net/rds/send.c

   1 /*
   2  * Copyright (c) 2006 Oracle.  All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  *
  32  */
  33 #include <linux/kernel.h>
  34 #include <linux/gfp.h>
  35 #include <net/sock.h>
  36 #include <linux/in.h>
  37 #include <linux/list.h>
  38
  39 #include "rds.h"
  40
  41 /* When transmitting messages in rds_send_xmit, we need to emerge from
  42  * time to time and briefly release the CPU. Otherwise the softlock watchdog
  43  * will kick our shin.
  44  * Also, it seems fairer to not let one busy connection stall all the
  45  * others.
  46  *
  47  * send_batch_count is the number of times we'll loop in send_xmit. Setting
  48  * it to 0 will restore the old behavior (where we looped until we had
  49  * drained the queue).
  50  */
  51 static int send_batch_count = 64;
  52 module_param(send_batch_count, int, 0444);
  53 MODULE_PARM_DESC(send_batch_count, " batch factor when working the send queue");
  54
  55 /*
  56  * Reset the send state. Caller must hold c_send_lock when calling here.
  57  */
  58 void rds_send_reset(struct rds_connection *conn)
  59 {
  60         struct rds_message *rm, *tmp;
  61         unsigned long flags;
  62
  63         if (conn->c_xmit_rm) {
  64                 /* Tell the user the RDMA op is no longer mapped by the
  65                  * transport. This isn't entirely true (it's flushed out
  66                  * independently) but as the connection is down, there's
  67                  * no ongoing RDMA to/from that memory */
  68                 rds_message_unmapped(conn->c_xmit_rm);
  69                 rds_message_put(conn->c_xmit_rm);
  70                 conn->c_xmit_rm = NULL;
  71         }
  72         conn->c_xmit_sg = 0;
  73         conn->c_xmit_hdr_off = 0;
  74         conn->c_xmit_data_off = 0;
  75         conn->c_xmit_atomic_sent = 0;
  76         conn->c_xmit_rdma_sent = 0;
  77         conn->c_xmit_data_sent = 0;
  78
  79         conn->c_map_queued = 0;
  80
  81         conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
  82         conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
  83
  84         /* Mark messages as retransmissions, and move them to the send q */
  85         spin_lock_irqsave(&conn->c_lock, flags);
  86         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
  87                 set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
  88                 set_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags);
  89         }
  90         list_splice_init(&conn->c_retrans, &conn->c_send_queue);
  91         spin_unlock_irqrestore(&conn->c_lock, flags);
  92 }
  93
  94 /*
  95  * We're making the concious trade-off here to only send one message
  96  * down the connection at a time.
  97  *   Pro:
  98  *      - tx queueing is a simple fifo list
  99  *      - reassembly is optional and easily done by transports per conn
 100  *      - no per flow rx lookup at all, straight to the socket
 101  *      - less per-frag memory and wire overhead
 102  *   Con:
 103  *      - queued acks can be delayed behind large messages
 104  *   Depends:
 105  *      - small message latency is higher behind queued large messages
 106  *      - large message latency isn't starved by intervening small sends
 107  */
 108 int rds_send_xmit(struct rds_connection *conn)
 109 {
 110         struct rds_message *rm;
 111         unsigned long flags;
 112         unsigned int tmp;
 113         unsigned int send_quota = send_batch_count;
 114         struct scatterlist *sg;
 115         int ret = 0;
 116         int was_empty = 0;
 117         LIST_HEAD(to_be_dropped);
 118
 119         if (!rds_conn_up(conn))
 120                 goto out;
 121
 122         /*
 123          * sendmsg calls here after having queued its message on the send
 124          * queue.  We only have one task feeding the connection at a time.  If
 125          * another thread is already feeding the queue then we back off.  This
 126          * avoids blocking the caller and trading per-connection data between
 127          * caches per message.
 128          */
 129         if (!spin_trylock_irqsave(&conn->c_send_lock, flags)) {
 130                 rds_stats_inc(s_send_lock_contention);
 131                 ret = -ENOMEM;
 132                 goto out;
 133         }
 134
 135         if (conn->c_trans->xmit_prepare)
 136                 conn->c_trans->xmit_prepare(conn);
 137
 138         /*
 139          * spin trying to push headers and data down the connection until
 140          * the connection doesn't make forward progress.
 141          */
 142         while (--send_quota) {
 143
 144                 rm = conn->c_xmit_rm;
 145
 146                 /*
 147                  * If between sending messages, we can send a pending congestion
 148                  * map update.
 149                  */
 150                 if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) {
 151                         rm = rds_cong_update_alloc(conn);
 152                         if (IS_ERR(rm)) {
 153                                 ret = PTR_ERR(rm);
 154                                 break;
 155                         }
 156                         rm->data.op_active = 1;
 157
 158                         conn->c_xmit_rm = rm;
 159                 }
 160
 161                 /*
 162                  * If not already working on one, grab the next message.
