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/moduleparam.h>
  35 #include <linux/gfp.h>
  36 #include <net/sock.h>
  37 #include <linux/in.h>
  38 #include <linux/list.h>
  39 #include <linux/ratelimit.h>
  40 #include <linux/export.h>
  41 #include <linux/sizes.h>
  42
  43 #include "rds.h"
  44
  45 /* When transmitting messages in rds_send_xmit, we need to emerge from
  46  * time to time and briefly release the CPU. Otherwise the softlock watchdog
  47  * will kick our shin.
  48  * Also, it seems fairer to not let one busy connection stall all the
  49  * others.
  50  *
  51  * send_batch_count is the number of times we'll loop in send_xmit. Setting
  52  * it to 0 will restore the old behavior (where we looped until we had
  53  * drained the queue).
  54  */
  55 static int send_batch_count = SZ_1K;
  56 module_param(send_batch_count, int, 0444);
  57 MODULE_PARM_DESC(send_batch_count, " batch factor when working the send queue");
  58
  59 static void rds_send_remove_from_sock(struct list_head *messages, int status);
  60
  61 /*
  62  * Reset the send state.  Callers must ensure that this doesn't race with
  63  * rds_send_xmit().
  64  */
  65 void rds_send_reset(struct rds_connection *conn)
  66 {
  67         struct rds_message *rm, *tmp;
  68         unsigned long flags;
  69
  70         if (conn->c_xmit_rm) {
  71                 rm = conn->c_xmit_rm;
  72                 conn->c_xmit_rm = NULL;
  73                 /* Tell the user the RDMA op is no longer mapped by the
  74                  * transport. This isn't entirely true (it's flushed out
  75                  * independently) but as the connection is down, there's
  76                  * no ongoing RDMA to/from that memory */
  77                 rds_message_unmapped(rm);
  78                 rds_message_put(rm);
  79         }
  80
  81         conn->c_xmit_sg = 0;
  82         conn->c_xmit_hdr_off = 0;
  83         conn->c_xmit_data_off = 0;
  84         conn->c_xmit_atomic_sent = 0;
  85         conn->c_xmit_rdma_sent = 0;
  86         conn->c_xmit_data_sent = 0;
  87
  88         conn->c_map_queued = 0;
  89
  90         conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
  91         conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
  92
  93         /* Mark messages as retransmissions, and move them to the send q */
  94         spin_lock_irqsave(&conn->c_lock, flags);
  95         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
  96                 set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
  97                 set_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags);
  98         }
  99         list_splice_init(&conn->c_retrans, &conn->c_send_queue);
 100         spin_unlock_irqrestore(&conn->c_lock, flags);
 101 }
 102 EXPORT_SYMBOL_GPL(rds_send_reset);
 103
 104 static int acquire_in_xmit(struct rds_connection *conn)
 105 {
 106         return test_and_set_bit(RDS_IN_XMIT, &conn->c_flags) == 0;
 107 }
 108
 109 static void release_in_xmit(struct rds_connection *conn)
 110 {
 111         clear_bit(RDS_IN_XMIT, &conn->c_flags);
 112         smp_mb__after_atomic();
 113         /*
 114          * We don't use wait_on_bit()/wake_up_bit() because our waking is in a
 115          * hot path and finding waiters is very rare.  We don't want to walk
 116          * the system-wide hashed waitqueue buckets in the fast path only to
 117          * almost never find waiters.
 118          */
 119         if (waitqueue_active(&conn->c_waitq))
 120                 wake_up_all(&conn->c_waitq);
 121 }
 122
 123 /*
 124  * We're making the conscious trade-off here to only send one message
 125  * down the connection at a time.
 126  *   Pro:
 127  *      - tx queueing is a simple fifo list
 128  *      - reassembly is optional and easily done by transports per conn
 129  *      - no per flow rx lookup at all, straight to the socket
 130  *      - less per-frag memory and wire overhead
 131  *   Con:
 132  *      - queued acks can be delayed behind large messages
 133  *   Depends:
 134  *      - small message latency is higher behind queued large messages
 135  *      - large message latency isn't starved by intervening small sends
 136  */
 137 int rds_send_xmit(struct rds_connection *conn)
 138 {
 139         struct rds_message *rm;
 140         unsigned long flags;
 141         unsigned int tmp;
 142         struct scatterlist *sg;
 143         int ret = 0;
 144         LIST_HEAD(to_be_dropped);
 145         int batch_count;
 146         unsigned long send_gen = 0;
 147
 148 restart:
 149         batch_count = 0;
 150
 151         /*
 152          * sendmsg calls here after having queued its message on the send
 153          * queue.  We only have one task feeding the connection at a time.  If
 154          * another thread is already feeding the queue then we back off.  This
 155          * avoids blocking the caller and trading per-connection data between
 156          * caches per message.
 157          */
 158         if (!acquire_in_xmit(conn)) {
 159                 rds_stats_inc(s_send_lock_contention);
 160                 ret = -ENOMEM;
 161                 goto out;
 162         }
 163
 164         /*
 165          * we record the send generation after doing the xmit acquire.
 166          * if someone else manages to jump in and do some work, we'll use
 167          * this to avoid a goto restart farther down.
