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(&conn->c_lock);
 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(&conn->c_lock);
 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(&conn->c_lock);
 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(&conn->c_lock);
 206                                 continue;
 207                         }
 208
 209                         /* Require an ACK every once in a while */
 210                         len = ntohl(rm->m_inc.i_hdr.h_len);
 211                         if (conn->c_unacked_packets == 0 ||
 212                             conn->c_unacked_bytes < len) {
 213                                 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
 214
 215                                 conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
 216                                 conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
 217                                 rds_stats_inc(s_send_ack_required);
 218                         } else {
 219                                 conn->c_unacked_bytes -= len;
 220                                 conn->c_unacked_packets--;
 221                         }
 222
 223                         conn->c_xmit_rm = rm;
 224                 }
 225
 226                 /* The transport either sends the whole rdma or none of it */
 227                 if (rm->rdma.op_active && !conn->c_xmit_rdma_sent) {
 228                         rm->m_final_op = &rm->rdma;
 229                         ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
 230                         if (ret)
 231                                 break;
 232                         conn->c_xmit_rdma_sent = 1;
 233
 234                         /* The transport owns the mapped memory for now.
 235                          * You can't unmap it while it's on the send queue */
 236                         set_bit(RDS_MSG_MAPPED, &rm->m_flags);
 237                 }
 238
 239                 if (rm->atomic.op_active && !conn->c_xmit_atomic_sent) {
 240                         rm->m_final_op = &rm->atomic;
 241                         ret = conn->c_trans->xmit_atomic(conn, &rm->atomic);
 242                         if (ret)
 243                                 break;
 244                         conn->c_xmit_atomic_sent = 1;
 245
 246                         /* The transport owns the mapped memory for now.
 247                          * You can't unmap it while it's on the send queue */
 248                         set_bit(RDS_MSG_MAPPED, &rm->m_flags);
 249                 }
 250
 251                 /*
 252                  * A number of cases require an RDS header to be sent
 253                  * even if there is no data.
 254                  * We permit 0-byte sends; rds-ping depends on this.
 255                  * However, if there are exclusively attached silent ops,
 256                  * we skip the hdr/data send, to enable silent operation.
 257                  */
 258                 if (rm->data.op_nents == 0) {
 259                         int ops_present;
 260                         int all_ops_are_silent = 1;
 261
 262                         ops_present = (rm->atomic.op_active || rm->rdma.op_active);
 263                         if (rm->atomic.op_active && !rm->atomic.op_silent)
 264                                 all_ops_are_silent = 0;
 265                         if (rm->rdma.op_active && !rm->rdma.op_silent)
 266                                 all_ops_are_silent = 0;
 267
 268                         if (ops_present && all_ops_are_silent
 269                             && !rm->m_rdma_cookie)
 270                                 rm->data.op_active = 0;
 271                 }
 272
 273                 if (rm->data.op_active && !conn->c_xmit_data_sent) {
 274                         rm->m_final_op = &rm->data;
 275                         ret = conn->c_trans->xmit(conn, rm,
 276                                                   conn->c_xmit_hdr_off,
 277                                                   conn->c_xmit_sg,
 278                                                   conn->c_xmit_data_off);
 279                         if (ret <= 0)
 280                                 break;
 281
 282                         if (conn->c_xmit_hdr_off < sizeof(struct rds_header)) {
 283                                 tmp = min_t(int, ret,
 284                                             sizeof(struct rds_header) -
 285                                             conn->c_xmit_hdr_off);
 286                                 conn->c_xmit_hdr_off += tmp;
 287                                 ret -= tmp;
 288                         }
 289
 290                         sg = &rm->data.op_sg[conn->c_xmit_sg];
 291                         while (ret) {
 292                                 tmp = min_t(int, ret, sg->length -
 293                                                       conn->c_xmit_data_off);
 294                                 conn->c_xmit_data_off += tmp;
 295                                 ret -= tmp;
 296                                 if (conn->c_xmit_data_off == sg->length) {
 297                                         conn->c_xmit_data_off = 0;
 298                                         sg++;
 299                                         conn->c_xmit_sg++;
 300                                         BUG_ON(ret != 0 &&
 301                                                conn->c_xmit_sg == rm->data.op_nents);
 302                                 }
 303                         }
 304
 305                         if (conn->c_xmit_hdr_off == sizeof(struct rds_header) &&
 306                             (conn->c_xmit_sg == rm->data.op_nents))
 307                                 conn->c_xmit_data_sent = 1;
 308                 }
 309
 310                 /*
 311                  * A rm will only take multiple times through this loop
 312                  * if there is a data op. Thus, if the data is sent (or there was
 313                  * none), then we're done with the rm.
