then we're done with the rm.
 329                  */
 330                 if (!rm->data.op_active || conn->c_xmit_data_sent) {
 331                         conn->c_xmit_rm = NULL;
 332                         conn->c_xmit_sg = 0;
 333                         conn->c_xmit_hdr_off = 0;
 334                         conn->c_xmit_data_off = 0;
 335                         conn->c_xmit_rdma_sent = 0;
 336                         conn->c_xmit_atomic_sent = 0;
 337                         conn->c_xmit_data_sent = 0;
 338
 339                         rds_message_put(rm);
 340                 }
 341         }
 342
 343         /* Nuke any messages we decided not to retransmit. */
 344         if (!list_empty(&to_be_dropped))
 345                 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
 346
 347         if (conn->c_trans->xmit_complete)
 348                 conn->c_trans->xmit_complete(conn);
 349
 350         /*
 351          * We might be racing with another sender who queued a message but
 352          * backed off on noticing that we held the c_send_lock.  If we check
 353          * for queued messages after dropping the sem then either we'll
 354          * see the queued message or the queuer will get the sem.  If we
 355          * notice the queued message then we trigger an immediate retry.
 356          *
 357          * We need to be careful only to do this when we stopped processing
 358          * the send queue because it was empty.  It's the only way we
 359          * stop processing the loop when the transport hasn't taken
 360          * responsibility for forward progress.
 361          */
 362         mutex_unlock(&conn->c_send_lock);
 363
 364         if (send_quota == 0 && !was_empty) {
 365                 /* We exhausted the send quota, but there's work left to
 366                  * do. Return and (re-)schedule the send worker.
 367                  */
 368                 ret = -EAGAIN;
 369         }
 370
 371         if (ret == 0 && was_empty) {
 372                 /* A simple bit test would be way faster than taking the
 373                  * spin lock */
 374                 spin_lock_irqsave(&conn->c_lock, flags);
 375                 if (!list_empty(&conn->c_send_queue)) {
 376                         rds_stats_inc(s_send_sem_queue_raced);
 377                         ret = -EAGAIN;
 378                 }
 379                 spin_unlock_irqrestore(&conn->c_lock, flags);
 380         }
 381 out:
 382         return ret;
 383 }
 384
 385 static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
 386 {
 387         u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
 388
 389         assert_spin_locked(&rs->rs_lock);
 390
 391         BUG_ON(rs->rs_snd_bytes < len);
 392         rs->rs_snd_bytes -= len;
 393
 394         if (rs->rs_snd_bytes == 0)
 395                 rds_stats_inc(s_send_queue_empty);
 396 }
 397
 398 static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
 399                                     is_acked_func is_acked)
 400 {
 401         if (is_acked)
 402                 return is_acked(rm, ack);
 403         return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
 404 }
 405
 406 /*
 407  * Returns true if there are no messages on the send and retransmit queues
 408  * which have a sequence number greater than or equal to the given sequence
 409  * number.
 410  */
 411 int rds_send_acked_before(struct rds_connection *conn, u64 seq)
 412 {
 413         struct rds_message *rm, *tmp;
 414         int ret = 1;
 415
 416         spin_lock(&conn->c_lock);
 417
 418         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
 419                 if (be64_to_cpu(rm->m_inc.i_hdr.h_sequence) < seq)
 420                         ret = 0;
 421                 break;
 422         }
 423
 424         list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
 425                 if (be64_to_cpu(rm->m_inc.i_hdr.h_sequence) < seq)
 426                         ret = 0;
 427                 break;
 428         }
 429
 430         spin_unlock(&conn->c_lock);
 431
 432         return ret;
 433 }
 434
 435 /*
 436  * This is pretty similar to what happens below in the ACK
 437  * handling code - except that we call here as soon as we get
 438  * the IB send completion on the RDMA op and the accompanying
 439  * message.
