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
 103 static int acquire_in_xmit(struct rds_connection *conn)
 104 {
 105         return test_and_set_bit(RDS_IN_XMIT, &conn->c_flags) == 0;
 106 }
 107
 108 static void release_in_xmit(struct rds_connection *conn)
 109 {
 110         clear_bit(RDS_IN_XMIT, &conn->c_flags);
 111         smp_mb__after_atomic();
 112         /*
 113          * We don't use wait_on_bit()/wake_up_bit() because our waking is in a
 114          * hot path and finding waiters is very rare.  We don't want to walk
 115          * the system-wide hashed waitqueue buckets in the fast path only to
 116          * almost never find waiters.
 117          */
 118         if (waitqueue_active(&conn->c_waitq))
 119                 wake_up_all(&conn->c_waitq);
 120 }
 121
 122 /*
 123  * We're making the conscious trade-off here to only send one message
 124  * down the connection at a time.
 125  *   Pro:
 126  *      - tx queueing is a simple fifo list
 127  *      - reassembly is optional and easily done by transports per conn
 128  *      - no per flow rx lookup at all, straight to the socket
 129  *      - less per-frag memory and wire overhead
 130  *   Con:
 131  *      - queued acks can be delayed behind large messages
 132  *   Depends:
 133  *      - small message latency is higher behind queued large messages
 134  *      - large message latency isn't starved by intervening small sends
 135  */
 136 int rds_send_xmit(struct rds_connection *conn)
 137 {
 138         struct rds_message *rm;
 139         unsigned long flags;
 140         unsigned int tmp;
 141         struct scatterlist *sg;
 142         int ret = 0;
 143         LIST_HEAD(to_be_dropped);
 144         int batch_count;
 145         unsigned long send_gen = 0;
 146
 147 restart:
 148         batch_count = 0;
 149
 150         /*
 151          * sendmsg calls here after having queued its message on the send
 152          * queue.  We only have one task feeding the connection at a time.  If
 153          * another thread is already feeding the queue then we back off.  This
 154          * avoids blocking the caller and trading per-connection data between
 155          * caches per message.
 156          */
 157         if (!acquire_in_xmit(conn)) {
 158                 rds_stats_inc(s_send_lock_contention);
 159                 ret = -ENOMEM;
 160                 goto out;
 161         }
 162
 163         /*
 164          * we record the send generation after doing the xmit acquire.
 165          * if someone else manages to jump in and do some work, we'll use
 166          * this to avoid a goto restart farther down.
 167          *
 168          * The acquire_in_xmit() check above ensures that only one
 169          * caller can increment c_send_gen at any time.
 170          */
 171         conn->c_send_gen++;
 172         send_gen = conn->c_send_gen;
 173
 174         /*
 175          * rds_conn_shutdown() sets the conn state and then tests RDS_IN_XMIT,
 176          * we do the opposite to avoid races.
 177          */
 178         if (!rds_conn_up(conn)) {
 179                 release_in_xmit(conn);
 180                 ret = 0;
 181                 goto out;
 182         }
 183
 184         if (conn->c_trans->xmit_prepare)
 185                 conn->c_trans->xmit_prepare(conn);
 186
 187         /*
 188          * spin trying to push headers and data down the connection until
 189          * the connection doesn't make forward progress.
 190          */
 191         while (1) {
 192
 193                 rm = conn->c_xmit_rm;
 194
 195                 /*
 196                  * If between sending messages, we can send a pending congestion
 197                  * map update.
 198                  */
 199                 if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) {
 200                         rm = rds_cong_update_alloc(conn);
 201                         if (IS_ERR(rm)) {
 202                                 ret = PTR_ERR(rm);
 203                                 break;
 204                         }
 205                         rm->data.op_active = 1;
 206
 207                         conn->c_xmit_rm = rm;
 208                 }
 209
 210                 /*
 211                  * If not already working on one, grab the next message.
 212                  *
 213                  * c_xmit_rm holds a ref while we're sending this message down
 214                  * the connction.  We can use this ref while holding the
 215                  * send_sem.. rds_send_reset() is serialized with it.
 216                  */
 217                 if (!rm) {
 218                         unsigned int len;
 219
 220                         batch_count++;
 221
 222                         /* we want to process as big a batch as we can, but
 223                          * we also want to avoid softlockups.  If we've been
 224                          * through a lot of messages, lets back off and see
 225                          * if anyone else jumps in
 226                          */
 227                         if (batch_count >= send_batch_count)
 228                                 goto over_batch;
 229
 230                         spin_lock_irqsave(&conn->c_lock, flags);
 231
 232                         if (!list_empty(&conn->c_send_queue)) {
 233                                 rm = list_entry(conn->c_send_queue.next,
 234                                                 struct rds_message,
 235                                                 m_conn_item);
 236                                 rds_message_addref(rm);
 237
 238                                 /*
 239                                  * Move the message from the send queue to the retransmit
 240                                  * list right away.