 314                  */
 315                 if (!rm->data.op_active || conn->c_xmit_data_sent) {
 316                         conn->c_xmit_rm = NULL;
 317                         conn->c_xmit_sg = 0;
 318                         conn->c_xmit_hdr_off = 0;
 319                         conn->c_xmit_data_off = 0;
 320                         conn->c_xmit_rdma_sent = 0;
 321                         conn->c_xmit_atomic_sent = 0;
 322                         conn->c_xmit_data_sent = 0;
 323
 324                         rds_message_put(rm);
 325                 }
 326         }
 327
 328         if (conn->c_trans->xmit_complete)
 329                 conn->c_trans->xmit_complete(conn);
 330
 331         /*
 332          * We might be racing with another sender who queued a message but
 333          * backed off on noticing that we held the c_send_lock.  If we check
 334          * for queued messages after dropping the sem then either we'll
 335          * see the queued message or the queuer will get the sem.  If we
 336          * notice the queued message then we trigger an immediate retry.
 337          *
 338          * We need to be careful only to do this when we stopped processing
 339          * the send queue because it was empty.  It's the only way we
 340          * stop processing the loop when the transport hasn't taken
 341          * responsibility for forward progress.
 342          */
 343         spin_unlock_irqrestore(&conn->c_send_lock, flags);
 344
 345         /* Nuke any messages we decided not to retransmit. */
 346         if (!list_empty(&to_be_dropped)) {
 347                 /* irqs on here, so we can put(), unlike above */
 348                 list_for_each_entry(rm, &to_be_dropped, m_conn_item)
 349                         rds_message_put(rm);
 350                 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
 351         }
 352
 353         if (send_quota == 0 && !was_empty) {
 354                 /* We exhausted the send quota, but there's work left to
 355                  * do. Return and (re-)schedule the send worker.
 356                  */
 357                 ret = -EAGAIN;
 358         }
 359
 360         if (ret == 0 && was_empty) {
 361                 /* A simple bit test would be way faster than taking the
 362                  * spin lock */
 363                 spin_lock_irqsave(&conn->c_lock, flags);
 364                 if (!list_empty(&conn->c_send_queue)) {
 365                         rds_stats_inc(s_send_lock_queue_raced);
 366                         ret = -EAGAIN;
 367                 }
 368                 spin_unlock_irqrestore(&conn->c_lock, flags);
 369         }
 370 out:
 371         return ret;
 372 }
 373
 374 static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
 375 {
 376         u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
 377
 378         assert_spin_locked(&rs->rs_lock);
 379
 380         BUG_ON(rs->rs_snd_bytes < len);
 381         rs->rs_snd_bytes -= len;
 382
 383         if (rs->rs_snd_bytes == 0)
 384                 rds_stats_inc(s_send_queue_empty);
 385 }
 386
 387 static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
 388                                     is_acked_func is_acked)
 389 {
 390         if (is_acked)
 391                 return is_acked(rm, ack);
 392         return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
 393 }
 394
 395 /*
 396  * Returns true if there are no messages on the send and retransmit queues
 397  * which have a sequence number greater than or equal to the given sequence
 398  * number.