 163                  *
 164                  * c_xmit_rm holds a ref while we're sending this message down
 165                  * the connction.  We can use this ref while holding the
 166                  * send_sem.. rds_send_reset() is serialized with it.
 167                  */
 168                 if (!rm) {
 169                         unsigned int len;
 170
 171                         spin_lock_irqsave(&conn->c_lock, flags);
 172
 173                         if (!list_empty(&conn->c_send_queue)) {
 174                                 rm = list_entry(conn->c_send_queue.next,
 175                                                 struct rds_message,
 176                                                 m_conn_item);
 177                                 rds_message_addref(rm);
 178
 179                                 /*
 180                                  * Move the message from the send queue to the retransmit
 181                                  * list right away.
 182                                  */
 183                                 list_move_tail(&rm->m_conn_item, &conn->c_retrans);
 184                         }
 185
 186                         spin_unlock_irqrestore(&conn->c_lock, flags);
 187
 188                         if (!rm) {
 189                                 was_empty = 1;
 190                                 break;
 191                         }
 192
 193                         /* Unfortunately, the way Infiniband deals with
 194                          * RDMA to a bad MR key is by moving the entire
 195                          * queue pair to error state. We cold possibly
 196                          * recover from that, but right now we drop the
 197                          * connection.
 198                          * Therefore, we never retransmit messages with RDMA ops.
 199                          */
 200                         if (rm->rdma.op_active &&
 201                             test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags)) {
 202                                 spin_lock_irqsave(&conn->c_lock, flags);
 203                                 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags))
 204                                         list_move(&rm->m_conn_item, &to_be_dropped);
 205                                 spin_unlock_irqrestore(&conn->c_lock, flags);
 206                                 rds_message_put(rm);
 207                                 continue;
 208                         }
 209
 210                         /* Require an ACK every once in a while */
 211                         len = ntohl(rm->m_inc.i_hdr.h_len);
 212                         if (conn->c_unacked_packets == 0 ||
 213                             conn->c_unacked_bytes < len) {
 214                                 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
 215
 216                                 conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
 217                                 conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
 218                                 rds_stats_inc(s_send_ack_required);
 219                         } else {
 220                                 conn->c_unacked_bytes -= len;
 221                                 conn->c_unacked_packets--;
 222                         }
 223
 224                         conn->c_xmit_rm = rm;
 225                 }
 226
 227                 /* The transport either sends the whole rdma or none of it */
 228                 if (rm->rdma.op_active && !conn->c_xmit_rdma_sent) {
 229                         rm->m_final_op = &rm->rdma;
 230                         ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
 231                         if (ret)
 232                                 break;
 233                         conn->c_xmit_rdma_sent = 1;
 234
 235                         /* The transport owns the mapped memory for now.
 236                          * You can't unmap it while it's on the send queue */
 237                         set_bit(RDS_MSG_MAPPED, &rm->m_flags);
 238                 }
 239
 240                 if (rm->atomic.op_active && !conn->c_xmit_atomic_sent) {
 241                         rm->m_final_op = &rm->atomic;
 242                         ret = conn->c_trans->xmit_atomic(conn, &rm->atomic);
 243                         if (ret)
 244                                 break;
 245                         conn->c_xmit_atomic_sent = 1;
 246
 247                         /* The transport owns the mapped memory for now.
 248                          * You can't unmap it while it's on the send queue */
 249                         set_bit(RDS_MSG_MAPPED, &rm->m_flags);
 250                 }
 251
 252                 /*
 253                  * A number of cases require an RDS header to be sent
 254                  * even if there is no data.
 255                  * We permit 0-byte sends; rds-ping depends on this.
 256                  * However, if there are exclusively attached silent ops,
 257                  * we skip the hdr/data send, to enable silent operation.