 440  */
 441 void rds_rdma_send_complete(struct rds_message *rm, int status)
 442 {
 443         struct rds_sock *rs = NULL;
 444         struct rm_rdma_op *ro;
 445         struct rds_notifier *notifier;
 446         unsigned long flags;
 447
 448         spin_lock_irqsave(&rm->m_rs_lock, flags);
 449
 450         ro = &rm->rdma;
 451         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
 452             ro->op_active && ro->op_notify && ro->op_notifier) {
 453                 notifier = ro->op_notifier;
 454                 rs = rm->m_rs;
 455                 sock_hold(rds_rs_to_sk(rs));
 456
 457                 notifier->n_status = status;
 458                 spin_lock(&rs->rs_lock);
 459                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
 460                 spin_unlock(&rs->rs_lock);
 461
 462                 ro->op_notifier = NULL;
 463         }
 464
 465         spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 466
 467         if (rs) {
 468                 rds_wake_sk_sleep(rs);
 469                 sock_put(rds_rs_to_sk(rs));
 470         }
 471 }
 472 EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
 473
 474 /*
 475  * Just like above, except looks at atomic op
 476  */
 477 void rds_atomic_send_complete(struct rds_message *rm, int status)
 478 {
 479         struct rds_sock *rs = NULL;
 480         struct rm_atomic_op *ao;
 481         struct rds_notifier *notifier;
 482
 483         spin_lock(&rm->m_rs_lock);
 484
 485         ao = &rm->atomic;
 486         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
 487             && ao->op_active && ao->op_notify && ao->op_notifier) {
 488                 notifier = ao->op_notifier;
 489                 rs = rm->m_rs;
 490                 sock_hold(rds_rs_to_sk(rs));
 491
 492                 notifier->n_status = status;
 493                 spin_lock(&rs->rs_lock);
 494                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
 495                 spin_unlock(&rs->rs_lock);
 496
 497                 ao->op_notifier = NULL;
 498         }
 499
 500         spin_unlock(&rm->m_rs_lock);
 501
 502         if (rs) {
 503                 rds_wake_sk_sleep(rs);
 504                 sock_put(rds_rs_to_sk(rs));
 505         }
 506 }
 507 EXPORT_SYMBOL_GPL(rds_atomic_send_complete);
 508
 509 /*
 510  * This is the same as rds_rdma_send_complete except we
 511  * don't do any locking - we have all the ingredients (message,
 512  * socket, socket lock) and can just move the notifier.
 513  */
 514 static inline void
 515 __rds_rdma_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
 516 {
 517         struct rm_rdma_op *ro;
 518
 519         ro = &rm->rdma;
 520         if (ro->op_active && ro->op_notify && ro->op_notifier) {
 521                 ro->op_notifier->n_status = status;
 522                 list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue);
 523                 ro->op_notifier = NULL;
 524         }
 525
 526         /* No need to wake the app - caller does this */
 527 }
 528
 529 /*
 530  * This is called from the IB send completion when we detect
 531  * a RDMA operation that failed with remote access error.
 532  * So speed is not an issue here.
 533  */
 534 struct rds_message *rds_send_get_message(struct rds_connection *conn,
 535                                          struct rm_rdma_op *op)
 536 {
 537         struct rds_message *rm, *tmp, *found = NULL;
 538         unsigned long flags;
 539
 540         spin_lock_irqsave(&conn->c_lock, flags);
 541
 542         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
 543                 if (&rm->rdma == op) {
 544                         atomic_inc(&rm->m_refcount);
 545                         found = rm;
 546                         goto out;
 547                 }
 548         }
 549
 550         list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
 551                 if (&rm->rdma == op) {
 552                         atomic_inc(&rm->m_refcount);
 553                         found = rm;
 554                         break;
 555                 }
 556         }
 557
 558 out:
 559         spin_unlock_irqrestore(&conn->c_lock, flags);
 560
 561         return found;
 562 }
 563 EXPORT_SYMBOL_GPL(rds_send_get_message);
 564
 565 /*
 566  * This removes messages from the socket's list if they're on it.  The list
 567  * argument must be private to the caller, we must be able to modify it
 568  * without locks.  The messages must have a reference held for their
 569  * position on the list.  This function will drop that reference after
 570  * removing the messages from the 'messages' list regardless of if it found
 571  * the messages on the socket list or not.
