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         struct scatterlist *sg;
 114         int ret = 0;
 115         int gen = 0;
 116         LIST_HEAD(to_be_dropped);
 117
 118 restart:
 119         if (!rds_conn_up(conn))
 120                 goto out;
 121
 122         /*
 123          * sendmsg calls here after having queued its message on the send
 124          * queue.  We only have one task feeding the connection at a time.  If
 125          * another thread is already feeding the queue then we back off.  This
 126          * avoids blocking the caller and trading per-connection data between
 127          * caches per message.
 128          */
 129         if (!spin_trylock_irqsave(&conn->c_send_lock, flags)) {
 130                 rds_stats_inc(s_send_lock_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         gen = atomic_inc_return(&conn->c_send_generation);
 139
 140         /*
 141          * spin trying to push headers and data down the connection until
 142          * the connection doesn't make forward progress.
 143          */
 144         while (1) {
 145
 146                 rm = conn->c_xmit_rm;
 147
 148                 /*
 149                  * If between sending messages, we can send a pending congestion
 150                  * map update.
 151                  */
 152                 if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) {
 153                         rm = rds_cong_update_alloc(conn);
 154                         if (IS_ERR(rm)) {
 155                                 ret = PTR_ERR(rm);
 156                                 break;
 157                         }
 158                         rm->data.op_active = 1;
 159
 160                         conn->c_xmit_rm = rm;
 161                 }
 162
 163                 /*
 164                  * If not already working on one, grab the next message.
 165                  *
 166                  * c_xmit_rm holds a ref while we're sending this message down
 167                  * the connction.  We can use this ref while holding the
 168                  * send_sem.. rds_send_reset() is serialized with it.
 169                  */
 170                 if (!rm) {
 171                         unsigned int len;
 172
 173                         spin_lock(&conn->c_lock);
 174
 175                         if (!list_empty(&conn->c_send_queue)) {
 176                                 rm = list_entry(conn->c_send_queue.next,
 177                                                 struct rds_message,
 178                                                 m_conn_item);
 179                                 rds_message_addref(rm);
 180
 181                                 /*
 182                                  * Move the message from the send queue to the retransmit
 183                                  * list right away.
 184                                  */
 185                                 list_move_tail(&rm->m_conn_item, &conn->c_retrans);
 186                         }
 187
 188                         spin_unlock(&conn->c_lock);
 189
 190                         if (!rm)
 191                                 break;
 192
 193                         /* Unfortunately, the way Infiniband deals with
 194                          * RDMA to a bad MR key is by moving the entire
 195                          * queue pair to error state. We cold possibly
 196                          * recover from that, but right now we drop the
 197                          * connection.
 198                          * Therefore, we never retransmit messages with RDMA ops.
 199                          */
 200                         if (rm->rdma.op_active &&
 201                             test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags)) {
 202                                 spin_lock(&conn->c_lock);
 203                                 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags))
 204                                         list_move(&rm->m_conn_item, &to_be_dropped);
 205                                 spin_unlock(&conn->c_lock);
 206                                 continue;
 207                         }
 208
 209                         /* Require an ACK every once in a while */
 210                         len = ntohl(rm->m_inc.i_hdr.h_len);
 211                         if (conn->c_unacked_packets == 0 ||
 212                             conn->c_unacked_bytes < len) {
 213                                 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
 214
 215                                 conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
 216                                 conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
 217                                 rds_stats_inc(s_send_ack_required);
 218                         } else {
 219                                 conn->c_unacked_bytes -= len;
 220                                 conn->c_unacked_packets--;
 221                         }
 222
 223                         conn->c_xmit_rm = rm;
 224                 }
 225
 226                 /* The transport either sends the whole rdma or none of it */
 227                 if (rm->rdma.op_active && !conn->c_xmit_rdma_sent) {
 228                         rm->m_final_op = &rm->rdma;
 229                         ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
 230                         if (ret)
 231                                 break;
 232                         conn->c_xmit_rdma_sent = 1;
 233
 234                         /* The transport owns the mapped memory for now.
 235                          * You can't unmap it while it's on the send queue */
 236                         set_bit(RDS_MSG_MAPPED, &rm->m_flags);
 237                 }
 238
 239                 if (rm->atomic.op_active && !conn->c_xmit_atomic_sent) {
 240                         rm->m_final_op = &rm->atomic;
 241                         ret = conn->c_trans->xmit_atomic(conn, &rm->atomic);
 242                         if (ret)
 243                                 break;
 244                         conn->c_xmit_atomic_sent = 1;
 245
 246                         /* The transport owns the mapped memory for now.