 258                  */
 259                 if (rm->data.op_nents == 0) {
 260                         int ops_present;
 261                         int all_ops_are_silent = 1;
 262
 263                         ops_present = (rm->atomic.op_active || rm->rdma.op_active);
 264                         if (rm->atomic.op_active && !rm->atomic.op_silent)
 265                                 all_ops_are_silent = 0;
 266                         if (rm->rdma.op_active && !rm->rdma.op_silent)
 267                                 all_ops_are_silent = 0;
 268
 269                         if (ops_present && all_ops_are_silent
 270                             && !rm->m_rdma_cookie)
 271                                 rm->data.op_active = 0;
 272                 }
 273
 274                 if (rm->data.op_active && !conn->c_xmit_data_sent) {
 275                         rm->m_final_op = &rm->data;
 276                         ret = conn->c_trans->xmit(conn, rm,
 277                                                   conn->c_xmit_hdr_off,
 278                                                   conn->c_xmit_sg,
 279                                                   conn->c_xmit_data_off);
 280                         if (ret <= 0)
 281                                 break;
 282
 283                         if (conn->c_xmit_hdr_off < sizeof(struct rds_header)) {
 284                                 tmp = min_t(int, ret,
 285                                             sizeof(struct rds_header) -
 286                                             conn->c_xmit_hdr_off);
 287                                 conn->c_xmit_hdr_off += tmp;
 288                                 ret -= tmp;
 289                         }
 290
 291                         sg = &rm->data.op_sg[conn->c_xmit_sg];
 292                         while (ret) {
 293                                 tmp = min_t(int, ret, sg->length -
 294                                                       conn->c_xmit_data_off);
 295                                 conn->c_xmit_data_off += tmp;
 296                                 ret -= tmp;
 297                                 if (conn->c_xmit_data_off == sg->length) {
 298                                         conn->c_xmit_data_off = 0;
 299                                         sg++;
 300                                         conn->c_xmit_sg++;
 301                                         BUG_ON(ret != 0 &&
 302                                                conn->c_xmit_sg == rm->data.op_nents);
 303                                 }
 304                         }
 305
 306                         if (conn->c_xmit_hdr_off == sizeof(struct rds_header) &&
 307                             (conn->c_xmit_sg == rm->data.op_nents))
 308                                 conn->c_xmit_data_sent = 1;
 309                 }
 310
 311                 /*
 312                  * A rm will only take multiple times through this loop
 313                  * if there is a data op. Thus, if the data is sent (or there was
 314                  * none), then we're done with the rm.
 315                  */
 316                 if (!rm->data.op_active || conn->c_xmit_data_sent) {
 317                         conn->c_xmit_rm = NULL;
 318                         conn->c_xmit_sg = 0;
 319                         conn->c_xmit_hdr_off = 0;
 320                         conn->c_xmit_data_off = 0;
 321                         conn->c_xmit_rdma_sent = 0;
 322                         conn->c_xmit_atomic_sent = 0;
 323                         conn->c_xmit_data_sent = 0;
 324
 325                         rds_message_put(rm);
 326                 }
 327         }
 328
 329         /* Nuke any messages we decided not to retransmit. */
 330         if (!list_empty(&to_be_dropped))
 331                 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
 332
 333         if (conn->c_trans->xmit_complete)
 334                 conn->c_trans->xmit_complete(conn);
 335
 336         /*
 337          * We might be racing with another sender who queued a message but
 338          * backed off on noticing that we held the c_send_lock.  If we check
 339          * for queued messages after dropping the sem then either we'll
 340          * see the queued message or the queuer will get the sem.  If we
 341          * notice the queued message then we trigger an immediate retry.
 342          *
 343          * We need to be careful only to do this when we stopped processing
 344          * the send queue because it was empty.  It's the only way we
 345          * stop processing the loop when the transport hasn't taken
 346          * responsibility for forward progress.
 347          */
 348         spin_unlock_irqrestore(&conn->c_send_lock, flags);
 349
 350         if (send_quota == 0 && !was_empty) {
 351                 /* We exhausted the send quota, but there's work left to
 352                  * do. Return and (re-)schedule the send worker.
 353                  */
 354                 ret = -EAGAIN;
 355         }
 356
 357         if (ret == 0 && was_empty) {
 358                 /* A simple bit test would be way faster than taking the
 359                  * spin lock */
 360                 spin_lock_irqsave(&conn->c_lock, flags);
 361                 if (!list_empty(&conn->c_send_queue)) {
 362                         rds_stats_inc(s_send_lock_queue_raced);
 363                         ret = -EAGAIN;
 364                 }
 365                 spin_unlock_irqrestore(&conn->c_lock, flags);
 366         }
 367 out:
 368         return ret;
 369 }
 370
 371 static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
 372 {
 373         u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
 374
 375         assert_spin_locked(&rs->rs_lock);
 376
 377         BUG_ON(rs->rs_snd_bytes < len);
 378         rs->rs_snd_bytes -= len;
 379
 380         if (rs->rs_snd_bytes == 0)
 381                 rds_stats_inc(s_send_queue_empty);
 382 }
 383
 384 static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
 385                                     is_acked_func is_acked)
 386 {
 387         if (is_acked)
 388                 return is_acked(rm, ack);
 389         return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
 390 }
 391
 392 /*
 393  * Returns true if there are no messages on the send and retransmit queues
 394  * which have a sequence number greater than or equal to the given sequence
 395  * number.
 396  */
 397 int rds_send_acked_before(struct rds_connection *conn, u64 seq)
 398 {
 399         struct rds_message *rm, *tmp;
 400         int ret = 1;
 401
 402         spin_lock(&conn->c_lock);
 403
 404         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
 405                 if (be64_to_cpu(rm->m_inc.i_hdr.h_sequence) < seq)
 406                         ret = 0;
 407                 break;
 408         }
 409
 410         list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
 411                 if (be64_to_cpu(rm->m_inc.i_hdr.h_sequence) < seq)
 412                         ret = 0;
 413                 break;
 414         }
 415
 416         spin_unlock(&conn->c_lock);
 417
 418         return ret;
 419 }
 420
 421 /*
 422  * This is pretty similar to what happens below in the ACK
 423  * handling code - except that we call here as soon as we get
 424  * the IB send completion on the RDMA op and the accompanying
 425  * message.