 399  */
 400 int rds_send_acked_before(struct rds_connection *conn, u64 seq)
 401 {
 402         struct rds_message *rm, *tmp;
 403         int ret = 1;
 404
 405         spin_lock(&conn->c_lock);
 406
 407         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
 408                 if (be64_to_cpu(rm->m_inc.i_hdr.h_sequence) < seq)
 409                         ret = 0;
 410                 break;
 411         }
 412
 413         list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
 414                 if (be64_to_cpu(rm->m_inc.i_hdr.h_sequence) < seq)
 415                         ret = 0;
 416                 break;
 417         }
 418
 419         spin_unlock(&conn->c_lock);
 420
 421         return ret;
 422 }
 423
 424 /*
 425  * This is pretty similar to what happens below in the ACK
 426  * handling code - except that we call here as soon as we get
 427  * the IB send completion on the RDMA op and the accompanying
 428  * message.
 429  */
 430 void rds_rdma_send_complete(struct rds_message *rm, int status)
 431 {
 432         struct rds_sock *rs = NULL;
 433         struct rm_rdma_op *ro;
 434         struct rds_notifier *notifier;
 435         unsigned long flags;
 436
 437         spin_lock_irqsave(&rm->m_rs_lock, flags);
 438
 439         ro = &rm->rdma;
 440         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
 441             ro->op_active && ro->op_notify && ro->op_notifier) {
 442                 notifier = ro->op_notifier;
 443                 rs = rm->m_rs;
 444                 sock_hold(rds_rs_to_sk(rs));
 445
 446                 notifier->n_status = status;
 447                 spin_lock(&rs->rs_lock);
 448                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
 449                 spin_unlock(&rs->rs_lock);
 450
 451                 ro->op_notifier = NULL;
 452         }
 453
 454         spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 455
 456         if (rs) {
 457                 rds_wake_sk_sleep(rs);
 458                 sock_put(rds_rs_to_sk(rs));
 459         }
 460 }
 461 EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
 462
 463 /*
 464  * Just like above, except looks at atomic op
 465  */
 466 void rds_atomic_send_complete(struct rds_message *rm, int status)
 467 {
 468         struct rds_sock *rs = NULL;
 469         struct rm_atomic_op *ao;
 470         struct rds_notifier *notifier;
 471         unsigned long flags;
 472
 473         spin_lock_irqsave(&rm->m_rs_lock, flags);
 474
 475         ao = &rm->atomic;
 476         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
 477             && ao->op_active && ao->op_notify && ao->op_notifier) {
 478                 notifier = ao->op_notifier;
 479                 rs = rm->m_rs;
 480                 sock_hold(rds_rs_to_sk(rs));
 481
 482                 notifier->n_status = status;
 483                 spin_lock(&rs->rs_lock);
 484                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
 485                 spin_unlock(&rs->rs_lock);
 486
 487                 ao->op_notifier = NULL;
 488         }
 489
 490         spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 491
 492         if (rs) {
 493                 rds_wake_sk_sleep(rs);
 494                 sock_put(rds_rs_to_sk(rs));
 495         }
 496 }
 497 EXPORT_SYMBOL_GPL(rds_atomic_send_complete);
 498
 499 /*
 500  * This is the same as rds_rdma_send_complete except we
 501  * don't do any locking - we have all the ingredients (message,
 502  * socket, socket lock) and can just move the notifier.
 503  */
 504 static inline void
 505 __rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
 506 {
 507         struct rm_rdma_op *ro;
 508         struct rm_atomic_op *ao;
 509
 510         ro = &rm->rdma;
 511         if (ro->op_active && ro->op_notify && ro->op_notifier) {
 512                 ro->op_notifier->n_status = status;
 513                 list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue);
 514                 ro->op_notifier = NULL;
 515         }
 516
 517         ao = &rm->atomic;
 518         if (ao->op_active && ao->op_notify && ao->op_notifier) {
 519                 ao->op_notifier->n_status = status;
 520                 list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue);
 521                 ao->op_notifier = NULL;
 522         }
 523
 524         /* No need to wake the app - caller does this */
 525 }
 526
 527 /*
 528  * This is called from the IB send completion when we detect
 529  * a RDMA operation that failed with remote access error.
 530  * So speed is not an issue here.