 168          *
 169          * The acquire_in_xmit() check above ensures that only one
 170          * caller can increment c_send_gen at any time.
 171          */
 172         conn->c_send_gen++;
 173         send_gen = conn->c_send_gen;
 174
 175         /*
 176          * rds_conn_shutdown() sets the conn state and then tests RDS_IN_XMIT,
 177          * we do the opposite to avoid races.
 178          */
 179         if (!rds_conn_up(conn)) {
 180                 release_in_xmit(conn);
 181                 ret = 0;
 182                 goto out;
 183         }
 184
 185         if (conn->c_trans->xmit_prepare)
 186                 conn->c_trans->xmit_prepare(conn);
 187
 188         /*
 189          * spin trying to push headers and data down the connection until
 190          * the connection doesn't make forward progress.
 191          */
 192         while (1) {
 193
 194                 rm = conn->c_xmit_rm;
 195
 196                 /*
 197                  * If between sending messages, we can send a pending congestion
 198                  * map update.
 199                  */
 200                 if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) {
 201                         rm = rds_cong_update_alloc(conn);
 202                         if (IS_ERR(rm)) {
 203                                 ret = PTR_ERR(rm);
 204                                 break;
 205                         }
 206                         rm->data.op_active = 1;
 207
 208                         conn->c_xmit_rm = rm;
 209                 }
 210
 211                 /*
 212                  * If not already working on one, grab the next message.
 213                  *
 214                  * c_xmit_rm holds a ref while we're sending this message down
 215                  * the connction.  We can use this ref while holding the
 216                  * send_sem.. rds_send_reset() is serialized with it.
 217                  */
 218                 if (!rm) {
 219                         unsigned int len;
 220
 221                         batch_count++;
 222
 223                         /* we want to process as big a batch as we can, but
 224                          * we also want to avoid softlockups.  If we've been
 225                          * through a lot of messages, lets back off and see
 226                          * if anyone else jumps in
 227                          */
 228                         if (batch_count >= send_batch_count)
 229                                 goto over_batch;
 230
 231                         spin_lock_irqsave(&conn->c_lock, flags);
 232
 233                         if (!list_empty(&conn->c_send_queue)) {
 234                                 rm = list_entry(conn->c_send_queue.next,
 235                                                 struct rds_message,
 236                                                 m_conn_item);
 237                                 rds_message_addref(rm);
 238
 239                                 /*
 240                                  * Move the message from the send queue to the retransmit
 241                                  * list right away.
 242                                  */
 243                                 list_move_tail(&rm->m_conn_item, &conn->c_retrans);
 244                         }
 245
 246                         spin_unlock_irqrestore(&conn->c_lock, flags);
 247
 248                         if (!rm)
 249                                 break;
 250
 251                         /* Unfortunately, the way Infiniband deals with
 252                          * RDMA to a bad MR key is by moving the entire
 253                          * queue pair to error state. We cold possibly
 254                          * recover from that, but right now we drop the
 255                          * connection.
 256                          * Therefore, we never retransmit messages with RDMA ops.
 257                          */
 258                         if (rm->rdma.op_active &&
 259                             test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags)) {
 260                                 spin_lock_irqsave(&conn->c_lock, flags);
 261                                 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags))
 262                                         list_move(&rm->m_conn_item, &to_be_dropped);
 263                                 spin_unlock_irqrestore(&conn->c_lock, flags);
 264                                 continue;
 265                         }
 266
 267                         /* Require an ACK every once in a while */
 268                         len = ntohl(rm->m_inc.i_hdr.h_len);
 269                         if (conn->c_unacked_packets == 0 ||
 270                             conn->c_unacked_bytes < len) {
 271                                 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
 272
 273                                 conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
 274                                 conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
 275                                 rds_stats_inc(s_send_ack_required);
 276                         } else {
 277                                 conn->c_unacked_bytes -= len;
 278                                 conn->c_unacked_packets--;
 279                         }
 280
 281                         conn->c_xmit_rm = rm;
 282                 }
 283
 284                 /* The transport either sends the whole rdma or none of it */
 285                 if (rm->rdma.op_active && !conn->c_xmit_rdma_sent) {
 286                         rm->m_final_op = &rm->rdma;
 287                         /* The transport owns the mapped memory for now.
 288                          * You can't unmap it while it's on the send queue
 289                          */
 290                         set_bit(RDS_MSG_MAPPED, &rm->m_flags);
 291                         ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
 292                         if (ret) {
 293                                 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
 294                                 wake_up_interruptible(&rm->m_flush_wait);
 295                                 break;
 296                         }
 297                         conn->c_xmit_rdma_sent = 1;
 298
 299                 }
 300
 301                 if (rm->atomic.op_active && !conn->c_xmit_atomic_sent) {
 302                         rm->m_final_op = &rm->atomic;
 303                         /* The transport owns the mapped memory for now.