 426  */
 427 void rds_rdma_send_complete(struct rds_message *rm, int status)
 428 {
 429         struct rds_sock *rs = NULL;
 430         struct rm_rdma_op *ro;
 431         struct rds_notifier *notifier;
 432         unsigned long flags;
 433
 434         spin_lock_irqsave(&rm->m_rs_lock, flags);
 435
 436         ro = &rm->rdma;
 437         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
 438             ro->op_active && ro->op_notify && ro->op_notifier) {
 439                 notifier = ro->op_notifier;
 440                 rs = rm->m_rs;
 441                 sock_hold(rds_rs_to_sk(rs));
 442
 443                 notifier->n_status = status;
 444                 spin_lock(&rs->rs_lock);
 445                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
 446                 spin_unlock(&rs->rs_lock);
 447
 448                 ro->op_notifier = NULL;
 449         }
 450
 451         spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 452
 453         if (rs) {
 454                 rds_wake_sk_sleep(rs);
 455                 sock_put(rds_rs_to_sk(rs));
 456         }
 457 }
 458 EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
 459
 460 /*
 461  * Just like above, except looks at atomic op
 462  */
 463 void rds_atomic_send_complete(struct rds_message *rm, int status)
 464 {
 465         struct rds_sock *rs = NULL;
 466         struct rm_atomic_op *ao;
 467         struct rds_notifier *notifier;
 468
 469         spin_lock(&rm->m_rs_lock);
 470
 471         ao = &rm->atomic;
 472         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
 473             && ao->op_active && ao->op_notify && ao->op_notifier) {
 474                 notifier = ao->op_notifier;
 475                 rs = rm->m_rs;
 476                 sock_hold(rds_rs_to_sk(rs));
 477
 478                 notifier->n_status = status;
 479                 spin_lock(&rs->rs_lock);
 480                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
 481                 spin_unlock(&rs->rs_lock);
 482
 483                 ao->op_notifier = NULL;
 484         }
 485
 486         spin_unlock(&rm->m_rs_lock);
 487
 488         if (rs) {
 489                 rds_wake_sk_sleep(rs);
 490                 sock_put(rds_rs_to_sk(rs));
 491         }
 492 }
 493 EXPORT_SYMBOL_GPL(rds_atomic_send_complete);
 494
 495 /*
 496  * This is the same as rds_rdma_send_complete except we
 497  * don't do any locking - we have all the ingredients (message,
 498  * socket, socket lock) and can just move the notifier.
 499  */
 500 static inline void
 501 __rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
 502 {
 503         struct rm_rdma_op *ro;
 504         struct rm_atomic_op *ao;
 505
 506         ro = &rm->rdma;
 507         if (ro->op_active && ro->op_notify && ro->op_notifier) {
 508                 ro->op_notifier->n_status = status;
 509                 list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue);
 510                 ro->op_notifier = NULL;
 511         }
 512
 513         ao = &rm->atomic;
 514         if (ao->op_active && ao->op_notify && ao->op_notifier) {
 515                 ao->op_notifier->n_status = status;
 516                 list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue);
 517                 ao->op_notifier = NULL;
 518         }
 519
 520         /* No need to wake the app - caller does this */
 521 }
 522
 523 /*
 524  * This is called from the IB send completion when we detect
 525  * a RDMA operation that failed with remote access error.
 526  * So speed is not an issue here.
 527  */
 528 struct rds_message *rds_send_get_message(struct rds_connection *conn,
 529                                          struct rm_rdma_op *op)
 530 {
 531         struct rds_message *rm, *tmp, *found = NULL;
 532         unsigned long flags;
 533
 534         spin_lock_irqsave(&conn->c_lock, flags);
 535
 536         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
 537                 if (&rm->rdma == op) {
 538                         atomic_inc(&rm->m_refcount);
 539                         found = rm;
 540                         goto out;
 541                 }
 542         }
 543
 544         list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
 545                 if (&rm->rdma == op) {
 546                         atomic_inc(&rm->m_refcount);
 547                         found = rm;
 548                         break;
 549                 }
 550         }
 551
 552 out:
 553         spin_unlock_irqrestore(&conn->c_lock, flags);
 554
 555         return found;
 556 }
 557 EXPORT_SYMBOL_GPL(rds_send_get_message);
 558
 559 /*
 560  * This removes messages from the socket's list if they're on it.  The list
 561  * argument must be private to the caller, we must be able to modify it
 562  * without locks.  The messages must have a reference held for their
 563  * position on the list.  This function will drop that reference after
 564  * removing the messages from the 'messages' list regardless of if it found
 565  * the messages on the socket list or not.