 241                                  */
 242                                 list_move_tail(&rm->m_conn_item, &conn->c_retrans);
 243                         }
 244
 245                         spin_unlock_irqrestore(&conn->c_lock, flags);
 246
 247                         if (!rm)
 248                                 break;
 249
 250                         /* Unfortunately, the way Infiniband deals with
 251                          * RDMA to a bad MR key is by moving the entire
 252                          * queue pair to error state. We cold possibly
 253                          * recover from that, but right now we drop the
 254                          * connection.
 255                          * Therefore, we never retransmit messages with RDMA ops.
 256                          */
 257                         if (rm->rdma.op_active &&
 258                             test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags)) {
 259                                 spin_lock_irqsave(&conn->c_lock, flags);
 260                                 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags))
 261                                         list_move(&rm->m_conn_item, &to_be_dropped);
 262                                 spin_unlock_irqrestore(&conn->c_lock, flags);
 263                                 continue;
 264                         }
 265
 266                         /* Require an ACK every once in a while */
 267                         len = ntohl(rm->m_inc.i_hdr.h_len);
 268                         if (conn->c_unacked_packets == 0 ||
 269                             conn->c_unacked_bytes < len) {
 270                                 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
 271
 272                                 conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
 273                                 conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
 274                                 rds_stats_inc(s_send_ack_required);
 275                         } else {
 276                                 conn->c_unacked_bytes -= len;
 277                                 conn->c_unacked_packets--;
 278                         }
 279
 280                         conn->c_xmit_rm = rm;
 281                 }
 282
 283                 /* The transport either sends the whole rdma or none of it */
 284                 if (rm->rdma.op_active && !conn->c_xmit_rdma_sent) {
 285                         rm->m_final_op = &rm->rdma;
 286                         /* The transport owns the mapped memory for now.
 287                          * You can't unmap it while it's on the send queue
 288                          */
 289                         set_bit(RDS_MSG_MAPPED, &rm->m_flags);
 290                         ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
 291                         if (ret) {
 292                                 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
 293                                 wake_up_interruptible(&rm->m_flush_wait);
 294                                 break;
 295                         }
 296                         conn->c_xmit_rdma_sent = 1;
 297
 298                 }
 299
 300                 if (rm->atomic.op_active && !conn->c_xmit_atomic_sent) {
 301                         rm->m_final_op = &rm->atomic;
 302                         /* The transport owns the mapped memory for now.
 303                          * You can't unmap it while it's on the send queue
 304                          */
 305                         set_bit(RDS_MSG_MAPPED, &rm->m_flags);
 306                         ret = conn->c_trans->xmit_atomic(conn, &rm->atomic);
 307                         if (ret) {
 308                                 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
 309                                 wake_up_interruptible(&rm->m_flush_wait);
 310                                 break;
 311                         }
 312                         conn->c_xmit_atomic_sent = 1;
 313
 314                 }
 315
 316                 /*
 317                  * A number of cases require an RDS header to be sent
 318                  * even if there is no data.
 319                  * We permit 0-byte sends; rds-ping depends on this.
 320                  * However, if there are exclusively attached silent ops,
 321                  * we skip the hdr/data send, to enable silent operation.
 322                  */
 323                 if (rm->data.op_nents == 0) {
 324                         int ops_present;
 325                         int all_ops_are_silent = 1;
 326
 327                         ops_present = (rm->atomic.op_active || rm->rdma.op_active);
 328                         if (rm->atomic.op_active && !rm->atomic.op_silent)
 329                                 all_ops_are_silent = 0;
 330                         if (rm->rdma.op_active && !rm->rdma.op_silent)
 331                                 all_ops_are_silent = 0;
 332
 333                         if (ops_present && all_ops_are_silent
 334                             && !rm->m_rdma_cookie)
 335                                 rm->data.op_active = 0;
 336                 }
 337
 338                 if (rm->data.op_active && !conn->c_xmit_data_sent) {
 339                         rm->m_final_op = &rm->data;
 340                         ret = conn->c_trans->xmit(conn, rm,
 341                                                   conn->c_xmit_hdr_off,
 342                                                   conn->c_xmit_sg,
 343                                                   conn->c_xmit_data_off);
 344                         if (ret <= 0)
 345                                 break;
 346
 347                         if (conn->c_xmit_hdr_off < sizeof(struct rds_header)) {
 348                                 tmp = min_t(int, ret,
 349                                             sizeof(struct rds_header) -
 350                                             conn->c_xmit_hdr_off);
 351                                 conn->c_xmit_hdr_off += tmp;
 352                                 ret -= tmp;
 353                         }
 354
 355                         sg = &rm->data.op_sg[conn->c_xmit_sg];
 356                         while (ret) {
 357                                 tmp = min_t(int, ret, sg->length -
 358                                                       conn->c_xmit_data_off);
 359                                 conn->c_xmit_data_off += tmp;
 360                                 ret -= tmp;
 361                                 if (conn->c_xmit_data_off == sg->length) {
 362                                         conn->c_xmit_data_off = 0;
 363                                         sg++;
 364                                         conn->c_xmit_sg++;
 365                                         BUG_ON(ret != 0 &&
 366                                                conn->c_xmit_sg == rm->data.op_nents);
 367                                 }
 368                         }
 369
 370                         if (conn->c_xmit_hdr_off == sizeof(struct rds_header) &&
 371                             (conn->c_xmit_sg == rm->data.op_nents))
 372                                 conn->c_xmit_data_sent = 1;
 373                 }
 374
 375                 /*
 376                  * A rm will only take multiple times through this loop
 377                  * if there is a data op. Thus, if the data is sent (or there was
 378                  * none), then we're done with the rm.