 304                          * You can't unmap it while it's on the send queue
 305                          */
 306                         set_bit(RDS_MSG_MAPPED, &rm->m_flags);
 307                         ret = conn->c_trans->xmit_atomic(conn, &rm->atomic);
 308                         if (ret) {
 309                                 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
 310                                 wake_up_interruptible(&rm->m_flush_wait);
 311                                 break;
 312                         }
 313                         conn->c_xmit_atomic_sent = 1;
 314
 315                 }
 316
 317                 /*
 318                  * A number of cases require an RDS header to be sent
 319                  * even if there is no data.
 320                  * We permit 0-byte sends; rds-ping depends on this.
 321                  * However, if there are exclusively attached silent ops,
 322                  * we skip the hdr/data send, to enable silent operation.
 323                  */
 324                 if (rm->data.op_nents == 0) {
 325                         int ops_present;
 326                         int all_ops_are_silent = 1;
 327
 328                         ops_present = (rm->atomic.op_active || rm->rdma.op_active);
 329                         if (rm->atomic.op_active && !rm->atomic.op_silent)
 330                                 all_ops_are_silent = 0;
 331                         if (rm->rdma.op_active && !rm->rdma.op_silent)
 332                                 all_ops_are_silent = 0;
 333
 334                         if (ops_present && all_ops_are_silent
 335                             && !rm->m_rdma_cookie)
 336                                 rm->data.op_active = 0;
 337                 }
 338
 339                 if (rm->data.op_active && !conn->c_xmit_data_sent) {
 340                         rm->m_final_op = &rm->data;
 341                         ret = conn->c_trans->xmit(conn, rm,
 342                                                   conn->c_xmit_hdr_off,
 343                                                   conn->c_xmit_sg,
 344                                                   conn->c_xmit_data_off);
 345                         if (ret <= 0)
 346                                 break;
 347
 348                         if (conn->c_xmit_hdr_off < sizeof(struct rds_header)) {
 349                                 tmp = min_t(int, ret,
 350                                             sizeof(struct rds_header) -
 351                                             conn->c_xmit_hdr_off);
 352                                 conn->c_xmit_hdr_off += tmp;
 353                                 ret -= tmp;
 354                         }
 355
 356                         sg = &rm->data.op_sg[conn->c_xmit_sg];
 357                         while (ret) {
 358                                 tmp = min_t(int, ret, sg->length -
 359                                                       conn->c_xmit_data_off);
 360                                 conn->c_xmit_data_off += tmp;
 361                                 ret -= tmp;
 362                                 if (conn->c_xmit_data_off == sg->length) {
 363                                         conn->c_xmit_data_off = 0;
 364                                         sg++;
 365                                         conn->c_xmit_sg++;
 366                                         BUG_ON(ret != 0 &&
 367                                                conn->c_xmit_sg == rm->data.op_nents);
 368                                 }
 369                         }
 370
 371                         if (conn->c_xmit_hdr_off == sizeof(struct rds_header) &&
 372                             (conn->c_xmit_sg == rm->data.op_nents))
 373                                 conn->c_xmit_data_sent = 1;
 374                 }
 375
 376                 /*
 377                  * A rm will only take multiple times through this loop
 378                  * if there is a data op. Thus, if the data is sent (or there was
 379                  * none), then we're done with the rm.
 380                  */
 381                 if (!rm->data.op_active || conn->c_xmit_data_sent) {
 382                         conn->c_xmit_rm = NULL;
 383                         conn->c_xmit_sg = 0;
 384                         conn->c_xmit_hdr_off = 0;
 385                         conn->c_xmit_data_off = 0;
 386                         conn->c_xmit_rdma_sent = 0;
 387                         conn->c_xmit_atomic_sent = 0;
 388                         conn->c_xmit_data_sent = 0;
 389
 390                         rds_message_put(rm);
 391                 }
 392         }
 393
 394 over_batch:
 395         if (conn->c_trans->xmit_complete)
 396                 conn->c_trans->xmit_complete(conn);
 397         release_in_xmit(conn);
 398
 399         /* Nuke any messages we decided not to retransmit. */
 400         if (!list_empty(&to_be_dropped)) {
 401                 /* irqs on here, so we can put(), unlike above */
 402                 list_for_each_entry(rm, &to_be_dropped, m_conn_item)
 403                         rds_message_put(rm);
 404                 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
 405         }
 406
 407         /*
 408          * Other senders can queue a message after we last test the send queue
 409          * but before we clear RDS_IN_XMIT.  In that case they'd back off and
 410          * not try and send their newly queued message.  We need to check the
 411          * send queue after having cleared RDS_IN_XMIT so that their message
 412          * doesn't get stuck on the send queue.
 413          *
 414          * If the transport cannot continue (i.e ret != 0), then it must
 415          * call us when more room is available, such as from the tx
 416          * completion handler.
