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