 379                  */
 380                 if (!rm->data.op_active || conn->c_xmit_data_sent) {
 381                         conn->c_xmit_rm = NULL;
 382                         conn->c_xmit_sg = 0;
 383                         conn->c_xmit_hdr_off = 0;
 384                         conn->c_xmit_data_off = 0;
 385                         conn->c_xmit_rdma_sent = 0;
 386                         conn->c_xmit_atomic_sent = 0;
 387                         conn->c_xmit_data_sent = 0;
 388
 389                         rds_message_put(rm);
 390                 }
 391         }
 392
 393 over_batch:
 394         if (conn->c_trans->xmit_complete)
 395                 conn->c_trans->xmit_complete(conn);
 396         release_in_xmit(conn);
 397
 398         /* Nuke any messages we decided not to retransmit. */
 399         if (!list_empty(&to_be_dropped)) {
 400                 /* irqs on here, so we can put(), unlike above */
 401                 list_for_each_entry(rm, &to_be_dropped, m_conn_item)
 402                         rds_message_put(rm);
 403                 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
 404         }
 405
 406         /*
 407          * Other senders can queue a message after we last test the send queue
 408          * but before we clear RDS_IN_XMIT.  In that case they'd back off and
 409          * not try and send their newly queued message.  We need to check the
 410          * send queue after having cleared RDS_IN_XMIT so that their message
 411          * doesn't get stuck on the send queue.
 412          *
 413          * If the transport cannot continue (i.e ret != 0), then it must
 414          * call us when more room is available, such as from the tx
 415          * completion handler.
 416          *
 417          * We have an extra generation check here so that if someone manages
 418          * to jump in after our release_in_xmit, we'll see that they have done
 419          * some work and we will skip our goto
 420          */
 421         if (ret == 0) {
 422                 smp_mb();
 423                 if ((test_bit(0, &conn->c_map_queued) ||
 424                      !list_empty(&conn->c_send_queue)) &&
 425                     send_gen == conn->c_send_gen) {
 426                         rds_stats_inc(s_send_lock_queue_raced);
 427                         if (batch_count < send_batch_count)
 428                                 goto restart;
 429                         queue_delayed_work(rds_wq, &conn->c_send_w, 1);
 430                 }
 431         }
 432 out:
 433         return ret;
 434 }
 435 EXPORT_SYMBOL_GPL(rds_send_xmit);
 436
 437 static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
 438 {
 439         u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
 440
 441         assert_spin_locked(&rs->rs_lock);
 442
 443         BUG_ON(rs->rs_snd_bytes < len);
 444         rs->rs_snd_bytes -= len;
 445
 446         if (rs->rs_snd_bytes == 0)
 447                 rds_stats_inc(s_send_queue_empty);
 448 }
 449
 450 static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
 451                                     is_acked_func is_acked)
 452 {
 453         if (is_acked)
 454                 return is_acked(rm, ack);
 455         return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
 456 }
 457
 458 /*
 459  * This is pretty similar to what happens below in the ACK
 460  * handling code - except that we call here as soon as we get
 461  * the IB send completion on the RDMA op and the accompanying
 462  * message.