 417          *
 418          * We have an extra generation check here so that if someone manages
 419          * to jump in after our release_in_xmit, we'll see that they have done
 420          * some work and we will skip our goto
 421          */
 422         if (ret == 0) {
 423                 smp_mb();
 424                 if ((test_bit(0, &conn->c_map_queued) ||
 425                      !list_empty(&conn->c_send_queue)) &&
 426                     send_gen == conn->c_send_gen) {
 427                         rds_stats_inc(s_send_lock_queue_raced);
 428                         if (batch_count < send_batch_count)
 429                                 goto restart;
 430                         queue_delayed_work(rds_wq, &conn->c_send_w, 1);
 431                 }
 432         }
 433 out:
 434         return ret;
 435 }
 436 EXPORT_SYMBOL_GPL(rds_send_xmit);
 437
 438 static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
 439 {
 440         u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
 441
 442         assert_spin_locked(&rs->rs_lock);
 443
 444         BUG_ON(rs->rs_snd_bytes < len);
 445         rs->rs_snd_bytes -= len;
 446
 447         if (rs->rs_snd_bytes == 0)
 448                 rds_stats_inc(s_send_queue_empty);
 449 }
 450
 451 static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
 452                                     is_acked_func is_acked)
 453 {
 454         if (is_acked)
 455                 return is_acked(rm, ack);
 456         return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
 457 }
 458
 459 /*
 460  * This is pretty similar to what happens below in the ACK
 461  * handling code - except that we call here as soon as we get
 462  * the IB send completion on the RDMA op and the accompanying
 463  * message.
 464  */
 465 void rds_rdma_send_complete(struct rds_message *rm, int status)
 466 {
 467         struct rds_sock *rs = NULL;
 468         struct rm_rdma_op *ro;
 469         struct rds_notifier *notifier;
 470         unsigned long flags;
 471
 472         spin_lock_irqsave(&rm->m_rs_lock, flags);
 473
 474         ro = &rm->rdma;
 475         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
 476             ro->op_active && ro->op_notify && ro->op_notifier) {
 477                 notifier = ro->op_notifier;
 478                 rs = rm->m_rs;
 479                 sock_hold(rds_rs_to_sk(rs));
 480
 481                 notifier->n_status = status;
 482                 spin_lock(&rs->rs_lock);
 483                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
 484                 spin_unlock(&rs->rs_lock);
 485
 486                 ro->op_notifier = NULL;
 487         }
 488
 489         spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 490
 491         if (rs) {
 492                 rds_wake_sk_sleep(rs);
 493                 sock_put(rds_rs_to_sk(rs));
 494         }
 495 }
 496 EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
 497
 498 /*
 499  * Just like above, except looks at atomic op
 500  */
 501 void rds_atomic_send_complete(struct rds_message *rm, int status)
 502 {
 503         struct rds_sock *rs = NULL;
 504         struct rm_atomic_op *ao;
 505         struct rds_notifier *notifier;
 506         unsigned long flags;
 507
 508         spin_lock_irqsave(&rm->m_rs_lock, flags);
 509
 510         ao = &rm->atomic;
 511         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
 512             && ao->op_active && ao->op_notify && ao->op_notifier) {
 513                 notifier = ao->op_notifier;
 514                 rs = rm->m_rs;
 515                 sock_hold(rds_rs_to_sk(rs));
 516
 517                 notifier->n_status = status;
 518                 spin_lock(&rs->rs_lock);
 519                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
 520                 spin_unlock(&rs->rs_lock);
 521
 522                 ao->op_notifier = NULL;
 523         }
 524
 525         spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 526
 527         if (rs) {
 528                 rds_wake_sk_sleep(rs);
 529                 sock_put(rds_rs_to_sk(rs));
 530         }
 531 }
 532 EXPORT_SYMBOL_GPL(rds_atomic_send_complete);
 533
 534 /*
 535  * This is the same as rds_rdma_send_complete except we
 536  * don't do any locking - we have all the ingredients (message,
 537  * socket, socket lock) and can just move the notifier.
 538  */
 539 static inline void
 540 __rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
 541 {
 542         struct rm_rdma_op *ro;
 543         struct rm_atomic_op *ao;
 544
 545         ro = &rm->rdma;
 546         if (ro->op_active && ro->op_notify && ro->op_notifier) {
 547                 ro->op_notifier->n_status = status;
 548                 list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue);
 549                 ro->op_notifier = NULL;
 550         }
 551
 552         ao = &rm->atomic;
 553         if (ao->op_active && ao->op_notify && ao->op_notifier) {
 554                 ao->op_notifier->n_status = status;
 555                 list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue);
 556                 ao->op_notifier = NULL;
 557         }
 558
 559         /* No need to wake the app - caller does this */
 560 }
 561
 562 /*
 563  * This is called from the IB send completion when we detect
 564  * a RDMA operation that failed with remote access error.
 565  * So speed is not an issue here.
 566  */
 567 struct rds_message *rds_send_get_message(struct rds_connection *conn,
 568                                          struct rm_rdma_op *op)
 569 {
 570         struct rds_message *rm, *tmp, *found = NULL;
 571         unsigned long flags;
 572
 573         spin_lock_irqsave(&conn->c_lock, flags);
 574
 575         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
 576                 if (&rm->rdma == op) {
 577                         atomic_inc(&rm->m_refcount);
 578                         found = rm;
 579                         goto out;
 580                 }
 581         }
 582
 583         list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
 584                 if (&rm->rdma == op) {
 585                         atomic_inc(&rm->m_refcount);
 586                         found = rm;
 587                         break;
 588                 }
 589         }
 590
 591 out:
 592         spin_unlock_irqrestore(&conn->c_lock, flags);
 593
 594         return found;
 595 }
 596 EXPORT_SYMBOL_GPL(rds_send_get_message);
 597
 598 /*
 599  * This removes messages from the socket's list if they're on it.  The list
 600  * argument must be private to the caller, we must be able to modify it
 601  * without locks.  The messages must have a reference held for their
 602  * position on the list.  This function will drop that reference after
 603  * removing the messages from the 'messages' list regardless of if it found
 604  * the messages on the socket list or not.