 566  */
 567 void rds_send_remove_from_sock(struct list_head *messages, int status)
 568 {
 569         unsigned long flags;
 570         struct rds_sock *rs = NULL;
 571         struct rds_message *rm;
 572
 573         while (!list_empty(messages)) {
 574                 int was_on_sock = 0;
 575
 576                 rm = list_entry(messages->next, struct rds_message,
 577                                 m_conn_item);
 578                 list_del_init(&rm->m_conn_item);
 579
 580                 /*
 581                  * If we see this flag cleared then we're *sure* that someone
 582                  * else beat us to removing it from the sock.  If we race
 583                  * with their flag update we'll get the lock and then really
 584                  * see that the flag has been cleared.
 585                  *
 586                  * The message spinlock makes sure nobody clears rm->m_rs
 587                  * while we're messing with it. It does not prevent the
 588                  * message from being removed from the socket, though.
 589                  */
 590                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 591                 if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags))
 592                         goto unlock_and_drop;
 593
 594                 if (rs != rm->m_rs) {
 595                         if (rs) {
 596                                 rds_wake_sk_sleep(rs);
 597                                 sock_put(rds_rs_to_sk(rs));
 598                         }
 599                         rs = rm->m_rs;
 600                         sock_hold(rds_rs_to_sk(rs));
 601                 }
 602                 spin_lock(&rs->rs_lock);
 603
 604                 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
 605                         struct rm_rdma_op *ro = &rm->rdma;
 606                         struct rds_notifier *notifier;
 607
 608                         list_del_init(&rm->m_sock_item);
 609                         rds_send_sndbuf_remove(rs, rm);
 610
 611                         if (ro->op_active && ro->op_notifier &&
 612                                (ro->op_notify || (ro->op_recverr && status))) {
 613                                 notifier = ro->op_notifier;
 614                                 list_add_tail(&notifier->n_list,
 615                                                 &rs->rs_notify_queue);
 616                                 if (!notifier->n_status)
 617                                         notifier->n_status = status;
 618                                 rm->rdma.op_notifier = NULL;
 619                         }
 620                         was_on_sock = 1;
 621                         rm->m_rs = NULL;
 622                 }
 623                 spin_unlock(&rs->rs_lock);
 624
 625 unlock_and_drop:
 626                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 627                 rds_message_put(rm);
 628                 if (was_on_sock)
 629                         rds_message_put(rm);
 630         }
 631
 632         if (rs) {
 633                 rds_wake_sk_sleep(rs);
 634                 sock_put(rds_rs_to_sk(rs));
 635         }
 636 }
 637
 638 /*
 639  * Transports call here when they've determined that the receiver queued
 640  * messages up to, and including, the given sequence number.  Messages are
 641  * moved to the retrans queue when rds_send_xmit picks them off the send
 642  * queue. This means that in the TCP case, the message may not have been
 643  * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
 644  * checks the RDS_MSG_HAS_ACK_SEQ bit.
 645  *
 646  * XXX It's not clear to me how this is safely serialized with socket
 647  * destruction.  Maybe it should bail if it sees SOCK_DEAD.
 648  */
 649 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
 650                          is_acked_func is_acked)
 651 {
 652         struct rds_message *rm, *tmp;
 653         unsigned long flags;
 654         LIST_HEAD(list);
 655
 656         spin_lock_irqsave(&conn->c_lock, flags);
 657
 658         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
 659                 if (!rds_send_is_acked(rm, ack, is_acked))
 660                         break;
 661
 662                 list_move(&rm->m_conn_item, &list);
 663                 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
 664         }
 665
 666         /* order flag updates with spin locks */
 667         if (!list_empty(&list))
 668                 smp_mb__after_clear_bit();
 669
 670         spin_unlock_irqrestore(&conn->c_lock, flags);
 671
 672         /* now remove the messages from the sock list as needed */
 673         rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
 674 }
 675 EXPORT_SYMBOL_GPL(rds_send_drop_acked);
 676
 677 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
 678 {
 679         struct rds_message *rm, *tmp;
 680         struct rds_connection *conn;
 681         unsigned long flags;
 682         LIST_HEAD(list);
 683
 684         /* get all the messages we're dropping under the rs lock */
 685         spin_lock_irqsave(&rs->rs_lock, flags);
 686
 687         list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) {
 688                 if (dest && (dest->sin_addr.s_addr != rm->m_daddr ||
 689                              dest->sin_port != rm->m_inc.i_hdr.h_dport))
 690                         continue;
 691
 692                 list_move(&rm->m_sock_item, &list);
 693                 rds_send_sndbuf_remove(rs, rm);
 694                 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
 695         }
 696
 697         /* order flag updates with the rs lock */
 698         smp_mb__after_clear_bit();
 699
 700         spin_unlock_irqrestore(&rs->rs_lock, flags);
 701
 702         if (list_empty(&list))
 703                 return;
 704
 705         /* Remove the messages from the conn */
 706         list_for_each_entry(rm, &list, m_sock_item) {
 707
 708                 conn = rm->m_inc.i_conn;
 709
 710                 spin_lock_irqsave(&conn->c_lock, flags);
 711                 /*
 712                  * Maybe someone else beat us to removing rm from the conn.