 463  */
 464 void rds_rdma_send_complete(struct rds_message *rm, int status)
 465 {
 466         struct rds_sock *rs = NULL;
 467         struct rm_rdma_op *ro;
 468         struct rds_notifier *notifier;
 469         unsigned long flags;
 470         unsigned int notify = 0;
 471
 472         spin_lock_irqsave(&rm->m_rs_lock, flags);
 473
 474         notify =  rm->rdma.op_notify | rm->data.op_notify;
 475         ro = &rm->rdma;
 476         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
 477             ro->op_active && notify && ro->op_notifier) {
 478                 notifier = ro->op_notifier;
 479                 rs = rm->m_rs;
 480                 sock_hold(rds_rs_to_sk(rs));
 481
 482                 notifier->n_status = status;
 483                 spin_lock(&rs->rs_lock);
 484                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
 485                 spin_unlock(&rs->rs_lock);
 486
 487                 ro->op_notifier = NULL;
 488         }
 489
 490         spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 491
 492         if (rs) {
 493                 rds_wake_sk_sleep(rs);
 494                 sock_put(rds_rs_to_sk(rs));
 495         }
 496 }
 497 EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
 498
 499 /*
 500  * Just like above, except looks at atomic op
 501  */
 502 void rds_atomic_send_complete(struct rds_message *rm, int status)
 503 {
 504         struct rds_sock *rs = NULL;
 505         struct rm_atomic_op *ao;
 506         struct rds_notifier *notifier;
 507         unsigned long flags;
 508
 509         spin_lock_irqsave(&rm->m_rs_lock, flags);
 510
 511         ao = &rm->atomic;
 512         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
 513             && ao->op_active && ao->op_notify && ao->op_notifier) {
 514                 notifier = ao->op_notifier;
 515                 rs = rm->m_rs;
 516                 sock_hold(rds_rs_to_sk(rs));
 517
 518                 notifier->n_status = status;
 519                 spin_lock(&rs->rs_lock);
 520                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
 521                 spin_unlock(&rs->rs_lock);
 522
 523                 ao->op_notifier = NULL;
 524         }
 525
 526         spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 527
 528         if (rs) {
 529                 rds_wake_sk_sleep(rs);
 530                 sock_put(rds_rs_to_sk(rs));
 531         }
 532 }
 533 EXPORT_SYMBOL_GPL(rds_atomic_send_complete);
 534
 535 /*
 536  * This is the same as rds_rdma_send_complete except we
 537  * don't do any locking - we have all the ingredients (message,
 538  * socket, socket lock) and can just move the notifier.
 539  */
 540 static inline void
 541 __rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
 542 {
 543         struct rm_rdma_op *ro;
 544         struct rm_atomic_op *ao;
 545
 546         ro = &rm->rdma;
 547         if (ro->op_active && ro->op_notify && ro->op_notifier) {
 548                 ro->op_notifier->n_status = status;
 549                 list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue);
 550                 ro->op_notifier = NULL;
 551         }
 552
 553         ao = &rm->atomic;
 554         if (ao->op_active && ao->op_notify && ao->op_notifier) {
 555                 ao->op_notifier->n_status = status;
 556                 list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue);
 557                 ao->op_notifier = NULL;
 558         }
 559
 560         /* No need to wake the app - caller does this */
 561 }
 562
 563 /*
 564  * This is called from the IB send completion when we detect
 565  * a RDMA operation that failed with remote access error.
 566  * So speed is not an issue here.
 567  */
 568 struct rds_message *rds_send_get_message(struct rds_connection *conn,
 569                                          struct rm_rdma_op *op)
 570 {
 571         struct rds_message *rm, *tmp, *found = NULL;
 572         unsigned long flags;
 573
 574         spin_lock_irqsave(&conn->c_lock, flags);
 575
 576         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
 577                 if (&rm->rdma == op) {
 578                         atomic_inc(&rm->m_refcount);
 579                         found = rm;
 580                         goto out;
 581                 }
 582         }
 583
 584         list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
 585                 if (&rm->rdma == op) {
 586                         atomic_inc(&rm->m_refcount);
 587                         found = rm;
 588                         break;
 589                 }
 590         }
 591
 592 out:
 593         spin_unlock_irqrestore(&conn->c_lock, flags);
 594
 595         return found;
 596 }
 597 EXPORT_SYMBOL_GPL(rds_send_get_message);
 598
 599 /*
 600  * This removes messages from the socket's list if they're on it.  The list
 601  * argument must be private to the caller, we must be able to modify it
 602  * without locks.  The messages must have a reference held for their
 603  * position on the list.  This function will drop that reference after
 604  * removing the messages from the 'messages' list regardless of if it found
 605  * the messages on the socket list or not.
 606  */
 607 static void rds_send_remove_from_sock(struct list_head *messages, int status)
 608 {
 609         unsigned long flags;
 610         struct rds_sock *rs = NULL;
 611         struct rds_message *rm;
 612
 613         while (!list_empty(messages)) {
 614                 int was_on_sock = 0;
 615
 616                 rm = list_entry(messages->next, struct rds_message,
 617                                 m_conn_item);
 618                 list_del_init(&rm->m_conn_item);
 619
 620                 /*
 621                  * If we see this flag cleared then we're *sure* that someone
 622                  * else beat us to removing it from the sock.  If we race
 623                  * with their flag update we'll get the lock and then really
 624                  * see that the flag has been cleared.
 625                  *
 626                  * The message spinlock makes sure nobody clears rm->m_rs
 627                  * while we're messing with it. It does not prevent the
 628                  * message from being removed from the socket, though.