 572  */
 573 void rds_send_remove_from_sock(struct list_head *messages, int status)
 574 {
 575         unsigned long flags;
 576         struct rds_sock *rs = NULL;
 577         struct rds_message *rm;
 578
 579         while (!list_empty(messages)) {
 580                 int was_on_sock = 0;
 581
 582                 rm = list_entry(messages->next, struct rds_message,
 583                                 m_conn_item);
 584                 list_del_init(&rm->m_conn_item);
 585
 586                 /*
 587                  * If we see this flag cleared then we're *sure* that someone
 588                  * else beat us to removing it from the sock.  If we race
 589                  * with their flag update we'll get the lock and then really
 590                  * see that the flag has been cleared.
 591                  *
 592                  * The message spinlock makes sure nobody clears rm->m_rs
 593                  * while we're messing with it. It does not prevent the
 594                  * message from being removed from the socket, though.
 595                  */
 596                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 597                 if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags))
 598                         goto unlock_and_drop;
 599
 600                 if (rs != rm->m_rs) {
 601                         if (rs) {
 602                                 rds_wake_sk_sleep(rs);
 603                                 sock_put(rds_rs_to_sk(rs));
 604                         }
 605                         rs = rm->m_rs;
 606                         sock_hold(rds_rs_to_sk(rs));
 607                 }
 608                 spin_lock(&rs->rs_lock);
 609
 610                 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
 611                         struct rm_rdma_op *ro = &rm->rdma;
 612                         struct rds_notifier *notifier;
 613
 614                         list_del_init(&rm->m_sock_item);
 615                         rds_send_sndbuf_remove(rs, rm);
 616
 617                         if (ro->op_active && ro->op_notifier &&
 618                                (ro->op_notify || (ro->op_recverr && status))) {
 619                                 notifier = ro->op_notifier;
 620                                 list_add_tail(&notifier->n_list,
 621                                                 &rs->rs_notify_queue);
 622                                 if (!notifier->n_status)
 623                                         notifier->n_status = status;
 624                                 rm->rdma.op_notifier = NULL;
 625                         }
 626                         was_on_sock = 1;
 627                         rm->m_rs = NULL;
 628                 }
 629                 spin_unlock(&rs->rs_lock);
 630
 631 unlock_and_drop:
 632                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 633                 rds_message_put(rm);
 634                 if (was_on_sock)
 635                         rds_message_put(rm);
 636         }
 637
 638         if (rs) {
 639                 rds_wake_sk_sleep(rs);
 640                 sock_put(rds_rs_to_sk(rs));
 641         }
 642 }
 643
 644 /*
 645  * Transports call here when they've determined that the receiver queued
 646  * messages up to, and including, the given sequence number.  Messages are
 647  * moved to the retrans queue when rds_send_xmit picks them off the send
 648  * queue. This means that in the TCP case, the message may not have been
 649  * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
 650  * checks the RDS_MSG_HAS_ACK_SEQ bit.
 651  *
 652  * XXX It's not clear to me how this is safely serialized with socket
 653  * destruction.  Maybe it should bail if it sees SOCK_DEAD.