 531  */
 532 struct rds_message *rds_send_get_message(struct rds_connection *conn,
 533                                          struct rm_rdma_op *op)
 534 {
 535         struct rds_message *rm, *tmp, *found = NULL;
 536         unsigned long flags;
 537
 538         spin_lock_irqsave(&conn->c_lock, flags);
 539
 540         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
 541                 if (&rm->rdma == op) {
 542                         atomic_inc(&rm->m_refcount);
 543                         found = rm;
 544                         goto out;
 545                 }
 546         }
 547
 548         list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
 549                 if (&rm->rdma == op) {
 550                         atomic_inc(&rm->m_refcount);
 551                         found = rm;
 552                         break;
 553                 }
 554         }
 555
 556 out:
 557         spin_unlock_irqrestore(&conn->c_lock, flags);
 558
 559         return found;
 560 }
 561 EXPORT_SYMBOL_GPL(rds_send_get_message);
 562
 563 /*
 564  * This removes messages from the socket's list if they're on it.  The list
 565  * argument must be private to the caller, we must be able to modify it
 566  * without locks.  The messages must have a reference held for their
 567  * position on the list.  This function will drop that reference after
 568  * removing the messages from the 'messages' list regardless of if it found
 569  * the messages on the socket list or not.
 570  */
 571 void rds_send_remove_from_sock(struct list_head *messages, int status)
 572 {
 573         unsigned long flags;
 574         struct rds_sock *rs = NULL;
 575         struct rds_message *rm;
 576
 577         while (!list_empty(messages)) {
 578                 int was_on_sock = 0;
 579
 580                 rm = list_entry(messages->next, struct rds_message,
 581                                 m_conn_item);
 582                 list_del_init(&rm->m_conn_item);
 583
 584                 /*
 585                  * If we see this flag cleared then we're *sure* that someone
 586                  * else beat us to removing it from the sock.  If we race
 587                  * with their flag update we'll get the lock and then really
 588                  * see that the flag has been cleared.
 589                  *
 590                  * The message spinlock makes sure nobody clears rm->m_rs
 591                  * while we're messing with it. It does not prevent the
 592                  * message from being removed from the socket, though.
 593                  */
 594                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 595                 if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags))
 596                         goto unlock_and_drop;
 597
 598                 if (rs != rm->m_rs) {
 599                         if (rs) {
 600                                 rds_wake_sk_sleep(rs);
 601                                 sock_put(rds_rs_to_sk(rs));
 602                         }
 603                         rs = rm->m_rs;
 604                         sock_hold(rds_rs_to_sk(rs));
 605                 }
 606                 spin_lock(&rs->rs_lock);
 607
 608                 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
 609                         struct rm_rdma_op *ro = &rm->rdma;
 610                         struct rds_notifier *notifier;
 611
 612                         list_del_init(&rm->m_sock_item);
 613                         rds_send_sndbuf_remove(rs, rm);
 614
 615                         if (ro->op_active && ro->op_notifier &&
 616                                (ro->op_notify || (ro->op_recverr && status))) {
 617                                 notifier = ro->op_notifier;
 618                                 list_add_tail(&notifier->n_list,
 619                                                 &rs->rs_notify_queue);
 620                                 if (!notifier->n_status)
 621                                         notifier->n_status = status;
 622                                 rm->rdma.op_notifier = NULL;
 623                         }
 624                         was_on_sock = 1;
 625                         rm->m_rs = NULL;
 626                 }
 627                 spin_unlock(&rs->rs_lock);
 628
 629 unlock_and_drop:
 630                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 631                 rds_message_put(rm);
 632                 if (was_on_sock)
 633                         rds_message_put(rm);
 634         }
 635
 636         if (rs) {
 637                 rds_wake_sk_sleep(rs);
 638                 sock_put(rds_rs_to_sk(rs));
 639         }
 640 }
 641
 642 /*
 643  * Transports call here when they've determined that the receiver queued
 644  * messages up to, and including, the given sequence number.  Messages are
 645  * moved to the retrans queue when rds_send_xmit picks them off the send
 646  * queue. This means that in the TCP case, the message may not have been
 647  * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
 648  * checks the RDS_MSG_HAS_ACK_SEQ bit.
 649  *
 650  * XXX It's not clear to me how this is safely serialized with socket
 651  * destruction.  Maybe it should bail if it sees SOCK_DEAD.