 629                  */
 630                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 631                 if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags))
 632                         goto unlock_and_drop;
 633
 634                 if (rs != rm->m_rs) {
 635                         if (rs) {
 636                                 rds_wake_sk_sleep(rs);
 637                                 sock_put(rds_rs_to_sk(rs));
 638                         }
 639                         rs = rm->m_rs;
 640                         if (rs)
 641                                 sock_hold(rds_rs_to_sk(rs));
 642                 }
 643                 if (!rs)
 644                         goto unlock_and_drop;
 645                 spin_lock(&rs->rs_lock);
 646
 647                 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
 648                         struct rm_rdma_op *ro = &rm->rdma;
 649                         struct rds_notifier *notifier;
 650
 651                         list_del_init(&rm->m_sock_item);
 652                         rds_send_sndbuf_remove(rs, rm);
 653
 654                         if (ro->op_active && ro->op_notifier &&
 655                                (ro->op_notify || (ro->op_recverr && status))) {
 656                                 notifier = ro->op_notifier;
 657                                 list_add_tail(&notifier->n_list,
 658                                                 &rs->rs_notify_queue);
 659                                 if (!notifier->n_status)
 660                                         notifier->n_status = status;
 661                                 rm->rdma.op_notifier = NULL;
 662                         }
 663                         was_on_sock = 1;
 664                         rm->m_rs = NULL;
 665                 }
 666                 spin_unlock(&rs->rs_lock);
 667
 668 unlock_and_drop:
 669                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 670                 rds_message_put(rm);
 671                 if (was_on_sock)
 672                         rds_message_put(rm);
 673         }
 674
 675         if (rs) {
 676                 rds_wake_sk_sleep(rs);
 677                 sock_put(rds_rs_to_sk(rs));
 678         }
 679 }
 680
 681 /*
 682  * Transports call here when they've determined that the receiver queued
 683  * messages up to, and including, the given sequence number.  Messages are
 684  * moved to the retrans queue when rds_send_xmit picks them off the send
 685  * queue. This means that in the TCP case, the message may not have been
 686  * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
 687  * checks the RDS_MSG_HAS_ACK_SEQ bit.
 688  */
 689 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
 690                          is_acked_func is_acked)
 691 {
 692         struct rds_message *rm, *tmp;
 693         unsigned long flags;
 694         LIST_HEAD(list);
 695
 696         spin_lock_irqsave(&conn->c_lock, flags);
 697
 698         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
 699                 if (!rds_send_is_acked(rm, ack, is_acked))
 700                         break;
 701
 702                 list_move(&rm->m_conn_item, &list);
 703                 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
 704         }
 705
 706         /* order flag updates with spin locks */
 707         if (!list_empty(&list))
 708                 smp_mb__after_atomic();
 709
 710         spin_unlock_irqrestore(&conn->c_lock, flags);
 711
 712         /* now remove the messages from the sock list as needed */
 713         rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
 714 }
 715 EXPORT_SYMBOL_GPL(rds_send_drop_acked);
 716
 717 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
 718 {
 719         struct rds_message *rm, *tmp;
 720         struct rds_connection *conn;
 721         unsigned long flags;
 722         LIST_HEAD(list);
 723
 724         /* get all the messages we're dropping under the rs lock */
 725         spin_lock_irqsave(&rs->rs_lock, flags);
 726
 727         list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) {
 728                 if (dest && (dest->sin_addr.s_addr != rm->m_daddr ||
 729                              dest->sin_port != rm->m_inc.i_hdr.h_dport))
 730                         continue;
 731
 732                 list_move(&rm->m_sock_item, &list);
 733                 rds_send_sndbuf_remove(rs, rm);
 734                 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
 735         }
 736
 737         /* order flag updates with the rs lock */
 738         smp_mb__after_atomic();
 739
 740         spin_unlock_irqrestore(&rs->rs_lock, flags);
 741
 742         if (list_empty(&list))
 743                 return;
 744
 745         /* Remove the messages from the conn */
 746         list_for_each_entry(rm, &list, m_sock_item) {
 747
 748                 conn = rm->m_inc.i_conn;
 749
 750                 spin_lock_irqsave(&conn->c_lock, flags);
 751                 /*
 752                  * Maybe someone else beat us to removing rm from the conn.
 753                  * If we race with their flag update we'll get the lock and
 754                  * then really see that the flag has been cleared.
 755                  */
 756                 if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
 757                         spin_unlock_irqrestore(&conn->c_lock, flags);
 758                         spin_lock_irqsave(&rm->m_rs_lock, flags);
 759                         rm->m_rs = NULL;
 760                         spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 761                         continue;
 762                 }
 763                 list_del_init(&rm->m_conn_item);
 764                 spin_unlock_irqrestore(&conn->c_lock, flags);
 765
 766                 /*
 767                  * Couldn't grab m_rs_lock in top loop (lock ordering),
 768                  * but we can now.