 713                  * If we race with their flag update we'll get the lock and
 714                  * then really see that the flag has been cleared.
 715                  */
 716                 if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
 717                         spin_unlock_irqrestore(&conn->c_lock, flags);
 718                         continue;
 719                 }
 720                 list_del_init(&rm->m_conn_item);
 721                 spin_unlock_irqrestore(&conn->c_lock, flags);
 722
 723                 /*
 724                  * Couldn't grab m_rs_lock in top loop (lock ordering),
 725                  * but we can now.
 726                  */
 727                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 728
 729                 spin_lock(&rs->rs_lock);
 730                 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
 731                 spin_unlock(&rs->rs_lock);
 732
 733                 rm->m_rs = NULL;
 734                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 735
 736                 rds_message_put(rm);
 737         }
 738
 739         rds_wake_sk_sleep(rs);
 740
 741         while (!list_empty(&list)) {
 742                 rm = list_entry(list.next, struct rds_message, m_sock_item);
 743                 list_del_init(&rm->m_sock_item);
 744
 745                 rds_message_wait(rm);
 746                 rds_message_put(rm);
 747         }
 748 }
 749
 750 /*
 751  * we only want this to fire once so we use the callers 'queued'.  It's
 752  * possible that another thread can race with us and remove the
 753  * message from the flow with RDS_CANCEL_SENT_TO.
 754  */
 755 static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
 756                              struct rds_message *rm, __be16 sport,
 757                              __be16 dport, int *queued)
 758 {
 759         unsigned long flags;
 760         u32 len;
 761
 762         if (*queued)
 763                 goto out;
 764
 765         len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
 766
 767         /* this is the only place which holds both the socket's rs_lock
 768          * and the connection's c_lock */
 769         spin_lock_irqsave(&rs->rs_lock, flags);
 770
 771         /*
 772          * If there is a little space in sndbuf, we don't queue anything,
 773          * and userspace gets -EAGAIN. But poll() indicates there's send
 774          * room. This can lead to bad behavior (spinning) if snd_bytes isn't
 775          * freed up by incoming acks. So we check the *old* value of
 776          * rs_snd_bytes here to allow the last msg to exceed the buffer,
 777          * and poll() now knows no more data can be sent.
 778          */
 779         if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
 780                 rs->rs_snd_bytes += len;
 781
 782                 /* let recv side know we are close to send space exhaustion.
 783                  * This is probably not the optimal way to do it, as this
 784                  * means we set the flag on *all* messages as soon as our
 785                  * throughput hits a certain threshold.
 786                  */
 787                 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
 788                         __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
 789
 790                 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
 791                 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
 792                 rds_message_addref(rm);
 793                 rm->m_rs = rs;
 794
 795                 /* The code ordering is a little weird, but we're
 796                    trying to minimize the time we hold c_lock */
 797                 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
 798                 rm->m_inc.i_conn = conn;
 799                 rds_message_addref(rm);
 800
 801                 spin_lock(&conn->c_lock);
 802                 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(conn->c_next_tx_seq++);
 803                 list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
 804                 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
 805                 spin_unlock(&conn->c_lock);
 806
 807                 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
 808                          rm, len, rs, rs->rs_snd_bytes,
 809                          (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
 810
 811                 *queued = 1;
 812         }
 813
 814         spin_unlock_irqrestore(&rs->rs_lock, flags);
 815 out:
 816         return *queued;
 817 }
 818
 819 /*
 820  * rds_message is getting to be quite complicated, and we'd like to allocate
 821  * it all in one go. This figures out how big it needs to be up front.