 605  */
 606 static void rds_send_remove_from_sock(struct list_head *messages, int status)
 607 {
 608         unsigned long flags;
 609         struct rds_sock *rs = NULL;
 610         struct rds_message *rm;
 611
 612         while (!list_empty(messages)) {
 613                 int was_on_sock = 0;
 614
 615                 rm = list_entry(messages->next, struct rds_message,
 616                                 m_conn_item);
 617                 list_del_init(&rm->m_conn_item);
 618
 619                 /*
 620                  * If we see this flag cleared then we're *sure* that someone
 621                  * else beat us to removing it from the sock.  If we race
 622                  * with their flag update we'll get the lock and then really
 623                  * see that the flag has been cleared.
 624                  *
 625                  * The message spinlock makes sure nobody clears rm->m_rs
 626                  * while we're messing with it. It does not prevent the
 627                  * message from being removed from the socket, though.
 628                  */
 629                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 630                 if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags))
 631                         goto unlock_and_drop;
 632
 633                 if (rs != rm->m_rs) {
 634                         if (rs) {
 635                                 rds_wake_sk_sleep(rs);
 636                                 sock_put(rds_rs_to_sk(rs));
 637                         }
 638                         rs = rm->m_rs;
 639                         if (rs)
 640                                 sock_hold(rds_rs_to_sk(rs));
 641                 }
 642                 if (!rs)
 643                         goto unlock_and_drop;
 644                 spin_lock(&rs->rs_lock);
 645
 646                 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
 647                         struct rm_rdma_op *ro = &rm->rdma;
 648                         struct rds_notifier *notifier;
 649
 650                         list_del_init(&rm->m_sock_item);
 651                         rds_send_sndbuf_remove(rs, rm);
 652
 653                         if (ro->op_active && ro->op_notifier &&
 654                                (ro->op_notify || (ro->op_recverr && status))) {
 655                                 notifier = ro->op_notifier;
 656                                 list_add_tail(&notifier->n_list,
 657                                                 &rs->rs_notify_queue);
 658                                 if (!notifier->n_status)
 659                                         notifier->n_status = status;
 660                                 rm->rdma.op_notifier = NULL;
 661                         }
 662                         was_on_sock = 1;
 663                         rm->m_rs = NULL;
 664                 }
 665                 spin_unlock(&rs->rs_lock);
 666
 667 unlock_and_drop:
 668                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 669                 rds_message_put(rm);
 670                 if (was_on_sock)
 671                         rds_message_put(rm);
 672         }
 673
 674         if (rs) {
 675                 rds_wake_sk_sleep(rs);
 676                 sock_put(rds_rs_to_sk(rs));
 677         }
 678 }
 679
 680 /*
 681  * Transports call here when they've determined that the receiver queued
 682  * messages up to, and including, the given sequence number.  Messages are
 683  * moved to the retrans queue when rds_send_xmit picks them off the send
 684  * queue. This means that in the TCP case, the message may not have been
 685  * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
 686  * checks the RDS_MSG_HAS_ACK_SEQ bit.
 687  */
 688 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
 689                          is_acked_func is_acked)
 690 {
 691         struct rds_message *rm, *tmp;
 692         unsigned long flags;
 693         LIST_HEAD(list);
 694
 695         spin_lock_irqsave(&conn->c_lock, flags);
 696
 697         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
 698                 if (!rds_send_is_acked(rm, ack, is_acked))
 699                         break;
 700
 701                 list_move(&rm->m_conn_item, &list);
 702                 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
 703         }
 704
 705         /* order flag updates with spin locks */
 706         if (!list_empty(&list))
 707                 smp_mb__after_atomic();
 708
 709         spin_unlock_irqrestore(&conn->c_lock, flags);
 710
 711         /* now remove the messages from the sock list as needed */
 712         rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
 713 }
 714 EXPORT_SYMBOL_GPL(rds_send_drop_acked);
 715
 716 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
 717 {
 718         struct rds_message *rm, *tmp;
 719         struct rds_connection *conn;
 720         unsigned long flags;
 721         LIST_HEAD(list);
 722
 723         /* get all the messages we're dropping under the rs lock */
 724         spin_lock_irqsave(&rs->rs_lock, flags);
 725
 726         list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) {
 727                 if (dest && (dest->sin_addr.s_addr != rm->m_daddr ||
 728                              dest->sin_port != rm->m_inc.i_hdr.h_dport))
 729                         continue;
 730
 731                 list_move(&rm->m_sock_item, &list);
 732                 rds_send_sndbuf_remove(rs, rm);
 733                 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
 734         }
 735
 736         /* order flag updates with the rs lock */
 737         smp_mb__after_atomic();
 738
 739         spin_unlock_irqrestore(&rs->rs_lock, flags);
 740
 741         if (list_empty(&list))
 742                 return;
 743
 744         /* Remove the messages from the conn */
 745         list_for_each_entry(rm, &list, m_sock_item) {
 746
 747                 conn = rm->m_inc.i_conn;
 748
 749                 spin_lock_irqsave(&conn->c_lock, flags);
 750                 /*
 751                  * Maybe someone else beat us to removing rm from the conn.