 769                  */
 770                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 771
 772                 spin_lock(&rs->rs_lock);
 773                 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
 774                 spin_unlock(&rs->rs_lock);
 775
 776                 rm->m_rs = NULL;
 777                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 778
 779                 rds_message_put(rm);
 780         }
 781
 782         rds_wake_sk_sleep(rs);
 783
 784         while (!list_empty(&list)) {
 785                 rm = list_entry(list.next, struct rds_message, m_sock_item);
 786                 list_del_init(&rm->m_sock_item);
 787                 rds_message_wait(rm);
 788
 789                 /* just in case the code above skipped this message
 790                  * because RDS_MSG_ON_CONN wasn't set, run it again here
 791                  * taking m_rs_lock is the only thing that keeps us
 792                  * from racing with ack processing.
 793                  */
 794                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 795
 796                 spin_lock(&rs->rs_lock);
 797                 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
 798                 spin_unlock(&rs->rs_lock);
 799
 800                 rm->m_rs = NULL;
 801                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 802
 803                 rds_message_put(rm);
 804         }
 805 }
 806
 807 /*
 808  * we only want this to fire once so we use the callers 'queued'.  It's
 809  * possible that another thread can race with us and remove the
 810  * message from the flow with RDS_CANCEL_SENT_TO.
 811  */
 812 static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
 813                              struct rds_message *rm, __be16 sport,
 814                              __be16 dport, int *queued)
 815 {
 816         unsigned long flags;
 817         u32 len;
 818
 819         if (*queued)
 820                 goto out;
 821
 822         len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
 823
 824         /* this is the only place which holds both the socket's rs_lock
 825          * and the connection's c_lock */
 826         spin_lock_irqsave(&rs->rs_lock, flags);
 827
 828         /*
 829          * If there is a little space in sndbuf, we don't queue anything,
 830          * and userspace gets -EAGAIN. But poll() indicates there's send
 831          * room. This can lead to bad behavior (spinning) if snd_bytes isn't
 832          * freed up by incoming acks. So we check the *old* value of
 833          * rs_snd_bytes here to allow the last msg to exceed the buffer,
 834          * and poll() now knows no more data can be sent.
 835          */
 836         if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
 837                 rs->rs_snd_bytes += len;
 838
 839                 /* let recv side know we are close to send space exhaustion.
 840                  * This is probably not the optimal way to do it, as this
 841                  * means we set the flag on *all* messages as soon as our
 842                  * throughput hits a certain threshold.
 843                  */
 844                 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
 845                         __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
 846
 847                 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
 848                 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
 849                 rds_message_addref(rm);
 850                 rm->m_rs = rs;
 851
 852                 /* The code ordering is a little weird, but we're
 853                    trying to minimize the time we hold c_lock */
 854                 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
 855                 rm->m_inc.i_conn = conn;
 856                 rds_message_addref(rm);
 857
 858                 spin_lock(&conn->c_lock);
 859                 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(conn->c_next_tx_seq++);
 860                 list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
 861                 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
 862                 spin_unlock(&conn->c_lock);
 863
 864                 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
 865                          rm, len, rs, rs->rs_snd_bytes,
 866                          (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
 867
 868                 *queued = 1;
 869         }
 870
 871         spin_unlock_irqrestore(&rs->rs_lock, flags);
 872 out:
 873         return *queued;
 874 }
 875
 876 /*
 877  * rds_message is getting to be quite complicated, and we'd like to allocate
 878  * it all in one go. This figures out how big it needs to be up front.
 879  */
 880 static int rds_rm_size(struct msghdr *msg, int data_len)
 881 {
 882         struct cmsghdr *cmsg;
 883         int size = 0;
 884         int cmsg_groups = 0;
 885         int retval;
 886
 887         for_each_cmsghdr(cmsg, msg) {
 888                 if (!CMSG_OK(msg, cmsg))
 889                         return -EINVAL;
 890
 891                 if (cmsg->cmsg_level != SOL_RDS)
 892                         continue;
 893
 894                 switch (cmsg->cmsg_type) {
 895                 case RDS_CMSG_RDMA_ARGS:
 896                         cmsg_groups |= 1;
 897                         retval = rds_rdma_extra_size(CMSG_DATA(cmsg));
 898                         if (retval < 0)
 899                                 return retval;
 900                         size += retval;
 901
 902                         break;
 903
 904                 case RDS_CMSG_RDMA_DEST:
 905                 case RDS_CMSG_RDMA_MAP:
 906                         cmsg_groups |= 2;