 654  */
 655 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
 656                          is_acked_func is_acked)
 657 {
 658         struct rds_message *rm, *tmp;
 659         unsigned long flags;
 660         LIST_HEAD(list);
 661
 662         spin_lock_irqsave(&conn->c_lock, flags);
 663
 664         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
 665                 if (!rds_send_is_acked(rm, ack, is_acked))
 666                         break;
 667
 668                 list_move(&rm->m_conn_item, &list);
 669                 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
 670         }
 671
 672         /* order flag updates with spin locks */
 673         if (!list_empty(&list))
 674                 smp_mb__after_clear_bit();
 675
 676         spin_unlock_irqrestore(&conn->c_lock, flags);
 677
 678         /* now remove the messages from the sock list as needed */
 679         rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
 680 }
 681 EXPORT_SYMBOL_GPL(rds_send_drop_acked);
 682
 683 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
 684 {
 685         struct rds_message *rm, *tmp;
 686         struct rds_connection *conn;
 687         unsigned long flags;
 688         LIST_HEAD(list);
 689
 690         /* get all the messages we're dropping under the rs lock */
 691         spin_lock_irqsave(&rs->rs_lock, flags);
 692
 693         list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) {
 694                 if (dest && (dest->sin_addr.s_addr != rm->m_daddr ||
 695                              dest->sin_port != rm->m_inc.i_hdr.h_dport))
 696                         continue;
 697
 698                 list_move(&rm->m_sock_item, &list);
 699                 rds_send_sndbuf_remove(rs, rm);
 700                 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
 701         }
 702
 703         /* order flag updates with the rs lock */
 704         smp_mb__after_clear_bit();
 705
 706         spin_unlock_irqrestore(&rs->rs_lock, flags);
 707
 708         if (list_empty(&list))
 709                 return;
 710
 711         /* Remove the messages from the conn */
 712         list_for_each_entry(rm, &list, m_sock_item) {
 713
 714                 conn = rm->m_inc.i_conn;
 715
 716                 spin_lock_irqsave(&conn->c_lock, flags);
 717                 /*
 718                  * Maybe someone else beat us to removing rm from the conn.
 719                  * If we race with their flag update we'll get the lock and
 720                  * then really see that the flag has been cleared.
 721                  */
 722                 if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
 723                         spin_unlock_irqrestore(&conn->c_lock, flags);
 724                         continue;
 725                 }
 726                 list_del_init(&rm->m_conn_item);
 727                 spin_unlock_irqrestore(&conn->c_lock, flags);
 728
 729                 /*
 730                  * Couldn't grab m_rs_lock in top loop (lock ordering),
 731                  * but we can now.
 732                  */
 733                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 734
 735                 spin_lock(&rs->rs_lock);
 736                 __rds_rdma_send_complete(rs, rm, RDS_RDMA_CANCELED);
 737                 spin_unlock(&rs->rs_lock);
 738
 739                 rm->m_rs = NULL;
 740                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 741
 742                 rds_message_put(rm);
 743         }
 744
 745         rds_wake_sk_sleep(rs);
 746
 747         while (!list_empty(&list)) {
 748                 rm = list_entry(list.next, struct rds_message, m_sock_item);
 749                 list_del_init(&rm->m_sock_item);
 750
 751                 rds_message_wait(rm);
 752                 rds_message_put(rm);
 753         }
 754 }
 755
 756 /*
 757  * we only want this to fire once so we use the callers 'queued'.  It's
 758  * possible that another thread can race with us and remove the
 759  * message from the flow with RDS_CANCEL_SENT_TO.
 760  */
 761 static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
 762                              struct rds_message *rm, __be16 sport,
 763                              __be16 dport, int *queued)
 764 {
 765         unsigned long flags;
 766         u32 len;
 767
 768         if (*queued)
 769                 goto out;
 770
 771         len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
 772
 773         /* this is the only place which holds both the socket's rs_lock
 774          * and the connection's c_lock */
 775         spin_lock_irqsave(&rs->rs_lock, flags);
 776
 777         /*
 778          * If there is a little space in sndbuf, we don't queue anything,
 779          * and userspace gets -EAGAIN. But poll() indicates there's send
 780          * room. This can lead to bad behavior (spinning) if snd_bytes isn't
 781          * freed up by incoming acks. So we check the *old* value of
 782          * rs_snd_bytes here to allow the last msg to exceed the buffer,
 783          * and poll() now knows no more data can be sent.