 752                  * If we race with their flag update we'll get the lock and
 753                  * then really see that the flag has been cleared.
 754                  */
 755                 if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
 756                         spin_unlock_irqrestore(&conn->c_lock, flags);
 757                         spin_lock_irqsave(&rm->m_rs_lock, flags);
 758                         rm->m_rs = NULL;
 759                         spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 760                         continue;
 761                 }
 762                 list_del_init(&rm->m_conn_item);
 763                 spin_unlock_irqrestore(&conn->c_lock, flags);
 764
 765                 /*
 766                  * Couldn't grab m_rs_lock in top loop (lock ordering),
 767                  * but we can now.
 768                  */
 769                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 770
 771                 spin_lock(&rs->rs_lock);
 772                 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
 773                 spin_unlock(&rs->rs_lock);
 774
 775                 rm->m_rs = NULL;
 776                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 777
 778                 rds_message_put(rm);
 779         }
 780
 781         rds_wake_sk_sleep(rs);
 782
 783         while (!list_empty(&list)) {
 784                 rm = list_entry(list.next, struct rds_message, m_sock_item);
 785                 list_del_init(&rm->m_sock_item);
 786                 rds_message_wait(rm);
 787
 788                 /* just in case the code above skipped this message
 789                  * because RDS_MSG_ON_CONN wasn't set, run it again here
 790                  * taking m_rs_lock is the only thing that keeps us
 791                  * from racing with ack processing.
 792                  */
 793                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 794
 795                 spin_lock(&rs->rs_lock);
 796                 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
 797                 spin_unlock(&rs->rs_lock);
 798
 799                 rm->m_rs = NULL;
 800                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 801
 802                 rds_message_put(rm);
 803         }
 804 }
 805
 806 /*
 807  * we only want this to fire once so we use the callers 'queued'.  It's
 808  * possible that another thread can race with us and remove the
 809  * message from the flow with RDS_CANCEL_SENT_TO.
 810  */
 811 static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
 812                              struct rds_message *rm, __be16 sport,
 813                              __be16 dport, int *queued)
 814 {
 815         unsigned long flags;
 816         u32 len;
 817
 818         if (*queued)
 819                 goto out;
 820
 821         len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
 822
 823         /* this is the only place which holds both the socket's rs_lock
 824          * and the connection's c_lock */
 825         spin_lock_irqsave(&rs->rs_lock, flags);
 826
 827         /*
 828          * If there is a little space in sndbuf, we don't queue anything,
 829          * and userspace gets -EAGAIN. But poll() indicates there's send
 830          * room. This can lead to bad behavior (spinning) if snd_bytes isn't
 831          * freed up by incoming acks. So we check the *old* value of
 832          * rs_snd_bytes here to allow the last msg to exceed the buffer,
 833          * and poll() now knows no more data can be sent.
 834          */
 835         if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
 836                 rs->rs_snd_bytes += len;
 837
 838                 /* let recv side know we are close to send space exhaustion.
 839                  * This is probably not the optimal way to do it, as this
 840                  * means we set the flag on *all* messages as soon as our
 841                  * throughput hits a certain threshold.
 842                  */
 843                 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
 844                         __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
 845
 846                 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
 847                 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
 848                 rds_message_addref(rm);
 849                 rm->m_rs = rs;
 850
 851                 /* The code ordering is a little weird, but we're
 852                    trying to minimize the time we hold c_lock */
 853                 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
 854                 rm->m_inc.i_conn = conn;
 855                 rds_message_addref(rm);
 856
 857                 spin_lock(&conn->c_lock);
 858                 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(conn->c_next_tx_seq++);
 859                 list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
 860                 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
 861                 spin_unlock(&conn->c_lock);
 862
 863                 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
 864                          rm, len, rs, rs->rs_snd_bytes,
 865                          (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
 866
 867                 *queued = 1;
 868         }
 869
 870         spin_unlock_irqrestore(&rs->rs_lock, flags);
 871 out:
 872         return *queued;
 873 }
 874
 875 /*
 876  * rds_message is getting to be quite complicated, and we'd like to allocate
 877  * it all in one go. This figures out how big it needs to be up front.