 907                         /* these are valid but do no add any size */
 908                         break;
 909
 910                 case RDS_CMSG_ATOMIC_CSWP:
 911                 case RDS_CMSG_ATOMIC_FADD:
 912                 case RDS_CMSG_MASKED_ATOMIC_CSWP:
 913                 case RDS_CMSG_MASKED_ATOMIC_FADD:
 914                         cmsg_groups |= 1;
 915                         size += sizeof(struct scatterlist);
 916                         break;
 917
 918                 default:
 919                         return -EINVAL;
 920                 }
 921
 922         }
 923
 924         size += ceil(data_len, PAGE_SIZE) * sizeof(struct scatterlist);
 925
 926         /* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */
 927         if (cmsg_groups == 3)
 928                 return -EINVAL;
 929
 930         return size;
 931 }
 932
 933 static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
 934                          struct msghdr *msg, int *allocated_mr)
 935 {
 936         struct cmsghdr *cmsg;
 937         int ret = 0;
 938
 939         for_each_cmsghdr(cmsg, msg) {
 940                 if (!CMSG_OK(msg, cmsg))
 941                         return -EINVAL;
 942
 943                 if (cmsg->cmsg_level != SOL_RDS)
 944                         continue;
 945
 946                 /* As a side effect, RDMA_DEST and RDMA_MAP will set
 947                  * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
 948                  */
 949                 switch (cmsg->cmsg_type) {
 950                 case RDS_CMSG_RDMA_ARGS:
 951                         ret = rds_cmsg_rdma_args(rs, rm, cmsg);
 952                         break;
 953
 954                 case RDS_CMSG_RDMA_DEST:
 955                         ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
 956                         break;
 957
 958                 case RDS_CMSG_RDMA_MAP:
 959                         ret = rds_cmsg_rdma_map(rs, rm, cmsg);
 960                         if (!ret)
 961                                 *allocated_mr = 1;
 962                         else if (ret == -ENODEV)
 963                                 /* Accommodate the get_mr() case which can fail
 964                                  * if connection isn't established yet.
 965                                  */
 966                                 ret = -EAGAIN;
 967                         break;
 968                 case RDS_CMSG_ATOMIC_CSWP:
 969                 case RDS_CMSG_ATOMIC_FADD:
 970                 case RDS_CMSG_MASKED_ATOMIC_CSWP:
 971                 case RDS_CMSG_MASKED_ATOMIC_FADD:
 972                         ret = rds_cmsg_atomic(rs, rm, cmsg);
 973                         break;
 974
 975                 default:
 976                         return -EINVAL;
 977                 }
 978
 979                 if (ret)
 980                         break;
 981         }
 982
 983         return ret;
 984 }
 985
 986 int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len)
 987 {
 988         struct sock *sk = sock->sk;
 989         struct rds_sock *rs = rds_sk_to_rs(sk);
 990         DECLARE_SOCKADDR(struct sockaddr_in *, usin, msg->msg_name);
 991         __be32 daddr;
 992         __be16 dport;
 993         struct rds_message *rm = NULL;
 994         struct rds_connection *conn;
 995         int ret = 0;
 996         int queued = 0, allocated_mr = 0;
 997         int nonblock = msg->msg_flags & MSG_DONTWAIT;
 998         long timeo = sock_sndtimeo(sk, nonblock);
 999
1000         /* Mirror Linux UDP mirror of BSD error message compatibility */
1001         /* XXX: Perhaps MSG_MORE someday */
1002         if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT)) {
1003                 ret = -EOPNOTSUPP;
1004                 goto out;
1005         }
1006
1007         if (msg->msg_namelen) {
1008                 /* XXX fail non-unicast destination IPs? */
1009                 if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) {
1010                         ret = -EINVAL;
1011                         goto out;
1012                 }
1013                 daddr = usin->sin_addr.s_addr;
1014                 dport = usin->sin_port;
1015         } else {
1016                 /* We only care about consistency with ->connect() */
1017                 lock_sock(sk);
1018                 daddr = rs->rs_conn_addr;
1019                 dport = rs->rs_conn_port;
1020                 release_sock(sk);
1021         }
1022
1023         lock_sock(sk);
1024         if (daddr == 0 || rs->rs_bound_addr == 0) {
1025                 release_sock(sk);
1026                 ret = -ENOTCONN; /* XXX not a great errno */
1027                 goto out;
1028         }
1029         release_sock(sk);
1030
1031         if (payload_len > rds_sk_sndbuf(rs)) {
1032                 ret = -EMSGSIZE;
1033                 goto out;
1034         }
1035
1036         /* size of rm including all sgs */
1037         ret = rds_rm_size(msg, payload_len);
1038         if (ret < 0)
1039                 goto out;
1040
1041         rm = rds_message_alloc(ret, GFP_KERNEL);
1042         if (!rm) {
1043                 ret = -ENOMEM;
1044                 goto out;
1045         }
1046
1047         /* Attach data to the rm */
1048         if (payload_len) {
1049                 rm->data.op_sg = rds_message_alloc_sgs(rm, ceil(payload_len, PAGE_SIZE));
1050                 if (!rm->data.op_sg) {
1051                         ret = -ENOMEM;
1052                         goto out;
1053                 }
1054                 ret = rds_message_copy_from_user(rm, &msg->msg_iter);
1055                 if (ret)
1056                         goto out;
1057         }
1058         rm->data.op_active = 1;
1059
1060         rm->m_daddr = daddr;
1061
1062         /* rds_conn_create has a spinlock that runs with IRQ off.