 784          */
 785         if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
 786                 rs->rs_snd_bytes += len;
 787
 788                 /* let recv side know we are close to send space exhaustion.
 789                  * This is probably not the optimal way to do it, as this
 790                  * means we set the flag on *all* messages as soon as our
 791                  * throughput hits a certain threshold.
 792                  */
 793                 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
 794                         __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
 795
 796                 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
 797                 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
 798                 rds_message_addref(rm);
 799                 rm->m_rs = rs;
 800
 801                 /* The code ordering is a little weird, but we're
 802                    trying to minimize the time we hold c_lock */
 803                 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
 804                 rm->m_inc.i_conn = conn;
 805                 rds_message_addref(rm);
 806
 807                 spin_lock(&conn->c_lock);
 808                 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(conn->c_next_tx_seq++);
 809                 list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
 810                 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
 811                 spin_unlock(&conn->c_lock);
 812
 813                 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
 814                          rm, len, rs, rs->rs_snd_bytes,
 815                          (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
 816
 817                 *queued = 1;
 818         }
 819
 820         spin_unlock_irqrestore(&rs->rs_lock, flags);
 821 out:
 822         return *queued;
 823 }
 824
 825 /*
 826  * rds_message is getting to be quite complicated, and we'd like to allocate
 827  * it all in one go. This figures out how big it needs to be up front.
 828  */
 829 static int rds_rm_size(struct msghdr *msg, int data_len)
 830 {
 831         struct cmsghdr *cmsg;
 832         int size = 0;
 833         int retval;
 834
 835         for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
 836                 if (!CMSG_OK(msg, cmsg))
 837                         return -EINVAL;
 838
 839                 if (cmsg->cmsg_level != SOL_RDS)
 840                         continue;
 841
 842                 switch (cmsg->cmsg_type) {
 843                 case RDS_CMSG_RDMA_ARGS:
 844                         retval = rds_rdma_extra_size(CMSG_DATA(cmsg));
 845                         if (retval < 0)
 846                                 return retval;
 847                         size += retval;
 848                         break;
 849
 850                 case RDS_CMSG_RDMA_DEST:
 851                 case RDS_CMSG_RDMA_MAP:
 852                         /* these are valid but do no add any size */
 853                         break;
 854
 855                 case RDS_CMSG_ATOMIC_CSWP:
 856                 case RDS_CMSG_ATOMIC_FADD:
 857                         size += sizeof(struct scatterlist);
 858                         break;
 859
 860                 default:
 861                         return -EINVAL;
 862                 }
 863
 864         }
 865
 866         size += ceil(data_len, PAGE_SIZE) * sizeof(struct scatterlist);
 867
 868         return size;
 869 }
 870
 871 static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
 872                          struct msghdr *msg, int *allocated_mr)
 873 {
 874         struct cmsghdr *cmsg;
 875         int ret = 0;
 876
 877         for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
 878                 if (!CMSG_OK(msg, cmsg))
 879                         return -EINVAL;
 880
 881                 if (cmsg->cmsg_level != SOL_RDS)
 882                         continue;
 883
 884                 /* As a side effect, RDMA_DEST and RDMA_MAP will set
 885                  * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
 886                  */
 887                 switch (cmsg->cmsg_type) {
 888                 case RDS_CMSG_RDMA_ARGS:
 889                         ret = rds_cmsg_rdma_args(rs, rm, cmsg);
 890                         break;
 891
 892                 case RDS_CMSG_RDMA_DEST:
 893                         ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
 894                         break;
 895
 896                 case RDS_CMSG_RDMA_MAP:
 897                         ret = rds_cmsg_rdma_map(rs, rm, cmsg);
 898                         if (!ret)
 899                                 *allocated_mr = 1;
 900                         break;
 901                 case RDS_CMSG_ATOMIC_CSWP:
 902                 case RDS_CMSG_ATOMIC_FADD:
 903                         ret = rds_cmsg_atomic(rs, rm, cmsg);
 904                         break;
 905
 906                 default:
 907                         return -EINVAL;
 908                 }
 909
 910                 if (ret)
 911                         break;
 912         }
 913
 914         return ret;
 915 }
 916
 917 int rds_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
 918                 size_t payload_len)
 919 {
 920         struct sock *sk = sock->sk;
 921         struct rds_sock *rs = rds_sk_to_rs(sk);
 922         struct sockaddr_in *usin = (struct sockaddr_in *)msg->msg_name;
 923         __be32 daddr;
 924         __be16 dport;
 925         struct rds_message *rm = NULL;
 926         struct rds_connection *conn;
 927         int ret = 0;
 928         int queued = 0, allocated_mr = 0;
 929         int nonblock = msg->msg_flags & MSG_DONTWAIT;
 930         long timeo = sock_sndtimeo(sk, nonblock);
 931
 932         /* Mirror Linux UDP mirror of BSD error message compatibility */
 933         /* XXX: Perhaps MSG_MORE someday */
 934         if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT)) {
 935                 printk(KERN_INFO "msg_flags 0x%08X\n", msg->msg_flags);
 936                 ret = -EOPNOTSUPP;
 937                 goto out;
 938         }
 939
 940         if (msg->msg_namelen) {
 941                 /* XXX fail non-unicast destination IPs? */
 942                 if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) {
 943                         ret = -EINVAL;
 944                         goto out;
 945                 }
 946                 daddr = usin->sin_addr.s_addr;
 947                 dport = usin->sin_port;
 948         } else {
 949                 /* We only care about consistency with ->connect() */
 950                 lock_sock(sk);
 951                 daddr = rs->rs_conn_addr;
 952                 dport = rs->rs_conn_port;
 953                 release_sock(sk);
 954         }
 955
 956         /* racing with another thread binding seems ok here */
 957         if (daddr == 0 || rs->rs_bound_addr == 0) {
 958                 ret = -ENOTCONN; /* XXX not a great errno */
 959                 goto out;
 960         }
 961
 962         /* size of rm including all sgs */
 963         ret = rds_rm_size(msg, payload_len);
 964         if (ret < 0)
 965                 goto out;
 966
 967         rm = rds_message_alloc(ret, GFP_KERNEL);
 968         if (!rm) {
 969                 ret = -ENOMEM;
 970                 goto out;
 971         }
 972
 973         /* Attach data to the rm */
 974         if (payload_len) {
 975                 rm->data.op_sg = rds_message_alloc_sgs(rm, ceil(payload_len, PAGE_SIZE));
 976                 ret = rds_message_copy_from_user(rm, msg->msg_iov, payload_len);
 977                 if (ret)
 978                         goto out;
 979         }
 980         rm->data.op_active = 1;
 981
 982         rm->m_daddr = daddr;
 983
 984         /* rds_conn_create has a spinlock that runs with IRQ off.
 985          * Caching the conn in the socket helps a lot. */
 986         if (rs->rs_conn && rs->rs_conn->c_faddr == daddr)
 987                 conn = rs->rs_conn;
 988         else {
 989                 conn = rds_conn_create_outgoing(rs->rs_bound_addr, daddr,
 990                                         rs->rs_transport,
 991                                         sock->sk->sk_allocation);
 992                 if (IS_ERR(conn)) {
 993                         ret = PTR_ERR(conn);
 994                         goto out;
 995                 }
 996                 rs->rs_conn = conn;
 997         }
 998
 999         /* Parse any control messages the user may have included. */
1000         ret = rds_cmsg_send(rs, rm, msg, &allocated_mr);
1001         if (ret)
1002                 goto out;
1003
1004         if ((rm->m_rdma_cookie || rm->rdma.op_active) &&
1005             !conn->c_trans->xmit_rdma) {
1006                 if (printk_ratelimit())
1007                         printk(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
1008                                &rm->rdma, conn->c_trans->xmit_rdma);
1009                 ret = -EOPNOTSUPP;
1010                 goto out;
1011         }
1012
1013         if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
1014                 if (printk_ratelimit())
1015                         printk(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
1016                                &rm->atomic, conn->c_trans->xmit_atomic);
1017                 ret = -EOPNOTSUPP;
1018                 goto out;
1019         }
1020
1021         /* If the connection is down, trigger a connect. We may
1022          * have scheduled a delayed reconnect however - in this case
1023          * we should not interfere.
