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