1063          * Caching the conn in the socket helps a lot. */
1064         if (rs->rs_conn && rs->rs_conn->c_faddr == daddr)
1065                 conn = rs->rs_conn;
1066         else {
1067                 conn = rds_conn_create_outgoing(sock_net(sock->sk),
1068                                                 rs->rs_bound_addr, daddr,
1069                                         rs->rs_transport,
1070                                         sock->sk->sk_allocation);
1071                 if (IS_ERR(conn)) {
1072                         ret = PTR_ERR(conn);
1073                         goto out;
1074                 }
1075                 rs->rs_conn = conn;
1076         }
1077
1078         /* Parse any control messages the user may have included. */
1079         ret = rds_cmsg_send(rs, rm, msg, &allocated_mr);
1080         if (ret) {
1081                 /* Trigger connection so that its ready for the next retry */
1082                 if (ret ==  -EAGAIN)
1083                         rds_conn_connect_if_down(conn);
1084                 goto out;
1085         }
1086
1087         if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) {
1088                 printk_ratelimited(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
1089                                &rm->rdma, conn->c_trans->xmit_rdma);
1090                 ret = -EOPNOTSUPP;
1091                 goto out;
1092         }
1093
1094         if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
1095                 printk_ratelimited(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
1096                                &rm->atomic, conn->c_trans->xmit_atomic);
1097                 ret = -EOPNOTSUPP;
1098                 goto out;
1099         }
1100
1101         rds_conn_connect_if_down(conn);
1102
1103         ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
1104         if (ret) {
1105                 rs->rs_seen_congestion = 1;
1106                 goto out;
1107         }
1108
1109         while (!rds_send_queue_rm(rs, conn, rm, rs->rs_bound_port,
1110                                   dport, &queued)) {
1111                 rds_stats_inc(s_send_queue_full);
1112
1113                 if (nonblock) {
1114                         ret = -EAGAIN;
1115                         goto out;
1116                 }
1117
1118                 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
1119                                         rds_send_queue_rm(rs, conn, rm,
1120                                                           rs->rs_bound_port,
1121                                                           dport,
1122                                                           &queued),
1123                                         timeo);
1124                 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
1125                 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
1126                         continue;
1127
1128                 ret = timeo;
1129                 if (ret == 0)
1130                         ret = -ETIMEDOUT;
1131                 goto out;
1132         }
1133
1134         /*
1135          * By now we've committed to the send.  We reuse rds_send_worker()
1136          * to retry sends in the rds thread if the transport asks us to.
1137          */
1138         rds_stats_inc(s_send_queued);
1139
1140         ret = rds_send_xmit(conn);
1141         if (ret == -ENOMEM || ret == -EAGAIN)
1142                 queue_delayed_work(rds_wq, &conn->c_send_w, 1);
1143
1144         rds_message_put(rm);
1145         return payload_len;
1146
1147 out:
1148         /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1149          * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1150          * or in any other way, we need to destroy the MR again */
1151         if (allocated_mr)
1152                 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
1153
1154         if (rm)
1155                 rds_message_put(rm);
1156         return ret;
1157 }
1158
1159 /*
1160  * Reply to a ping packet.
1161  */
1162 int
1163 rds_send_pong(struct rds_connection *conn, __be16 dport)
1164 {
1165         struct rds_message *rm;
1166         unsigned long flags;
1167         int ret = 0;
1168
1169         rm = rds_message_alloc(0, GFP_ATOMIC);
1170         if (!rm) {
1171                 ret = -ENOMEM;
1172                 goto out;
1173         }
1174
1175         rm->m_daddr = conn->c_faddr;
1176         rm->data.op_active = 1;
1177
1178         rds_conn_connect_if_down(conn);
1179
1180         ret = rds_cong_wait(conn->c_fcong, dport, 1, NULL);
1181         if (ret)
1182                 goto out;
1183
1184         spin_lock_irqsave(&conn->c_lock, flags);
1185         list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
1186         set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
1187         rds_message_addref(rm);
1188         rm->m_inc.i_conn = conn;
1189
1190         rds_message_populate_header(&rm->m_inc.i_hdr, 0, dport,
1191                                     conn->c_next_tx_seq);
1192         conn->c_next_tx_seq++;
1193         spin_unlock_irqrestore(&conn->c_lock, flags);
1194
1195         rds_stats_inc(s_send_queued);
1196         rds_stats_inc(s_send_pong);
1197
1198         /* schedule the send work on rds_wq */
1199         queue_delayed_work(rds_wq, &conn->c_send_w, 1);
1200
1201         rds_message_put(rm);
1202         return 0;
1203
1204 out:
1205         if (rm)
1206                 rds_message_put(rm);
1207         return ret;
1208 }