 247                          * You can't unmap it while it's on the send queue */
 248                         set_bit(RDS_MSG_MAPPED, &rm->m_flags);
 249                 }
 250
 251                 /*
 252                  * A number of cases require an RDS header to be sent
 253                  * even if there is no data.
 254                  * We permit 0-byte sends; rds-ping depends on this.
 255                  * However, if there are exclusively attached silent ops,
 256                  * we skip the hdr/data send, to enable silent operation.
 257                  */
 258                 if (rm->data.op_nents == 0) {
 259                         int ops_present;
 260                         int all_ops_are_silent = 1;
 261
 262                         ops_present = (rm->atomic.op_active || rm->rdma.op_active);
 263                         if (rm->atomic.op_active && !rm->atomic.op_silent)
 264                                 all_ops_are_silent = 0;
 265                         if (rm->rdma.op_active && !rm->rdma.op_silent)
 266                                 all_ops_are_silent = 0;
 267
 268                         if (ops_present && all_ops_are_silent
 269                             && !rm->m_rdma_cookie)
 270                                 rm->data.op_active = 0;
 271                 }
 272
 273                 if (rm->data.op_active && !conn->c_xmit_data_sent) {
 274                         rm->m_final_op = &rm->data;
 275                         ret = conn->c_trans->xmit(conn, rm,
 276                                                   conn->c_xmit_hdr_off,
 277                                                   conn->c_xmit_sg,
 278                                                   conn->c_xmit_data_off);
 279                         if (ret <= 0)
 280                                 break;
 281
 282                         if (conn->c_xmit_hdr_off < sizeof(struct rds_header)) {
 283                                 tmp = min_t(int, ret,
 284                                             sizeof(struct rds_header) -
 285                                             conn->c_xmit_hdr_off);
 286                                 conn->c_xmit_hdr_off += tmp;
 287                                 ret -= tmp;
 288                         }
 289
 290                         sg = &rm->data.op_sg[conn->c_xmit_sg];
 291                         while (ret) {
 292                                 tmp = min_t(int, ret, sg->length -
 293                                                       conn->c_xmit_data_off);
 294                                 conn->c_xmit_data_off += tmp;
 295                                 ret -= tmp;
 296                                 if (conn->c_xmit_data_off == sg->length) {
 297                                         conn->c_xmit_data_off = 0;
 298                                         sg++;
 299                                         conn->c_xmit_sg++;
 300                                         BUG_ON(ret != 0 &&
 301                                                conn->c_xmit_sg == rm->data.op_nents);
 302                                 }
 303                         }
 304
 305                         if (conn->c_xmit_hdr_off == sizeof(struct rds_header) &&
 306                             (conn->c_xmit_sg == rm->data.op_nents))
 307                                 conn->c_xmit_data_sent = 1;
 308                 }
 309
 310                 /*
 311                  * A rm will only take multiple times through this loop
 312                  * if there is a data op. Thus, if the data is sent (or there was
 313                  * none), then we're done with the rm.
 314                  */
 315                 if (!rm->data.op_active || conn->c_xmit_data_sent) {
 316                         conn->c_xmit_rm = NULL;
 317                         conn->c_xmit_sg = 0;
 318                         conn->c_xmit_hdr_off = 0;
 319                         conn->c_xmit_data_off = 0;
 320                         conn->c_xmit_rdma_sent = 0;
 321                         conn->c_xmit_atomic_sent = 0;
 322                         conn->c_xmit_data_sent = 0;
 323
 324                         rds_message_put(rm);
 325                 }
 326         }
 327
 328         if (conn->c_trans->xmit_complete)
 329                 conn->c_trans->xmit_complete(conn);
 330
 331         /*
 332          * We might be racing with another sender who queued a message but
 333          * backed off on noticing that we held the c_send_lock.  If we check
 334          * for queued messages after dropping the sem then either we'll
 335          * see the queued message or the queuer will get the sem.  If we
 336          * notice the queued message then we trigger an immediate retry.
 337          *
 338          * We need to be careful only to do this when we stopped processing
 339          * the send queue because it was empty.  It's the only way we
 340          * stop processing the loop when the transport hasn't taken
 341          * responsibility for forward progress.