 878  */
 879 static int rds_rm_size(struct msghdr *msg, int data_len)
 880 {
 881         struct cmsghdr *cmsg;
 882         int size = 0;
 883         int cmsg_groups = 0;
 884         int retval;
 885
 886         for_each_cmsghdr(cmsg, msg) {
 887                 if (!CMSG_OK(msg, cmsg))
 888                         return -EINVAL;
 889
 890                 if (cmsg->cmsg_level != SOL_RDS)
 891                         continue;
 892
 893                 switch (cmsg->cmsg_type) {
 894                 case RDS_CMSG_RDMA_ARGS:
 895                         cmsg_groups |= 1;
 896                         retval = rds_rdma_extra_size(CMSG_DATA(cmsg));
 897                         if (retval < 0)
 898                                 return retval;
 899                         size += retval;
 900
 901                         break;
 902
 903                 case RDS_CMSG_RDMA_DEST:
 904                 case RDS_CMSG_RDMA_MAP:
 905                         cmsg_groups |= 2;
 906                         /* these are valid but do no add any size */
 907                         break;
 908
 909                 case RDS_CMSG_ATOMIC_CSWP:
 910                 case RDS_CMSG_ATOMIC_FADD:
 911                 case RDS_CMSG_MASKED_ATOMIC_CSWP:
 912                 case RDS_CMSG_MASKED_ATOMIC_FADD:
 913                         cmsg_groups |= 1;
 914                         size += sizeof(struct scatterlist);
 915                         break;
 916
 917                 default:
 918                         return -EINVAL;
 919                 }
 920
 921         }
 922
 923         size += ceil(data_len, PAGE_SIZE) * sizeof(struct scatterlist);
 924
 925         /* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */
 926         if (cmsg_groups == 3)
 927                 return -EINVAL;
 928
 929         return size;
 930 }
 931
 932 static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
 933                          struct msghdr *msg, int *allocated_mr)
 934 {
 935         struct cmsghdr *cmsg;
 936         int ret = 0;
 937
 938         for_each_cmsghdr(cmsg, msg) {
 939                 if (!CMSG_OK(msg, cmsg))
 940                         return -EINVAL;
 941
 942                 if (cmsg->cmsg_level != SOL_RDS)
 943                         continue;
 944
 945                 /* As a side effect, RDMA_DEST and RDMA_MAP will set
 946                  * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
 947                  */
 948                 switch (cmsg->cmsg_type) {
 949                 case RDS_CMSG_RDMA_ARGS:
 950                         ret = rds_cmsg_rdma_args(rs, rm, cmsg);
 951                         break;
 952
 953                 case RDS_CMSG_RDMA_DEST:
 954                         ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
 955                         break;
 956
 957                 case RDS_CMSG_RDMA_MAP:
 958                         ret = rds_cmsg_rdma_map(rs, rm, cmsg);
 959                         if (!ret)
 960                                 *allocated_mr = 1;
 961                         break;
 962                 case RDS_CMSG_ATOMIC_CSWP:
 963                 case RDS_CMSG_ATOMIC_FADD:
 964                 case RDS_CMSG_MASKED_ATOMIC_CSWP:
 965                 case RDS_CMSG_MASKED_ATOMIC_FADD:
 966                         ret = rds_cmsg_atomic(rs, rm, cmsg);
 967                         break;
 968
 969                 default:
 970                         return -EINVAL;
 971                 }
 972
 973                 if (ret)
 974                         break;
 975         }
 976
 977         return ret;
 978 }
 979
 980 int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len)
 981 {
 982         struct sock *sk = sock->sk;
 983         struct rds_sock *rs = rds_sk_to_rs(sk);
 984         DECLARE_SOCKADDR(struct sockaddr_in *, usin, msg->msg_name);
 985         __be32 daddr;
 986         __be16 dport;
 987         struct rds_message *rm = NULL;
 988         struct rds_connection *conn;
 989         int ret = 0;
 990         int queued = 0, allocated_mr = 0;
 991         int nonblock = msg->msg_flags & MSG_DONTWAIT;
 992         long timeo = sock_sndtimeo(sk, nonblock);
 993
 994         /* Mirror Linux UDP mirror of BSD error message compatibility */
 995         /* XXX: Perhaps MSG_MORE someday */
 996         if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT)) {
 997                 ret = -EOPNOTSUPP;
 998                 goto out;
 999         }
1000
1001         if (msg->msg_namelen) {
1002                 /* XXX fail non-unicast destination IPs? */
1003                 if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) {
1004                         ret = -EINVAL;
1005                         goto out;
1006                 }
1007                 daddr = usin->sin_addr.s_addr;
1008                 dport = usin->sin_port;
1009         } else {
1010                 /* We only care about consistency with ->connect() */
1011                 lock_sock(sk);
1012                 daddr = rs->rs_conn_addr;
1013                 dport = rs->rs_conn_port;
1014                 release_sock(sk);
1015         }
1016
1017         lock_sock(sk);
1018         if (daddr == 0 || rs->rs_bound_addr == 0) {
1019                 release_sock(sk);
1020                 ret = -ENOTCONN; /* XXX not a great errno */
1021                 goto out;
1022         }
1023         release_sock(sk);
1024
1025         if (payload_len > rds_sk_sndbuf(rs)) {
1026                 ret = -EMSGSIZE;
1027                 goto out;
1028         }
1029
1030         /* size of rm including all sgs */
1031         ret = rds_rm_size(msg, payload_len);
1032         if (ret < 0)
1033                 goto out;
1034
1035         rm = rds_message_alloc(ret, GFP_KERNEL);
1036         if (!rm) {
1037                 ret = -ENOMEM;
1038                 goto out;
1039         }
1040
1041         /* Attach data to the rm */
1042         if (payload_len) {
1043                 rm->data.op_sg = rds_message_alloc_sgs(rm, ceil(payload_len, PAGE_SIZE));
1044                 if (!rm->data.op_sg) {
1045                         ret = -ENOMEM;
1046                         goto out;
1047                 }
1048                 ret = rds_message_copy_from_user(rm, &msg->msg_iter);
1049                 if (ret)
1050                         goto out;
1051         }
1052         rm->data.op_active = 1;
1053
1054         rm->m_daddr = daddr;
1055
1056         /* rds_conn_create has a spinlock that runs with IRQ off.
