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