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