 342          */
 343         spin_unlock_irqrestore(&conn->c_send_lock, flags);
 344
 345         /* Nuke any messages we decided not to retransmit. */
 346         if (!list_empty(&to_be_dropped)) {
 347                 /* irqs on here, so we can put(), unlike above */
 348                 list_for_each_entry(rm, &to_be_dropped, m_conn_item)
 349                         rds_message_put(rm);
 350                 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
 351         }
 352
 353         /*
 354          * Other senders will see we have c_send_lock and exit. We
 355          * need to recheck the send queue and race again for c_send_lock
 356          * to make sure messages don't just sit on the send queue.
 357          *
 358          * If the transport cannot continue (i.e ret != 0), then it must
 359          * call us when more room is available, such as from the tx
 360          * completion handler.
 361          */
 362         if (ret == 0) {
 363                 /* A simple bit test would be way faster than taking the
 364                  * spin lock */
 365                 smp_mb();
 366                 if (!list_empty(&conn->c_send_queue)) {
 367                         rds_stats_inc(s_send_lock_queue_raced);
 368                         if (gen == atomic_read(&conn->c_send_generation)) {
 369                                 goto restart;
 370                         }
 371                 }
 372         }
 373 out:
 374         return ret;
 375 }
 376
 377 static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
 378 {
 379         u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
 380
 381         assert_spin_locked(&rs->rs_lock);
 382
 383         BUG_ON(rs->rs_snd_bytes < len);
 384         rs->rs_snd_bytes -= len;
 385
 386         if (rs->rs_snd_bytes == 0)
 387                 rds_stats_inc(s_send_queue_empty);
 388 }
 389
 390 static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
 391                                     is_acked_func is_acked)
 392 {
 393         if (is_acked)
 394                 return is_acked(rm, ack);
 395         return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
 396 }
 397
 398 /*
 399  * Returns true if there are no messages on the send and retransmit queues
 400  * which have a sequence number greater than or equal to the given sequence
 401  * number.
 402  */
 403 int rds_send_acked_before(struct rds_connection *conn, u64 seq)
 404 {
 405         struct rds_message *rm, *tmp;
 406         int ret = 1;
 407
 408         spin_lock(&conn->c_lock);
 409
 410         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
 411                 if (be64_to_cpu(rm->m_inc.i_hdr.h_sequence) < seq)
 412                         ret = 0;
 413                 break;
 414         }
 415
 416         list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
 417                 if (be64_to_cpu(rm->m_inc.i_hdr.h_sequence) < seq)
 418                         ret = 0;
 419                 break;
 420         }
 421
 422         spin_unlock(&conn->c_lock);
 423
 424         return ret;
 425 }
 426
 427 /*
 428  * This is pretty similar to what happens below in the ACK
 429  * handling code - except that we call here as soon as we get
 430  * the IB send completion on the RDMA op and the accompanying
 431  * message.
 432  */
 433 void rds_rdma_send_complete(struct rds_message *rm, int status)
 434 {
 435         struct rds_sock *rs = NULL;
 436         struct rm_rdma_op *ro;
 437         struct rds_notifier *notifier;
 438         unsigned long flags;
 439
 440         spin_lock_irqsave(&rm->m_rs_lock, flags);
 441
 442         ro = &rm->rdma;
 443         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
 444             ro->op_active && ro->op_notify && ro->op_notifier) {
 445                 notifier = ro->op_notifier;
 446                 rs = rm->m_rs;
 447                 sock_hold(rds_rs_to_sk(rs));
 448
 449                 notifier->n_status = status;
 450                 spin_lock(&rs->rs_lock);
 451                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
 452                 spin_unlock(&rs->rs_lock);
 453
 454                 ro->op_notifier = NULL;
 455         }
 456
 457         spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 458
 459         if (rs) {
 460                 rds_wake_sk_sleep(rs);
 461                 sock_put(rds_rs_to_sk(rs));
 462         }
 463 }
 464 EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
 465
 466 /*
 467  * Just like above, except looks at atomic op
 468  */
 469 void rds_atomic_send_complete(struct rds_message *rm, int status)
 470 {
 471         struct rds_sock *rs = NULL;
 472         struct rm_atomic_op *ao;
 473         struct rds_notifier *notifier;
 474         unsigned long flags;
 475
 476         spin_lock_irqsave(&rm->m_rs_lock, flags);
 477
 478         ao = &rm->atomic;
 479         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
 480             && ao->op_active && ao->op_notify && ao->op_notifier) {
 481                 notifier = ao->op_notifier;
 482                 rs = rm->m_rs;
 483                 sock_hold(rds_rs_to_sk(rs));
 484
 485                 notifier->n_status = status;
 486                 spin_lock(&rs->rs_lock);
 487                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
 488                 spin_unlock(&rs->rs_lock);
 489
 490                 ao->op_notifier = NULL;
 491         }
 492
 493         spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 494
 495         if (rs) {
 496                 rds_wake_sk_sleep(rs);
 497                 sock_put(rds_rs_to_sk(rs));
 498         }
 499 }
 500 EXPORT_SYMBOL_GPL(rds_atomic_send_complete);
 501
 502 /*
 503  * This is the same as rds_rdma_send_complete except we
 504  * don't do any locking - we have all the ingredients (message,
 505  * socket, socket lock) and can just move the notifier.