 652  */
 653 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
 654                          is_acked_func is_acked)
 655 {
 656         struct rds_message *rm, *tmp;
 657         unsigned long flags;
 658         LIST_HEAD(list);
 659
 660         spin_lock_irqsave(&conn->c_lock, flags);
 661
 662         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
 663                 if (!rds_send_is_acked(rm, ack, is_acked))
 664                         break;
 665
 666                 list_move(&rm->m_conn_item, &list);
 667                 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
 668         }
 669
 670         /* order flag updates with spin locks */
 671         if (!list_empty(&list))
 672                 smp_mb__after_clear_bit();
 673
 674         spin_unlock_irqrestore(&conn->c_lock, flags);
 675
 676         /* now remove the messages from the sock list as needed */
 677         rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
 678 }
 679 EXPORT_SYMBOL_GPL(rds_send_drop_acked);
 680
 681 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
 682 {
 683         struct rds_message *rm, *tmp;
 684         struct rds_connection *conn;
 685         unsigned long flags;
 686         LIST_HEAD(list);
 687
 688         /* get all the messages we're dropping under the rs lock */
 689         spin_lock_irqsave(&rs->rs_lock, flags);
 690
 691         list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) {
 692                 if (dest && (dest->sin_addr.s_addr != rm->m_daddr ||
 693                              dest->sin_port != rm->m_inc.i_hdr.h_dport))
 694                         continue;
 695
 696                 list_move(&rm->m_sock_item, &list);
 697                 rds_send_sndbuf_remove(rs, rm);
 698                 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
 699         }
 700
 701         /* order flag updates with the rs lock */
 702         smp_mb__after_clear_bit();
 703
 704         spin_unlock_irqrestore(&rs->rs_lock, flags);
 705
 706         if (list_empty(&list))
 707                 return;
 708
 709         /* Remove the messages from the conn */
 710         list_for_each_entry(rm, &list, m_sock_item) {
 711
 712                 conn = rm->m_inc.i_conn;
 713
 714                 spin_lock_irqsave(&conn->c_lock, flags);
 715                 /*
 716                  * Maybe someone else beat us to removing rm from the conn.
 717                  * If we race with their flag update we'll get the lock and
 718                  * then really see that the flag has been cleared.
 719                  */
 720                 if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
 721                         spin_unlock_irqrestore(&conn->c_lock, flags);
 722                         continue;
 723                 }
 724                 list_del_init(&rm->m_conn_item);
 725                 spin_unlock_irqrestore(&conn->c_lock, flags);
 726
 727                 /*
 728                  * Couldn't grab m_rs_lock in top loop (lock ordering),
 729                  * but we can now.
 730                  */
 731                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 732
 733                 spin_lock(&rs->rs_lock);
 734                 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
 735                 spin_unlock(&rs->rs_lock);
 736
 737                 rm->m_rs = NULL;
 738                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 739
 740                 rds_message_put(rm);
 741         }
 742
 743         rds_wake_sk_sleep(rs);
 744
 745         while (!list_empty(&list)) {
 746                 rm = list_entry(list.next, struct rds_message, m_sock_item);
 747                 list_del_init(&rm->m_sock_item);
 748
 749                 rds_message_wait(rm);
 750                 rds_message_put(rm);
 751         }
 752 }
 753
 754 /*
 755  * we only want this to fire once so we use the callers 'queued'.  It's
 756  * possible that another thread can race with us and remove the
 757  * message from the flow with RDS_CANCEL_SENT_TO.
 758  */
 759 static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
 760                              struct rds_message *rm, __be16 sport,
 761                              __be16 dport, int *queued)
 762 {
 763         unsigned long flags;
 764         u32 len;
 765
 766         if (*queued)
 767                 goto out;
 768
 769         len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
 770
 771         /* this is the only place which holds both the socket's rs_lock
 772          * and the connection's c_lock */
 773         spin_lock_irqsave(&rs->rs_lock, flags);
 774
 775         /*
 776          * If there is a little space in sndbuf, we don't queue anything,
 777          * and userspace gets -EAGAIN. But poll() indicates there's send
 778          * room. This can lead to bad behavior (spinning) if snd_bytes isn't
 779          * freed up by incoming acks. So we check the *old* value of
 780          * rs_snd_bytes here to allow the last msg to exceed the buffer,
 781          * and poll() now knows no more data can be sent.