1057          * Caching the conn in the socket helps a lot. */
1058         if (rs->rs_conn && rs->rs_conn->c_faddr == daddr)
1059                 conn = rs->rs_conn;
1060         else {
1061                 conn = rds_conn_create_outgoing(sock_net(sock->sk),
1062                                                 rs->rs_bound_addr, daddr,
1063                                         rs->rs_transport,
1064                                         sock->sk->sk_allocation);
1065                 if (IS_ERR(conn)) {
1066                         ret = PTR_ERR(conn);
1067                         goto out;
1068                 }
1069                 rs->rs_conn = conn;
1070         }
1071
1072         /* Parse any control messages the user may have included. */
1073         ret = rds_cmsg_send(rs, rm, msg, &allocated_mr);
1074         if (ret)
1075                 goto out;
1076
1077         if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) {
1078                 printk_ratelimited(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
1079                                &rm->rdma, conn->c_trans->xmit_rdma);
1080                 ret = -EOPNOTSUPP;
1081                 goto out;
1082         }
1083
1084         if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
1085                 printk_ratelimited(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
1086                                &rm->atomic, conn->c_trans->xmit_atomic);
1087                 ret = -EOPNOTSUPP;
1088                 goto out;
1089         }
1090
1091         rds_conn_connect_if_down(conn);
1092
1093         ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
1094         if (ret) {
1095                 rs->rs_seen_congestion = 1;
1096                 goto out;
1097         }
1098
1099         while (!rds_send_queue_rm(rs, conn, rm, rs->rs_bound_port,
1100                                   dport, &queued)) {
1101                 rds_stats_inc(s_send_queue_full);
1102
1103                 if (nonblock) {
1104                         ret = -EAGAIN;
1105                         goto out;
1106                 }
1107
1108                 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
1109                                         rds_send_queue_rm(rs, conn, rm,
1110                                                           rs->rs_bound_port,
1111                                                           dport,
1112                                                           &queued),
1113                                         timeo);
1114                 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
1115                 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
1116                         continue;
1117
1118                 ret = timeo;
1119                 if (ret == 0)
1120                         ret = -ETIMEDOUT;
1121                 goto out;
1122         }
1123
1124         /*
1125          * By now we've committed to the send.  We reuse rds_send_worker()
1126          * to retry sends in the rds thread if the transport asks us to.
1127          */
1128         rds_stats_inc(s_send_queued);
1129
1130         ret = rds_send_xmit(conn);
1131         if (ret == -ENOMEM || ret == -EAGAIN)
1132                 queue_delayed_work(rds_wq, &conn->c_send_w, 1);
1133
1134         rds_message_put(rm);
1135         return payload_len;
1136
1137 out:
1138         /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1139          * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1140          * or in any other way, we need to destroy the MR again */
1141         if (allocated_mr)
1142                 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
1143
1144         if (rm)
1145                 rds_message_put(rm);
1146         return ret;
1147 }
1148
1149 /*
1150  * Reply to a ping packet.
1151  */
1152 int
1153 rds_send_pong(struct rds_connection *conn, __be16 dport)
1154 {
1155         struct rds_message *rm;
1156         unsigned long flags;
1157         int ret = 0;
1158
1159         rm = rds_message_alloc(0, GFP_ATOMIC);
1160         if (!rm) {
1161                 ret = -ENOMEM;
1162                 goto out;
1163         }
1164
1165         rm->m_daddr = conn->c_faddr;
1166         rm->data.op_active = 1;
1167
1168         rds_conn_connect_if_down(conn);
1169
1170         ret = rds_cong_wait(conn->c_fcong, dport, 1, NULL);
1171         if (ret)
1172                 goto out;
1173
1174         spin_lock_irqsave(&conn->c_lock, flags);
1175         list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
1176         set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
1177         rds_message_addref(rm);
1178         rm->m_inc.i_conn = conn;
1179
1180         rds_message_populate_header(&rm->m_inc.i_hdr, 0, dport,
1181                                     conn->c_next_tx_seq);
1182         conn->c_next_tx_seq++;
1183         spin_unlock_irqrestore(&conn->c_lock, flags);
1184
1185         rds_stats_inc(s_send_queued);
1186         rds_stats_inc(s_send_pong);
1187
1188         /* schedule the send work on rds_wq */
1189         queue_delayed_work(rds_wq, &conn->c_send_w, 1);
1190
1191         rds_message_put(rm);
1192         return 0;
1193
1194 out:
1195         if (rm)
1196                 rds_message_put(rm);
1197         return ret;
1198 }