 506  */
 507 static inline void
 508 __rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
 509 {
 510         struct rm_rdma_op *ro;
 511         struct rm_atomic_op *ao;
 512
 513         ro = &rm->rdma;
 514         if (ro->op_active && ro->op_notify && ro->op_notifier) {
 515                 ro->op_notifier->n_status = status;
 516                 list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue);
 517                 ro->op_notifier = NULL;
 518         }
 519
 520         ao = &rm->atomic;
 521         if (ao->op_active && ao->op_notify && ao->op_notifier) {
 522                 ao->op_notifier->n_status = status;
 523                 list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue);
 524                 ao->op_notifier = NULL;
 525         }
 526
 527         /* No need to wake the app - caller does this */
 528 }
 529
 530 /*
 531  * This is called from the IB send completion when we detect
 532  * a RDMA operation that failed with remote access error.
 533  * So speed is not an issue here.
 534  */
 535 struct rds_message *rds_send_get_message(struct rds_connection *conn,
 536                                          struct rm_rdma_op *op)
 537 {
 538         struct rds_message *rm, *tmp, *found = NULL;
 539         unsigned long flags;
 540
 541         spin_lock_irqsave(&conn->c_lock, flags);
 542
 543         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
 544                 if (&rm->rdma == op) {
 545                         atomic_inc(&rm->m_refcount);
 546                         found = rm;
 547                         goto out;
 548                 }
 549         }
 550
 551         list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
 552                 if (&rm->rdma == op) {
 553                         atomic_inc(&rm->m_refcount);
 554                         found = rm;
 555                         break;
 556                 }
 557         }
 558
 559 out:
 560         spin_unlock_irqrestore(&conn->c_lock, flags);
 561
 562         return found;
 563 }
 564 EXPORT_SYMBOL_GPL(rds_send_get_message);
 565
 566 /*
 567  * This removes messages from the socket's list if they're on it.  The list
 568  * argument must be private to the caller, we must be able to modify it
 569  * without locks.  The messages must have a reference held for their
 570  * position on the list.  This function will drop that reference after
 571  * removing the messages from the 'messages' list regardless of if it found
 572  * the messages on the socket list or not.
 573  */
 574 void rds_send_remove_from_sock(struct list_head *messages, int status)
 575 {
 576         unsigned long flags;
 577         struct rds_sock *rs = NULL;
 578         struct rds_message *rm;
 579
 580         while (!list_empty(messages)) {
 581                 int was_on_sock = 0;
 582
 583                 rm = list_entry(messages->next, struct rds_message,
 584                                 m_conn_item);
 585                 list_del_init(&rm->m_conn_item);
 586
 587                 /*
 588                  * If we see this flag cleared then we're *sure* that someone
 589                  * else beat us to removing it from the sock.  If we race
 590                  * with their flag update we'll get the lock and then really
 591                  * see that the flag has been cleared.
 592                  *
 593                  * The message spinlock makes sure nobody clears rm->m_rs
 594                  * while we're messing with it. It does not prevent the
 595                  * message from being removed from the socket, though.