 822  */
 823 static int rds_rm_size(struct msghdr *msg, int data_len)
 824 {
 825         struct cmsghdr *cmsg;
 826         int size = 0;
 827         int cmsg_groups = 0;
 828         int retval;
 829
 830         for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
 831                 if (!CMSG_OK(msg, cmsg))
 832                         return -EINVAL;
 833
 834                 if (cmsg->cmsg_level != SOL_RDS)
 835                         continue;
 836
 837                 switch (cmsg->cmsg_type) {
 838                 case RDS_CMSG_RDMA_ARGS:
 839                         cmsg_groups |= 1;
 840                         retval = rds_rdma_extra_size(CMSG_DATA(cmsg));
 841                         if (retval < 0)
 842                                 return retval;
 843                         size += retval;
 844
 845                         break;
 846
 847                 case RDS_CMSG_RDMA_DEST:
 848                 case RDS_CMSG_RDMA_MAP:
 849                         cmsg_groups |= 2;
 850                         /* these are valid but do no add any size */
 851                         break;
 852
 853                 case RDS_CMSG_ATOMIC_CSWP:
 854                 case RDS_CMSG_ATOMIC_FADD:
 855                         cmsg_groups |= 1;
 856                         size += sizeof(struct scatterlist);
 857                         break;
 858
 859                 default:
 860                         return -EINVAL;
 861                 }
 862
 863         }
 864
 865         size += ceil(data_len, PAGE_SIZE) * sizeof(struct scatterlist);
 866
 867         /* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */
 868         if (cmsg_groups == 3)
 869                 return -EINVAL;
 870
 871         return size;
 872 }
 873
 874 static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
 875                          struct msghdr *msg, int *allocated_mr)
 876 {
 877         struct cmsghdr *cmsg;
 878         int ret = 0;
 879
 880         for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
 881                 if (!CMSG_OK(msg, cmsg))
 882                         return -EINVAL;
 883
 884                 if (cmsg->cmsg_level != SOL_RDS)
 885                         continue;
 886
 887                 /* As a side effect, RDMA_DEST and RDMA_MAP will set
 888                  * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
 889                  */
 890                 switch (cmsg->cmsg_type) {
 891                 case RDS_CMSG_RDMA_ARGS:
 892                         ret = rds_cmsg_rdma_args(rs, rm, cmsg);
 893                         break;
 894
 895                 case RDS_CMSG_RDMA_DEST:
 896                         ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
 897                         break;
 898
 899                 case RDS_CMSG_RDMA_MAP:
 900                         ret = rds_cmsg_rdma_map(rs, rm, cmsg);
 901                         if (!ret)
 902                                 *allocated_mr = 1;
 903                         break;
 904                 case RDS_CMSG_ATOMIC_CSWP:
 905                 case RDS_CMSG_ATOMIC_FADD:
 906                         ret = rds_cmsg_atomic(rs, rm, cmsg);
 907                         break;
 908
 909                 default:
 910                         return -EINVAL;
 911                 }
 912
 913                 if (ret)
 914                         break;
 915         }
 916
 917         return ret;
 918 }
 919
 920 int rds_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
 921                 size_t payload_len)
 922 {
 923         struct sock *sk = sock->sk;
 924         struct rds_sock *rs = rds_sk_to_rs(sk);
 925         struct sockaddr_in *usin = (struct sockaddr_in *)msg->msg_name;
 926         __be32 daddr;
 927         __be16 dport;
 928         struct rds_message *rm = NULL;
 929         struct rds_connection *conn;
 930         int ret = 0;
 931         int queued = 0, allocated_mr = 0;
 932         int nonblock = msg->msg_flags & MSG_DONTWAIT;
 933         long timeo = sock_sndtimeo(sk, nonblock);
 934
 935         /* Mirror Linux UDP mirror of BSD error message compatibility */
 936         /* XXX: Perhaps MSG_MORE someday */
 937         if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT)) {
 938                 printk(KERN_INFO "msg_flags 0x%08X\n", msg->msg_flags);
 939                 ret = -EOPNOTSUPP;
 940                 goto out;
 941         }
 942
 943         if (msg->msg_namelen) {
 944                 /* XXX fail non-unicast destination IPs? */
 945                 if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) {
 946                         ret = -EINVAL;
 947                         goto out;
 948                 }
 949                 daddr = usin->sin_addr.s_addr;
 950                 dport = usin->sin_port;
 951         } else {
 952                 /* We only care about consistency with ->connect() */
 953                 lock_sock(sk);
 954                 daddr = rs->rs_conn_addr;
 955                 dport = rs->rs_conn_port;
 956                 release_sock(sk);
 957         }
 958
 959         /* racing with another thread binding seems ok here */
 960         if (daddr == 0 || rs->rs_bound_addr == 0) {
 961                 ret = -ENOTCONN; /* XXX not a great errno */
 962                 goto out;
 963         }
 964
 965         /* size of rm including all sgs */
 966         ret = rds_rm_size(msg, payload_len);
 967         if (ret < 0)
 968                 goto out;
 969
 970         rm = rds_message_alloc(ret, GFP_KERNEL);
 971         if (!rm) {
 972                 ret = -ENOMEM;
 973                 goto out;
 974         }
 975
 976         /* Attach data to the rm */
 977         if (payload_len) {
 978                 rm->data.op_sg = rds_message_alloc_sgs(rm, ceil(payload_len, PAGE_SIZE));
 979                 ret = rds_message_copy_from_user(rm, msg->msg_iov, payload_len);
 980                 if (ret)
 981                         goto out;
 982         }
 983         rm->data.op_active = 1;
 984
 985         rm->m_daddr = daddr;
 986
 987         /* rds_conn_create has a spinlock that runs with IRQ off.