 782          */
 783         if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
 784                 rs->rs_snd_bytes += len;
 785
 786                 /* let recv side know we are close to send space exhaustion.
 787                  * This is probably not the optimal way to do it, as this
 788                  * means we set the flag on *all* messages as soon as our
 789                  * throughput hits a certain threshold.
 790                  */
 791                 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
 792                         __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
 793
 794                 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
 795                 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
 796                 rds_message_addref(rm);
 797                 rm->m_rs = rs;
 798
 799                 /* The code ordering is a little weird, but we're
 800                    trying to minimize the time we hold c_lock */
 801                 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
 802                 rm->m_inc.i_conn = conn;
 803                 rds_message_addref(rm);
 804
 805                 spin_lock(&conn->c_lock);
 806                 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(conn->c_next_tx_seq++);
 807                 list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
 808                 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
 809                 spin_unlock(&conn->c_lock);
 810
 811                 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
 812                          rm, len, rs, rs->rs_snd_bytes,
 813                          (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
 814
 815                 *queued = 1;
 816         }
 817
 818         spin_unlock_irqrestore(&rs->rs_lock, flags);
 819 out:
 820         return *queued;
 821 }
 822
 823 /*
 824  * rds_message is getting to be quite complicated, and we'd like to allocate
 825  * it all in one go. This figures out how big it needs to be up front.
 826  */
 827 static int rds_rm_size(struct msghdr *msg, int data_len)
 828 {
 829         struct cmsghdr *cmsg;
 830         int size = 0;
 831         int cmsg_groups = 0;
 832         int retval;
 833
 834         for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
 835                 if (!CMSG_OK(msg, cmsg))
 836                         return -EINVAL;
 837
 838                 if (cmsg->cmsg_level != SOL_RDS)
 839                         continue;
 840
 841                 switch (cmsg->cmsg_type) {
 842                 case RDS_CMSG_RDMA_ARGS:
 843                         cmsg_groups |= 1;
 844                         retval = rds_rdma_extra_size(CMSG_DATA(cmsg));
 845                         if (retval < 0)
 846                                 return retval;
 847                         size += retval;
 848
 849                         break;
 850
 851                 case RDS_CMSG_RDMA_DEST:
 852                 case RDS_CMSG_RDMA_MAP:
 853                         cmsg_groups |= 2;
 854                         /* these are valid but do no add any size */
 855                         break;
 856
 857                 case RDS_CMSG_ATOMIC_CSWP:
 858                 case RDS_CMSG_ATOMIC_FADD:
 859                         cmsg_groups |= 1;
 860                         size += sizeof(struct scatterlist);
 861                         break;
 862
 863                 default:
 864                         return -EINVAL;
 865                 }
 866
 867         }
 868
 869         size += ceil(data_len, PAGE_SIZE) * sizeof(struct scatterlist);
 870
 871         /* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */
 872         if (cmsg_groups == 3)
 873                 return -EINVAL;
 874
 875         return size;
 876 }
 877
 878 static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
 879                          struct msghdr *msg, int *allocated_mr)
 880 {
 881         struct cmsghdr *cmsg;
 882         int ret = 0;
 883
 884         for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
 885                 if (!CMSG_OK(msg, cmsg))
 886                         return -EINVAL;
 887
 888                 if (cmsg->cmsg_level != SOL_RDS)
 889                         continue;
 890
 891                 /* As a side effect, RDMA_DEST and RDMA_MAP will set
 892                  * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
 893                  */
 894                 switch (cmsg->cmsg_type) {
 895                 case RDS_CMSG_RDMA_ARGS:
 896                         ret = rds_cmsg_rdma_args(rs, rm, cmsg);
 897                         break;
 898
 899                 case RDS_CMSG_RDMA_DEST:
 900                         ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
 901                         break;
 902
 903                 case RDS_CMSG_RDMA_MAP:
 904                         ret = rds_cmsg_rdma_map(rs, rm, cmsg);