 596                  */
 597                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 598                 if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags))
 599                         goto unlock_and_drop;
 600
 601                 if (rs != rm->m_rs) {
 602                         if (rs) {
 603                                 rds_wake_sk_sleep(rs);
 604                                 sock_put(rds_rs_to_sk(rs));
 605                         }
 606                         rs = rm->m_rs;
 607                         sock_hold(rds_rs_to_sk(rs));
 608                 }
 609                 spin_lock(&rs->rs_lock);
 610
 611                 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
 612                         struct rm_rdma_op *ro = &rm->rdma;
 613                         struct rds_notifier *notifier;
 614
 615                         list_del_init(&rm->m_sock_item);
 616                         rds_send_sndbuf_remove(rs, rm);
 617
 618                         if (ro->op_active && ro->op_notifier &&
 619                                (ro->op_notify || (ro->op_recverr && status))) {
 620                                 notifier = ro->op_notifier;
 621                                 list_add_tail(&notifier->n_list,
 622                                                 &rs->rs_notify_queue);
 623                                 if (!notifier->n_status)
 624                                         notifier->n_status = status;
 625                                 rm->rdma.op_notifier = NULL;
 626                         }
 627                         was_on_sock = 1;
 628                         rm->m_rs = NULL;
 629                 }
 630                 spin_unlock(&rs->rs_lock);
 631
 632 unlock_and_drop:
 633                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 634                 rds_message_put(rm);
 635                 if (was_on_sock)
 636                         rds_message_put(rm);
 637         }
 638
 639         if (rs) {
 640                 rds_wake_sk_sleep(rs);
 641                 sock_put(rds_rs_to_sk(rs));
 642         }
 643 }
 644
 645 /*
 646  * Transports call here when they've determined that the receiver queued
 647  * messages up to, and including, the given sequence number.  Messages are
 648  * moved to the retrans queue when rds_send_xmit picks them off the send
 649  * queue. This means that in the TCP case, the message may not have been
 650  * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
 651  * checks the RDS_MSG_HAS_ACK_SEQ bit.
 652  *
 653  * XXX It's not clear to me how this is safely serialized with socket
 654  * destruction.  Maybe it should bail if it sees SOCK_DEAD.
 655  */
 656 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
 657                          is_acked_func is_acked)
 658 {
 659         struct rds_message *rm, *tmp;
 660         unsigned long flags;
 661         LIST_HEAD(list);
 662
 663         spin_lock_irqsave(&conn->c_lock, flags);
 664
 665         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
 666                 if (!rds_send_is_acked(rm, ack, is_acked))
 667                         break;
 668
 669                 list_move(&rm->m_conn_item, &list);
 670                 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
 671         }
 672
 673         /* order flag updates with spin locks */
 674         if (!list_empty(&list))
 675                 smp_mb__after_clear_bit();
 676
 677         spin_unlock_irqrestore(&conn->c_lock, flags);
 678
 679         /* now remove the messages from the sock list as needed */
 680         rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
 681 }
 682 EXPORT_SYMBOL_GPL(rds_send_drop_acked);
 683
 684 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
 685 {
 686         struct rds_message *rm, *tmp;
 687         struct rds_connection *conn;
 688         unsigned long flags;
 689         LIST_HEAD(list);
 690
 691         /* get all the messages we're dropping under the rs lock */
 692         spin_lock_irqsave(&rs->rs_lock, flags);
 693
 694         list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) {
 695                 if (dest && (dest->sin_addr.s_addr != rm->m_daddr ||
 696                              dest->sin_port != rm->m_inc.i_hdr.h_dport))
 697                         continue;
 698
 699                 list_move(&rm->m_sock_item, &list);
 700                 rds_send_sndbuf_remove(rs, rm);
 701                 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
 702         }
 703
 704         /* order flag updates with the rs lock */
 705         smp_mb__after_clear_bit();
 706
 707         spin_unlock_irqrestore(&rs->rs_lock, flags);
 708
 709         if (list_empty(&list))
 710                 return;
 711
 712         /* Remove the messages from the conn */
 713         list_for_each_entry(rm, &list, m_sock_item) {
 714
 715                 conn = rm->m_inc.i_conn;
 716
 717                 spin_lock_irqsave(&conn->c_lock, flags);
 718                 /*
 719                  * Maybe someone else beat us to removing rm from the conn.
 720                  * If we race with their flag update we'll get the lock and
 721                  * then really see that the flag has been cleared.
 722                  */
 723                 if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
 724                         spin_unlock_irqrestore(&conn->c_lock, flags);
 725                         continue;
 726                 }
 727                 list_del_init(&rm->m_conn_item);
 728                 spin_unlock_irqrestore(&conn->c_lock, flags);
 729
 730                 /*
 731                  * Couldn't grab m_rs_lock in top loop (lock ordering),
 732                  * but we can now.