 988          * Caching the conn in the socket helps a lot. */
 989         if (rs->rs_conn && rs->rs_conn->c_faddr == daddr)
 990                 conn = rs->rs_conn;
 991         else {
 992                 conn = rds_conn_create_outgoing(rs->rs_bound_addr, daddr,
 993                                         rs->rs_transport,
 994                                         sock->sk->sk_allocation);
 995                 if (IS_ERR(conn)) {
 996                         ret = PTR_ERR(conn);
 997                         goto out;
 998                 }
 999                 rs->rs_conn = conn;
1000         }
1001
1002         /* Parse any control messages the user may have included. */
1003         ret = rds_cmsg_send(rs, rm, msg, &allocated_mr);
1004         if (ret)
1005                 goto out;
1006
1007         if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) {
1008                 if (printk_ratelimit())
1009                         printk(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
1010                                &rm->rdma, conn->c_trans->xmit_rdma);
1011                 ret = -EOPNOTSUPP;
1012                 goto out;
1013         }
1014
1015         if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
1016                 if (printk_ratelimit())
1017                         printk(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
1018                                &rm->atomic, conn->c_trans->xmit_atomic);
1019                 ret = -EOPNOTSUPP;
1020                 goto out;
1021         }
1022
1023         /* If the connection is down, trigger a connect. We may
1024          * have scheduled a delayed reconnect however - in this case
1025          * we should not interfere.
1026          */
1027         if (rds_conn_state(conn) == RDS_CONN_DOWN &&
1028             !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
1029                 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
1030
1031         ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
1032         if (ret) {
1033                 rs->rs_seen_congestion = 1;
1034                 goto out;
1035         }
1036
1037         while (!rds_send_queue_rm(rs, conn, rm, rs->rs_bound_port,
1038                                   dport, &queued)) {
1039                 rds_stats_inc(s_send_queue_full);
1040                 /* XXX make sure this is reasonable */
1041                 if (payload_len > rds_sk_sndbuf(rs)) {
1042                         ret = -EMSGSIZE;
1043                         goto out;
1044                 }
1045                 if (nonblock) {
1046                         ret = -EAGAIN;
1047                         goto out;
1048                 }
1049
1050                 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
1051                                         rds_send_queue_rm(rs, conn, rm,
1052                                                           rs->rs_bound_port,
1053                                                           dport,
1054                                                           &queued),
1055                                         timeo);
1056                 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
1057                 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
1058                         continue;
1059
1060                 ret = timeo;
1061                 if (ret == 0)
1062                         ret = -ETIMEDOUT;
1063                 goto out;
1064         }
1065
1066         /*
1067          * By now we've committed to the send.  We reuse rds_send_worker()
1068          * to retry sends in the rds thread if the transport asks us to.
1069          */
1070         rds_stats_inc(s_send_queued);
1071
1072         if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
1073                 rds_send_worker(&conn->c_send_w.work);
1074
1075         rds_message_put(rm);
1076         return payload_len;
1077
1078 out:
1079         /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1080          * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1081          * or in any other way, we need to destroy the MR again */
1082         if (allocated_mr)
1083                 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
1084
1085         if (rm)
1086                 rds_message_put(rm);
1087         return ret;
1088 }
1089
1090 /*
1091  * Reply to a ping packet.
1092  */
1093 int
1094 rds_send_pong(struct rds_connection *conn, __be16 dport)
1095 {
1096         struct rds_message *rm;
1097         unsigned long flags;
1098         int ret = 0;
1099
1100         rm = rds_message_alloc(0, GFP_ATOMIC);
1101         if (!rm) {
1102                 ret = -ENOMEM;
1103                 goto out;
1104         }
1105
1106         rm->m_daddr = conn->c_faddr;
1107
1108         /* If the connection is down, trigger a connect. We may
1109          * have scheduled a delayed reconnect however - in this case
1110          * we should not interfere.
1111          */
1112         if (rds_conn_state(conn) == RDS_CONN_DOWN &&
1113             !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
1114                 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
1115
1116         ret = rds_cong_wait(conn->c_fcong, dport, 1, NULL);
1117         if (ret)
1118                 goto out;
1119
1120         spin_lock_irqsave(&conn->c_lock, flags);
1121         list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
1122         set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
1123         rds_message_addref(rm);
1124         rm->m_inc.i_conn = conn;
1125
1126         rds_message_populate_header(&rm->m_inc.i_hdr, 0, dport,
1127                                     conn->c_next_tx_seq);
1128         conn->c_next_tx_seq++;
1129         spin_unlock_irqrestore(&conn->c_lock, flags);
1130
1131         rds_stats_inc(s_send_queued);
1132         rds_stats_inc(s_send_pong);
1133
1134         queue_delayed_work(rds_wq, &conn->c_send_w, 0);
1135         rds_message_put(rm);
1136         return 0;
1137
1138 out:
1139         if (rm)
1140                 rds_message_put(rm);
1141         return ret;
1142 }