1024          */
1025         if (rds_conn_state(conn) == RDS_CONN_DOWN &&
1026             !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
1027                 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
1028
1029         ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
1030         if (ret) {
1031                 rs->rs_seen_congestion = 1;
1032                 goto out;
1033         }
1034
1035         while (!rds_send_queue_rm(rs, conn, rm, rs->rs_bound_port,
1036                                   dport, &queued)) {
1037                 rds_stats_inc(s_send_queue_full);
1038                 /* XXX make sure this is reasonable */
1039                 if (payload_len > rds_sk_sndbuf(rs)) {
1040                         ret = -EMSGSIZE;
1041                         goto out;
1042                 }
1043                 if (nonblock) {
1044                         ret = -EAGAIN;
1045                         goto out;
1046                 }
1047
1048                 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
1049                                         rds_send_queue_rm(rs, conn, rm,
1050                                                           rs->rs_bound_port,
1051                                                           dport,
1052                                                           &queued),
1053                                         timeo);
1054                 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
1055                 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
1056                         continue;
1057
1058                 ret = timeo;
1059                 if (ret == 0)
1060                         ret = -ETIMEDOUT;
1061                 goto out;
1062         }
1063
1064         /*
1065          * By now we've committed to the send.  We reuse rds_send_worker()
1066          * to retry sends in the rds thread if the transport asks us to.
1067          */
1068         rds_stats_inc(s_send_queued);
1069
1070         if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
1071                 rds_send_worker(&conn->c_send_w.work);
1072
1073         rds_message_put(rm);
1074         return payload_len;
1075
1076 out:
1077         /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1078          * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1079          * or in any other way, we need to destroy the MR again */
1080         if (allocated_mr)
1081                 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
1082
1083         if (rm)
1084                 rds_message_put(rm);
1085         return ret;
1086 }
1087
1088 /*
1089  * Reply to a ping packet.
1090  */
1091 int
1092 rds_send_pong(struct rds_connection *conn, __be16 dport)
1093 {
1094         struct rds_message *rm;
1095         unsigned long flags;
1096         int ret = 0;
1097
1098         rm = rds_message_alloc(0, GFP_ATOMIC);
1099         if (!rm) {
1100                 ret = -ENOMEM;
1101                 goto out;
1102         }
1103
1104         rm->m_daddr = conn->c_faddr;
1105
1106         /* If the connection is down, trigger a connect. We may
1107          * have scheduled a delayed reconnect however - in this case
1108          * we should not interfere.
1109          */
1110         if (rds_conn_state(conn) == RDS_CONN_DOWN &&
1111             !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
1112                 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
1113
1114         ret = rds_cong_wait(conn->c_fcong, dport, 1, NULL);
1115         if (ret)
1116                 goto out;
1117
1118         spin_lock_irqsave(&conn->c_lock, flags);
1119         list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
1120         set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
1121         rds_message_addref(rm);
1122         rm->m_inc.i_conn = conn;
1123
1124         rds_message_populate_header(&rm->m_inc.i_hdr, 0, dport,
1125                                     conn->c_next_tx_seq);
1126         conn->c_next_tx_seq++;
1127         spin_unlock_irqrestore(&conn->c_lock, flags);
1128
1129         rds_stats_inc(s_send_queued);
1130         rds_stats_inc(s_send_pong);
1131
1132         queue_delayed_work(rds_wq, &conn->c_send_w, 0);
1133         rds_message_put(rm);
1134         return 0;
1135
1136 out:
1137         if (rm)
1138                 rds_message_put(rm);
1139         return ret;
1140 }