 733                  */
 734                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 735
 736                 spin_lock(&rs->rs_lock);
 737                 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
 738                 spin_unlock(&rs->rs_lock);
 739
 740                 rm->m_rs = NULL;
 741                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 742
 743                 rds_message_put(rm);
 744         }
 745
 746         rds_wake_sk_sleep(rs);
 747
 748         while (!list_empty(&list)) {
 749                 rm = list_entry(list.next, struct rds_message, m_sock_item);
 750                 list_del_init(&rm->m_sock_item);
 751
 752                 rds_message_wait(rm);
 753                 rds_message_put(rm);
 754         }
 755 }
 756
 757 /*
 758  * we only want this to fire once so we use the callers 'queued'.  It's
 759  * possible that another thread can race with us and remove the
 760  * message from the flow with RDS_CANCEL_SENT_TO.
 761  */
 762 static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
 763                              struct rds_message *rm, __be16 sport,
 764                              __be16 dport, int *queued)
 765 {
 766         unsigned long flags;
 767         u32 len;
 768
 769         if (*queued)
 770                 goto out;
 771
 772         len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
 773
 774         /* this is the only place which holds both the socket's rs_lock
 775          * and the connection's c_lock */
 776         spin_lock_irqsave(&rs->rs_lock, flags);
 777
 778         /*
 779          * If there is a little space in sndbuf, we don't queue anything,
 780          * and userspace gets -EAGAIN. But poll() indicates there's send
 781          * room. This can lead to bad behavior (spinning) if snd_bytes isn't
 782          * freed up by incoming acks. So we check the *old* value of
 783          * rs_snd_bytes here to allow the last msg to exceed the buffer,
 784          * and poll() now knows no more data can be sent.
 785          */
 786         if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
 787                 rs->rs_snd_bytes += len;
 788
 789                 /* let recv side know we are close to send space exhaustion.
 790                  * This is probably not the optimal way to do it, as this
 791                  * means we set the flag on *all* messages as soon as our
 792                  * throughput hits a certain threshold.
 793                  */
 794                 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
 795                         __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
 796
 797                 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
 798                 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
 799                 rds_message_addref(rm);
 800                 rm->m_rs = rs;
 801
 802                 /* The code ordering is a little weird, but we're
 803                    trying to minimize the time we hold c_lock */
 804                 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
 805                 rm->m_inc.i_conn = conn;
 806                 rds_message_addref(rm);
 807
 808                 spin_lock(&conn->c_lock);
 809                 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(conn->c_next_tx_seq++);
 810                 list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
 811                 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
 812                 spin_unlock(&conn->c_lock);
 813
 814                 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
 815                          rm, len, rs, rs->rs_snd_bytes,
 816                          (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
 817
 818                 *queued = 1;
 819         }
 820
 821         spin_unlock_irqrestore(&rs->rs_lock, flags);
 822 out:
 823         return *queued;
 824 }
 825
 826 /*
 827  * rds_message is getting to be quite complicated, and we'd like to allocate
 828  * it all in one go. This figures out how big it needs to be up front.
 829  */
 830 static int rds_rm_size(struct msghdr *msg, int data_len)
 831 {
 832         struct cmsghdr *cmsg;
 833         int size = 0;
 834         int cmsg_groups = 0;
 835         int retval;
 836
 837         for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
 838                 if (!CMSG_OK(msg, cmsg))
 839                         return -EINVAL;
 840
 841                 if (cmsg->cmsg_level != SOL_RDS)
 842                         continue;
 843
 844                 switch (cmsg->cmsg_type) {
 845                 case RDS_CMSG_RDMA_ARGS:
 846                         cmsg_groups |= 1;
 847                         retval = rds_rdma_extra_size(CMSG_DATA(cmsg));
 848                         if (retval < 0)
 849                                 return retval;
 850                         size += retval;
 851
 852                         break;
 853
 854                 case RDS_CMSG_RDMA_DEST:
 855                 case RDS_CMSG_RDMA_MAP:
 856                         cmsg_groups |= 2;