 905                         if (!ret)
 906                                 *allocated_mr = 1;
 907                         break;
 908                 case RDS_CMSG_ATOMIC_CSWP:
 909                 case RDS_CMSG_ATOMIC_FADD:
 910                         ret = rds_cmsg_atomic(rs, rm, cmsg);
 911                         break;
 912
 913                 default:
 914                         return -EINVAL;
 915                 }
 916
 917                 if (ret)
 918                         break;
 919         }
 920
 921         return ret;
 922 }
 923
 924 int rds_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
 925                 size_t payload_len)
 926 {
 927         struct sock *sk = sock->sk;
 928         struct rds_sock *rs = rds_sk_to_rs(sk);
 929         struct sockaddr_in *usin = (struct sockaddr_in *)msg->msg_name;
 930         __be32 daddr;
 931         __be16 dport;
 932         struct rds_message *rm = NULL;
 933         struct rds_connection *conn;
 934         int ret = 0;
 935         int queued = 0, allocated_mr = 0;
 936         int nonblock = msg->msg_flags & MSG_DONTWAIT;
 937         long timeo = sock_sndtimeo(sk, nonblock);
 938
 939         /* Mirror Linux UDP mirror of BSD error message compatibility */
 940         /* XXX: Perhaps MSG_MORE someday */
 941         if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT)) {
 942                 printk(KERN_INFO "msg_flags 0x%08X\n", msg->msg_flags);
 943                 ret = -EOPNOTSUPP;
 944                 goto out;
 945         }
 946
 947         if (msg->msg_namelen) {
 948                 /* XXX fail non-unicast destination IPs? */
 949                 if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) {
 950                         ret = -EINVAL;
 951                         goto out;
 952                 }
 953                 daddr = usin->sin_addr.s_addr;
 954                 dport = usin->sin_port;
 955         } else {
 956                 /* We only care about consistency with ->connect() */
 957                 lock_sock(sk);
 958                 daddr = rs->rs_conn_addr;
 959                 dport = rs->rs_conn_port;
 960                 release_sock(sk);
 961         }
 962
 963         /* racing with another thread binding seems ok here */
 964         if (daddr == 0 || rs->rs_bound_addr == 0) {
 965                 ret = -ENOTCONN; /* XXX not a great errno */
 966                 goto out;
 967         }
 968
 969         /* size of rm including all sgs */
 970         ret = rds_rm_size(msg, payload_len);
 971         if (ret < 0)
 972                 goto out;
 973
 974         rm = rds_message_alloc(ret, GFP_KERNEL);
 975         if (!rm) {
 976                 ret = -ENOMEM;
 977                 goto out;
 978         }
 979
 980         /* Attach data to the rm */
 981         if (payload_len) {
 982                 rm->data.op_sg = rds_message_alloc_sgs(rm, ceil(payload_len, PAGE_SIZE));
 983                 ret = rds_message_copy_from_user(rm, msg->msg_iov, payload_len);
 984                 if (ret)
 985                         goto out;
 986         }
 987         rm->data.op_active = 1;
 988
 989         rm->m_daddr = daddr;
 990
 991         /* rds_conn_create has a spinlock that runs with IRQ off.
 992          * Caching the conn in the socket helps a lot. */
 993         if (rs->rs_conn && rs->rs_conn->c_faddr == daddr)
 994                 conn = rs->rs_conn;
 995         else {
 996                 conn = rds_conn_create_outgoing(rs->rs_bound_addr, daddr,
 997                                         rs->rs_transport,
 998                                         sock->sk->sk_allocation);
 999                 if (IS_ERR(conn)) {
1000                         ret = PTR_ERR(conn);
1001                         goto out;
1002                 }
1003                 rs->rs_conn = conn;
1004         }
1005
1006         /* Parse any control messages the user may have included. */
1007         ret = rds_cmsg_send(rs, rm, msg, &allocated_mr);
1008         if (ret)
1009                 goto out;
1010
1011         if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) {
1012                 if (printk_ratelimit())
1013                         printk(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
1014                                &rm->rdma, conn->c_trans->xmit_rdma);
1015                 ret = -EOPNOTSUPP;
1016                 goto out;
1017         }
1018
1019         if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
1020                 if (printk_ratelimit())
1021                         printk(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
1022                                &rm->atomic, conn->c_trans->xmit_atomic);
1023                 ret = -EOPNOTSUPP;
1024                 goto out;
1025         }
1026
1027         /* If the connection is down, trigger a connect. We may
1028          * have scheduled a delayed reconnect however - in this case
1029          * we should not interfere.
1030          */
1031         if (rds_conn_state(conn) == RDS_CONN_DOWN &&
1032             !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
1033                 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
1034
1035         ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
1036         if (ret) {
1037                 rs->rs_seen_congestion = 1;
1038                 goto out;
1039         }
1040
1041         while (!rds_send_queue_rm(rs, conn, rm, rs->rs_bound_port,
1042                                   dport, &queued)) {
1043                 rds_stats_inc(s_send_queue_full);
1044                 /* XXX make sure this is reasonable */
1045                 if (payload_len > rds_sk_sndbuf(rs)) {
1046                         ret = -EMSGSIZE;
1047                         goto out;
1048                 }
1049                 if (nonblock) {
1050                         ret = -EAGAIN;
1051                         goto out;
1052                 }
1053
1054                 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
1055                                         rds_send_queue_rm(rs, conn, rm,
1056                                                           rs->rs_bound_port,
1057                                                           dport,
1058                                                           &queued),
1059                                         timeo);
1060                 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
1061                 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
1062                         continue;
1063
1064                 ret = timeo;
1065                 if (ret == 0)
1066                         ret = -ETIMEDOUT;
1067                 goto out;
1068         }
1069
1070         /*
1071          * By now we've committed to the send.  We reuse rds_send_worker()
1072          * to retry sends in the rds thread if the transport asks us to.
1073          */
1074         rds_stats_inc(s_send_queued);
1075
1076         if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
1077                 rds_send_xmit(conn);
1078
1079         rds_message_put(rm);
1080         return payload_len;
1081
1082 out:
1083         /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1084          * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1085          * or in any other way, we need to destroy the MR again */
1086         if (allocated_mr)
1087                 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
1088
1089         if (rm)
1090                 rds_message_put(rm);
1091         return ret;
1092 }
1093
1094 /*
1095  * Reply to a ping packet.
1096  */
1097 int
1098 rds_send_pong(struct rds_connection *conn, __be16 dport)
1099 {
1100         struct rds_message *rm;
1101         unsigned long flags;
1102         int ret = 0;
1103
1104         rm = rds_message_alloc(0, GFP_ATOMIC);
1105         if (!rm) {
1106                 ret = -ENOMEM;
1107                 goto out;
1108         }
1109
1110         rm->m_daddr = conn->c_faddr;
1111
1112         /* If the connection is down, trigger a connect. We may
1113          * have scheduled a delayed reconnect however - in this case
1114          * we should not interfere.
1115          */
1116         if (rds_conn_state(conn) == RDS_CONN_DOWN &&
1117             !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
1118                 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
1119
1120         ret = rds_cong_wait(conn->c_fcong, dport, 1, NULL);
1121         if (ret)
1122                 goto out;
1123
1124         spin_lock_irqsave(&conn->c_lock, flags);
1125         list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
1126         set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
1127         rds_message_addref(rm);
1128         rm->m_inc.i_conn = conn;
1129
1130         rds_message_populate_header(&rm->m_inc.i_hdr, 0, dport,
1131                                     conn->c_next_tx_seq);
1132         conn->c_next_tx_seq++;
1133         spin_unlock_irqrestore(&conn->c_lock, flags);
1134
1135         rds_stats_inc(s_send_queued);
1136         rds_stats_inc(s_send_pong);
1137
1138         queue_delayed_work(rds_wq, &conn->c_send_w, 0);
1139         rds_message_put(rm);
1140         return 0;
1141
1142 out:
1143         if (rm)
1144                 rds_message_put(rm);
1145         return ret;
1146 }