 857                         /* these are valid but do no add any size */
 858                         break;
 859
 860                 case RDS_CMSG_ATOMIC_CSWP:
 861                 case RDS_CMSG_ATOMIC_FADD:
 862                         cmsg_groups |= 1;
 863                         size += sizeof(struct scatterlist);
 864                         break;
 865
 866                 default:
 867                         return -EINVAL;
 868                 }
 869
 870         }
 871
 872         size += ceil(data_len, PAGE_SIZE) * sizeof(struct scatterlist);
 873
 874         /* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */
 875         if (cmsg_groups == 3)
 876                 return -EINVAL;
 877
 878         return size;
 879 }
 880
 881 static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
 882                          struct msghdr *msg, int *allocated_mr)
 883 {
 884         struct cmsghdr *cmsg;
 885         int ret = 0;
 886
 887         for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
 888                 if (!CMSG_OK(msg, cmsg))
 889                         return -EINVAL;
 890
 891                 if (cmsg->cmsg_level != SOL_RDS)
 892                         continue;
 893
 894                 /* As a side effect, RDMA_DEST and RDMA_MAP will set
 895                  * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
 896                  */
 897                 switch (cmsg->cmsg_type) {
 898                 case RDS_CMSG_RDMA_ARGS:
 899                         ret = rds_cmsg_rdma_args(rs, rm, cmsg);
 900                         break;
 901
 902                 case RDS_CMSG_RDMA_DEST:
 903                         ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
 904                         break;
 905
 906                 case RDS_CMSG_RDMA_MAP:
 907                         ret = rds_cmsg_rdma_map(rs, rm, cmsg);
 908                         if (!ret)
 909                                 *allocated_mr = 1;
 910                         break;
 911                 case RDS_CMSG_ATOMIC_CSWP:
 912                 case RDS_CMSG_ATOMIC_FADD:
 913                         ret = rds_cmsg_atomic(rs, rm, cmsg);
 914                         break;
 915
 916                 default:
 917                         return -EINVAL;
 918                 }
 919
 920                 if (ret)
 921                         break;
 922         }
 923
 924         return ret;
 925 }
 926
 927 int rds_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
 928                 size_t payload_len)
 929 {
 930         struct sock *sk = sock->sk;
 931         struct rds_sock *rs = rds_sk_to_rs(sk);
 932         struct sockaddr_in *usin = (struct sockaddr_in *)msg->msg_name;
 933         __be32 daddr;
 934         __be16 dport;
 935         struct rds_message *rm = NULL;
 936         struct rds_connection *conn;
 937         int ret = 0;
 938         int queued = 0, allocated_mr = 0;
 939         int nonblock = msg->msg_flags & MSG_DONTWAIT;
 940         long timeo = sock_sndtimeo(sk, nonblock);
 941
 942         /* Mirror Linux UDP mirror of BSD error message compatibility */
 943         /* XXX: Perhaps MSG_MORE someday */
 944         if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT)) {
 945                 printk(KERN_INFO "msg_flags 0x%08X\n", msg->msg_flags);
 946                 ret = -EOPNOTSUPP;
 947                 goto out;
 948         }
 949
 950         if (msg->msg_namelen) {
 951                 /* XXX fail non-unicast destination IPs? */
 952                 if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) {
 953                         ret = -EINVAL;
 954                         goto out;
 955                 }
 956                 daddr = usin->sin_addr.s_addr;
 957                 dport = usin->sin_port;
 958         } else {
 959                 /* We only care about consistency with ->connect() */
 960                 lock_sock(sk);
 961                 daddr = rs->rs_conn_addr;
 962                 dport = rs->rs_conn_port;
 963                 release_sock(sk);
 964         }
 965
 966         /* racing with another thread binding seems ok here */
 967         if (daddr == 0 || rs->rs_bound_addr == 0) {
 968                 ret = -ENOTCONN; /* XXX not a great errno */
 969                 goto out;
 970         }
 971
 972         /* size of rm including all sgs */
 973         ret = rds_rm_size(msg, payload_len);
 974         if (ret < 0)
 975                 goto out;
 976
 977         rm = rds_message_alloc(ret, GFP_KERNEL);
 978         if (!rm) {
 979                 ret = -ENOMEM;
 980                 goto out;
 981         }
 982
 983         /* Attach data to the rm */
 984         if (payload_len) {
 985                 rm->data.op_sg = rds_message_alloc_sgs(rm, ceil(payload_len, PAGE_SIZE));
 986                 ret = rds_message_copy_from_user(rm, msg->msg_iov, payload_len);
 987                 if (ret)
 988                         goto out;
 989         }
 990         rm->data.op_active = 1;
 991
 992         rm->m_daddr = daddr;
 993
 994         /* rds_conn_create has a spinlock that runs with IRQ off.
 995          * Caching the conn in the socket helps a lot. */
 996         if (rs->rs_conn && rs->rs_conn->c_faddr == daddr)
 997                 conn = rs->rs_conn;
 998         else {
 999                 conn = rds_conn_create_outgoing(rs->rs_bound_addr, daddr,
1000                                         rs->rs_transport,
1001                                         sock->sk->sk_allocation);
1002                 if (IS_ERR(conn)) {
1003                         ret = PTR_ERR(conn);
1004                         goto out;
1005                 }
1006                 rs->rs_conn = conn;
1007         }
1008
1009         /* Parse any control messages the user may have included. */
1010         ret = rds_cmsg_send(rs, rm, msg, &allocated_mr);
1011         if (ret)
1012                 goto out;
1013
1014         if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) {
1015                 if (printk_ratelimit())
1016                         printk(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
1017                                &rm->rdma, conn->c_trans->xmit_rdma);
1018                 ret = -EOPNOTSUPP;
1019                 goto out;
1020         }
1021
1022         if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
1023                 if (printk_ratelimit())
1024                         printk(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
1025                                &rm->atomic, conn->c_trans->xmit_atomic);
1026                 ret = -EOPNOTSUPP;
1027                 goto out;
1028         }
1029
1030         /* If the connection is down, trigger a connect. We may
1031          * have scheduled a delayed reconnect however - in this case
1032          * we should not interfere.
1033          */
1034         if (rds_conn_state(conn) == RDS_CONN_DOWN &&
1035             !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
1036                 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
1037
1038         ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
1039         if (ret) {
1040                 rs->rs_seen_congestion = 1;
1041                 goto out;
1042         }
1043
1044         while (!rds_send_queue_rm(rs, conn, rm, rs->rs_bound_port,
1045                                   dport, &queued)) {
1046                 rds_stats_inc(s_send_queue_full);
1047                 /* XXX make sure this is reasonable */
1048                 if (payload_len > rds_sk_sndbuf(rs)) {
1049                         ret = -EMSGSIZE;
1050                         goto out;
1051                 }
1052                 if (nonblock) {
1053                         ret = -EAGAIN;
1054                         goto out;
1055                 }
1056
1057                 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
1058                                         rds_send_queue_rm(rs, conn, rm,
1059                                                           rs->rs_bound_port,
1060                                                           dport,
1061                                                           &queued),
1062                                         timeo);
1063                 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
1064                 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
1065                         continue;
1066
1067                 ret = timeo;
1068                 if (ret == 0)
1069                         ret = -ETIMEDOUT;
1070                 goto out;
1071         }
1072
1073         /*
1074          * By now we've committed to the send.  We reuse rds_send_worker()
1075          * to retry sends in the rds thread if the transport asks us to.
1076          */
1077         rds_stats_inc(s_send_queued);
1078
1079         if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
1080                 rds_send_xmit(conn);
1081
1082         rds_message_put(rm);
1083         return payload_len;
1084
1085 out:
1086         /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1087          * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1088          * or in any other way, we need to destroy the MR again */
1089         if (allocated_mr)
1090                 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
1091
1092         if (rm)
1093                 rds_message_put(rm);
1094         return ret;
1095 }
1096
1097 /*
1098  * Reply to a ping packet.
1099  */
1100 int
1101 rds_send_pong(struct rds_connection *conn, __be16 dport)
1102 {
1103         struct rds_message *rm;
1104         unsigned long flags;
1105         int ret = 0;
1106
1107         rm = rds_message_alloc(0, GFP_ATOMIC);
1108         if (!rm) {
1109                 ret = -ENOMEM;
1110                 goto out;
1111         }
1112
1113         rm->m_daddr = conn->c_faddr;
1114         rm->data.op_active = 1;
1115
1116         /* If the connection is down, trigger a connect. We may
1117          * have scheduled a delayed reconnect however - in this case
1118          * we should not interfere.
1119          */
1120         if (rds_conn_state(conn) == RDS_CONN_DOWN &&
1121             !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
1122                 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
1123
1124         ret = rds_cong_wait(conn->c_fcong, dport, 1, NULL);
1125         if (ret)
1126                 goto out;
1127
1128         spin_lock_irqsave(&conn->c_lock, flags);
1129         list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
1130         set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
1131         rds_message_addref(rm);
1132         rm->m_inc.i_conn = conn;
1133
1134         rds_message_populate_header(&rm->m_inc.i_hdr, 0, dport,
1135                                     conn->c_next_tx_seq);
1136         conn->c_next_tx_seq++;
1137         spin_unlock_irqrestore(&conn->c_lock, flags);
1138
1139         rds_stats_inc(s_send_queued);
1140         rds_stats_inc(s_send_pong);
1141
1142         if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
1143                 rds_send_xmit(conn);
1144
1145         rds_message_put(rm);
1146         return 0;
1147
1148 out:
1149         if (rm)
1150                 rds_message_put(rm);
1151         return ret;
1152 }